xref: /openbmc/linux/drivers/infiniband/core/verbs.c (revision 7bcae826)
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 *ib_create_ah(struct ib_pd *pd, struct ib_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 		atomic_inc(&pd->usecnt);
325 	}
326 
327 	return ah;
328 }
329 EXPORT_SYMBOL(ib_create_ah);
330 
331 int ib_get_rdma_header_version(const union rdma_network_hdr *hdr)
332 {
333 	const struct iphdr *ip4h = (struct iphdr *)&hdr->roce4grh;
334 	struct iphdr ip4h_checked;
335 	const struct ipv6hdr *ip6h = (struct ipv6hdr *)&hdr->ibgrh;
336 
337 	/* If it's IPv6, the version must be 6, otherwise, the first
338 	 * 20 bytes (before the IPv4 header) are garbled.
339 	 */
340 	if (ip6h->version != 6)
341 		return (ip4h->version == 4) ? 4 : 0;
342 	/* version may be 6 or 4 because the first 20 bytes could be garbled */
343 
344 	/* RoCE v2 requires no options, thus header length
345 	 * must be 5 words
346 	 */
347 	if (ip4h->ihl != 5)
348 		return 6;
349 
350 	/* Verify checksum.
351 	 * We can't write on scattered buffers so we need to copy to
352 	 * temp buffer.
353 	 */
354 	memcpy(&ip4h_checked, ip4h, sizeof(ip4h_checked));
355 	ip4h_checked.check = 0;
356 	ip4h_checked.check = ip_fast_csum((u8 *)&ip4h_checked, 5);
357 	/* if IPv4 header checksum is OK, believe it */
358 	if (ip4h->check == ip4h_checked.check)
359 		return 4;
360 	return 6;
361 }
362 EXPORT_SYMBOL(ib_get_rdma_header_version);
363 
364 static enum rdma_network_type ib_get_net_type_by_grh(struct ib_device *device,
365 						     u8 port_num,
366 						     const struct ib_grh *grh)
367 {
368 	int grh_version;
369 
370 	if (rdma_protocol_ib(device, port_num))
371 		return RDMA_NETWORK_IB;
372 
373 	grh_version = ib_get_rdma_header_version((union rdma_network_hdr *)grh);
374 
375 	if (grh_version == 4)
376 		return RDMA_NETWORK_IPV4;
377 
378 	if (grh->next_hdr == IPPROTO_UDP)
379 		return RDMA_NETWORK_IPV6;
380 
381 	return RDMA_NETWORK_ROCE_V1;
382 }
383 
384 struct find_gid_index_context {
385 	u16 vlan_id;
386 	enum ib_gid_type gid_type;
387 };
388 
389 static bool find_gid_index(const union ib_gid *gid,
390 			   const struct ib_gid_attr *gid_attr,
391 			   void *context)
392 {
393 	struct find_gid_index_context *ctx =
394 		(struct find_gid_index_context *)context;
395 
396 	if (ctx->gid_type != gid_attr->gid_type)
397 		return false;
398 
399 	if ((!!(ctx->vlan_id != 0xffff) == !is_vlan_dev(gid_attr->ndev)) ||
400 	    (is_vlan_dev(gid_attr->ndev) &&
401 	     vlan_dev_vlan_id(gid_attr->ndev) != ctx->vlan_id))
402 		return false;
403 
404 	return true;
405 }
406 
407 static int get_sgid_index_from_eth(struct ib_device *device, u8 port_num,
408 				   u16 vlan_id, const union ib_gid *sgid,
409 				   enum ib_gid_type gid_type,
410 				   u16 *gid_index)
411 {
412 	struct find_gid_index_context context = {.vlan_id = vlan_id,
413 						 .gid_type = gid_type};
414 
415 	return ib_find_gid_by_filter(device, sgid, port_num, find_gid_index,
416 				     &context, gid_index);
417 }
418 
419 int ib_get_gids_from_rdma_hdr(const union rdma_network_hdr *hdr,
420 			      enum rdma_network_type net_type,
421 			      union ib_gid *sgid, union ib_gid *dgid)
422 {
423 	struct sockaddr_in  src_in;
424 	struct sockaddr_in  dst_in;
425 	__be32 src_saddr, dst_saddr;
426 
427 	if (!sgid || !dgid)
428 		return -EINVAL;
429 
430 	if (net_type == RDMA_NETWORK_IPV4) {
431 		memcpy(&src_in.sin_addr.s_addr,
432 		       &hdr->roce4grh.saddr, 4);
433 		memcpy(&dst_in.sin_addr.s_addr,
434 		       &hdr->roce4grh.daddr, 4);
435 		src_saddr = src_in.sin_addr.s_addr;
436 		dst_saddr = dst_in.sin_addr.s_addr;
437 		ipv6_addr_set_v4mapped(src_saddr,
438 				       (struct in6_addr *)sgid);
439 		ipv6_addr_set_v4mapped(dst_saddr,
440 				       (struct in6_addr *)dgid);
441 		return 0;
442 	} else if (net_type == RDMA_NETWORK_IPV6 ||
443 		   net_type == RDMA_NETWORK_IB) {
444 		*dgid = hdr->ibgrh.dgid;
445 		*sgid = hdr->ibgrh.sgid;
446 		return 0;
447 	} else {
448 		return -EINVAL;
449 	}
450 }
451 EXPORT_SYMBOL(ib_get_gids_from_rdma_hdr);
452 
453 int ib_init_ah_from_wc(struct ib_device *device, u8 port_num,
454 		       const struct ib_wc *wc, const struct ib_grh *grh,
455 		       struct ib_ah_attr *ah_attr)
456 {
457 	u32 flow_class;
458 	u16 gid_index;
459 	int ret;
460 	enum rdma_network_type net_type = RDMA_NETWORK_IB;
461 	enum ib_gid_type gid_type = IB_GID_TYPE_IB;
462 	int hoplimit = 0xff;
463 	union ib_gid dgid;
464 	union ib_gid sgid;
465 
466 	memset(ah_attr, 0, sizeof *ah_attr);
467 	if (rdma_cap_eth_ah(device, port_num)) {
468 		if (wc->wc_flags & IB_WC_WITH_NETWORK_HDR_TYPE)
469 			net_type = wc->network_hdr_type;
470 		else
471 			net_type = ib_get_net_type_by_grh(device, port_num, grh);
472 		gid_type = ib_network_to_gid_type(net_type);
473 	}
474 	ret = ib_get_gids_from_rdma_hdr((union rdma_network_hdr *)grh, net_type,
475 					&sgid, &dgid);
476 	if (ret)
477 		return ret;
478 
479 	if (rdma_protocol_roce(device, port_num)) {
480 		int if_index = 0;
481 		u16 vlan_id = wc->wc_flags & IB_WC_WITH_VLAN ?
482 				wc->vlan_id : 0xffff;
483 		struct net_device *idev;
484 		struct net_device *resolved_dev;
485 
486 		if (!(wc->wc_flags & IB_WC_GRH))
487 			return -EPROTOTYPE;
488 
489 		if (!device->get_netdev)
490 			return -EOPNOTSUPP;
491 
492 		idev = device->get_netdev(device, port_num);
493 		if (!idev)
494 			return -ENODEV;
495 
496 		ret = rdma_addr_find_l2_eth_by_grh(&dgid, &sgid,
497 						   ah_attr->dmac,
498 						   wc->wc_flags & IB_WC_WITH_VLAN ?
499 						   NULL : &vlan_id,
500 						   &if_index, &hoplimit);
501 		if (ret) {
502 			dev_put(idev);
503 			return ret;
504 		}
505 
506 		resolved_dev = dev_get_by_index(&init_net, if_index);
507 		if (resolved_dev->flags & IFF_LOOPBACK) {
508 			dev_put(resolved_dev);
509 			resolved_dev = idev;
510 			dev_hold(resolved_dev);
511 		}
512 		rcu_read_lock();
513 		if (resolved_dev != idev && !rdma_is_upper_dev_rcu(idev,
514 								   resolved_dev))
515 			ret = -EHOSTUNREACH;
516 		rcu_read_unlock();
517 		dev_put(idev);
518 		dev_put(resolved_dev);
519 		if (ret)
520 			return ret;
521 
522 		ret = get_sgid_index_from_eth(device, port_num, vlan_id,
523 					      &dgid, gid_type, &gid_index);
524 		if (ret)
525 			return ret;
526 	}
527 
528 	ah_attr->dlid = wc->slid;
529 	ah_attr->sl = wc->sl;
530 	ah_attr->src_path_bits = wc->dlid_path_bits;
531 	ah_attr->port_num = port_num;
532 
533 	if (wc->wc_flags & IB_WC_GRH) {
534 		ah_attr->ah_flags = IB_AH_GRH;
535 		ah_attr->grh.dgid = sgid;
536 
537 		if (!rdma_cap_eth_ah(device, port_num)) {
538 			if (dgid.global.interface_id != cpu_to_be64(IB_SA_WELL_KNOWN_GUID)) {
539 				ret = ib_find_cached_gid_by_port(device, &dgid,
540 								 IB_GID_TYPE_IB,
541 								 port_num, NULL,
542 								 &gid_index);
543 				if (ret)
544 					return ret;
545 			} else {
546 				gid_index = 0;
547 			}
548 		}
549 
550 		ah_attr->grh.sgid_index = (u8) gid_index;
551 		flow_class = be32_to_cpu(grh->version_tclass_flow);
552 		ah_attr->grh.flow_label = flow_class & 0xFFFFF;
553 		ah_attr->grh.hop_limit = hoplimit;
554 		ah_attr->grh.traffic_class = (flow_class >> 20) & 0xFF;
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 ib_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 ib_create_ah(pd, &ah_attr);
571 }
572 EXPORT_SYMBOL(ib_create_ah_from_wc);
573 
574 int ib_modify_ah(struct ib_ah *ah, struct ib_ah_attr *ah_attr)
575 {
576 	return ah->device->modify_ah ?
