xref: /openbmc/linux/drivers/infiniband/core/verbs.c (revision 828ff2ad)
1 /*
2  * Copyright (c) 2004 Mellanox Technologies Ltd.  All rights reserved.
3  * Copyright (c) 2004 Infinicon Corporation.  All rights reserved.
4  * Copyright (c) 2004 Intel Corporation.  All rights reserved.
5  * Copyright (c) 2004 Topspin Corporation.  All rights reserved.
6  * Copyright (c) 2004 Voltaire Corporation.  All rights reserved.
7  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
8  * Copyright (c) 2005, 2006 Cisco Systems.  All rights reserved.
9  *
10  * This software is available to you under a choice of one of two
11  * licenses.  You may choose to be licensed under the terms of the GNU
12  * General Public License (GPL) Version 2, available from the file
13  * COPYING in the main directory of this source tree, or the
14  * OpenIB.org BSD license below:
15  *
16  *     Redistribution and use in source and binary forms, with or
17  *     without modification, are permitted provided that the following
18  *     conditions are met:
19  *
20  *      - Redistributions of source code must retain the above
21  *        copyright notice, this list of conditions and the following
22  *        disclaimer.
23  *
24  *      - Redistributions in binary form must reproduce the above
25  *        copyright notice, this list of conditions and the following
26  *        disclaimer in the documentation and/or other materials
27  *        provided with the distribution.
28  *
29  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
30  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
31  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
32  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
33  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
34  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
35  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
36  * SOFTWARE.
37  */
38 
39 #include <linux/errno.h>
40 #include <linux/err.h>
41 #include <linux/export.h>
42 #include <linux/string.h>
43 #include <linux/slab.h>
44 #include <linux/in.h>
45 #include <linux/in6.h>
46 #include <net/addrconf.h>
47 #include <linux/security.h>
48 
49 #include <rdma/ib_verbs.h>
50 #include <rdma/ib_cache.h>
51 #include <rdma/ib_addr.h>
52 #include <rdma/rw.h>
53 
54 #include "core_priv.h"
55 
56 static int ib_resolve_eth_dmac(struct ib_device *device,
57 			       struct rdma_ah_attr *ah_attr);
58 
59 static const char * const ib_events[] = {
60 	[IB_EVENT_CQ_ERR]		= "CQ error",
61 	[IB_EVENT_QP_FATAL]		= "QP fatal error",
62 	[IB_EVENT_QP_REQ_ERR]		= "QP request error",
63 	[IB_EVENT_QP_ACCESS_ERR]	= "QP access error",
64 	[IB_EVENT_COMM_EST]		= "communication established",
65 	[IB_EVENT_SQ_DRAINED]		= "send queue drained",
66 	[IB_EVENT_PATH_MIG]		= "path migration successful",
67 	[IB_EVENT_PATH_MIG_ERR]		= "path migration error",
68 	[IB_EVENT_DEVICE_FATAL]		= "device fatal error",
69 	[IB_EVENT_PORT_ACTIVE]		= "port active",
70 	[IB_EVENT_PORT_ERR]		= "port error",
71 	[IB_EVENT_LID_CHANGE]		= "LID change",
72 	[IB_EVENT_PKEY_CHANGE]		= "P_key change",
73 	[IB_EVENT_SM_CHANGE]		= "SM change",
74 	[IB_EVENT_SRQ_ERR]		= "SRQ error",
75 	[IB_EVENT_SRQ_LIMIT_REACHED]	= "SRQ limit reached",
76 	[IB_EVENT_QP_LAST_WQE_REACHED]	= "last WQE reached",
77 	[IB_EVENT_CLIENT_REREGISTER]	= "client reregister",
78 	[IB_EVENT_GID_CHANGE]		= "GID changed",
79 };
80 
81 const char *__attribute_const__ ib_event_msg(enum ib_event_type event)
82 {
83 	size_t index = event;
84 
85 	return (index < ARRAY_SIZE(ib_events) && ib_events[index]) ?
86 			ib_events[index] : "unrecognized event";
87 }
88 EXPORT_SYMBOL(ib_event_msg);
89 
90 static const char * const wc_statuses[] = {
91 	[IB_WC_SUCCESS]			= "success",
92 	[IB_WC_LOC_LEN_ERR]		= "local length error",
93 	[IB_WC_LOC_QP_OP_ERR]		= "local QP operation error",
94 	[IB_WC_LOC_EEC_OP_ERR]		= "local EE context operation error",
95 	[IB_WC_LOC_PROT_ERR]		= "local protection error",
96 	[IB_WC_WR_FLUSH_ERR]		= "WR flushed",
97 	[IB_WC_MW_BIND_ERR]		= "memory management operation error",
98 	[IB_WC_BAD_RESP_ERR]		= "bad response error",
99 	[IB_WC_LOC_ACCESS_ERR]		= "local access error",
100 	[IB_WC_REM_INV_REQ_ERR]		= "invalid request error",
101 	[IB_WC_REM_ACCESS_ERR]		= "remote access error",
102 	[IB_WC_REM_OP_ERR]		= "remote operation error",
103 	[IB_WC_RETRY_EXC_ERR]		= "transport retry counter exceeded",
104 	[IB_WC_RNR_RETRY_EXC_ERR]	= "RNR retry counter exceeded",
105 	[IB_WC_LOC_RDD_VIOL_ERR]	= "local RDD violation error",
106 	[IB_WC_REM_INV_RD_REQ_ERR]	= "remote invalid RD request",
107 	[IB_WC_REM_ABORT_ERR]		= "operation aborted",
108 	[IB_WC_INV_EECN_ERR]		= "invalid EE context number",
109 	[IB_WC_INV_EEC_STATE_ERR]	= "invalid EE context state",
110 	[IB_WC_FATAL_ERR]		= "fatal error",
111 	[IB_WC_RESP_TIMEOUT_ERR]	= "response timeout error",
112 	[IB_WC_GENERAL_ERR]		= "general error",
113 };
114 
115 const char *__attribute_const__ ib_wc_status_msg(enum ib_wc_status status)
116 {
117 	size_t index = status;
118 
119 	return (index < ARRAY_SIZE(wc_statuses) && wc_statuses[index]) ?
120 			wc_statuses[index] : "unrecognized status";
121 }
122 EXPORT_SYMBOL(ib_wc_status_msg);
123 
124 __attribute_const__ int ib_rate_to_mult(enum ib_rate rate)
125 {
126 	switch (rate) {
127 	case IB_RATE_2_5_GBPS: return   1;
128 	case IB_RATE_5_GBPS:   return   2;
129 	case IB_RATE_10_GBPS:  return   4;
130 	case IB_RATE_20_GBPS:  return   8;
131 	case IB_RATE_30_GBPS:  return  12;
132 	case IB_RATE_40_GBPS:  return  16;
133 	case IB_RATE_60_GBPS:  return  24;
134 	case IB_RATE_80_GBPS:  return  32;
135 	case IB_RATE_120_GBPS: return  48;
136 	case IB_RATE_14_GBPS:  return   6;
137 	case IB_RATE_56_GBPS:  return  22;
138 	case IB_RATE_112_GBPS: return  45;
139 	case IB_RATE_168_GBPS: return  67;
140 	case IB_RATE_25_GBPS:  return  10;
141 	case IB_RATE_100_GBPS: return  40;
142 	case IB_RATE_200_GBPS: return  80;
143 	case IB_RATE_300_GBPS: return 120;
144 	default:	       return  -1;
145 	}
146 }
147 EXPORT_SYMBOL(ib_rate_to_mult);
148 
149 __attribute_const__ enum ib_rate mult_to_ib_rate(int mult)
150 {
151 	switch (mult) {
152 	case 1:   return IB_RATE_2_5_GBPS;
153 	case 2:   return IB_RATE_5_GBPS;
154 	case 4:   return IB_RATE_10_GBPS;
155 	case 8:   return IB_RATE_20_GBPS;
156 	case 12:  return IB_RATE_30_GBPS;
157 	case 16:  return IB_RATE_40_GBPS;
158 	case 24:  return IB_RATE_60_GBPS;
159 	case 32:  return IB_RATE_80_GBPS;
160 	case 48:  return IB_RATE_120_GBPS;
161 	case 6:   return IB_RATE_14_GBPS;
162 	case 22:  return IB_RATE_56_GBPS;
163 	case 45:  return IB_RATE_112_GBPS;
164 	case 67:  return IB_RATE_168_GBPS;
165 	case 10:  return IB_RATE_25_GBPS;
166 	case 40:  return IB_RATE_100_GBPS;
167 	case 80:  return IB_RATE_200_GBPS;
168 	case 120: return IB_RATE_300_GBPS;
169 	default:  return IB_RATE_PORT_CURRENT;
170 	}
171 }
172 EXPORT_SYMBOL(mult_to_ib_rate);
173 
174 __attribute_const__ int ib_rate_to_mbps(enum ib_rate rate)
175 {
176 	switch (rate) {
177 	case IB_RATE_2_5_GBPS: return 2500;
178 	case IB_RATE_5_GBPS:   return 5000;
179 	case IB_RATE_10_GBPS:  return 10000;
180 	case IB_RATE_20_GBPS:  return 20000;
181 	case IB_RATE_30_GBPS:  return 30000;
182 	case IB_RATE_40_GBPS:  return 40000;
183 	case IB_RATE_60_GBPS:  return 60000;
184 	case IB_RATE_80_GBPS:  return 80000;
185 	case IB_RATE_120_GBPS: return 120000;
186 	case IB_RATE_14_GBPS:  return 14062;
187 	case IB_RATE_56_GBPS:  return 56250;
188 	case IB_RATE_112_GBPS: return 112500;
189 	case IB_RATE_168_GBPS: return 168750;
190 	case IB_RATE_25_GBPS:  return 25781;
191 	case IB_RATE_100_GBPS: return 103125;
192 	case IB_RATE_200_GBPS: return 206250;
193 	case IB_RATE_300_GBPS: return 309375;
194 	default:	       return -1;
195 	}
196 }
197 EXPORT_SYMBOL(ib_rate_to_mbps);
198 
199 __attribute_const__ enum rdma_transport_type
200 rdma_node_get_transport(enum rdma_node_type node_type)
201 {
202 
203 	if (node_type == RDMA_NODE_USNIC)
204 		return RDMA_TRANSPORT_USNIC;
205 	if (node_type == RDMA_NODE_USNIC_UDP)
206 		return RDMA_TRANSPORT_USNIC_UDP;
207 	if (node_type == RDMA_NODE_RNIC)
208 		return RDMA_TRANSPORT_IWARP;
209 
210 	return RDMA_TRANSPORT_IB;
211 }
212 EXPORT_SYMBOL(rdma_node_get_transport);
213 
214 enum rdma_link_layer rdma_port_get_link_layer(struct ib_device *device, u8 port_num)
215 {
216 	enum rdma_transport_type lt;
217 	if (device->get_link_layer)
218 		return device->get_link_layer(device, port_num);
219 
220 	lt = rdma_node_get_transport(device->node_type);
221 	if (lt == RDMA_TRANSPORT_IB)
222 		return IB_LINK_LAYER_INFINIBAND;
223 
224 	return IB_LINK_LAYER_ETHERNET;
225 }
226 EXPORT_SYMBOL(rdma_port_get_link_layer);
227 
228 /* Protection domains */
229 
230 /**
231  * ib_alloc_pd - Allocates an unused protection domain.
232  * @device: The device on which to allocate the protection domain.
233  *
234  * A protection domain object provides an association between QPs, shared
235  * receive queues, address handles, memory regions, and memory windows.
236  *
237  * Every PD has a local_dma_lkey which can be used as the lkey value for local
238  * memory operations.
239  */
240 struct ib_pd *__ib_alloc_pd(struct ib_device *device, unsigned int flags,
241 		const char *caller)
242 {
243 	struct ib_pd *pd;
244 	int mr_access_flags = 0;
245 
246 	pd = device->alloc_pd(device, NULL, NULL);
247 	if (IS_ERR(pd))
248 		return pd;
249 
250 	pd->device = device;
251 	pd->uobject = NULL;
252 	pd->__internal_mr = NULL;
253 	atomic_set(&pd->usecnt, 0);
254 	pd->flags = flags;
255 
256 	if (device->attrs.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY)
257 		pd->local_dma_lkey = device->local_dma_lkey;
258 	else
259 		mr_access_flags |= IB_ACCESS_LOCAL_WRITE;
260 
261 	if (flags & IB_PD_UNSAFE_GLOBAL_RKEY) {
262 		pr_warn("%s: enabling unsafe global rkey\n", caller);
263 		mr_access_flags |= IB_ACCESS_REMOTE_READ | IB_ACCESS_REMOTE_WRITE;
264 	}
265 
266 	pd->res.type = RDMA_RESTRACK_PD;
267 	rdma_restrack_set_task(&pd->res, caller);
268 	rdma_restrack_add(&pd->res);
269 
270 	if (mr_access_flags) {
271 		struct ib_mr *mr;
272 
273 		mr = pd->device->get_dma_mr(pd, mr_access_flags);
274 		if (IS_ERR(mr)) {
275 			ib_dealloc_pd(pd);
276 			return ERR_CAST(mr);
277 		}
278 
279 		mr->device	= pd->device;
280 		mr->pd		= pd;
281 		mr->uobject	= NULL;
282 		mr->need_inval	= false;
283 
284 		pd->__internal_mr = mr;
285 
286 		if (!(device->attrs.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY))
287 			pd->local_dma_lkey = pd->__internal_mr->lkey;
288 
289 		if (flags & IB_PD_UNSAFE_GLOBAL_RKEY)
290 			pd->unsafe_global_rkey = pd->__internal_mr->rkey;
291 	}
292 
293 	return pd;
294 }
295 EXPORT_SYMBOL(__ib_alloc_pd);
296 
297 /**
298  * ib_dealloc_pd - Deallocates a protection domain.
299  * @pd: The protection domain to deallocate.
300  *
301  * It is an error to call this function while any resources in the pd still
302  * exist.  The caller is responsible to synchronously destroy them and
303  * guarantee no new allocations will happen.
