]> CyberLeo.Net >> Repos - FreeBSD/releng/10.2.git/blob - sys/ofed/drivers/infiniband/hw/mthca/mthca_main.c
- Copy stable/10@285827 to releng/10.2 in preparation for 10.2-RC1
[FreeBSD/releng/10.2.git] / sys / ofed / drivers / infiniband / hw / mthca / mthca_main.c
1 /*
2  * Copyright (c) 2004, 2005 Topspin Communications.  All rights reserved.
3  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
4  * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
5  *
6  * This software is available to you under a choice of one of two
7  * licenses.  You may choose to be licensed under the terms of the GNU
8  * General Public License (GPL) Version 2, available from the file
9  * COPYING in the main directory of this source tree, or the
10  * OpenIB.org BSD license below:
11  *
12  *     Redistribution and use in source and binary forms, with or
13  *     without modification, are permitted provided that the following
14  *     conditions are met:
15  *
16  *      - Redistributions of source code must retain the above
17  *        copyright notice, this list of conditions and the following
18  *        disclaimer.
19  *
20  *      - Redistributions in binary form must reproduce the above
21  *        copyright notice, this list of conditions and the following
22  *        disclaimer in the documentation and/or other materials
23  *        provided with the distribution.
24  *
25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
32  * SOFTWARE.
33  */
34
35 #include <linux/module.h>
36 #include <linux/errno.h>
37 #include <linux/pci.h>
38 #include <linux/interrupt.h>
39
40 #include "mthca_dev.h"
41 #include "mthca_config_reg.h"
42 #include "mthca_cmd.h"
43 #include "mthca_profile.h"
44 #include "mthca_memfree.h"
45 #include "mthca_wqe.h"
46
47 MODULE_AUTHOR("Roland Dreier");
48 MODULE_DESCRIPTION("Mellanox InfiniBand HCA low-level driver");
49 MODULE_LICENSE("Dual BSD/GPL");
50 MODULE_VERSION(DRV_VERSION);
51
52 #ifdef CONFIG_INFINIBAND_MTHCA_DEBUG
53
54 int mthca_debug_level = 0;
55 module_param_named(debug_level, mthca_debug_level, int, 0644);
56 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
57
58 #endif /* CONFIG_INFINIBAND_MTHCA_DEBUG */
59
60 #ifdef CONFIG_PCI_MSI
61
62 static int msi_x = 1;
63 module_param(msi_x, int, 0444);
64 MODULE_PARM_DESC(msi_x, "attempt to use MSI-X if nonzero");
65
66 #else /* CONFIG_PCI_MSI */
67
68 #define msi_x (0)
69
70 #endif /* CONFIG_PCI_MSI */
71
72 static int tune_pci = 0;
73 module_param(tune_pci, int, 0444);
74 MODULE_PARM_DESC(tune_pci, "increase PCI burst from the default set by BIOS if nonzero");
75
76 DEFINE_MUTEX(mthca_device_mutex);
77
78 #define MTHCA_DEFAULT_NUM_QP            (1 << 16)
79 #define MTHCA_DEFAULT_RDB_PER_QP        (1 << 2)
80 #define MTHCA_DEFAULT_NUM_CQ            (1 << 16)
81 #define MTHCA_DEFAULT_NUM_MCG           (1 << 13)
82 #define MTHCA_DEFAULT_NUM_MPT           (1 << 17)
83 #define MTHCA_DEFAULT_NUM_MTT           (1 << 20)
84 #define MTHCA_DEFAULT_NUM_UDAV          (1 << 15)
85 #define MTHCA_DEFAULT_NUM_RESERVED_MTTS (1 << 18)
86 #define MTHCA_DEFAULT_NUM_UARC_SIZE     (1 << 18)
87
88 static struct mthca_profile hca_profile = {
89         .num_qp             = MTHCA_DEFAULT_NUM_QP,
90         .rdb_per_qp         = MTHCA_DEFAULT_RDB_PER_QP,
91         .num_cq             = MTHCA_DEFAULT_NUM_CQ,
92         .num_mcg            = MTHCA_DEFAULT_NUM_MCG,
93         .num_mpt            = MTHCA_DEFAULT_NUM_MPT,
94         .num_mtt            = MTHCA_DEFAULT_NUM_MTT,
95         .num_udav           = MTHCA_DEFAULT_NUM_UDAV,          /* Tavor only */
96         .fmr_reserved_mtts  = MTHCA_DEFAULT_NUM_RESERVED_MTTS, /* Tavor only */
97         .uarc_size          = MTHCA_DEFAULT_NUM_UARC_SIZE,     /* Arbel only */
98 };
99
100 module_param_named(num_qp, hca_profile.num_qp, int, 0444);
101 MODULE_PARM_DESC(num_qp, "maximum number of QPs per HCA");
102
103 module_param_named(rdb_per_qp, hca_profile.rdb_per_qp, int, 0444);
104 MODULE_PARM_DESC(rdb_per_qp, "number of RDB buffers per QP");
105
106 module_param_named(num_cq, hca_profile.num_cq, int, 0444);
107 MODULE_PARM_DESC(num_cq, "maximum number of CQs per HCA");
108
109 module_param_named(num_mcg, hca_profile.num_mcg, int, 0444);
110 MODULE_PARM_DESC(num_mcg, "maximum number of multicast groups per HCA");
111
112 module_param_named(num_mpt, hca_profile.num_mpt, int, 0444);
113 MODULE_PARM_DESC(num_mpt,
114                 "maximum number of memory protection table entries per HCA");
115
116 module_param_named(num_mtt, hca_profile.num_mtt, int, 0444);
117 MODULE_PARM_DESC(num_mtt,
118                  "maximum number of memory translation table segments per HCA");
119
120 module_param_named(num_udav, hca_profile.num_udav, int, 0444);
121 MODULE_PARM_DESC(num_udav, "maximum number of UD address vectors per HCA");
122
123 module_param_named(fmr_reserved_mtts, hca_profile.fmr_reserved_mtts, int, 0444);
124 MODULE_PARM_DESC(fmr_reserved_mtts,
125                  "number of memory translation table segments reserved for FMR");
126
127 static int log_mtts_per_seg;
128 module_param_named(log_mtts_per_seg, log_mtts_per_seg, int, 0444);
129 MODULE_PARM_DESC(log_mtts_per_seg, "Log2 number of MTT entries per segment (1-5)");
130
131 static char mthca_version[] __devinitdata =
132         DRV_NAME ": Mellanox InfiniBand HCA driver v"
133         DRV_VERSION " (" DRV_RELDATE ")\n";
134
135 static int mthca_tune_pci(struct mthca_dev *mdev)
136 {
137         if (!tune_pci)
138                 return 0;
139
140         /* First try to max out Read Byte Count */
141         if (pci_find_capability(mdev->pdev, PCI_CAP_ID_PCIX)) {
142                 if (pcix_set_mmrbc(mdev->pdev, pcix_get_max_mmrbc(mdev->pdev))) {
143                         mthca_err(mdev, "Couldn't set PCI-X max read count, "
