]> www.pilppa.org Git - linux-2.6-omap-h63xx.git/blob - drivers/infiniband/hw/mthca/mthca_provider.c
[IB] mthca: report page size capability
[linux-2.6-omap-h63xx.git] / drivers / infiniband / hw / mthca / mthca_provider.c
1 /*
2  * Copyright (c) 2004, 2005 Topspin Communications.  All rights reserved.
3  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
4  * Copyright (c) 2005 Cisco Systems. All rights reserved.
5  * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
6  * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
7  *
8  * This software is available to you under a choice of one of two
9  * licenses.  You may choose to be licensed under the terms of the GNU
10  * General Public License (GPL) Version 2, available from the file
11  * COPYING in the main directory of this source tree, or the
12  * OpenIB.org BSD license below:
13  *
14  *     Redistribution and use in source and binary forms, with or
15  *     without modification, are permitted provided that the following
16  *     conditions are met:
17  *
18  *      - Redistributions of source code must retain the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer.
21  *
22  *      - Redistributions in binary form must reproduce the above
23  *        copyright notice, this list of conditions and the following
24  *        disclaimer in the documentation and/or other materials
25  *        provided with the distribution.
26  *
27  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
28  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
29  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
30  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
31  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
32  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
33  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
34  * SOFTWARE.
35  *
36  * $Id: mthca_provider.c 1397 2004-12-28 05:09:00Z roland $
37  */
38
39 #include <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 int mthca_query_device(struct ib_device *ibdev,
49                               struct ib_device_attr *props)
50 {
51         struct ib_smp *in_mad  = NULL;
52         struct ib_smp *out_mad = NULL;
53         int err = -ENOMEM;
54         struct mthca_dev* mdev = to_mdev(ibdev);
55
56         u8 status;
57
58         in_mad  = kmalloc(sizeof *in_mad, GFP_KERNEL);
59         out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
60         if (!in_mad || !out_mad)
61                 goto out;
62
63         memset(props, 0, sizeof *props);
64
65         props->fw_ver              = mdev->fw_ver;
66
67         memset(in_mad, 0, sizeof *in_mad);
68         in_mad->base_version       = 1;
69         in_mad->mgmt_class         = IB_MGMT_CLASS_SUBN_LID_ROUTED;
70         in_mad->class_version      = 1;
71         in_mad->method             = IB_MGMT_METHOD_GET;
72         in_mad->attr_id            = IB_SMP_ATTR_NODE_INFO;
73
74         err = mthca_MAD_IFC(mdev, 1, 1,
75                             1, NULL, NULL, in_mad, out_mad,
76                             &status);
77         if (err)
78                 goto out;
79         if (status) {
80                 err = -EINVAL;
81                 goto out;
82         }
83
84         props->device_cap_flags    = mdev->device_cap_flags;
85         props->vendor_id           = be32_to_cpup((__be32 *) (out_mad->data + 36)) &
86                 0xffffff;
87         props->vendor_part_id      = be16_to_cpup((__be16 *) (out_mad->data + 30));
88         props->hw_ver              = be32_to_cpup((__be32 *) (out_mad->data + 32));
89         memcpy(&props->sys_image_guid, out_mad->data +  4, 8);
90         memcpy(&props->node_guid,      out_mad->data + 12, 8);
91
92         props->max_mr_size         = ~0ull;
93         props->page_size_cap       = mdev->limits.page_size_cap;
94         props->max_qp              = mdev->limits.num_qps - mdev->limits.reserved_qps;
95         props->max_qp_wr           = mdev->limits.max_wqes;
96         props->max_sge             = mdev->limits.max_sg;
97         props->max_cq              = mdev->limits.num_cqs - mdev->limits.reserved_cqs;
98         props->max_cqe             = mdev->limits.max_cqes;
99         props->max_mr              = mdev->limits.num_mpts - mdev->limits.reserved_mrws;
100         props->max_pd              = mdev->limits.num_pds - mdev->limits.reserved_pds;
101         props->max_qp_rd_atom      = 1 << mdev->qp_table.rdb_shift;
102         props->max_qp_init_rd_atom = mdev->limits.max_qp_init_rdma;
103         props->max_res_rd_atom     = props->max_qp_rd_atom * props->max_qp;
104         props->max_srq             = mdev->limits.num_srqs - mdev->limits.reserved_srqs;
105         props->max_srq_wr          = mdev->limits.max_srq_wqes;
106         props->max_srq_sge         = mdev->limits.max_sg;
107         props->local_ca_ack_delay  = mdev->limits.local_ca_ack_delay;
108         props->atomic_cap          = mdev->limits.flags & DEV_LIM_FLAG_ATOMIC ? 
