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 #include <rdma/lag.h>
  54
  55 #include "core_priv.h"
  56 #include <trace/events/rdma_core.h>
  57
  58 static int ib_resolve_eth_dmac(struct ib_device *device,
  59                                struct rdma_ah_attr *ah_attr);
  60
  61 static const char * const ib_events[] = {
  62         [IB_EVENT_CQ_ERR]               = "CQ error",
  63         [IB_EVENT_QP_FATAL]             = "QP fatal error",
  64         [IB_EVENT_QP_REQ_ERR]           = "QP request error",
  65         [IB_EVENT_QP_ACCESS_ERR]        = "QP access error",
  66         [IB_EVENT_COMM_EST]             = "communication established",
  67         [IB_EVENT_SQ_DRAINED]           = "send queue drained",
  68         [IB_EVENT_PATH_MIG]             = "path migration successful",
  69         [IB_EVENT_PATH_MIG_ERR]         = "path migration error",
  70         [IB_EVENT_DEVICE_FATAL]         = "device fatal error",
  71         [IB_EVENT_PORT_ACTIVE]          = "port active",
  72         [IB_EVENT_PORT_ERR]             = "port error",
  73         [IB_EVENT_LID_CHANGE]           = "LID change",
  74         [IB_EVENT_PKEY_CHANGE]          = "P_key change",
  75         [IB_EVENT_SM_CHANGE]            = "SM change",
  76         [IB_EVENT_SRQ_ERR]              = "SRQ error",
  77         [IB_EVENT_SRQ_LIMIT_REACHED]    = "SRQ limit reached",
  78         [IB_EVENT_QP_LAST_WQE_REACHED]  = "last WQE reached",
  79         [IB_EVENT_CLIENT_REREGISTER]    = "client reregister",
  80         [IB_EVENT_GID_CHANGE]           = "GID changed",
  81 };
  82
  83 const char *__attribute_const__ ib_event_msg(enum ib_event_type event)
  84 {
  85         size_t index = event;
  86
  87         return (index < ARRAY_SIZE(ib_events) && ib_events[index]) ?
  88                         ib_events[index] : "unrecognized event";
  89 }
  90 EXPORT_SYMBOL(ib_event_msg);
  91
  92 static const char * const wc_statuses[] = {
  93         [IB_WC_SUCCESS]                 = "success",
  94         [IB_WC_LOC_LEN_ERR]             = "local length error",
  95         [IB_WC_LOC_QP_OP_ERR]           = "local QP operation error",
  96         [IB_WC_LOC_EEC_OP_ERR]          = "local EE context operation error",
  97         [IB_WC_LOC_PROT_ERR]            = "local protection error",
  98         [IB_WC_WR_FLUSH_ERR]            = "WR flushed",
  99         [IB_WC_MW_BIND_ERR]             = "memory bind operation error",
 100         [IB_WC_BAD_RESP_ERR]            = "bad response error",
 101         [IB_WC_LOC_ACCESS_ERR]          = "local access error",
 102         [IB_WC_REM_INV_REQ_ERR]         = "remote invalid request error",
 103         [IB_WC_REM_ACCESS_ERR]          = "remote access error",
 104         [IB_WC_REM_OP_ERR]              = "remote operation error",
 105         [IB_WC_RETRY_EXC_ERR]           = "transport retry counter exceeded",
 106         [IB_WC_RNR_RETRY_EXC_ERR]       = "RNR retry counter exceeded",
 107         [IB_WC_LOC_RDD_VIOL_ERR]        = "local RDD violation error",
 108         [IB_WC_REM_INV_RD_REQ_ERR]      = "remote invalid RD request",
 109         [IB_WC_REM_ABORT_ERR]           = "operation aborted",
 110         [IB_WC_INV_EECN_ERR]            = "invalid EE context number",
 111         [IB_WC_INV_EEC_STATE_ERR]       = "invalid EE context state",
 112         [IB_WC_FATAL_ERR]               = "fatal error",
 113         [IB_WC_RESP_TIMEOUT_ERR]        = "response timeout error",
 114         [IB_WC_GENERAL_ERR]             = "general error",
 115 };
 116
 117 const char *__attribute_const__ ib_wc_status_msg(enum ib_wc_status status)
 118 {
 119         size_t index = status;
 120
 121         return (index < ARRAY_SIZE(wc_statuses) && wc_statuses[index]) ?
 122                         wc_statuses[index] : "unrecognized status";
 123 }
 124 EXPORT_SYMBOL(ib_wc_status_msg);
 125
 126 __attribute_const__ int ib_rate_to_mult(enum ib_rate rate)
 127 {
 128         switch (rate) {
 129         case IB_RATE_2_5_GBPS: return   1;
 130         case IB_RATE_5_GBPS:   return   2;
 131         case IB_RATE_10_GBPS:  return   4;
 132         case IB_RATE_20_GBPS:  return   8;
 133         case IB_RATE_30_GBPS:  return  12;
 134         case IB_RATE_40_GBPS:  return  16;
 135         case IB_RATE_60_GBPS:  return  24;
 136         case IB_RATE_80_GBPS:  return  32;
 137         case IB_RATE_120_GBPS: return  48;
 138         case IB_RATE_14_GBPS:  return   6;
 139         case IB_RATE_56_GBPS:  return  22;
 140         case IB_RATE_112_GBPS: return  45;
 141         case IB_RATE_168_GBPS: return  67;
 142         case IB_RATE_25_GBPS:  return  10;
 143         case IB_RATE_100_GBPS: return  40;
 144         case IB_RATE_200_GBPS: return  80;
 145         case IB_RATE_300_GBPS: return 120;
 146         case IB_RATE_28_GBPS:  return  11;
 147         case IB_RATE_50_GBPS:  return  20;
 148         case IB_RATE_400_GBPS: return 160;
 149         case IB_RATE_600_GBPS: return 240;
 150         default:               return  -1;
 151         }
 152 }
 153 EXPORT_SYMBOL(ib_rate_to_mult);
 154
 155 __attribute_const__ enum ib_rate mult_to_ib_rate(int mult)
 156 {
 157         switch (mult) {
 158         case 1:   return IB_RATE_2_5_GBPS;
 159         case 2:   return IB_RATE_5_GBPS;
 160         case 4:   return IB_RATE_10_GBPS;
 161         case 8:   return IB_RATE_20_GBPS;
 162         case 12:  return IB_RATE_30_GBPS;
 163         case 16:  return IB_RATE_40_GBPS;
 164         case 24:  return IB_RATE_60_GBPS;
 165         case 32:  return IB_RATE_80_GBPS;
 166         case 48:  return IB_RATE_120_GBPS;
 167         case 6:   return IB_RATE_14_GBPS;
 168         case 22:  return IB_RATE_56_GBPS;
 169         case 45:  return IB_RATE_112_GBPS;
 170         case 67:  return IB_RATE_168_GBPS;
 171         case 10:  return IB_RATE_25_GBPS;
 172         case 40:  return IB_RATE_100_GBPS;
 173         case 80:  return IB_RATE_200_GBPS;
 174         case 120: return IB_RATE_300_GBPS;
 175         case 11:  return IB_RATE_28_GBPS;
 176         case 20:  return IB_RATE_50_GBPS;
 177         case 160: return IB_RATE_400_GBPS;
 178         case 240: return IB_RATE_600_GBPS;
 179         default:  return IB_RATE_PORT_CURRENT;
 180         }
 181 }
 182 EXPORT_SYMBOL(mult_to_ib_rate);
 183
 184 __attribute_const__ int ib_rate_to_mbps(enum ib_rate rate)
 185 {
 186         switch (rate) {
 187         case IB_RATE_2_5_GBPS: return 2500;
 188         case IB_RATE_5_GBPS:   return 5000;
 189         case IB_RATE_10_GBPS:  return 10000;
 190         case IB_RATE_20_GBPS:  return 20000;
 191         case IB_RATE_30_GBPS:  return 30000;
 192         case IB_RATE_40_GBPS:  return 40000;
 193         case IB_RATE_60_GBPS:  return 60000;
 194         case IB_RATE_80_GBPS:  return 80000;
 195         case IB_RATE_120_GBPS: return 120000;
 196         case IB_RATE_14_GBPS:  return 14062;
 197         case IB_RATE_56_GBPS:  return 56250;
 198         case IB_RATE_112_GBPS: return 112500;
 199         case IB_RATE_168_GBPS: return 168750;
 200         case IB_RATE_25_GBPS:  return 25781;
 201         case IB_RATE_100_GBPS: return 103125;
 202         case IB_RATE_200_GBPS: return 206250;
 203         case IB_RATE_300_GBPS: return 309375;
 204         case IB_RATE_28_GBPS:  return 28125;
 205         case IB_RATE_50_GBPS:  return 53125;
 206         case IB_RATE_400_GBPS: return 425000;
 207         case IB_RATE_600_GBPS: return 637500;
 208         default:               return -1;
 209         }
 210 }
 211 EXPORT_SYMBOL(ib_rate_to_mbps);
 212
 213 __attribute_const__ enum rdma_transport_type
 214 rdma_node_get_transport(unsigned int node_type)
 215 {
 216
 217         if (node_type == RDMA_NODE_USNIC)
 218                 return RDMA_TRANSPORT_USNIC;
 219         if (node_type == RDMA_NODE_USNIC_UDP)
 220                 return RDMA_TRANSPORT_USNIC_UDP;
 221         if (node_type == RDMA_NODE_RNIC)
 222                 return RDMA_TRANSPORT_IWARP;
 223         if (node_type == RDMA_NODE_UNSPECIFIED)
 224                 return RDMA_TRANSPORT_UNSPECIFIED;
 225
 226         return RDMA_TRANSPORT_IB;
 227 }
 228 EXPORT_SYMBOL(rdma_node_get_transport);
 229
 230 enum rdma_link_layer rdma_port_get_link_layer(struct ib_device *device,
 231                                               u32 port_num)
 232 {
 233         enum rdma_transport_type lt;
 234         if (device->ops.get_link_layer)
 235                 return device->ops.get_link_layer(device, port_num);
 236
 237         lt = rdma_node_get_transport(device->node_type);
 238         if (lt == RDMA_TRANSPORT_IB)
 239                 return IB_LINK_LAYER_INFINIBAND;
 240
 241         return IB_LINK_LAYER_ETHERNET;
 242 }
 243 EXPORT_SYMBOL(rdma_port_get_link_layer);
 244
 245 /* Protection domains */
 246
 247 /**
 248  * __ib_alloc_pd - Allocates an unused protection domain.
 249  * @device: The device on which to allocate the protection domain.
 250  * @flags: protection domain flags
 251  * @caller: caller's build-time module name
 252  *
 253  * A protection domain object provides an association between QPs, shared
 254  * receive queues, address handles, memory regions, and memory windows.
 255  *
 256  * Every PD has a local_dma_lkey which can be used as the lkey value for local
 257  * memory operations.
 258  */
 259 struct ib_pd *__ib_alloc_pd(struct ib_device *device, unsigned int flags,
 260                 const char *caller)
 261 {
 262         struct ib_pd *pd;
 263         int mr_access_flags = 0;
 264         int ret;
 265
 266         pd = rdma_zalloc_drv_obj(device, ib_pd);
 267         if (!pd)
 268                 return ERR_PTR(-ENOMEM);
 269
 270         pd->device = device;
 271         pd->flags = flags;
 272
 273         rdma_restrack_new(&pd->res, RDMA_RESTRACK_PD);
 274         rdma_restrack_set_name(&pd->res, caller);
 275
 276         ret = device->ops.alloc_pd(pd, NULL);
 277         if (ret) {
 278                 rdma_restrack_put(&pd->res);
 279                 kfree(pd);
 280                 return ERR_PTR(ret);
 281         }
 282         rdma_restrack_add(&pd->res);
 283
 284         if (device->attrs.kernel_cap_flags & IBK_LOCAL_DMA_LKEY)
 285                 pd->local_dma_lkey = device->local_dma_lkey;
 286         else
 287                 mr_access_flags |= IB_ACCESS_LOCAL_WRITE;
 288
 289         if (flags & IB_PD_UNSAFE_GLOBAL_RKEY) {
 290                 pr_warn("%s: enabling unsafe global rkey\n", caller);
 291                 mr_access_flags |= IB_ACCESS_REMOTE_READ | IB_ACCESS_REMOTE_WRITE;
 292         }
 293
 294         if (mr_access_flags) {
 295                 struct ib_mr *mr;
 296
 297                 mr = pd->device->ops.get_dma_mr(pd, mr_access_flags);
 298                 if (IS_ERR(mr)) {
 299                         ib_dealloc_pd(pd);
 300                         return ERR_CAST(mr);
 301                 }
 302
 303                 mr->device      = pd->device;
 304                 mr->pd          = pd;
 305                 mr->type        = IB_MR_TYPE_DMA;
 306                 mr->uobject     = NULL;
 307                 mr->need_inval  = false;
 308
 309                 pd->__internal_mr = mr;
 310
 311                 if (!(device->attrs.kernel_cap_flags & IBK_LOCAL_DMA_LKEY))
 312                         pd->local_dma_lkey = pd->__internal_mr->lkey;
 313
 314                 if (flags & IB_PD_UNSAFE_GLOBAL_RKEY)
 315                         pd->unsafe_global_rkey = pd->__internal_mr->rkey;
 316         }
 317
 318         return pd;
 319 }
 320 EXPORT_SYMBOL(__ib_alloc_pd);
 321
 322 /**
 323  * ib_dealloc_pd_user - Deallocates a protection domain.
 324  * @pd: The protection domain to deallocate.
 325  * @udata: Valid user data or NULL for kernel object
 326  *
 327  * It is an error to call this function while any resources in the pd still
 328  * exist.  The caller is responsible to synchronously destroy them and
 329  * guarantee no new allocations will happen.
 330  */
 331 int ib_dealloc_pd_user(struct ib_pd *pd, struct ib_udata *udata)
 332 {
 333         int ret;
 334
 335         if (pd->__internal_mr) {
 336                 ret = pd->device->ops.dereg_mr(pd->__internal_mr, NULL);
 337                 WARN_ON(ret);
 338                 pd->__internal_mr = NULL;
 339         }
 340
 341         ret = pd->device->ops.dealloc_pd(pd, udata);
 342         if (ret)
 343                 return ret;
 344
 345         rdma_restrack_del(&pd->res);
 346         kfree(pd);
 347         return ret;
 348 }
 349 EXPORT_SYMBOL(ib_dealloc_pd_user);
 350
 351 /* Address handles */
 352
 353 /**
 354  * rdma_copy_ah_attr - Copy rdma ah attribute from source to destination.
 355  * @dest:       Pointer to destination ah_attr. Contents of the destination
 356  *              pointer is assumed to be invalid and attribute are overwritten.
 357  * @src:        Pointer to source ah_attr.
 358  */
 359 void rdma_copy_ah_attr(struct rdma_ah_attr *dest,
 360                        const struct rdma_ah_attr *src)
 361 {
 362         *dest = *src;
 363         if (dest->grh.sgid_attr)
 364                 rdma_hold_gid_attr(dest->grh.sgid_attr);
 365 }
 366 EXPORT_SYMBOL(rdma_copy_ah_attr);
 367
 368 /**
 369  * rdma_replace_ah_attr - Replace valid ah_attr with new new one.
 370  * @old:        Pointer to existing ah_attr which needs to be replaced.
 371  *              old is assumed to be valid or zero'd
 372  * @new:        Pointer to the new ah_attr.
 373  *
 374  * rdma_replace_ah_attr() first releases any reference in the old ah_attr if
 375  * old the ah_attr is valid; after that it copies the new attribute and holds
 376  * the reference to the replaced ah_attr.
 377  */
 378 void rdma_replace_ah_attr(struct rdma_ah_attr *old,
 379                           const struct rdma_ah_attr *new)
 380 {
 381         rdma_destroy_ah_attr(old);
 382         *old = *new;
 383         if (old->grh.sgid_attr)
 384                 rdma_hold_gid_attr(old->grh.sgid_attr);
 385 }
 386 EXPORT_SYMBOL(rdma_replace_ah_attr);
 387
 388 /**
 389  * rdma_move_ah_attr - Move ah_attr pointed by source to destination.
