IB/mthca: Give reserved MTTs a separate cache line
[safe/jmp/linux-2.6] / 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  * $Id: mthca_main.c 1396 2004-12-28 04:10:27Z roland $
35  */
36
37 #include <linux/module.h>
38 #include <linux/init.h>
39 #include <linux/errno.h>
40 #include <linux/pci.h>
41 #include <linux/interrupt.h>
42
43 #include "mthca_dev.h"
44 #include "mthca_config_reg.h"
45 #include "mthca_cmd.h"
46 #include "mthca_profile.h"
47 #include "mthca_memfree.h"
48
49 MODULE_AUTHOR("Roland Dreier");
50 MODULE_DESCRIPTION("Mellanox InfiniBand HCA low-level driver");
51 MODULE_LICENSE("Dual BSD/GPL");
52 MODULE_VERSION(DRV_VERSION);
53
54 #ifdef CONFIG_INFINIBAND_MTHCA_DEBUG
55
56 int mthca_debug_level = 0;
57 module_param_named(debug_level, mthca_debug_level, int, 0644);
58 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
59
60 #endif /* CONFIG_INFINIBAND_MTHCA_DEBUG */
61
62 #ifdef CONFIG_PCI_MSI
63
64 static int msi_x = 0;
65 module_param(msi_x, int, 0444);
66 MODULE_PARM_DESC(msi_x, "attempt to use MSI-X if nonzero");
67
68 static int msi = 0;
69 module_param(msi, int, 0444);
70 MODULE_PARM_DESC(msi, "attempt to use MSI if nonzero");
71
72 #else /* CONFIG_PCI_MSI */
73
74 #define msi_x (0)
75 #define msi   (0)
76
77 #endif /* CONFIG_PCI_MSI */
78
79 static int tune_pci = 0;
80 module_param(tune_pci, int, 0444);
81 MODULE_PARM_DESC(tune_pci, "increase PCI burst from the default set by BIOS if nonzero");
82
83 DEFINE_MUTEX(mthca_device_mutex);
84
85 #define MTHCA_DEFAULT_NUM_QP            (1 << 16)
86 #define MTHCA_DEFAULT_RDB_PER_QP        (1 << 2)
87 #define MTHCA_DEFAULT_NUM_CQ            (1 << 16)
88 #define MTHCA_DEFAULT_NUM_MCG           (1 << 13)
89 #define MTHCA_DEFAULT_NUM_MPT           (1 << 17)
90 #define MTHCA_DEFAULT_NUM_MTT           (1 << 20)
91 #define MTHCA_DEFAULT_NUM_UDAV          (1 << 15)
92 #define MTHCA_DEFAULT_NUM_RESERVED_MTTS (1 << 18)
93 #define MTHCA_DEFAULT_NUM_UARC_SIZE     (1 << 18)
94
95 static struct mthca_profile hca_profile = {
96         .num_qp             = MTHCA_DEFAULT_NUM_QP,
97         .rdb_per_qp         = MTHCA_DEFAULT_RDB_PER_QP,
98         .num_cq             = MTHCA_DEFAULT_NUM_CQ,
99         .num_mcg            = MTHCA_DEFAULT_NUM_MCG,
100         .num_mpt            = MTHCA_DEFAULT_NUM_MPT,
101         .num_mtt            = MTHCA_DEFAULT_NUM_MTT,
102         .num_udav           = MTHCA_DEFAULT_NUM_UDAV,          /* Tavor only */
103         .fmr_reserved_mtts  = MTHCA_DEFAULT_NUM_RESERVED_MTTS, /* Tavor only */
104         .uarc_size          = MTHCA_DEFAULT_NUM_UARC_SIZE,     /* Arbel only */
105 };
106
107 module_param_named(num_qp, hca_profile.num_qp, int, 0444);
108 MODULE_PARM_DESC(num_qp, "maximum number of QPs per HCA");
109
110 module_param_named(rdb_per_qp, hca_profile.rdb_per_qp, int, 0444);
111 MODULE_PARM_DESC(rdb_per_qp, "number of RDB buffers per QP");
112
113 module_param_named(num_cq, hca_profile.num_cq, int, 0444);
114 MODULE_PARM_DESC(num_cq, "maximum number of CQs per HCA");
115
116 module_param_named(num_mcg, hca_profile.num_mcg, int, 0444);
117 MODULE_PARM_DESC(num_mcg, "maximum number of multicast groups per HCA");
118
119 module_param_named(num_mpt, hca_profile.num_mpt, int, 0444);
120 MODULE_PARM_DESC(num_mpt,
121                 "maximum number of memory protection table entries per HCA");
122
123 module_param_named(num_mtt, hca_profile.num_mtt, int, 0444);
124 MODULE_PARM_DESC(num_mtt,
125                  "maximum number of memory translation table segments per HCA");
126
127 module_param_named(num_udav, hca_profile.num_udav, int, 0444);
128 MODULE_PARM_DESC(num_udav, "maximum number of UD address vectors per HCA");
129
130 module_param_named(fmr_reserved_mtts, hca_profile.fmr_reserved_mtts, int, 0444);
131 MODULE_PARM_DESC(fmr_reserved_mtts,
132                  "number of memory translation table segments reserved for FMR");
133
134 static const char mthca_version[] __devinitdata =
135         DRV_NAME ": Mellanox InfiniBand HCA driver v"
136         DRV_VERSION " (" DRV_RELDATE ")\n";
137
138 static int mthca_tune_pci(struct mthca_dev *mdev)
139 {
140         int cap;
141         u16 val;
142
143         if (!tune_pci)
144                 return 0;
145
146         /* First try to max out Read Byte Count */
147         cap = pci_find_capability(mdev->pdev, PCI_CAP_ID_PCIX);
148         if (cap) {
149                 if (pci_read_config_word(mdev->pdev, cap + PCI_X_CMD, &val)) {
150                         mthca_err(mdev, "Couldn't read PCI-X command register, "
151                                   "aborting.\n");
152                         return -ENODEV;
153                 }
154                 val = (val & ~PCI_X_CMD_MAX_READ) | (3 << 2);
155                 if (pci_write_config_word(mdev->pdev, cap + PCI_X_CMD, val)) {
156                         mthca_err(mdev, "Couldn't write PCI-X command register, "
157                                   "aborting.\n");
158                         return -ENODEV;
159                 }
160         } else if (!(mdev->mthca_flags & MTHCA_FLAG_PCIE))
161                 mthca_info(mdev, "No PCI-X capability, not setting RBC.\n");
162
163         cap = pci_find_capability(mdev->pdev, PCI_CAP_ID_EXP);
164         if (cap) {
165                 if (pci_read_config_word(mdev->pdev, cap + PCI_EXP_DEVCTL, &val)) {
166                         mthca_err(mdev, "Couldn't read PCI Express device control "
167                                   "register, aborting.\n");
168                         return -ENODEV;
169                 }
170                 val = (val & ~PCI_EXP_DEVCTL_READRQ) | (5 << 12);
171                 if (pci_write_config_word(mdev->pdev, cap + PCI_EXP_DEVCTL, val)) {
172                         mthca_err(mdev, "Couldn't write PCI Express device control "
173                                   "register, aborting.\n");
174                         return -ENODEV;
175                 }
176         } else if (mdev->mthca_flags & MTHCA_FLAG_PCIE)
177                 mthca_info(mdev, "No PCI Express capability, "
178                            "not setting Max Read Request Size.\n");
179
180         return 0;
181 }
182
183 static int mthca_dev_lim(struct mthca_dev *mdev, struct mthca_dev_lim *dev_lim)
184 {
185         int err;
186         u8 status;
187
188         err = mthca_QUERY_DEV_LIM(mdev, dev_lim, &status);
189         if (err) {
190                 mthca_err(mdev, "QUERY_DEV_LIM command failed, aborting.\n");
191                 return err;
192         }
193         if (status) {
194                 mthca_err(mdev, "QUERY_DEV_LIM returned status 0x%02x, "
195                           "aborting.\n", status);
196                 return -EINVAL;
197         }
198         if (dev_lim->min_page_sz > PAGE_SIZE) {
199                 mthca_err(mdev, "HCA minimum page size of %d bigger than "
200                           "kernel PAGE_SIZE of %ld, aborting.\n",
201                           dev_lim->min_page_sz, PAGE_SIZE);
