xref: /openbmc/linux/drivers/infiniband/core/verbs.c (revision 11a163f2)
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 #include <rdma/lag.h>
54 
55 #include "core_priv.h"
56 #include <trace/events/rdma_core.h>
57 
58 static int ib_resolve_eth_dmac(struct ib_device *device,
59 			       struct rdma_ah_attr *ah_attr);
60 
61 static const char * const ib_events[] = {
62 	[IB_EVENT_CQ_ERR]		= "CQ error",
63 	[IB_EVENT_QP_FATAL]		= "QP fatal error",
64 	[IB_EVENT_QP_REQ_ERR]		= "QP request error",
65 	[IB_EVENT_QP_ACCESS_ERR]	= "QP access error",
66 	[IB_EVENT_COMM_EST]		= "communication established",
67 	[IB_EVENT_SQ_DRAINED]		= "send queue drained",
68 	[IB_EVENT_PATH_MIG]		= "path migration successful",
69 	[IB_EVENT_PATH_MIG_ERR]		= "path migration error",
70 	[IB_EVENT_DEVICE_FATAL]		= "device fatal error",
71 	[IB_EVENT_PORT_ACTIVE]		= "port active",
72 	[IB_EVENT_PORT_ERR]		= "port error",
73 	[IB_EVENT_LID_CHANGE]		= "LID change",
74 	[IB_EVENT_PKEY_CHANGE]		= "P_key change",
75 	[IB_EVENT_SM_CHANGE]		= "SM change",
76 	[IB_EVENT_SRQ_ERR]		= "SRQ error",
77 	[IB_EVENT_SRQ_LIMIT_REACHED]	= "SRQ limit reached",
78 	[IB_EVENT_QP_LAST_WQE_REACHED]	= "last WQE reached",
79 	[IB_EVENT_CLIENT_REREGISTER]	= "client reregister",
80 	[IB_EVENT_GID_CHANGE]		= "GID changed",
81 };
82 
83 const char *__attribute_const__ ib_event_msg(enum ib_event_type event)
84 {
85 	size_t index = event;
86 
87 	return (index < ARRAY_SIZE(ib_events) && ib_events[index]) ?
88 			ib_events[index] : "unrecognized event";
89 }
90 EXPORT_SYMBOL(ib_event_msg);
91 
92 static const char * const wc_statuses[] = {
93 	[IB_WC_SUCCESS]			= "success",
94 	[IB_WC_LOC_LEN_ERR]		= "local length error",
95 	[IB_WC_LOC_QP_OP_ERR]		= "local QP operation error",
96 	[IB_WC_LOC_EEC_OP_ERR]		= "local EE context operation error",
97 	[IB_WC_LOC_PROT_ERR]		= "local protection error",
98 	[IB_WC_WR_FLUSH_ERR]		= "WR flushed",
99 	[IB_WC_MW_BIND_ERR]		= "memory management operation error",
100 	[IB_WC_BAD_RESP_ERR]		= "bad response error",
101 	[IB_WC_LOC_ACCESS_ERR]		= "local access error",
102 	[IB_WC_REM_INV_REQ_ERR]		= "invalid request error",
103 	[IB_WC_REM_ACCESS_ERR]		= "remote access error",
104 	[IB_WC_REM_OP_ERR]		= "remote operation error",
105 	[IB_WC_RETRY_EXC_ERR]		= "transport retry counter exceeded",
106 	[IB_WC_RNR_RETRY_EXC_ERR]	= "RNR retry counter exceeded",
107 	[IB_WC_LOC_RDD_VIOL_ERR]	= "local RDD violation error",
108 	[IB_WC_REM_INV_RD_REQ_ERR]	= "remote invalid RD request",
109 	[IB_WC_REM_ABORT_ERR]		= "operation aborted",
110 	[IB_WC_INV_EECN_ERR]		= "invalid EE context number",
111 	[IB_WC_INV_EEC_STATE_ERR]	= "invalid EE context state",
112 	[IB_WC_FATAL_ERR]		= "fatal error",
113 	[IB_WC_RESP_TIMEOUT_ERR]	= "response timeout error",
114 	[IB_WC_GENERAL_ERR]		= "general error",
115 };
116 
117 const char *__attribute_const__ ib_wc_status_msg(enum ib_wc_status status)
118 {
119 	size_t index = status;
120 
121 	return (index < ARRAY_SIZE(wc_statuses) && wc_statuses[index]) ?
122 			wc_statuses[index] : "unrecognized status";
123 }
124 EXPORT_SYMBOL(ib_wc_status_msg);
125 
126 __attribute_const__ int ib_rate_to_mult(enum ib_rate rate)
127 {
128 	switch (rate) {
129 	case IB_RATE_2_5_GBPS: return   1;
130 	case IB_RATE_5_GBPS:   return   2;
131 	case IB_RATE_10_GBPS:  return   4;
132 	case IB_RATE_20_GBPS:  return   8;
133 	case IB_RATE_30_GBPS:  return  12;
134 	case IB_RATE_40_GBPS:  return  16;
135 	case IB_RATE_60_GBPS:  return  24;
136 	case IB_RATE_80_GBPS:  return  32;
137 	case IB_RATE_120_GBPS: return  48;
138 	case IB_RATE_14_GBPS:  return   6;
139 	case IB_RATE_56_GBPS:  return  22;
140 	case IB_RATE_112_GBPS: return  45;
141 	case IB_RATE_168_GBPS: return  67;
142 	case IB_RATE_25_GBPS:  return  10;
143 	case IB_RATE_100_GBPS: return  40;
144 	case IB_RATE_200_GBPS: return  80;
145 	case IB_RATE_300_GBPS: return 120;
146 	case IB_RATE_28_GBPS:  return  11;
147 	case IB_RATE_50_GBPS:  return  20;
148 	case IB_RATE_400_GBPS: return 160;
149 	case IB_RATE_600_GBPS: return 240;
150 	default:	       return  -1;
151 	}
152 }
153 EXPORT_SYMBOL(ib_rate_to_mult);
154 
155 __attribute_const__ enum ib_rate mult_to_ib_rate(int mult)
156 {
157 	switch (mult) {
158 	case 1:   return IB_RATE_2_5_GBPS;
159 	case 2:   return IB_RATE_5_GBPS;
160 	case 4:   return IB_RATE_10_GBPS;
161 	case 8:   return IB_RATE_20_GBPS;
162 	case 12:  return IB_RATE_30_GBPS;
163 	case 16:  return IB_RATE_40_GBPS;
164 	case 24:  return IB_RATE_60_GBPS;
165 	case 32:  return IB_RATE_80_GBPS;
166 	case 48:  return IB_RATE_120_GBPS;
167 	case 6:   return IB_RATE_14_GBPS;
168 	case 22:  return IB_RATE_56_GBPS;
169 	case 45:  return IB_RATE_112_GBPS;
170 	case 67:  return IB_RATE_168_GBPS;
171 	case 10:  return IB_RATE_25_GBPS;
172 	case 40:  return IB_RATE_100_GBPS;
173 	case 80:  return IB_RATE_200_GBPS;
174 	case 120: return IB_RATE_300_GBPS;
175 	case 11:  return IB_RATE_28_GBPS;
176 	case 20:  return IB_RATE_50_GBPS;
177 	case 160: return IB_RATE_400_GBPS;
178 	case 240: return IB_RATE_600_GBPS;
179 	default:  return IB_RATE_PORT_CURRENT;
180 	}
181 }
182 EXPORT_SYMBOL(mult_to_ib_rate);
183 
184 __attribute_const__ int ib_rate_to_mbps(enum ib_rate rate)
185 {
186 	switch (rate) {
187 	case IB_RATE_2_5_GBPS: return 2500;
188 	case IB_RATE_5_GBPS:   return 5000;
189 	case IB_RATE_10_GBPS:  return 10000;
190 	case IB_RATE_20_GBPS:  return 20000;
191 	case IB_RATE_30_GBPS:  return 30000;
192 	case IB_RATE_40_GBPS:  return 40000;
193 	case IB_RATE_60_GBPS:  return 60000;
194 	case IB_RATE_80_GBPS:  return 80000;
195 	case IB_RATE_120_GBPS: return 120000;
196 	case IB_RATE_14_GBPS:  return 14062;
197 	case IB_RATE_56_GBPS:  return 56250;
198 	case IB_RATE_112_GBPS: return 112500;
199 	case IB_RATE_168_GBPS: return 168750;
200 	case IB_RATE_25_GBPS:  return 25781;
201 	case IB_RATE_100_GBPS: return 103125;
202 	case IB_RATE_200_GBPS: return 206250;
203 	case IB_RATE_300_GBPS: return 309375;
204 	case IB_RATE_28_GBPS:  return 28125;
205 	case IB_RATE_50_GBPS:  return 53125;
206 	case IB_RATE_400_GBPS: return 425000;
207 	case IB_RATE_600_GBPS: return 637500;
208 	default:	       return -1;
209 	}
210 }
211 EXPORT_SYMBOL(ib_rate_to_mbps);
212 
213 __attribute_const__ enum rdma_transport_type
214 rdma_node_get_transport(unsigned int node_type)
215 {
216 
217 	if (node_type == RDMA_NODE_USNIC)
218 		return RDMA_TRANSPORT_USNIC;
219 	if (node_type == RDMA_NODE_USNIC_UDP)
220 		return RDMA_TRANSPORT_USNIC_UDP;
221 	if (node_type == RDMA_NODE_RNIC)
222 		return RDMA_TRANSPORT_IWARP;
223 	if (node_type == RDMA_NODE_UNSPECIFIED)
224 		return RDMA_TRANSPORT_UNSPECIFIED;
225 
226 	return RDMA_TRANSPORT_IB;
227 }
228 EXPORT_SYMBOL(rdma_node_get_transport);
229 
230 enum rdma_link_layer rdma_port_get_link_layer(struct ib_device *device, u8 port_num)
231 {
232 	enum rdma_transport_type lt;
233 	if (device->ops.get_link_layer)
234 		return device->ops.get_link_layer(device, port_num);
235 
236 	lt = rdma_node_get_transport(device->node_type);
237 	if (lt == RDMA_TRANSPORT_IB)
238 		return IB_LINK_LAYER_INFINIBAND;
239 
240 	return IB_LINK_LAYER_ETHERNET;
241 }
242 EXPORT_SYMBOL(rdma_port_get_link_layer);
243 
244 /* Protection domains */
245 
246 /**
247  * ib_alloc_pd - Allocates an unused protection domain.
248  * @device: The device on which to allocate the protection domain.
249  * @flags: protection domain flags
250  * @caller: caller's build-time module name
251  *
252  * A protection domain object provides an association between QPs, shared
253  * receive queues, address handles, memory regions, and memory windows.
254  *
255  * Every PD has a local_dma_lkey which can be used as the lkey value for local
256  * memory operations.
257  */
258 struct ib_pd *__ib_alloc_pd(struct ib_device *device, unsigned int flags,
259 		const char *caller)
260 {
261 	struct ib_pd *pd;
262 	int mr_access_flags = 0;
263 	int ret;
264 
265 	pd = rdma_zalloc_drv_obj(device, ib_pd);
266 	if (!pd)
267 		return ERR_PTR(-ENOMEM);
268 
269 	pd->device = device;
270 	pd->uobject = NULL;
271 	pd->__internal_mr = NULL;
272 	atomic_set(&pd->usecnt, 0);
273 	pd->flags = flags;
274 
275 	rdma_restrack_new(&pd->res, RDMA_RESTRACK_PD);
276 	rdma_restrack_set_name(&pd->res, caller);
277 
278 	ret = device->ops.alloc_pd(pd, NULL);
279 	if (ret) {
280 		rdma_restrack_put(&pd->res);
281 		kfree(pd);
282 		return ERR_PTR(ret);
283 	}
284 	rdma_restrack_add(&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 int 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 	ret = pd->device->ops.dealloc_pd(pd, udata);
348 	if (ret)
349 		return ret;
350 
351 	rdma_restrack_del(&pd->res);
352 	kfree(pd);
353 	return ret;
354 }
355 EXPORT_SYMBOL(ib_dealloc_pd_user);
356 
357 /* Address handles */
358 
359 /**
360  * rdma_copy_ah_attr - Copy rdma ah attribute from source to destination.
361  * @dest:       Pointer to destination ah_attr. Contents of the destination
362  *              pointer is assumed to be invalid and attribute are overwritten.
363  * @src:        Pointer to source ah_attr.
364  */
365 void rdma_copy_ah_attr(struct rdma_ah_attr *dest,
366 		       const struct rdma_ah_attr *src)
367 {
368 	*dest = *src;
369 	if (dest->grh.sgid_attr)
370 		rdma_hold_gid_attr(dest->grh.sgid_attr);
371 }
372 EXPORT_SYMBOL(rdma_copy_ah_attr);
373 
374 /**
375  * rdma_replace_ah_attr - Replace valid ah_attr with new new one.
376  * @old:        Pointer to existing ah_attr which needs to be replaced.
377  *              old is assumed to be valid or zero'd
378  * @new:        Pointer to the new ah_attr.
379  *
380  * rdma_replace_ah_attr() first releases any reference in the old ah_attr if
381  * old the ah_attr is valid; after that it copies the new attribute and holds
382  * the reference to the replaced ah_attr.
383  */
384 void rdma_replace_ah_attr(struct rdma_ah_attr *old,
385 			  const struct rdma_ah_attr *new)
386 {
387 	rdma_destroy_ah_attr(old);
388 	*old = *new;
389 	if (old->grh.sgid_attr)
390 		rdma_hold_gid_attr(old->grh.sgid_attr);
391 }
392 EXPORT_SYMBOL(rdma_replace_ah_attr);
393 
394 /**
395  * rdma_move_ah_attr - Move ah_attr pointed by source to destination.
396  * @dest:       Pointer to destination ah_attr to copy to.
397  *              dest is assumed to be valid or zero'd
398  * @src:        Pointer to the new ah_attr.
399  *
400  * rdma_move_ah_attr() first releases any reference in the destination ah_attr
401  * if it is valid. This also transfers ownership of internal references from
402  * src to dest, making src invalid in the process. No new reference of the src
403  * ah_attr is taken.
404  */
405 void rdma_move_ah_attr(struct rdma_ah_attr *dest, struct rdma_ah_attr *src)
406 {
407 	rdma_destroy_ah_attr(dest);
408 	*dest = *src;
409 	src->grh.sgid_attr = NULL;
410 }
411 EXPORT_SYMBOL(rdma_move_ah_attr);
412 
413 /*
414  * Validate that the rdma_ah_attr is valid for the device before passing it
415  * off to the driver.
416  */
417 static int rdma_check_ah_attr(struct ib_device *device,
418 			      struct rdma_ah_attr *ah_attr)
419 {
420 	if (!rdma_is_port_valid(device, ah_attr->port_num))
421 		return -EINVAL;
422 
423 	if ((rdma_is_grh_required(device, ah_attr->port_num) ||
424 	     ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE) &&
425 	    !(ah_attr->ah_flags & IB_AH_GRH))
426 		return -EINVAL;
427 
428 	if (ah_attr->grh.sgid_attr) {
429 		/*
430 		 * Make sure the passed sgid_attr is consistent with the
431 		 * parameters
432 		 */
433 		if (ah_attr->grh.sgid_attr->index != ah_attr->grh.sgid_index ||
434 		    ah_attr->grh.sgid_attr->port_num != ah_attr->port_num)
435 			return -EINVAL;
436 	}
437 	return 0;
438 }
439 
440 /*
441  * If the ah requires a GRH then ensure that sgid_attr pointer is filled in.
