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