xref: /openbmc/linux/drivers/infiniband/core/verbs.c (revision 8e74a48d)
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 	default:	       return  -1;
151 	}
152 }
153 EXPORT_SYMBOL(ib_rate_to_mult);
154 
155 __attribute_const__ enum ib_rate mult_to_ib_rate(int mult)
156 {
157 	switch (mult) {
158 	case 1:   return IB_RATE_2_5_GBPS;
159 	case 2:   return IB_RATE_5_GBPS;
160 	case 4:   return IB_RATE_10_GBPS;
161 	case 8:   return IB_RATE_20_GBPS;
162 	case 12:  return IB_RATE_30_GBPS;
163 	case 16:  return IB_RATE_40_GBPS;
164 	case 24:  return IB_RATE_60_GBPS;
165 	case 32:  return IB_RATE_80_GBPS;
166 	case 48:  return IB_RATE_120_GBPS;
167 	case 6:   return IB_RATE_14_GBPS;
168 	case 22:  return IB_RATE_56_GBPS;
169 	case 45:  return IB_RATE_112_GBPS;
170 	case 67:  return IB_RATE_168_GBPS;
171 	case 10:  return IB_RATE_25_GBPS;
172 	case 40:  return IB_RATE_100_GBPS;
173 	case 80:  return IB_RATE_200_GBPS;
174 	case 120: return IB_RATE_300_GBPS;
175 	case 11:  return IB_RATE_28_GBPS;
176 	case 20:  return IB_RATE_50_GBPS;
177 	case 160: return IB_RATE_400_GBPS;
178 	case 240: return IB_RATE_600_GBPS;
179 	default:  return IB_RATE_PORT_CURRENT;
180 	}
181 }
182 EXPORT_SYMBOL(mult_to_ib_rate);
183 
184 __attribute_const__ int ib_rate_to_mbps(enum ib_rate rate)
185 {
186 	switch (rate) {
187 	case IB_RATE_2_5_GBPS: return 2500;
188 	case IB_RATE_5_GBPS:   return 5000;
189 	case IB_RATE_10_GBPS:  return 10000;
190 	case IB_RATE_20_GBPS:  return 20000;
191 	case IB_RATE_30_GBPS:  return 30000;
192 	case IB_RATE_40_GBPS:  return 40000;
193 	case IB_RATE_60_GBPS:  return 60000;
194 	case IB_RATE_80_GBPS:  return 80000;
195 	case IB_RATE_120_GBPS: return 120000;
196 	case IB_RATE_14_GBPS:  return 14062;
197 	case IB_RATE_56_GBPS:  return 56250;
198 	case IB_RATE_112_GBPS: return 112500;
199 	case IB_RATE_168_GBPS: return 168750;
200 	case IB_RATE_25_GBPS:  return 25781;
201 	case IB_RATE_100_GBPS: return 103125;
202 	case IB_RATE_200_GBPS: return 206250;
203 	case IB_RATE_300_GBPS: return 309375;
204 	case IB_RATE_28_GBPS:  return 28125;
205 	case IB_RATE_50_GBPS:  return 53125;
206 	case IB_RATE_400_GBPS: return 425000;
207 	case IB_RATE_600_GBPS: return 637500;
208 	default:	       return -1;
209 	}
210 }
211 EXPORT_SYMBOL(ib_rate_to_mbps);
212 
213 __attribute_const__ enum rdma_transport_type
214 rdma_node_get_transport(unsigned int node_type)
215 {
216 
217 	if (node_type == RDMA_NODE_USNIC)
218 		return RDMA_TRANSPORT_USNIC;
219 	if (node_type == RDMA_NODE_USNIC_UDP)
220 		return RDMA_TRANSPORT_USNIC_UDP;
221 	if (node_type == RDMA_NODE_RNIC)
222 		return RDMA_TRANSPORT_IWARP;
223 	if (node_type == RDMA_NODE_UNSPECIFIED)
224 		return RDMA_TRANSPORT_UNSPECIFIED;
225 
226 	return RDMA_TRANSPORT_IB;
227 }
228 EXPORT_SYMBOL(rdma_node_get_transport);
229 
230 enum rdma_link_layer rdma_port_get_link_layer(struct ib_device *device,
231 					      u32 port_num)
232 {
233 	enum rdma_transport_type lt;
234 	if (device->ops.get_link_layer)
235 		return device->ops.get_link_layer(device, port_num);
236 
237 	lt = rdma_node_get_transport(device->node_type);
238 	if (lt == RDMA_TRANSPORT_IB)
239 		return IB_LINK_LAYER_INFINIBAND;
240 
241 	return IB_LINK_LAYER_ETHERNET;
242 }
243 EXPORT_SYMBOL(rdma_port_get_link_layer);
244 
245 /* Protection domains */
246 
247 /**
248  * __ib_alloc_pd - Allocates an unused protection domain.
249  * @device: The device on which to allocate the protection domain.
250  * @flags: protection domain flags
251  * @caller: caller's build-time module name
252  *
253  * A protection domain object provides an association between QPs, shared
254  * receive queues, address handles, memory regions, and memory windows.
255  *
256  * Every PD has a local_dma_lkey which can be used as the lkey value for local
257  * memory operations.
258  */
259 struct ib_pd *__ib_alloc_pd(struct ib_device *device, unsigned int flags,
260 		const char *caller)
261 {
262 	struct ib_pd *pd;
263 	int mr_access_flags = 0;
264 	int ret;
265 
266 	pd = rdma_zalloc_drv_obj(device, ib_pd);
267 	if (!pd)
268 		return ERR_PTR(-ENOMEM);
269 
270 	pd->device = device;
271 	pd->uobject = NULL;
272 	pd->__internal_mr = NULL;
273 	atomic_set(&pd->usecnt, 0);
274 	pd->flags = flags;
275 
276 	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.device_cap_flags & IB_DEVICE_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.device_cap_flags & IB_DEVICE_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 	/* uverbs manipulates usecnt with proper locking, while the kabi
345 	 * requires the caller to guarantee we can't race here.
346 	 */
347 	WARN_ON(atomic_read(&pd->usecnt));
348 
349 	ret = pd->device->ops.dealloc_pd(pd, udata);
350 	if (ret)
351 		return ret;
352 
353 	rdma_restrack_del(&pd->res);
354 	kfree(pd);
355 	return ret;
356 }
357 EXPORT_SYMBOL(ib_dealloc_pd_user);
358 
359 /* Address handles */
360 
361 /**
362  * rdma_copy_ah_attr - Copy rdma ah attribute from source to destination.
363  * @dest:       Pointer to destination ah_attr. Contents of the destination
364  *              pointer is assumed to be invalid and attribute are overwritten.
365  * @src:        Pointer to source ah_attr.
366  */
367 void rdma_copy_ah_attr(struct rdma_ah_attr *dest,
368 		       const struct rdma_ah_attr *src)
369 {
370 	*dest = *src;
371 	if (dest->grh.sgid_attr)
372 		rdma_hold_gid_attr(dest->grh.sgid_attr);
373 }
374 EXPORT_SYMBOL(rdma_copy_ah_attr);
375 
376 /**
377  * rdma_replace_ah_attr - Replace valid ah_attr with new new one.
378  * @old:        Pointer to existing ah_attr which needs to be replaced.
379  *              old is assumed to be valid or zero'd
380  * @new:        Pointer to the new ah_attr.
381  *
382  * rdma_replace_ah_attr() first releases any reference in the old ah_attr if
383  * old the ah_attr is valid; after that it copies the new attribute and holds
384  * the reference to the replaced ah_attr.
385  */
386 void rdma_replace_ah_attr(struct rdma_ah_attr *old,
387 			  const struct rdma_ah_attr *new)
388 {
389 	rdma_destroy_ah_attr(old);
390 	*old = *new;
391 	if (old->grh.sgid_attr)
392 		rdma_hold_gid_attr(old->grh.sgid_attr);
393 }
394 EXPORT_SYMBOL(rdma_replace_ah_attr);
395 
396 /**
397  * rdma_move_ah_attr - Move ah_attr pointed by source to destination.
398  * @dest:       Pointer to destination ah_attr to copy to.
399  *              dest is assumed to be valid or zero'd
400  * @src:        Pointer to the new ah_attr.
401  *
402  * rdma_move_ah_attr() first releases any reference in the destination ah_attr
403  * if it is valid. This also transfers ownership of internal references from
404  * src to dest, making src invalid in the process. No new reference of the src
405  * ah_attr is taken.
406  */
407 void rdma_move_ah_attr(struct rdma_ah_attr *dest, struct rdma_ah_attr *src)
408 {
409 	rdma_destroy_ah_attr(dest);
410 	*dest = *src;
411 	src->grh.sgid_attr = NULL;
412 }
413 EXPORT_SYMBOL(rdma_move_ah_attr);
414 
415 /*
416  * Validate that the rdma_ah_attr is valid for the device before passing it
417  * off to the driver.
418  */
419 static int rdma_check_ah_attr(struct ib_device *device,
420 			      struct rdma_ah_attr *ah_attr)
421 {
422 	if (!rdma_is_port_valid(device, ah_attr->port_num))
423 		return -EINVAL;
424 
425 	if ((rdma_is_grh_required(device, ah_attr->port_num) ||
426 	     ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE) &&
427 	    !(ah_attr->ah_flags & IB_AH_GRH))
428 		return -EINVAL;
429 
430 	if (ah_attr->grh.sgid_attr) {
431 		/*
432 		 * Make sure the passed sgid_attr is consistent with the
433 		 * parameters
434 		 */
435 		if (ah_attr->grh.sgid_attr->index != ah_attr->grh.sgid_index ||
436 		    ah_attr->grh.sgid_attr->port_num != ah_attr->port_num)
437 			return -EINVAL;
438 	}
439 	return 0;
440 }
441 
442 /*
443  * If the ah requires a GRH then ensure that sgid_attr pointer is filled in.
444  * On success the caller is responsible to call rdma_unfill_sgid_attr().
445  */
446 static int rdma_fill_sgid_attr(struct ib_device *device,
447 			       struct rdma_ah_attr *ah_attr,
448 			       const struct ib_gid_attr **old_sgid_attr)
449 {
450 	const struct ib_gid_attr *sgid_attr;
451 	struct ib_global_route *grh;
452 	int ret;
453 
454 	*old_sgid_attr = ah_attr->grh.sgid_attr;
455 
456 	ret = rdma_check_ah_attr(device, ah_attr);
457 	if (ret)
458 		return ret;
459 
460 	if (!(ah_attr->ah_flags & IB_AH_GRH))
461 		return 0;
462 
463 	grh = rdma_ah_retrieve_grh(ah_attr);
464 	if (grh->sgid_attr)
465 		return 0;
466 
467 	sgid_attr =
468 		rdma_get_gid_attr(device, ah_attr->port_num, grh->sgid_index);
469 	if (IS_ERR(sgid_attr))
470 		return PTR_ERR(sgid_attr);
471 
472 	/* Move ownerhip of the kref into the ah_attr */
473 	grh->sgid_attr = sgid_attr;
474 	return 0;
475 }
476 
477 static void rdma_unfill_sgid_attr(struct rdma_ah_attr *ah_attr,
478 				  const struct ib_gid_attr *old_sgid_attr)
479 {
480 	/*
481 	 * Fill didn't change anything, the caller retains ownership of
482 	 * whatever it passed
483 	 */
484 	if (ah_attr->grh.sgid_attr == old_sgid_attr)
485 		return;
486 
487 	/*
488 	 * Otherwise, we need to undo what rdma_fill_sgid_attr so the caller
489 	 * doesn't see any change in the rdma_ah_attr. If we get here
490 	 * old_sgid_attr is NULL.
491 	 */
492 	rdma_destroy_ah_attr(ah_attr);
493 }
494 
495 static const struct ib_gid_attr *
496 rdma_update_sgid_attr(struct rdma_ah_attr *ah_attr,
497 		      const struct ib_gid_attr *old_attr)
498 {
499 	if (old_attr)
500 		rdma_put_gid_attr(old_attr);
501 	if (ah_attr->ah_flags & IB_AH_GRH) {
502 		rdma_hold_gid_attr(ah_attr->grh.sgid_attr);
503 		return ah_attr->grh.sgid_attr;
504 	}
505 	return NULL;
506 }
507 
508 static struct ib_ah *_rdma_create_ah(struct ib_pd *pd,
509 				     struct rdma_ah_attr *ah_attr,
510 				     u32 flags,
511 				     struct ib_udata *udata,
512 				     struct net_device *xmit_slave)
513 {
514 	struct rdma_ah_init_attr init_attr = {};
515 	struct ib_device *device = pd->device;
516 	struct ib_ah *ah;
517 	int ret;
518 
519 	might_sleep_if(flags & RDMA_CREATE_AH_SLEEPABLE);
520 
521 	if (!udata && !device->ops.create_ah)
522 		return ERR_PTR(-EOPNOTSUPP);
523 
524 	ah = rdma_zalloc_drv_obj_gfp(
525 		device, ib_ah,
526 		(flags & RDMA_CREATE_AH_SLEEPABLE) ? GFP_KERNEL : GFP_ATOMIC);
527 	if (!ah)
528 		return ERR_PTR(-ENOMEM);
529 
530 	ah->device = device;
531 	ah->pd = pd;
532 	ah->type = ah_attr->type;
533 	ah->sgid_attr = rdma_update_sgid_attr(ah_attr, NULL);
534 	init_attr.ah_attr = ah_attr;
535 	init_attr.flags = flags;
536 	init_attr.xmit_slave = xmit_slave;
537 
538 	if (udata)
539 		ret = device->ops.create_user_ah(ah, &init_attr, udata);
540 	else
541 		ret = device->ops.create_ah(ah, &init_attr, NULL);
542 	if (ret) {
543 		kfree(ah);
544 		return ERR_PTR(ret);
545 	}
546 
547 	atomic_inc(&pd->usecnt);
548 	return ah;
549 }
550 
551 /**
552  * rdma_create_ah - Creates an address handle for the
553  * given address vector.
554  * @pd: The protection domain associated with the address handle.
555  * @ah_attr: The attributes of the address vector.
556  * @flags: Create address handle flags (see enum rdma_create_ah_flags).
557  *
558  * It returns 0 on success and returns appropriate error code on error.
559  * The address handle is used to reference a local or global destination
560  * in all UD QP post sends.
561  */
562 struct ib_ah *rdma_create_ah(struct ib_pd *pd, struct rdma_ah_attr *ah_attr,
563 			     u32 flags)
564 {
565 	const struct ib_gid_attr *old_sgid_attr;
566 	struct net_device *slave;
567 	struct ib_ah *ah;
568 	int ret;
569 
570 	ret = rdma_fill_sgid_attr(pd->device, ah_attr, &old_sgid_attr);
571 	if (ret)
572 		return ERR_PTR(ret);
573 	slave = rdma_lag_get_ah_roce_slave(pd->device, ah_attr,
574 					   (flags & RDMA_CREATE_AH_SLEEPABLE) ?
