1/*
2 * Copyright (c) 2004, 2005 Topspin Communications.  All rights reserved.
3 * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
4 * Copyright (c) 2005, 2006 Cisco Systems.  All rights reserved.
5 * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
6 * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
7 *
8 * This software is available to you under a choice of one of two
9 * licenses.  You may choose to be licensed under the terms of the GNU
10 * General Public License (GPL) Version 2, available from the file
11 * COPYING in the main directory of this source tree, or the
12 * OpenIB.org BSD license below:
13 *
14 *     Redistribution and use in source and binary forms, with or
15 *     without modification, are permitted provided that the following
16 *     conditions are met:
17 *
18 *      - Redistributions of source code must retain the above
19 *        copyright notice, this list of conditions and the following
20 *        disclaimer.
21 *
22 *      - Redistributions in binary form must reproduce the above
23 *        copyright notice, this list of conditions and the following
24 *        disclaimer in the documentation and/or other materials
25 *        provided with the distribution.
26 *
27 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
28 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
29 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
30 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
31 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
32 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
33 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
34 * SOFTWARE.
35 *
36 * $Id: mthca_provider.c,v 1.1.1.1 2007/08/03 18:52:32 Exp $
37 */
38
39#include <rdma/ib_smi.h>
40#include <rdma/ib_umem.h>
41#include <rdma/ib_user_verbs.h>
42#include <linux/mm.h>
43
44#include "mthca_dev.h"
45#include "mthca_cmd.h"
46#include "mthca_user.h"
47#include "mthca_memfree.h"
48
49static void init_query_mad(struct ib_smp *mad)
50{
51	mad->base_version  = 1;
52	mad->mgmt_class    = IB_MGMT_CLASS_SUBN_LID_ROUTED;
53	mad->class_version = 1;
54	mad->method    	   = IB_MGMT_METHOD_GET;
55}
56
57static int mthca_query_device(struct ib_device *ibdev,
58			      struct ib_device_attr *props)
59{
60	struct ib_smp *in_mad  = NULL;
61	struct ib_smp *out_mad = NULL;
62	int err = -ENOMEM;
63	struct mthca_dev* mdev = to_mdev(ibdev);
64
65	u8 status;
66
67	in_mad  = kzalloc(sizeof *in_mad, GFP_KERNEL);
68	out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
69	if (!in_mad || !out_mad)
70		goto out;
71
72	memset(props, 0, sizeof *props);
73
74	props->fw_ver              = mdev->fw_ver;
75
76	init_query_mad(in_mad);
77	in_mad->attr_id = IB_SMP_ATTR_NODE_INFO;
78
79	err = mthca_MAD_IFC(mdev, 1, 1,
80			    1, NULL, NULL, in_mad, out_mad,
81			    &status);
82	if (err)
83		goto out;
84	if (status) {
85		err = -EINVAL;
86		goto out;
87	}
88
89	props->device_cap_flags    = mdev->device_cap_flags;
90	props->vendor_id           = be32_to_cpup((__be32 *) (out_mad->data + 36)) &
91		0xffffff;
92	props->vendor_part_id      = be16_to_cpup((__be16 *) (out_mad->data + 30));
93	props->hw_ver              = be32_to_cpup((__be32 *) (out_mad->data + 32));
94	memcpy(&props->sys_image_guid, out_mad->data +  4, 8);
95
96	props->max_mr_size         = ~0ull;
97	props->page_size_cap       = mdev->limits.page_size_cap;
98	props->max_qp              = mdev->limits.num_qps - mdev->limits.reserved_qps;
99	props->max_qp_wr           = mdev->limits.max_wqes;
100	props->max_sge             = mdev->limits.max_sg;
101	props->max_cq              = mdev->limits.num_cqs - mdev->limits.reserved_cqs;
102	props->max_cqe             = mdev->limits.max_cqes;
103	props->max_mr              = mdev->limits.num_mpts - mdev->limits.reserved_mrws;
104	props->max_pd              = mdev->limits.num_pds - mdev->limits.reserved_pds;
105	props->max_qp_rd_atom      = 1 << mdev->qp_table.rdb_shift;
106	props->max_qp_init_rd_atom = mdev->limits.max_qp_init_rdma;
107	props->max_res_rd_atom     = props->max_qp_rd_atom * props->max_qp;
108	props->max_srq             = mdev->limits.num_srqs - mdev->limits.reserved_srqs;
109	props->max_srq_wr          = mdev->limits.max_srq_wqes;
110	props->max_srq_sge         = mdev->limits.max_srq_sge;
111	props->local_ca_ack_delay  = mdev->limits.local_ca_ack_delay;
112	props->atomic_cap          = mdev->limits.flags & DEV_LIM_FLAG_ATOMIC ?
113					IB_ATOMIC_HCA : IB_ATOMIC_NONE;
114	props->max_pkeys           = mdev->limits.pkey_table_len;
115	props->max_mcast_grp       = mdev->limits.num_mgms + mdev->limits.num_amgms;
116	props->max_mcast_qp_attach = MTHCA_QP_PER_MGM;
117	props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
118					   props->max_mcast_grp;
119	/*
120	 * If Sinai memory key optimization is being used, then only
121	 * the 8-bit key portion will change.  For other HCAs, the
122	 * unused index bits will also be used for FMR remapping.
