Pull vector-domain into release branch
[linux-2.6] / drivers / net / mlx4 / 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  * Copyright (c) 2006, 2007 Cisco Systems, Inc. All rights reserved.
6  *
7  * This software is available to you under a choice of one of two
8  * licenses.  You may choose to be licensed under the terms of the GNU
9  * General Public License (GPL) Version 2, available from the file
10  * COPYING in the main directory of this source tree, or the
11  * OpenIB.org BSD license below:
12  *
13  *     Redistribution and use in source and binary forms, with or
14  *     without modification, are permitted provided that the following
15  *     conditions are met:
16  *
17  *      - Redistributions of source code must retain the above
18  *        copyright notice, this list of conditions and the following
19  *        disclaimer.
20  *
21  *      - Redistributions in binary form must reproduce the above
22  *        copyright notice, this list of conditions and the following
23  *        disclaimer in the documentation and/or other materials
24  *        provided with the distribution.
25  *
26  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
27  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
28  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
29  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
30  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
31  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
32  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
33  * SOFTWARE.
34  */
35
36 #include <linux/module.h>
37 #include <linux/init.h>
38 #include <linux/errno.h>
39 #include <linux/pci.h>
40 #include <linux/dma-mapping.h>
41
42 #include <linux/mlx4/device.h>
43 #include <linux/mlx4/doorbell.h>
44
45 #include "mlx4.h"
46 #include "fw.h"
47 #include "icm.h"
48
49 MODULE_AUTHOR("Roland Dreier");
50 MODULE_DESCRIPTION("Mellanox ConnectX HCA low-level driver");
51 MODULE_LICENSE("Dual BSD/GPL");
52 MODULE_VERSION(DRV_VERSION);
53
54 #ifdef CONFIG_MLX4_DEBUG
55
56 int mlx4_debug_level = 0;
57 module_param_named(debug_level, mlx4_debug_level, int, 0644);
58 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
59
60 #endif /* CONFIG_MLX4_DEBUG */
61
62 #ifdef CONFIG_PCI_MSI
63
64 static int msi_x;
65 module_param(msi_x, int, 0444);
66 MODULE_PARM_DESC(msi_x, "attempt to use MSI-X if nonzero");
67
68 #else /* CONFIG_PCI_MSI */
69
70 #define msi_x (0)
71
72 #endif /* CONFIG_PCI_MSI */
73
74 static const char mlx4_version[] __devinitdata =
75         DRV_NAME ": Mellanox ConnectX core driver v"
76         DRV_VERSION " (" DRV_RELDATE ")\n";
77
78 static struct mlx4_profile default_profile = {
79         .num_qp         = 1 << 16,
80         .num_srq        = 1 << 16,
81         .rdmarc_per_qp  = 1 << 4,
82         .num_cq         = 1 << 16,
83         .num_mcg        = 1 << 13,
84         .num_mpt        = 1 << 17,
85         .num_mtt        = 1 << 20,
86 };
87
88 static int __devinit mlx4_dev_cap(struct mlx4_dev *dev, struct mlx4_dev_cap *dev_cap)
89 {
90         int err;
91         int i;
92
93         err = mlx4_QUERY_DEV_CAP(dev, dev_cap);
94         if (err) {
95                 mlx4_err(dev, "QUERY_DEV_CAP command failed, aborting.\n");
96                 return err;
97         }
98
99         if (dev_cap->min_page_sz > PAGE_SIZE) {
100                 mlx4_err(dev, "HCA minimum page size of %d bigger than "
101                          "kernel PAGE_SIZE of %ld, aborting.\n",
102                          dev_cap->min_page_sz, PAGE_SIZE);
103                 return -ENODEV;
104         }
105         if (dev_cap->num_ports > MLX4_MAX_PORTS) {
106                 mlx4_err(dev, "HCA has %d ports, but we only support %d, "
107                          "aborting.\n",
108                          dev_cap->num_ports, MLX4_MAX_PORTS);
109                 return -ENODEV;
110         }
111
112         if (dev_cap->uar_size > pci_resource_len(dev->pdev, 2)) {
113                 mlx4_err(dev, "HCA reported UAR size of 0x%x bigger than "
114                          "PCI resource 2 size of 0x%llx, aborting.\n",
115                          dev_cap->uar_size,
116                          (unsigned long long) pci_resource_len(dev->pdev, 2));
117                 return -ENODEV;
118         }
119
120         dev->caps.num_ports          = dev_cap->num_ports;
121         for (i = 1; i <= dev->caps.num_ports; ++i) {