144                                 "aborting.\n");
145                         return -ENODEV;
146                 }
147         } else if (!(mdev->mthca_flags & MTHCA_FLAG_PCIE))
148                 mthca_info(mdev, "No PCI-X capability, not setting RBC.\n");
149
150         if (pci_find_capability(mdev->pdev, PCI_CAP_ID_EXP)) {
151                 if (pcie_set_readrq(mdev->pdev, 4096)) {
152                         mthca_err(mdev, "Couldn't write PCI Express read request, "
153                                 "aborting.\n");
154                         return -ENODEV;
155                 }
156         } else if (mdev->mthca_flags & MTHCA_FLAG_PCIE)
157                 mthca_info(mdev, "No PCI Express capability, "
158                            "not setting Max Read Request Size.\n");
159
160         return 0;
161 }
162
163 static int mthca_dev_lim(struct mthca_dev *mdev, struct mthca_dev_lim *dev_lim)
164 {
165         int err;
166         u8 status;
167
168         mdev->limits.mtt_seg_size = (1 << log_mtts_per_seg) * 8;
169         err = mthca_QUERY_DEV_LIM(mdev, dev_lim, &status);
170         if (err) {
171                 mthca_err(mdev, "QUERY_DEV_LIM command failed, aborting.\n");
172                 return err;
173         }
174         if (status) {
175                 mthca_err(mdev, "QUERY_DEV_LIM returned status 0x%02x, "
176                           "aborting.\n", status);
177                 return -EINVAL;
178         }
179         if (dev_lim->min_page_sz > PAGE_SIZE) {
180                 mthca_err(mdev, "HCA minimum page size of %d bigger than "
181                           "kernel PAGE_SIZE of %d, aborting.\n",
182                           dev_lim->min_page_sz, PAGE_SIZE);
183                 return -ENODEV;
184         }
185         if (dev_lim->num_ports > MTHCA_MAX_PORTS) {
186                 mthca_err(mdev, "HCA has %d ports, but we only support %d, "
187                           "aborting.\n",
188                           dev_lim->num_ports, MTHCA_MAX_PORTS);
189                 return -ENODEV;
190         }
191
192         if (dev_lim->uar_size > pci_resource_len(mdev->pdev, 2)) {
193                 mthca_err(mdev, "HCA reported UAR size of 0x%x bigger than "
194                           "PCI resource 2 size of 0x%llx, aborting.\n",
195                           dev_lim->uar_size,
196                           (unsigned long long)pci_resource_len(mdev->pdev, 2));
197                 return -ENODEV;
198         }
199
200         mdev->limits.num_ports          = dev_lim->num_ports;
201         mdev->limits.vl_cap             = dev_lim->max_vl;
202         mdev->limits.mtu_cap            = dev_lim->max_mtu;
203         mdev->limits.gid_table_len      = dev_lim->max_gids;
204         mdev->limits.pkey_table_len     = dev_lim->max_pkeys;
205         mdev->limits.local_ca_ack_delay = dev_lim->local_ca_ack_delay;
206         /*
207          * Need to allow for worst case send WQE overhead and check
208          * whether max_desc_sz imposes a lower limit than max_sg; UD
209          * send has the biggest overhead.
210          */
211         mdev->limits.max_sg             = min_t(int, dev_lim->max_sg,
212                                               (dev_lim->max_desc_sz -
213                                                sizeof (struct mthca_next_seg) -
214                                                (mthca_is_memfree(mdev) ?
215                                                 sizeof (struct mthca_arbel_ud_seg) :
216                                                 sizeof (struct mthca_tavor_ud_seg))) /
217                                                 sizeof (struct mthca_data_seg));
218         mdev->limits.max_wqes           = dev_lim->max_qp_sz;
219         mdev->limits.max_qp_init_rdma   = dev_lim->max_requester_per_qp;
220         mdev->limits.reserved_qps       = dev_lim->reserved_qps;
221         mdev->limits.max_srq_wqes       = dev_lim->max_srq_sz;
222         mdev->limits.reserved_srqs      = dev_lim->reserved_srqs;
223         mdev->limits.reserved_eecs      = dev_lim->reserved_eecs;
224         mdev->limits.max_desc_sz        = dev_lim->max_desc_sz;
225         mdev->limits.max_srq_sge        = mthca_max_srq_sge(mdev);
226         /*
227          * Subtract 1 from the limit because we need to allocate a
228          * spare CQE so the HCA HW can tell the difference between an
229          * empty CQ and a full CQ.
230          */
231         mdev->limits.max_cqes           = dev_lim->max_cq_sz - 1;
232         mdev->limits.reserved_cqs       = dev_lim->reserved_cqs;
233         mdev->limits.reserved_eqs       = dev_lim->reserved_eqs;
234         mdev->limits.reserved_mtts      = dev_lim->reserved_mtts;
235         mdev->limits.reserved_mrws      = dev_lim->reserved_mrws;
236         mdev->limits.reserved_uars      = dev_lim->reserved_uars;
237         mdev->limits.reserved_pds       = dev_lim->reserved_pds;
238         mdev->limits.port_width_cap     = dev_lim->max_port_width;
239         mdev->limits.page_size_cap      = ~(u32) (dev_lim->min_page_sz - 1);
240         mdev->limits.flags              = dev_lim->flags;
241         /*
242          * For old FW that doesn't return static rate support, use a
243          * value of 0x3 (only static rate values of 0 or 1 are handled),
244          * except on Sinai, where even old FW can handle static rate
245          * values of 2 and 3.
246          */
247         if (dev_lim->stat_rate_support)
248                 mdev->limits.stat_rate_support = dev_lim->stat_rate_support;
249         else if (mdev->mthca_flags & MTHCA_FLAG_SINAI_OPT)
250                 mdev->limits.stat_rate_support = 0xf;
251         else
252                 mdev->limits.stat_rate_support = 0x3;
253
254         /* IB_DEVICE_RESIZE_MAX_WR not supported by driver.