577 		ah->device->modify_ah(ah, ah_attr) :
578 		-ENOSYS;
579 }
580 EXPORT_SYMBOL(ib_modify_ah);
581 
582 int ib_query_ah(struct ib_ah *ah, struct ib_ah_attr *ah_attr)
583 {
584 	return ah->device->query_ah ?
585 		ah->device->query_ah(ah, ah_attr) :
586 		-ENOSYS;
587 }
588 EXPORT_SYMBOL(ib_query_ah);
589 
590 int ib_destroy_ah(struct ib_ah *ah)
591 {
592 	struct ib_pd *pd;
593 	int ret;
594 
595 	pd = ah->pd;
596 	ret = ah->device->destroy_ah(ah);
597 	if (!ret)
598 		atomic_dec(&pd->usecnt);
599 
600 	return ret;
601 }
602 EXPORT_SYMBOL(ib_destroy_ah);
603 
604 /* Shared receive queues */
605 
606 struct ib_srq *ib_create_srq(struct ib_pd *pd,
607 			     struct ib_srq_init_attr *srq_init_attr)
608 {
609 	struct ib_srq *srq;
610 
611 	if (!pd->device->create_srq)
612 		return ERR_PTR(-ENOSYS);
613 
614 	srq = pd->device->create_srq(pd, srq_init_attr, NULL);
615 
616 	if (!IS_ERR(srq)) {
617 		srq->device    	   = pd->device;
618 		srq->pd        	   = pd;
619 		srq->uobject       = NULL;
620 		srq->event_handler = srq_init_attr->event_handler;
621 		srq->srq_context   = srq_init_attr->srq_context;
622 		srq->srq_type      = srq_init_attr->srq_type;
623 		if (srq->srq_type == IB_SRQT_XRC) {
624 			srq->ext.xrc.xrcd = srq_init_attr->ext.xrc.xrcd;
625 			srq->ext.xrc.cq   = srq_init_attr->ext.xrc.cq;
626 			atomic_inc(&srq->ext.xrc.xrcd->usecnt);
627 			atomic_inc(&srq->ext.xrc.cq->usecnt);
628 		}
629 		atomic_inc(&pd->usecnt);
630 		atomic_set(&srq->usecnt, 0);
631 	}
632 
633 	return srq;
634 }
635 EXPORT_SYMBOL(ib_create_srq);
636 
637 int ib_modify_srq(struct ib_srq *srq,
638 		  struct ib_srq_attr *srq_attr,
639 		  enum ib_srq_attr_mask srq_attr_mask)
640 {
641 	return srq->device->modify_srq ?
642 		srq->device->modify_srq(srq, srq_attr, srq_attr_mask, NULL) :
643 		-ENOSYS;
644 }
645 EXPORT_SYMBOL(ib_modify_srq);
646 
647 int ib_query_srq(struct ib_srq *srq,
648 		 struct ib_srq_attr *srq_attr)
649 {
650 	return srq->device->query_srq ?
651 		srq->device->query_srq(srq, srq_attr) : -ENOSYS;
652 }
653 EXPORT_SYMBOL(ib_query_srq);
654 
655 int ib_destroy_srq(struct ib_srq *srq)
656 {
657 	struct ib_pd *pd;
658 	enum ib_srq_type srq_type;
659 	struct ib_xrcd *uninitialized_var(xrcd);
660 	struct ib_cq *uninitialized_var(cq);
661 	int ret;
662 
663 	if (atomic_read(&srq->usecnt))
664 		return -EBUSY;
665 
666 	pd = srq->pd;
667 	srq_type = srq->srq_type;
668 	if (srq_type == IB_SRQT_XRC) {
669 		xrcd = srq->ext.xrc.xrcd;
670 		cq = srq->ext.xrc.cq;
671 	}
672 
673 	ret = srq->device->destroy_srq(srq);
674 	if (!ret) {
675 		atomic_dec(&pd->usecnt);
676 		if (srq_type == IB_SRQT_XRC) {
677 			atomic_dec(&xrcd->usecnt);
678 			atomic_dec(&cq->usecnt);
679 		}
680 	}
681 
682 	return ret;
683 }
684 EXPORT_SYMBOL(ib_destroy_srq);
685 
686 /* Queue pairs */
687 
688 static void __ib_shared_qp_event_handler(struct ib_event *event, void *context)
689 {
690 	struct ib_qp *qp = context;
691 	unsigned long flags;
692 
693 	spin_lock_irqsave(&qp->device->event_handler_lock, flags);
694 	list_for_each_entry(event->element.qp, &qp->open_list, open_list)
695 		if (event->element.qp->event_handler)
696 			event->element.qp->event_handler(event, event->element.qp->qp_context);
697 	spin_unlock_irqrestore(&qp->device->event_handler_lock, flags);
698 }
699 
700 static void __ib_insert_xrcd_qp(struct ib_xrcd *xrcd, struct ib_qp *qp)
701 {
702 	mutex_lock(&xrcd->tgt_qp_mutex);
703 	list_add(&qp->xrcd_list, &xrcd->tgt_qp_list);
704 	mutex_unlock(&xrcd->tgt_qp_mutex);
705 }
706 
707 static struct ib_qp *__ib_open_qp(struct ib_qp *real_qp,
708 				  void (*event_handler)(struct ib_event *, void *),
709 				  void *qp_context)
710 {
711 	struct ib_qp *qp;
712 	unsigned long flags;
713 
714 	qp = kzalloc(sizeof *qp, GFP_KERNEL);
715 	if (!qp)
716 		return ERR_PTR(-ENOMEM);
717 
718 	qp->real_qp = real_qp;
719 	atomic_inc(&real_qp->usecnt);
720 	qp->device = real_qp->device;
721 	qp->event_handler = event_handler;
722 	qp->qp_context = qp_context;
723 	qp->qp_num = real_qp->qp_num;
724 	qp->qp_type = real_qp->qp_type;
725 
726 	spin_lock_irqsave(&real_qp->device->event_handler_lock, flags);
727 	list_add(&qp->open_list, &real_qp->open_list);
728 	spin_unlock_irqrestore(&real_qp->device->event_handler_lock, flags);
729 
730 	return qp;
731 }
732 
733 struct ib_qp *ib_open_qp(struct ib_xrcd *xrcd,
734 			 struct ib_qp_open_attr *qp_open_attr)
735 {
736 	struct ib_qp *qp, *real_qp;
737 
738 	if (qp_open_attr->qp_type != IB_QPT_XRC_TGT)
739 		return ERR_PTR(-EINVAL);
740 
741 	qp = ERR_PTR(-EINVAL);
742 	mutex_lock(&xrcd->tgt_qp_mutex);
743 	list_for_each_entry(real_qp, &xrcd->tgt_qp_list, xrcd_list) {
744 		if (real_qp->qp_num == qp_open_attr->qp_num) {
745 			qp = __ib_open_qp(real_qp, qp_open_attr->event_handler,
746 					  qp_open_attr->qp_context);
747 			break;
748 		}
749 	}
750 	mutex_unlock(&xrcd->tgt_qp_mutex);
751 	return qp;
752 }
753 EXPORT_SYMBOL(ib_open_qp);
754 
755 static struct ib_qp *ib_create_xrc_qp(struct ib_qp *qp,
756 		struct ib_qp_init_attr *qp_init_attr)
757 {
758 	struct ib_qp *real_qp = qp;
759 
760 	qp->event_handler = __ib_shared_qp_event_handler;
761 	qp->qp_context = qp;
762 	qp->pd = NULL;
763 	qp->send_cq = qp->recv_cq = NULL;
764 	qp->srq = NULL;
765 	qp->xrcd = qp_init_attr->xrcd;
766 	atomic_inc(&qp_init_attr->xrcd->usecnt);
767 	INIT_LIST_HEAD(&qp->open_list);
768 
769 	qp = __ib_open_qp(real_qp, qp_init_attr->event_handler,
770 			  qp_init_attr->qp_context);
771 	if (!IS_ERR(qp))
772 		__ib_insert_xrcd_qp(qp_init_attr->xrcd, real_qp);
773 	else
774 		real_qp->device->destroy_qp(real_qp);
775 	return qp;
776 }
777 
778 struct ib_qp *ib_create_qp(struct ib_pd *pd,
779 			   struct ib_qp_init_attr *qp_init_attr)
780 {
781 	struct ib_device *device = pd ? pd->device : qp_init_attr->xrcd->device;
782 	struct ib_qp *qp;
783 	int ret;
784 
785 	if (qp_init_attr->rwq_ind_tbl &&
786 	    (qp_init_attr->recv_cq ||
787 	    qp_init_attr->srq || qp_init_attr->cap.max_recv_wr ||
788 	    qp_init_attr->cap.max_recv_sge))
789 		return ERR_PTR(-EINVAL);
790 
791 	/*
792 	 * If the callers is using the RDMA API calculate the resources
793 	 * needed for the RDMA READ/WRITE operations.