122                 dev->caps.vl_cap[i]         = dev_cap->max_vl[i];
123                 dev->caps.mtu_cap[i]        = dev_cap->max_mtu[i];
124                 dev->caps.gid_table_len[i]  = dev_cap->max_gids[i];
125                 dev->caps.pkey_table_len[i] = dev_cap->max_pkeys[i];
126                 dev->caps.port_width_cap[i] = dev_cap->max_port_width[i];
127         }
128
129         dev->caps.num_uars           = dev_cap->uar_size / PAGE_SIZE;
130         dev->caps.local_ca_ack_delay = dev_cap->local_ca_ack_delay;
131         dev->caps.bf_reg_size        = dev_cap->bf_reg_size;
132         dev->caps.bf_regs_per_page   = dev_cap->bf_regs_per_page;
133         dev->caps.max_sq_sg          = dev_cap->max_sq_sg;
134         dev->caps.max_rq_sg          = dev_cap->max_rq_sg;
135         dev->caps.max_wqes           = dev_cap->max_qp_sz;
136         dev->caps.max_qp_init_rdma   = dev_cap->max_requester_per_qp;
137         dev->caps.reserved_qps       = dev_cap->reserved_qps;
138         dev->caps.max_srq_wqes       = dev_cap->max_srq_sz;
139         dev->caps.max_srq_sge        = dev_cap->max_rq_sg - 1;
140         dev->caps.reserved_srqs      = dev_cap->reserved_srqs;
141         dev->caps.max_sq_desc_sz     = dev_cap->max_sq_desc_sz;
142         dev->caps.max_rq_desc_sz     = dev_cap->max_rq_desc_sz;
143         dev->caps.num_qp_per_mgm     = MLX4_QP_PER_MGM;
144         /*
145          * Subtract 1 from the limit because we need to allocate a
146          * spare CQE so the HCA HW can tell the difference between an
147          * empty CQ and a full CQ.
148          */
149         dev->caps.max_cqes           = dev_cap->max_cq_sz - 1;
150         dev->caps.reserved_cqs       = dev_cap->reserved_cqs;
151         dev->caps.reserved_eqs       = dev_cap->reserved_eqs;
152         dev->caps.reserved_mtts      = dev_cap->reserved_mtts;
153         dev->caps.reserved_mrws      = dev_cap->reserved_mrws;
154         dev->caps.reserved_uars      = dev_cap->reserved_uars;
155         dev->caps.reserved_pds       = dev_cap->reserved_pds;
156         dev->caps.mtt_entry_sz       = MLX4_MTT_ENTRY_PER_SEG * dev_cap->mtt_entry_sz;
157         dev->caps.max_msg_sz         = dev_cap->max_msg_sz;
158         dev->caps.page_size_cap      = ~(u32) (dev_cap->min_page_sz - 1);
159         dev->caps.flags              = dev_cap->flags;
160         dev->caps.stat_rate_support  = dev_cap->stat_rate_support;
161
162         return 0;
163 }
164
165 static int __devinit mlx4_load_fw(struct mlx4_dev *dev)
166 {
167         struct mlx4_priv *priv = mlx4_priv(dev);
168         int err;
169
170         priv->fw.fw_icm = mlx4_alloc_icm(dev, priv->fw.fw_pages,
171                                          GFP_HIGHUSER | __GFP_NOWARN);
172         if (!priv->fw.fw_icm) {
173                 mlx4_err(dev, "Couldn't allocate FW area, aborting.\n");
174                 return -ENOMEM;
175         }
176
177         err = mlx4_MAP_FA(dev, priv->fw.fw_icm);
178         if (err) {
179                 mlx4_err(dev, "MAP_FA command failed, aborting.\n");
180                 goto err_free;
181         }
182
183         err = mlx4_RUN_FW(dev);
184         if (err) {
185                 mlx4_err(dev, "RUN_FW command failed, aborting.\n");
186                 goto err_unmap_fa;
187         }
188
189         return 0;
190
191 err_unmap_fa:
192         mlx4_UNMAP_FA(dev);
193
194 err_free:
195         mlx4_free_icm(dev, priv->fw.fw_icm);
196         return err;
197 }
198
199 static int __devinit mlx4_init_cmpt_table(struct mlx4_dev *dev, u64 cmpt_base,
200                                           int cmpt_entry_sz)
201 {
202         struct mlx4_priv *priv = mlx4_priv(dev);
203         int err;
204
205         err = mlx4_init_icm_table(dev, &priv->qp_table.cmpt_table,
206                                   cmpt_base +
207                                   ((u64) (MLX4_CMPT_TYPE_QP *
208                                           cmpt_entry_sz) << MLX4_CMPT_SHIFT),
209                                   cmpt_entry_sz, dev->caps.num_qps,
210                                   dev->caps.reserved_qps, 0);
211         if (err)
212                 goto err;
213
214         err = mlx4_init_icm_table(dev, &priv->srq_table.cmpt_table,
215                                   cmpt_base +
216                                   ((u64) (MLX4_CMPT_TYPE_SRQ *
