xref: /freebsd/sys/dev/mthca/mthca_provider.c (revision 5956d97f4b3204318ceb6aa9c77bd0bc6ea87a41)
1 /*
2  * Copyright (c) 2004, 2005 Topspin Communications.  All rights reserved.
3  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
4  * Copyright (c) 2005, 2006 Cisco Systems.  All rights reserved.
5  * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
6  * Copyright (c) 2004 Voltaire, Inc. All rights reserved.
7  *
8  * This software is available to you under a choice of one of two
9  * licenses.  You may choose to be licensed under the terms of the GNU
10  * General Public License (GPL) Version 2, available from the file
11  * COPYING in the main directory of this source tree, or the
12  * OpenIB.org BSD license below:
13  *
14  *     Redistribution and use in source and binary forms, with or
15  *     without modification, are permitted provided that the following
16  *     conditions are met:
17  *
18  *      - Redistributions of source code must retain the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer.
21  *
22  *      - Redistributions in binary form must reproduce the above
23  *        copyright notice, this list of conditions and the following
24  *        disclaimer in the documentation and/or other materials
25  *        provided with the distribution.
26  *
27  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
28  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
29  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
30  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
31  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
32  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
33  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
34  * SOFTWARE.
35  */
36 
37 #include <rdma/ib_smi.h>
38 #include <rdma/ib_umem.h>
39 #include <rdma/ib_user_verbs.h>
40 #include <rdma/uverbs_ioctl.h>
41 
42 #include <linux/sched.h>
43 #include <linux/slab.h>
44 #include <linux/mm.h>
45 #include <linux/fs.h>
46 
47 #include "mthca_dev.h"
48 #include "mthca_cmd.h"
49 #include <rdma/mthca-abi.h>
50 #include "mthca_memfree.h"
51 
52 static void init_query_mad(struct ib_smp *mad)
53 {
54 	mad->base_version  = 1;
55 	mad->mgmt_class    = IB_MGMT_CLASS_SUBN_LID_ROUTED;
56 	mad->class_version = 1;
57 	mad->method    	   = IB_MGMT_METHOD_GET;
58 }
59 
60 static int mthca_query_device(struct ib_device *ibdev, struct ib_device_attr *props,
61 			      struct ib_udata *uhw)
62 {
63 	struct ib_smp *in_mad  = NULL;
64 	struct ib_smp *out_mad = NULL;
65 	int err = -ENOMEM;
66 	struct mthca_dev *mdev = to_mdev(ibdev);
67 
68 	if (uhw->inlen || uhw->outlen)
69 		return -EINVAL;
70 
71 	in_mad  = kzalloc(sizeof *in_mad, GFP_KERNEL);
72 	out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
73 	if (!in_mad || !out_mad)
74 		goto out;
75 
76 	memset(props, 0, sizeof *props);
77 
78 	props->fw_ver              = mdev->fw_ver;
79 
80 	init_query_mad(in_mad);
81 	in_mad->attr_id = IB_SMP_ATTR_NODE_INFO;
82 
83 	err = mthca_MAD_IFC(mdev, 1, 1,
84 			    1, NULL, NULL, in_mad, out_mad);
85 	if (err)
86 		goto out;
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_sge_rd          = props->max_sge;
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 
137 static 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 
144 	in_mad  = kzalloc(sizeof *in_mad, GFP_KERNEL);
145 	out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
146 	if (!in_mad || !out_mad)
147 		goto out;
148 
149 	memset(props, 0, sizeof *props);
150 
151 	init_query_mad(in_mad);
152 	in_mad->attr_id  = IB_SMP_ATTR_PORT_INFO;
153 	in_mad->attr_mod = cpu_to_be32(port);
154 
155 	err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
156 			    port, NULL, NULL, in_mad, out_mad);
157 	if (err)
158 		goto out;
159 
160 	props->lid               = be16_to_cpup((__be16 *) (out_mad->data + 16));
161 	props->lmc               = out_mad->data[34] & 0x7;
162 	props->sm_lid            = be16_to_cpup((__be16 *) (out_mad->data + 18));
163 	props->sm_sl             = out_mad->data[36] & 0xf;
164 	props->state             = out_mad->data[32] & 0xf;
165 	props->phys_state        = out_mad->data[33] >> 4;
166 	props->port_cap_flags    = be32_to_cpup((__be32 *) (out_mad->data + 20));
167 	props->gid_tbl_len       = to_mdev(ibdev)->limits.gid_table_len;
168 	props->max_msg_sz        = 0x80000000;
169 	props->pkey_tbl_len      = to_mdev(ibdev)->limits.pkey_table_len;
170 	props->bad_pkey_cntr     = be16_to_cpup((__be16 *) (out_mad->data + 46));
171 	props->qkey_viol_cntr    = be16_to_cpup((__be16 *) (out_mad->data + 48));
172 	props->active_width      = out_mad->data[31] & 0xf;
173 	props->active_speed      = out_mad->data[35] >> 4;
174 	props->max_mtu           = out_mad->data[41] & 0xf;
175 	props->active_mtu        = out_mad->data[36] >> 4;
176 	props->subnet_timeout    = out_mad->data[51] & 0x1f;
177 	props->max_vl_num        = out_mad->data[37] >> 4;
178 	props->init_type_reply   = out_mad->data[41] >> 4;
179 
180  out:
181 	kfree(in_mad);
182 	kfree(out_mad);
