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 int mthca_query_device(struct ib_device *ibdev, struct ib_device_attr *props,
54 			      struct ib_udata *uhw)
55 {
56 	struct ib_smp *in_mad;
57 	struct ib_smp *out_mad;
58 	int err = -ENOMEM;
59 	struct mthca_dev *mdev = to_mdev(ibdev);
60 
61 	if (uhw->inlen || uhw->outlen)
62 		return -EINVAL;
63 
64 	in_mad  = kzalloc(sizeof *in_mad, GFP_KERNEL);
65 	out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
66 	if (!in_mad || !out_mad)
67 		goto out;
68 
69 	memset(props, 0, sizeof *props);
70 
71 	props->fw_ver              = mdev->fw_ver;
72 
73 	ib_init_query_mad(in_mad);
74 	in_mad->attr_id = IB_SMP_ATTR_NODE_INFO;
75 
76 	err = mthca_MAD_IFC(mdev, 1, 1,
77 			    1, NULL, NULL, in_mad, out_mad);
78 	if (err)
79 		goto out;
80 
81 	props->device_cap_flags    = mdev->device_cap_flags;
82 	props->vendor_id           = be32_to_cpup((__be32 *) (out_mad->data + 36)) &
83 		0xffffff;
84 	props->vendor_part_id      = be16_to_cpup((__be16 *) (out_mad->data + 30));
85 	props->hw_ver              = be32_to_cpup((__be32 *) (out_mad->data + 32));
86 	memcpy(&props->sys_image_guid, out_mad->data +  4, 8);
87 
88 	props->max_mr_size         = ~0ull;
89 	props->page_size_cap       = mdev->limits.page_size_cap;
90 	props->max_qp              = mdev->limits.num_qps - mdev->limits.reserved_qps;
91 	props->max_qp_wr           = mdev->limits.max_wqes;
92 	props->max_send_sge        = mdev->limits.max_sg;
93 	props->max_recv_sge        = mdev->limits.max_sg;
94 	props->max_sge_rd          = mdev->limits.max_sg;
95 	props->max_cq              = mdev->limits.num_cqs - mdev->limits.reserved_cqs;
96 	props->max_cqe             = mdev->limits.max_cqes;
97 	props->max_mr              = mdev->limits.num_mpts - mdev->limits.reserved_mrws;
98 	props->max_pd              = mdev->limits.num_pds - mdev->limits.reserved_pds;
99 	props->max_qp_rd_atom      = 1 << mdev->qp_table.rdb_shift;
100 	props->max_qp_init_rd_atom = mdev->limits.max_qp_init_rdma;
101 	props->max_res_rd_atom     = props->max_qp_rd_atom * props->max_qp;
102 	props->max_srq             = mdev->limits.num_srqs - mdev->limits.reserved_srqs;
103 	props->max_srq_wr          = mdev->limits.max_srq_wqes;
104 	props->max_srq_sge         = mdev->limits.max_srq_sge;
105 	props->local_ca_ack_delay  = mdev->limits.local_ca_ack_delay;
106 	props->atomic_cap          = mdev->limits.flags & DEV_LIM_FLAG_ATOMIC ?
107 					IB_ATOMIC_HCA : IB_ATOMIC_NONE;
108 	props->max_pkeys           = mdev->limits.pkey_table_len;
109 	props->max_mcast_grp       = mdev->limits.num_mgms + mdev->limits.num_amgms;
110 	props->max_mcast_qp_attach = MTHCA_QP_PER_MGM;
111 	props->max_total_mcast_qp_attach = props->max_mcast_qp_attach *
112 					   props->max_mcast_grp;
113 
114 	err = 0;
115  out:
116 	kfree(in_mad);
117 	kfree(out_mad);
118 	return err;
119 }
120 
121 static int mthca_query_port(struct ib_device *ibdev,
122 			    u32 port, struct ib_port_attr *props)
123 {
124 	struct ib_smp *in_mad;
125 	struct ib_smp *out_mad;
126 	int err = -ENOMEM;
127 
128 	in_mad  = kzalloc(sizeof *in_mad, GFP_KERNEL);
129 	out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
130 	if (!in_mad || !out_mad)
131 		goto out;
132 
133 	/* props being zeroed by the caller, avoid zeroing it here */
134 
135 	ib_init_query_mad(in_mad);
136 	in_mad->attr_id  = IB_SMP_ATTR_PORT_INFO;
137 	in_mad->attr_mod = cpu_to_be32(port);
138 
139 	err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
140 			    port, NULL, NULL, in_mad, out_mad);
141 	if (err)
142 		goto out;
143 
144 	props->lid               = be16_to_cpup((__be16 *) (out_mad->data + 16));
145 	props->lmc               = out_mad->data[34] & 0x7;
146 	props->sm_lid            = be16_to_cpup((__be16 *) (out_mad->data + 18));
147 	props->sm_sl             = out_mad->data[36] & 0xf;
148 	props->state             = out_mad->data[32] & 0xf;
149 	props->phys_state        = out_mad->data[33] >> 4;
150 	props->port_cap_flags    = be32_to_cpup((__be32 *) (out_mad->data + 20));
