git.oblomov.eu Git - linux-2.6/blob - drivers/infiniband/hw/mthca/mthca_main.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 Mellanox Technologies. All rights reserved.
   5  *
   6  * This software is available to you under a choice of one of two
   7  * licenses.  You may choose to be licensed under the terms of the GNU
   8  * General Public License (GPL) Version 2, available from the file
   9  * COPYING in the main directory of this source tree, or the
  10  * OpenIB.org BSD license below:
  11  *
  12  *     Redistribution and use in source and binary forms, with or
  13  *     without modification, are permitted provided that the following
  14  *     conditions are met:
  15  *
  16  *      - Redistributions of source code must retain the above
  17  *        copyright notice, this list of conditions and the following
  18  *        disclaimer.
  19  *
  20  *      - Redistributions in binary form must reproduce the above
  21  *        copyright notice, this list of conditions and the following
  22  *        disclaimer in the documentation and/or other materials
  23  *        provided with the distribution.
  24  *
  25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  32  * SOFTWARE.
  33  */
  34
  35 #include <linux/module.h>
  36 #include <linux/init.h>
  37 #include <linux/errno.h>
  38 #include <linux/pci.h>
  39 #include <linux/interrupt.h>
  40
  41 #include "mthca_dev.h"
  42 #include "mthca_config_reg.h"
  43 #include "mthca_cmd.h"
  44 #include "mthca_profile.h"
  45 #include "mthca_memfree.h"
  46 #include "mthca_wqe.h"
  47
  48 MODULE_AUTHOR("Roland Dreier");
  49 MODULE_DESCRIPTION("Mellanox InfiniBand HCA low-level driver");
  50 MODULE_LICENSE("Dual BSD/GPL");
  51 MODULE_VERSION(DRV_VERSION);
  52
  53 #ifdef CONFIG_INFINIBAND_MTHCA_DEBUG
  54
  55 int mthca_debug_level = 0;
  56 module_param_named(debug_level, mthca_debug_level, int, 0644);
  57 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
  58
  59 #endif /* CONFIG_INFINIBAND_MTHCA_DEBUG */
  60
  61 #ifdef CONFIG_PCI_MSI
  62
  63 static int msi_x = 1;
  64 module_param(msi_x, int, 0444);
  65 MODULE_PARM_DESC(msi_x, "attempt to use MSI-X if nonzero");
  66
  67 #else /* CONFIG_PCI_MSI */
  68
  69 #define msi_x (0)
  70
  71 #endif /* CONFIG_PCI_MSI */
  72
  73 static int tune_pci = 0;
  74 module_param(tune_pci, int, 0444);
  75 MODULE_PARM_DESC(tune_pci, "increase PCI burst from the default set by BIOS if nonzero");
  76
  77 DEFINE_MUTEX(mthca_device_mutex);
  78
  79 #define MTHCA_DEFAULT_NUM_QP            (1 << 16)
  80 #define MTHCA_DEFAULT_RDB_PER_QP        (1 << 2)
  81 #define MTHCA_DEFAULT_NUM_CQ            (1 << 16)
  82 #define MTHCA_DEFAULT_NUM_MCG           (1 << 13)
  83 #define MTHCA_DEFAULT_NUM_MPT           (1 << 17)
  84 #define MTHCA_DEFAULT_NUM_MTT           (1 << 20)
  85 #define MTHCA_DEFAULT_NUM_UDAV          (1 << 15)
  86 #define MTHCA_DEFAULT_NUM_RESERVED_MTTS (1 << 18)
  87 #define MTHCA_DEFAULT_NUM_UARC_SIZE     (1 << 18)
  88
  89 static struct mthca_profile hca_profile = {
  90         .num_qp             = MTHCA_DEFAULT_NUM_QP,
  91         .rdb_per_qp         = MTHCA_DEFAULT_RDB_PER_QP,
  92         .num_cq             = MTHCA_DEFAULT_NUM_CQ,
  93         .num_mcg            = MTHCA_DEFAULT_NUM_MCG,
  94         .num_mpt            = MTHCA_DEFAULT_NUM_MPT,
  95         .num_mtt            = MTHCA_DEFAULT_NUM_MTT,
  96         .num_udav           = MTHCA_DEFAULT_NUM_UDAV,          /* Tavor only */
  97         .fmr_reserved_mtts  = MTHCA_DEFAULT_NUM_RESERVED_MTTS, /* Tavor only */
  98         .uarc_size          = MTHCA_DEFAULT_NUM_UARC_SIZE,     /* Arbel only */
  99 };
 100
 101 module_param_named(num_qp, hca_profile.num_qp, int, 0444);
 102 MODULE_PARM_DESC(num_qp, "maximum number of QPs per HCA");
 103
 104 module_param_named(rdb_per_qp, hca_profile.rdb_per_qp, int, 0444);
 105 MODULE_PARM_DESC(rdb_per_qp, "number of RDB buffers per QP");
 106
 107 module_param_named(num_cq, hca_profile.num_cq, int, 0444);
 108 MODULE_PARM_DESC(num_cq, "maximum number of CQs per HCA");
 109
 110 module_param_named(num_mcg, hca_profile.num_mcg, int, 0444);
 111 MODULE_PARM_DESC(num_mcg, "maximum number of multicast groups per HCA");
 112
 113 module_param_named(num_mpt, hca_profile.num_mpt, int, 0444);
 114 MODULE_PARM_DESC(num_mpt,
 115                 "maximum number of memory protection table entries per HCA");
 116
 117 module_param_named(num_mtt, hca_profile.num_mtt, int, 0444);
 118 MODULE_PARM_DESC(num_mtt,
 119                  "maximum number of memory translation table segments per HCA");
 120
 121 module_param_named(num_udav, hca_profile.num_udav, int, 0444);
 122 MODULE_PARM_DESC(num_udav, "maximum number of UD address vectors per HCA");
 123
 124 module_param_named(fmr_reserved_mtts, hca_profile.fmr_reserved_mtts, int, 0444);
 125 MODULE_PARM_DESC(fmr_reserved_mtts,
 126                  "number of memory translation table segments reserved for FMR");
 127
 128 static char mthca_version[] __devinitdata =
 129         DRV_NAME ": Mellanox InfiniBand HCA driver v"
 130         DRV_VERSION " (" DRV_RELDATE ")\n";
 131
 132 static int mthca_tune_pci(struct mthca_dev *mdev)
 133 {
 134         if (!tune_pci)
 135                 return 0;
 136
 137         /* First try to max out Read Byte Count */
 138         if (pci_find_capability(mdev->pdev, PCI_CAP_ID_PCIX)) {
 139                 if (pcix_set_mmrbc(mdev->pdev, pcix_get_max_mmrbc(mdev->pdev))) {
 140                         mthca_err(mdev, "Couldn't set PCI-X max read count, "
 141                                 "aborting.\n");
 142                         return -ENODEV;
 143                 }
 144         } else if (!(mdev->mthca_flags & MTHCA_FLAG_PCIE))
