xref: /linux/drivers/infiniband/core/verbs.c (revision a36e9f5cfe9eb3a1dce8769c7058251c42705357)
1 /*
2  * Copyright (c) 2004 Mellanox Technologies Ltd.  All rights reserved.
3  * Copyright (c) 2004 Infinicon Corporation.  All rights reserved.
4  * Copyright (c) 2004 Intel Corporation.  All rights reserved.
5  * Copyright (c) 2004 Topspin Corporation.  All rights reserved.
6  * Copyright (c) 2004 Voltaire Corporation.  All rights reserved.
7  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
8  * Copyright (c) 2005, 2006 Cisco Systems.  All rights reserved.
9  *
10  * This software is available to you under a choice of one of two
11  * licenses.  You may choose to be licensed under the terms of the GNU
12  * General Public License (GPL) Version 2, available from the file
13  * COPYING in the main directory of this source tree, or the
14  * OpenIB.org BSD license below:
15  *
16  *     Redistribution and use in source and binary forms, with or
17  *     without modification, are permitted provided that the following
18  *     conditions are met:
19  *
20  *      - Redistributions of source code must retain the above
21  *        copyright notice, this list of conditions and the following
22  *        disclaimer.
23  *
24  *      - Redistributions in binary form must reproduce the above
25  *        copyright notice, this list of conditions and the following
26  *        disclaimer in the documentation and/or other materials
27  *        provided with the distribution.
28  *
29  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
30  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
31  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
32  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
33  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
34  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
35  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
36  * SOFTWARE.
37  */
38 
39 #include <linux/errno.h>
40 #include <linux/err.h>
41 #include <linux/export.h>
42 #include <linux/string.h>
43 #include <linux/slab.h>
44 #include <linux/in.h>
45 #include <linux/in6.h>
46 #include <net/addrconf.h>
47 #include <linux/security.h>
48 
49 #include <rdma/ib_verbs.h>
50 #include <rdma/ib_cache.h>
51 #include <rdma/ib_addr.h>
52 #include <rdma/rw.h>
53 #include <rdma/lag.h>
54 
55 #include "core_priv.h"
56 #include <trace/events/rdma_core.h>
57 
58 static int ib_resolve_eth_dmac(struct ib_device *device,
59 			       struct rdma_ah_attr *ah_attr);
60 
61 static const char * const ib_events[] = {
62 	[IB_EVENT_CQ_ERR]		= "CQ error",
63 	[IB_EVENT_QP_FATAL]		= "QP fatal error",
64 	[IB_EVENT_QP_REQ_ERR]		= "QP request error",
65 	[IB_EVENT_QP_ACCESS_ERR]	= "QP access error",
66 	[IB_EVENT_COMM_EST]		= "communication established",
67 	[IB_EVENT_SQ_DRAINED]		= "send queue drained",
68 	[IB_EVENT_PATH_MIG]		= "path migration successful",
69 	[IB_EVENT_PATH_MIG_ERR]		= "path migration error",
70 	[IB_EVENT_DEVICE_FATAL]		= "device fatal error",
71 	[IB_EVENT_PORT_ACTIVE]		= "port active",
72 	[IB_EVENT_PORT_ERR]		= "port error",
73 	[IB_EVENT_LID_CHANGE]		= "LID change",
74 	[IB_EVENT_PKEY_CHANGE]		= "P_key change",
75 	[IB_EVENT_SM_CHANGE]		= "SM change",
76 	[IB_EVENT_SRQ_ERR]		= "SRQ error",
77 	[IB_EVENT_SRQ_LIMIT_REACHED]	= "SRQ limit reached",
78 	[IB_EVENT_QP_LAST_WQE_REACHED]	= "last WQE reached",
79 	[IB_EVENT_CLIENT_REREGISTER]	= "client reregister",
80 	[IB_EVENT_GID_CHANGE]		= "GID changed",
81 };
82 
83 const char *__attribute_const__ ib_event_msg(enum ib_event_type event)
84 {
85 	size_t index = event;
86 
87 	return (index < ARRAY_SIZE(ib_events) && ib_events[index]) ?
88 			ib_events[index] : "unrecognized event";
89 }
90 EXPORT_SYMBOL(ib_event_msg);
91 
92 static const char * const wc_statuses[] = {
93 	[IB_WC_SUCCESS]			= "success",
94 	[IB_WC_LOC_LEN_ERR]		= "local length error",
95 	[IB_WC_LOC_QP_OP_ERR]		= "local QP operation error",
96 	[IB_WC_LOC_EEC_OP_ERR]		= "local EE context operation error",
97 	[IB_WC_LOC_PROT_ERR]		= "local protection error",
98 	[IB_WC_WR_FLUSH_ERR]		= "WR flushed",
99 	[IB_WC_MW_BIND_ERR]		= "memory bind operation error",
100 	[IB_WC_BAD_RESP_ERR]		= "bad response error",
101 	[IB_WC_LOC_ACCESS_ERR]		= "local access error",
102 	[IB_WC_REM_INV_REQ_ERR]		= "remote invalid request error",
103 	[IB_WC_REM_ACCESS_ERR]		= "remote access error",
104 	[IB_WC_REM_OP_ERR]		= "remote operation error",
105 	[IB_WC_RETRY_EXC_ERR]		= "transport retry counter exceeded",
106 	[IB_WC_RNR_RETRY_EXC_ERR]	= "RNR retry counter exceeded",
107 	[IB_WC_LOC_RDD_VIOL_ERR]	= "local RDD violation error",
108 	[IB_WC_REM_INV_RD_REQ_ERR]	= "remote invalid RD request",
109 	[IB_WC_REM_ABORT_ERR]		= "operation aborted",
110 	[IB_WC_INV_EECN_ERR]		= "invalid EE context number",
111 	[IB_WC_INV_EEC_STATE_ERR]	= "invalid EE context state",
112 	[IB_WC_FATAL_ERR]		= "fatal error",
113 	[IB_WC_RESP_TIMEOUT_ERR]	= "response timeout error",
114 	[IB_WC_GENERAL_ERR]		= "general error",
115 };
116 
117 const char *__attribute_const__ ib_wc_status_msg(enum ib_wc_status status)
118 {
119 	size_t index = status;
120 
121 	return (index < ARRAY_SIZE(wc_statuses) && wc_statuses[index]) ?
122 			wc_statuses[index] : "unrecognized status";
123 }
124 EXPORT_SYMBOL(ib_wc_status_msg);
125 
126 __attribute_const__ int ib_rate_to_mult(enum ib_rate rate)
127 {
128 	switch (rate) {
129 	case IB_RATE_2_5_GBPS: return   1;
130 	case IB_RATE_5_GBPS:   return   2;
131 	case IB_RATE_10_GBPS:  return   4;
132 	case IB_RATE_20_GBPS:  return   8;
133 	case IB_RATE_30_GBPS:  return  12;
134 	case IB_RATE_40_GBPS:  return  16;
135 	case IB_RATE_60_GBPS:  return  24;
136 	case IB_RATE_80_GBPS:  return  32;
137 	case IB_RATE_120_GBPS: return  48;
138 	case IB_RATE_14_GBPS:  return   6;
139 	case IB_RATE_56_GBPS:  return  22;
140 	case IB_RATE_112_GBPS: return  45;
141 	case IB_RATE_168_GBPS: return  67;
142 	case IB_RATE_25_GBPS:  return  10;
143 	case IB_RATE_100_GBPS: return  40;
144 	case IB_RATE_200_GBPS: return  80;
145 	case IB_RATE_300_GBPS: return 120;
146 	case IB_RATE_28_GBPS:  return  11;
147 	case IB_RATE_50_GBPS:  return  20;
148 	case IB_RATE_400_GBPS: return 160;
149 	case IB_RATE_600_GBPS: return 240;
150 	case IB_RATE_800_GBPS: return 320;
151 	default:	       return  -1;
152 	}
153 }
154 EXPORT_SYMBOL(ib_rate_to_mult);
155 
156 __attribute_const__ enum ib_rate mult_to_ib_rate(int mult)
157 {
158 	switch (mult) {
159 	case 1:   return IB_RATE_2_5_GBPS;
160 	case 2:   return IB_RATE_5_GBPS;
161 	case 4:   return IB_RATE_10_GBPS;
162 	case 8:   return IB_RATE_20_GBPS;
163 	case 12:  return IB_RATE_30_GBPS;
164 	case 16:  return IB_RATE_40_GBPS;
165 	case 24:  return IB_RATE_60_GBPS;
166 	case 32:  return IB_RATE_80_GBPS;
167 	case 48:  return IB_RATE_120_GBPS;
168 	case 6:   return IB_RATE_14_GBPS;
169 	case 22:  return IB_RATE_56_GBPS;
170 	case 45:  return IB_RATE_112_GBPS;
171 	case 67:  return IB_RATE_168_GBPS;
172 	case 10:  return IB_RATE_25_GBPS;
173 	case 40:  return IB_RATE_100_GBPS;
174 	case 80:  return IB_RATE_200_GBPS;
175 	case 120: return IB_RATE_300_GBPS;
176 	case 11:  return IB_RATE_28_GBPS;
177 	case 20:  return IB_RATE_50_GBPS;
178 	case 160: return IB_RATE_400_GBPS;
179 	case 240: return IB_RATE_600_GBPS;
180 	case 320: return IB_RATE_800_GBPS;
181 	default:  return IB_RATE_PORT_CURRENT;
182 	}
183 }
184 EXPORT_SYMBOL(mult_to_ib_rate);
185 
186 __attribute_const__ int ib_rate_to_mbps(enum ib_rate rate)
187 {
188 	switch (rate) {
189 	case IB_RATE_2_5_GBPS: return 2500;
190 	case IB_RATE_5_GBPS:   return 5000;
191 	case IB_RATE_10_GBPS:  return 10000;
192 	case IB_RATE_20_GBPS:  return 20000;
193 	case IB_RATE_30_GBPS:  return 30000;
194 	case IB_RATE_40_GBPS:  return 40000;
195 	case IB_RATE_60_GBPS:  return 60000;
196 	case IB_RATE_80_GBPS:  return 80000;
197 	case IB_RATE_120_GBPS: return 120000;
198 	case IB_RATE_14_GBPS:  return 14062;
199 	case IB_RATE_56_GBPS:  return 56250;
200 	case IB_RATE_112_GBPS: return 112500;
201 	case IB_RATE_168_GBPS: return 168750;
202 	case IB_RATE_25_GBPS:  return 25781;
203 	case IB_RATE_100_GBPS: return 103125;
204 	case IB_RATE_200_GBPS: return 206250;
205 	case IB_RATE_300_GBPS: return 309375;
206 	case IB_RATE_28_GBPS:  return 28125;
207 	case IB_RATE_50_GBPS:  return 53125;
208 	case IB_RATE_400_GBPS: return 425000;
209 	case IB_RATE_600_GBPS: return 637500;
210 	case IB_RATE_800_GBPS: return 850000;
211 	default:	       return -1;
212 	}
213 }
214 EXPORT_SYMBOL(ib_rate_to_mbps);
215 
216 __attribute_const__ enum rdma_transport_type
217 rdma_node_get_transport(unsigned int node_type)
218 {
219 
220 	if (node_type == RDMA_NODE_USNIC)
221 		return RDMA_TRANSPORT_USNIC;
222 	if (node_type == RDMA_NODE_USNIC_UDP)
223 		return RDMA_TRANSPORT_USNIC_UDP;
224 	if (node_type == RDMA_NODE_RNIC)
225 		return RDMA_TRANSPORT_IWARP;
226 	if (node_type == RDMA_NODE_UNSPECIFIED)
227 		return RDMA_TRANSPORT_UNSPECIFIED;
228 
229 	return RDMA_TRANSPORT_IB;
230 }
231 EXPORT_SYMBOL(rdma_node_get_transport);
232 
233 enum rdma_link_layer rdma_port_get_link_layer(struct ib_device *device,
234 					      u32 port_num)
235 {
236 	enum rdma_transport_type lt;
237 	if (device->ops.get_link_layer)
238 		return device->ops.get_link_layer(device, port_num);
239 
240 	lt = rdma_node_get_transport(device->node_type);
241 	if (lt == RDMA_TRANSPORT_IB)
242 		return IB_LINK_LAYER_INFINIBAND;
243 
244 	return IB_LINK_LAYER_ETHERNET;
245 }
246 EXPORT_SYMBOL(rdma_port_get_link_layer);
247 
248 /* Protection domains */
249 
250 /**
251  * __ib_alloc_pd - Allocates an unused protection domain.
252  * @device: The device on which to allocate the protection domain.
253  * @flags: protection domain flags
254  * @caller: caller's build-time module name
255  *
256  * A protection domain object provides an association between QPs, shared
257  * receive queues, address handles, memory regions, and memory windows.
258  *
259  * Every PD has a local_dma_lkey which can be used as the lkey value for local
260  * memory operations.
261  */
262 struct ib_pd *__ib_alloc_pd(struct ib_device *device, unsigned int flags,
263 		const char *caller)
264 {
265 	struct ib_pd *pd;
266 	int mr_access_flags = 0;
267 	int ret;
268 
269 	pd = rdma_zalloc_drv_obj(device, ib_pd);
270 	if (!pd)
271 		return ERR_PTR(-ENOMEM);
272 
273 	pd->device = device;
274 	pd->flags = flags;
275 
276 	rdma_restrack_new(&pd->res, RDMA_RESTRACK_PD);
277 	rdma_restrack_set_name(&pd->res, caller);
278 
279 	ret = device->ops.alloc_pd(pd, NULL);
280 	if (ret) {
281 		rdma_restrack_put(&pd->res);
282 		kfree(pd);
283 		return ERR_PTR(ret);
284 	}
285 	rdma_restrack_add(&pd->res);
286 
287 	if (device->attrs.kernel_cap_flags & IBK_LOCAL_DMA_LKEY)
288 		pd->local_dma_lkey = device->local_dma_lkey;
289 	else
290 		mr_access_flags |= IB_ACCESS_LOCAL_WRITE;
291 
292 	if (flags & IB_PD_UNSAFE_GLOBAL_RKEY) {
293 		pr_warn("%s: enabling unsafe global rkey\n", caller);
294 		mr_access_flags |= IB_ACCESS_REMOTE_READ | IB_ACCESS_REMOTE_WRITE;
295 	}
296 
297 	if (mr_access_flags) {
298 		struct ib_mr *mr;
299 
300 		mr = pd->device->ops.get_dma_mr(pd, mr_access_flags);
301 		if (IS_ERR(mr)) {
302 			ib_dealloc_pd(pd);
303 			return ERR_CAST(mr);
304 		}
305 
306 		mr->device	= pd->device;
307 		mr->pd		= pd;
308 		mr->type        = IB_MR_TYPE_DMA;
309 		mr->uobject	= NULL;
310 		mr->need_inval	= false;
311 
312 		pd->__internal_mr = mr;
313 
314 		if (!(device->attrs.kernel_cap_flags & IBK_LOCAL_DMA_LKEY))
315 			pd->local_dma_lkey = pd->__internal_mr->lkey;
316 
317 		if (flags & IB_PD_UNSAFE_GLOBAL_RKEY)
318 			pd->unsafe_global_rkey = pd->__internal_mr->rkey;
319 	}
320 
321 	return pd;
322 }
323 EXPORT_SYMBOL(__ib_alloc_pd);
324 
325 /**
326  * ib_dealloc_pd_user - Deallocates a protection domain.
327  * @pd: The protection domain to deallocate.
328  * @udata: Valid user data or NULL for kernel object
329  *
330  * It is an error to call this function while any resources in the pd still
331  * exist.  The caller is responsible to synchronously destroy them and
332  * guarantee no new allocations will happen.
333  */
334 int ib_dealloc_pd_user(struct ib_pd *pd, struct ib_udata *udata)
335 {
336 	int ret;
337 
338 	if (pd->__internal_mr) {
339 		ret = pd->device->ops.dereg_mr(pd->__internal_mr, NULL);
340 		WARN_ON(ret);
341 		pd->__internal_mr = NULL;
342 	}
343 
344 	ret = pd->device->ops.dealloc_pd(pd, udata);
345 	if (ret)
346 		return ret;
347 
348 	rdma_restrack_del(&pd->res);
349 	kfree(pd);
350 	return ret;
351 }
352 EXPORT_SYMBOL(ib_dealloc_pd_user);
353 
354 /* Address handles */
355 
356 /**
357  * rdma_copy_ah_attr - Copy rdma ah attribute from source to destination.
358  * @dest:       Pointer to destination ah_attr. Contents of the destination
359  *              pointer is assumed to be invalid and attribute are overwritten.
360  * @src:        Pointer to source ah_attr.
361  */
362 void rdma_copy_ah_attr(struct rdma_ah_attr *dest,
363 		       const struct rdma_ah_attr *src)
364 {
365 	*dest = *src;
366 	if (dest->grh.sgid_attr)
367 		rdma_hold_gid_attr(dest->grh.sgid_attr);
368 }
369 EXPORT_SYMBOL(rdma_copy_ah_attr);
370 
371 /**
372  * rdma_replace_ah_attr - Replace valid ah_attr with new one.