183 	return err;
184 }
185 
186 static int mthca_modify_device(struct ib_device *ibdev,
187 			       int mask,
188 			       struct ib_device_modify *props)
189 {
190 	if (mask & ~IB_DEVICE_MODIFY_NODE_DESC)
191 		return -EOPNOTSUPP;
192 
193 	if (mask & IB_DEVICE_MODIFY_NODE_DESC) {
194 		if (mutex_lock_interruptible(&to_mdev(ibdev)->cap_mask_mutex))
195 			return -ERESTARTSYS;
196 		memcpy(ibdev->node_desc, props->node_desc,
197 		       IB_DEVICE_NODE_DESC_MAX);
198 		mutex_unlock(&to_mdev(ibdev)->cap_mask_mutex);
199 	}
200 
201 	return 0;
202 }
203 
204 static int mthca_modify_port(struct ib_device *ibdev,
205 			     u8 port, int port_modify_mask,
206 			     struct ib_port_modify *props)
207 {
208 	struct mthca_set_ib_param set_ib;
209 	struct ib_port_attr attr;
210 	int err;
211 
212 	if (mutex_lock_interruptible(&to_mdev(ibdev)->cap_mask_mutex))
213 		return -ERESTARTSYS;
214 
215 	err = mthca_query_port(ibdev, port, &attr);
216 	if (err)
217 		goto out;
218 
219 	set_ib.set_si_guid     = 0;
220 	set_ib.reset_qkey_viol = !!(port_modify_mask & IB_PORT_RESET_QKEY_CNTR);
221 
222 	set_ib.cap_mask = (attr.port_cap_flags | props->set_port_cap_mask) &
223 		~props->clr_port_cap_mask;
224 
225 	err = mthca_SET_IB(to_mdev(ibdev), &set_ib, port);
226 	if (err)
227 		goto out;
228 out:
229 	mutex_unlock(&to_mdev(ibdev)->cap_mask_mutex);
230 	return err;
231 }
232 
233 static int mthca_query_pkey(struct ib_device *ibdev,
234 			    u8 port, u16 index, u16 *pkey)
235 {
236 	struct ib_smp *in_mad  = NULL;
237 	struct ib_smp *out_mad = NULL;
238 	int err = -ENOMEM;
239 
240 	in_mad  = kzalloc(sizeof *in_mad, GFP_KERNEL);
241 	out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
242 	if (!in_mad || !out_mad)
243 		goto out;
244 
245 	init_query_mad(in_mad);
246 	in_mad->attr_id  = IB_SMP_ATTR_PKEY_TABLE;
247 	in_mad->attr_mod = cpu_to_be32(index / 32);
248 
249 	err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
250 			    port, NULL, NULL, in_mad, out_mad);
251 	if (err)
252 		goto out;
253 
254 	*pkey = be16_to_cpu(((__be16 *) out_mad->data)[index % 32]);
255 
256  out:
257 	kfree(in_mad);
258 	kfree(out_mad);
259 	return err;
260 }
261 
262 static int mthca_query_gid(struct ib_device *ibdev, u8 port,
263 			   int index, union ib_gid *gid)
264 {
265 	struct ib_smp *in_mad  = NULL;
266 	struct ib_smp *out_mad = NULL;
267 	int err = -ENOMEM;
268 
269 	in_mad  = kzalloc(sizeof *in_mad, GFP_KERNEL);
270 	out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
271 	if (!in_mad || !out_mad)
272 		goto out;
273 
274 	init_query_mad(in_mad);
275 	in_mad->attr_id  = IB_SMP_ATTR_PORT_INFO;
276 	in_mad->attr_mod = cpu_to_be32(port);
277 
278 	err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
279 			    port, NULL, NULL, in_mad, out_mad);
280 	if (err)
281 		goto out;
282 
283 	memcpy(gid->raw, out_mad->data + 8, 8);
284 
285 	init_query_mad(in_mad);
286 	in_mad->attr_id  = IB_SMP_ATTR_GUID_INFO;
287 	in_mad->attr_mod = cpu_to_be32(index / 8);
288 
289 	err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
290 			    port, NULL, NULL, in_mad, out_mad);
291 	if (err)
292 		goto out;
293 
294 	memcpy(gid->raw + 8, out_mad->data + (index % 8) * 8, 8);
295 
296  out:
297 	kfree(in_mad);
298 	kfree(out_mad);
299 	return err;
300 }
301 
302 static int mthca_alloc_ucontext(struct ib_ucontext *uctx,
303 				struct ib_udata *udata)
304 {
305 	struct ib_device *ibdev = uctx->device;
306 	struct mthca_alloc_ucontext_resp uresp = {};
307 	struct mthca_ucontext *context = to_mucontext(uctx);
308 	int                              err;
309 
310 	if (!(to_mdev(ibdev)->active))
311 		return -EAGAIN;
312 
313 	uresp.qp_tab_size = to_mdev(ibdev)->limits.num_qps;
314 	if (mthca_is_memfree(to_mdev(ibdev)))
315 		uresp.uarc_size = to_mdev(ibdev)->uar_table.uarc_size;
316 	else
317 		uresp.uarc_size = 0;
318 
319 	err = mthca_uar_alloc(to_mdev(ibdev), &context->uar);
320 	if (err)
321 		return err;
322 
323 	context->db_tab = mthca_init_user_db_tab(to_mdev(ibdev));
324 	if (IS_ERR(context->db_tab)) {
325 		err = PTR_ERR(context->db_tab);
326 		mthca_uar_free(to_mdev(ibdev), &context->uar);
327 		return err;
328 	}
329 
330 	if (ib_copy_to_udata(udata, &uresp, sizeof(uresp))) {
331 		mthca_cleanup_user_db_tab(to_mdev(ibdev), &context->uar, context->db_tab);
332 		mthca_uar_free(to_mdev(ibdev), &context->uar);
333 		return -EFAULT;
334 	}
335 
336 	context->reg_mr_warned = 0;
337 
338 	return 0;
339 }
340 
341 static void mthca_dealloc_ucontext(struct ib_ucontext *context)
342 {
343 	mthca_cleanup_user_db_tab(to_mdev(context->device), &to_mucontext(context)->uar,
344 				  to_mucontext(context)->db_tab);
345 	mthca_uar_free(to_mdev(context->device), &to_mucontext(context)->uar);
346 }
347 
348 static int mthca_mmap_uar(struct ib_ucontext *context,
349 			  struct vm_area_struct *vma)
350 {
351 	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
352 		return -EINVAL;
353 
354 	vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
