xref: /openbmc/linux/drivers/infiniband/core/verbs.c (revision 6aa7de05)
1 /*
2  * Copyright (c) 2004 Mellanox Technologies Ltd.  All rights reserved.
3  * Copyright (c) 2004 Infinicon Corporation.  All rights reserved.
4  * Copyright (c) 2004 Intel Corporation.  All rights reserved.
5  * Copyright (c) 2004 Topspin Corporation.  All rights reserved.
6  * Copyright (c) 2004 Voltaire Corporation.  All rights reserved.
7  * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
8  * Copyright (c) 2005, 2006 Cisco Systems.  All rights reserved.
9  *
10  * This software is available to you under a choice of one of two
11  * licenses.  You may choose to be licensed under the terms of the GNU
12  * General Public License (GPL) Version 2, available from the file
13  * COPYING in the main directory of this source tree, or the
14  * OpenIB.org BSD license below:
15  *
16  *     Redistribution and use in source and binary forms, with or
17  *     without modification, are permitted provided that the following
18  *     conditions are met:
19  *
20  *      - Redistributions of source code must retain the above
21  *        copyright notice, this list of conditions and the following
22  *        disclaimer.
23  *
24  *      - Redistributions in binary form must reproduce the above
25  *        copyright notice, this list of conditions and the following
26  *        disclaimer in the documentation and/or other materials
27  *        provided with the distribution.
28  *
29  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
30  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
31  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
32  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
33  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
34  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
35  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
36  * SOFTWARE.
37  */
38 
39 #include <linux/errno.h>
40 #include <linux/err.h>
41 #include <linux/export.h>
42 #include <linux/string.h>
43 #include <linux/slab.h>
44 #include <linux/in.h>
45 #include <linux/in6.h>
46 #include <net/addrconf.h>
47 #include <linux/security.h>
48 
49 #include <rdma/ib_verbs.h>
50 #include <rdma/ib_cache.h>
51 #include <rdma/ib_addr.h>
52 #include <rdma/rw.h>
53 
54 #include "core_priv.h"
55 
56 static const char * const ib_events[] = {
57 	[IB_EVENT_CQ_ERR]		= "CQ error",
58 	[IB_EVENT_QP_FATAL]		= "QP fatal error",
59 	[IB_EVENT_QP_REQ_ERR]		= "QP request error",
60 	[IB_EVENT_QP_ACCESS_ERR]	= "QP access error",
61 	[IB_EVENT_COMM_EST]		= "communication established",
62 	[IB_EVENT_SQ_DRAINED]		= "send queue drained",
63 	[IB_EVENT_PATH_MIG]		= "path migration successful",
64 	[IB_EVENT_PATH_MIG_ERR]		= "path migration error",
65 	[IB_EVENT_DEVICE_FATAL]		= "device fatal error",
66 	[IB_EVENT_PORT_ACTIVE]		= "port active",
67 	[IB_EVENT_PORT_ERR]		= "port error",
68 	[IB_EVENT_LID_CHANGE]		= "LID change",
69 	[IB_EVENT_PKEY_CHANGE]		= "P_key change",
70 	[IB_EVENT_SM_CHANGE]		= "SM change",
71 	[IB_EVENT_SRQ_ERR]		= "SRQ error",
72 	[IB_EVENT_SRQ_LIMIT_REACHED]	= "SRQ limit reached",
73 	[IB_EVENT_QP_LAST_WQE_REACHED]	= "last WQE reached",
74 	[IB_EVENT_CLIENT_REREGISTER]	= "client reregister",
75 	[IB_EVENT_GID_CHANGE]		= "GID changed",
76 };
77 
78 const char *__attribute_const__ ib_event_msg(enum ib_event_type event)
79 {
80 	size_t index = event;
81 
82 	return (index < ARRAY_SIZE(ib_events) && ib_events[index]) ?
83 			ib_events[index] : "unrecognized event";
84 }
85 EXPORT_SYMBOL(ib_event_msg);
86 
87 static const char * const wc_statuses[] = {
88 	[IB_WC_SUCCESS]			= "success",
89 	[IB_WC_LOC_LEN_ERR]		= "local length error",
90 	[IB_WC_LOC_QP_OP_ERR]		= "local QP operation error",
91 	[IB_WC_LOC_EEC_OP_ERR]		= "local EE context operation error",
92 	[IB_WC_LOC_PROT_ERR]		= "local protection error",
93 	[IB_WC_WR_FLUSH_ERR]		= "WR flushed",
94 	[IB_WC_MW_BIND_ERR]		= "memory management operation error",
95 	[IB_WC_BAD_RESP_ERR]		= "bad response error",
96 	[IB_WC_LOC_ACCESS_ERR]		= "local access error",
97 	[IB_WC_REM_INV_REQ_ERR]		= "invalid request error",
98 	[IB_WC_REM_ACCESS_ERR]		= "remote access error",
99 	[IB_WC_REM_OP_ERR]		= "remote operation error",
100 	[IB_WC_RETRY_EXC_ERR]		= "transport retry counter exceeded",
101 	[IB_WC_RNR_RETRY_EXC_ERR]	= "RNR retry counter exceeded",
102 	[IB_WC_LOC_RDD_VIOL_ERR]	= "local RDD violation error",
103 	[IB_WC_REM_INV_RD_REQ_ERR]	= "remote invalid RD request",
104 	[IB_WC_REM_ABORT_ERR]		= "operation aborted",
105 	[IB_WC_INV_EECN_ERR]		= "invalid EE context number",
106 	[IB_WC_INV_EEC_STATE_ERR]	= "invalid EE context state",
107 	[IB_WC_FATAL_ERR]		= "fatal error",
108 	[IB_WC_RESP_TIMEOUT_ERR]	= "response timeout error",
109 	[IB_WC_GENERAL_ERR]		= "general error",
110 };
111 
112 const char *__attribute_const__ ib_wc_status_msg(enum ib_wc_status status)
113 {
114 	size_t index = status;
115 
116 	return (index < ARRAY_SIZE(wc_statuses) && wc_statuses[index]) ?
117 			wc_statuses[index] : "unrecognized status";
118 }
119 EXPORT_SYMBOL(ib_wc_status_msg);
120 
121 __attribute_const__ int ib_rate_to_mult(enum ib_rate rate)
122 {
123 	switch (rate) {
124 	case IB_RATE_2_5_GBPS: return  1;
125 	case IB_RATE_5_GBPS:   return  2;
126 	case IB_RATE_10_GBPS:  return  4;
127 	case IB_RATE_20_GBPS:  return  8;
128 	case IB_RATE_30_GBPS:  return 12;
129 	case IB_RATE_40_GBPS:  return 16;
130 	case IB_RATE_60_GBPS:  return 24;
131 	case IB_RATE_80_GBPS:  return 32;
132 	case IB_RATE_120_GBPS: return 48;
133 	default:	       return -1;
134 	}
135 }
136 EXPORT_SYMBOL(ib_rate_to_mult);
137 
138 __attribute_const__ enum ib_rate mult_to_ib_rate(int mult)
139 {
140 	switch (mult) {
141 	case 1:  return IB_RATE_2_5_GBPS;
142 	case 2:  return IB_RATE_5_GBPS;
143 	case 4:  return IB_RATE_10_GBPS;
144 	case 8:  return IB_RATE_20_GBPS;
145 	case 12: return IB_RATE_30_GBPS;
146 	case 16: return IB_RATE_40_GBPS;
147 	case 24: return IB_RATE_60_GBPS;
148 	case 32: return IB_RATE_80_GBPS;
149 	case 48: return IB_RATE_120_GBPS;
150 	default: return IB_RATE_PORT_CURRENT;
151 	}
152 }
153 EXPORT_SYMBOL(mult_to_ib_rate);
154 
155 __attribute_const__ int ib_rate_to_mbps(enum ib_rate rate)
156 {
157 	switch (rate) {
158 	case IB_RATE_2_5_GBPS: return 2500;
159 	case IB_RATE_5_GBPS:   return 5000;
160 	case IB_RATE_10_GBPS:  return 10000;
161 	case IB_RATE_20_GBPS:  return 20000;
162 	case IB_RATE_30_GBPS:  return 30000;
163 	case IB_RATE_40_GBPS:  return 40000;
164 	case IB_RATE_60_GBPS:  return 60000;
165 	case IB_RATE_80_GBPS:  return 80000;
166 	case IB_RATE_120_GBPS: return 120000;
167 	case IB_RATE_14_GBPS:  return 14062;
168 	case IB_RATE_56_GBPS:  return 56250;
169 	case IB_RATE_112_GBPS: return 112500;
170 	case IB_RATE_168_GBPS: return 168750;
171 	case IB_RATE_25_GBPS:  return 25781;
172 	case IB_RATE_100_GBPS: return 103125;
173 	case IB_RATE_200_GBPS: return 206250;
174 	case IB_RATE_300_GBPS: return 309375;
175 	default:	       return -1;
176 	}
177 }
178 EXPORT_SYMBOL(ib_rate_to_mbps);
179 
180 __attribute_const__ enum rdma_transport_type
181 rdma_node_get_transport(enum rdma_node_type node_type)
182 {
183 
184 	if (node_type == RDMA_NODE_USNIC)
185 		return RDMA_TRANSPORT_USNIC;
186 	if (node_type == RDMA_NODE_USNIC_UDP)
187 		return RDMA_TRANSPORT_USNIC_UDP;
188 	if (node_type == RDMA_NODE_RNIC)
189 		return RDMA_TRANSPORT_IWARP;
190 
191 	return RDMA_TRANSPORT_IB;
192 }
193 EXPORT_SYMBOL(rdma_node_get_transport);
194 
195 enum rdma_link_layer rdma_port_get_link_layer(struct ib_device *device, u8 port_num)
196 {
197 	enum rdma_transport_type lt;
198 	if (device->get_link_layer)
199 		return device->get_link_layer(device, port_num);
200 
201 	lt = rdma_node_get_transport(device->node_type);
202 	if (lt == RDMA_TRANSPORT_IB)
203 		return IB_LINK_LAYER_INFINIBAND;
204 
205 	return IB_LINK_LAYER_ETHERNET;
206 }
207 EXPORT_SYMBOL(rdma_port_get_link_layer);
208 
209 /* Protection domains */
210 
211 /**
212  * ib_alloc_pd - Allocates an unused protection domain.
213  * @device: The device on which to allocate the protection domain.
214  *
215  * A protection domain object provides an association between QPs, shared
216  * receive queues, address handles, memory regions, and memory windows.
217  *
218  * Every PD has a local_dma_lkey which can be used as the lkey value for local
219  * memory operations.
220  */
221 struct ib_pd *__ib_alloc_pd(struct ib_device *device, unsigned int flags,
222 		const char *caller)
223 {
224 	struct ib_pd *pd;
225 	int mr_access_flags = 0;
226 
227 	pd = device->alloc_pd(device, NULL, NULL);
228 	if (IS_ERR(pd))
229 		return pd;
230 
231 	pd->device = device;
232 	pd->uobject = NULL;
233 	pd->__internal_mr = NULL;
234 	atomic_set(&pd->usecnt, 0);
235 	pd->flags = flags;
236 
237 	if (device->attrs.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY)
238 		pd->local_dma_lkey = device->local_dma_lkey;
239 	else
240 		mr_access_flags |= IB_ACCESS_LOCAL_WRITE;
241 
242 	if (flags & IB_PD_UNSAFE_GLOBAL_RKEY) {
243 		pr_warn("%s: enabling unsafe global rkey\n", caller);
244 		mr_access_flags |= IB_ACCESS_REMOTE_READ | IB_ACCESS_REMOTE_WRITE;
245 	}
246 
247 	if (mr_access_flags) {
248 		struct ib_mr *mr;
249 
250 		mr = pd->device->get_dma_mr(pd, mr_access_flags);
251 		if (IS_ERR(mr)) {
252 			ib_dealloc_pd(pd);
253 			return ERR_CAST(mr);
254 		}
255 
256 		mr->device	= pd->device;
257 		mr->pd		= pd;
258 		mr->uobject	= NULL;
259 		mr->need_inval	= false;
260 
261 		pd->__internal_mr = mr;
262 
263 		if (!(device->attrs.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY))
264 			pd->local_dma_lkey = pd->__internal_mr->lkey;
265 
266 		if (flags & IB_PD_UNSAFE_GLOBAL_RKEY)
267 			pd->unsafe_global_rkey = pd->__internal_mr->rkey;
268 	}
269 
270 	return pd;
271 }
272 EXPORT_SYMBOL(__ib_alloc_pd);
273 
274 /**
275  * ib_dealloc_pd - Deallocates a protection domain.