373  * @old:        Pointer to existing ah_attr which needs to be replaced.
374  *              old is assumed to be valid or zero'd
375  * @new:        Pointer to the new ah_attr.
376  *
377  * rdma_replace_ah_attr() first releases any reference in the old ah_attr if
378  * old the ah_attr is valid; after that it copies the new attribute and holds
379  * the reference to the replaced ah_attr.
380  */
381 void rdma_replace_ah_attr(struct rdma_ah_attr *old,
382 			  const struct rdma_ah_attr *new)
383 {
384 	rdma_destroy_ah_attr(old);
385 	*old = *new;
386 	if (old->grh.sgid_attr)
387 		rdma_hold_gid_attr(old->grh.sgid_attr);
388 }
389 EXPORT_SYMBOL(rdma_replace_ah_attr);
390 
391 /**
392  * rdma_move_ah_attr - Move ah_attr pointed by source to destination.
393  * @dest:       Pointer to destination ah_attr to copy to.
394  *              dest is assumed to be valid or zero'd
395  * @src:        Pointer to the new ah_attr.
396  *
397  * rdma_move_ah_attr() first releases any reference in the destination ah_attr
398  * if it is valid. This also transfers ownership of internal references from
399  * src to dest, making src invalid in the process. No new reference of the src
400  * ah_attr is taken.
401  */
402 void rdma_move_ah_attr(struct rdma_ah_attr *dest, struct rdma_ah_attr *src)
403 {
404 	rdma_destroy_ah_attr(dest);
405 	*dest = *src;
406 	src->grh.sgid_attr = NULL;
407 }
408 EXPORT_SYMBOL(rdma_move_ah_attr);
409 
410 /*
411  * Validate that the rdma_ah_attr is valid for the device before passing it
412  * off to the driver.
413  */
414 static int rdma_check_ah_attr(struct ib_device *device,
415 			      struct rdma_ah_attr *ah_attr)
416 {
417 	if (!rdma_is_port_valid(device, ah_attr->port_num))
418 		return -EINVAL;
419 
420 	if ((rdma_is_grh_required(device, ah_attr->port_num) ||
421 	     ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE) &&
422 	    !(ah_attr->ah_flags & IB_AH_GRH))
423 		return -EINVAL;
424 
425 	if (ah_attr->grh.sgid_attr) {
426 		/*
427 		 * Make sure the passed sgid_attr is consistent with the
428 		 * parameters
429 		 */
430 		if (ah_attr->grh.sgid_attr->index != ah_attr->grh.sgid_index ||
431 		    ah_attr->grh.sgid_attr->port_num != ah_attr->port_num)
432 			return -EINVAL;
433 	}
434 	return 0;
435 }
436 
437 /*
438  * If the ah requires a GRH then ensure that sgid_attr pointer is filled in.
439  * On success the caller is responsible to call rdma_unfill_sgid_attr().
440  */
441 static int rdma_fill_sgid_attr(struct ib_device *device,
442 			       struct rdma_ah_attr *ah_attr,
443 			       const struct ib_gid_attr **old_sgid_attr)
444 {
445 	const struct ib_gid_attr *sgid_attr;
446 	struct ib_global_route *grh;
447 	int ret;
448 
449 	*old_sgid_attr = ah_attr->grh.sgid_attr;
450 
451 	ret = rdma_check_ah_attr(device, ah_attr);
452 	if (ret)
453 		return ret;
454 
455 	if (!(ah_attr->ah_flags & IB_AH_GRH))
456 		return 0;
457 
458 	grh = rdma_ah_retrieve_grh(ah_attr);
459 	if (grh->sgid_attr)
460 		return 0;
461 
462 	sgid_attr =
463 		rdma_get_gid_attr(device, ah_attr->port_num, grh->sgid_index);
464 	if (IS_ERR(sgid_attr))
465 		return PTR_ERR(sgid_attr);
466 
467 	/* Move ownerhip of the kref into the ah_attr */
468 	grh->sgid_attr = sgid_attr;
469 	return 0;
470 }
471 
472 static void rdma_unfill_sgid_attr(struct rdma_ah_attr *ah_attr,
473 				  const struct ib_gid_attr *old_sgid_attr)
474 {
475 	/*
476 	 * Fill didn't change anything, the caller retains ownership of
477 	 * whatever it passed
478 	 */
479 	if (ah_attr->grh.sgid_attr == old_sgid_attr)
480 		return;
481 
482 	/*
483 	 * Otherwise, we need to undo what rdma_fill_sgid_attr so the caller
484 	 * doesn't see any change in the rdma_ah_attr. If we get here
485 	 * old_sgid_attr is NULL.
486 	 */
487 	rdma_destroy_ah_attr(ah_attr);
488 }
489 
490 static const struct ib_gid_attr *
491 rdma_update_sgid_attr(struct rdma_ah_attr *ah_attr,
492 		      const struct ib_gid_attr *old_attr)
493 {
494 	if (old_attr)
495 		rdma_put_gid_attr(old_attr);
496 	if (ah_attr->ah_flags & IB_AH_GRH) {
497 		rdma_hold_gid_attr(ah_attr->grh.sgid_attr);
498 		return ah_attr->grh.sgid_attr;
499 	}
500 	return NULL;
501 }
502 
503 static struct ib_ah *_rdma_create_ah(struct ib_pd *pd,
504 				     struct rdma_ah_attr *ah_attr,
505 				     u32 flags,
506 				     struct ib_udata *udata,
507 				     struct net_device *xmit_slave)
508 {
509 	struct rdma_ah_init_attr init_attr = {};
510 	struct ib_device *device = pd->device;
511 	struct ib_ah *ah;
512 	int ret;
513 
514 	might_sleep_if(flags & RDMA_CREATE_AH_SLEEPABLE);
515 
516 	if (!udata && !device->ops.create_ah)
517 		return ERR_PTR(-EOPNOTSUPP);
518 
519 	ah = rdma_zalloc_drv_obj_gfp(
520 		device, ib_ah,
521 		(flags & RDMA_CREATE_AH_SLEEPABLE) ? GFP_KERNEL : GFP_ATOMIC);
522 	if (!ah)
523 		return ERR_PTR(-ENOMEM);
524 
525 	ah->device = device;
526 	ah->pd = pd;
527 	ah->type = ah_attr->type;
528 	ah->sgid_attr = rdma_update_sgid_attr(ah_attr, NULL);
529 	init_attr.ah_attr = ah_attr;
530 	init_attr.flags = flags;
531 	init_attr.xmit_slave = xmit_slave;
532 
533 	if (udata)
534 		ret = device->ops.create_user_ah(ah, &init_attr, udata);
535 	else
536 		ret = device->ops.create_ah(ah, &init_attr, NULL);
537 	if (ret) {
538 		if (ah->sgid_attr)
539 			rdma_put_gid_attr(ah->sgid_attr);
540 		kfree(ah);
541 		return ERR_PTR(ret);
542 	}
543 
544 	atomic_inc(&pd->usecnt);
545 	return ah;
546 }
547 
548 /**
549  * rdma_create_ah - Creates an address handle for the
550  * given address vector.
551  * @pd: The protection domain associated with the address handle.
552  * @ah_attr: The attributes of the address vector.
553  * @flags: Create address handle flags (see enum rdma_create_ah_flags).
554  *
555  * It returns 0 on success and returns appropriate error code on error.
556  * The address handle is used to reference a local or global destination
557  * in all UD QP post sends.
558  */
559 struct ib_ah *rdma_create_ah(struct ib_pd *pd, struct rdma_ah_attr *ah_attr,
560 			     u32 flags)
561 {
562 	const struct ib_gid_attr *old_sgid_attr;
563 	struct net_device *slave;
564 	struct ib_ah *ah;
565 	int ret;
566 
567 	ret = rdma_fill_sgid_attr(pd->device, ah_attr, &old_sgid_attr);
568 	if (ret)
569 		return ERR_PTR(ret);
570 	slave = rdma_lag_get_ah_roce_slave(pd->device, ah_attr,
571 					   (flags & RDMA_CREATE_AH_SLEEPABLE) ?
572 					   GFP_KERNEL : GFP_ATOMIC);
573 	if (IS_ERR(slave)) {
574 		rdma_unfill_sgid_attr(ah_attr, old_sgid_attr);
575 		return (void *)slave;
576 	}
577 	ah = _rdma_create_ah(pd, ah_attr, flags, NULL, slave);
578 	rdma_lag_put_ah_roce_slave(slave);
579 	rdma_unfill_sgid_attr(ah_attr, old_sgid_attr);
580 	return ah;
581 }
582 EXPORT_SYMBOL(rdma_create_ah);
583 
584 /**
585  * rdma_create_user_ah - Creates an address handle for the
586  * given address vector.
587  * It resolves destination mac address for ah attribute of RoCE type.
588  * @pd: The protection domain associated with the address handle.
589  * @ah_attr: The attributes of the address vector.
590  * @udata: pointer to user's input output buffer information need by
591  *         provider driver.
592  *
593  * It returns 0 on success and returns appropriate error code on error.
594  * The address handle is used to reference a local or global destination
595  * in all UD QP post sends.
596  */
597 struct ib_ah *rdma_create_user_ah(struct ib_pd *pd,
598 				  struct rdma_ah_attr *ah_attr,
599 				  struct ib_udata *udata)
600 {
601 	const struct ib_gid_attr *old_sgid_attr;
602 	struct ib_ah *ah;
603 	int err;
604 
605 	err = rdma_fill_sgid_attr(pd->device, ah_attr, &old_sgid_attr);
606 	if (err)
607 		return ERR_PTR(err);
608 
609 	if (ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE) {
610 		err = ib_resolve_eth_dmac(pd->device, ah_attr);
611 		if (err) {
612 			ah = ERR_PTR(err);
613 			goto out;
614 		}
615 	}
616 
617 	ah = _rdma_create_ah(pd, ah_attr, RDMA_CREATE_AH_SLEEPABLE,
618 			     udata, NULL);
619 
620 out:
621 	rdma_unfill_sgid_attr(ah_attr, old_sgid_attr);
622 	return ah;
623 }
624 EXPORT_SYMBOL(rdma_create_user_ah);
625 
626 int ib_get_rdma_header_version(const union rdma_network_hdr *hdr)
627 {
628 	const struct iphdr *ip4h = (struct iphdr *)&hdr->roce4grh;
629 	struct iphdr ip4h_checked;
630 	const struct ipv6hdr *ip6h = (struct ipv6hdr *)&hdr->ibgrh;
631 
632 	/* If it's IPv6, the version must be 6, otherwise, the first
633 	 * 20 bytes (before the IPv4 header) are garbled.
634 	 */
635 	if (ip6h->version != 6)
636 		return (ip4h->version == 4) ? 4 : 0;
637 	/* version may be 6 or 4 because the first 20 bytes could be garbled */
638 
639 	/* RoCE v2 requires no options, thus header length
640 	 * must be 5 words
641 	 */
642 	if (ip4h->ihl != 5)
643 		return 6;
644 
645 	/* Verify checksum.
646 	 * We can't write on scattered buffers so we need to copy to
647 	 * temp buffer.
648 	 */
649 	memcpy(&ip4h_checked, ip4h, sizeof(ip4h_checked));
650 	ip4h_checked.check = 0;
651 	ip4h_checked.check = ip_fast_csum((u8 *)&ip4h_checked, 5);
652 	/* if IPv4 header checksum is OK, believe it */
653 	if (ip4h->check == ip4h_checked.check)
654 		return 4;
655 	return 6;
656 }
657 EXPORT_SYMBOL(ib_get_rdma_header_version);
658 
659 static enum rdma_network_type ib_get_net_type_by_grh(struct ib_device *device,
660 						     u32 port_num,
661 						     const struct ib_grh *grh)
662 {
663 	int grh_version;
664 
665 	if (rdma_protocol_ib(device, port_num))
666 		return RDMA_NETWORK_IB;
667 
668 	grh_version = ib_get_rdma_header_version((union rdma_network_hdr *)grh);
669 
670 	if (grh_version == 4)
671 		return RDMA_NETWORK_IPV4;
672 
673 	if (grh->next_hdr == IPPROTO_UDP)
674 		return RDMA_NETWORK_IPV6;
675 
676 	return RDMA_NETWORK_ROCE_V1;
677 }
678 
679 struct find_gid_index_context {
680 	u16 vlan_id;
681 	enum ib_gid_type gid_type;
682 };
683 
684 static bool find_gid_index(const union ib_gid *gid,
685 			   const struct ib_gid_attr *gid_attr,
686 			   void *context)
687 {
688 	struct find_gid_index_context *ctx = context;
689 	u16 vlan_id = 0xffff;
690 	int ret;
691 
692 	if (ctx->gid_type != gid_attr->gid_type)
693 		return false;
694 
695 	ret = rdma_read_gid_l2_fields(gid_attr, &vlan_id, NULL);
696 	if (ret)
697 		return false;
698 
699 	return ctx->vlan_id == vlan_id;
700 }
701 
702 static const struct ib_gid_attr *
703 get_sgid_attr_from_eth(struct ib_device *device, u32 port_num,
704 		       u16 vlan_id, const union ib_gid *sgid,
705 		       enum ib_gid_type gid_type)
706 {
707 	struct find_gid_index_context context = {.vlan_id = vlan_id,
708 						 .gid_type = gid_type};
709 
710 	return rdma_find_gid_by_filter(device, sgid, port_num, find_gid_index,
711 				       &context);
712 }
713 
714 int ib_get_gids_from_rdma_hdr(const union rdma_network_hdr *hdr,
715 			      enum rdma_network_type net_type,
716 			      union ib_gid *sgid, union ib_gid *dgid)
717 {
718 	struct sockaddr_in  src_in;
719 	struct sockaddr_in  dst_in;
720 	__be32 src_saddr, dst_saddr;
721 
722 	if (!sgid || !dgid)
723 		return -EINVAL;
724 
725 	if (net_type == RDMA_NETWORK_IPV4) {
726 		memcpy(&src_in.sin_addr.s_addr,
727 		       &hdr->roce4grh.saddr, 4);
728 		memcpy(&dst_in.sin_addr.s_addr,
729 		       &hdr->roce4grh.daddr, 4);
730 		src_saddr = src_in.sin_addr.s_addr;
731 		dst_saddr = dst_in.sin_addr.s_addr;
732 		ipv6_addr_set_v4mapped(src_saddr,
733 				       (struct in6_addr *)sgid);
734 		ipv6_addr_set_v4mapped(dst_saddr,
735 				       (struct in6_addr *)dgid);
736 		return 0;
737 	} else if (net_type == RDMA_NETWORK_IPV6 ||
738 		   net_type == RDMA_NETWORK_IB || RDMA_NETWORK_ROCE_V1) {
739 		*dgid = hdr->ibgrh.dgid;
740 		*sgid = hdr->ibgrh.sgid;
741 		return 0;
742 	} else {
743 		return -EINVAL;
744 	}
745 }
746 EXPORT_SYMBOL(ib_get_gids_from_rdma_hdr);
747 
748 /* Resolve destination mac address and hop limit for unicast destination
749  * GID entry, considering the source GID entry as well.
750  * ah_attribute must have valid port_num, sgid_index.
751  */
752 static int ib_resolve_unicast_gid_dmac(struct ib_device *device,
753 				       struct rdma_ah_attr *ah_attr)
754 {
755 	struct ib_global_route *grh = rdma_ah_retrieve_grh(ah_attr);
756 	const struct ib_gid_attr *sgid_attr = grh->sgid_attr;
757 	int hop_limit = 0xff;
758 	int ret = 0;
759 
760 	/* If destination is link local and source GID is RoCEv1,
761 	 * IP stack is not used.
762 	 */
763 	if (rdma_link_local_addr((struct in6_addr *)grh->dgid.raw) &&
764 	    sgid_attr->gid_type == IB_GID_TYPE_ROCE) {
765 		rdma_get_ll_mac((struct in6_addr *)grh->dgid.raw,
766 				ah_attr->roce.dmac);
767 		return ret;
768 	}
769 
770 	ret = rdma_addr_find_l2_eth_by_grh(&sgid_attr->gid, &grh->dgid,
771 					   ah_attr->roce.dmac,
772 					   sgid_attr, &hop_limit);
773 
774 	grh->hop_limit = hop_limit;
775 	return ret;
776 }
777 
778 /*
779  * This function initializes address handle attributes from the incoming packet.
780  * Incoming packet has dgid of the receiver node on which this code is
781  * getting executed and, sgid contains the GID of the sender.
782  *
783  * When resolving mac address of destination, the arrived dgid is used
784  * as sgid and, sgid is used as dgid because sgid contains destinations
785  * GID whom to respond to.