 145                 mthca_info(mdev, "No PCI-X capability, not setting RBC.\n");
 146
 147         if (pci_find_capability(mdev->pdev, PCI_CAP_ID_EXP)) {
 148                 if (pcie_set_readrq(mdev->pdev, 4096)) {
 149                         mthca_err(mdev, "Couldn't write PCI Express read request, "
 150                                 "aborting.\n");
 151                         return -ENODEV;
 152                 }
 153         } else if (mdev->mthca_flags & MTHCA_FLAG_PCIE)
 154                 mthca_info(mdev, "No PCI Express capability, "
 155                            "not setting Max Read Request Size.\n");
 156
 157         return 0;
 158 }
 159
 160 static int mthca_dev_lim(struct mthca_dev *mdev, struct mthca_dev_lim *dev_lim)
 161 {
 162         int err;
 163         u8 status;
 164
 165         err = mthca_QUERY_DEV_LIM(mdev, dev_lim, &status);
 166         if (err) {
 167                 mthca_err(mdev, "QUERY_DEV_LIM command failed, aborting.\n");
 168                 return err;
 169         }
 170         if (status) {
 171                 mthca_err(mdev, "QUERY_DEV_LIM returned status 0x%02x, "
 172                           "aborting.\n", status);
 173                 return -EINVAL;
 174         }
 175         if (dev_lim->min_page_sz > PAGE_SIZE) {
 176                 mthca_err(mdev, "HCA minimum page size of %d bigger than "
 177                           "kernel PAGE_SIZE of %ld, aborting.\n",
 178                           dev_lim->min_page_sz, PAGE_SIZE);
 179                 return -ENODEV;
 180         }
 181         if (dev_lim->num_ports > MTHCA_MAX_PORTS) {
 182                 mthca_err(mdev, "HCA has %d ports, but we only support %d, "
 183                           "aborting.\n",
 184                           dev_lim->num_ports, MTHCA_MAX_PORTS);
 185                 return -ENODEV;
 186         }
 187
 188         if (dev_lim->uar_size > pci_resource_len(mdev->pdev, 2)) {
 189                 mthca_err(mdev, "HCA reported UAR size of 0x%x bigger than "
 190                           "PCI resource 2 size of 0x%llx, aborting.\n",
 191                           dev_lim->uar_size,
 192                           (unsigned long long)pci_resource_len(mdev->pdev, 2));
 193                 return -ENODEV;
 194         }
 195
 196         mdev->limits.num_ports          = dev_lim->num_ports;
 197         mdev->limits.vl_cap             = dev_lim->max_vl;
 198         mdev->limits.mtu_cap            = dev_lim->max_mtu;
 199         mdev->limits.gid_table_len      = dev_lim->max_gids;
 200         mdev->limits.pkey_table_len     = dev_lim->max_pkeys;
 201         mdev->limits.local_ca_ack_delay = dev_lim->local_ca_ack_delay;
 202         /*
 203          * Need to allow for worst case send WQE overhead and check
 204          * whether max_desc_sz imposes a lower limit than max_sg; UD
 205          * send has the biggest overhead.
 206          */
 207         mdev->limits.max_sg             = min_t(int, dev_lim->max_sg,
 208                                               (dev_lim->max_desc_sz -
 209                                                sizeof (struct mthca_next_seg) -
 210                                                (mthca_is_memfree(mdev) ?
 211                                                 sizeof (struct mthca_arbel_ud_seg) :
 212                                                 sizeof (struct mthca_tavor_ud_seg))) /
 213                                                 sizeof (struct mthca_data_seg));
 214         mdev->limits.max_wqes           = dev_lim->max_qp_sz;
 215         mdev->limits.max_qp_init_rdma   = dev_lim->max_requester_per_qp;
 216         mdev->limits.reserved_qps       = dev_lim->reserved_qps;
 217         mdev->limits.max_srq_wqes       = dev_lim->max_srq_sz;
 218         mdev->limits.reserved_srqs      = dev_lim->reserved_srqs;
 219         mdev->limits.reserved_eecs      = dev_lim->reserved_eecs;
 220         mdev->limits.max_desc_sz        = dev_lim->max_desc_sz;
 221         mdev->limits.max_srq_sge        = mthca_max_srq_sge(mdev);
 222         /*
 223          * Subtract 1 from the limit because we need to allocate a
 224          * spare CQE so the HCA HW can tell the difference between an
 225          * empty CQ and a full CQ.
 226          */
 227         mdev->limits.max_cqes           = dev_lim->max_cq_sz - 1;
 228         mdev->limits.reserved_cqs       = dev_lim->reserved_cqs;
 229         mdev->limits.reserved_eqs       = dev_lim->reserved_eqs;
 230         mdev->limits.reserved_mtts      = dev_lim->reserved_mtts;
 231         mdev->limits.reserved_mrws      = dev_lim->reserved_mrws;
 232         mdev->limits.reserved_uars      = dev_lim->reserved_uars;
 233         mdev->limits.reserved_pds       = dev_lim->reserved_pds;
 234         mdev->limits.port_width_cap     = dev_lim->max_port_width;
 235         mdev->limits.page_size_cap      = ~(u32) (dev_lim->min_page_sz - 1);
 236         mdev->limits.flags              = dev_lim->flags;
 237         /*
 238          * For old FW that doesn't return static rate support, use a
 239          * value of 0x3 (only static rate values of 0 or 1 are handled),
 240          * except on Sinai, where even old FW can handle static rate
 241          * values of 2 and 3.
 242          */
 243         if (dev_lim->stat_rate_support)
 244                 mdev->limits.stat_rate_support = dev_lim->stat_rate_support;
 245         else if (mdev->mthca_flags & MTHCA_FLAG_SINAI_OPT)
 246                 mdev->limits.stat_rate_support = 0xf;
 247         else
 248                 mdev->limits.stat_rate_support = 0x3;
 249
 250         /* IB_DEVICE_RESIZE_MAX_WR not supported by driver.
