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

   1 /*-
   2  * SPDX-License-Identifier: BSD-2-Clause OR GPL-2.0
   3  *
   4  * Copyright (c) 2004 Mellanox Technologies Ltd.  All rights reserved.
   5  * Copyright (c) 2004 Infinicon Corporation.  All rights reserved.
   6  * Copyright (c) 2004 Intel Corporation.  All rights reserved.
   7  * Copyright (c) 2004 Topspin Corporation.  All rights reserved.
   8  * Copyright (c) 2004 Voltaire Corporation.  All rights reserved.
   9  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
  10  * Copyright (c) 2005, 2006 Cisco Systems.  All rights reserved.
  11  *
  12  * This software is available to you under a choice of one of two
  13  * licenses.  You may choose to be licensed under the terms of the GNU
  14  * General Public License (GPL) Version 2, available from the file
  15  * COPYING in the main directory of this source tree, or the
  16  * OpenIB.org BSD license below:
  17  *
  18  *     Redistribution and use in source and binary forms, with or
  19  *     without modification, are permitted provided that the following
  20  *     conditions are met:
  21  *
  22  *      - Redistributions of source code must retain the above
  23  *        copyright notice, this list of conditions and the following
  24  *        disclaimer.
  25  *
  26  *      - Redistributions in binary form must reproduce the above
  27  *        copyright notice, this list of conditions and the following
  28  *        disclaimer in the documentation and/or other materials
  29  *        provided with the distribution.
  30  *
  31  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  32  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  33  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  34  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  35  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  36  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  37  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  38  * SOFTWARE.
  39  */
  40
  41 #include <sys/cdefs.h>
  42 __FBSDID("$FreeBSD$");
  43
  44 #include <linux/errno.h>
  45 #include <linux/err.h>
  46 #include <linux/string.h>
  47 #include <linux/slab.h>
  48 #include <linux/in.h>
  49 #include <linux/in6.h>
  50
  51 #include <rdma/ib_verbs.h>
  52 #include <rdma/ib_cache.h>
  53 #include <rdma/ib_addr.h>
  54
  55 #include <netinet/ip.h>
  56 #include <netinet/ip6.h>
  57
  58 #include <machine/in_cksum.h>
  59
  60 #include "core_priv.h"
  61
  62 static const char * const ib_events[] = {
  63         [IB_EVENT_CQ_ERR]               = "CQ error",
  64         [IB_EVENT_QP_FATAL]             = "QP fatal error",
  65         [IB_EVENT_QP_REQ_ERR]           = "QP request error",
  66         [IB_EVENT_QP_ACCESS_ERR]        = "QP access error",
  67         [IB_EVENT_COMM_EST]             = "communication established",
  68         [IB_EVENT_SQ_DRAINED]           = "send queue drained",
  69         [IB_EVENT_PATH_MIG]             = "path migration successful",
  70         [IB_EVENT_PATH_MIG_ERR]         = "path migration error",
  71         [IB_EVENT_DEVICE_FATAL]         = "device fatal error",
  72         [IB_EVENT_PORT_ACTIVE]          = "port active",
  73         [IB_EVENT_PORT_ERR]             = "port error",
  74         [IB_EVENT_LID_CHANGE]           = "LID change",
  75         [IB_EVENT_PKEY_CHANGE]          = "P_key change",
  76         [IB_EVENT_SM_CHANGE]            = "SM change",
  77         [IB_EVENT_SRQ_ERR]              = "SRQ error",
  78         [IB_EVENT_SRQ_LIMIT_REACHED]    = "SRQ limit reached",
  79         [IB_EVENT_QP_LAST_WQE_REACHED]  = "last WQE reached",
  80         [IB_EVENT_CLIENT_REREGISTER]    = "client reregister",
  81         [IB_EVENT_GID_CHANGE]           = "GID changed",
  82 };
  83
  84 const char *__attribute_const__ ib_event_msg(enum ib_event_type event)
  85 {
  86         size_t index = event;
  87
  88         return (index < ARRAY_SIZE(ib_events) && ib_events[index]) ?
  89                         ib_events[index] : "unrecognized event";
  90 }
  91 EXPORT_SYMBOL(ib_event_msg);
  92
  93 static const char * const wc_statuses[] = {
  94         [IB_WC_SUCCESS]                 = "success",
  95         [IB_WC_LOC_LEN_ERR]             = "local length error",
  96         [IB_WC_LOC_QP_OP_ERR]           = "local QP operation error",
  97         [IB_WC_LOC_EEC_OP_ERR]          = "local EE context operation error",
  98         [IB_WC_LOC_PROT_ERR]            = "local protection error",
  99         [IB_WC_WR_FLUSH_ERR]            = "WR flushed",
 100         [IB_WC_MW_BIND_ERR]             = "memory management operation error",
 101         [IB_WC_BAD_RESP_ERR]            = "bad response error",
 102         [IB_WC_LOC_ACCESS_ERR]          = "local access error",
 103         [IB_WC_REM_INV_REQ_ERR]         = "invalid request error",
 104         [IB_WC_REM_ACCESS_ERR]          = "remote access error",
 105         [IB_WC_REM_OP_ERR]              = "remote operation error",
 106         [IB_WC_RETRY_EXC_ERR]           = "transport retry counter exceeded",
 107         [IB_WC_RNR_RETRY_EXC_ERR]       = "RNR retry counter exceeded",
 108         [IB_WC_LOC_RDD_VIOL_ERR]        = "local RDD violation error",
 109         [IB_WC_REM_INV_RD_REQ_ERR]      = "remote invalid RD request",
 110         [IB_WC_REM_ABORT_ERR]           = "operation aborted",
 111         [IB_WC_INV_EECN_ERR]            = "invalid EE context number",
 112         [IB_WC_INV_EEC_STATE_ERR]       = "invalid EE context state",
 113         [IB_WC_FATAL_ERR]               = "fatal error",
 114         [IB_WC_RESP_TIMEOUT_ERR]        = "response timeout error",
 115         [IB_WC_GENERAL_ERR]             = "general error",
 116 };
 117
 118 const char *__attribute_const__ ib_wc_status_msg(enum ib_wc_status status)
 119 {
 120         size_t index = status;
 121
 122         return (index < ARRAY_SIZE(wc_statuses) && wc_statuses[index]) ?
 123                         wc_statuses[index] : "unrecognized status";
 124 }
 125 EXPORT_SYMBOL(ib_wc_status_msg);
 126
 127 __attribute_const__ int ib_rate_to_mult(enum ib_rate rate)
 128 {
 129         switch (rate) {
 130         case IB_RATE_2_5_GBPS: return  1;
 131         case IB_RATE_5_GBPS:   return  2;
 132         case IB_RATE_10_GBPS:  return  4;
 133         case IB_RATE_20_GBPS:  return  8;
 134         case IB_RATE_30_GBPS:  return 12;
 135         case IB_RATE_40_GBPS:  return 16;
 136         case IB_RATE_60_GBPS:  return 24;
 137         case IB_RATE_80_GBPS:  return 32;
 138         case IB_RATE_120_GBPS: return 48;
 139         case IB_RATE_28_GBPS:  return  11;
 140         case IB_RATE_50_GBPS:  return  20;
 141         case IB_RATE_400_GBPS: return 160;
 142         case IB_RATE_600_GBPS: return 240;
 143         default:               return -1;
 144         }
 145 }
 146 EXPORT_SYMBOL(ib_rate_to_mult);
 147
 148 __attribute_const__ enum ib_rate mult_to_ib_rate(int mult)
 149 {
 150         switch (mult) {
 151         case 1:  return IB_RATE_2_5_GBPS;
 152         case 2:  return IB_RATE_5_GBPS;
 153         case 4:  return IB_RATE_10_GBPS;
 154         case 8:  return IB_RATE_20_GBPS;
 155         case 12: return IB_RATE_30_GBPS;
 156         case 16: return IB_RATE_40_GBPS;
 157         case 24: return IB_RATE_60_GBPS;
 158         case 32: return IB_RATE_80_GBPS;
 159         case 48: return IB_RATE_120_GBPS;
 160         case 6:   return IB_RATE_14_GBPS;
 161         case 22:  return IB_RATE_56_GBPS;
 162         case 45:  return IB_RATE_112_GBPS;
 163         case 67:  return IB_RATE_168_GBPS;
 164         case 10:  return IB_RATE_25_GBPS;
 165         case 40:  return IB_RATE_100_GBPS;
 166         case 80:  return IB_RATE_200_GBPS;
 167         case 120: return IB_RATE_300_GBPS;
 168         case 11:  return IB_RATE_28_GBPS;
 169         case 20:  return IB_RATE_50_GBPS;
 170         case 160: return IB_RATE_400_GBPS;
 171         case 240: return IB_RATE_600_GBPS;
 172         default: return IB_RATE_PORT_CURRENT;
 173         }
 174 }
 175 EXPORT_SYMBOL(mult_to_ib_rate);
 176
 177 __attribute_const__ int ib_rate_to_mbps(enum ib_rate rate)
 178 {
 179         switch (rate) {
 180         case IB_RATE_2_5_GBPS: return 2500;
 181         case IB_RATE_5_GBPS:   return 5000;
 182         case IB_RATE_10_GBPS:  return 10000;
 183         case IB_RATE_20_GBPS:  return 20000;
 184         case IB_RATE_30_GBPS:  return 30000;
 185         case IB_RATE_40_GBPS:  return 40000;
 186         case IB_RATE_60_GBPS:  return 60000;
 187         case IB_RATE_80_GBPS:  return 80000;
 188         case IB_RATE_120_GBPS: return 120000;
 189         case IB_RATE_14_GBPS:  return 14062;
 190         case IB_RATE_56_GBPS:  return 56250;
 191         case IB_RATE_112_GBPS: return 112500;
 192         case IB_RATE_168_GBPS: return 168750;
 193         case IB_RATE_25_GBPS:  return 25781;
 194         case IB_RATE_100_GBPS: return 103125;
 195         case IB_RATE_200_GBPS: return 206250;
 196         case IB_RATE_300_GBPS: return 309375;
 197         case IB_RATE_28_GBPS:  return 28125;
 198         case IB_RATE_50_GBPS:  return 53125;
 199         case IB_RATE_400_GBPS: return 425000;
 200         case IB_RATE_600_GBPS: return 637500;
 201         default:               return -1;
 202         }
 203 }
 204 EXPORT_SYMBOL(ib_rate_to_mbps);
 205
 206 __attribute_const__ enum rdma_transport_type
 207 rdma_node_get_transport(enum rdma_node_type node_type)
 208 {
 209         switch (node_type) {
 210         case RDMA_NODE_IB_CA:
 211         case RDMA_NODE_IB_SWITCH:
 212         case RDMA_NODE_IB_ROUTER:
 213                 return RDMA_TRANSPORT_IB;
 214         case RDMA_NODE_RNIC:
 215                 return RDMA_TRANSPORT_IWARP;
 216         case RDMA_NODE_USNIC:
 217                 return RDMA_TRANSPORT_USNIC;
 218         case RDMA_NODE_USNIC_UDP:
 219                 return RDMA_TRANSPORT_USNIC_UDP;
 220         default:
 221                 BUG();
 222                 return 0;
 223         }
 224 }
 225 EXPORT_SYMBOL(rdma_node_get_transport);
 226
 227 enum rdma_link_layer rdma_port_get_link_layer(struct ib_device *device, u8 port_num)
 228 {
 229         if (device->get_link_layer)
 230                 return device->get_link_layer(device, port_num);
 231
 232         switch (rdma_node_get_transport(device->node_type)) {
 233         case RDMA_TRANSPORT_IB:
 234                 return IB_LINK_LAYER_INFINIBAND;
 235         case RDMA_TRANSPORT_IWARP:
 236         case RDMA_TRANSPORT_USNIC:
 237         case RDMA_TRANSPORT_USNIC_UDP:
 238                 return IB_LINK_LAYER_ETHERNET;
 239         default:
 240                 return IB_LINK_LAYER_UNSPECIFIED;
 241         }
 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  *
 251  * A protection domain object provides an association between QPs, shared
 252  * receive queues, address handles, memory regions, and memory windows.