575 					   GFP_KERNEL : GFP_ATOMIC);
576 	if (IS_ERR(slave)) {
577 		rdma_unfill_sgid_attr(ah_attr, old_sgid_attr);
578 		return (void *)slave;
579 	}
580 	ah = _rdma_create_ah(pd, ah_attr, flags, NULL, slave);
581 	rdma_lag_put_ah_roce_slave(slave);
582 	rdma_unfill_sgid_attr(ah_attr, old_sgid_attr);
583 	return ah;
584 }
585 EXPORT_SYMBOL(rdma_create_ah);
586 
587 /**
588  * rdma_create_user_ah - Creates an address handle for the
589  * given address vector.
590  * It resolves destination mac address for ah attribute of RoCE type.
591  * @pd: The protection domain associated with the address handle.
592  * @ah_attr: The attributes of the address vector.
593  * @udata: pointer to user's input output buffer information need by
594  *         provider driver.
595  *
596  * It returns 0 on success and returns appropriate error code on error.
597  * The address handle is used to reference a local or global destination
598  * in all UD QP post sends.
599  */
600 struct ib_ah *rdma_create_user_ah(struct ib_pd *pd,
601 				  struct rdma_ah_attr *ah_attr,
602 				  struct ib_udata *udata)
603 {
604 	const struct ib_gid_attr *old_sgid_attr;
605 	struct ib_ah *ah;
606 	int err;
607 
608 	err = rdma_fill_sgid_attr(pd->device, ah_attr, &old_sgid_attr);
609 	if (err)
610 		return ERR_PTR(err);
611 
612 	if (ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE) {
613 		err = ib_resolve_eth_dmac(pd->device, ah_attr);
614 		if (err) {
615 			ah = ERR_PTR(err);
616 			goto out;
617 		}
618 	}
619 
620 	ah = _rdma_create_ah(pd, ah_attr, RDMA_CREATE_AH_SLEEPABLE,
621 			     udata, NULL);
622 
623 out:
624 	rdma_unfill_sgid_attr(ah_attr, old_sgid_attr);
625 	return ah;
626 }
627 EXPORT_SYMBOL(rdma_create_user_ah);
628 
629 int ib_get_rdma_header_version(const union rdma_network_hdr *hdr)
630 {
631 	const struct iphdr *ip4h = (struct iphdr *)&hdr->roce4grh;
632 	struct iphdr ip4h_checked;
633 	const struct ipv6hdr *ip6h = (struct ipv6hdr *)&hdr->ibgrh;
634 
635 	/* If it's IPv6, the version must be 6, otherwise, the first
636 	 * 20 bytes (before the IPv4 header) are garbled.
637 	 */
638 	if (ip6h->version != 6)
639 		return (ip4h->version == 4) ? 4 : 0;
640 	/* version may be 6 or 4 because the first 20 bytes could be garbled */
641 
642 	/* RoCE v2 requires no options, thus header length
643 	 * must be 5 words
644 	 */
645 	if (ip4h->ihl != 5)
646 		return 6;
647 
648 	/* Verify checksum.
649 	 * We can't write on scattered buffers so we need to copy to
650 	 * temp buffer.
651 	 */
652 	memcpy(&ip4h_checked, ip4h, sizeof(ip4h_checked));
653 	ip4h_checked.check = 0;
654 	ip4h_checked.check = ip_fast_csum((u8 *)&ip4h_checked, 5);
655 	/* if IPv4 header checksum is OK, believe it */
656 	if (ip4h->check == ip4h_checked.check)
657 		return 4;
658 	return 6;
659 }
660 EXPORT_SYMBOL(ib_get_rdma_header_version);
661 
662 static enum rdma_network_type ib_get_net_type_by_grh(struct ib_device *device,
663 						     u32 port_num,
664 						     const struct ib_grh *grh)
665 {
666 	int grh_version;
667 
668 	if (rdma_protocol_ib(device, port_num))
669 		return RDMA_NETWORK_IB;
670 
671 	grh_version = ib_get_rdma_header_version((union rdma_network_hdr *)grh);
672 
673 	if (grh_version == 4)
674 		return RDMA_NETWORK_IPV4;
675 
676 	if (grh->next_hdr == IPPROTO_UDP)
677 		return RDMA_NETWORK_IPV6;
678 
679 	return RDMA_NETWORK_ROCE_V1;
680 }
681 
682 struct find_gid_index_context {
683 	u16 vlan_id;
684 	enum ib_gid_type gid_type;
685 };
686 
687 static bool find_gid_index(const union ib_gid *gid,
688 			   const struct ib_gid_attr *gid_attr,
689 			   void *context)
690 {
691 	struct find_gid_index_context *ctx = context;
692 	u16 vlan_id = 0xffff;
693 	int ret;
694 
695 	if (ctx->gid_type != gid_attr->gid_type)
696 		return false;
697 
698 	ret = rdma_read_gid_l2_fields(gid_attr, &vlan_id, NULL);
699 	if (ret)
700 		return false;
701 
702 	return ctx->vlan_id == vlan_id;
703 }
704 
705 static const struct ib_gid_attr *
706 get_sgid_attr_from_eth(struct ib_device *device, u32 port_num,
707 		       u16 vlan_id, const union ib_gid *sgid,
708 		       enum ib_gid_type gid_type)
709 {
710 	struct find_gid_index_context context = {.vlan_id = vlan_id,
711 						 .gid_type = gid_type};
712 
713 	return rdma_find_gid_by_filter(device, sgid, port_num, find_gid_index,
714 				       &context);
715 }
716 
717 int ib_get_gids_from_rdma_hdr(const union rdma_network_hdr *hdr,
718 			      enum rdma_network_type net_type,
719 			      union ib_gid *sgid, union ib_gid *dgid)
720 {
721 	struct sockaddr_in  src_in;
722 	struct sockaddr_in  dst_in;
723 	__be32 src_saddr, dst_saddr;
724 
725 	if (!sgid || !dgid)
726 		return -EINVAL;
727 
728 	if (net_type == RDMA_NETWORK_IPV4) {
729 		memcpy(&src_in.sin_addr.s_addr,
730 		       &hdr->roce4grh.saddr, 4);
731 		memcpy(&dst_in.sin_addr.s_addr,
732 		       &hdr->roce4grh.daddr, 4);
733 		src_saddr = src_in.sin_addr.s_addr;
734 		dst_saddr = dst_in.sin_addr.s_addr;
735 		ipv6_addr_set_v4mapped(src_saddr,
736 				       (struct in6_addr *)sgid);
737 		ipv6_addr_set_v4mapped(dst_saddr,
738 				       (struct in6_addr *)dgid);
739 		return 0;
740 	} else if (net_type == RDMA_NETWORK_IPV6 ||
741 		   net_type == RDMA_NETWORK_IB || RDMA_NETWORK_ROCE_V1) {
742 		*dgid = hdr->ibgrh.dgid;
743 		*sgid = hdr->ibgrh.sgid;
744 		return 0;
745 	} else {
746 		return -EINVAL;
747 	}
748 }
749 EXPORT_SYMBOL(ib_get_gids_from_rdma_hdr);
750 
751 /* Resolve destination mac address and hop limit for unicast destination
752  * GID entry, considering the source GID entry as well.
753  * ah_attribute must have have valid port_num, sgid_index.
754  */
755 static int ib_resolve_unicast_gid_dmac(struct ib_device *device,
756 				       struct rdma_ah_attr *ah_attr)
757 {
758 	struct ib_global_route *grh = rdma_ah_retrieve_grh(ah_attr);
759 	const struct ib_gid_attr *sgid_attr = grh->sgid_attr;
760 	int hop_limit = 0xff;
761 	int ret = 0;
762 
763 	/* If destination is link local and source GID is RoCEv1,
764 	 * IP stack is not used.
765 	 */
766 	if (rdma_link_local_addr((struct in6_addr *)grh->dgid.raw) &&
767 	    sgid_attr->gid_type == IB_GID_TYPE_ROCE) {
768 		rdma_get_ll_mac((struct in6_addr *)grh->dgid.raw,
769 				ah_attr->roce.dmac);
770 		return ret;
771 	}
772 
773 	ret = rdma_addr_find_l2_eth_by_grh(&sgid_attr->gid, &grh->dgid,
774 					   ah_attr->roce.dmac,
775 					   sgid_attr, &hop_limit);
776 
777 	grh->hop_limit = hop_limit;
778 	return ret;
779 }
780 
781 /*
782  * This function initializes address handle attributes from the incoming packet.
783  * Incoming packet has dgid of the receiver node on which this code is
784  * getting executed and, sgid contains the GID of the sender.
785  *
786  * When resolving mac address of destination, the arrived dgid is used
787  * as sgid and, sgid is used as dgid because sgid contains destinations
788  * GID whom to respond to.
789  *
790  * On success the caller is responsible to call rdma_destroy_ah_attr on the
791  * attr.
792  */
793 int ib_init_ah_attr_from_wc(struct ib_device *device, u32 port_num,
794 			    const struct ib_wc *wc, const struct ib_grh *grh,
795 			    struct rdma_ah_attr *ah_attr)
796 {
797 	u32 flow_class;
798 	int ret;
799 	enum rdma_network_type net_type = RDMA_NETWORK_IB;
800 	enum ib_gid_type gid_type = IB_GID_TYPE_IB;
801 	const struct ib_gid_attr *sgid_attr;
802 	int hoplimit = 0xff;
803 	union ib_gid dgid;
804 	union ib_gid sgid;
805 
806 	might_sleep();
807 
808 	memset(ah_attr, 0, sizeof *ah_attr);
809 	ah_attr->type = rdma_ah_find_type(device, port_num);
810 	if (rdma_cap_eth_ah(device, port_num)) {
811 		if (wc->wc_flags & IB_WC_WITH_NETWORK_HDR_TYPE)
812 			net_type = wc->network_hdr_type;
813 		else
814 			net_type = ib_get_net_type_by_grh(device, port_num, grh);
815 		gid_type = ib_network_to_gid_type(net_type);
816 	}
817 	ret = ib_get_gids_from_rdma_hdr((union rdma_network_hdr *)grh, net_type,
818 					&sgid, &dgid);
819 	if (ret)
820 		return ret;
821 
822 	rdma_ah_set_sl(ah_attr, wc->sl);
823 	rdma_ah_set_port_num(ah_attr, port_num);
824 
825 	if (rdma_protocol_roce(device, port_num)) {
826 		u16 vlan_id = wc->wc_flags & IB_WC_WITH_VLAN ?
827 				wc->vlan_id : 0xffff;
828 
829 		if (!(wc->wc_flags & IB_WC_GRH))
830 			return -EPROTOTYPE;
831 
832 		sgid_attr = get_sgid_attr_from_eth(device, port_num,
833 						   vlan_id, &dgid,
834 						   gid_type);
835 		if (IS_ERR(sgid_attr))
836 			return PTR_ERR(sgid_attr);
837 
838 		flow_class = be32_to_cpu(grh->version_tclass_flow);
839 		rdma_move_grh_sgid_attr(ah_attr,
840 					&sgid,
841 					flow_class & 0xFFFFF,
842 					hoplimit,
843 					(flow_class >> 20) & 0xFF,
844 					sgid_attr);
845 
846 		ret = ib_resolve_unicast_gid_dmac(device, ah_attr);
847 		if (ret)
848 			rdma_destroy_ah_attr(ah_attr);
849 
850 		return ret;
851 	} else {
852 		rdma_ah_set_dlid(ah_attr, wc->slid);
853 		rdma_ah_set_path_bits(ah_attr, wc->dlid_path_bits);
854 
855 		if ((wc->wc_flags & IB_WC_GRH) == 0)
856 			return 0;
857 
858 		if (dgid.global.interface_id !=
859 					cpu_to_be64(IB_SA_WELL_KNOWN_GUID)) {
860 			sgid_attr = rdma_find_gid_by_port(
861 				device, &dgid, IB_GID_TYPE_IB, port_num, NULL);
862 		} else
863 			sgid_attr = rdma_get_gid_attr(device, port_num, 0);
864 
865 		if (IS_ERR(sgid_attr))
866 			return PTR_ERR(sgid_attr);
867 		flow_class = be32_to_cpu(grh->version_tclass_flow);
868 		rdma_move_grh_sgid_attr(ah_attr,
869 					&sgid,
870 					flow_class & 0xFFFFF,
871 					hoplimit,
872 					(flow_class >> 20) & 0xFF,
873 					sgid_attr);
874 
875 		return 0;
876 	}
877 }
878 EXPORT_SYMBOL(ib_init_ah_attr_from_wc);
879 
880 /**
881  * rdma_move_grh_sgid_attr - Sets the sgid attribute of GRH, taking ownership
882  * of the reference
883  *
884  * @attr:	Pointer to AH attribute structure
885  * @dgid:	Destination GID
886  * @flow_label:	Flow label
887  * @hop_limit:	Hop limit
888  * @traffic_class: traffic class
889  * @sgid_attr:	Pointer to SGID attribute
890  *
891  * This takes ownership of the sgid_attr reference. The caller must ensure
892  * rdma_destroy_ah_attr() is called before destroying the rdma_ah_attr after
893  * calling this function.
894  */
895 void rdma_move_grh_sgid_attr(struct rdma_ah_attr *attr, union ib_gid *dgid,
896 			     u32 flow_label, u8 hop_limit, u8 traffic_class,
897 			     const struct ib_gid_attr *sgid_attr)
898 {
899 	rdma_ah_set_grh(attr, dgid, flow_label, sgid_attr->index, hop_limit,
900 			traffic_class);
901 	attr->grh.sgid_attr = sgid_attr;
902 }
903 EXPORT_SYMBOL(rdma_move_grh_sgid_attr);
904 
905 /**
906  * rdma_destroy_ah_attr - Release reference to SGID attribute of
907  * ah attribute.
908  * @ah_attr: Pointer to ah attribute
909  *
910  * Release reference to the SGID attribute of the ah attribute if it is
911  * non NULL. It is safe to call this multiple times, and safe to call it on
912  * a zero initialized ah_attr.
913  */
914 void rdma_destroy_ah_attr(struct rdma_ah_attr *ah_attr)
915 {
916 	if (ah_attr->grh.sgid_attr) {
917 		rdma_put_gid_attr(ah_attr->grh.sgid_attr);
918 		ah_attr->grh.sgid_attr = NULL;
919 	}
920 }
921 EXPORT_SYMBOL(rdma_destroy_ah_attr);
922 
923 struct ib_ah *ib_create_ah_from_wc(struct ib_pd *pd, const struct ib_wc *wc,
924 				   const struct ib_grh *grh, u32 port_num)
925 {
926 	struct rdma_ah_attr ah_attr;
927 	struct ib_ah *ah;
928 	int ret;
929 
930 	ret = ib_init_ah_attr_from_wc(pd->device, port_num, wc, grh, &ah_attr);
931 	if (ret)
932 		return ERR_PTR(ret);
933 
934 	ah = rdma_create_ah(pd, &ah_attr, RDMA_CREATE_AH_SLEEPABLE);
935 
936 	rdma_destroy_ah_attr(&ah_attr);
937 	return ah;
938 }
939 EXPORT_SYMBOL(ib_create_ah_from_wc);
940 
941 int rdma_modify_ah(struct ib_ah *ah, struct rdma_ah_attr *ah_attr)
942 {
943 	const struct ib_gid_attr *old_sgid_attr;
944 	int ret;
945 
946 	if (ah->type != ah_attr->type)
947 		return -EINVAL;
948 
949 	ret = rdma_fill_sgid_attr(ah->device, ah_attr, &old_sgid_attr);
950 	if (ret)
951 		return ret;
952 
953 	ret = ah->device->ops.modify_ah ?
