SAFE public projects git trees. - safe/jmp/linux-2.6/blob - 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/slab.h>
  43 #include <linux/mm.h>
  44
  45 #include "mthca_dev.h"
  46 #include "mthca_cmd.h"
  47 #include "mthca_user.h"
  48 #include "mthca_memfree.h"
  49
  50 static void init_query_mad(struct ib_smp *mad)
  51 {
  52         mad->base_version  = 1;
  53         mad->mgmt_class    = IB_MGMT_CLASS_SUBN_LID_ROUTED;
  54         mad->class_version = 1;
  55         mad->method        = IB_MGMT_METHOD_GET;
  56 }
  57
  58 static int mthca_query_device(struct ib_device *ibdev,
  59                               struct ib_device_attr *props)
  60 {
  61         struct ib_smp *in_mad  = NULL;
  62         struct ib_smp *out_mad = NULL;
  63         int err = -ENOMEM;
  64         struct mthca_dev *mdev = to_mdev(ibdev);
  65
  66         u8 status;
  67
  68         in_mad  = kzalloc(sizeof *in_mad, GFP_KERNEL);
  69         out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
  70         if (!in_mad || !out_mad)
  71                 goto out;
  72
  73         memset(props, 0, sizeof *props);
  74
  75         props->fw_ver              = mdev->fw_ver;
  76
  77         init_query_mad(in_mad);
  78         in_mad->attr_id = IB_SMP_ATTR_NODE_INFO;
  79
  80         err = mthca_MAD_IFC(mdev, 1, 1,
  81                             1, NULL, NULL, in_mad, out_mad,
  82                             &status);
  83         if (err)
  84                 goto out;
  85         if (status) {
  86                 err = -EINVAL;
  87                 goto out;
  88         }
  89
  90         props->device_cap_flags    = mdev->device_cap_flags;
  91         props->vendor_id           = be32_to_cpup((__be32 *) (out_mad->data + 36)) &
  92                 0xffffff;
  93         props->vendor_part_id      = be16_to_cpup((__be16 *) (out_mad->data + 30));
  94         props->hw_ver              = be32_to_cpup((__be32 *) (out_mad->data + 32));
  95         memcpy(&props->sys_image_guid, out_mad->data +  4, 8);
  96
  97         props->max_mr_size         = ~0ull;
  98         props->page_size_cap       = mdev->limits.page_size_cap;
  99         props->max_qp              = mdev->limits.num_qps - mdev->limits.reserved_qps;
 100         props->max_qp_wr           = mdev->limits.max_wqes;
 101         props->max_sge             = mdev->limits.max_sg;
 102         props->max_cq              = mdev->limits.num_cqs - mdev->limits.reserved_cqs;
 103         props->max_cqe             = mdev->limits.max_cqes;
 104         props->max_mr              = mdev->limits.num_mpts - mdev->limits.reserved_mrws;
 105         props->max_pd              = mdev->limits.num_pds - mdev->limits.reserved_pds;
 106         props->max_qp_rd_atom      = 1 << mdev->qp_table.rdb_shift;
 107         props->max_qp_init_rd_atom = mdev->limits.max_qp_init_rdma;
 108         props->max_res_rd_atom     = props->max_qp_rd_atom * props->max_qp;
 109         props->max_srq             = mdev->limits.num_srqs - mdev->limits.reserved_srqs;
 110         props->max_srq_wr          = mdev->limits.max_srq_wqes;
 111         props->max_srq_sge         = mdev->limits.max_srq_sge;
 112         props->local_ca_ack_delay  = mdev->limits.local_ca_ack_delay;
 113         props->atomic_cap          = mdev->limits.flags & DEV_LIM_FLAG_ATOMIC ?
 114                                         IB_ATOMIC_HCA : IB_ATOMIC_NONE;
 115         props->max_pkeys           = mdev->limits.pkey_table_len;
 116         props->max_mcast_grp       = mdev->limits.num_mgms + mdev->limits.num_amgms;
 117         props->max_mcast_qp_attach = MTHCA_QP_PER_MGM;
 118         props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
 119                                            props->max_mcast_grp;
 120         /*
 121          * If Sinai memory key optimization is being used, then only
 122          * the 8-bit key portion will change.  For other HCAs, the
 123          * unused index bits will also be used for FMR remapping.
 124          */
 125         if (mdev->mthca_flags & MTHCA_FLAG_SINAI_OPT)
 126                 props->max_map_per_fmr = 255;
 127         else
 128                 props->max_map_per_fmr =
 129                         (1 << (32 - ilog2(mdev->limits.num_mpts))) - 1;
 130
 131         err = 0;
 132  out:
 133         kfree(in_mad);
 134         kfree(out_mad);
 135         return err;
 136 }
 137
 138 static int mthca_query_port(struct ib_device *ibdev,
 139                             u8 port, struct ib_port_attr *props)
 140 {
 141         struct ib_smp *in_mad  = NULL;
 142         struct ib_smp *out_mad = NULL;
 143         int err = -ENOMEM;
 144         u8 status;
 145
 146         in_mad  = kzalloc(sizeof *in_mad, GFP_KERNEL);
 147         out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
 148         if (!in_mad || !out_mad)
 149                 goto out;
 150
 151         memset(props, 0, sizeof *props);
 152
 153         init_query_mad(in_mad);
 154         in_mad->attr_id  = IB_SMP_ATTR_PORT_INFO;
 155         in_mad->attr_mod = cpu_to_be32(port);
 156
 157         err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
 158                             port, NULL, NULL, in_mad, out_mad,
 159                             &status);
 160         if (err)
 161                 goto out;
 162         if (status) {
 163                 err = -EINVAL;
 164                 goto out;
 165         }
 166
 167         props->lid               = be16_to_cpup((__be16 *) (out_mad->data + 16));
 168         props->lmc               = out_mad->data[34] & 0x7;
 169         props->sm_lid            = be16_to_cpup((__be16 *) (out_mad->data + 18));
 170         props->sm_sl             = out_mad->data[36] & 0xf;
 171         props->state             = out_mad->data[32] & 0xf;
 172         props->phys_state        = out_mad->data[33] >> 4;
 173         props->port_cap_flags    = be32_to_cpup((__be32 *) (out_mad->data + 20));
 174         props->gid_tbl_len       = to_mdev(ibdev)->limits.gid_table_len;
 175         props->max_msg_sz        = 0x80000000;
 176         props->pkey_tbl_len      = to_mdev(ibdev)->limits.pkey_table_len;
 177         props->bad_pkey_cntr     = be16_to_cpup((__be16 *) (out_mad->data + 46));
 178         props->qkey_viol_cntr    = be16_to_cpup((__be16 *) (out_mad->data + 48));
 179         props->active_width      = out_mad->data[31] & 0xf;
 180         props->active_speed      = out_mad->data[35] >> 4;
 181         props->max_mtu           = out_mad->data[41] & 0xf;
 182         props->active_mtu        = out_mad->data[36] >> 4;