123	 */
124	if (mdev->mthca_flags & MTHCA_FLAG_SINAI_OPT)
125		props->max_map_per_fmr = 255;
126	else
127		props->max_map_per_fmr =
128			(1 << (32 - ilog2(mdev->limits.num_mpts))) - 1;
129
130	err = 0;
131 out:
132	kfree(in_mad);
133	kfree(out_mad);
134	return err;
135}
136
137static int mthca_query_port(struct ib_device *ibdev,
138			    u8 port, struct ib_port_attr *props)
139{
140	struct ib_smp *in_mad  = NULL;
141	struct ib_smp *out_mad = NULL;
142	int err = -ENOMEM;
143	u8 status;
144
145	in_mad  = kzalloc(sizeof *in_mad, GFP_KERNEL);
146	out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
147	if (!in_mad || !out_mad)
148		goto out;
149
150	memset(props, 0, sizeof *props);
151
152	init_query_mad(in_mad);
153	in_mad->attr_id  = IB_SMP_ATTR_PORT_INFO;
154	in_mad->attr_mod = cpu_to_be32(port);
155
156	err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
157			    port, NULL, NULL, in_mad, out_mad,
158			    &status);
159	if (err)
160		goto out;
161	if (status) {
162		err = -EINVAL;
163		goto out;
164	}
165
166	props->lid               = be16_to_cpup((__be16 *) (out_mad->data + 16));
167	props->lmc               = out_mad->data[34] & 0x7;
168	props->sm_lid            = be16_to_cpup((__be16 *) (out_mad->data + 18));
169	props->sm_sl             = out_mad->data[36] & 0xf;
170	props->state             = out_mad->data[32] & 0xf;
171	props->phys_state        = out_mad->data[33] >> 4;
172	props->port_cap_flags    = be32_to_cpup((__be32 *) (out_mad->data + 20));
173	props->gid_tbl_len       = to_mdev(ibdev)->limits.gid_table_len;
174	props->max_msg_sz        = 0x80000000;
175	props->pkey_tbl_len      = to_mdev(ibdev)->limits.pkey_table_len;
176	props->bad_pkey_cntr     = be16_to_cpup((__be16 *) (out_mad->data + 46));
177	props->qkey_viol_cntr    = be16_to_cpup((__be16 *) (out_mad->data + 48));
178	props->active_width      = out_mad->data[31] & 0xf;
179	props->active_speed      = out_mad->data[35] >> 4;
180	props->max_mtu           = out_mad->data[41] & 0xf;
181	props->active_mtu        = out_mad->data[36] >> 4;
182	props->subnet_timeout    = out_mad->data[51] & 0x1f;
183	props->max_vl_num        = out_mad->data[37] >> 4;
184	props->init_type_reply   = out_mad->data[41] >> 4;
185
186 out:
187	kfree(in_mad);
188	kfree(out_mad);
189	return err;
190}
191
192static int mthca_modify_device(struct ib_device *ibdev,
193			       int mask,
194			       struct ib_device_modify *props)
195{
196	if (mask & ~IB_DEVICE_MODIFY_NODE_DESC)
197		return -EOPNOTSUPP;
198
199	if (mask & IB_DEVICE_MODIFY_NODE_DESC) {
200		if (mutex_lock_interruptible(&to_mdev(ibdev)->cap_mask_mutex))
201			return -ERESTARTSYS;
202		memcpy(ibdev->node_desc, props->node_desc, 64);
203		mutex_unlock(&to_mdev(ibdev)->cap_mask_mutex);
204	}
205
206	return 0;
207}
208
209static int mthca_modify_port(struct ib_device *ibdev,
210			     u8 port, int port_modify_mask,
211			     struct ib_port_modify *props)
212{
213	struct mthca_set_ib_param set_ib;
214	struct ib_port_attr attr;
215	int err;
216	u8 status;
217
218	if (mutex_lock_interruptible(&to_mdev(ibdev)->cap_mask_mutex))
219		return -ERESTARTSYS;
220
221	err = mthca_query_port(ibdev, port, &attr);
222	if (err)
223		goto out;
224
225	set_ib.set_si_guid     = 0;
226	set_ib.reset_qkey_viol = !!(port_modify_mask & IB_PORT_RESET_QKEY_CNTR);
227
228	set_ib.cap_mask = (attr.port_cap_flags | props->set_port_cap_mask) &
229		~props->clr_port_cap_mask;
230
231	err = mthca_SET_IB(to_mdev(ibdev), &set_ib, port, &status);
232	if (err)
233		goto out;
234	if (status) {
235		err = -EINVAL;
236		goto out;
237	}
238
239out:
240	mutex_unlock(&to_mdev(ibdev)->cap_mask_mutex);
241	return err;
242}
243
244static int mthca_query_pkey(struct ib_device *ibdev,
245			    u8 port, u16 index, u16 *pkey)
246{
247	struct ib_smp *in_mad  = NULL;
248	struct ib_smp *out_mad = NULL;
249	int err = -ENOMEM;
250	u8 status;
251
252	in_mad  = kzalloc(sizeof *in_mad, GFP_KERNEL);
253	out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
254	if (!in_mad || !out_mad)
255		goto out;
256
257	init_query_mad(in_mad);
258	in_mad->attr_id  = IB_SMP_ATTR_PKEY_TABLE;
259	in_mad->attr_mod = cpu_to_be32(index / 32);
260
261	err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
262			    port, NULL, NULL, in_mad, out_mad,
263			    &status);
264	if (err)
265		goto out;
266	if (status) {
267		err = -EINVAL;
268		goto out;
269	}
270
271	*pkey = be16_to_cpu(((__be16 *) out_mad->data)[index % 32]);
272
273 out:
274	kfree(in_mad);
275	kfree(out_mad);
276	return err;
277}
278
279static int mthca_query_gid(struct ib_device *ibdev, u8 port,
280			   int index, union ib_gid *gid)
281{
282	struct ib_smp *in_mad  = NULL;
283	struct ib_smp *out_mad = NULL;
284	int err = -ENOMEM;
285	u8 status;
286
287	in_mad  = kzalloc(sizeof *in_mad, GFP_KERNEL);
288	out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
289	if (!in_mad || !out_mad)
290		goto out;
291
292	init_query_mad(in_mad);
293	in_mad->attr_id  = IB_SMP_ATTR_PORT_INFO;
294	in_mad->attr_mod = cpu_to_be32(port);
295
296	err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
297			    port, NULL, NULL, in_mad, out_mad,
