RDMA: iWARP Core Changes.
[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, 2006 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 4859 2006-01-09 21:55:10Z roland $
37  */
38
39 #include <rdma/ib_smi.h>
40 #include <rdma/ib_user_verbs.h>
41 #include <linux/mm.h>
42
43 #include "mthca_dev.h"
44 #include "mthca_cmd.h"
45 #include "mthca_user.h"
46 #include "mthca_memfree.h"
47
48 static void init_query_mad(struct ib_smp *mad)
49 {
50         mad->base_version  = 1;
51         mad->mgmt_class    = IB_MGMT_CLASS_SUBN_LID_ROUTED;
52         mad->class_version = 1;
53         mad->method        = IB_MGMT_METHOD_GET;
54 }
55
56 static int mthca_query_device(struct ib_device *ibdev,
57                               struct ib_device_attr *props)
58 {
59         struct ib_smp *in_mad  = NULL;
60         struct ib_smp *out_mad = NULL;
61         int err = -ENOMEM;
62         struct mthca_dev* mdev = to_mdev(ibdev);
63
64         u8 status;
65
66         in_mad  = kzalloc(sizeof *in_mad, GFP_KERNEL);
67         out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
68         if (!in_mad || !out_mad)
69                 goto out;
70
71         memset(props, 0, sizeof *props);
72
73         props->fw_ver              = mdev->fw_ver;
74
75         init_query_mad(in_mad);
76         in_mad->attr_id = IB_SMP_ATTR_NODE_INFO;
77
78         err = mthca_MAD_IFC(mdev, 1, 1,
79                             1, NULL, NULL, in_mad, out_mad,
80                             &status);
81         if (err)
82                 goto out;
83         if (status) {
84                 err = -EINVAL;
85                 goto out;
86         }
87
88         props->device_cap_flags    = mdev->device_cap_flags;
89         props->vendor_id           = be32_to_cpup((__be32 *) (out_mad->data + 36)) &
90                 0xffffff;
91         props->vendor_part_id      = be16_to_cpup((__be16 *) (out_mad->data + 30));
92         props->hw_ver              = be32_to_cpup((__be32 *) (out_mad->data + 32));
93         memcpy(&props->sys_image_guid, out_mad->data +  4, 8);
94
95         props->max_mr_size         = ~0ull;
96         props->page_size_cap       = mdev->limits.page_size_cap;
97         props->max_qp              = mdev->limits.num_qps - mdev->limits.reserved_qps;
98         props->max_qp_wr           = mdev->limits.max_wqes;
99         props->max_sge             = mdev->limits.max_sg;
100         props->max_cq              = mdev->limits.num_cqs - mdev->limits.reserved_cqs;
101         props->max_cqe             = mdev->limits.max_cqes;
102         props->max_mr              = mdev->limits.num_mpts - mdev->limits.reserved_mrws;
103         props->max_pd              = mdev->limits.num_pds - mdev->limits.reserved_pds;
104         props->max_qp_rd_atom      = 1 << mdev->qp_table.rdb_shift;
105         props->max_qp_init_rd_atom = mdev->limits.max_qp_init_rdma;
106         props->max_res_rd_atom     = props->max_qp_rd_atom * props->max_qp;
107         props->max_srq             = mdev->limits.num_srqs - mdev->limits.reserved_srqs;
108         props->max_srq_wr          = mdev->limits.max_srq_wqes;
109         props->max_srq_sge         = mdev->limits.max_srq_sge;
110         props->local_ca_ack_delay  = mdev->limits.local_ca_ack_delay;
111         props->atomic_cap          = mdev->limits.flags & DEV_LIM_FLAG_ATOMIC ?
112                                         IB_ATOMIC_HCA : IB_ATOMIC_NONE;
113         props->max_pkeys           = mdev->limits.pkey_table_len;
114         props->max_mcast_grp       = mdev->limits.num_mgms + mdev->limits.num_amgms;
115         props->max_mcast_qp_attach = MTHCA_QP_PER_MGM;
116         props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
117                                            props->max_mcast_grp;
118         /*
119          * If Sinai memory key optimization is being used, then only
120          * the 8-bit key portion will change.  For other HCAs, the
121          * unused index bits will also be used for FMR remapping.
122          */
123         if (mdev->mthca_flags & MTHCA_FLAG_SINAI_OPT)
124                 props->max_map_per_fmr = 255;
125         else
126                 props->max_map_per_fmr =
127                         (1 << (32 - long_log2(mdev->limits.num_mpts))) - 1;
128
129         err = 0;
130  out:
131         kfree(in_mad);
132         kfree(out_mad);
133         return err;
134 }
135
136 static int mthca_query_port(struct ib_device *ibdev,
137                             u8 port, struct ib_port_attr *props)
138 {
139         struct ib_smp *in_mad  = NULL;
140         struct ib_smp *out_mad = NULL;
141         int err = -ENOMEM;
142         u8 status;
143
144         in_mad  = kzalloc(sizeof *in_mad, GFP_KERNEL);
145         out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
146         if (!in_mad || !out_mad)
147                 goto out;
148
149         memset(props, 0, sizeof *props);
150
151         init_query_mad(in_mad);
152         in_mad->attr_id  = IB_SMP_ATTR_PORT_INFO;
153         in_mad->attr_mod = cpu_to_be32(port);
154
155         err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
156                             port, NULL, NULL, in_mad, out_mad,
157                             &status);
158         if (err)
159                 goto out;
160         if (status) {
161                 err = -EINVAL;
162                 goto out;
163         }
164
165         props->lid               = be16_to_cpup((__be16 *) (out_mad->data + 16));
166         props->lmc               = out_mad->data[34] & 0x7;
167         props->sm_lid            = be16_to_cpup((__be16 *) (out_mad->data + 18));
168         props->sm_sl             = out_mad->data[36] & 0xf;
169         props->state             = out_mad->data[32] & 0xf;
170         props->phys_state        = out_mad->data[33] >> 4;
171         props->port_cap_flags    = be32_to_cpup((__be32 *) (out_mad->data + 20));
172         props->gid_tbl_len       = to_mdev(ibdev)->limits.gid_table_len;
173         props->max_msg_sz        = 0x80000000;
174         props->pkey_tbl_len      = to_mdev(ibdev)->limits.pkey_table_len;
175         props->bad_pkey_cntr     = be16_to_cpup((__be16 *) (out_mad->data + 46));
176         props->qkey_viol_cntr    = be16_to_cpup((__be16 *) (out_mad->data + 48));
177         props->active_width      = out_mad->data[31] & 0xf;
178         props->active_speed      = out_mad->data[35] >> 4;
179         props->max_mtu           = out_mad->data[41] & 0xf;
180         props->active_mtu        = out_mad->data[36] >> 4;
