RDMA/core: Add SRQ type field
[linux-3.10.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
37 #include <rdma/ib_smi.h>
38 #include <rdma/ib_umem.h>
39 #include <rdma/ib_user_verbs.h>
40
41 #include <linux/sched.h>
42 #include <linux/slab.h>
43 #include <linux/mm.h>
44
45 #include "mthca_dev.h"
46 #include "mthca_cmd.h"
47 #include "mthca_user.h"
48 #include "mthca_memfree.h"
49
50 static void init_query_mad(struct ib_smp *mad)
51 {
52         mad->base_version  = 1;
53         mad->mgmt_class    = IB_MGMT_CLASS_SUBN_LID_ROUTED;
54         mad->class_version = 1;
55         mad->method        = IB_MGMT_METHOD_GET;
56 }
57
58 static int mthca_query_device(struct ib_device *ibdev,
59                               struct ib_device_attr *props)
60 {
61         struct ib_smp *in_mad  = NULL;
62         struct ib_smp *out_mad = NULL;
63         int err = -ENOMEM;
64         struct mthca_dev *mdev = to_mdev(ibdev);
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         if (err)
81                 goto out;
82
83         props->device_cap_flags    = mdev->device_cap_flags;
84         props->vendor_id           = be32_to_cpup((__be32 *) (out_mad->data + 36)) &
85                 0xffffff;
86         props->vendor_part_id      = be16_to_cpup((__be16 *) (out_mad->data + 30));
87         props->hw_ver              = be32_to_cpup((__be32 *) (out_mad->data + 32));
88         memcpy(&props->sys_image_guid, out_mad->data +  4, 8);
89
90         props->max_mr_size         = ~0ull;
91         props->page_size_cap       = mdev->limits.page_size_cap;
92         props->max_qp              = mdev->limits.num_qps - mdev->limits.reserved_qps;
93         props->max_qp_wr           = mdev->limits.max_wqes;
94         props->max_sge             = mdev->limits.max_sg;
95         props->max_cq              = mdev->limits.num_cqs - mdev->limits.reserved_cqs;
96         props->max_cqe             = mdev->limits.max_cqes;
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 = mdev->limits.max_qp_init_rdma;
101         props->max_res_rd_atom     = props->max_qp_rd_atom * props->max_qp;
102         props->max_srq             = mdev->limits.num_srqs - mdev->limits.reserved_srqs;
103         props->max_srq_wr          = mdev->limits.max_srq_wqes;
104         props->max_srq_sge         = mdev->limits.max_srq_sge;
105         props->local_ca_ack_delay  = mdev->limits.local_ca_ack_delay;
106         props->atomic_cap          = mdev->limits.flags & DEV_LIM_FLAG_ATOMIC ?
107                                         IB_ATOMIC_HCA : IB_ATOMIC_NONE;
108         props->max_pkeys           = mdev->limits.pkey_table_len;
109         props->max_mcast_grp       = mdev->limits.num_mgms + mdev->limits.num_amgms;
110         props->max_mcast_qp_attach = MTHCA_QP_PER_MGM;
111         props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
112                                            props->max_mcast_grp;
113         /*
114          * If Sinai memory key optimization is being used, then only
115          * the 8-bit key portion will change.  For other HCAs, the
116          * unused index bits will also be used for FMR remapping.
117          */
118         if (mdev->mthca_flags & MTHCA_FLAG_SINAI_OPT)
119                 props->max_map_per_fmr = 255;
120         else
121                 props->max_map_per_fmr =
122                         (1 << (32 - ilog2(mdev->limits.num_mpts))) - 1;
123
124         err = 0;
125  out:
126         kfree(in_mad);
127         kfree(out_mad);
128         return err;
129 }
130
131 static int mthca_query_port(struct ib_device *ibdev,
132                             u8 port, struct ib_port_attr *props)
133 {
134         struct ib_smp *in_mad  = NULL;
135         struct ib_smp *out_mad = NULL;
136         int err = -ENOMEM;
137
138         in_mad  = kzalloc(sizeof *in_mad, GFP_KERNEL);
139         out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
140         if (!in_mad || !out_mad)
141                 goto out;
142
143         memset(props, 0, sizeof *props);
144
145         init_query_mad(in_mad);
146         in_mad->attr_id  = IB_SMP_ATTR_PORT_INFO;
147         in_mad->attr_mod = cpu_to_be32(port);
148
149         err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
150                             port, NULL, NULL, in_mad, out_mad);
151         if (err)
152                 goto out;
153
154         props->lid               = be16_to_cpup((__be16 *) (out_mad->data + 16));
155         props->lmc               = out_mad->data[34] & 0x7;
156         props->sm_lid            = be16_to_cpup((__be16 *) (out_mad->data + 18));
157         props->sm_sl             = out_mad->data[36] & 0xf;
158         props->state             = out_mad->data[32] & 0xf;
159         props->phys_state        = out_mad->data[33] >> 4;
160         props->port_cap_flags    = be32_to_cpup((__be32 *) (out_mad->data + 20));
161         props->gid_tbl_len       = to_mdev(ibdev)->limits.gid_table_len;
162         props->max_msg_sz        = 0x80000000;
163         props->pkey_tbl_len      = to_mdev(ibdev)->limits.pkey_table_len;
164         props->bad_pkey_cntr     = be16_to_cpup((__be16 *) (out_mad->data + 46));
165         props->qkey_viol_cntr    = be16_to_cpup((__be16 *) (out_mad->data + 48));
166         props->active_width      = out_mad->data[31] & 0xf;
167         props->active_speed      = out_mad->data[35] >> 4;
168         props->max_mtu           = out_mad->data[41] & 0xf;
169         props->active_mtu        = out_mad->data[36] >> 4;
170         props->subnet_timeout    = out_mad->data[51] & 0x1f;
171         props->max_vl_num        = out_mad->data[37] >> 4;
172         props->init_type_reply   = out_mad->data[41] >> 4;
173
174  out:
175         kfree(in_mad);
176         kfree(out_mad);
177         return err;
178 }
179
180 static int mthca_modify_device(struct ib_device *ibdev,
181                                int mask,
182                                struct ib_device_modify *props)
183 {
184         if (mask & ~IB_DEVICE_MODIFY_NODE_DESC)
185                 return -EOPNOTSUPP;
186
187         if (mask & IB_DEVICE_MODIFY_NODE_DESC) {
188                 if (mutex_lock_interruptible(&to_mdev(ibdev)->cap_mask_mutex))
189                         return -ERESTARTSYS;
190                 memcpy(ibdev->node_desc, props->node_desc, 64);
191                 mutex_unlock(&to_mdev(ibdev)->cap_mask_mutex);