442  * On success the caller is responsible to call rdma_unfill_sgid_attr().
443  */
444 static int rdma_fill_sgid_attr(struct ib_device *device,
445 			       struct rdma_ah_attr *ah_attr,
446 			       const struct ib_gid_attr **old_sgid_attr)
447 {
448 	const struct ib_gid_attr *sgid_attr;
449 	struct ib_global_route *grh;
450 	int ret;
451 
452 	*old_sgid_attr = ah_attr->grh.sgid_attr;
453 
454 	ret = rdma_check_ah_attr(device, ah_attr);
455 	if (ret)
456 		return ret;
457 
458 	if (!(ah_attr->ah_flags & IB_AH_GRH))
459 		return 0;
460 
461 	grh = rdma_ah_retrieve_grh(ah_attr);
462 	if (grh->sgid_attr)
463 		return 0;
464 
465 	sgid_attr =
466 		rdma_get_gid_attr(device, ah_attr->port_num, grh->sgid_index);
467 	if (IS_ERR(sgid_attr))
468 		return PTR_ERR(sgid_attr);
469 
470 	/* Move ownerhip of the kref into the ah_attr */
471 	grh->sgid_attr = sgid_attr;
472 	return 0;
473 }
474 
475 static void rdma_unfill_sgid_attr(struct rdma_ah_attr *ah_attr,
476 				  const struct ib_gid_attr *old_sgid_attr)
477 {
478 	/*
479 	 * Fill didn't change anything, the caller retains ownership of
480 	 * whatever it passed
481 	 */
482 	if (ah_attr->grh.sgid_attr == old_sgid_attr)
483 		return;
484 
485 	/*
486 	 * Otherwise, we need to undo what rdma_fill_sgid_attr so the caller
487 	 * doesn't see any change in the rdma_ah_attr. If we get here
488 	 * old_sgid_attr is NULL.
489 	 */
490 	rdma_destroy_ah_attr(ah_attr);
491 }
492 
493 static const struct ib_gid_attr *
494 rdma_update_sgid_attr(struct rdma_ah_attr *ah_attr,
495 		      const struct ib_gid_attr *old_attr)
496 {
497 	if (old_attr)
498 		rdma_put_gid_attr(old_attr);
499 	if (ah_attr->ah_flags & IB_AH_GRH) {
500 		rdma_hold_gid_attr(ah_attr->grh.sgid_attr);
501 		return ah_attr->grh.sgid_attr;
502 	}
503 	return NULL;
504 }
505 
506 static struct ib_ah *_rdma_create_ah(struct ib_pd *pd,
507 				     struct rdma_ah_attr *ah_attr,
508 				     u32 flags,
509 				     struct ib_udata *udata,
510 				     struct net_device *xmit_slave)
511 {
512 	struct rdma_ah_init_attr init_attr = {};
513 	struct ib_device *device = pd->device;
514 	struct ib_ah *ah;
515 	int ret;
516 
517 	might_sleep_if(flags & RDMA_CREATE_AH_SLEEPABLE);
518 
519 	if (!device->ops.create_ah)
520 		return ERR_PTR(-EOPNOTSUPP);
521 
522 	ah = rdma_zalloc_drv_obj_gfp(
523 		device, ib_ah,
524 		(flags & RDMA_CREATE_AH_SLEEPABLE) ? GFP_KERNEL : GFP_ATOMIC);
525 	if (!ah)
526 		return ERR_PTR(-ENOMEM);
527 
528 	ah->device = device;
529 	ah->pd = pd;
530 	ah->type = ah_attr->type;
531 	ah->sgid_attr = rdma_update_sgid_attr(ah_attr, NULL);
532 	init_attr.ah_attr = ah_attr;
533 	init_attr.flags = flags;
534 	init_attr.xmit_slave = xmit_slave;
535 
536 	ret = device->ops.create_ah(ah, &init_attr, udata);
537 	if (ret) {
538 		kfree(ah);
539 		return ERR_PTR(ret);
540 	}
541 
542 	atomic_inc(&pd->usecnt);
543 	return ah;
544 }
545 
546 /**
547  * rdma_create_ah - Creates an address handle for the
548  * given address vector.
549  * @pd: The protection domain associated with the address handle.
550  * @ah_attr: The attributes of the address vector.
551  * @flags: Create address handle flags (see enum rdma_create_ah_flags).
552  *
553  * It returns 0 on success and returns appropriate error code on error.
554  * The address handle is used to reference a local or global destination
555  * in all UD QP post sends.
556  */
557 struct ib_ah *rdma_create_ah(struct ib_pd *pd, struct rdma_ah_attr *ah_attr,
558 			     u32 flags)
559 {
560 	const struct ib_gid_attr *old_sgid_attr;
561 	struct net_device *slave;
562 	struct ib_ah *ah;
563 	int ret;
564 
565 	ret = rdma_fill_sgid_attr(pd->device, ah_attr, &old_sgid_attr);
566 	if (ret)
567 		return ERR_PTR(ret);
568 	slave = rdma_lag_get_ah_roce_slave(pd->device, ah_attr,
569 					   (flags & RDMA_CREATE_AH_SLEEPABLE) ?
570 					   GFP_KERNEL : GFP_ATOMIC);
571 	if (IS_ERR(slave)) {
572 		rdma_unfill_sgid_attr(ah_attr, old_sgid_attr);
573 		return (void *)slave;
574 	}
575 	ah = _rdma_create_ah(pd, ah_attr, flags, NULL, slave);
576 	rdma_lag_put_ah_roce_slave(slave);
577 	rdma_unfill_sgid_attr(ah_attr, old_sgid_attr);
578 	return ah;
579 }
580 EXPORT_SYMBOL(rdma_create_ah);
581 
582 /**
583  * rdma_create_user_ah - Creates an address handle for the
584  * given address vector.
585  * It resolves destination mac address for ah attribute of RoCE type.
586  * @pd: The protection domain associated with the address handle.
587  * @ah_attr: The attributes of the address vector.
588  * @udata: pointer to user's input output buffer information need by
589  *         provider driver.
590  *
591  * It returns 0 on success and returns appropriate error code on error.
592  * The address handle is used to reference a local or global destination
593  * in all UD QP post sends.
594  */
595 struct ib_ah *rdma_create_user_ah(struct ib_pd *pd,
596 				  struct rdma_ah_attr *ah_attr,
597 				  struct ib_udata *udata)
598 {
599 	const struct ib_gid_attr *old_sgid_attr;
600 	struct ib_ah *ah;
601 	int err;
602 
603 	err = rdma_fill_sgid_attr(pd->device, ah_attr, &old_sgid_attr);
604 	if (err)
605 		return ERR_PTR(err);
606 
607 	if (ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE) {
608 		err = ib_resolve_eth_dmac(pd->device, ah_attr);
609 		if (err) {
610 			ah = ERR_PTR(err);
611 			goto out;
612 		}
613 	}
614 
615 	ah = _rdma_create_ah(pd, ah_attr, RDMA_CREATE_AH_SLEEPABLE,
616 			     udata, NULL);
617 
618 out:
619 	rdma_unfill_sgid_attr(ah_attr, old_sgid_attr);
620 	return ah;
621 }
622 EXPORT_SYMBOL(rdma_create_user_ah);
623 
624 int ib_get_rdma_header_version(const union rdma_network_hdr *hdr)
625 {
626 	const struct iphdr *ip4h = (struct iphdr *)&hdr->roce4grh;
627 	struct iphdr ip4h_checked;
628 	const struct ipv6hdr *ip6h = (struct ipv6hdr *)&hdr->ibgrh;
629 
630 	/* If it's IPv6, the version must be 6, otherwise, the first
631 	 * 20 bytes (before the IPv4 header) are garbled.
632 	 */
633 	if (ip6h->version != 6)
634 		return (ip4h->version == 4) ? 4 : 0;
635 	/* version may be 6 or 4 because the first 20 bytes could be garbled */
636 
637 	/* RoCE v2 requires no options, thus header length
638 	 * must be 5 words
639 	 */
640 	if (ip4h->ihl != 5)
641 		return 6;
642 
643 	/* Verify checksum.
644 	 * We can't write on scattered buffers so we need to copy to
645 	 * temp buffer.
646 	 */
647 	memcpy(&ip4h_checked, ip4h, sizeof(ip4h_checked));
648 	ip4h_checked.check = 0;
649 	ip4h_checked.check = ip_fast_csum((u8 *)&ip4h_checked, 5);
650 	/* if IPv4 header checksum is OK, believe it */
651 	if (ip4h->check == ip4h_checked.check)
652 		return 4;
653 	return 6;
654 }
655 EXPORT_SYMBOL(ib_get_rdma_header_version);
656 
657 static enum rdma_network_type ib_get_net_type_by_grh(struct ib_device *device,
658 						     u8 port_num,
659 						     const struct ib_grh *grh)
660 {
661 	int grh_version;
662 
663 	if (rdma_protocol_ib(device, port_num))
664 		return RDMA_NETWORK_IB;
665 
666 	grh_version = ib_get_rdma_header_version((union rdma_network_hdr *)grh);
667 
668 	if (grh_version == 4)
669 		return RDMA_NETWORK_IPV4;
670 
671 	if (grh->next_hdr == IPPROTO_UDP)
672 		return RDMA_NETWORK_IPV6;
673 
674 	return RDMA_NETWORK_ROCE_V1;
675 }
676 
677 struct find_gid_index_context {
678 	u16 vlan_id;
679 	enum ib_gid_type gid_type;
680 };
681 
682 static bool find_gid_index(const union ib_gid *gid,
683 			   const struct ib_gid_attr *gid_attr,
684 			   void *context)
685 {
686 	struct find_gid_index_context *ctx = context;
687 	u16 vlan_id = 0xffff;
688 	int ret;
689 
690 	if (ctx->gid_type != gid_attr->gid_type)
691 		return false;
692 
693 	ret = rdma_read_gid_l2_fields(gid_attr, &vlan_id, NULL);
694 	if (ret)
695 		return false;
696 
697 	return ctx->vlan_id == vlan_id;
698 }
699 
700 static const struct ib_gid_attr *
701 get_sgid_attr_from_eth(struct ib_device *device, u8 port_num,
702 		       u16 vlan_id, const union ib_gid *sgid,
703 		       enum ib_gid_type gid_type)
704 {
705 	struct find_gid_index_context context = {.vlan_id = vlan_id,
706 						 .gid_type = gid_type};
707 
708 	return rdma_find_gid_by_filter(device, sgid, port_num, find_gid_index,
709 				       &context);
710 }
711 
712 int ib_get_gids_from_rdma_hdr(const union rdma_network_hdr *hdr,
713 			      enum rdma_network_type net_type,
714 			      union ib_gid *sgid, union ib_gid *dgid)
715 {
716 	struct sockaddr_in  src_in;
717 	struct sockaddr_in  dst_in;
718 	__be32 src_saddr, dst_saddr;
719 
720 	if (!sgid || !dgid)
721 		return -EINVAL;
722 
723 	if (net_type == RDMA_NETWORK_IPV4) {
724 		memcpy(&src_in.sin_addr.s_addr,
725 		       &hdr->roce4grh.saddr, 4);
726 		memcpy(&dst_in.sin_addr.s_addr,
727 		       &hdr->roce4grh.daddr, 4);
728 		src_saddr = src_in.sin_addr.s_addr;
729 		dst_saddr = dst_in.sin_addr.s_addr;
730 		ipv6_addr_set_v4mapped(src_saddr,
731 				       (struct in6_addr *)sgid);
732 		ipv6_addr_set_v4mapped(dst_saddr,
733 				       (struct in6_addr *)dgid);
734 		return 0;
735 	} else if (net_type == RDMA_NETWORK_IPV6 ||
736 		   net_type == RDMA_NETWORK_IB || RDMA_NETWORK_ROCE_V1) {
737 		*dgid = hdr->ibgrh.dgid;
738 		*sgid = hdr->ibgrh.sgid;
739 		return 0;
740 	} else {
741 		return -EINVAL;
742 	}
743 }
744 EXPORT_SYMBOL(ib_get_gids_from_rdma_hdr);
745 
746 /* Resolve destination mac address and hop limit for unicast destination
747  * GID entry, considering the source GID entry as well.
748  * ah_attribute must have have valid port_num, sgid_index.
749  */
750 static int ib_resolve_unicast_gid_dmac(struct ib_device *device,
751 				       struct rdma_ah_attr *ah_attr)
752 {
753 	struct ib_global_route *grh = rdma_ah_retrieve_grh(ah_attr);
754 	const struct ib_gid_attr *sgid_attr = grh->sgid_attr;
755 	int hop_limit = 0xff;
756 	int ret = 0;
757 
758 	/* If destination is link local and source GID is RoCEv1,
759 	 * IP stack is not used.
760 	 */
761 	if (rdma_link_local_addr((struct in6_addr *)grh->dgid.raw) &&
762 	    sgid_attr->gid_type == IB_GID_TYPE_ROCE) {
763 		rdma_get_ll_mac((struct in6_addr *)grh->dgid.raw,
764 				ah_attr->roce.dmac);
765 		return ret;
766 	}
767 
768 	ret = rdma_addr_find_l2_eth_by_grh(&sgid_attr->gid, &grh->dgid,
769 					   ah_attr->roce.dmac,
770 					   sgid_attr, &hop_limit);
771 
772 	grh->hop_limit = hop_limit;
773 	return ret;
774 }
775 
776 /*
777  * This function initializes address handle attributes from the incoming packet.
778  * Incoming packet has dgid of the receiver node on which this code is
779  * getting executed and, sgid contains the GID of the sender.
780  *
781  * When resolving mac address of destination, the arrived dgid is used
782  * as sgid and, sgid is used as dgid because sgid contains destinations
783  * GID whom to respond to.
784  *
785  * On success the caller is responsible to call rdma_destroy_ah_attr on the
786  * attr.
787  */
788 int ib_init_ah_attr_from_wc(struct ib_device *device, u8 port_num,
789 			    const struct ib_wc *wc, const struct ib_grh *grh,
790 			    struct rdma_ah_attr *ah_attr)
791 {
792 	u32 flow_class;
793 	int ret;
794 	enum rdma_network_type net_type = RDMA_NETWORK_IB;
795 	enum ib_gid_type gid_type = IB_GID_TYPE_IB;
796 	const struct ib_gid_attr *sgid_attr;
797 	int hoplimit = 0xff;
798 	union ib_gid dgid;
799 	union ib_gid sgid;
800 
801 	might_sleep();
802 
803 	memset(ah_attr, 0, sizeof *ah_attr);
804 	ah_attr->type = rdma_ah_find_type(device, port_num);
805 	if (rdma_cap_eth_ah(device, port_num)) {
806 		if (wc->wc_flags & IB_WC_WITH_NETWORK_HDR_TYPE)
807 			net_type = wc->network_hdr_type;
808 		else
809 			net_type = ib_get_net_type_by_grh(device, port_num, grh);
810 		gid_type = ib_network_to_gid_type(net_type);
811 	}
812 	ret = ib_get_gids_from_rdma_hdr((union rdma_network_hdr *)grh, net_type,
813 					&sgid, &dgid);
814 	if (ret)
815 		return ret;
816 
817 	rdma_ah_set_sl(ah_attr, wc->sl);
818 	rdma_ah_set_port_num(ah_attr, port_num);
819 
820 	if (rdma_protocol_roce(device, port_num)) {
821 		u16 vlan_id = wc->wc_flags & IB_WC_WITH_VLAN ?