304  */
305 void ib_dealloc_pd(struct ib_pd *pd)
306 {
307 	int ret;
308 
309 	if (pd->__internal_mr) {
310 		ret = pd->device->dereg_mr(pd->__internal_mr);
311 		WARN_ON(ret);
312 		pd->__internal_mr = NULL;
313 	}
314 
315 	/* uverbs manipulates usecnt with proper locking, while the kabi
316 	   requires the caller to guarantee we can't race here. */
317 	WARN_ON(atomic_read(&pd->usecnt));
318 
319 	rdma_restrack_del(&pd->res);
320 	/* Making delalloc_pd a void return is a WIP, no driver should return
321 	   an error here. */
322 	ret = pd->device->dealloc_pd(pd);
323 	WARN_ONCE(ret, "Infiniband HW driver failed dealloc_pd");
324 }
325 EXPORT_SYMBOL(ib_dealloc_pd);
326 
327 /* Address handles */
328 
329 /**
330  * rdma_copy_ah_attr - Copy rdma ah attribute from source to destination.
331  * @dest:       Pointer to destination ah_attr. Contents of the destination
332  *              pointer is assumed to be invalid and attribute are overwritten.
333  * @src:        Pointer to source ah_attr.
334  */
335 void rdma_copy_ah_attr(struct rdma_ah_attr *dest,
336 		       const struct rdma_ah_attr *src)
337 {
338 	*dest = *src;
339 	if (dest->grh.sgid_attr)
340 		rdma_hold_gid_attr(dest->grh.sgid_attr);
341 }
342 EXPORT_SYMBOL(rdma_copy_ah_attr);
343 
344 /**
345  * rdma_replace_ah_attr - Replace valid ah_attr with new new one.
346  * @old:        Pointer to existing ah_attr which needs to be replaced.
347  *              old is assumed to be valid or zero'd
348  * @new:        Pointer to the new ah_attr.
349  *
350  * rdma_replace_ah_attr() first releases any reference in the old ah_attr if
351  * old the ah_attr is valid; after that it copies the new attribute and holds
352  * the reference to the replaced ah_attr.
353  */
354 void rdma_replace_ah_attr(struct rdma_ah_attr *old,
355 			  const struct rdma_ah_attr *new)
356 {
357 	rdma_destroy_ah_attr(old);
358 	*old = *new;
359 	if (old->grh.sgid_attr)
360 		rdma_hold_gid_attr(old->grh.sgid_attr);
361 }
362 EXPORT_SYMBOL(rdma_replace_ah_attr);
363 
364 /**
365  * rdma_move_ah_attr - Move ah_attr pointed by source to destination.
366  * @dest:       Pointer to destination ah_attr to copy to.
367  *              dest is assumed to be valid or zero'd
368  * @src:        Pointer to the new ah_attr.
369  *
370  * rdma_move_ah_attr() first releases any reference in the destination ah_attr
371  * if it is valid. This also transfers ownership of internal references from
372  * src to dest, making src invalid in the process. No new reference of the src
373  * ah_attr is taken.
374  */
375 void rdma_move_ah_attr(struct rdma_ah_attr *dest, struct rdma_ah_attr *src)
376 {
377 	rdma_destroy_ah_attr(dest);
378 	*dest = *src;
379 	src->grh.sgid_attr = NULL;
380 }
381 EXPORT_SYMBOL(rdma_move_ah_attr);
382 
383 /*
384  * Validate that the rdma_ah_attr is valid for the device before passing it
385  * off to the driver.
386  */
387 static int rdma_check_ah_attr(struct ib_device *device,
388 			      struct rdma_ah_attr *ah_attr)
389 {
390 	if (!rdma_is_port_valid(device, ah_attr->port_num))
391 		return -EINVAL;
392 
393 	if ((rdma_is_grh_required(device, ah_attr->port_num) ||
394 	     ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE) &&
395 	    !(ah_attr->ah_flags & IB_AH_GRH))
396 		return -EINVAL;
397 
398 	if (ah_attr->grh.sgid_attr) {
399 		/*
400 		 * Make sure the passed sgid_attr is consistent with the
401 		 * parameters
402 		 */
403 		if (ah_attr->grh.sgid_attr->index != ah_attr->grh.sgid_index ||
404 		    ah_attr->grh.sgid_attr->port_num != ah_attr->port_num)
405 			return -EINVAL;
406 	}
407 	return 0;
408 }
409 
410 /*
411  * If the ah requires a GRH then ensure that sgid_attr pointer is filled in.
412  * On success the caller is responsible to call rdma_unfill_sgid_attr().
413  */
414 static int rdma_fill_sgid_attr(struct ib_device *device,
415 			       struct rdma_ah_attr *ah_attr,
416 			       const struct ib_gid_attr **old_sgid_attr)
417 {
418 	const struct ib_gid_attr *sgid_attr;
419 	struct ib_global_route *grh;
420 	int ret;
421 
422 	*old_sgid_attr = ah_attr->grh.sgid_attr;
423 
424 	ret = rdma_check_ah_attr(device, ah_attr);
425 	if (ret)
426 		return ret;
427 
428 	if (!(ah_attr->ah_flags & IB_AH_GRH))
429 		return 0;
430 
431 	grh = rdma_ah_retrieve_grh(ah_attr);
432 	if (grh->sgid_attr)
433 		return 0;
434 
435 	sgid_attr =
436 		rdma_get_gid_attr(device, ah_attr->port_num, grh->sgid_index);
437 	if (IS_ERR(sgid_attr))
438 		return PTR_ERR(sgid_attr);
439 
440 	/* Move ownerhip of the kref into the ah_attr */
441 	grh->sgid_attr = sgid_attr;
442 	return 0;
443 }
444 
445 static void rdma_unfill_sgid_attr(struct rdma_ah_attr *ah_attr,
446 				  const struct ib_gid_attr *old_sgid_attr)
447 {
448 	/*
449 	 * Fill didn't change anything, the caller retains ownership of
450 	 * whatever it passed
451 	 */
452 	if (ah_attr->grh.sgid_attr == old_sgid_attr)
453 		return;
454 
455 	/*
456 	 * Otherwise, we need to undo what rdma_fill_sgid_attr so the caller
457 	 * doesn't see any change in the rdma_ah_attr. If we get here
458 	 * old_sgid_attr is NULL.
459 	 */
460 	rdma_destroy_ah_attr(ah_attr);
461 }
462 
463 static const struct ib_gid_attr *
464 rdma_update_sgid_attr(struct rdma_ah_attr *ah_attr,
465 		      const struct ib_gid_attr *old_attr)
466 {
467 	if (old_attr)
468 		rdma_put_gid_attr(old_attr);
469 	if (ah_attr->ah_flags & IB_AH_GRH) {
470 		rdma_hold_gid_attr(ah_attr->grh.sgid_attr);
471 		return ah_attr->grh.sgid_attr;
472 	}
473 	return NULL;
474 }
475 
476 static struct ib_ah *_rdma_create_ah(struct ib_pd *pd,
477 				     struct rdma_ah_attr *ah_attr,
478 				     struct ib_udata *udata)
479 {
480 	struct ib_ah *ah;
481 
482 	if (!pd->device->create_ah)
483 		return ERR_PTR(-EOPNOTSUPP);
484 
485 	ah = pd->device->create_ah(pd, ah_attr, udata);
486 
487 	if (!IS_ERR(ah)) {
488 		ah->device  = pd->device;
489 		ah->pd      = pd;
490 		ah->uobject = NULL;
491 		ah->type    = ah_attr->type;
492 		ah->sgid_attr = rdma_update_sgid_attr(ah_attr, NULL);
493 
494 		atomic_inc(&pd->usecnt);
495 	}
496 
497 	return ah;
498 }
499 
500 /**
501  * rdma_create_ah - Creates an address handle for the
502  * given address vector.
503  * @pd: The protection domain associated with the address handle.
504  * @ah_attr: The attributes of the address vector.
505  *
506  * It returns 0 on success and returns appropriate error code on error.
507  * The address handle is used to reference a local or global destination
508  * in all UD QP post sends.
509  */
510 struct ib_ah *rdma_create_ah(struct ib_pd *pd, struct rdma_ah_attr *ah_attr)
511 {
512 	const struct ib_gid_attr *old_sgid_attr;
513 	struct ib_ah *ah;
514 	int ret;
515 
516 	ret = rdma_fill_sgid_attr(pd->device, ah_attr, &old_sgid_attr);
517 	if (ret)
518 		return ERR_PTR(ret);
519 
520 	ah = _rdma_create_ah(pd, ah_attr, NULL);
521 
522 	rdma_unfill_sgid_attr(ah_attr, old_sgid_attr);
523 	return ah;
524 }
525 EXPORT_SYMBOL(rdma_create_ah);
526 
527 /**
528  * rdma_create_user_ah - Creates an address handle for the
529  * given address vector.
530  * It resolves destination mac address for ah attribute of RoCE type.
531  * @pd: The protection domain associated with the address handle.
532  * @ah_attr: The attributes of the address vector.
533  * @udata: pointer to user's input output buffer information need by
534  *         provider driver.
535  *
536  * It returns 0 on success and returns appropriate error code on error.
537  * The address handle is used to reference a local or global destination
538  * in all UD QP post sends.
539  */
540 struct ib_ah *rdma_create_user_ah(struct ib_pd *pd,
541 				  struct rdma_ah_attr *ah_attr,
542 				  struct ib_udata *udata)
543 {
544 	const struct ib_gid_attr *old_sgid_attr;
545 	struct ib_ah *ah;
546 	int err;
547 
548 	err = rdma_fill_sgid_attr(pd->device, ah_attr, &old_sgid_attr);
549 	if (err)
550 		return ERR_PTR(err);
551 
552 	if (ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE) {
553 		err = ib_resolve_eth_dmac(pd->device, ah_attr);
554 		if (err) {
555 			ah = ERR_PTR(err);
556 			goto out;
557 		}
558 	}
559 
560 	ah = _rdma_create_ah(pd, ah_attr, udata);
561 
562 out:
563 	rdma_unfill_sgid_attr(ah_attr, old_sgid_attr);
564 	return ah;
565 }
566 EXPORT_SYMBOL(rdma_create_user_ah);
567 
568 int ib_get_rdma_header_version(const union rdma_network_hdr *hdr)
569 {
570 	const struct iphdr *ip4h = (struct iphdr *)&hdr->roce4grh;
571 	struct iphdr ip4h_checked;
572 	const struct ipv6hdr *ip6h = (struct ipv6hdr *)&hdr->ibgrh;
573 
574 	/* If it's IPv6, the version must be 6, otherwise, the first
575 	 * 20 bytes (before the IPv4 header) are garbled.
576 	 */
577 	if (ip6h->version != 6)
578 		return (ip4h->version == 4) ? 4 : 0;
579 	/* version may be 6 or 4 because the first 20 bytes could be garbled */
580 
581 	/* RoCE v2 requires no options, thus header length
582 	 * must be 5 words
583 	 */
584 	if (ip4h->ihl != 5)
585 		return 6;
586 
587 	/* Verify checksum.
588 	 * We can't write on scattered buffers so we need to copy to
589 	 * temp buffer.
590 	 */
591 	memcpy(&ip4h_checked, ip4h, sizeof(ip4h_checked));
592 	ip4h_checked.check = 0;
593 	ip4h_checked.check = ip_fast_csum((u8 *)&ip4h_checked, 5);
594 	/* if IPv4 header checksum is OK, believe it */
595 	if (ip4h->check == ip4h_checked.check)
596 		return 4;
597 	return 6;
598 }
599 EXPORT_SYMBOL(ib_get_rdma_header_version);
600 
601 static enum rdma_network_type ib_get_net_type_by_grh(struct ib_device *device,
602 						     u8 port_num,
603 						     const struct ib_grh *grh)
604 {
605 	int grh_version;
606 
607 	if (rdma_protocol_ib(device, port_num))
608 		return RDMA_NETWORK_IB;
609 
610 	grh_version = ib_get_rdma_header_version((union rdma_network_hdr *)grh);
611 
612 	if (grh_version == 4)
613 		return RDMA_NETWORK_IPV4;
614 
615 	if (grh->next_hdr == IPPROTO_UDP)
616 		return RDMA_NETWORK_IPV6;
617 
618 	return RDMA_NETWORK_ROCE_V1;
619 }
620 
621 struct find_gid_index_context {
622 	u16 vlan_id;
623 	enum ib_gid_type gid_type;
624 };
625 
626 static bool find_gid_index(const union ib_gid *gid,
627 			   const struct ib_gid_attr *gid_attr,
628 			   void *context)
629 {
630 	struct find_gid_index_context *ctx = context;
631 
632 	if (ctx->gid_type != gid_attr->gid_type)
633 		return false;
634 
635 	if ((!!(ctx->vlan_id != 0xffff) == !is_vlan_dev(gid_attr->ndev)) ||
636 	    (is_vlan_dev(gid_attr->ndev) &&
637 	     vlan_dev_vlan_id(gid_attr->ndev) != ctx->vlan_id))
638 		return false;
639 
640 	return true;
641 }
642 
643 static const struct ib_gid_attr *
644 get_sgid_attr_from_eth(struct ib_device *device, u8 port_num,
645 		       u16 vlan_id, const union ib_gid *sgid,
646 		       enum ib_gid_type gid_type)
647 {
648 	struct find_gid_index_context context = {.vlan_id = vlan_id,
649 						 .gid_type = gid_type};
650 
651 	return rdma_find_gid_by_filter(device, sgid, port_num, find_gid_index,
652 				       &context);
653 }
654 
655 int ib_get_gids_from_rdma_hdr(const union rdma_network_hdr *hdr,
656 			      enum rdma_network_type net_type,
657 			      union ib_gid *sgid, union ib_gid *dgid)
658 {
659 	struct sockaddr_in  src_in;
660 	struct sockaddr_in  dst_in;
661 	__be32 src_saddr, dst_saddr;
662 
663 	if (!sgid || !dgid)
664 		return -EINVAL;
665 
666 	if (net_type == RDMA_NETWORK_IPV4) {
667 		memcpy(&src_in.sin_addr.s_addr,
668 		       &hdr->roce4grh.saddr, 4);
669 		memcpy(&dst_in.sin_addr.s_addr,
670 		       &hdr->roce4grh.daddr, 4);
671 		src_saddr = src_in.sin_addr.s_addr;
672 		dst_saddr = dst_in.sin_addr.s_addr;
673 		ipv6_addr_set_v4mapped(src_saddr,
674 				       (struct in6_addr *)sgid);
675 		ipv6_addr_set_v4mapped(dst_saddr,
676 				       (struct in6_addr *)dgid);
677 		return 0;
678 	} else if (net_type == RDMA_NETWORK_IPV6 ||
679 		   net_type == RDMA_NETWORK_IB) {
680 		*dgid = hdr->ibgrh.dgid;
681 		*sgid = hdr->ibgrh.sgid;
682 		return 0;
683 	} else {
684 		return -EINVAL;
685 	}
686 }
687 EXPORT_SYMBOL(ib_get_gids_from_rdma_hdr);
688 
689 /* Resolve destination mac address and hop limit for unicast destination
690  * GID entry, considering the source GID entry as well.
