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.
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:
16 * Redistribution and use in source and binary forms, with or
17 * without modification, are permitted provided that the following
20 * - Redistributions of source code must retain the above
21 * copyright notice, this list of conditions and the following
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.
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
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>
45 #include <linux/in6.h>
46 #include <net/addrconf.h>
47 #include <linux/security.h>
49 #include <rdma/ib_verbs.h>
50 #include <rdma/ib_cache.h>
51 #include <rdma/ib_addr.h>
54 #include "core_priv.h"
56 static int ib_resolve_eth_dmac(struct ib_device *device,
57 struct rdma_ah_attr *ah_attr);
59 static const char * const ib_events[] = {
60 [IB_EVENT_CQ_ERR] = "CQ error",
61 [IB_EVENT_QP_FATAL] = "QP fatal error",
62 [IB_EVENT_QP_REQ_ERR] = "QP request error",
63 [IB_EVENT_QP_ACCESS_ERR] = "QP access error",
64 [IB_EVENT_COMM_EST] = "communication established",
65 [IB_EVENT_SQ_DRAINED] = "send queue drained",
66 [IB_EVENT_PATH_MIG] = "path migration successful",
67 [IB_EVENT_PATH_MIG_ERR] = "path migration error",
68 [IB_EVENT_DEVICE_FATAL] = "device fatal error",
69 [IB_EVENT_PORT_ACTIVE] = "port active",
70 [IB_EVENT_PORT_ERR] = "port error",
71 [IB_EVENT_LID_CHANGE] = "LID change",
72 [IB_EVENT_PKEY_CHANGE] = "P_key change",
73 [IB_EVENT_SM_CHANGE] = "SM change",
74 [IB_EVENT_SRQ_ERR] = "SRQ error",
75 [IB_EVENT_SRQ_LIMIT_REACHED] = "SRQ limit reached",
76 [IB_EVENT_QP_LAST_WQE_REACHED] = "last WQE reached",
77 [IB_EVENT_CLIENT_REREGISTER] = "client reregister",
78 [IB_EVENT_GID_CHANGE] = "GID changed",
81 const char *__attribute_const__ ib_event_msg(enum ib_event_type event)
85 return (index < ARRAY_SIZE(ib_events) && ib_events[index]) ?
86 ib_events[index] : "unrecognized event";
88 EXPORT_SYMBOL(ib_event_msg);
90 static const char * const wc_statuses[] = {
91 [IB_WC_SUCCESS] = "success",
92 [IB_WC_LOC_LEN_ERR] = "local length error",
93 [IB_WC_LOC_QP_OP_ERR] = "local QP operation error",
94 [IB_WC_LOC_EEC_OP_ERR] = "local EE context operation error",
95 [IB_WC_LOC_PROT_ERR] = "local protection error",
96 [IB_WC_WR_FLUSH_ERR] = "WR flushed",
97 [IB_WC_MW_BIND_ERR] = "memory management operation error",
98 [IB_WC_BAD_RESP_ERR] = "bad response error",
99 [IB_WC_LOC_ACCESS_ERR] = "local access error",
100 [IB_WC_REM_INV_REQ_ERR] = "invalid request error",
101 [IB_WC_REM_ACCESS_ERR] = "remote access error",
102 [IB_WC_REM_OP_ERR] = "remote operation error",
103 [IB_WC_RETRY_EXC_ERR] = "transport retry counter exceeded",
104 [IB_WC_RNR_RETRY_EXC_ERR] = "RNR retry counter exceeded",
105 [IB_WC_LOC_RDD_VIOL_ERR] = "local RDD violation error",
106 [IB_WC_REM_INV_RD_REQ_ERR] = "remote invalid RD request",
107 [IB_WC_REM_ABORT_ERR] = "operation aborted",
108 [IB_WC_INV_EECN_ERR] = "invalid EE context number",
109 [IB_WC_INV_EEC_STATE_ERR] = "invalid EE context state",
110 [IB_WC_FATAL_ERR] = "fatal error",
111 [IB_WC_RESP_TIMEOUT_ERR] = "response timeout error",
112 [IB_WC_GENERAL_ERR] = "general error",
115 const char *__attribute_const__ ib_wc_status_msg(enum ib_wc_status status)
117 size_t index = status;
119 return (index < ARRAY_SIZE(wc_statuses) && wc_statuses[index]) ?