151 	props->gid_tbl_len       = to_mdev(ibdev)->limits.gid_table_len;
152 	props->max_msg_sz        = 0x80000000;
153 	props->pkey_tbl_len      = to_mdev(ibdev)->limits.pkey_table_len;
154 	props->bad_pkey_cntr     = be16_to_cpup((__be16 *) (out_mad->data + 46));
155 	props->qkey_viol_cntr    = be16_to_cpup((__be16 *) (out_mad->data + 48));
156 	props->active_width      = out_mad->data[31] & 0xf;
157 	props->active_speed      = out_mad->data[35] >> 4;
158 	props->max_mtu           = out_mad->data[41] & 0xf;
159 	props->active_mtu        = out_mad->data[36] >> 4;
160 	props->subnet_timeout    = out_mad->data[51] & 0x1f;
161 	props->max_vl_num        = out_mad->data[37] >> 4;
162 	props->init_type_reply   = out_mad->data[41] >> 4;
163 
164  out:
165 	kfree(in_mad);
166 	kfree(out_mad);
167 	return err;
168 }
169 
170 static int mthca_modify_device(struct ib_device *ibdev,
171 			       int mask,
172 			       struct ib_device_modify *props)
173 {
174 	if (mask & ~IB_DEVICE_MODIFY_NODE_DESC)
175 		return -EOPNOTSUPP;
176 
177 	if (mask & IB_DEVICE_MODIFY_NODE_DESC) {
178 		if (mutex_lock_interruptible(&to_mdev(ibdev)->cap_mask_mutex))
179 			return -ERESTARTSYS;
180 		memcpy(ibdev->node_desc, props->node_desc,
181 		       IB_DEVICE_NODE_DESC_MAX);
182 		mutex_unlock(&to_mdev(ibdev)->cap_mask_mutex);
183 	}
184 
185 	return 0;
186 }
187 
188 static int mthca_modify_port(struct ib_device *ibdev,
189 			     u32 port, int port_modify_mask,
190 			     struct ib_port_modify *props)
191 {
192 	struct mthca_set_ib_param set_ib;
193 	struct ib_port_attr attr;
194 	int err;
195 
196 	if (mutex_lock_interruptible(&to_mdev(ibdev)->cap_mask_mutex))
197 		return -ERESTARTSYS;
198 
199 	err = ib_query_port(ibdev, port, &attr);
200 	if (err)
201 		goto out;
202 
203 	set_ib.set_si_guid     = 0;
204 	set_ib.reset_qkey_viol = !!(port_modify_mask & IB_PORT_RESET_QKEY_CNTR);
205 
206 	set_ib.cap_mask = (attr.port_cap_flags | props->set_port_cap_mask) &
207 		~props->clr_port_cap_mask;
208 
209 	err = mthca_SET_IB(to_mdev(ibdev), &set_ib, port);
210 	if (err)
211 		goto out;
212 out:
213 	mutex_unlock(&to_mdev(ibdev)->cap_mask_mutex);
214 	return err;
215 }
216 
217 static int mthca_query_pkey(struct ib_device *ibdev,
218 			    u32 port, u16 index, u16 *pkey)
219 {
220 	struct ib_smp *in_mad;
221 	struct ib_smp *out_mad;
222 	int err = -ENOMEM;
223 
224 	in_mad  = kzalloc(sizeof *in_mad, GFP_KERNEL);
225 	out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
226 	if (!in_mad || !out_mad)
227 		goto out;
228 
229 	ib_init_query_mad(in_mad);
230 	in_mad->attr_id  = IB_SMP_ATTR_PKEY_TABLE;
231 	in_mad->attr_mod = cpu_to_be32(index / 32);
232 
233 	err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
234 			    port, NULL, NULL, in_mad, out_mad);
235 	if (err)
236 		goto out;
237 
238 	*pkey = be16_to_cpu(((__be16 *) out_mad->data)[index % 32]);
239 
240  out:
241 	kfree(in_mad);
242 	kfree(out_mad);
243 	return err;
244 }
245 
246 static int mthca_query_gid(struct ib_device *ibdev, u32 port,
247 			   int index, union ib_gid *gid)
248 {
249 	struct ib_smp *in_mad;
250 	struct ib_smp *out_mad;
251 	int err = -ENOMEM;
252 
253 	in_mad  = kzalloc(sizeof *in_mad, GFP_KERNEL);
254 	out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
255 	if (!in_mad || !out_mad)
256 		goto out;
257 
258 	ib_init_query_mad(in_mad);
259 	in_mad->attr_id  = IB_SMP_ATTR_PORT_INFO;
260 	in_mad->attr_mod = cpu_to_be32(port);
261 
262 	err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
263 			    port, NULL, NULL, in_mad, out_mad);
264 	if (err)
265 		goto out;
266 
267 	memcpy(gid->raw, out_mad->data + 8, 8);
268 
269 	ib_init_query_mad(in_mad);
270 	in_mad->attr_id  = IB_SMP_ATTR_GUID_INFO;
271 	in_mad->attr_mod = cpu_to_be32(index / 8);
272 
273 	err = mthca_MAD_IFC(to_mdev(ibdev), 1, 1,
274 			    port, NULL, NULL, in_mad, out_mad);
275 	if (err)
276 		goto out;
277 
278 	memcpy(gid->raw + 8, out_mad->data + (index % 8) * 8, 8);
279 
280  out:
281 	kfree(in_mad);
282 	kfree(out_mad);
283 	return err;
284 }
285 
286 static int mthca_alloc_ucontext(struct ib_ucontext *uctx,
