drivers/infiniband/core/verbs.c

   1 /*
   2  * Copyright (c) 2004 Mellanox Technologies Ltd.  All rights reserved.
   3  * Copyright (c) 2004 Infinicon Corporation.  All rights reserved.
   4  * Copyright (c) 2004 Intel Corporation.  All rights reserved.
   5  * Copyright (c) 2004 Topspin Corporation.  All rights reserved.
   6  * Copyright (c) 2004 Voltaire Corporation.  All rights reserved.
   7  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
   8  * Copyright (c) 2005, 2006 Cisco Systems.  All rights reserved.
   9  *
  10  * This software is available to you under a choice of one of two
  11  * licenses.  You may choose to be licensed under the terms of the GNU
  12  * General Public License (GPL) Version 2, available from the file
  13  * COPYING in the main directory of this source tree, or the
  14  * OpenIB.org BSD license below:
  15  *
  16  *     Redistribution and use in source and binary forms, with or
  17  *     without modification, are permitted provided that the following
  18  *     conditions are met:
  19  *
  20  *      - Redistributions of source code must retain the above
  21  *        copyright notice, this list of conditions and the following
  22  *        disclaimer.
  23  *
  24  *      - Redistributions in binary form must reproduce the above
  25  *        copyright notice, this list of conditions and the following
  26  *        disclaimer in the documentation and/or other materials
  27  *        provided with the distribution.
  28  *
  29  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  30  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  31  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  32  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  33  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  34  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  35  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  36  * SOFTWARE.
  37  */
  38
  39 #include <linux/errno.h>
  40 #include <linux/err.h>
  41 #include <linux/export.h>
  42 #include <linux/string.h>
  43 #include <linux/slab.h>
  44 #include <linux/in.h>
  45 #include <linux/in6.h>
  46 #include <net/addrconf.h>
  47 #include <linux/security.h>
  48
  49 #include <rdma/ib_verbs.h>
  50 #include <rdma/ib_cache.h>
  51 #include <rdma/ib_addr.h>
  52 #include <rdma/rw.h>
  53
  54 #include "core_priv.h"
  55
  56 static int ib_resolve_eth_dmac(struct ib_device *device,
  57                                struct rdma_ah_attr *ah_attr);
  58
  59 static const char * const ib_events[] = {
  60         [IB_EVENT_CQ_ERR]               = "CQ error",
  61         [IB_EVENT_QP_FATAL]             = "QP fatal error",
  62         [IB_EVENT_QP_REQ_ERR]           = "QP request error",
  63         [IB_EVENT_QP_ACCESS_ERR]        = "QP access error",
  64         [IB_EVENT_COMM_EST]             = "communication established",
  65         [IB_EVENT_SQ_DRAINED]           = "send queue drained",
  66         [IB_EVENT_PATH_MIG]             = "path migration successful",
  67         [IB_EVENT_PATH_MIG_ERR]         = "path migration error",
  68         [IB_EVENT_DEVICE_FATAL]         = "device fatal error",
  69         [IB_EVENT_PORT_ACTIVE]          = "port active",
  70         [IB_EVENT_PORT_ERR]             = "port error",
  71         [IB_EVENT_LID_CHANGE]           = "LID change",
  72         [IB_EVENT_PKEY_CHANGE]          = "P_key change",
  73         [IB_EVENT_SM_CHANGE]            = "SM change",
  74         [IB_EVENT_SRQ_ERR]              = "SRQ error",
  75         [IB_EVENT_SRQ_LIMIT_REACHED]    = "SRQ limit reached",
  76         [IB_EVENT_QP_LAST_WQE_REACHED]  = "last WQE reached",
  77         [IB_EVENT_CLIENT_REREGISTER]    = "client reregister",
  78         [IB_EVENT_GID_CHANGE]           = "GID changed",
  79 };
  80
  81 const char *__attribute_const__ ib_event_msg(enum ib_event_type event)
  82 {
  83         size_t index = event;
  84
  85         return (index < ARRAY_SIZE(ib_events) && ib_events[index]) ?
  86                         ib_events[index] : "unrecognized event";
  87 }
  88 EXPORT_SYMBOL(ib_event_msg);
  89
  90 static const char * const wc_statuses[] = {
  91         [IB_WC_SUCCESS]                 = "success",
  92         [IB_WC_LOC_LEN_ERR]             = "local length error",
  93         [IB_WC_LOC_QP_OP_ERR]           = "local QP operation error",
  94         [IB_WC_LOC_EEC_OP_ERR]          = "local EE context operation error",
  95         [IB_WC_LOC_PROT_ERR]            = "local protection error",
  96         [IB_WC_WR_FLUSH_ERR]            = "WR flushed",
  97         [IB_WC_MW_BIND_ERR]             = "memory management operation error",
  98         [IB_WC_BAD_RESP_ERR]            = "bad response error",
  99         [IB_WC_LOC_ACCESS_ERR]          = "local access error",
 100         [IB_WC_REM_INV_REQ_ERR]         = "invalid request error",
 101         [IB_WC_REM_ACCESS_ERR]          = "remote access error",
 102         [IB_WC_REM_OP_ERR]              = "remote operation error",
 103         [IB_WC_RETRY_EXC_ERR]           = "transport retry counter exceeded",
 104         [IB_WC_RNR_RETRY_EXC_ERR]       = "RNR retry counter exceeded",
 105         [IB_WC_LOC_RDD_VIOL_ERR]        = "local RDD violation error",
 106         [IB_WC_REM_INV_RD_REQ_ERR]      = "remote invalid RD request",
 107         [IB_WC_REM_ABORT_ERR]           = "operation aborted",
 108         [IB_WC_INV_EECN_ERR]            = "invalid EE context number",
 109         [IB_WC_INV_EEC_STATE_ERR]       = "invalid EE context state",
 110         [IB_WC_FATAL_ERR]               = "fatal error",
 111         [IB_WC_RESP_TIMEOUT_ERR]        = "response timeout error",
 112         [IB_WC_GENERAL_ERR]             = "general error",
 113 };
 114
 115 const char *__attribute_const__ ib_wc_status_msg(enum ib_wc_status status)
 116 {
 117         size_t index = status;
 118
 119         return (index < ARRAY_SIZE(wc_statuses) && wc_statuses[index]) ?
 120                         wc_statuses[index] : "unrecognized status";
 121 }
 122 EXPORT_SYMBOL(ib_wc_status_msg);
 123
 124 __attribute_const__ int ib_rate_to_mult(enum ib_rate rate)
 125 {
 126         switch (rate) {
 127         case IB_RATE_2_5_GBPS: return   1;
 128         case IB_RATE_5_GBPS:   return   2;
 129         case IB_RATE_10_GBPS:  return   4;
 130         case IB_RATE_20_GBPS:  return   8;
 131         case IB_RATE_30_GBPS:  return  12;
 132         case IB_RATE_40_GBPS:  return  16;
 133         case IB_RATE_60_GBPS:  return  24;
 134         case IB_RATE_80_GBPS:  return  32;
 135         case IB_RATE_120_GBPS: return  48;
 136         case IB_RATE_14_GBPS:  return   6;
 137         case IB_RATE_56_GBPS:  return  22;
 138         case IB_RATE_112_GBPS: return  45;
 139         case IB_RATE_168_GBPS: return  67;
 140         case IB_RATE_25_GBPS:  return  10;
 141         case IB_RATE_100_GBPS: return  40;
 142         case IB_RATE_200_GBPS: return  80;
 143         case IB_RATE_300_GBPS: return 120;
 144         case IB_RATE_28_GBPS:  return  11;
 145         case IB_RATE_50_GBPS:  return  20;
 146         case IB_RATE_400_GBPS: return 160;
 147         case IB_RATE_600_GBPS: return 240;
 148         default:               return  -1;
 149         }
 150 }
 151 EXPORT_SYMBOL(ib_rate_to_mult);
 152
 153 __attribute_const__ enum ib_rate mult_to_ib_rate(int mult)
 154 {
 155         switch (mult) {
 156         case 1:   return IB_RATE_2_5_GBPS;
 157         case 2:   return IB_RATE_5_GBPS;
 158         case 4:   return IB_RATE_10_GBPS;
 159         case 8:   return IB_RATE_20_GBPS;
 160         case 12:  return IB_RATE_30_GBPS;
 161         case 16:  return IB_RATE_40_GBPS;
 162         case 24:  return IB_RATE_60_GBPS;
 163         case 32:  return IB_RATE_80_GBPS;
 164         case 48:  return IB_RATE_120_GBPS;
 165         case 6:   return IB_RATE_14_GBPS;
 166         case 22:  return IB_RATE_56_GBPS;
 167         case 45:  return IB_RATE_112_GBPS;
 168         case 67:  return IB_RATE_168_GBPS;
 169         case 10:  return IB_RATE_25_GBPS;
 170         case 40:  return IB_RATE_100_GBPS;
 171         case 80:  return IB_RATE_200_GBPS;
 172         case 120: return IB_RATE_300_GBPS;
 173         case 11:  return IB_RATE_28_GBPS;
 174         case 20:  return IB_RATE_50_GBPS;
 175         case 160: return IB_RATE_400_GBPS;
 176         case 240: return IB_RATE_600_GBPS;
 177         default:  return IB_RATE_PORT_CURRENT;
 178         }
 179 }
 180 EXPORT_SYMBOL(mult_to_ib_rate);
 181
 182 __attribute_const__ int ib_rate_to_mbps(enum ib_rate rate)
 183 {
 184         switch (rate) {
 185         case IB_RATE_2_5_GBPS: return 2500;
 186         case IB_RATE_5_GBPS:   return 5000;
 187         case IB_RATE_10_GBPS:  return 10000;
 188         case IB_RATE_20_GBPS:  return 20000;
 189         case IB_RATE_30_GBPS:  return 30000;
 190         case IB_RATE_40_GBPS:  return 40000;
 191         case IB_RATE_60_GBPS:  return 60000;
 192         case IB_RATE_80_GBPS:  return 80000;
 193         case IB_RATE_120_GBPS: return 120000;
 194         case IB_RATE_14_GBPS:  return 14062;
 195         case IB_RATE_56_GBPS:  return 56250;
 196         case IB_RATE_112_GBPS: return 112500;
 197         case IB_RATE_168_GBPS: return 168750;
 198         case IB_RATE_25_GBPS:  return 25781;
 199         case IB_RATE_100_GBPS: return 103125;
 200         case IB_RATE_200_GBPS: return 206250;
 201         case IB_RATE_300_GBPS: return 309375;
 202         case IB_RATE_28_GBPS:  return 28125;
 203         case IB_RATE_50_GBPS:  return 53125;
 204         case IB_RATE_400_GBPS: return 425000;
 205         case IB_RATE_600_GBPS: return 637500;
 206         default:               return -1;
 207         }
 208 }
 209 EXPORT_SYMBOL(ib_rate_to_mbps);
 210
 211 __attribute_const__ enum rdma_transport_type
 212 rdma_node_get_transport(enum rdma_node_type node_type)
 213 {
 214
 215         if (node_type == RDMA_NODE_USNIC)
 216                 return RDMA_TRANSPORT_USNIC;
 217         if (node_type == RDMA_NODE_USNIC_UDP)
 218                 return RDMA_TRANSPORT_USNIC_UDP;
 219         if (node_type == RDMA_NODE_RNIC)
 220                 return RDMA_TRANSPORT_IWARP;
 221         if (node_type == RDMA_NODE_UNSPECIFIED)
 222                 return RDMA_TRANSPORT_UNSPECIFIED;
 223
 224         return RDMA_TRANSPORT_IB;
 225 }
 226 EXPORT_SYMBOL(rdma_node_get_transport);
 227
 228 enum rdma_link_layer rdma_port_get_link_layer(struct ib_device *device, u8 port_num)
 229 {
 230         enum rdma_transport_type lt;
 231         if (device->ops.get_link_layer)
 232                 return device->ops.get_link_layer(device, port_num);
 233
 234         lt = rdma_node_get_transport(device->node_type);
 235         if (lt == RDMA_TRANSPORT_IB)
 236                 return IB_LINK_LAYER_INFINIBAND;
 237
 238         return IB_LINK_LAYER_ETHERNET;
 239 }
 240 EXPORT_SYMBOL(rdma_port_get_link_layer);
 241
 242 /* Protection domains */
 243
 244 /**
 245  * ib_alloc_pd - Allocates an unused protection domain.
 246  * @device: The device on which to allocate the protection domain.
 247  *
 248  * A protection domain object provides an association between QPs, shared
 249  * receive queues, address handles, memory regions, and memory windows.
 250  *
 251  * Every PD has a local_dma_lkey which can be used as the lkey value for local
 252  * memory operations.
 253  */
 254 struct ib_pd *__ib_alloc_pd(struct ib_device *device, unsigned int flags,
 255                 const char *caller)
 256 {
 257         struct ib_pd *pd;
 258         int mr_access_flags = 0;
 259         int ret;
 260
 261         pd = rdma_zalloc_drv_obj(device, ib_pd);
 262         if (!pd)
 263                 return ERR_PTR(-ENOMEM);
 264
 265         pd->device = device;
 266         pd->uobject = NULL;
 267         pd->__internal_mr = NULL;
 268         atomic_set(&pd->usecnt, 0);
 269         pd->flags = flags;
 270
 271         pd->res.type = RDMA_RESTRACK_PD;
 272         rdma_restrack_set_task(&pd->res, caller);
 273
 274         ret = device->ops.alloc_pd(pd, NULL);
 275         if (ret) {
 276                 kfree(pd);
 277                 return ERR_PTR(ret);
 278         }
 279         rdma_restrack_kadd(&pd->res);
 280
 281         if (device->attrs.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY)
 282                 pd->local_dma_lkey = device->local_dma_lkey;
 283         else
 284                 mr_access_flags |= IB_ACCESS_LOCAL_WRITE;
 285
 286         if (flags & IB_PD_UNSAFE_GLOBAL_RKEY) {
 287                 pr_warn("%s: enabling unsafe global rkey\n", caller);
 288                 mr_access_flags |= IB_ACCESS_REMOTE_READ | IB_ACCESS_REMOTE_WRITE;
 289         }
 290
 291         if (mr_access_flags) {
 292                 struct ib_mr *mr;
 293
 294                 mr = pd->device->ops.get_dma_mr(pd, mr_access_flags);
 295                 if (IS_ERR(mr)) {
 296                         ib_dealloc_pd(pd);
 297                         return ERR_CAST(mr);
 298                 }
 299
 300                 mr->device      = pd->device;
 301                 mr->pd          = pd;
 302                 mr->uobject     = NULL;
 303                 mr->need_inval  = false;
 304
 305                 pd->__internal_mr = mr;
 306
 307                 if (!(device->attrs.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY))
 308                         pd->local_dma_lkey = pd->__internal_mr->lkey;
 309
 310                 if (flags & IB_PD_UNSAFE_GLOBAL_RKEY)
 311                         pd->unsafe_global_rkey = pd->__internal_mr->rkey;
 312         }
 313
 314         return pd;
 315 }
 316 EXPORT_SYMBOL(__ib_alloc_pd);
 317
 318 /**
 319  * ib_dealloc_pd - Deallocates a protection domain.