355 
356 	if (io_remap_pfn_range(vma, vma->vm_start,
357 			       to_mucontext(context)->uar.pfn,
358 			       PAGE_SIZE, vma->vm_page_prot))
359 		return -EAGAIN;
360 
361 	return 0;
362 }
363 
364 static int mthca_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
365 {
366 	struct ib_device *ibdev = ibpd->device;
367 	struct mthca_pd *pd = to_mpd(ibpd);
368 	int err;
369 
370 	err = mthca_pd_alloc(to_mdev(ibdev), !udata, pd);
371 	if (err)
372 		return err;
373 
374 	if (udata) {
375 		if (ib_copy_to_udata(udata, &pd->pd_num, sizeof (__u32))) {
376 			mthca_pd_free(to_mdev(ibdev), pd);
377 			return -EFAULT;
378 		}
379 	}
380 
381 	return 0;
382 }
383 
384 static void mthca_dealloc_pd(struct ib_pd *pd, struct ib_udata *udata)
385 {
386 	mthca_pd_free(to_mdev(pd->device), to_mpd(pd));
387 }
388 
389 static int mthca_ah_create(struct ib_ah *ibah,
390 			   struct ib_ah_attr *init_attr, u32 flags,
391 			   struct ib_udata *udata)
392 
393 {
394 	struct mthca_ah *ah = to_mah(ibah);
395 
396 	return mthca_create_ah(to_mdev(ibah->device), to_mpd(ibah->pd),
397 			       init_attr, ah);
398 }
399 
400 static void mthca_ah_destroy(struct ib_ah *ah, u32 flags)
401 {
402 	mthca_destroy_ah(to_mdev(ah->device), to_mah(ah));
403 }
404 
405 static int mthca_create_srq(struct ib_srq *ibsrq,
406 			    struct ib_srq_init_attr *init_attr,
407 			    struct ib_udata *udata)
408 {
409 	struct mthca_create_srq ucmd;
410 	struct mthca_ucontext *context = rdma_udata_to_drv_context(
411 		udata, struct mthca_ucontext, ibucontext);
412 	struct mthca_srq *srq = to_msrq(ibsrq);
413 	int err;
414 
415 	if (init_attr->srq_type != IB_SRQT_BASIC)
416 		return -EOPNOTSUPP;
417 
418 	if (udata) {
419 		if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd)))
420 			return -EFAULT;
421 
422 		err = mthca_map_user_db(to_mdev(ibsrq->device), &context->uar,
423 					context->db_tab, ucmd.db_index,
424 					ucmd.db_page);
425 
426 		if (err)
427 			return err;
428 
429 		srq->mr.ibmr.lkey = ucmd.lkey;
430 		srq->db_index     = ucmd.db_index;
431 	}
432 
433 	err = mthca_alloc_srq(to_mdev(ibsrq->device), to_mpd(ibsrq->pd),
434 			      &init_attr->attr, srq, udata);
435 
436 	if (err && udata)
437 		mthca_unmap_user_db(to_mdev(ibsrq->device), &context->uar,
438 				    context->db_tab, ucmd.db_index);
439 
440 	if (err)
441 		return err;
442 
443 	if (context && ib_copy_to_udata(udata, &srq->srqn, sizeof(__u32))) {
444 		mthca_free_srq(to_mdev(ibsrq->device), srq);
445 		return -EFAULT;
446 	}
447 
448 	return 0;
449 }
450 
451 static void mthca_destroy_srq(struct ib_srq *srq, struct ib_udata *udata)
452 {
453 	if (udata) {
454 		struct mthca_ucontext *context =
455 			rdma_udata_to_drv_context(
456 				udata,
457 				struct mthca_ucontext,
458 				ibucontext);
459 
460 		mthca_unmap_user_db(to_mdev(srq->device), &context->uar,
461 				    context->db_tab, to_msrq(srq)->db_index);
462 	}
463 
464 	mthca_free_srq(to_mdev(srq->device), to_msrq(srq));
465 }
466 
467 static struct ib_qp *mthca_create_qp(struct ib_pd *pd,
468 				     struct ib_qp_init_attr *init_attr,
469 				     struct ib_udata *udata)
470 {
471 	struct mthca_ucontext *context = rdma_udata_to_drv_context(
472 		udata, struct mthca_ucontext, ibucontext);
473 	struct mthca_create_qp ucmd;
474 	struct mthca_qp *qp;
475 	int err;
476 
477 	if (init_attr->create_flags)
478 		return ERR_PTR(-EINVAL);
479 
480 	switch (init_attr->qp_type) {
481 	case IB_QPT_RC:
482 	case IB_QPT_UC:
483 	case IB_QPT_UD:
484 	{
485 		qp = kzalloc(sizeof(*qp), GFP_KERNEL);
486 		if (!qp)
487 			return ERR_PTR(-ENOMEM);
488 
489 		if (udata) {
490 			if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
491 				kfree(qp);
492 				return ERR_PTR(-EFAULT);
493 			}
494 
495 			err = mthca_map_user_db(to_mdev(pd->device), &context->uar,
496 						context->db_tab,
497 						ucmd.sq_db_index, ucmd.sq_db_page);
498 			if (err) {
499 				kfree(qp);
500 				return ERR_PTR(err);
501 			}
502 
503 			err = mthca_map_user_db(to_mdev(pd->device), &context->uar,
504 						context->db_tab,
505 						ucmd.rq_db_index, ucmd.rq_db_page);
506 			if (err) {
507 				mthca_unmap_user_db(to_mdev(pd->device),
508 						    &context->uar,
509 						    context->db_tab,
510 						    ucmd.sq_db_index);
511 				kfree(qp);
512 				return ERR_PTR(err);
513 			}
514 
515 			qp->mr.ibmr.lkey = ucmd.lkey;
516 			qp->sq.db_index  = ucmd.sq_db_index;
517 			qp->rq.db_index  = ucmd.rq_db_index;
518 		}
519 
520 		err = mthca_alloc_qp(to_mdev(pd->device), to_mpd(pd),
521 				     to_mcq(init_attr->send_cq),
522 				     to_mcq(init_attr->recv_cq),
523 				     init_attr->qp_type, init_attr->sq_sig_type,
524 				     &init_attr->cap, qp, udata);
525 
526 		if (err && udata) {
527 			mthca_unmap_user_db(to_mdev(pd->device),
528 					    &context->uar,
529 					    context->db_tab,
530 					    ucmd.sq_db_index);
531 			mthca_unmap_user_db(to_mdev(pd->device),
532 					    &context->uar,
533 					    context->db_tab,
534 					    ucmd.rq_db_index);
