Merge master.kernel.org:/pub/scm/linux/kernel/git/davej/x86
[sfrench/cifs-2.6.git] / drivers / infiniband / hw / mthca / mthca_provider.c
1 /*
2  * Copyright (c) 2004, 2005 Topspin Communications.  All rights reserved.
3  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
4  * Copyright (c) 2005 Cisco Systems. All rights reserved.
5  * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
6  * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
7  *
8  * This software is available to you under a choice of one of two
9  * licenses.  You may choose to be licensed under the terms of the GNU
10  * General Public License (GPL) Version 2, available from the file
11  * COPYING in the main directory of this source tree, or the
12  * OpenIB.org BSD license below:
13  *
14  *     Redistribution and use in source and binary forms, with or
15  *     without modification, are permitted provided that the following
16  *     conditions are met:
17  *
18  *      - Redistributions of source code must retain the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer.
21  *
22  *      - Redistributions in binary form must reproduce the above
23  *        copyright notice, this list of conditions and the following
24  *        disclaimer in the documentation and/or other materials
25  *        provided with the distribution.
26  *
27  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
28  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
29  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
30  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
31  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
32  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
33  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
34  * SOFTWARE.
35  *
36  * $Id: mthca_provider.c 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_sg;
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         err = 0;
120  out:
121         kfree(in_mad);
122         kfree(out_mad);
123         return err;
124 }
125
126 static int mthca_query_port(struct ib_device *ibdev,
127                             u8 port, struct ib_port_attr *props)
128 {
129         struct ib_smp *in_mad  = NULL;
130         struct ib_smp *out_mad = NULL;
131         int err = -ENOMEM;
132         u8 status;
133
134         in_mad  = kzalloc(sizeof *in_mad, GFP_KERNEL);
135         out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
136         if (!in_mad || !out_mad)
137                 goto out;
138
139         memset(props, 0, sizeof *props);
140
141         init_query_mad(in_mad);
142         in_mad->attr_id  = IB_SMP_ATTR_PORT_INFO;
143         in_mad->attr_mod = cpu_to_be32(port);
144
145         err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
146                             port, NULL, NULL, in_mad, out_mad,
147                             &status);
148         if (err)
149                 goto out;
150         if (status) {
151                 err = -EINVAL;
152                 goto out;
153         }
154
155         props->lid               = be16_to_cpup((__be16 *) (out_mad->data + 16));
156         props->lmc               = out_mad->data[34] & 0x7;
157         props->sm_lid            = be16_to_cpup((__be16 *) (out_mad->data + 18));
158         props->sm_sl             = out_mad->data[36] & 0xf;
159         props->state             = out_mad->data[32] & 0xf;
160         props->phys_state        = out_mad->data[33] >> 4;
161         props->port_cap_flags    = be32_to_cpup((__be32 *) (out_mad->data + 20));
162         props->gid_tbl_len       = to_mdev(ibdev)->limits.gid_table_len;
163         props->max_msg_sz        = 0x80000000;
164         props->pkey_tbl_len      = to_mdev(ibdev)->limits.pkey_table_len;
165         props->bad_pkey_cntr     = be16_to_cpup((__be16 *) (out_mad->data + 46));
166         props->qkey_viol_cntr    = be16_to_cpup((__be16 *) (out_mad->data + 48));
167         props->active_width      = out_mad->data[31] & 0xf;
168         props->active_speed      = out_mad->data[35] >> 4;
169         props->max_mtu           = out_mad->data[41] & 0xf;
170         props->active_mtu        = out_mad->data[36] >> 4;
171         props->subnet_timeout    = out_mad->data[51] & 0x1f;
172
173  out:
174         kfree(in_mad);
175         kfree(out_mad);
176         return err;
177 }
178
179 static int mthca_modify_port(struct ib_device *ibdev,
180                              u8 port, int port_modify_mask,
181                              struct ib_port_modify *props)
182 {
183         struct mthca_set_ib_param set_ib;
184         struct ib_port_attr attr;
185         int err;
186         u8 status;
187
188         if (down_interruptible(&to_mdev(ibdev)->cap_mask_mutex))
189                 return -ERESTARTSYS;
190
191         err = mthca_query_port(ibdev, port, &attr);
192         if (err)
193                 goto out;
194
195         set_ib.set_si_guid     = 0;
196         set_ib.reset_qkey_viol = !!(port_modify_mask & IB_PORT_RESET_QKEY_CNTR);
197
198         set_ib.cap_mask = (attr.port_cap_flags | props->set_port_cap_mask) &
199                 ~props->clr_port_cap_mask;
200
201         err = mthca_SET_IB(to_mdev(ibdev), &set_ib, port, &status);
202         if (err)
203                 goto out;
204         if (status) {
205                 err = -EINVAL;
206                 goto out;
207         }
208
209 out:
210         up(&to_mdev(ibdev)->cap_mask_mutex);
211         return err;
212 }
213
214 static int mthca_query_pkey(struct ib_device *ibdev,
