xref: /openbmc/linux/drivers/infiniband/core/verbs.c (revision 57ee11ea)
1 /*
2  * Copyright (c) 2004 Mellanox Technologies Ltd.  All rights reserved.
3  * Copyright (c) 2004 Infinicon Corporation.  All rights reserved.
4  * Copyright (c) 2004 Intel Corporation.  All rights reserved.
5  * Copyright (c) 2004 Topspin Corporation.  All rights reserved.
6  * Copyright (c) 2004 Voltaire Corporation.  All rights reserved.
7  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
8  * Copyright (c) 2005, 2006 Cisco Systems.  All rights reserved.
9  *
10  * This software is available to you under a choice of one of two
11  * licenses.  You may choose to be licensed under the terms of the GNU
12  * General Public License (GPL) Version 2, available from the file
13  * COPYING in the main directory of this source tree, or the
14  * OpenIB.org BSD license below:
15  *
16  *     Redistribution and use in source and binary forms, with or
17  *     without modification, are permitted provided that the following
18  *     conditions are met:
19  *
20  *      - Redistributions of source code must retain the above
21  *        copyright notice, this list of conditions and the following
22  *        disclaimer.
23  *
24  *      - Redistributions in binary form must reproduce the above
25  *        copyright notice, this list of conditions and the following
26  *        disclaimer in the documentation and/or other materials
27  *        provided with the distribution.
28  *
29  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
30  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
31  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
32  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
33  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
34  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
35  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
36  * SOFTWARE.
37  */
38 
39 #include <linux/errno.h>
40 #include <linux/err.h>
41 #include <linux/export.h>
42 #include <linux/string.h>
43 #include <linux/slab.h>
44 #include <linux/in.h>
45 #include <linux/in6.h>
46 #include <net/addrconf.h>
47 #include <linux/security.h>
48 
49 #include <rdma/ib_verbs.h>
50 #include <rdma/ib_cache.h>
51 #include <rdma/ib_addr.h>
52 #include <rdma/rw.h>
53 
54 #include "core_priv.h"
55 
56 static const char * const ib_events[] = {
57 	[IB_EVENT_CQ_ERR]		= "CQ error",
58 	[IB_EVENT_QP_FATAL]		= "QP fatal error",
59 	[IB_EVENT_QP_REQ_ERR]		= "QP request error",
60 	[IB_EVENT_QP_ACCESS_ERR]	= "QP access error",
61 	[IB_EVENT_COMM_EST]		= "communication established",
62 	[IB_EVENT_SQ_DRAINED]		= "send queue drained",
63 	[IB_EVENT_PATH_MIG]		= "path migration successful",
64 	[IB_EVENT_PATH_MIG_ERR]		= "path migration error",
65 	[IB_EVENT_DEVICE_FATAL]		= "device fatal error",
66 	[IB_EVENT_PORT_ACTIVE]		= "port active",
67 	[IB_EVENT_PORT_ERR]		= "port error",
68 	[IB_EVENT_LID_CHANGE]		= "LID change",
69 	[IB_EVENT_PKEY_CHANGE]		= "P_key change",
70 	[IB_EVENT_SM_CHANGE]		= "SM change",
71 	[IB_EVENT_SRQ_ERR]		= "SRQ error",
72 	[IB_EVENT_SRQ_LIMIT_REACHED]	= "SRQ limit reached",
73 	[IB_EVENT_QP_LAST_WQE_REACHED]	= "last WQE reached",
74 	[IB_EVENT_CLIENT_REREGISTER]	= "client reregister",
75 	[IB_EVENT_GID_CHANGE]		= "GID changed",
76 };
77 
78 const char *__attribute_const__ ib_event_msg(enum ib_event_type event)
79 {
80 	size_t index = event;
81 
82 	return (index < ARRAY_SIZE(ib_events) && ib_events[index]) ?
83 			ib_events[index] : "unrecognized event";
84 }
85 EXPORT_SYMBOL(ib_event_msg);
86 
87 static const char * const wc_statuses[] = {
88 	[IB_WC_SUCCESS]			= "success",
89 	[IB_WC_LOC_LEN_ERR]		= "local length error",
90 	[IB_WC_LOC_QP_OP_ERR]		= "local QP operation error",
91 	[IB_WC_LOC_EEC_OP_ERR]		= "local EE context operation error",
92 	[IB_WC_LOC_PROT_ERR]		= "local protection error",
93 	[IB_WC_WR_FLUSH_ERR]		= "WR flushed",
94 	[IB_WC_MW_BIND_ERR]		= "memory management operation error",
95 	[IB_WC_BAD_RESP_ERR]		= "bad response error",
96 	[IB_WC_LOC_ACCESS_ERR]		= "local access error",
97 	[IB_WC_REM_INV_REQ_ERR]		= "invalid request error",
98 	[IB_WC_REM_ACCESS_ERR]		= "remote access error",
99 	[IB_WC_REM_OP_ERR]		= "remote operation error",
100 	[IB_WC_RETRY_EXC_ERR]		= "transport retry counter exceeded",
101 	[IB_WC_RNR_RETRY_EXC_ERR]	= "RNR retry counter exceeded",
102 	[IB_WC_LOC_RDD_VIOL_ERR]	= "local RDD violation error",
103 	[IB_WC_REM_INV_RD_REQ_ERR]	= "remote invalid RD request",
104 	[IB_WC_REM_ABORT_ERR]		= "operation aborted",
105 	[IB_WC_INV_EECN_ERR]		= "invalid EE context number",
106 	[IB_WC_INV_EEC_STATE_ERR]	= "invalid EE context state",
107 	[IB_WC_FATAL_ERR]		= "fatal error",
108 	[IB_WC_RESP_TIMEOUT_ERR]	= "response timeout error",
109 	[IB_WC_GENERAL_ERR]		= "general error",
110 };
111 
112 const char *__attribute_const__ ib_wc_status_msg(enum ib_wc_status status)
113 {
114 	size_t index = status;
115 
116 	return (index < ARRAY_SIZE(wc_statuses) && wc_statuses[index]) ?
117 			wc_statuses[index] : "unrecognized status";
118 }
119 EXPORT_SYMBOL(ib_wc_status_msg);
120 
121 __attribute_const__ int ib_rate_to_mult(enum ib_rate rate)
122 {
123 	switch (rate) {
124 	case IB_RATE_2_5_GBPS: return  1;
125 	case IB_RATE_5_GBPS:   return  2;
126 	case IB_RATE_10_GBPS:  return  4;
127 	case IB_RATE_20_GBPS:  return  8;
128 	case IB_RATE_30_GBPS:  return 12;
129 	case IB_RATE_40_GBPS:  return 16;
130 	case IB_RATE_60_GBPS:  return 24;
131 	case IB_RATE_80_GBPS:  return 32;
132 	case IB_RATE_120_GBPS: return 48;
133 	default:	       return -1;
134 	}
135 }
136 EXPORT_SYMBOL(ib_rate_to_mult);
137 
138 __attribute_const__ enum ib_rate mult_to_ib_rate(int mult)
139 {
140 	switch (mult) {
141 	case 1:  return IB_RATE_2_5_GBPS;
142 	case 2:  return IB_RATE_5_GBPS;
143 	case 4:  return IB_RATE_10_GBPS;
144 	case 8:  return IB_RATE_20_GBPS;
145 	case 12: return IB_RATE_30_GBPS;
146 	case 16: return IB_RATE_40_GBPS;
147 	case 24: return IB_RATE_60_GBPS;
148 	case 32: return IB_RATE_80_GBPS;
149 	case 48: return IB_RATE_120_GBPS;
150 	default: return IB_RATE_PORT_CURRENT;
151 	}
152 }
153 EXPORT_SYMBOL(mult_to_ib_rate);
154 
155 __attribute_const__ int ib_rate_to_mbps(enum ib_rate rate)
156 {
157 	switch (rate) {
158 	case IB_RATE_2_5_GBPS: return 2500;
159 	case IB_RATE_5_GBPS:   return 5000;
160 	case IB_RATE_10_GBPS:  return 10000;
161 	case IB_RATE_20_GBPS:  return 20000;
162 	case IB_RATE_30_GBPS:  return 30000;
163 	case IB_RATE_40_GBPS:  return 40000;
164 	case IB_RATE_60_GBPS:  return 60000;
165 	case IB_RATE_80_GBPS:  return 80000;
166 	case IB_RATE_120_GBPS: return 120000;
167 	case IB_RATE_14_GBPS:  return 14062;
168 	case IB_RATE_56_GBPS:  return 56250;
169 	case IB_RATE_112_GBPS: return 112500;
170 	case IB_RATE_168_GBPS: return 168750;
171 	case IB_RATE_25_GBPS:  return 25781;
172 	case IB_RATE_100_GBPS: return 103125;
173 	case IB_RATE_200_GBPS: return 206250;
174 	case IB_RATE_300_GBPS: return 309375;
175 	default:	       return -1;
176 	}
177 }
178 EXPORT_SYMBOL(ib_rate_to_mbps);
179 
180 __attribute_const__ enum rdma_transport_type
181 rdma_node_get_transport(enum rdma_node_type node_type)
182 {
183 	switch (node_type) {
184 	case RDMA_NODE_IB_CA:
185 	case RDMA_NODE_IB_SWITCH:
186 	case RDMA_NODE_IB_ROUTER:
187 		return RDMA_TRANSPORT_IB;
188 	case RDMA_NODE_RNIC:
189 		return RDMA_TRANSPORT_IWARP;
190 	case RDMA_NODE_USNIC:
191 		return RDMA_TRANSPORT_USNIC;
192 	case RDMA_NODE_USNIC_UDP:
193 		return RDMA_TRANSPORT_USNIC_UDP;
194 	default:
195 		BUG();
196 		return 0;
197 	}
198 }
199 EXPORT_SYMBOL(rdma_node_get_transport);
200 
201 enum rdma_link_layer rdma_port_get_link_layer(struct ib_device *device, u8 port_num)
202 {
203 	if (device->get_link_layer)
204 		return device->get_link_layer(device, port_num);
205 
206 	switch (rdma_node_get_transport(device->node_type)) {
207 	case RDMA_TRANSPORT_IB:
208 		return IB_LINK_LAYER_INFINIBAND;
209 	case RDMA_TRANSPORT_IWARP:
210 	case RDMA_TRANSPORT_USNIC:
211 	case RDMA_TRANSPORT_USNIC_UDP:
212 		return IB_LINK_LAYER_ETHERNET;
213 	default:
214 		return IB_LINK_LAYER_UNSPECIFIED;
215 	}
216 }
217 EXPORT_SYMBOL(rdma_port_get_link_layer);
218 
219 /* Protection domains */
220 
221 /**
222  * ib_alloc_pd - Allocates an unused protection domain.
223  * @device: The device on which to allocate the protection domain.
224  *
225  * A protection domain object provides an association between QPs, shared
226  * receive queues, address handles, memory regions, and memory windows.
227  *
228  * Every PD has a local_dma_lkey which can be used as the lkey value for local
229  * memory operations.