535 		}
536 
537 		qp->ibqp.qp_num = qp->qpn;
538 		break;
539 	}
540 	case IB_QPT_SMI:
541 	case IB_QPT_GSI:
542 	{
543 		qp = kzalloc(sizeof(*qp), GFP_KERNEL);
544 		if (!qp)
545 			return ERR_PTR(-ENOMEM);
546 
547 		qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 : 1;
548 
549 		err = mthca_alloc_sqp(to_mdev(pd->device), to_mpd(pd),
550 				      to_mcq(init_attr->send_cq),
551 				      to_mcq(init_attr->recv_cq),
552 				      init_attr->sq_sig_type, &init_attr->cap,
553 				      qp->ibqp.qp_num, init_attr->port_num,
554 				      to_msqp(qp), udata);
555 		break;
556 	}
557 	default:
558 		/* Don't support raw QPs */
559 		return ERR_PTR(-ENOSYS);
560 	}
561 
562 	if (err) {
563 		kfree(qp);
564 		return ERR_PTR(err);
565 	}
566 
567 	init_attr->cap.max_send_wr     = qp->sq.max;
568 	init_attr->cap.max_recv_wr     = qp->rq.max;
569 	init_attr->cap.max_send_sge    = qp->sq.max_gs;
570 	init_attr->cap.max_recv_sge    = qp->rq.max_gs;
571 	init_attr->cap.max_inline_data = qp->max_inline_data;
572 
573 	return &qp->ibqp;
574 }
575 
576 static int mthca_destroy_qp(struct ib_qp *qp, struct ib_udata *udata)
577 {
578 	if (udata) {
579 		struct mthca_ucontext *context =
580 			rdma_udata_to_drv_context(
581 				udata,
582 				struct mthca_ucontext,
583 				ibucontext);
584 
585 		mthca_unmap_user_db(to_mdev(qp->device),
586 				    &context->uar,
587 				    context->db_tab,
588 				    to_mqp(qp)->sq.db_index);
589 		mthca_unmap_user_db(to_mdev(qp->device),
590 				    &context->uar,
591 				    context->db_tab,
592 				    to_mqp(qp)->rq.db_index);
593 	}
594 	mthca_free_qp(to_mdev(qp->device), to_mqp(qp));
595 	kfree(to_mqp(qp));
596 	return 0;
597 }
598 
599 static int mthca_create_cq(struct ib_cq *ibcq,
600 			   const struct ib_cq_init_attr *attr,
601 			   struct ib_udata *udata)
602 {
603 	struct ib_device *ibdev = ibcq->device;
604 	int entries = attr->cqe;
605 	struct mthca_create_cq ucmd;
606 	struct mthca_cq *cq;
607 	int nent;
608 	int err;
609 	struct mthca_ucontext *context = rdma_udata_to_drv_context(
610 		udata, struct mthca_ucontext, ibucontext);
611 
612 	if (attr->flags)
613 		return -EOPNOTSUPP;
614 
615 	if (entries < 1 || entries > to_mdev(ibdev)->limits.max_cqes)
616 		return -EINVAL;
617 
618 	if (udata) {
619 		if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd)))
620 			return -EFAULT;
621 
622 		err = mthca_map_user_db(to_mdev(ibdev), &context->uar,
623 					context->db_tab, ucmd.set_db_index,
624 					ucmd.set_db_page);
625 		if (err)
626 			return err;
627 
628 		err = mthca_map_user_db(to_mdev(ibdev), &context->uar,
629 					context->db_tab, ucmd.arm_db_index,
630 					ucmd.arm_db_page);
631 		if (err)
632 			goto err_unmap_set;
633 	}
634 
635 	cq = to_mcq(ibcq);
636 
637 	if (udata) {
638 		cq->buf.mr.ibmr.lkey = ucmd.lkey;
639 		cq->set_ci_db_index  = ucmd.set_db_index;
640 		cq->arm_db_index     = ucmd.arm_db_index;
641 	}
642 
643 	for (nent = 1; nent <= entries; nent <<= 1)
644 		; /* nothing */
645 
646 	err = mthca_init_cq(to_mdev(ibdev), nent, context,
647 			    udata ? ucmd.pdn : to_mdev(ibdev)->driver_pd.pd_num,
648 			    cq);
649 	if (err)
650 		goto err_unmap_arm;
651 
652 	if (udata && ib_copy_to_udata(udata, &cq->cqn, sizeof(__u32))) {
653 		mthca_free_cq(to_mdev(ibdev), cq);
654 		err = -EFAULT;
655 		goto err_unmap_arm;
656 	}
657 
658 	cq->resize_buf = NULL;
659 
660 	return 0;
661 
662 err_unmap_arm:
663 	if (udata)
664 		mthca_unmap_user_db(to_mdev(ibdev), &context->uar,
665 				    context->db_tab, ucmd.arm_db_index);
666 
667 err_unmap_set:
668 	if (udata)
669 		mthca_unmap_user_db(to_mdev(ibdev), &context->uar,
670 				    context->db_tab, ucmd.set_db_index);
671 
672 	return err;
673 }
674 
675 static int mthca_alloc_resize_buf(struct mthca_dev *dev, struct mthca_cq *cq,
676 				  int entries)
677 {
678 	int ret;
679 
680 	spin_lock_irq(&cq->lock);
681 	if (cq->resize_buf) {
682 		ret = -EBUSY;
683 		goto unlock;
684 	}
685 
686 	cq->resize_buf = kmalloc(sizeof *cq->resize_buf, GFP_ATOMIC);
687 	if (!cq->resize_buf) {
688 		ret = -ENOMEM;
689 		goto unlock;
690 	}
691 
692 	cq->resize_buf->state = CQ_RESIZE_ALLOC;
693 
694 	ret = 0;
695 
696 unlock:
697 	spin_unlock_irq(&cq->lock);
698 
699 	if (ret)
700 		return ret;
701 
702 	ret = mthca_alloc_cq_buf(dev, &cq->resize_buf->buf, entries);
703 	if (ret) {
704 		spin_lock_irq(&cq->lock);
705 		kfree(cq->resize_buf);
706 		cq->resize_buf = NULL;
707 		spin_unlock_irq(&cq->lock);
708 		return ret;
709 	}
710 
711 	cq->resize_buf->cqe = entries - 1;
712 
713 	spin_lock_irq(&cq->lock);
714 	cq->resize_buf->state = CQ_RESIZE_READY;
715 	spin_unlock_irq(&cq->lock);
716 
717 	return 0;
718 }
719 
720 static int mthca_resize_cq(struct ib_cq *ibcq, int entries, struct ib_udata *udata)
721 {
722 	struct mthca_dev *dev = to_mdev(ibcq->device);
723 	struct mthca_cq *cq = to_mcq(ibcq);
724 	struct mthca_resize_cq ucmd;
725 	u32 lkey;
726 	int ret;
727 
728 	if (entries < 1 || entries > dev->limits.max_cqes)
729 		return -EINVAL;