276  * @pd: The protection domain to deallocate.
277  *
278  * It is an error to call this function while any resources in the pd still
279  * exist.  The caller is responsible to synchronously destroy them and
280  * guarantee no new allocations will happen.
281  */
282 void ib_dealloc_pd(struct ib_pd *pd)
283 {
284 	int ret;
285 
286 	if (pd->__internal_mr) {
287 		ret = pd->device->dereg_mr(pd->__internal_mr);
288 		WARN_ON(ret);
289 		pd->__internal_mr = NULL;
290 	}
291 
292 	/* uverbs manipulates usecnt with proper locking, while the kabi
293 	   requires the caller to guarantee we can't race here. */
294 	WARN_ON(atomic_read(&pd->usecnt));
295 
296 	/* Making delalloc_pd a void return is a WIP, no driver should return
297 	   an error here. */
298 	ret = pd->device->dealloc_pd(pd);
299 	WARN_ONCE(ret, "Infiniband HW driver failed dealloc_pd");
300 }
301 EXPORT_SYMBOL(ib_dealloc_pd);
302 
303 /* Address handles */
304 
305 struct ib_ah *rdma_create_ah(struct ib_pd *pd, struct rdma_ah_attr *ah_attr)
306 {
307 	struct ib_ah *ah;
308 
309 	ah = pd->device->create_ah(pd, ah_attr, NULL);
310 
311 	if (!IS_ERR(ah)) {
312 		ah->device  = pd->device;
313 		ah->pd      = pd;
314 		ah->uobject = NULL;
315 		ah->type    = ah_attr->type;
316 		atomic_inc(&pd->usecnt);
317 	}
318 
319 	return ah;
320 }
321 EXPORT_SYMBOL(rdma_create_ah);
322 
323 int ib_get_rdma_header_version(const union rdma_network_hdr *hdr)
324 {
325 	const struct iphdr *ip4h = (struct iphdr *)&hdr->roce4grh;
326 	struct iphdr ip4h_checked;
327 	const struct ipv6hdr *ip6h = (struct ipv6hdr *)&hdr->ibgrh;
328 
329 	/* If it's IPv6, the version must be 6, otherwise, the first
330 	 * 20 bytes (before the IPv4 header) are garbled.
331 	 */
332 	if (ip6h->version != 6)
333 		return (ip4h->version == 4) ? 4 : 0;
334 	/* version may be 6 or 4 because the first 20 bytes could be garbled */
335 
336 	/* RoCE v2 requires no options, thus header length
337 	 * must be 5 words
338 	 */
339 	if (ip4h->ihl != 5)
340 		return 6;
341 
342 	/* Verify checksum.
343 	 * We can't write on scattered buffers so we need to copy to
344 	 * temp buffer.
345 	 */
346 	memcpy(&ip4h_checked, ip4h, sizeof(ip4h_checked));
347 	ip4h_checked.check = 0;
348 	ip4h_checked.check = ip_fast_csum((u8 *)&ip4h_checked, 5);
349 	/* if IPv4 header checksum is OK, believe it */
350 	if (ip4h->check == ip4h_checked.check)
351 		return 4;
352 	return 6;
353 }
354 EXPORT_SYMBOL(ib_get_rdma_header_version);
355 
356 static enum rdma_network_type ib_get_net_type_by_grh(struct ib_device *device,
357 						     u8 port_num,
358 						     const struct ib_grh *grh)
359 {
360 	int grh_version;
361 
362 	if (rdma_protocol_ib(device, port_num))
363 		return RDMA_NETWORK_IB;
364 
365 	grh_version = ib_get_rdma_header_version((union rdma_network_hdr *)grh);
366 
367 	if (grh_version == 4)
368 		return RDMA_NETWORK_IPV4;
369 
370 	if (grh->next_hdr == IPPROTO_UDP)
371 		return RDMA_NETWORK_IPV6;
372 
373 	return RDMA_NETWORK_ROCE_V1;
374 }
375 
376 struct find_gid_index_context {
377 	u16 vlan_id;
378 	enum ib_gid_type gid_type;
379 };
380 
381 static bool find_gid_index(const union ib_gid *gid,
382 			   const struct ib_gid_attr *gid_attr,
383 			   void *context)
384 {
385 	struct find_gid_index_context *ctx =
386 		(struct find_gid_index_context *)context;
387 
388 	if (ctx->gid_type != gid_attr->gid_type)
389 		return false;
390 
391 	if ((!!(ctx->vlan_id != 0xffff) == !is_vlan_dev(gid_attr->ndev)) ||
392 	    (is_vlan_dev(gid_attr->ndev) &&
393 	     vlan_dev_vlan_id(gid_attr->ndev) != ctx->vlan_id))
394 		return false;
395 
396 	return true;
397 }
398 
399 static int get_sgid_index_from_eth(struct ib_device *device, u8 port_num,
400 				   u16 vlan_id, const union ib_gid *sgid,
401 				   enum ib_gid_type gid_type,
402 				   u16 *gid_index)
403 {
404 	struct find_gid_index_context context = {.vlan_id = vlan_id,
405 						 .gid_type = gid_type};
406 
407 	return ib_find_gid_by_filter(device, sgid, port_num, find_gid_index,
408 				     &context, gid_index);
409 }
410 
411 int ib_get_gids_from_rdma_hdr(const union rdma_network_hdr *hdr,
412 			      enum rdma_network_type net_type,
413 			      union ib_gid *sgid, union ib_gid *dgid)
414 {
415 	struct sockaddr_in  src_in;
416 	struct sockaddr_in  dst_in;
417 	__be32 src_saddr, dst_saddr;
418 
419 	if (!sgid || !dgid)
420 		return -EINVAL;
421 
422 	if (net_type == RDMA_NETWORK_IPV4) {
423 		memcpy(&src_in.sin_addr.s_addr,
424 		       &hdr->roce4grh.saddr, 4);
425 		memcpy(&dst_in.sin_addr.s_addr,
426 		       &hdr->roce4grh.daddr, 4);
427 		src_saddr = src_in.sin_addr.s_addr;
428 		dst_saddr = dst_in.sin_addr.s_addr;
429 		ipv6_addr_set_v4mapped(src_saddr,
430 				       (struct in6_addr *)sgid);
431 		ipv6_addr_set_v4mapped(dst_saddr,
432 				       (struct in6_addr *)dgid);
433 		return 0;
434 	} else if (net_type == RDMA_NETWORK_IPV6 ||
435 		   net_type == RDMA_NETWORK_IB) {
436 		*dgid = hdr->ibgrh.dgid;
437 		*sgid = hdr->ibgrh.sgid;
438 		return 0;
439 	} else {
440 		return -EINVAL;
441 	}
442 }
443 EXPORT_SYMBOL(ib_get_gids_from_rdma_hdr);
444 
445 /*
446  * This function creates ah from the incoming packet.
447  * Incoming packet has dgid of the receiver node on which this code is
448  * getting executed and, sgid contains the GID of the sender.
449  *
450  * When resolving mac address of destination, the arrived dgid is used
451  * as sgid and, sgid is used as dgid because sgid contains destinations
452  * GID whom to respond to.
453  *
454  * This is why when calling rdma_addr_find_l2_eth_by_grh() function, the
455  * position of arguments dgid and sgid do not match the order of the
456  * parameters.
457  */
458 int ib_init_ah_from_wc(struct ib_device *device, u8 port_num,
459 		       const struct ib_wc *wc, const struct ib_grh *grh,
460 		       struct rdma_ah_attr *ah_attr)
461 {
462 	u32 flow_class;
463 	u16 gid_index;
464 	int ret;
465 	enum rdma_network_type net_type = RDMA_NETWORK_IB;
466 	enum ib_gid_type gid_type = IB_GID_TYPE_IB;
467 	int hoplimit = 0xff;
468 	union ib_gid dgid;
469 	union ib_gid sgid;
470 
471 	might_sleep();
472 
473 	memset(ah_attr, 0, sizeof *ah_attr);
474 	ah_attr->type = rdma_ah_find_type(device, port_num);
475 	if (rdma_cap_eth_ah(device, port_num)) {
476 		if (wc->wc_flags & IB_WC_WITH_NETWORK_HDR_TYPE)
477 			net_type = wc->network_hdr_type;
478 		else
479 			net_type = ib_get_net_type_by_grh(device, port_num, grh);
480 		gid_type = ib_network_to_gid_type(net_type);
481 	}
482 	ret = ib_get_gids_from_rdma_hdr((union rdma_network_hdr *)grh, net_type,
483 					&sgid, &dgid);
484 	if (ret)
485 		return ret;
486 
487 	if (rdma_protocol_roce(device, port_num)) {
488 		int if_index = 0;
489 		u16 vlan_id = wc->wc_flags & IB_WC_WITH_VLAN ?
490 				wc->vlan_id : 0xffff;
491 		struct net_device *idev;
492 		struct net_device *resolved_dev;
493 
494 		if (!(wc->wc_flags & IB_WC_GRH))
495 			return -EPROTOTYPE;
496 
497 		if (!device->get_netdev)
498 			return -EOPNOTSUPP;
499 
500 		idev = device->get_netdev(device, port_num);
501 		if (!idev)
502 			return -ENODEV;
503 
504 		ret = rdma_addr_find_l2_eth_by_grh(&dgid, &sgid,
505 						   ah_attr->roce.dmac,
506 						   wc->wc_flags & IB_WC_WITH_VLAN ?
507 						   NULL : &vlan_id,
508 						   &if_index, &hoplimit);
509 		if (ret) {
510 			dev_put(idev);
511 			return ret;
512 		}
513 
514 		resolved_dev = dev_get_by_index(&init_net, if_index);
515 		rcu_read_lock();
516 		if (resolved_dev != idev && !rdma_is_upper_dev_rcu(idev,
517 								   resolved_dev))
518 			ret = -EHOSTUNREACH;
519 		rcu_read_unlock();
520 		dev_put(idev);
521 		dev_put(resolved_dev);
522 		if (ret)
523 			return ret;
524 
525 		ret = get_sgid_index_from_eth(device, port_num, vlan_id,
526 					      &dgid, gid_type, &gid_index);
527 		if (ret)
528 			return ret;
529 	}
530 
531 	rdma_ah_set_dlid(ah_attr, wc->slid);
532 	rdma_ah_set_sl(ah_attr, wc->sl);
533 	rdma_ah_set_path_bits(ah_attr, wc->dlid_path_bits);
534 	rdma_ah_set_port_num(ah_attr, port_num);
535 
536 	if (wc->wc_flags & IB_WC_GRH) {
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 		flow_class = be32_to_cpu(grh->version_tclass_flow);
551 		rdma_ah_set_grh(ah_attr, &sgid,
552 				flow_class & 0xFFFFF,
553 				(u8)gid_index, hoplimit,
554 				(flow_class >> 20) & 0xFF);
555 
556 	}
557 	return 0;
558 }
559 EXPORT_SYMBOL(ib_init_ah_from_wc);
560 
561 struct ib_ah *ib_create_ah_from_wc(struct ib_pd *pd, const struct ib_wc *wc,
562 				   const struct ib_grh *grh, u8 port_num)
563 {
564 	struct rdma_ah_attr ah_attr;
565 	int ret;
566 
567 	ret = ib_init_ah_from_wc(pd->device, port_num, wc, grh, &ah_attr);
568 	if (ret)
569 		return ERR_PTR(ret);
570 
571 	return rdma_create_ah(pd, &ah_attr);
572 }
573 EXPORT_SYMBOL(ib_create_ah_from_wc);
574 
575 int rdma_modify_ah(struct ib_ah *ah, struct rdma_ah_attr *ah_attr)
576 {
577 	if (ah->type != ah_attr->type)
578 		return -EINVAL;
579 
580 	return ah->device->modify_ah ?