 320  * @pd: The protection domain to deallocate.
 321  * @udata: Valid user data or NULL for kernel object
 322  *
 323  * It is an error to call this function while any resources in the pd still
 324  * exist.  The caller is responsible to synchronously destroy them and
 325  * guarantee no new allocations will happen.
 326  */
 327 void ib_dealloc_pd_user(struct ib_pd *pd, struct ib_udata *udata)
 328 {
 329         int ret;
 330
 331         if (pd->__internal_mr) {
 332                 ret = pd->device->ops.dereg_mr(pd->__internal_mr, NULL);
 333                 WARN_ON(ret);
 334                 pd->__internal_mr = NULL;
 335         }
 336
 337         /* uverbs manipulates usecnt with proper locking, while the kabi
 338            requires the caller to guarantee we can't race here. */
 339         WARN_ON(atomic_read(&pd->usecnt));
 340
 341         rdma_restrack_del(&pd->res);
 342         pd->device->ops.dealloc_pd(pd, udata);
 343         kfree(pd);
 344 }
 345 EXPORT_SYMBOL(ib_dealloc_pd_user);
 346
 347 /* Address handles */
 348
 349 /**
 350  * rdma_copy_ah_attr - Copy rdma ah attribute from source to destination.
 351  * @dest:       Pointer to destination ah_attr. Contents of the destination
 352  *              pointer is assumed to be invalid and attribute are overwritten.
 353  * @src:        Pointer to source ah_attr.
 354  */
 355 void rdma_copy_ah_attr(struct rdma_ah_attr *dest,
 356                        const struct rdma_ah_attr *src)
 357 {
 358         *dest = *src;
 359         if (dest->grh.sgid_attr)
 360                 rdma_hold_gid_attr(dest->grh.sgid_attr);
 361 }
 362 EXPORT_SYMBOL(rdma_copy_ah_attr);
 363
 364 /**
 365  * rdma_replace_ah_attr - Replace valid ah_attr with new new one.
 366  * @old:        Pointer to existing ah_attr which needs to be replaced.
 367  *              old is assumed to be valid or zero'd
 368  * @new:        Pointer to the new ah_attr.
 369  *
 370  * rdma_replace_ah_attr() first releases any reference in the old ah_attr if
 371  * old the ah_attr is valid; after that it copies the new attribute and holds
 372  * the reference to the replaced ah_attr.
 373  */
 374 void rdma_replace_ah_attr(struct rdma_ah_attr *old,
 375                           const struct rdma_ah_attr *new)
 376 {
 377         rdma_destroy_ah_attr(old);
 378         *old = *new;
 379         if (old->grh.sgid_attr)
 380                 rdma_hold_gid_attr(old->grh.sgid_attr);
 381 }
 382 EXPORT_SYMBOL(rdma_replace_ah_attr);
 383
 384 /**
 385  * rdma_move_ah_attr - Move ah_attr pointed by source to destination.
 386  * @dest:       Pointer to destination ah_attr to copy to.
 387  *              dest is assumed to be valid or zero'd
 388  * @src:        Pointer to the new ah_attr.
 389  *
 390  * rdma_move_ah_attr() first releases any reference in the destination ah_attr
 391  * if it is valid. This also transfers ownership of internal references from
 392  * src to dest, making src invalid in the process. No new reference of the src
 393  * ah_attr is taken.
 394  */
 395 void rdma_move_ah_attr(struct rdma_ah_attr *dest, struct rdma_ah_attr *src)
 396 {
 397         rdma_destroy_ah_attr(dest);
 398         *dest = *src;
 399         src->grh.sgid_attr = NULL;
 400 }
 401 EXPORT_SYMBOL(rdma_move_ah_attr);
 402
 403 /*
 404  * Validate that the rdma_ah_attr is valid for the device before passing it
 405  * off to the driver.
 406  */
 407 static int rdma_check_ah_attr(struct ib_device *device,
 408                               struct rdma_ah_attr *ah_attr)
 409 {
 410         if (!rdma_is_port_valid(device, ah_attr->port_num))
 411                 return -EINVAL;
 412
 413         if ((rdma_is_grh_required(device, ah_attr->port_num) ||
 414              ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE) &&
 415             !(ah_attr->ah_flags & IB_AH_GRH))
 416                 return -EINVAL;
 417
 418         if (ah_attr->grh.sgid_attr) {
 419                 /*
 420                  * Make sure the passed sgid_attr is consistent with the
 421                  * parameters
 422                  */
 423                 if (ah_attr->grh.sgid_attr->index != ah_attr->grh.sgid_index ||
 424                     ah_attr->grh.sgid_attr->port_num != ah_attr->port_num)
 425                         return -EINVAL;
 426         }
 427         return 0;
 428 }
 429
 430 /*
 431  * If the ah requires a GRH then ensure that sgid_attr pointer is filled in.
 432  * On success the caller is responsible to call rdma_unfill_sgid_attr().
 433  */
 434 static int rdma_fill_sgid_attr(struct ib_device *device,
 435                                struct rdma_ah_attr *ah_attr,
 436                                const struct ib_gid_attr **old_sgid_attr)
 437 {
 438         const struct ib_gid_attr *sgid_attr;
 439         struct ib_global_route *grh;
 440         int ret;
 441
 442         *old_sgid_attr = ah_attr->grh.sgid_attr;
 443
 444         ret = rdma_check_ah_attr(device, ah_attr);
 445         if (ret)
 446                 return ret;
 447
 448         if (!(ah_attr->ah_flags & IB_AH_GRH))
 449                 return 0;
 450
 451         grh = rdma_ah_retrieve_grh(ah_attr);
 452         if (grh->sgid_attr)
 453                 return 0;
 454
 455         sgid_attr =
 456                 rdma_get_gid_attr(device, ah_attr->port_num, grh->sgid_index);
 457         if (IS_ERR(sgid_attr))
 458                 return PTR_ERR(sgid_attr);
 459
 460         /* Move ownerhip of the kref into the ah_attr */
 461         grh->sgid_attr = sgid_attr;
 462         return 0;
 463 }
 464
 465 static void rdma_unfill_sgid_attr(struct rdma_ah_attr *ah_attr,
 466                                   const struct ib_gid_attr *old_sgid_attr)
 467 {
 468         /*
 469          * Fill didn't change anything, the caller retains ownership of
 470          * whatever it passed
 471          */
 472         if (ah_attr->grh.sgid_attr == old_sgid_attr)
 473                 return;
 474
 475         /*
 476          * Otherwise, we need to undo what rdma_fill_sgid_attr so the caller
 477          * doesn't see any change in the rdma_ah_attr. If we get here
 478          * old_sgid_attr is NULL.
 479          */
 480         rdma_destroy_ah_attr(ah_attr);
 481 }
 482
 483 static const struct ib_gid_attr *
 484 rdma_update_sgid_attr(struct rdma_ah_attr *ah_attr,
 485                       const struct ib_gid_attr *old_attr)
 486 {
 487         if (old_attr)
 488                 rdma_put_gid_attr(old_attr);
 489         if (ah_attr->ah_flags & IB_AH_GRH) {
 490                 rdma_hold_gid_attr(ah_attr->grh.sgid_attr);
 491                 return ah_attr->grh.sgid_attr;
 492         }
 493         return NULL;
 494 }
 495
 496 static struct ib_ah *_rdma_create_ah(struct ib_pd *pd,
 497                                      struct rdma_ah_attr *ah_attr,
 498                                      u32 flags,
 499                                      struct ib_udata *udata)
 500 {
 501         struct ib_device *device = pd->device;
 502         struct ib_ah *ah;
 503         int ret;
 504
 505         might_sleep_if(flags & RDMA_CREATE_AH_SLEEPABLE);
 506
 507         if (!device->ops.create_ah)
 508                 return ERR_PTR(-EOPNOTSUPP);
 509
 510         ah = rdma_zalloc_drv_obj_gfp(
 511                 device, ib_ah,
 512                 (flags & RDMA_CREATE_AH_SLEEPABLE) ? GFP_KERNEL : GFP_ATOMIC);
 513         if (!ah)
 514                 return ERR_PTR(-ENOMEM);
 515
 516         ah->device = device;
 517         ah->pd = pd;
 518         ah->type = ah_attr->type;
 519         ah->sgid_attr = rdma_update_sgid_attr(ah_attr, NULL);
 520
 521         ret = device->ops.create_ah(ah, ah_attr, flags, udata);
 522         if (ret) {
 523                 kfree(ah);
 524                 return ERR_PTR(ret);
 525         }
 526
 527         atomic_inc(&pd->usecnt);
 528         return ah;
 529 }
 530
 531 /**
 532  * rdma_create_ah - Creates an address handle for the
 533  * given address vector.
 534  * @pd: The protection domain associated with the address handle.
 535  * @ah_attr: The attributes of the address vector.
 536  * @flags: Create address handle flags (see enum rdma_create_ah_flags).
 537  *
 538  * It returns 0 on success and returns appropriate error code on error.
 539  * The address handle is used to reference a local or global destination
 540  * in all UD QP post sends.
 541  */
 542 struct ib_ah *rdma_create_ah(struct ib_pd *pd, struct rdma_ah_attr *ah_attr,
 543                              u32 flags)
 544 {
 545         const struct ib_gid_attr *old_sgid_attr;
 546         struct ib_ah *ah;
 547         int ret;
 548
 549         ret = rdma_fill_sgid_attr(pd->device, ah_attr, &old_sgid_attr);
 550         if (ret)
 551                 return ERR_PTR(ret);
 552
 553         ah = _rdma_create_ah(pd, ah_attr, flags, NULL);
 554
 555         rdma_unfill_sgid_attr(ah_attr, old_sgid_attr);
 556         return ah;
 557 }
 558 EXPORT_SYMBOL(rdma_create_ah);
 559
 560 /**
 561  * rdma_create_user_ah - Creates an address handle for the
 562  * given address vector.
 563  * It resolves destination mac address for ah attribute of RoCE type.
 564  * @pd: The protection domain associated with the address handle.
 565  * @ah_attr: The attributes of the address vector.
 566  * @udata: pointer to user's input output buffer information need by
 567  *         provider driver.
 568  *
 569  * It returns 0 on success and returns appropriate error code on error.
 570  * The address handle is used to reference a local or global destination
 571  * in all UD QP post sends.
 572  */
 573 struct ib_ah *rdma_create_user_ah(struct ib_pd *pd,
 574                                   struct rdma_ah_attr *ah_attr,
 575                                   struct ib_udata *udata)
 576 {
 577         const struct ib_gid_attr *old_sgid_attr;
 578         struct ib_ah *ah;
 579         int err;
 580
 581         err = rdma_fill_sgid_attr(pd->device, ah_attr, &old_sgid_attr);
 582         if (err)
 583                 return ERR_PTR(err);
 584
 585         if (ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE) {
 586                 err = ib_resolve_eth_dmac(pd->device, ah_attr);
 587                 if (err) {
 588                         ah = ERR_PTR(err);
 589                         goto out;
 590                 }
 591         }
 592
 593         ah = _rdma_create_ah(pd, ah_attr, RDMA_CREATE_AH_SLEEPABLE, udata);
 594
 595 out:
 596         rdma_unfill_sgid_attr(ah_attr, old_sgid_attr);
 597         return ah;
 598 }
 599 EXPORT_SYMBOL(rdma_create_user_ah);
 600
 601 int ib_get_rdma_header_version(const union rdma_network_hdr *hdr)
 602 {
 603         const struct iphdr *ip4h = (struct iphdr *)&hdr->roce4grh;
 604         struct iphdr ip4h_checked;
 605         const struct ipv6hdr *ip6h = (struct ipv6hdr *)&hdr->ibgrh;
 606
 607         /* If it's IPv6, the version must be 6, otherwise, the first
 608          * 20 bytes (before the IPv4 header) are garbled.
 609          */
 610         if (ip6h->version != 6)
 611                 return (ip4h->version == 4) ? 4 : 0;
 612         /* version may be 6 or 4 because the first 20 bytes could be garbled */
 613
 614         /* RoCE v2 requires no options, thus header length
 615          * must be 5 words
 616          */
 617         if (ip4h->ihl != 5)
 618                 return 6;
 619
 620         /* Verify checksum.
 621          * We can't write on scattered buffers so we need to copy to
 622          * temp buffer.
 623          */
 624         memcpy(&ip4h_checked, ip4h, sizeof(ip4h_checked));
 625         ip4h_checked.check = 0;
 626         ip4h_checked.check = ip_fast_csum((u8 *)&ip4h_checked, 5);
 627         /* if IPv4 header checksum is OK, believe it */
 628         if (ip4h->check == ip4h_checked.check)
 629                 return 4;
 630         return 6;
 631 }
 632 EXPORT_SYMBOL(ib_get_rdma_header_version);
 633
 634 static enum rdma_network_type ib_get_net_type_by_grh(struct ib_device *device,
 635                                                      u8 port_num,
 636                                                      const struct ib_grh *grh)
 637 {
 638         int grh_version;
 639
 640         if (rdma_protocol_ib(device, port_num))
 641                 return RDMA_NETWORK_IB;
 642
 643         grh_version = ib_get_rdma_header_version((union rdma_network_hdr *)grh);
 644
 645         if (grh_version == 4)
 646                 return RDMA_NETWORK_IPV4;
 647
 648         if (grh->next_hdr == IPPROTO_UDP)
 649                 return RDMA_NETWORK_IPV6;
 650
 651         return RDMA_NETWORK_ROCE_V1;
 652 }
 653
 654 struct find_gid_index_context {
 655         u16 vlan_id;
 656         enum ib_gid_type gid_type;
 657 };
 658
 659 static bool find_gid_index(const union ib_gid *gid,
 660                            const struct ib_gid_attr *gid_attr,
 661                            void *context)
 662 {
 663         struct find_gid_index_context *ctx = context;
 664
 665         if (ctx->gid_type != gid_attr->gid_type)
 666                 return false;
 667
 668         if ((!!(ctx->vlan_id != 0xffff) == !is_vlan_dev(gid_attr->ndev)) ||
 669             (is_vlan_dev(gid_attr->ndev) &&
 670              vlan_dev_vlan_id(gid_attr->ndev) != ctx->vlan_id))
 671                 return false;
 672
 673         return true;
 674 }
 675
 676 static const struct ib_gid_attr *
 677 get_sgid_attr_from_eth(struct ib_device *device, u8 port_num,
 678                        u16 vlan_id, const union ib_gid *sgid,
 679                        enum ib_gid_type gid_type)
 680 {
 681         struct find_gid_index_context context = {.vlan_id = vlan_id,
 682                                                  .gid_type = gid_type};
 683
 684         return rdma_find_gid_by_filter(device, sgid, port_num, find_gid_index,
 685                                        &context);
 686 }
 687
 688 int ib_get_gids_from_rdma_hdr(const union rdma_network_hdr *hdr,
 689                               enum rdma_network_type net_type,
 690                               union ib_gid *sgid, union ib_gid *dgid)
 691 {
 692         struct sockaddr_in  src_in;
 693         struct sockaddr_in  dst_in;
 694         __be32 src_saddr, dst_saddr;
 695
 696         if (!sgid || !dgid)
 697                 return -EINVAL;
 698
 699         if (net_type == RDMA_NETWORK_IPV4) {
 700                 memcpy(&src_in.sin_addr.s_addr,
 701                        &hdr->roce4grh.saddr, 4);
 702                 memcpy(&dst_in.sin_addr.s_addr,
 703                        &hdr->roce4grh.daddr, 4);
 704                 src_saddr = src_in.sin_addr.s_addr;
 705                 dst_saddr = dst_in.sin_addr.s_addr;
 706                 ipv6_addr_set_v4mapped(src_saddr,
 707                                        (struct in6_addr *)sgid);
 708                 ipv6_addr_set_v4mapped(dst_saddr,
 709                                        (struct in6_addr *)dgid);
 710                 return 0;
 711         } else if (net_type == RDMA_NETWORK_IPV6 ||
 712                    net_type == RDMA_NETWORK_IB) {
 713                 *dgid = hdr->ibgrh.dgid;
 714                 *sgid = hdr->ibgrh.sgid;
 715                 return 0;
 716         } else {
 717                 return -EINVAL;
 718         }
 719 }
 720 EXPORT_SYMBOL(ib_get_gids_from_rdma_hdr);
 721
 722 /* Resolve destination mac address and hop limit for unicast destination
 723  * GID entry, considering the source GID entry as well.