 253  *
 254  * Every PD has a local_dma_lkey which can be used as the lkey value for local
 255  * memory operations.
 256  */
 257 struct ib_pd *__ib_alloc_pd(struct ib_device *device, unsigned int flags,
 258                 const char *caller)
 259 {
 260         struct ib_pd *pd;
 261         int mr_access_flags = 0;
 262
 263         pd = device->alloc_pd(device, NULL, NULL);
 264         if (IS_ERR(pd))
 265                 return pd;
 266
 267         pd->device = device;
 268         pd->uobject = NULL;
 269         pd->__internal_mr = NULL;
 270         atomic_set(&pd->usecnt, 0);
 271         pd->flags = flags;
 272
 273         if (device->attrs.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY)
 274                 pd->local_dma_lkey = device->local_dma_lkey;
 275         else
 276                 mr_access_flags |= IB_ACCESS_LOCAL_WRITE;
 277
 278         if (flags & IB_PD_UNSAFE_GLOBAL_RKEY) {
 279                 pr_warn("%s: enabling unsafe global rkey\n", caller);
 280                 mr_access_flags |= IB_ACCESS_REMOTE_READ | IB_ACCESS_REMOTE_WRITE;
 281         }
 282
 283         if (mr_access_flags) {
 284                 struct ib_mr *mr;
 285
 286                 mr = pd->device->get_dma_mr(pd, mr_access_flags);
 287                 if (IS_ERR(mr)) {
 288                         ib_dealloc_pd(pd);
 289                         return ERR_CAST(mr);
 290                 }
 291
 292                 mr->device      = pd->device;
 293                 mr->pd          = pd;
 294                 mr->uobject     = NULL;
 295                 mr->need_inval  = false;
 296
 297                 pd->__internal_mr = mr;
 298
 299                 if (!(device->attrs.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY))
 300                         pd->local_dma_lkey = pd->__internal_mr->lkey;
 301
 302                 if (flags & IB_PD_UNSAFE_GLOBAL_RKEY)
 303                         pd->unsafe_global_rkey = pd->__internal_mr->rkey;
 304         }
 305
 306         return pd;
 307 }
 308 EXPORT_SYMBOL(__ib_alloc_pd);
 309
 310 /**
 311  * ib_dealloc_pd - Deallocates a protection domain.
 312  * @pd: The protection domain to deallocate.
 313  *
 314  * It is an error to call this function while any resources in the pd still
 315  * exist.  The caller is responsible to synchronously destroy them and
 316  * guarantee no new allocations will happen.
 317  */
 318 void ib_dealloc_pd(struct ib_pd *pd)
 319 {
 320         int ret;
 321
 322         if (pd->__internal_mr) {
 323                 ret = pd->device->dereg_mr(pd->__internal_mr);
 324                 WARN_ON(ret);
 325                 pd->__internal_mr = NULL;
 326         }
 327
 328         /* uverbs manipulates usecnt with proper locking, while the kabi
 329            requires the caller to guarantee we can't race here. */
 330         WARN_ON(atomic_read(&pd->usecnt));
 331
 332         /* Making delalloc_pd a void return is a WIP, no driver should return
 333            an error here. */
 334         ret = pd->device->dealloc_pd(pd);
 335         WARN_ONCE(ret, "Infiniband HW driver failed dealloc_pd");
 336 }
 337 EXPORT_SYMBOL(ib_dealloc_pd);
 338
 339 /* Address handles */
 340
 341 struct ib_ah *ib_create_ah(struct ib_pd *pd, struct ib_ah_attr *ah_attr)
 342 {
 343         struct ib_ah *ah;
 344
 345         ah = pd->device->create_ah(pd, ah_attr, NULL);
 346
 347         if (!IS_ERR(ah)) {
 348                 ah->device  = pd->device;
 349                 ah->pd      = pd;
 350                 ah->uobject = NULL;
 351                 atomic_inc(&pd->usecnt);
 352         }
 353
 354         return ah;
 355 }
 356 EXPORT_SYMBOL(ib_create_ah);
 357
 358 static int ib_get_header_version(const union rdma_network_hdr *hdr)
 359 {
 360         const struct ip *ip4h = (const struct ip *)&hdr->roce4grh;
 361         struct ip ip4h_checked;
 362         const struct ip6_hdr *ip6h = (const struct ip6_hdr *)&hdr->ibgrh;
 363
 364         /* If it's IPv6, the version must be 6, otherwise, the first
 365          * 20 bytes (before the IPv4 header) are garbled.
 366          */
 367         if ((ip6h->ip6_vfc & IPV6_VERSION_MASK) != IPV6_VERSION)
 368                 return (ip4h->ip_v == 4) ? 4 : 0;
 369         /* version may be 6 or 4 because the first 20 bytes could be garbled */
 370
 371         /* RoCE v2 requires no options, thus header length
 372          * must be 5 words
 373          */
 374         if (ip4h->ip_hl != 5)
 375                 return 6;
 376
 377         /* Verify checksum.
 378          * We can't write on scattered buffers so we need to copy to
 379          * temp buffer.
 380          */
 381         memcpy(&ip4h_checked, ip4h, sizeof(ip4h_checked));
 382         ip4h_checked.ip_sum = 0;
 383 #if defined(INET) || defined(INET6)
 384         ip4h_checked.ip_sum = in_cksum_hdr(&ip4h_checked);
 385 #endif
 386         /* if IPv4 header checksum is OK, believe it */
 387         if (ip4h->ip_sum == ip4h_checked.ip_sum)
 388                 return 4;
 389         return 6;
 390 }
 391
 392 static enum rdma_network_type ib_get_net_type_by_grh(struct ib_device *device,
 393                                                      u8 port_num,
 394                                                      const struct ib_grh *grh)
 395 {
 396         int grh_version;
 397
 398         if (rdma_protocol_ib(device, port_num))
 399                 return RDMA_NETWORK_IB;
 400
 401         grh_version = ib_get_header_version((const union rdma_network_hdr *)grh);
 402
 403         if (grh_version == 4)
 404                 return RDMA_NETWORK_IPV4;
 405
 406         if (grh->next_hdr == IPPROTO_UDP)
 407                 return RDMA_NETWORK_IPV6;
 408
 409         return RDMA_NETWORK_ROCE_V1;
 410 }
 411
 412 struct find_gid_index_context {
 413         u16 vlan_id;
 414         enum ib_gid_type gid_type;
 415 };
 416
 417 static bool find_gid_index(const union ib_gid *gid,
 418                            const struct ib_gid_attr *gid_attr,
 419                            void *context)
 420 {
 421         struct find_gid_index_context *ctx =
 422                 (struct find_gid_index_context *)context;
 423
 424         if (ctx->gid_type != gid_attr->gid_type)
 425                 return false;
 426         if (rdma_vlan_dev_vlan_id(gid_attr->ndev) != ctx->vlan_id)
 427                 return false;
 428         return true;
 429 }
 430
 431 static int get_sgid_index_from_eth(struct ib_device *device, u8 port_num,
 432                                    u16 vlan_id, const union ib_gid *sgid,
 433                                    enum ib_gid_type gid_type,
 434                                    u16 *gid_index)
 435 {
 436         struct find_gid_index_context context = {.vlan_id = vlan_id,
 437                                                  .gid_type = gid_type};
 438
 439         return ib_find_gid_by_filter(device, sgid, port_num, find_gid_index,
 440                                      &context, gid_index);
 441 }
 442
 443 static int get_gids_from_rdma_hdr(const union rdma_network_hdr *hdr,
 444                                   enum rdma_network_type net_type,
 445                                   union ib_gid *sgid, union ib_gid *dgid)
 446 {
 447         struct sockaddr_in  src_in;
 448         struct sockaddr_in  dst_in;
 449         __be32 src_saddr, dst_saddr;
 450
 451         if (!sgid || !dgid)
 452                 return -EINVAL;
 453
 454         if (net_type == RDMA_NETWORK_IPV4) {
 455                 memcpy(&src_in.sin_addr.s_addr,
 456                        &hdr->roce4grh.ip_src, 4);
 457                 memcpy(&dst_in.sin_addr.s_addr,
 458                        &hdr->roce4grh.ip_dst, 4);
 459                 src_saddr = src_in.sin_addr.s_addr;
 460                 dst_saddr = dst_in.sin_addr.s_addr;
 461                 ipv6_addr_set_v4mapped(src_saddr,
 462                                        (struct in6_addr *)sgid);
 463                 ipv6_addr_set_v4mapped(dst_saddr,
 464                                        (struct in6_addr *)dgid);
 465                 return 0;
 466         } else if (net_type == RDMA_NETWORK_IPV6 ||
 467                    net_type == RDMA_NETWORK_IB) {
 468                 *dgid = hdr->ibgrh.dgid;
 469                 *sgid = hdr->ibgrh.sgid;
 470                 return 0;
 471         } else {
 472                 return -EINVAL;
 473         }
 474 }
 475
 476 int ib_init_ah_from_wc(struct ib_device *device, u8 port_num,
 477                        const struct ib_wc *wc, const struct ib_grh *grh,
 478                        struct ib_ah_attr *ah_attr)
 479 {
 480         u32 flow_class;
 481         u16 gid_index = 0;
 482         int ret;
 483         enum rdma_network_type net_type = RDMA_NETWORK_IB;
 484         enum ib_gid_type gid_type = IB_GID_TYPE_IB;
 485         int hoplimit = 0xff;
 486         union ib_gid dgid;
 487         union ib_gid sgid;
 488
 489         memset(ah_attr, 0, sizeof *ah_attr);
 490         if (rdma_cap_eth_ah(device, port_num)) {
 491                 if (wc->wc_flags & IB_WC_WITH_NETWORK_HDR_TYPE)
 492                         net_type = wc->network_hdr_type;
 493                 else
 494                         net_type = ib_get_net_type_by_grh(device, port_num, grh);
 495                 gid_type = ib_network_to_gid_type(net_type);
 496         }
 497         ret = get_gids_from_rdma_hdr((const union rdma_network_hdr *)grh, net_type,
 498                                      &sgid, &dgid);
 499         if (ret)
 500                 return ret;
 501
 502         if (rdma_protocol_roce(device, port_num)) {
 503                 struct ib_gid_attr dgid_attr;
 504                 const u16 vlan_id = (wc->wc_flags & IB_WC_WITH_VLAN) ?
