34e6b8685ba3d18cf08c63304010e9f364ca672c
[linux-2.6.git] / drivers / infiniband / hw / mthca / mthca_provider.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 Cisco Systems. All rights reserved.
5  * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
6  * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
7  *
8  * This software is available to you under a choice of one of two
9  * licenses.  You may choose to be licensed under the terms of the GNU
10  * General Public License (GPL) Version 2, available from the file
11  * COPYING in the main directory of this source tree, or the
12  * OpenIB.org BSD license below:
13  *
14  *     Redistribution and use in source and binary forms, with or
15  *     without modification, are permitted provided that the following
16  *     conditions are met:
17  *
18  *      - Redistributions of source code must retain the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer.
21  *
22  *      - Redistributions in binary form must reproduce the above
23  *        copyright notice, this list of conditions and the following
24  *        disclaimer in the documentation and/or other materials
25  *        provided with the distribution.
26  *
27  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
28  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
29  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
30  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
31  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
32  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
33  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
34  * SOFTWARE.
35  *
36  * $Id: mthca_provider.c 1397 2004-12-28 05:09:00Z roland $
37  */
38
39 #include <ib_smi.h>
40 #include <linux/mm.h>
41
42 #include "mthca_dev.h"
43 #include "mthca_cmd.h"
44 #include "mthca_user.h"
45 #include "mthca_memfree.h"
46
47 static int mthca_query_device(struct ib_device *ibdev,
48                               struct ib_device_attr *props)
49 {
50         struct ib_smp *in_mad  = NULL;
51         struct ib_smp *out_mad = NULL;
52         int err = -ENOMEM;
53         struct mthca_dev* mdev = to_mdev(ibdev);
54
55         u8 status;
56
57         in_mad  = kmalloc(sizeof *in_mad, GFP_KERNEL);
58         out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
59         if (!in_mad || !out_mad)
60                 goto out;
61
62         memset(props, 0, sizeof *props);
63
64         props->fw_ver              = mdev->fw_ver;
65
66         memset(in_mad, 0, sizeof *in_mad);
67         in_mad->base_version       = 1;
68         in_mad->mgmt_class         = IB_MGMT_CLASS_SUBN_LID_ROUTED;
69         in_mad->class_version      = 1;
70         in_mad->method             = IB_MGMT_METHOD_GET;
71         in_mad->attr_id            = IB_SMP_ATTR_NODE_INFO;
72
73         err = mthca_MAD_IFC(mdev, 1, 1,
74                             1, NULL, NULL, in_mad, out_mad,
75                             &status);
76         if (err)
77                 goto out;
78         if (status) {
79                 err = -EINVAL;
80                 goto out;
81         }
82
83         props->device_cap_flags    = mdev->device_cap_flags;
84         props->vendor_id           = be32_to_cpup((u32 *) (out_mad->data + 36)) &
85                 0xffffff;
86         props->vendor_part_id      = be16_to_cpup((u16 *) (out_mad->data + 30));
87         props->hw_ver              = be16_to_cpup((u16 *) (out_mad->data + 32));
88         memcpy(&props->sys_image_guid, out_mad->data +  4, 8);
89         memcpy(&props->node_guid,      out_mad->data + 12, 8);
90
91         props->max_mr_size         = ~0ull;
92         props->max_qp              = mdev->limits.num_qps - mdev->limits.reserved_qps;
93         props->max_qp_wr           = 0xffff;
94         props->max_sge             = mdev->limits.max_sg;
95         props->max_cq              = mdev->limits.num_cqs - mdev->limits.reserved_cqs;
96         props->max_cqe             = 0xffff;
97         props->max_mr              = mdev->limits.num_mpts - mdev->limits.reserved_mrws;
98         props->max_pd              = mdev->limits.num_pds - mdev->limits.reserved_pds;
99         props->max_qp_rd_atom      = 1 << mdev->qp_table.rdb_shift;
100         props->max_qp_init_rd_atom = 1 << mdev->qp_table.rdb_shift;
101         props->local_ca_ack_delay  = mdev->limits.local_ca_ack_delay;
102
103         err = 0;
104  out:
105         kfree(in_mad);
106         kfree(out_mad);
107         return err;
108 }
109
110 static int mthca_query_port(struct ib_device *ibdev,
111                             u8 port, struct ib_port_attr *props)
112 {
113         struct ib_smp *in_mad  = NULL;
114         struct ib_smp *out_mad = NULL;
115         int err = -ENOMEM;
116         u8 status;
117
118         in_mad  = kmalloc(sizeof *in_mad, GFP_KERNEL);
119         out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
120         if (!in_mad || !out_mad)
121                 goto out;
122
123         memset(in_mad, 0, sizeof *in_mad);
124         in_mad->base_version       = 1;
125         in_mad->mgmt_class         = IB_MGMT_CLASS_SUBN_LID_ROUTED;
126         in_mad->class_version      = 1;
127         in_mad->method             = IB_MGMT_METHOD_GET;
128         in_mad->attr_id            = IB_SMP_ATTR_PORT_INFO;