298			    &status);
299	if (err)
300		goto out;
301	if (status) {
302		err = -EINVAL;
303		goto out;
304	}
305
306	memcpy(gid->raw, out_mad->data + 8, 8);
307
308	init_query_mad(in_mad);
309	in_mad->attr_id  = IB_SMP_ATTR_GUID_INFO;
310	in_mad->attr_mod = cpu_to_be32(index / 8);
311
312	err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
313			    port, NULL, NULL, in_mad, out_mad,
314			    &status);
315	if (err)
316		goto out;
317	if (status) {
318		err = -EINVAL;
319		goto out;
320	}
321
322	memcpy(gid->raw + 8, out_mad->data + (index % 8) * 8, 8);
323
324 out:
325	kfree(in_mad);
326	kfree(out_mad);
327	return err;
328}
329
330static struct ib_ucontext *mthca_alloc_ucontext(struct ib_device *ibdev,
331						struct ib_udata *udata)
332{
333	struct mthca_alloc_ucontext_resp uresp;
334	struct mthca_ucontext           *context;
335	int                              err;
336
337	memset(&uresp, 0, sizeof uresp);
338
339	uresp.qp_tab_size = to_mdev(ibdev)->limits.num_qps;
340	if (mthca_is_memfree(to_mdev(ibdev)))
341		uresp.uarc_size = to_mdev(ibdev)->uar_table.uarc_size;
342	else
343		uresp.uarc_size = 0;
344
345	context = kmalloc(sizeof *context, GFP_KERNEL);
346	if (!context)
347		return ERR_PTR(-ENOMEM);
348
349	err = mthca_uar_alloc(to_mdev(ibdev), &context->uar);
350	if (err) {
351		kfree(context);
352		return ERR_PTR(err);
353	}
354
355	context->db_tab = mthca_init_user_db_tab(to_mdev(ibdev));
356	if (IS_ERR(context->db_tab)) {
357		err = PTR_ERR(context->db_tab);
358		mthca_uar_free(to_mdev(ibdev), &context->uar);
359		kfree(context);
360		return ERR_PTR(err);
361	}
362
363	if (ib_copy_to_udata(udata, &uresp, sizeof uresp)) {
364		mthca_cleanup_user_db_tab(to_mdev(ibdev), &context->uar, context->db_tab);
365		mthca_uar_free(to_mdev(ibdev), &context->uar);
366		kfree(context);
367		return ERR_PTR(-EFAULT);
368	}
369
370	return &context->ibucontext;
371}
372
373static int mthca_dealloc_ucontext(struct ib_ucontext *context)
374{
375	mthca_cleanup_user_db_tab(to_mdev(context->device), &to_mucontext(context)->uar,
376				  to_mucontext(context)->db_tab);
377	mthca_uar_free(to_mdev(context->device), &to_mucontext(context)->uar);
378	kfree(to_mucontext(context));
379
380	return 0;
381}
382
383static int mthca_mmap_uar(struct ib_ucontext *context,
384			  struct vm_area_struct *vma)
385{
386	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
387		return -EINVAL;
388
389	vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
390
391	if (io_remap_pfn_range(vma, vma->vm_start,
392			       to_mucontext(context)->uar.pfn,
393			       PAGE_SIZE, vma->vm_page_prot))
394		return -EAGAIN;
395
396	return 0;
397}
398
399static struct ib_pd *mthca_alloc_pd(struct ib_device *ibdev,
400				    struct ib_ucontext *context,
401				    struct ib_udata *udata)
402{
403	struct mthca_pd *pd;
404	int err;
405
406	pd = kmalloc(sizeof *pd, GFP_KERNEL);
407	if (!pd)
408		return ERR_PTR(-ENOMEM);
409
410	err = mthca_pd_alloc(to_mdev(ibdev), !context, pd);
411	if (err) {
412		kfree(pd);
413		return ERR_PTR(err);
414	}
415
416	if (context) {
417		if (ib_copy_to_udata(udata, &pd->pd_num, sizeof (__u32))) {
418			mthca_pd_free(to_mdev(ibdev), pd);
419			kfree(pd);
420			return ERR_PTR(-EFAULT);
421		}
422	}
423
424	return &pd->ibpd;
425}
426
427static int mthca_dealloc_pd(struct ib_pd *pd)
428{
429	mthca_pd_free(to_mdev(pd->device), to_mpd(pd));
430	kfree(pd);
431
432	return 0;
433}
434
435static struct ib_ah *mthca_ah_create(struct ib_pd *pd,
436				     struct ib_ah_attr *ah_attr)
437{
438	int err;
439	struct mthca_ah *ah;
440
441	ah = kmalloc(sizeof *ah, GFP_ATOMIC);
442	if (!ah)
443		return ERR_PTR(-ENOMEM);
444
445	err = mthca_create_ah(to_mdev(pd->device), to_mpd(pd), ah_attr, ah);
446	if (err) {
447		kfree(ah);
448		return ERR_PTR(err);
449	}
450
451	return &ah->ibah;
452}
453
454static int mthca_ah_destroy(struct ib_ah *ah)
455{
456	mthca_destroy_ah(to_mdev(ah->device), to_mah(ah));
457	kfree(ah);
458
459	return 0;
460}
461
462static struct ib_srq *mthca_create_srq(struct ib_pd *pd,
463				       struct ib_srq_init_attr *init_attr,
464				       struct ib_udata *udata)
465{
466	struct mthca_create_srq ucmd;
467	struct mthca_ucontext *context = NULL;
468	struct mthca_srq *srq;
469	int err;
470
471	srq = kmalloc(sizeof *srq, GFP_KERNEL);
472	if (!srq)
473		return ERR_PTR(-ENOMEM);
474
475	if (pd->uobject) {
476		context = to_mucontext(pd->uobject->context);
477
478		if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
479			err = -EFAULT;
480			goto err_free;
481		}
482
483		err = mthca_map_user_db(to_mdev(pd->device), &context->uar,
484					context->db_tab, ucmd.db_index,
485					ucmd.db_page);
486
487		if (err)
488			goto err_free;
489
490		srq->mr.ibmr.lkey = ucmd.lkey;
491		srq->db_index     = ucmd.db_index;
492	}
493
494	err = mthca_alloc_srq(to_mdev(pd->device), to_mpd(pd),
495			      &init_attr->attr, srq);
496
497	if (err && pd->uobject)
498		mthca_unmap_user_db(to_mdev(pd->device), &context->uar,
499				    context->db_tab, ucmd.db_index);
500
501	if (err)
502		goto err_free;
503
