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
  37 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  38
  39 #include <linux/dma-mapping.h>
  40 #include <linux/slab.h>
  41 #include <linux/module.h>
  42 #include <rdma/ib_cache.h>
  43
  44 #include "mad_priv.h"
  45 #include "mad_rmpp.h"
  46 #include "smi.h"
  47 #include "agent.h"
  48
  49 MODULE_LICENSE("Dual BSD/GPL");
  50 MODULE_DESCRIPTION("kernel IB MAD API");
  51 MODULE_AUTHOR("Hal Rosenstock");
  52 MODULE_AUTHOR("Sean Hefty");
  53
  54 static int mad_sendq_size = IB_MAD_QP_SEND_SIZE;
  55 static int mad_recvq_size = IB_MAD_QP_RECV_SIZE;
  56
  57 module_param_named(send_queue_size, mad_sendq_size, int, 0444);
  58 MODULE_PARM_DESC(send_queue_size, "Size of send queue in number of work requests");
  59 module_param_named(recv_queue_size, mad_recvq_size, int, 0444);
  60 MODULE_PARM_DESC(recv_queue_size, "Size of receive queue in number of work requests");
  61
  62 static struct kmem_cache *ib_mad_cache;
  63
  64 static struct list_head ib_mad_port_list;
  65 static u32 ib_mad_client_id = 0;
  66
  67 /* Port list lock */
  68 static DEFINE_SPINLOCK(ib_mad_port_list_lock);
  69
  70 /* Forward declarations */
  71 static int method_in_use(struct ib_mad_mgmt_method_table **method,
  72                          struct ib_mad_reg_req *mad_reg_req);
  73 static void remove_mad_reg_req(struct ib_mad_agent_private *priv);
  74 static struct ib_mad_agent_private *find_mad_agent(
  75                                         struct ib_mad_port_private *port_priv,
  76                                         struct ib_mad *mad);
  77 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
  78                                     struct ib_mad_private *mad);
  79 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv);
  80 static void timeout_sends(struct work_struct *work);
  81 static void local_completions(struct work_struct *work);
  82 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
  83                               struct ib_mad_agent_private *agent_priv,
  84                               u8 mgmt_class);
  85 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
  86                            struct ib_mad_agent_private *agent_priv);
  87
  88 /*
  89  * Returns a ib_mad_port_private structure or NULL for a device/port
  90  * Assumes ib_mad_port_list_lock is being held
  91  */
  92 static inline struct ib_mad_port_private *
  93 __ib_get_mad_port(struct ib_device *device, int port_num)
  94 {
  95         struct ib_mad_port_private *entry;
  96
  97         list_for_each_entry(entry, &ib_mad_port_list, port_list) {
  98                 if (entry->device == device && entry->port_num == port_num)
  99                         return entry;
 100         }
 101         return NULL;
 102 }
 103
 104 /*
 105  * Wrapper function to return a ib_mad_port_private structure or NULL
 106  * for a device/port
 107  */
 108 static inline struct ib_mad_port_private *
 109 ib_get_mad_port(struct ib_device *device, int port_num)
 110 {
 111         struct ib_mad_port_private *entry;
 112         unsigned long flags;
 113
 114         spin_lock_irqsave(&ib_mad_port_list_lock, flags);
 115         entry = __ib_get_mad_port(device, port_num);
 116         spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
 117
 118         return entry;
 119 }
 120
 121 static inline u8 convert_mgmt_class(u8 mgmt_class)
 122 {
 123         /* Alias IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE to 0 */
 124         return mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE ?
 125                 0 : mgmt_class;
 126 }
 127
 128 static int get_spl_qp_index(enum ib_qp_type qp_type)
 129 {
 130         switch (qp_type)
 131         {
 132         case IB_QPT_SMI:
 133                 return 0;
 134         case IB_QPT_GSI:
 135                 return 1;
 136         default:
 137                 return -1;
 138         }
 139 }
 140
 141 static int vendor_class_index(u8 mgmt_class)
 142 {
 143         return mgmt_class - IB_MGMT_CLASS_VENDOR_RANGE2_START;
 144 }
 145
 146 static int is_vendor_class(u8 mgmt_class)
 147 {
 148         if ((mgmt_class < IB_MGMT_CLASS_VENDOR_RANGE2_START) ||
 149             (mgmt_class > IB_MGMT_CLASS_VENDOR_RANGE2_END))
 150                 return 0;
 151         return 1;
 152 }
 153
 154 static int is_vendor_oui(char *oui)
 155 {
 156         if (oui[0] || oui[1] || oui[2])
 157                 return 1;
 158         return 0;
 159 }
 160
 161 static int is_vendor_method_in_use(
 162                 struct ib_mad_mgmt_vendor_class *vendor_class,
 163                 struct ib_mad_reg_req *mad_reg_req)
 164 {
 165         struct ib_mad_mgmt_method_table *method;
 166         int i;
 167
 168         for (i = 0; i < MAX_MGMT_OUI; i++) {
 169                 if (!memcmp(vendor_class->oui[i], mad_reg_req->oui, 3)) {
 170                         method = vendor_class->method_table[i];
 171                         if (method) {
 172                                 if (method_in_use(&method, mad_reg_req))
 173                                         return 1;
 174                                 else
 175                                         break;
 176                         }
 177                 }
 178         }
 179         return 0;
 180 }
 181
 182 int ib_response_mad(struct ib_mad *mad)
 183 {
 184         return ((mad->mad_hdr.method & IB_MGMT_METHOD_RESP) ||
 185                 (mad->mad_hdr.method == IB_MGMT_METHOD_TRAP_REPRESS) ||
 186                 ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_BM) &&
 187                  (mad->mad_hdr.attr_mod & IB_BM_ATTR_MOD_RESP)));
 188 }
 189 EXPORT_SYMBOL(ib_response_mad);
 190
 191 /*
 192  * ib_register_mad_agent - Register to send/receive MADs
 193  */
 194 struct ib_mad_agent *ib_register_mad_agent(struct ib_device *device,
 195                                            u8 port_num,
 196                                            enum ib_qp_type qp_type,
 197                                            struct ib_mad_reg_req *mad_reg_req,
 198                                            u8 rmpp_version,
 199                                            ib_mad_send_handler send_handler,
 200                                            ib_mad_recv_handler recv_handler,
 201                                            void *context)
 202 {
 203         struct ib_mad_port_private *port_priv;
 204         struct ib_mad_agent *ret = ERR_PTR(-EINVAL);
 205         struct ib_mad_agent_private *mad_agent_priv;
 206         struct ib_mad_reg_req *reg_req = NULL;
 207         struct ib_mad_mgmt_class_table *class;
 208         struct ib_mad_mgmt_vendor_class_table *vendor;
 209         struct ib_mad_mgmt_vendor_class *vendor_class;
 210         struct ib_mad_mgmt_method_table *method;
 211         int ret2, qpn;
 212         unsigned long flags;
 213         u8 mgmt_class, vclass;
 214
 215         /* Validate parameters */
 216         qpn = get_spl_qp_index(qp_type);
 217         if (qpn == -1)
 218                 goto error1;
 219
 220         if (rmpp_version && rmpp_version != IB_MGMT_RMPP_VERSION)
 221                 goto error1;
 222
 223         /* Validate MAD registration request if supplied */
 224         if (mad_reg_req) {
 225                 if (mad_reg_req->mgmt_class_version >= MAX_MGMT_VERSION)
 226                         goto error1;
 227                 if (!recv_handler)
 228                         goto error1;
 229                 if (mad_reg_req->mgmt_class >= MAX_MGMT_CLASS) {
 230                         /*
 231                          * IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE is the only
 232                          * one in this range currently allowed
 233                          */
 234                         if (mad_reg_req->mgmt_class !=
 235                             IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
 236                                 goto error1;
 237                 } else if (mad_reg_req->mgmt_class == 0) {
 238                         /*
 239                          * Class 0 is reserved in IBA and is used for
 240                          * aliasing of IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE
 241                          */
 242                         goto error1;
 243                 } else if (is_vendor_class(mad_reg_req->mgmt_class)) {
 244                         /*
 245                          * If class is in "new" vendor range,
 246                          * ensure supplied OUI is not zero
 247                          */
 248                         if (!is_vendor_oui(mad_reg_req->oui))
 249                                 goto error1;
 250                 }
 251                 /* Make sure class supplied is consistent with RMPP */
 252                 if (!ib_is_mad_class_rmpp(mad_reg_req->mgmt_class)) {
 253                         if (rmpp_version)
 254                                 goto error1;
 255                 }
 256                 /* Make sure class supplied is consistent with QP type */
 257                 if (qp_type == IB_QPT_SMI) {
 258                         if ((mad_reg_req->mgmt_class !=
 259                                         IB_MGMT_CLASS_SUBN_LID_ROUTED) &&
 260                             (mad_reg_req->mgmt_class !=
 261                                         IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE))
 262                                 goto error1;
 263                 } else {
 264                         if ((mad_reg_req->mgmt_class ==
 265                                         IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
 266                             (mad_reg_req->mgmt_class ==
 267                                         IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE))
 268                                 goto error1;
 269                 }
 270         } else {
 271                 /* No registration request supplied */
 272                 if (!send_handler)
 273                         goto error1;
 274         }
 275
 276         /* Validate device and port */
 277         port_priv = ib_get_mad_port(device, port_num);
 278         if (!port_priv) {
 279                 ret = ERR_PTR(-ENODEV);
 280                 goto error1;
 281         }
 282
 283         /* Verify the QP requested is supported.  For example, Ethernet devices
 284          * will not have QP0 */
 285         if (!port_priv->qp_info[qpn].qp) {
 286                 ret = ERR_PTR(-EPROTONOSUPPORT);
 287                 goto error1;
 288         }
 289
 290         /* Allocate structures */
 291         mad_agent_priv = kzalloc(sizeof *mad_agent_priv, GFP_KERNEL);
 292         if (!mad_agent_priv) {
 293                 ret = ERR_PTR(-ENOMEM);
 294                 goto error1;
 295         }
 296
 297         mad_agent_priv->agent.mr = ib_get_dma_mr(port_priv->qp_info[qpn].qp->pd,
 298                                                  IB_ACCESS_LOCAL_WRITE);
 299         if (IS_ERR(mad_agent_priv->agent.mr)) {
 300                 ret = ERR_PTR(-ENOMEM);
 301                 goto error2;
 302         }
 303
 304         if (mad_reg_req) {
 305                 reg_req = kmemdup(mad_reg_req, sizeof *reg_req, GFP_KERNEL);
 306                 if (!reg_req) {
 307                         ret = ERR_PTR(-ENOMEM);
 308                         goto error3;
 309                 }
 310         }
 311
 312         /* Now, fill in the various structures */
 313         mad_agent_priv->qp_info = &port_priv->qp_info[qpn];
 314         mad_agent_priv->reg_req = reg_req;
 315         mad_agent_priv->agent.rmpp_version = rmpp_version;
 316         mad_agent_priv->agent.device = device;
 317         mad_agent_priv->agent.recv_handler = recv_handler;
 318         mad_agent_priv->agent.send_handler = send_handler;
 319         mad_agent_priv->agent.context = context;
 320         mad_agent_priv->agent.qp = port_priv->qp_info[qpn].qp;
 321         mad_agent_priv->agent.port_num = port_num;
 322         spin_lock_init(&mad_agent_priv->lock);
 323         INIT_LIST_HEAD(&mad_agent_priv->send_list);
 324         INIT_LIST_HEAD(&mad_agent_priv->wait_list);
 325         INIT_LIST_HEAD(&mad_agent_priv->done_list);
 326         INIT_LIST_HEAD(&mad_agent_priv->rmpp_list);
 327         INIT_DELAYED_WORK(&mad_agent_priv->timed_work, timeout_sends);
 328         INIT_LIST_HEAD(&mad_agent_priv->local_list);
 329         INIT_WORK(&mad_agent_priv->local_work, local_completions);
 330         atomic_set(&mad_agent_priv->refcount, 1);
 331         init_completion(&mad_agent_priv->comp);
 332
 333         spin_lock_irqsave(&port_priv->reg_lock, flags);
 334         mad_agent_priv->agent.hi_tid = ++ib_mad_client_id;
 335
 336         /*
 337          * Make sure MAD registration (if supplied)
 338          * is non overlapping with any existing ones
 339          */
 340         if (mad_reg_req) {
 341                 mgmt_class = convert_mgmt_class(mad_reg_req->mgmt_class);
 342                 if (!is_vendor_class(mgmt_class)) {
 343                         class = port_priv->version[mad_reg_req->
 344                                                    mgmt_class_version].class;
 345                         if (class) {
 346                                 method = class->method_table[mgmt_class];
 347                                 if (method) {
 348                                         if (method_in_use(&method,
 349                                                            mad_reg_req))
 350                                                 goto error4;
 351                                 }
 352                         }
 353                         ret2 = add_nonoui_reg_req(mad_reg_req, mad_agent_priv,
 354                                                   mgmt_class);
 355                 } else {
 356                         /* "New" vendor class range */
 357                         vendor = port_priv->version[mad_reg_req->
 358                                                     mgmt_class_version].vendor;
 359                         if (vendor) {
 360                                 vclass = vendor_class_index(mgmt_class);
 361                                 vendor_class = vendor->vendor_class[vclass];
 362                                 if (vendor_class) {
 363                                         if (is_vendor_method_in_use(
 364                                                         vendor_class,
 365                                                         mad_reg_req))
 366                                                 goto error4;
 367                                 }
 368                         }
 369                         ret2 = add_oui_reg_req(mad_reg_req, mad_agent_priv);
 370                 }
 371                 if (ret2) {
 372                         ret = ERR_PTR(ret2);
 373                         goto error4;
 374                 }
 375         }
 376
 377         /* Add mad agent into port's agent list */
 378         list_add_tail(&mad_agent_priv->agent_list, &port_priv->agent_list);
 379         spin_unlock_irqrestore(&port_priv->reg_lock, flags);
 380
 381         return &mad_agent_priv->agent;
 382
 383 error4:
 384         spin_unlock_irqrestore(&port_priv->reg_lock, flags);
 385         kfree(reg_req);
 386 error3:
 387         ib_dereg_mr(mad_agent_priv->agent.mr);
 388 error2:
 389         kfree(mad_agent_priv);
 390 error1:
 391         return ret;
 392 }
 393 EXPORT_SYMBOL(ib_register_mad_agent);
 394
 395 static inline int is_snooping_sends(int mad_snoop_flags)
 396 {
 397         return (mad_snoop_flags &
 398                 (/*IB_MAD_SNOOP_POSTED_SENDS |
 399                  IB_MAD_SNOOP_RMPP_SENDS |*/
 400                  IB_MAD_SNOOP_SEND_COMPLETIONS /*|
 401                  IB_MAD_SNOOP_RMPP_SEND_COMPLETIONS*/));
 402 }
 403
 404 static inline int is_snooping_recvs(int mad_snoop_flags)
 405 {
 406         return (mad_snoop_flags &
