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           = mdev->limits.max_wqes;
  94         props->max_sge             = mdev->limits.max_sg;
  95         props->max_cq              = mdev->limits.num_cqs - mdev->limits.reserved_cqs;
  96         props->max_cqe             = mdev->limits.max_cqes;
  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 = mdev->limits.max_qp_init_rdma;
 101         props->max_res_rd_atom     = props->max_qp_rd_atom * props->max_qp;
 102         props->max_srq             = mdev->limits.num_srqs - mdev->limits.reserved_srqs;
 103         props->max_srq_wr          = mdev->limits.max_srq_wqes;
 104         props->max_srq_sge         = mdev->limits.max_sg;
 105         props->local_ca_ack_delay  = mdev->limits.local_ca_ack_delay;
 106         props->atomic_cap          = mdev->limits.flags & DEV_LIM_FLAG_ATOMIC ?
 107                                         IB_ATOMIC_HCA : IB_ATOMIC_NONE;
 108         props->max_pkeys           = mdev->limits.pkey_table_len;
 109         props->max_mcast_grp       = mdev->limits.num_mgms + mdev->limits.num_amgms;
 110         props->max_mcast_qp_attach = MTHCA_QP_PER_MGM;
 111         props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
 112                                            props->max_mcast_grp;
 113
 114         err = 0;
 115  out:
 116         kfree(in_mad);
 117         kfree(out_mad);
 118         return err;
 119 }
 120
 121 static int mthca_query_port(struct ib_device *ibdev,
 122                             u8 port, struct ib_port_attr *props)
 123 {
 124         struct ib_smp *in_mad  = NULL;
 125         struct ib_smp *out_mad = NULL;
 126         int err = -ENOMEM;
 127         u8 status;
 128
 129         in_mad  = kmalloc(sizeof *in_mad, GFP_KERNEL);
 130         out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
 131         if (!in_mad || !out_mad)
 132                 goto out;
 133
 134         memset(props, 0, sizeof *props);
 135
 136         memset(in_mad, 0, sizeof *in_mad);
 137         in_mad->base_version       = 1;
 138         in_mad->mgmt_class         = IB_MGMT_CLASS_SUBN_LID_ROUTED;
 139         in_mad->class_version      = 1;
 140         in_mad->method             = IB_MGMT_METHOD_GET;
 141         in_mad->attr_id            = IB_SMP_ATTR_PORT_INFO;
 142         in_mad->attr_mod           = cpu_to_be32(port);
 143
 144         err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
 145                             port, NULL, NULL, in_mad, out_mad,
 146                             &status);
 147         if (err)
 148                 goto out;
 149         if (status) {
 150                 err = -EINVAL;
 151                 goto out;
 152         }
 153
 154         props->lid               = be16_to_cpup((__be16 *) (out_mad->data + 16));
 155         props->lmc               = out_mad->data[34] & 0x7;
 156         props->sm_lid            = be16_to_cpup((__be16 *) (out_mad->data + 18));
 157         props->sm_sl             = out_mad->data[36] & 0xf;
 158         props->state             = out_mad->data[32] & 0xf;
 159         props->phys_state        = out_mad->data[33] >> 4;
 160         props->port_cap_flags    = be32_to_cpup((__be32 *) (out_mad->data + 20));
 161         props->gid_tbl_len       = to_mdev(ibdev)->limits.gid_table_len;
 162         props->max_msg_sz        = 0x80000000;
 163         props->pkey_tbl_len      = to_mdev(ibdev)->limits.pkey_table_len;
 164         props->bad_pkey_cntr     = be16_to_cpup((__be16 *) (out_mad->data + 46));
 165         props->qkey_viol_cntr    = be16_to_cpup((__be16 *) (out_mad->data + 48));
 166         props->active_width      = out_mad->data[31] & 0xf;
 167         props->active_speed      = out_mad->data[35] >> 4;
 168         props->max_mtu           = out_mad->data[41] & 0xf;
 169         props->active_mtu        = out_mad->data[36] >> 4;
 170         props->subnet_timeout    = out_mad->data[51] & 0x1f;
 171
 172  out:
 173         kfree(in_mad);
 174         kfree(out_mad);
 175         return err;
 176 }
 177
 178 static int mthca_modify_port(struct ib_device *ibdev,
 179                              u8 port, int port_modify_mask,
 180                              struct ib_port_modify *props)
 181 {
 182         struct mthca_set_ib_param set_ib;
 183         struct ib_port_attr attr;
 184         int err;
 185         u8 status;
 186
 187         if (down_interruptible(&to_mdev(ibdev)->cap_mask_mutex))
 188                 return -ERESTARTSYS;
 189
 190         err = mthca_query_port(ibdev, port, &attr);
 191         if (err)
 192                 goto out;
 193
 194         set_ib.set_si_guid     = 0;
 195         set_ib.reset_qkey_viol = !!(port_modify_mask & IB_PORT_RESET_QKEY_CNTR);
 196
 197         set_ib.cap_mask = (attr.port_cap_flags | props->set_port_cap_mask) &
 198                 ~props->clr_port_cap_mask;
 199
 200         err = mthca_SET_IB(to_mdev(ibdev), &set_ib, port, &status);
 201         if (err)
 202                 goto out;
 203         if (status) {
 204                 err = -EINVAL;
 205                 goto out;
 206         }
 207
 208 out:
 209         up(&to_mdev(ibdev)->cap_mask_mutex);
 210         return err;
 211 }
 212
 213 static int mthca_query_pkey(struct ib_device *ibdev,
 214                             u8 port, u16 index, u16 *pkey)
 215 {
 216         struct ib_smp *in_mad  = NULL;
