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  * Copyright (c) 2014 Intel Corporation.  All rights reserved.
   7  *
   8  * This software is available to you under a choice of one of two
   9  * licenses.  You may choose to be licensed under the terms of the GNU
  10  * General Public License (GPL) Version 2, available from the file
  11  * COPYING in the main directory of this source tree, or the
  12  * OpenIB.org BSD license below:
  13  *
  14  *     Redistribution and use in source and binary forms, with or
  15  *     without modification, are permitted provided that the following
  16  *     conditions are met:
  17  *
  18  *      - Redistributions of source code must retain the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer.
  21  *
  22  *      - Redistributions in binary form must reproduce the above
  23  *        copyright notice, this list of conditions and the following
  24  *        disclaimer in the documentation and/or other materials
  25  *        provided with the distribution.
  26  *
  27  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  28  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  29  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  30  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  31  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  32  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  33  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  34  * SOFTWARE.
  35  *
  36  */
  37
  38 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  39
  40 #include <linux/dma-mapping.h>
  41 #include <linux/idr.h>
  42 #include <linux/slab.h>
  43 #include <linux/module.h>
  44 #include <linux/security.h>
  45 #include <rdma/ib_cache.h>
  46
  47 #include "mad_priv.h"
  48 #include "core_priv.h"
  49 #include "mad_rmpp.h"
  50 #include "smi.h"
  51 #include "opa_smi.h"
  52 #include "agent.h"
  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 /*
  63  * The mlx4 driver uses the top byte to distinguish which virtual function
  64  * generated the MAD, so we must avoid using it.
  65  */
  66 #define AGENT_ID_LIMIT          (1 << 24)
  67 static DEFINE_IDR(ib_mad_clients);
  68 static struct list_head ib_mad_port_list;
  69
  70 /* Port list lock */
  71 static DEFINE_SPINLOCK(ib_mad_port_list_lock);
  72
  73 /* Forward declarations */
  74 static int method_in_use(struct ib_mad_mgmt_method_table **method,
  75                          struct ib_mad_reg_req *mad_reg_req);
  76 static void remove_mad_reg_req(struct ib_mad_agent_private *priv);
  77 static struct ib_mad_agent_private *find_mad_agent(
  78                                         struct ib_mad_port_private *port_priv,
  79                                         const struct ib_mad_hdr *mad);
  80 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
  81                                     struct ib_mad_private *mad);
  82 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv);
  83 static void timeout_sends(struct work_struct *work);
  84 static void local_completions(struct work_struct *work);
  85 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
  86                               struct ib_mad_agent_private *agent_priv,
  87                               u8 mgmt_class);
  88 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
  89                            struct ib_mad_agent_private *agent_priv);
  90 static bool ib_mad_send_error(struct ib_mad_port_private *port_priv,
  91                               struct ib_wc *wc);
  92 static void ib_mad_send_done(struct ib_cq *cq, struct ib_wc *wc);
  93
  94 /*
  95  * Returns a ib_mad_port_private structure or NULL for a device/port
  96  * Assumes ib_mad_port_list_lock is being held
  97  */
  98 static inline struct ib_mad_port_private *
  99 __ib_get_mad_port(struct ib_device *device, int port_num)
 100 {
 101         struct ib_mad_port_private *entry;
 102
 103         list_for_each_entry(entry, &ib_mad_port_list, port_list) {
 104                 if (entry->device == device && entry->port_num == port_num)
 105                         return entry;
 106         }
 107         return NULL;
 108 }
 109
 110 /*
 111  * Wrapper function to return a ib_mad_port_private structure or NULL
 112  * for a device/port
 113  */
 114 static inline struct ib_mad_port_private *
 115 ib_get_mad_port(struct ib_device *device, int port_num)
 116 {
 117         struct ib_mad_port_private *entry;
 118         unsigned long flags;
 119
 120         spin_lock_irqsave(&ib_mad_port_list_lock, flags);
 121         entry = __ib_get_mad_port(device, port_num);
 122         spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
 123
 124         return entry;
 125 }
 126
 127 static inline u8 convert_mgmt_class(u8 mgmt_class)
 128 {
 129         /* Alias IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE to 0 */
 130         return mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE ?
 131                 0 : mgmt_class;
 132 }
 133
 134 static int get_spl_qp_index(enum ib_qp_type qp_type)
 135 {
 136         switch (qp_type)
 137         {
 138         case IB_QPT_SMI:
 139                 return 0;
 140         case IB_QPT_GSI:
 141                 return 1;
 142         default:
 143                 return -1;
 144         }
 145 }
 146
 147 static int vendor_class_index(u8 mgmt_class)
 148 {
 149         return mgmt_class - IB_MGMT_CLASS_VENDOR_RANGE2_START;
 150 }
 151
 152 static int is_vendor_class(u8 mgmt_class)
 153 {
 154         if ((mgmt_class < IB_MGMT_CLASS_VENDOR_RANGE2_START) ||
 155             (mgmt_class > IB_MGMT_CLASS_VENDOR_RANGE2_END))
 156                 return 0;
 157         return 1;
 158 }
 159
 160 static int is_vendor_oui(char *oui)
 161 {
 162         if (oui[0] || oui[1] || oui[2])
 163                 return 1;
 164         return 0;
 165 }
 166
 167 static int is_vendor_method_in_use(
 168                 struct ib_mad_mgmt_vendor_class *vendor_class,
 169                 struct ib_mad_reg_req *mad_reg_req)
 170 {
 171         struct ib_mad_mgmt_method_table *method;
 172         int i;
 173
 174         for (i = 0; i < MAX_MGMT_OUI; i++) {
 175                 if (!memcmp(vendor_class->oui[i], mad_reg_req->oui, 3)) {
 176                         method = vendor_class->method_table[i];
 177                         if (method) {
 178                                 if (method_in_use(&method, mad_reg_req))
 179                                         return 1;
 180                                 else
 181                                         break;
 182                         }
 183                 }
 184         }
 185         return 0;
 186 }
 187
 188 int ib_response_mad(const struct ib_mad_hdr *hdr)
 189 {
 190         return ((hdr->method & IB_MGMT_METHOD_RESP) ||
 191                 (hdr->method == IB_MGMT_METHOD_TRAP_REPRESS) ||
 192                 ((hdr->mgmt_class == IB_MGMT_CLASS_BM) &&
 193                  (hdr->attr_mod & IB_BM_ATTR_MOD_RESP)));
 194 }
 195 EXPORT_SYMBOL(ib_response_mad);
 196
 197 /*
 198  * ib_register_mad_agent - Register to send/receive MADs
 199  *
 200  * Context: Process context.
 201  */
 202 struct ib_mad_agent *ib_register_mad_agent(struct ib_device *device,
 203                                            u8 port_num,
 204                                            enum ib_qp_type qp_type,
 205                                            struct ib_mad_reg_req *mad_reg_req,
 206                                            u8 rmpp_version,
 207                                            ib_mad_send_handler send_handler,
 208                                            ib_mad_recv_handler recv_handler,
 209                                            void *context,
 210                                            u32 registration_flags)
 211 {
 212         struct ib_mad_port_private *port_priv;
 213         struct ib_mad_agent *ret = ERR_PTR(-EINVAL);
 214         struct ib_mad_agent_private *mad_agent_priv;
 215         struct ib_mad_reg_req *reg_req = NULL;
 216         struct ib_mad_mgmt_class_table *class;
 217         struct ib_mad_mgmt_vendor_class_table *vendor;
 218         struct ib_mad_mgmt_vendor_class *vendor_class;
 219         struct ib_mad_mgmt_method_table *method;
 220         int ret2, qpn;
 221         u8 mgmt_class, vclass;
 222
 223         /* Validate parameters */
 224         qpn = get_spl_qp_index(qp_type);
 225         if (qpn == -1) {
 226                 dev_notice(&device->dev,
 227                            "ib_register_mad_agent: invalid QP Type %d\n",
 228                            qp_type);
 229                 goto error1;
 230         }
 231
 232         if (rmpp_version && rmpp_version != IB_MGMT_RMPP_VERSION) {
 233                 dev_notice(&device->dev,
 234                            "ib_register_mad_agent: invalid RMPP Version %u\n",
 235                            rmpp_version);
 236                 goto error1;
 237         }
 238
 239         /* Validate MAD registration request if supplied */
 240         if (mad_reg_req) {
 241                 if (mad_reg_req->mgmt_class_version >= MAX_MGMT_VERSION) {
 242                         dev_notice(&device->dev,
 243                                    "ib_register_mad_agent: invalid Class Version %u\n",
 244                                    mad_reg_req->mgmt_class_version);
 245                         goto error1;
 246                 }
 247                 if (!recv_handler) {
 248                         dev_notice(&device->dev,
 249                                    "ib_register_mad_agent: no recv_handler\n");
 250                         goto error1;
 251                 }
 252                 if (mad_reg_req->mgmt_class >= MAX_MGMT_CLASS) {
 253                         /*
 254                          * IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE is the only
 255                          * one in this range currently allowed
 256                          */
 257                         if (mad_reg_req->mgmt_class !=
 258                             IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
 259                                 dev_notice(&device->dev,
 260                                            "ib_register_mad_agent: Invalid Mgmt Class 0x%x\n",
 261                                            mad_reg_req->mgmt_class);
 262                                 goto error1;
 263                         }
 264                 } else if (mad_reg_req->mgmt_class == 0) {
 265                         /*
 266                          * Class 0 is reserved in IBA and is used for
 267                          * aliasing of IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE
 268                          */
 269                         dev_notice(&device->dev,
 270                                    "ib_register_mad_agent: Invalid Mgmt Class 0\n");
 271                         goto error1;
 272                 } else if (is_vendor_class(mad_reg_req->mgmt_class)) {
 273                         /*
 274                          * If class is in "new" vendor range,
 275                          * ensure supplied OUI is not zero
 276                          */
 277                         if (!is_vendor_oui(mad_reg_req->oui)) {
 278                                 dev_notice(&device->dev,
 279                                            "ib_register_mad_agent: No OUI specified for class 0x%x\n",
 280                                            mad_reg_req->mgmt_class);
 281                                 goto error1;
 282                         }
 283                 }
 284                 /* Make sure class supplied is consistent with RMPP */
 285                 if (!ib_is_mad_class_rmpp(mad_reg_req->mgmt_class)) {
 286                         if (rmpp_version) {
 287                                 dev_notice(&device->dev,
 288                                            "ib_register_mad_agent: RMPP version for non-RMPP class 0x%x\n",
 289                                            mad_reg_req->mgmt_class);
 290                                 goto error1;
 291                         }
 292                 }
 293
 294                 /* Make sure class supplied is consistent with QP type */
 295                 if (qp_type == IB_QPT_SMI) {
 296                         if ((mad_reg_req->mgmt_class !=
 297                                         IB_MGMT_CLASS_SUBN_LID_ROUTED) &&
 298                             (mad_reg_req->mgmt_class !=
 299                                         IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
 300                                 dev_notice(&device->dev,
 301                                            "ib_register_mad_agent: Invalid SM QP type: class 0x%x\n",
 302                                            mad_reg_req->mgmt_class);
 303                                 goto error1;
 304                         }
 305                 } else {
 306                         if ((mad_reg_req->mgmt_class ==
 307                                         IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
 308                             (mad_reg_req->mgmt_class ==
 309                                         IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
 310                                 dev_notice(&device->dev,
 311                                            "ib_register_mad_agent: Invalid GS QP type: class 0x%x\n",
 312                                            mad_reg_req->mgmt_class);
 313                                 goto error1;
 314                         }
 315                 }
 316         } else {
 317                 /* No registration request supplied */
 318                 if (!send_handler)
 319                         goto error1;
 320                 if (registration_flags & IB_MAD_USER_RMPP)
 321                         goto error1;
 322         }
 323
 324         /* Validate device and port */
 325         port_priv = ib_get_mad_port(device, port_num);
 326         if (!port_priv) {
 327                 dev_notice(&device->dev,
 328                            "ib_register_mad_agent: Invalid port %d\n",
 329                            port_num);
 330                 ret = ERR_PTR(-ENODEV);
 331                 goto error1;
 332         }
 333
 334         /* Verify the QP requested is supported.  For example, Ethernet devices
 335          * will not have QP0 */
 336         if (!port_priv->qp_info[qpn].qp) {
 337                 dev_notice(&device->dev,
 338                            "ib_register_mad_agent: QP %d not supported\n", qpn);
 339                 ret = ERR_PTR(-EPROTONOSUPPORT);
 340                 goto error1;
 341         }
 342
 343         /* Allocate structures */
 344         mad_agent_priv = kzalloc(sizeof *mad_agent_priv, GFP_KERNEL);
 345         if (!mad_agent_priv) {
 346                 ret = ERR_PTR(-ENOMEM);
 347                 goto error1;
 348         }
 349
 350         if (mad_reg_req) {
 351                 reg_req = kmemdup(mad_reg_req, sizeof *reg_req, GFP_KERNEL);
 352                 if (!reg_req) {
 353                         ret = ERR_PTR(-ENOMEM);
 354                         goto error3;
 355                 }
 356         }
 357
 358         /* Now, fill in the various structures */
 359         mad_agent_priv->qp_info = &port_priv->qp_info[qpn];
 360         mad_agent_priv->reg_req = reg_req;
 361         mad_agent_priv->agent.rmpp_version = rmpp_version;
 362         mad_agent_priv->agent.device = device;
 363         mad_agent_priv->agent.recv_handler = recv_handler;
 364         mad_agent_priv->agent.send_handler = send_handler;
 365         mad_agent_priv->agent.context = context;
 366         mad_agent_priv->agent.qp = port_priv->qp_info[qpn].qp;
 367         mad_agent_priv->agent.port_num = port_num;
 368         mad_agent_priv->agent.flags = registration_flags;
 369         spin_lock_init(&mad_agent_priv->lock);
 370         INIT_LIST_HEAD(&mad_agent_priv->send_list);
 371         INIT_LIST_HEAD(&mad_agent_priv->wait_list);
 372         INIT_LIST_HEAD(&mad_agent_priv->done_list);
 373         INIT_LIST_HEAD(&mad_agent_priv->rmpp_list);
 374         INIT_DELAYED_WORK(&mad_agent_priv->timed_work, timeout_sends);
 375         INIT_LIST_HEAD(&mad_agent_priv->local_list);
 376         INIT_WORK(&mad_agent_priv->local_work, local_completions);
 377         atomic_set(&mad_agent_priv->refcount, 1);
 378         init_completion(&mad_agent_priv->comp);
 379
 380         ret2 = ib_mad_agent_security_setup(&mad_agent_priv->agent, qp_type);
 381         if (ret2) {
 382                 ret = ERR_PTR(ret2);
 383                 goto error4;
 384         }
 385
 386         idr_preload(GFP_KERNEL);
 387         idr_lock(&ib_mad_clients);
 388         ret2 = idr_alloc_cyclic(&ib_mad_clients, mad_agent_priv, 0,
 389                         AGENT_ID_LIMIT, GFP_ATOMIC);
 390         idr_unlock(&ib_mad_clients);
 391         idr_preload_end();
 392
 393         if (ret2 < 0) {
 394                 ret = ERR_PTR(ret2);
 395                 goto error5;
 396         }
 397         mad_agent_priv->agent.hi_tid = ret2;
 398
 399         /*
 400          * Make sure MAD registration (if supplied)
 401          * is non overlapping with any existing ones
