xref: /openbmc/linux/drivers/infiniband/core/verbs.c (revision 5edb7691)
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 #include <trace/events/rdma_core.h>
56 
57 #include <trace/events/rdma_core.h>
58 
59 static int ib_resolve_eth_dmac(struct ib_device *device,
60 			       struct rdma_ah_attr *ah_attr);
61 
62 static const char * const ib_events[] = {
63 	[IB_EVENT_CQ_ERR]		= "CQ error",
64 	[IB_EVENT_QP_FATAL]		= "QP fatal error",
65 	[IB_EVENT_QP_REQ_ERR]		= "QP request error",
66 	[IB_EVENT_QP_ACCESS_ERR]	= "QP access error",
67 	[IB_EVENT_COMM_EST]		= "communication established",
68 	[IB_EVENT_SQ_DRAINED]		= "send queue drained",
69 	[IB_EVENT_PATH_MIG]		= "path migration successful",
70 	[IB_EVENT_PATH_MIG_ERR]		= "path migration error",
71 	[IB_EVENT_DEVICE_FATAL]		= "device fatal error",
72 	[IB_EVENT_PORT_ACTIVE]		= "port active",
73 	[IB_EVENT_PORT_ERR]		= "port error",
74 	[IB_EVENT_LID_CHANGE]		= "LID change",
75 	[IB_EVENT_PKEY_CHANGE]		= "P_key change",
76 	[IB_EVENT_SM_CHANGE]		= "SM change",
77 	[IB_EVENT_SRQ_ERR]		= "SRQ error",
78 	[IB_EVENT_SRQ_LIMIT_REACHED]	= "SRQ limit reached",
79 	[IB_EVENT_QP_LAST_WQE_REACHED]	= "last WQE reached",
80 	[IB_EVENT_CLIENT_REREGISTER]	= "client reregister",
81 	[IB_EVENT_GID_CHANGE]		= "GID changed",
82 };
83 
84 const char *__attribute_const__ ib_event_msg(enum ib_event_type event)
85 {
86 	size_t index = event;
87 
88 	return (index < ARRAY_SIZE(ib_events) && ib_events[index]) ?
89 			ib_events[index] : "unrecognized event";
90 }
91 EXPORT_SYMBOL(ib_event_msg);
92 
93 static const char * const wc_statuses[] = {
94 	[IB_WC_SUCCESS]			= "success",
95 	[IB_WC_LOC_LEN_ERR]		= "local length error",
96 	[IB_WC_LOC_QP_OP_ERR]		= "local QP operation error",
97 	[IB_WC_LOC_EEC_OP_ERR]		= "local EE context operation error",
98 	[IB_WC_LOC_PROT_ERR]		= "local protection error",
99 	[IB_WC_WR_FLUSH_ERR]		= "WR flushed",
100 	[IB_WC_MW_BIND_ERR]		= "memory management operation error",
101 	[IB_WC_BAD_RESP_ERR]		= "bad response error",
102 	[IB_WC_LOC_ACCESS_ERR]		= "local access error",
103 	[IB_WC_REM_INV_REQ_ERR]		= "invalid request error",
104 	[IB_WC_REM_ACCESS_ERR]		= "remote access error",
105 	[IB_WC_REM_OP_ERR]		= "remote operation error",
106 	[IB_WC_RETRY_EXC_ERR]		= "transport retry counter exceeded",
107 	[IB_WC_RNR_RETRY_EXC_ERR]	= "RNR retry counter exceeded",
108 	[IB_WC_LOC_RDD_VIOL_ERR]	= "local RDD violation error",
109 	[IB_WC_REM_INV_RD_REQ_ERR]	= "remote invalid RD request",
110 	[IB_WC_REM_ABORT_ERR]		= "operation aborted",
111 	[IB_WC_INV_EECN_ERR]		= "invalid EE context number",
112 	[IB_WC_INV_EEC_STATE_ERR]	= "invalid EE context state",
113 	[IB_WC_FATAL_ERR]		= "fatal error",
114 	[IB_WC_RESP_TIMEOUT_ERR]	= "response timeout error",
115 	[IB_WC_GENERAL_ERR]		= "general error",
116 };
117 
118 const char *__attribute_const__ ib_wc_status_msg(enum ib_wc_status status)
119 {
120 	size_t index = status;
121 
122 	return (index < ARRAY_SIZE(wc_statuses) && wc_statuses[index]) ?
123 			wc_statuses[index] : "unrecognized status";
124 }
125 EXPORT_SYMBOL(ib_wc_status_msg);
126 
127 __attribute_const__ int ib_rate_to_mult(enum ib_rate rate)
128 {
129 	switch (rate) {
130 	case IB_RATE_2_5_GBPS: return   1;
131 	case IB_RATE_5_GBPS:   return   2;
132 	case IB_RATE_10_GBPS:  return   4;
133 	case IB_RATE_20_GBPS:  return   8;
134 	case IB_RATE_30_GBPS:  return  12;
135 	case IB_RATE_40_GBPS:  return  16;
136 	case IB_RATE_60_GBPS:  return  24;
137 	case IB_RATE_80_GBPS:  return  32;
138 	case IB_RATE_120_GBPS: return  48;
139 	case IB_RATE_14_GBPS:  return   6;
140 	case IB_RATE_56_GBPS:  return  22;
141 	case IB_RATE_112_GBPS: return  45;
142 	case IB_RATE_168_GBPS: return  67;
143 	case IB_RATE_25_GBPS:  return  10;
144 	case IB_RATE_100_GBPS: return  40;
145 	case IB_RATE_200_GBPS: return  80;
146 	case IB_RATE_300_GBPS: return 120;
147 	case IB_RATE_28_GBPS:  return  11;
148 	case IB_RATE_50_GBPS:  return  20;
149 	case IB_RATE_400_GBPS: return 160;
150 	case IB_RATE_600_GBPS: return 240;
151 	default:	       return  -1;
152 	}
153 }
154 EXPORT_SYMBOL(ib_rate_to_mult);
155 
156 __attribute_const__ enum ib_rate mult_to_ib_rate(int mult)
157 {
158 	switch (mult) {
159 	case 1:   return IB_RATE_2_5_GBPS;
160 	case 2:   return IB_RATE_5_GBPS;
161 	case 4:   return IB_RATE_10_GBPS;
162 	case 8:   return IB_RATE_20_GBPS;
163 	case 12:  return IB_RATE_30_GBPS;
164 	case 16:  return IB_RATE_40_GBPS;
165 	case 24:  return IB_RATE_60_GBPS;
166 	case 32:  return IB_RATE_80_GBPS;
167 	case 48:  return IB_RATE_120_GBPS;
168 	case 6:   return IB_RATE_14_GBPS;
169 	case 22:  return IB_RATE_56_GBPS;
170 	case 45:  return IB_RATE_112_GBPS;
171 	case 67:  return IB_RATE_168_GBPS;
172 	case 10:  return IB_RATE_25_GBPS;
173 	case 40:  return IB_RATE_100_GBPS;
174 	case 80:  return IB_RATE_200_GBPS;
175 	case 120: return IB_RATE_300_GBPS;
176 	case 11:  return IB_RATE_28_GBPS;
177 	case 20:  return IB_RATE_50_GBPS;
178 	case 160: return IB_RATE_400_GBPS;
179 	case 240: return IB_RATE_600_GBPS;
180 	default:  return IB_RATE_PORT_CURRENT;
181 	}
182 }
183 EXPORT_SYMBOL(mult_to_ib_rate);
184 
185 __attribute_const__ int ib_rate_to_mbps(enum ib_rate rate)
186 {
187 	switch (rate) {
188 	case IB_RATE_2_5_GBPS: return 2500;
189 	case IB_RATE_5_GBPS:   return 5000;
190 	case IB_RATE_10_GBPS:  return 10000;
191 	case IB_RATE_20_GBPS:  return 20000;
192 	case IB_RATE_30_GBPS:  return 30000;
193 	case IB_RATE_40_GBPS:  return 40000;
194 	case IB_RATE_60_GBPS:  return 60000;
195 	case IB_RATE_80_GBPS:  return 80000;
196 	case IB_RATE_120_GBPS: return 120000;
197 	case IB_RATE_14_GBPS:  return 14062;
198 	case IB_RATE_56_GBPS:  return 56250;
199 	case IB_RATE_112_GBPS: return 112500;
200 	case IB_RATE_168_GBPS: return 168750;
201 	case IB_RATE_25_GBPS:  return 25781;
202 	case IB_RATE_100_GBPS: return 103125;
203 	case IB_RATE_200_GBPS: return 206250;
204 	case IB_RATE_300_GBPS: return 309375;
205 	case IB_RATE_28_GBPS:  return 28125;
206 	case IB_RATE_50_GBPS:  return 53125;
207 	case IB_RATE_400_GBPS: return 425000;
208 	case IB_RATE_600_GBPS: return 637500;
209 	default:	       return -1;
210 	}
211 }
212 EXPORT_SYMBOL(ib_rate_to_mbps);
213 
214 __attribute_const__ enum rdma_transport_type
215 rdma_node_get_transport(unsigned int node_type)
216 {
217 
218 	if (node_type == RDMA_NODE_USNIC)
219 		return RDMA_TRANSPORT_USNIC;
220 	if (node_type == RDMA_NODE_USNIC_UDP)
221 		return RDMA_TRANSPORT_USNIC_UDP;
222 	if (node_type == RDMA_NODE_RNIC)
223 		return RDMA_TRANSPORT_IWARP;
224 	if (node_type == RDMA_NODE_UNSPECIFIED)
225 		return RDMA_TRANSPORT_UNSPECIFIED;
226 
227 	return RDMA_TRANSPORT_IB;
228 }
229 EXPORT_SYMBOL(rdma_node_get_transport);
230 
231 enum rdma_link_layer rdma_port_get_link_layer(struct ib_device *device, u8 port_num)
232 {
233 	enum rdma_transport_type lt;
234 	if (device->ops.get_link_layer)
235 		return device->ops.get_link_layer(device, port_num);
236 
237 	lt = rdma_node_get_transport(device->node_type);
238 	if (lt == RDMA_TRANSPORT_IB)
239 		return IB_LINK_LAYER_INFINIBAND;
240 
241 	return IB_LINK_LAYER_ETHERNET;
242 }
243 EXPORT_SYMBOL(rdma_port_get_link_layer);
244 
245 /* Protection domains */
246 
247 /**
248  * ib_alloc_pd - Allocates an unused protection domain.
249  * @device: The device on which to allocate the protection domain.
250  * @flags: protection domain flags
251  * @caller: caller's build-time module name
252  *
253  * A protection domain object provides an association between QPs, shared
254  * receive queues, address handles, memory regions, and memory windows.
255  *
256  * Every PD has a local_dma_lkey which can be used as the lkey value for local
257  * memory operations.
258  */
259 struct ib_pd *__ib_alloc_pd(struct ib_device *device, unsigned int flags,
260 		const char *caller)
261 {
262 	struct ib_pd *pd;
263 	int mr_access_flags = 0;
264 	int ret;
265 
266 	pd = rdma_zalloc_drv_obj(device, ib_pd);
267 	if (!pd)
268 		return ERR_PTR(-ENOMEM);
269 
270 	pd->device = device;
271 	pd->uobject = NULL;
272 	pd->__internal_mr = NULL;
273 	atomic_set(&pd->usecnt, 0);
274 	pd->flags = flags;
275 
276 	pd->res.type = RDMA_RESTRACK_PD;
277 	rdma_restrack_set_task(&pd->res, caller);
278 
279 	ret = device->ops.alloc_pd(pd, NULL);
280 	if (ret) {
281 		kfree(pd);
282 		return ERR_PTR(ret);
283 	}
284 	rdma_restrack_kadd(&pd->res);
285 
286 	if (device->attrs.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY)
287 		pd->local_dma_lkey = device->local_dma_lkey;
288 	else
289 		mr_access_flags |= IB_ACCESS_LOCAL_WRITE;
290 
291 	if (flags & IB_PD_UNSAFE_GLOBAL_RKEY) {
292 		pr_warn("%s: enabling unsafe global rkey\n", caller);
293 		mr_access_flags |= IB_ACCESS_REMOTE_READ | IB_ACCESS_REMOTE_WRITE;
294 	}
295 
296 	if (mr_access_flags) {
297 		struct ib_mr *mr;
298 
299 		mr = pd->device->ops.get_dma_mr(pd, mr_access_flags);
300 		if (IS_ERR(mr)) {
301 			ib_dealloc_pd(pd);
302 			return ERR_CAST(mr);
303 		}
304 
305 		mr->device	= pd->device;
306 		mr->pd		= pd;
307 		mr->type        = IB_MR_TYPE_DMA;
308 		mr->uobject	= NULL;
309 		mr->need_inval	= false;
310 
311 		pd->__internal_mr = mr;
312 
313 		if (!(device->attrs.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY))
314 			pd->local_dma_lkey = pd->__internal_mr->lkey;
315 
316 		if (flags & IB_PD_UNSAFE_GLOBAL_RKEY)
317 			pd->unsafe_global_rkey = pd->__internal_mr->rkey;
318 	}
319 
320 	return pd;
321 }
322 EXPORT_SYMBOL(__ib_alloc_pd);
323 
324 /**
325  * ib_dealloc_pd_user - Deallocates a protection domain.
326  * @pd: The protection domain to deallocate.
327  * @udata: Valid user data or NULL for kernel object
328  *
329  * It is an error to call this function while any resources in the pd still
330  * exist.  The caller is responsible to synchronously destroy them and
331  * guarantee no new allocations will happen.
332  */
333 void ib_dealloc_pd_user(struct ib_pd *pd, struct ib_udata *udata)
334 {
335 	int ret;
336 
337 	if (pd->__internal_mr) {
338 		ret = pd->device->ops.dereg_mr(pd->__internal_mr, NULL);
339 		WARN_ON(ret);
340 		pd->__internal_mr = NULL;
341 	}
342 
343 	/* uverbs manipulates usecnt with proper locking, while the kabi
344 	   requires the caller to guarantee we can't race here. */
345 	WARN_ON(atomic_read(&pd->usecnt));
346 
347 	rdma_restrack_del(&pd->res);
348 	pd->device->ops.dealloc_pd(pd, udata);
349 	kfree(pd);
350 }
351 EXPORT_SYMBOL(ib_dealloc_pd_user);
352 
353 /* Address handles */
354 
355 /**
356  * rdma_copy_ah_attr - Copy rdma ah attribute from source to destination.
357  * @dest:       Pointer to destination ah_attr. Contents of the destination
358  *              pointer is assumed to be invalid and attribute are overwritten.
359  * @src:        Pointer to source ah_attr.
360  */
361 void rdma_copy_ah_attr(struct rdma_ah_attr *dest,
362 		       const struct rdma_ah_attr *src)
363 {
364 	*dest = *src;
365 	if (dest->grh.sgid_attr)
366 		rdma_hold_gid_attr(dest->grh.sgid_attr);
367 }
368 EXPORT_SYMBOL(rdma_copy_ah_attr);
369 
370 /**
371  * rdma_replace_ah_attr - Replace valid ah_attr with new new one.
372  * @old:        Pointer to existing ah_attr which needs to be replaced.
373  *              old is assumed to be valid or zero'd
374  * @new:        Pointer to the new ah_attr.
375  *
376  * rdma_replace_ah_attr() first releases any reference in the old ah_attr if
377  * old the ah_attr is valid; after that it copies the new attribute and holds
378  * the reference to the replaced ah_attr.