822 				wc->vlan_id : 0xffff;
823 
824 		if (!(wc->wc_flags & IB_WC_GRH))
825 			return -EPROTOTYPE;
826 
827 		sgid_attr = get_sgid_attr_from_eth(device, port_num,
828 						   vlan_id, &dgid,
829 						   gid_type);
830 		if (IS_ERR(sgid_attr))
831 			return PTR_ERR(sgid_attr);
832 
833 		flow_class = be32_to_cpu(grh->version_tclass_flow);
834 		rdma_move_grh_sgid_attr(ah_attr,
835 					&sgid,
836 					flow_class & 0xFFFFF,
837 					hoplimit,
838 					(flow_class >> 20) & 0xFF,
839 					sgid_attr);
840 
841 		ret = ib_resolve_unicast_gid_dmac(device, ah_attr);
842 		if (ret)
843 			rdma_destroy_ah_attr(ah_attr);
844 
845 		return ret;
846 	} else {
847 		rdma_ah_set_dlid(ah_attr, wc->slid);
848 		rdma_ah_set_path_bits(ah_attr, wc->dlid_path_bits);
849 
850 		if ((wc->wc_flags & IB_WC_GRH) == 0)
851 			return 0;
852 
853 		if (dgid.global.interface_id !=
854 					cpu_to_be64(IB_SA_WELL_KNOWN_GUID)) {
855 			sgid_attr = rdma_find_gid_by_port(
856 				device, &dgid, IB_GID_TYPE_IB, port_num, NULL);
857 		} else
858 			sgid_attr = rdma_get_gid_attr(device, port_num, 0);
859 
860 		if (IS_ERR(sgid_attr))
861 			return PTR_ERR(sgid_attr);
862 		flow_class = be32_to_cpu(grh->version_tclass_flow);
863 		rdma_move_grh_sgid_attr(ah_attr,
864 					&sgid,
865 					flow_class & 0xFFFFF,
866 					hoplimit,
867 					(flow_class >> 20) & 0xFF,
868 					sgid_attr);
869 
870 		return 0;
871 	}
872 }
873 EXPORT_SYMBOL(ib_init_ah_attr_from_wc);
874 
875 /**
876  * rdma_move_grh_sgid_attr - Sets the sgid attribute of GRH, taking ownership
877  * of the reference
878  *
879  * @attr:	Pointer to AH attribute structure
880  * @dgid:	Destination GID
881  * @flow_label:	Flow label
882  * @hop_limit:	Hop limit
883  * @traffic_class: traffic class
884  * @sgid_attr:	Pointer to SGID attribute
885  *
886  * This takes ownership of the sgid_attr reference. The caller must ensure
887  * rdma_destroy_ah_attr() is called before destroying the rdma_ah_attr after
888  * calling this function.
889  */
890 void rdma_move_grh_sgid_attr(struct rdma_ah_attr *attr, union ib_gid *dgid,
891 			     u32 flow_label, u8 hop_limit, u8 traffic_class,
892 			     const struct ib_gid_attr *sgid_attr)
893 {
894 	rdma_ah_set_grh(attr, dgid, flow_label, sgid_attr->index, hop_limit,
895 			traffic_class);
896 	attr->grh.sgid_attr = sgid_attr;
897 }
898 EXPORT_SYMBOL(rdma_move_grh_sgid_attr);
899 
900 /**
901  * rdma_destroy_ah_attr - Release reference to SGID attribute of
902  * ah attribute.
903  * @ah_attr: Pointer to ah attribute
904  *
905  * Release reference to the SGID attribute of the ah attribute if it is
906  * non NULL. It is safe to call this multiple times, and safe to call it on
907  * a zero initialized ah_attr.
908  */
909 void rdma_destroy_ah_attr(struct rdma_ah_attr *ah_attr)
910 {
911 	if (ah_attr->grh.sgid_attr) {
912 		rdma_put_gid_attr(ah_attr->grh.sgid_attr);
913 		ah_attr->grh.sgid_attr = NULL;
914 	}
915 }
916 EXPORT_SYMBOL(rdma_destroy_ah_attr);
917 
918 struct ib_ah *ib_create_ah_from_wc(struct ib_pd *pd, const struct ib_wc *wc,
919 				   const struct ib_grh *grh, u8 port_num)
920 {
921 	struct rdma_ah_attr ah_attr;
922 	struct ib_ah *ah;
923 	int ret;
924 
925 	ret = ib_init_ah_attr_from_wc(pd->device, port_num, wc, grh, &ah_attr);
926 	if (ret)
927 		return ERR_PTR(ret);
928 
929 	ah = rdma_create_ah(pd, &ah_attr, RDMA_CREATE_AH_SLEEPABLE);
930 
931 	rdma_destroy_ah_attr(&ah_attr);
932 	return ah;
933 }
934 EXPORT_SYMBOL(ib_create_ah_from_wc);
935 
936 int rdma_modify_ah(struct ib_ah *ah, struct rdma_ah_attr *ah_attr)
937 {
938 	const struct ib_gid_attr *old_sgid_attr;
939 	int ret;
940 
941 	if (ah->type != ah_attr->type)
942 		return -EINVAL;
943 
944 	ret = rdma_fill_sgid_attr(ah->device, ah_attr, &old_sgid_attr);
945 	if (ret)
946 		return ret;
947 
948 	ret = ah->device->ops.modify_ah ?
949 		ah->device->ops.modify_ah(ah, ah_attr) :
950 		-EOPNOTSUPP;
951 
952 	ah->sgid_attr = rdma_update_sgid_attr(ah_attr, ah->sgid_attr);
953 	rdma_unfill_sgid_attr(ah_attr, old_sgid_attr);
954 	return ret;
955 }
956 EXPORT_SYMBOL(rdma_modify_ah);
957 
958 int rdma_query_ah(struct ib_ah *ah, struct rdma_ah_attr *ah_attr)
959 {
960 	ah_attr->grh.sgid_attr = NULL;
961 
962 	return ah->device->ops.query_ah ?
963 		ah->device->ops.query_ah(ah, ah_attr) :
964 		-EOPNOTSUPP;
965 }
966 EXPORT_SYMBOL(rdma_query_ah);
967 
968 int rdma_destroy_ah_user(struct ib_ah *ah, u32 flags, struct ib_udata *udata)
969 {
970 	const struct ib_gid_attr *sgid_attr = ah->sgid_attr;
971 	struct ib_pd *pd;
972 	int ret;
973 
974 	might_sleep_if(flags & RDMA_DESTROY_AH_SLEEPABLE);
975 
976 	pd = ah->pd;
977 
978 	ret = ah->device->ops.destroy_ah(ah, flags);
979 	if (ret)
980 		return ret;
981 
982 	atomic_dec(&pd->usecnt);
983 	if (sgid_attr)
984 		rdma_put_gid_attr(sgid_attr);
985 
986 	kfree(ah);
987 	return ret;
988 }
989 EXPORT_SYMBOL(rdma_destroy_ah_user);
990 
991 /* Shared receive queues */
992 
993 /**
994  * ib_create_srq_user - Creates a SRQ associated with the specified protection
995  *   domain.
996  * @pd: The protection domain associated with the SRQ.
997  * @srq_init_attr: A list of initial attributes required to create the
998  *   SRQ.  If SRQ creation succeeds, then the attributes are updated to
999  *   the actual capabilities of the created SRQ.
1000  * @uobject: uobject pointer if this is not a kernel SRQ
1001  * @udata: udata pointer if this is not a kernel SRQ
1002  *
1003  * srq_attr->max_wr and srq_attr->max_sge are read the determine the
1004  * requested size of the SRQ, and set to the actual values allocated
1005  * on return.  If ib_create_srq() succeeds, then max_wr and max_sge
1006  * will always be at least as large as the requested values.
1007  */
1008 struct ib_srq *ib_create_srq_user(struct ib_pd *pd,
1009 				  struct ib_srq_init_attr *srq_init_attr,
1010 				  struct ib_usrq_object *uobject,
1011 				  struct ib_udata *udata)
1012 {
1013 	struct ib_srq *srq;
1014 	int ret;
1015 
1016 	srq = rdma_zalloc_drv_obj(pd->device, ib_srq);
1017 	if (!srq)
1018 		return ERR_PTR(-ENOMEM);
1019 
1020 	srq->device = pd->device;
1021 	srq->pd = pd;
1022 	srq->event_handler = srq_init_attr->event_handler;
1023 	srq->srq_context = srq_init_attr->srq_context;
1024 	srq->srq_type = srq_init_attr->srq_type;
1025 	srq->uobject = uobject;
1026 
1027 	if (ib_srq_has_cq(srq->srq_type)) {
1028 		srq->ext.cq = srq_init_attr->ext.cq;
1029 		atomic_inc(&srq->ext.cq->usecnt);
1030 	}
1031 	if (srq->srq_type == IB_SRQT_XRC) {
1032 		srq->ext.xrc.xrcd = srq_init_attr->ext.xrc.xrcd;
1033 		atomic_inc(&srq->ext.xrc.xrcd->usecnt);
1034 	}
1035 	atomic_inc(&pd->usecnt);
1036 
1037 	ret = pd->device->ops.create_srq(srq, srq_init_attr, udata);
1038 	if (ret) {
1039 		atomic_dec(&srq->pd->usecnt);
1040 		if (srq->srq_type == IB_SRQT_XRC)
1041 			atomic_dec(&srq->ext.xrc.xrcd->usecnt);
1042 		if (ib_srq_has_cq(srq->srq_type))
1043 			atomic_dec(&srq->ext.cq->usecnt);
1044 		kfree(srq);
1045 		return ERR_PTR(ret);
1046 	}
1047 
1048 	return srq;
1049 }
1050 EXPORT_SYMBOL(ib_create_srq_user);
1051 
1052 int ib_modify_srq(struct ib_srq *srq,
1053 		  struct ib_srq_attr *srq_attr,
1054 		  enum ib_srq_attr_mask srq_attr_mask)
1055 {
1056 	return srq->device->ops.modify_srq ?
1057 		srq->device->ops.modify_srq(srq, srq_attr, srq_attr_mask,
1058 					    NULL) : -EOPNOTSUPP;
1059 }
1060 EXPORT_SYMBOL(ib_modify_srq);
1061 
1062 int ib_query_srq(struct ib_srq *srq,
1063 		 struct ib_srq_attr *srq_attr)
1064 {
1065 	return srq->device->ops.query_srq ?
1066 		srq->device->ops.query_srq(srq, srq_attr) : -EOPNOTSUPP;
1067 }
1068 EXPORT_SYMBOL(ib_query_srq);
1069 
1070 int ib_destroy_srq_user(struct ib_srq *srq, struct ib_udata *udata)
1071 {
1072 	int ret;
1073 
1074 	if (atomic_read(&srq->usecnt))
1075 		return -EBUSY;
1076 
1077 	ret = srq->device->ops.destroy_srq(srq, udata);
1078 	if (ret)
1079 		return ret;
1080 
1081 	atomic_dec(&srq->pd->usecnt);
1082 	if (srq->srq_type == IB_SRQT_XRC)
1083 		atomic_dec(&srq->ext.xrc.xrcd->usecnt);
1084 	if (ib_srq_has_cq(srq->srq_type))
1085 		atomic_dec(&srq->ext.cq->usecnt);
1086 	kfree(srq);
1087 
1088 	return ret;
1089 }
1090 EXPORT_SYMBOL(ib_destroy_srq_user);
1091 
1092 /* Queue pairs */
1093 
1094 static void __ib_shared_qp_event_handler(struct ib_event *event, void *context)
1095 {
1096 	struct ib_qp *qp = context;
1097 	unsigned long flags;
1098 
1099 	spin_lock_irqsave(&qp->device->qp_open_list_lock, flags);
1100 	list_for_each_entry(event->element.qp, &qp->open_list, open_list)
1101 		if (event->element.qp->event_handler)
1102 			event->element.qp->event_handler(event, event->element.qp->qp_context);
1103 	spin_unlock_irqrestore(&qp->device->qp_open_list_lock, flags);
1104 }
1105 
1106 static struct ib_qp *__ib_open_qp(struct ib_qp *real_qp,
1107 				  void (*event_handler)(struct ib_event *, void *),
1108 				  void *qp_context)
1109 {
1110 	struct ib_qp *qp;
1111 	unsigned long flags;
1112 	int err;
1113 
1114 	qp = kzalloc(sizeof *qp, GFP_KERNEL);
1115 	if (!qp)
1116 		return ERR_PTR(-ENOMEM);
1117 
1118 	qp->real_qp = real_qp;
1119 	err = ib_open_shared_qp_security(qp, real_qp->device);
1120 	if (err) {
1121 		kfree(qp);
1122 		return ERR_PTR(err);
1123 	}
1124 
1125 	qp->real_qp = real_qp;
1126 	atomic_inc(&real_qp->usecnt);
1127 	qp->device = real_qp->device;
1128 	qp->event_handler = event_handler;
1129 	qp->qp_context = qp_context;
1130 	qp->qp_num = real_qp->qp_num;
1131 	qp->qp_type = real_qp->qp_type;
1132 
1133 	spin_lock_irqsave(&real_qp->device->qp_open_list_lock, flags);
1134 	list_add(&qp->open_list, &real_qp->open_list);
1135 	spin_unlock_irqrestore(&real_qp->device->qp_open_list_lock, flags);
1136 
1137 	return qp;
1138 }
1139 
1140 struct ib_qp *ib_open_qp(struct ib_xrcd *xrcd,
1141 			 struct ib_qp_open_attr *qp_open_attr)
1142 {
1143 	struct ib_qp *qp, *real_qp;
1144 
1145 	if (qp_open_attr->qp_type != IB_QPT_XRC_TGT)
1146 		return ERR_PTR(-EINVAL);
1147 
1148 	down_read(&xrcd->tgt_qps_rwsem);
1149 	real_qp = xa_load(&xrcd->tgt_qps, qp_open_attr->qp_num);
1150 	if (!real_qp) {
1151 		up_read(&xrcd->tgt_qps_rwsem);
1152 		return ERR_PTR(-EINVAL);
1153 	}
1154 	qp = __ib_open_qp(real_qp, qp_open_attr->event_handler,
1155 			  qp_open_attr->qp_context);
1156 	up_read(&xrcd->tgt_qps_rwsem);
1157 	return qp;
1158 }
1159 EXPORT_SYMBOL(ib_open_qp);
1160 
1161 static struct ib_qp *create_xrc_qp_user(struct ib_qp *qp,
1162 					struct ib_qp_init_attr *qp_init_attr)
1163 {
1164 	struct ib_qp *real_qp = qp;
1165 	int err;
1166 
1167 	qp->event_handler = __ib_shared_qp_event_handler;
1168 	qp->qp_context = qp;
1169 	qp->pd = NULL;
1170 	qp->send_cq = qp->recv_cq = NULL;
1171 	qp->srq = NULL;
1172 	qp->xrcd = qp_init_attr->xrcd;
1173 	atomic_inc(&qp_init_attr->xrcd->usecnt);
1174 	INIT_LIST_HEAD(&qp->open_list);
1175 
1176 	qp = __ib_open_qp(real_qp, qp_init_attr->event_handler,
1177 			  qp_init_attr->qp_context);
1178 	if (IS_ERR(qp))
1179 		return qp;
1180 
1181 	err = xa_err(xa_store(&qp_init_attr->xrcd->tgt_qps, real_qp->qp_num,
1182 			      real_qp, GFP_KERNEL));
1183 	if (err) {
1184 		ib_close_qp(qp);
1185 		return ERR_PTR(err);
1186 	}
1187 	return qp;
1188 }
1189 
1190 /**
1191  * ib_create_qp - Creates a kernel QP associated with the specified protection
1192  *   domain.