 505                                 wc->vlan_id : 0xffff;
 506
 507                 if (!(wc->wc_flags & IB_WC_GRH))
 508                         return -EPROTOTYPE;
 509
 510                 ret = get_sgid_index_from_eth(device, port_num, vlan_id,
 511                                               &dgid, gid_type, &gid_index);
 512                 if (ret)
 513                         return ret;
 514
 515                 ret = ib_get_cached_gid(device, port_num, gid_index, &dgid, &dgid_attr);
 516                 if (ret)
 517                         return ret;
 518
 519                 if (dgid_attr.ndev == NULL)
 520                         return -ENODEV;
 521
 522                 ret = rdma_addr_find_l2_eth_by_grh(&dgid, &sgid, ah_attr->dmac,
 523                     dgid_attr.ndev, &hoplimit);
 524
 525                 dev_put(dgid_attr.ndev);
 526                 if (ret)
 527                         return ret;
 528         }
 529
 530         ah_attr->dlid = wc->slid;
 531         ah_attr->sl = wc->sl;
 532         ah_attr->src_path_bits = wc->dlid_path_bits;
 533         ah_attr->port_num = port_num;
 534
 535         if (wc->wc_flags & IB_WC_GRH) {
 536                 ah_attr->ah_flags = IB_AH_GRH;
 537                 ah_attr->grh.dgid = sgid;
 538
 539                 if (!rdma_cap_eth_ah(device, port_num)) {
 540                         if (dgid.global.interface_id != cpu_to_be64(IB_SA_WELL_KNOWN_GUID)) {
 541                                 ret = ib_find_cached_gid_by_port(device, &dgid,
 542                                                                  IB_GID_TYPE_IB,
 543                                                                  port_num, NULL,
 544                                                                  &gid_index);
 545                                 if (ret)
 546                                         return ret;
 547                         }
 548                 }
 549
 550                 ah_attr->grh.sgid_index = (u8) gid_index;
 551                 flow_class = be32_to_cpu(grh->version_tclass_flow);
 552                 ah_attr->grh.flow_label = flow_class & 0xFFFFF;
 553                 ah_attr->grh.hop_limit = hoplimit;
 554                 ah_attr->grh.traffic_class = (flow_class >> 20) & 0xFF;
 555         }
 556         return 0;
 557 }
 558 EXPORT_SYMBOL(ib_init_ah_from_wc);
 559
 560 struct ib_ah *ib_create_ah_from_wc(struct ib_pd *pd, const struct ib_wc *wc,
 561                                    const struct ib_grh *grh, u8 port_num)
 562 {
 563         struct ib_ah_attr ah_attr;
 564         int ret;
 565
 566         ret = ib_init_ah_from_wc(pd->device, port_num, wc, grh, &ah_attr);
 567         if (ret)
 568                 return ERR_PTR(ret);
 569
 570         return ib_create_ah(pd, &ah_attr);
 571 }
 572 EXPORT_SYMBOL(ib_create_ah_from_wc);
 573
 574 int ib_modify_ah(struct ib_ah *ah, struct ib_ah_attr *ah_attr)
 575 {
 576         return ah->device->modify_ah ?
 577                 ah->device->modify_ah(ah, ah_attr) :
 578                 -ENOSYS;
 579 }
 580 EXPORT_SYMBOL(ib_modify_ah);
 581
 582 int ib_query_ah(struct ib_ah *ah, struct ib_ah_attr *ah_attr)
 583 {
 584         return ah->device->query_ah ?
 585                 ah->device->query_ah(ah, ah_attr) :
 586                 -ENOSYS;
 587 }
 588 EXPORT_SYMBOL(ib_query_ah);
 589
 590 int ib_destroy_ah(struct ib_ah *ah)
 591 {
 592         struct ib_pd *pd;
 593         int ret;
 594
 595         pd = ah->pd;
 596         ret = ah->device->destroy_ah(ah);
 597         if (!ret)
 598                 atomic_dec(&pd->usecnt);
 599
 600         return ret;
 601 }
 602 EXPORT_SYMBOL(ib_destroy_ah);
 603
 604 /* Shared receive queues */
 605
 606 struct ib_srq *ib_create_srq(struct ib_pd *pd,
 607                              struct ib_srq_init_attr *srq_init_attr)
 608 {
 609         struct ib_srq *srq;
 610
 611         if (!pd->device->create_srq)
 612                 return ERR_PTR(-ENOSYS);
 613
 614         srq = pd->device->create_srq(pd, srq_init_attr, NULL);
 615
 616         if (!IS_ERR(srq)) {
 617                 srq->device        = pd->device;
 618                 srq->pd            = pd;
 619                 srq->uobject       = NULL;
 620                 srq->event_handler = srq_init_attr->event_handler;
 621                 srq->srq_context   = srq_init_attr->srq_context;
 622                 srq->srq_type      = srq_init_attr->srq_type;
 623                 if (srq->srq_type == IB_SRQT_XRC) {
 624                         srq->ext.xrc.xrcd = srq_init_attr->ext.xrc.xrcd;
 625                         srq->ext.xrc.cq   = srq_init_attr->ext.xrc.cq;
 626                         atomic_inc(&srq->ext.xrc.xrcd->usecnt);
 627                         atomic_inc(&srq->ext.xrc.cq->usecnt);
 628                 }
 629                 atomic_inc(&pd->usecnt);
 630                 atomic_set(&srq->usecnt, 0);
 631         }
 632
 633         return srq;
 634 }
 635 EXPORT_SYMBOL(ib_create_srq);
 636
 637 int ib_modify_srq(struct ib_srq *srq,
 638                   struct ib_srq_attr *srq_attr,
 639                   enum ib_srq_attr_mask srq_attr_mask)
 640 {
 641         return srq->device->modify_srq ?
 642                 srq->device->modify_srq(srq, srq_attr, srq_attr_mask, NULL) :
 643                 -ENOSYS;
 644 }
 645 EXPORT_SYMBOL(ib_modify_srq);
 646
 647 int ib_query_srq(struct ib_srq *srq,
 648                  struct ib_srq_attr *srq_attr)
 649 {
 650         return srq->device->query_srq ?
 651                 srq->device->query_srq(srq, srq_attr) : -ENOSYS;
 652 }
 653 EXPORT_SYMBOL(ib_query_srq);
 654
 655 int ib_destroy_srq(struct ib_srq *srq)
 656 {
 657         struct ib_pd *pd;
 658         enum ib_srq_type srq_type;
 659         struct ib_xrcd *uninitialized_var(xrcd);
 660         struct ib_cq *uninitialized_var(cq);
 661         int ret;
 662
 663         if (atomic_read(&srq->usecnt))
 664                 return -EBUSY;
 665
 666         pd = srq->pd;
 667         srq_type = srq->srq_type;
 668         if (srq_type == IB_SRQT_XRC) {
 669                 xrcd = srq->ext.xrc.xrcd;
 670                 cq = srq->ext.xrc.cq;
 671         }
 672
 673         ret = srq->device->destroy_srq(srq);
 674         if (!ret) {
 675                 atomic_dec(&pd->usecnt);
 676                 if (srq_type == IB_SRQT_XRC) {
 677                         atomic_dec(&xrcd->usecnt);
 678                         atomic_dec(&cq->usecnt);
 679                 }
 680         }
 681
 682         return ret;
 683 }
 684 EXPORT_SYMBOL(ib_destroy_srq);
 685
 686 /* Queue pairs */
 687
 688 static void __ib_shared_qp_event_handler(struct ib_event *event, void *context)
 689 {
 690         struct ib_qp *qp = context;
 691         unsigned long flags;
 692
 693         spin_lock_irqsave(&qp->device->event_handler_lock, flags);
 694         list_for_each_entry(event->element.qp, &qp->open_list, open_list)
 695                 if (event->element.qp->event_handler)
 696                         event->element.qp->event_handler(event, event->element.qp->qp_context);
 697         spin_unlock_irqrestore(&qp->device->event_handler_lock, flags);
 698 }
 699
 700 static void __ib_insert_xrcd_qp(struct ib_xrcd *xrcd, struct ib_qp *qp)
 701 {
 702         mutex_lock(&xrcd->tgt_qp_mutex);
 703         list_add(&qp->xrcd_list, &xrcd->tgt_qp_list);
 704         mutex_unlock(&xrcd->tgt_qp_mutex);
 705 }
 706
 707 static struct ib_qp *__ib_open_qp(struct ib_qp *real_qp,
 708                                   void (*event_handler)(struct ib_event *, void *),
 709                                   void *qp_context)
 710 {
 711         struct ib_qp *qp;
 712         unsigned long flags;
 713
 714         qp = kzalloc(sizeof *qp, GFP_KERNEL);
 715         if (!qp)
 716                 return ERR_PTR(-ENOMEM);
 717
 718         qp->real_qp = real_qp;
 719         atomic_inc(&real_qp->usecnt);
 720         qp->device = real_qp->device;
 721         qp->event_handler = event_handler;
 722         qp->qp_context = qp_context;
 723         qp->qp_num = real_qp->qp_num;
 724         qp->qp_type = real_qp->qp_type;
 725
 726         spin_lock_irqsave(&real_qp->device->event_handler_lock, flags);
 727         list_add(&qp->open_list, &real_qp->open_list);
 728         spin_unlock_irqrestore(&real_qp->device->event_handler_lock, flags);
 729
 730         return qp;
 731 }
 732
 733 struct ib_qp *ib_open_qp(struct ib_xrcd *xrcd,
 734                          struct ib_qp_open_attr *qp_open_attr)
 735 {
 736         struct ib_qp *qp, *real_qp;
 737
 738         if (qp_open_attr->qp_type != IB_QPT_XRC_TGT)
 739                 return ERR_PTR(-EINVAL);
 740
 741         qp = ERR_PTR(-EINVAL);
 742         mutex_lock(&xrcd->tgt_qp_mutex);
 743         list_for_each_entry(real_qp, &xrcd->tgt_qp_list, xrcd_list) {
 744                 if (real_qp->qp_num == qp_open_attr->qp_num) {
 745                         qp = __ib_open_qp(real_qp, qp_open_attr->event_handler,
 746                                           qp_open_attr->qp_context);
 747                         break;
 748                 }
 749         }
 750         mutex_unlock(&xrcd->tgt_qp_mutex);
 751         return qp;
 752 }
 753 EXPORT_SYMBOL(ib_open_qp);
 754
 755 static struct ib_qp *ib_create_xrc_qp(struct ib_qp *qp,
 756                 struct ib_qp_init_attr *qp_init_attr)
 757 {
 758         struct ib_qp *real_qp = qp;
 759
 760         qp->event_handler = __ib_shared_qp_event_handler;
 761         qp->qp_context = qp;
 762         qp->pd = NULL;
 763         qp->send_cq = qp->recv_cq = NULL;
 764         qp->srq = NULL;
 765         qp->xrcd = qp_init_attr->xrcd;
 766         atomic_inc(&qp_init_attr->xrcd->usecnt);
 767         INIT_LIST_HEAD(&qp->open_list);
 768
 769         qp = __ib_open_qp(real_qp, qp_init_attr->event_handler,
 770                           qp_init_attr->qp_context);
 771         if (!IS_ERR(qp))
 772                 __ib_insert_xrcd_qp(qp_init_attr->xrcd, real_qp);
 773         else
 774                 real_qp->device->destroy_qp(real_qp);
 775         return qp;
 776 }
 777
 778 struct ib_qp *ib_create_qp(struct ib_pd *pd,
 779                            struct ib_qp_init_attr *qp_init_attr)
 780 {
 781         struct ib_device *device = pd ? pd->device : qp_init_attr->xrcd->device;
 782         struct ib_qp *qp;
 783
 784         if (qp_init_attr->rwq_ind_tbl &&
 785             (qp_init_attr->recv_cq ||
 786             qp_init_attr->srq || qp_init_attr->cap.max_recv_wr ||
 787             qp_init_attr->cap.max_recv_sge))
 788                 return ERR_PTR(-EINVAL);
 789
 790         qp = device->create_qp(pd, qp_init_attr, NULL);
 791         if (IS_ERR(qp))
 792                 return qp;
 793
 794         qp->device     = device;
 795         qp->real_qp    = qp;
 796         qp->uobject    = NULL;
 797         qp->qp_type    = qp_init_attr->qp_type;
 798         qp->rwq_ind_tbl = qp_init_attr->rwq_ind_tbl;
 799
 800         atomic_set(&qp->usecnt, 0);
 801         spin_lock_init(&qp->mr_lock);
 802
 803         if (qp_init_attr->qp_type == IB_QPT_XRC_TGT)
 804                 return ib_create_xrc_qp(qp, qp_init_attr);
 805
 806         qp->event_handler = qp_init_attr->event_handler;
 807         qp->qp_context = qp_init_attr->qp_context;
 808         if (qp_init_attr->qp_type == IB_QPT_XRC_INI) {
 809                 qp->recv_cq = NULL;
 810                 qp->srq = NULL;
 811         } else {
 812                 qp->recv_cq = qp_init_attr->recv_cq;
 813                 if (qp_init_attr->recv_cq)
 814                         atomic_inc(&qp_init_attr->recv_cq->usecnt);
 815                 qp->srq = qp_init_attr->srq;
 816                 if (qp->srq)
 817                         atomic_inc(&qp_init_attr->srq->usecnt);
 818         }
 819
 820         qp->pd      = pd;
 821         qp->send_cq = qp_init_attr->send_cq;
 822         qp->xrcd    = NULL;
 823
 824         atomic_inc(&pd->usecnt);
 825         if (qp_init_attr->send_cq)
 826                 atomic_inc(&qp_init_attr->send_cq->usecnt);
 827         if (qp_init_attr->rwq_ind_tbl)
 828                 atomic_inc(&qp->rwq_ind_tbl->usecnt);
 829
 830         /*
 831          * Note: all hw drivers guarantee that max_send_sge is lower than
 832          * the device RDMA WRITE SGE limit but not all hw drivers ensure that
 833          * max_send_sge <= max_sge_rd.