 217         struct ib_smp *out_mad = NULL;
 218         int err = -ENOMEM;
 219         u8 status;
 220
 221         in_mad  = kmalloc(sizeof *in_mad, GFP_KERNEL);
 222         out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
 223         if (!in_mad || !out_mad)
 224                 goto out;
 225
 226         memset(in_mad, 0, sizeof *in_mad);
 227         in_mad->base_version       = 1;
 228         in_mad->mgmt_class         = IB_MGMT_CLASS_SUBN_LID_ROUTED;
 229         in_mad->class_version      = 1;
 230         in_mad->method             = IB_MGMT_METHOD_GET;
 231         in_mad->attr_id            = IB_SMP_ATTR_PKEY_TABLE;
 232         in_mad->attr_mod           = cpu_to_be32(index / 32);
 233
 234         err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
 235                             port, NULL, NULL, in_mad, out_mad,
 236                             &status);
 237         if (err)
 238                 goto out;
 239         if (status) {
 240                 err = -EINVAL;
 241                 goto out;
 242         }
 243
 244         *pkey = be16_to_cpu(((__be16 *) out_mad->data)[index % 32]);
 245
 246  out:
 247         kfree(in_mad);
 248         kfree(out_mad);
 249         return err;
 250 }
 251
 252 static int mthca_query_gid(struct ib_device *ibdev, u8 port,
 253                            int index, union ib_gid *gid)
 254 {
 255         struct ib_smp *in_mad  = NULL;
 256         struct ib_smp *out_mad = NULL;
 257         int err = -ENOMEM;
 258         u8 status;
 259
 260         in_mad  = kmalloc(sizeof *in_mad, GFP_KERNEL);
 261         out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
 262         if (!in_mad || !out_mad)
 263                 goto out;
 264
 265         memset(in_mad, 0, sizeof *in_mad);
 266         in_mad->base_version       = 1;
 267         in_mad->mgmt_class         = IB_MGMT_CLASS_SUBN_LID_ROUTED;
 268         in_mad->class_version      = 1;
 269         in_mad->method             = IB_MGMT_METHOD_GET;
 270         in_mad->attr_id            = IB_SMP_ATTR_PORT_INFO;
 271         in_mad->attr_mod           = cpu_to_be32(port);
 272
 273         err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
 274                             port, NULL, NULL, in_mad, out_mad,
 275                             &status);
 276         if (err)
 277                 goto out;
 278         if (status) {
 279                 err = -EINVAL;
 280                 goto out;
 281         }
 282
 283         memcpy(gid->raw, out_mad->data + 8, 8);
 284
 285         memset(in_mad, 0, sizeof *in_mad);
 286         in_mad->base_version       = 1;
 287         in_mad->mgmt_class         = IB_MGMT_CLASS_SUBN_LID_ROUTED;
 288         in_mad->class_version      = 1;
 289         in_mad->method             = IB_MGMT_METHOD_GET;
 290         in_mad->attr_id            = IB_SMP_ATTR_GUID_INFO;
 291         in_mad->attr_mod           = cpu_to_be32(index / 8);
 292
 293         err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
 294                             port, NULL, NULL, in_mad, out_mad,
 295                             &status);
 296         if (err)
 297                 goto out;
 298         if (status) {
 299                 err = -EINVAL;
 300                 goto out;
 301         }
 302
 303         memcpy(gid->raw + 8, out_mad->data + (index % 8) * 16, 8);
 304
 305  out:
 306         kfree(in_mad);
 307         kfree(out_mad);
 308         return err;
 309 }
 310
 311 static struct ib_ucontext *mthca_alloc_ucontext(struct ib_device *ibdev,
 312                                                 struct ib_udata *udata)
 313 {
 314         struct mthca_alloc_ucontext_resp uresp;
 315         struct mthca_ucontext           *context;
 316         int                              err;
 317
 318         memset(&uresp, 0, sizeof uresp);
 319
 320         uresp.qp_tab_size = to_mdev(ibdev)->limits.num_qps;
 321         if (mthca_is_memfree(to_mdev(ibdev)))
 322                 uresp.uarc_size = to_mdev(ibdev)->uar_table.uarc_size;
 323         else
 324                 uresp.uarc_size = 0;
 325
 326         context = kmalloc(sizeof *context, GFP_KERNEL);
 327         if (!context)
 328                 return ERR_PTR(-ENOMEM);
 329
 330         err = mthca_uar_alloc(to_mdev(ibdev), &context->uar);
 331         if (err) {
 332                 kfree(context);
 333                 return ERR_PTR(err);
 334         }
 335
 336         context->db_tab = mthca_init_user_db_tab(to_mdev(ibdev));
 337         if (IS_ERR(context->db_tab)) {
 338                 err = PTR_ERR(context->db_tab);
 339                 mthca_uar_free(to_mdev(ibdev), &context->uar);
 340                 kfree(context);
 341                 return ERR_PTR(err);
 342         }
 343
 344         if (ib_copy_to_udata(udata, &uresp, sizeof uresp)) {
 345                 mthca_cleanup_user_db_tab(to_mdev(ibdev), &context->uar, context->db_tab);
 346                 mthca_uar_free(to_mdev(ibdev), &context->uar);
 347                 kfree(context);
 348                 return ERR_PTR(-EFAULT);
 349         }
 350
 351         return &context->ibucontext;
 352 }
 353
 354 static int mthca_dealloc_ucontext(struct ib_ucontext *context)
 355 {
 356         mthca_cleanup_user_db_tab(to_mdev(context->device), &to_mucontext(context)->uar,
 357                                   to_mucontext(context)->db_tab);