215                             u8 port, u16 index, u16 *pkey)
216 {
217         struct ib_smp *in_mad  = NULL;
218         struct ib_smp *out_mad = NULL;
219         int err = -ENOMEM;
220         u8 status;
221
222         in_mad  = kzalloc(sizeof *in_mad, GFP_KERNEL);
223         out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
224         if (!in_mad || !out_mad)
225                 goto out;
226
227         init_query_mad(in_mad);
228         in_mad->attr_id  = IB_SMP_ATTR_PKEY_TABLE;
229         in_mad->attr_mod = cpu_to_be32(index / 32);
230
231         err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
232                             port, NULL, NULL, in_mad, out_mad,
233                             &status);
234         if (err)
235                 goto out;
236         if (status) {
237                 err = -EINVAL;
238                 goto out;
239         }
240
241         *pkey = be16_to_cpu(((__be16 *) out_mad->data)[index % 32]);
242
243  out:
244         kfree(in_mad);
245         kfree(out_mad);
246         return err;
247 }
248
249 static int mthca_query_gid(struct ib_device *ibdev, u8 port,
250                            int index, union ib_gid *gid)
251 {
252         struct ib_smp *in_mad  = NULL;
253         struct ib_smp *out_mad = NULL;
254         int err = -ENOMEM;
255         u8 status;
256
257         in_mad  = kzalloc(sizeof *in_mad, GFP_KERNEL);
258         out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
259         if (!in_mad || !out_mad)
260                 goto out;
261
262         init_query_mad(in_mad);
263         in_mad->attr_id  = IB_SMP_ATTR_PORT_INFO;
264         in_mad->attr_mod = cpu_to_be32(port);
265
266         err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
267                             port, NULL, NULL, in_mad, out_mad,
268                             &status);
269         if (err)
270                 goto out;
271         if (status) {
272                 err = -EINVAL;
273                 goto out;
274         }
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                             &status);
285         if (err)
286                 goto out;
287         if (status) {
288                 err = -EINVAL;
289                 goto out;
290         }
291
292         memcpy(gid->raw + 8, out_mad->data + (index % 8) * 16, 8);
293
294  out:
295         kfree(in_mad);
296         kfree(out_mad);
297         return err;
298 }
299
300 static struct ib_ucontext *mthca_alloc_ucontext(struct ib_device *ibdev,
301                                                 struct ib_udata *udata)
302 {
303         struct mthca_alloc_ucontext_resp uresp;
304         struct mthca_ucontext           *context;
305         int                              err;
306
307         memset(&uresp, 0, sizeof uresp);
308
309         uresp.qp_tab_size = to_mdev(ibdev)->limits.num_qps;
310         if (mthca_is_memfree(to_mdev(ibdev)))
311                 uresp.uarc_size = to_mdev(ibdev)->uar_table.uarc_size;
312         else
313                 uresp.uarc_size = 0;
314
315         context = kmalloc(sizeof *context, GFP_KERNEL);
316         if (!context)
317                 return ERR_PTR(-ENOMEM);
318
319         err = mthca_uar_alloc(to_mdev(ibdev), &context->uar);
320         if (err) {
321                 kfree(context);
322                 return ERR_PTR(err);
323         }
324
325         context->db_tab = mthca_init_user_db_tab(to_mdev(ibdev));
326         if (IS_ERR(context->db_tab)) {
327                 err = PTR_ERR(context->db_tab);
328                 mthca_uar_free(to_mdev(ibdev), &context->uar);
329                 kfree(context);
330                 return ERR_PTR(err);
331         }
332
333         if (ib_copy_to_udata(udata, &uresp, sizeof uresp)) {
334                 mthca_cleanup_user_db_tab(to_mdev(ibdev), &context->uar, context->db_tab);
335                 mthca_uar_free(to_mdev(ibdev), &context->uar);
336                 kfree(context);
337                 return ERR_PTR(-EFAULT);
338         }
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         srq = kmalloc(sizeof *srq, GFP_KERNEL);
442         if (!srq)
443                 return ERR_PTR(-ENOMEM);
444
445         if (pd->uobject) {
446                 context = to_mucontext(pd->uobject->context);
447
448                 if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
449                         err = -EFAULT;
450                         goto err_free;
451                 }
452
453                 err = mthca_map_user_db(to_mdev(pd->device), &context->uar,
454                                         context->db_tab, ucmd.db_index,
455                                         ucmd.db_page);
456
457                 if (err)
458                         goto err_free;
459
460                 srq->mr.ibmr.lkey = ucmd.lkey;
461                 srq->db_index     = ucmd.db_index;
462         }
463
464         err = mthca_alloc_srq(to_mdev(pd->device), to_mpd(pd),
465                               &init_attr->attr, srq);
466
467         if (err && pd->uobject)
468                 mthca_unmap_user_db(to_mdev(pd->device), &context->uar,
469                                     context->db_tab, ucmd.db_index);
470
471         if (err)
472                 goto err_free;
473
474         if (context && ib_copy_to_udata(udata, &srq->srqn, sizeof (__u32))) {
475                 mthca_free_srq(to_mdev(pd->device), srq);
476                 err = -EFAULT;
477                 goto err_free;
478         }
479
