err.no Git - linux-2.6/blob - drivers/infiniband/hw/mthca/mthca_provider.c

   1 /*
   2  * Copyright (c) 2004, 2005 Topspin Communications.  All rights reserved.
   3  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
   4  * Copyright (c) 2005 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  * $Id: mthca_provider.c 1397 2004-12-28 05:09:00Z roland $
  37  */
  38
  39 #include <rdma/ib_smi.h>
  40 #include <linux/mm.h>
  41
  42 #include "mthca_dev.h"
  43 #include "mthca_cmd.h"
  44 #include "mthca_user.h"
  45 #include "mthca_memfree.h"
  46
  47 static int mthca_query_device(struct ib_device *ibdev,
  48                               struct ib_device_attr *props)
  49 {
  50         struct ib_smp *in_mad  = NULL;
  51         struct ib_smp *out_mad = NULL;
  52         int err = -ENOMEM;
  53         struct mthca_dev* mdev = to_mdev(ibdev);
  54
  55         u8 status;
  56
  57         in_mad  = kmalloc(sizeof *in_mad, GFP_KERNEL);
  58         out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
  59         if (!in_mad || !out_mad)
  60                 goto out;
  61
  62         memset(props, 0, sizeof *props);
  63
  64         props->fw_ver              = mdev->fw_ver;
  65
  66         memset(in_mad, 0, sizeof *in_mad);
  67         in_mad->base_version       = 1;
  68         in_mad->mgmt_class         = IB_MGMT_CLASS_SUBN_LID_ROUTED;
  69         in_mad->class_version      = 1;
  70         in_mad->method             = IB_MGMT_METHOD_GET;
  71         in_mad->attr_id            = IB_SMP_ATTR_NODE_INFO;
  72
  73         err = mthca_MAD_IFC(mdev, 1, 1,
  74                             1, NULL, NULL, in_mad, out_mad,
  75                             &status);
  76         if (err)
  77                 goto out;
  78         if (status) {
  79                 err = -EINVAL;
  80                 goto out;
  81         }
  82
  83         props->device_cap_flags    = mdev->device_cap_flags;
  84         props->vendor_id           = be32_to_cpup((__be32 *) (out_mad->data + 36)) &
  85                 0xffffff;
  86         props->vendor_part_id      = be16_to_cpup((__be16 *) (out_mad->data + 30));
  87         props->hw_ver              = be32_to_cpup((__be32 *) (out_mad->data + 32));
  88         memcpy(&props->sys_image_guid, out_mad->data +  4, 8);
  89         memcpy(&props->node_guid,      out_mad->data + 12, 8);
  90
  91         props->max_mr_size         = ~0ull;
  92         props->max_qp              = mdev->limits.num_qps - mdev->limits.reserved_qps;
  93         props->max_qp_wr           = 0xffff;
  94         props->max_sge             = mdev->limits.max_sg;
  95         props->max_cq              = mdev->limits.num_cqs - mdev->limits.reserved_cqs;
  96         props->max_cqe             = 0xffff;
  97         props->max_mr              = mdev->limits.num_mpts - mdev->limits.reserved_mrws;
  98         props->max_pd              = mdev->limits.num_pds - mdev->limits.reserved_pds;
  99         props->max_qp_rd_atom      = 1 << mdev->qp_table.rdb_shift;
 100         props->max_qp_init_rd_atom = 1 << mdev->qp_table.rdb_shift;
 101         props->local_ca_ack_delay  = mdev->limits.local_ca_ack_delay;
 102         props->atomic_cap          = mdev->limits.flags & DEV_LIM_FLAG_ATOMIC ?
 103                                         IB_ATOMIC_HCA : IB_ATOMIC_NONE;
 104
 105         err = 0;
 106  out:
 107         kfree(in_mad);
 108         kfree(out_mad);
 109         return err;
 110 }
 111
 112 static int mthca_query_port(struct ib_device *ibdev,
 113                             u8 port, struct ib_port_attr *props)
 114 {
 115         struct ib_smp *in_mad  = NULL;
 116         struct ib_smp *out_mad = NULL;
 117         int err = -ENOMEM;
 118         u8 status;
 119
 120         in_mad  = kmalloc(sizeof *in_mad, GFP_KERNEL);
 121         out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
 122         if (!in_mad || !out_mad)
 123                 goto out;
 124
 125         memset(props, 0, sizeof *props);
 126
 127         memset(in_mad, 0, sizeof *in_mad);
 128         in_mad->base_version       = 1;
 129         in_mad->mgmt_class         = IB_MGMT_CLASS_SUBN_LID_ROUTED;
 130         in_mad->class_version      = 1;
 131         in_mad->method             = IB_MGMT_METHOD_GET;
 132         in_mad->attr_id            = IB_SMP_ATTR_PORT_INFO;
 133         in_mad->attr_mod           = cpu_to_be32(port);
 134
 135         err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
 136                             port, NULL, NULL, in_mad, out_mad,
 137                             &status);
 138         if (err)
 139                 goto out;
 140         if (status) {
 141                 err = -EINVAL;
 142                 goto out;
 143         }
 144
 145         props->lid               = be16_to_cpup((__be16 *) (out_mad->data + 16));
 146         props->lmc               = out_mad->data[34] & 0x7;
 147         props->sm_lid            = be16_to_cpup((__be16 *) (out_mad->data + 18));
 148         props->sm_sl             = out_mad->data[36] & 0xf;
 149         props->state             = out_mad->data[32] & 0xf;
 150         props->phys_state        = out_mad->data[33] >> 4;
 151         props->port_cap_flags    = be32_to_cpup((__be32 *) (out_mad->data + 20));
 152         props->gid_tbl_len       = to_mdev(ibdev)->limits.gid_table_len;
 153         props->max_msg_sz        = 0x80000000;
 154         props->pkey_tbl_len      = to_mdev(ibdev)->limits.pkey_table_len;
 155         props->bad_pkey_cntr     = be16_to_cpup((__be16 *) (out_mad->data + 46));
 156         props->qkey_viol_cntr    = be16_to_cpup((__be16 *) (out_mad->data + 48));
 157         props->active_width      = out_mad->data[31] & 0xf;
