xref: /linux/drivers/infiniband/core/verbs.c (revision e58e871becec2d3b04ed91c0c16fe8deac9c9dfa)
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 
48 #include <rdma/ib_verbs.h>
49 #include <rdma/ib_cache.h>
50 #include <rdma/ib_addr.h>
51 #include <rdma/rw.h>
52 
53 #include "core_priv.h"
54 
55 static const char * const ib_events[] = {
56 	[IB_EVENT_CQ_ERR]		= "CQ error",
57 	[IB_EVENT_QP_FATAL]		= "QP fatal error",
58 	[IB_EVENT_QP_REQ_ERR]		= "QP request error",
59 	[IB_EVENT_QP_ACCESS_ERR]	= "QP access error",
60 	[IB_EVENT_COMM_EST]		= "communication established",
61 	[IB_EVENT_SQ_DRAINED]		= "send queue drained",
62 	[IB_EVENT_PATH_MIG]		= "path migration successful",
63 	[IB_EVENT_PATH_MIG_ERR]		= "path migration error",
64 	[IB_EVENT_DEVICE_FATAL]		= "device fatal error",
65 	[IB_EVENT_PORT_ACTIVE]		= "port active",
66 	[IB_EVENT_PORT_ERR]		= "port error",
67 	[IB_EVENT_LID_CHANGE]		= "LID change",
68 	[IB_EVENT_PKEY_CHANGE]		= "P_key change",
69 	[IB_EVENT_SM_CHANGE]		= "SM change",
70 	[IB_EVENT_SRQ_ERR]		= "SRQ error",
71 	[IB_EVENT_SRQ_LIMIT_REACHED]	= "SRQ limit reached",
72 	[IB_EVENT_QP_LAST_WQE_REACHED]	= "last WQE reached",
73 	[IB_EVENT_CLIENT_REREGISTER]	= "client reregister",
74 	[IB_EVENT_GID_CHANGE]		= "GID changed",
75 };
76 
77 const char *__attribute_const__ ib_event_msg(enum ib_event_type event)
78 {
79 	size_t index = event;
80 
81 	return (index < ARRAY_SIZE(ib_events) && ib_events[index]) ?
82 			ib_events[index] : "unrecognized event";
83 }
84 EXPORT_SYMBOL(ib_event_msg);
85 
86 static const char * const wc_statuses[] = {
87 	[IB_WC_SUCCESS]			= "success",
88 	[IB_WC_LOC_LEN_ERR]		= "local length error",
89 	[IB_WC_LOC_QP_OP_ERR]		= "local QP operation error",
90 	[IB_WC_LOC_EEC_OP_ERR]		= "local EE context operation error",
91 	[IB_WC_LOC_PROT_ERR]		= "local protection error",
92 	[IB_WC_WR_FLUSH_ERR]		= "WR flushed",
93 	[IB_WC_MW_BIND_ERR]		= "memory management operation error",
94 	[IB_WC_BAD_RESP_ERR]		= "bad response error",
95 	[IB_WC_LOC_ACCESS_ERR]		= "local access error",
96 	[IB_WC_REM_INV_REQ_ERR]		= "invalid request error",
97 	[IB_WC_REM_ACCESS_ERR]		= "remote access error",
98 	[IB_WC_REM_OP_ERR]		= "remote operation error",
99 	[IB_WC_RETRY_EXC_ERR]		= "transport retry counter exceeded",
100 	[IB_WC_RNR_RETRY_EXC_ERR]	= "RNR retry counter exceeded",
101 	[IB_WC_LOC_RDD_VIOL_ERR]	= "local RDD violation error",
102 	[IB_WC_REM_INV_RD_REQ_ERR]	= "remote invalid RD request",
103 	[IB_WC_REM_ABORT_ERR]		= "operation aborted",
104 	[IB_WC_INV_EECN_ERR]		= "invalid EE context number",
105 	[IB_WC_INV_EEC_STATE_ERR]	= "invalid EE context state",
106 	[IB_WC_FATAL_ERR]		= "fatal error",
107 	[IB_WC_RESP_TIMEOUT_ERR]	= "response timeout error",
108 	[IB_WC_GENERAL_ERR]		= "general error",
109 };
110 
111 const char *__attribute_const__ ib_wc_status_msg(enum ib_wc_status status)
112 {
113 	size_t index = status;
114 
115 	return (index < ARRAY_SIZE(wc_statuses) && wc_statuses[index]) ?
116 			wc_statuses[index] : "unrecognized status";
117 }
118 EXPORT_SYMBOL(ib_wc_status_msg);
119 
120 __attribute_const__ int ib_rate_to_mult(enum ib_rate rate)
121 {
122 	switch (rate) {
123 	case IB_RATE_2_5_GBPS: return  1;
124 	case IB_RATE_5_GBPS:   return  2;
125 	case IB_RATE_10_GBPS:  return  4;
126 	case IB_RATE_20_GBPS:  return  8;
127 	case IB_RATE_30_GBPS:  return 12;
128 	case IB_RATE_40_GBPS:  return 16;
129 	case IB_RATE_60_GBPS:  return 24;
130 	case IB_RATE_80_GBPS:  return 32;
131 	case IB_RATE_120_GBPS: return 48;
132 	default:	       return -1;
133 	}
134 }
135 EXPORT_SYMBOL(ib_rate_to_mult);
136 
137 __attribute_const__ enum ib_rate mult_to_ib_rate(int mult)
138 {
139 	switch (mult) {
140 	case 1:  return IB_RATE_2_5_GBPS;
141 	case 2:  return IB_RATE_5_GBPS;
142 	case 4:  return IB_RATE_10_GBPS;
143 	case 8:  return IB_RATE_20_GBPS;
144 	case 12: return IB_RATE_30_GBPS;
145 	case 16: return IB_RATE_40_GBPS;
146 	case 24: return IB_RATE_60_GBPS;
147 	case 32: return IB_RATE_80_GBPS;
148 	case 48: return IB_RATE_120_GBPS;
149 	default: return IB_RATE_PORT_CURRENT;
150 	}
151 }
152 EXPORT_SYMBOL(mult_to_ib_rate);
153 
154 __attribute_const__ int ib_rate_to_mbps(enum ib_rate rate)
155 {
156 	switch (rate) {
157 	case IB_RATE_2_5_GBPS: return 2500;
158 	case IB_RATE_5_GBPS:   return 5000;
159 	case IB_RATE_10_GBPS:  return 10000;
160 	case IB_RATE_20_GBPS:  return 20000;
161 	case IB_RATE_30_GBPS:  return 30000;
162 	case IB_RATE_40_GBPS:  return 40000;
163 	case IB_RATE_60_GBPS:  return 60000;
164 	case IB_RATE_80_GBPS:  return 80000;
165 	case IB_RATE_120_GBPS: return 120000;
166 	case IB_RATE_14_GBPS:  return 14062;
167 	case IB_RATE_56_GBPS:  return 56250;
168 	case IB_RATE_112_GBPS: return 112500;
169 	case IB_RATE_168_GBPS: return 168750;
170 	case IB_RATE_25_GBPS:  return 25781;
171 	case IB_RATE_100_GBPS: return 103125;
172 	case IB_RATE_200_GBPS: return 206250;
173 	case IB_RATE_300_GBPS: return 309375;
174 	default:	       return -1;
175 	}
176 }
177 EXPORT_SYMBOL(ib_rate_to_mbps);
178 
179 __attribute_const__ enum rdma_transport_type
180 rdma_node_get_transport(enum rdma_node_type node_type)
181 {
182 	switch (node_type) {
183 	case RDMA_NODE_IB_CA:
184 	case RDMA_NODE_IB_SWITCH:
185 	case RDMA_NODE_IB_ROUTER:
186 		return RDMA_TRANSPORT_IB;
187 	case RDMA_NODE_RNIC:
188 		return RDMA_TRANSPORT_IWARP;
189 	case RDMA_NODE_USNIC:
190 		return RDMA_TRANSPORT_USNIC;
191 	case RDMA_NODE_USNIC_UDP:
192 		return RDMA_TRANSPORT_USNIC_UDP;
193 	default:
194 		BUG();
195 		return 0;
196 	}
197 }
198 EXPORT_SYMBOL(rdma_node_get_transport);
199 
200 enum rdma_link_layer rdma_port_get_link_layer(struct ib_device *device, u8 port_num)
201 {
202 	if (device->get_link_layer)
203 		return device->get_link_layer(device, port_num);
204 
205 	switch (rdma_node_get_transport(device->node_type)) {
206 	case RDMA_TRANSPORT_IB:
207 		return IB_LINK_LAYER_INFINIBAND;
208 	case RDMA_TRANSPORT_IWARP:
209 	case RDMA_TRANSPORT_USNIC:
210 	case RDMA_TRANSPORT_USNIC_UDP:
211 		return IB_LINK_LAYER_ETHERNET;
212 	default:
213 		return IB_LINK_LAYER_UNSPECIFIED;
214 	}
215 }
216 EXPORT_SYMBOL(rdma_port_get_link_layer);
217 
218 /* Protection domains */
219 
220 /**
221  * ib_alloc_pd - Allocates an unused protection domain.
222  * @device: The device on which to allocate the protection domain.
223  *
224  * A protection domain object provides an association between QPs, shared
225  * receive queues, address handles, memory regions, and memory windows.
226  *
227  * Every PD has a local_dma_lkey which can be used as the lkey value for local
228  * memory operations.
229  */
230 struct ib_pd *__ib_alloc_pd(struct ib_device *device, unsigned int flags,
231 		const char *caller)
232 {
233 	struct ib_pd *pd;
234 	int mr_access_flags = 0;
235 
236 	pd = device->alloc_pd(device, NULL, NULL);
237 	if (IS_ERR(pd))
238 		return pd;
239 
240 	pd->device = device;
241 	pd->uobject = NULL;
242 	pd->__internal_mr = NULL;
243 	atomic_set(&pd->usecnt, 0);
244 	pd->flags = flags;
245 
246 	if (device->attrs.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY)
247 		pd->local_dma_lkey = device->local_dma_lkey;
248 	else
249 		mr_access_flags |= IB_ACCESS_LOCAL_WRITE;
250 
251 	if (flags & IB_PD_UNSAFE_GLOBAL_RKEY) {
252 		pr_warn("%s: enabling unsafe global rkey\n", caller);
253 		mr_access_flags |= IB_ACCESS_REMOTE_READ | IB_ACCESS_REMOTE_WRITE;
254 	}
255 
256 	if (mr_access_flags) {
257 		struct ib_mr *mr;
258 
259 		mr = pd->device->get_dma_mr(pd, mr_access_flags);
260 		if (IS_ERR(mr)) {
261 			ib_dealloc_pd(pd);
262 			return ERR_CAST(mr);
263 		}
264 
265 		mr->device	= pd->device;
266 		mr->pd		= pd;
267 		mr->uobject	= NULL;
268 		mr->need_inval	= false;
269 
270 		pd->__internal_mr = mr;
271 
272 		if (!(device->attrs.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY))
273 			pd->local_dma_lkey = pd->__internal_mr->lkey;
274 
275 		if (flags & IB_PD_UNSAFE_GLOBAL_RKEY)
276 			pd->unsafe_global_rkey = pd->__internal_mr->rkey;
277 	}
278 
279 	return pd;
280 }
281 EXPORT_SYMBOL(__ib_alloc_pd);
282 
283 /**
284  * ib_dealloc_pd - Deallocates a protection domain.