480         return &srq->ibsrq;
481
482 err_free:
483         kfree(srq);
484
485         return ERR_PTR(err);
486 }
487
488 static int mthca_destroy_srq(struct ib_srq *srq)
489 {
490         struct mthca_ucontext *context;
491
492         if (srq->uobject) {
493                 context = to_mucontext(srq->uobject->context);
494
495                 mthca_unmap_user_db(to_mdev(srq->device), &context->uar,
496                                     context->db_tab, to_msrq(srq)->db_index);
497         }
498
499         mthca_free_srq(to_mdev(srq->device), to_msrq(srq));
500         kfree(srq);
501
502         return 0;
503 }
504
505 static struct ib_qp *mthca_create_qp(struct ib_pd *pd,
506                                      struct ib_qp_init_attr *init_attr,
507                                      struct ib_udata *udata)
508 {
509         struct mthca_create_qp ucmd;
510         struct mthca_qp *qp;
511         int err;
512
513         switch (init_attr->qp_type) {
514         case IB_QPT_RC:
515         case IB_QPT_UC:
516         case IB_QPT_UD:
517         {
518                 struct mthca_ucontext *context;
519
520                 qp = kmalloc(sizeof *qp, GFP_KERNEL);
521                 if (!qp)
522                         return ERR_PTR(-ENOMEM);
523
524                 if (pd->uobject) {
525                         context = to_mucontext(pd->uobject->context);
526
527                         if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
528                                 kfree(qp);
529                                 return ERR_PTR(-EFAULT);
530                         }
531
532                         err = mthca_map_user_db(to_mdev(pd->device), &context->uar,
533                                                 context->db_tab,
534                                                 ucmd.sq_db_index, ucmd.sq_db_page);
535                         if (err) {
536                                 kfree(qp);
537                                 return ERR_PTR(err);
538                         }
539
540                         err = mthca_map_user_db(to_mdev(pd->device), &context->uar,
541                                                 context->db_tab,
542                                                 ucmd.rq_db_index, ucmd.rq_db_page);
543                         if (err) {
544                                 mthca_unmap_user_db(to_mdev(pd->device),
545                                                     &context->uar,
546                                                     context->db_tab,
547                                                     ucmd.sq_db_index);
548                                 kfree(qp);
549                                 return ERR_PTR(err);
550                         }
551
552                         qp->mr.ibmr.lkey = ucmd.lkey;
553                         qp->sq.db_index  = ucmd.sq_db_index;
554                         qp->rq.db_index  = ucmd.rq_db_index;
555                 }
556
557                 err = mthca_alloc_qp(to_mdev(pd->device), to_mpd(pd),
558                                      to_mcq(init_attr->send_cq),
559                                      to_mcq(init_attr->recv_cq),
560                                      init_attr->qp_type, init_attr->sq_sig_type,
561                                      &init_attr->cap, qp);
562
563                 if (err && pd->uobject) {
564                         context = to_mucontext(pd->uobject->context);
565
566                         mthca_unmap_user_db(to_mdev(pd->device),
567                                             &context->uar,
568                                             context->db_tab,
569                                             ucmd.sq_db_index);
570                         mthca_unmap_user_db(to_mdev(pd->device),
571                                             &context->uar,
572                                             context->db_tab,
573                                             ucmd.rq_db_index);
574                 }
575
576                 qp->ibqp.qp_num = qp->qpn;
577                 break;
578         }
579         case IB_QPT_SMI:
580         case IB_QPT_GSI:
581         {
582                 /* Don't allow userspace to create special QPs */
583                 if (pd->uobject)
584                         return ERR_PTR(-EINVAL);
585
586                 qp = kmalloc(sizeof (struct mthca_sqp), GFP_KERNEL);
587                 if (!qp)
588                         return ERR_PTR(-ENOMEM);
589
590                 qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 : 1;
591
592                 err = mthca_alloc_sqp(to_mdev(pd->device), to_mpd(pd),
593                                       to_mcq(init_attr->send_cq),
594                                       to_mcq(init_attr->recv_cq),
595                                       init_attr->sq_sig_type, &init_attr->cap,
596                                       qp->ibqp.qp_num, init_attr->port_num,
597                                       to_msqp(qp));
598                 break;
599         }
600         default:
601                 /* Don't support raw QPs */
602                 return ERR_PTR(-ENOSYS);
603         }
604
605         if (err) {
606                 kfree(qp);
607                 return ERR_PTR(err);
608         }
609
610         init_attr->cap.max_send_wr     = qp->sq.max;
611         init_attr->cap.max_recv_wr     = qp->rq.max;
612         init_attr->cap.max_send_sge    = qp->sq.max_gs;
613         init_attr->cap.max_recv_sge    = qp->rq.max_gs;