581 		ah->device->modify_ah(ah, ah_attr) :
582 		-ENOSYS;
583 }
584 EXPORT_SYMBOL(rdma_modify_ah);
585 
586 int rdma_query_ah(struct ib_ah *ah, struct rdma_ah_attr *ah_attr)
587 {
588 	return ah->device->query_ah ?
589 		ah->device->query_ah(ah, ah_attr) :
590 		-ENOSYS;
591 }
592 EXPORT_SYMBOL(rdma_query_ah);
593 
594 int rdma_destroy_ah(struct ib_ah *ah)
595 {
596 	struct ib_pd *pd;
597 	int ret;
598 
599 	pd = ah->pd;
600 	ret = ah->device->destroy_ah(ah);
601 	if (!ret)
602 		atomic_dec(&pd->usecnt);
603 
604 	return ret;
605 }
606 EXPORT_SYMBOL(rdma_destroy_ah);
607 
608 /* Shared receive queues */
609 
610 struct ib_srq *ib_create_srq(struct ib_pd *pd,
611 			     struct ib_srq_init_attr *srq_init_attr)
612 {
613 	struct ib_srq *srq;
614 
615 	if (!pd->device->create_srq)
616 		return ERR_PTR(-ENOSYS);
617 
618 	srq = pd->device->create_srq(pd, srq_init_attr, NULL);
619 
620 	if (!IS_ERR(srq)) {
621 		srq->device    	   = pd->device;
622 		srq->pd        	   = pd;
623 		srq->uobject       = NULL;
624 		srq->event_handler = srq_init_attr->event_handler;
625 		srq->srq_context   = srq_init_attr->srq_context;
626 		srq->srq_type      = srq_init_attr->srq_type;
627 		if (ib_srq_has_cq(srq->srq_type)) {
628 			srq->ext.cq   = srq_init_attr->ext.cq;
629 			atomic_inc(&srq->ext.cq->usecnt);
630 		}
631 		if (srq->srq_type == IB_SRQT_XRC) {
632 			srq->ext.xrc.xrcd = srq_init_attr->ext.xrc.xrcd;
633 			atomic_inc(&srq->ext.xrc.xrcd->usecnt);
634 		}
635 		atomic_inc(&pd->usecnt);
636 		atomic_set(&srq->usecnt, 0);
637 	}
638 
639 	return srq;
640 }
641 EXPORT_SYMBOL(ib_create_srq);
642 
643 int ib_modify_srq(struct ib_srq *srq,
644 		  struct ib_srq_attr *srq_attr,
645 		  enum ib_srq_attr_mask srq_attr_mask)
646 {
647 	return srq->device->modify_srq ?
648 		srq->device->modify_srq(srq, srq_attr, srq_attr_mask, NULL) :
649 		-ENOSYS;
650 }
651 EXPORT_SYMBOL(ib_modify_srq);
652 
653 int ib_query_srq(struct ib_srq *srq,
654 		 struct ib_srq_attr *srq_attr)
655 {
656 	return srq->device->query_srq ?
657 		srq->device->query_srq(srq, srq_attr) : -ENOSYS;
658 }
659 EXPORT_SYMBOL(ib_query_srq);
660 
661 int ib_destroy_srq(struct ib_srq *srq)
662 {
663 	struct ib_pd *pd;
664 	enum ib_srq_type srq_type;
665 	struct ib_xrcd *uninitialized_var(xrcd);
666 	struct ib_cq *uninitialized_var(cq);
667 	int ret;
668 
669 	if (atomic_read(&srq->usecnt))
670 		return -EBUSY;
671 
672 	pd = srq->pd;
673 	srq_type = srq->srq_type;
674 	if (ib_srq_has_cq(srq_type))
675 		cq = srq->ext.cq;
676 	if (srq_type == IB_SRQT_XRC)
677 		xrcd = srq->ext.xrc.xrcd;
678 
679 	ret = srq->device->destroy_srq(srq);
680 	if (!ret) {
681 		atomic_dec(&pd->usecnt);
682 		if (srq_type == IB_SRQT_XRC)
683 			atomic_dec(&xrcd->usecnt);
684 		if (ib_srq_has_cq(srq_type))
685 			atomic_dec(&cq->usecnt);
686 	}
687 
688 	return ret;
689 }
690 EXPORT_SYMBOL(ib_destroy_srq);
691 
692 /* Queue pairs */
693 
694 static void __ib_shared_qp_event_handler(struct ib_event *event, void *context)
695 {
696 	struct ib_qp *qp = context;
697 	unsigned long flags;
698 
699 	spin_lock_irqsave(&qp->device->event_handler_lock, flags);
700 	list_for_each_entry(event->element.qp, &qp->open_list, open_list)
701 		if (event->element.qp->event_handler)
702 			event->element.qp->event_handler(event, event->element.qp->qp_context);
703 	spin_unlock_irqrestore(&qp->device->event_handler_lock, flags);
704 }
705 
706 static void __ib_insert_xrcd_qp(struct ib_xrcd *xrcd, struct ib_qp *qp)
707 {
708 	mutex_lock(&xrcd->tgt_qp_mutex);
709 	list_add(&qp->xrcd_list, &xrcd->tgt_qp_list);
710 	mutex_unlock(&xrcd->tgt_qp_mutex);
711 }
712 
713 static struct ib_qp *__ib_open_qp(struct ib_qp *real_qp,
714 				  void (*event_handler)(struct ib_event *, void *),
715 				  void *qp_context)
716 {
717 	struct ib_qp *qp;
718 	unsigned long flags;
719 	int err;
720 
721 	qp = kzalloc(sizeof *qp, GFP_KERNEL);
722 	if (!qp)
723 		return ERR_PTR(-ENOMEM);
724 
725 	qp->real_qp = real_qp;
726 	err = ib_open_shared_qp_security(qp, real_qp->device);
727 	if (err) {
728 		kfree(qp);
729 		return ERR_PTR(err);
730 	}
731 
732 	qp->real_qp = real_qp;
733 	atomic_inc(&real_qp->usecnt);
734 	qp->device = real_qp->device;
735 	qp->event_handler = event_handler;
736 	qp->qp_context = qp_context;
737 	qp->qp_num = real_qp->qp_num;
738 	qp->qp_type = real_qp->qp_type;
739 
740 	spin_lock_irqsave(&real_qp->device->event_handler_lock, flags);
741 	list_add(&qp->open_list, &real_qp->open_list);
742 	spin_unlock_irqrestore(&real_qp->device->event_handler_lock, flags);
743 
744 	return qp;
745 }
746 
747 struct ib_qp *ib_open_qp(struct ib_xrcd *xrcd,
748 			 struct ib_qp_open_attr *qp_open_attr)
749 {
750 	struct ib_qp *qp, *real_qp;
751 
752 	if (qp_open_attr->qp_type != IB_QPT_XRC_TGT)
753 		return ERR_PTR(-EINVAL);
754 
755 	qp = ERR_PTR(-EINVAL);
756 	mutex_lock(&xrcd->tgt_qp_mutex);
757 	list_for_each_entry(real_qp, &xrcd->tgt_qp_list, xrcd_list) {
758 		if (real_qp->qp_num == qp_open_attr->qp_num) {
759 			qp = __ib_open_qp(real_qp, qp_open_attr->event_handler,
760 					  qp_open_attr->qp_context);
761 			break;
762 		}
763 	}
764 	mutex_unlock(&xrcd->tgt_qp_mutex);
765 	return qp;
766 }
767 EXPORT_SYMBOL(ib_open_qp);
768 
769 static struct ib_qp *ib_create_xrc_qp(struct ib_qp *qp,
770 		struct ib_qp_init_attr *qp_init_attr)
771 {
772 	struct ib_qp *real_qp = qp;
773 
774 	qp->event_handler = __ib_shared_qp_event_handler;
775 	qp->qp_context = qp;
776 	qp->pd = NULL;
777 	qp->send_cq = qp->recv_cq = NULL;
778 	qp->srq = NULL;
779 	qp->xrcd = qp_init_attr->xrcd;
780 	atomic_inc(&qp_init_attr->xrcd->usecnt);
781 	INIT_LIST_HEAD(&qp->open_list);
782 
783 	qp = __ib_open_qp(real_qp, qp_init_attr->event_handler,
784 			  qp_init_attr->qp_context);
785 	if (!IS_ERR(qp))
786 		__ib_insert_xrcd_qp(qp_init_attr->xrcd, real_qp);
787 	else
788 		real_qp->device->destroy_qp(real_qp);
789 	return qp;
790 }
791 
792 struct ib_qp *ib_create_qp(struct ib_pd *pd,
793 			   struct ib_qp_init_attr *qp_init_attr)
794 {
795 	struct ib_device *device = pd ? pd->device : qp_init_attr->xrcd->device;
796 	struct ib_qp *qp;
797 	int ret;
798 
799 	if (qp_init_attr->rwq_ind_tbl &&
800 	    (qp_init_attr->recv_cq ||
801 	    qp_init_attr->srq || qp_init_attr->cap.max_recv_wr ||
802 	    qp_init_attr->cap.max_recv_sge))
803 		return ERR_PTR(-EINVAL);
804 
805 	/*
806 	 * If the callers is using the RDMA API calculate the resources
807 	 * needed for the RDMA READ/WRITE operations.
808 	 *
809 	 * Note that these callers need to pass in a port number.
810 	 */
811 	if (qp_init_attr->cap.max_rdma_ctxs)
812 		rdma_rw_init_qp(device, qp_init_attr);
813 
814 	qp = device->create_qp(pd, qp_init_attr, NULL);
815 	if (IS_ERR(qp))
816 		return qp;
817 
818 	ret = ib_create_qp_security(qp, device);
819 	if (ret) {
820 		ib_destroy_qp(qp);
821 		return ERR_PTR(ret);
822 	}
823 
824 	qp->device     = device;
825 	qp->real_qp    = qp;
826 	qp->uobject    = NULL;
827 	qp->qp_type    = qp_init_attr->qp_type;
828 	qp->rwq_ind_tbl = qp_init_attr->rwq_ind_tbl;
829 
830 	atomic_set(&qp->usecnt, 0);
831 	qp->mrs_used = 0;
832 	spin_lock_init(&qp->mr_lock);
833 	INIT_LIST_HEAD(&qp->rdma_mrs);
834 	INIT_LIST_HEAD(&qp->sig_mrs);
835 	qp->port = 0;
836 
837 	if (qp_init_attr->qp_type == IB_QPT_XRC_TGT)
838 		return ib_create_xrc_qp(qp, qp_init_attr);
839 
840 	qp->event_handler = qp_init_attr->event_handler;
841 	qp->qp_context = qp_init_attr->qp_context;
842 	if (qp_init_attr->qp_type == IB_QPT_XRC_INI) {
843 		qp->recv_cq = NULL;
844 		qp->srq = NULL;
845 	} else {
846 		qp->recv_cq = qp_init_attr->recv_cq;
847 		if (qp_init_attr->recv_cq)
848 			atomic_inc(&qp_init_attr->recv_cq->usecnt);
849 		qp->srq = qp_init_attr->srq;
850 		if (qp->srq)
851 			atomic_inc(&qp_init_attr->srq->usecnt);
852 	}
853 
854 	qp->pd	    = pd;
855 	qp->send_cq = qp_init_attr->send_cq;
856 	qp->xrcd    = NULL;
857 
858 	atomic_inc(&pd->usecnt);
859 	if (qp_init_attr->send_cq)
860 		atomic_inc(&qp_init_attr->send_cq->usecnt);
861 	if (qp_init_attr->rwq_ind_tbl)
862 		atomic_inc(&qp->rwq_ind_tbl->usecnt);
863 
864 	if (qp_init_attr->cap.max_rdma_ctxs) {
865 		ret = rdma_rw_init_mrs(qp, qp_init_attr);
866 		if (ret) {
867 			pr_err("failed to init MR pool ret= %d\n", ret);
868 			ib_destroy_qp(qp);
869 			return ERR_PTR(ret);
870 		}
871 	}
872 
873 	/*
874 	 * Note: all hw drivers guarantee that max_send_sge is lower than
875 	 * the device RDMA WRITE SGE limit but not all hw drivers ensure that
876 	 * max_send_sge <= max_sge_rd.