 407                 (IB_MAD_SNOOP_RECVS /*|
 408                  IB_MAD_SNOOP_RMPP_RECVS*/));
 409 }
 410
 411 static int register_snoop_agent(struct ib_mad_qp_info *qp_info,
 412                                 struct ib_mad_snoop_private *mad_snoop_priv)
 413 {
 414         struct ib_mad_snoop_private **new_snoop_table;
 415         unsigned long flags;
 416         int i;
 417
 418         spin_lock_irqsave(&qp_info->snoop_lock, flags);
 419         /* Check for empty slot in array. */
 420         for (i = 0; i < qp_info->snoop_table_size; i++)
 421                 if (!qp_info->snoop_table[i])
 422                         break;
 423
 424         if (i == qp_info->snoop_table_size) {
 425                 /* Grow table. */
 426                 new_snoop_table = krealloc(qp_info->snoop_table,
 427                                            sizeof mad_snoop_priv *
 428                                            (qp_info->snoop_table_size + 1),
 429                                            GFP_ATOMIC);
 430                 if (!new_snoop_table) {
 431                         i = -ENOMEM;
 432                         goto out;
 433                 }
 434
 435                 qp_info->snoop_table = new_snoop_table;
 436                 qp_info->snoop_table_size++;
 437         }
 438         qp_info->snoop_table[i] = mad_snoop_priv;
 439         atomic_inc(&qp_info->snoop_count);
 440 out:
 441         spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
 442         return i;
 443 }
 444
 445 struct ib_mad_agent *ib_register_mad_snoop(struct ib_device *device,
 446                                            u8 port_num,
 447                                            enum ib_qp_type qp_type,
 448                                            int mad_snoop_flags,
 449                                            ib_mad_snoop_handler snoop_handler,
 450                                            ib_mad_recv_handler recv_handler,
 451                                            void *context)
 452 {
 453         struct ib_mad_port_private *port_priv;
 454         struct ib_mad_agent *ret;
 455         struct ib_mad_snoop_private *mad_snoop_priv;
 456         int qpn;
 457
 458         /* Validate parameters */
 459         if ((is_snooping_sends(mad_snoop_flags) && !snoop_handler) ||
 460             (is_snooping_recvs(mad_snoop_flags) && !recv_handler)) {
 461                 ret = ERR_PTR(-EINVAL);
 462                 goto error1;
 463         }
 464         qpn = get_spl_qp_index(qp_type);
 465         if (qpn == -1) {
 466                 ret = ERR_PTR(-EINVAL);
 467                 goto error1;
 468         }
 469         port_priv = ib_get_mad_port(device, port_num);
 470         if (!port_priv) {
 471                 ret = ERR_PTR(-ENODEV);
 472                 goto error1;
 473         }
 474         /* Allocate structures */
 475         mad_snoop_priv = kzalloc(sizeof *mad_snoop_priv, GFP_KERNEL);
 476         if (!mad_snoop_priv) {
 477                 ret = ERR_PTR(-ENOMEM);
 478                 goto error1;
 479         }
 480
 481         /* Now, fill in the various structures */
 482         mad_snoop_priv->qp_info = &port_priv->qp_info[qpn];
 483         mad_snoop_priv->agent.device = device;
 484         mad_snoop_priv->agent.recv_handler = recv_handler;
 485         mad_snoop_priv->agent.snoop_handler = snoop_handler;
 486         mad_snoop_priv->agent.context = context;
 487         mad_snoop_priv->agent.qp = port_priv->qp_info[qpn].qp;
 488         mad_snoop_priv->agent.port_num = port_num;
 489         mad_snoop_priv->mad_snoop_flags = mad_snoop_flags;
 490         init_completion(&mad_snoop_priv->comp);
 491         mad_snoop_priv->snoop_index = register_snoop_agent(
 492                                                 &port_priv->qp_info[qpn],
 493                                                 mad_snoop_priv);
 494         if (mad_snoop_priv->snoop_index < 0) {
 495                 ret = ERR_PTR(mad_snoop_priv->snoop_index);
 496                 goto error2;
 497         }
 498
 499         atomic_set(&mad_snoop_priv->refcount, 1);
 500         return &mad_snoop_priv->agent;
 501
 502 error2:
 503         kfree(mad_snoop_priv);
 504 error1:
 505         return ret;
 506 }
 507 EXPORT_SYMBOL(ib_register_mad_snoop);
 508
 509 static inline void deref_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
 510 {
 511         if (atomic_dec_and_test(&mad_agent_priv->refcount))
 512                 complete(&mad_agent_priv->comp);
 513 }
 514
 515 static inline void deref_snoop_agent(struct ib_mad_snoop_private *mad_snoop_priv)
 516 {
 517         if (atomic_dec_and_test(&mad_snoop_priv->refcount))
 518                 complete(&mad_snoop_priv->comp);
 519 }
 520
 521 static void unregister_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
 522 {
 523         struct ib_mad_port_private *port_priv;
 524         unsigned long flags;
 525
 526         /* Note that we could still be handling received MADs */
 527
 528         /*
 529          * Canceling all sends results in dropping received response
 530          * MADs, preventing us from queuing additional work
 531          */
 532         cancel_mads(mad_agent_priv);
 533         port_priv = mad_agent_priv->qp_info->port_priv;
 534         cancel_delayed_work(&mad_agent_priv->timed_work);
 535
 536         spin_lock_irqsave(&port_priv->reg_lock, flags);
 537         remove_mad_reg_req(mad_agent_priv);
 538         list_del(&mad_agent_priv->agent_list);
 539         spin_unlock_irqrestore(&port_priv->reg_lock, flags);
 540
 541         flush_workqueue(port_priv->wq);
 542         ib_cancel_rmpp_recvs(mad_agent_priv);
 543
 544         deref_mad_agent(mad_agent_priv);
 545         wait_for_completion(&mad_agent_priv->comp);
 546
 547         kfree(mad_agent_priv->reg_req);
 548         ib_dereg_mr(mad_agent_priv->agent.mr);
 549         kfree(mad_agent_priv);
 550 }
 551
 552 static void unregister_mad_snoop(struct ib_mad_snoop_private *mad_snoop_priv)
 553 {
 554         struct ib_mad_qp_info *qp_info;
 555         unsigned long flags;
 556
 557         qp_info = mad_snoop_priv->qp_info;
 558         spin_lock_irqsave(&qp_info->snoop_lock, flags);
 559         qp_info->snoop_table[mad_snoop_priv->snoop_index] = NULL;
 560         atomic_dec(&qp_info->snoop_count);
 561         spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
 562
 563         deref_snoop_agent(mad_snoop_priv);
 564         wait_for_completion(&mad_snoop_priv->comp);
 565
 566         kfree(mad_snoop_priv);
 567 }
 568
 569 /*
 570  * ib_unregister_mad_agent - Unregisters a client from using MAD services
 571  */
 572 int ib_unregister_mad_agent(struct ib_mad_agent *mad_agent)
 573 {
 574         struct ib_mad_agent_private *mad_agent_priv;
 575         struct ib_mad_snoop_private *mad_snoop_priv;
 576
 577         /* If the TID is zero, the agent can only snoop. */
 578         if (mad_agent->hi_tid) {
 579                 mad_agent_priv = container_of(mad_agent,
 580                                               struct ib_mad_agent_private,
 581                                               agent);
 582                 unregister_mad_agent(mad_agent_priv);
 583         } else {
 584                 mad_snoop_priv = container_of(mad_agent,
 585                                               struct ib_mad_snoop_private,
 586                                               agent);
 587                 unregister_mad_snoop(mad_snoop_priv);
 588         }
 589         return 0;
 590 }
 591 EXPORT_SYMBOL(ib_unregister_mad_agent);
 592
 593 static void dequeue_mad(struct ib_mad_list_head *mad_list)
 594 {
 595         struct ib_mad_queue *mad_queue;
 596         unsigned long flags;
 597
 598         BUG_ON(!mad_list->mad_queue);
 599         mad_queue = mad_list->mad_queue;
 600         spin_lock_irqsave(&mad_queue->lock, flags);
 601         list_del(&mad_list->list);
 602         mad_queue->count--;
 603         spin_unlock_irqrestore(&mad_queue->lock, flags);
 604 }
 605
 606 static void snoop_send(struct ib_mad_qp_info *qp_info,
 607                        struct ib_mad_send_buf *send_buf,
 608                        struct ib_mad_send_wc *mad_send_wc,
 609                        int mad_snoop_flags)
 610 {
 611         struct ib_mad_snoop_private *mad_snoop_priv;
 612         unsigned long flags;
 613         int i;
 614
 615         spin_lock_irqsave(&qp_info->snoop_lock, flags);
 616         for (i = 0; i < qp_info->snoop_table_size; i++) {
 617                 mad_snoop_priv = qp_info->snoop_table[i];
 618                 if (!mad_snoop_priv ||
 619                     !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
 620                         continue;
 621
 622                 atomic_inc(&mad_snoop_priv->refcount);
 623                 spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
 624                 mad_snoop_priv->agent.snoop_handler(&mad_snoop_priv->agent,
 625                                                     send_buf, mad_send_wc);
 626                 deref_snoop_agent(mad_snoop_priv);
 627                 spin_lock_irqsave(&qp_info->snoop_lock, flags);
 628         }
 629         spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
 630 }
 631
 632 static void snoop_recv(struct ib_mad_qp_info *qp_info,
 633                        struct ib_mad_recv_wc *mad_recv_wc,
 634                        int mad_snoop_flags)
 635 {
 636         struct ib_mad_snoop_private *mad_snoop_priv;
 637         unsigned long flags;
 638         int i;
 639
 640         spin_lock_irqsave(&qp_info->snoop_lock, flags);
 641         for (i = 0; i < qp_info->snoop_table_size; i++) {
 642                 mad_snoop_priv = qp_info->snoop_table[i];
 643                 if (!mad_snoop_priv ||
 644                     !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
 645                         continue;
 646
 647                 atomic_inc(&mad_snoop_priv->refcount);
 648                 spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
 649                 mad_snoop_priv->agent.recv_handler(&mad_snoop_priv->agent,
 650                                                    mad_recv_wc);
 651                 deref_snoop_agent(mad_snoop_priv);
 652                 spin_lock_irqsave(&qp_info->snoop_lock, flags);
 653         }
 654         spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
 655 }
 656
 657 static void build_smp_wc(struct ib_qp *qp,
 658                          u64 wr_id, u16 slid, u16 pkey_index, u8 port_num,
 659                          struct ib_wc *wc)
 660 {
 661         memset(wc, 0, sizeof *wc);
 662         wc->wr_id = wr_id;
 663         wc->status = IB_WC_SUCCESS;
 664         wc->opcode = IB_WC_RECV;
 665         wc->pkey_index = pkey_index;
 666         wc->byte_len = sizeof(struct ib_mad) + sizeof(struct ib_grh);
 667         wc->src_qp = IB_QP0;
 668         wc->qp = qp;
 669         wc->slid = slid;
 670         wc->sl = 0;
 671         wc->dlid_path_bits = 0;
 672         wc->port_num = port_num;
 673 }
 674
 675 /*
 676  * Return 0 if SMP is to be sent
 677  * Return 1 if SMP was consumed locally (whether or not solicited)
 678  * Return < 0 if error
 679  */
 680 static int handle_outgoing_dr_smp(struct ib_mad_agent_private *mad_agent_priv,
 681                                   struct ib_mad_send_wr_private *mad_send_wr)
 682 {
 683         int ret = 0;
 684         struct ib_smp *smp = mad_send_wr->send_buf.mad;
 685         unsigned long flags;
 686         struct ib_mad_local_private *local;
 687         struct ib_mad_private *mad_priv;
 688         struct ib_mad_port_private *port_priv;
 689         struct ib_mad_agent_private *recv_mad_agent = NULL;
 690         struct ib_device *device = mad_agent_priv->agent.device;
 691         u8 port_num;
 692         struct ib_wc mad_wc;
 693         struct ib_send_wr *send_wr = &mad_send_wr->send_wr;
 694
 695         if (device->node_type == RDMA_NODE_IB_SWITCH &&
 696             smp->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
 697                 port_num = send_wr->wr.ud.port_num;
 698         else
 699                 port_num = mad_agent_priv->agent.port_num;
 700
 701         /*
 702          * Directed route handling starts if the initial LID routed part of
 703          * a request or the ending LID routed part of a response is empty.
 704          * If we are at the start of the LID routed part, don't update the
 705          * hop_ptr or hop_cnt.  See section 14.2.2, Vol 1 IB spec.
 706          */
 707         if ((ib_get_smp_direction(smp) ? smp->dr_dlid : smp->dr_slid) ==
 708              IB_LID_PERMISSIVE &&
 709              smi_handle_dr_smp_send(smp, device->node_type, port_num) ==
 710              IB_SMI_DISCARD) {
 711                 ret = -EINVAL;
 712                 dev_err(&device->dev, "Invalid directed route\n");
 713                 goto out;
 714         }
 715
 716         /* Check to post send on QP or process locally */
 717         if (smi_check_local_smp(smp, device) == IB_SMI_DISCARD &&
 718             smi_check_local_returning_smp(smp, device) == IB_SMI_DISCARD)
 719                 goto out;
 720
 721         local = kmalloc(sizeof *local, GFP_ATOMIC);
 722         if (!local) {
 723                 ret = -ENOMEM;
 724                 dev_err(&device->dev, "No memory for ib_mad_local_private\n");
 725                 goto out;
 726         }
 727         local->mad_priv = NULL;
 728         local->recv_mad_agent = NULL;
 729         mad_priv = kmem_cache_alloc(ib_mad_cache, GFP_ATOMIC);
 730         if (!mad_priv) {
 731                 ret = -ENOMEM;
 732                 dev_err(&device->dev, "No memory for local response MAD\n");
 733                 kfree(local);
 734                 goto out;
 735         }
 736
 737         build_smp_wc(mad_agent_priv->agent.qp,
 738                      send_wr->wr_id, be16_to_cpu(smp->dr_slid),
 739                      send_wr->wr.ud.pkey_index,
 740                      send_wr->wr.ud.port_num, &mad_wc);
 741
 742         /* No GRH for DR SMP */
 743         ret = device->process_mad(device, 0, port_num, &mad_wc, NULL,
 744                                   (struct ib_mad *)smp,
 745                                   (struct ib_mad *)&mad_priv->mad);
 746         switch (ret)
 747         {
 748         case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY:
 749                 if (ib_response_mad(&mad_priv->mad.mad) &&
 750                     mad_agent_priv->agent.recv_handler) {
 751                         local->mad_priv = mad_priv;
 752                         local->recv_mad_agent = mad_agent_priv;
 753                         /*
 754                          * Reference MAD agent until receive
 755                          * side of local completion handled
 756                          */
 757                         atomic_inc(&mad_agent_priv->refcount);
 758                 } else
 759                         kmem_cache_free(ib_mad_cache, mad_priv);
 760                 break;
 761         case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED:
 762                 kmem_cache_free(ib_mad_cache, mad_priv);
 763                 break;
 764         case IB_MAD_RESULT_SUCCESS:
 765                 /* Treat like an incoming receive MAD */
 766                 port_priv = ib_get_mad_port(mad_agent_priv->agent.device,
 767                                             mad_agent_priv->agent.port_num);
 768                 if (port_priv) {
 769                         memcpy(&mad_priv->mad.mad, smp, sizeof(struct ib_mad));
 770                         recv_mad_agent = find_mad_agent(port_priv,
 771                                                         &mad_priv->mad.mad);
 772                 }
 773                 if (!port_priv || !recv_mad_agent) {
 774                         /*
 775                          * No receiving agent so drop packet and
 776                          * generate send completion.
