drivers/infiniband/hw/mthca/mthca_provider.c

   1 /*
   2  * Copyright (c) 2004, 2005 Topspin Communications.  All rights reserved.
   3  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
   4  * Copyright (c) 2005, 2006 Cisco Systems.  All rights reserved.
   5  * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
   6  * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
   7  *
   8  * This software is available to you under a choice of one of two
   9  * licenses.  You may choose to be licensed under the terms of the GNU
  10  * General Public License (GPL) Version 2, available from the file
  11  * COPYING in the main directory of this source tree, or the
  12  * OpenIB.org BSD license below:
  13  *
  14  *     Redistribution and use in source and binary forms, with or
  15  *     without modification, are permitted provided that the following
  16  *     conditions are met:
  17  *
  18  *      - Redistributions of source code must retain the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer.
  21  *
  22  *      - Redistributions in binary form must reproduce the above
  23  *        copyright notice, this list of conditions and the following
  24  *        disclaimer in the documentation and/or other materials
  25  *        provided with the distribution.
  26  *
  27  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  28  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  29  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  30  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  31  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  32  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  33  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  34  * SOFTWARE.
  35  */
  36
  37 #include <rdma/ib_smi.h>
  38 #include <rdma/ib_umem.h>
  39 #include <rdma/ib_user_verbs.h>
  40
  41 #include <linux/sched.h>
  42 #include <linux/mm.h>
  43
  44 #include "mthca_dev.h"
  45 #include "mthca_cmd.h"
  46 #include "mthca_user.h"
  47 #include "mthca_memfree.h"
  48
  49 static void init_query_mad(struct ib_smp *mad)
  50 {
  51         mad->base_version  = 1;
  52         mad->mgmt_class    = IB_MGMT_CLASS_SUBN_LID_ROUTED;
  53         mad->class_version = 1;
  54         mad->method        = IB_MGMT_METHOD_GET;
  55 }
  56
  57 static int mthca_query_device(struct ib_device *ibdev,
  58                               struct ib_device_attr *props)
  59 {
  60         struct ib_smp *in_mad  = NULL;
  61         struct ib_smp *out_mad = NULL;
  62         int err = -ENOMEM;
  63         struct mthca_dev *mdev = to_mdev(ibdev);
  64
  65         u8 status;
  66
  67         in_mad  = kzalloc(sizeof *in_mad, GFP_KERNEL);
  68         out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
  69         if (!in_mad || !out_mad)
  70                 goto out;
  71
  72         memset(props, 0, sizeof *props);
  73
  74         props->fw_ver              = mdev->fw_ver;
  75
  76         init_query_mad(in_mad);
  77         in_mad->attr_id = IB_SMP_ATTR_NODE_INFO;
  78
  79         err = mthca_MAD_IFC(mdev, 1, 1,
  80                             1, NULL, NULL, in_mad, out_mad,
  81                             &status);
  82         if (err)
  83                 goto out;
  84         if (status) {
  85                 err = -EINVAL;
  86                 goto out;
  87         }
  88
  89         props->device_cap_flags    = mdev->device_cap_flags;
  90         props->vendor_id           = be32_to_cpup((__be32 *) (out_mad->data + 36)) &
  91                 0xffffff;
  92         props->vendor_part_id      = be16_to_cpup((__be16 *) (out_mad->data + 30));
  93         props->hw_ver              = be32_to_cpup((__be32 *) (out_mad->data + 32));
  94         memcpy(&props->sys_image_guid, out_mad->data +  4, 8);
  95
  96         props->max_mr_size         = ~0ull;
  97         props->page_size_cap       = mdev->limits.page_size_cap;
  98         props->max_qp              = mdev->limits.num_qps - mdev->limits.reserved_qps;
  99         props->max_qp_wr           = mdev->limits.max_wqes;
 100         props->max_sge             = mdev->limits.max_sg;
 101         props->max_cq              = mdev->limits.num_cqs - mdev->limits.reserved_cqs;
 102         props->max_cqe             = mdev->limits.max_cqes;
 103         props->max_mr              = mdev->limits.num_mpts - mdev->limits.reserved_mrws;
 104         props->max_pd              = mdev->limits.num_pds - mdev->limits.reserved_pds;
 105         props->max_qp_rd_atom      = 1 << mdev->qp_table.rdb_shift;
 106         props->max_qp_init_rd_atom = mdev->limits.max_qp_init_rdma;
 107         props->max_res_rd_atom     = props->max_qp_rd_atom * props->max_qp;
 108         props->max_srq             = mdev->limits.num_srqs - mdev->limits.reserved_srqs;
 109         props->max_srq_wr          = mdev->limits.max_srq_wqes;
 110         props->max_srq_sge         = mdev->limits.max_srq_sge;
 111         props->local_ca_ack_delay  = mdev->limits.local_ca_ack_delay;
 112         props->atomic_cap          = mdev->limits.flags & DEV_LIM_FLAG_ATOMIC ?
 113                                         IB_ATOMIC_HCA : IB_ATOMIC_NONE;
 114         props->max_pkeys           = mdev->limits.pkey_table_len;
 115         props->max_mcast_grp       = mdev->limits.num_mgms + mdev->limits.num_amgms;
 116         props->max_mcast_qp_attach = MTHCA_QP_PER_MGM;
 117         props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
 118                                            props->max_mcast_grp;
 119         /*
 120          * If Sinai memory key optimization is being used, then only
 121          * the 8-bit key portion will change.  For other HCAs, the
 122          * unused index bits will also be used for FMR remapping.