129         in_mad->attr_mod           = cpu_to_be32(port);
130
131         err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
132                             port, NULL, NULL, in_mad, out_mad,
133                             &status);
134         if (err)
135                 goto out;
136         if (status) {
137                 err = -EINVAL;
138                 goto out;
139         }
140
141         props->lid               = be16_to_cpup((u16 *) (out_mad->data + 16));
142         props->lmc               = out_mad->data[34] & 0x7;
143         props->sm_lid            = be16_to_cpup((u16 *) (out_mad->data + 18));
144         props->sm_sl             = out_mad->data[36] & 0xf;
145         props->state             = out_mad->data[32] & 0xf;
146         props->phys_state        = out_mad->data[33] >> 4;
147         props->port_cap_flags    = be32_to_cpup((u32 *) (out_mad->data + 20));
148         props->gid_tbl_len       = to_mdev(ibdev)->limits.gid_table_len;
149         props->pkey_tbl_len      = to_mdev(ibdev)->limits.pkey_table_len;
150         props->qkey_viol_cntr    = be16_to_cpup((u16 *) (out_mad->data + 48));
151         props->active_width      = out_mad->data[31] & 0xf;
152         props->active_speed      = out_mad->data[35] >> 4;
153
154  out:
155         kfree(in_mad);
156         kfree(out_mad);
157         return err;
158 }
159
160 static int mthca_modify_port(struct ib_device *ibdev,
161                              u8 port, int port_modify_mask,
162                              struct ib_port_modify *props)
163 {
164         struct mthca_set_ib_param set_ib;
165         struct ib_port_attr attr;
166         int err;
167         u8 status;
168
169         if (down_interruptible(&to_mdev(ibdev)->cap_mask_mutex))
170                 return -ERESTARTSYS;
171
172         err = mthca_query_port(ibdev, port, &attr);
173         if (err)
174                 goto out;
175
176         set_ib.set_si_guid     = 0;
177         set_ib.reset_qkey_viol = !!(port_modify_mask & IB_PORT_RESET_QKEY_CNTR);
178
179         set_ib.cap_mask = (attr.port_cap_flags | props->set_port_cap_mask) &
180                 ~props->clr_port_cap_mask;
181
182         err = mthca_SET_IB(to_mdev(ibdev), &set_ib, port, &status);
183         if (err)
184                 goto out;
185         if (status) {
186                 err = -EINVAL;
187                 goto out;
188         }
189
190 out:
191         up(&to_mdev(ibdev)->cap_mask_mutex);
192         return err;
193 }
194
195 static int mthca_query_pkey(struct ib_device *ibdev,
196                             u8 port, u16 index, u16 *pkey)
197 {
198         struct ib_smp *in_mad  = NULL;
199         struct ib_smp *out_mad = NULL;
200         int err = -ENOMEM;
201         u8 status;
202
203         in_mad  = kmalloc(sizeof *in_mad, GFP_KERNEL);
204         out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
205         if (!in_mad || !out_mad)
206                 goto out;
207
208         memset(in_mad, 0, sizeof *in_mad);
209         in_mad->base_version       = 1;
210         in_mad->mgmt_class         = IB_MGMT_CLASS_SUBN_LID_ROUTED;
211         in_mad->class_version      = 1;
212         in_mad->method             = IB_MGMT_METHOD_GET;
213         in_mad->attr_id            = IB_SMP_ATTR_PKEY_TABLE;
214         in_mad->attr_mod           = cpu_to_be32(index / 32);
215
216         err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
217                             port, NULL, NULL, in_mad, out_mad,
218                             &status);
219         if (err)
220                 goto out;
221         if (status) {
222                 err = -EINVAL;
223                 goto out;
224         }
225
226         *pkey = be16_to_cpu(((u16 *) out_mad->data)[index % 32]);
227
228  out:
229         kfree(in_mad);
230         kfree(out_mad);
231         return err;
232 }
233
234 static int mthca_query_gid(struct ib_device *ibdev, u8 port,
235                            int index, union ib_gid *gid)
236 {
237         struct ib_smp *in_mad  = NULL;
238         struct ib_smp *out_mad = NULL;
239         int err = -ENOMEM;
240         u8 status;
241
242         in_mad  = kmalloc(sizeof *in_mad, GFP_KERNEL);
243         out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
244         if (!in_mad || !out_mad)
245                 goto out;
246
247         memset(in_mad, 0, sizeof *in_mad);
248         in_mad->base_version       = 1;
249         in_mad->mgmt_class         = IB_MGMT_CLASS_SUBN_LID_ROUTED;
250         in_mad->class_version      = 1;
251         in_mad->method             = IB_MGMT_METHOD_GET;
252         in_mad->attr_id            = IB_SMP_ATTR_PORT_INFO;
253         in_mad->attr_mod           = cpu_to_be32(port);
254
255         err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
256                             port, NULL, NULL, in_mad, out_mad,
257                             &status);
258         if (err)
259                 goto out;
260         if (status) {
261                 err = -EINVAL;
262                 goto out;
263         }
264
265         memcpy(gid->raw, out_mad->data + 8, 8);
266
267         memset(in_mad, 0, sizeof *in_mad);
268         in_mad->base_version       = 1;
269         in_mad->mgmt_class         = IB_MGMT_CLASS_SUBN_LID_ROUTED;