730 
731 	mutex_lock(&cq->mutex);
732 
733 	entries = roundup_pow_of_two(entries + 1);
734 	if (entries == ibcq->cqe + 1) {
735 		ret = 0;
736 		goto out;
737 	}
738 
739 	if (cq->is_kernel) {
740 		ret = mthca_alloc_resize_buf(dev, cq, entries);
741 		if (ret)
742 			goto out;
743 		lkey = cq->resize_buf->buf.mr.ibmr.lkey;
744 	} else {
745 		if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
746 			ret = -EFAULT;
747 			goto out;
748 		}
749 		lkey = ucmd.lkey;
750 	}
751 
752 	ret = mthca_RESIZE_CQ(dev, cq->cqn, lkey, ilog2(entries));
753 
754 	if (ret) {
755 		if (cq->resize_buf) {
756 			mthca_free_cq_buf(dev, &cq->resize_buf->buf,
757 					  cq->resize_buf->cqe);
758 			kfree(cq->resize_buf);
759 			spin_lock_irq(&cq->lock);
760 			cq->resize_buf = NULL;
761 			spin_unlock_irq(&cq->lock);
762 		}
763 		goto out;
764 	}
765 
766 	if (cq->is_kernel) {
767 		struct mthca_cq_buf tbuf;
768 		int tcqe;
769 
770 		spin_lock_irq(&cq->lock);
771 		if (cq->resize_buf->state == CQ_RESIZE_READY) {
772 			mthca_cq_resize_copy_cqes(cq);
773 			tbuf         = cq->buf;
774 			tcqe         = cq->ibcq.cqe;
775 			cq->buf      = cq->resize_buf->buf;
776 			cq->ibcq.cqe = cq->resize_buf->cqe;
777 		} else {
778 			tbuf = cq->resize_buf->buf;
779 			tcqe = cq->resize_buf->cqe;
780 		}
781 
782 		kfree(cq->resize_buf);
783 		cq->resize_buf = NULL;
784 		spin_unlock_irq(&cq->lock);
785 
786 		mthca_free_cq_buf(dev, &tbuf, tcqe);
787 	} else
788 		ibcq->cqe = entries - 1;
789 
790 out:
791 	mutex_unlock(&cq->mutex);
792 
793 	return ret;
794 }
795 
796 static void mthca_destroy_cq(struct ib_cq *cq, struct ib_udata *udata)
797 {
798 	if (udata) {
799 		struct mthca_ucontext *context =
800 			rdma_udata_to_drv_context(
801 				udata,
802 				struct mthca_ucontext,
803 				ibucontext);
804 
805 		mthca_unmap_user_db(to_mdev(cq->device),
806 				    &context->uar,
807 				    context->db_tab,
808 				    to_mcq(cq)->arm_db_index);
809 		mthca_unmap_user_db(to_mdev(cq->device),
810 				    &context->uar,
811 				    context->db_tab,
812 				    to_mcq(cq)->set_ci_db_index);
813 	}
814 	mthca_free_cq(to_mdev(cq->device), to_mcq(cq));
815 }
816 
817 static inline u32 convert_access(int acc)
818 {
819 	return (acc & IB_ACCESS_REMOTE_ATOMIC ? MTHCA_MPT_FLAG_ATOMIC       : 0) |
820 	       (acc & IB_ACCESS_REMOTE_WRITE  ? MTHCA_MPT_FLAG_REMOTE_WRITE : 0) |
821 	       (acc & IB_ACCESS_REMOTE_READ   ? MTHCA_MPT_FLAG_REMOTE_READ  : 0) |
822 	       (acc & IB_ACCESS_LOCAL_WRITE   ? MTHCA_MPT_FLAG_LOCAL_WRITE  : 0) |
823 	       MTHCA_MPT_FLAG_LOCAL_READ;
824 }
825 
826 static struct ib_mr *mthca_get_dma_mr(struct ib_pd *pd, int acc)
827 {
828 	struct mthca_mr *mr;
829 	int err;
830 
831 	mr = kmalloc(sizeof *mr, GFP_KERNEL);
832 	if (!mr)
833 		return ERR_PTR(-ENOMEM);
834 
835 	err = mthca_mr_alloc_notrans(to_mdev(pd->device),
836 				     to_mpd(pd)->pd_num,
837 				     convert_access(acc), mr);
838 
839 	if (err) {
840 		kfree(mr);
841 		return ERR_PTR(err);
842 	}
843 
844 	mr->umem = NULL;
845 
846 	return &mr->ibmr;
847 }
848 
849 static struct ib_mr *mthca_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
850 				       u64 virt, int acc, struct ib_udata *udata)
851 {
852 	struct mthca_dev *dev = to_mdev(pd->device);
853 	struct scatterlist *sg;
854 	struct mthca_mr *mr;
855 	struct mthca_reg_mr ucmd;
856 	u64 *pages;
857 	int shift, n, len;
858 	int i, k, entry;
859 	int err = 0;
860 	int write_mtt_size;
861 
862 	if (udata->inlen - sizeof (struct ib_uverbs_cmd_hdr) < sizeof ucmd) {
863 		if (!to_mucontext(pd->uobject->context)->reg_mr_warned) {
864 			mthca_warn(dev, "Process '%s' did not pass in MR attrs.\n",
865 				   current->comm);
866 			mthca_warn(dev, "  Update libmthca to fix this.\n");
867 		}
868 		++to_mucontext(pd->uobject->context)->reg_mr_warned;
869 		ucmd.mr_attrs = 0;
870 	} else if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd))
871 		return ERR_PTR(-EFAULT);
872 
873 	mr = kmalloc(sizeof *mr, GFP_KERNEL);
874 	if (!mr)
875 		return ERR_PTR(-ENOMEM);
876 
877 	mr->umem = ib_umem_get(pd->uobject->context, start, length, acc,
878 			       ucmd.mr_attrs & MTHCA_MR_DMASYNC);
879 
880 	if (IS_ERR(mr->umem)) {
881 		err = PTR_ERR(mr->umem);
882 		goto err;
883 	}
884 
885 	shift = ffs(mr->umem->page_size) - 1;
886 	n = mr->umem->nmap;
887 
888 	mr->mtt = mthca_alloc_mtt(dev, n);
889 	if (IS_ERR(mr->mtt)) {
890 		err = PTR_ERR(mr->mtt);
891 		goto err_umem;
892 	}
893 
894 	pages = (u64 *) __get_free_page(GFP_KERNEL);
895 	if (!pages) {
896 		err = -ENOMEM;
897 		goto err_mtt;
898 	}
899 
900 	i = n = 0;
901 
902 	write_mtt_size = min(mthca_write_mtt_size(dev), (int) (PAGE_SIZE / sizeof *pages));
903 
904 	for_each_sg(mr->umem->sg_head.sgl, sg, mr->umem->nmap, entry) {
905 		len = sg_dma_len(sg) >> shift;
906 		for (k = 0; k < len; ++k) {
907 			pages[i++] = sg_dma_address(sg) +
908 				mr->umem->page_size * k;
909 			/*
910 			 * Be friendly to write_mtt and pass it chunks
911 			 * of appropriate size.