255            May be doable since hardware supports it for SRQ.
256
257            IB_DEVICE_N_NOTIFY_CQ is supported by hardware but not by driver.
258
259            IB_DEVICE_SRQ_RESIZE is supported by hardware but SRQ is not
260            supported by driver. */
261         mdev->device_cap_flags = IB_DEVICE_CHANGE_PHY_PORT |
262                 IB_DEVICE_PORT_ACTIVE_EVENT |
263                 IB_DEVICE_SYS_IMAGE_GUID |
264                 IB_DEVICE_RC_RNR_NAK_GEN;
265
266         if (dev_lim->flags & DEV_LIM_FLAG_BAD_PKEY_CNTR)
267                 mdev->device_cap_flags |= IB_DEVICE_BAD_PKEY_CNTR;
268
269         if (dev_lim->flags & DEV_LIM_FLAG_BAD_QKEY_CNTR)
270                 mdev->device_cap_flags |= IB_DEVICE_BAD_QKEY_CNTR;
271
272         if (dev_lim->flags & DEV_LIM_FLAG_RAW_MULTI)
273                 mdev->device_cap_flags |= IB_DEVICE_RAW_MULTI;
274
275         if (dev_lim->flags & DEV_LIM_FLAG_AUTO_PATH_MIG)
276                 mdev->device_cap_flags |= IB_DEVICE_AUTO_PATH_MIG;
277
278         if (dev_lim->flags & DEV_LIM_FLAG_UD_AV_PORT_ENFORCE)
279                 mdev->device_cap_flags |= IB_DEVICE_UD_AV_PORT_ENFORCE;
280
281         if (dev_lim->flags & DEV_LIM_FLAG_SRQ)
282                 mdev->mthca_flags |= MTHCA_FLAG_SRQ;
283
284         if (mthca_is_memfree(mdev))
285                 if (dev_lim->flags & DEV_LIM_FLAG_IPOIB_CSUM)
286                         mdev->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
287
288         return 0;
289 }
290
291 static int mthca_init_tavor(struct mthca_dev *mdev)
292 {
293         s64 size;
294         u8 status;
295         int err;
296         struct mthca_dev_lim        dev_lim;
297         struct mthca_profile        profile;
298         struct mthca_init_hca_param init_hca;
299
300         err = mthca_SYS_EN(mdev, &status);
301         if (err) {
302                 mthca_err(mdev, "SYS_EN command failed, aborting.\n");
303                 return err;
304         }
305         if (status) {
306                 mthca_err(mdev, "SYS_EN returned status 0x%02x, "
307                           "aborting.\n", status);
308                 return -EINVAL;
309         }
310
311         err = mthca_QUERY_FW(mdev, &status);
312         if (err) {
313                 mthca_err(mdev, "QUERY_FW command failed, aborting.\n");
314                 goto err_disable;
315         }
316         if (status) {
317                 mthca_err(mdev, "QUERY_FW returned status 0x%02x, "
318                           "aborting.\n", status);
319                 err = -EINVAL;
320                 goto err_disable;
321         }
322         err = mthca_QUERY_DDR(mdev, &status);
323         if (err) {
324                 mthca_err(mdev, "QUERY_DDR command failed, aborting.\n");
325                 goto err_disable;
326         }
327         if (status) {
328                 mthca_err(mdev, "QUERY_DDR returned status 0x%02x, "
329                           "aborting.\n", status);
330                 err = -EINVAL;
331                 goto err_disable;
332         }
333
334         err = mthca_dev_lim(mdev, &dev_lim);
335         if (err) {
336                 mthca_err(mdev, "QUERY_DEV_LIM command failed, aborting.\n");
337                 goto err_disable;
338         }
339
340         profile = hca_profile;
341         profile.num_uar   = dev_lim.uar_size / PAGE_SIZE;
342         profile.uarc_size = 0;
343         if (mdev->mthca_flags & MTHCA_FLAG_SRQ)
344                 profile.num_srq = dev_lim.max_srqs;
345
346         size = mthca_make_profile(mdev, &profile, &dev_lim, &init_hca);
347         if (size < 0) {
348                 err = size;
349                 goto err_disable;
350         }
351
352         err = mthca_INIT_HCA(mdev, &init_hca, &status);
353         if (err) {
354                 mthca_err(mdev, "INIT_HCA command failed, aborting.\n");
355                 goto err_disable;
356         }
357         if (status) {
358                 mthca_err(mdev, "INIT_HCA returned status 0x%02x, "
359                           "aborting.\n", status);
360                 err = -EINVAL;
361                 goto err_disable;
362         }
363
364         return 0;
365
366 err_disable:
367         mthca_SYS_DIS(mdev, &status);
368
369         return err;
370 }
371
372 static int mthca_load_fw(struct mthca_dev *mdev)
373 {
374         u8 status;
375         int err;
376
377         /* FIXME: use HCA-attached memory for FW if present */
378
379         mdev->fw.arbel.fw_icm =
380                 mthca_alloc_icm(mdev, mdev->fw.arbel.fw_pages,
381                                 GFP_HIGHUSER | __GFP_NOWARN, 0);
382         if (!mdev->fw.arbel.fw_icm) {
383                 mthca_err(mdev, "Couldn't allocate FW area, aborting.\n");
384                 return -ENOMEM;
385         }
386
387         err = mthca_MAP_FA(mdev, mdev->fw.arbel.fw_icm, &status);
388         if (err) {
389                 mthca_err(mdev, "MAP_FA command failed, aborting.\n");
390                 goto err_free;
391         }
392         if (status) {
393                 mthca_err(mdev, "MAP_FA returned status 0x%02x, aborting.\n", status);
394                 err = -EINVAL;
395                 goto err_free;
396         }
397         err = mthca_RUN_FW(mdev, &status);
398         if (err) {
399                 mthca_err(mdev, "RUN_FW command failed, aborting.\n");
400                 goto err_unmap_fa;
401         }
402         if (status) {
403                 mthca_err(mdev, "RUN_FW returned status 0x%02x, aborting.\n", status);
404                 err = -EINVAL;
405                 goto err_unmap_fa;
406         }
407
408         return 0;
409
410 err_unmap_fa:
411         mthca_UNMAP_FA(mdev, &status);
412
413 err_free:
414         mthca_free_icm(mdev, mdev->fw.arbel.fw_icm, 0);
415         return err;
416 }
417
418 static int mthca_init_icm(struct mthca_dev *mdev,
419                           struct mthca_dev_lim *dev_lim,
420                           struct mthca_init_hca_param *init_hca,
421                           u64 icm_size)
422 {
423         u64 aux_pages;
424         u8 status;
425         int err;
426
427         err = mthca_SET_ICM_SIZE(mdev, icm_size, &aux_pages, &status);
428         if (err) {
429                 mthca_err(mdev, "SET_ICM_SIZE command failed, aborting.\n");
430                 return err;
431         }
432         if (status) {
433                 mthca_err(mdev, "SET_ICM_SIZE returned status 0x%02x, "
434                           "aborting.\n", status);
435                 return -EINVAL;
436         }
437
438         mthca_dbg(mdev, "%lld KB of HCA context requires %lld KB aux memory.\n",
439                   (unsigned long long) icm_size >> 10,
440                   (unsigned long long) aux_pages << 2);
441
442         mdev->fw.arbel.aux_icm = mthca_alloc_icm(mdev, aux_pages,
443                                                  GFP_HIGHUSER | __GFP_NOWARN, 0);
444         if (!mdev->fw.arbel.aux_icm) {