202                 return -ENODEV;
203         }
204         if (dev_lim->num_ports > MTHCA_MAX_PORTS) {
205                 mthca_err(mdev, "HCA has %d ports, but we only support %d, "
206                           "aborting.\n",
207                           dev_lim->num_ports, MTHCA_MAX_PORTS);
208                 return -ENODEV;
209         }
210
211         if (dev_lim->uar_size > pci_resource_len(mdev->pdev, 2)) {
212                 mthca_err(mdev, "HCA reported UAR size of 0x%x bigger than "
213                           "PCI resource 2 size of 0x%llx, aborting.\n",
214                           dev_lim->uar_size,
215                           (unsigned long long)pci_resource_len(mdev->pdev, 2));
216                 return -ENODEV;
217         }
218
219         mdev->limits.num_ports          = dev_lim->num_ports;
220         mdev->limits.vl_cap             = dev_lim->max_vl;
221         mdev->limits.mtu_cap            = dev_lim->max_mtu;
222         mdev->limits.gid_table_len      = dev_lim->max_gids;
223         mdev->limits.pkey_table_len     = dev_lim->max_pkeys;
224         mdev->limits.local_ca_ack_delay = dev_lim->local_ca_ack_delay;
225         mdev->limits.max_sg             = dev_lim->max_sg;
226         mdev->limits.max_wqes           = dev_lim->max_qp_sz;
227         mdev->limits.max_qp_init_rdma   = dev_lim->max_requester_per_qp;
228         mdev->limits.reserved_qps       = dev_lim->reserved_qps;
229         mdev->limits.max_srq_wqes       = dev_lim->max_srq_sz;
230         mdev->limits.reserved_srqs      = dev_lim->reserved_srqs;
231         mdev->limits.reserved_eecs      = dev_lim->reserved_eecs;
232         mdev->limits.max_desc_sz        = dev_lim->max_desc_sz;
233         mdev->limits.max_srq_sge        = mthca_max_srq_sge(mdev);
234         /*
235          * Subtract 1 from the limit because we need to allocate a
236          * spare CQE so the HCA HW can tell the difference between an
237          * empty CQ and a full CQ.
238          */
239         mdev->limits.max_cqes           = dev_lim->max_cq_sz - 1;
240         mdev->limits.reserved_cqs       = dev_lim->reserved_cqs;
241         mdev->limits.reserved_eqs       = dev_lim->reserved_eqs;
242         mdev->limits.reserved_mtts      = dev_lim->reserved_mtts;
243         mdev->limits.reserved_mrws      = dev_lim->reserved_mrws;
244         mdev->limits.reserved_uars      = dev_lim->reserved_uars;
245         mdev->limits.reserved_pds       = dev_lim->reserved_pds;
246         mdev->limits.port_width_cap     = dev_lim->max_port_width;
247         mdev->limits.page_size_cap      = ~(u32) (dev_lim->min_page_sz - 1);
248         mdev->limits.flags              = dev_lim->flags;
249         /*
250          * For old FW that doesn't return static rate support, use a
251          * value of 0x3 (only static rate values of 0 or 1 are handled),
252          * except on Sinai, where even old FW can handle static rate
253          * values of 2 and 3.
254          */
255         if (dev_lim->stat_rate_support)
256                 mdev->limits.stat_rate_support = dev_lim->stat_rate_support;
257         else if (mdev->mthca_flags & MTHCA_FLAG_SINAI_OPT)
258                 mdev->limits.stat_rate_support = 0xf;
259         else
260                 mdev->limits.stat_rate_support = 0x3;
261
262         /* IB_DEVICE_RESIZE_MAX_WR not supported by driver.
263            May be doable since hardware supports it for SRQ.
264
265            IB_DEVICE_N_NOTIFY_CQ is supported by hardware but not by driver.
266
267            IB_DEVICE_SRQ_RESIZE is supported by hardware but SRQ is not
268            supported by driver. */
269         mdev->device_cap_flags = IB_DEVICE_CHANGE_PHY_PORT |
270                 IB_DEVICE_PORT_ACTIVE_EVENT |
271                 IB_DEVICE_SYS_IMAGE_GUID |
272                 IB_DEVICE_RC_RNR_NAK_GEN;
273
274         if (dev_lim->flags & DEV_LIM_FLAG_BAD_PKEY_CNTR)
275                 mdev->device_cap_flags |= IB_DEVICE_BAD_PKEY_CNTR;
276
277         if (dev_lim->flags & DEV_LIM_FLAG_BAD_QKEY_CNTR)
278                 mdev->device_cap_flags |= IB_DEVICE_BAD_QKEY_CNTR;
279
280         if (dev_lim->flags & DEV_LIM_FLAG_RAW_MULTI)
281                 mdev->device_cap_flags |= IB_DEVICE_RAW_MULTI;
282
283         if (dev_lim->flags & DEV_LIM_FLAG_AUTO_PATH_MIG)
284                 mdev->device_cap_flags |= IB_DEVICE_AUTO_PATH_MIG;
285
286         if (dev_lim->flags & DEV_LIM_FLAG_UD_AV_PORT_ENFORCE)
287                 mdev->device_cap_flags |= IB_DEVICE_UD_AV_PORT_ENFORCE;
288
289         if (dev_lim->flags & DEV_LIM_FLAG_SRQ)
290                 mdev->mthca_flags |= MTHCA_FLAG_SRQ;
291
292         return 0;
293 }
294
295 static int mthca_init_tavor(struct mthca_dev *mdev)
296 {
297         u8 status;
298         int err;
299         struct mthca_dev_lim        dev_lim;
300         struct mthca_profile        profile;
301         struct mthca_init_hca_param init_hca;
302
303         err = mthca_SYS_EN(mdev, &status);
304         if (err) {
305                 mthca_err(mdev, "SYS_EN command failed, aborting.\n");
306                 return err;
307         }
308         if (status) {
309                 mthca_err(mdev, "SYS_EN returned status 0x%02x, "
310                           "aborting.\n", status);
311                 return -EINVAL;
312         }
313
314         err = mthca_QUERY_FW(mdev, &status);
315         if (err) {
316                 mthca_err(mdev, "QUERY_FW command failed, aborting.\n");
317                 goto err_disable;
318         }
319         if (status) {
320                 mthca_err(mdev, "QUERY_FW returned status 0x%02x, "
321                           "aborting.\n", status);
322                 err = -EINVAL;
323                 goto err_disable;
324         }
325         err = mthca_QUERY_DDR(mdev, &status);
326         if (err) {
327                 mthca_err(mdev, "QUERY_DDR command failed, aborting.\n");
328                 goto err_disable;
329         }
330         if (status) {
331                 mthca_err(mdev, "QUERY_DDR returned status 0x%02x, "
332                           "aborting.\n", status);
333                 err = -EINVAL;
334                 goto err_disable;
335         }
336
337         err = mthca_dev_lim(mdev, &dev_lim);
338         if (err) {
339                 mthca_err(mdev, "QUERY_DEV_LIM command failed, aborting.\n");
340                 goto err_disable;
341         }
342
343         profile = hca_profile;
344         profile.num_uar   = dev_lim.uar_size / PAGE_SIZE;
345         profile.uarc_size = 0;
346         if (mdev->mthca_flags & MTHCA_FLAG_SRQ)
347                 profile.num_srq = dev_lim.max_srqs;
348
349         err = mthca_make_profile(mdev, &profile, &dev_lim, &init_hca);
350         if (err < 0)
351                 goto err_disable;
352
353         err = mthca_INIT_HCA(mdev, &init_hca, &status);
354         if (err) {
355                 mthca_err(mdev, "INIT_HCA command failed, aborting.\n");
356                 goto err_disable;
357         }