109                                         IB_ATOMIC_HCA : IB_ATOMIC_NONE;
110         props->max_pkeys           = mdev->limits.pkey_table_len;
111         props->max_mcast_grp       = mdev->limits.num_mgms + mdev->limits.num_amgms;
112         props->max_mcast_qp_attach = MTHCA_QP_PER_MGM;
113         props->max_total_mcast_qp_attach = props->max_mcast_qp_attach * 
114                                            props->max_mcast_grp;
115
116         err = 0;
117  out:
118         kfree(in_mad);
119         kfree(out_mad);
120         return err;
121 }
122
123 static int mthca_query_port(struct ib_device *ibdev,
124                             u8 port, struct ib_port_attr *props)
125 {
126         struct ib_smp *in_mad  = NULL;
127         struct ib_smp *out_mad = NULL;
128         int err = -ENOMEM;
129         u8 status;
130
131         in_mad  = kmalloc(sizeof *in_mad, GFP_KERNEL);
132         out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
133         if (!in_mad || !out_mad)
134                 goto out;
135
136         memset(props, 0, sizeof *props);
137
138         memset(in_mad, 0, sizeof *in_mad);
139         in_mad->base_version       = 1;
140         in_mad->mgmt_class         = IB_MGMT_CLASS_SUBN_LID_ROUTED;
141         in_mad->class_version      = 1;
142         in_mad->method             = IB_MGMT_METHOD_GET;
143         in_mad->attr_id            = IB_SMP_ATTR_PORT_INFO;
144         in_mad->attr_mod           = cpu_to_be32(port);
145
146         err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
147                             port, NULL, NULL, in_mad, out_mad,
148                             &status);
149         if (err)
150                 goto out;
151         if (status) {
152                 err = -EINVAL;
153                 goto out;
154         }
155
156         props->lid               = be16_to_cpup((__be16 *) (out_mad->data + 16));
157         props->lmc               = out_mad->data[34] & 0x7;
158         props->sm_lid            = be16_to_cpup((__be16 *) (out_mad->data + 18));
159         props->sm_sl             = out_mad->data[36] & 0xf;
160         props->state             = out_mad->data[32] & 0xf;
161         props->phys_state        = out_mad->data[33] >> 4;
162         props->port_cap_flags    = be32_to_cpup((__be32 *) (out_mad->data + 20));
163         props->gid_tbl_len       = to_mdev(ibdev)->limits.gid_table_len;
164         props->max_msg_sz        = 0x80000000;
165         props->pkey_tbl_len      = to_mdev(ibdev)->limits.pkey_table_len;
166         props->bad_pkey_cntr     = be16_to_cpup((__be16 *) (out_mad->data + 46));
167         props->qkey_viol_cntr    = be16_to_cpup((__be16 *) (out_mad->data + 48));
168         props->active_width      = out_mad->data[31] & 0xf;
169         props->active_speed      = out_mad->data[35] >> 4;
170         props->max_mtu           = out_mad->data[41] & 0xf;
171         props->active_mtu        = out_mad->data[36] >> 4;
172         props->subnet_timeout    = out_mad->data[51] & 0x1f;
173
174  out:
175         kfree(in_mad);
176         kfree(out_mad);
177         return err;
178 }
179
180 static int mthca_modify_port(struct ib_device *ibdev,
181                              u8 port, int port_modify_mask,
182                              struct ib_port_modify *props)
183 {
184         struct mthca_set_ib_param set_ib;
185         struct ib_port_attr attr;
186         int err;
187         u8 status;
188
189         if (down_interruptible(&to_mdev(ibdev)->cap_mask_mutex))
190                 return -ERESTARTSYS;
191
192         err = mthca_query_port(ibdev, port, &attr);
193         if (err)
194                 goto out;
195
196         set_ib.set_si_guid     = 0;
197         set_ib.reset_qkey_viol = !!(port_modify_mask & IB_PORT_RESET_QKEY_CNTR);
198
199         set_ib.cap_mask = (attr.port_cap_flags | props->set_port_cap_mask) &
200                 ~props->clr_port_cap_mask;
201
202         err = mthca_SET_IB(to_mdev(ibdev), &set_ib, port, &status);
203         if (err)
204                 goto out;
205         if (status) {
206                 err = -EINVAL;
207                 goto out;
208         }
209
210 out:
211         up(&to_mdev(ibdev)->cap_mask_mutex);
212         return err;
213 }
214
215 static int mthca_query_pkey(struct ib_device *ibdev,
216                             u8 port, u16 index, u16 *pkey)
217 {
218         struct ib_smp *in_mad  = NULL;
219         struct ib_smp *out_mad = NULL;
220         int err = -ENOMEM;
221         u8 status;
222
223         in_mad  = kmalloc(sizeof *in_mad, GFP_KERNEL);
224         out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
225         if (!in_mad || !out_mad)
226                 goto out;
227
228         memset(in_mad, 0, sizeof *in_mad);
229         in_mad->base_version       = 1;
230         in_mad->mgmt_class         = IB_MGMT_CLASS_SUBN_LID_ROUTED;
231         in_mad->class_version      = 1;
232         in_mad->method             = IB_MGMT_METHOD_GET;
233         in_mad->attr_id            = IB_SMP_ATTR_PKEY_TABLE;
234         in_mad->attr_mod           = cpu_to_be32(index / 32);
235
236         err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
237                             port, NULL, NULL, in_mad, out_mad,
238                             &status);
239         if (err)
240                 goto out;
241         if (status) {
242                 err = -EINVAL;
243                 goto out;
244         }
245
246         *pkey = be16_to_cpu(((__be16 *) out_mad->data)[index % 32]);
247
248  out:
249         kfree(in_mad);
250         kfree(out_mad);
251         return err;
252 }
253
254 static int mthca_query_gid(struct ib_device *ibdev, u8 port,
255                            int index, union ib_gid *gid)
256 {
257         struct ib_smp *in_mad  = NULL;
258         struct ib_smp *out_mad = NULL;
259         int err = -ENOMEM;
260         u8 status;
261
262         in_mad  = kmalloc(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         memset(in_mad, 0, sizeof *in_mad);