 183         props->subnet_timeout    = out_mad->data[51] & 0x1f;
 184         props->max_vl_num        = out_mad->data[37] >> 4;
 185         props->init_type_reply   = out_mad->data[41] >> 4;
 186
 187  out:
 188         kfree(in_mad);
 189         kfree(out_mad);
 190         return err;
 191 }
 192
 193 static int mthca_modify_device(struct ib_device *ibdev,
 194                                int mask,
 195                                struct ib_device_modify *props)
 196 {
 197         if (mask & ~IB_DEVICE_MODIFY_NODE_DESC)
 198                 return -EOPNOTSUPP;
 199
 200         if (mask & IB_DEVICE_MODIFY_NODE_DESC) {
 201                 if (mutex_lock_interruptible(&to_mdev(ibdev)->cap_mask_mutex))
 202                         return -ERESTARTSYS;
 203                 memcpy(ibdev->node_desc, props->node_desc, 64);
 204                 mutex_unlock(&to_mdev(ibdev)->cap_mask_mutex);
 205         }
 206
 207         return 0;
 208 }
 209
 210 static int mthca_modify_port(struct ib_device *ibdev,
 211                              u8 port, int port_modify_mask,
 212                              struct ib_port_modify *props)
 213 {
 214         struct mthca_set_ib_param set_ib;
 215         struct ib_port_attr attr;
 216         int err;
 217         u8 status;
 218
 219         if (mutex_lock_interruptible(&to_mdev(ibdev)->cap_mask_mutex))
 220                 return -ERESTARTSYS;
 221
 222         err = mthca_query_port(ibdev, port, &attr);
 223         if (err)
 224                 goto out;
 225
 226         set_ib.set_si_guid     = 0;
 227         set_ib.reset_qkey_viol = !!(port_modify_mask & IB_PORT_RESET_QKEY_CNTR);
 228
 229         set_ib.cap_mask = (attr.port_cap_flags | props->set_port_cap_mask) &
 230                 ~props->clr_port_cap_mask;
 231
 232         err = mthca_SET_IB(to_mdev(ibdev), &set_ib, port, &status);
 233         if (err)
 234                 goto out;
 235         if (status) {
 236                 err = -EINVAL;
 237                 goto out;
 238         }
 239
 240 out:
 241         mutex_unlock(&to_mdev(ibdev)->cap_mask_mutex);
 242         return err;
 243 }
 244
 245 static int mthca_query_pkey(struct ib_device *ibdev,
 246                             u8 port, u16 index, u16 *pkey)
 247 {
 248         struct ib_smp *in_mad  = NULL;
 249         struct ib_smp *out_mad = NULL;
 250         int err = -ENOMEM;
 251         u8 status;
 252
 253         in_mad  = kzalloc(sizeof *in_mad, GFP_KERNEL);
 254         out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
 255         if (!in_mad || !out_mad)
 256                 goto out;
 257
 258         init_query_mad(in_mad);
 259         in_mad->attr_id  = IB_SMP_ATTR_PKEY_TABLE;
 260         in_mad->attr_mod = cpu_to_be32(index / 32);
 261
 262         err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
 263                             port, NULL, NULL, in_mad, out_mad,
 264                             &status);
 265         if (err)
 266                 goto out;
 267         if (status) {
 268                 err = -EINVAL;
 269                 goto out;
 270         }
 271
 272         *pkey = be16_to_cpu(((__be16 *) out_mad->data)[index % 32]);
 273
 274  out:
 275         kfree(in_mad);
 276         kfree(out_mad);
 277         return err;
 278 }
 279
 280 static int mthca_query_gid(struct ib_device *ibdev, u8 port,
 281                            int index, union ib_gid *gid)
 282 {
 283         struct ib_smp *in_mad  = NULL;
 284         struct ib_smp *out_mad = NULL;
 285         int err = -ENOMEM;
 286         u8 status;
 287
 288         in_mad  = kzalloc(sizeof *in_mad, GFP_KERNEL);
 289         out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
 290         if (!in_mad || !out_mad)
 291                 goto out;
 292
 293         init_query_mad(in_mad);
 294         in_mad->attr_id  = IB_SMP_ATTR_PORT_INFO;
 295         in_mad->attr_mod = cpu_to_be32(port);
 296
 297         err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
 298                             port, NULL, NULL, in_mad, out_mad,
 299                             &status);
 300         if (err)
 301                 goto out;
 302         if (status) {
 303                 err = -EINVAL;
 304                 goto out;
 305         }
 306
 307         memcpy(gid->raw, out_mad->data + 8, 8);
 308
 309         init_query_mad(in_mad);
 310         in_mad->attr_id  = IB_SMP_ATTR_GUID_INFO;
 311         in_mad->attr_mod = cpu_to_be32(index / 8);
 312
 313         err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
 314                             port, NULL, NULL, in_mad, out_mad,
 315                             &status);
 316         if (err)
 317                 goto out;
 318         if (status) {
 319                 err = -EINVAL;
 320                 goto out;
 321         }
 322
 323         memcpy(gid->raw + 8, out_mad->data + (index % 8) * 8, 8);
 324
 325  out:
 326         kfree(in_mad);
 327         kfree(out_mad);
 328         return err;
 329 }
 330
 331 static struct ib_ucontext *mthca_alloc_ucontext(struct ib_device *ibdev,
 332                                                 struct ib_udata *udata)
 333 {
 334         struct mthca_alloc_ucontext_resp uresp;
 335         struct mthca_ucontext           *context;
 336         int                              err;
 337
 338         if (!(to_mdev(ibdev)->active))
 339                 return ERR_PTR(-EAGAIN);
 340
 341         memset(&uresp, 0, sizeof uresp);
 342
 343         uresp.qp_tab_size = to_mdev(ibdev)->limits.num_qps;
 344         if (mthca_is_memfree(to_mdev(ibdev)))
 345                 uresp.uarc_size = to_mdev(ibdev)->uar_table.uarc_size;
 346         else
 347                 uresp.uarc_size = 0;
 348
 349         context = kmalloc(sizeof *context, GFP_KERNEL);
 350         if (!context)
 351                 return ERR_PTR(-ENOMEM);
 352
 353         err = mthca_uar_alloc(to_mdev(ibdev), &context->uar);
 354         if (err) {
 355                 kfree(context);
 356                 return ERR_PTR(err);
 357         }
 358
 359         context->db_tab = mthca_init_user_db_tab(to_mdev(ibdev));
 360         if (IS_ERR(context->db_tab)) {
 361                 err = PTR_ERR(context->db_tab);
 362                 mthca_uar_free(to_mdev(ibdev), &context->uar);
 363                 kfree(context);
 364                 return ERR_PTR(err);
 365         }
 366
 367         if (ib_copy_to_udata(udata, &uresp, sizeof uresp)) {
 368                 mthca_cleanup_user_db_tab(to_mdev(ibdev), &context->uar, context->db_tab);