285  * @pd: The protection domain to deallocate.
286  *
287  * It is an error to call this function while any resources in the pd still
288  * exist.  The caller is responsible to synchronously destroy them and
289  * guarantee no new allocations will happen.
290  */
291 void ib_dealloc_pd(struct ib_pd *pd)
292 {
293 	int ret;
294 
295 	if (pd->__internal_mr) {
296 		ret = pd->device->dereg_mr(pd->__internal_mr);
297 		WARN_ON(ret);
298 		pd->__internal_mr = NULL;
299 	}
300 
301 	/* uverbs manipulates usecnt with proper locking, while the kabi
302 	   requires the caller to guarantee we can't race here. */
303 	WARN_ON(atomic_read(&pd->usecnt));
304 
305 	/* Making delalloc_pd a void return is a WIP, no driver should return
306 	   an error here. */
307 	ret = pd->device->dealloc_pd(pd);
308 	WARN_ONCE(ret, "Infiniband HW driver failed dealloc_pd");
309 }
310 EXPORT_SYMBOL(ib_dealloc_pd);
311 
312 /* Address handles */
313 
314 struct ib_ah *rdma_create_ah(struct ib_pd *pd, struct rdma_ah_attr *ah_attr)
315 {
316 	struct ib_ah *ah;
317 
318 	ah = pd->device->create_ah(pd, ah_attr, NULL);
319 
320 	if (!IS_ERR(ah)) {
321 		ah->device  = pd->device;
322 		ah->pd      = pd;
323 		ah->uobject = NULL;
324 		ah->type    = ah_attr->type;
325 		atomic_inc(&pd->usecnt);
326 	}
327 
328 	return ah;
329 }
330 EXPORT_SYMBOL(rdma_create_ah);
331 
332 int ib_get_rdma_header_version(const union rdma_network_hdr *hdr)
333 {
334 	const struct iphdr *ip4h = (struct iphdr *)&hdr->roce4grh;
335 	struct iphdr ip4h_checked;
336 	const struct ipv6hdr *ip6h = (struct ipv6hdr *)&hdr->ibgrh;
337 
338 	/* If it's IPv6, the version must be 6, otherwise, the first
339 	 * 20 bytes (before the IPv4 header) are garbled.
340 	 */
341 	if (ip6h->version != 6)
342 		return (ip4h->version == 4) ? 4 : 0;
343 	/* version may be 6 or 4 because the first 20 bytes could be garbled */
344 
345 	/* RoCE v2 requires no options, thus header length
346 	 * must be 5 words
347 	 */
348 	if (ip4h->ihl != 5)
349 		return 6;
350 
351 	/* Verify checksum.
352 	 * We can't write on scattered buffers so we need to copy to
353 	 * temp buffer.
354 	 */
355 	memcpy(&ip4h_checked, ip4h, sizeof(ip4h_checked));
356 	ip4h_checked.check = 0;
357 	ip4h_checked.check = ip_fast_csum((u8 *)&ip4h_checked, 5);
358 	/* if IPv4 header checksum is OK, believe it */
359 	if (ip4h->check == ip4h_checked.check)
360 		return 4;
361 	return 6;
362 }
363 EXPORT_SYMBOL(ib_get_rdma_header_version);
364 
365 static enum rdma_network_type ib_get_net_type_by_grh(struct ib_device *device,
366 						     u8 port_num,
367 						     const struct ib_grh *grh)
368 {
369 	int grh_version;
370 
371 	if (rdma_protocol_ib(device, port_num))
372 		return RDMA_NETWORK_IB;
373 
374 	grh_version = ib_get_rdma_header_version((union rdma_network_hdr *)grh);
375 
376 	if (grh_version == 4)
377 		return RDMA_NETWORK_IPV4;
378 
379 	if (grh->next_hdr == IPPROTO_UDP)
380 		return RDMA_NETWORK_IPV6;
381 
382 	return RDMA_NETWORK_ROCE_V1;
383 }
384 
385 struct find_gid_index_context {
386 	u16 vlan_id;
387 	enum ib_gid_type gid_type;
388 };
389 
390 static bool find_gid_index(const union ib_gid *gid,
391 			   const struct ib_gid_attr *gid_attr,
392 			   void *context)
393 {
394 	struct find_gid_index_context *ctx =
395 		(struct find_gid_index_context *)context;
396 
397 	if (ctx->gid_type != gid_attr->gid_type)
398 		return false;
399 
400 	if ((!!(ctx->vlan_id != 0xffff) == !is_vlan_dev(gid_attr->ndev)) ||
401 	    (is_vlan_dev(gid_attr->ndev) &&
402 	     vlan_dev_vlan_id(gid_attr->ndev) != ctx->vlan_id))
403 		return false;
404 
405 	return true;
406 }
407 
408 static int get_sgid_index_from_eth(struct ib_device *device, u8 port_num,
409 				   u16 vlan_id, const union ib_gid *sgid,
410 				   enum ib_gid_type gid_type,
411 				   u16 *gid_index)
412 {
413 	struct find_gid_index_context context = {.vlan_id = vlan_id,
414 						 .gid_type = gid_type};
415 
416 	return ib_find_gid_by_filter(device, sgid, port_num, find_gid_index,
417 				     &context, gid_index);
418 }
419 
420 int ib_get_gids_from_rdma_hdr(const union rdma_network_hdr *hdr,
421 			      enum rdma_network_type net_type,
422 			      union ib_gid *sgid, union ib_gid *dgid)
423 {
424 	struct sockaddr_in  src_in;
425 	struct sockaddr_in  dst_in;
426 	__be32 src_saddr, dst_saddr;
427 
428 	if (!sgid || !dgid)
429 		return -EINVAL;
430 
431 	if (net_type == RDMA_NETWORK_IPV4) {
432 		memcpy(&src_in.sin_addr.s_addr,
433 		       &hdr->roce4grh.saddr, 4);
434 		memcpy(&dst_in.sin_addr.s_addr,
435 		       &hdr->roce4grh.daddr, 4);
436 		src_saddr = src_in.sin_addr.s_addr;
437 		dst_saddr = dst_in.sin_addr.s_addr;
438 		ipv6_addr_set_v4mapped(src_saddr,
439 				       (struct in6_addr *)sgid);
440 		ipv6_addr_set_v4mapped(dst_saddr,
441 				       (struct in6_addr *)dgid);
442 		return 0;
443 	} else if (net_type == RDMA_NETWORK_IPV6 ||
444 		   net_type == RDMA_NETWORK_IB) {
445 		*dgid = hdr->ibgrh.dgid;
446 		*sgid = hdr->ibgrh.sgid;
447 		return 0;
448 	} else {
449 		return -EINVAL;
450 	}
451 }
452 EXPORT_SYMBOL(ib_get_gids_from_rdma_hdr);
453 
454 int ib_init_ah_from_wc(struct ib_device *device, u8 port_num,
455 		       const struct ib_wc *wc, const struct ib_grh *grh,
456 		       struct rdma_ah_attr *ah_attr)
457 {
458 	u32 flow_class;
459 	u16 gid_index;
460 	int ret;
461 	enum rdma_network_type net_type = RDMA_NETWORK_IB;
462 	enum ib_gid_type gid_type = IB_GID_TYPE_IB;
463 	int hoplimit = 0xff;
464 	union ib_gid dgid;
465 	union ib_gid sgid;
466 
467 	memset(ah_attr, 0, sizeof *ah_attr);
468 	ah_attr->type = rdma_ah_find_type(device, port_num);
469 	if (rdma_cap_eth_ah(device, port_num)) {
470 		if (wc->wc_flags & IB_WC_WITH_NETWORK_HDR_TYPE)
471 			net_type = wc->network_hdr_type;
472 		else
473 			net_type = ib_get_net_type_by_grh(device, port_num, grh);
474 		gid_type = ib_network_to_gid_type(net_type);
475 	}
476 	ret = ib_get_gids_from_rdma_hdr((union rdma_network_hdr *)grh, net_type,
477 					&sgid, &dgid);
478 	if (ret)
479 		return ret;
480 
481 	if (rdma_protocol_roce(device, port_num)) {
482 		int if_index = 0;
483 		u16 vlan_id = wc->wc_flags & IB_WC_WITH_VLAN ?
484 				wc->vlan_id : 0xffff;
485 		struct net_device *idev;
486 		struct net_device *resolved_dev;
487 
488 		if (!(wc->wc_flags & IB_WC_GRH))
489 			return -EPROTOTYPE;
490 
491 		if (!device->get_netdev)
492 			return -EOPNOTSUPP;
493 
494 		idev = device->get_netdev(device, port_num);
495 		if (!idev)
496 			return -ENODEV;
497 
498 		ret = rdma_addr_find_l2_eth_by_grh(&dgid, &sgid,
499 						   ah_attr->roce.dmac,
500 						   wc->wc_flags & IB_WC_WITH_VLAN ?
501 						   NULL : &vlan_id,
502 						   &if_index, &hoplimit);
503 		if (ret) {
504 			dev_put(idev);
505 			return ret;
506 		}
507 
508 		resolved_dev = dev_get_by_index(&init_net, if_index);
509 		if (resolved_dev->flags & IFF_LOOPBACK) {
510 			dev_put(resolved_dev);
511 			resolved_dev = idev;
512 			dev_hold(resolved_dev);
513 		}
514 		rcu_read_lock();
515 		if (resolved_dev != idev && !rdma_is_upper_dev_rcu(idev,
516 								   resolved_dev))
517 			ret = -EHOSTUNREACH;
518 		rcu_read_unlock();
519 		dev_put(idev);
520 		dev_put(resolved_dev);
521 		if (ret)
522 			return ret;
523 
524 		ret = get_sgid_index_from_eth(device, port_num, vlan_id,
525 					      &dgid, gid_type, &gid_index);
526 		if (ret)
527 			return ret;
528 	}
529 
530 	rdma_ah_set_dlid(ah_attr, wc->slid);
531 	rdma_ah_set_sl(ah_attr, wc->sl);
532 	rdma_ah_set_path_bits(ah_attr, wc->dlid_path_bits);
533 	rdma_ah_set_port_num(ah_attr, port_num);
534 
535 	if (wc->wc_flags & IB_WC_GRH) {
536 		if (!rdma_cap_eth_ah(device, port_num)) {
537 			if (dgid.global.interface_id != cpu_to_be64(IB_SA_WELL_KNOWN_GUID)) {
538 				ret = ib_find_cached_gid_by_port(device, &dgid,
539 								 IB_GID_TYPE_IB,
540 								 port_num, NULL,
541 								 &gid_index);
542 				if (ret)
543 					return ret;
544 			} else {
545 				gid_index = 0;
546 			}
547 		}
548 
549 		flow_class = be32_to_cpu(grh->version_tclass_flow);
550 		rdma_ah_set_grh(ah_attr, &sgid,
551 				flow_class & 0xFFFFF,
552 				(u8)gid_index, hoplimit,
553 				(flow_class >> 20) & 0xFF);
554 
555 	}
556 	return 0;
557 }
558 EXPORT_SYMBOL(ib_init_ah_from_wc);
559 
560 struct ib_ah *ib_create_ah_from_wc(struct ib_pd *pd, const struct ib_wc *wc,
561 				   const struct ib_grh *grh, u8 port_num)
562 {
563 	struct rdma_ah_attr ah_attr;
564 	int ret;
565 
566 	ret = ib_init_ah_from_wc(pd->device, port_num, wc, grh, &ah_attr);
567 	if (ret)
568 		return ERR_PTR(ret);
569 
570 	return rdma_create_ah(pd, &ah_attr);
571 }
572 EXPORT_SYMBOL(ib_create_ah_from_wc);
573 
574 int rdma_modify_ah(struct ib_ah *ah, struct rdma_ah_attr *ah_attr)
575 {
576 	if (ah->type != ah_attr->type)
577 		return -EINVAL;
578 
579 	return ah->device->modify_ah ?
