xref: /linux/drivers/infiniband/hw/mthca/mthca_provider.c (revision ca55b2fef3a9373fcfc30f82fd26bc7fccbda732)
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 
41 #include <linux/sched.h>
42 #include <linux/slab.h>
43 #include <linux/stat.h>
44 #include <linux/mm.h>
45 #include <linux/export.h>
46 
47 #include "mthca_dev.h"
48 #include "mthca_cmd.h"
49 #include "mthca_user.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, 64);
197 		mutex_unlock(&to_mdev(ibdev)->cap_mask_mutex);
198 	}
199 
200 	return 0;
201 }
202 
203 static int mthca_modify_port(struct ib_device *ibdev,
204 			     u8 port, int port_modify_mask,
205 			     struct ib_port_modify *props)
206 {
207 	struct mthca_set_ib_param set_ib;
208 	struct ib_port_attr attr;
209 	int err;
210 
211 	if (mutex_lock_interruptible(&to_mdev(ibdev)->cap_mask_mutex))
212 		return -ERESTARTSYS;
213 
214 	err = mthca_query_port(ibdev, port, &attr);
215 	if (err)
216 		goto out;
217 
218 	set_ib.set_si_guid     = 0;
219 	set_ib.reset_qkey_viol = !!(port_modify_mask & IB_PORT_RESET_QKEY_CNTR);
220 
221 	set_ib.cap_mask = (attr.port_cap_flags | props->set_port_cap_mask) &
222 		~props->clr_port_cap_mask;
223 
224 	err = mthca_SET_IB(to_mdev(ibdev), &set_ib, port);
225 	if (err)
226 		goto out;
227 out:
228 	mutex_unlock(&to_mdev(ibdev)->cap_mask_mutex);
229 	return err;
230 }
231 
232 static int mthca_query_pkey(struct ib_device *ibdev,
233 			    u8 port, u16 index, u16 *pkey)
234 {
235 	struct ib_smp *in_mad  = NULL;
236 	struct ib_smp *out_mad = NULL;
237 	int err = -ENOMEM;
238 
239 	in_mad  = kzalloc(sizeof *in_mad, GFP_KERNEL);
240 	out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
241 	if (!in_mad || !out_mad)
242 		goto out;
243 
244 	init_query_mad(in_mad);
245 	in_mad->attr_id  = IB_SMP_ATTR_PKEY_TABLE;
246 	in_mad->attr_mod = cpu_to_be32(index / 32);
247 
248 	err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
249 			    port, NULL, NULL, in_mad, out_mad);
250 	if (err)
251 		goto out;
252 
253 	*pkey = be16_to_cpu(((__be16 *) out_mad->data)[index % 32]);
254 
255  out:
256 	kfree(in_mad);
257 	kfree(out_mad);
258 	return err;
259 }
260 
261 static int mthca_query_gid(struct ib_device *ibdev, u8 port,
262 			   int index, union ib_gid *gid)
263 {
264 	struct ib_smp *in_mad  = NULL;
265 	struct ib_smp *out_mad = NULL;
266 	int err = -ENOMEM;
267 
268 	in_mad  = kzalloc(sizeof *in_mad, GFP_KERNEL);
269 	out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
270 	if (!in_mad || !out_mad)
271 		goto out;
272 
273 	init_query_mad(in_mad);
274 	in_mad->attr_id  = IB_SMP_ATTR_PORT_INFO;
275 	in_mad->attr_mod = cpu_to_be32(port);
276 
277 	err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
278 			    port, NULL, NULL, in_mad, out_mad);
279 	if (err)
280 		goto out;
281 
282 	memcpy(gid->raw, out_mad->data + 8, 8);
283 
284 	init_query_mad(in_mad);
285 	in_mad->attr_id  = IB_SMP_ATTR_GUID_INFO;
286 	in_mad->attr_mod = cpu_to_be32(index / 8);
287 
288 	err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
289 			    port, NULL, NULL, in_mad, out_mad);
290 	if (err)
291 		goto out;
292 
293 	memcpy(gid->raw + 8, out_mad->data + (index % 8) * 8, 8);
294 
295  out:
296 	kfree(in_mad);
297 	kfree(out_mad);
298 	return err;
299 }
300 
301 static struct ib_ucontext *mthca_alloc_ucontext(struct ib_device *ibdev,
302 						struct ib_udata *udata)
303 {
304 	struct mthca_alloc_ucontext_resp uresp;
305 	struct mthca_ucontext           *context;
306 	int                              err;
307 
308 	if (!(to_mdev(ibdev)->active))
309 		return ERR_PTR(-EAGAIN);
310 
311 	memset(&uresp, 0, sizeof uresp);
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 	context = kmalloc(sizeof *context, GFP_KERNEL);
320 	if (!context)
321 		return ERR_PTR(-ENOMEM);
322 
323 	err = mthca_uar_alloc(to_mdev(ibdev), &context->uar);
324 	if (err) {
325 		kfree(context);
326 		return ERR_PTR(err);
327 	}
328 
329 	context->db_tab = mthca_init_user_db_tab(to_mdev(ibdev));
330 	if (IS_ERR(context->db_tab)) {
331 		err = PTR_ERR(context->db_tab);
332 		mthca_uar_free(to_mdev(ibdev), &context->uar);
333 		kfree(context);
334 		return ERR_PTR(err);
335 	}
336 
337 	if (ib_copy_to_udata(udata, &uresp, sizeof uresp)) {
338 		mthca_cleanup_user_db_tab(to_mdev(ibdev), &context->uar, context->db_tab);
339 		mthca_uar_free(to_mdev(ibdev), &context->uar);
340 		kfree(context);
341 		return ERR_PTR(-EFAULT);
342 	}
343 
344 	context->reg_mr_warned = 0;
345 
346 	return &context->ibucontext;
347 }
348 
349 static int mthca_dealloc_ucontext(struct ib_ucontext *context)
350 {
351 	mthca_cleanup_user_db_tab(to_mdev(context->device), &to_mucontext(context)->uar,