287 				struct ib_udata *udata)
288 {
289 	struct ib_device *ibdev = uctx->device;
290 	struct mthca_alloc_ucontext_resp uresp = {};
291 	struct mthca_ucontext *context = to_mucontext(uctx);
292 	int                              err;
293 
294 	if (!(to_mdev(ibdev)->active))
295 		return -EAGAIN;
296 
297 	uresp.qp_tab_size = to_mdev(ibdev)->limits.num_qps;
298 	if (mthca_is_memfree(to_mdev(ibdev)))
299 		uresp.uarc_size = to_mdev(ibdev)->uar_table.uarc_size;
300 	else
301 		uresp.uarc_size = 0;
302 
303 	err = mthca_uar_alloc(to_mdev(ibdev), &context->uar);
304 	if (err)
305 		return err;
306 
307 	context->db_tab = mthca_init_user_db_tab(to_mdev(ibdev));
308 	if (IS_ERR(context->db_tab)) {
309 		err = PTR_ERR(context->db_tab);
310 		mthca_uar_free(to_mdev(ibdev), &context->uar);
311 		return err;
312 	}
313 
314 	if (ib_copy_to_udata(udata, &uresp, sizeof(uresp))) {
315 		mthca_cleanup_user_db_tab(to_mdev(ibdev), &context->uar, context->db_tab);
316 		mthca_uar_free(to_mdev(ibdev), &context->uar);
317 		return -EFAULT;
318 	}
319 
320 	context->reg_mr_warned = 0;
321 
322 	return 0;
323 }
324 
325 static void mthca_dealloc_ucontext(struct ib_ucontext *context)
326 {
327 	mthca_cleanup_user_db_tab(to_mdev(context->device), &to_mucontext(context)->uar,
328 				  to_mucontext(context)->db_tab);
329 	mthca_uar_free(to_mdev(context->device), &to_mucontext(context)->uar);
330 }
331 
332 static int mthca_mmap_uar(struct ib_ucontext *context,
333 			  struct vm_area_struct *vma)
334 {
335 	if (vma->vm_end - vma->vm_start != PAGE_SIZE)
336 		return -EINVAL;
337 
338 	vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
339 
340 	if (io_remap_pfn_range(vma, vma->vm_start,
341 			       to_mucontext(context)->uar.pfn,
342 			       PAGE_SIZE, vma->vm_page_prot))
343 		return -EAGAIN;
344 
345 	return 0;
346 }
347 
348 static int mthca_alloc_pd(struct ib_pd *ibpd, struct ib_udata *udata)
349 {
350 	struct ib_device *ibdev = ibpd->device;
351 	struct mthca_pd *pd = to_mpd(ibpd);
352 	int err;
353 
354 	err = mthca_pd_alloc(to_mdev(ibdev), !udata, pd);
355 	if (err)
356 		return err;
357 
358 	if (udata) {
359 		if (ib_copy_to_udata(udata, &pd->pd_num, sizeof (__u32))) {
360 			mthca_pd_free(to_mdev(ibdev), pd);
361 			return -EFAULT;
362 		}
363 	}
364 
365 	return 0;
366 }
367 
368 static int mthca_dealloc_pd(struct ib_pd *pd, struct ib_udata *udata)
369 {
370 	mthca_pd_free(to_mdev(pd->device), to_mpd(pd));
371 	return 0;
372 }
373 
374 static int mthca_ah_create(struct ib_ah *ibah,
375 			   struct rdma_ah_init_attr *init_attr,
376 			   struct ib_udata *udata)
377 
378 {
379 	struct mthca_ah *ah = to_mah(ibah);
380 
381 	return mthca_create_ah(to_mdev(ibah->device), to_mpd(ibah->pd),
382 			       init_attr->ah_attr, ah);
383 }
384 
385 static int mthca_ah_destroy(struct ib_ah *ah, u32 flags)
386 {
387 	mthca_destroy_ah(to_mdev(ah->device), to_mah(ah));
388 	return 0;
389 }
390 
391 static int mthca_create_srq(struct ib_srq *ibsrq,
392 			    struct ib_srq_init_attr *init_attr,
393 			    struct ib_udata *udata)
394 {
395 	struct mthca_create_srq ucmd;
396 	struct mthca_ucontext *context = rdma_udata_to_drv_context(
397 		udata, struct mthca_ucontext, ibucontext);
398 	struct mthca_srq *srq = to_msrq(ibsrq);
399 	int err;
400 
401 	if (init_attr->srq_type != IB_SRQT_BASIC)
402 		return -EOPNOTSUPP;
403 
404 	if (udata) {
405 		if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd)))
406 			return -EFAULT;
407 
408 		err = mthca_map_user_db(to_mdev(ibsrq->device), &context->uar,
409 					context->db_tab, ucmd.db_index,
410 					ucmd.db_page);
411 
412 		if (err)
413 			return err;
414 
415 		srq->mr.ibmr.lkey = ucmd.lkey;
416 		srq->db_index     = ucmd.db_index;
417 	}
418 
419 	err = mthca_alloc_srq(to_mdev(ibsrq->device), to_mpd(ibsrq->pd),
420 			      &init_attr->attr, srq, udata);
421 
422 	if (err && udata)
423 		mthca_unmap_user_db(to_mdev(ibsrq->device), &context->uar,
424 				    context->db_tab, ucmd.db_index);
425 
426 	if (err)
427 		return err;
428 