580 		ah->device->modify_ah(ah, ah_attr) :
581 		-ENOSYS;
582 }
583 EXPORT_SYMBOL(rdma_modify_ah);
584 
585 int rdma_query_ah(struct ib_ah *ah, struct rdma_ah_attr *ah_attr)
586 {
587 	return ah->device->query_ah ?
588 		ah->device->query_ah(ah, ah_attr) :
589 		-ENOSYS;
590 }
591 EXPORT_SYMBOL(rdma_query_ah);
592 
593 int rdma_destroy_ah(struct ib_ah *ah)
594 {
595 	struct ib_pd *pd;
596 	int ret;
597 
598 	pd = ah->pd;
599 	ret = ah->device->destroy_ah(ah);
600 	if (!ret)
601 		atomic_dec(&pd->usecnt);
602 
603 	return ret;
604 }
605 EXPORT_SYMBOL(rdma_destroy_ah);
606 
607 /* Shared receive queues */
608 
609 struct ib_srq *ib_create_srq(struct ib_pd *pd,
610 			     struct ib_srq_init_attr *srq_init_attr)
611 {
612 	struct ib_srq *srq;
613 
614 	if (!pd->device->create_srq)
615 		return ERR_PTR(-ENOSYS);
616 
617 	srq = pd->device->create_srq(pd, srq_init_attr, NULL);
618 
619 	if (!IS_ERR(srq)) {
620 		srq->device    	   = pd->device;
621 		srq->pd        	   = pd;
622 		srq->uobject       = NULL;
623 		srq->event_handler = srq_init_attr->event_handler;
624 		srq->srq_context   = srq_init_attr->srq_context;
625 		srq->srq_type      = srq_init_attr->srq_type;
626 		if (srq->srq_type == IB_SRQT_XRC) {
627 			srq->ext.xrc.xrcd = srq_init_attr->ext.xrc.xrcd;
628 			srq->ext.xrc.cq   = srq_init_attr->ext.xrc.cq;
629 			atomic_inc(&srq->ext.xrc.xrcd->usecnt);
630 			atomic_inc(&srq->ext.xrc.cq->usecnt);
631 		}
632 		atomic_inc(&pd->usecnt);
633 		atomic_set(&srq->usecnt, 0);
634 	}
635 
636 	return srq;
637 }
638 EXPORT_SYMBOL(ib_create_srq);
639 
640 int ib_modify_srq(struct ib_srq *srq,
641 		  struct ib_srq_attr *srq_attr,
642 		  enum ib_srq_attr_mask srq_attr_mask)
643 {
644 	return srq->device->modify_srq ?
645 		srq->device->modify_srq(srq, srq_attr, srq_attr_mask, NULL) :
646 		-ENOSYS;
647 }
648 EXPORT_SYMBOL(ib_modify_srq);
649 
650 int ib_query_srq(struct ib_srq *srq,
651 		 struct ib_srq_attr *srq_attr)
652 {
653 	return srq->device->query_srq ?
654 		srq->device->query_srq(srq, srq_attr) : -ENOSYS;
655 }
656 EXPORT_SYMBOL(ib_query_srq);
657 
658 int ib_destroy_srq(struct ib_srq *srq)
659 {
660 	struct ib_pd *pd;
661 	enum ib_srq_type srq_type;
662 	struct ib_xrcd *uninitialized_var(xrcd);
663 	struct ib_cq *uninitialized_var(cq);
664 	int ret;
665 
666 	if (atomic_read(&srq->usecnt))
667 		return -EBUSY;
668 
669 	pd = srq->pd;
670 	srq_type = srq->srq_type;
671 	if (srq_type == IB_SRQT_XRC) {
672 		xrcd = srq->ext.xrc.xrcd;
673 		cq = srq->ext.xrc.cq;
674 	}
675 
676 	ret = srq->device->destroy_srq(srq);
677 	if (!ret) {
678 		atomic_dec(&pd->usecnt);
679 		if (srq_type == IB_SRQT_XRC) {
680 			atomic_dec(&xrcd->usecnt);
681 			atomic_dec(&cq->usecnt);
682 		}
683 	}
684 
685 	return ret;
686 }
687 EXPORT_SYMBOL(ib_destroy_srq);
688 
689 /* Queue pairs */
690 
691 static void __ib_shared_qp_event_handler(struct ib_event *event, void *context)
692 {
693 	struct ib_qp *qp = context;
694 	unsigned long flags;
695 
696 	spin_lock_irqsave(&qp->device->event_handler_lock, flags);
697 	list_for_each_entry(event->element.qp, &qp->open_list, open_list)
698 		if (event->element.qp->event_handler)
699 			event->element.qp->event_handler(event, event->element.qp->qp_context);
700 	spin_unlock_irqrestore(&qp->device->event_handler_lock, flags);
701 }
702 
703 static void __ib_insert_xrcd_qp(struct ib_xrcd *xrcd, struct ib_qp *qp)
704 {
705 	mutex_lock(&xrcd->tgt_qp_mutex);
706 	list_add(&qp->xrcd_list, &xrcd->tgt_qp_list);
707 	mutex_unlock(&xrcd->tgt_qp_mutex);
708 }
709 
710 static struct ib_qp *__ib_open_qp(struct ib_qp *real_qp,
711 				  void (*event_handler)(struct ib_event *, void *),
712 				  void *qp_context)
713 {
714 	struct ib_qp *qp;
715 	unsigned long flags;
716 
717 	qp = kzalloc(sizeof *qp, GFP_KERNEL);
718 	if (!qp)
719 		return ERR_PTR(-ENOMEM);
720 
721 	qp->real_qp = real_qp;
722 	atomic_inc(&real_qp->usecnt);
723 	qp->device = real_qp->device;
724 	qp->event_handler = event_handler;
725 	qp->qp_context = qp_context;
726 	qp->qp_num = real_qp->qp_num;
727 	qp->qp_type = real_qp->qp_type;
728 
729 	spin_lock_irqsave(&real_qp->device->event_handler_lock, flags);
730 	list_add(&qp->open_list, &real_qp->open_list);
731 	spin_unlock_irqrestore(&real_qp->device->event_handler_lock, flags);
732 
733 	return qp;
734 }
735 
736 struct ib_qp *ib_open_qp(struct ib_xrcd *xrcd,
737 			 struct ib_qp_open_attr *qp_open_attr)
738 {
739 	struct ib_qp *qp, *real_qp;
740 
741 	if (qp_open_attr->qp_type != IB_QPT_XRC_TGT)
742 		return ERR_PTR(-EINVAL);
743 
744 	qp = ERR_PTR(-EINVAL);
745 	mutex_lock(&xrcd->tgt_qp_mutex);
746 	list_for_each_entry(real_qp, &xrcd->tgt_qp_list, xrcd_list) {
747 		if (real_qp->qp_num == qp_open_attr->qp_num) {
748 			qp = __ib_open_qp(real_qp, qp_open_attr->event_handler,
749 					  qp_open_attr->qp_context);
750 			break;
751 		}
752 	}
753 	mutex_unlock(&xrcd->tgt_qp_mutex);
754 	return qp;
755 }
756 EXPORT_SYMBOL(ib_open_qp);
757 
758 static struct ib_qp *ib_create_xrc_qp(struct ib_qp *qp,
759 		struct ib_qp_init_attr *qp_init_attr)
760 {
761 	struct ib_qp *real_qp = qp;
762 
763 	qp->event_handler = __ib_shared_qp_event_handler;
764 	qp->qp_context = qp;
765 	qp->pd = NULL;
766 	qp->send_cq = qp->recv_cq = NULL;
767 	qp->srq = NULL;
768 	qp->xrcd = qp_init_attr->xrcd;
769 	atomic_inc(&qp_init_attr->xrcd->usecnt);
770 	INIT_LIST_HEAD(&qp->open_list);
771 
772 	qp = __ib_open_qp(real_qp, qp_init_attr->event_handler,
773 			  qp_init_attr->qp_context);
774 	if (!IS_ERR(qp))
775 		__ib_insert_xrcd_qp(qp_init_attr->xrcd, real_qp);
776 	else
777 		real_qp->device->destroy_qp(real_qp);
778 	return qp;
779 }
780 
781 struct ib_qp *ib_create_qp(struct ib_pd *pd,
782 			   struct ib_qp_init_attr *qp_init_attr)
783 {
784 	struct ib_device *device = pd ? pd->device : qp_init_attr->xrcd->device;
785 	struct ib_qp *qp;
786 	int ret;
787 
788 	if (qp_init_attr->rwq_ind_tbl &&
789 	    (qp_init_attr->recv_cq ||
790 	    qp_init_attr->srq || qp_init_attr->cap.max_recv_wr ||
791 	    qp_init_attr->cap.max_recv_sge))
792 		return ERR_PTR(-EINVAL);
793 
794 	/*
795 	 * If the callers is using the RDMA API calculate the resources
796 	 * needed for the RDMA READ/WRITE operations.
797 	 *
798 	 * Note that these callers need to pass in a port number.