230  */
231 struct ib_pd *__ib_alloc_pd(struct ib_device *device, unsigned int flags,
232 		const char *caller)
233 {
234 	struct ib_pd *pd;
235 	int mr_access_flags = 0;
236 
237 	pd = device->alloc_pd(device, NULL, NULL);
238 	if (IS_ERR(pd))
239 		return pd;
240 
241 	pd->device = device;
242 	pd->uobject = NULL;
243 	pd->__internal_mr = NULL;
244 	atomic_set(&pd->usecnt, 0);
245 	pd->flags = flags;
246 
247 	if (device->attrs.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY)
248 		pd->local_dma_lkey = device->local_dma_lkey;
249 	else
250 		mr_access_flags |= IB_ACCESS_LOCAL_WRITE;
251 
252 	if (flags & IB_PD_UNSAFE_GLOBAL_RKEY) {
253 		pr_warn("%s: enabling unsafe global rkey\n", caller);
254 		mr_access_flags |= IB_ACCESS_REMOTE_READ | IB_ACCESS_REMOTE_WRITE;
255 	}
256 
257 	if (mr_access_flags) {
258 		struct ib_mr *mr;
259 
260 		mr = pd->device->get_dma_mr(pd, mr_access_flags);
261 		if (IS_ERR(mr)) {
262 			ib_dealloc_pd(pd);
263 			return ERR_CAST(mr);
264 		}
265 
266 		mr->device	= pd->device;
267 		mr->pd		= pd;
268 		mr->uobject	= NULL;
269 		mr->need_inval	= false;
270 
271 		pd->__internal_mr = mr;
272 
273 		if (!(device->attrs.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY))
274 			pd->local_dma_lkey = pd->__internal_mr->lkey;
275 
276 		if (flags & IB_PD_UNSAFE_GLOBAL_RKEY)
277 			pd->unsafe_global_rkey = pd->__internal_mr->rkey;
278 	}
279 
280 	return pd;
281 }
282 EXPORT_SYMBOL(__ib_alloc_pd);
283 
284 /**
285  * ib_dealloc_pd - Deallocates a protection domain.
286  * @pd: The protection domain to deallocate.
287  *
288  * It is an error to call this function while any resources in the pd still
289  * exist.  The caller is responsible to synchronously destroy them and
290  * guarantee no new allocations will happen.
291  */
292 void ib_dealloc_pd(struct ib_pd *pd)
293 {
294 	int ret;
295 
296 	if (pd->__internal_mr) {
297 		ret = pd->device->dereg_mr(pd->__internal_mr);
298 		WARN_ON(ret);
299 		pd->__internal_mr = NULL;
300 	}
301 
302 	/* uverbs manipulates usecnt with proper locking, while the kabi
303 	   requires the caller to guarantee we can't race here. */
304 	WARN_ON(atomic_read(&pd->usecnt));
305 
306 	/* Making delalloc_pd a void return is a WIP, no driver should return
307 	   an error here. */
308 	ret = pd->device->dealloc_pd(pd);
309 	WARN_ONCE(ret, "Infiniband HW driver failed dealloc_pd");
310 }
311 EXPORT_SYMBOL(ib_dealloc_pd);
312 
313 /* Address handles */
314 
315 struct ib_ah *rdma_create_ah(struct ib_pd *pd, struct rdma_ah_attr *ah_attr)
316 {
317 	struct ib_ah *ah;
318 
319 	ah = pd->device->create_ah(pd, ah_attr, NULL);
320 
321 	if (!IS_ERR(ah)) {
322 		ah->device  = pd->device;
323 		ah->pd      = pd;
324 		ah->uobject = NULL;
325 		ah->type    = ah_attr->type;
326 		atomic_inc(&pd->usecnt);
327 	}
328 
329 	return ah;
330 }
331 EXPORT_SYMBOL(rdma_create_ah);
332 
333 int ib_get_rdma_header_version(const union rdma_network_hdr *hdr)
334 {
335 	const struct iphdr *ip4h = (struct iphdr *)&hdr->roce4grh;
336 	struct iphdr ip4h_checked;
337 	const struct ipv6hdr *ip6h = (struct ipv6hdr *)&hdr->ibgrh;
338 
339 	/* If it's IPv6, the version must be 6, otherwise, the first
340 	 * 20 bytes (before the IPv4 header) are garbled.
341 	 */
342 	if (ip6h->version != 6)
343 		return (ip4h->version == 4) ? 4 : 0;
344 	/* version may be 6 or 4 because the first 20 bytes could be garbled */
345 
346 	/* RoCE v2 requires no options, thus header length
347 	 * must be 5 words
348 	 */
349 	if (ip4h->ihl != 5)
350 		return 6;
351 
352 	/* Verify checksum.
353 	 * We can't write on scattered buffers so we need to copy to
354 	 * temp buffer.
355 	 */
356 	memcpy(&ip4h_checked, ip4h, sizeof(ip4h_checked));
357 	ip4h_checked.check = 0;
358 	ip4h_checked.check = ip_fast_csum((u8 *)&ip4h_checked, 5);
359 	/* if IPv4 header checksum is OK, believe it */
360 	if (ip4h->check == ip4h_checked.check)
361 		return 4;
362 	return 6;
363 }
364 EXPORT_SYMBOL(ib_get_rdma_header_version);
365 
366 static enum rdma_network_type ib_get_net_type_by_grh(struct ib_device *device,
367 						     u8 port_num,
368 						     const struct ib_grh *grh)
369 {
370 	int grh_version;
371 
372 	if (rdma_protocol_ib(device, port_num))
373 		return RDMA_NETWORK_IB;
374 
375 	grh_version = ib_get_rdma_header_version((union rdma_network_hdr *)grh);
376 
377 	if (grh_version == 4)
378 		return RDMA_NETWORK_IPV4;
379 
380 	if (grh->next_hdr == IPPROTO_UDP)
381 		return RDMA_NETWORK_IPV6;
382 
383 	return RDMA_NETWORK_ROCE_V1;
384 }
385 
386 struct find_gid_index_context {
387 	u16 vlan_id;
388 	enum ib_gid_type gid_type;
389 };
390 
391 static bool find_gid_index(const union ib_gid *gid,
392 			   const struct ib_gid_attr *gid_attr,
393 			   void *context)
394 {
395 	struct find_gid_index_context *ctx =
396 		(struct find_gid_index_context *)context;
397 
398 	if (ctx->gid_type != gid_attr->gid_type)
399 		return false;
400 
401 	if ((!!(ctx->vlan_id != 0xffff) == !is_vlan_dev(gid_attr->ndev)) ||
402 	    (is_vlan_dev(gid_attr->ndev) &&
403 	     vlan_dev_vlan_id(gid_attr->ndev) != ctx->vlan_id))
404 		return false;
405 
406 	return true;
407 }
408 
409 static int get_sgid_index_from_eth(struct ib_device *device, u8 port_num,
410 				   u16 vlan_id, const union ib_gid *sgid,
411 				   enum ib_gid_type gid_type,
412 				   u16 *gid_index)
413 {
414 	struct find_gid_index_context context = {.vlan_id = vlan_id,
415 						 .gid_type = gid_type};
416 
417 	return ib_find_gid_by_filter(device, sgid, port_num, find_gid_index,
418 				     &context, gid_index);
419 }
420 
421 int ib_get_gids_from_rdma_hdr(const union rdma_network_hdr *hdr,
422 			      enum rdma_network_type net_type,
423 			      union ib_gid *sgid, union ib_gid *dgid)
424 {
425 	struct sockaddr_in  src_in;
426 	struct sockaddr_in  dst_in;
427 	__be32 src_saddr, dst_saddr;
428 
429 	if (!sgid || !dgid)
430 		return -EINVAL;
431 
432 	if (net_type == RDMA_NETWORK_IPV4) {
433 		memcpy(&src_in.sin_addr.s_addr,
434 		       &hdr->roce4grh.saddr, 4);
435 		memcpy(&dst_in.sin_addr.s_addr,
436 		       &hdr->roce4grh.daddr, 4);
437 		src_saddr = src_in.sin_addr.s_addr;
438 		dst_saddr = dst_in.sin_addr.s_addr;
439 		ipv6_addr_set_v4mapped(src_saddr,
440 				       (struct in6_addr *)sgid);
441 		ipv6_addr_set_v4mapped(dst_saddr,
442 				       (struct in6_addr *)dgid);
443 		return 0;
444 	} else if (net_type == RDMA_NETWORK_IPV6 ||
445 		   net_type == RDMA_NETWORK_IB) {
446 		*dgid = hdr->ibgrh.dgid;
447 		*sgid = hdr->ibgrh.sgid;
448 		return 0;
449 	} else {
450 		return -EINVAL;
451 	}
452 }
453 EXPORT_SYMBOL(ib_get_gids_from_rdma_hdr);
454 
455 /*
456  * This function creates ah from the incoming packet.
457  * Incoming packet has dgid of the receiver node on which this code is
458  * getting executed and, sgid contains the GID of the sender.
459  *
460  * When resolving mac address of destination, the arrived dgid is used
461  * as sgid and, sgid is used as dgid because sgid contains destinations
462  * GID whom to respond to.
463  *
464  * This is why when calling rdma_addr_find_l2_eth_by_grh() function, the
465  * position of arguments dgid and sgid do not match the order of the
466  * parameters.
467  */
468 int ib_init_ah_from_wc(struct ib_device *device, u8 port_num,
469 		       const struct ib_wc *wc, const struct ib_grh *grh,
470 		       struct rdma_ah_attr *ah_attr)
471 {
472 	u32 flow_class;
473 	u16 gid_index;
474 	int ret;
475 	enum rdma_network_type net_type = RDMA_NETWORK_IB;
476 	enum ib_gid_type gid_type = IB_GID_TYPE_IB;
477 	int hoplimit = 0xff;
478 	union ib_gid dgid;
479 	union ib_gid sgid;
480 
481 	memset(ah_attr, 0, sizeof *ah_attr);
482 	ah_attr->type = rdma_ah_find_type(device, port_num);
483 	if (rdma_cap_eth_ah(device, port_num)) {
484 		if (wc->wc_flags & IB_WC_WITH_NETWORK_HDR_TYPE)
485 			net_type = wc->network_hdr_type;
486 		else
487 			net_type = ib_get_net_type_by_grh(device, port_num, grh);
488 		gid_type = ib_network_to_gid_type(net_type);
489 	}
490 	ret = ib_get_gids_from_rdma_hdr((union rdma_network_hdr *)grh, net_type,
491 					&sgid, &dgid);
492 	if (ret)
493 		return ret;
494 
495 	if (rdma_protocol_roce(device, port_num)) {
496 		int if_index = 0;
497 		u16 vlan_id = wc->wc_flags & IB_WC_WITH_VLAN ?
498 				wc->vlan_id : 0xffff;
499 		struct net_device *idev;
500 		struct net_device *resolved_dev;
501 
502 		if (!(wc->wc_flags & IB_WC_GRH))
503 			return -EPROTOTYPE;
504 
505 		if (!device->get_netdev)
506 			return -EOPNOTSUPP;
507 
508 		idev = device->get_netdev(device, port_num);
509 		if (!idev)
510 			return -ENODEV;
511 
512 		ret = rdma_addr_find_l2_eth_by_grh(&dgid, &sgid,
513 						   ah_attr->roce.dmac,
514 						   wc->wc_flags & IB_WC_WITH_VLAN ?