504	if (context && ib_copy_to_udata(udata, &srq->srqn, sizeof (__u32))) {
505		mthca_free_srq(to_mdev(pd->device), srq);
506		err = -EFAULT;
507		goto err_free;
508	}
509
510	return &srq->ibsrq;
511
512err_free:
513	kfree(srq);
514
515	return ERR_PTR(err);
516}
517
518static int mthca_destroy_srq(struct ib_srq *srq)
519{
520	struct mthca_ucontext *context;
521
522	if (srq->uobject) {
523		context = to_mucontext(srq->uobject->context);
524
525		mthca_unmap_user_db(to_mdev(srq->device), &context->uar,
526				    context->db_tab, to_msrq(srq)->db_index);
527	}
528
529	mthca_free_srq(to_mdev(srq->device), to_msrq(srq));
530	kfree(srq);
531
532	return 0;
533}
534
535static struct ib_qp *mthca_create_qp(struct ib_pd *pd,
536				     struct ib_qp_init_attr *init_attr,
537				     struct ib_udata *udata)
538{
539	struct mthca_create_qp ucmd;
540	struct mthca_qp *qp;
541	int err;
542
543	switch (init_attr->qp_type) {
544	case IB_QPT_RC:
545	case IB_QPT_UC:
546	case IB_QPT_UD:
547	{
548		struct mthca_ucontext *context;
549
550		qp = kmalloc(sizeof *qp, GFP_KERNEL);
551		if (!qp)
552			return ERR_PTR(-ENOMEM);
553
554		if (pd->uobject) {
555			context = to_mucontext(pd->uobject->context);
556
557			if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
558				kfree(qp);
559				return ERR_PTR(-EFAULT);
560			}
561
562			err = mthca_map_user_db(to_mdev(pd->device), &context->uar,
563						context->db_tab,
564						ucmd.sq_db_index, ucmd.sq_db_page);
565			if (err) {
566				kfree(qp);
567				return ERR_PTR(err);
568			}
569
570			err = mthca_map_user_db(to_mdev(pd->device), &context->uar,
571						context->db_tab,
572						ucmd.rq_db_index, ucmd.rq_db_page);
573			if (err) {
574				mthca_unmap_user_db(to_mdev(pd->device),
575						    &context->uar,
576						    context->db_tab,
577						    ucmd.sq_db_index);
578				kfree(qp);
579				return ERR_PTR(err);
580			}
581
582			qp->mr.ibmr.lkey = ucmd.lkey;
583			qp->sq.db_index  = ucmd.sq_db_index;
584			qp->rq.db_index  = ucmd.rq_db_index;
585		}
586
587		err = mthca_alloc_qp(to_mdev(pd->device), to_mpd(pd),
588				     to_mcq(init_attr->send_cq),
589				     to_mcq(init_attr->recv_cq),
590				     init_attr->qp_type, init_attr->sq_sig_type,
591				     &init_attr->cap, qp);
592
593		if (err && pd->uobject) {
594			context = to_mucontext(pd->uobject->context);
595
596			mthca_unmap_user_db(to_mdev(pd->device),
597					    &context->uar,
598					    context->db_tab,
599					    ucmd.sq_db_index);
600			mthca_unmap_user_db(to_mdev(pd->device),
601					    &context->uar,
602					    context->db_tab,
603					    ucmd.rq_db_index);
604		}
605
606		qp->ibqp.qp_num = qp->qpn;
607		break;
608	}
609	case IB_QPT_SMI:
610	case IB_QPT_GSI:
611	{
612		/* Don't allow userspace to create special QPs */
613		if (pd->uobject)
614			return ERR_PTR(-EINVAL);
615
616		qp = kmalloc(sizeof (struct mthca_sqp), GFP_KERNEL);
617		if (!qp)
618			return ERR_PTR(-ENOMEM);
619
620		qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 : 1;
621
622		err = mthca_alloc_sqp(to_mdev(pd->device), to_mpd(pd),
623				      to_mcq(init_attr->send_cq),
624				      to_mcq(init_attr->recv_cq),
625				      init_attr->sq_sig_type, &init_attr->cap,
626				      qp->ibqp.qp_num, init_attr->port_num,
627				      to_msqp(qp));
628		break;
629	}
630	default:
631		/* Don't support raw QPs */
632		return ERR_PTR(-ENOSYS);
633	}
634
635	if (err) {
636		kfree(qp);
637		return ERR_PTR(err);
638	}
639
640	init_attr->cap.max_send_wr     = qp->sq.max;
641	init_attr->cap.max_recv_wr     = qp->rq.max;
642	init_attr->cap.max_send_sge    = qp->sq.max_gs;
643	init_attr->cap.max_recv_sge    = qp->rq.max_gs;
644	init_attr->cap.max_inline_data = qp->max_inline_data;
645
646	return &qp->ibqp;
647}
648
649static int mthca_destroy_qp(struct ib_qp *qp)
650{
651	if (qp->uobject) {
652		mthca_unmap_user_db(to_mdev(qp->device),
653				    &to_mucontext(qp->uobject->context)->uar,
654				    to_mucontext(qp->uobject->context)->db_tab,
655				    to_mqp(qp)->sq.db_index);
656		mthca_unmap_user_db(to_mdev(qp->device),
657				    &to_mucontext(qp->uobject->context)->uar,
658				    to_mucontext(qp->uobject->context)->db_tab,
659				    to_mqp(qp)->rq.db_index);
660	}
661	mthca_free_qp(to_mdev(qp->device), to_mqp(qp));
662	kfree(qp);
663	return 0;
664}
665
666static struct ib_cq *mthca_create_cq(struct ib_device *ibdev, int entries,
667				     int comp_vector,
668				     struct ib_ucontext *context,
669				     struct ib_udata *udata)
670{
671	struct mthca_create_cq ucmd;
672	struct mthca_cq *cq;
673	int nent;
674	int err;
675
676	if (entries < 1 || entries > to_mdev(ibdev)->limits.max_cqes)
677		return ERR_PTR(-EINVAL);
678
679	if (context) {
680		if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd))
681			return ERR_PTR(-EFAULT);
682
683		err = mthca_map_user_db(to_mdev(ibdev), &to_mucontext(context)->uar,
684					to_mucontext(context)->db_tab,
685					ucmd.set_db_index, ucmd.set_db_page);
686		if (err)
687			return ERR_PTR(err);
688
689		err = mthca_map_user_db(to_mdev(ibdev), &to_mucontext(context)->uar,
690					to_mucontext(context)->db_tab,
