drivers/infiniband/core/mad.c

   1 /*
   2  * Copyright (c) 2004-2007 Voltaire, Inc. All rights reserved.
   3  * Copyright (c) 2005 Intel Corporation.  All rights reserved.
   4  * Copyright (c) 2005 Mellanox Technologies Ltd.  All rights reserved.
   5  * Copyright (c) 2009 HNR Consulting. All rights reserved.
   6  * Copyright (c) 2014 Intel Corporation.  All rights reserved.
   7  *
   8  * This software is available to you under a choice of one of two
   9  * licenses.  You may choose to be licensed under the terms of the GNU
  10  * General Public License (GPL) Version 2, available from the file
  11  * COPYING in the main directory of this source tree, or the
  12  * OpenIB.org BSD license below:
  13  *
  14  *     Redistribution and use in source and binary forms, with or
  15  *     without modification, are permitted provided that the following
  16  *     conditions are met:
  17  *
  18  *      - Redistributions of source code must retain the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer.
  21  *
  22  *      - Redistributions in binary form must reproduce the above
  23  *        copyright notice, this list of conditions and the following
  24  *        disclaimer in the documentation and/or other materials
  25  *        provided with the distribution.
  26  *
  27  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  28  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  29  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  30  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  31  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  32  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  33  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  34  * SOFTWARE.
  35  *
  36  */
  37
  38 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  39
  40 #include <linux/dma-mapping.h>
  41 #include <linux/slab.h>
  42 #include <linux/module.h>
  43 #include <linux/security.h>
  44 #include <rdma/ib_cache.h>
  45
  46 #include "mad_priv.h"
  47 #include "core_priv.h"
  48 #include "mad_rmpp.h"
  49 #include "smi.h"
  50 #include "opa_smi.h"
  51 #include "agent.h"
  52 #include "core_priv.h"
  53
  54 static int mad_sendq_size = IB_MAD_QP_SEND_SIZE;
  55 static int mad_recvq_size = IB_MAD_QP_RECV_SIZE;
  56
  57 module_param_named(send_queue_size, mad_sendq_size, int, 0444);
  58 MODULE_PARM_DESC(send_queue_size, "Size of send queue in number of work requests");
  59 module_param_named(recv_queue_size, mad_recvq_size, int, 0444);
  60 MODULE_PARM_DESC(recv_queue_size, "Size of receive queue in number of work requests");
  61
  62 static struct list_head ib_mad_port_list;
  63 static atomic_t ib_mad_client_id = ATOMIC_INIT(0);
  64
  65 /* Port list lock */
  66 static DEFINE_SPINLOCK(ib_mad_port_list_lock);
  67
  68 /* Forward declarations */
  69 static int method_in_use(struct ib_mad_mgmt_method_table **method,
  70                          struct ib_mad_reg_req *mad_reg_req);
  71 static void remove_mad_reg_req(struct ib_mad_agent_private *priv);
  72 static struct ib_mad_agent_private *find_mad_agent(
  73                                         struct ib_mad_port_private *port_priv,
  74                                         const struct ib_mad_hdr *mad);
  75 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
  76                                     struct ib_mad_private *mad);
  77 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv);
  78 static void timeout_sends(struct work_struct *work);
  79 static void local_completions(struct work_struct *work);
  80 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
  81                               struct ib_mad_agent_private *agent_priv,
  82                               u8 mgmt_class);
  83 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
  84                            struct ib_mad_agent_private *agent_priv);
  85 static bool ib_mad_send_error(struct ib_mad_port_private *port_priv,
  86                               struct ib_wc *wc);
  87 static void ib_mad_send_done(struct ib_cq *cq, struct ib_wc *wc);
  88
  89 /*
  90  * Returns a ib_mad_port_private structure or NULL for a device/port
  91  * Assumes ib_mad_port_list_lock is being held
  92  */
  93 static inline struct ib_mad_port_private *
  94 __ib_get_mad_port(struct ib_device *device, int port_num)
  95 {
  96         struct ib_mad_port_private *entry;
  97
  98         list_for_each_entry(entry, &ib_mad_port_list, port_list) {
  99                 if (entry->device == device && entry->port_num == port_num)
 100                         return entry;
 101         }
 102         return NULL;
 103 }
 104
 105 /*
 106  * Wrapper function to return a ib_mad_port_private structure or NULL
 107  * for a device/port
 108  */
 109 static inline struct ib_mad_port_private *
 110 ib_get_mad_port(struct ib_device *device, int port_num)
 111 {
 112         struct ib_mad_port_private *entry;
 113         unsigned long flags;
 114
 115         spin_lock_irqsave(&ib_mad_port_list_lock, flags);
 116         entry = __ib_get_mad_port(device, port_num);
 117         spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
 118
 119         return entry;
 120 }
 121
 122 static inline u8 convert_mgmt_class(u8 mgmt_class)
 123 {
 124         /* Alias IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE to 0 */
 125         return mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE ?
 126                 0 : mgmt_class;
 127 }
 128
 129 static int get_spl_qp_index(enum ib_qp_type qp_type)
 130 {
 131         switch (qp_type)
 132         {
 133         case IB_QPT_SMI:
 134                 return 0;
 135         case IB_QPT_GSI:
 136                 return 1;
 137         default:
 138                 return -1;
 139         }
 140 }
 141
 142 static int vendor_class_index(u8 mgmt_class)
 143 {
 144         return mgmt_class - IB_MGMT_CLASS_VENDOR_RANGE2_START;
 145 }
 146
 147 static int is_vendor_class(u8 mgmt_class)
 148 {
 149         if ((mgmt_class < IB_MGMT_CLASS_VENDOR_RANGE2_START) ||
 150             (mgmt_class > IB_MGMT_CLASS_VENDOR_RANGE2_END))
 151                 return 0;
 152         return 1;
 153 }
 154
 155 static int is_vendor_oui(char *oui)
 156 {
 157         if (oui[0] || oui[1] || oui[2])
 158                 return 1;
 159         return 0;
 160 }
 161
 162 static int is_vendor_method_in_use(
 163                 struct ib_mad_mgmt_vendor_class *vendor_class,
 164                 struct ib_mad_reg_req *mad_reg_req)
 165 {
 166         struct ib_mad_mgmt_method_table *method;
 167         int i;
 168
 169         for (i = 0; i < MAX_MGMT_OUI; i++) {
 170                 if (!memcmp(vendor_class->oui[i], mad_reg_req->oui, 3)) {
 171                         method = vendor_class->method_table[i];
 172                         if (method) {
 173                                 if (method_in_use(&method, mad_reg_req))
 174                                         return 1;
 175                                 else
 176                                         break;
 177                         }
 178                 }
 179         }
 180         return 0;
 181 }
 182
 183 int ib_response_mad(const struct ib_mad_hdr *hdr)
 184 {
 185         return ((hdr->method & IB_MGMT_METHOD_RESP) ||
 186                 (hdr->method == IB_MGMT_METHOD_TRAP_REPRESS) ||
 187                 ((hdr->mgmt_class == IB_MGMT_CLASS_BM) &&
 188                  (hdr->attr_mod & IB_BM_ATTR_MOD_RESP)));
 189 }
 190 EXPORT_SYMBOL(ib_response_mad);
 191
 192 /*
 193  * ib_register_mad_agent - Register to send/receive MADs
 194  */
 195 struct ib_mad_agent *ib_register_mad_agent(struct ib_device *device,
 196                                            u8 port_num,
 197                                            enum ib_qp_type qp_type,
 198                                            struct ib_mad_reg_req *mad_reg_req,
 199                                            u8 rmpp_version,
 200                                            ib_mad_send_handler send_handler,
 201                                            ib_mad_recv_handler recv_handler,
 202                                            void *context,
 203                                            u32 registration_flags)
 204 {
 205         struct ib_mad_port_private *port_priv;
 206         struct ib_mad_agent *ret = ERR_PTR(-EINVAL);
 207         struct ib_mad_agent_private *mad_agent_priv;
 208         struct ib_mad_reg_req *reg_req = NULL;
 209         struct ib_mad_mgmt_class_table *class;
 210         struct ib_mad_mgmt_vendor_class_table *vendor;
 211         struct ib_mad_mgmt_vendor_class *vendor_class;
 212         struct ib_mad_mgmt_method_table *method;
 213         int ret2, qpn;
 214         unsigned long flags;
 215         u8 mgmt_class, vclass;
 216
 217         /* Validate parameters */
 218         qpn = get_spl_qp_index(qp_type);
 219         if (qpn == -1) {
 220                 dev_dbg_ratelimited(&device->dev, "%s: invalid QP Type %d\n",
 221                                     __func__, qp_type);
 222                 goto error1;
 223         }
 224
 225         if (rmpp_version && rmpp_version != IB_MGMT_RMPP_VERSION) {
 226                 dev_dbg_ratelimited(&device->dev,
 227                                     "%s: invalid RMPP Version %u\n",
 228                                     __func__, rmpp_version);
 229                 goto error1;
 230         }
 231
 232         /* Validate MAD registration request if supplied */
 233         if (mad_reg_req) {
 234                 if (mad_reg_req->mgmt_class_version >= MAX_MGMT_VERSION) {
 235                         dev_dbg_ratelimited(&device->dev,
 236                                             "%s: invalid Class Version %u\n",
 237                                             __func__,
 238                                             mad_reg_req->mgmt_class_version);
 239                         goto error1;
 240                 }
 241                 if (!recv_handler) {
 242                         dev_dbg_ratelimited(&device->dev,
 243                                             "%s: no recv_handler\n", __func__);
 244                         goto error1;
 245                 }
 246                 if (mad_reg_req->mgmt_class >= MAX_MGMT_CLASS) {
 247                         /*
 248                          * IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE is the only
 249                          * one in this range currently allowed
 250                          */
 251                         if (mad_reg_req->mgmt_class !=
 252                             IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
 253                                 dev_dbg_ratelimited(&device->dev,
 254                                         "%s: Invalid Mgmt Class 0x%x\n",
 255                                         __func__, mad_reg_req->mgmt_class);
 256                                 goto error1;
 257                         }
 258                 } else if (mad_reg_req->mgmt_class == 0) {
 259                         /*
 260                          * Class 0 is reserved in IBA and is used for
 261                          * aliasing of IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE
 262                          */
 263                         dev_dbg_ratelimited(&device->dev,
 264                                             "%s: Invalid Mgmt Class 0\n",
 265                                             __func__);
 266                         goto error1;
 267                 } else if (is_vendor_class(mad_reg_req->mgmt_class)) {
 268                         /*
 269                          * If class is in "new" vendor range,
 270                          * ensure supplied OUI is not zero
 271                          */
 272                         if (!is_vendor_oui(mad_reg_req->oui)) {
 273                                 dev_dbg_ratelimited(&device->dev,
 274                                         "%s: No OUI specified for class 0x%x\n",
 275                                         __func__,
 276                                         mad_reg_req->mgmt_class);
 277                                 goto error1;
 278                         }
 279                 }
 280                 /* Make sure class supplied is consistent with RMPP */
 281                 if (!ib_is_mad_class_rmpp(mad_reg_req->mgmt_class)) {
 282                         if (rmpp_version) {
 283                                 dev_dbg_ratelimited(&device->dev,
 284                                         "%s: RMPP version for non-RMPP class 0x%x\n",
 285                                         __func__, mad_reg_req->mgmt_class);
 286                                 goto error1;
 287                         }
 288                 }
 289
 290                 /* Make sure class supplied is consistent with QP type */
 291                 if (qp_type == IB_QPT_SMI) {
 292                         if ((mad_reg_req->mgmt_class !=
 293                                         IB_MGMT_CLASS_SUBN_LID_ROUTED) &&
 294                             (mad_reg_req->mgmt_class !=
 295                                         IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
 296                                 dev_dbg_ratelimited(&device->dev,
 297                                         "%s: Invalid SM QP type: class 0x%x\n",
 298                                         __func__, mad_reg_req->mgmt_class);
 299                                 goto error1;
 300                         }
 301                 } else {
 302                         if ((mad_reg_req->mgmt_class ==
 303                                         IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
 304                             (mad_reg_req->mgmt_class ==
 305                                         IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
 306                                 dev_dbg_ratelimited(&device->dev,
 307                                         "%s: Invalid GS QP type: class 0x%x\n",
 308                                         __func__, mad_reg_req->mgmt_class);
 309                                 goto error1;
 310                         }
 311                 }
 312         } else {
 313                 /* No registration request supplied */
 314                 if (!send_handler)
 315                         goto error1;
 316                 if (registration_flags & IB_MAD_USER_RMPP)
 317                         goto error1;
 318         }
 319
 320         /* Validate device and port */
 321         port_priv = ib_get_mad_port(device, port_num);
 322         if (!port_priv) {
 323                 dev_dbg_ratelimited(&device->dev, "%s: Invalid port %d\n",
 324                                     __func__, port_num);
 325                 ret = ERR_PTR(-ENODEV);
 326                 goto error1;
 327         }
 328
 329         /* Verify the QP requested is supported. For example, Ethernet devices
 330          * will not have QP0.