 358         mthca_uar_free(to_mdev(context->device), &to_mucontext(context)->uar);
 359         kfree(to_mucontext(context));
 360
 361         return 0;
 362 }
 363
 364 static int mthca_mmap_uar(struct ib_ucontext *context,
 365                           struct vm_area_struct *vma)
 366 {
 367         if (vma->vm_end - vma->vm_start != PAGE_SIZE)
 368                 return -EINVAL;
 369
 370         vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
 371
 372         if (io_remap_pfn_range(vma, vma->vm_start,
 373                                to_mucontext(context)->uar.pfn,
 374                                PAGE_SIZE, vma->vm_page_prot))
 375                 return -EAGAIN;
 376
 377         return 0;
 378 }
 379
 380 static struct ib_pd *mthca_alloc_pd(struct ib_device *ibdev,
 381                                     struct ib_ucontext *context,
 382                                     struct ib_udata *udata)
 383 {
 384         struct mthca_pd *pd;
 385         int err;
 386
 387         pd = kmalloc(sizeof *pd, GFP_KERNEL);
 388         if (!pd)
 389                 return ERR_PTR(-ENOMEM);
 390
 391         err = mthca_pd_alloc(to_mdev(ibdev), !context, pd);
 392         if (err) {
 393                 kfree(pd);
 394                 return ERR_PTR(err);
 395         }
 396
 397         if (context) {
 398                 if (ib_copy_to_udata(udata, &pd->pd_num, sizeof (__u32))) {
 399                         mthca_pd_free(to_mdev(ibdev), pd);
 400                         kfree(pd);
 401                         return ERR_PTR(-EFAULT);
 402                 }
 403         }
 404
 405         return &pd->ibpd;
 406 }
 407
 408 static int mthca_dealloc_pd(struct ib_pd *pd)
 409 {
 410         mthca_pd_free(to_mdev(pd->device), to_mpd(pd));
 411         kfree(pd);
 412
 413         return 0;
 414 }
 415
 416 static struct ib_ah *mthca_ah_create(struct ib_pd *pd,
 417                                      struct ib_ah_attr *ah_attr)
 418 {
 419         int err;
 420         struct mthca_ah *ah;
 421
 422         ah = kmalloc(sizeof *ah, GFP_ATOMIC);
 423         if (!ah)
 424                 return ERR_PTR(-ENOMEM);
 425
 426         err = mthca_create_ah(to_mdev(pd->device), to_mpd(pd), ah_attr, ah);
 427         if (err) {
 428                 kfree(ah);
 429                 return ERR_PTR(err);
 430         }
 431
 432         return &ah->ibah;
 433 }
 434
 435 static int mthca_ah_destroy(struct ib_ah *ah)
 436 {
 437         mthca_destroy_ah(to_mdev(ah->device), to_mah(ah));
 438         kfree(ah);
 439
 440         return 0;
 441 }
 442
 443 static struct ib_srq *mthca_create_srq(struct ib_pd *pd,
 444                                        struct ib_srq_init_attr *init_attr,
 445                                        struct ib_udata *udata)
 446 {
 447         struct mthca_create_srq ucmd;
 448         struct mthca_ucontext *context = NULL;
 449         struct mthca_srq *srq;
 450         int err;
 451
 452         srq = kmalloc(sizeof *srq, GFP_KERNEL);
 453         if (!srq)
 454                 return ERR_PTR(-ENOMEM);
 455
 456         if (pd->uobject) {
 457                 context = to_mucontext(pd->uobject->context);
 458
 459                 if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd))
 460                         return ERR_PTR(-EFAULT);
 461
 462                 err = mthca_map_user_db(to_mdev(pd->device), &context->uar,
 463                                         context->db_tab, ucmd.db_index,
 464                                         ucmd.db_page);
 465
 466                 if (err)
 467                         goto err_free;
 468
 469                 srq->mr.ibmr.lkey = ucmd.lkey;
 470                 srq->db_index     = ucmd.db_index;
 471         }
 472
 473         err = mthca_alloc_srq(to_mdev(pd->device), to_mpd(pd),
 474                               &init_attr->attr, srq);
 475
 476         if (err && pd->uobject)
 477                 mthca_unmap_user_db(to_mdev(pd->device), &context->uar,
 478                                     context->db_tab, ucmd.db_index);
 479
 480         if (err)
 481                 goto err_free;
 482
 483         if (context && ib_copy_to_udata(udata, &srq->srqn, sizeof (__u32))) {
 484                 mthca_free_srq(to_mdev(pd->device), srq);
 485                 err = -EFAULT;
 486                 goto err_free;
 487         }
 488
 489         return &srq->ibsrq;
 490
 491 err_free:
 492         kfree(srq);
 493
 494         return ERR_PTR(err);
 495 }
 496
 497 static int mthca_destroy_srq(struct ib_srq *srq)
 498 {
 499         struct mthca_ucontext *context;
 500
 501         if (srq->uobject) {
 502                 context = to_mucontext(srq->uobject->context);
 503
 504                 mthca_unmap_user_db(to_mdev(srq->device), &context->uar,
 505                                     context->db_tab, to_msrq(srq)->db_index);
 506         }
 507
 508         mthca_free_srq(to_mdev(srq->device), to_msrq(srq));
 509         kfree(srq);
 510
 511         return 0;
 512 }
 513
 514 static struct ib_qp *mthca_create_qp(struct ib_pd *pd,
 515                                      struct ib_qp_init_attr *init_attr,
 516                                      struct ib_udata *udata)
 517 {
 518         struct mthca_create_qp ucmd;
 519         struct mthca_qp *qp;
 520         int err;
 521