268         in_mad->base_version       = 1;
269         in_mad->mgmt_class         = IB_MGMT_CLASS_SUBN_LID_ROUTED;
270         in_mad->class_version      = 1;
271         in_mad->method             = IB_MGMT_METHOD_GET;
272         in_mad->attr_id            = IB_SMP_ATTR_PORT_INFO;
273         in_mad->attr_mod           = cpu_to_be32(port);
274
275         err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
276                             port, NULL, NULL, in_mad, out_mad,
277                             &status);
278         if (err)
279                 goto out;
280         if (status) {
281                 err = -EINVAL;
282                 goto out;
283         }
284
285         memcpy(gid->raw, out_mad->data + 8, 8);
286
287         memset(in_mad, 0, sizeof *in_mad);
288         in_mad->base_version       = 1;
289         in_mad->mgmt_class         = IB_MGMT_CLASS_SUBN_LID_ROUTED;
290         in_mad->class_version      = 1;
291         in_mad->method             = IB_MGMT_METHOD_GET;
292         in_mad->attr_id            = IB_SMP_ATTR_GUID_INFO;
293         in_mad->attr_mod           = cpu_to_be32(index / 8);
294
295         err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
296                             port, NULL, NULL, in_mad, out_mad,
297                             &status);
298         if (err)
299                 goto out;
300         if (status) {
301                 err = -EINVAL;
302                 goto out;
303         }
304
305         memcpy(gid->raw + 8, out_mad->data + (index % 8) * 16, 8);
306
307  out:
308         kfree(in_mad);
309         kfree(out_mad);
310         return err;
311 }
312
313 static struct ib_ucontext *mthca_alloc_ucontext(struct ib_device *ibdev,
314                                                 struct ib_udata *udata)
315 {
316         struct mthca_alloc_ucontext_resp uresp;
317         struct mthca_ucontext           *context;
318         int                              err;
319
320         memset(&uresp, 0, sizeof uresp);
321
322         uresp.qp_tab_size = to_mdev(ibdev)->limits.num_qps;
323         if (mthca_is_memfree(to_mdev(ibdev)))
324                 uresp.uarc_size = to_mdev(ibdev)->uar_table.uarc_size;
325         else
326                 uresp.uarc_size = 0;
327
328         context = kmalloc(sizeof *context, GFP_KERNEL);
329         if (!context)
330                 return ERR_PTR(-ENOMEM);
331
332         err = mthca_uar_alloc(to_mdev(ibdev), &context->uar);
333         if (err) {
334                 kfree(context);
335                 return ERR_PTR(err);
336         }
337
338         context->db_tab = mthca_init_user_db_tab(to_mdev(ibdev));
339         if (IS_ERR(context->db_tab)) {
340                 err = PTR_ERR(context->db_tab);
341                 mthca_uar_free(to_mdev(ibdev), &context->uar);
342                 kfree(context);
343                 return ERR_PTR(err);
344         }
345
346         if (ib_copy_to_udata(udata, &uresp, sizeof uresp)) {
347                 mthca_cleanup_user_db_tab(to_mdev(ibdev), &context->uar, context->db_tab);
348                 mthca_uar_free(to_mdev(ibdev), &context->uar);
349                 kfree(context);
350                 return ERR_PTR(-EFAULT);
351         }
352
353         return &context->ibucontext;
354 }
355
356 static int mthca_dealloc_ucontext(struct ib_ucontext *context)
357 {
358         mthca_cleanup_user_db_tab(to_mdev(context->device), &to_mucontext(context)->uar,
359                                   to_mucontext(context)->db_tab);
360         mthca_uar_free(to_mdev(context->device), &to_mucontext(context)->uar);
361         kfree(to_mucontext(context));
362
363         return 0;
364 }
365
366 static int mthca_mmap_uar(struct ib_ucontext *context,
367                           struct vm_area_struct *vma)
368 {
369         if (vma->vm_end - vma->vm_start != PAGE_SIZE)
370                 return -EINVAL;
371
372         vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
373
374         if (io_remap_pfn_range(vma, vma->vm_start,
375                                to_mucontext(context)->uar.pfn,
376                                PAGE_SIZE, vma->vm_page_prot))
377                 return -EAGAIN;
378
379         return 0;
380 }
381
382 static struct ib_pd *mthca_alloc_pd(struct ib_device *ibdev,
383                                     struct ib_ucontext *context,
384                                     struct ib_udata *udata)
385 {
386         struct mthca_pd *pd;
387         int err;
388
389         pd = kmalloc(sizeof *pd, GFP_KERNEL);
390         if (!pd)
391                 return ERR_PTR(-ENOMEM);
392
393         err = mthca_pd_alloc(to_mdev(ibdev), !context, pd);
394         if (err) {
395                 kfree(pd);
396                 return ERR_PTR(err);
397         }
398
399         if (context) {
400                 if (ib_copy_to_udata(udata, &pd->pd_num, sizeof (__u32))) {
401                         mthca_pd_free(to_mdev(ibdev), pd);
402                         kfree(pd);
403                         return ERR_PTR(-EFAULT);
404                 }
405         }
406
407         return &pd->ibpd;
408 }
409
410 static int mthca_dealloc_pd(struct ib_pd *pd)
411 {
412         mthca_pd_free(to_mdev(pd->device), to_mpd(pd));
413         kfree(pd);
414
415         return 0;
416 }
417
418 static struct ib_ah *mthca_ah_create(struct ib_pd *pd,
419                                      struct ib_ah_attr *ah_attr)
420 {
421         int err;
422         struct mthca_ah *ah;
423
424         ah = kmalloc(sizeof *ah, GFP_ATOMIC);
425         if (!ah)
426                 return ERR_PTR(-ENOMEM);
427
428         err = mthca_create_ah(to_mdev(pd->device), to_mpd(pd), ah_attr, ah);
429         if (err) {
430                 kfree(ah);
431                 return ERR_PTR(err);
432         }
433
434         return &ah->ibah;
435 }
436
437 static int mthca_ah_destroy(struct ib_ah *ah)
438 {
439         mthca_destroy_ah(to_mdev(ah->device), to_mah(ah));