691  * ah_attribute must have have valid port_num, sgid_index.
692  */
693 static int ib_resolve_unicast_gid_dmac(struct ib_device *device,
694 				       struct rdma_ah_attr *ah_attr)
695 {
696 	struct ib_global_route *grh = rdma_ah_retrieve_grh(ah_attr);
697 	const struct ib_gid_attr *sgid_attr = grh->sgid_attr;
698 	int hop_limit = 0xff;
699 	int ret = 0;
700 
701 	/* If destination is link local and source GID is RoCEv1,
702 	 * IP stack is not used.
703 	 */
704 	if (rdma_link_local_addr((struct in6_addr *)grh->dgid.raw) &&
705 	    sgid_attr->gid_type == IB_GID_TYPE_ROCE) {
706 		rdma_get_ll_mac((struct in6_addr *)grh->dgid.raw,
707 				ah_attr->roce.dmac);
708 		return ret;
709 	}
710 
711 	ret = rdma_addr_find_l2_eth_by_grh(&sgid_attr->gid, &grh->dgid,
712 					   ah_attr->roce.dmac,
713 					   sgid_attr, &hop_limit);
714 
715 	grh->hop_limit = hop_limit;
716 	return ret;
717 }
718 
719 /*
720  * This function initializes address handle attributes from the incoming packet.
721  * Incoming packet has dgid of the receiver node on which this code is
722  * getting executed and, sgid contains the GID of the sender.
723  *
724  * When resolving mac address of destination, the arrived dgid is used
725  * as sgid and, sgid is used as dgid because sgid contains destinations
726  * GID whom to respond to.
727  *
728  * On success the caller is responsible to call rdma_destroy_ah_attr on the
729  * attr.
730  */
731 int ib_init_ah_attr_from_wc(struct ib_device *device, u8 port_num,
732 			    const struct ib_wc *wc, const struct ib_grh *grh,
733 			    struct rdma_ah_attr *ah_attr)
734 {
735 	u32 flow_class;
736 	int ret;
737 	enum rdma_network_type net_type = RDMA_NETWORK_IB;
738 	enum ib_gid_type gid_type = IB_GID_TYPE_IB;
739 	const struct ib_gid_attr *sgid_attr;
740 	int hoplimit = 0xff;
741 	union ib_gid dgid;
742 	union ib_gid sgid;
743 
744 	might_sleep();
745 
746 	memset(ah_attr, 0, sizeof *ah_attr);
747 	ah_attr->type = rdma_ah_find_type(device, port_num);
748 	if (rdma_cap_eth_ah(device, port_num)) {
749 		if (wc->wc_flags & IB_WC_WITH_NETWORK_HDR_TYPE)
750 			net_type = wc->network_hdr_type;
751 		else
752 			net_type = ib_get_net_type_by_grh(device, port_num, grh);
753 		gid_type = ib_network_to_gid_type(net_type);
754 	}
755 	ret = ib_get_gids_from_rdma_hdr((union rdma_network_hdr *)grh, net_type,
756 					&sgid, &dgid);
757 	if (ret)
758 		return ret;
759 
760 	rdma_ah_set_sl(ah_attr, wc->sl);
761 	rdma_ah_set_port_num(ah_attr, port_num);
762 
763 	if (rdma_protocol_roce(device, port_num)) {
764 		u16 vlan_id = wc->wc_flags & IB_WC_WITH_VLAN ?
765 				wc->vlan_id : 0xffff;
766 
767 		if (!(wc->wc_flags & IB_WC_GRH))
768 			return -EPROTOTYPE;
769 
770 		sgid_attr = get_sgid_attr_from_eth(device, port_num,
771 						   vlan_id, &dgid,
772 						   gid_type);
773 		if (IS_ERR(sgid_attr))
774 			return PTR_ERR(sgid_attr);
775 
776 		flow_class = be32_to_cpu(grh->version_tclass_flow);
777 		rdma_move_grh_sgid_attr(ah_attr,
778 					&sgid,
779 					flow_class & 0xFFFFF,
780 					hoplimit,
781 					(flow_class >> 20) & 0xFF,
782 					sgid_attr);
783 
784 		ret = ib_resolve_unicast_gid_dmac(device, ah_attr);
785 		if (ret)
786 			rdma_destroy_ah_attr(ah_attr);
787 
788 		return ret;
789 	} else {
790 		rdma_ah_set_dlid(ah_attr, wc->slid);
791 		rdma_ah_set_path_bits(ah_attr, wc->dlid_path_bits);
792 
793 		if ((wc->wc_flags & IB_WC_GRH) == 0)
794 			return 0;
795 
796 		if (dgid.global.interface_id !=
797 					cpu_to_be64(IB_SA_WELL_KNOWN_GUID)) {
798 			sgid_attr = rdma_find_gid_by_port(
799 				device, &dgid, IB_GID_TYPE_IB, port_num, NULL);
800 		} else
801 			sgid_attr = rdma_get_gid_attr(device, port_num, 0);
802 
803 		if (IS_ERR(sgid_attr))
804 			return PTR_ERR(sgid_attr);
805 		flow_class = be32_to_cpu(grh->version_tclass_flow);
806 		rdma_move_grh_sgid_attr(ah_attr,
807 					&sgid,
808 					flow_class & 0xFFFFF,
809 					hoplimit,
810 					(flow_class >> 20) & 0xFF,
811 					sgid_attr);
812 
813 		return 0;
814 	}
815 }
816 EXPORT_SYMBOL(ib_init_ah_attr_from_wc);
817 
818 /**
819  * rdma_move_grh_sgid_attr - Sets the sgid attribute of GRH, taking ownership
820  * of the reference
821  *
822  * @attr:	Pointer to AH attribute structure
823  * @dgid:	Destination GID
824  * @flow_label:	Flow label
825  * @hop_limit:	Hop limit
826  * @traffic_class: traffic class
827  * @sgid_attr:	Pointer to SGID attribute
828  *
829  * This takes ownership of the sgid_attr reference. The caller must ensure
830  * rdma_destroy_ah_attr() is called before destroying the rdma_ah_attr after
831  * calling this function.
832  */
833 void rdma_move_grh_sgid_attr(struct rdma_ah_attr *attr, union ib_gid *dgid,
834 			     u32 flow_label, u8 hop_limit, u8 traffic_class,
835 			     const struct ib_gid_attr *sgid_attr)
836 {
837 	rdma_ah_set_grh(attr, dgid, flow_label, sgid_attr->index, hop_limit,
838 			traffic_class);
839 	attr->grh.sgid_attr = sgid_attr;
840 }
841 EXPORT_SYMBOL(rdma_move_grh_sgid_attr);
842 
843 /**
844  * rdma_destroy_ah_attr - Release reference to SGID attribute of
845  * ah attribute.
846  * @ah_attr: Pointer to ah attribute
847  *
848  * Release reference to the SGID attribute of the ah attribute if it is
849  * non NULL. It is safe to call this multiple times, and safe to call it on
850  * a zero initialized ah_attr.
851  */
852 void rdma_destroy_ah_attr(struct rdma_ah_attr *ah_attr)
853 {
854 	if (ah_attr->grh.sgid_attr) {
855 		rdma_put_gid_attr(ah_attr->grh.sgid_attr);
856 		ah_attr->grh.sgid_attr = NULL;
857 	}
858 }
859 EXPORT_SYMBOL(rdma_destroy_ah_attr);
860 
861 struct ib_ah *ib_create_ah_from_wc(struct ib_pd *pd, const struct ib_wc *wc,
862 				   const struct ib_grh *grh, u8 port_num)
863 {
864 	struct rdma_ah_attr ah_attr;
865 	struct ib_ah *ah;
866 	int ret;
867 
868 	ret = ib_init_ah_attr_from_wc(pd->device, port_num, wc, grh, &ah_attr);
869 	if (ret)
870 		return ERR_PTR(ret);
871 
872 	ah = rdma_create_ah(pd, &ah_attr);
873 
874 	rdma_destroy_ah_attr(&ah_attr);
875 	return ah;
876 }
877 EXPORT_SYMBOL(ib_create_ah_from_wc);
878 
879 int rdma_modify_ah(struct ib_ah *ah, struct rdma_ah_attr *ah_attr)
880 {
881 	const struct ib_gid_attr *old_sgid_attr;
882 	int ret;
883 
884 	if (ah->type != ah_attr->type)
885 		return -EINVAL;
886 
887 	ret = rdma_fill_sgid_attr(ah->device, ah_attr, &old_sgid_attr);
888 	if (ret)
889 		return ret;
890 
891 	ret = ah->device->modify_ah ?
892 		ah->device->modify_ah(ah, ah_attr) :
893 		-EOPNOTSUPP;
894 
895 	ah->sgid_attr = rdma_update_sgid_attr(ah_attr, ah->sgid_attr);
896 	rdma_unfill_sgid_attr(ah_attr, old_sgid_attr);
897 	return ret;
898 }
899 EXPORT_SYMBOL(rdma_modify_ah);
900 
901 int rdma_query_ah(struct ib_ah *ah, struct rdma_ah_attr *ah_attr)
902 {
903 	ah_attr->grh.sgid_attr = NULL;
904 
905 	return ah->device->query_ah ?
906 		ah->device->query_ah(ah, ah_attr) :
907 		-EOPNOTSUPP;
908 }
909 EXPORT_SYMBOL(rdma_query_ah);
910 
911 int rdma_destroy_ah(struct ib_ah *ah)
912 {
913 	const struct ib_gid_attr *sgid_attr = ah->sgid_attr;
914 	struct ib_pd *pd;
915 	int ret;
916 
917 	pd = ah->pd;
918 	ret = ah->device->destroy_ah(ah);
919 	if (!ret) {
920 		atomic_dec(&pd->usecnt);
921 		if (sgid_attr)
922 			rdma_put_gid_attr(sgid_attr);
923 	}
924 
925 	return ret;
926 }
927 EXPORT_SYMBOL(rdma_destroy_ah);
928 
929 /* Shared receive queues */
930 
931 struct ib_srq *ib_create_srq(struct ib_pd *pd,
932 			     struct ib_srq_init_attr *srq_init_attr)
933 {
934 	struct ib_srq *srq;
935 
936 	if (!pd->device->create_srq)
937 		return ERR_PTR(-EOPNOTSUPP);
938 
939 	srq = pd->device->create_srq(pd, srq_init_attr, NULL);
940 
941 	if (!IS_ERR(srq)) {
942 		srq->device    	   = pd->device;
943 		srq->pd        	   = pd;
944 		srq->uobject       = NULL;
945 		srq->event_handler = srq_init_attr->event_handler;
946 		srq->srq_context   = srq_init_attr->srq_context;
947 		srq->srq_type      = srq_init_attr->srq_type;
948 		if (ib_srq_has_cq(srq->srq_type)) {
949 			srq->ext.cq   = srq_init_attr->ext.cq;
950 			atomic_inc(&srq->ext.cq->usecnt);
951 		}
952 		if (srq->srq_type == IB_SRQT_XRC) {
953 			srq->ext.xrc.xrcd = srq_init_attr->ext.xrc.xrcd;
954 			atomic_inc(&srq->ext.xrc.xrcd->usecnt);
955 		}
956 		atomic_inc(&pd->usecnt);
957 		atomic_set(&srq->usecnt, 0);
958 	}
959 
960 	return srq;
961 }
962 EXPORT_SYMBOL(ib_create_srq);
963 
964 int ib_modify_srq(struct ib_srq *srq,
965 		  struct ib_srq_attr *srq_attr,
966 		  enum ib_srq_attr_mask srq_attr_mask)
967 {
968 	return srq->device->modify_srq ?
969 		srq->device->modify_srq(srq, srq_attr, srq_attr_mask, NULL) :
970 		-EOPNOTSUPP;
971 }
972 EXPORT_SYMBOL(ib_modify_srq);
973 
974 int ib_query_srq(struct ib_srq *srq,
975 		 struct ib_srq_attr *srq_attr)
976 {
977 	return srq->device->query_srq ?