 402          */
 403         spin_lock_irq(&port_priv->reg_lock);
 404         if (mad_reg_req) {
 405                 mgmt_class = convert_mgmt_class(mad_reg_req->mgmt_class);
 406                 if (!is_vendor_class(mgmt_class)) {
 407                         class = port_priv->version[mad_reg_req->
 408                                                    mgmt_class_version].class;
 409                         if (class) {
 410                                 method = class->method_table[mgmt_class];
 411                                 if (method) {
 412                                         if (method_in_use(&method,
 413                                                            mad_reg_req))
 414                                                 goto error6;
 415                                 }
 416                         }
 417                         ret2 = add_nonoui_reg_req(mad_reg_req, mad_agent_priv,
 418                                                   mgmt_class);
 419                 } else {
 420                         /* "New" vendor class range */
 421                         vendor = port_priv->version[mad_reg_req->
 422                                                     mgmt_class_version].vendor;
 423                         if (vendor) {
 424                                 vclass = vendor_class_index(mgmt_class);
 425                                 vendor_class = vendor->vendor_class[vclass];
 426                                 if (vendor_class) {
 427                                         if (is_vendor_method_in_use(
 428                                                         vendor_class,
 429                                                         mad_reg_req))
 430                                                 goto error6;
 431                                 }
 432                         }
 433                         ret2 = add_oui_reg_req(mad_reg_req, mad_agent_priv);
 434                 }
 435                 if (ret2) {
 436                         ret = ERR_PTR(ret2);
 437                         goto error6;
 438                 }
 439         }
 440         spin_unlock_irq(&port_priv->reg_lock);
 441
 442         return &mad_agent_priv->agent;
 443 error6:
 444         spin_unlock_irq(&port_priv->reg_lock);
 445         idr_lock(&ib_mad_clients);
 446         idr_remove(&ib_mad_clients, mad_agent_priv->agent.hi_tid);
 447         idr_unlock(&ib_mad_clients);
 448 error5:
 449         ib_mad_agent_security_cleanup(&mad_agent_priv->agent);
 450 error4:
 451         kfree(reg_req);
 452 error3:
 453         kfree(mad_agent_priv);
 454 error1:
 455         return ret;
 456 }
 457 EXPORT_SYMBOL(ib_register_mad_agent);
 458
 459 static inline int is_snooping_sends(int mad_snoop_flags)
 460 {
 461         return (mad_snoop_flags &
 462                 (/*IB_MAD_SNOOP_POSTED_SENDS |
 463                  IB_MAD_SNOOP_RMPP_SENDS |*/
 464                  IB_MAD_SNOOP_SEND_COMPLETIONS /*|
 465                  IB_MAD_SNOOP_RMPP_SEND_COMPLETIONS*/));
 466 }
 467
 468 static inline int is_snooping_recvs(int mad_snoop_flags)
 469 {
 470         return (mad_snoop_flags &
 471                 (IB_MAD_SNOOP_RECVS /*|
 472                  IB_MAD_SNOOP_RMPP_RECVS*/));
 473 }
 474
 475 static int register_snoop_agent(struct ib_mad_qp_info *qp_info,
 476                                 struct ib_mad_snoop_private *mad_snoop_priv)
 477 {
 478         struct ib_mad_snoop_private **new_snoop_table;
 479         unsigned long flags;
 480         int i;
 481
 482         spin_lock_irqsave(&qp_info->snoop_lock, flags);
 483         /* Check for empty slot in array. */
 484         for (i = 0; i < qp_info->snoop_table_size; i++)
 485                 if (!qp_info->snoop_table[i])
 486                         break;
 487
 488         if (i == qp_info->snoop_table_size) {
 489                 /* Grow table. */
 490                 new_snoop_table = krealloc(qp_info->snoop_table,
 491                                            sizeof mad_snoop_priv *
 492                                            (qp_info->snoop_table_size + 1),
 493                                            GFP_ATOMIC);
 494                 if (!new_snoop_table) {
 495                         i = -ENOMEM;
 496                         goto out;
 497                 }
 498
 499                 qp_info->snoop_table = new_snoop_table;
 500                 qp_info->snoop_table_size++;
 501         }
 502         qp_info->snoop_table[i] = mad_snoop_priv;
 503         atomic_inc(&qp_info->snoop_count);
 504 out:
 505         spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
 506         return i;
 507 }
 508
 509 struct ib_mad_agent *ib_register_mad_snoop(struct ib_device *device,
 510                                            u8 port_num,
 511                                            enum ib_qp_type qp_type,
 512                                            int mad_snoop_flags,
 513                                            ib_mad_snoop_handler snoop_handler,
 514                                            ib_mad_recv_handler recv_handler,
 515                                            void *context)
 516 {
 517         struct ib_mad_port_private *port_priv;
 518         struct ib_mad_agent *ret;
 519         struct ib_mad_snoop_private *mad_snoop_priv;
 520         int qpn;
 521         int err;
 522
 523         /* Validate parameters */
 524         if ((is_snooping_sends(mad_snoop_flags) && !snoop_handler) ||
 525             (is_snooping_recvs(mad_snoop_flags) && !recv_handler)) {
 526                 ret = ERR_PTR(-EINVAL);
 527                 goto error1;
 528         }
 529         qpn = get_spl_qp_index(qp_type);
 530         if (qpn == -1) {
 531                 ret = ERR_PTR(-EINVAL);
 532                 goto error1;
 533         }
 534         port_priv = ib_get_mad_port(device, port_num);
 535         if (!port_priv) {
 536                 ret = ERR_PTR(-ENODEV);
 537                 goto error1;
 538         }
 539         /* Allocate structures */
 540         mad_snoop_priv = kzalloc(sizeof *mad_snoop_priv, GFP_KERNEL);
 541         if (!mad_snoop_priv) {
 542                 ret = ERR_PTR(-ENOMEM);
 543                 goto error1;
 544         }
 545
 546         /* Now, fill in the various structures */
 547         mad_snoop_priv->qp_info = &port_priv->qp_info[qpn];
 548         mad_snoop_priv->agent.device = device;
 549         mad_snoop_priv->agent.recv_handler = recv_handler;
 550         mad_snoop_priv->agent.snoop_handler = snoop_handler;
 551         mad_snoop_priv->agent.context = context;
 552         mad_snoop_priv->agent.qp = port_priv->qp_info[qpn].qp;
 553         mad_snoop_priv->agent.port_num = port_num;
 554         mad_snoop_priv->mad_snoop_flags = mad_snoop_flags;
 555         init_completion(&mad_snoop_priv->comp);
 556
 557         err = ib_mad_agent_security_setup(&mad_snoop_priv->agent, qp_type);
 558         if (err) {
 559                 ret = ERR_PTR(err);
 560                 goto error2;
 561         }
 562
 563         mad_snoop_priv->snoop_index = register_snoop_agent(
 564                                                 &port_priv->qp_info[qpn],
 565                                                 mad_snoop_priv);
 566         if (mad_snoop_priv->snoop_index < 0) {
 567                 ret = ERR_PTR(mad_snoop_priv->snoop_index);
 568                 goto error3;
 569         }
 570
 571         atomic_set(&mad_snoop_priv->refcount, 1);
 572         return &mad_snoop_priv->agent;
 573 error3:
 574         ib_mad_agent_security_cleanup(&mad_snoop_priv->agent);
 575 error2:
 576         kfree(mad_snoop_priv);
 577 error1:
 578         return ret;
 579 }
 580 EXPORT_SYMBOL(ib_register_mad_snoop);
 581
 582 static inline void deref_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
 583 {
 584         if (atomic_dec_and_test(&mad_agent_priv->refcount))
 585                 complete(&mad_agent_priv->comp);
 586 }
 587
 588 static inline void deref_snoop_agent(struct ib_mad_snoop_private *mad_snoop_priv)
 589 {
 590         if (atomic_dec_and_test(&mad_snoop_priv->refcount))
 591                 complete(&mad_snoop_priv->comp);
 592 }
 593
 594 static void unregister_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
 595 {
 596         struct ib_mad_port_private *port_priv;
 597
 598         /* Note that we could still be handling received MADs */
 599
 600         /*
 601          * Canceling all sends results in dropping received response
 602          * MADs, preventing us from queuing additional work
 603          */
 604         cancel_mads(mad_agent_priv);
 605         port_priv = mad_agent_priv->qp_info->port_priv;
 606         cancel_delayed_work(&mad_agent_priv->timed_work);
 607
 608         spin_lock_irq(&port_priv->reg_lock);
 609         remove_mad_reg_req(mad_agent_priv);
 610         spin_unlock_irq(&port_priv->reg_lock);
 611         idr_lock(&ib_mad_clients);
 612         idr_remove(&ib_mad_clients, mad_agent_priv->agent.hi_tid);
 613         idr_unlock(&ib_mad_clients);
 614
 615         flush_workqueue(port_priv->wq);
 616         ib_cancel_rmpp_recvs(mad_agent_priv);
 617
 618         deref_mad_agent(mad_agent_priv);
 619         wait_for_completion(&mad_agent_priv->comp);
 620
 621         ib_mad_agent_security_cleanup(&mad_agent_priv->agent);
 622
 623         kfree(mad_agent_priv->reg_req);
 624         kfree_rcu(mad_agent_priv, rcu);
 625 }
 626
 627 static void unregister_mad_snoop(struct ib_mad_snoop_private *mad_snoop_priv)
 628 {
 629         struct ib_mad_qp_info *qp_info;
 630         unsigned long flags;
 631
 632         qp_info = mad_snoop_priv->qp_info;
 633         spin_lock_irqsave(&qp_info->snoop_lock, flags);
 634         qp_info->snoop_table[mad_snoop_priv->snoop_index] = NULL;
 635         atomic_dec(&qp_info->snoop_count);
 636         spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
 637
 638         deref_snoop_agent(mad_snoop_priv);
 639         wait_for_completion(&mad_snoop_priv->comp);
 640
 641         ib_mad_agent_security_cleanup(&mad_snoop_priv->agent);
 642
 643         kfree(mad_snoop_priv);
 644 }
 645
 646 /*
 647  * ib_unregister_mad_agent - Unregisters a client from using MAD services
 648  *
 649  * Context: Process context.
 650  */
 651 void ib_unregister_mad_agent(struct ib_mad_agent *mad_agent)
 652 {
 653         struct ib_mad_agent_private *mad_agent_priv;
 654         struct ib_mad_snoop_private *mad_snoop_priv;
 655
 656         /* If the TID is zero, the agent can only snoop. */
 657         if (mad_agent->hi_tid) {
 658                 mad_agent_priv = container_of(mad_agent,
 659                                               struct ib_mad_agent_private,
 660                                               agent);
 661                 unregister_mad_agent(mad_agent_priv);
 662         } else {
 663                 mad_snoop_priv = container_of(mad_agent,
 664                                               struct ib_mad_snoop_private,
 665                                               agent);
 666                 unregister_mad_snoop(mad_snoop_priv);
 667         }
 668 }
 669 EXPORT_SYMBOL(ib_unregister_mad_agent);
 670
 671 static void dequeue_mad(struct ib_mad_list_head *mad_list)
 672 {
 673         struct ib_mad_queue *mad_queue;
 674         unsigned long flags;
 675
 676         mad_queue = mad_list->mad_queue;
 677         spin_lock_irqsave(&mad_queue->lock, flags);
 678         list_del(&mad_list->list);
 679         mad_queue->count--;
 680         spin_unlock_irqrestore(&mad_queue->lock, flags);
 681 }
 682
 683 static void snoop_send(struct ib_mad_qp_info *qp_info,
 684                        struct ib_mad_send_buf *send_buf,
 685                        struct ib_mad_send_wc *mad_send_wc,
 686                        int mad_snoop_flags)
 687 {
 688         struct ib_mad_snoop_private *mad_snoop_priv;
 689         unsigned long flags;
 690         int i;
 691
 692         spin_lock_irqsave(&qp_info->snoop_lock, flags);
 693         for (i = 0; i < qp_info->snoop_table_size; i++) {
 694                 mad_snoop_priv = qp_info->snoop_table[i];
 695                 if (!mad_snoop_priv ||
 696                     !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
 697                         continue;
 698
 699                 atomic_inc(&mad_snoop_priv->refcount);
 700                 spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
 701                 mad_snoop_priv->agent.snoop_handler(&mad_snoop_priv->agent,
 702                                                     send_buf, mad_send_wc);
 703                 deref_snoop_agent(mad_snoop_priv);
 704                 spin_lock_irqsave(&qp_info->snoop_lock, flags);
 705         }
 706         spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
 707 }
 708
 709 static void snoop_recv(struct ib_mad_qp_info *qp_info,
 710                        struct ib_mad_recv_wc *mad_recv_wc,
 711                        int mad_snoop_flags)
 712 {
 713         struct ib_mad_snoop_private *mad_snoop_priv;
 714         unsigned long flags;
 715         int i;
 716
 717         spin_lock_irqsave(&qp_info->snoop_lock, flags);
 718         for (i = 0; i < qp_info->snoop_table_size; i++) {
 719                 mad_snoop_priv = qp_info->snoop_table[i];
 720                 if (!mad_snoop_priv ||
 721                     !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
 722                         continue;
 723
 724                 atomic_inc(&mad_snoop_priv->refcount);
 725                 spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
 726                 mad_snoop_priv->agent.recv_handler(&mad_snoop_priv->agent, NULL,
 727                                                    mad_recv_wc);
 728                 deref_snoop_agent(mad_snoop_priv);
 729                 spin_lock_irqsave(&qp_info->snoop_lock, flags);
 730         }
 731         spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
 732 }
 733
 734 static void build_smp_wc(struct ib_qp *qp, struct ib_cqe *cqe, u16 slid,
 735                 u16 pkey_index, u8 port_num, struct ib_wc *wc)
 736 {
 737         memset(wc, 0, sizeof *wc);
 738         wc->wr_cqe = cqe;
 739         wc->status = IB_WC_SUCCESS;
 740         wc->opcode = IB_WC_RECV;
 741         wc->pkey_index = pkey_index;
 742         wc->byte_len = sizeof(struct ib_mad) + sizeof(struct ib_grh);
 743         wc->src_qp = IB_QP0;
 744         wc->qp = qp;
 745         wc->slid = slid;
 746         wc->sl = 0;
 747         wc->dlid_path_bits = 0;
 748         wc->port_num = port_num;
 749 }
 750
 751 static size_t mad_priv_size(const struct ib_mad_private *mp)
 752 {
 753         return sizeof(struct ib_mad_private) + mp->mad_size;
 754 }
 755
 756 static struct ib_mad_private *alloc_mad_private(size_t mad_size, gfp_t flags)
 757 {
 758         size_t size = sizeof(struct ib_mad_private) + mad_size;
 759         struct ib_mad_private *ret = kzalloc(size, flags);
 760
 761         if (ret)
 762                 ret->mad_size = mad_size;
 763
 764         return ret;
 765 }
 766
 767 static size_t port_mad_size(const struct ib_mad_port_private *port_priv)
 768 {
 769         return rdma_max_mad_size(port_priv->device, port_priv->port_num);
 770 }
 771
 772 static size_t mad_priv_dma_size(const struct ib_mad_private *mp)
 773 {
 774         return sizeof(struct ib_grh) + mp->mad_size;
 775 }
 776
 777 /*
 778  * Return 0 if SMP is to be sent
 779  * Return 1 if SMP was consumed locally (whether or not solicited)
 780  * Return < 0 if error
 781  */
 782 static int handle_outgoing_dr_smp(struct ib_mad_agent_private *mad_agent_priv,
 783                                   struct ib_mad_send_wr_private *mad_send_wr)
 784 {
 785         int ret = 0;
 786         struct ib_smp *smp = mad_send_wr->send_buf.mad;
 787         struct opa_smp *opa_smp = (struct opa_smp *)smp;
 788         unsigned long flags;
 789         struct ib_mad_local_private *local;
 790         struct ib_mad_private *mad_priv;
 791         struct ib_mad_port_private *port_priv;
 792         struct ib_mad_agent_private *recv_mad_agent = NULL;
 793         struct ib_device *device = mad_agent_priv->agent.device;
 794         u8 port_num;
 795         struct ib_wc mad_wc;
 796         struct ib_ud_wr *send_wr = &mad_send_wr->send_wr;
 797         size_t mad_size = port_mad_size(mad_agent_priv->qp_info->port_priv);
 798         u16 out_mad_pkey_index = 0;
 799         u16 drslid;
 800         bool opa = rdma_cap_opa_mad(mad_agent_priv->qp_info->port_priv->device,
 801                                     mad_agent_priv->qp_info->port_priv->port_num);
 802
 803         if (rdma_cap_ib_switch(device) &&
 804             smp->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
 805                 port_num = send_wr->port_num;
 806         else
 807                 port_num = mad_agent_priv->agent.port_num;
 808
 809         /*
 810          * Directed route handling starts if the initial LID routed part of
 811          * a request or the ending LID routed part of a response is empty.
 812          * If we are at the start of the LID routed part, don't update the
 813          * hop_ptr or hop_cnt.  See section 14.2.2, Vol 1 IB spec.