270         in_mad->class_version      = 1;
271         in_mad->method             = IB_MGMT_METHOD_GET;
272         in_mad->attr_id            = IB_SMP_ATTR_GUID_INFO;
273         in_mad->attr_mod           = cpu_to_be32(index / 8);
274
275         err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
276                             port, NULL, NULL, in_mad, out_mad,
277                             &status);
278         if (err)
279                 goto out;
280         if (status) {
281                 err = -EINVAL;
282                 goto out;
283         }
284
285         memcpy(gid->raw + 8, out_mad->data + (index % 8) * 16, 8);
286
287  out:
288         kfree(in_mad);
289         kfree(out_mad);
290         return err;
291 }
292
293 static struct ib_ucontext *mthca_alloc_ucontext(struct ib_device *ibdev,
294                                                 struct ib_udata *udata)
295 {
296         struct mthca_alloc_ucontext_resp uresp;
297         struct mthca_ucontext           *context;
298         int                              err;
299
300         memset(&uresp, 0, sizeof uresp);
301
302         uresp.qp_tab_size = to_mdev(ibdev)->limits.num_qps;
303         if (mthca_is_memfree(to_mdev(ibdev)))
304                 uresp.uarc_size = to_mdev(ibdev)->uar_table.uarc_size;
305         else
306                 uresp.uarc_size = 0;
307
308         context = kmalloc(sizeof *context, GFP_KERNEL);
309         if (!context)
310                 return ERR_PTR(-ENOMEM);
311
312         err = mthca_uar_alloc(to_mdev(ibdev), &context->uar);
313         if (err) {
314                 kfree(context);
315                 return ERR_PTR(err);
316         }
317
318         context->db_tab = mthca_init_user_db_tab(to_mdev(ibdev));
319         if (IS_ERR(context->db_tab)) {
320                 err = PTR_ERR(context->db_tab);
321                 mthca_uar_free(to_mdev(ibdev), &context->uar);
322                 kfree(context);
323                 return ERR_PTR(err);
324         }
325
326         if (ib_copy_to_udata(udata, &uresp, sizeof uresp)) {
327                 mthca_cleanup_user_db_tab(to_mdev(ibdev), &context->uar, context->db_tab);
328                 mthca_uar_free(to_mdev(ibdev), &context->uar);
329                 kfree(context);
330                 return ERR_PTR(-EFAULT);
331         }
332
333         return &context->ibucontext;
334 }
335
336 static int mthca_dealloc_ucontext(struct ib_ucontext *context)
337 {
338         mthca_cleanup_user_db_tab(to_mdev(context->device), &to_mucontext(context)->uar,
339                                   to_mucontext(context)->db_tab);
340         mthca_uar_free(to_mdev(context->device), &to_mucontext(context)->uar);
341         kfree(to_mucontext(context));
342
343         return 0;
344 }
345
346 static int mthca_mmap_uar(struct ib_ucontext *context,
347                           struct vm_area_struct *vma)
348 {
349         if (vma->vm_end - vma->vm_start != PAGE_SIZE)
350                 return -EINVAL;
351
352         vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
353
354         if (io_remap_pfn_range(vma, vma->vm_start,
355                                to_mucontext(context)->uar.pfn,
356                                PAGE_SIZE, vma->vm_page_prot))
357                 return -EAGAIN;
358
359         return 0;
360 }
361
362 static struct ib_pd *mthca_alloc_pd(struct ib_device *ibdev,
363                                     struct ib_ucontext *context,
364                                     struct ib_udata *udata)
365 {
366         struct mthca_pd *pd;
367         int err;
368
369         pd = kmalloc(sizeof *pd, GFP_KERNEL);
370         if (!pd)
371                 return ERR_PTR(-ENOMEM);
372
373         err = mthca_pd_alloc(to_mdev(ibdev), !context, pd);
374         if (err) {
375                 kfree(pd);
376                 return ERR_PTR(err);
377         }
378
379         if (context) {
380                 if (ib_copy_to_udata(udata, &pd->pd_num, sizeof (__u32))) {
381                         mthca_pd_free(to_mdev(ibdev), pd);
382                         kfree(pd);
383                         return ERR_PTR(-EFAULT);
384                 }
385         }
386
387         return &pd->ibpd;
388 }
389
390 static int mthca_dealloc_pd(struct ib_pd *pd)
391 {
392         mthca_pd_free(to_mdev(pd->device), to_mpd(pd));
393         kfree(pd);
394
395         return 0;
396 }
397
398 static struct ib_ah *mthca_ah_create(struct ib_pd *pd,
399                                      struct ib_ah_attr *ah_attr)
400 {
401         int err;
402         struct mthca_ah *ah;
403
404         ah = kmalloc(sizeof *ah, GFP_ATOMIC);
405         if (!ah)
406                 return ERR_PTR(-ENOMEM);
407
408         err = mthca_create_ah(to_mdev(pd->device), to_mpd(pd), ah_attr, ah);
409         if (err) {
410                 kfree(ah);
411                 return ERR_PTR(err);
412         }
413
414         return &ah->ibah;
415 }
416
417 static int mthca_ah_destroy(struct ib_ah *ah)
418 {
419         mthca_destroy_ah(to_mdev(ah->device), to_mah(ah));
420         kfree(ah);
421
422         return 0;
423 }
424
425 static struct ib_qp *mthca_create_qp(struct ib_pd *pd,
426                                      struct ib_qp_init_attr *init_attr,