978 		srq->device->query_srq(srq, srq_attr) : -EOPNOTSUPP;
979 }
980 EXPORT_SYMBOL(ib_query_srq);
981 
982 int ib_destroy_srq(struct ib_srq *srq)
983 {
984 	struct ib_pd *pd;
985 	enum ib_srq_type srq_type;
986 	struct ib_xrcd *uninitialized_var(xrcd);
987 	struct ib_cq *uninitialized_var(cq);
988 	int ret;
989 
990 	if (atomic_read(&srq->usecnt))
991 		return -EBUSY;
992 
993 	pd = srq->pd;
994 	srq_type = srq->srq_type;
995 	if (ib_srq_has_cq(srq_type))
996 		cq = srq->ext.cq;
997 	if (srq_type == IB_SRQT_XRC)
998 		xrcd = srq->ext.xrc.xrcd;
999 
1000 	ret = srq->device->destroy_srq(srq);
1001 	if (!ret) {
1002 		atomic_dec(&pd->usecnt);
1003 		if (srq_type == IB_SRQT_XRC)
1004 			atomic_dec(&xrcd->usecnt);
1005 		if (ib_srq_has_cq(srq_type))
1006 			atomic_dec(&cq->usecnt);
1007 	}
1008 
1009 	return ret;
1010 }
1011 EXPORT_SYMBOL(ib_destroy_srq);
1012 
1013 /* Queue pairs */
1014 
1015 static void __ib_shared_qp_event_handler(struct ib_event *event, void *context)
1016 {
1017 	struct ib_qp *qp = context;
1018 	unsigned long flags;
1019 
1020 	spin_lock_irqsave(&qp->device->event_handler_lock, flags);
1021 	list_for_each_entry(event->element.qp, &qp->open_list, open_list)
1022 		if (event->element.qp->event_handler)
1023 			event->element.qp->event_handler(event, event->element.qp->qp_context);
1024 	spin_unlock_irqrestore(&qp->device->event_handler_lock, flags);
1025 }
1026 
1027 static void __ib_insert_xrcd_qp(struct ib_xrcd *xrcd, struct ib_qp *qp)
1028 {
1029 	mutex_lock(&xrcd->tgt_qp_mutex);
1030 	list_add(&qp->xrcd_list, &xrcd->tgt_qp_list);
1031 	mutex_unlock(&xrcd->tgt_qp_mutex);
1032 }
1033 
1034 static struct ib_qp *__ib_open_qp(struct ib_qp *real_qp,
1035 				  void (*event_handler)(struct ib_event *, void *),
1036 				  void *qp_context)
1037 {
1038 	struct ib_qp *qp;
1039 	unsigned long flags;
1040 	int err;
1041 
1042 	qp = kzalloc(sizeof *qp, GFP_KERNEL);
1043 	if (!qp)
1044 		return ERR_PTR(-ENOMEM);
1045 
1046 	qp->real_qp = real_qp;
1047 	err = ib_open_shared_qp_security(qp, real_qp->device);
1048 	if (err) {
1049 		kfree(qp);
1050 		return ERR_PTR(err);
1051 	}
1052 
1053 	qp->real_qp = real_qp;
1054 	atomic_inc(&real_qp->usecnt);
1055 	qp->device = real_qp->device;
1056 	qp->event_handler = event_handler;
1057 	qp->qp_context = qp_context;
1058 	qp->qp_num = real_qp->qp_num;
1059 	qp->qp_type = real_qp->qp_type;
1060 
1061 	spin_lock_irqsave(&real_qp->device->event_handler_lock, flags);
1062 	list_add(&qp->open_list, &real_qp->open_list);
1063 	spin_unlock_irqrestore(&real_qp->device->event_handler_lock, flags);
1064 
1065 	return qp;
1066 }
1067 
1068 struct ib_qp *ib_open_qp(struct ib_xrcd *xrcd,
1069 			 struct ib_qp_open_attr *qp_open_attr)
1070 {
1071 	struct ib_qp *qp, *real_qp;
1072 
1073 	if (qp_open_attr->qp_type != IB_QPT_XRC_TGT)
1074 		return ERR_PTR(-EINVAL);
1075 
1076 	qp = ERR_PTR(-EINVAL);
1077 	mutex_lock(&xrcd->tgt_qp_mutex);
1078 	list_for_each_entry(real_qp, &xrcd->tgt_qp_list, xrcd_list) {
1079 		if (real_qp->qp_num == qp_open_attr->qp_num) {
1080 			qp = __ib_open_qp(real_qp, qp_open_attr->event_handler,
1081 					  qp_open_attr->qp_context);
1082 			break;
1083 		}
1084 	}
1085 	mutex_unlock(&xrcd->tgt_qp_mutex);
1086 	return qp;
1087 }
1088 EXPORT_SYMBOL(ib_open_qp);
1089 
1090 static struct ib_qp *ib_create_xrc_qp(struct ib_qp *qp,
1091 		struct ib_qp_init_attr *qp_init_attr)
1092 {
1093 	struct ib_qp *real_qp = qp;
1094 
1095 	qp->event_handler = __ib_shared_qp_event_handler;
1096 	qp->qp_context = qp;
1097 	qp->pd = NULL;
1098 	qp->send_cq = qp->recv_cq = NULL;
1099 	qp->srq = NULL;
1100 	qp->xrcd = qp_init_attr->xrcd;
1101 	atomic_inc(&qp_init_attr->xrcd->usecnt);
1102 	INIT_LIST_HEAD(&qp->open_list);
1103 
1104 	qp = __ib_open_qp(real_qp, qp_init_attr->event_handler,
1105 			  qp_init_attr->qp_context);
1106 	if (!IS_ERR(qp))
1107 		__ib_insert_xrcd_qp(qp_init_attr->xrcd, real_qp);
1108 	else
1109 		real_qp->device->destroy_qp(real_qp);
1110 	return qp;
1111 }
1112 
1113 struct ib_qp *ib_create_qp(struct ib_pd *pd,
1114 			   struct ib_qp_init_attr *qp_init_attr)
1115 {
1116 	struct ib_device *device = pd ? pd->device : qp_init_attr->xrcd->device;
1117 	struct ib_qp *qp;
1118 	int ret;
1119 
1120 	if (qp_init_attr->rwq_ind_tbl &&
1121 	    (qp_init_attr->recv_cq ||
1122 	    qp_init_attr->srq || qp_init_attr->cap.max_recv_wr ||
1123 	    qp_init_attr->cap.max_recv_sge))
1124 		return ERR_PTR(-EINVAL);
1125 
1126 	/*
1127 	 * If the callers is using the RDMA API calculate the resources
1128 	 * needed for the RDMA READ/WRITE operations.
1129 	 *
1130 	 * Note that these callers need to pass in a port number.
1131 	 */
1132 	if (qp_init_attr->cap.max_rdma_ctxs)
1133 		rdma_rw_init_qp(device, qp_init_attr);
1134 
1135 	qp = _ib_create_qp(device, pd, qp_init_attr, NULL, NULL);
1136 	if (IS_ERR(qp))
1137 		return qp;
1138 
1139 	ret = ib_create_qp_security(qp, device);
1140 	if (ret) {
1141 		ib_destroy_qp(qp);
1142 		return ERR_PTR(ret);
1143 	}
1144 
1145 	qp->real_qp    = qp;
1146 	qp->qp_type    = qp_init_attr->qp_type;
1147 	qp->rwq_ind_tbl = qp_init_attr->rwq_ind_tbl;
1148 
1149 	atomic_set(&qp->usecnt, 0);
1150 	qp->mrs_used = 0;
1151 	spin_lock_init(&qp->mr_lock);
1152 	INIT_LIST_HEAD(&qp->rdma_mrs);
1153 	INIT_LIST_HEAD(&qp->sig_mrs);
1154 	qp->port = 0;
1155 
1156 	if (qp_init_attr->qp_type == IB_QPT_XRC_TGT)
1157 		return ib_create_xrc_qp(qp, qp_init_attr);
1158 
1159 	qp->event_handler = qp_init_attr->event_handler;
1160 	qp->qp_context = qp_init_attr->qp_context;
1161 	if (qp_init_attr->qp_type == IB_QPT_XRC_INI) {
1162 		qp->recv_cq = NULL;
1163 		qp->srq = NULL;
1164 	} else {
1165 		qp->recv_cq = qp_init_attr->recv_cq;
1166 		if (qp_init_attr->recv_cq)
1167 			atomic_inc(&qp_init_attr->recv_cq->usecnt);
1168 		qp->srq = qp_init_attr->srq;
1169 		if (qp->srq)
1170 			atomic_inc(&qp_init_attr->srq->usecnt);
1171 	}
1172 
1173 	qp->send_cq = qp_init_attr->send_cq;
1174 	qp->xrcd    = NULL;
1175 
1176 	atomic_inc(&pd->usecnt);
1177 	if (qp_init_attr->send_cq)
1178 		atomic_inc(&qp_init_attr->send_cq->usecnt);
1179 	if (qp_init_attr->rwq_ind_tbl)
1180 		atomic_inc(&qp->rwq_ind_tbl->usecnt);
1181 
1182 	if (qp_init_attr->cap.max_rdma_ctxs) {
1183 		ret = rdma_rw_init_mrs(qp, qp_init_attr);
1184 		if (ret) {
1185 			pr_err("failed to init MR pool ret= %d\n", ret);
1186 			ib_destroy_qp(qp);
1187 			return ERR_PTR(ret);
1188 		}
1189 	}
1190 
1191 	/*
1192 	 * Note: all hw drivers guarantee that max_send_sge is lower than
1193 	 * the device RDMA WRITE SGE limit but not all hw drivers ensure that
1194 	 * max_send_sge <= max_sge_rd.
1195 	 */
1196 	qp->max_write_sge = qp_init_attr->cap.max_send_sge;
1197 	qp->max_read_sge = min_t(u32, qp_init_attr->cap.max_send_sge,
1198 				 device->attrs.max_sge_rd);
1199 
1200 	return qp;
1201 }
1202 EXPORT_SYMBOL(ib_create_qp);
1203 
1204 static const struct {
1205 	int			valid;
1206 	enum ib_qp_attr_mask	req_param[IB_QPT_MAX];
1207 	enum ib_qp_attr_mask	opt_param[IB_QPT_MAX];
1208 } qp_state_table[IB_QPS_ERR + 1][IB_QPS_ERR + 1] = {
1209 	[IB_QPS_RESET] = {
1210 		[IB_QPS_RESET] = { .valid = 1 },
1211 		[IB_QPS_INIT]  = {
1212 			.valid = 1,
1213 			.req_param = {
1214 				[IB_QPT_UD]  = (IB_QP_PKEY_INDEX		|
1215 						IB_QP_PORT			|
1216 						IB_QP_QKEY),
1217 				[IB_QPT_RAW_PACKET] = IB_QP_PORT,
1218 				[IB_QPT_UC]  = (IB_QP_PKEY_INDEX		|
1219 						IB_QP_PORT			|
1220 						IB_QP_ACCESS_FLAGS),
1221 				[IB_QPT_RC]  = (IB_QP_PKEY_INDEX		|
1222 						IB_QP_PORT			|
1223 						IB_QP_ACCESS_FLAGS),
1224 				[IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX		|
1225 						IB_QP_PORT			|
1226 						IB_QP_ACCESS_FLAGS),
1227 				[IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX		|
1228 						IB_QP_PORT			|
1229 						IB_QP_ACCESS_FLAGS),
1230 				[IB_QPT_SMI] = (IB_QP_PKEY_INDEX		|
1231 						IB_QP_QKEY),
1232 				[IB_QPT_GSI] = (IB_QP_PKEY_INDEX		|
1233 						IB_QP_QKEY),
1234 			}
1235 		},
1236 	},
1237 	[IB_QPS_INIT]  = {
1238 		[IB_QPS_RESET] = { .valid = 1 },
1239 		[IB_QPS_ERR] =   { .valid = 1 },
1240 		[IB_QPS_INIT]  = {
1241 			.valid = 1,
1242 			.opt_param = {
1243 				[IB_QPT_UD]  = (IB_QP_PKEY_INDEX		|
1244 						IB_QP_PORT			|
1245 						IB_QP_QKEY),
1246 				[IB_QPT_UC]  = (IB_QP_PKEY_INDEX		|
1247 						IB_QP_PORT			|
1248 						IB_QP_ACCESS_FLAGS),
1249 				[IB_QPT_RC]  = (IB_QP_PKEY_INDEX		|
1250 						IB_QP_PORT			|
1251 						IB_QP_ACCESS_FLAGS),
1252 				[IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX		|
1253 						IB_QP_PORT			|
1254 						IB_QP_ACCESS_FLAGS),
1255 				[IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX		|
1256 						IB_QP_PORT			|
1257 						IB_QP_ACCESS_FLAGS),
1258 				[IB_QPT_SMI] = (IB_QP_PKEY_INDEX		|
1259 						IB_QP_QKEY),
1260 				[IB_QPT_GSI] = (IB_QP_PKEY_INDEX		|
1261 						IB_QP_QKEY),
1262 			}
1263 		},
1264 		[IB_QPS_RTR]   = {
1265 			.valid = 1,
1266 			.req_param = {
1267 				[IB_QPT_UC]  = (IB_QP_AV			|
1268 						IB_QP_PATH_MTU			|
1269 						IB_QP_DEST_QPN			|
1270 						IB_QP_RQ_PSN),
1271 				[IB_QPT_RC]  = (IB_QP_AV			|
1272 						IB_QP_PATH_MTU			|
1273 						IB_QP_DEST_QPN			|
1274 						IB_QP_RQ_PSN			|
1275 						IB_QP_MAX_DEST_RD_ATOMIC	|
1276 						IB_QP_MIN_RNR_TIMER),
1277 				[IB_QPT_XRC_INI] = (IB_QP_AV			|
1278 						IB_QP_PATH_MTU			|
1279 						IB_QP_DEST_QPN			|
1280 						IB_QP_RQ_PSN),
1281 				[IB_QPT_XRC_TGT] = (IB_QP_AV			|
1282 						IB_QP_PATH_MTU			|
1283 						IB_QP_DEST_QPN			|
1284 						IB_QP_RQ_PSN			|
1285 						IB_QP_MAX_DEST_RD_ATOMIC	|
1286 						IB_QP_MIN_RNR_TIMER),
1287 			},
1288 			.opt_param = {
1289 				 [IB_QPT_UD]  = (IB_QP_PKEY_INDEX		|
1290 						 IB_QP_QKEY),
1291 				 [IB_QPT_UC]  = (IB_QP_ALT_PATH			|
1292 						 IB_QP_ACCESS_FLAGS		|
1293 						 IB_QP_PKEY_INDEX),
1294 				 [IB_QPT_RC]  = (IB_QP_ALT_PATH			|
1295 						 IB_QP_ACCESS_FLAGS		|
1296 						 IB_QP_PKEY_INDEX),
1297 				 [IB_QPT_XRC_INI] = (IB_QP_ALT_PATH		|
1298 						 IB_QP_ACCESS_FLAGS		|
1299 						 IB_QP_PKEY_INDEX),