954 		ah->device->ops.modify_ah(ah, ah_attr) :
955 		-EOPNOTSUPP;
956 
957 	ah->sgid_attr = rdma_update_sgid_attr(ah_attr, ah->sgid_attr);
958 	rdma_unfill_sgid_attr(ah_attr, old_sgid_attr);
959 	return ret;
960 }
961 EXPORT_SYMBOL(rdma_modify_ah);
962 
963 int rdma_query_ah(struct ib_ah *ah, struct rdma_ah_attr *ah_attr)
964 {
965 	ah_attr->grh.sgid_attr = NULL;
966 
967 	return ah->device->ops.query_ah ?
968 		ah->device->ops.query_ah(ah, ah_attr) :
969 		-EOPNOTSUPP;
970 }
971 EXPORT_SYMBOL(rdma_query_ah);
972 
973 int rdma_destroy_ah_user(struct ib_ah *ah, u32 flags, struct ib_udata *udata)
974 {
975 	const struct ib_gid_attr *sgid_attr = ah->sgid_attr;
976 	struct ib_pd *pd;
977 	int ret;
978 
979 	might_sleep_if(flags & RDMA_DESTROY_AH_SLEEPABLE);
980 
981 	pd = ah->pd;
982 
983 	ret = ah->device->ops.destroy_ah(ah, flags);
984 	if (ret)
985 		return ret;
986 
987 	atomic_dec(&pd->usecnt);
988 	if (sgid_attr)
989 		rdma_put_gid_attr(sgid_attr);
990 
991 	kfree(ah);
992 	return ret;
993 }
994 EXPORT_SYMBOL(rdma_destroy_ah_user);
995 
996 /* Shared receive queues */
997 
998 /**
999  * ib_create_srq_user - Creates a SRQ associated with the specified protection
1000  *   domain.
1001  * @pd: The protection domain associated with the SRQ.
1002  * @srq_init_attr: A list of initial attributes required to create the
1003  *   SRQ.  If SRQ creation succeeds, then the attributes are updated to
1004  *   the actual capabilities of the created SRQ.
1005  * @uobject: uobject pointer if this is not a kernel SRQ
1006  * @udata: udata pointer if this is not a kernel SRQ
1007  *
1008  * srq_attr->max_wr and srq_attr->max_sge are read the determine the
1009  * requested size of the SRQ, and set to the actual values allocated
1010  * on return.  If ib_create_srq() succeeds, then max_wr and max_sge
1011  * will always be at least as large as the requested values.
1012  */
1013 struct ib_srq *ib_create_srq_user(struct ib_pd *pd,
1014 				  struct ib_srq_init_attr *srq_init_attr,
1015 				  struct ib_usrq_object *uobject,
1016 				  struct ib_udata *udata)
1017 {
1018 	struct ib_srq *srq;
1019 	int ret;
1020 
1021 	srq = rdma_zalloc_drv_obj(pd->device, ib_srq);
1022 	if (!srq)
1023 		return ERR_PTR(-ENOMEM);
1024 
1025 	srq->device = pd->device;
1026 	srq->pd = pd;
1027 	srq->event_handler = srq_init_attr->event_handler;
1028 	srq->srq_context = srq_init_attr->srq_context;
1029 	srq->srq_type = srq_init_attr->srq_type;
1030 	srq->uobject = uobject;
1031 
1032 	if (ib_srq_has_cq(srq->srq_type)) {
1033 		srq->ext.cq = srq_init_attr->ext.cq;
1034 		atomic_inc(&srq->ext.cq->usecnt);
1035 	}
1036 	if (srq->srq_type == IB_SRQT_XRC) {
1037 		srq->ext.xrc.xrcd = srq_init_attr->ext.xrc.xrcd;
1038 		if (srq->ext.xrc.xrcd)
1039 			atomic_inc(&srq->ext.xrc.xrcd->usecnt);
1040 	}
1041 	atomic_inc(&pd->usecnt);
1042 
1043 	rdma_restrack_new(&srq->res, RDMA_RESTRACK_SRQ);
1044 	rdma_restrack_parent_name(&srq->res, &pd->res);
1045 
1046 	ret = pd->device->ops.create_srq(srq, srq_init_attr, udata);
1047 	if (ret) {
1048 		rdma_restrack_put(&srq->res);
1049 		atomic_dec(&srq->pd->usecnt);
1050 		if (srq->srq_type == IB_SRQT_XRC && srq->ext.xrc.xrcd)
1051 			atomic_dec(&srq->ext.xrc.xrcd->usecnt);
1052 		if (ib_srq_has_cq(srq->srq_type))
1053 			atomic_dec(&srq->ext.cq->usecnt);
1054 		kfree(srq);
1055 		return ERR_PTR(ret);
1056 	}
1057 
1058 	rdma_restrack_add(&srq->res);
1059 
1060 	return srq;
1061 }
1062 EXPORT_SYMBOL(ib_create_srq_user);
1063 
1064 int ib_modify_srq(struct ib_srq *srq,
1065 		  struct ib_srq_attr *srq_attr,
1066 		  enum ib_srq_attr_mask srq_attr_mask)
1067 {
1068 	return srq->device->ops.modify_srq ?
1069 		srq->device->ops.modify_srq(srq, srq_attr, srq_attr_mask,
1070 					    NULL) : -EOPNOTSUPP;
1071 }
1072 EXPORT_SYMBOL(ib_modify_srq);
1073 
1074 int ib_query_srq(struct ib_srq *srq,
1075 		 struct ib_srq_attr *srq_attr)
1076 {
1077 	return srq->device->ops.query_srq ?
1078 		srq->device->ops.query_srq(srq, srq_attr) : -EOPNOTSUPP;
1079 }
1080 EXPORT_SYMBOL(ib_query_srq);
1081 
1082 int ib_destroy_srq_user(struct ib_srq *srq, struct ib_udata *udata)
1083 {
1084 	int ret;
1085 
1086 	if (atomic_read(&srq->usecnt))
1087 		return -EBUSY;
1088 
1089 	ret = srq->device->ops.destroy_srq(srq, udata);
1090 	if (ret)
1091 		return ret;
1092 
1093 	atomic_dec(&srq->pd->usecnt);
1094 	if (srq->srq_type == IB_SRQT_XRC && srq->ext.xrc.xrcd)
1095 		atomic_dec(&srq->ext.xrc.xrcd->usecnt);
1096 	if (ib_srq_has_cq(srq->srq_type))
1097 		atomic_dec(&srq->ext.cq->usecnt);
1098 	rdma_restrack_del(&srq->res);
1099 	kfree(srq);
1100 
1101 	return ret;
1102 }
1103 EXPORT_SYMBOL(ib_destroy_srq_user);
1104 
1105 /* Queue pairs */
1106 
1107 static void __ib_shared_qp_event_handler(struct ib_event *event, void *context)
1108 {
1109 	struct ib_qp *qp = context;
1110 	unsigned long flags;
1111 
1112 	spin_lock_irqsave(&qp->device->qp_open_list_lock, flags);
1113 	list_for_each_entry(event->element.qp, &qp->open_list, open_list)
1114 		if (event->element.qp->event_handler)
1115 			event->element.qp->event_handler(event, event->element.qp->qp_context);
1116 	spin_unlock_irqrestore(&qp->device->qp_open_list_lock, flags);
1117 }
1118 
1119 static struct ib_qp *__ib_open_qp(struct ib_qp *real_qp,
1120 				  void (*event_handler)(struct ib_event *, void *),
1121 				  void *qp_context)
1122 {
1123 	struct ib_qp *qp;
1124 	unsigned long flags;
1125 	int err;
1126 
1127 	qp = kzalloc(sizeof *qp, GFP_KERNEL);
1128 	if (!qp)
1129 		return ERR_PTR(-ENOMEM);
1130 
1131 	qp->real_qp = real_qp;
1132 	err = ib_open_shared_qp_security(qp, real_qp->device);
1133 	if (err) {
1134 		kfree(qp);
1135 		return ERR_PTR(err);
1136 	}
1137 
1138 	qp->real_qp = real_qp;
1139 	atomic_inc(&real_qp->usecnt);
1140 	qp->device = real_qp->device;
1141 	qp->event_handler = event_handler;
1142 	qp->qp_context = qp_context;
1143 	qp->qp_num = real_qp->qp_num;
1144 	qp->qp_type = real_qp->qp_type;
1145 
1146 	spin_lock_irqsave(&real_qp->device->qp_open_list_lock, flags);
1147 	list_add(&qp->open_list, &real_qp->open_list);
1148 	spin_unlock_irqrestore(&real_qp->device->qp_open_list_lock, flags);
1149 
1150 	return qp;
1151 }
1152 
1153 struct ib_qp *ib_open_qp(struct ib_xrcd *xrcd,
1154 			 struct ib_qp_open_attr *qp_open_attr)
1155 {
1156 	struct ib_qp *qp, *real_qp;
1157 
1158 	if (qp_open_attr->qp_type != IB_QPT_XRC_TGT)
1159 		return ERR_PTR(-EINVAL);
1160 
1161 	down_read(&xrcd->tgt_qps_rwsem);
1162 	real_qp = xa_load(&xrcd->tgt_qps, qp_open_attr->qp_num);
1163 	if (!real_qp) {
1164 		up_read(&xrcd->tgt_qps_rwsem);
1165 		return ERR_PTR(-EINVAL);
1166 	}
1167 	qp = __ib_open_qp(real_qp, qp_open_attr->event_handler,
1168 			  qp_open_attr->qp_context);
1169 	up_read(&xrcd->tgt_qps_rwsem);
1170 	return qp;
1171 }
1172 EXPORT_SYMBOL(ib_open_qp);
1173 
1174 static struct ib_qp *create_xrc_qp_user(struct ib_qp *qp,
1175 					struct ib_qp_init_attr *qp_init_attr)
1176 {
1177 	struct ib_qp *real_qp = qp;
1178 	int err;
1179 
1180 	qp->event_handler = __ib_shared_qp_event_handler;
1181 	qp->qp_context = qp;
1182 	qp->pd = NULL;
1183 	qp->send_cq = qp->recv_cq = NULL;
1184 	qp->srq = NULL;
1185 	qp->xrcd = qp_init_attr->xrcd;
1186 	atomic_inc(&qp_init_attr->xrcd->usecnt);
1187 	INIT_LIST_HEAD(&qp->open_list);
1188 
1189 	qp = __ib_open_qp(real_qp, qp_init_attr->event_handler,
1190 			  qp_init_attr->qp_context);
1191 	if (IS_ERR(qp))
1192 		return qp;
1193 
1194 	err = xa_err(xa_store(&qp_init_attr->xrcd->tgt_qps, real_qp->qp_num,
1195 			      real_qp, GFP_KERNEL));
1196 	if (err) {
1197 		ib_close_qp(qp);
1198 		return ERR_PTR(err);
1199 	}
1200 	return qp;
1201 }
1202 
1203 static struct ib_qp *create_qp(struct ib_device *dev, struct ib_pd *pd,
1204 			       struct ib_qp_init_attr *attr,
1205 			       struct ib_udata *udata,
1206 			       struct ib_uqp_object *uobj, const char *caller)
1207 {
1208 	struct ib_udata dummy = {};
1209 	struct ib_qp *qp;
1210 	int ret;
1211 
1212 	if (!dev->ops.create_qp)
1213 		return ERR_PTR(-EOPNOTSUPP);
1214 
1215 	qp = rdma_zalloc_drv_obj_numa(dev, ib_qp);
1216 	if (!qp)
1217 		return ERR_PTR(-ENOMEM);
1218 
1219 	qp->device = dev;
1220 	qp->pd = pd;
1221 	qp->uobject = uobj;
1222 	qp->real_qp = qp;
1223 
1224 	qp->qp_type = attr->qp_type;
1225 	qp->rwq_ind_tbl = attr->rwq_ind_tbl;
1226 	qp->srq = attr->srq;
1227 	qp->event_handler = attr->event_handler;
1228 	qp->port = attr->port_num;
1229 	qp->qp_context = attr->qp_context;
1230 
1231 	spin_lock_init(&qp->mr_lock);
1232 	INIT_LIST_HEAD(&qp->rdma_mrs);
1233 	INIT_LIST_HEAD(&qp->sig_mrs);
1234 
1235 	qp->send_cq = attr->send_cq;
1236 	qp->recv_cq = attr->recv_cq;
1237 
1238 	rdma_restrack_new(&qp->res, RDMA_RESTRACK_QP);
1239 	WARN_ONCE(!udata && !caller, "Missing kernel QP owner");
1240 	rdma_restrack_set_name(&qp->res, udata ? NULL : caller);
1241 	ret = dev->ops.create_qp(qp, attr, udata);
1242 	if (ret)
1243 		goto err_create;
1244 
1245 	/*
1246 	 * TODO: The mlx4 internally overwrites send_cq and recv_cq.
1247 	 * Unfortunately, it is not an easy task to fix that driver.
1248 	 */
1249 	qp->send_cq = attr->send_cq;
1250 	qp->recv_cq = attr->recv_cq;
1251 
1252 	ret = ib_create_qp_security(qp, dev);
1253 	if (ret)
1254 		goto err_security;
1255 
1256 	rdma_restrack_add(&qp->res);
1257 	return qp;
1258 
1259 err_security:
1260 	qp->device->ops.destroy_qp(qp, udata ? &dummy : NULL);
1261 err_create:
1262 	rdma_restrack_put(&qp->res);
1263 	kfree(qp);
1264 	return ERR_PTR(ret);
1265 
1266 }
1267 
1268 /**
1269  * ib_create_qp_user - Creates a QP associated with the specified protection
1270  *   domain.
1271  * @dev: IB device
1272  * @pd: The protection domain associated with the QP.
1273  * @attr: A list of initial attributes required to create the
1274  *   QP.  If QP creation succeeds, then the attributes are updated to
1275  *   the actual capabilities of the created QP.