358         if (status) {
359                 mthca_err(mdev, "INIT_HCA returned status 0x%02x, "
360                           "aborting.\n", status);
361                 err = -EINVAL;
362                 goto err_disable;
363         }
364
365         return 0;
366
367 err_disable:
368         mthca_SYS_DIS(mdev, &status);
369
370         return err;
371 }
372
373 static int mthca_load_fw(struct mthca_dev *mdev)
374 {
375         u8 status;
376         int err;
377
378         /* FIXME: use HCA-attached memory for FW if present */
379
380         mdev->fw.arbel.fw_icm =
381                 mthca_alloc_icm(mdev, mdev->fw.arbel.fw_pages,
382                                 GFP_HIGHUSER | __GFP_NOWARN);
383         if (!mdev->fw.arbel.fw_icm) {
384                 mthca_err(mdev, "Couldn't allocate FW area, aborting.\n");
385                 return -ENOMEM;
386         }
387
388         err = mthca_MAP_FA(mdev, mdev->fw.arbel.fw_icm, &status);
389         if (err) {
390                 mthca_err(mdev, "MAP_FA command failed, aborting.\n");
391                 goto err_free;
392         }
393         if (status) {
394                 mthca_err(mdev, "MAP_FA returned status 0x%02x, aborting.\n", status);
395                 err = -EINVAL;
396                 goto err_free;
397         }
398         err = mthca_RUN_FW(mdev, &status);
399         if (err) {
400                 mthca_err(mdev, "RUN_FW command failed, aborting.\n");
401                 goto err_unmap_fa;
402         }
403         if (status) {
404                 mthca_err(mdev, "RUN_FW returned status 0x%02x, aborting.\n", status);
405                 err = -EINVAL;
406                 goto err_unmap_fa;
407         }
408
409         return 0;
410
411 err_unmap_fa:
412         mthca_UNMAP_FA(mdev, &status);
413
414 err_free:
415         mthca_free_icm(mdev, mdev->fw.arbel.fw_icm);
416         return err;
417 }
418
419 static int mthca_init_icm(struct mthca_dev *mdev,
420                           struct mthca_dev_lim *dev_lim,
421                           struct mthca_init_hca_param *init_hca,
422                           u64 icm_size)
423 {
424         u64 aux_pages;
425         u8 status;
426         int err;
427
428         err = mthca_SET_ICM_SIZE(mdev, icm_size, &aux_pages, &status);
429         if (err) {
430                 mthca_err(mdev, "SET_ICM_SIZE command failed, aborting.\n");
431                 return err;
432         }
433         if (status) {
434                 mthca_err(mdev, "SET_ICM_SIZE returned status 0x%02x, "
435                           "aborting.\n", status);
436                 return -EINVAL;
437         }
438
439         mthca_dbg(mdev, "%lld KB of HCA context requires %lld KB aux memory.\n",
440                   (unsigned long long) icm_size >> 10,
441                   (unsigned long long) aux_pages << 2);
442
443         mdev->fw.arbel.aux_icm = mthca_alloc_icm(mdev, aux_pages,
444                                                  GFP_HIGHUSER | __GFP_NOWARN);
445         if (!mdev->fw.arbel.aux_icm) {
446                 mthca_err(mdev, "Couldn't allocate aux memory, aborting.\n");
447                 return -ENOMEM;
448         }
449
450         err = mthca_MAP_ICM_AUX(mdev, mdev->fw.arbel.aux_icm, &status);
451         if (err) {
452                 mthca_err(mdev, "MAP_ICM_AUX command failed, aborting.\n");
453                 goto err_free_aux;
454         }
455         if (status) {
456                 mthca_err(mdev, "MAP_ICM_AUX returned status 0x%02x, aborting.\n", status);
457                 err = -EINVAL;
458                 goto err_free_aux;
459         }
460
461         err = mthca_map_eq_icm(mdev, init_hca->eqc_base);
462         if (err) {
463                 mthca_err(mdev, "Failed to map EQ context memory, aborting.\n");
464                 goto err_unmap_aux;
465         }
466
467         /* CPU writes to non-reserved MTTs, while HCA might DMA to reserved mtts */
468         mdev->limits.reserved_mtts = ALIGN(mdev->limits.reserved_mtts * MTHCA_MTT_SEG_SIZE,
469                                            dma_get_cache_alignment()) / MTHCA_MTT_SEG_SIZE;
470
471         mdev->mr_table.mtt_table = mthca_alloc_icm_table(mdev, init_hca->mtt_base,
472                                                          MTHCA_MTT_SEG_SIZE,
473                                                          mdev->limits.num_mtt_segs,
474                                                          mdev->limits.reserved_mtts, 1);
475         if (!mdev->mr_table.mtt_table) {
476                 mthca_err(mdev, "Failed to map MTT context memory, aborting.\n");
477                 err = -ENOMEM;
478                 goto err_unmap_eq;
479         }
480
481         mdev->mr_table.mpt_table = mthca_alloc_icm_table(mdev, init_hca->mpt_base,
482                                                          dev_lim->mpt_entry_sz,
483                                                          mdev->limits.num_mpts,
484                                                          mdev->limits.reserved_mrws, 1);
485         if (!mdev->mr_table.mpt_table) {
486                 mthca_err(mdev, "Failed to map MPT context memory, aborting.\n");
487                 err = -ENOMEM;
488                 goto err_unmap_mtt;
489         }
490
491         mdev->qp_table.qp_table = mthca_alloc_icm_table(mdev, init_hca->qpc_base,
492                                                         dev_lim->qpc_entry_sz,
493                                                         mdev->limits.num_qps,
494                                                         mdev->limits.reserved_qps, 0);
495         if (!mdev->qp_table.qp_table) {
496                 mthca_err(mdev, "Failed to map QP context memory, aborting.\n");
497                 err = -ENOMEM;
498                 goto err_unmap_mpt;
499         }
500
501         mdev->qp_table.eqp_table = mthca_alloc_icm_table(mdev, init_hca->eqpc_base,
502                                                          dev_lim->eqpc_entry_sz,
503                                                          mdev->limits.num_qps,
504                                                          mdev->limits.reserved_qps, 0);
505         if (!mdev->qp_table.eqp_table) {
506                 mthca_err(mdev, "Failed to map EQP context memory, aborting.\n");
507                 err = -ENOMEM;
508                 goto err_unmap_qp;
509         }
510
511         mdev->qp_table.rdb_table = mthca_alloc_icm_table(mdev, init_hca->rdb_base,
512                                                          MTHCA_RDB_ENTRY_SIZE,
513                                                          mdev->limits.num_qps <<
514                                                          mdev->qp_table.rdb_shift,
515                                                          0, 0);
516         if (!mdev->qp_table.rdb_table) {
517                 mthca_err(mdev, "Failed to map RDB context memory, aborting\n");
518                 err = -ENOMEM;