 522         switch (init_attr->qp_type) {
 523         case IB_QPT_RC:
 524         case IB_QPT_UC:
 525         case IB_QPT_UD:
 526         {
 527                 struct mthca_ucontext *context;
 528
 529                 qp = kmalloc(sizeof *qp, GFP_KERNEL);
 530                 if (!qp)
 531                         return ERR_PTR(-ENOMEM);
 532
 533                 if (pd->uobject) {
 534                         context = to_mucontext(pd->uobject->context);
 535
 536                         if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd))
 537                                 return ERR_PTR(-EFAULT);
 538
 539                         err = mthca_map_user_db(to_mdev(pd->device), &context->uar,
 540                                                 context->db_tab,
 541                                                 ucmd.sq_db_index, ucmd.sq_db_page);
 542                         if (err) {
 543                                 kfree(qp);
 544                                 return ERR_PTR(err);
 545                         }
 546
 547                         err = mthca_map_user_db(to_mdev(pd->device), &context->uar,
 548                                                 context->db_tab,
 549                                                 ucmd.rq_db_index, ucmd.rq_db_page);
 550                         if (err) {
 551                                 mthca_unmap_user_db(to_mdev(pd->device),
 552                                                     &context->uar,
 553                                                     context->db_tab,
 554                                                     ucmd.sq_db_index);
 555                                 kfree(qp);
 556                                 return ERR_PTR(err);
 557                         }
 558
 559                         qp->mr.ibmr.lkey = ucmd.lkey;
 560                         qp->sq.db_index  = ucmd.sq_db_index;
 561                         qp->rq.db_index  = ucmd.rq_db_index;
 562                 }
 563
 564                 err = mthca_alloc_qp(to_mdev(pd->device), to_mpd(pd),
 565                                      to_mcq(init_attr->send_cq),
 566                                      to_mcq(init_attr->recv_cq),
 567                                      init_attr->qp_type, init_attr->sq_sig_type,
 568                                      &init_attr->cap, qp);
 569
 570                 if (err && pd->uobject) {
 571                         context = to_mucontext(pd->uobject->context);
 572
 573                         mthca_unmap_user_db(to_mdev(pd->device),
 574                                             &context->uar,
 575                                             context->db_tab,
 576                                             ucmd.sq_db_index);
 577                         mthca_unmap_user_db(to_mdev(pd->device),
 578                                             &context->uar,
 579                                             context->db_tab,
 580                                             ucmd.rq_db_index);
 581                 }
 582
 583                 qp->ibqp.qp_num = qp->qpn;
 584                 break;
 585         }
 586         case IB_QPT_SMI:
 587         case IB_QPT_GSI:
 588         {
 589                 /* Don't allow userspace to create special QPs */
 590                 if (pd->uobject)
 591                         return ERR_PTR(-EINVAL);
 592
 593                 qp = kmalloc(sizeof (struct mthca_sqp), GFP_KERNEL);
 594                 if (!qp)
 595                         return ERR_PTR(-ENOMEM);
 596
 597                 qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 : 1;
 598
 599                 err = mthca_alloc_sqp(to_mdev(pd->device), to_mpd(pd),
 600                                       to_mcq(init_attr->send_cq),
 601                                       to_mcq(init_attr->recv_cq),
 602                                       init_attr->sq_sig_type, &init_attr->cap,
 603                                       qp->ibqp.qp_num, init_attr->port_num,
 604                                       to_msqp(qp));
 605                 break;
 606         }
 607         default:
 608                 /* Don't support raw QPs */
 609                 return ERR_PTR(-ENOSYS);
 610         }
 611
 612         if (err) {
 613                 kfree(qp);
 614                 return ERR_PTR(err);
 615         }
 616
 617         init_attr->cap.max_inline_data = 0;
 618         init_attr->cap.max_send_wr     = qp->sq.max;
 619         init_attr->cap.max_recv_wr     = qp->rq.max;
 620         init_attr->cap.max_send_sge    = qp->sq.max_gs;
 621         init_attr->cap.max_recv_sge    = qp->rq.max_gs;
 622
 623         return &qp->ibqp;
 624 }
 625
 626 static int mthca_destroy_qp(struct ib_qp *qp)
 627 {
 628         if (qp->uobject) {
 629                 mthca_unmap_user_db(to_mdev(qp->device),
 630                                     &to_mucontext(qp->uobject->context)->uar,
 631                                     to_mucontext(qp->uobject->context)->db_tab,
 632                                     to_mqp(qp)->sq.db_index);
 633                 mthca_unmap_user_db(to_mdev(qp->device),
 634                                     &to_mucontext(qp->uobject->context)->uar,
 635                                     to_mucontext(qp->uobject->context)->db_tab,
 636                                     to_mqp(qp)->rq.db_index);
