sys/ofed/drivers/infiniband/core/ib_mad.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-2-Clause OR GPL-2.0
   3  *
   4  * Copyright (c) 2004-2007 Voltaire, Inc. All rights reserved.
   5  * Copyright (c) 2005 Intel Corporation.  All rights reserved.
   6  * Copyright (c) 2005 Mellanox Technologies Ltd.  All rights reserved.
   7  * Copyright (c) 2009 HNR Consulting. All rights reserved.
   8  * Copyright (c) 2014 Intel Corporation.  All rights reserved.
   9  *
  10  * This software is available to you under a choice of one of two
  11  * licenses.  You may choose to be licensed under the terms of the GNU
  12  * General Public License (GPL) Version 2, available from the file
  13  * COPYING in the main directory of this source tree, or the
  14  * OpenIB.org BSD license below:
  15  *
  16  *     Redistribution and use in source and binary forms, with or
  17  *     without modification, are permitted provided that the following
  18  *     conditions are met:
  19  *
  20  *      - Redistributions of source code must retain the above
  21  *        copyright notice, this list of conditions and the following
  22  *        disclaimer.
  23  *
  24  *      - Redistributions in binary form must reproduce the above
  25  *        copyright notice, this list of conditions and the following
  26  *        disclaimer in the documentation and/or other materials
  27  *        provided with the distribution.
  28  *
  29  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  30  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  31  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  32  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  33  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  34  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  35  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  36  * SOFTWARE.
  37  */
  38
  39 #include <sys/cdefs.h>
  40 __FBSDID("$FreeBSD$");
  41
  42 #define LINUXKPI_PARAM_PREFIX ibcore_
  43 #define KBUILD_MODNAME "ibcore"
  44
  45 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  46
  47 #include <linux/dma-mapping.h>
  48 #include <linux/slab.h>
  49 #include <linux/module.h>
  50 #include <rdma/ib_cache.h>
  51
  52 #include "mad_priv.h"
  53 #include "mad_rmpp.h"
  54 #include "smi.h"
  55 #include "opa_smi.h"
  56 #include "agent.h"
  57 #include "core_priv.h"
  58
  59 static int mad_sendq_size = IB_MAD_QP_SEND_SIZE;
  60 static int mad_recvq_size = IB_MAD_QP_RECV_SIZE;
  61
  62 module_param_named(send_queue_size, mad_sendq_size, int, 0444);
  63 MODULE_PARM_DESC(send_queue_size, "Size of send queue in number of work requests");
  64 module_param_named(recv_queue_size, mad_recvq_size, int, 0444);
  65 MODULE_PARM_DESC(recv_queue_size, "Size of receive queue in number of work requests");
  66
  67 static struct list_head ib_mad_port_list;
  68 static u32 ib_mad_client_id = 0;
  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 struct ib_mad_agent *ib_register_mad_agent(struct ib_device *device,
 201                                            u8 port_num,
 202                                            enum ib_qp_type qp_type,
 203                                            struct ib_mad_reg_req *mad_reg_req,
 204                                            u8 rmpp_version,
 205                                            ib_mad_send_handler send_handler,
 206                                            ib_mad_recv_handler recv_handler,
 207                                            void *context,
 208                                            u32 registration_flags)
 209 {
 210         struct ib_mad_port_private *port_priv;
 211         struct ib_mad_agent *ret = ERR_PTR(-EINVAL);
 212         struct ib_mad_agent_private *mad_agent_priv;
 213         struct ib_mad_reg_req *reg_req = NULL;
 214         struct ib_mad_mgmt_class_table *class;
 215         struct ib_mad_mgmt_vendor_class_table *vendor;
 216         struct ib_mad_mgmt_vendor_class *vendor_class;
 217         struct ib_mad_mgmt_method_table *method;
 218         int ret2, qpn;
 219         unsigned long flags;
 220         u8 mgmt_class, vclass;
 221
 222         /* Validate parameters */
 223         qpn = get_spl_qp_index(qp_type);
 224         if (qpn == -1) {
 225                 dev_notice(&device->dev,
 226                            "ib_register_mad_agent: invalid QP Type %d\n",
 227                            qp_type);
 228                 goto error1;
 229         }
 230
 231         if (rmpp_version && rmpp_version != IB_MGMT_RMPP_VERSION) {
 232                 dev_notice(&device->dev,
 233                            "ib_register_mad_agent: invalid RMPP Version %u\n",
 234                            rmpp_version);
 235                 goto error1;
 236         }
 237
 238         /* Validate MAD registration request if supplied */
 239         if (mad_reg_req) {
 240                 if (mad_reg_req->mgmt_class_version >= MAX_MGMT_VERSION) {
 241                         dev_notice(&device->dev,
 242                                    "ib_register_mad_agent: invalid Class Version %u\n",
 243                                    mad_reg_req->mgmt_class_version);
 244                         goto error1;
 245                 }
 246                 if (!recv_handler) {
 247                         dev_notice(&device->dev,
 248                                    "ib_register_mad_agent: no recv_handler\n");
 249                         goto error1;
 250                 }
 251                 if (mad_reg_req->mgmt_class >= MAX_MGMT_CLASS) {
 252                         /*
 253                          * IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE is the only
 254                          * one in this range currently allowed
 255                          */
 256                         if (mad_reg_req->mgmt_class !=
 257                             IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
 258                                 dev_notice(&device->dev,
 259                                            "ib_register_mad_agent: Invalid Mgmt Class 0x%x\n",
 260                                            mad_reg_req->mgmt_class);
 261                                 goto error1;
 262                         }
 263                 } else if (mad_reg_req->mgmt_class == 0) {
 264                         /*
 265                          * Class 0 is reserved in IBA and is used for
 266                          * aliasing of IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE
 267                          */
 268                         dev_notice(&device->dev,
 269                                    "ib_register_mad_agent: Invalid Mgmt Class 0\n");
 270                         goto error1;
 271                 } else if (is_vendor_class(mad_reg_req->mgmt_class)) {
 272                         /*
 273                          * If class is in "new" vendor range,
 274                          * ensure supplied OUI is not zero
 275                          */
 276                         if (!is_vendor_oui(mad_reg_req->oui)) {
 277                                 dev_notice(&device->dev,
 278                                            "ib_register_mad_agent: No OUI specified for class 0x%x\n",
 279                                            mad_reg_req->mgmt_class);
 280                                 goto error1;
 281                         }
 282                 }
 283                 /* Make sure class supplied is consistent with RMPP */
 284                 if (!ib_is_mad_class_rmpp(mad_reg_req->mgmt_class)) {
 285                         if (rmpp_version) {
 286                                 dev_notice(&device->dev,
 287                                            "ib_register_mad_agent: RMPP version for non-RMPP class 0x%x\n",
 288                                            mad_reg_req->mgmt_class);
 289                                 goto error1;
 290                         }
 291                 }
 292
 293                 /* Make sure class supplied is consistent with QP type */
 294                 if (qp_type == IB_QPT_SMI) {
 295                         if ((mad_reg_req->mgmt_class !=
 296                                         IB_MGMT_CLASS_SUBN_LID_ROUTED) &&
 297                             (mad_reg_req->mgmt_class !=
 298                                         IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
 299                                 dev_notice(&device->dev,
 300                                            "ib_register_mad_agent: Invalid SM QP type: class 0x%x\n",
 301                                            mad_reg_req->mgmt_class);
 302                                 goto error1;
 303                         }
 304                 } else {
 305                         if ((mad_reg_req->mgmt_class ==
 306                                         IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
 307                             (mad_reg_req->mgmt_class ==
 308                                         IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
 309                                 dev_notice(&device->dev,
 310                                            "ib_register_mad_agent: Invalid GS QP type: class 0x%x\n",
 311                                            mad_reg_req->mgmt_class);
 312                                 goto error1;
 313                         }
 314                 }
 315         } else {
 316                 /* No registration request supplied */
 317                 if (!send_handler)
 318                         goto error1;
 319                 if (registration_flags & IB_MAD_USER_RMPP)
 320                         goto error1;
 321         }
 322
 323         /* Validate device and port */
 324         port_priv = ib_get_mad_port(device, port_num);
 325         if (!port_priv) {
 326                 dev_notice(&device->dev, "ib_register_mad_agent: Invalid port\n");
 327                 ret = ERR_PTR(-ENODEV);
 328                 goto error1;
 329         }
 330
 331         /* Verify the QP requested is supported.  For example, Ethernet devices
 332          * will not have QP0 */
 333         if (!port_priv->qp_info[qpn].qp) {
 334                 dev_notice(&device->dev,
 335                            "ib_register_mad_agent: QP %d not supported\n", qpn);
 336                 ret = ERR_PTR(-EPROTONOSUPPORT);
 337                 goto error1;
 338         }
 339
 340         /* Allocate structures */
 341         mad_agent_priv = kzalloc(sizeof *mad_agent_priv, GFP_KERNEL);
 342         if (!mad_agent_priv) {
 343                 ret = ERR_PTR(-ENOMEM);
 344                 goto error1;
 345         }
 346
 347         if (mad_reg_req) {
 348                 reg_req = kmemdup(mad_reg_req, sizeof *reg_req, GFP_KERNEL);
 349                 if (!reg_req) {
 350                         ret = ERR_PTR(-ENOMEM);
 351                         goto error3;
 352                 }
 353         }
 354
 355         /* Now, fill in the various structures */
 356         mad_agent_priv->qp_info = &port_priv->qp_info[qpn];
 357         mad_agent_priv->reg_req = reg_req;
 358         mad_agent_priv->agent.rmpp_version = rmpp_version;
 359         mad_agent_priv->agent.device = device;
 360         mad_agent_priv->agent.recv_handler = recv_handler;
 361         mad_agent_priv->agent.send_handler = send_handler;
 362         mad_agent_priv->agent.context = context;
 363         mad_agent_priv->agent.qp = port_priv->qp_info[qpn].qp;
 364         mad_agent_priv->agent.port_num = port_num;
 365         mad_agent_priv->agent.flags = registration_flags;
 366         spin_lock_init(&mad_agent_priv->lock);
 367         INIT_LIST_HEAD(&mad_agent_priv->send_list);
 368         INIT_LIST_HEAD(&mad_agent_priv->wait_list);
 369         INIT_LIST_HEAD(&mad_agent_priv->done_list);
 370         INIT_LIST_HEAD(&mad_agent_priv->rmpp_list);
 371         INIT_DELAYED_WORK(&mad_agent_priv->timed_work, timeout_sends);
 372         INIT_LIST_HEAD(&mad_agent_priv->local_list);
 373         INIT_WORK(&mad_agent_priv->local_work, local_completions);
 374         atomic_set(&mad_agent_priv->refcount, 1);
 375         init_completion(&mad_agent_priv->comp);
 376
 377         spin_lock_irqsave(&port_priv->reg_lock, flags);
 378         mad_agent_priv->agent.hi_tid = ++ib_mad_client_id;
 379
 380         /*
 381          * Make sure MAD registration (if supplied)
 382          * is non overlapping with any existing ones
 383          */
 384         if (mad_reg_req) {
 385                 mgmt_class = convert_mgmt_class(mad_reg_req->mgmt_class);
 386                 if (!is_vendor_class(mgmt_class)) {
 387                         class = port_priv->version[mad_reg_req->
 388                                                    mgmt_class_version].class;
 389                         if (class) {
 390                                 method = class->method_table[mgmt_class];
 391                                 if (method) {
 392                                         if (method_in_use(&method,
 393                                                            mad_reg_req))
 394                                                 goto error4;
 395                                 }
 396                         }
 397                         ret2 = add_nonoui_reg_req(mad_reg_req, mad_agent_priv,
 398                                                   mgmt_class);
 399                 } else {
 400                         /* "New" vendor class range */
 401                         vendor = port_priv->version[mad_reg_req->
 402                                                     mgmt_class_version].vendor;
 403                         if (vendor) {
 404                                 vclass = vendor_class_index(mgmt_class);
 405                                 vendor_class = vendor->vendor_class[vclass];
 406                                 if (vendor_class) {
 407                                         if (is_vendor_method_in_use(
 408                                                         vendor_class,
 409                                                         mad_reg_req))
 410                                                 goto error4;
 411                                 }
 412                         }
 413                         ret2 = add_oui_reg_req(mad_reg_req, mad_agent_priv);
 414                 }
 415                 if (ret2) {
 416                         ret = ERR_PTR(ret2);
 417                         goto error4;
 418                 }
 419         }
 420
 421         /* Add mad agent into port's agent list */
 422         list_add_tail(&mad_agent_priv->agent_list, &port_priv->agent_list);
 423         spin_unlock_irqrestore(&port_priv->reg_lock, flags);
 424
 425         return &mad_agent_priv->agent;
 426
 427 error4:
 428         spin_unlock_irqrestore(&port_priv->reg_lock, flags);
 429         kfree(reg_req);
 430 error3:
 431         kfree(mad_agent_priv);
 432 error1:
 433         return ret;
 434 }
 435 EXPORT_SYMBOL(ib_register_mad_agent);
 436
 437 static inline int is_snooping_sends(int mad_snoop_flags)
 438 {
 439         return (mad_snoop_flags &
 440                 (/*IB_MAD_SNOOP_POSTED_SENDS |
 441                  IB_MAD_SNOOP_RMPP_SENDS |*/
 442                  IB_MAD_SNOOP_SEND_COMPLETIONS /*|
 443                  IB_MAD_SNOOP_RMPP_SEND_COMPLETIONS*/));
 444 }
 445
 446 static inline int is_snooping_recvs(int mad_snoop_flags)
 447 {
 448         return (mad_snoop_flags &
 449                 (IB_MAD_SNOOP_RECVS /*|
 450                  IB_MAD_SNOOP_RMPP_RECVS*/));
 451 }
 452
 453 static int register_snoop_agent(struct ib_mad_qp_info *qp_info,
 454                                 struct ib_mad_snoop_private *mad_snoop_priv)
 455 {
 456         struct ib_mad_snoop_private **new_snoop_table;
 457         unsigned long flags;
 458         int i;
 459
 460         spin_lock_irqsave(&qp_info->snoop_lock, flags);
 461         /* Check for empty slot in array. */
 462         for (i = 0; i < qp_info->snoop_table_size; i++)
 463                 if (!qp_info->snoop_table[i])
 464                         break;
 465
 466         if (i == qp_info->snoop_table_size) {
 467                 /* Grow table. */
 468                 new_snoop_table = krealloc(qp_info->snoop_table,
 469                                            sizeof mad_snoop_priv *
 470                                            (qp_info->snoop_table_size + 1),
 471                                            GFP_ATOMIC);
 472                 if (!new_snoop_table) {
 473                         i = -ENOMEM;
 474                         goto out;
 475                 }
 476
 477                 qp_info->snoop_table = new_snoop_table;
 478                 qp_info->snoop_table_size++;
 479         }
 480         qp_info->snoop_table[i] = mad_snoop_priv;
 481         atomic_inc(&qp_info->snoop_count);
 482 out:
 483         spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
 484         return i;
 485 }
 486
 487 struct ib_mad_agent *ib_register_mad_snoop(struct ib_device *device,
 488                                            u8 port_num,
 489                                            enum ib_qp_type qp_type,
 490                                            int mad_snoop_flags,
