IB/mthca: Enable MSI-X by default
[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 = 1;
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 (deprecated, use MSI-X instead)");
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, 0);
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, 0);
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, 0);
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,
475                                                          1, 0);
476         if (!mdev->mr_table.mtt_table) {
477                 mthca_err(mdev, "Failed to map MTT context memory, aborting.\n");
478                 err = -ENOMEM;
479                 goto err_unmap_eq;
480         }
481
482         mdev->mr_table.mpt_table = mthca_alloc_icm_table(mdev, init_hca->mpt_base,
483                                                          dev_lim->mpt_entry_sz,
484                                                          mdev->limits.num_mpts,
485                                                          mdev->limits.reserved_mrws,
486                                                          1, 1);
487         if (!mdev->mr_table.mpt_table) {
488                 mthca_err(mdev, "Failed to map MPT context memory, aborting.\n");
489                 err = -ENOMEM;
490                 goto err_unmap_mtt;
491         }
492
493         mdev->qp_table.qp_table = mthca_alloc_icm_table(mdev, init_hca->qpc_base,
494                                                         dev_lim->qpc_entry_sz,
495                                                         mdev->limits.num_qps,
496                                                         mdev->limits.reserved_qps,
497                                                         0, 0);
498         if (!mdev->qp_table.qp_table) {
499                 mthca_err(mdev, "Failed to map QP context memory, aborting.\n");
500                 err = -ENOMEM;
501                 goto err_unmap_mpt;
502         }
503
504         mdev->qp_table.eqp_table = mthca_alloc_icm_table(mdev, init_hca->eqpc_base,
505                                                          dev_lim->eqpc_entry_sz,
506                                                          mdev->limits.num_qps,
507                                                          mdev->limits.reserved_qps,
508                                                          0, 0);
509         if (!mdev->qp_table.eqp_table) {
510                 mthca_err(mdev, "Failed to map EQP context memory, aborting.\n");
511                 err = -ENOMEM;
512                 goto err_unmap_qp;
513         }
514
515         mdev->qp_table.rdb_table = mthca_alloc_icm_table(mdev, init_hca->rdb_base,
516                                                          MTHCA_RDB_ENTRY_SIZE,
517                                                          mdev->limits.num_qps <<
518                                                          mdev->qp_table.rdb_shift, 0,
519                                                          0, 0);
520         if (!mdev->qp_table.rdb_table) {
521                 mthca_err(mdev, "Failed to map RDB context memory, aborting\n");
522                 err = -ENOMEM;
523                 goto err_unmap_eqp;
524         }
525
526        mdev->cq_table.table = mthca_alloc_icm_table(mdev, init_hca->cqc_base,
527                                                     dev_lim->cqc_entry_sz,
528                                                     mdev->limits.num_cqs,
529                                                     mdev->limits.reserved_cqs,
530                                                     0, 0);
531         if (!mdev->cq_table.table) {
532                 mthca_err(mdev, "Failed to map CQ context memory, aborting.\n");
533                 err = -ENOMEM;
534                 goto err_unmap_rdb;
535         }
536
537         if (mdev->mthca_flags & MTHCA_FLAG_SRQ) {
538                 mdev->srq_table.table =
539                         mthca_alloc_icm_table(mdev, init_hca->srqc_base,
540                                               dev_lim->srq_entry_sz,
541                                               mdev->limits.num_srqs,
542                                               mdev->limits.reserved_srqs,
543                                               0, 0);
544                 if (!mdev->srq_table.table) {
545                         mthca_err(mdev, "Failed to map SRQ context memory, "
546                                   "aborting.\n");
547                         err = -ENOMEM;
548                         goto err_unmap_cq;
549                 }
550         }
551
552         /*
553          * It's not strictly required, but for simplicity just map the
554          * whole multicast group table now.  The table isn't very big
555          * and it's a lot easier than trying to track ref counts.