912 			 */
913 			if (i == write_mtt_size) {
914 				err = mthca_write_mtt(dev, mr->mtt, n, pages, i);
915 				if (err)
916 					goto mtt_done;
917 				n += i;
918 				i = 0;
919 			}
920 		}
921 	}
922 
923 	if (i)
924 		err = mthca_write_mtt(dev, mr->mtt, n, pages, i);
925 mtt_done:
926 	free_page((unsigned long) pages);
927 	if (err)
928 		goto err_mtt;
929 
930 	err = mthca_mr_alloc(dev, to_mpd(pd)->pd_num, shift, virt, length,
931 			     convert_access(acc), mr);
932 
933 	if (err)
934 		goto err_mtt;
935 
936 	return &mr->ibmr;
937 
938 err_mtt:
939 	mthca_free_mtt(dev, mr->mtt);
940 
941 err_umem:
942 	ib_umem_release(mr->umem);
943 
944 err:
945 	kfree(mr);
946 	return ERR_PTR(err);
947 }
948 
949 static int mthca_dereg_mr(struct ib_mr *mr, struct ib_udata *udata)
950 {
951 	struct mthca_mr *mmr = to_mmr(mr);
952 
953 	mthca_free_mr(to_mdev(mr->device), mmr);
954 	ib_umem_release(mmr->umem);
955 	kfree(mmr);
956 
957 	return 0;
958 }
959 
960 static struct ib_fmr *mthca_alloc_fmr(struct ib_pd *pd, int mr_access_flags,
961 				      struct ib_fmr_attr *fmr_attr)
962 {
963 	struct mthca_fmr *fmr;
964 	int err;
965 
966 	fmr = kmalloc(sizeof *fmr, GFP_KERNEL);
967 	if (!fmr)
968 		return ERR_PTR(-ENOMEM);
969 
970 	memcpy(&fmr->attr, fmr_attr, sizeof *fmr_attr);
971 	err = mthca_fmr_alloc(to_mdev(pd->device), to_mpd(pd)->pd_num,
972 			     convert_access(mr_access_flags), fmr);
973 
974 	if (err) {
975 		kfree(fmr);
976 		return ERR_PTR(err);
977 	}
978 
979 	return &fmr->ibmr;
980 }
981 
982 static int mthca_dealloc_fmr(struct ib_fmr *fmr)
983 {
984 	struct mthca_fmr *mfmr = to_mfmr(fmr);
985 	int err;
986 
987 	err = mthca_free_fmr(to_mdev(fmr->device), mfmr);
988 	if (err)
989 		return err;
990 
991 	kfree(mfmr);
992 	return 0;
993 }
994 
995 static int mthca_unmap_fmr(struct list_head *fmr_list)
996 {
997 	struct ib_fmr *fmr;
998 	int err;
999 	struct mthca_dev *mdev = NULL;
1000 
1001 	list_for_each_entry(fmr, fmr_list, list) {
1002 		if (mdev && to_mdev(fmr->device) != mdev)
1003 			return -EINVAL;
1004 		mdev = to_mdev(fmr->device);
1005 	}
1006 
1007 	if (!mdev)
1008 		return 0;
1009 
1010 	if (mthca_is_memfree(mdev)) {
1011 		list_for_each_entry(fmr, fmr_list, list)
1012 			mthca_arbel_fmr_unmap(mdev, to_mfmr(fmr));
1013 
1014 		wmb();
1015 	} else
1016 		list_for_each_entry(fmr, fmr_list, list)
1017 			mthca_tavor_fmr_unmap(mdev, to_mfmr(fmr));
1018 
1019 	err = mthca_SYNC_TPT(mdev);
1020 	return err;
1021 }
1022 
1023 static ssize_t show_rev(struct device *device, struct device_attribute *attr,
1024 			char *buf)
1025 {
1026 	struct mthca_dev *dev =
1027 		container_of(device, struct mthca_dev, ib_dev.dev);
1028 	return sprintf(buf, "%x\n", dev->rev_id);
1029 }
1030 
1031 static ssize_t show_hca(struct device *device, struct device_attribute *attr,
1032 			char *buf)
1033 {
1034 	struct mthca_dev *dev =
1035 		container_of(device, struct mthca_dev, ib_dev.dev);
1036 	switch (dev->pdev->device) {
1037 	case PCI_DEVICE_ID_MELLANOX_TAVOR:
1038 		return sprintf(buf, "MT23108\n");
1039 	case PCI_DEVICE_ID_MELLANOX_ARBEL_COMPAT:
1040 		return sprintf(buf, "MT25208 (MT23108 compat mode)\n");
1041 	case PCI_DEVICE_ID_MELLANOX_ARBEL:
1042 		return sprintf(buf, "MT25208\n");
1043 	case PCI_DEVICE_ID_MELLANOX_SINAI:
1044 	case PCI_DEVICE_ID_MELLANOX_SINAI_OLD:
1045 		return sprintf(buf, "MT25204\n");
1046 	default:
1047 		return sprintf(buf, "unknown\n");
1048 	}
1049 }
1050 
1051 static ssize_t show_board(struct device *device, struct device_attribute *attr,
1052 			  char *buf)
1053 {
1054 	struct mthca_dev *dev =
1055 		container_of(device, struct mthca_dev, ib_dev.dev);
1056 	return sprintf(buf, "%.*s\n", MTHCA_BOARD_ID_LEN, dev->board_id);
1057 }
1058 
1059 static DEVICE_ATTR(hw_rev,   S_IRUGO, show_rev,    NULL);
1060 static DEVICE_ATTR(hca_type, S_IRUGO, show_hca,    NULL);
1061 static DEVICE_ATTR(board_id, S_IRUGO, show_board,  NULL);