352 				  to_mucontext(context)->db_tab);
353 	mthca_uar_free(to_mdev(context->device), &to_mucontext(context)->uar);
354 	kfree(to_mucontext(context));
355 
356 	return 0;
357 }
358 
359 static int mthca_mmap_uar(struct ib_ucontext *context,
360 			  struct vm_area_struct *vma)
361 {
362 	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
363 		return -EINVAL;
364 
365 	vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
366 
367 	if (io_remap_pfn_range(vma, vma->vm_start,
368 			       to_mucontext(context)->uar.pfn,
369 			       PAGE_SIZE, vma->vm_page_prot))
370 		return -EAGAIN;
371 
372 	return 0;
373 }
374 
375 static struct ib_pd *mthca_alloc_pd(struct ib_device *ibdev,
376 				    struct ib_ucontext *context,
377 				    struct ib_udata *udata)
378 {
379 	struct mthca_pd *pd;
380 	int err;
381 
382 	pd = kmalloc(sizeof *pd, GFP_KERNEL);
383 	if (!pd)
384 		return ERR_PTR(-ENOMEM);
385 
386 	err = mthca_pd_alloc(to_mdev(ibdev), !context, pd);
387 	if (err) {
388 		kfree(pd);
389 		return ERR_PTR(err);
390 	}
391 
392 	if (context) {
393 		if (ib_copy_to_udata(udata, &pd->pd_num, sizeof (__u32))) {
394 			mthca_pd_free(to_mdev(ibdev), pd);
395 			kfree(pd);
396 			return ERR_PTR(-EFAULT);
397 		}
398 	}
399 
400 	return &pd->ibpd;
401 }
402 
403 static int mthca_dealloc_pd(struct ib_pd *pd)
404 {
405 	mthca_pd_free(to_mdev(pd->device), to_mpd(pd));
406 	kfree(pd);
407 
408 	return 0;
409 }
410 
411 static struct ib_ah *mthca_ah_create(struct ib_pd *pd,
412 				     struct ib_ah_attr *ah_attr)
413 {
414 	int err;
415 	struct mthca_ah *ah;
416 
417 	ah = kmalloc(sizeof *ah, GFP_ATOMIC);
418 	if (!ah)
419 		return ERR_PTR(-ENOMEM);
420 
421 	err = mthca_create_ah(to_mdev(pd->device), to_mpd(pd), ah_attr, ah);
422 	if (err) {
423 		kfree(ah);
424 		return ERR_PTR(err);
425 	}
426 
427 	return &ah->ibah;
428 }
429 
430 static int mthca_ah_destroy(struct ib_ah *ah)
431 {
432 	mthca_destroy_ah(to_mdev(ah->device), to_mah(ah));
433 	kfree(ah);
434 
435 	return 0;
436 }
437 
438 static struct ib_srq *mthca_create_srq(struct ib_pd *pd,
439 				       struct ib_srq_init_attr *init_attr,
440 				       struct ib_udata *udata)
441 {
442 	struct mthca_create_srq ucmd;
443 	struct mthca_ucontext *context = NULL;
444 	struct mthca_srq *srq;
445 	int err;
446 
447 	if (init_attr->srq_type != IB_SRQT_BASIC)
448 		return ERR_PTR(-ENOSYS);
449 
450 	srq = kmalloc(sizeof *srq, GFP_KERNEL);
451 	if (!srq)
452 		return ERR_PTR(-ENOMEM);
453 
454 	if (pd->uobject) {
455 		context = to_mucontext(pd->uobject->context);
456 
457 		if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
458 			err = -EFAULT;
459 			goto err_free;
460 		}
461 
462 		err = mthca_map_user_db(to_mdev(pd->device), &context->uar,
463 					context->db_tab, ucmd.db_index,
464 					ucmd.db_page);
465 
466 		if (err)
467 			goto err_free;
468 
469 		srq->mr.ibmr.lkey = ucmd.lkey;
470 		srq->db_index     = ucmd.db_index;
471 	}
472 
473 	err = mthca_alloc_srq(to_mdev(pd->device), to_mpd(pd),
474 			      &init_attr->attr, srq);
475 
476 	if (err && pd->uobject)
477 		mthca_unmap_user_db(to_mdev(pd->device), &context->uar,
478 				    context->db_tab, ucmd.db_index);
479 
480 	if (err)
481 		goto err_free;
482 
483 	if (context && ib_copy_to_udata(udata, &srq->srqn, sizeof (__u32))) {
484 		mthca_free_srq(to_mdev(pd->device), srq);
485 		err = -EFAULT;
486 		goto err_free;
487 	}
488 
489 	return &srq->ibsrq;
490 
491 err_free:
492 	kfree(srq);
493 
494 	return ERR_PTR(err);
495 }
496 
497 static int mthca_destroy_srq(struct ib_srq *srq)
498 {
499 	struct mthca_ucontext *context;
500 
501 	if (srq->uobject) {
502 		context = to_mucontext(srq->uobject->context);
503 
504 		mthca_unmap_user_db(to_mdev(srq->device), &context->uar,
505 				    context->db_tab, to_msrq(srq)->db_index);
506 	}
507 
508 	mthca_free_srq(to_mdev(srq->device), to_msrq(srq));
509 	kfree(srq);
510 
511 	return 0;
512 }
513 
514 static struct ib_qp *mthca_create_qp(struct ib_pd *pd,
515 				     struct ib_qp_init_attr *init_attr,
516 				     struct ib_udata *udata)
517 {
518 	struct mthca_create_qp ucmd;
519 	struct mthca_qp *qp;
520 	int err;
521 
522 	if (init_attr->create_flags)
523 		return ERR_PTR(-EINVAL);
524 
525 	switch (init_attr->qp_type) {
526 	case IB_QPT_RC:
527 	case IB_QPT_UC:
528 	case IB_QPT_UD:
529 	{
530 		struct mthca_ucontext *context;
531 
532 		qp = kmalloc(sizeof *qp, GFP_KERNEL);
533 		if (!qp)
534 			return ERR_PTR(-ENOMEM);
535 
536 		if (pd->uobject) {
537 			context = to_mucontext(pd->uobject->context);
538 
539 			if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
540 				kfree(qp);