 123          */
 124         if (mdev->mthca_flags & MTHCA_FLAG_SINAI_OPT)
 125                 props->max_map_per_fmr = 255;
 126         else
 127                 props->max_map_per_fmr =
 128                         (1 << (32 - ilog2(mdev->limits.num_mpts))) - 1;
 129
 130         err = 0;
 131  out:
 132         kfree(in_mad);
 133         kfree(out_mad);
 134         return err;
 135 }
 136
 137 static int mthca_query_port(struct ib_device *ibdev,
 138                             u8 port, struct ib_port_attr *props)
 139 {
 140         struct ib_smp *in_mad  = NULL;
 141         struct ib_smp *out_mad = NULL;
 142         int err = -ENOMEM;
 143         u8 status;
 144
 145         in_mad  = kzalloc(sizeof *in_mad, GFP_KERNEL);
 146         out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
 147         if (!in_mad || !out_mad)
 148                 goto out;
 149
 150         memset(props, 0, sizeof *props);
 151
 152         init_query_mad(in_mad);
 153         in_mad->attr_id  = IB_SMP_ATTR_PORT_INFO;
 154         in_mad->attr_mod = cpu_to_be32(port);
 155
 156         err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
 157                             port, NULL, NULL, in_mad, out_mad,
 158                             &status);
 159         if (err)
 160                 goto out;
 161         if (status) {
 162                 err = -EINVAL;
 163                 goto out;
 164         }
 165
 166         props->lid               = be16_to_cpup((__be16 *) (out_mad->data + 16));
 167         props->lmc               = out_mad->data[34] & 0x7;
 168         props->sm_lid            = be16_to_cpup((__be16 *) (out_mad->data + 18));
 169         props->sm_sl             = out_mad->data[36] & 0xf;
 170         props->state             = out_mad->data[32] & 0xf;
 171         props->phys_state        = out_mad->data[33] >> 4;
 172         props->port_cap_flags    = be32_to_cpup((__be32 *) (out_mad->data + 20));
 173         props->gid_tbl_len       = to_mdev(ibdev)->limits.gid_table_len;
 174         props->max_msg_sz        = 0x80000000;
 175         props->pkey_tbl_len      = to_mdev(ibdev)->limits.pkey_table_len;
 176         props->bad_pkey_cntr     = be16_to_cpup((__be16 *) (out_mad->data + 46));
 177         props->qkey_viol_cntr    = be16_to_cpup((__be16 *) (out_mad->data + 48));
 178         props->active_width      = out_mad->data[31] & 0xf;
 179         props->active_speed      = out_mad->data[35] >> 4;
 180         props->max_mtu           = out_mad->data[41] & 0xf;
 181         props->active_mtu        = out_mad->data[36] >> 4;
 182         props->subnet_timeout    = out_mad->data[51] & 0x1f;
 183         props->max_vl_num        = out_mad->data[37] >> 4;
 184         props->init_type_reply   = out_mad->data[41] >> 4;
 185
 186  out:
 187         kfree(in_mad);
 188         kfree(out_mad);
 189         return err;
 190 }
 191
 192 static int mthca_modify_device(struct ib_device *ibdev,
 193                                int mask,
 194                                struct ib_device_modify *props)
 195 {
 196         if (mask & ~IB_DEVICE_MODIFY_NODE_DESC)
 197                 return -EOPNOTSUPP;
 198
 199         if (mask & IB_DEVICE_MODIFY_NODE_DESC) {
 200                 if (mutex_lock_interruptible(&to_mdev(ibdev)->cap_mask_mutex))
 201                         return -ERESTARTSYS;
 202                 memcpy(ibdev->node_desc, props->node_desc, 64);
 203                 mutex_unlock(&to_mdev(ibdev)->cap_mask_mutex);
 204         }
 205
 206         return 0;
 207 }
 208
 209 static int mthca_modify_port(struct ib_device *ibdev,
 210                              u8 port, int port_modify_mask,
 211                              struct ib_port_modify *props)
 212 {
 213         struct mthca_set_ib_param set_ib;
 214         struct ib_port_attr attr;
 215         int err;
 216         u8 status;
 217
 218         if (mutex_lock_interruptible(&to_mdev(ibdev)->cap_mask_mutex))
 219                 return -ERESTARTSYS;
 220
 221         err = mthca_query_port(ibdev, port, &attr);
 222         if (err)
 223                 goto out;
 224
 225         set_ib.set_si_guid     = 0;
 226         set_ib.reset_qkey_viol = !!(port_modify_mask & IB_PORT_RESET_QKEY_CNTR);
 227
 228         set_ib.cap_mask = (attr.port_cap_flags | props->set_port_cap_mask) &
 229                 ~props->clr_port_cap_mask;
 230
 231         err = mthca_SET_IB(to_mdev(ibdev), &set_ib, port, &status);
 232         if (err)
 233                 goto out;
 234         if (status) {
 235                 err = -EINVAL;
 236                 goto out;
 237         }
 238
 239 out:
 240         mutex_unlock(&to_mdev(ibdev)->cap_mask_mutex);
 241         return err;
 242 }
 243
 244 static int mthca_query_pkey(struct ib_device *ibdev,
 245                             u8 port, u16 index, u16 *pkey)
 246 {
 247         struct ib_smp *in_mad  = NULL;
 248         struct ib_smp *out_mad = NULL;
 249         int err = -ENOMEM;
 250         u8 status;
 251
 252         in_mad  = kzalloc(sizeof *in_mad, GFP_KERNEL);
 253         out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
 254         if (!in_mad || !out_mad)
 255                 goto out;
 256
 257         init_query_mad(in_mad);
 258         in_mad->attr_id  = IB_SMP_ATTR_PKEY_TABLE;
 259         in_mad->attr_mod = cpu_to_be32(index / 32);
 260
 261         err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
 262                             port, NULL, NULL, in_mad, out_mad,
 263                             &status);
 264         if (err)
 265                 goto out;
 266         if (status) {
 267                 err = -EINVAL;
 268                 goto out;
 269         }
 270
 271         *pkey = be16_to_cpu(((__be16 *) out_mad->data)[index % 32]);
 272
 273  out:
 274         kfree(in_mad);
 275         kfree(out_mad);
 276         return err;
 277 }
 278
 279 static int mthca_query_gid(struct ib_device *ibdev, u8 port,
 280                            int index, union ib_gid *gid)
 281 {
 282         struct ib_smp *in_mad  = NULL;
 283         struct ib_smp *out_mad = NULL;
 284         int err = -ENOMEM;
 285         u8 status;
 286
 287         in_mad  = kzalloc(sizeof *in_mad, GFP_KERNEL);
 288         out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
 289         if (!in_mad || !out_mad)
 290                 goto out;
 291
 292         init_query_mad(in_mad);
 293         in_mad->attr_id  = IB_SMP_ATTR_PORT_INFO;
 294         in_mad->attr_mod = cpu_to_be32(port);
 295
 296         err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
 297                             port, NULL, NULL, in_mad, out_mad,
 298                             &status);
 299         if (err)
 300                 goto out;
 301         if (status) {
 302                 err = -EINVAL;
 303                 goto out;
 304         }
 305
 306         memcpy(gid->raw, out_mad->data + 8, 8);
 307
 308         init_query_mad(in_mad);
 309         in_mad->attr_id  = IB_SMP_ATTR_GUID_INFO;
 310         in_mad->attr_mod = cpu_to_be32(index / 8);
 311
 312         err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