1193  * @pd: The protection domain associated with the QP.
1194  * @qp_init_attr: A list of initial attributes required to create the
1195  *   QP.  If QP creation succeeds, then the attributes are updated to
1196  *   the actual capabilities of the created QP.
1197  *
1198  * NOTE: for user qp use ib_create_qp_user with valid udata!
1199  */
1200 struct ib_qp *ib_create_qp(struct ib_pd *pd,
1201 			   struct ib_qp_init_attr *qp_init_attr)
1202 {
1203 	struct ib_device *device = pd ? pd->device : qp_init_attr->xrcd->device;
1204 	struct ib_qp *qp;
1205 	int ret;
1206 
1207 	if (qp_init_attr->rwq_ind_tbl &&
1208 	    (qp_init_attr->recv_cq ||
1209 	    qp_init_attr->srq || qp_init_attr->cap.max_recv_wr ||
1210 	    qp_init_attr->cap.max_recv_sge))
1211 		return ERR_PTR(-EINVAL);
1212 
1213 	if ((qp_init_attr->create_flags & IB_QP_CREATE_INTEGRITY_EN) &&
1214 	    !(device->attrs.device_cap_flags & IB_DEVICE_INTEGRITY_HANDOVER))
1215 		return ERR_PTR(-EINVAL);
1216 
1217 	/*
1218 	 * If the callers is using the RDMA API calculate the resources
1219 	 * needed for the RDMA READ/WRITE operations.
1220 	 *
1221 	 * Note that these callers need to pass in a port number.
1222 	 */
1223 	if (qp_init_attr->cap.max_rdma_ctxs)
1224 		rdma_rw_init_qp(device, qp_init_attr);
1225 
1226 	qp = _ib_create_qp(device, pd, qp_init_attr, NULL, NULL);
1227 	if (IS_ERR(qp))
1228 		return qp;
1229 
1230 	ret = ib_create_qp_security(qp, device);
1231 	if (ret)
1232 		goto err;
1233 
1234 	if (qp_init_attr->qp_type == IB_QPT_XRC_TGT) {
1235 		struct ib_qp *xrc_qp =
1236 			create_xrc_qp_user(qp, qp_init_attr);
1237 
1238 		if (IS_ERR(xrc_qp)) {
1239 			ret = PTR_ERR(xrc_qp);
1240 			goto err;
1241 		}
1242 		return xrc_qp;
1243 	}
1244 
1245 	qp->event_handler = qp_init_attr->event_handler;
1246 	qp->qp_context = qp_init_attr->qp_context;
1247 	if (qp_init_attr->qp_type == IB_QPT_XRC_INI) {
1248 		qp->recv_cq = NULL;
1249 		qp->srq = NULL;
1250 	} else {
1251 		qp->recv_cq = qp_init_attr->recv_cq;
1252 		if (qp_init_attr->recv_cq)
1253 			atomic_inc(&qp_init_attr->recv_cq->usecnt);
1254 		qp->srq = qp_init_attr->srq;
1255 		if (qp->srq)
1256 			atomic_inc(&qp_init_attr->srq->usecnt);
1257 	}
1258 
1259 	qp->send_cq = qp_init_attr->send_cq;
1260 	qp->xrcd    = NULL;
1261 
1262 	atomic_inc(&pd->usecnt);
1263 	if (qp_init_attr->send_cq)
1264 		atomic_inc(&qp_init_attr->send_cq->usecnt);
1265 	if (qp_init_attr->rwq_ind_tbl)
1266 		atomic_inc(&qp->rwq_ind_tbl->usecnt);
1267 
1268 	if (qp_init_attr->cap.max_rdma_ctxs) {
1269 		ret = rdma_rw_init_mrs(qp, qp_init_attr);
1270 		if (ret)
1271 			goto err;
1272 	}
1273 
1274 	/*
1275 	 * Note: all hw drivers guarantee that max_send_sge is lower than
1276 	 * the device RDMA WRITE SGE limit but not all hw drivers ensure that
1277 	 * max_send_sge <= max_sge_rd.
1278 	 */
1279 	qp->max_write_sge = qp_init_attr->cap.max_send_sge;
1280 	qp->max_read_sge = min_t(u32, qp_init_attr->cap.max_send_sge,
1281 				 device->attrs.max_sge_rd);
1282 	if (qp_init_attr->create_flags & IB_QP_CREATE_INTEGRITY_EN)
1283 		qp->integrity_en = true;
1284 
1285 	return qp;
1286 
1287 err:
1288 	ib_destroy_qp(qp);
1289 	return ERR_PTR(ret);
1290 
1291 }
1292 EXPORT_SYMBOL(ib_create_qp);
1293 
1294 static const struct {
1295 	int			valid;
1296 	enum ib_qp_attr_mask	req_param[IB_QPT_MAX];
1297 	enum ib_qp_attr_mask	opt_param[IB_QPT_MAX];
1298 } qp_state_table[IB_QPS_ERR + 1][IB_QPS_ERR + 1] = {
1299 	[IB_QPS_RESET] = {
1300 		[IB_QPS_RESET] = { .valid = 1 },
1301 		[IB_QPS_INIT]  = {
1302 			.valid = 1,
1303 			.req_param = {
1304 				[IB_QPT_UD]  = (IB_QP_PKEY_INDEX		|
1305 						IB_QP_PORT			|
1306 						IB_QP_QKEY),
1307 				[IB_QPT_RAW_PACKET] = IB_QP_PORT,
1308 				[IB_QPT_UC]  = (IB_QP_PKEY_INDEX		|
1309 						IB_QP_PORT			|
1310 						IB_QP_ACCESS_FLAGS),
1311 				[IB_QPT_RC]  = (IB_QP_PKEY_INDEX		|
1312 						IB_QP_PORT			|
1313 						IB_QP_ACCESS_FLAGS),
1314 				[IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX		|
1315 						IB_QP_PORT			|
1316 						IB_QP_ACCESS_FLAGS),
1317 				[IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX		|
1318 						IB_QP_PORT			|
1319 						IB_QP_ACCESS_FLAGS),
1320 				[IB_QPT_SMI] = (IB_QP_PKEY_INDEX		|
1321 						IB_QP_QKEY),
1322 				[IB_QPT_GSI] = (IB_QP_PKEY_INDEX		|
1323 						IB_QP_QKEY),
1324 			}
1325 		},
1326 	},
1327 	[IB_QPS_INIT]  = {
1328 		[IB_QPS_RESET] = { .valid = 1 },
1329 		[IB_QPS_ERR] =   { .valid = 1 },
1330 		[IB_QPS_INIT]  = {
1331 			.valid = 1,
1332 			.opt_param = {
1333 				[IB_QPT_UD]  = (IB_QP_PKEY_INDEX		|
1334 						IB_QP_PORT			|
1335 						IB_QP_QKEY),
1336 				[IB_QPT_UC]  = (IB_QP_PKEY_INDEX		|
1337 						IB_QP_PORT			|
1338 						IB_QP_ACCESS_FLAGS),
1339 				[IB_QPT_RC]  = (IB_QP_PKEY_INDEX		|
1340 						IB_QP_PORT			|
1341 						IB_QP_ACCESS_FLAGS),
1342 				[IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX		|
1343 						IB_QP_PORT			|
1344 						IB_QP_ACCESS_FLAGS),
1345 				[IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX		|
1346 						IB_QP_PORT			|
1347 						IB_QP_ACCESS_FLAGS),
1348 				[IB_QPT_SMI] = (IB_QP_PKEY_INDEX		|
1349 						IB_QP_QKEY),
1350 				[IB_QPT_GSI] = (IB_QP_PKEY_INDEX		|
1351 						IB_QP_QKEY),
1352 			}
1353 		},
1354 		[IB_QPS_RTR]   = {
1355 			.valid = 1,
1356 			.req_param = {
1357 				[IB_QPT_UC]  = (IB_QP_AV			|
1358 						IB_QP_PATH_MTU			|
1359 						IB_QP_DEST_QPN			|
1360 						IB_QP_RQ_PSN),
1361 				[IB_QPT_RC]  = (IB_QP_AV			|
1362 						IB_QP_PATH_MTU			|
1363 						IB_QP_DEST_QPN			|
1364 						IB_QP_RQ_PSN			|
1365 						IB_QP_MAX_DEST_RD_ATOMIC	|
1366 						IB_QP_MIN_RNR_TIMER),
1367 				[IB_QPT_XRC_INI] = (IB_QP_AV			|
1368 						IB_QP_PATH_MTU			|
1369 						IB_QP_DEST_QPN			|
1370 						IB_QP_RQ_PSN),
1371 				[IB_QPT_XRC_TGT] = (IB_QP_AV			|
1372 						IB_QP_PATH_MTU			|
1373 						IB_QP_DEST_QPN			|
1374 						IB_QP_RQ_PSN			|
1375 						IB_QP_MAX_DEST_RD_ATOMIC	|
1376 						IB_QP_MIN_RNR_TIMER),
1377 			},
1378 			.opt_param = {
1379 				 [IB_QPT_UD]  = (IB_QP_PKEY_INDEX		|
1380 						 IB_QP_QKEY),
1381 				 [IB_QPT_UC]  = (IB_QP_ALT_PATH			|
1382 						 IB_QP_ACCESS_FLAGS		|
1383 						 IB_QP_PKEY_INDEX),
1384 				 [IB_QPT_RC]  = (IB_QP_ALT_PATH			|
1385 						 IB_QP_ACCESS_FLAGS		|
1386 						 IB_QP_PKEY_INDEX),
1387 				 [IB_QPT_XRC_INI] = (IB_QP_ALT_PATH		|
1388 						 IB_QP_ACCESS_FLAGS		|
1389 						 IB_QP_PKEY_INDEX),
1390 				 [IB_QPT_XRC_TGT] = (IB_QP_ALT_PATH		|
1391 						 IB_QP_ACCESS_FLAGS		|
1392 						 IB_QP_PKEY_INDEX),
1393 				 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX		|
1394 						 IB_QP_QKEY),
1395 				 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX		|
1396 						 IB_QP_QKEY),
1397 			 },
1398 		},
1399 	},
1400 	[IB_QPS_RTR]   = {
1401 		[IB_QPS_RESET] = { .valid = 1 },
1402 		[IB_QPS_ERR] =   { .valid = 1 },
1403 		[IB_QPS_RTS]   = {
1404 			.valid = 1,
1405 			.req_param = {
1406 				[IB_QPT_UD]  = IB_QP_SQ_PSN,
1407 				[IB_QPT_UC]  = IB_QP_SQ_PSN,
1408 				[IB_QPT_RC]  = (IB_QP_TIMEOUT			|
1409 						IB_QP_RETRY_CNT			|
1410 						IB_QP_RNR_RETRY			|
1411 						IB_QP_SQ_PSN			|
1412 						IB_QP_MAX_QP_RD_ATOMIC),
1413 				[IB_QPT_XRC_INI] = (IB_QP_TIMEOUT		|
1414 						IB_QP_RETRY_CNT			|
1415 						IB_QP_RNR_RETRY			|
1416 						IB_QP_SQ_PSN			|
1417 						IB_QP_MAX_QP_RD_ATOMIC),
1418 				[IB_QPT_XRC_TGT] = (IB_QP_TIMEOUT		|
1419 						IB_QP_SQ_PSN),
1420 				[IB_QPT_SMI] = IB_QP_SQ_PSN,
1421 				[IB_QPT_GSI] = IB_QP_SQ_PSN,
1422 			},
1423 			.opt_param = {
1424 				 [IB_QPT_UD]  = (IB_QP_CUR_STATE		|
1425 						 IB_QP_QKEY),
1426 				 [IB_QPT_UC]  = (IB_QP_CUR_STATE		|
1427 						 IB_QP_ALT_PATH			|
1428 						 IB_QP_ACCESS_FLAGS		|
1429 						 IB_QP_PATH_MIG_STATE),
1430 				 [IB_QPT_RC]  = (IB_QP_CUR_STATE		|
1431 						 IB_QP_ALT_PATH			|
1432 						 IB_QP_ACCESS_FLAGS		|
1433 						 IB_QP_MIN_RNR_TIMER		|
1434 						 IB_QP_PATH_MIG_STATE),
1435 				 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE		|
1436 						 IB_QP_ALT_PATH			|
1437 						 IB_QP_ACCESS_FLAGS		|
1438 						 IB_QP_PATH_MIG_STATE),
1439 				 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE		|
1440 						 IB_QP_ALT_PATH			|
1441 						 IB_QP_ACCESS_FLAGS		|
1442 						 IB_QP_MIN_RNR_TIMER		|
1443 						 IB_QP_PATH_MIG_STATE),
1444 				 [IB_QPT_SMI] = (IB_QP_CUR_STATE		|
1445 						 IB_QP_QKEY),
1446 				 [IB_QPT_GSI] = (IB_QP_CUR_STATE		|
1447 						 IB_QP_QKEY),
1448 				 [IB_QPT_RAW_PACKET] = IB_QP_RATE_LIMIT,
1449 			 }
1450 		}
1451 	},
1452 	[IB_QPS_RTS]   = {
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_ACCESS_FLAGS		|
1462 						IB_QP_ALT_PATH			|
1463 						IB_QP_PATH_MIG_STATE),
1464 				[IB_QPT_RC]  = (IB_QP_CUR_STATE			|
1465 						IB_QP_ACCESS_FLAGS		|
1466 						IB_QP_ALT_PATH			|
1467 						IB_QP_PATH_MIG_STATE		|
1468 						IB_QP_MIN_RNR_TIMER),
1469 				[IB_QPT_XRC_INI] = (IB_QP_CUR_STATE		|
1470 						IB_QP_ACCESS_FLAGS		|
1471 						IB_QP_ALT_PATH			|
1472 						IB_QP_PATH_MIG_STATE),
1473 				[IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE		|
1474 						IB_QP_ACCESS_FLAGS		|
1475 						IB_QP_ALT_PATH			|
1476 						IB_QP_PATH_MIG_STATE		|
1477 						IB_QP_MIN_RNR_TIMER),
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 				[IB_QPT_RAW_PACKET] = IB_QP_RATE_LIMIT,
1483 			}
1484 		},
1485 		[IB_QPS_SQD]   = {
1486 			.valid = 1,
1487 			.opt_param = {
1488 				[IB_QPT_UD]  = IB_QP_EN_SQD_ASYNC_NOTIFY,
1489 				[IB_QPT_UC]  = IB_QP_EN_SQD_ASYNC_NOTIFY,
1490 				[IB_QPT_RC]  = IB_QP_EN_SQD_ASYNC_NOTIFY,
1491 				[IB_QPT_XRC_INI] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1492 				[IB_QPT_XRC_TGT] = IB_QP_EN_SQD_ASYNC_NOTIFY, /* ??? */
1493 				[IB_QPT_SMI] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1494 				[IB_QPT_GSI] = IB_QP_EN_SQD_ASYNC_NOTIFY
