xref: /linux/drivers/infiniband/core/verbs.c (revision 60a2f25de7b8b785baee2932db932ae9a5b8c86d)
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 bind 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]		= "remote 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 	case IB_RATE_800_GBPS: return 320;
151 	default:	       return  -1;
152 	}
153 }
154 EXPORT_SYMBOL(ib_rate_to_mult);
155 
156 __attribute_const__ enum ib_rate mult_to_ib_rate(int mult)
157 {
158 	switch (mult) {
159 	case 1:   return IB_RATE_2_5_GBPS;
160 	case 2:   return IB_RATE_5_GBPS;
161 	case 4:   return IB_RATE_10_GBPS;
162 	case 8:   return IB_RATE_20_GBPS;
163 	case 12:  return IB_RATE_30_GBPS;
164 	case 16:  return IB_RATE_40_GBPS;
165 	case 24:  return IB_RATE_60_GBPS;
166 	case 32:  return IB_RATE_80_GBPS;
167 	case 48:  return IB_RATE_120_GBPS;
168 	case 6:   return IB_RATE_14_GBPS;
169 	case 22:  return IB_RATE_56_GBPS;
170 	case 45:  return IB_RATE_112_GBPS;
171 	case 67:  return IB_RATE_168_GBPS;
172 	case 10:  return IB_RATE_25_GBPS;
173 	case 40:  return IB_RATE_100_GBPS;
174 	case 80:  return IB_RATE_200_GBPS;
175 	case 120: return IB_RATE_300_GBPS;
176 	case 11:  return IB_RATE_28_GBPS;
177 	case 20:  return IB_RATE_50_GBPS;
178 	case 160: return IB_RATE_400_GBPS;
179 	case 240: return IB_RATE_600_GBPS;
180 	case 320: return IB_RATE_800_GBPS;
181 	default:  return IB_RATE_PORT_CURRENT;
182 	}
183 }
184 EXPORT_SYMBOL(mult_to_ib_rate);
185 
186 __attribute_const__ int ib_rate_to_mbps(enum ib_rate rate)
187 {
188 	switch (rate) {
189 	case IB_RATE_2_5_GBPS: return 2500;
190 	case IB_RATE_5_GBPS:   return 5000;
191 	case IB_RATE_10_GBPS:  return 10000;
192 	case IB_RATE_20_GBPS:  return 20000;
193 	case IB_RATE_30_GBPS:  return 30000;
194 	case IB_RATE_40_GBPS:  return 40000;
195 	case IB_RATE_60_GBPS:  return 60000;
196 	case IB_RATE_80_GBPS:  return 80000;
197 	case IB_RATE_120_GBPS: return 120000;
198 	case IB_RATE_14_GBPS:  return 14062;
199 	case IB_RATE_56_GBPS:  return 56250;
200 	case IB_RATE_112_GBPS: return 112500;
201 	case IB_RATE_168_GBPS: return 168750;
202 	case IB_RATE_25_GBPS:  return 25781;
203 	case IB_RATE_100_GBPS: return 103125;
204 	case IB_RATE_200_GBPS: return 206250;
205 	case IB_RATE_300_GBPS: return 309375;
206 	case IB_RATE_28_GBPS:  return 28125;
207 	case IB_RATE_50_GBPS:  return 53125;
208 	case IB_RATE_400_GBPS: return 425000;
209 	case IB_RATE_600_GBPS: return 637500;
210 	case IB_RATE_800_GBPS: return 850000;
211 	default:	       return -1;
212 	}
213 }
214 EXPORT_SYMBOL(ib_rate_to_mbps);
215 
216 __attribute_const__ enum rdma_transport_type
217 rdma_node_get_transport(unsigned int node_type)
218 {
219 
220 	if (node_type == RDMA_NODE_USNIC)
221 		return RDMA_TRANSPORT_USNIC;
222 	if (node_type == RDMA_NODE_USNIC_UDP)
223 		return RDMA_TRANSPORT_USNIC_UDP;
224 	if (node_type == RDMA_NODE_RNIC)
225 		return RDMA_TRANSPORT_IWARP;
226 	if (node_type == RDMA_NODE_UNSPECIFIED)
227 		return RDMA_TRANSPORT_UNSPECIFIED;
228 
229 	return RDMA_TRANSPORT_IB;
230 }
231 EXPORT_SYMBOL(rdma_node_get_transport);
232 
233 enum rdma_link_layer rdma_port_get_link_layer(struct ib_device *device,
234 					      u32 port_num)
235 {
236 	enum rdma_transport_type lt;
237 	if (device->ops.get_link_layer)
238 		return device->ops.get_link_layer(device, port_num);
239 
240 	lt = rdma_node_get_transport(device->node_type);
241 	if (lt == RDMA_TRANSPORT_IB)
242 		return IB_LINK_LAYER_INFINIBAND;
243 
244 	return IB_LINK_LAYER_ETHERNET;
245 }
246 EXPORT_SYMBOL(rdma_port_get_link_layer);
247 
248 /* Protection domains */
249 
250 /**
251  * __ib_alloc_pd - Allocates an unused protection domain.
252  * @device: The device on which to allocate the protection domain.
253  * @flags: protection domain flags
254  * @caller: caller's build-time module name
255  *
256  * A protection domain object provides an association between QPs, shared
257  * receive queues, address handles, memory regions, and memory windows.
258  *
259  * Every PD has a local_dma_lkey which can be used as the lkey value for local
260  * memory operations.
261  */
262 struct ib_pd *__ib_alloc_pd(struct ib_device *device, unsigned int flags,
263 		const char *caller)
264 {
265 	struct ib_pd *pd;
266 	int mr_access_flags = 0;
267 	int ret;
268 
269 	pd = rdma_zalloc_drv_obj(device, ib_pd);
270 	if (!pd)
271 		return ERR_PTR(-ENOMEM);
272 
273 	pd->device = device;
274 	pd->flags = flags;
275 
276 	rdma_restrack_new(&pd->res, RDMA_RESTRACK_PD);
277 	rdma_restrack_set_name(&pd->res, caller);
278 
279 	ret = device->ops.alloc_pd(pd, NULL);
280 	if (ret) {
281 		rdma_restrack_put(&pd->res);
282 		kfree(pd);
283 		return ERR_PTR(ret);
284 	}
285 	rdma_restrack_add(&pd->res);
286 
287 	if (device->attrs.kernel_cap_flags & IBK_LOCAL_DMA_LKEY)
288 		pd->local_dma_lkey = device->local_dma_lkey;
289 	else
290 		mr_access_flags |= IB_ACCESS_LOCAL_WRITE;
291 
292 	if (flags & IB_PD_UNSAFE_GLOBAL_RKEY) {
293 		pr_warn("%s: enabling unsafe global rkey\n", caller);
294 		mr_access_flags |= IB_ACCESS_REMOTE_READ | IB_ACCESS_REMOTE_WRITE;
295 	}
296 
297 	if (mr_access_flags) {
298 		struct ib_mr *mr;
299 
300 		mr = pd->device->ops.get_dma_mr(pd, mr_access_flags);
301 		if (IS_ERR(mr)) {
302 			ib_dealloc_pd(pd);
303 			return ERR_CAST(mr);
304 		}
305 
306 		mr->device	= pd->device;
307 		mr->pd		= pd;
308 		mr->type        = IB_MR_TYPE_DMA;
309 		mr->uobject	= NULL;
310 		mr->need_inval	= false;
311 
312 		pd->__internal_mr = mr;
313 
314 		if (!(device->attrs.kernel_cap_flags & IBK_LOCAL_DMA_LKEY))
315 			pd->local_dma_lkey = pd->__internal_mr->lkey;
316 
317 		if (flags & IB_PD_UNSAFE_GLOBAL_RKEY)
318 			pd->unsafe_global_rkey = pd->__internal_mr->rkey;
319 	}
320 
321 	return pd;
322 }
323 EXPORT_SYMBOL(__ib_alloc_pd);
324 
325 /**
326  * ib_dealloc_pd_user - Deallocates a protection domain.
327  * @pd: The protection domain to deallocate.
328  * @udata: Valid user data or NULL for kernel object
329  *
330  * It is an error to call this function while any resources in the pd still
331  * exist.  The caller is responsible to synchronously destroy them and
332  * guarantee no new allocations will happen.
333  */
334 int ib_dealloc_pd_user(struct ib_pd *pd, struct ib_udata *udata)
335 {
336 	int ret;
337 
338 	if (pd->__internal_mr) {
339 		ret = pd->device->ops.dereg_mr(pd->__internal_mr, NULL);
340 		WARN_ON(ret);
341 		pd->__internal_mr = NULL;
342 	}
343 
344 	ret = pd->device->ops.dealloc_pd(pd, udata);
345 	if (ret)
346 		return ret;
347 
348 	rdma_restrack_del(&pd->res);
349 	kfree(pd);
350 	return ret;
351 }
352 EXPORT_SYMBOL(ib_dealloc_pd_user);
353 
354 /* Address handles */
355 
356 /**
357  * rdma_copy_ah_attr - Copy rdma ah attribute from source to destination.
358  * @dest:       Pointer to destination ah_attr. Contents of the destination
359  *              pointer is assumed to be invalid and attribute are overwritten.
360  * @src:        Pointer to source ah_attr.
361  */
362 void rdma_copy_ah_attr(struct rdma_ah_attr *dest,
363 		       const struct rdma_ah_attr *src)
364 {
365 	*dest = *src;
366 	if (dest->grh.sgid_attr)
367 		rdma_hold_gid_attr(dest->grh.sgid_attr);
368 }
369 EXPORT_SYMBOL(rdma_copy_ah_attr);
370 
371 /**
372  * rdma_replace_ah_attr - Replace valid ah_attr with new one.
373  * @old:        Pointer to existing ah_attr which needs to be replaced.
374  *              old is assumed to be valid or zero'd
375  * @new:        Pointer to the new ah_attr.
376  *
377  * rdma_replace_ah_attr() first releases any reference in the old ah_attr if
378  * old the ah_attr is valid; after that it copies the new attribute and holds
379  * the reference to the replaced ah_attr.
380  */
381 void rdma_replace_ah_attr(struct rdma_ah_attr *old,
382 			  const struct rdma_ah_attr *new)
383 {
384 	rdma_destroy_ah_attr(old);
385 	*old = *new;
386 	if (old->grh.sgid_attr)
387 		rdma_hold_gid_attr(old->grh.sgid_attr);
388 }
389 EXPORT_SYMBOL(rdma_replace_ah_attr);
390 
391 /**
392  * rdma_move_ah_attr - Move ah_attr pointed by source to destination.
393  * @dest:       Pointer to destination ah_attr to copy to.
394  *              dest is assumed to be valid or zero'd
395  * @src:        Pointer to the new ah_attr.
396  *
397  * rdma_move_ah_attr() first releases any reference in the destination ah_attr
398  * if it is valid. This also transfers ownership of internal references from
399  * src to dest, making src invalid in the process. No new reference of the src
400  * ah_attr is taken.
401  */
402 void rdma_move_ah_attr(struct rdma_ah_attr *dest, struct rdma_ah_attr *src)
403 {
404 	rdma_destroy_ah_attr(dest);
405 	*dest = *src;
406 	src->grh.sgid_attr = NULL;
407 }
408 EXPORT_SYMBOL(rdma_move_ah_attr);
409 
410 /*
411  * Validate that the rdma_ah_attr is valid for the device before passing it
412  * off to the driver.
413  */
414 static int rdma_check_ah_attr(struct ib_device *device,
415 			      struct rdma_ah_attr *ah_attr)
416 {
417 	if (!rdma_is_port_valid(device, ah_attr->port_num))
418 		return -EINVAL;
419 
420 	if ((rdma_is_grh_required(device, ah_attr->port_num) ||
421 	     ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE) &&
422 	    !(ah_attr->ah_flags & IB_AH_GRH))
423 		return -EINVAL;
424 
425 	if (ah_attr->grh.sgid_attr) {
426 		/*
427 		 * Make sure the passed sgid_attr is consistent with the
428 		 * parameters
429 		 */
430 		if (ah_attr->grh.sgid_attr->index != ah_attr->grh.sgid_index ||
431 		    ah_attr->grh.sgid_attr->port_num != ah_attr->port_num)
432 			return -EINVAL;
433 	}
434 	return 0;
435 }
436 
437 /*
438  * If the ah requires a GRH then ensure that sgid_attr pointer is filled in.
439  * On success the caller is responsible to call rdma_unfill_sgid_attr().
440  */
441 static int rdma_fill_sgid_attr(struct ib_device *device,
442 			       struct rdma_ah_attr *ah_attr,
443 			       const struct ib_gid_attr **old_sgid_attr)
444 {
445 	const struct ib_gid_attr *sgid_attr;
446 	struct ib_global_route *grh;
447 	int ret;
448 
449 	*old_sgid_attr = ah_attr->grh.sgid_attr;
450 
451 	ret = rdma_check_ah_attr(device, ah_attr);
452 	if (ret)
453 		return ret;
454 
455 	if (!(ah_attr->ah_flags & IB_AH_GRH))
456 		return 0;
457 
458 	grh = rdma_ah_retrieve_grh(ah_attr);
459 	if (grh->sgid_attr)
460 		return 0;
461 
462 	sgid_attr =
463 		rdma_get_gid_attr(device, ah_attr->port_num, grh->sgid_index);
464 	if (IS_ERR(sgid_attr))
465 		return PTR_ERR(sgid_attr);
466 
467 	/* Move ownerhip of the kref into the ah_attr */
468 	grh->sgid_attr = sgid_attr;
469 	return 0;
470 }
471 
472 static void rdma_unfill_sgid_attr(struct rdma_ah_attr *ah_attr,
473 				  const struct ib_gid_attr *old_sgid_attr)
474 {
475 	/*
476 	 * Fill didn't change anything, the caller retains ownership of
477 	 * whatever it passed
478 	 */
479 	if (ah_attr->grh.sgid_attr == old_sgid_attr)
480 		return;
481 
482 	/*
483 	 * Otherwise, we need to undo what rdma_fill_sgid_attr so the caller
484 	 * doesn't see any change in the rdma_ah_attr. If we get here
485 	 * old_sgid_attr is NULL.
486 	 */
487 	rdma_destroy_ah_attr(ah_attr);
488 }
489 
490 static const struct ib_gid_attr *
491 rdma_update_sgid_attr(struct rdma_ah_attr *ah_attr,
492 		      const struct ib_gid_attr *old_attr)
493 {
494 	if (old_attr)
495 		rdma_put_gid_attr(old_attr);
496 	if (ah_attr->ah_flags & IB_AH_GRH) {
497 		rdma_hold_gid_attr(ah_attr->grh.sgid_attr);
498 		return ah_attr->grh.sgid_attr;
499 	}
500 	return NULL;
501 }
502 
503 static struct ib_ah *_rdma_create_ah(struct ib_pd *pd,
504 				     struct rdma_ah_attr *ah_attr,
505 				     u32 flags,
506 				     struct ib_udata *udata,
507 				     struct net_device *xmit_slave)
508 {
509 	struct rdma_ah_init_attr init_attr = {};
510 	struct ib_device *device = pd->device;
511 	struct ib_ah *ah;
512 	int ret;
513 
514 	might_sleep_if(flags & RDMA_CREATE_AH_SLEEPABLE);
515 
516 	if (!udata && !device->ops.create_ah)
517 		return ERR_PTR(-EOPNOTSUPP);
518 
519 	ah = rdma_zalloc_drv_obj_gfp(
520 		device, ib_ah,
521 		(flags & RDMA_CREATE_AH_SLEEPABLE) ? GFP_KERNEL : GFP_ATOMIC);
522 	if (!ah)
523 		return ERR_PTR(-ENOMEM);
524 
525 	ah->device = device;
526 	ah->pd = pd;
527 	ah->type = ah_attr->type;
528 	ah->sgid_attr = rdma_update_sgid_attr(ah_attr, NULL);
529 	init_attr.ah_attr = ah_attr;
530 	init_attr.flags = flags;
531 	init_attr.xmit_slave = xmit_slave;
532 
533 	if (udata)
534 		ret = device->ops.create_user_ah(ah, &init_attr, udata);
535 	else
536 		ret = device->ops.create_ah(ah, &init_attr, NULL);
537 	if (ret) {
538 		if (ah->sgid_attr)
539 			rdma_put_gid_attr(ah->sgid_attr);
540 		kfree(ah);
541 		return ERR_PTR(ret);
542 	}
543 
544 	atomic_inc(&pd->usecnt);
545 	return ah;
546 }
547 
548 /**
549  * rdma_create_ah - Creates an address handle for the
550  * given address vector.