519                 goto err_unmap_eqp;
520         }
521
522        mdev->cq_table.table = mthca_alloc_icm_table(mdev, init_hca->cqc_base,
523                                                     dev_lim->cqc_entry_sz,
524                                                     mdev->limits.num_cqs,
525                                                     mdev->limits.reserved_cqs, 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, 0);
538                 if (!mdev->srq_table.table) {
539                         mthca_err(mdev, "Failed to map SRQ context memory, "
540                                   "aborting.\n");
541                         err = -ENOMEM;
542                         goto err_unmap_cq;
543                 }
544         }
545
546         /*
547          * It's not strictly required, but for simplicity just map the
548          * whole multicast group table now.  The table isn't very big
549          * and it's a lot easier than trying to track ref counts.
550          */
551         mdev->mcg_table.table = mthca_alloc_icm_table(mdev, init_hca->mc_base,
552                                                       MTHCA_MGM_ENTRY_SIZE,
553                                                       mdev->limits.num_mgms +
554                                                       mdev->limits.num_amgms,
555                                                       mdev->limits.num_mgms +
556                                                       mdev->limits.num_amgms,
557                                                       0);
558         if (!mdev->mcg_table.table) {
559                 mthca_err(mdev, "Failed to map MCG context memory, aborting.\n");
560                 err = -ENOMEM;
561                 goto err_unmap_srq;
562         }
563
564         return 0;
565
566 err_unmap_srq:
567         if (mdev->mthca_flags & MTHCA_FLAG_SRQ)
568                 mthca_free_icm_table(mdev, mdev->srq_table.table);
569
570 err_unmap_cq:
571         mthca_free_icm_table(mdev, mdev->cq_table.table);
572
573 err_unmap_rdb:
574         mthca_free_icm_table(mdev, mdev->qp_table.rdb_table);
575
576 err_unmap_eqp:
577         mthca_free_icm_table(mdev, mdev->qp_table.eqp_table);
578
579 err_unmap_qp:
580         mthca_free_icm_table(mdev, mdev->qp_table.qp_table);
581
582 err_unmap_mpt:
583         mthca_free_icm_table(mdev, mdev->mr_table.mpt_table);
584
585 err_unmap_mtt:
586         mthca_free_icm_table(mdev, mdev->mr_table.mtt_table);
587
588 err_unmap_eq:
589         mthca_unmap_eq_icm(mdev);
590
591 err_unmap_aux:
592         mthca_UNMAP_ICM_AUX(mdev, &status);
593
594 err_free_aux:
595         mthca_free_icm(mdev, mdev->fw.arbel.aux_icm);
596
597         return err;
598 }
599
600 static void mthca_free_icms(struct mthca_dev *mdev)
601 {
602         u8 status;
603
604         mthca_free_icm_table(mdev, mdev->mcg_table.table);
605         if (mdev->mthca_flags & MTHCA_FLAG_SRQ)
606                 mthca_free_icm_table(mdev, mdev->srq_table.table);
607         mthca_free_icm_table(mdev, mdev->cq_table.table);
608         mthca_free_icm_table(mdev, mdev->qp_table.rdb_table);
609         mthca_free_icm_table(mdev, mdev->qp_table.eqp_table);
610         mthca_free_icm_table(mdev, mdev->qp_table.qp_table);
611         mthca_free_icm_table(mdev, mdev->mr_table.mpt_table);
612         mthca_free_icm_table(mdev, mdev->mr_table.mtt_table);
613         mthca_unmap_eq_icm(mdev);
614
615         mthca_UNMAP_ICM_AUX(mdev, &status);
616         mthca_free_icm(mdev, mdev->fw.arbel.aux_icm);
617 }
618
619 static int mthca_init_arbel(struct mthca_dev *mdev)
620 {
621         struct mthca_dev_lim        dev_lim;
622         struct mthca_profile        profile;
623         struct mthca_init_hca_param init_hca;
624         u64 icm_size;
625         u8 status;
626         int err;
627
628         err = mthca_QUERY_FW(mdev, &status);
629         if (err) {
630                 mthca_err(mdev, "QUERY_FW command failed, aborting.\n");
631                 return err;
632         }
633         if (status) {
634                 mthca_err(mdev, "QUERY_FW returned status 0x%02x, "
635                           "aborting.\n", status);
636                 return -EINVAL;
637         }
638
639         err = mthca_ENABLE_LAM(mdev, &status);
640         if (err) {
641                 mthca_err(mdev, "ENABLE_LAM command failed, aborting.\n");
642                 return err;
643         }
644         if (status == MTHCA_CMD_STAT_LAM_NOT_PRE) {
645                 mthca_dbg(mdev, "No HCA-attached memory (running in MemFree mode)\n");
646                 mdev->mthca_flags |= MTHCA_FLAG_NO_LAM;
647         } else if (status) {
648                 mthca_err(mdev, "ENABLE_LAM returned status 0x%02x, "
649                           "aborting.\n", status);
650                 return -EINVAL;
651         }
652
653         err = mthca_load_fw(mdev);
654         if (err) {
655                 mthca_err(mdev, "Failed to start FW, aborting.\n");
656                 goto err_disable;
657         }
658
659         err = mthca_dev_lim(mdev, &dev_lim);
660         if (err) {
661                 mthca_err(mdev, "QUERY_DEV_LIM command failed, aborting.\n");
662                 goto err_stop_fw;
663         }
664
665         profile = hca_profile;
666         profile.num_uar  = dev_lim.uar_size / PAGE_SIZE;
667         profile.num_udav = 0;
668         if (mdev->mthca_flags & MTHCA_FLAG_SRQ)
669                 profile.num_srq = dev_lim.max_srqs;
670
671         icm_size = mthca_make_profile(mdev, &profile, &dev_lim, &init_hca);
672         if ((int) icm_size < 0) {
673                 err = icm_size;
674                 goto err_stop_fw;
675         }
676
677         err = mthca_init_icm(mdev, &dev_lim, &init_hca, icm_size);
678         if (err)
679                 goto err_stop_fw;
680
681         err = mthca_INIT_HCA(mdev, &init_hca, &status);
682         if (err) {
683                 mthca_err(mdev, "INIT_HCA command failed, aborting.\n");
684                 goto err_free_icm;
685         }
686         if (status) {
687                 mthca_err(mdev, "INIT_HCA returned status 0x%02x, "
688                           "aborting.\n", status);
689                 err = -EINVAL;
690                 goto err_free_icm;
691         }
692
693         return 0;
694
695 err_free_icm:
696         mthca_free_icms(mdev);
697
698 err_stop_fw:
699         mthca_UNMAP_FA(mdev, &status);
700         mthca_free_icm(mdev, mdev->fw.arbel.fw_icm);
701
702 err_disable:
703         if (!(mdev->mthca_flags & MTHCA_FLAG_NO_LAM))
704                 mthca_DISABLE_LAM(mdev, &status);
705
706         return err;
707 }
708
709 static void mthca_close_hca(struct mthca_dev *mdev)
710 {
711         u8 status;
712
713         mthca_CLOSE_HCA(mdev, 0, &status);
714
715         if (mthca_is_memfree(mdev)) {
716                 mthca_free_icms(mdev);
717
718                 mthca_UNMAP_FA(mdev, &status);
719                 mthca_free_icm(mdev, mdev->fw.arbel.fw_icm);
720
