[PATCH] IB/mthca: fix MTT allocation in mem-free mode
[safe/jmp/linux-2.6] / drivers / infiniband / hw / mthca / mthca_main.c
1 /*
2  * Copyright (c) 2004, 2005 Topspin Communications.  All rights reserved.
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the
8  * OpenIB.org BSD license below:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *        copyright notice, this list of conditions and the following
16  *        disclaimer.
17  *
18  *      - Redistributions in binary form must reproduce the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer in the documentation and/or other materials
21  *        provided with the distribution.
22  *
23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30  * SOFTWARE.
31  *
32  * $Id: mthca_main.c 1396 2004-12-28 04:10:27Z roland $
33  */
34
35 #include <linux/config.h>
36 #include <linux/version.h>
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_PCI_MSI
55
56 static int msi_x = 0;
57 module_param(msi_x, int, 0444);
58 MODULE_PARM_DESC(msi_x, "attempt to use MSI-X if nonzero");
59
60 static int msi = 0;
61 module_param(msi, int, 0444);
62 MODULE_PARM_DESC(msi, "attempt to use MSI if nonzero");
63
64 #else /* CONFIG_PCI_MSI */
65
66 #define msi_x (0)
67 #define msi   (0)
68
69 #endif /* CONFIG_PCI_MSI */
70
71 static const char mthca_version[] __devinitdata =
72         "ib_mthca: Mellanox InfiniBand HCA driver v"
73         DRV_VERSION " (" DRV_RELDATE ")\n";
74
75 static struct mthca_profile default_profile = {
76         .num_qp     = 1 << 16,
77         .rdb_per_qp = 4,
78         .num_cq     = 1 << 16,
79         .num_mcg    = 1 << 13,
80         .num_mpt    = 1 << 17,
81         .num_mtt    = 1 << 20,
82         .num_udav   = 1 << 15,  /* Tavor only */
83         .uarc_size  = 1 << 18,  /* Arbel only */
84 };
85
86 static int __devinit mthca_tune_pci(struct mthca_dev *mdev)
87 {
88         int cap;
89         u16 val;
90
91         /* First try to max out Read Byte Count */
92         cap = pci_find_capability(mdev->pdev, PCI_CAP_ID_PCIX);
93         if (cap) {
94                 if (pci_read_config_word(mdev->pdev, cap + PCI_X_CMD, &val)) {
95                         mthca_err(mdev, "Couldn't read PCI-X command register, "
96                                   "aborting.\n");
97                         return -ENODEV;
98                 }
99                 val = (val & ~PCI_X_CMD_MAX_READ) | (3 << 2);
100                 if (pci_write_config_word(mdev->pdev, cap + PCI_X_CMD, val)) {
101                         mthca_err(mdev, "Couldn't write PCI-X command register, "
102                                   "aborting.\n");
103                         return -ENODEV;
104                 }
105         } else if (mdev->hca_type == TAVOR)
106                 mthca_info(mdev, "No PCI-X capability, not setting RBC.\n");
107
108         cap = pci_find_capability(mdev->pdev, PCI_CAP_ID_EXP);
109         if (cap) {
110                 if (pci_read_config_word(mdev->pdev, cap + PCI_EXP_DEVCTL, &val)) {
111                         mthca_err(mdev, "Couldn't read PCI Express device control "
112                                   "register, aborting.\n");
113                         return -ENODEV;
114                 }
115                 val = (val & ~PCI_EXP_DEVCTL_READRQ) | (5 << 12);
116                 if (pci_write_config_word(mdev->pdev, cap + PCI_EXP_DEVCTL, val)) {
117                         mthca_err(mdev, "Couldn't write PCI Express device control "
118                                   "register, aborting.\n");
119                         return -ENODEV;
120                 }
121         } else if (mdev->hca_type == ARBEL_NATIVE ||
122                    mdev->hca_type == ARBEL_COMPAT)
123                 mthca_info(mdev, "No PCI Express capability, "
124                            "not setting Max Read Request Size.\n");
125
126         return 0;
127 }
128
129 static int __devinit mthca_dev_lim(struct mthca_dev *mdev, struct mthca_dev_lim *dev_lim)
130 {
131         int err;
132         u8 status;
133
134         err = mthca_QUERY_DEV_LIM(mdev, dev_lim, &status);
135         if (err) {
136                 mthca_err(mdev, "QUERY_DEV_LIM command failed, aborting.\n");
137                 return err;
138         }
139         if (status) {
140                 mthca_err(mdev, "QUERY_DEV_LIM returned status 0x%02x, "
141                           "aborting.\n", status);
142                 return -EINVAL;
143         }
144         if (dev_lim->min_page_sz > PAGE_SIZE) {
145                 mthca_err(mdev, "HCA minimum page size of %d bigger than "
146                           "kernel PAGE_SIZE of %ld, aborting.\n",
147                           dev_lim->min_page_sz, PAGE_SIZE);
148                 return -ENODEV;
149         }
150         if (dev_lim->num_ports > MTHCA_MAX_PORTS) {
151                 mthca_err(mdev, "HCA has %d ports, but we only support %d, "
152                           "aborting.\n",
153                           dev_lim->num_ports, MTHCA_MAX_PORTS);
154                 return -ENODEV;
155         }
156
157         mdev->limits.num_ports          = dev_lim->num_ports;
158         mdev->limits.vl_cap             = dev_lim->max_vl;
159         mdev->limits.mtu_cap            = dev_lim->max_mtu;
160         mdev->limits.gid_table_len      = dev_lim->max_gids;
161         mdev->limits.pkey_table_len     = dev_lim->max_pkeys;
162         mdev->limits.local_ca_ack_delay = dev_lim->local_ca_ack_delay;
163         mdev->limits.max_sg             = dev_lim->max_sg;
164         mdev->limits.reserved_qps       = dev_lim->reserved_qps;
165         mdev->limits.reserved_srqs      = dev_lim->reserved_srqs;
166         mdev->limits.reserved_eecs      = dev_lim->reserved_eecs;
167         mdev->limits.reserved_cqs       = dev_lim->reserved_cqs;
168         mdev->limits.reserved_eqs       = dev_lim->reserved_eqs;
169         mdev->limits.reserved_mtts      = dev_lim->reserved_mtts;
170         mdev->limits.reserved_mrws      = dev_lim->reserved_mrws;
171         mdev->limits.reserved_uars      = dev_lim->reserved_uars;
172         mdev->limits.reserved_pds       = dev_lim->reserved_pds;
173
174         /* IB_DEVICE_RESIZE_MAX_WR not supported by driver.