 331          */
 332         if (!port_priv->qp_info[qpn].qp) {
 333                 dev_dbg_ratelimited(&device->dev, "%s: QP %d not supported\n",
 334                                     __func__, qpn);
 335                 ret = ERR_PTR(-EPROTONOSUPPORT);
 336                 goto error1;
 337         }
 338
 339         /* Allocate structures */
 340         mad_agent_priv = kzalloc(sizeof *mad_agent_priv, GFP_KERNEL);
 341         if (!mad_agent_priv) {
 342                 ret = ERR_PTR(-ENOMEM);
 343                 goto error1;
 344         }
 345
 346         if (mad_reg_req) {
 347                 reg_req = kmemdup(mad_reg_req, sizeof *reg_req, GFP_KERNEL);
 348                 if (!reg_req) {
 349                         ret = ERR_PTR(-ENOMEM);
 350                         goto error3;
 351                 }
 352         }
 353
 354         /* Now, fill in the various structures */
 355         mad_agent_priv->qp_info = &port_priv->qp_info[qpn];
 356         mad_agent_priv->reg_req = reg_req;
 357         mad_agent_priv->agent.rmpp_version = rmpp_version;
 358         mad_agent_priv->agent.device = device;
 359         mad_agent_priv->agent.recv_handler = recv_handler;
 360         mad_agent_priv->agent.send_handler = send_handler;
 361         mad_agent_priv->agent.context = context;
 362         mad_agent_priv->agent.qp = port_priv->qp_info[qpn].qp;
 363         mad_agent_priv->agent.port_num = port_num;
 364         mad_agent_priv->agent.flags = registration_flags;
 365         spin_lock_init(&mad_agent_priv->lock);
 366         INIT_LIST_HEAD(&mad_agent_priv->send_list);
 367         INIT_LIST_HEAD(&mad_agent_priv->wait_list);
 368         INIT_LIST_HEAD(&mad_agent_priv->done_list);
 369         INIT_LIST_HEAD(&mad_agent_priv->rmpp_list);
 370         INIT_DELAYED_WORK(&mad_agent_priv->timed_work, timeout_sends);
 371         INIT_LIST_HEAD(&mad_agent_priv->local_list);
 372         INIT_WORK(&mad_agent_priv->local_work, local_completions);
 373         atomic_set(&mad_agent_priv->refcount, 1);
 374         init_completion(&mad_agent_priv->comp);
 375
 376         ret2 = ib_mad_agent_security_setup(&mad_agent_priv->agent, qp_type);
 377         if (ret2) {
 378                 ret = ERR_PTR(ret2);
 379                 goto error4;
 380         }
 381
 382         spin_lock_irqsave(&port_priv->reg_lock, flags);
 383         mad_agent_priv->agent.hi_tid = atomic_inc_return(&ib_mad_client_id);
 384
 385         /*
 386          * Make sure MAD registration (if supplied)
 387          * is non overlapping with any existing ones
 388          */
 389         if (mad_reg_req) {
 390                 mgmt_class = convert_mgmt_class(mad_reg_req->mgmt_class);
 391                 if (!is_vendor_class(mgmt_class)) {
 392                         class = port_priv->version[mad_reg_req->
 393                                                    mgmt_class_version].class;
 394                         if (class) {
 395                                 method = class->method_table[mgmt_class];
 396                                 if (method) {
 397                                         if (method_in_use(&method,
 398                                                            mad_reg_req))
 399                                                 goto error5;
 400                                 }
 401                         }
 402                         ret2 = add_nonoui_reg_req(mad_reg_req, mad_agent_priv,
 403                                                   mgmt_class);
 404                 } else {
 405                         /* "New" vendor class range */
 406                         vendor = port_priv->version[mad_reg_req->
 407                                                     mgmt_class_version].vendor;
 408                         if (vendor) {
 409                                 vclass = vendor_class_index(mgmt_class);
 410                                 vendor_class = vendor->vendor_class[vclass];
 411                                 if (vendor_class) {
 412                                         if (is_vendor_method_in_use(
 413                                                         vendor_class,
 414                                                         mad_reg_req))
 415                                                 goto error5;
 416                                 }
 417                         }
 418                         ret2 = add_oui_reg_req(mad_reg_req, mad_agent_priv);
 419                 }
 420                 if (ret2) {
 421                         ret = ERR_PTR(ret2);
 422                         goto error5;
 423                 }
 424         }
 425
 426         /* Add mad agent into port's agent list */
 427         list_add_tail(&mad_agent_priv->agent_list, &port_priv->agent_list);
 428         spin_unlock_irqrestore(&port_priv->reg_lock, flags);
 429
 430         return &mad_agent_priv->agent;
 431 error5:
 432         spin_unlock_irqrestore(&port_priv->reg_lock, flags);
 433         ib_mad_agent_security_cleanup(&mad_agent_priv->agent);
 434 error4:
 435         kfree(reg_req);
 436 error3:
 437         kfree(mad_agent_priv);
 438 error1:
 439         return ret;
 440 }
 441 EXPORT_SYMBOL(ib_register_mad_agent);
 442
 443 static inline int is_snooping_sends(int mad_snoop_flags)
 444 {
 445         return (mad_snoop_flags &
 446                 (/*IB_MAD_SNOOP_POSTED_SENDS |
 447                  IB_MAD_SNOOP_RMPP_SENDS |*/
 448                  IB_MAD_SNOOP_SEND_COMPLETIONS /*|
 449                  IB_MAD_SNOOP_RMPP_SEND_COMPLETIONS*/));
 450 }
 451
 452 static inline int is_snooping_recvs(int mad_snoop_flags)
 453 {
 454         return (mad_snoop_flags &
 455                 (IB_MAD_SNOOP_RECVS /*|
 456                  IB_MAD_SNOOP_RMPP_RECVS*/));
 457 }
 458
 459 static int register_snoop_agent(struct ib_mad_qp_info *qp_info,
 460                                 struct ib_mad_snoop_private *mad_snoop_priv)
 461 {
 462         struct ib_mad_snoop_private **new_snoop_table;
 463         unsigned long flags;
 464         int i;
 465
 466         spin_lock_irqsave(&qp_info->snoop_lock, flags);
 467         /* Check for empty slot in array. */
 468         for (i = 0; i < qp_info->snoop_table_size; i++)
 469                 if (!qp_info->snoop_table[i])
 470                         break;
 471
 472         if (i == qp_info->snoop_table_size) {
 473                 /* Grow table. */
 474                 new_snoop_table = krealloc(qp_info->snoop_table,
 475                                            sizeof mad_snoop_priv *
 476                                            (qp_info->snoop_table_size + 1),
 477                                            GFP_ATOMIC);
 478                 if (!new_snoop_table) {
 479                         i = -ENOMEM;
 480                         goto out;
 481                 }
 482
 483                 qp_info->snoop_table = new_snoop_table;
 484                 qp_info->snoop_table_size++;
 485         }
 486         qp_info->snoop_table[i] = mad_snoop_priv;
 487         atomic_inc(&qp_info->snoop_count);
 488 out:
 489         spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
 490         return i;
 491 }
 492
 493 struct ib_mad_agent *ib_register_mad_snoop(struct ib_device *device,
 494                                            u8 port_num,
 495                                            enum ib_qp_type qp_type,
 496                                            int mad_snoop_flags,
 497                                            ib_mad_snoop_handler snoop_handler,
 498                                            ib_mad_recv_handler recv_handler,
 499                                            void *context)
 500 {
 501         struct ib_mad_port_private *port_priv;
 502         struct ib_mad_agent *ret;
 503         struct ib_mad_snoop_private *mad_snoop_priv;
 504         int qpn;
 505         int err;
 506
 507         /* Validate parameters */
 508         if ((is_snooping_sends(mad_snoop_flags) && !snoop_handler) ||
 509             (is_snooping_recvs(mad_snoop_flags) && !recv_handler)) {
 510                 ret = ERR_PTR(-EINVAL);
 511                 goto error1;
 512         }
 513         qpn = get_spl_qp_index(qp_type);
 514         if (qpn == -1) {
 515                 ret = ERR_PTR(-EINVAL);
 516                 goto error1;
 517         }
 518         port_priv = ib_get_mad_port(device, port_num);
 519         if (!port_priv) {
 520                 ret = ERR_PTR(-ENODEV);
 521                 goto error1;
 522         }
 523         /* Allocate structures */
 524         mad_snoop_priv = kzalloc(sizeof *mad_snoop_priv, GFP_KERNEL);
 525         if (!mad_snoop_priv) {
 526                 ret = ERR_PTR(-ENOMEM);
 527                 goto error1;
 528         }
 529
 530         /* Now, fill in the various structures */
 531         mad_snoop_priv->qp_info = &port_priv->qp_info[qpn];
 532         mad_snoop_priv->agent.device = device;
 533         mad_snoop_priv->agent.recv_handler = recv_handler;
 534         mad_snoop_priv->agent.snoop_handler = snoop_handler;
 535         mad_snoop_priv->agent.context = context;
 536         mad_snoop_priv->agent.qp = port_priv->qp_info[qpn].qp;
 537         mad_snoop_priv->agent.port_num = port_num;
 538         mad_snoop_priv->mad_snoop_flags = mad_snoop_flags;
 539         init_completion(&mad_snoop_priv->comp);
 540
 541         err = ib_mad_agent_security_setup(&mad_snoop_priv->agent, qp_type);
 542         if (err) {
 543                 ret = ERR_PTR(err);
 544                 goto error2;
 545         }
 546
 547         mad_snoop_priv->snoop_index = register_snoop_agent(
 548                                                 &port_priv->qp_info[qpn],
 549                                                 mad_snoop_priv);
 550         if (mad_snoop_priv->snoop_index < 0) {
 551                 ret = ERR_PTR(mad_snoop_priv->snoop_index);
 552                 goto error3;
 553         }
 554
 555         atomic_set(&mad_snoop_priv->refcount, 1);
 556         return &mad_snoop_priv->agent;
 557 error3:
 558         ib_mad_agent_security_cleanup(&mad_snoop_priv->agent);
 559 error2:
 560         kfree(mad_snoop_priv);
 561 error1:
 562         return ret;
 563 }
 564 EXPORT_SYMBOL(ib_register_mad_snoop);
 565
 566 static inline void deref_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
 567 {
 568         if (atomic_dec_and_test(&mad_agent_priv->refcount))
 569                 complete(&mad_agent_priv->comp);
 570 }
 571
 572 static inline void deref_snoop_agent(struct ib_mad_snoop_private *mad_snoop_priv)
 573 {
 574         if (atomic_dec_and_test(&mad_snoop_priv->refcount))
 575                 complete(&mad_snoop_priv->comp);
 576 }
 577
 578 static void unregister_mad_agent(struct ib_mad_agent_private *mad_agent_priv)
 579 {
 580         struct ib_mad_port_private *port_priv;
 581         unsigned long flags;
 582
 583         /* Note that we could still be handling received MADs */
 584
 585         /*
 586          * Canceling all sends results in dropping received response
 587          * MADs, preventing us from queuing additional work
 588          */
 589         cancel_mads(mad_agent_priv);
 590         port_priv = mad_agent_priv->qp_info->port_priv;
 591         cancel_delayed_work(&mad_agent_priv->timed_work);
 592
 593         spin_lock_irqsave(&port_priv->reg_lock, flags);
 594         remove_mad_reg_req(mad_agent_priv);
 595         list_del(&mad_agent_priv->agent_list);
 596         spin_unlock_irqrestore(&port_priv->reg_lock, flags);
 597
 598         flush_workqueue(port_priv->wq);
 599         ib_cancel_rmpp_recvs(mad_agent_priv);
 600
 601         deref_mad_agent(mad_agent_priv);
 602         wait_for_completion(&mad_agent_priv->comp);
 603
 604         ib_mad_agent_security_cleanup(&mad_agent_priv->agent);
 605
 606         kfree(mad_agent_priv->reg_req);
 607         kfree(mad_agent_priv);
 608 }
 609
 610 static void unregister_mad_snoop(struct ib_mad_snoop_private *mad_snoop_priv)
 611 {
 612         struct ib_mad_qp_info *qp_info;
 613         unsigned long flags;
 614
 615         qp_info = mad_snoop_priv->qp_info;
 616         spin_lock_irqsave(&qp_info->snoop_lock, flags);
 617         qp_info->snoop_table[mad_snoop_priv->snoop_index] = NULL;
 618         atomic_dec(&qp_info->snoop_count);
 619         spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
 620
 621         deref_snoop_agent(mad_snoop_priv);
 622         wait_for_completion(&mad_snoop_priv->comp);
 623
 624         ib_mad_agent_security_cleanup(&mad_snoop_priv->agent);
 625
 626         kfree(mad_snoop_priv);
 627 }
 628
 629 /*
 630  * ib_unregister_mad_agent - Unregisters a client from using MAD services
 631  */
 632 void ib_unregister_mad_agent(struct ib_mad_agent *mad_agent)
 633 {
 634         struct ib_mad_agent_private *mad_agent_priv;
 635         struct ib_mad_snoop_private *mad_snoop_priv;
 636
 637         /* If the TID is zero, the agent can only snoop. */
 638         if (mad_agent->hi_tid) {
 639                 mad_agent_priv = container_of(mad_agent,
 640                                               struct ib_mad_agent_private,
 641                                               agent);
 642                 unregister_mad_agent(mad_agent_priv);
 643         } else {
 644                 mad_snoop_priv = container_of(mad_agent,
 645                                               struct ib_mad_snoop_private,
 646                                               agent);
 647                 unregister_mad_snoop(mad_snoop_priv);
 648         }
 649 }
 650 EXPORT_SYMBOL(ib_unregister_mad_agent);
 651
 652 static void dequeue_mad(struct ib_mad_list_head *mad_list)
 653 {
 654         struct ib_mad_queue *mad_queue;
 655         unsigned long flags;
 656
 657         BUG_ON(!mad_list->mad_queue);
 658         mad_queue = mad_list->mad_queue;
 659         spin_lock_irqsave(&mad_queue->lock, flags);
 660         list_del(&mad_list->list);
 661         mad_queue->count--;
 662         spin_unlock_irqrestore(&mad_queue->lock, flags);
 663 }
 664
 665 static void snoop_send(struct ib_mad_qp_info *qp_info,
 666                        struct ib_mad_send_buf *send_buf,
 667                        struct ib_mad_send_wc *mad_send_wc,
 668                        int mad_snoop_flags)
 669 {
 670         struct ib_mad_snoop_private *mad_snoop_priv;
 671         unsigned long flags;
 672         int i;
 673
 674         spin_lock_irqsave(&qp_info->snoop_lock, flags);
 675         for (i = 0; i < qp_info->snoop_table_size; i++) {
 676                 mad_snoop_priv = qp_info->snoop_table[i];
 677                 if (!mad_snoop_priv ||
 678                     !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
 679                         continue;
 680
 681                 atomic_inc(&mad_snoop_priv->refcount);
 682                 spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
 683                 mad_snoop_priv->agent.snoop_handler(&mad_snoop_priv->agent,
 684                                                     send_buf, mad_send_wc);
 685                 deref_snoop_agent(mad_snoop_priv);
 686                 spin_lock_irqsave(&qp_info->snoop_lock, flags);
 687         }
 688         spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
 689 }
 690
 691 static void snoop_recv(struct ib_mad_qp_info *qp_info,
 692                        struct ib_mad_recv_wc *mad_recv_wc,
 693                        int mad_snoop_flags)
 694 {
 695         struct ib_mad_snoop_private *mad_snoop_priv;
 696         unsigned long flags;
 697         int i;
 698
 699         spin_lock_irqsave(&qp_info->snoop_lock, flags);
 700         for (i = 0; i < qp_info->snoop_table_size; i++) {
 701                 mad_snoop_priv = qp_info->snoop_table[i];
 702                 if (!mad_snoop_priv ||
 703                     !(mad_snoop_priv->mad_snoop_flags & mad_snoop_flags))
 704                         continue;
 705
 706                 atomic_inc(&mad_snoop_priv->refcount);
 707                 spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
 708                 mad_snoop_priv->agent.recv_handler(&mad_snoop_priv->agent, NULL,
 709                                                    mad_recv_wc);
 710                 deref_snoop_agent(mad_snoop_priv);
 711                 spin_lock_irqsave(&qp_info->snoop_lock, flags);
 712         }
 713         spin_unlock_irqrestore(&qp_info->snoop_lock, flags);
 714 }
 715
 716 static void build_smp_wc(struct ib_qp *qp, struct ib_cqe *cqe, u16 slid,
 717                 u16 pkey_index, u8 port_num, struct ib_wc *wc)
 718 {
 719         memset(wc, 0, sizeof *wc);
 720         wc->wr_cqe = cqe;
 721         wc->status = IB_WC_SUCCESS;
 722         wc->opcode = IB_WC_RECV;
 723         wc->pkey_index = pkey_index;
 724         wc->byte_len = sizeof(struct ib_mad) + sizeof(struct ib_grh);
 725         wc->src_qp = IB_QP0;
 726         wc->qp = qp;
 727         wc->slid = slid;
 728         wc->sl = 0;
 729         wc->dlid_path_bits = 0;
 730         wc->port_num = port_num;
 731 }
 732
 733 static size_t mad_priv_size(const struct ib_mad_private *mp)
 734 {
 735         return sizeof(struct ib_mad_private) + mp->mad_size;
 736 }
 737
 738 static struct ib_mad_private *alloc_mad_private(size_t mad_size, gfp_t flags)
 739 {
 740         size_t size = sizeof(struct ib_mad_private) + mad_size;
 741         struct ib_mad_private *ret = kzalloc(size, flags);
 742
 743         if (ret)
 744                 ret->mad_size = mad_size;
 745
 746         return ret;
 747 }
 748
 749 static size_t port_mad_size(const struct ib_mad_port_private *port_priv)
 750 {
 751         return rdma_max_mad_size(port_priv->device, port_priv->port_num);
 752 }
 753
 754 static size_t mad_priv_dma_size(const struct ib_mad_private *mp)
 755 {
 756         return sizeof(struct ib_grh) + mp->mad_size;
 757 }
 758
 759 /*
 760  * Return 0 if SMP is to be sent
 761  * Return 1 if SMP was consumed locally (whether or not solicited)
 762  * Return < 0 if error
 763  */
 764 static int handle_outgoing_dr_smp(struct ib_mad_agent_private *mad_agent_priv,
 765                                   struct ib_mad_send_wr_private *mad_send_wr)
 766 {
 767         int ret = 0;
 768         struct ib_smp *smp = mad_send_wr->send_buf.mad;
 769         struct opa_smp *opa_smp = (struct opa_smp *)smp;
 770         unsigned long flags;
 771         struct ib_mad_local_private *local;
 772         struct ib_mad_private *mad_priv;
 773         struct ib_mad_port_private *port_priv;
 774         struct ib_mad_agent_private *recv_mad_agent = NULL;
 775         struct ib_device *device = mad_agent_priv->agent.device;
 776         u8 port_num;
 777         struct ib_wc mad_wc;
 778         struct ib_ud_wr *send_wr = &mad_send_wr->send_wr;
 779         size_t mad_size = port_mad_size(mad_agent_priv->qp_info->port_priv);
 780         u16 out_mad_pkey_index = 0;
 781         u16 drslid;
 782         bool opa = rdma_cap_opa_mad(mad_agent_priv->qp_info->port_priv->device,
 783                                     mad_agent_priv->qp_info->port_priv->port_num);
 784
 785         if (rdma_cap_ib_switch(device) &&
 786             smp->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
 787                 port_num = send_wr->port_num;
 788         else
 789                 port_num = mad_agent_priv->agent.port_num;
 790
 791         /*
 792          * Directed route handling starts if the initial LID routed part of
 793          * a request or the ending LID routed part of a response is empty.
 794          * If we are at the start of the LID routed part, don't update the
 795          * hop_ptr or hop_cnt.  See section 14.2.2, Vol 1 IB spec.