440         kfree(ah);
441
442         return 0;
443 }
444
445 static struct ib_srq *mthca_create_srq(struct ib_pd *pd,
446                                        struct ib_srq_init_attr *init_attr,
447                                        struct ib_udata *udata)
448 {
449         struct mthca_create_srq ucmd;
450         struct mthca_ucontext *context = NULL;
451         struct mthca_srq *srq;
452         int err;
453
454         srq = kmalloc(sizeof *srq, GFP_KERNEL);
455         if (!srq)
456                 return ERR_PTR(-ENOMEM);
457
458         if (pd->uobject) {
459                 context = to_mucontext(pd->uobject->context);
460
461                 if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd))
462                         return ERR_PTR(-EFAULT);
463
464                 err = mthca_map_user_db(to_mdev(pd->device), &context->uar,
465                                         context->db_tab, ucmd.db_index,
466                                         ucmd.db_page);
467
468                 if (err)
469                         goto err_free;
470
471                 srq->mr.ibmr.lkey = ucmd.lkey;
472                 srq->db_index     = ucmd.db_index;
473         }
474
475         err = mthca_alloc_srq(to_mdev(pd->device), to_mpd(pd),
476                               &init_attr->attr, srq);
477
478         if (err && pd->uobject)
479                 mthca_unmap_user_db(to_mdev(pd->device), &context->uar,
480                                     context->db_tab, ucmd.db_index);
481
482         if (err)
483                 goto err_free;
484
485         if (context && ib_copy_to_udata(udata, &srq->srqn, sizeof (__u32))) {
486                 mthca_free_srq(to_mdev(pd->device), srq);
487                 err = -EFAULT;
488                 goto err_free;
489         }
490
491         return &srq->ibsrq;
492
493 err_free:
494         kfree(srq);
495
496         return ERR_PTR(err);
497 }
498
499 static int mthca_destroy_srq(struct ib_srq *srq)
500 {
501         struct mthca_ucontext *context;
502
503         if (srq->uobject) {
504                 context = to_mucontext(srq->uobject->context);
505
506                 mthca_unmap_user_db(to_mdev(srq->device), &context->uar,
507                                     context->db_tab, to_msrq(srq)->db_index);
508         }
509
510         mthca_free_srq(to_mdev(srq->device), to_msrq(srq));
511         kfree(srq);
512
513         return 0;
514 }
515
516 static struct ib_qp *mthca_create_qp(struct ib_pd *pd,
517                                      struct ib_qp_init_attr *init_attr,
518                                      struct ib_udata *udata)
519 {
520         struct mthca_create_qp ucmd;
521         struct mthca_qp *qp;
522         int err;
523
524         switch (init_attr->qp_type) {
525         case IB_QPT_RC:
526         case IB_QPT_UC:
527         case IB_QPT_UD:
528         {
529                 struct mthca_ucontext *context;
530
531                 qp = kmalloc(sizeof *qp, GFP_KERNEL);
532                 if (!qp)
533                         return ERR_PTR(-ENOMEM);
534
535                 if (pd->uobject) {
536                         context = to_mucontext(pd->uobject->context);
537
538                         if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd))
539                                 return ERR_PTR(-EFAULT);
540
541                         err = mthca_map_user_db(to_mdev(pd->device), &context->uar,
542                                                 context->db_tab,
543                                                 ucmd.sq_db_index, ucmd.sq_db_page);
544                         if (err) {
545                                 kfree(qp);
546                                 return ERR_PTR(err);
547                         }
548
549                         err = mthca_map_user_db(to_mdev(pd->device), &context->uar,
550                                                 context->db_tab,
551                                                 ucmd.rq_db_index, ucmd.rq_db_page);
552                         if (err) {
553                                 mthca_unmap_user_db(to_mdev(pd->device),
554                                                     &context->uar,
555                                                     context->db_tab,
556                                                     ucmd.sq_db_index);
557                                 kfree(qp);
558                                 return ERR_PTR(err);
559                         }
560
561                         qp->mr.ibmr.lkey = ucmd.lkey;
562                         qp->sq.db_index  = ucmd.sq_db_index;
563                         qp->rq.db_index  = ucmd.rq_db_index;
564                 }
565
566                 err = mthca_alloc_qp(to_mdev(pd->device), to_mpd(pd),
567                                      to_mcq(init_attr->send_cq),
568                                      to_mcq(init_attr->recv_cq),
569                                      init_attr->qp_type, init_attr->sq_sig_type,
570                                      &init_attr->cap, qp);
571
572                 if (err && pd->uobject) {
573                         context = to_mucontext(pd->uobject->context);
574
575                         mthca_unmap_user_db(to_mdev(pd->device),
576                                             &context->uar,
577                                             context->db_tab,
578                                             ucmd.sq_db_index);
579                         mthca_unmap_user_db(to_mdev(pd->device),
580                                             &context->uar,
581                                             context->db_tab,
582                                             ucmd.rq_db_index);
583                 }
584
585                 qp->ibqp.qp_num = qp->qpn;
586                 break;
587         }
588         case IB_QPT_SMI:
589         case IB_QPT_GSI:
590         {
591                 /* Don't allow userspace to create special QPs */