120 wc_statuses[index] : "unrecognized status";
122 EXPORT_SYMBOL(ib_wc_status_msg);
124 __attribute_const__ int ib_rate_to_mult(enum ib_rate rate)
127 case IB_RATE_2_5_GBPS: return 1;
128 case IB_RATE_5_GBPS: return 2;
129 case IB_RATE_10_GBPS: return 4;
130 case IB_RATE_20_GBPS: return 8;
131 case IB_RATE_30_GBPS: return 12;
132 case IB_RATE_40_GBPS: return 16;
133 case IB_RATE_60_GBPS: return 24;
134 case IB_RATE_80_GBPS: return 32;
135 case IB_RATE_120_GBPS: return 48;
136 case IB_RATE_14_GBPS: return 6;
137 case IB_RATE_56_GBPS: return 22;
138 case IB_RATE_112_GBPS: return 45;
139 case IB_RATE_168_GBPS: return 67;
140 case IB_RATE_25_GBPS: return 10;
141 case IB_RATE_100_GBPS: return 40;
142 case IB_RATE_200_GBPS: return 80;
143 case IB_RATE_300_GBPS: return 120;
147 EXPORT_SYMBOL(ib_rate_to_mult);
149 __attribute_const__ enum ib_rate mult_to_ib_rate(int mult)
152 case 1: return IB_RATE_2_5_GBPS;
153 case 2: return IB_RATE_5_GBPS;
154 case 4: return IB_RATE_10_GBPS;
155 case 8: return IB_RATE_20_GBPS;
156 case 12: return IB_RATE_30_GBPS;
157 case 16: return IB_RATE_40_GBPS;
158 case 24: return IB_RATE_60_GBPS;
159 case 32: return IB_RATE_80_GBPS;
160 case 48: return IB_RATE_120_GBPS;
161 case 6: return IB_RATE_14_GBPS;
162 case 22: return IB_RATE_56_GBPS;
163 case 45: return IB_RATE_112_GBPS;
164 case 67: return IB_RATE_168_GBPS;
165 case 10: return IB_RATE_25_GBPS;
166 case 40: return IB_RATE_100_GBPS;
167 case 80: return IB_RATE_200_GBPS;
168 case 120: return IB_RATE_300_GBPS;
169 default: return IB_RATE_PORT_CURRENT;
172 EXPORT_SYMBOL(mult_to_ib_rate);
174 __attribute_const__ int ib_rate_to_mbps(enum ib_rate rate)
177 case IB_RATE_2_5_GBPS: return 2500;
178 case IB_RATE_5_GBPS: return 5000;
179 case IB_RATE_10_GBPS: return 10000;
180 case IB_RATE_20_GBPS: return 20000;
181 case IB_RATE_30_GBPS: return 30000;
182 case IB_RATE_40_GBPS: return 40000;
183 case IB_RATE_60_GBPS: return 60000;
184 case IB_RATE_80_GBPS: return 80000;
185 case IB_RATE_120_GBPS: return 120000;
186 case IB_RATE_14_GBPS: return 14062;
187 case IB_RATE_56_GBPS: return 56250;
188 case IB_RATE_112_GBPS: return 112500;
189 case IB_RATE_168_GBPS: return 168750;
190 case IB_RATE_25_GBPS: return 25781;
191 case IB_RATE_100_GBPS: return 103125;
192 case IB_RATE_200_GBPS: return 206250;
193 case IB_RATE_300_GBPS: return 309375;
197 EXPORT_SYMBOL(ib_rate_to_mbps);
199 __attribute_const__ enum rdma_transport_type
200 rdma_node_get_transport(enum rdma_node_type node_type)
203 if (node_type == RDMA_NODE_USNIC)
204 return RDMA_TRANSPORT_USNIC;
205 if (node_type == RDMA_NODE_USNIC_UDP)
206 return RDMA_TRANSPORT_USNIC_UDP;
207 if (node_type == RDMA_NODE_RNIC)
208 return RDMA_TRANSPORT_IWARP;
210 return RDMA_TRANSPORT_IB;
212 EXPORT_SYMBOL(rdma_node_get_transport);
214 enum rdma_link_layer rdma_port_get_link_layer(struct ib_device *device, u8 port_num)
216 enum rdma_transport_type lt;
217 if (device->get_link_layer)
218 return device->get_link_layer(device, port_num);
220 lt = rdma_node_get_transport(device->node_type);
221 if (lt == RDMA_TRANSPORT_IB)
222 return IB_LINK_LAYER_INFINIBAND;
224 return IB_LINK_LAYER_ETHERNET;
226 EXPORT_SYMBOL(rdma_port_get_link_layer);
228 /* Protection domains */
231 * ib_alloc_pd - Allocates an unused protection domain.