429 	if (context && ib_copy_to_udata(udata, &srq->srqn, sizeof(__u32))) {
430 		mthca_free_srq(to_mdev(ibsrq->device), srq);
431 		return -EFAULT;
432 	}
433 
434 	return 0;
435 }
436 
437 static int mthca_destroy_srq(struct ib_srq *srq, struct ib_udata *udata)
438 {
439 	if (udata) {
440 		struct mthca_ucontext *context =
441 			rdma_udata_to_drv_context(
442 				udata,
443 				struct mthca_ucontext,
444 				ibucontext);
445 
446 		mthca_unmap_user_db(to_mdev(srq->device), &context->uar,
447 				    context->db_tab, to_msrq(srq)->db_index);
448 	}
449 
450 	mthca_free_srq(to_mdev(srq->device), to_msrq(srq));
451 	return 0;
452 }
453 
454 static int mthca_create_qp(struct ib_qp *ibqp,
455 			   struct ib_qp_init_attr *init_attr,
456 			   struct ib_udata *udata)
457 {
458 	struct mthca_ucontext *context = rdma_udata_to_drv_context(
459 		udata, struct mthca_ucontext, ibucontext);
460 	struct mthca_create_qp ucmd;
461 	struct mthca_qp *qp = to_mqp(ibqp);
462 	struct mthca_dev *dev = to_mdev(ibqp->device);
463 	int err;
464 
465 	if (init_attr->create_flags)
466 		return -EOPNOTSUPP;
467 
468 	switch (init_attr->qp_type) {
469 	case IB_QPT_RC:
470 	case IB_QPT_UC:
471 	case IB_QPT_UD:
472 	{
473 		if (udata) {
474 			if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd)))
475 				return -EFAULT;
476 
477 			err = mthca_map_user_db(dev, &context->uar,
478 						context->db_tab,
479 						ucmd.sq_db_index,
480 						ucmd.sq_db_page);
481 			if (err)
482 				return err;
483 
484 			err = mthca_map_user_db(dev, &context->uar,
485 						context->db_tab,
486 						ucmd.rq_db_index,
487 						ucmd.rq_db_page);
488 			if (err) {
489 				mthca_unmap_user_db(dev, &context->uar,
490 						    context->db_tab,
491 						    ucmd.sq_db_index);
492 				return err;
493 			}
494 
495 			qp->mr.ibmr.lkey = ucmd.lkey;
496 			qp->sq.db_index  = ucmd.sq_db_index;
497 			qp->rq.db_index  = ucmd.rq_db_index;
498 		}
499 
500 		err = mthca_alloc_qp(dev, to_mpd(ibqp->pd),
501 				     to_mcq(init_attr->send_cq),
502 				     to_mcq(init_attr->recv_cq),
503 				     init_attr->qp_type, init_attr->sq_sig_type,
504 				     &init_attr->cap, qp, udata);
505 
506 		if (err && udata) {
507 			mthca_unmap_user_db(dev, &context->uar, context->db_tab,
508 					    ucmd.sq_db_index);
509 			mthca_unmap_user_db(dev, &context->uar, context->db_tab,
510 					    ucmd.rq_db_index);
511 		}
512 
513 		qp->ibqp.qp_num = qp->qpn;
514 		break;
515 	}
516 	case IB_QPT_SMI:
517 	case IB_QPT_GSI:
518 	{
519 		qp->sqp = kzalloc(sizeof(struct mthca_sqp), GFP_KERNEL);
520 		if (!qp->sqp)
521 			return -ENOMEM;
522 
523 		qp->ibqp.qp_num = init_attr->qp_type == IB_QPT_SMI ? 0 : 1;
524 
525 		err = mthca_alloc_sqp(dev, to_mpd(ibqp->pd),
526 				      to_mcq(init_attr->send_cq),
527 				      to_mcq(init_attr->recv_cq),
528 				      init_attr->sq_sig_type, &init_attr->cap,
529 				      qp->ibqp.qp_num, init_attr->port_num, qp,
530 				      udata);
531 		break;
532 	}
533 	default:
534 		/* Don't support raw QPs */
535 		return -EOPNOTSUPP;
536 	}
537 
538 	if (err) {
539 		kfree(qp->sqp);
540 		return err;
541 	}
542 
543 	init_attr->cap.max_send_wr     = qp->sq.max;
544 	init_attr->cap.max_recv_wr     = qp->rq.max;
545 	init_attr->cap.max_send_sge    = qp->sq.max_gs;
546 	init_attr->cap.max_recv_sge    = qp->rq.max_gs;
547 	init_attr->cap.max_inline_data = qp->max_inline_data;
548 
549 	return 0;
550 }
551 
552 static int mthca_destroy_qp(struct ib_qp *qp, struct ib_udata *udata)
553 {
554 	if (udata) {
555 		struct mthca_ucontext *context =
556 			rdma_udata_to_drv_context(
557 				udata,
558 				struct mthca_ucontext,
559 				ibucontext);
560 
561 		mthca_unmap_user_db(to_mdev(qp->device),
562 				    &context->uar,
563 				    context->db_tab,
564 				    to_mqp(qp)->sq.db_index);
565 		mthca_unmap_user_db(to_mdev(qp->device),
566 				    &context->uar,
567 				    context->db_tab,
568 				    to_mqp(qp)->rq.db_index);
569 	}
570 	mthca_free_qp(to_mdev(qp->device), to_mqp(qp));
571 	kfree(to_mqp(qp)->sqp);
572 	return 0;
573 }
574 
575 static int mthca_create_cq(struct ib_cq *ibcq,