592                 if (pd->uobject)
593                         return ERR_PTR(-EINVAL);
594
595                 qp = kmalloc(sizeof (struct mthca_sqp), GFP_KERNEL);
596                 if (!qp)
597                         return ERR_PTR(-ENOMEM);
598
599                 qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 : 1;
600
601                 err = mthca_alloc_sqp(to_mdev(pd->device), to_mpd(pd),
602                                       to_mcq(init_attr->send_cq),
603                                       to_mcq(init_attr->recv_cq),
604                                       init_attr->sq_sig_type, &init_attr->cap,
605                                       qp->ibqp.qp_num, init_attr->port_num,
606                                       to_msqp(qp));
607                 break;
608         }
609         default:
610                 /* Don't support raw QPs */
611                 return ERR_PTR(-ENOSYS);
612         }
613
614         if (err) {
615                 kfree(qp);
616                 return ERR_PTR(err);
617         }
618
619         init_attr->cap.max_inline_data = 0;
620         init_attr->cap.max_send_wr     = qp->sq.max;
621         init_attr->cap.max_recv_wr     = qp->rq.max;
622         init_attr->cap.max_send_sge    = qp->sq.max_gs;
623         init_attr->cap.max_recv_sge    = qp->rq.max_gs;
624
625         return &qp->ibqp;
626 }
627
628 static int mthca_destroy_qp(struct ib_qp *qp)
629 {
630         if (qp->uobject) {
631                 mthca_unmap_user_db(to_mdev(qp->device),
632                                     &to_mucontext(qp->uobject->context)->uar,
633                                     to_mucontext(qp->uobject->context)->db_tab,
634                                     to_mqp(qp)->sq.db_index);
635                 mthca_unmap_user_db(to_mdev(qp->device),
636                                     &to_mucontext(qp->uobject->context)->uar,
637                                     to_mucontext(qp->uobject->context)->db_tab,
638                                     to_mqp(qp)->rq.db_index);
639         }
640         mthca_free_qp(to_mdev(qp->device), to_mqp(qp));
641         kfree(qp);
642         return 0;
643 }
644
645 static struct ib_cq *mthca_create_cq(struct ib_device *ibdev, int entries,
646                                      struct ib_ucontext *context,
647                                      struct ib_udata *udata)
648 {
649         struct mthca_create_cq ucmd;
650         struct mthca_cq *cq;
651         int nent;
652         int err;
653
654         if (entries < 1 || entries > to_mdev(ibdev)->limits.max_cqes)
655                 return ERR_PTR(-EINVAL);
656
657         if (context) {
658                 if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd))
659                         return ERR_PTR(-EFAULT);
660
661                 err = mthca_map_user_db(to_mdev(ibdev), &to_mucontext(context)->uar,
662                                         to_mucontext(context)->db_tab,
663                                         ucmd.set_db_index, ucmd.set_db_page);
664                 if (err)
665                         return ERR_PTR(err);
666
667                 err = mthca_map_user_db(to_mdev(ibdev), &to_mucontext(context)->uar,
668                                         to_mucontext(context)->db_tab,
669                                         ucmd.arm_db_index, ucmd.arm_db_page);
670                 if (err)
671                         goto err_unmap_set;
672         }
673
674         cq = kmalloc(sizeof *cq, GFP_KERNEL);
675         if (!cq) {
676                 err = -ENOMEM;
677                 goto err_unmap_arm;
678         }
679
680         if (context) {
681                 cq->mr.ibmr.lkey    = ucmd.lkey;
682                 cq->set_ci_db_index = ucmd.set_db_index;
683                 cq->arm_db_index    = ucmd.arm_db_index;
684         }
685
686         for (nent = 1; nent <= entries; nent <<= 1)
687                 ; /* nothing */
688
689         err = mthca_init_cq(to_mdev(ibdev), nent,
690                             context ? to_mucontext(context) : NULL,
691                             context ? ucmd.pdn : to_mdev(ibdev)->driver_pd.pd_num,
692                             cq);
693         if (err)
694                 goto err_free;
695
696         if (context && ib_copy_to_udata(udata, &cq->cqn, sizeof (__u32))) {
697                 mthca_free_cq(to_mdev(ibdev), cq);
698                 goto err_free;
699         }
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_destroy_cq(struct ib_cq *cq)
720 {
721         if (cq->uobject) {
722                 mthca_unmap_user_db(to_mdev(cq->device),
723                                     &to_mucontext(cq->uobject->context)->uar,
724                                     to_mucontext(cq->uobject->context)->db_tab,
725                                     to_mcq(cq)->arm_db_index);
726                 mthca_unmap_user_db(to_mdev(cq->device),
727                                     &to_mucontext(cq->uobject->context)->uar,
728                                     to_mucontext(cq->uobject->context)->db_tab,
729                                     to_mcq(cq)->set_ci_db_index);
730         }
731         mthca_free_cq(to_mdev(cq->device), to_mcq(cq));
732         kfree(cq);
733
734         return 0;
735 }
736
737 static inline u32 convert_access(int acc)
738 {
739         return (acc & IB_ACCESS_REMOTE_ATOMIC ? MTHCA_MPT_FLAG_ATOMIC       : 0) |
740                (acc & IB_ACCESS_REMOTE_WRITE  ? MTHCA_MPT_FLAG_REMOTE_WRITE : 0) |
741                (acc & IB_ACCESS_REMOTE_READ   ? MTHCA_MPT_FLAG_REMOTE_READ  : 0) |
742                (acc & IB_ACCESS_LOCAL_WRITE   ? MTHCA_MPT_FLAG_LOCAL_WRITE  : 0) |
743                MTHCA_MPT_FLAG_LOCAL_READ;
744 }
745
746 static struct ib_mr *mthca_get_dma_mr(struct ib_pd *pd, int acc)
747 {
748         struct mthca_mr *mr;
749         int err;
750
751         mr = kmalloc(sizeof *mr, GFP_KERNEL);
752         if (!mr)
753                 return ERR_PTR(-ENOMEM);
754