 777                          */
 778                         kmem_cache_free(ib_mad_cache, mad_priv);
 779                         break;
 780                 }
 781                 local->mad_priv = mad_priv;
 782                 local->recv_mad_agent = recv_mad_agent;
 783                 break;
 784         default:
 785                 kmem_cache_free(ib_mad_cache, mad_priv);
 786                 kfree(local);
 787                 ret = -EINVAL;
 788                 goto out;
 789         }
 790
 791         local->mad_send_wr = mad_send_wr;
 792         /* Reference MAD agent until send side of local completion handled */
 793         atomic_inc(&mad_agent_priv->refcount);
 794         /* Queue local completion to local list */
 795         spin_lock_irqsave(&mad_agent_priv->lock, flags);
 796         list_add_tail(&local->completion_list, &mad_agent_priv->local_list);
 797         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
 798         queue_work(mad_agent_priv->qp_info->port_priv->wq,
 799                    &mad_agent_priv->local_work);
 800         ret = 1;
 801 out:
 802         return ret;
 803 }
 804
 805 static int get_pad_size(int hdr_len, int data_len)
 806 {
 807         int seg_size, pad;
 808
 809         seg_size = sizeof(struct ib_mad) - hdr_len;
 810         if (data_len && seg_size) {
 811                 pad = seg_size - data_len % seg_size;
 812                 return pad == seg_size ? 0 : pad;
 813         } else
 814                 return seg_size;
 815 }
 816
 817 static void free_send_rmpp_list(struct ib_mad_send_wr_private *mad_send_wr)
 818 {
 819         struct ib_rmpp_segment *s, *t;
 820
 821         list_for_each_entry_safe(s, t, &mad_send_wr->rmpp_list, list) {
 822                 list_del(&s->list);
 823                 kfree(s);
 824         }
 825 }
 826
 827 static int alloc_send_rmpp_list(struct ib_mad_send_wr_private *send_wr,
 828                                 gfp_t gfp_mask)
 829 {
 830         struct ib_mad_send_buf *send_buf = &send_wr->send_buf;
 831         struct ib_rmpp_mad *rmpp_mad = send_buf->mad;
 832         struct ib_rmpp_segment *seg = NULL;
 833         int left, seg_size, pad;
 834
 835         send_buf->seg_size = sizeof (struct ib_mad) - send_buf->hdr_len;
 836         seg_size = send_buf->seg_size;
 837         pad = send_wr->pad;
 838
 839         /* Allocate data segments. */
 840         for (left = send_buf->data_len + pad; left > 0; left -= seg_size) {
 841                 seg = kmalloc(sizeof (*seg) + seg_size, gfp_mask);
 842                 if (!seg) {
 843                         dev_err(&send_buf->mad_agent->device->dev,
 844                                 "alloc_send_rmpp_segs: RMPP mem alloc failed for len %zd, gfp %#x\n",
 845                                 sizeof (*seg) + seg_size, gfp_mask);
 846                         free_send_rmpp_list(send_wr);
 847                         return -ENOMEM;
 848                 }
 849                 seg->num = ++send_buf->seg_count;
 850                 list_add_tail(&seg->list, &send_wr->rmpp_list);
 851         }
 852
 853         /* Zero any padding */
 854         if (pad)
 855                 memset(seg->data + seg_size - pad, 0, pad);
 856
 857         rmpp_mad->rmpp_hdr.rmpp_version = send_wr->mad_agent_priv->
 858                                           agent.rmpp_version;
 859         rmpp_mad->rmpp_hdr.rmpp_type = IB_MGMT_RMPP_TYPE_DATA;
 860         ib_set_rmpp_flags(&rmpp_mad->rmpp_hdr, IB_MGMT_RMPP_FLAG_ACTIVE);
 861
 862         send_wr->cur_seg = container_of(send_wr->rmpp_list.next,
 863                                         struct ib_rmpp_segment, list);
 864         send_wr->last_ack_seg = send_wr->cur_seg;
 865         return 0;
 866 }
 867
 868 struct ib_mad_send_buf * ib_create_send_mad(struct ib_mad_agent *mad_agent,
 869                                             u32 remote_qpn, u16 pkey_index,
 870                                             int rmpp_active,
 871                                             int hdr_len, int data_len,
 872                                             gfp_t gfp_mask)
 873 {
 874         struct ib_mad_agent_private *mad_agent_priv;
 875         struct ib_mad_send_wr_private *mad_send_wr;
 876         int pad, message_size, ret, size;
 877         void *buf;
 878
 879         mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
 880                                       agent);
 881         pad = get_pad_size(hdr_len, data_len);
 882         message_size = hdr_len + data_len + pad;
 883
 884         if ((!mad_agent->rmpp_version &&
 885              (rmpp_active || message_size > sizeof(struct ib_mad))) ||
 886             (!rmpp_active && message_size > sizeof(struct ib_mad)))
 887                 return ERR_PTR(-EINVAL);
 888
 889         size = rmpp_active ? hdr_len : sizeof(struct ib_mad);
 890         buf = kzalloc(sizeof *mad_send_wr + size, gfp_mask);
 891         if (!buf)
 892                 return ERR_PTR(-ENOMEM);
 893
 894         mad_send_wr = buf + size;
 895         INIT_LIST_HEAD(&mad_send_wr->rmpp_list);
 896         mad_send_wr->send_buf.mad = buf;
 897         mad_send_wr->send_buf.hdr_len = hdr_len;
 898         mad_send_wr->send_buf.data_len = data_len;
 899         mad_send_wr->pad = pad;
 900
 901         mad_send_wr->mad_agent_priv = mad_agent_priv;
 902         mad_send_wr->sg_list[0].length = hdr_len;
 903         mad_send_wr->sg_list[0].lkey = mad_agent->mr->lkey;
 904         mad_send_wr->sg_list[1].length = sizeof(struct ib_mad) - hdr_len;
 905         mad_send_wr->sg_list[1].lkey = mad_agent->mr->lkey;
 906
 907         mad_send_wr->send_wr.wr_id = (unsigned long) mad_send_wr;
 908         mad_send_wr->send_wr.sg_list = mad_send_wr->sg_list;
 909         mad_send_wr->send_wr.num_sge = 2;
 910         mad_send_wr->send_wr.opcode = IB_WR_SEND;
 911         mad_send_wr->send_wr.send_flags = IB_SEND_SIGNALED;
 912         mad_send_wr->send_wr.wr.ud.remote_qpn = remote_qpn;
 913         mad_send_wr->send_wr.wr.ud.remote_qkey = IB_QP_SET_QKEY;
 914         mad_send_wr->send_wr.wr.ud.pkey_index = pkey_index;
 915
 916         if (rmpp_active) {
 917                 ret = alloc_send_rmpp_list(mad_send_wr, gfp_mask);
 918                 if (ret) {
 919                         kfree(buf);
 920                         return ERR_PTR(ret);
 921                 }
 922         }
 923
 924         mad_send_wr->send_buf.mad_agent = mad_agent;
 925         atomic_inc(&mad_agent_priv->refcount);
 926         return &mad_send_wr->send_buf;
 927 }
 928 EXPORT_SYMBOL(ib_create_send_mad);
 929
 930 int ib_get_mad_data_offset(u8 mgmt_class)
 931 {
 932         if (mgmt_class == IB_MGMT_CLASS_SUBN_ADM)
 933                 return IB_MGMT_SA_HDR;
 934         else if ((mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
 935                  (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
 936                  (mgmt_class == IB_MGMT_CLASS_BIS))
 937                 return IB_MGMT_DEVICE_HDR;
 938         else if ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
 939                  (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END))
 940                 return IB_MGMT_VENDOR_HDR;
 941         else
 942                 return IB_MGMT_MAD_HDR;
 943 }
 944 EXPORT_SYMBOL(ib_get_mad_data_offset);
 945
 946 int ib_is_mad_class_rmpp(u8 mgmt_class)
 947 {
 948         if ((mgmt_class == IB_MGMT_CLASS_SUBN_ADM) ||
 949             (mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
 950             (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
 951             (mgmt_class == IB_MGMT_CLASS_BIS) ||
 952             ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
 953              (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END)))
 954                 return 1;
 955         return 0;
 956 }
 957 EXPORT_SYMBOL(ib_is_mad_class_rmpp);
 958
 959 void *ib_get_rmpp_segment(struct ib_mad_send_buf *send_buf, int seg_num)
 960 {
 961         struct ib_mad_send_wr_private *mad_send_wr;
 962         struct list_head *list;
 963
 964         mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
 965                                    send_buf);
 966         list = &mad_send_wr->cur_seg->list;
 967
 968         if (mad_send_wr->cur_seg->num < seg_num) {
 969                 list_for_each_entry(mad_send_wr->cur_seg, list, list)
 970                         if (mad_send_wr->cur_seg->num == seg_num)
 971                                 break;
 972         } else if (mad_send_wr->cur_seg->num > seg_num) {
 973                 list_for_each_entry_reverse(mad_send_wr->cur_seg, list, list)
 974                         if (mad_send_wr->cur_seg->num == seg_num)
 975                                 break;
 976         }
 977         return mad_send_wr->cur_seg->data;
 978 }
 979 EXPORT_SYMBOL(ib_get_rmpp_segment);
 980
 981 static inline void *ib_get_payload(struct ib_mad_send_wr_private *mad_send_wr)
 982 {
 983         if (mad_send_wr->send_buf.seg_count)
 984                 return ib_get_rmpp_segment(&mad_send_wr->send_buf,
 985                                            mad_send_wr->seg_num);
 986         else
 987                 return mad_send_wr->send_buf.mad +
 988                        mad_send_wr->send_buf.hdr_len;
 989 }
 990
 991 void ib_free_send_mad(struct ib_mad_send_buf *send_buf)
 992 {
 993         struct ib_mad_agent_private *mad_agent_priv;
 994         struct ib_mad_send_wr_private *mad_send_wr;
 995
 996         mad_agent_priv = container_of(send_buf->mad_agent,
 997                                       struct ib_mad_agent_private, agent);
 998         mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
 999                                    send_buf);
1000
1001         free_send_rmpp_list(mad_send_wr);
1002         kfree(send_buf->mad);
1003         deref_mad_agent(mad_agent_priv);
1004 }
1005 EXPORT_SYMBOL(ib_free_send_mad);
1006
1007 int ib_send_mad(struct ib_mad_send_wr_private *mad_send_wr)
1008 {
1009         struct ib_mad_qp_info *qp_info;
1010         struct list_head *list;
1011         struct ib_send_wr *bad_send_wr;
1012         struct ib_mad_agent *mad_agent;
1013         struct ib_sge *sge;
1014         unsigned long flags;
1015         int ret;
1016
1017         /* Set WR ID to find mad_send_wr upon completion */
1018         qp_info = mad_send_wr->mad_agent_priv->qp_info;
1019         mad_send_wr->send_wr.wr_id = (unsigned long)&mad_send_wr->mad_list;
1020         mad_send_wr->mad_list.mad_queue = &qp_info->send_queue;
1021
1022         mad_agent = mad_send_wr->send_buf.mad_agent;
1023         sge = mad_send_wr->sg_list;
1024         sge[0].addr = ib_dma_map_single(mad_agent->device,
1025                                         mad_send_wr->send_buf.mad,
1026                                         sge[0].length,
1027                                         DMA_TO_DEVICE);
1028         if (unlikely(ib_dma_mapping_error(mad_agent->device, sge[0].addr)))
1029                 return -ENOMEM;
1030
1031         mad_send_wr->header_mapping = sge[0].addr;
1032
1033         sge[1].addr = ib_dma_map_single(mad_agent->device,
1034                                         ib_get_payload(mad_send_wr),
1035                                         sge[1].length,
1036                                         DMA_TO_DEVICE);
1037         if (unlikely(ib_dma_mapping_error(mad_agent->device, sge[1].addr))) {
1038                 ib_dma_unmap_single(mad_agent->device,
1039                                     mad_send_wr->header_mapping,
1040                                     sge[0].length, DMA_TO_DEVICE);
1041                 return -ENOMEM;
1042         }
1043         mad_send_wr->payload_mapping = sge[1].addr;
1044
1045         spin_lock_irqsave(&qp_info->send_queue.lock, flags);
1046         if (qp_info->send_queue.count < qp_info->send_queue.max_active) {
1047                 ret = ib_post_send(mad_agent->qp, &mad_send_wr->send_wr,
1048                                    &bad_send_wr);
1049                 list = &qp_info->send_queue.list;
1050         } else {
1051                 ret = 0;
1052                 list = &qp_info->overflow_list;
1053         }
1054
1055         if (!ret) {
1056                 qp_info->send_queue.count++;
1057                 list_add_tail(&mad_send_wr->mad_list.list, list);
1058         }
1059         spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
1060         if (ret) {
1061                 ib_dma_unmap_single(mad_agent->device,
1062                                     mad_send_wr->header_mapping,
1063                                     sge[0].length, DMA_TO_DEVICE);
1064                 ib_dma_unmap_single(mad_agent->device,
1065                                     mad_send_wr->payload_mapping,
1066                                     sge[1].length, DMA_TO_DEVICE);
1067         }
1068         return ret;
1069 }
1070
1071 /*
1072  * ib_post_send_mad - Posts MAD(s) to the send queue of the QP associated
1073  *  with the registered client
1074  */
1075 int ib_post_send_mad(struct ib_mad_send_buf *send_buf,
1076                      struct ib_mad_send_buf **bad_send_buf)
1077 {
1078         struct ib_mad_agent_private *mad_agent_priv;
1079         struct ib_mad_send_buf *next_send_buf;
1080         struct ib_mad_send_wr_private *mad_send_wr;
1081         unsigned long flags;
1082         int ret = -EINVAL;
1083
1084         /* Walk list of send WRs and post each on send list */
1085         for (; send_buf; send_buf = next_send_buf) {
1086
1087                 mad_send_wr = container_of(send_buf,
1088                                            struct ib_mad_send_wr_private,
1089                                            send_buf);
1090                 mad_agent_priv = mad_send_wr->mad_agent_priv;
1091
1092                 if (!send_buf->mad_agent->send_handler ||
1093                     (send_buf->timeout_ms &&
1094                      !send_buf->mad_agent->recv_handler)) {
1095                         ret = -EINVAL;
1096                         goto error;
1097                 }
1098
1099                 if (!ib_is_mad_class_rmpp(((struct ib_mad_hdr *) send_buf->mad)->mgmt_class)) {
1100                         if (mad_agent_priv->agent.rmpp_version) {
1101                                 ret = -EINVAL;
1102                                 goto error;
1103                         }
1104                 }
1105
1106                 /*
1107                  * Save pointer to next work request to post in case the
1108                  * current one completes, and the user modifies the work
1109                  * request associated with the completion
1110                  */
1111                 next_send_buf = send_buf->next;
1112                 mad_send_wr->send_wr.wr.ud.ah = send_buf->ah;
1113
1114                 if (((struct ib_mad_hdr *) send_buf->mad)->mgmt_class ==
1115                     IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
1116                         ret = handle_outgoing_dr_smp(mad_agent_priv,
1117                                                      mad_send_wr);
1118                         if (ret < 0)            /* error */
1119                                 goto error;
1120                         else if (ret == 1)      /* locally consumed */
1121                                 continue;
1122                 }
1123
1124                 mad_send_wr->tid = ((struct ib_mad_hdr *) send_buf->mad)->tid;
1125                 /* Timeout will be updated after send completes */
1126                 mad_send_wr->timeout = msecs_to_jiffies(send_buf->timeout_ms);
1127                 mad_send_wr->max_retries = send_buf->retries;
1128                 mad_send_wr->retries_left = send_buf->retries;
1129                 send_buf->retries = 0;
1130                 /* Reference for work request to QP + response */
1131                 mad_send_wr->refcount = 1 + (mad_send_wr->timeout > 0);
1132                 mad_send_wr->status = IB_WC_SUCCESS;
1133
1134                 /* Reference MAD agent until send completes */
1135                 atomic_inc(&mad_agent_priv->refcount);
1136                 spin_lock_irqsave(&mad_agent_priv->lock, flags);
1137                 list_add_tail(&mad_send_wr->agent_list,
1138                               &mad_agent_priv->send_list);
1139                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1140
1141                 if (mad_agent_priv->agent.rmpp_version) {
1142                         ret = ib_send_rmpp_mad(mad_send_wr);
1143                         if (ret >= 0 && ret != IB_RMPP_RESULT_CONSUMED)
1144                                 ret = ib_send_mad(mad_send_wr);
1145                 } else
1146                         ret = ib_send_mad(mad_send_wr);
1147                 if (ret < 0) {
1148                         /* Fail send request */
1149                         spin_lock_irqsave(&mad_agent_priv->lock, flags);
