drivers/infiniband/hw/vmw_pvrdma/pvrdma_verbs.c

   1 /*
   2  * Copyright (c) 2012-2016 VMware, Inc.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of EITHER the GNU General Public License
   6  * version 2 as published by the Free Software Foundation or the BSD
   7  * 2-Clause License. This program is distributed in the hope that it
   8  * will be useful, but WITHOUT ANY WARRANTY; WITHOUT EVEN THE IMPLIED
   9  * WARRANTY OF MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE.
  10  * See the GNU General Public License version 2 for more details at
  11  * http://www.gnu.org/licenses/old-licenses/gpl-2.0.en.html.
  12  *
  13  * You should have received a copy of the GNU General Public License
  14  * along with this program available in the file COPYING in the main
  15  * directory of this source tree.
  16  *
  17  * The BSD 2-Clause License
  18  *
  19  *     Redistribution and use in source and binary forms, with or
  20  *     without modification, are permitted provided that the following
  21  *     conditions are met:
  22  *
  23  *      - Redistributions of source code must retain the above
  24  *        copyright notice, this list of conditions and the following
  25  *        disclaimer.
  26  *
  27  *      - Redistributions in binary form must reproduce the above
  28  *        copyright notice, this list of conditions and the following
  29  *        disclaimer in the documentation and/or other materials
  30  *        provided with the distribution.
  31  *
  32  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  33  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  34  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  35  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
  36  * COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT,
  37  * INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
  38  * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
  39  * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  40  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
  41  * STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
  42  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED
  43  * OF THE POSSIBILITY OF SUCH DAMAGE.
  44  */
  45
  46 #include <asm/page.h>
  47 #include <linux/inet.h>
  48 #include <linux/io.h>
  49 #include <rdma/ib_addr.h>
  50 #include <rdma/ib_smi.h>
  51 #include <rdma/ib_user_verbs.h>
  52 #include <rdma/vmw_pvrdma-abi.h>
  53
  54 #include "pvrdma.h"
  55
  56 /**
  57  * pvrdma_query_device - query device
  58  * @ibdev: the device to query
  59  * @props: the device properties
  60  * @uhw: user data
  61  *
  62  * @return: 0 on success, otherwise negative errno
  63  */
  64 int pvrdma_query_device(struct ib_device *ibdev,
  65                         struct ib_device_attr *props,
  66                         struct ib_udata *uhw)
  67 {
  68         struct pvrdma_dev *dev = to_vdev(ibdev);
  69
  70         if (uhw->inlen || uhw->outlen)
  71                 return -EINVAL;
  72
  73         memset(props, 0, sizeof(*props));
  74
  75         props->fw_ver = dev->dsr->caps.fw_ver;
  76         props->sys_image_guid = dev->dsr->caps.sys_image_guid;
  77         props->max_mr_size = dev->dsr->caps.max_mr_size;
  78         props->page_size_cap = dev->dsr->caps.page_size_cap;
  79         props->vendor_id = dev->dsr->caps.vendor_id;
  80         props->vendor_part_id = dev->pdev->device;
  81         props->hw_ver = dev->dsr->caps.hw_ver;
  82         props->max_qp = dev->dsr->caps.max_qp;
  83         props->max_qp_wr = dev->dsr->caps.max_qp_wr;
  84         props->device_cap_flags = dev->dsr->caps.device_cap_flags;
  85         props->max_sge = dev->dsr->caps.max_sge;
  86         props->max_sge_rd = PVRDMA_GET_CAP(dev, dev->dsr->caps.max_sge,
  87                                            dev->dsr->caps.max_sge_rd);
  88         props->max_srq = dev->dsr->caps.max_srq;
  89         props->max_srq_wr = dev->dsr->caps.max_srq_wr;
  90         props->max_srq_sge = dev->dsr->caps.max_srq_sge;
  91         props->max_cq = dev->dsr->caps.max_cq;
  92         props->max_cqe = dev->dsr->caps.max_cqe;
  93         props->max_mr = dev->dsr->caps.max_mr;
  94         props->max_pd = dev->dsr->caps.max_pd;
  95         props->max_qp_rd_atom = dev->dsr->caps.max_qp_rd_atom;
  96         props->max_qp_init_rd_atom = dev->dsr->caps.max_qp_init_rd_atom;
  97         props->atomic_cap =
  98                 dev->dsr->caps.atomic_ops &
  99                 (PVRDMA_ATOMIC_OP_COMP_SWAP | PVRDMA_ATOMIC_OP_FETCH_ADD) ?