379  */
380 void rdma_replace_ah_attr(struct rdma_ah_attr *old,
381 			  const struct rdma_ah_attr *new)
382 {
383 	rdma_destroy_ah_attr(old);
384 	*old = *new;
385 	if (old->grh.sgid_attr)
386 		rdma_hold_gid_attr(old->grh.sgid_attr);
387 }
388 EXPORT_SYMBOL(rdma_replace_ah_attr);
389 
390 /**
391  * rdma_move_ah_attr - Move ah_attr pointed by source to destination.
392  * @dest:       Pointer to destination ah_attr to copy to.
393  *              dest is assumed to be valid or zero'd
394  * @src:        Pointer to the new ah_attr.
395  *
396  * rdma_move_ah_attr() first releases any reference in the destination ah_attr
397  * if it is valid. This also transfers ownership of internal references from
398  * src to dest, making src invalid in the process. No new reference of the src
399  * ah_attr is taken.
400  */
401 void rdma_move_ah_attr(struct rdma_ah_attr *dest, struct rdma_ah_attr *src)
402 {
403 	rdma_destroy_ah_attr(dest);
404 	*dest = *src;
405 	src->grh.sgid_attr = NULL;
406 }
407 EXPORT_SYMBOL(rdma_move_ah_attr);
408 
409 /*
410  * Validate that the rdma_ah_attr is valid for the device before passing it
411  * off to the driver.
412  */
413 static int rdma_check_ah_attr(struct ib_device *device,
414 			      struct rdma_ah_attr *ah_attr)
415 {
416 	if (!rdma_is_port_valid(device, ah_attr->port_num))
417 		return -EINVAL;
418 
419 	if ((rdma_is_grh_required(device, ah_attr->port_num) ||
420 	     ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE) &&
421 	    !(ah_attr->ah_flags & IB_AH_GRH))
422 		return -EINVAL;
423 
424 	if (ah_attr->grh.sgid_attr) {
425 		/*
426 		 * Make sure the passed sgid_attr is consistent with the
427 		 * parameters
428 		 */
429 		if (ah_attr->grh.sgid_attr->index != ah_attr->grh.sgid_index ||
430 		    ah_attr->grh.sgid_attr->port_num != ah_attr->port_num)
431 			return -EINVAL;
432 	}
433 	return 0;
434 }
435 
436 /*
437  * If the ah requires a GRH then ensure that sgid_attr pointer is filled in.
438  * On success the caller is responsible to call rdma_unfill_sgid_attr().
439  */
440 static int rdma_fill_sgid_attr(struct ib_device *device,
441 			       struct rdma_ah_attr *ah_attr,
442 			       const struct ib_gid_attr **old_sgid_attr)
443 {
444 	const struct ib_gid_attr *sgid_attr;
445 	struct ib_global_route *grh;
446 	int ret;
447 
448 	*old_sgid_attr = ah_attr->grh.sgid_attr;
449 
450 	ret = rdma_check_ah_attr(device, ah_attr);
451 	if (ret)
452 		return ret;
453 
454 	if (!(ah_attr->ah_flags & IB_AH_GRH))
455 		return 0;
456 
457 	grh = rdma_ah_retrieve_grh(ah_attr);
458 	if (grh->sgid_attr)
459 		return 0;
460 
461 	sgid_attr =
462 		rdma_get_gid_attr(device, ah_attr->port_num, grh->sgid_index);
463 	if (IS_ERR(sgid_attr))
464 		return PTR_ERR(sgid_attr);
465 
466 	/* Move ownerhip of the kref into the ah_attr */
467 	grh->sgid_attr = sgid_attr;
468 	return 0;
469 }
470 
471 static void rdma_unfill_sgid_attr(struct rdma_ah_attr *ah_attr,
472 				  const struct ib_gid_attr *old_sgid_attr)
473 {
474 	/*
475 	 * Fill didn't change anything, the caller retains ownership of
476 	 * whatever it passed
477 	 */
478 	if (ah_attr->grh.sgid_attr == old_sgid_attr)
479 		return;
480 
481 	/*
482 	 * Otherwise, we need to undo what rdma_fill_sgid_attr so the caller
483 	 * doesn't see any change in the rdma_ah_attr. If we get here
484 	 * old_sgid_attr is NULL.
485 	 */
486 	rdma_destroy_ah_attr(ah_attr);
487 }
488 
489 static const struct ib_gid_attr *
490 rdma_update_sgid_attr(struct rdma_ah_attr *ah_attr,
491 		      const struct ib_gid_attr *old_attr)
492 {
493 	if (old_attr)
494 		rdma_put_gid_attr(old_attr);
495 	if (ah_attr->ah_flags & IB_AH_GRH) {
496 		rdma_hold_gid_attr(ah_attr->grh.sgid_attr);
497 		return ah_attr->grh.sgid_attr;
498 	}
499 	return NULL;
500 }
501 
502 static struct ib_ah *_rdma_create_ah(struct ib_pd *pd,
503 				     struct rdma_ah_attr *ah_attr,
504 				     u32 flags,
505 				     struct ib_udata *udata)
506 {
507 	struct ib_device *device = pd->device;
508 	struct ib_ah *ah;
509 	int ret;
510 
511 	might_sleep_if(flags & RDMA_CREATE_AH_SLEEPABLE);
512 
513 	if (!device->ops.create_ah)
514 		return ERR_PTR(-EOPNOTSUPP);
515 
516 	ah = rdma_zalloc_drv_obj_gfp(
517 		device, ib_ah,
518 		(flags & RDMA_CREATE_AH_SLEEPABLE) ? GFP_KERNEL : GFP_ATOMIC);
519 	if (!ah)
520 		return ERR_PTR(-ENOMEM);
521 
522 	ah->device = device;
523 	ah->pd = pd;
524 	ah->type = ah_attr->type;
525 	ah->sgid_attr = rdma_update_sgid_attr(ah_attr, NULL);
526 
527 	ret = device->ops.create_ah(ah, ah_attr, flags, udata);
528 	if (ret) {
529 		kfree(ah);
530 		return ERR_PTR(ret);
531 	}
532 
533 	atomic_inc(&pd->usecnt);
534 	return ah;
535 }
536 
537 /**
538  * rdma_create_ah - Creates an address handle for the
539  * given address vector.
540  * @pd: The protection domain associated with the address handle.
541  * @ah_attr: The attributes of the address vector.
542  * @flags: Create address handle flags (see enum rdma_create_ah_flags).
543  *
544  * It returns 0 on success and returns appropriate error code on error.
545  * The address handle is used to reference a local or global destination
546  * in all UD QP post sends.
547  */
548 struct ib_ah *rdma_create_ah(struct ib_pd *pd, struct rdma_ah_attr *ah_attr,
549 			     u32 flags)
550 {
551 	const struct ib_gid_attr *old_sgid_attr;
552 	struct ib_ah *ah;
553 	int ret;
554 
555 	ret = rdma_fill_sgid_attr(pd->device, ah_attr, &old_sgid_attr);
556 	if (ret)
557 		return ERR_PTR(ret);
558 
559 	ah = _rdma_create_ah(pd, ah_attr, flags, NULL);
560 
561 	rdma_unfill_sgid_attr(ah_attr, old_sgid_attr);
562 	return ah;
563 }
564 EXPORT_SYMBOL(rdma_create_ah);
565 
566 /**
567  * rdma_create_user_ah - Creates an address handle for the
568  * given address vector.
569  * It resolves destination mac address for ah attribute of RoCE type.
570  * @pd: The protection domain associated with the address handle.
571  * @ah_attr: The attributes of the address vector.
572  * @udata: pointer to user's input output buffer information need by
573  *         provider driver.
574  *
575  * It returns 0 on success and returns appropriate error code on error.
576  * The address handle is used to reference a local or global destination
577  * in all UD QP post sends.
578  */
579 struct ib_ah *rdma_create_user_ah(struct ib_pd *pd,
580 				  struct rdma_ah_attr *ah_attr,
581 				  struct ib_udata *udata)
582 {
583 	const struct ib_gid_attr *old_sgid_attr;
584 	struct ib_ah *ah;
585 	int err;
586 
587 	err = rdma_fill_sgid_attr(pd->device, ah_attr, &old_sgid_attr);
588 	if (err)
589 		return ERR_PTR(err);
590 
591 	if (ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE) {
592 		err = ib_resolve_eth_dmac(pd->device, ah_attr);
593 		if (err) {
594 			ah = ERR_PTR(err);
595 			goto out;
596 		}
597 	}
598 
599 	ah = _rdma_create_ah(pd, ah_attr, RDMA_CREATE_AH_SLEEPABLE, udata);
600 
601 out:
602 	rdma_unfill_sgid_attr(ah_attr, old_sgid_attr);
603 	return ah;
604 }
605 EXPORT_SYMBOL(rdma_create_user_ah);
606 
607 int ib_get_rdma_header_version(const union rdma_network_hdr *hdr)
608 {
609 	const struct iphdr *ip4h = (struct iphdr *)&hdr->roce4grh;
610 	struct iphdr ip4h_checked;
611 	const struct ipv6hdr *ip6h = (struct ipv6hdr *)&hdr->ibgrh;
612 
613 	/* If it's IPv6, the version must be 6, otherwise, the first
614 	 * 20 bytes (before the IPv4 header) are garbled.
615 	 */
616 	if (ip6h->version != 6)
617 		return (ip4h->version == 4) ? 4 : 0;
618 	/* version may be 6 or 4 because the first 20 bytes could be garbled */
619 
620 	/* RoCE v2 requires no options, thus header length
621 	 * must be 5 words
622 	 */
623 	if (ip4h->ihl != 5)
624 		return 6;
625 
626 	/* Verify checksum.
627 	 * We can't write on scattered buffers so we need to copy to
628 	 * temp buffer.
629 	 */
630 	memcpy(&ip4h_checked, ip4h, sizeof(ip4h_checked));
631 	ip4h_checked.check = 0;
632 	ip4h_checked.check = ip_fast_csum((u8 *)&ip4h_checked, 5);
633 	/* if IPv4 header checksum is OK, believe it */
634 	if (ip4h->check == ip4h_checked.check)
635 		return 4;
636 	return 6;
637 }
638 EXPORT_SYMBOL(ib_get_rdma_header_version);
639 
640 static enum rdma_network_type ib_get_net_type_by_grh(struct ib_device *device,
641 						     u8 port_num,
642 						     const struct ib_grh *grh)
643 {
644 	int grh_version;
645 
646 	if (rdma_protocol_ib(device, port_num))
647 		return RDMA_NETWORK_IB;
648 
649 	grh_version = ib_get_rdma_header_version((union rdma_network_hdr *)grh);
650 
651 	if (grh_version == 4)
652 		return RDMA_NETWORK_IPV4;
653 
654 	if (grh->next_hdr == IPPROTO_UDP)
655 		return RDMA_NETWORK_IPV6;
656 
657 	return RDMA_NETWORK_ROCE_V1;
658 }
659 
660 struct find_gid_index_context {
661 	u16 vlan_id;
662 	enum ib_gid_type gid_type;
663 };
664 
665 static bool find_gid_index(const union ib_gid *gid,
666 			   const struct ib_gid_attr *gid_attr,
667 			   void *context)
668 {
669 	struct find_gid_index_context *ctx = context;
670 	u16 vlan_id = 0xffff;
671 	int ret;
672 
673 	if (ctx->gid_type != gid_attr->gid_type)
674 		return false;
675 
676 	ret = rdma_read_gid_l2_fields(gid_attr, &vlan_id, NULL);
677 	if (ret)
678 		return false;
679 
680 	return ctx->vlan_id == vlan_id;
681 }
682 
683 static const struct ib_gid_attr *
684 get_sgid_attr_from_eth(struct ib_device *device, u8 port_num,
685 		       u16 vlan_id, const union ib_gid *sgid,
686 		       enum ib_gid_type gid_type)
687 {
688 	struct find_gid_index_context context = {.vlan_id = vlan_id,
689 						 .gid_type = gid_type};
690 
691 	return rdma_find_gid_by_filter(device, sgid, port_num, find_gid_index,
692 				       &context);
693 }
694 
695 int ib_get_gids_from_rdma_hdr(const union rdma_network_hdr *hdr,
696 			      enum rdma_network_type net_type,
697 			      union ib_gid *sgid, union ib_gid *dgid)
698 {
699 	struct sockaddr_in  src_in;
700 	struct sockaddr_in  dst_in;
701 	__be32 src_saddr, dst_saddr;
702 
703 	if (!sgid || !dgid)
704 		return -EINVAL;
705 
706 	if (net_type == RDMA_NETWORK_IPV4) {
707 		memcpy(&src_in.sin_addr.s_addr,
708 		       &hdr->roce4grh.saddr, 4);
709 		memcpy(&dst_in.sin_addr.s_addr,
710 		       &hdr->roce4grh.daddr, 4);
711 		src_saddr = src_in.sin_addr.s_addr;
712 		dst_saddr = dst_in.sin_addr.s_addr;
713 		ipv6_addr_set_v4mapped(src_saddr,
714 				       (struct in6_addr *)sgid);
715 		ipv6_addr_set_v4mapped(dst_saddr,
716 				       (struct in6_addr *)dgid);
717 		return 0;
718 	} else if (net_type == RDMA_NETWORK_IPV6 ||
719 		   net_type == RDMA_NETWORK_IB) {
720 		*dgid = hdr->ibgrh.dgid;
721 		*sgid = hdr->ibgrh.sgid;
722 		return 0;
723 	} else {
724 		return -EINVAL;
725 	}
726 }
727 EXPORT_SYMBOL(ib_get_gids_from_rdma_hdr);
728 
729 /* Resolve destination mac address and hop limit for unicast destination
730  * GID entry, considering the source GID entry as well.
731  * ah_attribute must have have valid port_num, sgid_index.
732  */
733 static int ib_resolve_unicast_gid_dmac(struct ib_device *device,
734 				       struct rdma_ah_attr *ah_attr)
735 {
736 	struct ib_global_route *grh = rdma_ah_retrieve_grh(ah_attr);
737 	const struct ib_gid_attr *sgid_attr = grh->sgid_attr;
738 	int hop_limit = 0xff;
739 	int ret = 0;
740 
741 	/* If destination is link local and source GID is RoCEv1,
742 	 * IP stack is not used.
743 	 */
744 	if (rdma_link_local_addr((struct in6_addr *)grh->dgid.raw) &&
745 	    sgid_attr->gid_type == IB_GID_TYPE_ROCE) {
746 		rdma_get_ll_mac((struct in6_addr *)grh->dgid.raw,
747 				ah_attr->roce.dmac);
748 		return ret;
749 	}
750 
751 	ret = rdma_addr_find_l2_eth_by_grh(&sgid_attr->gid, &grh->dgid,
752 					   ah_attr->roce.dmac,
753 					   sgid_attr, &hop_limit);
754 
755 	grh->hop_limit = hop_limit;
756 	return ret;
757 }
758 
759 /*
760  * This function initializes address handle attributes from the incoming packet.
761  * Incoming packet has dgid of the receiver node on which this code is
762  * getting executed and, sgid contains the GID of the sender.
763  *
764  * When resolving mac address of destination, the arrived dgid is used
765  * as sgid and, sgid is used as dgid because sgid contains destinations
766  * GID whom to respond to.
767  *
768  * On success the caller is responsible to call rdma_destroy_ah_attr on the
769  * attr.