232 * @device: The device on which to allocate the protection domain.
234 * A protection domain object provides an association between QPs, shared
235 * receive queues, address handles, memory regions, and memory windows.
237 * Every PD has a local_dma_lkey which can be used as the lkey value for local
240 struct ib_pd *__ib_alloc_pd(struct ib_device *device, unsigned int flags,
244 int mr_access_flags = 0;
246 pd = device->alloc_pd(device, NULL, NULL);
252 pd->__internal_mr = NULL;
253 atomic_set(&pd->usecnt, 0);
256 if (device->attrs.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY)
257 pd->local_dma_lkey = device->local_dma_lkey;
259 mr_access_flags |= IB_ACCESS_LOCAL_WRITE;
261 if (flags & IB_PD_UNSAFE_GLOBAL_RKEY) {
262 pr_warn("%s: enabling unsafe global rkey\n", caller);
263 mr_access_flags |= IB_ACCESS_REMOTE_READ | IB_ACCESS_REMOTE_WRITE;
266 pd->res.type = RDMA_RESTRACK_PD;
267 pd->res.kern_name = caller;
268 rdma_restrack_add(&pd->res);
270 if (mr_access_flags) {
273 mr = pd->device->get_dma_mr(pd, mr_access_flags);
279 mr->device = pd->device;
282 mr->need_inval = false;
284 pd->__internal_mr = mr;
286 if (!(device->attrs.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY))
287 pd->local_dma_lkey = pd->__internal_mr->lkey;
289 if (flags & IB_PD_UNSAFE_GLOBAL_RKEY)
290 pd->unsafe_global_rkey = pd->__internal_mr->rkey;
295 EXPORT_SYMBOL(__ib_alloc_pd);
298 * ib_dealloc_pd - Deallocates a protection domain.
299 * @pd: The protection domain to deallocate.
301 * It is an error to call this function while any resources in the pd still
302 * exist. The caller is responsible to synchronously destroy them and
303 * guarantee no new allocations will happen.
305 void ib_dealloc_pd(struct ib_pd *pd)
309 if (pd->__internal_mr) {
310 ret = pd->device->dereg_mr(pd->__internal_mr);
312 pd->__internal_mr = NULL;
315 /* uverbs manipulates usecnt with proper locking, while the kabi
316 requires the caller to guarantee we can't race here. */
317 WARN_ON(atomic_read(&pd->usecnt));
319 rdma_restrack_del(&pd->res);
320 /* Making delalloc_pd a void return is a WIP, no driver should return
322 ret = pd->device->dealloc_pd(pd);
323 WARN_ONCE(ret, "Infiniband HW driver failed dealloc_pd");
325 EXPORT_SYMBOL(ib_dealloc_pd);
327 /* Address handles */
330 * rdma_copy_ah_attr - Copy rdma ah attribute from source to destination.
331 * @dest: Pointer to destination ah_attr. Contents of the destination
332 * pointer is assumed to be invalid and attribute are overwritten.
333 * @src: Pointer to source ah_attr.