 100                 IB_ATOMIC_HCA : IB_ATOMIC_NONE;
 101         props->masked_atomic_cap = props->atomic_cap;
 102         props->max_ah = dev->dsr->caps.max_ah;
 103         props->max_pkeys = dev->dsr->caps.max_pkeys;
 104         props->local_ca_ack_delay = dev->dsr->caps.local_ca_ack_delay;
 105         if ((dev->dsr->caps.bmme_flags & PVRDMA_BMME_FLAG_LOCAL_INV) &&
 106             (dev->dsr->caps.bmme_flags & PVRDMA_BMME_FLAG_REMOTE_INV) &&
 107             (dev->dsr->caps.bmme_flags & PVRDMA_BMME_FLAG_FAST_REG_WR)) {
 108                 props->device_cap_flags |= IB_DEVICE_MEM_MGT_EXTENSIONS;
 109                 props->max_fast_reg_page_list_len = PVRDMA_GET_CAP(dev,
 110                                 PVRDMA_MAX_FAST_REG_PAGES,
 111                                 dev->dsr->caps.max_fast_reg_page_list_len);
 112         }
 113
 114         props->device_cap_flags |= IB_DEVICE_PORT_ACTIVE_EVENT |
 115                                    IB_DEVICE_RC_RNR_NAK_GEN;
 116
 117         return 0;
 118 }
 119
 120 /**
 121  * pvrdma_query_port - query device port attributes
 122  * @ibdev: the device to query
 123  * @port: the port number
 124  * @props: the device properties
 125  *
 126  * @return: 0 on success, otherwise negative errno
 127  */
 128 int pvrdma_query_port(struct ib_device *ibdev, u8 port,
 129                       struct ib_port_attr *props)
 130 {
 131         struct pvrdma_dev *dev = to_vdev(ibdev);
 132         union pvrdma_cmd_req req;
 133         union pvrdma_cmd_resp rsp;
 134         struct pvrdma_cmd_query_port *cmd = &req.query_port;
 135         struct pvrdma_cmd_query_port_resp *resp = &rsp.query_port_resp;
 136         int err;
 137
 138         memset(cmd, 0, sizeof(*cmd));
 139         cmd->hdr.cmd = PVRDMA_CMD_QUERY_PORT;
 140         cmd->port_num = port;
 141
 142         err = pvrdma_cmd_post(dev, &req, &rsp, PVRDMA_CMD_QUERY_PORT_RESP);
 143         if (err < 0) {
 144                 dev_warn(&dev->pdev->dev,
 145                          "could not query port, error: %d\n", err);
 146                 return err;
 147         }
 148
 149         /* props being zeroed by the caller, avoid zeroing it here */
 150
 151         props->state = pvrdma_port_state_to_ib(resp->attrs.state);
 152         props->max_mtu = pvrdma_mtu_to_ib(resp->attrs.max_mtu);
 153         props->active_mtu = pvrdma_mtu_to_ib(resp->attrs.active_mtu);
 154         props->gid_tbl_len = resp->attrs.gid_tbl_len;
 155         props->port_cap_flags =
 156                 pvrdma_port_cap_flags_to_ib(resp->attrs.port_cap_flags);
 157         props->port_cap_flags |= IB_PORT_CM_SUP | IB_PORT_IP_BASED_GIDS;
 158         props->max_msg_sz = resp->attrs.max_msg_sz;
 159         props->bad_pkey_cntr = resp->attrs.bad_pkey_cntr;
 160         props->qkey_viol_cntr = resp->attrs.qkey_viol_cntr;
 161         props->pkey_tbl_len = resp->attrs.pkey_tbl_len;
 162         props->lid = resp->attrs.lid;
 163         props->sm_lid = resp->attrs.sm_lid;
 164         props->lmc = resp->attrs.lmc;
 165         props->max_vl_num = resp->attrs.max_vl_num;
 166         props->sm_sl = resp->attrs.sm_sl;
 167         props->subnet_timeout = resp->attrs.subnet_timeout;
 168         props->init_type_reply = resp->attrs.init_type_reply;
 169         props->active_width = pvrdma_port_width_to_ib(resp->attrs.active_width);
 170         props->active_speed = pvrdma_port_speed_to_ib(resp->attrs.active_speed);
 171         props->phys_state = resp->attrs.phys_state;
 172
 173         return 0;
 174 }
 175
 176 /**
 177  * pvrdma_query_gid - query device gid
 178  * @ibdev: the device to query
 179  * @port: the port number
 180  * @index: the index
 181  * @gid: the device gid value
 182  *
 183  * @return: 0 on success, otherwise negative errno
 184  */
 185 int pvrdma_query_gid(struct ib_device *ibdev, u8 port, int index,
 186                      union ib_gid *gid)
 187 {
 188         struct pvrdma_dev *dev = to_vdev(ibdev);
 189
 190         if (index >= dev->dsr->caps.gid_tbl_len)
 191                 return -EINVAL;