1495 			}
1496 		},
1497 	},
1498 	[IB_QPS_SQD]   = {
1499 		[IB_QPS_RESET] = { .valid = 1 },
1500 		[IB_QPS_ERR] =   { .valid = 1 },
1501 		[IB_QPS_RTS]   = {
1502 			.valid = 1,
1503 			.opt_param = {
1504 				[IB_QPT_UD]  = (IB_QP_CUR_STATE			|
1505 						IB_QP_QKEY),
1506 				[IB_QPT_UC]  = (IB_QP_CUR_STATE			|
1507 						IB_QP_ALT_PATH			|
1508 						IB_QP_ACCESS_FLAGS		|
1509 						IB_QP_PATH_MIG_STATE),
1510 				[IB_QPT_RC]  = (IB_QP_CUR_STATE			|
1511 						IB_QP_ALT_PATH			|
1512 						IB_QP_ACCESS_FLAGS		|
1513 						IB_QP_MIN_RNR_TIMER		|
1514 						IB_QP_PATH_MIG_STATE),
1515 				[IB_QPT_XRC_INI] = (IB_QP_CUR_STATE		|
1516 						IB_QP_ALT_PATH			|
1517 						IB_QP_ACCESS_FLAGS		|
1518 						IB_QP_PATH_MIG_STATE),
1519 				[IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE		|
1520 						IB_QP_ALT_PATH			|
1521 						IB_QP_ACCESS_FLAGS		|
1522 						IB_QP_MIN_RNR_TIMER		|
1523 						IB_QP_PATH_MIG_STATE),
1524 				[IB_QPT_SMI] = (IB_QP_CUR_STATE			|
1525 						IB_QP_QKEY),
1526 				[IB_QPT_GSI] = (IB_QP_CUR_STATE			|
1527 						IB_QP_QKEY),
1528 			}
1529 		},
1530 		[IB_QPS_SQD]   = {
1531 			.valid = 1,
1532 			.opt_param = {
1533 				[IB_QPT_UD]  = (IB_QP_PKEY_INDEX		|
1534 						IB_QP_QKEY),
1535 				[IB_QPT_UC]  = (IB_QP_AV			|
1536 						IB_QP_ALT_PATH			|
1537 						IB_QP_ACCESS_FLAGS		|
1538 						IB_QP_PKEY_INDEX		|
1539 						IB_QP_PATH_MIG_STATE),
1540 				[IB_QPT_RC]  = (IB_QP_PORT			|
1541 						IB_QP_AV			|
1542 						IB_QP_TIMEOUT			|
1543 						IB_QP_RETRY_CNT			|
1544 						IB_QP_RNR_RETRY			|
1545 						IB_QP_MAX_QP_RD_ATOMIC		|
1546 						IB_QP_MAX_DEST_RD_ATOMIC	|
1547 						IB_QP_ALT_PATH			|
1548 						IB_QP_ACCESS_FLAGS		|
1549 						IB_QP_PKEY_INDEX		|
1550 						IB_QP_MIN_RNR_TIMER		|
1551 						IB_QP_PATH_MIG_STATE),
1552 				[IB_QPT_XRC_INI] = (IB_QP_PORT			|
1553 						IB_QP_AV			|
1554 						IB_QP_TIMEOUT			|
1555 						IB_QP_RETRY_CNT			|
1556 						IB_QP_RNR_RETRY			|
1557 						IB_QP_MAX_QP_RD_ATOMIC		|
1558 						IB_QP_ALT_PATH			|
1559 						IB_QP_ACCESS_FLAGS		|
1560 						IB_QP_PKEY_INDEX		|
1561 						IB_QP_PATH_MIG_STATE),
1562 				[IB_QPT_XRC_TGT] = (IB_QP_PORT			|
1563 						IB_QP_AV			|
1564 						IB_QP_TIMEOUT			|
1565 						IB_QP_MAX_DEST_RD_ATOMIC	|
1566 						IB_QP_ALT_PATH			|
1567 						IB_QP_ACCESS_FLAGS		|
1568 						IB_QP_PKEY_INDEX		|
1569 						IB_QP_MIN_RNR_TIMER		|
1570 						IB_QP_PATH_MIG_STATE),
1571 				[IB_QPT_SMI] = (IB_QP_PKEY_INDEX		|
1572 						IB_QP_QKEY),
1573 				[IB_QPT_GSI] = (IB_QP_PKEY_INDEX		|
1574 						IB_QP_QKEY),
1575 			}
1576 		}
1577 	},
1578 	[IB_QPS_SQE]   = {
1579 		[IB_QPS_RESET] = { .valid = 1 },
1580 		[IB_QPS_ERR] =   { .valid = 1 },
1581 		[IB_QPS_RTS]   = {
1582 			.valid = 1,
1583 			.opt_param = {
1584 				[IB_QPT_UD]  = (IB_QP_CUR_STATE			|
1585 						IB_QP_QKEY),
1586 				[IB_QPT_UC]  = (IB_QP_CUR_STATE			|
1587 						IB_QP_ACCESS_FLAGS),
1588 				[IB_QPT_SMI] = (IB_QP_CUR_STATE			|
1589 						IB_QP_QKEY),
1590 				[IB_QPT_GSI] = (IB_QP_CUR_STATE			|
1591 						IB_QP_QKEY),
1592 			}
1593 		}
1594 	},
1595 	[IB_QPS_ERR] = {
1596 		[IB_QPS_RESET] = { .valid = 1 },
1597 		[IB_QPS_ERR] =   { .valid = 1 }
1598 	}
1599 };
1600 
1601 bool ib_modify_qp_is_ok(enum ib_qp_state cur_state, enum ib_qp_state next_state,
1602 			enum ib_qp_type type, enum ib_qp_attr_mask mask)
1603 {
1604 	enum ib_qp_attr_mask req_param, opt_param;
1605 
1606 	if (mask & IB_QP_CUR_STATE  &&
1607 	    cur_state != IB_QPS_RTR && cur_state != IB_QPS_RTS &&
1608 	    cur_state != IB_QPS_SQD && cur_state != IB_QPS_SQE)
1609 		return false;
1610 
1611 	if (!qp_state_table[cur_state][next_state].valid)
1612 		return false;
1613 
1614 	req_param = qp_state_table[cur_state][next_state].req_param[type];
1615 	opt_param = qp_state_table[cur_state][next_state].opt_param[type];
1616 
1617 	if ((mask & req_param) != req_param)
1618 		return false;
1619 
1620 	if (mask & ~(req_param | opt_param | IB_QP_STATE))
1621 		return false;
1622 
1623 	return true;
1624 }
1625 EXPORT_SYMBOL(ib_modify_qp_is_ok);
1626 
1627 /**
1628  * ib_resolve_eth_dmac - Resolve destination mac address
1629  * @device:		Device to consider
1630  * @ah_attr:		address handle attribute which describes the
1631  *			source and destination parameters
1632  * ib_resolve_eth_dmac() resolves destination mac address and L3 hop limit It
1633  * returns 0 on success or appropriate error code. It initializes the
1634  * necessary ah_attr fields when call is successful.
1635  */
1636 static int ib_resolve_eth_dmac(struct ib_device *device,
1637 			       struct rdma_ah_attr *ah_attr)
1638 {
1639 	int ret = 0;
1640 
1641 	if (rdma_is_multicast_addr((struct in6_addr *)ah_attr->grh.dgid.raw)) {
1642 		if (ipv6_addr_v4mapped((struct in6_addr *)ah_attr->grh.dgid.raw)) {
1643 			__be32 addr = 0;
1644 
1645 			memcpy(&addr, ah_attr->grh.dgid.raw + 12, 4);
1646 			ip_eth_mc_map(addr, (char *)ah_attr->roce.dmac);
1647 		} else {
1648 			ipv6_eth_mc_map((struct in6_addr *)ah_attr->grh.dgid.raw,
1649 					(char *)ah_attr->roce.dmac);
1650 		}
1651 	} else {
1652 		ret = ib_resolve_unicast_gid_dmac(device, ah_attr);
1653 	}
1654 	return ret;
1655 }
1656 
1657 static bool is_qp_type_connected(const struct ib_qp *qp)
1658 {
1659 	return (qp->qp_type == IB_QPT_UC ||
1660 		qp->qp_type == IB_QPT_RC ||
1661 		qp->qp_type == IB_QPT_XRC_INI ||
1662 		qp->qp_type == IB_QPT_XRC_TGT);
1663 }
1664 
1665 /**
1666  * IB core internal function to perform QP attributes modification.
1667  */
1668 static int _ib_modify_qp(struct ib_qp *qp, struct ib_qp_attr *attr,
1669 			 int attr_mask, struct ib_udata *udata)
1670 {
1671 	u8 port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
1672 	const struct ib_gid_attr *old_sgid_attr_av;
1673 	const struct ib_gid_attr *old_sgid_attr_alt_av;
1674 	int ret;
1675 
1676 	attr->xmit_slave = NULL;
1677 	if (attr_mask & IB_QP_AV) {
1678 		ret = rdma_fill_sgid_attr(qp->device, &attr->ah_attr,
1679 					  &old_sgid_attr_av);
1680 		if (ret)
1681 			return ret;
1682 
1683 		if (attr->ah_attr.type == RDMA_AH_ATTR_TYPE_ROCE &&
1684 		    is_qp_type_connected(qp)) {
1685 			struct net_device *slave;
1686 
1687 			/*
1688 			 * If the user provided the qp_attr then we have to
1689 			 * resolve it. Kerne users have to provide already
1690 			 * resolved rdma_ah_attr's.
1691 			 */
1692 			if (udata) {
1693 				ret = ib_resolve_eth_dmac(qp->device,
1694 							  &attr->ah_attr);
1695 				if (ret)
1696 					goto out_av;
1697 			}
1698 			slave = rdma_lag_get_ah_roce_slave(qp->device,
1699 							   &attr->ah_attr,
1700 							   GFP_KERNEL);
1701 			if (IS_ERR(slave))
1702 				goto out_av;
1703 			attr->xmit_slave = slave;
1704 		}
1705 	}
1706 	if (attr_mask & IB_QP_ALT_PATH) {
1707 		/*
1708 		 * FIXME: This does not track the migration state, so if the
1709 		 * user loads a new alternate path after the HW has migrated
1710 		 * from primary->alternate we will keep the wrong
1711 		 * references. This is OK for IB because the reference
1712 		 * counting does not serve any functional purpose.
1713 		 */
1714 		ret = rdma_fill_sgid_attr(qp->device, &attr->alt_ah_attr,
1715 					  &old_sgid_attr_alt_av);
1716 		if (ret)
1717 			goto out_av;
1718 
1719 		/*
1720 		 * Today the core code can only handle alternate paths and APM
1721 		 * for IB. Ban them in roce mode.
1722 		 */
1723 		if (!(rdma_protocol_ib(qp->device,
1724 				       attr->alt_ah_attr.port_num) &&
1725 		      rdma_protocol_ib(qp->device, port))) {
1726 			ret = -EINVAL;
1727 			goto out;
1728 		}
1729 	}
1730 
1731 	if (rdma_ib_or_roce(qp->device, port)) {
1732 		if (attr_mask & IB_QP_RQ_PSN && attr->rq_psn & ~0xffffff) {
1733 			dev_warn(&qp->device->dev,
1734 				 "%s rq_psn overflow, masking to 24 bits\n",
1735 				 __func__);
1736 			attr->rq_psn &= 0xffffff;
1737 		}
1738 
1739 		if (attr_mask & IB_QP_SQ_PSN && attr->sq_psn & ~0xffffff) {
1740 			dev_warn(&qp->device->dev,
1741 				 " %s sq_psn overflow, masking to 24 bits\n",
1742 				 __func__);
1743 			attr->sq_psn &= 0xffffff;
1744 		}
1745 	}
1746 
1747 	/*
1748 	 * Bind this qp to a counter automatically based on the rdma counter
1749 	 * rules. This only set in RST2INIT with port specified
1750 	 */
1751 	if (!qp->counter && (attr_mask & IB_QP_PORT) &&
1752 	    ((attr_mask & IB_QP_STATE) && attr->qp_state == IB_QPS_INIT))
1753 		rdma_counter_bind_qp_auto(qp, attr->port_num);
1754 
1755 	ret = ib_security_modify_qp(qp, attr, attr_mask, udata);
1756 	if (ret)
1757 		goto out;
1758 
1759 	if (attr_mask & IB_QP_PORT)
1760 		qp->port = attr->port_num;
1761 	if (attr_mask & IB_QP_AV)
1762 		qp->av_sgid_attr =
1763 			rdma_update_sgid_attr(&attr->ah_attr, qp->av_sgid_attr);
1764 	if (attr_mask & IB_QP_ALT_PATH)
1765 		qp->alt_path_sgid_attr = rdma_update_sgid_attr(
1766 			&attr->alt_ah_attr, qp->alt_path_sgid_attr);
1767 
1768 out:
1769 	if (attr_mask & IB_QP_ALT_PATH)
1770 		rdma_unfill_sgid_attr(&attr->alt_ah_attr, old_sgid_attr_alt_av);
1771 out_av:
1772 	if (attr_mask & IB_QP_AV) {
1773 		rdma_lag_put_ah_roce_slave(attr->xmit_slave);
1774 		rdma_unfill_sgid_attr(&attr->ah_attr, old_sgid_attr_av);
1775 	}
1776 	return ret;
1777 }
1778 
1779 /**
1780  * ib_modify_qp_with_udata - Modifies the attributes for the specified QP.
1781  * @ib_qp: The QP to modify.
1782  * @attr: On input, specifies the QP attributes to modify.  On output,
1783  *   the current values of selected QP attributes are returned.
1784  * @attr_mask: A bit-mask used to specify which attributes of the QP
1785  *   are being modified.
1786  * @udata: pointer to user's input output buffer information
1787  *   are being modified.
1788  * It returns 0 on success and returns appropriate error code on error.