877 	 */
878 	qp->max_write_sge = qp_init_attr->cap.max_send_sge;
879 	qp->max_read_sge = min_t(u32, qp_init_attr->cap.max_send_sge,
880 				 device->attrs.max_sge_rd);
881 
882 	return qp;
883 }
884 EXPORT_SYMBOL(ib_create_qp);
885 
886 static const struct {
887 	int			valid;
888 	enum ib_qp_attr_mask	req_param[IB_QPT_MAX];
889 	enum ib_qp_attr_mask	opt_param[IB_QPT_MAX];
890 } qp_state_table[IB_QPS_ERR + 1][IB_QPS_ERR + 1] = {
891 	[IB_QPS_RESET] = {
892 		[IB_QPS_RESET] = { .valid = 1 },
893 		[IB_QPS_INIT]  = {
894 			.valid = 1,
895 			.req_param = {
896 				[IB_QPT_UD]  = (IB_QP_PKEY_INDEX		|
897 						IB_QP_PORT			|
898 						IB_QP_QKEY),
899 				[IB_QPT_RAW_PACKET] = IB_QP_PORT,
900 				[IB_QPT_UC]  = (IB_QP_PKEY_INDEX		|
901 						IB_QP_PORT			|
902 						IB_QP_ACCESS_FLAGS),
903 				[IB_QPT_RC]  = (IB_QP_PKEY_INDEX		|
904 						IB_QP_PORT			|
905 						IB_QP_ACCESS_FLAGS),
906 				[IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX		|
907 						IB_QP_PORT			|
908 						IB_QP_ACCESS_FLAGS),
909 				[IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX		|
910 						IB_QP_PORT			|
911 						IB_QP_ACCESS_FLAGS),
912 				[IB_QPT_SMI] = (IB_QP_PKEY_INDEX		|
913 						IB_QP_QKEY),
914 				[IB_QPT_GSI] = (IB_QP_PKEY_INDEX		|
915 						IB_QP_QKEY),
916 			}
917 		},
918 	},
919 	[IB_QPS_INIT]  = {
920 		[IB_QPS_RESET] = { .valid = 1 },
921 		[IB_QPS_ERR] =   { .valid = 1 },
922 		[IB_QPS_INIT]  = {
923 			.valid = 1,
924 			.opt_param = {
925 				[IB_QPT_UD]  = (IB_QP_PKEY_INDEX		|
926 						IB_QP_PORT			|
927 						IB_QP_QKEY),
928 				[IB_QPT_UC]  = (IB_QP_PKEY_INDEX		|
929 						IB_QP_PORT			|
930 						IB_QP_ACCESS_FLAGS),
931 				[IB_QPT_RC]  = (IB_QP_PKEY_INDEX		|
932 						IB_QP_PORT			|
933 						IB_QP_ACCESS_FLAGS),
934 				[IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX		|
935 						IB_QP_PORT			|
936 						IB_QP_ACCESS_FLAGS),
937 				[IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX		|
938 						IB_QP_PORT			|
939 						IB_QP_ACCESS_FLAGS),
940 				[IB_QPT_SMI] = (IB_QP_PKEY_INDEX		|
941 						IB_QP_QKEY),
942 				[IB_QPT_GSI] = (IB_QP_PKEY_INDEX		|
943 						IB_QP_QKEY),
944 			}
945 		},
946 		[IB_QPS_RTR]   = {
947 			.valid = 1,
948 			.req_param = {
949 				[IB_QPT_UC]  = (IB_QP_AV			|
950 						IB_QP_PATH_MTU			|
951 						IB_QP_DEST_QPN			|
952 						IB_QP_RQ_PSN),
953 				[IB_QPT_RC]  = (IB_QP_AV			|
954 						IB_QP_PATH_MTU			|
955 						IB_QP_DEST_QPN			|
956 						IB_QP_RQ_PSN			|
957 						IB_QP_MAX_DEST_RD_ATOMIC	|
958 						IB_QP_MIN_RNR_TIMER),
959 				[IB_QPT_XRC_INI] = (IB_QP_AV			|
960 						IB_QP_PATH_MTU			|
961 						IB_QP_DEST_QPN			|
962 						IB_QP_RQ_PSN),
963 				[IB_QPT_XRC_TGT] = (IB_QP_AV			|
964 						IB_QP_PATH_MTU			|
965 						IB_QP_DEST_QPN			|
966 						IB_QP_RQ_PSN			|
967 						IB_QP_MAX_DEST_RD_ATOMIC	|
968 						IB_QP_MIN_RNR_TIMER),
969 			},
970 			.opt_param = {
971 				 [IB_QPT_UD]  = (IB_QP_PKEY_INDEX		|
972 						 IB_QP_QKEY),
973 				 [IB_QPT_UC]  = (IB_QP_ALT_PATH			|
974 						 IB_QP_ACCESS_FLAGS		|
975 						 IB_QP_PKEY_INDEX),
976 				 [IB_QPT_RC]  = (IB_QP_ALT_PATH			|
977 						 IB_QP_ACCESS_FLAGS		|
978 						 IB_QP_PKEY_INDEX),
979 				 [IB_QPT_XRC_INI] = (IB_QP_ALT_PATH		|
980 						 IB_QP_ACCESS_FLAGS		|
981 						 IB_QP_PKEY_INDEX),
982 				 [IB_QPT_XRC_TGT] = (IB_QP_ALT_PATH		|
983 						 IB_QP_ACCESS_FLAGS		|
984 						 IB_QP_PKEY_INDEX),
985 				 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX		|
986 						 IB_QP_QKEY),
987 				 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX		|
988 						 IB_QP_QKEY),
989 			 },
990 		},
991 	},
992 	[IB_QPS_RTR]   = {
993 		[IB_QPS_RESET] = { .valid = 1 },
994 		[IB_QPS_ERR] =   { .valid = 1 },
995 		[IB_QPS_RTS]   = {
996 			.valid = 1,
997 			.req_param = {
998 				[IB_QPT_UD]  = IB_QP_SQ_PSN,
999 				[IB_QPT_UC]  = IB_QP_SQ_PSN,
1000 				[IB_QPT_RC]  = (IB_QP_TIMEOUT			|
1001 						IB_QP_RETRY_CNT			|
1002 						IB_QP_RNR_RETRY			|
1003 						IB_QP_SQ_PSN			|
1004 						IB_QP_MAX_QP_RD_ATOMIC),
1005 				[IB_QPT_XRC_INI] = (IB_QP_TIMEOUT		|
1006 						IB_QP_RETRY_CNT			|
1007 						IB_QP_RNR_RETRY			|
1008 						IB_QP_SQ_PSN			|
1009 						IB_QP_MAX_QP_RD_ATOMIC),
1010 				[IB_QPT_XRC_TGT] = (IB_QP_TIMEOUT		|
1011 						IB_QP_SQ_PSN),
1012 				[IB_QPT_SMI] = IB_QP_SQ_PSN,
1013 				[IB_QPT_GSI] = IB_QP_SQ_PSN,
1014 			},
1015 			.opt_param = {
1016 				 [IB_QPT_UD]  = (IB_QP_CUR_STATE		|
1017 						 IB_QP_QKEY),
1018 				 [IB_QPT_UC]  = (IB_QP_CUR_STATE		|
1019 						 IB_QP_ALT_PATH			|
1020 						 IB_QP_ACCESS_FLAGS		|
1021 						 IB_QP_PATH_MIG_STATE),
1022 				 [IB_QPT_RC]  = (IB_QP_CUR_STATE		|
1023 						 IB_QP_ALT_PATH			|
1024 						 IB_QP_ACCESS_FLAGS		|
1025 						 IB_QP_MIN_RNR_TIMER		|
1026 						 IB_QP_PATH_MIG_STATE),
1027 				 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE		|
1028 						 IB_QP_ALT_PATH			|
1029 						 IB_QP_ACCESS_FLAGS		|
1030 						 IB_QP_PATH_MIG_STATE),
1031 				 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE		|
1032 						 IB_QP_ALT_PATH			|
1033 						 IB_QP_ACCESS_FLAGS		|
1034 						 IB_QP_MIN_RNR_TIMER		|
1035 						 IB_QP_PATH_MIG_STATE),
1036 				 [IB_QPT_SMI] = (IB_QP_CUR_STATE		|
1037 						 IB_QP_QKEY),
1038 				 [IB_QPT_GSI] = (IB_QP_CUR_STATE		|
1039 						 IB_QP_QKEY),
1040 				 [IB_QPT_RAW_PACKET] = IB_QP_RATE_LIMIT,
1041 			 }
1042 		}
1043 	},
1044 	[IB_QPS_RTS]   = {
1045 		[IB_QPS_RESET] = { .valid = 1 },
1046 		[IB_QPS_ERR] =   { .valid = 1 },
1047 		[IB_QPS_RTS]   = {
1048 			.valid = 1,
1049 			.opt_param = {
1050 				[IB_QPT_UD]  = (IB_QP_CUR_STATE			|
1051 						IB_QP_QKEY),
1052 				[IB_QPT_UC]  = (IB_QP_CUR_STATE			|
1053 						IB_QP_ACCESS_FLAGS		|
1054 						IB_QP_ALT_PATH			|
1055 						IB_QP_PATH_MIG_STATE),
1056 				[IB_QPT_RC]  = (IB_QP_CUR_STATE			|
1057 						IB_QP_ACCESS_FLAGS		|
1058 						IB_QP_ALT_PATH			|
1059 						IB_QP_PATH_MIG_STATE		|
1060 						IB_QP_MIN_RNR_TIMER),
1061 				[IB_QPT_XRC_INI] = (IB_QP_CUR_STATE		|
1062 						IB_QP_ACCESS_FLAGS		|
1063 						IB_QP_ALT_PATH			|
1064 						IB_QP_PATH_MIG_STATE),
1065 				[IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE		|
1066 						IB_QP_ACCESS_FLAGS		|
1067 						IB_QP_ALT_PATH			|
1068 						IB_QP_PATH_MIG_STATE		|
1069 						IB_QP_MIN_RNR_TIMER),
1070 				[IB_QPT_SMI] = (IB_QP_CUR_STATE			|
1071 						IB_QP_QKEY),
1072 				[IB_QPT_GSI] = (IB_QP_CUR_STATE			|
1073 						IB_QP_QKEY),
1074 				[IB_QPT_RAW_PACKET] = IB_QP_RATE_LIMIT,
1075 			}
1076 		},
1077 		[IB_QPS_SQD]   = {
1078 			.valid = 1,
1079 			.opt_param = {
1080 				[IB_QPT_UD]  = IB_QP_EN_SQD_ASYNC_NOTIFY,
1081 				[IB_QPT_UC]  = IB_QP_EN_SQD_ASYNC_NOTIFY,
1082 				[IB_QPT_RC]  = IB_QP_EN_SQD_ASYNC_NOTIFY,
1083 				[IB_QPT_XRC_INI] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1084 				[IB_QPT_XRC_TGT] = IB_QP_EN_SQD_ASYNC_NOTIFY, /* ??? */
1085 				[IB_QPT_SMI] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1086 				[IB_QPT_GSI] = IB_QP_EN_SQD_ASYNC_NOTIFY
1087 			}
1088 		},
1089 	},
1090 	[IB_QPS_SQD]   = {
1091 		[IB_QPS_RESET] = { .valid = 1 },
1092 		[IB_QPS_ERR] =   { .valid = 1 },
1093 		[IB_QPS_RTS]   = {
1094 			.valid = 1,
1095 			.opt_param = {
1096 				[IB_QPT_UD]  = (IB_QP_CUR_STATE			|
1097 						IB_QP_QKEY),
1098 				[IB_QPT_UC]  = (IB_QP_CUR_STATE			|
1099 						IB_QP_ALT_PATH			|
1100 						IB_QP_ACCESS_FLAGS		|
1101 						IB_QP_PATH_MIG_STATE),
1102 				[IB_QPT_RC]  = (IB_QP_CUR_STATE			|
1103 						IB_QP_ALT_PATH			|
1104 						IB_QP_ACCESS_FLAGS		|
1105 						IB_QP_MIN_RNR_TIMER		|
1106 						IB_QP_PATH_MIG_STATE),
1107 				[IB_QPT_XRC_INI] = (IB_QP_CUR_STATE		|
1108 						IB_QP_ALT_PATH			|
1109 						IB_QP_ACCESS_FLAGS		|
1110 						IB_QP_PATH_MIG_STATE),
1111 				[IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE		|
1112 						IB_QP_ALT_PATH			|
1113 						IB_QP_ACCESS_FLAGS		|
1114 						IB_QP_MIN_RNR_TIMER		|
1115 						IB_QP_PATH_MIG_STATE),
1116 				[IB_QPT_SMI] = (IB_QP_CUR_STATE			|
1117 						IB_QP_QKEY),