515 						   NULL : &vlan_id,
516 						   &if_index, &hoplimit);
517 		if (ret) {
518 			dev_put(idev);
519 			return ret;
520 		}
521 
522 		resolved_dev = dev_get_by_index(&init_net, if_index);
523 		rcu_read_lock();
524 		if (resolved_dev != idev && !rdma_is_upper_dev_rcu(idev,
525 								   resolved_dev))
526 			ret = -EHOSTUNREACH;
527 		rcu_read_unlock();
528 		dev_put(idev);
529 		dev_put(resolved_dev);
530 		if (ret)
531 			return ret;
532 
533 		ret = get_sgid_index_from_eth(device, port_num, vlan_id,
534 					      &dgid, gid_type, &gid_index);
535 		if (ret)
536 			return ret;
537 	}
538 
539 	rdma_ah_set_dlid(ah_attr, wc->slid);
540 	rdma_ah_set_sl(ah_attr, wc->sl);
541 	rdma_ah_set_path_bits(ah_attr, wc->dlid_path_bits);
542 	rdma_ah_set_port_num(ah_attr, port_num);
543 
544 	if (wc->wc_flags & IB_WC_GRH) {
545 		if (!rdma_cap_eth_ah(device, port_num)) {
546 			if (dgid.global.interface_id != cpu_to_be64(IB_SA_WELL_KNOWN_GUID)) {
547 				ret = ib_find_cached_gid_by_port(device, &dgid,
548 								 IB_GID_TYPE_IB,
549 								 port_num, NULL,
550 								 &gid_index);
551 				if (ret)
552 					return ret;
553 			} else {
554 				gid_index = 0;
555 			}
556 		}
557 
558 		flow_class = be32_to_cpu(grh->version_tclass_flow);
559 		rdma_ah_set_grh(ah_attr, &sgid,
560 				flow_class & 0xFFFFF,
561 				(u8)gid_index, hoplimit,
562 				(flow_class >> 20) & 0xFF);
563 
564 	}
565 	return 0;
566 }
567 EXPORT_SYMBOL(ib_init_ah_from_wc);
568 
569 struct ib_ah *ib_create_ah_from_wc(struct ib_pd *pd, const struct ib_wc *wc,
570 				   const struct ib_grh *grh, u8 port_num)
571 {
572 	struct rdma_ah_attr ah_attr;
573 	int ret;
574 
575 	ret = ib_init_ah_from_wc(pd->device, port_num, wc, grh, &ah_attr);
576 	if (ret)
577 		return ERR_PTR(ret);
578 
579 	return rdma_create_ah(pd, &ah_attr);
580 }
581 EXPORT_SYMBOL(ib_create_ah_from_wc);
582 
583 int rdma_modify_ah(struct ib_ah *ah, struct rdma_ah_attr *ah_attr)
584 {
585 	if (ah->type != ah_attr->type)
586 		return -EINVAL;
587 
588 	return ah->device->modify_ah ?
589 		ah->device->modify_ah(ah, ah_attr) :
590 		-ENOSYS;
591 }
592 EXPORT_SYMBOL(rdma_modify_ah);
593 
594 int rdma_query_ah(struct ib_ah *ah, struct rdma_ah_attr *ah_attr)
595 {
596 	return ah->device->query_ah ?
597 		ah->device->query_ah(ah, ah_attr) :
598 		-ENOSYS;
599 }
600 EXPORT_SYMBOL(rdma_query_ah);
601 
602 int rdma_destroy_ah(struct ib_ah *ah)
603 {
604 	struct ib_pd *pd;
605 	int ret;
606 
607 	pd = ah->pd;
608 	ret = ah->device->destroy_ah(ah);
609 	if (!ret)
610 		atomic_dec(&pd->usecnt);
611 
612 	return ret;
613 }
614 EXPORT_SYMBOL(rdma_destroy_ah);
615 
616 /* Shared receive queues */
617 
618 struct ib_srq *ib_create_srq(struct ib_pd *pd,
619 			     struct ib_srq_init_attr *srq_init_attr)
620 {
621 	struct ib_srq *srq;
622 
623 	if (!pd->device->create_srq)
624 		return ERR_PTR(-ENOSYS);
625 
626 	srq = pd->device->create_srq(pd, srq_init_attr, NULL);
627 
628 	if (!IS_ERR(srq)) {
629 		srq->device    	   = pd->device;
630 		srq->pd        	   = pd;
631 		srq->uobject       = NULL;
632 		srq->event_handler = srq_init_attr->event_handler;
633 		srq->srq_context   = srq_init_attr->srq_context;
634 		srq->srq_type      = srq_init_attr->srq_type;
635 		if (srq->srq_type == IB_SRQT_XRC) {
636 			srq->ext.xrc.xrcd = srq_init_attr->ext.xrc.xrcd;
637 			srq->ext.xrc.cq   = srq_init_attr->ext.xrc.cq;
638 			atomic_inc(&srq->ext.xrc.xrcd->usecnt);
639 			atomic_inc(&srq->ext.xrc.cq->usecnt);
640 		}
641 		atomic_inc(&pd->usecnt);
642 		atomic_set(&srq->usecnt, 0);
643 	}
644 
645 	return srq;
646 }
647 EXPORT_SYMBOL(ib_create_srq);
648 
649 int ib_modify_srq(struct ib_srq *srq,
650 		  struct ib_srq_attr *srq_attr,
651 		  enum ib_srq_attr_mask srq_attr_mask)
652 {
653 	return srq->device->modify_srq ?
654 		srq->device->modify_srq(srq, srq_attr, srq_attr_mask, NULL) :
655 		-ENOSYS;
656 }
657 EXPORT_SYMBOL(ib_modify_srq);
658 
659 int ib_query_srq(struct ib_srq *srq,
660 		 struct ib_srq_attr *srq_attr)
661 {
662 	return srq->device->query_srq ?
663 		srq->device->query_srq(srq, srq_attr) : -ENOSYS;
664 }
665 EXPORT_SYMBOL(ib_query_srq);
666 
667 int ib_destroy_srq(struct ib_srq *srq)
668 {
669 	struct ib_pd *pd;
670 	enum ib_srq_type srq_type;
671 	struct ib_xrcd *uninitialized_var(xrcd);
672 	struct ib_cq *uninitialized_var(cq);
673 	int ret;
674 
675 	if (atomic_read(&srq->usecnt))
676 		return -EBUSY;
677 
678 	pd = srq->pd;
679 	srq_type = srq->srq_type;
680 	if (srq_type == IB_SRQT_XRC) {
681 		xrcd = srq->ext.xrc.xrcd;
682 		cq = srq->ext.xrc.cq;
683 	}
684 
685 	ret = srq->device->destroy_srq(srq);
686 	if (!ret) {
687 		atomic_dec(&pd->usecnt);
688 		if (srq_type == IB_SRQT_XRC) {
689 			atomic_dec(&xrcd->usecnt);
690 			atomic_dec(&cq->usecnt);
691 		}
692 	}
693 
694 	return ret;
695 }
696 EXPORT_SYMBOL(ib_destroy_srq);
697 
698 /* Queue pairs */
699 
700 static void __ib_shared_qp_event_handler(struct ib_event *event, void *context)
701 {
702 	struct ib_qp *qp = context;
703 	unsigned long flags;
704 
705 	spin_lock_irqsave(&qp->device->event_handler_lock, flags);
706 	list_for_each_entry(event->element.qp, &qp->open_list, open_list)
707 		if (event->element.qp->event_handler)
708 			event->element.qp->event_handler(event, event->element.qp->qp_context);
709 	spin_unlock_irqrestore(&qp->device->event_handler_lock, flags);
710 }
711 
712 static void __ib_insert_xrcd_qp(struct ib_xrcd *xrcd, struct ib_qp *qp)
713 {
714 	mutex_lock(&xrcd->tgt_qp_mutex);
715 	list_add(&qp->xrcd_list, &xrcd->tgt_qp_list);
716 	mutex_unlock(&xrcd->tgt_qp_mutex);
717 }
718 
719 static struct ib_qp *__ib_open_qp(struct ib_qp *real_qp,
720 				  void (*event_handler)(struct ib_event *, void *),
721 				  void *qp_context)
722 {
723 	struct ib_qp *qp;
724 	unsigned long flags;
725 	int err;
726 
727 	qp = kzalloc(sizeof *qp, GFP_KERNEL);
728 	if (!qp)
729 		return ERR_PTR(-ENOMEM);
730 
731 	qp->real_qp = real_qp;
732 	err = ib_open_shared_qp_security(qp, real_qp->device);
733 	if (err) {
734 		kfree(qp);
735 		return ERR_PTR(err);
736 	}
737 
738 	qp->real_qp = real_qp;
739 	atomic_inc(&real_qp->usecnt);
740 	qp->device = real_qp->device;
741 	qp->event_handler = event_handler;
742 	qp->qp_context = qp_context;
743 	qp->qp_num = real_qp->qp_num;
744 	qp->qp_type = real_qp->qp_type;
745 
746 	spin_lock_irqsave(&real_qp->device->event_handler_lock, flags);
747 	list_add(&qp->open_list, &real_qp->open_list);
748 	spin_unlock_irqrestore(&real_qp->device->event_handler_lock, flags);
749 
750 	return qp;
751 }
752 
753 struct ib_qp *ib_open_qp(struct ib_xrcd *xrcd,
754 			 struct ib_qp_open_attr *qp_open_attr)
755 {
756 	struct ib_qp *qp, *real_qp;
757 
758 	if (qp_open_attr->qp_type != IB_QPT_XRC_TGT)
759 		return ERR_PTR(-EINVAL);
760 
761 	qp = ERR_PTR(-EINVAL);
762 	mutex_lock(&xrcd->tgt_qp_mutex);
763 	list_for_each_entry(real_qp, &xrcd->tgt_qp_list, xrcd_list) {
764 		if (real_qp->qp_num == qp_open_attr->qp_num) {
765 			qp = __ib_open_qp(real_qp, qp_open_attr->event_handler,
766 					  qp_open_attr->qp_context);
767 			break;
768 		}
769 	}
770 	mutex_unlock(&xrcd->tgt_qp_mutex);
771 	return qp;
772 }
773 EXPORT_SYMBOL(ib_open_qp);
774 
775 static struct ib_qp *ib_create_xrc_qp(struct ib_qp *qp,
776 		struct ib_qp_init_attr *qp_init_attr)
777 {
778 	struct ib_qp *real_qp = qp;
779 
780 	qp->event_handler = __ib_shared_qp_event_handler;
781 	qp->qp_context = qp;
782 	qp->pd = NULL;
783 	qp->send_cq = qp->recv_cq = NULL;
784 	qp->srq = NULL;
785 	qp->xrcd = qp_init_attr->xrcd;
786 	atomic_inc(&qp_init_attr->xrcd->usecnt);
787 	INIT_LIST_HEAD(&qp->open_list);
788 
789 	qp = __ib_open_qp(real_qp, qp_init_attr->event_handler,
790 			  qp_init_attr->qp_context);
791 	if (!IS_ERR(qp))
792 		__ib_insert_xrcd_qp(qp_init_attr->xrcd, real_qp);
793 	else
794 		real_qp->device->destroy_qp(real_qp);
795 	return qp;
796 }
797 
798 struct ib_qp *ib_create_qp(struct ib_pd *pd,
799 			   struct ib_qp_init_attr *qp_init_attr)
800 {
801 	struct ib_device *device = pd ? pd->device : qp_init_attr->xrcd->device;
802 	struct ib_qp *qp;
803 	int ret;
804 
805 	if (qp_init_attr->rwq_ind_tbl &&
806 	    (qp_init_attr->recv_cq ||
807 	    qp_init_attr->srq || qp_init_attr->cap.max_recv_wr ||
808 	    qp_init_attr->cap.max_recv_sge))
809 		return ERR_PTR(-EINVAL);
810 
811 	/*
812 	 * If the callers is using the RDMA API calculate the resources
813 	 * needed for the RDMA READ/WRITE operations.