 796          */
 797         if (opa && smp->class_version == OPA_SM_CLASS_VERSION) {
 798                 u32 opa_drslid;
 799
 800                 if ((opa_get_smp_direction(opa_smp)
 801                      ? opa_smp->route.dr.dr_dlid : opa_smp->route.dr.dr_slid) ==
 802                      OPA_LID_PERMISSIVE &&
 803                      opa_smi_handle_dr_smp_send(opa_smp,
 804                                                 rdma_cap_ib_switch(device),
 805                                                 port_num) == IB_SMI_DISCARD) {
 806                         ret = -EINVAL;
 807                         dev_err(&device->dev, "OPA Invalid directed route\n");
 808                         goto out;
 809                 }
 810                 opa_drslid = be32_to_cpu(opa_smp->route.dr.dr_slid);
 811                 if (opa_drslid != be32_to_cpu(OPA_LID_PERMISSIVE) &&
 812                     opa_drslid & 0xffff0000) {
 813                         ret = -EINVAL;
 814                         dev_err(&device->dev, "OPA Invalid dr_slid 0x%x\n",
 815                                opa_drslid);
 816                         goto out;
 817                 }
 818                 drslid = (u16)(opa_drslid & 0x0000ffff);
 819
 820                 /* Check to post send on QP or process locally */
 821                 if (opa_smi_check_local_smp(opa_smp, device) == IB_SMI_DISCARD &&
 822                     opa_smi_check_local_returning_smp(opa_smp, device) == IB_SMI_DISCARD)
 823                         goto out;
 824         } else {
 825                 if ((ib_get_smp_direction(smp) ? smp->dr_dlid : smp->dr_slid) ==
 826                      IB_LID_PERMISSIVE &&
 827                      smi_handle_dr_smp_send(smp, rdma_cap_ib_switch(device), port_num) ==
 828                      IB_SMI_DISCARD) {
 829                         ret = -EINVAL;
 830                         dev_err(&device->dev, "Invalid directed route\n");
 831                         goto out;
 832                 }
 833                 drslid = be16_to_cpu(smp->dr_slid);
 834
 835                 /* Check to post send on QP or process locally */
 836                 if (smi_check_local_smp(smp, device) == IB_SMI_DISCARD &&
 837                     smi_check_local_returning_smp(smp, device) == IB_SMI_DISCARD)
 838                         goto out;
 839         }
 840
 841         local = kmalloc(sizeof *local, GFP_ATOMIC);
 842         if (!local) {
 843                 ret = -ENOMEM;
 844                 goto out;
 845         }
 846         local->mad_priv = NULL;
 847         local->recv_mad_agent = NULL;
 848         mad_priv = alloc_mad_private(mad_size, GFP_ATOMIC);
 849         if (!mad_priv) {
 850                 ret = -ENOMEM;
 851                 kfree(local);
 852                 goto out;
 853         }
 854
 855         build_smp_wc(mad_agent_priv->agent.qp,
 856                      send_wr->wr.wr_cqe, drslid,
 857                      send_wr->pkey_index,
 858                      send_wr->port_num, &mad_wc);
 859
 860         if (opa && smp->base_version == OPA_MGMT_BASE_VERSION) {
 861                 mad_wc.byte_len = mad_send_wr->send_buf.hdr_len
 862                                         + mad_send_wr->send_buf.data_len
 863                                         + sizeof(struct ib_grh);
 864         }
 865
 866         /* No GRH for DR SMP */
 867         ret = device->process_mad(device, 0, port_num, &mad_wc, NULL,
 868                                   (const struct ib_mad_hdr *)smp, mad_size,
 869                                   (struct ib_mad_hdr *)mad_priv->mad,
 870                                   &mad_size, &out_mad_pkey_index);
 871         switch (ret)
 872         {
 873         case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_REPLY:
 874                 if (ib_response_mad((const struct ib_mad_hdr *)mad_priv->mad) &&
 875                     mad_agent_priv->agent.recv_handler) {
 876                         local->mad_priv = mad_priv;
 877                         local->recv_mad_agent = mad_agent_priv;
 878                         /*
 879                          * Reference MAD agent until receive
 880                          * side of local completion handled
 881                          */
 882                         atomic_inc(&mad_agent_priv->refcount);
 883                 } else
 884                         kfree(mad_priv);
 885                 break;
 886         case IB_MAD_RESULT_SUCCESS | IB_MAD_RESULT_CONSUMED:
 887                 kfree(mad_priv);
 888                 break;
 889         case IB_MAD_RESULT_SUCCESS:
 890                 /* Treat like an incoming receive MAD */
 891                 port_priv = ib_get_mad_port(mad_agent_priv->agent.device,
 892                                             mad_agent_priv->agent.port_num);
 893                 if (port_priv) {
 894                         memcpy(mad_priv->mad, smp, mad_priv->mad_size);
 895                         recv_mad_agent = find_mad_agent(port_priv,
 896                                                         (const struct ib_mad_hdr *)mad_priv->mad);
 897                 }
 898                 if (!port_priv || !recv_mad_agent) {
 899                         /*
 900                          * No receiving agent so drop packet and
 901                          * generate send completion.
 902                          */
 903                         kfree(mad_priv);
 904                         break;
 905                 }
 906                 local->mad_priv = mad_priv;
 907                 local->recv_mad_agent = recv_mad_agent;
 908                 break;
 909         default:
 910                 kfree(mad_priv);
 911                 kfree(local);
 912                 ret = -EINVAL;
 913                 goto out;
 914         }
 915
 916         local->mad_send_wr = mad_send_wr;
 917         if (opa) {
 918                 local->mad_send_wr->send_wr.pkey_index = out_mad_pkey_index;
 919                 local->return_wc_byte_len = mad_size;
 920         }
 921         /* Reference MAD agent until send side of local completion handled */
 922         atomic_inc(&mad_agent_priv->refcount);
 923         /* Queue local completion to local list */
 924         spin_lock_irqsave(&mad_agent_priv->lock, flags);
 925         list_add_tail(&local->completion_list, &mad_agent_priv->local_list);
 926         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
 927         queue_work(mad_agent_priv->qp_info->port_priv->wq,
 928                    &mad_agent_priv->local_work);
 929         ret = 1;
 930 out:
 931         return ret;
 932 }
 933
 934 static int get_pad_size(int hdr_len, int data_len, size_t mad_size)
 935 {
 936         int seg_size, pad;
 937
 938         seg_size = mad_size - hdr_len;
 939         if (data_len && seg_size) {
 940                 pad = seg_size - data_len % seg_size;
 941                 return pad == seg_size ? 0 : pad;
 942         } else
 943                 return seg_size;
 944 }
 945
 946 static void free_send_rmpp_list(struct ib_mad_send_wr_private *mad_send_wr)
 947 {
 948         struct ib_rmpp_segment *s, *t;
 949
 950         list_for_each_entry_safe(s, t, &mad_send_wr->rmpp_list, list) {
 951                 list_del(&s->list);
 952                 kfree(s);
 953         }
 954 }
 955
 956 static int alloc_send_rmpp_list(struct ib_mad_send_wr_private *send_wr,
 957                                 size_t mad_size, gfp_t gfp_mask)
 958 {
 959         struct ib_mad_send_buf *send_buf = &send_wr->send_buf;
 960         struct ib_rmpp_mad *rmpp_mad = send_buf->mad;
 961         struct ib_rmpp_segment *seg = NULL;
 962         int left, seg_size, pad;
 963
 964         send_buf->seg_size = mad_size - send_buf->hdr_len;
 965         send_buf->seg_rmpp_size = mad_size - IB_MGMT_RMPP_HDR;
 966         seg_size = send_buf->seg_size;
 967         pad = send_wr->pad;
 968
 969         /* Allocate data segments. */
 970         for (left = send_buf->data_len + pad; left > 0; left -= seg_size) {
 971                 seg = kmalloc(sizeof (*seg) + seg_size, gfp_mask);
 972                 if (!seg) {
 973                         free_send_rmpp_list(send_wr);
 974                         return -ENOMEM;
 975                 }
 976                 seg->num = ++send_buf->seg_count;
 977                 list_add_tail(&seg->list, &send_wr->rmpp_list);
 978         }
 979
 980         /* Zero any padding */
 981         if (pad)
 982                 memset(seg->data + seg_size - pad, 0, pad);
 983
 984         rmpp_mad->rmpp_hdr.rmpp_version = send_wr->mad_agent_priv->
 985                                           agent.rmpp_version;
 986         rmpp_mad->rmpp_hdr.rmpp_type = IB_MGMT_RMPP_TYPE_DATA;
 987         ib_set_rmpp_flags(&rmpp_mad->rmpp_hdr, IB_MGMT_RMPP_FLAG_ACTIVE);
 988
 989         send_wr->cur_seg = container_of(send_wr->rmpp_list.next,
 990                                         struct ib_rmpp_segment, list);
 991         send_wr->last_ack_seg = send_wr->cur_seg;
 992         return 0;
 993 }
 994
 995 int ib_mad_kernel_rmpp_agent(const struct ib_mad_agent *agent)
 996 {
 997         return agent->rmpp_version && !(agent->flags & IB_MAD_USER_RMPP);
 998 }
 999 EXPORT_SYMBOL(ib_mad_kernel_rmpp_agent);
1000
1001 struct ib_mad_send_buf * ib_create_send_mad(struct ib_mad_agent *mad_agent,
1002                                             u32 remote_qpn, u16 pkey_index,
1003                                             int rmpp_active,
1004                                             int hdr_len, int data_len,
1005                                             gfp_t gfp_mask,
1006                                             u8 base_version)
1007 {
1008         struct ib_mad_agent_private *mad_agent_priv;
1009         struct ib_mad_send_wr_private *mad_send_wr;
1010         int pad, message_size, ret, size;
1011         void *buf;
1012         size_t mad_size;
1013         bool opa;
1014
1015         mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
1016                                       agent);
1017
1018         opa = rdma_cap_opa_mad(mad_agent->device, mad_agent->port_num);
1019
1020         if (opa && base_version == OPA_MGMT_BASE_VERSION)
1021                 mad_size = sizeof(struct opa_mad);
1022         else
1023                 mad_size = sizeof(struct ib_mad);
1024
1025         pad = get_pad_size(hdr_len, data_len, mad_size);
1026         message_size = hdr_len + data_len + pad;
1027
1028         if (ib_mad_kernel_rmpp_agent(mad_agent)) {
1029                 if (!rmpp_active && message_size > mad_size)
1030                         return ERR_PTR(-EINVAL);
1031         } else
1032                 if (rmpp_active || message_size > mad_size)
1033                         return ERR_PTR(-EINVAL);
1034
1035         size = rmpp_active ? hdr_len : mad_size;
1036         buf = kzalloc(sizeof *mad_send_wr + size, gfp_mask);
1037         if (!buf)
1038                 return ERR_PTR(-ENOMEM);
1039
1040         mad_send_wr = buf + size;
1041         INIT_LIST_HEAD(&mad_send_wr->rmpp_list);
1042         mad_send_wr->send_buf.mad = buf;
1043         mad_send_wr->send_buf.hdr_len = hdr_len;
1044         mad_send_wr->send_buf.data_len = data_len;
1045         mad_send_wr->pad = pad;
1046
1047         mad_send_wr->mad_agent_priv = mad_agent_priv;
1048         mad_send_wr->sg_list[0].length = hdr_len;
1049         mad_send_wr->sg_list[0].lkey = mad_agent->qp->pd->local_dma_lkey;
1050
1051         /* OPA MADs don't have to be the full 2048 bytes */
1052         if (opa && base_version == OPA_MGMT_BASE_VERSION &&
1053             data_len < mad_size - hdr_len)
1054                 mad_send_wr->sg_list[1].length = data_len;
1055         else
1056                 mad_send_wr->sg_list[1].length = mad_size - hdr_len;
1057
1058         mad_send_wr->sg_list[1].lkey = mad_agent->qp->pd->local_dma_lkey;
1059
1060         mad_send_wr->mad_list.cqe.done = ib_mad_send_done;
1061
1062         mad_send_wr->send_wr.wr.wr_cqe = &mad_send_wr->mad_list.cqe;
1063         mad_send_wr->send_wr.wr.sg_list = mad_send_wr->sg_list;
1064         mad_send_wr->send_wr.wr.num_sge = 2;
1065         mad_send_wr->send_wr.wr.opcode = IB_WR_SEND;
1066         mad_send_wr->send_wr.wr.send_flags = IB_SEND_SIGNALED;
1067         mad_send_wr->send_wr.remote_qpn = remote_qpn;
1068         mad_send_wr->send_wr.remote_qkey = IB_QP_SET_QKEY;
1069         mad_send_wr->send_wr.pkey_index = pkey_index;
1070
1071         if (rmpp_active) {
1072                 ret = alloc_send_rmpp_list(mad_send_wr, mad_size, gfp_mask);
1073                 if (ret) {
1074                         kfree(buf);
1075                         return ERR_PTR(ret);
1076                 }
1077         }
1078
1079         mad_send_wr->send_buf.mad_agent = mad_agent;
1080         atomic_inc(&mad_agent_priv->refcount);
1081         return &mad_send_wr->send_buf;
1082 }
1083 EXPORT_SYMBOL(ib_create_send_mad);
1084
1085 int ib_get_mad_data_offset(u8 mgmt_class)
1086 {
1087         if (mgmt_class == IB_MGMT_CLASS_SUBN_ADM)
1088                 return IB_MGMT_SA_HDR;
1089         else if ((mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
1090                  (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
1091                  (mgmt_class == IB_MGMT_CLASS_BIS))
1092                 return IB_MGMT_DEVICE_HDR;
1093         else if ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
1094                  (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END))
1095                 return IB_MGMT_VENDOR_HDR;
1096         else
1097                 return IB_MGMT_MAD_HDR;
1098 }
1099 EXPORT_SYMBOL(ib_get_mad_data_offset);
1100
1101 int ib_is_mad_class_rmpp(u8 mgmt_class)
1102 {
1103         if ((mgmt_class == IB_MGMT_CLASS_SUBN_ADM) ||
1104             (mgmt_class == IB_MGMT_CLASS_DEVICE_MGMT) ||
1105             (mgmt_class == IB_MGMT_CLASS_DEVICE_ADM) ||
1106             (mgmt_class == IB_MGMT_CLASS_BIS) ||
1107             ((mgmt_class >= IB_MGMT_CLASS_VENDOR_RANGE2_START) &&
1108              (mgmt_class <= IB_MGMT_CLASS_VENDOR_RANGE2_END)))
1109                 return 1;
1110         return 0;
1111 }
1112 EXPORT_SYMBOL(ib_is_mad_class_rmpp);
1113
1114 void *ib_get_rmpp_segment(struct ib_mad_send_buf *send_buf, int seg_num)
1115 {
1116         struct ib_mad_send_wr_private *mad_send_wr;
1117         struct list_head *list;
1118
1119         mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
1120                                    send_buf);
1121         list = &mad_send_wr->cur_seg->list;
1122
1123         if (mad_send_wr->cur_seg->num < seg_num) {
1124                 list_for_each_entry(mad_send_wr->cur_seg, list, list)
1125                         if (mad_send_wr->cur_seg->num == seg_num)
1126                                 break;
1127         } else if (mad_send_wr->cur_seg->num > seg_num) {
1128                 list_for_each_entry_reverse(mad_send_wr->cur_seg, list, list)
1129                         if (mad_send_wr->cur_seg->num == seg_num)
1130                                 break;
1131         }
1132         return mad_send_wr->cur_seg->data;
1133 }
1134 EXPORT_SYMBOL(ib_get_rmpp_segment);
1135
1136 static inline void *ib_get_payload(struct ib_mad_send_wr_private *mad_send_wr)
1137 {
1138         if (mad_send_wr->send_buf.seg_count)
1139                 return ib_get_rmpp_segment(&mad_send_wr->send_buf,
1140                                            mad_send_wr->seg_num);
1141         else
1142                 return mad_send_wr->send_buf.mad +
1143                        mad_send_wr->send_buf.hdr_len;
1144 }
1145
1146 void ib_free_send_mad(struct ib_mad_send_buf *send_buf)
1147 {
1148         struct ib_mad_agent_private *mad_agent_priv;
1149         struct ib_mad_send_wr_private *mad_send_wr;
1150
1151         mad_agent_priv = container_of(send_buf->mad_agent,
1152                                       struct ib_mad_agent_private, agent);
1153         mad_send_wr = container_of(send_buf, struct ib_mad_send_wr_private,
1154                                    send_buf);
1155
1156         free_send_rmpp_list(mad_send_wr);
1157         kfree(send_buf->mad);
1158         deref_mad_agent(mad_agent_priv);
1159 }
1160 EXPORT_SYMBOL(ib_free_send_mad);
1161
1162 int ib_send_mad(struct ib_mad_send_wr_private *mad_send_wr)
1163 {
1164         struct ib_mad_qp_info *qp_info;
1165         struct list_head *list;
1166         struct ib_send_wr *bad_send_wr;
1167         struct ib_mad_agent *mad_agent;
1168         struct ib_sge *sge;
1169         unsigned long flags;
1170         int ret;
1171
1172         /* Set WR ID to find mad_send_wr upon completion */
1173         qp_info = mad_send_wr->mad_agent_priv->qp_info;
1174         mad_send_wr->mad_list.mad_queue = &qp_info->send_queue;
1175         mad_send_wr->mad_list.cqe.done = ib_mad_send_done;
1176         mad_send_wr->send_wr.wr.wr_cqe = &mad_send_wr->mad_list.cqe;
1177
1178         mad_agent = mad_send_wr->send_buf.mad_agent;
1179         sge = mad_send_wr->sg_list;
1180         sge[0].addr = ib_dma_map_single(mad_agent->device,
1181                                         mad_send_wr->send_buf.mad,
1182                                         sge[0].length,
1183                                         DMA_TO_DEVICE);
1184         if (unlikely(ib_dma_mapping_error(mad_agent->device, sge[0].addr)))
1185                 return -ENOMEM;
1186
1187         mad_send_wr->header_mapping = sge[0].addr;
1188
1189         sge[1].addr = ib_dma_map_single(mad_agent->device,
1190                                         ib_get_payload(mad_send_wr),
1191                                         sge[1].length,
1192                                         DMA_TO_DEVICE);
1193         if (unlikely(ib_dma_mapping_error(mad_agent->device, sge[1].addr))) {
1194                 ib_dma_unmap_single(mad_agent->device,
1195                                     mad_send_wr->header_mapping,
1196                                     sge[0].length, DMA_TO_DEVICE);
1197                 return -ENOMEM;
1198         }
1199         mad_send_wr->payload_mapping = sge[1].addr;
1200
1201         spin_lock_irqsave(&qp_info->send_queue.lock, flags);
1202         if (qp_info->send_queue.count < qp_info->send_queue.max_active) {
1203                 ret = ib_post_send(mad_agent->qp, &mad_send_wr->send_wr.wr,
1204                                    &bad_send_wr);
1205                 list = &qp_info->send_queue.list;
1206         } else {
1207                 ret = 0;
1208                 list = &qp_info->overflow_list;
1209         }
1210
1211         if (!ret) {
1212                 qp_info->send_queue.count++;
1213                 list_add_tail(&mad_send_wr->mad_list.list, list);
1214         }
1215         spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
1216         if (ret) {
1217                 ib_dma_unmap_single(mad_agent->device,
1218                                     mad_send_wr->header_mapping,
1219                                     sge[0].length, DMA_TO_DEVICE);
1220                 ib_dma_unmap_single(mad_agent->device,
1221                                     mad_send_wr->payload_mapping,
1222                                     sge[1].length, DMA_TO_DEVICE);
1223         }
1224         return ret;
1225 }
1226
1227 /*
1228  * ib_post_send_mad - Posts MAD(s) to the send queue of the QP associated
1229  *  with the registered client
1230  */
1231 int ib_post_send_mad(struct ib_mad_send_buf *send_buf,
1232                      struct ib_mad_send_buf **bad_send_buf)
1233 {
1234         struct ib_mad_agent_private *mad_agent_priv;
1235         struct ib_mad_send_buf *next_send_buf;
1236         struct ib_mad_send_wr_private *mad_send_wr;
1237         unsigned long flags;
1238         int ret = -EINVAL;
1239
1240         /* Walk list of send WRs and post each on send list */
1241         for (; send_buf; send_buf = next_send_buf) {
1242                 mad_send_wr = container_of(send_buf,
1243                                            struct ib_mad_send_wr_private,