 251            May be doable since hardware supports it for SRQ.
 252
 253            IB_DEVICE_N_NOTIFY_CQ is supported by hardware but not by driver.
 254
 255            IB_DEVICE_SRQ_RESIZE is supported by hardware but SRQ is not
 256            supported by driver. */
 257         mdev->device_cap_flags = IB_DEVICE_CHANGE_PHY_PORT |
 258                 IB_DEVICE_PORT_ACTIVE_EVENT |
 259                 IB_DEVICE_SYS_IMAGE_GUID |
 260                 IB_DEVICE_RC_RNR_NAK_GEN;
 261
 262         if (dev_lim->flags & DEV_LIM_FLAG_BAD_PKEY_CNTR)
 263                 mdev->device_cap_flags |= IB_DEVICE_BAD_PKEY_CNTR;
 264
 265         if (dev_lim->flags & DEV_LIM_FLAG_BAD_QKEY_CNTR)
 266                 mdev->device_cap_flags |= IB_DEVICE_BAD_QKEY_CNTR;
 267
 268         if (dev_lim->flags & DEV_LIM_FLAG_RAW_MULTI)
 269                 mdev->device_cap_flags |= IB_DEVICE_RAW_MULTI;
 270
 271         if (dev_lim->flags & DEV_LIM_FLAG_AUTO_PATH_MIG)
 272                 mdev->device_cap_flags |= IB_DEVICE_AUTO_PATH_MIG;
 273
 274         if (dev_lim->flags & DEV_LIM_FLAG_UD_AV_PORT_ENFORCE)
 275                 mdev->device_cap_flags |= IB_DEVICE_UD_AV_PORT_ENFORCE;
 276
 277         if (dev_lim->flags & DEV_LIM_FLAG_SRQ)
 278                 mdev->mthca_flags |= MTHCA_FLAG_SRQ;
 279
 280         if (mthca_is_memfree(mdev))
 281                 if (dev_lim->flags & DEV_LIM_FLAG_IPOIB_CSUM)
 282                         mdev->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
 283
 284         return 0;
 285 }
 286
 287 static int mthca_init_tavor(struct mthca_dev *mdev)
 288 {
 289         s64 size;
 290         u8 status;
 291         int err;
 292         struct mthca_dev_lim        dev_lim;
 293         struct mthca_profile        profile;
 294         struct mthca_init_hca_param init_hca;
 295
 296         err = mthca_SYS_EN(mdev, &status);
 297         if (err) {
 298                 mthca_err(mdev, "SYS_EN command failed, aborting.\n");
 299                 return err;
 300         }
 301         if (status) {
 302                 mthca_err(mdev, "SYS_EN returned status 0x%02x, "
 303                           "aborting.\n", status);
 304                 return -EINVAL;
 305         }
 306
 307         err = mthca_QUERY_FW(mdev, &status);
 308         if (err) {
 309                 mthca_err(mdev, "QUERY_FW command failed, aborting.\n");
 310                 goto err_disable;
 311         }
 312         if (status) {
 313                 mthca_err(mdev, "QUERY_FW returned status 0x%02x, "
 314                           "aborting.\n", status);
 315                 err = -EINVAL;
 316                 goto err_disable;
 317         }
 318         err = mthca_QUERY_DDR(mdev, &status);
 319         if (err) {
 320                 mthca_err(mdev, "QUERY_DDR command failed, aborting.\n");
 321                 goto err_disable;
 322         }
 323         if (status) {
 324                 mthca_err(mdev, "QUERY_DDR returned status 0x%02x, "
 325                           "aborting.\n", status);
 326                 err = -EINVAL;
 327                 goto err_disable;
 328         }
 329
 330         err = mthca_dev_lim(mdev, &dev_lim);
 331         if (err) {
 332                 mthca_err(mdev, "QUERY_DEV_LIM command failed, aborting.\n");
 333                 goto err_disable;
 334         }
 335
 336         profile = hca_profile;
 337         profile.num_uar   = dev_lim.uar_size / PAGE_SIZE;
 338         profile.uarc_size = 0;
 339         if (mdev->mthca_flags & MTHCA_FLAG_SRQ)
 340                 profile.num_srq = dev_lim.max_srqs;
 341
 342         size = mthca_make_profile(mdev, &profile, &dev_lim, &init_hca);
 343         if (size < 0) {
 344                 err = size;
 345                 goto err_disable;
 346         }
 347
 348         err = mthca_INIT_HCA(mdev, &init_hca, &status);
 349         if (err) {
 350                 mthca_err(mdev, "INIT_HCA command failed, aborting.\n");
 351                 goto err_disable;
 352         }
 353         if (status) {
 354                 mthca_err(mdev, "INIT_HCA returned status 0x%02x, "
 355                           "aborting.\n", status);
 356                 err = -EINVAL;
 357                 goto err_disable;
 358         }
 359
 360         return 0;
 361
 362 err_disable:
 363         mthca_SYS_DIS(mdev, &status);
 364
 365         return err;
 366 }
 367
 368 static int mthca_load_fw(struct mthca_dev *mdev)
 369 {
 370         u8 status;
 371         int err;
 372
 373         /* FIXME: use HCA-attached memory for FW if present */
 374
 375         mdev->fw.arbel.fw_icm =
 376                 mthca_alloc_icm(mdev, mdev->fw.arbel.fw_pages,
 377                                 GFP_HIGHUSER | __GFP_NOWARN, 0);
 378         if (!mdev->fw.arbel.fw_icm) {
 379                 mthca_err(mdev, "Couldn't allocate FW area, aborting.\n");
 380                 return -ENOMEM;
 381         }
 382
 383         err = mthca_MAP_FA(mdev, mdev->fw.arbel.fw_icm, &status);
 384         if (err) {
 385                 mthca_err(mdev, "MAP_FA command failed, aborting.\n");
 386                 goto err_free;
 387         }
 388         if (status) {
 389                 mthca_err(mdev, "MAP_FA returned status 0x%02x, aborting.\n", status);
 390                 err = -EINVAL;
 391                 goto err_free;
 392         }
 393         err = mthca_RUN_FW(mdev, &status);
 394         if (err) {
 395                 mthca_err(mdev, "RUN_FW command failed, aborting.\n");
 396                 goto err_unmap_fa;
 397         }
 398         if (status) {
 399                 mthca_err(mdev, "RUN_FW returned status 0x%02x, aborting.\n", status);
 400                 err = -EINVAL;
 401                 goto err_unmap_fa;
 402         }
 403
 404         return 0;
 405
 406 err_unmap_fa:
 407         mthca_UNMAP_FA(mdev, &status);
 408
 409 err_free:
 410         mthca_free_icm(mdev, mdev->fw.arbel.fw_icm, 0);
 411         return err;
 412 }
 413
 414 static int mthca_init_icm(struct mthca_dev *mdev,
 415                           struct mthca_dev_lim *dev_lim,
 416                           struct mthca_init_hca_param *init_hca,
 417                           u64 icm_size)
 418 {
 419         u64 aux_pages;
 420         u8 status;
 421         int err;
 422
 423         err = mthca_SET_ICM_SIZE(mdev, icm_size, &aux_pages, &status);
 424         if (err) {
 425                 mthca_err(mdev, "SET_ICM_SIZE command failed, aborting.\n");
 426                 return err;
 427         }
 428         if (status) {
 429                 mthca_err(mdev, "SET_ICM_SIZE returned status 0x%02x, "
 430                           "aborting.\n", status);
 431                 return -EINVAL;
 432         }
 433
 434         mthca_dbg(mdev, "%lld KB of HCA context requires %lld KB aux memory.\n",
 435                   (unsigned long long) icm_size >> 10,
 436                   (unsigned long long) aux_pages << 2);
 437
 438         mdev->fw.arbel.aux_icm = mthca_alloc_icm(mdev, aux_pages,
 439                                                  GFP_HIGHUSER | __GFP_NOWARN, 0);
 440         if (!mdev->fw.arbel.aux_icm) {
 441                 mthca_err(mdev, "Couldn't allocate aux memory, aborting.\n");
 442                 return -ENOMEM;
 443         }
 444
 445         err = mthca_MAP_ICM_AUX(mdev, mdev->fw.arbel.aux_icm, &status);
 446         if (err) {