 491                                            ib_mad_snoop_handler snoop_handler,
 492                                            ib_mad_recv_handler recv_handler,
 493                                            void *context)
 494 {
 495         struct ib_mad_port_private *port_priv;
 496         struct ib_mad_agent *ret;
 497         struct ib_mad_snoop_private *mad_snoop_priv;
 498         int qpn;
 499
 500         /* Validate parameters */
 501         if ((is_snooping_sends(mad_snoop_flags) && !snoop_handler) ||
 502             (is_snooping_recvs(mad_snoop_flags) && !recv_handler)) {
 503                 ret = ERR_PTR(-EINVAL);
 504                 goto error1;
 505         }
 506         qpn = get_spl_qp_index(qp_type);
 507         if (qpn == -1) {
 508                 ret = ERR_PTR(-EINVAL);
 509                 goto error1;
 510         }
 511         port_priv = ib_get_mad_port(device, port_num);
 512         if (!port_priv) {
 513                 ret = ERR_PTR(-ENODEV);
 514                 goto error1;
 515         }
 516         /* Allocate structures */
 517         mad_snoop_priv = kzalloc(sizeof *mad_snoop_priv, GFP_KERNEL);
 518         if (!mad_snoop_priv) {
 519                 ret = ERR_PTR(-ENOMEM);
 520                 goto error1;
 521         }
 522
 523         /* Now, fill in the various structures */
 524         mad_snoop_priv->qp_info = &port_priv->qp_info[qpn];
 525         mad_snoop_priv->agent.device = device;
 526         mad_snoop_priv->agent.recv_handler = recv_handler;
 527         mad_snoop_priv->agent.snoop_handler = snoop_handler;
 528         mad_snoop_priv->agent.context = context;
 529         mad_snoop_priv->agent.qp = port_priv->qp_info[qpn].qp;
 530         mad_snoop_priv->agent.port_num = port_num;
 531         mad_snoop_priv->mad_snoop_flags = mad_snoop_flags;
 532         init_completion(&mad_snoop_priv->comp);
 533         mad_snoop_priv->snoop_index = register_snoop_agent(
 534                                                 &port_priv->qp_info[qpn],
 535                                                 mad_snoop_priv);
 536         if (mad_snoop_priv->snoop_index < 0) {
 537                 ret = ERR_PTR(mad_snoop_priv->snoop_index);
 538                 goto error2;
 539         }
 540
 541         atomic_set(&mad_snoop_priv->refcount, 1);
 542         return &mad_snoop_priv->agent;
 543
 544 error2:
 545         kfree(mad_snoop_priv);
 546 error1:
 547         return ret;
 548 }
 549 EXPORT_SYMBOL(ib_register_mad_snoop);
 550
 551 static inline void deref_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
 552 {
 553         if (atomic_dec_and_test(&mad_agent_priv->refcount))
 554                 complete(&mad_agent_priv->comp);
 555 }
 556
 557 static inline void deref_snoop_agent(struct ib_mad_snoop_private *mad_snoop_priv)
 558 {
 559         if (atomic_dec_and_test(&mad_snoop_priv->refcount))
 560                 complete(&mad_snoop_priv->comp);
 561 }
 562
 563 static void unregister_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
 564 {
 565         struct ib_mad_port_private *port_priv;
 566         unsigned long flags;
 567
 568         /* Note that we could still be handling received MADs */
 569
 570         /*
 571          * Canceling all sends results in dropping received response
 572          * MADs, preventing us from queuing additional work
 573          */
 574         cancel_mads(mad_agent_priv);
 575         port_priv = mad_agent_priv->qp_info->port_priv;
 576         cancel_delayed_work_sync(&mad_agent_priv->timed_work);
 577
 578         spin_lock_irqsave(&port_priv->reg_lock, flags);
 579         remove_mad_reg_req(mad_agent_priv);
 580         list_del(&mad_agent_priv->agent_list);
 581         spin_unlock_irqrestore(&port_priv->reg_lock, flags);
 582
 583         flush_workqueue(port_priv->wq);
 584         ib_cancel_rmpp_recvs(mad_agent_priv);
 585
 586         deref_mad_agent(mad_agent_priv);
 587         wait_for_completion(&mad_agent_priv->comp);
 588
 589         kfree(mad_agent_priv->reg_req);
 590         kfree(mad_agent_priv);
 591 }
 592
 593 static void unregister_mad_snoop(struct ib_mad_snoop_private *mad_snoop_priv)
 594 {
 595         struct ib_mad_qp_info *qp_info;
 596         unsigned long flags;
 597
 598         qp_info = mad_snoop_priv->qp_info;
 599         spin_lock_irqsave(&qp_info->snoop_lock, flags);
 600         qp_info->snoop_table[mad_snoop_priv->snoop_index] = NULL;
 601         atomic_dec(&qp_info->snoop_count);
 602         spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
 603
 604         deref_snoop_agent(mad_snoop_priv);
 605         wait_for_completion(&mad_snoop_priv->comp);
 606
 607         kfree(mad_snoop_priv);
 608 }
 609
 610 /*
 611  * ib_unregister_mad_agent - Unregisters a client from using MAD services
 612  */
 613 int ib_unregister_mad_agent(struct ib_mad_agent *mad_agent)
 614 {
 615         struct ib_mad_agent_private *mad_agent_priv;
 616         struct ib_mad_snoop_private *mad_snoop_priv;
 617
 618         /* If the TID is zero, the agent can only snoop. */
 619         if (mad_agent->hi_tid) {
 620                 mad_agent_priv = container_of(mad_agent,
 621                                               struct ib_mad_agent_private,
 622                                               agent);
 623                 unregister_mad_agent(mad_agent_priv);
 624         } else {
 625                 mad_snoop_priv = container_of(mad_agent,
 626                                               struct ib_mad_snoop_private,
 627                                               agent);
 628                 unregister_mad_snoop(mad_snoop_priv);
 629         }
 630         return 0;
 631 }
 632 EXPORT_SYMBOL(ib_unregister_mad_agent);
 633
 634 static void dequeue_mad(struct ib_mad_list_head *mad_list)
 635 {
 636         struct ib_mad_queue *mad_queue;
 637         unsigned long flags;
 638
 639         BUG_ON(!mad_list->mad_queue);
 640         mad_queue = mad_list->mad_queue;
 641         spin_lock_irqsave(&mad_queue->lock, flags);
 642         list_del(&mad_list->list);
 643         mad_queue->count--;
 644         spin_unlock_irqrestore(&mad_queue->lock, flags);
 645 }
 646
 647 static void snoop_send(struct ib_mad_qp_info *qp_info,
 648                        struct ib_mad_send_buf *send_buf,
 649                        struct ib_mad_send_wc *mad_send_wc,
 650                        int mad_snoop_flags)
 651 {
 652         struct ib_mad_snoop_private *mad_snoop_priv;
 653         unsigned long flags;
 654         int i;
 655
 656         spin_lock_irqsave(&qp_info->snoop_lock, flags);
 657         for (i = 0; i < qp_info->snoop_table_size; i++) {
 658                 mad_snoop_priv = qp_info->snoop_table[i];
 659                 if (!mad_snoop_priv ||
 660                     !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
 661                         continue;
 662
 663                 atomic_inc(&mad_snoop_priv->refcount);
 664                 spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
 665                 mad_snoop_priv->agent.snoop_handler(&mad_snoop_priv->agent,
 666                                                     send_buf, mad_send_wc);
 667                 deref_snoop_agent(mad_snoop_priv);
 668                 spin_lock_irqsave(&qp_info->snoop_lock, flags);
 669         }
 670         spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
 671 }
 672
 673 static void snoop_recv(struct ib_mad_qp_info *qp_info,
 674                        struct ib_mad_recv_wc *mad_recv_wc,
 675                        int mad_snoop_flags)
 676 {
 677         struct ib_mad_snoop_private *mad_snoop_priv;
 678         unsigned long flags;
 679         int i;
 680
 681         spin_lock_irqsave(&qp_info->snoop_lock, flags);
 682         for (i = 0; i < qp_info->snoop_table_size; i++) {
 683                 mad_snoop_priv = qp_info->snoop_table[i];
 684                 if (!mad_snoop_priv ||
 685                     !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
 686                         continue;
 687
 688                 atomic_inc(&mad_snoop_priv->refcount);
 689                 spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
 690                 mad_snoop_priv->agent.recv_handler(&mad_snoop_priv->agent, NULL,
 691                                                    mad_recv_wc);
 692                 deref_snoop_agent(mad_snoop_priv);
 693                 spin_lock_irqsave(&qp_info->snoop_lock, flags);
 694         }
 695         spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
 696 }
 697
 698 static void build_smp_wc(struct ib_qp *qp, struct ib_cqe *cqe, u16 slid,
 699                 u16 pkey_index, u8 port_num, struct ib_wc *wc)
 700 {
 701         memset(wc, 0, sizeof *wc);
 702         wc->wr_cqe = cqe;
 703         wc->status = IB_WC_SUCCESS;
 704         wc->opcode = IB_WC_RECV;
 705         wc->pkey_index = pkey_index;
 706         wc->byte_len = sizeof(struct ib_mad) + sizeof(struct ib_grh);
 707         wc->src_qp = IB_QP0;
 708         wc->qp = qp;
 709         wc->slid = slid;
 710         wc->sl = 0;
 711         wc->dlid_path_bits = 0;
 712         wc->port_num = port_num;
 713 }
 714
 715 static size_t mad_priv_size(const struct ib_mad_private *mp)
 716 {
 717         return sizeof(struct ib_mad_private) + mp->mad_size;
 718 }
 719
 720 static struct ib_mad_private *alloc_mad_private(size_t mad_size, gfp_t flags)
 721 {
 722         size_t size = sizeof(struct ib_mad_private) + mad_size;
 723         struct ib_mad_private *ret = kzalloc(size, flags);
 724
 725         if (ret)
 726                 ret->mad_size = mad_size;
 727
 728         return ret;
 729 }
 730
 731 static size_t port_mad_size(const struct ib_mad_port_private *port_priv)
 732 {
 733         return rdma_max_mad_size(port_priv->device, port_priv->port_num);
 734 }
 735
 736 static size_t mad_priv_dma_size(const struct ib_mad_private *mp)
 737 {
 738         return sizeof(struct ib_grh) + mp->mad_size;
 739 }
 740
 741 /*
 742  * Return 0 if SMP is to be sent
 743  * Return 1 if SMP was consumed locally (whether or not solicited)
 744  * Return < 0 if error
 745  */
 746 static int handle_outgoing_dr_smp(struct ib_mad_agent_private *mad_agent_priv,
 747                                   struct ib_mad_send_wr_private *mad_send_wr)
 748 {
 749         int ret = 0;
 750         struct ib_smp *smp = mad_send_wr->send_buf.mad;
 751         struct opa_smp *opa_smp = (struct opa_smp *)smp;
 752         unsigned long flags;
 753         struct ib_mad_local_private *local;
 754         struct ib_mad_private *mad_priv;
 755         struct ib_mad_port_private *port_priv;
 756         struct ib_mad_agent_private *recv_mad_agent = NULL;
 757         struct ib_device *device = mad_agent_priv->agent.device;
 758         u8 port_num;
 759         struct ib_wc mad_wc;
 760         struct ib_ud_wr *send_wr = &mad_send_wr->send_wr;
 761         size_t mad_size = port_mad_size(mad_agent_priv->qp_info->port_priv);
 762         u16 out_mad_pkey_index = 0;
 763         u16 drslid;
 764         bool opa = rdma_cap_opa_mad(mad_agent_priv->qp_info->port_priv->device,
 765                                     mad_agent_priv->qp_info->port_priv->port_num);
 766
 767         if (rdma_cap_ib_switch(device) &&
 768             smp->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
 769                 port_num = send_wr->port_num;
 770         else
 771                 port_num = mad_agent_priv->agent.port_num;
 772
 773         /*
 774          * Directed route handling starts if the initial LID routed part of
 775          * a request or the ending LID routed part of a response is empty.
 776          * If we are at the start of the LID routed part, don't update the
 777          * hop_ptr or hop_cnt.  See section 14.2.2, Vol 1 IB spec.
 778          */
 779         if (opa && smp->class_version == OPA_SMP_CLASS_VERSION) {
 780                 u32 opa_drslid;
 781
 782                 if ((opa_get_smp_direction(opa_smp)
 783                      ? opa_smp->route.dr.dr_dlid : opa_smp->route.dr.dr_slid) ==
 784                      OPA_LID_PERMISSIVE &&
 785                      opa_smi_handle_dr_smp_send(opa_smp,
 786                                                 rdma_cap_ib_switch(device),
 787                                                 port_num) == IB_SMI_DISCARD) {
 788                         ret = -EINVAL;
 789                         dev_err(&device->dev, "OPA Invalid directed route\n");
 790                         goto out;
 791                 }
 792                 opa_drslid = be32_to_cpu(opa_smp->route.dr.dr_slid);
 793                 if (opa_drslid != be32_to_cpu(OPA_LID_PERMISSIVE) &&
 794                     opa_drslid & 0xffff0000) {
 795                         ret = -EINVAL;
 796                         dev_err(&device->dev, "OPA Invalid dr_slid 0x%x\n",
 797                                opa_drslid);
 798                         goto out;
 799                 }
 800                 drslid = (u16)(opa_drslid & 0x0000ffff);
 801
 802                 /* Check to post send on QP or process locally */
 803                 if (opa_smi_check_local_smp(opa_smp, device) == IB_SMI_DISCARD &&
 804                     opa_smi_check_local_returning_smp(opa_smp, device) == IB_SMI_DISCARD)
 805                         goto out;
 806         } else {
 807                 if ((ib_get_smp_direction(smp) ? smp->dr_dlid : smp->dr_slid) ==
 808                      IB_LID_PERMISSIVE &&
 809                      smi_handle_dr_smp_send(smp, rdma_cap_ib_switch(device), port_num) ==
 810                      IB_SMI_DISCARD) {
 811                         ret = -EINVAL;
 812                         dev_err(&device->dev, "Invalid directed route\n");
 813                         goto out;
 814                 }
 815                 drslid = be16_to_cpu(smp->dr_slid);
 816
 817                 /* Check to post send on QP or process locally */
 818                 if (smi_check_local_smp(smp, device) == IB_SMI_DISCARD &&
 819                     smi_check_local_returning_smp(smp, device) == IB_SMI_DISCARD)
 820                         goto out;
 821         }
 822
 823         local = kmalloc(sizeof *local, GFP_ATOMIC);
 824         if (!local) {
 825                 ret = -ENOMEM;
 826                 dev_err(&device->dev, "No memory for ib_mad_local_private\n");
 827                 goto out;
 828         }
 829         local->mad_priv = NULL;
 830         local->recv_mad_agent = NULL;
 831         mad_priv = alloc_mad_private(mad_size, GFP_ATOMIC);
 832         if (!mad_priv) {
 833                 ret = -ENOMEM;
 834                 dev_err(&device->dev, "No memory for local response MAD\n");
 835                 kfree(local);
 836                 goto out;
 837         }
 838
 839         build_smp_wc(mad_agent_priv->agent.qp,
 840                      send_wr->wr.wr_cqe, drslid,
 841                      send_wr->pkey_index,
 842                      send_wr->port_num, &mad_wc);
 843
 844         if (opa && smp->base_version == OPA_MGMT_BASE_VERSION) {
 845                 mad_wc.byte_len = mad_send_wr->send_buf.hdr_len
 846                                         + mad_send_wr->send_buf.data_len
 847                                         + sizeof(struct ib_grh);
 848         }
 849
 850         /* No GRH for DR SMP */
 851         ret = device->process_mad(device, 0, port_num, &mad_wc, NULL,
 852                                   (const struct ib_mad_hdr *)smp, mad_size,
 853                                   (struct ib_mad_hdr *)mad_priv->mad,
 854                                   &mad_size, &out_mad_pkey_index);
 855         switch (ret)
 856         {
 857         case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY:
 858                 if (ib_response_mad((const struct ib_mad_hdr *)mad_priv->mad) &&
 859                     mad_agent_priv->agent.recv_handler) {
 860                         local->mad_priv = mad_priv;
 861                         local->recv_mad_agent = mad_agent_priv;
 862                         /*
 863                          * Reference MAD agent until receive
 864                          * side of local completion handled
 865                          */
 866                         atomic_inc(&mad_agent_priv->refcount);
 867                 } else
 868                         kfree(mad_priv);
 869                 break;
 870         case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED:
 871                 kfree(mad_priv);
 872                 break;
 873         case IB_MAD_RESULT_SUCCESS:
 874                 /* Treat like an incoming receive MAD */
 875                 port_priv = ib_get_mad_port(mad_agent_priv->agent.device,
 876                                             mad_agent_priv->agent.port_num);
 877                 if (port_priv) {
 878                         memcpy(mad_priv->mad, smp, mad_priv->mad_size);
 879                         recv_mad_agent = find_mad_agent(port_priv,
 880                                                         (const struct ib_mad_hdr *)mad_priv->mad);
 881                 }
 882                 if (!port_priv || !recv_mad_agent) {
 883                         /*
 884                          * No receiving agent so drop packet and
 885                          * generate send completion.