175            May be doable since hardware supports it for SRQ.
176
177            IB_DEVICE_N_NOTIFY_CQ is supported by hardware but not by driver.
178
179            IB_DEVICE_SRQ_RESIZE is supported by hardware but SRQ is not
180            supported by driver. */
181         mdev->device_cap_flags = IB_DEVICE_CHANGE_PHY_PORT |
182                 IB_DEVICE_PORT_ACTIVE_EVENT |
183                 IB_DEVICE_SYS_IMAGE_GUID |
184                 IB_DEVICE_RC_RNR_NAK_GEN;
185
186         if (dev_lim->flags & DEV_LIM_FLAG_BAD_PKEY_CNTR)
187                 mdev->device_cap_flags |= IB_DEVICE_BAD_PKEY_CNTR;
188
189         if (dev_lim->flags & DEV_LIM_FLAG_BAD_QKEY_CNTR)
190                 mdev->device_cap_flags |= IB_DEVICE_BAD_QKEY_CNTR;
191
192         if (dev_lim->flags & DEV_LIM_FLAG_RAW_MULTI)
193                 mdev->device_cap_flags |= IB_DEVICE_RAW_MULTI;
194
195         if (dev_lim->flags & DEV_LIM_FLAG_AUTO_PATH_MIG)
196                 mdev->device_cap_flags |= IB_DEVICE_AUTO_PATH_MIG;
197
198         if (dev_lim->flags & DEV_LIM_FLAG_UD_AV_PORT_ENFORCE)
199                 mdev->device_cap_flags |= IB_DEVICE_UD_AV_PORT_ENFORCE;
200
201         if (dev_lim->flags & DEV_LIM_FLAG_SRQ)
202                 mdev->mthca_flags |= MTHCA_FLAG_SRQ;
203
204         return 0;
205 }
206
207 static int __devinit mthca_init_tavor(struct mthca_dev *mdev)
208 {
209         u8 status;
210         int err;
211         struct mthca_dev_lim        dev_lim;
212         struct mthca_profile        profile;
213         struct mthca_init_hca_param init_hca;
214         struct mthca_adapter        adapter;
215
216         err = mthca_SYS_EN(mdev, &status);
217         if (err) {
218                 mthca_err(mdev, "SYS_EN command failed, aborting.\n");
219                 return err;
220         }
221         if (status) {
222                 mthca_err(mdev, "SYS_EN returned status 0x%02x, "
223                           "aborting.\n", status);
224                 return -EINVAL;
225         }
226
227         err = mthca_QUERY_FW(mdev, &status);
228         if (err) {
229                 mthca_err(mdev, "QUERY_FW command failed, aborting.\n");
230                 goto err_disable;
231         }
232         if (status) {
233                 mthca_err(mdev, "QUERY_FW returned status 0x%02x, "
234                           "aborting.\n", status);
235                 err = -EINVAL;
236                 goto err_disable;
237         }
238         err = mthca_QUERY_DDR(mdev, &status);
239         if (err) {
240                 mthca_err(mdev, "QUERY_DDR command failed, aborting.\n");
241                 goto err_disable;
242         }
243         if (status) {
244                 mthca_err(mdev, "QUERY_DDR returned status 0x%02x, "
245                           "aborting.\n", status);
246                 err = -EINVAL;
247                 goto err_disable;
248         }
249
250         err = mthca_dev_lim(mdev, &dev_lim);
251
252         profile = default_profile;
253         profile.num_uar   = dev_lim.uar_size / PAGE_SIZE;
254         profile.uarc_size = 0;
255
256         err = mthca_make_profile(mdev, &profile, &dev_lim, &init_hca);
257         if (err < 0)
258                 goto err_disable;
259
260         err = mthca_INIT_HCA(mdev, &init_hca, &status);
261         if (err) {
262                 mthca_err(mdev, "INIT_HCA command failed, aborting.\n");
263                 goto err_disable;
264         }
265         if (status) {
266                 mthca_err(mdev, "INIT_HCA returned status 0x%02x, "
267                           "aborting.\n", status);
268                 err = -EINVAL;
269                 goto err_disable;
270         }
271
272         err = mthca_QUERY_ADAPTER(mdev, &adapter, &status);
273         if (err) {
274                 mthca_err(mdev, "QUERY_ADAPTER command failed, aborting.\n");
275                 goto err_close;
276         }
277         if (status) {
278                 mthca_err(mdev, "QUERY_ADAPTER returned status 0x%02x, "
279                           "aborting.\n", status);
280                 err = -EINVAL;
281                 goto err_close;
282         }
283
284         mdev->eq_table.inta_pin = adapter.inta_pin;
285         mdev->rev_id            = adapter.revision_id;
286
287         return 0;
288
289 err_close:
290         mthca_CLOSE_HCA(mdev, 0, &status);
291
292 err_disable:
293         mthca_SYS_DIS(mdev, &status);
294
295         return err;
296 }
297
298 static int __devinit mthca_load_fw(struct mthca_dev *mdev)
299 {
300         u8 status;
301         int err;
302
303         /* FIXME: use HCA-attached memory for FW if present */
304
305         mdev->fw.arbel.fw_icm =
306                 mthca_alloc_icm(mdev, mdev->fw.arbel.fw_pages,
307                                 GFP_HIGHUSER | __GFP_NOWARN);
308         if (!mdev->fw.arbel.fw_icm) {
309                 mthca_err(mdev, "Couldn't allocate FW area, aborting.\n");
310                 return -ENOMEM;
311         }
312
313         err = mthca_MAP_FA(mdev, mdev->fw.arbel.fw_icm, &status);
314         if (err) {
315                 mthca_err(mdev, "MAP_FA command failed, aborting.\n");
316                 goto err_free;
317         }
318         if (status) {
319                 mthca_err(mdev, "MAP_FA returned status 0x%02x, aborting.\n", status);
320                 err = -EINVAL;
321                 goto err_free;
322         }
323         err = mthca_RUN_FW(mdev, &status);
324         if (err) {
325                 mthca_err(mdev, "RUN_FW command failed, aborting.\n");