576 			   const struct ib_cq_init_attr *attr,
577 			   struct ib_udata *udata)
578 {
579 	struct ib_device *ibdev = ibcq->device;
580 	int entries = attr->cqe;
581 	struct mthca_create_cq ucmd;
582 	struct mthca_cq *cq;
583 	int nent;
584 	int err;
585 	struct mthca_ucontext *context = rdma_udata_to_drv_context(
586 		udata, struct mthca_ucontext, ibucontext);
587 
588 	if (attr->flags)
589 		return -EOPNOTSUPP;
590 
591 	if (entries < 1 || entries > to_mdev(ibdev)->limits.max_cqes)
592 		return -EINVAL;
593 
594 	if (udata) {
595 		if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd)))
596 			return -EFAULT;
597 
598 		err = mthca_map_user_db(to_mdev(ibdev), &context->uar,
599 					context->db_tab, ucmd.set_db_index,
600 					ucmd.set_db_page);
601 		if (err)
602 			return err;
603 
604 		err = mthca_map_user_db(to_mdev(ibdev), &context->uar,
605 					context->db_tab, ucmd.arm_db_index,
606 					ucmd.arm_db_page);
607 		if (err)
608 			goto err_unmap_set;
609 	}
610 
611 	cq = to_mcq(ibcq);
612 
613 	if (udata) {
614 		cq->buf.mr.ibmr.lkey = ucmd.lkey;
615 		cq->set_ci_db_index  = ucmd.set_db_index;
616 		cq->arm_db_index     = ucmd.arm_db_index;
617 	}
618 
619 	for (nent = 1; nent <= entries; nent <<= 1)
620 		; /* nothing */
621 
622 	err = mthca_init_cq(to_mdev(ibdev), nent, context,
623 			    udata ? ucmd.pdn : to_mdev(ibdev)->driver_pd.pd_num,
624 			    cq);
625 	if (err)
626 		goto err_unmap_arm;
627 
628 	if (udata && ib_copy_to_udata(udata, &cq->cqn, sizeof(__u32))) {
629 		mthca_free_cq(to_mdev(ibdev), cq);
630 		err = -EFAULT;
631 		goto err_unmap_arm;
632 	}
633 
634 	cq->resize_buf = NULL;
635 
636 	return 0;
637 
638 err_unmap_arm:
639 	if (udata)
640 		mthca_unmap_user_db(to_mdev(ibdev), &context->uar,
641 				    context->db_tab, ucmd.arm_db_index);
642 
643 err_unmap_set:
644 	if (udata)
645 		mthca_unmap_user_db(to_mdev(ibdev), &context->uar,
646 				    context->db_tab, ucmd.set_db_index);
647 
648 	return err;
649 }
650 
651 static int mthca_alloc_resize_buf(struct mthca_dev *dev, struct mthca_cq *cq,
652 				  int entries)
653 {
654 	int ret;
655 
656 	spin_lock_irq(&cq->lock);
657 	if (cq->resize_buf) {
658 		ret = -EBUSY;
659 		goto unlock;
660 	}
661 
662 	cq->resize_buf = kmalloc(sizeof *cq->resize_buf, GFP_ATOMIC);
663 	if (!cq->resize_buf) {
664 		ret = -ENOMEM;
665 		goto unlock;
666 	}
667 
668 	cq->resize_buf->state = CQ_RESIZE_ALLOC;
669 
670 	ret = 0;
671 
672 unlock:
673 	spin_unlock_irq(&cq->lock);
674 
675 	if (ret)
676 		return ret;
677 
678 	ret = mthca_alloc_cq_buf(dev, &cq->resize_buf->buf, entries);
679 	if (ret) {
680 		spin_lock_irq(&cq->lock);
681 		kfree(cq->resize_buf);
682 		cq->resize_buf = NULL;
683 		spin_unlock_irq(&cq->lock);
684 		return ret;
685 	}
686 
687 	cq->resize_buf->cqe = entries - 1;
688 
689 	spin_lock_irq(&cq->lock);
690 	cq->resize_buf->state = CQ_RESIZE_READY;
691 	spin_unlock_irq(&cq->lock);
692 
693 	return 0;
694 }
695 
696 static int mthca_resize_cq(struct ib_cq *ibcq, int entries, struct ib_udata *udata)
697 {
698 	struct mthca_dev *dev = to_mdev(ibcq->device);
699 	struct mthca_cq *cq = to_mcq(ibcq);
700 	struct mthca_resize_cq ucmd;
701 	u32 lkey;
702 	int ret;
703 
704 	if (entries < 1 || entries > dev->limits.max_cqes)
705 		return -EINVAL;
706 
707 	mutex_lock(&cq->mutex);
708 
709 	entries = roundup_pow_of_two(entries + 1);
710 	if (entries == ibcq->cqe + 1) {
711 		ret = 0;
712 		goto out;
713 	}
714 
715 	if (cq->is_kernel) {
716 		ret = mthca_alloc_resize_buf(dev, cq, entries);
717 		if (ret)
718 			goto out;
719 		lkey = cq->resize_buf->buf.mr.ibmr.lkey;
720 	} else {
721 		if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd)) {
722 			ret = -EFAULT;
723 			goto out;
724 		}
725 		lkey = ucmd.lkey;
726 	}
727 
728 	ret = mthca_RESIZE_CQ(dev, cq->cqn, lkey, ilog2(entries));
729 
730 	if (ret) {
731 		if (cq->resize_buf) {
732 			mthca_free_cq_buf(dev, &cq->resize_buf->buf,
733 					  cq->resize_buf->cqe);
734 			kfree(cq->resize_buf);
735 			spin_lock_irq(&cq->lock);
736 			cq->resize_buf = NULL;