217                                           cmpt_entry_sz) << MLX4_CMPT_SHIFT),
218                                   cmpt_entry_sz, dev->caps.num_srqs,
219                                   dev->caps.reserved_srqs, 0);
220         if (err)
221                 goto err_qp;
222
223         err = mlx4_init_icm_table(dev, &priv->cq_table.cmpt_table,
224                                   cmpt_base +
225                                   ((u64) (MLX4_CMPT_TYPE_CQ *
226                                           cmpt_entry_sz) << MLX4_CMPT_SHIFT),
227                                   cmpt_entry_sz, dev->caps.num_cqs,
228                                   dev->caps.reserved_cqs, 0);
229         if (err)
230                 goto err_srq;
231
232         err = mlx4_init_icm_table(dev, &priv->eq_table.cmpt_table,
233                                   cmpt_base +
234                                   ((u64) (MLX4_CMPT_TYPE_EQ *
235                                           cmpt_entry_sz) << MLX4_CMPT_SHIFT),
236                                   cmpt_entry_sz,
237                                   roundup_pow_of_two(MLX4_NUM_EQ +
238                                                      dev->caps.reserved_eqs),
239                                   MLX4_NUM_EQ + dev->caps.reserved_eqs, 0);
240         if (err)
241                 goto err_cq;
242
243         return 0;
244
245 err_cq:
246         mlx4_cleanup_icm_table(dev, &priv->cq_table.cmpt_table);
247
248 err_srq:
249         mlx4_cleanup_icm_table(dev, &priv->srq_table.cmpt_table);
250
251 err_qp:
252         mlx4_cleanup_icm_table(dev, &priv->qp_table.cmpt_table);
253
254 err:
255         return err;
256 }
257
258 static int __devinit mlx4_init_icm(struct mlx4_dev *dev,
259                                    struct mlx4_dev_cap *dev_cap,
260                                    struct mlx4_init_hca_param *init_hca,
261                                    u64 icm_size)
262 {
263         struct mlx4_priv *priv = mlx4_priv(dev);
264         u64 aux_pages;
265         int err;
266
267         err = mlx4_SET_ICM_SIZE(dev, icm_size, &aux_pages);
268         if (err) {
269                 mlx4_err(dev, "SET_ICM_SIZE command failed, aborting.\n");
270                 return err;
271         }
272
273         mlx4_dbg(dev, "%lld KB of HCA context requires %lld KB aux memory.\n",
274                  (unsigned long long) icm_size >> 10,
275                  (unsigned long long) aux_pages << 2);
276
277         priv->fw.aux_icm = mlx4_alloc_icm(dev, aux_pages,
278                                           GFP_HIGHUSER | __GFP_NOWARN);
279         if (!priv->fw.aux_icm) {
280                 mlx4_err(dev, "Couldn't allocate aux memory, aborting.\n");
281                 return -ENOMEM;
282         }
283
284         err = mlx4_MAP_ICM_AUX(dev, priv->fw.aux_icm);
285         if (err) {
286                 mlx4_err(dev, "MAP_ICM_AUX command failed, aborting.\n");
287                 goto err_free_aux;
288         }
289
290         err = mlx4_init_cmpt_table(dev, init_hca->cmpt_base, dev_cap->cmpt_entry_sz);
291         if (err) {
292                 mlx4_err(dev, "Failed to map cMPT context memory, aborting.\n");
293                 goto err_unmap_aux;
294         }
295
296         err = mlx4_map_eq_icm(dev, init_hca->eqc_base);
297         if (err) {
298                 mlx4_err(dev, "Failed to map EQ context memory, aborting.\n");
299                 goto err_unmap_cmpt;
300         }
301
302         err = mlx4_init_icm_table(dev, &priv->mr_table.mtt_table,
303                                   init_hca->mtt_base,
304                                   dev->caps.mtt_entry_sz,
305                                   dev->caps.num_mtt_segs,
306                                   dev->caps.reserved_mtts, 1);
307         if (err) {
308                 mlx4_err(dev, "Failed to map MTT context memory, aborting.\n");
309                 goto err_unmap_eq;
310         }
311
312         err = mlx4_init_icm_table(dev, &priv->mr_table.dmpt_table,
313                                   init_hca->dmpt_base,
314                                   dev_cap->dmpt_entry_sz,
315                                   dev->caps.num_mpts,
316                                   dev->caps.reserved_mrws, 1);
317         if (err) {
318                 mlx4_err(dev, "Failed to map dMPT context memory, aborting.\n");
319                 goto err_unmap_mtt;