770  */
771 int ib_init_ah_attr_from_wc(struct ib_device *device, u8 port_num,
772 			    const struct ib_wc *wc, const struct ib_grh *grh,
773 			    struct rdma_ah_attr *ah_attr)
774 {
775 	u32 flow_class;
776 	int ret;
777 	enum rdma_network_type net_type = RDMA_NETWORK_IB;
778 	enum ib_gid_type gid_type = IB_GID_TYPE_IB;
779 	const struct ib_gid_attr *sgid_attr;
780 	int hoplimit = 0xff;
781 	union ib_gid dgid;
782 	union ib_gid sgid;
783 
784 	might_sleep();
785 
786 	memset(ah_attr, 0, sizeof *ah_attr);
787 	ah_attr->type = rdma_ah_find_type(device, port_num);
788 	if (rdma_cap_eth_ah(device, port_num)) {
789 		if (wc->wc_flags & IB_WC_WITH_NETWORK_HDR_TYPE)
790 			net_type = wc->network_hdr_type;
791 		else
792 			net_type = ib_get_net_type_by_grh(device, port_num, grh);
793 		gid_type = ib_network_to_gid_type(net_type);
794 	}
795 	ret = ib_get_gids_from_rdma_hdr((union rdma_network_hdr *)grh, net_type,
796 					&sgid, &dgid);
797 	if (ret)
798 		return ret;
799 
800 	rdma_ah_set_sl(ah_attr, wc->sl);
801 	rdma_ah_set_port_num(ah_attr, port_num);
802 
803 	if (rdma_protocol_roce(device, port_num)) {
804 		u16 vlan_id = wc->wc_flags & IB_WC_WITH_VLAN ?
805 				wc->vlan_id : 0xffff;
806 
807 		if (!(wc->wc_flags & IB_WC_GRH))
808 			return -EPROTOTYPE;
809 
810 		sgid_attr = get_sgid_attr_from_eth(device, port_num,
811 						   vlan_id, &dgid,
812 						   gid_type);
813 		if (IS_ERR(sgid_attr))
814 			return PTR_ERR(sgid_attr);
815 
816 		flow_class = be32_to_cpu(grh->version_tclass_flow);
817 		rdma_move_grh_sgid_attr(ah_attr,
818 					&sgid,
819 					flow_class & 0xFFFFF,
820 					hoplimit,
821 					(flow_class >> 20) & 0xFF,
822 					sgid_attr);
823 
824 		ret = ib_resolve_unicast_gid_dmac(device, ah_attr);
825 		if (ret)
826 			rdma_destroy_ah_attr(ah_attr);
827 
828 		return ret;
829 	} else {
830 		rdma_ah_set_dlid(ah_attr, wc->slid);
831 		rdma_ah_set_path_bits(ah_attr, wc->dlid_path_bits);
832 
833 		if ((wc->wc_flags & IB_WC_GRH) == 0)
834 			return 0;
835 
836 		if (dgid.global.interface_id !=
837 					cpu_to_be64(IB_SA_WELL_KNOWN_GUID)) {
838 			sgid_attr = rdma_find_gid_by_port(
839 				device, &dgid, IB_GID_TYPE_IB, port_num, NULL);
840 		} else
841 			sgid_attr = rdma_get_gid_attr(device, port_num, 0);
842 
843 		if (IS_ERR(sgid_attr))
844 			return PTR_ERR(sgid_attr);
845 		flow_class = be32_to_cpu(grh->version_tclass_flow);
846 		rdma_move_grh_sgid_attr(ah_attr,
847 					&sgid,
848 					flow_class & 0xFFFFF,
849 					hoplimit,
850 					(flow_class >> 20) & 0xFF,
851 					sgid_attr);
852 
853 		return 0;
854 	}
855 }
856 EXPORT_SYMBOL(ib_init_ah_attr_from_wc);
857 
858 /**
859  * rdma_move_grh_sgid_attr - Sets the sgid attribute of GRH, taking ownership
860  * of the reference
861  *
862  * @attr:	Pointer to AH attribute structure
863  * @dgid:	Destination GID
864  * @flow_label:	Flow label
865  * @hop_limit:	Hop limit
866  * @traffic_class: traffic class
867  * @sgid_attr:	Pointer to SGID attribute
868  *
869  * This takes ownership of the sgid_attr reference. The caller must ensure
870  * rdma_destroy_ah_attr() is called before destroying the rdma_ah_attr after
871  * calling this function.
872  */
873 void rdma_move_grh_sgid_attr(struct rdma_ah_attr *attr, union ib_gid *dgid,
874 			     u32 flow_label, u8 hop_limit, u8 traffic_class,
875 			     const struct ib_gid_attr *sgid_attr)
876 {
877 	rdma_ah_set_grh(attr, dgid, flow_label, sgid_attr->index, hop_limit,
878 			traffic_class);
879 	attr->grh.sgid_attr = sgid_attr;
880 }
881 EXPORT_SYMBOL(rdma_move_grh_sgid_attr);
882 
883 /**
884  * rdma_destroy_ah_attr - Release reference to SGID attribute of
885  * ah attribute.
886  * @ah_attr: Pointer to ah attribute
887  *
888  * Release reference to the SGID attribute of the ah attribute if it is
889  * non NULL. It is safe to call this multiple times, and safe to call it on
890  * a zero initialized ah_attr.
891  */
892 void rdma_destroy_ah_attr(struct rdma_ah_attr *ah_attr)
893 {
894 	if (ah_attr->grh.sgid_attr) {
895 		rdma_put_gid_attr(ah_attr->grh.sgid_attr);
896 		ah_attr->grh.sgid_attr = NULL;
897 	}
898 }
899 EXPORT_SYMBOL(rdma_destroy_ah_attr);
900 
901 struct ib_ah *ib_create_ah_from_wc(struct ib_pd *pd, const struct ib_wc *wc,
902 				   const struct ib_grh *grh, u8 port_num)
903 {
904 	struct rdma_ah_attr ah_attr;
905 	struct ib_ah *ah;
906 	int ret;
907 
908 	ret = ib_init_ah_attr_from_wc(pd->device, port_num, wc, grh, &ah_attr);
909 	if (ret)
910 		return ERR_PTR(ret);
911 
912 	ah = rdma_create_ah(pd, &ah_attr, RDMA_CREATE_AH_SLEEPABLE);
913 
914 	rdma_destroy_ah_attr(&ah_attr);
915 	return ah;
916 }
917 EXPORT_SYMBOL(ib_create_ah_from_wc);
918 
919 int rdma_modify_ah(struct ib_ah *ah, struct rdma_ah_attr *ah_attr)
920 {
921 	const struct ib_gid_attr *old_sgid_attr;
922 	int ret;
923 
924 	if (ah->type != ah_attr->type)
925 		return -EINVAL;
926 
927 	ret = rdma_fill_sgid_attr(ah->device, ah_attr, &old_sgid_attr);
928 	if (ret)
929 		return ret;
930 
931 	ret = ah->device->ops.modify_ah ?
932 		ah->device->ops.modify_ah(ah, ah_attr) :
933 		-EOPNOTSUPP;
934 
935 	ah->sgid_attr = rdma_update_sgid_attr(ah_attr, ah->sgid_attr);
936 	rdma_unfill_sgid_attr(ah_attr, old_sgid_attr);
937 	return ret;
938 }
939 EXPORT_SYMBOL(rdma_modify_ah);
940 
941 int rdma_query_ah(struct ib_ah *ah, struct rdma_ah_attr *ah_attr)
942 {
943 	ah_attr->grh.sgid_attr = NULL;
944 
945 	return ah->device->ops.query_ah ?
946 		ah->device->ops.query_ah(ah, ah_attr) :
947 		-EOPNOTSUPP;
948 }
949 EXPORT_SYMBOL(rdma_query_ah);
950 
951 int rdma_destroy_ah_user(struct ib_ah *ah, u32 flags, struct ib_udata *udata)
952 {
953 	const struct ib_gid_attr *sgid_attr = ah->sgid_attr;
954 	struct ib_pd *pd;
955 
956 	might_sleep_if(flags & RDMA_DESTROY_AH_SLEEPABLE);
957 
958 	pd = ah->pd;
959 
960 	ah->device->ops.destroy_ah(ah, flags);
961 	atomic_dec(&pd->usecnt);
962 	if (sgid_attr)
963 		rdma_put_gid_attr(sgid_attr);
964 
965 	kfree(ah);
966 	return 0;
967 }
968 EXPORT_SYMBOL(rdma_destroy_ah_user);
969 
970 /* Shared receive queues */
971 
972 struct ib_srq *ib_create_srq(struct ib_pd *pd,
973 			     struct ib_srq_init_attr *srq_init_attr)
974 {
975 	struct ib_srq *srq;
976 	int ret;
977 
978 	if (!pd->device->ops.create_srq)
979 		return ERR_PTR(-EOPNOTSUPP);
980 
981 	srq = rdma_zalloc_drv_obj(pd->device, ib_srq);
982 	if (!srq)
983 		return ERR_PTR(-ENOMEM);
984 
985 	srq->device = pd->device;
986 	srq->pd = pd;
987 	srq->event_handler = srq_init_attr->event_handler;
988 	srq->srq_context = srq_init_attr->srq_context;
989 	srq->srq_type = srq_init_attr->srq_type;
990 
991 	if (ib_srq_has_cq(srq->srq_type)) {
992 		srq->ext.cq = srq_init_attr->ext.cq;
993 		atomic_inc(&srq->ext.cq->usecnt);
994 	}
995 	if (srq->srq_type == IB_SRQT_XRC) {
996 		srq->ext.xrc.xrcd = srq_init_attr->ext.xrc.xrcd;
997 		atomic_inc(&srq->ext.xrc.xrcd->usecnt);
998 	}
999 	atomic_inc(&pd->usecnt);
1000 
1001 	ret = pd->device->ops.create_srq(srq, srq_init_attr, NULL);
1002 	if (ret) {
1003 		atomic_dec(&srq->pd->usecnt);
1004 		if (srq->srq_type == IB_SRQT_XRC)
1005 			atomic_dec(&srq->ext.xrc.xrcd->usecnt);
1006 		if (ib_srq_has_cq(srq->srq_type))
1007 			atomic_dec(&srq->ext.cq->usecnt);
1008 		kfree(srq);
1009 		return ERR_PTR(ret);
1010 	}
1011 
1012 	return srq;
1013 }
1014 EXPORT_SYMBOL(ib_create_srq);
1015 
1016 int ib_modify_srq(struct ib_srq *srq,
1017 		  struct ib_srq_attr *srq_attr,
1018 		  enum ib_srq_attr_mask srq_attr_mask)
1019 {
1020 	return srq->device->ops.modify_srq ?
1021 		srq->device->ops.modify_srq(srq, srq_attr, srq_attr_mask,
1022 					    NULL) : -EOPNOTSUPP;
1023 }
1024 EXPORT_SYMBOL(ib_modify_srq);
1025 
1026 int ib_query_srq(struct ib_srq *srq,
1027 		 struct ib_srq_attr *srq_attr)
1028 {
1029 	return srq->device->ops.query_srq ?
1030 		srq->device->ops.query_srq(srq, srq_attr) : -EOPNOTSUPP;
1031 }
1032 EXPORT_SYMBOL(ib_query_srq);
1033 
1034 int ib_destroy_srq_user(struct ib_srq *srq, struct ib_udata *udata)
1035 {
1036 	if (atomic_read(&srq->usecnt))
1037 		return -EBUSY;
1038 
1039 	srq->device->ops.destroy_srq(srq, udata);
1040 
1041 	atomic_dec(&srq->pd->usecnt);
1042 	if (srq->srq_type == IB_SRQT_XRC)
1043 		atomic_dec(&srq->ext.xrc.xrcd->usecnt);
1044 	if (ib_srq_has_cq(srq->srq_type))
1045 		atomic_dec(&srq->ext.cq->usecnt);
1046 	kfree(srq);
1047 
1048 	return 0;
1049 }
1050 EXPORT_SYMBOL(ib_destroy_srq_user);
1051 
1052 /* Queue pairs */
1053 
1054 static void __ib_shared_qp_event_handler(struct ib_event *event, void *context)
1055 {
1056 	struct ib_qp *qp = context;
1057 	unsigned long flags;
1058 
1059 	spin_lock_irqsave(&qp->device->qp_open_list_lock, flags);
1060 	list_for_each_entry(event->element.qp, &qp->open_list, open_list)
1061 		if (event->element.qp->event_handler)
1062 			event->element.qp->event_handler(event, event->element.qp->qp_context);
1063 	spin_unlock_irqrestore(&qp->device->qp_open_list_lock, flags);
1064 }
1065 
1066 static void __ib_insert_xrcd_qp(struct ib_xrcd *xrcd, struct ib_qp *qp)
1067 {
1068 	mutex_lock(&xrcd->tgt_qp_mutex);
1069 	list_add(&qp->xrcd_list, &xrcd->tgt_qp_list);
1070 	mutex_unlock(&xrcd->tgt_qp_mutex);
1071 }
1072 
1073 static struct ib_qp *__ib_open_qp(struct ib_qp *real_qp,
1074 				  void (*event_handler)(struct ib_event *, void *),
1075 				  void *qp_context)
1076 {
1077 	struct ib_qp *qp;
1078 	unsigned long flags;
1079 	int err;
1080 
1081 	qp = kzalloc(sizeof *qp, GFP_KERNEL);
1082 	if (!qp)
1083 		return ERR_PTR(-ENOMEM);
1084 
1085 	qp->real_qp = real_qp;
1086 	err = ib_open_shared_qp_security(qp, real_qp->device);
1087 	if (err) {
1088 		kfree(qp);
1089 		return ERR_PTR(err);
1090 	}
1091 
1092 	qp->real_qp = real_qp;
1093 	atomic_inc(&real_qp->usecnt);
1094 	qp->device = real_qp->device;
1095 	qp->event_handler = event_handler;
1096 	qp->qp_context = qp_context;
1097 	qp->qp_num = real_qp->qp_num;
1098 	qp->qp_type = real_qp->qp_type;
1099 
1100 	spin_lock_irqsave(&real_qp->device->qp_open_list_lock, flags);
1101 	list_add(&qp->open_list, &real_qp->open_list);
1102 	spin_unlock_irqrestore(&real_qp->device->qp_open_list_lock, flags);
1103 
1104 	return qp;
1105 }
1106 
1107 struct ib_qp *ib_open_qp(struct ib_xrcd *xrcd,
1108 			 struct ib_qp_open_attr *qp_open_attr)
1109 {
1110 	struct ib_qp *qp, *real_qp;
1111 
1112 	if (qp_open_attr->qp_type != IB_QPT_XRC_TGT)
1113 		return ERR_PTR(-EINVAL);
1114 
1115 	qp = ERR_PTR(-EINVAL);
1116 	mutex_lock(&xrcd->tgt_qp_mutex);
1117 	list_for_each_entry(real_qp, &xrcd->tgt_qp_list, xrcd_list) {
1118 		if (real_qp->qp_num == qp_open_attr->qp_num) {
1119 			qp = __ib_open_qp(real_qp, qp_open_attr->event_handler,
1120 					  qp_open_attr->qp_context);
1121 			break;
1122 		}
1123 	}
1124 	mutex_unlock(&xrcd->tgt_qp_mutex);
1125 	return qp;
1126 }
1127 EXPORT_SYMBOL(ib_open_qp);
1128 
1129 static struct ib_qp *create_xrc_qp_user(struct ib_qp *qp,
1130 					struct ib_qp_init_attr *qp_init_attr,
1131 					struct ib_udata *udata)
1132 {
1133 	struct ib_qp *real_qp = qp;
1134 
1135 	qp->event_handler = __ib_shared_qp_event_handler;
1136 	qp->qp_context = qp;
1137 	qp->pd = NULL;
1138 	qp->send_cq = qp->recv_cq = NULL;
1139 	qp->srq = NULL;
1140 	qp->xrcd = qp_init_attr->xrcd;
1141 	atomic_inc(&qp_init_attr->xrcd->usecnt);
1142 	INIT_LIST_HEAD(&qp->open_list);
1143 
1144 	qp = __ib_open_qp(real_qp, qp_init_attr->event_handler,
1145 			  qp_init_attr->qp_context);
1146 	if (IS_ERR(qp))
1147 		return qp;
1148 
1149 	__ib_insert_xrcd_qp(qp_init_attr->xrcd, real_qp);
1150 	return qp;
1151 }
1152 
1153 struct ib_qp *ib_create_qp_user(struct ib_pd *pd,
1154 				struct ib_qp_init_attr *qp_init_attr,
1155 				struct ib_udata *udata)
1156 {
1157 	struct ib_device *device = pd ? pd->device : qp_init_attr->xrcd->device;
1158 	struct ib_qp *qp;
1159 	int ret;
1160 
1161 	if (qp_init_attr->rwq_ind_tbl &&
1162 	    (qp_init_attr->recv_cq ||
1163 	    qp_init_attr->srq || qp_init_attr->cap.max_recv_wr ||
1164 	    qp_init_attr->cap.max_recv_sge))
1165 		return ERR_PTR(-EINVAL);
1166 
1167 	if ((qp_init_attr->create_flags & IB_QP_CREATE_INTEGRITY_EN) &&
1168 	    !(device->attrs.device_cap_flags & IB_DEVICE_INTEGRITY_HANDOVER))
1169 		return ERR_PTR(-EINVAL);
1170 
1171 	/*
1172 	 * If the callers is using the RDMA API calculate the resources
1173 	 * needed for the RDMA READ/WRITE operations.