 637         }
 638         mthca_free_qp(to_mdev(qp->device), to_mqp(qp));
 639         kfree(qp);
 640         return 0;
 641 }
 642
 643 static struct ib_cq *mthca_create_cq(struct ib_device *ibdev, int entries,
 644                                      struct ib_ucontext *context,
 645                                      struct ib_udata *udata)
 646 {
 647         struct mthca_create_cq ucmd;
 648         struct mthca_cq *cq;
 649         int nent;
 650         int err;
 651
 652         if (entries < 1 || entries > to_mdev(ibdev)->limits.max_cqes)
 653                 return ERR_PTR(-EINVAL);
 654
 655         if (context) {
 656                 if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd))
 657                         return ERR_PTR(-EFAULT);
 658
 659                 err = mthca_map_user_db(to_mdev(ibdev), &to_mucontext(context)->uar,
 660                                         to_mucontext(context)->db_tab,
 661                                         ucmd.set_db_index, ucmd.set_db_page);
 662                 if (err)
 663                         return ERR_PTR(err);
 664
 665                 err = mthca_map_user_db(to_mdev(ibdev), &to_mucontext(context)->uar,
 666                                         to_mucontext(context)->db_tab,
 667                                         ucmd.arm_db_index, ucmd.arm_db_page);
 668                 if (err)
 669                         goto err_unmap_set;
 670         }
 671
 672         cq = kmalloc(sizeof *cq, GFP_KERNEL);
 673         if (!cq) {
 674                 err = -ENOMEM;
 675                 goto err_unmap_arm;
 676         }
 677
 678         if (context) {
 679                 cq->mr.ibmr.lkey    = ucmd.lkey;
 680                 cq->set_ci_db_index = ucmd.set_db_index;
 681                 cq->arm_db_index    = ucmd.arm_db_index;
 682         }
 683
 684         for (nent = 1; nent <= entries; nent <<= 1)
 685                 ; /* nothing */
 686
 687         err = mthca_init_cq(to_mdev(ibdev), nent,
 688                             context ? to_mucontext(context) : NULL,
 689                             context ? ucmd.pdn : to_mdev(ibdev)->driver_pd.pd_num,
 690                             cq);
 691         if (err)
 692                 goto err_free;
 693
 694         if (context && ib_copy_to_udata(udata, &cq->cqn, sizeof (__u32))) {
 695                 mthca_free_cq(to_mdev(ibdev), cq);
 696                 goto err_free;
 697         }
 698
 699         return &cq->ibcq;
 700
 701 err_free:
 702         kfree(cq);
 703
 704 err_unmap_arm:
 705         if (context)
 706                 mthca_unmap_user_db(to_mdev(ibdev), &to_mucontext(context)->uar,
 707                                     to_mucontext(context)->db_tab, ucmd.arm_db_index);
 708
 709 err_unmap_set:
 710         if (context)
 711                 mthca_unmap_user_db(to_mdev(ibdev), &to_mucontext(context)->uar,
 712                                     to_mucontext(context)->db_tab, ucmd.set_db_index);
 713
 714         return ERR_PTR(err);
 715 }
 716
 717 static int mthca_destroy_cq(struct ib_cq *cq)
 718 {
 719         if (cq->uobject) {
 720                 mthca_unmap_user_db(to_mdev(cq->device),
 721                                     &to_mucontext(cq->uobject->context)->uar,
 722                                     to_mucontext(cq->uobject->context)->db_tab,
 723                                     to_mcq(cq)->arm_db_index);
 724                 mthca_unmap_user_db(to_mdev(cq->device),
 725                                     &to_mucontext(cq->uobject->context)->uar,
 726                                     to_mucontext(cq->uobject->context)->db_tab,
 727                                     to_mcq(cq)->set_ci_db_index);
 728         }
 729         mthca_free_cq(to_mdev(cq->device), to_mcq(cq));
 730         kfree(cq);
 731
 732         return 0;
 733 }
 734
 735 static inline u32 convert_access(int acc)
 736 {
 737         return (acc & IB_ACCESS_REMOTE_ATOMIC ? MTHCA_MPT_FLAG_ATOMIC       : 0) |
 738                (acc & IB_ACCESS_REMOTE_WRITE  ? MTHCA_MPT_FLAG_REMOTE_WRITE : 0) |
 739                (acc & IB_ACCESS_REMOTE_READ   ? MTHCA_MPT_FLAG_REMOTE_READ  : 0) |
 740                (acc & IB_ACCESS_LOCAL_WRITE   ? MTHCA_MPT_FLAG_LOCAL_WRITE  : 0) |
 741                MTHCA_MPT_FLAG_LOCAL_READ;
 742 }
 743
 744 static struct ib_mr *mthca_get_dma_mr(struct ib_pd *pd, int acc)
 745 {
 746         struct mthca_mr *mr;
 747         int err;
 748
 749         mr = kmalloc(sizeof *mr, GFP_KERNEL);
 750         if (!mr)
 751                 return ERR_PTR(-ENOMEM);
 752
 753         err = mthca_mr_alloc_notrans(to_mdev(pd->device),
 754                                      to_mpd(pd)->pd_num,
 755                                      convert_access(acc), mr);
 756
 757         if (err) {
 758                 kfree(mr);
 759                 return ERR_PTR(err);
 760         }
 761
 762         return &mr->ibmr;
 763 }
 764
 765 static struct ib_mr *mthca_reg_phys_mr(struct ib_pd       *pd,
 766                                        struct ib_phys_buf *buffer_list,
 767                                        int                 num_phys_buf,
 768                                        int                 acc,
 769                                        u64                *iova_start)
 770 {
 771         struct mthca_mr *mr;
 772         u64 *page_list;