 158         props->active_speed      = out_mad->data[35] >> 4;
 159         props->max_mtu           = out_mad->data[41] & 0xf;
 160         props->active_mtu        = out_mad->data[36] >> 4;
 161         props->subnet_timeout    = out_mad->data[51] & 0x1f;
 162
 163  out:
 164         kfree(in_mad);
 165         kfree(out_mad);
 166         return err;
 167 }
 168
 169 static int mthca_modify_port(struct ib_device *ibdev,
 170                              u8 port, int port_modify_mask,
 171                              struct ib_port_modify *props)
 172 {
 173         struct mthca_set_ib_param set_ib;
 174         struct ib_port_attr attr;
 175         int err;
 176         u8 status;
 177
 178         if (down_interruptible(&to_mdev(ibdev)->cap_mask_mutex))
 179                 return -ERESTARTSYS;
 180
 181         err = mthca_query_port(ibdev, port, &attr);
 182         if (err)
 183                 goto out;
 184
 185         set_ib.set_si_guid     = 0;
 186         set_ib.reset_qkey_viol = !!(port_modify_mask & IB_PORT_RESET_QKEY_CNTR);
 187
 188         set_ib.cap_mask = (attr.port_cap_flags | props->set_port_cap_mask) &
 189                 ~props->clr_port_cap_mask;
 190
 191         err = mthca_SET_IB(to_mdev(ibdev), &set_ib, port, &status);
 192         if (err)
 193                 goto out;
 194         if (status) {
 195                 err = -EINVAL;
 196                 goto out;
 197         }
 198
 199 out:
 200         up(&to_mdev(ibdev)->cap_mask_mutex);
 201         return err;
 202 }
 203
 204 static int mthca_query_pkey(struct ib_device *ibdev,
 205                             u8 port, u16 index, u16 *pkey)
 206 {
 207         struct ib_smp *in_mad  = NULL;
 208         struct ib_smp *out_mad = NULL;
 209         int err = -ENOMEM;
 210         u8 status;
 211
 212         in_mad  = kmalloc(sizeof *in_mad, GFP_KERNEL);
 213         out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
 214         if (!in_mad || !out_mad)
 215                 goto out;
 216
 217         memset(in_mad, 0, sizeof *in_mad);
 218         in_mad->base_version       = 1;
 219         in_mad->mgmt_class         = IB_MGMT_CLASS_SUBN_LID_ROUTED;
 220         in_mad->class_version      = 1;
 221         in_mad->method             = IB_MGMT_METHOD_GET;
 222         in_mad->attr_id            = IB_SMP_ATTR_PKEY_TABLE;
 223         in_mad->attr_mod           = cpu_to_be32(index / 32);
 224
 225         err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
 226                             port, NULL, NULL, in_mad, out_mad,
 227                             &status);
 228         if (err)
 229                 goto out;
 230         if (status) {
 231                 err = -EINVAL;
 232                 goto out;
 233         }
 234
 235         *pkey = be16_to_cpu(((__be16 *) out_mad->data)[index % 32]);
 236
 237  out:
 238         kfree(in_mad);
 239         kfree(out_mad);
 240         return err;
 241 }
 242
 243 static int mthca_query_gid(struct ib_device *ibdev, u8 port,
 244                            int index, union ib_gid *gid)
 245 {
 246         struct ib_smp *in_mad  = NULL;
 247         struct ib_smp *out_mad = NULL;
 248         int err = -ENOMEM;
 249         u8 status;
 250
 251         in_mad  = kmalloc(sizeof *in_mad, GFP_KERNEL);
 252         out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
 253         if (!in_mad || !out_mad)
 254                 goto out;
 255
 256         memset(in_mad, 0, sizeof *in_mad);
 257         in_mad->base_version       = 1;
 258         in_mad->mgmt_class         = IB_MGMT_CLASS_SUBN_LID_ROUTED;
 259         in_mad->class_version      = 1;
 260         in_mad->method             = IB_MGMT_METHOD_GET;
 261         in_mad->attr_id            = IB_SMP_ATTR_PORT_INFO;
 262         in_mad->attr_mod           = cpu_to_be32(port);
 263
 264         err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
 265                             port, NULL, NULL, in_mad, out_mad,
 266                             &status);
 267         if (err)
 268                 goto out;
 269         if (status) {
 270                 err = -EINVAL;
 271                 goto out;
 272         }
 273
 274         memcpy(gid->raw, out_mad->data + 8, 8);
 275
 276         memset(in_mad, 0, sizeof *in_mad);
 277         in_mad->base_version       = 1;
 278         in_mad->mgmt_class         = IB_MGMT_CLASS_SUBN_LID_ROUTED;
 279         in_mad->class_version      = 1;
 280         in_mad->method             = IB_MGMT_METHOD_GET;
 281         in_mad->attr_id            = IB_SMP_ATTR_GUID_INFO;
 282         in_mad->attr_mod           = cpu_to_be32(index / 8);
 283
 284         err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
 285                             port, NULL, NULL, in_mad, out_mad,
 286                             &status);
 287         if (err)
 288                 goto out;
 289         if (status) {
 290                 err = -EINVAL;
 291                 goto out;
 292         }
 293
 294         memcpy(gid->raw + 8, out_mad->data + (index % 8) * 16, 8);
 295
 296  out:
 297         kfree(in_mad);
 298         kfree(out_mad);
 299         return err;
 300 }
 301
 302 static struct ib_ucontext *mthca_alloc_ucontext(struct ib_device *ibdev,
 303                                                 struct ib_udata *udata)
 304 {
 305         struct mthca_alloc_ucontext_resp uresp;
 306         struct mthca_ucontext           *context;
 307         int                              err;
 308
 309         memset(&uresp, 0, sizeof uresp);
 310
 311         uresp.qp_tab_size = to_mdev(ibdev)->limits.num_qps;
 312         if (mthca_is_memfree(to_mdev(ibdev)))