1174 	 *
1175 	 * Note that these callers need to pass in a port number.
1176 	 */
1177 	if (qp_init_attr->cap.max_rdma_ctxs)
1178 		rdma_rw_init_qp(device, qp_init_attr);
1179 
1180 	qp = _ib_create_qp(device, pd, qp_init_attr, NULL, NULL);
1181 	if (IS_ERR(qp))
1182 		return qp;
1183 
1184 	ret = ib_create_qp_security(qp, device);
1185 	if (ret)
1186 		goto err;
1187 
1188 	if (qp_init_attr->qp_type == IB_QPT_XRC_TGT) {
1189 		struct ib_qp *xrc_qp =
1190 			create_xrc_qp_user(qp, qp_init_attr, udata);
1191 
1192 		if (IS_ERR(xrc_qp)) {
1193 			ret = PTR_ERR(xrc_qp);
1194 			goto err;
1195 		}
1196 		return xrc_qp;
1197 	}
1198 
1199 	qp->event_handler = qp_init_attr->event_handler;
1200 	qp->qp_context = qp_init_attr->qp_context;
1201 	if (qp_init_attr->qp_type == IB_QPT_XRC_INI) {
1202 		qp->recv_cq = NULL;
1203 		qp->srq = NULL;
1204 	} else {
1205 		qp->recv_cq = qp_init_attr->recv_cq;
1206 		if (qp_init_attr->recv_cq)
1207 			atomic_inc(&qp_init_attr->recv_cq->usecnt);
1208 		qp->srq = qp_init_attr->srq;
1209 		if (qp->srq)
1210 			atomic_inc(&qp_init_attr->srq->usecnt);
1211 	}
1212 
1213 	qp->send_cq = qp_init_attr->send_cq;
1214 	qp->xrcd    = NULL;
1215 
1216 	atomic_inc(&pd->usecnt);
1217 	if (qp_init_attr->send_cq)
1218 		atomic_inc(&qp_init_attr->send_cq->usecnt);
1219 	if (qp_init_attr->rwq_ind_tbl)
1220 		atomic_inc(&qp->rwq_ind_tbl->usecnt);
1221 
1222 	if (qp_init_attr->cap.max_rdma_ctxs) {
1223 		ret = rdma_rw_init_mrs(qp, qp_init_attr);
1224 		if (ret)
1225 			goto err;
1226 	}
1227 
1228 	/*
1229 	 * Note: all hw drivers guarantee that max_send_sge is lower than
1230 	 * the device RDMA WRITE SGE limit but not all hw drivers ensure that
1231 	 * max_send_sge <= max_sge_rd.
1232 	 */
1233 	qp->max_write_sge = qp_init_attr->cap.max_send_sge;
1234 	qp->max_read_sge = min_t(u32, qp_init_attr->cap.max_send_sge,
1235 				 device->attrs.max_sge_rd);
1236 	if (qp_init_attr->create_flags & IB_QP_CREATE_INTEGRITY_EN)
1237 		qp->integrity_en = true;
1238 
1239 	return qp;
1240 
1241 err:
1242 	ib_destroy_qp(qp);
1243 	return ERR_PTR(ret);
1244 
1245 }
1246 EXPORT_SYMBOL(ib_create_qp_user);
1247 
1248 static const struct {
1249 	int			valid;
1250 	enum ib_qp_attr_mask	req_param[IB_QPT_MAX];
1251 	enum ib_qp_attr_mask	opt_param[IB_QPT_MAX];
1252 } qp_state_table[IB_QPS_ERR + 1][IB_QPS_ERR + 1] = {
1253 	[IB_QPS_RESET] = {
1254 		[IB_QPS_RESET] = { .valid = 1 },
1255 		[IB_QPS_INIT]  = {
1256 			.valid = 1,
1257 			.req_param = {
1258 				[IB_QPT_UD]  = (IB_QP_PKEY_INDEX		|
1259 						IB_QP_PORT			|
1260 						IB_QP_QKEY),
1261 				[IB_QPT_RAW_PACKET] = IB_QP_PORT,
1262 				[IB_QPT_UC]  = (IB_QP_PKEY_INDEX		|
1263 						IB_QP_PORT			|
1264 						IB_QP_ACCESS_FLAGS),
1265 				[IB_QPT_RC]  = (IB_QP_PKEY_INDEX		|
1266 						IB_QP_PORT			|
1267 						IB_QP_ACCESS_FLAGS),
1268 				[IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX		|
1269 						IB_QP_PORT			|
1270 						IB_QP_ACCESS_FLAGS),
1271 				[IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX		|
1272 						IB_QP_PORT			|
1273 						IB_QP_ACCESS_FLAGS),
1274 				[IB_QPT_SMI] = (IB_QP_PKEY_INDEX		|
1275 						IB_QP_QKEY),
1276 				[IB_QPT_GSI] = (IB_QP_PKEY_INDEX		|
1277 						IB_QP_QKEY),
1278 			}
1279 		},
1280 	},
1281 	[IB_QPS_INIT]  = {
1282 		[IB_QPS_RESET] = { .valid = 1 },
1283 		[IB_QPS_ERR] =   { .valid = 1 },
1284 		[IB_QPS_INIT]  = {
1285 			.valid = 1,
1286 			.opt_param = {
1287 				[IB_QPT_UD]  = (IB_QP_PKEY_INDEX		|
1288 						IB_QP_PORT			|
1289 						IB_QP_QKEY),
1290 				[IB_QPT_UC]  = (IB_QP_PKEY_INDEX		|
1291 						IB_QP_PORT			|
1292 						IB_QP_ACCESS_FLAGS),
1293 				[IB_QPT_RC]  = (IB_QP_PKEY_INDEX		|
1294 						IB_QP_PORT			|
1295 						IB_QP_ACCESS_FLAGS),
1296 				[IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX		|
1297 						IB_QP_PORT			|
1298 						IB_QP_ACCESS_FLAGS),
1299 				[IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX		|
1300 						IB_QP_PORT			|
1301 						IB_QP_ACCESS_FLAGS),
1302 				[IB_QPT_SMI] = (IB_QP_PKEY_INDEX		|
1303 						IB_QP_QKEY),
1304 				[IB_QPT_GSI] = (IB_QP_PKEY_INDEX		|
1305 						IB_QP_QKEY),
1306 			}
1307 		},
1308 		[IB_QPS_RTR]   = {
1309 			.valid = 1,
1310 			.req_param = {
1311 				[IB_QPT_UC]  = (IB_QP_AV			|
1312 						IB_QP_PATH_MTU			|
1313 						IB_QP_DEST_QPN			|
1314 						IB_QP_RQ_PSN),
1315 				[IB_QPT_RC]  = (IB_QP_AV			|
1316 						IB_QP_PATH_MTU			|
1317 						IB_QP_DEST_QPN			|
1318 						IB_QP_RQ_PSN			|
1319 						IB_QP_MAX_DEST_RD_ATOMIC	|
1320 						IB_QP_MIN_RNR_TIMER),
1321 				[IB_QPT_XRC_INI] = (IB_QP_AV			|
1322 						IB_QP_PATH_MTU			|
1323 						IB_QP_DEST_QPN			|
1324 						IB_QP_RQ_PSN),
1325 				[IB_QPT_XRC_TGT] = (IB_QP_AV			|
1326 						IB_QP_PATH_MTU			|
1327 						IB_QP_DEST_QPN			|
1328 						IB_QP_RQ_PSN			|
1329 						IB_QP_MAX_DEST_RD_ATOMIC	|
1330 						IB_QP_MIN_RNR_TIMER),
1331 			},
1332 			.opt_param = {
1333 				 [IB_QPT_UD]  = (IB_QP_PKEY_INDEX		|
1334 						 IB_QP_QKEY),
1335 				 [IB_QPT_UC]  = (IB_QP_ALT_PATH			|
1336 						 IB_QP_ACCESS_FLAGS		|
1337 						 IB_QP_PKEY_INDEX),
1338 				 [IB_QPT_RC]  = (IB_QP_ALT_PATH			|
1339 						 IB_QP_ACCESS_FLAGS		|
1340 						 IB_QP_PKEY_INDEX),
1341 				 [IB_QPT_XRC_INI] = (IB_QP_ALT_PATH		|
1342 						 IB_QP_ACCESS_FLAGS		|
1343 						 IB_QP_PKEY_INDEX),
1344 				 [IB_QPT_XRC_TGT] = (IB_QP_ALT_PATH		|
1345 						 IB_QP_ACCESS_FLAGS		|
1346 						 IB_QP_PKEY_INDEX),
1347 				 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX		|
1348 						 IB_QP_QKEY),
1349 				 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX		|
1350 						 IB_QP_QKEY),
1351 			 },
1352 		},
1353 	},
1354 	[IB_QPS_RTR]   = {
1355 		[IB_QPS_RESET] = { .valid = 1 },
1356 		[IB_QPS_ERR] =   { .valid = 1 },
1357 		[IB_QPS_RTS]   = {
1358 			.valid = 1,
1359 			.req_param = {
1360 				[IB_QPT_UD]  = IB_QP_SQ_PSN,
1361 				[IB_QPT_UC]  = IB_QP_SQ_PSN,
1362 				[IB_QPT_RC]  = (IB_QP_TIMEOUT			|
1363 						IB_QP_RETRY_CNT			|
1364 						IB_QP_RNR_RETRY			|
1365 						IB_QP_SQ_PSN			|
1366 						IB_QP_MAX_QP_RD_ATOMIC),
1367 				[IB_QPT_XRC_INI] = (IB_QP_TIMEOUT		|
1368 						IB_QP_RETRY_CNT			|
1369 						IB_QP_RNR_RETRY			|
1370 						IB_QP_SQ_PSN			|
1371 						IB_QP_MAX_QP_RD_ATOMIC),
1372 				[IB_QPT_XRC_TGT] = (IB_QP_TIMEOUT		|
1373 						IB_QP_SQ_PSN),
1374 				[IB_QPT_SMI] = IB_QP_SQ_PSN,
1375 				[IB_QPT_GSI] = IB_QP_SQ_PSN,
1376 			},
1377 			.opt_param = {
1378 				 [IB_QPT_UD]  = (IB_QP_CUR_STATE		|
1379 						 IB_QP_QKEY),
1380 				 [IB_QPT_UC]  = (IB_QP_CUR_STATE		|
1381 						 IB_QP_ALT_PATH			|
1382 						 IB_QP_ACCESS_FLAGS		|
1383 						 IB_QP_PATH_MIG_STATE),
1384 				 [IB_QPT_RC]  = (IB_QP_CUR_STATE		|
1385 						 IB_QP_ALT_PATH			|
1386 						 IB_QP_ACCESS_FLAGS		|
1387 						 IB_QP_MIN_RNR_TIMER		|
1388 						 IB_QP_PATH_MIG_STATE),
1389 				 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE		|
1390 						 IB_QP_ALT_PATH			|
1391 						 IB_QP_ACCESS_FLAGS		|
1392 						 IB_QP_PATH_MIG_STATE),
1393 				 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE		|
1394 						 IB_QP_ALT_PATH			|
1395 						 IB_QP_ACCESS_FLAGS		|
1396 						 IB_QP_MIN_RNR_TIMER		|
1397 						 IB_QP_PATH_MIG_STATE),
1398 				 [IB_QPT_SMI] = (IB_QP_CUR_STATE		|
1399 						 IB_QP_QKEY),
1400 				 [IB_QPT_GSI] = (IB_QP_CUR_STATE		|
1401 						 IB_QP_QKEY),
1402 				 [IB_QPT_RAW_PACKET] = IB_QP_RATE_LIMIT,
1403 			 }
1404 		}
1405 	},
1406 	[IB_QPS_RTS]   = {
1407 		[IB_QPS_RESET] = { .valid = 1 },
1408 		[IB_QPS_ERR] =   { .valid = 1 },
1409 		[IB_QPS_RTS]   = {
1410 			.valid = 1,
1411 			.opt_param = {
1412 				[IB_QPT_UD]  = (IB_QP_CUR_STATE			|
1413 						IB_QP_QKEY),
1414 				[IB_QPT_UC]  = (IB_QP_CUR_STATE			|
1415 						IB_QP_ACCESS_FLAGS		|
1416 						IB_QP_ALT_PATH			|
1417 						IB_QP_PATH_MIG_STATE),
1418 				[IB_QPT_RC]  = (IB_QP_CUR_STATE			|
1419 						IB_QP_ACCESS_FLAGS		|
1420 						IB_QP_ALT_PATH			|
1421 						IB_QP_PATH_MIG_STATE		|
1422 						IB_QP_MIN_RNR_TIMER),
1423 				[IB_QPT_XRC_INI] = (IB_QP_CUR_STATE		|
1424 						IB_QP_ACCESS_FLAGS		|
1425 						IB_QP_ALT_PATH			|
1426 						IB_QP_PATH_MIG_STATE),
1427 				[IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE		|
1428 						IB_QP_ACCESS_FLAGS		|
1429 						IB_QP_ALT_PATH			|
1430 						IB_QP_PATH_MIG_STATE		|
1431 						IB_QP_MIN_RNR_TIMER),
1432 				[IB_QPT_SMI] = (IB_QP_CUR_STATE			|
1433 						IB_QP_QKEY),
1434 				[IB_QPT_GSI] = (IB_QP_CUR_STATE			|
1435 						IB_QP_QKEY),
1436 				[IB_QPT_RAW_PACKET] = IB_QP_RATE_LIMIT,
1437 			}
1438 		},
1439 		[IB_QPS_SQD]   = {
1440 			.valid = 1,
1441 			.opt_param = {
1442 				[IB_QPT_UD]  = IB_QP_EN_SQD_ASYNC_NOTIFY,
1443 				[IB_QPT_UC]  = IB_QP_EN_SQD_ASYNC_NOTIFY,
1444 				[IB_QPT_RC]  = IB_QP_EN_SQD_ASYNC_NOTIFY,
1445 				[IB_QPT_XRC_INI] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1446 				[IB_QPT_XRC_TGT] = IB_QP_EN_SQD_ASYNC_NOTIFY, /* ??? */
1447 				[IB_QPT_SMI] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1448 				[IB_QPT_GSI] = IB_QP_EN_SQD_ASYNC_NOTIFY
1449 			}
1450 		},
1451 	},
1452 	[IB_QPS_SQD]   = {
1453 		[IB_QPS_RESET] = { .valid = 1 },
1454 		[IB_QPS_ERR] =   { .valid = 1 },
1455 		[IB_QPS_RTS]   = {
1456 			.valid = 1,
1457 			.opt_param = {
1458 				[IB_QPT_UD]  = (IB_QP_CUR_STATE			|
1459 						IB_QP_QKEY),
1460 				[IB_QPT_UC]  = (IB_QP_CUR_STATE			|
1461 						IB_QP_ALT_PATH			|
1462 						IB_QP_ACCESS_FLAGS		|
1463 						IB_QP_PATH_MIG_STATE),
1464 				[IB_QPT_RC]  = (IB_QP_CUR_STATE			|
1465 						IB_QP_ALT_PATH			|
1466 						IB_QP_ACCESS_FLAGS		|
1467 						IB_QP_MIN_RNR_TIMER		|
1468 						IB_QP_PATH_MIG_STATE),
1469 				[IB_QPT_XRC_INI] = (IB_QP_CUR_STATE		|
1470 						IB_QP_ALT_PATH			|
1471 						IB_QP_ACCESS_FLAGS		|
1472 						IB_QP_PATH_MIG_STATE),
1473 				[IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE		|
1474 						IB_QP_ALT_PATH			|
1475 						IB_QP_ACCESS_FLAGS		|
1476 						IB_QP_MIN_RNR_TIMER		|
1477 						IB_QP_PATH_MIG_STATE),
1478 				[IB_QPT_SMI] = (IB_QP_CUR_STATE			|
1479 						IB_QP_QKEY),
1480 				[IB_QPT_GSI] = (IB_QP_CUR_STATE			|
1481 						IB_QP_QKEY),
1482 			}
1483 		},
1484 		[IB_QPS_SQD]   = {
1485 			.valid = 1,
1486 			.opt_param = {
1487 				[IB_QPT_UD]  = (IB_QP_PKEY_INDEX		|
1488 						IB_QP_QKEY),
1489 				[IB_QPT_UC]  = (IB_QP_AV			|
1490 						IB_QP_ALT_PATH			|
