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