814 	 *
815 	 * Note that these callers need to pass in a port number.
816 	 */
817 	if (qp_init_attr->cap.max_rdma_ctxs)
818 		rdma_rw_init_qp(device, qp_init_attr);
819 
820 	qp = device->create_qp(pd, qp_init_attr, NULL);
821 	if (IS_ERR(qp))
822 		return qp;
823 
824 	ret = ib_create_qp_security(qp, device);
825 	if (ret) {
826 		ib_destroy_qp(qp);
827 		return ERR_PTR(ret);
828 	}
829 
830 	qp->device     = device;
831 	qp->real_qp    = qp;
832 	qp->uobject    = NULL;
833 	qp->qp_type    = qp_init_attr->qp_type;
834 	qp->rwq_ind_tbl = qp_init_attr->rwq_ind_tbl;
835 
836 	atomic_set(&qp->usecnt, 0);
837 	qp->mrs_used = 0;
838 	spin_lock_init(&qp->mr_lock);
839 	INIT_LIST_HEAD(&qp->rdma_mrs);
840 	INIT_LIST_HEAD(&qp->sig_mrs);
841 
842 	if (qp_init_attr->qp_type == IB_QPT_XRC_TGT)
843 		return ib_create_xrc_qp(qp, qp_init_attr);
844 
845 	qp->event_handler = qp_init_attr->event_handler;
846 	qp->qp_context = qp_init_attr->qp_context;
847 	if (qp_init_attr->qp_type == IB_QPT_XRC_INI) {
848 		qp->recv_cq = NULL;
849 		qp->srq = NULL;
850 	} else {
851 		qp->recv_cq = qp_init_attr->recv_cq;
852 		if (qp_init_attr->recv_cq)
853 			atomic_inc(&qp_init_attr->recv_cq->usecnt);
854 		qp->srq = qp_init_attr->srq;
855 		if (qp->srq)
856 			atomic_inc(&qp_init_attr->srq->usecnt);
857 	}
858 
859 	qp->pd	    = pd;
860 	qp->send_cq = qp_init_attr->send_cq;
861 	qp->xrcd    = NULL;
862 
863 	atomic_inc(&pd->usecnt);
864 	if (qp_init_attr->send_cq)
865 		atomic_inc(&qp_init_attr->send_cq->usecnt);
866 	if (qp_init_attr->rwq_ind_tbl)
867 		atomic_inc(&qp->rwq_ind_tbl->usecnt);
868 
869 	if (qp_init_attr->cap.max_rdma_ctxs) {
870 		ret = rdma_rw_init_mrs(qp, qp_init_attr);
871 		if (ret) {
872 			pr_err("failed to init MR pool ret= %d\n", ret);
873 			ib_destroy_qp(qp);
874 			return ERR_PTR(ret);
875 		}
876 	}
877 
878 	/*
879 	 * Note: all hw drivers guarantee that max_send_sge is lower than
880 	 * the device RDMA WRITE SGE limit but not all hw drivers ensure that
881 	 * max_send_sge <= max_sge_rd.
882 	 */
883 	qp->max_write_sge = qp_init_attr->cap.max_send_sge;
884 	qp->max_read_sge = min_t(u32, qp_init_attr->cap.max_send_sge,
885 				 device->attrs.max_sge_rd);
886 
887 	return qp;
888 }
889 EXPORT_SYMBOL(ib_create_qp);
890 
891 static const struct {
892 	int			valid;
893 	enum ib_qp_attr_mask	req_param[IB_QPT_MAX];
894 	enum ib_qp_attr_mask	opt_param[IB_QPT_MAX];
895 } qp_state_table[IB_QPS_ERR + 1][IB_QPS_ERR + 1] = {
896 	[IB_QPS_RESET] = {
897 		[IB_QPS_RESET] = { .valid = 1 },
898 		[IB_QPS_ERR] =   { .valid = 1 },
899 		[IB_QPS_INIT]  = {
900 			.valid = 1,
901 			.req_param = {
902 				[IB_QPT_UD]  = (IB_QP_PKEY_INDEX		|
903 						IB_QP_PORT			|
904 						IB_QP_QKEY),
905 				[IB_QPT_RAW_PACKET] = IB_QP_PORT,
906 				[IB_QPT_UC]  = (IB_QP_PKEY_INDEX		|
907 						IB_QP_PORT			|
908 						IB_QP_ACCESS_FLAGS),
909 				[IB_QPT_RC]  = (IB_QP_PKEY_INDEX		|
910 						IB_QP_PORT			|
911 						IB_QP_ACCESS_FLAGS),
912 				[IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX		|
913 						IB_QP_PORT			|
914 						IB_QP_ACCESS_FLAGS),
915 				[IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX		|
916 						IB_QP_PORT			|
917 						IB_QP_ACCESS_FLAGS),
918 				[IB_QPT_SMI] = (IB_QP_PKEY_INDEX		|
919 						IB_QP_QKEY),
920 				[IB_QPT_GSI] = (IB_QP_PKEY_INDEX		|
921 						IB_QP_QKEY),
922 			}
923 		},
924 	},
925 	[IB_QPS_INIT]  = {
926 		[IB_QPS_RESET] = { .valid = 1 },
927 		[IB_QPS_ERR] =   { .valid = 1 },
928 		[IB_QPS_INIT]  = {
929 			.valid = 1,
930 			.opt_param = {
931 				[IB_QPT_UD]  = (IB_QP_PKEY_INDEX		|
932 						IB_QP_PORT			|
933 						IB_QP_QKEY),
934 				[IB_QPT_UC]  = (IB_QP_PKEY_INDEX		|
935 						IB_QP_PORT			|
936 						IB_QP_ACCESS_FLAGS),
937 				[IB_QPT_RC]  = (IB_QP_PKEY_INDEX		|
938 						IB_QP_PORT			|
939 						IB_QP_ACCESS_FLAGS),
940 				[IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX		|
941 						IB_QP_PORT			|
942 						IB_QP_ACCESS_FLAGS),
943 				[IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX		|
944 						IB_QP_PORT			|
945 						IB_QP_ACCESS_FLAGS),
946 				[IB_QPT_SMI] = (IB_QP_PKEY_INDEX		|
947 						IB_QP_QKEY),
948 				[IB_QPT_GSI] = (IB_QP_PKEY_INDEX		|
949 						IB_QP_QKEY),
950 			}
951 		},
952 		[IB_QPS_RTR]   = {
953 			.valid = 1,
954 			.req_param = {
955 				[IB_QPT_UC]  = (IB_QP_AV			|
956 						IB_QP_PATH_MTU			|
957 						IB_QP_DEST_QPN			|
958 						IB_QP_RQ_PSN),
959 				[IB_QPT_RC]  = (IB_QP_AV			|
960 						IB_QP_PATH_MTU			|
961 						IB_QP_DEST_QPN			|
962 						IB_QP_RQ_PSN			|
963 						IB_QP_MAX_DEST_RD_ATOMIC	|
964 						IB_QP_MIN_RNR_TIMER),
965 				[IB_QPT_XRC_INI] = (IB_QP_AV			|
966 						IB_QP_PATH_MTU			|
967 						IB_QP_DEST_QPN			|
968 						IB_QP_RQ_PSN),
969 				[IB_QPT_XRC_TGT] = (IB_QP_AV			|
970 						IB_QP_PATH_MTU			|
971 						IB_QP_DEST_QPN			|
972 						IB_QP_RQ_PSN			|
973 						IB_QP_MAX_DEST_RD_ATOMIC	|
974 						IB_QP_MIN_RNR_TIMER),
975 			},
976 			.opt_param = {
977 				 [IB_QPT_UD]  = (IB_QP_PKEY_INDEX		|
978 						 IB_QP_QKEY),
979 				 [IB_QPT_UC]  = (IB_QP_ALT_PATH			|
980 						 IB_QP_ACCESS_FLAGS		|
981 						 IB_QP_PKEY_INDEX),
982 				 [IB_QPT_RC]  = (IB_QP_ALT_PATH			|
983 						 IB_QP_ACCESS_FLAGS		|
984 						 IB_QP_PKEY_INDEX),
985 				 [IB_QPT_XRC_INI] = (IB_QP_ALT_PATH		|
986 						 IB_QP_ACCESS_FLAGS		|
987 						 IB_QP_PKEY_INDEX),
988 				 [IB_QPT_XRC_TGT] = (IB_QP_ALT_PATH		|
989 						 IB_QP_ACCESS_FLAGS		|
990 						 IB_QP_PKEY_INDEX),
991 				 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX		|
992 						 IB_QP_QKEY),
993 				 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX		|
994 						 IB_QP_QKEY),
995 			 },
996 		},
997 	},
998 	[IB_QPS_RTR]   = {
999 		[IB_QPS_RESET] = { .valid = 1 },
1000 		[IB_QPS_ERR] =   { .valid = 1 },
1001 		[IB_QPS_RTS]   = {
1002 			.valid = 1,
1003 			.req_param = {
1004 				[IB_QPT_UD]  = IB_QP_SQ_PSN,
1005 				[IB_QPT_UC]  = IB_QP_SQ_PSN,
1006 				[IB_QPT_RC]  = (IB_QP_TIMEOUT			|
1007 						IB_QP_RETRY_CNT			|
1008 						IB_QP_RNR_RETRY			|
1009 						IB_QP_SQ_PSN			|
1010 						IB_QP_MAX_QP_RD_ATOMIC),
1011 				[IB_QPT_XRC_INI] = (IB_QP_TIMEOUT		|
1012 						IB_QP_RETRY_CNT			|
1013 						IB_QP_RNR_RETRY			|
1014 						IB_QP_SQ_PSN			|
1015 						IB_QP_MAX_QP_RD_ATOMIC),
1016 				[IB_QPT_XRC_TGT] = (IB_QP_TIMEOUT		|
1017 						IB_QP_SQ_PSN),
1018 				[IB_QPT_SMI] = IB_QP_SQ_PSN,
1019 				[IB_QPT_GSI] = IB_QP_SQ_PSN,
1020 			},
1021 			.opt_param = {
1022 				 [IB_QPT_UD]  = (IB_QP_CUR_STATE		|
1023 						 IB_QP_QKEY),
1024 				 [IB_QPT_UC]  = (IB_QP_CUR_STATE		|
1025 						 IB_QP_ALT_PATH			|
1026 						 IB_QP_ACCESS_FLAGS		|
1027 						 IB_QP_PATH_MIG_STATE),
1028 				 [IB_QPT_RC]  = (IB_QP_CUR_STATE		|
1029 						 IB_QP_ALT_PATH			|
1030 						 IB_QP_ACCESS_FLAGS		|
1031 						 IB_QP_MIN_RNR_TIMER		|
1032 						 IB_QP_PATH_MIG_STATE),
1033 				 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE		|
1034 						 IB_QP_ALT_PATH			|
1035 						 IB_QP_ACCESS_FLAGS		|
1036 						 IB_QP_PATH_MIG_STATE),
1037 				 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE		|
1038 						 IB_QP_ALT_PATH			|
1039 						 IB_QP_ACCESS_FLAGS		|
1040 						 IB_QP_MIN_RNR_TIMER		|
1041 						 IB_QP_PATH_MIG_STATE),
1042 				 [IB_QPT_SMI] = (IB_QP_CUR_STATE		|
1043 						 IB_QP_QKEY),
1044 				 [IB_QPT_GSI] = (IB_QP_CUR_STATE		|