1118 				[IB_QPT_GSI] = (IB_QP_CUR_STATE			|
1119 						IB_QP_QKEY),
1120 			}
1121 		},
1122 		[IB_QPS_SQD]   = {
1123 			.valid = 1,
1124 			.opt_param = {
1125 				[IB_QPT_UD]  = (IB_QP_PKEY_INDEX		|
1126 						IB_QP_QKEY),
1127 				[IB_QPT_UC]  = (IB_QP_AV			|
1128 						IB_QP_ALT_PATH			|
1129 						IB_QP_ACCESS_FLAGS		|
1130 						IB_QP_PKEY_INDEX		|
1131 						IB_QP_PATH_MIG_STATE),
1132 				[IB_QPT_RC]  = (IB_QP_PORT			|
1133 						IB_QP_AV			|
1134 						IB_QP_TIMEOUT			|
1135 						IB_QP_RETRY_CNT			|
1136 						IB_QP_RNR_RETRY			|
1137 						IB_QP_MAX_QP_RD_ATOMIC		|
1138 						IB_QP_MAX_DEST_RD_ATOMIC	|
1139 						IB_QP_ALT_PATH			|
1140 						IB_QP_ACCESS_FLAGS		|
1141 						IB_QP_PKEY_INDEX		|
1142 						IB_QP_MIN_RNR_TIMER		|
1143 						IB_QP_PATH_MIG_STATE),
1144 				[IB_QPT_XRC_INI] = (IB_QP_PORT			|
1145 						IB_QP_AV			|
1146 						IB_QP_TIMEOUT			|
1147 						IB_QP_RETRY_CNT			|
1148 						IB_QP_RNR_RETRY			|
1149 						IB_QP_MAX_QP_RD_ATOMIC		|
1150 						IB_QP_ALT_PATH			|
1151 						IB_QP_ACCESS_FLAGS		|
1152 						IB_QP_PKEY_INDEX		|
1153 						IB_QP_PATH_MIG_STATE),
1154 				[IB_QPT_XRC_TGT] = (IB_QP_PORT			|
1155 						IB_QP_AV			|
1156 						IB_QP_TIMEOUT			|
1157 						IB_QP_MAX_DEST_RD_ATOMIC	|
1158 						IB_QP_ALT_PATH			|
1159 						IB_QP_ACCESS_FLAGS		|
1160 						IB_QP_PKEY_INDEX		|
1161 						IB_QP_MIN_RNR_TIMER		|
1162 						IB_QP_PATH_MIG_STATE),
1163 				[IB_QPT_SMI] = (IB_QP_PKEY_INDEX		|
1164 						IB_QP_QKEY),
1165 				[IB_QPT_GSI] = (IB_QP_PKEY_INDEX		|
1166 						IB_QP_QKEY),
1167 			}
1168 		}
1169 	},
1170 	[IB_QPS_SQE]   = {
1171 		[IB_QPS_RESET] = { .valid = 1 },
1172 		[IB_QPS_ERR] =   { .valid = 1 },
1173 		[IB_QPS_RTS]   = {
1174 			.valid = 1,
1175 			.opt_param = {
1176 				[IB_QPT_UD]  = (IB_QP_CUR_STATE			|
1177 						IB_QP_QKEY),
1178 				[IB_QPT_UC]  = (IB_QP_CUR_STATE			|
1179 						IB_QP_ACCESS_FLAGS),
1180 				[IB_QPT_SMI] = (IB_QP_CUR_STATE			|
1181 						IB_QP_QKEY),
1182 				[IB_QPT_GSI] = (IB_QP_CUR_STATE			|
1183 						IB_QP_QKEY),
1184 			}
1185 		}
1186 	},
1187 	[IB_QPS_ERR] = {
1188 		[IB_QPS_RESET] = { .valid = 1 },
1189 		[IB_QPS_ERR] =   { .valid = 1 }
1190 	}
1191 };
1192 
1193 int ib_modify_qp_is_ok(enum ib_qp_state cur_state, enum ib_qp_state next_state,
1194 		       enum ib_qp_type type, enum ib_qp_attr_mask mask,
1195 		       enum rdma_link_layer ll)
1196 {
1197 	enum ib_qp_attr_mask req_param, opt_param;
1198 
1199 	if (cur_state  < 0 || cur_state  > IB_QPS_ERR ||
1200 	    next_state < 0 || next_state > IB_QPS_ERR)
1201 		return 0;
1202 
1203 	if (mask & IB_QP_CUR_STATE  &&
1204 	    cur_state != IB_QPS_RTR && cur_state != IB_QPS_RTS &&
1205 	    cur_state != IB_QPS_SQD && cur_state != IB_QPS_SQE)
1206 		return 0;
1207 
1208 	if (!qp_state_table[cur_state][next_state].valid)
1209 		return 0;
1210 
1211 	req_param = qp_state_table[cur_state][next_state].req_param[type];
1212 	opt_param = qp_state_table[cur_state][next_state].opt_param[type];
1213 
1214 	if ((mask & req_param) != req_param)
1215 		return 0;
1216 
1217 	if (mask & ~(req_param | opt_param | IB_QP_STATE))
1218 		return 0;
1219 
1220 	return 1;
1221 }
1222 EXPORT_SYMBOL(ib_modify_qp_is_ok);
1223 
1224 int ib_resolve_eth_dmac(struct ib_device *device,
1225 			struct rdma_ah_attr *ah_attr)
1226 {
1227 	int           ret = 0;
1228 	struct ib_global_route *grh;
1229 
1230 	if (!rdma_is_port_valid(device, rdma_ah_get_port_num(ah_attr)))
1231 		return -EINVAL;
1232 
1233 	if (ah_attr->type != RDMA_AH_ATTR_TYPE_ROCE)
1234 		return 0;
1235 
1236 	grh = rdma_ah_retrieve_grh(ah_attr);
1237 
1238 	if (rdma_link_local_addr((struct in6_addr *)grh->dgid.raw)) {
1239 		rdma_get_ll_mac((struct in6_addr *)grh->dgid.raw,
1240 				ah_attr->roce.dmac);
1241 		return 0;
1242 	}
1243 	if (rdma_is_multicast_addr((struct in6_addr *)ah_attr->grh.dgid.raw)) {
1244 		if (ipv6_addr_v4mapped((struct in6_addr *)ah_attr->grh.dgid.raw)) {
1245 			__be32 addr = 0;
1246 
1247 			memcpy(&addr, ah_attr->grh.dgid.raw + 12, 4);
1248 			ip_eth_mc_map(addr, (char *)ah_attr->roce.dmac);
1249 		} else {
1250 			ipv6_eth_mc_map((struct in6_addr *)ah_attr->grh.dgid.raw,
1251 					(char *)ah_attr->roce.dmac);
1252 		}
1253 	} else {
1254 		union ib_gid		sgid;
1255 		struct ib_gid_attr	sgid_attr;
1256 		int			ifindex;
1257 		int			hop_limit;
1258 
1259 		ret = ib_query_gid(device,
1260 				   rdma_ah_get_port_num(ah_attr),
1261 				   grh->sgid_index,
1262 				   &sgid, &sgid_attr);
1263 
1264 		if (ret || !sgid_attr.ndev) {
1265 			if (!ret)
1266 				ret = -ENXIO;
1267 			goto out;
1268 		}
1269 
1270 		ifindex = sgid_attr.ndev->ifindex;
1271 
1272 		ret =
1273 		rdma_addr_find_l2_eth_by_grh(&sgid, &grh->dgid,
1274 					     ah_attr->roce.dmac,
1275 					     NULL, &ifindex, &hop_limit);
1276 
1277 		dev_put(sgid_attr.ndev);
1278 
1279 		grh->hop_limit = hop_limit;
1280 	}
1281 out:
1282 	return ret;
1283 }
1284 EXPORT_SYMBOL(ib_resolve_eth_dmac);
1285 
1286 /**
1287  * ib_modify_qp_with_udata - Modifies the attributes for the specified QP.
1288  * @qp: The QP to modify.
1289  * @attr: On input, specifies the QP attributes to modify.  On output,
1290  *   the current values of selected QP attributes are returned.
1291  * @attr_mask: A bit-mask used to specify which attributes of the QP
1292  *   are being modified.
1293  * @udata: pointer to user's input output buffer information
1294  *   are being modified.
1295  * It returns 0 on success and returns appropriate error code on error.
1296  */
1297 int ib_modify_qp_with_udata(struct ib_qp *qp, struct ib_qp_attr *attr,
1298 			    int attr_mask, struct ib_udata *udata)
1299 {
1300 	int ret;
1301 
1302 	if (attr_mask & IB_QP_AV) {
1303 		ret = ib_resolve_eth_dmac(qp->device, &attr->ah_attr);
1304 		if (ret)
1305 			return ret;
1306 	}
1307 	ret = ib_security_modify_qp(qp, attr, attr_mask, udata);
1308 	if (!ret && (attr_mask & IB_QP_PORT))
1309 		qp->port = attr->port_num;
1310 
1311 	return ret;
1312 }
1313 EXPORT_SYMBOL(ib_modify_qp_with_udata);
1314 
1315 int ib_get_eth_speed(struct ib_device *dev, u8 port_num, u8 *speed, u8 *width)
1316 {
1317 	int rc;
1318 	u32 netdev_speed;
1319 	struct net_device *netdev;
1320 	struct ethtool_link_ksettings lksettings;
1321 
1322 	if (rdma_port_get_link_layer(dev, port_num) != IB_LINK_LAYER_ETHERNET)
1323 		return -EINVAL;
1324 
1325 	if (!dev->get_netdev)
1326 		return -EOPNOTSUPP;
1327 
1328 	netdev = dev->get_netdev(dev, port_num);
1329 	if (!netdev)
1330 		return -ENODEV;
1331 
1332 	rtnl_lock();
1333 	rc = __ethtool_get_link_ksettings(netdev, &lksettings);
1334 	rtnl_unlock();
1335 
1336 	dev_put(netdev);
1337 
1338 	if (!rc) {
1339 		netdev_speed = lksettings.base.speed;
1340 	} else {
1341 		netdev_speed = SPEED_1000;
1342 		pr_warn("%s speed is unknown, defaulting to %d\n", netdev->name,
1343 			netdev_speed);
1344 	}
1345 
1346 	if (netdev_speed <= SPEED_1000) {
1347 		*width = IB_WIDTH_1X;
1348 		*speed = IB_SPEED_SDR;
1349 	} else if (netdev_speed <= SPEED_10000) {
1350 		*width = IB_WIDTH_1X;
1351 		*speed = IB_SPEED_FDR10;
1352 	} else if (netdev_speed <= SPEED_20000) {
1353 		*width = IB_WIDTH_4X;
1354 		*speed = IB_SPEED_DDR;
1355 	} else if (netdev_speed <= SPEED_25000) {
1356 		*width = IB_WIDTH_1X;
1357 		*speed = IB_SPEED_EDR;
1358 	} else if (netdev_speed <= SPEED_40000) {
1359 		*width = IB_WIDTH_4X;
1360 		*speed = IB_SPEED_FDR10;
1361 	} else {
1362 		*width = IB_WIDTH_4X;
1363 		*speed = IB_SPEED_EDR;
1364 	}
1365 
1366 	return 0;
1367 }
1368 EXPORT_SYMBOL(ib_get_eth_speed);
1369 
1370 int ib_modify_qp(struct ib_qp *qp,
1371 		 struct ib_qp_attr *qp_attr,
1372 		 int qp_attr_mask)
1373 {
1374 	return ib_modify_qp_with_udata(qp, qp_attr, qp_attr_mask, NULL);
1375 }
1376 EXPORT_SYMBOL(ib_modify_qp);
1377 
1378 int ib_query_qp(struct ib_qp *qp,
1379 		struct ib_qp_attr *qp_attr,
1380 		int qp_attr_mask,
1381 		struct ib_qp_init_attr *qp_init_attr)
1382 {
1383 	return qp->device->query_qp ?