1300 				 [IB_QPT_XRC_TGT] = (IB_QP_ALT_PATH		|
1301 						 IB_QP_ACCESS_FLAGS		|
1302 						 IB_QP_PKEY_INDEX),
1303 				 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX		|
1304 						 IB_QP_QKEY),
1305 				 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX		|
1306 						 IB_QP_QKEY),
1307 			 },
1308 		},
1309 	},
1310 	[IB_QPS_RTR]   = {
1311 		[IB_QPS_RESET] = { .valid = 1 },
1312 		[IB_QPS_ERR] =   { .valid = 1 },
1313 		[IB_QPS_RTS]   = {
1314 			.valid = 1,
1315 			.req_param = {
1316 				[IB_QPT_UD]  = IB_QP_SQ_PSN,
1317 				[IB_QPT_UC]  = IB_QP_SQ_PSN,
1318 				[IB_QPT_RC]  = (IB_QP_TIMEOUT			|
1319 						IB_QP_RETRY_CNT			|
1320 						IB_QP_RNR_RETRY			|
1321 						IB_QP_SQ_PSN			|
1322 						IB_QP_MAX_QP_RD_ATOMIC),
1323 				[IB_QPT_XRC_INI] = (IB_QP_TIMEOUT		|
1324 						IB_QP_RETRY_CNT			|
1325 						IB_QP_RNR_RETRY			|
1326 						IB_QP_SQ_PSN			|
1327 						IB_QP_MAX_QP_RD_ATOMIC),
1328 				[IB_QPT_XRC_TGT] = (IB_QP_TIMEOUT		|
1329 						IB_QP_SQ_PSN),
1330 				[IB_QPT_SMI] = IB_QP_SQ_PSN,
1331 				[IB_QPT_GSI] = IB_QP_SQ_PSN,
1332 			},
1333 			.opt_param = {
1334 				 [IB_QPT_UD]  = (IB_QP_CUR_STATE		|
1335 						 IB_QP_QKEY),
1336 				 [IB_QPT_UC]  = (IB_QP_CUR_STATE		|
1337 						 IB_QP_ALT_PATH			|
1338 						 IB_QP_ACCESS_FLAGS		|
1339 						 IB_QP_PATH_MIG_STATE),
1340 				 [IB_QPT_RC]  = (IB_QP_CUR_STATE		|
1341 						 IB_QP_ALT_PATH			|
1342 						 IB_QP_ACCESS_FLAGS		|
1343 						 IB_QP_MIN_RNR_TIMER		|
1344 						 IB_QP_PATH_MIG_STATE),
1345 				 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE		|
1346 						 IB_QP_ALT_PATH			|
1347 						 IB_QP_ACCESS_FLAGS		|
1348 						 IB_QP_PATH_MIG_STATE),
1349 				 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE		|
1350 						 IB_QP_ALT_PATH			|
1351 						 IB_QP_ACCESS_FLAGS		|
1352 						 IB_QP_MIN_RNR_TIMER		|
1353 						 IB_QP_PATH_MIG_STATE),
1354 				 [IB_QPT_SMI] = (IB_QP_CUR_STATE		|
1355 						 IB_QP_QKEY),
1356 				 [IB_QPT_GSI] = (IB_QP_CUR_STATE		|
1357 						 IB_QP_QKEY),
1358 				 [IB_QPT_RAW_PACKET] = IB_QP_RATE_LIMIT,
1359 			 }
1360 		}
1361 	},
1362 	[IB_QPS_RTS]   = {
1363 		[IB_QPS_RESET] = { .valid = 1 },
1364 		[IB_QPS_ERR] =   { .valid = 1 },
1365 		[IB_QPS_RTS]   = {
1366 			.valid = 1,
1367 			.opt_param = {
1368 				[IB_QPT_UD]  = (IB_QP_CUR_STATE			|
1369 						IB_QP_QKEY),
1370 				[IB_QPT_UC]  = (IB_QP_CUR_STATE			|
1371 						IB_QP_ACCESS_FLAGS		|
1372 						IB_QP_ALT_PATH			|
1373 						IB_QP_PATH_MIG_STATE),
1374 				[IB_QPT_RC]  = (IB_QP_CUR_STATE			|
1375 						IB_QP_ACCESS_FLAGS		|
1376 						IB_QP_ALT_PATH			|
1377 						IB_QP_PATH_MIG_STATE		|
1378 						IB_QP_MIN_RNR_TIMER),
1379 				[IB_QPT_XRC_INI] = (IB_QP_CUR_STATE		|
1380 						IB_QP_ACCESS_FLAGS		|
1381 						IB_QP_ALT_PATH			|
1382 						IB_QP_PATH_MIG_STATE),
1383 				[IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE		|
1384 						IB_QP_ACCESS_FLAGS		|
1385 						IB_QP_ALT_PATH			|
1386 						IB_QP_PATH_MIG_STATE		|
1387 						IB_QP_MIN_RNR_TIMER),
1388 				[IB_QPT_SMI] = (IB_QP_CUR_STATE			|
1389 						IB_QP_QKEY),
1390 				[IB_QPT_GSI] = (IB_QP_CUR_STATE			|
1391 						IB_QP_QKEY),
1392 				[IB_QPT_RAW_PACKET] = IB_QP_RATE_LIMIT,
1393 			}
1394 		},
1395 		[IB_QPS_SQD]   = {
1396 			.valid = 1,
1397 			.opt_param = {
1398 				[IB_QPT_UD]  = IB_QP_EN_SQD_ASYNC_NOTIFY,
1399 				[IB_QPT_UC]  = IB_QP_EN_SQD_ASYNC_NOTIFY,
1400 				[IB_QPT_RC]  = IB_QP_EN_SQD_ASYNC_NOTIFY,
1401 				[IB_QPT_XRC_INI] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1402 				[IB_QPT_XRC_TGT] = IB_QP_EN_SQD_ASYNC_NOTIFY, /* ??? */
1403 				[IB_QPT_SMI] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1404 				[IB_QPT_GSI] = IB_QP_EN_SQD_ASYNC_NOTIFY
1405 			}
1406 		},
1407 	},
1408 	[IB_QPS_SQD]   = {
1409 		[IB_QPS_RESET] = { .valid = 1 },
1410 		[IB_QPS_ERR] =   { .valid = 1 },
1411 		[IB_QPS_RTS]   = {
1412 			.valid = 1,
1413 			.opt_param = {
1414 				[IB_QPT_UD]  = (IB_QP_CUR_STATE			|
1415 						IB_QP_QKEY),
1416 				[IB_QPT_UC]  = (IB_QP_CUR_STATE			|
1417 						IB_QP_ALT_PATH			|
1418 						IB_QP_ACCESS_FLAGS		|
1419 						IB_QP_PATH_MIG_STATE),
1420 				[IB_QPT_RC]  = (IB_QP_CUR_STATE			|
1421 						IB_QP_ALT_PATH			|
1422 						IB_QP_ACCESS_FLAGS		|
1423 						IB_QP_MIN_RNR_TIMER		|
1424 						IB_QP_PATH_MIG_STATE),
1425 				[IB_QPT_XRC_INI] = (IB_QP_CUR_STATE		|
1426 						IB_QP_ALT_PATH			|
1427 						IB_QP_ACCESS_FLAGS		|
1428 						IB_QP_PATH_MIG_STATE),
1429 				[IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE		|
1430 						IB_QP_ALT_PATH			|
1431 						IB_QP_ACCESS_FLAGS		|
1432 						IB_QP_MIN_RNR_TIMER		|
1433 						IB_QP_PATH_MIG_STATE),
1434 				[IB_QPT_SMI] = (IB_QP_CUR_STATE			|
1435 						IB_QP_QKEY),
1436 				[IB_QPT_GSI] = (IB_QP_CUR_STATE			|
1437 						IB_QP_QKEY),
1438 			}
1439 		},
1440 		[IB_QPS_SQD]   = {
1441 			.valid = 1,
1442 			.opt_param = {
1443 				[IB_QPT_UD]  = (IB_QP_PKEY_INDEX		|
1444 						IB_QP_QKEY),
1445 				[IB_QPT_UC]  = (IB_QP_AV			|
1446 						IB_QP_ALT_PATH			|
1447 						IB_QP_ACCESS_FLAGS		|
1448 						IB_QP_PKEY_INDEX		|
1449 						IB_QP_PATH_MIG_STATE),
1450 				[IB_QPT_RC]  = (IB_QP_PORT			|
1451 						IB_QP_AV			|
1452 						IB_QP_TIMEOUT			|
1453 						IB_QP_RETRY_CNT			|
1454 						IB_QP_RNR_RETRY			|
1455 						IB_QP_MAX_QP_RD_ATOMIC		|
1456 						IB_QP_MAX_DEST_RD_ATOMIC	|
1457 						IB_QP_ALT_PATH			|
1458 						IB_QP_ACCESS_FLAGS		|
1459 						IB_QP_PKEY_INDEX		|
1460 						IB_QP_MIN_RNR_TIMER		|
1461 						IB_QP_PATH_MIG_STATE),
1462 				[IB_QPT_XRC_INI] = (IB_QP_PORT			|
1463 						IB_QP_AV			|
1464 						IB_QP_TIMEOUT			|
1465 						IB_QP_RETRY_CNT			|
1466 						IB_QP_RNR_RETRY			|
1467 						IB_QP_MAX_QP_RD_ATOMIC		|
1468 						IB_QP_ALT_PATH			|
1469 						IB_QP_ACCESS_FLAGS		|
1470 						IB_QP_PKEY_INDEX		|
1471 						IB_QP_PATH_MIG_STATE),
1472 				[IB_QPT_XRC_TGT] = (IB_QP_PORT			|
1473 						IB_QP_AV			|
1474 						IB_QP_TIMEOUT			|
1475 						IB_QP_MAX_DEST_RD_ATOMIC	|
1476 						IB_QP_ALT_PATH			|
1477 						IB_QP_ACCESS_FLAGS		|
1478 						IB_QP_PKEY_INDEX		|
1479 						IB_QP_MIN_RNR_TIMER		|
1480 						IB_QP_PATH_MIG_STATE),
1481 				[IB_QPT_SMI] = (IB_QP_PKEY_INDEX		|
1482 						IB_QP_QKEY),
1483 				[IB_QPT_GSI] = (IB_QP_PKEY_INDEX		|
1484 						IB_QP_QKEY),
1485 			}
1486 		}
1487 	},
1488 	[IB_QPS_SQE]   = {
1489 		[IB_QPS_RESET] = { .valid = 1 },
1490 		[IB_QPS_ERR] =   { .valid = 1 },
1491 		[IB_QPS_RTS]   = {
1492 			.valid = 1,
1493 			.opt_param = {
1494 				[IB_QPT_UD]  = (IB_QP_CUR_STATE			|
1495 						IB_QP_QKEY),
1496 				[IB_QPT_UC]  = (IB_QP_CUR_STATE			|
1497 						IB_QP_ACCESS_FLAGS),
1498 				[IB_QPT_SMI] = (IB_QP_CUR_STATE			|
1499 						IB_QP_QKEY),
1500 				[IB_QPT_GSI] = (IB_QP_CUR_STATE			|
1501 						IB_QP_QKEY),
1502 			}
1503 		}
1504 	},
1505 	[IB_QPS_ERR] = {
1506 		[IB_QPS_RESET] = { .valid = 1 },
1507 		[IB_QPS_ERR] =   { .valid = 1 }
1508 	}
1509 };
1510 
1511 bool ib_modify_qp_is_ok(enum ib_qp_state cur_state, enum ib_qp_state next_state,
1512 			enum ib_qp_type type, enum ib_qp_attr_mask mask)
1513 {
1514 	enum ib_qp_attr_mask req_param, opt_param;
1515 
1516 	if (mask & IB_QP_CUR_STATE  &&
1517 	    cur_state != IB_QPS_RTR && cur_state != IB_QPS_RTS &&
1518 	    cur_state != IB_QPS_SQD && cur_state != IB_QPS_SQE)
1519 		return false;
1520 
1521 	if (!qp_state_table[cur_state][next_state].valid)
1522 		return false;
1523 
1524 	req_param = qp_state_table[cur_state][next_state].req_param[type];
1525 	opt_param = qp_state_table[cur_state][next_state].opt_param[type];
1526 
1527 	if ((mask & req_param) != req_param)
1528 		return false;
1529 
1530 	if (mask & ~(req_param | opt_param | IB_QP_STATE))
1531 		return false;
1532 
1533 	return true;
1534 }
1535 EXPORT_SYMBOL(ib_modify_qp_is_ok);
1536 
1537 /**
1538  * ib_resolve_eth_dmac - Resolve destination mac address
1539  * @device:		Device to consider
1540  * @ah_attr:		address handle attribute which describes the
1541  *			source and destination parameters
1542  * ib_resolve_eth_dmac() resolves destination mac address and L3 hop limit It
1543  * returns 0 on success or appropriate error code. It initializes the
1544  * necessary ah_attr fields when call is successful.
1545  */
1546 static int ib_resolve_eth_dmac(struct ib_device *device,
1547 			       struct rdma_ah_attr *ah_attr)
1548 {
1549 	int ret = 0;
1550 
1551 	if (rdma_is_multicast_addr((struct in6_addr *)ah_attr->grh.dgid.raw)) {
1552 		if (ipv6_addr_v4mapped((struct in6_addr *)ah_attr->grh.dgid.raw)) {
1553 			__be32 addr = 0;
1554 
1555 			memcpy(&addr, ah_attr->grh.dgid.raw + 12, 4);
1556 			ip_eth_mc_map(addr, (char *)ah_attr->roce.dmac);
1557 		} else {
1558 			ipv6_eth_mc_map((struct in6_addr *)ah_attr->grh.dgid.raw,
1559 					(char *)ah_attr->roce.dmac);
1560 		}
1561 	} else {
1562 		ret = ib_resolve_unicast_gid_dmac(device, ah_attr);
1563 	}
1564 	return ret;
1565 }
1566 
1567 static bool is_qp_type_connected(const struct ib_qp *qp)
1568 {
1569 	return (qp->qp_type == IB_QPT_UC ||
1570 		qp->qp_type == IB_QPT_RC ||
1571 		qp->qp_type == IB_QPT_XRC_INI ||
1572 		qp->qp_type == IB_QPT_XRC_TGT);
1573 }
1574 
1575 /**
1576  * IB core internal function to perform QP attributes modification.
1577  */
1578 static int _ib_modify_qp(struct ib_qp *qp, struct ib_qp_attr *attr,
1579 			 int attr_mask, struct ib_udata *udata)
1580 {
1581 	u8 port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
1582 	const struct ib_gid_attr *old_sgid_attr_av;
1583 	const struct ib_gid_attr *old_sgid_attr_alt_av;
1584 	int ret;
1585 
1586 	if (attr_mask & IB_QP_AV) {
1587 		ret = rdma_fill_sgid_attr(qp->device, &attr->ah_attr,
1588 					  &old_sgid_attr_av);
1589 		if (ret)
1590 			return ret;
1591 	}
1592 	if (attr_mask & IB_QP_ALT_PATH) {
1593 		/*
1594 		 * FIXME: This does not track the migration state, so if the
1595 		 * user loads a new alternate path after the HW has migrated
1596 		 * from primary->alternate we will keep the wrong
1597 		 * references. This is OK for IB because the reference
1598 		 * counting does not serve any functional purpose.