541 				return ERR_PTR(-EFAULT);
542 			}
543 
544 			err = mthca_map_user_db(to_mdev(pd->device), &context->uar,
545 						context->db_tab,
546 						ucmd.sq_db_index, ucmd.sq_db_page);
547 			if (err) {
548 				kfree(qp);
549 				return ERR_PTR(err);
550 			}
551 
552 			err = mthca_map_user_db(to_mdev(pd->device), &context->uar,
553 						context->db_tab,
554 						ucmd.rq_db_index, ucmd.rq_db_page);
555 			if (err) {
556 				mthca_unmap_user_db(to_mdev(pd->device),
557 						    &context->uar,
558 						    context->db_tab,
559 						    ucmd.sq_db_index);
560 				kfree(qp);
561 				return ERR_PTR(err);
562 			}
563 
564 			qp->mr.ibmr.lkey = ucmd.lkey;
565 			qp->sq.db_index  = ucmd.sq_db_index;
566 			qp->rq.db_index  = ucmd.rq_db_index;
567 		}
568 
569 		err = mthca_alloc_qp(to_mdev(pd->device), to_mpd(pd),
570 				     to_mcq(init_attr->send_cq),
571 				     to_mcq(init_attr->recv_cq),
572 				     init_attr->qp_type, init_attr->sq_sig_type,
573 				     &init_attr->cap, qp);
574 
575 		if (err && pd->uobject) {
576 			context = to_mucontext(pd->uobject->context);
577 
578 			mthca_unmap_user_db(to_mdev(pd->device),
579 					    &context->uar,
580 					    context->db_tab,
581 					    ucmd.sq_db_index);
582 			mthca_unmap_user_db(to_mdev(pd->device),
583 					    &context->uar,
584 					    context->db_tab,
585 					    ucmd.rq_db_index);
586 		}
587 
588 		qp->ibqp.qp_num = qp->qpn;
589 		break;
590 	}
591 	case IB_QPT_SMI:
592 	case IB_QPT_GSI:
593 	{
594 		/* Don't allow userspace to create special QPs */
595 		if (pd->uobject)
596 			return ERR_PTR(-EINVAL);
597 
598 		qp = kmalloc(sizeof (struct mthca_sqp), GFP_KERNEL);
599 		if (!qp)
600 			return ERR_PTR(-ENOMEM);
601 
602 		qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 : 1;
603 
604 		err = mthca_alloc_sqp(to_mdev(pd->device), to_mpd(pd),
605 				      to_mcq(init_attr->send_cq),
606 				      to_mcq(init_attr->recv_cq),
607 				      init_attr->sq_sig_type, &init_attr->cap,
608 				      qp->ibqp.qp_num, init_attr->port_num,
609 				      to_msqp(qp));
610 		break;
611 	}
612 	default:
613 		/* Don't support raw QPs */
614 		return ERR_PTR(-ENOSYS);
615 	}
616 
617 	if (err) {
618 		kfree(qp);
619 		return ERR_PTR(err);
620 	}
621 
622 	init_attr->cap.max_send_wr     = qp->sq.max;
623 	init_attr->cap.max_recv_wr     = qp->rq.max;
624 	init_attr->cap.max_send_sge    = qp->sq.max_gs;
625 	init_attr->cap.max_recv_sge    = qp->rq.max_gs;
626 	init_attr->cap.max_inline_data = qp->max_inline_data;
627 
628 	return &qp->ibqp;
629 }
630 
631 static int mthca_destroy_qp(struct ib_qp *qp)
632 {
633 	if (qp->uobject) {
634 		mthca_unmap_user_db(to_mdev(qp->device),
635 				    &to_mucontext(qp->uobject->context)->uar,
636 				    to_mucontext(qp->uobject->context)->db_tab,
637 				    to_mqp(qp)->sq.db_index);
638 		mthca_unmap_user_db(to_mdev(qp->device),
639 				    &to_mucontext(qp->uobject->context)->uar,
640 				    to_mucontext(qp->uobject->context)->db_tab,
641 				    to_mqp(qp)->rq.db_index);
642 	}
643 	mthca_free_qp(to_mdev(qp->device), to_mqp(qp));
644 	kfree(qp);
645 	return 0;
646 }
647 
648 static struct ib_cq *mthca_create_cq(struct ib_device *ibdev,
649 				     const struct ib_cq_init_attr *attr,
650 				     struct ib_ucontext *context,
651 				     struct ib_udata *udata)
652 {
653 	int entries = attr->cqe;
654 	struct mthca_create_cq ucmd;
655 	struct mthca_cq *cq;
656 	int nent;
657 	int err;
658 
659 	if (attr->flags)
660 		return ERR_PTR(-EINVAL);
661 
662 	if (entries < 1 || entries > to_mdev(ibdev)->limits.max_cqes)
663 		return ERR_PTR(-EINVAL);
664 
665 	if (context) {
666 		if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd))
667 			return ERR_PTR(-EFAULT);
668 
669 		err = mthca_map_user_db(to_mdev(ibdev), &to_mucontext(context)->uar,
670 					to_mucontext(context)->db_tab,
671 					ucmd.set_db_index, ucmd.set_db_page);
672 		if (err)
673 			return ERR_PTR(err);
674 
675 		err = mthca_map_user_db(to_mdev(ibdev), &to_mucontext(context)->uar,
676 					to_mucontext(context)->db_tab,
677 					ucmd.arm_db_index, ucmd.arm_db_page);
678 		if (err)
679 			goto err_unmap_set;
680 	}
681 
682 	cq = kmalloc(sizeof *cq, GFP_KERNEL);
683 	if (!cq) {
684 		err = -ENOMEM;
685 		goto err_unmap_arm;
686 	}
687 
688 	if (context) {
689 		cq->buf.mr.ibmr.lkey = ucmd.lkey;
690 		cq->set_ci_db_index  = ucmd.set_db_index;
691 		cq->arm_db_index     = ucmd.arm_db_index;
692 	}
693 
694 	for (nent = 1; nent <= entries; nent <<= 1)
695 		; /* nothing */
696 
697 	err = mthca_init_cq(to_mdev(ibdev), nent,
698 			    context ? to_mucontext(context) : NULL,
699 			    context ? ucmd.pdn : to_mdev(ibdev)->driver_pd.pd_num,