794 	 *
795 	 * Note that these callers need to pass in a port number.
796 	 */
797 	if (qp_init_attr->cap.max_rdma_ctxs)
798 		rdma_rw_init_qp(device, qp_init_attr);
799 
800 	qp = device->create_qp(pd, qp_init_attr, NULL);
801 	if (IS_ERR(qp))
802 		return qp;
803 
804 	qp->device     = device;
805 	qp->real_qp    = qp;
806 	qp->uobject    = NULL;
807 	qp->qp_type    = qp_init_attr->qp_type;
808 	qp->rwq_ind_tbl = qp_init_attr->rwq_ind_tbl;
809 
810 	atomic_set(&qp->usecnt, 0);
811 	qp->mrs_used = 0;
812 	spin_lock_init(&qp->mr_lock);
813 	INIT_LIST_HEAD(&qp->rdma_mrs);
814 	INIT_LIST_HEAD(&qp->sig_mrs);
815 
816 	if (qp_init_attr->qp_type == IB_QPT_XRC_TGT)
817 		return ib_create_xrc_qp(qp, qp_init_attr);
818 
819 	qp->event_handler = qp_init_attr->event_handler;
820 	qp->qp_context = qp_init_attr->qp_context;
821 	if (qp_init_attr->qp_type == IB_QPT_XRC_INI) {
822 		qp->recv_cq = NULL;
823 		qp->srq = NULL;
824 	} else {
825 		qp->recv_cq = qp_init_attr->recv_cq;
826 		if (qp_init_attr->recv_cq)
827 			atomic_inc(&qp_init_attr->recv_cq->usecnt);
828 		qp->srq = qp_init_attr->srq;
829 		if (qp->srq)
830 			atomic_inc(&qp_init_attr->srq->usecnt);
831 	}
832 
833 	qp->pd	    = pd;
834 	qp->send_cq = qp_init_attr->send_cq;
835 	qp->xrcd    = NULL;
836 
837 	atomic_inc(&pd->usecnt);
838 	if (qp_init_attr->send_cq)
839 		atomic_inc(&qp_init_attr->send_cq->usecnt);
840 	if (qp_init_attr->rwq_ind_tbl)
841 		atomic_inc(&qp->rwq_ind_tbl->usecnt);
842 
843 	if (qp_init_attr->cap.max_rdma_ctxs) {
844 		ret = rdma_rw_init_mrs(qp, qp_init_attr);
845 		if (ret) {
846 			pr_err("failed to init MR pool ret= %d\n", ret);
847 			ib_destroy_qp(qp);
848 			return ERR_PTR(ret);
849 		}
850 	}
851 
852 	/*
853 	 * Note: all hw drivers guarantee that max_send_sge is lower than
854 	 * the device RDMA WRITE SGE limit but not all hw drivers ensure that
855 	 * max_send_sge <= max_sge_rd.
856 	 */
857 	qp->max_write_sge = qp_init_attr->cap.max_send_sge;
858 	qp->max_read_sge = min_t(u32, qp_init_attr->cap.max_send_sge,
859 				 device->attrs.max_sge_rd);
860 
861 	return qp;
862 }
863 EXPORT_SYMBOL(ib_create_qp);
864 
865 static const struct {
866 	int			valid;
867 	enum ib_qp_attr_mask	req_param[IB_QPT_MAX];
868 	enum ib_qp_attr_mask	opt_param[IB_QPT_MAX];
869 } qp_state_table[IB_QPS_ERR + 1][IB_QPS_ERR + 1] = {
870 	[IB_QPS_RESET] = {
871 		[IB_QPS_RESET] = { .valid = 1 },
872 		[IB_QPS_INIT]  = {
873 			.valid = 1,
874 			.req_param = {
875 				[IB_QPT_UD]  = (IB_QP_PKEY_INDEX		|
876 						IB_QP_PORT			|
877 						IB_QP_QKEY),
878 				[IB_QPT_RAW_PACKET] = IB_QP_PORT,
879 				[IB_QPT_UC]  = (IB_QP_PKEY_INDEX		|
880 						IB_QP_PORT			|
881 						IB_QP_ACCESS_FLAGS),
882 				[IB_QPT_RC]  = (IB_QP_PKEY_INDEX		|
883 						IB_QP_PORT			|
884 						IB_QP_ACCESS_FLAGS),
885 				[IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX		|
886 						IB_QP_PORT			|
887 						IB_QP_ACCESS_FLAGS),
888 				[IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX		|
889 						IB_QP_PORT			|
890 						IB_QP_ACCESS_FLAGS),
891 				[IB_QPT_SMI] = (IB_QP_PKEY_INDEX		|
892 						IB_QP_QKEY),
893 				[IB_QPT_GSI] = (IB_QP_PKEY_INDEX		|
894 						IB_QP_QKEY),
895 			}
896 		},
897 	},
898 	[IB_QPS_INIT]  = {
899 		[IB_QPS_RESET] = { .valid = 1 },
900 		[IB_QPS_ERR] =   { .valid = 1 },
901 		[IB_QPS_INIT]  = {
902 			.valid = 1,
903 			.opt_param = {
904 				[IB_QPT_UD]  = (IB_QP_PKEY_INDEX		|
905 						IB_QP_PORT			|
906 						IB_QP_QKEY),
907 				[IB_QPT_UC]  = (IB_QP_PKEY_INDEX		|
908 						IB_QP_PORT			|
909 						IB_QP_ACCESS_FLAGS),
910 				[IB_QPT_RC]  = (IB_QP_PKEY_INDEX		|
911 						IB_QP_PORT			|
912 						IB_QP_ACCESS_FLAGS),
913 				[IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX		|
914 						IB_QP_PORT			|
915 						IB_QP_ACCESS_FLAGS),
916 				[IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX		|
917 						IB_QP_PORT			|
918 						IB_QP_ACCESS_FLAGS),
919 				[IB_QPT_SMI] = (IB_QP_PKEY_INDEX		|
920 						IB_QP_QKEY),
921 				[IB_QPT_GSI] = (IB_QP_PKEY_INDEX		|
922 						IB_QP_QKEY),
923 			}
924 		},
925 		[IB_QPS_RTR]   = {
926 			.valid = 1,
927 			.req_param = {
928 				[IB_QPT_UC]  = (IB_QP_AV			|
929 						IB_QP_PATH_MTU			|
930 						IB_QP_DEST_QPN			|
931 						IB_QP_RQ_PSN),
932 				[IB_QPT_RC]  = (IB_QP_AV			|
933 						IB_QP_PATH_MTU			|
934 						IB_QP_DEST_QPN			|
935 						IB_QP_RQ_PSN			|
936 						IB_QP_MAX_DEST_RD_ATOMIC	|
937 						IB_QP_MIN_RNR_TIMER),
938 				[IB_QPT_XRC_INI] = (IB_QP_AV			|
939 						IB_QP_PATH_MTU			|
940 						IB_QP_DEST_QPN			|
941 						IB_QP_RQ_PSN),
942 				[IB_QPT_XRC_TGT] = (IB_QP_AV			|
943 						IB_QP_PATH_MTU			|
944 						IB_QP_DEST_QPN			|
945 						IB_QP_RQ_PSN			|
946 						IB_QP_MAX_DEST_RD_ATOMIC	|
947 						IB_QP_MIN_RNR_TIMER),
948 			},
949 			.opt_param = {
950 				 [IB_QPT_UD]  = (IB_QP_PKEY_INDEX		|
951 						 IB_QP_QKEY),
952 				 [IB_QPT_UC]  = (IB_QP_ALT_PATH			|
953 						 IB_QP_ACCESS_FLAGS		|
954 						 IB_QP_PKEY_INDEX),
955 				 [IB_QPT_RC]  = (IB_QP_ALT_PATH			|
956 						 IB_QP_ACCESS_FLAGS		|
957 						 IB_QP_PKEY_INDEX),
958 				 [IB_QPT_XRC_INI] = (IB_QP_ALT_PATH		|
959 						 IB_QP_ACCESS_FLAGS		|
960 						 IB_QP_PKEY_INDEX),
961 				 [IB_QPT_XRC_TGT] = (IB_QP_ALT_PATH		|
962 						 IB_QP_ACCESS_FLAGS		|
963 						 IB_QP_PKEY_INDEX),
964 				 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX		|
965 						 IB_QP_QKEY),
966 				 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX		|
967 						 IB_QP_QKEY),
968 			 },
969 		},
970 	},
971 	[IB_QPS_RTR]   = {
972 		[IB_QPS_RESET] = { .valid = 1 },
973 		[IB_QPS_ERR] =   { .valid = 1 },
974 		[IB_QPS_RTS]   = {
975 			.valid = 1,
976 			.req_param = {
977 				[IB_QPT_UD]  = IB_QP_SQ_PSN,
978 				[IB_QPT_UC]  = IB_QP_SQ_PSN,
979 				[IB_QPT_RC]  = (IB_QP_TIMEOUT			|
980 						IB_QP_RETRY_CNT			|
981 						IB_QP_RNR_RETRY			|
982 						IB_QP_SQ_PSN			|
983 						IB_QP_MAX_QP_RD_ATOMIC),
984 				[IB_QPT_XRC_INI] = (IB_QP_TIMEOUT		|
985 						IB_QP_RETRY_CNT			|