1599 		 */
1600 		ret = rdma_fill_sgid_attr(qp->device, &attr->alt_ah_attr,
1601 					  &old_sgid_attr_alt_av);
1602 		if (ret)
1603 			goto out_av;
1604 
1605 		/*
1606 		 * Today the core code can only handle alternate paths and APM
1607 		 * for IB. Ban them in roce mode.
1608 		 */
1609 		if (!(rdma_protocol_ib(qp->device,
1610 				       attr->alt_ah_attr.port_num) &&
1611 		      rdma_protocol_ib(qp->device, port))) {
1612 			ret = EINVAL;
1613 			goto out;
1614 		}
1615 	}
1616 
1617 	/*
1618 	 * If the user provided the qp_attr then we have to resolve it. Kernel
1619 	 * users have to provide already resolved rdma_ah_attr's
1620 	 */
1621 	if (udata && (attr_mask & IB_QP_AV) &&
1622 	    attr->ah_attr.type == RDMA_AH_ATTR_TYPE_ROCE &&
1623 	    is_qp_type_connected(qp)) {
1624 		ret = ib_resolve_eth_dmac(qp->device, &attr->ah_attr);
1625 		if (ret)
1626 			goto out;
1627 	}
1628 
1629 	if (rdma_ib_or_roce(qp->device, port)) {
1630 		if (attr_mask & IB_QP_RQ_PSN && attr->rq_psn & ~0xffffff) {
1631 			dev_warn(&qp->device->dev,
1632 				 "%s rq_psn overflow, masking to 24 bits\n",
1633 				 __func__);
1634 			attr->rq_psn &= 0xffffff;
1635 		}
1636 
1637 		if (attr_mask & IB_QP_SQ_PSN && attr->sq_psn & ~0xffffff) {
1638 			dev_warn(&qp->device->dev,
1639 				 " %s sq_psn overflow, masking to 24 bits\n",
1640 				 __func__);
1641 			attr->sq_psn &= 0xffffff;
1642 		}
1643 	}
1644 
1645 	ret = ib_security_modify_qp(qp, attr, attr_mask, udata);
1646 	if (ret)
1647 		goto out;
1648 
1649 	if (attr_mask & IB_QP_PORT)
1650 		qp->port = attr->port_num;
1651 	if (attr_mask & IB_QP_AV)
1652 		qp->av_sgid_attr =
1653 			rdma_update_sgid_attr(&attr->ah_attr, qp->av_sgid_attr);
1654 	if (attr_mask & IB_QP_ALT_PATH)
1655 		qp->alt_path_sgid_attr = rdma_update_sgid_attr(
1656 			&attr->alt_ah_attr, qp->alt_path_sgid_attr);
1657 
1658 out:
1659 	if (attr_mask & IB_QP_ALT_PATH)
1660 		rdma_unfill_sgid_attr(&attr->alt_ah_attr, old_sgid_attr_alt_av);
1661 out_av:
1662 	if (attr_mask & IB_QP_AV)
1663 		rdma_unfill_sgid_attr(&attr->ah_attr, old_sgid_attr_av);
1664 	return ret;
1665 }
1666 
1667 /**
1668  * ib_modify_qp_with_udata - Modifies the attributes for the specified QP.
1669  * @ib_qp: The QP to modify.
1670  * @attr: On input, specifies the QP attributes to modify.  On output,
1671  *   the current values of selected QP attributes are returned.
1672  * @attr_mask: A bit-mask used to specify which attributes of the QP
1673  *   are being modified.
1674  * @udata: pointer to user's input output buffer information
1675  *   are being modified.
1676  * It returns 0 on success and returns appropriate error code on error.
1677  */
1678 int ib_modify_qp_with_udata(struct ib_qp *ib_qp, struct ib_qp_attr *attr,
1679 			    int attr_mask, struct ib_udata *udata)
1680 {
1681 	return _ib_modify_qp(ib_qp->real_qp, attr, attr_mask, udata);
1682 }
1683 EXPORT_SYMBOL(ib_modify_qp_with_udata);
1684 
1685 int ib_get_eth_speed(struct ib_device *dev, u8 port_num, u8 *speed, u8 *width)
1686 {
1687 	int rc;
1688 	u32 netdev_speed;
1689 	struct net_device *netdev;
1690 	struct ethtool_link_ksettings lksettings;
1691 
1692 	if (rdma_port_get_link_layer(dev, port_num) != IB_LINK_LAYER_ETHERNET)
1693 		return -EINVAL;
1694 
1695 	if (!dev->get_netdev)
1696 		return -EOPNOTSUPP;
1697 
1698 	netdev = dev->get_netdev(dev, port_num);
1699 	if (!netdev)
1700 		return -ENODEV;
1701 
1702 	rtnl_lock();
1703 	rc = __ethtool_get_link_ksettings(netdev, &lksettings);
1704 	rtnl_unlock();
1705 
1706 	dev_put(netdev);
1707 
1708 	if (!rc) {
1709 		netdev_speed = lksettings.base.speed;
1710 	} else {
1711 		netdev_speed = SPEED_1000;
1712 		pr_warn("%s speed is unknown, defaulting to %d\n", netdev->name,
1713 			netdev_speed);
1714 	}
1715 
1716 	if (netdev_speed <= SPEED_1000) {
1717 		*width = IB_WIDTH_1X;
1718 		*speed = IB_SPEED_SDR;
1719 	} else if (netdev_speed <= SPEED_10000) {
1720 		*width = IB_WIDTH_1X;
1721 		*speed = IB_SPEED_FDR10;
1722 	} else if (netdev_speed <= SPEED_20000) {
1723 		*width = IB_WIDTH_4X;
1724 		*speed = IB_SPEED_DDR;
1725 	} else if (netdev_speed <= SPEED_25000) {
1726 		*width = IB_WIDTH_1X;
1727 		*speed = IB_SPEED_EDR;
1728 	} else if (netdev_speed <= SPEED_40000) {
1729 		*width = IB_WIDTH_4X;
1730 		*speed = IB_SPEED_FDR10;
1731 	} else {
1732 		*width = IB_WIDTH_4X;
1733 		*speed = IB_SPEED_EDR;
1734 	}
1735 
1736 	return 0;
1737 }
1738 EXPORT_SYMBOL(ib_get_eth_speed);
1739 
1740 int ib_modify_qp(struct ib_qp *qp,
1741 		 struct ib_qp_attr *qp_attr,
1742 		 int qp_attr_mask)
1743 {
1744 	return _ib_modify_qp(qp->real_qp, qp_attr, qp_attr_mask, NULL);
1745 }
1746 EXPORT_SYMBOL(ib_modify_qp);
1747 
1748 int ib_query_qp(struct ib_qp *qp,
1749 		struct ib_qp_attr *qp_attr,
1750 		int qp_attr_mask,
1751 		struct ib_qp_init_attr *qp_init_attr)
1752 {
1753 	qp_attr->ah_attr.grh.sgid_attr = NULL;
1754 	qp_attr->alt_ah_attr.grh.sgid_attr = NULL;
1755 
1756 	return qp->device->query_qp ?
1757 		qp->device->query_qp(qp->real_qp, qp_attr, qp_attr_mask, qp_init_attr) :
1758 		-EOPNOTSUPP;
1759 }
1760 EXPORT_SYMBOL(ib_query_qp);
1761 
1762 int ib_close_qp(struct ib_qp *qp)
1763 {
1764 	struct ib_qp *real_qp;
1765 	unsigned long flags;
1766 
1767 	real_qp = qp->real_qp;
1768 	if (real_qp == qp)
1769 		return -EINVAL;
1770 
1771 	spin_lock_irqsave(&real_qp->device->event_handler_lock, flags);
1772 	list_del(&qp->open_list);
1773 	spin_unlock_irqrestore(&real_qp->device->event_handler_lock, flags);
1774 
1775 	atomic_dec(&real_qp->usecnt);
1776 	if (qp->qp_sec)
1777 		ib_close_shared_qp_security(qp->qp_sec);
1778 	kfree(qp);
1779 
1780 	return 0;
1781 }
1782 EXPORT_SYMBOL(ib_close_qp);
1783 
1784 static int __ib_destroy_shared_qp(struct ib_qp *qp)
1785 {
1786 	struct ib_xrcd *xrcd;
1787 	struct ib_qp *real_qp;
1788 	int ret;
1789 
1790 	real_qp = qp->real_qp;
1791 	xrcd = real_qp->xrcd;
1792 
1793 	mutex_lock(&xrcd->tgt_qp_mutex);
1794 	ib_close_qp(qp);
1795 	if (atomic_read(&real_qp->usecnt) == 0)
1796 		list_del(&real_qp->xrcd_list);
1797 	else
1798 		real_qp = NULL;
1799 	mutex_unlock(&xrcd->tgt_qp_mutex);
1800 
1801 	if (real_qp) {
1802 		ret = ib_destroy_qp(real_qp);
1803 		if (!ret)
1804 			atomic_dec(&xrcd->usecnt);
1805 		else
1806 			__ib_insert_xrcd_qp(xrcd, real_qp);
1807 	}
1808 
1809 	return 0;
1810 }
1811 
1812 int ib_destroy_qp(struct ib_qp *qp)
1813 {
1814 	const struct ib_gid_attr *alt_path_sgid_attr = qp->alt_path_sgid_attr;
1815 	const struct ib_gid_attr *av_sgid_attr = qp->av_sgid_attr;
1816 	struct ib_pd *pd;
1817 	struct ib_cq *scq, *rcq;
1818 	struct ib_srq *srq;
1819 	struct ib_rwq_ind_table *ind_tbl;
1820 	struct ib_qp_security *sec;
1821 	int ret;
1822 
1823 	WARN_ON_ONCE(qp->mrs_used > 0);
1824 
1825 	if (atomic_read(&qp->usecnt))
1826 		return -EBUSY;
1827 
1828 	if (qp->real_qp != qp)
1829 		return __ib_destroy_shared_qp(qp);
1830 
1831 	pd   = qp->pd;
1832 	scq  = qp->send_cq;
1833 	rcq  = qp->recv_cq;
1834 	srq  = qp->srq;
1835 	ind_tbl = qp->rwq_ind_tbl;
1836 	sec  = qp->qp_sec;
1837 	if (sec)
1838 		ib_destroy_qp_security_begin(sec);
1839 
1840 	if (!qp->uobject)
1841 		rdma_rw_cleanup_mrs(qp);
1842 
1843 	rdma_restrack_del(&qp->res);
1844 	ret = qp->device->destroy_qp(qp);
1845 	if (!ret) {
1846 		if (alt_path_sgid_attr)
1847 			rdma_put_gid_attr(alt_path_sgid_attr);
1848 		if (av_sgid_attr)
1849 			rdma_put_gid_attr(av_sgid_attr);
1850 		if (pd)
1851 			atomic_dec(&pd->usecnt);
1852 		if (scq)
1853 			atomic_dec(&scq->usecnt);
1854 		if (rcq)
1855 			atomic_dec(&rcq->usecnt);
1856 		if (srq)
1857 			atomic_dec(&srq->usecnt);
1858 		if (ind_tbl)
1859 			atomic_dec(&ind_tbl->usecnt);
1860 		if (sec)
1861 			ib_destroy_qp_security_end(sec);
1862 	} else {
1863 		if (sec)
1864 			ib_destroy_qp_security_abort(sec);
1865 	}
1866 
1867 	return ret;
1868 }
1869 EXPORT_SYMBOL(ib_destroy_qp);
1870 
1871 /* Completion queues */
1872 
1873 struct ib_cq *__ib_create_cq(struct ib_device *device,
1874 			     ib_comp_handler comp_handler,
1875 			     void (*event_handler)(struct ib_event *, void *),
1876 			     void *cq_context,
1877 			     const struct ib_cq_init_attr *cq_attr,
1878 			     const char *caller)
1879 {
1880 	struct ib_cq *cq;
1881 
1882 	cq = device->create_cq(device, cq_attr, NULL, NULL);
1883 
1884 	if (!IS_ERR(cq)) {
1885 		cq->device        = device;
1886 		cq->uobject       = NULL;
1887 		cq->comp_handler  = comp_handler;
1888 		cq->event_handler = event_handler;
1889 		cq->cq_context    = cq_context;
1890 		atomic_set(&cq->usecnt, 0);
1891 		cq->res.type = RDMA_RESTRACK_CQ;
1892 		rdma_restrack_set_task(&cq->res, caller);
1893 		rdma_restrack_add(&cq->res);
1894 	}
1895 
1896 	return cq;
1897 }
1898 EXPORT_SYMBOL(__ib_create_cq);
1899 
1900 int rdma_set_cq_moderation(struct ib_cq *cq, u16 cq_count, u16 cq_period)
1901 {
1902 	return cq->device->modify_cq ?
1903 		cq->device->modify_cq(cq, cq_count, cq_period) : -EOPNOTSUPP;
1904 }
1905 EXPORT_SYMBOL(rdma_set_cq_moderation);
1906 
1907 int ib_destroy_cq(struct ib_cq *cq)
1908 {
1909 	if (atomic_read(&cq->usecnt))
1910 		return -EBUSY;
1911 
1912 	rdma_restrack_del(&cq->res);
1913 	return cq->device->destroy_cq(cq);
1914 }
1915 EXPORT_SYMBOL(ib_destroy_cq);
1916 
1917 int ib_resize_cq(struct ib_cq *cq, int cqe)
1918 {
1919 	return cq->device->resize_cq ?
1920 		cq->device->resize_cq(cq, cqe, NULL) : -EOPNOTSUPP;
1921 }
1922 EXPORT_SYMBOL(ib_resize_cq);
1923 
1924 /* Memory regions */
1925 
1926 int ib_dereg_mr(struct ib_mr *mr)
1927 {
1928 	struct ib_pd *pd = mr->pd;
1929 	struct ib_dm *dm = mr->dm;
1930 	int ret;
1931 
1932 	rdma_restrack_del(&mr->res);
1933 	ret = mr->device->dereg_mr(mr);
1934 	if (!ret) {
1935 		atomic_dec(&pd->usecnt);
1936 		if (dm)
1937 			atomic_dec(&dm->usecnt);
1938 	}
1939 
1940 	return ret;
1941 }
1942 EXPORT_SYMBOL(ib_dereg_mr);
1943 
1944 /**
1945  * ib_alloc_mr() - Allocates a memory region
1946  * @pd:            protection domain associated with the region
1947  * @mr_type:       memory region type
1948  * @max_num_sg:    maximum sg entries available for registration.