556          */
557         mdev->mcg_table.table = mthca_alloc_icm_table(mdev, init_hca->mc_base,
558                                                       MTHCA_MGM_ENTRY_SIZE,
559                                                       mdev->limits.num_mgms +
560                                                       mdev->limits.num_amgms,
561                                                       mdev->limits.num_mgms +
562                                                       mdev->limits.num_amgms,
563                                                       0, 0);
564         if (!mdev->mcg_table.table) {
565                 mthca_err(mdev, "Failed to map MCG context memory, aborting.\n");
566                 err = -ENOMEM;
567                 goto err_unmap_srq;
568         }
569
570         return 0;
571
572 err_unmap_srq:
573         if (mdev->mthca_flags & MTHCA_FLAG_SRQ)
574                 mthca_free_icm_table(mdev, mdev->srq_table.table);
575
576 err_unmap_cq:
577         mthca_free_icm_table(mdev, mdev->cq_table.table);
578
579 err_unmap_rdb:
580         mthca_free_icm_table(mdev, mdev->qp_table.rdb_table);
581
582 err_unmap_eqp:
583         mthca_free_icm_table(mdev, mdev->qp_table.eqp_table);
584
585 err_unmap_qp:
586         mthca_free_icm_table(mdev, mdev->qp_table.qp_table);
587
588 err_unmap_mpt:
589         mthca_free_icm_table(mdev, mdev->mr_table.mpt_table);
590
591 err_unmap_mtt:
592         mthca_free_icm_table(mdev, mdev->mr_table.mtt_table);
593
594 err_unmap_eq:
595         mthca_unmap_eq_icm(mdev);
596
597 err_unmap_aux:
598         mthca_UNMAP_ICM_AUX(mdev, &status);
599
600 err_free_aux:
601         mthca_free_icm(mdev, mdev->fw.arbel.aux_icm, 0);
602
603         return err;
604 }
605
606 static void mthca_free_icms(struct mthca_dev *mdev)
607 {
608         u8 status;
609
610         mthca_free_icm_table(mdev, mdev->mcg_table.table);
611         if (mdev->mthca_flags & MTHCA_FLAG_SRQ)
612                 mthca_free_icm_table(mdev, mdev->srq_table.table);
613         mthca_free_icm_table(mdev, mdev->cq_table.table);
614         mthca_free_icm_table(mdev, mdev->qp_table.rdb_table);
615         mthca_free_icm_table(mdev, mdev->qp_table.eqp_table);
616         mthca_free_icm_table(mdev, mdev->qp_table.qp_table);
617         mthca_free_icm_table(mdev, mdev->mr_table.mpt_table);
618         mthca_free_icm_table(mdev, mdev->mr_table.mtt_table);
619         mthca_unmap_eq_icm(mdev);
620
621         mthca_UNMAP_ICM_AUX(mdev, &status);
622         mthca_free_icm(mdev, mdev->fw.arbel.aux_icm, 0);
623 }
624
625 static int mthca_init_arbel(struct mthca_dev *mdev)
626 {
627         struct mthca_dev_lim        dev_lim;
628         struct mthca_profile        profile;
629         struct mthca_init_hca_param init_hca;
630         u64 icm_size;
631         u8 status;
632         int err;
633
634         err = mthca_QUERY_FW(mdev, &status);
635         if (err) {
636                 mthca_err(mdev, "QUERY_FW command failed, aborting.\n");
637                 return err;
638         }
639         if (status) {
640                 mthca_err(mdev, "QUERY_FW returned status 0x%02x, "
641                           "aborting.\n", status);
642                 return -EINVAL;
643         }
644
645         err = mthca_ENABLE_LAM(mdev, &status);
646         if (err) {
647                 mthca_err(mdev, "ENABLE_LAM command failed, aborting.\n");
648                 return err;
649         }
650         if (status == MTHCA_CMD_STAT_LAM_NOT_PRE) {
651                 mthca_dbg(mdev, "No HCA-attached memory (running in MemFree mode)\n");
652                 mdev->mthca_flags |= MTHCA_FLAG_NO_LAM;
653         } else if (status) {
654                 mthca_err(mdev, "ENABLE_LAM returned status 0x%02x, "
655                           "aborting.\n", status);
656                 return -EINVAL;
657         }
658
659         err = mthca_load_fw(mdev);
660         if (err) {
661                 mthca_err(mdev, "Failed to start FW, aborting.\n");
662                 goto err_disable;
663         }
664
665         err = mthca_dev_lim(mdev, &dev_lim);
666         if (err) {
667                 mthca_err(mdev, "QUERY_DEV_LIM command failed, aborting.\n");
668                 goto err_stop_fw;
669         }
670
671         profile = hca_profile;
672         profile.num_uar  = dev_lim.uar_size / PAGE_SIZE;
673         profile.num_udav = 0;
674         if (mdev->mthca_flags & MTHCA_FLAG_SRQ)
675                 profile.num_srq = dev_lim.max_srqs;
676
677         icm_size = mthca_make_profile(mdev, &profile, &dev_lim, &init_hca);
678         if ((int) icm_size < 0) {
679                 err = icm_size;
680                 goto err_stop_fw;
681         }
682
683         err = mthca_init_icm(mdev, &dev_lim, &init_hca, icm_size);
684         if (err)
685                 goto err_stop_fw;
686
687         err = mthca_INIT_HCA(mdev, &init_hca, &status);
688         if (err) {
689                 mthca_err(mdev, "INIT_HCA command failed, aborting.\n");