614         init_attr->cap.max_inline_data = qp->max_inline_data;
615
616         return &qp->ibqp;
617 }
618
619 static int mthca_destroy_qp(struct ib_qp *qp)
620 {
621         if (qp->uobject) {
622                 mthca_unmap_user_db(to_mdev(qp->device),
623                                     &to_mucontext(qp->uobject->context)->uar,
624                                     to_mucontext(qp->uobject->context)->db_tab,
625                                     to_mqp(qp)->sq.db_index);
626                 mthca_unmap_user_db(to_mdev(qp->device),
627                                     &to_mucontext(qp->uobject->context)->uar,
628                                     to_mucontext(qp->uobject->context)->db_tab,
629                                     to_mqp(qp)->rq.db_index);
630         }
631         mthca_free_qp(to_mdev(qp->device), to_mqp(qp));
632         kfree(qp);
633         return 0;
634 }
635
636 static struct ib_cq *mthca_create_cq(struct ib_device *ibdev, int entries,
637                                      struct ib_ucontext *context,
638                                      struct ib_udata *udata)
639 {
640         struct mthca_create_cq ucmd;
641         struct mthca_cq *cq;
642         int nent;
643         int err;
644
645         if (entries < 1 || entries > to_mdev(ibdev)->limits.max_cqes)
646                 return ERR_PTR(-EINVAL);
647
648         if (context) {
649                 if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd))
650                         return ERR_PTR(-EFAULT);
651
652                 err = mthca_map_user_db(to_mdev(ibdev), &to_mucontext(context)->uar,
653                                         to_mucontext(context)->db_tab,
654                                         ucmd.set_db_index, ucmd.set_db_page);
655                 if (err)
656                         return ERR_PTR(err);
657
658                 err = mthca_map_user_db(to_mdev(ibdev), &to_mucontext(context)->uar,
659                                         to_mucontext(context)->db_tab,
660                                         ucmd.arm_db_index, ucmd.arm_db_page);
661                 if (err)
662                         goto err_unmap_set;
663         }
664
665         cq = kmalloc(sizeof *cq, GFP_KERNEL);
666         if (!cq) {
667                 err = -ENOMEM;
668                 goto err_unmap_arm;
669         }
670
671         if (context) {
672                 cq->mr.ibmr.lkey    = ucmd.lkey;
673                 cq->set_ci_db_index = ucmd.set_db_index;
674                 cq->arm_db_index    = ucmd.arm_db_index;
675         }
676
677         for (nent = 1; nent <= entries; nent <<= 1)
678                 ; /* nothing */
679
680         err = mthca_init_cq(to_mdev(ibdev), nent,
681                             context ? to_mucontext(context) : NULL,
682                             context ? ucmd.pdn : to_mdev(ibdev)->driver_pd.pd_num,
683                             cq);
684         if (err)
685                 goto err_free;
686
687         if (context && ib_copy_to_udata(udata, &cq->cqn, sizeof (__u32))) {
688                 mthca_free_cq(to_mdev(ibdev), cq);
689                 goto err_free;
690         }
691
692         return &cq->ibcq;
693
694 err_free:
695         kfree(cq);
696
697 err_unmap_arm:
698         if (context)
699                 mthca_unmap_user_db(to_mdev(ibdev), &to_mucontext(context)->uar,
700                                     to_mucontext(context)->db_tab, ucmd.arm_db_index);
701
702 err_unmap_set:
703         if (context)
704                 mthca_unmap_user_db(to_mdev(ibdev), &to_mucontext(context)->uar,
705                                     to_mucontext(context)->db_tab, ucmd.set_db_index);
706
707         return ERR_PTR(err);
708 }
709
710 static int mthca_destroy_cq(struct ib_cq *cq)
711 {
712         if (cq->uobject) {
713                 mthca_unmap_user_db(to_mdev(cq->device),
714                                     &to_mucontext(cq->uobject->context)->uar,
715                                     to_mucontext(cq->uobject->context)->db_tab,
716                                     to_mcq(cq)->arm_db_index);
717                 mthca_unmap_user_db(to_mdev(cq->device),
718                                     &to_mucontext(cq->uobject->context)->uar,
719                                     to_mucontext(cq->uobject->context)->db_tab,
720                                     to_mcq(cq)->set_ci_db_index);
721         }
722         mthca_free_cq(to_mdev(cq->device), to_mcq(cq));
723         kfree(cq);
724
725         return 0;
726 }
727
728 static inline u32 convert_access(int acc)
729 {
730         return (acc & IB_ACCESS_REMOTE_ATOMIC ? MTHCA_MPT_FLAG_ATOMIC       : 0) |
731                (acc & IB_ACCESS_REMOTE_WRITE  ? MTHCA_MPT_FLAG_REMOTE_WRITE : 0) |
732                (acc & IB_ACCESS_REMOTE_READ   ? MTHCA_MPT_FLAG_REMOTE_READ  : 0) |
733                (acc & IB_ACCESS_LOCAL_WRITE   ? MTHCA_MPT_FLAG_LOCAL_WRITE  : 0) |
734                MTHCA_MPT_FLAG_LOCAL_READ;
735 }
736
737 static struct ib_mr *mthca_get_dma_mr(struct ib_pd *pd, int acc)
738 {
739         struct mthca_mr *mr;
740         int err;
741
742         mr = kmalloc(sizeof *mr, GFP_KERNEL);
743         if (!mr)
744                 return ERR_PTR(-ENOMEM);
745
746         err = mthca_mr_alloc_notrans(to_mdev(pd->device),
747                                      to_mpd(pd)->pd_num,
748                                      convert_access(acc), mr);
749
750         if (err) {
751                 kfree(mr);