700 			    cq);
701 	if (err)
702 		goto err_free;
703 
704 	if (context && ib_copy_to_udata(udata, &cq->cqn, sizeof (__u32))) {
705 		mthca_free_cq(to_mdev(ibdev), cq);
706 		err = -EFAULT;
707 		goto err_free;
708 	}
709 
710 	cq->resize_buf = NULL;
711 
712 	return &cq->ibcq;
713 
714 err_free:
715 	kfree(cq);
716 
717 err_unmap_arm:
718 	if (context)
719 		mthca_unmap_user_db(to_mdev(ibdev), &to_mucontext(context)->uar,
720 				    to_mucontext(context)->db_tab, ucmd.arm_db_index);
721 
722 err_unmap_set:
723 	if (context)
724 		mthca_unmap_user_db(to_mdev(ibdev), &to_mucontext(context)->uar,
725 				    to_mucontext(context)->db_tab, ucmd.set_db_index);
726 
727 	return ERR_PTR(err);
728 }
729 
730 static int mthca_alloc_resize_buf(struct mthca_dev *dev, struct mthca_cq *cq,
731 				  int entries)
732 {
733 	int ret;
734 
735 	spin_lock_irq(&cq->lock);
736 	if (cq->resize_buf) {
737 		ret = -EBUSY;
738 		goto unlock;
739 	}
740 
741 	cq->resize_buf = kmalloc(sizeof *cq->resize_buf, GFP_ATOMIC);
742 	if (!cq->resize_buf) {
743 		ret = -ENOMEM;
744 		goto unlock;
745 	}
746 
747 	cq->resize_buf->state = CQ_RESIZE_ALLOC;
748 
749 	ret = 0;
750 
751 unlock:
752 	spin_unlock_irq(&cq->lock);
753 
754 	if (ret)
755 		return ret;
756 
757 	ret = mthca_alloc_cq_buf(dev, &cq->resize_buf->buf, entries);
758 	if (ret) {
759 		spin_lock_irq(&cq->lock);
760 		kfree(cq->resize_buf);
761 		cq->resize_buf = NULL;
762 		spin_unlock_irq(&cq->lock);
763 		return ret;
764 	}
765 
766 	cq->resize_buf->cqe = entries - 1;
767 
768 	spin_lock_irq(&cq->lock);
769 	cq->resize_buf->state = CQ_RESIZE_READY;
770 	spin_unlock_irq(&cq->lock);
771 
772 	return 0;
773 }
774 
775 static int mthca_resize_cq(struct ib_cq *ibcq, int entries, struct ib_udata *udata)
776 {
777 	struct mthca_dev *dev = to_mdev(ibcq->device);
778 	struct mthca_cq *cq = to_mcq(ibcq);
779 	struct mthca_resize_cq ucmd;
780 	u32 lkey;
781 	int ret;
782 
783 	if (entries < 1 || entries > dev->limits.max_cqes)
784 		return -EINVAL;
785 
786 	mutex_lock(&cq->mutex);
787 
788 	entries = roundup_pow_of_two(entries + 1);
789 	if (entries == ibcq->cqe + 1) {
790 		ret = 0;
791 		goto out;
792 	}
793 
794 	if (cq->is_kernel) {
795 		ret = mthca_alloc_resize_buf(dev, cq, entries);
796 		if (ret)
797 			goto out;
798 		lkey = cq->resize_buf->buf.mr.ibmr.lkey;
799 	} else {
800 		if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
801 			ret = -EFAULT;
802 			goto out;
803 		}
804 		lkey = ucmd.lkey;
805 	}
806 
807 	ret = mthca_RESIZE_CQ(dev, cq->cqn, lkey, ilog2(entries));
808 
809 	if (ret) {
810 		if (cq->resize_buf) {
811 			mthca_free_cq_buf(dev, &cq->resize_buf->buf,
812 					  cq->resize_buf->cqe);
813 			kfree(cq->resize_buf);
814 			spin_lock_irq(&cq->lock);
815 			cq->resize_buf = NULL;
816 			spin_unlock_irq(&cq->lock);
817 		}
818 		goto out;
819 	}
820 
821 	if (cq->is_kernel) {
822 		struct mthca_cq_buf tbuf;
823 		int tcqe;
824 
825 		spin_lock_irq(&cq->lock);
826 		if (cq->resize_buf->state == CQ_RESIZE_READY) {
827 			mthca_cq_resize_copy_cqes(cq);
828 			tbuf         = cq->buf;
829 			tcqe         = cq->ibcq.cqe;
830 			cq->buf      = cq->resize_buf->buf;
831 			cq->ibcq.cqe = cq->resize_buf->cqe;
832 		} else {
833 			tbuf = cq->resize_buf->buf;
834 			tcqe = cq->resize_buf->cqe;
835 		}
836 
837 		kfree(cq->resize_buf);
838 		cq->resize_buf = NULL;
839 		spin_unlock_irq(&cq->lock);
840 
841 		mthca_free_cq_buf(dev, &tbuf, tcqe);
842 	} else
843 		ibcq->cqe = entries - 1;
844 
845 out:
846 	mutex_unlock(&cq->mutex);
847 
848 	return ret;
849 }
850 
851 static int mthca_destroy_cq(struct ib_cq *cq)
852 {
853 	if (cq->uobject) {
854 		mthca_unmap_user_db(to_mdev(cq->device),
855 				    &to_mucontext(cq->uobject->context)->uar,
856 				    to_mucontext(cq->uobject->context)->db_tab,
857 				    to_mcq(cq)->arm_db_index);
858 		mthca_unmap_user_db(to_mdev(cq->device),
859 				    &to_mucontext(cq->uobject->context)->uar,
860 				    to_mucontext(cq->uobject->context)->db_tab,
861 				    to_mcq(cq)->set_ci_db_index);
862 	}
863 	mthca_free_cq(to_mdev(cq->device), to_mcq(cq));
864 	kfree(cq);
865 
866 	return 0;
867 }
868 
869 static inline u32 convert_access(int acc)
870 {
871 	return (acc & IB_ACCESS_REMOTE_ATOMIC ? MTHCA_MPT_FLAG_ATOMIC       : 0) |
872 	       (acc & IB_ACCESS_REMOTE_WRITE  ? MTHCA_MPT_FLAG_REMOTE_WRITE : 0) |
873 	       (acc & IB_ACCESS_REMOTE_READ   ? MTHCA_MPT_FLAG_REMOTE_READ  : 0) |
874 	       (acc & IB_ACCESS_LOCAL_WRITE   ? MTHCA_MPT_FLAG_LOCAL_WRITE  : 0) |
875 	       MTHCA_MPT_FLAG_LOCAL_READ;
876 }
877 
878 static struct ib_mr *mthca_get_dma_mr(struct ib_pd *pd, int acc)