1949  *
1950  * Notes:
1951  * Memory registeration page/sg lists must not exceed max_num_sg.
1952  * For mr_type IB_MR_TYPE_MEM_REG, the total length cannot exceed
1953  * max_num_sg * used_page_size.
1954  *
1955  */
1956 struct ib_mr *ib_alloc_mr(struct ib_pd *pd,
1957 			  enum ib_mr_type mr_type,
1958 			  u32 max_num_sg)
1959 {
1960 	struct ib_mr *mr;
1961 
1962 	if (!pd->device->alloc_mr)
1963 		return ERR_PTR(-EOPNOTSUPP);
1964 
1965 	mr = pd->device->alloc_mr(pd, mr_type, max_num_sg);
1966 	if (!IS_ERR(mr)) {
1967 		mr->device  = pd->device;
1968 		mr->pd      = pd;
1969 		mr->dm      = NULL;
1970 		mr->uobject = NULL;
1971 		atomic_inc(&pd->usecnt);
1972 		mr->need_inval = false;
1973 		mr->res.type = RDMA_RESTRACK_MR;
1974 		rdma_restrack_add(&mr->res);
1975 	}
1976 
1977 	return mr;
1978 }
1979 EXPORT_SYMBOL(ib_alloc_mr);
1980 
1981 /* "Fast" memory regions */
1982 
1983 struct ib_fmr *ib_alloc_fmr(struct ib_pd *pd,
1984 			    int mr_access_flags,
1985 			    struct ib_fmr_attr *fmr_attr)
1986 {
1987 	struct ib_fmr *fmr;
1988 
1989 	if (!pd->device->alloc_fmr)
1990 		return ERR_PTR(-EOPNOTSUPP);
1991 
1992 	fmr = pd->device->alloc_fmr(pd, mr_access_flags, fmr_attr);
1993 	if (!IS_ERR(fmr)) {
1994 		fmr->device = pd->device;
1995 		fmr->pd     = pd;
1996 		atomic_inc(&pd->usecnt);
1997 	}
1998 
1999 	return fmr;
2000 }
2001 EXPORT_SYMBOL(ib_alloc_fmr);
2002 
2003 int ib_unmap_fmr(struct list_head *fmr_list)
2004 {
2005 	struct ib_fmr *fmr;
2006 
2007 	if (list_empty(fmr_list))
2008 		return 0;
2009 
2010 	fmr = list_entry(fmr_list->next, struct ib_fmr, list);
2011 	return fmr->device->unmap_fmr(fmr_list);
2012 }
2013 EXPORT_SYMBOL(ib_unmap_fmr);
2014 
2015 int ib_dealloc_fmr(struct ib_fmr *fmr)
2016 {
2017 	struct ib_pd *pd;
2018 	int ret;
2019 
2020 	pd = fmr->pd;
2021 	ret = fmr->device->dealloc_fmr(fmr);
2022 	if (!ret)
2023 		atomic_dec(&pd->usecnt);
2024 
2025 	return ret;
2026 }
2027 EXPORT_SYMBOL(ib_dealloc_fmr);
2028 
2029 /* Multicast groups */
2030 
2031 static bool is_valid_mcast_lid(struct ib_qp *qp, u16 lid)
2032 {
2033 	struct ib_qp_init_attr init_attr = {};
2034 	struct ib_qp_attr attr = {};
2035 	int num_eth_ports = 0;
2036 	int port;
2037 
2038 	/* If QP state >= init, it is assigned to a port and we can check this
2039 	 * port only.
2040 	 */
2041 	if (!ib_query_qp(qp, &attr, IB_QP_STATE | IB_QP_PORT, &init_attr)) {
2042 		if (attr.qp_state >= IB_QPS_INIT) {
2043 			if (rdma_port_get_link_layer(qp->device, attr.port_num) !=
2044 			    IB_LINK_LAYER_INFINIBAND)
2045 				return true;
2046 			goto lid_check;
2047 		}
2048 	}
2049 
2050 	/* Can't get a quick answer, iterate over all ports */
2051 	for (port = 0; port < qp->device->phys_port_cnt; port++)
2052 		if (rdma_port_get_link_layer(qp->device, port) !=
2053 		    IB_LINK_LAYER_INFINIBAND)
2054 			num_eth_ports++;
2055 
2056 	/* If we have at lease one Ethernet port, RoCE annex declares that
2057 	 * multicast LID should be ignored. We can't tell at this step if the
2058 	 * QP belongs to an IB or Ethernet port.
2059 	 */
2060 	if (num_eth_ports)
2061 		return true;
2062 
2063 	/* If all the ports are IB, we can check according to IB spec. */
2064 lid_check:
2065 	return !(lid < be16_to_cpu(IB_MULTICAST_LID_BASE) ||
2066 		 lid == be16_to_cpu(IB_LID_PERMISSIVE));
2067 }
2068 
2069 int ib_attach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid)
2070 {
2071 	int ret;
2072 
2073 	if (!qp->device->attach_mcast)
2074 		return -EOPNOTSUPP;
2075 
2076 	if (!rdma_is_multicast_addr((struct in6_addr *)gid->raw) ||
2077 	    qp->qp_type != IB_QPT_UD || !is_valid_mcast_lid(qp, lid))
2078 		return -EINVAL;
2079 
2080 	ret = qp->device->attach_mcast(qp, gid, lid);
2081 	if (!ret)
2082 		atomic_inc(&qp->usecnt);
2083 	return ret;
2084 }
2085 EXPORT_SYMBOL(ib_attach_mcast);
2086 
2087 int ib_detach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid)
2088 {
2089 	int ret;
2090 
2091 	if (!qp->device->detach_mcast)
2092 		return -EOPNOTSUPP;
2093 
2094 	if (!rdma_is_multicast_addr((struct in6_addr *)gid->raw) ||
2095 	    qp->qp_type != IB_QPT_UD || !is_valid_mcast_lid(qp, lid))
2096 		return -EINVAL;
2097 
2098 	ret = qp->device->detach_mcast(qp, gid, lid);
2099 	if (!ret)
2100 		atomic_dec(&qp->usecnt);
2101 	return ret;
2102 }
2103 EXPORT_SYMBOL(ib_detach_mcast);
2104 
2105 struct ib_xrcd *__ib_alloc_xrcd(struct ib_device *device, const char *caller)
2106 {
2107 	struct ib_xrcd *xrcd;
2108 
2109 	if (!device->alloc_xrcd)
2110 		return ERR_PTR(-EOPNOTSUPP);
2111 
2112 	xrcd = device->alloc_xrcd(device, NULL, NULL);
2113 	if (!IS_ERR(xrcd)) {
2114 		xrcd->device = device;
2115 		xrcd->inode = NULL;
2116 		atomic_set(&xrcd->usecnt, 0);
2117 		mutex_init(&xrcd->tgt_qp_mutex);
2118 		INIT_LIST_HEAD(&xrcd->tgt_qp_list);
2119 	}
2120 
2121 	return xrcd;
2122 }
2123 EXPORT_SYMBOL(__ib_alloc_xrcd);
2124 
2125 int ib_dealloc_xrcd(struct ib_xrcd *xrcd)
2126 {
2127 	struct ib_qp *qp;
2128 	int ret;
2129 
2130 	if (atomic_read(&xrcd->usecnt))
2131 		return -EBUSY;
2132 
2133 	while (!list_empty(&xrcd->tgt_qp_list)) {
2134 		qp = list_entry(xrcd->tgt_qp_list.next, struct ib_qp, xrcd_list);
2135 		ret = ib_destroy_qp(qp);
2136 		if (ret)
2137 			return ret;
2138 	}
2139 
2140 	return xrcd->device->dealloc_xrcd(xrcd);
2141 }
2142 EXPORT_SYMBOL(ib_dealloc_xrcd);
2143 
2144 /**
2145  * ib_create_wq - Creates a WQ associated with the specified protection
2146  * domain.
2147  * @pd: The protection domain associated with the WQ.
2148  * @wq_attr: A list of initial attributes required to create the
2149  * WQ. If WQ creation succeeds, then the attributes are updated to
2150  * the actual capabilities of the created WQ.
2151  *
2152  * wq_attr->max_wr and wq_attr->max_sge determine
2153  * the requested size of the WQ, and set to the actual values allocated
2154  * on return.
2155  * If ib_create_wq() succeeds, then max_wr and max_sge will always be
2156  * at least as large as the requested values.
2157  */
2158 struct ib_wq *ib_create_wq(struct ib_pd *pd,
2159 			   struct ib_wq_init_attr *wq_attr)
2160 {
2161 	struct ib_wq *wq;
2162 
2163 	if (!pd->device->create_wq)
2164 		return ERR_PTR(-EOPNOTSUPP);
2165 
2166 	wq = pd->device->create_wq(pd, wq_attr, NULL);
2167 	if (!IS_ERR(wq)) {
2168 		wq->event_handler = wq_attr->event_handler;
2169 		wq->wq_context = wq_attr->wq_context;
2170 		wq->wq_type = wq_attr->wq_type;
2171 		wq->cq = wq_attr->cq;
2172 		wq->device = pd->device;
2173 		wq->pd = pd;
2174 		wq->uobject = NULL;
2175 		atomic_inc(&pd->usecnt);
2176 		atomic_inc(&wq_attr->cq->usecnt);
2177 		atomic_set(&wq->usecnt, 0);
2178 	}
2179 	return wq;
2180 }
2181 EXPORT_SYMBOL(ib_create_wq);
2182 
2183 /**
2184  * ib_destroy_wq - Destroys the specified WQ.
2185  * @wq: The WQ to destroy.
2186  */
2187 int ib_destroy_wq(struct ib_wq *wq)
2188 {
2189 	int err;
2190 	struct ib_cq *cq = wq->cq;
2191 	struct ib_pd *pd = wq->pd;
2192 
2193 	if (atomic_read(&wq->usecnt))
2194 		return -EBUSY;
2195 
2196 	err = wq->device->destroy_wq(wq);
2197 	if (!err) {
2198 		atomic_dec(&pd->usecnt);
2199 		atomic_dec(&cq->usecnt);
2200 	}
2201 	return err;
2202 }
2203 EXPORT_SYMBOL(ib_destroy_wq);
2204 
2205 /**
2206  * ib_modify_wq - Modifies the specified WQ.
2207  * @wq: The WQ to modify.
2208  * @wq_attr: On input, specifies the WQ attributes to modify.
2209  * @wq_attr_mask: A bit-mask used to specify which attributes of the WQ
2210  *   are being modified.
2211  * On output, the current values of selected WQ attributes are returned.
2212  */
2213 int ib_modify_wq(struct ib_wq *wq, struct ib_wq_attr *wq_attr,
2214 		 u32 wq_attr_mask)
2215 {
2216 	int err;
2217 
2218 	if (!wq->device->modify_wq)
2219 		return -EOPNOTSUPP;
2220 
2221 	err = wq->device->modify_wq(wq, wq_attr, wq_attr_mask, NULL);
2222 	return err;
2223 }
2224 EXPORT_SYMBOL(ib_modify_wq);
2225 
2226 /*
2227  * ib_create_rwq_ind_table - Creates a RQ Indirection Table.
2228  * @device: The device on which to create the rwq indirection table.
2229  * @ib_rwq_ind_table_init_attr: A list of initial attributes required to
2230  * create the Indirection Table.
2231  *
2232  * Note: The life time of ib_rwq_ind_table_init_attr->ind_tbl is not less
2233  *	than the created ib_rwq_ind_table object and the caller is responsible
2234  *	for its memory allocation/free.
2235  */
2236 struct ib_rwq_ind_table *ib_create_rwq_ind_table(struct ib_device *device,
2237 						 struct ib_rwq_ind_table_init_attr *init_attr)
2238 {
2239 	struct ib_rwq_ind_table *rwq_ind_table;
2240 	int i;
2241 	u32 table_size;
2242 
2243 	if (!device->create_rwq_ind_table)
2244 		return ERR_PTR(-EOPNOTSUPP);
2245 
2246 	table_size = (1 << init_attr->log_ind_tbl_size);
2247 	rwq_ind_table = device->create_rwq_ind_table(device,
2248 				init_attr, NULL);
2249 	if (IS_ERR(rwq_ind_table))
2250 		return rwq_ind_table;
2251 
2252 	rwq_ind_table->ind_tbl = init_attr->ind_tbl;
2253 	rwq_ind_table->log_ind_tbl_size = init_attr->log_ind_tbl_size;
2254 	rwq_ind_table->device = device;
2255 	rwq_ind_table->uobject = NULL;
2256 	atomic_set(&rwq_ind_table->usecnt, 0);
2257 
2258 	for (i = 0; i < table_size; i++)
2259 		atomic_inc(&rwq_ind_table->ind_tbl[i]->usecnt);
2260 
2261 	return rwq_ind_table;
2262 }
2263 EXPORT_SYMBOL(ib_create_rwq_ind_table);
2264 
2265 /*
2266  * ib_destroy_rwq_ind_table - Destroys the specified Indirection Table.
2267  * @wq_ind_table: The Indirection Table to destroy.
2268 */
2269 int ib_destroy_rwq_ind_table(struct ib_rwq_ind_table *rwq_ind_table)
2270 {
2271 	int err, i;
2272 	u32 table_size = (1 << rwq_ind_table->log_ind_tbl_size);
2273 	struct ib_wq **ind_tbl = rwq_ind_table->ind_tbl;
2274 
2275 	if (atomic_read(&rwq_ind_table->usecnt))
2276 		return -EBUSY;
2277 
2278 	err = rwq_ind_table->device->destroy_rwq_ind_table(rwq_ind_table);
2279 	if (!err) {
2280 		for (i = 0; i < table_size; i++)
2281 			atomic_dec(&ind_tbl[i]->usecnt);
2282 	}
2283 
2284 	return err;
2285 }
2286 EXPORT_SYMBOL(ib_destroy_rwq_ind_table);
2287 
2288 int ib_check_mr_status(struct ib_mr *mr, u32 check_mask,
2289 		       struct ib_mr_status *mr_status)
2290 {
2291 	return mr->device->check_mr_status ?