335 void rdma_copy_ah_attr(struct rdma_ah_attr *dest,
336 const struct rdma_ah_attr *src)
339 if (dest->grh.sgid_attr)
340 rdma_hold_gid_attr(dest->grh.sgid_attr);
342 EXPORT_SYMBOL(rdma_copy_ah_attr);
345 * rdma_replace_ah_attr - Replace valid ah_attr with new new one.
346 * @old: Pointer to existing ah_attr which needs to be replaced.
347 * old is assumed to be valid or zero'd
348 * @new: Pointer to the new ah_attr.
350 * rdma_replace_ah_attr() first releases any reference in the old ah_attr if
351 * old the ah_attr is valid; after that it copies the new attribute and holds
352 * the reference to the replaced ah_attr.
354 void rdma_replace_ah_attr(struct rdma_ah_attr *old,
355 const struct rdma_ah_attr *new)
357 rdma_destroy_ah_attr(old);
359 if (old->grh.sgid_attr)
360 rdma_hold_gid_attr(old->grh.sgid_attr);
362 EXPORT_SYMBOL(rdma_replace_ah_attr);
365 * rdma_move_ah_attr - Move ah_attr pointed by source to destination.
366 * @dest: Pointer to destination ah_attr to copy to.
367 * dest is assumed to be valid or zero'd
368 * @src: Pointer to the new ah_attr.
370 * rdma_move_ah_attr() first releases any reference in the destination ah_attr
371 * if it is valid. This also transfers ownership of internal references from
372 * src to dest, making src invalid in the process. No new reference of the src
375 void rdma_move_ah_attr(struct rdma_ah_attr *dest, struct rdma_ah_attr *src)
377 rdma_destroy_ah_attr(dest);
379 src->grh.sgid_attr = NULL;
381 EXPORT_SYMBOL(rdma_move_ah_attr);
384 * Validate that the rdma_ah_attr is valid for the device before passing it
387 static int rdma_check_ah_attr(struct ib_device *device,
388 struct rdma_ah_attr *ah_attr)
390 if (!rdma_is_port_valid(device, ah_attr->port_num))
393 if ((rdma_is_grh_required(device, ah_attr->port_num) ||
394 ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE) &&
395 !(ah_attr->ah_flags & IB_AH_GRH))
398 if (ah_attr->grh.sgid_attr) {
400 * Make sure the passed sgid_attr is consistent with the
403 if (ah_attr->grh.sgid_attr->index != ah_attr->grh.sgid_index ||
404 ah_attr->grh.sgid_attr->port_num != ah_attr->port_num)
411 * If the ah requires a GRH then ensure that sgid_attr pointer is filled in.
412 * On success the caller is responsible to call rdma_unfill_sgid_attr().
414 static int rdma_fill_sgid_attr(struct ib_device *device,
415 struct rdma_ah_attr *ah_attr,
416 const struct ib_gid_attr **old_sgid_attr)
418 const struct ib_gid_attr *sgid_attr;
419 struct ib_global_route *grh;
422 *old_sgid_attr = ah_attr->grh.sgid_attr;
424 ret = rdma_check_ah_attr(device, ah_attr);
428 if (!(ah_attr->ah_flags & IB_AH_GRH))
431 grh = rdma_ah_retrieve_grh(ah_attr);
436 rdma_get_gid_attr(device, ah_attr->port_num, grh->sgid_index);
437 if (IS_ERR(sgid_attr))
438 return PTR_ERR(sgid_attr);
440 /* Move ownerhip of the kref into the ah_attr */
441 grh->sgid_attr = sgid_attr;
445 static void rdma_unfill_sgid_attr(struct rdma_ah_attr *ah_attr,
446 const struct ib_gid_attr *old_sgid_attr)
449 * Fill didn't change anything, the caller retains ownership of
452 if (ah_attr->grh.sgid_attr == old_sgid_attr)
456 * Otherwise, we need to undo what rdma_fill_sgid_attr so the caller
457 * doesn't see any change in the rdma_ah_attr. If we get here
458 * old_sgid_attr is NULL.