1244                                            send_buf);
1245                 mad_agent_priv = mad_send_wr->mad_agent_priv;
1246
1247                 ret = ib_mad_enforce_security(mad_agent_priv,
1248                                               mad_send_wr->send_wr.pkey_index);
1249                 if (ret)
1250                         goto error;
1251
1252                 if (!send_buf->mad_agent->send_handler ||
1253                     (send_buf->timeout_ms &&
1254                      !send_buf->mad_agent->recv_handler)) {
1255                         ret = -EINVAL;
1256                         goto error;
1257                 }
1258
1259                 if (!ib_is_mad_class_rmpp(((struct ib_mad_hdr *) send_buf->mad)->mgmt_class)) {
1260                         if (mad_agent_priv->agent.rmpp_version) {
1261                                 ret = -EINVAL;
1262                                 goto error;
1263                         }
1264                 }
1265
1266                 /*
1267                  * Save pointer to next work request to post in case the
1268                  * current one completes, and the user modifies the work
1269                  * request associated with the completion
1270                  */
1271                 next_send_buf = send_buf->next;
1272                 mad_send_wr->send_wr.ah = send_buf->ah;
1273
1274                 if (((struct ib_mad_hdr *) send_buf->mad)->mgmt_class ==
1275                     IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
1276                         ret = handle_outgoing_dr_smp(mad_agent_priv,
1277                                                      mad_send_wr);
1278                         if (ret < 0)            /* error */
1279                                 goto error;
1280                         else if (ret == 1)      /* locally consumed */
1281                                 continue;
1282                 }
1283
1284                 mad_send_wr->tid = ((struct ib_mad_hdr *) send_buf->mad)->tid;
1285                 /* Timeout will be updated after send completes */
1286                 mad_send_wr->timeout = msecs_to_jiffies(send_buf->timeout_ms);
1287                 mad_send_wr->max_retries = send_buf->retries;
1288                 mad_send_wr->retries_left = send_buf->retries;
1289                 send_buf->retries = 0;
1290                 /* Reference for work request to QP + response */
1291                 mad_send_wr->refcount = 1 + (mad_send_wr->timeout > 0);
1292                 mad_send_wr->status = IB_WC_SUCCESS;
1293
1294                 /* Reference MAD agent until send completes */
1295                 atomic_inc(&mad_agent_priv->refcount);
1296                 spin_lock_irqsave(&mad_agent_priv->lock, flags);
1297                 list_add_tail(&mad_send_wr->agent_list,
1298                               &mad_agent_priv->send_list);
1299                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1300
1301                 if (ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)) {
1302                         ret = ib_send_rmpp_mad(mad_send_wr);
1303                         if (ret >= 0 && ret != IB_RMPP_RESULT_CONSUMED)
1304                                 ret = ib_send_mad(mad_send_wr);
1305                 } else
1306                         ret = ib_send_mad(mad_send_wr);
1307                 if (ret < 0) {
1308                         /* Fail send request */
1309                         spin_lock_irqsave(&mad_agent_priv->lock, flags);
1310                         list_del(&mad_send_wr->agent_list);
1311                         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
1312                         atomic_dec(&mad_agent_priv->refcount);
1313                         goto error;
1314                 }
1315         }
1316         return 0;
1317 error:
1318         if (bad_send_buf)
1319                 *bad_send_buf = send_buf;
1320         return ret;
1321 }
1322 EXPORT_SYMBOL(ib_post_send_mad);
1323
1324 /*
1325  * ib_free_recv_mad - Returns data buffers used to receive
1326  *  a MAD to the access layer
1327  */
1328 void ib_free_recv_mad(struct ib_mad_recv_wc *mad_recv_wc)
1329 {
1330         struct ib_mad_recv_buf *mad_recv_buf, *temp_recv_buf;
1331         struct ib_mad_private_header *mad_priv_hdr;
1332         struct ib_mad_private *priv;
1333         struct list_head free_list;
1334
1335         INIT_LIST_HEAD(&free_list);
1336         list_splice_init(&mad_recv_wc->rmpp_list, &free_list);
1337
1338         list_for_each_entry_safe(mad_recv_buf, temp_recv_buf,
1339                                         &free_list, list) {
1340                 mad_recv_wc = container_of(mad_recv_buf, struct ib_mad_recv_wc,
1341                                            recv_buf);
1342                 mad_priv_hdr = container_of(mad_recv_wc,
1343                                             struct ib_mad_private_header,
1344                                             recv_wc);
1345                 priv = container_of(mad_priv_hdr, struct ib_mad_private,
1346                                     header);
1347                 kfree(priv);
1348         }
1349 }
1350 EXPORT_SYMBOL(ib_free_recv_mad);
1351
1352 struct ib_mad_agent *ib_redirect_mad_qp(struct ib_qp *qp,
1353                                         u8 rmpp_version,
1354                                         ib_mad_send_handler send_handler,
1355                                         ib_mad_recv_handler recv_handler,
1356                                         void *context)
1357 {
1358         return ERR_PTR(-EINVAL);        /* XXX: for now */
1359 }
1360 EXPORT_SYMBOL(ib_redirect_mad_qp);
1361
1362 int ib_process_mad_wc(struct ib_mad_agent *mad_agent,
1363                       struct ib_wc *wc)
1364 {
1365         dev_err(&mad_agent->device->dev,
1366                 "ib_process_mad_wc() not implemented yet\n");
1367         return 0;
1368 }
1369 EXPORT_SYMBOL(ib_process_mad_wc);
1370
1371 static int method_in_use(struct ib_mad_mgmt_method_table **method,
1372                          struct ib_mad_reg_req *mad_reg_req)
1373 {
1374         int i;
1375
1376         for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS) {
1377                 if ((*method)->agent[i]) {
1378                         pr_err("Method %d already in use\n", i);
1379                         return -EINVAL;
1380                 }
1381         }
1382         return 0;
1383 }
1384
1385 static int allocate_method_table(struct ib_mad_mgmt_method_table **method)
1386 {
1387         /* Allocate management method table */
1388         *method = kzalloc(sizeof **method, GFP_ATOMIC);
1389         return (*method) ? 0 : (-ENOMEM);
1390 }
1391
1392 /*
1393  * Check to see if there are any methods still in use
1394  */
1395 static int check_method_table(struct ib_mad_mgmt_method_table *method)
1396 {
1397         int i;
1398
1399         for (i = 0; i < IB_MGMT_MAX_METHODS; i++)
1400                 if (method->agent[i])
1401                         return 1;
1402         return 0;
1403 }
1404
1405 /*
1406  * Check to see if there are any method tables for this class still in use
1407  */
1408 static int check_class_table(struct ib_mad_mgmt_class_table *class)
1409 {
1410         int i;
1411
1412         for (i = 0; i < MAX_MGMT_CLASS; i++)
1413                 if (class->method_table[i])
1414                         return 1;
1415         return 0;
1416 }
1417
1418 static int check_vendor_class(struct ib_mad_mgmt_vendor_class *vendor_class)
1419 {
1420         int i;
1421
1422         for (i = 0; i < MAX_MGMT_OUI; i++)
1423                 if (vendor_class->method_table[i])
1424                         return 1;
1425         return 0;
1426 }
1427
1428 static int find_vendor_oui(struct ib_mad_mgmt_vendor_class *vendor_class,
1429                            const char *oui)
1430 {
1431         int i;
1432
1433         for (i = 0; i < MAX_MGMT_OUI; i++)
1434                 /* Is there matching OUI for this vendor class ? */
1435                 if (!memcmp(vendor_class->oui[i], oui, 3))
1436                         return i;
1437
1438         return -1;
1439 }
1440
1441 static int check_vendor_table(struct ib_mad_mgmt_vendor_class_table *vendor)
1442 {
1443         int i;
1444
1445         for (i = 0; i < MAX_MGMT_VENDOR_RANGE2; i++)
1446                 if (vendor->vendor_class[i])
1447                         return 1;
1448
1449         return 0;
1450 }
1451
1452 static void remove_methods_mad_agent(struct ib_mad_mgmt_method_table *method,
1453                                      struct ib_mad_agent_private *agent)
1454 {
1455         int i;
1456
1457         /* Remove any methods for this mad agent */
1458         for (i = 0; i < IB_MGMT_MAX_METHODS; i++) {
1459                 if (method->agent[i] == agent) {
1460                         method->agent[i] = NULL;
1461                 }
1462         }
1463 }
1464
1465 static int add_nonoui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1466                               struct ib_mad_agent_private *agent_priv,
1467                               u8 mgmt_class)
1468 {
1469         struct ib_mad_port_private *port_priv;
1470         struct ib_mad_mgmt_class_table **class;
1471         struct ib_mad_mgmt_method_table **method;
1472         int i, ret;
1473
1474         port_priv = agent_priv->qp_info->port_priv;
1475         class = &port_priv->version[mad_reg_req->mgmt_class_version].class;
1476         if (!*class) {
1477                 /* Allocate management class table for "new" class version */
1478                 *class = kzalloc(sizeof **class, GFP_ATOMIC);
1479                 if (!*class) {
1480                         ret = -ENOMEM;
1481                         goto error1;
1482                 }
1483
1484                 /* Allocate method table for this management class */
1485                 method = &(*class)->method_table[mgmt_class];
1486                 if ((ret = allocate_method_table(method)))
1487                         goto error2;
1488         } else {
1489                 method = &(*class)->method_table[mgmt_class];
1490                 if (!*method) {
1491                         /* Allocate method table for this management class */
1492                         if ((ret = allocate_method_table(method)))
1493                                 goto error1;
1494                 }
1495         }
1496
1497         /* Now, make sure methods are not already in use */
1498         if (method_in_use(method, mad_reg_req))
1499                 goto error3;
1500
1501         /* Finally, add in methods being registered */
1502         for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS)
1503                 (*method)->agent[i] = agent_priv;
1504
1505         return 0;
1506
1507 error3:
1508         /* Remove any methods for this mad agent */
1509         remove_methods_mad_agent(*method, agent_priv);
1510         /* Now, check to see if there are any methods in use */
1511         if (!check_method_table(*method)) {
1512                 /* If not, release management method table */
1513                 kfree(*method);
1514                 *method = NULL;
1515         }
1516         ret = -EINVAL;
1517         goto error1;
1518 error2:
1519         kfree(*class);
1520         *class = NULL;
1521 error1:
1522         return ret;
1523 }
1524
1525 static int add_oui_reg_req(struct ib_mad_reg_req *mad_reg_req,
1526                            struct ib_mad_agent_private *agent_priv)
1527 {
1528         struct ib_mad_port_private *port_priv;
1529         struct ib_mad_mgmt_vendor_class_table **vendor_table;
1530         struct ib_mad_mgmt_vendor_class_table *vendor = NULL;
1531         struct ib_mad_mgmt_vendor_class *vendor_class = NULL;
1532         struct ib_mad_mgmt_method_table **method;
1533         int i, ret = -ENOMEM;
1534         u8 vclass;
1535
1536         /* "New" vendor (with OUI) class */
1537         vclass = vendor_class_index(mad_reg_req->mgmt_class);
1538         port_priv = agent_priv->qp_info->port_priv;
1539         vendor_table = &port_priv->version[
1540                                 mad_reg_req->mgmt_class_version].vendor;
1541         if (!*vendor_table) {
1542                 /* Allocate mgmt vendor class table for "new" class version */
1543                 vendor = kzalloc(sizeof *vendor, GFP_ATOMIC);
1544                 if (!vendor)
1545                         goto error1;
1546
1547                 *vendor_table = vendor;
1548         }
1549         if (!(*vendor_table)->vendor_class[vclass]) {
1550                 /* Allocate table for this management vendor class */
1551                 vendor_class = kzalloc(sizeof *vendor_class, GFP_ATOMIC);
1552                 if (!vendor_class)
1553                         goto error2;
1554
1555                 (*vendor_table)->vendor_class[vclass] = vendor_class;
1556         }
1557         for (i = 0; i < MAX_MGMT_OUI; i++) {
1558                 /* Is there matching OUI for this vendor class ? */
1559                 if (!memcmp((*vendor_table)->vendor_class[vclass]->oui[i],
1560                             mad_reg_req->oui, 3)) {
1561                         method = &(*vendor_table)->vendor_class[
1562                                                 vclass]->method_table[i];
1563                         if (!*method)
1564                                 goto error3;
1565                         goto check_in_use;
1566                 }
1567         }
1568         for (i = 0; i < MAX_MGMT_OUI; i++) {
1569                 /* OUI slot available ? */
1570                 if (!is_vendor_oui((*vendor_table)->vendor_class[
1571                                 vclass]->oui[i])) {
1572                         method = &(*vendor_table)->vendor_class[
1573                                 vclass]->method_table[i];
1574                         /* Allocate method table for this OUI */
1575                         if (!*method) {
1576                                 ret = allocate_method_table(method);
1577                                 if (ret)
1578                                         goto error3;
1579                         }
1580                         memcpy((*vendor_table)->vendor_class[vclass]->oui[i],
1581                                mad_reg_req->oui, 3);
1582                         goto check_in_use;
1583                 }
1584         }
1585         dev_err(&agent_priv->agent.device->dev, "All OUI slots in use\n");
1586         goto error3;
1587
1588 check_in_use:
1589         /* Now, make sure methods are not already in use */
1590         if (method_in_use(method, mad_reg_req))
1591                 goto error4;
1592
1593         /* Finally, add in methods being registered */
1594         for_each_set_bit(i, mad_reg_req->method_mask, IB_MGMT_MAX_METHODS)
1595                 (*method)->agent[i] = agent_priv;
1596
1597         return 0;
1598
1599 error4:
1600         /* Remove any methods for this mad agent */
1601         remove_methods_mad_agent(*method, agent_priv);
1602         /* Now, check to see if there are any methods in use */
1603         if (!check_method_table(*method)) {
1604                 /* If not, release management method table */
1605                 kfree(*method);
1606                 *method = NULL;
1607         }
1608         ret = -EINVAL;
1609 error3:
1610         if (vendor_class) {
1611                 (*vendor_table)->vendor_class[vclass] = NULL;
1612                 kfree(vendor_class);
1613         }
1614 error2:
1615         if (vendor) {
1616                 *vendor_table = NULL;
1617                 kfree(vendor);
1618         }
1619 error1:
1620         return ret;
1621 }
1622
1623 static void remove_mad_reg_req(struct ib_mad_agent_private *agent_priv)
1624 {
1625         struct ib_mad_port_private *port_priv;
1626         struct ib_mad_mgmt_class_table *class;
1627         struct ib_mad_mgmt_method_table *method;
1628         struct ib_mad_mgmt_vendor_class_table *vendor;
1629         struct ib_mad_mgmt_vendor_class *vendor_class;
1630         int index;
1631         u8 mgmt_class;
1632
1633         /*
1634          * Was MAD registration request supplied
1635          * with original registration ?
1636          */
1637         if (!agent_priv->reg_req) {
1638                 goto out;
1639         }
1640
1641         port_priv = agent_priv->qp_info->port_priv;
1642         mgmt_class = convert_mgmt_class(agent_priv->reg_req->mgmt_class);
1643         class = port_priv->version[
1644                         agent_priv->reg_req->mgmt_class_version].class;
1645         if (!class)
1646                 goto vendor_check;
1647
1648         method = class->method_table[mgmt_class];
1649         if (method) {
1650                 /* Remove any methods for this mad agent */
1651                 remove_methods_mad_agent(method, agent_priv);
1652                 /* Now, check to see if there are any methods still in use */
1653                 if (!check_method_table(method)) {
1654                         /* If not, release management method table */
1655                         kfree(method);
1656                         class->method_table[mgmt_class] = NULL;
1657                         /* Any management classes left ? */
1658                         if (!check_class_table(class)) {
1659                                 /* If not, release management class table */
1660                                 kfree(class);
1661                                 port_priv->version[
1662                                         agent_priv->reg_req->
1663                                         mgmt_class_version].class = NULL;
1664                         }
1665                 }
1666         }
1667
1668 vendor_check:
1669         if (!is_vendor_class(mgmt_class))
1670                 goto out;
1671
1672         /* normalize mgmt_class to vendor range 2 */
1673         mgmt_class = vendor_class_index(agent_priv->reg_req->mgmt_class);
1674         vendor = port_priv->version[
1675                         agent_priv->reg_req->mgmt_class_version].vendor;
1676
1677         if (!vendor)
1678                 goto out;
1679
1680         vendor_class = vendor->vendor_class[mgmt_class];
1681         if (vendor_class) {
1682                 index = find_vendor_oui(vendor_class, agent_priv->reg_req->oui);
1683                 if (index < 0)
1684                         goto out;
1685                 method = vendor_class->method_table[index];
1686                 if (method) {
1687                         /* Remove any methods for this mad agent */
1688                         remove_methods_mad_agent(method, agent_priv);
1689                         /*
1690                          * Now, check to see if there are
1691                          * any methods still in use
1692                          */
1693                         if (!check_method_table(method)) {
1694                                 /* If not, release management method table */
1695                                 kfree(method);
1696                                 vendor_class->method_table[index] = NULL;
1697                                 memset(vendor_class->oui[index], 0, 3);
1698                                 /* Any OUIs left ? */
1699                                 if (!check_vendor_class(vendor_class)) {
1700                                         /* If not, release vendor class table */
1701                                         kfree(vendor_class);
1702                                         vendor->vendor_class[mgmt_class] = NULL;
1703                                         /* Any other vendor classes left ? */
1704                                         if (!check_vendor_table(vendor)) {
1705                                                 kfree(vendor);
1706                                                 port_priv->version[
1707                                                         agent_priv->reg_req->
1708                                                         mgmt_class_version].