445                 mthca_err(mdev, "Couldn't allocate aux memory, aborting.\n");
446                 return -ENOMEM;
447         }
448
449         err = mthca_MAP_ICM_AUX(mdev, mdev->fw.arbel.aux_icm, &status);
450         if (err) {
451                 mthca_err(mdev, "MAP_ICM_AUX command failed, aborting.\n");
452                 goto err_free_aux;
453         }
454         if (status) {
455                 mthca_err(mdev, "MAP_ICM_AUX returned status 0x%02x, aborting.\n", status);
456                 err = -EINVAL;
457                 goto err_free_aux;
458         }
459
460         err = mthca_map_eq_icm(mdev, init_hca->eqc_base);
461         if (err) {
462                 mthca_err(mdev, "Failed to map EQ context memory, aborting.\n");
463                 goto err_unmap_aux;
464         }
465
466         /* CPU writes to non-reserved MTTs, while HCA might DMA to reserved mtts */
467         mdev->limits.reserved_mtts = ALIGN(mdev->limits.reserved_mtts * mdev->limits.mtt_seg_size,
468                                            dma_get_cache_alignment()) / mdev->limits.mtt_seg_size;
469
470         mdev->mr_table.mtt_table = mthca_alloc_icm_table(mdev, init_hca->mtt_base,
471                                                          mdev->limits.mtt_seg_size,
472                                                          mdev->limits.num_mtt_segs,
473                                                          mdev->limits.reserved_mtts,
474                                                          1, 0);
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,
485                                                          1, 1);
486         if (!mdev->mr_table.mpt_table) {
487                 mthca_err(mdev, "Failed to map MPT context memory, aborting.\n");
488                 err = -ENOMEM;
489                 goto err_unmap_mtt;
490         }
491
492         mdev->qp_table.qp_table = mthca_alloc_icm_table(mdev, init_hca->qpc_base,
493                                                         dev_lim->qpc_entry_sz,
494                                                         mdev->limits.num_qps,
495                                                         mdev->limits.reserved_qps,
496                                                         0, 0);
497         if (!mdev->qp_table.qp_table) {
498                 mthca_err(mdev, "Failed to map QP context memory, aborting.\n");
499                 err = -ENOMEM;
500                 goto err_unmap_mpt;
501         }
502
503         mdev->qp_table.eqp_table = mthca_alloc_icm_table(mdev, init_hca->eqpc_base,
504                                                          dev_lim->eqpc_entry_sz,
505                                                          mdev->limits.num_qps,
506                                                          mdev->limits.reserved_qps,
507                                                          0, 0);
508         if (!mdev->qp_table.eqp_table) {
509                 mthca_err(mdev, "Failed to map EQP context memory, aborting.\n");
510                 err = -ENOMEM;
511                 goto err_unmap_qp;
512         }
513
514         mdev->qp_table.rdb_table = mthca_alloc_icm_table(mdev, init_hca->rdb_base,
515                                                          MTHCA_RDB_ENTRY_SIZE,
516                                                          mdev->limits.num_qps <<
517                                                          mdev->qp_table.rdb_shift, 0,
518                                                          0, 0);
519         if (!mdev->qp_table.rdb_table) {
520                 mthca_err(mdev, "Failed to map RDB context memory, aborting\n");
521                 err = -ENOMEM;
522                 goto err_unmap_eqp;
523         }
524
525        mdev->cq_table.table = mthca_alloc_icm_table(mdev, init_hca->cqc_base,
526                                                     dev_lim->cqc_entry_sz,
527                                                     mdev->limits.num_cqs,
528                                                     mdev->limits.reserved_cqs,
529                                                     0, 0);
530         if (!mdev->cq_table.table) {
531                 mthca_err(mdev, "Failed to map CQ context memory, aborting.\n");
532                 err = -ENOMEM;
533                 goto err_unmap_rdb;
534         }
535
536         if (mdev->mthca_flags & MTHCA_FLAG_SRQ) {
537                 mdev->srq_table.table =
538                         mthca_alloc_icm_table(mdev, init_hca->srqc_base,
539                                               dev_lim->srq_entry_sz,
540                                               mdev->limits.num_srqs,
541                                               mdev->limits.reserved_srqs,
542                                               0, 0);
543                 if (!mdev->srq_table.table) {
544                         mthca_err(mdev, "Failed to map SRQ context memory, "
545                                   "aborting.\n");
546                         err = -ENOMEM;
547                         goto err_unmap_cq;
548                 }
549         }
550
551         /*
552          * It's not strictly required, but for simplicity just map the
553          * whole multicast group table now.  The table isn't very big
554          * and it's a lot easier than trying to track ref counts.
555          */
556         mdev->mcg_table.table = mthca_alloc_icm_table(mdev, init_hca->mc_base,
557                                                       MTHCA_MGM_ENTRY_SIZE,
558                                                       mdev->limits.num_mgms +
559                                                       mdev->limits.num_amgms,
560                                                       mdev->limits.num_mgms +
561                                                       mdev->limits.num_amgms,
562                                                       0, 0);
563         if (!mdev->mcg_table.table) {
564                 mthca_err(mdev, "Failed to map MCG context memory, aborting.\n");
565                 err = -ENOMEM;
566                 goto err_unmap_srq;
567         }
568
569         return 0;
570
571 err_unmap_srq:
572         if (mdev->mthca_flags & MTHCA_FLAG_SRQ)
573                 mthca_free_icm_table(mdev, mdev->srq_table.table);
574
575 err_unmap_cq:
576         mthca_free_icm_table(mdev, mdev->cq_table.table);
577
578 err_unmap_rdb:
579         mthca_free_icm_table(mdev, mdev->qp_table.rdb_table);
580
581 err_unmap_eqp:
582         mthca_free_icm_table(mdev, mdev->qp_table.eqp_table);
583
584 err_unmap_qp:
585         mthca_free_icm_table(mdev, mdev->qp_table.qp_table);
586
587 err_unmap_mpt:
588         mthca_free_icm_table(mdev, mdev->mr_table.mpt_table);
589
590 err_unmap_mtt:
591         mthca_free_icm_table(mdev, mdev->mr_table.mtt_table);
592
593 err_unmap_eq:
594         mthca_unmap_eq_icm(mdev);
595
596 err_unmap_aux:
597         mthca_UNMAP_ICM_AUX(mdev, &status);