 724  * ah_attribute must have have valid port_num, sgid_index.
 725  */
 726 static int ib_resolve_unicast_gid_dmac(struct ib_device *device,
 727                                        struct rdma_ah_attr *ah_attr)
 728 {
 729         struct ib_global_route *grh = rdma_ah_retrieve_grh(ah_attr);
 730         const struct ib_gid_attr *sgid_attr = grh->sgid_attr;
 731         int hop_limit = 0xff;
 732         int ret = 0;
 733
 734         /* If destination is link local and source GID is RoCEv1,
 735          * IP stack is not used.
 736          */
 737         if (rdma_link_local_addr((struct in6_addr *)grh->dgid.raw) &&
 738             sgid_attr->gid_type == IB_GID_TYPE_ROCE) {
 739                 rdma_get_ll_mac((struct in6_addr *)grh->dgid.raw,
 740                                 ah_attr->roce.dmac);
 741                 return ret;
 742         }
 743
 744         ret = rdma_addr_find_l2_eth_by_grh(&sgid_attr->gid, &grh->dgid,
 745                                            ah_attr->roce.dmac,
 746                                            sgid_attr, &hop_limit);
 747
 748         grh->hop_limit = hop_limit;
 749         return ret;
 750 }
 751
 752 /*
 753  * This function initializes address handle attributes from the incoming packet.
 754  * Incoming packet has dgid of the receiver node on which this code is
 755  * getting executed and, sgid contains the GID of the sender.
 756  *
 757  * When resolving mac address of destination, the arrived dgid is used
 758  * as sgid and, sgid is used as dgid because sgid contains destinations
 759  * GID whom to respond to.
 760  *
 761  * On success the caller is responsible to call rdma_destroy_ah_attr on the
 762  * attr.
 763  */
 764 int ib_init_ah_attr_from_wc(struct ib_device *device, u8 port_num,
 765                             const struct ib_wc *wc, const struct ib_grh *grh,
 766                             struct rdma_ah_attr *ah_attr)
 767 {
 768         u32 flow_class;
 769         int ret;
 770         enum rdma_network_type net_type = RDMA_NETWORK_IB;
 771         enum ib_gid_type gid_type = IB_GID_TYPE_IB;
 772         const struct ib_gid_attr *sgid_attr;
 773         int hoplimit = 0xff;
 774         union ib_gid dgid;
 775         union ib_gid sgid;
 776
 777         might_sleep();
 778
 779         memset(ah_attr, 0, sizeof *ah_attr);
 780         ah_attr->type = rdma_ah_find_type(device, port_num);
 781         if (rdma_cap_eth_ah(device, port_num)) {
 782                 if (wc->wc_flags & IB_WC_WITH_NETWORK_HDR_TYPE)
 783                         net_type = wc->network_hdr_type;
 784                 else
 785                         net_type = ib_get_net_type_by_grh(device, port_num, grh);
 786                 gid_type = ib_network_to_gid_type(net_type);
 787         }
 788         ret = ib_get_gids_from_rdma_hdr((union rdma_network_hdr *)grh, net_type,
 789                                         &sgid, &dgid);
 790         if (ret)
 791                 return ret;
 792
 793         rdma_ah_set_sl(ah_attr, wc->sl);
 794         rdma_ah_set_port_num(ah_attr, port_num);
 795
 796         if (rdma_protocol_roce(device, port_num)) {
 797                 u16 vlan_id = wc->wc_flags & IB_WC_WITH_VLAN ?
 798                                 wc->vlan_id : 0xffff;
 799
 800                 if (!(wc->wc_flags & IB_WC_GRH))
 801                         return -EPROTOTYPE;
 802
 803                 sgid_attr = get_sgid_attr_from_eth(device, port_num,
 804                                                    vlan_id, &dgid,
 805                                                    gid_type);
 806                 if (IS_ERR(sgid_attr))
 807                         return PTR_ERR(sgid_attr);
 808
 809                 flow_class = be32_to_cpu(grh->version_tclass_flow);
 810                 rdma_move_grh_sgid_attr(ah_attr,
 811                                         &sgid,
 812                                         flow_class & 0xFFFFF,
 813                                         hoplimit,
 814                                         (flow_class >> 20) & 0xFF,
 815                                         sgid_attr);
 816
 817                 ret = ib_resolve_unicast_gid_dmac(device, ah_attr);
 818                 if (ret)
 819                         rdma_destroy_ah_attr(ah_attr);
 820
 821                 return ret;
 822         } else {
 823                 rdma_ah_set_dlid(ah_attr, wc->slid);
 824                 rdma_ah_set_path_bits(ah_attr, wc->dlid_path_bits);
 825
 826                 if ((wc->wc_flags & IB_WC_GRH) == 0)
 827                         return 0;
 828
 829                 if (dgid.global.interface_id !=
 830                                         cpu_to_be64(IB_SA_WELL_KNOWN_GUID)) {
 831                         sgid_attr = rdma_find_gid_by_port(
 832                                 device, &dgid, IB_GID_TYPE_IB, port_num, NULL);
 833                 } else
 834                         sgid_attr = rdma_get_gid_attr(device, port_num, 0);
 835
 836                 if (IS_ERR(sgid_attr))
 837                         return PTR_ERR(sgid_attr);
 838                 flow_class = be32_to_cpu(grh->version_tclass_flow);
 839                 rdma_move_grh_sgid_attr(ah_attr,
 840                                         &sgid,
 841                                         flow_class & 0xFFFFF,
 842                                         hoplimit,
 843                                         (flow_class >> 20) & 0xFF,
 844                                         sgid_attr);
 845
 846                 return 0;
 847         }
 848 }
 849 EXPORT_SYMBOL(ib_init_ah_attr_from_wc);
 850
 851 /**
 852  * rdma_move_grh_sgid_attr - Sets the sgid attribute of GRH, taking ownership
 853  * of the reference
 854  *
 855  * @attr:       Pointer to AH attribute structure
 856  * @dgid:       Destination GID
 857  * @flow_label: Flow label
 858  * @hop_limit:  Hop limit
 859  * @traffic_class: traffic class
 860  * @sgid_attr:  Pointer to SGID attribute
 861  *
 862  * This takes ownership of the sgid_attr reference. The caller must ensure
 863  * rdma_destroy_ah_attr() is called before destroying the rdma_ah_attr after
 864  * calling this function.
 865  */
 866 void rdma_move_grh_sgid_attr(struct rdma_ah_attr *attr, union ib_gid *dgid,
 867                              u32 flow_label, u8 hop_limit, u8 traffic_class,
 868                              const struct ib_gid_attr *sgid_attr)
 869 {
 870         rdma_ah_set_grh(attr, dgid, flow_label, sgid_attr->index, hop_limit,
 871                         traffic_class);
 872         attr->grh.sgid_attr = sgid_attr;
 873 }
 874 EXPORT_SYMBOL(rdma_move_grh_sgid_attr);
 875
 876 /**
 877  * rdma_destroy_ah_attr - Release reference to SGID attribute of
 878  * ah attribute.
 879  * @ah_attr: Pointer to ah attribute
 880  *
 881  * Release reference to the SGID attribute of the ah attribute if it is
 882  * non NULL. It is safe to call this multiple times, and safe to call it on
 883  * a zero initialized ah_attr.
 884  */
 885 void rdma_destroy_ah_attr(struct rdma_ah_attr *ah_attr)
 886 {
 887         if (ah_attr->grh.sgid_attr) {
 888                 rdma_put_gid_attr(ah_attr->grh.sgid_attr);
 889                 ah_attr->grh.sgid_attr = NULL;
 890         }
 891 }
 892 EXPORT_SYMBOL(rdma_destroy_ah_attr);
 893
 894 struct ib_ah *ib_create_ah_from_wc(struct ib_pd *pd, const struct ib_wc *wc,
 895                                    const struct ib_grh *grh, u8 port_num)
 896 {
 897         struct rdma_ah_attr ah_attr;
 898         struct ib_ah *ah;
 899         int ret;
 900
 901         ret = ib_init_ah_attr_from_wc(pd->device, port_num, wc, grh, &ah_attr);
 902         if (ret)
 903                 return ERR_PTR(ret);
 904
 905         ah = rdma_create_ah(pd, &ah_attr, RDMA_CREATE_AH_SLEEPABLE);
 906
 907         rdma_destroy_ah_attr(&ah_attr);
 908         return ah;
 909 }
 910 EXPORT_SYMBOL(ib_create_ah_from_wc);
 911
 912 int rdma_modify_ah(struct ib_ah *ah, struct rdma_ah_attr *ah_attr)
 913 {
 914         const struct ib_gid_attr *old_sgid_attr;
 915         int ret;
 916
 917         if (ah->type != ah_attr->type)
 918                 return -EINVAL;
 919
 920         ret = rdma_fill_sgid_attr(ah->device, ah_attr, &old_sgid_attr);
 921         if (ret)
 922                 return ret;
 923
 924         ret = ah->device->ops.modify_ah ?
 925                 ah->device->ops.modify_ah(ah, ah_attr) :
 926                 -EOPNOTSUPP;
 927
 928         ah->sgid_attr = rdma_update_sgid_attr(ah_attr, ah->sgid_attr);
 929         rdma_unfill_sgid_attr(ah_attr, old_sgid_attr);
 930         return ret;
 931 }
 932 EXPORT_SYMBOL(rdma_modify_ah);
 933
 934 int rdma_query_ah(struct ib_ah *ah, struct rdma_ah_attr *ah_attr)
 935 {
 936         ah_attr->grh.sgid_attr = NULL;
 937
 938         return ah->device->ops.query_ah ?
 939                 ah->device->ops.query_ah(ah, ah_attr) :
 940                 -EOPNOTSUPP;
 941 }
 942 EXPORT_SYMBOL(rdma_query_ah);
 943
 944 int rdma_destroy_ah_user(struct ib_ah *ah, u32 flags, struct ib_udata *udata)
 945 {
 946         const struct ib_gid_attr *sgid_attr = ah->sgid_attr;
 947         struct ib_pd *pd;
 948
 949         might_sleep_if(flags & RDMA_DESTROY_AH_SLEEPABLE);
 950
 951         pd = ah->pd;
 952
 953         ah->device->ops.destroy_ah(ah, flags);
 954         atomic_dec(&pd->usecnt);
 955         if (sgid_attr)
 956                 rdma_put_gid_attr(sgid_attr);
 957
 958         kfree(ah);
 959         return 0;
 960 }
 961 EXPORT_SYMBOL(rdma_destroy_ah_user);
 962
 963 /* Shared receive queues */
 964
 965 struct ib_srq *ib_create_srq(struct ib_pd *pd,
 966                              struct ib_srq_init_attr *srq_init_attr)
 967 {
 968         struct ib_srq *srq;
 969         int ret;
 970
 971         if (!pd->device->ops.create_srq)
 972                 return ERR_PTR(-EOPNOTSUPP);
 973
 974         srq = rdma_zalloc_drv_obj(pd->device, ib_srq);
 975         if (!srq)
 976                 return ERR_PTR(-ENOMEM);
 977
 978         srq->device = pd->device;
 979         srq->pd = pd;
 980         srq->event_handler = srq_init_attr->event_handler;
 981         srq->srq_context = srq_init_attr->srq_context;
 982         srq->srq_type = srq_init_attr->srq_type;
 983
 984         if (ib_srq_has_cq(srq->srq_type)) {
 985                 srq->ext.cq = srq_init_attr->ext.cq;
 986                 atomic_inc(&srq->ext.cq->usecnt);
 987         }
 988         if (srq->srq_type == IB_SRQT_XRC) {
 989                 srq->ext.xrc.xrcd = srq_init_attr->ext.xrc.xrcd;
 990                 atomic_inc(&srq->ext.xrc.xrcd->usecnt);
 991         }
 992         atomic_inc(&pd->usecnt);
 993
 994         ret = pd->device->ops.create_srq(srq, srq_init_attr, NULL);
 995         if (ret) {
 996                 atomic_dec(&srq->pd->usecnt);
 997                 if (srq->srq_type == IB_SRQT_XRC)
 998                         atomic_dec(&srq->ext.xrc.xrcd->usecnt);
 999                 if (ib_srq_has_cq(srq->srq_type))
1000                         atomic_dec(&srq->ext.cq->usecnt);
1001                 kfree(srq);
1002                 return ERR_PTR(ret);
1003         }
1004
1005         return srq;
1006 }
1007 EXPORT_SYMBOL(ib_create_srq);
1008
1009 int ib_modify_srq(struct ib_srq *srq,
1010                   struct ib_srq_attr *srq_attr,
1011                   enum ib_srq_attr_mask srq_attr_mask)
1012 {
1013         return srq->device->ops.modify_srq ?
1014                 srq->device->ops.modify_srq(srq, srq_attr, srq_attr_mask,
1015                                             NULL) : -EOPNOTSUPP;
1016 }
1017 EXPORT_SYMBOL(ib_modify_srq);
1018
1019 int ib_query_srq(struct ib_srq *srq,
1020                  struct ib_srq_attr *srq_attr)
1021 {
1022         return srq->device->ops.query_srq ?