551  * @pd: The protection domain associated with the address handle.
552  * @ah_attr: The attributes of the address vector.
553  * @flags: Create address handle flags (see enum rdma_create_ah_flags).
554  *
555  * It returns 0 on success and returns appropriate error code on error.
556  * The address handle is used to reference a local or global destination
557  * in all UD QP post sends.
558  */
559 struct ib_ah *rdma_create_ah(struct ib_pd *pd, struct rdma_ah_attr *ah_attr,
560 			     u32 flags)
561 {
562 	const struct ib_gid_attr *old_sgid_attr;
563 	struct net_device *slave;
564 	struct ib_ah *ah;
565 	int ret;
566 
567 	ret = rdma_fill_sgid_attr(pd->device, ah_attr, &old_sgid_attr);
568 	if (ret)
569 		return ERR_PTR(ret);
570 	slave = rdma_lag_get_ah_roce_slave(pd->device, ah_attr,
571 					   (flags & RDMA_CREATE_AH_SLEEPABLE) ?
572 					   GFP_KERNEL : GFP_ATOMIC);
573 	if (IS_ERR(slave)) {
574 		rdma_unfill_sgid_attr(ah_attr, old_sgid_attr);
575 		return (void *)slave;
576 	}
577 	ah = _rdma_create_ah(pd, ah_attr, flags, NULL, slave);
578 	rdma_lag_put_ah_roce_slave(slave);
579 	rdma_unfill_sgid_attr(ah_attr, old_sgid_attr);
580 	return ah;
581 }
582 EXPORT_SYMBOL(rdma_create_ah);
583 
584 /**
585  * rdma_create_user_ah - Creates an address handle for the
586  * given address vector.
587  * It resolves destination mac address for ah attribute of RoCE type.
588  * @pd: The protection domain associated with the address handle.
589  * @ah_attr: The attributes of the address vector.
590  * @udata: pointer to user's input output buffer information need by
591  *         provider driver.
592  *
593  * It returns 0 on success and returns appropriate error code on error.
594  * The address handle is used to reference a local or global destination
595  * in all UD QP post sends.
596  */
597 struct ib_ah *rdma_create_user_ah(struct ib_pd *pd,
598 				  struct rdma_ah_attr *ah_attr,
599 				  struct ib_udata *udata)
600 {
601 	const struct ib_gid_attr *old_sgid_attr;
602 	struct ib_ah *ah;
603 	int err;
604 
605 	err = rdma_fill_sgid_attr(pd->device, ah_attr, &old_sgid_attr);
606 	if (err)
607 		return ERR_PTR(err);
608 
609 	if (ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE) {
610 		err = ib_resolve_eth_dmac(pd->device, ah_attr);
611 		if (err) {
612 			ah = ERR_PTR(err);
613 			goto out;
614 		}
615 	}
616 
617 	ah = _rdma_create_ah(pd, ah_attr, RDMA_CREATE_AH_SLEEPABLE,
618 			     udata, NULL);
619 
620 out:
621 	rdma_unfill_sgid_attr(ah_attr, old_sgid_attr);
622 	return ah;
623 }
624 EXPORT_SYMBOL(rdma_create_user_ah);
625 
626 int ib_get_rdma_header_version(const union rdma_network_hdr *hdr)
627 {
628 	const struct iphdr *ip4h = (struct iphdr *)&hdr->roce4grh;
629 	struct iphdr ip4h_checked;
630 	const struct ipv6hdr *ip6h = (struct ipv6hdr *)&hdr->ibgrh;
631 
632 	/* If it's IPv6, the version must be 6, otherwise, the first
633 	 * 20 bytes (before the IPv4 header) are garbled.
634 	 */
635 	if (ip6h->version != 6)
636 		return (ip4h->version == 4) ? 4 : 0;
637 	/* version may be 6 or 4 because the first 20 bytes could be garbled */
638 
639 	/* RoCE v2 requires no options, thus header length
640 	 * must be 5 words
641 	 */
642 	if (ip4h->ihl != 5)
643 		return 6;
644 
645 	/* Verify checksum.
646 	 * We can't write on scattered buffers so we need to copy to
647 	 * temp buffer.
648 	 */
649 	memcpy(&ip4h_checked, ip4h, sizeof(ip4h_checked));
650 	ip4h_checked.check = 0;
651 	ip4h_checked.check = ip_fast_csum((u8 *)&ip4h_checked, 5);
652 	/* if IPv4 header checksum is OK, believe it */
653 	if (ip4h->check == ip4h_checked.check)
654 		return 4;
655 	return 6;
656 }
657 EXPORT_SYMBOL(ib_get_rdma_header_version);
658 
659 static enum rdma_network_type ib_get_net_type_by_grh(struct ib_device *device,
660 						     u32 port_num,
661 						     const struct ib_grh *grh)
662 {
663 	int grh_version;
664 
665 	if (rdma_protocol_ib(device, port_num))
666 		return RDMA_NETWORK_IB;
667 
668 	grh_version = ib_get_rdma_header_version((union rdma_network_hdr *)grh);
669 
670 	if (grh_version == 4)
671 		return RDMA_NETWORK_IPV4;
672 
673 	if (grh->next_hdr == IPPROTO_UDP)
674 		return RDMA_NETWORK_IPV6;
675 
676 	return RDMA_NETWORK_ROCE_V1;
677 }
678 
679 struct find_gid_index_context {
680 	u16 vlan_id;
681 	enum ib_gid_type gid_type;
682 };
683 
684 static bool find_gid_index(const union ib_gid *gid,
685 			   const struct ib_gid_attr *gid_attr,
686 			   void *context)
687 {
688 	struct find_gid_index_context *ctx = context;
689 	u16 vlan_id = 0xffff;
690 	int ret;
691 
692 	if (ctx->gid_type != gid_attr->gid_type)
693 		return false;
694 
695 	ret = rdma_read_gid_l2_fields(gid_attr, &vlan_id, NULL);
696 	if (ret)
697 		return false;
698 
699 	return ctx->vlan_id == vlan_id;
700 }
701 
702 static const struct ib_gid_attr *
703 get_sgid_attr_from_eth(struct ib_device *device, u32 port_num,
704 		       u16 vlan_id, const union ib_gid *sgid,
705 		       enum ib_gid_type gid_type)
706 {
707 	struct find_gid_index_context context = {.vlan_id = vlan_id,
708 						 .gid_type = gid_type};
709 
710 	return rdma_find_gid_by_filter(device, sgid, port_num, find_gid_index,
711 				       &context);
712 }
713 
714 int ib_get_gids_from_rdma_hdr(const union rdma_network_hdr *hdr,
715 			      enum rdma_network_type net_type,
716 			      union ib_gid *sgid, union ib_gid *dgid)
717 {
718 	struct sockaddr_in  src_in;
719 	struct sockaddr_in  dst_in;
720 	__be32 src_saddr, dst_saddr;
721 
722 	if (!sgid || !dgid)
723 		return -EINVAL;
724 
725 	if (net_type == RDMA_NETWORK_IPV4) {
726 		memcpy(&src_in.sin_addr.s_addr,
727 		       &hdr->roce4grh.saddr, 4);
728 		memcpy(&dst_in.sin_addr.s_addr,
729 		       &hdr->roce4grh.daddr, 4);
730 		src_saddr = src_in.sin_addr.s_addr;
731 		dst_saddr = dst_in.sin_addr.s_addr;
732 		ipv6_addr_set_v4mapped(src_saddr,
733 				       (struct in6_addr *)sgid);
734 		ipv6_addr_set_v4mapped(dst_saddr,
735 				       (struct in6_addr *)dgid);
736 		return 0;
737 	} else if (net_type == RDMA_NETWORK_IPV6 ||
738 		   net_type == RDMA_NETWORK_IB || RDMA_NETWORK_ROCE_V1) {
739 		*dgid = hdr->ibgrh.dgid;
740 		*sgid = hdr->ibgrh.sgid;
741 		return 0;
742 	} else {
743 		return -EINVAL;
744 	}
745 }
746 EXPORT_SYMBOL(ib_get_gids_from_rdma_hdr);
747 
748 /* Resolve destination mac address and hop limit for unicast destination
749  * GID entry, considering the source GID entry as well.
750  * ah_attribute must have valid port_num, sgid_index.
751  */
752 static int ib_resolve_unicast_gid_dmac(struct ib_device *device,
753 				       struct rdma_ah_attr *ah_attr)
754 {
755 	struct ib_global_route *grh = rdma_ah_retrieve_grh(ah_attr);
756 	const struct ib_gid_attr *sgid_attr = grh->sgid_attr;
757 	int hop_limit = 0xff;
758 	int ret = 0;
759 
760 	/* If destination is link local and source GID is RoCEv1,
761 	 * IP stack is not used.
762 	 */
763 	if (rdma_link_local_addr((struct in6_addr *)grh->dgid.raw) &&
764 	    sgid_attr->gid_type == IB_GID_TYPE_ROCE) {
765 		rdma_get_ll_mac((struct in6_addr *)grh->dgid.raw,
766 				ah_attr->roce.dmac);
767 		return ret;
768 	}
769 
770 	ret = rdma_addr_find_l2_eth_by_grh(&sgid_attr->gid, &grh->dgid,
771 					   ah_attr->roce.dmac,
772 					   sgid_attr, &hop_limit);
773 
774 	grh->hop_limit = hop_limit;
775 	return ret;
776 }
777 
778 /*
779  * This function initializes address handle attributes from the incoming packet.
780  * Incoming packet has dgid of the receiver node on which this code is
781  * getting executed and, sgid contains the GID of the sender.
782  *
783  * When resolving mac address of destination, the arrived dgid is used
784  * as sgid and, sgid is used as dgid because sgid contains destinations
785  * GID whom to respond to.
786  *
787  * On success the caller is responsible to call rdma_destroy_ah_attr on the
788  * attr.
789  */
790 int ib_init_ah_attr_from_wc(struct ib_device *device, u32 port_num,
791 			    const struct ib_wc *wc, const struct ib_grh *grh,
792 			    struct rdma_ah_attr *ah_attr)
793 {
794 	u32 flow_class;
795 	int ret;
796 	enum rdma_network_type net_type = RDMA_NETWORK_IB;
797 	enum ib_gid_type gid_type = IB_GID_TYPE_IB;
798 	const struct ib_gid_attr *sgid_attr;
799 	int hoplimit = 0xff;
800 	union ib_gid dgid;
801 	union ib_gid sgid;
802 
803 	might_sleep();
804 
805 	memset(ah_attr, 0, sizeof *ah_attr);
806 	ah_attr->type = rdma_ah_find_type(device, port_num);
807 	if (rdma_cap_eth_ah(device, port_num)) {
808 		if (wc->wc_flags & IB_WC_WITH_NETWORK_HDR_TYPE)
809 			net_type = wc->network_hdr_type;
810 		else
811 			net_type = ib_get_net_type_by_grh(device, port_num, grh);
812 		gid_type = ib_network_to_gid_type(net_type);
813 	}
814 	ret = ib_get_gids_from_rdma_hdr((union rdma_network_hdr *)grh, net_type,
815 					&sgid, &dgid);
816 	if (ret)
817 		return ret;
818 
819 	rdma_ah_set_sl(ah_attr, wc->sl);
820 	rdma_ah_set_port_num(ah_attr, port_num);
821 
822 	if (rdma_protocol_roce(device, port_num)) {
823 		u16 vlan_id = wc->wc_flags & IB_WC_WITH_VLAN ?
824 				wc->vlan_id : 0xffff;
825 
826 		if (!(wc->wc_flags & IB_WC_GRH))
827 			return -EPROTOTYPE;
828 
829 		sgid_attr = get_sgid_attr_from_eth(device, port_num,
830 						   vlan_id, &dgid,
831 						   gid_type);
832 		if (IS_ERR(sgid_attr))
833 			return PTR_ERR(sgid_attr);
834 
835 		flow_class = be32_to_cpu(grh->version_tclass_flow);
836 		rdma_move_grh_sgid_attr(ah_attr,
837 					&sgid,
838 					flow_class & 0xFFFFF,
839 					hoplimit,
840 					(flow_class >> 20) & 0xFF,
841 					sgid_attr);
842 
843 		ret = ib_resolve_unicast_gid_dmac(device, ah_attr);
844 		if (ret)
845 			rdma_destroy_ah_attr(ah_attr);
846 
847 		return ret;
848 	} else {
849 		rdma_ah_set_dlid(ah_attr, wc->slid);
850 		rdma_ah_set_path_bits(ah_attr, wc->dlid_path_bits);
851 
852 		if ((wc->wc_flags & IB_WC_GRH) == 0)
853 			return 0;
854 
855 		if (dgid.global.interface_id !=
856 					cpu_to_be64(IB_SA_WELL_KNOWN_GUID)) {
857 			sgid_attr = rdma_find_gid_by_port(
858 				device, &dgid, IB_GID_TYPE_IB, port_num, NULL);
859 		} else
860 			sgid_attr = rdma_get_gid_attr(device, port_num, 0);
861 
862 		if (IS_ERR(sgid_attr))
863 			return PTR_ERR(sgid_attr);
864 		flow_class = be32_to_cpu(grh->version_tclass_flow);
865 		rdma_move_grh_sgid_attr(ah_attr,
866 					&sgid,
867 					flow_class & 0xFFFFF,
868 					hoplimit,
869 					(flow_class >> 20) & 0xFF,
870 					sgid_attr);
871 
872 		return 0;
873 	}
874 }
875 EXPORT_SYMBOL(ib_init_ah_attr_from_wc);
876 
877 /**
878  * rdma_move_grh_sgid_attr - Sets the sgid attribute of GRH, taking ownership
879  * of the reference
880  *
881  * @attr:	Pointer to AH attribute structure
882  * @dgid:	Destination GID
883  * @flow_label:	Flow label
884  * @hop_limit:	Hop limit
885  * @traffic_class: traffic class
886  * @sgid_attr:	Pointer to SGID attribute
887  *
888  * This takes ownership of the sgid_attr reference. The caller must ensure
889  * rdma_destroy_ah_attr() is called before destroying the rdma_ah_attr after
890  * calling this function.
891  */
892 void rdma_move_grh_sgid_attr(struct rdma_ah_attr *attr, union ib_gid *dgid,
893 			     u32 flow_label, u8 hop_limit, u8 traffic_class,
894 			     const struct ib_gid_attr *sgid_attr)
895 {
896 	rdma_ah_set_grh(attr, dgid, flow_label, sgid_attr->index, hop_limit,
897 			traffic_class);
898 	attr->grh.sgid_attr = sgid_attr;
899 }
900 EXPORT_SYMBOL(rdma_move_grh_sgid_attr);
901 
902 /**
903  * rdma_destroy_ah_attr - Release reference to SGID attribute of
904  * ah attribute.