 390  * @dest:       Pointer to destination ah_attr to copy to.
 391  *              dest is assumed to be valid or zero'd
 392  * @src:        Pointer to the new ah_attr.
 393  *
 394  * rdma_move_ah_attr() first releases any reference in the destination ah_attr
 395  * if it is valid. This also transfers ownership of internal references from
 396  * src to dest, making src invalid in the process. No new reference of the src
 397  * ah_attr is taken.
 398  */
 399 void rdma_move_ah_attr(struct rdma_ah_attr *dest, struct rdma_ah_attr *src)
 400 {
 401         rdma_destroy_ah_attr(dest);
 402         *dest = *src;
 403         src->grh.sgid_attr = NULL;
 404 }
 405 EXPORT_SYMBOL(rdma_move_ah_attr);
 406
 407 /*
 408  * Validate that the rdma_ah_attr is valid for the device before passing it
 409  * off to the driver.
 410  */
 411 static int rdma_check_ah_attr(struct ib_device *device,
 412                               struct rdma_ah_attr *ah_attr)
 413 {
 414         if (!rdma_is_port_valid(device, ah_attr->port_num))
 415                 return -EINVAL;
 416
 417         if ((rdma_is_grh_required(device, ah_attr->port_num) ||
 418              ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE) &&
 419             !(ah_attr->ah_flags & IB_AH_GRH))
 420                 return -EINVAL;
 421
 422         if (ah_attr->grh.sgid_attr) {
 423                 /*
 424                  * Make sure the passed sgid_attr is consistent with the
 425                  * parameters
 426                  */
 427                 if (ah_attr->grh.sgid_attr->index != ah_attr->grh.sgid_index ||
 428                     ah_attr->grh.sgid_attr->port_num != ah_attr->port_num)
 429                         return -EINVAL;
 430         }
 431         return 0;
 432 }
 433
 434 /*
 435  * If the ah requires a GRH then ensure that sgid_attr pointer is filled in.
 436  * On success the caller is responsible to call rdma_unfill_sgid_attr().
 437  */
 438 static int rdma_fill_sgid_attr(struct ib_device *device,
 439                                struct rdma_ah_attr *ah_attr,
 440                                const struct ib_gid_attr **old_sgid_attr)
 441 {
 442         const struct ib_gid_attr *sgid_attr;
 443         struct ib_global_route *grh;
 444         int ret;
 445
 446         *old_sgid_attr = ah_attr->grh.sgid_attr;
 447
 448         ret = rdma_check_ah_attr(device, ah_attr);
 449         if (ret)
 450                 return ret;
 451
 452         if (!(ah_attr->ah_flags & IB_AH_GRH))
 453                 return 0;
 454
 455         grh = rdma_ah_retrieve_grh(ah_attr);
 456         if (grh->sgid_attr)
 457                 return 0;
 458
 459         sgid_attr =
 460                 rdma_get_gid_attr(device, ah_attr->port_num, grh->sgid_index);
 461         if (IS_ERR(sgid_attr))
 462                 return PTR_ERR(sgid_attr);
 463
 464         /* Move ownerhip of the kref into the ah_attr */
 465         grh->sgid_attr = sgid_attr;
 466         return 0;
 467 }
 468
 469 static void rdma_unfill_sgid_attr(struct rdma_ah_attr *ah_attr,
 470                                   const struct ib_gid_attr *old_sgid_attr)
 471 {
 472         /*
 473          * Fill didn't change anything, the caller retains ownership of
 474          * whatever it passed
 475          */
 476         if (ah_attr->grh.sgid_attr == old_sgid_attr)
 477                 return;
 478
 479         /*
 480          * Otherwise, we need to undo what rdma_fill_sgid_attr so the caller
 481          * doesn't see any change in the rdma_ah_attr. If we get here
 482          * old_sgid_attr is NULL.
 483          */
 484         rdma_destroy_ah_attr(ah_attr);
 485 }
 486
 487 static const struct ib_gid_attr *
 488 rdma_update_sgid_attr(struct rdma_ah_attr *ah_attr,
 489                       const struct ib_gid_attr *old_attr)
 490 {
 491         if (old_attr)
 492                 rdma_put_gid_attr(old_attr);
 493         if (ah_attr->ah_flags & IB_AH_GRH) {
 494                 rdma_hold_gid_attr(ah_attr->grh.sgid_attr);
 495                 return ah_attr->grh.sgid_attr;
 496         }
 497         return NULL;
 498 }
 499
 500 static struct ib_ah *_rdma_create_ah(struct ib_pd *pd,
 501                                      struct rdma_ah_attr *ah_attr,
 502                                      u32 flags,
 503                                      struct ib_udata *udata,
 504                                      struct net_device *xmit_slave)
 505 {
 506         struct rdma_ah_init_attr init_attr = {};
 507         struct ib_device *device = pd->device;
 508         struct ib_ah *ah;
 509         int ret;
 510
 511         might_sleep_if(flags & RDMA_CREATE_AH_SLEEPABLE);
 512
 513         if (!udata && !device->ops.create_ah)
 514                 return ERR_PTR(-EOPNOTSUPP);
 515
 516         ah = rdma_zalloc_drv_obj_gfp(
 517                 device, ib_ah,
 518                 (flags & RDMA_CREATE_AH_SLEEPABLE) ? GFP_KERNEL : GFP_ATOMIC);
 519         if (!ah)
 520                 return ERR_PTR(-ENOMEM);
 521
 522         ah->device = device;
 523         ah->pd = pd;
 524         ah->type = ah_attr->type;
 525         ah->sgid_attr = rdma_update_sgid_attr(ah_attr, NULL);
 526         init_attr.ah_attr = ah_attr;
 527         init_attr.flags = flags;
 528         init_attr.xmit_slave = xmit_slave;
 529
 530         if (udata)
 531                 ret = device->ops.create_user_ah(ah, &init_attr, udata);
 532         else
 533                 ret = device->ops.create_ah(ah, &init_attr, NULL);
 534         if (ret) {
 535                 if (ah->sgid_attr)
 536                         rdma_put_gid_attr(ah->sgid_attr);
 537                 kfree(ah);
 538                 return ERR_PTR(ret);
 539         }
 540
 541         atomic_inc(&pd->usecnt);
 542         return ah;
 543 }
 544
 545 /**
 546  * rdma_create_ah - Creates an address handle for the
 547  * given address vector.
 548  * @pd: The protection domain associated with the address handle.
 549  * @ah_attr: The attributes of the address vector.
 550  * @flags: Create address handle flags (see enum rdma_create_ah_flags).
 551  *
 552  * It returns 0 on success and returns appropriate error code on error.
 553  * The address handle is used to reference a local or global destination
 554  * in all UD QP post sends.
 555  */
 556 struct ib_ah *rdma_create_ah(struct ib_pd *pd, struct rdma_ah_attr *ah_attr,
 557                              u32 flags)
 558 {
 559         const struct ib_gid_attr *old_sgid_attr;
 560         struct net_device *slave;
 561         struct ib_ah *ah;
 562         int ret;
 563
 564         ret = rdma_fill_sgid_attr(pd->device, ah_attr, &old_sgid_attr);
 565         if (ret)
 566                 return ERR_PTR(ret);
 567         slave = rdma_lag_get_ah_roce_slave(pd->device, ah_attr,
 568                                            (flags & RDMA_CREATE_AH_SLEEPABLE) ?
 569                                            GFP_KERNEL : GFP_ATOMIC);
 570         if (IS_ERR(slave)) {
 571                 rdma_unfill_sgid_attr(ah_attr, old_sgid_attr);
 572                 return (void *)slave;
 573         }
 574         ah = _rdma_create_ah(pd, ah_attr, flags, NULL, slave);
 575         rdma_lag_put_ah_roce_slave(slave);
 576         rdma_unfill_sgid_attr(ah_attr, old_sgid_attr);
 577         return ah;
 578 }
 579 EXPORT_SYMBOL(rdma_create_ah);
 580
 581 /**
 582  * rdma_create_user_ah - Creates an address handle for the
 583  * given address vector.
 584  * It resolves destination mac address for ah attribute of RoCE type.
 585  * @pd: The protection domain associated with the address handle.
 586  * @ah_attr: The attributes of the address vector.
 587  * @udata: pointer to user's input output buffer information need by
 588  *         provider driver.
 589  *
 590  * It returns 0 on success and returns appropriate error code on error.
 591  * The address handle is used to reference a local or global destination
 592  * in all UD QP post sends.
 593  */
 594 struct ib_ah *rdma_create_user_ah(struct ib_pd *pd,
 595                                   struct rdma_ah_attr *ah_attr,
 596                                   struct ib_udata *udata)
 597 {
 598         const struct ib_gid_attr *old_sgid_attr;
 599         struct ib_ah *ah;
 600         int err;
 601
 602         err = rdma_fill_sgid_attr(pd->device, ah_attr, &old_sgid_attr);
 603         if (err)
 604                 return ERR_PTR(err);
 605
 606         if (ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE) {
 607                 err = ib_resolve_eth_dmac(pd->device, ah_attr);
 608                 if (err) {
 609                         ah = ERR_PTR(err);
 610                         goto out;
 611                 }
 612         }
 613
 614         ah = _rdma_create_ah(pd, ah_attr, RDMA_CREATE_AH_SLEEPABLE,
 615                              udata, NULL);
 616
 617 out:
 618         rdma_unfill_sgid_attr(ah_attr, old_sgid_attr);
 619         return ah;
 620 }
 621 EXPORT_SYMBOL(rdma_create_user_ah);
 622
 623 int ib_get_rdma_header_version(const union rdma_network_hdr *hdr)
 624 {
 625         const struct iphdr *ip4h = (struct iphdr *)&hdr->roce4grh;
 626         struct iphdr ip4h_checked;
 627         const struct ipv6hdr *ip6h = (struct ipv6hdr *)&hdr->ibgrh;
 628
 629         /* If it's IPv6, the version must be 6, otherwise, the first
 630          * 20 bytes (before the IPv4 header) are garbled.
 631          */
 632         if (ip6h->version != 6)
 633                 return (ip4h->version == 4) ? 4 : 0;
 634         /* version may be 6 or 4 because the first 20 bytes could be garbled */
 635
 636         /* RoCE v2 requires no options, thus header length
 637          * must be 5 words
 638          */
 639         if (ip4h->ihl != 5)
 640                 return 6;
 641
 642         /* Verify checksum.
 643          * We can't write on scattered buffers so we need to copy to
 644          * temp buffer.
 645          */
 646         memcpy(&ip4h_checked, ip4h, sizeof(ip4h_checked));
 647         ip4h_checked.check = 0;
 648         ip4h_checked.check = ip_fast_csum((u8 *)&ip4h_checked, 5);
 649         /* if IPv4 header checksum is OK, believe it */
 650         if (ip4h->check == ip4h_checked.check)
 651                 return 4;
 652         return 6;
 653 }
 654 EXPORT_SYMBOL(ib_get_rdma_header_version);
 655
 656 static enum rdma_network_type ib_get_net_type_by_grh(struct ib_device *device,
 657                                                      u32 port_num,
 658                                                      const struct ib_grh *grh)
 659 {
 660         int grh_version;
 661
 662         if (rdma_protocol_ib(device, port_num))
 663                 return RDMA_NETWORK_IB;
 664
 665         grh_version = ib_get_rdma_header_version((union rdma_network_hdr *)grh);
 666
 667         if (grh_version == 4)
 668                 return RDMA_NETWORK_IPV4;
 669
 670         if (grh->next_hdr == IPPROTO_UDP)
 671                 return RDMA_NETWORK_IPV6;
 672
 673         return RDMA_NETWORK_ROCE_V1;
 674 }
 675
 676 struct find_gid_index_context {
 677         u16 vlan_id;
 678         enum ib_gid_type gid_type;
 679 };
 680
 681 static bool find_gid_index(const union ib_gid *gid,
 682                            const struct ib_gid_attr *gid_attr,
 683                            void *context)
 684 {
 685         struct find_gid_index_context *ctx = context;
 686         u16 vlan_id = 0xffff;
 687         int ret;
 688
 689         if (ctx->gid_type != gid_attr->gid_type)
 690                 return false;
 691
 692         ret = rdma_read_gid_l2_fields(gid_attr, &vlan_id, NULL);
 693         if (ret)
 694                 return false;
 695
 696         return ctx->vlan_id == vlan_id;
 697 }
 698
 699 static const struct ib_gid_attr *
 700 get_sgid_attr_from_eth(struct ib_device *device, u32 port_num,
 701                        u16 vlan_id, const union ib_gid *sgid,
 702                        enum ib_gid_type gid_type)
 703 {
 704         struct find_gid_index_context context = {.vlan_id = vlan_id,
 705                                                  .gid_type = gid_type};
 706
 707         return rdma_find_gid_by_filter(device, sgid, port_num, find_gid_index,
 708                                        &context);
 709 }
 710
 711 int ib_get_gids_from_rdma_hdr(const union rdma_network_hdr *hdr,
 712                               enum rdma_network_type net_type,
 713                               union ib_gid *sgid, union ib_gid *dgid)
 714 {
 715         struct sockaddr_in  src_in;
 716         struct sockaddr_in  dst_in;
 717         __be32 src_saddr, dst_saddr;
 718
 719         if (!sgid || !dgid)
 720                 return -EINVAL;
 721
 722         if (net_type == RDMA_NETWORK_IPV4) {
 723                 memcpy(&src_in.sin_addr.s_addr,
 724                        &hdr->roce4grh.saddr, 4);
 725                 memcpy(&dst_in.sin_addr.s_addr,
 726                        &hdr->roce4grh.daddr, 4);
 727                 src_saddr = src_in.sin_addr.s_addr;
 728                 dst_saddr = dst_in.sin_addr.s_addr;
 729                 ipv6_addr_set_v4mapped(src_saddr,
 730                                        (struct in6_addr *)sgid);
 731                 ipv6_addr_set_v4mapped(dst_saddr,
 732                                        (struct in6_addr *)dgid);
 733                 return 0;
 734         } else if (net_type == RDMA_NETWORK_IPV6 ||
 735                    net_type == RDMA_NETWORK_IB || RDMA_NETWORK_ROCE_V1) {
 736                 *dgid = hdr->ibgrh.dgid;
 737                 *sgid = hdr->ibgrh.sgid;
 738                 return 0;
 739         } else {
 740                 return -EINVAL;
 741         }
 742 }
 743 EXPORT_SYMBOL(ib_get_gids_from_rdma_hdr);
 744
 745 /* Resolve destination mac address and hop limit for unicast destination
 746  * GID entry, considering the source GID entry as well.
 747  * ah_attribute must have have valid port_num, sgid_index.
 748  */
 749 static int ib_resolve_unicast_gid_dmac(struct ib_device *device,
 750                                        struct rdma_ah_attr *ah_attr)
 751 {
 752         struct ib_global_route *grh = rdma_ah_retrieve_grh(ah_attr);
 753         const struct ib_gid_attr *sgid_attr = grh->sgid_attr;
 754         int hop_limit = 0xff;
 755         int ret = 0;
 756
 757         /* If destination is link local and source GID is RoCEv1,
 758          * IP stack is not used.
 759          */
 760         if (rdma_link_local_addr((struct in6_addr *)grh->dgid.raw) &&
 761             sgid_attr->gid_type == IB_GID_TYPE_ROCE) {
 762                 rdma_get_ll_mac((struct in6_addr *)grh->dgid.raw,
 763                                 ah_attr->roce.dmac);
 764                 return ret;
 765         }
 766
 767         ret = rdma_addr_find_l2_eth_by_grh(&sgid_attr->gid, &grh->dgid,
 768                                            ah_attr->roce.dmac,
 769                                            sgid_attr, &hop_limit);
 770
 771         grh->hop_limit = hop_limit;
 772         return ret;
 773 }
 774
 775 /*
 776  * This function initializes address handle attributes from the incoming packet.