181         props->subnet_timeout    = out_mad->data[51] & 0x1f;
182
183  out:
184         kfree(in_mad);
185         kfree(out_mad);
186         return err;
187 }
188
189 static int mthca_modify_device(struct ib_device *ibdev,
190                                int mask,
191                                struct ib_device_modify *props)
192 {
193         if (mask & ~IB_DEVICE_MODIFY_NODE_DESC)
194                 return -EOPNOTSUPP;
195
196         if (mask & IB_DEVICE_MODIFY_NODE_DESC) {
197                 if (mutex_lock_interruptible(&to_mdev(ibdev)->cap_mask_mutex))
198                         return -ERESTARTSYS;
199                 memcpy(ibdev->node_desc, props->node_desc, 64);
200                 mutex_unlock(&to_mdev(ibdev)->cap_mask_mutex);
201         }
202
203         return 0;
204 }
205
206 static int mthca_modify_port(struct ib_device *ibdev,
207                              u8 port, int port_modify_mask,
208                              struct ib_port_modify *props)
209 {
210         struct mthca_set_ib_param set_ib;
211         struct ib_port_attr attr;
212         int err;
213         u8 status;
214
215         if (mutex_lock_interruptible(&to_mdev(ibdev)->cap_mask_mutex))
216                 return -ERESTARTSYS;
217
218         err = mthca_query_port(ibdev, port, &attr);
219         if (err)
220                 goto out;
221
222         set_ib.set_si_guid     = 0;
223         set_ib.reset_qkey_viol = !!(port_modify_mask & IB_PORT_RESET_QKEY_CNTR);
224
225         set_ib.cap_mask = (attr.port_cap_flags | props->set_port_cap_mask) &
226                 ~props->clr_port_cap_mask;
227
228         err = mthca_SET_IB(to_mdev(ibdev), &set_ib, port, &status);
229         if (err)
230                 goto out;
231         if (status) {
232                 err = -EINVAL;
233                 goto out;
234         }
235
236 out:
237         mutex_unlock(&to_mdev(ibdev)->cap_mask_mutex);
238         return err;
239 }
240
241 static int mthca_query_pkey(struct ib_device *ibdev,
242                             u8 port, u16 index, u16 *pkey)
243 {
244         struct ib_smp *in_mad  = NULL;
245         struct ib_smp *out_mad = NULL;
246         int err = -ENOMEM;
247         u8 status;
248
249         in_mad  = kzalloc(sizeof *in_mad, GFP_KERNEL);
250         out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
251         if (!in_mad || !out_mad)
252                 goto out;
253
254         init_query_mad(in_mad);
255         in_mad->attr_id  = IB_SMP_ATTR_PKEY_TABLE;
256         in_mad->attr_mod = cpu_to_be32(index / 32);
257
258         err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
259                             port, NULL, NULL, in_mad, out_mad,
260                             &status);
261         if (err)
262                 goto out;
263         if (status) {
264                 err = -EINVAL;
265                 goto out;
266         }
267
268         *pkey = be16_to_cpu(((__be16 *) out_mad->data)[index % 32]);
269
270  out:
271         kfree(in_mad);
272         kfree(out_mad);
273         return err;
274 }
275
276 static int mthca_query_gid(struct ib_device *ibdev, u8 port,
277                            int index, union ib_gid *gid)
278 {
279         struct ib_smp *in_mad  = NULL;
280         struct ib_smp *out_mad = NULL;
281         int err = -ENOMEM;
282         u8 status;
283
284         in_mad  = kzalloc(sizeof *in_mad, GFP_KERNEL);
285         out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
286         if (!in_mad || !out_mad)
287                 goto out;
288
289         init_query_mad(in_mad);
290         in_mad->attr_id  = IB_SMP_ATTR_PORT_INFO;
291         in_mad->attr_mod = cpu_to_be32(port);
292
293         err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
294                             port, NULL, NULL, in_mad, out_mad,
295                             &status);
296         if (err)
297                 goto out;
298         if (status) {
299                 err = -EINVAL;
300                 goto out;
301         }
302
303         memcpy(gid->raw, out_mad->data + 8, 8);
304
305         init_query_mad(in_mad);
306         in_mad->attr_id  = IB_SMP_ATTR_GUID_INFO;
307         in_mad->attr_mod = cpu_to_be32(index / 8);
308
309         err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
310                             port, NULL, NULL, in_mad, out_mad,
311                             &status);
312         if (err)
313                 goto out;
314         if (status) {
315                 err = -EINVAL;
316                 goto out;
317         }
318
319         memcpy(gid->raw + 8, out_mad->data + (index % 8) * 8, 8);
320
321  out:
322         kfree(in_mad);
323         kfree(out_mad);
324         return err;
325 }
326
327 static struct ib_ucontext *mthca_alloc_ucontext(struct ib_device *ibdev,
328                                                 struct ib_udata *udata)
329 {
330         struct mthca_alloc_ucontext_resp uresp;
331         struct mthca_ucontext           *context;
332         int                              err;
333
334         memset(&uresp, 0, sizeof uresp);
335
336         uresp.qp_tab_size = to_mdev(ibdev)->limits.num_qps;
337         if (mthca_is_memfree(to_mdev(ibdev)))
338                 uresp.uarc_size = to_mdev(ibdev)->uar_table.uarc_size;
339         else
340                 uresp.uarc_size = 0;
341
342         context = kmalloc(sizeof *context, GFP_KERNEL);
343         if (!context)
344                 return ERR_PTR(-ENOMEM);
345
346         err = mthca_uar_alloc(to_mdev(ibdev), &context->uar);
347         if (err) {
348                 kfree(context);
349                 return ERR_PTR(err);
350         }
351
352         context->db_tab = mthca_init_user_db_tab(to_mdev(ibdev));
353         if (IS_ERR(context->db_tab)) {
354                 err = PTR_ERR(context->db_tab);
355                 mthca_uar_free(to_mdev(ibdev), &context->uar);
356                 kfree(context);
357                 return ERR_PTR(err);
358         }
359
360         if (ib_copy_to_udata(udata, &uresp, sizeof uresp)) {
361                 mthca_cleanup_user_db_tab(to_mdev(ibdev), &context->uar, context->db_tab);
362                 mthca_uar_free(to_mdev(ibdev), &context->uar);
363                 kfree(context);
364                 return ERR_PTR(-EFAULT);
365         }
366
367         return &context->ibucontext;
368 }
369
370 static int mthca_dealloc_ucontext(struct ib_ucontext *context)
371 {