1384 		qp->device->query_qp(qp->real_qp, qp_attr, qp_attr_mask, qp_init_attr) :
1385 		-ENOSYS;
1386 }
1387 EXPORT_SYMBOL(ib_query_qp);
1388 
1389 int ib_close_qp(struct ib_qp *qp)
1390 {
1391 	struct ib_qp *real_qp;
1392 	unsigned long flags;
1393 
1394 	real_qp = qp->real_qp;
1395 	if (real_qp == qp)
1396 		return -EINVAL;
1397 
1398 	spin_lock_irqsave(&real_qp->device->event_handler_lock, flags);
1399 	list_del(&qp->open_list);
1400 	spin_unlock_irqrestore(&real_qp->device->event_handler_lock, flags);
1401 
1402 	atomic_dec(&real_qp->usecnt);
1403 	ib_close_shared_qp_security(qp->qp_sec);
1404 	kfree(qp);
1405 
1406 	return 0;
1407 }
1408 EXPORT_SYMBOL(ib_close_qp);
1409 
1410 static int __ib_destroy_shared_qp(struct ib_qp *qp)
1411 {
1412 	struct ib_xrcd *xrcd;
1413 	struct ib_qp *real_qp;
1414 	int ret;
1415 
1416 	real_qp = qp->real_qp;
1417 	xrcd = real_qp->xrcd;
1418 
1419 	mutex_lock(&xrcd->tgt_qp_mutex);
1420 	ib_close_qp(qp);
1421 	if (atomic_read(&real_qp->usecnt) == 0)
1422 		list_del(&real_qp->xrcd_list);
1423 	else
1424 		real_qp = NULL;
1425 	mutex_unlock(&xrcd->tgt_qp_mutex);
1426 
1427 	if (real_qp) {
1428 		ret = ib_destroy_qp(real_qp);
1429 		if (!ret)
1430 			atomic_dec(&xrcd->usecnt);
1431 		else
1432 			__ib_insert_xrcd_qp(xrcd, real_qp);
1433 	}
1434 
1435 	return 0;
1436 }
1437 
1438 int ib_destroy_qp(struct ib_qp *qp)
1439 {
1440 	struct ib_pd *pd;
1441 	struct ib_cq *scq, *rcq;
1442 	struct ib_srq *srq;
1443 	struct ib_rwq_ind_table *ind_tbl;
1444 	struct ib_qp_security *sec;
1445 	int ret;
1446 
1447 	WARN_ON_ONCE(qp->mrs_used > 0);
1448 
1449 	if (atomic_read(&qp->usecnt))
1450 		return -EBUSY;
1451 
1452 	if (qp->real_qp != qp)
1453 		return __ib_destroy_shared_qp(qp);
1454 
1455 	pd   = qp->pd;
1456 	scq  = qp->send_cq;
1457 	rcq  = qp->recv_cq;
1458 	srq  = qp->srq;
1459 	ind_tbl = qp->rwq_ind_tbl;
1460 	sec  = qp->qp_sec;
1461 	if (sec)
1462 		ib_destroy_qp_security_begin(sec);
1463 
1464 	if (!qp->uobject)
1465 		rdma_rw_cleanup_mrs(qp);
1466 
1467 	ret = qp->device->destroy_qp(qp);
1468 	if (!ret) {
1469 		if (pd)
1470 			atomic_dec(&pd->usecnt);
1471 		if (scq)
1472 			atomic_dec(&scq->usecnt);
1473 		if (rcq)
1474 			atomic_dec(&rcq->usecnt);
1475 		if (srq)
1476 			atomic_dec(&srq->usecnt);
1477 		if (ind_tbl)
1478 			atomic_dec(&ind_tbl->usecnt);
1479 		if (sec)
1480 			ib_destroy_qp_security_end(sec);
1481 	} else {
1482 		if (sec)
1483 			ib_destroy_qp_security_abort(sec);
1484 	}
1485 
1486 	return ret;
1487 }
1488 EXPORT_SYMBOL(ib_destroy_qp);
1489 
1490 /* Completion queues */
1491 
1492 struct ib_cq *ib_create_cq(struct ib_device *device,
1493 			   ib_comp_handler comp_handler,
1494 			   void (*event_handler)(struct ib_event *, void *),
1495 			   void *cq_context,
1496 			   const struct ib_cq_init_attr *cq_attr)
1497 {
1498 	struct ib_cq *cq;
1499 
1500 	cq = device->create_cq(device, cq_attr, NULL, NULL);
1501 
1502 	if (!IS_ERR(cq)) {
1503 		cq->device        = device;
1504 		cq->uobject       = NULL;
1505 		cq->comp_handler  = comp_handler;
1506 		cq->event_handler = event_handler;
1507 		cq->cq_context    = cq_context;
1508 		atomic_set(&cq->usecnt, 0);
1509 	}
1510 
1511 	return cq;
1512 }
1513 EXPORT_SYMBOL(ib_create_cq);
1514 
1515 int ib_modify_cq(struct ib_cq *cq, u16 cq_count, u16 cq_period)
1516 {
1517 	return cq->device->modify_cq ?
1518 		cq->device->modify_cq(cq, cq_count, cq_period) : -ENOSYS;
1519 }
1520 EXPORT_SYMBOL(ib_modify_cq);
1521 
1522 int ib_destroy_cq(struct ib_cq *cq)
1523 {
1524 	if (atomic_read(&cq->usecnt))
1525 		return -EBUSY;
1526 
1527 	return cq->device->destroy_cq(cq);
1528 }
1529 EXPORT_SYMBOL(ib_destroy_cq);
1530 
1531 int ib_resize_cq(struct ib_cq *cq, int cqe)
1532 {
1533 	return cq->device->resize_cq ?
1534 		cq->device->resize_cq(cq, cqe, NULL) : -ENOSYS;
1535 }
1536 EXPORT_SYMBOL(ib_resize_cq);
1537 
1538 /* Memory regions */
1539 
1540 int ib_dereg_mr(struct ib_mr *mr)
1541 {
1542 	struct ib_pd *pd = mr->pd;
1543 	int ret;
1544 
1545 	ret = mr->device->dereg_mr(mr);
1546 	if (!ret)
1547 		atomic_dec(&pd->usecnt);
1548 
1549 	return ret;
1550 }
1551 EXPORT_SYMBOL(ib_dereg_mr);
1552 
1553 /**
1554  * ib_alloc_mr() - Allocates a memory region
1555  * @pd:            protection domain associated with the region
1556  * @mr_type:       memory region type
1557  * @max_num_sg:    maximum sg entries available for registration.
1558  *
1559  * Notes:
1560  * Memory registeration page/sg lists must not exceed max_num_sg.
1561  * For mr_type IB_MR_TYPE_MEM_REG, the total length cannot exceed
1562  * max_num_sg * used_page_size.
1563  *
1564  */
1565 struct ib_mr *ib_alloc_mr(struct ib_pd *pd,
1566 			  enum ib_mr_type mr_type,
1567 			  u32 max_num_sg)
1568 {
1569 	struct ib_mr *mr;
1570 
1571 	if (!pd->device->alloc_mr)
1572 		return ERR_PTR(-ENOSYS);
1573 
1574 	mr = pd->device->alloc_mr(pd, mr_type, max_num_sg);
1575 	if (!IS_ERR(mr)) {
1576 		mr->device  = pd->device;
1577 		mr->pd      = pd;
1578 		mr->uobject = NULL;
1579 		atomic_inc(&pd->usecnt);
1580 		mr->need_inval = false;
1581 	}
1582 
1583 	return mr;
1584 }
1585 EXPORT_SYMBOL(ib_alloc_mr);
1586 
1587 /* "Fast" memory regions */
1588 
1589 struct ib_fmr *ib_alloc_fmr(struct ib_pd *pd,
1590 			    int mr_access_flags,
1591 			    struct ib_fmr_attr *fmr_attr)
1592 {
1593 	struct ib_fmr *fmr;
1594 
1595 	if (!pd->device->alloc_fmr)
1596 		return ERR_PTR(-ENOSYS);
1597 
1598 	fmr = pd->device->alloc_fmr(pd, mr_access_flags, fmr_attr);
1599 	if (!IS_ERR(fmr)) {
1600 		fmr->device = pd->device;
1601 		fmr->pd     = pd;
1602 		atomic_inc(&pd->usecnt);
1603 	}
1604 
1605 	return fmr;
1606 }
1607 EXPORT_SYMBOL(ib_alloc_fmr);
1608 
1609 int ib_unmap_fmr(struct list_head *fmr_list)
1610 {
1611 	struct ib_fmr *fmr;
1612 
1613 	if (list_empty(fmr_list))
1614 		return 0;
1615 
1616 	fmr = list_entry(fmr_list->next, struct ib_fmr, list);
1617 	return fmr->device->unmap_fmr(fmr_list);
1618 }
1619 EXPORT_SYMBOL(ib_unmap_fmr);
1620 
1621 int ib_dealloc_fmr(struct ib_fmr *fmr)
1622 {
1623 	struct ib_pd *pd;
1624 	int ret;
1625 
1626 	pd = fmr->pd;
1627 	ret = fmr->device->dealloc_fmr(fmr);
1628 	if (!ret)
1629 		atomic_dec(&pd->usecnt);
1630 
1631 	return ret;
1632 }
1633 EXPORT_SYMBOL(ib_dealloc_fmr);
1634 
1635 /* Multicast groups */
1636 
1637 static bool is_valid_mcast_lid(struct ib_qp *qp, u16 lid)
1638 {
1639 	struct ib_qp_init_attr init_attr = {};
1640 	struct ib_qp_attr attr = {};
1641 	int num_eth_ports = 0;
1642 	int port;
1643 
1644 	/* If QP state >= init, it is assigned to a port and we can check this
1645 	 * port only.