879 {
880 	struct mthca_mr *mr;
881 	int err;
882 
883 	mr = kmalloc(sizeof *mr, GFP_KERNEL);
884 	if (!mr)
885 		return ERR_PTR(-ENOMEM);
886 
887 	err = mthca_mr_alloc_notrans(to_mdev(pd->device),
888 				     to_mpd(pd)->pd_num,
889 				     convert_access(acc), mr);
890 
891 	if (err) {
892 		kfree(mr);
893 		return ERR_PTR(err);
894 	}
895 
896 	mr->umem = NULL;
897 
898 	return &mr->ibmr;
899 }
900 
901 static struct ib_mr *mthca_reg_phys_mr(struct ib_pd       *pd,
902 				       struct ib_phys_buf *buffer_list,
903 				       int                 num_phys_buf,
904 				       int                 acc,
905 				       u64                *iova_start)
906 {
907 	struct mthca_mr *mr;
908 	u64 *page_list;
909 	u64 total_size;
910 	unsigned long mask;
911 	int shift;
912 	int npages;
913 	int err;
914 	int i, j, n;
915 
916 	mask = buffer_list[0].addr ^ *iova_start;
917 	total_size = 0;
918 	for (i = 0; i < num_phys_buf; ++i) {
919 		if (i != 0)
920 			mask |= buffer_list[i].addr;
921 		if (i != num_phys_buf - 1)
922 			mask |= buffer_list[i].addr + buffer_list[i].size;
923 
924 		total_size += buffer_list[i].size;
925 	}
926 
927 	if (mask & ~PAGE_MASK)
928 		return ERR_PTR(-EINVAL);
929 
930 	shift = __ffs(mask | 1 << 31);
931 
932 	buffer_list[0].size += buffer_list[0].addr & ((1ULL << shift) - 1);
933 	buffer_list[0].addr &= ~0ull << shift;
934 
935 	mr = kmalloc(sizeof *mr, GFP_KERNEL);
936 	if (!mr)
937 		return ERR_PTR(-ENOMEM);
938 
939 	npages = 0;
940 	for (i = 0; i < num_phys_buf; ++i)
941 		npages += (buffer_list[i].size + (1ULL << shift) - 1) >> shift;
942 
943 	if (!npages)
944 		return &mr->ibmr;
945 
946 	page_list = kmalloc(npages * sizeof *page_list, GFP_KERNEL);
947 	if (!page_list) {
948 		kfree(mr);
949 		return ERR_PTR(-ENOMEM);
950 	}
951 
952 	n = 0;
953 	for (i = 0; i < num_phys_buf; ++i)
954 		for (j = 0;
955 		     j < (buffer_list[i].size + (1ULL << shift) - 1) >> shift;
956 		     ++j)
957 			page_list[n++] = buffer_list[i].addr + ((u64) j << shift);
958 
959 	mthca_dbg(to_mdev(pd->device), "Registering memory at %llx (iova %llx) "
960 		  "in PD %x; shift %d, npages %d.\n",
961 		  (unsigned long long) buffer_list[0].addr,
962 		  (unsigned long long) *iova_start,
963 		  to_mpd(pd)->pd_num,
964 		  shift, npages);
965 
966 	err = mthca_mr_alloc_phys(to_mdev(pd->device),
967 				  to_mpd(pd)->pd_num,
968 				  page_list, shift, npages,
969 				  *iova_start, total_size,
970 				  convert_access(acc), mr);
971 
972 	if (err) {
973 		kfree(page_list);
974 		kfree(mr);
975 		return ERR_PTR(err);
976 	}
977 
978 	kfree(page_list);
979 	mr->umem = NULL;
980 
981 	return &mr->ibmr;
982 }
983 
984 static struct ib_mr *mthca_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
985 				       u64 virt, int acc, struct ib_udata *udata)
986 {
987 	struct mthca_dev *dev = to_mdev(pd->device);
988 	struct scatterlist *sg;
989 	struct mthca_mr *mr;
990 	struct mthca_reg_mr ucmd;
991 	u64 *pages;
992 	int shift, n, len;
993 	int i, k, entry;
994 	int err = 0;
995 	int write_mtt_size;
996 
997 	if (udata->inlen - sizeof (struct ib_uverbs_cmd_hdr) < sizeof ucmd) {
998 		if (!to_mucontext(pd->uobject->context)->reg_mr_warned) {
999 			mthca_warn(dev, "Process '%s' did not pass in MR attrs.\n",
1000 				   current->comm);
1001 			mthca_warn(dev, "  Update libmthca to fix this.\n");
1002 		}
1003 		++to_mucontext(pd->uobject->context)->reg_mr_warned;
1004 		ucmd.mr_attrs = 0;
1005 	} else if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd))
1006 		return ERR_PTR(-EFAULT);
1007 
1008 	mr = kmalloc(sizeof *mr, GFP_KERNEL);
1009 	if (!mr)
1010 		return ERR_PTR(-ENOMEM);
1011 
1012 	mr->umem = ib_umem_get(pd->uobject->context, start, length, acc,
1013 			       ucmd.mr_attrs & MTHCA_MR_DMASYNC);
1014 
1015 	if (IS_ERR(mr->umem)) {
1016 		err = PTR_ERR(mr->umem);
1017 		goto err;
1018 	}
1019 
1020 	shift = ffs(mr->umem->page_size) - 1;
1021 	n = mr->umem->nmap;
1022 
1023 	mr->mtt = mthca_alloc_mtt(dev, n);
1024 	if (IS_ERR(mr->mtt)) {
1025 		err = PTR_ERR(mr->mtt);
1026 		goto err_umem;
1027 	}
1028 
1029 	pages = (u64 *) __get_free_page(GFP_KERNEL);
1030 	if (!pages) {
1031 		err = -ENOMEM;
1032 		goto err_mtt;
1033 	}
1034 
1035 	i = n = 0;
1036 
1037 	write_mtt_size = min(mthca_write_mtt_size(dev), (int) (PAGE_SIZE / sizeof *pages));
1038 
1039 	for_each_sg(mr->umem->sg_head.sgl, sg, mr->umem->nmap, entry) {
1040 		len = sg_dma_len(sg) >> shift;
1041 		for (k = 0; k < len; ++k) {
1042 			pages[i++] = sg_dma_address(sg) +
1043 				mr->umem->page_size * k;
1044 			/*
1045 			 * Be friendly to write_mtt and pass it chunks
1046 			 * of appropriate size.