320         }
321
322         err = mlx4_init_icm_table(dev, &priv->qp_table.qp_table,
323                                   init_hca->qpc_base,
324                                   dev_cap->qpc_entry_sz,
325                                   dev->caps.num_qps,
326                                   dev->caps.reserved_qps, 0);
327         if (err) {
328                 mlx4_err(dev, "Failed to map QP context memory, aborting.\n");
329                 goto err_unmap_dmpt;
330         }
331
332         err = mlx4_init_icm_table(dev, &priv->qp_table.auxc_table,
333                                   init_hca->auxc_base,
334                                   dev_cap->aux_entry_sz,
335                                   dev->caps.num_qps,
336                                   dev->caps.reserved_qps, 0);
337         if (err) {
338                 mlx4_err(dev, "Failed to map AUXC context memory, aborting.\n");
339                 goto err_unmap_qp;
340         }
341
342         err = mlx4_init_icm_table(dev, &priv->qp_table.altc_table,
343                                   init_hca->altc_base,
344                                   dev_cap->altc_entry_sz,
345                                   dev->caps.num_qps,
346                                   dev->caps.reserved_qps, 0);
347         if (err) {
348                 mlx4_err(dev, "Failed to map ALTC context memory, aborting.\n");
349                 goto err_unmap_auxc;
350         }
351
352         err = mlx4_init_icm_table(dev, &priv->qp_table.rdmarc_table,
353                                   init_hca->rdmarc_base,
354                                   dev_cap->rdmarc_entry_sz << priv->qp_table.rdmarc_shift,
355                                   dev->caps.num_qps,
356                                   dev->caps.reserved_qps, 0);
357         if (err) {
358                 mlx4_err(dev, "Failed to map RDMARC context memory, aborting\n");
359                 goto err_unmap_altc;
360         }
361
362         err = mlx4_init_icm_table(dev, &priv->cq_table.table,
363                                   init_hca->cqc_base,
364                                   dev_cap->cqc_entry_sz,
365                                   dev->caps.num_cqs,
366                                   dev->caps.reserved_cqs, 0);
367         if (err) {
368                 mlx4_err(dev, "Failed to map CQ context memory, aborting.\n");
369                 goto err_unmap_rdmarc;
370         }
371
372         err = mlx4_init_icm_table(dev, &priv->srq_table.table,
373                                   init_hca->srqc_base,
374                                   dev_cap->srq_entry_sz,
375                                   dev->caps.num_srqs,
376                                   dev->caps.reserved_srqs, 0);
377         if (err) {
378                 mlx4_err(dev, "Failed to map SRQ context memory, aborting.\n");
379                 goto err_unmap_cq;
380         }
381
382         /*
383          * It's not strictly required, but for simplicity just map the
384          * whole multicast group table now.  The table isn't very big
385          * and it's a lot easier than trying to track ref counts.
386          */
387         err = mlx4_init_icm_table(dev, &priv->mcg_table.table,
388                                   init_hca->mc_base, MLX4_MGM_ENTRY_SIZE,
389                                   dev->caps.num_mgms + dev->caps.num_amgms,
390                                   dev->caps.num_mgms + dev->caps.num_amgms,
391                                   0);
392         if (err) {
393                 mlx4_err(dev, "Failed to map MCG context memory, aborting.\n");
394                 goto err_unmap_srq;
395         }
396
397         return 0;
398
399 err_unmap_srq:
400         mlx4_cleanup_icm_table(dev, &priv->srq_table.table);
401
402 err_unmap_cq:
403         mlx4_cleanup_icm_table(dev, &priv->cq_table.table);
404
405 err_unmap_rdmarc:
406         mlx4_cleanup_icm_table(dev, &priv->qp_table.rdmarc_table);
407
408 err_unmap_altc:
409         mlx4_cleanup_icm_table(dev, &priv->qp_table.altc_table);
410
411 err_unmap_auxc:
412         mlx4_cleanup_icm_table(dev, &priv->qp_table.auxc_table);
413
414 err_unmap_qp:
415         mlx4_cleanup_icm_table(dev, &priv->qp_table.qp_table);
416
417 err_unmap_dmpt:
418         mlx4_cleanup_icm_table(dev, &priv->mr_table.dmpt_table);
419
420 err_unmap_mtt:
421         mlx4_cleanup_icm_table(dev, &priv->mr_table.mtt_table);
422
423 err_unmap_eq:
424         mlx4_unmap_eq_icm(dev);
425
426 err_unmap_cmpt:
427         mlx4_cleanup_icm_table(dev, &priv->eq_table.cmpt_table);
428         mlx4_cleanup_icm_table(dev, &priv->cq_table.cmpt_table);
429         mlx4_cleanup_icm_table(dev, &priv->srq_table.cmpt_table);
430         mlx4_cleanup_icm_table(dev, &priv->qp_table.cmpt_table);
431
432 err_unmap_aux:
433         mlx4_UNMAP_ICM_AUX(dev);
434
435 err_free_aux:
436         mlx4_free_icm(dev, priv->fw.aux_icm);
437
438         return err;
439 }
440
441 static void mlx4_free_icms(struct mlx4_dev *dev)
442 {
443         struct mlx4_priv *priv = mlx4_priv(dev);
444
445         mlx4_cleanup_icm_table(dev, &priv->mcg_table.table);
446         mlx4_cleanup_icm_table(dev, &priv->srq_table.table);
447         mlx4_cleanup_icm_table(dev, &priv->cq_table.table);
448         mlx4_cleanup_icm_table(dev, &priv->qp_table.rdmarc_table);
449         mlx4_cleanup_icm_table(dev, &priv->qp_table.altc_table);
450         mlx4_cleanup_icm_table(dev, &priv->qp_table.auxc_table);
451         mlx4_cleanup_icm_table(dev, &priv->qp_table.qp_table);
452         mlx4_cleanup_icm_table(dev, &priv->mr_table.dmpt_table);
453         mlx4_cleanup_icm_table(dev, &priv->mr_table.mtt_table);
454         mlx4_cleanup_icm_table(dev, &priv->eq_table.cmpt_table);
455         mlx4_cleanup_icm_table(dev, &priv->cq_table.cmpt_table);
456         mlx4_cleanup_icm_table(dev, &priv->srq_table.cmpt_table);
457         mlx4_cleanup_icm_table(dev, &priv->qp_table.cmpt_table);
458         mlx4_unmap_eq_icm(dev);
459
460         mlx4_UNMAP_ICM_AUX(dev);
461         mlx4_free_icm(dev, priv->fw.aux_icm);
462 }
463
464 static void mlx4_close_hca(struct mlx4_dev *dev)
465 {
466         mlx4_CLOSE_HCA(dev, 0);
467         mlx4_free_icms(dev);
468         mlx4_UNMAP_FA(dev);
469         mlx4_free_icm(dev, mlx4_priv(dev)->fw.fw_icm);
470 }
471
472 static int __devinit mlx4_init_hca(struct mlx4_dev *dev)
473 {
474         struct mlx4_priv          *priv = mlx4_priv(dev);
475         struct mlx4_adapter        adapter;
476         struct mlx4_dev_cap        dev_cap;
477         struct mlx4_profile        profile;
478         struct mlx4_init_hca_param init_hca;
479         u64 icm_size;
480         int err;
481
482         err = mlx4_QUERY_FW(dev);
483         if (err) {
484                 mlx4_err(dev, "QUERY_FW command failed, aborting.\n");
485                 return err;
486         }
487
488         err = mlx4_load_fw(dev);
489         if (err) {
490                 mlx4_err(dev, "Failed to start FW, aborting.\n");
491                 return err;
492         }
493
494         err = mlx4_dev_cap(dev, &dev_cap);
495         if (err) {
496                 mlx4_err(dev, "QUERY_DEV_CAP command failed, aborting.\n");
497                 goto err_stop_fw;
498         }
499
500         profile = default_profile;
501
502         icm_size = mlx4_make_profile(dev, &profile, &dev_cap, &init_hca);
503         if ((long long) icm_size < 0) {
504                 err = icm_size;
505                 goto err_stop_fw;
506         }
507
508         init_hca.log_uar_sz = ilog2(dev->caps.num_uars);
509
510         err = mlx4_init_icm(dev, &dev_cap, &init_hca, icm_size);
511         if (err)
512                 goto err_stop_fw;
513
514         err = mlx4_INIT_HCA(dev, &init_hca);
515         if (err) {
516                 mlx4_err(dev, "INIT_HCA command failed, aborting.\n");
517                 goto err_free_icm;
518         }
519
520         err = mlx4_QUERY_ADAPTER(dev, &adapter);
521         if (err) {
522                 mlx4_err(dev, "QUERY_ADAPTER command failed, aborting.\n");
523                 goto err_close;
524         }
525
526         priv->eq_table.inta_pin = adapter.inta_pin;
527         priv->rev_id            = adapter.revision_id;
528         memcpy(priv->board_id, adapter.board_id, sizeof priv->board_id);
529
530         return 0;
531
532 err_close:
533         mlx4_close_hca(dev);
534
535 err_free_icm:
536         mlx4_free_icms(dev);
537
538 err_stop_fw:
539         mlx4_UNMAP_FA(dev);
540         mlx4_free_icm(dev, priv->fw.fw_icm);
541
542         return err;
543 }
544
545 static int __devinit mlx4_setup_hca(struct mlx4_dev *dev)
546 {
547         struct mlx4_priv *priv = mlx4_priv(dev);
548         int err;
549
550         err = mlx4_init_uar_table(dev);