1150                         list_del(&mad_send_wr->agent_list);
1151                         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1152                         atomic_dec(&mad_agent_priv->refcount);
1153                         goto error;
1154                 }
1155         }
1156         return 0;
1157 error:
1158         if (bad_send_buf)
1159                 *bad_send_buf = send_buf;
1160         return ret;
1161 }
1162 EXPORT_SYMBOL(ib_post_send_mad);
1163
1164 /*
1165  * ib_free_recv_mad - Returns data buffers used to receive
1166  *  a MAD to the access layer
1167  */
1168 void ib_free_recv_mad(struct ib_mad_recv_wc *mad_recv_wc)
1169 {
1170         struct ib_mad_recv_buf *mad_recv_buf, *temp_recv_buf;
1171         struct ib_mad_private_header *mad_priv_hdr;
1172         struct ib_mad_private *priv;
1173         struct list_head free_list;
1174
1175         INIT_LIST_HEAD(&free_list);
1176         list_splice_init(&mad_recv_wc->rmpp_list, &free_list);
1177
1178         list_for_each_entry_safe(mad_recv_buf, temp_recv_buf,
1179                                         &free_list, list) {
1180                 mad_recv_wc = container_of(mad_recv_buf, struct ib_mad_recv_wc,
1181                                            recv_buf);
1182                 mad_priv_hdr = container_of(mad_recv_wc,
1183                                             struct ib_mad_private_header,
1184                                             recv_wc);
1185                 priv = container_of(mad_priv_hdr, struct ib_mad_private,
1186                                     header);
1187                 kmem_cache_free(ib_mad_cache, priv);
1188         }
1189 }
1190 EXPORT_SYMBOL(ib_free_recv_mad);
1191
1192 struct ib_mad_agent *ib_redirect_mad_qp(struct ib_qp *qp,
1193                                         u8 rmpp_version,
1194                                         ib_mad_send_handler send_handler,
1195                                         ib_mad_recv_handler recv_handler,
1196                                         void *context)
1197 {
1198         return ERR_PTR(-EINVAL);        /* XXX: for now */
1199 }
1200 EXPORT_SYMBOL(ib_redirect_mad_qp);
1201
1202 int ib_process_mad_wc(struct ib_mad_agent *mad_agent,
1203                       struct ib_wc *wc)
1204 {
1205         dev_err(&mad_agent->device->dev,
1206                 "ib_process_mad_wc() not implemented yet\n");
1207         return 0;
1208 }
1209 EXPORT_SYMBOL(ib_process_mad_wc);
1210
1211 static int method_in_use(struct ib_mad_mgmt_method_table **method,
1212                          struct ib_mad_reg_req *mad_reg_req)
1213 {
1214         int i;
1215
1216         for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS) {
1217                 if ((*method)->agent[i]) {
1218                         pr_err("Method %d already in use\n", i);
1219                         return -EINVAL;
1220                 }
1221         }
1222         return 0;
1223 }
1224
1225 static int allocate_method_table(struct ib_mad_mgmt_method_table **method)
1226 {
1227         /* Allocate management method table */
1228         *method = kzalloc(sizeof **method, GFP_ATOMIC);
1229         if (!*method) {
1230                 pr_err("No memory for ib_mad_mgmt_method_table\n");
1231                 return -ENOMEM;
1232         }
1233
1234         return 0;
1235 }
1236
1237 /*
1238  * Check to see if there are any methods still in use
1239  */
1240 static int check_method_table(struct ib_mad_mgmt_method_table *method)
1241 {
1242         int i;
1243
1244         for (i = 0; i < IB_MGMT_MAX_METHODS; i++)
1245                 if (method->agent[i])
1246                         return 1;
1247         return 0;
1248 }
1249
1250 /*
1251  * Check to see if there are any method tables for this class still in use
1252  */
1253 static int check_class_table(struct ib_mad_mgmt_class_table *class)
1254 {
1255         int i;
1256
1257         for (i = 0; i < MAX_MGMT_CLASS; i++)
1258                 if (class->method_table[i])
1259                         return 1;
1260         return 0;
1261 }
1262
1263 static int check_vendor_class(struct ib_mad_mgmt_vendor_class *vendor_class)
1264 {
1265         int i;
1266
1267         for (i = 0; i < MAX_MGMT_OUI; i++)
1268                 if (vendor_class->method_table[i])
1269                         return 1;
1270         return 0;
1271 }
1272
1273 static int find_vendor_oui(struct ib_mad_mgmt_vendor_class *vendor_class,
1274                            char *oui)
1275 {
1276         int i;
1277
1278         for (i = 0; i < MAX_MGMT_OUI; i++)
1279                 /* Is there matching OUI for this vendor class ? */
1280                 if (!memcmp(vendor_class->oui[i], oui, 3))
1281                         return i;
1282
1283         return -1;
1284 }
1285
1286 static int check_vendor_table(struct ib_mad_mgmt_vendor_class_table *vendor)
1287 {
1288         int i;
1289
1290         for (i = 0; i < MAX_MGMT_VENDOR_RANGE2; i++)
1291                 if (vendor->vendor_class[i])
1292                         return 1;
1293
1294         return 0;
1295 }
1296
1297 static void remove_methods_mad_agent(struct ib_mad_mgmt_method_table *method,
1298                                      struct ib_mad_agent_private *agent)
1299 {
1300         int i;
1301
1302         /* Remove any methods for this mad agent */
1303         for (i = 0; i < IB_MGMT_MAX_METHODS; i++) {
1304                 if (method->agent[i] == agent) {
1305                         method->agent[i] = NULL;
1306                 }
1307         }
1308 }
1309
1310 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1311                               struct ib_mad_agent_private *agent_priv,
1312                               u8 mgmt_class)
1313 {
1314         struct ib_mad_port_private *port_priv;
1315         struct ib_mad_mgmt_class_table **class;
1316         struct ib_mad_mgmt_method_table **method;
1317         int i, ret;
1318
1319         port_priv = agent_priv->qp_info->port_priv;
1320         class = &port_priv->version[mad_reg_req->mgmt_class_version].class;
1321         if (!*class) {
1322                 /* Allocate management class table for "new" class version */
1323                 *class = kzalloc(sizeof **class, GFP_ATOMIC);
1324                 if (!*class) {
1325                         dev_err(&agent_priv->agent.device->dev,
1326                                 "No memory for ib_mad_mgmt_class_table\n");
1327                         ret = -ENOMEM;
1328                         goto error1;
1329                 }
1330
1331                 /* Allocate method table for this management class */
1332                 method = &(*class)->method_table[mgmt_class];
1333                 if ((ret = allocate_method_table(method)))
1334                         goto error2;
1335         } else {
1336                 method = &(*class)->method_table[mgmt_class];
1337                 if (!*method) {
1338                         /* Allocate method table for this management class */
1339                         if ((ret = allocate_method_table(method)))
1340                                 goto error1;
1341                 }
1342         }
1343
1344         /* Now, make sure methods are not already in use */
1345         if (method_in_use(method, mad_reg_req))
1346                 goto error3;
1347
1348         /* Finally, add in methods being registered */
1349         for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS)
1350                 (*method)->agent[i] = agent_priv;
1351
1352         return 0;
1353
1354 error3:
1355         /* Remove any methods for this mad agent */
1356         remove_methods_mad_agent(*method, agent_priv);
1357         /* Now, check to see if there are any methods in use */
1358         if (!check_method_table(*method)) {
1359                 /* If not, release management method table */
1360                 kfree(*method);
1361                 *method = NULL;
1362         }
1363         ret = -EINVAL;
1364         goto error1;
1365 error2:
1366         kfree(*class);
1367         *class = NULL;
1368 error1:
1369         return ret;
1370 }
1371
1372 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1373                            struct ib_mad_agent_private *agent_priv)
1374 {
1375         struct ib_mad_port_private *port_priv;
1376         struct ib_mad_mgmt_vendor_class_table **vendor_table;
1377         struct ib_mad_mgmt_vendor_class_table *vendor = NULL;
1378         struct ib_mad_mgmt_vendor_class *vendor_class = NULL;
1379         struct ib_mad_mgmt_method_table **method;
1380         int i, ret = -ENOMEM;
1381         u8 vclass;
1382
1383         /* "New" vendor (with OUI) class */
1384         vclass = vendor_class_index(mad_reg_req->mgmt_class);
1385         port_priv = agent_priv->qp_info->port_priv;
1386         vendor_table = &port_priv->version[
1387                                 mad_reg_req->mgmt_class_version].vendor;
1388         if (!*vendor_table) {
1389                 /* Allocate mgmt vendor class table for "new" class version */
1390                 vendor = kzalloc(sizeof *vendor, GFP_ATOMIC);
1391                 if (!vendor) {
1392                         dev_err(&agent_priv->agent.device->dev,
1393                                 "No memory for ib_mad_mgmt_vendor_class_table\n");
1394                         goto error1;
1395                 }
1396
1397                 *vendor_table = vendor;
1398         }
1399         if (!(*vendor_table)->vendor_class[vclass]) {
1400                 /* Allocate table for this management vendor class */
1401                 vendor_class = kzalloc(sizeof *vendor_class, GFP_ATOMIC);
1402                 if (!vendor_class) {
1403                         dev_err(&agent_priv->agent.device->dev,
1404                                 "No memory for ib_mad_mgmt_vendor_class\n");
1405                         goto error2;
1406                 }
1407
1408                 (*vendor_table)->vendor_class[vclass] = vendor_class;
1409         }
1410         for (i = 0; i < MAX_MGMT_OUI; i++) {
1411                 /* Is there matching OUI for this vendor class ? */
1412                 if (!memcmp((*vendor_table)->vendor_class[vclass]->oui[i],
1413                             mad_reg_req->oui, 3)) {
1414                         method = &(*vendor_table)->vendor_class[
1415                                                 vclass]->method_table[i];
1416                         BUG_ON(!*method);
1417                         goto check_in_use;
1418                 }
1419         }
1420         for (i = 0; i < MAX_MGMT_OUI; i++) {
1421                 /* OUI slot available ? */
1422                 if (!is_vendor_oui((*vendor_table)->vendor_class[
1423                                 vclass]->oui[i])) {
1424                         method = &(*vendor_table)->vendor_class[
1425                                 vclass]->method_table[i];
1426                         BUG_ON(*method);
1427                         /* Allocate method table for this OUI */
1428                         if ((ret = allocate_method_table(method)))
1429                                 goto error3;
1430                         memcpy((*vendor_table)->vendor_class[vclass]->oui[i],
1431                                mad_reg_req->oui, 3);
1432                         goto check_in_use;
1433                 }
1434         }
1435         dev_err(&agent_priv->agent.device->dev, "All OUI slots in use\n");
1436         goto error3;
1437
1438 check_in_use:
1439         /* Now, make sure methods are not already in use */
1440         if (method_in_use(method, mad_reg_req))
1441                 goto error4;
1442
1443         /* Finally, add in methods being registered */
1444         for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS)
1445                 (*method)->agent[i] = agent_priv;
1446
1447         return 0;
1448
1449 error4:
1450         /* Remove any methods for this mad agent */
1451         remove_methods_mad_agent(*method, agent_priv);
1452         /* Now, check to see if there are any methods in use */
1453         if (!check_method_table(*method)) {
1454                 /* If not, release management method table */
1455                 kfree(*method);
1456                 *method = NULL;
1457         }
1458         ret = -EINVAL;
1459 error3:
1460         if (vendor_class) {
1461                 (*vendor_table)->vendor_class[vclass] = NULL;
1462                 kfree(vendor_class);
1463         }
1464 error2:
1465         if (vendor) {
1466                 *vendor_table = NULL;
1467                 kfree(vendor);
1468         }
1469 error1:
1470         return ret;
1471 }
1472
1473 static void remove_mad_reg_req(struct ib_mad_agent_private *agent_priv)
1474 {
1475         struct ib_mad_port_private *port_priv;
1476         struct ib_mad_mgmt_class_table *class;
1477         struct ib_mad_mgmt_method_table *method;
1478         struct ib_mad_mgmt_vendor_class_table *vendor;
1479         struct ib_mad_mgmt_vendor_class *vendor_class;
1480         int index;
1481         u8 mgmt_class;
1482
1483         /*
1484          * Was MAD registration request supplied
1485          * with original registration ?
1486          */
1487         if (!agent_priv->reg_req) {
1488                 goto out;
1489         }
1490
1491         port_priv = agent_priv->qp_info->port_priv;
1492         mgmt_class = convert_mgmt_class(agent_priv->reg_req->mgmt_class);
1493         class = port_priv->version[
1494                         agent_priv->reg_req->mgmt_class_version].class;
1495         if (!class)
1496                 goto vendor_check;
1497
1498         method = class->method_table[mgmt_class];
1499         if (method) {
1500                 /* Remove any methods for this mad agent */
1501                 remove_methods_mad_agent(method, agent_priv);
1502                 /* Now, check to see if there are any methods still in use */
1503                 if (!check_method_table(method)) {
1504                         /* If not, release management method table */
1505                          kfree(method);
1506                          class->method_table[mgmt_class] = NULL;
1507                          /* Any management classes left ? */
1508                         if (!check_class_table(class)) {
1509                                 /* If not, release management class table */
1510                                 kfree(class);
1511                                 port_priv->version[
1512                                         agent_priv->reg_req->
1513                                         mgmt_class_version].class = NULL;
1514                         }
1515                 }
1516         }
1517
1518 vendor_check:
1519         if (!is_vendor_class(mgmt_class))
1520                 goto out;
1521
1522         /* normalize mgmt_class to vendor range 2 */
1523         mgmt_class = vendor_class_index(agent_priv->reg_req->mgmt_class);
1524         vendor = port_priv->version[
1525                         agent_priv->reg_req->mgmt_class_version].vendor;
1526
1527         if (!vendor)
1528                 goto out;
1529
1530         vendor_class = vendor->vendor_class[mgmt_class];
1531         if (vendor_class) {
1532                 index = find_vendor_oui(vendor_class, agent_priv->reg_req->oui);
1533                 if (index < 0)
1534                         goto out;
1535                 method = vendor_class->method_table[index];
1536                 if (method) {
1537                         /* Remove any methods for this mad agent */
1538                         remove_methods_mad_agent(method, agent_priv);
1539                         /*
1540                          * Now, check to see if there are
1541                          * any methods still in use
1542                          */
1543                         if (!check_method_table(method)) {
1544                                 /* If not, release management method table */
1545                                 kfree(method);
1546                                 vendor_class->method_table[index] = NULL;
1547                                 memset(vendor_class->oui[index], 0, 3);
1548                                 /* Any OUIs left ? */
1549                                 if (!check_vendor_class(vendor_class)) {
1550                                         /* If not, release vendor class table */
1551                                         kfree(vendor_class);
1552                                         vendor->vendor_class[mgmt_class] = NULL;
1553                                         /* Any other vendor classes left ? */
1554                                         if (!check_vendor_table(vendor)) {
1555                                                 kfree(vendor);
1556                                                 port_priv->version[
1557                                                         agent_priv->reg_req->
1558                                                         mgmt_class_version].