786  *
787  * On success the caller is responsible to call rdma_destroy_ah_attr on the
788  * attr.
789  */
790 int ib_init_ah_attr_from_wc(struct ib_device *device, u32 port_num,
791 			    const struct ib_wc *wc, const struct ib_grh *grh,
792 			    struct rdma_ah_attr *ah_attr)
793 {
794 	u32 flow_class;
795 	int ret;
796 	enum rdma_network_type net_type = RDMA_NETWORK_IB;
797 	enum ib_gid_type gid_type = IB_GID_TYPE_IB;
798 	const struct ib_gid_attr *sgid_attr;
799 	int hoplimit = 0xff;
800 	union ib_gid dgid;
801 	union ib_gid sgid;
802 
803 	might_sleep();
804 
805 	memset(ah_attr, 0, sizeof *ah_attr);
806 	ah_attr->type = rdma_ah_find_type(device, port_num);
807 	if (rdma_cap_eth_ah(device, port_num)) {
808 		if (wc->wc_flags & IB_WC_WITH_NETWORK_HDR_TYPE)
809 			net_type = wc->network_hdr_type;
810 		else
811 			net_type = ib_get_net_type_by_grh(device, port_num, grh);
812 		gid_type = ib_network_to_gid_type(net_type);
813 	}
814 	ret = ib_get_gids_from_rdma_hdr((union rdma_network_hdr *)grh, net_type,
815 					&sgid, &dgid);
816 	if (ret)
817 		return ret;
818 
819 	rdma_ah_set_sl(ah_attr, wc->sl);
820 	rdma_ah_set_port_num(ah_attr, port_num);
821 
822 	if (rdma_protocol_roce(device, port_num)) {
823 		u16 vlan_id = wc->wc_flags & IB_WC_WITH_VLAN ?
824 				wc->vlan_id : 0xffff;
825 
826 		if (!(wc->wc_flags & IB_WC_GRH))
827 			return -EPROTOTYPE;
828 
829 		sgid_attr = get_sgid_attr_from_eth(device, port_num,
830 						   vlan_id, &dgid,
831 						   gid_type);
832 		if (IS_ERR(sgid_attr))
833 			return PTR_ERR(sgid_attr);
834 
835 		flow_class = be32_to_cpu(grh->version_tclass_flow);
836 		rdma_move_grh_sgid_attr(ah_attr,
837 					&sgid,
838 					flow_class & 0xFFFFF,
839 					hoplimit,
840 					(flow_class >> 20) & 0xFF,
841 					sgid_attr);
842 
843 		ret = ib_resolve_unicast_gid_dmac(device, ah_attr);
844 		if (ret)
845 			rdma_destroy_ah_attr(ah_attr);
846 
847 		return ret;
848 	} else {
849 		rdma_ah_set_dlid(ah_attr, wc->slid);
850 		rdma_ah_set_path_bits(ah_attr, wc->dlid_path_bits);
851 
852 		if ((wc->wc_flags & IB_WC_GRH) == 0)
853 			return 0;
854 
855 		if (dgid.global.interface_id !=
856 					cpu_to_be64(IB_SA_WELL_KNOWN_GUID)) {
857 			sgid_attr = rdma_find_gid_by_port(
858 				device, &dgid, IB_GID_TYPE_IB, port_num, NULL);
859 		} else
860 			sgid_attr = rdma_get_gid_attr(device, port_num, 0);
861 
862 		if (IS_ERR(sgid_attr))
863 			return PTR_ERR(sgid_attr);
864 		flow_class = be32_to_cpu(grh->version_tclass_flow);
865 		rdma_move_grh_sgid_attr(ah_attr,
866 					&sgid,
867 					flow_class & 0xFFFFF,
868 					hoplimit,
869 					(flow_class >> 20) & 0xFF,
870 					sgid_attr);
871 
872 		return 0;
873 	}
874 }
875 EXPORT_SYMBOL(ib_init_ah_attr_from_wc);
876 
877 /**
878  * rdma_move_grh_sgid_attr - Sets the sgid attribute of GRH, taking ownership
879  * of the reference
880  *
881  * @attr:	Pointer to AH attribute structure
882  * @dgid:	Destination GID
883  * @flow_label:	Flow label
884  * @hop_limit:	Hop limit
885  * @traffic_class: traffic class
886  * @sgid_attr:	Pointer to SGID attribute
887  *
888  * This takes ownership of the sgid_attr reference. The caller must ensure
889  * rdma_destroy_ah_attr() is called before destroying the rdma_ah_attr after
890  * calling this function.
891  */
892 void rdma_move_grh_sgid_attr(struct rdma_ah_attr *attr, union ib_gid *dgid,
893 			     u32 flow_label, u8 hop_limit, u8 traffic_class,
894 			     const struct ib_gid_attr *sgid_attr)
895 {
896 	rdma_ah_set_grh(attr, dgid, flow_label, sgid_attr->index, hop_limit,
897 			traffic_class);
898 	attr->grh.sgid_attr = sgid_attr;
899 }
900 EXPORT_SYMBOL(rdma_move_grh_sgid_attr);
901 
902 /**
903  * rdma_destroy_ah_attr - Release reference to SGID attribute of
904  * ah attribute.
905  * @ah_attr: Pointer to ah attribute
906  *
907  * Release reference to the SGID attribute of the ah attribute if it is
908  * non NULL. It is safe to call this multiple times, and safe to call it on
909  * a zero initialized ah_attr.
910  */
911 void rdma_destroy_ah_attr(struct rdma_ah_attr *ah_attr)
912 {
913 	if (ah_attr->grh.sgid_attr) {
914 		rdma_put_gid_attr(ah_attr->grh.sgid_attr);
915 		ah_attr->grh.sgid_attr = NULL;
916 	}
917 }
918 EXPORT_SYMBOL(rdma_destroy_ah_attr);
919 
920 struct ib_ah *ib_create_ah_from_wc(struct ib_pd *pd, const struct ib_wc *wc,
921 				   const struct ib_grh *grh, u32 port_num)
922 {
923 	struct rdma_ah_attr ah_attr;
924 	struct ib_ah *ah;
925 	int ret;
926 
927 	ret = ib_init_ah_attr_from_wc(pd->device, port_num, wc, grh, &ah_attr);
928 	if (ret)
929 		return ERR_PTR(ret);
930 
931 	ah = rdma_create_ah(pd, &ah_attr, RDMA_CREATE_AH_SLEEPABLE);
932 
933 	rdma_destroy_ah_attr(&ah_attr);
934 	return ah;
935 }
936 EXPORT_SYMBOL(ib_create_ah_from_wc);
937 
938 int rdma_modify_ah(struct ib_ah *ah, struct rdma_ah_attr *ah_attr)
939 {
940 	const struct ib_gid_attr *old_sgid_attr;
941 	int ret;
942 
943 	if (ah->type != ah_attr->type)
944 		return -EINVAL;
945 
946 	ret = rdma_fill_sgid_attr(ah->device, ah_attr, &old_sgid_attr);
947 	if (ret)
948 		return ret;
949 
950 	ret = ah->device->ops.modify_ah ?
951 		ah->device->ops.modify_ah(ah, ah_attr) :
952 		-EOPNOTSUPP;
953 
954 	ah->sgid_attr = rdma_update_sgid_attr(ah_attr, ah->sgid_attr);
955 	rdma_unfill_sgid_attr(ah_attr, old_sgid_attr);
956 	return ret;
957 }
958 EXPORT_SYMBOL(rdma_modify_ah);
959 
960 int rdma_query_ah(struct ib_ah *ah, struct rdma_ah_attr *ah_attr)
961 {
962 	ah_attr->grh.sgid_attr = NULL;
963 
964 	return ah->device->ops.query_ah ?
965 		ah->device->ops.query_ah(ah, ah_attr) :
966 		-EOPNOTSUPP;
967 }
968 EXPORT_SYMBOL(rdma_query_ah);
969 
970 int rdma_destroy_ah_user(struct ib_ah *ah, u32 flags, struct ib_udata *udata)
971 {
972 	const struct ib_gid_attr *sgid_attr = ah->sgid_attr;
973 	struct ib_pd *pd;
974 	int ret;
975 
976 	might_sleep_if(flags & RDMA_DESTROY_AH_SLEEPABLE);
977 
978 	pd = ah->pd;
979 
980 	ret = ah->device->ops.destroy_ah(ah, flags);
981 	if (ret)
982 		return ret;
983 
984 	atomic_dec(&pd->usecnt);
985 	if (sgid_attr)
986 		rdma_put_gid_attr(sgid_attr);
987 
988 	kfree(ah);
989 	return ret;
990 }
991 EXPORT_SYMBOL(rdma_destroy_ah_user);
992 
993 /* Shared receive queues */
994 
995 /**
996  * ib_create_srq_user - Creates a SRQ associated with the specified protection
997  *   domain.
998  * @pd: The protection domain associated with the SRQ.
999  * @srq_init_attr: A list of initial attributes required to create the
1000  *   SRQ.  If SRQ creation succeeds, then the attributes are updated to
1001  *   the actual capabilities of the created SRQ.
1002  * @uobject: uobject pointer if this is not a kernel SRQ
1003  * @udata: udata pointer if this is not a kernel SRQ
1004  *
1005  * srq_attr->max_wr and srq_attr->max_sge are read the determine the
1006  * requested size of the SRQ, and set to the actual values allocated
1007  * on return.  If ib_create_srq() succeeds, then max_wr and max_sge
1008  * will always be at least as large as the requested values.
1009  */
1010 struct ib_srq *ib_create_srq_user(struct ib_pd *pd,
1011 				  struct ib_srq_init_attr *srq_init_attr,
1012 				  struct ib_usrq_object *uobject,
1013 				  struct ib_udata *udata)
1014 {
1015 	struct ib_srq *srq;
1016 	int ret;
1017 
1018 	srq = rdma_zalloc_drv_obj(pd->device, ib_srq);
1019 	if (!srq)
1020 		return ERR_PTR(-ENOMEM);
1021 
1022 	srq->device = pd->device;
1023 	srq->pd = pd;
1024 	srq->event_handler = srq_init_attr->event_handler;
1025 	srq->srq_context = srq_init_attr->srq_context;
1026 	srq->srq_type = srq_init_attr->srq_type;
1027 	srq->uobject = uobject;
1028 
1029 	if (ib_srq_has_cq(srq->srq_type)) {
1030 		srq->ext.cq = srq_init_attr->ext.cq;
1031 		atomic_inc(&srq->ext.cq->usecnt);
1032 	}
1033 	if (srq->srq_type == IB_SRQT_XRC) {
1034 		srq->ext.xrc.xrcd = srq_init_attr->ext.xrc.xrcd;
1035 		if (srq->ext.xrc.xrcd)
1036 			atomic_inc(&srq->ext.xrc.xrcd->usecnt);
1037 	}
1038 	atomic_inc(&pd->usecnt);
1039 
1040 	rdma_restrack_new(&srq->res, RDMA_RESTRACK_SRQ);
1041 	rdma_restrack_parent_name(&srq->res, &pd->res);
1042 
1043 	ret = pd->device->ops.create_srq(srq, srq_init_attr, udata);
1044 	if (ret) {
1045 		rdma_restrack_put(&srq->res);
1046 		atomic_dec(&pd->usecnt);
1047 		if (srq->srq_type == IB_SRQT_XRC && srq->ext.xrc.xrcd)
1048 			atomic_dec(&srq->ext.xrc.xrcd->usecnt);
1049 		if (ib_srq_has_cq(srq->srq_type))
1050 			atomic_dec(&srq->ext.cq->usecnt);
1051 		kfree(srq);
1052 		return ERR_PTR(ret);
1053 	}
1054 
1055 	rdma_restrack_add(&srq->res);
1056 
1057 	return srq;
1058 }
1059 EXPORT_SYMBOL(ib_create_srq_user);
1060 
1061 int ib_modify_srq(struct ib_srq *srq,
1062 		  struct ib_srq_attr *srq_attr,
1063 		  enum ib_srq_attr_mask srq_attr_mask)
1064 {
1065 	return srq->device->ops.modify_srq ?
1066 		srq->device->ops.modify_srq(srq, srq_attr, srq_attr_mask,
1067 					    NULL) : -EOPNOTSUPP;
1068 }
1069 EXPORT_SYMBOL(ib_modify_srq);
1070 
1071 int ib_query_srq(struct ib_srq *srq,
1072 		 struct ib_srq_attr *srq_attr)
1073 {
1074 	return srq->device->ops.query_srq ?
1075 		srq->device->ops.query_srq(srq, srq_attr) : -EOPNOTSUPP;
1076 }
1077 EXPORT_SYMBOL(ib_query_srq);
1078 
1079 int ib_destroy_srq_user(struct ib_srq *srq, struct ib_udata *udata)
1080 {
1081 	int ret;
1082 
1083 	if (atomic_read(&srq->usecnt))
1084 		return -EBUSY;
1085 
1086 	ret = srq->device->ops.destroy_srq(srq, udata);
1087 	if (ret)
1088 		return ret;
1089 
1090 	atomic_dec(&srq->pd->usecnt);
1091 	if (srq->srq_type == IB_SRQT_XRC && srq->ext.xrc.xrcd)
1092 		atomic_dec(&srq->ext.xrc.xrcd->usecnt);
1093 	if (ib_srq_has_cq(srq->srq_type))
1094 		atomic_dec(&srq->ext.cq->usecnt);
1095 	rdma_restrack_del(&srq->res);
1096 	kfree(srq);
1097 
1098 	return ret;
1099 }
1100 EXPORT_SYMBOL(ib_destroy_srq_user);
1101 
1102 /* Queue pairs */
1103 
1104 static void __ib_qp_event_handler(struct ib_event *event, void *context)
1105 {
1106 	struct ib_qp *qp = event->element.qp;
1107 
1108 	if (event->event == IB_EVENT_QP_LAST_WQE_REACHED)
1109 		complete(&qp->srq_completion);
1110 	if (qp->registered_event_handler)
1111 		qp->registered_event_handler(event, qp->qp_context);
1112 }
1113 
1114 static void __ib_shared_qp_event_handler(struct ib_event *event, void *context)
1115 {
1116 	struct ib_qp *qp = context;
1117 	unsigned long flags;
1118 
1119 	spin_lock_irqsave(&qp->device->qp_open_list_lock, flags);
1120 	list_for_each_entry(event->element.qp, &qp->open_list, open_list)
1121 		if (event->element.qp->event_handler)
1122 			event->element.qp->event_handler(event, event->element.qp->qp_context);
1123 	spin_unlock_irqrestore(&qp->device->qp_open_list_lock, flags);
1124 }
1125 
1126 static struct ib_qp *__ib_open_qp(struct ib_qp *real_qp,
1127 				  void (*event_handler)(struct ib_event *, void *),
1128 				  void *qp_context)
1129 {
1130 	struct ib_qp *qp;
1131 	unsigned long flags;
1132 	int err;
1133 
1134 	qp = kzalloc(sizeof *qp, GFP_KERNEL);
1135 	if (!qp)
1136 		return ERR_PTR(-ENOMEM);
1137 
1138 	qp->real_qp = real_qp;
1139 	err = ib_open_shared_qp_security(qp, real_qp->device);
1140 	if (err) {
1141 		kfree(qp);
1142 		return ERR_PTR(err);
1143 	}
1144 
1145 	qp->real_qp = real_qp;
1146 	atomic_inc(&real_qp->usecnt);
1147 	qp->device = real_qp->device;
1148 	qp->event_handler = event_handler;
1149 	qp->qp_context = qp_context;
1150 	qp->qp_num = real_qp->qp_num;
1151 	qp->qp_type = real_qp->qp_type;
1152 
1153 	spin_lock_irqsave(&real_qp->device->qp_open_list_lock, flags);
1154 	list_add(&qp->open_list, &real_qp->open_list);
1155 	spin_unlock_irqrestore(&real_qp->device->qp_open_list_lock, flags);
1156 
1157 	return qp;
1158 }
1159 
1160 struct ib_qp *ib_open_qp(struct ib_xrcd *xrcd,
1161 			 struct ib_qp_open_attr *qp_open_attr)
1162 {
1163 	struct ib_qp *qp, *real_qp;
1164 
1165 	if (qp_open_attr->qp_type != IB_QPT_XRC_TGT)
1166 		return ERR_PTR(-EINVAL);
1167 
1168 	down_read(&xrcd->tgt_qps_rwsem);
1169 	real_qp = xa_load(&xrcd->tgt_qps, qp_open_attr->qp_num);
1170 	if (!real_qp) {
1171 		up_read(&xrcd->tgt_qps_rwsem);
1172 		return ERR_PTR(-EINVAL);
1173 	}
1174 	qp = __ib_open_qp(real_qp, qp_open_attr->event_handler,
1175 			  qp_open_attr->qp_context);
1176 	up_read(&xrcd->tgt_qps_rwsem);
1177 	return qp;
1178 }
1179 EXPORT_SYMBOL(ib_open_qp);
1180 
1181 static struct ib_qp *create_xrc_qp_user(struct ib_qp *qp,
1182 					struct ib_qp_init_attr *qp_init_attr)
1183 {
1184 	struct ib_qp *real_qp = qp;
1185 	int err;
1186 
1187 	qp->event_handler = __ib_shared_qp_event_handler;
1188 	qp->qp_context = qp;
1189 	qp->pd = NULL;
1190 	qp->send_cq = qp->recv_cq = NULL;
1191 	qp->srq = NULL;
1192 	qp->xrcd = qp_init_attr->xrcd;
1193 	atomic_inc(&qp_init_attr->xrcd->usecnt);
1194 	INIT_LIST_HEAD(&qp->open_list);
1195 
1196 	qp = __ib_open_qp(real_qp, qp_init_attr->event_handler,
1197 			  qp_init_attr->qp_context);
1198 	if (IS_ERR(qp))
1199 		return qp;
1200 
1201 	err = xa_err(xa_store(&qp_init_attr->xrcd->tgt_qps, real_qp->qp_num,
1202 			      real_qp, GFP_KERNEL));
1203 	if (err) {
1204 		ib_close_qp(qp);
1205 		return ERR_PTR(err);
1206 	}
1207 	return qp;
1208 }
1209 
1210 static struct ib_qp *create_qp(struct ib_device *dev, struct ib_pd *pd,
1211 			       struct ib_qp_init_attr *attr,
1212 			       struct ib_udata *udata,
1213 			       struct ib_uqp_object *uobj, const char *caller)
1214 {
1215 	struct ib_udata dummy = {};
1216 	struct ib_qp *qp;
1217 	int ret;
1218 
1219 	if (!dev->ops.create_qp)
1220 		return ERR_PTR(-EOPNOTSUPP);
1221 
1222 	qp = rdma_zalloc_drv_obj_numa(dev, ib_qp);
1223 	if (!qp)
1224 		return ERR_PTR(-ENOMEM);
1225 
1226 	qp->device = dev;
1227 	qp->pd = pd;
1228 	qp->uobject = uobj;
1229 	qp->real_qp = qp;
1230 
1231 	qp->qp_type = attr->qp_type;
1232 	qp->rwq_ind_tbl = attr->rwq_ind_tbl;
1233 	qp->srq = attr->srq;
1234 	qp->event_handler = __ib_qp_event_handler;
1235 	qp->registered_event_handler = attr->event_handler;
1236 	qp->port = attr->port_num;
1237 	qp->qp_context = attr->qp_context;
1238 
1239 	spin_lock_init(&qp->mr_lock);
1240 	INIT_LIST_HEAD(&qp->rdma_mrs);
1241 	INIT_LIST_HEAD(&qp->sig_mrs);
1242 	init_completion(&qp->srq_completion);
1243 
1244 	qp->send_cq = attr->send_cq;
1245 	qp->recv_cq = attr->recv_cq;
1246 
1247 	rdma_restrack_new(&qp->res, RDMA_RESTRACK_QP);
1248 	WARN_ONCE(!udata && !caller, "Missing kernel QP owner");
1249 	rdma_restrack_set_name(&qp->res, udata ? NULL : caller);
1250 	ret = dev->ops.create_qp(qp, attr, udata);
1251 	if (ret)
1252 		goto err_create;
1253 
1254 	/*
1255 	 * TODO: The mlx4 internally overwrites send_cq and recv_cq.