799 	 */
800 	if (qp_init_attr->cap.max_rdma_ctxs)
801 		rdma_rw_init_qp(device, qp_init_attr);
802 
803 	qp = device->create_qp(pd, qp_init_attr, NULL);
804 	if (IS_ERR(qp))
805 		return qp;
806 
807 	qp->device     = device;
808 	qp->real_qp    = qp;
809 	qp->uobject    = NULL;
810 	qp->qp_type    = qp_init_attr->qp_type;
811 	qp->rwq_ind_tbl = qp_init_attr->rwq_ind_tbl;
812 
813 	atomic_set(&qp->usecnt, 0);
814 	qp->mrs_used = 0;
815 	spin_lock_init(&qp->mr_lock);
816 	INIT_LIST_HEAD(&qp->rdma_mrs);
817 	INIT_LIST_HEAD(&qp->sig_mrs);
818 
819 	if (qp_init_attr->qp_type == IB_QPT_XRC_TGT)
820 		return ib_create_xrc_qp(qp, qp_init_attr);
821 
822 	qp->event_handler = qp_init_attr->event_handler;
823 	qp->qp_context = qp_init_attr->qp_context;
824 	if (qp_init_attr->qp_type == IB_QPT_XRC_INI) {
825 		qp->recv_cq = NULL;
826 		qp->srq = NULL;
827 	} else {
828 		qp->recv_cq = qp_init_attr->recv_cq;
829 		if (qp_init_attr->recv_cq)
830 			atomic_inc(&qp_init_attr->recv_cq->usecnt);
831 		qp->srq = qp_init_attr->srq;
832 		if (qp->srq)
833 			atomic_inc(&qp_init_attr->srq->usecnt);
834 	}
835 
836 	qp->pd	    = pd;
837 	qp->send_cq = qp_init_attr->send_cq;
838 	qp->xrcd    = NULL;
839 
840 	atomic_inc(&pd->usecnt);
841 	if (qp_init_attr->send_cq)
842 		atomic_inc(&qp_init_attr->send_cq->usecnt);
843 	if (qp_init_attr->rwq_ind_tbl)
844 		atomic_inc(&qp->rwq_ind_tbl->usecnt);
845 
846 	if (qp_init_attr->cap.max_rdma_ctxs) {
847 		ret = rdma_rw_init_mrs(qp, qp_init_attr);
848 		if (ret) {
849 			pr_err("failed to init MR pool ret= %d\n", ret);
850 			ib_destroy_qp(qp);
851 			return ERR_PTR(ret);
852 		}
853 	}
854 
855 	/*
856 	 * Note: all hw drivers guarantee that max_send_sge is lower than
857 	 * the device RDMA WRITE SGE limit but not all hw drivers ensure that
858 	 * max_send_sge <= max_sge_rd.
859 	 */
860 	qp->max_write_sge = qp_init_attr->cap.max_send_sge;
861 	qp->max_read_sge = min_t(u32, qp_init_attr->cap.max_send_sge,
862 				 device->attrs.max_sge_rd);
863 
864 	return qp;
865 }
866 EXPORT_SYMBOL(ib_create_qp);
867 
868 static const struct {
869 	int			valid;
870 	enum ib_qp_attr_mask	req_param[IB_QPT_MAX];
871 	enum ib_qp_attr_mask	opt_param[IB_QPT_MAX];
872 } qp_state_table[IB_QPS_ERR + 1][IB_QPS_ERR + 1] = {
873 	[IB_QPS_RESET] = {
874 		[IB_QPS_RESET] = { .valid = 1 },
875 		[IB_QPS_INIT]  = {
876 			.valid = 1,
877 			.req_param = {
878 				[IB_QPT_UD]  = (IB_QP_PKEY_INDEX		|
879 						IB_QP_PORT			|
880 						IB_QP_QKEY),
881 				[IB_QPT_RAW_PACKET] = IB_QP_PORT,
882 				[IB_QPT_UC]  = (IB_QP_PKEY_INDEX		|
883 						IB_QP_PORT			|
884 						IB_QP_ACCESS_FLAGS),
885 				[IB_QPT_RC]  = (IB_QP_PKEY_INDEX		|
886 						IB_QP_PORT			|
887 						IB_QP_ACCESS_FLAGS),
888 				[IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX		|
889 						IB_QP_PORT			|
890 						IB_QP_ACCESS_FLAGS),
891 				[IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX		|
892 						IB_QP_PORT			|
893 						IB_QP_ACCESS_FLAGS),
894 				[IB_QPT_SMI] = (IB_QP_PKEY_INDEX		|
895 						IB_QP_QKEY),
896 				[IB_QPT_GSI] = (IB_QP_PKEY_INDEX		|
897 						IB_QP_QKEY),
898 			}
899 		},
900 	},
901 	[IB_QPS_INIT]  = {
902 		[IB_QPS_RESET] = { .valid = 1 },
903 		[IB_QPS_ERR] =   { .valid = 1 },
904 		[IB_QPS_INIT]  = {
905 			.valid = 1,
906 			.opt_param = {
907 				[IB_QPT_UD]  = (IB_QP_PKEY_INDEX		|
908 						IB_QP_PORT			|
909 						IB_QP_QKEY),
910 				[IB_QPT_UC]  = (IB_QP_PKEY_INDEX		|
911 						IB_QP_PORT			|
912 						IB_QP_ACCESS_FLAGS),
913 				[IB_QPT_RC]  = (IB_QP_PKEY_INDEX		|
914 						IB_QP_PORT			|
915 						IB_QP_ACCESS_FLAGS),
916 				[IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX		|
917 						IB_QP_PORT			|
918 						IB_QP_ACCESS_FLAGS),
919 				[IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX		|
920 						IB_QP_PORT			|
921 						IB_QP_ACCESS_FLAGS),
922 				[IB_QPT_SMI] = (IB_QP_PKEY_INDEX		|
923 						IB_QP_QKEY),
924 				[IB_QPT_GSI] = (IB_QP_PKEY_INDEX		|
925 						IB_QP_QKEY),
926 			}
927 		},
928 		[IB_QPS_RTR]   = {
929 			.valid = 1,
930 			.req_param = {
931 				[IB_QPT_UC]  = (IB_QP_AV			|
932 						IB_QP_PATH_MTU			|
933 						IB_QP_DEST_QPN			|
934 						IB_QP_RQ_PSN),
935 				[IB_QPT_RC]  = (IB_QP_AV			|
936 						IB_QP_PATH_MTU			|
937 						IB_QP_DEST_QPN			|
938 						IB_QP_RQ_PSN			|
939 						IB_QP_MAX_DEST_RD_ATOMIC	|
940 						IB_QP_MIN_RNR_TIMER),
941 				[IB_QPT_XRC_INI] = (IB_QP_AV			|
942 						IB_QP_PATH_MTU			|
943 						IB_QP_DEST_QPN			|
944 						IB_QP_RQ_PSN),
945 				[IB_QPT_XRC_TGT] = (IB_QP_AV			|
946 						IB_QP_PATH_MTU			|
947 						IB_QP_DEST_QPN			|
948 						IB_QP_RQ_PSN			|
949 						IB_QP_MAX_DEST_RD_ATOMIC	|
950 						IB_QP_MIN_RNR_TIMER),
951 			},
952 			.opt_param = {
953 				 [IB_QPT_UD]  = (IB_QP_PKEY_INDEX		|
954 						 IB_QP_QKEY),
955 				 [IB_QPT_UC]  = (IB_QP_ALT_PATH			|
956 						 IB_QP_ACCESS_FLAGS		|
957 						 IB_QP_PKEY_INDEX),
958 				 [IB_QPT_RC]  = (IB_QP_ALT_PATH			|
959 						 IB_QP_ACCESS_FLAGS		|
960 						 IB_QP_PKEY_INDEX),
961 				 [IB_QPT_XRC_INI] = (IB_QP_ALT_PATH		|
962 						 IB_QP_ACCESS_FLAGS		|
963 						 IB_QP_PKEY_INDEX),
964 				 [IB_QPT_XRC_TGT] = (IB_QP_ALT_PATH		|
965 						 IB_QP_ACCESS_FLAGS		|
966 						 IB_QP_PKEY_INDEX),
967 				 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX		|
968 						 IB_QP_QKEY),
969 				 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX		|
970 						 IB_QP_QKEY),
971 			 },
972 		},
973 	},
974 	[IB_QPS_RTR]   = {
975 		[IB_QPS_RESET] = { .valid = 1 },
976 		[IB_QPS_ERR] =   { .valid = 1 },
977 		[IB_QPS_RTS]   = {
978 			.valid = 1,
979 			.req_param = {
980 				[IB_QPT_UD]  = IB_QP_SQ_PSN,
981 				[IB_QPT_UC]  = IB_QP_SQ_PSN,
982 				[IB_QPT_RC]  = (IB_QP_TIMEOUT			|
983 						IB_QP_RETRY_CNT			|
984 						IB_QP_RNR_RETRY			|
985 						IB_QP_SQ_PSN			|
986 						IB_QP_MAX_QP_RD_ATOMIC),
987 				[IB_QPT_XRC_INI] = (IB_QP_TIMEOUT		|
988 						IB_QP_RETRY_CNT			|
989 						IB_QP_RNR_RETRY			|
990 						IB_QP_SQ_PSN			|
991 						IB_QP_MAX_QP_RD_ATOMIC),
992 				[IB_QPT_XRC_TGT] = (IB_QP_TIMEOUT		|
993 						IB_QP_SQ_PSN),
994 				[IB_QPT_SMI] = IB_QP_SQ_PSN,
995 				[IB_QPT_GSI] = IB_QP_SQ_PSN,
996 			},
997 			.opt_param = {
998 				 [IB_QPT_UD]  = (IB_QP_CUR_STATE		|
999 						 IB_QP_QKEY),
1000 				 [IB_QPT_UC]  = (IB_QP_CUR_STATE		|
1001 						 IB_QP_ALT_PATH			|
1002 						 IB_QP_ACCESS_FLAGS		|
1003 						 IB_QP_PATH_MIG_STATE),
1004 				 [IB_QPT_RC]  = (IB_QP_CUR_STATE		|
1005 						 IB_QP_ALT_PATH			|
1006 						 IB_QP_ACCESS_FLAGS		|
1007 						 IB_QP_MIN_RNR_TIMER		|
1008 						 IB_QP_PATH_MIG_STATE),
1009 				 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE		|
1010 						 IB_QP_ALT_PATH			|
1011 						 IB_QP_ACCESS_FLAGS		|
1012 						 IB_QP_PATH_MIG_STATE),
1013 				 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE		|
1014 						 IB_QP_ALT_PATH			|
1015 						 IB_QP_ACCESS_FLAGS		|
1016 						 IB_QP_MIN_RNR_TIMER		|
1017 						 IB_QP_PATH_MIG_STATE),
1018 				 [IB_QPT_SMI] = (IB_QP_CUR_STATE		|
1019 						 IB_QP_QKEY),
1020 				 [IB_QPT_GSI] = (IB_QP_CUR_STATE		|
1021 						 IB_QP_QKEY),
1022 				 [IB_QPT_RAW_PACKET] = IB_QP_RATE_LIMIT,
1023 			 }
1024 		}
1025 	},
1026 	[IB_QPS_RTS]   = {
1027 		[IB_QPS_RESET] = { .valid = 1 },
1028 		[IB_QPS_ERR] =   { .valid = 1 },
1029 		[IB_QPS_RTS]   = {
1030 			.valid = 1,
1031 			.opt_param = {
1032 				[IB_QPT_UD]  = (IB_QP_CUR_STATE			|
1033 						IB_QP_QKEY),
1034 				[IB_QPT_UC]  = (IB_QP_CUR_STATE			|