 814          */
 815         if (opa && smp->class_version == OPA_SM_CLASS_VERSION) {
 816                 u32 opa_drslid;
 817
 818                 if ((opa_get_smp_direction(opa_smp)
 819                      ? opa_smp->route.dr.dr_dlid : opa_smp->route.dr.dr_slid) ==
 820                      OPA_LID_PERMISSIVE &&
 821                      opa_smi_handle_dr_smp_send(opa_smp,
 822                                                 rdma_cap_ib_switch(device),
 823                                                 port_num) == IB_SMI_DISCARD) {
 824                         ret = -EINVAL;
 825                         dev_err(&device->dev, "OPA Invalid directed route\n");
 826                         goto out;
 827                 }
 828                 opa_drslid = be32_to_cpu(opa_smp->route.dr.dr_slid);
 829                 if (opa_drslid != be32_to_cpu(OPA_LID_PERMISSIVE) &&
 830                     opa_drslid & 0xffff0000) {
 831                         ret = -EINVAL;
 832                         dev_err(&device->dev, "OPA Invalid dr_slid 0x%x\n",
 833                                opa_drslid);
 834                         goto out;
 835                 }
 836                 drslid = (u16)(opa_drslid & 0x0000ffff);
 837
 838                 /* Check to post send on QP or process locally */
 839                 if (opa_smi_check_local_smp(opa_smp, device) == IB_SMI_DISCARD &&
 840                     opa_smi_check_local_returning_smp(opa_smp, device) == IB_SMI_DISCARD)
 841                         goto out;
 842         } else {
 843                 if ((ib_get_smp_direction(smp) ? smp->dr_dlid : smp->dr_slid) ==
 844                      IB_LID_PERMISSIVE &&
 845                      smi_handle_dr_smp_send(smp, rdma_cap_ib_switch(device), port_num) ==
 846                      IB_SMI_DISCARD) {
 847                         ret = -EINVAL;
 848                         dev_err(&device->dev, "Invalid directed route\n");
 849                         goto out;
 850                 }
 851                 drslid = be16_to_cpu(smp->dr_slid);
 852
 853                 /* Check to post send on QP or process locally */
 854                 if (smi_check_local_smp(smp, device) == IB_SMI_DISCARD &&
 855                     smi_check_local_returning_smp(smp, device) == IB_SMI_DISCARD)
 856                         goto out;
 857         }
 858
 859         local = kmalloc(sizeof *local, GFP_ATOMIC);
 860         if (!local) {
 861                 ret = -ENOMEM;
 862                 goto out;
 863         }
 864         local->mad_priv = NULL;
 865         local->recv_mad_agent = NULL;
 866         mad_priv = alloc_mad_private(mad_size, GFP_ATOMIC);
 867         if (!mad_priv) {
 868                 ret = -ENOMEM;
 869                 kfree(local);
 870                 goto out;
 871         }
 872
 873         build_smp_wc(mad_agent_priv->agent.qp,
 874                      send_wr->wr.wr_cqe, drslid,
 875                      send_wr->pkey_index,
 876                      send_wr->port_num, &mad_wc);
 877
 878         if (opa && smp->base_version == OPA_MGMT_BASE_VERSION) {
 879                 mad_wc.byte_len = mad_send_wr->send_buf.hdr_len
 880                                         + mad_send_wr->send_buf.data_len
 881                                         + sizeof(struct ib_grh);
 882         }
 883
 884         /* No GRH for DR SMP */
 885         ret = device->process_mad(device, 0, port_num, &mad_wc, NULL,
 886                                   (const struct ib_mad_hdr *)smp, mad_size,
 887                                   (struct ib_mad_hdr *)mad_priv->mad,
 888                                   &mad_size, &out_mad_pkey_index);
 889         switch (ret)
 890         {
 891         case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY:
 892                 if (ib_response_mad((const struct ib_mad_hdr *)mad_priv->mad) &&
 893                     mad_agent_priv->agent.recv_handler) {
 894                         local->mad_priv = mad_priv;
 895                         local->recv_mad_agent = mad_agent_priv;
 896                         /*
 897                          * Reference MAD agent until receive
 898                          * side of local completion handled
 899                          */
 900                         atomic_inc(&mad_agent_priv->refcount);
 901                 } else
 902                         kfree(mad_priv);
 903                 break;
 904         case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED:
 905                 kfree(mad_priv);
 906                 break;
 907         case IB_MAD_RESULT_SUCCESS:
 908                 /* Treat like an incoming receive MAD */
 909                 port_priv = ib_get_mad_port(mad_agent_priv->agent.device,
 910                                             mad_agent_priv->agent.port_num);
 911                 if (port_priv) {
 912                         memcpy(mad_priv->mad, smp, mad_priv->mad_size);
 913                         recv_mad_agent = find_mad_agent(port_priv,
 914                                                         (const struct ib_mad_hdr *)mad_priv->mad);
 915                 }
 916                 if (!port_priv || !recv_mad_agent) {
 917                         /*
 918                          * No receiving agent so drop packet and
 919                          * generate send completion.
 920                          */
 921                         kfree(mad_priv);
 922                         break;
 923                 }
 924                 local->mad_priv = mad_priv;
 925                 local->recv_mad_agent = recv_mad_agent;
 926                 break;
 927         default:
 928                 kfree(mad_priv);
 929                 kfree(local);
 930                 ret = -EINVAL;
 931                 goto out;
 932         }
 933
 934         local->mad_send_wr = mad_send_wr;
 935         if (opa) {
 936                 local->mad_send_wr->send_wr.pkey_index = out_mad_pkey_index;
 937                 local->return_wc_byte_len = mad_size;
 938         }
 939         /* Reference MAD agent until send side of local completion handled */
 940         atomic_inc(&mad_agent_priv->refcount);
 941         /* Queue local completion to local list */
 942         spin_lock_irqsave(&mad_agent_priv->lock, flags);
 943         list_add_tail(&local->completion_list, &mad_agent_priv->local_list);
 944         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
 945         queue_work(mad_agent_priv->qp_info->port_priv->wq,
 946                    &mad_agent_priv->local_work);
 947         ret = 1;
 948 out:
 949         return ret;
 950 }
 951
 952 static int get_pad_size(int hdr_len, int data_len, size_t mad_size)
 953 {
 954         int seg_size, pad;
 955
 956         seg_size = mad_size - hdr_len;
 957         if (data_len && seg_size) {
 958                 pad = seg_size - data_len % seg_size;
 959                 return pad == seg_size ? 0 : pad;
 960         } else
 961                 return seg_size;
 962 }
 963
 964 static void free_send_rmpp_list(struct ib_mad_send_wr_private *mad_send_wr)
 965 {
 966         struct ib_rmpp_segment *s, *t;
 967
 968         list_for_each_entry_safe(s, t, &mad_send_wr->rmpp_list, list) {
 969                 list_del(&s->list);
 970                 kfree(s);
 971         }
 972 }
 973
 974 static int alloc_send_rmpp_list(struct ib_mad_send_wr_private *send_wr,
 975                                 size_t mad_size, gfp_t gfp_mask)
 976 {
 977         struct ib_mad_send_buf *send_buf = &send_wr->send_buf;
 978         struct ib_rmpp_mad *rmpp_mad = send_buf->mad;
 979         struct ib_rmpp_segment *seg = NULL;
 980         int left, seg_size, pad;
 981
 982         send_buf->seg_size = mad_size - send_buf->hdr_len;
 983         send_buf->seg_rmpp_size = mad_size - IB_MGMT_RMPP_HDR;
 984         seg_size = send_buf->seg_size;
 985         pad = send_wr->pad;
 986
 987         /* Allocate data segments. */
 988         for (left = send_buf->data_len + pad; left > 0; left -= seg_size) {
 989                 seg = kmalloc(sizeof (*seg) + seg_size, gfp_mask);
 990                 if (!seg) {
 991                         free_send_rmpp_list(send_wr);
 992                         return -ENOMEM;
 993                 }
 994                 seg->num = ++send_buf->seg_count;
 995                 list_add_tail(&seg->list, &send_wr->rmpp_list);
 996         }
 997
 998         /* Zero any padding */
 999         if (pad)
1000                 memset(seg->data + seg_size - pad, 0, pad);
1001
1002         rmpp_mad->rmpp_hdr.rmpp_version = send_wr->mad_agent_priv->
1003                                           agent.rmpp_version;
1004         rmpp_mad->rmpp_hdr.rmpp_type = IB_MGMT_RMPP_TYPE_DATA;
1005         ib_set_rmpp_flags(&rmpp_mad->rmpp_hdr, IB_MGMT_RMPP_FLAG_ACTIVE);
1006
1007         send_wr->cur_seg = container_of(send_wr->rmpp_list.next,
1008                                         struct ib_rmpp_segment, list);
1009         send_wr->last_ack_seg = send_wr->cur_seg;
1010         return 0;
1011 }
1012
1013 int ib_mad_kernel_rmpp_agent(const struct ib_mad_agent *agent)
1014 {
1015         return agent->rmpp_version && !(agent->flags & IB_MAD_USER_RMPP);
1016 }
1017 EXPORT_SYMBOL(ib_mad_kernel_rmpp_agent);
1018
1019 struct ib_mad_send_buf * ib_create_send_mad(struct ib_mad_agent *mad_agent,
1020                                             u32 remote_qpn, u16 pkey_index,
1021                                             int rmpp_active,
1022                                             int hdr_len, int data_len,
1023                                             gfp_t gfp_mask,
1024                                             u8 base_version)
1025 {
1026         struct ib_mad_agent_private *mad_agent_priv;
1027         struct ib_mad_send_wr_private *mad_send_wr;
1028         int pad, message_size, ret, size;
1029         void *buf;
1030         size_t mad_size;
1031         bool opa;
1032
1033         mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
1034                                       agent);
1035
1036         opa = rdma_cap_opa_mad(mad_agent->device, mad_agent->port_num);
1037
1038         if (opa && base_version == OPA_MGMT_BASE_VERSION)
1039                 mad_size = sizeof(struct opa_mad);
1040         else
1041                 mad_size = sizeof(struct ib_mad);
1042
1043         pad = get_pad_size(hdr_len, data_len, mad_size);
1044         message_size = hdr_len + data_len + pad;
1045
1046         if (ib_mad_kernel_rmpp_agent(mad_agent)) {
1047                 if (!rmpp_active && message_size > mad_size)
1048                         return ERR_PTR(-EINVAL);
1049         } else
1050                 if (rmpp_active || message_size > mad_size)
1051                         return ERR_PTR(-EINVAL);
1052
1053         size = rmpp_active ? hdr_len : mad_size;
1054         buf = kzalloc(sizeof *mad_send_wr + size, gfp_mask);
1055         if (!buf)
1056                 return ERR_PTR(-ENOMEM);
1057
1058         mad_send_wr = buf + size;
1059         INIT_LIST_HEAD(&mad_send_wr->rmpp_list);
1060         mad_send_wr->send_buf.mad = buf;
1061         mad_send_wr->send_buf.hdr_len = hdr_len;
1062         mad_send_wr->send_buf.data_len = data_len;
1063         mad_send_wr->pad = pad;
1064
1065         mad_send_wr->mad_agent_priv = mad_agent_priv;
1066         mad_send_wr->sg_list[0].length = hdr_len;
1067         mad_send_wr->sg_list[0].lkey = mad_agent->qp->pd->local_dma_lkey;
1068
1069         /* OPA MADs don't have to be the full 2048 bytes */
1070         if (opa && base_version == OPA_MGMT_BASE_VERSION &&
1071             data_len < mad_size - hdr_len)
1072                 mad_send_wr->sg_list[1].length = data_len;
1073         else
1074                 mad_send_wr->sg_list[1].length = mad_size - hdr_len;
1075
1076         mad_send_wr->sg_list[1].lkey = mad_agent->qp->pd->local_dma_lkey;
1077
1078         mad_send_wr->mad_list.cqe.done = ib_mad_send_done;
1079
1080         mad_send_wr->send_wr.wr.wr_cqe = &mad_send_wr->mad_list.cqe;
1081         mad_send_wr->send_wr.wr.sg_list = mad_send_wr->sg_list;
1082         mad_send_wr->send_wr.wr.num_sge = 2;
1083         mad_send_wr->send_wr.wr.opcode = IB_WR_SEND;
1084         mad_send_wr->send_wr.wr.send_flags = IB_SEND_SIGNALED;
1085         mad_send_wr->send_wr.remote_qpn = remote_qpn;
1086         mad_send_wr->send_wr.remote_qkey = IB_QP_SET_QKEY;
1087         mad_send_wr->send_wr.pkey_index = pkey_index;
1088
1089         if (rmpp_active) {
1090                 ret = alloc_send_rmpp_list(mad_send_wr, mad_size, gfp_mask);
1091                 if (ret) {
1092                         kfree(buf);
1093                         return ERR_PTR(ret);
1094                 }
1095         }
1096
1097         mad_send_wr->send_buf.mad_agent = mad_agent;
1098         atomic_inc(&mad_agent_priv->refcount);
1099         return &mad_send_wr->send_buf;
1100 }
1101 EXPORT_SYMBOL(ib_create_send_mad);
1102
1103 int ib_get_mad_data_offset(u8 mgmt_class)
1104 {
1105         if (mgmt_class == IB_MGMT_CLASS_SUBN_ADM)
1106                 return IB_MGMT_SA_HDR;
1107         else if ((mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
1108                  (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
1109                  (mgmt_class == IB_MGMT_CLASS_BIS))
1110                 return IB_MGMT_DEVICE_HDR;
1111         else if ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
1112                  (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END))
1113                 return IB_MGMT_VENDOR_HDR;
1114         else
1115                 return IB_MGMT_MAD_HDR;
1116 }
1117 EXPORT_SYMBOL(ib_get_mad_data_offset);
1118
1119 int ib_is_mad_class_rmpp(u8 mgmt_class)
1120 {
1121         if ((mgmt_class == IB_MGMT_CLASS_SUBN_ADM) ||
1122             (mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
1123             (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
1124             (mgmt_class == IB_MGMT_CLASS_BIS) ||
1125             ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
1126              (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END)))
1127                 return 1;
1128         return 0;
1129 }
1130 EXPORT_SYMBOL(ib_is_mad_class_rmpp);
1131
1132 void *ib_get_rmpp_segment(struct ib_mad_send_buf *send_buf, int seg_num)
1133 {
1134         struct ib_mad_send_wr_private *mad_send_wr;
1135         struct list_head *list;
1136
1137         mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
1138                                    send_buf);
1139         list = &mad_send_wr->cur_seg->list;
1140
1141         if (mad_send_wr->cur_seg->num < seg_num) {
1142                 list_for_each_entry(mad_send_wr->cur_seg, list, list)
1143                         if (mad_send_wr->cur_seg->num == seg_num)
1144                                 break;
1145         } else if (mad_send_wr->cur_seg->num > seg_num) {
1146                 list_for_each_entry_reverse(mad_send_wr->cur_seg, list, list)
1147                         if (mad_send_wr->cur_seg->num == seg_num)
1148                                 break;
1149         }
1150         return mad_send_wr->cur_seg->data;
1151 }
1152 EXPORT_SYMBOL(ib_get_rmpp_segment);
1153
1154 static inline void *ib_get_payload(struct ib_mad_send_wr_private *mad_send_wr)
1155 {
1156         if (mad_send_wr->send_buf.seg_count)
1157                 return ib_get_rmpp_segment(&mad_send_wr->send_buf,
1158                                            mad_send_wr->seg_num);
1159         else
1160                 return mad_send_wr->send_buf.mad +
1161                        mad_send_wr->send_buf.hdr_len;
1162 }
1163
1164 void ib_free_send_mad(struct ib_mad_send_buf *send_buf)
1165 {
1166         struct ib_mad_agent_private *mad_agent_priv;
1167         struct ib_mad_send_wr_private *mad_send_wr;
1168
1169         mad_agent_priv = container_of(send_buf->mad_agent,
1170                                       struct ib_mad_agent_private, agent);
1171         mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
1172                                    send_buf);
1173
1174         free_send_rmpp_list(mad_send_wr);
1175         kfree(send_buf->mad);
1176         deref_mad_agent(mad_agent_priv);
1177 }
1178 EXPORT_SYMBOL(ib_free_send_mad);
1179
1180 int ib_send_mad(struct ib_mad_send_wr_private *mad_send_wr)
1181 {
1182         struct ib_mad_qp_info *qp_info;
1183         struct list_head *list;
1184         struct ib_mad_agent *mad_agent;
1185         struct ib_sge *sge;
1186         unsigned long flags;
1187         int ret;
1188
1189         /* Set WR ID to find mad_send_wr upon completion */
1190         qp_info = mad_send_wr->mad_agent_priv->qp_info;
1191         mad_send_wr->mad_list.mad_queue = &qp_info->send_queue;
1192         mad_send_wr->mad_list.cqe.done = ib_mad_send_done;
1193         mad_send_wr->send_wr.wr.wr_cqe = &mad_send_wr->mad_list.cqe;
1194
1195         mad_agent = mad_send_wr->send_buf.mad_agent;
1196         sge = mad_send_wr->sg_list;
1197         sge[0].addr = ib_dma_map_single(mad_agent->device,
1198                                         mad_send_wr->send_buf.mad,
1199                                         sge[0].length,
1200                                         DMA_TO_DEVICE);
1201         if (unlikely(ib_dma_mapping_error(mad_agent->device, sge[0].addr)))
1202                 return -ENOMEM;
1203
1204         mad_send_wr->header_mapping = sge[0].addr;
1205
1206         sge[1].addr = ib_dma_map_single(mad_agent->device,
1207                                         ib_get_payload(mad_send_wr),
1208                                         sge[1].length,
1209                                         DMA_TO_DEVICE);
1210         if (unlikely(ib_dma_mapping_error(mad_agent->device, sge[1].addr))) {
1211                 ib_dma_unmap_single(mad_agent->device,
1212                                     mad_send_wr->header_mapping,
1213                                     sge[0].length, DMA_TO_DEVICE);
1214                 return -ENOMEM;
1215         }
1216         mad_send_wr->payload_mapping = sge[1].addr;
1217
1218         spin_lock_irqsave(&qp_info->send_queue.lock, flags);
1219         if (qp_info->send_queue.count < qp_info->send_queue.max_active) {
1220                 ret = ib_post_send(mad_agent->qp, &mad_send_wr->send_wr.wr,
1221                                    NULL);
1222                 list = &qp_info->send_queue.list;
1223         } else {
1224                 ret = 0;
1225                 list = &qp_info->overflow_list;
1226         }
1227
1228         if (!ret) {
1229                 qp_info->send_queue.count++;
1230                 list_add_tail(&mad_send_wr->mad_list.list, list);
1231         }
1232         spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
1233         if (ret) {
1234                 ib_dma_unmap_single(mad_agent->device,
1235                                     mad_send_wr->header_mapping,
1236                                     sge[0].length, DMA_TO_DEVICE);
1237                 ib_dma_unmap_single(mad_agent->device,
1238                                     mad_send_wr->payload_mapping,
1239                                     sge[1].length, DMA_TO_DEVICE);
1240         }
1241         return ret;
1242 }
1243
1244 /*
1245  * ib_post_send_mad - Posts MAD(s) to the send queue of the QP associated
1246  *  with the registered client
1247  */
1248 int ib_post_send_mad(struct ib_mad_send_buf *send_buf,
1249                      struct ib_mad_send_buf **bad_send_buf)
1250 {
1251         struct ib_mad_agent_private *mad_agent_priv;
1252         struct ib_mad_send_buf *next_send_buf;
1253         struct ib_mad_send_wr_private *mad_send_wr;
1254         unsigned long flags;
1255         int ret = -EINVAL;
1256
1257         /* Walk list of send WRs and post each on send list */
1258         for (; send_buf; send_buf = next_send_buf) {
1259                 mad_send_wr = container_of(send_buf,