 834          */
 835         qp->max_write_sge = qp_init_attr->cap.max_send_sge;
 836         qp->max_read_sge = min_t(u32, qp_init_attr->cap.max_send_sge,
 837                                  device->attrs.max_sge_rd);
 838
 839         return qp;
 840 }
 841 EXPORT_SYMBOL(ib_create_qp);
 842
 843 static const struct {
 844         int                     valid;
 845         enum ib_qp_attr_mask    req_param[IB_QPT_MAX];
 846         enum ib_qp_attr_mask    opt_param[IB_QPT_MAX];
 847 } qp_state_table[IB_QPS_ERR + 1][IB_QPS_ERR + 1] = {
 848         [IB_QPS_RESET] = {
 849                 [IB_QPS_RESET] = { .valid = 1 },
 850                 [IB_QPS_INIT]  = {
 851                         .valid = 1,
 852                         .req_param = {
 853                                 [IB_QPT_UD]  = (IB_QP_PKEY_INDEX                |
 854                                                 IB_QP_PORT                      |
 855                                                 IB_QP_QKEY),
 856                                 [IB_QPT_RAW_PACKET] = IB_QP_PORT,
 857                                 [IB_QPT_UC]  = (IB_QP_PKEY_INDEX                |
 858                                                 IB_QP_PORT                      |
 859                                                 IB_QP_ACCESS_FLAGS),
 860                                 [IB_QPT_RC]  = (IB_QP_PKEY_INDEX                |
 861                                                 IB_QP_PORT                      |
 862                                                 IB_QP_ACCESS_FLAGS),
 863                                 [IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX            |
 864                                                 IB_QP_PORT                      |
 865                                                 IB_QP_ACCESS_FLAGS),
 866                                 [IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX            |
 867                                                 IB_QP_PORT                      |
 868                                                 IB_QP_ACCESS_FLAGS),
 869                                 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX                |
 870                                                 IB_QP_QKEY),
 871                                 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX                |
 872                                                 IB_QP_QKEY),
 873                         }
 874                 },
 875         },
 876         [IB_QPS_INIT]  = {
 877                 [IB_QPS_RESET] = { .valid = 1 },
 878                 [IB_QPS_ERR] =   { .valid = 1 },
 879                 [IB_QPS_INIT]  = {
 880                         .valid = 1,
 881                         .opt_param = {
 882                                 [IB_QPT_UD]  = (IB_QP_PKEY_INDEX                |
 883                                                 IB_QP_PORT                      |
 884                                                 IB_QP_QKEY),
 885                                 [IB_QPT_UC]  = (IB_QP_PKEY_INDEX                |
 886                                                 IB_QP_PORT                      |
 887                                                 IB_QP_ACCESS_FLAGS),
 888                                 [IB_QPT_RC]  = (IB_QP_PKEY_INDEX                |
 889                                                 IB_QP_PORT                      |
 890                                                 IB_QP_ACCESS_FLAGS),
 891                                 [IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX            |
 892                                                 IB_QP_PORT                      |
 893                                                 IB_QP_ACCESS_FLAGS),
 894                                 [IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX            |
 895                                                 IB_QP_PORT                      |
 896                                                 IB_QP_ACCESS_FLAGS),
 897                                 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX                |
 898                                                 IB_QP_QKEY),
 899                                 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX                |
 900                                                 IB_QP_QKEY),
 901                         }
 902                 },
 903                 [IB_QPS_RTR]   = {
 904                         .valid = 1,
 905                         .req_param = {
 906                                 [IB_QPT_UC]  = (IB_QP_AV                        |
 907                                                 IB_QP_PATH_MTU                  |
 908                                                 IB_QP_DEST_QPN                  |
 909                                                 IB_QP_RQ_PSN),
 910                                 [IB_QPT_RC]  = (IB_QP_AV                        |
 911                                                 IB_QP_PATH_MTU                  |
 912                                                 IB_QP_DEST_QPN                  |
 913                                                 IB_QP_RQ_PSN                    |
 914                                                 IB_QP_MAX_DEST_RD_ATOMIC        |
 915                                                 IB_QP_MIN_RNR_TIMER),
 916                                 [IB_QPT_XRC_INI] = (IB_QP_AV                    |
 917                                                 IB_QP_PATH_MTU                  |
 918                                                 IB_QP_DEST_QPN                  |
 919                                                 IB_QP_RQ_PSN),
 920                                 [IB_QPT_XRC_TGT] = (IB_QP_AV                    |
 921                                                 IB_QP_PATH_MTU                  |
 922                                                 IB_QP_DEST_QPN                  |
 923                                                 IB_QP_RQ_PSN                    |
 924                                                 IB_QP_MAX_DEST_RD_ATOMIC        |
 925                                                 IB_QP_MIN_RNR_TIMER),
 926                         },
 927                         .opt_param = {
 928                                  [IB_QPT_UD]  = (IB_QP_PKEY_INDEX               |
 929                                                  IB_QP_QKEY),
 930                                  [IB_QPT_UC]  = (IB_QP_ALT_PATH                 |
 931                                                  IB_QP_ACCESS_FLAGS             |
 932                                                  IB_QP_PKEY_INDEX),
 933                                  [IB_QPT_RC]  = (IB_QP_ALT_PATH                 |
 934                                                  IB_QP_ACCESS_FLAGS             |
 935                                                  IB_QP_PKEY_INDEX),
 936                                  [IB_QPT_XRC_INI] = (IB_QP_ALT_PATH             |
 937                                                  IB_QP_ACCESS_FLAGS             |
 938                                                  IB_QP_PKEY_INDEX),
 939                                  [IB_QPT_XRC_TGT] = (IB_QP_ALT_PATH             |
 940                                                  IB_QP_ACCESS_FLAGS             |
 941                                                  IB_QP_PKEY_INDEX),
 942                                  [IB_QPT_SMI] = (IB_QP_PKEY_INDEX               |
 943                                                  IB_QP_QKEY),
 944                                  [IB_QPT_GSI] = (IB_QP_PKEY_INDEX               |
 945                                                  IB_QP_QKEY),
 946                          },
 947                 },
 948         },
 949         [IB_QPS_RTR]   = {
 950                 [IB_QPS_RESET] = { .valid = 1 },
 951                 [IB_QPS_ERR] =   { .valid = 1 },
 952                 [IB_QPS_RTS]   = {
 953                         .valid = 1,
 954                         .req_param = {
 955                                 [IB_QPT_UD]  = IB_QP_SQ_PSN,
 956                                 [IB_QPT_UC]  = IB_QP_SQ_PSN,
 957                                 [IB_QPT_RC]  = (IB_QP_TIMEOUT                   |
 958                                                 IB_QP_RETRY_CNT                 |
 959                                                 IB_QP_RNR_RETRY                 |
 960                                                 IB_QP_SQ_PSN                    |
 961                                                 IB_QP_MAX_QP_RD_ATOMIC),
 962                                 [IB_QPT_XRC_INI] = (IB_QP_TIMEOUT               |
 963                                                 IB_QP_RETRY_CNT                 |
 964                                                 IB_QP_RNR_RETRY                 |
 965                                                 IB_QP_SQ_PSN                    |
 966                                                 IB_QP_MAX_QP_RD_ATOMIC),
 967                                 [IB_QPT_XRC_TGT] = (IB_QP_TIMEOUT               |
 968                                                 IB_QP_SQ_PSN),
 969                                 [IB_QPT_SMI] = IB_QP_SQ_PSN,
 970                                 [IB_QPT_GSI] = IB_QP_SQ_PSN,
 971                         },
 972                         .opt_param = {
 973                                  [IB_QPT_UD]  = (IB_QP_CUR_STATE                |
 974                                                  IB_QP_QKEY),
 975                                  [IB_QPT_UC]  = (IB_QP_CUR_STATE                |
 976                                                  IB_QP_ALT_PATH                 |
 977                                                  IB_QP_ACCESS_FLAGS             |
 978                                                  IB_QP_PATH_MIG_STATE),
 979                                  [IB_QPT_RC]  = (IB_QP_CUR_STATE                |
 980                                                  IB_QP_ALT_PATH                 |
 981                                                  IB_QP_ACCESS_FLAGS             |
 982                                                  IB_QP_MIN_RNR_TIMER            |
 983                                                  IB_QP_PATH_MIG_STATE),
 984                                  [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE            |
 985                                                  IB_QP_ALT_PATH                 |
 986                                                  IB_QP_ACCESS_FLAGS             |
 987                                                  IB_QP_PATH_MIG_STATE),
 988                                  [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE            |
 989                                                  IB_QP_ALT_PATH                 |
 990                                                  IB_QP_ACCESS_FLAGS             |
 991                                                  IB_QP_MIN_RNR_TIMER            |
 992                                                  IB_QP_PATH_MIG_STATE),
 993                                  [IB_QPT_SMI] = (IB_QP_CUR_STATE                |
 994                                                  IB_QP_QKEY),
 995                                  [IB_QPT_GSI] = (IB_QP_CUR_STATE                |
 996                                                  IB_QP_QKEY),
 997                          }
 998                 }
 999         },
1000         [IB_QPS_RTS]   = {
1001                 [IB_QPS_RESET] = { .valid = 1 },
1002                 [IB_QPS_ERR] =   { .valid = 1 },
1003                 [IB_QPS_RTS]   = {
1004                         .valid = 1,
1005                         .opt_param = {
1006                                 [IB_QPT_UD]  = (IB_QP_CUR_STATE                 |
1007                                                 IB_QP_QKEY),
1008                                 [IB_QPT_UC]  = (IB_QP_CUR_STATE                 |
1009                                                 IB_QP_ACCESS_FLAGS              |
1010                                                 IB_QP_ALT_PATH                  |
1011                                                 IB_QP_PATH_MIG_STATE),
1012                                 [IB_QPT_RC]  = (IB_QP_CUR_STATE                 |
1013                                                 IB_QP_ACCESS_FLAGS              |
1014                                                 IB_QP_ALT_PATH                  |
1015                                                 IB_QP_PATH_MIG_STATE            |
1016                                                 IB_QP_MIN_RNR_TIMER),
1017                                 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE             |
1018                                                 IB_QP_ACCESS_FLAGS              |