 313                 uresp.uarc_size = to_mdev(ibdev)->uar_table.uarc_size;
 314         else
 315                 uresp.uarc_size = 0;
 316
 317         context = kmalloc(sizeof *context, GFP_KERNEL);
 318         if (!context)
 319                 return ERR_PTR(-ENOMEM);
 320
 321         err = mthca_uar_alloc(to_mdev(ibdev), &context->uar);
 322         if (err) {
 323                 kfree(context);
 324                 return ERR_PTR(err);
 325         }
 326
 327         context->db_tab = mthca_init_user_db_tab(to_mdev(ibdev));
 328         if (IS_ERR(context->db_tab)) {
 329                 err = PTR_ERR(context->db_tab);
 330                 mthca_uar_free(to_mdev(ibdev), &context->uar);
 331                 kfree(context);
 332                 return ERR_PTR(err);
 333         }
 334
 335         if (ib_copy_to_udata(udata, &uresp, sizeof uresp)) {
 336                 mthca_cleanup_user_db_tab(to_mdev(ibdev), &context->uar, context->db_tab);
 337                 mthca_uar_free(to_mdev(ibdev), &context->uar);
 338                 kfree(context);
 339                 return ERR_PTR(-EFAULT);
 340         }
 341
 342         return &context->ibucontext;
 343 }
 344
 345 static int mthca_dealloc_ucontext(struct ib_ucontext *context)
 346 {
 347         mthca_cleanup_user_db_tab(to_mdev(context->device), &to_mucontext(context)->uar,
 348                                   to_mucontext(context)->db_tab);
 349         mthca_uar_free(to_mdev(context->device), &to_mucontext(context)->uar);
 350         kfree(to_mucontext(context));
 351
 352         return 0;
 353 }
 354
 355 static int mthca_mmap_uar(struct ib_ucontext *context,
 356                           struct vm_area_struct *vma)
 357 {
 358         if (vma->vm_end - vma->vm_start != PAGE_SIZE)
 359                 return -EINVAL;
 360
 361         vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
 362
 363         if (io_remap_pfn_range(vma, vma->vm_start,
 364                                to_mucontext(context)->uar.pfn,
 365                                PAGE_SIZE, vma->vm_page_prot))
 366                 return -EAGAIN;
 367
 368         return 0;
 369 }
 370
 371 static struct ib_pd *mthca_alloc_pd(struct ib_device *ibdev,
 372                                     struct ib_ucontext *context,
 373                                     struct ib_udata *udata)
 374 {
 375         struct mthca_pd *pd;
 376         int err;
 377
 378         pd = kmalloc(sizeof *pd, GFP_KERNEL);
 379         if (!pd)
 380                 return ERR_PTR(-ENOMEM);
 381
 382         err = mthca_pd_alloc(to_mdev(ibdev), !context, pd);
 383         if (err) {
 384                 kfree(pd);
 385                 return ERR_PTR(err);
 386         }
 387
 388         if (context) {
 389                 if (ib_copy_to_udata(udata, &pd->pd_num, sizeof (__u32))) {
 390                         mthca_pd_free(to_mdev(ibdev), pd);
 391                         kfree(pd);
 392                         return ERR_PTR(-EFAULT);
 393                 }
 394         }
 395
 396         return &pd->ibpd;
 397 }
 398
 399 static int mthca_dealloc_pd(struct ib_pd *pd)
 400 {
 401         mthca_pd_free(to_mdev(pd->device), to_mpd(pd));
 402         kfree(pd);
 403
 404         return 0;
 405 }
 406
 407 static struct ib_ah *mthca_ah_create(struct ib_pd *pd,
 408                                      struct ib_ah_attr *ah_attr)
 409 {
 410         int err;
 411         struct mthca_ah *ah;
 412
 413         ah = kmalloc(sizeof *ah, GFP_ATOMIC);
 414         if (!ah)
 415                 return ERR_PTR(-ENOMEM);
 416
 417         err = mthca_create_ah(to_mdev(pd->device), to_mpd(pd), ah_attr, ah);
 418         if (err) {
 419                 kfree(ah);
 420                 return ERR_PTR(err);
 421         }
 422
 423         return &ah->ibah;
 424 }
 425
 426 static int mthca_ah_destroy(struct ib_ah *ah)
 427 {
 428         mthca_destroy_ah(to_mdev(ah->device), to_mah(ah));
 429         kfree(ah);
 430
 431         return 0;
 432 }
 433
 434 static struct ib_srq *mthca_create_srq(struct ib_pd *pd,
 435                                        struct ib_srq_init_attr *init_attr,
 436                                        struct ib_udata *udata)
 437 {
 438         struct mthca_create_srq ucmd;
 439         struct mthca_ucontext *context = NULL;
 440         struct mthca_srq *srq;
 441         int err;
 442
 443         srq = kmalloc(sizeof *srq, GFP_KERNEL);
 444         if (!srq)
 445                 return ERR_PTR(-ENOMEM);
 446
 447         if (pd->uobject) {
 448                 context = to_mucontext(pd->uobject->context);
 449
 450                 if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd))
 451                         return ERR_PTR(-EFAULT);
 452
 453                 err = mthca_map_user_db(to_mdev(pd->device), &context->uar,
 454                                         context->db_tab, ucmd.db_index,
 455                                         ucmd.db_page);
 456
 457                 if (err)
 458                         goto err_free;
 459
 460                 srq->mr.ibmr.lkey = ucmd.lkey;
 461                 srq->db_index     = ucmd.db_index;
 462         }
 463
 464         err = mthca_alloc_srq(to_mdev(pd->device), to_mpd(pd),
 465                               &init_attr->attr, srq);
 466
 467         if (err && pd->uobject)
 468                 mthca_unmap_user_db(to_mdev(pd->device), &context->uar,
 469                                     context->db_tab, ucmd.db_index);
 470
 471         if (err)
 472                 goto err_free;
 473
 474         if (context && ib_copy_to_udata(udata, &srq->srqn, sizeof (__u32))) {
 475                 mthca_free_srq(to_mdev(pd->device), srq);