372         mthca_cleanup_user_db_tab(to_mdev(context->device), &to_mucontext(context)->uar,
373                                   to_mucontext(context)->db_tab);
374         mthca_uar_free(to_mdev(context->device), &to_mucontext(context)->uar);
375         kfree(to_mucontext(context));
376
377         return 0;
378 }
379
380 static int mthca_mmap_uar(struct ib_ucontext *context,
381                           struct vm_area_struct *vma)
382 {
383         if (vma->vm_end - vma->vm_start != PAGE_SIZE)
384                 return -EINVAL;
385
386         vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
387
388         if (io_remap_pfn_range(vma, vma->vm_start,
389                                to_mucontext(context)->uar.pfn,
390                                PAGE_SIZE, vma->vm_page_prot))
391                 return -EAGAIN;
392
393         return 0;
394 }
395
396 static struct ib_pd *mthca_alloc_pd(struct ib_device *ibdev,
397                                     struct ib_ucontext *context,
398                                     struct ib_udata *udata)
399 {
400         struct mthca_pd *pd;
401         int err;
402
403         pd = kmalloc(sizeof *pd, GFP_KERNEL);
404         if (!pd)
405                 return ERR_PTR(-ENOMEM);
406
407         err = mthca_pd_alloc(to_mdev(ibdev), !context, pd);
408         if (err) {
409                 kfree(pd);
410                 return ERR_PTR(err);
411         }
412
413         if (context) {
414                 if (ib_copy_to_udata(udata, &pd->pd_num, sizeof (__u32))) {
415                         mthca_pd_free(to_mdev(ibdev), pd);
416                         kfree(pd);
417                         return ERR_PTR(-EFAULT);
418                 }
419         }
420
421         return &pd->ibpd;
422 }
423
424 static int mthca_dealloc_pd(struct ib_pd *pd)
425 {
426         mthca_pd_free(to_mdev(pd->device), to_mpd(pd));
427         kfree(pd);
428
429         return 0;
430 }
431
432 static struct ib_ah *mthca_ah_create(struct ib_pd *pd,
433                                      struct ib_ah_attr *ah_attr)
434 {
435         int err;
436         struct mthca_ah *ah;
437
438         ah = kmalloc(sizeof *ah, GFP_ATOMIC);
439         if (!ah)
440                 return ERR_PTR(-ENOMEM);
441
442         err = mthca_create_ah(to_mdev(pd->device), to_mpd(pd), ah_attr, ah);
443         if (err) {
444                 kfree(ah);
445                 return ERR_PTR(err);
446         }
447
448         return &ah->ibah;
449 }
450
451 static int mthca_ah_destroy(struct ib_ah *ah)
452 {
453         mthca_destroy_ah(to_mdev(ah->device), to_mah(ah));
454         kfree(ah);
455
456         return 0;
457 }
458
459 static struct ib_srq *mthca_create_srq(struct ib_pd *pd,
460                                        struct ib_srq_init_attr *init_attr,
461                                        struct ib_udata *udata)
462 {
463         struct mthca_create_srq ucmd;
464         struct mthca_ucontext *context = NULL;
465         struct mthca_srq *srq;
466         int err;
467
468         srq = kmalloc(sizeof *srq, GFP_KERNEL);
469         if (!srq)
470                 return ERR_PTR(-ENOMEM);
471
472         if (pd->uobject) {
473                 context = to_mucontext(pd->uobject->context);
474
475                 if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
476                         err = -EFAULT;
477                         goto err_free;
478                 }
479
480                 err = mthca_map_user_db(to_mdev(pd->device), &context->uar,
481                                         context->db_tab, ucmd.db_index,
482                                         ucmd.db_page);
483
484                 if (err)
485                         goto err_free;
486
487                 srq->mr.ibmr.lkey = ucmd.lkey;
488                 srq->db_index     = ucmd.db_index;
489         }
490
491         err = mthca_alloc_srq(to_mdev(pd->device), to_mpd(pd),
492                               &init_attr->attr, srq);
493
494         if (err && pd->uobject)
495                 mthca_unmap_user_db(to_mdev(pd->device), &context->uar,
496                                     context->db_tab, ucmd.db_index);
497
498         if (err)
499                 goto err_free;
500
501         if (context && ib_copy_to_udata(udata, &srq->srqn, sizeof (__u32))) {
502                 mthca_free_srq(to_mdev(pd->device), srq);
503                 err = -EFAULT;
504                 goto err_free;
505         }
506
507         return &srq->ibsrq;
508
509 err_free:
510         kfree(srq);
511
512         return ERR_PTR(err);
513 }
514
515 static int mthca_destroy_srq(struct ib_srq *srq)
516 {
517         struct mthca_ucontext *context;
518
519         if (srq->uobject) {
520                 context = to_mucontext(srq->uobject->context);
521
522                 mthca_unmap_user_db(to_mdev(srq->device), &context->uar,
523                                     context->db_tab, to_msrq(srq)->db_index);
524         }
525
526         mthca_free_srq(to_mdev(srq->device), to_msrq(srq));
527         kfree(srq);
528
529         return 0;
530 }
531
532 static struct ib_qp *mthca_create_qp(struct ib_pd *pd,
533                                      struct ib_qp_init_attr *init_attr,
534                                      struct ib_udata *udata)
535 {
536         struct mthca_create_qp ucmd;
537         struct mthca_qp *qp;
538         int err;
539
540         switch (init_attr->qp_type) {
541         case IB_QPT_RC:
542         case IB_QPT_UC:
543         case IB_QPT_UD:
544         {
545                 struct mthca_ucontext *context;
546
547                 qp = kmalloc(sizeof *qp, GFP_KERNEL);
548                 if (!qp)
549                         return ERR_PTR(-ENOMEM);
550
551                 if (pd->uobject) {
552                         context = to_mucontext(pd->uobject->context);
553
554                         if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
555                                 kfree(qp);
556                                 return ERR_PTR(-EFAULT);
557                         }
558
559                         err = mthca_map_user_db(to_mdev(pd->device), &context->uar,
560                                                 context->db_tab,
561                                                 ucmd.sq_db_index, ucmd.sq_db_page);
562                         if (err) {
563                                 kfree(qp);
564                                 return ERR_PTR(err);
565                         }
566