721                 if (!(mdev->mthca_flags & MTHCA_FLAG_NO_LAM))
722                         mthca_DISABLE_LAM(mdev, &status);
723         } else
724                 mthca_SYS_DIS(mdev, &status);
725 }
726
727 static int mthca_init_hca(struct mthca_dev *mdev)
728 {
729         u8 status;
730         int err;
731         struct mthca_adapter adapter;
732
733         if (mthca_is_memfree(mdev))
734                 err = mthca_init_arbel(mdev);
735         else
736                 err = mthca_init_tavor(mdev);
737
738         if (err)
739                 return err;
740
741         err = mthca_QUERY_ADAPTER(mdev, &adapter, &status);
742         if (err) {
743                 mthca_err(mdev, "QUERY_ADAPTER command failed, aborting.\n");
744                 goto err_close;
745         }
746         if (status) {
747                 mthca_err(mdev, "QUERY_ADAPTER returned status 0x%02x, "
748                           "aborting.\n", status);
749                 err = -EINVAL;
750                 goto err_close;
751         }
752
753         mdev->eq_table.inta_pin = adapter.inta_pin;
754         mdev->rev_id            = adapter.revision_id;
755         memcpy(mdev->board_id, adapter.board_id, sizeof mdev->board_id);
756
757         return 0;
758
759 err_close:
760         mthca_close_hca(mdev);
761         return err;
762 }
763
764 static int mthca_setup_hca(struct mthca_dev *dev)
765 {
766         int err;
767         u8 status;
768
769         MTHCA_INIT_DOORBELL_LOCK(&dev->doorbell_lock);
770
771         err = mthca_init_uar_table(dev);
772         if (err) {
773                 mthca_err(dev, "Failed to initialize "
774                           "user access region table, aborting.\n");
775                 return err;
776         }
777
778         err = mthca_uar_alloc(dev, &dev->driver_uar);
779         if (err) {
780                 mthca_err(dev, "Failed to allocate driver access region, "
781                           "aborting.\n");
782                 goto err_uar_table_free;
783         }
784
785         dev->kar = ioremap(dev->driver_uar.pfn << PAGE_SHIFT, PAGE_SIZE);
786         if (!dev->kar) {
787                 mthca_err(dev, "Couldn't map kernel access region, "
788                           "aborting.\n");
789                 err = -ENOMEM;
790                 goto err_uar_free;
791         }
792
793         err = mthca_init_pd_table(dev);
794         if (err) {
795                 mthca_err(dev, "Failed to initialize "
796                           "protection domain table, aborting.\n");
797                 goto err_kar_unmap;
798         }
799
800         err = mthca_init_mr_table(dev);
801         if (err) {
802                 mthca_err(dev, "Failed to initialize "
803                           "memory region table, aborting.\n");
804                 goto err_pd_table_free;
805         }
806
807         err = mthca_pd_alloc(dev, 1, &dev->driver_pd);
808         if (err) {
809                 mthca_err(dev, "Failed to create driver PD, "
810                           "aborting.\n");
811                 goto err_mr_table_free;
812         }
813
814         err = mthca_init_eq_table(dev);
815         if (err) {
816                 mthca_err(dev, "Failed to initialize "
817                           "event queue table, aborting.\n");
818                 goto err_pd_free;
819         }
820
821         err = mthca_cmd_use_events(dev);
822         if (err) {
823                 mthca_err(dev, "Failed to switch to event-driven "
824                           "firmware commands, aborting.\n");
825                 goto err_eq_table_free;
826         }
827
828         err = mthca_NOP(dev, &status);
829         if (err || status) {
830                 mthca_err(dev, "NOP command failed to generate interrupt (IRQ %d), aborting.\n",
831                           dev->mthca_flags & MTHCA_FLAG_MSI_X ?
832                           dev->eq_table.eq[MTHCA_EQ_CMD].msi_x_vector :
833                           dev->pdev->irq);
834                 if (dev->mthca_flags & (MTHCA_FLAG_MSI | MTHCA_FLAG_MSI_X))
835                         mthca_err(dev, "Try again with MSI/MSI-X disabled.\n");
836                 else
837                         mthca_err(dev, "BIOS or ACPI interrupt routing problem?\n");
838
839                 goto err_cmd_poll;
840         }
841
842         mthca_dbg(dev, "NOP command IRQ test passed\n");
843
844         err = mthca_init_cq_table(dev);
845         if (err) {
846                 mthca_err(dev, "Failed to initialize "
847                           "completion queue table, aborting.\n");
848                 goto err_cmd_poll;
849         }
850
851         err = mthca_init_srq_table(dev);
852         if (err) {
853                 mthca_err(dev, "Failed to initialize "
854                           "shared receive queue table, aborting.\n");
855                 goto err_cq_table_free;
856         }
857
858         err = mthca_init_qp_table(dev);
859         if (err) {
860                 mthca_err(dev, "Failed to initialize "
861                           "queue pair table, aborting.\n");
862                 goto err_srq_table_free;
863         }
864
865         err = mthca_init_av_table(dev);
866         if (err) {
867                 mthca_err(dev, "Failed to initialize "
868                           "address vector table, aborting.\n");
869                 goto err_qp_table_free;
870         }
871
872         err = mthca_init_mcg_table(dev);
873         if (err) {
874                 mthca_err(dev, "Failed to initialize "
875                           "multicast group table, aborting.\n");
876                 goto err_av_table_free;
877         }
878
879         return 0;
880
881 err_av_table_free:
882         mthca_cleanup_av_table(dev);
883
884 err_qp_table_free:
885         mthca_cleanup_qp_table(dev);
886
887 err_srq_table_free:
888         mthca_cleanup_srq_table(dev);
889
890 err_cq_table_free:
891         mthca_cleanup_cq_table(dev);
892
893 err_cmd_poll:
894         mthca_cmd_use_polling(dev);
895
896 err_eq_table_free:
897         mthca_cleanup_eq_table(dev);
898
899 err_pd_free:
900         mthca_pd_free(dev, &dev->driver_pd);
901
902 err_mr_table_free:
903         mthca_cleanup_mr_table(dev);
904
905 err_pd_table_free:
906         mthca_cleanup_pd_table(dev);
907
908 err_kar_unmap:
909         iounmap(dev->kar);
910
911 err_uar_free:
912         mthca_uar_free(dev, &dev->driver_uar);
913
914 err_uar_table_free:
915         mthca_cleanup_uar_table(dev);
916         return err;
917 }
918
919 static int mthca_request_regions(struct pci_dev *pdev, int ddr_hidden)
920 {
921         int err;
922
923         /*
924          * We can't just use pci_request_regions() because the MSI-X
925          * table is right in the middle of the first BAR.  If we did
926          * pci_request_region and grab all of the first BAR, then
927          * setting up MSI-X would fail, since the PCI core wants to do
928          * request_mem_region on the MSI-X vector table.