905  * @ah_attr: Pointer to ah attribute
906  *
907  * Release reference to the SGID attribute of the ah attribute if it is
908  * non NULL. It is safe to call this multiple times, and safe to call it on
909  * a zero initialized ah_attr.
910  */
911 void rdma_destroy_ah_attr(struct rdma_ah_attr *ah_attr)
912 {
913 	if (ah_attr->grh.sgid_attr) {
914 		rdma_put_gid_attr(ah_attr->grh.sgid_attr);
915 		ah_attr->grh.sgid_attr = NULL;
916 	}
917 }
918 EXPORT_SYMBOL(rdma_destroy_ah_attr);
919 
920 struct ib_ah *ib_create_ah_from_wc(struct ib_pd *pd, const struct ib_wc *wc,
921 				   const struct ib_grh *grh, u32 port_num)
922 {
923 	struct rdma_ah_attr ah_attr;
924 	struct ib_ah *ah;
925 	int ret;
926 
927 	ret = ib_init_ah_attr_from_wc(pd->device, port_num, wc, grh, &ah_attr);
928 	if (ret)
929 		return ERR_PTR(ret);
930 
931 	ah = rdma_create_ah(pd, &ah_attr, RDMA_CREATE_AH_SLEEPABLE);
932 
933 	rdma_destroy_ah_attr(&ah_attr);
934 	return ah;
935 }
936 EXPORT_SYMBOL(ib_create_ah_from_wc);
937 
938 int rdma_modify_ah(struct ib_ah *ah, struct rdma_ah_attr *ah_attr)
939 {
940 	const struct ib_gid_attr *old_sgid_attr;
941 	int ret;
942 
943 	if (ah->type != ah_attr->type)
944 		return -EINVAL;
945 
946 	ret = rdma_fill_sgid_attr(ah->device, ah_attr, &old_sgid_attr);
947 	if (ret)
948 		return ret;
949 
950 	ret = ah->device->ops.modify_ah ?
951 		ah->device->ops.modify_ah(ah, ah_attr) :
952 		-EOPNOTSUPP;
953 
954 	ah->sgid_attr = rdma_update_sgid_attr(ah_attr, ah->sgid_attr);
955 	rdma_unfill_sgid_attr(ah_attr, old_sgid_attr);
956 	return ret;
957 }
958 EXPORT_SYMBOL(rdma_modify_ah);
959 
960 int rdma_query_ah(struct ib_ah *ah, struct rdma_ah_attr *ah_attr)
961 {
962 	ah_attr->grh.sgid_attr = NULL;
963 
964 	return ah->device->ops.query_ah ?
965 		ah->device->ops.query_ah(ah, ah_attr) :
966 		-EOPNOTSUPP;
967 }
968 EXPORT_SYMBOL(rdma_query_ah);
969 
970 int rdma_destroy_ah_user(struct ib_ah *ah, u32 flags, struct ib_udata *udata)
971 {
972 	const struct ib_gid_attr *sgid_attr = ah->sgid_attr;
973 	struct ib_pd *pd;
974 	int ret;
975 
976 	might_sleep_if(flags & RDMA_DESTROY_AH_SLEEPABLE);
977 
978 	pd = ah->pd;
979 
980 	ret = ah->device->ops.destroy_ah(ah, flags);
981 	if (ret)
982 		return ret;
983 
984 	atomic_dec(&pd->usecnt);
985 	if (sgid_attr)
986 		rdma_put_gid_attr(sgid_attr);
987 
988 	kfree(ah);
989 	return ret;
990 }
991 EXPORT_SYMBOL(rdma_destroy_ah_user);
992 
993 /* Shared receive queues */
994 
995 /**
996  * ib_create_srq_user - Creates a SRQ associated with the specified protection
997  *   domain.
998  * @pd: The protection domain associated with the SRQ.
999  * @srq_init_attr: A list of initial attributes required to create the
1000  *   SRQ.  If SRQ creation succeeds, then the attributes are updated to
1001  *   the actual capabilities of the created SRQ.
1002  * @uobject: uobject pointer if this is not a kernel SRQ
1003  * @udata: udata pointer if this is not a kernel SRQ
1004  *
1005  * srq_attr->max_wr and srq_attr->max_sge are read the determine the
1006  * requested size of the SRQ, and set to the actual values allocated
1007  * on return.  If ib_create_srq() succeeds, then max_wr and max_sge
1008  * will always be at least as large as the requested values.
1009  */
1010 struct ib_srq *ib_create_srq_user(struct ib_pd *pd,
1011 				  struct ib_srq_init_attr *srq_init_attr,
1012 				  struct ib_usrq_object *uobject,
1013 				  struct ib_udata *udata)
1014 {
1015 	struct ib_srq *srq;
1016 	int ret;
1017 
1018 	srq = rdma_zalloc_drv_obj(pd->device, ib_srq);
1019 	if (!srq)
1020 		return ERR_PTR(-ENOMEM);
1021 
1022 	srq->device = pd->device;
1023 	srq->pd = pd;
1024 	srq->event_handler = srq_init_attr->event_handler;
1025 	srq->srq_context = srq_init_attr->srq_context;
1026 	srq->srq_type = srq_init_attr->srq_type;
1027 	srq->uobject = uobject;
1028 
1029 	if (ib_srq_has_cq(srq->srq_type)) {
1030 		srq->ext.cq = srq_init_attr->ext.cq;
1031 		atomic_inc(&srq->ext.cq->usecnt);
1032 	}
1033 	if (srq->srq_type == IB_SRQT_XRC) {
1034 		srq->ext.xrc.xrcd = srq_init_attr->ext.xrc.xrcd;
1035 		if (srq->ext.xrc.xrcd)
1036 			atomic_inc(&srq->ext.xrc.xrcd->usecnt);
1037 	}
1038 	atomic_inc(&pd->usecnt);
1039 
1040 	rdma_restrack_new(&srq->res, RDMA_RESTRACK_SRQ);
1041 	rdma_restrack_parent_name(&srq->res, &pd->res);
1042 
1043 	ret = pd->device->ops.create_srq(srq, srq_init_attr, udata);
1044 	if (ret) {
1045 		rdma_restrack_put(&srq->res);
1046 		atomic_dec(&pd->usecnt);
1047 		if (srq->srq_type == IB_SRQT_XRC && srq->ext.xrc.xrcd)
1048 			atomic_dec(&srq->ext.xrc.xrcd->usecnt);
1049 		if (ib_srq_has_cq(srq->srq_type))
1050 			atomic_dec(&srq->ext.cq->usecnt);
1051 		kfree(srq);
1052 		return ERR_PTR(ret);
1053 	}
1054 
1055 	rdma_restrack_add(&srq->res);
1056 
1057 	return srq;
1058 }
1059 EXPORT_SYMBOL(ib_create_srq_user);
1060 
1061 int ib_modify_srq(struct ib_srq *srq,
1062 		  struct ib_srq_attr *srq_attr,
1063 		  enum ib_srq_attr_mask srq_attr_mask)
1064 {
1065 	return srq->device->ops.modify_srq ?
1066 		srq->device->ops.modify_srq(srq, srq_attr, srq_attr_mask,
1067 					    NULL) : -EOPNOTSUPP;
1068 }
1069 EXPORT_SYMBOL(ib_modify_srq);
1070 
1071 int ib_query_srq(struct ib_srq *srq,
1072 		 struct ib_srq_attr *srq_attr)
1073 {
1074 	return srq->device->ops.query_srq ?
1075 		srq->device->ops.query_srq(srq, srq_attr) : -EOPNOTSUPP;
1076 }
1077 EXPORT_SYMBOL(ib_query_srq);
1078 
1079 int ib_destroy_srq_user(struct ib_srq *srq, struct ib_udata *udata)
1080 {
1081 	int ret;
1082 
1083 	if (atomic_read(&srq->usecnt))
1084 		return -EBUSY;
1085 
1086 	ret = srq->device->ops.destroy_srq(srq, udata);
1087 	if (ret)
1088 		return ret;
1089 
1090 	atomic_dec(&srq->pd->usecnt);
1091 	if (srq->srq_type == IB_SRQT_XRC && srq->ext.xrc.xrcd)
1092 		atomic_dec(&srq->ext.xrc.xrcd->usecnt);
1093 	if (ib_srq_has_cq(srq->srq_type))
1094 		atomic_dec(&srq->ext.cq->usecnt);
1095 	rdma_restrack_del(&srq->res);
1096 	kfree(srq);
1097 
1098 	return ret;
1099 }
1100 EXPORT_SYMBOL(ib_destroy_srq_user);
1101 
1102 /* Queue pairs */
1103 
1104 static void __ib_shared_qp_event_handler(struct ib_event *event, void *context)
1105 {
1106 	struct ib_qp *qp = context;
1107 	unsigned long flags;
1108 
1109 	spin_lock_irqsave(&qp->device->qp_open_list_lock, flags);
1110 	list_for_each_entry(event->element.qp, &qp->open_list, open_list)
1111 		if (event->element.qp->event_handler)
1112 			event->element.qp->event_handler(event, event->element.qp->qp_context);
1113 	spin_unlock_irqrestore(&qp->device->qp_open_list_lock, flags);
1114 }
1115 
1116 static struct ib_qp *__ib_open_qp(struct ib_qp *real_qp,
1117 				  void (*event_handler)(struct ib_event *, void *),
1118 				  void *qp_context)
1119 {
1120 	struct ib_qp *qp;
1121 	unsigned long flags;
1122 	int err;
1123 
1124 	qp = kzalloc(sizeof *qp, GFP_KERNEL);
1125 	if (!qp)
1126 		return ERR_PTR(-ENOMEM);
1127 
1128 	qp->real_qp = real_qp;
1129 	err = ib_open_shared_qp_security(qp, real_qp->device);
1130 	if (err) {
1131 		kfree(qp);
1132 		return ERR_PTR(err);
1133 	}
1134 
1135 	qp->real_qp = real_qp;
1136 	atomic_inc(&real_qp->usecnt);
1137 	qp->device = real_qp->device;
1138 	qp->event_handler = event_handler;
1139 	qp->qp_context = qp_context;
1140 	qp->qp_num = real_qp->qp_num;
1141 	qp->qp_type = real_qp->qp_type;
1142 
1143 	spin_lock_irqsave(&real_qp->device->qp_open_list_lock, flags);
1144 	list_add(&qp->open_list, &real_qp->open_list);
1145 	spin_unlock_irqrestore(&real_qp->device->qp_open_list_lock, flags);
1146 
1147 	return qp;
1148 }
1149 
1150 struct ib_qp *ib_open_qp(struct ib_xrcd *xrcd,
1151 			 struct ib_qp_open_attr *qp_open_attr)
1152 {
1153 	struct ib_qp *qp, *real_qp;
1154 
1155 	if (qp_open_attr->qp_type != IB_QPT_XRC_TGT)
1156 		return ERR_PTR(-EINVAL);
1157 
1158 	down_read(&xrcd->tgt_qps_rwsem);
1159 	real_qp = xa_load(&xrcd->tgt_qps, qp_open_attr->qp_num);
1160 	if (!real_qp) {
1161 		up_read(&xrcd->tgt_qps_rwsem);
1162 		return ERR_PTR(-EINVAL);
1163 	}
1164 	qp = __ib_open_qp(real_qp, qp_open_attr->event_handler,
1165 			  qp_open_attr->qp_context);
1166 	up_read(&xrcd->tgt_qps_rwsem);
1167 	return qp;
1168 }
1169 EXPORT_SYMBOL(ib_open_qp);
1170 
1171 static struct ib_qp *create_xrc_qp_user(struct ib_qp *qp,
1172 					struct ib_qp_init_attr *qp_init_attr)
1173 {
1174 	struct ib_qp *real_qp = qp;
1175 	int err;
1176 
1177 	qp->event_handler = __ib_shared_qp_event_handler;
1178 	qp->qp_context = qp;
1179 	qp->pd = NULL;
1180 	qp->send_cq = qp->recv_cq = NULL;
1181 	qp->srq = NULL;
1182 	qp->xrcd = qp_init_attr->xrcd;
1183 	atomic_inc(&qp_init_attr->xrcd->usecnt);
1184 	INIT_LIST_HEAD(&qp->open_list);
1185 
1186 	qp = __ib_open_qp(real_qp, qp_init_attr->event_handler,
1187 			  qp_init_attr->qp_context);
1188 	if (IS_ERR(qp))
1189 		return qp;
1190 
1191 	err = xa_err(xa_store(&qp_init_attr->xrcd->tgt_qps, real_qp->qp_num,
1192 			      real_qp, GFP_KERNEL));
1193 	if (err) {
1194 		ib_close_qp(qp);
1195 		return ERR_PTR(err);
1196 	}
1197 	return qp;
1198 }
1199 
1200 static struct ib_qp *create_qp(struct ib_device *dev, struct ib_pd *pd,
1201 			       struct ib_qp_init_attr *attr,
1202 			       struct ib_udata *udata,
1203 			       struct ib_uqp_object *uobj, const char *caller)
1204 {
1205 	struct ib_udata dummy = {};
1206 	struct ib_qp *qp;
1207 	int ret;
1208 
1209 	if (!dev->ops.create_qp)
1210 		return ERR_PTR(-EOPNOTSUPP);
1211 
1212 	qp = rdma_zalloc_drv_obj_numa(dev, ib_qp);
1213 	if (!qp)
1214 		return ERR_PTR(-ENOMEM);
1215 
1216 	qp->device = dev;
1217 	qp->pd = pd;
1218 	qp->uobject = uobj;
1219 	qp->real_qp = qp;
1220 
1221 	qp->qp_type = attr->qp_type;
1222 	qp->rwq_ind_tbl = attr->rwq_ind_tbl;
1223 	qp->srq = attr->srq;
1224 	qp->event_handler = attr->event_handler;
1225 	qp->port = attr->port_num;
1226 	qp->qp_context = attr->qp_context;
1227 
1228 	spin_lock_init(&qp->mr_lock);
1229 	INIT_LIST_HEAD(&qp->rdma_mrs);
1230 	INIT_LIST_HEAD(&qp->sig_mrs);
1231 
1232 	qp->send_cq = attr->send_cq;
1233 	qp->recv_cq = attr->recv_cq;
1234 
1235 	rdma_restrack_new(&qp->res, RDMA_RESTRACK_QP);
1236 	WARN_ONCE(!udata && !caller, "Missing kernel QP owner");
1237 	rdma_restrack_set_name(&qp->res, udata ? NULL : caller);
1238 	ret = dev->ops.create_qp(qp, attr, udata);
1239 	if (ret)
1240 		goto err_create;
1241 
1242 	/*
1243 	 * TODO: The mlx4 internally overwrites send_cq and recv_cq.
1244 	 * Unfortunately, it is not an easy task to fix that driver.
1245 	 */
1246 	qp->send_cq = attr->send_cq;
1247 	qp->recv_cq = attr->recv_cq;
1248 
1249 	ret = ib_create_qp_security(qp, dev);
1250 	if (ret)
1251 		goto err_security;
1252 
1253 	rdma_restrack_add(&qp->res);
1254 	return qp;
1255 
1256 err_security:
1257 	qp->device->ops.destroy_qp(qp, udata ? &dummy : NULL);
1258 err_create:
1259 	rdma_restrack_put(&qp->res);
1260 	kfree(qp);
1261 	return ERR_PTR(ret);
1262 
1263 }
1264 
1265 /**
1266  * ib_create_qp_user - Creates a QP associated with the specified protection
1267  *   domain.
1268  * @dev: IB device
1269  * @pd: The protection domain associated with the QP.