 313                             port, NULL, NULL, in_mad, out_mad,
 314                             &status);
 315         if (err)
 316                 goto out;
 317         if (status) {
 318                 err = -EINVAL;
 319                 goto out;
 320         }
 321
 322         memcpy(gid->raw + 8, out_mad->data + (index % 8) * 8, 8);
 323
 324  out:
 325         kfree(in_mad);
 326         kfree(out_mad);
 327         return err;
 328 }
 329
 330 static struct ib_ucontext *mthca_alloc_ucontext(struct ib_device *ibdev,
 331                                                 struct ib_udata *udata)
 332 {
 333         struct mthca_alloc_ucontext_resp uresp;
 334         struct mthca_ucontext           *context;
 335         int                              err;
 336
 337         memset(&uresp, 0, sizeof uresp);
 338
 339         uresp.qp_tab_size = to_mdev(ibdev)->limits.num_qps;
 340         if (mthca_is_memfree(to_mdev(ibdev)))
 341                 uresp.uarc_size = to_mdev(ibdev)->uar_table.uarc_size;
 342         else
 343                 uresp.uarc_size = 0;
 344
 345         context = kmalloc(sizeof *context, GFP_KERNEL);
 346         if (!context)
 347                 return ERR_PTR(-ENOMEM);
 348
 349         err = mthca_uar_alloc(to_mdev(ibdev), &context->uar);
 350         if (err) {
 351                 kfree(context);
 352                 return ERR_PTR(err);
 353         }
 354
 355         context->db_tab = mthca_init_user_db_tab(to_mdev(ibdev));
 356         if (IS_ERR(context->db_tab)) {
 357                 err = PTR_ERR(context->db_tab);
 358                 mthca_uar_free(to_mdev(ibdev), &context->uar);
 359                 kfree(context);
 360                 return ERR_PTR(err);
 361         }
 362
 363         if (ib_copy_to_udata(udata, &uresp, sizeof uresp)) {
 364                 mthca_cleanup_user_db_tab(to_mdev(ibdev), &context->uar, context->db_tab);
 365                 mthca_uar_free(to_mdev(ibdev), &context->uar);
 366                 kfree(context);
 367                 return ERR_PTR(-EFAULT);
 368         }
 369
 370         context->reg_mr_warned = 0;
 371
 372         return &context->ibucontext;
 373 }
 374
 375 static int mthca_dealloc_ucontext(struct ib_ucontext *context)
 376 {
 377         mthca_cleanup_user_db_tab(to_mdev(context->device), &to_mucontext(context)->uar,
 378                                   to_mucontext(context)->db_tab);
 379         mthca_uar_free(to_mdev(context->device), &to_mucontext(context)->uar);
 380         kfree(to_mucontext(context));
 381
 382         return 0;
 383 }
 384
 385 static int mthca_mmap_uar(struct ib_ucontext *context,
 386                           struct vm_area_struct *vma)
 387 {
 388         if (vma->vm_end - vma->vm_start != PAGE_SIZE)
 389                 return -EINVAL;
 390
 391         vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
 392
 393         if (io_remap_pfn_range(vma, vma->vm_start,
 394                                to_mucontext(context)->uar.pfn,
 395                                PAGE_SIZE, vma->vm_page_prot))
 396                 return -EAGAIN;
 397
 398         return 0;
 399 }
 400
 401 static struct ib_pd *mthca_alloc_pd(struct ib_device *ibdev,
 402                                     struct ib_ucontext *context,
 403                                     struct ib_udata *udata)
 404 {
 405         struct mthca_pd *pd;
 406         int err;
 407
 408         pd = kmalloc(sizeof *pd, GFP_KERNEL);
 409         if (!pd)
 410                 return ERR_PTR(-ENOMEM);
 411
 412         err = mthca_pd_alloc(to_mdev(ibdev), !context, pd);
 413         if (err) {
 414                 kfree(pd);
 415                 return ERR_PTR(err);
 416         }
 417
 418         if (context) {
 419                 if (ib_copy_to_udata(udata, &pd->pd_num, sizeof (__u32))) {
 420                         mthca_pd_free(to_mdev(ibdev), pd);
 421                         kfree(pd);
 422                         return ERR_PTR(-EFAULT);
 423                 }
 424         }
 425
 426         return &pd->ibpd;
 427 }
 428
 429 static int mthca_dealloc_pd(struct ib_pd *pd)
 430 {
 431         mthca_pd_free(to_mdev(pd->device), to_mpd(pd));
 432         kfree(pd);
 433
 434         return 0;
 435 }
 436
 437 static struct ib_ah *mthca_ah_create(struct ib_pd *pd,
 438                                      struct ib_ah_attr *ah_attr)
 439 {
 440         int err;
 441         struct mthca_ah *ah;
 442
 443         ah = kmalloc(sizeof *ah, GFP_ATOMIC);
 444         if (!ah)
 445                 return ERR_PTR(-ENOMEM);
 446
 447         err = mthca_create_ah(to_mdev(pd->device), to_mpd(pd), ah_attr, ah);
 448         if (err) {
 449                 kfree(ah);
 450                 return ERR_PTR(err);
 451         }
 452
 453         return &ah->ibah;
 454 }
 455
 456 static int mthca_ah_destroy(struct ib_ah *ah)
 457 {
 458         mthca_destroy_ah(to_mdev(ah->device), to_mah(ah));
 459         kfree(ah);
 460
 461         return 0;
 462 }
 463
 464 static struct ib_srq *mthca_create_srq(struct ib_pd *pd,
 465                                        struct ib_srq_init_attr *init_attr,
 466                                        struct ib_udata *udata)
 467 {
 468         struct mthca_create_srq ucmd;
 469         struct mthca_ucontext *context = NULL;
 470         struct mthca_srq *srq;
 471         int err;
 472
 473         srq = kmalloc(sizeof *srq, GFP_KERNEL);
 474         if (!srq)
 475                 return ERR_PTR(-ENOMEM);
 476
 477         if (pd->uobject) {
 478                 context = to_mucontext(pd->uobject->context);
 479
 480                 if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
 481                         err = -EFAULT;
 482                         goto err_free;
 483                 }
 484
 485                 err = mthca_map_user_db(to_mdev(pd->device), &context->uar,
 486                                         context->db_tab, ucmd.db_index,
 487                                         ucmd.db_page);
 488
 489                 if (err)
 490                         goto err_free;
 491
 492                 srq->mr.ibmr.lkey = ucmd.lkey;
 493                 srq->db_index     = ucmd.db_index;
 494         }
 495
 496         err = mthca_alloc_srq(to_mdev(pd->device), to_mpd(pd),
 497                               &init_attr->attr, srq);
 498
 499         if (err && pd->uobject)
 500                 mthca_unmap_user_db(to_mdev(pd->device), &context->uar,
 501                                     context->db_tab, ucmd.db_index);
 502
 503         if (err)
 504                 goto err_free;
 505
 506         if (context && ib_copy_to_udata(udata, &srq->srqn, sizeof (__u32))) {
 507                 mthca_free_srq(to_mdev(pd->device), srq);
 508                 err = -EFAULT;
 509                 goto err_free;
 510         }
 511
 512         return &srq->ibsrq;
 513
 514 err_free:
 515         kfree(srq);
 516
 517         return ERR_PTR(err);
 518 }
 519
 520 static int mthca_destroy_srq(struct ib_srq *srq)
 521 {
 522         struct mthca_ucontext *context;
 523
 524         if (srq->uobject) {