 886                          */
 887                         kfree(mad_priv);
 888                         break;
 889                 }
 890                 local->mad_priv = mad_priv;
 891                 local->recv_mad_agent = recv_mad_agent;
 892                 break;
 893         default:
 894                 kfree(mad_priv);
 895                 kfree(local);
 896                 ret = -EINVAL;
 897                 goto out;
 898         }
 899
 900         local->mad_send_wr = mad_send_wr;
 901         if (opa) {
 902                 local->mad_send_wr->send_wr.pkey_index = out_mad_pkey_index;
 903                 local->return_wc_byte_len = mad_size;
 904         }
 905         /* Reference MAD agent until send side of local completion handled */
 906         atomic_inc(&mad_agent_priv->refcount);
 907         /* Queue local completion to local list */
 908         spin_lock_irqsave(&mad_agent_priv->lock, flags);
 909         list_add_tail(&local->completion_list, &mad_agent_priv->local_list);
 910         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
 911         queue_work(mad_agent_priv->qp_info->port_priv->wq,
 912                    &mad_agent_priv->local_work);
 913         ret = 1;
 914 out:
 915         return ret;
 916 }
 917
 918 static int get_pad_size(int hdr_len, int data_len, size_t mad_size)
 919 {
 920         int seg_size, pad;
 921
 922         seg_size = mad_size - hdr_len;
 923         if (data_len && seg_size) {
 924                 pad = seg_size - data_len % seg_size;
 925                 return pad == seg_size ? 0 : pad;
 926         } else
 927                 return seg_size;
 928 }
 929
 930 static void free_send_rmpp_list(struct ib_mad_send_wr_private *mad_send_wr)
 931 {
 932         struct ib_rmpp_segment *s, *t;
 933
 934         list_for_each_entry_safe(s, t, &mad_send_wr->rmpp_list, list) {
 935                 list_del(&s->list);
 936                 kfree(s);
 937         }
 938 }
 939
 940 static int alloc_send_rmpp_list(struct ib_mad_send_wr_private *send_wr,
 941                                 size_t mad_size, gfp_t gfp_mask)
 942 {
 943         struct ib_mad_send_buf *send_buf = &send_wr->send_buf;
 944         struct ib_rmpp_mad *rmpp_mad = send_buf->mad;
 945         struct ib_rmpp_segment *seg = NULL;
 946         int left, seg_size, pad;
 947
 948         send_buf->seg_size = mad_size - send_buf->hdr_len;
 949         send_buf->seg_rmpp_size = mad_size - IB_MGMT_RMPP_HDR;
 950         seg_size = send_buf->seg_size;
 951         pad = send_wr->pad;
 952
 953         /* Allocate data segments. */
 954         for (left = send_buf->data_len + pad; left > 0; left -= seg_size) {
 955                 seg = kmalloc(sizeof (*seg) + seg_size, gfp_mask);
 956                 if (!seg) {
 957                         dev_err(&send_buf->mad_agent->device->dev,
 958                                 "alloc_send_rmpp_segs: RMPP mem alloc failed for len %zd, gfp %#x\n",
 959                                 sizeof (*seg) + seg_size, gfp_mask);
 960                         free_send_rmpp_list(send_wr);
 961                         return -ENOMEM;
 962                 }
 963                 seg->num = ++send_buf->seg_count;
 964                 list_add_tail(&seg->list, &send_wr->rmpp_list);
 965         }
 966
 967         /* Zero any padding */
 968         if (pad)
 969                 memset(seg->data + seg_size - pad, 0, pad);
 970
 971         rmpp_mad->rmpp_hdr.rmpp_version = send_wr->mad_agent_priv->
 972                                           agent.rmpp_version;
 973         rmpp_mad->rmpp_hdr.rmpp_type = IB_MGMT_RMPP_TYPE_DATA;
 974         ib_set_rmpp_flags(&rmpp_mad->rmpp_hdr, IB_MGMT_RMPP_FLAG_ACTIVE);
 975
 976         send_wr->cur_seg = container_of(send_wr->rmpp_list.next,
 977                                         struct ib_rmpp_segment, list);
 978         send_wr->last_ack_seg = send_wr->cur_seg;
 979         return 0;
 980 }
 981
 982 int ib_mad_kernel_rmpp_agent(const struct ib_mad_agent *agent)
 983 {
 984         return agent->rmpp_version && !(agent->flags & IB_MAD_USER_RMPP);
 985 }
 986 EXPORT_SYMBOL(ib_mad_kernel_rmpp_agent);
 987
 988 struct ib_mad_send_buf * ib_create_send_mad(struct ib_mad_agent *mad_agent,
 989                                             u32 remote_qpn, u16 pkey_index,
 990                                             int rmpp_active,
 991                                             int hdr_len, int data_len,
 992                                             gfp_t gfp_mask,
 993                                             u8 base_version)
 994 {
 995         struct ib_mad_agent_private *mad_agent_priv;
 996         struct ib_mad_send_wr_private *mad_send_wr;
 997         int pad, message_size, ret, size;
 998         void *buf;
 999         size_t mad_size;
1000         bool opa;
1001
1002         mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
1003                                       agent);
1004
1005         opa = rdma_cap_opa_mad(mad_agent->device, mad_agent->port_num);
1006
1007         if (opa && base_version == OPA_MGMT_BASE_VERSION)
1008                 mad_size = sizeof(struct opa_mad);
1009         else
1010                 mad_size = sizeof(struct ib_mad);
1011
1012         pad = get_pad_size(hdr_len, data_len, mad_size);
1013         message_size = hdr_len + data_len + pad;
1014
1015         if (ib_mad_kernel_rmpp_agent(mad_agent)) {
1016                 if (!rmpp_active && message_size > mad_size)
1017                         return ERR_PTR(-EINVAL);
1018         } else
1019                 if (rmpp_active || message_size > mad_size)
1020                         return ERR_PTR(-EINVAL);
1021
1022         size = rmpp_active ? hdr_len : mad_size;
1023         buf = kzalloc(sizeof *mad_send_wr + size, gfp_mask);
1024         if (!buf)
1025                 return ERR_PTR(-ENOMEM);
1026
1027         mad_send_wr = (struct ib_mad_send_wr_private *)((char *)buf + size);
1028         INIT_LIST_HEAD(&mad_send_wr->rmpp_list);
1029         mad_send_wr->send_buf.mad = buf;
1030         mad_send_wr->send_buf.hdr_len = hdr_len;
1031         mad_send_wr->send_buf.data_len = data_len;
1032         mad_send_wr->pad = pad;
1033
1034         mad_send_wr->mad_agent_priv = mad_agent_priv;
1035         mad_send_wr->sg_list[0].length = hdr_len;
1036         mad_send_wr->sg_list[0].lkey = mad_agent->qp->pd->local_dma_lkey;
1037
1038         /* OPA MADs don't have to be the full 2048 bytes */
1039         if (opa && base_version == OPA_MGMT_BASE_VERSION &&
1040             data_len < mad_size - hdr_len)
1041                 mad_send_wr->sg_list[1].length = data_len;
1042         else
1043                 mad_send_wr->sg_list[1].length = mad_size - hdr_len;
1044
1045         mad_send_wr->sg_list[1].lkey = mad_agent->qp->pd->local_dma_lkey;
1046
1047         mad_send_wr->mad_list.cqe.done = ib_mad_send_done;
1048
1049         mad_send_wr->send_wr.wr.wr_cqe = &mad_send_wr->mad_list.cqe;
1050         mad_send_wr->send_wr.wr.sg_list = mad_send_wr->sg_list;
1051         mad_send_wr->send_wr.wr.num_sge = 2;
1052         mad_send_wr->send_wr.wr.opcode = IB_WR_SEND;
1053         mad_send_wr->send_wr.wr.send_flags = IB_SEND_SIGNALED;
1054         mad_send_wr->send_wr.remote_qpn = remote_qpn;
1055         mad_send_wr->send_wr.remote_qkey = IB_QP_SET_QKEY;
1056         mad_send_wr->send_wr.pkey_index = pkey_index;
1057
1058         if (rmpp_active) {
1059                 ret = alloc_send_rmpp_list(mad_send_wr, mad_size, gfp_mask);
1060                 if (ret) {
1061                         kfree(buf);
1062                         return ERR_PTR(ret);
1063                 }
1064         }
1065
1066         mad_send_wr->send_buf.mad_agent = mad_agent;
1067         atomic_inc(&mad_agent_priv->refcount);
1068         return &mad_send_wr->send_buf;
1069 }
1070 EXPORT_SYMBOL(ib_create_send_mad);
1071
1072 int ib_get_mad_data_offset(u8 mgmt_class)
1073 {
1074         if (mgmt_class == IB_MGMT_CLASS_SUBN_ADM)
1075                 return IB_MGMT_SA_HDR;
1076         else if ((mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
1077                  (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
1078                  (mgmt_class == IB_MGMT_CLASS_BIS))
1079                 return IB_MGMT_DEVICE_HDR;
1080         else if ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
1081                  (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END))
1082                 return IB_MGMT_VENDOR_HDR;
1083         else
1084                 return IB_MGMT_MAD_HDR;
1085 }
1086 EXPORT_SYMBOL(ib_get_mad_data_offset);
1087
1088 int ib_is_mad_class_rmpp(u8 mgmt_class)
1089 {
1090         if ((mgmt_class == IB_MGMT_CLASS_SUBN_ADM) ||
1091             (mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
1092             (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
1093             (mgmt_class == IB_MGMT_CLASS_BIS) ||
1094             ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
1095              (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END)))
1096                 return 1;
1097         return 0;
1098 }
1099 EXPORT_SYMBOL(ib_is_mad_class_rmpp);
1100
1101 void *ib_get_rmpp_segment(struct ib_mad_send_buf *send_buf, int seg_num)
1102 {
1103         struct ib_mad_send_wr_private *mad_send_wr;
1104         struct list_head *list;
1105
1106         mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
1107                                    send_buf);
1108         list = &mad_send_wr->cur_seg->list;
1109
1110         if (mad_send_wr->cur_seg->num < seg_num) {
1111                 list_for_each_entry(mad_send_wr->cur_seg, list, list)
1112                         if (mad_send_wr->cur_seg->num == seg_num)
1113                                 break;
1114         } else if (mad_send_wr->cur_seg->num > seg_num) {
1115                 list_for_each_entry_reverse(mad_send_wr->cur_seg, list, list)
1116                         if (mad_send_wr->cur_seg->num == seg_num)
1117                                 break;
1118         }
1119         return mad_send_wr->cur_seg->data;
1120 }
1121 EXPORT_SYMBOL(ib_get_rmpp_segment);
1122
1123 static inline void *ib_get_payload(struct ib_mad_send_wr_private *mad_send_wr)
1124 {
1125         if (mad_send_wr->send_buf.seg_count)
1126                 return ib_get_rmpp_segment(&mad_send_wr->send_buf,
1127                                            mad_send_wr->seg_num);
1128         else
1129                 return (char *)mad_send_wr->send_buf.mad +
1130                        mad_send_wr->send_buf.hdr_len;
1131 }
1132
1133 void ib_free_send_mad(struct ib_mad_send_buf *send_buf)
1134 {
1135         struct ib_mad_agent_private *mad_agent_priv;
1136         struct ib_mad_send_wr_private *mad_send_wr;
1137
1138         mad_agent_priv = container_of(send_buf->mad_agent,
1139                                       struct ib_mad_agent_private, agent);
1140         mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
1141                                    send_buf);
1142
1143         free_send_rmpp_list(mad_send_wr);
1144         kfree(send_buf->mad);
1145         deref_mad_agent(mad_agent_priv);
1146 }
1147 EXPORT_SYMBOL(ib_free_send_mad);
1148
1149 int ib_send_mad(struct ib_mad_send_wr_private *mad_send_wr)
1150 {
1151         struct ib_mad_qp_info *qp_info;
1152         struct list_head *list;
1153         struct ib_send_wr *bad_send_wr;
1154         struct ib_mad_agent *mad_agent;
1155         struct ib_sge *sge;
1156         unsigned long flags;
1157         int ret;
1158
1159         /* Set WR ID to find mad_send_wr upon completion */
1160         qp_info = mad_send_wr->mad_agent_priv->qp_info;
1161         mad_send_wr->mad_list.mad_queue = &qp_info->send_queue;
1162         mad_send_wr->mad_list.cqe.done = ib_mad_send_done;
1163         mad_send_wr->send_wr.wr.wr_cqe = &mad_send_wr->mad_list.cqe;
1164
1165         mad_agent = mad_send_wr->send_buf.mad_agent;
1166         sge = mad_send_wr->sg_list;
1167         sge[0].addr = ib_dma_map_single(mad_agent->device,
1168                                         mad_send_wr->send_buf.mad,
1169                                         sge[0].length,
1170                                         DMA_TO_DEVICE);
1171         if (unlikely(ib_dma_mapping_error(mad_agent->device, sge[0].addr)))
1172                 return -ENOMEM;
1173
1174         mad_send_wr->header_mapping = sge[0].addr;
1175
1176         sge[1].addr = ib_dma_map_single(mad_agent->device,
1177                                         ib_get_payload(mad_send_wr),
1178                                         sge[1].length,
1179                                         DMA_TO_DEVICE);
1180         if (unlikely(ib_dma_mapping_error(mad_agent->device, sge[1].addr))) {
1181                 ib_dma_unmap_single(mad_agent->device,
1182                                     mad_send_wr->header_mapping,
1183                                     sge[0].length, DMA_TO_DEVICE);
1184                 return -ENOMEM;
1185         }
1186         mad_send_wr->payload_mapping = sge[1].addr;
1187
1188         spin_lock_irqsave(&qp_info->send_queue.lock, flags);
1189         if (qp_info->send_queue.count < qp_info->send_queue.max_active) {
1190                 ret = ib_post_send(mad_agent->qp, &mad_send_wr->send_wr.wr,
1191                                    &bad_send_wr);
1192                 list = &qp_info->send_queue.list;
1193         } else {
1194                 ret = 0;
1195                 list = &qp_info->overflow_list;
1196         }
1197
1198         if (!ret) {
1199                 qp_info->send_queue.count++;
1200                 list_add_tail(&mad_send_wr->mad_list.list, list);
1201         }
1202         spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
1203         if (ret) {
1204                 ib_dma_unmap_single(mad_agent->device,
1205                                     mad_send_wr->header_mapping,
1206                                     sge[0].length, DMA_TO_DEVICE);
1207                 ib_dma_unmap_single(mad_agent->device,
1208                                     mad_send_wr->payload_mapping,
1209                                     sge[1].length, DMA_TO_DEVICE);
1210         }
1211         return ret;
1212 }
1213
1214 /*
1215  * ib_post_send_mad - Posts MAD(s) to the send queue of the QP associated
1216  *  with the registered client
1217  */
1218 int ib_post_send_mad(struct ib_mad_send_buf *send_buf,
1219                      struct ib_mad_send_buf **bad_send_buf)
1220 {
1221         struct ib_mad_agent_private *mad_agent_priv;
1222         struct ib_mad_send_buf *next_send_buf;
1223         struct ib_mad_send_wr_private *mad_send_wr;
1224         unsigned long flags;
1225         int ret = -EINVAL;
1226
1227         /* Walk list of send WRs and post each on send list */
1228         for (; send_buf; send_buf = next_send_buf) {
1229
1230                 mad_send_wr = container_of(send_buf,
1231                                            struct ib_mad_send_wr_private,
1232                                            send_buf);
1233                 mad_agent_priv = mad_send_wr->mad_agent_priv;
1234
1235                 if (!send_buf->mad_agent->send_handler ||
1236                     (send_buf->timeout_ms &&
1237                      !send_buf->mad_agent->recv_handler)) {
1238                         ret = -EINVAL;
1239                         goto error;
1240                 }
1241
1242                 if (!ib_is_mad_class_rmpp(((struct ib_mad_hdr *) send_buf->mad)->mgmt_class)) {
1243                         if (mad_agent_priv->agent.rmpp_version) {
1244                                 ret = -EINVAL;
1245                                 goto error;
1246                         }
1247                 }
1248
1249                 /*
1250                  * Save pointer to next work request to post in case the
1251                  * current one completes, and the user modifies the work
1252                  * request associated with the completion
1253                  */
1254                 next_send_buf = send_buf->next;
1255                 mad_send_wr->send_wr.ah = send_buf->ah;
1256
1257                 if (((struct ib_mad_hdr *) send_buf->mad)->mgmt_class ==
1258                     IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
1259                         ret = handle_outgoing_dr_smp(mad_agent_priv,
1260                                                      mad_send_wr);
1261                         if (ret < 0)            /* error */
1262                                 goto error;
1263                         else if (ret == 1)      /* locally consumed */
1264                                 continue;
1265                 }
1266
1267                 mad_send_wr->tid = ((struct ib_mad_hdr *) send_buf->mad)->tid;
1268                 /* Timeout will be updated after send completes */
1269                 mad_send_wr->timeout = msecs_to_jiffies(send_buf->timeout_ms);
1270                 mad_send_wr->max_retries = send_buf->retries;
1271                 mad_send_wr->retries_left = send_buf->retries;
1272                 send_buf->retries = 0;
1273                 /* Reference for work request to QP + response */
1274                 mad_send_wr->refcount = 1 + (mad_send_wr->timeout > 0);
1275                 mad_send_wr->status = IB_WC_SUCCESS;
1276
1277                 /* Reference MAD agent until send completes */
1278                 atomic_inc(&mad_agent_priv->refcount);
1279                 spin_lock_irqsave(&mad_agent_priv->lock, flags);
1280                 list_add_tail(&mad_send_wr->agent_list,
1281                               &mad_agent_priv->send_list);
1282                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1283
1284                 if (ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)) {
1285                         ret = ib_send_rmpp_mad(mad_send_wr);
1286                         if (ret >= 0 && ret != IB_RMPP_RESULT_CONSUMED)
1287                                 ret = ib_send_mad(mad_send_wr);