1062 
1063 static struct device_attribute *mthca_dev_attributes[] = {
1064 	&dev_attr_hw_rev,
1065 	&dev_attr_hca_type,
1066 	&dev_attr_board_id
1067 };
1068 
1069 static int mthca_init_node_data(struct mthca_dev *dev)
1070 {
1071 	struct ib_smp *in_mad  = NULL;
1072 	struct ib_smp *out_mad = NULL;
1073 	int err = -ENOMEM;
1074 
1075 	in_mad  = kzalloc(sizeof *in_mad, GFP_KERNEL);
1076 	out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
1077 	if (!in_mad || !out_mad)
1078 		goto out;
1079 
1080 	init_query_mad(in_mad);
1081 	in_mad->attr_id = IB_SMP_ATTR_NODE_DESC;
1082 
1083 	err = mthca_MAD_IFC(dev, 1, 1,
1084 			    1, NULL, NULL, in_mad, out_mad);
1085 	if (err)
1086 		goto out;
1087 
1088 	memcpy(dev->ib_dev.node_desc, out_mad->data, IB_DEVICE_NODE_DESC_MAX);
1089 
1090 	in_mad->attr_id = IB_SMP_ATTR_NODE_INFO;
1091 
1092 	err = mthca_MAD_IFC(dev, 1, 1,
1093 			    1, NULL, NULL, in_mad, out_mad);
1094 	if (err)
1095 		goto out;
1096 
1097 	if (mthca_is_memfree(dev))
1098 		dev->rev_id = be32_to_cpup((__be32 *) (out_mad->data + 32));
1099 	memcpy(&dev->ib_dev.node_guid, out_mad->data + 12, 8);
1100 
1101 out:
1102 	kfree(in_mad);
1103 	kfree(out_mad);
1104 	return err;
1105 }
1106 
1107 static int mthca_port_immutable(struct ib_device *ibdev, u8 port_num,
1108 			        struct ib_port_immutable *immutable)
1109 {
1110 	struct ib_port_attr attr;
1111 	int err;
1112 
1113 	err = mthca_query_port(ibdev, port_num, &attr);
1114 	if (err)
1115 		return err;
1116 
1117 	immutable->pkey_tbl_len = attr.pkey_tbl_len;
1118 	immutable->gid_tbl_len = attr.gid_tbl_len;
1119 	immutable->core_cap_flags = RDMA_CORE_PORT_IBA_IB;
1120 	immutable->max_mad_size = IB_MGMT_MAD_SIZE;
1121 
1122 	return 0;
1123 }
1124 
1125 static void get_dev_fw_str(struct ib_device *device, char *str,
1126 			   size_t str_len)
1127 {
1128 	struct mthca_dev *dev =
1129 		container_of(device, struct mthca_dev, ib_dev);
1130 	snprintf(str, str_len, "%d.%d.%d",
1131 		 (int) (dev->fw_ver >> 32),
1132 		 (int) (dev->fw_ver >> 16) & 0xffff,
1133 		 (int) dev->fw_ver & 0xffff);
1134 }
1135 
1136 int mthca_register_device(struct mthca_dev *dev)
1137 {
1138 	int ret;
1139 	int i;
1140 
1141 	ret = mthca_init_node_data(dev);
1142 	if (ret)
1143 		return ret;
1144 
1145 #define	mthca_ib_ah mthca_ah
1146 #define	mthca_ib_cq mthca_cq
1147 #define	mthca_ib_pd mthca_pd
1148 #define	mthca_ib_qp mthca_qp
1149 #define	mthca_ib_srq mthca_srq
1150 #define	mthca_ib_ucontext mthca_ucontext
1151 	INIT_IB_DEVICE_OPS(&dev->ib_dev.ops, mthca, MTHCA);
1152 	strlcpy(dev->ib_dev.name, "mthca%d", IB_DEVICE_NAME_MAX);
1153 	dev->ib_dev.owner                = THIS_MODULE;
1154 
1155 	dev->ib_dev.uverbs_abi_ver	 = MTHCA_UVERBS_ABI_VERSION;
1156 	dev->ib_dev.uverbs_cmd_mask	 =
1157 		(1ull << IB_USER_VERBS_CMD_GET_CONTEXT)		|
1158 		(1ull << IB_USER_VERBS_CMD_QUERY_DEVICE)	|
1159 		(1ull << IB_USER_VERBS_CMD_QUERY_PORT)		|
1160 		(1ull << IB_USER_VERBS_CMD_ALLOC_PD)		|
1161 		(1ull << IB_USER_VERBS_CMD_DEALLOC_PD)		|
1162 		(1ull << IB_USER_VERBS_CMD_REG_MR)		|
1163 		(1ull << IB_USER_VERBS_CMD_DEREG_MR)		|
1164 		(1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL)	|
1165 		(1ull << IB_USER_VERBS_CMD_CREATE_CQ)		|
1166 		(1ull << IB_USER_VERBS_CMD_RESIZE_CQ)		|
1167 		(1ull << IB_USER_VERBS_CMD_DESTROY_CQ)		|
1168 		(1ull << IB_USER_VERBS_CMD_CREATE_QP)		|
1169 		(1ull << IB_USER_VERBS_CMD_QUERY_QP)		|
1170 		(1ull << IB_USER_VERBS_CMD_MODIFY_QP)		|
1171 		(1ull << IB_USER_VERBS_CMD_DESTROY_QP)		|
1172 		(1ull << IB_USER_VERBS_CMD_ATTACH_MCAST)	|
1173 		(1ull << IB_USER_VERBS_CMD_DETACH_MCAST);
1174 	dev->ib_dev.node_type            = RDMA_NODE_IB_CA;
1175 	dev->ib_dev.phys_port_cnt        = dev->limits.num_ports;
1176 	dev->ib_dev.num_comp_vectors     = 1;
1177 	dev->ib_dev.dma_device           = &dev->pdev->dev;
1178 	dev->ib_dev.query_device         = mthca_query_device;