1256 	 * Unfortunately, it is not an easy task to fix that driver.
1257 	 */
1258 	qp->send_cq = attr->send_cq;
1259 	qp->recv_cq = attr->recv_cq;
1260 
1261 	ret = ib_create_qp_security(qp, dev);
1262 	if (ret)
1263 		goto err_security;
1264 
1265 	rdma_restrack_add(&qp->res);
1266 	return qp;
1267 
1268 err_security:
1269 	qp->device->ops.destroy_qp(qp, udata ? &dummy : NULL);
1270 err_create:
1271 	rdma_restrack_put(&qp->res);
1272 	kfree(qp);
1273 	return ERR_PTR(ret);
1274 
1275 }
1276 
1277 /**
1278  * ib_create_qp_user - Creates a QP associated with the specified protection
1279  *   domain.
1280  * @dev: IB device
1281  * @pd: The protection domain associated with the QP.
1282  * @attr: A list of initial attributes required to create the
1283  *   QP.  If QP creation succeeds, then the attributes are updated to
1284  *   the actual capabilities of the created QP.
1285  * @udata: User data
1286  * @uobj: uverbs obect
1287  * @caller: caller's build-time module name
1288  */
1289 struct ib_qp *ib_create_qp_user(struct ib_device *dev, struct ib_pd *pd,
1290 				struct ib_qp_init_attr *attr,
1291 				struct ib_udata *udata,
1292 				struct ib_uqp_object *uobj, const char *caller)
1293 {
1294 	struct ib_qp *qp, *xrc_qp;
1295 
1296 	if (attr->qp_type == IB_QPT_XRC_TGT)
1297 		qp = create_qp(dev, pd, attr, NULL, NULL, caller);
1298 	else
1299 		qp = create_qp(dev, pd, attr, udata, uobj, NULL);
1300 	if (attr->qp_type != IB_QPT_XRC_TGT || IS_ERR(qp))
1301 		return qp;
1302 
1303 	xrc_qp = create_xrc_qp_user(qp, attr);
1304 	if (IS_ERR(xrc_qp)) {
1305 		ib_destroy_qp(qp);
1306 		return xrc_qp;
1307 	}
1308 
1309 	xrc_qp->uobject = uobj;
1310 	return xrc_qp;
1311 }
1312 EXPORT_SYMBOL(ib_create_qp_user);
1313 
1314 void ib_qp_usecnt_inc(struct ib_qp *qp)
1315 {
1316 	if (qp->pd)
1317 		atomic_inc(&qp->pd->usecnt);
1318 	if (qp->send_cq)
1319 		atomic_inc(&qp->send_cq->usecnt);
1320 	if (qp->recv_cq)
1321 		atomic_inc(&qp->recv_cq->usecnt);
1322 	if (qp->srq)
1323 		atomic_inc(&qp->srq->usecnt);
1324 	if (qp->rwq_ind_tbl)
1325 		atomic_inc(&qp->rwq_ind_tbl->usecnt);
1326 }
1327 EXPORT_SYMBOL(ib_qp_usecnt_inc);
1328 
1329 void ib_qp_usecnt_dec(struct ib_qp *qp)
1330 {
1331 	if (qp->rwq_ind_tbl)
1332 		atomic_dec(&qp->rwq_ind_tbl->usecnt);
1333 	if (qp->srq)
1334 		atomic_dec(&qp->srq->usecnt);
1335 	if (qp->recv_cq)
1336 		atomic_dec(&qp->recv_cq->usecnt);
1337 	if (qp->send_cq)
1338 		atomic_dec(&qp->send_cq->usecnt);
1339 	if (qp->pd)
1340 		atomic_dec(&qp->pd->usecnt);
1341 }
1342 EXPORT_SYMBOL(ib_qp_usecnt_dec);
1343 
1344 struct ib_qp *ib_create_qp_kernel(struct ib_pd *pd,
1345 				  struct ib_qp_init_attr *qp_init_attr,
1346 				  const char *caller)
1347 {
1348 	struct ib_device *device = pd->device;
1349 	struct ib_qp *qp;
1350 	int ret;
1351 
1352 	/*
1353 	 * If the callers is using the RDMA API calculate the resources
1354 	 * needed for the RDMA READ/WRITE operations.
1355 	 *
1356 	 * Note that these callers need to pass in a port number.
1357 	 */
1358 	if (qp_init_attr->cap.max_rdma_ctxs)
1359 		rdma_rw_init_qp(device, qp_init_attr);
1360 
1361 	qp = create_qp(device, pd, qp_init_attr, NULL, NULL, caller);
1362 	if (IS_ERR(qp))
1363 		return qp;
1364 
1365 	ib_qp_usecnt_inc(qp);
1366 
1367 	if (qp_init_attr->cap.max_rdma_ctxs) {
1368 		ret = rdma_rw_init_mrs(qp, qp_init_attr);
1369 		if (ret)
1370 			goto err;
1371 	}
1372 
1373 	/*
1374 	 * Note: all hw drivers guarantee that max_send_sge is lower than
1375 	 * the device RDMA WRITE SGE limit but not all hw drivers ensure that
1376 	 * max_send_sge <= max_sge_rd.
1377 	 */
1378 	qp->max_write_sge = qp_init_attr->cap.max_send_sge;
1379 	qp->max_read_sge = min_t(u32, qp_init_attr->cap.max_send_sge,
1380 				 device->attrs.max_sge_rd);
1381 	if (qp_init_attr->create_flags & IB_QP_CREATE_INTEGRITY_EN)
1382 		qp->integrity_en = true;
1383 
1384 	return qp;
1385 
1386 err:
1387 	ib_destroy_qp(qp);
1388 	return ERR_PTR(ret);
1389 
1390 }
1391 EXPORT_SYMBOL(ib_create_qp_kernel);
1392 
1393 static const struct {
1394 	int			valid;
1395 	enum ib_qp_attr_mask	req_param[IB_QPT_MAX];
1396 	enum ib_qp_attr_mask	opt_param[IB_QPT_MAX];
1397 } qp_state_table[IB_QPS_ERR + 1][IB_QPS_ERR + 1] = {
1398 	[IB_QPS_RESET] = {
1399 		[IB_QPS_RESET] = { .valid = 1 },
1400 		[IB_QPS_INIT]  = {
1401 			.valid = 1,
1402 			.req_param = {
1403 				[IB_QPT_UD]  = (IB_QP_PKEY_INDEX		|
1404 						IB_QP_PORT			|
1405 						IB_QP_QKEY),
1406 				[IB_QPT_RAW_PACKET] = IB_QP_PORT,
1407 				[IB_QPT_UC]  = (IB_QP_PKEY_INDEX		|
1408 						IB_QP_PORT			|
1409 						IB_QP_ACCESS_FLAGS),
1410 				[IB_QPT_RC]  = (IB_QP_PKEY_INDEX		|
1411 						IB_QP_PORT			|
1412 						IB_QP_ACCESS_FLAGS),
1413 				[IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX		|
1414 						IB_QP_PORT			|
1415 						IB_QP_ACCESS_FLAGS),
1416 				[IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX		|
1417 						IB_QP_PORT			|
1418 						IB_QP_ACCESS_FLAGS),
1419 				[IB_QPT_SMI] = (IB_QP_PKEY_INDEX		|
1420 						IB_QP_QKEY),
1421 				[IB_QPT_GSI] = (IB_QP_PKEY_INDEX		|
1422 						IB_QP_QKEY),
1423 			}
1424 		},
1425 	},
1426 	[IB_QPS_INIT]  = {
1427 		[IB_QPS_RESET] = { .valid = 1 },
1428 		[IB_QPS_ERR] =   { .valid = 1 },
1429 		[IB_QPS_INIT]  = {
1430 			.valid = 1,
1431 			.opt_param = {
1432 				[IB_QPT_UD]  = (IB_QP_PKEY_INDEX		|
1433 						IB_QP_PORT			|
1434 						IB_QP_QKEY),
1435 				[IB_QPT_UC]  = (IB_QP_PKEY_INDEX		|
1436 						IB_QP_PORT			|
1437 						IB_QP_ACCESS_FLAGS),
1438 				[IB_QPT_RC]  = (IB_QP_PKEY_INDEX		|
1439 						IB_QP_PORT			|
1440 						IB_QP_ACCESS_FLAGS),
1441 				[IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX		|
1442 						IB_QP_PORT			|
1443 						IB_QP_ACCESS_FLAGS),
1444 				[IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX		|
1445 						IB_QP_PORT			|
1446 						IB_QP_ACCESS_FLAGS),
1447 				[IB_QPT_SMI] = (IB_QP_PKEY_INDEX		|
1448 						IB_QP_QKEY),
1449 				[IB_QPT_GSI] = (IB_QP_PKEY_INDEX		|
1450 						IB_QP_QKEY),
1451 			}
1452 		},
1453 		[IB_QPS_RTR]   = {
1454 			.valid = 1,
1455 			.req_param = {
1456 				[IB_QPT_UC]  = (IB_QP_AV			|
1457 						IB_QP_PATH_MTU			|
1458 						IB_QP_DEST_QPN			|
1459 						IB_QP_RQ_PSN),
1460 				[IB_QPT_RC]  = (IB_QP_AV			|
1461 						IB_QP_PATH_MTU			|
1462 						IB_QP_DEST_QPN			|
1463 						IB_QP_RQ_PSN			|
1464 						IB_QP_MAX_DEST_RD_ATOMIC	|
1465 						IB_QP_MIN_RNR_TIMER),
1466 				[IB_QPT_XRC_INI] = (IB_QP_AV			|
1467 						IB_QP_PATH_MTU			|
1468 						IB_QP_DEST_QPN			|
1469 						IB_QP_RQ_PSN),
1470 				[IB_QPT_XRC_TGT] = (IB_QP_AV			|
1471 						IB_QP_PATH_MTU			|
1472 						IB_QP_DEST_QPN			|
1473 						IB_QP_RQ_PSN			|
1474 						IB_QP_MAX_DEST_RD_ATOMIC	|
1475 						IB_QP_MIN_RNR_TIMER),
1476 			},
1477 			.opt_param = {
1478 				 [IB_QPT_UD]  = (IB_QP_PKEY_INDEX		|
1479 						 IB_QP_QKEY),
1480 				 [IB_QPT_UC]  = (IB_QP_ALT_PATH			|
1481 						 IB_QP_ACCESS_FLAGS		|
1482 						 IB_QP_PKEY_INDEX),
1483 				 [IB_QPT_RC]  = (IB_QP_ALT_PATH			|
1484 						 IB_QP_ACCESS_FLAGS		|
1485 						 IB_QP_PKEY_INDEX),
1486 				 [IB_QPT_XRC_INI] = (IB_QP_ALT_PATH		|
1487 						 IB_QP_ACCESS_FLAGS		|
1488 						 IB_QP_PKEY_INDEX),
1489 				 [IB_QPT_XRC_TGT] = (IB_QP_ALT_PATH		|
1490 						 IB_QP_ACCESS_FLAGS		|
1491 						 IB_QP_PKEY_INDEX),
1492 				 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX		|
1493 						 IB_QP_QKEY),
1494 				 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX		|
1495 						 IB_QP_QKEY),
1496 			 },
1497 		},
1498 	},
1499 	[IB_QPS_RTR]   = {
1500 		[IB_QPS_RESET] = { .valid = 1 },
1501 		[IB_QPS_ERR] =   { .valid = 1 },
1502 		[IB_QPS_RTS]   = {
1503 			.valid = 1,
1504 			.req_param = {
1505 				[IB_QPT_UD]  = IB_QP_SQ_PSN,
1506 				[IB_QPT_UC]  = IB_QP_SQ_PSN,
1507 				[IB_QPT_RC]  = (IB_QP_TIMEOUT			|
1508 						IB_QP_RETRY_CNT			|
1509 						IB_QP_RNR_RETRY			|
1510 						IB_QP_SQ_PSN			|
1511 						IB_QP_MAX_QP_RD_ATOMIC),
1512 				[IB_QPT_XRC_INI] = (IB_QP_TIMEOUT		|
1513 						IB_QP_RETRY_CNT			|
1514 						IB_QP_RNR_RETRY			|
1515 						IB_QP_SQ_PSN			|
1516 						IB_QP_MAX_QP_RD_ATOMIC),
1517 				[IB_QPT_XRC_TGT] = (IB_QP_TIMEOUT		|
1518 						IB_QP_SQ_PSN),
1519 				[IB_QPT_SMI] = IB_QP_SQ_PSN,
1520 				[IB_QPT_GSI] = IB_QP_SQ_PSN,
1521 			},
1522 			.opt_param = {
1523 				 [IB_QPT_UD]  = (IB_QP_CUR_STATE		|
1524 						 IB_QP_QKEY),
1525 				 [IB_QPT_UC]  = (IB_QP_CUR_STATE		|
1526 						 IB_QP_ALT_PATH			|
1527 						 IB_QP_ACCESS_FLAGS		|
1528 						 IB_QP_PATH_MIG_STATE),
1529 				 [IB_QPT_RC]  = (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_XRC_INI] = (IB_QP_CUR_STATE		|
1535 						 IB_QP_ALT_PATH			|
1536 						 IB_QP_ACCESS_FLAGS		|
1537 						 IB_QP_PATH_MIG_STATE),
1538 				 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE		|
1539 						 IB_QP_ALT_PATH			|
1540 						 IB_QP_ACCESS_FLAGS		|
1541 						 IB_QP_MIN_RNR_TIMER		|
1542 						 IB_QP_PATH_MIG_STATE),
1543 				 [IB_QPT_SMI] = (IB_QP_CUR_STATE		|
1544 						 IB_QP_QKEY),
1545 				 [IB_QPT_GSI] = (IB_QP_CUR_STATE		|
1546 						 IB_QP_QKEY),
1547 				 [IB_QPT_RAW_PACKET] = IB_QP_RATE_LIMIT,
1548 			 }
1549 		}
1550 	},
1551 	[IB_QPS_RTS]   = {
1552 		[IB_QPS_RESET] = { .valid = 1 },
1553 		[IB_QPS_ERR] =   { .valid = 1 },
1554 		[IB_QPS_RTS]   = {
1555 			.valid = 1,
1556 			.opt_param = {
1557 				[IB_QPT_UD]  = (IB_QP_CUR_STATE			|
1558 						IB_QP_QKEY),
1559 				[IB_QPT_UC]  = (IB_QP_CUR_STATE			|
1560 						IB_QP_ACCESS_FLAGS		|
1561 						IB_QP_ALT_PATH			|
1562 						IB_QP_PATH_MIG_STATE),
1563 				[IB_QPT_RC]  = (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_XRC_INI] = (IB_QP_CUR_STATE		|
1569 						IB_QP_ACCESS_FLAGS		|
1570 						IB_QP_ALT_PATH			|
1571 						IB_QP_PATH_MIG_STATE),
1572 				[IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE		|
1573 						IB_QP_ACCESS_FLAGS		|
1574 						IB_QP_ALT_PATH			|
1575 						IB_QP_PATH_MIG_STATE		|
1576 						IB_QP_MIN_RNR_TIMER),
1577 				[IB_QPT_SMI] = (IB_QP_CUR_STATE			|
1578 						IB_QP_QKEY),
1579 				[IB_QPT_GSI] = (IB_QP_CUR_STATE			|
1580 						IB_QP_QKEY),
1581 				[IB_QPT_RAW_PACKET] = IB_QP_RATE_LIMIT,
1582 			}
1583 		},