1276  * @udata: User data
1277  * @uobj: uverbs obect
1278  * @caller: caller's build-time module name
1279  */
1280 struct ib_qp *ib_create_qp_user(struct ib_device *dev, struct ib_pd *pd,
1281 				struct ib_qp_init_attr *attr,
1282 				struct ib_udata *udata,
1283 				struct ib_uqp_object *uobj, const char *caller)
1284 {
1285 	struct ib_qp *qp, *xrc_qp;
1286 
1287 	if (attr->qp_type == IB_QPT_XRC_TGT)
1288 		qp = create_qp(dev, pd, attr, NULL, NULL, caller);
1289 	else
1290 		qp = create_qp(dev, pd, attr, udata, uobj, NULL);
1291 	if (attr->qp_type != IB_QPT_XRC_TGT || IS_ERR(qp))
1292 		return qp;
1293 
1294 	xrc_qp = create_xrc_qp_user(qp, attr);
1295 	if (IS_ERR(xrc_qp)) {
1296 		ib_destroy_qp(qp);
1297 		return xrc_qp;
1298 	}
1299 
1300 	xrc_qp->uobject = uobj;
1301 	return xrc_qp;
1302 }
1303 EXPORT_SYMBOL(ib_create_qp_user);
1304 
1305 void ib_qp_usecnt_inc(struct ib_qp *qp)
1306 {
1307 	if (qp->pd)
1308 		atomic_inc(&qp->pd->usecnt);
1309 	if (qp->send_cq)
1310 		atomic_inc(&qp->send_cq->usecnt);
1311 	if (qp->recv_cq)
1312 		atomic_inc(&qp->recv_cq->usecnt);
1313 	if (qp->srq)
1314 		atomic_inc(&qp->srq->usecnt);
1315 	if (qp->rwq_ind_tbl)
1316 		atomic_inc(&qp->rwq_ind_tbl->usecnt);
1317 }
1318 EXPORT_SYMBOL(ib_qp_usecnt_inc);
1319 
1320 void ib_qp_usecnt_dec(struct ib_qp *qp)
1321 {
1322 	if (qp->rwq_ind_tbl)
1323 		atomic_dec(&qp->rwq_ind_tbl->usecnt);
1324 	if (qp->srq)
1325 		atomic_dec(&qp->srq->usecnt);
1326 	if (qp->recv_cq)
1327 		atomic_dec(&qp->recv_cq->usecnt);
1328 	if (qp->send_cq)
1329 		atomic_dec(&qp->send_cq->usecnt);
1330 	if (qp->pd)
1331 		atomic_dec(&qp->pd->usecnt);
1332 }
1333 EXPORT_SYMBOL(ib_qp_usecnt_dec);
1334 
1335 struct ib_qp *ib_create_qp_kernel(struct ib_pd *pd,
1336 				  struct ib_qp_init_attr *qp_init_attr,
1337 				  const char *caller)
1338 {
1339 	struct ib_device *device = pd->device;
1340 	struct ib_qp *qp;
1341 	int ret;
1342 
1343 	/*
1344 	 * If the callers is using the RDMA API calculate the resources
1345 	 * needed for the RDMA READ/WRITE operations.
1346 	 *
1347 	 * Note that these callers need to pass in a port number.
1348 	 */
1349 	if (qp_init_attr->cap.max_rdma_ctxs)
1350 		rdma_rw_init_qp(device, qp_init_attr);
1351 
1352 	qp = create_qp(device, pd, qp_init_attr, NULL, NULL, caller);
1353 	if (IS_ERR(qp))
1354 		return qp;
1355 
1356 	ib_qp_usecnt_inc(qp);
1357 
1358 	if (qp_init_attr->cap.max_rdma_ctxs) {
1359 		ret = rdma_rw_init_mrs(qp, qp_init_attr);
1360 		if (ret)
1361 			goto err;
1362 	}
1363 
1364 	/*
1365 	 * Note: all hw drivers guarantee that max_send_sge is lower than
1366 	 * the device RDMA WRITE SGE limit but not all hw drivers ensure that
1367 	 * max_send_sge <= max_sge_rd.
1368 	 */
1369 	qp->max_write_sge = qp_init_attr->cap.max_send_sge;
1370 	qp->max_read_sge = min_t(u32, qp_init_attr->cap.max_send_sge,
1371 				 device->attrs.max_sge_rd);
1372 	if (qp_init_attr->create_flags & IB_QP_CREATE_INTEGRITY_EN)
1373 		qp->integrity_en = true;
1374 
1375 	return qp;
1376 
1377 err:
1378 	ib_destroy_qp(qp);
1379 	return ERR_PTR(ret);
1380 
1381 }
1382 EXPORT_SYMBOL(ib_create_qp_kernel);
1383 
1384 static const struct {
1385 	int			valid;
1386 	enum ib_qp_attr_mask	req_param[IB_QPT_MAX];
1387 	enum ib_qp_attr_mask	opt_param[IB_QPT_MAX];
1388 } qp_state_table[IB_QPS_ERR + 1][IB_QPS_ERR + 1] = {
1389 	[IB_QPS_RESET] = {
1390 		[IB_QPS_RESET] = { .valid = 1 },
1391 		[IB_QPS_INIT]  = {
1392 			.valid = 1,
1393 			.req_param = {
1394 				[IB_QPT_UD]  = (IB_QP_PKEY_INDEX		|
1395 						IB_QP_PORT			|
1396 						IB_QP_QKEY),
1397 				[IB_QPT_RAW_PACKET] = IB_QP_PORT,
1398 				[IB_QPT_UC]  = (IB_QP_PKEY_INDEX		|
1399 						IB_QP_PORT			|
1400 						IB_QP_ACCESS_FLAGS),
1401 				[IB_QPT_RC]  = (IB_QP_PKEY_INDEX		|
1402 						IB_QP_PORT			|
1403 						IB_QP_ACCESS_FLAGS),
1404 				[IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX		|
1405 						IB_QP_PORT			|
1406 						IB_QP_ACCESS_FLAGS),
1407 				[IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX		|
1408 						IB_QP_PORT			|
1409 						IB_QP_ACCESS_FLAGS),
1410 				[IB_QPT_SMI] = (IB_QP_PKEY_INDEX		|
1411 						IB_QP_QKEY),
1412 				[IB_QPT_GSI] = (IB_QP_PKEY_INDEX		|
1413 						IB_QP_QKEY),
1414 			}
1415 		},
1416 	},
1417 	[IB_QPS_INIT]  = {
1418 		[IB_QPS_RESET] = { .valid = 1 },
1419 		[IB_QPS_ERR] =   { .valid = 1 },
1420 		[IB_QPS_INIT]  = {
1421 			.valid = 1,
1422 			.opt_param = {
1423 				[IB_QPT_UD]  = (IB_QP_PKEY_INDEX		|
1424 						IB_QP_PORT			|
1425 						IB_QP_QKEY),
1426 				[IB_QPT_UC]  = (IB_QP_PKEY_INDEX		|
1427 						IB_QP_PORT			|
1428 						IB_QP_ACCESS_FLAGS),
1429 				[IB_QPT_RC]  = (IB_QP_PKEY_INDEX		|
1430 						IB_QP_PORT			|
1431 						IB_QP_ACCESS_FLAGS),
1432 				[IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX		|
1433 						IB_QP_PORT			|
1434 						IB_QP_ACCESS_FLAGS),
1435 				[IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX		|
1436 						IB_QP_PORT			|
1437 						IB_QP_ACCESS_FLAGS),
1438 				[IB_QPT_SMI] = (IB_QP_PKEY_INDEX		|
1439 						IB_QP_QKEY),
1440 				[IB_QPT_GSI] = (IB_QP_PKEY_INDEX		|
1441 						IB_QP_QKEY),
1442 			}
1443 		},
1444 		[IB_QPS_RTR]   = {
1445 			.valid = 1,
1446 			.req_param = {
1447 				[IB_QPT_UC]  = (IB_QP_AV			|
1448 						IB_QP_PATH_MTU			|
1449 						IB_QP_DEST_QPN			|
1450 						IB_QP_RQ_PSN),
1451 				[IB_QPT_RC]  = (IB_QP_AV			|
1452 						IB_QP_PATH_MTU			|
1453 						IB_QP_DEST_QPN			|
1454 						IB_QP_RQ_PSN			|
1455 						IB_QP_MAX_DEST_RD_ATOMIC	|
1456 						IB_QP_MIN_RNR_TIMER),
1457 				[IB_QPT_XRC_INI] = (IB_QP_AV			|
1458 						IB_QP_PATH_MTU			|
1459 						IB_QP_DEST_QPN			|
1460 						IB_QP_RQ_PSN),
1461 				[IB_QPT_XRC_TGT] = (IB_QP_AV			|
1462 						IB_QP_PATH_MTU			|
1463 						IB_QP_DEST_QPN			|
1464 						IB_QP_RQ_PSN			|
1465 						IB_QP_MAX_DEST_RD_ATOMIC	|
1466 						IB_QP_MIN_RNR_TIMER),
1467 			},
1468 			.opt_param = {
1469 				 [IB_QPT_UD]  = (IB_QP_PKEY_INDEX		|
1470 						 IB_QP_QKEY),
1471 				 [IB_QPT_UC]  = (IB_QP_ALT_PATH			|
1472 						 IB_QP_ACCESS_FLAGS		|
1473 						 IB_QP_PKEY_INDEX),
1474 				 [IB_QPT_RC]  = (IB_QP_ALT_PATH			|
1475 						 IB_QP_ACCESS_FLAGS		|
1476 						 IB_QP_PKEY_INDEX),
1477 				 [IB_QPT_XRC_INI] = (IB_QP_ALT_PATH		|
1478 						 IB_QP_ACCESS_FLAGS		|
1479 						 IB_QP_PKEY_INDEX),
1480 				 [IB_QPT_XRC_TGT] = (IB_QP_ALT_PATH		|
1481 						 IB_QP_ACCESS_FLAGS		|
1482 						 IB_QP_PKEY_INDEX),
1483 				 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX		|
1484 						 IB_QP_QKEY),
1485 				 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX		|
1486 						 IB_QP_QKEY),
1487 			 },
1488 		},
1489 	},
1490 	[IB_QPS_RTR]   = {
1491 		[IB_QPS_RESET] = { .valid = 1 },
1492 		[IB_QPS_ERR] =   { .valid = 1 },
1493 		[IB_QPS_RTS]   = {
1494 			.valid = 1,
1495 			.req_param = {
1496 				[IB_QPT_UD]  = IB_QP_SQ_PSN,
1497 				[IB_QPT_UC]  = IB_QP_SQ_PSN,
1498 				[IB_QPT_RC]  = (IB_QP_TIMEOUT			|
1499 						IB_QP_RETRY_CNT			|
1500 						IB_QP_RNR_RETRY			|
1501 						IB_QP_SQ_PSN			|
1502 						IB_QP_MAX_QP_RD_ATOMIC),
1503 				[IB_QPT_XRC_INI] = (IB_QP_TIMEOUT		|
1504 						IB_QP_RETRY_CNT			|
1505 						IB_QP_RNR_RETRY			|
1506 						IB_QP_SQ_PSN			|
1507 						IB_QP_MAX_QP_RD_ATOMIC),
1508 				[IB_QPT_XRC_TGT] = (IB_QP_TIMEOUT		|
1509 						IB_QP_SQ_PSN),
1510 				[IB_QPT_SMI] = IB_QP_SQ_PSN,
1511 				[IB_QPT_GSI] = IB_QP_SQ_PSN,
1512 			},
1513 			.opt_param = {
1514 				 [IB_QPT_UD]  = (IB_QP_CUR_STATE		|
1515 						 IB_QP_QKEY),
1516 				 [IB_QPT_UC]  = (IB_QP_CUR_STATE		|
1517 						 IB_QP_ALT_PATH			|
1518 						 IB_QP_ACCESS_FLAGS		|
1519 						 IB_QP_PATH_MIG_STATE),
1520 				 [IB_QPT_RC]  = (IB_QP_CUR_STATE		|
1521 						 IB_QP_ALT_PATH			|
1522 						 IB_QP_ACCESS_FLAGS		|
1523 						 IB_QP_MIN_RNR_TIMER		|
1524 						 IB_QP_PATH_MIG_STATE),
1525 				 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE		|
1526 						 IB_QP_ALT_PATH			|
1527 						 IB_QP_ACCESS_FLAGS		|
1528 						 IB_QP_PATH_MIG_STATE),
1529 				 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE		|
1530 						 IB_QP_ALT_PATH			|
1531 						 IB_QP_ACCESS_FLAGS		|
1532 						 IB_QP_MIN_RNR_TIMER		|
1533 						 IB_QP_PATH_MIG_STATE),
1534 				 [IB_QPT_SMI] = (IB_QP_CUR_STATE		|
1535 						 IB_QP_QKEY),
1536 				 [IB_QPT_GSI] = (IB_QP_CUR_STATE		|
1537 						 IB_QP_QKEY),
1538 				 [IB_QPT_RAW_PACKET] = IB_QP_RATE_LIMIT,
1539 			 }
1540 		}
1541 	},
1542 	[IB_QPS_RTS]   = {
1543 		[IB_QPS_RESET] = { .valid = 1 },
1544 		[IB_QPS_ERR] =   { .valid = 1 },
1545 		[IB_QPS_RTS]   = {
1546 			.valid = 1,
1547 			.opt_param = {
1548 				[IB_QPT_UD]  = (IB_QP_CUR_STATE			|
1549 						IB_QP_QKEY),
1550 				[IB_QPT_UC]  = (IB_QP_CUR_STATE			|
1551 						IB_QP_ACCESS_FLAGS		|
1552 						IB_QP_ALT_PATH			|
1553 						IB_QP_PATH_MIG_STATE),
1554 				[IB_QPT_RC]  = (IB_QP_CUR_STATE			|
1555 						IB_QP_ACCESS_FLAGS		|
1556 						IB_QP_ALT_PATH			|
1557 						IB_QP_PATH_MIG_STATE		|
1558 						IB_QP_MIN_RNR_TIMER),
1559 				[IB_QPT_XRC_INI] = (IB_QP_CUR_STATE		|
1560 						IB_QP_ACCESS_FLAGS		|
1561 						IB_QP_ALT_PATH			|
1562 						IB_QP_PATH_MIG_STATE),
1563 				[IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE		|
1564 						IB_QP_ACCESS_FLAGS		|
1565 						IB_QP_ALT_PATH			|
1566 						IB_QP_PATH_MIG_STATE		|
1567 						IB_QP_MIN_RNR_TIMER),
1568 				[IB_QPT_SMI] = (IB_QP_CUR_STATE			|
1569 						IB_QP_QKEY),
1570 				[IB_QPT_GSI] = (IB_QP_CUR_STATE			|
1571 						IB_QP_QKEY),
1572 				[IB_QPT_RAW_PACKET] = IB_QP_RATE_LIMIT,
1573 			}
1574 		},
1575 		[IB_QPS_SQD]   = {
1576 			.valid = 1,
1577 			.opt_param = {
1578 				[IB_QPT_UD]  = IB_QP_EN_SQD_ASYNC_NOTIFY,
1579 				[IB_QPT_UC]  = IB_QP_EN_SQD_ASYNC_NOTIFY,
1580 				[IB_QPT_RC]  = IB_QP_EN_SQD_ASYNC_NOTIFY,
1581 				[IB_QPT_XRC_INI] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1582 				[IB_QPT_XRC_TGT] = IB_QP_EN_SQD_ASYNC_NOTIFY, /* ??? */