1179 	dev->ib_dev.query_port           = mthca_query_port;
1180 	dev->ib_dev.modify_device        = mthca_modify_device;
1181 	dev->ib_dev.modify_port          = mthca_modify_port;
1182 	dev->ib_dev.query_pkey           = mthca_query_pkey;
1183 	dev->ib_dev.query_gid            = mthca_query_gid;
1184 	dev->ib_dev.alloc_ucontext       = mthca_alloc_ucontext;
1185 	dev->ib_dev.dealloc_ucontext     = mthca_dealloc_ucontext;
1186 	dev->ib_dev.mmap                 = mthca_mmap_uar;
1187 	dev->ib_dev.alloc_pd             = mthca_alloc_pd;
1188 	dev->ib_dev.dealloc_pd           = mthca_dealloc_pd;
1189 	dev->ib_dev.create_ah            = mthca_ah_create;
1190 	dev->ib_dev.query_ah             = mthca_ah_query;
1191 	dev->ib_dev.destroy_ah           = mthca_ah_destroy;
1192 
1193 	if (dev->mthca_flags & MTHCA_FLAG_SRQ) {
1194 		dev->ib_dev.create_srq           = mthca_create_srq;
1195 		dev->ib_dev.modify_srq           = mthca_modify_srq;
1196 		dev->ib_dev.query_srq            = mthca_query_srq;
1197 		dev->ib_dev.destroy_srq          = mthca_destroy_srq;
1198 		dev->ib_dev.uverbs_cmd_mask	|=
1199 			(1ull << IB_USER_VERBS_CMD_CREATE_SRQ)		|
1200 			(1ull << IB_USER_VERBS_CMD_MODIFY_SRQ)		|
1201 			(1ull << IB_USER_VERBS_CMD_QUERY_SRQ)		|
1202 			(1ull << IB_USER_VERBS_CMD_DESTROY_SRQ);
1203 
1204 		if (mthca_is_memfree(dev))
1205 			dev->ib_dev.post_srq_recv = mthca_arbel_post_srq_recv;
1206 		else
1207 			dev->ib_dev.post_srq_recv = mthca_tavor_post_srq_recv;
1208 	}
1209 
1210 	dev->ib_dev.create_qp            = mthca_create_qp;
1211 	dev->ib_dev.modify_qp            = mthca_modify_qp;
1212 	dev->ib_dev.query_qp             = mthca_query_qp;
1213 	dev->ib_dev.destroy_qp           = mthca_destroy_qp;
1214 	dev->ib_dev.create_cq            = mthca_create_cq;
1215 	dev->ib_dev.resize_cq            = mthca_resize_cq;
1216 	dev->ib_dev.destroy_cq           = mthca_destroy_cq;
1217 	dev->ib_dev.poll_cq              = mthca_poll_cq;
1218 	dev->ib_dev.get_dma_mr           = mthca_get_dma_mr;
1219 	dev->ib_dev.reg_user_mr          = mthca_reg_user_mr;
1220 	dev->ib_dev.dereg_mr             = mthca_dereg_mr;
1221 	dev->ib_dev.get_port_immutable   = mthca_port_immutable;
1222 	dev->ib_dev.get_dev_fw_str       = get_dev_fw_str;
1223 
1224 	if (dev->mthca_flags & MTHCA_FLAG_FMR) {
1225 		dev->ib_dev.alloc_fmr            = mthca_alloc_fmr;
1226 		dev->ib_dev.unmap_fmr            = mthca_unmap_fmr;
1227 		dev->ib_dev.dealloc_fmr          = mthca_dealloc_fmr;
1228 		if (mthca_is_memfree(dev))
1229 			dev->ib_dev.map_phys_fmr = mthca_arbel_map_phys_fmr;
1230 		else
1231 			dev->ib_dev.map_phys_fmr = mthca_tavor_map_phys_fmr;
1232 	}
1233 
1234 	dev->ib_dev.attach_mcast         = mthca_multicast_attach;
1235 	dev->ib_dev.detach_mcast         = mthca_multicast_detach;
1236 	dev->ib_dev.process_mad          = mthca_process_mad;
1237 
1238 	if (mthca_is_memfree(dev)) {
1239 		dev->ib_dev.req_notify_cq = mthca_arbel_arm_cq;
1240 		dev->ib_dev.post_send     = mthca_arbel_post_send;
1241 		dev->ib_dev.post_recv     = mthca_arbel_post_receive;
1242 	} else {
1243 		dev->ib_dev.req_notify_cq = mthca_tavor_arm_cq;
1244 		dev->ib_dev.post_send     = mthca_tavor_post_send;
1245 		dev->ib_dev.post_recv     = mthca_tavor_post_receive;
1246 	}
1247 
1248 	mutex_init(&dev->cap_mask_mutex);
1249 
1250 	ret = ib_register_device(&dev->ib_dev, NULL);
1251 	if (ret)
1252 		return ret;
1253 
1254 	for (i = 0; i < ARRAY_SIZE(mthca_dev_attributes); ++i) {
1255 		ret = device_create_file(&dev->ib_dev.dev,
1256 					 mthca_dev_attributes[i]);
1257 		if (ret) {
1258 			ib_unregister_device(&dev->ib_dev);
1259 			return ret;
1260 		}
1261 	}
1262 
1263 	mthca_start_catas_poll(dev);
1264 
1265 	return 0;
1266 }
1267 
1268 void mthca_unregister_device(struct mthca_dev *dev)
1269 {
1270 	mthca_stop_catas_poll(dev);
1271 	ib_unregister_device(&dev->ib_dev);
1272 }
1273