326                 goto err_unmap_fa;
327         }
328         if (status) {
329                 mthca_err(mdev, "RUN_FW returned status 0x%02x, aborting.\n", status);
330                 err = -EINVAL;
331                 goto err_unmap_fa;
332         }
333
334         return 0;
335
336 err_unmap_fa:
337         mthca_UNMAP_FA(mdev, &status);
338
339 err_free:
340         mthca_free_icm(mdev, mdev->fw.arbel.fw_icm);
341         return err;
342 }
343
344 static int __devinit mthca_init_icm(struct mthca_dev *mdev,
345                                     struct mthca_dev_lim *dev_lim,
346                                     struct mthca_init_hca_param *init_hca,
347                                     u64 icm_size)
348 {
349         u64 aux_pages;
350         u8 status;
351         int err;
352
353         err = mthca_SET_ICM_SIZE(mdev, icm_size, &aux_pages, &status);
354         if (err) {
355                 mthca_err(mdev, "SET_ICM_SIZE command failed, aborting.\n");
356                 return err;
357         }
358         if (status) {
359                 mthca_err(mdev, "SET_ICM_SIZE returned status 0x%02x, "
360                           "aborting.\n", status);
361                 return -EINVAL;
362         }
363
364         mthca_dbg(mdev, "%lld KB of HCA context requires %lld KB aux memory.\n",
365                   (unsigned long long) icm_size >> 10,
366                   (unsigned long long) aux_pages << 2);
367
368         mdev->fw.arbel.aux_icm = mthca_alloc_icm(mdev, aux_pages,
369                                                  GFP_HIGHUSER | __GFP_NOWARN);
370         if (!mdev->fw.arbel.aux_icm) {
371                 mthca_err(mdev, "Couldn't allocate aux memory, aborting.\n");
372                 return -ENOMEM;
373         }
374
375         err = mthca_MAP_ICM_AUX(mdev, mdev->fw.arbel.aux_icm, &status);
376         if (err) {
377                 mthca_err(mdev, "MAP_ICM_AUX command failed, aborting.\n");
378                 goto err_free_aux;
379         }
380         if (status) {
381                 mthca_err(mdev, "MAP_ICM_AUX returned status 0x%02x, aborting.\n", status);
382                 err = -EINVAL;
383                 goto err_free_aux;
384         }
385
386         err = mthca_map_eq_icm(mdev, init_hca->eqc_base);
387         if (err) {
388                 mthca_err(mdev, "Failed to map EQ context memory, aborting.\n");
389                 goto err_unmap_aux;
390         }
391
392         mdev->mr_table.mtt_table = mthca_alloc_icm_table(mdev, init_hca->mtt_base,
393                                                          MTHCA_MTT_SEG_SIZE,
394                                                          mdev->limits.num_mtt_segs,
395                                                          mdev->limits.reserved_mtts, 1);
396         if (!mdev->mr_table.mtt_table) {
397                 mthca_err(mdev, "Failed to map MTT context memory, aborting.\n");
398                 err = -ENOMEM;
399                 goto err_unmap_eq;
400         }
401
402         mdev->mr_table.mpt_table = mthca_alloc_icm_table(mdev, init_hca->mpt_base,
403                                                          dev_lim->mpt_entry_sz,
404                                                          mdev->limits.num_mpts,
405                                                          mdev->limits.reserved_mrws, 1);
406         if (!mdev->mr_table.mpt_table) {
407                 mthca_err(mdev, "Failed to map MPT context memory, aborting.\n");
408                 err = -ENOMEM;
409                 goto err_unmap_mtt;
410         }
411
412         mdev->qp_table.qp_table = mthca_alloc_icm_table(mdev, init_hca->qpc_base,
413                                                         dev_lim->qpc_entry_sz,
414                                                         mdev->limits.num_qps,
415                                                         mdev->limits.reserved_qps, 0);
416         if (!mdev->qp_table.qp_table) {
417                 mthca_err(mdev, "Failed to map QP context memory, aborting.\n");
418                 err = -ENOMEM;
419                 goto err_unmap_mpt;
420         }
421
422         mdev->qp_table.eqp_table = mthca_alloc_icm_table(mdev, init_hca->eqpc_base,
423                                                          dev_lim->eqpc_entry_sz,
424                                                          mdev->limits.num_qps,
425                                                          mdev->limits.reserved_qps, 0);
426         if (!mdev->qp_table.eqp_table) {
427                 mthca_err(mdev, "Failed to map EQP context memory, aborting.\n");
428                 err = -ENOMEM;
429                 goto err_unmap_qp;
430         }
431
432         mdev->cq_table.table = mthca_alloc_icm_table(mdev, init_hca->cqc_base,
433                                                      dev_lim->cqc_entry_sz,
434                                                      mdev->limits.num_cqs,
435                                                      mdev->limits.reserved_cqs, 0);
436         if (!mdev->cq_table.table) {
437                 mthca_err(mdev, "Failed to map CQ context memory, aborting.\n");
438                 err = -ENOMEM;
439                 goto err_unmap_eqp;
440         }
441
442         /*
443          * It's not strictly required, but for simplicity just map the
444          * whole multicast group table now.  The table isn't very big
445          * and it's a lot easier than trying to track ref counts.