1045 						 IB_QP_QKEY),
1046 				 [IB_QPT_RAW_PACKET] = IB_QP_RATE_LIMIT,
1047 			 }
1048 		}
1049 	},
1050 	[IB_QPS_RTS]   = {
1051 		[IB_QPS_RESET] = { .valid = 1 },
1052 		[IB_QPS_ERR] =   { .valid = 1 },
1053 		[IB_QPS_RTS]   = {
1054 			.valid = 1,
1055 			.opt_param = {
1056 				[IB_QPT_UD]  = (IB_QP_CUR_STATE			|
1057 						IB_QP_QKEY),
1058 				[IB_QPT_UC]  = (IB_QP_CUR_STATE			|
1059 						IB_QP_ACCESS_FLAGS		|
1060 						IB_QP_ALT_PATH			|
1061 						IB_QP_PATH_MIG_STATE),
1062 				[IB_QPT_RC]  = (IB_QP_CUR_STATE			|
1063 						IB_QP_ACCESS_FLAGS		|
1064 						IB_QP_ALT_PATH			|
1065 						IB_QP_PATH_MIG_STATE		|
1066 						IB_QP_MIN_RNR_TIMER),
1067 				[IB_QPT_XRC_INI] = (IB_QP_CUR_STATE		|
1068 						IB_QP_ACCESS_FLAGS		|
1069 						IB_QP_ALT_PATH			|
1070 						IB_QP_PATH_MIG_STATE),
1071 				[IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE		|
1072 						IB_QP_ACCESS_FLAGS		|
1073 						IB_QP_ALT_PATH			|
1074 						IB_QP_PATH_MIG_STATE		|
1075 						IB_QP_MIN_RNR_TIMER),
1076 				[IB_QPT_SMI] = (IB_QP_CUR_STATE			|
1077 						IB_QP_QKEY),
1078 				[IB_QPT_GSI] = (IB_QP_CUR_STATE			|
1079 						IB_QP_QKEY),
1080 				[IB_QPT_RAW_PACKET] = IB_QP_RATE_LIMIT,
1081 			}
1082 		},
1083 		[IB_QPS_SQD]   = {
1084 			.valid = 1,
1085 			.opt_param = {
1086 				[IB_QPT_UD]  = IB_QP_EN_SQD_ASYNC_NOTIFY,
1087 				[IB_QPT_UC]  = IB_QP_EN_SQD_ASYNC_NOTIFY,
1088 				[IB_QPT_RC]  = IB_QP_EN_SQD_ASYNC_NOTIFY,
1089 				[IB_QPT_XRC_INI] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1090 				[IB_QPT_XRC_TGT] = IB_QP_EN_SQD_ASYNC_NOTIFY, /* ??? */
1091 				[IB_QPT_SMI] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1092 				[IB_QPT_GSI] = IB_QP_EN_SQD_ASYNC_NOTIFY
1093 			}
1094 		},
1095 	},
1096 	[IB_QPS_SQD]   = {
1097 		[IB_QPS_RESET] = { .valid = 1 },
1098 		[IB_QPS_ERR] =   { .valid = 1 },
1099 		[IB_QPS_RTS]   = {
1100 			.valid = 1,
1101 			.opt_param = {
1102 				[IB_QPT_UD]  = (IB_QP_CUR_STATE			|
1103 						IB_QP_QKEY),
1104 				[IB_QPT_UC]  = (IB_QP_CUR_STATE			|
1105 						IB_QP_ALT_PATH			|
1106 						IB_QP_ACCESS_FLAGS		|
1107 						IB_QP_PATH_MIG_STATE),
1108 				[IB_QPT_RC]  = (IB_QP_CUR_STATE			|
1109 						IB_QP_ALT_PATH			|
1110 						IB_QP_ACCESS_FLAGS		|
1111 						IB_QP_MIN_RNR_TIMER		|
1112 						IB_QP_PATH_MIG_STATE),
1113 				[IB_QPT_XRC_INI] = (IB_QP_CUR_STATE		|
1114 						IB_QP_ALT_PATH			|
1115 						IB_QP_ACCESS_FLAGS		|
1116 						IB_QP_PATH_MIG_STATE),
1117 				[IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE		|
1118 						IB_QP_ALT_PATH			|
1119 						IB_QP_ACCESS_FLAGS		|
1120 						IB_QP_MIN_RNR_TIMER		|
1121 						IB_QP_PATH_MIG_STATE),
1122 				[IB_QPT_SMI] = (IB_QP_CUR_STATE			|
1123 						IB_QP_QKEY),
1124 				[IB_QPT_GSI] = (IB_QP_CUR_STATE			|
1125 						IB_QP_QKEY),
1126 			}
1127 		},
1128 		[IB_QPS_SQD]   = {
1129 			.valid = 1,
1130 			.opt_param = {
1131 				[IB_QPT_UD]  = (IB_QP_PKEY_INDEX		|
1132 						IB_QP_QKEY),
1133 				[IB_QPT_UC]  = (IB_QP_AV			|
1134 						IB_QP_ALT_PATH			|
1135 						IB_QP_ACCESS_FLAGS		|
1136 						IB_QP_PKEY_INDEX		|
1137 						IB_QP_PATH_MIG_STATE),
1138 				[IB_QPT_RC]  = (IB_QP_PORT			|
1139 						IB_QP_AV			|
1140 						IB_QP_TIMEOUT			|
1141 						IB_QP_RETRY_CNT			|
1142 						IB_QP_RNR_RETRY			|
1143 						IB_QP_MAX_QP_RD_ATOMIC		|
1144 						IB_QP_MAX_DEST_RD_ATOMIC	|
1145 						IB_QP_ALT_PATH			|
1146 						IB_QP_ACCESS_FLAGS		|
1147 						IB_QP_PKEY_INDEX		|
1148 						IB_QP_MIN_RNR_TIMER		|
1149 						IB_QP_PATH_MIG_STATE),
1150 				[IB_QPT_XRC_INI] = (IB_QP_PORT			|
1151 						IB_QP_AV			|
1152 						IB_QP_TIMEOUT			|
1153 						IB_QP_RETRY_CNT			|
1154 						IB_QP_RNR_RETRY			|
1155 						IB_QP_MAX_QP_RD_ATOMIC		|
1156 						IB_QP_ALT_PATH			|
1157 						IB_QP_ACCESS_FLAGS		|
1158 						IB_QP_PKEY_INDEX		|
1159 						IB_QP_PATH_MIG_STATE),
1160 				[IB_QPT_XRC_TGT] = (IB_QP_PORT			|
1161 						IB_QP_AV			|
1162 						IB_QP_TIMEOUT			|
1163 						IB_QP_MAX_DEST_RD_ATOMIC	|
1164 						IB_QP_ALT_PATH			|
1165 						IB_QP_ACCESS_FLAGS		|
1166 						IB_QP_PKEY_INDEX		|
1167 						IB_QP_MIN_RNR_TIMER		|
1168 						IB_QP_PATH_MIG_STATE),
1169 				[IB_QPT_SMI] = (IB_QP_PKEY_INDEX		|
1170 						IB_QP_QKEY),
1171 				[IB_QPT_GSI] = (IB_QP_PKEY_INDEX		|
1172 						IB_QP_QKEY),
1173 			}
1174 		}
1175 	},
1176 	[IB_QPS_SQE]   = {
1177 		[IB_QPS_RESET] = { .valid = 1 },
1178 		[IB_QPS_ERR] =   { .valid = 1 },
1179 		[IB_QPS_RTS]   = {
1180 			.valid = 1,
1181 			.opt_param = {
1182 				[IB_QPT_UD]  = (IB_QP_CUR_STATE			|
1183 						IB_QP_QKEY),
1184 				[IB_QPT_UC]  = (IB_QP_CUR_STATE			|
1185 						IB_QP_ACCESS_FLAGS),
1186 				[IB_QPT_SMI] = (IB_QP_CUR_STATE			|
1187 						IB_QP_QKEY),
1188 				[IB_QPT_GSI] = (IB_QP_CUR_STATE			|
1189 						IB_QP_QKEY),
1190 			}
1191 		}
1192 	},
1193 	[IB_QPS_ERR] = {
1194 		[IB_QPS_RESET] = { .valid = 1 },
1195 		[IB_QPS_ERR] =   { .valid = 1 }
1196 	}
1197 };
1198 
1199 int ib_modify_qp_is_ok(enum ib_qp_state cur_state, enum ib_qp_state next_state,
1200 		       enum ib_qp_type type, enum ib_qp_attr_mask mask,
1201 		       enum rdma_link_layer ll)
1202 {
1203 	enum ib_qp_attr_mask req_param, opt_param;
1204 
1205 	if (cur_state  < 0 || cur_state  > IB_QPS_ERR ||
1206 	    next_state < 0 || next_state > IB_QPS_ERR)
1207 		return 0;
1208 
1209 	if (mask & IB_QP_CUR_STATE  &&
1210 	    cur_state != IB_QPS_RTR && cur_state != IB_QPS_RTS &&
1211 	    cur_state != IB_QPS_SQD && cur_state != IB_QPS_SQE)
1212 		return 0;
1213 
1214 	if (!qp_state_table[cur_state][next_state].valid)
1215 		return 0;
1216 
1217 	req_param = qp_state_table[cur_state][next_state].req_param[type];
1218 	opt_param = qp_state_table[cur_state][next_state].opt_param[type];
1219 
1220 	if ((mask & req_param) != req_param)
1221 		return 0;
1222 
1223 	if (mask & ~(req_param | opt_param | IB_QP_STATE))
1224 		return 0;
1225 
1226 	return 1;
1227 }
1228 EXPORT_SYMBOL(ib_modify_qp_is_ok);
1229 
1230 int ib_resolve_eth_dmac(struct ib_device *device,
1231 			struct rdma_ah_attr *ah_attr)
1232 {
1233 	int           ret = 0;
1234 	struct ib_global_route *grh;
1235 
1236 	if (!rdma_is_port_valid(device, rdma_ah_get_port_num(ah_attr)))
1237 		return -EINVAL;
1238 
1239 	if (ah_attr->type != RDMA_AH_ATTR_TYPE_ROCE)
1240 		return 0;
1241 
1242 	grh = rdma_ah_retrieve_grh(ah_attr);
1243 
1244 	if (rdma_link_local_addr((struct in6_addr *)grh->dgid.raw)) {
1245 		rdma_get_ll_mac((struct in6_addr *)grh->dgid.raw,
1246 				ah_attr->roce.dmac);
1247 	} else {
1248 		union ib_gid		sgid;
1249 		struct ib_gid_attr	sgid_attr;
1250 		int			ifindex;
1251 		int			hop_limit;
1252 
1253 		ret = ib_query_gid(device,
1254 				   rdma_ah_get_port_num(ah_attr),
1255 				   grh->sgid_index,
1256 				   &sgid, &sgid_attr);
1257 
1258 		if (ret || !sgid_attr.ndev) {
1259 			if (!ret)
1260 				ret = -ENXIO;
1261 			goto out;
1262 		}
1263 
1264 		ifindex = sgid_attr.ndev->ifindex;
1265 
1266 		ret =
1267 		rdma_addr_find_l2_eth_by_grh(&sgid, &grh->dgid,
1268 					     ah_attr->roce.dmac,
1269 					     NULL, &ifindex, &hop_limit);
1270 
1271 		dev_put(sgid_attr.ndev);
1272 
1273 		grh->hop_limit = hop_limit;
1274 	}
1275 out:
1276 	return ret;
1277 }
1278 EXPORT_SYMBOL(ib_resolve_eth_dmac);
1279 
1280 /**
1281  * ib_modify_qp_with_udata - Modifies the attributes for the specified QP.