427                                      struct ib_udata *udata)
428 {
429         struct mthca_create_qp ucmd;
430         struct mthca_qp *qp;
431         int err;
432
433         switch (init_attr->qp_type) {
434         case IB_QPT_RC:
435         case IB_QPT_UC:
436         case IB_QPT_UD:
437         {
438                 struct mthca_ucontext *context;
439
440                 qp = kmalloc(sizeof *qp, GFP_KERNEL);
441                 if (!qp)
442                         return ERR_PTR(-ENOMEM);
443
444                 if (pd->uobject) {
445                         context = to_mucontext(pd->uobject->context);
446
447                         if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd))
448                                 return ERR_PTR(-EFAULT);
449
450                         err = mthca_map_user_db(to_mdev(pd->device), &context->uar,
451                                                 context->db_tab,
452                                                 ucmd.sq_db_index, ucmd.sq_db_page);
453                         if (err) {
454                                 kfree(qp);
455                                 return ERR_PTR(err);
456                         }
457
458                         err = mthca_map_user_db(to_mdev(pd->device), &context->uar,
459                                                 context->db_tab,
460                                                 ucmd.rq_db_index, ucmd.rq_db_page);
461                         if (err) {
462                                 mthca_unmap_user_db(to_mdev(pd->device),
463                                                     &context->uar,
464                                                     context->db_tab,
465                                                     ucmd.sq_db_index);
466                                 kfree(qp);
467                                 return ERR_PTR(err);
468                         }
469
470                         qp->mr.ibmr.lkey = ucmd.lkey;
471                         qp->sq.db_index  = ucmd.sq_db_index;
472                         qp->rq.db_index  = ucmd.rq_db_index;
473                 }
474
475                 err = mthca_alloc_qp(to_mdev(pd->device), to_mpd(pd),
476                                      to_mcq(init_attr->send_cq),
477                                      to_mcq(init_attr->recv_cq),
478                                      init_attr->qp_type, init_attr->sq_sig_type,
479                                      &init_attr->cap, qp);
480
481                 if (err && pd->uobject) {
482                         context = to_mucontext(pd->uobject->context);
483
484                         mthca_unmap_user_db(to_mdev(pd->device),
485                                             &context->uar,
486                                             context->db_tab,
487                                             ucmd.sq_db_index);
488                         mthca_unmap_user_db(to_mdev(pd->device),
489                                             &context->uar,
490                                             context->db_tab,
491                                             ucmd.rq_db_index);
492                 }
493
494                 qp->ibqp.qp_num = qp->qpn;
495                 break;
496         }
497         case IB_QPT_SMI:
498         case IB_QPT_GSI:
499         {
500                 /* Don't allow userspace to create special QPs */
501                 if (pd->uobject)
502                         return ERR_PTR(-EINVAL);
503
504                 qp = kmalloc(sizeof (struct mthca_sqp), GFP_KERNEL);
505                 if (!qp)
506                         return ERR_PTR(-ENOMEM);
507
508                 qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 : 1;
509
510                 err = mthca_alloc_sqp(to_mdev(pd->device), to_mpd(pd),
511                                       to_mcq(init_attr->send_cq),
512                                       to_mcq(init_attr->recv_cq),
513                                       init_attr->sq_sig_type, &init_attr->cap,
514                                       qp->ibqp.qp_num, init_attr->port_num,
515                                       to_msqp(qp));
516                 break;
517         }
518         default:
519                 /* Don't support raw QPs */
520                 return ERR_PTR(-ENOSYS);
521         }
522
523         if (err) {
524                 kfree(qp);
525                 return ERR_PTR(err);
526         }
527
528         init_attr->cap.max_inline_data = 0;
529         init_attr->cap.max_send_wr     = qp->sq.max;
530         init_attr->cap.max_recv_wr     = qp->rq.max;
531         init_attr->cap.max_send_sge    = qp->sq.max_gs;
532         init_attr->cap.max_recv_sge    = qp->rq.max_gs;
533
534         return &qp->ibqp;
535 }
536
537 static int mthca_destroy_qp(struct ib_qp *qp)
538 {
539         if (qp->uobject) {
540                 mthca_unmap_user_db(to_mdev(qp->device),
541                                     &to_mucontext(qp->uobject->context)->uar,
542                                     to_mucontext(qp->uobject->context)->db_tab,
543                                     to_mqp(qp)->sq.db_index);
544                 mthca_unmap_user_db(to_mdev(qp->device),
545                                     &to_mucontext(qp->uobject->context)->uar,