755         err = mthca_mr_alloc_notrans(to_mdev(pd->device),
756                                      to_mpd(pd)->pd_num,
757                                      convert_access(acc), mr);
758
759         if (err) {
760                 kfree(mr);
761                 return ERR_PTR(err);
762         }
763
764         return &mr->ibmr;
765 }
766
767 static struct ib_mr *mthca_reg_phys_mr(struct ib_pd       *pd,
768                                        struct ib_phys_buf *buffer_list,
769                                        int                 num_phys_buf,
770                                        int                 acc,
771                                        u64                *iova_start)
772 {
773         struct mthca_mr *mr;
774         u64 *page_list;
775         u64 total_size;
776         u64 mask;
777         int shift;
778         int npages;
779         int err;
780         int i, j, n;
781
782         /* First check that we have enough alignment */
783         if ((*iova_start & ~PAGE_MASK) != (buffer_list[0].addr & ~PAGE_MASK))
784                 return ERR_PTR(-EINVAL);
785
786         if (num_phys_buf > 1 &&
787             ((buffer_list[0].addr + buffer_list[0].size) & ~PAGE_MASK))
788                 return ERR_PTR(-EINVAL);
789
790         mask = 0;
791         total_size = 0;
792         for (i = 0; i < num_phys_buf; ++i) {
793                 if (i != 0 && buffer_list[i].addr & ~PAGE_MASK)
794                         return ERR_PTR(-EINVAL);
795                 if (i != 0 && i != num_phys_buf - 1 &&
796                     (buffer_list[i].size & ~PAGE_MASK))
797                         return ERR_PTR(-EINVAL);
798
799                 total_size += buffer_list[i].size;
800                 if (i > 0)
801                         mask |= buffer_list[i].addr;
802         }
803
804         /* Find largest page shift we can use to cover buffers */
805         for (shift = PAGE_SHIFT; shift < 31; ++shift)
806                 if (num_phys_buf > 1) {
807                         if ((1ULL << shift) & mask)
808                                 break;
809                 } else {
810                         if (1ULL << shift >=
811                             buffer_list[0].size +
812                             (buffer_list[0].addr & ((1ULL << shift) - 1)))
813                                 break;
814                 }
815
816         buffer_list[0].size += buffer_list[0].addr & ((1ULL << shift) - 1);
817         buffer_list[0].addr &= ~0ull << shift;
818
819         mr = kmalloc(sizeof *mr, GFP_KERNEL);
820         if (!mr)
821                 return ERR_PTR(-ENOMEM);
822
823         npages = 0;
824         for (i = 0; i < num_phys_buf; ++i)
825                 npages += (buffer_list[i].size + (1ULL << shift) - 1) >> shift;
826
827         if (!npages)
828                 return &mr->ibmr;
829
830         page_list = kmalloc(npages * sizeof *page_list, GFP_KERNEL);
831         if (!page_list) {
832                 kfree(mr);
833                 return ERR_PTR(-ENOMEM);
834         }
835
836         n = 0;
837         for (i = 0; i < num_phys_buf; ++i)
838                 for (j = 0;
839                      j < (buffer_list[i].size + (1ULL << shift) - 1) >> shift;
840                      ++j)
841                         page_list[n++] = buffer_list[i].addr + ((u64) j << shift);
842
843         mthca_dbg(to_mdev(pd->device), "Registering memory at %llx (iova %llx) "
844                   "in PD %x; shift %d, npages %d.\n",
845                   (unsigned long long) buffer_list[0].addr,
846                   (unsigned long long) *iova_start,
847                   to_mpd(pd)->pd_num,
848                   shift, npages);
849
850         err = mthca_mr_alloc_phys(to_mdev(pd->device),
851                                   to_mpd(pd)->pd_num,
852                                   page_list, shift, npages,
853                                   *iova_start, total_size,
854                                   convert_access(acc), mr);
855
856         if (err) {
857                 kfree(page_list);
858                 kfree(mr);
859                 return ERR_PTR(err);
860         }
861
862         kfree(page_list);
863         return &mr->ibmr;
864 }
865
866 static struct ib_mr *mthca_reg_user_mr(struct ib_pd *pd, struct ib_umem *region,
867                                        int acc, struct ib_udata *udata)
868 {
869         struct mthca_dev *dev = to_mdev(pd->device);
870         struct ib_umem_chunk *chunk;
871         struct mthca_mr *mr;
872         u64 *pages;
873         int shift, n, len;
874         int i, j, k;
875         int err = 0;
876
877         shift = ffs(region->page_size) - 1;
878
879         mr = kmalloc(sizeof *mr, GFP_KERNEL);
880         if (!mr)
881                 return ERR_PTR(-ENOMEM);
882
883         n = 0;
884         list_for_each_entry(chunk, &region->chunk_list, list)
885                 n += chunk->nents;
886
887         mr->mtt = mthca_alloc_mtt(dev, n);
888         if (IS_ERR(mr->mtt)) {
889                 err = PTR_ERR(mr->mtt);
890                 goto err;
891         }
892
893         pages = (u64 *) __get_free_page(GFP_KERNEL);
894         if (!pages) {
895                 err = -ENOMEM;
896                 goto err_mtt;
897         }
898
899         i = n = 0;
900
901         list_for_each_entry(chunk, &region->chunk_list, list)
902                 for (j = 0; j < chunk->nmap; ++j) {
903                         len = sg_dma_len(&chunk->page_list[j]) >> shift;
904                         for (k = 0; k < len; ++k) {
905                                 pages[i++] = sg_dma_address(&chunk->page_list[j]) +
906                                         region->page_size * k;
907                                 /*
908                                  * Be friendly to WRITE_MTT command
909                                  * and leave two empty slots for the
910                                  * index and reserved fields of the
911                                  * mailbox.