737 			spin_unlock_irq(&cq->lock);
738 		}
739 		goto out;
740 	}
741 
742 	if (cq->is_kernel) {
743 		struct mthca_cq_buf tbuf;
744 		int tcqe;
745 
746 		spin_lock_irq(&cq->lock);
747 		if (cq->resize_buf->state == CQ_RESIZE_READY) {
748 			mthca_cq_resize_copy_cqes(cq);
749 			tbuf         = cq->buf;
750 			tcqe         = cq->ibcq.cqe;
751 			cq->buf      = cq->resize_buf->buf;
752 			cq->ibcq.cqe = cq->resize_buf->cqe;
753 		} else {
754 			tbuf = cq->resize_buf->buf;
755 			tcqe = cq->resize_buf->cqe;
756 		}
757 
758 		kfree(cq->resize_buf);
759 		cq->resize_buf = NULL;
760 		spin_unlock_irq(&cq->lock);
761 
762 		mthca_free_cq_buf(dev, &tbuf, tcqe);
763 	} else
764 		ibcq->cqe = entries - 1;
765 
766 out:
767 	mutex_unlock(&cq->mutex);
768 
769 	return ret;
770 }
771 
772 static int mthca_destroy_cq(struct ib_cq *cq, struct ib_udata *udata)
773 {
774 	if (udata) {
775 		struct mthca_ucontext *context =
776 			rdma_udata_to_drv_context(
777 				udata,
778 				struct mthca_ucontext,
779 				ibucontext);
780 
781 		mthca_unmap_user_db(to_mdev(cq->device),
782 				    &context->uar,
783 				    context->db_tab,
784 				    to_mcq(cq)->arm_db_index);
785 		mthca_unmap_user_db(to_mdev(cq->device),
786 				    &context->uar,
787 				    context->db_tab,
788 				    to_mcq(cq)->set_ci_db_index);
789 	}
790 	mthca_free_cq(to_mdev(cq->device), to_mcq(cq));
791 	return 0;
792 }
793 
794 static inline u32 convert_access(int acc)
795 {
796 	return (acc & IB_ACCESS_REMOTE_ATOMIC ? MTHCA_MPT_FLAG_ATOMIC       : 0) |
797 	       (acc & IB_ACCESS_REMOTE_WRITE  ? MTHCA_MPT_FLAG_REMOTE_WRITE : 0) |
798 	       (acc & IB_ACCESS_REMOTE_READ   ? MTHCA_MPT_FLAG_REMOTE_READ  : 0) |
799 	       (acc & IB_ACCESS_LOCAL_WRITE   ? MTHCA_MPT_FLAG_LOCAL_WRITE  : 0) |
800 	       MTHCA_MPT_FLAG_LOCAL_READ;
801 }
802 
803 static struct ib_mr *mthca_get_dma_mr(struct ib_pd *pd, int acc)
804 {
805 	struct mthca_mr *mr;
806 	int err;
807 
808 	mr = kmalloc(sizeof *mr, GFP_KERNEL);
809 	if (!mr)
810 		return ERR_PTR(-ENOMEM);
811 
812 	err = mthca_mr_alloc_notrans(to_mdev(pd->device),
813 				     to_mpd(pd)->pd_num,
814 				     convert_access(acc), mr);
815 
816 	if (err) {
817 		kfree(mr);
818 		return ERR_PTR(err);
819 	}
820 
821 	mr->umem = NULL;
822 
823 	return &mr->ibmr;
824 }
825 
826 static struct ib_mr *mthca_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
827 				       u64 virt, int acc, struct ib_udata *udata)
828 {
829 	struct mthca_dev *dev = to_mdev(pd->device);
830 	struct ib_block_iter biter;
831 	struct mthca_ucontext *context = rdma_udata_to_drv_context(
832 		udata, struct mthca_ucontext, ibucontext);
833 	struct mthca_mr *mr;
834 	struct mthca_reg_mr ucmd;
835 	u64 *pages;
836 	int n, i;
837 	int err = 0;
838 	int write_mtt_size;
839 
840 	if (udata->inlen < sizeof ucmd) {
841 		if (!context->reg_mr_warned) {
842 			mthca_warn(dev, "Process '%s' did not pass in MR attrs.\n",
843 				   current->comm);
844 			mthca_warn(dev, "  Update libmthca to fix this.\n");
845 		}
846 		++context->reg_mr_warned;
847 		ucmd.mr_attrs = 0;
848 	} else if (ib_copy_from_udata(&ucmd, udata, sizeof ucmd))
849 		return ERR_PTR(-EFAULT);
850 
851 	mr = kmalloc(sizeof *mr, GFP_KERNEL);
852 	if (!mr)
853 		return ERR_PTR(-ENOMEM);
854 
855 	mr->umem = ib_umem_get(pd->device, start, length, acc);
856 	if (IS_ERR(mr->umem)) {
857 		err = PTR_ERR(mr->umem);
858 		goto err;
859 	}
860 
861 	n = ib_umem_num_dma_blocks(mr->umem, PAGE_SIZE);
862 
863 	mr->mtt = mthca_alloc_mtt(dev, n);
864 	if (IS_ERR(mr->mtt)) {
865 		err = PTR_ERR(mr->mtt);
866 		goto err_umem;
867 	}
868 
869 	pages = (u64 *) __get_free_page(GFP_KERNEL);
870 	if (!pages) {
871 		err = -ENOMEM;
872 		goto err_mtt;
873 	}
874 
875 	i = n = 0;
876 
877 	write_mtt_size = min(mthca_write_mtt_size(dev), (int) (PAGE_SIZE / sizeof *pages));
878 
879 	rdma_umem_for_each_dma_block(mr->umem, &biter, PAGE_SIZE) {
880 		pages[i++] = rdma_block_iter_dma_address(&biter);
881 
882 		/*
883 		 * Be friendly to write_mtt and pass it chunks
884 		 * of appropriate size.