691					ucmd.arm_db_index, ucmd.arm_db_page);
692		if (err)
693			goto err_unmap_set;
694	}
695
696	cq = kmalloc(sizeof *cq, GFP_KERNEL);
697	if (!cq) {
698		err = -ENOMEM;
699		goto err_unmap_arm;
700	}
701
702	if (context) {
703		cq->buf.mr.ibmr.lkey = ucmd.lkey;
704		cq->set_ci_db_index  = ucmd.set_db_index;
705		cq->arm_db_index     = ucmd.arm_db_index;
706	}
707
708	for (nent = 1; nent <= entries; nent <<= 1)
709		; /* nothing */
710
711	err = mthca_init_cq(to_mdev(ibdev), nent,
712			    context ? to_mucontext(context) : NULL,
713			    context ? ucmd.pdn : to_mdev(ibdev)->driver_pd.pd_num,
714			    cq);
715	if (err)
716		goto err_free;
717
718	if (context && ib_copy_to_udata(udata, &cq->cqn, sizeof (__u32))) {
719		mthca_free_cq(to_mdev(ibdev), cq);
720		goto err_free;
721	}
722
723	cq->resize_buf = NULL;
724
725	return &cq->ibcq;
726
727err_free:
728	kfree(cq);
729
730err_unmap_arm:
731	if (context)
732		mthca_unmap_user_db(to_mdev(ibdev), &to_mucontext(context)->uar,
733				    to_mucontext(context)->db_tab, ucmd.arm_db_index);
734
735err_unmap_set:
736	if (context)
737		mthca_unmap_user_db(to_mdev(ibdev), &to_mucontext(context)->uar,
738				    to_mucontext(context)->db_tab, ucmd.set_db_index);
739
740	return ERR_PTR(err);
741}
742
743static int mthca_alloc_resize_buf(struct mthca_dev *dev, struct mthca_cq *cq,
744				  int entries)
745{
746	int ret;
747
748	spin_lock_irq(&cq->lock);
749	if (cq->resize_buf) {
750		ret = -EBUSY;
751		goto unlock;
752	}
753
754	cq->resize_buf = kmalloc(sizeof *cq->resize_buf, GFP_ATOMIC);
755	if (!cq->resize_buf) {
756		ret = -ENOMEM;
757		goto unlock;
758	}
759
760	cq->resize_buf->state = CQ_RESIZE_ALLOC;
761
762	ret = 0;
763
764unlock:
765	spin_unlock_irq(&cq->lock);
766
767	if (ret)
768		return ret;
769
770	ret = mthca_alloc_cq_buf(dev, &cq->resize_buf->buf, entries);
771	if (ret) {
772		spin_lock_irq(&cq->lock);
773		kfree(cq->resize_buf);
774		cq->resize_buf = NULL;
775		spin_unlock_irq(&cq->lock);
776		return ret;
777	}
778
779	cq->resize_buf->cqe = entries - 1;
780
781	spin_lock_irq(&cq->lock);
782	cq->resize_buf->state = CQ_RESIZE_READY;
783	spin_unlock_irq(&cq->lock);
784
785	return 0;
786}
787
788static int mthca_resize_cq(struct ib_cq *ibcq, int entries, struct ib_udata *udata)
789{
790	struct mthca_dev *dev = to_mdev(ibcq->device);
791	struct mthca_cq *cq = to_mcq(ibcq);
792	struct mthca_resize_cq ucmd;
793	u32 lkey;
794	u8 status;
795	int ret;
796
797	if (entries < 1 || entries > dev->limits.max_cqes)
798		return -EINVAL;
799
800	mutex_lock(&cq->mutex);
801
802	entries = roundup_pow_of_two(entries + 1);
803	if (entries == ibcq->cqe + 1) {
804		ret = 0;
805		goto out;
806	}
807
808	if (cq->is_kernel) {
809		ret = mthca_alloc_resize_buf(dev, cq, entries);
810		if (ret)
811			goto out;
812		lkey = cq->resize_buf->buf.mr.ibmr.lkey;
813	} else {
814		if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
815			ret = -EFAULT;
816			goto out;
817		}
818		lkey = ucmd.lkey;
819	}
820
821	ret = mthca_RESIZE_CQ(dev, cq->cqn, lkey, ilog2(entries), &status);
822	if (status)
823		ret = -EINVAL;
824
825	if (ret) {
826		if (cq->resize_buf) {
827			mthca_free_cq_buf(dev, &cq->resize_buf->buf,
828					  cq->resize_buf->cqe);
829			kfree(cq->resize_buf);
830			spin_lock_irq(&cq->lock);
831			cq->resize_buf = NULL;
832			spin_unlock_irq(&cq->lock);
833		}
834		goto out;
835	}
836
837	if (cq->is_kernel) {
838		struct mthca_cq_buf tbuf;
839		int tcqe;
840
841		spin_lock_irq(&cq->lock);
842		if (cq->resize_buf->state == CQ_RESIZE_READY) {
843			mthca_cq_resize_copy_cqes(cq);
844			tbuf         = cq->buf;
845			tcqe         = cq->ibcq.cqe;
846			cq->buf      = cq->resize_buf->buf;
847			cq->ibcq.cqe = cq->resize_buf->cqe;
848		} else {
849			tbuf = cq->resize_buf->buf;
850			tcqe = cq->resize_buf->cqe;
851		}
852
853		kfree(cq->resize_buf);
854		cq->resize_buf = NULL;
855		spin_unlock_irq(&cq->lock);
856
857		mthca_free_cq_buf(dev, &tbuf, tcqe);
858	} else
859		ibcq->cqe = entries - 1;
860
861out:
862	mutex_unlock(&cq->mutex);
863
864	return ret;
865}
866
867static int mthca_destroy_cq(struct ib_cq *cq)
868{
869	if (cq->uobject) {
870		mthca_unmap_user_db(to_mdev(cq->device),
871				    &to_mucontext(cq->uobject->context)->uar,
872				    to_mucontext(cq->uobject->context)->db_tab,
873				    to_mcq(cq)->arm_db_index);
874		mthca_unmap_user_db(to_mdev(cq->device),
875				    &to_mucontext(cq->uobject->context)->uar,
876				    to_mucontext(cq->uobject->context)->db_tab,
877				    to_mcq(cq)->set_ci_db_index);
878	}
879	mthca_free_cq(to_mdev(cq->device), to_mcq(cq));
880	kfree(cq);
881
882	return 0;
883}
884
885static inline u32 convert_access(int acc)
886{
887	return (acc & IB_ACCESS_REMOTE_ATOMIC ? MTHCA_MPT_FLAG_ATOMIC       : 0) |
888	       (acc & IB_ACCESS_REMOTE_WRITE  ? MTHCA_MPT_FLAG_REMOTE_WRITE : 0) |
889	       (acc & IB_ACCESS_REMOTE_READ   ? MTHCA_MPT_FLAG_REMOTE_READ  : 0) |
890	       (acc & IB_ACCESS_LOCAL_WRITE   ? MTHCA_MPT_FLAG_LOCAL_WRITE  : 0) |
891	       MTHCA_MPT_FLAG_LOCAL_READ;
892}
893
894static struct ib_mr *mthca_get_dma_mr(struct ib_pd *pd, int acc)