1035 						IB_QP_ACCESS_FLAGS		|
1036 						IB_QP_ALT_PATH			|
1037 						IB_QP_PATH_MIG_STATE),
1038 				[IB_QPT_RC]  = (IB_QP_CUR_STATE			|
1039 						IB_QP_ACCESS_FLAGS		|
1040 						IB_QP_ALT_PATH			|
1041 						IB_QP_PATH_MIG_STATE		|
1042 						IB_QP_MIN_RNR_TIMER),
1043 				[IB_QPT_XRC_INI] = (IB_QP_CUR_STATE		|
1044 						IB_QP_ACCESS_FLAGS		|
1045 						IB_QP_ALT_PATH			|
1046 						IB_QP_PATH_MIG_STATE),
1047 				[IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE		|
1048 						IB_QP_ACCESS_FLAGS		|
1049 						IB_QP_ALT_PATH			|
1050 						IB_QP_PATH_MIG_STATE		|
1051 						IB_QP_MIN_RNR_TIMER),
1052 				[IB_QPT_SMI] = (IB_QP_CUR_STATE			|
1053 						IB_QP_QKEY),
1054 				[IB_QPT_GSI] = (IB_QP_CUR_STATE			|
1055 						IB_QP_QKEY),
1056 				[IB_QPT_RAW_PACKET] = IB_QP_RATE_LIMIT,
1057 			}
1058 		},
1059 		[IB_QPS_SQD]   = {
1060 			.valid = 1,
1061 			.opt_param = {
1062 				[IB_QPT_UD]  = IB_QP_EN_SQD_ASYNC_NOTIFY,
1063 				[IB_QPT_UC]  = IB_QP_EN_SQD_ASYNC_NOTIFY,
1064 				[IB_QPT_RC]  = IB_QP_EN_SQD_ASYNC_NOTIFY,
1065 				[IB_QPT_XRC_INI] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1066 				[IB_QPT_XRC_TGT] = IB_QP_EN_SQD_ASYNC_NOTIFY, /* ??? */
1067 				[IB_QPT_SMI] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1068 				[IB_QPT_GSI] = IB_QP_EN_SQD_ASYNC_NOTIFY
1069 			}
1070 		},
1071 	},
1072 	[IB_QPS_SQD]   = {
1073 		[IB_QPS_RESET] = { .valid = 1 },
1074 		[IB_QPS_ERR] =   { .valid = 1 },
1075 		[IB_QPS_RTS]   = {
1076 			.valid = 1,
1077 			.opt_param = {
1078 				[IB_QPT_UD]  = (IB_QP_CUR_STATE			|
1079 						IB_QP_QKEY),
1080 				[IB_QPT_UC]  = (IB_QP_CUR_STATE			|
1081 						IB_QP_ALT_PATH			|
1082 						IB_QP_ACCESS_FLAGS		|
1083 						IB_QP_PATH_MIG_STATE),
1084 				[IB_QPT_RC]  = (IB_QP_CUR_STATE			|
1085 						IB_QP_ALT_PATH			|
1086 						IB_QP_ACCESS_FLAGS		|
1087 						IB_QP_MIN_RNR_TIMER		|
1088 						IB_QP_PATH_MIG_STATE),
1089 				[IB_QPT_XRC_INI] = (IB_QP_CUR_STATE		|
1090 						IB_QP_ALT_PATH			|
1091 						IB_QP_ACCESS_FLAGS		|
1092 						IB_QP_PATH_MIG_STATE),
1093 				[IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE		|
1094 						IB_QP_ALT_PATH			|
1095 						IB_QP_ACCESS_FLAGS		|
1096 						IB_QP_MIN_RNR_TIMER		|
1097 						IB_QP_PATH_MIG_STATE),
1098 				[IB_QPT_SMI] = (IB_QP_CUR_STATE			|
1099 						IB_QP_QKEY),
1100 				[IB_QPT_GSI] = (IB_QP_CUR_STATE			|
1101 						IB_QP_QKEY),
1102 			}
1103 		},
1104 		[IB_QPS_SQD]   = {
1105 			.valid = 1,
1106 			.opt_param = {
1107 				[IB_QPT_UD]  = (IB_QP_PKEY_INDEX		|
1108 						IB_QP_QKEY),
1109 				[IB_QPT_UC]  = (IB_QP_AV			|
1110 						IB_QP_ALT_PATH			|
1111 						IB_QP_ACCESS_FLAGS		|
1112 						IB_QP_PKEY_INDEX		|
1113 						IB_QP_PATH_MIG_STATE),
1114 				[IB_QPT_RC]  = (IB_QP_PORT			|
1115 						IB_QP_AV			|
1116 						IB_QP_TIMEOUT			|
1117 						IB_QP_RETRY_CNT			|
1118 						IB_QP_RNR_RETRY			|
1119 						IB_QP_MAX_QP_RD_ATOMIC		|
1120 						IB_QP_MAX_DEST_RD_ATOMIC	|
1121 						IB_QP_ALT_PATH			|
1122 						IB_QP_ACCESS_FLAGS		|
1123 						IB_QP_PKEY_INDEX		|
1124 						IB_QP_MIN_RNR_TIMER		|
1125 						IB_QP_PATH_MIG_STATE),
1126 				[IB_QPT_XRC_INI] = (IB_QP_PORT			|
1127 						IB_QP_AV			|
1128 						IB_QP_TIMEOUT			|
1129 						IB_QP_RETRY_CNT			|
1130 						IB_QP_RNR_RETRY			|
1131 						IB_QP_MAX_QP_RD_ATOMIC		|
1132 						IB_QP_ALT_PATH			|
1133 						IB_QP_ACCESS_FLAGS		|
1134 						IB_QP_PKEY_INDEX		|
1135 						IB_QP_PATH_MIG_STATE),
1136 				[IB_QPT_XRC_TGT] = (IB_QP_PORT			|
1137 						IB_QP_AV			|
1138 						IB_QP_TIMEOUT			|
1139 						IB_QP_MAX_DEST_RD_ATOMIC	|
1140 						IB_QP_ALT_PATH			|
1141 						IB_QP_ACCESS_FLAGS		|
1142 						IB_QP_PKEY_INDEX		|
1143 						IB_QP_MIN_RNR_TIMER		|
1144 						IB_QP_PATH_MIG_STATE),
1145 				[IB_QPT_SMI] = (IB_QP_PKEY_INDEX		|
1146 						IB_QP_QKEY),
1147 				[IB_QPT_GSI] = (IB_QP_PKEY_INDEX		|
1148 						IB_QP_QKEY),
1149 			}
1150 		}
1151 	},
1152 	[IB_QPS_SQE]   = {
1153 		[IB_QPS_RESET] = { .valid = 1 },
1154 		[IB_QPS_ERR] =   { .valid = 1 },
1155 		[IB_QPS_RTS]   = {
1156 			.valid = 1,
1157 			.opt_param = {
1158 				[IB_QPT_UD]  = (IB_QP_CUR_STATE			|
1159 						IB_QP_QKEY),
1160 				[IB_QPT_UC]  = (IB_QP_CUR_STATE			|
1161 						IB_QP_ACCESS_FLAGS),
1162 				[IB_QPT_SMI] = (IB_QP_CUR_STATE			|
1163 						IB_QP_QKEY),
1164 				[IB_QPT_GSI] = (IB_QP_CUR_STATE			|
1165 						IB_QP_QKEY),
1166 			}
1167 		}
1168 	},
1169 	[IB_QPS_ERR] = {
1170 		[IB_QPS_RESET] = { .valid = 1 },
1171 		[IB_QPS_ERR] =   { .valid = 1 }
1172 	}
1173 };
1174 
1175 int ib_modify_qp_is_ok(enum ib_qp_state cur_state, enum ib_qp_state next_state,
1176 		       enum ib_qp_type type, enum ib_qp_attr_mask mask,
1177 		       enum rdma_link_layer ll)
1178 {
1179 	enum ib_qp_attr_mask req_param, opt_param;
1180 
1181 	if (cur_state  < 0 || cur_state  > IB_QPS_ERR ||
1182 	    next_state < 0 || next_state > IB_QPS_ERR)
1183 		return 0;
1184 
1185 	if (mask & IB_QP_CUR_STATE  &&
1186 	    cur_state != IB_QPS_RTR && cur_state != IB_QPS_RTS &&
1187 	    cur_state != IB_QPS_SQD && cur_state != IB_QPS_SQE)
1188 		return 0;
1189 
1190 	if (!qp_state_table[cur_state][next_state].valid)
1191 		return 0;
1192 
1193 	req_param = qp_state_table[cur_state][next_state].req_param[type];
1194 	opt_param = qp_state_table[cur_state][next_state].opt_param[type];
1195 
1196 	if ((mask & req_param) != req_param)
1197 		return 0;
1198 
1199 	if (mask & ~(req_param | opt_param | IB_QP_STATE))
1200 		return 0;
1201 
1202 	return 1;
1203 }
1204 EXPORT_SYMBOL(ib_modify_qp_is_ok);
1205 
1206 int ib_resolve_eth_dmac(struct ib_device *device,
1207 			struct rdma_ah_attr *ah_attr)
1208 {
1209 	int           ret = 0;
1210 	struct ib_global_route *grh;
1211 
1212 	if (!rdma_is_port_valid(device, rdma_ah_get_port_num(ah_attr)))
1213 		return -EINVAL;
1214 
1215 	if (ah_attr->type != RDMA_AH_ATTR_TYPE_ROCE)
1216 		return 0;
1217 
1218 	grh = rdma_ah_retrieve_grh(ah_attr);
1219 
1220 	if (rdma_link_local_addr((struct in6_addr *)grh->dgid.raw)) {
1221 		rdma_get_ll_mac((struct in6_addr *)grh->dgid.raw,
1222 				ah_attr->roce.dmac);
1223 	} else {
1224 		union ib_gid		sgid;
1225 		struct ib_gid_attr	sgid_attr;
1226 		int			ifindex;
1227 		int			hop_limit;
1228 
1229 		ret = ib_query_gid(device,
1230 				   rdma_ah_get_port_num(ah_attr),
1231 				   grh->sgid_index,
1232 				   &sgid, &sgid_attr);
1233 
1234 		if (ret || !sgid_attr.ndev) {
1235 			if (!ret)
1236 				ret = -ENXIO;
1237 			goto out;
1238 		}
1239 
1240 		ifindex = sgid_attr.ndev->ifindex;
1241 
1242 		ret =
1243 		rdma_addr_find_l2_eth_by_grh(&sgid, &grh->dgid,
1244 					     ah_attr->roce.dmac,
1245 					     NULL, &ifindex, &hop_limit);
1246 
1247 		dev_put(sgid_attr.ndev);
1248 
1249 		grh->hop_limit = hop_limit;
1250 	}
1251 out:
1252 	return ret;
1253 }
1254 EXPORT_SYMBOL(ib_resolve_eth_dmac);
1255 
1256 int ib_modify_qp(struct ib_qp *qp,
1257 		 struct ib_qp_attr *qp_attr,
1258 		 int qp_attr_mask)
1259 {
1260 
1261 	if (qp_attr_mask & IB_QP_AV) {
1262 		int ret;
1263 
1264 		ret = ib_resolve_eth_dmac(qp->device, &qp_attr->ah_attr);
1265 		if (ret)
1266 			return ret;
1267 	}
1268 
1269 	return qp->device->modify_qp(qp->real_qp, qp_attr, qp_attr_mask, NULL);
1270 }
1271 EXPORT_SYMBOL(ib_modify_qp);
1272 
1273 int ib_query_qp(struct ib_qp *qp,
1274 		struct ib_qp_attr *qp_attr,
1275 		int qp_attr_mask,
1276 		struct ib_qp_init_attr *qp_init_attr)
1277 {
1278 	return qp->device->query_qp ?