1023                 srq->device->ops.query_srq(srq, srq_attr) : -EOPNOTSUPP;
1024 }
1025 EXPORT_SYMBOL(ib_query_srq);
1026
1027 int ib_destroy_srq_user(struct ib_srq *srq, struct ib_udata *udata)
1028 {
1029         if (atomic_read(&srq->usecnt))
1030                 return -EBUSY;
1031
1032         srq->device->ops.destroy_srq(srq, udata);
1033
1034         atomic_dec(&srq->pd->usecnt);
1035         if (srq->srq_type == IB_SRQT_XRC)
1036                 atomic_dec(&srq->ext.xrc.xrcd->usecnt);
1037         if (ib_srq_has_cq(srq->srq_type))
1038                 atomic_dec(&srq->ext.cq->usecnt);
1039         kfree(srq);
1040
1041         return 0;
1042 }
1043 EXPORT_SYMBOL(ib_destroy_srq_user);
1044
1045 /* Queue pairs */
1046
1047 static void __ib_shared_qp_event_handler(struct ib_event *event, void *context)
1048 {
1049         struct ib_qp *qp = context;
1050         unsigned long flags;
1051
1052         spin_lock_irqsave(&qp->device->event_handler_lock, flags);
1053         list_for_each_entry(event->element.qp, &qp->open_list, open_list)
1054                 if (event->element.qp->event_handler)
1055                         event->element.qp->event_handler(event, event->element.qp->qp_context);
1056         spin_unlock_irqrestore(&qp->device->event_handler_lock, flags);
1057 }
1058
1059 static void __ib_insert_xrcd_qp(struct ib_xrcd *xrcd, struct ib_qp *qp)
1060 {
1061         mutex_lock(&xrcd->tgt_qp_mutex);
1062         list_add(&qp->xrcd_list, &xrcd->tgt_qp_list);
1063         mutex_unlock(&xrcd->tgt_qp_mutex);
1064 }
1065
1066 static struct ib_qp *__ib_open_qp(struct ib_qp *real_qp,
1067                                   void (*event_handler)(struct ib_event *, void *),
1068                                   void *qp_context)
1069 {
1070         struct ib_qp *qp;
1071         unsigned long flags;
1072         int err;
1073
1074         qp = kzalloc(sizeof *qp, GFP_KERNEL);
1075         if (!qp)
1076                 return ERR_PTR(-ENOMEM);
1077
1078         qp->real_qp = real_qp;
1079         err = ib_open_shared_qp_security(qp, real_qp->device);
1080         if (err) {
1081                 kfree(qp);
1082                 return ERR_PTR(err);
1083         }
1084
1085         qp->real_qp = real_qp;
1086         atomic_inc(&real_qp->usecnt);
1087         qp->device = real_qp->device;
1088         qp->event_handler = event_handler;
1089         qp->qp_context = qp_context;
1090         qp->qp_num = real_qp->qp_num;
1091         qp->qp_type = real_qp->qp_type;
1092
1093         spin_lock_irqsave(&real_qp->device->event_handler_lock, flags);
1094         list_add(&qp->open_list, &real_qp->open_list);
1095         spin_unlock_irqrestore(&real_qp->device->event_handler_lock, flags);
1096
1097         return qp;
1098 }
1099
1100 struct ib_qp *ib_open_qp(struct ib_xrcd *xrcd,
1101                          struct ib_qp_open_attr *qp_open_attr)
1102 {
1103         struct ib_qp *qp, *real_qp;
1104
1105         if (qp_open_attr->qp_type != IB_QPT_XRC_TGT)
1106                 return ERR_PTR(-EINVAL);
1107
1108         qp = ERR_PTR(-EINVAL);
1109         mutex_lock(&xrcd->tgt_qp_mutex);
1110         list_for_each_entry(real_qp, &xrcd->tgt_qp_list, xrcd_list) {
1111                 if (real_qp->qp_num == qp_open_attr->qp_num) {
1112                         qp = __ib_open_qp(real_qp, qp_open_attr->event_handler,
1113                                           qp_open_attr->qp_context);
1114                         break;
1115                 }
1116         }
1117         mutex_unlock(&xrcd->tgt_qp_mutex);
1118         return qp;
1119 }
1120 EXPORT_SYMBOL(ib_open_qp);
1121
1122 static struct ib_qp *create_xrc_qp_user(struct ib_qp *qp,
1123                                         struct ib_qp_init_attr *qp_init_attr,
1124                                         struct ib_udata *udata)
1125 {
1126         struct ib_qp *real_qp = qp;
1127
1128         qp->event_handler = __ib_shared_qp_event_handler;
1129         qp->qp_context = qp;
1130         qp->pd = NULL;
1131         qp->send_cq = qp->recv_cq = NULL;
1132         qp->srq = NULL;
1133         qp->xrcd = qp_init_attr->xrcd;
1134         atomic_inc(&qp_init_attr->xrcd->usecnt);
1135         INIT_LIST_HEAD(&qp->open_list);
1136
1137         qp = __ib_open_qp(real_qp, qp_init_attr->event_handler,
1138                           qp_init_attr->qp_context);
1139         if (IS_ERR(qp))
1140                 return qp;
1141
1142         __ib_insert_xrcd_qp(qp_init_attr->xrcd, real_qp);
1143         return qp;
1144 }
1145
1146 struct ib_qp *ib_create_qp_user(struct ib_pd *pd,
1147                                 struct ib_qp_init_attr *qp_init_attr,
1148                                 struct ib_udata *udata)
1149 {
1150         struct ib_device *device = pd ? pd->device : qp_init_attr->xrcd->device;
1151         struct ib_qp *qp;
1152         int ret;
1153
1154         if (qp_init_attr->rwq_ind_tbl &&
1155             (qp_init_attr->recv_cq ||
1156             qp_init_attr->srq || qp_init_attr->cap.max_recv_wr ||
1157             qp_init_attr->cap.max_recv_sge))
1158                 return ERR_PTR(-EINVAL);
1159
1160         /*
1161          * If the callers is using the RDMA API calculate the resources
1162          * needed for the RDMA READ/WRITE operations.
1163          *
1164          * Note that these callers need to pass in a port number.
1165          */
1166         if (qp_init_attr->cap.max_rdma_ctxs)
1167                 rdma_rw_init_qp(device, qp_init_attr);
1168
1169         qp = _ib_create_qp(device, pd, qp_init_attr, NULL, NULL);
1170         if (IS_ERR(qp))
1171                 return qp;
1172
1173         ret = ib_create_qp_security(qp, device);
1174         if (ret)
1175                 goto err;
1176
1177         qp->qp_type    = qp_init_attr->qp_type;
1178         qp->rwq_ind_tbl = qp_init_attr->rwq_ind_tbl;
1179
1180         atomic_set(&qp->usecnt, 0);
1181         qp->mrs_used = 0;
1182         spin_lock_init(&qp->mr_lock);
1183         INIT_LIST_HEAD(&qp->rdma_mrs);
1184         INIT_LIST_HEAD(&qp->sig_mrs);
1185         qp->port = 0;
1186
1187         if (qp_init_attr->qp_type == IB_QPT_XRC_TGT) {
1188                 struct ib_qp *xrc_qp =
1189                         create_xrc_qp_user(qp, qp_init_attr, udata);
1190
1191                 if (IS_ERR(xrc_qp)) {
1192                         ret = PTR_ERR(xrc_qp);
1193                         goto err;
1194                 }
1195                 return xrc_qp;
1196         }
1197
1198         qp->event_handler = qp_init_attr->event_handler;
1199         qp->qp_context = qp_init_attr->qp_context;
1200         if (qp_init_attr->qp_type == IB_QPT_XRC_INI) {
1201                 qp->recv_cq = NULL;
1202                 qp->srq = NULL;
1203         } else {
1204                 qp->recv_cq = qp_init_attr->recv_cq;
1205                 if (qp_init_attr->recv_cq)
1206                         atomic_inc(&qp_init_attr->recv_cq->usecnt);
1207                 qp->srq = qp_init_attr->srq;
1208                 if (qp->srq)
1209                         atomic_inc(&qp_init_attr->srq->usecnt);
1210         }
1211
1212         qp->send_cq = qp_init_attr->send_cq;
1213         qp->xrcd    = NULL;
1214
1215         atomic_inc(&pd->usecnt);
1216         if (qp_init_attr->send_cq)
1217                 atomic_inc(&qp_init_attr->send_cq->usecnt);
1218         if (qp_init_attr->rwq_ind_tbl)
1219                 atomic_inc(&qp->rwq_ind_tbl->usecnt);
1220
1221         if (qp_init_attr->cap.max_rdma_ctxs) {
1222                 ret = rdma_rw_init_mrs(qp, qp_init_attr);
1223                 if (ret)
1224                         goto err;
1225         }
1226
1227         /*
1228          * Note: all hw drivers guarantee that max_send_sge is lower than
1229          * the device RDMA WRITE SGE limit but not all hw drivers ensure that
1230          * max_send_sge <= max_sge_rd.
1231          */
1232         qp->max_write_sge = qp_init_attr->cap.max_send_sge;
1233         qp->max_read_sge = min_t(u32, qp_init_attr->cap.max_send_sge,
1234                                  device->attrs.max_sge_rd);
1235
1236         return qp;
1237
1238 err:
1239         ib_destroy_qp(qp);
1240         return ERR_PTR(ret);
1241
1242 }
1243 EXPORT_SYMBOL(ib_create_qp_user);
1244
1245 static const struct {
1246         int                     valid;
1247         enum ib_qp_attr_mask    req_param[IB_QPT_MAX];
1248         enum ib_qp_attr_mask    opt_param[IB_QPT_MAX];
1249 } qp_state_table[IB_QPS_ERR + 1][IB_QPS_ERR + 1] = {
1250         [IB_QPS_RESET] = {
1251                 [IB_QPS_RESET] = { .valid = 1 },
1252                 [IB_QPS_INIT]  = {
1253                         .valid = 1,
1254                         .req_param = {
1255                                 [IB_QPT_UD]  = (IB_QP_PKEY_INDEX                |
1256                                                 IB_QP_PORT                      |
1257                                                 IB_QP_QKEY),
1258                                 [IB_QPT_RAW_PACKET] = IB_QP_PORT,
1259                                 [IB_QPT_UC]  = (IB_QP_PKEY_INDEX                |
1260                                                 IB_QP_PORT                      |
1261                                                 IB_QP_ACCESS_FLAGS),
1262                                 [IB_QPT_RC]  = (IB_QP_PKEY_INDEX                |
1263                                                 IB_QP_PORT                      |
1264                                                 IB_QP_ACCESS_FLAGS),
1265                                 [IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX            |
1266                                                 IB_QP_PORT                      |
1267                                                 IB_QP_ACCESS_FLAGS),
1268                                 [IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX            |
1269                                                 IB_QP_PORT                      |
1270                                                 IB_QP_ACCESS_FLAGS),
1271                                 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX                |
1272                                                 IB_QP_QKEY),
1273                                 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX                |
1274                                                 IB_QP_QKEY),
1275                         }
1276                 },
1277         },
1278         [IB_QPS_INIT]  = {
1279                 [IB_QPS_RESET] = { .valid = 1 },
1280                 [IB_QPS_ERR] =   { .valid = 1 },
1281                 [IB_QPS_INIT]  = {
1282                         .valid = 1,
1283                         .opt_param = {
1284                                 [IB_QPT_UD]  = (IB_QP_PKEY_INDEX                |
1285                                                 IB_QP_PORT                      |
1286                                                 IB_QP_QKEY),
1287                                 [IB_QPT_UC]  = (IB_QP_PKEY_INDEX                |
1288                                                 IB_QP_PORT                      |
1289                                                 IB_QP_ACCESS_FLAGS),
1290                                 [IB_QPT_RC]  = (IB_QP_PKEY_INDEX                |
1291                                                 IB_QP_PORT                      |
1292                                                 IB_QP_ACCESS_FLAGS),
1293                                 [IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX            |
1294                                                 IB_QP_PORT                      |
1295                                                 IB_QP_ACCESS_FLAGS),
1296                                 [IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX            |
1297                                                 IB_QP_PORT                      |
1298                                                 IB_QP_ACCESS_FLAGS),
1299                                 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX                |
1300                                                 IB_QP_QKEY),
1301                                 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX                |
1302                                                 IB_QP_QKEY),
1303                         }
1304                 },
1305                 [IB_QPS_RTR]   = {
1306                         .valid = 1,
1307                         .req_param = {
1308                                 [IB_QPT_UC]  = (IB_QP_AV                        |
1309                                                 IB_QP_PATH_MTU                  |
1310                                                 IB_QP_DEST_QPN                  |
1311                                                 IB_QP_RQ_PSN),
1312                                 [IB_QPT_RC]  = (IB_QP_AV                        |
1313                                                 IB_QP_PATH_MTU                  |
1314                                                 IB_QP_DEST_QPN                  |
1315                                                 IB_QP_RQ_PSN                    |
1316                                                 IB_QP_MAX_DEST_RD_ATOMIC        |
1317                                                 IB_QP_MIN_RNR_TIMER),
1318                                 [IB_QPT_XRC_INI] = (IB_QP_AV                    |
1319                                                 IB_QP_PATH_MTU                  |
1320                                                 IB_QP_DEST_QPN                  |
1321                                                 IB_QP_RQ_PSN),
1322                                 [IB_QPT_XRC_TGT] = (IB_QP_AV                    |
1323                                                 IB_QP_PATH_MTU                  |
1324                                                 IB_QP_DEST_QPN                  |
1325                                                 IB_QP_RQ_PSN                    |
1326                                                 IB_QP_MAX_DEST_RD_ATOMIC        |
1327                                                 IB_QP_MIN_RNR_TIMER),
1328                         },
1329                         .opt_param = {
1330                                  [IB_QPT_UD]  = (IB_QP_PKEY_INDEX               |
1331                                                  IB_QP_QKEY),
1332                                  [IB_QPT_UC]  = (IB_QP_ALT_PATH                 |
1333                                                  IB_QP_ACCESS_FLAGS             |
1334                                                  IB_QP_PKEY_INDEX),
1335                                  [IB_QPT_RC]  = (IB_QP_ALT_PATH                 |
1336                                                  IB_QP_ACCESS_FLAGS             |
1337                                                  IB_QP_PKEY_INDEX),
1338                                  [IB_QPT_XRC_INI] = (IB_QP_ALT_PATH             |
1339                                                  IB_QP_ACCESS_FLAGS             |
1340                                                  IB_QP_PKEY_INDEX),
1341                                  [IB_QPT_XRC_TGT] = (IB_QP_ALT_PATH             |
1342                                                  IB_QP_ACCESS_FLAGS             |
1343                                                  IB_QP_PKEY_INDEX),
1344                                  [IB_QPT_SMI] = (IB_QP_PKEY_INDEX               |
1345                                                  IB_QP_QKEY),
1346                                  [IB_QPT_GSI] = (IB_QP_PKEY_INDEX               |
1347                                                  IB_QP_QKEY),
1348                          },
1349                 },
1350         },
1351         [IB_QPS_RTR]   = {
1352                 [IB_QPS_RESET] = { .valid = 1 },
1353                 [IB_QPS_ERR] =   { .valid = 1 },
1354                 [IB_QPS_RTS]   = {
1355                         .valid = 1,
1356                         .req_param = {
1357                                 [IB_QPT_UD]  = IB_QP_SQ_PSN,
1358                                 [IB_QPT_UC]  = IB_QP_SQ_PSN,
1359                                 [IB_QPT_RC]  = (IB_QP_TIMEOUT                   |
1360                                                 IB_QP_RETRY_CNT                 |