885 		 */
886 		if (i == write_mtt_size) {
887 			err = mthca_write_mtt(dev, mr->mtt, n, pages, i);
888 			if (err)
889 				goto mtt_done;
890 			n += i;
891 			i = 0;
892 		}
893 	}
894 
895 	if (i)
896 		err = mthca_write_mtt(dev, mr->mtt, n, pages, i);
897 mtt_done:
898 	free_page((unsigned long) pages);
899 	if (err)
900 		goto err_mtt;
901 
902 	err = mthca_mr_alloc(dev, to_mpd(pd)->pd_num, PAGE_SHIFT, virt, length,
903 			     convert_access(acc), mr);
904 
905 	if (err)
906 		goto err_mtt;
907 
908 	return &mr->ibmr;
909 
910 err_mtt:
911 	mthca_free_mtt(dev, mr->mtt);
912 
913 err_umem:
914 	ib_umem_release(mr->umem);
915 
916 err:
917 	kfree(mr);
918 	return ERR_PTR(err);
919 }
920 
921 static int mthca_dereg_mr(struct ib_mr *mr, struct ib_udata *udata)
922 {
923 	struct mthca_mr *mmr = to_mmr(mr);
924 
925 	mthca_free_mr(to_mdev(mr->device), mmr);
926 	ib_umem_release(mmr->umem);
927 	kfree(mmr);
928 
929 	return 0;
930 }
931 
932 static ssize_t hw_rev_show(struct device *device,
933 			   struct device_attribute *attr, char *buf)
934 {
935 	struct mthca_dev *dev =
936 		rdma_device_to_drv_device(device, struct mthca_dev, ib_dev);
937 
938 	return sysfs_emit(buf, "%x\n", dev->rev_id);
939 }
940 static DEVICE_ATTR_RO(hw_rev);
941 
942 static const char *hca_type_string(int hca_type)
943 {
944 	switch (hca_type) {
945 	case PCI_DEVICE_ID_MELLANOX_TAVOR:
946 		return "MT23108";
947 	case PCI_DEVICE_ID_MELLANOX_ARBEL_COMPAT:
948 		return "MT25208 (MT23108 compat mode)";
949 	case PCI_DEVICE_ID_MELLANOX_ARBEL:
950 		return "MT25208";
951 	case PCI_DEVICE_ID_MELLANOX_SINAI:
952 	case PCI_DEVICE_ID_MELLANOX_SINAI_OLD:
953 		return "MT25204";
954 	}
955 
956 	return "unknown";
957 }
958 
959 static ssize_t hca_type_show(struct device *device,
960 			     struct device_attribute *attr, char *buf)
961 {
962 	struct mthca_dev *dev =
963 		rdma_device_to_drv_device(device, struct mthca_dev, ib_dev);
964 
965 	return sysfs_emit(buf, "%s\n", hca_type_string(dev->pdev->device));
966 }
967 static DEVICE_ATTR_RO(hca_type);
968 
969 static ssize_t board_id_show(struct device *device,
970 			     struct device_attribute *attr, char *buf)
971 {
972 	struct mthca_dev *dev =
973 		rdma_device_to_drv_device(device, struct mthca_dev, ib_dev);
974 
975 	return sysfs_emit(buf, "%.*s\n", MTHCA_BOARD_ID_LEN, dev->board_id);
976 }
977 static DEVICE_ATTR_RO(board_id);
978 
979 static struct attribute *mthca_dev_attributes[] = {
980 	&dev_attr_hw_rev.attr,
981 	&dev_attr_hca_type.attr,
982 	&dev_attr_board_id.attr,
983 	NULL
984 };
985 
986 static const struct attribute_group mthca_attr_group = {
987 	.attrs = mthca_dev_attributes,
988 };
989 
990 static int mthca_init_node_data(struct mthca_dev *dev)
991 {
992 	struct ib_smp *in_mad;
993 	struct ib_smp *out_mad;
994 	int err = -ENOMEM;
995 
996 	in_mad  = kzalloc(sizeof *in_mad, GFP_KERNEL);
997 	out_mad = kmalloc(sizeof *out_mad, GFP_KERNEL);
998 	if (!in_mad || !out_mad)
999 		goto out;
1000 
1001 	ib_init_query_mad(in_mad);
1002 	in_mad->attr_id = IB_SMP_ATTR_NODE_DESC;
1003 
1004 	err = mthca_MAD_IFC(dev, 1, 1,
1005 			    1, NULL, NULL, in_mad, out_mad);
1006 	if (err)
1007 		goto out;
1008 
1009 	memcpy(dev->ib_dev.node_desc, out_mad->data, IB_DEVICE_NODE_DESC_MAX);
1010 
1011 	in_mad->attr_id = IB_SMP_ATTR_NODE_INFO;
1012 
1013 	err = mthca_MAD_IFC(dev, 1, 1,
1014 			    1, NULL, NULL, in_mad, out_mad);
1015 	if (err)
1016 		goto out;
1017 
1018 	if (mthca_is_memfree(dev))
1019 		dev->rev_id = be32_to_cpup((__be32 *) (out_mad->data + 32));
1020 	memcpy(&dev->ib_dev.node_guid, out_mad->data + 12, 8);
1021 
1022 out:
1023 	kfree(in_mad);
1024 	kfree(out_mad);
1025 	return err;
1026 }
1027 
1028 static int mthca_port_immutable(struct ib_device *ibdev, u32 port_num,
1029 			        struct ib_port_immutable *immutable)
1030 {
1031 	struct ib_port_attr attr;
1032 	int err;
1033 
1034 	immutable->core_cap_flags = RDMA_CORE_PORT_IBA_IB;
1035 
1036 	err = ib_query_port(ibdev, port_num, &attr);
1037 	if (err)
1038 		return err;
1039 
1040 	immutable->pkey_tbl_len = attr.pkey_tbl_len;
1041 	immutable->gid_tbl_len = attr.gid_tbl_len;
1042 	immutable->max_mad_size = IB_MGMT_MAD_SIZE;
1043 
1044 	return 0;
1045 }
1046 
1047 static void get_dev_fw_str(struct ib_device *device, char *str)
1048 {
1049 	struct mthca_dev *dev =