1789  */
1790 int ib_modify_qp_with_udata(struct ib_qp *ib_qp, struct ib_qp_attr *attr,
1791 			    int attr_mask, struct ib_udata *udata)
1792 {
1793 	return _ib_modify_qp(ib_qp->real_qp, attr, attr_mask, udata);
1794 }
1795 EXPORT_SYMBOL(ib_modify_qp_with_udata);
1796 
1797 int ib_get_eth_speed(struct ib_device *dev, u8 port_num, u16 *speed, u8 *width)
1798 {
1799 	int rc;
1800 	u32 netdev_speed;
1801 	struct net_device *netdev;
1802 	struct ethtool_link_ksettings lksettings;
1803 
1804 	if (rdma_port_get_link_layer(dev, port_num) != IB_LINK_LAYER_ETHERNET)
1805 		return -EINVAL;
1806 
1807 	netdev = ib_device_get_netdev(dev, port_num);
1808 	if (!netdev)
1809 		return -ENODEV;
1810 
1811 	rtnl_lock();
1812 	rc = __ethtool_get_link_ksettings(netdev, &lksettings);
1813 	rtnl_unlock();
1814 
1815 	dev_put(netdev);
1816 
1817 	if (!rc && lksettings.base.speed != (u32)SPEED_UNKNOWN) {
1818 		netdev_speed = lksettings.base.speed;
1819 	} else {
1820 		netdev_speed = SPEED_1000;
1821 		pr_warn("%s speed is unknown, defaulting to %d\n", netdev->name,
1822 			netdev_speed);
1823 	}
1824 
1825 	if (netdev_speed <= SPEED_1000) {
1826 		*width = IB_WIDTH_1X;
1827 		*speed = IB_SPEED_SDR;
1828 	} else if (netdev_speed <= SPEED_10000) {
1829 		*width = IB_WIDTH_1X;
1830 		*speed = IB_SPEED_FDR10;
1831 	} else if (netdev_speed <= SPEED_20000) {
1832 		*width = IB_WIDTH_4X;
1833 		*speed = IB_SPEED_DDR;
1834 	} else if (netdev_speed <= SPEED_25000) {
1835 		*width = IB_WIDTH_1X;
1836 		*speed = IB_SPEED_EDR;
1837 	} else if (netdev_speed <= SPEED_40000) {
1838 		*width = IB_WIDTH_4X;
1839 		*speed = IB_SPEED_FDR10;
1840 	} else {
1841 		*width = IB_WIDTH_4X;
1842 		*speed = IB_SPEED_EDR;
1843 	}
1844 
1845 	return 0;
1846 }
1847 EXPORT_SYMBOL(ib_get_eth_speed);
1848 
1849 int ib_modify_qp(struct ib_qp *qp,
1850 		 struct ib_qp_attr *qp_attr,
1851 		 int qp_attr_mask)
1852 {
1853 	return _ib_modify_qp(qp->real_qp, qp_attr, qp_attr_mask, NULL);
1854 }
1855 EXPORT_SYMBOL(ib_modify_qp);
1856 
1857 int ib_query_qp(struct ib_qp *qp,
1858 		struct ib_qp_attr *qp_attr,
1859 		int qp_attr_mask,
1860 		struct ib_qp_init_attr *qp_init_attr)
1861 {
1862 	qp_attr->ah_attr.grh.sgid_attr = NULL;
1863 	qp_attr->alt_ah_attr.grh.sgid_attr = NULL;
1864 
1865 	return qp->device->ops.query_qp ?
1866 		qp->device->ops.query_qp(qp->real_qp, qp_attr, qp_attr_mask,
1867 					 qp_init_attr) : -EOPNOTSUPP;
1868 }
1869 EXPORT_SYMBOL(ib_query_qp);
1870 
1871 int ib_close_qp(struct ib_qp *qp)
1872 {
1873 	struct ib_qp *real_qp;
1874 	unsigned long flags;
1875 
1876 	real_qp = qp->real_qp;
1877 	if (real_qp == qp)
1878 		return -EINVAL;
1879 
1880 	spin_lock_irqsave(&real_qp->device->qp_open_list_lock, flags);
1881 	list_del(&qp->open_list);
1882 	spin_unlock_irqrestore(&real_qp->device->qp_open_list_lock, flags);
1883 
1884 	atomic_dec(&real_qp->usecnt);
1885 	if (qp->qp_sec)
1886 		ib_close_shared_qp_security(qp->qp_sec);
1887 	kfree(qp);
1888 
1889 	return 0;
1890 }
1891 EXPORT_SYMBOL(ib_close_qp);
1892 
1893 static int __ib_destroy_shared_qp(struct ib_qp *qp)
1894 {
1895 	struct ib_xrcd *xrcd;
1896 	struct ib_qp *real_qp;
1897 	int ret;
1898 
1899 	real_qp = qp->real_qp;
1900 	xrcd = real_qp->xrcd;
1901 	down_write(&xrcd->tgt_qps_rwsem);
1902 	ib_close_qp(qp);
1903 	if (atomic_read(&real_qp->usecnt) == 0)
1904 		xa_erase(&xrcd->tgt_qps, real_qp->qp_num);
1905 	else
1906 		real_qp = NULL;
1907 	up_write(&xrcd->tgt_qps_rwsem);
1908 
1909 	if (real_qp) {
1910 		ret = ib_destroy_qp(real_qp);
1911 		if (!ret)
1912 			atomic_dec(&xrcd->usecnt);
1913 	}
1914 
1915 	return 0;
1916 }
1917 
1918 int ib_destroy_qp_user(struct ib_qp *qp, struct ib_udata *udata)
1919 {
1920 	const struct ib_gid_attr *alt_path_sgid_attr = qp->alt_path_sgid_attr;
1921 	const struct ib_gid_attr *av_sgid_attr = qp->av_sgid_attr;
1922 	struct ib_pd *pd;
1923 	struct ib_cq *scq, *rcq;
1924 	struct ib_srq *srq;
1925 	struct ib_rwq_ind_table *ind_tbl;
1926 	struct ib_qp_security *sec;
1927 	int ret;
1928 
1929 	WARN_ON_ONCE(qp->mrs_used > 0);
1930 
1931 	if (atomic_read(&qp->usecnt))
1932 		return -EBUSY;
1933 
1934 	if (qp->real_qp != qp)
1935 		return __ib_destroy_shared_qp(qp);
1936 
1937 	pd   = qp->pd;
1938 	scq  = qp->send_cq;
1939 	rcq  = qp->recv_cq;
1940 	srq  = qp->srq;
1941 	ind_tbl = qp->rwq_ind_tbl;
1942 	sec  = qp->qp_sec;
1943 	if (sec)
1944 		ib_destroy_qp_security_begin(sec);
1945 
1946 	if (!qp->uobject)
1947 		rdma_rw_cleanup_mrs(qp);
1948 
1949 	rdma_counter_unbind_qp(qp, true);
1950 	rdma_restrack_del(&qp->res);
1951 	ret = qp->device->ops.destroy_qp(qp, udata);
1952 	if (!ret) {
1953 		if (alt_path_sgid_attr)
1954 			rdma_put_gid_attr(alt_path_sgid_attr);
1955 		if (av_sgid_attr)
1956 			rdma_put_gid_attr(av_sgid_attr);
1957 		if (pd)
1958 			atomic_dec(&pd->usecnt);
1959 		if (scq)
1960 			atomic_dec(&scq->usecnt);
1961 		if (rcq)
1962 			atomic_dec(&rcq->usecnt);
1963 		if (srq)
1964 			atomic_dec(&srq->usecnt);
1965 		if (ind_tbl)
1966 			atomic_dec(&ind_tbl->usecnt);
1967 		if (sec)
1968 			ib_destroy_qp_security_end(sec);
1969 	} else {
1970 		if (sec)
1971 			ib_destroy_qp_security_abort(sec);
1972 	}
1973 
1974 	return ret;
1975 }
1976 EXPORT_SYMBOL(ib_destroy_qp_user);
1977 
1978 /* Completion queues */
1979 
1980 struct ib_cq *__ib_create_cq(struct ib_device *device,
1981 			     ib_comp_handler comp_handler,
1982 			     void (*event_handler)(struct ib_event *, void *),
1983 			     void *cq_context,
1984 			     const struct ib_cq_init_attr *cq_attr,
1985 			     const char *caller)
1986 {
1987 	struct ib_cq *cq;
1988 	int ret;
1989 
1990 	cq = rdma_zalloc_drv_obj(device, ib_cq);
1991 	if (!cq)
1992 		return ERR_PTR(-ENOMEM);
1993 
1994 	cq->device = device;
1995 	cq->uobject = NULL;
1996 	cq->comp_handler = comp_handler;
1997 	cq->event_handler = event_handler;
1998 	cq->cq_context = cq_context;
1999 	atomic_set(&cq->usecnt, 0);
2000 
2001 	rdma_restrack_new(&cq->res, RDMA_RESTRACK_CQ);
2002 	rdma_restrack_set_name(&cq->res, caller);
2003 
2004 	ret = device->ops.create_cq(cq, cq_attr, NULL);
2005 	if (ret) {
2006 		rdma_restrack_put(&cq->res);
2007 		kfree(cq);
2008 		return ERR_PTR(ret);
2009 	}
2010 
2011 	rdma_restrack_add(&cq->res);
2012 	return cq;
2013 }
2014 EXPORT_SYMBOL(__ib_create_cq);
2015 
2016 int rdma_set_cq_moderation(struct ib_cq *cq, u16 cq_count, u16 cq_period)
2017 {
2018 	if (cq->shared)
2019 		return -EOPNOTSUPP;
2020 
2021 	return cq->device->ops.modify_cq ?
2022 		cq->device->ops.modify_cq(cq, cq_count,
2023 					  cq_period) : -EOPNOTSUPP;
2024 }
2025 EXPORT_SYMBOL(rdma_set_cq_moderation);
2026 
2027 int ib_destroy_cq_user(struct ib_cq *cq, struct ib_udata *udata)
2028 {
2029 	int ret;
2030 
2031 	if (WARN_ON_ONCE(cq->shared))
2032 		return -EOPNOTSUPP;
2033 
2034 	if (atomic_read(&cq->usecnt))
2035 		return -EBUSY;
2036 
2037 	ret = cq->device->ops.destroy_cq(cq, udata);
2038 	if (ret)
2039 		return ret;
2040 
2041 	rdma_restrack_del(&cq->res);
2042 	kfree(cq);
2043 	return ret;
2044 }
2045 EXPORT_SYMBOL(ib_destroy_cq_user);
2046 
2047 int ib_resize_cq(struct ib_cq *cq, int cqe)
2048 {
2049 	if (cq->shared)
2050 		return -EOPNOTSUPP;
2051 
2052 	return cq->device->ops.resize_cq ?
2053 		cq->device->ops.resize_cq(cq, cqe, NULL) : -EOPNOTSUPP;
2054 }
2055 EXPORT_SYMBOL(ib_resize_cq);
2056 
2057 /* Memory regions */
2058 
2059 struct ib_mr *ib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
2060 			     u64 virt_addr, int access_flags)
2061 {
2062 	struct ib_mr *mr;
2063 
2064 	if (access_flags & IB_ACCESS_ON_DEMAND) {
2065 		if (!(pd->device->attrs.device_cap_flags &
2066 		      IB_DEVICE_ON_DEMAND_PAGING)) {
2067 			pr_debug("ODP support not available\n");
2068 			return ERR_PTR(-EINVAL);
2069 		}
2070 	}
2071 
2072 	mr = pd->device->ops.reg_user_mr(pd, start, length, virt_addr,
2073 					 access_flags, NULL);
2074 
2075 	if (IS_ERR(mr))
2076 		return mr;
2077 
2078 	mr->device = pd->device;
2079 	mr->pd = pd;
2080 	mr->dm = NULL;
2081 	atomic_inc(&pd->usecnt);
2082 
2083 	rdma_restrack_new(&mr->res, RDMA_RESTRACK_MR);
2084 	rdma_restrack_parent_name(&mr->res, &pd->res);
2085 	rdma_restrack_add(&mr->res);
2086 
2087 	return mr;
2088 }
2089 EXPORT_SYMBOL(ib_reg_user_mr);
2090 
2091 int ib_advise_mr(struct ib_pd *pd, enum ib_uverbs_advise_mr_advice advice,
2092 		 u32 flags, struct ib_sge *sg_list, u32 num_sge)
2093 {
2094 	if (!pd->device->ops.advise_mr)
2095 		return -EOPNOTSUPP;
2096 
2097 	if (!num_sge)
2098 		return 0;
2099 
2100 	return pd->device->ops.advise_mr(pd, advice, flags, sg_list, num_sge,
2101 					 NULL);
2102 }
2103 EXPORT_SYMBOL(ib_advise_mr);
2104 
2105 int ib_dereg_mr_user(struct ib_mr *mr, struct ib_udata *udata)
2106 {
2107 	struct ib_pd *pd = mr->pd;
2108 	struct ib_dm *dm = mr->dm;
2109 	struct ib_sig_attrs *sig_attrs = mr->sig_attrs;
2110 	int ret;
2111 
2112 	trace_mr_dereg(mr);
2113 	rdma_restrack_del(&mr->res);
2114 	ret = mr->device->ops.dereg_mr(mr, udata);
2115 	if (!ret) {
2116 		atomic_dec(&pd->usecnt);
2117 		if (dm)
2118 			atomic_dec(&dm->usecnt);
2119 		kfree(sig_attrs);
2120 	}
2121 
2122 	return ret;
2123 }
2124 EXPORT_SYMBOL(ib_dereg_mr_user);
2125 
2126 /**
2127  * ib_alloc_mr() - Allocates a memory region
2128  * @pd:            protection domain associated with the region
2129  * @mr_type:       memory region type
2130  * @max_num_sg:    maximum sg entries available for registration.
2131  *
2132  * Notes:
2133  * Memory registeration page/sg lists must not exceed max_num_sg.
2134  * For mr_type IB_MR_TYPE_MEM_REG, the total length cannot exceed
2135  * max_num_sg * used_page_size.
2136  *
2137  */
2138 struct ib_mr *ib_alloc_mr(struct ib_pd *pd, enum ib_mr_type mr_type,
2139 			  u32 max_num_sg)
2140 {
2141 	struct ib_mr *mr;
2142 
2143 	if (!pd->device->ops.alloc_mr) {
2144 		mr = ERR_PTR(-EOPNOTSUPP);
2145 		goto out;
2146 	}
2147 
2148 	if (mr_type == IB_MR_TYPE_INTEGRITY) {
2149 		WARN_ON_ONCE(1);
2150 		mr = ERR_PTR(-EINVAL);
2151 		goto out;
2152 	}
2153 
2154 	mr = pd->device->ops.alloc_mr(pd, mr_type, max_num_sg);
2155 	if (IS_ERR(mr))
2156 		goto out;
2157 
2158 	mr->device = pd->device;
2159 	mr->pd = pd;
2160 	mr->dm = NULL;
2161 	mr->uobject = NULL;
2162 	atomic_inc(&pd->usecnt);
2163 	mr->need_inval = false;
2164 	mr->type = mr_type;
2165 	mr->sig_attrs = NULL;
2166 
2167 	rdma_restrack_new(&mr->res, RDMA_RESTRACK_MR);
2168 	rdma_restrack_parent_name(&mr->res, &pd->res);
2169 	rdma_restrack_add(&mr->res);
2170 out:
2171 	trace_mr_alloc(pd, mr_type, max_num_sg, mr);
2172 	return mr;
2173 }
2174 EXPORT_SYMBOL(ib_alloc_mr);
2175 
2176 /**
2177  * ib_alloc_mr_integrity() - Allocates an integrity memory region
2178  * @pd:                      protection domain associated with the region
2179  * @max_num_data_sg:         maximum data sg entries available for registration
2180  * @max_num_meta_sg:         maximum metadata sg entries available for
2181  *                           registration
2182  *
2183  * Notes:
2184  * Memory registration page/sg lists must not exceed max_num_sg,
2185  * also the integrity page/sg lists must not exceed max_num_meta_sg.