 773         u64 total_size;
 774         u64 mask;
 775         int shift;
 776         int npages;
 777         int err;
 778         int i, j, n;
 779
 780         /* First check that we have enough alignment */
 781         if ((*iova_start & ~PAGE_MASK) != (buffer_list[0].addr & ~PAGE_MASK))
 782                 return ERR_PTR(-EINVAL);
 783
 784         if (num_phys_buf > 1 &&
 785             ((buffer_list[0].addr + buffer_list[0].size) & ~PAGE_MASK))
 786                 return ERR_PTR(-EINVAL);
 787
 788         mask = 0;
 789         total_size = 0;
 790         for (i = 0; i < num_phys_buf; ++i) {
 791                 if (i != 0 && buffer_list[i].addr & ~PAGE_MASK)
 792                         return ERR_PTR(-EINVAL);
 793                 if (i != 0 && i != num_phys_buf - 1 &&
 794                     (buffer_list[i].size & ~PAGE_MASK))
 795                         return ERR_PTR(-EINVAL);
 796
 797                 total_size += buffer_list[i].size;
 798                 if (i > 0)
 799                         mask |= buffer_list[i].addr;
 800         }
 801
 802         /* Find largest page shift we can use to cover buffers */
 803         for (shift = PAGE_SHIFT; shift < 31; ++shift)
 804                 if (num_phys_buf > 1) {
 805                         if ((1ULL << shift) & mask)
 806                                 break;
 807                 } else {
 808                         if (1ULL << shift >=
 809                             buffer_list[0].size +
 810                             (buffer_list[0].addr & ((1ULL << shift) - 1)))
 811                                 break;
 812                 }
 813
 814         buffer_list[0].size += buffer_list[0].addr & ((1ULL << shift) - 1);
 815         buffer_list[0].addr &= ~0ull << shift;
 816
 817         mr = kmalloc(sizeof *mr, GFP_KERNEL);
 818         if (!mr)
 819                 return ERR_PTR(-ENOMEM);
 820
 821         npages = 0;
 822         for (i = 0; i < num_phys_buf; ++i)
 823                 npages += (buffer_list[i].size + (1ULL << shift) - 1) >> shift;
 824
 825         if (!npages)
 826                 return &mr->ibmr;
 827
 828         page_list = kmalloc(npages * sizeof *page_list, GFP_KERNEL);
 829         if (!page_list) {
 830                 kfree(mr);
 831                 return ERR_PTR(-ENOMEM);
 832         }
 833
 834         n = 0;
 835         for (i = 0; i < num_phys_buf; ++i)
 836                 for (j = 0;
 837                      j < (buffer_list[i].size + (1ULL << shift) - 1) >> shift;
 838                      ++j)
 839                         page_list[n++] = buffer_list[i].addr + ((u64) j << shift);
 840
 841         mthca_dbg(to_mdev(pd->device), "Registering memory at %llx (iova %llx) "
 842                   "in PD %x; shift %d, npages %d.\n",
 843                   (unsigned long long) buffer_list[0].addr,
 844                   (unsigned long long) *iova_start,
 845                   to_mpd(pd)->pd_num,
 846                   shift, npages);
 847
 848         err = mthca_mr_alloc_phys(to_mdev(pd->device),
 849                                   to_mpd(pd)->pd_num,
 850                                   page_list, shift, npages,
 851                                   *iova_start, total_size,
 852                                   convert_access(acc), mr);
 853
 854         if (err) {
 855                 kfree(page_list);
 856                 kfree(mr);
 857                 return ERR_PTR(err);
 858         }
 859
 860         kfree(page_list);
 861         return &mr->ibmr;
 862 }
 863
 864 static struct ib_mr *mthca_reg_user_mr(struct ib_pd *pd, struct ib_umem *region,
 865                                        int acc, struct ib_udata *udata)
 866 {
 867         struct mthca_dev *dev = to_mdev(pd->device);
 868         struct ib_umem_chunk *chunk;
 869         struct mthca_mr *mr;
 870         u64 *pages;
 871         int shift, n, len;
 872         int i, j, k;
 873         int err = 0;
 874
 875         shift = ffs(region->page_size) - 1;
 876
 877         mr = kmalloc(sizeof *mr, GFP_KERNEL);
 878         if (!mr)
 879                 return ERR_PTR(-ENOMEM);
 880
 881         n = 0;
 882         list_for_each_entry(chunk, &region->chunk_list, list)
 883                 n += chunk->nents;
 884
 885         mr->mtt = mthca_alloc_mtt(dev, n);
 886         if (IS_ERR(mr->mtt)) {
 887                 err = PTR_ERR(mr->mtt);
 888                 goto err;
 889         }
 890
 891         pages = (u64 *) __get_free_page(GFP_KERNEL);
 892         if (!pages) {
 893                 err = -ENOMEM;
 894                 goto err_mtt;
 895         }
 896
 897         i = n = 0;
 898
 899         list_for_each_entry(chunk, &region->chunk_list, list)
 900                 for (j = 0; j < chunk->nmap; ++j) {
 901                         len = sg_dma_len(&chunk->page_list[j]) >> shift;
 902                         for (k = 0; k < len; ++k) {
 903                                 pages[i++] = sg_dma_address(&chunk->page_list[j]) +
 904                                         region->page_size * k;
 905                                 /*
 906                                  * Be friendly to WRITE_MTT command
 907                                  * and leave two empty slots for the
 908                                  * index and reserved fields of the
 909                                  * mailbox.