 447                 mthca_err(mdev, "MAP_ICM_AUX command failed, aborting.\n");
 448                 goto err_free_aux;
 449         }
 450         if (status) {
 451                 mthca_err(mdev, "MAP_ICM_AUX returned status 0x%02x, aborting.\n", status);
 452                 err = -EINVAL;
 453                 goto err_free_aux;
 454         }
 455
 456         err = mthca_map_eq_icm(mdev, init_hca->eqc_base);
 457         if (err) {
 458                 mthca_err(mdev, "Failed to map EQ context memory, aborting.\n");
 459                 goto err_unmap_aux;
 460         }
 461
 462         /* CPU writes to non-reserved MTTs, while HCA might DMA to reserved mtts */
 463         mdev->limits.reserved_mtts = ALIGN(mdev->limits.reserved_mtts * MTHCA_MTT_SEG_SIZE,
 464                                            dma_get_cache_alignment()) / MTHCA_MTT_SEG_SIZE;
 465
 466         mdev->mr_table.mtt_table = mthca_alloc_icm_table(mdev, init_hca->mtt_base,
 467                                                          MTHCA_MTT_SEG_SIZE,
 468                                                          mdev->limits.num_mtt_segs,
 469                                                          mdev->limits.reserved_mtts,
 470                                                          1, 0);
 471         if (!mdev->mr_table.mtt_table) {
 472                 mthca_err(mdev, "Failed to map MTT context memory, aborting.\n");
 473                 err = -ENOMEM;
 474                 goto err_unmap_eq;
 475         }
 476
 477         mdev->mr_table.mpt_table = mthca_alloc_icm_table(mdev, init_hca->mpt_base,
 478                                                          dev_lim->mpt_entry_sz,
 479                                                          mdev->limits.num_mpts,
 480                                                          mdev->limits.reserved_mrws,
 481                                                          1, 1);
 482         if (!mdev->mr_table.mpt_table) {
 483                 mthca_err(mdev, "Failed to map MPT context memory, aborting.\n");
 484                 err = -ENOMEM;
 485                 goto err_unmap_mtt;
 486         }
 487
 488         mdev->qp_table.qp_table = mthca_alloc_icm_table(mdev, init_hca->qpc_base,
 489                                                         dev_lim->qpc_entry_sz,
 490                                                         mdev->limits.num_qps,
 491                                                         mdev->limits.reserved_qps,
 492                                                         0, 0);
 493         if (!mdev->qp_table.qp_table) {
 494                 mthca_err(mdev, "Failed to map QP context memory, aborting.\n");
 495                 err = -ENOMEM;
 496                 goto err_unmap_mpt;
 497         }
 498
 499         mdev->qp_table.eqp_table = mthca_alloc_icm_table(mdev, init_hca->eqpc_base,
 500                                                          dev_lim->eqpc_entry_sz,
 501                                                          mdev->limits.num_qps,
 502                                                          mdev->limits.reserved_qps,
 503                                                          0, 0);
 504         if (!mdev->qp_table.eqp_table) {
 505                 mthca_err(mdev, "Failed to map EQP context memory, aborting.\n");
 506                 err = -ENOMEM;
 507                 goto err_unmap_qp;
 508         }
 509
 510         mdev->qp_table.rdb_table = mthca_alloc_icm_table(mdev, init_hca->rdb_base,
 511                                                          MTHCA_RDB_ENTRY_SIZE,
 512                                                          mdev->limits.num_qps <<
 513                                                          mdev->qp_table.rdb_shift, 0,
 514                                                          0, 0);
 515         if (!mdev->qp_table.rdb_table) {
 516                 mthca_err(mdev, "Failed to map RDB context memory, aborting\n");
 517                 err = -ENOMEM;
 518                 goto err_unmap_eqp;
 519         }
 520
 521        mdev->cq_table.table = mthca_alloc_icm_table(mdev, init_hca->cqc_base,
 522                                                     dev_lim->cqc_entry_sz,
 523                                                     mdev->limits.num_cqs,
 524                                                     mdev->limits.reserved_cqs,
 525                                                     0, 0);
 526         if (!mdev->cq_table.table) {
 527                 mthca_err(mdev, "Failed to map CQ context memory, aborting.\n");
 528                 err = -ENOMEM;
 529                 goto err_unmap_rdb;
 530         }
 531
 532         if (mdev->mthca_flags & MTHCA_FLAG_SRQ) {
 533                 mdev->srq_table.table =
 534                         mthca_alloc_icm_table(mdev, init_hca->srqc_base,
 535                                               dev_lim->srq_entry_sz,
 536                                               mdev->limits.num_srqs,
 537                                               mdev->limits.reserved_srqs,
 538                                               0, 0);
 539                 if (!mdev->srq_table.table) {
 540                         mthca_err(mdev, "Failed to map SRQ context memory, "
 541                                   "aborting.\n");
 542                         err = -ENOMEM;
 543                         goto err_unmap_cq;
 544                 }
 545         }
 546
 547         /*
 548          * It's not strictly required, but for simplicity just map the
 549          * whole multicast group table now.  The table isn't very big
 550          * and it's a lot easier than trying to track ref counts.
 551          */
 552         mdev->mcg_table.table = mthca_alloc_icm_table(mdev, init_hca->mc_base,
 553                                                       MTHCA_MGM_ENTRY_SIZE,
 554                                                       mdev->limits.num_mgms +
 555                                                       mdev->limits.num_amgms,
 556                                                       mdev->limits.num_mgms +
 557                                                       mdev->limits.num_amgms,
 558                                                       0, 0);
 559         if (!mdev->mcg_table.table) {
 560                 mthca_err(mdev, "Failed to map MCG context memory, aborting.\n");
 561                 err = -ENOMEM;
 562                 goto err_unmap_srq;
 563         }
 564
 565         return 0;
 566
 567 err_unmap_srq:
 568         if (mdev->mthca_flags & MTHCA_FLAG_SRQ)
 569                 mthca_free_icm_table(mdev, mdev->srq_table.table);
 570
 571 err_unmap_cq:
 572         mthca_free_icm_table(mdev, mdev->cq_table.table);
 573
 574 err_unmap_rdb:
 575         mthca_free_icm_table(mdev, mdev->qp_table.rdb_table);
 576
 577 err_unmap_eqp:
 578         mthca_free_icm_table(mdev, mdev->qp_table.eqp_table);
 579
 580 err_unmap_qp:
 581         mthca_free_icm_table(mdev, mdev->qp_table.qp_table);
 582
 583 err_unmap_mpt:
 584         mthca_free_icm_table(mdev, mdev->mr_table.mpt_table);
 585
 586 err_unmap_mtt:
 587         mthca_free_icm_table(mdev, mdev->mr_table.mtt_table);
 588
 589 err_unmap_eq:
 590         mthca_unmap_eq_icm(mdev);
 591
 592 err_unmap_aux:
 593         mthca_UNMAP_ICM_AUX(mdev, &status);
 594
 595 err_free_aux:
 596         mthca_free_icm(mdev, mdev->fw.arbel.aux_icm, 0);
 597
 598         return err;
 599 }
 600
 601 static void mthca_free_icms(struct mthca_dev *mdev)
 602 {
 603         u8 status;
 604
 605         mthca_free_icm_table(mdev, mdev->mcg_table.table);
 606         if (mdev->mthca_flags & MTHCA_FLAG_SRQ)
 607                 mthca_free_icm_table(mdev, mdev->srq_table.table);
 608         mthca_free_icm_table(mdev, mdev->cq_table.table);
 609         mthca_free_icm_table(mdev, mdev->qp_table.rdb_table);
 610         mthca_free_icm_table(mdev, mdev->qp_table.eqp_table);
 611         mthca_free_icm_table(mdev, mdev->qp_table.qp_table);
 612         mthca_free_icm_table(mdev, mdev->mr_table.mpt_table);
 613         mthca_free_icm_table(mdev, mdev->mr_table.mtt_table);
 614         mthca_unmap_eq_icm(mdev);
 615
 616         mthca_UNMAP_ICM_AUX(mdev, &status);
 617         mthca_free_icm(mdev, mdev->fw.arbel.aux_icm, 0);
 618 }
 619
 620 static int mthca_init_arbel(struct mthca_dev *mdev)
 621 {
 622         struct mthca_dev_lim        dev_lim;
 623         struct mthca_profile        profile;
 624         struct mthca_init_hca_param init_hca;
 625         s64 icm_size;
 626         u8 status;
 627         int err;
 628
 629         err = mthca_QUERY_FW(mdev, &status);
 630         if (err) {
 631                 mthca_err(mdev, "QUERY_FW command failed, aborting.\n");
 632                 return err;
 633         }
 634         if (status) {
 635                 mthca_err(mdev, "QUERY_FW returned status 0x%02x, "
 636                           "aborting.\n", status);
 637                 return -EINVAL;
 638         }
 639
 640         err = mthca_ENABLE_LAM(mdev, &status);
 641         if (err) {
 642                 mthca_err(mdev, "ENABLE_LAM command failed, aborting.\n");
 643                 return err;
 644         }
 645         if (status == MTHCA_CMD_STAT_LAM_NOT_PRE) {
 646                 mthca_dbg(mdev, "No HCA-attached memory (running in MemFree mode)\n");
 647                 mdev->mthca_flags |= MTHCA_FLAG_NO_LAM;
 648         } else if (status) {
 649                 mthca_err(mdev, "ENABLE_LAM returned status 0x%02x, "
 650                           "aborting.\n", status);
 651                 return -EINVAL;
 652         }
 653
 654         err = mthca_load_fw(mdev);
 655         if (err) {
 656                 mthca_err(mdev, "Failed to start FW, aborting.\n");
 657                 goto err_disable;
 658         }
 659
 660         err = mthca_dev_lim(mdev, &dev_lim);
 661         if (err) {
 662                 mthca_err(mdev, "QUERY_DEV_LIM command failed, aborting.\n");
 663                 goto err_stop_fw;
 664         }
 665
 666         profile = hca_profile;
 667         profile.num_uar  = dev_lim.uar_size / PAGE_SIZE;
 668         profile.num_udav = 0;
 669         if (mdev->mthca_flags & MTHCA_FLAG_SRQ)
 670                 profile.num_srq = dev_lim.max_srqs;
 671
 672         icm_size = mthca_make_profile(mdev, &profile, &dev_lim, &init_hca);
 673         if (icm_size < 0) {
 674                 err = icm_size;
 675                 goto err_stop_fw;
 676         }
 677
 678         err = mthca_init_icm(mdev, &dev_lim, &init_hca, icm_size);
 679         if (err)
 680                 goto err_stop_fw;
 681
 682         err = mthca_INIT_HCA(mdev, &init_hca, &status);
 683         if (err) {
 684                 mthca_err(mdev, "INIT_HCA command failed, aborting.\n");
 685                 goto err_free_icm;
 686         }
 687         if (status) {
 688                 mthca_err(mdev, "INIT_HCA returned status 0x%02x, "
 689                           "aborting.\n", status);
 690                 err = -EINVAL;
 691                 goto err_free_icm;
 692         }
 693
 694         return 0;
 695
 696 err_free_icm:
 697         mthca_free_icms(mdev);
 698
 699 err_stop_fw:
 700         mthca_UNMAP_FA(mdev, &status);
 701         mthca_free_icm(mdev, mdev->fw.arbel.fw_icm, 0);
 702
 703 err_disable:
 704         if (!(mdev->mthca_flags & MTHCA_FLAG_NO_LAM))
 705                 mthca_DISABLE_LAM(mdev, &status);
 706
 707         return err;
 708 }
 709
 710 static void mthca_close_hca(struct mthca_dev *mdev)
 711 {
 712         u8 status;
 713
 714         mthca_CLOSE_HCA(mdev, 0, &status);
 715
 716         if (mthca_is_memfree(mdev)) {
 717                 mthca_free_icms(mdev);
 718
 719                 mthca_UNMAP_FA(mdev, &status);
 720                 mthca_free_icm(mdev, mdev->fw.arbel.fw_icm, 0);
 721
 722                 if (!(mdev->mthca_flags & MTHCA_FLAG_NO_LAM))
 723                         mthca_DISABLE_LAM(mdev, &status);
 724         } else
 725                 mthca_SYS_DIS(mdev, &status);
 726 }
 727
 728 static int mthca_init_hca(struct mthca_dev *mdev)
 729 {
 730         u8 status;
 731         int err;
 732         struct mthca_adapter adapter;
 733
 734         if (mthca_is_memfree(mdev))
 735                 err = mthca_init_arbel(mdev);
 736         else
 737                 err = mthca_init_tavor(mdev);
 738
 739         if (err)
 740                 return err;
 741
 742         err = mthca_QUERY_ADAPTER(mdev, &adapter, &status);
 743         if (err) {
 744                 mthca_err(mdev, "QUERY_ADAPTER command failed, aborting.\n");
 745                 goto err_close;
 746         }
 747         if (status) {
 748                 mthca_err(mdev, "QUERY_ADAPTER returned status 0x%02x, "
 749                           "aborting.\n", status);
 750                 err = -EINVAL;
 751                 goto err_close;
 752         }
 753
 754         mdev->eq_table.inta_pin = adapter.inta_pin;
 755         if (!mthca_is_memfree(mdev))
 756                 mdev->rev_id = adapter.revision_id;
 757         memcpy(mdev->board_id, adapter.board_id, sizeof mdev->board_id);
 758
 759         return 0;
 760
 761 err_close:
 762         mthca_close_hca(mdev);
 763         return err;
 764 }
 765
 766 static int mthca_setup_hca(struct mthca_dev *dev)
 767 {
 768         int err;
 769         u8 status;
 770
 771         MTHCA_INIT_DOORBELL_LOCK(&dev->doorbell_lock);
 772
 773         err = mthca_init_uar_table(dev);
 774         if (err) {
 775                 mthca_err(dev, "Failed to initialize "
 776                           "user access region table, aborting.\n");
 777                 return err;
 778         }
 779
 780         err = mthca_uar_alloc(dev, &dev->driver_uar);
 781         if (err) {
 782                 mthca_err(dev, "Failed to allocate driver access region, "
 783                           "aborting.\n");
 784                 goto err_uar_table_free;
 785         }
 786
 787         dev->kar = ioremap(dev->driver_uar.pfn << PAGE_SHIFT, PAGE_SIZE);
 788         if (!dev->kar) {
 789                 mthca_err(dev, "Couldn't map kernel access region, "
 790                           "aborting.\n");
 791                 err = -ENOMEM;
 792                 goto err_uar_free;
 793         }
 794
 795         err = mthca_init_pd_table(dev);
 796         if (err) {
 797                 mthca_err(dev, "Failed to initialize "
 798                           "protection domain table, aborting.\n");
 799                 goto err_kar_unmap;
 800         }
 801
 802         err = mthca_init_mr_table(dev);
 803         if (err) {
 804                 mthca_err(dev, "Failed to initialize "
 805                           "memory region table, aborting.\n");