 777  * Incoming packet has dgid of the receiver node on which this code is
 778  * getting executed and, sgid contains the GID of the sender.
 779  *
 780  * When resolving mac address of destination, the arrived dgid is used
 781  * as sgid and, sgid is used as dgid because sgid contains destinations
 782  * GID whom to respond to.
 783  *
 784  * On success the caller is responsible to call rdma_destroy_ah_attr on the
 785  * attr.
 786  */
 787 int ib_init_ah_attr_from_wc(struct ib_device *device, u32 port_num,
 788                             const struct ib_wc *wc, const struct ib_grh *grh,
 789                             struct rdma_ah_attr *ah_attr)
 790 {
 791         u32 flow_class;
 792         int ret;
 793         enum rdma_network_type net_type = RDMA_NETWORK_IB;
 794         enum ib_gid_type gid_type = IB_GID_TYPE_IB;
 795         const struct ib_gid_attr *sgid_attr;
 796         int hoplimit = 0xff;
 797         union ib_gid dgid;
 798         union ib_gid sgid;
 799
 800         might_sleep();
 801
 802         memset(ah_attr, 0, sizeof *ah_attr);
 803         ah_attr->type = rdma_ah_find_type(device, port_num);
 804         if (rdma_cap_eth_ah(device, port_num)) {
 805                 if (wc->wc_flags & IB_WC_WITH_NETWORK_HDR_TYPE)
 806                         net_type = wc->network_hdr_type;
 807                 else
 808                         net_type = ib_get_net_type_by_grh(device, port_num, grh);
 809                 gid_type = ib_network_to_gid_type(net_type);
 810         }
 811         ret = ib_get_gids_from_rdma_hdr((union rdma_network_hdr *)grh, net_type,
 812                                         &sgid, &dgid);
 813         if (ret)
 814                 return ret;
 815
 816         rdma_ah_set_sl(ah_attr, wc->sl);
 817         rdma_ah_set_port_num(ah_attr, port_num);
 818
 819         if (rdma_protocol_roce(device, port_num)) {
 820                 u16 vlan_id = wc->wc_flags & IB_WC_WITH_VLAN ?
 821                                 wc->vlan_id : 0xffff;
 822
 823                 if (!(wc->wc_flags & IB_WC_GRH))
 824                         return -EPROTOTYPE;
 825
 826                 sgid_attr = get_sgid_attr_from_eth(device, port_num,
 827                                                    vlan_id, &dgid,
 828                                                    gid_type);
 829                 if (IS_ERR(sgid_attr))
 830                         return PTR_ERR(sgid_attr);
 831
 832                 flow_class = be32_to_cpu(grh->version_tclass_flow);
 833                 rdma_move_grh_sgid_attr(ah_attr,
 834                                         &sgid,
 835                                         flow_class & 0xFFFFF,
 836                                         hoplimit,
 837                                         (flow_class >> 20) & 0xFF,
 838                                         sgid_attr);
 839
 840                 ret = ib_resolve_unicast_gid_dmac(device, ah_attr);
 841                 if (ret)
 842                         rdma_destroy_ah_attr(ah_attr);
 843
 844                 return ret;
 845         } else {
 846                 rdma_ah_set_dlid(ah_attr, wc->slid);
 847                 rdma_ah_set_path_bits(ah_attr, wc->dlid_path_bits);
 848
 849                 if ((wc->wc_flags & IB_WC_GRH) == 0)
 850                         return 0;
 851
 852                 if (dgid.global.interface_id !=
 853                                         cpu_to_be64(IB_SA_WELL_KNOWN_GUID)) {
 854                         sgid_attr = rdma_find_gid_by_port(
 855                                 device, &dgid, IB_GID_TYPE_IB, port_num, NULL);
 856                 } else
 857                         sgid_attr = rdma_get_gid_attr(device, port_num, 0);
 858
 859                 if (IS_ERR(sgid_attr))
 860                         return PTR_ERR(sgid_attr);
 861                 flow_class = be32_to_cpu(grh->version_tclass_flow);
 862                 rdma_move_grh_sgid_attr(ah_attr,
 863                                         &sgid,
 864                                         flow_class & 0xFFFFF,
 865                                         hoplimit,
 866                                         (flow_class >> 20) & 0xFF,
 867                                         sgid_attr);
 868
 869                 return 0;
 870         }
 871 }
 872 EXPORT_SYMBOL(ib_init_ah_attr_from_wc);
 873
 874 /**
 875  * rdma_move_grh_sgid_attr - Sets the sgid attribute of GRH, taking ownership
 876  * of the reference
 877  *
 878  * @attr:       Pointer to AH attribute structure
 879  * @dgid:       Destination GID
 880  * @flow_label: Flow label
 881  * @hop_limit:  Hop limit
 882  * @traffic_class: traffic class
 883  * @sgid_attr:  Pointer to SGID attribute
 884  *
 885  * This takes ownership of the sgid_attr reference. The caller must ensure
 886  * rdma_destroy_ah_attr() is called before destroying the rdma_ah_attr after
 887  * calling this function.
 888  */
 889 void rdma_move_grh_sgid_attr(struct rdma_ah_attr *attr, union ib_gid *dgid,
 890                              u32 flow_label, u8 hop_limit, u8 traffic_class,
 891                              const struct ib_gid_attr *sgid_attr)
 892 {
 893         rdma_ah_set_grh(attr, dgid, flow_label, sgid_attr->index, hop_limit,
 894                         traffic_class);
 895         attr->grh.sgid_attr = sgid_attr;
 896 }
 897 EXPORT_SYMBOL(rdma_move_grh_sgid_attr);
 898
 899 /**
 900  * rdma_destroy_ah_attr - Release reference to SGID attribute of
 901  * ah attribute.
 902  * @ah_attr: Pointer to ah attribute
 903  *
 904  * Release reference to the SGID attribute of the ah attribute if it is
 905  * non NULL. It is safe to call this multiple times, and safe to call it on
 906  * a zero initialized ah_attr.
 907  */
 908 void rdma_destroy_ah_attr(struct rdma_ah_attr *ah_attr)
 909 {
 910         if (ah_attr->grh.sgid_attr) {
 911                 rdma_put_gid_attr(ah_attr->grh.sgid_attr);
 912                 ah_attr->grh.sgid_attr = NULL;
 913         }
 914 }
 915 EXPORT_SYMBOL(rdma_destroy_ah_attr);
 916
 917 struct ib_ah *ib_create_ah_from_wc(struct ib_pd *pd, const struct ib_wc *wc,
 918                                    const struct ib_grh *grh, u32 port_num)
 919 {
 920         struct rdma_ah_attr ah_attr;
 921         struct ib_ah *ah;
 922         int ret;
 923
 924         ret = ib_init_ah_attr_from_wc(pd->device, port_num, wc, grh, &ah_attr);
 925         if (ret)
 926                 return ERR_PTR(ret);
 927
 928         ah = rdma_create_ah(pd, &ah_attr, RDMA_CREATE_AH_SLEEPABLE);
 929
 930         rdma_destroy_ah_attr(&ah_attr);
 931         return ah;
 932 }
 933 EXPORT_SYMBOL(ib_create_ah_from_wc);
 934
 935 int rdma_modify_ah(struct ib_ah *ah, struct rdma_ah_attr *ah_attr)
 936 {
 937         const struct ib_gid_attr *old_sgid_attr;
 938         int ret;
 939
 940         if (ah->type != ah_attr->type)
 941                 return -EINVAL;
 942
 943         ret = rdma_fill_sgid_attr(ah->device, ah_attr, &old_sgid_attr);
 944         if (ret)
 945                 return ret;
 946
 947         ret = ah->device->ops.modify_ah ?
 948                 ah->device->ops.modify_ah(ah, ah_attr) :
 949                 -EOPNOTSUPP;
 950
 951         ah->sgid_attr = rdma_update_sgid_attr(ah_attr, ah->sgid_attr);
 952         rdma_unfill_sgid_attr(ah_attr, old_sgid_attr);
 953         return ret;
 954 }
 955 EXPORT_SYMBOL(rdma_modify_ah);
 956
 957 int rdma_query_ah(struct ib_ah *ah, struct rdma_ah_attr *ah_attr)
 958 {
 959         ah_attr->grh.sgid_attr = NULL;
 960
 961         return ah->device->ops.query_ah ?
 962                 ah->device->ops.query_ah(ah, ah_attr) :
 963                 -EOPNOTSUPP;
 964 }
 965 EXPORT_SYMBOL(rdma_query_ah);
 966
 967 int rdma_destroy_ah_user(struct ib_ah *ah, u32 flags, struct ib_udata *udata)
 968 {
 969         const struct ib_gid_attr *sgid_attr = ah->sgid_attr;
 970         struct ib_pd *pd;
 971         int ret;
 972
 973         might_sleep_if(flags & RDMA_DESTROY_AH_SLEEPABLE);
 974
 975         pd = ah->pd;
 976
 977         ret = ah->device->ops.destroy_ah(ah, flags);
 978         if (ret)
 979                 return ret;
 980
 981         atomic_dec(&pd->usecnt);
 982         if (sgid_attr)
 983                 rdma_put_gid_attr(sgid_attr);
 984
 985         kfree(ah);
 986         return ret;
 987 }
 988 EXPORT_SYMBOL(rdma_destroy_ah_user);
 989
 990 /* Shared receive queues */
 991
 992 /**
 993  * ib_create_srq_user - Creates a SRQ associated with the specified protection
 994  *   domain.
 995  * @pd: The protection domain associated with the SRQ.
 996  * @srq_init_attr: A list of initial attributes required to create the
 997  *   SRQ.  If SRQ creation succeeds, then the attributes are updated to
 998  *   the actual capabilities of the created SRQ.
 999  * @uobject: uobject pointer if this is not a kernel SRQ
1000  * @udata: udata pointer if this is not a kernel SRQ
1001  *
1002  * srq_attr->max_wr and srq_attr->max_sge are read the determine the
1003  * requested size of the SRQ, and set to the actual values allocated
1004  * on return.  If ib_create_srq() succeeds, then max_wr and max_sge
1005  * will always be at least as large as the requested values.
1006  */
1007 struct ib_srq *ib_create_srq_user(struct ib_pd *pd,
1008                                   struct ib_srq_init_attr *srq_init_attr,
1009                                   struct ib_usrq_object *uobject,
1010                                   struct ib_udata *udata)
1011 {
1012         struct ib_srq *srq;
1013         int ret;
1014
1015         srq = rdma_zalloc_drv_obj(pd->device, ib_srq);
1016         if (!srq)
1017                 return ERR_PTR(-ENOMEM);
1018
1019         srq->device = pd->device;
1020         srq->pd = pd;
1021         srq->event_handler = srq_init_attr->event_handler;
1022         srq->srq_context = srq_init_attr->srq_context;
1023         srq->srq_type = srq_init_attr->srq_type;
1024         srq->uobject = uobject;
1025
1026         if (ib_srq_has_cq(srq->srq_type)) {
1027                 srq->ext.cq = srq_init_attr->ext.cq;
1028                 atomic_inc(&srq->ext.cq->usecnt);
1029         }
1030         if (srq->srq_type == IB_SRQT_XRC) {
1031                 srq->ext.xrc.xrcd = srq_init_attr->ext.xrc.xrcd;
1032                 if (srq->ext.xrc.xrcd)
1033                         atomic_inc(&srq->ext.xrc.xrcd->usecnt);
1034         }
1035         atomic_inc(&pd->usecnt);
1036
1037         rdma_restrack_new(&srq->res, RDMA_RESTRACK_SRQ);
1038         rdma_restrack_parent_name(&srq->res, &pd->res);
1039
1040         ret = pd->device->ops.create_srq(srq, srq_init_attr, udata);
1041         if (ret) {
1042                 rdma_restrack_put(&srq->res);
1043                 atomic_dec(&pd->usecnt);
1044                 if (srq->srq_type == IB_SRQT_XRC && srq->ext.xrc.xrcd)
1045                         atomic_dec(&srq->ext.xrc.xrcd->usecnt);
1046                 if (ib_srq_has_cq(srq->srq_type))
1047                         atomic_dec(&srq->ext.cq->usecnt);
1048                 kfree(srq);
1049                 return ERR_PTR(ret);
1050         }
1051
1052         rdma_restrack_add(&srq->res);
1053
1054         return srq;
1055 }
1056 EXPORT_SYMBOL(ib_create_srq_user);
1057
1058 int ib_modify_srq(struct ib_srq *srq,
1059                   struct ib_srq_attr *srq_attr,
1060                   enum ib_srq_attr_mask srq_attr_mask)
1061 {
1062         return srq->device->ops.modify_srq ?
1063                 srq->device->ops.modify_srq(srq, srq_attr, srq_attr_mask,
1064                                             NULL) : -EOPNOTSUPP;
1065 }
1066 EXPORT_SYMBOL(ib_modify_srq);
1067
1068 int ib_query_srq(struct ib_srq *srq,
1069                  struct ib_srq_attr *srq_attr)
1070 {
1071         return srq->device->ops.query_srq ?