1260                                            struct ib_mad_send_wr_private,
1261                                            send_buf);
1262                 mad_agent_priv = mad_send_wr->mad_agent_priv;
1263
1264                 ret = ib_mad_enforce_security(mad_agent_priv,
1265                                               mad_send_wr->send_wr.pkey_index);
1266                 if (ret)
1267                         goto error;
1268
1269                 if (!send_buf->mad_agent->send_handler ||
1270                     (send_buf->timeout_ms &&
1271                      !send_buf->mad_agent->recv_handler)) {
1272                         ret = -EINVAL;
1273                         goto error;
1274                 }
1275
1276                 if (!ib_is_mad_class_rmpp(((struct ib_mad_hdr *) send_buf->mad)->mgmt_class)) {
1277                         if (mad_agent_priv->agent.rmpp_version) {
1278                                 ret = -EINVAL;
1279                                 goto error;
1280                         }
1281                 }
1282
1283                 /*
1284                  * Save pointer to next work request to post in case the
1285                  * current one completes, and the user modifies the work
1286                  * request associated with the completion
1287                  */
1288                 next_send_buf = send_buf->next;
1289                 mad_send_wr->send_wr.ah = send_buf->ah;
1290
1291                 if (((struct ib_mad_hdr *) send_buf->mad)->mgmt_class ==
1292                     IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
1293                         ret = handle_outgoing_dr_smp(mad_agent_priv,
1294                                                      mad_send_wr);
1295                         if (ret < 0)            /* error */
1296                                 goto error;
1297                         else if (ret == 1)      /* locally consumed */
1298                                 continue;
1299                 }
1300
1301                 mad_send_wr->tid = ((struct ib_mad_hdr *) send_buf->mad)->tid;
1302                 /* Timeout will be updated after send completes */
1303                 mad_send_wr->timeout = msecs_to_jiffies(send_buf->timeout_ms);
1304                 mad_send_wr->max_retries = send_buf->retries;
1305                 mad_send_wr->retries_left = send_buf->retries;
1306                 send_buf->retries = 0;
1307                 /* Reference for work request to QP + response */
1308                 mad_send_wr->refcount = 1 + (mad_send_wr->timeout > 0);
1309                 mad_send_wr->status = IB_WC_SUCCESS;
1310
1311                 /* Reference MAD agent until send completes */
1312                 atomic_inc(&mad_agent_priv->refcount);
1313                 spin_lock_irqsave(&mad_agent_priv->lock, flags);
1314                 list_add_tail(&mad_send_wr->agent_list,
1315                               &mad_agent_priv->send_list);
1316                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1317
1318                 if (ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)) {
1319                         ret = ib_send_rmpp_mad(mad_send_wr);
1320                         if (ret >= 0 && ret != IB_RMPP_RESULT_CONSUMED)
1321                                 ret = ib_send_mad(mad_send_wr);
1322                 } else
1323                         ret = ib_send_mad(mad_send_wr);
1324                 if (ret < 0) {
1325                         /* Fail send request */
1326                         spin_lock_irqsave(&mad_agent_priv->lock, flags);
1327                         list_del(&mad_send_wr->agent_list);
1328                         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1329                         atomic_dec(&mad_agent_priv->refcount);
1330                         goto error;
1331                 }
1332         }
1333         return 0;
1334 error:
1335         if (bad_send_buf)
1336                 *bad_send_buf = send_buf;
1337         return ret;
1338 }
1339 EXPORT_SYMBOL(ib_post_send_mad);
1340
1341 /*
1342  * ib_free_recv_mad - Returns data buffers used to receive
1343  *  a MAD to the access layer
1344  */
1345 void ib_free_recv_mad(struct ib_mad_recv_wc *mad_recv_wc)
1346 {
1347         struct ib_mad_recv_buf *mad_recv_buf, *temp_recv_buf;
1348         struct ib_mad_private_header *mad_priv_hdr;
1349         struct ib_mad_private *priv;
1350         struct list_head free_list;
1351
1352         INIT_LIST_HEAD(&free_list);
1353         list_splice_init(&mad_recv_wc->rmpp_list, &free_list);
1354
1355         list_for_each_entry_safe(mad_recv_buf, temp_recv_buf,
1356                                         &free_list, list) {
1357                 mad_recv_wc = container_of(mad_recv_buf, struct ib_mad_recv_wc,
1358                                            recv_buf);
1359                 mad_priv_hdr = container_of(mad_recv_wc,
1360                                             struct ib_mad_private_header,
1361                                             recv_wc);
1362                 priv = container_of(mad_priv_hdr, struct ib_mad_private,
1363                                     header);
1364                 kfree(priv);
1365         }
1366 }
1367 EXPORT_SYMBOL(ib_free_recv_mad);
1368
1369 struct ib_mad_agent *ib_redirect_mad_qp(struct ib_qp *qp,
1370                                         u8 rmpp_version,
1371                                         ib_mad_send_handler send_handler,
1372                                         ib_mad_recv_handler recv_handler,
1373                                         void *context)
1374 {
1375         return ERR_PTR(-EINVAL);        /* XXX: for now */
1376 }
1377 EXPORT_SYMBOL(ib_redirect_mad_qp);
1378
1379 int ib_process_mad_wc(struct ib_mad_agent *mad_agent,
1380                       struct ib_wc *wc)
1381 {
1382         dev_err(&mad_agent->device->dev,
1383                 "ib_process_mad_wc() not implemented yet\n");
1384         return 0;
1385 }
1386 EXPORT_SYMBOL(ib_process_mad_wc);
1387
1388 static int method_in_use(struct ib_mad_mgmt_method_table **method,
1389                          struct ib_mad_reg_req *mad_reg_req)
1390 {
1391         int i;
1392
1393         for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS) {
1394                 if ((*method)->agent[i]) {
1395                         pr_err("Method %d already in use\n", i);
1396                         return -EINVAL;
1397                 }
1398         }
1399         return 0;
1400 }
1401
1402 static int allocate_method_table(struct ib_mad_mgmt_method_table **method)
1403 {
1404         /* Allocate management method table */
1405         *method = kzalloc(sizeof **method, GFP_ATOMIC);
1406         return (*method) ? 0 : (-ENOMEM);
1407 }
1408
1409 /*
1410  * Check to see if there are any methods still in use
1411  */
1412 static int check_method_table(struct ib_mad_mgmt_method_table *method)
1413 {
1414         int i;
1415
1416         for (i = 0; i < IB_MGMT_MAX_METHODS; i++)
1417                 if (method->agent[i])
1418                         return 1;
1419         return 0;
1420 }
1421
1422 /*
1423  * Check to see if there are any method tables for this class still in use
1424  */
1425 static int check_class_table(struct ib_mad_mgmt_class_table *class)
1426 {
1427         int i;
1428
1429         for (i = 0; i < MAX_MGMT_CLASS; i++)
1430                 if (class->method_table[i])
1431                         return 1;
1432         return 0;
1433 }
1434
1435 static int check_vendor_class(struct ib_mad_mgmt_vendor_class *vendor_class)
1436 {
1437         int i;
1438
1439         for (i = 0; i < MAX_MGMT_OUI; i++)
1440                 if (vendor_class->method_table[i])
1441                         return 1;
1442         return 0;
1443 }
1444
1445 static int find_vendor_oui(struct ib_mad_mgmt_vendor_class *vendor_class,
1446                            const char *oui)
1447 {
1448         int i;
1449
1450         for (i = 0; i < MAX_MGMT_OUI; i++)
1451                 /* Is there matching OUI for this vendor class ? */
1452                 if (!memcmp(vendor_class->oui[i], oui, 3))
1453                         return i;
1454
1455         return -1;
1456 }
1457
1458 static int check_vendor_table(struct ib_mad_mgmt_vendor_class_table *vendor)
1459 {
1460         int i;
1461
1462         for (i = 0; i < MAX_MGMT_VENDOR_RANGE2; i++)
1463                 if (vendor->vendor_class[i])
1464                         return 1;
1465
1466         return 0;
1467 }
1468
1469 static void remove_methods_mad_agent(struct ib_mad_mgmt_method_table *method,
1470                                      struct ib_mad_agent_private *agent)
1471 {
1472         int i;
1473
1474         /* Remove any methods for this mad agent */
1475         for (i = 0; i < IB_MGMT_MAX_METHODS; i++) {
1476                 if (method->agent[i] == agent) {
1477                         method->agent[i] = NULL;
1478                 }
1479         }
1480 }
1481
1482 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1483                               struct ib_mad_agent_private *agent_priv,
1484                               u8 mgmt_class)
1485 {
1486         struct ib_mad_port_private *port_priv;
1487         struct ib_mad_mgmt_class_table **class;
1488         struct ib_mad_mgmt_method_table **method;
1489         int i, ret;
1490
1491         port_priv = agent_priv->qp_info->port_priv;
1492         class = &port_priv->version[mad_reg_req->mgmt_class_version].class;
1493         if (!*class) {
1494                 /* Allocate management class table for "new" class version */
1495                 *class = kzalloc(sizeof **class, GFP_ATOMIC);
1496                 if (!*class) {
1497                         ret = -ENOMEM;
1498                         goto error1;
1499                 }
1500
1501                 /* Allocate method table for this management class */
1502                 method = &(*class)->method_table[mgmt_class];
1503                 if ((ret = allocate_method_table(method)))
1504                         goto error2;
1505         } else {
1506                 method = &(*class)->method_table[mgmt_class];
1507                 if (!*method) {
1508                         /* Allocate method table for this management class */
1509                         if ((ret = allocate_method_table(method)))
1510                                 goto error1;
1511                 }
1512         }
1513
1514         /* Now, make sure methods are not already in use */
1515         if (method_in_use(method, mad_reg_req))
1516                 goto error3;
1517
1518         /* Finally, add in methods being registered */
1519         for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS)
1520                 (*method)->agent[i] = agent_priv;
1521
1522         return 0;
1523
1524 error3:
1525         /* Remove any methods for this mad agent */
1526         remove_methods_mad_agent(*method, agent_priv);
1527         /* Now, check to see if there are any methods in use */
1528         if (!check_method_table(*method)) {
1529                 /* If not, release management method table */
1530                 kfree(*method);
1531                 *method = NULL;
1532         }
1533         ret = -EINVAL;
1534         goto error1;
1535 error2:
1536         kfree(*class);
1537         *class = NULL;
1538 error1:
1539         return ret;
1540 }
1541
1542 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1543                            struct ib_mad_agent_private *agent_priv)
1544 {
1545         struct ib_mad_port_private *port_priv;
1546         struct ib_mad_mgmt_vendor_class_table **vendor_table;
1547         struct ib_mad_mgmt_vendor_class_table *vendor = NULL;
1548         struct ib_mad_mgmt_vendor_class *vendor_class = NULL;
1549         struct ib_mad_mgmt_method_table **method;
1550         int i, ret = -ENOMEM;
1551         u8 vclass;
1552
1553         /* "New" vendor (with OUI) class */
1554         vclass = vendor_class_index(mad_reg_req->mgmt_class);
1555         port_priv = agent_priv->qp_info->port_priv;
1556         vendor_table = &port_priv->version[
1557                                 mad_reg_req->mgmt_class_version].vendor;
1558         if (!*vendor_table) {
1559                 /* Allocate mgmt vendor class table for "new" class version */
1560                 vendor = kzalloc(sizeof *vendor, GFP_ATOMIC);
1561                 if (!vendor)
1562                         goto error1;
1563
1564                 *vendor_table = vendor;
1565         }
1566         if (!(*vendor_table)->vendor_class[vclass]) {
1567                 /* Allocate table for this management vendor class */
1568                 vendor_class = kzalloc(sizeof *vendor_class, GFP_ATOMIC);
1569                 if (!vendor_class)
1570                         goto error2;
1571
1572                 (*vendor_table)->vendor_class[vclass] = vendor_class;
1573         }
1574         for (i = 0; i < MAX_MGMT_OUI; i++) {
1575                 /* Is there matching OUI for this vendor class ? */
1576                 if (!memcmp((*vendor_table)->vendor_class[vclass]->oui[i],
1577                             mad_reg_req->oui, 3)) {
1578                         method = &(*vendor_table)->vendor_class[
1579                                                 vclass]->method_table[i];
1580                         if (!*method)
1581                                 goto error3;
1582                         goto check_in_use;
1583                 }
1584         }
1585         for (i = 0; i < MAX_MGMT_OUI; i++) {
1586                 /* OUI slot available ? */
1587                 if (!is_vendor_oui((*vendor_table)->vendor_class[
1588                                 vclass]->oui[i])) {
1589                         method = &(*vendor_table)->vendor_class[
1590                                 vclass]->method_table[i];
1591                         /* Allocate method table for this OUI */
1592                         if (!*method) {
1593                                 ret = allocate_method_table(method);
1594                                 if (ret)
1595                                         goto error3;
1596                         }
1597                         memcpy((*vendor_table)->vendor_class[vclass]->oui[i],
1598                                mad_reg_req->oui, 3);
1599                         goto check_in_use;
1600                 }
1601         }
1602         dev_err(&agent_priv->agent.device->dev, "All OUI slots in use\n");
1603         goto error3;
1604
1605 check_in_use:
1606         /* Now, make sure methods are not already in use */
1607         if (method_in_use(method, mad_reg_req))
1608                 goto error4;
1609
1610         /* Finally, add in methods being registered */
1611         for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS)
1612                 (*method)->agent[i] = agent_priv;
1613
1614         return 0;
1615
1616 error4:
1617         /* Remove any methods for this mad agent */
1618         remove_methods_mad_agent(*method, agent_priv);
1619         /* Now, check to see if there are any methods in use */
1620         if (!check_method_table(*method)) {
1621                 /* If not, release management method table */
1622                 kfree(*method);
1623                 *method = NULL;
1624         }
1625         ret = -EINVAL;
1626 error3:
1627         if (vendor_class) {
1628                 (*vendor_table)->vendor_class[vclass] = NULL;
1629                 kfree(vendor_class);
1630         }
1631 error2:
1632         if (vendor) {
1633                 *vendor_table = NULL;
1634                 kfree(vendor);
1635         }
1636 error1:
1637         return ret;
1638 }
1639
1640 static void remove_mad_reg_req(struct ib_mad_agent_private *agent_priv)
1641 {
1642         struct ib_mad_port_private *port_priv;
1643         struct ib_mad_mgmt_class_table *class;
1644         struct ib_mad_mgmt_method_table *method;
1645         struct ib_mad_mgmt_vendor_class_table *vendor;
1646         struct ib_mad_mgmt_vendor_class *vendor_class;
1647         int index;
1648         u8 mgmt_class;
1649
1650         /*
1651          * Was MAD registration request supplied
1652          * with original registration ?
1653          */
1654         if (!agent_priv->reg_req) {
1655                 goto out;
1656         }
1657
1658         port_priv = agent_priv->qp_info->port_priv;
1659         mgmt_class = convert_mgmt_class(agent_priv->reg_req->mgmt_class);
1660         class = port_priv->version[
1661                         agent_priv->reg_req->mgmt_class_version].class;
1662         if (!class)
1663                 goto vendor_check;
1664
1665         method = class->method_table[mgmt_class];
1666         if (method) {
1667                 /* Remove any methods for this mad agent */
1668                 remove_methods_mad_agent(method, agent_priv);
1669                 /* Now, check to see if there are any methods still in use */
1670                 if (!check_method_table(method)) {
1671                         /* If not, release management method table */
1672                         kfree(method);
1673                         class->method_table[mgmt_class] = NULL;
1674                         /* Any management classes left ? */
1675                         if (!check_class_table(class)) {
1676                                 /* If not, release management class table */
1677                                 kfree(class);
1678                                 port_priv->version[
1679                                         agent_priv->reg_req->
1680                                         mgmt_class_version].class = NULL;
1681                         }
1682                 }
1683         }
1684
1685 vendor_check:
1686         if (!is_vendor_class(mgmt_class))
1687                 goto out;
1688
1689         /* normalize mgmt_class to vendor range 2 */
1690         mgmt_class = vendor_class_index(agent_priv->reg_req->mgmt_class);
1691         vendor = port_priv->version[
1692                         agent_priv->reg_req->mgmt_class_version].vendor;
1693
1694         if (!vendor)
1695                 goto out;
1696
1697         vendor_class = vendor->vendor_class[mgmt_class];
1698         if (vendor_class) {
1699                 index = find_vendor_oui(vendor_class, agent_priv->reg_req->oui);
1700                 if (index < 0)
1701                         goto out;
1702                 method = vendor_class->method_table[index];
1703                 if (method) {
1704                         /* Remove any methods for this mad agent */
1705                         remove_methods_mad_agent(method, agent_priv);
1706                         /*
1707                          * Now, check to see if there are
1708                          * any methods still in use
1709                          */
1710                         if (!check_method_table(method)) {
1711                                 /* If not, release management method table */
1712                                 kfree(method);
1713                                 vendor_class->method_table[index] = NULL;
1714                                 memset(vendor_class->oui[index], 0, 3);
1715                                 /* Any OUIs left ? */
1716                                 if (!check_vendor_class(vendor_class)) {
1717                                         /* If not, release vendor class table */
1718                                         kfree(vendor_class);
1719                                         vendor->vendor_class[mgmt_class] = NULL;
1720                                         /* Any other vendor classes left ? */
1721                                         if (!check_vendor_table(vendor)) {
1722                                                 kfree(vendor);
1723                                                 port_priv->version[
1724                                                         agent_priv->reg_req->
1725                                                         mgmt_class_version].