986 						IB_QP_RNR_RETRY			|
987 						IB_QP_SQ_PSN			|
988 						IB_QP_MAX_QP_RD_ATOMIC),
989 				[IB_QPT_XRC_TGT] = (IB_QP_TIMEOUT		|
990 						IB_QP_SQ_PSN),
991 				[IB_QPT_SMI] = IB_QP_SQ_PSN,
992 				[IB_QPT_GSI] = IB_QP_SQ_PSN,
993 			},
994 			.opt_param = {
995 				 [IB_QPT_UD]  = (IB_QP_CUR_STATE		|
996 						 IB_QP_QKEY),
997 				 [IB_QPT_UC]  = (IB_QP_CUR_STATE		|
998 						 IB_QP_ALT_PATH			|
999 						 IB_QP_ACCESS_FLAGS		|
1000 						 IB_QP_PATH_MIG_STATE),
1001 				 [IB_QPT_RC]  = (IB_QP_CUR_STATE		|
1002 						 IB_QP_ALT_PATH			|
1003 						 IB_QP_ACCESS_FLAGS		|
1004 						 IB_QP_MIN_RNR_TIMER		|
1005 						 IB_QP_PATH_MIG_STATE),
1006 				 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE		|
1007 						 IB_QP_ALT_PATH			|
1008 						 IB_QP_ACCESS_FLAGS		|
1009 						 IB_QP_PATH_MIG_STATE),
1010 				 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE		|
1011 						 IB_QP_ALT_PATH			|
1012 						 IB_QP_ACCESS_FLAGS		|
1013 						 IB_QP_MIN_RNR_TIMER		|
1014 						 IB_QP_PATH_MIG_STATE),
1015 				 [IB_QPT_SMI] = (IB_QP_CUR_STATE		|
1016 						 IB_QP_QKEY),
1017 				 [IB_QPT_GSI] = (IB_QP_CUR_STATE		|
1018 						 IB_QP_QKEY),
1019 				 [IB_QPT_RAW_PACKET] = IB_QP_RATE_LIMIT,
1020 			 }
1021 		}
1022 	},
1023 	[IB_QPS_RTS]   = {
1024 		[IB_QPS_RESET] = { .valid = 1 },
1025 		[IB_QPS_ERR] =   { .valid = 1 },
1026 		[IB_QPS_RTS]   = {
1027 			.valid = 1,
1028 			.opt_param = {
1029 				[IB_QPT_UD]  = (IB_QP_CUR_STATE			|
1030 						IB_QP_QKEY),
1031 				[IB_QPT_UC]  = (IB_QP_CUR_STATE			|
1032 						IB_QP_ACCESS_FLAGS		|
1033 						IB_QP_ALT_PATH			|
1034 						IB_QP_PATH_MIG_STATE),
1035 				[IB_QPT_RC]  = (IB_QP_CUR_STATE			|
1036 						IB_QP_ACCESS_FLAGS		|
1037 						IB_QP_ALT_PATH			|
1038 						IB_QP_PATH_MIG_STATE		|
1039 						IB_QP_MIN_RNR_TIMER),
1040 				[IB_QPT_XRC_INI] = (IB_QP_CUR_STATE		|
1041 						IB_QP_ACCESS_FLAGS		|
1042 						IB_QP_ALT_PATH			|
1043 						IB_QP_PATH_MIG_STATE),
1044 				[IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE		|
1045 						IB_QP_ACCESS_FLAGS		|
1046 						IB_QP_ALT_PATH			|
1047 						IB_QP_PATH_MIG_STATE		|
1048 						IB_QP_MIN_RNR_TIMER),
1049 				[IB_QPT_SMI] = (IB_QP_CUR_STATE			|
1050 						IB_QP_QKEY),
1051 				[IB_QPT_GSI] = (IB_QP_CUR_STATE			|
1052 						IB_QP_QKEY),
1053 				[IB_QPT_RAW_PACKET] = IB_QP_RATE_LIMIT,
1054 			}
1055 		},
1056 		[IB_QPS_SQD]   = {
1057 			.valid = 1,
1058 			.opt_param = {
1059 				[IB_QPT_UD]  = IB_QP_EN_SQD_ASYNC_NOTIFY,
1060 				[IB_QPT_UC]  = IB_QP_EN_SQD_ASYNC_NOTIFY,
1061 				[IB_QPT_RC]  = IB_QP_EN_SQD_ASYNC_NOTIFY,
1062 				[IB_QPT_XRC_INI] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1063 				[IB_QPT_XRC_TGT] = IB_QP_EN_SQD_ASYNC_NOTIFY, /* ??? */
1064 				[IB_QPT_SMI] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1065 				[IB_QPT_GSI] = IB_QP_EN_SQD_ASYNC_NOTIFY
1066 			}
1067 		},
1068 	},
1069 	[IB_QPS_SQD]   = {
1070 		[IB_QPS_RESET] = { .valid = 1 },
1071 		[IB_QPS_ERR] =   { .valid = 1 },
1072 		[IB_QPS_RTS]   = {
1073 			.valid = 1,
1074 			.opt_param = {
1075 				[IB_QPT_UD]  = (IB_QP_CUR_STATE			|
1076 						IB_QP_QKEY),
1077 				[IB_QPT_UC]  = (IB_QP_CUR_STATE			|
1078 						IB_QP_ALT_PATH			|
1079 						IB_QP_ACCESS_FLAGS		|
1080 						IB_QP_PATH_MIG_STATE),
1081 				[IB_QPT_RC]  = (IB_QP_CUR_STATE			|
1082 						IB_QP_ALT_PATH			|
1083 						IB_QP_ACCESS_FLAGS		|
1084 						IB_QP_MIN_RNR_TIMER		|
1085 						IB_QP_PATH_MIG_STATE),
1086 				[IB_QPT_XRC_INI] = (IB_QP_CUR_STATE		|
1087 						IB_QP_ALT_PATH			|
1088 						IB_QP_ACCESS_FLAGS		|
1089 						IB_QP_PATH_MIG_STATE),
1090 				[IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE		|
1091 						IB_QP_ALT_PATH			|
1092 						IB_QP_ACCESS_FLAGS		|
1093 						IB_QP_MIN_RNR_TIMER		|
1094 						IB_QP_PATH_MIG_STATE),
1095 				[IB_QPT_SMI] = (IB_QP_CUR_STATE			|
1096 						IB_QP_QKEY),
1097 				[IB_QPT_GSI] = (IB_QP_CUR_STATE			|
1098 						IB_QP_QKEY),
1099 			}
1100 		},
1101 		[IB_QPS_SQD]   = {
1102 			.valid = 1,
1103 			.opt_param = {
1104 				[IB_QPT_UD]  = (IB_QP_PKEY_INDEX		|
1105 						IB_QP_QKEY),
1106 				[IB_QPT_UC]  = (IB_QP_AV			|
1107 						IB_QP_ALT_PATH			|
1108 						IB_QP_ACCESS_FLAGS		|
1109 						IB_QP_PKEY_INDEX		|
1110 						IB_QP_PATH_MIG_STATE),
1111 				[IB_QPT_RC]  = (IB_QP_PORT			|
1112 						IB_QP_AV			|
1113 						IB_QP_TIMEOUT			|
1114 						IB_QP_RETRY_CNT			|
1115 						IB_QP_RNR_RETRY			|
1116 						IB_QP_MAX_QP_RD_ATOMIC		|
1117 						IB_QP_MAX_DEST_RD_ATOMIC	|
1118 						IB_QP_ALT_PATH			|
1119 						IB_QP_ACCESS_FLAGS		|
1120 						IB_QP_PKEY_INDEX		|
1121 						IB_QP_MIN_RNR_TIMER		|
1122 						IB_QP_PATH_MIG_STATE),
1123 				[IB_QPT_XRC_INI] = (IB_QP_PORT			|
1124 						IB_QP_AV			|
1125 						IB_QP_TIMEOUT			|
1126 						IB_QP_RETRY_CNT			|
1127 						IB_QP_RNR_RETRY			|
1128 						IB_QP_MAX_QP_RD_ATOMIC		|
1129 						IB_QP_ALT_PATH			|
1130 						IB_QP_ACCESS_FLAGS		|
1131 						IB_QP_PKEY_INDEX		|
1132 						IB_QP_PATH_MIG_STATE),
1133 				[IB_QPT_XRC_TGT] = (IB_QP_PORT			|
1134 						IB_QP_AV			|
1135 						IB_QP_TIMEOUT			|
1136 						IB_QP_MAX_DEST_RD_ATOMIC	|
1137 						IB_QP_ALT_PATH			|
1138 						IB_QP_ACCESS_FLAGS		|
1139 						IB_QP_PKEY_INDEX		|
1140 						IB_QP_MIN_RNR_TIMER		|
1141 						IB_QP_PATH_MIG_STATE),
1142 				[IB_QPT_SMI] = (IB_QP_PKEY_INDEX		|
1143 						IB_QP_QKEY),
1144 				[IB_QPT_GSI] = (IB_QP_PKEY_INDEX		|
1145 						IB_QP_QKEY),
1146 			}
1147 		}
1148 	},
1149 	[IB_QPS_SQE]   = {
1150 		[IB_QPS_RESET] = { .valid = 1 },
1151 		[IB_QPS_ERR] =   { .valid = 1 },
1152 		[IB_QPS_RTS]   = {
1153 			.valid = 1,
1154 			.opt_param = {
1155 				[IB_QPT_UD]  = (IB_QP_CUR_STATE			|
1156 						IB_QP_QKEY),
1157 				[IB_QPT_UC]  = (IB_QP_CUR_STATE			|
1158 						IB_QP_ACCESS_FLAGS),
1159 				[IB_QPT_SMI] = (IB_QP_CUR_STATE			|
1160 						IB_QP_QKEY),
1161 				[IB_QPT_GSI] = (IB_QP_CUR_STATE			|
1162 						IB_QP_QKEY),
1163 			}
1164 		}
1165 	},
1166 	[IB_QPS_ERR] = {
1167 		[IB_QPS_RESET] = { .valid = 1 },