1282  * @qp: The QP to modify.
1283  * @attr: On input, specifies the QP attributes to modify.  On output,
1284  *   the current values of selected QP attributes are returned.
1285  * @attr_mask: A bit-mask used to specify which attributes of the QP
1286  *   are being modified.
1287  * @udata: pointer to user's input output buffer information
1288  *   are being modified.
1289  * It returns 0 on success and returns appropriate error code on error.
1290  */
1291 int ib_modify_qp_with_udata(struct ib_qp *qp, struct ib_qp_attr *attr,
1292 			    int attr_mask, struct ib_udata *udata)
1293 {
1294 	int ret;
1295 
1296 	if (attr_mask & IB_QP_AV) {
1297 		ret = ib_resolve_eth_dmac(qp->device, &attr->ah_attr);
1298 		if (ret)
1299 			return ret;
1300 	}
1301 	return ib_security_modify_qp(qp, attr, attr_mask, udata);
1302 }
1303 EXPORT_SYMBOL(ib_modify_qp_with_udata);
1304 
1305 int ib_modify_qp(struct ib_qp *qp,
1306 		 struct ib_qp_attr *qp_attr,
1307 		 int qp_attr_mask)
1308 {
1309 	return ib_modify_qp_with_udata(qp, qp_attr, qp_attr_mask, NULL);
1310 }
1311 EXPORT_SYMBOL(ib_modify_qp);
1312 
1313 int ib_query_qp(struct ib_qp *qp,
1314 		struct ib_qp_attr *qp_attr,
1315 		int qp_attr_mask,
1316 		struct ib_qp_init_attr *qp_init_attr)
1317 {
1318 	return qp->device->query_qp ?
1319 		qp->device->query_qp(qp->real_qp, qp_attr, qp_attr_mask, qp_init_attr) :
1320 		-ENOSYS;
1321 }
1322 EXPORT_SYMBOL(ib_query_qp);
1323 
1324 int ib_close_qp(struct ib_qp *qp)
1325 {
1326 	struct ib_qp *real_qp;
1327 	unsigned long flags;
1328 
1329 	real_qp = qp->real_qp;
1330 	if (real_qp == qp)
1331 		return -EINVAL;
1332 
1333 	spin_lock_irqsave(&real_qp->device->event_handler_lock, flags);
1334 	list_del(&qp->open_list);
1335 	spin_unlock_irqrestore(&real_qp->device->event_handler_lock, flags);
1336 
1337 	atomic_dec(&real_qp->usecnt);
1338 	ib_close_shared_qp_security(qp->qp_sec);
1339 	kfree(qp);
1340 
1341 	return 0;
1342 }
1343 EXPORT_SYMBOL(ib_close_qp);
1344 
1345 static int __ib_destroy_shared_qp(struct ib_qp *qp)
1346 {
1347 	struct ib_xrcd *xrcd;
1348 	struct ib_qp *real_qp;
1349 	int ret;
1350 
1351 	real_qp = qp->real_qp;
1352 	xrcd = real_qp->xrcd;
1353 
1354 	mutex_lock(&xrcd->tgt_qp_mutex);
1355 	ib_close_qp(qp);
1356 	if (atomic_read(&real_qp->usecnt) == 0)
1357 		list_del(&real_qp->xrcd_list);
1358 	else
1359 		real_qp = NULL;
1360 	mutex_unlock(&xrcd->tgt_qp_mutex);
1361 
1362 	if (real_qp) {
1363 		ret = ib_destroy_qp(real_qp);
1364 		if (!ret)
1365 			atomic_dec(&xrcd->usecnt);
1366 		else
1367 			__ib_insert_xrcd_qp(xrcd, real_qp);
1368 	}
1369 
1370 	return 0;
1371 }
1372 
1373 int ib_destroy_qp(struct ib_qp *qp)
1374 {
1375 	struct ib_pd *pd;
1376 	struct ib_cq *scq, *rcq;
1377 	struct ib_srq *srq;
1378 	struct ib_rwq_ind_table *ind_tbl;
1379 	struct ib_qp_security *sec;
1380 	int ret;
1381 
1382 	WARN_ON_ONCE(qp->mrs_used > 0);
1383 
1384 	if (atomic_read(&qp->usecnt))
1385 		return -EBUSY;
1386 
1387 	if (qp->real_qp != qp)
1388 		return __ib_destroy_shared_qp(qp);
1389 
1390 	pd   = qp->pd;
1391 	scq  = qp->send_cq;
1392 	rcq  = qp->recv_cq;
1393 	srq  = qp->srq;
1394 	ind_tbl = qp->rwq_ind_tbl;
1395 	sec  = qp->qp_sec;
1396 	if (sec)
1397 		ib_destroy_qp_security_begin(sec);
1398 
1399 	if (!qp->uobject)
1400 		rdma_rw_cleanup_mrs(qp);
1401 
1402 	ret = qp->device->destroy_qp(qp);
1403 	if (!ret) {
1404 		if (pd)
1405 			atomic_dec(&pd->usecnt);
1406 		if (scq)
1407 			atomic_dec(&scq->usecnt);
1408 		if (rcq)
1409 			atomic_dec(&rcq->usecnt);
1410 		if (srq)
1411 			atomic_dec(&srq->usecnt);
1412 		if (ind_tbl)
1413 			atomic_dec(&ind_tbl->usecnt);
1414 		if (sec)
1415 			ib_destroy_qp_security_end(sec);
1416 	} else {
1417 		if (sec)
1418 			ib_destroy_qp_security_abort(sec);
1419 	}
1420 
1421 	return ret;
1422 }
1423 EXPORT_SYMBOL(ib_destroy_qp);
1424 
1425 /* Completion queues */
1426 
1427 struct ib_cq *ib_create_cq(struct ib_device *device,
1428 			   ib_comp_handler comp_handler,
1429 			   void (*event_handler)(struct ib_event *, void *),
1430 			   void *cq_context,
1431 			   const struct ib_cq_init_attr *cq_attr)
1432 {
1433 	struct ib_cq *cq;
1434 
1435 	cq = device->create_cq(device, cq_attr, NULL, NULL);
1436 
1437 	if (!IS_ERR(cq)) {
1438 		cq->device        = device;
1439 		cq->uobject       = NULL;
1440 		cq->comp_handler  = comp_handler;
1441 		cq->event_handler = event_handler;
1442 		cq->cq_context    = cq_context;
1443 		atomic_set(&cq->usecnt, 0);
1444 	}
1445 
1446 	return cq;
1447 }
1448 EXPORT_SYMBOL(ib_create_cq);
1449 
1450 int ib_modify_cq(struct ib_cq *cq, u16 cq_count, u16 cq_period)
1451 {
1452 	return cq->device->modify_cq ?
1453 		cq->device->modify_cq(cq, cq_count, cq_period) : -ENOSYS;
1454 }
1455 EXPORT_SYMBOL(ib_modify_cq);
1456 
1457 int ib_destroy_cq(struct ib_cq *cq)
1458 {
1459 	if (atomic_read(&cq->usecnt))
1460 		return -EBUSY;
1461 
1462 	return cq->device->destroy_cq(cq);
1463 }
1464 EXPORT_SYMBOL(ib_destroy_cq);
1465 
1466 int ib_resize_cq(struct ib_cq *cq, int cqe)
1467 {
1468 	return cq->device->resize_cq ?
1469 		cq->device->resize_cq(cq, cqe, NULL) : -ENOSYS;
1470 }
1471 EXPORT_SYMBOL(ib_resize_cq);
1472 
1473 /* Memory regions */
1474 
1475 int ib_dereg_mr(struct ib_mr *mr)
1476 {
1477 	struct ib_pd *pd = mr->pd;
1478 	int ret;
1479 
1480 	ret = mr->device->dereg_mr(mr);
1481 	if (!ret)
1482 		atomic_dec(&pd->usecnt);
1483 
1484 	return ret;
1485 }
1486 EXPORT_SYMBOL(ib_dereg_mr);
1487 
1488 /**
1489  * ib_alloc_mr() - Allocates a memory region
1490  * @pd:            protection domain associated with the region
1491  * @mr_type:       memory region type
1492  * @max_num_sg:    maximum sg entries available for registration.
1493  *
1494  * Notes:
1495  * Memory registeration page/sg lists must not exceed max_num_sg.
1496  * For mr_type IB_MR_TYPE_MEM_REG, the total length cannot exceed
1497  * max_num_sg * used_page_size.