1288                 } else
1289                         ret = ib_send_mad(mad_send_wr);
1290                 if (ret < 0) {
1291                         /* Fail send request */
1292                         spin_lock_irqsave(&mad_agent_priv->lock, flags);
1293                         list_del(&mad_send_wr->agent_list);
1294                         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1295                         atomic_dec(&mad_agent_priv->refcount);
1296                         goto error;
1297                 }
1298         }
1299         return 0;
1300 error:
1301         if (bad_send_buf)
1302                 *bad_send_buf = send_buf;
1303         return ret;
1304 }
1305 EXPORT_SYMBOL(ib_post_send_mad);
1306
1307 /*
1308  * ib_free_recv_mad - Returns data buffers used to receive
1309  *  a MAD to the access layer
1310  */
1311 void ib_free_recv_mad(struct ib_mad_recv_wc *mad_recv_wc)
1312 {
1313         struct ib_mad_recv_buf *mad_recv_buf, *temp_recv_buf;
1314         struct ib_mad_private_header *mad_priv_hdr;
1315         struct ib_mad_private *priv;
1316         struct list_head free_list;
1317
1318         INIT_LIST_HEAD(&free_list);
1319         list_splice_init(&mad_recv_wc->rmpp_list, &free_list);
1320
1321         list_for_each_entry_safe(mad_recv_buf, temp_recv_buf,
1322                                         &free_list, list) {
1323                 mad_recv_wc = container_of(mad_recv_buf, struct ib_mad_recv_wc,
1324                                            recv_buf);
1325                 mad_priv_hdr = container_of(mad_recv_wc,
1326                                             struct ib_mad_private_header,
1327                                             recv_wc);
1328                 priv = container_of(mad_priv_hdr, struct ib_mad_private,
1329                                     header);
1330                 kfree(priv);
1331         }
1332 }
1333 EXPORT_SYMBOL(ib_free_recv_mad);
1334
1335 struct ib_mad_agent *ib_redirect_mad_qp(struct ib_qp *qp,
1336                                         u8 rmpp_version,
1337                                         ib_mad_send_handler send_handler,
1338                                         ib_mad_recv_handler recv_handler,
1339                                         void *context)
1340 {
1341         return ERR_PTR(-EINVAL);        /* XXX: for now */
1342 }
1343 EXPORT_SYMBOL(ib_redirect_mad_qp);
1344
1345 int ib_process_mad_wc(struct ib_mad_agent *mad_agent,
1346                       struct ib_wc *wc)
1347 {
1348         dev_err(&mad_agent->device->dev,
1349                 "ib_process_mad_wc() not implemented yet\n");
1350         return 0;
1351 }
1352 EXPORT_SYMBOL(ib_process_mad_wc);
1353
1354 static int method_in_use(struct ib_mad_mgmt_method_table **method,
1355                          struct ib_mad_reg_req *mad_reg_req)
1356 {
1357         int i;
1358
1359         for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS) {
1360                 if ((*method)->agent[i]) {
1361                         pr_err("Method %d already in use\n", i);
1362                         return -EINVAL;
1363                 }
1364         }
1365         return 0;
1366 }
1367
1368 static int allocate_method_table(struct ib_mad_mgmt_method_table **method)
1369 {
1370         /* Allocate management method table */
1371         *method = kzalloc(sizeof **method, GFP_ATOMIC);
1372         if (!*method) {
1373                 pr_err("No memory for ib_mad_mgmt_method_table\n");
1374                 return -ENOMEM;
1375         }
1376
1377         return 0;
1378 }
1379
1380 /*
1381  * Check to see if there are any methods still in use
1382  */
1383 static int check_method_table(struct ib_mad_mgmt_method_table *method)
1384 {
1385         int i;
1386
1387         for (i = 0; i < IB_MGMT_MAX_METHODS; i++)
1388                 if (method->agent[i])
1389                         return 1;
1390         return 0;
1391 }
1392
1393 /*
1394  * Check to see if there are any method tables for this class still in use
1395  */
1396 static int check_class_table(struct ib_mad_mgmt_class_table *class)
1397 {
1398         int i;
1399
1400         for (i = 0; i < MAX_MGMT_CLASS; i++)
1401                 if (class->method_table[i])
1402                         return 1;
1403         return 0;
1404 }
1405
1406 static int check_vendor_class(struct ib_mad_mgmt_vendor_class *vendor_class)
1407 {
1408         int i;
1409
1410         for (i = 0; i < MAX_MGMT_OUI; i++)
1411                 if (vendor_class->method_table[i])
1412                         return 1;
1413         return 0;
1414 }
1415
1416 static int find_vendor_oui(struct ib_mad_mgmt_vendor_class *vendor_class,
1417                            const char *oui)
1418 {
1419         int i;
1420
1421         for (i = 0; i < MAX_MGMT_OUI; i++)
1422                 /* Is there matching OUI for this vendor class ? */
1423                 if (!memcmp(vendor_class->oui[i], oui, 3))
1424                         return i;
1425
1426         return -1;
1427 }
1428
1429 static int check_vendor_table(struct ib_mad_mgmt_vendor_class_table *vendor)
1430 {
1431         int i;
1432
1433         for (i = 0; i < MAX_MGMT_VENDOR_RANGE2; i++)
1434                 if (vendor->vendor_class[i])
1435                         return 1;
1436
1437         return 0;
1438 }
1439
1440 static void remove_methods_mad_agent(struct ib_mad_mgmt_method_table *method,
1441                                      struct ib_mad_agent_private *agent)
1442 {
1443         int i;
1444
1445         /* Remove any methods for this mad agent */
1446         for (i = 0; i < IB_MGMT_MAX_METHODS; i++) {
1447                 if (method->agent[i] == agent) {
1448                         method->agent[i] = NULL;
1449                 }
1450         }
1451 }
1452
1453 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1454                               struct ib_mad_agent_private *agent_priv,
1455                               u8 mgmt_class)
1456 {
1457         struct ib_mad_port_private *port_priv;
1458         struct ib_mad_mgmt_class_table **class;
1459         struct ib_mad_mgmt_method_table **method;
1460         int i, ret;
1461
1462         port_priv = agent_priv->qp_info->port_priv;
1463         class = &port_priv->version[mad_reg_req->mgmt_class_version].class;
1464         if (!*class) {
1465                 /* Allocate management class table for "new" class version */
1466                 *class = kzalloc(sizeof **class, GFP_ATOMIC);
1467                 if (!*class) {
1468                         dev_err(&agent_priv->agent.device->dev,
1469                                 "No memory for ib_mad_mgmt_class_table\n");
1470                         ret = -ENOMEM;
1471                         goto error1;
1472                 }
1473
1474                 /* Allocate method table for this management class */
1475                 method = &(*class)->method_table[mgmt_class];
1476                 if ((ret = allocate_method_table(method)))
1477                         goto error2;
1478         } else {
1479                 method = &(*class)->method_table[mgmt_class];
1480                 if (!*method) {
1481                         /* Allocate method table for this management class */
1482                         if ((ret = allocate_method_table(method)))
1483                                 goto error1;
1484                 }
1485         }
1486
1487         /* Now, make sure methods are not already in use */
1488         if (method_in_use(method, mad_reg_req))
1489                 goto error3;
1490
1491         /* Finally, add in methods being registered */
1492         for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS)
1493                 (*method)->agent[i] = agent_priv;
1494
1495         return 0;
1496
1497 error3:
1498         /* Remove any methods for this mad agent */
1499         remove_methods_mad_agent(*method, agent_priv);
1500         /* Now, check to see if there are any methods in use */
1501         if (!check_method_table(*method)) {
1502                 /* If not, release management method table */
1503                 kfree(*method);
1504                 *method = NULL;
1505         }
1506         ret = -EINVAL;
1507         goto error1;
1508 error2:
1509         kfree(*class);
1510         *class = NULL;
1511 error1:
1512         return ret;
1513 }
1514
1515 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1516                            struct ib_mad_agent_private *agent_priv)
1517 {
1518         struct ib_mad_port_private *port_priv;
1519         struct ib_mad_mgmt_vendor_class_table **vendor_table;
1520         struct ib_mad_mgmt_vendor_class_table *vendor = NULL;
1521         struct ib_mad_mgmt_vendor_class *vendor_class = NULL;
1522         struct ib_mad_mgmt_method_table **method;
1523         int i, ret = -ENOMEM;
1524         u8 vclass;
1525
1526         /* "New" vendor (with OUI) class */
1527         vclass = vendor_class_index(mad_reg_req->mgmt_class);
1528         port_priv = agent_priv->qp_info->port_priv;
1529         vendor_table = &port_priv->version[
1530                                 mad_reg_req->mgmt_class_version].vendor;
1531         if (!*vendor_table) {
1532                 /* Allocate mgmt vendor class table for "new" class version */
1533                 vendor = kzalloc(sizeof *vendor, GFP_ATOMIC);
1534                 if (!vendor) {
1535                         dev_err(&agent_priv->agent.device->dev,
1536                                 "No memory for ib_mad_mgmt_vendor_class_table\n");
1537                         goto error1;
1538                 }
1539
1540                 *vendor_table = vendor;
1541         }
1542         if (!(*vendor_table)->vendor_class[vclass]) {
1543                 /* Allocate table for this management vendor class */
1544                 vendor_class = kzalloc(sizeof *vendor_class, GFP_ATOMIC);
1545                 if (!vendor_class) {
1546                         dev_err(&agent_priv->agent.device->dev,
1547                                 "No memory for ib_mad_mgmt_vendor_class\n");
1548                         goto error2;
1549                 }
1550
1551                 (*vendor_table)->vendor_class[vclass] = vendor_class;
1552         }
1553         for (i = 0; i < MAX_MGMT_OUI; i++) {
1554                 /* Is there matching OUI for this vendor class ? */
1555                 if (!memcmp((*vendor_table)->vendor_class[vclass]->oui[i],
1556                             mad_reg_req->oui, 3)) {
1557                         method = &(*vendor_table)->vendor_class[
1558                                                 vclass]->method_table[i];
1559                         BUG_ON(!*method);
1560                         goto check_in_use;
1561                 }
1562         }
1563         for (i = 0; i < MAX_MGMT_OUI; i++) {
1564                 /* OUI slot available ? */
1565                 if (!is_vendor_oui((*vendor_table)->vendor_class[
1566                                 vclass]->oui[i])) {
1567                         method = &(*vendor_table)->vendor_class[
1568                                 vclass]->method_table[i];
1569                         BUG_ON(*method);
1570                         /* Allocate method table for this OUI */
1571                         if ((ret = allocate_method_table(method)))
1572                                 goto error3;
1573                         memcpy((*vendor_table)->vendor_class[vclass]->oui[i],
1574                                mad_reg_req->oui, 3);
1575                         goto check_in_use;
1576                 }
1577         }
1578         dev_err(&agent_priv->agent.device->dev, "All OUI slots in use\n");
1579         goto error3;
1580
1581 check_in_use:
1582         /* Now, make sure methods are not already in use */
1583         if (method_in_use(method, mad_reg_req))
1584                 goto error4;
1585
1586         /* Finally, add in methods being registered */
1587         for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS)
1588                 (*method)->agent[i] = agent_priv;
1589
1590         return 0;
1591
1592 error4:
1593         /* Remove any methods for this mad agent */
1594         remove_methods_mad_agent(*method, agent_priv);
1595         /* Now, check to see if there are any methods in use */
1596         if (!check_method_table(*method)) {
1597                 /* If not, release management method table */
1598                 kfree(*method);
1599                 *method = NULL;
1600         }
1601         ret = -EINVAL;
1602 error3:
1603         if (vendor_class) {
1604                 (*vendor_table)->vendor_class[vclass] = NULL;
1605                 kfree(vendor_class);
1606         }
1607 error2:
1608         if (vendor) {
1609                 *vendor_table = NULL;
1610                 kfree(vendor);
1611         }
1612 error1:
1613         return ret;
1614 }
1615
1616 static void remove_mad_reg_req(struct ib_mad_agent_private *agent_priv)
1617 {
1618         struct ib_mad_port_private *port_priv;
1619         struct ib_mad_mgmt_class_table *class;
1620         struct ib_mad_mgmt_method_table *method;
1621         struct ib_mad_mgmt_vendor_class_table *vendor;
1622         struct ib_mad_mgmt_vendor_class *vendor_class;
1623         int index;
1624         u8 mgmt_class;
1625
1626         /*
1627          * Was MAD registration request supplied
1628          * with original registration ?
1629          */
1630         if (!agent_priv->reg_req) {
1631                 goto out;
1632         }
1633
1634         port_priv = agent_priv->qp_info->port_priv;
1635         mgmt_class = convert_mgmt_class(agent_priv->reg_req->mgmt_class);
1636         class = port_priv->version[
1637                         agent_priv->reg_req->mgmt_class_version].class;
1638         if (!class)
1639                 goto vendor_check;
1640
1641         method = class->method_table[mgmt_class];
1642         if (method) {
1643                 /* Remove any methods for this mad agent */
1644                 remove_methods_mad_agent(method, agent_priv);
1645                 /* Now, check to see if there are any methods still in use */
1646                 if (!check_method_table(method)) {
1647                         /* If not, release management method table */
1648                         kfree(method);
1649                         class->method_table[mgmt_class] = NULL;
1650                         /* Any management classes left ? */
1651                         if (!check_class_table(class)) {
1652                                 /* If not, release management class table */
1653                                 kfree(class);
1654                                 port_priv->version[
1655                                         agent_priv->reg_req->
1656                                         mgmt_class_version].class = NULL;
1657                         }
1658                 }
1659         }
1660
1661 vendor_check:
1662         if (!is_vendor_class(mgmt_class))
1663                 goto out;
1664
1665         /* normalize mgmt_class to vendor range 2 */
1666         mgmt_class = vendor_class_index(agent_priv->reg_req->mgmt_class);
1667         vendor = port_priv->version[
1668                         agent_priv->reg_req->mgmt_class_version].vendor;
1669
1670         if (!vendor)
1671                 goto out;
1672
1673         vendor_class = vendor->vendor_class[mgmt_class];
1674         if (vendor_class) {
1675                 index = find_vendor_oui(vendor_class, agent_priv->reg_req->oui);
1676                 if (index < 0)
1677                         goto out;
1678                 method = vendor_class->method_table[index];
1679                 if (method) {
1680                         /* Remove any methods for this mad agent */
1681                         remove_methods_mad_agent(method, agent_priv);
1682                         /*
1683                          * Now, check to see if there are
1684                          * any methods still in use
1685                          */
1686                         if (!check_method_table(method)) {
1687                                 /* If not, release management method table */
1688                                 kfree(method);
1689                                 vendor_class->method_table[index] = NULL;
1690                                 memset(vendor_class->oui[index], 0, 3);
1691                                 /* Any OUIs left ? */
1692                                 if (!check_vendor_class(vendor_class)) {
1693                                         /* If not, release vendor class table */
1694                                         kfree(vendor_class);
1695                                         vendor->vendor_class[mgmt_class] = NULL;
1696                                         /* Any other vendor classes left ? */
1697                                         if (!check_vendor_table(vendor)) {
1698                                                 kfree(vendor);
1699                                                 port_priv->version[
1700                                                         agent_priv->reg_req->
1701                                                         mgmt_class_version].
1702                                                         vendor = NULL;
1703                                         }
1704                                 }
1705                         }
1706                 }
1707         }
1708
1709 out:
1710         return;
1711 }
1712
1713 static struct ib_mad_agent_private *
1714 find_mad_agent(struct ib_mad_port_private *port_priv,
1715                const struct ib_mad_hdr *mad_hdr)
1716 {
1717         struct ib_mad_agent_private *mad_agent = NULL;
1718         unsigned long flags;
1719
1720         spin_lock_irqsave(&port_priv->reg_lock, flags);
1721         if (ib_response_mad(mad_hdr)) {
1722                 u32 hi_tid;
1723                 struct ib_mad_agent_private *entry;
1724
1725                 /*
1726                  * Routing is based on high 32 bits of transaction ID
1727                  * of MAD.