1072                 srq->device->ops.query_srq(srq, srq_attr) : -EOPNOTSUPP;
1073 }
1074 EXPORT_SYMBOL(ib_query_srq);
1075
1076 int ib_destroy_srq_user(struct ib_srq *srq, struct ib_udata *udata)
1077 {
1078         int ret;
1079
1080         if (atomic_read(&srq->usecnt))
1081                 return -EBUSY;
1082
1083         ret = srq->device->ops.destroy_srq(srq, udata);
1084         if (ret)
1085                 return ret;
1086
1087         atomic_dec(&srq->pd->usecnt);
1088         if (srq->srq_type == IB_SRQT_XRC && srq->ext.xrc.xrcd)
1089                 atomic_dec(&srq->ext.xrc.xrcd->usecnt);
1090         if (ib_srq_has_cq(srq->srq_type))
1091                 atomic_dec(&srq->ext.cq->usecnt);
1092         rdma_restrack_del(&srq->res);
1093         kfree(srq);
1094
1095         return ret;
1096 }
1097 EXPORT_SYMBOL(ib_destroy_srq_user);
1098
1099 /* Queue pairs */
1100
1101 static void __ib_shared_qp_event_handler(struct ib_event *event, void *context)
1102 {
1103         struct ib_qp *qp = context;
1104         unsigned long flags;
1105
1106         spin_lock_irqsave(&qp->device->qp_open_list_lock, flags);
1107         list_for_each_entry(event->element.qp, &qp->open_list, open_list)
1108                 if (event->element.qp->event_handler)
1109                         event->element.qp->event_handler(event, event->element.qp->qp_context);
1110         spin_unlock_irqrestore(&qp->device->qp_open_list_lock, flags);
1111 }
1112
1113 static struct ib_qp *__ib_open_qp(struct ib_qp *real_qp,
1114                                   void (*event_handler)(struct ib_event *, void *),
1115                                   void *qp_context)
1116 {
1117         struct ib_qp *qp;
1118         unsigned long flags;
1119         int err;
1120
1121         qp = kzalloc(sizeof *qp, GFP_KERNEL);
1122         if (!qp)
1123                 return ERR_PTR(-ENOMEM);
1124
1125         qp->real_qp = real_qp;
1126         err = ib_open_shared_qp_security(qp, real_qp->device);
1127         if (err) {
1128                 kfree(qp);
1129                 return ERR_PTR(err);
1130         }
1131
1132         qp->real_qp = real_qp;
1133         atomic_inc(&real_qp->usecnt);
1134         qp->device = real_qp->device;
1135         qp->event_handler = event_handler;
1136         qp->qp_context = qp_context;
1137         qp->qp_num = real_qp->qp_num;
1138         qp->qp_type = real_qp->qp_type;
1139
1140         spin_lock_irqsave(&real_qp->device->qp_open_list_lock, flags);
1141         list_add(&qp->open_list, &real_qp->open_list);
1142         spin_unlock_irqrestore(&real_qp->device->qp_open_list_lock, flags);
1143
1144         return qp;
1145 }
1146
1147 struct ib_qp *ib_open_qp(struct ib_xrcd *xrcd,
1148                          struct ib_qp_open_attr *qp_open_attr)
1149 {
1150         struct ib_qp *qp, *real_qp;
1151
1152         if (qp_open_attr->qp_type != IB_QPT_XRC_TGT)
1153                 return ERR_PTR(-EINVAL);
1154
1155         down_read(&xrcd->tgt_qps_rwsem);
1156         real_qp = xa_load(&xrcd->tgt_qps, qp_open_attr->qp_num);
1157         if (!real_qp) {
1158                 up_read(&xrcd->tgt_qps_rwsem);
1159                 return ERR_PTR(-EINVAL);
1160         }
1161         qp = __ib_open_qp(real_qp, qp_open_attr->event_handler,
1162                           qp_open_attr->qp_context);
1163         up_read(&xrcd->tgt_qps_rwsem);
1164         return qp;
1165 }
1166 EXPORT_SYMBOL(ib_open_qp);
1167
1168 static struct ib_qp *create_xrc_qp_user(struct ib_qp *qp,
1169                                         struct ib_qp_init_attr *qp_init_attr)
1170 {
1171         struct ib_qp *real_qp = qp;
1172         int err;
1173
1174         qp->event_handler = __ib_shared_qp_event_handler;
1175         qp->qp_context = qp;
1176         qp->pd = NULL;
1177         qp->send_cq = qp->recv_cq = NULL;
1178         qp->srq = NULL;
1179         qp->xrcd = qp_init_attr->xrcd;
1180         atomic_inc(&qp_init_attr->xrcd->usecnt);
1181         INIT_LIST_HEAD(&qp->open_list);
1182
1183         qp = __ib_open_qp(real_qp, qp_init_attr->event_handler,
1184                           qp_init_attr->qp_context);
1185         if (IS_ERR(qp))
1186                 return qp;
1187
1188         err = xa_err(xa_store(&qp_init_attr->xrcd->tgt_qps, real_qp->qp_num,
1189                               real_qp, GFP_KERNEL));
1190         if (err) {
1191                 ib_close_qp(qp);
1192                 return ERR_PTR(err);
1193         }
1194         return qp;
1195 }
1196
1197 static struct ib_qp *create_qp(struct ib_device *dev, struct ib_pd *pd,
1198                                struct ib_qp_init_attr *attr,
1199                                struct ib_udata *udata,
1200                                struct ib_uqp_object *uobj, const char *caller)
1201 {
1202         struct ib_udata dummy = {};
1203         struct ib_qp *qp;
1204         int ret;
1205
1206         if (!dev->ops.create_qp)
1207                 return ERR_PTR(-EOPNOTSUPP);
1208
1209         qp = rdma_zalloc_drv_obj_numa(dev, ib_qp);
1210         if (!qp)
1211                 return ERR_PTR(-ENOMEM);
1212
1213         qp->device = dev;
1214         qp->pd = pd;
1215         qp->uobject = uobj;
1216         qp->real_qp = qp;
1217
1218         qp->qp_type = attr->qp_type;
1219         qp->rwq_ind_tbl = attr->rwq_ind_tbl;
1220         qp->srq = attr->srq;
1221         qp->event_handler = attr->event_handler;
1222         qp->port = attr->port_num;
1223         qp->qp_context = attr->qp_context;
1224
1225         spin_lock_init(&qp->mr_lock);
1226         INIT_LIST_HEAD(&qp->rdma_mrs);
1227         INIT_LIST_HEAD(&qp->sig_mrs);
1228
1229         qp->send_cq = attr->send_cq;
1230         qp->recv_cq = attr->recv_cq;
1231
1232         rdma_restrack_new(&qp->res, RDMA_RESTRACK_QP);
1233         WARN_ONCE(!udata && !caller, "Missing kernel QP owner");
1234         rdma_restrack_set_name(&qp->res, udata ? NULL : caller);
1235         ret = dev->ops.create_qp(qp, attr, udata);
1236         if (ret)
1237                 goto err_create;
1238
1239         /*
1240          * TODO: The mlx4 internally overwrites send_cq and recv_cq.
1241          * Unfortunately, it is not an easy task to fix that driver.
1242          */
1243         qp->send_cq = attr->send_cq;
1244         qp->recv_cq = attr->recv_cq;
1245
1246         ret = ib_create_qp_security(qp, dev);
1247         if (ret)
1248                 goto err_security;
1249
1250         rdma_restrack_add(&qp->res);
1251         return qp;
1252
1253 err_security:
1254         qp->device->ops.destroy_qp(qp, udata ? &dummy : NULL);
1255 err_create:
1256         rdma_restrack_put(&qp->res);
1257         kfree(qp);
1258         return ERR_PTR(ret);
1259
1260 }
1261
1262 /**
1263  * ib_create_qp_user - Creates a QP associated with the specified protection
1264  *   domain.
1265  * @dev: IB device
1266  * @pd: The protection domain associated with the QP.
1267  * @attr: A list of initial attributes required to create the
1268  *   QP.  If QP creation succeeds, then the attributes are updated to
1269  *   the actual capabilities of the created QP.
1270  * @udata: User data
1271  * @uobj: uverbs obect
1272  * @caller: caller's build-time module name
1273  */
1274 struct ib_qp *ib_create_qp_user(struct ib_device *dev, struct ib_pd *pd,
1275                                 struct ib_qp_init_attr *attr,
1276                                 struct ib_udata *udata,
1277                                 struct ib_uqp_object *uobj, const char *caller)
1278 {
1279         struct ib_qp *qp, *xrc_qp;
1280
1281         if (attr->qp_type == IB_QPT_XRC_TGT)
1282                 qp = create_qp(dev, pd, attr, NULL, NULL, caller);
1283         else
1284                 qp = create_qp(dev, pd, attr, udata, uobj, NULL);
1285         if (attr->qp_type != IB_QPT_XRC_TGT || IS_ERR(qp))
1286                 return qp;
1287
1288         xrc_qp = create_xrc_qp_user(qp, attr);
1289         if (IS_ERR(xrc_qp)) {
1290                 ib_destroy_qp(qp);
1291                 return xrc_qp;
1292         }
1293
1294         xrc_qp->uobject = uobj;
1295         return xrc_qp;
1296 }
1297 EXPORT_SYMBOL(ib_create_qp_user);
1298
1299 void ib_qp_usecnt_inc(struct ib_qp *qp)
1300 {
1301         if (qp->pd)
1302                 atomic_inc(&qp->pd->usecnt);
1303         if (qp->send_cq)
1304                 atomic_inc(&qp->send_cq->usecnt);
1305         if (qp->recv_cq)
1306                 atomic_inc(&qp->recv_cq->usecnt);
1307         if (qp->srq)
1308                 atomic_inc(&qp->srq->usecnt);
1309         if (qp->rwq_ind_tbl)
1310                 atomic_inc(&qp->rwq_ind_tbl->usecnt);
1311 }
1312 EXPORT_SYMBOL(ib_qp_usecnt_inc);
1313
1314 void ib_qp_usecnt_dec(struct ib_qp *qp)
1315 {
1316         if (qp->rwq_ind_tbl)
1317                 atomic_dec(&qp->rwq_ind_tbl->usecnt);
1318         if (qp->srq)
1319                 atomic_dec(&qp->srq->usecnt);
1320         if (qp->recv_cq)
1321                 atomic_dec(&qp->recv_cq->usecnt);
1322         if (qp->send_cq)
1323                 atomic_dec(&qp->send_cq->usecnt);
1324         if (qp->pd)
1325                 atomic_dec(&qp->pd->usecnt);
1326 }
1327 EXPORT_SYMBOL(ib_qp_usecnt_dec);
1328
1329 struct ib_qp *ib_create_qp_kernel(struct ib_pd *pd,
1330                                   struct ib_qp_init_attr *qp_init_attr,
1331                                   const char *caller)
1332 {
1333         struct ib_device *device = pd->device;
1334         struct ib_qp *qp;
1335         int ret;
1336
1337         /*
1338          * If the callers is using the RDMA API calculate the resources
1339          * needed for the RDMA READ/WRITE operations.
1340          *
1341          * Note that these callers need to pass in a port number.
1342          */
1343         if (qp_init_attr->cap.max_rdma_ctxs)
1344                 rdma_rw_init_qp(device, qp_init_attr);
1345
1346         qp = create_qp(device, pd, qp_init_attr, NULL, NULL, caller);
1347         if (IS_ERR(qp))
1348                 return qp;
1349
1350         ib_qp_usecnt_inc(qp);
1351
1352         if (qp_init_attr->cap.max_rdma_ctxs) {
1353                 ret = rdma_rw_init_mrs(qp, qp_init_attr);
1354                 if (ret)
1355                         goto err;
1356         }
1357
1358         /*
1359          * Note: all hw drivers guarantee that max_send_sge is lower than
1360          * the device RDMA WRITE SGE limit but not all hw drivers ensure that
1361          * max_send_sge <= max_sge_rd.
1362          */
1363         qp->max_write_sge = qp_init_attr->cap.max_send_sge;
1364         qp->max_read_sge = min_t(u32, qp_init_attr->cap.max_send_sge,
1365                                  device->attrs.max_sge_rd);
1366         if (qp_init_attr->create_flags & IB_QP_CREATE_INTEGRITY_EN)
1367                 qp->integrity_en = true;
1368
1369         return qp;
1370
1371 err:
1372         ib_destroy_qp(qp);
1373         return ERR_PTR(ret);
1374
1375 }
1376 EXPORT_SYMBOL(ib_create_qp_kernel);
1377
1378 static const struct {
1379         int                     valid;
1380         enum ib_qp_attr_mask    req_param[IB_QPT_MAX];
1381         enum ib_qp_attr_mask    opt_param[IB_QPT_MAX];
1382 } qp_state_table[IB_QPS_ERR + 1][IB_QPS_ERR + 1] = {
1383         [IB_QPS_RESET] = {
1384                 [IB_QPS_RESET] = { .valid = 1 },
1385                 [IB_QPS_INIT]  = {
1386                         .valid = 1,
1387                         .req_param = {
1388                                 [IB_QPT_UD]  = (IB_QP_PKEY_INDEX                |
1389                                                 IB_QP_PORT                      |
1390                                                 IB_QP_QKEY),
1391                                 [IB_QPT_RAW_PACKET] = IB_QP_PORT,
1392                                 [IB_QPT_UC]  = (IB_QP_PKEY_INDEX                |
1393                                                 IB_QP_PORT                      |
1394                                                 IB_QP_ACCESS_FLAGS),
1395                                 [IB_QPT_RC]  = (IB_QP_PKEY_INDEX                |
1396                                                 IB_QP_PORT                      |
1397                                                 IB_QP_ACCESS_FLAGS),
1398                                 [IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX            |
1399                                                 IB_QP_PORT                      |
1400                                                 IB_QP_ACCESS_FLAGS),
1401                                 [IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX            |
1402                                                 IB_QP_PORT                      |
1403                                                 IB_QP_ACCESS_FLAGS),
1404                                 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX                |
1405                                                 IB_QP_QKEY),
1406                                 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX                |
1407                                                 IB_QP_QKEY),
1408                         }
1409                 },
1410         },
1411         [IB_QPS_INIT]  = {
1412                 [IB_QPS_RESET] = { .valid = 1 },
1413                 [IB_QPS_ERR] =   { .valid = 1 },
1414                 [IB_QPS_INIT]  = {
1415                         .valid = 1,
1416                         .opt_param = {
1417                                 [IB_QPT_UD]  = (IB_QP_PKEY_INDEX                |
1418                                                 IB_QP_PORT                      |
1419                                                 IB_QP_QKEY),
1420                                 [IB_QPT_UC]  = (IB_QP_PKEY_INDEX                |
1421                                                 IB_QP_PORT                      |
1422                                                 IB_QP_ACCESS_FLAGS),
1423                                 [IB_QPT_RC]  = (IB_QP_PKEY_INDEX                |
1424                                                 IB_QP_PORT                      |
1425                                                 IB_QP_ACCESS_FLAGS),
1426                                 [IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX            |
1427                                                 IB_QP_PORT                      |
1428                                                 IB_QP_ACCESS_FLAGS),
1429                                 [IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX            |
1430                                                 IB_QP_PORT                      |
1431                                                 IB_QP_ACCESS_FLAGS),
1432                                 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX                |
1433                                                 IB_QP_QKEY),
1434                                 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX                |
1435                                                 IB_QP_QKEY),
1436                         }
1437                 },
1438                 [IB_QPS_RTR]   = {
1439                         .valid = 1,
1440                         .req_param = {
1441                                 [IB_QPT_UC]  = (IB_QP_AV                        |
1442                                                 IB_QP_PATH_MTU                  |
1443                                                 IB_QP_DEST_QPN                  |
1444                                                 IB_QP_RQ_PSN),
1445                                 [IB_QPT_RC]  = (IB_QP_AV                        |
1446                                                 IB_QP_PATH_MTU                  |
1447                                                 IB_QP_DEST_QPN                  |
1448                                                 IB_QP_RQ_PSN                    |
1449                                                 IB_QP_MAX_DEST_RD_ATOMIC        |
1450                                                 IB_QP_MIN_RNR_TIMER),
1451                                 [IB_QPT_XRC_INI] = (IB_QP_AV                    |
1452                                                 IB_QP_PATH_MTU                  |
1453                                                 IB_QP_DEST_QPN                  |
1454                                                 IB_QP_RQ_PSN),
1455                                 [IB_QPT_XRC_TGT] = (IB_QP_AV                    |
1456                                                 IB_QP_PATH_MTU                  |
1457                                                 IB_QP_DEST_QPN                  |
1458                                                 IB_QP_RQ_PSN                    |
1459                                                 IB_QP_MAX_DEST_RD_ATOMIC        |
1460                                                 IB_QP_MIN_RNR_TIMER),
1461                         },
1462                         .opt_param = {
1463                                  [IB_QPT_UD]  = (IB_QP_PKEY_INDEX               |
1464                                                  IB_QP_QKEY),
1465                                  [IB_QPT_UC]  = (IB_QP_ALT_PATH                 |
1466                                                  IB_QP_ACCESS_FLAGS             |
1467                                                  IB_QP_PKEY_INDEX),
1468                                  [IB_QPT_RC]  = (IB_QP_ALT_PATH                 |
1469                                                  IB_QP_ACCESS_FLAGS             |
1470                                                  IB_QP_PKEY_INDEX),
1471                                  [IB_QPT_XRC_INI] = (IB_QP_ALT_PATH             |
1472                                                  IB_QP_ACCESS_FLAGS             |
1473                                                  IB_QP_PKEY_INDEX),
1474                                  [IB_QPT_XRC_TGT] = (IB_QP_ALT_PATH             |
1475                                                  IB_QP_ACCESS_FLAGS             |
1476                                                  IB_QP_PKEY_INDEX),
1477                                  [IB_QPT_SMI] = (IB_QP_PKEY_INDEX               |
1478                                                  IB_QP_QKEY),
1479                                  [IB_QPT_GSI] = (IB_QP_PKEY_INDEX               |
1480                                                  IB_QP_QKEY),
1481                          },
1482                 },
1483         },
1484         [IB_QPS_RTR]   = {
1485                 [IB_QPS_RESET] = { .valid = 1 },
1486                 [IB_QPS_ERR] =   { .valid = 1 },
1487                 [IB_QPS_RTS]   = {
1488                         .valid = 1,
1489                         .req_param = {
1490                                 [IB_QPT_UD]  = IB_QP_SQ_PSN,
1491                                 [IB_QPT_UC]  = IB_QP_SQ_PSN,
1492                                 [IB_QPT_RC]  = (IB_QP_TIMEOUT                   |
1493                                                 IB_QP_RETRY_CNT                 |
1494                                                 IB_QP_RNR_RETRY                 |