 192
 193         memcpy(gid, &dev->sgid_tbl[index], sizeof(union ib_gid));
 194
 195         return 0;
 196 }
 197
 198 /**
 199  * pvrdma_query_pkey - query device port's P_Key table
 200  * @ibdev: the device to query
 201  * @port: the port number
 202  * @index: the index
 203  * @pkey: the device P_Key value
 204  *
 205  * @return: 0 on success, otherwise negative errno
 206  */
 207 int pvrdma_query_pkey(struct ib_device *ibdev, u8 port, u16 index,
 208                       u16 *pkey)
 209 {
 210         int err = 0;
 211         union pvrdma_cmd_req req;
 212         union pvrdma_cmd_resp rsp;
 213         struct pvrdma_cmd_query_pkey *cmd = &req.query_pkey;
 214
 215         memset(cmd, 0, sizeof(*cmd));
 216         cmd->hdr.cmd = PVRDMA_CMD_QUERY_PKEY;
 217         cmd->port_num = port;
 218         cmd->index = index;
 219
 220         err = pvrdma_cmd_post(to_vdev(ibdev), &req, &rsp,
 221                               PVRDMA_CMD_QUERY_PKEY_RESP);
 222         if (err < 0) {
 223                 dev_warn(&to_vdev(ibdev)->pdev->dev,
 224                          "could not query pkey, error: %d\n", err);
 225                 return err;
 226         }
 227
 228         *pkey = rsp.query_pkey_resp.pkey;
 229
 230         return 0;
 231 }
 232
 233 enum rdma_link_layer pvrdma_port_link_layer(struct ib_device *ibdev,
 234                                             u8 port)
 235 {
 236         return IB_LINK_LAYER_ETHERNET;
 237 }
 238
 239 int pvrdma_modify_device(struct ib_device *ibdev, int mask,
 240                          struct ib_device_modify *props)
 241 {
 242         unsigned long flags;
 243
 244         if (mask & ~(IB_DEVICE_MODIFY_SYS_IMAGE_GUID |
 245                      IB_DEVICE_MODIFY_NODE_DESC)) {
 246                 dev_warn(&to_vdev(ibdev)->pdev->dev,
 247                          "unsupported device modify mask %#x\n", mask);
 248                 return -EOPNOTSUPP;
 249         }
 250
 251         if (mask & IB_DEVICE_MODIFY_NODE_DESC) {
 252                 spin_lock_irqsave(&to_vdev(ibdev)->desc_lock, flags);
 253                 memcpy(ibdev->node_desc, props->node_desc, 64);
 254                 spin_unlock_irqrestore(&to_vdev(ibdev)->desc_lock, flags);
 255         }
 256
 257         if (mask & IB_DEVICE_MODIFY_SYS_IMAGE_GUID) {
 258                 mutex_lock(&to_vdev(ibdev)->port_mutex);
 259                 to_vdev(ibdev)->sys_image_guid =
 260                         cpu_to_be64(props->sys_image_guid);
 261                 mutex_unlock(&to_vdev(ibdev)->port_mutex);
 262         }
 263
 264         return 0;
 265 }
 266
 267 /**
 268  * pvrdma_modify_port - modify device port attributes
 269  * @ibdev: the device to modify
 270  * @port: the port number
 271  * @mask: attributes to modify
 272  * @props: the device properties
 273  *
 274  * @return: 0 on success, otherwise negative errno
 275  */
 276 int pvrdma_modify_port(struct ib_device *ibdev, u8 port, int mask,
 277                        struct ib_port_modify *props)
 278 {
 279         struct ib_port_attr attr;
 280         struct pvrdma_dev *vdev = to_vdev(ibdev);
 281         int ret;
 282
 283         if (mask & ~IB_PORT_SHUTDOWN) {
 284                 dev_warn(&vdev->pdev->dev,
 285                          "unsupported port modify mask %#x\n", mask);
 286                 return -EOPNOTSUPP;
 287         }
 288
 289         mutex_lock(&vdev->port_mutex);
 290         ret = ib_query_port(ibdev, port, &attr);
 291         if (ret)
 292                 goto out;
 293
 294         vdev->port_cap_mask |= props->set_port_cap_mask;
 295         vdev->port_cap_mask &= ~props->clr_port_cap_mask;
 296
 297         if (mask & IB_PORT_SHUTDOWN)
 298                 vdev->ib_active = false;
 299
 300 out:
 301         mutex_unlock(&vdev->port_mutex);
 302         return ret;
 303 }
 304
 305 /**
 306  * pvrdma_alloc_ucontext - allocate ucontext
 307  * @ibdev: the IB device
 308  * @udata: user data
 309  *
 310  * @return: the ib_ucontext pointer on success, otherwise errno.