1728                  */
1729                 hi_tid = be64_to_cpu(mad_hdr->tid) >> 32;
1730                 list_for_each_entry(entry, &port_priv->agent_list, agent_list) {
1731                         if (entry->agent.hi_tid == hi_tid) {
1732                                 mad_agent = entry;
1733                                 break;
1734                         }
1735                 }
1736         } else {
1737                 struct ib_mad_mgmt_class_table *class;
1738                 struct ib_mad_mgmt_method_table *method;
1739                 struct ib_mad_mgmt_vendor_class_table *vendor;
1740                 struct ib_mad_mgmt_vendor_class *vendor_class;
1741                 const struct ib_vendor_mad *vendor_mad;
1742                 int index;
1743
1744                 /*
1745                  * Routing is based on version, class, and method
1746                  * For "newer" vendor MADs, also based on OUI
1747                  */
1748                 if (mad_hdr->class_version >= MAX_MGMT_VERSION)
1749                         goto out;
1750                 if (!is_vendor_class(mad_hdr->mgmt_class)) {
1751                         class = port_priv->version[
1752                                         mad_hdr->class_version].class;
1753                         if (!class)
1754                                 goto out;
1755                         if (convert_mgmt_class(mad_hdr->mgmt_class) >=
1756                             ARRAY_SIZE(class->method_table))
1757                                 goto out;
1758                         method = class->method_table[convert_mgmt_class(
1759                                                         mad_hdr->mgmt_class)];
1760                         if (method)
1761                                 mad_agent = method->agent[mad_hdr->method &
1762                                                           ~IB_MGMT_METHOD_RESP];
1763                 } else {
1764                         vendor = port_priv->version[
1765                                         mad_hdr->class_version].vendor;
1766                         if (!vendor)
1767                                 goto out;
1768                         vendor_class = vendor->vendor_class[vendor_class_index(
1769                                                 mad_hdr->mgmt_class)];
1770                         if (!vendor_class)
1771                                 goto out;
1772                         /* Find matching OUI */
1773                         vendor_mad = (const struct ib_vendor_mad *)mad_hdr;
1774                         index = find_vendor_oui(vendor_class, vendor_mad->oui);
1775                         if (index == -1)
1776                                 goto out;
1777                         method = vendor_class->method_table[index];
1778                         if (method) {
1779                                 mad_agent = method->agent[mad_hdr->method &
1780                                                           ~IB_MGMT_METHOD_RESP];
1781                         }
1782                 }
1783         }
1784
1785         if (mad_agent) {
1786                 if (mad_agent->agent.recv_handler)
1787                         atomic_inc(&mad_agent->refcount);
1788                 else {
1789                         dev_notice(&port_priv->device->dev,
1790                                    "No receive handler for client %p on port %d\n",
1791                                    &mad_agent->agent, port_priv->port_num);
1792                         mad_agent = NULL;
1793                 }
1794         }
1795 out:
1796         spin_unlock_irqrestore(&port_priv->reg_lock, flags);
1797
1798         return mad_agent;
1799 }
1800
1801 static int validate_mad(const struct ib_mad_hdr *mad_hdr,
1802                         const struct ib_mad_qp_info *qp_info,
1803                         bool opa)
1804 {
1805         int valid = 0;
1806         u32 qp_num = qp_info->qp->qp_num;
1807
1808         /* Make sure MAD base version is understood */
1809         if (mad_hdr->base_version != IB_MGMT_BASE_VERSION &&
1810             (!opa || mad_hdr->base_version != OPA_MGMT_BASE_VERSION)) {
1811                 pr_err("MAD received with unsupported base version %d %s\n",
1812                        mad_hdr->base_version, opa ? "(opa)" : "");
1813                 goto out;
1814         }
1815
1816         /* Filter SMI packets sent to other than QP0 */
1817         if ((mad_hdr->mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
1818             (mad_hdr->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
1819                 if (qp_num == 0)
1820                         valid = 1;
1821         } else {
1822                 /* CM attributes other than ClassPortInfo only use Send method */
1823                 if ((mad_hdr->mgmt_class == IB_MGMT_CLASS_CM) &&
1824                     (mad_hdr->attr_id != IB_MGMT_CLASSPORTINFO_ATTR_ID) &&
1825                     (mad_hdr->method != IB_MGMT_METHOD_SEND))
1826                         goto out;
1827                 /* Filter GSI packets sent to QP0 */
1828                 if (qp_num != 0)
1829                         valid = 1;
1830         }
1831
1832 out:
1833         return valid;
1834 }
1835
1836 static int is_rmpp_data_mad(const struct ib_mad_agent_private *mad_agent_priv,
1837                             const struct ib_mad_hdr *mad_hdr)
1838 {
1839         const struct ib_rmpp_mad *rmpp_mad;
1840
1841         rmpp_mad = (const struct ib_rmpp_mad *)mad_hdr;
1842         return !mad_agent_priv->agent.rmpp_version ||
1843                 !ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent) ||
1844                 !(ib_get_rmpp_flags(&rmpp_mad->rmpp_hdr) &
1845                                     IB_MGMT_RMPP_FLAG_ACTIVE) ||
1846                 (rmpp_mad->rmpp_hdr.rmpp_type == IB_MGMT_RMPP_TYPE_DATA);
1847 }
1848
1849 static inline int rcv_has_same_class(const struct ib_mad_send_wr_private *wr,
1850                                      const struct ib_mad_recv_wc *rwc)
1851 {
1852         return ((struct ib_mad_hdr *)(wr->send_buf.mad))->mgmt_class ==
1853                 rwc->recv_buf.mad->mad_hdr.mgmt_class;
1854 }
1855
1856 static inline int rcv_has_same_gid(const struct ib_mad_agent_private *mad_agent_priv,
1857                                    const struct ib_mad_send_wr_private *wr,
1858                                    const struct ib_mad_recv_wc *rwc )
1859 {
1860         struct ib_ah_attr attr;
1861         u8 send_resp, rcv_resp;
1862         union ib_gid sgid;
1863         struct ib_device *device = mad_agent_priv->agent.device;
1864         u8 port_num = mad_agent_priv->agent.port_num;
1865         u8 lmc;
1866
1867         send_resp = ib_response_mad((struct ib_mad_hdr *)wr->send_buf.mad);
1868         rcv_resp = ib_response_mad(&rwc->recv_buf.mad->mad_hdr);
1869
1870         if (send_resp == rcv_resp)
1871                 /* both requests, or both responses. GIDs different */
1872                 return 0;
1873
1874         if (ib_query_ah(wr->send_buf.ah, &attr))
1875                 /* Assume not equal, to avoid false positives. */
1876                 return 0;
1877
1878         if (!!(attr.ah_flags & IB_AH_GRH) !=
1879             !!(rwc->wc->wc_flags & IB_WC_GRH))
1880                 /* one has GID, other does not.  Assume different */
1881                 return 0;
1882
1883         if (!send_resp && rcv_resp) {
1884                 /* is request/response. */
1885                 if (!(attr.ah_flags & IB_AH_GRH)) {
1886                         if (ib_get_cached_lmc(device, port_num, &lmc))
1887                                 return 0;
1888                         return (!lmc || !((attr.src_path_bits ^
1889                                            rwc->wc->dlid_path_bits) &
1890                                           ((1 << lmc) - 1)));
1891                 } else {
1892                         if (ib_get_cached_gid(device, port_num,
1893                                               attr.grh.sgid_index, &sgid, NULL))
1894                                 return 0;
1895                         return !memcmp(sgid.raw, rwc->recv_buf.grh->dgid.raw,
1896                                        16);
1897                 }
1898         }
1899
1900         if (!(attr.ah_flags & IB_AH_GRH))
1901                 return attr.dlid == rwc->wc->slid;
1902         else
1903                 return !memcmp(attr.grh.dgid.raw, rwc->recv_buf.grh->sgid.raw,
1904                                16);
1905 }
1906
1907 static inline int is_direct(u8 class)
1908 {
1909         return (class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE);
1910 }
1911
1912 struct ib_mad_send_wr_private*
1913 ib_find_send_mad(const struct ib_mad_agent_private *mad_agent_priv,
1914                  const struct ib_mad_recv_wc *wc)
1915 {
1916         struct ib_mad_send_wr_private *wr;
1917         const struct ib_mad_hdr *mad_hdr;
1918
1919         mad_hdr = &wc->recv_buf.mad->mad_hdr;
1920
1921         list_for_each_entry(wr, &mad_agent_priv->wait_list, agent_list) {
1922                 if ((wr->tid == mad_hdr->tid) &&
1923                     rcv_has_same_class(wr, wc) &&
1924                     /*
1925                      * Don't check GID for direct routed MADs.
1926                      * These might have permissive LIDs.
1927                      */
1928                     (is_direct(mad_hdr->mgmt_class) ||
1929                      rcv_has_same_gid(mad_agent_priv, wr, wc)))
1930                         return (wr->status == IB_WC_SUCCESS) ? wr : NULL;
1931         }
1932
1933         /*
1934          * It's possible to receive the response before we've
1935          * been notified that the send has completed
1936          */
1937         list_for_each_entry(wr, &mad_agent_priv->send_list, agent_list) {
1938                 if (is_rmpp_data_mad(mad_agent_priv, wr->send_buf.mad) &&
1939                     wr->tid == mad_hdr->tid &&
1940                     wr->timeout &&
1941                     rcv_has_same_class(wr, wc) &&
1942                     /*
1943                      * Don't check GID for direct routed MADs.
1944                      * These might have permissive LIDs.
1945                      */
1946                     (is_direct(mad_hdr->mgmt_class) ||
1947                      rcv_has_same_gid(mad_agent_priv, wr, wc)))
1948                         /* Verify request has not been canceled */
1949                         return (wr->status == IB_WC_SUCCESS) ? wr : NULL;
1950         }
1951         return NULL;
1952 }
1953
1954 void ib_mark_mad_done(struct ib_mad_send_wr_private *mad_send_wr)
1955 {
1956         mad_send_wr->timeout = 0;
1957         if (mad_send_wr->refcount == 1)
1958                 list_move_tail(&mad_send_wr->agent_list,
1959                               &mad_send_wr->mad_agent_priv->done_list);
1960 }
1961
1962 static void ib_mad_complete_recv(struct ib_mad_agent_private *mad_agent_priv,
1963                                  struct ib_mad_recv_wc *mad_recv_wc)
1964 {
1965         struct ib_mad_send_wr_private *mad_send_wr;
1966         struct ib_mad_send_wc mad_send_wc;
1967         unsigned long flags;
1968
1969         INIT_LIST_HEAD(&mad_recv_wc->rmpp_list);
1970         list_add(&mad_recv_wc->recv_buf.list, &mad_recv_wc->rmpp_list);
1971         if (ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)) {
1972                 mad_recv_wc = ib_process_rmpp_recv_wc(mad_agent_priv,
1973                                                       mad_recv_wc);
1974                 if (!mad_recv_wc) {
1975                         deref_mad_agent(mad_agent_priv);
1976                         return;
1977                 }
1978         }
1979
1980         /* Complete corresponding request */
1981         if (ib_response_mad(&mad_recv_wc->recv_buf.mad->mad_hdr)) {
1982                 spin_lock_irqsave(&mad_agent_priv->lock, flags);
1983                 mad_send_wr = ib_find_send_mad(mad_agent_priv, mad_recv_wc);
1984                 if (!mad_send_wr) {
1985                         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1986                         if (!ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)
1987                            && ib_is_mad_class_rmpp(mad_recv_wc->recv_buf.mad->mad_hdr.mgmt_class)
1988                            && (ib_get_rmpp_flags(&((struct ib_rmpp_mad *)mad_recv_wc->recv_buf.mad)->rmpp_hdr)
1989                                         & IB_MGMT_RMPP_FLAG_ACTIVE)) {
1990                                 /* user rmpp is in effect
1991                                  * and this is an active RMPP MAD
1992                                  */
1993                                 mad_agent_priv->agent.recv_handler(
1994                                                 &mad_agent_priv->agent, NULL,
1995                                                 mad_recv_wc);
1996                                 atomic_dec(&mad_agent_priv->refcount);
1997                         } else {
1998                                 /* not user rmpp, revert to normal behavior and
1999                                  * drop the mad */
2000                                 ib_free_recv_mad(mad_recv_wc);
2001                                 deref_mad_agent(mad_agent_priv);
2002                                 return;
2003                         }
2004                 } else {
2005                         ib_mark_mad_done(mad_send_wr);
2006                         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2007
2008                         /* Defined behavior is to complete response before request */
2009                         mad_agent_priv->agent.recv_handler(
2010                                         &mad_agent_priv->agent,
2011                                         &mad_send_wr->send_buf,
2012                                         mad_recv_wc);
2013                         atomic_dec(&mad_agent_priv->refcount);
2014
2015                         mad_send_wc.status = IB_WC_SUCCESS;
2016                         mad_send_wc.vendor_err = 0;
2017                         mad_send_wc.send_buf = &mad_send_wr->send_buf;
2018                         ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
2019                 }
2020         } else {
2021                 mad_agent_priv->agent.recv_handler(&mad_agent_priv->agent, NULL,
2022                                                    mad_recv_wc);
2023                 deref_mad_agent(mad_agent_priv);
2024         }
2025 }
2026
2027 static enum smi_action handle_ib_smi(const struct ib_mad_port_private *port_priv,
2028                                      const struct ib_mad_qp_info *qp_info,
2029                                      const struct ib_wc *wc,
2030                                      int port_num,
2031                                      struct ib_mad_private *recv,
2032                                      struct ib_mad_private *response)
2033 {
2034         enum smi_forward_action retsmi;
2035         struct ib_smp *smp = (struct ib_smp *)recv->mad;
2036
2037         if (smi_handle_dr_smp_recv(smp,
2038                                    rdma_cap_ib_switch(port_priv->device),
2039                                    port_num,
2040                                    port_priv->device->phys_port_cnt) ==
2041                                    IB_SMI_DISCARD)
2042                 return IB_SMI_DISCARD;
2043
2044         retsmi = smi_check_forward_dr_smp(smp);
2045         if (retsmi == IB_SMI_LOCAL)
2046                 return IB_SMI_HANDLE;
2047
2048         if (retsmi == IB_SMI_SEND) { /* don't forward */
2049                 if (smi_handle_dr_smp_send(smp,
2050                                            rdma_cap_ib_switch(port_priv->device),
2051                                            port_num) == IB_SMI_DISCARD)
2052                         return IB_SMI_DISCARD;
2053
2054                 if (smi_check_local_smp(smp, port_priv->device) == IB_SMI_DISCARD)
2055                         return IB_SMI_DISCARD;
2056         } else if (rdma_cap_ib_switch(port_priv->device)) {
2057                 /* forward case for switches */
2058                 memcpy(response, recv, mad_priv_size(response));
2059                 response->header.recv_wc.wc = &response->header.wc;
2060                 response->header.recv_wc.recv_buf.mad = (struct ib_mad *)response->mad;
2061                 response->header.recv_wc.recv_buf.grh = &response->grh;
2062
2063                 agent_send_response((const struct ib_mad_hdr *)response->mad,
2064                                     &response->grh, wc,
2065                                     port_priv->device,
2066                                     smi_get_fwd_port(smp),
2067                                     qp_info->qp->qp_num,
2068                                     response->mad_size,
2069                                     false);
2070
2071                 return IB_SMI_DISCARD;
2072         }
2073         return IB_SMI_HANDLE;
2074 }
2075
2076 static bool generate_unmatched_resp(const struct ib_mad_private *recv,
2077                                     struct ib_mad_private *response,
2078                                     size_t *resp_len, bool opa)
2079 {
2080         const struct ib_mad_hdr *recv_hdr = (const struct ib_mad_hdr *)recv->mad;
2081         struct ib_mad_hdr *resp_hdr = (struct ib_mad_hdr *)response->mad;
2082
2083         if (recv_hdr->method == IB_MGMT_METHOD_GET ||
2084             recv_hdr->method == IB_MGMT_METHOD_SET) {
2085                 memcpy(response, recv, mad_priv_size(response));
2086                 response->header.recv_wc.wc = &response->header.wc;
2087                 response->header.recv_wc.recv_buf.mad = (struct ib_mad *)response->mad;
2088                 response->header.recv_wc.recv_buf.grh = &response->grh;
2089                 resp_hdr->method = IB_MGMT_METHOD_GET_RESP;
2090                 resp_hdr->status = cpu_to_be16(IB_MGMT_MAD_STATUS_UNSUPPORTED_METHOD_ATTRIB);
2091                 if (recv_hdr->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
2092                         resp_hdr->status |= IB_SMP_DIRECTION;
2093
2094                 if (opa && recv_hdr->base_version == OPA_MGMT_BASE_VERSION) {