551         if (err) {
552                 mlx4_err(dev, "Failed to initialize "
553                          "user access region table, aborting.\n");
554                 return err;
555         }
556
557         err = mlx4_uar_alloc(dev, &priv->driver_uar);
558         if (err) {
559                 mlx4_err(dev, "Failed to allocate driver access region, "
560                          "aborting.\n");
561                 goto err_uar_table_free;
562         }
563
564         priv->kar = ioremap(priv->driver_uar.pfn << PAGE_SHIFT, PAGE_SIZE);
565         if (!priv->kar) {
566                 mlx4_err(dev, "Couldn't map kernel access region, "
567                          "aborting.\n");
568                 err = -ENOMEM;
569                 goto err_uar_free;
570         }
571
572         err = mlx4_init_pd_table(dev);
573         if (err) {
574                 mlx4_err(dev, "Failed to initialize "
575                          "protection domain table, aborting.\n");
576                 goto err_kar_unmap;
577         }
578
579         err = mlx4_init_mr_table(dev);
580         if (err) {
581                 mlx4_err(dev, "Failed to initialize "
582                          "memory region table, aborting.\n");
583                 goto err_pd_table_free;
584         }
585
586         err = mlx4_init_eq_table(dev);
587         if (err) {
588                 mlx4_err(dev, "Failed to initialize "
589                          "event queue table, aborting.\n");
590                 goto err_mr_table_free;
591         }
592
593         err = mlx4_cmd_use_events(dev);
594         if (err) {
595                 mlx4_err(dev, "Failed to switch to event-driven "
596                          "firmware commands, aborting.\n");
597                 goto err_eq_table_free;
598         }
599
600         err = mlx4_NOP(dev);
601         if (err) {
602                 mlx4_err(dev, "NOP command failed to generate interrupt "
603                          "(IRQ %d), aborting.\n",
604                          priv->eq_table.eq[MLX4_EQ_ASYNC].irq);
605                 if (dev->flags & MLX4_FLAG_MSI_X)
606                         mlx4_err(dev, "Try again with MSI-X disabled.\n");
607                 else
608                         mlx4_err(dev, "BIOS or ACPI interrupt routing problem?\n");
609
610                 goto err_cmd_poll;
611         }
612
613         mlx4_dbg(dev, "NOP command IRQ test passed\n");
614
615         err = mlx4_init_cq_table(dev);
616         if (err) {
617                 mlx4_err(dev, "Failed to initialize "
618                          "completion queue table, aborting.\n");
619                 goto err_cmd_poll;
620         }
621
622         err = mlx4_init_srq_table(dev);
623         if (err) {
624                 mlx4_err(dev, "Failed to initialize "
625                          "shared receive queue table, aborting.\n");
626                 goto err_cq_table_free;
627         }
628
629         err = mlx4_init_qp_table(dev);
630         if (err) {
631                 mlx4_err(dev, "Failed to initialize "
632                          "queue pair table, aborting.\n");
633                 goto err_srq_table_free;
634         }
635
636         err = mlx4_init_mcg_table(dev);
637         if (err) {
638                 mlx4_err(dev, "Failed to initialize "
639                          "multicast group table, aborting.\n");
640                 goto err_qp_table_free;
641         }
642
643         return 0;
644
645 err_qp_table_free:
646         mlx4_cleanup_qp_table(dev);
647
648 err_srq_table_free:
649         mlx4_cleanup_srq_table(dev);
650
651 err_cq_table_free:
652         mlx4_cleanup_cq_table(dev);
653
654 err_cmd_poll:
655         mlx4_cmd_use_polling(dev);
656
657 err_eq_table_free:
658         mlx4_cleanup_eq_table(dev);
659
660 err_mr_table_free:
661         mlx4_cleanup_mr_table(dev);
662
663 err_pd_table_free:
664         mlx4_cleanup_pd_table(dev);
665
666 err_kar_unmap:
667         iounmap(priv->kar);
668
669 err_uar_free:
670         mlx4_uar_free(dev, &priv->driver_uar);
671
672 err_uar_table_free:
673         mlx4_cleanup_uar_table(dev);
674         return err;
675 }
676
677 static void __devinit mlx4_enable_msi_x(struct mlx4_dev *dev)
678 {
679         struct mlx4_priv *priv = mlx4_priv(dev);
680         struct msix_entry entries[MLX4_NUM_EQ];
681         int err;
682         int i;
683
684         if (msi_x) {