460 rdma_destroy_ah_attr(ah_attr);
463 static const struct ib_gid_attr *
464 rdma_update_sgid_attr(struct rdma_ah_attr *ah_attr,
465 const struct ib_gid_attr *old_attr)
468 rdma_put_gid_attr(old_attr);
469 if (ah_attr->ah_flags & IB_AH_GRH) {
470 rdma_hold_gid_attr(ah_attr->grh.sgid_attr);
471 return ah_attr->grh.sgid_attr;
476 static struct ib_ah *_rdma_create_ah(struct ib_pd *pd,
477 struct rdma_ah_attr *ah_attr,
478 struct ib_udata *udata)
482 if (!pd->device->create_ah)
483 return ERR_PTR(-EOPNOTSUPP);
485 ah = pd->device->create_ah(pd, ah_attr, udata);
488 ah->device = pd->device;
491 ah->type = ah_attr->type;
492 ah->sgid_attr = rdma_update_sgid_attr(ah_attr, NULL);
494 atomic_inc(&pd->usecnt);
501 * rdma_create_ah - Creates an address handle for the
502 * given address vector.
503 * @pd: The protection domain associated with the address handle.
504 * @ah_attr: The attributes of the address vector.
506 * It returns 0 on success and returns appropriate error code on error.
507 * The address handle is used to reference a local or global destination
508 * in all UD QP post sends.
510 struct ib_ah *rdma_create_ah(struct ib_pd *pd, struct rdma_ah_attr *ah_attr)
512 const struct ib_gid_attr *old_sgid_attr;
516 ret = rdma_fill_sgid_attr(pd->device, ah_attr, &old_sgid_attr);
520 ah = _rdma_create_ah(pd, ah_attr, NULL);
522 rdma_unfill_sgid_attr(ah_attr, old_sgid_attr);
525 EXPORT_SYMBOL(rdma_create_ah);
528 * rdma_create_user_ah - Creates an address handle for the
529 * given address vector.
530 * It resolves destination mac address for ah attribute of RoCE type.
531 * @pd: The protection domain associated with the address handle.
532 * @ah_attr: The attributes of the address vector.
533 * @udata: pointer to user's input output buffer information need by
536 * It returns 0 on success and returns appropriate error code on error.
537 * The address handle is used to reference a local or global destination
538 * in all UD QP post sends.
540 struct ib_ah *rdma_create_user_ah(struct ib_pd *pd,
541 struct rdma_ah_attr *ah_attr,
542 struct ib_udata *udata)
544 const struct ib_gid_attr *old_sgid_attr;
548 err = rdma_fill_sgid_attr(pd->device, ah_attr, &old_sgid_attr);
552 if (ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE) {
553 err = ib_resolve_eth_dmac(pd->device, ah_attr);
560 ah = _rdma_create_ah(pd, ah_attr, udata);
563 rdma_unfill_sgid_attr(ah_attr, old_sgid_attr);
566 EXPORT_SYMBOL(rdma_create_user_ah);
568 int ib_get_rdma_header_version(const union rdma_network_hdr *hdr)
570 const struct iphdr *ip4h = (struct iphdr *)&hdr->roce4grh;
571 struct iphdr ip4h_checked;
572 const struct ipv6hdr *ip6h = (struct ipv6hdr *)&hdr->ibgrh;
574 /* If it's IPv6, the version must be 6, otherwise, the first
575 * 20 bytes (before the IPv4 header) are garbled.