1559                                                         vendor = NULL;
1560                                         }
1561                                 }
1562                         }
1563                 }
1564         }
1565
1566 out:
1567         return;
1568 }
1569
1570 static struct ib_mad_agent_private *
1571 find_mad_agent(struct ib_mad_port_private *port_priv,
1572                struct ib_mad *mad)
1573 {
1574         struct ib_mad_agent_private *mad_agent = NULL;
1575         unsigned long flags;
1576
1577         spin_lock_irqsave(&port_priv->reg_lock, flags);
1578         if (ib_response_mad(mad)) {
1579                 u32 hi_tid;
1580                 struct ib_mad_agent_private *entry;
1581
1582                 /*
1583                  * Routing is based on high 32 bits of transaction ID
1584                  * of MAD.
1585                  */
1586                 hi_tid = be64_to_cpu(mad->mad_hdr.tid) >> 32;
1587                 list_for_each_entry(entry, &port_priv->agent_list, agent_list) {
1588                         if (entry->agent.hi_tid == hi_tid) {
1589                                 mad_agent = entry;
1590                                 break;
1591                         }
1592                 }
1593         } else {
1594                 struct ib_mad_mgmt_class_table *class;
1595                 struct ib_mad_mgmt_method_table *method;
1596                 struct ib_mad_mgmt_vendor_class_table *vendor;
1597                 struct ib_mad_mgmt_vendor_class *vendor_class;
1598                 struct ib_vendor_mad *vendor_mad;
1599                 int index;
1600
1601                 /*
1602                  * Routing is based on version, class, and method
1603                  * For "newer" vendor MADs, also based on OUI
1604                  */
1605                 if (mad->mad_hdr.class_version >= MAX_MGMT_VERSION)
1606                         goto out;
1607                 if (!is_vendor_class(mad->mad_hdr.mgmt_class)) {
1608                         class = port_priv->version[
1609                                         mad->mad_hdr.class_version].class;
1610                         if (!class)
1611                                 goto out;
1612                         if (convert_mgmt_class(mad->mad_hdr.mgmt_class) >=
1613                             IB_MGMT_MAX_METHODS)
1614                                 goto out;
1615                         method = class->method_table[convert_mgmt_class(
1616                                                         mad->mad_hdr.mgmt_class)];
1617                         if (method)
1618                                 mad_agent = method->agent[mad->mad_hdr.method &
1619                                                           ~IB_MGMT_METHOD_RESP];
1620                 } else {
1621                         vendor = port_priv->version[
1622                                         mad->mad_hdr.class_version].vendor;
1623                         if (!vendor)
1624                                 goto out;
1625                         vendor_class = vendor->vendor_class[vendor_class_index(
1626                                                 mad->mad_hdr.mgmt_class)];
1627                         if (!vendor_class)
1628                                 goto out;
1629                         /* Find matching OUI */
1630                         vendor_mad = (struct ib_vendor_mad *)mad;
1631                         index = find_vendor_oui(vendor_class, vendor_mad->oui);
1632                         if (index == -1)
1633                                 goto out;
1634                         method = vendor_class->method_table[index];
1635                         if (method) {
1636                                 mad_agent = method->agent[mad->mad_hdr.method &
1637                                                           ~IB_MGMT_METHOD_RESP];
1638                         }
1639                 }
1640         }
1641
1642         if (mad_agent) {
1643                 if (mad_agent->agent.recv_handler)
1644                         atomic_inc(&mad_agent->refcount);
1645                 else {
1646                         dev_notice(&port_priv->device->dev,
1647                                    "No receive handler for client %p on port %d\n",
1648                                    &mad_agent->agent, port_priv->port_num);
1649                         mad_agent = NULL;
1650                 }
1651         }
1652 out:
1653         spin_unlock_irqrestore(&port_priv->reg_lock, flags);
1654
1655         return mad_agent;
1656 }
1657
1658 static int validate_mad(struct ib_mad *mad, u32 qp_num)
1659 {
1660         int valid = 0;
1661
1662         /* Make sure MAD base version is understood */
1663         if (mad->mad_hdr.base_version != IB_MGMT_BASE_VERSION) {
1664                 pr_err("MAD received with unsupported base version %d\n",
1665                         mad->mad_hdr.base_version);
1666                 goto out;
1667         }
1668
1669         /* Filter SMI packets sent to other than QP0 */
1670         if ((mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
1671             (mad->mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
1672                 if (qp_num == 0)
1673                         valid = 1;
1674         } else {
1675                 /* Filter GSI packets sent to QP0 */
1676                 if (qp_num != 0)
1677                         valid = 1;
1678         }
1679
1680 out:
1681         return valid;
1682 }
1683
1684 static int is_data_mad(struct ib_mad_agent_private *mad_agent_priv,
1685                        struct ib_mad_hdr *mad_hdr)
1686 {
1687         struct ib_rmpp_mad *rmpp_mad;
1688
1689         rmpp_mad = (struct ib_rmpp_mad *)mad_hdr;
1690         return !mad_agent_priv->agent.rmpp_version ||
1691                 !(ib_get_rmpp_flags(&rmpp_mad->rmpp_hdr) &
1692                                     IB_MGMT_RMPP_FLAG_ACTIVE) ||
1693                 (rmpp_mad->rmpp_hdr.rmpp_type == IB_MGMT_RMPP_TYPE_DATA);
1694 }
1695
1696 static inline int rcv_has_same_class(struct ib_mad_send_wr_private *wr,
1697                                      struct ib_mad_recv_wc *rwc)
1698 {
1699         return ((struct ib_mad *)(wr->send_buf.mad))->mad_hdr.mgmt_class ==
1700                 rwc->recv_buf.mad->mad_hdr.mgmt_class;
1701 }
1702
1703 static inline int rcv_has_same_gid(struct ib_mad_agent_private *mad_agent_priv,
1704                                    struct ib_mad_send_wr_private *wr,
1705                                    struct ib_mad_recv_wc *rwc )
1706 {
1707         struct ib_ah_attr attr;
1708         u8 send_resp, rcv_resp;
1709         union ib_gid sgid;
1710         struct ib_device *device = mad_agent_priv->agent.device;
1711         u8 port_num = mad_agent_priv->agent.port_num;
1712         u8 lmc;
1713
1714         send_resp = ib_response_mad((struct ib_mad *)wr->send_buf.mad);
1715         rcv_resp = ib_response_mad(rwc->recv_buf.mad);
1716
1717         if (send_resp == rcv_resp)
1718                 /* both requests, or both responses. GIDs different */
1719                 return 0;
1720
1721         if (ib_query_ah(wr->send_buf.ah, &attr))
1722                 /* Assume not equal, to avoid false positives. */
1723                 return 0;
1724
1725         if (!!(attr.ah_flags & IB_AH_GRH) !=
1726             !!(rwc->wc->wc_flags & IB_WC_GRH))
1727                 /* one has GID, other does not.  Assume different */
1728                 return 0;
1729
1730         if (!send_resp && rcv_resp) {
1731                 /* is request/response. */
1732                 if (!(attr.ah_flags & IB_AH_GRH)) {
1733                         if (ib_get_cached_lmc(device, port_num, &lmc))
1734                                 return 0;
1735                         return (!lmc || !((attr.src_path_bits ^
1736                                            rwc->wc->dlid_path_bits) &
1737                                           ((1 << lmc) - 1)));
1738                 } else {
1739                         if (ib_get_cached_gid(device, port_num,
1740                                               attr.grh.sgid_index, &sgid))
1741                                 return 0;
1742                         return !memcmp(sgid.raw, rwc->recv_buf.grh->dgid.raw,
1743                                        16);
1744                 }
1745         }
1746
1747         if (!(attr.ah_flags & IB_AH_GRH))
1748                 return attr.dlid == rwc->wc->slid;
1749         else
1750                 return !memcmp(attr.grh.dgid.raw, rwc->recv_buf.grh->sgid.raw,
1751                                16);
1752 }
1753
1754 static inline int is_direct(u8 class)
1755 {
1756         return (class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE);
1757 }
1758
1759 struct ib_mad_send_wr_private*
1760 ib_find_send_mad(struct ib_mad_agent_private *mad_agent_priv,
1761                  struct ib_mad_recv_wc *wc)
1762 {
1763         struct ib_mad_send_wr_private *wr;
1764         struct ib_mad *mad;
1765
1766         mad = (struct ib_mad *)wc->recv_buf.mad;
1767
1768         list_for_each_entry(wr, &mad_agent_priv->wait_list, agent_list) {
1769                 if ((wr->tid == mad->mad_hdr.tid) &&
1770                     rcv_has_same_class(wr, wc) &&
1771                     /*
1772                      * Don't check GID for direct routed MADs.
1773                      * These might have permissive LIDs.
1774                      */
1775                     (is_direct(wc->recv_buf.mad->mad_hdr.mgmt_class) ||
1776                      rcv_has_same_gid(mad_agent_priv, wr, wc)))
1777                         return (wr->status == IB_WC_SUCCESS) ? wr : NULL;
1778         }
1779
1780         /*
1781          * It's possible to receive the response before we've
1782          * been notified that the send has completed
1783          */
1784         list_for_each_entry(wr, &mad_agent_priv->send_list, agent_list) {
1785                 if (is_data_mad(mad_agent_priv, wr->send_buf.mad) &&
1786                     wr->tid == mad->mad_hdr.tid &&
1787                     wr->timeout &&
1788                     rcv_has_same_class(wr, wc) &&
1789                     /*
1790                      * Don't check GID for direct routed MADs.
1791                      * These might have permissive LIDs.
1792                      */
1793                     (is_direct(wc->recv_buf.mad->mad_hdr.mgmt_class) ||
1794                      rcv_has_same_gid(mad_agent_priv, wr, wc)))
1795                         /* Verify request has not been canceled */
1796                         return (wr->status == IB_WC_SUCCESS) ? wr : NULL;
1797         }
1798         return NULL;
1799 }
1800
1801 void ib_mark_mad_done(struct ib_mad_send_wr_private *mad_send_wr)
1802 {
1803         mad_send_wr->timeout = 0;
1804         if (mad_send_wr->refcount == 1)
1805                 list_move_tail(&mad_send_wr->agent_list,
1806                               &mad_send_wr->mad_agent_priv->done_list);
1807 }
1808
1809 static void ib_mad_complete_recv(struct ib_mad_agent_private *mad_agent_priv,
1810                                  struct ib_mad_recv_wc *mad_recv_wc)
1811 {
1812         struct ib_mad_send_wr_private *mad_send_wr;
1813         struct ib_mad_send_wc mad_send_wc;
1814         unsigned long flags;
1815
1816         INIT_LIST_HEAD(&mad_recv_wc->rmpp_list);
1817         list_add(&mad_recv_wc->recv_buf.list, &mad_recv_wc->rmpp_list);
1818         if (mad_agent_priv->agent.rmpp_version) {
1819                 mad_recv_wc = ib_process_rmpp_recv_wc(mad_agent_priv,
1820                                                       mad_recv_wc);
1821                 if (!mad_recv_wc) {
1822                         deref_mad_agent(mad_agent_priv);
1823                         return;
1824                 }
1825         }
1826
1827         /* Complete corresponding request */
1828         if (ib_response_mad(mad_recv_wc->recv_buf.mad)) {
1829                 spin_lock_irqsave(&mad_agent_priv->lock, flags);
1830                 mad_send_wr = ib_find_send_mad(mad_agent_priv, mad_recv_wc);
1831                 if (!mad_send_wr) {
1832                         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1833                         ib_free_recv_mad(mad_recv_wc);
1834                         deref_mad_agent(mad_agent_priv);
1835                         return;
1836                 }
1837                 ib_mark_mad_done(mad_send_wr);
1838                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1839
1840                 /* Defined behavior is to complete response before request */
1841                 mad_recv_wc->wc->wr_id = (unsigned long) &mad_send_wr->send_buf;
1842                 mad_agent_priv->agent.recv_handler(&mad_agent_priv->agent,
1843                                                    mad_recv_wc);
1844                 atomic_dec(&mad_agent_priv->refcount);
1845
1846                 mad_send_wc.status = IB_WC_SUCCESS;
1847                 mad_send_wc.vendor_err = 0;
1848                 mad_send_wc.send_buf = &mad_send_wr->send_buf;
1849                 ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
1850         } else {
1851                 mad_agent_priv->agent.recv_handler(&mad_agent_priv->agent,
1852                                                    mad_recv_wc);
1853                 deref_mad_agent(mad_agent_priv);
1854         }
1855 }
1856
1857 static bool generate_unmatched_resp(struct ib_mad_private *recv,
1858                                     struct ib_mad_private *response)
1859 {
1860         if (recv->mad.mad.mad_hdr.method == IB_MGMT_METHOD_GET ||
1861             recv->mad.mad.mad_hdr.method == IB_MGMT_METHOD_SET) {
1862                 memcpy(response, recv, sizeof *response);
1863                 response->header.recv_wc.wc = &response->header.wc;
1864                 response->header.recv_wc.recv_buf.mad = &response->mad.mad;
1865                 response->header.recv_wc.recv_buf.grh = &response->grh;
1866                 response->mad.mad.mad_hdr.method = IB_MGMT_METHOD_GET_RESP;
1867                 response->mad.mad.mad_hdr.status =
1868                         cpu_to_be16(IB_MGMT_MAD_STATUS_UNSUPPORTED_METHOD_ATTRIB);
1869                 if (recv->mad.mad.mad_hdr.mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
1870                         response->mad.mad.mad_hdr.status |= IB_SMP_DIRECTION;
1871
1872                 return true;
1873         } else {
1874                 return false;
1875         }
1876 }
1877 static void ib_mad_recv_done_handler(struct ib_mad_port_private *port_priv,
1878                                      struct ib_wc *wc)
1879 {
1880         struct ib_mad_qp_info *qp_info;
1881         struct ib_mad_private_header *mad_priv_hdr;
1882         struct ib_mad_private *recv, *response = NULL;
1883         struct ib_mad_list_head *mad_list;
1884         struct ib_mad_agent_private *mad_agent;
1885         int port_num;
1886         int ret = IB_MAD_RESULT_SUCCESS;
1887
1888         mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
1889         qp_info = mad_list->mad_queue->qp_info;
1890         dequeue_mad(mad_list);
1891
1892         mad_priv_hdr = container_of(mad_list, struct ib_mad_private_header,
1893                                     mad_list);
1894         recv = container_of(mad_priv_hdr, struct ib_mad_private, header);
1895         ib_dma_unmap_single(port_priv->device,
1896                             recv->header.mapping,
1897                             sizeof(struct ib_mad_private) -
1898                               sizeof(struct ib_mad_private_header),
1899                             DMA_FROM_DEVICE);
1900
1901         /* Setup MAD receive work completion from "normal" work completion */
1902         recv->header.wc = *wc;
1903         recv->header.recv_wc.wc = &recv->header.wc;
1904         recv->header.recv_wc.mad_len = sizeof(struct ib_mad);
1905         recv->header.recv_wc.recv_buf.mad = &recv->mad.mad;
1906         recv->header.recv_wc.recv_buf.grh = &recv->grh;
1907
1908         if (atomic_read(&qp_info->snoop_count))
1909                 snoop_recv(qp_info, &recv->header.recv_wc, IB_MAD_SNOOP_RECVS);
1910
1911         /* Validate MAD */
1912         if (!validate_mad(&recv->mad.mad, qp_info->qp->qp_num))
1913                 goto out;
1914
1915         response = kmem_cache_alloc(ib_mad_cache, GFP_KERNEL);
1916         if (!response) {
1917                 dev_err(&port_priv->device->dev,
1918                         "ib_mad_recv_done_handler no memory for response buffer\n");
1919                 goto out;
1920         }
1921
1922         if (port_priv->device->node_type == RDMA_NODE_IB_SWITCH)
1923                 port_num = wc->port_num;
1924         else
1925                 port_num = port_priv->port_num;
1926
1927         if (recv->mad.mad.mad_hdr.mgmt_class ==
1928             IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
1929                 enum smi_forward_action retsmi;
1930
1931                 if (smi_handle_dr_smp_recv(&recv->mad.smp,