1709                                                         vendor = NULL;
1710                                         }
1711                                 }
1712                         }
1713                 }
1714         }
1715
1716 out:
1717         return;
1718 }
1719
1720 static struct ib_mad_agent_private *
1721 find_mad_agent(struct ib_mad_port_private *port_priv,
1722                const struct ib_mad_hdr *mad_hdr)
1723 {
1724         struct ib_mad_agent_private *mad_agent = NULL;
1725         unsigned long flags;
1726
1727         spin_lock_irqsave(&port_priv->reg_lock, flags);
1728         if (ib_response_mad(mad_hdr)) {
1729                 u32 hi_tid;
1730                 struct ib_mad_agent_private *entry;
1731
1732                 /*
1733                  * Routing is based on high 32 bits of transaction ID
1734                  * of MAD.
1735                  */
1736                 hi_tid = be64_to_cpu(mad_hdr->tid) >> 32;
1737                 list_for_each_entry(entry, &port_priv->agent_list, agent_list) {
1738                         if (entry->agent.hi_tid == hi_tid) {
1739                                 mad_agent = entry;
1740                                 break;
1741                         }
1742                 }
1743         } else {
1744                 struct ib_mad_mgmt_class_table *class;
1745                 struct ib_mad_mgmt_method_table *method;
1746                 struct ib_mad_mgmt_vendor_class_table *vendor;
1747                 struct ib_mad_mgmt_vendor_class *vendor_class;
1748                 const struct ib_vendor_mad *vendor_mad;
1749                 int index;
1750
1751                 /*
1752                  * Routing is based on version, class, and method
1753                  * For "newer" vendor MADs, also based on OUI
1754                  */
1755                 if (mad_hdr->class_version >= MAX_MGMT_VERSION)
1756                         goto out;
1757                 if (!is_vendor_class(mad_hdr->mgmt_class)) {
1758                         class = port_priv->version[
1759                                         mad_hdr->class_version].class;
1760                         if (!class)
1761                                 goto out;
1762                         if (convert_mgmt_class(mad_hdr->mgmt_class) >=
1763                             ARRAY_SIZE(class->method_table))
1764                                 goto out;
1765                         method = class->method_table[convert_mgmt_class(
1766                                                         mad_hdr->mgmt_class)];
1767                         if (method)
1768                                 mad_agent = method->agent[mad_hdr->method &
1769                                                           ~IB_MGMT_METHOD_RESP];
1770                 } else {
1771                         vendor = port_priv->version[
1772                                         mad_hdr->class_version].vendor;
1773                         if (!vendor)
1774                                 goto out;
1775                         vendor_class = vendor->vendor_class[vendor_class_index(
1776                                                 mad_hdr->mgmt_class)];
1777                         if (!vendor_class)
1778                                 goto out;
1779                         /* Find matching OUI */
1780                         vendor_mad = (const struct ib_vendor_mad *)mad_hdr;
1781                         index = find_vendor_oui(vendor_class, vendor_mad->oui);
1782                         if (index == -1)
1783                                 goto out;
1784                         method = vendor_class->method_table[index];
1785                         if (method) {
1786                                 mad_agent = method->agent[mad_hdr->method &
1787                                                           ~IB_MGMT_METHOD_RESP];
1788                         }
1789                 }
1790         }
1791
1792         if (mad_agent) {
1793                 if (mad_agent->agent.recv_handler)
1794                         atomic_inc(&mad_agent->refcount);
1795                 else {
1796                         dev_notice(&port_priv->device->dev,
1797                                    "No receive handler for client %p on port %d\n",
1798                                    &mad_agent->agent, port_priv->port_num);
1799                         mad_agent = NULL;
1800                 }
1801         }
1802 out:
1803         spin_unlock_irqrestore(&port_priv->reg_lock, flags);
1804
1805         return mad_agent;
1806 }
1807
1808 static int validate_mad(const struct ib_mad_hdr *mad_hdr,
1809                         const struct ib_mad_qp_info *qp_info,
1810                         bool opa)
1811 {
1812         int valid = 0;
1813         u32 qp_num = qp_info->qp->qp_num;
1814
1815         /* Make sure MAD base version is understood */
1816         if (mad_hdr->base_version != IB_MGMT_BASE_VERSION &&
1817             (!opa || mad_hdr->base_version != OPA_MGMT_BASE_VERSION)) {
1818                 pr_err("MAD received with unsupported base version %d %s\n",
1819                        mad_hdr->base_version, opa ? "(opa)" : "");
1820                 goto out;
1821         }
1822
1823         /* Filter SMI packets sent to other than QP0 */
1824         if ((mad_hdr->mgmt_class == IB_MGMT_CLASS_SUBN_LID_ROUTED) ||
1825             (mad_hdr->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)) {
1826                 if (qp_num == 0)
1827                         valid = 1;
1828         } else {
1829                 /* CM attributes other than ClassPortInfo only use Send method */
1830                 if ((mad_hdr->mgmt_class == IB_MGMT_CLASS_CM) &&
1831                     (mad_hdr->attr_id != IB_MGMT_CLASSPORTINFO_ATTR_ID) &&
1832                     (mad_hdr->method != IB_MGMT_METHOD_SEND))
1833                         goto out;
1834                 /* Filter GSI packets sent to QP0 */
1835                 if (qp_num != 0)
1836                         valid = 1;
1837         }
1838
1839 out:
1840         return valid;
1841 }
1842
1843 static int is_rmpp_data_mad(const struct ib_mad_agent_private *mad_agent_priv,
1844                             const struct ib_mad_hdr *mad_hdr)
1845 {
1846         struct ib_rmpp_mad *rmpp_mad;
1847
1848         rmpp_mad = (struct ib_rmpp_mad *)mad_hdr;
1849         return !mad_agent_priv->agent.rmpp_version ||
1850                 !ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent) ||
1851                 !(ib_get_rmpp_flags(&rmpp_mad->rmpp_hdr) &
1852                                     IB_MGMT_RMPP_FLAG_ACTIVE) ||
1853                 (rmpp_mad->rmpp_hdr.rmpp_type == IB_MGMT_RMPP_TYPE_DATA);
1854 }
1855
1856 static inline int rcv_has_same_class(const struct ib_mad_send_wr_private *wr,
1857                                      const struct ib_mad_recv_wc *rwc)
1858 {
1859         return ((struct ib_mad_hdr *)(wr->send_buf.mad))->mgmt_class ==
1860                 rwc->recv_buf.mad->mad_hdr.mgmt_class;
1861 }
1862
1863 static inline int rcv_has_same_gid(const struct ib_mad_agent_private *mad_agent_priv,
1864                                    const struct ib_mad_send_wr_private *wr,
1865                                    const struct ib_mad_recv_wc *rwc )
1866 {
1867         struct rdma_ah_attr attr;
1868         u8 send_resp, rcv_resp;
1869         union ib_gid sgid;
1870         struct ib_device *device = mad_agent_priv->agent.device;
1871         u8 port_num = mad_agent_priv->agent.port_num;
1872         u8 lmc;
1873         bool has_grh;
1874
1875         send_resp = ib_response_mad((struct ib_mad_hdr *)wr->send_buf.mad);
1876         rcv_resp = ib_response_mad(&rwc->recv_buf.mad->mad_hdr);
1877
1878         if (send_resp == rcv_resp)
1879                 /* both requests, or both responses. GIDs different */
1880                 return 0;
1881
1882         if (rdma_query_ah(wr->send_buf.ah, &attr))
1883                 /* Assume not equal, to avoid false positives. */
1884                 return 0;
1885
1886         has_grh = !!(rdma_ah_get_ah_flags(&attr) & IB_AH_GRH);
1887         if (has_grh != !!(rwc->wc->wc_flags & IB_WC_GRH))
1888                 /* one has GID, other does not.  Assume different */
1889                 return 0;
1890
1891         if (!send_resp && rcv_resp) {
1892                 /* is request/response. */
1893                 if (!has_grh) {
1894                         if (ib_get_cached_lmc(device, port_num, &lmc))
1895                                 return 0;
1896                         return (!lmc || !((rdma_ah_get_path_bits(&attr) ^
1897                                            rwc->wc->dlid_path_bits) &
1898                                           ((1 << lmc) - 1)));
1899                 } else {
1900                         const struct ib_global_route *grh =
1901                                         rdma_ah_read_grh(&attr);
1902
1903                         if (ib_get_cached_gid(device, port_num,
1904                                               grh->sgid_index, &sgid, NULL))
1905                                 return 0;
1906                         return !memcmp(sgid.raw, rwc->recv_buf.grh->dgid.raw,
1907                                        16);
1908                 }
1909         }
1910
1911         if (!has_grh)
1912                 return rdma_ah_get_dlid(&attr) == rwc->wc->slid;
1913         else
1914                 return !memcmp(rdma_ah_read_grh(&attr)->dgid.raw,
1915                                rwc->recv_buf.grh->sgid.raw,
1916                                16);
1917 }
1918
1919 static inline int is_direct(u8 class)
1920 {
1921         return (class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE);
1922 }
1923
1924 struct ib_mad_send_wr_private*
1925 ib_find_send_mad(const struct ib_mad_agent_private *mad_agent_priv,
1926                  const struct ib_mad_recv_wc *wc)
1927 {
1928         struct ib_mad_send_wr_private *wr;
1929         const struct ib_mad_hdr *mad_hdr;
1930
1931         mad_hdr = &wc->recv_buf.mad->mad_hdr;
1932
1933         list_for_each_entry(wr, &mad_agent_priv->wait_list, agent_list) {
1934                 if ((wr->tid == mad_hdr->tid) &&
1935                     rcv_has_same_class(wr, wc) &&
1936                     /*
1937                      * Don't check GID for direct routed MADs.
1938                      * These might have permissive LIDs.
1939                      */
1940                     (is_direct(mad_hdr->mgmt_class) ||
1941                      rcv_has_same_gid(mad_agent_priv, wr, wc)))
1942                         return (wr->status == IB_WC_SUCCESS) ? wr : NULL;
1943         }
1944
1945         /*
1946          * It's possible to receive the response before we've
1947          * been notified that the send has completed
1948          */
1949         list_for_each_entry(wr, &mad_agent_priv->send_list, agent_list) {
1950                 if (is_rmpp_data_mad(mad_agent_priv, wr->send_buf.mad) &&
1951                     wr->tid == mad_hdr->tid &&
1952                     wr->timeout &&
1953                     rcv_has_same_class(wr, wc) &&
1954                     /*
1955                      * Don't check GID for direct routed MADs.
1956                      * These might have permissive LIDs.
1957                      */
1958                     (is_direct(mad_hdr->mgmt_class) ||
1959                      rcv_has_same_gid(mad_agent_priv, wr, wc)))
1960                         /* Verify request has not been canceled */
1961                         return (wr->status == IB_WC_SUCCESS) ? wr : NULL;
1962         }
1963         return NULL;
1964 }
1965
1966 void ib_mark_mad_done(struct ib_mad_send_wr_private *mad_send_wr)
1967 {
1968         mad_send_wr->timeout = 0;
1969         if (mad_send_wr->refcount == 1)
1970                 list_move_tail(&mad_send_wr->agent_list,
1971                               &mad_send_wr->mad_agent_priv->done_list);
1972 }
1973
1974 static void ib_mad_complete_recv(struct ib_mad_agent_private *mad_agent_priv,
1975                                  struct ib_mad_recv_wc *mad_recv_wc)
1976 {
1977         struct ib_mad_send_wr_private *mad_send_wr;
1978         struct ib_mad_send_wc mad_send_wc;
1979         unsigned long flags;
1980         int ret;
1981
1982         INIT_LIST_HEAD(&mad_recv_wc->rmpp_list);
1983         ret = ib_mad_enforce_security(mad_agent_priv,
1984                                       mad_recv_wc->wc->pkey_index);
1985         if (ret) {
1986                 ib_free_recv_mad(mad_recv_wc);
1987                 deref_mad_agent(mad_agent_priv);
1988                 return;
1989         }
1990
1991         list_add(&mad_recv_wc->recv_buf.list, &mad_recv_wc->rmpp_list);
1992         if (ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)) {
1993                 mad_recv_wc = ib_process_rmpp_recv_wc(mad_agent_priv,
1994                                                       mad_recv_wc);
1995                 if (!mad_recv_wc) {
1996                         deref_mad_agent(mad_agent_priv);
1997                         return;
1998                 }
1999         }
2000
2001         /* Complete corresponding request */
2002         if (ib_response_mad(&mad_recv_wc->recv_buf.mad->mad_hdr)) {
2003                 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2004                 mad_send_wr = ib_find_send_mad(mad_agent_priv, mad_recv_wc);
2005                 if (!mad_send_wr) {
2006                         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2007                         if (!ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)
2008                            && ib_is_mad_class_rmpp(mad_recv_wc->recv_buf.mad->mad_hdr.mgmt_class)
2009                            && (ib_get_rmpp_flags(&((struct ib_rmpp_mad *)mad_recv_wc->recv_buf.mad)->rmpp_hdr)
2010                                         & IB_MGMT_RMPP_FLAG_ACTIVE)) {
2011                                 /* user rmpp is in effect
2012                                  * and this is an active RMPP MAD
2013                                  */
2014                                 mad_agent_priv->agent.recv_handler(
2015                                                 &mad_agent_priv->agent, NULL,
2016                                                 mad_recv_wc);
2017                                 atomic_dec(&mad_agent_priv->refcount);
2018                         } else {
2019                                 /* not user rmpp, revert to normal behavior and
2020                                  * drop the mad */
2021                                 ib_free_recv_mad(mad_recv_wc);
2022                                 deref_mad_agent(mad_agent_priv);
2023                                 return;
2024                         }
2025                 } else {
2026                         ib_mark_mad_done(mad_send_wr);
2027                         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2028
2029                         /* Defined behavior is to complete response before request */
2030                         mad_agent_priv->agent.recv_handler(
2031                                         &mad_agent_priv->agent,
2032                                         &mad_send_wr->send_buf,
2033                                         mad_recv_wc);
2034                         atomic_dec(&mad_agent_priv->refcount);
2035
2036                         mad_send_wc.status = IB_WC_SUCCESS;
2037                         mad_send_wc.vendor_err = 0;
2038                         mad_send_wc.send_buf = &mad_send_wr->send_buf;
2039                         ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
2040                 }
2041         } else {
2042                 mad_agent_priv->agent.recv_handler(&mad_agent_priv->agent, NULL,
2043                                                    mad_recv_wc);
2044                 deref_mad_agent(mad_agent_priv);
2045         }
2046
2047         return;
2048 }
2049
2050 static enum smi_action handle_ib_smi(const struct ib_mad_port_private *port_priv,
2051                                      const struct ib_mad_qp_info *qp_info,
2052                                      const struct ib_wc *wc,
2053                                      int port_num,
2054                                      struct ib_mad_private *recv,
2055                                      struct ib_mad_private *response)
2056 {
2057         enum smi_forward_action retsmi;
2058         struct ib_smp *smp = (struct ib_smp *)recv->mad;
2059
2060         if (smi_handle_dr_smp_recv(smp,
2061                                    rdma_cap_ib_switch(port_priv->device),
2062                                    port_num,
2063                                    port_priv->device->phys_port_cnt) ==
2064                                    IB_SMI_DISCARD)
2065                 return IB_SMI_DISCARD;
2066
2067         retsmi = smi_check_forward_dr_smp(smp);
2068         if (retsmi == IB_SMI_LOCAL)
2069                 return IB_SMI_HANDLE;
2070
2071         if (retsmi == IB_SMI_SEND) { /* don't forward */
2072                 if (smi_handle_dr_smp_send(smp,
2073                                            rdma_cap_ib_switch(port_priv->device),
2074                                            port_num) == IB_SMI_DISCARD)
2075                         return IB_SMI_DISCARD;
2076
2077                 if (smi_check_local_smp(smp, port_priv->device) == IB_SMI_DISCARD)
2078                         return IB_SMI_DISCARD;
2079         } else if (rdma_cap_ib_switch(port_priv->device)) {
2080                 /* forward case for switches */
2081                 memcpy(response, recv, mad_priv_size(response));
2082                 response->header.recv_wc.wc = &response->header.wc;
2083                 response->header.recv_wc.recv_buf.mad = (struct ib_mad *)response->mad;
2084                 response->header.recv_wc.recv_buf.grh = &response->grh;
2085
2086                 agent_send_response((const struct ib_mad_hdr *)response->mad,
2087                                     &response->grh, wc,
2088                                     port_priv->device,
2089                                     smi_get_fwd_port(smp),
2090                                     qp_info->qp->qp_num,
2091                                     response->mad_size,
2092                                     false);
2093
2094                 return IB_SMI_DISCARD;
2095         }
2096         return IB_SMI_HANDLE;
2097 }
2098