752                 return ERR_PTR(err);
753         }
754
755         return &mr->ibmr;
756 }
757
758 static struct ib_mr *mthca_reg_phys_mr(struct ib_pd       *pd,
759                                        struct ib_phys_buf *buffer_list,
760                                        int                 num_phys_buf,
761                                        int                 acc,
762                                        u64                *iova_start)
763 {
764         struct mthca_mr *mr;
765         u64 *page_list;
766         u64 total_size;
767         u64 mask;
768         int shift;
769         int npages;
770         int err;
771         int i, j, n;
772
773         /* First check that we have enough alignment */
774         if ((*iova_start & ~PAGE_MASK) != (buffer_list[0].addr & ~PAGE_MASK))
775                 return ERR_PTR(-EINVAL);
776
777         mask = 0;
778         total_size = 0;
779         for (i = 0; i < num_phys_buf; ++i) {
780                 if (i != 0)
781                         mask |= buffer_list[i].addr;
782                 if (i != num_phys_buf - 1)
783                         mask |= buffer_list[i].addr + buffer_list[i].size;
784
785                 total_size += buffer_list[i].size;
786         }
787
788         if (mask & ~PAGE_MASK)
789                 return ERR_PTR(-EINVAL);
790
791         /* Find largest page shift we can use to cover buffers */
792         for (shift = PAGE_SHIFT; shift < 31; ++shift)
793                 if (num_phys_buf > 1) {
794                         if ((1ULL << shift) & mask)
795                                 break;
796                 } else {
797                         if (1ULL << shift >=
798                             buffer_list[0].size +
799                             (buffer_list[0].addr & ((1ULL << shift) - 1)))
800                                 break;
801                 }
802
803         buffer_list[0].size += buffer_list[0].addr & ((1ULL << shift) - 1);
804         buffer_list[0].addr &= ~0ull << shift;
805
806         mr = kmalloc(sizeof *mr, GFP_KERNEL);
807         if (!mr)
808                 return ERR_PTR(-ENOMEM);
809
810         npages = 0;
811         for (i = 0; i < num_phys_buf; ++i)
812                 npages += (buffer_list[i].size + (1ULL << shift) - 1) >> shift;
813
814         if (!npages)
815                 return &mr->ibmr;
816
817         page_list = kmalloc(npages * sizeof *page_list, GFP_KERNEL);
818         if (!page_list) {
819                 kfree(mr);
820                 return ERR_PTR(-ENOMEM);
821         }
822
823         n = 0;
824         for (i = 0; i < num_phys_buf; ++i)
825                 for (j = 0;
826                      j < (buffer_list[i].size + (1ULL << shift) - 1) >> shift;
827                      ++j)
828                         page_list[n++] = buffer_list[i].addr + ((u64) j << shift);
829
830         mthca_dbg(to_mdev(pd->device), "Registering memory at %llx (iova %llx) "
831                   "in PD %x; shift %d, npages %d.\n",
832                   (unsigned long long) buffer_list[0].addr,
833                   (unsigned long long) *iova_start,
834                   to_mpd(pd)->pd_num,
835                   shift, npages);
836
837         err = mthca_mr_alloc_phys(to_mdev(pd->device),
838                                   to_mpd(pd)->pd_num,
839                                   page_list, shift, npages,
840                                   *iova_start, total_size,
841                                   convert_access(acc), mr);
842
843         if (err) {
844                 kfree(page_list);
845                 kfree(mr);
846                 return ERR_PTR(err);
847         }
848
849         kfree(page_list);
850         return &mr->ibmr;
851 }
852
853 static struct ib_mr *mthca_reg_user_mr(struct ib_pd *pd, struct ib_umem *region,
854                                        int acc, struct ib_udata *udata)
855 {
856         struct mthca_dev *dev = to_mdev(pd->device);
857         struct ib_umem_chunk *chunk;
858         struct mthca_mr *mr;
859         u64 *pages;
860         int shift, n, len;
861         int i, j, k;
862         int err = 0;
863
864         shift = ffs(region->page_size) - 1;
865
866         mr = kmalloc(sizeof *mr, GFP_KERNEL);
867         if (!mr)
868                 return ERR_PTR(-ENOMEM);
869
870         n = 0;
871         list_for_each_entry(chunk, &region->chunk_list, list)
872                 n += chunk->nents;
873
874         mr->mtt = mthca_alloc_mtt(dev, n);
875         if (IS_ERR(mr->mtt)) {
876                 err = PTR_ERR(mr->mtt);
877                 goto err;
878         }
879
880         pages = (u64 *) __get_free_page(GFP_KERNEL);
881         if (!pages) {
882                 err = -ENOMEM;
883                 goto err_mtt;
884         }
885
886         i = n = 0;
887
888         list_for_each_entry(chunk, &region->chunk_list, list)
889                 for (j = 0; j < chunk->nmap; ++j) {
890                         len = sg_dma_len(&chunk->page_list[j]) >> shift;
891                         for (k = 0; k < len; ++k) {
892                                 pages[i++] = sg_dma_address(&chunk->page_list[j]) +
893                                         region->page_size * k;
894                                 /*
895                                  * Be friendly to WRITE_MTT command
896                                  * and leave two empty slots for the
897                                  * index and reserved fields of the
898                                  * mailbox.