1047 			 */
1048 			if (i == write_mtt_size) {
1049 				err = mthca_write_mtt(dev, mr->mtt, n, pages, i);
1050 				if (err)
1051 					goto mtt_done;
1052 				n += i;
1053 				i = 0;
1054 			}
1055 		}
1056 	}
1057 
1058 	if (i)
1059 		err = mthca_write_mtt(dev, mr->mtt, n, pages, i);
1060 mtt_done:
1061 	free_page((unsigned long) pages);
1062 	if (err)
1063 		goto err_mtt;
1064 
1065 	err = mthca_mr_alloc(dev, to_mpd(pd)->pd_num, shift, virt, length,
1066 			     convert_access(acc), mr);
1067 
1068 	if (err)
1069 		goto err_mtt;
1070 
1071 	return &mr->ibmr;
1072 
1073 err_mtt:
1074 	mthca_free_mtt(dev, mr->mtt);
1075 
1076 err_umem:
1077 	ib_umem_release(mr->umem);
1078 
1079 err:
1080 	kfree(mr);
1081 	return ERR_PTR(err);
1082 }
1083 
1084 static int mthca_dereg_mr(struct ib_mr *mr)
1085 {
1086 	struct mthca_mr *mmr = to_mmr(mr);
1087 
1088 	mthca_free_mr(to_mdev(mr->device), mmr);
1089 	if (mmr->umem)
1090 		ib_umem_release(mmr->umem);
1091 	kfree(mmr);
1092 
1093 	return 0;
1094 }
1095 
1096 static struct ib_fmr *mthca_alloc_fmr(struct ib_pd *pd, int mr_access_flags,
1097 				      struct ib_fmr_attr *fmr_attr)
1098 {
1099 	struct mthca_fmr *fmr;
1100 	int err;
1101 
1102 	fmr = kmalloc(sizeof *fmr, GFP_KERNEL);
1103 	if (!fmr)
1104 		return ERR_PTR(-ENOMEM);
1105 
1106 	memcpy(&fmr->attr, fmr_attr, sizeof *fmr_attr);
1107 	err = mthca_fmr_alloc(to_mdev(pd->device), to_mpd(pd)->pd_num,
1108 			     convert_access(mr_access_flags), fmr);
1109 
1110 	if (err) {
1111 		kfree(fmr);
1112 		return ERR_PTR(err);
1113 	}
1114 
1115 	return &fmr->ibmr;
1116 }
1117 
1118 static int mthca_dealloc_fmr(struct ib_fmr *fmr)
1119 {
1120 	struct mthca_fmr *mfmr = to_mfmr(fmr);
1121 	int err;
1122 
1123 	err = mthca_free_fmr(to_mdev(fmr->device), mfmr);
1124 	if (err)
1125 		return err;
1126 
1127 	kfree(mfmr);
1128 	return 0;
1129 }
1130 
1131 static int mthca_unmap_fmr(struct list_head *fmr_list)
1132 {
1133 	struct ib_fmr *fmr;
1134 	int err;
1135 	struct mthca_dev *mdev = NULL;
1136 
1137 	list_for_each_entry(fmr, fmr_list, list) {
1138 		if (mdev && to_mdev(fmr->device) != mdev)
1139 			return -EINVAL;
1140 		mdev = to_mdev(fmr->device);
1141 	}
1142 
1143 	if (!mdev)
1144 		return 0;
1145 
1146 	if (mthca_is_memfree(mdev)) {
1147 		list_for_each_entry(fmr, fmr_list, list)
1148 			mthca_arbel_fmr_unmap(mdev, to_mfmr(fmr));
1149 
1150 		wmb();
1151 	} else
1152 		list_for_each_entry(fmr, fmr_list, list)
1153 			mthca_tavor_fmr_unmap(mdev, to_mfmr(fmr));
1154 
1155 	err = mthca_SYNC_TPT(mdev);
1156 	return err;
1157 }
1158 
1159 static ssize_t show_rev(struct device *device, struct device_attribute *attr,
1160 			char *buf)
1161 {
1162 	struct mthca_dev *dev =
1163 		container_of(device, struct mthca_dev, ib_dev.dev);
1164 	return sprintf(buf, "%x\n", dev->rev_id);
1165 }
1166 
1167 static ssize_t show_fw_ver(struct device *device, struct device_attribute *attr,
1168 			   char *buf)
1169 {
1170 	struct mthca_dev *dev =
1171 		container_of(device, struct mthca_dev, ib_dev.dev);
1172 	return sprintf(buf, "%d.%d.%d\n", (int) (dev->fw_ver >> 32),
1173 		       (int) (dev->fw_ver >> 16) & 0xffff,
1174 		       (int) dev->fw_ver & 0xffff);
1175 }
1176 
1177 static ssize_t show_hca(struct device *device, struct device_attribute *attr,
1178 			char *buf)
1179 {
1180 	struct mthca_dev *dev =
1181 		container_of(device, struct mthca_dev, ib_dev.dev);
1182 	switch (dev->pdev->device) {
1183 	case PCI_DEVICE_ID_MELLANOX_TAVOR:
1184 		return sprintf(buf, "MT23108\n");
1185 	case PCI_DEVICE_ID_MELLANOX_ARBEL_COMPAT:
1186 		return sprintf(buf, "MT25208 (MT23108 compat mode)\n");
1187 	case PCI_DEVICE_ID_MELLANOX_ARBEL:
1188 		return sprintf(buf, "MT25208\n");
1189 	case PCI_DEVICE_ID_MELLANOX_SINAI:
1190 	case PCI_DEVICE_ID_MELLANOX_SINAI_OLD:
1191 		return sprintf(buf, "MT25204\n");
1192 	default:
1193 		return sprintf(buf, "unknown\n");