2095                         if (recv_hdr->mgmt_class ==
2096                             IB_MGMT_CLASS_SUBN_LID_ROUTED ||
2097                             recv_hdr->mgmt_class ==
2098                             IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
2099                                 *resp_len = opa_get_smp_header_size(
2100                                                         (const struct opa_smp *)recv->mad);
2101                         else
2102                                 *resp_len = sizeof(struct ib_mad_hdr);
2103                 }
2104
2105                 return true;
2106         } else {
2107                 return false;
2108         }
2109 }
2110
2111 static enum smi_action
2112 handle_opa_smi(struct ib_mad_port_private *port_priv,
2113                struct ib_mad_qp_info *qp_info,
2114                struct ib_wc *wc,
2115                int port_num,
2116                struct ib_mad_private *recv,
2117                struct ib_mad_private *response)
2118 {
2119         enum smi_forward_action retsmi;
2120         struct opa_smp *smp = (struct opa_smp *)recv->mad;
2121
2122         if (opa_smi_handle_dr_smp_recv(smp,
2123                                    rdma_cap_ib_switch(port_priv->device),
2124                                    port_num,
2125                                    port_priv->device->phys_port_cnt) ==
2126                                    IB_SMI_DISCARD)
2127                 return IB_SMI_DISCARD;
2128
2129         retsmi = opa_smi_check_forward_dr_smp(smp);
2130         if (retsmi == IB_SMI_LOCAL)
2131                 return IB_SMI_HANDLE;
2132
2133         if (retsmi == IB_SMI_SEND) { /* don't forward */
2134                 if (opa_smi_handle_dr_smp_send(smp,
2135                                            rdma_cap_ib_switch(port_priv->device),
2136                                            port_num) == IB_SMI_DISCARD)
2137                         return IB_SMI_DISCARD;
2138
2139                 if (opa_smi_check_local_smp(smp, port_priv->device) ==
2140                     IB_SMI_DISCARD)
2141                         return IB_SMI_DISCARD;
2142
2143         } else if (rdma_cap_ib_switch(port_priv->device)) {
2144                 /* forward case for switches */
2145                 memcpy(response, recv, mad_priv_size(response));
2146                 response->header.recv_wc.wc = &response->header.wc;
2147                 response->header.recv_wc.recv_buf.opa_mad =
2148                                 (struct opa_mad *)response->mad;
2149                 response->header.recv_wc.recv_buf.grh = &response->grh;
2150
2151                 agent_send_response((const struct ib_mad_hdr *)response->mad,
2152                                     &response->grh, wc,
2153                                     port_priv->device,
2154                                     opa_smi_get_fwd_port(smp),
2155                                     qp_info->qp->qp_num,
2156                                     recv->header.wc.byte_len,
2157                                     true);
2158
2159                 return IB_SMI_DISCARD;
2160         }
2161
2162         return IB_SMI_HANDLE;
2163 }
2164
2165 static enum smi_action
2166 handle_smi(struct ib_mad_port_private *port_priv,
2167            struct ib_mad_qp_info *qp_info,
2168            struct ib_wc *wc,
2169            int port_num,
2170            struct ib_mad_private *recv,
2171            struct ib_mad_private *response,
2172            bool opa)
2173 {
2174         struct ib_mad_hdr *mad_hdr = (struct ib_mad_hdr *)recv->mad;
2175
2176         if (opa && mad_hdr->base_version == OPA_MGMT_BASE_VERSION &&
2177             mad_hdr->class_version == OPA_SMI_CLASS_VERSION)
2178                 return handle_opa_smi(port_priv, qp_info, wc, port_num, recv,
2179                                       response);
2180
2181         return handle_ib_smi(port_priv, qp_info, wc, port_num, recv, response);
2182 }
2183
2184 static void ib_mad_recv_done(struct ib_cq *cq, struct ib_wc *wc)
2185 {
2186         struct ib_mad_port_private *port_priv = cq->cq_context;
2187         struct ib_mad_list_head *mad_list =
2188                 container_of(wc->wr_cqe, struct ib_mad_list_head, cqe);
2189         struct ib_mad_qp_info *qp_info;
2190         struct ib_mad_private_header *mad_priv_hdr;
2191         struct ib_mad_private *recv, *response = NULL;
2192         struct ib_mad_agent_private *mad_agent;
2193         int port_num;
2194         int ret = IB_MAD_RESULT_SUCCESS;
2195         size_t mad_size;
2196         u16 resp_mad_pkey_index = 0;
2197         bool opa;
2198
2199         if (list_empty_careful(&port_priv->port_list))
2200                 return;
2201
2202         if (wc->status != IB_WC_SUCCESS) {
2203                 /*
2204                  * Receive errors indicate that the QP has entered the error
2205                  * state - error handling/shutdown code will cleanup
2206                  */
2207                 return;
2208         }
2209
2210         qp_info = mad_list->mad_queue->qp_info;
2211         dequeue_mad(mad_list);
2212
2213         opa = rdma_cap_opa_mad(qp_info->port_priv->device,
2214                                qp_info->port_priv->port_num);
2215
2216         mad_priv_hdr = container_of(mad_list, struct ib_mad_private_header,
2217                                     mad_list);
2218         recv = container_of(mad_priv_hdr, struct ib_mad_private, header);
2219         ib_dma_unmap_single(port_priv->device,
2220                             recv->header.mapping,
2221                             mad_priv_dma_size(recv),
2222                             DMA_FROM_DEVICE);
2223
2224         /* Setup MAD receive work completion from "normal" work completion */
2225         recv->header.wc = *wc;
2226         recv->header.recv_wc.wc = &recv->header.wc;
2227
2228         if (opa && ((struct ib_mad_hdr *)(recv->mad))->base_version == OPA_MGMT_BASE_VERSION) {
2229                 recv->header.recv_wc.mad_len = wc->byte_len - sizeof(struct ib_grh);
2230                 recv->header.recv_wc.mad_seg_size = sizeof(struct opa_mad);
2231         } else {
2232                 recv->header.recv_wc.mad_len = sizeof(struct ib_mad);
2233                 recv->header.recv_wc.mad_seg_size = sizeof(struct ib_mad);
2234         }
2235
2236         recv->header.recv_wc.recv_buf.mad = (struct ib_mad *)recv->mad;
2237         recv->header.recv_wc.recv_buf.grh = &recv->grh;
2238
2239         if (atomic_read(&qp_info->snoop_count))
2240                 snoop_recv(qp_info, &recv->header.recv_wc, IB_MAD_SNOOP_RECVS);
2241
2242         /* Validate MAD */
2243         if (!validate_mad((const struct ib_mad_hdr *)recv->mad, qp_info, opa))
2244                 goto out;
2245
2246         mad_size = recv->mad_size;
2247         response = alloc_mad_private(mad_size, GFP_KERNEL);
2248         if (!response) {
2249                 dev_err(&port_priv->device->dev,
2250                         "%s: no memory for response buffer\n", __func__);
2251                 goto out;
2252         }
2253
2254         if (rdma_cap_ib_switch(port_priv->device))
2255                 port_num = wc->port_num;
2256         else
2257                 port_num = port_priv->port_num;
2258
2259         if (((struct ib_mad_hdr *)recv->mad)->mgmt_class ==
2260             IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
2261                 if (handle_smi(port_priv, qp_info, wc, port_num, recv,
2262                                response, opa)
2263                     == IB_SMI_DISCARD)
2264                         goto out;
2265         }
2266
2267         /* Give driver "right of first refusal" on incoming MAD */
2268         if (port_priv->device->process_mad) {
2269                 ret = port_priv->device->process_mad(port_priv->device, 0,
2270                                                      port_priv->port_num,
2271                                                      wc, &recv->grh,
2272                                                      (const struct ib_mad_hdr *)recv->mad,
2273                                                      recv->mad_size,
2274                                                      (struct ib_mad_hdr *)response->mad,
2275                                                      &mad_size, &resp_mad_pkey_index);
2276
2277                 if (opa)
2278                         wc->pkey_index = resp_mad_pkey_index;
2279
2280                 if (ret & IB_MAD_RESULT_SUCCESS) {
2281                         if (ret & IB_MAD_RESULT_CONSUMED)
2282                                 goto out;
2283                         if (ret & IB_MAD_RESULT_REPLY) {
2284                                 agent_send_response((const struct ib_mad_hdr *)response->mad,
2285                                                     &recv->grh, wc,
2286                                                     port_priv->device,
2287                                                     port_num,
2288                                                     qp_info->qp->qp_num,
2289                                                     mad_size, opa);
2290                                 goto out;
2291                         }
2292                 }
2293         }
2294
2295         mad_agent = find_mad_agent(port_priv, (const struct ib_mad_hdr *)recv->mad);
2296         if (mad_agent) {
2297                 ib_mad_complete_recv(mad_agent, &recv->header.recv_wc);
2298                 /*
2299                  * recv is freed up in error cases in ib_mad_complete_recv
2300                  * or via recv_handler in ib_mad_complete_recv()
2301                  */
2302                 recv = NULL;
2303         } else if ((ret & IB_MAD_RESULT_SUCCESS) &&
2304                    generate_unmatched_resp(recv, response, &mad_size, opa)) {
2305                 agent_send_response((const struct ib_mad_hdr *)response->mad, &recv->grh, wc,
2306                                     port_priv->device, port_num,
2307                                     qp_info->qp->qp_num, mad_size, opa);
2308         }
2309
2310 out:
2311         /* Post another receive request for this QP */
2312         if (response) {
2313                 ib_mad_post_receive_mads(qp_info, response);
2314                 kfree(recv);
2315         } else
2316                 ib_mad_post_receive_mads(qp_info, recv);
2317 }
2318
2319 static void adjust_timeout(struct ib_mad_agent_private *mad_agent_priv)
2320 {
2321         struct ib_mad_send_wr_private *mad_send_wr;
2322         unsigned long delay;
2323
2324         if (list_empty(&mad_agent_priv->wait_list)) {
2325                 cancel_delayed_work(&mad_agent_priv->timed_work);
2326         } else {
2327                 mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
2328                                          struct ib_mad_send_wr_private,
2329                                          agent_list);
2330
2331                 if (time_after(mad_agent_priv->timeout,
2332                                mad_send_wr->timeout)) {
2333                         mad_agent_priv->timeout = mad_send_wr->timeout;
2334                         delay = mad_send_wr->timeout - jiffies;
2335                         if ((long)delay <= 0)
2336                                 delay = 1;
2337                         mod_delayed_work(mad_agent_priv->qp_info->port_priv->wq,
2338                                          &mad_agent_priv->timed_work, delay);
2339                 }
2340         }
2341 }
2342
2343 static void wait_for_response(struct ib_mad_send_wr_private *mad_send_wr)
2344 {
2345         struct ib_mad_agent_private *mad_agent_priv;
2346         struct ib_mad_send_wr_private *temp_mad_send_wr;
2347         struct list_head *list_item;
2348         unsigned long delay;
2349
2350         mad_agent_priv = mad_send_wr->mad_agent_priv;
2351         list_del(&mad_send_wr->agent_list);
2352
2353         delay = mad_send_wr->timeout;
2354         mad_send_wr->timeout += jiffies;
2355
2356         if (delay) {
2357                 list_for_each_prev(list_item, &mad_agent_priv->wait_list) {
2358                         temp_mad_send_wr = list_entry(list_item,
2359                                                 struct ib_mad_send_wr_private,
2360                                                 agent_list);
2361                         if (time_after(mad_send_wr->timeout,
2362                                        temp_mad_send_wr->timeout))
2363                                 break;
2364                 }
2365         }
2366         else
2367                 list_item = &mad_agent_priv->wait_list;
2368         list_add(&mad_send_wr->agent_list, list_item);
2369
2370         /* Reschedule a work item if we have a shorter timeout */
2371         if (mad_agent_priv->wait_list.next == &mad_send_wr->agent_list)
2372                 mod_delayed_work(mad_agent_priv->qp_info->port_priv->wq,
2373                                  &mad_agent_priv->timed_work, delay);
2374 }
2375
2376 void ib_reset_mad_timeout(struct ib_mad_send_wr_private *mad_send_wr,
2377                           int timeout_ms)
2378 {
2379         mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2380         wait_for_response(mad_send_wr);
2381 }
2382
2383 /*
2384  * Process a send work completion
2385  */
2386 void ib_mad_complete_send_wr(struct ib_mad_send_wr_private *mad_send_wr,
2387                              struct ib_mad_send_wc *mad_send_wc)
2388 {
2389         struct ib_mad_agent_private     *mad_agent_priv;
2390         unsigned long                   flags;
2391         int                             ret;
2392
2393         mad_agent_priv = mad_send_wr->mad_agent_priv;
2394         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2395         if (ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)) {
2396                 ret = ib_process_rmpp_send_wc(mad_send_wr, mad_send_wc);
2397                 if (ret == IB_RMPP_RESULT_CONSUMED)
2398                         goto done;
2399         } else
2400                 ret = IB_RMPP_RESULT_UNHANDLED;
2401
2402         if (mad_send_wc->status != IB_WC_SUCCESS &&
2403             mad_send_wr->status == IB_WC_SUCCESS) {
2404                 mad_send_wr->status = mad_send_wc->status;
2405                 mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2406         }
2407
2408         if (--mad_send_wr->refcount > 0) {
2409                 if (mad_send_wr->refcount == 1 && mad_send_wr->timeout &&
2410                     mad_send_wr->status == IB_WC_SUCCESS) {
2411                         wait_for_response(mad_send_wr);
2412                 }
2413                 goto done;
2414         }
2415
2416         /* Remove send from MAD agent and notify client of completion */
2417         list_del(&mad_send_wr->agent_list);
2418         adjust_timeout(mad_agent_priv);
2419         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2420
2421         if (mad_send_wr->status != IB_WC_SUCCESS )
2422                 mad_send_wc->status = mad_send_wr->status;
2423         if (ret == IB_RMPP_RESULT_INTERNAL)
2424                 ib_rmpp_send_handler(mad_send_wc);
2425         else
2426                 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2427                                                    mad_send_wc);
2428
2429         /* Release reference on agent taken when sending */
2430         deref_mad_agent(mad_agent_priv);
2431         return;
2432 done:
2433         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2434 }
2435
2436 static void ib_mad_send_done(struct ib_cq *cq, struct ib_wc *wc)
2437 {
2438         struct ib_mad_port_private *port_priv = cq->cq_context;
2439         struct ib_mad_list_head *mad_list =
2440                 container_of(wc->wr_cqe, struct ib_mad_list_head, cqe);
2441         struct ib_mad_send_wr_private   *mad_send_wr, *queued_send_wr;
2442         struct ib_mad_qp_info           *qp_info;
2443         struct ib_mad_queue             *send_queue;
2444         struct ib_send_wr               *bad_send_wr;
2445         struct ib_mad_send_wc           mad_send_wc;
2446         unsigned long flags;
2447         int ret;
2448
2449         if (list_empty_careful(&port_priv->port_list))
2450                 return;
2451
2452         if (wc->status != IB_WC_SUCCESS) {
2453                 if (!ib_mad_send_error(port_priv, wc))
2454                         return;
2455         }
2456
2457         mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2458                                    mad_list);
2459         send_queue = mad_list->mad_queue;
2460         qp_info = send_queue->qp_info;
2461
2462 retry:
2463         ib_dma_unmap_single(mad_send_wr->send_buf.mad_agent->device,
2464                             mad_send_wr->header_mapping,
2465                             mad_send_wr->sg_list[0].length, DMA_TO_DEVICE);
2466         ib_dma_unmap_single(mad_send_wr->send_buf.mad_agent->device,
2467                             mad_send_wr->payload_mapping,
2468                             mad_send_wr->sg_list[1].length, DMA_TO_DEVICE);
2469         queued_send_wr = NULL;
2470         spin_lock_irqsave(&send_queue->lock, flags);
2471         list_del(&mad_list->list);
2472
2473         /* Move queued send to the send queue */
2474         if (send_queue->count-- > send_queue->max_active) {
2475                 mad_list = container_of(qp_info->overflow_list.next,
2476                                         struct ib_mad_list_head, list);
2477                 queued_send_wr = container_of(mad_list,
2478                                         struct ib_mad_send_wr_private,
2479                                         mad_list);
2480                 list_move_tail(&mad_list->list, &send_queue->list);
2481         }
2482         spin_unlock_irqrestore(&send_queue->lock, flags);
2483
2484         mad_send_wc.send_buf = &mad_send_wr->send_buf;
2485         mad_send_wc.status = wc->status;
2486         mad_send_wc.vendor_err = wc->vendor_err;
2487         if (atomic_read(&qp_info->snoop_count))
2488                 snoop_send(qp_info, &mad_send_wr->send_buf, &mad_send_wc,
2489                            IB_MAD_SNOOP_SEND_COMPLETIONS);