895{
896	struct mthca_mr *mr;
897	int err;
898
899	mr = kmalloc(sizeof *mr, GFP_KERNEL);
900	if (!mr)
901		return ERR_PTR(-ENOMEM);
902
903	err = mthca_mr_alloc_notrans(to_mdev(pd->device),
904				     to_mpd(pd)->pd_num,
905				     convert_access(acc), mr);
906
907	if (err) {
908		kfree(mr);
909		return ERR_PTR(err);
910	}
911
912	mr->umem = NULL;
913
914	return &mr->ibmr;
915}
916
917static struct ib_mr *mthca_reg_phys_mr(struct ib_pd       *pd,
918				       struct ib_phys_buf *buffer_list,
919				       int                 num_phys_buf,
920				       int                 acc,
921				       u64                *iova_start)
922{
923	struct mthca_mr *mr;
924	u64 *page_list;
925	u64 total_size;
926	u64 mask;
927	int shift;
928	int npages;
929	int err;
930	int i, j, n;
931
932	/* First check that we have enough alignment */
933	if ((*iova_start & ~PAGE_MASK) != (buffer_list[0].addr & ~PAGE_MASK))
934		return ERR_PTR(-EINVAL);
935
936	mask = 0;
937	total_size = 0;
938	for (i = 0; i < num_phys_buf; ++i) {
939		if (i != 0)
940			mask |= buffer_list[i].addr;
941		if (i != num_phys_buf - 1)
942			mask |= buffer_list[i].addr + buffer_list[i].size;
943
944		total_size += buffer_list[i].size;
945	}
946
947	if (mask & ~PAGE_MASK)
948		return ERR_PTR(-EINVAL);
949
950	/* Find largest page shift we can use to cover buffers */
951	for (shift = PAGE_SHIFT; shift < 31; ++shift)
952		if (num_phys_buf > 1) {
953			if ((1ULL << shift) & mask)
954				break;
955		} else {
956			if (1ULL << shift >=
957			    buffer_list[0].size +
958			    (buffer_list[0].addr & ((1ULL << shift) - 1)))
959				break;
960		}
961
962	buffer_list[0].size += buffer_list[0].addr & ((1ULL << shift) - 1);
963	buffer_list[0].addr &= ~0ull << shift;
964
965	mr = kmalloc(sizeof *mr, GFP_KERNEL);
966	if (!mr)
967		return ERR_PTR(-ENOMEM);
968
969	npages = 0;
970	for (i = 0; i < num_phys_buf; ++i)
971		npages += (buffer_list[i].size + (1ULL << shift) - 1) >> shift;
972
973	if (!npages)
974		return &mr->ibmr;
975
976	page_list = kmalloc(npages * sizeof *page_list, GFP_KERNEL);
977	if (!page_list) {
978		kfree(mr);
979		return ERR_PTR(-ENOMEM);
980	}
981
982	n = 0;
983	for (i = 0; i < num_phys_buf; ++i)
984		for (j = 0;
985		     j < (buffer_list[i].size + (1ULL << shift) - 1) >> shift;
986		     ++j)
987			page_list[n++] = buffer_list[i].addr + ((u64) j << shift);
988
989	mthca_dbg(to_mdev(pd->device), "Registering memory at %llx (iova %llx) "
990		  "in PD %x; shift %d, npages %d.\n",
991		  (unsigned long long) buffer_list[0].addr,
992		  (unsigned long long) *iova_start,
993		  to_mpd(pd)->pd_num,
994		  shift, npages);
995
996	err = mthca_mr_alloc_phys(to_mdev(pd->device),
997				  to_mpd(pd)->pd_num,
998				  page_list, shift, npages,
999				  *iova_start, total_size,
1000				  convert_access(acc), mr);
1001
1002	if (err) {
1003		kfree(page_list);
1004		kfree(mr);
1005		return ERR_PTR(err);
1006	}
1007
1008	kfree(page_list);
1009	mr->umem = NULL;
1010
1011	return &mr->ibmr;
1012}
1013
1014static struct ib_mr *mthca_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
1015				       u64 virt, int acc, struct ib_udata *udata)
1016{
1017	struct mthca_dev *dev = to_mdev(pd->device);
1018	struct ib_umem_chunk *chunk;
1019	struct mthca_mr *mr;
1020	u64 *pages;
1021	int shift, n, len;
1022	int i, j, k;
1023	int err = 0;
1024	int write_mtt_size;
1025
1026	mr = kmalloc(sizeof *mr, GFP_KERNEL);
1027	if (!mr)
1028		return ERR_PTR(-ENOMEM);
1029
1030	mr->umem = ib_umem_get(pd->uobject->context, start, length, acc);
1031	if (IS_ERR(mr->umem)) {
1032		err = PTR_ERR(mr->umem);
1033		goto err;
1034	}
1035
1036	shift = ffs(mr->umem->page_size) - 1;
1037
1038	n = 0;
1039	list_for_each_entry(chunk, &mr->umem->chunk_list, list)
1040		n += chunk->nents;
1041
1042	mr->mtt = mthca_alloc_mtt(dev, n);
1043	if (IS_ERR(mr->mtt)) {
1044		err = PTR_ERR(mr->mtt);
1045		goto err_umem;
1046	}
1047
1048	pages = (u64 *) __get_free_page(GFP_KERNEL);
1049	if (!pages) {
1050		err = -ENOMEM;
1051		goto err_mtt;
1052	}
1053
1054	i = n = 0;
1055
1056	write_mtt_size = min(mthca_write_mtt_size(dev), (int) (PAGE_SIZE / sizeof *pages));
1057
1058	list_for_each_entry(chunk, &mr->umem->chunk_list, list)
1059		for (j = 0; j < chunk->nmap; ++j) {
1060			len = sg_dma_len(&chunk->page_list[j]) >> shift;
1061			for (k = 0; k < len; ++k) {
1062				pages[i++] = sg_dma_address(&chunk->page_list[j]) +
1063					mr->umem->page_size * k;
1064				/*
1065				 * Be friendly to write_mtt and pass it chunks
1066				 * of appropriate size.
1067				 */
1068				if (i == write_mtt_size) {
1069					err = mthca_write_mtt(dev, mr->mtt, n, pages, i);
1070					if (err)
1071						goto mtt_done;
1072					n += i;
1073					i = 0;
1074				}
1075			}
1076		}
1077
1078	if (i)
1079		err = mthca_write_mtt(dev, mr->mtt, n, pages, i);
1080mtt_done:
1081	free_page((unsigned long) pages);
1082	if (err)
1083		goto err_mtt;
1084
1085	err = mthca_mr_alloc(dev, to_mpd(pd)->pd_num, shift, virt, length,
1086			     convert_access(acc), mr);
1087
1088	if (err)
1089		goto err_mtt;
1090
1091	return &mr->ibmr;
1092
1093err_mtt:
1094	mthca_free_mtt(dev, mr->mtt);