2186  *
2187  */
2188 struct ib_mr *ib_alloc_mr_integrity(struct ib_pd *pd,
2189 				    u32 max_num_data_sg,
2190 				    u32 max_num_meta_sg)
2191 {
2192 	struct ib_mr *mr;
2193 	struct ib_sig_attrs *sig_attrs;
2194 
2195 	if (!pd->device->ops.alloc_mr_integrity ||
2196 	    !pd->device->ops.map_mr_sg_pi) {
2197 		mr = ERR_PTR(-EOPNOTSUPP);
2198 		goto out;
2199 	}
2200 
2201 	if (!max_num_meta_sg) {
2202 		mr = ERR_PTR(-EINVAL);
2203 		goto out;
2204 	}
2205 
2206 	sig_attrs = kzalloc(sizeof(struct ib_sig_attrs), GFP_KERNEL);
2207 	if (!sig_attrs) {
2208 		mr = ERR_PTR(-ENOMEM);
2209 		goto out;
2210 	}
2211 
2212 	mr = pd->device->ops.alloc_mr_integrity(pd, max_num_data_sg,
2213 						max_num_meta_sg);
2214 	if (IS_ERR(mr)) {
2215 		kfree(sig_attrs);
2216 		goto out;
2217 	}
2218 
2219 	mr->device = pd->device;
2220 	mr->pd = pd;
2221 	mr->dm = NULL;
2222 	mr->uobject = NULL;
2223 	atomic_inc(&pd->usecnt);
2224 	mr->need_inval = false;
2225 	mr->type = IB_MR_TYPE_INTEGRITY;
2226 	mr->sig_attrs = sig_attrs;
2227 
2228 	rdma_restrack_new(&mr->res, RDMA_RESTRACK_MR);
2229 	rdma_restrack_parent_name(&mr->res, &pd->res);
2230 	rdma_restrack_add(&mr->res);
2231 out:
2232 	trace_mr_integ_alloc(pd, max_num_data_sg, max_num_meta_sg, mr);
2233 	return mr;
2234 }
2235 EXPORT_SYMBOL(ib_alloc_mr_integrity);
2236 
2237 /* Multicast groups */
2238 
2239 static bool is_valid_mcast_lid(struct ib_qp *qp, u16 lid)
2240 {
2241 	struct ib_qp_init_attr init_attr = {};
2242 	struct ib_qp_attr attr = {};
2243 	int num_eth_ports = 0;
2244 	int port;
2245 
2246 	/* If QP state >= init, it is assigned to a port and we can check this
2247 	 * port only.
2248 	 */
2249 	if (!ib_query_qp(qp, &attr, IB_QP_STATE | IB_QP_PORT, &init_attr)) {
2250 		if (attr.qp_state >= IB_QPS_INIT) {
2251 			if (rdma_port_get_link_layer(qp->device, attr.port_num) !=
2252 			    IB_LINK_LAYER_INFINIBAND)
2253 				return true;
2254 			goto lid_check;
2255 		}
2256 	}
2257 
2258 	/* Can't get a quick answer, iterate over all ports */
2259 	for (port = 0; port < qp->device->phys_port_cnt; port++)
2260 		if (rdma_port_get_link_layer(qp->device, port) !=
2261 		    IB_LINK_LAYER_INFINIBAND)
2262 			num_eth_ports++;
2263 
2264 	/* If we have at lease one Ethernet port, RoCE annex declares that
2265 	 * multicast LID should be ignored. We can't tell at this step if the
2266 	 * QP belongs to an IB or Ethernet port.
2267 	 */
2268 	if (num_eth_ports)
2269 		return true;
2270 
2271 	/* If all the ports are IB, we can check according to IB spec. */
2272 lid_check:
2273 	return !(lid < be16_to_cpu(IB_MULTICAST_LID_BASE) ||
2274 		 lid == be16_to_cpu(IB_LID_PERMISSIVE));
2275 }
2276 
2277 int ib_attach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid)
2278 {
2279 	int ret;
2280 
2281 	if (!qp->device->ops.attach_mcast)
2282 		return -EOPNOTSUPP;
2283 
2284 	if (!rdma_is_multicast_addr((struct in6_addr *)gid->raw) ||
2285 	    qp->qp_type != IB_QPT_UD || !is_valid_mcast_lid(qp, lid))
2286 		return -EINVAL;
2287 
2288 	ret = qp->device->ops.attach_mcast(qp, gid, lid);
2289 	if (!ret)
2290 		atomic_inc(&qp->usecnt);
2291 	return ret;
2292 }
2293 EXPORT_SYMBOL(ib_attach_mcast);
2294 
2295 int ib_detach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid)
2296 {
2297 	int ret;
2298 
2299 	if (!qp->device->ops.detach_mcast)
2300 		return -EOPNOTSUPP;
2301 
2302 	if (!rdma_is_multicast_addr((struct in6_addr *)gid->raw) ||
2303 	    qp->qp_type != IB_QPT_UD || !is_valid_mcast_lid(qp, lid))
2304 		return -EINVAL;
2305 
2306 	ret = qp->device->ops.detach_mcast(qp, gid, lid);
2307 	if (!ret)
2308 		atomic_dec(&qp->usecnt);
2309 	return ret;
2310 }
2311 EXPORT_SYMBOL(ib_detach_mcast);
2312 
2313 /**
2314  * ib_alloc_xrcd_user - Allocates an XRC domain.
2315  * @device: The device on which to allocate the XRC domain.
2316  * @inode: inode to connect XRCD
2317  * @udata: Valid user data or NULL for kernel object
2318  */
2319 struct ib_xrcd *ib_alloc_xrcd_user(struct ib_device *device,
2320 				   struct inode *inode, struct ib_udata *udata)
2321 {
2322 	struct ib_xrcd *xrcd;
2323 	int ret;
2324 
2325 	if (!device->ops.alloc_xrcd)
2326 		return ERR_PTR(-EOPNOTSUPP);
2327 
2328 	xrcd = rdma_zalloc_drv_obj(device, ib_xrcd);
2329 	if (!xrcd)
2330 		return ERR_PTR(-ENOMEM);
2331 
2332 	xrcd->device = device;
2333 	xrcd->inode = inode;
2334 	atomic_set(&xrcd->usecnt, 0);
2335 	init_rwsem(&xrcd->tgt_qps_rwsem);
2336 	xa_init(&xrcd->tgt_qps);
2337 
2338 	ret = device->ops.alloc_xrcd(xrcd, udata);
2339 	if (ret)
2340 		goto err;
2341 	return xrcd;
2342 err:
2343 	kfree(xrcd);
2344 	return ERR_PTR(ret);
2345 }
2346 EXPORT_SYMBOL(ib_alloc_xrcd_user);
2347 
2348 /**
2349  * ib_dealloc_xrcd_user - Deallocates an XRC domain.
2350  * @xrcd: The XRC domain to deallocate.
2351  * @udata: Valid user data or NULL for kernel object
2352  */
2353 int ib_dealloc_xrcd_user(struct ib_xrcd *xrcd, struct ib_udata *udata)
2354 {
2355 	int ret;
2356 
2357 	if (atomic_read(&xrcd->usecnt))
2358 		return -EBUSY;
2359 
2360 	WARN_ON(!xa_empty(&xrcd->tgt_qps));
2361 	ret = xrcd->device->ops.dealloc_xrcd(xrcd, udata);
2362 	if (ret)
2363 		return ret;
2364 	kfree(xrcd);
2365 	return ret;
2366 }
2367 EXPORT_SYMBOL(ib_dealloc_xrcd_user);
2368 
2369 /**
2370  * ib_create_wq - Creates a WQ associated with the specified protection
2371  * domain.
2372  * @pd: The protection domain associated with the WQ.
2373  * @wq_attr: A list of initial attributes required to create the
2374  * WQ. If WQ creation succeeds, then the attributes are updated to
2375  * the actual capabilities of the created WQ.
2376  *
2377  * wq_attr->max_wr and wq_attr->max_sge determine
2378  * the requested size of the WQ, and set to the actual values allocated
2379  * on return.
2380  * If ib_create_wq() succeeds, then max_wr and max_sge will always be
2381  * at least as large as the requested values.
2382  */
2383 struct ib_wq *ib_create_wq(struct ib_pd *pd,
2384 			   struct ib_wq_init_attr *wq_attr)
2385 {
2386 	struct ib_wq *wq;
2387 
2388 	if (!pd->device->ops.create_wq)
2389 		return ERR_PTR(-EOPNOTSUPP);
2390 
2391 	wq = pd->device->ops.create_wq(pd, wq_attr, NULL);
2392 	if (!IS_ERR(wq)) {
2393 		wq->event_handler = wq_attr->event_handler;
2394 		wq->wq_context = wq_attr->wq_context;
2395 		wq->wq_type = wq_attr->wq_type;
2396 		wq->cq = wq_attr->cq;
2397 		wq->device = pd->device;
2398 		wq->pd = pd;
2399 		wq->uobject = NULL;
2400 		atomic_inc(&pd->usecnt);
2401 		atomic_inc(&wq_attr->cq->usecnt);
2402 		atomic_set(&wq->usecnt, 0);
2403 	}
2404 	return wq;
2405 }
2406 EXPORT_SYMBOL(ib_create_wq);
2407 
2408 /**
2409  * ib_destroy_wq_user - Destroys the specified user WQ.
2410  * @wq: The WQ to destroy.
2411  * @udata: Valid user data
2412  */
2413 int ib_destroy_wq_user(struct ib_wq *wq, struct ib_udata *udata)
2414 {
2415 	struct ib_cq *cq = wq->cq;
2416 	struct ib_pd *pd = wq->pd;
2417 	int ret;
2418 
2419 	if (atomic_read(&wq->usecnt))
2420 		return -EBUSY;
2421 
2422 	ret = wq->device->ops.destroy_wq(wq, udata);
2423 	if (ret)
2424 		return ret;
2425 
2426 	atomic_dec(&pd->usecnt);
2427 	atomic_dec(&cq->usecnt);
2428 	return ret;
2429 }
2430 EXPORT_SYMBOL(ib_destroy_wq_user);
2431 
2432 /**
2433  * ib_modify_wq - Modifies the specified WQ.
2434  * @wq: The WQ to modify.
2435  * @wq_attr: On input, specifies the WQ attributes to modify.
2436  * @wq_attr_mask: A bit-mask used to specify which attributes of the WQ
2437  *   are being modified.
2438  * On output, the current values of selected WQ attributes are returned.
2439  */
2440 int ib_modify_wq(struct ib_wq *wq, struct ib_wq_attr *wq_attr,
2441 		 u32 wq_attr_mask)
2442 {
2443 	int err;
2444 
2445 	if (!wq->device->ops.modify_wq)
2446 		return -EOPNOTSUPP;
2447 
2448 	err = wq->device->ops.modify_wq(wq, wq_attr, wq_attr_mask, NULL);
2449 	return err;
2450 }
2451 EXPORT_SYMBOL(ib_modify_wq);
2452 
2453 int ib_check_mr_status(struct ib_mr *mr, u32 check_mask,
2454 		       struct ib_mr_status *mr_status)
2455 {
2456 	if (!mr->device->ops.check_mr_status)
2457 		return -EOPNOTSUPP;
2458 
2459 	return mr->device->ops.check_mr_status(mr, check_mask, mr_status);
2460 }
2461 EXPORT_SYMBOL(ib_check_mr_status);
2462 
2463 int ib_set_vf_link_state(struct ib_device *device, int vf, u8 port,
2464 			 int state)
2465 {
2466 	if (!device->ops.set_vf_link_state)
2467 		return -EOPNOTSUPP;
2468 
2469 	return device->ops.set_vf_link_state(device, vf, port, state);
2470 }
2471 EXPORT_SYMBOL(ib_set_vf_link_state);
2472 
2473 int ib_get_vf_config(struct ib_device *device, int vf, u8 port,
2474 		     struct ifla_vf_info *info)
2475 {
2476 	if (!device->ops.get_vf_config)
2477 		return -EOPNOTSUPP;
2478 
2479 	return device->ops.get_vf_config(device, vf, port, info);
2480 }
2481 EXPORT_SYMBOL(ib_get_vf_config);
2482 
2483 int ib_get_vf_stats(struct ib_device *device, int vf, u8 port,
2484 		    struct ifla_vf_stats *stats)
2485 {
2486 	if (!device->ops.get_vf_stats)
2487 		return -EOPNOTSUPP;
2488 
2489 	return device->ops.get_vf_stats(device, vf, port, stats);
2490 }
2491 EXPORT_SYMBOL(ib_get_vf_stats);
2492 
2493 int ib_set_vf_guid(struct ib_device *device, int vf, u8 port, u64 guid,
2494 		   int type)
2495 {
2496 	if (!device->ops.set_vf_guid)
2497 		return -EOPNOTSUPP;
2498 
2499 	return device->ops.set_vf_guid(device, vf, port, guid, type);
2500 }
2501 EXPORT_SYMBOL(ib_set_vf_guid);
2502 
2503 int ib_get_vf_guid(struct ib_device *device, int vf, u8 port,
2504 		   struct ifla_vf_guid *node_guid,
2505 		   struct ifla_vf_guid *port_guid)
2506 {
2507 	if (!device->ops.get_vf_guid)
2508 		return -EOPNOTSUPP;
2509 
2510 	return device->ops.get_vf_guid(device, vf, port, node_guid, port_guid);
2511 }
2512 EXPORT_SYMBOL(ib_get_vf_guid);
2513 /**
2514  * ib_map_mr_sg_pi() - Map the dma mapped SG lists for PI (protection
2515  *     information) and set an appropriate memory region for registration.
2516  * @mr:             memory region
2517  * @data_sg:        dma mapped scatterlist for data
2518  * @data_sg_nents:  number of entries in data_sg
2519  * @data_sg_offset: offset in bytes into data_sg
2520  * @meta_sg:        dma mapped scatterlist for metadata
2521  * @meta_sg_nents:  number of entries in meta_sg
2522  * @meta_sg_offset: offset in bytes into meta_sg
2523  * @page_size:      page vector desired page size
2524  *
2525  * Constraints:
2526  * - The MR must be allocated with type IB_MR_TYPE_INTEGRITY.
2527  *
2528  * Return: 0 on success.
2529  *
2530  * After this completes successfully, the  memory region
2531  * is ready for registration.