1726                                                         vendor = NULL;
1727                                         }
1728                                 }
1729                         }
1730                 }
1731         }
1732
1733 out:
1734         return;
1735 }
1736
1737 static struct ib_mad_agent_private *
1738 find_mad_agent(struct ib_mad_port_private *port_priv,
1739                const struct ib_mad_hdr *mad_hdr)
1740 {
1741         struct ib_mad_agent_private *mad_agent = NULL;
1742         unsigned long flags;
1743
1744         if (ib_response_mad(mad_hdr)) {
1745                 u32 hi_tid;
1746
1747                 /*
1748                  * Routing is based on high 32 bits of transaction ID
1749                  * of MAD.
1750                  */
1751                 hi_tid = be64_to_cpu(mad_hdr->tid) >> 32;
1752                 rcu_read_lock();
1753                 mad_agent = idr_find(&ib_mad_clients, hi_tid);
1754                 if (mad_agent && !atomic_inc_not_zero(&mad_agent->refcount))
1755                         mad_agent = NULL;
1756                 rcu_read_unlock();
1757         } else {
1758                 struct ib_mad_mgmt_class_table *class;
1759                 struct ib_mad_mgmt_method_table *method;
1760                 struct ib_mad_mgmt_vendor_class_table *vendor;
1761                 struct ib_mad_mgmt_vendor_class *vendor_class;
1762                 const struct ib_vendor_mad *vendor_mad;
1763                 int index;
1764
1765                 spin_lock_irqsave(&port_priv->reg_lock, flags);
1766                 /*
1767                  * Routing is based on version, class, and method
1768                  * For "newer" vendor MADs, also based on OUI
1769                  */
1770                 if (mad_hdr->class_version >= MAX_MGMT_VERSION)
1771                         goto out;
1772                 if (!is_vendor_class(mad_hdr->mgmt_class)) {
1773                         class = port_priv->version[
1774                                         mad_hdr->class_version].class;
1775                         if (!class)
1776                                 goto out;
1777                         if (convert_mgmt_class(mad_hdr->mgmt_class) >=
1778                             ARRAY_SIZE(class->method_table))
1779                                 goto out;
1780                         method = class->method_table[convert_mgmt_class(
1781                                                         mad_hdr->mgmt_class)];
1782                         if (method)
1783                                 mad_agent = method->agent[mad_hdr->method &
1784                                                           ~IB_MGMT_METHOD_RESP];
1785                 } else {
1786                         vendor = port_priv->version[
1787                                         mad_hdr->class_version].vendor;
1788                         if (!vendor)
1789                                 goto out;
1790                         vendor_class = vendor->vendor_class[vendor_class_index(
1791                                                 mad_hdr->mgmt_class)];
1792                         if (!vendor_class)
1793                                 goto out;
1794                         /* Find matching OUI */
1795                         vendor_mad = (const struct ib_vendor_mad *)mad_hdr;
1796                         index = find_vendor_oui(vendor_class, vendor_mad->oui);
1797                         if (index == -1)
1798                                 goto out;
1799                         method = vendor_class->method_table[index];
1800                         if (method) {
1801                                 mad_agent = method->agent[mad_hdr->method &
1802                                                           ~IB_MGMT_METHOD_RESP];
1803                         }
1804                 }
1805                 if (mad_agent)
1806                         atomic_inc(&mad_agent->refcount);
1807 out:
1808                 spin_unlock_irqrestore(&port_priv->reg_lock, flags);
1809         }
1810
1811         if (mad_agent && !mad_agent->agent.recv_handler) {
1812                 dev_notice(&port_priv->device->dev,
1813                            "No receive handler for client %p on port %d\n",
1814                            &mad_agent->agent, port_priv->port_num);
1815                 deref_mad_agent(mad_agent);
1816                 mad_agent = NULL;
1817         }
1818
1819         return mad_agent;
1820 }
1821
1822 static int validate_mad(const struct ib_mad_hdr *mad_hdr,
1823                         const struct ib_mad_qp_info *qp_info,
1824                         bool opa)
1825 {
1826         int valid = 0;
1827         u32 qp_num = qp_info->qp->qp_num;
1828
1829         /* Make sure MAD base version is understood */
1830         if (mad_hdr->base_version != IB_MGMT_BASE_VERSION &&
1831             (!opa || mad_hdr->base_version != OPA_MGMT_BASE_VERSION)) {
1832                 pr_err("MAD received with unsupported base version %d %s\n",
1833                        mad_hdr->base_version, opa ? "(opa)" : "");
1834                 goto out;
1835         }
1836
1837         /* Filter SMI packets sent to other than QP0 */
1838         if ((mad_hdr->mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
1839             (mad_hdr->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
1840                 if (qp_num == 0)
1841                         valid = 1;
1842         } else {
1843                 /* CM attributes other than ClassPortInfo only use Send method */
1844                 if ((mad_hdr->mgmt_class == IB_MGMT_CLASS_CM) &&
1845                     (mad_hdr->attr_id != IB_MGMT_CLASSPORTINFO_ATTR_ID) &&
1846                     (mad_hdr->method != IB_MGMT_METHOD_SEND))
1847                         goto out;
1848                 /* Filter GSI packets sent to QP0 */
1849                 if (qp_num != 0)
1850                         valid = 1;
1851         }
1852
1853 out:
1854         return valid;
1855 }
1856
1857 static int is_rmpp_data_mad(const struct ib_mad_agent_private *mad_agent_priv,
1858                             const struct ib_mad_hdr *mad_hdr)
1859 {
1860         struct ib_rmpp_mad *rmpp_mad;
1861
1862         rmpp_mad = (struct ib_rmpp_mad *)mad_hdr;
1863         return !mad_agent_priv->agent.rmpp_version ||
1864                 !ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent) ||
1865                 !(ib_get_rmpp_flags(&rmpp_mad->rmpp_hdr) &
1866                                     IB_MGMT_RMPP_FLAG_ACTIVE) ||
1867                 (rmpp_mad->rmpp_hdr.rmpp_type == IB_MGMT_RMPP_TYPE_DATA);
1868 }
1869
1870 static inline int rcv_has_same_class(const struct ib_mad_send_wr_private *wr,
1871                                      const struct ib_mad_recv_wc *rwc)
1872 {
1873         return ((struct ib_mad_hdr *)(wr->send_buf.mad))->mgmt_class ==
1874                 rwc->recv_buf.mad->mad_hdr.mgmt_class;
1875 }
1876
1877 static inline int rcv_has_same_gid(const struct ib_mad_agent_private *mad_agent_priv,
1878                                    const struct ib_mad_send_wr_private *wr,
1879                                    const struct ib_mad_recv_wc *rwc )
1880 {
1881         struct rdma_ah_attr attr;
1882         u8 send_resp, rcv_resp;
1883         union ib_gid sgid;
1884         struct ib_device *device = mad_agent_priv->agent.device;
1885         u8 port_num = mad_agent_priv->agent.port_num;
1886         u8 lmc;
1887         bool has_grh;
1888
1889         send_resp = ib_response_mad((struct ib_mad_hdr *)wr->send_buf.mad);
1890         rcv_resp = ib_response_mad(&rwc->recv_buf.mad->mad_hdr);
1891
1892         if (send_resp == rcv_resp)
1893                 /* both requests, or both responses. GIDs different */
1894                 return 0;
1895
1896         if (rdma_query_ah(wr->send_buf.ah, &attr))
1897                 /* Assume not equal, to avoid false positives. */
1898                 return 0;
1899
1900         has_grh = !!(rdma_ah_get_ah_flags(&attr) & IB_AH_GRH);
1901         if (has_grh != !!(rwc->wc->wc_flags & IB_WC_GRH))
1902                 /* one has GID, other does not.  Assume different */
1903                 return 0;
1904
1905         if (!send_resp && rcv_resp) {
1906                 /* is request/response. */
1907                 if (!has_grh) {
1908                         if (ib_get_cached_lmc(device, port_num, &lmc))
1909                                 return 0;
1910                         return (!lmc || !((rdma_ah_get_path_bits(&attr) ^
1911                                            rwc->wc->dlid_path_bits) &
1912                                           ((1 << lmc) - 1)));
1913                 } else {
1914                         const struct ib_global_route *grh =
1915                                         rdma_ah_read_grh(&attr);
1916
1917                         if (rdma_query_gid(device, port_num,
1918                                            grh->sgid_index, &sgid))
1919                                 return 0;
1920                         return !memcmp(sgid.raw, rwc->recv_buf.grh->dgid.raw,
1921                                        16);
1922                 }
1923         }
1924
1925         if (!has_grh)
1926                 return rdma_ah_get_dlid(&attr) == rwc->wc->slid;
1927         else
1928                 return !memcmp(rdma_ah_read_grh(&attr)->dgid.raw,
1929                                rwc->recv_buf.grh->sgid.raw,
1930                                16);
1931 }
1932
1933 static inline int is_direct(u8 class)
1934 {
1935         return (class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE);
1936 }
1937
1938 struct ib_mad_send_wr_private*
1939 ib_find_send_mad(const struct ib_mad_agent_private *mad_agent_priv,
1940                  const struct ib_mad_recv_wc *wc)
1941 {
1942         struct ib_mad_send_wr_private *wr;
1943         const struct ib_mad_hdr *mad_hdr;
1944
1945         mad_hdr = &wc->recv_buf.mad->mad_hdr;
1946
1947         list_for_each_entry(wr, &mad_agent_priv->wait_list, agent_list) {
1948                 if ((wr->tid == mad_hdr->tid) &&
1949                     rcv_has_same_class(wr, wc) &&
1950                     /*
1951                      * Don't check GID for direct routed MADs.
1952                      * These might have permissive LIDs.
1953                      */
1954                     (is_direct(mad_hdr->mgmt_class) ||
1955                      rcv_has_same_gid(mad_agent_priv, wr, wc)))
1956                         return (wr->status == IB_WC_SUCCESS) ? wr : NULL;
1957         }
1958
1959         /*
1960          * It's possible to receive the response before we've
1961          * been notified that the send has completed
1962          */
1963         list_for_each_entry(wr, &mad_agent_priv->send_list, agent_list) {
1964                 if (is_rmpp_data_mad(mad_agent_priv, wr->send_buf.mad) &&
1965                     wr->tid == mad_hdr->tid &&
1966                     wr->timeout &&
1967                     rcv_has_same_class(wr, wc) &&
1968                     /*
1969                      * Don't check GID for direct routed MADs.
1970                      * These might have permissive LIDs.
1971                      */
1972                     (is_direct(mad_hdr->mgmt_class) ||
1973                      rcv_has_same_gid(mad_agent_priv, wr, wc)))
1974                         /* Verify request has not been canceled */
1975                         return (wr->status == IB_WC_SUCCESS) ? wr : NULL;
1976         }
1977         return NULL;
1978 }
1979
1980 void ib_mark_mad_done(struct ib_mad_send_wr_private *mad_send_wr)
1981 {
1982         mad_send_wr->timeout = 0;
1983         if (mad_send_wr->refcount == 1)
1984                 list_move_tail(&mad_send_wr->agent_list,
1985                               &mad_send_wr->mad_agent_priv->done_list);
1986 }
1987
1988 static void ib_mad_complete_recv(struct ib_mad_agent_private *mad_agent_priv,
1989                                  struct ib_mad_recv_wc *mad_recv_wc)
1990 {
1991         struct ib_mad_send_wr_private *mad_send_wr;
1992         struct ib_mad_send_wc mad_send_wc;
1993         unsigned long flags;
1994         int ret;
1995
1996         INIT_LIST_HEAD(&mad_recv_wc->rmpp_list);
1997         ret = ib_mad_enforce_security(mad_agent_priv,
1998                                       mad_recv_wc->wc->pkey_index);
1999         if (ret) {
2000                 ib_free_recv_mad(mad_recv_wc);
2001                 deref_mad_agent(mad_agent_priv);
2002                 return;
2003         }
2004
2005         list_add(&mad_recv_wc->recv_buf.list, &mad_recv_wc->rmpp_list);
2006         if (ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)) {
2007                 mad_recv_wc = ib_process_rmpp_recv_wc(mad_agent_priv,
2008                                                       mad_recv_wc);
2009                 if (!mad_recv_wc) {
2010                         deref_mad_agent(mad_agent_priv);
2011                         return;
2012                 }
2013         }
2014
2015         /* Complete corresponding request */
2016         if (ib_response_mad(&mad_recv_wc->recv_buf.mad->mad_hdr)) {
2017                 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2018                 mad_send_wr = ib_find_send_mad(mad_agent_priv, mad_recv_wc);
2019                 if (!mad_send_wr) {
2020                         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2021                         if (!ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)
2022                            && ib_is_mad_class_rmpp(mad_recv_wc->recv_buf.mad->mad_hdr.mgmt_class)
2023                            && (ib_get_rmpp_flags(&((struct ib_rmpp_mad *)mad_recv_wc->recv_buf.mad)->rmpp_hdr)
2024                                         & IB_MGMT_RMPP_FLAG_ACTIVE)) {
2025                                 /* user rmpp is in effect
2026                                  * and this is an active RMPP MAD
2027                                  */
2028                                 mad_agent_priv->agent.recv_handler(
2029                                                 &mad_agent_priv->agent, NULL,
2030                                                 mad_recv_wc);
2031                                 atomic_dec(&mad_agent_priv->refcount);
2032                         } else {
2033                                 /* not user rmpp, revert to normal behavior and
2034                                  * drop the mad */
2035                                 ib_free_recv_mad(mad_recv_wc);
2036                                 deref_mad_agent(mad_agent_priv);
2037                                 return;
2038                         }
2039                 } else {
2040                         ib_mark_mad_done(mad_send_wr);
2041                         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2042
2043                         /* Defined behavior is to complete response before request */
2044                         mad_agent_priv->agent.recv_handler(
2045                                         &mad_agent_priv->agent,
2046                                         &mad_send_wr->send_buf,
2047                                         mad_recv_wc);
2048                         atomic_dec(&mad_agent_priv->refcount);
2049
2050                         mad_send_wc.status = IB_WC_SUCCESS;
2051                         mad_send_wc.vendor_err = 0;
2052                         mad_send_wc.send_buf = &mad_send_wr->send_buf;
2053                         ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
2054                 }
2055         } else {
2056                 mad_agent_priv->agent.recv_handler(&mad_agent_priv->agent, NULL,
2057                                                    mad_recv_wc);
2058                 deref_mad_agent(mad_agent_priv);
2059         }
2060
2061         return;
2062 }
2063
2064 static enum smi_action handle_ib_smi(const struct ib_mad_port_private *port_priv,
2065                                      const struct ib_mad_qp_info *qp_info,
2066                                      const struct ib_wc *wc,
2067                                      int port_num,
2068                                      struct ib_mad_private *recv,
2069                                      struct ib_mad_private *response)
2070 {
2071         enum smi_forward_action retsmi;
2072         struct ib_smp *smp = (struct ib_smp *)recv->mad;
2073
2074         if (smi_handle_dr_smp_recv(smp,
2075                                    rdma_cap_ib_switch(port_priv->device),
2076                                    port_num,
2077                                    port_priv->device->phys_port_cnt) ==
2078                                    IB_SMI_DISCARD)
2079                 return IB_SMI_DISCARD;
2080
2081         retsmi = smi_check_forward_dr_smp(smp);
2082         if (retsmi == IB_SMI_LOCAL)
2083                 return IB_SMI_HANDLE;
2084
2085         if (retsmi == IB_SMI_SEND) { /* don't forward */
2086                 if (smi_handle_dr_smp_send(smp,
2087                                            rdma_cap_ib_switch(port_priv->device),
2088                                            port_num) == IB_SMI_DISCARD)
2089                         return IB_SMI_DISCARD;
2090
2091                 if (smi_check_local_smp(smp, port_priv->device) == IB_SMI_DISCARD)
2092                         return IB_SMI_DISCARD;
2093         } else if (rdma_cap_ib_switch(port_priv->device)) {
2094                 /* forward case for switches */
2095                 memcpy(response, recv, mad_priv_size(response));
2096                 response->header.recv_wc.wc = &response->header.wc;
2097                 response->header.recv_wc.recv_buf.mad = (struct ib_mad *)response->mad;
2098                 response->header.recv_wc.recv_buf.grh = &response->grh;
2099
2100                 agent_send_response((const struct ib_mad_hdr *)response->mad,
2101                                     &response->grh, wc,
2102                                     port_priv->device,
2103                                     smi_get_fwd_port(smp),
2104                                     qp_info->qp->qp_num,
2105                                     response->mad_size,
2106                                     false);
2107
2108                 return IB_SMI_DISCARD;
2109         }
2110         return IB_SMI_HANDLE;
2111 }
2112
2113 static bool generate_unmatched_resp(const struct ib_mad_private *recv,