1932                                            port_priv->device->node_type,
1933                                            port_num,
1934                                            port_priv->device->phys_port_cnt) ==
1935                                            IB_SMI_DISCARD)
1936                         goto out;
1937
1938                 retsmi = smi_check_forward_dr_smp(&recv->mad.smp);
1939                 if (retsmi == IB_SMI_LOCAL)
1940                         goto local;
1941
1942                 if (retsmi == IB_SMI_SEND) { /* don't forward */
1943                         if (smi_handle_dr_smp_send(&recv->mad.smp,
1944                                                    port_priv->device->node_type,
1945                                                    port_num) == IB_SMI_DISCARD)
1946                                 goto out;
1947
1948                         if (smi_check_local_smp(&recv->mad.smp, port_priv->device) == IB_SMI_DISCARD)
1949                                 goto out;
1950                 } else if (port_priv->device->node_type == RDMA_NODE_IB_SWITCH) {
1951                         /* forward case for switches */
1952                         memcpy(response, recv, sizeof(*response));
1953                         response->header.recv_wc.wc = &response->header.wc;
1954                         response->header.recv_wc.recv_buf.mad = &response->mad.mad;
1955                         response->header.recv_wc.recv_buf.grh = &response->grh;
1956
1957                         agent_send_response(&response->mad.mad,
1958                                             &response->grh, wc,
1959                                             port_priv->device,
1960                                             smi_get_fwd_port(&recv->mad.smp),
1961                                             qp_info->qp->qp_num);
1962
1963                         goto out;
1964                 }
1965         }
1966
1967 local:
1968         /* Give driver "right of first refusal" on incoming MAD */
1969         if (port_priv->device->process_mad) {
1970                 ret = port_priv->device->process_mad(port_priv->device, 0,
1971                                                      port_priv->port_num,
1972                                                      wc, &recv->grh,
1973                                                      &recv->mad.mad,
1974                                                      &response->mad.mad);
1975                 if (ret & IB_MAD_RESULT_SUCCESS) {
1976                         if (ret & IB_MAD_RESULT_CONSUMED)
1977                                 goto out;
1978                         if (ret & IB_MAD_RESULT_REPLY) {
1979                                 agent_send_response(&response->mad.mad,
1980                                                     &recv->grh, wc,
1981                                                     port_priv->device,
1982                                                     port_num,
1983                                                     qp_info->qp->qp_num);
1984                                 goto out;
1985                         }
1986                 }
1987         }
1988
1989         mad_agent = find_mad_agent(port_priv, &recv->mad.mad);
1990         if (mad_agent) {
1991                 ib_mad_complete_recv(mad_agent, &recv->header.recv_wc);
1992                 /*
1993                  * recv is freed up in error cases in ib_mad_complete_recv
1994                  * or via recv_handler in ib_mad_complete_recv()
1995                  */
1996                 recv = NULL;
1997         } else if ((ret & IB_MAD_RESULT_SUCCESS) &&
1998                    generate_unmatched_resp(recv, response)) {
1999                 agent_send_response(&response->mad.mad, &recv->grh, wc,
2000                                     port_priv->device, port_num, qp_info->qp->qp_num);
2001         }
2002
2003 out:
2004         /* Post another receive request for this QP */
2005         if (response) {
2006                 ib_mad_post_receive_mads(qp_info, response);
2007                 if (recv)
2008                         kmem_cache_free(ib_mad_cache, recv);
2009         } else
2010                 ib_mad_post_receive_mads(qp_info, recv);
2011 }
2012
2013 static void adjust_timeout(struct ib_mad_agent_private *mad_agent_priv)
2014 {
2015         struct ib_mad_send_wr_private *mad_send_wr;
2016         unsigned long delay;
2017
2018         if (list_empty(&mad_agent_priv->wait_list)) {
2019                 cancel_delayed_work(&mad_agent_priv->timed_work);
2020         } else {
2021                 mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
2022                                          struct ib_mad_send_wr_private,
2023                                          agent_list);
2024
2025                 if (time_after(mad_agent_priv->timeout,
2026                                mad_send_wr->timeout)) {
2027                         mad_agent_priv->timeout = mad_send_wr->timeout;
2028                         delay = mad_send_wr->timeout - jiffies;
2029                         if ((long)delay <= 0)
2030                                 delay = 1;
2031                         mod_delayed_work(mad_agent_priv->qp_info->port_priv->wq,
2032                                          &mad_agent_priv->timed_work, delay);
2033                 }
2034         }
2035 }
2036
2037 static void wait_for_response(struct ib_mad_send_wr_private *mad_send_wr)
2038 {
2039         struct ib_mad_agent_private *mad_agent_priv;
2040         struct ib_mad_send_wr_private *temp_mad_send_wr;
2041         struct list_head *list_item;
2042         unsigned long delay;
2043
2044         mad_agent_priv = mad_send_wr->mad_agent_priv;
2045         list_del(&mad_send_wr->agent_list);
2046
2047         delay = mad_send_wr->timeout;
2048         mad_send_wr->timeout += jiffies;
2049
2050         if (delay) {
2051                 list_for_each_prev(list_item, &mad_agent_priv->wait_list) {
2052                         temp_mad_send_wr = list_entry(list_item,
2053                                                 struct ib_mad_send_wr_private,
2054                                                 agent_list);
2055                         if (time_after(mad_send_wr->timeout,
2056                                        temp_mad_send_wr->timeout))
2057                                 break;
2058                 }
2059         }
2060         else
2061                 list_item = &mad_agent_priv->wait_list;
2062         list_add(&mad_send_wr->agent_list, list_item);
2063
2064         /* Reschedule a work item if we have a shorter timeout */
2065         if (mad_agent_priv->wait_list.next == &mad_send_wr->agent_list)
2066                 mod_delayed_work(mad_agent_priv->qp_info->port_priv->wq,
2067                                  &mad_agent_priv->timed_work, delay);
2068 }
2069
2070 void ib_reset_mad_timeout(struct ib_mad_send_wr_private *mad_send_wr,
2071                           int timeout_ms)
2072 {
2073         mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2074         wait_for_response(mad_send_wr);
2075 }
2076
2077 /*
2078  * Process a send work completion
2079  */
2080 void ib_mad_complete_send_wr(struct ib_mad_send_wr_private *mad_send_wr,
2081                              struct ib_mad_send_wc *mad_send_wc)
2082 {
2083         struct ib_mad_agent_private     *mad_agent_priv;
2084         unsigned long                   flags;
2085         int                             ret;
2086
2087         mad_agent_priv = mad_send_wr->mad_agent_priv;
2088         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2089         if (mad_agent_priv->agent.rmpp_version) {
2090                 ret = ib_process_rmpp_send_wc(mad_send_wr, mad_send_wc);
2091                 if (ret == IB_RMPP_RESULT_CONSUMED)
2092                         goto done;
2093         } else
2094                 ret = IB_RMPP_RESULT_UNHANDLED;
2095
2096         if (mad_send_wc->status != IB_WC_SUCCESS &&
2097             mad_send_wr->status == IB_WC_SUCCESS) {
2098                 mad_send_wr->status = mad_send_wc->status;
2099                 mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2100         }
2101
2102         if (--mad_send_wr->refcount > 0) {
2103                 if (mad_send_wr->refcount == 1 && mad_send_wr->timeout &&
2104                     mad_send_wr->status == IB_WC_SUCCESS) {
2105                         wait_for_response(mad_send_wr);
2106                 }
2107                 goto done;
2108         }
2109
2110         /* Remove send from MAD agent and notify client of completion */
2111         list_del(&mad_send_wr->agent_list);
2112         adjust_timeout(mad_agent_priv);
2113         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2114
2115         if (mad_send_wr->status != IB_WC_SUCCESS )
2116                 mad_send_wc->status = mad_send_wr->status;
2117         if (ret == IB_RMPP_RESULT_INTERNAL)
2118                 ib_rmpp_send_handler(mad_send_wc);
2119         else
2120                 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2121                                                    mad_send_wc);
2122
2123         /* Release reference on agent taken when sending */
2124         deref_mad_agent(mad_agent_priv);
2125         return;
2126 done:
2127         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2128 }
2129
2130 static void ib_mad_send_done_handler(struct ib_mad_port_private *port_priv,
2131                                      struct ib_wc *wc)
2132 {
2133         struct ib_mad_send_wr_private   *mad_send_wr, *queued_send_wr;
2134         struct ib_mad_list_head         *mad_list;
2135         struct ib_mad_qp_info           *qp_info;
2136         struct ib_mad_queue             *send_queue;
2137         struct ib_send_wr               *bad_send_wr;
2138         struct ib_mad_send_wc           mad_send_wc;
2139         unsigned long flags;
2140         int ret;
2141
2142         mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
2143         mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2144                                    mad_list);
2145         send_queue = mad_list->mad_queue;
2146         qp_info = send_queue->qp_info;
2147
2148 retry:
2149         ib_dma_unmap_single(mad_send_wr->send_buf.mad_agent->device,
2150                             mad_send_wr->header_mapping,
2151                             mad_send_wr->sg_list[0].length, DMA_TO_DEVICE);
2152         ib_dma_unmap_single(mad_send_wr->send_buf.mad_agent->device,
2153                             mad_send_wr->payload_mapping,
2154                             mad_send_wr->sg_list[1].length, DMA_TO_DEVICE);
2155         queued_send_wr = NULL;
2156         spin_lock_irqsave(&send_queue->lock, flags);
2157         list_del(&mad_list->list);
2158
2159         /* Move queued send to the send queue */
2160         if (send_queue->count-- > send_queue->max_active) {
2161                 mad_list = container_of(qp_info->overflow_list.next,
2162                                         struct ib_mad_list_head, list);
2163                 queued_send_wr = container_of(mad_list,
2164                                         struct ib_mad_send_wr_private,
2165                                         mad_list);
2166                 list_move_tail(&mad_list->list, &send_queue->list);
2167         }
2168         spin_unlock_irqrestore(&send_queue->lock, flags);
2169
2170         mad_send_wc.send_buf = &mad_send_wr->send_buf;
2171         mad_send_wc.status = wc->status;
2172         mad_send_wc.vendor_err = wc->vendor_err;
2173         if (atomic_read(&qp_info->snoop_count))
2174                 snoop_send(qp_info, &mad_send_wr->send_buf, &mad_send_wc,
2175                            IB_MAD_SNOOP_SEND_COMPLETIONS);
2176         ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
2177
2178         if (queued_send_wr) {
2179                 ret = ib_post_send(qp_info->qp, &queued_send_wr->send_wr,
2180                                    &bad_send_wr);
2181                 if (ret) {
2182                         dev_err(&port_priv->device->dev,
2183                                 "ib_post_send failed: %d\n", ret);
2184                         mad_send_wr = queued_send_wr;
2185                         wc->status = IB_WC_LOC_QP_OP_ERR;
2186                         goto retry;
2187                 }
2188         }
2189 }
2190
2191 static void mark_sends_for_retry(struct ib_mad_qp_info *qp_info)
2192 {
2193         struct ib_mad_send_wr_private *mad_send_wr;
2194         struct ib_mad_list_head *mad_list;
2195         unsigned long flags;
2196
2197         spin_lock_irqsave(&qp_info->send_queue.lock, flags);
2198         list_for_each_entry(mad_list, &qp_info->send_queue.list, list) {
2199                 mad_send_wr = container_of(mad_list,
2200                                            struct ib_mad_send_wr_private,
2201                                            mad_list);
2202                 mad_send_wr->retry = 1;
2203         }
2204         spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
2205 }
2206
2207 static void mad_error_handler(struct ib_mad_port_private *port_priv,
2208                               struct ib_wc *wc)
2209 {
2210         struct ib_mad_list_head *mad_list;
2211         struct ib_mad_qp_info *qp_info;
2212         struct ib_mad_send_wr_private *mad_send_wr;
2213         int ret;
2214
2215         /* Determine if failure was a send or receive */
2216         mad_list = (struct ib_mad_list_head *)(unsigned long)wc->wr_id;
2217         qp_info = mad_list->mad_queue->qp_info;
2218         if (mad_list->mad_queue == &qp_info->recv_queue)
2219                 /*
2220                  * Receive errors indicate that the QP has entered the error
2221                  * state - error handling/shutdown code will cleanup
2222                  */
2223                 return;
2224
2225         /*
2226          * Send errors will transition the QP to SQE - move
2227          * QP to RTS and repost flushed work requests
2228          */
2229         mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2230                                    mad_list);
2231         if (wc->status == IB_WC_WR_FLUSH_ERR) {
2232                 if (mad_send_wr->retry) {
2233                         /* Repost send */
2234                         struct ib_send_wr *bad_send_wr;
2235
2236                         mad_send_wr->retry = 0;
2237                         ret = ib_post_send(qp_info->qp, &mad_send_wr->send_wr,
2238                                         &bad_send_wr);
2239                         if (ret)
2240                                 ib_mad_send_done_handler(port_priv, wc);
2241                 } else
2242                         ib_mad_send_done_handler(port_priv, wc);
2243         } else {
2244                 struct ib_qp_attr *attr;
2245
2246                 /* Transition QP to RTS and fail offending send */
2247                 attr = kmalloc(sizeof *attr, GFP_KERNEL);
2248                 if (attr) {
2249                         attr->qp_state = IB_QPS_RTS;
2250                         attr->cur_qp_state = IB_QPS_SQE;
2251                         ret = ib_modify_qp(qp_info->qp, attr,
2252                                            IB_QP_STATE | IB_QP_CUR_STATE);
2253                         kfree(attr);
2254                         if (ret)
2255                                 dev_err(&port_priv->device->dev,
2256                                         "mad_error_handler - ib_modify_qp to RTS : %d\n",
2257                                         ret);
2258                         else
2259                                 mark_sends_for_retry(qp_info);
2260                 }
2261                 ib_mad_send_done_handler(port_priv, wc);
2262         }
2263 }
2264
2265 /*
2266  * IB MAD completion callback
2267  */
2268 static void ib_mad_completion_handler(struct work_struct *work)
2269 {
2270         struct ib_mad_port_private *port_priv;
2271         struct ib_wc wc;
2272
2273         port_priv = container_of(work, struct ib_mad_port_private, work);
2274         ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP);
2275
2276         while (ib_poll_cq(port_priv->cq, 1, &wc) == 1) {
2277                 if (wc.status == IB_WC_SUCCESS) {
2278                         switch (wc.opcode) {
2279                         case IB_WC_SEND:
2280                                 ib_mad_send_done_handler(port_priv, &wc);
2281                                 break;
2282                         case IB_WC_RECV:
2283                                 ib_mad_recv_done_handler(port_priv, &wc);
2284                                 break;
2285                         default:
2286                                 BUG_ON(1);
2287                                 break;
2288                         }
2289                 } else
2290                         mad_error_handler(port_priv, &wc);
2291         }
2292 }
2293
2294 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv)
2295 {
2296         unsigned long flags;