 476                 err = -EFAULT;
 477                 goto err_free;
 478         }
 479
 480         return &srq->ibsrq;
 481
 482 err_free:
 483         kfree(srq);
 484
 485         return ERR_PTR(err);
 486 }
 487
 488 static int mthca_destroy_srq(struct ib_srq *srq)
 489 {
 490         struct mthca_ucontext *context;
 491
 492         if (srq->uobject) {
 493                 context = to_mucontext(srq->uobject->context);
 494
 495                 mthca_unmap_user_db(to_mdev(srq->device), &context->uar,
 496                                     context->db_tab, to_msrq(srq)->db_index);
 497         }
 498
 499         mthca_free_srq(to_mdev(srq->device), to_msrq(srq));
 500         kfree(srq);
 501
 502         return 0;
 503 }
 504
 505 static struct ib_qp *mthca_create_qp(struct ib_pd *pd,
 506                                      struct ib_qp_init_attr *init_attr,
 507                                      struct ib_udata *udata)
 508 {
 509         struct mthca_create_qp ucmd;
 510         struct mthca_qp *qp;
 511         int err;
 512
 513         switch (init_attr->qp_type) {
 514         case IB_QPT_RC:
 515         case IB_QPT_UC:
 516         case IB_QPT_UD:
 517         {
 518                 struct mthca_ucontext *context;
 519
 520                 qp = kmalloc(sizeof *qp, GFP_KERNEL);
 521                 if (!qp)
 522                         return ERR_PTR(-ENOMEM);
 523
 524                 if (pd->uobject) {
 525                         context = to_mucontext(pd->uobject->context);
 526
 527                         if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd))
 528                                 return ERR_PTR(-EFAULT);
 529
 530                         err = mthca_map_user_db(to_mdev(pd->device), &context->uar,
 531                                                 context->db_tab,
 532                                                 ucmd.sq_db_index, ucmd.sq_db_page);
 533                         if (err) {
 534                                 kfree(qp);
 535                                 return ERR_PTR(err);
 536                         }
 537
 538                         err = mthca_map_user_db(to_mdev(pd->device), &context->uar,
 539                                                 context->db_tab,
 540                                                 ucmd.rq_db_index, ucmd.rq_db_page);
 541                         if (err) {
 542                                 mthca_unmap_user_db(to_mdev(pd->device),
 543                                                     &context->uar,
 544                                                     context->db_tab,
 545                                                     ucmd.sq_db_index);
 546                                 kfree(qp);
 547                                 return ERR_PTR(err);
 548                         }
 549
 550                         qp->mr.ibmr.lkey = ucmd.lkey;
 551                         qp->sq.db_index  = ucmd.sq_db_index;
 552                         qp->rq.db_index  = ucmd.rq_db_index;
 553                 }
 554
 555                 err = mthca_alloc_qp(to_mdev(pd->device), to_mpd(pd),
 556                                      to_mcq(init_attr->send_cq),
 557                                      to_mcq(init_attr->recv_cq),
 558                                      init_attr->qp_type, init_attr->sq_sig_type,
 559                                      &init_attr->cap, qp);
 560
 561                 if (err && pd->uobject) {
 562                         context = to_mucontext(pd->uobject->context);
 563
 564                         mthca_unmap_user_db(to_mdev(pd->device),
 565                                             &context->uar,
 566                                             context->db_tab,
 567                                             ucmd.sq_db_index);
 568                         mthca_unmap_user_db(to_mdev(pd->device),
 569                                             &context->uar,
 570                                             context->db_tab,
 571                                             ucmd.rq_db_index);
 572                 }
 573
 574                 qp->ibqp.qp_num = qp->qpn;
 575                 break;
 576         }
 577         case IB_QPT_SMI:
 578         case IB_QPT_GSI:
 579         {
 580                 /* Don't allow userspace to create special QPs */
 581                 if (pd->uobject)
 582                         return ERR_PTR(-EINVAL);
 583
 584                 qp = kmalloc(sizeof (struct mthca_sqp), GFP_KERNEL);
 585                 if (!qp)
 586                         return ERR_PTR(-ENOMEM);
 587
 588                 qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 : 1;
 589
 590                 err = mthca_alloc_sqp(to_mdev(pd->device), to_mpd(pd),
 591                                       to_mcq(init_attr->send_cq),
 592                                       to_mcq(init_attr->recv_cq),
 593                                       init_attr->sq_sig_type, &init_attr->cap,
 594                                       qp->ibqp.qp_num, init_attr->port_num,
 595                                       to_msqp(qp));
 596                 break;
 597         }
 598         default:
 599                 /* Don't support raw QPs */
 600                 return ERR_PTR(-ENOSYS);
 601         }
 602
 603         if (err) {
 604                 kfree(qp);
 605                 return ERR_PTR(err);
 606         }
 607
 608         init_attr->cap.max_inline_data = 0;
 609         init_attr->cap.max_send_wr     = qp->sq.max;
 610         init_attr->cap.max_recv_wr     = qp->rq.max;
 611         init_attr->cap.max_send_sge    = qp->sq.max_gs;