2114                                     struct ib_mad_private *response,
2115                                     size_t *resp_len, bool opa)
2116 {
2117         const struct ib_mad_hdr *recv_hdr = (const struct ib_mad_hdr *)recv->mad;
2118         struct ib_mad_hdr *resp_hdr = (struct ib_mad_hdr *)response->mad;
2119
2120         if (recv_hdr->method == IB_MGMT_METHOD_GET ||
2121             recv_hdr->method == IB_MGMT_METHOD_SET) {
2122                 memcpy(response, recv, mad_priv_size(response));
2123                 response->header.recv_wc.wc = &response->header.wc;
2124                 response->header.recv_wc.recv_buf.mad = (struct ib_mad *)response->mad;
2125                 response->header.recv_wc.recv_buf.grh = &response->grh;
2126                 resp_hdr->method = IB_MGMT_METHOD_GET_RESP;
2127                 resp_hdr->status = cpu_to_be16(IB_MGMT_MAD_STATUS_UNSUPPORTED_METHOD_ATTRIB);
2128                 if (recv_hdr->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
2129                         resp_hdr->status |= IB_SMP_DIRECTION;
2130
2131                 if (opa && recv_hdr->base_version == OPA_MGMT_BASE_VERSION) {
2132                         if (recv_hdr->mgmt_class ==
2133                             IB_MGMT_CLASS_SUBN_LID_ROUTED ||
2134                             recv_hdr->mgmt_class ==
2135                             IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
2136                                 *resp_len = opa_get_smp_header_size(
2137                                                         (struct opa_smp *)recv->mad);
2138                         else
2139                                 *resp_len = sizeof(struct ib_mad_hdr);
2140                 }
2141
2142                 return true;
2143         } else {
2144                 return false;
2145         }
2146 }
2147
2148 static enum smi_action
2149 handle_opa_smi(struct ib_mad_port_private *port_priv,
2150                struct ib_mad_qp_info *qp_info,
2151                struct ib_wc *wc,
2152                int port_num,
2153                struct ib_mad_private *recv,
2154                struct ib_mad_private *response)
2155 {
2156         enum smi_forward_action retsmi;
2157         struct opa_smp *smp = (struct opa_smp *)recv->mad;
2158
2159         if (opa_smi_handle_dr_smp_recv(smp,
2160                                    rdma_cap_ib_switch(port_priv->device),
2161                                    port_num,
2162                                    port_priv->device->phys_port_cnt) ==
2163                                    IB_SMI_DISCARD)
2164                 return IB_SMI_DISCARD;
2165
2166         retsmi = opa_smi_check_forward_dr_smp(smp);
2167         if (retsmi == IB_SMI_LOCAL)
2168                 return IB_SMI_HANDLE;
2169
2170         if (retsmi == IB_SMI_SEND) { /* don't forward */
2171                 if (opa_smi_handle_dr_smp_send(smp,
2172                                            rdma_cap_ib_switch(port_priv->device),
2173                                            port_num) == IB_SMI_DISCARD)
2174                         return IB_SMI_DISCARD;
2175
2176                 if (opa_smi_check_local_smp(smp, port_priv->device) ==
2177                     IB_SMI_DISCARD)
2178                         return IB_SMI_DISCARD;
2179
2180         } else if (rdma_cap_ib_switch(port_priv->device)) {
2181                 /* forward case for switches */
2182                 memcpy(response, recv, mad_priv_size(response));
2183                 response->header.recv_wc.wc = &response->header.wc;
2184                 response->header.recv_wc.recv_buf.opa_mad =
2185                                 (struct opa_mad *)response->mad;
2186                 response->header.recv_wc.recv_buf.grh = &response->grh;
2187
2188                 agent_send_response((const struct ib_mad_hdr *)response->mad,
2189                                     &response->grh, wc,
2190                                     port_priv->device,
2191                                     opa_smi_get_fwd_port(smp),
2192                                     qp_info->qp->qp_num,
2193                                     recv->header.wc.byte_len,
2194                                     true);
2195
2196                 return IB_SMI_DISCARD;
2197         }
2198
2199         return IB_SMI_HANDLE;
2200 }
2201
2202 static enum smi_action
2203 handle_smi(struct ib_mad_port_private *port_priv,
2204            struct ib_mad_qp_info *qp_info,
2205            struct ib_wc *wc,
2206            int port_num,
2207            struct ib_mad_private *recv,
2208            struct ib_mad_private *response,
2209            bool opa)
2210 {
2211         struct ib_mad_hdr *mad_hdr = (struct ib_mad_hdr *)recv->mad;
2212
2213         if (opa && mad_hdr->base_version == OPA_MGMT_BASE_VERSION &&
2214             mad_hdr->class_version == OPA_SM_CLASS_VERSION)
2215                 return handle_opa_smi(port_priv, qp_info, wc, port_num, recv,
2216                                       response);
2217
2218         return handle_ib_smi(port_priv, qp_info, wc, port_num, recv, response);
2219 }
2220
2221 static void ib_mad_recv_done(struct ib_cq *cq, struct ib_wc *wc)
2222 {
2223         struct ib_mad_port_private *port_priv = cq->cq_context;
2224         struct ib_mad_list_head *mad_list =
2225                 container_of(wc->wr_cqe, struct ib_mad_list_head, cqe);
2226         struct ib_mad_qp_info *qp_info;
2227         struct ib_mad_private_header *mad_priv_hdr;
2228         struct ib_mad_private *recv, *response = NULL;
2229         struct ib_mad_agent_private *mad_agent;
2230         int port_num;
2231         int ret = IB_MAD_RESULT_SUCCESS;
2232         size_t mad_size;
2233         u16 resp_mad_pkey_index = 0;
2234         bool opa;
2235
2236         if (list_empty_careful(&port_priv->port_list))
2237                 return;
2238
2239         if (wc->status != IB_WC_SUCCESS) {
2240                 /*
2241                  * Receive errors indicate that the QP has entered the error
2242                  * state - error handling/shutdown code will cleanup
2243                  */
2244                 return;
2245         }
2246
2247         qp_info = mad_list->mad_queue->qp_info;
2248         dequeue_mad(mad_list);
2249
2250         opa = rdma_cap_opa_mad(qp_info->port_priv->device,
2251                                qp_info->port_priv->port_num);
2252
2253         mad_priv_hdr = container_of(mad_list, struct ib_mad_private_header,
2254                                     mad_list);
2255         recv = container_of(mad_priv_hdr, struct ib_mad_private, header);
2256         ib_dma_unmap_single(port_priv->device,
2257                             recv->header.mapping,
2258                             mad_priv_dma_size(recv),
2259                             DMA_FROM_DEVICE);
2260
2261         /* Setup MAD receive work completion from "normal" work completion */
2262         recv->header.wc = *wc;
2263         recv->header.recv_wc.wc = &recv->header.wc;
2264
2265         if (opa && ((struct ib_mad_hdr *)(recv->mad))->base_version == OPA_MGMT_BASE_VERSION) {
2266                 recv->header.recv_wc.mad_len = wc->byte_len - sizeof(struct ib_grh);
2267                 recv->header.recv_wc.mad_seg_size = sizeof(struct opa_mad);
2268         } else {
2269                 recv->header.recv_wc.mad_len = sizeof(struct ib_mad);
2270                 recv->header.recv_wc.mad_seg_size = sizeof(struct ib_mad);
2271         }
2272
2273         recv->header.recv_wc.recv_buf.mad = (struct ib_mad *)recv->mad;
2274         recv->header.recv_wc.recv_buf.grh = &recv->grh;
2275
2276         if (atomic_read(&qp_info->snoop_count))
2277                 snoop_recv(qp_info, &recv->header.recv_wc, IB_MAD_SNOOP_RECVS);
2278
2279         /* Validate MAD */
2280         if (!validate_mad((const struct ib_mad_hdr *)recv->mad, qp_info, opa))
2281                 goto out;
2282
2283         mad_size = recv->mad_size;
2284         response = alloc_mad_private(mad_size, GFP_KERNEL);
2285         if (!response)
2286                 goto out;
2287
2288         if (rdma_cap_ib_switch(port_priv->device))
2289                 port_num = wc->port_num;
2290         else
2291                 port_num = port_priv->port_num;
2292
2293         if (((struct ib_mad_hdr *)recv->mad)->mgmt_class ==
2294             IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
2295                 if (handle_smi(port_priv, qp_info, wc, port_num, recv,
2296                                response, opa)
2297                     == IB_SMI_DISCARD)
2298                         goto out;
2299         }
2300
2301         /* Give driver "right of first refusal" on incoming MAD */
2302         if (port_priv->device->process_mad) {
2303                 ret = port_priv->device->process_mad(port_priv->device, 0,
2304                                                      port_priv->port_num,
2305                                                      wc, &recv->grh,
2306                                                      (const struct ib_mad_hdr *)recv->mad,
2307                                                      recv->mad_size,
2308                                                      (struct ib_mad_hdr *)response->mad,
2309                                                      &mad_size, &resp_mad_pkey_index);
2310
2311                 if (opa)
2312                         wc->pkey_index = resp_mad_pkey_index;
2313
2314                 if (ret & IB_MAD_RESULT_SUCCESS) {
2315                         if (ret & IB_MAD_RESULT_CONSUMED)
2316                                 goto out;
2317                         if (ret & IB_MAD_RESULT_REPLY) {
2318                                 agent_send_response((const struct ib_mad_hdr *)response->mad,
2319                                                     &recv->grh, wc,
2320                                                     port_priv->device,
2321                                                     port_num,
2322                                                     qp_info->qp->qp_num,
2323                                                     mad_size, opa);
2324                                 goto out;
2325                         }
2326                 }
2327         }
2328
2329         mad_agent = find_mad_agent(port_priv, (const struct ib_mad_hdr *)recv->mad);
2330         if (mad_agent) {
2331                 ib_mad_complete_recv(mad_agent, &recv->header.recv_wc);
2332                 /*
2333                  * recv is freed up in error cases in ib_mad_complete_recv
2334                  * or via recv_handler in ib_mad_complete_recv()
2335                  */
2336                 recv = NULL;
2337         } else if ((ret & IB_MAD_RESULT_SUCCESS) &&
2338                    generate_unmatched_resp(recv, response, &mad_size, opa)) {
2339                 agent_send_response((const struct ib_mad_hdr *)response->mad, &recv->grh, wc,
2340                                     port_priv->device, port_num,
2341                                     qp_info->qp->qp_num, mad_size, opa);
2342         }
2343
2344 out:
2345         /* Post another receive request for this QP */
2346         if (response) {
2347                 ib_mad_post_receive_mads(qp_info, response);
2348                 kfree(recv);
2349         } else
2350                 ib_mad_post_receive_mads(qp_info, recv);
2351 }
2352
2353 static void adjust_timeout(struct ib_mad_agent_private *mad_agent_priv)
2354 {
2355         struct ib_mad_send_wr_private *mad_send_wr;
2356         unsigned long delay;
2357
2358         if (list_empty(&mad_agent_priv->wait_list)) {
2359                 cancel_delayed_work(&mad_agent_priv->timed_work);
2360         } else {
2361                 mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
2362                                          struct ib_mad_send_wr_private,
2363                                          agent_list);
2364
2365                 if (time_after(mad_agent_priv->timeout,
2366                                mad_send_wr->timeout)) {
2367                         mad_agent_priv->timeout = mad_send_wr->timeout;
2368                         delay = mad_send_wr->timeout - jiffies;
2369                         if ((long)delay <= 0)
2370                                 delay = 1;
2371                         mod_delayed_work(mad_agent_priv->qp_info->port_priv->wq,
2372                                          &mad_agent_priv->timed_work, delay);
2373                 }
2374         }
2375 }
2376
2377 static void wait_for_response(struct ib_mad_send_wr_private *mad_send_wr)
2378 {
2379         struct ib_mad_agent_private *mad_agent_priv;
2380         struct ib_mad_send_wr_private *temp_mad_send_wr;
2381         struct list_head *list_item;
2382         unsigned long delay;
2383
2384         mad_agent_priv = mad_send_wr->mad_agent_priv;
2385         list_del(&mad_send_wr->agent_list);
2386
2387         delay = mad_send_wr->timeout;
2388         mad_send_wr->timeout += jiffies;
2389
2390         if (delay) {
2391                 list_for_each_prev(list_item, &mad_agent_priv->wait_list) {
2392                         temp_mad_send_wr = list_entry(list_item,
2393                                                 struct ib_mad_send_wr_private,
2394                                                 agent_list);
2395                         if (time_after(mad_send_wr->timeout,
2396                                        temp_mad_send_wr->timeout))
2397                                 break;
2398                 }
2399         }
2400         else
2401                 list_item = &mad_agent_priv->wait_list;
2402         list_add(&mad_send_wr->agent_list, list_item);
2403
2404         /* Reschedule a work item if we have a shorter timeout */
2405         if (mad_agent_priv->wait_list.next == &mad_send_wr->agent_list)
2406                 mod_delayed_work(mad_agent_priv->qp_info->port_priv->wq,
2407                                  &mad_agent_priv->timed_work, delay);
2408 }
2409
2410 void ib_reset_mad_timeout(struct ib_mad_send_wr_private *mad_send_wr,
2411                           int timeout_ms)
2412 {
2413         mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2414         wait_for_response(mad_send_wr);
2415 }
2416
2417 /*
2418  * Process a send work completion
2419  */
2420 void ib_mad_complete_send_wr(struct ib_mad_send_wr_private *mad_send_wr,
2421                              struct ib_mad_send_wc *mad_send_wc)
2422 {
2423         struct ib_mad_agent_private     *mad_agent_priv;
2424         unsigned long                   flags;
2425         int                             ret;
2426
2427         mad_agent_priv = mad_send_wr->mad_agent_priv;
2428         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2429         if (ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)) {
2430                 ret = ib_process_rmpp_send_wc(mad_send_wr, mad_send_wc);
2431                 if (ret == IB_RMPP_RESULT_CONSUMED)
2432                         goto done;
2433         } else
2434                 ret = IB_RMPP_RESULT_UNHANDLED;
2435
2436         if (mad_send_wc->status != IB_WC_SUCCESS &&
2437             mad_send_wr->status == IB_WC_SUCCESS) {
2438                 mad_send_wr->status = mad_send_wc->status;
2439                 mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2440         }
2441
2442         if (--mad_send_wr->refcount > 0) {
2443                 if (mad_send_wr->refcount == 1 && mad_send_wr->timeout &&
2444                     mad_send_wr->status == IB_WC_SUCCESS) {
2445                         wait_for_response(mad_send_wr);
2446                 }
2447                 goto done;
2448         }
2449
2450         /* Remove send from MAD agent and notify client of completion */
2451         list_del(&mad_send_wr->agent_list);
2452         adjust_timeout(mad_agent_priv);
2453         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2454
2455         if (mad_send_wr->status != IB_WC_SUCCESS )
2456                 mad_send_wc->status = mad_send_wr->status;
2457         if (ret == IB_RMPP_RESULT_INTERNAL)
2458                 ib_rmpp_send_handler(mad_send_wc);
2459         else
2460                 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2461                                                    mad_send_wc);
2462
2463         /* Release reference on agent taken when sending */
2464         deref_mad_agent(mad_agent_priv);
2465         return;
2466 done:
2467         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2468 }
2469
2470 static void ib_mad_send_done(struct ib_cq *cq, struct ib_wc *wc)
2471 {
2472         struct ib_mad_port_private *port_priv = cq->cq_context;
2473         struct ib_mad_list_head *mad_list =
2474                 container_of(wc->wr_cqe, struct ib_mad_list_head, cqe);
2475         struct ib_mad_send_wr_private   *mad_send_wr, *queued_send_wr;
2476         struct ib_mad_qp_info           *qp_info;
2477         struct ib_mad_queue             *send_queue;
2478         struct ib_mad_send_wc           mad_send_wc;
2479         unsigned long flags;
2480         int ret;
2481
2482         if (list_empty_careful(&port_priv->port_list))
2483                 return;
2484
2485         if (wc->status != IB_WC_SUCCESS) {
2486                 if (!ib_mad_send_error(port_priv, wc))
2487                         return;
2488         }
2489
2490         mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2491                                    mad_list);
2492         send_queue = mad_list->mad_queue;
2493         qp_info = send_queue->qp_info;
2494
2495 retry:
2496         ib_dma_unmap_single(mad_send_wr->send_buf.mad_agent->device,
2497                             mad_send_wr->header_mapping,
2498                             mad_send_wr->sg_list[0].length, DMA_TO_DEVICE);
2499         ib_dma_unmap_single(mad_send_wr->send_buf.mad_agent->device,
2500                             mad_send_wr->payload_mapping,
2501                             mad_send_wr->sg_list[1].length, DMA_TO_DEVICE);
2502         queued_send_wr = NULL;
2503         spin_lock_irqsave(&send_queue->lock, flags);
2504         list_del(&mad_list->list);
2505
2506         /* Move queued send to the send queue */
2507         if (send_queue->count-- > send_queue->max_active) {
2508                 mad_list = container_of(qp_info->overflow_list.next,