1583 				[IB_QPT_SMI] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1584 				[IB_QPT_GSI] = IB_QP_EN_SQD_ASYNC_NOTIFY
1585 			}
1586 		},
1587 	},
1588 	[IB_QPS_SQD]   = {
1589 		[IB_QPS_RESET] = { .valid = 1 },
1590 		[IB_QPS_ERR] =   { .valid = 1 },
1591 		[IB_QPS_RTS]   = {
1592 			.valid = 1,
1593 			.opt_param = {
1594 				[IB_QPT_UD]  = (IB_QP_CUR_STATE			|
1595 						IB_QP_QKEY),
1596 				[IB_QPT_UC]  = (IB_QP_CUR_STATE			|
1597 						IB_QP_ALT_PATH			|
1598 						IB_QP_ACCESS_FLAGS		|
1599 						IB_QP_PATH_MIG_STATE),
1600 				[IB_QPT_RC]  = (IB_QP_CUR_STATE			|
1601 						IB_QP_ALT_PATH			|
1602 						IB_QP_ACCESS_FLAGS		|
1603 						IB_QP_MIN_RNR_TIMER		|
1604 						IB_QP_PATH_MIG_STATE),
1605 				[IB_QPT_XRC_INI] = (IB_QP_CUR_STATE		|
1606 						IB_QP_ALT_PATH			|
1607 						IB_QP_ACCESS_FLAGS		|
1608 						IB_QP_PATH_MIG_STATE),
1609 				[IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE		|
1610 						IB_QP_ALT_PATH			|
1611 						IB_QP_ACCESS_FLAGS		|
1612 						IB_QP_MIN_RNR_TIMER		|
1613 						IB_QP_PATH_MIG_STATE),
1614 				[IB_QPT_SMI] = (IB_QP_CUR_STATE			|
1615 						IB_QP_QKEY),
1616 				[IB_QPT_GSI] = (IB_QP_CUR_STATE			|
1617 						IB_QP_QKEY),
1618 			}
1619 		},
1620 		[IB_QPS_SQD]   = {
1621 			.valid = 1,
1622 			.opt_param = {
1623 				[IB_QPT_UD]  = (IB_QP_PKEY_INDEX		|
1624 						IB_QP_QKEY),
1625 				[IB_QPT_UC]  = (IB_QP_AV			|
1626 						IB_QP_ALT_PATH			|
1627 						IB_QP_ACCESS_FLAGS		|
1628 						IB_QP_PKEY_INDEX		|
1629 						IB_QP_PATH_MIG_STATE),
1630 				[IB_QPT_RC]  = (IB_QP_PORT			|
1631 						IB_QP_AV			|
1632 						IB_QP_TIMEOUT			|
1633 						IB_QP_RETRY_CNT			|
1634 						IB_QP_RNR_RETRY			|
1635 						IB_QP_MAX_QP_RD_ATOMIC		|
1636 						IB_QP_MAX_DEST_RD_ATOMIC	|
1637 						IB_QP_ALT_PATH			|
1638 						IB_QP_ACCESS_FLAGS		|
1639 						IB_QP_PKEY_INDEX		|
1640 						IB_QP_MIN_RNR_TIMER		|
1641 						IB_QP_PATH_MIG_STATE),
1642 				[IB_QPT_XRC_INI] = (IB_QP_PORT			|
1643 						IB_QP_AV			|
1644 						IB_QP_TIMEOUT			|
1645 						IB_QP_RETRY_CNT			|
1646 						IB_QP_RNR_RETRY			|
1647 						IB_QP_MAX_QP_RD_ATOMIC		|
1648 						IB_QP_ALT_PATH			|
1649 						IB_QP_ACCESS_FLAGS		|
1650 						IB_QP_PKEY_INDEX		|
1651 						IB_QP_PATH_MIG_STATE),
1652 				[IB_QPT_XRC_TGT] = (IB_QP_PORT			|
1653 						IB_QP_AV			|
1654 						IB_QP_TIMEOUT			|
1655 						IB_QP_MAX_DEST_RD_ATOMIC	|
1656 						IB_QP_ALT_PATH			|
1657 						IB_QP_ACCESS_FLAGS		|
1658 						IB_QP_PKEY_INDEX		|
1659 						IB_QP_MIN_RNR_TIMER		|
1660 						IB_QP_PATH_MIG_STATE),
1661 				[IB_QPT_SMI] = (IB_QP_PKEY_INDEX		|
1662 						IB_QP_QKEY),
1663 				[IB_QPT_GSI] = (IB_QP_PKEY_INDEX		|
1664 						IB_QP_QKEY),
1665 			}
1666 		}
1667 	},
1668 	[IB_QPS_SQE]   = {
1669 		[IB_QPS_RESET] = { .valid = 1 },
1670 		[IB_QPS_ERR] =   { .valid = 1 },
1671 		[IB_QPS_RTS]   = {
1672 			.valid = 1,
1673 			.opt_param = {
1674 				[IB_QPT_UD]  = (IB_QP_CUR_STATE			|
1675 						IB_QP_QKEY),
1676 				[IB_QPT_UC]  = (IB_QP_CUR_STATE			|
1677 						IB_QP_ACCESS_FLAGS),
1678 				[IB_QPT_SMI] = (IB_QP_CUR_STATE			|
1679 						IB_QP_QKEY),
1680 				[IB_QPT_GSI] = (IB_QP_CUR_STATE			|
1681 						IB_QP_QKEY),
1682 			}
1683 		}
1684 	},
1685 	[IB_QPS_ERR] = {
1686 		[IB_QPS_RESET] = { .valid = 1 },
1687 		[IB_QPS_ERR] =   { .valid = 1 }
1688 	}
1689 };
1690 
1691 bool ib_modify_qp_is_ok(enum ib_qp_state cur_state, enum ib_qp_state next_state,
1692 			enum ib_qp_type type, enum ib_qp_attr_mask mask)
1693 {
1694 	enum ib_qp_attr_mask req_param, opt_param;
1695 
1696 	if (mask & IB_QP_CUR_STATE  &&
1697 	    cur_state != IB_QPS_RTR && cur_state != IB_QPS_RTS &&
1698 	    cur_state != IB_QPS_SQD && cur_state != IB_QPS_SQE)
1699 		return false;
1700 
1701 	if (!qp_state_table[cur_state][next_state].valid)
1702 		return false;
1703 
1704 	req_param = qp_state_table[cur_state][next_state].req_param[type];
1705 	opt_param = qp_state_table[cur_state][next_state].opt_param[type];
1706 
1707 	if ((mask & req_param) != req_param)
1708 		return false;
1709 
1710 	if (mask & ~(req_param | opt_param | IB_QP_STATE))
1711 		return false;
1712 
1713 	return true;
1714 }
1715 EXPORT_SYMBOL(ib_modify_qp_is_ok);
1716 
1717 /**
1718  * ib_resolve_eth_dmac - Resolve destination mac address
1719  * @device:		Device to consider
1720  * @ah_attr:		address handle attribute which describes the
1721  *			source and destination parameters
1722  * ib_resolve_eth_dmac() resolves destination mac address and L3 hop limit It
1723  * returns 0 on success or appropriate error code. It initializes the
1724  * necessary ah_attr fields when call is successful.
1725  */
1726 static int ib_resolve_eth_dmac(struct ib_device *device,
1727 			       struct rdma_ah_attr *ah_attr)
1728 {
1729 	int ret = 0;
1730 
1731 	if (rdma_is_multicast_addr((struct in6_addr *)ah_attr->grh.dgid.raw)) {
1732 		if (ipv6_addr_v4mapped((struct in6_addr *)ah_attr->grh.dgid.raw)) {
1733 			__be32 addr = 0;
1734 
1735 			memcpy(&addr, ah_attr->grh.dgid.raw + 12, 4);
1736 			ip_eth_mc_map(addr, (char *)ah_attr->roce.dmac);
1737 		} else {
1738 			ipv6_eth_mc_map((struct in6_addr *)ah_attr->grh.dgid.raw,
1739 					(char *)ah_attr->roce.dmac);
1740 		}
1741 	} else {
1742 		ret = ib_resolve_unicast_gid_dmac(device, ah_attr);
1743 	}
1744 	return ret;
1745 }
1746 
1747 static bool is_qp_type_connected(const struct ib_qp *qp)
1748 {
1749 	return (qp->qp_type == IB_QPT_UC ||
1750 		qp->qp_type == IB_QPT_RC ||
1751 		qp->qp_type == IB_QPT_XRC_INI ||
1752 		qp->qp_type == IB_QPT_XRC_TGT);
1753 }
1754 
1755 /*
1756  * IB core internal function to perform QP attributes modification.
1757  */
1758 static int _ib_modify_qp(struct ib_qp *qp, struct ib_qp_attr *attr,
1759 			 int attr_mask, struct ib_udata *udata)
1760 {
1761 	u32 port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
1762 	const struct ib_gid_attr *old_sgid_attr_av;
1763 	const struct ib_gid_attr *old_sgid_attr_alt_av;
1764 	int ret;
1765 
1766 	attr->xmit_slave = NULL;
1767 	if (attr_mask & IB_QP_AV) {
1768 		ret = rdma_fill_sgid_attr(qp->device, &attr->ah_attr,
1769 					  &old_sgid_attr_av);
1770 		if (ret)
1771 			return ret;
1772 
1773 		if (attr->ah_attr.type == RDMA_AH_ATTR_TYPE_ROCE &&
1774 		    is_qp_type_connected(qp)) {
1775 			struct net_device *slave;
1776 
1777 			/*
1778 			 * If the user provided the qp_attr then we have to
1779 			 * resolve it. Kerne users have to provide already
1780 			 * resolved rdma_ah_attr's.
1781 			 */
1782 			if (udata) {
1783 				ret = ib_resolve_eth_dmac(qp->device,
1784 							  &attr->ah_attr);
1785 				if (ret)
1786 					goto out_av;
1787 			}
1788 			slave = rdma_lag_get_ah_roce_slave(qp->device,
1789 							   &attr->ah_attr,
1790 							   GFP_KERNEL);
1791 			if (IS_ERR(slave)) {
1792 				ret = PTR_ERR(slave);
1793 				goto out_av;
1794 			}
1795 			attr->xmit_slave = slave;
1796 		}
1797 	}
1798 	if (attr_mask & IB_QP_ALT_PATH) {
1799 		/*
1800 		 * FIXME: This does not track the migration state, so if the
1801 		 * user loads a new alternate path after the HW has migrated
1802 		 * from primary->alternate we will keep the wrong
1803 		 * references. This is OK for IB because the reference
1804 		 * counting does not serve any functional purpose.
1805 		 */
1806 		ret = rdma_fill_sgid_attr(qp->device, &attr->alt_ah_attr,
1807 					  &old_sgid_attr_alt_av);
1808 		if (ret)
1809 			goto out_av;
1810 
1811 		/*
1812 		 * Today the core code can only handle alternate paths and APM
1813 		 * for IB. Ban them in roce mode.
1814 		 */
1815 		if (!(rdma_protocol_ib(qp->device,
1816 				       attr->alt_ah_attr.port_num) &&
1817 		      rdma_protocol_ib(qp->device, port))) {
1818 			ret = -EINVAL;
1819 			goto out;
1820 		}
1821 	}
1822 
1823 	if (rdma_ib_or_roce(qp->device, port)) {
1824 		if (attr_mask & IB_QP_RQ_PSN && attr->rq_psn & ~0xffffff) {
1825 			dev_warn(&qp->device->dev,
1826 				 "%s rq_psn overflow, masking to 24 bits\n",
1827 				 __func__);
1828 			attr->rq_psn &= 0xffffff;
1829 		}
1830 
1831 		if (attr_mask & IB_QP_SQ_PSN && attr->sq_psn & ~0xffffff) {
1832 			dev_warn(&qp->device->dev,
1833 				 " %s sq_psn overflow, masking to 24 bits\n",
1834 				 __func__);
1835 			attr->sq_psn &= 0xffffff;
1836 		}
1837 	}
1838 
1839 	/*
1840 	 * Bind this qp to a counter automatically based on the rdma counter
1841 	 * rules. This only set in RST2INIT with port specified
1842 	 */
1843 	if (!qp->counter && (attr_mask & IB_QP_PORT) &&
1844 	    ((attr_mask & IB_QP_STATE) && attr->qp_state == IB_QPS_INIT))
1845 		rdma_counter_bind_qp_auto(qp, attr->port_num);
1846 
1847 	ret = ib_security_modify_qp(qp, attr, attr_mask, udata);
1848 	if (ret)
1849 		goto out;
1850 
1851 	if (attr_mask & IB_QP_PORT)
1852 		qp->port = attr->port_num;
1853 	if (attr_mask & IB_QP_AV)
1854 		qp->av_sgid_attr =
1855 			rdma_update_sgid_attr(&attr->ah_attr, qp->av_sgid_attr);
1856 	if (attr_mask & IB_QP_ALT_PATH)
1857 		qp->alt_path_sgid_attr = rdma_update_sgid_attr(
1858 			&attr->alt_ah_attr, qp->alt_path_sgid_attr);
1859 
1860 out:
1861 	if (attr_mask & IB_QP_ALT_PATH)
1862 		rdma_unfill_sgid_attr(&attr->alt_ah_attr, old_sgid_attr_alt_av);
1863 out_av:
1864 	if (attr_mask & IB_QP_AV) {
1865 		rdma_lag_put_ah_roce_slave(attr->xmit_slave);
1866 		rdma_unfill_sgid_attr(&attr->ah_attr, old_sgid_attr_av);
1867 	}
1868 	return ret;
1869 }
1870 
1871 /**
1872  * ib_modify_qp_with_udata - Modifies the attributes for the specified QP.
1873  * @ib_qp: The QP to modify.
1874  * @attr: On input, specifies the QP attributes to modify.  On output,
1875  *   the current values of selected QP attributes are returned.
1876  * @attr_mask: A bit-mask used to specify which attributes of the QP
1877  *   are being modified.
1878  * @udata: pointer to user's input output buffer information
1879  *   are being modified.
1880  * It returns 0 on success and returns appropriate error code on error.