 612         init_attr->cap.max_recv_sge    = qp->rq.max_gs;
 613
 614         return &qp->ibqp;
 615 }
 616
 617 static int mthca_destroy_qp(struct ib_qp *qp)
 618 {
 619         if (qp->uobject) {
 620                 mthca_unmap_user_db(to_mdev(qp->device),
 621                                     &to_mucontext(qp->uobject->context)->uar,
 622                                     to_mucontext(qp->uobject->context)->db_tab,
 623                                     to_mqp(qp)->sq.db_index);
 624                 mthca_unmap_user_db(to_mdev(qp->device),
 625                                     &to_mucontext(qp->uobject->context)->uar,
 626                                     to_mucontext(qp->uobject->context)->db_tab,
 627                                     to_mqp(qp)->rq.db_index);
 628         }
 629         mthca_free_qp(to_mdev(qp->device), to_mqp(qp));
 630         kfree(qp);
 631         return 0;
 632 }
 633
 634 static struct ib_cq *mthca_create_cq(struct ib_device *ibdev, int entries,
 635                                      struct ib_ucontext *context,
 636                                      struct ib_udata *udata)
 637 {
 638         struct mthca_create_cq ucmd;
 639         struct mthca_cq *cq;
 640         int nent;
 641         int err;
 642
 643         if (context) {
 644                 if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd))
 645                         return ERR_PTR(-EFAULT);
 646
 647                 err = mthca_map_user_db(to_mdev(ibdev), &to_mucontext(context)->uar,
 648                                         to_mucontext(context)->db_tab,
 649                                         ucmd.set_db_index, ucmd.set_db_page);
 650                 if (err)
 651                         return ERR_PTR(err);
 652
 653                 err = mthca_map_user_db(to_mdev(ibdev), &to_mucontext(context)->uar,
 654                                         to_mucontext(context)->db_tab,
 655                                         ucmd.arm_db_index, ucmd.arm_db_page);
 656                 if (err)
 657                         goto err_unmap_set;
 658         }
 659
 660         cq = kmalloc(sizeof *cq, GFP_KERNEL);
 661         if (!cq) {
 662                 err = -ENOMEM;
 663                 goto err_unmap_arm;
 664         }
 665
 666         if (context) {
 667                 cq->mr.ibmr.lkey    = ucmd.lkey;
 668                 cq->set_ci_db_index = ucmd.set_db_index;
 669                 cq->arm_db_index    = ucmd.arm_db_index;
 670         }
 671
 672         for (nent = 1; nent <= entries; nent <<= 1)
 673                 ; /* nothing */
 674
 675         err = mthca_init_cq(to_mdev(ibdev), nent,
 676                             context ? to_mucontext(context) : NULL,
 677                             context ? ucmd.pdn : to_mdev(ibdev)->driver_pd.pd_num,
 678                             cq);
 679         if (err)
 680                 goto err_free;
 681
 682         if (context && ib_copy_to_udata(udata, &cq->cqn, sizeof (__u32))) {
 683                 mthca_free_cq(to_mdev(ibdev), cq);
 684                 goto err_free;
 685         }
 686
 687         return &cq->ibcq;
 688
 689 err_free:
 690         kfree(cq);
 691
 692 err_unmap_arm:
 693         if (context)
 694                 mthca_unmap_user_db(to_mdev(ibdev), &to_mucontext(context)->uar,
 695                                     to_mucontext(context)->db_tab, ucmd.arm_db_index);
 696
 697 err_unmap_set:
 698         if (context)
 699                 mthca_unmap_user_db(to_mdev(ibdev), &to_mucontext(context)->uar,
 700                                     to_mucontext(context)->db_tab, ucmd.set_db_index);
 701
 702         return ERR_PTR(err);
 703 }
 704
 705 static int mthca_destroy_cq(struct ib_cq *cq)
 706 {
 707         if (cq->uobject) {
 708                 mthca_unmap_user_db(to_mdev(cq->device),
 709                                     &to_mucontext(cq->uobject->context)->uar,
 710                                     to_mucontext(cq->uobject->context)->db_tab,
 711                                     to_mcq(cq)->arm_db_index);
 712                 mthca_unmap_user_db(to_mdev(cq->device),
 713                                     &to_mucontext(cq->uobject->context)->uar,
 714                                     to_mucontext(cq->uobject->context)->db_tab,
 715                                     to_mcq(cq)->set_ci_db_index);
 716         }
 717         mthca_free_cq(to_mdev(cq->device), to_mcq(cq));
 718         kfree(cq);
 719
 720         return 0;
 721 }
 722
 723 static inline u32 convert_access(int acc)
 724 {
 725         return (acc & IB_ACCESS_REMOTE_ATOMIC ? MTHCA_MPT_FLAG_ATOMIC       : 0) |
 726                (acc & IB_ACCESS_REMOTE_WRITE  ? MTHCA_MPT_FLAG_REMOTE_WRITE : 0) |
 727                (acc & IB_ACCESS_REMOTE_READ   ? MTHCA_MPT_FLAG_REMOTE_READ  : 0) |
 728                (acc & IB_ACCESS_LOCAL_WRITE   ? MTHCA_MPT_FLAG_LOCAL_WRITE  : 0) |
 729                MTHCA_MPT_FLAG_LOCAL_READ;
 730 }
 731
 732 static struct ib_mr *mthca_get_dma_mr(struct ib_pd *pd, int acc)
 733 {
 734         struct mthca_mr *mr;
 735         int err;
 736
 737         mr = kmalloc(sizeof *mr, GFP_KERNEL);
 738         if (!mr)
 739                 return ERR_PTR(-ENOMEM);
 740
 741         err = mthca_mr_alloc_notrans(to_mdev(pd->device),
 742                                      to_mpd(pd)->pd_num,
 743                                      convert_access(acc), mr);
 744
 745         if (err) {
 746                 kfree(mr);
 747                 return ERR_PTR(err);
 748         }
 749
 750         return &mr->ibmr;
 751 }
 752
 753 static struct ib_mr *mthca_reg_phys_mr(struct ib_pd       *pd,
 754                                        struct ib_phys_buf *buffer_list,