899                                  */
900                                 if (i == PAGE_SIZE / sizeof (u64) - 2) {
901                                         err = mthca_write_mtt(dev, mr->mtt,
902                                                               n, pages, i);
903                                         if (err)
904                                                 goto mtt_done;
905                                         n += i;
906                                         i = 0;
907                                 }
908                         }
909                 }
910
911         if (i)
912                 err = mthca_write_mtt(dev, mr->mtt, n, pages, i);
913 mtt_done:
914         free_page((unsigned long) pages);
915         if (err)
916                 goto err_mtt;
917
918         err = mthca_mr_alloc(dev, to_mpd(pd)->pd_num, shift, region->virt_base,
919                              region->length, convert_access(acc), mr);
920
921         if (err)
922                 goto err_mtt;
923
924         return &mr->ibmr;
925
926 err_mtt:
927         mthca_free_mtt(dev, mr->mtt);
928
929 err:
930         kfree(mr);
931         return ERR_PTR(err);
932 }
933
934 static int mthca_dereg_mr(struct ib_mr *mr)
935 {
936         struct mthca_mr *mmr = to_mmr(mr);
937         mthca_free_mr(to_mdev(mr->device), mmr);
938         kfree(mmr);
939         return 0;
940 }
941
942 static struct ib_fmr *mthca_alloc_fmr(struct ib_pd *pd, int mr_access_flags,
943                                       struct ib_fmr_attr *fmr_attr)
944 {
945         struct mthca_fmr *fmr;
946         int err;
947
948         fmr = kmalloc(sizeof *fmr, GFP_KERNEL);
949         if (!fmr)
950                 return ERR_PTR(-ENOMEM);
951
952         memcpy(&fmr->attr, fmr_attr, sizeof *fmr_attr);
953         err = mthca_fmr_alloc(to_mdev(pd->device), to_mpd(pd)->pd_num,
954                              convert_access(mr_access_flags), fmr);
955
956         if (err) {
957                 kfree(fmr);
958                 return ERR_PTR(err);
959         }
960
961         return &fmr->ibmr;
962 }
963
964 static int mthca_dealloc_fmr(struct ib_fmr *fmr)
965 {
966         struct mthca_fmr *mfmr = to_mfmr(fmr);
967         int err;
968
969         err = mthca_free_fmr(to_mdev(fmr->device), mfmr);
970         if (err)
971                 return err;
972
973         kfree(mfmr);
974         return 0;
975 }
976
977 static int mthca_unmap_fmr(struct list_head *fmr_list)
978 {
979         struct ib_fmr *fmr;
980         int err;
981         u8 status;
982         struct mthca_dev *mdev = NULL;
983
984         list_for_each_entry(fmr, fmr_list, list) {
985                 if (mdev && to_mdev(fmr->device) != mdev)
986                         return -EINVAL;
987                 mdev = to_mdev(fmr->device);
988         }
989
990         if (!mdev)
991                 return 0;
992
993         if (mthca_is_memfree(mdev)) {
994                 list_for_each_entry(fmr, fmr_list, list)
995                         mthca_arbel_fmr_unmap(mdev, to_mfmr(fmr));
996
997                 wmb();
998         } else
999                 list_for_each_entry(fmr, fmr_list, list)
1000                         mthca_tavor_fmr_unmap(mdev, to_mfmr(fmr));
1001
1002         err = mthca_SYNC_TPT(mdev, &status);
1003         if (err)
1004                 return err;
1005         if (status)
1006                 return -EINVAL;
1007         return 0;
1008 }
1009
1010 static ssize_t show_rev(struct class_device *cdev, char *buf)
1011 {
1012         struct mthca_dev *dev = container_of(cdev, struct mthca_dev, ib_dev.class_dev);
1013         return sprintf(buf, "%x\n", dev->rev_id);
1014 }
1015
1016 static ssize_t show_fw_ver(struct class_device *cdev, char *buf)
1017 {
1018         struct mthca_dev *dev = container_of(cdev, struct mthca_dev, ib_dev.class_dev);
1019         return sprintf(buf, "%d.%d.%d\n", (int) (dev->fw_ver >> 32),
1020                        (int) (dev->fw_ver >> 16) & 0xffff,
1021                        (int) dev->fw_ver & 0xffff);
1022 }
1023
1024 static ssize_t show_hca(struct class_device *cdev, char *buf)
1025 {
1026         struct mthca_dev *dev = container_of(cdev, struct mthca_dev, ib_dev.class_dev);
1027         switch (dev->pdev->device) {