912                                  */
913                                 if (i == PAGE_SIZE / sizeof (u64) - 2) {
914                                         err = mthca_write_mtt(dev, mr->mtt,
915                                                               n, pages, i);
916                                         if (err)
917                                                 goto mtt_done;
918                                         n += i;
919                                         i = 0;
920                                 }
921                         }
922                 }
923
924         if (i)
925                 err = mthca_write_mtt(dev, mr->mtt, n, pages, i);
926 mtt_done:
927         free_page((unsigned long) pages);
928         if (err)
929                 goto err_mtt;
930
931         err = mthca_mr_alloc(dev, to_mpd(pd)->pd_num, shift, region->virt_base,
932                              region->length, convert_access(acc), mr);
933
934         if (err)
935                 goto err_mtt;
936
937         return &mr->ibmr;
938
939 err_mtt:
940         mthca_free_mtt(dev, mr->mtt);
941
942 err:
943         kfree(mr);
944         return ERR_PTR(err);
945 }
946
947 static int mthca_dereg_mr(struct ib_mr *mr)
948 {
949         struct mthca_mr *mmr = to_mmr(mr);
950         mthca_free_mr(to_mdev(mr->device), mmr);
951         kfree(mmr);
952         return 0;
953 }
954
955 static struct ib_fmr *mthca_alloc_fmr(struct ib_pd *pd, int mr_access_flags,
956                                       struct ib_fmr_attr *fmr_attr)
957 {
958         struct mthca_fmr *fmr;
959         int err;
960
961         fmr = kmalloc(sizeof *fmr, GFP_KERNEL);
962         if (!fmr)
963                 return ERR_PTR(-ENOMEM);
964
965         memcpy(&fmr->attr, fmr_attr, sizeof *fmr_attr);
966         err = mthca_fmr_alloc(to_mdev(pd->device), to_mpd(pd)->pd_num,
967                              convert_access(mr_access_flags), fmr);
968
969         if (err) {
970                 kfree(fmr);
971                 return ERR_PTR(err);
972         }
973
974         return &fmr->ibmr;
975 }
976
977 static int mthca_dealloc_fmr(struct ib_fmr *fmr)
978 {
979         struct mthca_fmr *mfmr = to_mfmr(fmr);
980         int err;
981
982         err = mthca_free_fmr(to_mdev(fmr->device), mfmr);
983         if (err)
984                 return err;
985
986         kfree(mfmr);
987         return 0;
988 }
989
990 static int mthca_unmap_fmr(struct list_head *fmr_list)
991 {
992         struct ib_fmr *fmr;
993         int err;
994         u8 status;
995         struct mthca_dev *mdev = NULL;
996
997         list_for_each_entry(fmr, fmr_list, list) {
998                 if (mdev && to_mdev(fmr->device) != mdev)
999                         return -EINVAL;
1000                 mdev = to_mdev(fmr->device);
1001         }
1002
1003         if (!mdev)
1004                 return 0;
1005
1006         if (mthca_is_memfree(mdev)) {
1007                 list_for_each_entry(fmr, fmr_list, list)
1008                         mthca_arbel_fmr_unmap(mdev, to_mfmr(fmr));
1009
1010                 wmb();
1011         } else
1012                 list_for_each_entry(fmr, fmr_list, list)
1013                         mthca_tavor_fmr_unmap(mdev, to_mfmr(fmr));
1014
1015         err = mthca_SYNC_TPT(mdev, &status);
1016         if (err)
1017                 return err;
1018         if (status)
1019                 return -EINVAL;
1020         return 0;
1021 }
1022
1023 static ssize_t show_rev(struct class_device *cdev, char *buf)
1024 {
1025         struct mthca_dev *dev = container_of(cdev, struct mthca_dev, ib_dev.class_dev);
1026         return sprintf(buf, "%x\n", dev->rev_id);
1027 }
1028
1029 static ssize_t show_fw_ver(struct class_device *cdev, char *buf)
1030 {
1031         struct mthca_dev *dev = container_of(cdev, struct mthca_dev, ib_dev.class_dev);
1032         return sprintf(buf, "%d.%d.%d\n", (int) (dev->fw_ver >> 32),
1033                        (int) (dev->fw_ver >> 16) & 0xffff,
1034                        (int) dev->fw_ver & 0xffff);
1035 }
1036
1037 static ssize_t show_hca(struct class_device *cdev, char *buf)
1038 {
1039         struct mthca_dev *dev = container_of(cdev, struct mthca_dev, ib_dev.class_dev);
1040         switch (dev->pdev->device) {
1041         case PCI_DEVICE_ID_MELLANOX_TAVOR:
1042                 return sprintf(buf, "MT23108\n");
1043         case PCI_DEVICE_ID_MELLANOX_ARBEL_COMPAT:
1044                 return sprintf(buf, "MT25208 (MT23108 compat mode)\n");
1045         case PCI_DEVICE_ID_MELLANOX_ARBEL:
1046                 return sprintf(buf, "MT25208\n");
1047         case PCI_DEVICE_ID_MELLANOX_SINAI:
1048         case PCI_DEVICE_ID_MELLANOX_SINAI_OLD:
1049                 return sprintf(buf, "MT25204\n");
1050         default:
1051                 return sprintf(buf, "unknown\n");
1052         }
1053 }
1054
1055 static ssize_t show_board(struct class_device *cdev, char *buf)
1056 {
1057         struct mthca_dev *dev = container_of(cdev, struct mthca_dev, ib_dev.class_dev);
1058         return sprintf(buf, "%.*s\n", MTHCA_BOARD_ID_LEN, dev->board_id);
1059 }
1060
1061 static CLASS_DEVICE_ATTR(hw_rev,   S_IRUGO, show_rev,    NULL);
1062 static CLASS_DEVICE_ATTR(fw_ver,   S_IRUGO, show_fw_ver, NULL);
1063 static CLASS_DEVICE_ATTR(hca_type, S_IRUGO, show_hca,    NULL);
1064 static CLASS_DEVICE_ATTR(board_id, S_IRUGO, show_board,  NULL);
1065
1066 static struct class_device_attribute *mthca_class_attributes[] = {
1067         &class_device_attr_hw_rev,
1068         &class_device_attr_fw_ver,
1069         &class_device_attr_hca_type,
1070         &class_device_attr_board_id
1071 };
1072
1073 int mthca_register_device(struct mthca_dev *dev)
1074 {
1075         int ret;
1076         int i;
1077
1078         strlcpy(dev->ib_dev.name, "mthca%d", IB_DEVICE_NAME_MAX);