1881  */
1882 int ib_modify_qp_with_udata(struct ib_qp *ib_qp, struct ib_qp_attr *attr,
1883 			    int attr_mask, struct ib_udata *udata)
1884 {
1885 	return _ib_modify_qp(ib_qp->real_qp, attr, attr_mask, udata);
1886 }
1887 EXPORT_SYMBOL(ib_modify_qp_with_udata);
1888 
1889 int ib_get_eth_speed(struct ib_device *dev, u32 port_num, u16 *speed, u8 *width)
1890 {
1891 	int rc;
1892 	u32 netdev_speed;
1893 	struct net_device *netdev;
1894 	struct ethtool_link_ksettings lksettings;
1895 
1896 	if (rdma_port_get_link_layer(dev, port_num) != IB_LINK_LAYER_ETHERNET)
1897 		return -EINVAL;
1898 
1899 	netdev = ib_device_get_netdev(dev, port_num);
1900 	if (!netdev)
1901 		return -ENODEV;
1902 
1903 	rtnl_lock();
1904 	rc = __ethtool_get_link_ksettings(netdev, &lksettings);
1905 	rtnl_unlock();
1906 
1907 	dev_put(netdev);
1908 
1909 	if (!rc && lksettings.base.speed != (u32)SPEED_UNKNOWN) {
1910 		netdev_speed = lksettings.base.speed;
1911 	} else {
1912 		netdev_speed = SPEED_1000;
1913 		pr_warn("%s speed is unknown, defaulting to %u\n", netdev->name,
1914 			netdev_speed);
1915 	}
1916 
1917 	if (netdev_speed <= SPEED_1000) {
1918 		*width = IB_WIDTH_1X;
1919 		*speed = IB_SPEED_SDR;
1920 	} else if (netdev_speed <= SPEED_10000) {
1921 		*width = IB_WIDTH_1X;
1922 		*speed = IB_SPEED_FDR10;
1923 	} else if (netdev_speed <= SPEED_20000) {
1924 		*width = IB_WIDTH_4X;
1925 		*speed = IB_SPEED_DDR;
1926 	} else if (netdev_speed <= SPEED_25000) {
1927 		*width = IB_WIDTH_1X;
1928 		*speed = IB_SPEED_EDR;
1929 	} else if (netdev_speed <= SPEED_40000) {
1930 		*width = IB_WIDTH_4X;
1931 		*speed = IB_SPEED_FDR10;
1932 	} else {
1933 		*width = IB_WIDTH_4X;
1934 		*speed = IB_SPEED_EDR;
1935 	}
1936 
1937 	return 0;
1938 }
1939 EXPORT_SYMBOL(ib_get_eth_speed);
1940 
1941 int ib_modify_qp(struct ib_qp *qp,
1942 		 struct ib_qp_attr *qp_attr,
1943 		 int qp_attr_mask)
1944 {
1945 	return _ib_modify_qp(qp->real_qp, qp_attr, qp_attr_mask, NULL);
1946 }
1947 EXPORT_SYMBOL(ib_modify_qp);
1948 
1949 int ib_query_qp(struct ib_qp *qp,
1950 		struct ib_qp_attr *qp_attr,
1951 		int qp_attr_mask,
1952 		struct ib_qp_init_attr *qp_init_attr)
1953 {
1954 	qp_attr->ah_attr.grh.sgid_attr = NULL;
1955 	qp_attr->alt_ah_attr.grh.sgid_attr = NULL;
1956 
1957 	return qp->device->ops.query_qp ?
1958 		qp->device->ops.query_qp(qp->real_qp, qp_attr, qp_attr_mask,
1959 					 qp_init_attr) : -EOPNOTSUPP;
1960 }
1961 EXPORT_SYMBOL(ib_query_qp);
1962 
1963 int ib_close_qp(struct ib_qp *qp)
1964 {
1965 	struct ib_qp *real_qp;
1966 	unsigned long flags;
1967 
1968 	real_qp = qp->real_qp;
1969 	if (real_qp == qp)
1970 		return -EINVAL;
1971 
1972 	spin_lock_irqsave(&real_qp->device->qp_open_list_lock, flags);
1973 	list_del(&qp->open_list);
1974 	spin_unlock_irqrestore(&real_qp->device->qp_open_list_lock, flags);
1975 
1976 	atomic_dec(&real_qp->usecnt);
1977 	if (qp->qp_sec)
1978 		ib_close_shared_qp_security(qp->qp_sec);
1979 	kfree(qp);
1980 
1981 	return 0;
1982 }
1983 EXPORT_SYMBOL(ib_close_qp);
1984 
1985 static int __ib_destroy_shared_qp(struct ib_qp *qp)
1986 {
1987 	struct ib_xrcd *xrcd;
1988 	struct ib_qp *real_qp;
1989 	int ret;
1990 
1991 	real_qp = qp->real_qp;
1992 	xrcd = real_qp->xrcd;
1993 	down_write(&xrcd->tgt_qps_rwsem);
1994 	ib_close_qp(qp);
1995 	if (atomic_read(&real_qp->usecnt) == 0)
1996 		xa_erase(&xrcd->tgt_qps, real_qp->qp_num);
1997 	else
1998 		real_qp = NULL;
1999 	up_write(&xrcd->tgt_qps_rwsem);
2000 
2001 	if (real_qp) {
2002 		ret = ib_destroy_qp(real_qp);
2003 		if (!ret)
2004 			atomic_dec(&xrcd->usecnt);
2005 	}
2006 
2007 	return 0;
2008 }
2009 
2010 int ib_destroy_qp_user(struct ib_qp *qp, struct ib_udata *udata)
2011 {
2012 	const struct ib_gid_attr *alt_path_sgid_attr = qp->alt_path_sgid_attr;
2013 	const struct ib_gid_attr *av_sgid_attr = qp->av_sgid_attr;
2014 	struct ib_qp_security *sec;
2015 	int ret;
2016 
2017 	WARN_ON_ONCE(qp->mrs_used > 0);
2018 
2019 	if (atomic_read(&qp->usecnt))
2020 		return -EBUSY;
2021 
2022 	if (qp->real_qp != qp)
2023 		return __ib_destroy_shared_qp(qp);
2024 
2025 	sec  = qp->qp_sec;
2026 	if (sec)
2027 		ib_destroy_qp_security_begin(sec);
2028 
2029 	if (!qp->uobject)
2030 		rdma_rw_cleanup_mrs(qp);
2031 
2032 	rdma_counter_unbind_qp(qp, true);
2033 	ret = qp->device->ops.destroy_qp(qp, udata);
2034 	if (ret) {
2035 		if (sec)
2036 			ib_destroy_qp_security_abort(sec);
2037 		return ret;
2038 	}
2039 
2040 	if (alt_path_sgid_attr)
2041 		rdma_put_gid_attr(alt_path_sgid_attr);
2042 	if (av_sgid_attr)
2043 		rdma_put_gid_attr(av_sgid_attr);
2044 
2045 	ib_qp_usecnt_dec(qp);
2046 	if (sec)
2047 		ib_destroy_qp_security_end(sec);
2048 
2049 	rdma_restrack_del(&qp->res);
2050 	kfree(qp);
2051 	return ret;
2052 }
2053 EXPORT_SYMBOL(ib_destroy_qp_user);
2054 
2055 /* Completion queues */
2056 
2057 struct ib_cq *__ib_create_cq(struct ib_device *device,
2058 			     ib_comp_handler comp_handler,
2059 			     void (*event_handler)(struct ib_event *, void *),
2060 			     void *cq_context,
2061 			     const struct ib_cq_init_attr *cq_attr,
2062 			     const char *caller)
2063 {
2064 	struct ib_cq *cq;
2065 	int ret;
2066 
2067 	cq = rdma_zalloc_drv_obj(device, ib_cq);
2068 	if (!cq)
2069 		return ERR_PTR(-ENOMEM);
2070 
2071 	cq->device = device;
2072 	cq->uobject = NULL;
2073 	cq->comp_handler = comp_handler;
2074 	cq->event_handler = event_handler;
2075 	cq->cq_context = cq_context;
2076 	atomic_set(&cq->usecnt, 0);
2077 
2078 	rdma_restrack_new(&cq->res, RDMA_RESTRACK_CQ);
2079 	rdma_restrack_set_name(&cq->res, caller);
2080 
2081 	ret = device->ops.create_cq(cq, cq_attr, NULL);
2082 	if (ret) {
2083 		rdma_restrack_put(&cq->res);
2084 		kfree(cq);
2085 		return ERR_PTR(ret);
2086 	}
2087 
2088 	rdma_restrack_add(&cq->res);
2089 	return cq;
2090 }
2091 EXPORT_SYMBOL(__ib_create_cq);
2092 
2093 int rdma_set_cq_moderation(struct ib_cq *cq, u16 cq_count, u16 cq_period)
2094 {
2095 	if (cq->shared)
2096 		return -EOPNOTSUPP;
2097 
2098 	return cq->device->ops.modify_cq ?
2099 		cq->device->ops.modify_cq(cq, cq_count,
2100 					  cq_period) : -EOPNOTSUPP;
2101 }
2102 EXPORT_SYMBOL(rdma_set_cq_moderation);
2103 
2104 int ib_destroy_cq_user(struct ib_cq *cq, struct ib_udata *udata)
2105 {
2106 	int ret;
2107 
2108 	if (WARN_ON_ONCE(cq->shared))
2109 		return -EOPNOTSUPP;
2110 
2111 	if (atomic_read(&cq->usecnt))
2112 		return -EBUSY;
2113 
2114 	ret = cq->device->ops.destroy_cq(cq, udata);
2115 	if (ret)
2116 		return ret;
2117 
2118 	rdma_restrack_del(&cq->res);
2119 	kfree(cq);
2120 	return ret;
2121 }
2122 EXPORT_SYMBOL(ib_destroy_cq_user);
2123 
2124 int ib_resize_cq(struct ib_cq *cq, int cqe)
2125 {
2126 	if (cq->shared)
2127 		return -EOPNOTSUPP;
2128 
2129 	return cq->device->ops.resize_cq ?
2130 		cq->device->ops.resize_cq(cq, cqe, NULL) : -EOPNOTSUPP;
2131 }
2132 EXPORT_SYMBOL(ib_resize_cq);
2133 
2134 /* Memory regions */
2135 
2136 struct ib_mr *ib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
2137 			     u64 virt_addr, int access_flags)
2138 {
2139 	struct ib_mr *mr;
2140 
2141 	if (access_flags & IB_ACCESS_ON_DEMAND) {
2142 		if (!(pd->device->attrs.device_cap_flags &
2143 		      IB_DEVICE_ON_DEMAND_PAGING)) {
2144 			pr_debug("ODP support not available\n");
2145 			return ERR_PTR(-EINVAL);
2146 		}
2147 	}
2148 
2149 	mr = pd->device->ops.reg_user_mr(pd, start, length, virt_addr,
2150 					 access_flags, NULL);
2151 
2152 	if (IS_ERR(mr))
2153 		return mr;
2154 
2155 	mr->device = pd->device;
2156 	mr->pd = pd;
2157 	mr->dm = NULL;
2158 	atomic_inc(&pd->usecnt);
2159 
2160 	rdma_restrack_new(&mr->res, RDMA_RESTRACK_MR);
2161 	rdma_restrack_parent_name(&mr->res, &pd->res);
2162 	rdma_restrack_add(&mr->res);
2163 
2164 	return mr;
2165 }
2166 EXPORT_SYMBOL(ib_reg_user_mr);
2167 
2168 int ib_advise_mr(struct ib_pd *pd, enum ib_uverbs_advise_mr_advice advice,
2169 		 u32 flags, struct ib_sge *sg_list, u32 num_sge)
2170 {
2171 	if (!pd->device->ops.advise_mr)
2172 		return -EOPNOTSUPP;
2173 
2174 	if (!num_sge)
2175 		return 0;
2176 
2177 	return pd->device->ops.advise_mr(pd, advice, flags, sg_list, num_sge,
2178 					 NULL);
2179 }
2180 EXPORT_SYMBOL(ib_advise_mr);
2181 
2182 int ib_dereg_mr_user(struct ib_mr *mr, struct ib_udata *udata)
2183 {
2184 	struct ib_pd *pd = mr->pd;
2185 	struct ib_dm *dm = mr->dm;
2186 	struct ib_sig_attrs *sig_attrs = mr->sig_attrs;
2187 	int ret;
2188 
2189 	trace_mr_dereg(mr);
2190 	rdma_restrack_del(&mr->res);
2191 	ret = mr->device->ops.dereg_mr(mr, udata);
2192 	if (!ret) {
2193 		atomic_dec(&pd->usecnt);
2194 		if (dm)
2195 			atomic_dec(&dm->usecnt);
2196 		kfree(sig_attrs);
2197 	}
2198 
2199 	return ret;
2200 }
2201 EXPORT_SYMBOL(ib_dereg_mr_user);
2202 
2203 /**
2204  * ib_alloc_mr() - Allocates a memory region
2205  * @pd:            protection domain associated with the region
2206  * @mr_type:       memory region type
2207  * @max_num_sg:    maximum sg entries available for registration.
2208  *
2209  * Notes:
2210  * Memory registeration page/sg lists must not exceed max_num_sg.
2211  * For mr_type IB_MR_TYPE_MEM_REG, the total length cannot exceed
2212  * max_num_sg * used_page_size.
2213  *
2214  */
2215 struct ib_mr *ib_alloc_mr(struct ib_pd *pd, enum ib_mr_type mr_type,
2216 			  u32 max_num_sg)
2217 {
2218 	struct ib_mr *mr;
2219 
2220 	if (!pd->device->ops.alloc_mr) {
2221 		mr = ERR_PTR(-EOPNOTSUPP);
2222 		goto out;
2223 	}
2224 
2225 	if (mr_type == IB_MR_TYPE_INTEGRITY) {
2226 		WARN_ON_ONCE(1);
2227 		mr = ERR_PTR(-EINVAL);
2228 		goto out;
2229 	}
2230 
2231 	mr = pd->device->ops.alloc_mr(pd, mr_type, max_num_sg);
2232 	if (IS_ERR(mr))
2233 		goto out;
2234 
2235 	mr->device = pd->device;
2236 	mr->pd = pd;
2237 	mr->dm = NULL;
2238 	mr->uobject = NULL;
2239 	atomic_inc(&pd->usecnt);
2240 	mr->need_inval = false;
2241 	mr->type = mr_type;
2242 	mr->sig_attrs = NULL;
2243 
2244 	rdma_restrack_new(&mr->res, RDMA_RESTRACK_MR);
2245 	rdma_restrack_parent_name(&mr->res, &pd->res);
2246 	rdma_restrack_add(&mr->res);
2247 out:
2248 	trace_mr_alloc(pd, mr_type, max_num_sg, mr);
2249 	return mr;
2250 }
2251 EXPORT_SYMBOL(ib_alloc_mr);
2252 
2253 /**
2254  * ib_alloc_mr_integrity() - Allocates an integrity memory region
2255  * @pd:                      protection domain associated with the region
2256  * @max_num_data_sg:         maximum data sg entries available for registration
2257  * @max_num_meta_sg:         maximum metadata sg entries available for
2258  *                           registration
2259  *
2260  * Notes:
2261  * Memory registration page/sg lists must not exceed max_num_sg,
2262  * also the integrity page/sg lists must not exceed max_num_meta_sg.