 369                 mthca_uar_free(to_mdev(ibdev), &context->uar);
 370                 kfree(context);
 371                 return ERR_PTR(-EFAULT);
 372         }
 373
 374         context->reg_mr_warned = 0;
 375
 376         return &context->ibucontext;
 377 }
 378
 379 static int mthca_dealloc_ucontext(struct ib_ucontext *context)
 380 {
 381         mthca_cleanup_user_db_tab(to_mdev(context->device), &to_mucontext(context)->uar,
 382                                   to_mucontext(context)->db_tab);
 383         mthca_uar_free(to_mdev(context->device), &to_mucontext(context)->uar);
 384         kfree(to_mucontext(context));
 385
 386         return 0;
 387 }
 388
 389 static int mthca_mmap_uar(struct ib_ucontext *context,
 390                           struct vm_area_struct *vma)
 391 {
 392         if (vma->vm_end - vma->vm_start != PAGE_SIZE)
 393                 return -EINVAL;
 394
 395         vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
 396
 397         if (io_remap_pfn_range(vma, vma->vm_start,
 398                                to_mucontext(context)->uar.pfn,
 399                                PAGE_SIZE, vma->vm_page_prot))
 400                 return -EAGAIN;
 401
 402         return 0;
 403 }
 404
 405 static struct ib_pd *mthca_alloc_pd(struct ib_device *ibdev,
 406                                     struct ib_ucontext *context,
 407                                     struct ib_udata *udata)
 408 {
 409         struct mthca_pd *pd;
 410         int err;
 411
 412         pd = kmalloc(sizeof *pd, GFP_KERNEL);
 413         if (!pd)
 414                 return ERR_PTR(-ENOMEM);
 415
 416         err = mthca_pd_alloc(to_mdev(ibdev), !context, pd);
 417         if (err) {
 418                 kfree(pd);
 419                 return ERR_PTR(err);
 420         }
 421
 422         if (context) {
 423                 if (ib_copy_to_udata(udata, &pd->pd_num, sizeof (__u32))) {
 424                         mthca_pd_free(to_mdev(ibdev), pd);
 425                         kfree(pd);
 426                         return ERR_PTR(-EFAULT);
 427                 }
 428         }
 429
 430         return &pd->ibpd;
 431 }
 432
 433 static int mthca_dealloc_pd(struct ib_pd *pd)
 434 {
 435         mthca_pd_free(to_mdev(pd->device), to_mpd(pd));
 436         kfree(pd);
 437
 438         return 0;
 439 }
 440
 441 static struct ib_ah *mthca_ah_create(struct ib_pd *pd,
 442                                      struct ib_ah_attr *ah_attr)
 443 {
 444         int err;
 445         struct mthca_ah *ah;
 446
 447         ah = kmalloc(sizeof *ah, GFP_ATOMIC);
 448         if (!ah)
 449                 return ERR_PTR(-ENOMEM);
 450
 451         err = mthca_create_ah(to_mdev(pd->device), to_mpd(pd), ah_attr, ah);
 452         if (err) {
 453                 kfree(ah);
 454                 return ERR_PTR(err);
 455         }
 456
 457         return &ah->ibah;
 458 }
 459
 460 static int mthca_ah_destroy(struct ib_ah *ah)
 461 {
 462         mthca_destroy_ah(to_mdev(ah->device), to_mah(ah));
 463         kfree(ah);
 464
 465         return 0;
 466 }
 467
 468 static struct ib_srq *mthca_create_srq(struct ib_pd *pd,
 469                                        struct ib_srq_init_attr *init_attr,
 470                                        struct ib_udata *udata)
 471 {
 472         struct mthca_create_srq ucmd;
 473         struct mthca_ucontext *context = NULL;
 474         struct mthca_srq *srq;
 475         int err;
 476
 477         srq = kmalloc(sizeof *srq, GFP_KERNEL);
 478         if (!srq)
 479                 return ERR_PTR(-ENOMEM);
 480
 481         if (pd->uobject) {
 482                 context = to_mucontext(pd->uobject->context);
 483
 484                 if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
 485                         err = -EFAULT;
 486                         goto err_free;
 487                 }
 488
 489                 err = mthca_map_user_db(to_mdev(pd->device), &context->uar,
 490                                         context->db_tab, ucmd.db_index,
 491                                         ucmd.db_page);
 492
 493                 if (err)
 494                         goto err_free;
 495
 496                 srq->mr.ibmr.lkey = ucmd.lkey;
 497                 srq->db_index     = ucmd.db_index;
 498         }
 499
 500         err = mthca_alloc_srq(to_mdev(pd->device), to_mpd(pd),
 501                               &init_attr->attr, srq);
 502
 503         if (err && pd->uobject)
 504                 mthca_unmap_user_db(to_mdev(pd->device), &context->uar,
 505                                     context->db_tab, ucmd.db_index);
 506
 507         if (err)
 508                 goto err_free;
 509
 510         if (context && ib_copy_to_udata(udata, &srq->srqn, sizeof (__u32))) {
 511                 mthca_free_srq(to_mdev(pd->device), srq);
 512                 err = -EFAULT;
 513                 goto err_free;
 514         }
 515
 516         return &srq->ibsrq;
 517
 518 err_free:
 519         kfree(srq);
 520
 521         return ERR_PTR(err);
 522 }
 523
 524 static int mthca_destroy_srq(struct ib_srq *srq)
 525 {
 526         struct mthca_ucontext *context;
 527
 528         if (srq->uobject) {
 529                 context = to_mucontext(srq->uobject->context);
 530
 531                 mthca_unmap_user_db(to_mdev(srq->device), &context->uar,
 532                                     context->db_tab, to_msrq(srq)->db_index);
 533         }
 534
 535         mthca_free_srq(to_mdev(srq->device), to_msrq(srq));
 536         kfree(srq);
 537
 538         return 0;
 539 }
 540
 541 static struct ib_qp *mthca_create_qp(struct ib_pd *pd,
 542                                      struct ib_qp_init_attr *init_attr,
 543                                      struct ib_udata *udata)
 544 {
 545         struct mthca_create_qp ucmd;
 546         struct mthca_qp *qp;
 547         int err;
 548
 549         if (init_attr->create_flags)
 550                 return ERR_PTR(-EINVAL);
 551
 552         switch (init_attr->qp_type) {
 553         case IB_QPT_RC:
 554         case IB_QPT_UC:
 555         case IB_QPT_UD:
 556         {
 557                 struct mthca_ucontext *context;
 558
 559                 qp = kmalloc(sizeof *qp, GFP_KERNEL);
 560                 if (!qp)
 561                         return ERR_PTR(-ENOMEM);
 562
 563                 if (pd->uobject) {
 564                         context = to_mucontext(pd->uobject->context);
 565
 566                         if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