1495                                                 IB_QP_SQ_PSN                    |
1496                                                 IB_QP_MAX_QP_RD_ATOMIC),
1497                                 [IB_QPT_XRC_INI] = (IB_QP_TIMEOUT               |
1498                                                 IB_QP_RETRY_CNT                 |
1499                                                 IB_QP_RNR_RETRY                 |
1500                                                 IB_QP_SQ_PSN                    |
1501                                                 IB_QP_MAX_QP_RD_ATOMIC),
1502                                 [IB_QPT_XRC_TGT] = (IB_QP_TIMEOUT               |
1503                                                 IB_QP_SQ_PSN),
1504                                 [IB_QPT_SMI] = IB_QP_SQ_PSN,
1505                                 [IB_QPT_GSI] = IB_QP_SQ_PSN,
1506                         },
1507                         .opt_param = {
1508                                  [IB_QPT_UD]  = (IB_QP_CUR_STATE                |
1509                                                  IB_QP_QKEY),
1510                                  [IB_QPT_UC]  = (IB_QP_CUR_STATE                |
1511                                                  IB_QP_ALT_PATH                 |
1512                                                  IB_QP_ACCESS_FLAGS             |
1513                                                  IB_QP_PATH_MIG_STATE),
1514                                  [IB_QPT_RC]  = (IB_QP_CUR_STATE                |
1515                                                  IB_QP_ALT_PATH                 |
1516                                                  IB_QP_ACCESS_FLAGS             |
1517                                                  IB_QP_MIN_RNR_TIMER            |
1518                                                  IB_QP_PATH_MIG_STATE),
1519                                  [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE            |
1520                                                  IB_QP_ALT_PATH                 |
1521                                                  IB_QP_ACCESS_FLAGS             |
1522                                                  IB_QP_PATH_MIG_STATE),
1523                                  [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE            |
1524                                                  IB_QP_ALT_PATH                 |
1525                                                  IB_QP_ACCESS_FLAGS             |
1526                                                  IB_QP_MIN_RNR_TIMER            |
1527                                                  IB_QP_PATH_MIG_STATE),
1528                                  [IB_QPT_SMI] = (IB_QP_CUR_STATE                |
1529                                                  IB_QP_QKEY),
1530                                  [IB_QPT_GSI] = (IB_QP_CUR_STATE                |
1531                                                  IB_QP_QKEY),
1532                                  [IB_QPT_RAW_PACKET] = IB_QP_RATE_LIMIT,
1533                          }
1534                 }
1535         },
1536         [IB_QPS_RTS]   = {
1537                 [IB_QPS_RESET] = { .valid = 1 },
1538                 [IB_QPS_ERR] =   { .valid = 1 },
1539                 [IB_QPS_RTS]   = {
1540                         .valid = 1,
1541                         .opt_param = {
1542                                 [IB_QPT_UD]  = (IB_QP_CUR_STATE                 |
1543                                                 IB_QP_QKEY),
1544                                 [IB_QPT_UC]  = (IB_QP_CUR_STATE                 |
1545                                                 IB_QP_ACCESS_FLAGS              |
1546                                                 IB_QP_ALT_PATH                  |
1547                                                 IB_QP_PATH_MIG_STATE),
1548                                 [IB_QPT_RC]  = (IB_QP_CUR_STATE                 |
1549                                                 IB_QP_ACCESS_FLAGS              |
1550                                                 IB_QP_ALT_PATH                  |
1551                                                 IB_QP_PATH_MIG_STATE            |
1552                                                 IB_QP_MIN_RNR_TIMER),
1553                                 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE             |
1554                                                 IB_QP_ACCESS_FLAGS              |
1555                                                 IB_QP_ALT_PATH                  |
1556                                                 IB_QP_PATH_MIG_STATE),
1557                                 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE             |
1558                                                 IB_QP_ACCESS_FLAGS              |
1559                                                 IB_QP_ALT_PATH                  |
1560                                                 IB_QP_PATH_MIG_STATE            |
1561                                                 IB_QP_MIN_RNR_TIMER),
1562                                 [IB_QPT_SMI] = (IB_QP_CUR_STATE                 |
1563                                                 IB_QP_QKEY),
1564                                 [IB_QPT_GSI] = (IB_QP_CUR_STATE                 |
1565                                                 IB_QP_QKEY),
1566                                 [IB_QPT_RAW_PACKET] = IB_QP_RATE_LIMIT,
1567                         }
1568                 },
1569                 [IB_QPS_SQD]   = {
1570                         .valid = 1,
1571                         .opt_param = {
1572                                 [IB_QPT_UD]  = IB_QP_EN_SQD_ASYNC_NOTIFY,
1573                                 [IB_QPT_UC]  = IB_QP_EN_SQD_ASYNC_NOTIFY,
1574                                 [IB_QPT_RC]  = IB_QP_EN_SQD_ASYNC_NOTIFY,
1575                                 [IB_QPT_XRC_INI] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1576                                 [IB_QPT_XRC_TGT] = IB_QP_EN_SQD_ASYNC_NOTIFY, /* ??? */
1577                                 [IB_QPT_SMI] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1578                                 [IB_QPT_GSI] = IB_QP_EN_SQD_ASYNC_NOTIFY
1579                         }
1580                 },
1581         },
1582         [IB_QPS_SQD]   = {
1583                 [IB_QPS_RESET] = { .valid = 1 },
1584                 [IB_QPS_ERR] =   { .valid = 1 },
1585                 [IB_QPS_RTS]   = {
1586                         .valid = 1,
1587                         .opt_param = {
1588                                 [IB_QPT_UD]  = (IB_QP_CUR_STATE                 |
1589                                                 IB_QP_QKEY),
1590                                 [IB_QPT_UC]  = (IB_QP_CUR_STATE                 |
1591                                                 IB_QP_ALT_PATH                  |
1592                                                 IB_QP_ACCESS_FLAGS              |
1593                                                 IB_QP_PATH_MIG_STATE),
1594                                 [IB_QPT_RC]  = (IB_QP_CUR_STATE                 |
1595                                                 IB_QP_ALT_PATH                  |
1596                                                 IB_QP_ACCESS_FLAGS              |
1597                                                 IB_QP_MIN_RNR_TIMER             |
1598                                                 IB_QP_PATH_MIG_STATE),
1599                                 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE             |
1600                                                 IB_QP_ALT_PATH                  |
1601                                                 IB_QP_ACCESS_FLAGS              |
1602                                                 IB_QP_PATH_MIG_STATE),
1603                                 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE             |
1604                                                 IB_QP_ALT_PATH                  |
1605                                                 IB_QP_ACCESS_FLAGS              |
1606                                                 IB_QP_MIN_RNR_TIMER             |
1607                                                 IB_QP_PATH_MIG_STATE),
1608                                 [IB_QPT_SMI] = (IB_QP_CUR_STATE                 |
1609                                                 IB_QP_QKEY),
1610                                 [IB_QPT_GSI] = (IB_QP_CUR_STATE                 |
1611                                                 IB_QP_QKEY),
1612                         }
1613                 },
1614                 [IB_QPS_SQD]   = {
1615                         .valid = 1,
1616                         .opt_param = {
1617                                 [IB_QPT_UD]  = (IB_QP_PKEY_INDEX                |
1618                                                 IB_QP_QKEY),
1619                                 [IB_QPT_UC]  = (IB_QP_AV                        |
1620                                                 IB_QP_ALT_PATH                  |
1621                                                 IB_QP_ACCESS_FLAGS              |
1622                                                 IB_QP_PKEY_INDEX                |
1623                                                 IB_QP_PATH_MIG_STATE),
1624                                 [IB_QPT_RC]  = (IB_QP_PORT                      |
1625                                                 IB_QP_AV                        |
1626                                                 IB_QP_TIMEOUT                   |
1627                                                 IB_QP_RETRY_CNT                 |
1628                                                 IB_QP_RNR_RETRY                 |
1629                                                 IB_QP_MAX_QP_RD_ATOMIC          |
1630                                                 IB_QP_MAX_DEST_RD_ATOMIC        |
1631                                                 IB_QP_ALT_PATH                  |
1632                                                 IB_QP_ACCESS_FLAGS              |
1633                                                 IB_QP_PKEY_INDEX                |
1634                                                 IB_QP_MIN_RNR_TIMER             |
1635                                                 IB_QP_PATH_MIG_STATE),
1636                                 [IB_QPT_XRC_INI] = (IB_QP_PORT                  |
1637                                                 IB_QP_AV                        |
1638                                                 IB_QP_TIMEOUT                   |
1639                                                 IB_QP_RETRY_CNT                 |
1640                                                 IB_QP_RNR_RETRY                 |
1641                                                 IB_QP_MAX_QP_RD_ATOMIC          |
1642                                                 IB_QP_ALT_PATH                  |
1643                                                 IB_QP_ACCESS_FLAGS              |
1644                                                 IB_QP_PKEY_INDEX                |
1645                                                 IB_QP_PATH_MIG_STATE),
1646                                 [IB_QPT_XRC_TGT] = (IB_QP_PORT                  |
1647                                                 IB_QP_AV                        |
1648                                                 IB_QP_TIMEOUT                   |
1649                                                 IB_QP_MAX_DEST_RD_ATOMIC        |
1650                                                 IB_QP_ALT_PATH                  |
1651                                                 IB_QP_ACCESS_FLAGS              |
1652                                                 IB_QP_PKEY_INDEX                |
1653                                                 IB_QP_MIN_RNR_TIMER             |
1654                                                 IB_QP_PATH_MIG_STATE),
1655                                 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX                |
1656                                                 IB_QP_QKEY),
1657                                 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX                |
1658                                                 IB_QP_QKEY),
1659                         }
1660                 }
1661         },
1662         [IB_QPS_SQE]   = {
1663                 [IB_QPS_RESET] = { .valid = 1 },
1664                 [IB_QPS_ERR] =   { .valid = 1 },
1665                 [IB_QPS_RTS]   = {
1666                         .valid = 1,
1667                         .opt_param = {
1668                                 [IB_QPT_UD]  = (IB_QP_CUR_STATE                 |
1669                                                 IB_QP_QKEY),
1670                                 [IB_QPT_UC]  = (IB_QP_CUR_STATE                 |
1671                                                 IB_QP_ACCESS_FLAGS),
1672                                 [IB_QPT_SMI] = (IB_QP_CUR_STATE                 |
1673                                                 IB_QP_QKEY),
1674                                 [IB_QPT_GSI] = (IB_QP_CUR_STATE                 |
1675                                                 IB_QP_QKEY),
1676                         }
1677                 }
1678         },
1679         [IB_QPS_ERR] = {
1680                 [IB_QPS_RESET] = { .valid = 1 },
1681                 [IB_QPS_ERR] =   { .valid = 1 }
1682         }
1683 };
1684
1685 bool ib_modify_qp_is_ok(enum ib_qp_state cur_state, enum ib_qp_state next_state,
1686                         enum ib_qp_type type, enum ib_qp_attr_mask mask)
1687 {
1688         enum ib_qp_attr_mask req_param, opt_param;
1689
1690         if (mask & IB_QP_CUR_STATE  &&
1691             cur_state != IB_QPS_RTR && cur_state != IB_QPS_RTS &&
1692             cur_state != IB_QPS_SQD && cur_state != IB_QPS_SQE)
1693                 return false;
1694
1695         if (!qp_state_table[cur_state][next_state].valid)
1696                 return false;
1697
1698         req_param = qp_state_table[cur_state][next_state].req_param[type];
1699         opt_param = qp_state_table[cur_state][next_state].opt_param[type];
1700
1701         if ((mask & req_param) != req_param)
1702                 return false;
1703
1704         if (mask & ~(req_param | opt_param | IB_QP_STATE))
1705                 return false;
1706
1707         return true;
1708 }
1709 EXPORT_SYMBOL(ib_modify_qp_is_ok);
1710
1711 /**
1712  * ib_resolve_eth_dmac - Resolve destination mac address
1713  * @device:             Device to consider
1714  * @ah_attr:            address handle attribute which describes the
1715  *                      source and destination parameters
1716  * ib_resolve_eth_dmac() resolves destination mac address and L3 hop limit It
1717  * returns 0 on success or appropriate error code. It initializes the
1718  * necessary ah_attr fields when call is successful.
1719  */
1720 static int ib_resolve_eth_dmac(struct ib_device *device,
1721                                struct rdma_ah_attr *ah_attr)
1722 {
1723         int ret = 0;
1724
1725         if (rdma_is_multicast_addr((struct in6_addr *)ah_attr->grh.dgid.raw)) {
1726                 if (ipv6_addr_v4mapped((struct in6_addr *)ah_attr->grh.dgid.raw)) {
1727                         __be32 addr = 0;
1728
1729                         memcpy(&addr, ah_attr->grh.dgid.raw + 12, 4);
1730                         ip_eth_mc_map(addr, (char *)ah_attr->roce.dmac);
1731                 } else {
1732                         ipv6_eth_mc_map((struct in6_addr *)ah_attr->grh.dgid.raw,
1733                                         (char *)ah_attr->roce.dmac);
1734                 }
1735         } else {
1736                 ret = ib_resolve_unicast_gid_dmac(device, ah_attr);
1737         }
1738         return ret;
1739 }
1740
1741 static bool is_qp_type_connected(const struct ib_qp *qp)
1742 {
1743         return (qp->qp_type == IB_QPT_UC ||
1744                 qp->qp_type == IB_QPT_RC ||
1745                 qp->qp_type == IB_QPT_XRC_INI ||
1746                 qp->qp_type == IB_QPT_XRC_TGT);
1747 }
1748
1749 /*
1750  * IB core internal function to perform QP attributes modification.
1751  */
1752 static int _ib_modify_qp(struct ib_qp *qp, struct ib_qp_attr *attr,
1753                          int attr_mask, struct ib_udata *udata)
1754 {
1755         u32 port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
1756         const struct ib_gid_attr *old_sgid_attr_av;
1757         const struct ib_gid_attr *old_sgid_attr_alt_av;
1758         int ret;
1759
1760         attr->xmit_slave = NULL;
1761         if (attr_mask & IB_QP_AV) {
1762                 ret = rdma_fill_sgid_attr(qp->device, &attr->ah_attr,
1763                                           &old_sgid_attr_av);
1764                 if (ret)
1765                         return ret;
1766
1767                 if (attr->ah_attr.type == RDMA_AH_ATTR_TYPE_ROCE &&
1768                     is_qp_type_connected(qp)) {
1769                         struct net_device *slave;
1770
1771                         /*
1772                          * If the user provided the qp_attr then we have to
1773                          * resolve it. Kerne users have to provide already
1774                          * resolved rdma_ah_attr's.
1775                          */
1776                         if (udata) {
1777                                 ret = ib_resolve_eth_dmac(qp->device,
1778                                                           &attr->ah_attr);
1779                                 if (ret)
1780                                         goto out_av;
1781                         }
1782                         slave = rdma_lag_get_ah_roce_slave(qp->device,
1783                                                            &attr->ah_attr,
1784                                                            GFP_KERNEL);
1785                         if (IS_ERR(slave)) {
1786                                 ret = PTR_ERR(slave);
1787                                 goto out_av;
1788                         }
1789                         attr->xmit_slave = slave;
1790                 }
1791         }
1792         if (attr_mask & IB_QP_ALT_PATH) {
1793                 /*
1794                  * FIXME: This does not track the migration state, so if the
1795                  * user loads a new alternate path after the HW has migrated
1796                  * from primary->alternate we will keep the wrong
1797                  * references. This is OK for IB because the reference
1798                  * counting does not serve any functional purpose.
1799                  */
1800                 ret = rdma_fill_sgid_attr(qp->device, &attr->alt_ah_attr,
1801                                           &old_sgid_attr_alt_av);
1802                 if (ret)
1803                         goto out_av;
1804
1805                 /*
1806                  * Today the core code can only handle alternate paths and APM
1807                  * for IB. Ban them in roce mode.