446          */
447         mdev->mcg_table.table = mthca_alloc_icm_table(mdev, init_hca->mc_base,
448                                                       MTHCA_MGM_ENTRY_SIZE,
449                                                       mdev->limits.num_mgms +
450                                                       mdev->limits.num_amgms,
451                                                       mdev->limits.num_mgms +
452                                                       mdev->limits.num_amgms,
453                                                       0);
454         if (!mdev->mcg_table.table) {
455                 mthca_err(mdev, "Failed to map MCG context memory, aborting.\n");
456                 err = -ENOMEM;
457                 goto err_unmap_cq;
458         }
459
460         return 0;
461
462 err_unmap_cq:
463         mthca_free_icm_table(mdev, mdev->cq_table.table);
464
465 err_unmap_eqp:
466         mthca_free_icm_table(mdev, mdev->qp_table.eqp_table);
467
468 err_unmap_qp:
469         mthca_free_icm_table(mdev, mdev->qp_table.qp_table);
470
471 err_unmap_mpt:
472         mthca_free_icm_table(mdev, mdev->mr_table.mpt_table);
473
474 err_unmap_mtt:
475         mthca_free_icm_table(mdev, mdev->mr_table.mtt_table);
476
477 err_unmap_eq:
478         mthca_unmap_eq_icm(mdev);
479
480 err_unmap_aux:
481         mthca_UNMAP_ICM_AUX(mdev, &status);
482
483 err_free_aux:
484         mthca_free_icm(mdev, mdev->fw.arbel.aux_icm);
485
486         return err;
487 }
488
489 static int __devinit mthca_init_arbel(struct mthca_dev *mdev)
490 {
491         struct mthca_dev_lim        dev_lim;
492         struct mthca_profile        profile;
493         struct mthca_init_hca_param init_hca;
494         struct mthca_adapter        adapter;
495         u64 icm_size;
496         u8 status;
497         int err;
498
499         err = mthca_QUERY_FW(mdev, &status);
500         if (err) {
501                 mthca_err(mdev, "QUERY_FW command failed, aborting.\n");
502                 return err;
503         }
504         if (status) {
505                 mthca_err(mdev, "QUERY_FW returned status 0x%02x, "
506                           "aborting.\n", status);
507                 return -EINVAL;
508         }
509
510         err = mthca_ENABLE_LAM(mdev, &status);
511         if (err) {
512                 mthca_err(mdev, "ENABLE_LAM command failed, aborting.\n");
513                 return err;
514         }
515         if (status == MTHCA_CMD_STAT_LAM_NOT_PRE) {
516                 mthca_dbg(mdev, "No HCA-attached memory (running in MemFree mode)\n");
517                 mdev->mthca_flags |= MTHCA_FLAG_NO_LAM;
518         } else if (status) {
519                 mthca_err(mdev, "ENABLE_LAM returned status 0x%02x, "
520                           "aborting.\n", status);
521                 return -EINVAL;
522         }
523
524         err = mthca_load_fw(mdev);
525         if (err) {
526                 mthca_err(mdev, "Failed to start FW, aborting.\n");
527                 goto err_disable;
528         }
529
530         err = mthca_dev_lim(mdev, &dev_lim);
531         if (err) {
532                 mthca_err(mdev, "QUERY_DEV_LIM command failed, aborting.\n");
533                 goto err_stop_fw;
534         }
535
536         profile = default_profile;
537         profile.num_uar  = dev_lim.uar_size / PAGE_SIZE;
538         profile.num_udav = 0;
539
540         icm_size = mthca_make_profile(mdev, &profile, &dev_lim, &init_hca);
541         if ((int) icm_size < 0) {
542                 err = icm_size;
543                 goto err_stop_fw;
544         }
545
546         err = mthca_init_icm(mdev, &dev_lim, &init_hca, icm_size);
547         if (err)
548                 goto err_stop_fw;
549
550         err = mthca_INIT_HCA(mdev, &init_hca, &status);
551         if (err) {
552                 mthca_err(mdev, "INIT_HCA command failed, aborting.\n");
553                 goto err_free_icm;
554         }
555         if (status) {
556                 mthca_err(mdev, "INIT_HCA returned status 0x%02x, "
557                           "aborting.\n", status);
558                 err = -EINVAL;
559                 goto err_free_icm;
560         }