598
599 err_free_aux:
600         mthca_free_icm(mdev, mdev->fw.arbel.aux_icm, 0);
601
602         return err;
603 }
604
605 static void mthca_free_icms(struct mthca_dev *mdev)
606 {
607         u8 status;
608
609         mthca_free_icm_table(mdev, mdev->mcg_table.table);
610         if (mdev->mthca_flags & MTHCA_FLAG_SRQ)
611                 mthca_free_icm_table(mdev, mdev->srq_table.table);
612         mthca_free_icm_table(mdev, mdev->cq_table.table);
613         mthca_free_icm_table(mdev, mdev->qp_table.rdb_table);
614         mthca_free_icm_table(mdev, mdev->qp_table.eqp_table);
615         mthca_free_icm_table(mdev, mdev->qp_table.qp_table);
616         mthca_free_icm_table(mdev, mdev->mr_table.mpt_table);
617         mthca_free_icm_table(mdev, mdev->mr_table.mtt_table);
618         mthca_unmap_eq_icm(mdev);
619
620         mthca_UNMAP_ICM_AUX(mdev, &status);
621         mthca_free_icm(mdev, mdev->fw.arbel.aux_icm, 0);
622 }
623
624 static int mthca_init_arbel(struct mthca_dev *mdev)
625 {
626         struct mthca_dev_lim        dev_lim;
627         struct mthca_profile        profile;
628         struct mthca_init_hca_param init_hca;
629         s64 icm_size;
630         u8 status;
631         int err;
632
633         err = mthca_QUERY_FW(mdev, &status);
634         if (err) {
635                 mthca_err(mdev, "QUERY_FW command failed, aborting.\n");
636                 return err;
637         }
638         if (status) {
639                 mthca_err(mdev, "QUERY_FW returned status 0x%02x, "
640                           "aborting.\n", status);
641                 return -EINVAL;
642         }
643
644         err = mthca_ENABLE_LAM(mdev, &status);
645         if (err) {
646                 mthca_err(mdev, "ENABLE_LAM command failed, aborting.\n");
647                 return err;
648         }
649         if (status == MTHCA_CMD_STAT_LAM_NOT_PRE) {
650                 mthca_dbg(mdev, "No HCA-attached memory (running in MemFree mode)\n");
651                 mdev->mthca_flags |= MTHCA_FLAG_NO_LAM;
652         } else if (status) {
653                 mthca_err(mdev, "ENABLE_LAM returned status 0x%02x, "
654                           "aborting.\n", status);
655                 return -EINVAL;
656         }
657
658         err = mthca_load_fw(mdev);
659         if (err) {
660                 mthca_err(mdev, "Failed to start FW, aborting.\n");
661                 goto err_disable;
662         }
663
664         err = mthca_dev_lim(mdev, &dev_lim);
665         if (err) {
666                 mthca_err(mdev, "QUERY_DEV_LIM command failed, aborting.\n");
667                 goto err_stop_fw;
668         }
669
670         profile = hca_profile;
671         profile.num_uar  = dev_lim.uar_size / PAGE_SIZE;
672         profile.num_udav = 0;
673         if (mdev->mthca_flags & MTHCA_FLAG_SRQ)
674                 profile.num_srq = dev_lim.max_srqs;
675
676         icm_size = mthca_make_profile(mdev, &profile, &dev_lim, &init_hca);
677         if (icm_size < 0) {
678                 err = icm_size;
679                 goto err_stop_fw;
680         }
681
682         err = mthca_init_icm(mdev, &dev_lim, &init_hca, icm_size);
683         if (err)
684                 goto err_stop_fw;
685
686         err = mthca_INIT_HCA(mdev, &init_hca, &status);
687         if (err) {
688                 mthca_err(mdev, "INIT_HCA command failed, aborting.\n");
689                 goto err_free_icm;
690         }
691         if (status) {
692                 mthca_err(mdev, "INIT_HCA returned status 0x%02x, "
693                           "aborting.\n", status);
694                 err = -EINVAL;
695                 goto err_free_icm;
696         }
697
698         return 0;
699
700 err_free_icm:
701         mthca_free_icms(mdev);
702
703 err_stop_fw:
704         mthca_UNMAP_FA(mdev, &status);
705         mthca_free_icm(mdev, mdev->fw.arbel.fw_icm, 0);
706
707 err_disable:
708         if (!(mdev->mthca_flags & MTHCA_FLAG_NO_LAM))
709                 mthca_DISABLE_LAM(mdev, &status);
710
711         return err;
712 }
713
714 static void mthca_close_hca(struct mthca_dev *mdev)
715 {
716         u8 status;
717
718         mthca_CLOSE_HCA(mdev, 0, &status);
719
720         if (mthca_is_memfree(mdev)) {
721                 mthca_free_icms(mdev);
722
723                 mthca_UNMAP_FA(mdev, &status);
724                 mthca_free_icm(mdev, mdev->fw.arbel.fw_icm, 0);
725
726                 if (!(mdev->mthca_flags & MTHCA_FLAG_NO_LAM))
727                         mthca_DISABLE_LAM(mdev, &status);
728         } else
729                 mthca_SYS_DIS(mdev, &status);
730 }
731
732 static int mthca_init_hca(struct mthca_dev *mdev)
733 {
734         u8 status;
735         int err;
736         struct mthca_adapter adapter;
737
738         if (mthca_is_memfree(mdev))
739                 err = mthca_init_arbel(mdev);
740         else
741                 err = mthca_init_tavor(mdev);
742
743         if (err)
744                 return err;
745
746         err = mthca_QUERY_ADAPTER(mdev, &adapter, &status);
747         if (err) {
748                 mthca_err(mdev, "QUERY_ADAPTER command failed, aborting.\n");
749                 goto err_close;
750         }
751         if (status) {
752                 mthca_err(mdev, "QUERY_ADAPTER returned status 0x%02x, "
753                           "aborting.\n", status);
754                 err = -EINVAL;
755                 goto err_close;
756         }
757
758         mdev->eq_table.inta_pin = adapter.inta_pin;
759         if (!mthca_is_memfree(mdev))
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_X) {
837                         mthca_warn(dev, "NOP command failed to generate interrupt "
838                                    "(IRQ %d).\n",
839                                    dev->eq_table.eq[MTHCA_EQ_CMD].msi_x_vector);
840                         mthca_warn(dev, "Trying again with MSI-X disabled.\n");
841                 } else {
842                         mthca_err(dev, "NOP command failed to generate interrupt "
843                                   "(IRQ %d), aborting.\n",
844                                   dev->pdev->irq);
845                         mthca_err(dev, "BIOS or ACPI interrupt routing problem?\n");
846                 }
847
848                 goto err_cmd_poll;
849         }
850
851         mthca_dbg(dev, "NOP command IRQ test passed\n");
852
853         err = mthca_init_cq_table(dev);
854         if (err) {
855                 mthca_err(dev, "Failed to initialize "
856                           "completion queue table, aborting.\n");
857                 goto err_cmd_poll;
858         }
859
860         err = mthca_init_srq_table(dev);
861         if (err) {
862                 mthca_err(dev, "Failed to initialize "
863                           "shared receive queue table, aborting.\n");
864                 goto err_cq_table_free;
865         }
866
867         err = mthca_init_qp_table(dev);
868         if (err) {