2263  *
2264  */
2265 struct ib_mr *ib_alloc_mr_integrity(struct ib_pd *pd,
2266 				    u32 max_num_data_sg,
2267 				    u32 max_num_meta_sg)
2268 {
2269 	struct ib_mr *mr;
2270 	struct ib_sig_attrs *sig_attrs;
2271 
2272 	if (!pd->device->ops.alloc_mr_integrity ||
2273 	    !pd->device->ops.map_mr_sg_pi) {
2274 		mr = ERR_PTR(-EOPNOTSUPP);
2275 		goto out;
2276 	}
2277 
2278 	if (!max_num_meta_sg) {
2279 		mr = ERR_PTR(-EINVAL);
2280 		goto out;
2281 	}
2282 
2283 	sig_attrs = kzalloc(sizeof(struct ib_sig_attrs), GFP_KERNEL);
2284 	if (!sig_attrs) {
2285 		mr = ERR_PTR(-ENOMEM);
2286 		goto out;
2287 	}
2288 
2289 	mr = pd->device->ops.alloc_mr_integrity(pd, max_num_data_sg,
2290 						max_num_meta_sg);
2291 	if (IS_ERR(mr)) {
2292 		kfree(sig_attrs);
2293 		goto out;
2294 	}
2295 
2296 	mr->device = pd->device;
2297 	mr->pd = pd;
2298 	mr->dm = NULL;
2299 	mr->uobject = NULL;
2300 	atomic_inc(&pd->usecnt);
2301 	mr->need_inval = false;
2302 	mr->type = IB_MR_TYPE_INTEGRITY;
2303 	mr->sig_attrs = sig_attrs;
2304 
2305 	rdma_restrack_new(&mr->res, RDMA_RESTRACK_MR);
2306 	rdma_restrack_parent_name(&mr->res, &pd->res);
2307 	rdma_restrack_add(&mr->res);
2308 out:
2309 	trace_mr_integ_alloc(pd, max_num_data_sg, max_num_meta_sg, mr);
2310 	return mr;
2311 }
2312 EXPORT_SYMBOL(ib_alloc_mr_integrity);
2313 
2314 /* Multicast groups */
2315 
2316 static bool is_valid_mcast_lid(struct ib_qp *qp, u16 lid)
2317 {
2318 	struct ib_qp_init_attr init_attr = {};
2319 	struct ib_qp_attr attr = {};
2320 	int num_eth_ports = 0;
2321 	unsigned int port;
2322 
2323 	/* If QP state >= init, it is assigned to a port and we can check this
2324 	 * port only.
2325 	 */
2326 	if (!ib_query_qp(qp, &attr, IB_QP_STATE | IB_QP_PORT, &init_attr)) {
2327 		if (attr.qp_state >= IB_QPS_INIT) {
2328 			if (rdma_port_get_link_layer(qp->device, attr.port_num) !=
2329 			    IB_LINK_LAYER_INFINIBAND)
2330 				return true;
2331 			goto lid_check;
2332 		}
2333 	}
2334 
2335 	/* Can't get a quick answer, iterate over all ports */
2336 	rdma_for_each_port(qp->device, port)
2337 		if (rdma_port_get_link_layer(qp->device, port) !=
2338 		    IB_LINK_LAYER_INFINIBAND)
2339 			num_eth_ports++;
2340 
2341 	/* If we have at lease one Ethernet port, RoCE annex declares that
2342 	 * multicast LID should be ignored. We can't tell at this step if the
2343 	 * QP belongs to an IB or Ethernet port.
2344 	 */
2345 	if (num_eth_ports)
2346 		return true;
2347 
2348 	/* If all the ports are IB, we can check according to IB spec. */
2349 lid_check:
2350 	return !(lid < be16_to_cpu(IB_MULTICAST_LID_BASE) ||
2351 		 lid == be16_to_cpu(IB_LID_PERMISSIVE));
2352 }
2353 
2354 int ib_attach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid)
2355 {
2356 	int ret;
2357 
2358 	if (!qp->device->ops.attach_mcast)
2359 		return -EOPNOTSUPP;
2360 
2361 	if (!rdma_is_multicast_addr((struct in6_addr *)gid->raw) ||
2362 	    qp->qp_type != IB_QPT_UD || !is_valid_mcast_lid(qp, lid))
2363 		return -EINVAL;
2364 
2365 	ret = qp->device->ops.attach_mcast(qp, gid, lid);
2366 	if (!ret)
2367 		atomic_inc(&qp->usecnt);
2368 	return ret;
2369 }
2370 EXPORT_SYMBOL(ib_attach_mcast);
2371 
2372 int ib_detach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid)
2373 {
2374 	int ret;
2375 
2376 	if (!qp->device->ops.detach_mcast)
2377 		return -EOPNOTSUPP;
2378 
2379 	if (!rdma_is_multicast_addr((struct in6_addr *)gid->raw) ||
2380 	    qp->qp_type != IB_QPT_UD || !is_valid_mcast_lid(qp, lid))
2381 		return -EINVAL;
2382 
2383 	ret = qp->device->ops.detach_mcast(qp, gid, lid);
2384 	if (!ret)
2385 		atomic_dec(&qp->usecnt);
2386 	return ret;
2387 }
2388 EXPORT_SYMBOL(ib_detach_mcast);
2389 
2390 /**
2391  * ib_alloc_xrcd_user - Allocates an XRC domain.
2392  * @device: The device on which to allocate the XRC domain.
2393  * @inode: inode to connect XRCD
2394  * @udata: Valid user data or NULL for kernel object
2395  */
2396 struct ib_xrcd *ib_alloc_xrcd_user(struct ib_device *device,
2397 				   struct inode *inode, struct ib_udata *udata)
2398 {
2399 	struct ib_xrcd *xrcd;
2400 	int ret;
2401 
2402 	if (!device->ops.alloc_xrcd)
2403 		return ERR_PTR(-EOPNOTSUPP);
2404 
2405 	xrcd = rdma_zalloc_drv_obj(device, ib_xrcd);
2406 	if (!xrcd)
2407 		return ERR_PTR(-ENOMEM);
2408 
2409 	xrcd->device = device;
2410 	xrcd->inode = inode;
2411 	atomic_set(&xrcd->usecnt, 0);
2412 	init_rwsem(&xrcd->tgt_qps_rwsem);
2413 	xa_init(&xrcd->tgt_qps);
2414 
2415 	ret = device->ops.alloc_xrcd(xrcd, udata);
2416 	if (ret)
2417 		goto err;
2418 	return xrcd;
2419 err:
2420 	kfree(xrcd);
2421 	return ERR_PTR(ret);
2422 }
2423 EXPORT_SYMBOL(ib_alloc_xrcd_user);
2424 
2425 /**
2426  * ib_dealloc_xrcd_user - Deallocates an XRC domain.
2427  * @xrcd: The XRC domain to deallocate.
2428  * @udata: Valid user data or NULL for kernel object
2429  */
2430 int ib_dealloc_xrcd_user(struct ib_xrcd *xrcd, struct ib_udata *udata)
2431 {
2432 	int ret;
2433 
2434 	if (atomic_read(&xrcd->usecnt))
2435 		return -EBUSY;
2436 
2437 	WARN_ON(!xa_empty(&xrcd->tgt_qps));
2438 	ret = xrcd->device->ops.dealloc_xrcd(xrcd, udata);
2439 	if (ret)
2440 		return ret;
2441 	kfree(xrcd);
2442 	return ret;
2443 }
2444 EXPORT_SYMBOL(ib_dealloc_xrcd_user);
2445 
2446 /**
2447  * ib_create_wq - Creates a WQ associated with the specified protection
2448  * domain.
2449  * @pd: The protection domain associated with the WQ.
2450  * @wq_attr: A list of initial attributes required to create the
2451  * WQ. If WQ creation succeeds, then the attributes are updated to
2452  * the actual capabilities of the created WQ.
2453  *
2454  * wq_attr->max_wr and wq_attr->max_sge determine
2455  * the requested size of the WQ, and set to the actual values allocated
2456  * on return.
2457  * If ib_create_wq() succeeds, then max_wr and max_sge will always be
2458  * at least as large as the requested values.
2459  */
2460 struct ib_wq *ib_create_wq(struct ib_pd *pd,
2461 			   struct ib_wq_init_attr *wq_attr)
2462 {
2463 	struct ib_wq *wq;
2464 
2465 	if (!pd->device->ops.create_wq)
2466 		return ERR_PTR(-EOPNOTSUPP);
2467 
2468 	wq = pd->device->ops.create_wq(pd, wq_attr, NULL);
2469 	if (!IS_ERR(wq)) {
2470 		wq->event_handler = wq_attr->event_handler;
2471 		wq->wq_context = wq_attr->wq_context;
2472 		wq->wq_type = wq_attr->wq_type;
2473 		wq->cq = wq_attr->cq;
2474 		wq->device = pd->device;
2475 		wq->pd = pd;
2476 		wq->uobject = NULL;
2477 		atomic_inc(&pd->usecnt);
2478 		atomic_inc(&wq_attr->cq->usecnt);
2479 		atomic_set(&wq->usecnt, 0);
2480 	}
2481 	return wq;
2482 }
2483 EXPORT_SYMBOL(ib_create_wq);
2484 
2485 /**
2486  * ib_destroy_wq_user - Destroys the specified user WQ.
2487  * @wq: The WQ to destroy.
2488  * @udata: Valid user data
2489  */
2490 int ib_destroy_wq_user(struct ib_wq *wq, struct ib_udata *udata)
2491 {
2492 	struct ib_cq *cq = wq->cq;
2493 	struct ib_pd *pd = wq->pd;
2494 	int ret;
2495 
2496 	if (atomic_read(&wq->usecnt))
2497 		return -EBUSY;
2498 
2499 	ret = wq->device->ops.destroy_wq(wq, udata);
2500 	if (ret)
2501 		return ret;
2502 
2503 	atomic_dec(&pd->usecnt);
2504 	atomic_dec(&cq->usecnt);
2505 	return ret;
2506 }
2507 EXPORT_SYMBOL(ib_destroy_wq_user);
2508 
2509 int ib_check_mr_status(struct ib_mr *mr, u32 check_mask,
2510 		       struct ib_mr_status *mr_status)
2511 {
2512 	if (!mr->device->ops.check_mr_status)
2513 		return -EOPNOTSUPP;
2514 
2515 	return mr->device->ops.check_mr_status(mr, check_mask, mr_status);
2516 }
2517 EXPORT_SYMBOL(ib_check_mr_status);
2518 
2519 int ib_set_vf_link_state(struct ib_device *device, int vf, u32 port,
2520 			 int state)
2521 {
2522 	if (!device->ops.set_vf_link_state)
2523 		return -EOPNOTSUPP;
2524 
2525 	return device->ops.set_vf_link_state(device, vf, port, state);
2526 }
2527 EXPORT_SYMBOL(ib_set_vf_link_state);
2528 
2529 int ib_get_vf_config(struct ib_device *device, int vf, u32 port,
2530 		     struct ifla_vf_info *info)
2531 {
2532 	if (!device->ops.get_vf_config)
2533 		return -EOPNOTSUPP;
2534 
2535 	return device->ops.get_vf_config(device, vf, port, info);
2536 }
2537 EXPORT_SYMBOL(ib_get_vf_config);
2538 
2539 int ib_get_vf_stats(struct ib_device *device, int vf, u32 port,
2540 		    struct ifla_vf_stats *stats)
2541 {
2542 	if (!device->ops.get_vf_stats)
2543 		return -EOPNOTSUPP;
2544 
2545 	return device->ops.get_vf_stats(device, vf, port, stats);
2546 }
2547 EXPORT_SYMBOL(ib_get_vf_stats);
2548 
2549 int ib_set_vf_guid(struct ib_device *device, int vf, u32 port, u64 guid,
2550 		   int type)
2551 {
2552 	if (!device->ops.set_vf_guid)
2553 		return -EOPNOTSUPP;
2554 
2555 	return device->ops.set_vf_guid(device, vf, port, guid, type);
2556 }
2557 EXPORT_SYMBOL(ib_set_vf_guid);
2558 
2559 int ib_get_vf_guid(struct ib_device *device, int vf, u32 port,
2560 		   struct ifla_vf_guid *node_guid,
2561 		   struct ifla_vf_guid *port_guid)
2562 {
2563 	if (!device->ops.get_vf_guid)
2564 		return -EOPNOTSUPP;
2565 
2566 	return device->ops.get_vf_guid(device, vf, port, node_guid, port_guid);
2567 }
2568 EXPORT_SYMBOL(ib_get_vf_guid);
2569 /**
2570  * ib_map_mr_sg_pi() - Map the dma mapped SG lists for PI (protection
2571  *     information) and set an appropriate memory region for registration.
2572  * @mr:             memory region
2573  * @data_sg:        dma mapped scatterlist for data
2574  * @data_sg_nents:  number of entries in data_sg
2575  * @data_sg_offset: offset in bytes into data_sg
2576  * @meta_sg:        dma mapped scatterlist for metadata
2577  * @meta_sg_nents:  number of entries in meta_sg
2578  * @meta_sg_offset: offset in bytes into meta_sg
2579  * @page_size:      page vector desired page size
2580  *
2581  * Constraints:
2582  * - The MR must be allocated with type IB_MR_TYPE_INTEGRITY.
2583  *
2584  * Return: 0 on success.
2585  *
2586  * After this completes successfully, the  memory region
2587  * is ready for registration.
2588  */
2589 int ib_map_mr_sg_pi(struct ib_mr *mr, struct scatterlist *data_sg,
2590 		    int data_sg_nents, unsigned int *data_sg_offset,
2591 		    struct scatterlist *meta_sg, int meta_sg_nents,
2592 		    unsigned int *meta_sg_offset, unsigned int page_size)
2593 {
2594 	if (unlikely(!mr->device->ops.map_mr_sg_pi ||
2595 		     WARN_ON_ONCE(mr->type != IB_MR_TYPE_INTEGRITY)))
2596 		return -EOPNOTSUPP;
2597 
2598 	mr->page_size = page_size;
2599 
2600 	return mr->device->ops.map_mr_sg_pi(mr, data_sg, data_sg_nents,
2601 					    data_sg_offset, meta_sg,
2602 					    meta_sg_nents, meta_sg_offset);
2603 }
2604 EXPORT_SYMBOL(ib_map_mr_sg_pi);
2605 
2606 /**
2607  * ib_map_mr_sg() - Map the largest prefix of a dma mapped SG list
2608  *     and set it the memory region.
2609  * @mr:            memory region
2610  * @sg:            dma mapped scatterlist
2611  * @sg_nents:      number of entries in sg
2612  * @sg_offset:     offset in bytes into sg
2613  * @page_size:     page vector desired page size
2614  *
2615  * Constraints:
2616  *
2617  * - The first sg element is allowed to have an offset.
2618  * - Each sg element must either be aligned to page_size or virtually
2619  *   contiguous to the previous element. In case an sg element has a
2620  *   non-contiguous offset, the mapping prefix will not include it.
2621  * - The last sg element is allowed to have length less than page_size.
2622  * - If sg_nents total byte length exceeds the mr max_num_sge * page_size
2623  *   then only max_num_sg entries will be mapped.
2624  * - If the MR was allocated with type IB_MR_TYPE_SG_GAPS, none of these
2625  *   constraints holds and the page_size argument is ignored.
2626  *
2627  * Returns the number of sg elements that were mapped to the memory region.
2628  *
2629  * After this completes successfully, the  memory region
2630  * is ready for registration.