2532  */
2533 int ib_map_mr_sg_pi(struct ib_mr *mr, struct scatterlist *data_sg,
2534 		    int data_sg_nents, unsigned int *data_sg_offset,
2535 		    struct scatterlist *meta_sg, int meta_sg_nents,
2536 		    unsigned int *meta_sg_offset, unsigned int page_size)
2537 {
2538 	if (unlikely(!mr->device->ops.map_mr_sg_pi ||
2539 		     WARN_ON_ONCE(mr->type != IB_MR_TYPE_INTEGRITY)))
2540 		return -EOPNOTSUPP;
2541 
2542 	mr->page_size = page_size;
2543 
2544 	return mr->device->ops.map_mr_sg_pi(mr, data_sg, data_sg_nents,
2545 					    data_sg_offset, meta_sg,
2546 					    meta_sg_nents, meta_sg_offset);
2547 }
2548 EXPORT_SYMBOL(ib_map_mr_sg_pi);
2549 
2550 /**
2551  * ib_map_mr_sg() - Map the largest prefix of a dma mapped SG list
2552  *     and set it the memory region.
2553  * @mr:            memory region
2554  * @sg:            dma mapped scatterlist
2555  * @sg_nents:      number of entries in sg
2556  * @sg_offset:     offset in bytes into sg
2557  * @page_size:     page vector desired page size
2558  *
2559  * Constraints:
2560  *
2561  * - The first sg element is allowed to have an offset.
2562  * - Each sg element must either be aligned to page_size or virtually
2563  *   contiguous to the previous element. In case an sg element has a
2564  *   non-contiguous offset, the mapping prefix will not include it.
2565  * - The last sg element is allowed to have length less than page_size.
2566  * - If sg_nents total byte length exceeds the mr max_num_sge * page_size
2567  *   then only max_num_sg entries will be mapped.
2568  * - If the MR was allocated with type IB_MR_TYPE_SG_GAPS, none of these
2569  *   constraints holds and the page_size argument is ignored.
2570  *
2571  * Returns the number of sg elements that were mapped to the memory region.
2572  *
2573  * After this completes successfully, the  memory region
2574  * is ready for registration.
2575  */
2576 int ib_map_mr_sg(struct ib_mr *mr, struct scatterlist *sg, int sg_nents,
2577 		 unsigned int *sg_offset, unsigned int page_size)
2578 {
2579 	if (unlikely(!mr->device->ops.map_mr_sg))
2580 		return -EOPNOTSUPP;
2581 
2582 	mr->page_size = page_size;
2583 
2584 	return mr->device->ops.map_mr_sg(mr, sg, sg_nents, sg_offset);
2585 }
2586 EXPORT_SYMBOL(ib_map_mr_sg);
2587 
2588 /**
2589  * ib_sg_to_pages() - Convert the largest prefix of a sg list
2590  *     to a page vector
2591  * @mr:            memory region
2592  * @sgl:           dma mapped scatterlist
2593  * @sg_nents:      number of entries in sg
2594  * @sg_offset_p:   ==== =======================================================
2595  *                 IN   start offset in bytes into sg
2596  *                 OUT  offset in bytes for element n of the sg of the first
2597  *                      byte that has not been processed where n is the return
2598  *                      value of this function.
2599  *                 ==== =======================================================
2600  * @set_page:      driver page assignment function pointer
2601  *
2602  * Core service helper for drivers to convert the largest
2603  * prefix of given sg list to a page vector. The sg list
2604  * prefix converted is the prefix that meet the requirements
2605  * of ib_map_mr_sg.
2606  *
2607  * Returns the number of sg elements that were assigned to
2608  * a page vector.
2609  */
2610 int ib_sg_to_pages(struct ib_mr *mr, struct scatterlist *sgl, int sg_nents,
2611 		unsigned int *sg_offset_p, int (*set_page)(struct ib_mr *, u64))
2612 {
2613 	struct scatterlist *sg;
2614 	u64 last_end_dma_addr = 0;
2615 	unsigned int sg_offset = sg_offset_p ? *sg_offset_p : 0;
2616 	unsigned int last_page_off = 0;
2617 	u64 page_mask = ~((u64)mr->page_size - 1);
2618 	int i, ret;
2619 
2620 	if (unlikely(sg_nents <= 0 || sg_offset > sg_dma_len(&sgl[0])))
2621 		return -EINVAL;
2622 
2623 	mr->iova = sg_dma_address(&sgl[0]) + sg_offset;
2624 	mr->length = 0;
2625 
2626 	for_each_sg(sgl, sg, sg_nents, i) {
2627 		u64 dma_addr = sg_dma_address(sg) + sg_offset;
2628 		u64 prev_addr = dma_addr;
2629 		unsigned int dma_len = sg_dma_len(sg) - sg_offset;
2630 		u64 end_dma_addr = dma_addr + dma_len;
2631 		u64 page_addr = dma_addr & page_mask;
2632 
2633 		/*
2634 		 * For the second and later elements, check whether either the
2635 		 * end of element i-1 or the start of element i is not aligned
2636 		 * on a page boundary.
2637 		 */
2638 		if (i && (last_page_off != 0 || page_addr != dma_addr)) {
2639 			/* Stop mapping if there is a gap. */
2640 			if (last_end_dma_addr != dma_addr)
2641 				break;
2642 
2643 			/*
2644 			 * Coalesce this element with the last. If it is small
2645 			 * enough just update mr->length. Otherwise start
2646 			 * mapping from the next page.
2647 			 */
2648 			goto next_page;
2649 		}
2650 
2651 		do {
2652 			ret = set_page(mr, page_addr);
2653 			if (unlikely(ret < 0)) {
2654 				sg_offset = prev_addr - sg_dma_address(sg);
2655 				mr->length += prev_addr - dma_addr;
2656 				if (sg_offset_p)
2657 					*sg_offset_p = sg_offset;
2658 				return i || sg_offset ? i : ret;
2659 			}
2660 			prev_addr = page_addr;
2661 next_page:
2662 			page_addr += mr->page_size;
2663 		} while (page_addr < end_dma_addr);
2664 
2665 		mr->length += dma_len;
2666 		last_end_dma_addr = end_dma_addr;
2667 		last_page_off = end_dma_addr & ~page_mask;
2668 
2669 		sg_offset = 0;
2670 	}
2671 
2672 	if (sg_offset_p)
2673 		*sg_offset_p = 0;
2674 	return i;
2675 }
2676 EXPORT_SYMBOL(ib_sg_to_pages);
2677 
2678 struct ib_drain_cqe {
2679 	struct ib_cqe cqe;
2680 	struct completion done;
2681 };
2682 
2683 static void ib_drain_qp_done(struct ib_cq *cq, struct ib_wc *wc)
2684 {
2685 	struct ib_drain_cqe *cqe = container_of(wc->wr_cqe, struct ib_drain_cqe,
2686 						cqe);
2687 
2688 	complete(&cqe->done);
2689 }
2690 
2691 /*
2692  * Post a WR and block until its completion is reaped for the SQ.
2693  */
2694 static void __ib_drain_sq(struct ib_qp *qp)
2695 {
2696 	struct ib_cq *cq = qp->send_cq;
2697 	struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
2698 	struct ib_drain_cqe sdrain;
2699 	struct ib_rdma_wr swr = {
2700 		.wr = {
2701 			.next = NULL,
2702 			{ .wr_cqe	= &sdrain.cqe, },
2703 			.opcode	= IB_WR_RDMA_WRITE,
2704 		},
2705 	};
2706 	int ret;
2707 
2708 	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
2709 	if (ret) {
2710 		WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
2711 		return;
2712 	}
2713 
2714 	sdrain.cqe.done = ib_drain_qp_done;
2715 	init_completion(&sdrain.done);
2716 
2717 	ret = ib_post_send(qp, &swr.wr, NULL);
2718 	if (ret) {
2719 		WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
2720 		return;
2721 	}
2722 
2723 	if (cq->poll_ctx == IB_POLL_DIRECT)
2724 		while (wait_for_completion_timeout(&sdrain.done, HZ / 10) <= 0)
2725 			ib_process_cq_direct(cq, -1);
2726 	else
2727 		wait_for_completion(&sdrain.done);
2728 }
2729 
2730 /*
2731  * Post a WR and block until its completion is reaped for the RQ.
2732  */
2733 static void __ib_drain_rq(struct ib_qp *qp)
2734 {
2735 	struct ib_cq *cq = qp->recv_cq;
2736 	struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
2737 	struct ib_drain_cqe rdrain;
2738 	struct ib_recv_wr rwr = {};
2739 	int ret;
2740 
2741 	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
2742 	if (ret) {
2743 		WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
2744 		return;
2745 	}
2746 
2747 	rwr.wr_cqe = &rdrain.cqe;
2748 	rdrain.cqe.done = ib_drain_qp_done;
2749 	init_completion(&rdrain.done);
2750 
2751 	ret = ib_post_recv(qp, &rwr, NULL);
2752 	if (ret) {
2753 		WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
2754 		return;
2755 	}
2756 
2757 	if (cq->poll_ctx == IB_POLL_DIRECT)
2758 		while (wait_for_completion_timeout(&rdrain.done, HZ / 10) <= 0)
2759 			ib_process_cq_direct(cq, -1);
2760 	else
2761 		wait_for_completion(&rdrain.done);
2762 }
2763 
2764 /**
2765  * ib_drain_sq() - Block until all SQ CQEs have been consumed by the
2766  *		   application.
2767  * @qp:            queue pair to drain
2768  *
2769  * If the device has a provider-specific drain function, then
2770  * call that.  Otherwise call the generic drain function
2771  * __ib_drain_sq().
2772  *
2773  * The caller must:
2774  *
2775  * ensure there is room in the CQ and SQ for the drain work request and
2776  * completion.
2777  *
2778  * allocate the CQ using ib_alloc_cq().
2779  *
2780  * ensure that there are no other contexts that are posting WRs concurrently.
2781  * Otherwise the drain is not guaranteed.
2782  */
2783 void ib_drain_sq(struct ib_qp *qp)
2784 {
2785 	if (qp->device->ops.drain_sq)
2786 		qp->device->ops.drain_sq(qp);
2787 	else
2788 		__ib_drain_sq(qp);
2789 	trace_cq_drain_complete(qp->send_cq);
2790 }
2791 EXPORT_SYMBOL(ib_drain_sq);
2792 
2793 /**
2794  * ib_drain_rq() - Block until all RQ CQEs have been consumed by the
2795  *		   application.
2796  * @qp:            queue pair to drain
2797  *
2798  * If the device has a provider-specific drain function, then
2799  * call that.  Otherwise call the generic drain function
2800  * __ib_drain_rq().
2801  *
2802  * The caller must:
2803  *
2804  * ensure there is room in the CQ and RQ for the drain work request and
2805  * completion.
2806  *
2807  * allocate the CQ using ib_alloc_cq().
2808  *
2809  * ensure that there are no other contexts that are posting WRs concurrently.
2810  * Otherwise the drain is not guaranteed.
2811  */
2812 void ib_drain_rq(struct ib_qp *qp)
2813 {
2814 	if (qp->device->ops.drain_rq)
2815 		qp->device->ops.drain_rq(qp);
2816 	else
2817 		__ib_drain_rq(qp);
2818 	trace_cq_drain_complete(qp->recv_cq);
2819 }
2820 EXPORT_SYMBOL(ib_drain_rq);
2821 
2822 /**
2823  * ib_drain_qp() - Block until all CQEs have been consumed by the
2824  *		   application on both the RQ and SQ.
2825  * @qp:            queue pair to drain
2826  *
2827  * The caller must:
2828  *
2829  * ensure there is room in the CQ(s), SQ, and RQ for drain work requests
2830  * and completions.
2831  *
2832  * allocate the CQs using ib_alloc_cq().
2833  *
2834  * ensure that there are no other contexts that are posting WRs concurrently.
2835  * Otherwise the drain is not guaranteed.
2836  */
2837 void ib_drain_qp(struct ib_qp *qp)
2838 {
2839 	ib_drain_sq(qp);
2840 	if (!qp->srq)
2841 		ib_drain_rq(qp);
2842 }
2843 EXPORT_SYMBOL(ib_drain_qp);
2844 
2845 struct net_device *rdma_alloc_netdev(struct ib_device *device, u8 port_num,
2846 				     enum rdma_netdev_t type, const char *name,
2847 				     unsigned char name_assign_type,
2848 				     void (*setup)(struct net_device *))
2849 {
2850 	struct rdma_netdev_alloc_params params;
2851 	struct net_device *netdev;
2852 	int rc;
2853 
2854 	if (!device->ops.rdma_netdev_get_params)
2855 		return ERR_PTR(-EOPNOTSUPP);
2856 
2857 	rc = device->ops.rdma_netdev_get_params(device, port_num, type,
2858 						&params);
2859 	if (rc)
2860 		return ERR_PTR(rc);
2861 
2862 	netdev = alloc_netdev_mqs(params.sizeof_priv, name, name_assign_type,
2863 				  setup, params.txqs, params.rxqs);
2864 	if (!netdev)
2865 		return ERR_PTR(-ENOMEM);
2866 
2867 	return netdev;
2868 }
2869 EXPORT_SYMBOL(rdma_alloc_netdev);
2870 
2871 int rdma_init_netdev(struct ib_device *device, u8 port_num,
2872 		     enum rdma_netdev_t type, const char *name,
2873 		     unsigned char name_assign_type,
2874 		     void (*setup)(struct net_device *),
2875 		     struct net_device *netdev)
2876 {
2877 	struct rdma_netdev_alloc_params params;
2878 	int rc;
2879 
2880 	if (!device->ops.rdma_netdev_get_params)
2881 		return -EOPNOTSUPP;
2882 
2883 	rc = device->ops.rdma_netdev_get_params(device, port_num, type,
2884 						&params);
2885 	if (rc)
2886 		return rc;
2887 
2888 	return params.initialize_rdma_netdev(device, port_num,
2889 					     netdev, params.param);
2890 }
2891 EXPORT_SYMBOL(rdma_init_netdev);
2892 
2893 void __rdma_block_iter_start(struct ib_block_iter *biter,
2894 			     struct scatterlist *sglist, unsigned int nents,
2895 			     unsigned long pgsz)
2896 {
2897 	memset(biter, 0, sizeof(struct ib_block_iter));
2898 	biter->__sg = sglist;
2899 	biter->__sg_nents = nents;
2900 
2901 	/* Driver provides best block size to use */
2902 	biter->__pg_bit = __fls(pgsz);
2903 }
2904 EXPORT_SYMBOL(__rdma_block_iter_start);
2905 
2906 bool __rdma_block_iter_next(struct ib_block_iter *biter)
2907 {
2908 	unsigned int block_offset;
2909 
2910 	if (!biter->__sg_nents || !biter->__sg)
2911 		return false;
2912 
2913 	biter->__dma_addr = sg_dma_address(biter->__sg) + biter->__sg_advance;
2914 	block_offset = biter->__dma_addr & (BIT_ULL(biter->__pg_bit) - 1);
2915 	biter->__sg_advance += BIT_ULL(biter->__pg_bit) - block_offset;
2916 
2917 	if (biter->__sg_advance >= sg_dma_len(biter->__sg)) {
2918 		biter->__sg_advance = 0;
2919 		biter->__sg = sg_next(biter->__sg);
2920 		biter->__sg_nents--;
2921 	}
2922 
2923 	return true;
2924 }
2925 EXPORT_SYMBOL(__rdma_block_iter_next);
2926