1019                                                 IB_QP_ALT_PATH                  |
1020                                                 IB_QP_PATH_MIG_STATE),
1021                                 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE             |
1022                                                 IB_QP_ACCESS_FLAGS              |
1023                                                 IB_QP_ALT_PATH                  |
1024                                                 IB_QP_PATH_MIG_STATE            |
1025                                                 IB_QP_MIN_RNR_TIMER),
1026                                 [IB_QPT_SMI] = (IB_QP_CUR_STATE                 |
1027                                                 IB_QP_QKEY),
1028                                 [IB_QPT_GSI] = (IB_QP_CUR_STATE                 |
1029                                                 IB_QP_QKEY),
1030                         }
1031                 },
1032                 [IB_QPS_SQD]   = {
1033                         .valid = 1,
1034                         .opt_param = {
1035                                 [IB_QPT_UD]  = IB_QP_EN_SQD_ASYNC_NOTIFY,
1036                                 [IB_QPT_UC]  = IB_QP_EN_SQD_ASYNC_NOTIFY,
1037                                 [IB_QPT_RC]  = IB_QP_EN_SQD_ASYNC_NOTIFY,
1038                                 [IB_QPT_XRC_INI] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1039                                 [IB_QPT_XRC_TGT] = IB_QP_EN_SQD_ASYNC_NOTIFY, /* ??? */
1040                                 [IB_QPT_SMI] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1041                                 [IB_QPT_GSI] = IB_QP_EN_SQD_ASYNC_NOTIFY
1042                         }
1043                 },
1044         },
1045         [IB_QPS_SQD]   = {
1046                 [IB_QPS_RESET] = { .valid = 1 },
1047                 [IB_QPS_ERR] =   { .valid = 1 },
1048                 [IB_QPS_RTS]   = {
1049                         .valid = 1,
1050                         .opt_param = {
1051                                 [IB_QPT_UD]  = (IB_QP_CUR_STATE                 |
1052                                                 IB_QP_QKEY),
1053                                 [IB_QPT_UC]  = (IB_QP_CUR_STATE                 |
1054                                                 IB_QP_ALT_PATH                  |
1055                                                 IB_QP_ACCESS_FLAGS              |
1056                                                 IB_QP_PATH_MIG_STATE),
1057                                 [IB_QPT_RC]  = (IB_QP_CUR_STATE                 |
1058                                                 IB_QP_ALT_PATH                  |
1059                                                 IB_QP_ACCESS_FLAGS              |
1060                                                 IB_QP_MIN_RNR_TIMER             |
1061                                                 IB_QP_PATH_MIG_STATE),
1062                                 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE             |
1063                                                 IB_QP_ALT_PATH                  |
1064                                                 IB_QP_ACCESS_FLAGS              |
1065                                                 IB_QP_PATH_MIG_STATE),
1066                                 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE             |
1067                                                 IB_QP_ALT_PATH                  |
1068                                                 IB_QP_ACCESS_FLAGS              |
1069                                                 IB_QP_MIN_RNR_TIMER             |
1070                                                 IB_QP_PATH_MIG_STATE),
1071                                 [IB_QPT_SMI] = (IB_QP_CUR_STATE                 |
1072                                                 IB_QP_QKEY),
1073                                 [IB_QPT_GSI] = (IB_QP_CUR_STATE                 |
1074                                                 IB_QP_QKEY),
1075                         }
1076                 },
1077                 [IB_QPS_SQD]   = {
1078                         .valid = 1,
1079                         .opt_param = {
1080                                 [IB_QPT_UD]  = (IB_QP_PKEY_INDEX                |
1081                                                 IB_QP_QKEY),
1082                                 [IB_QPT_UC]  = (IB_QP_AV                        |
1083                                                 IB_QP_ALT_PATH                  |
1084                                                 IB_QP_ACCESS_FLAGS              |
1085                                                 IB_QP_PKEY_INDEX                |
1086                                                 IB_QP_PATH_MIG_STATE),
1087                                 [IB_QPT_RC]  = (IB_QP_PORT                      |
1088                                                 IB_QP_AV                        |
1089                                                 IB_QP_TIMEOUT                   |
1090                                                 IB_QP_RETRY_CNT                 |
1091                                                 IB_QP_RNR_RETRY                 |
1092                                                 IB_QP_MAX_QP_RD_ATOMIC          |
1093                                                 IB_QP_MAX_DEST_RD_ATOMIC        |
1094                                                 IB_QP_ALT_PATH                  |
1095                                                 IB_QP_ACCESS_FLAGS              |
1096                                                 IB_QP_PKEY_INDEX                |
1097                                                 IB_QP_MIN_RNR_TIMER             |
1098                                                 IB_QP_PATH_MIG_STATE),
1099                                 [IB_QPT_XRC_INI] = (IB_QP_PORT                  |
1100                                                 IB_QP_AV                        |
1101                                                 IB_QP_TIMEOUT                   |
1102                                                 IB_QP_RETRY_CNT                 |
1103                                                 IB_QP_RNR_RETRY                 |
1104                                                 IB_QP_MAX_QP_RD_ATOMIC          |
1105                                                 IB_QP_ALT_PATH                  |
1106                                                 IB_QP_ACCESS_FLAGS              |
1107                                                 IB_QP_PKEY_INDEX                |
1108                                                 IB_QP_PATH_MIG_STATE),
1109                                 [IB_QPT_XRC_TGT] = (IB_QP_PORT                  |
1110                                                 IB_QP_AV                        |
1111                                                 IB_QP_TIMEOUT                   |
1112                                                 IB_QP_MAX_DEST_RD_ATOMIC        |
1113                                                 IB_QP_ALT_PATH                  |
1114                                                 IB_QP_ACCESS_FLAGS              |
1115                                                 IB_QP_PKEY_INDEX                |
1116                                                 IB_QP_MIN_RNR_TIMER             |
1117                                                 IB_QP_PATH_MIG_STATE),
1118                                 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX                |
1119                                                 IB_QP_QKEY),
1120                                 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX                |
1121                                                 IB_QP_QKEY),
1122                         }
1123                 }
1124         },
1125         [IB_QPS_SQE]   = {
1126                 [IB_QPS_RESET] = { .valid = 1 },
1127                 [IB_QPS_ERR] =   { .valid = 1 },
1128                 [IB_QPS_RTS]   = {
1129                         .valid = 1,
1130                         .opt_param = {
1131                                 [IB_QPT_UD]  = (IB_QP_CUR_STATE                 |
1132                                                 IB_QP_QKEY),
1133                                 [IB_QPT_UC]  = (IB_QP_CUR_STATE                 |
1134                                                 IB_QP_ACCESS_FLAGS),
1135                                 [IB_QPT_SMI] = (IB_QP_CUR_STATE                 |
1136                                                 IB_QP_QKEY),
1137                                 [IB_QPT_GSI] = (IB_QP_CUR_STATE                 |
1138                                                 IB_QP_QKEY),
1139                         }
1140                 }
1141         },
1142         [IB_QPS_ERR] = {
1143                 [IB_QPS_RESET] = { .valid = 1 },
1144                 [IB_QPS_ERR] =   { .valid = 1 }
1145         }
1146 };
1147
1148 int ib_modify_qp_is_ok(enum ib_qp_state cur_state, enum ib_qp_state next_state,
1149                        enum ib_qp_type type, enum ib_qp_attr_mask mask,
1150                        enum rdma_link_layer ll)
1151 {
1152         enum ib_qp_attr_mask req_param, opt_param;
1153
1154         if (cur_state  < 0 || cur_state  > IB_QPS_ERR ||
1155             next_state < 0 || next_state > IB_QPS_ERR)
1156                 return 0;
1157
1158         if (mask & IB_QP_CUR_STATE  &&
1159             cur_state != IB_QPS_RTR && cur_state != IB_QPS_RTS &&
1160             cur_state != IB_QPS_SQD && cur_state != IB_QPS_SQE)
1161                 return 0;
1162
1163         if (!qp_state_table[cur_state][next_state].valid)
1164                 return 0;
1165
1166         req_param = qp_state_table[cur_state][next_state].req_param[type];
1167         opt_param = qp_state_table[cur_state][next_state].opt_param[type];
1168
1169         if ((mask & req_param) != req_param)
1170                 return 0;
1171
1172         if (mask & ~(req_param | opt_param | IB_QP_STATE))
1173                 return 0;
1174
1175         return 1;
1176 }
1177 EXPORT_SYMBOL(ib_modify_qp_is_ok);
1178
1179 int ib_resolve_eth_dmac(struct ib_device *device,
1180                         struct ib_ah_attr *ah_attr)
1181 {
1182         struct ib_gid_attr sgid_attr;
1183         union ib_gid sgid;
1184         int hop_limit;
1185         int ret;
1186
1187         if (ah_attr->port_num < rdma_start_port(device) ||
1188             ah_attr->port_num > rdma_end_port(device))
1189                 return -EINVAL;
1190
1191         if (!rdma_cap_eth_ah(device, ah_attr->port_num))
1192                 return 0;
1193
1194         if (rdma_is_multicast_addr((struct in6_addr *)ah_attr->grh.dgid.raw)) {
1195                 if (ipv6_addr_v4mapped((struct in6_addr *)ah_attr->grh.dgid.raw)) {
1196                         __be32 addr = 0;
1197
1198                         memcpy(&addr, ah_attr->grh.dgid.raw + 12, 4);
1199                         ip_eth_mc_map(addr, (char *)ah_attr->dmac);
1200                 } else {
1201                         ipv6_eth_mc_map((struct in6_addr *)ah_attr->grh.dgid.raw,
1202                                         (char *)ah_attr->dmac);
1203                 }
1204                 return 0;
1205         }
1206
1207         ret = ib_query_gid(device,
1208                            ah_attr->port_num,
1209                            ah_attr->grh.sgid_index,
1210                            &sgid, &sgid_attr);
1211         if (ret != 0)
1212                 return (ret);
1213         if (!sgid_attr.ndev)
1214                 return -ENXIO;
1215
1216         ret = rdma_addr_find_l2_eth_by_grh(&sgid,
1217                                            &ah_attr->grh.dgid,
1218                                            ah_attr->dmac,
1219                                            sgid_attr.ndev, &hop_limit);
1220         dev_put(sgid_attr.ndev);
1221
1222         ah_attr->grh.hop_limit = hop_limit;
1223         return ret;
1224 }
1225 EXPORT_SYMBOL(ib_resolve_eth_dmac);
1226
1227
1228 int ib_modify_qp(struct ib_qp *qp,
1229                  struct ib_qp_attr *qp_attr,
1230                  int qp_attr_mask)
1231 {
1232         if (qp_attr_mask & IB_QP_AV) {
1233                 int ret;
1234
1235                 ret = ib_resolve_eth_dmac(qp->device, &qp_attr->ah_attr);
1236                 if (ret)
1237                         return ret;
1238         }
1239
1240         return qp->device->modify_qp(qp->real_qp, qp_attr, qp_attr_mask, NULL);
1241 }
1242 EXPORT_SYMBOL(ib_modify_qp);
1243
1244 int ib_query_qp(struct ib_qp *qp,
1245                 struct ib_qp_attr *qp_attr,
1246                 int qp_attr_mask,
1247                 struct ib_qp_init_attr *qp_init_attr)
1248 {
1249         return qp->device->query_qp ?