1028         case PCI_DEVICE_ID_MELLANOX_TAVOR:
1029                 return sprintf(buf, "MT23108\n");
1030         case PCI_DEVICE_ID_MELLANOX_ARBEL_COMPAT:
1031                 return sprintf(buf, "MT25208 (MT23108 compat mode)\n");
1032         case PCI_DEVICE_ID_MELLANOX_ARBEL:
1033                 return sprintf(buf, "MT25208\n");
1034         case PCI_DEVICE_ID_MELLANOX_SINAI:
1035         case PCI_DEVICE_ID_MELLANOX_SINAI_OLD:
1036                 return sprintf(buf, "MT25204\n");
1037         default:
1038                 return sprintf(buf, "unknown\n");
1039         }
1040 }
1041
1042 static ssize_t show_board(struct class_device *cdev, char *buf)
1043 {
1044         struct mthca_dev *dev = container_of(cdev, struct mthca_dev, ib_dev.class_dev);
1045         return sprintf(buf, "%.*s\n", MTHCA_BOARD_ID_LEN, dev->board_id);
1046 }
1047
1048 static CLASS_DEVICE_ATTR(hw_rev,   S_IRUGO, show_rev,    NULL);
1049 static CLASS_DEVICE_ATTR(fw_ver,   S_IRUGO, show_fw_ver, NULL);
1050 static CLASS_DEVICE_ATTR(hca_type, S_IRUGO, show_hca,    NULL);
1051 static CLASS_DEVICE_ATTR(board_id, S_IRUGO, show_board,  NULL);
1052
1053 static struct class_device_attribute *mthca_class_attributes[] = {
1054         &class_device_attr_hw_rev,
1055         &class_device_attr_fw_ver,
1056         &class_device_attr_hca_type,
1057         &class_device_attr_board_id
1058 };
1059
1060 static int mthca_init_node_data(struct mthca_dev *dev)
1061 {
1062         struct ib_smp *in_mad  = NULL;
1063         struct ib_smp *out_mad = NULL;
1064         int err = -ENOMEM;
1065         u8 status;
1066
1067         in_mad  = kzalloc(sizeof *in_mad, GFP_KERNEL);
1068         out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
1069         if (!in_mad || !out_mad)
1070                 goto out;
1071
1072         init_query_mad(in_mad);
1073         in_mad->attr_id = IB_SMP_ATTR_NODE_INFO;
1074
1075         err = mthca_MAD_IFC(dev, 1, 1,
1076                             1, NULL, NULL, in_mad, out_mad,
1077                             &status);
1078         if (err)
1079                 goto out;
1080         if (status) {
1081                 err = -EINVAL;
1082                 goto out;
1083         }
1084
1085         memcpy(&dev->ib_dev.node_guid, out_mad->data + 12, 8);
1086
1087 out:
1088         kfree(in_mad);
1089         kfree(out_mad);
1090         return err;
1091 }
1092
1093 int mthca_register_device(struct mthca_dev *dev)
1094 {
1095         int ret;
1096         int i;
1097
1098         ret = mthca_init_node_data(dev);
1099         if (ret)
1100                 return ret;
1101
1102         strlcpy(dev->ib_dev.name, "mthca%d", IB_DEVICE_NAME_MAX);
1103         dev->ib_dev.owner                = THIS_MODULE;
1104
1105         dev->ib_dev.uverbs_abi_ver       = MTHCA_UVERBS_ABI_VERSION;
1106         dev->ib_dev.uverbs_cmd_mask      =
1107                 (1ull << IB_USER_VERBS_CMD_GET_CONTEXT)         |
1108                 (1ull << IB_USER_VERBS_CMD_QUERY_DEVICE)        |
1109                 (1ull << IB_USER_VERBS_CMD_QUERY_PORT)          |
1110                 (1ull << IB_USER_VERBS_CMD_ALLOC_PD)            |
1111                 (1ull << IB_USER_VERBS_CMD_DEALLOC_PD)          |
1112                 (1ull << IB_USER_VERBS_CMD_REG_MR)              |
1113                 (1ull << IB_USER_VERBS_CMD_DEREG_MR)            |
1114                 (1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL) |
1115                 (1ull << IB_USER_VERBS_CMD_CREATE_CQ)           |
1116                 (1ull << IB_USER_VERBS_CMD_DESTROY_CQ)          |
1117                 (1ull << IB_USER_VERBS_CMD_CREATE_QP)           |
1118                 (1ull << IB_USER_VERBS_CMD_MODIFY_QP)           |
1119                 (1ull << IB_USER_VERBS_CMD_DESTROY_QP)          |
1120                 (1ull << IB_USER_VERBS_CMD_ATTACH_MCAST)        |
1121                 (1ull << IB_USER_VERBS_CMD_DETACH_MCAST)        |
1122                 (1ull << IB_USER_VERBS_CMD_CREATE_SRQ)          |
1123                 (1ull << IB_USER_VERBS_CMD_MODIFY_SRQ)          |
1124                 (1ull << IB_USER_VERBS_CMD_DESTROY_SRQ);
1125         dev->ib_dev.node_type            = IB_NODE_CA;