690                 goto err_free_icm;
691         }
692         if (status) {
693                 mthca_err(mdev, "INIT_HCA returned status 0x%02x, "
694                           "aborting.\n", status);
695                 err = -EINVAL;
696                 goto err_free_icm;
697         }
698
699         return 0;
700
701 err_free_icm:
702         mthca_free_icms(mdev);
703
704 err_stop_fw:
705         mthca_UNMAP_FA(mdev, &status);
706         mthca_free_icm(mdev, mdev->fw.arbel.fw_icm, 0);
707
708 err_disable:
709         if (!(mdev->mthca_flags & MTHCA_FLAG_NO_LAM))
710                 mthca_DISABLE_LAM(mdev, &status);
711
712         return err;
713 }
714
715 static void mthca_close_hca(struct mthca_dev *mdev)
716 {
717         u8 status;
718
719         mthca_CLOSE_HCA(mdev, 0, &status);
720
721         if (mthca_is_memfree(mdev)) {
722                 mthca_free_icms(mdev);
723
724                 mthca_UNMAP_FA(mdev, &status);
725                 mthca_free_icm(mdev, mdev->fw.arbel.fw_icm, 0);
726
727                 if (!(mdev->mthca_flags & MTHCA_FLAG_NO_LAM))
728                         mthca_DISABLE_LAM(mdev, &status);
729         } else
730                 mthca_SYS_DIS(mdev, &status);
731 }
732
733 static int mthca_init_hca(struct mthca_dev *mdev)
734 {
735         u8 status;
736         int err;
737         struct mthca_adapter adapter;
738
739         if (mthca_is_memfree(mdev))
740                 err = mthca_init_arbel(mdev);
741         else
742                 err = mthca_init_tavor(mdev);
743
744         if (err)
745                 return err;
746
747         err = mthca_QUERY_ADAPTER(mdev, &adapter, &status);
748         if (err) {
749                 mthca_err(mdev, "QUERY_ADAPTER command failed, aborting.\n");
750                 goto err_close;
751         }
752         if (status) {
753                 mthca_err(mdev, "QUERY_ADAPTER returned status 0x%02x, "
754                           "aborting.\n", status);
755                 err = -EINVAL;
756                 goto err_close;
757         }
758
759         mdev->eq_table.inta_pin = adapter.inta_pin;
760         mdev->rev_id            = adapter.revision_id;
761         memcpy(mdev->board_id, adapter.board_id, sizeof mdev->board_id);
762
763         return 0;
764
765 err_close:
766         mthca_close_hca(mdev);
767         return err;
768 }
769
770 static int mthca_setup_hca(struct mthca_dev *dev)
771 {
772         int err;
773         u8 status;
774
775         MTHCA_INIT_DOORBELL_LOCK(&dev->doorbell_lock);
776
777         err = mthca_init_uar_table(dev);
778         if (err) {
779                 mthca_err(dev, "Failed to initialize "
780                           "user access region table, aborting.\n");
781                 return err;
782         }
783
784         err = mthca_uar_alloc(dev, &dev->driver_uar);
785         if (err) {
786                 mthca_err(dev, "Failed to allocate driver access region, "
787                           "aborting.\n");
788                 goto err_uar_table_free;
789         }
790
791         dev->kar = ioremap(dev->driver_uar.pfn << PAGE_SHIFT, PAGE_SIZE);
792         if (!dev->kar) {
793                 mthca_err(dev, "Couldn't map kernel access region, "
794                           "aborting.\n");
795                 err = -ENOMEM;
796                 goto err_uar_free;
797         }
798
799         err = mthca_init_pd_table(dev);
800         if (err) {
801                 mthca_err(dev, "Failed to initialize "
802                           "protection domain table, aborting.\n");
803                 goto err_kar_unmap;
804         }
805
806         err = mthca_init_mr_table(dev);
807         if (err) {
808                 mthca_err(dev, "Failed to initialize "
809                           "memory region table, aborting.\n");
810                 goto err_pd_table_free;
811         }
812
813         err = mthca_pd_alloc(dev, 1, &dev->driver_pd);
814         if (err) {
815                 mthca_err(dev, "Failed to create driver PD, "
816                           "aborting.\n");
817                 goto err_mr_table_free;
818         }
819
820         err = mthca_init_eq_table(dev);
821         if (err) {
822                 mthca_err(dev, "Failed to initialize "
823                           "event queue table, aborting.\n");
824                 goto err_pd_free;
825         }
826
827         err = mthca_cmd_use_events(dev);
828         if (err) {
829                 mthca_err(dev, "Failed to switch to event-driven "
830                           "firmware commands, aborting.\n");
831                 goto err_eq_table_free;
832         }
833
834         err = mthca_NOP(dev, &status);
835         if (err || status) {
836                 if (dev->mthca_flags & (MTHCA_FLAG_MSI | MTHCA_FLAG_MSI_X)) {
837                         mthca_warn(dev, "NOP command failed to generate interrupt "
838                                    "(IRQ %d).\n",
839                                    dev->mthca_flags & MTHCA_FLAG_MSI_X ?