561
562         err = mthca_QUERY_ADAPTER(mdev, &adapter, &status);
563         if (err) {
564                 mthca_err(mdev, "QUERY_ADAPTER command failed, aborting.\n");
565                 goto err_free_icm;
566         }
567         if (status) {
568                 mthca_err(mdev, "QUERY_ADAPTER returned status 0x%02x, "
569                           "aborting.\n", status);
570                 err = -EINVAL;
571                 goto err_free_icm;
572         }
573
574         mdev->eq_table.inta_pin = adapter.inta_pin;
575         mdev->rev_id            = adapter.revision_id;
576
577         return 0;
578
579 err_free_icm:
580         mthca_free_icm_table(mdev, mdev->cq_table.table);
581         mthca_free_icm_table(mdev, mdev->qp_table.eqp_table);
582         mthca_free_icm_table(mdev, mdev->qp_table.qp_table);
583         mthca_free_icm_table(mdev, mdev->mr_table.mpt_table);
584         mthca_free_icm_table(mdev, mdev->mr_table.mtt_table);
585         mthca_unmap_eq_icm(mdev);
586
587         mthca_UNMAP_ICM_AUX(mdev, &status);
588         mthca_free_icm(mdev, mdev->fw.arbel.aux_icm);
589
590 err_stop_fw:
591         mthca_UNMAP_FA(mdev, &status);
592         mthca_free_icm(mdev, mdev->fw.arbel.fw_icm);
593
594 err_disable:
595         if (!(mdev->mthca_flags & MTHCA_FLAG_NO_LAM))
596                 mthca_DISABLE_LAM(mdev, &status);
597
598         return err;
599 }
600
601 static int __devinit mthca_init_hca(struct mthca_dev *mdev)
602 {
603         if (mdev->hca_type == ARBEL_NATIVE)
604                 return mthca_init_arbel(mdev);
605         else
606                 return mthca_init_tavor(mdev);
607 }
608
609 static int __devinit mthca_setup_hca(struct mthca_dev *dev)
610 {
611         int err;
612         u8 status;
613
614         MTHCA_INIT_DOORBELL_LOCK(&dev->doorbell_lock);
615
616         err = mthca_init_uar_table(dev);
617         if (err) {
618                 mthca_err(dev, "Failed to initialize "
619                           "user access region table, aborting.\n");
620                 return err;
621         }
622
623         err = mthca_uar_alloc(dev, &dev->driver_uar);
624         if (err) {
625                 mthca_err(dev, "Failed to allocate driver access region, "
626                           "aborting.\n");
627                 goto err_uar_table_free;
628         }
629
630         dev->kar = ioremap(dev->driver_uar.pfn << PAGE_SHIFT, PAGE_SIZE);
631         if (!dev->kar) {
632                 mthca_err(dev, "Couldn't map kernel access region, "
633                           "aborting.\n");
634                 err = -ENOMEM;
635                 goto err_uar_free;
636         }
637
638         err = mthca_init_pd_table(dev);
639         if (err) {
640                 mthca_err(dev, "Failed to initialize "
641                           "protection domain table, aborting.\n");
642                 goto err_kar_unmap;
643         }
644
645         err = mthca_init_mr_table(dev);
646         if (err) {
647                 mthca_err(dev, "Failed to initialize "
648                           "memory region table, aborting.\n");
649                 goto err_pd_table_free;
650         }
651
652         err = mthca_pd_alloc(dev, &dev->driver_pd);
653         if (err) {
654                 mthca_err(dev, "Failed to create driver PD, "
655                           "aborting.\n");
656                 goto err_mr_table_free;
657         }
658
659         err = mthca_init_eq_table(dev);
660         if (err) {
661                 mthca_err(dev, "Failed to initialize "
662                           "event queue table, aborting.\n");
663                 goto err_pd_free;
664         }
665
666         err = mthca_cmd_use_events(dev);
667         if (err) {
668                 mthca_err(dev, "Failed to switch to event-driven "
669                           "firmware commands, aborting.\n");
670                 goto err_eq_table_free;
671         }
672
673         err = mthca_NOP(dev, &status);
674         if (err || status) {
675                 mthca_err(dev, "NOP command failed to generate interrupt (IRQ %d), aborting.\n",
676                           dev->mthca_flags & MTHCA_FLAG_MSI_X ?