2099 static bool generate_unmatched_resp(const struct ib_mad_private *recv,
2100                                     struct ib_mad_private *response,
2101                                     size_t *resp_len, bool opa)
2102 {
2103         const struct ib_mad_hdr *recv_hdr = (const struct ib_mad_hdr *)recv->mad;
2104         struct ib_mad_hdr *resp_hdr = (struct ib_mad_hdr *)response->mad;
2105
2106         if (recv_hdr->method == IB_MGMT_METHOD_GET ||
2107             recv_hdr->method == IB_MGMT_METHOD_SET) {
2108                 memcpy(response, recv, mad_priv_size(response));
2109                 response->header.recv_wc.wc = &response->header.wc;
2110                 response->header.recv_wc.recv_buf.mad = (struct ib_mad *)response->mad;
2111                 response->header.recv_wc.recv_buf.grh = &response->grh;
2112                 resp_hdr->method = IB_MGMT_METHOD_GET_RESP;
2113                 resp_hdr->status = cpu_to_be16(IB_MGMT_MAD_STATUS_UNSUPPORTED_METHOD_ATTRIB);
2114                 if (recv_hdr->mgmt_class == IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
2115                         resp_hdr->status |= IB_SMP_DIRECTION;
2116
2117                 if (opa && recv_hdr->base_version == OPA_MGMT_BASE_VERSION) {
2118                         if (recv_hdr->mgmt_class ==
2119                             IB_MGMT_CLASS_SUBN_LID_ROUTED ||
2120                             recv_hdr->mgmt_class ==
2121                             IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE)
2122                                 *resp_len = opa_get_smp_header_size(
2123                                                         (struct opa_smp *)recv->mad);
2124                         else
2125                                 *resp_len = sizeof(struct ib_mad_hdr);
2126                 }
2127
2128                 return true;
2129         } else {
2130                 return false;
2131         }
2132 }
2133
2134 static enum smi_action
2135 handle_opa_smi(struct ib_mad_port_private *port_priv,
2136                struct ib_mad_qp_info *qp_info,
2137                struct ib_wc *wc,
2138                int port_num,
2139                struct ib_mad_private *recv,
2140                struct ib_mad_private *response)
2141 {
2142         enum smi_forward_action retsmi;
2143         struct opa_smp *smp = (struct opa_smp *)recv->mad;
2144
2145         if (opa_smi_handle_dr_smp_recv(smp,
2146                                    rdma_cap_ib_switch(port_priv->device),
2147                                    port_num,
2148                                    port_priv->device->phys_port_cnt) ==
2149                                    IB_SMI_DISCARD)
2150                 return IB_SMI_DISCARD;
2151
2152         retsmi = opa_smi_check_forward_dr_smp(smp);
2153         if (retsmi == IB_SMI_LOCAL)
2154                 return IB_SMI_HANDLE;
2155
2156         if (retsmi == IB_SMI_SEND) { /* don't forward */
2157                 if (opa_smi_handle_dr_smp_send(smp,
2158                                            rdma_cap_ib_switch(port_priv->device),
2159                                            port_num) == IB_SMI_DISCARD)
2160                         return IB_SMI_DISCARD;
2161
2162                 if (opa_smi_check_local_smp(smp, port_priv->device) ==
2163                     IB_SMI_DISCARD)
2164                         return IB_SMI_DISCARD;
2165
2166         } else if (rdma_cap_ib_switch(port_priv->device)) {
2167                 /* forward case for switches */
2168                 memcpy(response, recv, mad_priv_size(response));
2169                 response->header.recv_wc.wc = &response->header.wc;
2170                 response->header.recv_wc.recv_buf.opa_mad =
2171                                 (struct opa_mad *)response->mad;
2172                 response->header.recv_wc.recv_buf.grh = &response->grh;
2173
2174                 agent_send_response((const struct ib_mad_hdr *)response->mad,
2175                                     &response->grh, wc,
2176                                     port_priv->device,
2177                                     opa_smi_get_fwd_port(smp),
2178                                     qp_info->qp->qp_num,
2179                                     recv->header.wc.byte_len,
2180                                     true);
2181
2182                 return IB_SMI_DISCARD;
2183         }
2184
2185         return IB_SMI_HANDLE;
2186 }
2187
2188 static enum smi_action
2189 handle_smi(struct ib_mad_port_private *port_priv,
2190            struct ib_mad_qp_info *qp_info,
2191            struct ib_wc *wc,
2192            int port_num,
2193            struct ib_mad_private *recv,
2194            struct ib_mad_private *response,
2195            bool opa)
2196 {
2197         struct ib_mad_hdr *mad_hdr = (struct ib_mad_hdr *)recv->mad;
2198
2199         if (opa && mad_hdr->base_version == OPA_MGMT_BASE_VERSION &&
2200             mad_hdr->class_version == OPA_SM_CLASS_VERSION)
2201                 return handle_opa_smi(port_priv, qp_info, wc, port_num, recv,
2202                                       response);
2203
2204         return handle_ib_smi(port_priv, qp_info, wc, port_num, recv, response);
2205 }
2206
2207 static void ib_mad_recv_done(struct ib_cq *cq, struct ib_wc *wc)
2208 {
2209         struct ib_mad_port_private *port_priv = cq->cq_context;
2210         struct ib_mad_list_head *mad_list =
2211                 container_of(wc->wr_cqe, struct ib_mad_list_head, cqe);
2212         struct ib_mad_qp_info *qp_info;
2213         struct ib_mad_private_header *mad_priv_hdr;
2214         struct ib_mad_private *recv, *response = NULL;
2215         struct ib_mad_agent_private *mad_agent;
2216         int port_num;
2217         int ret = IB_MAD_RESULT_SUCCESS;
2218         size_t mad_size;
2219         u16 resp_mad_pkey_index = 0;
2220         bool opa;
2221
2222         if (list_empty_careful(&port_priv->port_list))
2223                 return;
2224
2225         if (wc->status != IB_WC_SUCCESS) {
2226                 /*
2227                  * Receive errors indicate that the QP has entered the error
2228                  * state - error handling/shutdown code will cleanup
2229                  */
2230                 return;
2231         }
2232
2233         qp_info = mad_list->mad_queue->qp_info;
2234         dequeue_mad(mad_list);
2235
2236         opa = rdma_cap_opa_mad(qp_info->port_priv->device,
2237                                qp_info->port_priv->port_num);
2238
2239         mad_priv_hdr = container_of(mad_list, struct ib_mad_private_header,
2240                                     mad_list);
2241         recv = container_of(mad_priv_hdr, struct ib_mad_private, header);
2242         ib_dma_unmap_single(port_priv->device,
2243                             recv->header.mapping,
2244                             mad_priv_dma_size(recv),
2245                             DMA_FROM_DEVICE);
2246
2247         /* Setup MAD receive work completion from "normal" work completion */
2248         recv->header.wc = *wc;
2249         recv->header.recv_wc.wc = &recv->header.wc;
2250
2251         if (opa && ((struct ib_mad_hdr *)(recv->mad))->base_version == OPA_MGMT_BASE_VERSION) {
2252                 recv->header.recv_wc.mad_len = wc->byte_len - sizeof(struct ib_grh);
2253                 recv->header.recv_wc.mad_seg_size = sizeof(struct opa_mad);
2254         } else {
2255                 recv->header.recv_wc.mad_len = sizeof(struct ib_mad);
2256                 recv->header.recv_wc.mad_seg_size = sizeof(struct ib_mad);
2257         }
2258
2259         recv->header.recv_wc.recv_buf.mad = (struct ib_mad *)recv->mad;
2260         recv->header.recv_wc.recv_buf.grh = &recv->grh;
2261
2262         if (atomic_read(&qp_info->snoop_count))
2263                 snoop_recv(qp_info, &recv->header.recv_wc, IB_MAD_SNOOP_RECVS);
2264
2265         /* Validate MAD */
2266         if (!validate_mad((const struct ib_mad_hdr *)recv->mad, qp_info, opa))
2267                 goto out;
2268
2269         mad_size = recv->mad_size;
2270         response = alloc_mad_private(mad_size, GFP_KERNEL);
2271         if (!response)
2272                 goto out;
2273
2274         if (rdma_cap_ib_switch(port_priv->device))
2275                 port_num = wc->port_num;
2276         else
2277                 port_num = port_priv->port_num;
2278
2279         if (((struct ib_mad_hdr *)recv->mad)->mgmt_class ==
2280             IB_MGMT_CLASS_SUBN_DIRECTED_ROUTE) {
2281                 if (handle_smi(port_priv, qp_info, wc, port_num, recv,
2282                                response, opa)
2283                     == IB_SMI_DISCARD)
2284                         goto out;
2285         }
2286
2287         /* Give driver "right of first refusal" on incoming MAD */
2288         if (port_priv->device->process_mad) {
2289                 ret = port_priv->device->process_mad(port_priv->device, 0,
2290                                                      port_priv->port_num,
2291                                                      wc, &recv->grh,
2292                                                      (const struct ib_mad_hdr *)recv->mad,
2293                                                      recv->mad_size,
2294                                                      (struct ib_mad_hdr *)response->mad,
2295                                                      &mad_size, &resp_mad_pkey_index);
2296
2297                 if (opa)
2298                         wc->pkey_index = resp_mad_pkey_index;
2299
2300                 if (ret & IB_MAD_RESULT_SUCCESS) {
2301                         if (ret & IB_MAD_RESULT_CONSUMED)
2302                                 goto out;
2303                         if (ret & IB_MAD_RESULT_REPLY) {
2304                                 agent_send_response((const struct ib_mad_hdr *)response->mad,
2305                                                     &recv->grh, wc,
2306                                                     port_priv->device,
2307                                                     port_num,
2308                                                     qp_info->qp->qp_num,
2309                                                     mad_size, opa);
2310                                 goto out;
2311                         }
2312                 }
2313         }
2314
2315         mad_agent = find_mad_agent(port_priv, (const struct ib_mad_hdr *)recv->mad);
2316         if (mad_agent) {
2317                 ib_mad_complete_recv(mad_agent, &recv->header.recv_wc);
2318                 /*
2319                  * recv is freed up in error cases in ib_mad_complete_recv
2320                  * or via recv_handler in ib_mad_complete_recv()
2321                  */
2322                 recv = NULL;
2323         } else if ((ret & IB_MAD_RESULT_SUCCESS) &&
2324                    generate_unmatched_resp(recv, response, &mad_size, opa)) {
2325                 agent_send_response((const struct ib_mad_hdr *)response->mad, &recv->grh, wc,
2326                                     port_priv->device, port_num,
2327                                     qp_info->qp->qp_num, mad_size, opa);
2328         }
2329
2330 out:
2331         /* Post another receive request for this QP */
2332         if (response) {
2333                 ib_mad_post_receive_mads(qp_info, response);
2334                 kfree(recv);
2335         } else
2336                 ib_mad_post_receive_mads(qp_info, recv);
2337 }
2338
2339 static void adjust_timeout(struct ib_mad_agent_private *mad_agent_priv)
2340 {
2341         struct ib_mad_send_wr_private *mad_send_wr;
2342         unsigned long delay;
2343
2344         if (list_empty(&mad_agent_priv->wait_list)) {
2345                 cancel_delayed_work(&mad_agent_priv->timed_work);
2346         } else {
2347                 mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
2348                                          struct ib_mad_send_wr_private,
2349                                          agent_list);
2350
2351                 if (time_after(mad_agent_priv->timeout,
2352                                mad_send_wr->timeout)) {
2353                         mad_agent_priv->timeout = mad_send_wr->timeout;
2354                         delay = mad_send_wr->timeout - jiffies;
2355                         if ((long)delay <= 0)
2356                                 delay = 1;
2357                         mod_delayed_work(mad_agent_priv->qp_info->port_priv->wq,
2358                                          &mad_agent_priv->timed_work, delay);
2359                 }
2360         }
2361 }
2362
2363 static void wait_for_response(struct ib_mad_send_wr_private *mad_send_wr)
2364 {
2365         struct ib_mad_agent_private *mad_agent_priv;
2366         struct ib_mad_send_wr_private *temp_mad_send_wr;
2367         struct list_head *list_item;
2368         unsigned long delay;
2369
2370         mad_agent_priv = mad_send_wr->mad_agent_priv;
2371         list_del(&mad_send_wr->agent_list);
2372
2373         delay = mad_send_wr->timeout;
2374         mad_send_wr->timeout += jiffies;
2375
2376         if (delay) {
2377                 list_for_each_prev(list_item, &mad_agent_priv->wait_list) {
2378                         temp_mad_send_wr = list_entry(list_item,
2379                                                 struct ib_mad_send_wr_private,
2380                                                 agent_list);
2381                         if (time_after(mad_send_wr->timeout,
2382                                        temp_mad_send_wr->timeout))
2383                                 break;
2384                 }
2385         }
2386         else
2387                 list_item = &mad_agent_priv->wait_list;
2388         list_add(&mad_send_wr->agent_list, list_item);
2389
2390         /* Reschedule a work item if we have a shorter timeout */
2391         if (mad_agent_priv->wait_list.next == &mad_send_wr->agent_list)
2392                 mod_delayed_work(mad_agent_priv->qp_info->port_priv->wq,
2393                                  &mad_agent_priv->timed_work, delay);
2394 }
2395
2396 void ib_reset_mad_timeout(struct ib_mad_send_wr_private *mad_send_wr,
2397                           int timeout_ms)
2398 {
2399         mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2400         wait_for_response(mad_send_wr);
2401 }
2402
2403 /*
2404  * Process a send work completion
2405  */
2406 void ib_mad_complete_send_wr(struct ib_mad_send_wr_private *mad_send_wr,
2407                              struct ib_mad_send_wc *mad_send_wc)
2408 {
2409         struct ib_mad_agent_private     *mad_agent_priv;
2410         unsigned long                   flags;
2411         int                             ret;
2412
2413         mad_agent_priv = mad_send_wr->mad_agent_priv;
2414         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2415         if (ib_mad_kernel_rmpp_agent(&mad_agent_priv->agent)) {
2416                 ret = ib_process_rmpp_send_wc(mad_send_wr, mad_send_wc);
2417                 if (ret == IB_RMPP_RESULT_CONSUMED)
2418                         goto done;
2419         } else
2420                 ret = IB_RMPP_RESULT_UNHANDLED;
2421
2422         if (mad_send_wc->status != IB_WC_SUCCESS &&
2423             mad_send_wr->status == IB_WC_SUCCESS) {
2424                 mad_send_wr->status = mad_send_wc->status;
2425                 mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2426         }
2427
2428         if (--mad_send_wr->refcount > 0) {
2429                 if (mad_send_wr->refcount == 1 && mad_send_wr->timeout &&
2430                     mad_send_wr->status == IB_WC_SUCCESS) {
2431                         wait_for_response(mad_send_wr);
2432                 }
2433                 goto done;
2434         }
2435
2436         /* Remove send from MAD agent and notify client of completion */
2437         list_del(&mad_send_wr->agent_list);
2438         adjust_timeout(mad_agent_priv);
2439         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2440
2441         if (mad_send_wr->status != IB_WC_SUCCESS )
2442                 mad_send_wc->status = mad_send_wr->status;
2443         if (ret == IB_RMPP_RESULT_INTERNAL)
2444                 ib_rmpp_send_handler(mad_send_wc);
2445         else
2446                 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2447                                                    mad_send_wc);
2448
2449         /* Release reference on agent taken when sending */
2450         deref_mad_agent(mad_agent_priv);
2451         return;
2452 done:
2453         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2454 }
2455
2456 static void ib_mad_send_done(struct ib_cq *cq, struct ib_wc *wc)
2457 {
2458         struct ib_mad_port_private *port_priv = cq->cq_context;
2459         struct ib_mad_list_head *mad_list =
2460                 container_of(wc->wr_cqe, struct ib_mad_list_head, cqe);
2461         struct ib_mad_send_wr_private   *mad_send_wr, *queued_send_wr;
2462         struct ib_mad_qp_info           *qp_info;
2463         struct ib_mad_queue             *send_queue;
2464         struct ib_send_wr               *bad_send_wr;
2465         struct ib_mad_send_wc           mad_send_wc;
2466         unsigned long flags;
2467         int ret;
2468
2469         if (list_empty_careful(&port_priv->port_list))
2470                 return;
2471
2472         if (wc->status != IB_WC_SUCCESS) {
2473                 if (!ib_mad_send_error(port_priv, wc))
2474                         return;
2475         }
2476
2477         mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2478                                    mad_list);
2479         send_queue = mad_list->mad_queue;
2480         qp_info = send_queue->qp_info;
2481
2482 retry:
2483         ib_dma_unmap_single(mad_send_wr->send_buf.mad_agent->device,
2484                             mad_send_wr->header_mapping,
2485                             mad_send_wr->sg_list[0].length, DMA_TO_DEVICE);
2486         ib_dma_unmap_single(mad_send_wr->send_buf.mad_agent->device,
2487                             mad_send_wr->payload_mapping,
2488                             mad_send_wr->sg_list[1].length, DMA_TO_DEVICE);
2489         queued_send_wr = NULL;
2490         spin_lock_irqsave(&send_queue->lock, flags);
2491         list_del(&mad_list->list);
2492
2493         /* Move queued send to the send queue */
2494         if (send_queue->count-- > send_queue->max_active) {
2495                 mad_list = container_of(qp_info->overflow_list.next,