 806                 goto err_pd_table_free;
 807         }
 808
 809         err = mthca_pd_alloc(dev, 1, &dev->driver_pd);
 810         if (err) {
 811                 mthca_err(dev, "Failed to create driver PD, "
 812                           "aborting.\n");
 813                 goto err_mr_table_free;
 814         }
 815
 816         err = mthca_init_eq_table(dev);
 817         if (err) {
 818                 mthca_err(dev, "Failed to initialize "
 819                           "event queue table, aborting.\n");
 820                 goto err_pd_free;
 821         }
 822
 823         err = mthca_cmd_use_events(dev);
 824         if (err) {
 825                 mthca_err(dev, "Failed to switch to event-driven "
 826                           "firmware commands, aborting.\n");
 827                 goto err_eq_table_free;
 828         }
 829
 830         err = mthca_NOP(dev, &status);
 831         if (err || status) {
 832                 if (dev->mthca_flags & MTHCA_FLAG_MSI_X) {
 833                         mthca_warn(dev, "NOP command failed to generate interrupt "
 834                                    "(IRQ %d).\n",
 835                                    dev->eq_table.eq[MTHCA_EQ_CMD].msi_x_vector);
 836                         mthca_warn(dev, "Trying again with MSI-X disabled.\n");
 837                 } else {
 838                         mthca_err(dev, "NOP command failed to generate interrupt "
 839                                   "(IRQ %d), aborting.\n",
 840                                   dev->pdev->irq);
 841                         mthca_err(dev, "BIOS or ACPI interrupt routing problem?\n");
 842                 }
 843
 844                 goto err_cmd_poll;
 845         }
 846
 847         mthca_dbg(dev, "NOP command IRQ test passed\n");
 848
 849         err = mthca_init_cq_table(dev);
 850         if (err) {
 851                 mthca_err(dev, "Failed to initialize "
 852                           "completion queue table, aborting.\n");
 853                 goto err_cmd_poll;
 854         }
 855
 856         err = mthca_init_srq_table(dev);
 857         if (err) {
 858                 mthca_err(dev, "Failed to initialize "
 859                           "shared receive queue table, aborting.\n");
 860                 goto err_cq_table_free;
 861         }
 862
 863         err = mthca_init_qp_table(dev);
 864         if (err) {
 865                 mthca_err(dev, "Failed to initialize "
 866                           "queue pair table, aborting.\n");
 867                 goto err_srq_table_free;
 868         }
 869
 870         err = mthca_init_av_table(dev);
 871         if (err) {
 872                 mthca_err(dev, "Failed to initialize "
 873                           "address vector table, aborting.\n");
 874                 goto err_qp_table_free;
 875         }
 876
 877         err = mthca_init_mcg_table(dev);
 878         if (err) {
 879                 mthca_err(dev, "Failed to initialize "
 880                           "multicast group table, aborting.\n");
 881                 goto err_av_table_free;
 882         }
 883
 884         return 0;
 885
 886 err_av_table_free:
 887         mthca_cleanup_av_table(dev);
 888
 889 err_qp_table_free:
 890         mthca_cleanup_qp_table(dev);
 891
 892 err_srq_table_free:
 893         mthca_cleanup_srq_table(dev);
 894
 895 err_cq_table_free:
 896         mthca_cleanup_cq_table(dev);
 897
 898 err_cmd_poll:
 899         mthca_cmd_use_polling(dev);
 900
 901 err_eq_table_free:
 902         mthca_cleanup_eq_table(dev);
 903
 904 err_pd_free:
 905         mthca_pd_free(dev, &dev->driver_pd);
 906
 907 err_mr_table_free:
 908         mthca_cleanup_mr_table(dev);
 909
 910 err_pd_table_free:
 911         mthca_cleanup_pd_table(dev);
 912
 913 err_kar_unmap:
 914         iounmap(dev->kar);
 915
 916 err_uar_free:
 917         mthca_uar_free(dev, &dev->driver_uar);
 918
 919 err_uar_table_free:
 920         mthca_cleanup_uar_table(dev);
 921         return err;
 922 }
 923
 924 static int mthca_request_regions(struct pci_dev *pdev, int ddr_hidden)
 925 {
 926         int err;
 927
 928         /*
 929          * We can't just use pci_request_regions() because the MSI-X
 930          * table is right in the middle of the first BAR.  If we did
 931          * pci_request_region and grab all of the first BAR, then
 932          * setting up MSI-X would fail, since the PCI core wants to do
 933          * request_mem_region on the MSI-X vector table.
 934          *
 935          * So just request what we need right now, and request any
 936          * other regions we need when setting up EQs.
 937          */
 938         if (!request_mem_region(pci_resource_start(pdev, 0) + MTHCA_HCR_BASE,
 939                                 MTHCA_HCR_SIZE, DRV_NAME))
 940                 return -EBUSY;
 941
 942         err = pci_request_region(pdev, 2, DRV_NAME);
 943         if (err)
 944                 goto err_bar2_failed;
 945
 946         if (!ddr_hidden) {
 947                 err = pci_request_region(pdev, 4, DRV_NAME);
 948                 if (err)
 949                         goto err_bar4_failed;
 950         }
 951
 952         return 0;
 953
 954 err_bar4_failed:
 955         pci_release_region(pdev, 2);
 956
 957 err_bar2_failed:
 958         release_mem_region(pci_resource_start(pdev, 0) + MTHCA_HCR_BASE,
 959                            MTHCA_HCR_SIZE);
 960
 961         return err;
 962 }
 963
 964 static void mthca_release_regions(struct pci_dev *pdev,
 965                                   int ddr_hidden)
 966 {
 967         if (!ddr_hidden)
 968                 pci_release_region(pdev, 4);
 969
 970         pci_release_region(pdev, 2);
 971
 972         release_mem_region(pci_resource_start(pdev, 0) + MTHCA_HCR_BASE,
 973                            MTHCA_HCR_SIZE);
 974 }
 975
 976 static int mthca_enable_msi_x(struct mthca_dev *mdev)
 977 {
 978         struct msix_entry entries[3];
 979         int err;
 980
 981         entries[0].entry = 0;
 982         entries[1].entry = 1;
 983         entries[2].entry = 2;
 984
 985         err = pci_enable_msix(mdev->pdev, entries, ARRAY_SIZE(entries));
 986         if (err) {
 987                 if (err > 0)
 988                         mthca_info(mdev, "Only %d MSI-X vectors available, "
 989                                    "not using MSI-X\n", err);
 990                 return err;
 991         }
 992
 993         mdev->eq_table.eq[MTHCA_EQ_COMP ].msi_x_vector = entries[0].vector;
 994         mdev->eq_table.eq[MTHCA_EQ_ASYNC].msi_x_vector = entries[1].vector;
 995         mdev->eq_table.eq[MTHCA_EQ_CMD  ].msi_x_vector = entries[2].vector;
 996
 997         return 0;
 998 }
 999
1000 /* Types of supported HCA */
1001 enum {
1002         TAVOR,                  /* MT23108                        */
1003         ARBEL_COMPAT,           /* MT25208 in Tavor compat mode   */
1004         ARBEL_NATIVE,           /* MT25208 with extended features */
1005         SINAI                   /* MT25204 */
1006 };
1007
1008 #define MTHCA_FW_VER(major, minor, subminor) \
1009         (((u64) (major) << 32) | ((u64) (minor) << 16) | (u64) (subminor))
1010
1011 static struct {
1012         u64 latest_fw;
1013         u32 flags;
1014 } mthca_hca_table[] = {
1015         [TAVOR]        = { .latest_fw = MTHCA_FW_VER(3, 5, 0),
1016                            .flags     = 0 },