1250                 qp->device->query_qp(qp->real_qp, qp_attr, qp_attr_mask, qp_init_attr) :
1251                 -ENOSYS;
1252 }
1253 EXPORT_SYMBOL(ib_query_qp);
1254
1255 int ib_close_qp(struct ib_qp *qp)
1256 {
1257         struct ib_qp *real_qp;
1258         unsigned long flags;
1259
1260         real_qp = qp->real_qp;
1261         if (real_qp == qp)
1262                 return -EINVAL;
1263
1264         spin_lock_irqsave(&real_qp->device->event_handler_lock, flags);
1265         list_del(&qp->open_list);
1266         spin_unlock_irqrestore(&real_qp->device->event_handler_lock, flags);
1267
1268         atomic_dec(&real_qp->usecnt);
1269         kfree(qp);
1270
1271         return 0;
1272 }
1273 EXPORT_SYMBOL(ib_close_qp);
1274
1275 static int __ib_destroy_shared_qp(struct ib_qp *qp)
1276 {
1277         struct ib_xrcd *xrcd;
1278         struct ib_qp *real_qp;
1279         int ret;
1280
1281         real_qp = qp->real_qp;
1282         xrcd = real_qp->xrcd;
1283
1284         mutex_lock(&xrcd->tgt_qp_mutex);
1285         ib_close_qp(qp);
1286         if (atomic_read(&real_qp->usecnt) == 0)
1287                 list_del(&real_qp->xrcd_list);
1288         else
1289                 real_qp = NULL;
1290         mutex_unlock(&xrcd->tgt_qp_mutex);
1291
1292         if (real_qp) {
1293                 ret = ib_destroy_qp(real_qp);
1294                 if (!ret)
1295                         atomic_dec(&xrcd->usecnt);
1296                 else
1297                         __ib_insert_xrcd_qp(xrcd, real_qp);
1298         }
1299
1300         return 0;
1301 }
1302
1303 int ib_destroy_qp(struct ib_qp *qp)
1304 {
1305         struct ib_pd *pd;
1306         struct ib_cq *scq, *rcq;
1307         struct ib_srq *srq;
1308         struct ib_rwq_ind_table *ind_tbl;
1309         int ret;
1310
1311         if (atomic_read(&qp->usecnt))
1312                 return -EBUSY;
1313
1314         if (qp->real_qp != qp)
1315                 return __ib_destroy_shared_qp(qp);
1316
1317         pd   = qp->pd;
1318         scq  = qp->send_cq;
1319         rcq  = qp->recv_cq;
1320         srq  = qp->srq;
1321         ind_tbl = qp->rwq_ind_tbl;
1322
1323         ret = qp->device->destroy_qp(qp);
1324         if (!ret) {
1325                 if (pd)
1326                         atomic_dec(&pd->usecnt);
1327                 if (scq)
1328                         atomic_dec(&scq->usecnt);
1329                 if (rcq)
1330                         atomic_dec(&rcq->usecnt);
1331                 if (srq)
1332                         atomic_dec(&srq->usecnt);
1333                 if (ind_tbl)
1334                         atomic_dec(&ind_tbl->usecnt);
1335         }
1336
1337         return ret;
1338 }
1339 EXPORT_SYMBOL(ib_destroy_qp);
1340
1341 /* Completion queues */
1342
1343 struct ib_cq *ib_create_cq(struct ib_device *device,
1344                            ib_comp_handler comp_handler,
1345                            void (*event_handler)(struct ib_event *, void *),
1346                            void *cq_context,
1347                            const struct ib_cq_init_attr *cq_attr)
1348 {
1349         struct ib_cq *cq;
1350
1351         cq = device->create_cq(device, cq_attr, NULL, NULL);
1352
1353         if (!IS_ERR(cq)) {
1354                 cq->device        = device;
1355                 cq->uobject       = NULL;
1356                 cq->comp_handler  = comp_handler;
1357                 cq->event_handler = event_handler;
1358                 cq->cq_context    = cq_context;
1359                 atomic_set(&cq->usecnt, 0);
1360         }
1361
1362         return cq;
1363 }
1364 EXPORT_SYMBOL(ib_create_cq);
1365
1366 int ib_modify_cq(struct ib_cq *cq, u16 cq_count, u16 cq_period)
1367 {
1368         return cq->device->modify_cq ?
1369                 cq->device->modify_cq(cq, cq_count, cq_period) : -ENOSYS;
1370 }
1371 EXPORT_SYMBOL(ib_modify_cq);
1372
1373 int ib_destroy_cq(struct ib_cq *cq)
1374 {
1375         if (atomic_read(&cq->usecnt))
1376                 return -EBUSY;
1377
1378         return cq->device->destroy_cq(cq);
1379 }
1380 EXPORT_SYMBOL(ib_destroy_cq);
1381
1382 int ib_resize_cq(struct ib_cq *cq, int cqe)
1383 {
1384         return cq->device->resize_cq ?
1385                 cq->device->resize_cq(cq, cqe, NULL) : -ENOSYS;
1386 }
1387 EXPORT_SYMBOL(ib_resize_cq);
1388
1389 /* Memory regions */
1390
1391 int ib_dereg_mr(struct ib_mr *mr)
1392 {
1393         struct ib_pd *pd = mr->pd;
1394         int ret;
1395
1396         ret = mr->device->dereg_mr(mr);
1397         if (!ret)
1398                 atomic_dec(&pd->usecnt);
1399
1400         return ret;
1401 }
1402 EXPORT_SYMBOL(ib_dereg_mr);
1403
1404 /**
1405  * ib_alloc_mr() - Allocates a memory region
1406  * @pd:            protection domain associated with the region
1407  * @mr_type:       memory region type
1408  * @max_num_sg:    maximum sg entries available for registration.
1409  *
1410  * Notes:
1411  * Memory registeration page/sg lists must not exceed max_num_sg.
1412  * For mr_type IB_MR_TYPE_MEM_REG, the total length cannot exceed
1413  * max_num_sg * used_page_size.
1414  *
1415  */
1416 struct ib_mr *ib_alloc_mr(struct ib_pd *pd,
1417                           enum ib_mr_type mr_type,
1418                           u32 max_num_sg)
1419 {
1420         struct ib_mr *mr;
1421
1422         if (!pd->device->alloc_mr)
1423                 return ERR_PTR(-ENOSYS);
1424
1425         mr = pd->device->alloc_mr(pd, mr_type, max_num_sg);
1426         if (!IS_ERR(mr)) {
1427                 mr->device  = pd->device;
1428                 mr->pd      = pd;
1429                 mr->uobject = NULL;
1430                 atomic_inc(&pd->usecnt);
1431                 mr->need_inval = false;
1432         }
1433
1434         return mr;
1435 }
1436 EXPORT_SYMBOL(ib_alloc_mr);
1437
1438 /* "Fast" memory regions */
1439
1440 struct ib_fmr *ib_alloc_fmr(struct ib_pd *pd,
1441                             int mr_access_flags,
1442                             struct ib_fmr_attr *fmr_attr)
1443 {
1444         struct ib_fmr *fmr;
1445
1446         if (!pd->device->alloc_fmr)
1447                 return ERR_PTR(-ENOSYS);
1448
1449         fmr = pd->device->alloc_fmr(pd, mr_access_flags, fmr_attr);
1450         if (!IS_ERR(fmr)) {
1451                 fmr->device = pd->device;
1452                 fmr->pd     = pd;
1453                 atomic_inc(&pd->usecnt);
1454         }
1455
1456         return fmr;
1457 }
1458 EXPORT_SYMBOL(ib_alloc_fmr);
1459
1460 int ib_unmap_fmr(struct list_head *fmr_list)
1461 {
1462         struct ib_fmr *fmr;
1463
1464         if (list_empty(fmr_list))
1465                 return 0;
1466
1467         fmr = list_entry(fmr_list->next, struct ib_fmr, list);
1468         return fmr->device->unmap_fmr(fmr_list);
1469 }
1470 EXPORT_SYMBOL(ib_unmap_fmr);
1471
1472 int ib_dealloc_fmr(struct ib_fmr *fmr)
1473 {
1474         struct ib_pd *pd;
1475         int ret;
1476
1477         pd = fmr->pd;
1478         ret = fmr->device->dealloc_fmr(fmr);
1479         if (!ret)
1480                 atomic_dec(&pd->usecnt);
1481
1482         return ret;
1483 }
1484 EXPORT_SYMBOL(ib_dealloc_fmr);
1485
1486 /* Multicast groups */
1487
1488 static bool is_valid_mcast_lid(struct ib_qp *qp, u16 lid)
1489 {
1490         struct ib_qp_init_attr init_attr = {};
1491         struct ib_qp_attr attr = {};
1492         int num_eth_ports = 0;
1493         int port;
1494
1495         /* If QP state >= init, it is assigned to a port and we can check this
1496          * port only.
1497          */
1498         if (!ib_query_qp(qp, &attr, IB_QP_STATE | IB_QP_PORT, &init_attr)) {
1499                 if (attr.qp_state >= IB_QPS_INIT) {
1500                         if (rdma_port_get_link_layer(qp->device, attr.port_num) !=
1501                             IB_LINK_LAYER_INFINIBAND)
1502                                 return true;
1503                         goto lid_check;
1504                 }
1505         }
1506
1507         /* Can't get a quick answer, iterate over all ports */
1508         for (port = 0; port < qp->device->phys_port_cnt; port++)
1509                 if (rdma_port_get_link_layer(qp->device, port) !=
1510                     IB_LINK_LAYER_INFINIBAND)
1511                         num_eth_ports++;
1512
1513         /* If we have at lease one Ethernet port, RoCE annex declares that
1514          * multicast LID should be ignored. We can't tell at this step if the
1515          * QP belongs to an IB or Ethernet port.