1584 		[IB_QPS_SQD]   = {
1585 			.valid = 1,
1586 			.opt_param = {
1587 				[IB_QPT_UD]  = IB_QP_EN_SQD_ASYNC_NOTIFY,
1588 				[IB_QPT_UC]  = IB_QP_EN_SQD_ASYNC_NOTIFY,
1589 				[IB_QPT_RC]  = IB_QP_EN_SQD_ASYNC_NOTIFY,
1590 				[IB_QPT_XRC_INI] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1591 				[IB_QPT_XRC_TGT] = IB_QP_EN_SQD_ASYNC_NOTIFY, /* ??? */
1592 				[IB_QPT_SMI] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1593 				[IB_QPT_GSI] = IB_QP_EN_SQD_ASYNC_NOTIFY
1594 			}
1595 		},
1596 	},
1597 	[IB_QPS_SQD]   = {
1598 		[IB_QPS_RESET] = { .valid = 1 },
1599 		[IB_QPS_ERR] =   { .valid = 1 },
1600 		[IB_QPS_RTS]   = {
1601 			.valid = 1,
1602 			.opt_param = {
1603 				[IB_QPT_UD]  = (IB_QP_CUR_STATE			|
1604 						IB_QP_QKEY),
1605 				[IB_QPT_UC]  = (IB_QP_CUR_STATE			|
1606 						IB_QP_ALT_PATH			|
1607 						IB_QP_ACCESS_FLAGS		|
1608 						IB_QP_PATH_MIG_STATE),
1609 				[IB_QPT_RC]  = (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_XRC_INI] = (IB_QP_CUR_STATE		|
1615 						IB_QP_ALT_PATH			|
1616 						IB_QP_ACCESS_FLAGS		|
1617 						IB_QP_PATH_MIG_STATE),
1618 				[IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE		|
1619 						IB_QP_ALT_PATH			|
1620 						IB_QP_ACCESS_FLAGS		|
1621 						IB_QP_MIN_RNR_TIMER		|
1622 						IB_QP_PATH_MIG_STATE),
1623 				[IB_QPT_SMI] = (IB_QP_CUR_STATE			|
1624 						IB_QP_QKEY),
1625 				[IB_QPT_GSI] = (IB_QP_CUR_STATE			|
1626 						IB_QP_QKEY),
1627 			}
1628 		},
1629 		[IB_QPS_SQD]   = {
1630 			.valid = 1,
1631 			.opt_param = {
1632 				[IB_QPT_UD]  = (IB_QP_PKEY_INDEX		|
1633 						IB_QP_QKEY),
1634 				[IB_QPT_UC]  = (IB_QP_AV			|
1635 						IB_QP_ALT_PATH			|
1636 						IB_QP_ACCESS_FLAGS		|
1637 						IB_QP_PKEY_INDEX		|
1638 						IB_QP_PATH_MIG_STATE),
1639 				[IB_QPT_RC]  = (IB_QP_PORT			|
1640 						IB_QP_AV			|
1641 						IB_QP_TIMEOUT			|
1642 						IB_QP_RETRY_CNT			|
1643 						IB_QP_RNR_RETRY			|
1644 						IB_QP_MAX_QP_RD_ATOMIC		|
1645 						IB_QP_MAX_DEST_RD_ATOMIC	|
1646 						IB_QP_ALT_PATH			|
1647 						IB_QP_ACCESS_FLAGS		|
1648 						IB_QP_PKEY_INDEX		|
1649 						IB_QP_MIN_RNR_TIMER		|
1650 						IB_QP_PATH_MIG_STATE),
1651 				[IB_QPT_XRC_INI] = (IB_QP_PORT			|
1652 						IB_QP_AV			|
1653 						IB_QP_TIMEOUT			|
1654 						IB_QP_RETRY_CNT			|
1655 						IB_QP_RNR_RETRY			|
1656 						IB_QP_MAX_QP_RD_ATOMIC		|
1657 						IB_QP_ALT_PATH			|
1658 						IB_QP_ACCESS_FLAGS		|
1659 						IB_QP_PKEY_INDEX		|
1660 						IB_QP_PATH_MIG_STATE),
1661 				[IB_QPT_XRC_TGT] = (IB_QP_PORT			|
1662 						IB_QP_AV			|
1663 						IB_QP_TIMEOUT			|
1664 						IB_QP_MAX_DEST_RD_ATOMIC	|
1665 						IB_QP_ALT_PATH			|
1666 						IB_QP_ACCESS_FLAGS		|
1667 						IB_QP_PKEY_INDEX		|
1668 						IB_QP_MIN_RNR_TIMER		|
1669 						IB_QP_PATH_MIG_STATE),
1670 				[IB_QPT_SMI] = (IB_QP_PKEY_INDEX		|
1671 						IB_QP_QKEY),
1672 				[IB_QPT_GSI] = (IB_QP_PKEY_INDEX		|
1673 						IB_QP_QKEY),
1674 			}
1675 		}
1676 	},
1677 	[IB_QPS_SQE]   = {
1678 		[IB_QPS_RESET] = { .valid = 1 },
1679 		[IB_QPS_ERR] =   { .valid = 1 },
1680 		[IB_QPS_RTS]   = {
1681 			.valid = 1,
1682 			.opt_param = {
1683 				[IB_QPT_UD]  = (IB_QP_CUR_STATE			|
1684 						IB_QP_QKEY),
1685 				[IB_QPT_UC]  = (IB_QP_CUR_STATE			|
1686 						IB_QP_ACCESS_FLAGS),
1687 				[IB_QPT_SMI] = (IB_QP_CUR_STATE			|
1688 						IB_QP_QKEY),
1689 				[IB_QPT_GSI] = (IB_QP_CUR_STATE			|
1690 						IB_QP_QKEY),
1691 			}
1692 		}
1693 	},
1694 	[IB_QPS_ERR] = {
1695 		[IB_QPS_RESET] = { .valid = 1 },
1696 		[IB_QPS_ERR] =   { .valid = 1 }
1697 	}
1698 };
1699 
1700 bool ib_modify_qp_is_ok(enum ib_qp_state cur_state, enum ib_qp_state next_state,
1701 			enum ib_qp_type type, enum ib_qp_attr_mask mask)
1702 {
1703 	enum ib_qp_attr_mask req_param, opt_param;
1704 
1705 	if (mask & IB_QP_CUR_STATE  &&
1706 	    cur_state != IB_QPS_RTR && cur_state != IB_QPS_RTS &&
1707 	    cur_state != IB_QPS_SQD && cur_state != IB_QPS_SQE)
1708 		return false;
1709 
1710 	if (!qp_state_table[cur_state][next_state].valid)
1711 		return false;
1712 
1713 	req_param = qp_state_table[cur_state][next_state].req_param[type];
1714 	opt_param = qp_state_table[cur_state][next_state].opt_param[type];
1715 
1716 	if ((mask & req_param) != req_param)
1717 		return false;
1718 
1719 	if (mask & ~(req_param | opt_param | IB_QP_STATE))
1720 		return false;
1721 
1722 	return true;
1723 }
1724 EXPORT_SYMBOL(ib_modify_qp_is_ok);
1725 
1726 /**
1727  * ib_resolve_eth_dmac - Resolve destination mac address
1728  * @device:		Device to consider
1729  * @ah_attr:		address handle attribute which describes the
1730  *			source and destination parameters
1731  * ib_resolve_eth_dmac() resolves destination mac address and L3 hop limit It
1732  * returns 0 on success or appropriate error code. It initializes the
1733  * necessary ah_attr fields when call is successful.
1734  */
1735 static int ib_resolve_eth_dmac(struct ib_device *device,
1736 			       struct rdma_ah_attr *ah_attr)
1737 {
1738 	int ret = 0;
1739 
1740 	if (rdma_is_multicast_addr((struct in6_addr *)ah_attr->grh.dgid.raw)) {
1741 		if (ipv6_addr_v4mapped((struct in6_addr *)ah_attr->grh.dgid.raw)) {
1742 			__be32 addr = 0;
1743 
1744 			memcpy(&addr, ah_attr->grh.dgid.raw + 12, 4);
1745 			ip_eth_mc_map(addr, (char *)ah_attr->roce.dmac);
1746 		} else {
1747 			ipv6_eth_mc_map((struct in6_addr *)ah_attr->grh.dgid.raw,
1748 					(char *)ah_attr->roce.dmac);
1749 		}
1750 	} else {
1751 		ret = ib_resolve_unicast_gid_dmac(device, ah_attr);
1752 	}
1753 	return ret;
1754 }
1755 
1756 static bool is_qp_type_connected(const struct ib_qp *qp)
1757 {
1758 	return (qp->qp_type == IB_QPT_UC ||
1759 		qp->qp_type == IB_QPT_RC ||
1760 		qp->qp_type == IB_QPT_XRC_INI ||
1761 		qp->qp_type == IB_QPT_XRC_TGT);
1762 }
1763 
1764 /*
1765  * IB core internal function to perform QP attributes modification.
1766  */
1767 static int _ib_modify_qp(struct ib_qp *qp, struct ib_qp_attr *attr,
1768 			 int attr_mask, struct ib_udata *udata)
1769 {
1770 	u32 port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
1771 	const struct ib_gid_attr *old_sgid_attr_av;
1772 	const struct ib_gid_attr *old_sgid_attr_alt_av;
1773 	int ret;
1774 
1775 	attr->xmit_slave = NULL;
1776 	if (attr_mask & IB_QP_AV) {
1777 		ret = rdma_fill_sgid_attr(qp->device, &attr->ah_attr,
1778 					  &old_sgid_attr_av);
1779 		if (ret)
1780 			return ret;
1781 
1782 		if (attr->ah_attr.type == RDMA_AH_ATTR_TYPE_ROCE &&
1783 		    is_qp_type_connected(qp)) {
1784 			struct net_device *slave;
1785 
1786 			/*
1787 			 * If the user provided the qp_attr then we have to
1788 			 * resolve it. Kerne users have to provide already
1789 			 * resolved rdma_ah_attr's.
1790 			 */
1791 			if (udata) {
1792 				ret = ib_resolve_eth_dmac(qp->device,
1793 							  &attr->ah_attr);
1794 				if (ret)
1795 					goto out_av;
1796 			}
1797 			slave = rdma_lag_get_ah_roce_slave(qp->device,
1798 							   &attr->ah_attr,
1799 							   GFP_KERNEL);
1800 			if (IS_ERR(slave)) {
1801 				ret = PTR_ERR(slave);
1802 				goto out_av;
1803 			}
1804 			attr->xmit_slave = slave;
1805 		}
1806 	}
1807 	if (attr_mask & IB_QP_ALT_PATH) {
1808 		/*
1809 		 * FIXME: This does not track the migration state, so if the
1810 		 * user loads a new alternate path after the HW has migrated
1811 		 * from primary->alternate we will keep the wrong
1812 		 * references. This is OK for IB because the reference
1813 		 * counting does not serve any functional purpose.
1814 		 */
1815 		ret = rdma_fill_sgid_attr(qp->device, &attr->alt_ah_attr,
1816 					  &old_sgid_attr_alt_av);
1817 		if (ret)
1818 			goto out_av;
1819 
1820 		/*
1821 		 * Today the core code can only handle alternate paths and APM
1822 		 * for IB. Ban them in roce mode.
1823 		 */
1824 		if (!(rdma_protocol_ib(qp->device,
1825 				       attr->alt_ah_attr.port_num) &&
1826 		      rdma_protocol_ib(qp->device, port))) {
1827 			ret = -EINVAL;
1828 			goto out;
1829 		}
1830 	}
1831 
1832 	if (rdma_ib_or_roce(qp->device, port)) {
1833 		if (attr_mask & IB_QP_RQ_PSN && attr->rq_psn & ~0xffffff) {
1834 			dev_warn(&qp->device->dev,
1835 				 "%s rq_psn overflow, masking to 24 bits\n",
1836 				 __func__);
1837 			attr->rq_psn &= 0xffffff;
1838 		}
1839 
1840 		if (attr_mask & IB_QP_SQ_PSN && attr->sq_psn & ~0xffffff) {
1841 			dev_warn(&qp->device->dev,
1842 				 " %s sq_psn overflow, masking to 24 bits\n",
1843 				 __func__);
1844 			attr->sq_psn &= 0xffffff;
1845 		}
1846 	}
1847 
1848 	/*
1849 	 * Bind this qp to a counter automatically based on the rdma counter
1850 	 * rules. This only set in RST2INIT with port specified
1851 	 */
1852 	if (!qp->counter && (attr_mask & IB_QP_PORT) &&
1853 	    ((attr_mask & IB_QP_STATE) && attr->qp_state == IB_QPS_INIT))
1854 		rdma_counter_bind_qp_auto(qp, attr->port_num);
1855 
1856 	ret = ib_security_modify_qp(qp, attr, attr_mask, udata);
1857 	if (ret)
1858 		goto out;
1859 
1860 	if (attr_mask & IB_QP_PORT)
1861 		qp->port = attr->port_num;
1862 	if (attr_mask & IB_QP_AV)
1863 		qp->av_sgid_attr =
1864 			rdma_update_sgid_attr(&attr->ah_attr, qp->av_sgid_attr);
1865 	if (attr_mask & IB_QP_ALT_PATH)
1866 		qp->alt_path_sgid_attr = rdma_update_sgid_attr(
1867 			&attr->alt_ah_attr, qp->alt_path_sgid_attr);
1868 
1869 out:
1870 	if (attr_mask & IB_QP_ALT_PATH)
1871 		rdma_unfill_sgid_attr(&attr->alt_ah_attr, old_sgid_attr_alt_av);
1872 out_av:
1873 	if (attr_mask & IB_QP_AV) {
1874 		rdma_lag_put_ah_roce_slave(attr->xmit_slave);
1875 		rdma_unfill_sgid_attr(&attr->ah_attr, old_sgid_attr_av);
1876 	}
1877 	return ret;
1878 }
1879 
1880 /**
1881  * ib_modify_qp_with_udata - Modifies the attributes for the specified QP.
1882  * @ib_qp: The QP to modify.
1883  * @attr: On input, specifies the QP attributes to modify.  On output,
1884  *   the current values of selected QP attributes are returned.
1885  * @attr_mask: A bit-mask used to specify which attributes of the QP
1886  *   are being modified.
1887  * @udata: pointer to user's input output buffer information
1888  *   are being modified.
1889  * It returns 0 on success and returns appropriate error code on error.