2297         struct ib_mad_send_wr_private *mad_send_wr, *temp_mad_send_wr;
2298         struct ib_mad_send_wc mad_send_wc;
2299         struct list_head cancel_list;
2300
2301         INIT_LIST_HEAD(&cancel_list);
2302
2303         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2304         list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2305                                  &mad_agent_priv->send_list, agent_list) {
2306                 if (mad_send_wr->status == IB_WC_SUCCESS) {
2307                         mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2308                         mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2309                 }
2310         }
2311
2312         /* Empty wait list to prevent receives from finding a request */
2313         list_splice_init(&mad_agent_priv->wait_list, &cancel_list);
2314         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2315
2316         /* Report all cancelled requests */
2317         mad_send_wc.status = IB_WC_WR_FLUSH_ERR;
2318         mad_send_wc.vendor_err = 0;
2319
2320         list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2321                                  &cancel_list, agent_list) {
2322                 mad_send_wc.send_buf = &mad_send_wr->send_buf;
2323                 list_del(&mad_send_wr->agent_list);
2324                 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2325                                                    &mad_send_wc);
2326                 atomic_dec(&mad_agent_priv->refcount);
2327         }
2328 }
2329
2330 static struct ib_mad_send_wr_private*
2331 find_send_wr(struct ib_mad_agent_private *mad_agent_priv,
2332              struct ib_mad_send_buf *send_buf)
2333 {
2334         struct ib_mad_send_wr_private *mad_send_wr;
2335
2336         list_for_each_entry(mad_send_wr, &mad_agent_priv->wait_list,
2337                             agent_list) {
2338                 if (&mad_send_wr->send_buf == send_buf)
2339                         return mad_send_wr;
2340         }
2341
2342         list_for_each_entry(mad_send_wr, &mad_agent_priv->send_list,
2343                             agent_list) {
2344                 if (is_data_mad(mad_agent_priv, mad_send_wr->send_buf.mad) &&
2345                     &mad_send_wr->send_buf == send_buf)
2346                         return mad_send_wr;
2347         }
2348         return NULL;
2349 }
2350
2351 int ib_modify_mad(struct ib_mad_agent *mad_agent,
2352                   struct ib_mad_send_buf *send_buf, u32 timeout_ms)
2353 {
2354         struct ib_mad_agent_private *mad_agent_priv;
2355         struct ib_mad_send_wr_private *mad_send_wr;
2356         unsigned long flags;
2357         int active;
2358
2359         mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
2360                                       agent);
2361         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2362         mad_send_wr = find_send_wr(mad_agent_priv, send_buf);
2363         if (!mad_send_wr || mad_send_wr->status != IB_WC_SUCCESS) {
2364                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2365                 return -EINVAL;
2366         }
2367
2368         active = (!mad_send_wr->timeout || mad_send_wr->refcount > 1);
2369         if (!timeout_ms) {
2370                 mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2371                 mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2372         }
2373
2374         mad_send_wr->send_buf.timeout_ms = timeout_ms;
2375         if (active)
2376                 mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2377         else
2378                 ib_reset_mad_timeout(mad_send_wr, timeout_ms);
2379
2380         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2381         return 0;
2382 }
2383 EXPORT_SYMBOL(ib_modify_mad);
2384
2385 void ib_cancel_mad(struct ib_mad_agent *mad_agent,
2386                    struct ib_mad_send_buf *send_buf)
2387 {
2388         ib_modify_mad(mad_agent, send_buf, 0);
2389 }
2390 EXPORT_SYMBOL(ib_cancel_mad);
2391
2392 static void local_completions(struct work_struct *work)
2393 {
2394         struct ib_mad_agent_private *mad_agent_priv;
2395         struct ib_mad_local_private *local;
2396         struct ib_mad_agent_private *recv_mad_agent;
2397         unsigned long flags;
2398         int free_mad;
2399         struct ib_wc wc;
2400         struct ib_mad_send_wc mad_send_wc;
2401
2402         mad_agent_priv =
2403                 container_of(work, struct ib_mad_agent_private, local_work);
2404
2405         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2406         while (!list_empty(&mad_agent_priv->local_list)) {
2407                 local = list_entry(mad_agent_priv->local_list.next,
2408                                    struct ib_mad_local_private,
2409                                    completion_list);
2410                 list_del(&local->completion_list);
2411                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2412                 free_mad = 0;
2413                 if (local->mad_priv) {
2414                         recv_mad_agent = local->recv_mad_agent;
2415                         if (!recv_mad_agent) {
2416                                 dev_err(&mad_agent_priv->agent.device->dev,
2417                                         "No receive MAD agent for local completion\n");
2418                                 free_mad = 1;
2419                                 goto local_send_completion;
2420                         }
2421
2422                         /*
2423                          * Defined behavior is to complete response
2424                          * before request
2425                          */
2426                         build_smp_wc(recv_mad_agent->agent.qp,
2427                                      (unsigned long) local->mad_send_wr,
2428                                      be16_to_cpu(IB_LID_PERMISSIVE),
2429                                      0, recv_mad_agent->agent.port_num, &wc);
2430
2431                         local->mad_priv->header.recv_wc.wc = &wc;
2432                         local->mad_priv->header.recv_wc.mad_len =
2433                                                 sizeof(struct ib_mad);
2434                         INIT_LIST_HEAD(&local->mad_priv->header.recv_wc.rmpp_list);
2435                         list_add(&local->mad_priv->header.recv_wc.recv_buf.list,
2436                                  &local->mad_priv->header.recv_wc.rmpp_list);
2437                         local->mad_priv->header.recv_wc.recv_buf.grh = NULL;
2438                         local->mad_priv->header.recv_wc.recv_buf.mad =
2439                                                 &local->mad_priv->mad.mad;
2440                         if (atomic_read(&recv_mad_agent->qp_info->snoop_count))
2441                                 snoop_recv(recv_mad_agent->qp_info,
2442                                           &local->mad_priv->header.recv_wc,
2443                                            IB_MAD_SNOOP_RECVS);
2444                         recv_mad_agent->agent.recv_handler(
2445                                                 &recv_mad_agent->agent,
2446                                                 &local->mad_priv->header.recv_wc);
2447                         spin_lock_irqsave(&recv_mad_agent->lock, flags);
2448                         atomic_dec(&recv_mad_agent->refcount);
2449                         spin_unlock_irqrestore(&recv_mad_agent->lock, flags);
2450                 }
2451
2452 local_send_completion:
2453                 /* Complete send */
2454                 mad_send_wc.status = IB_WC_SUCCESS;
2455                 mad_send_wc.vendor_err = 0;
2456                 mad_send_wc.send_buf = &local->mad_send_wr->send_buf;
2457                 if (atomic_read(&mad_agent_priv->qp_info->snoop_count))
2458                         snoop_send(mad_agent_priv->qp_info,
2459                                    &local->mad_send_wr->send_buf,
2460                                    &mad_send_wc, IB_MAD_SNOOP_SEND_COMPLETIONS);
2461                 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2462                                                    &mad_send_wc);
2463
2464                 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2465                 atomic_dec(&mad_agent_priv->refcount);
2466                 if (free_mad)
2467                         kmem_cache_free(ib_mad_cache, local->mad_priv);
2468                 kfree(local);
2469         }
2470         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2471 }
2472
2473 static int retry_send(struct ib_mad_send_wr_private *mad_send_wr)
2474 {
2475         int ret;
2476
2477         if (!mad_send_wr->retries_left)
2478                 return -ETIMEDOUT;
2479
2480         mad_send_wr->retries_left--;
2481         mad_send_wr->send_buf.retries++;
2482
2483         mad_send_wr->timeout = msecs_to_jiffies(mad_send_wr->send_buf.timeout_ms);
2484
2485         if (mad_send_wr->mad_agent_priv->agent.rmpp_version) {
2486                 ret = ib_retry_rmpp(mad_send_wr);
2487                 switch (ret) {
2488                 case IB_RMPP_RESULT_UNHANDLED:
2489                         ret = ib_send_mad(mad_send_wr);
2490                         break;
2491                 case IB_RMPP_RESULT_CONSUMED:
2492                         ret = 0;
2493                         break;
2494                 default:
2495                         ret = -ECOMM;
2496                         break;
2497                 }
2498         } else
2499                 ret = ib_send_mad(mad_send_wr);
2500
2501         if (!ret) {
2502                 mad_send_wr->refcount++;
2503                 list_add_tail(&mad_send_wr->agent_list,
2504                               &mad_send_wr->mad_agent_priv->send_list);
2505         }
2506         return ret;
2507 }
2508
2509 static void timeout_sends(struct work_struct *work)
2510 {
2511         struct ib_mad_agent_private *mad_agent_priv;
2512         struct ib_mad_send_wr_private *mad_send_wr;
2513         struct ib_mad_send_wc mad_send_wc;
2514         unsigned long flags, delay;
2515
2516         mad_agent_priv = container_of(work, struct ib_mad_agent_private,
2517                                       timed_work.work);
2518         mad_send_wc.vendor_err = 0;
2519
2520         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2521         while (!list_empty(&mad_agent_priv->wait_list)) {
2522                 mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
2523                                          struct ib_mad_send_wr_private,
2524                                          agent_list);
2525
2526                 if (time_after(mad_send_wr->timeout, jiffies)) {
2527                         delay = mad_send_wr->timeout - jiffies;
2528                         if ((long)delay <= 0)
2529                                 delay = 1;
2530                         queue_delayed_work(mad_agent_priv->qp_info->
2531                                            port_priv->wq,
2532                                            &mad_agent_priv->timed_work, delay);
2533                         break;
2534                 }
2535
2536                 list_del(&mad_send_wr->agent_list);
2537                 if (mad_send_wr->status == IB_WC_SUCCESS &&
2538                     !retry_send(mad_send_wr))
2539                         continue;
2540
2541                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2542
2543                 if (mad_send_wr->status == IB_WC_SUCCESS)
2544                         mad_send_wc.status = IB_WC_RESP_TIMEOUT_ERR;
2545                 else
2546                         mad_send_wc.status = mad_send_wr->status;
2547                 mad_send_wc.send_buf = &mad_send_wr->send_buf;
2548                 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2549                                                    &mad_send_wc);
2550
2551                 atomic_dec(&mad_agent_priv->refcount);
2552                 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2553         }
2554         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2555 }
2556
2557 static void ib_mad_thread_completion_handler(struct ib_cq *cq, void *arg)
2558 {
2559         struct ib_mad_port_private *port_priv = cq->cq_context;
2560         unsigned long flags;
2561
2562         spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2563         if (!list_empty(&port_priv->port_list))
2564                 queue_work(port_priv->wq, &port_priv->work);
2565         spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2566 }
2567
2568 /*
2569  * Allocate receive MADs and post receive WRs for them
2570  */
2571 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
2572                                     struct ib_mad_private *mad)
2573 {
2574         unsigned long flags;
2575         int post, ret;
2576         struct ib_mad_private *mad_priv;
2577         struct ib_sge sg_list;
2578         struct ib_recv_wr recv_wr, *bad_recv_wr;
2579         struct ib_mad_queue *recv_queue = &qp_info->recv_queue;
2580
2581         /* Initialize common scatter list fields */
2582         sg_list.length = sizeof *mad_priv - sizeof mad_priv->header;
2583         sg_list.lkey = (*qp_info->port_priv->mr).lkey;
2584
2585         /* Initialize common receive WR fields */
2586         recv_wr.next = NULL;
2587         recv_wr.sg_list = &sg_list;
2588         recv_wr.num_sge = 1;
2589
2590         do {
2591                 /* Allocate and map receive buffer */
2592                 if (mad) {
2593                         mad_priv = mad;
2594                         mad = NULL;
2595                 } else {
2596                         mad_priv = kmem_cache_alloc(ib_mad_cache, GFP_KERNEL);
2597                         if (!mad_priv) {
2598                                 dev_err(&qp_info->port_priv->device->dev,
2599                                         "No memory for receive buffer\n");
2600                                 ret = -ENOMEM;
2601                                 break;
2602                         }
2603                 }
2604                 sg_list.addr = ib_dma_map_single(qp_info->port_priv->device,
2605                                                  &mad_priv->grh,
2606                                                  sizeof *mad_priv -
2607                                                    sizeof mad_priv->header,
2608                                                  DMA_FROM_DEVICE);
2609                 if (unlikely(ib_dma_mapping_error(qp_info->port_priv->device,
2610                                                   sg_list.addr))) {
2611                         ret = -ENOMEM;
2612                         break;
2613                 }
2614                 mad_priv->header.mapping = sg_list.addr;
2615                 recv_wr.wr_id = (unsigned long)&mad_priv->header.mad_list;
2616                 mad_priv->header.mad_list.mad_queue = recv_queue;
2617
2618                 /* Post receive WR */
2619                 spin_lock_irqsave(&recv_queue->lock, flags);
2620                 post = (++recv_queue->count < recv_queue->max_active);
2621                 list_add_tail(&mad_priv->header.mad_list.list, &recv_queue->list);
2622                 spin_unlock_irqrestore(&recv_queue->lock, flags);
2623                 ret = ib_post_recv(qp_info->qp, &recv_wr, &bad_recv_wr);
2624                 if (ret) {
2625                         spin_lock_irqsave(&recv_queue->lock, flags);
2626                         list_del(&mad_priv->header.mad_list.list);
2627                         recv_queue->count--;
2628                         spin_unlock_irqrestore(&recv_queue->lock, flags);
2629                         ib_dma_unmap_single(qp_info->port_priv->device,
2630                                             mad_priv->header.mapping,
2631                                             sizeof *mad_priv -
2632                                               sizeof mad_priv->header,
2633                                             DMA_FROM_DEVICE);
2634                         kmem_cache_free(ib_mad_cache, mad_priv);
2635                         dev_err(&qp_info->port_priv->device->dev,
2636                                 "ib_post_recv failed: %d\n", ret);
2637                         break;
2638                 }
2639         } while (post);
2640
2641         return ret;
2642 }
2643
2644 /*
2645  * Return all the posted receive MADs
2646  */
2647 static void cleanup_recv_queue(struct ib_mad_qp_info *qp_info)
2648 {
2649         struct ib_mad_private_header *mad_priv_hdr;
2650         struct ib_mad_private *recv;
2651         struct ib_mad_list_head *mad_list;
2652
2653         if (!qp_info->qp)
2654                 return;
2655
2656         while (!list_empty(&qp_info->recv_queue.list)) {
2657
2658                 mad_list = list_entry(qp_info->recv_queue.list.next,
2659                                       struct ib_mad_list_head, list);
2660                 mad_priv_hdr = container_of(mad_list,