1498  *
1499  */
1500 struct ib_mr *ib_alloc_mr(struct ib_pd *pd,
1501 			  enum ib_mr_type mr_type,
1502 			  u32 max_num_sg)
1503 {
1504 	struct ib_mr *mr;
1505 
1506 	if (!pd->device->alloc_mr)
1507 		return ERR_PTR(-ENOSYS);
1508 
1509 	mr = pd->device->alloc_mr(pd, mr_type, max_num_sg);
1510 	if (!IS_ERR(mr)) {
1511 		mr->device  = pd->device;
1512 		mr->pd      = pd;
1513 		mr->uobject = NULL;
1514 		atomic_inc(&pd->usecnt);
1515 		mr->need_inval = false;
1516 	}
1517 
1518 	return mr;
1519 }
1520 EXPORT_SYMBOL(ib_alloc_mr);
1521 
1522 /* "Fast" memory regions */
1523 
1524 struct ib_fmr *ib_alloc_fmr(struct ib_pd *pd,
1525 			    int mr_access_flags,
1526 			    struct ib_fmr_attr *fmr_attr)
1527 {
1528 	struct ib_fmr *fmr;
1529 
1530 	if (!pd->device->alloc_fmr)
1531 		return ERR_PTR(-ENOSYS);
1532 
1533 	fmr = pd->device->alloc_fmr(pd, mr_access_flags, fmr_attr);
1534 	if (!IS_ERR(fmr)) {
1535 		fmr->device = pd->device;
1536 		fmr->pd     = pd;
1537 		atomic_inc(&pd->usecnt);
1538 	}
1539 
1540 	return fmr;
1541 }
1542 EXPORT_SYMBOL(ib_alloc_fmr);
1543 
1544 int ib_unmap_fmr(struct list_head *fmr_list)
1545 {
1546 	struct ib_fmr *fmr;
1547 
1548 	if (list_empty(fmr_list))
1549 		return 0;
1550 
1551 	fmr = list_entry(fmr_list->next, struct ib_fmr, list);
1552 	return fmr->device->unmap_fmr(fmr_list);
1553 }
1554 EXPORT_SYMBOL(ib_unmap_fmr);
1555 
1556 int ib_dealloc_fmr(struct ib_fmr *fmr)
1557 {
1558 	struct ib_pd *pd;
1559 	int ret;
1560 
1561 	pd = fmr->pd;
1562 	ret = fmr->device->dealloc_fmr(fmr);
1563 	if (!ret)
1564 		atomic_dec(&pd->usecnt);
1565 
1566 	return ret;
1567 }
1568 EXPORT_SYMBOL(ib_dealloc_fmr);
1569 
1570 /* Multicast groups */
1571 
1572 int ib_attach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid)
1573 {
1574 	int ret;
1575 
1576 	if (!qp->device->attach_mcast)
1577 		return -ENOSYS;
1578 	if (gid->raw[0] != 0xff || qp->qp_type != IB_QPT_UD ||
1579 	    lid < be16_to_cpu(IB_MULTICAST_LID_BASE) ||
1580 	    lid == be16_to_cpu(IB_LID_PERMISSIVE))
1581 		return -EINVAL;
1582 
1583 	ret = qp->device->attach_mcast(qp, gid, lid);
1584 	if (!ret)
1585 		atomic_inc(&qp->usecnt);
1586 	return ret;
1587 }
1588 EXPORT_SYMBOL(ib_attach_mcast);
1589 
1590 int ib_detach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid)
1591 {
1592 	int ret;
1593 
1594 	if (!qp->device->detach_mcast)
1595 		return -ENOSYS;
1596 	if (gid->raw[0] != 0xff || qp->qp_type != IB_QPT_UD ||
1597 	    lid < be16_to_cpu(IB_MULTICAST_LID_BASE) ||
1598 	    lid == be16_to_cpu(IB_LID_PERMISSIVE))
1599 		return -EINVAL;
1600 
1601 	ret = qp->device->detach_mcast(qp, gid, lid);
1602 	if (!ret)
1603 		atomic_dec(&qp->usecnt);
1604 	return ret;
1605 }
1606 EXPORT_SYMBOL(ib_detach_mcast);
1607 
1608 struct ib_xrcd *ib_alloc_xrcd(struct ib_device *device)
1609 {
1610 	struct ib_xrcd *xrcd;
1611 
1612 	if (!device->alloc_xrcd)
1613 		return ERR_PTR(-ENOSYS);
1614 
1615 	xrcd = device->alloc_xrcd(device, NULL, NULL);
1616 	if (!IS_ERR(xrcd)) {
1617 		xrcd->device = device;
1618 		xrcd->inode = NULL;
1619 		atomic_set(&xrcd->usecnt, 0);
1620 		mutex_init(&xrcd->tgt_qp_mutex);
1621 		INIT_LIST_HEAD(&xrcd->tgt_qp_list);
1622 	}
1623 
1624 	return xrcd;
1625 }
1626 EXPORT_SYMBOL(ib_alloc_xrcd);
1627 
1628 int ib_dealloc_xrcd(struct ib_xrcd *xrcd)
1629 {
1630 	struct ib_qp *qp;
1631 	int ret;
1632 
1633 	if (atomic_read(&xrcd->usecnt))
1634 		return -EBUSY;
1635 
1636 	while (!list_empty(&xrcd->tgt_qp_list)) {
1637 		qp = list_entry(xrcd->tgt_qp_list.next, struct ib_qp, xrcd_list);
1638 		ret = ib_destroy_qp(qp);
1639 		if (ret)
1640 			return ret;
1641 	}
1642 
1643 	return xrcd->device->dealloc_xrcd(xrcd);
1644 }
1645 EXPORT_SYMBOL(ib_dealloc_xrcd);
1646 
1647 /**
1648  * ib_create_wq - Creates a WQ associated with the specified protection
1649  * domain.
1650  * @pd: The protection domain associated with the WQ.
1651  * @wq_init_attr: A list of initial attributes required to create the
1652  * WQ. If WQ creation succeeds, then the attributes are updated to
1653  * the actual capabilities of the created WQ.
1654  *
1655  * wq_init_attr->max_wr and wq_init_attr->max_sge determine
1656  * the requested size of the WQ, and set to the actual values allocated
1657  * on return.
1658  * If ib_create_wq() succeeds, then max_wr and max_sge will always be
1659  * at least as large as the requested values.
1660  */
1661 struct ib_wq *ib_create_wq(struct ib_pd *pd,
1662 			   struct ib_wq_init_attr *wq_attr)
1663 {
1664 	struct ib_wq *wq;
1665 
1666 	if (!pd->device->create_wq)
1667 		return ERR_PTR(-ENOSYS);
1668 
1669 	wq = pd->device->create_wq(pd, wq_attr, NULL);
1670 	if (!IS_ERR(wq)) {
1671 		wq->event_handler = wq_attr->event_handler;
1672 		wq->wq_context = wq_attr->wq_context;
1673 		wq->wq_type = wq_attr->wq_type;
1674 		wq->cq = wq_attr->cq;
1675 		wq->device = pd->device;
1676 		wq->pd = pd;
1677 		wq->uobject = NULL;
1678 		atomic_inc(&pd->usecnt);
1679 		atomic_inc(&wq_attr->cq->usecnt);
1680 		atomic_set(&wq->usecnt, 0);
1681 	}
1682 	return wq;
1683 }
1684 EXPORT_SYMBOL(ib_create_wq);
1685 
1686 /**
1687  * ib_destroy_wq - Destroys the specified WQ.
1688  * @wq: The WQ to destroy.
1689  */
1690 int ib_destroy_wq(struct ib_wq *wq)
1691 {
1692 	int err;
1693 	struct ib_cq *cq = wq->cq;
1694 	struct ib_pd *pd = wq->pd;
1695 
1696 	if (atomic_read(&wq->usecnt))
1697 		return -EBUSY;
1698 
1699 	err = wq->device->destroy_wq(wq);
1700 	if (!err) {
1701 		atomic_dec(&pd->usecnt);
1702 		atomic_dec(&cq->usecnt);
1703 	}
1704 	return err;
1705 }
1706 EXPORT_SYMBOL(ib_destroy_wq);
1707 
1708 /**
1709  * ib_modify_wq - Modifies the specified WQ.
1710  * @wq: The WQ to modify.
1711  * @wq_attr: On input, specifies the WQ attributes to modify.
1712  * @wq_attr_mask: A bit-mask used to specify which attributes of the WQ
1713  *   are being modified.
1714  * On output, the current values of selected WQ attributes are returned.
1715  */
1716 int ib_modify_wq(struct ib_wq *wq, struct ib_wq_attr *wq_attr,
1717 		 u32 wq_attr_mask)
1718 {
1719 	int err;
1720 
1721 	if (!wq->device->modify_wq)
1722 		return -ENOSYS;
1723 
1724 	err = wq->device->modify_wq(wq, wq_attr, wq_attr_mask, NULL);
1725 	return err;
1726 }
1727 EXPORT_SYMBOL(ib_modify_wq);
1728 
1729 /*
1730  * ib_create_rwq_ind_table - Creates a RQ Indirection Table.
1731  * @device: The device on which to create the rwq indirection table.
1732  * @ib_rwq_ind_table_init_attr: A list of initial attributes required to
1733  * create the Indirection Table.
1734  *
1735  * Note: The life time of ib_rwq_ind_table_init_attr->ind_tbl is not less
1736  *	than the created ib_rwq_ind_table object and the caller is responsible
1737  *	for its memory allocation/free.
1738  */
1739 struct ib_rwq_ind_table *ib_create_rwq_ind_table(struct ib_device *device,
1740 						 struct ib_rwq_ind_table_init_attr *init_attr)
1741 {
1742 	struct ib_rwq_ind_table *rwq_ind_table;
1743 	int i;
1744 	u32 table_size;
1745 
1746 	if (!device->create_rwq_ind_table)
1747 		return ERR_PTR(-ENOSYS);
1748 
1749 	table_size = (1 << init_attr->log_ind_tbl_size);
1750 	rwq_ind_table = device->create_rwq_ind_table(device,
1751 				init_attr, NULL);
1752 	if (IS_ERR(rwq_ind_table))
1753 		return rwq_ind_table;
1754 
1755 	rwq_ind_table->ind_tbl = init_attr->ind_tbl;
1756 	rwq_ind_table->log_ind_tbl_size = init_attr->log_ind_tbl_size;
1757 	rwq_ind_table->device = device;
1758 	rwq_ind_table->uobject = NULL;
1759 	atomic_set(&rwq_ind_table->usecnt, 0);
1760 
1761 	for (i = 0; i < table_size; i++)
1762 		atomic_inc(&rwq_ind_table->ind_tbl[i]->usecnt);
1763 
1764 	return rwq_ind_table;
1765 }
1766 EXPORT_SYMBOL(ib_create_rwq_ind_table);
1767 
1768 /*
1769  * ib_destroy_rwq_ind_table - Destroys the specified Indirection Table.
1770  * @wq_ind_table: The Indirection Table to destroy.
1771 */
1772 int ib_destroy_rwq_ind_table(struct ib_rwq_ind_table *rwq_ind_table)
1773 {
1774 	int err, i;
1775 	u32 table_size = (1 << rwq_ind_table->log_ind_tbl_size);
1776 	struct ib_wq **ind_tbl = rwq_ind_table->ind_tbl;
1777 
1778 	if (atomic_read(&rwq_ind_table->usecnt))
1779 		return -EBUSY;
1780 
1781 	err = rwq_ind_table->device->destroy_rwq_ind_table(rwq_ind_table);
1782 	if (!err) {
1783 		for (i = 0; i < table_size; i++)
1784 			atomic_dec(&ind_tbl[i]->usecnt);
1785 	}
1786 
1787 	return err;
1788 }
1789 EXPORT_SYMBOL(ib_destroy_rwq_ind_table);
1790 
1791 struct ib_flow *ib_create_flow(struct ib_qp *qp,
1792 			       struct ib_flow_attr *flow_attr,
1793 			       int domain)
1794 {
1795 	struct ib_flow *flow_id;
1796 	if (!qp->device->create_flow)
1797 		return ERR_PTR(-ENOSYS);
1798 
1799 	flow_id = qp->device->create_flow(qp, flow_attr, domain);
1800 	if (!IS_ERR(flow_id)) {
1801 		atomic_inc(&qp->usecnt);
1802 		flow_id->qp = qp;
1803 	}
1804 	return flow_id;
1805 }
1806 EXPORT_SYMBOL(ib_create_flow);
1807 
1808 int ib_destroy_flow(struct ib_flow *flow_id)
1809 {
1810 	int err;
1811 	struct ib_qp *qp = flow_id->qp;
1812 
1813 	err = qp->device->destroy_flow(flow_id);
1814 	if (!err)
1815 		atomic_dec(&qp->usecnt);
1816 	return err;
1817 }
1818 EXPORT_SYMBOL(ib_destroy_flow);
1819 
1820 int ib_check_mr_status(struct ib_mr *mr, u32 check_mask,
1821 		       struct ib_mr_status *mr_status)
1822 {
1823 	return mr->device->check_mr_status ?
1824 		mr->device->check_mr_status(mr, check_mask, mr_status) : -ENOSYS;
1825 }
1826 EXPORT_SYMBOL(ib_check_mr_status);
1827 
1828 int ib_set_vf_link_state(struct ib_device *device, int vf, u8 port,
1829 			 int state)
1830 {
1831 	if (!device->set_vf_link_state)
1832 		return -ENOSYS;
1833 
1834 	return device->set_vf_link_state(device, vf, port, state);
1835 }
1836 EXPORT_SYMBOL(ib_set_vf_link_state);
1837 
1838 int ib_get_vf_config(struct ib_device *device, int vf, u8 port,
1839 		     struct ifla_vf_info *info)
1840 {
1841 	if (!device->get_vf_config)
1842 		return -ENOSYS;
1843 
1844 	return device->get_vf_config(device, vf, port, info);
1845 }
1846 EXPORT_SYMBOL(ib_get_vf_config);
1847 
1848 int ib_get_vf_stats(struct ib_device *device, int vf, u8 port,
1849 		    struct ifla_vf_stats *stats)
1850 {
1851 	if (!device->get_vf_stats)
1852 		return -ENOSYS;
1853 
1854 	return device->get_vf_stats(device, vf, port, stats);
1855 }
1856 EXPORT_SYMBOL(ib_get_vf_stats);
1857 
1858 int ib_set_vf_guid(struct ib_device *device, int vf, u8 port, u64 guid,
1859 		   int type)
1860 {
1861 	if (!device->set_vf_guid)
1862 		return -ENOSYS;
1863 
1864 	return device->set_vf_guid(device, vf, port, guid, type);
1865 }
1866 EXPORT_SYMBOL(ib_set_vf_guid);
1867 
1868 /**
1869  * ib_map_mr_sg() - Map the largest prefix of a dma mapped SG list
1870  *     and set it the memory region.