1361                                                 IB_QP_RNR_RETRY                 |
1362                                                 IB_QP_SQ_PSN                    |
1363                                                 IB_QP_MAX_QP_RD_ATOMIC),
1364                                 [IB_QPT_XRC_INI] = (IB_QP_TIMEOUT               |
1365                                                 IB_QP_RETRY_CNT                 |
1366                                                 IB_QP_RNR_RETRY                 |
1367                                                 IB_QP_SQ_PSN                    |
1368                                                 IB_QP_MAX_QP_RD_ATOMIC),
1369                                 [IB_QPT_XRC_TGT] = (IB_QP_TIMEOUT               |
1370                                                 IB_QP_SQ_PSN),
1371                                 [IB_QPT_SMI] = IB_QP_SQ_PSN,
1372                                 [IB_QPT_GSI] = IB_QP_SQ_PSN,
1373                         },
1374                         .opt_param = {
1375                                  [IB_QPT_UD]  = (IB_QP_CUR_STATE                |
1376                                                  IB_QP_QKEY),
1377                                  [IB_QPT_UC]  = (IB_QP_CUR_STATE                |
1378                                                  IB_QP_ALT_PATH                 |
1379                                                  IB_QP_ACCESS_FLAGS             |
1380                                                  IB_QP_PATH_MIG_STATE),
1381                                  [IB_QPT_RC]  = (IB_QP_CUR_STATE                |
1382                                                  IB_QP_ALT_PATH                 |
1383                                                  IB_QP_ACCESS_FLAGS             |
1384                                                  IB_QP_MIN_RNR_TIMER            |
1385                                                  IB_QP_PATH_MIG_STATE),
1386                                  [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE            |
1387                                                  IB_QP_ALT_PATH                 |
1388                                                  IB_QP_ACCESS_FLAGS             |
1389                                                  IB_QP_PATH_MIG_STATE),
1390                                  [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE            |
1391                                                  IB_QP_ALT_PATH                 |
1392                                                  IB_QP_ACCESS_FLAGS             |
1393                                                  IB_QP_MIN_RNR_TIMER            |
1394                                                  IB_QP_PATH_MIG_STATE),
1395                                  [IB_QPT_SMI] = (IB_QP_CUR_STATE                |
1396                                                  IB_QP_QKEY),
1397                                  [IB_QPT_GSI] = (IB_QP_CUR_STATE                |
1398                                                  IB_QP_QKEY),
1399                                  [IB_QPT_RAW_PACKET] = IB_QP_RATE_LIMIT,
1400                          }
1401                 }
1402         },
1403         [IB_QPS_RTS]   = {
1404                 [IB_QPS_RESET] = { .valid = 1 },
1405                 [IB_QPS_ERR] =   { .valid = 1 },
1406                 [IB_QPS_RTS]   = {
1407                         .valid = 1,
1408                         .opt_param = {
1409                                 [IB_QPT_UD]  = (IB_QP_CUR_STATE                 |
1410                                                 IB_QP_QKEY),
1411                                 [IB_QPT_UC]  = (IB_QP_CUR_STATE                 |
1412                                                 IB_QP_ACCESS_FLAGS              |
1413                                                 IB_QP_ALT_PATH                  |
1414                                                 IB_QP_PATH_MIG_STATE),
1415                                 [IB_QPT_RC]  = (IB_QP_CUR_STATE                 |
1416                                                 IB_QP_ACCESS_FLAGS              |
1417                                                 IB_QP_ALT_PATH                  |
1418                                                 IB_QP_PATH_MIG_STATE            |
1419                                                 IB_QP_MIN_RNR_TIMER),
1420                                 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE             |
1421                                                 IB_QP_ACCESS_FLAGS              |
1422                                                 IB_QP_ALT_PATH                  |
1423                                                 IB_QP_PATH_MIG_STATE),
1424                                 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE             |
1425                                                 IB_QP_ACCESS_FLAGS              |
1426                                                 IB_QP_ALT_PATH                  |
1427                                                 IB_QP_PATH_MIG_STATE            |
1428                                                 IB_QP_MIN_RNR_TIMER),
1429                                 [IB_QPT_SMI] = (IB_QP_CUR_STATE                 |
1430                                                 IB_QP_QKEY),
1431                                 [IB_QPT_GSI] = (IB_QP_CUR_STATE                 |
1432                                                 IB_QP_QKEY),
1433                                 [IB_QPT_RAW_PACKET] = IB_QP_RATE_LIMIT,
1434                         }
1435                 },
1436                 [IB_QPS_SQD]   = {
1437                         .valid = 1,
1438                         .opt_param = {
1439                                 [IB_QPT_UD]  = IB_QP_EN_SQD_ASYNC_NOTIFY,
1440                                 [IB_QPT_UC]  = IB_QP_EN_SQD_ASYNC_NOTIFY,
1441                                 [IB_QPT_RC]  = IB_QP_EN_SQD_ASYNC_NOTIFY,
1442                                 [IB_QPT_XRC_INI] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1443                                 [IB_QPT_XRC_TGT] = IB_QP_EN_SQD_ASYNC_NOTIFY, /* ??? */
1444                                 [IB_QPT_SMI] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1445                                 [IB_QPT_GSI] = IB_QP_EN_SQD_ASYNC_NOTIFY
1446                         }
1447                 },
1448         },
1449         [IB_QPS_SQD]   = {
1450                 [IB_QPS_RESET] = { .valid = 1 },
1451                 [IB_QPS_ERR] =   { .valid = 1 },
1452                 [IB_QPS_RTS]   = {
1453                         .valid = 1,
1454                         .opt_param = {
1455                                 [IB_QPT_UD]  = (IB_QP_CUR_STATE                 |
1456                                                 IB_QP_QKEY),
1457                                 [IB_QPT_UC]  = (IB_QP_CUR_STATE                 |
1458                                                 IB_QP_ALT_PATH                  |
1459                                                 IB_QP_ACCESS_FLAGS              |
1460                                                 IB_QP_PATH_MIG_STATE),
1461                                 [IB_QPT_RC]  = (IB_QP_CUR_STATE                 |
1462                                                 IB_QP_ALT_PATH                  |
1463                                                 IB_QP_ACCESS_FLAGS              |
1464                                                 IB_QP_MIN_RNR_TIMER             |
1465                                                 IB_QP_PATH_MIG_STATE),
1466                                 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE             |
1467                                                 IB_QP_ALT_PATH                  |
1468                                                 IB_QP_ACCESS_FLAGS              |
1469                                                 IB_QP_PATH_MIG_STATE),
1470                                 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE             |
1471                                                 IB_QP_ALT_PATH                  |
1472                                                 IB_QP_ACCESS_FLAGS              |
1473                                                 IB_QP_MIN_RNR_TIMER             |
1474                                                 IB_QP_PATH_MIG_STATE),
1475                                 [IB_QPT_SMI] = (IB_QP_CUR_STATE                 |
1476                                                 IB_QP_QKEY),
1477                                 [IB_QPT_GSI] = (IB_QP_CUR_STATE                 |
1478                                                 IB_QP_QKEY),
1479                         }
1480                 },
1481                 [IB_QPS_SQD]   = {
1482                         .valid = 1,
1483                         .opt_param = {
1484                                 [IB_QPT_UD]  = (IB_QP_PKEY_INDEX                |
1485                                                 IB_QP_QKEY),
1486                                 [IB_QPT_UC]  = (IB_QP_AV                        |
1487                                                 IB_QP_ALT_PATH                  |
1488                                                 IB_QP_ACCESS_FLAGS              |
1489                                                 IB_QP_PKEY_INDEX                |
1490                                                 IB_QP_PATH_MIG_STATE),
1491                                 [IB_QPT_RC]  = (IB_QP_PORT                      |
1492                                                 IB_QP_AV                        |
1493                                                 IB_QP_TIMEOUT                   |
1494                                                 IB_QP_RETRY_CNT                 |
1495                                                 IB_QP_RNR_RETRY                 |
1496                                                 IB_QP_MAX_QP_RD_ATOMIC          |
1497                                                 IB_QP_MAX_DEST_RD_ATOMIC        |
1498                                                 IB_QP_ALT_PATH                  |
1499                                                 IB_QP_ACCESS_FLAGS              |
1500                                                 IB_QP_PKEY_INDEX                |
1501                                                 IB_QP_MIN_RNR_TIMER             |
1502                                                 IB_QP_PATH_MIG_STATE),
1503                                 [IB_QPT_XRC_INI] = (IB_QP_PORT                  |
1504                                                 IB_QP_AV                        |
1505                                                 IB_QP_TIMEOUT                   |
1506                                                 IB_QP_RETRY_CNT                 |
1507                                                 IB_QP_RNR_RETRY                 |
1508                                                 IB_QP_MAX_QP_RD_ATOMIC          |
1509                                                 IB_QP_ALT_PATH                  |
1510                                                 IB_QP_ACCESS_FLAGS              |
1511                                                 IB_QP_PKEY_INDEX                |
1512                                                 IB_QP_PATH_MIG_STATE),
1513                                 [IB_QPT_XRC_TGT] = (IB_QP_PORT                  |
1514                                                 IB_QP_AV                        |
1515                                                 IB_QP_TIMEOUT                   |
1516                                                 IB_QP_MAX_DEST_RD_ATOMIC        |
1517                                                 IB_QP_ALT_PATH                  |
1518                                                 IB_QP_ACCESS_FLAGS              |
1519                                                 IB_QP_PKEY_INDEX                |
1520                                                 IB_QP_MIN_RNR_TIMER             |
1521                                                 IB_QP_PATH_MIG_STATE),
1522                                 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX                |
1523                                                 IB_QP_QKEY),
1524                                 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX                |
1525                                                 IB_QP_QKEY),
1526                         }
1527                 }
1528         },
1529         [IB_QPS_SQE]   = {
1530                 [IB_QPS_RESET] = { .valid = 1 },
1531                 [IB_QPS_ERR] =   { .valid = 1 },
1532                 [IB_QPS_RTS]   = {
1533                         .valid = 1,
1534                         .opt_param = {
1535                                 [IB_QPT_UD]  = (IB_QP_CUR_STATE                 |
1536                                                 IB_QP_QKEY),
1537                                 [IB_QPT_UC]  = (IB_QP_CUR_STATE                 |
1538                                                 IB_QP_ACCESS_FLAGS),
1539                                 [IB_QPT_SMI] = (IB_QP_CUR_STATE                 |
1540                                                 IB_QP_QKEY),
1541                                 [IB_QPT_GSI] = (IB_QP_CUR_STATE                 |
1542                                                 IB_QP_QKEY),
1543                         }
1544                 }
1545         },
1546         [IB_QPS_ERR] = {
1547                 [IB_QPS_RESET] = { .valid = 1 },
1548                 [IB_QPS_ERR] =   { .valid = 1 }
1549         }
1550 };
1551
1552 bool ib_modify_qp_is_ok(enum ib_qp_state cur_state, enum ib_qp_state next_state,
1553                         enum ib_qp_type type, enum ib_qp_attr_mask mask)
1554 {
1555         enum ib_qp_attr_mask req_param, opt_param;
1556
1557         if (mask & IB_QP_CUR_STATE  &&
1558             cur_state != IB_QPS_RTR && cur_state != IB_QPS_RTS &&
1559             cur_state != IB_QPS_SQD && cur_state != IB_QPS_SQE)
1560                 return false;
1561
1562         if (!qp_state_table[cur_state][next_state].valid)
1563                 return false;
1564
1565         req_param = qp_state_table[cur_state][next_state].req_param[type];
1566         opt_param = qp_state_table[cur_state][next_state].opt_param[type];
1567
1568         if ((mask & req_param) != req_param)
1569                 return false;
1570
1571         if (mask & ~(req_param | opt_param | IB_QP_STATE))
1572                 return false;
1573
1574         return true;
1575 }
1576 EXPORT_SYMBOL(ib_modify_qp_is_ok);
1577
1578 /**
1579  * ib_resolve_eth_dmac - Resolve destination mac address
1580  * @device:             Device to consider
1581  * @ah_attr:            address handle attribute which describes the
1582  *                      source and destination parameters
1583  * ib_resolve_eth_dmac() resolves destination mac address and L3 hop limit It
1584  * returns 0 on success or appropriate error code. It initializes the
1585  * necessary ah_attr fields when call is successful.
1586  */
1587 static int ib_resolve_eth_dmac(struct ib_device *device,
1588                                struct rdma_ah_attr *ah_attr)
1589 {
1590         int ret = 0;
1591
1592         if (rdma_is_multicast_addr((struct in6_addr *)ah_attr->grh.dgid.raw)) {
1593                 if (ipv6_addr_v4mapped((struct in6_addr *)ah_attr->grh.dgid.raw)) {
1594                         __be32 addr = 0;
1595
1596                         memcpy(&addr, ah_attr->grh.dgid.raw + 12, 4);
1597                         ip_eth_mc_map(addr, (char *)ah_attr->roce.dmac);
1598                 } else {
1599                         ipv6_eth_mc_map((struct in6_addr *)ah_attr->grh.dgid.raw,
1600                                         (char *)ah_attr->roce.dmac);
1601                 }
1602         } else {
1603                 ret = ib_resolve_unicast_gid_dmac(device, ah_attr);
1604         }
1605         return ret;
1606 }
1607
1608 static bool is_qp_type_connected(const struct ib_qp *qp)
1609 {
1610         return (qp->qp_type == IB_QPT_UC ||
1611                 qp->qp_type == IB_QPT_RC ||
1612                 qp->qp_type == IB_QPT_XRC_INI ||
1613                 qp->qp_type == IB_QPT_XRC_TGT);
1614 }
1615
1616 /**
1617  * IB core internal function to perform QP attributes modification.