192         }
193
194         return 0;
195 }
196
197 static int mthca_modify_port(struct ib_device *ibdev,
198                              u8 port, int port_modify_mask,
199                              struct ib_port_modify *props)
200 {
201         struct mthca_set_ib_param set_ib;
202         struct ib_port_attr attr;
203         int err;
204
205         if (mutex_lock_interruptible(&to_mdev(ibdev)->cap_mask_mutex))
206                 return -ERESTARTSYS;
207
208         err = mthca_query_port(ibdev, port, &attr);
209         if (err)
210                 goto out;
211
212         set_ib.set_si_guid     = 0;
213         set_ib.reset_qkey_viol = !!(port_modify_mask & IB_PORT_RESET_QKEY_CNTR);
214
215         set_ib.cap_mask = (attr.port_cap_flags | props->set_port_cap_mask) &
216                 ~props->clr_port_cap_mask;
217
218         err = mthca_SET_IB(to_mdev(ibdev), &set_ib, port);
219         if (err)
220                 goto out;
221 out:
222         mutex_unlock(&to_mdev(ibdev)->cap_mask_mutex);
223         return err;
224 }
225
226 static int mthca_query_pkey(struct ib_device *ibdev,
227                             u8 port, u16 index, u16 *pkey)
228 {
229         struct ib_smp *in_mad  = NULL;
230         struct ib_smp *out_mad = NULL;
231         int err = -ENOMEM;
232
233         in_mad  = kzalloc(sizeof *in_mad, GFP_KERNEL);
234         out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
235         if (!in_mad || !out_mad)
236                 goto out;
237
238         init_query_mad(in_mad);
239         in_mad->attr_id  = IB_SMP_ATTR_PKEY_TABLE;
240         in_mad->attr_mod = cpu_to_be32(index / 32);
241
242         err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
243                             port, NULL, NULL, in_mad, out_mad);
244         if (err)
245                 goto out;
246
247         *pkey = be16_to_cpu(((__be16 *) out_mad->data)[index % 32]);
248
249  out:
250         kfree(in_mad);
251         kfree(out_mad);
252         return err;
253 }
254
255 static int mthca_query_gid(struct ib_device *ibdev, u8 port,
256                            int index, union ib_gid *gid)
257 {
258         struct ib_smp *in_mad  = NULL;
259         struct ib_smp *out_mad = NULL;
260         int err = -ENOMEM;
261
262         in_mad  = kzalloc(sizeof *in_mad, GFP_KERNEL);
263         out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
264         if (!in_mad || !out_mad)
265                 goto out;
266
267         init_query_mad(in_mad);
268         in_mad->attr_id  = IB_SMP_ATTR_PORT_INFO;
269         in_mad->attr_mod = cpu_to_be32(port);
270
271         err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
272                             port, NULL, NULL, in_mad, out_mad);
273         if (err)
274                 goto out;
275
276         memcpy(gid->raw, out_mad->data + 8, 8);
277
278         init_query_mad(in_mad);
279         in_mad->attr_id  = IB_SMP_ATTR_GUID_INFO;
280         in_mad->attr_mod = cpu_to_be32(index / 8);
281
282         err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
283                             port, NULL, NULL, in_mad, out_mad);
284         if (err)
285                 goto out;
286
287         memcpy(gid->raw + 8, out_mad->data + (index % 8) * 8, 8);
288
289  out:
290         kfree(in_mad);
291         kfree(out_mad);
292         return err;
293 }
294
295 static struct ib_ucontext *mthca_alloc_ucontext(struct ib_device *ibdev,
296                                                 struct ib_udata *udata)
297 {
298         struct mthca_alloc_ucontext_resp uresp;
299         struct mthca_ucontext           *context;
300         int                              err;
301
302         if (!(to_mdev(ibdev)->active))
303                 return ERR_PTR(-EAGAIN);
304
305         memset(&uresp, 0, sizeof uresp);
306
307         uresp.qp_tab_size = to_mdev(ibdev)->limits.num_qps;
308         if (mthca_is_memfree(to_mdev(ibdev)))
309                 uresp.uarc_size = to_mdev(ibdev)->uar_table.uarc_size;
310         else
311                 uresp.uarc_size = 0;
312
313         context = kmalloc(sizeof *context, GFP_KERNEL);
314         if (!context)
315                 return ERR_PTR(-ENOMEM);
316
317         err = mthca_uar_alloc(to_mdev(ibdev), &context->uar);
318         if (err) {
319                 kfree(context);
320                 return ERR_PTR(err);
321         }
322
323         context->db_tab = mthca_init_user_db_tab(to_mdev(ibdev));
324         if (IS_ERR(context->db_tab)) {
325                 err = PTR_ERR(context->db_tab);
326                 mthca_uar_free(to_mdev(ibdev), &context->uar);
327                 kfree(context);
328                 return ERR_PTR(err);
329         }
330
331         if (ib_copy_to_udata(udata, &uresp, sizeof uresp)) {
332                 mthca_cleanup_user_db_tab(to_mdev(ibdev), &context->uar, context->db_tab);
333                 mthca_uar_free(to_mdev(ibdev), &context->uar);
334                 kfree(context);
335                 return ERR_PTR(-EFAULT);
336         }
337
338         context->reg_mr_warned = 0;
339
340         return &context->ibucontext;
341 }
342
343 static int mthca_dealloc_ucontext(struct ib_ucontext *context)
344 {
345         mthca_cleanup_user_db_tab(to_mdev(context->device), &to_mucontext(context)->uar,
346                                   to_mucontext(context)->db_tab);
347         mthca_uar_free(to_mdev(context->device), &to_mucontext(context)->uar);
348         kfree(to_mucontext(context));
349
350         return 0;
351 }
352
353 static int mthca_mmap_uar(struct ib_ucontext *context,
354                           struct vm_area_struct *vma)
355 {
356         if (vma->vm_end - vma->vm_start != PAGE_SIZE)
357                 return -EINVAL;
358
359         vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
360
361         if (io_remap_pfn_range(vma, vma->vm_start,
362                                to_mucontext(context)->uar.pfn,
363                                PAGE_SIZE, vma->vm_page_prot))
364                 return -EAGAIN;
365
366         return 0;
367 }
368
369 static struct ib_pd *mthca_alloc_pd(struct ib_device *ibdev,
370                                     struct ib_ucontext *context,
371                                     struct ib_udata *udata)
372 {
373         struct mthca_pd *pd;
374         int err;
375
376         pd = kmalloc(sizeof *pd, GFP_KERNEL);