567                         err = mthca_map_user_db(to_mdev(pd->device), &context->uar,
568                                                 context->db_tab,
569                                                 ucmd.rq_db_index, ucmd.rq_db_page);
570                         if (err) {
571                                 mthca_unmap_user_db(to_mdev(pd->device),
572                                                     &context->uar,
573                                                     context->db_tab,
574                                                     ucmd.sq_db_index);
575                                 kfree(qp);
576                                 return ERR_PTR(err);
577                         }
578
579                         qp->mr.ibmr.lkey = ucmd.lkey;
580                         qp->sq.db_index  = ucmd.sq_db_index;
581                         qp->rq.db_index  = ucmd.rq_db_index;
582                 }
583
584                 err = mthca_alloc_qp(to_mdev(pd->device), to_mpd(pd),
585                                      to_mcq(init_attr->send_cq),
586                                      to_mcq(init_attr->recv_cq),
587                                      init_attr->qp_type, init_attr->sq_sig_type,
588                                      &init_attr->cap, qp);
589
590                 if (err && pd->uobject) {
591                         context = to_mucontext(pd->uobject->context);
592
593                         mthca_unmap_user_db(to_mdev(pd->device),
594                                             &context->uar,
595                                             context->db_tab,
596                                             ucmd.sq_db_index);
597                         mthca_unmap_user_db(to_mdev(pd->device),
598                                             &context->uar,
599                                             context->db_tab,
600                                             ucmd.rq_db_index);
601                 }
602
603                 qp->ibqp.qp_num = qp->qpn;
604                 break;
605         }
606         case IB_QPT_SMI:
607         case IB_QPT_GSI:
608         {
609                 /* Don't allow userspace to create special QPs */
610                 if (pd->uobject)
611                         return ERR_PTR(-EINVAL);
612
613                 qp = kmalloc(sizeof (struct mthca_sqp), GFP_KERNEL);
614                 if (!qp)
615                         return ERR_PTR(-ENOMEM);
616
617                 qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 : 1;
618
619                 err = mthca_alloc_sqp(to_mdev(pd->device), to_mpd(pd),
620                                       to_mcq(init_attr->send_cq),
621                                       to_mcq(init_attr->recv_cq),
622                                       init_attr->sq_sig_type, &init_attr->cap,
623                                       qp->ibqp.qp_num, init_attr->port_num,
624                                       to_msqp(qp));
625                 break;
626         }
627         default:
628                 /* Don't support raw QPs */
629                 return ERR_PTR(-ENOSYS);
630         }
631
632         if (err) {
633                 kfree(qp);
634                 return ERR_PTR(err);
635         }
636
637         init_attr->cap.max_send_wr     = qp->sq.max;
638         init_attr->cap.max_recv_wr     = qp->rq.max;
639         init_attr->cap.max_send_sge    = qp->sq.max_gs;
640         init_attr->cap.max_recv_sge    = qp->rq.max_gs;
641         init_attr->cap.max_inline_data = qp->max_inline_data;
642
643         return &qp->ibqp;
644 }
645
646 static int mthca_destroy_qp(struct ib_qp *qp)
647 {
648         if (qp->uobject) {
649                 mthca_unmap_user_db(to_mdev(qp->device),
650                                     &to_mucontext(qp->uobject->context)->uar,
651                                     to_mucontext(qp->uobject->context)->db_tab,
652                                     to_mqp(qp)->sq.db_index);
653                 mthca_unmap_user_db(to_mdev(qp->device),
654                                     &to_mucontext(qp->uobject->context)->uar,
655                                     to_mucontext(qp->uobject->context)->db_tab,
656                                     to_mqp(qp)->rq.db_index);
657         }
658         mthca_free_qp(to_mdev(qp->device), to_mqp(qp));
659         kfree(qp);
660         return 0;
661 }
662
663 static struct ib_cq *mthca_create_cq(struct ib_device *ibdev, int entries,
664                                      struct ib_ucontext *context,
665                                      struct ib_udata *udata)
666 {
667         struct mthca_create_cq ucmd;
668         struct mthca_cq *cq;
669         int nent;
670         int err;
671
672         if (entries < 1 || entries > to_mdev(ibdev)->limits.max_cqes)
673                 return ERR_PTR(-EINVAL);
674
675         if (context) {
676                 if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd))
677                         return ERR_PTR(-EFAULT);
678
679                 err = mthca_map_user_db(to_mdev(ibdev), &to_mucontext(context)->uar,
680                                         to_mucontext(context)->db_tab,
681                                         ucmd.set_db_index, ucmd.set_db_page);
682                 if (err)
683                         return ERR_PTR(err);
684
685                 err = mthca_map_user_db(to_mdev(ibdev), &to_mucontext(context)->uar,
686                                         to_mucontext(context)->db_tab,
687                                         ucmd.arm_db_index, ucmd.arm_db_page);
688                 if (err)
689                         goto err_unmap_set;
690         }
691
692         cq = kmalloc(sizeof *cq, GFP_KERNEL);
693         if (!cq) {
694                 err = -ENOMEM;
695                 goto err_unmap_arm;
696         }
697
698         if (context) {
699                 cq->buf.mr.ibmr.lkey = ucmd.lkey;
700                 cq->set_ci_db_index  = ucmd.set_db_index;
701                 cq->arm_db_index     = ucmd.arm_db_index;
702         }
703
704         for (nent = 1; nent <= entries; nent <<= 1)
705                 ; /* nothing */
706
707         err = mthca_init_cq(to_mdev(ibdev), nent,
708                             context ? to_mucontext(context) : NULL,
709                             context ? ucmd.pdn : to_mdev(ibdev)->driver_pd.pd_num,
710                             cq);
711         if (err)
712                 goto err_free;
713
714         if (context && ib_copy_to_udata(udata, &cq->cqn, sizeof (__u32))) {