2509                                         struct ib_mad_list_head, list);
2510                 queued_send_wr = container_of(mad_list,
2511                                         struct ib_mad_send_wr_private,
2512                                         mad_list);
2513                 list_move_tail(&mad_list->list, &send_queue->list);
2514         }
2515         spin_unlock_irqrestore(&send_queue->lock, flags);
2516
2517         mad_send_wc.send_buf = &mad_send_wr->send_buf;
2518         mad_send_wc.status = wc->status;
2519         mad_send_wc.vendor_err = wc->vendor_err;
2520         if (atomic_read(&qp_info->snoop_count))
2521                 snoop_send(qp_info, &mad_send_wr->send_buf, &mad_send_wc,
2522                            IB_MAD_SNOOP_SEND_COMPLETIONS);
2523         ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
2524
2525         if (queued_send_wr) {
2526                 ret = ib_post_send(qp_info->qp, &queued_send_wr->send_wr.wr,
2527                                    NULL);
2528                 if (ret) {
2529                         dev_err(&port_priv->device->dev,
2530                                 "ib_post_send failed: %d\n", ret);
2531                         mad_send_wr = queued_send_wr;
2532                         wc->status = IB_WC_LOC_QP_OP_ERR;
2533                         goto retry;
2534                 }
2535         }
2536 }
2537
2538 static void mark_sends_for_retry(struct ib_mad_qp_info *qp_info)
2539 {
2540         struct ib_mad_send_wr_private *mad_send_wr;
2541         struct ib_mad_list_head *mad_list;
2542         unsigned long flags;
2543
2544         spin_lock_irqsave(&qp_info->send_queue.lock, flags);
2545         list_for_each_entry(mad_list, &qp_info->send_queue.list, list) {
2546                 mad_send_wr = container_of(mad_list,
2547                                            struct ib_mad_send_wr_private,
2548                                            mad_list);
2549                 mad_send_wr->retry = 1;
2550         }
2551         spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
2552 }
2553
2554 static bool ib_mad_send_error(struct ib_mad_port_private *port_priv,
2555                 struct ib_wc *wc)
2556 {
2557         struct ib_mad_list_head *mad_list =
2558                 container_of(wc->wr_cqe, struct ib_mad_list_head, cqe);
2559         struct ib_mad_qp_info *qp_info = mad_list->mad_queue->qp_info;
2560         struct ib_mad_send_wr_private *mad_send_wr;
2561         int ret;
2562
2563         /*
2564          * Send errors will transition the QP to SQE - move
2565          * QP to RTS and repost flushed work requests
2566          */
2567         mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2568                                    mad_list);
2569         if (wc->status == IB_WC_WR_FLUSH_ERR) {
2570                 if (mad_send_wr->retry) {
2571                         /* Repost send */
2572                         mad_send_wr->retry = 0;
2573                         ret = ib_post_send(qp_info->qp, &mad_send_wr->send_wr.wr,
2574                                            NULL);
2575                         if (!ret)
2576                                 return false;
2577                 }
2578         } else {
2579                 struct ib_qp_attr *attr;
2580
2581                 /* Transition QP to RTS and fail offending send */
2582                 attr = kmalloc(sizeof *attr, GFP_KERNEL);
2583                 if (attr) {
2584                         attr->qp_state = IB_QPS_RTS;
2585                         attr->cur_qp_state = IB_QPS_SQE;
2586                         ret = ib_modify_qp(qp_info->qp, attr,
2587                                            IB_QP_STATE | IB_QP_CUR_STATE);
2588                         kfree(attr);
2589                         if (ret)
2590                                 dev_err(&port_priv->device->dev,
2591                                         "%s - ib_modify_qp to RTS: %d\n",
2592                                         __func__, ret);
2593                         else
2594                                 mark_sends_for_retry(qp_info);
2595                 }
2596         }
2597
2598         return true;
2599 }
2600
2601 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv)
2602 {
2603         unsigned long flags;
2604         struct ib_mad_send_wr_private *mad_send_wr, *temp_mad_send_wr;
2605         struct ib_mad_send_wc mad_send_wc;
2606         struct list_head cancel_list;
2607
2608         INIT_LIST_HEAD(&cancel_list);
2609
2610         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2611         list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2612                                  &mad_agent_priv->send_list, agent_list) {
2613                 if (mad_send_wr->status == IB_WC_SUCCESS) {
2614                         mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2615                         mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2616                 }
2617         }
2618
2619         /* Empty wait list to prevent receives from finding a request */
2620         list_splice_init(&mad_agent_priv->wait_list, &cancel_list);
2621         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2622
2623         /* Report all cancelled requests */
2624         mad_send_wc.status = IB_WC_WR_FLUSH_ERR;
2625         mad_send_wc.vendor_err = 0;
2626
2627         list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2628                                  &cancel_list, agent_list) {
2629                 mad_send_wc.send_buf = &mad_send_wr->send_buf;
2630                 list_del(&mad_send_wr->agent_list);
2631                 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2632                                                    &mad_send_wc);
2633                 atomic_dec(&mad_agent_priv->refcount);
2634         }
2635 }
2636
2637 static struct ib_mad_send_wr_private*
2638 find_send_wr(struct ib_mad_agent_private *mad_agent_priv,
2639              struct ib_mad_send_buf *send_buf)
2640 {
2641         struct ib_mad_send_wr_private *mad_send_wr;
2642
2643         list_for_each_entry(mad_send_wr, &mad_agent_priv->wait_list,
2644                             agent_list) {
2645                 if (&mad_send_wr->send_buf == send_buf)
2646                         return mad_send_wr;
2647         }
2648
2649         list_for_each_entry(mad_send_wr, &mad_agent_priv->send_list,
2650                             agent_list) {
2651                 if (is_rmpp_data_mad(mad_agent_priv,
2652                                      mad_send_wr->send_buf.mad) &&
2653                     &mad_send_wr->send_buf == send_buf)
2654                         return mad_send_wr;
2655         }
2656         return NULL;
2657 }
2658
2659 int ib_modify_mad(struct ib_mad_agent *mad_agent,
2660                   struct ib_mad_send_buf *send_buf, u32 timeout_ms)
2661 {
2662         struct ib_mad_agent_private *mad_agent_priv;
2663         struct ib_mad_send_wr_private *mad_send_wr;
2664         unsigned long flags;
2665         int active;
2666
2667         mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
2668                                       agent);
2669         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2670         mad_send_wr = find_send_wr(mad_agent_priv, send_buf);
2671         if (!mad_send_wr || mad_send_wr->status != IB_WC_SUCCESS) {
2672                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2673                 return -EINVAL;
2674         }
2675
2676         active = (!mad_send_wr->timeout || mad_send_wr->refcount > 1);
2677         if (!timeout_ms) {
2678                 mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2679                 mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2680         }
2681
2682         mad_send_wr->send_buf.timeout_ms = timeout_ms;
2683         if (active)
2684                 mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2685         else
2686                 ib_reset_mad_timeout(mad_send_wr, timeout_ms);
2687
2688         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2689         return 0;
2690 }
2691 EXPORT_SYMBOL(ib_modify_mad);
2692
2693 void ib_cancel_mad(struct ib_mad_agent *mad_agent,
2694                    struct ib_mad_send_buf *send_buf)
2695 {
2696         ib_modify_mad(mad_agent, send_buf, 0);
2697 }
2698 EXPORT_SYMBOL(ib_cancel_mad);
2699
2700 static void local_completions(struct work_struct *work)
2701 {
2702         struct ib_mad_agent_private *mad_agent_priv;
2703         struct ib_mad_local_private *local;
2704         struct ib_mad_agent_private *recv_mad_agent;
2705         unsigned long flags;
2706         int free_mad;
2707         struct ib_wc wc;
2708         struct ib_mad_send_wc mad_send_wc;
2709         bool opa;
2710
2711         mad_agent_priv =
2712                 container_of(work, struct ib_mad_agent_private, local_work);
2713
2714         opa = rdma_cap_opa_mad(mad_agent_priv->qp_info->port_priv->device,
2715                                mad_agent_priv->qp_info->port_priv->port_num);
2716
2717         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2718         while (!list_empty(&mad_agent_priv->local_list)) {
2719                 local = list_entry(mad_agent_priv->local_list.next,
2720                                    struct ib_mad_local_private,
2721                                    completion_list);
2722                 list_del(&local->completion_list);
2723                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2724                 free_mad = 0;
2725                 if (local->mad_priv) {
2726                         u8 base_version;
2727                         recv_mad_agent = local->recv_mad_agent;
2728                         if (!recv_mad_agent) {
2729                                 dev_err(&mad_agent_priv->agent.device->dev,
2730                                         "No receive MAD agent for local completion\n");
2731                                 free_mad = 1;
2732                                 goto local_send_completion;
2733                         }
2734
2735                         /*
2736                          * Defined behavior is to complete response
2737                          * before request
2738                          */
2739                         build_smp_wc(recv_mad_agent->agent.qp,
2740                                      local->mad_send_wr->send_wr.wr.wr_cqe,
2741                                      be16_to_cpu(IB_LID_PERMISSIVE),
2742                                      local->mad_send_wr->send_wr.pkey_index,
2743                                      recv_mad_agent->agent.port_num, &wc);
2744
2745                         local->mad_priv->header.recv_wc.wc = &wc;
2746
2747                         base_version = ((struct ib_mad_hdr *)(local->mad_priv->mad))->base_version;
2748                         if (opa && base_version == OPA_MGMT_BASE_VERSION) {
2749                                 local->mad_priv->header.recv_wc.mad_len = local->return_wc_byte_len;
2750                                 local->mad_priv->header.recv_wc.mad_seg_size = sizeof(struct opa_mad);
2751                         } else {
2752                                 local->mad_priv->header.recv_wc.mad_len = sizeof(struct ib_mad);
2753                                 local->mad_priv->header.recv_wc.mad_seg_size = sizeof(struct ib_mad);
2754                         }
2755
2756                         INIT_LIST_HEAD(&local->mad_priv->header.recv_wc.rmpp_list);
2757                         list_add(&local->mad_priv->header.recv_wc.recv_buf.list,
2758                                  &local->mad_priv->header.recv_wc.rmpp_list);
2759                         local->mad_priv->header.recv_wc.recv_buf.grh = NULL;
2760                         local->mad_priv->header.recv_wc.recv_buf.mad =
2761                                                 (struct ib_mad *)local->mad_priv->mad;
2762                         if (atomic_read(&recv_mad_agent->qp_info->snoop_count))
2763                                 snoop_recv(recv_mad_agent->qp_info,
2764                                           &local->mad_priv->header.recv_wc,
2765                                            IB_MAD_SNOOP_RECVS);
2766                         recv_mad_agent->agent.recv_handler(
2767                                                 &recv_mad_agent->agent,
2768                                                 &local->mad_send_wr->send_buf,
2769                                                 &local->mad_priv->header.recv_wc);
2770                         spin_lock_irqsave(&recv_mad_agent->lock, flags);
2771                         atomic_dec(&recv_mad_agent->refcount);
2772                         spin_unlock_irqrestore(&recv_mad_agent->lock, flags);
2773                 }
2774
2775 local_send_completion:
2776                 /* Complete send */
2777                 mad_send_wc.status = IB_WC_SUCCESS;
2778                 mad_send_wc.vendor_err = 0;
2779                 mad_send_wc.send_buf = &local->mad_send_wr->send_buf;
2780                 if (atomic_read(&mad_agent_priv->qp_info->snoop_count))
2781                         snoop_send(mad_agent_priv->qp_info,
2782                                    &local->mad_send_wr->send_buf,
2783                                    &mad_send_wc, IB_MAD_SNOOP_SEND_COMPLETIONS);
2784                 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2785                                                    &mad_send_wc);
2786
2787                 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2788                 atomic_dec(&mad_agent_priv->refcount);
2789                 if (free_mad)
2790                         kfree(local->mad_priv);
2791                 kfree(local);
2792         }
2793         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2794 }
2795
2796 static int retry_send(struct ib_mad_send_wr_private *mad_send_wr)
2797 {
2798         int ret;
2799
2800         if (!mad_send_wr->retries_left)
2801                 return -ETIMEDOUT;
2802
2803         mad_send_wr->retries_left--;
2804         mad_send_wr->send_buf.retries++;
2805
2806         mad_send_wr->timeout = msecs_to_jiffies(mad_send_wr->send_buf.timeout_ms);
2807
2808         if (ib_mad_kernel_rmpp_agent(&mad_send_wr->mad_agent_priv->agent)) {
2809                 ret = ib_retry_rmpp(mad_send_wr);
2810                 switch (ret) {
2811                 case IB_RMPP_RESULT_UNHANDLED:
2812                         ret = ib_send_mad(mad_send_wr);
2813                         break;
2814                 case IB_RMPP_RESULT_CONSUMED:
2815                         ret = 0;
2816                         break;
2817                 default:
2818                         ret = -ECOMM;
2819                         break;
2820                 }
2821         } else
2822                 ret = ib_send_mad(mad_send_wr);
2823
2824         if (!ret) {
2825                 mad_send_wr->refcount++;
2826                 list_add_tail(&mad_send_wr->agent_list,
2827                               &mad_send_wr->mad_agent_priv->send_list);
2828         }
2829         return ret;
2830 }
2831
2832 static void timeout_sends(struct work_struct *work)
2833 {
2834         struct ib_mad_agent_private *mad_agent_priv;
2835         struct ib_mad_send_wr_private *mad_send_wr;
2836         struct ib_mad_send_wc mad_send_wc;
2837         unsigned long flags, delay;
2838
2839         mad_agent_priv = container_of(work, struct ib_mad_agent_private,
2840                                       timed_work.work);
2841         mad_send_wc.vendor_err = 0;
2842
2843         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2844         while (!list_empty(&mad_agent_priv->wait_list)) {
2845                 mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
2846                                          struct ib_mad_send_wr_private,
2847                                          agent_list);
2848
2849                 if (time_after(mad_send_wr->timeout, jiffies)) {
2850                         delay = mad_send_wr->timeout - jiffies;
2851                         if ((long)delay <= 0)
2852                                 delay = 1;
2853                         queue_delayed_work(mad_agent_priv->qp_info->
2854                                            port_priv->wq,
2855                                            &mad_agent_priv->timed_work, delay);
2856                         break;
2857                 }
2858
2859                 list_del(&mad_send_wr->agent_list);
2860                 if (mad_send_wr->status == IB_WC_SUCCESS &&
2861                     !retry_send(mad_send_wr))
2862                         continue;
2863
2864                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2865
2866                 if (mad_send_wr->status == IB_WC_SUCCESS)
2867                         mad_send_wc.status = IB_WC_RESP_TIMEOUT_ERR;
2868                 else
2869                         mad_send_wc.status = mad_send_wr->status;
2870                 mad_send_wc.send_buf = &mad_send_wr->send_buf;
2871                 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2872                                                    &mad_send_wc);
2873
2874                 atomic_dec(&mad_agent_priv->refcount);
2875                 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2876         }
2877         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2878 }
2879
2880 /*
2881  * Allocate receive MADs and post receive WRs for them
2882  */
2883 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
2884                                     struct ib_mad_private *mad)
2885 {
2886         unsigned long flags;
2887         int post, ret;
2888         struct ib_mad_private *mad_priv;
2889         struct ib_sge sg_list;
2890         struct ib_recv_wr recv_wr;
2891         struct ib_mad_queue *recv_queue = &qp_info->recv_queue;
2892
2893         /* Initialize common scatter list fields */
2894         sg_list.lkey = qp_info->port_priv->pd->local_dma_lkey;
2895
2896         /* Initialize common receive WR fields */
2897         recv_wr.next = NULL;
2898         recv_wr.sg_list = &sg_list;
2899         recv_wr.num_sge = 1;
2900
2901         do {
2902                 /* Allocate and map receive buffer */
2903                 if (mad) {
2904                         mad_priv = mad;
2905                         mad = NULL;
2906                 } else {