1516          */
1517         if (num_eth_ports)
1518                 return true;
1519
1520         /* If all the ports are IB, we can check according to IB spec. */
1521 lid_check:
1522         return !(lid < be16_to_cpu(IB_MULTICAST_LID_BASE) ||
1523                  lid == be16_to_cpu(IB_LID_PERMISSIVE));
1524 }
1525
1526 int ib_attach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid)
1527 {
1528         int ret;
1529
1530         if (!qp->device->attach_mcast)
1531                 return -ENOSYS;
1532
1533         if (!rdma_is_multicast_addr((struct in6_addr *)gid->raw) ||
1534             qp->qp_type != IB_QPT_UD || !is_valid_mcast_lid(qp, lid))
1535                 return -EINVAL;
1536
1537         ret = qp->device->attach_mcast(qp, gid, lid);
1538         if (!ret)
1539                 atomic_inc(&qp->usecnt);
1540         return ret;
1541 }
1542 EXPORT_SYMBOL(ib_attach_mcast);
1543
1544 int ib_detach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid)
1545 {
1546         int ret;
1547
1548         if (!qp->device->detach_mcast)
1549                 return -ENOSYS;
1550
1551         if (!rdma_is_multicast_addr((struct in6_addr *)gid->raw) ||
1552             qp->qp_type != IB_QPT_UD || !is_valid_mcast_lid(qp, lid))
1553                 return -EINVAL;
1554
1555         ret = qp->device->detach_mcast(qp, gid, lid);
1556         if (!ret)
1557                 atomic_dec(&qp->usecnt);
1558         return ret;
1559 }
1560 EXPORT_SYMBOL(ib_detach_mcast);
1561
1562 struct ib_xrcd *ib_alloc_xrcd(struct ib_device *device)
1563 {
1564         struct ib_xrcd *xrcd;
1565
1566         if (!device->alloc_xrcd)
1567                 return ERR_PTR(-ENOSYS);
1568
1569         xrcd = device->alloc_xrcd(device, NULL, NULL);
1570         if (!IS_ERR(xrcd)) {
1571                 xrcd->device = device;
1572                 xrcd->inode = NULL;
1573                 atomic_set(&xrcd->usecnt, 0);
1574                 mutex_init(&xrcd->tgt_qp_mutex);
1575                 INIT_LIST_HEAD(&xrcd->tgt_qp_list);
1576         }
1577
1578         return xrcd;
1579 }
1580 EXPORT_SYMBOL(ib_alloc_xrcd);
1581
1582 int ib_dealloc_xrcd(struct ib_xrcd *xrcd)
1583 {
1584         struct ib_qp *qp;
1585         int ret;
1586
1587         if (atomic_read(&xrcd->usecnt))
1588                 return -EBUSY;
1589
1590         while (!list_empty(&xrcd->tgt_qp_list)) {
1591                 qp = list_entry(xrcd->tgt_qp_list.next, struct ib_qp, xrcd_list);
1592                 ret = ib_destroy_qp(qp);
1593                 if (ret)
1594                         return ret;
1595         }
1596
1597         return xrcd->device->dealloc_xrcd(xrcd);
1598 }
1599 EXPORT_SYMBOL(ib_dealloc_xrcd);
1600
1601 /**
1602  * ib_create_wq - Creates a WQ associated with the specified protection
1603  * domain.
1604  * @pd: The protection domain associated with the WQ.
1605  * @wq_init_attr: A list of initial attributes required to create the
1606  * WQ. If WQ creation succeeds, then the attributes are updated to
1607  * the actual capabilities of the created WQ.
1608  *
1609  * wq_init_attr->max_wr and wq_init_attr->max_sge determine
1610  * the requested size of the WQ, and set to the actual values allocated
1611  * on return.
1612  * If ib_create_wq() succeeds, then max_wr and max_sge will always be
1613  * at least as large as the requested values.
1614  */
1615 struct ib_wq *ib_create_wq(struct ib_pd *pd,
1616                            struct ib_wq_init_attr *wq_attr)
1617 {
1618         struct ib_wq *wq;
1619
1620         if (!pd->device->create_wq)
1621                 return ERR_PTR(-ENOSYS);
1622
1623         wq = pd->device->create_wq(pd, wq_attr, NULL);
1624         if (!IS_ERR(wq)) {
1625                 wq->event_handler = wq_attr->event_handler;
1626                 wq->wq_context = wq_attr->wq_context;
1627                 wq->wq_type = wq_attr->wq_type;
1628                 wq->cq = wq_attr->cq;
1629                 wq->device = pd->device;
1630                 wq->pd = pd;
1631                 wq->uobject = NULL;
1632                 atomic_inc(&pd->usecnt);
1633                 atomic_inc(&wq_attr->cq->usecnt);
1634                 atomic_set(&wq->usecnt, 0);
1635         }
1636         return wq;
1637 }
1638 EXPORT_SYMBOL(ib_create_wq);
1639
1640 /**
1641  * ib_destroy_wq - Destroys the specified WQ.
1642  * @wq: The WQ to destroy.
1643  */
1644 int ib_destroy_wq(struct ib_wq *wq)
1645 {
1646         int err;
1647         struct ib_cq *cq = wq->cq;
1648         struct ib_pd *pd = wq->pd;
1649
1650         if (atomic_read(&wq->usecnt))
1651                 return -EBUSY;
1652
1653         err = wq->device->destroy_wq(wq);
1654         if (!err) {
1655                 atomic_dec(&pd->usecnt);
1656                 atomic_dec(&cq->usecnt);
1657         }
1658         return err;
1659 }
1660 EXPORT_SYMBOL(ib_destroy_wq);
1661
1662 /**
1663  * ib_modify_wq - Modifies the specified WQ.
1664  * @wq: The WQ to modify.
1665  * @wq_attr: On input, specifies the WQ attributes to modify.
1666  * @wq_attr_mask: A bit-mask used to specify which attributes of the WQ
1667  *   are being modified.
1668  * On output, the current values of selected WQ attributes are returned.
1669  */
1670 int ib_modify_wq(struct ib_wq *wq, struct ib_wq_attr *wq_attr,
1671                  u32 wq_attr_mask)
1672 {
1673         int err;
1674
1675         if (!wq->device->modify_wq)
1676                 return -ENOSYS;
1677
1678         err = wq->device->modify_wq(wq, wq_attr, wq_attr_mask, NULL);
1679         return err;
1680 }
1681 EXPORT_SYMBOL(ib_modify_wq);
1682
1683 /*
1684  * ib_create_rwq_ind_table - Creates a RQ Indirection Table.
1685  * @device: The device on which to create the rwq indirection table.
1686  * @ib_rwq_ind_table_init_attr: A list of initial attributes required to
1687  * create the Indirection Table.
1688  *
1689  * Note: The life time of ib_rwq_ind_table_init_attr->ind_tbl is not less
1690  *      than the created ib_rwq_ind_table object and the caller is responsible
1691  *      for its memory allocation/free.
1692  */
1693 struct ib_rwq_ind_table *ib_create_rwq_ind_table(struct ib_device *device,
1694                                                  struct ib_rwq_ind_table_init_attr *init_attr)
1695 {
1696         struct ib_rwq_ind_table *rwq_ind_table;
1697         int i;
1698         u32 table_size;
1699
1700         if (!device->create_rwq_ind_table)
1701                 return ERR_PTR(-ENOSYS);
1702
1703         table_size = (1 << init_attr->log_ind_tbl_size);
1704         rwq_ind_table = device->create_rwq_ind_table(device,
1705                                 init_attr, NULL);
1706         if (IS_ERR(rwq_ind_table))
1707                 return rwq_ind_table;
1708
1709         rwq_ind_table->ind_tbl = init_attr->ind_tbl;
1710         rwq_ind_table->log_ind_tbl_size = init_attr->log_ind_tbl_size;
1711         rwq_ind_table->device = device;
1712         rwq_ind_table->uobject = NULL;
1713         atomic_set(&rwq_ind_table->usecnt, 0);
1714
1715         for (i = 0; i < table_size; i++)
1716                 atomic_inc(&rwq_ind_table->ind_tbl[i]->usecnt);
1717
1718         return rwq_ind_table;
1719 }
1720 EXPORT_SYMBOL(ib_create_rwq_ind_table);
1721
1722 /*
1723  * ib_destroy_rwq_ind_table - Destroys the specified Indirection Table.
1724  * @wq_ind_table: The Indirection Table to destroy.
1725 */
1726 int ib_destroy_rwq_ind_table(struct ib_rwq_ind_table *rwq_ind_table)
1727 {
1728         int err, i;
1729         u32 table_size = (1 << rwq_ind_table->log_ind_tbl_size);
1730         struct ib_wq **ind_tbl = rwq_ind_table->ind_tbl;
1731
1732         if (atomic_read(&rwq_ind_table->usecnt))
1733                 return -EBUSY;
1734
1735         err = rwq_ind_table->device->destroy_rwq_ind_table(rwq_ind_table);
1736         if (!err) {
1737                 for (i = 0; i < table_size; i++)
1738                         atomic_dec(&ind_tbl[i]->usecnt);
1739         }
1740
1741         return err;
1742 }
1743 EXPORT_SYMBOL(ib_destroy_rwq_ind_table);
1744
1745 struct ib_flow *ib_create_flow(struct ib_qp *qp,
1746                                struct ib_flow_attr *flow_attr,
1747                                int domain)
1748 {
1749         struct ib_flow *flow_id;
1750         if (!qp->device->create_flow)
1751                 return ERR_PTR(-ENOSYS);
1752
1753         flow_id = qp->device->create_flow(qp, flow_attr, domain);
1754         if (!IS_ERR(flow_id))
1755                 atomic_inc(&qp->usecnt);
1756         return flow_id;
1757 }
1758 EXPORT_SYMBOL(ib_create_flow);
1759
1760 int ib_destroy_flow(struct ib_flow *flow_id)
1761 {
1762         int err;
1763         struct ib_qp *qp = flow_id->qp;
1764
1765         err = qp->device->destroy_flow(flow_id);
1766         if (!err)
1767                 atomic_dec(&qp->usecnt);
1768         return err;
1769 }
1770 EXPORT_SYMBOL(ib_destroy_flow);
1771
1772 int ib_check_mr_status(struct ib_mr *mr, u32 check_mask,
1773                        struct ib_mr_status *mr_status)
1774 {
1775         return mr->device->check_mr_status ?
1776                 mr->device->check_mr_status(mr, check_mask, mr_status) : -ENOSYS;
1777 }
1778 EXPORT_SYMBOL(ib_check_mr_status);
1779
1780 int ib_set_vf_link_state(struct ib_device *device, int vf, u8 port,
1781                          int state)
1782 {
1783         if (!device->set_vf_link_state)
1784                 return -ENOSYS;
1785
1786         return device->set_vf_link_state(device, vf, port, state);
1787 }
1788 EXPORT_SYMBOL(ib_set_vf_link_state);
1789
1790 int ib_get_vf_config(struct ib_device *device, int vf, u8 port,
1791                      struct ifla_vf_info *info)
1792 {
1793         if (!device->get_vf_config)
1794                 return -ENOSYS;
1795
1796         return device->get_vf_config(device, vf, port, info);
1797 }
1798 EXPORT_SYMBOL(ib_get_vf_config);
1799
1800 int ib_get_vf_stats(struct ib_device *device, int vf, u8 port,
1801                     struct ifla_vf_stats *stats)
1802 {
1803         if (!device->get_vf_stats)
1804                 return -ENOSYS;
1805
1806         return device->get_vf_stats(device, vf, port, stats);
1807 }
1808 EXPORT_SYMBOL(ib_get_vf_stats);
1809
1810 int ib_set_vf_guid(struct ib_device *device, int vf, u8 port, u64 guid,
1811                    int type)
1812 {
1813         if (!device->set_vf_guid)
1814                 return -ENOSYS;
1815
1816         return device->set_vf_guid(device, vf, port, guid, type);
1817 }
1818 EXPORT_SYMBOL(ib_set_vf_guid);
1819
1820 /**
1821  * ib_map_mr_sg() - Map the largest prefix of a dma mapped SG list
1822  *     and set it the memory region.