1871  * @mr:            memory region
1872  * @sg:            dma mapped scatterlist
1873  * @sg_nents:      number of entries in sg
1874  * @sg_offset:     offset in bytes into sg
1875  * @page_size:     page vector desired page size
1876  *
1877  * Constraints:
1878  * - The first sg element is allowed to have an offset.
1879  * - Each sg element must either be aligned to page_size or virtually
1880  *   contiguous to the previous element. In case an sg element has a
1881  *   non-contiguous offset, the mapping prefix will not include it.
1882  * - The last sg element is allowed to have length less than page_size.
1883  * - If sg_nents total byte length exceeds the mr max_num_sge * page_size
1884  *   then only max_num_sg entries will be mapped.
1885  * - If the MR was allocated with type IB_MR_TYPE_SG_GAPS, none of these
1886  *   constraints holds and the page_size argument is ignored.
1887  *
1888  * Returns the number of sg elements that were mapped to the memory region.
1889  *
1890  * After this completes successfully, the  memory region
1891  * is ready for registration.
1892  */
1893 int ib_map_mr_sg(struct ib_mr *mr, struct scatterlist *sg, int sg_nents,
1894 		 unsigned int *sg_offset, unsigned int page_size)
1895 {
1896 	if (unlikely(!mr->device->map_mr_sg))
1897 		return -ENOSYS;
1898 
1899 	mr->page_size = page_size;
1900 
1901 	return mr->device->map_mr_sg(mr, sg, sg_nents, sg_offset);
1902 }
1903 EXPORT_SYMBOL(ib_map_mr_sg);
1904 
1905 /**
1906  * ib_sg_to_pages() - Convert the largest prefix of a sg list
1907  *     to a page vector
1908  * @mr:            memory region
1909  * @sgl:           dma mapped scatterlist
1910  * @sg_nents:      number of entries in sg
1911  * @sg_offset_p:   IN:  start offset in bytes into sg
1912  *                 OUT: offset in bytes for element n of the sg of the first
1913  *                      byte that has not been processed where n is the return
1914  *                      value of this function.
1915  * @set_page:      driver page assignment function pointer
1916  *
1917  * Core service helper for drivers to convert the largest
1918  * prefix of given sg list to a page vector. The sg list
1919  * prefix converted is the prefix that meet the requirements
1920  * of ib_map_mr_sg.
1921  *
1922  * Returns the number of sg elements that were assigned to
1923  * a page vector.
1924  */
1925 int ib_sg_to_pages(struct ib_mr *mr, struct scatterlist *sgl, int sg_nents,
1926 		unsigned int *sg_offset_p, int (*set_page)(struct ib_mr *, u64))
1927 {
1928 	struct scatterlist *sg;
1929 	u64 last_end_dma_addr = 0;
1930 	unsigned int sg_offset = sg_offset_p ? *sg_offset_p : 0;
1931 	unsigned int last_page_off = 0;
1932 	u64 page_mask = ~((u64)mr->page_size - 1);
1933 	int i, ret;
1934 
1935 	if (unlikely(sg_nents <= 0 || sg_offset > sg_dma_len(&sgl[0])))
1936 		return -EINVAL;
1937 
1938 	mr->iova = sg_dma_address(&sgl[0]) + sg_offset;
1939 	mr->length = 0;
1940 
1941 	for_each_sg(sgl, sg, sg_nents, i) {
1942 		u64 dma_addr = sg_dma_address(sg) + sg_offset;
1943 		u64 prev_addr = dma_addr;
1944 		unsigned int dma_len = sg_dma_len(sg) - sg_offset;
1945 		u64 end_dma_addr = dma_addr + dma_len;
1946 		u64 page_addr = dma_addr & page_mask;
1947 
1948 		/*
1949 		 * For the second and later elements, check whether either the
1950 		 * end of element i-1 or the start of element i is not aligned
1951 		 * on a page boundary.
1952 		 */
1953 		if (i && (last_page_off != 0 || page_addr != dma_addr)) {
1954 			/* Stop mapping if there is a gap. */
1955 			if (last_end_dma_addr != dma_addr)
1956 				break;
1957 
1958 			/*
1959 			 * Coalesce this element with the last. If it is small
1960 			 * enough just update mr->length. Otherwise start
1961 			 * mapping from the next page.
1962 			 */
1963 			goto next_page;
1964 		}
1965 
1966 		do {
1967 			ret = set_page(mr, page_addr);
1968 			if (unlikely(ret < 0)) {
1969 				sg_offset = prev_addr - sg_dma_address(sg);
1970 				mr->length += prev_addr - dma_addr;
1971 				if (sg_offset_p)
1972 					*sg_offset_p = sg_offset;
1973 				return i || sg_offset ? i : ret;
1974 			}
1975 			prev_addr = page_addr;
1976 next_page:
1977 			page_addr += mr->page_size;
1978 		} while (page_addr < end_dma_addr);
1979 
1980 		mr->length += dma_len;
1981 		last_end_dma_addr = end_dma_addr;
1982 		last_page_off = end_dma_addr & ~page_mask;
1983 
1984 		sg_offset = 0;
1985 	}
1986 
1987 	if (sg_offset_p)
1988 		*sg_offset_p = 0;
1989 	return i;
1990 }
1991 EXPORT_SYMBOL(ib_sg_to_pages);
1992 
1993 struct ib_drain_cqe {
1994 	struct ib_cqe cqe;
1995 	struct completion done;
1996 };
1997 
1998 static void ib_drain_qp_done(struct ib_cq *cq, struct ib_wc *wc)
1999 {
2000 	struct ib_drain_cqe *cqe = container_of(wc->wr_cqe, struct ib_drain_cqe,
2001 						cqe);
2002 
2003 	complete(&cqe->done);
2004 }
2005 
2006 /*
2007  * Post a WR and block until its completion is reaped for the SQ.
2008  */
2009 static void __ib_drain_sq(struct ib_qp *qp)
2010 {
2011 	struct ib_cq *cq = qp->send_cq;
2012 	struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
2013 	struct ib_drain_cqe sdrain;
2014 	struct ib_send_wr swr = {}, *bad_swr;
2015 	int ret;
2016 
2017 	swr.wr_cqe = &sdrain.cqe;
2018 	sdrain.cqe.done = ib_drain_qp_done;
2019 	init_completion(&sdrain.done);
2020 
2021 	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
2022 	if (ret) {
2023 		WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
2024 		return;
2025 	}
2026 
2027 	ret = ib_post_send(qp, &swr, &bad_swr);
2028 	if (ret) {
2029 		WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
2030 		return;
2031 	}
2032 
2033 	if (cq->poll_ctx == IB_POLL_DIRECT)
2034 		while (wait_for_completion_timeout(&sdrain.done, HZ / 10) <= 0)
2035 			ib_process_cq_direct(cq, -1);
2036 	else
2037 		wait_for_completion(&sdrain.done);
2038 }
2039 
2040 /*
2041  * Post a WR and block until its completion is reaped for the RQ.
2042  */
2043 static void __ib_drain_rq(struct ib_qp *qp)
2044 {
2045 	struct ib_cq *cq = qp->recv_cq;
2046 	struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
2047 	struct ib_drain_cqe rdrain;
2048 	struct ib_recv_wr rwr = {}, *bad_rwr;
2049 	int ret;
2050 
2051 	rwr.wr_cqe = &rdrain.cqe;
2052 	rdrain.cqe.done = ib_drain_qp_done;
2053 	init_completion(&rdrain.done);
2054 
2055 	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
2056 	if (ret) {
2057 		WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
2058 		return;
2059 	}
2060 
2061 	ret = ib_post_recv(qp, &rwr, &bad_rwr);
2062 	if (ret) {
2063 		WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
2064 		return;
2065 	}
2066 
2067 	if (cq->poll_ctx == IB_POLL_DIRECT)
2068 		while (wait_for_completion_timeout(&rdrain.done, HZ / 10) <= 0)
2069 			ib_process_cq_direct(cq, -1);
2070 	else
2071 		wait_for_completion(&rdrain.done);
2072 }
2073 
2074 /**
2075  * ib_drain_sq() - Block until all SQ CQEs have been consumed by the
2076  *		   application.
2077  * @qp:            queue pair to drain
2078  *
2079  * If the device has a provider-specific drain function, then
2080  * call that.  Otherwise call the generic drain function
2081  * __ib_drain_sq().
2082  *
2083  * The caller must:
2084  *
2085  * ensure there is room in the CQ and SQ for the drain work request and
2086  * completion.
2087  *
2088  * allocate the CQ using ib_alloc_cq().
2089  *
2090  * ensure that there are no other contexts that are posting WRs concurrently.
2091  * Otherwise the drain is not guaranteed.
2092  */
2093 void ib_drain_sq(struct ib_qp *qp)
2094 {
2095 	if (qp->device->drain_sq)
2096 		qp->device->drain_sq(qp);
2097 	else
2098 		__ib_drain_sq(qp);
2099 }
2100 EXPORT_SYMBOL(ib_drain_sq);
2101 
2102 /**
2103  * ib_drain_rq() - Block until all RQ CQEs have been consumed by the
2104  *		   application.
2105  * @qp:            queue pair to drain
2106  *
2107  * If the device has a provider-specific drain function, then
2108  * call that.  Otherwise call the generic drain function
2109  * __ib_drain_rq().
2110  *
2111  * The caller must:
2112  *
2113  * ensure there is room in the CQ and RQ for the drain work request and
2114  * completion.
2115  *
2116  * allocate the CQ using ib_alloc_cq().
2117  *
2118  * ensure that there are no other contexts that are posting WRs concurrently.
2119  * Otherwise the drain is not guaranteed.
2120  */
2121 void ib_drain_rq(struct ib_qp *qp)
2122 {
2123 	if (qp->device->drain_rq)
2124 		qp->device->drain_rq(qp);
2125 	else
2126 		__ib_drain_rq(qp);
2127 }
2128 EXPORT_SYMBOL(ib_drain_rq);
2129 
2130 /**
2131  * ib_drain_qp() - Block until all CQEs have been consumed by the
2132  *		   application on both the RQ and SQ.
2133  * @qp:            queue pair to drain
2134  *
2135  * The caller must:
2136  *
2137  * ensure there is room in the CQ(s), SQ, and RQ for drain work requests
2138  * and completions.
2139  *
2140  * allocate the CQs using ib_alloc_cq().
2141  *
2142  * ensure that there are no other contexts that are posting WRs concurrently.
2143  * Otherwise the drain is not guaranteed.
2144  */
2145 void ib_drain_qp(struct ib_qp *qp)
2146 {
2147 	ib_drain_sq(qp);
2148 	if (!qp->srq)
2149 		ib_drain_rq(qp);
2150 }
2151 EXPORT_SYMBOL(ib_drain_qp);
2152