840                                    dev->eq_table.eq[MTHCA_EQ_CMD].msi_x_vector :
841                                    dev->pdev->irq);
842                         mthca_warn(dev, "Trying again with MSI/MSI-X disabled.\n");
843                 } else {
844                         mthca_err(dev, "NOP command failed to generate interrupt "
845                                   "(IRQ %d), aborting.\n",
846                                   dev->pdev->irq);
847                         mthca_err(dev, "BIOS or ACPI interrupt routing problem?\n");
848                 }
849
850                 goto err_cmd_poll;
851         }
852
853         mthca_dbg(dev, "NOP command IRQ test passed\n");
854
855         err = mthca_init_cq_table(dev);
856         if (err) {
857                 mthca_err(dev, "Failed to initialize "
858                           "completion queue table, aborting.\n");
859                 goto err_cmd_poll;
860         }
861
862         err = mthca_init_srq_table(dev);
863         if (err) {
864                 mthca_err(dev, "Failed to initialize "
865                           "shared receive queue table, aborting.\n");
866                 goto err_cq_table_free;
867         }
868
869         err = mthca_init_qp_table(dev);
870         if (err) {
871                 mthca_err(dev, "Failed to initialize "
872                           "queue pair table, aborting.\n");
873                 goto err_srq_table_free;
874         }
875
876         err = mthca_init_av_table(dev);
877         if (err) {
878                 mthca_err(dev, "Failed to initialize "
879                           "address vector table, aborting.\n");
880                 goto err_qp_table_free;
881         }
882
883         err = mthca_init_mcg_table(dev);
884         if (err) {
885                 mthca_err(dev, "Failed to initialize "
886                           "multicast group table, aborting.\n");
887                 goto err_av_table_free;
888         }
889
890         return 0;
891
892 err_av_table_free:
893         mthca_cleanup_av_table(dev);
894
895 err_qp_table_free:
896         mthca_cleanup_qp_table(dev);
897
898 err_srq_table_free:
899         mthca_cleanup_srq_table(dev);
900
901 err_cq_table_free:
902         mthca_cleanup_cq_table(dev);
903
904 err_cmd_poll:
905         mthca_cmd_use_polling(dev);
906
907 err_eq_table_free:
908         mthca_cleanup_eq_table(dev);
909
910 err_pd_free:
911         mthca_pd_free(dev, &dev->driver_pd);
912
913 err_mr_table_free:
914         mthca_cleanup_mr_table(dev);
915
916 err_pd_table_free:
917         mthca_cleanup_pd_table(dev);
918
919 err_kar_unmap:
920         iounmap(dev->kar);
921
922 err_uar_free:
923         mthca_uar_free(dev, &dev->driver_uar);
924
925 err_uar_table_free:
926         mthca_cleanup_uar_table(dev);
927         return err;
928 }
929
930 static int mthca_request_regions(struct pci_dev *pdev, int ddr_hidden)
931 {
932         int err;
933
934         /*
935          * We can't just use pci_request_regions() because the MSI-X
936          * table is right in the middle of the first BAR.  If we did
937          * pci_request_region and grab all of the first BAR, then
938          * setting up MSI-X would fail, since the PCI core wants to do
939          * request_mem_region on the MSI-X vector table.
940          *
941          * So just request what we need right now, and request any
942          * other regions we need when setting up EQs.
943          */
944         if (!request_mem_region(pci_resource_start(pdev, 0) + MTHCA_HCR_BASE,
945                                 MTHCA_HCR_SIZE, DRV_NAME))
946                 return -EBUSY;
947
948         err = pci_request_region(pdev, 2, DRV_NAME);
949         if (err)
950                 goto err_bar2_failed;
951
952         if (!ddr_hidden) {
953                 err = pci_request_region(pdev, 4, DRV_NAME);
954                 if (err)
955                         goto err_bar4_failed;
956         }
957
958         return 0;
959
960 err_bar4_failed:
961         pci_release_region(pdev, 2);
962
963 err_bar2_failed:
964         release_mem_region(pci_resource_start(pdev, 0) + MTHCA_HCR_BASE,
965                            MTHCA_HCR_SIZE);
966
967         return err;
968 }
969
970 static void mthca_release_regions(struct pci_dev *pdev,
971                                   int ddr_hidden)
972 {
973         if (!ddr_hidden)
974                 pci_release_region(pdev, 4);
975
976         pci_release_region(pdev, 2);
977
978         release_mem_region(pci_resource_start(pdev, 0) + MTHCA_HCR_BASE,
979                            MTHCA_HCR_SIZE);
980 }
981
982 static int mthca_enable_msi_x(struct mthca_dev *mdev)
983 {
984         struct msix_entry entries[3];
985         int err;
986
987         entries[0].entry = 0;
988         entries[1].entry = 1;
989         entries[2].entry = 2;
990
991         err = pci_enable_msix(mdev->pdev, entries, ARRAY_SIZE(entries));
992         if (err) {
993                 if (err > 0)
994                         mthca_info(mdev, "Only %d MSI-X vectors available, "