 755                                        int                 num_phys_buf,
 756                                        int                 acc,
 757                                        u64                *iova_start)
 758 {
 759         struct mthca_mr *mr;
 760         u64 *page_list;
 761         u64 total_size;
 762         u64 mask;
 763         int shift;
 764         int npages;
 765         int err;
 766         int i, j, n;
 767
 768         /* First check that we have enough alignment */
 769         if ((*iova_start & ~PAGE_MASK) != (buffer_list[0].addr & ~PAGE_MASK))
 770                 return ERR_PTR(-EINVAL);
 771
 772         if (num_phys_buf > 1 &&
 773             ((buffer_list[0].addr + buffer_list[0].size) & ~PAGE_MASK))
 774                 return ERR_PTR(-EINVAL);
 775
 776         mask = 0;
 777         total_size = 0;
 778         for (i = 0; i < num_phys_buf; ++i) {
 779                 if (i != 0 && buffer_list[i].addr & ~PAGE_MASK)
 780                         return ERR_PTR(-EINVAL);
 781                 if (i != 0 && i != num_phys_buf - 1 &&
 782                     (buffer_list[i].size & ~PAGE_MASK))
 783                         return ERR_PTR(-EINVAL);
 784
 785                 total_size += buffer_list[i].size;
 786                 if (i > 0)
 787                         mask |= buffer_list[i].addr;
 788         }
 789
 790         /* Find largest page shift we can use to cover buffers */
 791         for (shift = PAGE_SHIFT; shift < 31; ++shift)
 792                 if (num_phys_buf > 1) {
 793                         if ((1ULL << shift) & mask)
 794                                 break;
 795                 } else {
 796                         if (1ULL << shift >=
 797                             buffer_list[0].size +
 798                             (buffer_list[0].addr & ((1ULL << shift) - 1)))
 799                                 break;
 800                 }
 801
 802         buffer_list[0].size += buffer_list[0].addr & ((1ULL << shift) - 1);
 803         buffer_list[0].addr &= ~0ull << shift;
 804
 805         mr = kmalloc(sizeof *mr, GFP_KERNEL);
 806         if (!mr)
 807                 return ERR_PTR(-ENOMEM);
 808
 809         npages = 0;
 810         for (i = 0; i < num_phys_buf; ++i)
 811                 npages += (buffer_list[i].size + (1ULL << shift) - 1) >> shift;
 812
 813         if (!npages)
 814                 return &mr->ibmr;
 815
 816         page_list = kmalloc(npages * sizeof *page_list, GFP_KERNEL);
 817         if (!page_list) {
 818                 kfree(mr);
 819                 return ERR_PTR(-ENOMEM);
 820         }
 821
 822         n = 0;
 823         for (i = 0; i < num_phys_buf; ++i)
 824                 for (j = 0;
 825                      j < (buffer_list[i].size + (1ULL << shift) - 1) >> shift;
 826                      ++j)
 827                         page_list[n++] = buffer_list[i].addr + ((u64) j << shift);
 828
 829         mthca_dbg(to_mdev(pd->device), "Registering memory at %llx (iova %llx) "
 830                   "in PD %x; shift %d, npages %d.\n",
 831                   (unsigned long long) buffer_list[0].addr,
 832                   (unsigned long long) *iova_start,
 833                   to_mpd(pd)->pd_num,
 834                   shift, npages);
 835
 836         err = mthca_mr_alloc_phys(to_mdev(pd->device),
 837                                   to_mpd(pd)->pd_num,
 838                                   page_list, shift, npages,
 839                                   *iova_start, total_size,
 840                                   convert_access(acc), mr);
 841
 842         if (err) {
 843                 kfree(page_list);
 844                 kfree(mr);
 845                 return ERR_PTR(err);
 846         }
 847
 848         kfree(page_list);
 849         return &mr->ibmr;
 850 }
 851
 852 static struct ib_mr *mthca_reg_user_mr(struct ib_pd *pd, struct ib_umem *region,
 853                                        int acc, struct ib_udata *udata)
 854 {
 855         struct mthca_dev *dev = to_mdev(pd->device);
 856         struct ib_umem_chunk *chunk;
 857         struct mthca_mr *mr;
 858         u64 *pages;
 859         int shift, n, len;
 860         int i, j, k;
 861         int err = 0;
 862
 863         shift = ffs(region->page_size) - 1;
 864
 865         mr = kmalloc(sizeof *mr, GFP_KERNEL);
 866         if (!mr)
 867                 return ERR_PTR(-ENOMEM);
 868
 869         n = 0;
 870         list_for_each_entry(chunk, &region->chunk_list, list)
 871                 n += chunk->nents;
 872
 873         mr->mtt = mthca_alloc_mtt(dev, n);
 874         if (IS_ERR(mr->mtt)) {
 875                 err = PTR_ERR(mr->mtt);
 876                 goto err;
 877         }
 878
 879         pages = (u64 *) __get_free_page(GFP_KERNEL);
 880         if (!pages) {
 881                 err = -ENOMEM;
 882                 goto err_mtt;
 883         }
 884
 885         i = n = 0;
 886
 887         list_for_each_entry(chunk, &region->chunk_list, list)
 888                 for (j = 0; j < chunk->nmap; ++j) {
 889                         len = sg_dma_len(&chunk->page_list[j]) >> shift;
 890                         for (k = 0; k < len; ++k) {
 891                                 pages[i++] = sg_dma_address(&chunk->page_list[j]) +
 892                                         region->page_size * k;
 893                                 /*
 894                                  * Be friendly to WRITE_MTT command
 895                                  * and leave two empty slots for the
 896                                  * index and reserved fields of the
 897                                  * mailbox.