1270  * @attr: A list of initial attributes required to create the
1271  *   QP.  If QP creation succeeds, then the attributes are updated to
1272  *   the actual capabilities of the created QP.
1273  * @udata: User data
1274  * @uobj: uverbs obect
1275  * @caller: caller's build-time module name
1276  */
1277 struct ib_qp *ib_create_qp_user(struct ib_device *dev, struct ib_pd *pd,
1278 				struct ib_qp_init_attr *attr,
1279 				struct ib_udata *udata,
1280 				struct ib_uqp_object *uobj, const char *caller)
1281 {
1282 	struct ib_qp *qp, *xrc_qp;
1283 
1284 	if (attr->qp_type == IB_QPT_XRC_TGT)
1285 		qp = create_qp(dev, pd, attr, NULL, NULL, caller);
1286 	else
1287 		qp = create_qp(dev, pd, attr, udata, uobj, NULL);
1288 	if (attr->qp_type != IB_QPT_XRC_TGT || IS_ERR(qp))
1289 		return qp;
1290 
1291 	xrc_qp = create_xrc_qp_user(qp, attr);
1292 	if (IS_ERR(xrc_qp)) {
1293 		ib_destroy_qp(qp);
1294 		return xrc_qp;
1295 	}
1296 
1297 	xrc_qp->uobject = uobj;
1298 	return xrc_qp;
1299 }
1300 EXPORT_SYMBOL(ib_create_qp_user);
1301 
1302 void ib_qp_usecnt_inc(struct ib_qp *qp)
1303 {
1304 	if (qp->pd)
1305 		atomic_inc(&qp->pd->usecnt);
1306 	if (qp->send_cq)
1307 		atomic_inc(&qp->send_cq->usecnt);
1308 	if (qp->recv_cq)
1309 		atomic_inc(&qp->recv_cq->usecnt);
1310 	if (qp->srq)
1311 		atomic_inc(&qp->srq->usecnt);
1312 	if (qp->rwq_ind_tbl)
1313 		atomic_inc(&qp->rwq_ind_tbl->usecnt);
1314 }
1315 EXPORT_SYMBOL(ib_qp_usecnt_inc);
1316 
1317 void ib_qp_usecnt_dec(struct ib_qp *qp)
1318 {
1319 	if (qp->rwq_ind_tbl)
1320 		atomic_dec(&qp->rwq_ind_tbl->usecnt);
1321 	if (qp->srq)
1322 		atomic_dec(&qp->srq->usecnt);
1323 	if (qp->recv_cq)
1324 		atomic_dec(&qp->recv_cq->usecnt);
1325 	if (qp->send_cq)
1326 		atomic_dec(&qp->send_cq->usecnt);
1327 	if (qp->pd)
1328 		atomic_dec(&qp->pd->usecnt);
1329 }
1330 EXPORT_SYMBOL(ib_qp_usecnt_dec);
1331 
1332 struct ib_qp *ib_create_qp_kernel(struct ib_pd *pd,
1333 				  struct ib_qp_init_attr *qp_init_attr,
1334 				  const char *caller)
1335 {
1336 	struct ib_device *device = pd->device;
1337 	struct ib_qp *qp;
1338 	int ret;
1339 
1340 	/*
1341 	 * If the callers is using the RDMA API calculate the resources
1342 	 * needed for the RDMA READ/WRITE operations.
1343 	 *
1344 	 * Note that these callers need to pass in a port number.
1345 	 */
1346 	if (qp_init_attr->cap.max_rdma_ctxs)
1347 		rdma_rw_init_qp(device, qp_init_attr);
1348 
1349 	qp = create_qp(device, pd, qp_init_attr, NULL, NULL, caller);
1350 	if (IS_ERR(qp))
1351 		return qp;
1352 
1353 	ib_qp_usecnt_inc(qp);
1354 
1355 	if (qp_init_attr->cap.max_rdma_ctxs) {
1356 		ret = rdma_rw_init_mrs(qp, qp_init_attr);
1357 		if (ret)
1358 			goto err;
1359 	}
1360 
1361 	/*
1362 	 * Note: all hw drivers guarantee that max_send_sge is lower than
1363 	 * the device RDMA WRITE SGE limit but not all hw drivers ensure that
1364 	 * max_send_sge <= max_sge_rd.
1365 	 */
1366 	qp->max_write_sge = qp_init_attr->cap.max_send_sge;
1367 	qp->max_read_sge = min_t(u32, qp_init_attr->cap.max_send_sge,
1368 				 device->attrs.max_sge_rd);
1369 	if (qp_init_attr->create_flags & IB_QP_CREATE_INTEGRITY_EN)
1370 		qp->integrity_en = true;
1371 
1372 	return qp;
1373 
1374 err:
1375 	ib_destroy_qp(qp);
1376 	return ERR_PTR(ret);
1377 
1378 }
1379 EXPORT_SYMBOL(ib_create_qp_kernel);
1380 
1381 static const struct {
1382 	int			valid;
1383 	enum ib_qp_attr_mask	req_param[IB_QPT_MAX];
1384 	enum ib_qp_attr_mask	opt_param[IB_QPT_MAX];
1385 } qp_state_table[IB_QPS_ERR + 1][IB_QPS_ERR + 1] = {
1386 	[IB_QPS_RESET] = {
1387 		[IB_QPS_RESET] = { .valid = 1 },
1388 		[IB_QPS_INIT]  = {
1389 			.valid = 1,
1390 			.req_param = {
1391 				[IB_QPT_UD]  = (IB_QP_PKEY_INDEX		|
1392 						IB_QP_PORT			|
1393 						IB_QP_QKEY),
1394 				[IB_QPT_RAW_PACKET] = IB_QP_PORT,
1395 				[IB_QPT_UC]  = (IB_QP_PKEY_INDEX		|
1396 						IB_QP_PORT			|
1397 						IB_QP_ACCESS_FLAGS),
1398 				[IB_QPT_RC]  = (IB_QP_PKEY_INDEX		|
1399 						IB_QP_PORT			|
1400 						IB_QP_ACCESS_FLAGS),
1401 				[IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX		|
1402 						IB_QP_PORT			|
1403 						IB_QP_ACCESS_FLAGS),
1404 				[IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX		|
1405 						IB_QP_PORT			|
1406 						IB_QP_ACCESS_FLAGS),
1407 				[IB_QPT_SMI] = (IB_QP_PKEY_INDEX		|
1408 						IB_QP_QKEY),
1409 				[IB_QPT_GSI] = (IB_QP_PKEY_INDEX		|
1410 						IB_QP_QKEY),
1411 			}
1412 		},
1413 	},
1414 	[IB_QPS_INIT]  = {
1415 		[IB_QPS_RESET] = { .valid = 1 },
1416 		[IB_QPS_ERR] =   { .valid = 1 },
1417 		[IB_QPS_INIT]  = {
1418 			.valid = 1,
1419 			.opt_param = {
1420 				[IB_QPT_UD]  = (IB_QP_PKEY_INDEX		|
1421 						IB_QP_PORT			|
1422 						IB_QP_QKEY),
1423 				[IB_QPT_UC]  = (IB_QP_PKEY_INDEX		|
1424 						IB_QP_PORT			|
1425 						IB_QP_ACCESS_FLAGS),
1426 				[IB_QPT_RC]  = (IB_QP_PKEY_INDEX		|
1427 						IB_QP_PORT			|
1428 						IB_QP_ACCESS_FLAGS),
1429 				[IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX		|
1430 						IB_QP_PORT			|
1431 						IB_QP_ACCESS_FLAGS),
1432 				[IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX		|
1433 						IB_QP_PORT			|
1434 						IB_QP_ACCESS_FLAGS),
1435 				[IB_QPT_SMI] = (IB_QP_PKEY_INDEX		|
1436 						IB_QP_QKEY),
1437 				[IB_QPT_GSI] = (IB_QP_PKEY_INDEX		|
1438 						IB_QP_QKEY),
1439 			}
1440 		},
1441 		[IB_QPS_RTR]   = {
1442 			.valid = 1,
1443 			.req_param = {
1444 				[IB_QPT_UC]  = (IB_QP_AV			|
1445 						IB_QP_PATH_MTU			|
1446 						IB_QP_DEST_QPN			|
1447 						IB_QP_RQ_PSN),
1448 				[IB_QPT_RC]  = (IB_QP_AV			|
1449 						IB_QP_PATH_MTU			|
1450 						IB_QP_DEST_QPN			|
1451 						IB_QP_RQ_PSN			|
1452 						IB_QP_MAX_DEST_RD_ATOMIC	|
1453 						IB_QP_MIN_RNR_TIMER),
1454 				[IB_QPT_XRC_INI] = (IB_QP_AV			|
1455 						IB_QP_PATH_MTU			|
1456 						IB_QP_DEST_QPN			|
1457 						IB_QP_RQ_PSN),
1458 				[IB_QPT_XRC_TGT] = (IB_QP_AV			|
1459 						IB_QP_PATH_MTU			|
1460 						IB_QP_DEST_QPN			|
1461 						IB_QP_RQ_PSN			|
1462 						IB_QP_MAX_DEST_RD_ATOMIC	|
1463 						IB_QP_MIN_RNR_TIMER),
1464 			},
1465 			.opt_param = {
1466 				 [IB_QPT_UD]  = (IB_QP_PKEY_INDEX		|
1467 						 IB_QP_QKEY),
1468 				 [IB_QPT_UC]  = (IB_QP_ALT_PATH			|
1469 						 IB_QP_ACCESS_FLAGS		|
1470 						 IB_QP_PKEY_INDEX),
1471 				 [IB_QPT_RC]  = (IB_QP_ALT_PATH			|
1472 						 IB_QP_ACCESS_FLAGS		|
1473 						 IB_QP_PKEY_INDEX),
1474 				 [IB_QPT_XRC_INI] = (IB_QP_ALT_PATH		|
1475 						 IB_QP_ACCESS_FLAGS		|
1476 						 IB_QP_PKEY_INDEX),
1477 				 [IB_QPT_XRC_TGT] = (IB_QP_ALT_PATH		|
1478 						 IB_QP_ACCESS_FLAGS		|
1479 						 IB_QP_PKEY_INDEX),
1480 				 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX		|
1481 						 IB_QP_QKEY),
1482 				 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX		|
1483 						 IB_QP_QKEY),
1484 			 },
1485 		},
1486 	},
1487 	[IB_QPS_RTR]   = {
1488 		[IB_QPS_RESET] = { .valid = 1 },
1489 		[IB_QPS_ERR] =   { .valid = 1 },
1490 		[IB_QPS_RTS]   = {
1491 			.valid = 1,
1492 			.req_param = {
1493 				[IB_QPT_UD]  = IB_QP_SQ_PSN,
1494 				[IB_QPT_UC]  = IB_QP_SQ_PSN,
1495 				[IB_QPT_RC]  = (IB_QP_TIMEOUT			|
1496 						IB_QP_RETRY_CNT			|
1497 						IB_QP_RNR_RETRY			|
1498 						IB_QP_SQ_PSN			|
1499 						IB_QP_MAX_QP_RD_ATOMIC),
1500 				[IB_QPT_XRC_INI] = (IB_QP_TIMEOUT		|
1501 						IB_QP_RETRY_CNT			|
1502 						IB_QP_RNR_RETRY			|
1503 						IB_QP_SQ_PSN			|
1504 						IB_QP_MAX_QP_RD_ATOMIC),
1505 				[IB_QPT_XRC_TGT] = (IB_QP_TIMEOUT		|
1506 						IB_QP_SQ_PSN),
1507 				[IB_QPT_SMI] = IB_QP_SQ_PSN,
1508 				[IB_QPT_GSI] = IB_QP_SQ_PSN,
1509 			},
1510 			.opt_param = {
1511 				 [IB_QPT_UD]  = (IB_QP_CUR_STATE		|
1512 						 IB_QP_QKEY),
1513 				 [IB_QPT_UC]  = (IB_QP_CUR_STATE		|
1514 						 IB_QP_ALT_PATH			|
1515 						 IB_QP_ACCESS_FLAGS		|
1516 						 IB_QP_PATH_MIG_STATE),
1517 				 [IB_QPT_RC]  = (IB_QP_CUR_STATE		|
1518 						 IB_QP_ALT_PATH			|
1519 						 IB_QP_ACCESS_FLAGS		|
1520 						 IB_QP_MIN_RNR_TIMER		|
1521 						 IB_QP_PATH_MIG_STATE),
1522 				 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE		|
1523 						 IB_QP_ALT_PATH			|
1524 						 IB_QP_ACCESS_FLAGS		|
1525 						 IB_QP_PATH_MIG_STATE),
1526 				 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE		|
1527 						 IB_QP_ALT_PATH			|
1528 						 IB_QP_ACCESS_FLAGS		|
1529 						 IB_QP_MIN_RNR_TIMER		|
1530 						 IB_QP_PATH_MIG_STATE),
1531 				 [IB_QPT_SMI] = (IB_QP_CUR_STATE		|
1532 						 IB_QP_QKEY),
1533 				 [IB_QPT_GSI] = (IB_QP_CUR_STATE		|
1534 						 IB_QP_QKEY),
1535 				 [IB_QPT_RAW_PACKET] = IB_QP_RATE_LIMIT,
1536 			 }
1537 		}
1538 	},
1539 	[IB_QPS_RTS]   = {
1540 		[IB_QPS_RESET] = { .valid = 1 },
1541 		[IB_QPS_ERR] =   { .valid = 1 },
1542 		[IB_QPS_RTS]   = {
1543 			.valid = 1,
1544 			.opt_param = {
1545 				[IB_QPT_UD]  = (IB_QP_CUR_STATE			|
1546 						IB_QP_QKEY),
1547 				[IB_QPT_UC]  = (IB_QP_CUR_STATE			|
1548 						IB_QP_ACCESS_FLAGS		|
1549 						IB_QP_ALT_PATH			|
1550 						IB_QP_PATH_MIG_STATE),
1551 				[IB_QPT_RC]  = (IB_QP_CUR_STATE			|
1552 						IB_QP_ACCESS_FLAGS		|
1553 						IB_QP_ALT_PATH			|
1554 						IB_QP_PATH_MIG_STATE		|
1555 						IB_QP_MIN_RNR_TIMER),
1556 				[IB_QPT_XRC_INI] = (IB_QP_CUR_STATE		|
1557 						IB_QP_ACCESS_FLAGS		|
1558 						IB_QP_ALT_PATH			|
1559 						IB_QP_PATH_MIG_STATE),
1560 				[IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE		|
1561 						IB_QP_ACCESS_FLAGS		|
1562 						IB_QP_ALT_PATH			|
1563 						IB_QP_PATH_MIG_STATE		|
1564 						IB_QP_MIN_RNR_TIMER),
1565 				[IB_QPT_SMI] = (IB_QP_CUR_STATE			|
1566 						IB_QP_QKEY),
1567 				[IB_QPT_GSI] = (IB_QP_CUR_STATE			|
1568 						IB_QP_QKEY),
1569 				[IB_QPT_RAW_PACKET] = IB_QP_RATE_LIMIT,
1570 			}
1571 		},
1572 		[IB_QPS_SQD]   = {
1573 			.valid = 1,
1574 			.opt_param = {
1575 				[IB_QPT_UD]  = IB_QP_EN_SQD_ASYNC_NOTIFY,
1576 				[IB_QPT_UC]  = IB_QP_EN_SQD_ASYNC_NOTIFY,
1577 				[IB_QPT_RC]  = IB_QP_EN_SQD_ASYNC_NOTIFY,