995                                    "not using MSI-X\n", err);
996                 return err;
997         }
998
999         mdev->eq_table.eq[MTHCA_EQ_COMP ].msi_x_vector = entries[0].vector;
1000         mdev->eq_table.eq[MTHCA_EQ_ASYNC].msi_x_vector = entries[1].vector;
1001         mdev->eq_table.eq[MTHCA_EQ_CMD  ].msi_x_vector = entries[2].vector;
1002
1003         return 0;
1004 }
1005
1006 /* Types of supported HCA */
1007 enum {
1008         TAVOR,                  /* MT23108                        */
1009         ARBEL_COMPAT,           /* MT25208 in Tavor compat mode   */
1010         ARBEL_NATIVE,           /* MT25208 with extended features */
1011         SINAI                   /* MT25204 */
1012 };
1013
1014 #define MTHCA_FW_VER(major, minor, subminor) \
1015         (((u64) (major) << 32) | ((u64) (minor) << 16) | (u64) (subminor))
1016
1017 static struct {
1018         u64 latest_fw;
1019         u32 flags;
1020 } mthca_hca_table[] = {
1021         [TAVOR]        = { .latest_fw = MTHCA_FW_VER(3, 5, 0),
1022                            .flags     = 0 },
1023         [ARBEL_COMPAT] = { .latest_fw = MTHCA_FW_VER(4, 8, 200),
1024                            .flags     = MTHCA_FLAG_PCIE },
1025         [ARBEL_NATIVE] = { .latest_fw = MTHCA_FW_VER(5, 2, 0),
1026                            .flags     = MTHCA_FLAG_MEMFREE |
1027                                         MTHCA_FLAG_PCIE },
1028         [SINAI]        = { .latest_fw = MTHCA_FW_VER(1, 2, 0),
1029                            .flags     = MTHCA_FLAG_MEMFREE |
1030                                         MTHCA_FLAG_PCIE    |
1031                                         MTHCA_FLAG_SINAI_OPT }
1032 };
1033
1034 static int __mthca_init_one(struct pci_dev *pdev, int hca_type)
1035 {
1036         int ddr_hidden = 0;
1037         int err;
1038         struct mthca_dev *mdev;
1039
1040         printk(KERN_INFO PFX "Initializing %s\n",
1041                pci_name(pdev));
1042
1043         err = pci_enable_device(pdev);
1044         if (err) {
1045                 dev_err(&pdev->dev, "Cannot enable PCI device, "
1046                         "aborting.\n");
1047                 return err;
1048         }
1049
1050         /*
1051          * Check for BARs.  We expect 0: 1MB, 2: 8MB, 4: DDR (may not
1052          * be present)
1053          */
1054         if (!(pci_resource_flags(pdev, 0) & IORESOURCE_MEM) ||
1055             pci_resource_len(pdev, 0) != 1 << 20) {
1056                 dev_err(&pdev->dev, "Missing DCS, aborting.\n");
1057                 err = -ENODEV;
1058                 goto err_disable_pdev;
1059         }
1060         if (!(pci_resource_flags(pdev, 2) & IORESOURCE_MEM)) {
1061                 dev_err(&pdev->dev, "Missing UAR, aborting.\n");
1062                 err = -ENODEV;
1063                 goto err_disable_pdev;
1064         }
1065         if (!(pci_resource_flags(pdev, 4) & IORESOURCE_MEM))
1066                 ddr_hidden = 1;
1067
1068         err = mthca_request_regions(pdev, ddr_hidden);
1069         if (err) {
1070                 dev_err(&pdev->dev, "Cannot obtain PCI resources, "
1071                         "aborting.\n");
1072                 goto err_disable_pdev;
1073         }
1074
1075         pci_set_master(pdev);
1076
1077         err = pci_set_dma_mask(pdev, DMA_64BIT_MASK);
1078         if (err) {
1079                 dev_warn(&pdev->dev, "Warning: couldn't set 64-bit PCI DMA mask.\n");
1080                 err = pci_set_dma_mask(pdev, DMA_32BIT_MASK);
1081                 if (err) {
1082                         dev_err(&pdev->dev, "Can't set PCI DMA mask, aborting.\n");
1083                         goto err_free_res;
1084                 }
1085         }
1086         err = pci_set_consistent_dma_mask(pdev, DMA_64BIT_MASK);
1087         if (err) {
1088                 dev_warn(&pdev->dev, "Warning: couldn't set 64-bit "
1089                          "consistent PCI DMA mask.\n");
1090                 err = pci_set_consistent_dma_mask(pdev, DMA_32BIT_MASK);
1091                 if (err) {
1092                         dev_err(&pdev->dev, "Can't set consistent PCI DMA mask, "
1093                                 "aborting.\n");
1094                         goto err_free_res;
1095                 }
1096         }
1097
1098         mdev = (struct mthca_dev *) ib_alloc_device(sizeof *mdev);
1099         if (!mdev) {
1100                 dev_err(&pdev->dev, "Device struct alloc failed, "
1101                         "aborting.\n");
1102                 err = -ENOMEM;
1103                 goto err_free_res;
1104         }
1105
1106         mdev->pdev = pdev;
1107
1108         mdev->mthca_flags = mthca_hca_table[hca_type].flags;
1109         if (ddr_hidden)
1110                 mdev->mthca_flags |= MTHCA_FLAG_DDR_HIDDEN;
1111
1112         /*
1113          * Now reset the HCA before we touch the PCI capabilities or
1114          * attempt a firmware command, since a boot ROM may have left
1115          * the HCA in an undefined state.