 311  */
 312 struct ib_ucontext *pvrdma_alloc_ucontext(struct ib_device *ibdev,
 313                                           struct ib_udata *udata)
 314 {
 315         struct pvrdma_dev *vdev = to_vdev(ibdev);
 316         struct pvrdma_ucontext *context;
 317         union pvrdma_cmd_req req;
 318         union pvrdma_cmd_resp rsp;
 319         struct pvrdma_cmd_create_uc *cmd = &req.create_uc;
 320         struct pvrdma_cmd_create_uc_resp *resp = &rsp.create_uc_resp;
 321         struct pvrdma_alloc_ucontext_resp uresp = {0};
 322         int ret;
 323         void *ptr;
 324
 325         if (!vdev->ib_active)
 326                 return ERR_PTR(-EAGAIN);
 327
 328         context = kmalloc(sizeof(*context), GFP_KERNEL);
 329         if (!context)
 330                 return ERR_PTR(-ENOMEM);
 331
 332         context->dev = vdev;
 333         ret = pvrdma_uar_alloc(vdev, &context->uar);
 334         if (ret) {
 335                 kfree(context);
 336                 return ERR_PTR(-ENOMEM);
 337         }
 338
 339         /* get ctx_handle from host */
 340         memset(cmd, 0, sizeof(*cmd));
 341         cmd->pfn = context->uar.pfn;
 342         cmd->hdr.cmd = PVRDMA_CMD_CREATE_UC;
 343         ret = pvrdma_cmd_post(vdev, &req, &rsp, PVRDMA_CMD_CREATE_UC_RESP);
 344         if (ret < 0) {
 345                 dev_warn(&vdev->pdev->dev,
 346                          "could not create ucontext, error: %d\n", ret);
 347                 ptr = ERR_PTR(ret);
 348                 goto err;
 349         }
 350
 351         context->ctx_handle = resp->ctx_handle;
 352
 353         /* copy back to user */
 354         uresp.qp_tab_size = vdev->dsr->caps.max_qp;
 355         ret = ib_copy_to_udata(udata, &uresp, sizeof(uresp));
 356         if (ret) {
 357                 pvrdma_uar_free(vdev, &context->uar);
 358                 context->ibucontext.device = ibdev;
 359                 pvrdma_dealloc_ucontext(&context->ibucontext);
 360                 return ERR_PTR(-EFAULT);
 361         }
 362
 363         return &context->ibucontext;
 364
 365 err:
 366         pvrdma_uar_free(vdev, &context->uar);
 367         kfree(context);
 368         return ptr;
 369 }
 370
 371 /**
 372  * pvrdma_dealloc_ucontext - deallocate ucontext
 373  * @ibcontext: the ucontext
 374  *
 375  * @return: 0 on success, otherwise errno.