1017         [ARBEL_COMPAT] = { .latest_fw = MTHCA_FW_VER(4, 8, 200),
1018                            .flags     = MTHCA_FLAG_PCIE },
1019         [ARBEL_NATIVE] = { .latest_fw = MTHCA_FW_VER(5, 3, 0),
1020                            .flags     = MTHCA_FLAG_MEMFREE |
1021                                         MTHCA_FLAG_PCIE },
1022         [SINAI]        = { .latest_fw = MTHCA_FW_VER(1, 2, 0),
1023                            .flags     = MTHCA_FLAG_MEMFREE |
1024                                         MTHCA_FLAG_PCIE    |
1025                                         MTHCA_FLAG_SINAI_OPT }
1026 };
1027
1028 static int __mthca_init_one(struct pci_dev *pdev, int hca_type)
1029 {
1030         int ddr_hidden = 0;
1031         int err;
1032         struct mthca_dev *mdev;
1033
1034         printk(KERN_INFO PFX "Initializing %s\n",
1035                pci_name(pdev));
1036
1037         err = pci_enable_device(pdev);
1038         if (err) {
1039                 dev_err(&pdev->dev, "Cannot enable PCI device, "
1040                         "aborting.\n");
1041                 return err;
1042         }
1043
1044         /*
1045          * Check for BARs.  We expect 0: 1MB, 2: 8MB, 4: DDR (may not
1046          * be present)
1047          */
1048         if (!(pci_resource_flags(pdev, 0) & IORESOURCE_MEM) ||
1049             pci_resource_len(pdev, 0) != 1 << 20) {
1050                 dev_err(&pdev->dev, "Missing DCS, aborting.\n");
1051                 err = -ENODEV;
1052                 goto err_disable_pdev;
1053         }
1054         if (!(pci_resource_flags(pdev, 2) & IORESOURCE_MEM)) {
1055                 dev_err(&pdev->dev, "Missing UAR, aborting.\n");
1056                 err = -ENODEV;
1057                 goto err_disable_pdev;
1058         }
1059         if (!(pci_resource_flags(pdev, 4) & IORESOURCE_MEM))
1060                 ddr_hidden = 1;
1061
1062         err = mthca_request_regions(pdev, ddr_hidden);
1063         if (err) {
1064                 dev_err(&pdev->dev, "Cannot obtain PCI resources, "
1065                         "aborting.\n");
1066                 goto err_disable_pdev;
1067         }
1068
1069         pci_set_master(pdev);
1070
1071         err = pci_set_dma_mask(pdev, DMA_64BIT_MASK);
1072         if (err) {
1073                 dev_warn(&pdev->dev, "Warning: couldn't set 64-bit PCI DMA mask.\n");
1074                 err = pci_set_dma_mask(pdev, DMA_32BIT_MASK);
1075                 if (err) {
1076                         dev_err(&pdev->dev, "Can't set PCI DMA mask, aborting.\n");
1077                         goto err_free_res;
1078                 }
1079         }
1080         err = pci_set_consistent_dma_mask(pdev, DMA_64BIT_MASK);
1081         if (err) {
1082                 dev_warn(&pdev->dev, "Warning: couldn't set 64-bit "
1083                          "consistent PCI DMA mask.\n");
1084                 err = pci_set_consistent_dma_mask(pdev, DMA_32BIT_MASK);
1085                 if (err) {
1086                         dev_err(&pdev->dev, "Can't set consistent PCI DMA mask, "
1087                                 "aborting.\n");
1088                         goto err_free_res;
1089                 }
1090         }
1091
1092         mdev = (struct mthca_dev *) ib_alloc_device(sizeof *mdev);
1093         if (!mdev) {
1094                 dev_err(&pdev->dev, "Device struct alloc failed, "
1095                         "aborting.\n");
1096                 err = -ENOMEM;
1097                 goto err_free_res;
1098         }
1099
1100         mdev->pdev = pdev;
1101
1102         mdev->mthca_flags = mthca_hca_table[hca_type].flags;
1103         if (ddr_hidden)
1104                 mdev->mthca_flags |= MTHCA_FLAG_DDR_HIDDEN;
1105
1106         /*
1107          * Now reset the HCA before we touch the PCI capabilities or
1108          * attempt a firmware command, since a boot ROM may have left
1109          * the HCA in an undefined state.
1110          */
1111         err = mthca_reset(mdev);
1112         if (err) {
1113                 mthca_err(mdev, "Failed to reset HCA, aborting.\n");
1114                 goto err_free_dev;
1115         }
1116
1117         if (mthca_cmd_init(mdev)) {
1118                 mthca_err(mdev, "Failed to init command interface, aborting.\n");
1119                 goto err_free_dev;
1120         }
1121
1122         err = mthca_tune_pci(mdev);
1123         if (err)
1124                 goto err_cmd;
1125
1126         err = mthca_init_hca(mdev);
1127         if (err)
1128                 goto err_cmd;
1129
1130         if (mdev->fw_ver < mthca_hca_table[hca_type].latest_fw) {
1131                 mthca_warn(mdev, "HCA FW version %d.%d.%03d is old (%d.%d.%03d is current).\n",
1132                            (int) (mdev->fw_ver >> 32), (int) (mdev->fw_ver >> 16) & 0xffff,
1133                            (int) (mdev->fw_ver & 0xffff),
1134                            (int) (mthca_hca_table[hca_type].latest_fw >> 32),
1135                            (int) (mthca_hca_table[hca_type].latest_fw >> 16) & 0xffff,
1136                            (int) (mthca_hca_table[hca_type].latest_fw & 0xffff));
1137                 mthca_warn(mdev, "If you have problems, try updating your HCA FW.\n");
1138         }
1139
1140         if (msi_x && !mthca_enable_msi_x(mdev))
1141                 mdev->mthca_flags |= MTHCA_FLAG_MSI_X;
1142
1143         err = mthca_setup_hca(mdev);
1144         if (err == -EBUSY && (mdev->mthca_flags & MTHCA_FLAG_MSI_X)) {
1145                 if (mdev->mthca_flags & MTHCA_FLAG_MSI_X)
1146                         pci_disable_msix(pdev);
1147                 mdev->mthca_flags &= ~MTHCA_FLAG_MSI_X;
1148
1149                 err = mthca_setup_hca(mdev);
1150         }
1151
1152         if (err)
1153                 goto err_close;
1154
1155         err = mthca_register_device(mdev);
1156         if (err)
1157                 goto err_cleanup;
1158
1159         err = mthca_create_agents(mdev);
1160         if (err)
1161                 goto err_unregister;
1162
1163         pci_set_drvdata(pdev, mdev);
1164         mdev->hca_type = hca_type;
1165
1166         return 0;
1167
1168 err_unregister:
1169         mthca_unregister_device(mdev);
1170
1171 err_cleanup:
1172         mthca_cleanup_mcg_table(mdev);
1173         mthca_cleanup_av_table(mdev);
1174         mthca_cleanup_qp_table(mdev);
1175         mthca_cleanup_srq_table(mdev);
1176         mthca_cleanup_cq_table(mdev);
1177         mthca_cmd_use_polling(mdev);
1178         mthca_cleanup_eq_table(mdev);
1179
1180         mthca_pd_free(mdev, &mdev->driver_pd);
1181
1182         mthca_cleanup_mr_table(mdev);
1183         mthca_cleanup_pd_table(mdev);
1184         mthca_cleanup_uar_table(mdev);
1185
1186 err_close:
1187         if (mdev->mthca_flags & MTHCA_FLAG_MSI_X)
1188                 pci_disable_msix(pdev);
1189
1190         mthca_close_hca(mdev);
1191
1192 err_cmd:
1193         mthca_cmd_cleanup(mdev);
1194
1195 err_free_dev:
1196         ib_dealloc_device(&mdev->ib_dev);
1197
1198 err_free_res:
1199         mthca_release_regions(pdev, ddr_hidden);
1200
1201 err_disable_pdev:
1202         pci_disable_device(pdev);
1203         pci_set_drvdata(pdev, NULL);
1204         return err;
1205 }
1206
1207 static void __mthca_remove_one(struct pci_dev *pdev)
1208 {
1209         struct mthca_dev *mdev = pci_get_drvdata(pdev);
1210         u8 status;
1211         int p;
1212
1213         if (mdev) {