715                 mthca_free_cq(to_mdev(ibdev), cq);
716                 goto err_free;
717         }
718
719         cq->resize_buf = NULL;
720
721         return &cq->ibcq;
722
723 err_free:
724         kfree(cq);
725
726 err_unmap_arm:
727         if (context)
728                 mthca_unmap_user_db(to_mdev(ibdev), &to_mucontext(context)->uar,
729                                     to_mucontext(context)->db_tab, ucmd.arm_db_index);
730
731 err_unmap_set:
732         if (context)
733                 mthca_unmap_user_db(to_mdev(ibdev), &to_mucontext(context)->uar,
734                                     to_mucontext(context)->db_tab, ucmd.set_db_index);
735
736         return ERR_PTR(err);
737 }
738
739 static int mthca_alloc_resize_buf(struct mthca_dev *dev, struct mthca_cq *cq,
740                                   int entries)
741 {
742         int ret;
743
744         spin_lock_irq(&cq->lock);
745         if (cq->resize_buf) {
746                 ret = -EBUSY;
747                 goto unlock;
748         }
749
750         cq->resize_buf = kmalloc(sizeof *cq->resize_buf, GFP_ATOMIC);
751         if (!cq->resize_buf) {
752                 ret = -ENOMEM;
753                 goto unlock;
754         }
755
756         cq->resize_buf->state = CQ_RESIZE_ALLOC;
757
758         ret = 0;
759
760 unlock:
761         spin_unlock_irq(&cq->lock);
762
763         if (ret)
764                 return ret;
765
766         ret = mthca_alloc_cq_buf(dev, &cq->resize_buf->buf, entries);
767         if (ret) {
768                 spin_lock_irq(&cq->lock);
769                 kfree(cq->resize_buf);
770                 cq->resize_buf = NULL;
771                 spin_unlock_irq(&cq->lock);
772                 return ret;
773         }
774
775         cq->resize_buf->cqe = entries - 1;
776
777         spin_lock_irq(&cq->lock);
778         cq->resize_buf->state = CQ_RESIZE_READY;
779         spin_unlock_irq(&cq->lock);
780
781         return 0;
782 }
783
784 static int mthca_resize_cq(struct ib_cq *ibcq, int entries, struct ib_udata *udata)
785 {
786         struct mthca_dev *dev = to_mdev(ibcq->device);
787         struct mthca_cq *cq = to_mcq(ibcq);
788         struct mthca_resize_cq ucmd;
789         u32 lkey;
790         u8 status;
791         int ret;
792
793         if (entries < 1 || entries > dev->limits.max_cqes)
794                 return -EINVAL;
795
796         mutex_lock(&cq->mutex);
797
798         entries = roundup_pow_of_two(entries + 1);
799         if (entries == ibcq->cqe + 1) {
800                 ret = 0;
801                 goto out;
802         }
803
804         if (cq->is_kernel) {
805                 ret = mthca_alloc_resize_buf(dev, cq, entries);
806                 if (ret)
807                         goto out;
808                 lkey = cq->resize_buf->buf.mr.ibmr.lkey;
809         } else {
810                 if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
811                         ret = -EFAULT;
812                         goto out;
813                 }
814                 lkey = ucmd.lkey;
815         }
816
817         ret = mthca_RESIZE_CQ(dev, cq->cqn, lkey, long_log2(entries), &status);
818         if (status)
819                 ret = -EINVAL;
820
821         if (ret) {
822                 if (cq->resize_buf) {
823                         mthca_free_cq_buf(dev, &cq->resize_buf->buf,
824                                           cq->resize_buf->cqe);
825                         kfree(cq->resize_buf);
826                         spin_lock_irq(&cq->lock);
827                         cq->resize_buf = NULL;
828                         spin_unlock_irq(&cq->lock);
829                 }
830                 goto out;
831         }
832
833         if (cq->is_kernel) {
834                 struct mthca_cq_buf tbuf;
835                 int tcqe;
836
837                 spin_lock_irq(&cq->lock);
838                 if (cq->resize_buf->state == CQ_RESIZE_READY) {
839                         mthca_cq_resize_copy_cqes(cq);
840                         tbuf         = cq->buf;
841                         tcqe         = cq->ibcq.cqe;
842                         cq->buf      = cq->resize_buf->buf;
843                         cq->ibcq.cqe = cq->resize_buf->cqe;
844                 } else {
845                         tbuf = cq->resize_buf->buf;
846                         tcqe = cq->resize_buf->cqe;
847                 }
848
849                 kfree(cq->resize_buf);
850                 cq->resize_buf = NULL;
851                 spin_unlock_irq(&cq->lock);
852
853                 mthca_free_cq_buf(dev, &tbuf, tcqe);
854         } else
855                 ibcq->cqe = entries - 1;
856
857 out:
858         mutex_unlock(&cq->mutex);
859
860         return ret;
861 }
862
863 static int mthca_destroy_cq(struct ib_cq *cq)
864 {
865         if (cq->uobject) {
866                 mthca_unmap_user_db(to_mdev(cq->device),
867                                     &to_mucontext(cq->uobject->context)->uar,
868                                     to_mucontext(cq->uobject->context)->db_tab,
869                                     to_mcq(cq)->arm_db_index);
870                 mthca_unmap_user_db(to_mdev(cq->device),
871                                     &to_mucontext(cq->uobject->context)->uar,
872                                     to_mucontext(cq->uobject->context)->db_tab,
873                                     to_mcq(cq)->set_ci_db_index);
874         }
875         mthca_free_cq(to_mdev(cq->device), to_mcq(cq));
876         kfree(cq);
877
878         return 0;
879 }
880
881 static inline u32 convert_access(int acc)
882 {
883         return (acc & IB_ACCESS_REMOTE_ATOMIC ? MTHCA_MPT_FLAG_ATOMIC       : 0) |
884                (acc & IB_ACCESS_REMOTE_WRITE  ? MTHCA_MPT_FLAG_REMOTE_WRITE : 0) |
885                (acc & IB_ACCESS_REMOTE_READ   ? MTHCA_MPT_FLAG_REMOTE_READ  : 0) |
886                (acc & IB_ACCESS_LOCAL_WRITE   ? MTHCA_MPT_FLAG_LOCAL_WRITE  : 0) |
887                MTHCA_MPT_FLAG_LOCAL_READ;
888 }
889
890 static struct ib_mr *mthca_get_dma_mr(struct ib_pd *pd, int acc)
891 {
892         struct mthca_mr *mr;
893         int err;
894
895         mr = kmalloc(sizeof *mr, GFP_KERNEL);
896         if (!mr)
897                 return ERR_PTR(-ENOMEM);
898
899         err = mthca_mr_alloc_notrans(to_mdev(pd->device),
900                                      to_mpd(pd)->pd_num,
901                                      convert_access(acc), mr);
902
903         if (err) {