2907                         mad_priv = alloc_mad_private(port_mad_size(qp_info->port_priv),
2908                                                      GFP_ATOMIC);
2909                         if (!mad_priv) {
2910                                 ret = -ENOMEM;
2911                                 break;
2912                         }
2913                 }
2914                 sg_list.length = mad_priv_dma_size(mad_priv);
2915                 sg_list.addr = ib_dma_map_single(qp_info->port_priv->device,
2916                                                  &mad_priv->grh,
2917                                                  mad_priv_dma_size(mad_priv),
2918                                                  DMA_FROM_DEVICE);
2919                 if (unlikely(ib_dma_mapping_error(qp_info->port_priv->device,
2920                                                   sg_list.addr))) {
2921                         ret = -ENOMEM;
2922                         break;
2923                 }
2924                 mad_priv->header.mapping = sg_list.addr;
2925                 mad_priv->header.mad_list.mad_queue = recv_queue;
2926                 mad_priv->header.mad_list.cqe.done = ib_mad_recv_done;
2927                 recv_wr.wr_cqe = &mad_priv->header.mad_list.cqe;
2928
2929                 /* Post receive WR */
2930                 spin_lock_irqsave(&recv_queue->lock, flags);
2931                 post = (++recv_queue->count < recv_queue->max_active);
2932                 list_add_tail(&mad_priv->header.mad_list.list, &recv_queue->list);
2933                 spin_unlock_irqrestore(&recv_queue->lock, flags);
2934                 ret = ib_post_recv(qp_info->qp, &recv_wr, NULL);
2935                 if (ret) {
2936                         spin_lock_irqsave(&recv_queue->lock, flags);
2937                         list_del(&mad_priv->header.mad_list.list);
2938                         recv_queue->count--;
2939                         spin_unlock_irqrestore(&recv_queue->lock, flags);
2940                         ib_dma_unmap_single(qp_info->port_priv->device,
2941                                             mad_priv->header.mapping,
2942                                             mad_priv_dma_size(mad_priv),
2943                                             DMA_FROM_DEVICE);
2944                         kfree(mad_priv);
2945                         dev_err(&qp_info->port_priv->device->dev,
2946                                 "ib_post_recv failed: %d\n", ret);
2947                         break;
2948                 }
2949         } while (post);
2950
2951         return ret;
2952 }
2953
2954 /*
2955  * Return all the posted receive MADs
2956  */
2957 static void cleanup_recv_queue(struct ib_mad_qp_info *qp_info)
2958 {
2959         struct ib_mad_private_header *mad_priv_hdr;
2960         struct ib_mad_private *recv;
2961         struct ib_mad_list_head *mad_list;
2962
2963         if (!qp_info->qp)
2964                 return;
2965
2966         while (!list_empty(&qp_info->recv_queue.list)) {
2967
2968                 mad_list = list_entry(qp_info->recv_queue.list.next,
2969                                       struct ib_mad_list_head, list);
2970                 mad_priv_hdr = container_of(mad_list,
2971                                             struct ib_mad_private_header,
2972                                             mad_list);
2973                 recv = container_of(mad_priv_hdr, struct ib_mad_private,
2974                                     header);
2975
2976                 /* Remove from posted receive MAD list */
2977                 list_del(&mad_list->list);
2978
2979                 ib_dma_unmap_single(qp_info->port_priv->device,
2980                                     recv->header.mapping,
2981                                     mad_priv_dma_size(recv),
2982                                     DMA_FROM_DEVICE);
2983                 kfree(recv);
2984         }
2985
2986         qp_info->recv_queue.count = 0;
2987 }
2988
2989 /*
2990  * Start the port
2991  */
2992 static int ib_mad_port_start(struct ib_mad_port_private *port_priv)
2993 {
2994         int ret, i;
2995         struct ib_qp_attr *attr;
2996         struct ib_qp *qp;
2997         u16 pkey_index;
2998
2999         attr = kmalloc(sizeof *attr, GFP_KERNEL);
3000         if (!attr)
3001                 return -ENOMEM;
3002
3003         ret = ib_find_pkey(port_priv->device, port_priv->port_num,
3004                            IB_DEFAULT_PKEY_FULL, &pkey_index);
3005         if (ret)
3006                 pkey_index = 0;
3007
3008         for (i = 0; i < IB_MAD_QPS_CORE; i++) {
3009                 qp = port_priv->qp_info[i].qp;
3010                 if (!qp)
3011                         continue;
3012
3013                 /*
3014                  * PKey index for QP1 is irrelevant but
3015                  * one is needed for the Reset to Init transition
3016                  */
3017                 attr->qp_state = IB_QPS_INIT;
3018                 attr->pkey_index = pkey_index;
3019                 attr->qkey = (qp->qp_num == 0) ? 0 : IB_QP1_QKEY;
3020                 ret = ib_modify_qp(qp, attr, IB_QP_STATE |
3021                                              IB_QP_PKEY_INDEX | IB_QP_QKEY);
3022                 if (ret) {
3023                         dev_err(&port_priv->device->dev,
3024                                 "Couldn't change QP%d state to INIT: %d\n",
3025                                 i, ret);
3026                         goto out;
3027                 }
3028
3029                 attr->qp_state = IB_QPS_RTR;
3030                 ret = ib_modify_qp(qp, attr, IB_QP_STATE);
3031                 if (ret) {
3032                         dev_err(&port_priv->device->dev,
3033                                 "Couldn't change QP%d state to RTR: %d\n",
3034                                 i, ret);
3035                         goto out;
3036                 }
3037
3038                 attr->qp_state = IB_QPS_RTS;
3039                 attr->sq_psn = IB_MAD_SEND_Q_PSN;
3040                 ret = ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_SQ_PSN);
3041                 if (ret) {
3042                         dev_err(&port_priv->device->dev,
3043                                 "Couldn't change QP%d state to RTS: %d\n",
3044                                 i, ret);
3045                         goto out;
3046                 }
3047         }
3048
3049         ret = ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP);
3050         if (ret) {
3051                 dev_err(&port_priv->device->dev,
3052                         "Failed to request completion notification: %d\n",
3053                         ret);
3054                 goto out;
3055         }
3056
3057         for (i = 0; i < IB_MAD_QPS_CORE; i++) {
3058                 if (!port_priv->qp_info[i].qp)
3059                         continue;
3060
3061                 ret = ib_mad_post_receive_mads(&port_priv->qp_info[i], NULL);
3062                 if (ret) {
3063                         dev_err(&port_priv->device->dev,
3064                                 "Couldn't post receive WRs\n");
3065                         goto out;
3066                 }
3067         }
3068 out:
3069         kfree(attr);
3070         return ret;
3071 }
3072
3073 static void qp_event_handler(struct ib_event *event, void *qp_context)
3074 {
3075         struct ib_mad_qp_info   *qp_info = qp_context;
3076
3077         /* It's worse than that! He's dead, Jim! */
3078         dev_err(&qp_info->port_priv->device->dev,
3079                 "Fatal error (%d) on MAD QP (%d)\n",
3080                 event->event, qp_info->qp->qp_num);
3081 }
3082
3083 static void init_mad_queue(struct ib_mad_qp_info *qp_info,
3084                            struct ib_mad_queue *mad_queue)
3085 {
3086         mad_queue->qp_info = qp_info;
3087         mad_queue->count = 0;
3088         spin_lock_init(&mad_queue->lock);
3089         INIT_LIST_HEAD(&mad_queue->list);
3090 }
3091
3092 static void init_mad_qp(struct ib_mad_port_private *port_priv,
3093                         struct ib_mad_qp_info *qp_info)
3094 {
3095         qp_info->port_priv = port_priv;
3096         init_mad_queue(qp_info, &qp_info->send_queue);
3097         init_mad_queue(qp_info, &qp_info->recv_queue);
3098         INIT_LIST_HEAD(&qp_info->overflow_list);
3099         spin_lock_init(&qp_info->snoop_lock);
3100         qp_info->snoop_table = NULL;
3101         qp_info->snoop_table_size = 0;
3102         atomic_set(&qp_info->snoop_count, 0);
3103 }
3104
3105 static int create_mad_qp(struct ib_mad_qp_info *qp_info,
3106                          enum ib_qp_type qp_type)
3107 {
3108         struct ib_qp_init_attr  qp_init_attr;
3109         int ret;
3110
3111         memset(&qp_init_attr, 0, sizeof qp_init_attr);
3112         qp_init_attr.send_cq = qp_info->port_priv->cq;
3113         qp_init_attr.recv_cq = qp_info->port_priv->cq;
3114         qp_init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
3115         qp_init_attr.cap.max_send_wr = mad_sendq_size;
3116         qp_init_attr.cap.max_recv_wr = mad_recvq_size;
3117         qp_init_attr.cap.max_send_sge = IB_MAD_SEND_REQ_MAX_SG;
3118         qp_init_attr.cap.max_recv_sge = IB_MAD_RECV_REQ_MAX_SG;
3119         qp_init_attr.qp_type = qp_type;
3120         qp_init_attr.port_num = qp_info->port_priv->port_num;
3121         qp_init_attr.qp_context = qp_info;
3122         qp_init_attr.event_handler = qp_event_handler;
3123         qp_info->qp = ib_create_qp(qp_info->port_priv->pd, &qp_init_attr);
3124         if (IS_ERR(qp_info->qp)) {
3125                 dev_err(&qp_info->port_priv->device->dev,
3126                         "Couldn't create ib_mad QP%d\n",
3127                         get_spl_qp_index(qp_type));
3128                 ret = PTR_ERR(qp_info->qp);
3129                 goto error;
3130         }
3131         /* Use minimum queue sizes unless the CQ is resized */
3132         qp_info->send_queue.max_active = mad_sendq_size;
3133         qp_info->recv_queue.max_active = mad_recvq_size;
3134         return 0;
3135
3136 error:
3137         return ret;
3138 }
3139
3140 static void destroy_mad_qp(struct ib_mad_qp_info *qp_info)
3141 {
3142         if (!qp_info->qp)
3143                 return;
3144
3145         ib_destroy_qp(qp_info->qp);
3146         kfree(qp_info->snoop_table);
3147 }
3148
3149 /*
3150  * Open the port
3151  * Create the QP, PD, MR, and CQ if needed
3152  */
3153 static int ib_mad_port_open(struct ib_device *device,
3154                             int port_num)
3155 {
3156         int ret, cq_size;
3157         struct ib_mad_port_private *port_priv;
3158         unsigned long flags;
3159         char name[sizeof "ib_mad123"];
3160         int has_smi;
3161
3162         if (WARN_ON(rdma_max_mad_size(device, port_num) < IB_MGMT_MAD_SIZE))
3163                 return -EFAULT;
3164
3165         if (WARN_ON(rdma_cap_opa_mad(device, port_num) &&
3166                     rdma_max_mad_size(device, port_num) < OPA_MGMT_MAD_SIZE))
3167                 return -EFAULT;
3168
3169         /* Create new device info */
3170         port_priv = kzalloc(sizeof *port_priv, GFP_KERNEL);
3171         if (!port_priv)
3172                 return -ENOMEM;
3173
3174         port_priv->device = device;
3175         port_priv->port_num = port_num;
3176         spin_lock_init(&port_priv->reg_lock);
3177         init_mad_qp(port_priv, &port_priv->qp_info[0]);
3178         init_mad_qp(port_priv, &port_priv->qp_info[1]);
3179
3180         cq_size = mad_sendq_size + mad_recvq_size;
3181         has_smi = rdma_cap_ib_smi(device, port_num);
3182         if (has_smi)
3183                 cq_size *= 2;
3184
3185         port_priv->cq = ib_alloc_cq(port_priv->device, port_priv, cq_size, 0,
3186                         IB_POLL_WORKQUEUE);
3187         if (IS_ERR(port_priv->cq)) {
3188                 dev_err(&device->dev, "Couldn't create ib_mad CQ\n");
3189                 ret = PTR_ERR(port_priv->cq);
3190                 goto error3;
3191         }
3192
3193         port_priv->pd = ib_alloc_pd(device, 0);
3194         if (IS_ERR(port_priv->pd)) {
3195                 dev_err(&device->dev, "Couldn't create ib_mad PD\n");
3196                 ret = PTR_ERR(port_priv->pd);
3197                 goto error4;
3198         }
3199
3200         if (has_smi) {
3201                 ret = create_mad_qp(&port_priv->qp_info[0], IB_QPT_SMI);
3202                 if (ret)
3203                         goto error6;
3204         }
3205         ret = create_mad_qp(&port_priv->qp_info[1], IB_QPT_GSI);
3206         if (ret)
3207                 goto error7;
3208
3209         snprintf(name, sizeof name, "ib_mad%d", port_num);
3210         port_priv->wq = alloc_ordered_workqueue(name, WQ_MEM_RECLAIM);
3211         if (!port_priv->wq) {
3212                 ret = -ENOMEM;
3213                 goto error8;
3214         }
3215
3216         spin_lock_irqsave(&ib_mad_port_list_lock, flags);
3217         list_add_tail(&port_priv->port_list, &ib_mad_port_list);
3218         spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3219
3220         ret = ib_mad_port_start(port_priv);
3221         if (ret) {
3222                 dev_err(&device->dev, "Couldn't start port\n");
3223                 goto error9;
3224         }
3225
3226         return 0;
3227
3228 error9:
3229         spin_lock_irqsave(&ib_mad_port_list_lock, flags);
3230         list_del_init(&port_priv->port_list);
3231         spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3232
3233         destroy_workqueue(port_priv->wq);
3234 error8:
3235         destroy_mad_qp(&port_priv->qp_info[1]);
3236 error7:
3237         destroy_mad_qp(&port_priv->qp_info[0]);
3238 error6:
3239         ib_dealloc_pd(port_priv->pd);
3240 error4:
3241         ib_free_cq(port_priv->cq);
3242         cleanup_recv_queue(&port_priv->qp_info[1]);
3243         cleanup_recv_queue(&port_priv->qp_info[0]);
3244 error3:
3245         kfree(port_priv);
3246
3247         return ret;
3248 }
3249
3250 /*
3251  * Close the port
3252  * If there are no classes using the port, free the port
3253  * resources (CQ, MR, PD, QP) and remove the port's info structure
3254  */
3255 static int ib_mad_port_close(struct ib_device *device, int port_num)
3256 {
3257         struct ib_mad_port_private *port_priv;
3258         unsigned long flags;
3259
3260         spin_lock_irqsave(&ib_mad_port_list_lock, flags);
3261         port_priv = __ib_get_mad_port(device, port_num);
3262         if (port_priv == NULL) {
3263                 spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3264                 dev_err(&device->dev, "Port %d not found\n", port_num);
3265                 return -ENODEV;
3266         }
3267         list_del_init(&port_priv->port_list);
3268         spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3269
3270         destroy_workqueue(port_priv->wq);
3271         destroy_mad_qp(&port_priv->qp_info[1]);
3272         destroy_mad_qp(&port_priv->qp_info[0]);
3273         ib_dealloc_pd(port_priv->pd);
3274         ib_free_cq(port_priv->cq);
3275         cleanup_recv_queue(&port_priv->qp_info[1]);
3276         cleanup_recv_queue(&port_priv->qp_info[0]);
3277         /* XXX: Handle deallocation of MAD registration tables */
3278
3279         kfree(port_priv);
3280
3281         return 0;
3282 }
3283
3284 static void ib_mad_init_device(struct ib_device *device)
3285 {
3286         int start, i;
3287
3288         start = rdma_start_port(device);
3289
3290         for (i = start; i <= rdma_end_port(device); i++) {
3291                 if (!rdma_cap_ib_mad(device, i))
3292                         continue;
3293
3294                 if (ib_mad_port_open(device, i)) {
3295                         dev_err(&device->dev, "Couldn't open port %d\n", i);
3296                         goto error;
3297                 }
3298                 if (ib_agent_port_open(device, i)) {
3299                         dev_err(&device->dev,
3300                                 "Couldn't open port %d for agents\n", i);
3301                         goto error_agent;
3302                 }
3303         }
3304         return;
3305
3306 error_agent:
3307         if (ib_mad_port_close(device, i))
3308                 dev_err(&device->dev, "Couldn't close port %d\n", i);
3309
3310 error:
3311         while (--i >= start) {
3312                 if (!rdma_cap_ib_mad(device, i))
3313                         continue;
3314
3315                 if (ib_agent_port_close(device, i))
3316                         dev_err(&device->dev,
3317                                 "Couldn't close port %d for agents\n", i);
3318                 if (ib_mad_port_close(device, i))
3319                         dev_err(&device->dev, "Couldn't close port %d\n", i);
3320         }
3321 }
3322
3323 static void ib_mad_remove_device(struct ib_device *device, void *client_data)
3324 {
3325         int i;
3326
3327         for (i = rdma_start_port(device); i <= rdma_end_port(device); i++) {
3328                 if (!rdma_cap_ib_mad(device, i))
3329                         continue;
3330
3331                 if (ib_agent_port_close(device, i))
3332                         dev_err(&device->dev,
3333                                 "Couldn't close port %d for agents\n", i);
3334                 if (ib_mad_port_close(device, i))
3335                         dev_err(&device->dev, "Couldn't close port %d\n", i);
3336         }
3337 }
3338
3339 static struct ib_client mad_client = {
3340         .name   = "mad",
3341         .add = ib_mad_init_device,
3342         .remove = ib_mad_remove_device
3343 };
3344
3345 int ib_mad_init(void)
3346 {
3347         mad_recvq_size = min(mad_recvq_size, IB_MAD_QP_MAX_SIZE);
3348         mad_recvq_size = max(mad_recvq_size, IB_MAD_QP_MIN_SIZE);
3349
3350         mad_sendq_size = min(mad_sendq_size, IB_MAD_QP_MAX_SIZE);
3351         mad_sendq_size = max(mad_sendq_size, IB_MAD_QP_MIN_SIZE);
3352
3353         INIT_LIST_HEAD(&ib_mad_port_list);
3354
3355         /* Client ID 0 is used for snoop-only clients */
3356         idr_alloc(&ib_mad_clients, NULL, 0, 0, GFP_KERNEL);
3357
3358         if (ib_register_client(&mad_client)) {
3359                 pr_err("Couldn't register ib_mad client\n");
3360                 return -EINVAL;
3361         }
3362
3363         return 0;
3364 }
3365
3366 void ib_mad_cleanup(void)
3367 {
3368         ib_unregister_client(&mad_client);
3369 }