1279 		qp->device->query_qp(qp->real_qp, qp_attr, qp_attr_mask, qp_init_attr) :
1280 		-ENOSYS;
1281 }
1282 EXPORT_SYMBOL(ib_query_qp);
1283 
1284 int ib_close_qp(struct ib_qp *qp)
1285 {
1286 	struct ib_qp *real_qp;
1287 	unsigned long flags;
1288 
1289 	real_qp = qp->real_qp;
1290 	if (real_qp == qp)
1291 		return -EINVAL;
1292 
1293 	spin_lock_irqsave(&real_qp->device->event_handler_lock, flags);
1294 	list_del(&qp->open_list);
1295 	spin_unlock_irqrestore(&real_qp->device->event_handler_lock, flags);
1296 
1297 	atomic_dec(&real_qp->usecnt);
1298 	kfree(qp);
1299 
1300 	return 0;
1301 }
1302 EXPORT_SYMBOL(ib_close_qp);
1303 
1304 static int __ib_destroy_shared_qp(struct ib_qp *qp)
1305 {
1306 	struct ib_xrcd *xrcd;
1307 	struct ib_qp *real_qp;
1308 	int ret;
1309 
1310 	real_qp = qp->real_qp;
1311 	xrcd = real_qp->xrcd;
1312 
1313 	mutex_lock(&xrcd->tgt_qp_mutex);
1314 	ib_close_qp(qp);
1315 	if (atomic_read(&real_qp->usecnt) == 0)
1316 		list_del(&real_qp->xrcd_list);
1317 	else
1318 		real_qp = NULL;
1319 	mutex_unlock(&xrcd->tgt_qp_mutex);
1320 
1321 	if (real_qp) {
1322 		ret = ib_destroy_qp(real_qp);
1323 		if (!ret)
1324 			atomic_dec(&xrcd->usecnt);
1325 		else
1326 			__ib_insert_xrcd_qp(xrcd, real_qp);
1327 	}
1328 
1329 	return 0;
1330 }
1331 
1332 int ib_destroy_qp(struct ib_qp *qp)
1333 {
1334 	struct ib_pd *pd;
1335 	struct ib_cq *scq, *rcq;
1336 	struct ib_srq *srq;
1337 	struct ib_rwq_ind_table *ind_tbl;
1338 	int ret;
1339 
1340 	WARN_ON_ONCE(qp->mrs_used > 0);
1341 
1342 	if (atomic_read(&qp->usecnt))
1343 		return -EBUSY;
1344 
1345 	if (qp->real_qp != qp)
1346 		return __ib_destroy_shared_qp(qp);
1347 
1348 	pd   = qp->pd;
1349 	scq  = qp->send_cq;
1350 	rcq  = qp->recv_cq;
1351 	srq  = qp->srq;
1352 	ind_tbl = qp->rwq_ind_tbl;
1353 
1354 	if (!qp->uobject)
1355 		rdma_rw_cleanup_mrs(qp);
1356 
1357 	ret = qp->device->destroy_qp(qp);
1358 	if (!ret) {
1359 		if (pd)
1360 			atomic_dec(&pd->usecnt);
1361 		if (scq)
1362 			atomic_dec(&scq->usecnt);
1363 		if (rcq)
1364 			atomic_dec(&rcq->usecnt);
1365 		if (srq)
1366 			atomic_dec(&srq->usecnt);
1367 		if (ind_tbl)
1368 			atomic_dec(&ind_tbl->usecnt);
1369 	}
1370 
1371 	return ret;
1372 }
1373 EXPORT_SYMBOL(ib_destroy_qp);
1374 
1375 /* Completion queues */
1376 
1377 struct ib_cq *ib_create_cq(struct ib_device *device,
1378 			   ib_comp_handler comp_handler,
1379 			   void (*event_handler)(struct ib_event *, void *),
1380 			   void *cq_context,
1381 			   const struct ib_cq_init_attr *cq_attr)
1382 {
1383 	struct ib_cq *cq;
1384 
1385 	cq = device->create_cq(device, cq_attr, NULL, NULL);
1386 
1387 	if (!IS_ERR(cq)) {
1388 		cq->device        = device;
1389 		cq->uobject       = NULL;
1390 		cq->comp_handler  = comp_handler;
1391 		cq->event_handler = event_handler;
1392 		cq->cq_context    = cq_context;
1393 		atomic_set(&cq->usecnt, 0);
1394 	}
1395 
1396 	return cq;
1397 }
1398 EXPORT_SYMBOL(ib_create_cq);
1399 
1400 int ib_modify_cq(struct ib_cq *cq, u16 cq_count, u16 cq_period)
1401 {
1402 	return cq->device->modify_cq ?
1403 		cq->device->modify_cq(cq, cq_count, cq_period) : -ENOSYS;
1404 }
1405 EXPORT_SYMBOL(ib_modify_cq);
1406 
1407 int ib_destroy_cq(struct ib_cq *cq)
1408 {
1409 	if (atomic_read(&cq->usecnt))
1410 		return -EBUSY;
1411 
1412 	return cq->device->destroy_cq(cq);
1413 }
1414 EXPORT_SYMBOL(ib_destroy_cq);
1415 
1416 int ib_resize_cq(struct ib_cq *cq, int cqe)
1417 {
1418 	return cq->device->resize_cq ?
1419 		cq->device->resize_cq(cq, cqe, NULL) : -ENOSYS;
1420 }
1421 EXPORT_SYMBOL(ib_resize_cq);
1422 
1423 /* Memory regions */
1424 
1425 int ib_dereg_mr(struct ib_mr *mr)
1426 {
1427 	struct ib_pd *pd = mr->pd;
1428 	int ret;
1429 
1430 	ret = mr->device->dereg_mr(mr);
1431 	if (!ret)
1432 		atomic_dec(&pd->usecnt);
1433 
1434 	return ret;
1435 }
1436 EXPORT_SYMBOL(ib_dereg_mr);
1437 
1438 /**
1439  * ib_alloc_mr() - Allocates a memory region
1440  * @pd:            protection domain associated with the region
1441  * @mr_type:       memory region type
1442  * @max_num_sg:    maximum sg entries available for registration.
1443  *
1444  * Notes:
1445  * Memory registeration page/sg lists must not exceed max_num_sg.
1446  * For mr_type IB_MR_TYPE_MEM_REG, the total length cannot exceed
1447  * max_num_sg * used_page_size.
1448  *
1449  */
1450 struct ib_mr *ib_alloc_mr(struct ib_pd *pd,
1451 			  enum ib_mr_type mr_type,
1452 			  u32 max_num_sg)
1453 {
1454 	struct ib_mr *mr;
1455 
1456 	if (!pd->device->alloc_mr)
1457 		return ERR_PTR(-ENOSYS);
1458 
1459 	mr = pd->device->alloc_mr(pd, mr_type, max_num_sg);
1460 	if (!IS_ERR(mr)) {
1461 		mr->device  = pd->device;
1462 		mr->pd      = pd;
1463 		mr->uobject = NULL;
1464 		atomic_inc(&pd->usecnt);
1465 		mr->need_inval = false;
1466 	}
1467 
1468 	return mr;
1469 }
1470 EXPORT_SYMBOL(ib_alloc_mr);
1471 
1472 /* "Fast" memory regions */
1473 
1474 struct ib_fmr *ib_alloc_fmr(struct ib_pd *pd,
1475 			    int mr_access_flags,
1476 			    struct ib_fmr_attr *fmr_attr)
1477 {
1478 	struct ib_fmr *fmr;
1479 
1480 	if (!pd->device->alloc_fmr)
1481 		return ERR_PTR(-ENOSYS);
1482 
1483 	fmr = pd->device->alloc_fmr(pd, mr_access_flags, fmr_attr);
1484 	if (!IS_ERR(fmr)) {
1485 		fmr->device = pd->device;
1486 		fmr->pd     = pd;
1487 		atomic_inc(&pd->usecnt);
1488 	}
1489 
1490 	return fmr;
1491 }
1492 EXPORT_SYMBOL(ib_alloc_fmr);
1493 
1494 int ib_unmap_fmr(struct list_head *fmr_list)
1495 {
1496 	struct ib_fmr *fmr;
1497 
1498 	if (list_empty(fmr_list))
1499 		return 0;
1500 
1501 	fmr = list_entry(fmr_list->next, struct ib_fmr, list);
1502 	return fmr->device->unmap_fmr(fmr_list);
1503 }
1504 EXPORT_SYMBOL(ib_unmap_fmr);
1505 
1506 int ib_dealloc_fmr(struct ib_fmr *fmr)
1507 {
1508 	struct ib_pd *pd;
1509 	int ret;
1510 
1511 	pd = fmr->pd;
1512 	ret = fmr->device->dealloc_fmr(fmr);
1513 	if (!ret)
1514 		atomic_dec(&pd->usecnt);
1515 
1516 	return ret;
1517 }
1518 EXPORT_SYMBOL(ib_dealloc_fmr);
1519 
1520 /* Multicast groups */
1521 
1522 int ib_attach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid)
1523 {
1524 	int ret;
1525 
1526 	if (!qp->device->attach_mcast)
1527 		return -ENOSYS;
1528 	if (gid->raw[0] != 0xff || qp->qp_type != IB_QPT_UD ||
1529 	    lid < be16_to_cpu(IB_MULTICAST_LID_BASE) ||
1530 	    lid == be16_to_cpu(IB_LID_PERMISSIVE))
1531 		return -EINVAL;
1532 
1533 	ret = qp->device->attach_mcast(qp, gid, lid);
1534 	if (!ret)
1535 		atomic_inc(&qp->usecnt);
1536 	return ret;
1537 }
1538 EXPORT_SYMBOL(ib_attach_mcast);
1539 
1540 int ib_detach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid)
1541 {
1542 	int ret;
1543 
1544 	if (!qp->device->detach_mcast)
1545 		return -ENOSYS;
1546 	if (gid->raw[0] != 0xff || qp->qp_type != IB_QPT_UD ||
1547 	    lid < be16_to_cpu(IB_MULTICAST_LID_BASE) ||
1548 	    lid == be16_to_cpu(IB_LID_PERMISSIVE))
1549 		return -EINVAL;
1550 
1551 	ret = qp->device->detach_mcast(qp, gid, lid);
1552 	if (!ret)
1553 		atomic_dec(&qp->usecnt);
1554 	return ret;
1555 }
1556 EXPORT_SYMBOL(ib_detach_mcast);
1557 
1558 struct ib_xrcd *ib_alloc_xrcd(struct ib_device *device)
1559 {
1560 	struct ib_xrcd *xrcd;
1561 
1562 	if (!device->alloc_xrcd)
1563 		return ERR_PTR(-ENOSYS);
1564 
1565 	xrcd = device->alloc_xrcd(device, NULL, NULL);
1566 	if (!IS_ERR(xrcd)) {
1567 		xrcd->device = device;
1568 		xrcd->inode = NULL;
1569 		atomic_set(&xrcd->usecnt, 0);
1570 		mutex_init(&xrcd->tgt_qp_mutex);
1571 		INIT_LIST_HEAD(&xrcd->tgt_qp_list);
1572 	}
1573 
1574 	return xrcd;
1575 }
1576 EXPORT_SYMBOL(ib_alloc_xrcd);
1577 
1578 int ib_dealloc_xrcd(struct ib_xrcd *xrcd)
1579 {
1580 	struct ib_qp *qp;
1581 	int ret;
1582 
1583 	if (atomic_read(&xrcd->usecnt))
1584 		return -EBUSY;
1585 
1586 	while (!list_empty(&xrcd->tgt_qp_list)) {
1587 		qp = list_entry(xrcd->tgt_qp_list.next, struct ib_qp, xrcd_list);
1588 		ret = ib_destroy_qp(qp);
1589 		if (ret)
1590 			return ret;
1591 	}
1592 
1593 	return xrcd->device->dealloc_xrcd(xrcd);
1594 }
1595 EXPORT_SYMBOL(ib_dealloc_xrcd);
1596 
1597 /**
1598  * ib_create_wq - Creates a WQ associated with the specified protection
1599  * domain.