 567                                 kfree(qp);
 568                                 return ERR_PTR(-EFAULT);
 569                         }
 570
 571                         err = mthca_map_user_db(to_mdev(pd->device), &context->uar,
 572                                                 context->db_tab,
 573                                                 ucmd.sq_db_index, ucmd.sq_db_page);
 574                         if (err) {
 575                                 kfree(qp);
 576                                 return ERR_PTR(err);
 577                         }
 578
 579                         err = mthca_map_user_db(to_mdev(pd->device), &context->uar,
 580                                                 context->db_tab,
 581                                                 ucmd.rq_db_index, ucmd.rq_db_page);
 582                         if (err) {
 583                                 mthca_unmap_user_db(to_mdev(pd->device),
 584                                                     &context->uar,
 585                                                     context->db_tab,
 586                                                     ucmd.sq_db_index);
 587                                 kfree(qp);
 588                                 return ERR_PTR(err);
 589                         }
 590
 591                         qp->mr.ibmr.lkey = ucmd.lkey;
 592                         qp->sq.db_index  = ucmd.sq_db_index;
 593                         qp->rq.db_index  = ucmd.rq_db_index;
 594                 }
 595
 596                 err = mthca_alloc_qp(to_mdev(pd->device), to_mpd(pd),
 597                                      to_mcq(init_attr->send_cq),
 598                                      to_mcq(init_attr->recv_cq),
 599                                      init_attr->qp_type, init_attr->sq_sig_type,
 600                                      &init_attr->cap, qp);
 601
 602                 if (err && pd->uobject) {
 603                         context = to_mucontext(pd->uobject->context);
 604
 605                         mthca_unmap_user_db(to_mdev(pd->device),
 606                                             &context->uar,
 607                                             context->db_tab,
 608                                             ucmd.sq_db_index);
 609                         mthca_unmap_user_db(to_mdev(pd->device),
 610                                             &context->uar,
 611                                             context->db_tab,
 612                                             ucmd.rq_db_index);
 613                 }
 614
 615                 qp->ibqp.qp_num = qp->qpn;
 616                 break;
 617         }
 618         case IB_QPT_SMI:
 619         case IB_QPT_GSI:
 620         {
 621                 /* Don't allow userspace to create special QPs */
 622                 if (pd->uobject)
 623                         return ERR_PTR(-EINVAL);
 624
 625                 qp = kmalloc(sizeof (struct mthca_sqp), GFP_KERNEL);
 626                 if (!qp)
 627                         return ERR_PTR(-ENOMEM);
 628
 629                 qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 : 1;
 630
 631                 err = mthca_alloc_sqp(to_mdev(pd->device), to_mpd(pd),
 632                                       to_mcq(init_attr->send_cq),
 633                                       to_mcq(init_attr->recv_cq),
 634                                       init_attr->sq_sig_type, &init_attr->cap,
 635                                       qp->ibqp.qp_num, init_attr->port_num,
 636                                       to_msqp(qp));
 637                 break;
 638         }
 639         default:
 640                 /* Don't support raw QPs */
 641                 return ERR_PTR(-ENOSYS);
 642         }
 643
 644         if (err) {
 645                 kfree(qp);
 646                 return ERR_PTR(err);
 647         }
 648
 649         init_attr->cap.max_send_wr     = qp->sq.max;
 650         init_attr->cap.max_recv_wr     = qp->rq.max;
 651         init_attr->cap.max_send_sge    = qp->sq.max_gs;
 652         init_attr->cap.max_recv_sge    = qp->rq.max_gs;
 653         init_attr->cap.max_inline_data = qp->max_inline_data;
 654
 655         return &qp->ibqp;
 656 }
 657
 658 static int mthca_destroy_qp(struct ib_qp *qp)
 659 {
 660         if (qp->uobject) {
 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)->sq.db_index);
 665                 mthca_unmap_user_db(to_mdev(qp->device),
 666                                     &to_mucontext(qp->uobject->context)->uar,
 667                                     to_mucontext(qp->uobject->context)->db_tab,
 668                                     to_mqp(qp)->rq.db_index);
 669         }
 670         mthca_free_qp(to_mdev(qp->device), to_mqp(qp));
 671         kfree(qp);
 672         return 0;
 673 }
 674
 675 static struct ib_cq *mthca_create_cq(struct ib_device *ibdev, int entries,
 676                                      int comp_vector,
 677                                      struct ib_ucontext *context,
 678                                      struct ib_udata *udata)
 679 {
 680         struct mthca_create_cq ucmd;
 681         struct mthca_cq *cq;
 682         int nent;
 683         int err;
 684
 685         if (entries < 1 || entries > to_mdev(ibdev)->limits.max_cqes)
 686                 return ERR_PTR(-EINVAL);
 687
 688         if (context) {
 689                 if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd))
 690                         return ERR_PTR(-EFAULT);
 691
 692                 err = mthca_map_user_db(to_mdev(ibdev), &to_mucontext(context)->uar,
 693                                         to_mucontext(context)->db_tab,
 694                                         ucmd.set_db_index, ucmd.set_db_page);
 695                 if (err)
 696                         return ERR_PTR(err);
 697
 698                 err = mthca_map_user_db(to_mdev(ibdev), &to_mucontext(context)->uar,
 699                                         to_mucontext(context)->db_tab,
 700                                         ucmd.arm_db_index, ucmd.arm_db_page);
 701                 if (err)
 702                         goto err_unmap_set;
 703         }
 704
 705         cq = kmalloc(sizeof *cq, GFP_KERNEL);
 706         if (!cq) {
 707                 err = -ENOMEM;
 708                 goto err_unmap_arm;
 709         }
 710
 711         if (context) {
 712                 cq->buf.mr.ibmr.lkey = ucmd.lkey;
 713                 cq->set_ci_db_index  = ucmd.set_db_index;