546                                     to_mucontext(qp->uobject->context)->db_tab,
547                                     to_mqp(qp)->rq.db_index);
548         }
549         mthca_free_qp(to_mdev(qp->device), to_mqp(qp));
550         kfree(qp);
551         return 0;
552 }
553
554 static struct ib_cq *mthca_create_cq(struct ib_device *ibdev, int entries,
555                                      struct ib_ucontext *context,
556                                      struct ib_udata *udata)
557 {
558         struct mthca_create_cq ucmd;
559         struct mthca_cq *cq;
560         int nent;
561         int err;
562
563         if (context) {
564                 if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd))
565                         return ERR_PTR(-EFAULT);
566
567                 err = mthca_map_user_db(to_mdev(ibdev), &to_mucontext(context)->uar,
568                                         to_mucontext(context)->db_tab,
569                                         ucmd.set_db_index, ucmd.set_db_page);
570                 if (err)
571                         return ERR_PTR(err);
572
573                 err = mthca_map_user_db(to_mdev(ibdev), &to_mucontext(context)->uar,
574                                         to_mucontext(context)->db_tab,
575                                         ucmd.arm_db_index, ucmd.arm_db_page);
576                 if (err)
577                         goto err_unmap_set;
578         }
579
580         cq = kmalloc(sizeof *cq, GFP_KERNEL);
581         if (!cq) {
582                 err = -ENOMEM;
583                 goto err_unmap_arm;
584         }
585
586         if (context) {
587                 cq->mr.ibmr.lkey    = ucmd.lkey;
588                 cq->set_ci_db_index = ucmd.set_db_index;
589                 cq->arm_db_index    = ucmd.arm_db_index;
590         }
591
592         for (nent = 1; nent <= entries; nent <<= 1)
593                 ; /* nothing */
594
595         err = mthca_init_cq(to_mdev(ibdev), nent,
596                             context ? to_mucontext(context) : NULL,
597                             context ? ucmd.pdn : to_mdev(ibdev)->driver_pd.pd_num,
598                             cq);
599         if (err)
600                 goto err_free;
601
602         if (context && ib_copy_to_udata(udata, &cq->cqn, sizeof (__u32))) {
603                 mthca_free_cq(to_mdev(ibdev), cq);
604                 goto err_free;
605         }
606
607         return &cq->ibcq;
608
609 err_free:
610         kfree(cq);
611
612 err_unmap_arm:
613         if (context)
614                 mthca_unmap_user_db(to_mdev(ibdev), &to_mucontext(context)->uar,
615                                     to_mucontext(context)->db_tab, ucmd.arm_db_index);
616
617 err_unmap_set:
618         if (context)
619                 mthca_unmap_user_db(to_mdev(ibdev), &to_mucontext(context)->uar,
620                                     to_mucontext(context)->db_tab, ucmd.set_db_index);
621
622         return ERR_PTR(err);
623 }
624
625 static int mthca_destroy_cq(struct ib_cq *cq)
626 {
627         if (cq->uobject) {
628                 mthca_unmap_user_db(to_mdev(cq->device),
629                                     &to_mucontext(cq->uobject->context)->uar,
630                                     to_mucontext(cq->uobject->context)->db_tab,
631                                     to_mcq(cq)->arm_db_index);
632                 mthca_unmap_user_db(to_mdev(cq->device),
633                                     &to_mucontext(cq->uobject->context)->uar,
634                                     to_mucontext(cq->uobject->context)->db_tab,
635                                     to_mcq(cq)->set_ci_db_index);
636         }
637         mthca_free_cq(to_mdev(cq->device), to_mcq(cq));
638         kfree(cq);
639
640         return 0;
641 }
642
643 static inline u32 convert_access(int acc)
644 {
645         return (acc & IB_ACCESS_REMOTE_ATOMIC ? MTHCA_MPT_FLAG_ATOMIC       : 0) |
646                (acc & IB_ACCESS_REMOTE_WRITE  ? MTHCA_MPT_FLAG_REMOTE_WRITE : 0) |
647                (acc & IB_ACCESS_REMOTE_READ   ? MTHCA_MPT_FLAG_REMOTE_READ  : 0) |
648                (acc & IB_ACCESS_LOCAL_WRITE   ? MTHCA_MPT_FLAG_LOCAL_WRITE  : 0) |
649                MTHCA_MPT_FLAG_LOCAL_READ;
650 }
651
652 static struct ib_mr *mthca_get_dma_mr(struct ib_pd *pd, int acc)
653 {
654         struct mthca_mr *mr;
655         int err;
656
657         mr = kmalloc(sizeof *mr, GFP_KERNEL);
658         if (!mr)
659                 return ERR_PTR(-ENOMEM);
660
661         err = mthca_mr_alloc_notrans(to_mdev(pd->device),
662                                      to_mpd(pd)->pd_num,
663                                      convert_access(acc), mr);
664
665         if (err) {
666                 kfree(mr);
667                 return ERR_PTR(err);
668         }
669
670         return &mr->ibmr;
671 }
672
673 static struct ib_mr *mthca_reg_phys_mr(struct ib_pd       *pd,
674                                        struct ib_phys_buf *buffer_list,
675                                        int                 num_phys_buf,
676                                        int                 acc,
677                                        u64                *iova_start)
678 {
679         struct mthca_mr *mr;