1578 				[IB_QPT_XRC_INI] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1579 				[IB_QPT_XRC_TGT] = IB_QP_EN_SQD_ASYNC_NOTIFY, /* ??? */
1580 				[IB_QPT_SMI] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1581 				[IB_QPT_GSI] = IB_QP_EN_SQD_ASYNC_NOTIFY
1582 			}
1583 		},
1584 	},
1585 	[IB_QPS_SQD]   = {
1586 		[IB_QPS_RESET] = { .valid = 1 },
1587 		[IB_QPS_ERR] =   { .valid = 1 },
1588 		[IB_QPS_RTS]   = {
1589 			.valid = 1,
1590 			.opt_param = {
1591 				[IB_QPT_UD]  = (IB_QP_CUR_STATE			|
1592 						IB_QP_QKEY),
1593 				[IB_QPT_UC]  = (IB_QP_CUR_STATE			|
1594 						IB_QP_ALT_PATH			|
1595 						IB_QP_ACCESS_FLAGS		|
1596 						IB_QP_PATH_MIG_STATE),
1597 				[IB_QPT_RC]  = (IB_QP_CUR_STATE			|
1598 						IB_QP_ALT_PATH			|
1599 						IB_QP_ACCESS_FLAGS		|
1600 						IB_QP_MIN_RNR_TIMER		|
1601 						IB_QP_PATH_MIG_STATE),
1602 				[IB_QPT_XRC_INI] = (IB_QP_CUR_STATE		|
1603 						IB_QP_ALT_PATH			|
1604 						IB_QP_ACCESS_FLAGS		|
1605 						IB_QP_PATH_MIG_STATE),
1606 				[IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE		|
1607 						IB_QP_ALT_PATH			|
1608 						IB_QP_ACCESS_FLAGS		|
1609 						IB_QP_MIN_RNR_TIMER		|
1610 						IB_QP_PATH_MIG_STATE),
1611 				[IB_QPT_SMI] = (IB_QP_CUR_STATE			|
1612 						IB_QP_QKEY),
1613 				[IB_QPT_GSI] = (IB_QP_CUR_STATE			|
1614 						IB_QP_QKEY),
1615 			}
1616 		},
1617 		[IB_QPS_SQD]   = {
1618 			.valid = 1,
1619 			.opt_param = {
1620 				[IB_QPT_UD]  = (IB_QP_PKEY_INDEX		|
1621 						IB_QP_QKEY),
1622 				[IB_QPT_UC]  = (IB_QP_AV			|
1623 						IB_QP_ALT_PATH			|
1624 						IB_QP_ACCESS_FLAGS		|
1625 						IB_QP_PKEY_INDEX		|
1626 						IB_QP_PATH_MIG_STATE),
1627 				[IB_QPT_RC]  = (IB_QP_PORT			|
1628 						IB_QP_AV			|
1629 						IB_QP_TIMEOUT			|
1630 						IB_QP_RETRY_CNT			|
1631 						IB_QP_RNR_RETRY			|
1632 						IB_QP_MAX_QP_RD_ATOMIC		|
1633 						IB_QP_MAX_DEST_RD_ATOMIC	|
1634 						IB_QP_ALT_PATH			|
1635 						IB_QP_ACCESS_FLAGS		|
1636 						IB_QP_PKEY_INDEX		|
1637 						IB_QP_MIN_RNR_TIMER		|
1638 						IB_QP_PATH_MIG_STATE),
1639 				[IB_QPT_XRC_INI] = (IB_QP_PORT			|
1640 						IB_QP_AV			|
1641 						IB_QP_TIMEOUT			|
1642 						IB_QP_RETRY_CNT			|
1643 						IB_QP_RNR_RETRY			|
1644 						IB_QP_MAX_QP_RD_ATOMIC		|
1645 						IB_QP_ALT_PATH			|
1646 						IB_QP_ACCESS_FLAGS		|
1647 						IB_QP_PKEY_INDEX		|
1648 						IB_QP_PATH_MIG_STATE),
1649 				[IB_QPT_XRC_TGT] = (IB_QP_PORT			|
1650 						IB_QP_AV			|
1651 						IB_QP_TIMEOUT			|
1652 						IB_QP_MAX_DEST_RD_ATOMIC	|
1653 						IB_QP_ALT_PATH			|
1654 						IB_QP_ACCESS_FLAGS		|
1655 						IB_QP_PKEY_INDEX		|
1656 						IB_QP_MIN_RNR_TIMER		|
1657 						IB_QP_PATH_MIG_STATE),
1658 				[IB_QPT_SMI] = (IB_QP_PKEY_INDEX		|
1659 						IB_QP_QKEY),
1660 				[IB_QPT_GSI] = (IB_QP_PKEY_INDEX		|
1661 						IB_QP_QKEY),
1662 			}
1663 		}
1664 	},
1665 	[IB_QPS_SQE]   = {
1666 		[IB_QPS_RESET] = { .valid = 1 },
1667 		[IB_QPS_ERR] =   { .valid = 1 },
1668 		[IB_QPS_RTS]   = {
1669 			.valid = 1,
1670 			.opt_param = {
1671 				[IB_QPT_UD]  = (IB_QP_CUR_STATE			|
1672 						IB_QP_QKEY),
1673 				[IB_QPT_UC]  = (IB_QP_CUR_STATE			|
1674 						IB_QP_ACCESS_FLAGS),
1675 				[IB_QPT_SMI] = (IB_QP_CUR_STATE			|
1676 						IB_QP_QKEY),
1677 				[IB_QPT_GSI] = (IB_QP_CUR_STATE			|
1678 						IB_QP_QKEY),
1679 			}
1680 		}
1681 	},
1682 	[IB_QPS_ERR] = {
1683 		[IB_QPS_RESET] = { .valid = 1 },
1684 		[IB_QPS_ERR] =   { .valid = 1 }
1685 	}
1686 };
1687 
1688 bool ib_modify_qp_is_ok(enum ib_qp_state cur_state, enum ib_qp_state next_state,
1689 			enum ib_qp_type type, enum ib_qp_attr_mask mask)
1690 {
1691 	enum ib_qp_attr_mask req_param, opt_param;
1692 
1693 	if (mask & IB_QP_CUR_STATE  &&
1694 	    cur_state != IB_QPS_RTR && cur_state != IB_QPS_RTS &&
1695 	    cur_state != IB_QPS_SQD && cur_state != IB_QPS_SQE)
1696 		return false;
1697 
1698 	if (!qp_state_table[cur_state][next_state].valid)
1699 		return false;
1700 
1701 	req_param = qp_state_table[cur_state][next_state].req_param[type];
1702 	opt_param = qp_state_table[cur_state][next_state].opt_param[type];
1703 
1704 	if ((mask & req_param) != req_param)
1705 		return false;
1706 
1707 	if (mask & ~(req_param | opt_param | IB_QP_STATE))
1708 		return false;
1709 
1710 	return true;
1711 }
1712 EXPORT_SYMBOL(ib_modify_qp_is_ok);
1713 
1714 /**
1715  * ib_resolve_eth_dmac - Resolve destination mac address
1716  * @device:		Device to consider
1717  * @ah_attr:		address handle attribute which describes the
1718  *			source and destination parameters
1719  * ib_resolve_eth_dmac() resolves destination mac address and L3 hop limit It
1720  * returns 0 on success or appropriate error code. It initializes the
1721  * necessary ah_attr fields when call is successful.
1722  */
1723 static int ib_resolve_eth_dmac(struct ib_device *device,
1724 			       struct rdma_ah_attr *ah_attr)
1725 {
1726 	int ret = 0;
1727 
1728 	if (rdma_is_multicast_addr((struct in6_addr *)ah_attr->grh.dgid.raw)) {
1729 		if (ipv6_addr_v4mapped((struct in6_addr *)ah_attr->grh.dgid.raw)) {
1730 			__be32 addr = 0;
1731 
1732 			memcpy(&addr, ah_attr->grh.dgid.raw + 12, 4);
1733 			ip_eth_mc_map(addr, (char *)ah_attr->roce.dmac);
1734 		} else {
1735 			ipv6_eth_mc_map((struct in6_addr *)ah_attr->grh.dgid.raw,
1736 					(char *)ah_attr->roce.dmac);
1737 		}
1738 	} else {
1739 		ret = ib_resolve_unicast_gid_dmac(device, ah_attr);
1740 	}
1741 	return ret;
1742 }
1743 
1744 static bool is_qp_type_connected(const struct ib_qp *qp)
1745 {
1746 	return (qp->qp_type == IB_QPT_UC ||
1747 		qp->qp_type == IB_QPT_RC ||
1748 		qp->qp_type == IB_QPT_XRC_INI ||
1749 		qp->qp_type == IB_QPT_XRC_TGT);
1750 }
1751 
1752 /*
1753  * IB core internal function to perform QP attributes modification.
1754  */
1755 static int _ib_modify_qp(struct ib_qp *qp, struct ib_qp_attr *attr,
1756 			 int attr_mask, struct ib_udata *udata)
1757 {
1758 	u32 port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
1759 	const struct ib_gid_attr *old_sgid_attr_av;
1760 	const struct ib_gid_attr *old_sgid_attr_alt_av;
1761 	int ret;
1762 
1763 	attr->xmit_slave = NULL;
1764 	if (attr_mask & IB_QP_AV) {
1765 		ret = rdma_fill_sgid_attr(qp->device, &attr->ah_attr,
1766 					  &old_sgid_attr_av);
1767 		if (ret)
1768 			return ret;
1769 
1770 		if (attr->ah_attr.type == RDMA_AH_ATTR_TYPE_ROCE &&
1771 		    is_qp_type_connected(qp)) {
1772 			struct net_device *slave;
1773 
1774 			/*
1775 			 * If the user provided the qp_attr then we have to
1776 			 * resolve it. Kerne users have to provide already
1777 			 * resolved rdma_ah_attr's.
1778 			 */
1779 			if (udata) {
1780 				ret = ib_resolve_eth_dmac(qp->device,
1781 							  &attr->ah_attr);
1782 				if (ret)
1783 					goto out_av;
1784 			}
1785 			slave = rdma_lag_get_ah_roce_slave(qp->device,
1786 							   &attr->ah_attr,
1787 							   GFP_KERNEL);
1788 			if (IS_ERR(slave)) {
1789 				ret = PTR_ERR(slave);
1790 				goto out_av;
1791 			}
1792 			attr->xmit_slave = slave;
1793 		}
1794 	}
1795 	if (attr_mask & IB_QP_ALT_PATH) {
1796 		/*
1797 		 * FIXME: This does not track the migration state, so if the
1798 		 * user loads a new alternate path after the HW has migrated
1799 		 * from primary->alternate we will keep the wrong
1800 		 * references. This is OK for IB because the reference
1801 		 * counting does not serve any functional purpose.
1802 		 */
1803 		ret = rdma_fill_sgid_attr(qp->device, &attr->alt_ah_attr,
1804 					  &old_sgid_attr_alt_av);
1805 		if (ret)
1806 			goto out_av;
1807 
1808 		/*
1809 		 * Today the core code can only handle alternate paths and APM
1810 		 * for IB. Ban them in roce mode.
1811 		 */
1812 		if (!(rdma_protocol_ib(qp->device,
1813 				       attr->alt_ah_attr.port_num) &&
1814 		      rdma_protocol_ib(qp->device, port))) {
1815 			ret = -EINVAL;
1816 			goto out;
1817 		}
1818 	}
1819 
1820 	if (rdma_ib_or_roce(qp->device, port)) {
1821 		if (attr_mask & IB_QP_RQ_PSN && attr->rq_psn & ~0xffffff) {
1822 			dev_warn(&qp->device->dev,
1823 				 "%s rq_psn overflow, masking to 24 bits\n",
1824 				 __func__);
1825 			attr->rq_psn &= 0xffffff;
1826 		}
1827 
1828 		if (attr_mask & IB_QP_SQ_PSN && attr->sq_psn & ~0xffffff) {
1829 			dev_warn(&qp->device->dev,
1830 				 " %s sq_psn overflow, masking to 24 bits\n",
1831 				 __func__);
1832 			attr->sq_psn &= 0xffffff;
1833 		}
1834 	}
1835 
1836 	/*
1837 	 * Bind this qp to a counter automatically based on the rdma counter
1838 	 * rules. This only set in RST2INIT with port specified
1839 	 */
1840 	if (!qp->counter && (attr_mask & IB_QP_PORT) &&
1841 	    ((attr_mask & IB_QP_STATE) && attr->qp_state == IB_QPS_INIT))
1842 		rdma_counter_bind_qp_auto(qp, attr->port_num);
1843 
1844 	ret = ib_security_modify_qp(qp, attr, attr_mask, udata);
1845 	if (ret)
1846 		goto out;
1847 
1848 	if (attr_mask & IB_QP_PORT)
1849 		qp->port = attr->port_num;
1850 	if (attr_mask & IB_QP_AV)
1851 		qp->av_sgid_attr =
1852 			rdma_update_sgid_attr(&attr->ah_attr, qp->av_sgid_attr);
1853 	if (attr_mask & IB_QP_ALT_PATH)
1854 		qp->alt_path_sgid_attr = rdma_update_sgid_attr(
1855 			&attr->alt_ah_attr, qp->alt_path_sgid_attr);
1856 
1857 out:
1858 	if (attr_mask & IB_QP_ALT_PATH)
1859 		rdma_unfill_sgid_attr(&attr->alt_ah_attr, old_sgid_attr_alt_av);
1860 out_av:
1861 	if (attr_mask & IB_QP_AV) {
1862 		rdma_lag_put_ah_roce_slave(attr->xmit_slave);
1863 		rdma_unfill_sgid_attr(&attr->ah_attr, old_sgid_attr_av);
1864 	}
1865 	return ret;
1866 }
1867 
1868 /**
1869  * ib_modify_qp_with_udata - Modifies the attributes for the specified QP.
1870  * @ib_qp: The QP to modify.
1871  * @attr: On input, specifies the QP attributes to modify.  On output,
1872  *   the current values of selected QP attributes are returned.
1873  * @attr_mask: A bit-mask used to specify which attributes of the QP
1874  *   are being modified.
1875  * @udata: pointer to user's input output buffer information
1876  *   are being modified.
1877  * It returns 0 on success and returns appropriate error code on error.