377         if (!pd)
378                 return ERR_PTR(-ENOMEM);
379
380         err = mthca_pd_alloc(to_mdev(ibdev), !context, pd);
381         if (err) {
382                 kfree(pd);
383                 return ERR_PTR(err);
384         }
385
386         if (context) {
387                 if (ib_copy_to_udata(udata, &pd->pd_num, sizeof (__u32))) {
388                         mthca_pd_free(to_mdev(ibdev), pd);
389                         kfree(pd);
390                         return ERR_PTR(-EFAULT);
391                 }
392         }
393
394         return &pd->ibpd;
395 }
396
397 static int mthca_dealloc_pd(struct ib_pd *pd)
398 {
399         mthca_pd_free(to_mdev(pd->device), to_mpd(pd));
400         kfree(pd);
401
402         return 0;
403 }
404
405 static struct ib_ah *mthca_ah_create(struct ib_pd *pd,
406                                      struct ib_ah_attr *ah_attr)
407 {
408         int err;
409         struct mthca_ah *ah;
410
411         ah = kmalloc(sizeof *ah, GFP_ATOMIC);
412         if (!ah)
413                 return ERR_PTR(-ENOMEM);
414
415         err = mthca_create_ah(to_mdev(pd->device), to_mpd(pd), ah_attr, ah);
416         if (err) {
417                 kfree(ah);
418                 return ERR_PTR(err);
419         }
420
421         return &ah->ibah;
422 }
423
424 static int mthca_ah_destroy(struct ib_ah *ah)
425 {
426         mthca_destroy_ah(to_mdev(ah->device), to_mah(ah));
427         kfree(ah);
428
429         return 0;
430 }
431
432 static struct ib_srq *mthca_create_srq(struct ib_pd *pd,
433                                        struct ib_srq_init_attr *init_attr,
434                                        struct ib_udata *udata)
435 {
436         struct mthca_create_srq ucmd;
437         struct mthca_ucontext *context = NULL;
438         struct mthca_srq *srq;
439         int err;
440
441         if (init_attr->srq_type != IB_SRQT_BASIC)
442                 return ERR_PTR(-ENOSYS);
443
444         srq = kmalloc(sizeof *srq, GFP_KERNEL);
445         if (!srq)
446                 return ERR_PTR(-ENOMEM);
447
448         if (pd->uobject) {
449                 context = to_mucontext(pd->uobject->context);
450
451                 if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
452                         err = -EFAULT;
453                         goto err_free;
454                 }
455
456                 err = mthca_map_user_db(to_mdev(pd->device), &context->uar,
457                                         context->db_tab, ucmd.db_index,
458                                         ucmd.db_page);
459
460                 if (err)
461                         goto err_free;
462
463                 srq->mr.ibmr.lkey = ucmd.lkey;
464                 srq->db_index     = ucmd.db_index;
465         }
466
467         err = mthca_alloc_srq(to_mdev(pd->device), to_mpd(pd),
468                               &init_attr->attr, srq);
469
470         if (err && pd->uobject)
471                 mthca_unmap_user_db(to_mdev(pd->device), &context->uar,
472                                     context->db_tab, ucmd.db_index);
473
474         if (err)
475                 goto err_free;
476
477         if (context && ib_copy_to_udata(udata, &srq->srqn, sizeof (__u32))) {
478                 mthca_free_srq(to_mdev(pd->device), srq);
479                 err = -EFAULT;
480                 goto err_free;
481         }
482
483         return &srq->ibsrq;
484
485 err_free:
486         kfree(srq);
487
488         return ERR_PTR(err);
489 }
490
491 static int mthca_destroy_srq(struct ib_srq *srq)
492 {
493         struct mthca_ucontext *context;
494
495         if (srq->uobject) {
496                 context = to_mucontext(srq->uobject->context);
497
498                 mthca_unmap_user_db(to_mdev(srq->device), &context->uar,
499                                     context->db_tab, to_msrq(srq)->db_index);
500         }
501
502         mthca_free_srq(to_mdev(srq->device), to_msrq(srq));
503         kfree(srq);
504
505         return 0;
506 }
507
508 static struct ib_qp *mthca_create_qp(struct ib_pd *pd,
509                                      struct ib_qp_init_attr *init_attr,
510                                      struct ib_udata *udata)
511 {
512         struct mthca_create_qp ucmd;
513         struct mthca_qp *qp;
514         int err;
515
516         if (init_attr->create_flags)
517                 return ERR_PTR(-EINVAL);
518
519         switch (init_attr->qp_type) {
520         case IB_QPT_RC:
521         case IB_QPT_UC:
522         case IB_QPT_UD:
523         {
524                 struct mthca_ucontext *context;
525
526                 qp = kmalloc(sizeof *qp, GFP_KERNEL);
527                 if (!qp)
528                         return ERR_PTR(-ENOMEM);
529
530                 if (pd->uobject) {
531                         context = to_mucontext(pd->uobject->context);
532
533                         if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
534                                 kfree(qp);
535                                 return ERR_PTR(-EFAULT);
536                         }
537
538                         err = mthca_map_user_db(to_mdev(pd->device), &context->uar,
539                                                 context->db_tab,
540                                                 ucmd.sq_db_index, ucmd.sq_db_page);
541                         if (err) {
542                                 kfree(qp);
543                                 return ERR_PTR(err);
544                         }
545
546                         err = mthca_map_user_db(to_mdev(pd->device), &context->uar,
547                                                 context->db_tab,
548                                                 ucmd.rq_db_index, ucmd.rq_db_page);
549                         if (err) {
550                                 mthca_unmap_user_db(to_mdev(pd->device),
551                                                     &context->uar,
552                                                     context->db_tab,
553                                                     ucmd.sq_db_index);
554                                 kfree(qp);
555                                 return ERR_PTR(err);
556                         }
557
558                         qp->mr.ibmr.lkey = ucmd.lkey;
559                         qp->sq.db_index  = ucmd.sq_db_index;