 898                                  */
 899                                 if (i == PAGE_SIZE / sizeof (u64) - 2) {
 900                                         err = mthca_write_mtt(dev, mr->mtt,
 901                                                               n, pages, i);
 902                                         if (err)
 903                                                 goto mtt_done;
 904                                         n += i;
 905                                         i = 0;
 906                                 }
 907                         }
 908                 }
 909
 910         if (i)
 911                 err = mthca_write_mtt(dev, mr->mtt, n, pages, i);
 912 mtt_done:
 913         free_page((unsigned long) pages);
 914         if (err)
 915                 goto err_mtt;
 916
 917         err = mthca_mr_alloc(dev, to_mpd(pd)->pd_num, shift, region->virt_base,
 918                              region->length, convert_access(acc), mr);
 919
 920         if (err)
 921                 goto err_mtt;
 922
 923         return &mr->ibmr;
 924
 925 err_mtt:
 926         mthca_free_mtt(dev, mr->mtt);
 927
 928 err:
 929         kfree(mr);
 930         return ERR_PTR(err);
 931 }
 932
 933 static int mthca_dereg_mr(struct ib_mr *mr)
 934 {
 935         struct mthca_mr *mmr = to_mmr(mr);
 936         mthca_free_mr(to_mdev(mr->device), mmr);
 937         kfree(mmr);
 938         return 0;
 939 }
 940
 941 static struct ib_fmr *mthca_alloc_fmr(struct ib_pd *pd, int mr_access_flags,
 942                                       struct ib_fmr_attr *fmr_attr)
 943 {
 944         struct mthca_fmr *fmr;
 945         int err;
 946
 947         fmr = kmalloc(sizeof *fmr, GFP_KERNEL);
 948         if (!fmr)
 949                 return ERR_PTR(-ENOMEM);
 950
 951         memcpy(&fmr->attr, fmr_attr, sizeof *fmr_attr);
 952         err = mthca_fmr_alloc(to_mdev(pd->device), to_mpd(pd)->pd_num,
 953                              convert_access(mr_access_flags), fmr);
 954
 955         if (err) {
 956                 kfree(fmr);
 957                 return ERR_PTR(err);
 958         }
 959
 960         return &fmr->ibmr;
 961 }
 962
 963 static int mthca_dealloc_fmr(struct ib_fmr *fmr)
 964 {
 965         struct mthca_fmr *mfmr = to_mfmr(fmr);
 966         int err;
 967
 968         err = mthca_free_fmr(to_mdev(fmr->device), mfmr);
 969         if (err)
 970                 return err;
 971
 972         kfree(mfmr);
 973         return 0;
 974 }
 975
 976 static int mthca_unmap_fmr(struct list_head *fmr_list)
 977 {
 978         struct ib_fmr *fmr;
 979         int err;
 980         u8 status;
 981         struct mthca_dev *mdev = NULL;
 982
 983         list_for_each_entry(fmr, fmr_list, list) {
 984                 if (mdev && to_mdev(fmr->device) != mdev)
 985                         return -EINVAL;
 986                 mdev = to_mdev(fmr->device);
 987         }
 988
 989         if (!mdev)
 990                 return 0;
 991
 992         if (mthca_is_memfree(mdev)) {
 993                 list_for_each_entry(fmr, fmr_list, list)
 994                         mthca_arbel_fmr_unmap(mdev, to_mfmr(fmr));
 995
 996                 wmb();
 997         } else
 998                 list_for_each_entry(fmr, fmr_list, list)
 999                         mthca_tavor_fmr_unmap(mdev, to_mfmr(fmr));
1000
1001         err = mthca_SYNC_TPT(mdev, &status);
1002         if (err)
1003                 return err;
1004         if (status)
1005                 return -EINVAL;
1006         return 0;
1007 }
1008
1009 static ssize_t show_rev(struct class_device *cdev, char *buf)
1010 {
1011         struct mthca_dev *dev = container_of(cdev, struct mthca_dev, ib_dev.class_dev);
1012         return sprintf(buf, "%x\n", dev->rev_id);
1013 }
1014
1015 static ssize_t show_fw_ver(struct class_device *cdev, char *buf)
1016 {
1017         struct mthca_dev *dev = container_of(cdev, struct mthca_dev, ib_dev.class_dev);
1018         return sprintf(buf, "%x.%x.%x\n", (int) (dev->fw_ver >> 32),
1019                        (int) (dev->fw_ver >> 16) & 0xffff,
1020                        (int) dev->fw_ver & 0xffff);
1021 }
1022
1023 static ssize_t show_hca(struct class_device *cdev, char *buf)
1024 {
1025         struct mthca_dev *dev = container_of(cdev, struct mthca_dev, ib_dev.class_dev);
1026         switch (dev->pdev->device) {
1027         case PCI_DEVICE_ID_MELLANOX_TAVOR:
1028                 return sprintf(buf, "MT23108\n");
1029         case PCI_DEVICE_ID_MELLANOX_ARBEL_COMPAT:
1030                 return sprintf(buf, "MT25208 (MT23108 compat mode)\n");
1031         case PCI_DEVICE_ID_MELLANOX_ARBEL:
1032                 return sprintf(buf, "MT25208\n");
1033         case PCI_DEVICE_ID_MELLANOX_SINAI:
1034         case PCI_DEVICE_ID_MELLANOX_SINAI_OLD:
1035                 return sprintf(buf, "MT25204\n");
1036         default:
1037                 return sprintf(buf, "unknown\n");
1038         }
1039 }
1040
1041 static ssize_t show_board(struct class_device *cdev, char *buf)
1042 {
1043         struct mthca_dev *dev = container_of(cdev, struct mthca_dev, ib_dev.class_dev);