 376  */
 377 int pvrdma_dealloc_ucontext(struct ib_ucontext *ibcontext)
 378 {
 379         struct pvrdma_ucontext *context = to_vucontext(ibcontext);
 380         union pvrdma_cmd_req req;
 381         struct pvrdma_cmd_destroy_uc *cmd = &req.destroy_uc;
 382         int ret;
 383
 384         memset(cmd, 0, sizeof(*cmd));
 385         cmd->hdr.cmd = PVRDMA_CMD_DESTROY_UC;
 386         cmd->ctx_handle = context->ctx_handle;
 387
 388         ret = pvrdma_cmd_post(context->dev, &req, NULL, 0);
 389         if (ret < 0)
 390                 dev_warn(&context->dev->pdev->dev,
 391                          "destroy ucontext failed, error: %d\n", ret);
 392
 393         /* Free the UAR even if the device command failed */
 394         pvrdma_uar_free(to_vdev(ibcontext->device), &context->uar);
 395         kfree(context);
 396
 397         return ret;
 398 }
 399
 400 /**
 401  * pvrdma_mmap - create mmap region
 402  * @ibcontext: the user context
 403  * @vma: the VMA
 404  *
 405  * @return: 0 on success, otherwise errno.
 406  */
 407 int pvrdma_mmap(struct ib_ucontext *ibcontext, struct vm_area_struct *vma)
 408 {
 409         struct pvrdma_ucontext *context = to_vucontext(ibcontext);
 410         unsigned long start = vma->vm_start;
 411         unsigned long size = vma->vm_end - vma->vm_start;
 412         unsigned long offset = vma->vm_pgoff << PAGE_SHIFT;
 413
 414         dev_dbg(&context->dev->pdev->dev, "create mmap region\n");
 415
 416         if ((size != PAGE_SIZE) || (offset & ~PAGE_MASK)) {
 417                 dev_warn(&context->dev->pdev->dev,
 418                          "invalid params for mmap region\n");
 419                 return -EINVAL;
 420         }
 421
 422         /* Map UAR to kernel space, VM_LOCKED? */
 423         vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
 424         vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
 425         if (io_remap_pfn_range(vma, start, context->uar.pfn, size,
 426                                vma->vm_page_prot))
 427                 return -EAGAIN;
 428
 429         return 0;
 430 }
 431
 432 /**
 433  * pvrdma_alloc_pd - allocate protection domain
 434  * @ibdev: the IB device
 435  * @context: user context
 436  * @udata: user data
 437  *
 438  * @return: the ib_pd protection domain pointer on success, otherwise errno.
 439  */
 440 struct ib_pd *pvrdma_alloc_pd(struct ib_device *ibdev,
 441                               struct ib_ucontext *context,
 442                               struct ib_udata *udata)
 443 {
 444         struct pvrdma_pd *pd;
 445         struct pvrdma_dev *dev = to_vdev(ibdev);
 446         union pvrdma_cmd_req req;
 447         union pvrdma_cmd_resp rsp;
 448         struct pvrdma_cmd_create_pd *cmd = &req.create_pd;
 449         struct pvrdma_cmd_create_pd_resp *resp = &rsp.create_pd_resp;
 450         int ret;
 451         void *ptr;
 452
 453         /* Check allowed max pds */
 454         if (!atomic_add_unless(&dev->num_pds, 1, dev->dsr->caps.max_pd))
 455                 return ERR_PTR(-ENOMEM);
 456
 457         pd = kmalloc(sizeof(*pd), GFP_KERNEL);
 458         if (!pd) {
 459                 ptr = ERR_PTR(-ENOMEM);
 460                 goto err;
 461         }
 462
 463         memset(cmd, 0, sizeof(*cmd));
 464         cmd->hdr.cmd = PVRDMA_CMD_CREATE_PD;
 465         cmd->ctx_handle = (context) ? to_vucontext(context)->ctx_handle : 0;
 466         ret = pvrdma_cmd_post(dev, &req, &rsp, PVRDMA_CMD_CREATE_PD_RESP);
 467         if (ret < 0) {
 468                 dev_warn(&dev->pdev->dev,
 469                          "failed to allocate protection domain, error: %d\n",
 470                          ret);
 471                 ptr = ERR_PTR(ret);
 472                 goto freepd;
 473         }
 474
 475         pd->privileged = !context;
 476         pd->pd_handle = resp->pd_handle;
 477         pd->pdn = resp->pd_handle;
 478
 479         if (context) {
 480                 if (ib_copy_to_udata(udata, &pd->pdn, sizeof(__u32))) {
 481                         dev_warn(&dev->pdev->dev,
 482                                  "failed to copy back protection domain\n");
 483                         pvrdma_dealloc_pd(&pd->ibpd);
 484                         return ERR_PTR(-EFAULT);
 485                 }
 486         }
 487
 488         /* u32 pd handle */
 489         return &pd->ibpd;
 490
 491 freepd:
 492         kfree(pd);
 493 err:
 494         atomic_dec(&dev->num_pds);
 495         return ptr;
 496 }
 497
 498 /**
 499  * pvrdma_dealloc_pd - deallocate protection domain
 500  * @pd: the protection domain to be released
 501  *
 502  * @return: 0 on success, otherwise errno.