577 if (ip6h->version != 6)
578 return (ip4h->version == 4) ? 4 : 0;
579 /* version may be 6 or 4 because the first 20 bytes could be garbled */
581 /* RoCE v2 requires no options, thus header length
588 * We can't write on scattered buffers so we need to copy to
591 memcpy(&ip4h_checked, ip4h, sizeof(ip4h_checked));
592 ip4h_checked.check = 0;
593 ip4h_checked.check = ip_fast_csum((u8 *)&ip4h_checked, 5);
594 /* if IPv4 header checksum is OK, believe it */
595 if (ip4h->check == ip4h_checked.check)
599 EXPORT_SYMBOL(ib_get_rdma_header_version);
601 static enum rdma_network_type ib_get_net_type_by_grh(struct ib_device *device,
603 const struct ib_grh *grh)
607 if (rdma_protocol_ib(device, port_num))
608 return RDMA_NETWORK_IB;
610 grh_version = ib_get_rdma_header_version((union rdma_network_hdr *)grh);
612 if (grh_version == 4)
613 return RDMA_NETWORK_IPV4;
615 if (grh->next_hdr == IPPROTO_UDP)
616 return RDMA_NETWORK_IPV6;
618 return RDMA_NETWORK_ROCE_V1;
621 struct find_gid_index_context {
623 enum ib_gid_type gid_type;
626 static bool find_gid_index(const union ib_gid *gid,
627 const struct ib_gid_attr *gid_attr,
630 struct find_gid_index_context *ctx = context;
632 if (ctx->gid_type != gid_attr->gid_type)
635 if ((!!(ctx->vlan_id != 0xffff) == !is_vlan_dev(gid_attr->ndev)) ||
636 (is_vlan_dev(gid_attr->ndev) &&
637 vlan_dev_vlan_id(gid_attr->ndev) != ctx->vlan_id))
643 static const struct ib_gid_attr *
644 get_sgid_attr_from_eth(struct ib_device *device, u8 port_num,
645 u16 vlan_id, const union ib_gid *sgid,
646 enum ib_gid_type gid_type)
648 struct find_gid_index_context context = {.vlan_id = vlan_id,
649 .gid_type = gid_type};
651 return rdma_find_gid_by_filter(device, sgid, port_num, find_gid_index,
655 int ib_get_gids_from_rdma_hdr(const union rdma_network_hdr *hdr,
656 enum rdma_network_type net_type,
657 union ib_gid *sgid, union ib_gid *dgid)
659 struct sockaddr_in src_in;
660 struct sockaddr_in dst_in;
661 __be32 src_saddr, dst_saddr;
666 if (net_type == RDMA_NETWORK_IPV4) {
667 memcpy(&src_in.sin_addr.s_addr,
668 &hdr->roce4grh.saddr, 4);
669 memcpy(&dst_in.sin_addr.s_addr,
670 &hdr->roce4grh.daddr, 4);
671 src_saddr = src_in.sin_addr.s_addr;
672 dst_saddr = dst_in.sin_addr.s_addr;
673 ipv6_addr_set_v4mapped(src_saddr,
674 (struct in6_addr *)sgid);
675 ipv6_addr_set_v4mapped(dst_saddr,
676 (struct in6_addr *)dgid);
678 } else if (net_type == RDMA_NETWORK_IPV6 ||
679 net_type == RDMA_NETWORK_IB) {
680 *dgid = hdr->ibgrh.dgid;
681 *sgid = hdr->ibgrh.sgid;
687 EXPORT_SYMBOL(ib_get_gids_from_rdma_hdr);
689 /* Resolve destination mac address and hop limit for unicast destination
690 * GID entry, considering the source GID entry as well.
691 * ah_attribute must have have valid port_num, sgid_index.
693 static int ib_resolve_unicast_gid_dmac(struct ib_device *device,
694 struct rdma_ah_attr *ah_attr)
696 struct ib_global_route *grh = rdma_ah_retrieve_grh(ah_attr);
697 const struct ib_gid_attr *sgid_attr = grh->sgid_attr;
698 int hop_limit = 0xff;
701 /* If destination is link local and source GID is RoCEv1,
702 * IP stack is not used.
704 if (rdma_link_local_addr((struct in6_addr *)grh->dgid.raw) &&
705 sgid_attr->gid_type == IB_GID_TYPE_ROCE) {
706 rdma_get_ll_mac((struct in6_addr *)grh->dgid.raw,
711 ret = rdma_addr_find_l2_eth_by_grh(&sgid_attr->gid, &grh->dgid,
713 sgid_attr->ndev, &hop_limit);
715 grh->hop_limit = hop_limit;
720 * This function initializes address handle attributes from the incoming packet.