869                 mthca_err(dev, "Failed to initialize "
870                           "queue pair table, aborting.\n");
871                 goto err_srq_table_free;
872         }
873
874         err = mthca_init_av_table(dev);
875         if (err) {
876                 mthca_err(dev, "Failed to initialize "
877                           "address vector table, aborting.\n");
878                 goto err_qp_table_free;
879         }
880
881         err = mthca_init_mcg_table(dev);
882         if (err) {
883                 mthca_err(dev, "Failed to initialize "
884                           "multicast group table, aborting.\n");
885                 goto err_av_table_free;
886         }
887
888         return 0;
889
890 err_av_table_free:
891         mthca_cleanup_av_table(dev);
892
893 err_qp_table_free:
894         mthca_cleanup_qp_table(dev);
895
896 err_srq_table_free:
897         mthca_cleanup_srq_table(dev);
898
899 err_cq_table_free:
900         mthca_cleanup_cq_table(dev);
901
902 err_cmd_poll:
903         mthca_cmd_use_polling(dev);
904
905 err_eq_table_free:
906         mthca_cleanup_eq_table(dev);
907
908 err_pd_free:
909         mthca_pd_free(dev, &dev->driver_pd);
910
911 err_mr_table_free:
912         mthca_cleanup_mr_table(dev);
913
914 err_pd_table_free:
915         mthca_cleanup_pd_table(dev);
916
917 err_kar_unmap:
918         iounmap(dev->kar);
919
920 err_uar_free:
921         mthca_uar_free(dev, &dev->driver_uar);
922
923 err_uar_table_free:
924         mthca_cleanup_uar_table(dev);
925         return err;
926 }
927
928 static int mthca_enable_msi_x(struct mthca_dev *mdev)
929 {
930         struct msix_entry entries[3];
931         int err;
932
933         entries[0].entry = 0;
934         entries[1].entry = 1;
935         entries[2].entry = 2;
936
937         err = pci_enable_msix(mdev->pdev, entries, ARRAY_SIZE(entries));
938         if (err) {
939                 if (err > 0)
940                         mthca_info(mdev, "Only %d MSI-X vectors available, "
941                                    "not using MSI-X\n", err);
942                 return err;
943         }
944
945         mdev->eq_table.eq[MTHCA_EQ_COMP ].msi_x_vector = entries[0].vector;
946         mdev->eq_table.eq[MTHCA_EQ_ASYNC].msi_x_vector = entries[1].vector;
947         mdev->eq_table.eq[MTHCA_EQ_CMD  ].msi_x_vector = entries[2].vector;
948
949         return 0;
950 }
951
952 /* Types of supported HCA */
953 enum {
954         TAVOR,                  /* MT23108                        */
955         ARBEL_COMPAT,           /* MT25208 in Tavor compat mode   */
956         ARBEL_NATIVE,           /* MT25208 with extended features */
957         SINAI                   /* MT25204 */
958 };
959
960 #define MTHCA_FW_VER(major, minor, subminor) \
961         (((u64) (major) << 32) | ((u64) (minor) << 16) | (u64) (subminor))
962
963 static struct {
964         u64 latest_fw;
965         u32 flags;
966 } mthca_hca_table[] = {
967         [TAVOR]        = { .latest_fw = MTHCA_FW_VER(3, 5, 0),
968                            .flags     = 0 },
969         [ARBEL_COMPAT] = { .latest_fw = MTHCA_FW_VER(4, 8, 200),
970                            .flags     = MTHCA_FLAG_PCIE },
971         [ARBEL_NATIVE] = { .latest_fw = MTHCA_FW_VER(5, 3, 0),
972                            .flags     = MTHCA_FLAG_MEMFREE |
973                                         MTHCA_FLAG_PCIE },
974         [SINAI]        = { .latest_fw = MTHCA_FW_VER(1, 2, 0),
975                            .flags     = MTHCA_FLAG_MEMFREE |
976                                         MTHCA_FLAG_PCIE    |
977                                         MTHCA_FLAG_SINAI_OPT }
978 };
979
980 static int __mthca_init_one(struct pci_dev *pdev, int hca_type)
981 {
982         int ddr_hidden = 0;
983         int err;
984         struct mthca_dev *mdev;
985
986         printk(KERN_INFO PFX "Initializing %s\n",
987                pci_name(pdev));
988
989         err = pci_enable_device(pdev);
990         if (err) {
991                 dev_err(&pdev->dev, "Cannot enable PCI device, "
992                         "aborting.\n");
993                 return err;
994         }
995
996         /*
997          * Check for BARs.  We expect 0: 1MB, 2: 8MB, 4: DDR (may not
998          * be present)
999          */
1000         if (!(pci_resource_flags(pdev, 0) & IORESOURCE_MEM) ||
1001             pci_resource_len(pdev, 0) != 1 << 20) {
1002                 dev_err(&pdev->dev, "Missing DCS, aborting.\n");
1003                 err = -ENODEV;
1004                 goto err_disable_pdev;
1005         }
1006         if (!(pci_resource_flags(pdev, 2) & IORESOURCE_MEM)) {
1007                 dev_err(&pdev->dev, "Missing UAR, aborting.\n");
1008                 err = -ENODEV;
1009                 goto err_disable_pdev;
1010         }
1011         if (!(pci_resource_flags(pdev, 4) & IORESOURCE_MEM))
1012                 ddr_hidden = 1;
1013
1014         err = pci_request_regions(pdev, DRV_NAME);
1015         if (err) {
1016                 dev_err(&pdev->dev, "Cannot obtain PCI resources, "
1017                         "aborting.\n");
1018                 goto err_disable_pdev;
1019         }
1020
1021         pci_set_master(pdev);
1022
1023         err = pci_set_dma_mask(pdev, DMA_BIT_MASK(64));
1024         if (err) {
1025                 dev_warn(&pdev->dev, "Warning: couldn't set 64-bit PCI DMA mask.\n");
1026                 err = pci_set_dma_mask(pdev, DMA_BIT_MASK(32));
1027                 if (err) {
1028                         dev_err(&pdev->dev, "Can't set PCI DMA mask, aborting.\n");
1029                         goto err_free_res;
1030                 }
1031         }
1032         err = pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(64));
1033         if (err) {
1034                 dev_warn(&pdev->dev, "Warning: couldn't set 64-bit "
1035                          "consistent PCI DMA mask.\n");
1036                 err = pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(32));
1037                 if (err) {
1038                         dev_err(&pdev->dev, "Can't set consistent PCI DMA mask, "
1039                                 "aborting.\n");
1040                         goto err_free_res;
1041                 }
1042         }
1043
1044         mdev = (struct mthca_dev *) ib_alloc_device(sizeof *mdev);
1045         if (!mdev) {
1046                 dev_err(&pdev->dev, "Device struct alloc failed, "
1047                         "aborting.\n");
1048                 err = -ENOMEM;
1049                 goto err_free_res;
1050         }
1051
1052         mdev->pdev = pdev;
1053
1054         mdev->mthca_flags = mthca_hca_table[hca_type].flags;
1055         if (ddr_hidden)
1056                 mdev->mthca_flags |= MTHCA_FLAG_DDR_HIDDEN;
1057
1058         /*
1059          * Now reset the HCA before we touch the PCI capabilities or
1060          * attempt a firmware command, since a boot ROM may have left
1061          * the HCA in an undefined state.