 525                 context = to_mucontext(srq->uobject->context);
 526
 527                 mthca_unmap_user_db(to_mdev(srq->device), &context->uar,
 528                                     context->db_tab, to_msrq(srq)->db_index);
 529         }
 530
 531         mthca_free_srq(to_mdev(srq->device), to_msrq(srq));
 532         kfree(srq);
 533
 534         return 0;
 535 }
 536
 537 static struct ib_qp *mthca_create_qp(struct ib_pd *pd,
 538                                      struct ib_qp_init_attr *init_attr,
 539                                      struct ib_udata *udata)
 540 {
 541         struct mthca_create_qp ucmd;
 542         struct mthca_qp *qp;
 543         int err;
 544
 545         if (init_attr->create_flags)
 546                 return ERR_PTR(-EINVAL);
 547
 548         switch (init_attr->qp_type) {
 549         case IB_QPT_RC:
 550         case IB_QPT_UC:
 551         case IB_QPT_UD:
 552         {
 553                 struct mthca_ucontext *context;
 554
 555                 qp = kmalloc(sizeof *qp, GFP_KERNEL);
 556                 if (!qp)
 557                         return ERR_PTR(-ENOMEM);
 558
 559                 if (pd->uobject) {
 560                         context = to_mucontext(pd->uobject->context);
 561
 562                         if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
 563                                 kfree(qp);
 564                                 return ERR_PTR(-EFAULT);
 565                         }
 566
 567                         err = mthca_map_user_db(to_mdev(pd->device), &context->uar,
 568                                                 context->db_tab,
 569                                                 ucmd.sq_db_index, ucmd.sq_db_page);
 570                         if (err) {
 571                                 kfree(qp);
 572                                 return ERR_PTR(err);
 573                         }
 574
 575                         err = mthca_map_user_db(to_mdev(pd->device), &context->uar,
 576                                                 context->db_tab,
 577                                                 ucmd.rq_db_index, ucmd.rq_db_page);
 578                         if (err) {
 579                                 mthca_unmap_user_db(to_mdev(pd->device),
 580                                                     &context->uar,
 581                                                     context->db_tab,
 582                                                     ucmd.sq_db_index);
 583                                 kfree(qp);
 584                                 return ERR_PTR(err);
 585                         }
 586
 587                         qp->mr.ibmr.lkey = ucmd.lkey;
 588                         qp->sq.db_index  = ucmd.sq_db_index;
 589                         qp->rq.db_index  = ucmd.rq_db_index;
 590                 }
 591
 592                 err = mthca_alloc_qp(to_mdev(pd->device), to_mpd(pd),
 593                                      to_mcq(init_attr->send_cq),
 594                                      to_mcq(init_attr->recv_cq),
 595                                      init_attr->qp_type, init_attr->sq_sig_type,
 596                                      &init_attr->cap, qp);
 597
 598                 if (err && pd->uobject) {
 599                         context = to_mucontext(pd->uobject->context);
 600
 601                         mthca_unmap_user_db(to_mdev(pd->device),
 602                                             &context->uar,
 603                                             context->db_tab,
 604                                             ucmd.sq_db_index);
 605                         mthca_unmap_user_db(to_mdev(pd->device),
 606                                             &context->uar,
 607                                             context->db_tab,
 608                                             ucmd.rq_db_index);
 609                 }
 610
 611                 qp->ibqp.qp_num = qp->qpn;
 612                 break;
 613         }
 614         case IB_QPT_SMI:
 615         case IB_QPT_GSI:
 616         {
 617                 /* Don't allow userspace to create special QPs */
 618                 if (pd->uobject)
 619                         return ERR_PTR(-EINVAL);
 620
 621                 qp = kmalloc(sizeof (struct mthca_sqp), GFP_KERNEL);
 622                 if (!qp)
 623                         return ERR_PTR(-ENOMEM);
 624
 625                 qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 : 1;
 626
 627                 err = mthca_alloc_sqp(to_mdev(pd->device), to_mpd(pd),
 628                                       to_mcq(init_attr->send_cq),
 629                                       to_mcq(init_attr->recv_cq),
 630                                       init_attr->sq_sig_type, &init_attr->cap,
 631                                       qp->ibqp.qp_num, init_attr->port_num,
 632                                       to_msqp(qp));
 633                 break;
 634         }
 635         default:
 636                 /* Don't support raw QPs */
 637                 return ERR_PTR(-ENOSYS);
 638         }
 639
 640         if (err) {
 641                 kfree(qp);
 642                 return ERR_PTR(err);
 643         }
 644
 645         init_attr->cap.max_send_wr     = qp->sq.max;
 646         init_attr->cap.max_recv_wr     = qp->rq.max;
 647         init_attr->cap.max_send_sge    = qp->sq.max_gs;
 648         init_attr->cap.max_recv_sge    = qp->rq.max_gs;
 649         init_attr->cap.max_inline_data = qp->max_inline_data;
 650
 651         return &qp->ibqp;
 652 }
 653
 654 static int mthca_destroy_qp(struct ib_qp *qp)
 655 {
 656         if (qp->uobject) {
 657                 mthca_unmap_user_db(to_mdev(qp->device),
 658                                     &to_mucontext(qp->uobject->context)->uar,
 659                                     to_mucontext(qp->uobject->context)->db_tab,
 660                                     to_mqp(qp)->sq.db_index);
 661                 mthca_unmap_user_db(to_mdev(qp->device),
 662                                     &to_mucontext(qp->uobject->context)->uar,
 663                                     to_mucontext(qp->uobject->context)->db_tab,
 664                                     to_mqp(qp)->rq.db_index);
 665         }
 666         mthca_free_qp(to_mdev(qp->device), to_mqp(qp));
 667         kfree(qp);
 668         return 0;
 669 }
 670
 671 static struct ib_cq *mthca_create_cq(struct ib_device *ibdev, int entries,
 672                                      int comp_vector,
 673                                      struct ib_ucontext *context,
 674                                      struct ib_udata *udata)
 675 {
 676         struct mthca_create_cq ucmd;
 677         struct mthca_cq *cq;
 678         int nent;
 679         int err;
 680
 681         if (entries < 1 || entries > to_mdev(ibdev)->limits.max_cqes)
 682                 return ERR_PTR(-EINVAL);
 683
 684         if (context) {
 685                 if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd))
 686                         return ERR_PTR(-EFAULT);
 687
 688                 err = mthca_map_user_db(to_mdev(ibdev), &to_mucontext(context)->uar,