1823  * @mr:            memory region
1824  * @sg:            dma mapped scatterlist
1825  * @sg_nents:      number of entries in sg
1826  * @sg_offset:     offset in bytes into sg
1827  * @page_size:     page vector desired page size
1828  *
1829  * Constraints:
1830  * - The first sg element is allowed to have an offset.
1831  * - Each sg element must either be aligned to page_size or virtually
1832  *   contiguous to the previous element. In case an sg element has a
1833  *   non-contiguous offset, the mapping prefix will not include it.
1834  * - The last sg element is allowed to have length less than page_size.
1835  * - If sg_nents total byte length exceeds the mr max_num_sge * page_size
1836  *   then only max_num_sg entries will be mapped.
1837  * - If the MR was allocated with type IB_MR_TYPE_SG_GAPS, none of these
1838  *   constraints holds and the page_size argument is ignored.
1839  *
1840  * Returns the number of sg elements that were mapped to the memory region.
1841  *
1842  * After this completes successfully, the  memory region
1843  * is ready for registration.
1844  */
1845 int ib_map_mr_sg(struct ib_mr *mr, struct scatterlist *sg, int sg_nents,
1846                  unsigned int *sg_offset, unsigned int page_size)
1847 {
1848         if (unlikely(!mr->device->map_mr_sg))
1849                 return -ENOSYS;
1850
1851         mr->page_size = page_size;
1852
1853         return mr->device->map_mr_sg(mr, sg, sg_nents, sg_offset);
1854 }
1855 EXPORT_SYMBOL(ib_map_mr_sg);
1856
1857 /**
1858  * ib_sg_to_pages() - Convert the largest prefix of a sg list
1859  *     to a page vector
1860  * @mr:            memory region
1861  * @sgl:           dma mapped scatterlist
1862  * @sg_nents:      number of entries in sg
1863  * @sg_offset_p:   IN:  start offset in bytes into sg
1864  *                 OUT: offset in bytes for element n of the sg of the first
1865  *                      byte that has not been processed where n is the return
1866  *                      value of this function.
1867  * @set_page:      driver page assignment function pointer
1868  *
1869  * Core service helper for drivers to convert the largest
1870  * prefix of given sg list to a page vector. The sg list
1871  * prefix converted is the prefix that meet the requirements
1872  * of ib_map_mr_sg.
1873  *
1874  * Returns the number of sg elements that were assigned to
1875  * a page vector.
1876  */
1877 int ib_sg_to_pages(struct ib_mr *mr, struct scatterlist *sgl, int sg_nents,
1878                 unsigned int *sg_offset_p, int (*set_page)(struct ib_mr *, u64))
1879 {
1880         struct scatterlist *sg;
1881         u64 last_end_dma_addr = 0;
1882         unsigned int sg_offset = sg_offset_p ? *sg_offset_p : 0;
1883         unsigned int last_page_off = 0;
1884         u64 page_mask = ~((u64)mr->page_size - 1);
1885         int i, ret;
1886
1887         if (unlikely(sg_nents <= 0 || sg_offset > sg_dma_len(&sgl[0])))
1888                 return -EINVAL;
1889
1890         mr->iova = sg_dma_address(&sgl[0]) + sg_offset;
1891         mr->length = 0;
1892
1893         for_each_sg(sgl, sg, sg_nents, i) {
1894                 u64 dma_addr = sg_dma_address(sg) + sg_offset;
1895                 u64 prev_addr = dma_addr;
1896                 unsigned int dma_len = sg_dma_len(sg) - sg_offset;
1897                 u64 end_dma_addr = dma_addr + dma_len;
1898                 u64 page_addr = dma_addr & page_mask;
1899
1900                 /*
1901                  * For the second and later elements, check whether either the
1902                  * end of element i-1 or the start of element i is not aligned
1903                  * on a page boundary.
1904                  */
1905                 if (i && (last_page_off != 0 || page_addr != dma_addr)) {
1906                         /* Stop mapping if there is a gap. */
1907                         if (last_end_dma_addr != dma_addr)
1908                                 break;
1909
1910                         /*
1911                          * Coalesce this element with the last. If it is small
1912                          * enough just update mr->length. Otherwise start
1913                          * mapping from the next page.
1914                          */
1915                         goto next_page;
1916                 }
1917
1918                 do {
1919                         ret = set_page(mr, page_addr);
1920                         if (unlikely(ret < 0)) {
1921                                 sg_offset = prev_addr - sg_dma_address(sg);
1922                                 mr->length += prev_addr - dma_addr;
1923                                 if (sg_offset_p)
1924                                         *sg_offset_p = sg_offset;
1925                                 return i || sg_offset ? i : ret;
1926                         }
1927                         prev_addr = page_addr;
1928 next_page:
1929                         page_addr += mr->page_size;
1930                 } while (page_addr < end_dma_addr);
1931
1932                 mr->length += dma_len;
1933                 last_end_dma_addr = end_dma_addr;
1934                 last_page_off = end_dma_addr & ~page_mask;
1935
1936                 sg_offset = 0;
1937         }
1938
1939         if (sg_offset_p)
1940                 *sg_offset_p = 0;
1941         return i;
1942 }
1943 EXPORT_SYMBOL(ib_sg_to_pages);
1944
1945 struct ib_drain_cqe {
1946         struct ib_cqe cqe;
1947         struct completion done;
1948 };
1949
1950 static void ib_drain_qp_done(struct ib_cq *cq, struct ib_wc *wc)
1951 {
1952         struct ib_drain_cqe *cqe = container_of(wc->wr_cqe, struct ib_drain_cqe,
1953                                                 cqe);
1954
1955         complete(&cqe->done);
1956 }
1957
1958 /*
1959  * Post a WR and block until its completion is reaped for the SQ.
1960  */
1961 static void __ib_drain_sq(struct ib_qp *qp)
1962 {
1963         struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
1964         struct ib_drain_cqe sdrain;
1965         struct ib_send_wr *bad_swr;
1966         struct ib_rdma_wr swr = {
1967                 .wr = {
1968                         .opcode = IB_WR_RDMA_WRITE,
1969                         .wr_cqe = &sdrain.cqe,
1970                 },
1971         };
1972         int ret;
1973
1974         if (qp->send_cq->poll_ctx == IB_POLL_DIRECT) {
1975                 WARN_ONCE(qp->send_cq->poll_ctx == IB_POLL_DIRECT,
1976                           "IB_POLL_DIRECT poll_ctx not supported for drain\n");
1977                 return;
1978         }
1979
1980         sdrain.cqe.done = ib_drain_qp_done;
1981         init_completion(&sdrain.done);
1982
1983         ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
1984         if (ret) {
1985                 WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
1986                 return;
1987         }
1988
1989         ret = ib_post_send(qp, &swr.wr, &bad_swr);
1990         if (ret) {
1991                 WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
1992                 return;
1993         }
1994
1995         wait_for_completion(&sdrain.done);
1996 }
1997
1998 /*
1999  * Post a WR and block until its completion is reaped for the RQ.
2000  */
2001 static void __ib_drain_rq(struct ib_qp *qp)
2002 {
2003         struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
2004         struct ib_drain_cqe rdrain;
2005         struct ib_recv_wr rwr = {}, *bad_rwr;
2006         int ret;
2007
2008         if (qp->recv_cq->poll_ctx == IB_POLL_DIRECT) {
2009                 WARN_ONCE(qp->recv_cq->poll_ctx == IB_POLL_DIRECT,
2010                           "IB_POLL_DIRECT poll_ctx not supported for drain\n");
2011                 return;
2012         }
2013
2014         rwr.wr_cqe = &rdrain.cqe;
2015         rdrain.cqe.done = ib_drain_qp_done;
2016         init_completion(&rdrain.done);
2017
2018         ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
2019         if (ret) {
2020                 WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
2021                 return;
2022         }
2023
2024         ret = ib_post_recv(qp, &rwr, &bad_rwr);
2025         if (ret) {
2026                 WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
2027                 return;
2028         }
2029
2030         wait_for_completion(&rdrain.done);
2031 }
2032
2033 /**
2034  * ib_drain_sq() - Block until all SQ CQEs have been consumed by the
2035  *                 application.
2036  * @qp:            queue pair to drain
2037  *
2038  * If the device has a provider-specific drain function, then
2039  * call that.  Otherwise call the generic drain function
2040  * __ib_drain_sq().
2041  *
2042  * The caller must:
2043  *
2044  * ensure there is room in the CQ and SQ for the drain work request and
2045  * completion.
2046  *
2047  * allocate the CQ using ib_alloc_cq() and the CQ poll context cannot be
2048  * IB_POLL_DIRECT.
2049  *
2050  * ensure that there are no other contexts that are posting WRs concurrently.
2051  * Otherwise the drain is not guaranteed.
2052  */
2053 void ib_drain_sq(struct ib_qp *qp)
2054 {
2055         if (qp->device->drain_sq)
2056                 qp->device->drain_sq(qp);
2057         else
2058                 __ib_drain_sq(qp);
2059 }
2060 EXPORT_SYMBOL(ib_drain_sq);
2061
2062 /**
2063  * ib_drain_rq() - Block until all RQ CQEs have been consumed by the
2064  *                 application.
2065  * @qp:            queue pair to drain
2066  *
2067  * If the device has a provider-specific drain function, then
2068  * call that.  Otherwise call the generic drain function
2069  * __ib_drain_rq().
2070  *
2071  * The caller must:
2072  *
2073  * ensure there is room in the CQ and RQ for the drain work request and
2074  * completion.
2075  *
2076  * allocate the CQ using ib_alloc_cq() and the CQ poll context cannot be
2077  * IB_POLL_DIRECT.
2078  *
2079  * ensure that there are no other contexts that are posting WRs concurrently.
2080  * Otherwise the drain is not guaranteed.
2081  */
2082 void ib_drain_rq(struct ib_qp *qp)
2083 {
2084         if (qp->device->drain_rq)
2085                 qp->device->drain_rq(qp);
2086         else
2087                 __ib_drain_rq(qp);
2088 }
2089 EXPORT_SYMBOL(ib_drain_rq);
2090
2091 /**
2092  * ib_drain_qp() - Block until all CQEs have been consumed by the
2093  *                 application on both the RQ and SQ.
2094  * @qp:            queue pair to drain
2095  *
2096  * The caller must:
2097  *
2098  * ensure there is room in the CQ(s), SQ, and RQ for drain work requests
2099  * and completions.
2100  *
2101  * allocate the CQs using ib_alloc_cq() and the CQ poll context cannot be
2102  * IB_POLL_DIRECT.
2103  *
2104  * ensure that there are no other contexts that are posting WRs concurrently.
2105  * Otherwise the drain is not guaranteed.
2106  */
2107 void ib_drain_qp(struct ib_qp *qp)
2108 {
2109         ib_drain_sq(qp);
2110         if (!qp->srq)
2111                 ib_drain_rq(qp);
2112 }
2113 EXPORT_SYMBOL(ib_drain_qp);