 714                 cq->arm_db_index     = ucmd.arm_db_index;
 715         }
 716
 717         for (nent = 1; nent <= entries; nent <<= 1)
 718                 ; /* nothing */
 719
 720         err = mthca_init_cq(to_mdev(ibdev), nent,
 721                             context ? to_mucontext(context) : NULL,
 722                             context ? ucmd.pdn : to_mdev(ibdev)->driver_pd.pd_num,
 723                             cq);
 724         if (err)
 725                 goto err_free;
 726
 727         if (context && ib_copy_to_udata(udata, &cq->cqn, sizeof (__u32))) {
 728                 mthca_free_cq(to_mdev(ibdev), cq);
 729                 goto err_free;
 730         }
 731
 732         cq->resize_buf = NULL;
 733
 734         return &cq->ibcq;
 735
 736 err_free:
 737         kfree(cq);
 738
 739 err_unmap_arm:
 740         if (context)
 741                 mthca_unmap_user_db(to_mdev(ibdev), &to_mucontext(context)->uar,
 742                                     to_mucontext(context)->db_tab, ucmd.arm_db_index);
 743
 744 err_unmap_set:
 745         if (context)
 746                 mthca_unmap_user_db(to_mdev(ibdev), &to_mucontext(context)->uar,
 747                                     to_mucontext(context)->db_tab, ucmd.set_db_index);
 748
 749         return ERR_PTR(err);
 750 }
 751
 752 static int mthca_alloc_resize_buf(struct mthca_dev *dev, struct mthca_cq *cq,
 753                                   int entries)
 754 {
 755         int ret;
 756
 757         spin_lock_irq(&cq->lock);
 758         if (cq->resize_buf) {
 759                 ret = -EBUSY;
 760                 goto unlock;
 761         }
 762
 763         cq->resize_buf = kmalloc(sizeof *cq->resize_buf, GFP_ATOMIC);
 764         if (!cq->resize_buf) {
 765                 ret = -ENOMEM;
 766                 goto unlock;
 767         }
 768
 769         cq->resize_buf->state = CQ_RESIZE_ALLOC;
 770
 771         ret = 0;
 772
 773 unlock:
 774         spin_unlock_irq(&cq->lock);
 775
 776         if (ret)
 777                 return ret;
 778
 779         ret = mthca_alloc_cq_buf(dev, &cq->resize_buf->buf, entries);
 780         if (ret) {
 781                 spin_lock_irq(&cq->lock);
 782                 kfree(cq->resize_buf);
 783                 cq->resize_buf = NULL;
 784                 spin_unlock_irq(&cq->lock);
 785                 return ret;
 786         }
 787
 788         cq->resize_buf->cqe = entries - 1;
 789
 790         spin_lock_irq(&cq->lock);
 791         cq->resize_buf->state = CQ_RESIZE_READY;
 792         spin_unlock_irq(&cq->lock);
 793
 794         return 0;
 795 }
 796
 797 static int mthca_resize_cq(struct ib_cq *ibcq, int entries, struct ib_udata *udata)
 798 {
 799         struct mthca_dev *dev = to_mdev(ibcq->device);
 800         struct mthca_cq *cq = to_mcq(ibcq);
 801         struct mthca_resize_cq ucmd;
 802         u32 lkey;
 803         u8 status;
 804         int ret;
 805
 806         if (entries < 1 || entries > dev->limits.max_cqes)
 807                 return -EINVAL;
 808
 809         mutex_lock(&cq->mutex);
 810
 811         entries = roundup_pow_of_two(entries + 1);
 812         if (entries == ibcq->cqe + 1) {
 813                 ret = 0;
 814                 goto out;
 815         }
 816
 817         if (cq->is_kernel) {
 818                 ret = mthca_alloc_resize_buf(dev, cq, entries);
 819                 if (ret)
 820                         goto out;
 821                 lkey = cq->resize_buf->buf.mr.ibmr.lkey;
 822         } else {
 823                 if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
 824                         ret = -EFAULT;
 825                         goto out;
 826                 }
 827                 lkey = ucmd.lkey;
 828         }
 829
 830         ret = mthca_RESIZE_CQ(dev, cq->cqn, lkey, ilog2(entries), &status);
 831         if (status)
 832                 ret = -EINVAL;
 833
 834         if (ret) {
 835                 if (cq->resize_buf) {
 836                         mthca_free_cq_buf(dev, &cq->resize_buf->buf,
 837                                           cq->resize_buf->cqe);
 838                         kfree(cq->resize_buf);
 839                         spin_lock_irq(&cq->lock);
 840                         cq->resize_buf = NULL;
 841                         spin_unlock_irq(&cq->lock);
 842                 }
 843                 goto out;
 844         }
 845
 846         if (cq->is_kernel) {
 847                 struct mthca_cq_buf tbuf;
 848                 int tcqe;
 849
 850                 spin_lock_irq(&cq->lock);
 851                 if (cq->resize_buf->state == CQ_RESIZE_READY) {
 852                         mthca_cq_resize_copy_cqes(cq);
 853                         tbuf         = cq->buf;
 854                         tcqe         = cq->ibcq.cqe;
 855                         cq->buf      = cq->resize_buf->buf;
 856                         cq->ibcq.cqe = cq->resize_buf->cqe;
 857                 } else {
 858                         tbuf = cq->resize_buf->buf;
 859                         tcqe = cq->resize_buf->cqe;
 860                 }
 861
 862                 kfree(cq->resize_buf);
 863                 cq->resize_buf = NULL;
 864                 spin_unlock_irq(&cq->lock);
 865
 866                 mthca_free_cq_buf(dev, &tbuf, tcqe);
 867         } else
 868                 ibcq->cqe = entries - 1;
 869
 870 out:
 871         mutex_unlock(&cq->mutex);
 872
 873         return ret;
 874 }
 875
 876 static int mthca_destroy_cq(struct ib_cq *cq)
 877 {
 878         if (cq->uobject) {
 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)->arm_db_index);
 883                 mthca_unmap_user_db(to_mdev(cq->device),
 884                                     &to_mucontext(cq->uobject->context)->uar,
 885                                     to_mucontext(cq->uobject->context)->db_tab,
 886                                     to_mcq(cq)->set_ci_db_index);
 887         }
 888         mthca_free_cq(to_mdev(cq->device), to_mcq(cq));
 889         kfree(cq);
 890
 891         return 0;
 892 }
 893
 894 static inline u32 convert_access(int acc)
 895 {
 896         return (acc & IB_ACCESS_REMOTE_ATOMIC ? MTHCA_MPT_FLAG_ATOMIC       : 0) |
 897                (acc & IB_ACCESS_REMOTE_WRITE  ? MTHCA_MPT_FLAG_REMOTE_WRITE : 0) |
 898                (acc & IB_ACCESS_REMOTE_READ   ? MTHCA_MPT_FLAG_REMOTE_READ  : 0) |
 899                (acc & IB_ACCESS_LOCAL_WRITE   ? MTHCA_MPT_FLAG_LOCAL_WRITE  : 0) |
 900                MTHCA_MPT_FLAG_LOCAL_READ;
 901 }
 902
 903 static struct ib_mr *mthca_get_dma_mr(struct ib_pd *pd, int acc)