 689                                         to_mucontext(context)->db_tab,
 690                                         ucmd.set_db_index, ucmd.set_db_page);
 691                 if (err)
 692                         return ERR_PTR(err);
 693
 694                 err = mthca_map_user_db(to_mdev(ibdev), &to_mucontext(context)->uar,
 695                                         to_mucontext(context)->db_tab,
 696                                         ucmd.arm_db_index, ucmd.arm_db_page);
 697                 if (err)
 698                         goto err_unmap_set;
 699         }
 700
 701         cq = kmalloc(sizeof *cq, GFP_KERNEL);
 702         if (!cq) {
 703                 err = -ENOMEM;
 704                 goto err_unmap_arm;
 705         }
 706
 707         if (context) {
 708                 cq->buf.mr.ibmr.lkey = ucmd.lkey;
 709                 cq->set_ci_db_index  = ucmd.set_db_index;
 710                 cq->arm_db_index     = ucmd.arm_db_index;
 711         }
 712
 713         for (nent = 1; nent <= entries; nent <<= 1)
 714                 ; /* nothing */
 715
 716         err = mthca_init_cq(to_mdev(ibdev), nent,
 717                             context ? to_mucontext(context) : NULL,
 718                             context ? ucmd.pdn : to_mdev(ibdev)->driver_pd.pd_num,
 719                             cq);
 720         if (err)
 721                 goto err_free;
 722
 723         if (context && ib_copy_to_udata(udata, &cq->cqn, sizeof (__u32))) {
 724                 mthca_free_cq(to_mdev(ibdev), cq);
 725                 goto err_free;
 726         }
 727
 728         cq->resize_buf = NULL;
 729
 730         return &cq->ibcq;
 731
 732 err_free:
 733         kfree(cq);
 734
 735 err_unmap_arm:
 736         if (context)
 737                 mthca_unmap_user_db(to_mdev(ibdev), &to_mucontext(context)->uar,
 738                                     to_mucontext(context)->db_tab, ucmd.arm_db_index);
 739
 740 err_unmap_set:
 741         if (context)
 742                 mthca_unmap_user_db(to_mdev(ibdev), &to_mucontext(context)->uar,
 743                                     to_mucontext(context)->db_tab, ucmd.set_db_index);
 744
 745         return ERR_PTR(err);
 746 }
 747
 748 static int mthca_alloc_resize_buf(struct mthca_dev *dev, struct mthca_cq *cq,
 749                                   int entries)
 750 {
 751         int ret;
 752
 753         spin_lock_irq(&cq->lock);
 754         if (cq->resize_buf) {
 755                 ret = -EBUSY;
 756                 goto unlock;
 757         }
 758
 759         cq->resize_buf = kmalloc(sizeof *cq->resize_buf, GFP_ATOMIC);
 760         if (!cq->resize_buf) {
 761                 ret = -ENOMEM;
 762                 goto unlock;
 763         }
 764
 765         cq->resize_buf->state = CQ_RESIZE_ALLOC;
 766
 767         ret = 0;
 768
 769 unlock:
 770         spin_unlock_irq(&cq->lock);
 771
 772         if (ret)
 773                 return ret;
 774
 775         ret = mthca_alloc_cq_buf(dev, &cq->resize_buf->buf, entries);
 776         if (ret) {
 777                 spin_lock_irq(&cq->lock);
 778                 kfree(cq->resize_buf);
 779                 cq->resize_buf = NULL;
 780                 spin_unlock_irq(&cq->lock);
 781                 return ret;
 782         }
 783
 784         cq->resize_buf->cqe = entries - 1;
 785
 786         spin_lock_irq(&cq->lock);
 787         cq->resize_buf->state = CQ_RESIZE_READY;
 788         spin_unlock_irq(&cq->lock);
 789
 790         return 0;
 791 }
 792
 793 static int mthca_resize_cq(struct ib_cq *ibcq, int entries, struct ib_udata *udata)
 794 {
 795         struct mthca_dev *dev = to_mdev(ibcq->device);
 796         struct mthca_cq *cq = to_mcq(ibcq);
 797         struct mthca_resize_cq ucmd;
 798         u32 lkey;
 799         u8 status;
 800         int ret;
 801
 802         if (entries < 1 || entries > dev->limits.max_cqes)
 803                 return -EINVAL;
 804
 805         mutex_lock(&cq->mutex);
 806
 807         entries = roundup_pow_of_two(entries + 1);
 808         if (entries == ibcq->cqe + 1) {
 809                 ret = 0;
 810                 goto out;
 811         }
 812
 813         if (cq->is_kernel) {
 814                 ret = mthca_alloc_resize_buf(dev, cq, entries);
 815                 if (ret)
 816                         goto out;
 817                 lkey = cq->resize_buf->buf.mr.ibmr.lkey;
 818         } else {
 819                 if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
 820                         ret = -EFAULT;
 821                         goto out;
 822                 }
 823                 lkey = ucmd.lkey;
 824         }
 825
 826         ret = mthca_RESIZE_CQ(dev, cq->cqn, lkey, ilog2(entries), &status);
 827         if (status)
 828                 ret = -EINVAL;
 829
 830         if (ret) {
 831                 if (cq->resize_buf) {
 832                         mthca_free_cq_buf(dev, &cq->resize_buf->buf,
 833                                           cq->resize_buf->cqe);
 834                         kfree(cq->resize_buf);
 835                         spin_lock_irq(&cq->lock);
 836                         cq->resize_buf = NULL;
 837                         spin_unlock_irq(&cq->lock);
 838                 }
 839                 goto out;
 840         }
 841
 842         if (cq->is_kernel) {
 843                 struct mthca_cq_buf tbuf;
 844                 int tcqe;
 845
 846                 spin_lock_irq(&cq->lock);
 847                 if (cq->resize_buf->state == CQ_RESIZE_READY) {
 848                         mthca_cq_resize_copy_cqes(cq);
 849                         tbuf         = cq->buf;
 850                         tcqe         = cq->ibcq.cqe;
 851                         cq->buf      = cq->resize_buf->buf;
 852                         cq->ibcq.cqe = cq->resize_buf->cqe;
 853                 } else {
 854                         tbuf = cq->resize_buf->buf;
 855                         tcqe = cq->resize_buf->cqe;
 856                 }
 857
 858                 kfree(cq->resize_buf);
 859                 cq->resize_buf = NULL;
 860                 spin_unlock_irq(&cq->lock);
 861
 862                 mthca_free_cq_buf(dev, &tbuf, tcqe);
 863         } else
 864                 ibcq->cqe = entries - 1;
 865
 866 out:
 867         mutex_unlock(&cq->mutex);
 868
 869         return ret;
 870 }
 871
 872 static int mthca_destroy_cq(struct ib_cq *cq)
 873 {
 874         if (cq->uobject) {
 875                 mthca_unmap_user_db(to_mdev(cq->device),
 876                                     &to_mucontext(cq->uobject->context)->uar,
 877                                     to_mucontext(cq->uobject->context)->db_tab,
 878                                     to_mcq(cq)->arm_db_index);
 879                 mthca_unmap_user_db(to_mdev(cq->device),
 880                                     &to_mucontext(cq->uobject->context)->uar,
 881                                     to_mucontext(cq->uobject->context)->db_tab,
 882                                     to_mcq(cq)->set_ci_db_index);
 883         }
 884         mthca_free_cq(to_mdev(cq->device), to_mcq(cq));