2631  */
2632 int ib_map_mr_sg(struct ib_mr *mr, struct scatterlist *sg, int sg_nents,
2633 		 unsigned int *sg_offset, unsigned int page_size)
2634 {
2635 	if (unlikely(!mr->device->ops.map_mr_sg))
2636 		return -EOPNOTSUPP;
2637 
2638 	mr->page_size = page_size;
2639 
2640 	return mr->device->ops.map_mr_sg(mr, sg, sg_nents, sg_offset);
2641 }
2642 EXPORT_SYMBOL(ib_map_mr_sg);
2643 
2644 /**
2645  * ib_sg_to_pages() - Convert the largest prefix of a sg list
2646  *     to a page vector
2647  * @mr:            memory region
2648  * @sgl:           dma mapped scatterlist
2649  * @sg_nents:      number of entries in sg
2650  * @sg_offset_p:   ==== =======================================================
2651  *                 IN   start offset in bytes into sg
2652  *                 OUT  offset in bytes for element n of the sg of the first
2653  *                      byte that has not been processed where n is the return
2654  *                      value of this function.
2655  *                 ==== =======================================================
2656  * @set_page:      driver page assignment function pointer
2657  *
2658  * Core service helper for drivers to convert the largest
2659  * prefix of given sg list to a page vector. The sg list
2660  * prefix converted is the prefix that meet the requirements
2661  * of ib_map_mr_sg.
2662  *
2663  * Returns the number of sg elements that were assigned to
2664  * a page vector.
2665  */
2666 int ib_sg_to_pages(struct ib_mr *mr, struct scatterlist *sgl, int sg_nents,
2667 		unsigned int *sg_offset_p, int (*set_page)(struct ib_mr *, u64))
2668 {
2669 	struct scatterlist *sg;
2670 	u64 last_end_dma_addr = 0;
2671 	unsigned int sg_offset = sg_offset_p ? *sg_offset_p : 0;
2672 	unsigned int last_page_off = 0;
2673 	u64 page_mask = ~((u64)mr->page_size - 1);
2674 	int i, ret;
2675 
2676 	if (unlikely(sg_nents <= 0 || sg_offset > sg_dma_len(&sgl[0])))
2677 		return -EINVAL;
2678 
2679 	mr->iova = sg_dma_address(&sgl[0]) + sg_offset;
2680 	mr->length = 0;
2681 
2682 	for_each_sg(sgl, sg, sg_nents, i) {
2683 		u64 dma_addr = sg_dma_address(sg) + sg_offset;
2684 		u64 prev_addr = dma_addr;
2685 		unsigned int dma_len = sg_dma_len(sg) - sg_offset;
2686 		u64 end_dma_addr = dma_addr + dma_len;
2687 		u64 page_addr = dma_addr & page_mask;
2688 
2689 		/*
2690 		 * For the second and later elements, check whether either the
2691 		 * end of element i-1 or the start of element i is not aligned
2692 		 * on a page boundary.
2693 		 */
2694 		if (i && (last_page_off != 0 || page_addr != dma_addr)) {
2695 			/* Stop mapping if there is a gap. */
2696 			if (last_end_dma_addr != dma_addr)
2697 				break;
2698 
2699 			/*
2700 			 * Coalesce this element with the last. If it is small
2701 			 * enough just update mr->length. Otherwise start
2702 			 * mapping from the next page.
2703 			 */
2704 			goto next_page;
2705 		}
2706 
2707 		do {
2708 			ret = set_page(mr, page_addr);
2709 			if (unlikely(ret < 0)) {
2710 				sg_offset = prev_addr - sg_dma_address(sg);
2711 				mr->length += prev_addr - dma_addr;
2712 				if (sg_offset_p)
2713 					*sg_offset_p = sg_offset;
2714 				return i || sg_offset ? i : ret;
2715 			}
2716 			prev_addr = page_addr;
2717 next_page:
2718 			page_addr += mr->page_size;
2719 		} while (page_addr < end_dma_addr);
2720 
2721 		mr->length += dma_len;
2722 		last_end_dma_addr = end_dma_addr;
2723 		last_page_off = end_dma_addr & ~page_mask;
2724 
2725 		sg_offset = 0;
2726 	}
2727 
2728 	if (sg_offset_p)
2729 		*sg_offset_p = 0;
2730 	return i;
2731 }
2732 EXPORT_SYMBOL(ib_sg_to_pages);
2733 
2734 struct ib_drain_cqe {
2735 	struct ib_cqe cqe;
2736 	struct completion done;
2737 };
2738 
2739 static void ib_drain_qp_done(struct ib_cq *cq, struct ib_wc *wc)
2740 {
2741 	struct ib_drain_cqe *cqe = container_of(wc->wr_cqe, struct ib_drain_cqe,
2742 						cqe);
2743 
2744 	complete(&cqe->done);
2745 }
2746 
2747 /*
2748  * Post a WR and block until its completion is reaped for the SQ.
2749  */
2750 static void __ib_drain_sq(struct ib_qp *qp)
2751 {
2752 	struct ib_cq *cq = qp->send_cq;
2753 	struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
2754 	struct ib_drain_cqe sdrain;
2755 	struct ib_rdma_wr swr = {
2756 		.wr = {
2757 			.next = NULL,
2758 			{ .wr_cqe	= &sdrain.cqe, },
2759 			.opcode	= IB_WR_RDMA_WRITE,
2760 		},
2761 	};
2762 	int ret;
2763 
2764 	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
2765 	if (ret) {
2766 		WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
2767 		return;
2768 	}
2769 
2770 	sdrain.cqe.done = ib_drain_qp_done;
2771 	init_completion(&sdrain.done);
2772 
2773 	ret = ib_post_send(qp, &swr.wr, NULL);
2774 	if (ret) {
2775 		WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
2776 		return;
2777 	}
2778 
2779 	if (cq->poll_ctx == IB_POLL_DIRECT)
2780 		while (wait_for_completion_timeout(&sdrain.done, HZ / 10) <= 0)
2781 			ib_process_cq_direct(cq, -1);
2782 	else
2783 		wait_for_completion(&sdrain.done);
2784 }
2785 
2786 /*
2787  * Post a WR and block until its completion is reaped for the RQ.
2788  */
2789 static void __ib_drain_rq(struct ib_qp *qp)
2790 {
2791 	struct ib_cq *cq = qp->recv_cq;
2792 	struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
2793 	struct ib_drain_cqe rdrain;
2794 	struct ib_recv_wr rwr = {};
2795 	int ret;
2796 
2797 	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
2798 	if (ret) {
2799 		WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
2800 		return;
2801 	}
2802 
2803 	rwr.wr_cqe = &rdrain.cqe;
2804 	rdrain.cqe.done = ib_drain_qp_done;
2805 	init_completion(&rdrain.done);
2806 
2807 	ret = ib_post_recv(qp, &rwr, NULL);
2808 	if (ret) {
2809 		WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
2810 		return;
2811 	}
2812 
2813 	if (cq->poll_ctx == IB_POLL_DIRECT)
2814 		while (wait_for_completion_timeout(&rdrain.done, HZ / 10) <= 0)
2815 			ib_process_cq_direct(cq, -1);
2816 	else
2817 		wait_for_completion(&rdrain.done);
2818 }
2819 
2820 /**
2821  * ib_drain_sq() - Block until all SQ CQEs have been consumed by the
2822  *		   application.
2823  * @qp:            queue pair to drain
2824  *
2825  * If the device has a provider-specific drain function, then
2826  * call that.  Otherwise call the generic drain function
2827  * __ib_drain_sq().
2828  *
2829  * The caller must:
2830  *
2831  * ensure there is room in the CQ and SQ for the drain work request and
2832  * completion.
2833  *
2834  * allocate the CQ using ib_alloc_cq().
2835  *
2836  * ensure that there are no other contexts that are posting WRs concurrently.
2837  * Otherwise the drain is not guaranteed.
2838  */
2839 void ib_drain_sq(struct ib_qp *qp)
2840 {
2841 	if (qp->device->ops.drain_sq)
2842 		qp->device->ops.drain_sq(qp);
2843 	else
2844 		__ib_drain_sq(qp);
2845 	trace_cq_drain_complete(qp->send_cq);
2846 }
2847 EXPORT_SYMBOL(ib_drain_sq);
2848 
2849 /**
2850  * ib_drain_rq() - Block until all RQ CQEs have been consumed by the
2851  *		   application.
2852  * @qp:            queue pair to drain
2853  *
2854  * If the device has a provider-specific drain function, then
2855  * call that.  Otherwise call the generic drain function
2856  * __ib_drain_rq().
2857  *
2858  * The caller must:
2859  *
2860  * ensure there is room in the CQ and RQ for the drain work request and
2861  * completion.
2862  *
2863  * allocate the CQ using ib_alloc_cq().
2864  *
2865  * ensure that there are no other contexts that are posting WRs concurrently.
2866  * Otherwise the drain is not guaranteed.
2867  */
2868 void ib_drain_rq(struct ib_qp *qp)
2869 {
2870 	if (qp->device->ops.drain_rq)
2871 		qp->device->ops.drain_rq(qp);
2872 	else
2873 		__ib_drain_rq(qp);
2874 	trace_cq_drain_complete(qp->recv_cq);
2875 }
2876 EXPORT_SYMBOL(ib_drain_rq);
2877 
2878 /**
2879  * ib_drain_qp() - Block until all CQEs have been consumed by the
2880  *		   application on both the RQ and SQ.
2881  * @qp:            queue pair to drain
2882  *
2883  * The caller must:
2884  *
2885  * ensure there is room in the CQ(s), SQ, and RQ for drain work requests
2886  * and completions.
2887  *
2888  * allocate the CQs using ib_alloc_cq().
2889  *
2890  * ensure that there are no other contexts that are posting WRs concurrently.
2891  * Otherwise the drain is not guaranteed.
2892  */
2893 void ib_drain_qp(struct ib_qp *qp)
2894 {
2895 	ib_drain_sq(qp);
2896 	if (!qp->srq)
2897 		ib_drain_rq(qp);
2898 }
2899 EXPORT_SYMBOL(ib_drain_qp);
2900 
2901 struct net_device *rdma_alloc_netdev(struct ib_device *device, u32 port_num,
2902 				     enum rdma_netdev_t type, const char *name,
2903 				     unsigned char name_assign_type,
2904 				     void (*setup)(struct net_device *))
2905 {
2906 	struct rdma_netdev_alloc_params params;
2907 	struct net_device *netdev;
2908 	int rc;
2909 
2910 	if (!device->ops.rdma_netdev_get_params)
2911 		return ERR_PTR(-EOPNOTSUPP);
2912 
2913 	rc = device->ops.rdma_netdev_get_params(device, port_num, type,
2914 						&params);
2915 	if (rc)
2916 		return ERR_PTR(rc);
2917 
2918 	netdev = alloc_netdev_mqs(params.sizeof_priv, name, name_assign_type,
2919 				  setup, params.txqs, params.rxqs);
2920 	if (!netdev)
2921 		return ERR_PTR(-ENOMEM);
2922 
2923 	return netdev;
2924 }
2925 EXPORT_SYMBOL(rdma_alloc_netdev);
2926 
2927 int rdma_init_netdev(struct ib_device *device, u32 port_num,
2928 		     enum rdma_netdev_t type, const char *name,
2929 		     unsigned char name_assign_type,
2930 		     void (*setup)(struct net_device *),
2931 		     struct net_device *netdev)
2932 {
2933 	struct rdma_netdev_alloc_params params;
2934 	int rc;
2935 
2936 	if (!device->ops.rdma_netdev_get_params)
2937 		return -EOPNOTSUPP;
2938 
2939 	rc = device->ops.rdma_netdev_get_params(device, port_num, type,
2940 						&params);
2941 	if (rc)
2942 		return rc;
2943 
2944 	return params.initialize_rdma_netdev(device, port_num,
2945 					     netdev, params.param);
2946 }
2947 EXPORT_SYMBOL(rdma_init_netdev);
2948 
2949 void __rdma_block_iter_start(struct ib_block_iter *biter,
2950 			     struct scatterlist *sglist, unsigned int nents,
2951 			     unsigned long pgsz)
2952 {
2953 	memset(biter, 0, sizeof(struct ib_block_iter));
2954 	biter->__sg = sglist;
2955 	biter->__sg_nents = nents;
2956 
2957 	/* Driver provides best block size to use */
2958 	biter->__pg_bit = __fls(pgsz);
2959 }
2960 EXPORT_SYMBOL(__rdma_block_iter_start);
2961 
2962 bool __rdma_block_iter_next(struct ib_block_iter *biter)
2963 {
2964 	unsigned int block_offset;
2965 
2966 	if (!biter->__sg_nents || !biter->__sg)
2967 		return false;
2968 
2969 	biter->__dma_addr = sg_dma_address(biter->__sg) + biter->__sg_advance;
2970 	block_offset = biter->__dma_addr & (BIT_ULL(biter->__pg_bit) - 1);
2971 	biter->__sg_advance += BIT_ULL(biter->__pg_bit) - block_offset;
2972 
2973 	if (biter->__sg_advance >= sg_dma_len(biter->__sg)) {
2974 		biter->__sg_advance = 0;
2975 		biter->__sg = sg_next(biter->__sg);
2976 		biter->__sg_nents--;
2977 	}
2978 
2979 	return true;
2980 }
2981 EXPORT_SYMBOL(__rdma_block_iter_next);
2982 
2983 /**
2984  * rdma_alloc_hw_stats_struct - Helper function to allocate dynamic struct
2985  *   for the drivers.
2986  * @descs: array of static descriptors
2987  * @num_counters: number of elements in array
2988  * @lifespan: milliseconds between updates
2989  */
2990 struct rdma_hw_stats *rdma_alloc_hw_stats_struct(
2991 	const struct rdma_stat_desc *descs, int num_counters,
2992 	unsigned long lifespan)
2993 {
2994 	struct rdma_hw_stats *stats;
2995 
2996 	stats = kzalloc(struct_size(stats, value, num_counters), GFP_KERNEL);
2997 	if (!stats)
2998 		return NULL;
2999 
3000 	stats->is_disabled = kcalloc(BITS_TO_LONGS(num_counters),
3001 				     sizeof(*stats->is_disabled), GFP_KERNEL);
3002 	if (!stats->is_disabled)
3003 		goto err;
3004 
3005 	stats->descs = descs;
3006 	stats->num_counters = num_counters;
3007 	stats->lifespan = msecs_to_jiffies(lifespan);
3008 	mutex_init(&stats->lock);
3009 
3010 	return stats;
3011 
3012 err:
3013 	kfree(stats);
3014 	return NULL;
3015 }
3016 EXPORT_SYMBOL(rdma_alloc_hw_stats_struct);
3017 
3018 /**
3019  * rdma_free_hw_stats_struct - Helper function to release rdma_hw_stats
3020  * @stats: statistics to release
3021  */
3022 void rdma_free_hw_stats_struct(struct rdma_hw_stats *stats)
3023 {
3024 	if (!stats)
3025 		return;
3026 
3027 	kfree(stats->is_disabled);
3028 	kfree(stats);
3029 }
3030 EXPORT_SYMBOL(rdma_free_hw_stats_struct);
3031