1808                  */
1809                 if (!(rdma_protocol_ib(qp->device,
1810                                        attr->alt_ah_attr.port_num) &&
1811                       rdma_protocol_ib(qp->device, port))) {
1812                         ret = -EINVAL;
1813                         goto out;
1814                 }
1815         }
1816
1817         if (rdma_ib_or_roce(qp->device, port)) {
1818                 if (attr_mask & IB_QP_RQ_PSN && attr->rq_psn & ~0xffffff) {
1819                         dev_warn(&qp->device->dev,
1820                                  "%s rq_psn overflow, masking to 24 bits\n",
1821                                  __func__);
1822                         attr->rq_psn &= 0xffffff;
1823                 }
1824
1825                 if (attr_mask & IB_QP_SQ_PSN && attr->sq_psn & ~0xffffff) {
1826                         dev_warn(&qp->device->dev,
1827                                  " %s sq_psn overflow, masking to 24 bits\n",
1828                                  __func__);
1829                         attr->sq_psn &= 0xffffff;
1830                 }
1831         }
1832
1833         /*
1834          * Bind this qp to a counter automatically based on the rdma counter
1835          * rules. This only set in RST2INIT with port specified
1836          */
1837         if (!qp->counter && (attr_mask & IB_QP_PORT) &&
1838             ((attr_mask & IB_QP_STATE) && attr->qp_state == IB_QPS_INIT))
1839                 rdma_counter_bind_qp_auto(qp, attr->port_num);
1840
1841         ret = ib_security_modify_qp(qp, attr, attr_mask, udata);
1842         if (ret)
1843                 goto out;
1844
1845         if (attr_mask & IB_QP_PORT)
1846                 qp->port = attr->port_num;
1847         if (attr_mask & IB_QP_AV)
1848                 qp->av_sgid_attr =
1849                         rdma_update_sgid_attr(&attr->ah_attr, qp->av_sgid_attr);
1850         if (attr_mask & IB_QP_ALT_PATH)
1851                 qp->alt_path_sgid_attr = rdma_update_sgid_attr(
1852                         &attr->alt_ah_attr, qp->alt_path_sgid_attr);
1853
1854 out:
1855         if (attr_mask & IB_QP_ALT_PATH)
1856                 rdma_unfill_sgid_attr(&attr->alt_ah_attr, old_sgid_attr_alt_av);
1857 out_av:
1858         if (attr_mask & IB_QP_AV) {
1859                 rdma_lag_put_ah_roce_slave(attr->xmit_slave);
1860                 rdma_unfill_sgid_attr(&attr->ah_attr, old_sgid_attr_av);
1861         }
1862         return ret;
1863 }
1864
1865 /**
1866  * ib_modify_qp_with_udata - Modifies the attributes for the specified QP.
1867  * @ib_qp: The QP to modify.
1868  * @attr: On input, specifies the QP attributes to modify.  On output,
1869  *   the current values of selected QP attributes are returned.
1870  * @attr_mask: A bit-mask used to specify which attributes of the QP
1871  *   are being modified.
1872  * @udata: pointer to user's input output buffer information
1873  *   are being modified.
1874  * It returns 0 on success and returns appropriate error code on error.
1875  */
1876 int ib_modify_qp_with_udata(struct ib_qp *ib_qp, struct ib_qp_attr *attr,
1877                             int attr_mask, struct ib_udata *udata)
1878 {
1879         return _ib_modify_qp(ib_qp->real_qp, attr, attr_mask, udata);
1880 }
1881 EXPORT_SYMBOL(ib_modify_qp_with_udata);
1882
1883 int ib_get_eth_speed(struct ib_device *dev, u32 port_num, u16 *speed, u8 *width)
1884 {
1885         int rc;
1886         u32 netdev_speed;
1887         struct net_device *netdev;
1888         struct ethtool_link_ksettings lksettings;
1889
1890         if (rdma_port_get_link_layer(dev, port_num) != IB_LINK_LAYER_ETHERNET)
1891                 return -EINVAL;
1892
1893         netdev = ib_device_get_netdev(dev, port_num);
1894         if (!netdev)
1895                 return -ENODEV;
1896
1897         rtnl_lock();
1898         rc = __ethtool_get_link_ksettings(netdev, &lksettings);
1899         rtnl_unlock();
1900
1901         dev_put(netdev);
1902
1903         if (!rc && lksettings.base.speed != (u32)SPEED_UNKNOWN) {
1904                 netdev_speed = lksettings.base.speed;
1905         } else {
1906                 netdev_speed = SPEED_1000;
1907                 pr_warn("%s speed is unknown, defaulting to %u\n", netdev->name,
1908                         netdev_speed);
1909         }
1910
1911         if (netdev_speed <= SPEED_1000) {
1912                 *width = IB_WIDTH_1X;
1913                 *speed = IB_SPEED_SDR;
1914         } else if (netdev_speed <= SPEED_10000) {
1915                 *width = IB_WIDTH_1X;
1916                 *speed = IB_SPEED_FDR10;
1917         } else if (netdev_speed <= SPEED_20000) {
1918                 *width = IB_WIDTH_4X;
1919                 *speed = IB_SPEED_DDR;
1920         } else if (netdev_speed <= SPEED_25000) {
1921                 *width = IB_WIDTH_1X;
1922                 *speed = IB_SPEED_EDR;
1923         } else if (netdev_speed <= SPEED_40000) {
1924                 *width = IB_WIDTH_4X;
1925                 *speed = IB_SPEED_FDR10;
1926         } else {
1927                 *width = IB_WIDTH_4X;
1928                 *speed = IB_SPEED_EDR;
1929         }
1930
1931         return 0;
1932 }
1933 EXPORT_SYMBOL(ib_get_eth_speed);
1934
1935 int ib_modify_qp(struct ib_qp *qp,
1936                  struct ib_qp_attr *qp_attr,
1937                  int qp_attr_mask)
1938 {
1939         return _ib_modify_qp(qp->real_qp, qp_attr, qp_attr_mask, NULL);
1940 }
1941 EXPORT_SYMBOL(ib_modify_qp);
1942
1943 int ib_query_qp(struct ib_qp *qp,
1944                 struct ib_qp_attr *qp_attr,
1945                 int qp_attr_mask,
1946                 struct ib_qp_init_attr *qp_init_attr)
1947 {
1948         qp_attr->ah_attr.grh.sgid_attr = NULL;
1949         qp_attr->alt_ah_attr.grh.sgid_attr = NULL;
1950
1951         return qp->device->ops.query_qp ?
1952                 qp->device->ops.query_qp(qp->real_qp, qp_attr, qp_attr_mask,
1953                                          qp_init_attr) : -EOPNOTSUPP;
1954 }
1955 EXPORT_SYMBOL(ib_query_qp);
1956
1957 int ib_close_qp(struct ib_qp *qp)
1958 {
1959         struct ib_qp *real_qp;
1960         unsigned long flags;
1961
1962         real_qp = qp->real_qp;
1963         if (real_qp == qp)
1964                 return -EINVAL;
1965
1966         spin_lock_irqsave(&real_qp->device->qp_open_list_lock, flags);
1967         list_del(&qp->open_list);
1968         spin_unlock_irqrestore(&real_qp->device->qp_open_list_lock, flags);
1969
1970         atomic_dec(&real_qp->usecnt);
1971         if (qp->qp_sec)
1972                 ib_close_shared_qp_security(qp->qp_sec);
1973         kfree(qp);
1974
1975         return 0;
1976 }
1977 EXPORT_SYMBOL(ib_close_qp);
1978
1979 static int __ib_destroy_shared_qp(struct ib_qp *qp)
1980 {
1981         struct ib_xrcd *xrcd;
1982         struct ib_qp *real_qp;
1983         int ret;
1984
1985         real_qp = qp->real_qp;
1986         xrcd = real_qp->xrcd;
1987         down_write(&xrcd->tgt_qps_rwsem);
1988         ib_close_qp(qp);
1989         if (atomic_read(&real_qp->usecnt) == 0)
1990                 xa_erase(&xrcd->tgt_qps, real_qp->qp_num);
1991         else
1992                 real_qp = NULL;
1993         up_write(&xrcd->tgt_qps_rwsem);
1994
1995         if (real_qp) {
1996                 ret = ib_destroy_qp(real_qp);
1997                 if (!ret)
1998                         atomic_dec(&xrcd->usecnt);
1999         }
2000
2001         return 0;
2002 }
2003
2004 int ib_destroy_qp_user(struct ib_qp *qp, struct ib_udata *udata)
2005 {
2006         const struct ib_gid_attr *alt_path_sgid_attr = qp->alt_path_sgid_attr;
2007         const struct ib_gid_attr *av_sgid_attr = qp->av_sgid_attr;
2008         struct ib_qp_security *sec;
2009         int ret;
2010
2011         WARN_ON_ONCE(qp->mrs_used > 0);
2012
2013         if (atomic_read(&qp->usecnt))
2014                 return -EBUSY;
2015
2016         if (qp->real_qp != qp)
2017                 return __ib_destroy_shared_qp(qp);
2018
2019         sec  = qp->qp_sec;
2020         if (sec)
2021                 ib_destroy_qp_security_begin(sec);
2022
2023         if (!qp->uobject)
2024                 rdma_rw_cleanup_mrs(qp);
2025
2026         rdma_counter_unbind_qp(qp, true);
2027         ret = qp->device->ops.destroy_qp(qp, udata);
2028         if (ret) {
2029                 if (sec)
2030                         ib_destroy_qp_security_abort(sec);
2031                 return ret;
2032         }
2033
2034         if (alt_path_sgid_attr)
2035                 rdma_put_gid_attr(alt_path_sgid_attr);
2036         if (av_sgid_attr)
2037                 rdma_put_gid_attr(av_sgid_attr);
2038
2039         ib_qp_usecnt_dec(qp);
2040         if (sec)
2041                 ib_destroy_qp_security_end(sec);
2042
2043         rdma_restrack_del(&qp->res);
2044         kfree(qp);
2045         return ret;
2046 }
2047 EXPORT_SYMBOL(ib_destroy_qp_user);
2048
2049 /* Completion queues */
2050
2051 struct ib_cq *__ib_create_cq(struct ib_device *device,
2052                              ib_comp_handler comp_handler,
2053                              void (*event_handler)(struct ib_event *, void *),
2054                              void *cq_context,
2055                              const struct ib_cq_init_attr *cq_attr,
2056                              const char *caller)
2057 {
2058         struct ib_cq *cq;
2059         int ret;
2060
2061         cq = rdma_zalloc_drv_obj(device, ib_cq);
2062         if (!cq)
2063                 return ERR_PTR(-ENOMEM);
2064
2065         cq->device = device;
2066         cq->uobject = NULL;
2067         cq->comp_handler = comp_handler;
2068         cq->event_handler = event_handler;
2069         cq->cq_context = cq_context;
2070         atomic_set(&cq->usecnt, 0);
2071
2072         rdma_restrack_new(&cq->res, RDMA_RESTRACK_CQ);
2073         rdma_restrack_set_name(&cq->res, caller);
2074
2075         ret = device->ops.create_cq(cq, cq_attr, NULL);
2076         if (ret) {
2077                 rdma_restrack_put(&cq->res);
2078                 kfree(cq);
2079                 return ERR_PTR(ret);
2080         }
2081
2082         rdma_restrack_add(&cq->res);
2083         return cq;
2084 }
2085 EXPORT_SYMBOL(__ib_create_cq);
2086
2087 int rdma_set_cq_moderation(struct ib_cq *cq, u16 cq_count, u16 cq_period)
2088 {
2089         if (cq->shared)
2090                 return -EOPNOTSUPP;
2091
2092         return cq->device->ops.modify_cq ?
2093                 cq->device->ops.modify_cq(cq, cq_count,
2094                                           cq_period) : -EOPNOTSUPP;
2095 }
2096 EXPORT_SYMBOL(rdma_set_cq_moderation);
2097
2098 int ib_destroy_cq_user(struct ib_cq *cq, struct ib_udata *udata)
2099 {
2100         int ret;
2101
2102         if (WARN_ON_ONCE(cq->shared))
2103                 return -EOPNOTSUPP;
2104
2105         if (atomic_read(&cq->usecnt))
2106                 return -EBUSY;
2107
2108         ret = cq->device->ops.destroy_cq(cq, udata);
2109         if (ret)
2110                 return ret;
2111
2112         rdma_restrack_del(&cq->res);
2113         kfree(cq);
2114         return ret;
2115 }
2116 EXPORT_SYMBOL(ib_destroy_cq_user);
2117
2118 int ib_resize_cq(struct ib_cq *cq, int cqe)
2119 {
2120         if (cq->shared)
2121                 return -EOPNOTSUPP;
2122
2123         return cq->device->ops.resize_cq ?
2124                 cq->device->ops.resize_cq(cq, cqe, NULL) : -EOPNOTSUPP;
2125 }
2126 EXPORT_SYMBOL(ib_resize_cq);
2127
2128 /* Memory regions */
2129
2130 struct ib_mr *ib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
2131                              u64 virt_addr, int access_flags)
2132 {
2133         struct ib_mr *mr;
2134
2135         if (access_flags & IB_ACCESS_ON_DEMAND) {
2136                 if (!(pd->device->attrs.kernel_cap_flags &
2137                       IBK_ON_DEMAND_PAGING)) {
2138                         pr_debug("ODP support not available\n");
2139                         return ERR_PTR(-EINVAL);
2140                 }
2141         }
2142
2143         mr = pd->device->ops.reg_user_mr(pd, start, length, virt_addr,
2144                                          access_flags, NULL);
2145
2146         if (IS_ERR(mr))
2147                 return mr;
2148
2149         mr->device = pd->device;
2150         mr->type = IB_MR_TYPE_USER;
2151         mr->pd = pd;
2152         mr->dm = NULL;
2153         atomic_inc(&pd->usecnt);
2154         mr->iova =  virt_addr;
2155         mr->length = length;
2156
2157         rdma_restrack_new(&mr->res, RDMA_RESTRACK_MR);
2158         rdma_restrack_parent_name(&mr->res, &pd->res);
2159         rdma_restrack_add(&mr->res);
2160
2161         return mr;
2162 }
2163 EXPORT_SYMBOL(ib_reg_user_mr);
2164
2165 int ib_advise_mr(struct ib_pd *pd, enum ib_uverbs_advise_mr_advice advice,
2166                  u32 flags, struct ib_sge *sg_list, u32 num_sge)
2167 {
2168         if (!pd->device->ops.advise_mr)
2169                 return -EOPNOTSUPP;
2170
2171         if (!num_sge)
2172                 return 0;
2173
2174         return pd->device->ops.advise_mr(pd, advice, flags, sg_list, num_sge,
2175                                          NULL);
2176 }
2177 EXPORT_SYMBOL(ib_advise_mr);
2178
2179 int ib_dereg_mr_user(struct ib_mr *mr, struct ib_udata *udata)
2180 {
2181         struct ib_pd *pd = mr->pd;
2182         struct ib_dm *dm = mr->dm;
2183         struct ib_sig_attrs *sig_attrs = mr->sig_attrs;
2184         int ret;
2185
2186         trace_mr_dereg(mr);
2187         rdma_restrack_del(&mr->res);
2188         ret = mr->device->ops.dereg_mr(mr, udata);
2189         if (!ret) {
2190                 atomic_dec(&pd->usecnt);
2191                 if (dm)
2192                         atomic_dec(&dm->usecnt);
2193                 kfree(sig_attrs);
2194         }
2195
2196         return ret;
2197 }
2198 EXPORT_SYMBOL(ib_dereg_mr_user);
2199
2200 /**
2201  * ib_alloc_mr() - Allocates a memory region
2202  * @pd:            protection domain associated with the region
2203  * @mr_type:       memory region type
2204  * @max_num_sg:    maximum sg entries available for registration.
2205  *
2206  * Notes:
2207  * Memory registeration page/sg lists must not exceed max_num_sg.
2208  * For mr_type IB_MR_TYPE_MEM_REG, the total length cannot exceed
2209  * max_num_sg * used_page_size.