1194 	}
1195 }
1196 
1197 static ssize_t show_board(struct device *device, struct device_attribute *attr,
1198 			  char *buf)
1199 {
1200 	struct mthca_dev *dev =
1201 		container_of(device, struct mthca_dev, ib_dev.dev);
1202 	return sprintf(buf, "%.*s\n", MTHCA_BOARD_ID_LEN, dev->board_id);
1203 }
1204 
1205 static DEVICE_ATTR(hw_rev,   S_IRUGO, show_rev,    NULL);
1206 static DEVICE_ATTR(fw_ver,   S_IRUGO, show_fw_ver, NULL);
1207 static DEVICE_ATTR(hca_type, S_IRUGO, show_hca,    NULL);
1208 static DEVICE_ATTR(board_id, S_IRUGO, show_board,  NULL);
1209 
1210 static struct device_attribute *mthca_dev_attributes[] = {
1211 	&dev_attr_hw_rev,
1212 	&dev_attr_fw_ver,
1213 	&dev_attr_hca_type,
1214 	&dev_attr_board_id
1215 };
1216 
1217 static int mthca_init_node_data(struct mthca_dev *dev)
1218 {
1219 	struct ib_smp *in_mad  = NULL;
1220 	struct ib_smp *out_mad = NULL;
1221 	int err = -ENOMEM;
1222 
1223 	in_mad  = kzalloc(sizeof *in_mad, GFP_KERNEL);
1224 	out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
1225 	if (!in_mad || !out_mad)
1226 		goto out;
1227 
1228 	init_query_mad(in_mad);
1229 	in_mad->attr_id = IB_SMP_ATTR_NODE_DESC;
1230 
1231 	err = mthca_MAD_IFC(dev, 1, 1,
1232 			    1, NULL, NULL, in_mad, out_mad);
1233 	if (err)
1234 		goto out;
1235 
1236 	memcpy(dev->ib_dev.node_desc, out_mad->data, 64);
1237 
1238 	in_mad->attr_id = IB_SMP_ATTR_NODE_INFO;
1239 
1240 	err = mthca_MAD_IFC(dev, 1, 1,
1241 			    1, NULL, NULL, in_mad, out_mad);
1242 	if (err)
1243 		goto out;
1244 
1245 	if (mthca_is_memfree(dev))
1246 		dev->rev_id = be32_to_cpup((__be32 *) (out_mad->data + 32));
1247 	memcpy(&dev->ib_dev.node_guid, out_mad->data + 12, 8);
1248 
1249 out:
1250 	kfree(in_mad);
1251 	kfree(out_mad);
1252 	return err;
1253 }
1254 
1255 static int mthca_port_immutable(struct ib_device *ibdev, u8 port_num,
1256 			        struct ib_port_immutable *immutable)
1257 {
1258 	struct ib_port_attr attr;
1259 	int err;
1260 
1261 	err = mthca_query_port(ibdev, port_num, &attr);
1262 	if (err)
1263 		return err;
1264 
1265 	immutable->pkey_tbl_len = attr.pkey_tbl_len;
1266 	immutable->gid_tbl_len = attr.gid_tbl_len;
1267 	immutable->core_cap_flags = RDMA_CORE_PORT_IBA_IB;
1268 	immutable->max_mad_size = IB_MGMT_MAD_SIZE;
1269 
1270 	return 0;
1271 }
1272 
1273 int mthca_register_device(struct mthca_dev *dev)
1274 {
1275 	int ret;
1276 	int i;
1277 
1278 	ret = mthca_init_node_data(dev);
1279 	if (ret)
1280 		return ret;
1281 
1282 	strlcpy(dev->ib_dev.name, "mthca%d", IB_DEVICE_NAME_MAX);
1283 	dev->ib_dev.owner                = THIS_MODULE;
1284 
1285 	dev->ib_dev.uverbs_abi_ver	 = MTHCA_UVERBS_ABI_VERSION;
1286 	dev->ib_dev.uverbs_cmd_mask	 =
1287 		(1ull << IB_USER_VERBS_CMD_GET_CONTEXT)		|
1288 		(1ull << IB_USER_VERBS_CMD_QUERY_DEVICE)	|
1289 		(1ull << IB_USER_VERBS_CMD_QUERY_PORT)		|
1290 		(1ull << IB_USER_VERBS_CMD_ALLOC_PD)		|
1291 		(1ull << IB_USER_VERBS_CMD_DEALLOC_PD)		|
1292 		(1ull << IB_USER_VERBS_CMD_REG_MR)		|
1293 		(1ull << IB_USER_VERBS_CMD_DEREG_MR)		|
1294 		(1ull << IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL)	|
1295 		(1ull << IB_USER_VERBS_CMD_CREATE_CQ)		|
1296 		(1ull << IB_USER_VERBS_CMD_RESIZE_CQ)		|
1297 		(1ull << IB_USER_VERBS_CMD_DESTROY_CQ)		|
1298 		(1ull << IB_USER_VERBS_CMD_CREATE_QP)		|
1299 		(1ull << IB_USER_VERBS_CMD_QUERY_QP)		|
1300 		(1ull << IB_USER_VERBS_CMD_MODIFY_QP)		|
1301 		(1ull << IB_USER_VERBS_CMD_DESTROY_QP)		|
1302 		(1ull << IB_USER_VERBS_CMD_ATTACH_MCAST)	|
1303 		(1ull << IB_USER_VERBS_CMD_DETACH_MCAST);
1304 	dev->ib_dev.node_type            = RDMA_NODE_IB_CA;
1305 	dev->ib_dev.phys_port_cnt        = dev->limits.num_ports;
1306 	dev->ib_dev.num_comp_vectors     = 1;
1307 	dev->ib_dev.dma_device           = &dev->pdev->dev;
1308 	dev->ib_dev.query_device         = mthca_query_device;