 885         kfree(cq);
 886
 887         return 0;
 888 }
 889
 890 static inline u32 convert_access(int acc)
 891 {
 892         return (acc & IB_ACCESS_REMOTE_ATOMIC ? MTHCA_MPT_FLAG_ATOMIC       : 0) |
 893                (acc & IB_ACCESS_REMOTE_WRITE  ? MTHCA_MPT_FLAG_REMOTE_WRITE : 0) |
 894                (acc & IB_ACCESS_REMOTE_READ   ? MTHCA_MPT_FLAG_REMOTE_READ  : 0) |
 895                (acc & IB_ACCESS_LOCAL_WRITE   ? MTHCA_MPT_FLAG_LOCAL_WRITE  : 0) |
 896                MTHCA_MPT_FLAG_LOCAL_READ;
 897 }
 898
 899 static struct ib_mr *mthca_get_dma_mr(struct ib_pd *pd, int acc)
 900 {
 901         struct mthca_mr *mr;
 902         int err;
 903
 904         mr = kmalloc(sizeof *mr, GFP_KERNEL);
 905         if (!mr)
 906                 return ERR_PTR(-ENOMEM);
 907
 908         err = mthca_mr_alloc_notrans(to_mdev(pd->device),
 909                                      to_mpd(pd)->pd_num,
 910                                      convert_access(acc), mr);
 911
 912         if (err) {
 913                 kfree(mr);
 914                 return ERR_PTR(err);
 915         }
 916
 917         mr->umem = NULL;
 918
 919         return &mr->ibmr;
 920 }
 921
 922 static struct ib_mr *mthca_reg_phys_mr(struct ib_pd       *pd,
 923                                        struct ib_phys_buf *buffer_list,
 924                                        int                 num_phys_buf,
 925                                        int                 acc,
 926                                        u64                *iova_start)
 927 {
 928         struct mthca_mr *mr;
 929         u64 *page_list;
 930         u64 total_size;
 931         unsigned long mask;
 932         int shift;
 933         int npages;
 934         int err;
 935         int i, j, n;
 936
 937         mask = buffer_list[0].addr ^ *iova_start;
 938         total_size = 0;
 939         for (i = 0; i < num_phys_buf; ++i) {
 940                 if (i != 0)
 941                         mask |= buffer_list[i].addr;
 942                 if (i != num_phys_buf - 1)
 943                         mask |= buffer_list[i].addr + buffer_list[i].size;
 944
 945                 total_size += buffer_list[i].size;
 946         }
 947
 948         if (mask & ~PAGE_MASK)
 949                 return ERR_PTR(-EINVAL);
 950
 951         shift = __ffs(mask | 1 << 31);
 952
 953         buffer_list[0].size += buffer_list[0].addr & ((1ULL << shift) - 1);
 954         buffer_list[0].addr &= ~0ull << shift;
 955
 956         mr = kmalloc(sizeof *mr, GFP_KERNEL);
 957         if (!mr)
 958                 return ERR_PTR(-ENOMEM);
 959
 960         npages = 0;
 961         for (i = 0; i < num_phys_buf; ++i)
 962                 npages += (buffer_list[i].size + (1ULL << shift) - 1) >> shift;
 963
 964         if (!npages)
 965                 return &mr->ibmr;
 966
 967         page_list = kmalloc(npages * sizeof *page_list, GFP_KERNEL);
 968         if (!page_list) {
 969                 kfree(mr);
 970                 return ERR_PTR(-ENOMEM);
 971         }
 972
 973         n = 0;
 974         for (i = 0; i < num_phys_buf; ++i)
 975                 for (j = 0;
 976                      j < (buffer_list[i].size + (1ULL << shift) - 1) >> shift;
 977                      ++j)
 978                         page_list[n++] = buffer_list[i].addr + ((u64) j << shift);
 979
 980         mthca_dbg(to_mdev(pd->device), "Registering memory at %llx (iova %llx) "
 981                   "in PD %x; shift %d, npages %d.\n",
 982                   (unsigned long long) buffer_list[0].addr,
 983                   (unsigned long long) *iova_start,
 984                   to_mpd(pd)->pd_num,
 985                   shift, npages);
 986
 987         err = mthca_mr_alloc_phys(to_mdev(pd->device),
 988                                   to_mpd(pd)->pd_num,
 989                                   page_list, shift, npages,
 990                                   *iova_start, total_size,
 991                                   convert_access(acc), mr);
 992
 993         if (err) {
 994                 kfree(page_list);
 995                 kfree(mr);
 996                 return ERR_PTR(err);
 997         }
 998
 999         kfree(page_list);
1000         mr->umem = NULL;
1001
1002         return &mr->ibmr;
1003 }
1004
1005 static struct ib_mr *mthca_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
1006                                        u64 virt, int acc, struct ib_udata *udata)
1007 {
1008         struct mthca_dev *dev = to_mdev(pd->device);
1009         struct ib_umem_chunk *chunk;
1010         struct mthca_mr *mr;
1011         struct mthca_reg_mr ucmd;
1012         u64 *pages;
1013         int shift, n, len;
1014         int i, j, k;
1015         int err = 0;
1016         int write_mtt_size;
1017
1018         if (udata->inlen - sizeof (struct ib_uverbs_cmd_hdr) < sizeof ucmd) {
1019                 if (!to_mucontext(pd->uobject->context)->reg_mr_warned) {
1020                         mthca_warn(dev, "Process '%s' did not pass in MR attrs.\n",
1021                                    current->comm);
1022                         mthca_warn(dev, "  Update libmthca to fix this.\n");
1023                 }
1024                 ++to_mucontext(pd->uobject->context)->reg_mr_warned;
1025                 ucmd.mr_attrs = 0;
1026         } else if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd))
1027                 return ERR_PTR(-EFAULT);
1028
1029         mr = kmalloc(sizeof *mr, GFP_KERNEL);
1030         if (!mr)
1031                 return ERR_PTR(-ENOMEM);
1032
1033         mr->umem = ib_umem_get(pd->uobject->context, start, length, acc,
1034                                ucmd.mr_attrs & MTHCA_MR_DMASYNC);
1035
1036         if (IS_ERR(mr->umem)) {
1037                 err = PTR_ERR(mr->umem);
1038                 goto err;
1039         }
1040
1041         shift = ffs(mr->umem->page_size) - 1;
1042
1043         n = 0;
1044         list_for_each_entry(chunk, &mr->umem->chunk_list, list)
1045                 n += chunk->nents;
1046
1047         mr->mtt = mthca_alloc_mtt(dev, n);
1048         if (IS_ERR(mr->mtt)) {
1049                 err = PTR_ERR(mr->mtt);
1050                 goto err_umem;
1051         }
1052
1053         pages = (u64 *) __get_free_page(GFP_KERNEL);
1054         if (!pages) {
1055                 err = -ENOMEM;
1056                 goto err_mtt;
1057         }
1058
1059         i = n = 0;
1060
1061         write_mtt_size = min(mthca_write_mtt_size(dev), (int) (PAGE_SIZE / sizeof *pages));
1062
1063         list_for_each_entry(chunk, &mr->umem->chunk_list, list)
1064                 for (j = 0; j < chunk->nmap; ++j) {
1065                         len = sg_dma_len(&chunk->page_list[j]) >> shift;
1066                         for (k = 0; k < len; ++k) {
1067                                 pages[i++] = sg_dma_address(&chunk->page_list[j]) +
1068                                         mr->umem->page_size * k;
1069                                 /*
1070                                  * Be friendly to write_mtt and pass it chunks
1071                                  * of appropriate size.