677                           dev->eq_table.eq[MTHCA_EQ_CMD].msi_x_vector :
678                           dev->pdev->irq);
679                 if (dev->mthca_flags & (MTHCA_FLAG_MSI | MTHCA_FLAG_MSI_X))
680                         mthca_err(dev, "Try again with MSI/MSI-X disabled.\n");
681                 else
682                         mthca_err(dev, "BIOS or ACPI interrupt routing problem?\n");
683
684                 goto err_cmd_poll;
685         }
686
687         mthca_dbg(dev, "NOP command IRQ test passed\n");
688
689         err = mthca_init_cq_table(dev);
690         if (err) {
691                 mthca_err(dev, "Failed to initialize "
692                           "completion queue table, aborting.\n");
693                 goto err_cmd_poll;
694         }
695
696         err = mthca_init_qp_table(dev);
697         if (err) {
698                 mthca_err(dev, "Failed to initialize "
699                           "queue pair table, aborting.\n");
700                 goto err_cq_table_free;
701         }
702
703         err = mthca_init_av_table(dev);
704         if (err) {
705                 mthca_err(dev, "Failed to initialize "
706                           "address vector table, aborting.\n");
707                 goto err_qp_table_free;
708         }
709
710         err = mthca_init_mcg_table(dev);
711         if (err) {
712                 mthca_err(dev, "Failed to initialize "
713                           "multicast group table, aborting.\n");
714                 goto err_av_table_free;
715         }
716
717         return 0;
718
719 err_av_table_free:
720         mthca_cleanup_av_table(dev);
721
722 err_qp_table_free:
723         mthca_cleanup_qp_table(dev);
724
725 err_cq_table_free:
726         mthca_cleanup_cq_table(dev);
727
728 err_cmd_poll:
729         mthca_cmd_use_polling(dev);
730
731 err_eq_table_free:
732         mthca_cleanup_eq_table(dev);
733
734 err_pd_free:
735         mthca_pd_free(dev, &dev->driver_pd);
736
737 err_mr_table_free:
738         mthca_cleanup_mr_table(dev);
739
740 err_pd_table_free:
741         mthca_cleanup_pd_table(dev);
742
743 err_kar_unmap:
744         iounmap(dev->kar);
745
746 err_uar_free:
747         mthca_uar_free(dev, &dev->driver_uar);
748
749 err_uar_table_free:
750         mthca_cleanup_uar_table(dev);
751         return err;
752 }
753
754 static int __devinit mthca_request_regions(struct pci_dev *pdev,
755                                            int ddr_hidden)
756 {
757         int err;
758
759         /*
760          * We can't just use pci_request_regions() because the MSI-X
761          * table is right in the middle of the first BAR.  If we did
762          * pci_request_region and grab all of the first BAR, then
763          * setting up MSI-X would fail, since the PCI core wants to do
764          * request_mem_region on the MSI-X vector table.
765          *
766          * So just request what we need right now, and request any
767          * other regions we need when setting up EQs.
768          */
769         if (!request_mem_region(pci_resource_start(pdev, 0) + MTHCA_HCR_BASE,
770                                 MTHCA_HCR_SIZE, DRV_NAME))
771                 return -EBUSY;
772
773         err = pci_request_region(pdev, 2, DRV_NAME);
774         if (err)
775                 goto err_bar2_failed;
776
777         if (!ddr_hidden) {
778                 err = pci_request_region(pdev, 4, DRV_NAME);
779                 if (err)
780                         goto err_bar4_failed;
781         }
782
783         return 0;
784
785 err_bar4_failed:
786         pci_release_region(pdev, 2);
787
788 err_bar2_failed:
789         release_mem_region(pci_resource_start(pdev, 0) + MTHCA_HCR_BASE,
790                            MTHCA_HCR_SIZE);
791
792         return err;
793 }
794
795 static void mthca_release_regions(struct pci_dev *pdev,
796                                   int ddr_hidden)
797 {
798         if (!ddr_hidden)
799                 pci_release_region(pdev, 4);
800
801         pci_release_region(pdev, 2);
802
803         release_mem_region(pci_resource_start(pdev, 0) + MTHCA_HCR_BASE,
804                            MTHCA_HCR_SIZE);
805 }
806
807 static int __devinit mthca_enable_msi_x(struct mthca_dev *mdev)
808 {
809         struct msix_entry entries[3];
810         int err;
811
812         entries[0].entry = 0;
813         entries[1].entry = 1;
814         entries[2].entry = 2;
815
816         err = pci_enable_msix(mdev->pdev, entries, ARRAY_SIZE(entries));
817         if (err) {
818                 if (err > 0)
819                         mthca_info(mdev, "Only %d MSI-X vectors available, "
820                                    "not using MSI-X\n", err);
821                 return err;
822         }
823
824         mdev->eq_table.eq[MTHCA_EQ_COMP ].msi_x_vector = entries[0].vector;
825         mdev->eq_table.eq[MTHCA_EQ_ASYNC].msi_x_vector = entries[1].vector;
826         mdev->eq_table.eq[MTHCA_EQ_CMD  ].msi_x_vector = entries[2].vector;
827
828         return 0;
829 }
830
831 static void mthca_close_hca(struct mthca_dev *mdev)
832 {
833         u8 status;
834
835         mthca_CLOSE_HCA(mdev, 0, &status);
836
837         if (mdev->hca_type == ARBEL_NATIVE) {
838                 mthca_free_icm_table(mdev, mdev->cq_table.table);