1618  */
1619 static int _ib_modify_qp(struct ib_qp *qp, struct ib_qp_attr *attr,
1620                          int attr_mask, struct ib_udata *udata)
1621 {
1622         u8 port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
1623         const struct ib_gid_attr *old_sgid_attr_av;
1624         const struct ib_gid_attr *old_sgid_attr_alt_av;
1625         int ret;
1626
1627         if (attr_mask & IB_QP_AV) {
1628                 ret = rdma_fill_sgid_attr(qp->device, &attr->ah_attr,
1629                                           &old_sgid_attr_av);
1630                 if (ret)
1631                         return ret;
1632         }
1633         if (attr_mask & IB_QP_ALT_PATH) {
1634                 /*
1635                  * FIXME: This does not track the migration state, so if the
1636                  * user loads a new alternate path after the HW has migrated
1637                  * from primary->alternate we will keep the wrong
1638                  * references. This is OK for IB because the reference
1639                  * counting does not serve any functional purpose.
1640                  */
1641                 ret = rdma_fill_sgid_attr(qp->device, &attr->alt_ah_attr,
1642                                           &old_sgid_attr_alt_av);
1643                 if (ret)
1644                         goto out_av;
1645
1646                 /*
1647                  * Today the core code can only handle alternate paths and APM
1648                  * for IB. Ban them in roce mode.
1649                  */
1650                 if (!(rdma_protocol_ib(qp->device,
1651                                        attr->alt_ah_attr.port_num) &&
1652                       rdma_protocol_ib(qp->device, port))) {
1653                         ret = EINVAL;
1654                         goto out;
1655                 }
1656         }
1657
1658         /*
1659          * If the user provided the qp_attr then we have to resolve it. Kernel
1660          * users have to provide already resolved rdma_ah_attr's
1661          */
1662         if (udata && (attr_mask & IB_QP_AV) &&
1663             attr->ah_attr.type == RDMA_AH_ATTR_TYPE_ROCE &&
1664             is_qp_type_connected(qp)) {
1665                 ret = ib_resolve_eth_dmac(qp->device, &attr->ah_attr);
1666                 if (ret)
1667                         goto out;
1668         }
1669
1670         if (rdma_ib_or_roce(qp->device, port)) {
1671                 if (attr_mask & IB_QP_RQ_PSN && attr->rq_psn & ~0xffffff) {
1672                         dev_warn(&qp->device->dev,
1673                                  "%s rq_psn overflow, masking to 24 bits\n",
1674                                  __func__);
1675                         attr->rq_psn &= 0xffffff;
1676                 }
1677
1678                 if (attr_mask & IB_QP_SQ_PSN && attr->sq_psn & ~0xffffff) {
1679                         dev_warn(&qp->device->dev,
1680                                  " %s sq_psn overflow, masking to 24 bits\n",
1681                                  __func__);
1682                         attr->sq_psn &= 0xffffff;
1683                 }
1684         }
1685
1686         ret = ib_security_modify_qp(qp, attr, attr_mask, udata);
1687         if (ret)
1688                 goto out;
1689
1690         if (attr_mask & IB_QP_PORT)
1691                 qp->port = attr->port_num;
1692         if (attr_mask & IB_QP_AV)
1693                 qp->av_sgid_attr =
1694                         rdma_update_sgid_attr(&attr->ah_attr, qp->av_sgid_attr);
1695         if (attr_mask & IB_QP_ALT_PATH)
1696                 qp->alt_path_sgid_attr = rdma_update_sgid_attr(
1697                         &attr->alt_ah_attr, qp->alt_path_sgid_attr);
1698
1699 out:
1700         if (attr_mask & IB_QP_ALT_PATH)
1701                 rdma_unfill_sgid_attr(&attr->alt_ah_attr, old_sgid_attr_alt_av);
1702 out_av:
1703         if (attr_mask & IB_QP_AV)
1704                 rdma_unfill_sgid_attr(&attr->ah_attr, old_sgid_attr_av);
1705         return ret;
1706 }
1707
1708 /**
1709  * ib_modify_qp_with_udata - Modifies the attributes for the specified QP.
1710  * @ib_qp: The QP to modify.
1711  * @attr: On input, specifies the QP attributes to modify.  On output,
1712  *   the current values of selected QP attributes are returned.
1713  * @attr_mask: A bit-mask used to specify which attributes of the QP
1714  *   are being modified.
1715  * @udata: pointer to user's input output buffer information
1716  *   are being modified.
1717  * It returns 0 on success and returns appropriate error code on error.
1718  */
1719 int ib_modify_qp_with_udata(struct ib_qp *ib_qp, struct ib_qp_attr *attr,
1720                             int attr_mask, struct ib_udata *udata)
1721 {
1722         return _ib_modify_qp(ib_qp->real_qp, attr, attr_mask, udata);
1723 }
1724 EXPORT_SYMBOL(ib_modify_qp_with_udata);
1725
1726 int ib_get_eth_speed(struct ib_device *dev, u8 port_num, u8 *speed, u8 *width)
1727 {
1728         int rc;
1729         u32 netdev_speed;
1730         struct net_device *netdev;
1731         struct ethtool_link_ksettings lksettings;
1732
1733         if (rdma_port_get_link_layer(dev, port_num) != IB_LINK_LAYER_ETHERNET)
1734                 return -EINVAL;
1735
1736         netdev = ib_device_get_netdev(dev, port_num);
1737         if (!netdev)
1738                 return -ENODEV;
1739
1740         rtnl_lock();
1741         rc = __ethtool_get_link_ksettings(netdev, &lksettings);
1742         rtnl_unlock();
1743
1744         dev_put(netdev);
1745
1746         if (!rc) {
1747                 netdev_speed = lksettings.base.speed;
1748         } else {
1749                 netdev_speed = SPEED_1000;
1750                 pr_warn("%s speed is unknown, defaulting to %d\n", netdev->name,
1751                         netdev_speed);
1752         }
1753
1754         if (netdev_speed <= SPEED_1000) {
1755                 *width = IB_WIDTH_1X;
1756                 *speed = IB_SPEED_SDR;
1757         } else if (netdev_speed <= SPEED_10000) {
1758                 *width = IB_WIDTH_1X;
1759                 *speed = IB_SPEED_FDR10;
1760         } else if (netdev_speed <= SPEED_20000) {
1761                 *width = IB_WIDTH_4X;
1762                 *speed = IB_SPEED_DDR;
1763         } else if (netdev_speed <= SPEED_25000) {
1764                 *width = IB_WIDTH_1X;
1765                 *speed = IB_SPEED_EDR;
1766         } else if (netdev_speed <= SPEED_40000) {
1767                 *width = IB_WIDTH_4X;
1768                 *speed = IB_SPEED_FDR10;
1769         } else {
1770                 *width = IB_WIDTH_4X;
1771                 *speed = IB_SPEED_EDR;
1772         }
1773
1774         return 0;
1775 }
1776 EXPORT_SYMBOL(ib_get_eth_speed);
1777
1778 int ib_modify_qp(struct ib_qp *qp,
1779                  struct ib_qp_attr *qp_attr,
1780                  int qp_attr_mask)
1781 {
1782         return _ib_modify_qp(qp->real_qp, qp_attr, qp_attr_mask, NULL);
1783 }
1784 EXPORT_SYMBOL(ib_modify_qp);
1785
1786 int ib_query_qp(struct ib_qp *qp,
1787                 struct ib_qp_attr *qp_attr,
1788                 int qp_attr_mask,
1789                 struct ib_qp_init_attr *qp_init_attr)
1790 {
1791         qp_attr->ah_attr.grh.sgid_attr = NULL;
1792         qp_attr->alt_ah_attr.grh.sgid_attr = NULL;
1793
1794         return qp->device->ops.query_qp ?
1795                 qp->device->ops.query_qp(qp->real_qp, qp_attr, qp_attr_mask,
1796                                          qp_init_attr) : -EOPNOTSUPP;
1797 }
1798 EXPORT_SYMBOL(ib_query_qp);
1799
1800 int ib_close_qp(struct ib_qp *qp)
1801 {
1802         struct ib_qp *real_qp;
1803         unsigned long flags;
1804
1805         real_qp = qp->real_qp;
1806         if (real_qp == qp)
1807                 return -EINVAL;
1808
1809         spin_lock_irqsave(&real_qp->device->event_handler_lock, flags);
1810         list_del(&qp->open_list);
1811         spin_unlock_irqrestore(&real_qp->device->event_handler_lock, flags);
1812
1813         atomic_dec(&real_qp->usecnt);
1814         if (qp->qp_sec)
1815                 ib_close_shared_qp_security(qp->qp_sec);
1816         kfree(qp);
1817
1818         return 0;
1819 }
1820 EXPORT_SYMBOL(ib_close_qp);
1821
1822 static int __ib_destroy_shared_qp(struct ib_qp *qp)
1823 {
1824         struct ib_xrcd *xrcd;
1825         struct ib_qp *real_qp;
1826         int ret;
1827
1828         real_qp = qp->real_qp;
1829         xrcd = real_qp->xrcd;
1830
1831         mutex_lock(&xrcd->tgt_qp_mutex);
1832         ib_close_qp(qp);
1833         if (atomic_read(&real_qp->usecnt) == 0)
1834                 list_del(&real_qp->xrcd_list);
1835         else
1836                 real_qp = NULL;
1837         mutex_unlock(&xrcd->tgt_qp_mutex);
1838
1839         if (real_qp) {
1840                 ret = ib_destroy_qp(real_qp);
1841                 if (!ret)
1842                         atomic_dec(&xrcd->usecnt);
1843                 else
1844                         __ib_insert_xrcd_qp(xrcd, real_qp);
1845         }
1846
1847         return 0;
1848 }
1849
1850 int ib_destroy_qp_user(struct ib_qp *qp, struct ib_udata *udata)
1851 {
1852         const struct ib_gid_attr *alt_path_sgid_attr = qp->alt_path_sgid_attr;
1853         const struct ib_gid_attr *av_sgid_attr = qp->av_sgid_attr;
1854         struct ib_pd *pd;
1855         struct ib_cq *scq, *rcq;
1856         struct ib_srq *srq;
1857         struct ib_rwq_ind_table *ind_tbl;
1858         struct ib_qp_security *sec;
1859         int ret;
1860
1861         WARN_ON_ONCE(qp->mrs_used > 0);
1862
1863         if (atomic_read(&qp->usecnt))
1864                 return -EBUSY;
1865
1866         if (qp->real_qp != qp)
1867                 return __ib_destroy_shared_qp(qp);
1868
1869         pd   = qp->pd;
1870         scq  = qp->send_cq;
1871         rcq  = qp->recv_cq;
1872         srq  = qp->srq;
1873         ind_tbl = qp->rwq_ind_tbl;
1874         sec  = qp->qp_sec;
1875         if (sec)
1876                 ib_destroy_qp_security_begin(sec);
1877
1878         if (!qp->uobject)
1879                 rdma_rw_cleanup_mrs(qp);
1880
1881         rdma_restrack_del(&qp->res);
1882         ret = qp->device->ops.destroy_qp(qp, udata);
1883         if (!ret) {
1884                 if (alt_path_sgid_attr)
1885                         rdma_put_gid_attr(alt_path_sgid_attr);
1886                 if (av_sgid_attr)
1887                         rdma_put_gid_attr(av_sgid_attr);
1888                 if (pd)
1889                         atomic_dec(&pd->usecnt);
1890                 if (scq)
1891                         atomic_dec(&scq->usecnt);
1892                 if (rcq)
1893                         atomic_dec(&rcq->usecnt);
1894                 if (srq)
1895                         atomic_dec(&srq->usecnt);
1896                 if (ind_tbl)
1897                         atomic_dec(&ind_tbl->usecnt);
1898                 if (sec)
1899                         ib_destroy_qp_security_end(sec);
1900         } else {
1901                 if (sec)
1902                         ib_destroy_qp_security_abort(sec);
1903         }
1904
1905         return ret;
1906 }
1907 EXPORT_SYMBOL(ib_destroy_qp_user);
1908
1909 /* Completion queues */
1910
1911 struct ib_cq *__ib_create_cq(struct ib_device *device,
1912                              ib_comp_handler comp_handler,
1913                              void (*event_handler)(struct ib_event *, void *),
1914                              void *cq_context,
1915                              const struct ib_cq_init_attr *cq_attr,
1916                              const char *caller)
1917 {
1918         struct ib_cq *cq;
1919
1920         cq = device->ops.create_cq(device, cq_attr, NULL);
1921
1922         if (!IS_ERR(cq)) {
1923                 cq->device        = device;
1924                 cq->uobject       = NULL;
1925                 cq->comp_handler  = comp_handler;
1926                 cq->event_handler = event_handler;
1927                 cq->cq_context    = cq_context;
1928                 atomic_set(&cq->usecnt, 0);
1929                 cq->res.type = RDMA_RESTRACK_CQ;
1930                 rdma_restrack_set_task(&cq->res, caller);
1931                 rdma_restrack_kadd(&cq->res);
1932         }
1933
1934         return cq;
1935 }
1936 EXPORT_SYMBOL(__ib_create_cq);
1937
1938 int rdma_set_cq_moderation(struct ib_cq *cq, u16 cq_count, u16 cq_period)
1939 {
1940         return cq->device->ops.modify_cq ?
1941                 cq->device->ops.modify_cq(cq, cq_count,
1942                                           cq_period) : -EOPNOTSUPP;
1943 }
1944 EXPORT_SYMBOL(rdma_set_cq_moderation);
1945
1946 int ib_destroy_cq_user(struct ib_cq *cq, struct ib_udata *udata)
1947 {
1948         if (atomic_read(&cq->usecnt))
1949                 return -EBUSY;
1950
1951         rdma_restrack_del(&cq->res);
1952         return cq->device->ops.destroy_cq(cq, udata);
1953 }
1954 EXPORT_SYMBOL(ib_destroy_cq_user);
1955
1956 int ib_resize_cq(struct ib_cq *cq, int cqe)
1957 {
1958         return cq->device->ops.resize_cq ?
1959                 cq->device->ops.resize_cq(cq, cqe, NULL) : -EOPNOTSUPP;
1960 }
1961 EXPORT_SYMBOL(ib_resize_cq);
1962
1963 /* Memory regions */
1964
1965 int ib_dereg_mr_user(struct ib_mr *mr, struct ib_udata *udata)
1966 {
1967         struct ib_pd *pd = mr->pd;
1968         struct ib_dm *dm = mr->dm;
1969         int ret;
1970
1971         rdma_restrack_del(&mr->res);
1972         ret = mr->device->ops.dereg_mr(mr, udata);
1973         if (!ret) {
1974                 atomic_dec(&pd->usecnt);
1975                 if (dm)
1976                         atomic_dec(&dm->usecnt);
1977         }
1978
1979         return ret;
1980 }
1981 EXPORT_SYMBOL(ib_dereg_mr_user);
1982
1983 /**
1984  * ib_alloc_mr() - Allocates a memory region
1985  * @pd:            protection domain associated with the region
1986  * @mr_type:       memory region type
1987  * @max_num_sg:    maximum sg entries available for registration.