 503  */
 504 int pvrdma_dealloc_pd(struct ib_pd *pd)
 505 {
 506         struct pvrdma_dev *dev = to_vdev(pd->device);
 507         union pvrdma_cmd_req req;
 508         struct pvrdma_cmd_destroy_pd *cmd = &req.destroy_pd;
 509         int ret;
 510
 511         memset(cmd, 0, sizeof(*cmd));
 512         cmd->hdr.cmd = PVRDMA_CMD_DESTROY_PD;
 513         cmd->pd_handle = to_vpd(pd)->pd_handle;
 514
 515         ret = pvrdma_cmd_post(dev, &req, NULL, 0);
 516         if (ret)
 517                 dev_warn(&dev->pdev->dev,
 518                          "could not dealloc protection domain, error: %d\n",
 519                          ret);
 520
 521         kfree(to_vpd(pd));
 522         atomic_dec(&dev->num_pds);
 523
 524         return 0;
 525 }
 526
 527 /**
 528  * pvrdma_create_ah - create an address handle
 529  * @pd: the protection domain
 530  * @ah_attr: the attributes of the AH
 531  * @udata: user data blob
 532  *
 533  * @return: the ib_ah pointer on success, otherwise errno.
 534  */
 535 struct ib_ah *pvrdma_create_ah(struct ib_pd *pd, struct rdma_ah_attr *ah_attr,
 536                                struct ib_udata *udata)
 537 {
 538         struct pvrdma_dev *dev = to_vdev(pd->device);
 539         struct pvrdma_ah *ah;
 540         const struct ib_global_route *grh;
 541         u8 port_num = rdma_ah_get_port_num(ah_attr);
 542
 543         if (!(rdma_ah_get_ah_flags(ah_attr) & IB_AH_GRH))
 544                 return ERR_PTR(-EINVAL);
 545
 546         grh = rdma_ah_read_grh(ah_attr);
 547         if ((ah_attr->type != RDMA_AH_ATTR_TYPE_ROCE)  ||
 548             rdma_is_multicast_addr((struct in6_addr *)grh->dgid.raw))
 549                 return ERR_PTR(-EINVAL);
 550
 551         if (!atomic_add_unless(&dev->num_ahs, 1, dev->dsr->caps.max_ah))
 552                 return ERR_PTR(-ENOMEM);
 553
 554         ah = kzalloc(sizeof(*ah), GFP_KERNEL);
 555         if (!ah) {
 556                 atomic_dec(&dev->num_ahs);
 557                 return ERR_PTR(-ENOMEM);
 558         }
 559
 560         ah->av.port_pd = to_vpd(pd)->pd_handle | (port_num << 24);
 561         ah->av.src_path_bits = rdma_ah_get_path_bits(ah_attr);
 562         ah->av.src_path_bits |= 0x80;
 563         ah->av.gid_index = grh->sgid_index;
 564         ah->av.hop_limit = grh->hop_limit;
 565         ah->av.sl_tclass_flowlabel = (grh->traffic_class << 20) |
 566                                       grh->flow_label;
 567         memcpy(ah->av.dgid, grh->dgid.raw, 16);
 568         memcpy(ah->av.dmac, ah_attr->roce.dmac, ETH_ALEN);
 569
 570         ah->ibah.device = pd->device;
 571         ah->ibah.pd = pd;
 572         ah->ibah.uobject = NULL;
 573
 574         return &ah->ibah;
 575 }
 576
 577 /**
 578  * pvrdma_destroy_ah - destroy an address handle
 579  * @ah: the address handle to destroyed
 580  *
 581  * @return: 0 on success.
 582  */
 583 int pvrdma_destroy_ah(struct ib_ah *ah)
 584 {
 585         struct pvrdma_dev *dev = to_vdev(ah->device);
 586
 587         kfree(to_vah(ah));
 588         atomic_dec(&dev->num_ahs);
 589
 590         return 0;
 591 }