680         u64 *page_list;
681         u64 total_size;
682         u64 mask;
683         int shift;
684         int npages;
685         int err;
686         int i, j, n;
687
688         /* First check that we have enough alignment */
689         if ((*iova_start & ~PAGE_MASK) != (buffer_list[0].addr & ~PAGE_MASK))
690                 return ERR_PTR(-EINVAL);
691
692         if (num_phys_buf > 1 &&
693             ((buffer_list[0].addr + buffer_list[0].size) & ~PAGE_MASK))
694                 return ERR_PTR(-EINVAL);
695
696         mask = 0;
697         total_size = 0;
698         for (i = 0; i < num_phys_buf; ++i) {
699                 if (i != 0 && buffer_list[i].addr & ~PAGE_MASK)
700                         return ERR_PTR(-EINVAL);
701                 if (i != 0 && i != num_phys_buf - 1 &&
702                     (buffer_list[i].size & ~PAGE_MASK))
703                         return ERR_PTR(-EINVAL);
704
705                 total_size += buffer_list[i].size;
706                 if (i > 0)
707                         mask |= buffer_list[i].addr;
708         }
709
710         /* Find largest page shift we can use to cover buffers */
711         for (shift = PAGE_SHIFT; shift < 31; ++shift)
712                 if (num_phys_buf > 1) {
713                         if ((1ULL << shift) & mask)
714                                 break;
715                 } else {
716                         if (1ULL << shift >=
717                             buffer_list[0].size +
718                             (buffer_list[0].addr & ((1ULL << shift) - 1)))
719                                 break;
720                 }
721
722         buffer_list[0].size += buffer_list[0].addr & ((1ULL << shift) - 1);
723         buffer_list[0].addr &= ~0ull << shift;
724
725         mr = kmalloc(sizeof *mr, GFP_KERNEL);
726         if (!mr)
727                 return ERR_PTR(-ENOMEM);
728
729         npages = 0;
730         for (i = 0; i < num_phys_buf; ++i)
731                 npages += (buffer_list[i].size + (1ULL << shift) - 1) >> shift;
732
733         if (!npages)
734                 return &mr->ibmr;
735
736         page_list = kmalloc(npages * sizeof *page_list, GFP_KERNEL);
737         if (!page_list) {
738                 kfree(mr);
739                 return ERR_PTR(-ENOMEM);
740         }
741
742         n = 0;
743         for (i = 0; i < num_phys_buf; ++i)
744                 for (j = 0;
745                      j < (buffer_list[i].size + (1ULL << shift) - 1) >> shift;
746                      ++j)
747                         page_list[n++] = buffer_list[i].addr + ((u64) j << shift);
748
749         mthca_dbg(to_mdev(pd->device), "Registering memory at %llx (iova %llx) "
750                   "in PD %x; shift %d, npages %d.\n",
751                   (unsigned long long) buffer_list[0].addr,
752                   (unsigned long long) *iova_start,
753                   to_mpd(pd)->pd_num,
754                   shift, npages);
755
756         err = mthca_mr_alloc_phys(to_mdev(pd->device),
757                                   to_mpd(pd)->pd_num,
758                                   page_list, shift, npages,
759                                   *iova_start, total_size,
760                                   convert_access(acc), mr);
761
762         if (err) {
763                 kfree(page_list);
764                 kfree(mr);
765                 return ERR_PTR(err);
766         }
767
768         kfree(page_list);
769         return &mr->ibmr;
770 }
771
772 static struct ib_mr *mthca_reg_user_mr(struct ib_pd *pd, struct ib_umem *region,
773                                        int acc, struct ib_udata *udata)
774 {
775         struct mthca_dev *dev = to_mdev(pd->device);
776         struct ib_umem_chunk *chunk;
777         struct mthca_mr *mr;
778         u64 *pages;
779         int shift, n, len;
780         int i, j, k;
781         int err = 0;
782
783         shift = ffs(region->page_size) - 1;
784
785         mr = kmalloc(sizeof *mr, GFP_KERNEL);
786         if (!mr)
787                 return ERR_PTR(-ENOMEM);
788
789         n = 0;
790         list_for_each_entry(chunk, &region->chunk_list, list)
791                 n += chunk->nents;
792
793         mr->mtt = mthca_alloc_mtt(dev, n);
794         if (IS_ERR(mr->mtt)) {
795                 err = PTR_ERR(mr->mtt);
796                 goto err;
797         }
798
799         pages = (u64 *) __get_free_page(GFP_KERNEL);
800         if (!pages) {
801                 err = -ENOMEM;
802                 goto err_mtt;
803         }
804
805         i = n = 0;
806
807         list_for_each_entry(chunk, &region->chunk_list, list)
808                 for (j = 0; j < chunk->nmap; ++j) {
809                         len = sg_dma_len(&chunk->page_list[j]) >> shift;
810                         for (k = 0; k < len; ++k) {
811                                 pages[i++] = sg_dma_address(&chunk->page_list[j]) +
812                                         region->page_size * k;
813                                 /*
814                                  * Be friendly to WRITE_MTT command
815                                  * and leave two empty slots for the
816                                  * index and reserved fields of the
817                                  * mailbox.