1214                 mthca_free_agents(mdev);
1215                 mthca_unregister_device(mdev);
1216
1217                 for (p = 1; p <= mdev->limits.num_ports; ++p)
1218                         mthca_CLOSE_IB(mdev, p, &status);
1219
1220                 mthca_cleanup_mcg_table(mdev);
1221                 mthca_cleanup_av_table(mdev);
1222                 mthca_cleanup_qp_table(mdev);
1223                 mthca_cleanup_srq_table(mdev);
1224                 mthca_cleanup_cq_table(mdev);
1225                 mthca_cmd_use_polling(mdev);
1226                 mthca_cleanup_eq_table(mdev);
1227
1228                 mthca_pd_free(mdev, &mdev->driver_pd);
1229
1230                 mthca_cleanup_mr_table(mdev);
1231                 mthca_cleanup_pd_table(mdev);
1232
1233                 iounmap(mdev->kar);
1234                 mthca_uar_free(mdev, &mdev->driver_uar);
1235                 mthca_cleanup_uar_table(mdev);
1236                 mthca_close_hca(mdev);
1237                 mthca_cmd_cleanup(mdev);
1238
1239                 if (mdev->mthca_flags & MTHCA_FLAG_MSI_X)
1240                         pci_disable_msix(pdev);
1241
1242                 ib_dealloc_device(&mdev->ib_dev);
1243                 mthca_release_regions(pdev, mdev->mthca_flags &
1244                                       MTHCA_FLAG_DDR_HIDDEN);
1245                 pci_disable_device(pdev);
1246                 pci_set_drvdata(pdev, NULL);
1247         }
1248 }
1249
1250 int __mthca_restart_one(struct pci_dev *pdev)
1251 {
1252         struct mthca_dev *mdev;
1253         int hca_type;
1254
1255         mdev = pci_get_drvdata(pdev);
1256         if (!mdev)
1257                 return -ENODEV;
1258         hca_type = mdev->hca_type;
1259         __mthca_remove_one(pdev);
1260         return __mthca_init_one(pdev, hca_type);
1261 }
1262
1263 static int __devinit mthca_init_one(struct pci_dev *pdev,
1264                                     const struct pci_device_id *id)
1265 {
1266         static int mthca_version_printed = 0;
1267         int ret;
1268
1269         mutex_lock(&mthca_device_mutex);
1270
1271         if (!mthca_version_printed) {
1272                 printk(KERN_INFO "%s", mthca_version);
1273                 ++mthca_version_printed;
1274         }
1275
1276         if (id->driver_data >= ARRAY_SIZE(mthca_hca_table)) {
1277                 printk(KERN_ERR PFX "%s has invalid driver data %lx\n",
1278                        pci_name(pdev), id->driver_data);
1279                 mutex_unlock(&mthca_device_mutex);
1280                 return -ENODEV;
1281         }
1282
1283         ret = __mthca_init_one(pdev, id->driver_data);
1284
1285         mutex_unlock(&mthca_device_mutex);
1286
1287         return ret;
1288 }
1289
1290 static void __devexit mthca_remove_one(struct pci_dev *pdev)
1291 {
1292         mutex_lock(&mthca_device_mutex);
1293         __mthca_remove_one(pdev);
1294         mutex_unlock(&mthca_device_mutex);
1295 }
1296
1297 static struct pci_device_id mthca_pci_table[] = {
1298         { PCI_DEVICE(PCI_VENDOR_ID_MELLANOX, PCI_DEVICE_ID_MELLANOX_TAVOR),
1299           .driver_data = TAVOR },
1300         { PCI_DEVICE(PCI_VENDOR_ID_TOPSPIN, PCI_DEVICE_ID_MELLANOX_TAVOR),
1301           .driver_data = TAVOR },
1302         { PCI_DEVICE(PCI_VENDOR_ID_MELLANOX, PCI_DEVICE_ID_MELLANOX_ARBEL_COMPAT),
1303           .driver_data = ARBEL_COMPAT },
1304         { PCI_DEVICE(PCI_VENDOR_ID_TOPSPIN, PCI_DEVICE_ID_MELLANOX_ARBEL_COMPAT),
1305           .driver_data = ARBEL_COMPAT },
1306         { PCI_DEVICE(PCI_VENDOR_ID_MELLANOX, PCI_DEVICE_ID_MELLANOX_ARBEL),
1307           .driver_data = ARBEL_NATIVE },
1308         { PCI_DEVICE(PCI_VENDOR_ID_TOPSPIN, PCI_DEVICE_ID_MELLANOX_ARBEL),
1309           .driver_data = ARBEL_NATIVE },
1310         { PCI_DEVICE(PCI_VENDOR_ID_MELLANOX, PCI_DEVICE_ID_MELLANOX_SINAI),
1311           .driver_data = SINAI },
1312         { PCI_DEVICE(PCI_VENDOR_ID_TOPSPIN, PCI_DEVICE_ID_MELLANOX_SINAI),
1313           .driver_data = SINAI },
1314         { PCI_DEVICE(PCI_VENDOR_ID_MELLANOX, PCI_DEVICE_ID_MELLANOX_SINAI_OLD),
1315           .driver_data = SINAI },
1316         { PCI_DEVICE(PCI_VENDOR_ID_TOPSPIN, PCI_DEVICE_ID_MELLANOX_SINAI_OLD),
1317           .driver_data = SINAI },
1318         { 0, }
1319 };
1320
1321 MODULE_DEVICE_TABLE(pci, mthca_pci_table);
1322
1323 static struct pci_driver mthca_driver = {
1324         .name           = DRV_NAME,
1325         .id_table       = mthca_pci_table,
1326         .probe          = mthca_init_one,
1327         .remove         = __devexit_p(mthca_remove_one)
1328 };
1329
1330 static void __init __mthca_check_profile_val(const char *name, int *pval,
1331                                              int pval_default)
1332 {
1333         /* value must be positive and power of 2 */
1334         int old_pval = *pval;
1335
1336         if (old_pval <= 0)
1337                 *pval = pval_default;
1338         else
1339                 *pval = roundup_pow_of_two(old_pval);
1340
1341         if (old_pval != *pval) {
1342                 printk(KERN_WARNING PFX "Invalid value %d for %s in module parameter.\n",
1343                        old_pval, name);
1344                 printk(KERN_WARNING PFX "Corrected %s to %d.\n", name, *pval);
1345         }
1346 }
1347
1348 #define mthca_check_profile_val(name, default)                          \
1349         __mthca_check_profile_val(#name, &hca_profile.name, default)
1350
1351 static void __init mthca_validate_profile(void)
1352 {
1353         mthca_check_profile_val(num_qp,            MTHCA_DEFAULT_NUM_QP);
1354         mthca_check_profile_val(rdb_per_qp,        MTHCA_DEFAULT_RDB_PER_QP);
1355         mthca_check_profile_val(num_cq,            MTHCA_DEFAULT_NUM_CQ);
1356         mthca_check_profile_val(num_mcg,           MTHCA_DEFAULT_NUM_MCG);
1357         mthca_check_profile_val(num_mpt,           MTHCA_DEFAULT_NUM_MPT);
1358         mthca_check_profile_val(num_mtt,           MTHCA_DEFAULT_NUM_MTT);
1359         mthca_check_profile_val(num_udav,          MTHCA_DEFAULT_NUM_UDAV);
1360         mthca_check_profile_val(fmr_reserved_mtts, MTHCA_DEFAULT_NUM_RESERVED_MTTS);
1361
1362         if (hca_profile.fmr_reserved_mtts >= hca_profile.num_mtt) {
1363                 printk(KERN_WARNING PFX "Invalid fmr_reserved_mtts module parameter %d.\n",
1364                        hca_profile.fmr_reserved_mtts);
1365                 printk(KERN_WARNING PFX "(Must be smaller than num_mtt %d)\n",
1366                        hca_profile.num_mtt);
1367                 hca_profile.fmr_reserved_mtts = hca_profile.num_mtt / 2;
1368                 printk(KERN_WARNING PFX "Corrected fmr_reserved_mtts to %d.\n",
1369                        hca_profile.fmr_reserved_mtts);
1370         }
1371 }
1372
1373 static int __init mthca_init(void)
1374 {
1375         int ret;
1376
1377         mthca_validate_profile();
1378
1379         ret = mthca_catas_init();
1380         if (ret)
1381                 return ret;
1382
1383         ret = pci_register_driver(&mthca_driver);
1384         if (ret < 0) {
1385                 mthca_catas_cleanup();
1386                 return ret;
1387         }
1388
1389         return 0;
1390 }
1391
1392 static void __exit mthca_cleanup(void)
1393 {
1394         pci_unregister_driver(&mthca_driver);
1395         mthca_catas_cleanup();
1396 }
1397
1398 module_init(mthca_init);
1399 module_exit(mthca_cleanup);