1044         return sprintf(buf, "%.*s\n", MTHCA_BOARD_ID_LEN, dev->board_id);
1045 }
1046
1047 static CLASS_DEVICE_ATTR(hw_rev,   S_IRUGO, show_rev,    NULL);
1048 static CLASS_DEVICE_ATTR(fw_ver,   S_IRUGO, show_fw_ver, NULL);
1049 static CLASS_DEVICE_ATTR(hca_type, S_IRUGO, show_hca,    NULL);
1050 static CLASS_DEVICE_ATTR(board_id, S_IRUGO, show_board,  NULL);
1051
1052 static struct class_device_attribute *mthca_class_attributes[] = {
1053         &class_device_attr_hw_rev,
1054         &class_device_attr_fw_ver,
1055         &class_device_attr_hca_type,
1056         &class_device_attr_board_id
1057 };
1058
1059 int mthca_register_device(struct mthca_dev *dev)
1060 {
1061         int ret;
1062         int i;
1063
1064         strlcpy(dev->ib_dev.name, "mthca%d", IB_DEVICE_NAME_MAX);
1065         dev->ib_dev.owner                = THIS_MODULE;
1066
1067         dev->ib_dev.uverbs_abi_ver       = MTHCA_UVERBS_ABI_VERSION;
1068         dev->ib_dev.node_type            = IB_NODE_CA;
1069         dev->ib_dev.phys_port_cnt        = dev->limits.num_ports;
1070         dev->ib_dev.dma_device           = &dev->pdev->dev;
1071         dev->ib_dev.class_dev.dev        = &dev->pdev->dev;
1072         dev->ib_dev.query_device         = mthca_query_device;
1073         dev->ib_dev.query_port           = mthca_query_port;
1074         dev->ib_dev.modify_port          = mthca_modify_port;
1075         dev->ib_dev.query_pkey           = mthca_query_pkey;
1076         dev->ib_dev.query_gid            = mthca_query_gid;
1077         dev->ib_dev.alloc_ucontext       = mthca_alloc_ucontext;
1078         dev->ib_dev.dealloc_ucontext     = mthca_dealloc_ucontext;
1079         dev->ib_dev.mmap                 = mthca_mmap_uar;
1080         dev->ib_dev.alloc_pd             = mthca_alloc_pd;
1081         dev->ib_dev.dealloc_pd           = mthca_dealloc_pd;
1082         dev->ib_dev.create_ah            = mthca_ah_create;
1083         dev->ib_dev.destroy_ah           = mthca_ah_destroy;
1084
1085         if (dev->mthca_flags & MTHCA_FLAG_SRQ) {
1086                 dev->ib_dev.create_srq           = mthca_create_srq;
1087                 dev->ib_dev.destroy_srq          = mthca_destroy_srq;
1088
1089                 if (mthca_is_memfree(dev))
1090                         dev->ib_dev.post_srq_recv = mthca_arbel_post_srq_recv;
1091                 else
1092                         dev->ib_dev.post_srq_recv = mthca_tavor_post_srq_recv;
1093         }
1094
1095         dev->ib_dev.create_qp            = mthca_create_qp;
1096         dev->ib_dev.modify_qp            = mthca_modify_qp;
1097         dev->ib_dev.destroy_qp           = mthca_destroy_qp;
1098         dev->ib_dev.create_cq            = mthca_create_cq;
1099         dev->ib_dev.destroy_cq           = mthca_destroy_cq;
1100         dev->ib_dev.poll_cq              = mthca_poll_cq;
1101         dev->ib_dev.get_dma_mr           = mthca_get_dma_mr;
1102         dev->ib_dev.reg_phys_mr          = mthca_reg_phys_mr;
1103         dev->ib_dev.reg_user_mr          = mthca_reg_user_mr;
1104         dev->ib_dev.dereg_mr             = mthca_dereg_mr;
1105
1106         if (dev->mthca_flags & MTHCA_FLAG_FMR) {
1107                 dev->ib_dev.alloc_fmr            = mthca_alloc_fmr;
1108                 dev->ib_dev.unmap_fmr            = mthca_unmap_fmr;
1109                 dev->ib_dev.dealloc_fmr          = mthca_dealloc_fmr;
1110                 if (mthca_is_memfree(dev))
1111                         dev->ib_dev.map_phys_fmr = mthca_arbel_map_phys_fmr;
1112                 else
1113                         dev->ib_dev.map_phys_fmr = mthca_tavor_map_phys_fmr;
1114         }
1115
1116         dev->ib_dev.attach_mcast         = mthca_multicast_attach;
1117         dev->ib_dev.detach_mcast         = mthca_multicast_detach;
1118         dev->ib_dev.process_mad          = mthca_process_mad;
1119
1120         if (mthca_is_memfree(dev)) {
1121                 dev->ib_dev.req_notify_cq = mthca_arbel_arm_cq;
1122                 dev->ib_dev.post_send     = mthca_arbel_post_send;
1123                 dev->ib_dev.post_recv     = mthca_arbel_post_receive;
1124         } else {
1125                 dev->ib_dev.req_notify_cq = mthca_tavor_arm_cq;
1126                 dev->ib_dev.post_send     = mthca_tavor_post_send;
1127                 dev->ib_dev.post_recv     = mthca_tavor_post_receive;
1128         }
1129
1130         init_MUTEX(&dev->cap_mask_mutex);
1131
1132         ret = ib_register_device(&dev->ib_dev);
1133         if (ret)
1134                 return ret;
1135
1136         for (i = 0; i < ARRAY_SIZE(mthca_class_attributes); ++i) {
1137                 ret = class_device_create_file(&dev->ib_dev.class_dev,
1138                                                mthca_class_attributes[i]);
1139                 if (ret) {
1140                         ib_unregister_device(&dev->ib_dev);
1141                         return ret;
1142                 }
1143         }
1144
1145         return 0;
1146 }
1147
1148 void mthca_unregister_device(struct mthca_dev *dev)
1149 {
1150         ib_unregister_device(&dev->ib_dev);
1151 }