560                         qp->rq.db_index  = ucmd.rq_db_index;
561                 }
562
563                 err = mthca_alloc_qp(to_mdev(pd->device), to_mpd(pd),
564                                      to_mcq(init_attr->send_cq),
565                                      to_mcq(init_attr->recv_cq),
566                                      init_attr->qp_type, init_attr->sq_sig_type,
567                                      &init_attr->cap, qp);
568
569                 if (err && pd->uobject) {
570                         context = to_mucontext(pd->uobject->context);
571
572                         mthca_unmap_user_db(to_mdev(pd->device),
573                                             &context->uar,
574                                             context->db_tab,
575                                             ucmd.sq_db_index);
576                         mthca_unmap_user_db(to_mdev(pd->device),
577                                             &context->uar,
578                                             context->db_tab,
579                                             ucmd.rq_db_index);
580                 }
581
582                 qp->ibqp.qp_num = qp->qpn;
583                 break;
584         }
585         case IB_QPT_SMI:
586         case IB_QPT_GSI:
587         {
588                 /* Don't allow userspace to create special QPs */
589                 if (pd->uobject)
590                         return ERR_PTR(-EINVAL);
591
592                 qp = kmalloc(sizeof (struct mthca_sqp), GFP_KERNEL);
593                 if (!qp)
594                         return ERR_PTR(-ENOMEM);
595
596                 qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 : 1;
597
598                 err = mthca_alloc_sqp(to_mdev(pd->device), to_mpd(pd),
599                                       to_mcq(init_attr->send_cq),
600                                       to_mcq(init_attr->recv_cq),
601                                       init_attr->sq_sig_type, &init_attr->cap,
602                                       qp->ibqp.qp_num, init_attr->port_num,
603                                       to_msqp(qp));
604                 break;
605         }
606         default:
607                 /* Don't support raw QPs */
608                 return ERR_PTR(-ENOSYS);
609         }
610
611         if (err) {
612                 kfree(qp);
613                 return ERR_PTR(err);
614         }
615
616         init_attr->cap.max_send_wr     = qp->sq.max;
617         init_attr->cap.max_recv_wr     = qp->rq.max;
618         init_attr->cap.max_send_sge    = qp->sq.max_gs;
619         init_attr->cap.max_recv_sge    = qp->rq.max_gs;
620         init_attr->cap.max_inline_data = qp->max_inline_data;
621
622         return &qp->ibqp;
623 }
624
625 static int mthca_destroy_qp(struct ib_qp *qp)
626 {
627         if (qp->uobject) {
628                 mthca_unmap_user_db(to_mdev(qp->device),
629                                     &to_mucontext(qp->uobject->context)->uar,
630                                     to_mucontext(qp->uobject->context)->db_tab,
631                                     to_mqp(qp)->sq.db_index);
632                 mthca_unmap_user_db(to_mdev(qp->device),
633                                     &to_mucontext(qp->uobject->context)->uar,
634                                     to_mucontext(qp->uobject->context)->db_tab,
635                                     to_mqp(qp)->rq.db_index);
636         }
637         mthca_free_qp(to_mdev(qp->device), to_mqp(qp));
638         kfree(qp);
639         return 0;
640 }
641
642 static struct ib_cq *mthca_create_cq(struct ib_device *ibdev, int entries,
643                                      int comp_vector,
644                                      struct ib_ucontext *context,
645                                      struct ib_udata *udata)
646 {
647         struct mthca_create_cq ucmd;
648         struct mthca_cq *cq;
649         int nent;
650         int err;
651
652         if (entries < 1 || entries > to_mdev(ibdev)->limits.max_cqes)
653                 return ERR_PTR(-EINVAL);
654
655         if (context) {
656                 if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd))
657                         return ERR_PTR(-EFAULT);
658
659                 err = mthca_map_user_db(to_mdev(ibdev), &to_mucontext(context)->uar,
660                                         to_mucontext(context)->db_tab,
661                                         ucmd.set_db_index, ucmd.set_db_page);
662                 if (err)
663                         return ERR_PTR(err);
664
665                 err = mthca_map_user_db(to_mdev(ibdev), &to_mucontext(context)->uar,
666                                         to_mucontext(context)->db_tab,
667                                         ucmd.arm_db_index, ucmd.arm_db_page);
668                 if (err)
669                         goto err_unmap_set;
670         }
671
672         cq = kmalloc(sizeof *cq, GFP_KERNEL);
673         if (!cq) {
674                 err = -ENOMEM;
675                 goto err_unmap_arm;
676         }
677
678         if (context) {
679                 cq->buf.mr.ibmr.lkey = ucmd.lkey;
680                 cq->set_ci_db_index  = ucmd.set_db_index;
681                 cq->arm_db_index     = ucmd.arm_db_index;
682         }
683
684         for (nent = 1; nent <= entries; nent <<= 1)
685                 ; /* nothing */
686
687         err = mthca_init_cq(to_mdev(ibdev), nent,
688                             context ? to_mucontext(context) : NULL,
689                             context ? ucmd.pdn : to_mdev(ibdev)->driver_pd.pd_num,
690                             cq);
691         if (err)
692                 goto err_free;
693
694         if (context && ib_copy_to_udata(udata, &cq->cqn, sizeof (__u32))) {
695                 mthca_free_cq(to_mdev(ibdev), cq);
696                 goto err_free;
697         }
698
699         cq->resize_buf = NULL;
700
701         return &cq->ibcq;
702
703 err_free:
704         kfree(cq);
705
706 err_unmap_arm:
707         if (context)
708                 mthca_unmap_user_db(to_mdev(ibdev), &to_mucontext(context)->uar,
709                                     to_mucontext(context)->db_tab, ucmd.arm_db_index);
710
711 err_unmap_set:
712         if (context)
713                 mthca_unmap_user_db(to_mdev(ibdev), &to_mucontext(context)->uar,
714                                     to_mucontext(context)->db_tab, ucmd.set_db_index);