1168 		[IB_QPS_ERR] =   { .valid = 1 }
1169 	}
1170 };
1171 
1172 int ib_modify_qp_is_ok(enum ib_qp_state cur_state, enum ib_qp_state next_state,
1173 		       enum ib_qp_type type, enum ib_qp_attr_mask mask,
1174 		       enum rdma_link_layer ll)
1175 {
1176 	enum ib_qp_attr_mask req_param, opt_param;
1177 
1178 	if (cur_state  < 0 || cur_state  > IB_QPS_ERR ||
1179 	    next_state < 0 || next_state > IB_QPS_ERR)
1180 		return 0;
1181 
1182 	if (mask & IB_QP_CUR_STATE  &&
1183 	    cur_state != IB_QPS_RTR && cur_state != IB_QPS_RTS &&
1184 	    cur_state != IB_QPS_SQD && cur_state != IB_QPS_SQE)
1185 		return 0;
1186 
1187 	if (!qp_state_table[cur_state][next_state].valid)
1188 		return 0;
1189 
1190 	req_param = qp_state_table[cur_state][next_state].req_param[type];
1191 	opt_param = qp_state_table[cur_state][next_state].opt_param[type];
1192 
1193 	if ((mask & req_param) != req_param)
1194 		return 0;
1195 
1196 	if (mask & ~(req_param | opt_param | IB_QP_STATE))
1197 		return 0;
1198 
1199 	return 1;
1200 }
1201 EXPORT_SYMBOL(ib_modify_qp_is_ok);
1202 
1203 int ib_resolve_eth_dmac(struct ib_device *device,
1204 			struct ib_ah_attr *ah_attr)
1205 {
1206 	int           ret = 0;
1207 
1208 	if (!rdma_is_port_valid(device, ah_attr->port_num))
1209 		return -EINVAL;
1210 
1211 	if (!rdma_cap_eth_ah(device, ah_attr->port_num))
1212 		return 0;
1213 
1214 	if (rdma_link_local_addr((struct in6_addr *)ah_attr->grh.dgid.raw)) {
1215 		rdma_get_ll_mac((struct in6_addr *)ah_attr->grh.dgid.raw,
1216 				ah_attr->dmac);
1217 	} else {
1218 		union ib_gid		sgid;
1219 		struct ib_gid_attr	sgid_attr;
1220 		int			ifindex;
1221 		int			hop_limit;
1222 
1223 		ret = ib_query_gid(device,
1224 				   ah_attr->port_num,
1225 				   ah_attr->grh.sgid_index,
1226 				   &sgid, &sgid_attr);
1227 
1228 		if (ret || !sgid_attr.ndev) {
1229 			if (!ret)
1230 				ret = -ENXIO;
1231 			goto out;
1232 		}
1233 
1234 		ifindex = sgid_attr.ndev->ifindex;
1235 
1236 		ret = rdma_addr_find_l2_eth_by_grh(&sgid,
1237 						   &ah_attr->grh.dgid,
1238 						   ah_attr->dmac,
1239 						   NULL, &ifindex, &hop_limit);
1240 
1241 		dev_put(sgid_attr.ndev);
1242 
1243 		ah_attr->grh.hop_limit = hop_limit;
1244 	}
1245 out:
1246 	return ret;
1247 }
1248 EXPORT_SYMBOL(ib_resolve_eth_dmac);
1249 
1250 int ib_modify_qp(struct ib_qp *qp,
1251 		 struct ib_qp_attr *qp_attr,
1252 		 int qp_attr_mask)
1253 {
1254 
1255 	if (qp_attr_mask & IB_QP_AV) {
1256 		int ret;
1257 
1258 		ret = ib_resolve_eth_dmac(qp->device, &qp_attr->ah_attr);
1259 		if (ret)
1260 			return ret;
1261 	}
1262 
1263 	return qp->device->modify_qp(qp->real_qp, qp_attr, qp_attr_mask, NULL);
1264 }
1265 EXPORT_SYMBOL(ib_modify_qp);
1266 
1267 int ib_query_qp(struct ib_qp *qp,
1268 		struct ib_qp_attr *qp_attr,
1269 		int qp_attr_mask,
1270 		struct ib_qp_init_attr *qp_init_attr)
1271 {
1272 	return qp->device->query_qp ?
1273 		qp->device->query_qp(qp->real_qp, qp_attr, qp_attr_mask, qp_init_attr) :
1274 		-ENOSYS;
1275 }
1276 EXPORT_SYMBOL(ib_query_qp);
1277 
1278 int ib_close_qp(struct ib_qp *qp)
1279 {
1280 	struct ib_qp *real_qp;
1281 	unsigned long flags;
1282 
1283 	real_qp = qp->real_qp;
1284 	if (real_qp == qp)
1285 		return -EINVAL;
1286 
1287 	spin_lock_irqsave(&real_qp->device->event_handler_lock, flags);
1288 	list_del(&qp->open_list);
1289 	spin_unlock_irqrestore(&real_qp->device->event_handler_lock, flags);
1290 
1291 	atomic_dec(&real_qp->usecnt);
1292 	kfree(qp);
1293 
1294 	return 0;
1295 }
1296 EXPORT_SYMBOL(ib_close_qp);
1297 
1298 static int __ib_destroy_shared_qp(struct ib_qp *qp)
1299 {
1300 	struct ib_xrcd *xrcd;
1301 	struct ib_qp *real_qp;
1302 	int ret;
1303 
1304 	real_qp = qp->real_qp;
1305 	xrcd = real_qp->xrcd;
1306 
1307 	mutex_lock(&xrcd->tgt_qp_mutex);
1308 	ib_close_qp(qp);
1309 	if (atomic_read(&real_qp->usecnt) == 0)
1310 		list_del(&real_qp->xrcd_list);
1311 	else
1312 		real_qp = NULL;
1313 	mutex_unlock(&xrcd->tgt_qp_mutex);
1314 
1315 	if (real_qp) {
1316 		ret = ib_destroy_qp(real_qp);
1317 		if (!ret)
1318 			atomic_dec(&xrcd->usecnt);
1319 		else
1320 			__ib_insert_xrcd_qp(xrcd, real_qp);
1321 	}
1322 
1323 	return 0;
1324 }
1325 
1326 int ib_destroy_qp(struct ib_qp *qp)
1327 {
1328 	struct ib_pd *pd;
1329 	struct ib_cq *scq, *rcq;
1330 	struct ib_srq *srq;
1331 	struct ib_rwq_ind_table *ind_tbl;
1332 	int ret;
1333 
1334 	WARN_ON_ONCE(qp->mrs_used > 0);
1335 
1336 	if (atomic_read(&qp->usecnt))
1337 		return -EBUSY;
1338 
1339 	if (qp->real_qp != qp)
1340 		return __ib_destroy_shared_qp(qp);
1341 
1342 	pd   = qp->pd;
1343 	scq  = qp->send_cq;
1344 	rcq  = qp->recv_cq;
1345 	srq  = qp->srq;
1346 	ind_tbl = qp->rwq_ind_tbl;
1347 
1348 	if (!qp->uobject)
1349 		rdma_rw_cleanup_mrs(qp);
1350 
1351 	ret = qp->device->destroy_qp(qp);
1352 	if (!ret) {
1353 		if (pd)
1354 			atomic_dec(&pd->usecnt);
1355 		if (scq)
1356 			atomic_dec(&scq->usecnt);
1357 		if (rcq)
1358 			atomic_dec(&rcq->usecnt);
1359 		if (srq)
1360 			atomic_dec(&srq->usecnt);
1361 		if (ind_tbl)
1362 			atomic_dec(&ind_tbl->usecnt);
1363 	}
1364 
1365 	return ret;
1366 }
1367 EXPORT_SYMBOL(ib_destroy_qp);
1368 
1369 /* Completion queues */
1370 
1371 struct ib_cq *ib_create_cq(struct ib_device *device,
1372 			   ib_comp_handler comp_handler,
1373 			   void (*event_handler)(struct ib_event *, void *),
1374 			   void *cq_context,
1375 			   const struct ib_cq_init_attr *cq_attr)
1376 {
1377 	struct ib_cq *cq;
1378 
1379 	cq = device->create_cq(device, cq_attr, NULL, NULL);
1380 
1381 	if (!IS_ERR(cq)) {
1382 		cq->device        = device;
1383 		cq->uobject       = NULL;
1384 		cq->comp_handler  = comp_handler;
1385 		cq->event_handler = event_handler;
1386 		cq->cq_context    = cq_context;
1387 		atomic_set(&cq->usecnt, 0);
1388 	}
1389 
1390 	return cq;
1391 }
1392 EXPORT_SYMBOL(ib_create_cq);
1393 
1394 int ib_modify_cq(struct ib_cq *cq, u16 cq_count, u16 cq_period)
1395 {
1396 	return cq->device->modify_cq ?