1050 		container_of(device, struct mthca_dev, ib_dev);
1051 	snprintf(str, IB_FW_VERSION_NAME_MAX, "%d.%d.%d",
1052 		 (int) (dev->fw_ver >> 32),
1053 		 (int) (dev->fw_ver >> 16) & 0xffff,
1054 		 (int) dev->fw_ver & 0xffff);
1055 }
1056 
1057 static const struct ib_device_ops mthca_dev_ops = {
1058 	.owner = THIS_MODULE,
1059 	.driver_id = RDMA_DRIVER_MTHCA,
1060 	.uverbs_abi_ver = MTHCA_UVERBS_ABI_VERSION,
1061 	.uverbs_no_driver_id_binding = 1,
1062 
1063 	.alloc_pd = mthca_alloc_pd,
1064 	.alloc_ucontext = mthca_alloc_ucontext,
1065 	.attach_mcast = mthca_multicast_attach,
1066 	.create_ah = mthca_ah_create,
1067 	.create_cq = mthca_create_cq,
1068 	.create_qp = mthca_create_qp,
1069 	.dealloc_pd = mthca_dealloc_pd,
1070 	.dealloc_ucontext = mthca_dealloc_ucontext,
1071 	.dereg_mr = mthca_dereg_mr,
1072 	.destroy_ah = mthca_ah_destroy,
1073 	.destroy_cq = mthca_destroy_cq,
1074 	.destroy_qp = mthca_destroy_qp,
1075 	.detach_mcast = mthca_multicast_detach,
1076 	.device_group = &mthca_attr_group,
1077 	.get_dev_fw_str = get_dev_fw_str,
1078 	.get_dma_mr = mthca_get_dma_mr,
1079 	.get_port_immutable = mthca_port_immutable,
1080 	.mmap = mthca_mmap_uar,
1081 	.modify_device = mthca_modify_device,
1082 	.modify_port = mthca_modify_port,
1083 	.modify_qp = mthca_modify_qp,
1084 	.poll_cq = mthca_poll_cq,
1085 	.process_mad = mthca_process_mad,
1086 	.query_ah = mthca_ah_query,
1087 	.query_device = mthca_query_device,
1088 	.query_gid = mthca_query_gid,
1089 	.query_pkey = mthca_query_pkey,
1090 	.query_port = mthca_query_port,
1091 	.query_qp = mthca_query_qp,
1092 	.reg_user_mr = mthca_reg_user_mr,
1093 	.resize_cq = mthca_resize_cq,
1094 
1095 	INIT_RDMA_OBJ_SIZE(ib_ah, mthca_ah, ibah),
1096 	INIT_RDMA_OBJ_SIZE(ib_cq, mthca_cq, ibcq),
1097 	INIT_RDMA_OBJ_SIZE(ib_pd, mthca_pd, ibpd),
1098 	INIT_RDMA_OBJ_SIZE(ib_qp, mthca_qp, ibqp),
1099 	INIT_RDMA_OBJ_SIZE(ib_ucontext, mthca_ucontext, ibucontext),
1100 };
1101 
1102 static const struct ib_device_ops mthca_dev_arbel_srq_ops = {
1103 	.create_srq = mthca_create_srq,
1104 	.destroy_srq = mthca_destroy_srq,
1105 	.modify_srq = mthca_modify_srq,
1106 	.post_srq_recv = mthca_arbel_post_srq_recv,
1107 	.query_srq = mthca_query_srq,
1108 
1109 	INIT_RDMA_OBJ_SIZE(ib_srq, mthca_srq, ibsrq),
1110 };
1111 
1112 static const struct ib_device_ops mthca_dev_tavor_srq_ops = {
1113 	.create_srq = mthca_create_srq,
1114 	.destroy_srq = mthca_destroy_srq,
1115 	.modify_srq = mthca_modify_srq,
1116 	.post_srq_recv = mthca_tavor_post_srq_recv,
1117 	.query_srq = mthca_query_srq,
1118 
1119 	INIT_RDMA_OBJ_SIZE(ib_srq, mthca_srq, ibsrq),
1120 };
1121 
1122 static const struct ib_device_ops mthca_dev_arbel_ops = {
1123 	.post_recv = mthca_arbel_post_receive,
1124 	.post_send = mthca_arbel_post_send,
1125 	.req_notify_cq = mthca_arbel_arm_cq,
1126 };
1127 
1128 static const struct ib_device_ops mthca_dev_tavor_ops = {
1129 	.post_recv = mthca_tavor_post_receive,
1130 	.post_send = mthca_tavor_post_send,
1131 	.req_notify_cq = mthca_tavor_arm_cq,
1132 };
1133 
1134 int mthca_register_device(struct mthca_dev *dev)
1135 {
1136 	int ret;
1137 
1138 	ret = mthca_init_node_data(dev);
1139 	if (ret)
1140 		return ret;
1141 
1142 	dev->ib_dev.node_type            = RDMA_NODE_IB_CA;
1143 	dev->ib_dev.phys_port_cnt        = dev->limits.num_ports;
1144 	dev->ib_dev.num_comp_vectors     = 1;
1145 	dev->ib_dev.dev.parent           = &dev->pdev->dev;
1146 
1147 	if (dev->mthca_flags & MTHCA_FLAG_SRQ) {
1148 		if (mthca_is_memfree(dev))
1149 			ib_set_device_ops(&dev->ib_dev,
1150 					  &mthca_dev_arbel_srq_ops);
1151 		else
1152 			ib_set_device_ops(&dev->ib_dev,
1153 					  &mthca_dev_tavor_srq_ops);
1154 	}
1155 
1156 	ib_set_device_ops(&dev->ib_dev, &mthca_dev_ops);
1157 
1158 	if (mthca_is_memfree(dev))
1159 		ib_set_device_ops(&dev->ib_dev, &mthca_dev_arbel_ops);
1160 	else
1161 		ib_set_device_ops(&dev->ib_dev, &mthca_dev_tavor_ops);
1162 
1163 	mutex_init(&dev->cap_mask_mutex);
1164 
1165 	ret = ib_register_device(&dev->ib_dev, "mthca%d", &dev->pdev->dev);
1166 	if (ret)
1167 		return ret;
1168 
1169 	mthca_start_catas_poll(dev);
1170 
1171 	return 0;
1172 }
1173 
1174 void mthca_unregister_device(struct mthca_dev *dev)
1175 {
1176 	mthca_stop_catas_poll(dev);
1177 	ib_unregister_device(&dev->ib_dev);
1178 }
1179