1491 						IB_QP_ACCESS_FLAGS		|
1492 						IB_QP_PKEY_INDEX		|
1493 						IB_QP_PATH_MIG_STATE),
1494 				[IB_QPT_RC]  = (IB_QP_PORT			|
1495 						IB_QP_AV			|
1496 						IB_QP_TIMEOUT			|
1497 						IB_QP_RETRY_CNT			|
1498 						IB_QP_RNR_RETRY			|
1499 						IB_QP_MAX_QP_RD_ATOMIC		|
1500 						IB_QP_MAX_DEST_RD_ATOMIC	|
1501 						IB_QP_ALT_PATH			|
1502 						IB_QP_ACCESS_FLAGS		|
1503 						IB_QP_PKEY_INDEX		|
1504 						IB_QP_MIN_RNR_TIMER		|
1505 						IB_QP_PATH_MIG_STATE),
1506 				[IB_QPT_XRC_INI] = (IB_QP_PORT			|
1507 						IB_QP_AV			|
1508 						IB_QP_TIMEOUT			|
1509 						IB_QP_RETRY_CNT			|
1510 						IB_QP_RNR_RETRY			|
1511 						IB_QP_MAX_QP_RD_ATOMIC		|
1512 						IB_QP_ALT_PATH			|
1513 						IB_QP_ACCESS_FLAGS		|
1514 						IB_QP_PKEY_INDEX		|
1515 						IB_QP_PATH_MIG_STATE),
1516 				[IB_QPT_XRC_TGT] = (IB_QP_PORT			|
1517 						IB_QP_AV			|
1518 						IB_QP_TIMEOUT			|
1519 						IB_QP_MAX_DEST_RD_ATOMIC	|
1520 						IB_QP_ALT_PATH			|
1521 						IB_QP_ACCESS_FLAGS		|
1522 						IB_QP_PKEY_INDEX		|
1523 						IB_QP_MIN_RNR_TIMER		|
1524 						IB_QP_PATH_MIG_STATE),
1525 				[IB_QPT_SMI] = (IB_QP_PKEY_INDEX		|
1526 						IB_QP_QKEY),
1527 				[IB_QPT_GSI] = (IB_QP_PKEY_INDEX		|
1528 						IB_QP_QKEY),
1529 			}
1530 		}
1531 	},
1532 	[IB_QPS_SQE]   = {
1533 		[IB_QPS_RESET] = { .valid = 1 },
1534 		[IB_QPS_ERR] =   { .valid = 1 },
1535 		[IB_QPS_RTS]   = {
1536 			.valid = 1,
1537 			.opt_param = {
1538 				[IB_QPT_UD]  = (IB_QP_CUR_STATE			|
1539 						IB_QP_QKEY),
1540 				[IB_QPT_UC]  = (IB_QP_CUR_STATE			|
1541 						IB_QP_ACCESS_FLAGS),
1542 				[IB_QPT_SMI] = (IB_QP_CUR_STATE			|
1543 						IB_QP_QKEY),
1544 				[IB_QPT_GSI] = (IB_QP_CUR_STATE			|
1545 						IB_QP_QKEY),
1546 			}
1547 		}
1548 	},
1549 	[IB_QPS_ERR] = {
1550 		[IB_QPS_RESET] = { .valid = 1 },
1551 		[IB_QPS_ERR] =   { .valid = 1 }
1552 	}
1553 };
1554 
1555 bool ib_modify_qp_is_ok(enum ib_qp_state cur_state, enum ib_qp_state next_state,
1556 			enum ib_qp_type type, enum ib_qp_attr_mask mask)
1557 {
1558 	enum ib_qp_attr_mask req_param, opt_param;
1559 
1560 	if (mask & IB_QP_CUR_STATE  &&
1561 	    cur_state != IB_QPS_RTR && cur_state != IB_QPS_RTS &&
1562 	    cur_state != IB_QPS_SQD && cur_state != IB_QPS_SQE)
1563 		return false;
1564 
1565 	if (!qp_state_table[cur_state][next_state].valid)
1566 		return false;
1567 
1568 	req_param = qp_state_table[cur_state][next_state].req_param[type];
1569 	opt_param = qp_state_table[cur_state][next_state].opt_param[type];
1570 
1571 	if ((mask & req_param) != req_param)
1572 		return false;
1573 
1574 	if (mask & ~(req_param | opt_param | IB_QP_STATE))
1575 		return false;
1576 
1577 	return true;
1578 }
1579 EXPORT_SYMBOL(ib_modify_qp_is_ok);
1580 
1581 /**
1582  * ib_resolve_eth_dmac - Resolve destination mac address
1583  * @device:		Device to consider
1584  * @ah_attr:		address handle attribute which describes the
1585  *			source and destination parameters
1586  * ib_resolve_eth_dmac() resolves destination mac address and L3 hop limit It
1587  * returns 0 on success or appropriate error code. It initializes the
1588  * necessary ah_attr fields when call is successful.
1589  */
1590 static int ib_resolve_eth_dmac(struct ib_device *device,
1591 			       struct rdma_ah_attr *ah_attr)
1592 {
1593 	int ret = 0;
1594 
1595 	if (rdma_is_multicast_addr((struct in6_addr *)ah_attr->grh.dgid.raw)) {
1596 		if (ipv6_addr_v4mapped((struct in6_addr *)ah_attr->grh.dgid.raw)) {
1597 			__be32 addr = 0;
1598 
1599 			memcpy(&addr, ah_attr->grh.dgid.raw + 12, 4);
1600 			ip_eth_mc_map(addr, (char *)ah_attr->roce.dmac);
1601 		} else {
1602 			ipv6_eth_mc_map((struct in6_addr *)ah_attr->grh.dgid.raw,
1603 					(char *)ah_attr->roce.dmac);
1604 		}
1605 	} else {
1606 		ret = ib_resolve_unicast_gid_dmac(device, ah_attr);
1607 	}
1608 	return ret;
1609 }
1610 
1611 static bool is_qp_type_connected(const struct ib_qp *qp)
1612 {
1613 	return (qp->qp_type == IB_QPT_UC ||
1614 		qp->qp_type == IB_QPT_RC ||
1615 		qp->qp_type == IB_QPT_XRC_INI ||
1616 		qp->qp_type == IB_QPT_XRC_TGT);
1617 }
1618 
1619 /**
1620  * IB core internal function to perform QP attributes modification.
1621  */
1622 static int _ib_modify_qp(struct ib_qp *qp, struct ib_qp_attr *attr,
1623 			 int attr_mask, struct ib_udata *udata)
1624 {
1625 	u8 port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
1626 	const struct ib_gid_attr *old_sgid_attr_av;
1627 	const struct ib_gid_attr *old_sgid_attr_alt_av;
1628 	int ret;
1629 
1630 	if (attr_mask & IB_QP_AV) {
1631 		ret = rdma_fill_sgid_attr(qp->device, &attr->ah_attr,
1632 					  &old_sgid_attr_av);
1633 		if (ret)
1634 			return ret;
1635 	}
1636 	if (attr_mask & IB_QP_ALT_PATH) {
1637 		/*
1638 		 * FIXME: This does not track the migration state, so if the
1639 		 * user loads a new alternate path after the HW has migrated
1640 		 * from primary->alternate we will keep the wrong
1641 		 * references. This is OK for IB because the reference
1642 		 * counting does not serve any functional purpose.
1643 		 */
1644 		ret = rdma_fill_sgid_attr(qp->device, &attr->alt_ah_attr,
1645 					  &old_sgid_attr_alt_av);
1646 		if (ret)
1647 			goto out_av;
1648 
1649 		/*
1650 		 * Today the core code can only handle alternate paths and APM
1651 		 * for IB. Ban them in roce mode.
1652 		 */
1653 		if (!(rdma_protocol_ib(qp->device,
1654 				       attr->alt_ah_attr.port_num) &&
1655 		      rdma_protocol_ib(qp->device, port))) {
1656 			ret = EINVAL;
1657 			goto out;
1658 		}
1659 	}
1660 
1661 	/*
1662 	 * If the user provided the qp_attr then we have to resolve it. Kernel
1663 	 * users have to provide already resolved rdma_ah_attr's
1664 	 */
1665 	if (udata && (attr_mask & IB_QP_AV) &&
1666 	    attr->ah_attr.type == RDMA_AH_ATTR_TYPE_ROCE &&
1667 	    is_qp_type_connected(qp)) {
1668 		ret = ib_resolve_eth_dmac(qp->device, &attr->ah_attr);
1669 		if (ret)
1670 			goto out;
1671 	}
1672 
1673 	if (rdma_ib_or_roce(qp->device, port)) {
1674 		if (attr_mask & IB_QP_RQ_PSN && attr->rq_psn & ~0xffffff) {
1675 			dev_warn(&qp->device->dev,
1676 				 "%s rq_psn overflow, masking to 24 bits\n",
1677 				 __func__);
1678 			attr->rq_psn &= 0xffffff;
1679 		}
1680 
1681 		if (attr_mask & IB_QP_SQ_PSN && attr->sq_psn & ~0xffffff) {
1682 			dev_warn(&qp->device->dev,
1683 				 " %s sq_psn overflow, masking to 24 bits\n",
1684 				 __func__);
1685 			attr->sq_psn &= 0xffffff;
1686 		}
1687 	}
1688 
1689 	/*
1690 	 * Bind this qp to a counter automatically based on the rdma counter
1691 	 * rules. This only set in RST2INIT with port specified
1692 	 */
1693 	if (!qp->counter && (attr_mask & IB_QP_PORT) &&
1694 	    ((attr_mask & IB_QP_STATE) && attr->qp_state == IB_QPS_INIT))
1695 		rdma_counter_bind_qp_auto(qp, attr->port_num);
1696 
1697 	ret = ib_security_modify_qp(qp, attr, attr_mask, udata);
1698 	if (ret)
1699 		goto out;
1700 
1701 	if (attr_mask & IB_QP_PORT)
1702 		qp->port = attr->port_num;
1703 	if (attr_mask & IB_QP_AV)
1704 		qp->av_sgid_attr =
1705 			rdma_update_sgid_attr(&attr->ah_attr, qp->av_sgid_attr);
1706 	if (attr_mask & IB_QP_ALT_PATH)
1707 		qp->alt_path_sgid_attr = rdma_update_sgid_attr(
1708 			&attr->alt_ah_attr, qp->alt_path_sgid_attr);
1709 
1710 out:
1711 	if (attr_mask & IB_QP_ALT_PATH)
1712 		rdma_unfill_sgid_attr(&attr->alt_ah_attr, old_sgid_attr_alt_av);
1713 out_av:
1714 	if (attr_mask & IB_QP_AV)
1715 		rdma_unfill_sgid_attr(&attr->ah_attr, old_sgid_attr_av);
1716 	return ret;
1717 }
1718 
1719 /**
1720  * ib_modify_qp_with_udata - Modifies the attributes for the specified QP.
1721  * @ib_qp: The QP to modify.
1722  * @attr: On input, specifies the QP attributes to modify.  On output,
1723  *   the current values of selected QP attributes are returned.
1724  * @attr_mask: A bit-mask used to specify which attributes of the QP
1725  *   are being modified.
1726  * @udata: pointer to user's input output buffer information
1727  *   are being modified.
1728  * It returns 0 on success and returns appropriate error code on error.
1729  */
1730 int ib_modify_qp_with_udata(struct ib_qp *ib_qp, struct ib_qp_attr *attr,
1731 			    int attr_mask, struct ib_udata *udata)
1732 {
1733 	return _ib_modify_qp(ib_qp->real_qp, attr, attr_mask, udata);
1734 }
1735 EXPORT_SYMBOL(ib_modify_qp_with_udata);
1736 
1737 int ib_get_eth_speed(struct ib_device *dev, u8 port_num, u8 *speed, u8 *width)
1738 {
1739 	int rc;
1740 	u32 netdev_speed;
1741 	struct net_device *netdev;
1742 	struct ethtool_link_ksettings lksettings;
1743 
1744 	if (rdma_port_get_link_layer(dev, port_num) != IB_LINK_LAYER_ETHERNET)
1745 		return -EINVAL;
1746 
1747 	netdev = ib_device_get_netdev(dev, port_num);
1748 	if (!netdev)
1749 		return -ENODEV;
1750 
1751 	rtnl_lock();
1752 	rc = __ethtool_get_link_ksettings(netdev, &lksettings);
1753 	rtnl_unlock();
1754 
1755 	dev_put(netdev);
1756 
1757 	if (!rc) {
1758 		netdev_speed = lksettings.base.speed;
1759 	} else {
1760 		netdev_speed = SPEED_1000;
1761 		pr_warn("%s speed is unknown, defaulting to %d\n", netdev->name,
1762 			netdev_speed);
1763 	}
1764 
1765 	if (netdev_speed <= SPEED_1000) {
1766 		*width = IB_WIDTH_1X;
1767 		*speed = IB_SPEED_SDR;
1768 	} else if (netdev_speed <= SPEED_10000) {
1769 		*width = IB_WIDTH_1X;
1770 		*speed = IB_SPEED_FDR10;
1771 	} else if (netdev_speed <= SPEED_20000) {
1772 		*width = IB_WIDTH_4X;
1773 		*speed = IB_SPEED_DDR;
1774 	} else if (netdev_speed <= SPEED_25000) {
1775 		*width = IB_WIDTH_1X;
1776 		*speed = IB_SPEED_EDR;
1777 	} else if (netdev_speed <= SPEED_40000) {
1778 		*width = IB_WIDTH_4X;
1779 		*speed = IB_SPEED_FDR10;
1780 	} else {
1781 		*width = IB_WIDTH_4X;
1782 		*speed = IB_SPEED_EDR;
1783 	}
1784 
1785 	return 0;
1786 }
1787 EXPORT_SYMBOL(ib_get_eth_speed);
1788 
1789 int ib_modify_qp(struct ib_qp *qp,
1790 		 struct ib_qp_attr *qp_attr,
1791 		 int qp_attr_mask)
1792 {
1793 	return _ib_modify_qp(qp->real_qp, qp_attr, qp_attr_mask, NULL);
1794 }
1795 EXPORT_SYMBOL(ib_modify_qp);
1796 
1797 int ib_query_qp(struct ib_qp *qp,
1798 		struct ib_qp_attr *qp_attr,
1799 		int qp_attr_mask,
1800 		struct ib_qp_init_attr *qp_init_attr)
1801 {
1802 	qp_attr->ah_attr.grh.sgid_attr = NULL;
1803 	qp_attr->alt_ah_attr.grh.sgid_attr = NULL;
1804 
1805 	return qp->device->ops.query_qp ?