1646 	 */
1647 	if (!ib_query_qp(qp, &attr, IB_QP_STATE | IB_QP_PORT, &init_attr)) {
1648 		if (attr.qp_state >= IB_QPS_INIT) {
1649 			if (rdma_port_get_link_layer(qp->device, attr.port_num) !=
1650 			    IB_LINK_LAYER_INFINIBAND)
1651 				return true;
1652 			goto lid_check;
1653 		}
1654 	}
1655 
1656 	/* Can't get a quick answer, iterate over all ports */
1657 	for (port = 0; port < qp->device->phys_port_cnt; port++)
1658 		if (rdma_port_get_link_layer(qp->device, port) !=
1659 		    IB_LINK_LAYER_INFINIBAND)
1660 			num_eth_ports++;
1661 
1662 	/* If we have at lease one Ethernet port, RoCE annex declares that
1663 	 * multicast LID should be ignored. We can't tell at this step if the
1664 	 * QP belongs to an IB or Ethernet port.
1665 	 */
1666 	if (num_eth_ports)
1667 		return true;
1668 
1669 	/* If all the ports are IB, we can check according to IB spec. */
1670 lid_check:
1671 	return !(lid < be16_to_cpu(IB_MULTICAST_LID_BASE) ||
1672 		 lid == be16_to_cpu(IB_LID_PERMISSIVE));
1673 }
1674 
1675 int ib_attach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid)
1676 {
1677 	int ret;
1678 
1679 	if (!qp->device->attach_mcast)
1680 		return -ENOSYS;
1681 
1682 	if (!rdma_is_multicast_addr((struct in6_addr *)gid->raw) ||
1683 	    qp->qp_type != IB_QPT_UD || !is_valid_mcast_lid(qp, lid))
1684 		return -EINVAL;
1685 
1686 	ret = qp->device->attach_mcast(qp, gid, lid);
1687 	if (!ret)
1688 		atomic_inc(&qp->usecnt);
1689 	return ret;
1690 }
1691 EXPORT_SYMBOL(ib_attach_mcast);
1692 
1693 int ib_detach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid)
1694 {
1695 	int ret;
1696 
1697 	if (!qp->device->detach_mcast)
1698 		return -ENOSYS;
1699 
1700 	if (!rdma_is_multicast_addr((struct in6_addr *)gid->raw) ||
1701 	    qp->qp_type != IB_QPT_UD || !is_valid_mcast_lid(qp, lid))
1702 		return -EINVAL;
1703 
1704 	ret = qp->device->detach_mcast(qp, gid, lid);
1705 	if (!ret)
1706 		atomic_dec(&qp->usecnt);
1707 	return ret;
1708 }
1709 EXPORT_SYMBOL(ib_detach_mcast);
1710 
1711 struct ib_xrcd *ib_alloc_xrcd(struct ib_device *device)
1712 {
1713 	struct ib_xrcd *xrcd;
1714 
1715 	if (!device->alloc_xrcd)
1716 		return ERR_PTR(-ENOSYS);
1717 
1718 	xrcd = device->alloc_xrcd(device, NULL, NULL);
1719 	if (!IS_ERR(xrcd)) {
1720 		xrcd->device = device;
1721 		xrcd->inode = NULL;
1722 		atomic_set(&xrcd->usecnt, 0);
1723 		mutex_init(&xrcd->tgt_qp_mutex);
1724 		INIT_LIST_HEAD(&xrcd->tgt_qp_list);
1725 	}
1726 
1727 	return xrcd;
1728 }
1729 EXPORT_SYMBOL(ib_alloc_xrcd);
1730 
1731 int ib_dealloc_xrcd(struct ib_xrcd *xrcd)
1732 {
1733 	struct ib_qp *qp;
1734 	int ret;
1735 
1736 	if (atomic_read(&xrcd->usecnt))
1737 		return -EBUSY;
1738 
1739 	while (!list_empty(&xrcd->tgt_qp_list)) {
1740 		qp = list_entry(xrcd->tgt_qp_list.next, struct ib_qp, xrcd_list);
1741 		ret = ib_destroy_qp(qp);
1742 		if (ret)
1743 			return ret;
1744 	}
1745 
1746 	return xrcd->device->dealloc_xrcd(xrcd);
1747 }
1748 EXPORT_SYMBOL(ib_dealloc_xrcd);
1749 
1750 /**
1751  * ib_create_wq - Creates a WQ associated with the specified protection
1752  * domain.
1753  * @pd: The protection domain associated with the WQ.
1754  * @wq_init_attr: A list of initial attributes required to create the
1755  * WQ. If WQ creation succeeds, then the attributes are updated to
1756  * the actual capabilities of the created WQ.
1757  *
1758  * wq_init_attr->max_wr and wq_init_attr->max_sge determine
1759  * the requested size of the WQ, and set to the actual values allocated
1760  * on return.
1761  * If ib_create_wq() succeeds, then max_wr and max_sge will always be
1762  * at least as large as the requested values.
1763  */
1764 struct ib_wq *ib_create_wq(struct ib_pd *pd,
1765 			   struct ib_wq_init_attr *wq_attr)
1766 {
1767 	struct ib_wq *wq;
1768 
1769 	if (!pd->device->create_wq)
1770 		return ERR_PTR(-ENOSYS);
1771 
1772 	wq = pd->device->create_wq(pd, wq_attr, NULL);
1773 	if (!IS_ERR(wq)) {
1774 		wq->event_handler = wq_attr->event_handler;
1775 		wq->wq_context = wq_attr->wq_context;
1776 		wq->wq_type = wq_attr->wq_type;
1777 		wq->cq = wq_attr->cq;
1778 		wq->device = pd->device;
1779 		wq->pd = pd;
1780 		wq->uobject = NULL;
1781 		atomic_inc(&pd->usecnt);
1782 		atomic_inc(&wq_attr->cq->usecnt);
1783 		atomic_set(&wq->usecnt, 0);
1784 	}
1785 	return wq;
1786 }
1787 EXPORT_SYMBOL(ib_create_wq);
1788 
1789 /**
1790  * ib_destroy_wq - Destroys the specified WQ.
1791  * @wq: The WQ to destroy.
1792  */
1793 int ib_destroy_wq(struct ib_wq *wq)
1794 {
1795 	int err;
1796 	struct ib_cq *cq = wq->cq;
1797 	struct ib_pd *pd = wq->pd;
1798 
1799 	if (atomic_read(&wq->usecnt))
1800 		return -EBUSY;
1801 
1802 	err = wq->device->destroy_wq(wq);
1803 	if (!err) {
1804 		atomic_dec(&pd->usecnt);
1805 		atomic_dec(&cq->usecnt);
1806 	}
1807 	return err;
1808 }
1809 EXPORT_SYMBOL(ib_destroy_wq);
1810 
1811 /**
1812  * ib_modify_wq - Modifies the specified WQ.
1813  * @wq: The WQ to modify.
1814  * @wq_attr: On input, specifies the WQ attributes to modify.
1815  * @wq_attr_mask: A bit-mask used to specify which attributes of the WQ
1816  *   are being modified.
1817  * On output, the current values of selected WQ attributes are returned.
1818  */
1819 int ib_modify_wq(struct ib_wq *wq, struct ib_wq_attr *wq_attr,
1820 		 u32 wq_attr_mask)
1821 {
1822 	int err;
1823 
1824 	if (!wq->device->modify_wq)
1825 		return -ENOSYS;
1826 
1827 	err = wq->device->modify_wq(wq, wq_attr, wq_attr_mask, NULL);
1828 	return err;
1829 }
1830 EXPORT_SYMBOL(ib_modify_wq);
1831 
1832 /*
1833  * ib_create_rwq_ind_table - Creates a RQ Indirection Table.
1834  * @device: The device on which to create the rwq indirection table.
1835  * @ib_rwq_ind_table_init_attr: A list of initial attributes required to
1836  * create the Indirection Table.
1837  *
1838  * Note: The life time of ib_rwq_ind_table_init_attr->ind_tbl is not less
1839  *	than the created ib_rwq_ind_table object and the caller is responsible
1840  *	for its memory allocation/free.
1841  */
1842 struct ib_rwq_ind_table *ib_create_rwq_ind_table(struct ib_device *device,
1843 						 struct ib_rwq_ind_table_init_attr *init_attr)
1844 {
1845 	struct ib_rwq_ind_table *rwq_ind_table;
1846 	int i;
1847 	u32 table_size;
1848 
1849 	if (!device->create_rwq_ind_table)
1850 		return ERR_PTR(-ENOSYS);
1851 
1852 	table_size = (1 << init_attr->log_ind_tbl_size);
1853 	rwq_ind_table = device->create_rwq_ind_table(device,
1854 				init_attr, NULL);
1855 	if (IS_ERR(rwq_ind_table))
1856 		return rwq_ind_table;
1857 
1858 	rwq_ind_table->ind_tbl = init_attr->ind_tbl;
1859 	rwq_ind_table->log_ind_tbl_size = init_attr->log_ind_tbl_size;
1860 	rwq_ind_table->device = device;
1861 	rwq_ind_table->uobject = NULL;
1862 	atomic_set(&rwq_ind_table->usecnt, 0);
1863 
1864 	for (i = 0; i < table_size; i++)
1865 		atomic_inc(&rwq_ind_table->ind_tbl[i]->usecnt);
1866 
1867 	return rwq_ind_table;
1868 }
1869 EXPORT_SYMBOL(ib_create_rwq_ind_table);
1870 
1871 /*
1872  * ib_destroy_rwq_ind_table - Destroys the specified Indirection Table.
1873  * @wq_ind_table: The Indirection Table to destroy.
1874 */
1875 int ib_destroy_rwq_ind_table(struct ib_rwq_ind_table *rwq_ind_table)
1876 {
1877 	int err, i;
1878 	u32 table_size = (1 << rwq_ind_table->log_ind_tbl_size);
1879 	struct ib_wq **ind_tbl = rwq_ind_table->ind_tbl;
1880 
1881 	if (atomic_read(&rwq_ind_table->usecnt))
1882 		return -EBUSY;
1883 
1884 	err = rwq_ind_table->device->destroy_rwq_ind_table(rwq_ind_table);
1885 	if (!err) {
1886 		for (i = 0; i < table_size; i++)
1887 			atomic_dec(&ind_tbl[i]->usecnt);
1888 	}
1889 
1890 	return err;
1891 }
1892 EXPORT_SYMBOL(ib_destroy_rwq_ind_table);
1893 
1894 struct ib_flow *ib_create_flow(struct ib_qp *qp,
1895 			       struct ib_flow_attr *flow_attr,
1896 			       int domain)
1897 {
1898 	struct ib_flow *flow_id;
1899 	if (!qp->device->create_flow)
1900 		return ERR_PTR(-ENOSYS);
1901 
1902 	flow_id = qp->device->create_flow(qp, flow_attr, domain);
1903 	if (!IS_ERR(flow_id)) {
1904 		atomic_inc(&qp->usecnt);
1905 		flow_id->qp = qp;
1906 	}
1907 	return flow_id;
1908 }
1909 EXPORT_SYMBOL(ib_create_flow);
1910 
1911 int ib_destroy_flow(struct ib_flow *flow_id)
1912 {
1913 	int err;
1914 	struct ib_qp *qp = flow_id->qp;
1915 
1916 	err = qp->device->destroy_flow(flow_id);
1917 	if (!err)
1918 		atomic_dec(&qp->usecnt);
1919 	return err;
1920 }
1921 EXPORT_SYMBOL(ib_destroy_flow);
1922 
1923 int ib_check_mr_status(struct ib_mr *mr, u32 check_mask,
1924 		       struct ib_mr_status *mr_status)
1925 {
1926 	return mr->device->check_mr_status ?