1878  */
1879 int ib_modify_qp_with_udata(struct ib_qp *ib_qp, struct ib_qp_attr *attr,
1880 			    int attr_mask, struct ib_udata *udata)
1881 {
1882 	return _ib_modify_qp(ib_qp->real_qp, attr, attr_mask, udata);
1883 }
1884 EXPORT_SYMBOL(ib_modify_qp_with_udata);
1885 
1886 static void ib_get_width_and_speed(u32 netdev_speed, u32 lanes,
1887 				   u16 *speed, u8 *width)
1888 {
1889 	if (!lanes) {
1890 		if (netdev_speed <= SPEED_1000) {
1891 			*width = IB_WIDTH_1X;
1892 			*speed = IB_SPEED_SDR;
1893 		} else if (netdev_speed <= SPEED_10000) {
1894 			*width = IB_WIDTH_1X;
1895 			*speed = IB_SPEED_FDR10;
1896 		} else if (netdev_speed <= SPEED_20000) {
1897 			*width = IB_WIDTH_4X;
1898 			*speed = IB_SPEED_DDR;
1899 		} else if (netdev_speed <= SPEED_25000) {
1900 			*width = IB_WIDTH_1X;
1901 			*speed = IB_SPEED_EDR;
1902 		} else if (netdev_speed <= SPEED_40000) {
1903 			*width = IB_WIDTH_4X;
1904 			*speed = IB_SPEED_FDR10;
1905 		} else if (netdev_speed <= SPEED_50000) {
1906 			*width = IB_WIDTH_2X;
1907 			*speed = IB_SPEED_EDR;
1908 		} else if (netdev_speed <= SPEED_100000) {
1909 			*width = IB_WIDTH_4X;
1910 			*speed = IB_SPEED_EDR;
1911 		} else if (netdev_speed <= SPEED_200000) {
1912 			*width = IB_WIDTH_4X;
1913 			*speed = IB_SPEED_HDR;
1914 		} else {
1915 			*width = IB_WIDTH_4X;
1916 			*speed = IB_SPEED_NDR;
1917 		}
1918 
1919 		return;
1920 	}
1921 
1922 	switch (lanes) {
1923 	case 1:
1924 		*width = IB_WIDTH_1X;
1925 		break;
1926 	case 2:
1927 		*width = IB_WIDTH_2X;
1928 		break;
1929 	case 4:
1930 		*width = IB_WIDTH_4X;
1931 		break;
1932 	case 8:
1933 		*width = IB_WIDTH_8X;
1934 		break;
1935 	case 12:
1936 		*width = IB_WIDTH_12X;
1937 		break;
1938 	default:
1939 		*width = IB_WIDTH_1X;
1940 	}
1941 
1942 	switch (netdev_speed / lanes) {
1943 	case SPEED_2500:
1944 		*speed = IB_SPEED_SDR;
1945 		break;
1946 	case SPEED_5000:
1947 		*speed = IB_SPEED_DDR;
1948 		break;
1949 	case SPEED_10000:
1950 		*speed = IB_SPEED_FDR10;
1951 		break;
1952 	case SPEED_14000:
1953 		*speed = IB_SPEED_FDR;
1954 		break;
1955 	case SPEED_25000:
1956 		*speed = IB_SPEED_EDR;
1957 		break;
1958 	case SPEED_50000:
1959 		*speed = IB_SPEED_HDR;
1960 		break;
1961 	case SPEED_100000:
1962 		*speed = IB_SPEED_NDR;
1963 		break;
1964 	default:
1965 		*speed = IB_SPEED_SDR;
1966 	}
1967 }
1968 
1969 int ib_get_eth_speed(struct ib_device *dev, u32 port_num, u16 *speed, u8 *width)
1970 {
1971 	int rc;
1972 	u32 netdev_speed;
1973 	struct net_device *netdev;
1974 	struct ethtool_link_ksettings lksettings = {};
1975 
1976 	if (rdma_port_get_link_layer(dev, port_num) != IB_LINK_LAYER_ETHERNET)
1977 		return -EINVAL;
1978 
1979 	netdev = ib_device_get_netdev(dev, port_num);
1980 	if (!netdev)
1981 		return -ENODEV;
1982 
1983 	rtnl_lock();
1984 	rc = __ethtool_get_link_ksettings(netdev, &lksettings);
1985 	rtnl_unlock();
1986 
1987 	dev_put(netdev);
1988 
1989 	if (!rc && lksettings.base.speed != (u32)SPEED_UNKNOWN) {
1990 		netdev_speed = lksettings.base.speed;
1991 	} else {
1992 		netdev_speed = SPEED_1000;
1993 		if (rc)
1994 			pr_warn("%s speed is unknown, defaulting to %u\n",
1995 				netdev->name, netdev_speed);
1996 	}
1997 
1998 	ib_get_width_and_speed(netdev_speed, lksettings.lanes,
1999 			       speed, width);
2000 
2001 	return 0;
2002 }
2003 EXPORT_SYMBOL(ib_get_eth_speed);
2004 
2005 int ib_modify_qp(struct ib_qp *qp,
2006 		 struct ib_qp_attr *qp_attr,
2007 		 int qp_attr_mask)
2008 {
2009 	return _ib_modify_qp(qp->real_qp, qp_attr, qp_attr_mask, NULL);
2010 }
2011 EXPORT_SYMBOL(ib_modify_qp);
2012 
2013 int ib_query_qp(struct ib_qp *qp,
2014 		struct ib_qp_attr *qp_attr,
2015 		int qp_attr_mask,
2016 		struct ib_qp_init_attr *qp_init_attr)
2017 {
2018 	qp_attr->ah_attr.grh.sgid_attr = NULL;
2019 	qp_attr->alt_ah_attr.grh.sgid_attr = NULL;
2020 
2021 	return qp->device->ops.query_qp ?
2022 		qp->device->ops.query_qp(qp->real_qp, qp_attr, qp_attr_mask,
2023 					 qp_init_attr) : -EOPNOTSUPP;
2024 }
2025 EXPORT_SYMBOL(ib_query_qp);
2026 
2027 int ib_close_qp(struct ib_qp *qp)
2028 {
2029 	struct ib_qp *real_qp;
2030 	unsigned long flags;
2031 
2032 	real_qp = qp->real_qp;
2033 	if (real_qp == qp)
2034 		return -EINVAL;
2035 
2036 	spin_lock_irqsave(&real_qp->device->qp_open_list_lock, flags);
2037 	list_del(&qp->open_list);
2038 	spin_unlock_irqrestore(&real_qp->device->qp_open_list_lock, flags);
2039 
2040 	atomic_dec(&real_qp->usecnt);
2041 	if (qp->qp_sec)
2042 		ib_close_shared_qp_security(qp->qp_sec);
2043 	kfree(qp);
2044 
2045 	return 0;
2046 }
2047 EXPORT_SYMBOL(ib_close_qp);
2048 
2049 static int __ib_destroy_shared_qp(struct ib_qp *qp)
2050 {
2051 	struct ib_xrcd *xrcd;
2052 	struct ib_qp *real_qp;
2053 	int ret;
2054 
2055 	real_qp = qp->real_qp;
2056 	xrcd = real_qp->xrcd;
2057 	down_write(&xrcd->tgt_qps_rwsem);
2058 	ib_close_qp(qp);
2059 	if (atomic_read(&real_qp->usecnt) == 0)
2060 		xa_erase(&xrcd->tgt_qps, real_qp->qp_num);
2061 	else
2062 		real_qp = NULL;
2063 	up_write(&xrcd->tgt_qps_rwsem);
2064 
2065 	if (real_qp) {
2066 		ret = ib_destroy_qp(real_qp);
2067 		if (!ret)
2068 			atomic_dec(&xrcd->usecnt);
2069 	}
2070 
2071 	return 0;
2072 }
2073 
2074 int ib_destroy_qp_user(struct ib_qp *qp, struct ib_udata *udata)
2075 {
2076 	const struct ib_gid_attr *alt_path_sgid_attr = qp->alt_path_sgid_attr;
2077 	const struct ib_gid_attr *av_sgid_attr = qp->av_sgid_attr;
2078 	struct ib_qp_security *sec;
2079 	int ret;
2080 
2081 	WARN_ON_ONCE(qp->mrs_used > 0);
2082 
2083 	if (atomic_read(&qp->usecnt))
2084 		return -EBUSY;
2085 
2086 	if (qp->real_qp != qp)
2087 		return __ib_destroy_shared_qp(qp);
2088 
2089 	sec  = qp->qp_sec;
2090 	if (sec)
2091 		ib_destroy_qp_security_begin(sec);
2092 
2093 	if (!qp->uobject)
2094 		rdma_rw_cleanup_mrs(qp);
2095 
2096 	rdma_counter_unbind_qp(qp, true);
2097 	ret = qp->device->ops.destroy_qp(qp, udata);
2098 	if (ret) {
2099 		if (sec)
2100 			ib_destroy_qp_security_abort(sec);
2101 		return ret;
2102 	}
2103 
2104 	if (alt_path_sgid_attr)
2105 		rdma_put_gid_attr(alt_path_sgid_attr);
2106 	if (av_sgid_attr)
2107 		rdma_put_gid_attr(av_sgid_attr);
2108 
2109 	ib_qp_usecnt_dec(qp);
2110 	if (sec)
2111 		ib_destroy_qp_security_end(sec);
2112 
2113 	rdma_restrack_del(&qp->res);
2114 	kfree(qp);
2115 	return ret;
2116 }
2117 EXPORT_SYMBOL(ib_destroy_qp_user);
2118 
2119 /* Completion queues */
2120 
2121 struct ib_cq *__ib_create_cq(struct ib_device *device,
2122 			     ib_comp_handler comp_handler,
2123 			     void (*event_handler)(struct ib_event *, void *),
2124 			     void *cq_context,
2125 			     const struct ib_cq_init_attr *cq_attr,
2126 			     const char *caller)
2127 {
2128 	struct ib_cq *cq;
2129 	int ret;
2130 
2131 	cq = rdma_zalloc_drv_obj(device, ib_cq);
2132 	if (!cq)
2133 		return ERR_PTR(-ENOMEM);
2134 
2135 	cq->device = device;
2136 	cq->uobject = NULL;
2137 	cq->comp_handler = comp_handler;
2138 	cq->event_handler = event_handler;
2139 	cq->cq_context = cq_context;
2140 	atomic_set(&cq->usecnt, 0);
2141 
2142 	rdma_restrack_new(&cq->res, RDMA_RESTRACK_CQ);
2143 	rdma_restrack_set_name(&cq->res, caller);
2144 
2145 	ret = device->ops.create_cq(cq, cq_attr, NULL);
2146 	if (ret) {
2147 		rdma_restrack_put(&cq->res);
2148 		kfree(cq);
2149 		return ERR_PTR(ret);
2150 	}
2151 
2152 	rdma_restrack_add(&cq->res);
2153 	return cq;
2154 }
2155 EXPORT_SYMBOL(__ib_create_cq);
2156 
2157 int rdma_set_cq_moderation(struct ib_cq *cq, u16 cq_count, u16 cq_period)
2158 {
2159 	if (cq->shared)
2160 		return -EOPNOTSUPP;
2161 
2162 	return cq->device->ops.modify_cq ?
2163 		cq->device->ops.modify_cq(cq, cq_count,
2164 					  cq_period) : -EOPNOTSUPP;
2165 }
2166 EXPORT_SYMBOL(rdma_set_cq_moderation);
2167 
2168 int ib_destroy_cq_user(struct ib_cq *cq, struct ib_udata *udata)
2169 {
2170 	int ret;
2171 
2172 	if (WARN_ON_ONCE(cq->shared))
2173 		return -EOPNOTSUPP;
2174 
2175 	if (atomic_read(&cq->usecnt))
2176 		return -EBUSY;
2177 
2178 	ret = cq->device->ops.destroy_cq(cq, udata);
2179 	if (ret)
2180 		return ret;
2181 
2182 	rdma_restrack_del(&cq->res);
2183 	kfree(cq);
2184 	return ret;
2185 }
2186 EXPORT_SYMBOL(ib_destroy_cq_user);
2187 
2188 int ib_resize_cq(struct ib_cq *cq, int cqe)
2189 {
2190 	if (cq->shared)
2191 		return -EOPNOTSUPP;
2192 
2193 	return cq->device->ops.resize_cq ?
2194 		cq->device->ops.resize_cq(cq, cqe, NULL) : -EOPNOTSUPP;
2195 }
2196 EXPORT_SYMBOL(ib_resize_cq);
2197 
2198 /* Memory regions */
2199 
2200 struct ib_mr *ib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
2201 			     u64 virt_addr, int access_flags)
2202 {
2203 	struct ib_mr *mr;
2204 
2205 	if (access_flags & IB_ACCESS_ON_DEMAND) {
2206 		if (!(pd->device->attrs.kernel_cap_flags &
2207 		      IBK_ON_DEMAND_PAGING)) {
2208 			pr_debug("ODP support not available\n");
2209 			return ERR_PTR(-EINVAL);
2210 		}
2211 	}
2212 
2213 	mr = pd->device->ops.reg_user_mr(pd, start, length, virt_addr,
2214 					 access_flags, NULL);
2215 
2216 	if (IS_ERR(mr))
2217 		return mr;
2218 
2219 	mr->device = pd->device;
2220 	mr->type = IB_MR_TYPE_USER;
2221 	mr->pd = pd;
2222 	mr->dm = NULL;
2223 	atomic_inc(&pd->usecnt);
2224 	mr->iova =  virt_addr;
2225 	mr->length = length;
2226 
2227 	rdma_restrack_new(&mr->res, RDMA_RESTRACK_MR);
2228 	rdma_restrack_parent_name(&mr->res, &pd->res);
2229 	rdma_restrack_add(&mr->res);
2230 
2231 	return mr;
2232 }
2233 EXPORT_SYMBOL(ib_reg_user_mr);
2234 
2235 int ib_advise_mr(struct ib_pd *pd, enum ib_uverbs_advise_mr_advice advice,
2236 		 u32 flags, struct ib_sge *sg_list, u32 num_sge)
2237 {
2238 	if (!pd->device->ops.advise_mr)
2239 		return -EOPNOTSUPP;
2240 
2241 	if (!num_sge)
2242 		return 0;
2243 
2244 	return pd->device->ops.advise_mr(pd, advice, flags, sg_list, num_sge,
2245 					 NULL);
2246 }
2247 EXPORT_SYMBOL(ib_advise_mr);
2248 
2249 int ib_dereg_mr_user(struct ib_mr *mr, struct ib_udata *udata)
2250 {
2251 	struct ib_pd *pd = mr->pd;
2252 	struct ib_dm *dm = mr->dm;
2253 	struct ib_sig_attrs *sig_attrs = mr->sig_attrs;
2254 	int ret;
2255 
2256 	trace_mr_dereg(mr);
2257 	rdma_restrack_del(&mr->res);
2258 	ret = mr->device->ops.dereg_mr(mr, udata);
2259 	if (!ret) {
2260 		atomic_dec(&pd->usecnt);
2261 		if (dm)
2262 			atomic_dec(&dm->usecnt);
2263 		kfree(sig_attrs);
2264 	}
2265 
2266 	return ret;
2267 }
2268 EXPORT_SYMBOL(ib_dereg_mr_user);
2269 
2270 /**
2271  * ib_alloc_mr() - Allocates a memory region
2272  * @pd:            protection domain associated with the region
2273  * @mr_type:       memory region type
2274  * @max_num_sg:    maximum sg entries available for registration.
2275  *
2276  * Notes:
2277  * Memory registeration page/sg lists must not exceed max_num_sg.
2278  * For mr_type IB_MR_TYPE_MEM_REG, the total length cannot exceed
2279  * max_num_sg * used_page_size.
2280  *
2281  */
2282 struct ib_mr *ib_alloc_mr(struct ib_pd *pd, enum ib_mr_type mr_type,
2283 			  u32 max_num_sg)
2284 {
2285 	struct ib_mr *mr;
2286 
2287 	if (!pd->device->ops.alloc_mr) {
2288 		mr = ERR_PTR(-EOPNOTSUPP);
2289 		goto out;
2290 	}
2291 
2292 	if (mr_type == IB_MR_TYPE_INTEGRITY) {
2293 		WARN_ON_ONCE(1);
2294 		mr = ERR_PTR(-EINVAL);
2295 		goto out;
2296 	}
2297 
2298 	mr = pd->device->ops.alloc_mr(pd, mr_type, max_num_sg);
2299 	if (IS_ERR(mr))
2300 		goto out;
2301 
2302 	mr->device = pd->device;
2303 	mr->pd = pd;
2304 	mr->dm = NULL;
2305 	mr->uobject = NULL;
2306 	atomic_inc(&pd->usecnt);
2307 	mr->need_inval = false;
2308 	mr->type = mr_type;
2309 	mr->sig_attrs = NULL;
2310 
2311 	rdma_restrack_new(&mr->res, RDMA_RESTRACK_MR);
2312 	rdma_restrack_parent_name(&mr->res, &pd->res);
2313 	rdma_restrack_add(&mr->res);
2314 out:
2315 	trace_mr_alloc(pd, mr_type, max_num_sg, mr);
2316 	return mr;
2317 }
2318 EXPORT_SYMBOL(ib_alloc_mr);
2319 
2320 /**
2321  * ib_alloc_mr_integrity() - Allocates an integrity memory region
2322  * @pd:                      protection domain associated with the region
2323  * @max_num_data_sg:         maximum data sg entries available for registration
2324  * @max_num_meta_sg:         maximum metadata sg entries available for
2325  *                           registration
2326  *
2327  * Notes:
2328  * Memory registration page/sg lists must not exceed max_num_sg,
2329  * also the integrity page/sg lists must not exceed max_num_meta_sg.