 910                                  */
 911                                 if (i == PAGE_SIZE / sizeof (u64) - 2) {
 912                                         err = mthca_write_mtt(dev, mr->mtt,
 913                                                               n, pages, i);
 914                                         if (err)
 915                                                 goto mtt_done;
 916                                         n += i;
 917                                         i = 0;
 918                                 }
 919                         }
 920                 }
 921
 922         if (i)
 923                 err = mthca_write_mtt(dev, mr->mtt, n, pages, i);
 924 mtt_done:
 925         free_page((unsigned long) pages);
 926         if (err)
 927                 goto err_mtt;
 928
 929         err = mthca_mr_alloc(dev, to_mpd(pd)->pd_num, shift, region->virt_base,
 930                              region->length, convert_access(acc), mr);
 931
 932         if (err)
 933                 goto err_mtt;
 934
 935         return &mr->ibmr;
 936
 937 err_mtt:
 938         mthca_free_mtt(dev, mr->mtt);
 939
 940 err:
 941         kfree(mr);
 942         return ERR_PTR(err);
 943 }
 944
 945 static int mthca_dereg_mr(struct ib_mr *mr)
 946 {
 947         struct mthca_mr *mmr = to_mmr(mr);
 948         mthca_free_mr(to_mdev(mr->device), mmr);
 949         kfree(mmr);
 950         return 0;
 951 }
 952
 953 static struct ib_fmr *mthca_alloc_fmr(struct ib_pd *pd, int mr_access_flags,
 954                                       struct ib_fmr_attr *fmr_attr)
 955 {
 956         struct mthca_fmr *fmr;
 957         int err;
 958
 959         fmr = kmalloc(sizeof *fmr, GFP_KERNEL);
 960         if (!fmr)
 961                 return ERR_PTR(-ENOMEM);
 962
 963         memcpy(&fmr->attr, fmr_attr, sizeof *fmr_attr);
 964         err = mthca_fmr_alloc(to_mdev(pd->device), to_mpd(pd)->pd_num,
 965                              convert_access(mr_access_flags), fmr);
 966
 967         if (err) {
 968                 kfree(fmr);
 969                 return ERR_PTR(err);
 970         }
 971
 972         return &fmr->ibmr;
 973 }
 974
 975 static int mthca_dealloc_fmr(struct ib_fmr *fmr)
 976 {
 977         struct mthca_fmr *mfmr = to_mfmr(fmr);
 978         int err;
 979
 980         err = mthca_free_fmr(to_mdev(fmr->device), mfmr);
 981         if (err)
 982                 return err;
 983
 984         kfree(mfmr);
 985         return 0;
 986 }
 987
 988 static int mthca_unmap_fmr(struct list_head *fmr_list)
 989 {
 990         struct ib_fmr *fmr;
 991         int err;
 992         u8 status;
 993         struct mthca_dev *mdev = NULL;
 994
 995         list_for_each_entry(fmr, fmr_list, list) {
 996                 if (mdev && to_mdev(fmr->device) != mdev)
 997                         return -EINVAL;
 998                 mdev = to_mdev(fmr->device);
 999         }
1000
1001         if (!mdev)
1002                 return 0;
1003
1004         if (mthca_is_memfree(mdev)) {
1005                 list_for_each_entry(fmr, fmr_list, list)
1006                         mthca_arbel_fmr_unmap(mdev, to_mfmr(fmr));
1007
1008                 wmb();
1009         } else
1010                 list_for_each_entry(fmr, fmr_list, list)
1011                         mthca_tavor_fmr_unmap(mdev, to_mfmr(fmr));
1012
1013         err = mthca_SYNC_TPT(mdev, &status);
1014         if (err)
1015                 return err;
1016         if (status)
1017                 return -EINVAL;
1018         return 0;
1019 }
1020
1021 static ssize_t show_rev(struct class_device *cdev, char *buf)
1022 {
1023         struct mthca_dev *dev = container_of(cdev, struct mthca_dev, ib_dev.class_dev);
1024         return sprintf(buf, "%x\n", dev->rev_id);
1025 }
1026
1027 static ssize_t show_fw_ver(struct class_device *cdev, char *buf)
1028 {
1029         struct mthca_dev *dev = container_of(cdev, struct mthca_dev, ib_dev.class_dev);
1030         return sprintf(buf, "%x.%x.%x\n", (int) (dev->fw_ver >> 32),
1031                        (int) (dev->fw_ver >> 16) & 0xffff,
1032                        (int) dev->fw_ver & 0xffff);
1033 }
1034
1035 static ssize_t show_hca(struct class_device *cdev, char *buf)
1036 {
1037         struct mthca_dev *dev = container_of(cdev, struct mthca_dev, ib_dev.class_dev);
1038         switch (dev->pdev->device) {
1039         case PCI_DEVICE_ID_MELLANOX_TAVOR:
1040                 return sprintf(buf, "MT23108\n");
1041         case PCI_DEVICE_ID_MELLANOX_ARBEL_COMPAT:
1042                 return sprintf(buf, "MT25208 (MT23108 compat mode)\n");
1043         case PCI_DEVICE_ID_MELLANOX_ARBEL:
1044                 return sprintf(buf, "MT25208\n");
1045         case PCI_DEVICE_ID_MELLANOX_SINAI:
1046         case PCI_DEVICE_ID_MELLANOX_SINAI_OLD:
1047                 return sprintf(buf, "MT25204\n");
1048         default:
1049                 return sprintf(buf, "unknown\n");
1050         }
1051 }
1052
1053 static ssize_t show_board(struct class_device *cdev, char *buf)
1054 {
1055         struct mthca_dev *dev = container_of(cdev, struct mthca_dev, ib_dev.class_dev);