904                 kfree(mr);
905                 return ERR_PTR(err);
906         }
907
908         return &mr->ibmr;
909 }
910
911 static struct ib_mr *mthca_reg_phys_mr(struct ib_pd       *pd,
912                                        struct ib_phys_buf *buffer_list,
913                                        int                 num_phys_buf,
914                                        int                 acc,
915                                        u64                *iova_start)
916 {
917         struct mthca_mr *mr;
918         u64 *page_list;
919         u64 total_size;
920         u64 mask;
921         int shift;
922         int npages;
923         int err;
924         int i, j, n;
925
926         /* First check that we have enough alignment */
927         if ((*iova_start & ~PAGE_MASK) != (buffer_list[0].addr & ~PAGE_MASK))
928                 return ERR_PTR(-EINVAL);
929
930         mask = 0;
931         total_size = 0;
932         for (i = 0; i < num_phys_buf; ++i) {
933                 if (i != 0)
934                         mask |= buffer_list[i].addr;
935                 if (i != num_phys_buf - 1)
936                         mask |= buffer_list[i].addr + buffer_list[i].size;
937
938                 total_size += buffer_list[i].size;
939         }
940
941         if (mask & ~PAGE_MASK)
942                 return ERR_PTR(-EINVAL);
943
944         /* Find largest page shift we can use to cover buffers */
945         for (shift = PAGE_SHIFT; shift < 31; ++shift)
946                 if (num_phys_buf > 1) {
947                         if ((1ULL << shift) & mask)
948                                 break;
949                 } else {
950                         if (1ULL << shift >=
951                             buffer_list[0].size +
952                             (buffer_list[0].addr & ((1ULL << shift) - 1)))
953                                 break;
954                 }
955
956         buffer_list[0].size += buffer_list[0].addr & ((1ULL << shift) - 1);
957         buffer_list[0].addr &= ~0ull << shift;
958
959         mr = kmalloc(sizeof *mr, GFP_KERNEL);
960         if (!mr)
961                 return ERR_PTR(-ENOMEM);
962
963         npages = 0;
964         for (i = 0; i < num_phys_buf; ++i)
965                 npages += (buffer_list[i].size + (1ULL << shift) - 1) >> shift;
966
967         if (!npages)
968                 return &mr->ibmr;
969
970         page_list = kmalloc(npages * sizeof *page_list, GFP_KERNEL);
971         if (!page_list) {
972                 kfree(mr);
973                 return ERR_PTR(-ENOMEM);
974         }
975
976         n = 0;
977         for (i = 0; i < num_phys_buf; ++i)
978                 for (j = 0;
979                      j < (buffer_list[i].size + (1ULL << shift) - 1) >> shift;
980                      ++j)
981                         page_list[n++] = buffer_list[i].addr + ((u64) j << shift);
982
983         mthca_dbg(to_mdev(pd->device), "Registering memory at %llx (iova %llx) "
984                   "in PD %x; shift %d, npages %d.\n",
985                   (unsigned long long) buffer_list[0].addr,
986                   (unsigned long long) *iova_start,
987                   to_mpd(pd)->pd_num,
988                   shift, npages);
989
990         err = mthca_mr_alloc_phys(to_mdev(pd->device),
991                                   to_mpd(pd)->pd_num,
992                                   page_list, shift, npages,
993                                   *iova_start, total_size,
994                                   convert_access(acc), mr);
995
996         if (err) {
997                 kfree(page_list);
998                 kfree(mr);
999                 return ERR_PTR(err);
1000         }
1001
1002         kfree(page_list);
1003         return &mr->ibmr;
1004 }
1005
1006 static struct ib_mr *mthca_reg_user_mr(struct ib_pd *pd, struct ib_umem *region,
1007                                        int acc, struct ib_udata *udata)
1008 {
1009         struct mthca_dev *dev = to_mdev(pd->device);
1010         struct ib_umem_chunk *chunk;
1011         struct mthca_mr *mr;
1012         u64 *pages;
1013         int shift, n, len;
1014         int i, j, k;
1015         int err = 0;
1016
1017         shift = ffs(region->page_size) - 1;
1018
1019         mr = kmalloc(sizeof *mr, GFP_KERNEL);
1020         if (!mr)
1021                 return ERR_PTR(-ENOMEM);
1022
1023         n = 0;
1024         list_for_each_entry(chunk, &region->chunk_list, list)
1025                 n += chunk->nents;
1026
1027         mr->mtt = mthca_alloc_mtt(dev, n);
1028         if (IS_ERR(mr->mtt)) {
1029                 err = PTR_ERR(mr->mtt);
1030                 goto err;
1031         }
1032
1033         pages = (u64 *) __get_free_page(GFP_KERNEL);
1034         if (!pages) {
1035                 err = -ENOMEM;
1036                 goto err_mtt;
1037         }
1038
1039         i = n = 0;
1040
1041         list_for_each_entry(chunk, &region->chunk_list, list)
1042                 for (j = 0; j < chunk->nmap; ++j) {
1043                         len = sg_dma_len(&chunk->page_list[j]) >> shift;
1044                         for (k = 0; k < len; ++k) {
1045                                 pages[i++] = sg_dma_address(&chunk->page_list[j]) +
1046                                         region->page_size * k;
1047                                 /*
1048                                  * Be friendly to WRITE_MTT command
1049                                  * and leave two empty slots for the
1050                                  * index and reserved fields of the
1051                                  * mailbox.
1052                                  */
1053                                 if (i == PAGE_SIZE / sizeof (u64) - 2) {
1054                                         err = mthca_write_mtt(dev, mr->mtt,
1055                                                               n, pages, i);
1056                                         if (err)
1057                                                 goto mtt_done;
1058                                         n += i;
1059                                         i = 0;
1060                                 }
1061                         }
1062                 }
1063
1064         if (i)
1065                 err = mthca_write_mtt(dev, mr->mtt, n, pages, i);
1066 mtt_done:
1067         free_page((unsigned long) pages);
1068         if (err)
1069                 goto err_mtt;
1070
1071         err = mthca_mr_alloc(dev, to_mpd(pd)->pd_num, shift, region->virt_base,