1116          */
1117         err = mthca_reset(mdev);
1118         if (err) {
1119                 mthca_err(mdev, "Failed to reset HCA, aborting.\n");
1120                 goto err_free_dev;
1121         }
1122
1123         if (mthca_cmd_init(mdev)) {
1124                 mthca_err(mdev, "Failed to init command interface, aborting.\n");
1125                 goto err_free_dev;
1126         }
1127
1128         err = mthca_tune_pci(mdev);
1129         if (err)
1130                 goto err_cmd;
1131
1132         err = mthca_init_hca(mdev);
1133         if (err)
1134                 goto err_cmd;
1135
1136         if (mdev->fw_ver < mthca_hca_table[hca_type].latest_fw) {
1137                 mthca_warn(mdev, "HCA FW version %d.%d.%03d is old (%d.%d.%03d is current).\n",
1138                            (int) (mdev->fw_ver >> 32), (int) (mdev->fw_ver >> 16) & 0xffff,
1139                            (int) (mdev->fw_ver & 0xffff),
1140                            (int) (mthca_hca_table[hca_type].latest_fw >> 32),
1141                            (int) (mthca_hca_table[hca_type].latest_fw >> 16) & 0xffff,
1142                            (int) (mthca_hca_table[hca_type].latest_fw & 0xffff));
1143                 mthca_warn(mdev, "If you have problems, try updating your HCA FW.\n");
1144         }
1145
1146         if (msi_x && !mthca_enable_msi_x(mdev))
1147                 mdev->mthca_flags |= MTHCA_FLAG_MSI_X;
1148         else if (msi) {
1149                 static int warned;
1150
1151                 if (!warned) {
1152                         printk(KERN_WARNING PFX "WARNING: MSI support will be "
1153                                "removed from the ib_mthca driver in January 2008.\n");
1154                         printk(KERN_WARNING "    If you are using MSI and cannot "
1155                                "switch to MSI-X, please tell "
1156                                "<general@lists.openfabrics.org>.\n");
1157                         ++warned;
1158                 }
1159
1160                 if (!pci_enable_msi(pdev))
1161                         mdev->mthca_flags |= MTHCA_FLAG_MSI;
1162         }
1163
1164         err = mthca_setup_hca(mdev);
1165         if (err == -EBUSY && (mdev->mthca_flags & (MTHCA_FLAG_MSI | MTHCA_FLAG_MSI_X))) {
1166                 if (mdev->mthca_flags & MTHCA_FLAG_MSI_X)
1167                         pci_disable_msix(pdev);
1168                 if (mdev->mthca_flags & MTHCA_FLAG_MSI)
1169                         pci_disable_msi(pdev);
1170                 mdev->mthca_flags &= ~(MTHCA_FLAG_MSI_X | MTHCA_FLAG_MSI);
1171
1172                 err = mthca_setup_hca(mdev);
1173         }
1174
1175         if (err)
1176                 goto err_close;
1177
1178         err = mthca_register_device(mdev);
1179         if (err)
1180                 goto err_cleanup;
1181
1182         err = mthca_create_agents(mdev);
1183         if (err)
1184                 goto err_unregister;
1185
1186         pci_set_drvdata(pdev, mdev);
1187         mdev->hca_type = hca_type;
1188
1189         return 0;
1190
1191 err_unregister:
1192         mthca_unregister_device(mdev);
1193
1194 err_cleanup:
1195         mthca_cleanup_mcg_table(mdev);
1196         mthca_cleanup_av_table(mdev);
1197         mthca_cleanup_qp_table(mdev);
1198         mthca_cleanup_srq_table(mdev);
1199         mthca_cleanup_cq_table(mdev);
1200         mthca_cmd_use_polling(mdev);
1201         mthca_cleanup_eq_table(mdev);
1202
1203         mthca_pd_free(mdev, &mdev->driver_pd);
1204
1205         mthca_cleanup_mr_table(mdev);
1206         mthca_cleanup_pd_table(mdev);
1207         mthca_cleanup_uar_table(mdev);
1208
1209 err_close:
1210         if (mdev->mthca_flags & MTHCA_FLAG_MSI_X)
1211                 pci_disable_msix(pdev);
1212         if (mdev->mthca_flags & MTHCA_FLAG_MSI)
1213                 pci_disable_msi(pdev);
1214
1215         mthca_close_hca(mdev);
1216
1217 err_cmd:
1218         mthca_cmd_cleanup(mdev);
1219
1220 err_free_dev:
1221         ib_dealloc_device(&mdev->ib_dev);
1222
1223 err_free_res:
1224         mthca_release_regions(pdev, ddr_hidden);
1225
1226 err_disable_pdev:
1227         pci_disable_device(pdev);
1228         pci_set_drvdata(pdev, NULL);
1229         return err;
1230 }
1231
1232 static void __mthca_remove_one(struct pci_dev *pdev)
1233 {
1234         struct mthca_dev *mdev = pci_get_drvdata(pdev);
1235         u8 status;
1236         int p;
1237
1238         if (mdev) {
1239                 mthca_free_agents(mdev);
1240                 mthca_unregister_device(mdev);
1241
1242                 for (p = 1; p <= mdev->limits.num_ports; ++p)
1243                         mthca_CLOSE_IB(mdev, p, &status);
1244
1245                 mthca_cleanup_mcg_table(mdev);
1246                 mthca_cleanup_av_table(mdev);
1247                 mthca_cleanup_qp_table(mdev);
1248                 mthca_cleanup_srq_table(mdev);
1249                 mthca_cleanup_cq_table(mdev);
1250                 mthca_cmd_use_polling(mdev);
1251                 mthca_cleanup_eq_table(mdev);
1252
1253                 mthca_pd_free(mdev, &mdev->driver_pd);
1254
1255                 mthca_cleanup_mr_table(mdev);
1256                 mthca_cleanup_pd_table(mdev);
1257
1258                 iounmap(mdev->kar);
1259                 mthca_uar_free(mdev, &mdev->driver_uar);
1260                 mthca_cleanup_uar_table(mdev);
1261                 mthca_close_hca(mdev);
1262                 mthca_cmd_cleanup(mdev);
1263
1264                 if (mdev->mthca_flags & MTHCA_FLAG_MSI_X)
1265                         pci_disable_msix(pdev);
1266                 if (mdev->mthca_flags & MTHCA_FLAG_MSI)
1267                         pci_disable_msi(pdev);
1268
1269                 ib_dealloc_device(&mdev->ib_dev);
1270                 mthca_release_regions(pdev, mdev->mthca_flags &
1271                                       MTHCA_FLAG_DDR_HIDDEN);
1272                 pci_disable_device(pdev);
1273                 pci_set_drvdata(pdev, NULL);
1274         }
1275 }
1276
1277 int __mthca_restart_one(struct pci_dev *pdev)
1278 {
1279         struct mthca_dev *mdev;
1280         int hca_type;
1281
1282         mdev = pci_get_drvdata(pdev);
1283         if (!mdev)
1284                 return -ENODEV;
1285         hca_type = mdev->hca_type;
1286         __mthca_remove_one(pdev);
1287         return __mthca_init_one(pdev, hca_type);
1288 }
1289
1290 static int __devinit mthca_init_one(struct pci_dev *pdev,
1291                                     const struct pci_device_id *id)
1292 {
1293         static int mthca_version_printed = 0;
1294         int ret;
1295
1296         mutex_lock(&mthca_device_mutex);
1297
1298         if (!mthca_version_printed) {
1299                 printk(KERN_INFO "%s", mthca_version);
1300                 ++mthca_version_printed;
1301         }
1302
1303         if (id->driver_data >= ARRAY_SIZE(mthca_hca_table)) {
1304                 printk(KERN_ERR PFX "%s has invalid driver data %lx\n",
1305                        pci_name(pdev), id->driver_data);
1306                 mutex_unlock(&mthca_device_mutex);
1307                 return -ENODEV;
1308         }
1309
1310         ret = __mthca_init_one(pdev, id->driver_data);
1311
1312         mutex_unlock(&mthca_device_mutex);
1313
1314         return ret;
1315 }
1316
1317 static void __devexit mthca_remove_one(struct pci_dev *pdev)
1318 {
1319         mutex_lock(&mthca_device_mutex);
1320         __mthca_remove_one(pdev);
1321         mutex_unlock(&mthca_device_mutex);
1322 }
1323
1324 static struct pci_device_id mthca_pci_table[] = {
1325         { PCI_DEVICE(PCI_VENDOR_ID_MELLANOX, PCI_DEVICE_ID_MELLANOX_TAVOR),
1326           .driver_data = TAVOR },
1327         { PCI_DEVICE(PCI_VENDOR_ID_TOPSPIN, PCI_DEVICE_ID_MELLANOX_TAVOR),
1328           .driver_data = TAVOR },
1329         { PCI_DEVICE(PCI_VENDOR_ID_MELLANOX, PCI_DEVICE_ID_MELLANOX_ARBEL_COMPAT),
1330           .driver_data = ARBEL_COMPAT },