1806 		qp->device->ops.query_qp(qp->real_qp, qp_attr, qp_attr_mask,
1807 					 qp_init_attr) : -EOPNOTSUPP;
1808 }
1809 EXPORT_SYMBOL(ib_query_qp);
1810 
1811 int ib_close_qp(struct ib_qp *qp)
1812 {
1813 	struct ib_qp *real_qp;
1814 	unsigned long flags;
1815 
1816 	real_qp = qp->real_qp;
1817 	if (real_qp == qp)
1818 		return -EINVAL;
1819 
1820 	spin_lock_irqsave(&real_qp->device->qp_open_list_lock, flags);
1821 	list_del(&qp->open_list);
1822 	spin_unlock_irqrestore(&real_qp->device->qp_open_list_lock, flags);
1823 
1824 	atomic_dec(&real_qp->usecnt);
1825 	if (qp->qp_sec)
1826 		ib_close_shared_qp_security(qp->qp_sec);
1827 	kfree(qp);
1828 
1829 	return 0;
1830 }
1831 EXPORT_SYMBOL(ib_close_qp);
1832 
1833 static int __ib_destroy_shared_qp(struct ib_qp *qp)
1834 {
1835 	struct ib_xrcd *xrcd;
1836 	struct ib_qp *real_qp;
1837 	int ret;
1838 
1839 	real_qp = qp->real_qp;
1840 	xrcd = real_qp->xrcd;
1841 
1842 	mutex_lock(&xrcd->tgt_qp_mutex);
1843 	ib_close_qp(qp);
1844 	if (atomic_read(&real_qp->usecnt) == 0)
1845 		list_del(&real_qp->xrcd_list);
1846 	else
1847 		real_qp = NULL;
1848 	mutex_unlock(&xrcd->tgt_qp_mutex);
1849 
1850 	if (real_qp) {
1851 		ret = ib_destroy_qp(real_qp);
1852 		if (!ret)
1853 			atomic_dec(&xrcd->usecnt);
1854 		else
1855 			__ib_insert_xrcd_qp(xrcd, real_qp);
1856 	}
1857 
1858 	return 0;
1859 }
1860 
1861 int ib_destroy_qp_user(struct ib_qp *qp, struct ib_udata *udata)
1862 {
1863 	const struct ib_gid_attr *alt_path_sgid_attr = qp->alt_path_sgid_attr;
1864 	const struct ib_gid_attr *av_sgid_attr = qp->av_sgid_attr;
1865 	struct ib_pd *pd;
1866 	struct ib_cq *scq, *rcq;
1867 	struct ib_srq *srq;
1868 	struct ib_rwq_ind_table *ind_tbl;
1869 	struct ib_qp_security *sec;
1870 	int ret;
1871 
1872 	WARN_ON_ONCE(qp->mrs_used > 0);
1873 
1874 	if (atomic_read(&qp->usecnt))
1875 		return -EBUSY;
1876 
1877 	if (qp->real_qp != qp)
1878 		return __ib_destroy_shared_qp(qp);
1879 
1880 	pd   = qp->pd;
1881 	scq  = qp->send_cq;
1882 	rcq  = qp->recv_cq;
1883 	srq  = qp->srq;
1884 	ind_tbl = qp->rwq_ind_tbl;
1885 	sec  = qp->qp_sec;
1886 	if (sec)
1887 		ib_destroy_qp_security_begin(sec);
1888 
1889 	if (!qp->uobject)
1890 		rdma_rw_cleanup_mrs(qp);
1891 
1892 	rdma_counter_unbind_qp(qp, true);
1893 	rdma_restrack_del(&qp->res);
1894 	ret = qp->device->ops.destroy_qp(qp, udata);
1895 	if (!ret) {
1896 		if (alt_path_sgid_attr)
1897 			rdma_put_gid_attr(alt_path_sgid_attr);
1898 		if (av_sgid_attr)
1899 			rdma_put_gid_attr(av_sgid_attr);
1900 		if (pd)
1901 			atomic_dec(&pd->usecnt);
1902 		if (scq)
1903 			atomic_dec(&scq->usecnt);
1904 		if (rcq)
1905 			atomic_dec(&rcq->usecnt);
1906 		if (srq)
1907 			atomic_dec(&srq->usecnt);
1908 		if (ind_tbl)
1909 			atomic_dec(&ind_tbl->usecnt);
1910 		if (sec)
1911 			ib_destroy_qp_security_end(sec);
1912 	} else {
1913 		if (sec)
1914 			ib_destroy_qp_security_abort(sec);
1915 	}
1916 
1917 	return ret;
1918 }
1919 EXPORT_SYMBOL(ib_destroy_qp_user);
1920 
1921 /* Completion queues */
1922 
1923 struct ib_cq *__ib_create_cq(struct ib_device *device,
1924 			     ib_comp_handler comp_handler,
1925 			     void (*event_handler)(struct ib_event *, void *),
1926 			     void *cq_context,
1927 			     const struct ib_cq_init_attr *cq_attr,
1928 			     const char *caller)
1929 {
1930 	struct ib_cq *cq;
1931 	int ret;
1932 
1933 	cq = rdma_zalloc_drv_obj(device, ib_cq);
1934 	if (!cq)
1935 		return ERR_PTR(-ENOMEM);
1936 
1937 	cq->device = device;
1938 	cq->uobject = NULL;
1939 	cq->comp_handler = comp_handler;
1940 	cq->event_handler = event_handler;
1941 	cq->cq_context = cq_context;
1942 	atomic_set(&cq->usecnt, 0);
1943 	cq->res.type = RDMA_RESTRACK_CQ;
1944 	rdma_restrack_set_task(&cq->res, caller);
1945 
1946 	ret = device->ops.create_cq(cq, cq_attr, NULL);
1947 	if (ret) {
1948 		kfree(cq);
1949 		return ERR_PTR(ret);
1950 	}
1951 
1952 	rdma_restrack_kadd(&cq->res);
1953 	return cq;
1954 }
1955 EXPORT_SYMBOL(__ib_create_cq);
1956 
1957 int rdma_set_cq_moderation(struct ib_cq *cq, u16 cq_count, u16 cq_period)
1958 {
1959 	return cq->device->ops.modify_cq ?
1960 		cq->device->ops.modify_cq(cq, cq_count,
1961 					  cq_period) : -EOPNOTSUPP;
1962 }
1963 EXPORT_SYMBOL(rdma_set_cq_moderation);
1964 
1965 int ib_destroy_cq_user(struct ib_cq *cq, struct ib_udata *udata)
1966 {
1967 	if (atomic_read(&cq->usecnt))
1968 		return -EBUSY;
1969 
1970 	rdma_restrack_del(&cq->res);
1971 	cq->device->ops.destroy_cq(cq, udata);
1972 	kfree(cq);
1973 	return 0;
1974 }
1975 EXPORT_SYMBOL(ib_destroy_cq_user);
1976 
1977 int ib_resize_cq(struct ib_cq *cq, int cqe)
1978 {
1979 	return cq->device->ops.resize_cq ?
1980 		cq->device->ops.resize_cq(cq, cqe, NULL) : -EOPNOTSUPP;
1981 }
1982 EXPORT_SYMBOL(ib_resize_cq);
1983 
1984 /* Memory regions */
1985 
1986 struct ib_mr *ib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
1987 			     u64 virt_addr, int access_flags)
1988 {
1989 	struct ib_mr *mr;
1990 
1991 	if (access_flags & IB_ACCESS_ON_DEMAND) {
1992 		if (!(pd->device->attrs.device_cap_flags &
1993 		      IB_DEVICE_ON_DEMAND_PAGING)) {
1994 			pr_debug("ODP support not available\n");
1995 			return ERR_PTR(-EINVAL);
1996 		}
1997 	}
1998 
1999 	mr = pd->device->ops.reg_user_mr(pd, start, length, virt_addr,
2000 					 access_flags, NULL);
2001 
2002 	if (IS_ERR(mr))
2003 		return mr;
2004 
2005 	mr->device = pd->device;
2006 	mr->pd = pd;
2007 	mr->dm = NULL;
2008 	atomic_inc(&pd->usecnt);
2009 	mr->res.type = RDMA_RESTRACK_MR;
2010 	rdma_restrack_kadd(&mr->res);
2011 
2012 	return mr;
2013 }
2014 EXPORT_SYMBOL(ib_reg_user_mr);
2015 
2016 int ib_advise_mr(struct ib_pd *pd, enum ib_uverbs_advise_mr_advice advice,
2017 		 u32 flags, struct ib_sge *sg_list, u32 num_sge)
2018 {
2019 	if (!pd->device->ops.advise_mr)
2020 		return -EOPNOTSUPP;
2021 
2022 	return pd->device->ops.advise_mr(pd, advice, flags, sg_list, num_sge,
2023 					 NULL);
2024 }
2025 EXPORT_SYMBOL(ib_advise_mr);
2026 
2027 int ib_dereg_mr_user(struct ib_mr *mr, struct ib_udata *udata)
2028 {
2029 	struct ib_pd *pd = mr->pd;
2030 	struct ib_dm *dm = mr->dm;
2031 	struct ib_sig_attrs *sig_attrs = mr->sig_attrs;
2032 	int ret;
2033 
2034 	trace_mr_dereg(mr);
2035 	rdma_restrack_del(&mr->res);
2036 	ret = mr->device->ops.dereg_mr(mr, udata);
2037 	if (!ret) {
2038 		atomic_dec(&pd->usecnt);
2039 		if (dm)
2040 			atomic_dec(&dm->usecnt);
2041 		kfree(sig_attrs);
2042 	}
2043 
2044 	return ret;
2045 }
2046 EXPORT_SYMBOL(ib_dereg_mr_user);
2047 
2048 /**
2049  * ib_alloc_mr_user() - Allocates a memory region
2050  * @pd:            protection domain associated with the region
2051  * @mr_type:       memory region type
2052  * @max_num_sg:    maximum sg entries available for registration.
2053  * @udata:	   user data or null for kernel objects
2054  *
2055  * Notes:
2056  * Memory registeration page/sg lists must not exceed max_num_sg.
2057  * For mr_type IB_MR_TYPE_MEM_REG, the total length cannot exceed
2058  * max_num_sg * used_page_size.
2059  *
2060  */
2061 struct ib_mr *ib_alloc_mr_user(struct ib_pd *pd, enum ib_mr_type mr_type,
2062 			       u32 max_num_sg, struct ib_udata *udata)
2063 {
2064 	struct ib_mr *mr;
2065 
2066 	if (!pd->device->ops.alloc_mr) {
2067 		mr = ERR_PTR(-EOPNOTSUPP);
2068 		goto out;
2069 	}
2070 
2071 	if (mr_type == IB_MR_TYPE_INTEGRITY) {
2072 		WARN_ON_ONCE(1);
2073 		mr = ERR_PTR(-EINVAL);
2074 		goto out;
2075 	}
2076 
2077 	mr = pd->device->ops.alloc_mr(pd, mr_type, max_num_sg, udata);
2078 	if (!IS_ERR(mr)) {
2079 		mr->device  = pd->device;
2080 		mr->pd      = pd;
2081 		mr->dm      = NULL;
2082 		mr->uobject = NULL;
2083 		atomic_inc(&pd->usecnt);
2084 		mr->need_inval = false;
2085 		mr->res.type = RDMA_RESTRACK_MR;
2086 		rdma_restrack_kadd(&mr->res);
2087 		mr->type = mr_type;
2088 		mr->sig_attrs = NULL;
2089 	}
2090 
2091 out:
2092 	trace_mr_alloc(pd, mr_type, max_num_sg, mr);
2093 	return mr;
2094 }
2095 EXPORT_SYMBOL(ib_alloc_mr_user);
2096 
2097 /**
2098  * ib_alloc_mr_integrity() - Allocates an integrity memory region
2099  * @pd:                      protection domain associated with the region
2100  * @max_num_data_sg:         maximum data sg entries available for registration
2101  * @max_num_meta_sg:         maximum metadata sg entries available for
2102  *                           registration
2103  *
2104  * Notes:
2105  * Memory registration page/sg lists must not exceed max_num_sg,
2106  * also the integrity page/sg lists must not exceed max_num_meta_sg.
2107  *
2108  */
2109 struct ib_mr *ib_alloc_mr_integrity(struct ib_pd *pd,
2110 				    u32 max_num_data_sg,
2111 				    u32 max_num_meta_sg)
2112 {
2113 	struct ib_mr *mr;
2114 	struct ib_sig_attrs *sig_attrs;
2115 
2116 	if (!pd->device->ops.alloc_mr_integrity ||
2117 	    !pd->device->ops.map_mr_sg_pi) {
2118 		mr = ERR_PTR(-EOPNOTSUPP);
2119 		goto out;
2120 	}
2121 
2122 	if (!max_num_meta_sg) {
2123 		mr = ERR_PTR(-EINVAL);
2124 		goto out;
2125 	}
2126 
2127 	sig_attrs = kzalloc(sizeof(struct ib_sig_attrs), GFP_KERNEL);
2128 	if (!sig_attrs) {
2129 		mr = ERR_PTR(-ENOMEM);
2130 		goto out;
2131 	}
2132 
2133 	mr = pd->device->ops.alloc_mr_integrity(pd, max_num_data_sg,
2134 						max_num_meta_sg);
2135 	if (IS_ERR(mr)) {
2136 		kfree(sig_attrs);
2137 		goto out;
2138 	}
2139 
2140 	mr->device = pd->device;
2141 	mr->pd = pd;
2142 	mr->dm = NULL;
2143 	mr->uobject = NULL;
2144 	atomic_inc(&pd->usecnt);
2145 	mr->need_inval = false;
2146 	mr->res.type = RDMA_RESTRACK_MR;
2147 	rdma_restrack_kadd(&mr->res);
2148 	mr->type = IB_MR_TYPE_INTEGRITY;
2149 	mr->sig_attrs = sig_attrs;
2150 
2151 out:
2152 	trace_mr_integ_alloc(pd, max_num_data_sg, max_num_meta_sg, mr);
2153 	return mr;
2154 }
2155 EXPORT_SYMBOL(ib_alloc_mr_integrity);
2156 
2157 /* "Fast" memory regions */
2158 
2159 struct ib_fmr *ib_alloc_fmr(struct ib_pd *pd,
2160 			    int mr_access_flags,
2161 			    struct ib_fmr_attr *fmr_attr)
2162 {
2163 	struct ib_fmr *fmr;
2164 
2165 	if (!pd->device->ops.alloc_fmr)
2166 		return ERR_PTR(-EOPNOTSUPP);
2167 
2168 	fmr = pd->device->ops.alloc_fmr(pd, mr_access_flags, fmr_attr);
2169 	if (!IS_ERR(fmr)) {
2170 		fmr->device = pd->device;
2171 		fmr->pd     = pd;
2172 		atomic_inc(&pd->usecnt);
2173 	}
2174 
2175 	return fmr;
2176 }
2177 EXPORT_SYMBOL(ib_alloc_fmr);
2178 
2179 int ib_unmap_fmr(struct list_head *fmr_list)
2180 {
2181 	struct ib_fmr *fmr;
2182 
2183 	if (list_empty(fmr_list))
2184 		return 0;
2185 
2186 	fmr = list_entry(fmr_list->next, struct ib_fmr, list);
2187 	return fmr->device->ops.unmap_fmr(fmr_list);
2188 }
2189 EXPORT_SYMBOL(ib_unmap_fmr);
2190 
2191 int ib_dealloc_fmr(struct ib_fmr *fmr)
2192 {
2193 	struct ib_pd *pd;
2194 	int ret;
2195 
2196 	pd = fmr->pd;
2197 	ret = fmr->device->ops.dealloc_fmr(fmr);
2198 	if (!ret)
2199 		atomic_dec(&pd->usecnt);
2200 
2201 	return ret;
2202 }
2203 EXPORT_SYMBOL(ib_dealloc_fmr);
2204 
2205 /* Multicast groups */
2206 
2207 static bool is_valid_mcast_lid(struct ib_qp *qp, u16 lid)
2208 {
2209 	struct ib_qp_init_attr init_attr = {};
2210 	struct ib_qp_attr attr = {};
2211 	int num_eth_ports = 0;
2212 	int port;
2213 
2214 	/* If QP state >= init, it is assigned to a port and we can check this
2215 	 * port only.