1988  * @udata:         user data or null for kernel objects
1989  *
1990  * Notes:
1991  * Memory registeration page/sg lists must not exceed max_num_sg.
1992  * For mr_type IB_MR_TYPE_MEM_REG, the total length cannot exceed
1993  * max_num_sg * used_page_size.
1994  *
1995  */
1996 struct ib_mr *ib_alloc_mr_user(struct ib_pd *pd, enum ib_mr_type mr_type,
1997                                u32 max_num_sg, struct ib_udata *udata)
1998 {
1999         struct ib_mr *mr;
2000
2001         if (!pd->device->ops.alloc_mr)
2002                 return ERR_PTR(-EOPNOTSUPP);
2003
2004         mr = pd->device->ops.alloc_mr(pd, mr_type, max_num_sg, udata);
2005         if (!IS_ERR(mr)) {
2006                 mr->device  = pd->device;
2007                 mr->pd      = pd;
2008                 mr->dm      = NULL;
2009                 mr->uobject = NULL;
2010                 atomic_inc(&pd->usecnt);
2011                 mr->need_inval = false;
2012                 mr->res.type = RDMA_RESTRACK_MR;
2013                 rdma_restrack_kadd(&mr->res);
2014         }
2015
2016         return mr;
2017 }
2018 EXPORT_SYMBOL(ib_alloc_mr_user);
2019
2020 /* "Fast" memory regions */
2021
2022 struct ib_fmr *ib_alloc_fmr(struct ib_pd *pd,
2023                             int mr_access_flags,
2024                             struct ib_fmr_attr *fmr_attr)
2025 {
2026         struct ib_fmr *fmr;
2027
2028         if (!pd->device->ops.alloc_fmr)
2029                 return ERR_PTR(-EOPNOTSUPP);
2030
2031         fmr = pd->device->ops.alloc_fmr(pd, mr_access_flags, fmr_attr);
2032         if (!IS_ERR(fmr)) {
2033                 fmr->device = pd->device;
2034                 fmr->pd     = pd;
2035                 atomic_inc(&pd->usecnt);
2036         }
2037
2038         return fmr;
2039 }
2040 EXPORT_SYMBOL(ib_alloc_fmr);
2041
2042 int ib_unmap_fmr(struct list_head *fmr_list)
2043 {
2044         struct ib_fmr *fmr;
2045
2046         if (list_empty(fmr_list))
2047                 return 0;
2048
2049         fmr = list_entry(fmr_list->next, struct ib_fmr, list);
2050         return fmr->device->ops.unmap_fmr(fmr_list);
2051 }
2052 EXPORT_SYMBOL(ib_unmap_fmr);
2053
2054 int ib_dealloc_fmr(struct ib_fmr *fmr)
2055 {
2056         struct ib_pd *pd;
2057         int ret;
2058
2059         pd = fmr->pd;
2060         ret = fmr->device->ops.dealloc_fmr(fmr);
2061         if (!ret)
2062                 atomic_dec(&pd->usecnt);
2063
2064         return ret;
2065 }
2066 EXPORT_SYMBOL(ib_dealloc_fmr);
2067
2068 /* Multicast groups */
2069
2070 static bool is_valid_mcast_lid(struct ib_qp *qp, u16 lid)
2071 {
2072         struct ib_qp_init_attr init_attr = {};
2073         struct ib_qp_attr attr = {};
2074         int num_eth_ports = 0;
2075         int port;
2076
2077         /* If QP state >= init, it is assigned to a port and we can check this
2078          * port only.
2079          */
2080         if (!ib_query_qp(qp, &attr, IB_QP_STATE | IB_QP_PORT, &init_attr)) {
2081                 if (attr.qp_state >= IB_QPS_INIT) {
2082                         if (rdma_port_get_link_layer(qp->device, attr.port_num) !=
2083                             IB_LINK_LAYER_INFINIBAND)
2084                                 return true;
2085                         goto lid_check;
2086                 }
2087         }
2088
2089         /* Can't get a quick answer, iterate over all ports */
2090         for (port = 0; port < qp->device->phys_port_cnt; port++)
2091                 if (rdma_port_get_link_layer(qp->device, port) !=
2092                     IB_LINK_LAYER_INFINIBAND)
2093                         num_eth_ports++;
2094
2095         /* If we have at lease one Ethernet port, RoCE annex declares that
2096          * multicast LID should be ignored. We can't tell at this step if the
2097          * QP belongs to an IB or Ethernet port.
2098          */
2099         if (num_eth_ports)
2100                 return true;
2101
2102         /* If all the ports are IB, we can check according to IB spec. */
2103 lid_check:
2104         return !(lid < be16_to_cpu(IB_MULTICAST_LID_BASE) ||
2105                  lid == be16_to_cpu(IB_LID_PERMISSIVE));
2106 }
2107
2108 int ib_attach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid)
2109 {
2110         int ret;
2111
2112         if (!qp->device->ops.attach_mcast)
2113                 return -EOPNOTSUPP;
2114
2115         if (!rdma_is_multicast_addr((struct in6_addr *)gid->raw) ||
2116             qp->qp_type != IB_QPT_UD || !is_valid_mcast_lid(qp, lid))
2117                 return -EINVAL;
2118
2119         ret = qp->device->ops.attach_mcast(qp, gid, lid);
2120         if (!ret)
2121                 atomic_inc(&qp->usecnt);
2122         return ret;
2123 }
2124 EXPORT_SYMBOL(ib_attach_mcast);
2125
2126 int ib_detach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid)
2127 {
2128         int ret;
2129
2130         if (!qp->device->ops.detach_mcast)
2131                 return -EOPNOTSUPP;
2132
2133         if (!rdma_is_multicast_addr((struct in6_addr *)gid->raw) ||
2134             qp->qp_type != IB_QPT_UD || !is_valid_mcast_lid(qp, lid))
2135                 return -EINVAL;
2136
2137         ret = qp->device->ops.detach_mcast(qp, gid, lid);
2138         if (!ret)
2139                 atomic_dec(&qp->usecnt);
2140         return ret;
2141 }
2142 EXPORT_SYMBOL(ib_detach_mcast);
2143
2144 struct ib_xrcd *__ib_alloc_xrcd(struct ib_device *device, const char *caller)
2145 {
2146         struct ib_xrcd *xrcd;
2147
2148         if (!device->ops.alloc_xrcd)
2149                 return ERR_PTR(-EOPNOTSUPP);
2150
2151         xrcd = device->ops.alloc_xrcd(device, NULL);
2152         if (!IS_ERR(xrcd)) {
2153                 xrcd->device = device;
2154                 xrcd->inode = NULL;
2155                 atomic_set(&xrcd->usecnt, 0);
2156                 mutex_init(&xrcd->tgt_qp_mutex);
2157                 INIT_LIST_HEAD(&xrcd->tgt_qp_list);
2158         }
2159
2160         return xrcd;
2161 }
2162 EXPORT_SYMBOL(__ib_alloc_xrcd);
2163
2164 int ib_dealloc_xrcd(struct ib_xrcd *xrcd, struct ib_udata *udata)
2165 {
2166         struct ib_qp *qp;
2167         int ret;
2168
2169         if (atomic_read(&xrcd->usecnt))
2170                 return -EBUSY;
2171
2172         while (!list_empty(&xrcd->tgt_qp_list)) {
2173                 qp = list_entry(xrcd->tgt_qp_list.next, struct ib_qp, xrcd_list);
2174                 ret = ib_destroy_qp(qp);
2175                 if (ret)
2176                         return ret;
2177         }
2178
2179         return xrcd->device->ops.dealloc_xrcd(xrcd, udata);
2180 }
2181 EXPORT_SYMBOL(ib_dealloc_xrcd);
2182
2183 /**
2184  * ib_create_wq - Creates a WQ associated with the specified protection
2185  * domain.
2186  * @pd: The protection domain associated with the WQ.
2187  * @wq_attr: A list of initial attributes required to create the
2188  * WQ. If WQ creation succeeds, then the attributes are updated to
2189  * the actual capabilities of the created WQ.
2190  *
2191  * wq_attr->max_wr and wq_attr->max_sge determine
2192  * the requested size of the WQ, and set to the actual values allocated
2193  * on return.
2194  * If ib_create_wq() succeeds, then max_wr and max_sge will always be
2195  * at least as large as the requested values.
2196  */
2197 struct ib_wq *ib_create_wq(struct ib_pd *pd,
2198                            struct ib_wq_init_attr *wq_attr)
2199 {
2200         struct ib_wq *wq;
2201
2202         if (!pd->device->ops.create_wq)
2203                 return ERR_PTR(-EOPNOTSUPP);
2204
2205         wq = pd->device->ops.create_wq(pd, wq_attr, NULL);
2206         if (!IS_ERR(wq)) {
2207                 wq->event_handler = wq_attr->event_handler;
2208                 wq->wq_context = wq_attr->wq_context;
2209                 wq->wq_type = wq_attr->wq_type;
2210                 wq->cq = wq_attr->cq;
2211                 wq->device = pd->device;
2212                 wq->pd = pd;
2213                 wq->uobject = NULL;
2214                 atomic_inc(&pd->usecnt);
2215                 atomic_inc(&wq_attr->cq->usecnt);
2216                 atomic_set(&wq->usecnt, 0);
2217         }
2218         return wq;
2219 }
2220 EXPORT_SYMBOL(ib_create_wq);
2221
2222 /**
2223  * ib_destroy_wq - Destroys the specified user WQ.
2224  * @wq: The WQ to destroy.
2225  * @udata: Valid user data
2226  */
2227 int ib_destroy_wq(struct ib_wq *wq, struct ib_udata *udata)
2228 {
2229         int err;
2230         struct ib_cq *cq = wq->cq;
2231         struct ib_pd *pd = wq->pd;
2232
2233         if (atomic_read(&wq->usecnt))
2234                 return -EBUSY;
2235
2236         err = wq->device->ops.destroy_wq(wq, udata);
2237         if (!err) {
2238                 atomic_dec(&pd->usecnt);
2239                 atomic_dec(&cq->usecnt);
2240         }
2241         return err;
2242 }
2243 EXPORT_SYMBOL(ib_destroy_wq);
2244
2245 /**
2246  * ib_modify_wq - Modifies the specified WQ.
2247  * @wq: The WQ to modify.
2248  * @wq_attr: On input, specifies the WQ attributes to modify.
2249  * @wq_attr_mask: A bit-mask used to specify which attributes of the WQ
2250  *   are being modified.
2251  * On output, the current values of selected WQ attributes are returned.
2252  */
2253 int ib_modify_wq(struct ib_wq *wq, struct ib_wq_attr *wq_attr,
2254                  u32 wq_attr_mask)
2255 {
2256         int err;
2257
2258         if (!wq->device->ops.modify_wq)
2259                 return -EOPNOTSUPP;
2260
2261         err = wq->device->ops.modify_wq(wq, wq_attr, wq_attr_mask, NULL);
2262         return err;
2263 }
2264 EXPORT_SYMBOL(ib_modify_wq);
2265
2266 /*
2267  * ib_create_rwq_ind_table - Creates a RQ Indirection Table.
2268  * @device: The device on which to create the rwq indirection table.
2269  * @ib_rwq_ind_table_init_attr: A list of initial attributes required to
2270  * create the Indirection Table.
2271  *
2272  * Note: The life time of ib_rwq_ind_table_init_attr->ind_tbl is not less
2273  *      than the created ib_rwq_ind_table object and the caller is responsible
2274  *      for its memory allocation/free.
2275  */
2276 struct ib_rwq_ind_table *ib_create_rwq_ind_table(struct ib_device *device,
2277                                                  struct ib_rwq_ind_table_init_attr *init_attr)
2278 {
2279         struct ib_rwq_ind_table *rwq_ind_table;
2280         int i;
2281         u32 table_size;
2282
2283         if (!device->ops.create_rwq_ind_table)
2284                 return ERR_PTR(-EOPNOTSUPP);
2285
2286         table_size = (1 << init_attr->log_ind_tbl_size);
2287         rwq_ind_table = device->ops.create_rwq_ind_table(device,
2288                                                          init_attr, NULL);
2289         if (IS_ERR(rwq_ind_table))
2290                 return rwq_ind_table;
2291
2292         rwq_ind_table->ind_tbl = init_attr->ind_tbl;
2293         rwq_ind_table->log_ind_tbl_size = init_attr->log_ind_tbl_size;
2294         rwq_ind_table->device = device;
2295         rwq_ind_table->uobject = NULL;
2296         atomic_set(&rwq_ind_table->usecnt, 0);
2297
2298         for (i = 0; i < table_size; i++)
2299                 atomic_inc(&rwq_ind_table->ind_tbl[i]->usecnt);
2300
2301         return rwq_ind_table;
2302 }
2303 EXPORT_SYMBOL(ib_create_rwq_ind_table);
2304
2305 /*
2306  * ib_destroy_rwq_ind_table - Destroys the specified Indirection Table.
2307  * @wq_ind_table: The Indirection Table to destroy.
2308 */
2309 int ib_destroy_rwq_ind_table(struct ib_rwq_ind_table *rwq_ind_table)
2310 {
2311         int err, i;
2312         u32 table_size = (1 << rwq_ind_table->log_ind_tbl_size);
2313         struct ib_wq **ind_tbl = rwq_ind_table->ind_tbl;
2314
2315         if (atomic_read(&rwq_ind_table->usecnt))
2316                 return -EBUSY;
2317
2318         err = rwq_ind_table->device->ops.destroy_rwq_ind_table(rwq_ind_table);
2319         if (!err) {
2320                 for (i = 0; i < table_size; i++)
2321                         atomic_dec(&ind_tbl[i]->usecnt);
2322         }
2323
2324         return err;
2325 }
2326 EXPORT_SYMBOL(ib_destroy_rwq_ind_table);
2327
2328 int ib_check_mr_status(struct ib_mr *mr, u32 check_mask,
2329                        struct ib_mr_status *mr_status)
2330 {
2331         if (!mr->device->ops.check_mr_status)
2332                 return -EOPNOTSUPP;
2333
2334         return mr->device->ops.check_mr_status(mr, check_mask, mr_status);
2335 }
2336 EXPORT_SYMBOL(ib_check_mr_status);
2337
2338 int ib_set_vf_link_state(struct ib_device *device, int vf, u8 port,
2339                          int state)
2340 {
2341         if (!device->ops.set_vf_link_state)
2342                 return -EOPNOTSUPP;
2343
2344         return device->ops.set_vf_link_state(device, vf, port, state);
2345 }
2346 EXPORT_SYMBOL(ib_set_vf_link_state);
2347
2348 int ib_get_vf_config(struct ib_device *device, int vf, u8 port,
2349                      struct ifla_vf_info *info)
2350 {
2351         if (!device->ops.get_vf_config)
2352                 return -EOPNOTSUPP;
2353
2354         return device->ops.get_vf_config(device, vf, port, info);
2355 }
2356 EXPORT_SYMBOL(ib_get_vf_config);
2357
2358 int ib_get_vf_stats(struct ib_device *device, int vf, u8 port,
2359                     struct ifla_vf_stats *stats)
2360 {
2361         if (!device->ops.get_vf_stats)
2362                 return -EOPNOTSUPP;
2363
2364         return device->ops.get_vf_stats(device, vf, port, stats);
2365 }
2366 EXPORT_SYMBOL(ib_get_vf_stats);
2367
2368 int ib_set_vf_guid(struct ib_device *device, int vf, u8 port, u64 guid,
2369                    int type)
2370 {
2371         if (!device->ops.set_vf_guid)
2372                 return -EOPNOTSUPP;
2373
2374         return device->ops.set_vf_guid(device, vf, port, guid, type);
2375 }
2376 EXPORT_SYMBOL(ib_set_vf_guid);
2377
2378 /**
2379  * ib_map_mr_sg() - Map the largest prefix of a dma mapped SG list
2380  *     and set it the memory region.