715
716         return ERR_PTR(err);
717 }
718
719 static int mthca_alloc_resize_buf(struct mthca_dev *dev, struct mthca_cq *cq,
720                                   int entries)
721 {
722         int ret;
723
724         spin_lock_irq(&cq->lock);
725         if (cq->resize_buf) {
726                 ret = -EBUSY;
727                 goto unlock;
728         }
729
730         cq->resize_buf = kmalloc(sizeof *cq->resize_buf, GFP_ATOMIC);
731         if (!cq->resize_buf) {
732                 ret = -ENOMEM;
733                 goto unlock;
734         }
735
736         cq->resize_buf->state = CQ_RESIZE_ALLOC;
737
738         ret = 0;
739
740 unlock:
741         spin_unlock_irq(&cq->lock);
742
743         if (ret)
744                 return ret;
745
746         ret = mthca_alloc_cq_buf(dev, &cq->resize_buf->buf, entries);
747         if (ret) {
748                 spin_lock_irq(&cq->lock);
749                 kfree(cq->resize_buf);
750                 cq->resize_buf = NULL;
751                 spin_unlock_irq(&cq->lock);
752                 return ret;
753         }
754
755         cq->resize_buf->cqe = entries - 1;
756
757         spin_lock_irq(&cq->lock);
758         cq->resize_buf->state = CQ_RESIZE_READY;
759         spin_unlock_irq(&cq->lock);
760
761         return 0;
762 }
763
764 static int mthca_resize_cq(struct ib_cq *ibcq, int entries, struct ib_udata *udata)
765 {
766         struct mthca_dev *dev = to_mdev(ibcq->device);
767         struct mthca_cq *cq = to_mcq(ibcq);
768         struct mthca_resize_cq ucmd;
769         u32 lkey;
770         int ret;
771
772         if (entries < 1 || entries > dev->limits.max_cqes)
773                 return -EINVAL;
774
775         mutex_lock(&cq->mutex);
776
777         entries = roundup_pow_of_two(entries + 1);
778         if (entries == ibcq->cqe + 1) {
779                 ret = 0;
780                 goto out;
781         }
782
783         if (cq->is_kernel) {
784                 ret = mthca_alloc_resize_buf(dev, cq, entries);
785                 if (ret)
786                         goto out;
787                 lkey = cq->resize_buf->buf.mr.ibmr.lkey;
788         } else {
789                 if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
790                         ret = -EFAULT;
791                         goto out;
792                 }
793                 lkey = ucmd.lkey;
794         }
795
796         ret = mthca_RESIZE_CQ(dev, cq->cqn, lkey, ilog2(entries));
797
798         if (ret) {
799                 if (cq->resize_buf) {
800                         mthca_free_cq_buf(dev, &cq->resize_buf->buf,
801                                           cq->resize_buf->cqe);
802                         kfree(cq->resize_buf);
803                         spin_lock_irq(&cq->lock);
804                         cq->resize_buf = NULL;
805                         spin_unlock_irq(&cq->lock);
806                 }
807                 goto out;
808         }
809
810         if (cq->is_kernel) {
811                 struct mthca_cq_buf tbuf;
812                 int tcqe;
813
814                 spin_lock_irq(&cq->lock);
815                 if (cq->resize_buf->state == CQ_RESIZE_READY) {
816                         mthca_cq_resize_copy_cqes(cq);
817                         tbuf         = cq->buf;
818                         tcqe         = cq->ibcq.cqe;
819                         cq->buf      = cq->resize_buf->buf;
820                         cq->ibcq.cqe = cq->resize_buf->cqe;
821                 } else {
822                         tbuf = cq->resize_buf->buf;
823                         tcqe = cq->resize_buf->cqe;
824                 }
825
826                 kfree(cq->resize_buf);
827                 cq->resize_buf = NULL;
828                 spin_unlock_irq(&cq->lock);
829
830                 mthca_free_cq_buf(dev, &tbuf, tcqe);
831         } else
832                 ibcq->cqe = entries - 1;
833
834 out:
835         mutex_unlock(&cq->mutex);
836
837         return ret;
838 }
839
840 static int mthca_destroy_cq(struct ib_cq *cq)
841 {
842         if (cq->uobject) {
843                 mthca_unmap_user_db(to_mdev(cq->device),
844                                     &to_mucontext(cq->uobject->context)->uar,
845                                     to_mucontext(cq->uobject->context)->db_tab,
846                                     to_mcq(cq)->arm_db_index);
847                 mthca_unmap_user_db(to_mdev(cq->device),
848                                     &to_mucontext(cq->uobject->context)->uar,
849                                     to_mucontext(cq->uobject->context)->db_tab,
850                                     to_mcq(cq)->set_ci_db_index);
851         }
852         mthca_free_cq(to_mdev(cq->device), to_mcq(cq));
853         kfree(cq);
854
855         return 0;
856 }
857
858 static inline u32 convert_access(int acc)
859 {
860         return (acc & IB_ACCESS_REMOTE_ATOMIC ? MTHCA_MPT_FLAG_ATOMIC       : 0) |
861                (acc & IB_ACCESS_REMOTE_WRITE  ? MTHCA_MPT_FLAG_REMOTE_WRITE : 0) |
862                (acc & IB_ACCESS_REMOTE_READ   ? MTHCA_MPT_FLAG_REMOTE_READ  : 0) |
863                (acc & IB_ACCESS_LOCAL_WRITE   ? MTHCA_MPT_FLAG_LOCAL_WRITE  : 0) |
864                MTHCA_MPT_FLAG_LOCAL_READ;
865 }
866
867 static struct ib_mr *mthca_get_dma_mr(struct ib_pd *pd, int acc)
868 {
869         struct mthca_mr *mr;
870         int err;
871
872         mr = kmalloc(sizeof *mr, GFP_KERNEL);
873         if (!mr)
874                 return ERR_PTR(-ENOMEM);
875
876         err = mthca_mr_alloc_notrans(to_mdev(pd->device),
877                                      to_mpd(pd)->pd_num,
878                                      convert_access(acc), mr);
879
880         if (err) {
881                 kfree(mr);
882                 return ERR_PTR(err);
883         }
884
885         mr->umem = NULL;
886
887         return &mr->ibmr;
888 }
889
890 static struct ib_mr *mthca_reg_phys_mr(struct ib_pd       *pd,
891                                        struct ib_phys_buf *buffer_list,
892                                        int                 num_phys_buf,
893                                        int                 acc,
894                                        u64                *iova_start)
895 {
896         struct mthca_mr *mr;
897         u64 *page_list;
898         u64 total_size;
899         unsigned long mask;