839                 mthca_free_icm_table(mdev, mdev->qp_table.eqp_table);
840                 mthca_free_icm_table(mdev, mdev->qp_table.qp_table);
841                 mthca_free_icm_table(mdev, mdev->mr_table.mpt_table);
842                 mthca_free_icm_table(mdev, mdev->mr_table.mtt_table);
843                 mthca_unmap_eq_icm(mdev);
844
845                 mthca_UNMAP_ICM_AUX(mdev, &status);
846                 mthca_free_icm(mdev, mdev->fw.arbel.aux_icm);
847
848                 mthca_UNMAP_FA(mdev, &status);
849                 mthca_free_icm(mdev, mdev->fw.arbel.fw_icm);
850
851                 if (!(mdev->mthca_flags & MTHCA_FLAG_NO_LAM))
852                         mthca_DISABLE_LAM(mdev, &status);
853         } else
854                 mthca_SYS_DIS(mdev, &status);
855 }
856
857 static int __devinit mthca_init_one(struct pci_dev *pdev,
858                                     const struct pci_device_id *id)
859 {
860         static int mthca_version_printed = 0;
861         static int mthca_memfree_warned = 0;
862         int ddr_hidden = 0;
863         int err;
864         struct mthca_dev *mdev;
865
866         if (!mthca_version_printed) {
867                 printk(KERN_INFO "%s", mthca_version);
868                 ++mthca_version_printed;
869         }
870
871         printk(KERN_INFO PFX "Initializing %s (%s)\n",
872                pci_pretty_name(pdev), pci_name(pdev));
873
874         err = pci_enable_device(pdev);
875         if (err) {
876                 dev_err(&pdev->dev, "Cannot enable PCI device, "
877                         "aborting.\n");
878                 return err;
879         }
880
881         /*
882          * Check for BARs.  We expect 0: 1MB, 2: 8MB, 4: DDR (may not
883          * be present)
884          */
885         if (!(pci_resource_flags(pdev, 0) & IORESOURCE_MEM) ||
886             pci_resource_len(pdev, 0) != 1 << 20) {
887                 dev_err(&pdev->dev, "Missing DCS, aborting.");
888                 err = -ENODEV;
889                 goto err_disable_pdev;
890         }
891         if (!(pci_resource_flags(pdev, 2) & IORESOURCE_MEM) ||
892             pci_resource_len(pdev, 2) != 1 << 23) {
893                 dev_err(&pdev->dev, "Missing UAR, aborting.");
894                 err = -ENODEV;
895                 goto err_disable_pdev;
896         }
897         if (!(pci_resource_flags(pdev, 4) & IORESOURCE_MEM))
898                 ddr_hidden = 1;
899
900         err = mthca_request_regions(pdev, ddr_hidden);
901         if (err) {
902                 dev_err(&pdev->dev, "Cannot obtain PCI resources, "
903                         "aborting.\n");
904                 goto err_disable_pdev;
905         }
906
907         pci_set_master(pdev);
908
909         err = pci_set_dma_mask(pdev, DMA_64BIT_MASK);
910         if (err) {
911                 dev_warn(&pdev->dev, "Warning: couldn't set 64-bit PCI DMA mask.\n");
912                 err = pci_set_dma_mask(pdev, DMA_32BIT_MASK);
913                 if (err) {
914                         dev_err(&pdev->dev, "Can't set PCI DMA mask, aborting.\n");
915                         goto err_free_res;
916                 }
917         }
918         err = pci_set_consistent_dma_mask(pdev, DMA_64BIT_MASK);
919         if (err) {
920                 dev_warn(&pdev->dev, "Warning: couldn't set 64-bit "
921                          "consistent PCI DMA mask.\n");
922                 err = pci_set_consistent_dma_mask(pdev, DMA_32BIT_MASK);
923                 if (err) {
924                         dev_err(&pdev->dev, "Can't set consistent PCI DMA mask, "
925                                 "aborting.\n");
926                         goto err_free_res;
927                 }
928         }
929
930         mdev = (struct mthca_dev *) ib_alloc_device(sizeof *mdev);
931         if (!mdev) {
932                 dev_err(&pdev->dev, "Device struct alloc failed, "
933                         "aborting.\n");
934                 err = -ENOMEM;
935                 goto err_free_res;
936         }
937
938         mdev->pdev     = pdev;
939         mdev->hca_type = id->driver_data;
940
941         if (mdev->hca_type == ARBEL_NATIVE && !mthca_memfree_warned++)
942                 mthca_warn(mdev, "Warning: native MT25208 mode support is incomplete.  "
943                            "Your HCA may not work properly.\n");
944
945         if (ddr_hidden)
946                 mdev->mthca_flags |= MTHCA_FLAG_DDR_HIDDEN;
947
948         /*
949          * Now reset the HCA before we touch the PCI capabilities or
950          * attempt a firmware command, since a boot ROM may have left
951          * the HCA in an undefined state.
952          */
953         err = mthca_reset(mdev);
954         if (err) {
955                 mthca_err(mdev, "Failed to reset HCA, aborting.\n");
956                 goto err_free_dev;
957         }
958
959         if (msi_x && !mthca_enable_msi_x(mdev))
960                 mdev->mthca_flags |= MTHCA_FLAG_MSI_X;
961         if (msi && !(mdev->mthca_flags & MTHCA_FLAG_MSI_X) &&
962             !pci_enable_msi(pdev))
963                 mdev->mthca_flags |= MTHCA_FLAG_MSI;
964
965         sema_init(&mdev->cmd.hcr_sem, 1);
966         sema_init(&mdev->cmd.poll_sem, 1);
967         mdev->cmd.use_events = 0;
968