685                 for (i = 0; i < MLX4_NUM_EQ; ++i)
686                         entries[i].entry = i;
687
688                 err = pci_enable_msix(dev->pdev, entries, ARRAY_SIZE(entries));
689                 if (err) {
690                         if (err > 0)
691                                 mlx4_info(dev, "Only %d MSI-X vectors available, "
692                                           "not using MSI-X\n", err);
693                         goto no_msi;
694                 }
695
696                 for (i = 0; i < MLX4_NUM_EQ; ++i)
697                         priv->eq_table.eq[i].irq = entries[i].vector;
698
699                 dev->flags |= MLX4_FLAG_MSI_X;
700                 return;
701         }
702
703 no_msi:
704         for (i = 0; i < MLX4_NUM_EQ; ++i)
705                 priv->eq_table.eq[i].irq = dev->pdev->irq;
706 }
707
708 static int __devinit mlx4_init_one(struct pci_dev *pdev,
709                                    const struct pci_device_id *id)
710 {
711         static int mlx4_version_printed;
712         struct mlx4_priv *priv;
713         struct mlx4_dev *dev;
714         int err;
715
716         if (!mlx4_version_printed) {
717                 printk(KERN_INFO "%s", mlx4_version);
718                 ++mlx4_version_printed;
719         }
720
721         printk(KERN_INFO PFX "Initializing %s\n",
722                pci_name(pdev));
723
724         err = pci_enable_device(pdev);
725         if (err) {
726                 dev_err(&pdev->dev, "Cannot enable PCI device, "
727                         "aborting.\n");
728                 return err;
729         }
730
731         /*
732          * Check for BARs.  We expect 0: 1MB, 2: 8MB, 4: DDR (may not
733          * be present)
734          */
735         if (!(pci_resource_flags(pdev, 0) & IORESOURCE_MEM) ||
736             pci_resource_len(pdev, 0) != 1 << 20) {
737                 dev_err(&pdev->dev, "Missing DCS, aborting.\n");
738                 err = -ENODEV;
739                 goto err_disable_pdev;
740         }
741         if (!(pci_resource_flags(pdev, 2) & IORESOURCE_MEM)) {
742                 dev_err(&pdev->dev, "Missing UAR, aborting.\n");
743                 err = -ENODEV;
744                 goto err_disable_pdev;
745         }
746
747         err = pci_request_region(pdev, 0, DRV_NAME);
748         if (err) {
749                 dev_err(&pdev->dev, "Cannot request control region, aborting.\n");
750                 goto err_disable_pdev;
751         }
752
753         err = pci_request_region(pdev, 2, DRV_NAME);
754         if (err) {
755                 dev_err(&pdev->dev, "Cannot request UAR region, aborting.\n");
756                 goto err_release_bar0;
757         }
758
759         pci_set_master(pdev);
760
761         err = pci_set_dma_mask(pdev, DMA_64BIT_MASK);
762         if (err) {
763                 dev_warn(&pdev->dev, "Warning: couldn't set 64-bit PCI DMA mask.\n");
764                 err = pci_set_dma_mask(pdev, DMA_32BIT_MASK);
765                 if (err) {
766                         dev_err(&pdev->dev, "Can't set PCI DMA mask, aborting.\n");
767                         goto err_release_bar2;
768                 }
769         }
770         err = pci_set_consistent_dma_mask(pdev, DMA_64BIT_MASK);
771         if (err) {
772                 dev_warn(&pdev->dev, "Warning: couldn't set 64-bit "
773                          "consistent PCI DMA mask.\n");
774                 err = pci_set_consistent_dma_mask(pdev, DMA_32BIT_MASK);
775                 if (err) {
776                         dev_err(&pdev->dev, "Can't set consistent PCI DMA mask, "
777                                 "aborting.\n");
778                         goto err_release_bar2;
779                 }
780         }
781
782         priv = kzalloc(sizeof *priv, GFP_KERNEL);
783         if (!priv) {
784                 dev_err(&pdev->dev, "Device struct alloc failed, "
785                         "aborting.\n");
786                 err = -ENOMEM;
787                 goto err_release_bar2;
788         }
789
790         dev       = &priv->dev;
791         dev->pdev = pdev;
792         INIT_LIST_HEAD(&priv->ctx_list);
793         spin_lock_init(&priv->ctx_lock);
794
795         /*
796          * Now reset the HCA before we touch the PCI capabilities or
797          * attempt a firmware command, since a boot ROM may have left
798          * the HCA in an undefined state.