1397 		cq->device->modify_cq(cq, cq_count, cq_period) : -ENOSYS;
1398 }
1399 EXPORT_SYMBOL(ib_modify_cq);
1400 
1401 int ib_destroy_cq(struct ib_cq *cq)
1402 {
1403 	if (atomic_read(&cq->usecnt))
1404 		return -EBUSY;
1405 
1406 	return cq->device->destroy_cq(cq);
1407 }
1408 EXPORT_SYMBOL(ib_destroy_cq);
1409 
1410 int ib_resize_cq(struct ib_cq *cq, int cqe)
1411 {
1412 	return cq->device->resize_cq ?
1413 		cq->device->resize_cq(cq, cqe, NULL) : -ENOSYS;
1414 }
1415 EXPORT_SYMBOL(ib_resize_cq);
1416 
1417 /* Memory regions */
1418 
1419 int ib_dereg_mr(struct ib_mr *mr)
1420 {
1421 	struct ib_pd *pd = mr->pd;
1422 	int ret;
1423 
1424 	ret = mr->device->dereg_mr(mr);
1425 	if (!ret)
1426 		atomic_dec(&pd->usecnt);
1427 
1428 	return ret;
1429 }
1430 EXPORT_SYMBOL(ib_dereg_mr);
1431 
1432 /**
1433  * ib_alloc_mr() - Allocates a memory region
1434  * @pd:            protection domain associated with the region
1435  * @mr_type:       memory region type
1436  * @max_num_sg:    maximum sg entries available for registration.
1437  *
1438  * Notes:
1439  * Memory registeration page/sg lists must not exceed max_num_sg.
1440  * For mr_type IB_MR_TYPE_MEM_REG, the total length cannot exceed
1441  * max_num_sg * used_page_size.
1442  *
1443  */
1444 struct ib_mr *ib_alloc_mr(struct ib_pd *pd,
1445 			  enum ib_mr_type mr_type,
1446 			  u32 max_num_sg)
1447 {
1448 	struct ib_mr *mr;
1449 
1450 	if (!pd->device->alloc_mr)
1451 		return ERR_PTR(-ENOSYS);
1452 
1453 	mr = pd->device->alloc_mr(pd, mr_type, max_num_sg);
1454 	if (!IS_ERR(mr)) {
1455 		mr->device  = pd->device;
1456 		mr->pd      = pd;
1457 		mr->uobject = NULL;
1458 		atomic_inc(&pd->usecnt);
1459 		mr->need_inval = false;
1460 	}
1461 
1462 	return mr;
1463 }
1464 EXPORT_SYMBOL(ib_alloc_mr);
1465 
1466 /* "Fast" memory regions */
1467 
1468 struct ib_fmr *ib_alloc_fmr(struct ib_pd *pd,
1469 			    int mr_access_flags,
1470 			    struct ib_fmr_attr *fmr_attr)
1471 {
1472 	struct ib_fmr *fmr;
1473 
1474 	if (!pd->device->alloc_fmr)
1475 		return ERR_PTR(-ENOSYS);
1476 
1477 	fmr = pd->device->alloc_fmr(pd, mr_access_flags, fmr_attr);
1478 	if (!IS_ERR(fmr)) {
1479 		fmr->device = pd->device;
1480 		fmr->pd     = pd;
1481 		atomic_inc(&pd->usecnt);
1482 	}
1483 
1484 	return fmr;
1485 }
1486 EXPORT_SYMBOL(ib_alloc_fmr);
1487 
1488 int ib_unmap_fmr(struct list_head *fmr_list)
1489 {
1490 	struct ib_fmr *fmr;
1491 
1492 	if (list_empty(fmr_list))
1493 		return 0;
1494 
1495 	fmr = list_entry(fmr_list->next, struct ib_fmr, list);
1496 	return fmr->device->unmap_fmr(fmr_list);
1497 }
1498 EXPORT_SYMBOL(ib_unmap_fmr);
1499 
1500 int ib_dealloc_fmr(struct ib_fmr *fmr)
1501 {
1502 	struct ib_pd *pd;
1503 	int ret;
1504 
1505 	pd = fmr->pd;
1506 	ret = fmr->device->dealloc_fmr(fmr);
1507 	if (!ret)
1508 		atomic_dec(&pd->usecnt);
1509 
1510 	return ret;
1511 }
1512 EXPORT_SYMBOL(ib_dealloc_fmr);
1513 
1514 /* Multicast groups */
1515 
1516 int ib_attach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid)
1517 {
1518 	int ret;
1519 
1520 	if (!qp->device->attach_mcast)
1521 		return -ENOSYS;
1522 	if (gid->raw[0] != 0xff || qp->qp_type != IB_QPT_UD)
1523 		return -EINVAL;
1524 
1525 	ret = qp->device->attach_mcast(qp, gid, lid);
1526 	if (!ret)
1527 		atomic_inc(&qp->usecnt);
1528 	return ret;
1529 }
1530 EXPORT_SYMBOL(ib_attach_mcast);
1531 
1532 int ib_detach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid)
1533 {
1534 	int ret;
1535 
1536 	if (!qp->device->detach_mcast)
1537 		return -ENOSYS;
1538 	if (gid->raw[0] != 0xff || qp->qp_type != IB_QPT_UD)
1539 		return -EINVAL;
1540 
1541 	ret = qp->device->detach_mcast(qp, gid, lid);
1542 	if (!ret)
1543 		atomic_dec(&qp->usecnt);
1544 	return ret;
1545 }
1546 EXPORT_SYMBOL(ib_detach_mcast);
1547 
1548 struct ib_xrcd *ib_alloc_xrcd(struct ib_device *device)
1549 {
1550 	struct ib_xrcd *xrcd;
1551 
1552 	if (!device->alloc_xrcd)
1553 		return ERR_PTR(-ENOSYS);
1554 
1555 	xrcd = device->alloc_xrcd(device, NULL, NULL);
1556 	if (!IS_ERR(xrcd)) {
1557 		xrcd->device = device;
1558 		xrcd->inode = NULL;
1559 		atomic_set(&xrcd->usecnt, 0);
1560 		mutex_init(&xrcd->tgt_qp_mutex);
1561 		INIT_LIST_HEAD(&xrcd->tgt_qp_list);
1562 	}
1563 
1564 	return xrcd;
1565 }
1566 EXPORT_SYMBOL(ib_alloc_xrcd);
1567 
1568 int ib_dealloc_xrcd(struct ib_xrcd *xrcd)
1569 {
1570 	struct ib_qp *qp;
1571 	int ret;
1572 
1573 	if (atomic_read(&xrcd->usecnt))
1574 		return -EBUSY;
1575 
1576 	while (!list_empty(&xrcd->tgt_qp_list)) {
1577 		qp = list_entry(xrcd->tgt_qp_list.next, struct ib_qp, xrcd_list);
1578 		ret = ib_destroy_qp(qp);
1579 		if (ret)
1580 			return ret;
1581 	}
1582 
1583 	return xrcd->device->dealloc_xrcd(xrcd);
1584 }
1585 EXPORT_SYMBOL(ib_dealloc_xrcd);
1586 
1587 /**
1588  * ib_create_wq - Creates a WQ associated with the specified protection
1589  * domain.
1590  * @pd: The protection domain associated with the WQ.
1591  * @wq_init_attr: A list of initial attributes required to create the
1592  * WQ. If WQ creation succeeds, then the attributes are updated to
1593  * the actual capabilities of the created WQ.
1594  *
1595  * wq_init_attr->max_wr and wq_init_attr->max_sge determine
1596  * the requested size of the WQ, and set to the actual values allocated
1597  * on return.
1598  * If ib_create_wq() succeeds, then max_wr and max_sge will always be
1599  * at least as large as the requested values.
1600  */
1601 struct ib_wq *ib_create_wq(struct ib_pd *pd,
1602 			   struct ib_wq_init_attr *wq_attr)
1603 {
1604 	struct ib_wq *wq;
1605 
1606 	if (!pd->device->create_wq)
1607 		return ERR_PTR(-ENOSYS);
1608 
1609 	wq = pd->device->create_wq(pd, wq_attr, NULL);
1610 	if (!IS_ERR(wq)) {
1611 		wq->event_handler = wq_attr->event_handler;
1612 		wq->wq_context = wq_attr->wq_context;
1613 		wq->wq_type = wq_attr->wq_type;
1614 		wq->cq = wq_attr->cq;
1615 		wq->device = pd->device;
1616 		wq->pd = pd;
1617 		wq->uobject = NULL;
1618 		atomic_inc(&pd->usecnt);
1619 		atomic_inc(&wq_attr->cq->usecnt);
1620 		atomic_set(&wq->usecnt, 0);
1621 	}
1622 	return wq;
1623 }
1624 EXPORT_SYMBOL(ib_create_wq);
1625 
1626 /**
1627  * ib_destroy_wq - Destroys the specified WQ.