1095
1096err_umem:
1097	ib_umem_release(mr->umem);
1098
1099err:
1100	kfree(mr);
1101	return ERR_PTR(err);
1102}
1103
1104static int mthca_dereg_mr(struct ib_mr *mr)
1105{
1106	struct mthca_mr *mmr = to_mmr(mr);
1107
1108	mthca_free_mr(to_mdev(mr->device), mmr);
1109	if (mmr->umem)
1110		ib_umem_release(mmr->umem);
1111	kfree(mmr);
1112
1113	return 0;
1114}
1115
1116static struct ib_fmr *mthca_alloc_fmr(struct ib_pd *pd, int mr_access_flags,
1117				      struct ib_fmr_attr *fmr_attr)
1118{
1119	struct mthca_fmr *fmr;
1120	int err;
1121
1122	fmr = kmalloc(sizeof *fmr, GFP_KERNEL);
1123	if (!fmr)
1124		return ERR_PTR(-ENOMEM);
1125
1126	memcpy(&fmr->attr, fmr_attr, sizeof *fmr_attr);
1127	err = mthca_fmr_alloc(to_mdev(pd->device), to_mpd(pd)->pd_num,
1128			     convert_access(mr_access_flags), fmr);
1129
1130	if (err) {
1131		kfree(fmr);
1132		return ERR_PTR(err);
1133	}
1134
1135	return &fmr->ibmr;
1136}
1137
1138static int mthca_dealloc_fmr(struct ib_fmr *fmr)
1139{
1140	struct mthca_fmr *mfmr = to_mfmr(fmr);
1141	int err;
1142
1143	err = mthca_free_fmr(to_mdev(fmr->device), mfmr);
1144	if (err)
1145		return err;
1146
1147	kfree(mfmr);
1148	return 0;
1149}
1150
1151static int mthca_unmap_fmr(struct list_head *fmr_list)
1152{
1153	struct ib_fmr *fmr;
1154	int err;
1155	u8 status;
1156	struct mthca_dev *mdev = NULL;
1157
1158	list_for_each_entry(fmr, fmr_list, list) {
1159		if (mdev && to_mdev(fmr->device) != mdev)
1160			return -EINVAL;
1161		mdev = to_mdev(fmr->device);
1162	}
1163
1164	if (!mdev)
1165		return 0;
1166
1167	if (mthca_is_memfree(mdev)) {
1168		list_for_each_entry(fmr, fmr_list, list)
1169			mthca_arbel_fmr_unmap(mdev, to_mfmr(fmr));
1170
1171		wmb();
1172	} else
1173		list_for_each_entry(fmr, fmr_list, list)
1174			mthca_tavor_fmr_unmap(mdev, to_mfmr(fmr));
1175
1176	err = mthca_SYNC_TPT(mdev, &status);
1177	if (err)
1178		return err;
1179	if (status)
1180		return -EINVAL;
1181	return 0;
1182}
1183
1184static ssize_t show_rev(struct class_device *cdev, char *buf)
1185{
1186	struct mthca_dev *dev = container_of(cdev, struct mthca_dev, ib_dev.class_dev);
1187	return sprintf(buf, "%x\n", dev->rev_id);
1188}
1189
1190static ssize_t show_fw_ver(struct class_device *cdev, char *buf)
1191{
1192	struct mthca_dev *dev = container_of(cdev, struct mthca_dev, ib_dev.class_dev);
1193	return sprintf(buf, "%d.%d.%d\n", (int) (dev->fw_ver >> 32),
1194		       (int) (dev->fw_ver >> 16) & 0xffff,
1195		       (int) dev->fw_ver & 0xffff);
1196}
1197
1198static ssize_t show_hca(struct class_device *cdev, char *buf)
1199{
1200	struct mthca_dev *dev = container_of(cdev, struct mthca_dev, ib_dev.class_dev);
1201	switch (dev->pdev->device) {
1202	case PCI_DEVICE_ID_MELLANOX_TAVOR:
1203		return sprintf(buf, "MT23108\n");
1204	case PCI_DEVICE_ID_MELLANOX_ARBEL_COMPAT:
1205		return sprintf(buf, "MT25208 (MT23108 compat mode)\n");
1206	case PCI_DEVICE_ID_MELLANOX_ARBEL:
1207		return sprintf(buf, "MT25208\n");
1208	case PCI_DEVICE_ID_MELLANOX_SINAI:
1209	case PCI_DEVICE_ID_MELLANOX_SINAI_OLD:
1210		return sprintf(buf, "MT25204\n");
1211	default:
1212		return sprintf(buf, "unknown\n");
1213	}
1214}
1215
1216static ssize_t show_board(struct class_device *cdev, char *buf)
1217{
1218	struct mthca_dev *dev = container_of(cdev, struct mthca_dev, ib_dev.class_dev);
1219	return sprintf(buf, "%.*s\n", MTHCA_BOARD_ID_LEN, dev->board_id);
1220}
1221
1222static CLASS_DEVICE_ATTR(hw_rev,   S_IRUGO, show_rev,    NULL);
1223static CLASS_DEVICE_ATTR(fw_ver,   S_IRUGO, show_fw_ver, NULL);
1224static CLASS_DEVICE_ATTR(hca_type, S_IRUGO, show_hca,    NULL);
1225static CLASS_DEVICE_ATTR(board_id, S_IRUGO, show_board,  NULL);
1226
1227static struct class_device_attribute *mthca_class_attributes[] = {
1228	&class_device_attr_hw_rev,
1229	&class_device_attr_fw_ver,
1230	&class_device_attr_hca_type,
1231	&class_device_attr_board_id
1232};
1233
1234static int mthca_init_node_data(struct mthca_dev *dev)
1235{
1236	struct ib_smp *in_mad  = NULL;
1237	struct ib_smp *out_mad = NULL;
1238	int err = -ENOMEM;
1239	u8 status;
1240
1241	in_mad  = kzalloc(sizeof *in_mad, GFP_KERNEL);
1242	out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
1243	if (!in_mad || !out_mad)
1244		goto out;
1245
1246	init_query_mad(in_mad);
1247	in_mad->attr_id = IB_SMP_ATTR_NODE_DESC;
1248
1249	err = mthca_MAD_IFC(dev, 1, 1,
1250			    1, NULL, NULL, in_mad, out_mad,
1251			    &status);
1252	if (err)
1253		goto out;
1254	if (status) {
1255		err = -EINVAL;
1256		goto out;
1257	}
1258
1259	memcpy(dev->ib_dev.node_desc, out_mad->data, 64);
1260
1261	in_mad->attr_id = IB_SMP_ATTR_NODE_INFO;
1262
1263	err = mthca_MAD_IFC(dev, 1, 1,
1264			    1, NULL, NULL, in_mad, out_mad,
1265			    &status);
1266	if (err)
1267		goto out;
1268	if (status) {
1269		err = -EINVAL;
1270		goto out;
1271	}
1272
1273	memcpy(&dev->ib_dev.node_guid, out_mad->data + 12, 8);
1274
1275out:
1276	kfree(in_mad);
1277	kfree(out_mad);
1278	return err;
1279}
1280
1281int mthca_register_device(struct mthca_dev *dev)
1282{
1283	int ret;
1284	int i;
1285
1286	ret = mthca_init_node_data(dev);