1927 		mr->device->check_mr_status(mr, check_mask, mr_status) : -ENOSYS;
1928 }
1929 EXPORT_SYMBOL(ib_check_mr_status);
1930 
1931 int ib_set_vf_link_state(struct ib_device *device, int vf, u8 port,
1932 			 int state)
1933 {
1934 	if (!device->set_vf_link_state)
1935 		return -ENOSYS;
1936 
1937 	return device->set_vf_link_state(device, vf, port, state);
1938 }
1939 EXPORT_SYMBOL(ib_set_vf_link_state);
1940 
1941 int ib_get_vf_config(struct ib_device *device, int vf, u8 port,
1942 		     struct ifla_vf_info *info)
1943 {
1944 	if (!device->get_vf_config)
1945 		return -ENOSYS;
1946 
1947 	return device->get_vf_config(device, vf, port, info);
1948 }
1949 EXPORT_SYMBOL(ib_get_vf_config);
1950 
1951 int ib_get_vf_stats(struct ib_device *device, int vf, u8 port,
1952 		    struct ifla_vf_stats *stats)
1953 {
1954 	if (!device->get_vf_stats)
1955 		return -ENOSYS;
1956 
1957 	return device->get_vf_stats(device, vf, port, stats);
1958 }
1959 EXPORT_SYMBOL(ib_get_vf_stats);
1960 
1961 int ib_set_vf_guid(struct ib_device *device, int vf, u8 port, u64 guid,
1962 		   int type)
1963 {
1964 	if (!device->set_vf_guid)
1965 		return -ENOSYS;
1966 
1967 	return device->set_vf_guid(device, vf, port, guid, type);
1968 }
1969 EXPORT_SYMBOL(ib_set_vf_guid);
1970 
1971 /**
1972  * ib_map_mr_sg() - Map the largest prefix of a dma mapped SG list
1973  *     and set it the memory region.
1974  * @mr:            memory region
1975  * @sg:            dma mapped scatterlist
1976  * @sg_nents:      number of entries in sg
1977  * @sg_offset:     offset in bytes into sg
1978  * @page_size:     page vector desired page size
1979  *
1980  * Constraints:
1981  * - The first sg element is allowed to have an offset.
1982  * - Each sg element must either be aligned to page_size or virtually
1983  *   contiguous to the previous element. In case an sg element has a
1984  *   non-contiguous offset, the mapping prefix will not include it.
1985  * - The last sg element is allowed to have length less than page_size.
1986  * - If sg_nents total byte length exceeds the mr max_num_sge * page_size
1987  *   then only max_num_sg entries will be mapped.
1988  * - If the MR was allocated with type IB_MR_TYPE_SG_GAPS, none of these
1989  *   constraints holds and the page_size argument is ignored.
1990  *
1991  * Returns the number of sg elements that were mapped to the memory region.
1992  *
1993  * After this completes successfully, the  memory region
1994  * is ready for registration.
1995  */
1996 int ib_map_mr_sg(struct ib_mr *mr, struct scatterlist *sg, int sg_nents,
1997 		 unsigned int *sg_offset, unsigned int page_size)
1998 {
1999 	if (unlikely(!mr->device->map_mr_sg))
2000 		return -ENOSYS;
2001 
2002 	mr->page_size = page_size;
2003 
2004 	return mr->device->map_mr_sg(mr, sg, sg_nents, sg_offset);
2005 }
2006 EXPORT_SYMBOL(ib_map_mr_sg);
2007 
2008 /**
2009  * ib_sg_to_pages() - Convert the largest prefix of a sg list
2010  *     to a page vector
2011  * @mr:            memory region
2012  * @sgl:           dma mapped scatterlist
2013  * @sg_nents:      number of entries in sg
2014  * @sg_offset_p:   IN:  start offset in bytes into sg
2015  *                 OUT: offset in bytes for element n of the sg of the first
2016  *                      byte that has not been processed where n is the return
2017  *                      value of this function.
2018  * @set_page:      driver page assignment function pointer
2019  *
2020  * Core service helper for drivers to convert the largest
2021  * prefix of given sg list to a page vector. The sg list
2022  * prefix converted is the prefix that meet the requirements
2023  * of ib_map_mr_sg.
2024  *
2025  * Returns the number of sg elements that were assigned to
2026  * a page vector.
2027  */
2028 int ib_sg_to_pages(struct ib_mr *mr, struct scatterlist *sgl, int sg_nents,
2029 		unsigned int *sg_offset_p, int (*set_page)(struct ib_mr *, u64))
2030 {
2031 	struct scatterlist *sg;
2032 	u64 last_end_dma_addr = 0;
2033 	unsigned int sg_offset = sg_offset_p ? *sg_offset_p : 0;
2034 	unsigned int last_page_off = 0;
2035 	u64 page_mask = ~((u64)mr->page_size - 1);
2036 	int i, ret;
2037 
2038 	if (unlikely(sg_nents <= 0 || sg_offset > sg_dma_len(&sgl[0])))
2039 		return -EINVAL;
2040 
2041 	mr->iova = sg_dma_address(&sgl[0]) + sg_offset;
2042 	mr->length = 0;
2043 
2044 	for_each_sg(sgl, sg, sg_nents, i) {
2045 		u64 dma_addr = sg_dma_address(sg) + sg_offset;
2046 		u64 prev_addr = dma_addr;
2047 		unsigned int dma_len = sg_dma_len(sg) - sg_offset;
2048 		u64 end_dma_addr = dma_addr + dma_len;
2049 		u64 page_addr = dma_addr & page_mask;
2050 
2051 		/*
2052 		 * For the second and later elements, check whether either the
2053 		 * end of element i-1 or the start of element i is not aligned
2054 		 * on a page boundary.
2055 		 */
2056 		if (i && (last_page_off != 0 || page_addr != dma_addr)) {
2057 			/* Stop mapping if there is a gap. */
2058 			if (last_end_dma_addr != dma_addr)
2059 				break;
2060 
2061 			/*
2062 			 * Coalesce this element with the last. If it is small
2063 			 * enough just update mr->length. Otherwise start
2064 			 * mapping from the next page.
2065 			 */
2066 			goto next_page;
2067 		}
2068 
2069 		do {
2070 			ret = set_page(mr, page_addr);
2071 			if (unlikely(ret < 0)) {
2072 				sg_offset = prev_addr - sg_dma_address(sg);
2073 				mr->length += prev_addr - dma_addr;
2074 				if (sg_offset_p)
2075 					*sg_offset_p = sg_offset;
2076 				return i || sg_offset ? i : ret;
2077 			}
2078 			prev_addr = page_addr;
2079 next_page:
2080 			page_addr += mr->page_size;
2081 		} while (page_addr < end_dma_addr);
2082 
2083 		mr->length += dma_len;
2084 		last_end_dma_addr = end_dma_addr;
2085 		last_page_off = end_dma_addr & ~page_mask;
2086 
2087 		sg_offset = 0;
2088 	}
2089 
2090 	if (sg_offset_p)
2091 		*sg_offset_p = 0;
2092 	return i;
2093 }
2094 EXPORT_SYMBOL(ib_sg_to_pages);
2095 
2096 struct ib_drain_cqe {
2097 	struct ib_cqe cqe;
2098 	struct completion done;
2099 };
2100 
2101 static void ib_drain_qp_done(struct ib_cq *cq, struct ib_wc *wc)
2102 {
2103 	struct ib_drain_cqe *cqe = container_of(wc->wr_cqe, struct ib_drain_cqe,
2104 						cqe);
2105 
2106 	complete(&cqe->done);
2107 }
2108 
2109 /*
2110  * Post a WR and block until its completion is reaped for the SQ.
2111  */
2112 static void __ib_drain_sq(struct ib_qp *qp)
2113 {
2114 	struct ib_cq *cq = qp->send_cq;
2115 	struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
2116 	struct ib_drain_cqe sdrain;
2117 	struct ib_send_wr swr = {}, *bad_swr;
2118 	int ret;
2119 
2120 	swr.wr_cqe = &sdrain.cqe;
2121 	sdrain.cqe.done = ib_drain_qp_done;
2122 	init_completion(&sdrain.done);
2123 
2124 	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
2125 	if (ret) {
2126 		WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
2127 		return;
2128 	}
2129 
2130 	ret = ib_post_send(qp, &swr, &bad_swr);
2131 	if (ret) {
2132 		WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
2133 		return;
2134 	}
2135 
2136 	if (cq->poll_ctx == IB_POLL_DIRECT)
2137 		while (wait_for_completion_timeout(&sdrain.done, HZ / 10) <= 0)
2138 			ib_process_cq_direct(cq, -1);
2139 	else
2140 		wait_for_completion(&sdrain.done);
2141 }
2142 
2143 /*
2144  * Post a WR and block until its completion is reaped for the RQ.
2145  */
2146 static void __ib_drain_rq(struct ib_qp *qp)
2147 {
2148 	struct ib_cq *cq = qp->recv_cq;
2149 	struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
2150 	struct ib_drain_cqe rdrain;
2151 	struct ib_recv_wr rwr = {}, *bad_rwr;
2152 	int ret;
2153 
2154 	rwr.wr_cqe = &rdrain.cqe;
2155 	rdrain.cqe.done = ib_drain_qp_done;
2156 	init_completion(&rdrain.done);
2157 
2158 	ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
2159 	if (ret) {
2160 		WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
2161 		return;
2162 	}
2163 
2164 	ret = ib_post_recv(qp, &rwr, &bad_rwr);
2165 	if (ret) {
2166 		WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
2167 		return;
2168 	}
2169 
2170 	if (cq->poll_ctx == IB_POLL_DIRECT)
2171 		while (wait_for_completion_timeout(&rdrain.done, HZ / 10) <= 0)
2172 			ib_process_cq_direct(cq, -1);
2173 	else
2174 		wait_for_completion(&rdrain.done);
2175 }
2176 
2177 /**
2178  * ib_drain_sq() - Block until all SQ CQEs have been consumed by the
2179  *		   application.
2180  * @qp:            queue pair to drain
2181  *
2182  * If the device has a provider-specific drain function, then
2183  * call that.  Otherwise call the generic drain function
2184  * __ib_drain_sq().
2185  *
2186  * The caller must:
2187  *
2188  * ensure there is room in the CQ and SQ for the drain work request and
2189  * completion.
2190  *
2191  * allocate the CQ using ib_alloc_cq().
2192  *
2193  * ensure that there are no other contexts that are posting WRs concurrently.
2194  * Otherwise the drain is not guaranteed.
2195  */
2196 void ib_drain_sq(struct ib_qp *qp)
2197 {
2198 	if (qp->device->drain_sq)
2199 		qp->device->drain_sq(qp);
2200 	else
2201 		__ib_drain_sq(qp);
2202 }
2203 EXPORT_SYMBOL(ib_drain_sq);
2204 
2205 /**
2206  * ib_drain_rq() - Block until all RQ CQEs have been consumed by the
2207  *		   application.
2208  * @qp:            queue pair to drain
2209  *
2210  * If the device has a provider-specific drain function, then
2211  * call that.  Otherwise call the generic drain function
2212  * __ib_drain_rq().
2213  *
2214  * The caller must:
2215  *
2216  * ensure there is room in the CQ and RQ for the drain work request and
2217  * completion.
2218  *
2219  * allocate the CQ using ib_alloc_cq().
2220  *
2221  * ensure that there are no other contexts that are posting WRs concurrently.
2222  * Otherwise the drain is not guaranteed.
2223  */
2224 void ib_drain_rq(struct ib_qp *qp)
2225 {
2226 	if (qp->device->drain_rq)
2227 		qp->device->drain_rq(qp);
2228 	else
2229 		__ib_drain_rq(qp);
2230 }
2231 EXPORT_SYMBOL(ib_drain_rq);
2232 
2233 /**
2234  * ib_drain_qp() - Block until all CQEs have been consumed by the
2235  *		   application on both the RQ and SQ.
2236  * @qp:            queue pair to drain
2237  *
2238  * The caller must:
2239  *
2240  * ensure there is room in the CQ(s), SQ, and RQ for drain work requests
2241  * and completions.
2242  *
2243  * allocate the CQs using ib_alloc_cq().
2244  *
2245  * ensure that there are no other contexts that are posting WRs concurrently.
2246  * Otherwise the drain is not guaranteed.
2247  */
2248 void ib_drain_qp(struct ib_qp *qp)
2249 {
2250 	ib_drain_sq(qp);
2251 	if (!qp->srq)
2252 		ib_drain_rq(qp);
2253 }
2254 EXPORT_SYMBOL(ib_drain_qp);
2255