1890  */
1891 int ib_modify_qp_with_udata(struct ib_qp *ib_qp, struct ib_qp_attr *attr,
1892 			    int attr_mask, struct ib_udata *udata)
1893 {
1894 	return _ib_modify_qp(ib_qp->real_qp, attr, attr_mask, udata);
1895 }
1896 EXPORT_SYMBOL(ib_modify_qp_with_udata);
1897 
1898 static void ib_get_width_and_speed(u32 netdev_speed, u32 lanes,
1899 				   u16 *speed, u8 *width)
1900 {
1901 	if (!lanes) {
1902 		if (netdev_speed <= SPEED_1000) {
1903 			*width = IB_WIDTH_1X;
1904 			*speed = IB_SPEED_SDR;
1905 		} else if (netdev_speed <= SPEED_10000) {
1906 			*width = IB_WIDTH_1X;
1907 			*speed = IB_SPEED_FDR10;
1908 		} else if (netdev_speed <= SPEED_20000) {
1909 			*width = IB_WIDTH_4X;
1910 			*speed = IB_SPEED_DDR;
1911 		} else if (netdev_speed <= SPEED_25000) {
1912 			*width = IB_WIDTH_1X;
1913 			*speed = IB_SPEED_EDR;
1914 		} else if (netdev_speed <= SPEED_40000) {
1915 			*width = IB_WIDTH_4X;
1916 			*speed = IB_SPEED_FDR10;
1917 		} else if (netdev_speed <= SPEED_50000) {
1918 			*width = IB_WIDTH_2X;
1919 			*speed = IB_SPEED_EDR;
1920 		} else if (netdev_speed <= SPEED_100000) {
1921 			*width = IB_WIDTH_4X;
1922 			*speed = IB_SPEED_EDR;
1923 		} else if (netdev_speed <= SPEED_200000) {
1924 			*width = IB_WIDTH_4X;
1925 			*speed = IB_SPEED_HDR;
1926 		} else {
1927 			*width = IB_WIDTH_4X;
1928 			*speed = IB_SPEED_NDR;
1929 		}
1930 
1931 		return;
1932 	}
1933 
1934 	switch (lanes) {
1935 	case 1:
1936 		*width = IB_WIDTH_1X;
1937 		break;
1938 	case 2:
1939 		*width = IB_WIDTH_2X;
1940 		break;
1941 	case 4:
1942 		*width = IB_WIDTH_4X;
1943 		break;
1944 	case 8:
1945 		*width = IB_WIDTH_8X;
1946 		break;
1947 	case 12:
1948 		*width = IB_WIDTH_12X;
1949 		break;
1950 	default:
1951 		*width = IB_WIDTH_1X;
1952 	}
1953 
1954 	switch (netdev_speed / lanes) {
1955 	case SPEED_2500:
1956 		*speed = IB_SPEED_SDR;
1957 		break;
1958 	case SPEED_5000:
1959 		*speed = IB_SPEED_DDR;
1960 		break;
1961 	case SPEED_10000:
1962 		*speed = IB_SPEED_FDR10;
1963 		break;
1964 	case SPEED_14000:
1965 		*speed = IB_SPEED_FDR;
1966 		break;
1967 	case SPEED_25000:
1968 		*speed = IB_SPEED_EDR;
1969 		break;
1970 	case SPEED_50000:
1971 		*speed = IB_SPEED_HDR;
1972 		break;
1973 	case SPEED_100000:
1974 		*speed = IB_SPEED_NDR;
1975 		break;
1976 	default:
1977 		*speed = IB_SPEED_SDR;
1978 	}
1979 }
1980 
1981 int ib_get_eth_speed(struct ib_device *dev, u32 port_num, u16 *speed, u8 *width)
1982 {
1983 	int rc;
1984 	u32 netdev_speed;
1985 	struct net_device *netdev;
1986 	struct ethtool_link_ksettings lksettings = {};
1987 
1988 	if (rdma_port_get_link_layer(dev, port_num) != IB_LINK_LAYER_ETHERNET)
1989 		return -EINVAL;
1990 
1991 	netdev = ib_device_get_netdev(dev, port_num);
1992 	if (!netdev)
1993 		return -ENODEV;
1994 
1995 	rtnl_lock();
1996 	rc = __ethtool_get_link_ksettings(netdev, &lksettings);
1997 	rtnl_unlock();
1998 
1999 	dev_put(netdev);
2000 
2001 	if (!rc && lksettings.base.speed != (u32)SPEED_UNKNOWN) {
2002 		netdev_speed = lksettings.base.speed;
2003 	} else {
2004 		netdev_speed = SPEED_1000;
2005 		if (rc)
2006 			pr_warn("%s speed is unknown, defaulting to %u\n",
2007 				netdev->name, netdev_speed);
2008 	}
2009 
2010 	ib_get_width_and_speed(netdev_speed, lksettings.lanes,
2011 			       speed, width);
2012 
2013 	return 0;
2014 }
2015 EXPORT_SYMBOL(ib_get_eth_speed);
2016 
2017 int ib_modify_qp(struct ib_qp *qp,
2018 		 struct ib_qp_attr *qp_attr,
2019 		 int qp_attr_mask)
2020 {
2021 	return _ib_modify_qp(qp->real_qp, qp_attr, qp_attr_mask, NULL);
2022 }
2023 EXPORT_SYMBOL(ib_modify_qp);
2024 
2025 int ib_query_qp(struct ib_qp *qp,
2026 		struct ib_qp_attr *qp_attr,
2027 		int qp_attr_mask,
2028 		struct ib_qp_init_attr *qp_init_attr)
2029 {
2030 	qp_attr->ah_attr.grh.sgid_attr = NULL;
2031 	qp_attr->alt_ah_attr.grh.sgid_attr = NULL;
2032 
2033 	return qp->device->ops.query_qp ?
2034 		qp->device->ops.query_qp(qp->real_qp, qp_attr, qp_attr_mask,
2035 					 qp_init_attr) : -EOPNOTSUPP;
2036 }
2037 EXPORT_SYMBOL(ib_query_qp);
2038 
2039 int ib_close_qp(struct ib_qp *qp)
2040 {
2041 	struct ib_qp *real_qp;
2042 	unsigned long flags;
2043 
2044 	real_qp = qp->real_qp;
2045 	if (real_qp == qp)
2046 		return -EINVAL;
2047 
2048 	spin_lock_irqsave(&real_qp->device->qp_open_list_lock, flags);
2049 	list_del(&qp->open_list);
2050 	spin_unlock_irqrestore(&real_qp->device->qp_open_list_lock, flags);
2051 
2052 	atomic_dec(&real_qp->usecnt);
2053 	if (qp->qp_sec)
2054 		ib_close_shared_qp_security(qp->qp_sec);
2055 	kfree(qp);
2056 
2057 	return 0;
2058 }
2059 EXPORT_SYMBOL(ib_close_qp);
2060 
2061 static int __ib_destroy_shared_qp(struct ib_qp *qp)
2062 {
2063 	struct ib_xrcd *xrcd;
2064 	struct ib_qp *real_qp;
2065 	int ret;
2066 
2067 	real_qp = qp->real_qp;
2068 	xrcd = real_qp->xrcd;
2069 	down_write(&xrcd->tgt_qps_rwsem);
2070 	ib_close_qp(qp);
2071 	if (atomic_read(&real_qp->usecnt) == 0)
2072 		xa_erase(&xrcd->tgt_qps, real_qp->qp_num);
2073 	else
2074 		real_qp = NULL;
2075 	up_write(&xrcd->tgt_qps_rwsem);
2076 
2077 	if (real_qp) {
2078 		ret = ib_destroy_qp(real_qp);
2079 		if (!ret)
2080 			atomic_dec(&xrcd->usecnt);
2081 	}
2082 
2083 	return 0;
2084 }
2085 
2086 int ib_destroy_qp_user(struct ib_qp *qp, struct ib_udata *udata)
2087 {
2088 	const struct ib_gid_attr *alt_path_sgid_attr = qp->alt_path_sgid_attr;
2089 	const struct ib_gid_attr *av_sgid_attr = qp->av_sgid_attr;
2090 	struct ib_qp_security *sec;
2091 	int ret;
2092 
2093 	WARN_ON_ONCE(qp->mrs_used > 0);
2094 
2095 	if (atomic_read(&qp->usecnt))
2096 		return -EBUSY;
2097 
2098 	if (qp->real_qp != qp)
2099 		return __ib_destroy_shared_qp(qp);
2100 
2101 	sec  = qp->qp_sec;
2102 	if (sec)
2103 		ib_destroy_qp_security_begin(sec);
2104 
2105 	if (!qp->uobject)
2106 		rdma_rw_cleanup_mrs(qp);
2107 
2108 	rdma_counter_unbind_qp(qp, true);
2109 	ret = qp->device->ops.destroy_qp(qp, udata);
2110 	if (ret) {
2111 		if (sec)
2112 			ib_destroy_qp_security_abort(sec);
2113 		return ret;
2114 	}
2115 
2116 	if (alt_path_sgid_attr)
2117 		rdma_put_gid_attr(alt_path_sgid_attr);
2118 	if (av_sgid_attr)
2119 		rdma_put_gid_attr(av_sgid_attr);
2120 
2121 	ib_qp_usecnt_dec(qp);
2122 	if (sec)
2123 		ib_destroy_qp_security_end(sec);
2124 
2125 	rdma_restrack_del(&qp->res);
2126 	kfree(qp);
2127 	return ret;
2128 }
2129 EXPORT_SYMBOL(ib_destroy_qp_user);
2130 
2131 /* Completion queues */
2132 
2133 struct ib_cq *__ib_create_cq(struct ib_device *device,
2134 			     ib_comp_handler comp_handler,
2135 			     void (*event_handler)(struct ib_event *, void *),
2136 			     void *cq_context,
2137 			     const struct ib_cq_init_attr *cq_attr,
2138 			     const char *caller)
2139 {
2140 	struct ib_cq *cq;
2141 	int ret;
2142 
2143 	cq = rdma_zalloc_drv_obj(device, ib_cq);
2144 	if (!cq)
2145 		return ERR_PTR(-ENOMEM);
2146 
2147 	cq->device = device;
2148 	cq->uobject = NULL;
2149 	cq->comp_handler = comp_handler;
2150 	cq->event_handler = event_handler;
2151 	cq->cq_context = cq_context;
2152 	atomic_set(&cq->usecnt, 0);
2153 
2154 	rdma_restrack_new(&cq->res, RDMA_RESTRACK_CQ);
2155 	rdma_restrack_set_name(&cq->res, caller);
2156 
2157 	ret = device->ops.create_cq(cq, cq_attr, NULL);
2158 	if (ret) {
2159 		rdma_restrack_put(&cq->res);
2160 		kfree(cq);
2161 		return ERR_PTR(ret);
2162 	}
2163 
2164 	rdma_restrack_add(&cq->res);
2165 	return cq;
2166 }
2167 EXPORT_SYMBOL(__ib_create_cq);
2168 
2169 int rdma_set_cq_moderation(struct ib_cq *cq, u16 cq_count, u16 cq_period)
2170 {
2171 	if (cq->shared)
2172 		return -EOPNOTSUPP;
2173 
2174 	return cq->device->ops.modify_cq ?
2175 		cq->device->ops.modify_cq(cq, cq_count,
2176 					  cq_period) : -EOPNOTSUPP;
2177 }
2178 EXPORT_SYMBOL(rdma_set_cq_moderation);
2179 
2180 int ib_destroy_cq_user(struct ib_cq *cq, struct ib_udata *udata)
2181 {
2182 	int ret;
2183 
2184 	if (WARN_ON_ONCE(cq->shared))
2185 		return -EOPNOTSUPP;
2186 
2187 	if (atomic_read(&cq->usecnt))
2188 		return -EBUSY;
2189 
2190 	ret = cq->device->ops.destroy_cq(cq, udata);
2191 	if (ret)
2192 		return ret;
2193 
2194 	rdma_restrack_del(&cq->res);
2195 	kfree(cq);
2196 	return ret;
2197 }
2198 EXPORT_SYMBOL(ib_destroy_cq_user);
2199 
2200 int ib_resize_cq(struct ib_cq *cq, int cqe)
2201 {
2202 	if (cq->shared)
2203 		return -EOPNOTSUPP;
2204 
2205 	return cq->device->ops.resize_cq ?
2206 		cq->device->ops.resize_cq(cq, cqe, NULL) : -EOPNOTSUPP;
2207 }
2208 EXPORT_SYMBOL(ib_resize_cq);
2209 
2210 /* Memory regions */
2211 
2212 struct ib_mr *ib_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
2213 			     u64 virt_addr, int access_flags)
2214 {
2215 	struct ib_mr *mr;
2216 
2217 	if (access_flags & IB_ACCESS_ON_DEMAND) {
2218 		if (!(pd->device->attrs.kernel_cap_flags &
2219 		      IBK_ON_DEMAND_PAGING)) {
2220 			pr_debug("ODP support not available\n");
2221 			return ERR_PTR(-EINVAL);
2222 		}
2223 	}
2224 
2225 	mr = pd->device->ops.reg_user_mr(pd, start, length, virt_addr,
2226 					 access_flags, NULL);
2227 
2228 	if (IS_ERR(mr))
2229 		return mr;
2230 
2231 	mr->device = pd->device;
2232 	mr->type = IB_MR_TYPE_USER;
2233 	mr->pd = pd;
2234 	mr->dm = NULL;
2235 	atomic_inc(&pd->usecnt);
2236 	mr->iova =  virt_addr;
2237 	mr->length = length;
2238 
2239 	rdma_restrack_new(&mr->res, RDMA_RESTRACK_MR);
2240 	rdma_restrack_parent_name(&mr->res, &pd->res);
2241 	rdma_restrack_add(&mr->res);
2242 
2243 	return mr;
2244 }
2245 EXPORT_SYMBOL(ib_reg_user_mr);
2246 
2247 int ib_advise_mr(struct ib_pd *pd, enum ib_uverbs_advise_mr_advice advice,
2248 		 u32 flags, struct ib_sge *sg_list, u32 num_sge)
2249 {
2250 	if (!pd->device->ops.advise_mr)
2251 		return -EOPNOTSUPP;
2252 
2253 	if (!num_sge)
2254 		return 0;
2255 
2256 	return pd->device->ops.advise_mr(pd, advice, flags, sg_list, num_sge,
2257 					 NULL);
2258 }
2259 EXPORT_SYMBOL(ib_advise_mr);
2260 
2261 int ib_dereg_mr_user(struct ib_mr *mr, struct ib_udata *udata)
2262 {
2263 	struct ib_pd *pd = mr->pd;
2264 	struct ib_dm *dm = mr->dm;
2265 	struct ib_sig_attrs *sig_attrs = mr->sig_attrs;
2266 	int ret;
2267 
2268 	trace_mr_dereg(mr);
2269 	rdma_restrack_del(&mr->res);
2270 	ret = mr->device->ops.dereg_mr(mr, udata);
2271 	if (!ret) {
2272 		atomic_dec(&pd->usecnt);
2273 		if (dm)
2274 			atomic_dec(&dm->usecnt);
2275 		kfree(sig_attrs);
2276 	}
2277 
2278 	return ret;
2279 }
2280 EXPORT_SYMBOL(ib_dereg_mr_user);
2281 
2282 /**
2283  * ib_alloc_mr() - Allocates a memory region
2284  * @pd:            protection domain associated with the region
2285  * @mr_type:       memory region type
2286  * @max_num_sg:    maximum sg entries available for registration.
2287  *
2288  * Notes:
2289  * Memory registeration page/sg lists must not exceed max_num_sg.
2290  * For mr_type IB_MR_TYPE_MEM_REG, the total length cannot exceed
2291  * max_num_sg * used_page_size.
2292  *
2293  */
2294 struct ib_mr *ib_alloc_mr(struct ib_pd *pd, enum ib_mr_type mr_type,
2295 			  u32 max_num_sg)
2296 {
2297 	struct ib_mr *mr;
2298 
2299 	if (!pd->device->ops.alloc_mr) {
2300 		mr = ERR_PTR(-EOPNOTSUPP);
2301 		goto out;
2302 	}
2303 
2304 	if (mr_type == IB_MR_TYPE_INTEGRITY) {
2305 		WARN_ON_ONCE(1);
2306 		mr = ERR_PTR(-EINVAL);
2307 		goto out;
2308 	}
2309 
2310 	mr = pd->device->ops.alloc_mr(pd, mr_type, max_num_sg);
2311 	if (IS_ERR(mr))
2312 		goto out;
2313 
2314 	mr->device = pd->device;
2315 	mr->pd = pd;
2316 	mr->dm = NULL;
2317 	mr->uobject = NULL;
2318 	atomic_inc(&pd->usecnt);
2319 	mr->need_inval = false;
2320 	mr->type = mr_type;
2321 	mr->sig_attrs = NULL;
2322 
2323 	rdma_restrack_new(&mr->res, RDMA_RESTRACK_MR);
2324 	rdma_restrack_parent_name(&mr->res, &pd->res);
2325 	rdma_restrack_add(&mr->res);
2326 out:
2327 	trace_mr_alloc(pd, mr_type, max_num_sg, mr);
2328 	return mr;
2329 }
2330 EXPORT_SYMBOL(ib_alloc_mr);
2331 
2332 /**
2333  * ib_alloc_mr_integrity() - Allocates an integrity memory region
2334  * @pd:                      protection domain associated with the region
2335  * @max_num_data_sg:         maximum data sg entries available for registration
2336  * @max_num_meta_sg:         maximum metadata sg entries available for
2337  *                           registration
2338  *
2339  * Notes:
2340  * Memory registration page/sg lists must not exceed max_num_sg,
2341  * also the integrity page/sg lists must not exceed max_num_meta_sg.