2330  *
2331  */
2332 struct ib_mr *ib_alloc_mr_integrity(struct ib_pd *pd,
2333 				    u32 max_num_data_sg,
2334 				    u32 max_num_meta_sg)
2335 {
2336 	struct ib_mr *mr;
2337 	struct ib_sig_attrs *sig_attrs;
2338 
2339 	if (!pd->device->ops.alloc_mr_integrity ||
2340 	    !pd->device->ops.map_mr_sg_pi) {
2341 		mr = ERR_PTR(-EOPNOTSUPP);
2342 		goto out;
2343 	}
2344 
2345 	if (!max_num_meta_sg) {
2346 		mr = ERR_PTR(-EINVAL);
2347 		goto out;
2348 	}
2349 
2350 	sig_attrs = kzalloc(sizeof(struct ib_sig_attrs), GFP_KERNEL);
2351 	if (!sig_attrs) {
2352 		mr = ERR_PTR(-ENOMEM);
2353 		goto out;
2354 	}
2355 
2356 	mr = pd->device->ops.alloc_mr_integrity(pd, max_num_data_sg,
2357 						max_num_meta_sg);
2358 	if (IS_ERR(mr)) {
2359 		kfree(sig_attrs);
2360 		goto out;
2361 	}
2362 
2363 	mr->device = pd->device;
2364 	mr->pd = pd;
2365 	mr->dm = NULL;
2366 	mr->uobject = NULL;
2367 	atomic_inc(&pd->usecnt);
2368 	mr->need_inval = false;
2369 	mr->type = IB_MR_TYPE_INTEGRITY;
2370 	mr->sig_attrs = sig_attrs;
2371 
2372 	rdma_restrack_new(&mr->res, RDMA_RESTRACK_MR);
2373 	rdma_restrack_parent_name(&mr->res, &pd->res);
2374 	rdma_restrack_add(&mr->res);
2375 out:
2376 	trace_mr_integ_alloc(pd, max_num_data_sg, max_num_meta_sg, mr);
2377 	return mr;
2378 }
2379 EXPORT_SYMBOL(ib_alloc_mr_integrity);
2380 
2381 /* Multicast groups */
2382 
2383 static bool is_valid_mcast_lid(struct ib_qp *qp, u16 lid)
2384 {
2385 	struct ib_qp_init_attr init_attr = {};
2386 	struct ib_qp_attr attr = {};
2387 	int num_eth_ports = 0;
2388 	unsigned int port;
2389 
2390 	/* If QP state >= init, it is assigned to a port and we can check this
2391 	 * port only.
2392 	 */
2393 	if (!ib_query_qp(qp, &attr, IB_QP_STATE | IB_QP_PORT, &init_attr)) {
2394 		if (attr.qp_state >= IB_QPS_INIT) {
2395 			if (rdma_port_get_link_layer(qp->device, attr.port_num) !=
2396 			    IB_LINK_LAYER_INFINIBAND)
2397 				return true;
2398 			goto lid_check;
2399 		}
2400 	}
2401 
2402 	/* Can't get a quick answer, iterate over all ports */
2403 	rdma_for_each_port(qp->device, port)
2404 		if (rdma_port_get_link_layer(qp->device, port) !=
2405 		    IB_LINK_LAYER_INFINIBAND)
2406 			num_eth_ports++;
2407 
2408 	/* If we have at lease one Ethernet port, RoCE annex declares that
2409 	 * multicast LID should be ignored. We can't tell at this step if the
2410 	 * QP belongs to an IB or Ethernet port.
2411 	 */
2412 	if (num_eth_ports)
2413 		return true;
2414 
2415 	/* If all the ports are IB, we can check according to IB spec. */
2416 lid_check:
2417 	return !(lid < be16_to_cpu(IB_MULTICAST_LID_BASE) ||
2418 		 lid == be16_to_cpu(IB_LID_PERMISSIVE));
2419 }
2420 
2421 int ib_attach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid)
2422 {
2423 	int ret;
2424 
2425 	if (!qp->device->ops.attach_mcast)
2426 		return -EOPNOTSUPP;
2427 
2428 	if (!rdma_is_multicast_addr((struct in6_addr *)gid->raw) ||
2429 	    qp->qp_type != IB_QPT_UD || !is_valid_mcast_lid(qp, lid))
2430 		return -EINVAL;
2431 
2432 	ret = qp->device->ops.attach_mcast(qp, gid, lid);
2433 	if (!ret)
2434 		atomic_inc(&qp->usecnt);
2435 	return ret;
2436 }
2437 EXPORT_SYMBOL(ib_attach_mcast);
2438 
2439 int ib_detach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid)
2440 {
2441 	int ret;
2442 
2443 	if (!qp->device->ops.detach_mcast)
2444 		return -EOPNOTSUPP;
2445 
2446 	if (!rdma_is_multicast_addr((struct in6_addr *)gid->raw) ||
2447 	    qp->qp_type != IB_QPT_UD || !is_valid_mcast_lid(qp, lid))
2448 		return -EINVAL;
2449 
2450 	ret = qp->device->ops.detach_mcast(qp, gid, lid);
2451 	if (!ret)
2452 		atomic_dec(&qp->usecnt);
2453 	return ret;
2454 }
2455 EXPORT_SYMBOL(ib_detach_mcast);
2456 
2457 /**
2458  * ib_alloc_xrcd_user - Allocates an XRC domain.
2459  * @device: The device on which to allocate the XRC domain.
2460  * @inode: inode to connect XRCD
2461  * @udata: Valid user data or NULL for kernel object
2462  */
2463 struct ib_xrcd *ib_alloc_xrcd_user(struct ib_device *device,
2464 				   struct inode *inode, struct ib_udata *udata)
2465 {
2466 	struct ib_xrcd *xrcd;
2467 	int ret;
2468 
2469 	if (!device->ops.alloc_xrcd)
2470 		return ERR_PTR(-EOPNOTSUPP);
2471 
2472 	xrcd = rdma_zalloc_drv_obj(device, ib_xrcd);
2473 	if (!xrcd)
2474 		return ERR_PTR(-ENOMEM);
2475 
2476 	xrcd->device = device;
2477 	xrcd->inode = inode;
2478 	atomic_set(&xrcd->usecnt, 0);
2479 	init_rwsem(&xrcd->tgt_qps_rwsem);
2480 	xa_init(&xrcd->tgt_qps);
2481 
2482 	ret = device->ops.alloc_xrcd(xrcd, udata);
2483 	if (ret)
2484 		goto err;
2485 	return xrcd;
2486 err:
2487 	kfree(xrcd);
2488 	return ERR_PTR(ret);
2489 }
2490 EXPORT_SYMBOL(ib_alloc_xrcd_user);
2491 
2492 /**
2493  * ib_dealloc_xrcd_user - Deallocates an XRC domain.
2494  * @xrcd: The XRC domain to deallocate.
2495  * @udata: Valid user data or NULL for kernel object
2496  */
2497 int ib_dealloc_xrcd_user(struct ib_xrcd *xrcd, struct ib_udata *udata)
2498 {
2499 	int ret;
2500 
2501 	if (atomic_read(&xrcd->usecnt))
2502 		return -EBUSY;
2503 
2504 	WARN_ON(!xa_empty(&xrcd->tgt_qps));
2505 	ret = xrcd->device->ops.dealloc_xrcd(xrcd, udata);
2506 	if (ret)
2507 		return ret;
2508 	kfree(xrcd);
2509 	return ret;
2510 }
2511 EXPORT_SYMBOL(ib_dealloc_xrcd_user);
2512 
2513 /**
2514  * ib_create_wq - Creates a WQ associated with the specified protection
2515  * domain.
2516  * @pd: The protection domain associated with the WQ.
2517  * @wq_attr: A list of initial attributes required to create the
2518  * WQ. If WQ creation succeeds, then the attributes are updated to
2519  * the actual capabilities of the created WQ.
2520  *
2521  * wq_attr->max_wr and wq_attr->max_sge determine
2522  * the requested size of the WQ, and set to the actual values allocated
2523  * on return.
2524  * If ib_create_wq() succeeds, then max_wr and max_sge will always be
2525  * at least as large as the requested values.
2526  */
2527 struct ib_wq *ib_create_wq(struct ib_pd *pd,
2528 			   struct ib_wq_init_attr *wq_attr)
2529 {
2530 	struct ib_wq *wq;
2531 
2532 	if (!pd->device->ops.create_wq)
2533 		return ERR_PTR(-EOPNOTSUPP);
2534 
2535 	wq = pd->device->ops.create_wq(pd, wq_attr, NULL);
2536 	if (!IS_ERR(wq)) {
2537 		wq->event_handler = wq_attr->event_handler;
2538 		wq->wq_context = wq_attr->wq_context;
2539 		wq->wq_type = wq_attr->wq_type;
2540 		wq->cq = wq_attr->cq;
2541 		wq->device = pd->device;
2542 		wq->pd = pd;
2543 		wq->uobject = NULL;
2544 		atomic_inc(&pd->usecnt);
2545 		atomic_inc(&wq_attr->cq->usecnt);
2546 		atomic_set(&wq->usecnt, 0);
2547 	}
2548 	return wq;
2549 }
2550 EXPORT_SYMBOL(ib_create_wq);
2551 
2552 /**
2553  * ib_destroy_wq_user - Destroys the specified user WQ.
2554  * @wq: The WQ to destroy.
2555  * @udata: Valid user data
2556  */
2557 int ib_destroy_wq_user(struct ib_wq *wq, struct ib_udata *udata)
2558 {
2559 	struct ib_cq *cq = wq->cq;
2560 	struct ib_pd *pd = wq->pd;
2561 	int ret;
2562 
2563 	if (atomic_read(&wq->usecnt))
2564 		return -EBUSY;
2565 
2566 	ret = wq->device->ops.destroy_wq(wq, udata);
2567 	if (ret)
2568 		return ret;
2569 
2570 	atomic_dec(&pd->usecnt);
2571 	atomic_dec(&cq->usecnt);
2572 	return ret;
2573 }
2574 EXPORT_SYMBOL(ib_destroy_wq_user);
2575 
2576 int ib_check_mr_status(struct ib_mr *mr, u32 check_mask,
2577 		       struct ib_mr_status *mr_status)
2578 {
2579 	if (!mr->device->ops.check_mr_status)
2580 		return -EOPNOTSUPP;
2581 
2582 	return mr->device->ops.check_mr_status(mr, check_mask, mr_status);
2583 }
2584 EXPORT_SYMBOL(ib_check_mr_status);
2585 
2586 int ib_set_vf_link_state(struct ib_device *device, int vf, u32 port,
2587 			 int state)
2588 {
2589 	if (!device->ops.set_vf_link_state)
2590 		return -EOPNOTSUPP;
2591 
2592 	return device->ops.set_vf_link_state(device, vf, port, state);
2593 }
2594 EXPORT_SYMBOL(ib_set_vf_link_state);
2595 
2596 int ib_get_vf_config(struct ib_device *device, int vf, u32 port,
2597 		     struct ifla_vf_info *info)
2598 {
2599 	if (!device->ops.get_vf_config)
2600 		return -EOPNOTSUPP;
2601 
2602 	return device->ops.get_vf_config(device, vf, port, info);
2603 }
2604 EXPORT_SYMBOL(ib_get_vf_config);
2605 
2606 int ib_get_vf_stats(struct ib_device *device, int vf, u32 port,
2607 		    struct ifla_vf_stats *stats)
2608 {
2609 	if (!device->ops.get_vf_stats)
2610 		return -EOPNOTSUPP;
2611 
2612 	return device->ops.get_vf_stats(device, vf, port, stats);
2613 }
2614 EXPORT_SYMBOL(ib_get_vf_stats);
2615 
2616 int ib_set_vf_guid(struct ib_device *device, int vf, u32 port, u64 guid,
2617 		   int type)
2618 {
2619 	if (!device->ops.set_vf_guid)
2620 		return -EOPNOTSUPP;
2621 
2622 	return device->ops.set_vf_guid(device, vf, port, guid, type);
2623 }
2624 EXPORT_SYMBOL(ib_set_vf_guid);
2625 
2626 int ib_get_vf_guid(struct ib_device *device, int vf, u32 port,
2627 		   struct ifla_vf_guid *node_guid,
2628 		   struct ifla_vf_guid *port_guid)
2629 {
2630 	if (!device->ops.get_vf_guid)
2631 		return -EOPNOTSUPP;
2632 
2633 	return device->ops.get_vf_guid(device, vf, port, node_guid, port_guid);
2634 }
2635 EXPORT_SYMBOL(ib_get_vf_guid);
2636 /**
2637  * ib_map_mr_sg_pi() - Map the dma mapped SG lists for PI (protection
2638  *     information) and set an appropriate memory region for registration.
2639  * @mr:             memory region
2640  * @data_sg:        dma mapped scatterlist for data
2641  * @data_sg_nents:  number of entries in data_sg
2642  * @data_sg_offset: offset in bytes into data_sg
2643  * @meta_sg:        dma mapped scatterlist for metadata
2644  * @meta_sg_nents:  number of entries in meta_sg
2645  * @meta_sg_offset: offset in bytes into meta_sg
2646  * @page_size:      page vector desired page size
2647  *
2648  * Constraints:
2649  * - The MR must be allocated with type IB_MR_TYPE_INTEGRITY.
2650  *
2651  * Return: 0 on success.
2652  *
2653  * After this completes successfully, the  memory region
2654  * is ready for registration.
2655  */
2656 int ib_map_mr_sg_pi(struct ib_mr *mr, struct scatterlist *data_sg,
2657 		    int data_sg_nents, unsigned int *data_sg_offset,
2658 		    struct scatterlist *meta_sg, int meta_sg_nents,
2659 		    unsigned int *meta_sg_offset, unsigned int page_size)
2660 {
2661 	if (unlikely(!mr->device->ops.map_mr_sg_pi ||
2662 		     WARN_ON_ONCE(mr->type != IB_MR_TYPE_INTEGRITY)))
2663 		return -EOPNOTSUPP;
2664 
2665 	mr->page_size = page_size;
2666 
2667 	return mr->device->ops.map_mr_sg_pi(mr, data_sg, data_sg_nents,
2668 					    data_sg_offset, meta_sg,
2669 					    meta_sg_nents, meta_sg_offset);
2670 }
2671 EXPORT_SYMBOL(ib_map_mr_sg_pi);
2672 
2673 /**
2674  * ib_map_mr_sg() - Map the largest prefix of a dma mapped SG list
2675  *     and set it the memory region.
2676  * @mr:            memory region
2677  * @sg:            dma mapped scatterlist
2678  * @sg_nents:      number of entries in sg
2679  * @sg_offset:     offset in bytes into sg
2680  * @page_size:     page vector desired page size
2681  *
2682  * Constraints:
2683  *
2684  * - The first sg element is allowed to have an offset.
2685  * - Each sg element must either be aligned to page_size or virtually
2686  *   contiguous to the previous element. In case an sg element has a
2687  *   non-contiguous offset, the mapping prefix will not include it.
2688  * - The last sg element is allowed to have length less than page_size.
2689  * - If sg_nents total byte length exceeds the mr max_num_sge * page_size
2690  *   then only max_num_sg entries will be mapped.