2216 	 */
2217 	if (!ib_query_qp(qp, &attr, IB_QP_STATE | IB_QP_PORT, &init_attr)) {
2218 		if (attr.qp_state >= IB_QPS_INIT) {
2219 			if (rdma_port_get_link_layer(qp->device, attr.port_num) !=
2220 			    IB_LINK_LAYER_INFINIBAND)
2221 				return true;
2222 			goto lid_check;
2223 		}
2224 	}
2225 
2226 	/* Can't get a quick answer, iterate over all ports */
2227 	for (port = 0; port < qp->device->phys_port_cnt; port++)
2228 		if (rdma_port_get_link_layer(qp->device, port) !=
2229 		    IB_LINK_LAYER_INFINIBAND)
2230 			num_eth_ports++;
2231 
2232 	/* If we have at lease one Ethernet port, RoCE annex declares that
2233 	 * multicast LID should be ignored. We can't tell at this step if the
2234 	 * QP belongs to an IB or Ethernet port.
2235 	 */
2236 	if (num_eth_ports)
2237 		return true;
2238 
2239 	/* If all the ports are IB, we can check according to IB spec. */
2240 lid_check:
2241 	return !(lid < be16_to_cpu(IB_MULTICAST_LID_BASE) ||
2242 		 lid == be16_to_cpu(IB_LID_PERMISSIVE));
2243 }
2244 
2245 int ib_attach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid)
2246 {
2247 	int ret;
2248 
2249 	if (!qp->device->ops.attach_mcast)
2250 		return -EOPNOTSUPP;
2251 
2252 	if (!rdma_is_multicast_addr((struct in6_addr *)gid->raw) ||
2253 	    qp->qp_type != IB_QPT_UD || !is_valid_mcast_lid(qp, lid))
2254 		return -EINVAL;
2255 
2256 	ret = qp->device->ops.attach_mcast(qp, gid, lid);
2257 	if (!ret)
2258 		atomic_inc(&qp->usecnt);
2259 	return ret;
2260 }
2261 EXPORT_SYMBOL(ib_attach_mcast);
2262 
2263 int ib_detach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid)
2264 {
2265 	int ret;
2266 
2267 	if (!qp->device->ops.detach_mcast)
2268 		return -EOPNOTSUPP;
2269 
2270 	if (!rdma_is_multicast_addr((struct in6_addr *)gid->raw) ||
2271 	    qp->qp_type != IB_QPT_UD || !is_valid_mcast_lid(qp, lid))
2272 		return -EINVAL;
2273 
2274 	ret = qp->device->ops.detach_mcast(qp, gid, lid);
2275 	if (!ret)
2276 		atomic_dec(&qp->usecnt);
2277 	return ret;
2278 }
2279 EXPORT_SYMBOL(ib_detach_mcast);
2280 
2281 struct ib_xrcd *__ib_alloc_xrcd(struct ib_device *device, const char *caller)
2282 {
2283 	struct ib_xrcd *xrcd;
2284 
2285 	if (!device->ops.alloc_xrcd)
2286 		return ERR_PTR(-EOPNOTSUPP);
2287 
2288 	xrcd = device->ops.alloc_xrcd(device, NULL);
2289 	if (!IS_ERR(xrcd)) {
2290 		xrcd->device = device;
2291 		xrcd->inode = NULL;
2292 		atomic_set(&xrcd->usecnt, 0);
2293 		mutex_init(&xrcd->tgt_qp_mutex);
2294 		INIT_LIST_HEAD(&xrcd->tgt_qp_list);
2295 	}
2296 
2297 	return xrcd;
2298 }
2299 EXPORT_SYMBOL(__ib_alloc_xrcd);
2300 
2301 int ib_dealloc_xrcd(struct ib_xrcd *xrcd, struct ib_udata *udata)
2302 {
2303 	struct ib_qp *qp;
2304 	int ret;
2305 
2306 	if (atomic_read(&xrcd->usecnt))
2307 		return -EBUSY;
2308 
2309 	while (!list_empty(&xrcd->tgt_qp_list)) {
2310 		qp = list_entry(xrcd->tgt_qp_list.next, struct ib_qp, xrcd_list);
2311 		ret = ib_destroy_qp(qp);
2312 		if (ret)
2313 			return ret;
2314 	}
2315 	mutex_destroy(&xrcd->tgt_qp_mutex);
2316 
2317 	return xrcd->device->ops.dealloc_xrcd(xrcd, udata);
2318 }
2319 EXPORT_SYMBOL(ib_dealloc_xrcd);
2320 
2321 /**
2322  * ib_create_wq - Creates a WQ associated with the specified protection
2323  * domain.
2324  * @pd: The protection domain associated with the WQ.
2325  * @wq_attr: A list of initial attributes required to create the
2326  * WQ. If WQ creation succeeds, then the attributes are updated to
2327  * the actual capabilities of the created WQ.
2328  *
2329  * wq_attr->max_wr and wq_attr->max_sge determine
2330  * the requested size of the WQ, and set to the actual values allocated
2331  * on return.
2332  * If ib_create_wq() succeeds, then max_wr and max_sge will always be
2333  * at least as large as the requested values.
2334  */
2335 struct ib_wq *ib_create_wq(struct ib_pd *pd,
2336 			   struct ib_wq_init_attr *wq_attr)
2337 {
2338 	struct ib_wq *wq;
2339 
2340 	if (!pd->device->ops.create_wq)
2341 		return ERR_PTR(-EOPNOTSUPP);
2342 
2343 	wq = pd->device->ops.create_wq(pd, wq_attr, NULL);
2344 	if (!IS_ERR(wq)) {
2345 		wq->event_handler = wq_attr->event_handler;
2346 		wq->wq_context = wq_attr->wq_context;
2347 		wq->wq_type = wq_attr->wq_type;
2348 		wq->cq = wq_attr->cq;
2349 		wq->device = pd->device;
2350 		wq->pd = pd;
2351 		wq->uobject = NULL;
2352 		atomic_inc(&pd->usecnt);
2353 		atomic_inc(&wq_attr->cq->usecnt);
2354 		atomic_set(&wq->usecnt, 0);
2355 	}
2356 	return wq;
2357 }
2358 EXPORT_SYMBOL(ib_create_wq);
2359 
2360 /**
2361  * ib_destroy_wq - Destroys the specified user WQ.
2362  * @wq: The WQ to destroy.
2363  * @udata: Valid user data
2364  */
2365 int ib_destroy_wq(struct ib_wq *wq, struct ib_udata *udata)
2366 {
2367 	struct ib_cq *cq = wq->cq;
2368 	struct ib_pd *pd = wq->pd;
2369 
2370 	if (atomic_read(&wq->usecnt))
2371 		return -EBUSY;
2372 
2373 	wq->device->ops.destroy_wq(wq, udata);
2374 	atomic_dec(&pd->usecnt);
2375 	atomic_dec(&cq->usecnt);
2376 
2377 	return 0;
2378 }
2379 EXPORT_SYMBOL(ib_destroy_wq);
2380 
2381 /**
2382  * ib_modify_wq - Modifies the specified WQ.
2383  * @wq: The WQ to modify.
2384  * @wq_attr: On input, specifies the WQ attributes to modify.
2385  * @wq_attr_mask: A bit-mask used to specify which attributes of the WQ
2386  *   are being modified.
2387  * On output, the current values of selected WQ attributes are returned.
2388  */
2389 int ib_modify_wq(struct ib_wq *wq, struct ib_wq_attr *wq_attr,
2390 		 u32 wq_attr_mask)
2391 {
2392 	int err;
2393 
2394 	if (!wq->device->ops.modify_wq)
2395 		return -EOPNOTSUPP;
2396 
2397 	err = wq->device->ops.modify_wq(wq, wq_attr, wq_attr_mask, NULL);
2398 	return err;
2399 }
2400 EXPORT_SYMBOL(ib_modify_wq);
2401 
2402 /*
2403  * ib_create_rwq_ind_table - Creates a RQ Indirection Table.
2404  * @device: The device on which to create the rwq indirection table.
2405  * @ib_rwq_ind_table_init_attr: A list of initial attributes required to
2406  * create the Indirection Table.
2407  *
2408  * Note: The life time of ib_rwq_ind_table_init_attr->ind_tbl is not less
2409  *	than the created ib_rwq_ind_table object and the caller is responsible
2410  *	for its memory allocation/free.
2411  */
2412 struct ib_rwq_ind_table *ib_create_rwq_ind_table(struct ib_device *device,
2413 						 struct ib_rwq_ind_table_init_attr *init_attr)
2414 {
2415 	struct ib_rwq_ind_table *rwq_ind_table;
2416 	int i;
2417 	u32 table_size;
2418 
2419 	if (!device->ops.create_rwq_ind_table)
2420 		return ERR_PTR(-EOPNOTSUPP);
2421 
2422 	table_size = (1 << init_attr->log_ind_tbl_size);
2423 	rwq_ind_table = device->ops.create_rwq_ind_table(device,
2424 							 init_attr, NULL);
2425 	if (IS_ERR(rwq_ind_table))
2426 		return rwq_ind_table;
2427 
2428 	rwq_ind_table->ind_tbl = init_attr->ind_tbl;
2429 	rwq_ind_table->log_ind_tbl_size = init_attr->log_ind_tbl_size;
2430 	rwq_ind_table->device = device;
2431 	rwq_ind_table->uobject = NULL;
2432 	atomic_set(&rwq_ind_table->usecnt, 0);
2433 
2434 	for (i = 0; i < table_size; i++)
2435 		atomic_inc(&rwq_ind_table->ind_tbl[i]->usecnt);
2436 
2437 	return rwq_ind_table;
2438 }
2439 EXPORT_SYMBOL(ib_create_rwq_ind_table);
2440 
2441 /*
2442  * ib_destroy_rwq_ind_table - Destroys the specified Indirection Table.
2443  * @wq_ind_table: The Indirection Table to destroy.
2444 */
2445 int ib_destroy_rwq_ind_table(struct ib_rwq_ind_table *rwq_ind_table)
2446 {
2447 	int err, i;
2448 	u32 table_size = (1 << rwq_ind_table->log_ind_tbl_size);
2449 	struct ib_wq **ind_tbl = rwq_ind_table->ind_tbl;
2450 
2451 	if (atomic_read(&rwq_ind_table->usecnt))
2452 		return -EBUSY;
2453 
2454 	err = rwq_ind_table->device->ops.destroy_rwq_ind_table(rwq_ind_table);
2455 	if (!err) {
2456 		for (i = 0; i < table_size; i++)
2457 			atomic_dec(&ind_tbl[i]->usecnt);
2458 	}
2459 
2460 	return err;
2461 }
2462 EXPORT_SYMBOL(ib_destroy_rwq_ind_table);
2463 
2464 int ib_check_mr_status(struct ib_mr *mr, u32 check_mask,
2465 		       struct ib_mr_status *mr_status)
2466 {
2467 	if (!mr->device->ops.check_mr_status)
2468 		return -EOPNOTSUPP;
2469 
2470 	return mr->device->ops.check_mr_status(mr, check_mask, mr_status);
2471 }
2472 EXPORT_SYMBOL(ib_check_mr_status);
2473 
2474 int ib_set_vf_link_state(struct ib_device *device, int vf, u8 port,
2475 			 int state)
2476 {
2477 	if (!device->ops.set_vf_link_state)
2478 		return -EOPNOTSUPP;
2479 
2480 	return device->ops.set_vf_link_state(device, vf, port, state);
2481 }
2482 EXPORT_SYMBOL(ib_set_vf_link_state);
2483 
2484 int ib_get_vf_config(struct ib_device *device, int vf, u8 port,
2485 		     struct ifla_vf_info *info)
2486 {
2487 	if (!device->ops.get_vf_config)
2488 		return -EOPNOTSUPP;
2489 
2490 	return device->ops.get_vf_config(device, vf, port, info);
2491 }
2492 EXPORT_SYMBOL(ib_get_vf_config);
2493 
2494 int ib_get_vf_stats(struct ib_device *device, int vf, u8 port,
2495 		    struct ifla_vf_stats *stats)
2496 {
2497 	if (!device->ops.get_vf_stats)
2498 		return -EOPNOTSUPP;
2499 
2500 	return device->ops.get_vf_stats(device, vf, port, stats);
2501 }
2502 EXPORT_SYMBOL(ib_get_vf_stats);
2503 
2504 int ib_set_vf_guid(struct ib_device *device, int vf, u8 port, u64 guid,
2505 		   int type)
2506 {
2507 	if (!device->ops.set_vf_guid)
2508 		return -EOPNOTSUPP;
2509 
2510 	return device->ops.set_vf_guid(device, vf, port, guid, type);
2511 }
2512 EXPORT_SYMBOL(ib_set_vf_guid);
2513 
2514 int ib_get_vf_guid(struct ib_device *device, int vf, u8 port,
2515 		   struct ifla_vf_guid *node_guid,
2516 		   struct ifla_vf_guid *port_guid)
2517 {
2518 	if (!device->ops.get_vf_guid)
2519 		return -EOPNOTSUPP;
2520 
2521 	return device->ops.get_vf_guid(device, vf, port, node_guid, port_guid);
2522 }
2523 EXPORT_SYMBOL(ib_get_vf_guid);
2524 /**
2525  * ib_map_mr_sg_pi() - Map the dma mapped SG lists for PI (protection
2526  *     information) and set an appropriate memory region for registration.
2527  * @mr:             memory region
2528  * @data_sg:        dma mapped scatterlist for data
2529  * @data_sg_nents:  number of entries in data_sg
2530  * @data_sg_offset: offset in bytes into data_sg
2531  * @meta_sg:        dma mapped scatterlist for metadata
2532  * @meta_sg_nents:  number of entries in meta_sg
2533  * @meta_sg_offset: offset in bytes into meta_sg
2534  * @page_size:      page vector desired page size
2535  *
2536  * Constraints:
2537  * - The MR must be allocated with type IB_MR_TYPE_INTEGRITY.
