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