1072                                  */
1073                                 if (i == write_mtt_size) {
1074                                         err = mthca_write_mtt(dev, mr->mtt, n, pages, i);
1075                                         if (err)
1076                                                 goto mtt_done;
1077                                         n += i;
1078                                         i = 0;
1079                                 }
1080                         }
1081                 }
1082
1083         if (i)
1084                 err = mthca_write_mtt(dev, mr->mtt, n, pages, i);
1085 mtt_done:
1086         free_page((unsigned long) pages);
1087         if (err)
1088                 goto err_mtt;
1089
1090         err = mthca_mr_alloc(dev, to_mpd(pd)->pd_num, shift, virt, length,
1091                              convert_access(acc), mr);
1092
1093         if (err)
1094                 goto err_mtt;
1095
1096         return &mr->ibmr;
1097
1098 err_mtt:
1099         mthca_free_mtt(dev, mr->mtt);
1100
1101 err_umem:
1102         ib_umem_release(mr->umem);
1103
1104 err:
1105         kfree(mr);
1106         return ERR_PTR(err);
1107 }
1108
1109 static int mthca_dereg_mr(struct ib_mr *mr)
1110 {
1111         struct mthca_mr *mmr = to_mmr(mr);
1112
1113         mthca_free_mr(to_mdev(mr->device), mmr);
1114         if (mmr->umem)
1115                 ib_umem_release(mmr->umem);
1116         kfree(mmr);
1117
1118         return 0;
1119 }
1120
1121 static struct ib_fmr *mthca_alloc_fmr(struct ib_pd *pd, int mr_access_flags,
1122                                       struct ib_fmr_attr *fmr_attr)
1123 {
1124         struct mthca_fmr *fmr;
1125         int err;
1126
1127         fmr = kmalloc(sizeof *fmr, GFP_KERNEL);
1128         if (!fmr)
1129                 return ERR_PTR(-ENOMEM);
1130
1131         memcpy(&fmr->attr, fmr_attr, sizeof *fmr_attr);
1132         err = mthca_fmr_alloc(to_mdev(pd->device), to_mpd(pd)->pd_num,
1133                              convert_access(mr_access_flags), fmr);
1134
1135         if (err) {
1136                 kfree(fmr);
1137                 return ERR_PTR(err);
1138         }
1139
1140         return &fmr->ibmr;
1141 }
1142
1143 static int mthca_dealloc_fmr(struct ib_fmr *fmr)
1144 {
1145         struct mthca_fmr *mfmr = to_mfmr(fmr);
1146         int err;
1147
1148         err = mthca_free_fmr(to_mdev(fmr->device), mfmr);
1149         if (err)
1150                 return err;
1151
1152         kfree(mfmr);
1153         return 0;
1154 }
1155
1156 static int mthca_unmap_fmr(struct list_head *fmr_list)
1157 {
1158         struct ib_fmr *fmr;
1159         int err;
1160         u8 status;
1161         struct mthca_dev *mdev = NULL;
1162
1163         list_for_each_entry(fmr, fmr_list, list) {
1164                 if (mdev && to_mdev(fmr->device) != mdev)
1165                         return -EINVAL;
1166                 mdev = to_mdev(fmr->device);
1167         }
1168
1169         if (!mdev)
1170                 return 0;
1171
1172         if (mthca_is_memfree(mdev)) {
1173                 list_for_each_entry(fmr, fmr_list, list)
1174                         mthca_arbel_fmr_unmap(mdev, to_mfmr(fmr));
1175
1176                 wmb();
1177         } else
1178                 list_for_each_entry(fmr, fmr_list, list)
1179                         mthca_tavor_fmr_unmap(mdev, to_mfmr(fmr));
1180
1181         err = mthca_SYNC_TPT(mdev, &status);
1182         if (err)
1183                 return err;
1184         if (status)
1185                 return -EINVAL;
1186         return 0;
1187 }
1188
1189 static ssize_t show_rev(struct device *device, struct device_attribute *attr,
1190                         char *buf)
1191 {
1192         struct mthca_dev *dev =
1193                 container_of(device, struct mthca_dev, ib_dev.dev);
1194         return sprintf(buf, "%x\n", dev->rev_id);
1195 }
1196
1197 static ssize_t show_fw_ver(struct device *device, struct device_attribute *attr,
1198                            char *buf)
1199 {
1200         struct mthca_dev *dev =
1201                 container_of(device, struct mthca_dev, ib_dev.dev);
1202         return sprintf(buf, "%d.%d.%d\n", (int) (dev->fw_ver >> 32),
1203                        (int) (dev->fw_ver >> 16) & 0xffff,
1204                        (int) dev->fw_ver & 0xffff);
1205 }
1206
1207 static ssize_t show_hca(struct device *device, struct device_attribute *attr,
1208                         char *buf)
1209 {
1210         struct mthca_dev *dev =
1211                 container_of(device, struct mthca_dev, ib_dev.dev);
1212         switch (dev->pdev->device) {
1213         case PCI_DEVICE_ID_MELLANOX_TAVOR:
1214                 return sprintf(buf, "MT23108\n");
1215         case PCI_DEVICE_ID_MELLANOX_ARBEL_COMPAT:
1216                 return sprintf(buf, "MT25208 (MT23108 compat mode)\n");
1217         case PCI_DEVICE_ID_MELLANOX_ARBEL:
1218                 return sprintf(buf, "MT25208\n");
1219         case PCI_DEVICE_ID_MELLANOX_SINAI:
1220         case PCI_DEVICE_ID_MELLANOX_SINAI_OLD:
1221                 return sprintf(buf, "MT25204\n");
1222         default:
1223                 return sprintf(buf, "unknown\n");
1224         }
1225 }
1226
1227 static ssize_t show_board(struct device *device, struct device_attribute *attr,
1228                           char *buf)
1229 {
1230         struct mthca_dev *dev =
1231                 container_of(device, struct mthca_dev, ib_dev.dev);
1232         return sprintf(buf, "%.*s\n", MTHCA_BOARD_ID_LEN, dev->board_id);
1233 }
1234
1235 static DEVICE_ATTR(hw_rev,   S_IRUGO, show_rev,    NULL);
1236 static DEVICE_ATTR(fw_ver,   S_IRUGO, show_fw_ver, NULL);
1237 static DEVICE_ATTR(hca_type, S_IRUGO, show_hca,    NULL);
1238 static DEVICE_ATTR(board_id, S_IRUGO, show_board,  NULL);
1239
1240 static struct device_attribute *mthca_dev_attributes[] = {
1241         &dev_attr_hw_rev,
1242         &dev_attr_fw_ver,
1243         &dev_attr_hca_type,
1244         &dev_attr_board_id
1245 };
1246
1247 static int mthca_init_node_data(struct mthca_dev *dev)
1248 {
1249         struct ib_smp *in_mad  = NULL;
1250         struct ib_smp *out_mad = NULL;
1251         int err = -ENOMEM;
1252         u8 status;
1253
1254         in_mad  = kzalloc(sizeof *in_mad, GFP_KERNEL);
1255         out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
1256         if (!in_mad || !out_mad)
1257                 goto out;
1258
1259         init_query_mad(in_mad);
1260         in_mad->attr_id = IB_SMP_ATTR_NODE_DESC;
1261
1262         err = mthca_MAD_IFC(dev, 1, 1,
1263                             1, NULL, NULL, in_mad, out_mad,
1264                             &status);
1265         if (err)
1266                 goto out;
1267         if (status) {
1268                 err = -EINVAL;
1269                 goto out;
1270         }
1271
1272         memcpy(dev->ib_dev.node_desc, out_mad->data, 64);
1273
1274         in_mad->attr_id = IB_SMP_ATTR_NODE_INFO;
1275
1276         err = mthca_MAD_IFC(dev, 1, 1,
1277                             1, NULL, NULL, in_mad, out_mad,
1278                             &status);