1331         { PCI_DEVICE(PCI_VENDOR_ID_TOPSPIN, PCI_DEVICE_ID_MELLANOX_ARBEL_COMPAT),
1332           .driver_data = ARBEL_COMPAT },
1333         { PCI_DEVICE(PCI_VENDOR_ID_MELLANOX, PCI_DEVICE_ID_MELLANOX_ARBEL),
1334           .driver_data = ARBEL_NATIVE },
1335         { PCI_DEVICE(PCI_VENDOR_ID_TOPSPIN, PCI_DEVICE_ID_MELLANOX_ARBEL),
1336           .driver_data = ARBEL_NATIVE },
1337         { PCI_DEVICE(PCI_VENDOR_ID_MELLANOX, PCI_DEVICE_ID_MELLANOX_SINAI),
1338           .driver_data = SINAI },
1339         { PCI_DEVICE(PCI_VENDOR_ID_TOPSPIN, PCI_DEVICE_ID_MELLANOX_SINAI),
1340           .driver_data = SINAI },
1341         { PCI_DEVICE(PCI_VENDOR_ID_MELLANOX, PCI_DEVICE_ID_MELLANOX_SINAI_OLD),
1342           .driver_data = SINAI },
1343         { PCI_DEVICE(PCI_VENDOR_ID_TOPSPIN, PCI_DEVICE_ID_MELLANOX_SINAI_OLD),
1344           .driver_data = SINAI },
1345         { 0, }
1346 };
1347
1348 MODULE_DEVICE_TABLE(pci, mthca_pci_table);
1349
1350 static struct pci_driver mthca_driver = {
1351         .name           = DRV_NAME,
1352         .id_table       = mthca_pci_table,
1353         .probe          = mthca_init_one,
1354         .remove         = __devexit_p(mthca_remove_one)
1355 };
1356
1357 static void __init __mthca_check_profile_val(const char *name, int *pval,
1358                                              int pval_default)
1359 {
1360         /* value must be positive and power of 2 */
1361         int old_pval = *pval;
1362
1363         if (old_pval <= 0)
1364                 *pval = pval_default;
1365         else
1366                 *pval = roundup_pow_of_two(old_pval);
1367
1368         if (old_pval != *pval) {
1369                 printk(KERN_WARNING PFX "Invalid value %d for %s in module parameter.\n",
1370                        old_pval, name);
1371                 printk(KERN_WARNING PFX "Corrected %s to %d.\n", name, *pval);
1372         }
1373 }
1374
1375 #define mthca_check_profile_val(name, default)                          \
1376         __mthca_check_profile_val(#name, &hca_profile.name, default)
1377
1378 static void __init mthca_validate_profile(void)
1379 {
1380         mthca_check_profile_val(num_qp,            MTHCA_DEFAULT_NUM_QP);
1381         mthca_check_profile_val(rdb_per_qp,        MTHCA_DEFAULT_RDB_PER_QP);
1382         mthca_check_profile_val(num_cq,            MTHCA_DEFAULT_NUM_CQ);
1383         mthca_check_profile_val(num_mcg,           MTHCA_DEFAULT_NUM_MCG);
1384         mthca_check_profile_val(num_mpt,           MTHCA_DEFAULT_NUM_MPT);
1385         mthca_check_profile_val(num_mtt,           MTHCA_DEFAULT_NUM_MTT);
1386         mthca_check_profile_val(num_udav,          MTHCA_DEFAULT_NUM_UDAV);
1387         mthca_check_profile_val(fmr_reserved_mtts, MTHCA_DEFAULT_NUM_RESERVED_MTTS);
1388
1389         if (hca_profile.fmr_reserved_mtts >= hca_profile.num_mtt) {
1390                 printk(KERN_WARNING PFX "Invalid fmr_reserved_mtts module parameter %d.\n",
1391                        hca_profile.fmr_reserved_mtts);
1392                 printk(KERN_WARNING PFX "(Must be smaller than num_mtt %d)\n",
1393                        hca_profile.num_mtt);
1394                 hca_profile.fmr_reserved_mtts = hca_profile.num_mtt / 2;
1395                 printk(KERN_WARNING PFX "Corrected fmr_reserved_mtts to %d.\n",
1396                        hca_profile.fmr_reserved_mtts);
1397         }
1398 }
1399
1400 static int __init mthca_init(void)
1401 {
1402         int ret;
1403
1404         mthca_validate_profile();
1405
1406         ret = mthca_catas_init();
1407         if (ret)
1408                 return ret;
1409
1410         ret = pci_register_driver(&mthca_driver);
1411         if (ret < 0) {
1412                 mthca_catas_cleanup();
1413                 return ret;
1414         }
1415
1416         return 0;
1417 }
1418
1419 static void __exit mthca_cleanup(void)
1420 {
1421         pci_unregister_driver(&mthca_driver);
1422         mthca_catas_cleanup();
1423 }
1424
1425 module_init(mthca_init);
1426 module_exit(mthca_cleanup);