969         mdev->hcr = ioremap(pci_resource_start(pdev, 0) + MTHCA_HCR_BASE, MTHCA_HCR_SIZE);
970         if (!mdev->hcr) {
971                 mthca_err(mdev, "Couldn't map command register, "
972                           "aborting.\n");
973                 err = -ENOMEM;
974                 goto err_free_dev;
975         }
976
977         err = mthca_tune_pci(mdev);
978         if (err)
979                 goto err_iounmap;
980
981         err = mthca_init_hca(mdev);
982         if (err)
983                 goto err_iounmap;
984
985         err = mthca_setup_hca(mdev);
986         if (err)
987                 goto err_close;
988
989         err = mthca_register_device(mdev);
990         if (err)
991                 goto err_cleanup;
992
993         err = mthca_create_agents(mdev);
994         if (err)
995                 goto err_unregister;
996
997         pci_set_drvdata(pdev, mdev);
998
999         return 0;
1000
1001 err_unregister:
1002         mthca_unregister_device(mdev);
1003
1004 err_cleanup:
1005         mthca_cleanup_mcg_table(mdev);
1006         mthca_cleanup_av_table(mdev);
1007         mthca_cleanup_qp_table(mdev);
1008         mthca_cleanup_cq_table(mdev);
1009         mthca_cmd_use_polling(mdev);
1010         mthca_cleanup_eq_table(mdev);
1011
1012         mthca_pd_free(mdev, &mdev->driver_pd);
1013
1014         mthca_cleanup_mr_table(mdev);
1015         mthca_cleanup_pd_table(mdev);
1016         mthca_cleanup_uar_table(mdev);
1017
1018 err_close:
1019         mthca_close_hca(mdev);
1020
1021 err_iounmap:
1022         iounmap(mdev->hcr);
1023
1024 err_free_dev:
1025         if (mdev->mthca_flags & MTHCA_FLAG_MSI_X)
1026                 pci_disable_msix(pdev);
1027         if (mdev->mthca_flags & MTHCA_FLAG_MSI)
1028                 pci_disable_msi(pdev);
1029
1030         ib_dealloc_device(&mdev->ib_dev);
1031
1032 err_free_res:
1033         mthca_release_regions(pdev, ddr_hidden);
1034
1035 err_disable_pdev:
1036         pci_disable_device(pdev);
1037         pci_set_drvdata(pdev, NULL);
1038         return err;
1039 }
1040
1041 static void __devexit mthca_remove_one(struct pci_dev *pdev)
1042 {
1043         struct mthca_dev *mdev = pci_get_drvdata(pdev);
1044         u8 status;
1045         int p;
1046
1047         if (mdev) {
1048                 mthca_free_agents(mdev);
1049                 mthca_unregister_device(mdev);
1050
1051                 for (p = 1; p <= mdev->limits.num_ports; ++p)
1052                         mthca_CLOSE_IB(mdev, p, &status);
1053
1054                 mthca_cleanup_mcg_table(mdev);
1055                 mthca_cleanup_av_table(mdev);
1056                 mthca_cleanup_qp_table(mdev);
1057                 mthca_cleanup_cq_table(mdev);
1058                 mthca_cmd_use_polling(mdev);
1059                 mthca_cleanup_eq_table(mdev);
1060
1061                 mthca_pd_free(mdev, &mdev->driver_pd);
1062
1063                 mthca_cleanup_mr_table(mdev);
1064                 mthca_cleanup_pd_table(mdev);
1065
1066                 iounmap(mdev->kar);
1067                 mthca_uar_free(mdev, &mdev->driver_uar);
1068                 mthca_cleanup_uar_table(mdev);
1069
1070                 mthca_close_hca(mdev);
1071
1072                 iounmap(mdev->hcr);
1073
1074                 if (mdev->mthca_flags & MTHCA_FLAG_MSI_X)
1075                         pci_disable_msix(pdev);
1076                 if (mdev->mthca_flags & MTHCA_FLAG_MSI)
1077                         pci_disable_msi(pdev);
1078
1079                 ib_dealloc_device(&mdev->ib_dev);
1080                 mthca_release_regions(pdev, mdev->mthca_flags &
1081                                       MTHCA_FLAG_DDR_HIDDEN);
1082                 pci_disable_device(pdev);
1083                 pci_set_drvdata(pdev, NULL);
1084         }
1085 }
1086
1087 static struct pci_device_id mthca_pci_table[] = {
1088         { PCI_DEVICE(PCI_VENDOR_ID_MELLANOX, PCI_DEVICE_ID_MELLANOX_TAVOR),
1089           .driver_data = TAVOR },
1090         { PCI_DEVICE(PCI_VENDOR_ID_TOPSPIN, PCI_DEVICE_ID_MELLANOX_TAVOR),
1091           .driver_data = TAVOR },
1092         { PCI_DEVICE(PCI_VENDOR_ID_MELLANOX, PCI_DEVICE_ID_MELLANOX_ARBEL_COMPAT),
1093           .driver_data = ARBEL_COMPAT },
1094         { PCI_DEVICE(PCI_VENDOR_ID_TOPSPIN, PCI_DEVICE_ID_MELLANOX_ARBEL_COMPAT),
1095           .driver_data = ARBEL_COMPAT },
1096         { PCI_DEVICE(PCI_VENDOR_ID_MELLANOX, PCI_DEVICE_ID_MELLANOX_ARBEL),
1097           .driver_data = ARBEL_NATIVE },
1098         { PCI_DEVICE(PCI_VENDOR_ID_TOPSPIN, PCI_DEVICE_ID_MELLANOX_ARBEL),
1099           .driver_data = ARBEL_NATIVE },
1100         { 0, }
1101 };
1102
1103 MODULE_DEVICE_TABLE(pci, mthca_pci_table);
1104
1105 static struct pci_driver mthca_driver = {
1106         .name           = "ib_mthca",
1107         .id_table       = mthca_pci_table,
1108         .probe          = mthca_init_one,
1109         .remove         = __devexit_p(mthca_remove_one)
1110 };
1111
1112 static int __init mthca_init(void)
1113 {
1114         int ret;
1115
1116         ret = pci_register_driver(&mthca_driver);
1117         return ret < 0 ? ret : 0;
1118 }
1119
1120 static void __exit mthca_cleanup(void)
1121 {
1122         pci_unregister_driver(&mthca_driver);
1123 }
1124
1125 module_init(mthca_init);
1126 module_exit(mthca_cleanup);