 904 {
 905         struct mthca_mr *mr;
 906         int err;
 907
 908         mr = kmalloc(sizeof *mr, GFP_KERNEL);
 909         if (!mr)
 910                 return ERR_PTR(-ENOMEM);
 911
 912         err = mthca_mr_alloc_notrans(to_mdev(pd->device),
 913                                      to_mpd(pd)->pd_num,
 914                                      convert_access(acc), mr);
 915
 916         if (err) {
 917                 kfree(mr);
 918                 return ERR_PTR(err);
 919         }
 920
 921         mr->umem = NULL;
 922
 923         return &mr->ibmr;
 924 }
 925
 926 static struct ib_mr *mthca_reg_phys_mr(struct ib_pd       *pd,
 927                                        struct ib_phys_buf *buffer_list,
 928                                        int                 num_phys_buf,
 929                                        int                 acc,
 930                                        u64                *iova_start)
 931 {
 932         struct mthca_mr *mr;
 933         u64 *page_list;
 934         u64 total_size;
 935         unsigned long mask;
 936         int shift;
 937         int npages;
 938         int err;
 939         int i, j, n;
 940
 941         mask = buffer_list[0].addr ^ *iova_start;
 942         total_size = 0;
 943         for (i = 0; i < num_phys_buf; ++i) {
 944                 if (i != 0)
 945                         mask |= buffer_list[i].addr;
 946                 if (i != num_phys_buf - 1)
 947                         mask |= buffer_list[i].addr + buffer_list[i].size;
 948
 949                 total_size += buffer_list[i].size;
 950         }
 951
 952         if (mask & ~PAGE_MASK)
 953                 return ERR_PTR(-EINVAL);
 954
 955         shift = __ffs(mask | 1 << 31);
 956
 957         buffer_list[0].size += buffer_list[0].addr & ((1ULL << shift) - 1);
 958         buffer_list[0].addr &= ~0ull << shift;
 959
 960         mr = kmalloc(sizeof *mr, GFP_KERNEL);
 961         if (!mr)
 962                 return ERR_PTR(-ENOMEM);
 963
 964         npages = 0;
 965         for (i = 0; i < num_phys_buf; ++i)
 966                 npages += (buffer_list[i].size + (1ULL << shift) - 1) >> shift;
 967
 968         if (!npages)
 969                 return &mr->ibmr;
 970
 971         page_list = kmalloc(npages * sizeof *page_list, GFP_KERNEL);
 972         if (!page_list) {
 973                 kfree(mr);
 974                 return ERR_PTR(-ENOMEM);
 975         }
 976
 977         n = 0;
 978         for (i = 0; i < num_phys_buf; ++i)
 979                 for (j = 0;
 980                      j < (buffer_list[i].size + (1ULL << shift) - 1) >> shift;
 981                      ++j)
 982                         page_list[n++] = buffer_list[i].addr + ((u64) j << shift);
 983
 984         mthca_dbg(to_mdev(pd->device), "Registering memory at %llx (iova %llx) "
 985                   "in PD %x; shift %d, npages %d.\n",
 986                   (unsigned long long) buffer_list[0].addr,
 987                   (unsigned long long) *iova_start,
 988                   to_mpd(pd)->pd_num,
 989                   shift, npages);
 990
 991         err = mthca_mr_alloc_phys(to_mdev(pd->device),
 992                                   to_mpd(pd)->pd_num,
 993                                   page_list, shift, npages,
 994                                   *iova_start, total_size,
 995                                   convert_access(acc), mr);
 996
 997         if (err) {
 998                 kfree(page_list);
 999                 kfree(mr);
1000                 return ERR_PTR(err);
1001         }
1002
1003         kfree(page_list);
1004         mr->umem = NULL;
1005
1006         return &mr->ibmr;
1007 }
1008
1009 static struct ib_mr *mthca_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
1010                                        u64 virt, int acc, struct ib_udata *udata)
1011 {
1012         struct mthca_dev *dev = to_mdev(pd->device);
1013         struct ib_umem_chunk *chunk;
1014         struct mthca_mr *mr;
1015         struct mthca_reg_mr ucmd;
1016         u64 *pages;
1017         int shift, n, len;
1018         int i, j, k;
1019         int err = 0;
1020         int write_mtt_size;
1021
1022         if (udata->inlen - sizeof (struct ib_uverbs_cmd_hdr) < sizeof ucmd) {
1023                 if (!to_mucontext(pd->uobject->context)->reg_mr_warned) {
1024                         mthca_warn(dev, "Process '%s' did not pass in MR attrs.\n",
1025                                    current->comm);
1026                         mthca_warn(dev, "  Update libmthca to fix this.\n");
1027                 }
1028                 ++to_mucontext(pd->uobject->context)->reg_mr_warned;
1029                 ucmd.mr_attrs = 0;
1030         } else if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd))
1031                 return ERR_PTR(-EFAULT);
1032
1033         mr = kmalloc(sizeof *mr, GFP_KERNEL);
1034         if (!mr)
1035                 return ERR_PTR(-ENOMEM);
1036
1037         mr->umem = ib_umem_get(pd->uobject->context, start, length, acc,
1038                                ucmd.mr_attrs & MTHCA_MR_DMASYNC);
1039
1040         if (IS_ERR(mr->umem)) {
1041                 err = PTR_ERR(mr->umem);
1042                 goto err;
1043         }
1044
1045         shift = ffs(mr->umem->page_size) - 1;
1046
1047         n = 0;
1048         list_for_each_entry(chunk, &mr->umem->chunk_list, list)
1049                 n += chunk->nents;
1050
1051         mr->mtt = mthca_alloc_mtt(dev, n);
1052         if (IS_ERR(mr->mtt)) {
1053                 err = PTR_ERR(mr->mtt);
1054                 goto err_umem;
1055         }
1056
1057         pages = (u64 *) __get_free_page(GFP_KERNEL);
1058         if (!pages) {
1059                 err = -ENOMEM;
1060                 goto err_mtt;
1061         }
1062
1063         i = n = 0;
1064
1065         write_mtt_size = min(mthca_write_mtt_size(dev), (int) (PAGE_SIZE / sizeof *pages));
1066
1067         list_for_each_entry(chunk, &mr->umem->chunk_list, list)
1068                 for (j = 0; j < chunk->nmap; ++j) {
1069                         len = sg_dma_len(&chunk->page_list[j]) >> shift;
1070                         for (k = 0; k < len; ++k) {
1071                                 pages[i++] = sg_dma_address(&chunk->page_list[j]) +
1072                                         mr->umem->page_size * k;
1073                                 /*
1074                                  * Be friendly to write_mtt and pass it chunks
1075                                  * of appropriate size.