2381  * @mr:            memory region
2382  * @sg:            dma mapped scatterlist
2383  * @sg_nents:      number of entries in sg
2384  * @sg_offset:     offset in bytes into sg
2385  * @page_size:     page vector desired page size
2386  *
2387  * Constraints:
2388  * - The first sg element is allowed to have an offset.
2389  * - Each sg element must either be aligned to page_size or virtually
2390  *   contiguous to the previous element. In case an sg element has a
2391  *   non-contiguous offset, the mapping prefix will not include it.
2392  * - The last sg element is allowed to have length less than page_size.
2393  * - If sg_nents total byte length exceeds the mr max_num_sge * page_size
2394  *   then only max_num_sg entries will be mapped.
2395  * - If the MR was allocated with type IB_MR_TYPE_SG_GAPS, none of these
2396  *   constraints holds and the page_size argument is ignored.
2397  *
2398  * Returns the number of sg elements that were mapped to the memory region.
2399  *
2400  * After this completes successfully, the  memory region
2401  * is ready for registration.
2402  */
2403 int ib_map_mr_sg(struct ib_mr *mr, struct scatterlist *sg, int sg_nents,
2404                  unsigned int *sg_offset, unsigned int page_size)
2405 {
2406         if (unlikely(!mr->device->ops.map_mr_sg))
2407                 return -EOPNOTSUPP;
2408
2409         mr->page_size = page_size;
2410
2411         return mr->device->ops.map_mr_sg(mr, sg, sg_nents, sg_offset);
2412 }
2413 EXPORT_SYMBOL(ib_map_mr_sg);
2414
2415 /**
2416  * ib_sg_to_pages() - Convert the largest prefix of a sg list
2417  *     to a page vector
2418  * @mr:            memory region
2419  * @sgl:           dma mapped scatterlist
2420  * @sg_nents:      number of entries in sg
2421  * @sg_offset_p:   IN:  start offset in bytes into sg
2422  *                 OUT: offset in bytes for element n of the sg of the first
2423  *                      byte that has not been processed where n is the return
2424  *                      value of this function.
2425  * @set_page:      driver page assignment function pointer
2426  *
2427  * Core service helper for drivers to convert the largest
2428  * prefix of given sg list to a page vector. The sg list
2429  * prefix converted is the prefix that meet the requirements
2430  * of ib_map_mr_sg.
2431  *
2432  * Returns the number of sg elements that were assigned to
2433  * a page vector.
2434  */
2435 int ib_sg_to_pages(struct ib_mr *mr, struct scatterlist *sgl, int sg_nents,
2436                 unsigned int *sg_offset_p, int (*set_page)(struct ib_mr *, u64))
2437 {
2438         struct scatterlist *sg;
2439         u64 last_end_dma_addr = 0;
2440         unsigned int sg_offset = sg_offset_p ? *sg_offset_p : 0;
2441         unsigned int last_page_off = 0;
2442         u64 page_mask = ~((u64)mr->page_size - 1);
2443         int i, ret;
2444
2445         if (unlikely(sg_nents <= 0 || sg_offset > sg_dma_len(&sgl[0])))
2446                 return -EINVAL;
2447
2448         mr->iova = sg_dma_address(&sgl[0]) + sg_offset;
2449         mr->length = 0;
2450
2451         for_each_sg(sgl, sg, sg_nents, i) {
2452                 u64 dma_addr = sg_dma_address(sg) + sg_offset;
2453                 u64 prev_addr = dma_addr;
2454                 unsigned int dma_len = sg_dma_len(sg) - sg_offset;
2455                 u64 end_dma_addr = dma_addr + dma_len;
2456                 u64 page_addr = dma_addr & page_mask;
2457
2458                 /*
2459                  * For the second and later elements, check whether either the
2460                  * end of element i-1 or the start of element i is not aligned
2461                  * on a page boundary.
2462                  */
2463                 if (i && (last_page_off != 0 || page_addr != dma_addr)) {
2464                         /* Stop mapping if there is a gap. */
2465                         if (last_end_dma_addr != dma_addr)
2466                                 break;
2467
2468                         /*
2469                          * Coalesce this element with the last. If it is small
2470                          * enough just update mr->length. Otherwise start
2471                          * mapping from the next page.
2472                          */
2473                         goto next_page;
2474                 }
2475
2476                 do {
2477                         ret = set_page(mr, page_addr);
2478                         if (unlikely(ret < 0)) {
2479                                 sg_offset = prev_addr - sg_dma_address(sg);
2480                                 mr->length += prev_addr - dma_addr;
2481                                 if (sg_offset_p)
2482                                         *sg_offset_p = sg_offset;
2483                                 return i || sg_offset ? i : ret;
2484                         }
2485                         prev_addr = page_addr;
2486 next_page:
2487                         page_addr += mr->page_size;
2488                 } while (page_addr < end_dma_addr);
2489
2490                 mr->length += dma_len;
2491                 last_end_dma_addr = end_dma_addr;
2492                 last_page_off = end_dma_addr & ~page_mask;
2493
2494                 sg_offset = 0;
2495         }
2496
2497         if (sg_offset_p)
2498                 *sg_offset_p = 0;
2499         return i;
2500 }
2501 EXPORT_SYMBOL(ib_sg_to_pages);
2502
2503 struct ib_drain_cqe {
2504         struct ib_cqe cqe;
2505         struct completion done;
2506 };
2507
2508 static void ib_drain_qp_done(struct ib_cq *cq, struct ib_wc *wc)
2509 {
2510         struct ib_drain_cqe *cqe = container_of(wc->wr_cqe, struct ib_drain_cqe,
2511                                                 cqe);
2512
2513         complete(&cqe->done);
2514 }
2515
2516 /*
2517  * Post a WR and block until its completion is reaped for the SQ.
2518  */
2519 static void __ib_drain_sq(struct ib_qp *qp)
2520 {
2521         struct ib_cq *cq = qp->send_cq;
2522         struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
2523         struct ib_drain_cqe sdrain;
2524         struct ib_rdma_wr swr = {
2525                 .wr = {
2526                         .next = NULL,
2527                         { .wr_cqe       = &sdrain.cqe, },
2528                         .opcode = IB_WR_RDMA_WRITE,
2529                 },
2530         };
2531         int ret;
2532
2533         ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
2534         if (ret) {
2535                 WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
2536                 return;
2537         }
2538
2539         sdrain.cqe.done = ib_drain_qp_done;
2540         init_completion(&sdrain.done);
2541
2542         ret = ib_post_send(qp, &swr.wr, NULL);
2543         if (ret) {
2544                 WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
2545                 return;
2546         }
2547
2548         if (cq->poll_ctx == IB_POLL_DIRECT)
2549                 while (wait_for_completion_timeout(&sdrain.done, HZ / 10) <= 0)
2550                         ib_process_cq_direct(cq, -1);
2551         else
2552                 wait_for_completion(&sdrain.done);
2553 }
2554
2555 /*
2556  * Post a WR and block until its completion is reaped for the RQ.
2557  */
2558 static void __ib_drain_rq(struct ib_qp *qp)
2559 {
2560         struct ib_cq *cq = qp->recv_cq;
2561         struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
2562         struct ib_drain_cqe rdrain;
2563         struct ib_recv_wr rwr = {};
2564         int ret;
2565
2566         ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
2567         if (ret) {
2568                 WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
2569                 return;
2570         }
2571
2572         rwr.wr_cqe = &rdrain.cqe;
2573         rdrain.cqe.done = ib_drain_qp_done;
2574         init_completion(&rdrain.done);
2575
2576         ret = ib_post_recv(qp, &rwr, NULL);
2577         if (ret) {
2578                 WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
2579                 return;
2580         }
2581
2582         if (cq->poll_ctx == IB_POLL_DIRECT)
2583                 while (wait_for_completion_timeout(&rdrain.done, HZ / 10) <= 0)
2584                         ib_process_cq_direct(cq, -1);
2585         else
2586                 wait_for_completion(&rdrain.done);
2587 }
2588
2589 /**
2590  * ib_drain_sq() - Block until all SQ CQEs have been consumed by the
2591  *                 application.
2592  * @qp:            queue pair to drain
2593  *
2594  * If the device has a provider-specific drain function, then
2595  * call that.  Otherwise call the generic drain function
2596  * __ib_drain_sq().
2597  *
2598  * The caller must:
2599  *
2600  * ensure there is room in the CQ and SQ for the drain work request and
2601  * completion.
2602  *
2603  * allocate the CQ using ib_alloc_cq().
2604  *
2605  * ensure that there are no other contexts that are posting WRs concurrently.
2606  * Otherwise the drain is not guaranteed.
2607  */
2608 void ib_drain_sq(struct ib_qp *qp)
2609 {
2610         if (qp->device->ops.drain_sq)
2611                 qp->device->ops.drain_sq(qp);
2612         else
2613                 __ib_drain_sq(qp);
2614 }
2615 EXPORT_SYMBOL(ib_drain_sq);
2616
2617 /**
2618  * ib_drain_rq() - Block until all RQ CQEs have been consumed by the
2619  *                 application.
2620  * @qp:            queue pair to drain
2621  *
2622  * If the device has a provider-specific drain function, then
2623  * call that.  Otherwise call the generic drain function
2624  * __ib_drain_rq().
2625  *
2626  * The caller must:
2627  *
2628  * ensure there is room in the CQ and RQ for the drain work request and
2629  * completion.
2630  *
2631  * allocate the CQ using ib_alloc_cq().
2632  *
2633  * ensure that there are no other contexts that are posting WRs concurrently.
2634  * Otherwise the drain is not guaranteed.
2635  */
2636 void ib_drain_rq(struct ib_qp *qp)
2637 {
2638         if (qp->device->ops.drain_rq)
2639                 qp->device->ops.drain_rq(qp);
2640         else
2641                 __ib_drain_rq(qp);
2642 }
2643 EXPORT_SYMBOL(ib_drain_rq);
2644
2645 /**
2646  * ib_drain_qp() - Block until all CQEs have been consumed by the
2647  *                 application on both the RQ and SQ.
2648  * @qp:            queue pair to drain
2649  *
2650  * The caller must:
2651  *
2652  * ensure there is room in the CQ(s), SQ, and RQ for drain work requests
2653  * and completions.
2654  *
2655  * allocate the CQs using ib_alloc_cq().
2656  *
2657  * ensure that there are no other contexts that are posting WRs concurrently.
2658  * Otherwise the drain is not guaranteed.
2659  */
2660 void ib_drain_qp(struct ib_qp *qp)
2661 {
2662         ib_drain_sq(qp);
2663         if (!qp->srq)
2664                 ib_drain_rq(qp);
2665 }
2666 EXPORT_SYMBOL(ib_drain_qp);
2667
2668 struct net_device *rdma_alloc_netdev(struct ib_device *device, u8 port_num,
2669                                      enum rdma_netdev_t type, const char *name,
2670                                      unsigned char name_assign_type,
2671                                      void (*setup)(struct net_device *))
2672 {
2673         struct rdma_netdev_alloc_params params;
2674         struct net_device *netdev;
2675         int rc;
2676
2677         if (!device->ops.rdma_netdev_get_params)
2678                 return ERR_PTR(-EOPNOTSUPP);
2679
2680         rc = device->ops.rdma_netdev_get_params(device, port_num, type,
2681                                                 &params);
2682         if (rc)
2683                 return ERR_PTR(rc);
2684
2685         netdev = alloc_netdev_mqs(params.sizeof_priv, name, name_assign_type,
2686                                   setup, params.txqs, params.rxqs);
2687         if (!netdev)
2688                 return ERR_PTR(-ENOMEM);
2689
2690         return netdev;
2691 }
2692 EXPORT_SYMBOL(rdma_alloc_netdev);
2693
2694 int rdma_init_netdev(struct ib_device *device, u8 port_num,
2695                      enum rdma_netdev_t type, const char *name,
2696                      unsigned char name_assign_type,
2697                      void (*setup)(struct net_device *),
2698                      struct net_device *netdev)
2699 {
2700         struct rdma_netdev_alloc_params params;
2701         int rc;
2702
2703         if (!device->ops.rdma_netdev_get_params)
2704                 return -EOPNOTSUPP;
2705
2706         rc = device->ops.rdma_netdev_get_params(device, port_num, type,
2707                                                 &params);
2708         if (rc)
2709                 return rc;
2710
2711         return params.initialize_rdma_netdev(device, port_num,
2712                                              netdev, params.param);
2713 }
2714 EXPORT_SYMBOL(rdma_init_netdev);
2715
2716 void __rdma_block_iter_start(struct ib_block_iter *biter,
2717                              struct scatterlist *sglist, unsigned int nents,
2718                              unsigned long pgsz)
2719 {
2720         memset(biter, 0, sizeof(struct ib_block_iter));
2721         biter->__sg = sglist;
2722         biter->__sg_nents = nents;
2723
2724         /* Driver provides best block size to use */
2725         biter->__pg_bit = __fls(pgsz);
2726 }
2727 EXPORT_SYMBOL(__rdma_block_iter_start);
2728
2729 bool __rdma_block_iter_next(struct ib_block_iter *biter)
2730 {
2731         unsigned int block_offset;
2732
2733         if (!biter->__sg_nents || !biter->__sg)
2734                 return false;
2735
2736         biter->__dma_addr = sg_dma_address(biter->__sg) + biter->__sg_advance;
2737         block_offset = biter->__dma_addr & (BIT_ULL(biter->__pg_bit) - 1);
2738         biter->__sg_advance += BIT_ULL(biter->__pg_bit) - block_offset;
2739
2740         if (biter->__sg_advance >= sg_dma_len(biter->__sg)) {
2741                 biter->__sg_advance = 0;
2742                 biter->__sg = sg_next(biter->__sg);
2743                 biter->__sg_nents--;
2744         }
2745
2746         return true;
2747 }
2748 EXPORT_SYMBOL(__rdma_block_iter_next);