1072                              region->length, convert_access(acc), mr);
1073
1074         if (err)
1075                 goto err_mtt;
1076
1077         return &mr->ibmr;
1078
1079 err_mtt:
1080         mthca_free_mtt(dev, mr->mtt);
1081
1082 err:
1083         kfree(mr);
1084         return ERR_PTR(err);
1085 }
1086
1087 static int mthca_dereg_mr(struct ib_mr *mr)
1088 {
1089         struct mthca_mr *mmr = to_mmr(mr);
1090         mthca_free_mr(to_mdev(mr->device), mmr);
1091         kfree(mmr);
1092         return 0;
1093 }
1094
1095 static struct ib_fmr *mthca_alloc_fmr(struct ib_pd *pd, int mr_access_flags,
1096                                       struct ib_fmr_attr *fmr_attr)
1097 {
1098         struct mthca_fmr *fmr;
1099         int err;
1100
1101         fmr = kmalloc(sizeof *fmr, GFP_KERNEL);
1102         if (!fmr)
1103                 return ERR_PTR(-ENOMEM);
1104
1105         memcpy(&fmr->attr, fmr_attr, sizeof *fmr_attr);
1106         err = mthca_fmr_alloc(to_mdev(pd->device), to_mpd(pd)->pd_num,
1107                              convert_access(mr_access_flags), fmr);
1108
1109         if (err) {
1110                 kfree(fmr);
1111                 return ERR_PTR(err);
1112         }
1113
1114         return &fmr->ibmr;
1115 }
1116
1117 static int mthca_dealloc_fmr(struct ib_fmr *fmr)
1118 {
1119         struct mthca_fmr *mfmr = to_mfmr(fmr);
1120         int err;
1121
1122         err = mthca_free_fmr(to_mdev(fmr->device), mfmr);
1123         if (err)
1124                 return err;
1125
1126         kfree(mfmr);
1127         return 0;
1128 }
1129
1130 static int mthca_unmap_fmr(struct list_head *fmr_list)
1131 {
1132         struct ib_fmr *fmr;
1133         int err;
1134         u8 status;
1135         struct mthca_dev *mdev = NULL;
1136
1137         list_for_each_entry(fmr, fmr_list, list) {
1138                 if (mdev && to_mdev(fmr->device) != mdev)
1139                         return -EINVAL;
1140                 mdev = to_mdev(fmr->device);
1141         }
1142
1143         if (!mdev)
1144                 return 0;
1145
1146         if (mthca_is_memfree(mdev)) {
1147                 list_for_each_entry(fmr, fmr_list, list)
1148                         mthca_arbel_fmr_unmap(mdev, to_mfmr(fmr));
1149
1150                 wmb();
1151         } else
1152                 list_for_each_entry(fmr, fmr_list, list)
1153                         mthca_tavor_fmr_unmap(mdev, to_mfmr(fmr));
1154
1155         err = mthca_SYNC_TPT(mdev, &status);
1156         if (err)
1157                 return err;
1158         if (status)
1159                 return -EINVAL;
1160         return 0;
1161 }
1162
1163 static ssize_t show_rev(struct class_device *cdev, char *buf)
1164 {
1165         struct mthca_dev *dev = container_of(cdev, struct mthca_dev, ib_dev.class_dev);
1166         return sprintf(buf, "%x\n", dev->rev_id);
1167 }
1168
1169 static ssize_t show_fw_ver(struct class_device *cdev, char *buf)
1170 {
1171         struct mthca_dev *dev = container_of(cdev, struct mthca_dev, ib_dev.class_dev);
1172         return sprintf(buf, "%d.%d.%d\n", (int) (dev->fw_ver >> 32),
1173                        (int) (dev->fw_ver >> 16) & 0xffff,
1174                        (int) dev->fw_ver & 0xffff);
1175 }
1176
1177 static ssize_t show_hca(struct class_device *cdev, char *buf)
1178 {
1179         struct mthca_dev *dev = container_of(cdev, struct mthca_dev, ib_dev.class_dev);
1180         switch (dev->pdev->device) {
1181         case PCI_DEVICE_ID_MELLANOX_TAVOR:
1182                 return sprintf(buf, "MT23108\n");
1183         case PCI_DEVICE_ID_MELLANOX_ARBEL_COMPAT:
1184                 return sprintf(buf, "MT25208 (MT23108 compat mode)\n");
1185         case PCI_DEVICE_ID_MELLANOX_ARBEL:
1186                 return sprintf(buf, "MT25208\n");
1187         case PCI_DEVICE_ID_MELLANOX_SINAI:
1188         case PCI_DEVICE_ID_MELLANOX_SINAI_OLD:
1189                 return sprintf(buf, "MT25204\n");
1190         default:
1191                 return sprintf(buf, "unknown\n");
1192         }
1193 }
1194
1195 static ssize_t show_board(struct class_device *cdev, char *buf)
1196 {
1197         struct mthca_dev *dev = container_of(cdev, struct mthca_dev, ib_dev.class_dev);
1198         return sprintf(buf, "%.*s\n", MTHCA_BOARD_ID_LEN, dev->board_id);
1199 }
1200
1201 static CLASS_DEVICE_ATTR(hw_rev,   S_IRUGO, show_rev,    NULL);
1202 static CLASS_DEVICE_ATTR(fw_ver,   S_IRUGO, show_fw_ver, NULL);
1203 static CLASS_DEVICE_ATTR(hca_type, S_IRUGO, show_hca,    NULL);
1204 static CLASS_DEVICE_ATTR(board_id, S_IRUGO, show_board,  NULL);
1205
1206 static struct class_device_attribute *mthca_class_attributes[] = {
1207         &class_device_attr_hw_rev,
1208         &class_device_attr_fw_ver,
1209         &class_device_attr_hca_type,
1210         &class_device_attr_board_id
1211 };
1212
1213 static int mthca_init_node_data(struct mthca_dev *dev)
1214 {
1215         struct ib_smp *in_mad  = NULL;
1216         struct ib_smp *out_mad = NULL;
1217         int err = -ENOMEM;
1218         u8 status;
1219
1220         in_mad  = kzalloc(sizeof *in_mad, GFP_KERNEL);
1221         out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
1222         if (!in_mad || !out_mad)
1223                 goto out;
1224
1225         init_query_mad(in_mad);
1226         in_mad->attr_id = IB_SMP_ATTR_NODE_DESC;
1227
1228         err = mthca_MAD_IFC(dev, 1, 1,
1229                             1, NULL, NULL, in_mad, out_mad,
1230                             &status);
1231         if (err)
1232                 goto out;
1233         if (status) {
1234                 err = -EINVAL;
1235                 goto out;
1236         }
1237
1238         memcpy(dev->ib_dev.node_desc, out_mad->data, 64);
1239
1240         in_mad->attr_id = IB_SMP_ATTR_NODE_INFO;
1241
1242         err = mthca_MAD_IFC(dev, 1, 1,
1243                             1, NULL, NULL, in_mad, out_mad,
1244                             &status);
1245         if (err)
1246                 goto out;
1247         if (status) {
1248                 err = -EINVAL;
1249                 goto out;
1250         }
1251
1252         memcpy(&dev->ib_dev.node_guid, out_mad->data + 12, 8);
1253
1254 out:
1255         kfree(in_mad);
1256         kfree(out_mad);
1257         return err;
1258 }
1259
1260 int mthca_register_device(struct mthca_dev *dev)
1261 {
1262         int ret;
1263         int i;
1264
1265         ret = mthca_init_node_data(dev);
1266         if (ret)
1267                 return ret;