1062          */
1063         err = mthca_reset(mdev);
1064         if (err) {
1065                 mthca_err(mdev, "Failed to reset HCA, aborting.\n");
1066                 goto err_free_dev;
1067         }
1068
1069         if (mthca_cmd_init(mdev)) {
1070                 mthca_err(mdev, "Failed to init command interface, aborting.\n");
1071                 goto err_free_dev;
1072         }
1073
1074         err = mthca_tune_pci(mdev);
1075         if (err)
1076                 goto err_cmd;
1077
1078         err = mthca_init_hca(mdev);
1079         if (err)
1080                 goto err_cmd;
1081
1082         if (mdev->fw_ver < mthca_hca_table[hca_type].latest_fw) {
1083                 mthca_warn(mdev, "HCA FW version %d.%d.%03d is old (%d.%d.%03d is current).\n",
1084                            (int) (mdev->fw_ver >> 32), (int) (mdev->fw_ver >> 16) & 0xffff,
1085                            (int) (mdev->fw_ver & 0xffff),
1086                            (int) (mthca_hca_table[hca_type].latest_fw >> 32),
1087                            (int) (mthca_hca_table[hca_type].latest_fw >> 16) & 0xffff,
1088                            (int) (mthca_hca_table[hca_type].latest_fw & 0xffff));
1089                 mthca_warn(mdev, "If you have problems, try updating your HCA FW.\n");
1090         }
1091
1092         if (msi_x && !mthca_enable_msi_x(mdev))
1093                 mdev->mthca_flags |= MTHCA_FLAG_MSI_X;
1094
1095         err = mthca_setup_hca(mdev);
1096         if (err == -EBUSY && (mdev->mthca_flags & MTHCA_FLAG_MSI_X)) {
1097                 if (mdev->mthca_flags & MTHCA_FLAG_MSI_X)
1098                         pci_disable_msix(pdev);
1099                 mdev->mthca_flags &= ~MTHCA_FLAG_MSI_X;
1100
1101                 err = mthca_setup_hca(mdev);
1102         }
1103
1104         if (err)
1105                 goto err_close;
1106
1107         err = mthca_register_device(mdev);
1108         if (err)
1109                 goto err_cleanup;
1110
1111         err = mthca_create_agents(mdev);
1112         if (err)
1113                 goto err_unregister;
1114
1115         pci_set_drvdata(pdev, mdev);
1116         mdev->hca_type = hca_type;
1117
1118         mdev->active = 1;
1119
1120         return 0;
1121
1122 err_unregister:
1123         mthca_unregister_device(mdev);
1124
1125 err_cleanup:
1126         mthca_cleanup_mcg_table(mdev);
1127         mthca_cleanup_av_table(mdev);
1128         mthca_cleanup_qp_table(mdev);
1129         mthca_cleanup_srq_table(mdev);
1130         mthca_cleanup_cq_table(mdev);
1131         mthca_cmd_use_polling(mdev);
1132         mthca_cleanup_eq_table(mdev);
1133
1134         mthca_pd_free(mdev, &mdev->driver_pd);
1135
1136         mthca_cleanup_mr_table(mdev);
1137         mthca_cleanup_pd_table(mdev);
1138         mthca_cleanup_uar_table(mdev);
1139
1140 err_close:
1141         if (mdev->mthca_flags & MTHCA_FLAG_MSI_X)
1142                 pci_disable_msix(pdev);
1143
1144         mthca_close_hca(mdev);
1145
1146 err_cmd:
1147         mthca_cmd_cleanup(mdev);
1148
1149 err_free_dev:
1150         ib_dealloc_device(&mdev->ib_dev);
1151
1152 err_free_res:
1153         pci_release_regions(pdev);
1154
1155 err_disable_pdev:
1156         pci_disable_device(pdev);
1157         pci_set_drvdata(pdev, NULL);
1158         return err;
1159 }
1160
1161 static void __mthca_remove_one(struct pci_dev *pdev)
1162 {
1163         struct mthca_dev *mdev = pci_get_drvdata(pdev);
1164         u8 status;
1165         int p;
1166
1167         if (mdev) {
1168                 mthca_free_agents(mdev);
1169                 mthca_unregister_device(mdev);
1170
1171                 for (p = 1; p <= mdev->limits.num_ports; ++p)
1172                         mthca_CLOSE_IB(mdev, p, &status);
1173
1174                 mthca_cleanup_mcg_table(mdev);
1175                 mthca_cleanup_av_table(mdev);
1176                 mthca_cleanup_qp_table(mdev);
1177                 mthca_cleanup_srq_table(mdev);
1178                 mthca_cleanup_cq_table(mdev);
1179                 mthca_cmd_use_polling(mdev);
1180                 mthca_cleanup_eq_table(mdev);
1181
1182                 mthca_pd_free(mdev, &mdev->driver_pd);
1183
1184                 mthca_cleanup_mr_table(mdev);
1185                 mthca_cleanup_pd_table(mdev);
1186
1187                 iounmap(mdev->kar);
1188                 mthca_uar_free(mdev, &mdev->driver_uar);
1189                 mthca_cleanup_uar_table(mdev);
1190                 mthca_close_hca(mdev);
1191                 mthca_cmd_cleanup(mdev);
1192
1193                 if (mdev->mthca_flags & MTHCA_FLAG_MSI_X)
1194                         pci_disable_msix(pdev);
1195
1196                 ib_dealloc_device(&mdev->ib_dev);
1197                 pci_release_regions(pdev);
1198                 pci_disable_device(pdev);
1199                 pci_set_drvdata(pdev, NULL);
1200         }
1201 }
1202
1203 int __mthca_restart_one(struct pci_dev *pdev)
1204 {
1205         struct mthca_dev *mdev;
1206         int hca_type;
1207
1208         mdev = pci_get_drvdata(pdev);
1209         if (!mdev)
1210                 return -ENODEV;
1211         hca_type = mdev->hca_type;
1212         __mthca_remove_one(pdev);
1213         return __mthca_init_one(pdev, hca_type);
1214 }
1215
1216 static int __devinit mthca_init_one(struct pci_dev *pdev,
1217                                     const struct pci_device_id *id)
1218 {
1219         static int mthca_version_printed = 0;
1220         int ret;
1221
1222         mutex_lock(&mthca_device_mutex);
1223
1224         if (!mthca_version_printed) {
1225                 printk(KERN_INFO "%s", mthca_version);
1226                 ++mthca_version_printed;