929          *
930          * So just request what we need right now, and request any
931          * other regions we need when setting up EQs.
932          */
933         if (!request_mem_region(pci_resource_start(pdev, 0) + MTHCA_HCR_BASE,
934                                 MTHCA_HCR_SIZE, DRV_NAME))
935                 return -EBUSY;
936
937         err = pci_request_region(pdev, 2, DRV_NAME);
938         if (err)
939                 goto err_bar2_failed;
940
941         if (!ddr_hidden) {
942                 err = pci_request_region(pdev, 4, DRV_NAME);
943                 if (err)
944                         goto err_bar4_failed;
945         }
946
947         return 0;
948
949 err_bar4_failed:
950         pci_release_region(pdev, 2);
951
952 err_bar2_failed:
953         release_mem_region(pci_resource_start(pdev, 0) + MTHCA_HCR_BASE,
954                            MTHCA_HCR_SIZE);
955
956         return err;
957 }
958
959 static void mthca_release_regions(struct pci_dev *pdev,
960                                   int ddr_hidden)
961 {
962         if (!ddr_hidden)
963                 pci_release_region(pdev, 4);
964
965         pci_release_region(pdev, 2);
966
967         release_mem_region(pci_resource_start(pdev, 0) + MTHCA_HCR_BASE,
968                            MTHCA_HCR_SIZE);
969 }
970
971 static int mthca_enable_msi_x(struct mthca_dev *mdev)
972 {
973         struct msix_entry entries[3];
974         int err;
975
976         entries[0].entry = 0;
977         entries[1].entry = 1;
978         entries[2].entry = 2;
979
980         err = pci_enable_msix(mdev->pdev, entries, ARRAY_SIZE(entries));
981         if (err) {
982                 if (err > 0)
983                         mthca_info(mdev, "Only %d MSI-X vectors available, "
984                                    "not using MSI-X\n", err);
985                 return err;
986         }
987
988         mdev->eq_table.eq[MTHCA_EQ_COMP ].msi_x_vector = entries[0].vector;
989         mdev->eq_table.eq[MTHCA_EQ_ASYNC].msi_x_vector = entries[1].vector;
990         mdev->eq_table.eq[MTHCA_EQ_CMD  ].msi_x_vector = entries[2].vector;
991
992         return 0;
993 }
994
995 /* Types of supported HCA */
996 enum {
997         TAVOR,                  /* MT23108                        */
998         ARBEL_COMPAT,           /* MT25208 in Tavor compat mode   */
999         ARBEL_NATIVE,           /* MT25208 with extended features */
1000         SINAI                   /* MT25204 */
1001 };
1002
1003 #define MTHCA_FW_VER(major, minor, subminor) \
1004         (((u64) (major) << 32) | ((u64) (minor) << 16) | (u64) (subminor))
1005
1006 static struct {
1007         u64 latest_fw;
1008         u32 flags;
1009 } mthca_hca_table[] = {
1010         [TAVOR]        = { .latest_fw = MTHCA_FW_VER(3, 4, 0),
1011                            .flags     = 0 },
1012         [ARBEL_COMPAT] = { .latest_fw = MTHCA_FW_VER(4, 7, 600),
1013                            .flags     = MTHCA_FLAG_PCIE },
1014         [ARBEL_NATIVE] = { .latest_fw = MTHCA_FW_VER(5, 1, 400),
1015                            .flags     = MTHCA_FLAG_MEMFREE |
1016                                         MTHCA_FLAG_PCIE },
1017         [SINAI]        = { .latest_fw = MTHCA_FW_VER(1, 1, 0),
1018                            .flags     = MTHCA_FLAG_MEMFREE |
1019                                         MTHCA_FLAG_PCIE    |
1020                                         MTHCA_FLAG_SINAI_OPT }
1021 };
1022
1023 static int __mthca_init_one(struct pci_dev *pdev, int hca_type)
1024 {
1025         int ddr_hidden = 0;
1026         int err;
1027         struct mthca_dev *mdev;
1028
1029         printk(KERN_INFO PFX "Initializing %s\n",
1030                pci_name(pdev));
1031
1032         err = pci_enable_device(pdev);
1033         if (err) {
1034                 dev_err(&pdev->dev, "Cannot enable PCI device, "
1035                         "aborting.\n");
1036                 return err;
1037         }
1038
1039         /*
1040          * Check for BARs.  We expect 0: 1MB, 2: 8MB, 4: DDR (may not
1041          * be present)
1042          */
1043         if (!(pci_resource_flags(pdev, 0) & IORESOURCE_MEM) ||
1044             pci_resource_len(pdev, 0) != 1 << 20) {
1045                 dev_err(&pdev->dev, "Missing DCS, aborting.\n");
1046                 err = -ENODEV;
1047                 goto err_disable_pdev;
1048         }
1049         if (!(pci_resource_flags(pdev, 2) & IORESOURCE_MEM)) {
1050                 dev_err(&pdev->dev, "Missing UAR, aborting.\n");
1051                 err = -ENODEV;
1052                 goto err_disable_pdev;
1053         }
1054         if (!(pci_resource_flags(pdev, 4) & IORESOURCE_MEM))
1055                 ddr_hidden = 1;
1056
1057         err = mthca_request_regions(pdev, ddr_hidden);
1058         if (err) {
1059                 dev_err(&pdev->dev, "Cannot obtain PCI resources, "
1060                         "aborting.\n");
1061                 goto err_disable_pdev;
1062         }
1063
1064         pci_set_master(pdev);
1065
1066         err = pci_set_dma_mask(pdev, DMA_64BIT_MASK);
1067         if (err) {
1068                 dev_warn(&pdev->dev, "Warning: couldn't set 64-bit PCI DMA mask.\n");
1069                 err = pci_set_dma_mask(pdev, DMA_32BIT_MASK);
1070                 if (err) {
1071                         dev_err(&pdev->dev, "Can't set PCI DMA mask, aborting.\n");
1072                         goto err_free_res;
1073                 }
1074         }
1075         err = pci_set_consistent_dma_mask(pdev, DMA_64BIT_MASK);
1076         if (err) {
1077                 dev_warn(&pdev->dev, "Warning: couldn't set 64-bit "
1078                          "consistent PCI DMA mask.\n");
1079                 err = pci_set_consistent_dma_mask(pdev, DMA_32BIT_MASK);
1080                 if (err) {
1081                         dev_err(&pdev->dev, "Can't set consistent PCI DMA mask, "
1082                                 "aborting.\n");
1083                         goto err_free_res;
1084                 }
1085         }
1086
1087         mdev = (struct mthca_dev *) ib_alloc_device(sizeof *mdev);
1088         if (!mdev) {
1089                 dev_err(&pdev->dev, "Device struct alloc failed, "
1090                         "aborting.\n");
1091                 err = -ENOMEM;
1092                 goto err_free_res;
1093         }
1094
1095         mdev->pdev = pdev;
1096
1097         mdev->mthca_flags = mthca_hca_table[hca_type].flags;
1098         if (ddr_hidden)
1099                 mdev->mthca_flags |= MTHCA_FLAG_DDR_HIDDEN;
1100
1101         /*
1102          * Now reset the HCA before we touch the PCI capabilities or
1103          * attempt a firmware command, since a boot ROM may have left
1104          * the HCA in an undefined state.
