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