2342  *
2343  */
2344 struct ib_mr *ib_alloc_mr_integrity(struct ib_pd *pd,
2345 				    u32 max_num_data_sg,
2346 				    u32 max_num_meta_sg)
2347 {
2348 	struct ib_mr *mr;
2349 	struct ib_sig_attrs *sig_attrs;
2350 
2351 	if (!pd->device->ops.alloc_mr_integrity ||
2352 	    !pd->device->ops.map_mr_sg_pi) {
2353 		mr = ERR_PTR(-EOPNOTSUPP);
2354 		goto out;
2355 	}
2356 
2357 	if (!max_num_meta_sg) {
2358 		mr = ERR_PTR(-EINVAL);
2359 		goto out;
2360 	}
2361 
2362 	sig_attrs = kzalloc(sizeof(struct ib_sig_attrs), GFP_KERNEL);
2363 	if (!sig_attrs) {
2364 		mr = ERR_PTR(-ENOMEM);
2365 		goto out;
2366 	}
2367 
2368 	mr = pd->device->ops.alloc_mr_integrity(pd, max_num_data_sg,
2369 						max_num_meta_sg);
2370 	if (IS_ERR(mr)) {
2371 		kfree(sig_attrs);
2372 		goto out;
2373 	}
2374 
2375 	mr->device = pd->device;
2376 	mr->pd = pd;
2377 	mr->dm = NULL;
2378 	mr->uobject = NULL;
2379 	atomic_inc(&pd->usecnt);
2380 	mr->need_inval = false;
2381 	mr->type = IB_MR_TYPE_INTEGRITY;
2382 	mr->sig_attrs = sig_attrs;
2383 
2384 	rdma_restrack_new(&mr->res, RDMA_RESTRACK_MR);
2385 	rdma_restrack_parent_name(&mr->res, &pd->res);
2386 	rdma_restrack_add(&mr->res);
2387 out:
2388 	trace_mr_integ_alloc(pd, max_num_data_sg, max_num_meta_sg, mr);
2389 	return mr;
2390 }
2391 EXPORT_SYMBOL(ib_alloc_mr_integrity);
2392 
2393 /* Multicast groups */
2394 
2395 static bool is_valid_mcast_lid(struct ib_qp *qp, u16 lid)
2396 {
2397 	struct ib_qp_init_attr init_attr = {};
2398 	struct ib_qp_attr attr = {};
2399 	int num_eth_ports = 0;
2400 	unsigned int port;
2401 
2402 	/* If QP state >= init, it is assigned to a port and we can check this
2403 	 * port only.
2404 	 */
2405 	if (!ib_query_qp(qp, &attr, IB_QP_STATE | IB_QP_PORT, &init_attr)) {
2406 		if (attr.qp_state >= IB_QPS_INIT) {
2407 			if (rdma_port_get_link_layer(qp->device, attr.port_num) !=
2408 			    IB_LINK_LAYER_INFINIBAND)
2409 				return true;
2410 			goto lid_check;
2411 		}
2412 	}
2413 
2414 	/* Can't get a quick answer, iterate over all ports */
2415 	rdma_for_each_port(qp->device, port)
2416 		if (rdma_port_get_link_layer(qp->device, port) !=
2417 		    IB_LINK_LAYER_INFINIBAND)
2418 			num_eth_ports++;
2419 
2420 	/* If we have at lease one Ethernet port, RoCE annex declares that
2421 	 * multicast LID should be ignored. We can't tell at this step if the
2422 	 * QP belongs to an IB or Ethernet port.
2423 	 */
2424 	if (num_eth_ports)
2425 		return true;
2426 
2427 	/* If all the ports are IB, we can check according to IB spec. */
2428 lid_check:
2429 	return !(lid < be16_to_cpu(IB_MULTICAST_LID_BASE) ||
2430 		 lid == be16_to_cpu(IB_LID_PERMISSIVE));
2431 }
2432 
2433 int ib_attach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid)
2434 {
2435 	int ret;
2436 
2437 	if (!qp->device->ops.attach_mcast)
2438 		return -EOPNOTSUPP;
2439 
2440 	if (!rdma_is_multicast_addr((struct in6_addr *)gid->raw) ||
2441 	    qp->qp_type != IB_QPT_UD || !is_valid_mcast_lid(qp, lid))
2442 		return -EINVAL;
2443 
2444 	ret = qp->device->ops.attach_mcast(qp, gid, lid);
2445 	if (!ret)
2446 		atomic_inc(&qp->usecnt);
2447 	return ret;
2448 }
2449 EXPORT_SYMBOL(ib_attach_mcast);
2450 
2451 int ib_detach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid)
2452 {
2453 	int ret;
2454 
2455 	if (!qp->device->ops.detach_mcast)
2456 		return -EOPNOTSUPP;
2457 
2458 	if (!rdma_is_multicast_addr((struct in6_addr *)gid->raw) ||
2459 	    qp->qp_type != IB_QPT_UD || !is_valid_mcast_lid(qp, lid))
2460 		return -EINVAL;
2461 
2462 	ret = qp->device->ops.detach_mcast(qp, gid, lid);
2463 	if (!ret)
2464 		atomic_dec(&qp->usecnt);
2465 	return ret;
2466 }
2467 EXPORT_SYMBOL(ib_detach_mcast);
2468 
2469 /**
2470  * ib_alloc_xrcd_user - Allocates an XRC domain.
2471  * @device: The device on which to allocate the XRC domain.
2472  * @inode: inode to connect XRCD
2473  * @udata: Valid user data or NULL for kernel object
2474  */
2475 struct ib_xrcd *ib_alloc_xrcd_user(struct ib_device *device,
2476 				   struct inode *inode, struct ib_udata *udata)
2477 {
2478 	struct ib_xrcd *xrcd;
2479 	int ret;
2480 
2481 	if (!device->ops.alloc_xrcd)
2482 		return ERR_PTR(-EOPNOTSUPP);
2483 
2484 	xrcd = rdma_zalloc_drv_obj(device, ib_xrcd);
2485 	if (!xrcd)
2486 		return ERR_PTR(-ENOMEM);
2487 
2488 	xrcd->device = device;
2489 	xrcd->inode = inode;
2490 	atomic_set(&xrcd->usecnt, 0);
2491 	init_rwsem(&xrcd->tgt_qps_rwsem);
2492 	xa_init(&xrcd->tgt_qps);
2493 
2494 	ret = device->ops.alloc_xrcd(xrcd, udata);
2495 	if (ret)
2496 		goto err;
2497 	return xrcd;
2498 err:
2499 	kfree(xrcd);
2500 	return ERR_PTR(ret);
2501 }
2502 EXPORT_SYMBOL(ib_alloc_xrcd_user);
2503 
2504 /**
2505  * ib_dealloc_xrcd_user - Deallocates an XRC domain.
2506  * @xrcd: The XRC domain to deallocate.
2507  * @udata: Valid user data or NULL for kernel object
2508  */
2509 int ib_dealloc_xrcd_user(struct ib_xrcd *xrcd, struct ib_udata *udata)
2510 {
2511 	int ret;
2512 
2513 	if (atomic_read(&xrcd->usecnt))
2514 		return -EBUSY;
2515 
2516 	WARN_ON(!xa_empty(&xrcd->tgt_qps));
2517 	ret = xrcd->device->ops.dealloc_xrcd(xrcd, udata);
2518 	if (ret)
2519 		return ret;
2520 	kfree(xrcd);
2521 	return ret;
2522 }
2523 EXPORT_SYMBOL(ib_dealloc_xrcd_user);
2524 
2525 /**
2526  * ib_create_wq - Creates a WQ associated with the specified protection
2527  * domain.
2528  * @pd: The protection domain associated with the WQ.
2529  * @wq_attr: A list of initial attributes required to create the
2530  * WQ. If WQ creation succeeds, then the attributes are updated to
2531  * the actual capabilities of the created WQ.
2532  *
2533  * wq_attr->max_wr and wq_attr->max_sge determine
2534  * the requested size of the WQ, and set to the actual values allocated
2535  * on return.
2536  * If ib_create_wq() succeeds, then max_wr and max_sge will always be
2537  * at least as large as the requested values.
2538  */
2539 struct ib_wq *ib_create_wq(struct ib_pd *pd,
2540 			   struct ib_wq_init_attr *wq_attr)
2541 {
2542 	struct ib_wq *wq;
2543 
2544 	if (!pd->device->ops.create_wq)
2545 		return ERR_PTR(-EOPNOTSUPP);
2546 
2547 	wq = pd->device->ops.create_wq(pd, wq_attr, NULL);
2548 	if (!IS_ERR(wq)) {
2549 		wq->event_handler = wq_attr->event_handler;
2550 		wq->wq_context = wq_attr->wq_context;
2551 		wq->wq_type = wq_attr->wq_type;
2552 		wq->cq = wq_attr->cq;
2553 		wq->device = pd->device;
2554 		wq->pd = pd;
2555 		wq->uobject = NULL;
2556 		atomic_inc(&pd->usecnt);
2557 		atomic_inc(&wq_attr->cq->usecnt);
2558 		atomic_set(&wq->usecnt, 0);
2559 	}
2560 	return wq;
2561 }
2562 EXPORT_SYMBOL(ib_create_wq);
2563 
2564 /**
2565  * ib_destroy_wq_user - Destroys the specified user WQ.
2566  * @wq: The WQ to destroy.
2567  * @udata: Valid user data
2568  */
2569 int ib_destroy_wq_user(struct ib_wq *wq, struct ib_udata *udata)
2570 {
2571 	struct ib_cq *cq = wq->cq;
2572 	struct ib_pd *pd = wq->pd;
2573 	int ret;
2574 
2575 	if (atomic_read(&wq->usecnt))
2576 		return -EBUSY;
2577 
2578 	ret = wq->device->ops.destroy_wq(wq, udata);
2579 	if (ret)
2580 		return ret;
2581 
2582 	atomic_dec(&pd->usecnt);
2583 	atomic_dec(&cq->usecnt);
2584 	return ret;
2585 }
2586 EXPORT_SYMBOL(ib_destroy_wq_user);
2587 
2588 int ib_check_mr_status(struct ib_mr *mr, u32 check_mask,
2589 		       struct ib_mr_status *mr_status)
2590 {
2591 	if (!mr->device->ops.check_mr_status)
2592 		return -EOPNOTSUPP;
2593 
2594 	return mr->device->ops.check_mr_status(mr, check_mask, mr_status);
2595 }
2596 EXPORT_SYMBOL(ib_check_mr_status);
2597 
2598 int ib_set_vf_link_state(struct ib_device *device, int vf, u32 port,
2599 			 int state)
2600 {
2601 	if (!device->ops.set_vf_link_state)
2602 		return -EOPNOTSUPP;
2603 
2604 	return device->ops.set_vf_link_state(device, vf, port, state);
2605 }
2606 EXPORT_SYMBOL(ib_set_vf_link_state);
2607 
2608 int ib_get_vf_config(struct ib_device *device, int vf, u32 port,
2609 		     struct ifla_vf_info *info)
2610 {
2611 	if (!device->ops.get_vf_config)
2612 		return -EOPNOTSUPP;
2613 
2614 	return device->ops.get_vf_config(device, vf, port, info);
2615 }
2616 EXPORT_SYMBOL(ib_get_vf_config);
2617 
2618 int ib_get_vf_stats(struct ib_device *device, int vf, u32 port,
2619 		    struct ifla_vf_stats *stats)
2620 {
2621 	if (!device->ops.get_vf_stats)
2622 		return -EOPNOTSUPP;
2623 
2624 	return device->ops.get_vf_stats(device, vf, port, stats);
2625 }
2626 EXPORT_SYMBOL(ib_get_vf_stats);
2627 
2628 int ib_set_vf_guid(struct ib_device *device, int vf, u32 port, u64 guid,
2629 		   int type)
2630 {
2631 	if (!device->ops.set_vf_guid)
2632 		return -EOPNOTSUPP;
2633 
2634 	return device->ops.set_vf_guid(device, vf, port, guid, type);
2635 }
2636 EXPORT_SYMBOL(ib_set_vf_guid);
2637 
2638 int ib_get_vf_guid(struct ib_device *device, int vf, u32 port,
2639 		   struct ifla_vf_guid *node_guid,
2640 		   struct ifla_vf_guid *port_guid)
2641 {
2642 	if (!device->ops.get_vf_guid)
2643 		return -EOPNOTSUPP;
2644 
2645 	return device->ops.get_vf_guid(device, vf, port, node_guid, port_guid);
2646 }
2647 EXPORT_SYMBOL(ib_get_vf_guid);
2648 /**
2649  * ib_map_mr_sg_pi() - Map the dma mapped SG lists for PI (protection
2650  *     information) and set an appropriate memory region for registration.
2651  * @mr:             memory region
2652  * @data_sg:        dma mapped scatterlist for data
2653  * @data_sg_nents:  number of entries in data_sg
2654  * @data_sg_offset: offset in bytes into data_sg
2655  * @meta_sg:        dma mapped scatterlist for metadata
2656  * @meta_sg_nents:  number of entries in meta_sg
2657  * @meta_sg_offset: offset in bytes into meta_sg
2658  * @page_size:      page vector desired page size
2659  *
2660  * Constraints:
2661  * - The MR must be allocated with type IB_MR_TYPE_INTEGRITY.
2662  *
2663  * Return: 0 on success.
2664  *
2665  * After this completes successfully, the  memory region
2666  * is ready for registration.
2667  */
2668 int ib_map_mr_sg_pi(struct ib_mr *mr, struct scatterlist *data_sg,
2669 		    int data_sg_nents, unsigned int *data_sg_offset,
2670 		    struct scatterlist *meta_sg, int meta_sg_nents,
2671 		    unsigned int *meta_sg_offset, unsigned int page_size)
2672 {
2673 	if (unlikely(!mr->device->ops.map_mr_sg_pi ||
2674 		     WARN_ON_ONCE(mr->type != IB_MR_TYPE_INTEGRITY)))
2675 		return -EOPNOTSUPP;
2676 
2677 	mr->page_size = page_size;
2678 
2679 	return mr->device->ops.map_mr_sg_pi(mr, data_sg, data_sg_nents,
2680 					    data_sg_offset, meta_sg,
2681 					    meta_sg_nents, meta_sg_offset);
2682 }
2683 EXPORT_SYMBOL(ib_map_mr_sg_pi);
2684 
2685 /**
2686  * ib_map_mr_sg() - Map the largest prefix of a dma mapped SG list
2687  *     and set it the memory region.
2688  * @mr:            memory region
2689  * @sg:            dma mapped scatterlist
2690  * @sg_nents:      number of entries in sg
2691  * @sg_offset:     offset in bytes into sg
2692  * @page_size:     page vector desired page size
2693  *
2694  * Constraints:
2695  *
2696  * - The first sg element is allowed to have an offset.
2697  * - Each sg element must either be aligned to page_size or virtually
2698  *   contiguous to the previous element. In case an sg element has a
2699  *   non-contiguous offset, the mapping prefix will not include it.
2700  * - The last sg element is allowed to have length less than page_size.
2701  * - If sg_nents total byte length exceeds the mr max_num_sge * page_size
2702  *   then only max_num_sg entries will be mapped.
2703  * - If the MR was allocated with type IB_MR_TYPE_SG_GAPS, none of these
2704  *   constraints holds and the page_size argument is ignored.
2705  *
2706  * Returns the number of sg elements that were mapped to the memory region.
2707  *
2708  * After this completes successfully, the  memory region
2709  * is ready for registration.
2710  */
2711 int ib_map_mr_sg(struct ib_mr *mr, struct scatterlist *sg, int sg_nents,
2712 		 unsigned int *sg_offset, unsigned int page_size)
2713 {
2714 	if (unlikely(!mr->device->ops.map_mr_sg))
2715 		return -EOPNOTSUPP;
2716 
2717 	mr->page_size = page_size;
2718 
2719 	return mr->device->ops.map_mr_sg(mr, sg, sg_nents, sg_offset);
2720 }
2721 EXPORT_SYMBOL(ib_map_mr_sg);
2722 
2723 /**
2724  * ib_sg_to_pages() - Convert the largest prefix of a sg list
2725  *     to a page vector
2726  * @mr:            memory region
2727  * @sgl:           dma mapped scatterlist
2728  * @sg_nents:      number of entries in sg
2729  * @sg_offset_p:   ==== =======================================================
2730  *                 IN   start offset in bytes into sg
2731  *                 OUT  offset in bytes for element n of the sg of the first
2732  *                      byte that has not been processed where n is the return
2733  *                      value of this function.