1287	if (ret)
1288		return ret;
1289
1290	strlcpy(dev->ib_dev.name, "mthca%d", IB_DEVICE_NAME_MAX);
1291	dev->ib_dev.owner                = THIS_MODULE;
1292
1293	dev->ib_dev.uverbs_abi_ver	 = MTHCA_UVERBS_ABI_VERSION;
1294	dev->ib_dev.uverbs_cmd_mask	 =
1295		(1ull << IB_USER_VERBS_CMD_GET_CONTEXT)		|
1296		(1ull << IB_USER_VERBS_CMD_QUERY_DEVICE)	|
1297		(1ull << IB_USER_VERBS_CMD_QUERY_PORT)		|
1298		(1ull << IB_USER_VERBS_CMD_ALLOC_PD)		|
1299		(1ull << IB_USER_VERBS_CMD_DEALLOC_PD)		|
1300		(1ull << IB_USER_VERBS_CMD_REG_MR)		|
1301		(1ull << IB_USER_VERBS_CMD_DEREG_MR)		|
1302		(1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL)	|
1303		(1ull << IB_USER_VERBS_CMD_CREATE_CQ)		|
1304		(1ull << IB_USER_VERBS_CMD_RESIZE_CQ)		|
1305		(1ull << IB_USER_VERBS_CMD_DESTROY_CQ)		|
1306		(1ull << IB_USER_VERBS_CMD_CREATE_QP)		|
1307		(1ull << IB_USER_VERBS_CMD_QUERY_QP)		|
1308		(1ull << IB_USER_VERBS_CMD_MODIFY_QP)		|
1309		(1ull << IB_USER_VERBS_CMD_DESTROY_QP)		|
1310		(1ull << IB_USER_VERBS_CMD_ATTACH_MCAST)	|
1311		(1ull << IB_USER_VERBS_CMD_DETACH_MCAST);
1312	dev->ib_dev.node_type            = RDMA_NODE_IB_CA;
1313	dev->ib_dev.phys_port_cnt        = dev->limits.num_ports;
1314	dev->ib_dev.num_comp_vectors     = 1;
1315	dev->ib_dev.dma_device           = &dev->pdev->dev;
1316	dev->ib_dev.query_device         = mthca_query_device;
1317	dev->ib_dev.query_port           = mthca_query_port;
1318	dev->ib_dev.modify_device        = mthca_modify_device;
1319	dev->ib_dev.modify_port          = mthca_modify_port;
1320	dev->ib_dev.query_pkey           = mthca_query_pkey;
1321	dev->ib_dev.query_gid            = mthca_query_gid;
1322	dev->ib_dev.alloc_ucontext       = mthca_alloc_ucontext;
1323	dev->ib_dev.dealloc_ucontext     = mthca_dealloc_ucontext;
1324	dev->ib_dev.mmap                 = mthca_mmap_uar;
1325	dev->ib_dev.alloc_pd             = mthca_alloc_pd;
1326	dev->ib_dev.dealloc_pd           = mthca_dealloc_pd;
1327	dev->ib_dev.create_ah            = mthca_ah_create;
1328	dev->ib_dev.query_ah             = mthca_ah_query;
1329	dev->ib_dev.destroy_ah           = mthca_ah_destroy;
1330
1331	if (dev->mthca_flags & MTHCA_FLAG_SRQ) {
1332		dev->ib_dev.create_srq           = mthca_create_srq;
1333		dev->ib_dev.modify_srq           = mthca_modify_srq;
1334		dev->ib_dev.query_srq            = mthca_query_srq;
1335		dev->ib_dev.destroy_srq          = mthca_destroy_srq;
1336		dev->ib_dev.uverbs_cmd_mask	|=
1337			(1ull << IB_USER_VERBS_CMD_CREATE_SRQ)		|
1338			(1ull << IB_USER_VERBS_CMD_MODIFY_SRQ)		|
1339			(1ull << IB_USER_VERBS_CMD_QUERY_SRQ)		|
1340			(1ull << IB_USER_VERBS_CMD_DESTROY_SRQ);
1341
1342		if (mthca_is_memfree(dev))
1343			dev->ib_dev.post_srq_recv = mthca_arbel_post_srq_recv;
1344		else
1345			dev->ib_dev.post_srq_recv = mthca_tavor_post_srq_recv;
1346	}
1347
1348	dev->ib_dev.create_qp            = mthca_create_qp;
1349	dev->ib_dev.modify_qp            = mthca_modify_qp;
1350	dev->ib_dev.query_qp             = mthca_query_qp;
1351	dev->ib_dev.destroy_qp           = mthca_destroy_qp;
1352	dev->ib_dev.create_cq            = mthca_create_cq;
1353	dev->ib_dev.resize_cq            = mthca_resize_cq;
1354	dev->ib_dev.destroy_cq           = mthca_destroy_cq;
1355	dev->ib_dev.poll_cq              = mthca_poll_cq;
1356	dev->ib_dev.get_dma_mr           = mthca_get_dma_mr;
1357	dev->ib_dev.reg_phys_mr          = mthca_reg_phys_mr;
1358	dev->ib_dev.reg_user_mr          = mthca_reg_user_mr;
1359	dev->ib_dev.dereg_mr             = mthca_dereg_mr;
1360
1361	if (dev->mthca_flags & MTHCA_FLAG_FMR) {
1362		dev->ib_dev.alloc_fmr            = mthca_alloc_fmr;
1363		dev->ib_dev.unmap_fmr            = mthca_unmap_fmr;
1364		dev->ib_dev.dealloc_fmr          = mthca_dealloc_fmr;
1365		if (mthca_is_memfree(dev))
1366			dev->ib_dev.map_phys_fmr = mthca_arbel_map_phys_fmr;
1367		else
1368			dev->ib_dev.map_phys_fmr = mthca_tavor_map_phys_fmr;
1369	}
1370
1371	dev->ib_dev.attach_mcast         = mthca_multicast_attach;
1372	dev->ib_dev.detach_mcast         = mthca_multicast_detach;
1373	dev->ib_dev.process_mad          = mthca_process_mad;
1374
1375	if (mthca_is_memfree(dev)) {
1376		dev->ib_dev.req_notify_cq = mthca_arbel_arm_cq;
1377		dev->ib_dev.post_send     = mthca_arbel_post_send;
1378		dev->ib_dev.post_recv     = mthca_arbel_post_receive;
1379	} else {
1380		dev->ib_dev.req_notify_cq = mthca_tavor_arm_cq;
1381		dev->ib_dev.post_send     = mthca_tavor_post_send;
1382		dev->ib_dev.post_recv     = mthca_tavor_post_receive;
1383	}
1384
1385	mutex_init(&dev->cap_mask_mutex);
1386
1387	ret = ib_register_device(&dev->ib_dev);
1388	if (ret)
1389		return ret;
1390
1391	for (i = 0; i < ARRAY_SIZE(mthca_class_attributes); ++i) {
1392		ret = class_device_create_file(&dev->ib_dev.class_dev,
1393					       mthca_class_attributes[i]);
1394		if (ret) {
1395			ib_unregister_device(&dev->ib_dev);
1396			return ret;
1397		}
1398	}
1399
1400	mthca_start_catas_poll(dev);
1401
1402	return 0;
1403}
1404
1405void mthca_unregister_device(struct mthca_dev *dev)
1406{
1407	mthca_stop_catas_poll(dev);
1408	ib_unregister_device(&dev->ib_dev);
1409}
1410