2210  *
2211  */
2212 struct ib_mr *ib_alloc_mr(struct ib_pd *pd, enum ib_mr_type mr_type,
2213                           u32 max_num_sg)
2214 {
2215         struct ib_mr *mr;
2216
2217         if (!pd->device->ops.alloc_mr) {
2218                 mr = ERR_PTR(-EOPNOTSUPP);
2219                 goto out;
2220         }
2221
2222         if (mr_type == IB_MR_TYPE_INTEGRITY) {
2223                 WARN_ON_ONCE(1);
2224                 mr = ERR_PTR(-EINVAL);
2225                 goto out;
2226         }
2227
2228         mr = pd->device->ops.alloc_mr(pd, mr_type, max_num_sg);
2229         if (IS_ERR(mr))
2230                 goto out;
2231
2232         mr->device = pd->device;
2233         mr->pd = pd;
2234         mr->dm = NULL;
2235         mr->uobject = NULL;
2236         atomic_inc(&pd->usecnt);
2237         mr->need_inval = false;
2238         mr->type = mr_type;
2239         mr->sig_attrs = NULL;
2240
2241         rdma_restrack_new(&mr->res, RDMA_RESTRACK_MR);
2242         rdma_restrack_parent_name(&mr->res, &pd->res);
2243         rdma_restrack_add(&mr->res);
2244 out:
2245         trace_mr_alloc(pd, mr_type, max_num_sg, mr);
2246         return mr;
2247 }
2248 EXPORT_SYMBOL(ib_alloc_mr);
2249
2250 /**
2251  * ib_alloc_mr_integrity() - Allocates an integrity memory region
2252  * @pd:                      protection domain associated with the region
2253  * @max_num_data_sg:         maximum data sg entries available for registration
2254  * @max_num_meta_sg:         maximum metadata sg entries available for
2255  *                           registration
2256  *
2257  * Notes:
2258  * Memory registration page/sg lists must not exceed max_num_sg,
2259  * also the integrity page/sg lists must not exceed max_num_meta_sg.
2260  *
2261  */
2262 struct ib_mr *ib_alloc_mr_integrity(struct ib_pd *pd,
2263                                     u32 max_num_data_sg,
2264                                     u32 max_num_meta_sg)
2265 {
2266         struct ib_mr *mr;
2267         struct ib_sig_attrs *sig_attrs;
2268
2269         if (!pd->device->ops.alloc_mr_integrity ||
2270             !pd->device->ops.map_mr_sg_pi) {
2271                 mr = ERR_PTR(-EOPNOTSUPP);
2272                 goto out;
2273         }
2274
2275         if (!max_num_meta_sg) {
2276                 mr = ERR_PTR(-EINVAL);
2277                 goto out;
2278         }
2279
2280         sig_attrs = kzalloc(sizeof(struct ib_sig_attrs), GFP_KERNEL);
2281         if (!sig_attrs) {
2282                 mr = ERR_PTR(-ENOMEM);
2283                 goto out;
2284         }
2285
2286         mr = pd->device->ops.alloc_mr_integrity(pd, max_num_data_sg,
2287                                                 max_num_meta_sg);
2288         if (IS_ERR(mr)) {
2289                 kfree(sig_attrs);
2290                 goto out;
2291         }
2292
2293         mr->device = pd->device;
2294         mr->pd = pd;
2295         mr->dm = NULL;
2296         mr->uobject = NULL;
2297         atomic_inc(&pd->usecnt);
2298         mr->need_inval = false;
2299         mr->type = IB_MR_TYPE_INTEGRITY;
2300         mr->sig_attrs = sig_attrs;
2301
2302         rdma_restrack_new(&mr->res, RDMA_RESTRACK_MR);
2303         rdma_restrack_parent_name(&mr->res, &pd->res);
2304         rdma_restrack_add(&mr->res);
2305 out:
2306         trace_mr_integ_alloc(pd, max_num_data_sg, max_num_meta_sg, mr);
2307         return mr;
2308 }
2309 EXPORT_SYMBOL(ib_alloc_mr_integrity);
2310
2311 /* Multicast groups */
2312
2313 static bool is_valid_mcast_lid(struct ib_qp *qp, u16 lid)
2314 {
2315         struct ib_qp_init_attr init_attr = {};
2316         struct ib_qp_attr attr = {};
2317         int num_eth_ports = 0;
2318         unsigned int port;
2319
2320         /* If QP state >= init, it is assigned to a port and we can check this
2321          * port only.
2322          */
2323         if (!ib_query_qp(qp, &attr, IB_QP_STATE | IB_QP_PORT, &init_attr)) {
2324                 if (attr.qp_state >= IB_QPS_INIT) {
2325                         if (rdma_port_get_link_layer(qp->device, attr.port_num) !=
2326                             IB_LINK_LAYER_INFINIBAND)
2327                                 return true;
2328                         goto lid_check;
2329                 }
2330         }
2331
2332         /* Can't get a quick answer, iterate over all ports */
2333         rdma_for_each_port(qp->device, port)
2334                 if (rdma_port_get_link_layer(qp->device, port) !=
2335                     IB_LINK_LAYER_INFINIBAND)
2336                         num_eth_ports++;
2337
2338         /* If we have at lease one Ethernet port, RoCE annex declares that
2339          * multicast LID should be ignored. We can't tell at this step if the
2340          * QP belongs to an IB or Ethernet port.
2341          */
2342         if (num_eth_ports)
2343                 return true;
2344
2345         /* If all the ports are IB, we can check according to IB spec. */
2346 lid_check:
2347         return !(lid < be16_to_cpu(IB_MULTICAST_LID_BASE) ||
2348                  lid == be16_to_cpu(IB_LID_PERMISSIVE));
2349 }
2350
2351 int ib_attach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid)
2352 {
2353         int ret;
2354
2355         if (!qp->device->ops.attach_mcast)
2356                 return -EOPNOTSUPP;
2357
2358         if (!rdma_is_multicast_addr((struct in6_addr *)gid->raw) ||
2359             qp->qp_type != IB_QPT_UD || !is_valid_mcast_lid(qp, lid))
2360                 return -EINVAL;
2361
2362         ret = qp->device->ops.attach_mcast(qp, gid, lid);
2363         if (!ret)
2364                 atomic_inc(&qp->usecnt);
2365         return ret;
2366 }
2367 EXPORT_SYMBOL(ib_attach_mcast);
2368
2369 int ib_detach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid)
2370 {
2371         int ret;
2372
2373         if (!qp->device->ops.detach_mcast)
2374                 return -EOPNOTSUPP;
2375
2376         if (!rdma_is_multicast_addr((struct in6_addr *)gid->raw) ||
2377             qp->qp_type != IB_QPT_UD || !is_valid_mcast_lid(qp, lid))
2378                 return -EINVAL;
2379
2380         ret = qp->device->ops.detach_mcast(qp, gid, lid);
2381         if (!ret)
2382                 atomic_dec(&qp->usecnt);
2383         return ret;
2384 }
2385 EXPORT_SYMBOL(ib_detach_mcast);
2386
2387 /**
2388  * ib_alloc_xrcd_user - Allocates an XRC domain.
2389  * @device: The device on which to allocate the XRC domain.
2390  * @inode: inode to connect XRCD
2391  * @udata: Valid user data or NULL for kernel object
2392  */
2393 struct ib_xrcd *ib_alloc_xrcd_user(struct ib_device *device,
2394                                    struct inode *inode, struct ib_udata *udata)
2395 {
2396         struct ib_xrcd *xrcd;
2397         int ret;
2398
2399         if (!device->ops.alloc_xrcd)
2400                 return ERR_PTR(-EOPNOTSUPP);
2401
2402         xrcd = rdma_zalloc_drv_obj(device, ib_xrcd);
2403         if (!xrcd)
2404                 return ERR_PTR(-ENOMEM);
2405
2406         xrcd->device = device;
2407         xrcd->inode = inode;
2408         atomic_set(&xrcd->usecnt, 0);
2409         init_rwsem(&xrcd->tgt_qps_rwsem);
2410         xa_init(&xrcd->tgt_qps);
2411
2412         ret = device->ops.alloc_xrcd(xrcd, udata);
2413         if (ret)
2414                 goto err;
2415         return xrcd;
2416 err:
2417         kfree(xrcd);
2418         return ERR_PTR(ret);
2419 }
2420 EXPORT_SYMBOL(ib_alloc_xrcd_user);
2421
2422 /**
2423  * ib_dealloc_xrcd_user - Deallocates an XRC domain.
2424  * @xrcd: The XRC domain to deallocate.
2425  * @udata: Valid user data or NULL for kernel object
2426  */
2427 int ib_dealloc_xrcd_user(struct ib_xrcd *xrcd, struct ib_udata *udata)
2428 {
2429         int ret;
2430
2431         if (atomic_read(&xrcd->usecnt))
2432                 return -EBUSY;
2433
2434         WARN_ON(!xa_empty(&xrcd->tgt_qps));
2435         ret = xrcd->device->ops.dealloc_xrcd(xrcd, udata);
2436         if (ret)
2437                 return ret;
2438         kfree(xrcd);
2439         return ret;
2440 }
2441 EXPORT_SYMBOL(ib_dealloc_xrcd_user);
2442
2443 /**
2444  * ib_create_wq - Creates a WQ associated with the specified protection
2445  * domain.
2446  * @pd: The protection domain associated with the WQ.
2447  * @wq_attr: A list of initial attributes required to create the
2448  * WQ. If WQ creation succeeds, then the attributes are updated to
2449  * the actual capabilities of the created WQ.
2450  *
2451  * wq_attr->max_wr and wq_attr->max_sge determine
2452  * the requested size of the WQ, and set to the actual values allocated
2453  * on return.
2454  * If ib_create_wq() succeeds, then max_wr and max_sge will always be
2455  * at least as large as the requested values.
2456  */
2457 struct ib_wq *ib_create_wq(struct ib_pd *pd,
2458                            struct ib_wq_init_attr *wq_attr)
2459 {
2460         struct ib_wq *wq;
2461
2462         if (!pd->device->ops.create_wq)
2463                 return ERR_PTR(-EOPNOTSUPP);
2464
2465         wq = pd->device->ops.create_wq(pd, wq_attr, NULL);
2466         if (!IS_ERR(wq)) {
2467                 wq->event_handler = wq_attr->event_handler;
2468                 wq->wq_context = wq_attr->wq_context;
2469                 wq->wq_type = wq_attr->wq_type;
2470                 wq->cq = wq_attr->cq;
2471                 wq->device = pd->device;
2472                 wq->pd = pd;
2473                 wq->uobject = NULL;
2474                 atomic_inc(&pd->usecnt);
2475                 atomic_inc(&wq_attr->cq->usecnt);
2476                 atomic_set(&wq->usecnt, 0);
2477         }
2478         return wq;
2479 }
2480 EXPORT_SYMBOL(ib_create_wq);
2481
2482 /**
2483  * ib_destroy_wq_user - Destroys the specified user WQ.
2484  * @wq: The WQ to destroy.
2485  * @udata: Valid user data
2486  */
2487 int ib_destroy_wq_user(struct ib_wq *wq, struct ib_udata *udata)
2488 {
2489         struct ib_cq *cq = wq->cq;
2490         struct ib_pd *pd = wq->pd;
2491         int ret;
2492
2493         if (atomic_read(&wq->usecnt))
2494                 return -EBUSY;
2495
2496         ret = wq->device->ops.destroy_wq(wq, udata);
2497         if (ret)
2498                 return ret;
2499
2500         atomic_dec(&pd->usecnt);
2501         atomic_dec(&cq->usecnt);
2502         return ret;
2503 }
2504 EXPORT_SYMBOL(ib_destroy_wq_user);
2505
2506 int ib_check_mr_status(struct ib_mr *mr, u32 check_mask,
2507                        struct ib_mr_status *mr_status)
2508 {
2509         if (!mr->device->ops.check_mr_status)
2510                 return -EOPNOTSUPP;
2511
2512         return mr->device->ops.check_mr_status(mr, check_mask, mr_status);
2513 }
2514 EXPORT_SYMBOL(ib_check_mr_status);
2515
2516 int ib_set_vf_link_state(struct ib_device *device, int vf, u32 port,
2517                          int state)
2518 {
2519         if (!device->ops.set_vf_link_state)
2520                 return -EOPNOTSUPP;
2521
2522         return device->ops.set_vf_link_state(device, vf, port, state);
2523 }
2524 EXPORT_SYMBOL(ib_set_vf_link_state);
2525
2526 int ib_get_vf_config(struct ib_device *device, int vf, u32 port,
2527                      struct ifla_vf_info *info)
2528 {
2529         if (!device->ops.get_vf_config)
2530                 return -EOPNOTSUPP;
2531
2532         return device->ops.get_vf_config(device, vf, port, info);
2533 }
2534 EXPORT_SYMBOL(ib_get_vf_config);
2535
2536 int ib_get_vf_stats(struct ib_device *device, int vf, u32 port,
2537                     struct ifla_vf_stats *stats)
2538 {
2539         if (!device->ops.get_vf_stats)
2540                 return -EOPNOTSUPP;
2541
2542         return device->ops.get_vf_stats(device, vf, port, stats);
2543 }
2544 EXPORT_SYMBOL(ib_get_vf_stats);
2545
2546 int ib_set_vf_guid(struct ib_device *device, int vf, u32 port, u64 guid,
2547                    int type)
2548 {
2549         if (!device->ops.set_vf_guid)
2550                 return -EOPNOTSUPP;
2551
2552         return device->ops.set_vf_guid(device, vf, port, guid, type);
2553 }
2554 EXPORT_SYMBOL(ib_set_vf_guid);
2555
2556 int ib_get_vf_guid(struct ib_device *device, int vf, u32 port,
2557                    struct ifla_vf_guid *node_guid,
2558                    struct ifla_vf_guid *port_guid)
2559 {
2560         if (!device->ops.get_vf_guid)
2561                 return -EOPNOTSUPP;
2562
2563         return device->ops.get_vf_guid(device, vf, port, node_guid, port_guid);
2564 }
2565 EXPORT_SYMBOL(ib_get_vf_guid);
2566 /**
2567  * ib_map_mr_sg_pi() - Map the dma mapped SG lists for PI (protection
2568  *     information) and set an appropriate memory region for registration.
2569  * @mr:             memory region
2570  * @data_sg:        dma mapped scatterlist for data
2571  * @data_sg_nents:  number of entries in data_sg
2572  * @data_sg_offset: offset in bytes into data_sg
2573  * @meta_sg:        dma mapped scatterlist for metadata
2574  * @meta_sg_nents:  number of entries in meta_sg
2575  * @meta_sg_offset: offset in bytes into meta_sg
2576  * @page_size:      page vector desired page size
2577  *
2578  * Constraints:
2579  * - The MR must be allocated with type IB_MR_TYPE_INTEGRITY.
2580  *
2581  * Return: 0 on success.
2582  *
2583  * After this completes successfully, the  memory region
2584  * is ready for registration.
2585  */
2586 int ib_map_mr_sg_pi(struct ib_mr *mr, struct scatterlist *data_sg,
2587                     int data_sg_nents, unsigned int *data_sg_offset,
2588                     struct scatterlist *meta_sg, int meta_sg_nents,
2589                     unsigned int *meta_sg_offset, unsigned int page_size)
2590 {
2591         if (unlikely(!mr->device->ops.map_mr_sg_pi ||
2592                      WARN_ON_ONCE(mr->type != IB_MR_TYPE_INTEGRITY)))
2593                 return -EOPNOTSUPP;
2594
2595         mr->page_size = page_size;
2596
2597         return mr->device->ops.map_mr_sg_pi(mr, data_sg, data_sg_nents,
2598                                             data_sg_offset, meta_sg,
2599                                             meta_sg_nents, meta_sg_offset);
2600 }
2601 EXPORT_SYMBOL(ib_map_mr_sg_pi);
2602
2603 /**
2604  * ib_map_mr_sg() - Map the largest prefix of a dma mapped SG list
2605  *     and set it the memory region.