2691  * - If the MR was allocated with type IB_MR_TYPE_SG_GAPS, none of these
2692  *   constraints holds and the page_size argument is ignored.
2693  *
2694  * Returns the number of sg elements that were mapped to the memory region.
2695  *
2696  * After this completes successfully, the  memory region
2697  * is ready for registration.
2698  */
2699 int ib_map_mr_sg(struct ib_mr *mr, struct scatterlist *sg, int sg_nents,
2700 		 unsigned int *sg_offset, unsigned int page_size)
2701 {
2702 	if (unlikely(!mr->device->ops.map_mr_sg))
2703 		return -EOPNOTSUPP;
2704 
2705 	mr->page_size = page_size;
2706 
2707 	return mr->device->ops.map_mr_sg(mr, sg, sg_nents, sg_offset);
2708 }
2709 EXPORT_SYMBOL(ib_map_mr_sg);
2710 
2711 /**
2712  * ib_sg_to_pages() - Convert the largest prefix of a sg list
2713  *     to a page vector
2714  * @mr:            memory region
2715  * @sgl:           dma mapped scatterlist
2716  * @sg_nents:      number of entries in sg
2717  * @sg_offset_p:   ==== =======================================================
2718  *                 IN   start offset in bytes into sg
2719  *                 OUT  offset in bytes for element n of the sg of the first
2720  *                      byte that has not been processed where n is the return
2721  *                      value of this function.
2722  *                 ==== =======================================================
2723  * @set_page:      driver page assignment function pointer
2724  *
2725  * Core service helper for drivers to convert the largest
2726  * prefix of given sg list to a page vector. The sg list
2727  * prefix converted is the prefix that meet the requirements
2728  * of ib_map_mr_sg.
2729  *
2730  * Returns the number of sg elements that were assigned to
2731  * a page vector.
2732  */
2733 int ib_sg_to_pages(struct ib_mr *mr, struct scatterlist *sgl, int sg_nents,
2734 		unsigned int *sg_offset_p, int (*set_page)(struct ib_mr *, u64))
2735 {
2736 	struct scatterlist *sg;
2737 	u64 last_end_dma_addr = 0;
2738 	unsigned int sg_offset = sg_offset_p ? *sg_offset_p : 0;
2739 	unsigned int last_page_off = 0;
2740 	u64 page_mask = ~((u64)mr->page_size - 1);
2741 	int i, ret;
2742 
2743 	if (unlikely(sg_nents <= 0 || sg_offset > sg_dma_len(&sgl[0])))
2744 		return -EINVAL;
2745 
2746 	mr->iova = sg_dma_address(&sgl[0]) + sg_offset;
2747 	mr->length = 0;
2748 
2749 	for_each_sg(sgl, sg, sg_nents, i) {
2750 		u64 dma_addr = sg_dma_address(sg) + sg_offset;
2751 		u64 prev_addr = dma_addr;
2752 		unsigned int dma_len = sg_dma_len(sg) - sg_offset;
2753 		u64 end_dma_addr = dma_addr + dma_len;
2754 		u64 page_addr = dma_addr & page_mask;
2755 
2756 		/*
2757 		 * For the second and later elements, check whether either the
2758 		 * end of element i-1 or the start of element i is not aligned
2759 		 * on a page boundary.
2760 		 */
2761 		if (i && (last_page_off != 0 || page_addr != dma_addr)) {
2762 			/* Stop mapping if there is a gap. */
2763 			if (last_end_dma_addr != dma_addr)
2764 				break;
2765 
2766 			/*
2767 			 * Coalesce this element with the last. If it is small
2768 			 * enough just update mr->length. Otherwise start
2769 			 * mapping from the next page.
2770 			 */
2771 			goto next_page;
2772 		}
2773 
2774 		do {
2775 			ret = set_page(mr, page_addr);
2776 			if (unlikely(ret < 0)) {
2777 				sg_offset = prev_addr - sg_dma_address(sg);
2778 				mr->length += prev_addr - dma_addr;
2779 				if (sg_offset_p)
2780 					*sg_offset_p = sg_offset;
2781 				return i || sg_offset ? i : ret;
2782 			}
2783 			prev_addr = page_addr;
2784 next_page:
2785 			page_addr += mr->page_size;
2786 		} while (page_addr < end_dma_addr);
2787 
2788 		mr->length += dma_len;
2789 		last_end_dma_addr = end_dma_addr;
2790 		last_page_off = end_dma_addr & ~page_mask;
2791 
2792 		sg_offset = 0;
2793 	}
2794 
2795 	if (sg_offset_p)
2796 		*sg_offset_p = 0;
2797 	return i;
2798 }
2799 EXPORT_SYMBOL(ib_sg_to_pages);
2800 
2801 struct ib_drain_cqe {
2802 	struct ib_cqe cqe;
2803 	struct completion done;
2804 };
2805 
2806 static void ib_drain_qp_done(struct ib_cq *cq, struct ib_wc *wc)
2807 {
2808 	struct ib_drain_cqe *cqe = container_of(wc->wr_cqe, struct ib_drain_cqe,
2809 						cqe);
2810 
2811 	complete(&cqe->done);
2812 }
2813 
2814 /*
2815  * Post a WR and block until its completion is reaped for the SQ.
2816  */
2817 static void __ib_drain_sq(struct ib_qp *qp)
2818 {
2819 	struct ib_cq *cq = qp->send_cq;
2820 	struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
2821 	struct ib_drain_cqe sdrain;
2822 	struct ib_rdma_wr swr = {
2823 		.wr = {
2824 			.next = NULL,
2825 			{ .wr_cqe	= &sdrain.cqe, },
2826 			.opcode	= IB_WR_RDMA_WRITE,
2827 		},
2828 	};
2829 	int ret;
2830 
2831 	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
2832 	if (ret) {
2833 		WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
2834 		return;
2835 	}
2836 
2837 	sdrain.cqe.done = ib_drain_qp_done;
2838 	init_completion(&sdrain.done);
2839 
2840 	ret = ib_post_send(qp, &swr.wr, NULL);
2841 	if (ret) {
2842 		WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
2843 		return;
2844 	}
2845 
2846 	if (cq->poll_ctx == IB_POLL_DIRECT)
2847 		while (wait_for_completion_timeout(&sdrain.done, HZ / 10) <= 0)
2848 			ib_process_cq_direct(cq, -1);
2849 	else
2850 		wait_for_completion(&sdrain.done);
2851 }
2852 
2853 /*
2854  * Post a WR and block until its completion is reaped for the RQ.
2855  */
2856 static void __ib_drain_rq(struct ib_qp *qp)
2857 {
2858 	struct ib_cq *cq = qp->recv_cq;
2859 	struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
2860 	struct ib_drain_cqe rdrain;
2861 	struct ib_recv_wr rwr = {};
2862 	int ret;
2863 
2864 	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
2865 	if (ret) {
2866 		WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
2867 		return;
2868 	}
2869 
2870 	rwr.wr_cqe = &rdrain.cqe;
2871 	rdrain.cqe.done = ib_drain_qp_done;
2872 	init_completion(&rdrain.done);
2873 
2874 	ret = ib_post_recv(qp, &rwr, NULL);
2875 	if (ret) {
2876 		WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
2877 		return;
2878 	}
2879 
2880 	if (cq->poll_ctx == IB_POLL_DIRECT)
2881 		while (wait_for_completion_timeout(&rdrain.done, HZ / 10) <= 0)
2882 			ib_process_cq_direct(cq, -1);
2883 	else
2884 		wait_for_completion(&rdrain.done);
2885 }
2886 
2887 /**
2888  * ib_drain_sq() - Block until all SQ CQEs have been consumed by the
2889  *		   application.
2890  * @qp:            queue pair to drain
2891  *
2892  * If the device has a provider-specific drain function, then
2893  * call that.  Otherwise call the generic drain function
2894  * __ib_drain_sq().
2895  *
2896  * The caller must:
2897  *
2898  * ensure there is room in the CQ and SQ for the drain work request and
2899  * completion.
2900  *
2901  * allocate the CQ using ib_alloc_cq().
2902  *
2903  * ensure that there are no other contexts that are posting WRs concurrently.
2904  * Otherwise the drain is not guaranteed.
2905  */
2906 void ib_drain_sq(struct ib_qp *qp)
2907 {
2908 	if (qp->device->ops.drain_sq)
2909 		qp->device->ops.drain_sq(qp);
2910 	else
2911 		__ib_drain_sq(qp);
2912 	trace_cq_drain_complete(qp->send_cq);
2913 }
2914 EXPORT_SYMBOL(ib_drain_sq);
2915 
2916 /**
2917  * ib_drain_rq() - Block until all RQ CQEs have been consumed by the
2918  *		   application.
2919  * @qp:            queue pair to drain
2920  *
2921  * If the device has a provider-specific drain function, then
2922  * call that.  Otherwise call the generic drain function
2923  * __ib_drain_rq().
2924  *
2925  * The caller must:
2926  *
2927  * ensure there is room in the CQ and RQ for the drain work request and
2928  * completion.
2929  *
2930  * allocate the CQ using ib_alloc_cq().
2931  *
2932  * ensure that there are no other contexts that are posting WRs concurrently.
2933  * Otherwise the drain is not guaranteed.
2934  */
2935 void ib_drain_rq(struct ib_qp *qp)
2936 {
2937 	if (qp->device->ops.drain_rq)
2938 		qp->device->ops.drain_rq(qp);
2939 	else
2940 		__ib_drain_rq(qp);
2941 	trace_cq_drain_complete(qp->recv_cq);
2942 }
2943 EXPORT_SYMBOL(ib_drain_rq);
2944 
2945 /**
2946  * ib_drain_qp() - Block until all CQEs have been consumed by the
2947  *		   application on both the RQ and SQ.
2948  * @qp:            queue pair to drain
2949  *
2950  * The caller must:
2951  *
2952  * ensure there is room in the CQ(s), SQ, and RQ for drain work requests
2953  * and completions.
2954  *
2955  * allocate the CQs using ib_alloc_cq().
2956  *
2957  * ensure that there are no other contexts that are posting WRs concurrently.
2958  * Otherwise the drain is not guaranteed.
2959  */
2960 void ib_drain_qp(struct ib_qp *qp)
2961 {
2962 	ib_drain_sq(qp);
2963 	if (!qp->srq)
2964 		ib_drain_rq(qp);
2965 }
2966 EXPORT_SYMBOL(ib_drain_qp);
2967 
2968 struct net_device *rdma_alloc_netdev(struct ib_device *device, u32 port_num,
2969 				     enum rdma_netdev_t type, const char *name,
2970 				     unsigned char name_assign_type,
2971 				     void (*setup)(struct net_device *))
2972 {
2973 	struct rdma_netdev_alloc_params params;
2974 	struct net_device *netdev;
2975 	int rc;
2976 
2977 	if (!device->ops.rdma_netdev_get_params)
2978 		return ERR_PTR(-EOPNOTSUPP);
2979 
2980 	rc = device->ops.rdma_netdev_get_params(device, port_num, type,
2981 						&params);
2982 	if (rc)
2983 		return ERR_PTR(rc);
2984 
2985 	netdev = alloc_netdev_mqs(params.sizeof_priv, name, name_assign_type,
2986 				  setup, params.txqs, params.rxqs);
2987 	if (!netdev)
2988 		return ERR_PTR(-ENOMEM);
2989 
2990 	return netdev;
2991 }
2992 EXPORT_SYMBOL(rdma_alloc_netdev);
2993 
2994 int rdma_init_netdev(struct ib_device *device, u32 port_num,
2995 		     enum rdma_netdev_t type, const char *name,
2996 		     unsigned char name_assign_type,
2997 		     void (*setup)(struct net_device *),
2998 		     struct net_device *netdev)
2999 {
3000 	struct rdma_netdev_alloc_params params;
3001 	int rc;
3002 
3003 	if (!device->ops.rdma_netdev_get_params)
3004 		return -EOPNOTSUPP;
3005 
3006 	rc = device->ops.rdma_netdev_get_params(device, port_num, type,
3007 						&params);
3008 	if (rc)
3009 		return rc;
3010 
3011 	return params.initialize_rdma_netdev(device, port_num,
3012 					     netdev, params.param);
3013 }
3014 EXPORT_SYMBOL(rdma_init_netdev);
3015 
3016 void __rdma_block_iter_start(struct ib_block_iter *biter,
3017 			     struct scatterlist *sglist, unsigned int nents,
3018 			     unsigned long pgsz)
3019 {
3020 	memset(biter, 0, sizeof(struct ib_block_iter));
3021 	biter->__sg = sglist;
3022 	biter->__sg_nents = nents;
3023 
3024 	/* Driver provides best block size to use */
3025 	biter->__pg_bit = __fls(pgsz);
3026 }
3027 EXPORT_SYMBOL(__rdma_block_iter_start);
3028 
3029 bool __rdma_block_iter_next(struct ib_block_iter *biter)
3030 {
3031 	unsigned int block_offset;
3032 	unsigned int sg_delta;
3033 
3034 	if (!biter->__sg_nents || !biter->__sg)
3035 		return false;
3036 
3037 	biter->__dma_addr = sg_dma_address(biter->__sg) + biter->__sg_advance;
3038 	block_offset = biter->__dma_addr & (BIT_ULL(biter->__pg_bit) - 1);
3039 	sg_delta = BIT_ULL(biter->__pg_bit) - block_offset;
3040 
3041 	if (sg_dma_len(biter->__sg) - biter->__sg_advance > sg_delta) {
3042 		biter->__sg_advance += sg_delta;
3043 	} else {
3044 		biter->__sg_advance = 0;
3045 		biter->__sg = sg_next(biter->__sg);
3046 		biter->__sg_nents--;
3047 	}
3048 
3049 	return true;
3050 }
3051 EXPORT_SYMBOL(__rdma_block_iter_next);
3052 
3053 /**
3054  * rdma_alloc_hw_stats_struct - Helper function to allocate dynamic struct
3055  *   for the drivers.
3056  * @descs: array of static descriptors
3057  * @num_counters: number of elements in array
3058  * @lifespan: milliseconds between updates
3059  */
3060 struct rdma_hw_stats *rdma_alloc_hw_stats_struct(
3061 	const struct rdma_stat_desc *descs, int num_counters,
3062 	unsigned long lifespan)
3063 {
3064 	struct rdma_hw_stats *stats;
3065 
3066 	stats = kzalloc(struct_size(stats, value, num_counters), GFP_KERNEL);
3067 	if (!stats)
3068 		return NULL;
3069 
3070 	stats->is_disabled = kcalloc(BITS_TO_LONGS(num_counters),
3071 				     sizeof(*stats->is_disabled), GFP_KERNEL);
3072 	if (!stats->is_disabled)
3073 		goto err;
3074 
3075 	stats->descs = descs;
3076 	stats->num_counters = num_counters;
3077 	stats->lifespan = msecs_to_jiffies(lifespan);
3078 	mutex_init(&stats->lock);
3079 
3080 	return stats;
3081 
3082 err:
3083 	kfree(stats);
3084 	return NULL;
3085 }
3086 EXPORT_SYMBOL(rdma_alloc_hw_stats_struct);
3087 
3088 /**
3089  * rdma_free_hw_stats_struct - Helper function to release rdma_hw_stats
3090  * @stats: statistics to release
3091  */
3092 void rdma_free_hw_stats_struct(struct rdma_hw_stats *stats)
3093 {
3094 	if (!stats)
3095 		return;
3096 
3097 	kfree(stats->is_disabled);
3098 	kfree(stats);
3099 }
3100 EXPORT_SYMBOL(rdma_free_hw_stats_struct);
3101