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