2661                                             struct ib_mad_private_header,
2662                                             mad_list);
2663                 recv = container_of(mad_priv_hdr, struct ib_mad_private,
2664                                     header);
2665
2666                 /* Remove from posted receive MAD list */
2667                 list_del(&mad_list->list);
2668
2669                 ib_dma_unmap_single(qp_info->port_priv->device,
2670                                     recv->header.mapping,
2671                                     sizeof(struct ib_mad_private) -
2672                                       sizeof(struct ib_mad_private_header),
2673                                     DMA_FROM_DEVICE);
2674                 kmem_cache_free(ib_mad_cache, recv);
2675         }
2676
2677         qp_info->recv_queue.count = 0;
2678 }
2679
2680 /*
2681  * Start the port
2682  */
2683 static int ib_mad_port_start(struct ib_mad_port_private *port_priv)
2684 {
2685         int ret, i;
2686         struct ib_qp_attr *attr;
2687         struct ib_qp *qp;
2688         u16 pkey_index;
2689
2690         attr = kmalloc(sizeof *attr, GFP_KERNEL);
2691         if (!attr) {
2692                 dev_err(&port_priv->device->dev,
2693                         "Couldn't kmalloc ib_qp_attr\n");
2694                 return -ENOMEM;
2695         }
2696
2697         ret = ib_find_pkey(port_priv->device, port_priv->port_num,
2698                            IB_DEFAULT_PKEY_FULL, &pkey_index);
2699         if (ret)
2700                 pkey_index = 0;
2701
2702         for (i = 0; i < IB_MAD_QPS_CORE; i++) {
2703                 qp = port_priv->qp_info[i].qp;
2704                 if (!qp)
2705                         continue;
2706
2707                 /*
2708                  * PKey index for QP1 is irrelevant but
2709                  * one is needed for the Reset to Init transition
2710                  */
2711                 attr->qp_state = IB_QPS_INIT;
2712                 attr->pkey_index = pkey_index;
2713                 attr->qkey = (qp->qp_num == 0) ? 0 : IB_QP1_QKEY;
2714                 ret = ib_modify_qp(qp, attr, IB_QP_STATE |
2715                                              IB_QP_PKEY_INDEX | IB_QP_QKEY);
2716                 if (ret) {
2717                         dev_err(&port_priv->device->dev,
2718                                 "Couldn't change QP%d state to INIT: %d\n",
2719                                 i, ret);
2720                         goto out;
2721                 }
2722
2723                 attr->qp_state = IB_QPS_RTR;
2724                 ret = ib_modify_qp(qp, attr, IB_QP_STATE);
2725                 if (ret) {
2726                         dev_err(&port_priv->device->dev,
2727                                 "Couldn't change QP%d state to RTR: %d\n",
2728                                 i, ret);
2729                         goto out;
2730                 }
2731
2732                 attr->qp_state = IB_QPS_RTS;
2733                 attr->sq_psn = IB_MAD_SEND_Q_PSN;
2734                 ret = ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_SQ_PSN);
2735                 if (ret) {
2736                         dev_err(&port_priv->device->dev,
2737                                 "Couldn't change QP%d state to RTS: %d\n",
2738                                 i, ret);
2739                         goto out;
2740                 }
2741         }
2742
2743         ret = ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP);
2744         if (ret) {
2745                 dev_err(&port_priv->device->dev,
2746                         "Failed to request completion notification: %d\n",
2747                         ret);
2748                 goto out;
2749         }
2750
2751         for (i = 0; i < IB_MAD_QPS_CORE; i++) {
2752                 if (!port_priv->qp_info[i].qp)
2753                         continue;
2754
2755                 ret = ib_mad_post_receive_mads(&port_priv->qp_info[i], NULL);
2756                 if (ret) {
2757                         dev_err(&port_priv->device->dev,
2758                                 "Couldn't post receive WRs\n");
2759                         goto out;
2760                 }
2761         }
2762 out:
2763         kfree(attr);
2764         return ret;
2765 }
2766
2767 static void qp_event_handler(struct ib_event *event, void *qp_context)
2768 {
2769         struct ib_mad_qp_info   *qp_info = qp_context;
2770
2771         /* It's worse than that! He's dead, Jim! */
2772         dev_err(&qp_info->port_priv->device->dev,
2773                 "Fatal error (%d) on MAD QP (%d)\n",
2774                 event->event, qp_info->qp->qp_num);
2775 }
2776
2777 static void init_mad_queue(struct ib_mad_qp_info *qp_info,
2778                            struct ib_mad_queue *mad_queue)
2779 {
2780         mad_queue->qp_info = qp_info;
2781         mad_queue->count = 0;
2782         spin_lock_init(&mad_queue->lock);
2783         INIT_LIST_HEAD(&mad_queue->list);
2784 }
2785
2786 static void init_mad_qp(struct ib_mad_port_private *port_priv,
2787                         struct ib_mad_qp_info *qp_info)
2788 {
2789         qp_info->port_priv = port_priv;
2790         init_mad_queue(qp_info, &qp_info->send_queue);
2791         init_mad_queue(qp_info, &qp_info->recv_queue);
2792         INIT_LIST_HEAD(&qp_info->overflow_list);
2793         spin_lock_init(&qp_info->snoop_lock);
2794         qp_info->snoop_table = NULL;
2795         qp_info->snoop_table_size = 0;
2796         atomic_set(&qp_info->snoop_count, 0);
2797 }
2798
2799 static int create_mad_qp(struct ib_mad_qp_info *qp_info,
2800                          enum ib_qp_type qp_type)
2801 {
2802         struct ib_qp_init_attr  qp_init_attr;
2803         int ret;
2804
2805         memset(&qp_init_attr, 0, sizeof qp_init_attr);
2806         qp_init_attr.send_cq = qp_info->port_priv->cq;
2807         qp_init_attr.recv_cq = qp_info->port_priv->cq;
2808         qp_init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
2809         qp_init_attr.cap.max_send_wr = mad_sendq_size;
2810         qp_init_attr.cap.max_recv_wr = mad_recvq_size;
2811         qp_init_attr.cap.max_send_sge = IB_MAD_SEND_REQ_MAX_SG;
2812         qp_init_attr.cap.max_recv_sge = IB_MAD_RECV_REQ_MAX_SG;
2813         qp_init_attr.qp_type = qp_type;
2814         qp_init_attr.port_num = qp_info->port_priv->port_num;
2815         qp_init_attr.qp_context = qp_info;
2816         qp_init_attr.event_handler = qp_event_handler;
2817         qp_info->qp = ib_create_qp(qp_info->port_priv->pd, &qp_init_attr);
2818         if (IS_ERR(qp_info->qp)) {
2819                 dev_err(&qp_info->port_priv->device->dev,
2820                         "Couldn't create ib_mad QP%d\n",
2821                         get_spl_qp_index(qp_type));
2822                 ret = PTR_ERR(qp_info->qp);
2823                 goto error;
2824         }
2825         /* Use minimum queue sizes unless the CQ is resized */
2826         qp_info->send_queue.max_active = mad_sendq_size;
2827         qp_info->recv_queue.max_active = mad_recvq_size;
2828         return 0;
2829
2830 error:
2831         return ret;
2832 }
2833
2834 static void destroy_mad_qp(struct ib_mad_qp_info *qp_info)
2835 {
2836         if (!qp_info->qp)
2837                 return;
2838
2839         ib_destroy_qp(qp_info->qp);
2840         kfree(qp_info->snoop_table);
2841 }
2842
2843 /*
2844  * Open the port
2845  * Create the QP, PD, MR, and CQ if needed
2846  */
2847 static int ib_mad_port_open(struct ib_device *device,
2848                             int port_num)
2849 {
2850         int ret, cq_size;
2851         struct ib_mad_port_private *port_priv;
2852         unsigned long flags;
2853         char name[sizeof "ib_mad123"];
2854         int has_smi;
2855
2856         /* Create new device info */
2857         port_priv = kzalloc(sizeof *port_priv, GFP_KERNEL);
2858         if (!port_priv) {
2859                 dev_err(&device->dev, "No memory for ib_mad_port_private\n");
2860                 return -ENOMEM;
2861         }
2862
2863         port_priv->device = device;
2864         port_priv->port_num = port_num;
2865         spin_lock_init(&port_priv->reg_lock);
2866         INIT_LIST_HEAD(&port_priv->agent_list);
2867         init_mad_qp(port_priv, &port_priv->qp_info[0]);
2868         init_mad_qp(port_priv, &port_priv->qp_info[1]);
2869
2870         cq_size = mad_sendq_size + mad_recvq_size;
2871         has_smi = rdma_port_get_link_layer(device, port_num) == IB_LINK_LAYER_INFINIBAND;
2872         if (has_smi)
2873                 cq_size *= 2;
2874
2875         port_priv->cq = ib_create_cq(port_priv->device,
2876                                      ib_mad_thread_completion_handler,
2877                                      NULL, port_priv, cq_size, 0);
2878         if (IS_ERR(port_priv->cq)) {
2879                 dev_err(&device->dev, "Couldn't create ib_mad CQ\n");
2880                 ret = PTR_ERR(port_priv->cq);
2881                 goto error3;
2882         }
2883
2884         port_priv->pd = ib_alloc_pd(device);
2885         if (IS_ERR(port_priv->pd)) {
2886                 dev_err(&device->dev, "Couldn't create ib_mad PD\n");
2887                 ret = PTR_ERR(port_priv->pd);
2888                 goto error4;
2889         }
2890
2891         port_priv->mr = ib_get_dma_mr(port_priv->pd, IB_ACCESS_LOCAL_WRITE);
2892         if (IS_ERR(port_priv->mr)) {
2893                 dev_err(&device->dev, "Couldn't get ib_mad DMA MR\n");
2894                 ret = PTR_ERR(port_priv->mr);
2895                 goto error5;
2896         }
2897
2898         if (has_smi) {
2899                 ret = create_mad_qp(&port_priv->qp_info[0], IB_QPT_SMI);
2900                 if (ret)
2901                         goto error6;
2902         }
2903         ret = create_mad_qp(&port_priv->qp_info[1], IB_QPT_GSI);
2904         if (ret)
2905                 goto error7;
2906
2907         snprintf(name, sizeof name, "ib_mad%d", port_num);
2908         port_priv->wq = create_singlethread_workqueue(name);
2909         if (!port_priv->wq) {
2910                 ret = -ENOMEM;
2911                 goto error8;
2912         }
2913         INIT_WORK(&port_priv->work, ib_mad_completion_handler);
2914
2915         spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2916         list_add_tail(&port_priv->port_list, &ib_mad_port_list);
2917         spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2918
2919         ret = ib_mad_port_start(port_priv);
2920         if (ret) {
2921                 dev_err(&device->dev, "Couldn't start port\n");
2922                 goto error9;
2923         }
2924
2925         return 0;
2926
2927 error9:
2928         spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2929         list_del_init(&port_priv->port_list);
2930         spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2931
2932         destroy_workqueue(port_priv->wq);
2933 error8:
2934         destroy_mad_qp(&port_priv->qp_info[1]);
2935 error7:
2936         destroy_mad_qp(&port_priv->qp_info[0]);
2937 error6:
2938         ib_dereg_mr(port_priv->mr);
2939 error5:
2940         ib_dealloc_pd(port_priv->pd);
2941 error4:
2942         ib_destroy_cq(port_priv->cq);
2943         cleanup_recv_queue(&port_priv->qp_info[1]);
2944         cleanup_recv_queue(&port_priv->qp_info[0]);
2945 error3:
2946         kfree(port_priv);
2947
2948         return ret;
2949 }
2950
2951 /*
2952  * Close the port
2953  * If there are no classes using the port, free the port
2954  * resources (CQ, MR, PD, QP) and remove the port's info structure
2955  */
2956 static int ib_mad_port_close(struct ib_device *device, int port_num)
2957 {
2958         struct ib_mad_port_private *port_priv;
2959         unsigned long flags;
2960
2961         spin_lock_irqsave(&ib_mad_port_list_lock, flags);
2962         port_priv = __ib_get_mad_port(device, port_num);
2963         if (port_priv == NULL) {
2964                 spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2965                 dev_err(&device->dev, "Port %d not found\n", port_num);
2966                 return -ENODEV;
2967         }
2968         list_del_init(&port_priv->port_list);
2969         spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
2970
2971         destroy_workqueue(port_priv->wq);
2972         destroy_mad_qp(&port_priv->qp_info[1]);
2973         destroy_mad_qp(&port_priv->qp_info[0]);
2974         ib_dereg_mr(port_priv->mr);
2975         ib_dealloc_pd(port_priv->pd);
2976         ib_destroy_cq(port_priv->cq);
2977         cleanup_recv_queue(&port_priv->qp_info[1]);
2978         cleanup_recv_queue(&port_priv->qp_info[0]);
2979         /* XXX: Handle deallocation of MAD registration tables */
2980
2981         kfree(port_priv);
2982
2983         return 0;
2984 }
2985
2986 static void ib_mad_init_device(struct ib_device *device)
2987 {
2988         int start, end, i;
2989
2990         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
2991                 return;
2992
2993         if (device->node_type == RDMA_NODE_IB_SWITCH) {
2994                 start = 0;
2995                 end   = 0;
2996         } else {
2997                 start = 1;
2998                 end   = device->phys_port_cnt;
2999         }
3000
3001         for (i = start; i <= end; i++) {
3002                 if (ib_mad_port_open(device, i)) {
3003                         dev_err(&device->dev, "Couldn't open port %d\n", i);
3004                         goto error;
3005                 }
3006                 if (ib_agent_port_open(device, i)) {
3007                         dev_err(&device->dev,
3008                                 "Couldn't open port %d for agents\n", i);
3009                         goto error_agent;
3010                 }
3011         }
3012         return;
3013
3014 error_agent:
3015         if (ib_mad_port_close(device, i))
3016                 dev_err(&device->dev, "Couldn't close port %d\n", i);
3017
3018 error:
3019         i--;
3020
3021         while (i >= start) {
3022                 if (ib_agent_port_close(device, i))
3023                         dev_err(&device->dev,
3024                                 "Couldn't close port %d for agents\n", i);
3025                 if (ib_mad_port_close(device, i))
3026                         dev_err(&device->dev, "Couldn't close port %d\n", i);
3027                 i--;
3028         }
3029 }
3030
3031 static void ib_mad_remove_device(struct ib_device *device)
3032 {
3033         int i, num_ports, cur_port;
3034
3035         if (rdma_node_get_transport(device->node_type) != RDMA_TRANSPORT_IB)
3036                 return;
3037
3038         if (device->node_type == RDMA_NODE_IB_SWITCH) {
3039                 num_ports = 1;
3040                 cur_port = 0;
3041         } else {
3042                 num_ports = device->phys_port_cnt;
3043                 cur_port = 1;
3044         }
3045         for (i = 0; i < num_ports; i++, cur_port++) {
3046                 if (ib_agent_port_close(device, cur_port))
3047                         dev_err(&device->dev,
3048                                 "Couldn't close port %d for agents\n",
3049                                 cur_port);
3050                 if (ib_mad_port_close(device, cur_port))
3051                         dev_err(&device->dev, "Couldn't close port %d\n",
3052                                 cur_port);
3053         }
3054 }
3055
3056 static struct ib_client mad_client = {
3057         .name   = "mad",
3058         .add = ib_mad_init_device,
3059         .remove = ib_mad_remove_device
3060 };
3061
3062 static int __init ib_mad_init_module(void)
3063 {
3064         int ret;
3065
3066         mad_recvq_size = min(mad_recvq_size, IB_MAD_QP_MAX_SIZE);
3067         mad_recvq_size = max(mad_recvq_size, IB_MAD_QP_MIN_SIZE);
3068
3069         mad_sendq_size = min(mad_sendq_size, IB_MAD_QP_MAX_SIZE);
3070         mad_sendq_size = max(mad_sendq_size, IB_MAD_QP_MIN_SIZE);
3071
3072         ib_mad_cache = kmem_cache_create("ib_mad",
3073                                          sizeof(struct ib_mad_private),
3074                                          0,
3075                                          SLAB_HWCACHE_ALIGN,
3076                                          NULL);
3077         if (!ib_mad_cache) {
3078                 pr_err("Couldn't create ib_mad cache\n");
3079                 ret = -ENOMEM;
3080                 goto error1;
3081         }
3082
3083         INIT_LIST_HEAD(&ib_mad_port_list);
3084
3085         if (ib_register_client(&mad_client)) {
3086                 pr_err("Couldn't register ib_mad client\n");
3087                 ret = -EINVAL;
3088                 goto error2;
3089         }
3090
3091         return 0;
3092
3093 error2:
3094         kmem_cache_destroy(ib_mad_cache);
3095 error1:
3096         return ret;
3097 }
3098
3099 static void __exit ib_mad_cleanup_module(void)
3100 {
3101         ib_unregister_client(&mad_client);
3102         kmem_cache_destroy(ib_mad_cache);
3103 }
3104
3105 module_init(ib_mad_init_module);
3106 module_exit(ib_mad_cleanup_module);