2496                                         struct ib_mad_list_head, list);
2497                 queued_send_wr = container_of(mad_list,
2498                                         struct ib_mad_send_wr_private,
2499                                         mad_list);
2500                 list_move_tail(&mad_list->list, &send_queue->list);
2501         }
2502         spin_unlock_irqrestore(&send_queue->lock, flags);
2503
2504         mad_send_wc.send_buf = &mad_send_wr->send_buf;
2505         mad_send_wc.status = wc->status;
2506         mad_send_wc.vendor_err = wc->vendor_err;
2507         if (atomic_read(&qp_info->snoop_count))
2508                 snoop_send(qp_info, &mad_send_wr->send_buf, &mad_send_wc,
2509                            IB_MAD_SNOOP_SEND_COMPLETIONS);
2510         ib_mad_complete_send_wr(mad_send_wr, &mad_send_wc);
2511
2512         if (queued_send_wr) {
2513                 ret = ib_post_send(qp_info->qp, &queued_send_wr->send_wr.wr,
2514                                    &bad_send_wr);
2515                 if (ret) {
2516                         dev_err(&port_priv->device->dev,
2517                                 "ib_post_send failed: %d\n", ret);
2518                         mad_send_wr = queued_send_wr;
2519                         wc->status = IB_WC_LOC_QP_OP_ERR;
2520                         goto retry;
2521                 }
2522         }
2523 }
2524
2525 static void mark_sends_for_retry(struct ib_mad_qp_info *qp_info)
2526 {
2527         struct ib_mad_send_wr_private *mad_send_wr;
2528         struct ib_mad_list_head *mad_list;
2529         unsigned long flags;
2530
2531         spin_lock_irqsave(&qp_info->send_queue.lock, flags);
2532         list_for_each_entry(mad_list, &qp_info->send_queue.list, list) {
2533                 mad_send_wr = container_of(mad_list,
2534                                            struct ib_mad_send_wr_private,
2535                                            mad_list);
2536                 mad_send_wr->retry = 1;
2537         }
2538         spin_unlock_irqrestore(&qp_info->send_queue.lock, flags);
2539 }
2540
2541 static bool ib_mad_send_error(struct ib_mad_port_private *port_priv,
2542                 struct ib_wc *wc)
2543 {
2544         struct ib_mad_list_head *mad_list =
2545                 container_of(wc->wr_cqe, struct ib_mad_list_head, cqe);
2546         struct ib_mad_qp_info *qp_info = mad_list->mad_queue->qp_info;
2547         struct ib_mad_send_wr_private *mad_send_wr;
2548         int ret;
2549
2550         /*
2551          * Send errors will transition the QP to SQE - move
2552          * QP to RTS and repost flushed work requests
2553          */
2554         mad_send_wr = container_of(mad_list, struct ib_mad_send_wr_private,
2555                                    mad_list);
2556         if (wc->status == IB_WC_WR_FLUSH_ERR) {
2557                 if (mad_send_wr->retry) {
2558                         /* Repost send */
2559                         struct ib_send_wr *bad_send_wr;
2560
2561                         mad_send_wr->retry = 0;
2562                         ret = ib_post_send(qp_info->qp, &mad_send_wr->send_wr.wr,
2563                                         &bad_send_wr);
2564                         if (!ret)
2565                                 return false;
2566                 }
2567         } else {
2568                 struct ib_qp_attr *attr;
2569
2570                 /* Transition QP to RTS and fail offending send */
2571                 attr = kmalloc(sizeof *attr, GFP_KERNEL);
2572                 if (attr) {
2573                         attr->qp_state = IB_QPS_RTS;
2574                         attr->cur_qp_state = IB_QPS_SQE;
2575                         ret = ib_modify_qp(qp_info->qp, attr,
2576                                            IB_QP_STATE | IB_QP_CUR_STATE);
2577                         kfree(attr);
2578                         if (ret)
2579                                 dev_err(&port_priv->device->dev,
2580                                         "%s - ib_modify_qp to RTS: %d\n",
2581                                         __func__, ret);
2582                         else
2583                                 mark_sends_for_retry(qp_info);
2584                 }
2585         }
2586
2587         return true;
2588 }
2589
2590 static void cancel_mads(struct ib_mad_agent_private *mad_agent_priv)
2591 {
2592         unsigned long flags;
2593         struct ib_mad_send_wr_private *mad_send_wr, *temp_mad_send_wr;
2594         struct ib_mad_send_wc mad_send_wc;
2595         struct list_head cancel_list;
2596
2597         INIT_LIST_HEAD(&cancel_list);
2598
2599         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2600         list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2601                                  &mad_agent_priv->send_list, agent_list) {
2602                 if (mad_send_wr->status == IB_WC_SUCCESS) {
2603                         mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2604                         mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2605                 }
2606         }
2607
2608         /* Empty wait list to prevent receives from finding a request */
2609         list_splice_init(&mad_agent_priv->wait_list, &cancel_list);
2610         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2611
2612         /* Report all cancelled requests */
2613         mad_send_wc.status = IB_WC_WR_FLUSH_ERR;
2614         mad_send_wc.vendor_err = 0;
2615
2616         list_for_each_entry_safe(mad_send_wr, temp_mad_send_wr,
2617                                  &cancel_list, agent_list) {
2618                 mad_send_wc.send_buf = &mad_send_wr->send_buf;
2619                 list_del(&mad_send_wr->agent_list);
2620                 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2621                                                    &mad_send_wc);
2622                 atomic_dec(&mad_agent_priv->refcount);
2623         }
2624 }
2625
2626 static struct ib_mad_send_wr_private*
2627 find_send_wr(struct ib_mad_agent_private *mad_agent_priv,
2628              struct ib_mad_send_buf *send_buf)
2629 {
2630         struct ib_mad_send_wr_private *mad_send_wr;
2631
2632         list_for_each_entry(mad_send_wr, &mad_agent_priv->wait_list,
2633                             agent_list) {
2634                 if (&mad_send_wr->send_buf == send_buf)
2635                         return mad_send_wr;
2636         }
2637
2638         list_for_each_entry(mad_send_wr, &mad_agent_priv->send_list,
2639                             agent_list) {
2640                 if (is_rmpp_data_mad(mad_agent_priv,
2641                                      mad_send_wr->send_buf.mad) &&
2642                     &mad_send_wr->send_buf == send_buf)
2643                         return mad_send_wr;
2644         }
2645         return NULL;
2646 }
2647
2648 int ib_modify_mad(struct ib_mad_agent *mad_agent,
2649                   struct ib_mad_send_buf *send_buf, u32 timeout_ms)
2650 {
2651         struct ib_mad_agent_private *mad_agent_priv;
2652         struct ib_mad_send_wr_private *mad_send_wr;
2653         unsigned long flags;
2654         int active;
2655
2656         mad_agent_priv = container_of(mad_agent, struct ib_mad_agent_private,
2657                                       agent);
2658         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2659         mad_send_wr = find_send_wr(mad_agent_priv, send_buf);
2660         if (!mad_send_wr || mad_send_wr->status != IB_WC_SUCCESS) {
2661                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2662                 return -EINVAL;
2663         }
2664
2665         active = (!mad_send_wr->timeout || mad_send_wr->refcount > 1);
2666         if (!timeout_ms) {
2667                 mad_send_wr->status = IB_WC_WR_FLUSH_ERR;
2668                 mad_send_wr->refcount -= (mad_send_wr->timeout > 0);
2669         }
2670
2671         mad_send_wr->send_buf.timeout_ms = timeout_ms;
2672         if (active)
2673                 mad_send_wr->timeout = msecs_to_jiffies(timeout_ms);
2674         else
2675                 ib_reset_mad_timeout(mad_send_wr, timeout_ms);
2676
2677         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2678         return 0;
2679 }
2680 EXPORT_SYMBOL(ib_modify_mad);
2681
2682 void ib_cancel_mad(struct ib_mad_agent *mad_agent,
2683                    struct ib_mad_send_buf *send_buf)
2684 {
2685         ib_modify_mad(mad_agent, send_buf, 0);
2686 }
2687 EXPORT_SYMBOL(ib_cancel_mad);
2688
2689 static void local_completions(struct work_struct *work)
2690 {
2691         struct ib_mad_agent_private *mad_agent_priv;
2692         struct ib_mad_local_private *local;
2693         struct ib_mad_agent_private *recv_mad_agent;
2694         unsigned long flags;
2695         int free_mad;
2696         struct ib_wc wc;
2697         struct ib_mad_send_wc mad_send_wc;
2698         bool opa;
2699
2700         mad_agent_priv =
2701                 container_of(work, struct ib_mad_agent_private, local_work);
2702
2703         opa = rdma_cap_opa_mad(mad_agent_priv->qp_info->port_priv->device,
2704                                mad_agent_priv->qp_info->port_priv->port_num);
2705
2706         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2707         while (!list_empty(&mad_agent_priv->local_list)) {
2708                 local = list_entry(mad_agent_priv->local_list.next,
2709                                    struct ib_mad_local_private,
2710                                    completion_list);
2711                 list_del(&local->completion_list);
2712                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2713                 free_mad = 0;
2714                 if (local->mad_priv) {
2715                         u8 base_version;
2716                         recv_mad_agent = local->recv_mad_agent;
2717                         if (!recv_mad_agent) {
2718                                 dev_err(&mad_agent_priv->agent.device->dev,
2719                                         "No receive MAD agent for local completion\n");
2720                                 free_mad = 1;
2721                                 goto local_send_completion;
2722                         }
2723
2724                         /*
2725                          * Defined behavior is to complete response
2726                          * before request
2727                          */
2728                         build_smp_wc(recv_mad_agent->agent.qp,
2729                                      local->mad_send_wr->send_wr.wr.wr_cqe,
2730                                      be16_to_cpu(IB_LID_PERMISSIVE),
2731                                      local->mad_send_wr->send_wr.pkey_index,
2732                                      recv_mad_agent->agent.port_num, &wc);
2733
2734                         local->mad_priv->header.recv_wc.wc = &wc;
2735
2736                         base_version = ((struct ib_mad_hdr *)(local->mad_priv->mad))->base_version;
2737                         if (opa && base_version == OPA_MGMT_BASE_VERSION) {
2738                                 local->mad_priv->header.recv_wc.mad_len = local->return_wc_byte_len;
2739                                 local->mad_priv->header.recv_wc.mad_seg_size = sizeof(struct opa_mad);
2740                         } else {
2741                                 local->mad_priv->header.recv_wc.mad_len = sizeof(struct ib_mad);
2742                                 local->mad_priv->header.recv_wc.mad_seg_size = sizeof(struct ib_mad);
2743                         }
2744
2745                         INIT_LIST_HEAD(&local->mad_priv->header.recv_wc.rmpp_list);
2746                         list_add(&local->mad_priv->header.recv_wc.recv_buf.list,
2747                                  &local->mad_priv->header.recv_wc.rmpp_list);
2748                         local->mad_priv->header.recv_wc.recv_buf.grh = NULL;
2749                         local->mad_priv->header.recv_wc.recv_buf.mad =
2750                                                 (struct ib_mad *)local->mad_priv->mad;
2751                         if (atomic_read(&recv_mad_agent->qp_info->snoop_count))
2752                                 snoop_recv(recv_mad_agent->qp_info,
2753                                           &local->mad_priv->header.recv_wc,
2754                                            IB_MAD_SNOOP_RECVS);
2755                         recv_mad_agent->agent.recv_handler(
2756                                                 &recv_mad_agent->agent,
2757                                                 &local->mad_send_wr->send_buf,
2758                                                 &local->mad_priv->header.recv_wc);
2759                         spin_lock_irqsave(&recv_mad_agent->lock, flags);
2760                         atomic_dec(&recv_mad_agent->refcount);
2761                         spin_unlock_irqrestore(&recv_mad_agent->lock, flags);
2762                 }
2763
2764 local_send_completion:
2765                 /* Complete send */
2766                 mad_send_wc.status = IB_WC_SUCCESS;
2767                 mad_send_wc.vendor_err = 0;
2768                 mad_send_wc.send_buf = &local->mad_send_wr->send_buf;
2769                 if (atomic_read(&mad_agent_priv->qp_info->snoop_count))
2770                         snoop_send(mad_agent_priv->qp_info,
2771                                    &local->mad_send_wr->send_buf,
2772                                    &mad_send_wc, IB_MAD_SNOOP_SEND_COMPLETIONS);
2773                 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2774                                                    &mad_send_wc);
2775
2776                 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2777                 atomic_dec(&mad_agent_priv->refcount);
2778                 if (free_mad)
2779                         kfree(local->mad_priv);
2780                 kfree(local);
2781         }
2782         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2783 }
2784
2785 static int retry_send(struct ib_mad_send_wr_private *mad_send_wr)
2786 {
2787         int ret;
2788
2789         if (!mad_send_wr->retries_left)
2790                 return -ETIMEDOUT;
2791
2792         mad_send_wr->retries_left--;
2793         mad_send_wr->send_buf.retries++;
2794
2795         mad_send_wr->timeout = msecs_to_jiffies(mad_send_wr->send_buf.timeout_ms);
2796
2797         if (ib_mad_kernel_rmpp_agent(&mad_send_wr->mad_agent_priv->agent)) {
2798                 ret = ib_retry_rmpp(mad_send_wr);
2799                 switch (ret) {
2800                 case IB_RMPP_RESULT_UNHANDLED:
2801                         ret = ib_send_mad(mad_send_wr);
2802                         break;
2803                 case IB_RMPP_RESULT_CONSUMED:
2804                         ret = 0;
2805                         break;
2806                 default:
2807                         ret = -ECOMM;
2808                         break;
2809                 }
2810         } else
2811                 ret = ib_send_mad(mad_send_wr);
2812
2813         if (!ret) {
2814                 mad_send_wr->refcount++;
2815                 list_add_tail(&mad_send_wr->agent_list,
2816                               &mad_send_wr->mad_agent_priv->send_list);
2817         }
2818         return ret;
2819 }
2820
2821 static void timeout_sends(struct work_struct *work)
2822 {
2823         struct ib_mad_agent_private *mad_agent_priv;
2824         struct ib_mad_send_wr_private *mad_send_wr;
2825         struct ib_mad_send_wc mad_send_wc;
2826         unsigned long flags, delay;
2827
2828         mad_agent_priv = container_of(work, struct ib_mad_agent_private,
2829                                       timed_work.work);
2830         mad_send_wc.vendor_err = 0;
2831
2832         spin_lock_irqsave(&mad_agent_priv->lock, flags);
2833         while (!list_empty(&mad_agent_priv->wait_list)) {
2834                 mad_send_wr = list_entry(mad_agent_priv->wait_list.next,
2835                                          struct ib_mad_send_wr_private,
2836                                          agent_list);
2837
2838                 if (time_after(mad_send_wr->timeout, jiffies)) {
2839                         delay = mad_send_wr->timeout - jiffies;
2840                         if ((long)delay <= 0)
2841                                 delay = 1;
2842                         queue_delayed_work(mad_agent_priv->qp_info->
2843                                            port_priv->wq,
2844                                            &mad_agent_priv->timed_work, delay);
2845                         break;
2846                 }
2847
2848                 list_del(&mad_send_wr->agent_list);
2849                 if (mad_send_wr->status == IB_WC_SUCCESS &&
2850                     !retry_send(mad_send_wr))
2851                         continue;
2852
2853                 spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2854
2855                 if (mad_send_wr->status == IB_WC_SUCCESS)
2856                         mad_send_wc.status = IB_WC_RESP_TIMEOUT_ERR;
2857                 else
2858                         mad_send_wc.status = mad_send_wr->status;
2859                 mad_send_wc.send_buf = &mad_send_wr->send_buf;
2860                 mad_agent_priv->agent.send_handler(&mad_agent_priv->agent,
2861                                                    &mad_send_wc);
2862
2863                 atomic_dec(&mad_agent_priv->refcount);
2864                 spin_lock_irqsave(&mad_agent_priv->lock, flags);
2865         }
2866         spin_unlock_irqrestore(&mad_agent_priv->lock, flags);
2867 }
2868
2869 /*
2870  * Allocate receive MADs and post receive WRs for them
2871  */
2872 static int ib_mad_post_receive_mads(struct ib_mad_qp_info *qp_info,
2873                                     struct ib_mad_private *mad)
2874 {
2875         unsigned long flags;
2876         int post, ret;
2877         struct ib_mad_private *mad_priv;
2878         struct ib_sge sg_list;
2879         struct ib_recv_wr recv_wr, *bad_recv_wr;
2880         struct ib_mad_queue *recv_queue = &qp_info->recv_queue;
2881
2882         /* Initialize common scatter list fields */
2883         sg_list.lkey = qp_info->port_priv->pd->local_dma_lkey;
2884
2885         /* Initialize common receive WR fields */
2886         recv_wr.next = NULL;
2887         recv_wr.sg_list = &sg_list;
2888         recv_wr.num_sge = 1;
2889
2890         do {
2891                 /* Allocate and map receive buffer */
2892                 if (mad) {