799          */
800         err = mlx4_reset(dev);
801         if (err) {
802                 mlx4_err(dev, "Failed to reset HCA, aborting.\n");
803                 goto err_free_dev;
804         }
805
806         mlx4_enable_msi_x(dev);
807
808         if (mlx4_cmd_init(dev)) {
809                 mlx4_err(dev, "Failed to init command interface, aborting.\n");
810                 goto err_free_dev;
811         }
812
813         err = mlx4_init_hca(dev);
814         if (err)
815                 goto err_cmd;
816
817         err = mlx4_setup_hca(dev);
818         if (err)
819                 goto err_close;
820
821         err = mlx4_register_device(dev);
822         if (err)
823                 goto err_cleanup;
824
825         pci_set_drvdata(pdev, dev);
826
827         return 0;
828
829 err_cleanup:
830         mlx4_cleanup_mcg_table(dev);
831         mlx4_cleanup_qp_table(dev);
832         mlx4_cleanup_srq_table(dev);
833         mlx4_cleanup_cq_table(dev);
834         mlx4_cmd_use_polling(dev);
835         mlx4_cleanup_eq_table(dev);
836         mlx4_cleanup_mr_table(dev);
837         mlx4_cleanup_pd_table(dev);
838         mlx4_cleanup_uar_table(dev);
839
840 err_close:
841         mlx4_close_hca(dev);
842
843 err_cmd:
844         mlx4_cmd_cleanup(dev);
845
846 err_free_dev:
847         if (dev->flags & MLX4_FLAG_MSI_X)
848                 pci_disable_msix(pdev);
849
850         kfree(priv);
851
852 err_release_bar2:
853         pci_release_region(pdev, 2);
854
855 err_release_bar0:
856         pci_release_region(pdev, 0);
857
858 err_disable_pdev:
859         pci_disable_device(pdev);
860         pci_set_drvdata(pdev, NULL);
861         return err;
862 }
863
864 static void __devexit mlx4_remove_one(struct pci_dev *pdev)
865 {
866         struct mlx4_dev  *dev  = pci_get_drvdata(pdev);
867         struct mlx4_priv *priv = mlx4_priv(dev);
868         int p;
869
870         if (dev) {
871                 mlx4_unregister_device(dev);
872
873                 for (p = 1; p <= dev->caps.num_ports; ++p)
874                         mlx4_CLOSE_PORT(dev, p);
875
876                 mlx4_cleanup_mcg_table(dev);
877                 mlx4_cleanup_qp_table(dev);
878                 mlx4_cleanup_srq_table(dev);
879                 mlx4_cleanup_cq_table(dev);
880                 mlx4_cmd_use_polling(dev);
881                 mlx4_cleanup_eq_table(dev);
882                 mlx4_cleanup_mr_table(dev);
883                 mlx4_cleanup_pd_table(dev);
884
885                 iounmap(priv->kar);
886                 mlx4_uar_free(dev, &priv->driver_uar);
887                 mlx4_cleanup_uar_table(dev);
888                 mlx4_close_hca(dev);
889                 mlx4_cmd_cleanup(dev);
890
891                 if (dev->flags & MLX4_FLAG_MSI_X)
892                         pci_disable_msix(pdev);
893
894                 kfree(priv);
895                 pci_release_region(pdev, 2);
896                 pci_release_region(pdev, 0);
897                 pci_disable_device(pdev);
898                 pci_set_drvdata(pdev, NULL);
899         }
900 }
901
902 int mlx4_restart_one(struct pci_dev *pdev)
903 {
904         mlx4_remove_one(pdev);
905         return mlx4_init_one(pdev, NULL);
906 }
907
908 static struct pci_device_id mlx4_pci_table[] = {
909         { PCI_VDEVICE(MELLANOX, 0x6340) }, /* MT25408 "Hermon" SDR */
910         { PCI_VDEVICE(MELLANOX, 0x634a) }, /* MT25408 "Hermon" DDR */
911         { PCI_VDEVICE(MELLANOX, 0x6354) }, /* MT25408 "Hermon" QDR */
912         { PCI_VDEVICE(MELLANOX, 0x6732) }, /* MT25408 "Hermon" DDR PCIe gen2 */
913         { PCI_VDEVICE(MELLANOX, 0x673c) }, /* MT25408 "Hermon" QDR PCIe gen2 */
914         { 0, }
915 };
916
917 MODULE_DEVICE_TABLE(pci, mlx4_pci_table);
918
919 static struct pci_driver mlx4_driver = {
920         .name           = DRV_NAME,
921         .id_table       = mlx4_pci_table,
922         .probe          = mlx4_init_one,
923         .remove         = __devexit_p(mlx4_remove_one)
924 };
925
926 static int __init mlx4_init(void)
927 {
928         int ret;
929
930         ret = mlx4_catas_init();
931         if (ret)
932                 return ret;
933
934         ret = pci_register_driver(&mlx4_driver);
935         return ret < 0 ? ret : 0;
936 }
937
938 static void __exit mlx4_cleanup(void)
939 {
940         pci_unregister_driver(&mlx4_driver);
941         mlx4_catas_cleanup();
942 }
943
944 module_init(mlx4_init);
945 module_exit(mlx4_cleanup);