1056         return sprintf(buf, "%.*s\n", MTHCA_BOARD_ID_LEN, dev->board_id);
1057 }
1058
1059 static CLASS_DEVICE_ATTR(hw_rev,   S_IRUGO, show_rev,    NULL);
1060 static CLASS_DEVICE_ATTR(fw_ver,   S_IRUGO, show_fw_ver, NULL);
1061 static CLASS_DEVICE_ATTR(hca_type, S_IRUGO, show_hca,    NULL);
1062 static CLASS_DEVICE_ATTR(board_id, S_IRUGO, show_board,  NULL);
1063
1064 static struct class_device_attribute *mthca_class_attributes[] = {
1065         &class_device_attr_hw_rev,
1066         &class_device_attr_fw_ver,
1067         &class_device_attr_hca_type,
1068         &class_device_attr_board_id
1069 };
1070
1071 int mthca_register_device(struct mthca_dev *dev)
1072 {
1073         int ret;
1074         int i;
1075
1076         strlcpy(dev->ib_dev.name, "mthca%d", IB_DEVICE_NAME_MAX);
1077         dev->ib_dev.owner                = THIS_MODULE;
1078
1079         dev->ib_dev.uverbs_abi_ver       = MTHCA_UVERBS_ABI_VERSION;
1080         dev->ib_dev.node_type            = IB_NODE_CA;
1081         dev->ib_dev.phys_port_cnt        = dev->limits.num_ports;
1082         dev->ib_dev.dma_device           = &dev->pdev->dev;
1083         dev->ib_dev.class_dev.dev        = &dev->pdev->dev;
1084         dev->ib_dev.query_device         = mthca_query_device;
1085         dev->ib_dev.query_port           = mthca_query_port;
1086         dev->ib_dev.modify_port          = mthca_modify_port;
1087         dev->ib_dev.query_pkey           = mthca_query_pkey;
1088         dev->ib_dev.query_gid            = mthca_query_gid;
1089         dev->ib_dev.alloc_ucontext       = mthca_alloc_ucontext;
1090         dev->ib_dev.dealloc_ucontext     = mthca_dealloc_ucontext;
1091         dev->ib_dev.mmap                 = mthca_mmap_uar;
1092         dev->ib_dev.alloc_pd             = mthca_alloc_pd;
1093         dev->ib_dev.dealloc_pd           = mthca_dealloc_pd;
1094         dev->ib_dev.create_ah            = mthca_ah_create;
1095         dev->ib_dev.destroy_ah           = mthca_ah_destroy;
1096
1097         if (dev->mthca_flags & MTHCA_FLAG_SRQ) {
1098                 dev->ib_dev.create_srq           = mthca_create_srq;
1099                 dev->ib_dev.modify_srq           = mthca_modify_srq;
1100                 dev->ib_dev.destroy_srq          = mthca_destroy_srq;
1101
1102                 if (mthca_is_memfree(dev))
1103                         dev->ib_dev.post_srq_recv = mthca_arbel_post_srq_recv;
1104                 else
1105                         dev->ib_dev.post_srq_recv = mthca_tavor_post_srq_recv;
1106         }
1107
1108         dev->ib_dev.create_qp            = mthca_create_qp;
1109         dev->ib_dev.modify_qp            = mthca_modify_qp;
1110         dev->ib_dev.destroy_qp           = mthca_destroy_qp;
1111         dev->ib_dev.create_cq            = mthca_create_cq;
1112         dev->ib_dev.destroy_cq           = mthca_destroy_cq;
1113         dev->ib_dev.poll_cq              = mthca_poll_cq;
1114         dev->ib_dev.get_dma_mr           = mthca_get_dma_mr;
1115         dev->ib_dev.reg_phys_mr          = mthca_reg_phys_mr;
1116         dev->ib_dev.reg_user_mr          = mthca_reg_user_mr;
1117         dev->ib_dev.dereg_mr             = mthca_dereg_mr;
1118
1119         if (dev->mthca_flags & MTHCA_FLAG_FMR) {
1120                 dev->ib_dev.alloc_fmr            = mthca_alloc_fmr;
1121                 dev->ib_dev.unmap_fmr            = mthca_unmap_fmr;
1122                 dev->ib_dev.dealloc_fmr          = mthca_dealloc_fmr;
1123                 if (mthca_is_memfree(dev))
1124                         dev->ib_dev.map_phys_fmr = mthca_arbel_map_phys_fmr;
1125                 else
1126                         dev->ib_dev.map_phys_fmr = mthca_tavor_map_phys_fmr;
1127         }
1128
1129         dev->ib_dev.attach_mcast         = mthca_multicast_attach;
1130         dev->ib_dev.detach_mcast         = mthca_multicast_detach;
1131         dev->ib_dev.process_mad          = mthca_process_mad;
1132
1133         if (mthca_is_memfree(dev)) {
1134                 dev->ib_dev.req_notify_cq = mthca_arbel_arm_cq;
1135                 dev->ib_dev.post_send     = mthca_arbel_post_send;
1136                 dev->ib_dev.post_recv     = mthca_arbel_post_receive;
1137         } else {
1138                 dev->ib_dev.req_notify_cq = mthca_tavor_arm_cq;
1139                 dev->ib_dev.post_send     = mthca_tavor_post_send;
1140                 dev->ib_dev.post_recv     = mthca_tavor_post_receive;
1141         }
1142
1143         init_MUTEX(&dev->cap_mask_mutex);
1144
1145         ret = ib_register_device(&dev->ib_dev);
1146         if (ret)
1147                 return ret;
1148
1149         for (i = 0; i < ARRAY_SIZE(mthca_class_attributes); ++i) {
1150                 ret = class_device_create_file(&dev->ib_dev.class_dev,
1151                                                mthca_class_attributes[i]);
1152                 if (ret) {
1153                         ib_unregister_device(&dev->ib_dev);
1154                         return ret;
1155                 }
1156         }
1157
1158         return 0;
1159 }
1160
1161 void mthca_unregister_device(struct mthca_dev *dev)
1162 {
1163         ib_unregister_device(&dev->ib_dev);
1164 }