2734  *                 ==== =======================================================
2735  * @set_page:      driver page assignment function pointer
2736  *
2737  * Core service helper for drivers to convert the largest
2738  * prefix of given sg list to a page vector. The sg list
2739  * prefix converted is the prefix that meet the requirements
2740  * of ib_map_mr_sg.
2741  *
2742  * Returns the number of sg elements that were assigned to
2743  * a page vector.
2744  */
2745 int ib_sg_to_pages(struct ib_mr *mr, struct scatterlist *sgl, int sg_nents,
2746 		unsigned int *sg_offset_p, int (*set_page)(struct ib_mr *, u64))
2747 {
2748 	struct scatterlist *sg;
2749 	u64 last_end_dma_addr = 0;
2750 	unsigned int sg_offset = sg_offset_p ? *sg_offset_p : 0;
2751 	unsigned int last_page_off = 0;
2752 	u64 page_mask = ~((u64)mr->page_size - 1);
2753 	int i, ret;
2754 
2755 	if (unlikely(sg_nents <= 0 || sg_offset > sg_dma_len(&sgl[0])))
2756 		return -EINVAL;
2757 
2758 	mr->iova = sg_dma_address(&sgl[0]) + sg_offset;
2759 	mr->length = 0;
2760 
2761 	for_each_sg(sgl, sg, sg_nents, i) {
2762 		u64 dma_addr = sg_dma_address(sg) + sg_offset;
2763 		u64 prev_addr = dma_addr;
2764 		unsigned int dma_len = sg_dma_len(sg) - sg_offset;
2765 		u64 end_dma_addr = dma_addr + dma_len;
2766 		u64 page_addr = dma_addr & page_mask;
2767 
2768 		/*
2769 		 * For the second and later elements, check whether either the
2770 		 * end of element i-1 or the start of element i is not aligned
2771 		 * on a page boundary.
2772 		 */
2773 		if (i && (last_page_off != 0 || page_addr != dma_addr)) {
2774 			/* Stop mapping if there is a gap. */
2775 			if (last_end_dma_addr != dma_addr)
2776 				break;
2777 
2778 			/*
2779 			 * Coalesce this element with the last. If it is small
2780 			 * enough just update mr->length. Otherwise start
2781 			 * mapping from the next page.
2782 			 */
2783 			goto next_page;
2784 		}
2785 
2786 		do {
2787 			ret = set_page(mr, page_addr);
2788 			if (unlikely(ret < 0)) {
2789 				sg_offset = prev_addr - sg_dma_address(sg);
2790 				mr->length += prev_addr - dma_addr;
2791 				if (sg_offset_p)
2792 					*sg_offset_p = sg_offset;
2793 				return i || sg_offset ? i : ret;
2794 			}
2795 			prev_addr = page_addr;
2796 next_page:
2797 			page_addr += mr->page_size;
2798 		} while (page_addr < end_dma_addr);
2799 
2800 		mr->length += dma_len;
2801 		last_end_dma_addr = end_dma_addr;
2802 		last_page_off = end_dma_addr & ~page_mask;
2803 
2804 		sg_offset = 0;
2805 	}
2806 
2807 	if (sg_offset_p)
2808 		*sg_offset_p = 0;
2809 	return i;
2810 }
2811 EXPORT_SYMBOL(ib_sg_to_pages);
2812 
2813 struct ib_drain_cqe {
2814 	struct ib_cqe cqe;
2815 	struct completion done;
2816 };
2817 
2818 static void ib_drain_qp_done(struct ib_cq *cq, struct ib_wc *wc)
2819 {
2820 	struct ib_drain_cqe *cqe = container_of(wc->wr_cqe, struct ib_drain_cqe,
2821 						cqe);
2822 
2823 	complete(&cqe->done);
2824 }
2825 
2826 /*
2827  * Post a WR and block until its completion is reaped for the SQ.
2828  */
2829 static void __ib_drain_sq(struct ib_qp *qp)
2830 {
2831 	struct ib_cq *cq = qp->send_cq;
2832 	struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
2833 	struct ib_drain_cqe sdrain;
2834 	struct ib_rdma_wr swr = {
2835 		.wr = {
2836 			.next = NULL,
2837 			{ .wr_cqe	= &sdrain.cqe, },
2838 			.opcode	= IB_WR_RDMA_WRITE,
2839 		},
2840 	};
2841 	int ret;
2842 
2843 	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
2844 	if (ret) {
2845 		WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
2846 		return;
2847 	}
2848 
2849 	sdrain.cqe.done = ib_drain_qp_done;
2850 	init_completion(&sdrain.done);
2851 
2852 	ret = ib_post_send(qp, &swr.wr, NULL);
2853 	if (ret) {
2854 		WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
2855 		return;
2856 	}
2857 
2858 	if (cq->poll_ctx == IB_POLL_DIRECT)
2859 		while (wait_for_completion_timeout(&sdrain.done, HZ / 10) <= 0)
2860 			ib_process_cq_direct(cq, -1);
2861 	else
2862 		wait_for_completion(&sdrain.done);
2863 }
2864 
2865 /*
2866  * Post a WR and block until its completion is reaped for the RQ.
2867  */
2868 static void __ib_drain_rq(struct ib_qp *qp)
2869 {
2870 	struct ib_cq *cq = qp->recv_cq;
2871 	struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
2872 	struct ib_drain_cqe rdrain;
2873 	struct ib_recv_wr rwr = {};
2874 	int ret;
2875 
2876 	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
2877 	if (ret) {
2878 		WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
2879 		return;
2880 	}
2881 
2882 	rwr.wr_cqe = &rdrain.cqe;
2883 	rdrain.cqe.done = ib_drain_qp_done;
2884 	init_completion(&rdrain.done);
2885 
2886 	ret = ib_post_recv(qp, &rwr, NULL);
2887 	if (ret) {
2888 		WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
2889 		return;
2890 	}
2891 
2892 	if (cq->poll_ctx == IB_POLL_DIRECT)
2893 		while (wait_for_completion_timeout(&rdrain.done, HZ / 10) <= 0)
2894 			ib_process_cq_direct(cq, -1);
2895 	else
2896 		wait_for_completion(&rdrain.done);
2897 }
2898 
2899 /*
2900  * __ib_drain_srq() - Block until Last WQE Reached event arrives, or timeout
2901  *                    expires.
2902  * @qp:               queue pair associated with SRQ to drain
2903  *
2904  * Quoting 10.3.1 Queue Pair and EE Context States:
2905  *
2906  * Note, for QPs that are associated with an SRQ, the Consumer should take the
2907  * QP through the Error State before invoking a Destroy QP or a Modify QP to the
2908  * Reset State.  The Consumer may invoke the Destroy QP without first performing
2909  * a Modify QP to the Error State and waiting for the Affiliated Asynchronous
2910  * Last WQE Reached Event. However, if the Consumer does not wait for the
2911  * Affiliated Asynchronous Last WQE Reached Event, then WQE and Data Segment
2912  * leakage may occur. Therefore, it is good programming practice to tear down a
2913  * QP that is associated with an SRQ by using the following process:
2914  *
2915  * - Put the QP in the Error State
2916  * - Wait for the Affiliated Asynchronous Last WQE Reached Event;
2917  * - either:
2918  *       drain the CQ by invoking the Poll CQ verb and either wait for CQ
2919  *       to be empty or the number of Poll CQ operations has exceeded
2920  *       CQ capacity size;
2921  * - or
2922  *       post another WR that completes on the same CQ and wait for this
2923  *       WR to return as a WC;
2924  * - and then invoke a Destroy QP or Reset QP.
2925  *
2926  * We use the first option.
2927  */
2928 static void __ib_drain_srq(struct ib_qp *qp)
2929 {
2930 	struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
2931 	struct ib_cq *cq;
2932 	int n, polled = 0;
2933 	int ret;
2934 
2935 	if (!qp->srq) {
2936 		WARN_ONCE(1, "QP 0x%p is not associated with SRQ\n", qp);
2937 		return;
2938 	}
2939 
2940 	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
2941 	if (ret) {
2942 		WARN_ONCE(ret, "failed to drain shared recv queue: %d\n", ret);
2943 		return;
2944 	}
2945 
2946 	if (ib_srq_has_cq(qp->srq->srq_type)) {
2947 		cq = qp->srq->ext.cq;
2948 	} else if (qp->recv_cq) {
2949 		cq = qp->recv_cq;
2950 	} else {
2951 		WARN_ONCE(1, "QP 0x%p has no CQ associated with SRQ\n", qp);
2952 		return;
2953 	}
2954 
2955 	if (wait_for_completion_timeout(&qp->srq_completion, 60 * HZ) > 0) {
2956 		while (polled != cq->cqe) {
2957 			n = ib_process_cq_direct(cq, cq->cqe - polled);
2958 			if (!n)
2959 				return;
2960 			polled += n;
2961 		}
2962 	}
2963 }
2964 
2965 /**
2966  * ib_drain_sq() - Block until all SQ CQEs have been consumed by the
2967  *		   application.
2968  * @qp:            queue pair to drain
2969  *
2970  * If the device has a provider-specific drain function, then
2971  * call that.  Otherwise call the generic drain function
2972  * __ib_drain_sq().
2973  *
2974  * The caller must:
2975  *
2976  * ensure there is room in the CQ and SQ for the drain work request and
2977  * completion.
2978  *
2979  * allocate the CQ using ib_alloc_cq().
2980  *
2981  * ensure that there are no other contexts that are posting WRs concurrently.
2982  * Otherwise the drain is not guaranteed.
2983  */
2984 void ib_drain_sq(struct ib_qp *qp)
2985 {
2986 	if (qp->device->ops.drain_sq)
2987 		qp->device->ops.drain_sq(qp);
2988 	else
2989 		__ib_drain_sq(qp);
2990 	trace_cq_drain_complete(qp->send_cq);
2991 }
2992 EXPORT_SYMBOL(ib_drain_sq);
2993 
2994 /**
2995  * ib_drain_rq() - Block until all RQ CQEs have been consumed by the
2996  *		   application.
2997  * @qp:            queue pair to drain
2998  *
2999  * If the device has a provider-specific drain function, then
3000  * call that.  Otherwise call the generic drain function
3001  * __ib_drain_rq().
3002  *
3003  * The caller must:
3004  *
3005  * ensure there is room in the CQ and RQ for the drain work request and
3006  * completion.
3007  *
3008  * allocate the CQ using ib_alloc_cq().
3009  *
3010  * ensure that there are no other contexts that are posting WRs concurrently.
3011  * Otherwise the drain is not guaranteed.
3012  */
3013 void ib_drain_rq(struct ib_qp *qp)
3014 {
3015 	if (qp->device->ops.drain_rq)
3016 		qp->device->ops.drain_rq(qp);
3017 	else
3018 		__ib_drain_rq(qp);
3019 	trace_cq_drain_complete(qp->recv_cq);
3020 }
3021 EXPORT_SYMBOL(ib_drain_rq);
3022 
3023 /**
3024  * ib_drain_qp() - Block until all CQEs have been consumed by the
3025  *		   application on both the RQ and SQ.
3026  * @qp:            queue pair to drain
3027  *
3028  * The caller must:
3029  *
3030  * ensure there is room in the CQ(s), SQ, and RQ for drain work requests
3031  * and completions.
3032  *
3033  * allocate the CQs using ib_alloc_cq().
3034  *
3035  * ensure that there are no other contexts that are posting WRs concurrently.
3036  * Otherwise the drain is not guaranteed.
3037  */
3038 void ib_drain_qp(struct ib_qp *qp)
3039 {
3040 	ib_drain_sq(qp);
3041 	if (!qp->srq)
3042 		ib_drain_rq(qp);
3043 	else
3044 		__ib_drain_srq(qp);
3045 }
3046 EXPORT_SYMBOL(ib_drain_qp);
3047 
3048 struct net_device *rdma_alloc_netdev(struct ib_device *device, u32 port_num,
3049 				     enum rdma_netdev_t type, const char *name,
3050 				     unsigned char name_assign_type,
3051 				     void (*setup)(struct net_device *))
3052 {
3053 	struct rdma_netdev_alloc_params params;
3054 	struct net_device *netdev;
3055 	int rc;
3056 
3057 	if (!device->ops.rdma_netdev_get_params)
3058 		return ERR_PTR(-EOPNOTSUPP);
3059 
3060 	rc = device->ops.rdma_netdev_get_params(device, port_num, type,
3061 						&params);
3062 	if (rc)
3063 		return ERR_PTR(rc);
3064 
3065 	netdev = alloc_netdev_mqs(params.sizeof_priv, name, name_assign_type,
3066 				  setup, params.txqs, params.rxqs);
3067 	if (!netdev)
3068 		return ERR_PTR(-ENOMEM);
3069 
3070 	return netdev;
3071 }
3072 EXPORT_SYMBOL(rdma_alloc_netdev);
3073 
3074 int rdma_init_netdev(struct ib_device *device, u32 port_num,
3075 		     enum rdma_netdev_t type, const char *name,
3076 		     unsigned char name_assign_type,
3077 		     void (*setup)(struct net_device *),
3078 		     struct net_device *netdev)
3079 {
3080 	struct rdma_netdev_alloc_params params;
3081 	int rc;
3082 
3083 	if (!device->ops.rdma_netdev_get_params)
3084 		return -EOPNOTSUPP;
3085 
3086 	rc = device->ops.rdma_netdev_get_params(device, port_num, type,
3087 						&params);
3088 	if (rc)
3089 		return rc;
3090 
3091 	return params.initialize_rdma_netdev(device, port_num,
3092 					     netdev, params.param);
3093 }
3094 EXPORT_SYMBOL(rdma_init_netdev);
3095 
3096 void __rdma_block_iter_start(struct ib_block_iter *biter,
3097 			     struct scatterlist *sglist, unsigned int nents,
3098 			     unsigned long pgsz)
3099 {
3100 	memset(biter, 0, sizeof(struct ib_block_iter));
3101 	biter->__sg = sglist;
3102 	biter->__sg_nents = nents;
3103 
3104 	/* Driver provides best block size to use */
3105 	biter->__pg_bit = __fls(pgsz);
3106 }
3107 EXPORT_SYMBOL(__rdma_block_iter_start);
3108 
3109 bool __rdma_block_iter_next(struct ib_block_iter *biter)
3110 {
3111 	unsigned int block_offset;
3112 	unsigned int sg_delta;
3113 
3114 	if (!biter->__sg_nents || !biter->__sg)
3115 		return false;
3116 
3117 	biter->__dma_addr = sg_dma_address(biter->__sg) + biter->__sg_advance;
3118 	block_offset = biter->__dma_addr & (BIT_ULL(biter->__pg_bit) - 1);
3119 	sg_delta = BIT_ULL(biter->__pg_bit) - block_offset;
3120 
3121 	if (sg_dma_len(biter->__sg) - biter->__sg_advance > sg_delta) {
3122 		biter->__sg_advance += sg_delta;
3123 	} else {
3124 		biter->__sg_advance = 0;
3125 		biter->__sg = sg_next(biter->__sg);
3126 		biter->__sg_nents--;
3127 	}
3128 
3129 	return true;
3130 }
3131 EXPORT_SYMBOL(__rdma_block_iter_next);
3132 
3133 /**
3134  * rdma_alloc_hw_stats_struct - Helper function to allocate dynamic struct
3135  *   for the drivers.
3136  * @descs: array of static descriptors
3137  * @num_counters: number of elements in array
3138  * @lifespan: milliseconds between updates
3139  */
3140 struct rdma_hw_stats *rdma_alloc_hw_stats_struct(
3141 	const struct rdma_stat_desc *descs, int num_counters,
3142 	unsigned long lifespan)
3143 {
3144 	struct rdma_hw_stats *stats;
3145 
3146 	stats = kzalloc(struct_size(stats, value, num_counters), GFP_KERNEL);
3147 	if (!stats)
3148 		return NULL;
3149 
3150 	stats->is_disabled = kcalloc(BITS_TO_LONGS(num_counters),
3151 				     sizeof(*stats->is_disabled), GFP_KERNEL);
3152 	if (!stats->is_disabled)
3153 		goto err;
3154 
3155 	stats->descs = descs;
3156 	stats->num_counters = num_counters;
3157 	stats->lifespan = msecs_to_jiffies(lifespan);
3158 	mutex_init(&stats->lock);
3159 
3160 	return stats;
3161 
3162 err:
3163 	kfree(stats);
3164 	return NULL;
3165 }
3166 EXPORT_SYMBOL(rdma_alloc_hw_stats_struct);
3167 
3168 /**
3169  * rdma_free_hw_stats_struct - Helper function to release rdma_hw_stats
3170  * @stats: statistics to release
3171  */
3172 void rdma_free_hw_stats_struct(struct rdma_hw_stats *stats)
3173 {
3174 	if (!stats)
3175 		return;
3176 
3177 	kfree(stats->is_disabled);
3178 	kfree(stats);
3179 }
3180 EXPORT_SYMBOL(rdma_free_hw_stats_struct);
3181