900         int shift;
901         int npages;
902         int err;
903         int i, j, n;
904
905         mask = buffer_list[0].addr ^ *iova_start;
906         total_size = 0;
907         for (i = 0; i < num_phys_buf; ++i) {
908                 if (i != 0)
909                         mask |= buffer_list[i].addr;
910                 if (i != num_phys_buf - 1)
911                         mask |= buffer_list[i].addr + buffer_list[i].size;
912
913                 total_size += buffer_list[i].size;
914         }
915
916         if (mask & ~PAGE_MASK)
917                 return ERR_PTR(-EINVAL);
918
919         shift = __ffs(mask | 1 << 31);
920
921         buffer_list[0].size += buffer_list[0].addr & ((1ULL << shift) - 1);
922         buffer_list[0].addr &= ~0ull << shift;
923
924         mr = kmalloc(sizeof *mr, GFP_KERNEL);
925         if (!mr)
926                 return ERR_PTR(-ENOMEM);
927
928         npages = 0;
929         for (i = 0; i < num_phys_buf; ++i)
930                 npages += (buffer_list[i].size + (1ULL << shift) - 1) >> shift;
931
932         if (!npages)
933                 return &mr->ibmr;
934
935         page_list = kmalloc(npages * sizeof *page_list, GFP_KERNEL);
936         if (!page_list) {
937                 kfree(mr);
938                 return ERR_PTR(-ENOMEM);
939         }
940
941         n = 0;
942         for (i = 0; i < num_phys_buf; ++i)
943                 for (j = 0;
944                      j < (buffer_list[i].size + (1ULL << shift) - 1) >> shift;
945                      ++j)
946                         page_list[n++] = buffer_list[i].addr + ((u64) j << shift);
947
948         mthca_dbg(to_mdev(pd->device), "Registering memory at %llx (iova %llx) "
949                   "in PD %x; shift %d, npages %d.\n",
950                   (unsigned long long) buffer_list[0].addr,
951                   (unsigned long long) *iova_start,
952                   to_mpd(pd)->pd_num,
953                   shift, npages);
954
955         err = mthca_mr_alloc_phys(to_mdev(pd->device),
956                                   to_mpd(pd)->pd_num,
957                                   page_list, shift, npages,
958                                   *iova_start, total_size,
959                                   convert_access(acc), mr);
960
961         if (err) {
962                 kfree(page_list);
963                 kfree(mr);
964                 return ERR_PTR(err);
965         }
966
967         kfree(page_list);
968         mr->umem = NULL;
969
970         return &mr->ibmr;
971 }
972
973 static struct ib_mr *mthca_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
974                                        u64 virt, int acc, struct ib_udata *udata)
975 {
976         struct mthca_dev *dev = to_mdev(pd->device);
977         struct ib_umem_chunk *chunk;
978         struct mthca_mr *mr;
979         struct mthca_reg_mr ucmd;
980         u64 *pages;
981         int shift, n, len;
982         int i, j, k;
983         int err = 0;
984         int write_mtt_size;
985
986         if (udata->inlen - sizeof (struct ib_uverbs_cmd_hdr) < sizeof ucmd) {
987                 if (!to_mucontext(pd->uobject->context)->reg_mr_warned) {
988                         mthca_warn(dev, "Process '%s' did not pass in MR attrs.\n",
989                                    current->comm);
990                         mthca_warn(dev, "  Update libmthca to fix this.\n");
991                 }
992                 ++to_mucontext(pd->uobject->context)->reg_mr_warned;
993                 ucmd.mr_attrs = 0;
994         } else if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd))
995                 return ERR_PTR(-EFAULT);
996
997         mr = kmalloc(sizeof *mr, GFP_KERNEL);
998         if (!mr)
999                 return ERR_PTR(-ENOMEM);
1000
1001         mr->umem = ib_umem_get(pd->uobject->context, start, length, acc,
1002                                ucmd.mr_attrs & MTHCA_MR_DMASYNC);
1003
1004         if (IS_ERR(mr->umem)) {
1005                 err = PTR_ERR(mr->umem);
1006                 goto err;
1007         }
1008
1009         shift = ffs(mr->umem->page_size) - 1;
1010
1011         n = 0;
1012         list_for_each_entry(chunk, &mr->umem->chunk_list, list)
1013                 n += chunk->nents;
1014
1015         mr->mtt = mthca_alloc_mtt(dev, n);
1016         if (IS_ERR(mr->mtt)) {
1017                 err = PTR_ERR(mr->mtt);
1018                 goto err_umem;
1019         }
1020
1021         pages = (u64 *) __get_free_page(GFP_KERNEL);
1022         if (!pages) {
1023                 err = -ENOMEM;
1024                 goto err_mtt;
1025         }
1026
1027         i = n = 0;
1028
1029         write_mtt_size = min(mthca_write_mtt_size(dev), (int) (PAGE_SIZE / sizeof *pages));
1030
1031         list_for_each_entry(chunk, &mr->umem->chunk_list, list)
1032                 for (j = 0; j < chunk->nmap; ++j) {
1033                         len = sg_dma_len(&chunk->page_list[j]) >> shift;
1034                         for (k = 0; k < len; ++k) {
1035                                 pages[i++] = sg_dma_address(&chunk->page_list[j]) +
1036                                         mr->umem->page_size * k;
1037                                 /*
1038                                  * Be friendly to write_mtt and pass it chunks
1039                                  * of appropriate size.
1040                                  */
1041                                 if (i == write_mtt_size) {
1042                                         err = mthca_write_mtt(dev, mr->mtt, n, pages, i);
1043                                         if (err)
1044                                                 goto mtt_done;
1045                                         n += i;
1046                                         i = 0;
1047                                 }
1048                         }
1049                 }
1050
1051         if (i)
1052                 err = mthca_write_mtt(dev, mr->mtt, n, pages, i);
1053 mtt_done:
1054         free_page((unsigned long) pages);
1055         if (err)
1056                 goto err_mtt;
1057
1058         err = mthca_mr_alloc(dev, to_mpd(pd)->pd_num, shift, virt, length,
1059                              convert_access(acc), mr);
1060
1061         if (err)
1062                 goto err_mtt;
1063
1064         return &mr->ibmr;
1065
1066 err_mtt:
1067         mthca_free_mtt(dev, mr->mtt);