2606  * @mr:            memory region
2607  * @sg:            dma mapped scatterlist
2608  * @sg_nents:      number of entries in sg
2609  * @sg_offset:     offset in bytes into sg
2610  * @page_size:     page vector desired page size
2611  *
2612  * Constraints:
2613  *
2614  * - The first sg element is allowed to have an offset.
2615  * - Each sg element must either be aligned to page_size or virtually
2616  *   contiguous to the previous element. In case an sg element has a
2617  *   non-contiguous offset, the mapping prefix will not include it.
2618  * - The last sg element is allowed to have length less than page_size.
2619  * - If sg_nents total byte length exceeds the mr max_num_sge * page_size
2620  *   then only max_num_sg entries will be mapped.
2621  * - If the MR was allocated with type IB_MR_TYPE_SG_GAPS, none of these
2622  *   constraints holds and the page_size argument is ignored.
2623  *
2624  * Returns the number of sg elements that were mapped to the memory region.
2625  *
2626  * After this completes successfully, the  memory region
2627  * is ready for registration.
2628  */
2629 int ib_map_mr_sg(struct ib_mr *mr, struct scatterlist *sg, int sg_nents,
2630                  unsigned int *sg_offset, unsigned int page_size)
2631 {
2632         if (unlikely(!mr->device->ops.map_mr_sg))
2633                 return -EOPNOTSUPP;
2634
2635         mr->page_size = page_size;
2636
2637         return mr->device->ops.map_mr_sg(mr, sg, sg_nents, sg_offset);
2638 }
2639 EXPORT_SYMBOL(ib_map_mr_sg);
2640
2641 /**
2642  * ib_sg_to_pages() - Convert the largest prefix of a sg list
2643  *     to a page vector
2644  * @mr:            memory region
2645  * @sgl:           dma mapped scatterlist
2646  * @sg_nents:      number of entries in sg
2647  * @sg_offset_p:   ==== =======================================================
2648  *                 IN   start offset in bytes into sg
2649  *                 OUT  offset in bytes for element n of the sg of the first
2650  *                      byte that has not been processed where n is the return
2651  *                      value of this function.
2652  *                 ==== =======================================================
2653  * @set_page:      driver page assignment function pointer
2654  *
2655  * Core service helper for drivers to convert the largest
2656  * prefix of given sg list to a page vector. The sg list
2657  * prefix converted is the prefix that meet the requirements
2658  * of ib_map_mr_sg.
2659  *
2660  * Returns the number of sg elements that were assigned to
2661  * a page vector.
2662  */
2663 int ib_sg_to_pages(struct ib_mr *mr, struct scatterlist *sgl, int sg_nents,
2664                 unsigned int *sg_offset_p, int (*set_page)(struct ib_mr *, u64))
2665 {
2666         struct scatterlist *sg;
2667         u64 last_end_dma_addr = 0;
2668         unsigned int sg_offset = sg_offset_p ? *sg_offset_p : 0;
2669         unsigned int last_page_off = 0;
2670         u64 page_mask = ~((u64)mr->page_size - 1);
2671         int i, ret;
2672
2673         if (unlikely(sg_nents <= 0 || sg_offset > sg_dma_len(&sgl[0])))
2674                 return -EINVAL;
2675
2676         mr->iova = sg_dma_address(&sgl[0]) + sg_offset;
2677         mr->length = 0;
2678
2679         for_each_sg(sgl, sg, sg_nents, i) {
2680                 u64 dma_addr = sg_dma_address(sg) + sg_offset;
2681                 u64 prev_addr = dma_addr;
2682                 unsigned int dma_len = sg_dma_len(sg) - sg_offset;
2683                 u64 end_dma_addr = dma_addr + dma_len;
2684                 u64 page_addr = dma_addr & page_mask;
2685
2686                 /*
2687                  * For the second and later elements, check whether either the
2688                  * end of element i-1 or the start of element i is not aligned
2689                  * on a page boundary.
2690                  */
2691                 if (i && (last_page_off != 0 || page_addr != dma_addr)) {
2692                         /* Stop mapping if there is a gap. */
2693                         if (last_end_dma_addr != dma_addr)
2694                                 break;
2695
2696                         /*
2697                          * Coalesce this element with the last. If it is small
2698                          * enough just update mr->length. Otherwise start
2699                          * mapping from the next page.
2700                          */
2701                         goto next_page;
2702                 }
2703
2704                 do {
2705                         ret = set_page(mr, page_addr);
2706                         if (unlikely(ret < 0)) {
2707                                 sg_offset = prev_addr - sg_dma_address(sg);
2708                                 mr->length += prev_addr - dma_addr;
2709                                 if (sg_offset_p)
2710                                         *sg_offset_p = sg_offset;
2711                                 return i || sg_offset ? i : ret;
2712                         }
2713                         prev_addr = page_addr;
2714 next_page:
2715                         page_addr += mr->page_size;
2716                 } while (page_addr < end_dma_addr);
2717
2718                 mr->length += dma_len;
2719                 last_end_dma_addr = end_dma_addr;
2720                 last_page_off = end_dma_addr & ~page_mask;
2721
2722                 sg_offset = 0;
2723         }
2724
2725         if (sg_offset_p)
2726                 *sg_offset_p = 0;
2727         return i;
2728 }
2729 EXPORT_SYMBOL(ib_sg_to_pages);
2730
2731 struct ib_drain_cqe {
2732         struct ib_cqe cqe;
2733         struct completion done;
2734 };
2735
2736 static void ib_drain_qp_done(struct ib_cq *cq, struct ib_wc *wc)
2737 {
2738         struct ib_drain_cqe *cqe = container_of(wc->wr_cqe, struct ib_drain_cqe,
2739                                                 cqe);
2740
2741         complete(&cqe->done);
2742 }
2743
2744 /*
2745  * Post a WR and block until its completion is reaped for the SQ.
2746  */
2747 static void __ib_drain_sq(struct ib_qp *qp)
2748 {
2749         struct ib_cq *cq = qp->send_cq;
2750         struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
2751         struct ib_drain_cqe sdrain;
2752         struct ib_rdma_wr swr = {
2753                 .wr = {
2754                         .next = NULL,
2755                         { .wr_cqe       = &sdrain.cqe, },
2756                         .opcode = IB_WR_RDMA_WRITE,
2757                 },
2758         };
2759         int ret;
2760
2761         ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
2762         if (ret) {
2763                 WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
2764                 return;
2765         }
2766
2767         sdrain.cqe.done = ib_drain_qp_done;
2768         init_completion(&sdrain.done);
2769
2770         ret = ib_post_send(qp, &swr.wr, NULL);
2771         if (ret) {
2772                 WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
2773                 return;
2774         }
2775
2776         if (cq->poll_ctx == IB_POLL_DIRECT)
2777                 while (wait_for_completion_timeout(&sdrain.done, HZ / 10) <= 0)
2778                         ib_process_cq_direct(cq, -1);
2779         else
2780                 wait_for_completion(&sdrain.done);
2781 }
2782
2783 /*
2784  * Post a WR and block until its completion is reaped for the RQ.
2785  */
2786 static void __ib_drain_rq(struct ib_qp *qp)
2787 {
2788         struct ib_cq *cq = qp->recv_cq;
2789         struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
2790         struct ib_drain_cqe rdrain;
2791         struct ib_recv_wr rwr = {};
2792         int ret;
2793
2794         ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
2795         if (ret) {
2796                 WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
2797                 return;
2798         }
2799
2800         rwr.wr_cqe = &rdrain.cqe;
2801         rdrain.cqe.done = ib_drain_qp_done;
2802         init_completion(&rdrain.done);
2803
2804         ret = ib_post_recv(qp, &rwr, NULL);
2805         if (ret) {
2806                 WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
2807                 return;
2808         }
2809
2810         if (cq->poll_ctx == IB_POLL_DIRECT)
2811                 while (wait_for_completion_timeout(&rdrain.done, HZ / 10) <= 0)
2812                         ib_process_cq_direct(cq, -1);
2813         else
2814                 wait_for_completion(&rdrain.done);
2815 }
2816
2817 /**
2818  * ib_drain_sq() - Block until all SQ CQEs have been consumed by the
2819  *                 application.
2820  * @qp:            queue pair to drain
2821  *
2822  * If the device has a provider-specific drain function, then
2823  * call that.  Otherwise call the generic drain function
2824  * __ib_drain_sq().
2825  *
2826  * The caller must:
2827  *
2828  * ensure there is room in the CQ and SQ for the drain work request and
2829  * completion.
2830  *
2831  * allocate the CQ using ib_alloc_cq().
2832  *
2833  * ensure that there are no other contexts that are posting WRs concurrently.
2834  * Otherwise the drain is not guaranteed.
2835  */
2836 void ib_drain_sq(struct ib_qp *qp)
2837 {
2838         if (qp->device->ops.drain_sq)
2839                 qp->device->ops.drain_sq(qp);
2840         else
2841                 __ib_drain_sq(qp);
2842         trace_cq_drain_complete(qp->send_cq);
2843 }
2844 EXPORT_SYMBOL(ib_drain_sq);
2845
2846 /**
2847  * ib_drain_rq() - Block until all RQ CQEs have been consumed by the
2848  *                 application.
2849  * @qp:            queue pair to drain
2850  *
2851  * If the device has a provider-specific drain function, then
2852  * call that.  Otherwise call the generic drain function
2853  * __ib_drain_rq().
2854  *
2855  * The caller must:
2856  *
2857  * ensure there is room in the CQ and RQ for the drain work request and
2858  * completion.
2859  *
2860  * allocate the CQ using ib_alloc_cq().
2861  *
2862  * ensure that there are no other contexts that are posting WRs concurrently.
2863  * Otherwise the drain is not guaranteed.
2864  */
2865 void ib_drain_rq(struct ib_qp *qp)
2866 {
2867         if (qp->device->ops.drain_rq)
2868                 qp->device->ops.drain_rq(qp);
2869         else
2870                 __ib_drain_rq(qp);
2871         trace_cq_drain_complete(qp->recv_cq);
2872 }
2873 EXPORT_SYMBOL(ib_drain_rq);
2874
2875 /**
2876  * ib_drain_qp() - Block until all CQEs have been consumed by the
2877  *                 application on both the RQ and SQ.
2878  * @qp:            queue pair to drain
2879  *
2880  * The caller must:
2881  *
2882  * ensure there is room in the CQ(s), SQ, and RQ for drain work requests
2883  * and completions.
2884  *
2885  * allocate the CQs using ib_alloc_cq().
2886  *
2887  * ensure that there are no other contexts that are posting WRs concurrently.
2888  * Otherwise the drain is not guaranteed.
2889  */
2890 void ib_drain_qp(struct ib_qp *qp)
2891 {
2892         ib_drain_sq(qp);
2893         if (!qp->srq)
2894                 ib_drain_rq(qp);
2895 }
2896 EXPORT_SYMBOL(ib_drain_qp);
2897
2898 struct net_device *rdma_alloc_netdev(struct ib_device *device, u32 port_num,
2899                                      enum rdma_netdev_t type, const char *name,
2900                                      unsigned char name_assign_type,
2901                                      void (*setup)(struct net_device *))
2902 {
2903         struct rdma_netdev_alloc_params params;
2904         struct net_device *netdev;
2905         int rc;
2906
2907         if (!device->ops.rdma_netdev_get_params)
2908                 return ERR_PTR(-EOPNOTSUPP);
2909
2910         rc = device->ops.rdma_netdev_get_params(device, port_num, type,
2911                                                 &params);
2912         if (rc)
2913                 return ERR_PTR(rc);
2914
2915         netdev = alloc_netdev_mqs(params.sizeof_priv, name, name_assign_type,
2916                                   setup, params.txqs, params.rxqs);
2917         if (!netdev)
2918                 return ERR_PTR(-ENOMEM);
2919
2920         return netdev;
2921 }
2922 EXPORT_SYMBOL(rdma_alloc_netdev);
2923
2924 int rdma_init_netdev(struct ib_device *device, u32 port_num,
2925                      enum rdma_netdev_t type, const char *name,
2926                      unsigned char name_assign_type,
2927                      void (*setup)(struct net_device *),
2928                      struct net_device *netdev)
2929 {
2930         struct rdma_netdev_alloc_params params;
2931         int rc;
2932
2933         if (!device->ops.rdma_netdev_get_params)
2934                 return -EOPNOTSUPP;
2935
2936         rc = device->ops.rdma_netdev_get_params(device, port_num, type,
2937                                                 &params);
2938         if (rc)
2939                 return rc;
2940
2941         return params.initialize_rdma_netdev(device, port_num,
2942                                              netdev, params.param);
2943 }
2944 EXPORT_SYMBOL(rdma_init_netdev);
2945
2946 void __rdma_block_iter_start(struct ib_block_iter *biter,
2947                              struct scatterlist *sglist, unsigned int nents,
2948                              unsigned long pgsz)
2949 {
2950         memset(biter, 0, sizeof(struct ib_block_iter));
2951         biter->__sg = sglist;
2952         biter->__sg_nents = nents;
2953
2954         /* Driver provides best block size to use */
2955         biter->__pg_bit = __fls(pgsz);
2956 }
2957 EXPORT_SYMBOL(__rdma_block_iter_start);
2958
2959 bool __rdma_block_iter_next(struct ib_block_iter *biter)
2960 {
2961         unsigned int block_offset;
2962         unsigned int sg_delta;
2963
2964         if (!biter->__sg_nents || !biter->__sg)
2965                 return false;
2966
2967         biter->__dma_addr = sg_dma_address(biter->__sg) + biter->__sg_advance;
2968         block_offset = biter->__dma_addr & (BIT_ULL(biter->__pg_bit) - 1);
2969         sg_delta = BIT_ULL(biter->__pg_bit) - block_offset;
2970
2971         if (sg_dma_len(biter->__sg) - biter->__sg_advance > sg_delta) {
2972                 biter->__sg_advance += sg_delta;
2973         } else {
2974                 biter->__sg_advance = 0;
2975                 biter->__sg = sg_next(biter->__sg);
2976                 biter->__sg_nents--;
2977         }
2978
2979         return true;
2980 }
2981 EXPORT_SYMBOL(__rdma_block_iter_next);
2982
2983 /**
2984  * rdma_alloc_hw_stats_struct - Helper function to allocate dynamic struct
2985  *   for the drivers.
2986  * @descs: array of static descriptors
2987  * @num_counters: number of elements in array
2988  * @lifespan: milliseconds between updates
2989  */
2990 struct rdma_hw_stats *rdma_alloc_hw_stats_struct(
2991         const struct rdma_stat_desc *descs, int num_counters,
2992         unsigned long lifespan)
2993 {
2994         struct rdma_hw_stats *stats;
2995
2996         stats = kzalloc(struct_size(stats, value, num_counters), GFP_KERNEL);
2997         if (!stats)
2998                 return NULL;
2999
3000         stats->is_disabled = kcalloc(BITS_TO_LONGS(num_counters),
3001                                      sizeof(*stats->is_disabled), GFP_KERNEL);
3002         if (!stats->is_disabled)
3003                 goto err;
3004
3005         stats->descs = descs;
3006         stats->num_counters = num_counters;
3007         stats->lifespan = msecs_to_jiffies(lifespan);
3008         mutex_init(&stats->lock);
3009
3010         return stats;
3011
3012 err:
3013         kfree(stats);
3014         return NULL;
3015 }
3016 EXPORT_SYMBOL(rdma_alloc_hw_stats_struct);
3017
3018 /**
3019  * rdma_free_hw_stats_struct - Helper function to release rdma_hw_stats
3020  * @stats: statistics to release
3021  */
3022 void rdma_free_hw_stats_struct(struct rdma_hw_stats *stats)
3023 {
3024         if (!stats)
3025                 return;
3026
3027         kfree(stats->is_disabled);
3028         kfree(stats);
3029 }
3030 EXPORT_SYMBOL(rdma_free_hw_stats_struct);