2292 		mr->device->check_mr_status(mr, check_mask, mr_status) : -EOPNOTSUPP;
2293 }
2294 EXPORT_SYMBOL(ib_check_mr_status);
2295 
2296 int ib_set_vf_link_state(struct ib_device *device, int vf, u8 port,
2297 			 int state)
2298 {
2299 	if (!device->set_vf_link_state)
2300 		return -EOPNOTSUPP;
2301 
2302 	return device->set_vf_link_state(device, vf, port, state);
2303 }
2304 EXPORT_SYMBOL(ib_set_vf_link_state);
2305 
2306 int ib_get_vf_config(struct ib_device *device, int vf, u8 port,
2307 		     struct ifla_vf_info *info)
2308 {
2309 	if (!device->get_vf_config)
2310 		return -EOPNOTSUPP;
2311 
2312 	return device->get_vf_config(device, vf, port, info);
2313 }
2314 EXPORT_SYMBOL(ib_get_vf_config);
2315 
2316 int ib_get_vf_stats(struct ib_device *device, int vf, u8 port,
2317 		    struct ifla_vf_stats *stats)
2318 {
2319 	if (!device->get_vf_stats)
2320 		return -EOPNOTSUPP;
2321 
2322 	return device->get_vf_stats(device, vf, port, stats);
2323 }
2324 EXPORT_SYMBOL(ib_get_vf_stats);
2325 
2326 int ib_set_vf_guid(struct ib_device *device, int vf, u8 port, u64 guid,
2327 		   int type)
2328 {
2329 	if (!device->set_vf_guid)
2330 		return -EOPNOTSUPP;
2331 
2332 	return device->set_vf_guid(device, vf, port, guid, type);
2333 }
2334 EXPORT_SYMBOL(ib_set_vf_guid);
2335 
2336 /**
2337  * ib_map_mr_sg() - Map the largest prefix of a dma mapped SG list
2338  *     and set it the memory region.
2339  * @mr:            memory region
2340  * @sg:            dma mapped scatterlist
2341  * @sg_nents:      number of entries in sg
2342  * @sg_offset:     offset in bytes into sg
2343  * @page_size:     page vector desired page size
2344  *
2345  * Constraints:
2346  * - The first sg element is allowed to have an offset.
2347  * - Each sg element must either be aligned to page_size or virtually
2348  *   contiguous to the previous element. In case an sg element has a
2349  *   non-contiguous offset, the mapping prefix will not include it.
2350  * - The last sg element is allowed to have length less than page_size.
2351  * - If sg_nents total byte length exceeds the mr max_num_sge * page_size
2352  *   then only max_num_sg entries will be mapped.
2353  * - If the MR was allocated with type IB_MR_TYPE_SG_GAPS, none of these
2354  *   constraints holds and the page_size argument is ignored.
2355  *
2356  * Returns the number of sg elements that were mapped to the memory region.
2357  *
2358  * After this completes successfully, the  memory region
2359  * is ready for registration.
2360  */
2361 int ib_map_mr_sg(struct ib_mr *mr, struct scatterlist *sg, int sg_nents,
2362 		 unsigned int *sg_offset, unsigned int page_size)
2363 {
2364 	if (unlikely(!mr->device->map_mr_sg))
2365 		return -EOPNOTSUPP;
2366 
2367 	mr->page_size = page_size;
2368 
2369 	return mr->device->map_mr_sg(mr, sg, sg_nents, sg_offset);
2370 }
2371 EXPORT_SYMBOL(ib_map_mr_sg);
2372 
2373 /**
2374  * ib_sg_to_pages() - Convert the largest prefix of a sg list
2375  *     to a page vector
2376  * @mr:            memory region
2377  * @sgl:           dma mapped scatterlist
2378  * @sg_nents:      number of entries in sg
2379  * @sg_offset_p:   IN:  start offset in bytes into sg
2380  *                 OUT: offset in bytes for element n of the sg of the first
2381  *                      byte that has not been processed where n is the return
2382  *                      value of this function.
2383  * @set_page:      driver page assignment function pointer
2384  *
2385  * Core service helper for drivers to convert the largest
2386  * prefix of given sg list to a page vector. The sg list
2387  * prefix converted is the prefix that meet the requirements
2388  * of ib_map_mr_sg.
2389  *
2390  * Returns the number of sg elements that were assigned to
2391  * a page vector.
2392  */
2393 int ib_sg_to_pages(struct ib_mr *mr, struct scatterlist *sgl, int sg_nents,
2394 		unsigned int *sg_offset_p, int (*set_page)(struct ib_mr *, u64))
2395 {
2396 	struct scatterlist *sg;
2397 	u64 last_end_dma_addr = 0;
2398 	unsigned int sg_offset = sg_offset_p ? *sg_offset_p : 0;
2399 	unsigned int last_page_off = 0;
2400 	u64 page_mask = ~((u64)mr->page_size - 1);
2401 	int i, ret;
2402 
2403 	if (unlikely(sg_nents <= 0 || sg_offset > sg_dma_len(&sgl[0])))
2404 		return -EINVAL;
2405 
2406 	mr->iova = sg_dma_address(&sgl[0]) + sg_offset;
2407 	mr->length = 0;
2408 
2409 	for_each_sg(sgl, sg, sg_nents, i) {
2410 		u64 dma_addr = sg_dma_address(sg) + sg_offset;
2411 		u64 prev_addr = dma_addr;
2412 		unsigned int dma_len = sg_dma_len(sg) - sg_offset;
2413 		u64 end_dma_addr = dma_addr + dma_len;
2414 		u64 page_addr = dma_addr & page_mask;
2415 
2416 		/*
2417 		 * For the second and later elements, check whether either the
2418 		 * end of element i-1 or the start of element i is not aligned
2419 		 * on a page boundary.
2420 		 */
2421 		if (i && (last_page_off != 0 || page_addr != dma_addr)) {
2422 			/* Stop mapping if there is a gap. */
2423 			if (last_end_dma_addr != dma_addr)
2424 				break;
2425 
2426 			/*
2427 			 * Coalesce this element with the last. If it is small
2428 			 * enough just update mr->length. Otherwise start
2429 			 * mapping from the next page.
2430 			 */
2431 			goto next_page;
2432 		}
2433 
2434 		do {
2435 			ret = set_page(mr, page_addr);
2436 			if (unlikely(ret < 0)) {
2437 				sg_offset = prev_addr - sg_dma_address(sg);
2438 				mr->length += prev_addr - dma_addr;
2439 				if (sg_offset_p)
2440 					*sg_offset_p = sg_offset;
2441 				return i || sg_offset ? i : ret;
2442 			}
2443 			prev_addr = page_addr;
2444 next_page:
2445 			page_addr += mr->page_size;
2446 		} while (page_addr < end_dma_addr);
2447 
2448 		mr->length += dma_len;
2449 		last_end_dma_addr = end_dma_addr;
2450 		last_page_off = end_dma_addr & ~page_mask;
2451 
2452 		sg_offset = 0;
2453 	}
2454 
2455 	if (sg_offset_p)
2456 		*sg_offset_p = 0;
2457 	return i;
2458 }
2459 EXPORT_SYMBOL(ib_sg_to_pages);
2460 
2461 struct ib_drain_cqe {
2462 	struct ib_cqe cqe;
2463 	struct completion done;
2464 };
2465 
2466 static void ib_drain_qp_done(struct ib_cq *cq, struct ib_wc *wc)
2467 {
2468 	struct ib_drain_cqe *cqe = container_of(wc->wr_cqe, struct ib_drain_cqe,
2469 						cqe);
2470 
2471 	complete(&cqe->done);
2472 }
2473 
2474 /*
2475  * Post a WR and block until its completion is reaped for the SQ.
2476  */
2477 static void __ib_drain_sq(struct ib_qp *qp)
2478 {
2479 	struct ib_cq *cq = qp->send_cq;
2480 	struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
2481 	struct ib_drain_cqe sdrain;
2482 	struct ib_rdma_wr swr = {
2483 		.wr = {
2484 			.next = NULL,
2485 			{ .wr_cqe	= &sdrain.cqe, },
2486 			.opcode	= IB_WR_RDMA_WRITE,
2487 		},
2488 	};
2489 	int ret;
2490 
2491 	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
2492 	if (ret) {
2493 		WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
2494 		return;
2495 	}
2496 
2497 	sdrain.cqe.done = ib_drain_qp_done;
2498 	init_completion(&sdrain.done);
2499 
2500 	ret = ib_post_send(qp, &swr.wr, NULL);
2501 	if (ret) {
2502 		WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
2503 		return;
2504 	}
2505 
2506 	if (cq->poll_ctx == IB_POLL_DIRECT)
2507 		while (wait_for_completion_timeout(&sdrain.done, HZ / 10) <= 0)
2508 			ib_process_cq_direct(cq, -1);
2509 	else
2510 		wait_for_completion(&sdrain.done);
2511 }
2512 
2513 /*
2514  * Post a WR and block until its completion is reaped for the RQ.
2515  */
2516 static void __ib_drain_rq(struct ib_qp *qp)
2517 {
2518 	struct ib_cq *cq = qp->recv_cq;
2519 	struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
2520 	struct ib_drain_cqe rdrain;
2521 	struct ib_recv_wr rwr = {};
2522 	int ret;
2523 
2524 	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
2525 	if (ret) {
2526 		WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
2527 		return;
2528 	}
2529 
2530 	rwr.wr_cqe = &rdrain.cqe;
2531 	rdrain.cqe.done = ib_drain_qp_done;
2532 	init_completion(&rdrain.done);
2533 
2534 	ret = ib_post_recv(qp, &rwr, NULL);
2535 	if (ret) {
2536 		WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
2537 		return;
2538 	}
2539 
2540 	if (cq->poll_ctx == IB_POLL_DIRECT)
2541 		while (wait_for_completion_timeout(&rdrain.done, HZ / 10) <= 0)
2542 			ib_process_cq_direct(cq, -1);
2543 	else
2544 		wait_for_completion(&rdrain.done);
2545 }
2546 
2547 /**
2548  * ib_drain_sq() - Block until all SQ CQEs have been consumed by the
2549  *		   application.
2550  * @qp:            queue pair to drain
2551  *
2552  * If the device has a provider-specific drain function, then
2553  * call that.  Otherwise call the generic drain function
2554  * __ib_drain_sq().
2555  *
2556  * The caller must:
2557  *
2558  * ensure there is room in the CQ and SQ for the drain work request and
2559  * completion.
2560  *
2561  * allocate the CQ using ib_alloc_cq().
2562  *
2563  * ensure that there are no other contexts that are posting WRs concurrently.
2564  * Otherwise the drain is not guaranteed.
2565  */
2566 void ib_drain_sq(struct ib_qp *qp)
2567 {
2568 	if (qp->device->drain_sq)
2569 		qp->device->drain_sq(qp);
2570 	else
2571 		__ib_drain_sq(qp);
2572 }
2573 EXPORT_SYMBOL(ib_drain_sq);
2574 
2575 /**
2576  * ib_drain_rq() - Block until all RQ CQEs have been consumed by the
2577  *		   application.
2578  * @qp:            queue pair to drain
2579  *
2580  * If the device has a provider-specific drain function, then
2581  * call that.  Otherwise call the generic drain function
2582  * __ib_drain_rq().
2583  *
2584  * The caller must:
2585  *
2586  * ensure there is room in the CQ and RQ for the drain work request and
2587  * completion.
2588  *
2589  * allocate the CQ using ib_alloc_cq().
2590  *
2591  * ensure that there are no other contexts that are posting WRs concurrently.
2592  * Otherwise the drain is not guaranteed.
2593  */
2594 void ib_drain_rq(struct ib_qp *qp)
2595 {
2596 	if (qp->device->drain_rq)
2597 		qp->device->drain_rq(qp);
2598 	else
2599 		__ib_drain_rq(qp);
2600 }
2601 EXPORT_SYMBOL(ib_drain_rq);
2602 
2603 /**
2604  * ib_drain_qp() - Block until all CQEs have been consumed by the
2605  *		   application on both the RQ and SQ.
2606  * @qp:            queue pair to drain
2607  *
2608  * The caller must:
2609  *
2610  * ensure there is room in the CQ(s), SQ, and RQ for drain work requests
2611  * and completions.
2612  *
2613  * allocate the CQs using ib_alloc_cq().
2614  *
2615  * ensure that there are no other contexts that are posting WRs concurrently.
2616  * Otherwise the drain is not guaranteed.
2617  */
2618 void ib_drain_qp(struct ib_qp *qp)
2619 {
2620 	ib_drain_sq(qp);
2621 	if (!qp->srq)
2622 		ib_drain_rq(qp);
2623 }
2624 EXPORT_SYMBOL(ib_drain_qp);
2625 
2626 struct net_device *rdma_alloc_netdev(struct ib_device *device, u8 port_num,
2627 				     enum rdma_netdev_t type, const char *name,
2628 				     unsigned char name_assign_type,
2629 				     void (*setup)(struct net_device *))
2630 {
2631 	struct rdma_netdev_alloc_params params;
2632 	struct net_device *netdev;
2633 	int rc;
2634 
2635 	if (!device->rdma_netdev_get_params)
2636 		return ERR_PTR(-EOPNOTSUPP);
2637 
2638 	rc = device->rdma_netdev_get_params(device, port_num, type, &params);
2639 	if (rc)
2640 		return ERR_PTR(rc);
2641 
2642 	netdev = alloc_netdev_mqs(params.sizeof_priv, name, name_assign_type,
2643 				  setup, params.txqs, params.rxqs);
2644 	if (!netdev)
2645 		return ERR_PTR(-ENOMEM);
2646 
2647 	return netdev;
2648 }
2649 EXPORT_SYMBOL(rdma_alloc_netdev);
2650 
2651 int rdma_init_netdev(struct ib_device *device, u8 port_num,
2652 		     enum rdma_netdev_t type, const char *name,
2653 		     unsigned char name_assign_type,
2654 		     void (*setup)(struct net_device *),
2655 		     struct net_device *netdev)
2656 {
2657 	struct rdma_netdev_alloc_params params;
2658 	int rc;
2659 
2660 	if (!device->rdma_netdev_get_params)
2661 		return -EOPNOTSUPP;
2662 
2663 	rc = device->rdma_netdev_get_params(device, port_num, type, &params);
2664 	if (rc)
2665 		return rc;
2666 
2667 	return params.initialize_rdma_netdev(device, port_num,
2668 					     netdev, params.param);
2669 }
2670 EXPORT_SYMBOL(rdma_init_netdev);
2671