1600  * @pd: The protection domain associated with the WQ.
1601  * @wq_init_attr: A list of initial attributes required to create the
1602  * WQ. If WQ creation succeeds, then the attributes are updated to
1603  * the actual capabilities of the created WQ.
1604  *
1605  * wq_init_attr->max_wr and wq_init_attr->max_sge determine
1606  * the requested size of the WQ, and set to the actual values allocated
1607  * on return.
1608  * If ib_create_wq() succeeds, then max_wr and max_sge will always be
1609  * at least as large as the requested values.
1610  */
1611 struct ib_wq *ib_create_wq(struct ib_pd *pd,
1612 			   struct ib_wq_init_attr *wq_attr)
1613 {
1614 	struct ib_wq *wq;
1615 
1616 	if (!pd->device->create_wq)
1617 		return ERR_PTR(-ENOSYS);
1618 
1619 	wq = pd->device->create_wq(pd, wq_attr, NULL);
1620 	if (!IS_ERR(wq)) {
1621 		wq->event_handler = wq_attr->event_handler;
1622 		wq->wq_context = wq_attr->wq_context;
1623 		wq->wq_type = wq_attr->wq_type;
1624 		wq->cq = wq_attr->cq;
1625 		wq->device = pd->device;
1626 		wq->pd = pd;
1627 		wq->uobject = NULL;
1628 		atomic_inc(&pd->usecnt);
1629 		atomic_inc(&wq_attr->cq->usecnt);
1630 		atomic_set(&wq->usecnt, 0);
1631 	}
1632 	return wq;
1633 }
1634 EXPORT_SYMBOL(ib_create_wq);
1635 
1636 /**
1637  * ib_destroy_wq - Destroys the specified WQ.
1638  * @wq: The WQ to destroy.
1639  */
1640 int ib_destroy_wq(struct ib_wq *wq)
1641 {
1642 	int err;
1643 	struct ib_cq *cq = wq->cq;
1644 	struct ib_pd *pd = wq->pd;
1645 
1646 	if (atomic_read(&wq->usecnt))
1647 		return -EBUSY;
1648 
1649 	err = wq->device->destroy_wq(wq);
1650 	if (!err) {
1651 		atomic_dec(&pd->usecnt);
1652 		atomic_dec(&cq->usecnt);
1653 	}
1654 	return err;
1655 }
1656 EXPORT_SYMBOL(ib_destroy_wq);
1657 
1658 /**
1659  * ib_modify_wq - Modifies the specified WQ.
1660  * @wq: The WQ to modify.
1661  * @wq_attr: On input, specifies the WQ attributes to modify.
1662  * @wq_attr_mask: A bit-mask used to specify which attributes of the WQ
1663  *   are being modified.
1664  * On output, the current values of selected WQ attributes are returned.
1665  */
1666 int ib_modify_wq(struct ib_wq *wq, struct ib_wq_attr *wq_attr,
1667 		 u32 wq_attr_mask)
1668 {
1669 	int err;
1670 
1671 	if (!wq->device->modify_wq)
1672 		return -ENOSYS;
1673 
1674 	err = wq->device->modify_wq(wq, wq_attr, wq_attr_mask, NULL);
1675 	return err;
1676 }
1677 EXPORT_SYMBOL(ib_modify_wq);
1678 
1679 /*
1680  * ib_create_rwq_ind_table - Creates a RQ Indirection Table.
1681  * @device: The device on which to create the rwq indirection table.
1682  * @ib_rwq_ind_table_init_attr: A list of initial attributes required to
1683  * create the Indirection Table.
1684  *
1685  * Note: The life time of ib_rwq_ind_table_init_attr->ind_tbl is not less
1686  *	than the created ib_rwq_ind_table object and the caller is responsible
1687  *	for its memory allocation/free.
1688  */
1689 struct ib_rwq_ind_table *ib_create_rwq_ind_table(struct ib_device *device,
1690 						 struct ib_rwq_ind_table_init_attr *init_attr)
1691 {
1692 	struct ib_rwq_ind_table *rwq_ind_table;
1693 	int i;
1694 	u32 table_size;
1695 
1696 	if (!device->create_rwq_ind_table)
1697 		return ERR_PTR(-ENOSYS);
1698 
1699 	table_size = (1 << init_attr->log_ind_tbl_size);
1700 	rwq_ind_table = device->create_rwq_ind_table(device,
1701 				init_attr, NULL);
1702 	if (IS_ERR(rwq_ind_table))
1703 		return rwq_ind_table;
1704 
1705 	rwq_ind_table->ind_tbl = init_attr->ind_tbl;
1706 	rwq_ind_table->log_ind_tbl_size = init_attr->log_ind_tbl_size;
1707 	rwq_ind_table->device = device;
1708 	rwq_ind_table->uobject = NULL;
1709 	atomic_set(&rwq_ind_table->usecnt, 0);
1710 
1711 	for (i = 0; i < table_size; i++)
1712 		atomic_inc(&rwq_ind_table->ind_tbl[i]->usecnt);
1713 
1714 	return rwq_ind_table;
1715 }
1716 EXPORT_SYMBOL(ib_create_rwq_ind_table);
1717 
1718 /*
1719  * ib_destroy_rwq_ind_table - Destroys the specified Indirection Table.
1720  * @wq_ind_table: The Indirection Table to destroy.
1721 */
1722 int ib_destroy_rwq_ind_table(struct ib_rwq_ind_table *rwq_ind_table)
1723 {
1724 	int err, i;
1725 	u32 table_size = (1 << rwq_ind_table->log_ind_tbl_size);
1726 	struct ib_wq **ind_tbl = rwq_ind_table->ind_tbl;
1727 
1728 	if (atomic_read(&rwq_ind_table->usecnt))
1729 		return -EBUSY;
1730 
1731 	err = rwq_ind_table->device->destroy_rwq_ind_table(rwq_ind_table);
1732 	if (!err) {
1733 		for (i = 0; i < table_size; i++)
1734 			atomic_dec(&ind_tbl[i]->usecnt);
1735 	}
1736 
1737 	return err;
1738 }
1739 EXPORT_SYMBOL(ib_destroy_rwq_ind_table);
1740 
1741 struct ib_flow *ib_create_flow(struct ib_qp *qp,
1742 			       struct ib_flow_attr *flow_attr,
1743 			       int domain)
1744 {
1745 	struct ib_flow *flow_id;
1746 	if (!qp->device->create_flow)
1747 		return ERR_PTR(-ENOSYS);
1748 
1749 	flow_id = qp->device->create_flow(qp, flow_attr, domain);
1750 	if (!IS_ERR(flow_id)) {
1751 		atomic_inc(&qp->usecnt);
1752 		flow_id->qp = qp;
1753 	}
1754 	return flow_id;
1755 }
1756 EXPORT_SYMBOL(ib_create_flow);
1757 
1758 int ib_destroy_flow(struct ib_flow *flow_id)
1759 {
1760 	int err;
1761 	struct ib_qp *qp = flow_id->qp;
1762 
1763 	err = qp->device->destroy_flow(flow_id);
1764 	if (!err)
1765 		atomic_dec(&qp->usecnt);
1766 	return err;
1767 }
1768 EXPORT_SYMBOL(ib_destroy_flow);
1769 
1770 int ib_check_mr_status(struct ib_mr *mr, u32 check_mask,
1771 		       struct ib_mr_status *mr_status)
1772 {
1773 	return mr->device->check_mr_status ?
1774 		mr->device->check_mr_status(mr, check_mask, mr_status) : -ENOSYS;
1775 }
1776 EXPORT_SYMBOL(ib_check_mr_status);
1777 
1778 int ib_set_vf_link_state(struct ib_device *device, int vf, u8 port,
1779 			 int state)
1780 {
1781 	if (!device->set_vf_link_state)
1782 		return -ENOSYS;
1783 
1784 	return device->set_vf_link_state(device, vf, port, state);
1785 }
1786 EXPORT_SYMBOL(ib_set_vf_link_state);
1787 
1788 int ib_get_vf_config(struct ib_device *device, int vf, u8 port,
1789 		     struct ifla_vf_info *info)
1790 {
1791 	if (!device->get_vf_config)
1792 		return -ENOSYS;
1793 
1794 	return device->get_vf_config(device, vf, port, info);
1795 }
1796 EXPORT_SYMBOL(ib_get_vf_config);
1797 
1798 int ib_get_vf_stats(struct ib_device *device, int vf, u8 port,
1799 		    struct ifla_vf_stats *stats)
1800 {
1801 	if (!device->get_vf_stats)
1802 		return -ENOSYS;
1803 
1804 	return device->get_vf_stats(device, vf, port, stats);
1805 }
1806 EXPORT_SYMBOL(ib_get_vf_stats);
1807 
1808 int ib_set_vf_guid(struct ib_device *device, int vf, u8 port, u64 guid,
1809 		   int type)
1810 {
1811 	if (!device->set_vf_guid)
1812 		return -ENOSYS;
1813 
1814 	return device->set_vf_guid(device, vf, port, guid, type);
1815 }
1816 EXPORT_SYMBOL(ib_set_vf_guid);
1817 
1818 /**
1819  * ib_map_mr_sg() - Map the largest prefix of a dma mapped SG list
1820  *     and set it the memory region.