2490         ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
2491
2492         if (queued_send_wr) {
2493                 ret = ib_post_send(qp_info->qp, &queued_send_wr->send_wr.wr,
2494                                    &bad_send_wr);
2495                 if (ret) {
2496                         dev_err(&port_priv->device->dev,
2497                                 "ib_post_send failed: %d\n", ret);
2498                         mad_send_wr = queued_send_wr;
2499                         wc->status = IB_WC_LOC_QP_OP_ERR;
2500                         goto retry;
2501                 }
2502         }
2503 }
2504
2505 static void mark_sends_for_retry(struct ib_mad_qp_info *qp_info)
2506 {
2507         struct ib_mad_send_wr_private *mad_send_wr;
2508         struct ib_mad_list_head *mad_list;
2509         unsigned long flags;
2510
2511         spin_lock_irqsave(&qp_info->send_queue.lock, flags);
2512         list_for_each_entry(mad_list, &qp_info->send_queue.list, list) {
2513                 mad_send_wr = container_of(mad_list,
2514                                            struct ib_mad_send_wr_private,
2515                                            mad_list);
2516                 mad_send_wr->retry = 1;
2517         }
2518         spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
2519 }
2520
2521 static bool ib_mad_send_error(struct ib_mad_port_private *port_priv,
2522                 struct ib_wc *wc)
2523 {
2524         struct ib_mad_list_head *mad_list =
2525                 container_of(wc->wr_cqe, struct ib_mad_list_head, cqe);
2526         struct ib_mad_qp_info *qp_info = mad_list->mad_queue->qp_info;
2527         struct ib_mad_send_wr_private *mad_send_wr;
2528         int ret;
2529
2530         /*
2531          * Send errors will transition the QP to SQE - move
2532          * QP to RTS and repost flushed work requests
2533          */
2534         mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2535                                    mad_list);
2536         if (wc->status == IB_WC_WR_FLUSH_ERR) {
2537                 if (mad_send_wr->retry) {
2538                         /* Repost send */
2539                         struct ib_send_wr *bad_send_wr;
2540
2541                         mad_send_wr->retry = 0;
2542                         ret = ib_post_send(qp_info->qp, &mad_send_wr->send_wr.wr,
2543                                         &bad_send_wr);
2544                         if (!ret)
2545                                 return false;
2546                 }
2547         } else {
2548                 struct ib_qp_attr *attr;
2549
2550                 /* Transition QP to RTS and fail offending send */
2551                 attr = kmalloc(sizeof *attr, GFP_KERNEL);
2552                 if (attr) {
2553                         attr->qp_state = IB_QPS_RTS;
2554                         attr->cur_qp_state = IB_QPS_SQE;
2555                         ret = ib_modify_qp(qp_info->qp, attr,
2556                                            IB_QP_STATE | IB_QP_CUR_STATE);
2557                         kfree(attr);
2558                         if (ret)
2559                                 dev_err(&port_priv->device->dev,
2560                                         "%s - ib_modify_qp to RTS: %d\n",
2561                                         __func__, ret);
2562                         else
2563                                 mark_sends_for_retry(qp_info);
2564                 }
2565         }
2566
2567         return true;
2568 }
2569
2570 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv)
2571 {
2572         unsigned long flags;
2573         struct ib_mad_send_wr_private *mad_send_wr, *temp_mad_send_wr;
2574         struct ib_mad_send_wc mad_send_wc;
2575         struct list_head cancel_list;
2576
2577         INIT_LIST_HEAD(&cancel_list);
2578
2579         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2580         list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2581                                  &mad_agent_priv->send_list, agent_list) {
2582                 if (mad_send_wr->status == IB_WC_SUCCESS) {
2583                         mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2584                         mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2585                 }
2586         }
2587
2588         /* Empty wait list to prevent receives from finding a request */
2589         list_splice_init(&mad_agent_priv->wait_list, &cancel_list);
2590         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2591
2592         /* Report all cancelled requests */
2593         mad_send_wc.status = IB_WC_WR_FLUSH_ERR;
2594         mad_send_wc.vendor_err = 0;
2595
2596         list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2597                                  &cancel_list, agent_list) {
2598                 mad_send_wc.send_buf = &mad_send_wr->send_buf;
2599                 list_del(&mad_send_wr->agent_list);
2600                 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2601                                                    &mad_send_wc);
2602                 atomic_dec(&mad_agent_priv->refcount);
2603         }
2604 }
2605
2606 static struct ib_mad_send_wr_private*
2607 find_send_wr(struct ib_mad_agent_private *mad_agent_priv,
2608              struct ib_mad_send_buf *send_buf)
2609 {
2610         struct ib_mad_send_wr_private *mad_send_wr;
2611
2612         list_for_each_entry(mad_send_wr, &mad_agent_priv->wait_list,
2613                             agent_list) {
2614                 if (&mad_send_wr->send_buf == send_buf)
2615                         return mad_send_wr;
2616         }
2617
2618         list_for_each_entry(mad_send_wr, &mad_agent_priv->send_list,
2619                             agent_list) {
2620                 if (is_rmpp_data_mad(mad_agent_priv,
2621                                      mad_send_wr->send_buf.mad) &&
2622                     &mad_send_wr->send_buf == send_buf)
2623                         return mad_send_wr;
2624         }
2625         return NULL;
2626 }
2627
2628 int ib_modify_mad(struct ib_mad_agent *mad_agent,
2629                   struct ib_mad_send_buf *send_buf, u32 timeout_ms)
2630 {
2631         struct ib_mad_agent_private *mad_agent_priv;
2632         struct ib_mad_send_wr_private *mad_send_wr;
2633         unsigned long flags;
2634         int active;
2635
2636         mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
2637                                       agent);
2638         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2639         mad_send_wr = find_send_wr(mad_agent_priv, send_buf);
2640         if (!mad_send_wr || mad_send_wr->status != IB_WC_SUCCESS) {
2641                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2642                 return -EINVAL;
2643         }
2644
2645         active = (!mad_send_wr->timeout || mad_send_wr->refcount > 1);
2646         if (!timeout_ms) {
2647                 mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2648                 mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2649         }
2650
2651         mad_send_wr->send_buf.timeout_ms = timeout_ms;
2652         if (active)
2653                 mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2654         else
2655                 ib_reset_mad_timeout(mad_send_wr, timeout_ms);
2656
2657         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2658         return 0;
2659 }
2660 EXPORT_SYMBOL(ib_modify_mad);
2661
2662 void ib_cancel_mad(struct ib_mad_agent *mad_agent,
2663                    struct ib_mad_send_buf *send_buf)
2664 {
2665         ib_modify_mad(mad_agent, send_buf, 0);
2666 }
2667 EXPORT_SYMBOL(ib_cancel_mad);
2668
2669 static void local_completions(struct work_struct *work)
2670 {
2671         struct ib_mad_agent_private *mad_agent_priv;
2672         struct ib_mad_local_private *local;
2673         struct ib_mad_agent_private *recv_mad_agent;
2674         unsigned long flags;
2675         int free_mad;
2676         struct ib_wc wc;
2677         struct ib_mad_send_wc mad_send_wc;
2678         bool opa;
2679
2680         mad_agent_priv =
2681                 container_of(work, struct ib_mad_agent_private, local_work);
2682
2683         opa = rdma_cap_opa_mad(mad_agent_priv->qp_info->port_priv->device,
2684                                mad_agent_priv->qp_info->port_priv->port_num);
2685
2686         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2687         while (!list_empty(&mad_agent_priv->local_list)) {
2688                 local = list_entry(mad_agent_priv->local_list.next,
2689                                    struct ib_mad_local_private,
2690                                    completion_list);
2691                 list_del(&local->completion_list);
2692                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2693                 free_mad = 0;
2694                 if (local->mad_priv) {
2695                         u8 base_version;
2696                         recv_mad_agent = local->recv_mad_agent;
2697                         if (!recv_mad_agent) {
2698                                 dev_err(&mad_agent_priv->agent.device->dev,
2699                                         "No receive MAD agent for local completion\n");
2700                                 free_mad = 1;
2701                                 goto local_send_completion;
2702                         }
2703
2704                         /*
2705                          * Defined behavior is to complete response
2706                          * before request
2707                          */
2708                         build_smp_wc(recv_mad_agent->agent.qp,
2709                                      local->mad_send_wr->send_wr.wr.wr_cqe,
2710                                      be16_to_cpu(IB_LID_PERMISSIVE),
2711                                      local->mad_send_wr->send_wr.pkey_index,
2712                                      recv_mad_agent->agent.port_num, &wc);
2713
2714                         local->mad_priv->header.recv_wc.wc = &wc;
2715
2716                         base_version = ((struct ib_mad_hdr *)(local->mad_priv->mad))->base_version;
2717                         if (opa && base_version == OPA_MGMT_BASE_VERSION) {
2718                                 local->mad_priv->header.recv_wc.mad_len = local->return_wc_byte_len;
2719                                 local->mad_priv->header.recv_wc.mad_seg_size = sizeof(struct opa_mad);
2720                         } else {
2721                                 local->mad_priv->header.recv_wc.mad_len = sizeof(struct ib_mad);
2722                                 local->mad_priv->header.recv_wc.mad_seg_size = sizeof(struct ib_mad);
2723                         }
2724
2725                         INIT_LIST_HEAD(&local->mad_priv->header.recv_wc.rmpp_list);
2726                         list_add(&local->mad_priv->header.recv_wc.recv_buf.list,
2727                                  &local->mad_priv->header.recv_wc.rmpp_list);
2728                         local->mad_priv->header.recv_wc.recv_buf.grh = NULL;
2729                         local->mad_priv->header.recv_wc.recv_buf.mad =
2730                                                 (struct ib_mad *)local->mad_priv->mad;
2731                         if (atomic_read(&recv_mad_agent->qp_info->snoop_count))
2732                                 snoop_recv(recv_mad_agent->qp_info,
2733                                           &local->mad_priv->header.recv_wc,
2734                                            IB_MAD_SNOOP_RECVS);
2735                         recv_mad_agent->agent.recv_handler(
2736                                                 &recv_mad_agent->agent,
2737                                                 &local->mad_send_wr->send_buf,
2738                                                 &local->mad_priv->header.recv_wc);
2739                         spin_lock_irqsave(&recv_mad_agent->lock, flags);
2740                         atomic_dec(&recv_mad_agent->refcount);
2741                         spin_unlock_irqrestore(&recv_mad_agent->lock, flags);
2742                 }
2743
2744 local_send_completion:
2745                 /* Complete send */
2746                 mad_send_wc.status = IB_WC_SUCCESS;
2747                 mad_send_wc.vendor_err = 0;
2748                 mad_send_wc.send_buf = &local->mad_send_wr->send_buf;
2749                 if (atomic_read(&mad_agent_priv->qp_info->snoop_count))
2750                         snoop_send(mad_agent_priv->qp_info,
2751                                    &local->mad_send_wr->send_buf,
2752                                    &mad_send_wc, IB_MAD_SNOOP_SEND_COMPLETIONS);
2753                 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2754                                                    &mad_send_wc);
2755
2756                 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2757                 atomic_dec(&mad_agent_priv->refcount);
2758                 if (free_mad)
2759                         kfree(local->mad_priv);
2760                 kfree(local);
2761         }
2762         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2763 }
2764
2765 static int retry_send(struct ib_mad_send_wr_private *mad_send_wr)
2766 {
2767         int ret;
2768
2769         if (!mad_send_wr->retries_left)
2770                 return -ETIMEDOUT;
2771
2772         mad_send_wr->retries_left--;
2773         mad_send_wr->send_buf.retries++;
2774
2775         mad_send_wr->timeout = msecs_to_jiffies(mad_send_wr->send_buf.timeout_ms);
2776
2777         if (ib_mad_kernel_rmpp_agent(&mad_send_wr->mad_agent_priv->agent)) {
2778                 ret = ib_retry_rmpp(mad_send_wr);
2779                 switch (ret) {
2780                 case IB_RMPP_RESULT_UNHANDLED:
2781                         ret = ib_send_mad(mad_send_wr);
2782                         break;
2783                 case IB_RMPP_RESULT_CONSUMED:
2784                         ret = 0;
2785                         break;
2786                 default:
2787                         ret = -ECOMM;
2788                         break;
2789                 }
2790         } else
2791                 ret = ib_send_mad(mad_send_wr);
2792
2793         if (!ret) {
2794                 mad_send_wr->refcount++;
2795                 list_add_tail(&mad_send_wr->agent_list,
2796                               &mad_send_wr->mad_agent_priv->send_list);
2797         }
2798         return ret;
2799 }
2800
2801 static void timeout_sends(struct work_struct *work)
2802 {
2803         struct ib_mad_agent_private *mad_agent_priv;
2804         struct ib_mad_send_wr_private *mad_send_wr;
2805         struct ib_mad_send_wc mad_send_wc;
2806         unsigned long flags, delay;
2807
2808         mad_agent_priv = container_of(work, struct ib_mad_agent_private,
2809                                       timed_work.work);
2810         mad_send_wc.vendor_err = 0;
2811
2812         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2813         while (!list_empty(&mad_agent_priv->wait_list)) {
2814                 mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
2815                                          struct ib_mad_send_wr_private,
2816                                          agent_list);
2817
2818                 if (time_after(mad_send_wr->timeout, jiffies)) {
2819                         delay = mad_send_wr->timeout - jiffies;
2820                         if ((long)delay <= 0)
2821                                 delay = 1;
2822                         queue_delayed_work(mad_agent_priv->qp_info->
2823                                            port_priv->wq,
2824                                            &mad_agent_priv->timed_work, delay);
2825                         break;
2826                 }
2827
2828                 list_del(&mad_send_wr->agent_list);
2829                 if (mad_send_wr->status == IB_WC_SUCCESS &&
2830                     !retry_send(mad_send_wr))
2831                         continue;
2832
2833                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2834
2835                 if (mad_send_wr->status == IB_WC_SUCCESS)
2836                         mad_send_wc.status = IB_WC_RESP_TIMEOUT_ERR;
2837                 else
2838                         mad_send_wc.status = mad_send_wr->status;
2839                 mad_send_wc.send_buf = &mad_send_wr->send_buf;
2840                 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2841                                                    &mad_send_wc);
2842
2843                 atomic_dec(&mad_agent_priv->refcount);
2844                 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2845         }
2846         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2847 }
2848
2849 /*
2850  * Allocate receive MADs and post receive WRs for them
2851  */
2852 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
2853                                     struct ib_mad_private *mad)
2854 {
2855         unsigned long flags;
2856         int post, ret;
2857         struct ib_mad_private *mad_priv;
2858         struct ib_sge sg_list;
2859         struct ib_recv_wr recv_wr, *bad_recv_wr;
2860         struct ib_mad_queue *recv_queue = &qp_info->recv_queue;
2861
2862         /* Initialize common scatter list fields */
2863         sg_list.lkey = qp_info->port_priv->pd->local_dma_lkey;
2864
2865         /* Initialize common receive WR fields */
2866         recv_wr.next = NULL;
2867         recv_wr.sg_list = &sg_list;
2868         recv_wr.num_sge = 1;
2869
2870         do {
2871                 /* Allocate and map receive buffer */
2872                 if (mad) {
2873                         mad_priv = mad;
2874                         mad = NULL;
2875                 } else {
2876                         mad_priv = alloc_mad_private(port_mad_size(qp_info->port_priv),
2877                                                      GFP_ATOMIC);
2878                         if (!mad_priv) {
2879                                 dev_err(&qp_info->port_priv->device->dev,
2880                                         "No memory for receive buffer\n");
2881                                 ret = -ENOMEM;
2882                                 break;
2883                         }