1268
1269         strlcpy(dev->ib_dev.name, "mthca%d", IB_DEVICE_NAME_MAX);
1270         dev->ib_dev.owner                = THIS_MODULE;
1271
1272         dev->ib_dev.uverbs_abi_ver       = MTHCA_UVERBS_ABI_VERSION;
1273         dev->ib_dev.uverbs_cmd_mask      =
1274                 (1ull << IB_USER_VERBS_CMD_GET_CONTEXT)         |
1275                 (1ull << IB_USER_VERBS_CMD_QUERY_DEVICE)        |
1276                 (1ull << IB_USER_VERBS_CMD_QUERY_PORT)          |
1277                 (1ull << IB_USER_VERBS_CMD_ALLOC_PD)            |
1278                 (1ull << IB_USER_VERBS_CMD_DEALLOC_PD)          |
1279                 (1ull << IB_USER_VERBS_CMD_REG_MR)              |
1280                 (1ull << IB_USER_VERBS_CMD_DEREG_MR)            |
1281                 (1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL) |
1282                 (1ull << IB_USER_VERBS_CMD_CREATE_CQ)           |
1283                 (1ull << IB_USER_VERBS_CMD_RESIZE_CQ)           |
1284                 (1ull << IB_USER_VERBS_CMD_DESTROY_CQ)          |
1285                 (1ull << IB_USER_VERBS_CMD_CREATE_QP)           |
1286                 (1ull << IB_USER_VERBS_CMD_QUERY_QP)            |
1287                 (1ull << IB_USER_VERBS_CMD_MODIFY_QP)           |
1288                 (1ull << IB_USER_VERBS_CMD_DESTROY_QP)          |
1289                 (1ull << IB_USER_VERBS_CMD_ATTACH_MCAST)        |
1290                 (1ull << IB_USER_VERBS_CMD_DETACH_MCAST);
1291         dev->ib_dev.node_type            = RDMA_NODE_IB_CA;
1292         dev->ib_dev.phys_port_cnt        = dev->limits.num_ports;
1293         dev->ib_dev.dma_device           = &dev->pdev->dev;
1294         dev->ib_dev.class_dev.dev        = &dev->pdev->dev;
1295         dev->ib_dev.query_device         = mthca_query_device;
1296         dev->ib_dev.query_port           = mthca_query_port;
1297         dev->ib_dev.modify_device        = mthca_modify_device;
1298         dev->ib_dev.modify_port          = mthca_modify_port;
1299         dev->ib_dev.query_pkey           = mthca_query_pkey;
1300         dev->ib_dev.query_gid            = mthca_query_gid;
1301         dev->ib_dev.alloc_ucontext       = mthca_alloc_ucontext;
1302         dev->ib_dev.dealloc_ucontext     = mthca_dealloc_ucontext;
1303         dev->ib_dev.mmap                 = mthca_mmap_uar;
1304         dev->ib_dev.alloc_pd             = mthca_alloc_pd;
1305         dev->ib_dev.dealloc_pd           = mthca_dealloc_pd;
1306         dev->ib_dev.create_ah            = mthca_ah_create;
1307         dev->ib_dev.query_ah             = mthca_ah_query;
1308         dev->ib_dev.destroy_ah           = mthca_ah_destroy;
1309
1310         if (dev->mthca_flags & MTHCA_FLAG_SRQ) {
1311                 dev->ib_dev.create_srq           = mthca_create_srq;
1312                 dev->ib_dev.modify_srq           = mthca_modify_srq;
1313                 dev->ib_dev.query_srq            = mthca_query_srq;
1314                 dev->ib_dev.destroy_srq          = mthca_destroy_srq;
1315                 dev->ib_dev.uverbs_cmd_mask     |=
1316                         (1ull << IB_USER_VERBS_CMD_CREATE_SRQ)          |
1317                         (1ull << IB_USER_VERBS_CMD_MODIFY_SRQ)          |
1318                         (1ull << IB_USER_VERBS_CMD_QUERY_SRQ)           |
1319                         (1ull << IB_USER_VERBS_CMD_DESTROY_SRQ);
1320
1321                 if (mthca_is_memfree(dev))
1322                         dev->ib_dev.post_srq_recv = mthca_arbel_post_srq_recv;
1323                 else
1324                         dev->ib_dev.post_srq_recv = mthca_tavor_post_srq_recv;
1325         }
1326
1327         dev->ib_dev.create_qp            = mthca_create_qp;
1328         dev->ib_dev.modify_qp            = mthca_modify_qp;
1329         dev->ib_dev.query_qp             = mthca_query_qp;
1330         dev->ib_dev.destroy_qp           = mthca_destroy_qp;
1331         dev->ib_dev.create_cq            = mthca_create_cq;
1332         dev->ib_dev.resize_cq            = mthca_resize_cq;
1333         dev->ib_dev.destroy_cq           = mthca_destroy_cq;
1334         dev->ib_dev.poll_cq              = mthca_poll_cq;
1335         dev->ib_dev.get_dma_mr           = mthca_get_dma_mr;
1336         dev->ib_dev.reg_phys_mr          = mthca_reg_phys_mr;
1337         dev->ib_dev.reg_user_mr          = mthca_reg_user_mr;
1338         dev->ib_dev.dereg_mr             = mthca_dereg_mr;
1339
1340         if (dev->mthca_flags & MTHCA_FLAG_FMR) {
1341                 dev->ib_dev.alloc_fmr            = mthca_alloc_fmr;
1342                 dev->ib_dev.unmap_fmr            = mthca_unmap_fmr;
1343                 dev->ib_dev.dealloc_fmr          = mthca_dealloc_fmr;
1344                 if (mthca_is_memfree(dev))
1345                         dev->ib_dev.map_phys_fmr = mthca_arbel_map_phys_fmr;
1346                 else
1347                         dev->ib_dev.map_phys_fmr = mthca_tavor_map_phys_fmr;
1348         }
1349
1350         dev->ib_dev.attach_mcast         = mthca_multicast_attach;
1351         dev->ib_dev.detach_mcast         = mthca_multicast_detach;
1352         dev->ib_dev.process_mad          = mthca_process_mad;
1353
1354         if (mthca_is_memfree(dev)) {
1355                 dev->ib_dev.req_notify_cq = mthca_arbel_arm_cq;
1356                 dev->ib_dev.post_send     = mthca_arbel_post_send;
1357                 dev->ib_dev.post_recv     = mthca_arbel_post_receive;
1358         } else {
1359                 dev->ib_dev.req_notify_cq = mthca_tavor_arm_cq;
1360                 dev->ib_dev.post_send     = mthca_tavor_post_send;
1361                 dev->ib_dev.post_recv     = mthca_tavor_post_receive;
1362         }
1363
1364         mutex_init(&dev->cap_mask_mutex);
1365
1366         ret = ib_register_device(&dev->ib_dev);
1367         if (ret)
1368                 return ret;
1369
1370         for (i = 0; i < ARRAY_SIZE(mthca_class_attributes); ++i) {
1371                 ret = class_device_create_file(&dev->ib_dev.class_dev,
1372                                                mthca_class_attributes[i]);
1373                 if (ret) {
1374                         ib_unregister_device(&dev->ib_dev);
1375                         return ret;
1376                 }
1377         }
1378
1379         mthca_start_catas_poll(dev);
1380
1381         return 0;
1382 }
1383
1384 void mthca_unregister_device(struct mthca_dev *dev)
1385 {
1386         mthca_stop_catas_poll(dev);
1387         ib_unregister_device(&dev->ib_dev);
1388 }