1821  * @mr:            memory region
1822  * @sg:            dma mapped scatterlist
1823  * @sg_nents:      number of entries in sg
1824  * @sg_offset:     offset in bytes into sg
1825  * @page_size:     page vector desired page size
1826  *
1827  * Constraints:
1828  * - The first sg element is allowed to have an offset.
1829  * - Each sg element must either be aligned to page_size or virtually
1830  *   contiguous to the previous element. In case an sg element has a
1831  *   non-contiguous offset, the mapping prefix will not include it.
1832  * - The last sg element is allowed to have length less than page_size.
1833  * - If sg_nents total byte length exceeds the mr max_num_sge * page_size
1834  *   then only max_num_sg entries will be mapped.
1835  * - If the MR was allocated with type IB_MR_TYPE_SG_GAPS, none of these
1836  *   constraints holds and the page_size argument is ignored.
1837  *
1838  * Returns the number of sg elements that were mapped to the memory region.
1839  *
1840  * After this completes successfully, the  memory region
1841  * is ready for registration.
1842  */
1843 int ib_map_mr_sg(struct ib_mr *mr, struct scatterlist *sg, int sg_nents,
1844 		 unsigned int *sg_offset, unsigned int page_size)
1845 {
1846 	if (unlikely(!mr->device->map_mr_sg))
1847 		return -ENOSYS;
1848 
1849 	mr->page_size = page_size;
1850 
1851 	return mr->device->map_mr_sg(mr, sg, sg_nents, sg_offset);
1852 }
1853 EXPORT_SYMBOL(ib_map_mr_sg);
1854 
1855 /**
1856  * ib_sg_to_pages() - Convert the largest prefix of a sg list
1857  *     to a page vector
1858  * @mr:            memory region
1859  * @sgl:           dma mapped scatterlist
1860  * @sg_nents:      number of entries in sg
1861  * @sg_offset_p:   IN:  start offset in bytes into sg
1862  *                 OUT: offset in bytes for element n of the sg of the first
1863  *                      byte that has not been processed where n is the return
1864  *                      value of this function.
1865  * @set_page:      driver page assignment function pointer
1866  *
1867  * Core service helper for drivers to convert the largest
1868  * prefix of given sg list to a page vector. The sg list
1869  * prefix converted is the prefix that meet the requirements
1870  * of ib_map_mr_sg.
1871  *
1872  * Returns the number of sg elements that were assigned to
1873  * a page vector.
1874  */
1875 int ib_sg_to_pages(struct ib_mr *mr, struct scatterlist *sgl, int sg_nents,
1876 		unsigned int *sg_offset_p, int (*set_page)(struct ib_mr *, u64))
1877 {
1878 	struct scatterlist *sg;
1879 	u64 last_end_dma_addr = 0;
1880 	unsigned int sg_offset = sg_offset_p ? *sg_offset_p : 0;
1881 	unsigned int last_page_off = 0;
1882 	u64 page_mask = ~((u64)mr->page_size - 1);
1883 	int i, ret;
1884 
1885 	if (unlikely(sg_nents <= 0 || sg_offset > sg_dma_len(&sgl[0])))
1886 		return -EINVAL;
1887 
1888 	mr->iova = sg_dma_address(&sgl[0]) + sg_offset;
1889 	mr->length = 0;
1890 
1891 	for_each_sg(sgl, sg, sg_nents, i) {
1892 		u64 dma_addr = sg_dma_address(sg) + sg_offset;
1893 		u64 prev_addr = dma_addr;
1894 		unsigned int dma_len = sg_dma_len(sg) - sg_offset;
1895 		u64 end_dma_addr = dma_addr + dma_len;
1896 		u64 page_addr = dma_addr & page_mask;
1897 
1898 		/*
1899 		 * For the second and later elements, check whether either the
1900 		 * end of element i-1 or the start of element i is not aligned
1901 		 * on a page boundary.
1902 		 */
1903 		if (i && (last_page_off != 0 || page_addr != dma_addr)) {
1904 			/* Stop mapping if there is a gap. */
1905 			if (last_end_dma_addr != dma_addr)
1906 				break;
1907 
1908 			/*
1909 			 * Coalesce this element with the last. If it is small
1910 			 * enough just update mr->length. Otherwise start
1911 			 * mapping from the next page.
1912 			 */
1913 			goto next_page;
1914 		}
1915 
1916 		do {
1917 			ret = set_page(mr, page_addr);
1918 			if (unlikely(ret < 0)) {
1919 				sg_offset = prev_addr - sg_dma_address(sg);
1920 				mr->length += prev_addr - dma_addr;
1921 				if (sg_offset_p)
1922 					*sg_offset_p = sg_offset;
1923 				return i || sg_offset ? i : ret;
1924 			}
1925 			prev_addr = page_addr;
1926 next_page:
1927 			page_addr += mr->page_size;
1928 		} while (page_addr < end_dma_addr);
1929 
1930 		mr->length += dma_len;
1931 		last_end_dma_addr = end_dma_addr;
1932 		last_page_off = end_dma_addr & ~page_mask;
1933 
1934 		sg_offset = 0;
1935 	}
1936 
1937 	if (sg_offset_p)
1938 		*sg_offset_p = 0;
1939 	return i;
1940 }
1941 EXPORT_SYMBOL(ib_sg_to_pages);
1942 
1943 struct ib_drain_cqe {
1944 	struct ib_cqe cqe;
1945 	struct completion done;
1946 };
1947 
1948 static void ib_drain_qp_done(struct ib_cq *cq, struct ib_wc *wc)
1949 {
1950 	struct ib_drain_cqe *cqe = container_of(wc->wr_cqe, struct ib_drain_cqe,
1951 						cqe);
1952 
1953 	complete(&cqe->done);
1954 }
1955 
1956 /*
1957  * Post a WR and block until its completion is reaped for the SQ.
1958  */
1959 static void __ib_drain_sq(struct ib_qp *qp)
1960 {
1961 	struct ib_cq *cq = qp->send_cq;
1962 	struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
1963 	struct ib_drain_cqe sdrain;
1964 	struct ib_send_wr swr = {}, *bad_swr;
1965 	int ret;
1966 
1967 	swr.wr_cqe = &sdrain.cqe;
1968 	sdrain.cqe.done = ib_drain_qp_done;
1969 	init_completion(&sdrain.done);
1970 
1971 	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
1972 	if (ret) {
1973 		WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
1974 		return;
1975 	}
1976 
1977 	ret = ib_post_send(qp, &swr, &bad_swr);
1978 	if (ret) {
1979 		WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
1980 		return;
1981 	}
1982 
1983 	if (cq->poll_ctx == IB_POLL_DIRECT)
1984 		while (wait_for_completion_timeout(&sdrain.done, HZ / 10) <= 0)
1985 			ib_process_cq_direct(cq, -1);
1986 	else
1987 		wait_for_completion(&sdrain.done);
1988 }
1989 
1990 /*
1991  * Post a WR and block until its completion is reaped for the RQ.
1992  */
1993 static void __ib_drain_rq(struct ib_qp *qp)
1994 {
1995 	struct ib_cq *cq = qp->recv_cq;
1996 	struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
1997 	struct ib_drain_cqe rdrain;
1998 	struct ib_recv_wr rwr = {}, *bad_rwr;
1999 	int ret;
2000 
2001 	rwr.wr_cqe = &rdrain.cqe;
2002 	rdrain.cqe.done = ib_drain_qp_done;
2003 	init_completion(&rdrain.done);
2004 
2005 	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
2006 	if (ret) {
2007 		WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
2008 		return;
2009 	}
2010 
2011 	ret = ib_post_recv(qp, &rwr, &bad_rwr);
2012 	if (ret) {
2013 		WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
2014 		return;
2015 	}
2016 
2017 	if (cq->poll_ctx == IB_POLL_DIRECT)
2018 		while (wait_for_completion_timeout(&rdrain.done, HZ / 10) <= 0)
2019 			ib_process_cq_direct(cq, -1);
2020 	else
2021 		wait_for_completion(&rdrain.done);
2022 }
2023 
2024 /**
2025  * ib_drain_sq() - Block until all SQ CQEs have been consumed by the
2026  *		   application.
2027  * @qp:            queue pair to drain
2028  *
2029  * If the device has a provider-specific drain function, then
2030  * call that.  Otherwise call the generic drain function
2031  * __ib_drain_sq().
2032  *
2033  * The caller must:
2034  *
2035  * ensure there is room in the CQ and SQ for the drain work request and
2036  * completion.
2037  *
2038  * allocate the CQ using ib_alloc_cq().
2039  *
2040  * ensure that there are no other contexts that are posting WRs concurrently.
2041  * Otherwise the drain is not guaranteed.
2042  */
2043 void ib_drain_sq(struct ib_qp *qp)
2044 {
2045 	if (qp->device->drain_sq)
2046 		qp->device->drain_sq(qp);
2047 	else
2048 		__ib_drain_sq(qp);
2049 }
2050 EXPORT_SYMBOL(ib_drain_sq);
2051 
2052 /**
2053  * ib_drain_rq() - Block until all RQ CQEs have been consumed by the
2054  *		   application.
2055  * @qp:            queue pair to drain
2056  *
2057  * If the device has a provider-specific drain function, then
2058  * call that.  Otherwise call the generic drain function
2059  * __ib_drain_rq().
2060  *
2061  * The caller must:
2062  *
2063  * ensure there is room in the CQ and RQ for the drain work request and
2064  * completion.
2065  *
2066  * allocate the CQ using ib_alloc_cq().
2067  *
2068  * ensure that there are no other contexts that are posting WRs concurrently.
2069  * Otherwise the drain is not guaranteed.
2070  */
2071 void ib_drain_rq(struct ib_qp *qp)
2072 {
2073 	if (qp->device->drain_rq)
2074 		qp->device->drain_rq(qp);
2075 	else
2076 		__ib_drain_rq(qp);
2077 }
2078 EXPORT_SYMBOL(ib_drain_rq);
2079 
2080 /**
2081  * ib_drain_qp() - Block until all CQEs have been consumed by the
2082  *		   application on both the RQ and SQ.
2083  * @qp:            queue pair to drain
2084  *
2085  * The caller must:
2086  *
2087  * ensure there is room in the CQ(s), SQ, and RQ for drain work requests
2088  * and completions.
2089  *
2090  * allocate the CQs using ib_alloc_cq().
2091  *
2092  * ensure that there are no other contexts that are posting WRs concurrently.
2093  * Otherwise the drain is not guaranteed.
2094  */
2095 void ib_drain_qp(struct ib_qp *qp)
2096 {
2097 	ib_drain_sq(qp);
2098 	if (!qp->srq)
2099 		ib_drain_rq(qp);
2100 }
2101 EXPORT_SYMBOL(ib_drain_qp);
2102