1068
1069 err_umem:
1070         ib_umem_release(mr->umem);
1071
1072 err:
1073         kfree(mr);
1074         return ERR_PTR(err);
1075 }
1076
1077 static int mthca_dereg_mr(struct ib_mr *mr)
1078 {
1079         struct mthca_mr *mmr = to_mmr(mr);
1080
1081         mthca_free_mr(to_mdev(mr->device), mmr);
1082         if (mmr->umem)
1083                 ib_umem_release(mmr->umem);
1084         kfree(mmr);
1085
1086         return 0;
1087 }
1088
1089 static struct ib_fmr *mthca_alloc_fmr(struct ib_pd *pd, int mr_access_flags,
1090                                       struct ib_fmr_attr *fmr_attr)
1091 {
1092         struct mthca_fmr *fmr;
1093         int err;
1094
1095         fmr = kmalloc(sizeof *fmr, GFP_KERNEL);
1096         if (!fmr)
1097                 return ERR_PTR(-ENOMEM);
1098
1099         memcpy(&fmr->attr, fmr_attr, sizeof *fmr_attr);
1100         err = mthca_fmr_alloc(to_mdev(pd->device), to_mpd(pd)->pd_num,
1101                              convert_access(mr_access_flags), fmr);
1102
1103         if (err) {
1104                 kfree(fmr);
1105                 return ERR_PTR(err);
1106         }
1107
1108         return &fmr->ibmr;
1109 }
1110
1111 static int mthca_dealloc_fmr(struct ib_fmr *fmr)
1112 {
1113         struct mthca_fmr *mfmr = to_mfmr(fmr);
1114         int err;
1115
1116         err = mthca_free_fmr(to_mdev(fmr->device), mfmr);
1117         if (err)
1118                 return err;
1119
1120         kfree(mfmr);
1121         return 0;
1122 }
1123
1124 static int mthca_unmap_fmr(struct list_head *fmr_list)
1125 {
1126         struct ib_fmr *fmr;
1127         int err;
1128         struct mthca_dev *mdev = NULL;
1129
1130         list_for_each_entry(fmr, fmr_list, list) {
1131                 if (mdev && to_mdev(fmr->device) != mdev)
1132                         return -EINVAL;
1133                 mdev = to_mdev(fmr->device);
1134         }
1135
1136         if (!mdev)
1137                 return 0;
1138
1139         if (mthca_is_memfree(mdev)) {
1140                 list_for_each_entry(fmr, fmr_list, list)
1141                         mthca_arbel_fmr_unmap(mdev, to_mfmr(fmr));
1142
1143                 wmb();
1144         } else
1145                 list_for_each_entry(fmr, fmr_list, list)
1146                         mthca_tavor_fmr_unmap(mdev, to_mfmr(fmr));
1147
1148         err = mthca_SYNC_TPT(mdev);
1149         return err;
1150 }
1151
1152 static ssize_t show_rev(struct device *device, struct device_attribute *attr,
1153                         char *buf)
1154 {
1155         struct mthca_dev *dev =
1156                 container_of(device, struct mthca_dev, ib_dev.dev);
1157         return sprintf(buf, "%x\n", dev->rev_id);
1158 }
1159
1160 static ssize_t show_fw_ver(struct device *device, struct device_attribute *attr,
1161                            char *buf)
1162 {
1163         struct mthca_dev *dev =
1164                 container_of(device, struct mthca_dev, ib_dev.dev);
1165         return sprintf(buf, "%d.%d.%d\n", (int) (dev->fw_ver >> 32),
1166                        (int) (dev->fw_ver >> 16) & 0xffff,
1167                        (int) dev->fw_ver & 0xffff);
1168 }
1169
1170 static ssize_t show_hca(struct device *device, struct device_attribute *attr,
1171                         char *buf)
1172 {
1173         struct mthca_dev *dev =
1174                 container_of(device, struct mthca_dev, ib_dev.dev);
1175         switch (dev->pdev->device) {
1176         case PCI_DEVICE_ID_MELLANOX_TAVOR:
1177                 return sprintf(buf, "MT23108\n");
1178         case PCI_DEVICE_ID_MELLANOX_ARBEL_COMPAT:
1179                 return sprintf(buf, "MT25208 (MT23108 compat mode)\n");
1180         case PCI_DEVICE_ID_MELLANOX_ARBEL:
1181                 return sprintf(buf, "MT25208\n");
1182         case PCI_DEVICE_ID_MELLANOX_SINAI:
1183         case PCI_DEVICE_ID_MELLANOX_SINAI_OLD:
1184                 return sprintf(buf, "MT25204\n");
1185         default:
1186                 return sprintf(buf, "unknown\n");
1187         }
1188 }
1189
1190 static ssize_t show_board(struct device *device, struct device_attribute *attr,
1191                           char *buf)
1192 {
1193         struct mthca_dev *dev =
1194                 container_of(device, struct mthca_dev, ib_dev.dev);
1195         return sprintf(buf, "%.*s\n", MTHCA_BOARD_ID_LEN, dev->board_id);
1196 }
1197
1198 static DEVICE_ATTR(hw_rev,   S_IRUGO, show_rev,    NULL);
1199 static DEVICE_ATTR(fw_ver,   S_IRUGO, show_fw_ver, NULL);
1200 static DEVICE_ATTR(hca_type, S_IRUGO, show_hca,    NULL);
1201 static DEVICE_ATTR(board_id, S_IRUGO, show_board,  NULL);
1202
1203 static struct device_attribute *mthca_dev_attributes[] = {
1204         &dev_attr_hw_rev,
1205         &dev_attr_fw_ver,
1206         &dev_attr_hca_type,
1207         &dev_attr_board_id
1208 };
1209
1210 static int mthca_init_node_data(struct mthca_dev *dev)
1211 {
1212         struct ib_smp *in_mad  = NULL;
1213         struct ib_smp *out_mad = NULL;
1214         int err = -ENOMEM;
1215
1216         in_mad  = kzalloc(sizeof *in_mad, GFP_KERNEL);
1217         out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
1218         if (!in_mad || !out_mad)
1219                 goto out;
1220
1221         init_query_mad(in_mad);
1222         in_mad->attr_id = IB_SMP_ATTR_NODE_DESC;
1223
1224         err = mthca_MAD_IFC(dev, 1, 1,
1225                             1, NULL, NULL, in_mad, out_mad);
1226         if (err)
1227                 goto out;
1228
1229         memcpy(dev->ib_dev.node_desc, out_mad->data, 64);
1230
1231         in_mad->attr_id = IB_SMP_ATTR_NODE_INFO;
1232
1233         err = mthca_MAD_IFC(dev, 1, 1,
1234                             1, NULL, NULL, in_mad, out_mad);
1235         if (err)
1236                 goto out;
1237
1238         if (mthca_is_memfree(dev))
1239                 dev->rev_id = be32_to_cpup((__be32 *) (out_mad->data + 32));
1240         memcpy(&dev->ib_dev.node_guid, out_mad->data + 12, 8);
1241
1242 out:
1243         kfree(in_mad);
1244         kfree(out_mad);
1245         return err;
1246 }
1247
1248 int mthca_register_device(struct mthca_dev *dev)
1249 {
1250         int ret;
1251         int i;
1252
1253         ret = mthca_init_node_data(dev);
1254         if (ret)
1255                 return ret;
1256
1257         strlcpy(dev->ib_dev.name, "mthca%d", IB_DEVICE_NAME_MAX);