1076                                  */
1077                                 if (i == write_mtt_size) {
1078                                         err = mthca_write_mtt(dev, mr->mtt, n, pages, i);
1079                                         if (err)
1080                                                 goto mtt_done;
1081                                         n += i;
1082                                         i = 0;
1083                                 }
1084                         }
1085                 }
1086
1087         if (i)
1088                 err = mthca_write_mtt(dev, mr->mtt, n, pages, i);
1089 mtt_done:
1090         free_page((unsigned long) pages);
1091         if (err)
1092                 goto err_mtt;
1093
1094         err = mthca_mr_alloc(dev, to_mpd(pd)->pd_num, shift, virt, length,
1095                              convert_access(acc), mr);
1096
1097         if (err)
1098                 goto err_mtt;
1099
1100         return &mr->ibmr;
1101
1102 err_mtt:
1103         mthca_free_mtt(dev, mr->mtt);
1104
1105 err_umem:
1106         ib_umem_release(mr->umem);
1107
1108 err:
1109         kfree(mr);
1110         return ERR_PTR(err);
1111 }
1112
1113 static int mthca_dereg_mr(struct ib_mr *mr)
1114 {
1115         struct mthca_mr *mmr = to_mmr(mr);
1116
1117         mthca_free_mr(to_mdev(mr->device), mmr);
1118         if (mmr->umem)
1119                 ib_umem_release(mmr->umem);
1120         kfree(mmr);
1121
1122         return 0;
1123 }
1124
1125 static struct ib_fmr *mthca_alloc_fmr(struct ib_pd *pd, int mr_access_flags,
1126                                       struct ib_fmr_attr *fmr_attr)
1127 {
1128         struct mthca_fmr *fmr;
1129         int err;
1130
1131         fmr = kmalloc(sizeof *fmr, GFP_KERNEL);
1132         if (!fmr)
1133                 return ERR_PTR(-ENOMEM);
1134
1135         memcpy(&fmr->attr, fmr_attr, sizeof *fmr_attr);
1136         err = mthca_fmr_alloc(to_mdev(pd->device), to_mpd(pd)->pd_num,
1137                              convert_access(mr_access_flags), fmr);
1138
1139         if (err) {
1140                 kfree(fmr);
1141                 return ERR_PTR(err);
1142         }
1143
1144         return &fmr->ibmr;
1145 }
1146
1147 static int mthca_dealloc_fmr(struct ib_fmr *fmr)
1148 {
1149         struct mthca_fmr *mfmr = to_mfmr(fmr);
1150         int err;
1151
1152         err = mthca_free_fmr(to_mdev(fmr->device), mfmr);
1153         if (err)
1154                 return err;
1155
1156         kfree(mfmr);
1157         return 0;
1158 }
1159
1160 static int mthca_unmap_fmr(struct list_head *fmr_list)
1161 {
1162         struct ib_fmr *fmr;
1163         int err;
1164         u8 status;
1165         struct mthca_dev *mdev = NULL;
1166
1167         list_for_each_entry(fmr, fmr_list, list) {
1168                 if (mdev && to_mdev(fmr->device) != mdev)
1169                         return -EINVAL;
1170                 mdev = to_mdev(fmr->device);
1171         }
1172
1173         if (!mdev)
1174                 return 0;
1175
1176         if (mthca_is_memfree(mdev)) {
1177                 list_for_each_entry(fmr, fmr_list, list)
1178                         mthca_arbel_fmr_unmap(mdev, to_mfmr(fmr));
1179
1180                 wmb();
1181         } else
1182                 list_for_each_entry(fmr, fmr_list, list)
1183                         mthca_tavor_fmr_unmap(mdev, to_mfmr(fmr));
1184
1185         err = mthca_SYNC_TPT(mdev, &status);
1186         if (err)
1187                 return err;
1188         if (status)
1189                 return -EINVAL;
1190         return 0;
1191 }
1192
1193 static ssize_t show_rev(struct device *device, struct device_attribute *attr,
1194                         char *buf)
1195 {
1196         struct mthca_dev *dev =
1197                 container_of(device, struct mthca_dev, ib_dev.dev);
1198         return sprintf(buf, "%x\n", dev->rev_id);
1199 }
1200
1201 static ssize_t show_fw_ver(struct device *device, struct device_attribute *attr,
1202                            char *buf)
1203 {
1204         struct mthca_dev *dev =
1205                 container_of(device, struct mthca_dev, ib_dev.dev);
1206         return sprintf(buf, "%d.%d.%d\n", (int) (dev->fw_ver >> 32),
1207                        (int) (dev->fw_ver >> 16) & 0xffff,
1208                        (int) dev->fw_ver & 0xffff);
1209 }
1210
1211 static ssize_t show_hca(struct device *device, struct device_attribute *attr,
1212                         char *buf)
1213 {
1214         struct mthca_dev *dev =
1215                 container_of(device, struct mthca_dev, ib_dev.dev);
1216         switch (dev->pdev->device) {
1217         case PCI_DEVICE_ID_MELLANOX_TAVOR:
1218                 return sprintf(buf, "MT23108\n");
1219         case PCI_DEVICE_ID_MELLANOX_ARBEL_COMPAT:
1220                 return sprintf(buf, "MT25208 (MT23108 compat mode)\n");
1221         case PCI_DEVICE_ID_MELLANOX_ARBEL:
1222                 return sprintf(buf, "MT25208\n");
1223         case PCI_DEVICE_ID_MELLANOX_SINAI:
1224         case PCI_DEVICE_ID_MELLANOX_SINAI_OLD:
1225                 return sprintf(buf, "MT25204\n");
1226         default:
1227                 return sprintf(buf, "unknown\n");
1228         }
1229 }
1230
1231 static ssize_t show_board(struct device *device, struct device_attribute *attr,
1232                           char *buf)
1233 {
1234         struct mthca_dev *dev =
1235                 container_of(device, struct mthca_dev, ib_dev.dev);
1236         return sprintf(buf, "%.*s\n", MTHCA_BOARD_ID_LEN, dev->board_id);
1237 }
1238
1239 static DEVICE_ATTR(hw_rev,   S_IRUGO, show_rev,    NULL);
1240 static DEVICE_ATTR(fw_ver,   S_IRUGO, show_fw_ver, NULL);
1241 static DEVICE_ATTR(hca_type, S_IRUGO, show_hca,    NULL);
1242 static DEVICE_ATTR(board_id, S_IRUGO, show_board,  NULL);
1243
1244 static struct device_attribute *mthca_dev_attributes[] = {
1245         &dev_attr_hw_rev,
1246         &dev_attr_fw_ver,
1247         &dev_attr_hca_type,
1248         &dev_attr_board_id
1249 };
1250
1251 static int mthca_init_node_data(struct mthca_dev *dev)
1252 {
1253         struct ib_smp *in_mad  = NULL;
1254         struct ib_smp *out_mad = NULL;
1255         int err = -ENOMEM;
1256         u8 status;
1257
1258         in_mad  = kzalloc(sizeof *in_mad, GFP_KERNEL);
1259         out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
1260         if (!in_mad || !out_mad)
1261                 goto out;
1262
1263         init_query_mad(in_mad);
1264         in_mad->attr_id = IB_SMP_ATTR_NODE_DESC;
1265
1266         err = mthca_MAD_IFC(dev, 1, 1,
1267                             1, NULL, NULL, in_mad, out_mad,
1268                             &status);
1269         if (err)
1270                 goto out;
1271         if (status) {
1272                 err = -EINVAL;
1273                 goto out;
1274         }
1275
1276         memcpy(dev->ib_dev.node_desc, out_mad->data, 64);
1277
1278         in_mad->attr_id = IB_SMP_ATTR_NODE_INFO;
1279
1280         err = mthca_MAD_IFC(dev, 1, 1,
1281                             1, NULL, NULL, in_mad, out_mad,
1282                             &status);