1105          */
1106         err = mthca_reset(mdev);
1107         if (err) {
1108                 mthca_err(mdev, "Failed to reset HCA, aborting.\n");
1109                 goto err_free_dev;
1110         }
1111
1112         if (msi_x && !mthca_enable_msi_x(mdev))
1113                 mdev->mthca_flags |= MTHCA_FLAG_MSI_X;
1114         if (msi && !(mdev->mthca_flags & MTHCA_FLAG_MSI_X) &&
1115             !pci_enable_msi(pdev))
1116                 mdev->mthca_flags |= MTHCA_FLAG_MSI;
1117
1118         if (mthca_cmd_init(mdev)) {
1119                 mthca_err(mdev, "Failed to init command interface, aborting.\n");
1120                 goto err_free_dev;
1121         }
1122
1123         err = mthca_tune_pci(mdev);
1124         if (err)
1125                 goto err_cmd;
1126
1127         err = mthca_init_hca(mdev);
1128         if (err)
1129                 goto err_cmd;
1130
1131         if (mdev->fw_ver < mthca_hca_table[hca_type].latest_fw) {
1132                 mthca_warn(mdev, "HCA FW version %d.%d.%d is old (%d.%d.%d is current).\n",
1133                            (int) (mdev->fw_ver >> 32), (int) (mdev->fw_ver >> 16) & 0xffff,
1134                            (int) (mdev->fw_ver & 0xffff),
1135                            (int) (mthca_hca_table[hca_type].latest_fw >> 32),
1136                            (int) (mthca_hca_table[hca_type].latest_fw >> 16) & 0xffff,
1137                            (int) (mthca_hca_table[hca_type].latest_fw & 0xffff));
1138                 mthca_warn(mdev, "If you have problems, try updating your HCA FW.\n");
1139         }
1140
1141         err = mthca_setup_hca(mdev);
1142         if (err)
1143                 goto err_close;
1144
1145         err = mthca_register_device(mdev);
1146         if (err)
1147                 goto err_cleanup;
1148
1149         err = mthca_create_agents(mdev);
1150         if (err)
1151                 goto err_unregister;
1152
1153         pci_set_drvdata(pdev, mdev);
1154         mdev->hca_type = hca_type;
1155
1156         return 0;
1157
1158 err_unregister:
1159         mthca_unregister_device(mdev);
1160
1161 err_cleanup:
1162         mthca_cleanup_mcg_table(mdev);
1163         mthca_cleanup_av_table(mdev);
1164         mthca_cleanup_qp_table(mdev);
1165         mthca_cleanup_srq_table(mdev);
1166         mthca_cleanup_cq_table(mdev);
1167         mthca_cmd_use_polling(mdev);
1168         mthca_cleanup_eq_table(mdev);
1169
1170         mthca_pd_free(mdev, &mdev->driver_pd);
1171
1172         mthca_cleanup_mr_table(mdev);
1173         mthca_cleanup_pd_table(mdev);
1174         mthca_cleanup_uar_table(mdev);
1175
1176 err_close:
1177         mthca_close_hca(mdev);
1178
1179 err_cmd:
1180         mthca_cmd_cleanup(mdev);
1181
1182 err_free_dev:
1183         if (mdev->mthca_flags & MTHCA_FLAG_MSI_X)
1184                 pci_disable_msix(pdev);
1185         if (mdev->mthca_flags & MTHCA_FLAG_MSI)
1186                 pci_disable_msi(pdev);
1187
1188         ib_dealloc_device(&mdev->ib_dev);
1189
1190 err_free_res:
1191         mthca_release_regions(pdev, ddr_hidden);
1192
1193 err_disable_pdev:
1194         pci_disable_device(pdev);
1195         pci_set_drvdata(pdev, NULL);
1196         return err;
1197 }
1198
1199 static void __mthca_remove_one(struct pci_dev *pdev)
1200 {
1201         struct mthca_dev *mdev = pci_get_drvdata(pdev);
1202         u8 status;
1203         int p;
1204
1205         if (mdev) {
1206                 mthca_free_agents(mdev);
1207                 mthca_unregister_device(mdev);
1208
1209                 for (p = 1; p <= mdev->limits.num_ports; ++p)
1210                         mthca_CLOSE_IB(mdev, p, &status);
1211
1212                 mthca_cleanup_mcg_table(mdev);
1213                 mthca_cleanup_av_table(mdev);
1214                 mthca_cleanup_qp_table(mdev);
1215                 mthca_cleanup_srq_table(mdev);
1216                 mthca_cleanup_cq_table(mdev);
1217                 mthca_cmd_use_polling(mdev);
1218                 mthca_cleanup_eq_table(mdev);
1219
1220                 mthca_pd_free(mdev, &mdev->driver_pd);
1221
1222                 mthca_cleanup_mr_table(mdev);
1223                 mthca_cleanup_pd_table(mdev);
1224
1225                 iounmap(mdev->kar);
1226                 mthca_uar_free(mdev, &mdev->driver_uar);
1227                 mthca_cleanup_uar_table(mdev);
1228                 mthca_close_hca(mdev);
1229                 mthca_cmd_cleanup(mdev);
1230
1231                 if (mdev->mthca_flags & MTHCA_FLAG_MSI_X)
1232                         pci_disable_msix(pdev);
1233                 if (mdev->mthca_flags & MTHCA_FLAG_MSI)
1234                         pci_disable_msi(pdev);
1235
1236                 ib_dealloc_device(&mdev->ib_dev);
1237                 mthca_release_regions(pdev, mdev->mthca_flags &
1238                                       MTHCA_FLAG_DDR_HIDDEN);
1239                 pci_disable_device(pdev);
1240                 pci_set_drvdata(pdev, NULL);
1241         }
1242 }
1243
1244 int __mthca_restart_one(struct pci_dev *pdev)
1245 {
1246         struct mthca_dev *mdev;
1247
1248         mdev = pci_get_drvdata(pdev);
1249         if (!mdev)
1250                 return -ENODEV;
1251         __mthca_remove_one(pdev);
1252         return __mthca_init_one(pdev, mdev->hca_type);
1253 }
1254
1255 static int __devinit mthca_init_one(struct pci_dev *pdev,
1256                                     const struct pci_device_id *id)
1257 {