1227         }
1228
1229         if (id->driver_data >= ARRAY_SIZE(mthca_hca_table)) {
1230                 printk(KERN_ERR PFX "%s has invalid driver data %jx\n",
1231                        pci_name(pdev), (uintmax_t)id->driver_data);
1232                 mutex_unlock(&mthca_device_mutex);
1233                 return -ENODEV;
1234         }
1235
1236         ret = __mthca_init_one(pdev, id->driver_data);
1237
1238         mutex_unlock(&mthca_device_mutex);
1239
1240         return ret;
1241 }
1242
1243 static void __devexit mthca_remove_one(struct pci_dev *pdev)
1244 {
1245         mutex_lock(&mthca_device_mutex);
1246         __mthca_remove_one(pdev);
1247         mutex_unlock(&mthca_device_mutex);
1248 }
1249
1250 static struct pci_device_id mthca_pci_table[] = {
1251         { PCI_DEVICE(PCI_VENDOR_ID_MELLANOX, PCI_DEVICE_ID_MELLANOX_TAVOR),
1252           .driver_data = TAVOR },
1253         { PCI_DEVICE(PCI_VENDOR_ID_TOPSPIN, PCI_DEVICE_ID_MELLANOX_TAVOR),
1254           .driver_data = TAVOR },
1255         { PCI_DEVICE(PCI_VENDOR_ID_MELLANOX, PCI_DEVICE_ID_MELLANOX_ARBEL_COMPAT),
1256           .driver_data = ARBEL_COMPAT },
1257         { PCI_DEVICE(PCI_VENDOR_ID_TOPSPIN, PCI_DEVICE_ID_MELLANOX_ARBEL_COMPAT),
1258           .driver_data = ARBEL_COMPAT },
1259         { PCI_DEVICE(PCI_VENDOR_ID_MELLANOX, PCI_DEVICE_ID_MELLANOX_ARBEL),
1260           .driver_data = ARBEL_NATIVE },
1261         { PCI_DEVICE(PCI_VENDOR_ID_TOPSPIN, PCI_DEVICE_ID_MELLANOX_ARBEL),
1262           .driver_data = ARBEL_NATIVE },
1263         { PCI_DEVICE(PCI_VENDOR_ID_MELLANOX, PCI_DEVICE_ID_MELLANOX_SINAI),
1264           .driver_data = SINAI },
1265         { PCI_DEVICE(PCI_VENDOR_ID_TOPSPIN, PCI_DEVICE_ID_MELLANOX_SINAI),
1266           .driver_data = SINAI },
1267         { PCI_DEVICE(PCI_VENDOR_ID_MELLANOX, PCI_DEVICE_ID_MELLANOX_SINAI_OLD),
1268           .driver_data = SINAI },
1269         { PCI_DEVICE(PCI_VENDOR_ID_TOPSPIN, PCI_DEVICE_ID_MELLANOX_SINAI_OLD),
1270           .driver_data = SINAI },
1271         { 0, }
1272 };
1273
1274 MODULE_DEVICE_TABLE(pci, mthca_pci_table);
1275
1276 static struct pci_driver mthca_driver = {
1277         .name           = DRV_NAME,
1278         .id_table       = mthca_pci_table,
1279         .probe          = mthca_init_one,
1280         .remove         = __devexit_p(mthca_remove_one)
1281 };
1282
1283 static void __init __mthca_check_profile_val(const char *name, int *pval,
1284                                              int pval_default)
1285 {
1286         /* value must be positive and power of 2 */
1287         int old_pval = *pval;
1288
1289         if (old_pval <= 0)
1290                 *pval = pval_default;
1291         else
1292                 *pval = roundup_pow_of_two(old_pval);
1293
1294         if (old_pval != *pval) {
1295                 printk(KERN_WARNING PFX "Invalid value %d for %s in module parameter.\n",
1296                        old_pval, name);
1297                 printk(KERN_WARNING PFX "Corrected %s to %d.\n", name, *pval);
1298         }
1299 }
1300
1301 #define mthca_check_profile_val(name, default)                          \
1302         __mthca_check_profile_val(#name, &hca_profile.name, default)
1303
1304 static void __init mthca_validate_profile(void)
1305 {
1306         mthca_check_profile_val(num_qp,            MTHCA_DEFAULT_NUM_QP);
1307         mthca_check_profile_val(rdb_per_qp,        MTHCA_DEFAULT_RDB_PER_QP);
1308         mthca_check_profile_val(num_cq,            MTHCA_DEFAULT_NUM_CQ);
1309         mthca_check_profile_val(num_mcg,           MTHCA_DEFAULT_NUM_MCG);
1310         mthca_check_profile_val(num_mpt,           MTHCA_DEFAULT_NUM_MPT);
1311         mthca_check_profile_val(num_mtt,           MTHCA_DEFAULT_NUM_MTT);
1312         mthca_check_profile_val(num_udav,          MTHCA_DEFAULT_NUM_UDAV);
1313         mthca_check_profile_val(fmr_reserved_mtts, MTHCA_DEFAULT_NUM_RESERVED_MTTS);
1314
1315         if (hca_profile.fmr_reserved_mtts >= hca_profile.num_mtt) {
1316                 printk(KERN_WARNING PFX "Invalid fmr_reserved_mtts module parameter %d.\n",
1317                        hca_profile.fmr_reserved_mtts);
1318                 printk(KERN_WARNING PFX "(Must be smaller than num_mtt %d)\n",
1319                        hca_profile.num_mtt);
1320                 hca_profile.fmr_reserved_mtts = hca_profile.num_mtt / 2;
1321                 printk(KERN_WARNING PFX "Corrected fmr_reserved_mtts to %d.\n",
1322                        hca_profile.fmr_reserved_mtts);
1323         }
1324         if (log_mtts_per_seg == 0)
1325                 log_mtts_per_seg = ilog2(MTHCA_MTT_SEG_SIZE / 8);
1326         if ((log_mtts_per_seg < 1) || (log_mtts_per_seg > 5)) {
1327                 printk(KERN_WARNING PFX "bad log_mtts_per_seg (%d). Using default - %d\n",
1328                        log_mtts_per_seg, ilog2(MTHCA_MTT_SEG_SIZE / 8));
1329                 log_mtts_per_seg = ilog2(MTHCA_MTT_SEG_SIZE / 8);
1330         }
1331 }
1332
1333 static int __init mthca_init(void)
1334 {
1335         int ret;
1336
1337         mthca_validate_profile();
1338
1339         ret = mthca_catas_init();
1340         if (ret)
1341                 return ret;
1342
1343         ret = pci_register_driver(&mthca_driver);
1344         if (ret < 0) {
1345                 mthca_catas_cleanup();
1346                 return ret;
1347         }
1348
1349         return 0;
1350 }
1351
1352 static void __exit mthca_cleanup(void)
1353 {
1354         pci_unregister_driver(&mthca_driver);
1355         mthca_catas_cleanup();
1356 }
1357
1358 module_init_order(mthca_init, SI_ORDER_MIDDLE);
1359 module_exit(mthca_cleanup);