1283         if (err)
1284                 goto out;
1285         if (status) {
1286                 err = -EINVAL;
1287                 goto out;
1288         }
1289
1290         if (mthca_is_memfree(dev))
1291                 dev->rev_id = be32_to_cpup((__be32 *) (out_mad->data + 32));
1292         memcpy(&dev->ib_dev.node_guid, out_mad->data + 12, 8);
1293
1294 out:
1295         kfree(in_mad);
1296         kfree(out_mad);
1297         return err;
1298 }
1299
1300 int mthca_register_device(struct mthca_dev *dev)
1301 {
1302         int ret;
1303         int i;
1304
1305         ret = mthca_init_node_data(dev);
1306         if (ret)
1307                 return ret;
1308
1309         strlcpy(dev->ib_dev.name, "mthca%d", IB_DEVICE_NAME_MAX);
1310         dev->ib_dev.owner                = THIS_MODULE;
1311
1312         dev->ib_dev.uverbs_abi_ver       = MTHCA_UVERBS_ABI_VERSION;
1313         dev->ib_dev.uverbs_cmd_mask      =
1314                 (1ull << IB_USER_VERBS_CMD_GET_CONTEXT)         |
1315                 (1ull << IB_USER_VERBS_CMD_QUERY_DEVICE)        |
1316                 (1ull << IB_USER_VERBS_CMD_QUERY_PORT)          |
1317                 (1ull << IB_USER_VERBS_CMD_ALLOC_PD)            |
1318                 (1ull << IB_USER_VERBS_CMD_DEALLOC_PD)          |
1319                 (1ull << IB_USER_VERBS_CMD_REG_MR)              |
1320                 (1ull << IB_USER_VERBS_CMD_DEREG_MR)            |
1321                 (1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL) |
1322                 (1ull << IB_USER_VERBS_CMD_CREATE_CQ)           |
1323                 (1ull << IB_USER_VERBS_CMD_RESIZE_CQ)           |
1324                 (1ull << IB_USER_VERBS_CMD_DESTROY_CQ)          |
1325                 (1ull << IB_USER_VERBS_CMD_CREATE_QP)           |
1326                 (1ull << IB_USER_VERBS_CMD_QUERY_QP)            |
1327                 (1ull << IB_USER_VERBS_CMD_MODIFY_QP)           |
1328                 (1ull << IB_USER_VERBS_CMD_DESTROY_QP)          |
1329                 (1ull << IB_USER_VERBS_CMD_ATTACH_MCAST)        |
1330                 (1ull << IB_USER_VERBS_CMD_DETACH_MCAST);
1331         dev->ib_dev.node_type            = RDMA_NODE_IB_CA;
1332         dev->ib_dev.phys_port_cnt        = dev->limits.num_ports;
1333         dev->ib_dev.num_comp_vectors     = 1;
1334         dev->ib_dev.dma_device           = &dev->pdev->dev;
1335         dev->ib_dev.query_device         = mthca_query_device;
1336         dev->ib_dev.query_port           = mthca_query_port;
1337         dev->ib_dev.modify_device        = mthca_modify_device;
1338         dev->ib_dev.modify_port          = mthca_modify_port;
1339         dev->ib_dev.query_pkey           = mthca_query_pkey;
1340         dev->ib_dev.query_gid            = mthca_query_gid;
1341         dev->ib_dev.alloc_ucontext       = mthca_alloc_ucontext;
1342         dev->ib_dev.dealloc_ucontext     = mthca_dealloc_ucontext;
1343         dev->ib_dev.mmap                 = mthca_mmap_uar;
1344         dev->ib_dev.alloc_pd             = mthca_alloc_pd;
1345         dev->ib_dev.dealloc_pd           = mthca_dealloc_pd;
1346         dev->ib_dev.create_ah            = mthca_ah_create;
1347         dev->ib_dev.query_ah             = mthca_ah_query;
1348         dev->ib_dev.destroy_ah           = mthca_ah_destroy;
1349
1350         if (dev->mthca_flags & MTHCA_FLAG_SRQ) {
1351                 dev->ib_dev.create_srq           = mthca_create_srq;
1352                 dev->ib_dev.modify_srq           = mthca_modify_srq;
1353                 dev->ib_dev.query_srq            = mthca_query_srq;
1354                 dev->ib_dev.destroy_srq          = mthca_destroy_srq;
1355                 dev->ib_dev.uverbs_cmd_mask     |=
1356                         (1ull << IB_USER_VERBS_CMD_CREATE_SRQ)          |
1357                         (1ull << IB_USER_VERBS_CMD_MODIFY_SRQ)          |
1358                         (1ull << IB_USER_VERBS_CMD_QUERY_SRQ)           |
1359                         (1ull << IB_USER_VERBS_CMD_DESTROY_SRQ);
1360
1361                 if (mthca_is_memfree(dev))
1362                         dev->ib_dev.post_srq_recv = mthca_arbel_post_srq_recv;
1363                 else
1364                         dev->ib_dev.post_srq_recv = mthca_tavor_post_srq_recv;
1365         }
1366
1367         dev->ib_dev.create_qp            = mthca_create_qp;
1368         dev->ib_dev.modify_qp            = mthca_modify_qp;
1369         dev->ib_dev.query_qp             = mthca_query_qp;
1370         dev->ib_dev.destroy_qp           = mthca_destroy_qp;
1371         dev->ib_dev.create_cq            = mthca_create_cq;
1372         dev->ib_dev.resize_cq            = mthca_resize_cq;
1373         dev->ib_dev.destroy_cq           = mthca_destroy_cq;
1374         dev->ib_dev.poll_cq              = mthca_poll_cq;
1375         dev->ib_dev.get_dma_mr           = mthca_get_dma_mr;
1376         dev->ib_dev.reg_phys_mr          = mthca_reg_phys_mr;
1377         dev->ib_dev.reg_user_mr          = mthca_reg_user_mr;
1378         dev->ib_dev.dereg_mr             = mthca_dereg_mr;
1379
1380         if (dev->mthca_flags & MTHCA_FLAG_FMR) {
1381                 dev->ib_dev.alloc_fmr            = mthca_alloc_fmr;
1382                 dev->ib_dev.unmap_fmr            = mthca_unmap_fmr;
1383                 dev->ib_dev.dealloc_fmr          = mthca_dealloc_fmr;
1384                 if (mthca_is_memfree(dev))
1385                         dev->ib_dev.map_phys_fmr = mthca_arbel_map_phys_fmr;
1386                 else
1387                         dev->ib_dev.map_phys_fmr = mthca_tavor_map_phys_fmr;
1388         }
1389
1390         dev->ib_dev.attach_mcast         = mthca_multicast_attach;
1391         dev->ib_dev.detach_mcast         = mthca_multicast_detach;
1392         dev->ib_dev.process_mad          = mthca_process_mad;
1393
1394         if (mthca_is_memfree(dev)) {
1395                 dev->ib_dev.req_notify_cq = mthca_arbel_arm_cq;
1396                 dev->ib_dev.post_send     = mthca_arbel_post_send;
1397                 dev->ib_dev.post_recv     = mthca_arbel_post_receive;
1398         } else {
1399                 dev->ib_dev.req_notify_cq = mthca_tavor_arm_cq;
1400                 dev->ib_dev.post_send     = mthca_tavor_post_send;
1401                 dev->ib_dev.post_recv     = mthca_tavor_post_receive;
1402         }
1403
1404         mutex_init(&dev->cap_mask_mutex);
1405
1406         ret = ib_register_device(&dev->ib_dev);
1407         if (ret)
1408                 return ret;
1409
1410         for (i = 0; i < ARRAY_SIZE(mthca_dev_attributes); ++i) {
1411                 ret = device_create_file(&dev->ib_dev.dev,
1412                                          mthca_dev_attributes[i]);
1413                 if (ret) {
1414                         ib_unregister_device(&dev->ib_dev);
1415                         return ret;
1416                 }
1417         }
1418
1419         mthca_start_catas_poll(dev);
1420
1421         return 0;
1422 }
1423
1424 void mthca_unregister_device(struct mthca_dev *dev)
1425 {
1426         mthca_stop_catas_poll(dev);
1427         ib_unregister_device(&dev->ib_dev);
1428 }