1258         static int mthca_version_printed = 0;
1259         int ret;
1260
1261         mutex_lock(&mthca_device_mutex);
1262
1263         if (!mthca_version_printed) {
1264                 printk(KERN_INFO "%s", mthca_version);
1265                 ++mthca_version_printed;
1266         }
1267
1268         if (id->driver_data >= ARRAY_SIZE(mthca_hca_table)) {
1269                 printk(KERN_ERR PFX "%s has invalid driver data %lx\n",
1270                        pci_name(pdev), id->driver_data);
1271                 mutex_unlock(&mthca_device_mutex);
1272                 return -ENODEV;
1273         }
1274
1275         ret = __mthca_init_one(pdev, id->driver_data);
1276
1277         mutex_unlock(&mthca_device_mutex);
1278
1279         return ret;
1280 }
1281
1282 static void __devexit mthca_remove_one(struct pci_dev *pdev)
1283 {
1284         mutex_lock(&mthca_device_mutex);
1285         __mthca_remove_one(pdev);
1286         mutex_unlock(&mthca_device_mutex);
1287 }
1288
1289 static struct pci_device_id mthca_pci_table[] = {
1290         { PCI_DEVICE(PCI_VENDOR_ID_MELLANOX, PCI_DEVICE_ID_MELLANOX_TAVOR),
1291           .driver_data = TAVOR },
1292         { PCI_DEVICE(PCI_VENDOR_ID_TOPSPIN, PCI_DEVICE_ID_MELLANOX_TAVOR),
1293           .driver_data = TAVOR },
1294         { PCI_DEVICE(PCI_VENDOR_ID_MELLANOX, PCI_DEVICE_ID_MELLANOX_ARBEL_COMPAT),
1295           .driver_data = ARBEL_COMPAT },
1296         { PCI_DEVICE(PCI_VENDOR_ID_TOPSPIN, PCI_DEVICE_ID_MELLANOX_ARBEL_COMPAT),
1297           .driver_data = ARBEL_COMPAT },
1298         { PCI_DEVICE(PCI_VENDOR_ID_MELLANOX, PCI_DEVICE_ID_MELLANOX_ARBEL),
1299           .driver_data = ARBEL_NATIVE },
1300         { PCI_DEVICE(PCI_VENDOR_ID_TOPSPIN, PCI_DEVICE_ID_MELLANOX_ARBEL),
1301           .driver_data = ARBEL_NATIVE },
1302         { PCI_DEVICE(PCI_VENDOR_ID_MELLANOX, PCI_DEVICE_ID_MELLANOX_SINAI),
1303           .driver_data = SINAI },
1304         { PCI_DEVICE(PCI_VENDOR_ID_TOPSPIN, PCI_DEVICE_ID_MELLANOX_SINAI),
1305           .driver_data = SINAI },
1306         { PCI_DEVICE(PCI_VENDOR_ID_MELLANOX, PCI_DEVICE_ID_MELLANOX_SINAI_OLD),
1307           .driver_data = SINAI },
1308         { PCI_DEVICE(PCI_VENDOR_ID_TOPSPIN, PCI_DEVICE_ID_MELLANOX_SINAI_OLD),
1309           .driver_data = SINAI },
1310         { 0, }
1311 };
1312
1313 MODULE_DEVICE_TABLE(pci, mthca_pci_table);
1314
1315 static struct pci_driver mthca_driver = {
1316         .name           = DRV_NAME,
1317         .id_table       = mthca_pci_table,
1318         .probe          = mthca_init_one,
1319         .remove         = __devexit_p(mthca_remove_one)
1320 };
1321
1322 static void __init __mthca_check_profile_val(const char *name, int *pval,
1323                                              int pval_default)
1324 {
1325         /* value must be positive and power of 2 */
1326         int old_pval = *pval;
1327
1328         if (old_pval <= 0)
1329                 *pval = pval_default;
1330         else
1331                 *pval = roundup_pow_of_two(old_pval);
1332
1333         if (old_pval != *pval) {
1334                 printk(KERN_WARNING PFX "Invalid value %d for %s in module parameter.\n",
1335                        old_pval, name);
1336                 printk(KERN_WARNING PFX "Corrected %s to %d.\n", name, *pval);
1337         }
1338 }
1339
1340 #define mthca_check_profile_val(name, default)                          \
1341         __mthca_check_profile_val(#name, &hca_profile.name, default)
1342
1343 static void __init mthca_validate_profile(void)
1344 {
1345         mthca_check_profile_val(num_qp,            MTHCA_DEFAULT_NUM_QP);
1346         mthca_check_profile_val(rdb_per_qp,        MTHCA_DEFAULT_RDB_PER_QP);
1347         mthca_check_profile_val(num_cq,            MTHCA_DEFAULT_NUM_CQ);
1348         mthca_check_profile_val(num_mcg,           MTHCA_DEFAULT_NUM_MCG);
1349         mthca_check_profile_val(num_mpt,           MTHCA_DEFAULT_NUM_MPT);
1350         mthca_check_profile_val(num_mtt,           MTHCA_DEFAULT_NUM_MTT);
1351         mthca_check_profile_val(num_udav,          MTHCA_DEFAULT_NUM_UDAV);
1352         mthca_check_profile_val(fmr_reserved_mtts, MTHCA_DEFAULT_NUM_RESERVED_MTTS);
1353
1354         if (hca_profile.fmr_reserved_mtts >= hca_profile.num_mtt) {
1355                 printk(KERN_WARNING PFX "Invalid fmr_reserved_mtts module parameter %d.\n",
1356                        hca_profile.fmr_reserved_mtts);
1357                 printk(KERN_WARNING PFX "(Must be smaller than num_mtt %d)\n",
1358                        hca_profile.num_mtt);
1359                 hca_profile.fmr_reserved_mtts = hca_profile.num_mtt / 2;
1360                 printk(KERN_WARNING PFX "Corrected fmr_reserved_mtts to %d.\n",
1361                        hca_profile.fmr_reserved_mtts);
1362         }
1363 }
1364
1365 static int __init mthca_init(void)
1366 {
1367         int ret;
1368
1369         mthca_validate_profile();
1370
1371         ret = mthca_catas_init();
1372         if (ret)
1373                 return ret;
1374
1375         ret = pci_register_driver(&mthca_driver);
1376         if (ret < 0) {
1377                 mthca_catas_cleanup();
1378                 return ret;
1379         }
1380
1381         return 0;
1382 }
1383
1384 static void __exit mthca_cleanup(void)
1385 {
1386         pci_unregister_driver(&mthca_driver);
1387         mthca_catas_cleanup();
1388 }
1389
1390 module_init(mthca_init);
1391 module_exit(mthca_cleanup);