2538  *
2539  * Return: 0 on success.
2540  *
2541  * After this completes successfully, the  memory region
2542  * is ready for registration.
2543  */
2544 int ib_map_mr_sg_pi(struct ib_mr *mr, struct scatterlist *data_sg,
2545 		    int data_sg_nents, unsigned int *data_sg_offset,
2546 		    struct scatterlist *meta_sg, int meta_sg_nents,
2547 		    unsigned int *meta_sg_offset, unsigned int page_size)
2548 {
2549 	if (unlikely(!mr->device->ops.map_mr_sg_pi ||
2550 		     WARN_ON_ONCE(mr->type != IB_MR_TYPE_INTEGRITY)))
2551 		return -EOPNOTSUPP;
2552 
2553 	mr->page_size = page_size;
2554 
2555 	return mr->device->ops.map_mr_sg_pi(mr, data_sg, data_sg_nents,
2556 					    data_sg_offset, meta_sg,
2557 					    meta_sg_nents, meta_sg_offset);
2558 }
2559 EXPORT_SYMBOL(ib_map_mr_sg_pi);
2560 
2561 /**
2562  * ib_map_mr_sg() - Map the largest prefix of a dma mapped SG list
2563  *     and set it the memory region.
2564  * @mr:            memory region
2565  * @sg:            dma mapped scatterlist
2566  * @sg_nents:      number of entries in sg
2567  * @sg_offset:     offset in bytes into sg
2568  * @page_size:     page vector desired page size
2569  *
2570  * Constraints:
2571  * - The first sg element is allowed to have an offset.
2572  * - Each sg element must either be aligned to page_size or virtually
2573  *   contiguous to the previous element. In case an sg element has a
2574  *   non-contiguous offset, the mapping prefix will not include it.
2575  * - The last sg element is allowed to have length less than page_size.
2576  * - If sg_nents total byte length exceeds the mr max_num_sge * page_size
2577  *   then only max_num_sg entries will be mapped.
2578  * - If the MR was allocated with type IB_MR_TYPE_SG_GAPS, none of these
2579  *   constraints holds and the page_size argument is ignored.
2580  *
2581  * Returns the number of sg elements that were mapped to the memory region.
2582  *
2583  * After this completes successfully, the  memory region
2584  * is ready for registration.
2585  */
2586 int ib_map_mr_sg(struct ib_mr *mr, struct scatterlist *sg, int sg_nents,
2587 		 unsigned int *sg_offset, unsigned int page_size)
2588 {
2589 	if (unlikely(!mr->device->ops.map_mr_sg))
2590 		return -EOPNOTSUPP;
2591 
2592 	mr->page_size = page_size;
2593 
2594 	return mr->device->ops.map_mr_sg(mr, sg, sg_nents, sg_offset);
2595 }
2596 EXPORT_SYMBOL(ib_map_mr_sg);
2597 
2598 /**
2599  * ib_sg_to_pages() - Convert the largest prefix of a sg list
2600  *     to a page vector
2601  * @mr:            memory region
2602  * @sgl:           dma mapped scatterlist
2603  * @sg_nents:      number of entries in sg
2604  * @sg_offset_p:   IN:  start offset in bytes into sg
2605  *                 OUT: offset in bytes for element n of the sg of the first
2606  *                      byte that has not been processed where n is the return
2607  *                      value of this function.
2608  * @set_page:      driver page assignment function pointer
2609  *
2610  * Core service helper for drivers to convert the largest
2611  * prefix of given sg list to a page vector. The sg list
2612  * prefix converted is the prefix that meet the requirements
2613  * of ib_map_mr_sg.
2614  *
2615  * Returns the number of sg elements that were assigned to
2616  * a page vector.
2617  */
2618 int ib_sg_to_pages(struct ib_mr *mr, struct scatterlist *sgl, int sg_nents,
2619 		unsigned int *sg_offset_p, int (*set_page)(struct ib_mr *, u64))
2620 {
2621 	struct scatterlist *sg;
2622 	u64 last_end_dma_addr = 0;
2623 	unsigned int sg_offset = sg_offset_p ? *sg_offset_p : 0;
2624 	unsigned int last_page_off = 0;
2625 	u64 page_mask = ~((u64)mr->page_size - 1);
2626 	int i, ret;
2627 
2628 	if (unlikely(sg_nents <= 0 || sg_offset > sg_dma_len(&sgl[0])))
2629 		return -EINVAL;
2630 
2631 	mr->iova = sg_dma_address(&sgl[0]) + sg_offset;
2632 	mr->length = 0;
2633 
2634 	for_each_sg(sgl, sg, sg_nents, i) {
2635 		u64 dma_addr = sg_dma_address(sg) + sg_offset;
2636 		u64 prev_addr = dma_addr;
2637 		unsigned int dma_len = sg_dma_len(sg) - sg_offset;
2638 		u64 end_dma_addr = dma_addr + dma_len;
2639 		u64 page_addr = dma_addr & page_mask;
2640 
2641 		/*
2642 		 * For the second and later elements, check whether either the
2643 		 * end of element i-1 or the start of element i is not aligned
2644 		 * on a page boundary.
2645 		 */
2646 		if (i && (last_page_off != 0 || page_addr != dma_addr)) {
2647 			/* Stop mapping if there is a gap. */
2648 			if (last_end_dma_addr != dma_addr)
2649 				break;
2650 
2651 			/*
2652 			 * Coalesce this element with the last. If it is small
2653 			 * enough just update mr->length. Otherwise start
2654 			 * mapping from the next page.
2655 			 */
2656 			goto next_page;
2657 		}
2658 
2659 		do {
2660 			ret = set_page(mr, page_addr);
2661 			if (unlikely(ret < 0)) {
2662 				sg_offset = prev_addr - sg_dma_address(sg);
2663 				mr->length += prev_addr - dma_addr;
2664 				if (sg_offset_p)
2665 					*sg_offset_p = sg_offset;
2666 				return i || sg_offset ? i : ret;
2667 			}
2668 			prev_addr = page_addr;
2669 next_page:
2670 			page_addr += mr->page_size;
2671 		} while (page_addr < end_dma_addr);
2672 
2673 		mr->length += dma_len;
2674 		last_end_dma_addr = end_dma_addr;
2675 		last_page_off = end_dma_addr & ~page_mask;
2676 
2677 		sg_offset = 0;
2678 	}
2679 
2680 	if (sg_offset_p)
2681 		*sg_offset_p = 0;
2682 	return i;
2683 }
2684 EXPORT_SYMBOL(ib_sg_to_pages);
2685 
2686 struct ib_drain_cqe {
2687 	struct ib_cqe cqe;
2688 	struct completion done;
2689 };
2690 
2691 static void ib_drain_qp_done(struct ib_cq *cq, struct ib_wc *wc)
2692 {
2693 	struct ib_drain_cqe *cqe = container_of(wc->wr_cqe, struct ib_drain_cqe,
2694 						cqe);
2695 
2696 	complete(&cqe->done);
2697 }
2698 
2699 /*
2700  * Post a WR and block until its completion is reaped for the SQ.
2701  */
2702 static void __ib_drain_sq(struct ib_qp *qp)
2703 {
2704 	struct ib_cq *cq = qp->send_cq;
2705 	struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
2706 	struct ib_drain_cqe sdrain;
2707 	struct ib_rdma_wr swr = {
2708 		.wr = {
2709 			.next = NULL,
2710 			{ .wr_cqe	= &sdrain.cqe, },
2711 			.opcode	= IB_WR_RDMA_WRITE,
2712 		},
2713 	};
2714 	int ret;
2715 
2716 	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
2717 	if (ret) {
2718 		WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
2719 		return;
2720 	}
2721 
2722 	sdrain.cqe.done = ib_drain_qp_done;
2723 	init_completion(&sdrain.done);
2724 
2725 	ret = ib_post_send(qp, &swr.wr, NULL);
2726 	if (ret) {
2727 		WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
2728 		return;
2729 	}
2730 
2731 	if (cq->poll_ctx == IB_POLL_DIRECT)
2732 		while (wait_for_completion_timeout(&sdrain.done, HZ / 10) <= 0)
2733 			ib_process_cq_direct(cq, -1);
2734 	else
2735 		wait_for_completion(&sdrain.done);
2736 }
2737 
2738 /*
2739  * Post a WR and block until its completion is reaped for the RQ.
2740  */
2741 static void __ib_drain_rq(struct ib_qp *qp)
2742 {
2743 	struct ib_cq *cq = qp->recv_cq;
2744 	struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
2745 	struct ib_drain_cqe rdrain;
2746 	struct ib_recv_wr rwr = {};
2747 	int ret;
2748 
2749 	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
2750 	if (ret) {
2751 		WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
2752 		return;
2753 	}
2754 
2755 	rwr.wr_cqe = &rdrain.cqe;
2756 	rdrain.cqe.done = ib_drain_qp_done;
2757 	init_completion(&rdrain.done);
2758 
2759 	ret = ib_post_recv(qp, &rwr, NULL);
2760 	if (ret) {
2761 		WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
2762 		return;
2763 	}
2764 
2765 	if (cq->poll_ctx == IB_POLL_DIRECT)
2766 		while (wait_for_completion_timeout(&rdrain.done, HZ / 10) <= 0)
2767 			ib_process_cq_direct(cq, -1);
2768 	else
2769 		wait_for_completion(&rdrain.done);
2770 }
2771 
2772 /**
2773  * ib_drain_sq() - Block until all SQ CQEs have been consumed by the
2774  *		   application.
2775  * @qp:            queue pair to drain
2776  *
2777  * If the device has a provider-specific drain function, then
2778  * call that.  Otherwise call the generic drain function
2779  * __ib_drain_sq().
2780  *
2781  * The caller must:
2782  *
2783  * ensure there is room in the CQ and SQ for the drain work request and
2784  * completion.
2785  *
2786  * allocate the CQ using ib_alloc_cq().
2787  *
2788  * ensure that there are no other contexts that are posting WRs concurrently.
2789  * Otherwise the drain is not guaranteed.
2790  */
2791 void ib_drain_sq(struct ib_qp *qp)
2792 {
2793 	if (qp->device->ops.drain_sq)
2794 		qp->device->ops.drain_sq(qp);
2795 	else
2796 		__ib_drain_sq(qp);
2797 	trace_cq_drain_complete(qp->send_cq);
2798 }
2799 EXPORT_SYMBOL(ib_drain_sq);
2800 
2801 /**
2802  * ib_drain_rq() - Block until all RQ CQEs have been consumed by the
2803  *		   application.
2804  * @qp:            queue pair to drain
2805  *
2806  * If the device has a provider-specific drain function, then
2807  * call that.  Otherwise call the generic drain function
2808  * __ib_drain_rq().
2809  *
2810  * The caller must:
2811  *
2812  * ensure there is room in the CQ and RQ for the drain work request and
2813  * completion.
2814  *
2815  * allocate the CQ using ib_alloc_cq().
2816  *
2817  * ensure that there are no other contexts that are posting WRs concurrently.
2818  * Otherwise the drain is not guaranteed.
2819  */
2820 void ib_drain_rq(struct ib_qp *qp)
2821 {
2822 	if (qp->device->ops.drain_rq)
2823 		qp->device->ops.drain_rq(qp);
2824 	else
2825 		__ib_drain_rq(qp);
2826 	trace_cq_drain_complete(qp->recv_cq);
2827 }
2828 EXPORT_SYMBOL(ib_drain_rq);
2829 
2830 /**
2831  * ib_drain_qp() - Block until all CQEs have been consumed by the
2832  *		   application on both the RQ and SQ.
2833  * @qp:            queue pair to drain
2834  *
2835  * The caller must:
2836  *
2837  * ensure there is room in the CQ(s), SQ, and RQ for drain work requests
2838  * and completions.
2839  *
2840  * allocate the CQs using ib_alloc_cq().
2841  *
2842  * ensure that there are no other contexts that are posting WRs concurrently.
2843  * Otherwise the drain is not guaranteed.
2844  */
2845 void ib_drain_qp(struct ib_qp *qp)
2846 {
2847 	ib_drain_sq(qp);
2848 	if (!qp->srq)
2849 		ib_drain_rq(qp);
2850 }
2851 EXPORT_SYMBOL(ib_drain_qp);
2852 
2853 struct net_device *rdma_alloc_netdev(struct ib_device *device, u8 port_num,
2854 				     enum rdma_netdev_t type, const char *name,
2855 				     unsigned char name_assign_type,
2856 				     void (*setup)(struct net_device *))
2857 {
2858 	struct rdma_netdev_alloc_params params;
2859 	struct net_device *netdev;
2860 	int rc;
2861 
2862 	if (!device->ops.rdma_netdev_get_params)
2863 		return ERR_PTR(-EOPNOTSUPP);
2864 
2865 	rc = device->ops.rdma_netdev_get_params(device, port_num, type,
2866 						&params);
2867 	if (rc)
2868 		return ERR_PTR(rc);
2869 
2870 	netdev = alloc_netdev_mqs(params.sizeof_priv, name, name_assign_type,
2871 				  setup, params.txqs, params.rxqs);
2872 	if (!netdev)
2873 		return ERR_PTR(-ENOMEM);
2874 
2875 	return netdev;
2876 }
2877 EXPORT_SYMBOL(rdma_alloc_netdev);
2878 
2879 int rdma_init_netdev(struct ib_device *device, u8 port_num,
2880 		     enum rdma_netdev_t type, const char *name,
2881 		     unsigned char name_assign_type,
2882 		     void (*setup)(struct net_device *),
2883 		     struct net_device *netdev)
2884 {
2885 	struct rdma_netdev_alloc_params params;
2886 	int rc;
2887 
2888 	if (!device->ops.rdma_netdev_get_params)
2889 		return -EOPNOTSUPP;
2890 
2891 	rc = device->ops.rdma_netdev_get_params(device, port_num, type,
2892 						&params);
2893 	if (rc)
2894 		return rc;
2895 
2896 	return params.initialize_rdma_netdev(device, port_num,
2897 					     netdev, params.param);
2898 }
2899 EXPORT_SYMBOL(rdma_init_netdev);
2900 
2901 void __rdma_block_iter_start(struct ib_block_iter *biter,
2902 			     struct scatterlist *sglist, unsigned int nents,
2903 			     unsigned long pgsz)
2904 {
2905 	memset(biter, 0, sizeof(struct ib_block_iter));
2906 	biter->__sg = sglist;
2907 	biter->__sg_nents = nents;
2908 
2909 	/* Driver provides best block size to use */
2910 	biter->__pg_bit = __fls(pgsz);
2911 }
2912 EXPORT_SYMBOL(__rdma_block_iter_start);
2913 
2914 bool __rdma_block_iter_next(struct ib_block_iter *biter)
2915 {
2916 	unsigned int block_offset;
2917 
2918 	if (!biter->__sg_nents || !biter->__sg)
2919 		return false;
2920 
2921 	biter->__dma_addr = sg_dma_address(biter->__sg) + biter->__sg_advance;
2922 	block_offset = biter->__dma_addr & (BIT_ULL(biter->__pg_bit) - 1);
2923 	biter->__sg_advance += BIT_ULL(biter->__pg_bit) - block_offset;
2924 
2925 	if (biter->__sg_advance >= sg_dma_len(biter->__sg)) {
2926 		biter->__sg_advance = 0;
2927 		biter->__sg = sg_next(biter->__sg);
2928 		biter->__sg_nents--;
2929 	}
2930 
2931 	return true;
2932 }
2933 EXPORT_SYMBOL(__rdma_block_iter_next);
2934