2884                 }
2885                 sg_list.length = mad_priv_dma_size(mad_priv);
2886                 sg_list.addr = ib_dma_map_single(qp_info->port_priv->device,
2887                                                  &mad_priv->grh,
2888                                                  mad_priv_dma_size(mad_priv),
2889                                                  DMA_FROM_DEVICE);
2890                 if (unlikely(ib_dma_mapping_error(qp_info->port_priv->device,
2891                                                   sg_list.addr))) {
2892                         ret = -ENOMEM;
2893                         break;
2894                 }
2895                 mad_priv->header.mapping = sg_list.addr;
2896                 mad_priv->header.mad_list.mad_queue = recv_queue;
2897                 mad_priv->header.mad_list.cqe.done = ib_mad_recv_done;
2898                 recv_wr.wr_cqe = &mad_priv->header.mad_list.cqe;
2899
2900                 /* Post receive WR */
2901                 spin_lock_irqsave(&recv_queue->lock, flags);
2902                 post = (++recv_queue->count < recv_queue->max_active);
2903                 list_add_tail(&mad_priv->header.mad_list.list, &recv_queue->list);
2904                 spin_unlock_irqrestore(&recv_queue->lock, flags);
2905                 ret = ib_post_recv(qp_info->qp, &recv_wr, &bad_recv_wr);
2906                 if (ret) {
2907                         spin_lock_irqsave(&recv_queue->lock, flags);
2908                         list_del(&mad_priv->header.mad_list.list);
2909                         recv_queue->count--;
2910                         spin_unlock_irqrestore(&recv_queue->lock, flags);
2911                         ib_dma_unmap_single(qp_info->port_priv->device,
2912                                             mad_priv->header.mapping,
2913                                             mad_priv_dma_size(mad_priv),
2914                                             DMA_FROM_DEVICE);
2915                         kfree(mad_priv);
2916                         dev_err(&qp_info->port_priv->device->dev,
2917                                 "ib_post_recv failed: %d\n", ret);
2918                         break;
2919                 }
2920         } while (post);
2921
2922         return ret;
2923 }
2924
2925 /*
2926  * Return all the posted receive MADs
2927  */
2928 static void cleanup_recv_queue(struct ib_mad_qp_info *qp_info)
2929 {
2930         struct ib_mad_private_header *mad_priv_hdr;
2931         struct ib_mad_private *recv;
2932         struct ib_mad_list_head *mad_list;
2933
2934         if (!qp_info->qp)
2935                 return;
2936
2937         while (!list_empty(&qp_info->recv_queue.list)) {
2938
2939                 mad_list = list_entry(qp_info->recv_queue.list.next,
2940                                       struct ib_mad_list_head, list);
2941                 mad_priv_hdr = container_of(mad_list,
2942                                             struct ib_mad_private_header,
2943                                             mad_list);
2944                 recv = container_of(mad_priv_hdr, struct ib_mad_private,
2945                                     header);
2946
2947                 /* Remove from posted receive MAD list */
2948                 list_del(&mad_list->list);
2949
2950                 ib_dma_unmap_single(qp_info->port_priv->device,
2951                                     recv->header.mapping,
2952                                     mad_priv_dma_size(recv),
2953                                     DMA_FROM_DEVICE);
2954                 kfree(recv);
2955         }
2956
2957         qp_info->recv_queue.count = 0;
2958 }
2959
2960 /*
2961  * Start the port
2962  */
2963 static int ib_mad_port_start(struct ib_mad_port_private *port_priv)
2964 {
2965         int ret, i;
2966         struct ib_qp_attr *attr;
2967         struct ib_qp *qp;
2968         u16 pkey_index;
2969
2970         attr = kmalloc(sizeof *attr, GFP_KERNEL);
2971         if (!attr) {
2972                 dev_err(&port_priv->device->dev,
2973                         "Couldn't kmalloc ib_qp_attr\n");
2974                 return -ENOMEM;
2975         }
2976
2977         ret = ib_find_pkey(port_priv->device, port_priv->port_num,
2978                            IB_DEFAULT_PKEY_FULL, &pkey_index);
2979         if (ret)
2980                 pkey_index = 0;
2981
2982         for (i = 0; i < IB_MAD_QPS_CORE; i++) {
2983                 qp = port_priv->qp_info[i].qp;
2984                 if (!qp)
2985                         continue;
2986
2987                 /*
2988                  * PKey index for QP1 is irrelevant but
2989                  * one is needed for the Reset to Init transition
2990                  */
2991                 attr->qp_state = IB_QPS_INIT;
2992                 attr->pkey_index = pkey_index;
2993                 attr->qkey = (qp->qp_num == 0) ? 0 : IB_QP1_QKEY;
2994                 ret = ib_modify_qp(qp, attr, IB_QP_STATE |
2995                                              IB_QP_PKEY_INDEX | IB_QP_QKEY);
2996                 if (ret) {
2997                         dev_err(&port_priv->device->dev,
2998                                 "Couldn't change QP%d state to INIT: %d\n",
2999                                 i, ret);
3000                         goto out;
3001                 }
3002
3003                 attr->qp_state = IB_QPS_RTR;
3004                 ret = ib_modify_qp(qp, attr, IB_QP_STATE);
3005                 if (ret) {
3006                         dev_err(&port_priv->device->dev,
3007                                 "Couldn't change QP%d state to RTR: %d\n",
3008                                 i, ret);
3009                         goto out;
3010                 }
3011
3012                 attr->qp_state = IB_QPS_RTS;
3013                 attr->sq_psn = IB_MAD_SEND_Q_PSN;
3014                 ret = ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_SQ_PSN);
3015                 if (ret) {
3016                         dev_err(&port_priv->device->dev,
3017                                 "Couldn't change QP%d state to RTS: %d\n",
3018                                 i, ret);
3019                         goto out;
3020                 }
3021         }
3022
3023         ret = ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP);
3024         if (ret) {
3025                 dev_err(&port_priv->device->dev,
3026                         "Failed to request completion notification: %d\n",
3027                         ret);
3028                 goto out;
3029         }
3030
3031         for (i = 0; i < IB_MAD_QPS_CORE; i++) {
3032                 if (!port_priv->qp_info[i].qp)
3033                         continue;
3034
3035                 ret = ib_mad_post_receive_mads(&port_priv->qp_info[i], NULL);
3036                 if (ret) {
3037                         dev_err(&port_priv->device->dev,
3038                                 "Couldn't post receive WRs\n");
3039                         goto out;
3040                 }
3041         }
3042 out:
3043         kfree(attr);
3044         return ret;
3045 }
3046
3047 static void qp_event_handler(struct ib_event *event, void *qp_context)
3048 {
3049         struct ib_mad_qp_info   *qp_info = qp_context;
3050
3051         /* It's worse than that! He's dead, Jim! */
3052         dev_err(&qp_info->port_priv->device->dev,
3053                 "Fatal error (%d) on MAD QP (%d)\n",
3054                 event->event, qp_info->qp->qp_num);
3055 }
3056
3057 static void init_mad_queue(struct ib_mad_qp_info *qp_info,
3058                            struct ib_mad_queue *mad_queue)
3059 {
3060         mad_queue->qp_info = qp_info;
3061         mad_queue->count = 0;
3062         spin_lock_init(&mad_queue->lock);
3063         INIT_LIST_HEAD(&mad_queue->list);
3064 }
3065
3066 static void init_mad_qp(struct ib_mad_port_private *port_priv,
3067                         struct ib_mad_qp_info *qp_info)
3068 {
3069         qp_info->port_priv = port_priv;
3070         init_mad_queue(qp_info, &qp_info->send_queue);
3071         init_mad_queue(qp_info, &qp_info->recv_queue);
3072         INIT_LIST_HEAD(&qp_info->overflow_list);
3073         spin_lock_init(&qp_info->snoop_lock);
3074         qp_info->snoop_table = NULL;
3075         qp_info->snoop_table_size = 0;
3076         atomic_set(&qp_info->snoop_count, 0);
3077 }
3078
3079 static int create_mad_qp(struct ib_mad_qp_info *qp_info,
3080                          enum ib_qp_type qp_type)
3081 {
3082         struct ib_qp_init_attr  qp_init_attr;
3083         int ret;
3084
3085         memset(&qp_init_attr, 0, sizeof qp_init_attr);
3086         qp_init_attr.send_cq = qp_info->port_priv->cq;
3087         qp_init_attr.recv_cq = qp_info->port_priv->cq;
3088         qp_init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
3089         qp_init_attr.cap.max_send_wr = mad_sendq_size;
3090         qp_init_attr.cap.max_recv_wr = mad_recvq_size;
3091         qp_init_attr.cap.max_send_sge = IB_MAD_SEND_REQ_MAX_SG;
3092         qp_init_attr.cap.max_recv_sge = IB_MAD_RECV_REQ_MAX_SG;
3093         qp_init_attr.qp_type = qp_type;
3094         qp_init_attr.port_num = qp_info->port_priv->port_num;
3095         qp_init_attr.qp_context = qp_info;
3096         qp_init_attr.event_handler = qp_event_handler;
3097         qp_info->qp = ib_create_qp(qp_info->port_priv->pd, &qp_init_attr);
3098         if (IS_ERR(qp_info->qp)) {
3099                 dev_err(&qp_info->port_priv->device->dev,
3100                         "Couldn't create ib_mad QP%d\n",
3101                         get_spl_qp_index(qp_type));
3102                 ret = PTR_ERR(qp_info->qp);
3103                 goto error;
3104         }
3105         /* Use minimum queue sizes unless the CQ is resized */
3106         qp_info->send_queue.max_active = mad_sendq_size;
3107         qp_info->recv_queue.max_active = mad_recvq_size;
3108         return 0;
3109
3110 error:
3111         return ret;
3112 }
3113
3114 static void destroy_mad_qp(struct ib_mad_qp_info *qp_info)
3115 {
3116         if (!qp_info->qp)
3117                 return;
3118
3119         ib_destroy_qp(qp_info->qp);
3120         kfree(qp_info->snoop_table);
3121 }
3122
3123 /*
3124  * Open the port
3125  * Create the QP, PD, MR, and CQ if needed
3126  */
3127 static int ib_mad_port_open(struct ib_device *device,
3128                             int port_num)
3129 {
3130         int ret, cq_size;
3131         struct ib_mad_port_private *port_priv;
3132         unsigned long flags;
3133         char name[sizeof "ib_mad123"];
3134         int has_smi;
3135
3136         if (WARN_ON(rdma_max_mad_size(device, port_num) < IB_MGMT_MAD_SIZE))
3137                 return -EFAULT;
3138
3139         if (WARN_ON(rdma_cap_opa_mad(device, port_num) &&
3140                     rdma_max_mad_size(device, port_num) < OPA_MGMT_MAD_SIZE))
3141                 return -EFAULT;
3142
3143         /* Create new device info */
3144         port_priv = kzalloc(sizeof *port_priv, GFP_KERNEL);
3145         if (!port_priv) {
3146                 dev_err(&device->dev, "No memory for ib_mad_port_private\n");
3147                 return -ENOMEM;
3148         }
3149
3150         port_priv->device = device;
3151         port_priv->port_num = port_num;
3152         spin_lock_init(&port_priv->reg_lock);
3153         INIT_LIST_HEAD(&port_priv->agent_list);
3154         init_mad_qp(port_priv, &port_priv->qp_info[0]);
3155         init_mad_qp(port_priv, &port_priv->qp_info[1]);
3156
3157         cq_size = mad_sendq_size + mad_recvq_size;
3158         has_smi = rdma_cap_ib_smi(device, port_num);
3159         if (has_smi)
3160                 cq_size *= 2;
3161
3162         port_priv->cq = ib_alloc_cq(port_priv->device, port_priv, cq_size, 0,
3163                         IB_POLL_WORKQUEUE);
3164         if (IS_ERR(port_priv->cq)) {
3165                 dev_err(&device->dev, "Couldn't create ib_mad CQ\n");
3166                 ret = PTR_ERR(port_priv->cq);
3167                 goto error3;
3168         }
3169
3170         port_priv->pd = ib_alloc_pd(device, 0);
3171         if (IS_ERR(port_priv->pd)) {
3172                 dev_err(&device->dev, "Couldn't create ib_mad PD\n");
3173                 ret = PTR_ERR(port_priv->pd);
3174                 goto error4;
3175         }
3176
3177         if (has_smi) {
3178                 ret = create_mad_qp(&port_priv->qp_info[0], IB_QPT_SMI);
3179                 if (ret)
3180                         goto error6;
3181         }
3182         ret = create_mad_qp(&port_priv->qp_info[1], IB_QPT_GSI);
3183         if (ret)
3184                 goto error7;
3185
3186         snprintf(name, sizeof name, "ib_mad%d", port_num);
3187         port_priv->wq = alloc_ordered_workqueue(name, WQ_MEM_RECLAIM);
3188         if (!port_priv->wq) {
3189                 ret = -ENOMEM;
3190                 goto error8;
3191         }
3192
3193         spin_lock_irqsave(&ib_mad_port_list_lock, flags);
3194         list_add_tail(&port_priv->port_list, &ib_mad_port_list);
3195         spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3196
3197         ret = ib_mad_port_start(port_priv);
3198         if (ret) {
3199                 dev_err(&device->dev, "Couldn't start port\n");
3200                 goto error9;
3201         }
3202
3203         return 0;
3204
3205 error9:
3206         spin_lock_irqsave(&ib_mad_port_list_lock, flags);
3207         list_del_init(&port_priv->port_list);
3208         spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3209
3210         destroy_workqueue(port_priv->wq);
3211 error8:
3212         destroy_mad_qp(&port_priv->qp_info[1]);
3213 error7:
3214         destroy_mad_qp(&port_priv->qp_info[0]);
3215 error6:
3216         ib_dealloc_pd(port_priv->pd);
3217 error4:
3218         ib_free_cq(port_priv->cq);
3219         cleanup_recv_queue(&port_priv->qp_info[1]);
3220         cleanup_recv_queue(&port_priv->qp_info[0]);
3221 error3:
3222         kfree(port_priv);
3223
3224         return ret;
3225 }
3226
3227 /*
3228  * Close the port
3229  * If there are no classes using the port, free the port
3230  * resources (CQ, MR, PD, QP) and remove the port's info structure
3231  */
3232 static int ib_mad_port_close(struct ib_device *device, int port_num)
3233 {
3234         struct ib_mad_port_private *port_priv;
3235         unsigned long flags;
3236
3237         spin_lock_irqsave(&ib_mad_port_list_lock, flags);
3238         port_priv = __ib_get_mad_port(device, port_num);
3239         if (port_priv == NULL) {
3240                 spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3241                 dev_err(&device->dev, "Port %d not found\n", port_num);
3242                 return -ENODEV;
3243         }
3244         list_del_init(&port_priv->port_list);
3245         spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3246
3247         destroy_workqueue(port_priv->wq);
3248         destroy_mad_qp(&port_priv->qp_info[1]);
3249         destroy_mad_qp(&port_priv->qp_info[0]);
3250         ib_dealloc_pd(port_priv->pd);
3251         ib_free_cq(port_priv->cq);
3252         cleanup_recv_queue(&port_priv->qp_info[1]);
3253         cleanup_recv_queue(&port_priv->qp_info[0]);
3254         /* XXX: Handle deallocation of MAD registration tables */
3255
3256         kfree(port_priv);
3257
3258         return 0;
3259 }
3260
3261 static void ib_mad_init_device(struct ib_device *device)
3262 {
3263         int start, i;
3264
3265         start = rdma_start_port(device);
3266
3267         for (i = start; i <= rdma_end_port(device); i++) {
3268                 if (!rdma_cap_ib_mad(device, i))
3269                         continue;
3270
3271                 if (ib_mad_port_open(device, i)) {
3272                         dev_err(&device->dev, "Couldn't open port %d\n", i);
3273                         goto error;
3274                 }
3275                 if (ib_agent_port_open(device, i)) {
3276                         dev_err(&device->dev,
3277                                 "Couldn't open port %d for agents\n", i);
3278                         goto error_agent;
3279                 }
3280         }
3281         return;
3282
3283 error_agent:
3284         if (ib_mad_port_close(device, i))
3285                 dev_err(&device->dev, "Couldn't close port %d\n", i);
3286
3287 error:
3288         while (--i >= start) {
3289                 if (!rdma_cap_ib_mad(device, i))
3290                         continue;
3291
3292                 if (ib_agent_port_close(device, i))
3293                         dev_err(&device->dev,
3294                                 "Couldn't close port %d for agents\n", i);
3295                 if (ib_mad_port_close(device, i))
3296                         dev_err(&device->dev, "Couldn't close port %d\n", i);
3297         }
3298 }
3299
3300 static void ib_mad_remove_device(struct ib_device *device, void *client_data)
3301 {
3302         int i;
3303
3304         for (i = rdma_start_port(device); i <= rdma_end_port(device); i++) {
3305                 if (!rdma_cap_ib_mad(device, i))
3306                         continue;
3307
3308                 if (ib_agent_port_close(device, i))
3309                         dev_err(&device->dev,
3310                                 "Couldn't close port %d for agents\n", i);
3311                 if (ib_mad_port_close(device, i))
3312                         dev_err(&device->dev, "Couldn't close port %d\n", i);
3313         }
3314 }
3315
3316 static struct ib_client mad_client = {
3317         .name   = "mad",
3318         .add = ib_mad_init_device,
3319         .remove = ib_mad_remove_device
3320 };
3321
3322 int ib_mad_init(void)
3323 {
3324         mad_recvq_size = min(mad_recvq_size, IB_MAD_QP_MAX_SIZE);
3325         mad_recvq_size = max(mad_recvq_size, IB_MAD_QP_MIN_SIZE);
3326
3327         mad_sendq_size = min(mad_sendq_size, IB_MAD_QP_MAX_SIZE);
3328         mad_sendq_size = max(mad_sendq_size, IB_MAD_QP_MIN_SIZE);
3329
3330         INIT_LIST_HEAD(&ib_mad_port_list);
3331
3332         if (ib_register_client(&mad_client)) {
3333                 pr_err("Couldn't register ib_mad client\n");
3334                 return -EINVAL;
3335         }
3336
3337         return 0;
3338 }
3339
3340 void ib_mad_cleanup(void)
3341 {
3342         ib_unregister_client(&mad_client);
3343 }