818                                  */
819                                 if (i == PAGE_SIZE / sizeof (u64) - 2) {
820                                         err = mthca_write_mtt(dev, mr->mtt,
821                                                               n, pages, i);
822                                         if (err)
823                                                 goto mtt_done;
824                                         n += i;
825                                         i = 0;
826                                 }
827                         }
828                 }
829
830         if (i)
831                 err = mthca_write_mtt(dev, mr->mtt, n, pages, i);
832 mtt_done:
833         free_page((unsigned long) pages);
834         if (err)
835                 goto err_mtt;
836
837         err = mthca_mr_alloc(dev, to_mpd(pd)->pd_num, shift, region->virt_base,
838                              region->length, convert_access(acc), mr);
839
840         if (err)
841                 goto err_mtt;
842
843         return &mr->ibmr;
844
845 err_mtt:
846         mthca_free_mtt(dev, mr->mtt);
847
848 err:
849         kfree(mr);
850         return ERR_PTR(err);
851 }
852
853 static int mthca_dereg_mr(struct ib_mr *mr)
854 {
855         struct mthca_mr *mmr = to_mmr(mr);
856         mthca_free_mr(to_mdev(mr->device), mmr);
857         kfree(mmr);
858         return 0;
859 }
860
861 static struct ib_fmr *mthca_alloc_fmr(struct ib_pd *pd, int mr_access_flags,
862                                       struct ib_fmr_attr *fmr_attr)
863 {
864         struct mthca_fmr *fmr;
865         int err;
866
867         fmr = kmalloc(sizeof *fmr, GFP_KERNEL);
868         if (!fmr)
869                 return ERR_PTR(-ENOMEM);
870
871         memcpy(&fmr->attr, fmr_attr, sizeof *fmr_attr);
872         err = mthca_fmr_alloc(to_mdev(pd->device), to_mpd(pd)->pd_num,
873                              convert_access(mr_access_flags), fmr);
874
875         if (err) {
876                 kfree(fmr);
877                 return ERR_PTR(err);
878         }
879
880         return &fmr->ibmr;
881 }
882
883 static int mthca_dealloc_fmr(struct ib_fmr *fmr)
884 {
885         struct mthca_fmr *mfmr = to_mfmr(fmr);
886         int err;
887
888         err = mthca_free_fmr(to_mdev(fmr->device), mfmr);
889         if (err)
890                 return err;
891
892         kfree(mfmr);
893         return 0;
894 }
895
896 static int mthca_unmap_fmr(struct list_head *fmr_list)
897 {
898         struct ib_fmr *fmr;
899         int err;
900         u8 status;
901         struct mthca_dev *mdev = NULL;
902
903         list_for_each_entry(fmr, fmr_list, list) {
904                 if (mdev && to_mdev(fmr->device) != mdev)
905                         return -EINVAL;
906                 mdev = to_mdev(fmr->device);
907         }
908
909         if (!mdev)
910                 return 0;
911
912         if (mthca_is_memfree(mdev)) {
913                 list_for_each_entry(fmr, fmr_list, list)
914                         mthca_arbel_fmr_unmap(mdev, to_mfmr(fmr));
915
916                 wmb();
917         } else
918                 list_for_each_entry(fmr, fmr_list, list)
919                         mthca_tavor_fmr_unmap(mdev, to_mfmr(fmr));
920
921         err = mthca_SYNC_TPT(mdev, &status);
922         if (err)
923                 return err;
924         if (status)
925                 return -EINVAL;
926         return 0;
927 }
928
929 static ssize_t show_rev(struct class_device *cdev, char *buf)
930 {
931         struct mthca_dev *dev = container_of(cdev, struct mthca_dev, ib_dev.class_dev);
932         return sprintf(buf, "%x\n", dev->rev_id);
933 }
934
935 static ssize_t show_fw_ver(struct class_device *cdev, char *buf)
936 {
937         struct mthca_dev *dev = container_of(cdev, struct mthca_dev, ib_dev.class_dev);
938         return sprintf(buf, "%x.%x.%x\n", (int) (dev->fw_ver >> 32),
939                        (int) (dev->fw_ver >> 16) & 0xffff,
940                        (int) dev->fw_ver & 0xffff);
941 }
942
943 static ssize_t show_hca(struct class_device *cdev, char *buf)
944 {
945         struct mthca_dev *dev = container_of(cdev, struct mthca_dev, ib_dev.class_dev);
946         switch (dev->pdev->device) {
947         case PCI_DEVICE_ID_MELLANOX_TAVOR:
948                 return sprintf(buf, "MT23108\n");
949         case PCI_DEVICE_ID_MELLANOX_ARBEL_COMPAT:
950                 return sprintf(buf, "MT25208 (MT23108 compat mode)\n");
951         case PCI_DEVICE_ID_MELLANOX_ARBEL:
952                 return sprintf(buf, "MT25208\n");