1258         dev->ib_dev.owner                = THIS_MODULE;
1259
1260         dev->ib_dev.uverbs_abi_ver       = MTHCA_UVERBS_ABI_VERSION;
1261         dev->ib_dev.uverbs_cmd_mask      =
1262                 (1ull << IB_USER_VERBS_CMD_GET_CONTEXT)         |
1263                 (1ull << IB_USER_VERBS_CMD_QUERY_DEVICE)        |
1264                 (1ull << IB_USER_VERBS_CMD_QUERY_PORT)          |
1265                 (1ull << IB_USER_VERBS_CMD_ALLOC_PD)            |
1266                 (1ull << IB_USER_VERBS_CMD_DEALLOC_PD)          |
1267                 (1ull << IB_USER_VERBS_CMD_REG_MR)              |
1268                 (1ull << IB_USER_VERBS_CMD_DEREG_MR)            |
1269                 (1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL) |
1270                 (1ull << IB_USER_VERBS_CMD_CREATE_CQ)           |
1271                 (1ull << IB_USER_VERBS_CMD_RESIZE_CQ)           |
1272                 (1ull << IB_USER_VERBS_CMD_DESTROY_CQ)          |
1273                 (1ull << IB_USER_VERBS_CMD_CREATE_QP)           |
1274                 (1ull << IB_USER_VERBS_CMD_QUERY_QP)            |
1275                 (1ull << IB_USER_VERBS_CMD_MODIFY_QP)           |
1276                 (1ull << IB_USER_VERBS_CMD_DESTROY_QP)          |
1277                 (1ull << IB_USER_VERBS_CMD_ATTACH_MCAST)        |
1278                 (1ull << IB_USER_VERBS_CMD_DETACH_MCAST);
1279         dev->ib_dev.node_type            = RDMA_NODE_IB_CA;
1280         dev->ib_dev.phys_port_cnt        = dev->limits.num_ports;
1281         dev->ib_dev.num_comp_vectors     = 1;
1282         dev->ib_dev.dma_device           = &dev->pdev->dev;
1283         dev->ib_dev.query_device         = mthca_query_device;
1284         dev->ib_dev.query_port           = mthca_query_port;
1285         dev->ib_dev.modify_device        = mthca_modify_device;
1286         dev->ib_dev.modify_port          = mthca_modify_port;
1287         dev->ib_dev.query_pkey           = mthca_query_pkey;
1288         dev->ib_dev.query_gid            = mthca_query_gid;
1289         dev->ib_dev.alloc_ucontext       = mthca_alloc_ucontext;
1290         dev->ib_dev.dealloc_ucontext     = mthca_dealloc_ucontext;
1291         dev->ib_dev.mmap                 = mthca_mmap_uar;
1292         dev->ib_dev.alloc_pd             = mthca_alloc_pd;
1293         dev->ib_dev.dealloc_pd           = mthca_dealloc_pd;
1294         dev->ib_dev.create_ah            = mthca_ah_create;
1295         dev->ib_dev.query_ah             = mthca_ah_query;
1296         dev->ib_dev.destroy_ah           = mthca_ah_destroy;
1297
1298         if (dev->mthca_flags & MTHCA_FLAG_SRQ) {
1299                 dev->ib_dev.create_srq           = mthca_create_srq;
1300                 dev->ib_dev.modify_srq           = mthca_modify_srq;
1301                 dev->ib_dev.query_srq            = mthca_query_srq;
1302                 dev->ib_dev.destroy_srq          = mthca_destroy_srq;
1303                 dev->ib_dev.uverbs_cmd_mask     |=
1304                         (1ull << IB_USER_VERBS_CMD_CREATE_SRQ)          |
1305                         (1ull << IB_USER_VERBS_CMD_MODIFY_SRQ)          |
1306                         (1ull << IB_USER_VERBS_CMD_QUERY_SRQ)           |
1307                         (1ull << IB_USER_VERBS_CMD_DESTROY_SRQ);
1308
1309                 if (mthca_is_memfree(dev))
1310                         dev->ib_dev.post_srq_recv = mthca_arbel_post_srq_recv;
1311                 else
1312                         dev->ib_dev.post_srq_recv = mthca_tavor_post_srq_recv;
1313         }
1314
1315         dev->ib_dev.create_qp            = mthca_create_qp;
1316         dev->ib_dev.modify_qp            = mthca_modify_qp;
1317         dev->ib_dev.query_qp             = mthca_query_qp;
1318         dev->ib_dev.destroy_qp           = mthca_destroy_qp;
1319         dev->ib_dev.create_cq            = mthca_create_cq;
1320         dev->ib_dev.resize_cq            = mthca_resize_cq;
1321         dev->ib_dev.destroy_cq           = mthca_destroy_cq;
1322         dev->ib_dev.poll_cq              = mthca_poll_cq;
1323         dev->ib_dev.get_dma_mr           = mthca_get_dma_mr;
1324         dev->ib_dev.reg_phys_mr          = mthca_reg_phys_mr;
1325         dev->ib_dev.reg_user_mr          = mthca_reg_user_mr;
1326         dev->ib_dev.dereg_mr             = mthca_dereg_mr;
1327
1328         if (dev->mthca_flags & MTHCA_FLAG_FMR) {
1329                 dev->ib_dev.alloc_fmr            = mthca_alloc_fmr;
1330                 dev->ib_dev.unmap_fmr            = mthca_unmap_fmr;
1331                 dev->ib_dev.dealloc_fmr          = mthca_dealloc_fmr;
1332                 if (mthca_is_memfree(dev))
1333                         dev->ib_dev.map_phys_fmr = mthca_arbel_map_phys_fmr;
1334                 else
1335                         dev->ib_dev.map_phys_fmr = mthca_tavor_map_phys_fmr;
1336         }
1337
1338         dev->ib_dev.attach_mcast         = mthca_multicast_attach;
1339         dev->ib_dev.detach_mcast         = mthca_multicast_detach;
1340         dev->ib_dev.process_mad          = mthca_process_mad;
1341
1342         if (mthca_is_memfree(dev)) {
1343                 dev->ib_dev.req_notify_cq = mthca_arbel_arm_cq;
1344                 dev->ib_dev.post_send     = mthca_arbel_post_send;
1345                 dev->ib_dev.post_recv     = mthca_arbel_post_receive;
1346         } else {
1347                 dev->ib_dev.req_notify_cq = mthca_tavor_arm_cq;
1348                 dev->ib_dev.post_send     = mthca_tavor_post_send;
1349                 dev->ib_dev.post_recv     = mthca_tavor_post_receive;
1350         }
1351
1352         mutex_init(&dev->cap_mask_mutex);
1353
1354         ret = ib_register_device(&dev->ib_dev, NULL);
1355         if (ret)
1356                 return ret;
1357
1358         for (i = 0; i < ARRAY_SIZE(mthca_dev_attributes); ++i) {
1359                 ret = device_create_file(&dev->ib_dev.dev,
1360                                          mthca_dev_attributes[i]);
1361                 if (ret) {
1362                         ib_unregister_device(&dev->ib_dev);
1363                         return ret;
1364                 }
1365         }
1366
1367         mthca_start_catas_poll(dev);
1368
1369         return 0;
1370 }
1371
1372 void mthca_unregister_device(struct mthca_dev *dev)
1373 {
1374         mthca_stop_catas_poll(dev);
1375         ib_unregister_device(&dev->ib_dev);
1376 }