1126         dev->ib_dev.phys_port_cnt        = dev->limits.num_ports;
1127         dev->ib_dev.dma_device           = &dev->pdev->dev;
1128         dev->ib_dev.class_dev.dev        = &dev->pdev->dev;
1129         dev->ib_dev.query_device         = mthca_query_device;
1130         dev->ib_dev.query_port           = mthca_query_port;
1131         dev->ib_dev.modify_port          = mthca_modify_port;
1132         dev->ib_dev.query_pkey           = mthca_query_pkey;
1133         dev->ib_dev.query_gid            = mthca_query_gid;
1134         dev->ib_dev.alloc_ucontext       = mthca_alloc_ucontext;
1135         dev->ib_dev.dealloc_ucontext     = mthca_dealloc_ucontext;
1136         dev->ib_dev.mmap                 = mthca_mmap_uar;
1137         dev->ib_dev.alloc_pd             = mthca_alloc_pd;
1138         dev->ib_dev.dealloc_pd           = mthca_dealloc_pd;
1139         dev->ib_dev.create_ah            = mthca_ah_create;
1140         dev->ib_dev.destroy_ah           = mthca_ah_destroy;
1141
1142         if (dev->mthca_flags & MTHCA_FLAG_SRQ) {
1143                 dev->ib_dev.create_srq           = mthca_create_srq;
1144                 dev->ib_dev.modify_srq           = mthca_modify_srq;
1145                 dev->ib_dev.destroy_srq          = mthca_destroy_srq;
1146
1147                 if (mthca_is_memfree(dev))
1148                         dev->ib_dev.post_srq_recv = mthca_arbel_post_srq_recv;
1149                 else
1150                         dev->ib_dev.post_srq_recv = mthca_tavor_post_srq_recv;
1151         }
1152
1153         dev->ib_dev.create_qp            = mthca_create_qp;
1154         dev->ib_dev.modify_qp            = mthca_modify_qp;
1155         dev->ib_dev.destroy_qp           = mthca_destroy_qp;
1156         dev->ib_dev.create_cq            = mthca_create_cq;
1157         dev->ib_dev.destroy_cq           = mthca_destroy_cq;
1158         dev->ib_dev.poll_cq              = mthca_poll_cq;
1159         dev->ib_dev.get_dma_mr           = mthca_get_dma_mr;
1160         dev->ib_dev.reg_phys_mr          = mthca_reg_phys_mr;
1161         dev->ib_dev.reg_user_mr          = mthca_reg_user_mr;
1162         dev->ib_dev.dereg_mr             = mthca_dereg_mr;
1163
1164         if (dev->mthca_flags & MTHCA_FLAG_FMR) {
1165                 dev->ib_dev.alloc_fmr            = mthca_alloc_fmr;
1166                 dev->ib_dev.unmap_fmr            = mthca_unmap_fmr;
1167                 dev->ib_dev.dealloc_fmr          = mthca_dealloc_fmr;
1168                 if (mthca_is_memfree(dev))
1169                         dev->ib_dev.map_phys_fmr = mthca_arbel_map_phys_fmr;
1170                 else
1171                         dev->ib_dev.map_phys_fmr = mthca_tavor_map_phys_fmr;
1172         }
1173
1174         dev->ib_dev.attach_mcast         = mthca_multicast_attach;
1175         dev->ib_dev.detach_mcast         = mthca_multicast_detach;
1176         dev->ib_dev.process_mad          = mthca_process_mad;
1177
1178         if (mthca_is_memfree(dev)) {
1179                 dev->ib_dev.req_notify_cq = mthca_arbel_arm_cq;
1180                 dev->ib_dev.post_send     = mthca_arbel_post_send;
1181                 dev->ib_dev.post_recv     = mthca_arbel_post_receive;
1182         } else {
1183                 dev->ib_dev.req_notify_cq = mthca_tavor_arm_cq;
1184                 dev->ib_dev.post_send     = mthca_tavor_post_send;
1185                 dev->ib_dev.post_recv     = mthca_tavor_post_receive;
1186         }
1187
1188         init_MUTEX(&dev->cap_mask_mutex);
1189
1190         ret = ib_register_device(&dev->ib_dev);
1191         if (ret)
1192                 return ret;
1193
1194         for (i = 0; i < ARRAY_SIZE(mthca_class_attributes); ++i) {
1195                 ret = class_device_create_file(&dev->ib_dev.class_dev,
1196                                                mthca_class_attributes[i]);
1197                 if (ret) {
1198                         ib_unregister_device(&dev->ib_dev);
1199                         return ret;
1200                 }
1201         }
1202
1203         mthca_start_catas_poll(dev);
1204
1205         return 0;
1206 }
1207
1208 void mthca_unregister_device(struct mthca_dev *dev)
1209 {
1210         mthca_stop_catas_poll(dev);
1211         ib_unregister_device(&dev->ib_dev);
1212 }