953         case PCI_DEVICE_ID_MELLANOX_SINAI:
954         case PCI_DEVICE_ID_MELLANOX_SINAI_OLD:
955                 return sprintf(buf, "MT25204\n");
956         default:
957                 return sprintf(buf, "unknown\n");
958         }
959 }
960
961 static CLASS_DEVICE_ATTR(hw_rev,   S_IRUGO, show_rev,    NULL);
962 static CLASS_DEVICE_ATTR(fw_ver,   S_IRUGO, show_fw_ver, NULL);
963 static CLASS_DEVICE_ATTR(hca_type, S_IRUGO, show_hca,    NULL);
964
965 static struct class_device_attribute *mthca_class_attributes[] = {
966         &class_device_attr_hw_rev,
967         &class_device_attr_fw_ver,
968         &class_device_attr_hca_type
969 };
970
971 int mthca_register_device(struct mthca_dev *dev)
972 {
973         int ret;
974         int i;
975
976         strlcpy(dev->ib_dev.name, "mthca%d", IB_DEVICE_NAME_MAX);
977         dev->ib_dev.owner                = THIS_MODULE;
978
979         dev->ib_dev.node_type            = IB_NODE_CA;
980         dev->ib_dev.phys_port_cnt        = dev->limits.num_ports;
981         dev->ib_dev.dma_device           = &dev->pdev->dev;
982         dev->ib_dev.class_dev.dev        = &dev->pdev->dev;
983         dev->ib_dev.query_device         = mthca_query_device;
984         dev->ib_dev.query_port           = mthca_query_port;
985         dev->ib_dev.modify_port          = mthca_modify_port;
986         dev->ib_dev.query_pkey           = mthca_query_pkey;
987         dev->ib_dev.query_gid            = mthca_query_gid;
988         dev->ib_dev.alloc_ucontext       = mthca_alloc_ucontext;
989         dev->ib_dev.dealloc_ucontext     = mthca_dealloc_ucontext;
990         dev->ib_dev.mmap                 = mthca_mmap_uar;
991         dev->ib_dev.alloc_pd             = mthca_alloc_pd;
992         dev->ib_dev.dealloc_pd           = mthca_dealloc_pd;
993         dev->ib_dev.create_ah            = mthca_ah_create;
994         dev->ib_dev.destroy_ah           = mthca_ah_destroy;
995         dev->ib_dev.create_qp            = mthca_create_qp;
996         dev->ib_dev.modify_qp            = mthca_modify_qp;
997         dev->ib_dev.destroy_qp           = mthca_destroy_qp;
998         dev->ib_dev.create_cq            = mthca_create_cq;
999         dev->ib_dev.destroy_cq           = mthca_destroy_cq;
1000         dev->ib_dev.poll_cq              = mthca_poll_cq;
1001         dev->ib_dev.get_dma_mr           = mthca_get_dma_mr;
1002         dev->ib_dev.reg_phys_mr          = mthca_reg_phys_mr;
1003         dev->ib_dev.reg_user_mr          = mthca_reg_user_mr;
1004         dev->ib_dev.dereg_mr             = mthca_dereg_mr;
1005
1006         if (dev->mthca_flags & MTHCA_FLAG_FMR) {
1007                 dev->ib_dev.alloc_fmr            = mthca_alloc_fmr;
1008                 dev->ib_dev.unmap_fmr            = mthca_unmap_fmr;
1009                 dev->ib_dev.dealloc_fmr          = mthca_dealloc_fmr;
1010                 if (mthca_is_memfree(dev))
1011                         dev->ib_dev.map_phys_fmr = mthca_arbel_map_phys_fmr;
1012                 else
1013                         dev->ib_dev.map_phys_fmr = mthca_tavor_map_phys_fmr;
1014         }
1015
1016         dev->ib_dev.attach_mcast         = mthca_multicast_attach;
1017         dev->ib_dev.detach_mcast         = mthca_multicast_detach;
1018         dev->ib_dev.process_mad          = mthca_process_mad;
1019
1020         if (mthca_is_memfree(dev)) {
1021                 dev->ib_dev.req_notify_cq = mthca_arbel_arm_cq;
1022                 dev->ib_dev.post_send     = mthca_arbel_post_send;
1023                 dev->ib_dev.post_recv     = mthca_arbel_post_receive;
1024         } else {
1025                 dev->ib_dev.req_notify_cq = mthca_tavor_arm_cq;
1026                 dev->ib_dev.post_send     = mthca_tavor_post_send;
1027                 dev->ib_dev.post_recv     = mthca_tavor_post_receive;
1028         }
1029
1030         init_MUTEX(&dev->cap_mask_mutex);
1031
1032         ret = ib_register_device(&dev->ib_dev);
1033         if (ret)
1034                 return ret;
1035
1036         for (i = 0; i < ARRAY_SIZE(mthca_class_attributes); ++i) {
1037                 ret = class_device_create_file(&dev->ib_dev.class_dev,
1038                                                mthca_class_attributes[i]);
1039                 if (ret) {
1040                         ib_unregister_device(&dev->ib_dev);
1041                         return ret;
1042                 }
1043         }
1044
1045         return 0;
1046 }
1047
1048 void mthca_unregister_device(struct mthca_dev *dev)
1049 {
1050         ib_unregister_device(&dev->ib_dev);
1051 }