1079         dev->ib_dev.owner                = THIS_MODULE;
1080
1081         dev->ib_dev.uverbs_abi_ver       = MTHCA_UVERBS_ABI_VERSION;
1082         dev->ib_dev.uverbs_cmd_mask      =
1083                 (1ull << IB_USER_VERBS_CMD_GET_CONTEXT)         |
1084                 (1ull << IB_USER_VERBS_CMD_QUERY_DEVICE)        |
1085                 (1ull << IB_USER_VERBS_CMD_QUERY_PORT)          |
1086                 (1ull << IB_USER_VERBS_CMD_ALLOC_PD)            |
1087                 (1ull << IB_USER_VERBS_CMD_DEALLOC_PD)          |
1088                 (1ull << IB_USER_VERBS_CMD_REG_MR)              |
1089                 (1ull << IB_USER_VERBS_CMD_DEREG_MR)            |
1090                 (1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL) |
1091                 (1ull << IB_USER_VERBS_CMD_CREATE_CQ)           |
1092                 (1ull << IB_USER_VERBS_CMD_DESTROY_CQ)          |
1093                 (1ull << IB_USER_VERBS_CMD_CREATE_QP)           |
1094                 (1ull << IB_USER_VERBS_CMD_MODIFY_QP)           |
1095                 (1ull << IB_USER_VERBS_CMD_DESTROY_QP)          |
1096                 (1ull << IB_USER_VERBS_CMD_ATTACH_MCAST)        |
1097                 (1ull << IB_USER_VERBS_CMD_DETACH_MCAST)        |
1098                 (1ull << IB_USER_VERBS_CMD_CREATE_SRQ)          |
1099                 (1ull << IB_USER_VERBS_CMD_MODIFY_SRQ)          |
1100                 (1ull << IB_USER_VERBS_CMD_DESTROY_SRQ);
1101         dev->ib_dev.node_type            = IB_NODE_CA;
1102         dev->ib_dev.phys_port_cnt        = dev->limits.num_ports;
1103         dev->ib_dev.dma_device           = &dev->pdev->dev;
1104         dev->ib_dev.class_dev.dev        = &dev->pdev->dev;
1105         dev->ib_dev.query_device         = mthca_query_device;
1106         dev->ib_dev.query_port           = mthca_query_port;
1107         dev->ib_dev.modify_port          = mthca_modify_port;
1108         dev->ib_dev.query_pkey           = mthca_query_pkey;
1109         dev->ib_dev.query_gid            = mthca_query_gid;
1110         dev->ib_dev.alloc_ucontext       = mthca_alloc_ucontext;
1111         dev->ib_dev.dealloc_ucontext     = mthca_dealloc_ucontext;
1112         dev->ib_dev.mmap                 = mthca_mmap_uar;
1113         dev->ib_dev.alloc_pd             = mthca_alloc_pd;
1114         dev->ib_dev.dealloc_pd           = mthca_dealloc_pd;
1115         dev->ib_dev.create_ah            = mthca_ah_create;
1116         dev->ib_dev.destroy_ah           = mthca_ah_destroy;
1117
1118         if (dev->mthca_flags & MTHCA_FLAG_SRQ) {
1119                 dev->ib_dev.create_srq           = mthca_create_srq;
1120                 dev->ib_dev.modify_srq           = mthca_modify_srq;
1121                 dev->ib_dev.destroy_srq          = mthca_destroy_srq;
1122
1123                 if (mthca_is_memfree(dev))
1124                         dev->ib_dev.post_srq_recv = mthca_arbel_post_srq_recv;
1125                 else
1126                         dev->ib_dev.post_srq_recv = mthca_tavor_post_srq_recv;
1127         }
1128
1129         dev->ib_dev.create_qp            = mthca_create_qp;
1130         dev->ib_dev.modify_qp            = mthca_modify_qp;
1131         dev->ib_dev.destroy_qp           = mthca_destroy_qp;
1132         dev->ib_dev.create_cq            = mthca_create_cq;
1133         dev->ib_dev.destroy_cq           = mthca_destroy_cq;
1134         dev->ib_dev.poll_cq              = mthca_poll_cq;
1135         dev->ib_dev.get_dma_mr           = mthca_get_dma_mr;
1136         dev->ib_dev.reg_phys_mr          = mthca_reg_phys_mr;
1137         dev->ib_dev.reg_user_mr          = mthca_reg_user_mr;
1138         dev->ib_dev.dereg_mr             = mthca_dereg_mr;
1139
1140         if (dev->mthca_flags & MTHCA_FLAG_FMR) {
1141                 dev->ib_dev.alloc_fmr            = mthca_alloc_fmr;
1142                 dev->ib_dev.unmap_fmr            = mthca_unmap_fmr;
1143                 dev->ib_dev.dealloc_fmr          = mthca_dealloc_fmr;
1144                 if (mthca_is_memfree(dev))
1145                         dev->ib_dev.map_phys_fmr = mthca_arbel_map_phys_fmr;
1146                 else
1147                         dev->ib_dev.map_phys_fmr = mthca_tavor_map_phys_fmr;
1148         }
1149
1150         dev->ib_dev.attach_mcast         = mthca_multicast_attach;
1151         dev->ib_dev.detach_mcast         = mthca_multicast_detach;
1152         dev->ib_dev.process_mad          = mthca_process_mad;
1153
1154         if (mthca_is_memfree(dev)) {
1155                 dev->ib_dev.req_notify_cq = mthca_arbel_arm_cq;
1156                 dev->ib_dev.post_send     = mthca_arbel_post_send;
1157                 dev->ib_dev.post_recv     = mthca_arbel_post_receive;
1158         } else {
1159                 dev->ib_dev.req_notify_cq = mthca_tavor_arm_cq;
1160                 dev->ib_dev.post_send     = mthca_tavor_post_send;
1161                 dev->ib_dev.post_recv     = mthca_tavor_post_receive;
1162         }
1163
1164         init_MUTEX(&dev->cap_mask_mutex);
1165
1166         ret = ib_register_device(&dev->ib_dev);
1167         if (ret)
1168                 return ret;
1169
1170         for (i = 0; i < ARRAY_SIZE(mthca_class_attributes); ++i) {
1171                 ret = class_device_create_file(&dev->ib_dev.class_dev,
1172                                                mthca_class_attributes[i]);
1173                 if (ret) {
1174                         ib_unregister_device(&dev->ib_dev);
1175                         return ret;
1176                 }
1177         }
1178
1179         mthca_start_catas_poll(dev);
1180
1181         return 0;
1182 }
1183
1184 void mthca_unregister_device(struct mthca_dev *dev)
1185 {
1186         mthca_stop_catas_poll(dev);
1187         ib_unregister_device(&dev->ib_dev);
1188 }