2893                         mad_priv = mad;
2894                         mad = NULL;
2895                 } else {
2896                         mad_priv = alloc_mad_private(port_mad_size(qp_info->port_priv),
2897                                                      GFP_ATOMIC);
2898                         if (!mad_priv) {
2899                                 ret = -ENOMEM;
2900                                 break;
2901                         }
2902                 }
2903                 sg_list.length = mad_priv_dma_size(mad_priv);
2904                 sg_list.addr = ib_dma_map_single(qp_info->port_priv->device,
2905                                                  &mad_priv->grh,
2906                                                  mad_priv_dma_size(mad_priv),
2907                                                  DMA_FROM_DEVICE);
2908                 if (unlikely(ib_dma_mapping_error(qp_info->port_priv->device,
2909                                                   sg_list.addr))) {
2910                         kfree(mad_priv);
2911                         ret = -ENOMEM;
2912                         break;
2913                 }
2914                 mad_priv->header.mapping = sg_list.addr;
2915                 mad_priv->header.mad_list.mad_queue = recv_queue;
2916                 mad_priv->header.mad_list.cqe.done = ib_mad_recv_done;
2917                 recv_wr.wr_cqe = &mad_priv->header.mad_list.cqe;
2918
2919                 /* Post receive WR */
2920                 spin_lock_irqsave(&recv_queue->lock, flags);
2921                 post = (++recv_queue->count < recv_queue->max_active);
2922                 list_add_tail(&mad_priv->header.mad_list.list, &recv_queue->list);
2923                 spin_unlock_irqrestore(&recv_queue->lock, flags);
2924                 ret = ib_post_recv(qp_info->qp, &recv_wr, &bad_recv_wr);
2925                 if (ret) {
2926                         spin_lock_irqsave(&recv_queue->lock, flags);
2927                         list_del(&mad_priv->header.mad_list.list);
2928                         recv_queue->count--;
2929                         spin_unlock_irqrestore(&recv_queue->lock, flags);
2930                         ib_dma_unmap_single(qp_info->port_priv->device,
2931                                             mad_priv->header.mapping,
2932                                             mad_priv_dma_size(mad_priv),
2933                                             DMA_FROM_DEVICE);
2934                         kfree(mad_priv);
2935                         dev_err(&qp_info->port_priv->device->dev,
2936                                 "ib_post_recv failed: %d\n", ret);
2937                         break;
2938                 }
2939         } while (post);
2940
2941         return ret;
2942 }
2943
2944 /*
2945  * Return all the posted receive MADs
2946  */
2947 static void cleanup_recv_queue(struct ib_mad_qp_info *qp_info)
2948 {
2949         struct ib_mad_private_header *mad_priv_hdr;
2950         struct ib_mad_private *recv;
2951         struct ib_mad_list_head *mad_list;
2952
2953         if (!qp_info->qp)
2954                 return;
2955
2956         while (!list_empty(&qp_info->recv_queue.list)) {
2957
2958                 mad_list = list_entry(qp_info->recv_queue.list.next,
2959                                       struct ib_mad_list_head, list);
2960                 mad_priv_hdr = container_of(mad_list,
2961                                             struct ib_mad_private_header,
2962                                             mad_list);
2963                 recv = container_of(mad_priv_hdr, struct ib_mad_private,
2964                                     header);
2965
2966                 /* Remove from posted receive MAD list */
2967                 list_del(&mad_list->list);
2968
2969                 ib_dma_unmap_single(qp_info->port_priv->device,
2970                                     recv->header.mapping,
2971                                     mad_priv_dma_size(recv),
2972                                     DMA_FROM_DEVICE);
2973                 kfree(recv);
2974         }
2975
2976         qp_info->recv_queue.count = 0;
2977 }
2978
2979 /*
2980  * Start the port
2981  */
2982 static int ib_mad_port_start(struct ib_mad_port_private *port_priv)
2983 {
2984         int ret, i;
2985         struct ib_qp_attr *attr;
2986         struct ib_qp *qp;
2987         u16 pkey_index;
2988
2989         attr = kmalloc(sizeof *attr, GFP_KERNEL);
2990         if (!attr)
2991                 return -ENOMEM;
2992
2993         ret = ib_find_pkey(port_priv->device, port_priv->port_num,
2994                            IB_DEFAULT_PKEY_FULL, &pkey_index);
2995         if (ret)
2996                 pkey_index = 0;
2997
2998         for (i = 0; i < IB_MAD_QPS_CORE; i++) {
2999                 qp = port_priv->qp_info[i].qp;
3000                 if (!qp)
3001                         continue;
3002
3003                 /*
3004                  * PKey index for QP1 is irrelevant but
3005                  * one is needed for the Reset to Init transition
3006                  */
3007                 attr->qp_state = IB_QPS_INIT;
3008                 attr->pkey_index = pkey_index;
3009                 attr->qkey = (qp->qp_num == 0) ? 0 : IB_QP1_QKEY;
3010                 ret = ib_modify_qp(qp, attr, IB_QP_STATE |
3011                                              IB_QP_PKEY_INDEX | IB_QP_QKEY);
3012                 if (ret) {
3013                         dev_err(&port_priv->device->dev,
3014                                 "Couldn't change QP%d state to INIT: %d\n",
3015                                 i, ret);
3016                         goto out;
3017                 }
3018
3019                 attr->qp_state = IB_QPS_RTR;
3020                 ret = ib_modify_qp(qp, attr, IB_QP_STATE);
3021                 if (ret) {
3022                         dev_err(&port_priv->device->dev,
3023                                 "Couldn't change QP%d state to RTR: %d\n",
3024                                 i, ret);
3025                         goto out;
3026                 }
3027
3028                 attr->qp_state = IB_QPS_RTS;
3029                 attr->sq_psn = IB_MAD_SEND_Q_PSN;
3030                 ret = ib_modify_qp(qp, attr, IB_QP_STATE | IB_QP_SQ_PSN);
3031                 if (ret) {
3032                         dev_err(&port_priv->device->dev,
3033                                 "Couldn't change QP%d state to RTS: %d\n",
3034                                 i, ret);
3035                         goto out;
3036                 }
3037         }
3038
3039         ret = ib_req_notify_cq(port_priv->cq, IB_CQ_NEXT_COMP);
3040         if (ret) {
3041                 dev_err(&port_priv->device->dev,
3042                         "Failed to request completion notification: %d\n",
3043                         ret);
3044                 goto out;
3045         }
3046
3047         for (i = 0; i < IB_MAD_QPS_CORE; i++) {
3048                 if (!port_priv->qp_info[i].qp)
3049                         continue;
3050
3051                 ret = ib_mad_post_receive_mads(&port_priv->qp_info[i], NULL);
3052                 if (ret) {
3053                         dev_err(&port_priv->device->dev,
3054                                 "Couldn't post receive WRs\n");
3055                         goto out;
3056                 }
3057         }
3058 out:
3059         kfree(attr);
3060         return ret;
3061 }
3062
3063 static void qp_event_handler(struct ib_event *event, void *qp_context)
3064 {
3065         struct ib_mad_qp_info   *qp_info = qp_context;
3066
3067         /* It's worse than that! He's dead, Jim! */
3068         dev_err(&qp_info->port_priv->device->dev,
3069                 "Fatal error (%d) on MAD QP (%d)\n",
3070                 event->event, qp_info->qp->qp_num);
3071 }
3072
3073 static void init_mad_queue(struct ib_mad_qp_info *qp_info,
3074                            struct ib_mad_queue *mad_queue)
3075 {
3076         mad_queue->qp_info = qp_info;
3077         mad_queue->count = 0;
3078         spin_lock_init(&mad_queue->lock);
3079         INIT_LIST_HEAD(&mad_queue->list);
3080 }
3081
3082 static void init_mad_qp(struct ib_mad_port_private *port_priv,
3083                         struct ib_mad_qp_info *qp_info)
3084 {
3085         qp_info->port_priv = port_priv;
3086         init_mad_queue(qp_info, &qp_info->send_queue);
3087         init_mad_queue(qp_info, &qp_info->recv_queue);
3088         INIT_LIST_HEAD(&qp_info->overflow_list);
3089         spin_lock_init(&qp_info->snoop_lock);
3090         qp_info->snoop_table = NULL;
3091         qp_info->snoop_table_size = 0;
3092         atomic_set(&qp_info->snoop_count, 0);
3093 }
3094
3095 static int create_mad_qp(struct ib_mad_qp_info *qp_info,
3096                          enum ib_qp_type qp_type)
3097 {
3098         struct ib_qp_init_attr  qp_init_attr;
3099         int ret;
3100
3101         memset(&qp_init_attr, 0, sizeof qp_init_attr);
3102         qp_init_attr.send_cq = qp_info->port_priv->cq;
3103         qp_init_attr.recv_cq = qp_info->port_priv->cq;
3104         qp_init_attr.sq_sig_type = IB_SIGNAL_ALL_WR;
3105         qp_init_attr.cap.max_send_wr = mad_sendq_size;
3106         qp_init_attr.cap.max_recv_wr = mad_recvq_size;
3107         qp_init_attr.cap.max_send_sge = IB_MAD_SEND_REQ_MAX_SG;
3108         qp_init_attr.cap.max_recv_sge = IB_MAD_RECV_REQ_MAX_SG;
3109         qp_init_attr.qp_type = qp_type;
3110         qp_init_attr.port_num = qp_info->port_priv->port_num;
3111         qp_init_attr.qp_context = qp_info;
3112         qp_init_attr.event_handler = qp_event_handler;
3113         qp_info->qp = ib_create_qp(qp_info->port_priv->pd, &qp_init_attr);
3114         if (IS_ERR(qp_info->qp)) {
3115                 dev_err(&qp_info->port_priv->device->dev,
3116                         "Couldn't create ib_mad QP%d\n",
3117                         get_spl_qp_index(qp_type));
3118                 ret = PTR_ERR(qp_info->qp);
3119                 goto error;
3120         }
3121         /* Use minimum queue sizes unless the CQ is resized */
3122         qp_info->send_queue.max_active = mad_sendq_size;
3123         qp_info->recv_queue.max_active = mad_recvq_size;
3124         return 0;
3125
3126 error:
3127         return ret;
3128 }
3129
3130 static void destroy_mad_qp(struct ib_mad_qp_info *qp_info)
3131 {
3132         if (!qp_info->qp)
3133                 return;
3134
3135         ib_destroy_qp(qp_info->qp);
3136         kfree(qp_info->snoop_table);
3137 }
3138
3139 /*
3140  * Open the port
3141  * Create the QP, PD, MR, and CQ if needed
3142  */
3143 static int ib_mad_port_open(struct ib_device *device,
3144                             int port_num)
3145 {
3146         int ret, cq_size;
3147         struct ib_mad_port_private *port_priv;
3148         unsigned long flags;
3149         char name[sizeof "ib_mad123"];
3150         int has_smi;
3151
3152         if (WARN_ON(rdma_max_mad_size(device, port_num) < IB_MGMT_MAD_SIZE))
3153                 return -EFAULT;
3154
3155         if (WARN_ON(rdma_cap_opa_mad(device, port_num) &&
3156                     rdma_max_mad_size(device, port_num) < OPA_MGMT_MAD_SIZE))
3157                 return -EFAULT;
3158
3159         /* Create new device info */
3160         port_priv = kzalloc(sizeof *port_priv, GFP_KERNEL);
3161         if (!port_priv)
3162                 return -ENOMEM;
3163
3164         port_priv->device = device;
3165         port_priv->port_num = port_num;
3166         spin_lock_init(&port_priv->reg_lock);
3167         INIT_LIST_HEAD(&port_priv->agent_list);
3168         init_mad_qp(port_priv, &port_priv->qp_info[0]);
3169         init_mad_qp(port_priv, &port_priv->qp_info[1]);
3170
3171         cq_size = mad_sendq_size + mad_recvq_size;
3172         has_smi = rdma_cap_ib_smi(device, port_num);
3173         if (has_smi)
3174                 cq_size *= 2;
3175
3176         port_priv->pd = ib_alloc_pd(device, 0);
3177         if (IS_ERR(port_priv->pd)) {
3178                 dev_err(&device->dev, "Couldn't create ib_mad PD\n");
3179                 ret = PTR_ERR(port_priv->pd);
3180                 goto error3;
3181         }
3182
3183         port_priv->cq = ib_alloc_cq(port_priv->device, port_priv, cq_size, 0,
3184                         IB_POLL_UNBOUND_WORKQUEUE);
3185         if (IS_ERR(port_priv->cq)) {
3186                 dev_err(&device->dev, "Couldn't create ib_mad CQ\n");
3187                 ret = PTR_ERR(port_priv->cq);
3188                 goto error4;
3189         }
3190
3191         if (has_smi) {
3192                 ret = create_mad_qp(&port_priv->qp_info[0], IB_QPT_SMI);
3193                 if (ret)
3194                         goto error6;
3195         }
3196         ret = create_mad_qp(&port_priv->qp_info[1], IB_QPT_GSI);
3197         if (ret)
3198                 goto error7;
3199
3200         snprintf(name, sizeof name, "ib_mad%d", port_num);
3201         port_priv->wq = alloc_ordered_workqueue(name, WQ_MEM_RECLAIM);
3202         if (!port_priv->wq) {
3203                 ret = -ENOMEM;
3204                 goto error8;
3205         }
3206
3207         spin_lock_irqsave(&ib_mad_port_list_lock, flags);
3208         list_add_tail(&port_priv->port_list, &ib_mad_port_list);
3209         spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3210
3211         ret = ib_mad_port_start(port_priv);
3212         if (ret) {
3213                 dev_err(&device->dev, "Couldn't start port\n");
3214                 goto error9;
3215         }
3216
3217         return 0;
3218
3219 error9:
3220         spin_lock_irqsave(&ib_mad_port_list_lock, flags);
3221         list_del_init(&port_priv->port_list);
3222         spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3223
3224         destroy_workqueue(port_priv->wq);
3225 error8:
3226         destroy_mad_qp(&port_priv->qp_info[1]);
3227 error7:
3228         destroy_mad_qp(&port_priv->qp_info[0]);
3229 error6:
3230         ib_free_cq(port_priv->cq);
3231         cleanup_recv_queue(&port_priv->qp_info[1]);
3232         cleanup_recv_queue(&port_priv->qp_info[0]);
3233 error4:
3234         ib_dealloc_pd(port_priv->pd);
3235 error3:
3236         kfree(port_priv);
3237
3238         return ret;
3239 }
3240
3241 /*
3242  * Close the port
3243  * If there are no classes using the port, free the port
3244  * resources (CQ, MR, PD, QP) and remove the port's info structure
3245  */
3246 static int ib_mad_port_close(struct ib_device *device, int port_num)
3247 {
3248         struct ib_mad_port_private *port_priv;
3249         unsigned long flags;
3250
3251         spin_lock_irqsave(&ib_mad_port_list_lock, flags);
3252         port_priv = __ib_get_mad_port(device, port_num);
3253         if (port_priv == NULL) {
3254                 spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3255                 dev_err(&device->dev, "Port %d not found\n", port_num);
3256                 return -ENODEV;
3257         }
3258         list_del_init(&port_priv->port_list);
3259         spin_unlock_irqrestore(&ib_mad_port_list_lock, flags);
3260
3261         destroy_workqueue(port_priv->wq);
3262         destroy_mad_qp(&port_priv->qp_info[1]);
3263         destroy_mad_qp(&port_priv->qp_info[0]);
3264         ib_free_cq(port_priv->cq);
3265         ib_dealloc_pd(port_priv->pd);
3266         cleanup_recv_queue(&port_priv->qp_info[1]);
3267         cleanup_recv_queue(&port_priv->qp_info[0]);
3268         /* XXX: Handle deallocation of MAD registration tables */
3269
3270         kfree(port_priv);
3271
3272         return 0;
3273 }
3274
3275 static void ib_mad_init_device(struct ib_device *device)
3276 {
3277         int start, i;
3278
3279         start = rdma_start_port(device);
3280
3281         for (i = start; i <= rdma_end_port(device); i++) {
3282                 if (!rdma_cap_ib_mad(device, i))
3283                         continue;
3284
3285                 if (ib_mad_port_open(device, i)) {
3286                         dev_err(&device->dev, "Couldn't open port %d\n", i);
3287                         goto error;
3288                 }
3289                 if (ib_agent_port_open(device, i)) {
3290                         dev_err(&device->dev,
3291                                 "Couldn't open port %d for agents\n", i);
3292                         goto error_agent;
3293                 }
3294         }
3295         return;
3296
3297 error_agent:
3298         if (ib_mad_port_close(device, i))
3299                 dev_err(&device->dev, "Couldn't close port %d\n", i);
3300
3301 error:
3302         while (--i >= start) {
3303                 if (!rdma_cap_ib_mad(device, i))
3304                         continue;
3305
3306                 if (ib_agent_port_close(device, i))
3307                         dev_err(&device->dev,
3308                                 "Couldn't close port %d for agents\n", i);
3309                 if (ib_mad_port_close(device, i))
3310                         dev_err(&device->dev, "Couldn't close port %d\n", i);
3311         }
3312 }
3313
3314 static void ib_mad_remove_device(struct ib_device *device, void *client_data)
3315 {
3316         int i;
3317
3318         for (i = rdma_start_port(device); i <= rdma_end_port(device); i++) {
3319                 if (!rdma_cap_ib_mad(device, i))
3320                         continue;
3321
3322                 if (ib_agent_port_close(device, i))
3323                         dev_err(&device->dev,
3324                                 "Couldn't close port %d for agents\n", i);
3325                 if (ib_mad_port_close(device, i))
3326                         dev_err(&device->dev, "Couldn't close port %d\n", i);
3327         }
3328 }
3329
3330 static struct ib_client mad_client = {
3331         .name   = "mad",
3332         .add = ib_mad_init_device,
3333         .remove = ib_mad_remove_device
3334 };
3335
3336 int ib_mad_init(void)
3337 {
3338         mad_recvq_size = min(mad_recvq_size, IB_MAD_QP_MAX_SIZE);
3339         mad_recvq_size = max(mad_recvq_size, IB_MAD_QP_MIN_SIZE);
3340
3341         mad_sendq_size = min(mad_sendq_size, IB_MAD_QP_MAX_SIZE);
3342         mad_sendq_size = max(mad_sendq_size, IB_MAD_QP_MIN_SIZE);
3343
3344         INIT_LIST_HEAD(&ib_mad_port_list);
3345
3346         if (ib_register_client(&mad_client)) {
3347                 pr_err("Couldn't register ib_mad client\n");
3348                 return -EINVAL;
3349         }
3350
3351         return 0;
3352 }
3353
3354 void ib_mad_cleanup(void)
3355 {
3356         ib_unregister_client(&mad_client);
3357 }