1279         if (err)
1280                 goto out;
1281         if (status) {
1282                 err = -EINVAL;
1283                 goto out;
1284         }
1285
1286         if (mthca_is_memfree(dev))
1287                 dev->rev_id = be32_to_cpup((__be32 *) (out_mad->data + 32));
1288         memcpy(&dev->ib_dev.node_guid, out_mad->data + 12, 8);
1289
1290 out:
1291         kfree(in_mad);
1292         kfree(out_mad);
1293         return err;
1294 }
1295
1296 int mthca_register_device(struct mthca_dev *dev)
1297 {
1298         int ret;
1299         int i;
1300
1301         ret = mthca_init_node_data(dev);
1302         if (ret)
1303                 return ret;
1304
1305         strlcpy(dev->ib_dev.name, "mthca%d", IB_DEVICE_NAME_MAX);
1306         dev->ib_dev.owner                = THIS_MODULE;
1307
1308         dev->ib_dev.uverbs_abi_ver       = MTHCA_UVERBS_ABI_VERSION;
1309         dev->ib_dev.uverbs_cmd_mask      =
1310                 (1ull << IB_USER_VERBS_CMD_GET_CONTEXT)         |
1311                 (1ull << IB_USER_VERBS_CMD_QUERY_DEVICE)        |
1312                 (1ull << IB_USER_VERBS_CMD_QUERY_PORT)          |
1313                 (1ull << IB_USER_VERBS_CMD_ALLOC_PD)            |
1314                 (1ull << IB_USER_VERBS_CMD_DEALLOC_PD)          |
1315                 (1ull << IB_USER_VERBS_CMD_REG_MR)              |
1316                 (1ull << IB_USER_VERBS_CMD_DEREG_MR)            |
1317                 (1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL) |
1318                 (1ull << IB_USER_VERBS_CMD_CREATE_CQ)           |
1319                 (1ull << IB_USER_VERBS_CMD_RESIZE_CQ)           |
1320                 (1ull << IB_USER_VERBS_CMD_DESTROY_CQ)          |
1321                 (1ull << IB_USER_VERBS_CMD_CREATE_QP)           |
1322                 (1ull << IB_USER_VERBS_CMD_QUERY_QP)            |
1323                 (1ull << IB_USER_VERBS_CMD_MODIFY_QP)           |
1324                 (1ull << IB_USER_VERBS_CMD_DESTROY_QP)          |
1325                 (1ull << IB_USER_VERBS_CMD_ATTACH_MCAST)        |
1326                 (1ull << IB_USER_VERBS_CMD_DETACH_MCAST);
1327         dev->ib_dev.node_type            = RDMA_NODE_IB_CA;
1328         dev->ib_dev.phys_port_cnt        = dev->limits.num_ports;
1329         dev->ib_dev.num_comp_vectors     = 1;
1330         dev->ib_dev.dma_device           = &dev->pdev->dev;
1331         dev->ib_dev.query_device         = mthca_query_device;
1332         dev->ib_dev.query_port           = mthca_query_port;
1333         dev->ib_dev.modify_device        = mthca_modify_device;
1334         dev->ib_dev.modify_port          = mthca_modify_port;
1335         dev->ib_dev.query_pkey           = mthca_query_pkey;
1336         dev->ib_dev.query_gid            = mthca_query_gid;
1337         dev->ib_dev.alloc_ucontext       = mthca_alloc_ucontext;
1338         dev->ib_dev.dealloc_ucontext     = mthca_dealloc_ucontext;
1339         dev->ib_dev.mmap                 = mthca_mmap_uar;
1340         dev->ib_dev.alloc_pd             = mthca_alloc_pd;
1341         dev->ib_dev.dealloc_pd           = mthca_dealloc_pd;
1342         dev->ib_dev.create_ah            = mthca_ah_create;
1343         dev->ib_dev.query_ah             = mthca_ah_query;
1344         dev->ib_dev.destroy_ah           = mthca_ah_destroy;
1345
1346         if (dev->mthca_flags & MTHCA_FLAG_SRQ) {
1347                 dev->ib_dev.create_srq           = mthca_create_srq;
1348                 dev->ib_dev.modify_srq           = mthca_modify_srq;
1349                 dev->ib_dev.query_srq            = mthca_query_srq;
1350                 dev->ib_dev.destroy_srq          = mthca_destroy_srq;
1351                 dev->ib_dev.uverbs_cmd_mask     |=
1352                         (1ull << IB_USER_VERBS_CMD_CREATE_SRQ)          |
1353                         (1ull << IB_USER_VERBS_CMD_MODIFY_SRQ)          |
1354                         (1ull << IB_USER_VERBS_CMD_QUERY_SRQ)           |
1355                         (1ull << IB_USER_VERBS_CMD_DESTROY_SRQ);
1356
1357                 if (mthca_is_memfree(dev))
1358                         dev->ib_dev.post_srq_recv = mthca_arbel_post_srq_recv;
1359                 else
1360                         dev->ib_dev.post_srq_recv = mthca_tavor_post_srq_recv;
1361         }
1362
1363         dev->ib_dev.create_qp            = mthca_create_qp;
1364         dev->ib_dev.modify_qp            = mthca_modify_qp;
1365         dev->ib_dev.query_qp             = mthca_query_qp;
1366         dev->ib_dev.destroy_qp           = mthca_destroy_qp;
1367         dev->ib_dev.create_cq            = mthca_create_cq;
1368         dev->ib_dev.resize_cq            = mthca_resize_cq;
1369         dev->ib_dev.destroy_cq           = mthca_destroy_cq;
1370         dev->ib_dev.poll_cq              = mthca_poll_cq;
1371         dev->ib_dev.get_dma_mr           = mthca_get_dma_mr;
1372         dev->ib_dev.reg_phys_mr          = mthca_reg_phys_mr;
1373         dev->ib_dev.reg_user_mr          = mthca_reg_user_mr;
1374         dev->ib_dev.dereg_mr             = mthca_dereg_mr;
1375
1376         if (dev->mthca_flags & MTHCA_FLAG_FMR) {
1377                 dev->ib_dev.alloc_fmr            = mthca_alloc_fmr;
1378                 dev->ib_dev.unmap_fmr            = mthca_unmap_fmr;
1379                 dev->ib_dev.dealloc_fmr          = mthca_dealloc_fmr;
1380                 if (mthca_is_memfree(dev))
1381                         dev->ib_dev.map_phys_fmr = mthca_arbel_map_phys_fmr;
1382                 else
1383                         dev->ib_dev.map_phys_fmr = mthca_tavor_map_phys_fmr;
1384         }
1385
1386         dev->ib_dev.attach_mcast         = mthca_multicast_attach;
1387         dev->ib_dev.detach_mcast         = mthca_multicast_detach;
1388         dev->ib_dev.process_mad          = mthca_process_mad;
1389
1390         if (mthca_is_memfree(dev)) {
1391                 dev->ib_dev.req_notify_cq = mthca_arbel_arm_cq;
1392                 dev->ib_dev.post_send     = mthca_arbel_post_send;
1393                 dev->ib_dev.post_recv     = mthca_arbel_post_receive;
1394         } else {
1395                 dev->ib_dev.req_notify_cq = mthca_tavor_arm_cq;
1396                 dev->ib_dev.post_send     = mthca_tavor_post_send;
1397                 dev->ib_dev.post_recv     = mthca_tavor_post_receive;
1398         }
1399
1400         mutex_init(&dev->cap_mask_mutex);
1401
1402         ret = ib_register_device(&dev->ib_dev);
1403         if (ret)
1404                 return ret;
1405
1406         for (i = 0; i < ARRAY_SIZE(mthca_dev_attributes); ++i) {
1407                 ret = device_create_file(&dev->ib_dev.dev,
1408                                          mthca_dev_attributes[i]);
1409                 if (ret) {
1410                         ib_unregister_device(&dev->ib_dev);
1411                         return ret;
1412                 }
1413         }
1414
1415         mthca_start_catas_poll(dev);
1416
1417         return 0;
1418 }
1419
1420 void mthca_unregister_device(struct mthca_dev *dev)
1421 {
1422         mthca_stop_catas_poll(dev);
1423         ib_unregister_device(&dev->ib_dev);
1424 }