1628  * @wq: The WQ to destroy.
1629  */
1630 int ib_destroy_wq(struct ib_wq *wq)
1631 {
1632 	int err;
1633 	struct ib_cq *cq = wq->cq;
1634 	struct ib_pd *pd = wq->pd;
1635 
1636 	if (atomic_read(&wq->usecnt))
1637 		return -EBUSY;
1638 
1639 	err = wq->device->destroy_wq(wq);
1640 	if (!err) {
1641 		atomic_dec(&pd->usecnt);
1642 		atomic_dec(&cq->usecnt);
1643 	}
1644 	return err;
1645 }
1646 EXPORT_SYMBOL(ib_destroy_wq);
1647 
1648 /**
1649  * ib_modify_wq - Modifies the specified WQ.
1650  * @wq: The WQ to modify.
1651  * @wq_attr: On input, specifies the WQ attributes to modify.
1652  * @wq_attr_mask: A bit-mask used to specify which attributes of the WQ
1653  *   are being modified.
1654  * On output, the current values of selected WQ attributes are returned.
1655  */
1656 int ib_modify_wq(struct ib_wq *wq, struct ib_wq_attr *wq_attr,
1657 		 u32 wq_attr_mask)
1658 {
1659 	int err;
1660 
1661 	if (!wq->device->modify_wq)
1662 		return -ENOSYS;
1663 
1664 	err = wq->device->modify_wq(wq, wq_attr, wq_attr_mask, NULL);
1665 	return err;
1666 }
1667 EXPORT_SYMBOL(ib_modify_wq);
1668 
1669 /*
1670  * ib_create_rwq_ind_table - Creates a RQ Indirection Table.
1671  * @device: The device on which to create the rwq indirection table.
1672  * @ib_rwq_ind_table_init_attr: A list of initial attributes required to
1673  * create the Indirection Table.
1674  *
1675  * Note: The life time of ib_rwq_ind_table_init_attr->ind_tbl is not less
1676  *	than the created ib_rwq_ind_table object and the caller is responsible
1677  *	for its memory allocation/free.
1678  */
1679 struct ib_rwq_ind_table *ib_create_rwq_ind_table(struct ib_device *device,
1680 						 struct ib_rwq_ind_table_init_attr *init_attr)
1681 {
1682 	struct ib_rwq_ind_table *rwq_ind_table;
1683 	int i;
1684 	u32 table_size;
1685 
1686 	if (!device->create_rwq_ind_table)
1687 		return ERR_PTR(-ENOSYS);
1688 
1689 	table_size = (1 << init_attr->log_ind_tbl_size);
1690 	rwq_ind_table = device->create_rwq_ind_table(device,
1691 				init_attr, NULL);
1692 	if (IS_ERR(rwq_ind_table))
1693 		return rwq_ind_table;
1694 
1695 	rwq_ind_table->ind_tbl = init_attr->ind_tbl;
1696 	rwq_ind_table->log_ind_tbl_size = init_attr->log_ind_tbl_size;
1697 	rwq_ind_table->device = device;
1698 	rwq_ind_table->uobject = NULL;
1699 	atomic_set(&rwq_ind_table->usecnt, 0);
1700 
1701 	for (i = 0; i < table_size; i++)
1702 		atomic_inc(&rwq_ind_table->ind_tbl[i]->usecnt);
1703 
1704 	return rwq_ind_table;
1705 }
1706 EXPORT_SYMBOL(ib_create_rwq_ind_table);
1707 
1708 /*
1709  * ib_destroy_rwq_ind_table - Destroys the specified Indirection Table.
1710  * @wq_ind_table: The Indirection Table to destroy.
1711 */
1712 int ib_destroy_rwq_ind_table(struct ib_rwq_ind_table *rwq_ind_table)
1713 {
1714 	int err, i;
1715 	u32 table_size = (1 << rwq_ind_table->log_ind_tbl_size);
1716 	struct ib_wq **ind_tbl = rwq_ind_table->ind_tbl;
1717 
1718 	if (atomic_read(&rwq_ind_table->usecnt))
1719 		return -EBUSY;
1720 
1721 	err = rwq_ind_table->device->destroy_rwq_ind_table(rwq_ind_table);
1722 	if (!err) {
1723 		for (i = 0; i < table_size; i++)
1724 			atomic_dec(&ind_tbl[i]->usecnt);
1725 	}
1726 
1727 	return err;
1728 }
1729 EXPORT_SYMBOL(ib_destroy_rwq_ind_table);
1730 
1731 struct ib_flow *ib_create_flow(struct ib_qp *qp,
1732 			       struct ib_flow_attr *flow_attr,
1733 			       int domain)
1734 {
1735 	struct ib_flow *flow_id;
1736 	if (!qp->device->create_flow)
1737 		return ERR_PTR(-ENOSYS);
1738 
1739 	flow_id = qp->device->create_flow(qp, flow_attr, domain);
1740 	if (!IS_ERR(flow_id)) {
1741 		atomic_inc(&qp->usecnt);
1742 		flow_id->qp = qp;
1743 	}
1744 	return flow_id;
1745 }
1746 EXPORT_SYMBOL(ib_create_flow);
1747 
1748 int ib_destroy_flow(struct ib_flow *flow_id)
1749 {
1750 	int err;
1751 	struct ib_qp *qp = flow_id->qp;
1752 
1753 	err = qp->device->destroy_flow(flow_id);
1754 	if (!err)
1755 		atomic_dec(&qp->usecnt);
1756 	return err;
1757 }
1758 EXPORT_SYMBOL(ib_destroy_flow);
1759 
1760 int ib_check_mr_status(struct ib_mr *mr, u32 check_mask,
1761 		       struct ib_mr_status *mr_status)
1762 {
1763 	return mr->device->check_mr_status ?
1764 		mr->device->check_mr_status(mr, check_mask, mr_status) : -ENOSYS;
1765 }
1766 EXPORT_SYMBOL(ib_check_mr_status);
1767 
1768 int ib_set_vf_link_state(struct ib_device *device, int vf, u8 port,
1769 			 int state)
1770 {
1771 	if (!device->set_vf_link_state)
1772 		return -ENOSYS;
1773 
1774 	return device->set_vf_link_state(device, vf, port, state);
1775 }
1776 EXPORT_SYMBOL(ib_set_vf_link_state);
1777 
1778 int ib_get_vf_config(struct ib_device *device, int vf, u8 port,
1779 		     struct ifla_vf_info *info)
1780 {
1781 	if (!device->get_vf_config)
1782 		return -ENOSYS;
1783 
1784 	return device->get_vf_config(device, vf, port, info);
1785 }
1786 EXPORT_SYMBOL(ib_get_vf_config);
1787 
1788 int ib_get_vf_stats(struct ib_device *device, int vf, u8 port,
1789 		    struct ifla_vf_stats *stats)
1790 {
1791 	if (!device->get_vf_stats)
1792 		return -ENOSYS;
1793 
1794 	return device->get_vf_stats(device, vf, port, stats);
1795 }
1796 EXPORT_SYMBOL(ib_get_vf_stats);
1797 
1798 int ib_set_vf_guid(struct ib_device *device, int vf, u8 port, u64 guid,
1799 		   int type)
1800 {
1801 	if (!device->set_vf_guid)
1802 		return -ENOSYS;
1803 
1804 	return device->set_vf_guid(device, vf, port, guid, type);
1805 }
1806 EXPORT_SYMBOL(ib_set_vf_guid);
1807 
1808 /**
1809  * ib_map_mr_sg() - Map the largest prefix of a dma mapped SG list
1810  *     and set it the memory region.
1811  * @mr:            memory region
1812  * @sg:            dma mapped scatterlist
1813  * @sg_nents:      number of entries in sg
1814  * @sg_offset:     offset in bytes into sg
1815  * @page_size:     page vector desired page size
1816  *
1817  * Constraints:
1818  * - The first sg element is allowed to have an offset.
1819  * - Each sg element must either be aligned to page_size or virtually
1820  *   contiguous to the previous element. In case an sg element has a
1821  *   non-contiguous offset, the mapping prefix will not include it.
1822  * - The last sg element is allowed to have length less than page_size.
1823  * - If sg_nents total byte length exceeds the mr max_num_sge * page_size
1824  *   then only max_num_sg entries will be mapped.
1825  * - If the MR was allocated with type IB_MR_TYPE_SG_GAPS, none of these
1826  *   constraints holds and the page_size argument is ignored.
1827  *
1828  * Returns the number of sg elements that were mapped to the memory region.
1829  *
1830  * After this completes successfully, the  memory region
1831  * is ready for registration.
1832  */
1833 int ib_map_mr_sg(struct ib_mr *mr, struct scatterlist *sg, int sg_nents,
1834 		 unsigned int *sg_offset, unsigned int page_size)
1835 {
1836 	if (unlikely(!mr->device->map_mr_sg))
1837 		return -ENOSYS;
1838 
1839 	mr->page_size = page_size;
1840 
1841 	return mr->device->map_mr_sg(mr, sg, sg_nents, sg_offset);
1842 }
1843 EXPORT_SYMBOL(ib_map_mr_sg);
1844 
1845 /**
1846  * ib_sg_to_pages() - Convert the largest prefix of a sg list
1847  *     to a page vector
1848  * @mr:            memory region
1849  * @sgl:           dma mapped scatterlist
1850  * @sg_nents:      number of entries in sg
1851  * @sg_offset_p:   IN:  start offset in bytes into sg
1852  *                 OUT: offset in bytes for element n of the sg of the first
1853  *                      byte that has not been processed where n is the return
1854  *                      value of this function.