1309 	dev->ib_dev.query_port           = mthca_query_port;
1310 	dev->ib_dev.modify_device        = mthca_modify_device;
1311 	dev->ib_dev.modify_port          = mthca_modify_port;
1312 	dev->ib_dev.query_pkey           = mthca_query_pkey;
1313 	dev->ib_dev.query_gid            = mthca_query_gid;
1314 	dev->ib_dev.alloc_ucontext       = mthca_alloc_ucontext;
1315 	dev->ib_dev.dealloc_ucontext     = mthca_dealloc_ucontext;
1316 	dev->ib_dev.mmap                 = mthca_mmap_uar;
1317 	dev->ib_dev.alloc_pd             = mthca_alloc_pd;
1318 	dev->ib_dev.dealloc_pd           = mthca_dealloc_pd;
1319 	dev->ib_dev.create_ah            = mthca_ah_create;
1320 	dev->ib_dev.query_ah             = mthca_ah_query;
1321 	dev->ib_dev.destroy_ah           = mthca_ah_destroy;
1322 
1323 	if (dev->mthca_flags & MTHCA_FLAG_SRQ) {
1324 		dev->ib_dev.create_srq           = mthca_create_srq;
1325 		dev->ib_dev.modify_srq           = mthca_modify_srq;
1326 		dev->ib_dev.query_srq            = mthca_query_srq;
1327 		dev->ib_dev.destroy_srq          = mthca_destroy_srq;
1328 		dev->ib_dev.uverbs_cmd_mask	|=
1329 			(1ull << IB_USER_VERBS_CMD_CREATE_SRQ)		|
1330 			(1ull << IB_USER_VERBS_CMD_MODIFY_SRQ)		|
1331 			(1ull << IB_USER_VERBS_CMD_QUERY_SRQ)		|
1332 			(1ull << IB_USER_VERBS_CMD_DESTROY_SRQ);
1333 
1334 		if (mthca_is_memfree(dev))
1335 			dev->ib_dev.post_srq_recv = mthca_arbel_post_srq_recv;
1336 		else
1337 			dev->ib_dev.post_srq_recv = mthca_tavor_post_srq_recv;
1338 	}
1339 
1340 	dev->ib_dev.create_qp            = mthca_create_qp;
1341 	dev->ib_dev.modify_qp            = mthca_modify_qp;
1342 	dev->ib_dev.query_qp             = mthca_query_qp;
1343 	dev->ib_dev.destroy_qp           = mthca_destroy_qp;
1344 	dev->ib_dev.create_cq            = mthca_create_cq;
1345 	dev->ib_dev.resize_cq            = mthca_resize_cq;
1346 	dev->ib_dev.destroy_cq           = mthca_destroy_cq;
1347 	dev->ib_dev.poll_cq              = mthca_poll_cq;
1348 	dev->ib_dev.get_dma_mr           = mthca_get_dma_mr;
1349 	dev->ib_dev.reg_phys_mr          = mthca_reg_phys_mr;
1350 	dev->ib_dev.reg_user_mr          = mthca_reg_user_mr;
1351 	dev->ib_dev.dereg_mr             = mthca_dereg_mr;
1352 	dev->ib_dev.get_port_immutable   = mthca_port_immutable;
1353 
1354 	if (dev->mthca_flags & MTHCA_FLAG_FMR) {
1355 		dev->ib_dev.alloc_fmr            = mthca_alloc_fmr;
1356 		dev->ib_dev.unmap_fmr            = mthca_unmap_fmr;
1357 		dev->ib_dev.dealloc_fmr          = mthca_dealloc_fmr;
1358 		if (mthca_is_memfree(dev))
1359 			dev->ib_dev.map_phys_fmr = mthca_arbel_map_phys_fmr;
1360 		else
1361 			dev->ib_dev.map_phys_fmr = mthca_tavor_map_phys_fmr;
1362 	}
1363 
1364 	dev->ib_dev.attach_mcast         = mthca_multicast_attach;
1365 	dev->ib_dev.detach_mcast         = mthca_multicast_detach;
1366 	dev->ib_dev.process_mad          = mthca_process_mad;
1367 
1368 	if (mthca_is_memfree(dev)) {
1369 		dev->ib_dev.req_notify_cq = mthca_arbel_arm_cq;
1370 		dev->ib_dev.post_send     = mthca_arbel_post_send;
1371 		dev->ib_dev.post_recv     = mthca_arbel_post_receive;
1372 	} else {
1373 		dev->ib_dev.req_notify_cq = mthca_tavor_arm_cq;
1374 		dev->ib_dev.post_send     = mthca_tavor_post_send;
1375 		dev->ib_dev.post_recv     = mthca_tavor_post_receive;
1376 	}
1377 
1378 	mutex_init(&dev->cap_mask_mutex);
1379 
1380 	ret = ib_register_device(&dev->ib_dev, NULL);
1381 	if (ret)
1382 		return ret;
1383 
1384 	for (i = 0; i < ARRAY_SIZE(mthca_dev_attributes); ++i) {
1385 		ret = device_create_file(&dev->ib_dev.dev,
1386 					 mthca_dev_attributes[i]);
1387 		if (ret) {
1388 			ib_unregister_device(&dev->ib_dev);
1389 			return ret;
1390 		}
1391 	}
1392 
1393 	mthca_start_catas_poll(dev);
1394 
1395 	return 0;
1396 }
1397 
1398 void mthca_unregister_device(struct mthca_dev *dev)
1399 {
1400 	mthca_stop_catas_poll(dev);
1401 	ib_unregister_device(&dev->ib_dev);
1402 }
1403