1855  * @set_page:      driver page assignment function pointer
1856  *
1857  * Core service helper for drivers to convert the largest
1858  * prefix of given sg list to a page vector. The sg list
1859  * prefix converted is the prefix that meet the requirements
1860  * of ib_map_mr_sg.
1861  *
1862  * Returns the number of sg elements that were assigned to
1863  * a page vector.
1864  */
1865 int ib_sg_to_pages(struct ib_mr *mr, struct scatterlist *sgl, int sg_nents,
1866 		unsigned int *sg_offset_p, int (*set_page)(struct ib_mr *, u64))
1867 {
1868 	struct scatterlist *sg;
1869 	u64 last_end_dma_addr = 0;
1870 	unsigned int sg_offset = sg_offset_p ? *sg_offset_p : 0;
1871 	unsigned int last_page_off = 0;
1872 	u64 page_mask = ~((u64)mr->page_size - 1);
1873 	int i, ret;
1874 
1875 	if (unlikely(sg_nents <= 0 || sg_offset > sg_dma_len(&sgl[0])))
1876 		return -EINVAL;
1877 
1878 	mr->iova = sg_dma_address(&sgl[0]) + sg_offset;
1879 	mr->length = 0;
1880 
1881 	for_each_sg(sgl, sg, sg_nents, i) {
1882 		u64 dma_addr = sg_dma_address(sg) + sg_offset;
1883 		u64 prev_addr = dma_addr;
1884 		unsigned int dma_len = sg_dma_len(sg) - sg_offset;
1885 		u64 end_dma_addr = dma_addr + dma_len;
1886 		u64 page_addr = dma_addr & page_mask;
1887 
1888 		/*
1889 		 * For the second and later elements, check whether either the
1890 		 * end of element i-1 or the start of element i is not aligned
1891 		 * on a page boundary.
1892 		 */
1893 		if (i && (last_page_off != 0 || page_addr != dma_addr)) {
1894 			/* Stop mapping if there is a gap. */
1895 			if (last_end_dma_addr != dma_addr)
1896 				break;
1897 
1898 			/*
1899 			 * Coalesce this element with the last. If it is small
1900 			 * enough just update mr->length. Otherwise start
1901 			 * mapping from the next page.
1902 			 */
1903 			goto next_page;
1904 		}
1905 
1906 		do {
1907 			ret = set_page(mr, page_addr);
1908 			if (unlikely(ret < 0)) {
1909 				sg_offset = prev_addr - sg_dma_address(sg);
1910 				mr->length += prev_addr - dma_addr;
1911 				if (sg_offset_p)
1912 					*sg_offset_p = sg_offset;
1913 				return i || sg_offset ? i : ret;
1914 			}
1915 			prev_addr = page_addr;
1916 next_page:
1917 			page_addr += mr->page_size;
1918 		} while (page_addr < end_dma_addr);
1919 
1920 		mr->length += dma_len;
1921 		last_end_dma_addr = end_dma_addr;
1922 		last_page_off = end_dma_addr & ~page_mask;
1923 
1924 		sg_offset = 0;
1925 	}
1926 
1927 	if (sg_offset_p)
1928 		*sg_offset_p = 0;
1929 	return i;
1930 }
1931 EXPORT_SYMBOL(ib_sg_to_pages);
1932 
1933 struct ib_drain_cqe {
1934 	struct ib_cqe cqe;
1935 	struct completion done;
1936 };
1937 
1938 static void ib_drain_qp_done(struct ib_cq *cq, struct ib_wc *wc)
1939 {
1940 	struct ib_drain_cqe *cqe = container_of(wc->wr_cqe, struct ib_drain_cqe,
1941 						cqe);
1942 
1943 	complete(&cqe->done);
1944 }
1945 
1946 /*
1947  * Post a WR and block until its completion is reaped for the SQ.
1948  */
1949 static void __ib_drain_sq(struct ib_qp *qp)
1950 {
1951 	struct ib_cq *cq = qp->send_cq;
1952 	struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
1953 	struct ib_drain_cqe sdrain;
1954 	struct ib_send_wr swr = {}, *bad_swr;
1955 	int ret;
1956 
1957 	swr.wr_cqe = &sdrain.cqe;
1958 	sdrain.cqe.done = ib_drain_qp_done;
1959 	init_completion(&sdrain.done);
1960 
1961 	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
1962 	if (ret) {
1963 		WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
1964 		return;
1965 	}
1966 
1967 	ret = ib_post_send(qp, &swr, &bad_swr);
1968 	if (ret) {
1969 		WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
1970 		return;
1971 	}
1972 
1973 	if (cq->poll_ctx == IB_POLL_DIRECT)
1974 		while (wait_for_completion_timeout(&sdrain.done, HZ / 10) <= 0)
1975 			ib_process_cq_direct(cq, -1);
1976 	else
1977 		wait_for_completion(&sdrain.done);
1978 }
1979 
1980 /*
1981  * Post a WR and block until its completion is reaped for the RQ.
1982  */
1983 static void __ib_drain_rq(struct ib_qp *qp)
1984 {
1985 	struct ib_cq *cq = qp->recv_cq;
1986 	struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
1987 	struct ib_drain_cqe rdrain;
1988 	struct ib_recv_wr rwr = {}, *bad_rwr;
1989 	int ret;
1990 
1991 	rwr.wr_cqe = &rdrain.cqe;
1992 	rdrain.cqe.done = ib_drain_qp_done;
1993 	init_completion(&rdrain.done);
1994 
1995 	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
1996 	if (ret) {
1997 		WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
1998 		return;
1999 	}
2000 
2001 	ret = ib_post_recv(qp, &rwr, &bad_rwr);
2002 	if (ret) {
2003 		WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
2004 		return;
2005 	}
2006 
2007 	if (cq->poll_ctx == IB_POLL_DIRECT)
2008 		while (wait_for_completion_timeout(&rdrain.done, HZ / 10) <= 0)
2009 			ib_process_cq_direct(cq, -1);
2010 	else
2011 		wait_for_completion(&rdrain.done);
2012 }
2013 
2014 /**
2015  * ib_drain_sq() - Block until all SQ CQEs have been consumed by the
2016  *		   application.
2017  * @qp:            queue pair to drain
2018  *
2019  * If the device has a provider-specific drain function, then
2020  * call that.  Otherwise call the generic drain function
2021  * __ib_drain_sq().
2022  *
2023  * The caller must:
2024  *
2025  * ensure there is room in the CQ and SQ for the drain work request and
2026  * completion.
2027  *
2028  * allocate the CQ using ib_alloc_cq().
2029  *
2030  * ensure that there are no other contexts that are posting WRs concurrently.
2031  * Otherwise the drain is not guaranteed.
2032  */
2033 void ib_drain_sq(struct ib_qp *qp)
2034 {
2035 	if (qp->device->drain_sq)
2036 		qp->device->drain_sq(qp);
2037 	else
2038 		__ib_drain_sq(qp);
2039 }
2040 EXPORT_SYMBOL(ib_drain_sq);
2041 
2042 /**
2043  * ib_drain_rq() - Block until all RQ CQEs have been consumed by the
2044  *		   application.
2045  * @qp:            queue pair to drain
2046  *
2047  * If the device has a provider-specific drain function, then
2048  * call that.  Otherwise call the generic drain function
2049  * __ib_drain_rq().
2050  *
2051  * The caller must:
2052  *
2053  * ensure there is room in the CQ and RQ for the drain work request and
2054  * completion.
2055  *
2056  * allocate the CQ using ib_alloc_cq().
2057  *
2058  * ensure that there are no other contexts that are posting WRs concurrently.
2059  * Otherwise the drain is not guaranteed.
2060  */
2061 void ib_drain_rq(struct ib_qp *qp)
2062 {
2063 	if (qp->device->drain_rq)
2064 		qp->device->drain_rq(qp);
2065 	else
2066 		__ib_drain_rq(qp);
2067 }
2068 EXPORT_SYMBOL(ib_drain_rq);
2069 
2070 /**
2071  * ib_drain_qp() - Block until all CQEs have been consumed by the
2072  *		   application on both the RQ and SQ.
2073  * @qp:            queue pair to drain
2074  *
2075  * The caller must:
2076  *
2077  * ensure there is room in the CQ(s), SQ, and RQ for drain work requests
2078  * and completions.
2079  *
2080  * allocate the CQs using ib_alloc_cq().
2081  *
2082  * ensure that there are no other contexts that are posting WRs concurrently.
2083  * Otherwise the drain is not guaranteed.
2084  */
2085 void ib_drain_qp(struct ib_qp *qp)
2086 {
2087 	ib_drain_sq(qp);
2088 	if (!qp->srq)
2089 		ib_drain_rq(qp);
2090 }
2091 EXPORT_SYMBOL(ib_drain_qp);
2092