721 * Incoming packet has dgid of the receiver node on which this code is
722 * getting executed and, sgid contains the GID of the sender.
724 * When resolving mac address of destination, the arrived dgid is used
725 * as sgid and, sgid is used as dgid because sgid contains destinations
726 * GID whom to respond to.
728 * On success the caller is responsible to call rdma_destroy_ah_attr on the
731 int ib_init_ah_attr_from_wc(struct ib_device *device, u8 port_num,
732 const struct ib_wc *wc, const struct ib_grh *grh,
733 struct rdma_ah_attr *ah_attr)
737 enum rdma_network_type net_type = RDMA_NETWORK_IB;
738 enum ib_gid_type gid_type = IB_GID_TYPE_IB;
739 const struct ib_gid_attr *sgid_attr;
746 memset(ah_attr, 0, sizeof *ah_attr);
747 ah_attr->type = rdma_ah_find_type(device, port_num);
748 if (rdma_cap_eth_ah(device, port_num)) {
749 if (wc->wc_flags & IB_WC_WITH_NETWORK_HDR_TYPE)
750 net_type = wc->network_hdr_type;
752 net_type = ib_get_net_type_by_grh(device, port_num, grh);
753 gid_type = ib_network_to_gid_type(net_type);
755 ret = ib_get_gids_from_rdma_hdr((union rdma_network_hdr *)grh, net_type,
760 rdma_ah_set_sl(ah_attr, wc->sl);
761 rdma_ah_set_port_num(ah_attr, port_num);
763 if (rdma_protocol_roce(device, port_num)) {
764 u16 vlan_id = wc->wc_flags & IB_WC_WITH_VLAN ?
765 wc->vlan_id : 0xffff;
767 if (!(wc->wc_flags & IB_WC_GRH))
770 sgid_attr = get_sgid_attr_from_eth(device, port_num,
773 if (IS_ERR(sgid_attr))
774 return PTR_ERR(sgid_attr);
776 flow_class = be32_to_cpu(grh->version_tclass_flow);
777 rdma_move_grh_sgid_attr(ah_attr,
779 flow_class & 0xFFFFF,
781 (flow_class >> 20) & 0xFF,
784 ret = ib_resolve_unicast_gid_dmac(device, ah_attr);
786 rdma_destroy_ah_attr(ah_attr);
790 rdma_ah_set_dlid(ah_attr, wc->slid);
791 rdma_ah_set_path_bits(ah_attr, wc->dlid_path_bits);
793 if ((wc->wc_flags & IB_WC_GRH) == 0)
796 if (dgid.global.interface_id !=
797 cpu_to_be64(IB_SA_WELL_KNOWN_GUID)) {
798 sgid_attr = rdma_find_gid_by_port(
799 device, &dgid, IB_GID_TYPE_IB, port_num, NULL);
801 sgid_attr = rdma_get_gid_attr(device, port_num, 0);
803 if (IS_ERR(sgid_attr))
804 return PTR_ERR(sgid_attr);
805 flow_class = be32_to_cpu(grh->version_tclass_flow);
806 rdma_move_grh_sgid_attr(ah_attr,
808 flow_class & 0xFFFFF,
810 (flow_class >> 20) & 0xFF,
816 EXPORT_SYMBOL(ib_init_ah_attr_from_wc);
819 * rdma_move_grh_sgid_attr - Sets the sgid attribute of GRH, taking ownership
822 * @attr: Pointer to AH attribute structure
823 * @dgid: Destination GID
824 * @flow_label: Flow label
825 * @hop_limit: Hop limit
826 * @traffic_class: traffic class
827 * @sgid_attr: Pointer to SGID attribute
829 * This takes ownership of the sgid_attr reference. The caller must ensure
830 * rdma_destroy_ah_attr() is called before destroying the rdma_ah_attr after
831 * calling this function.
833 void rdma_move_grh_sgid_attr(struct rdma_ah_attr *attr, union ib_gid *dgid,
834 u32 flow_label, u8 hop_limit, u8 traffic_class,
835 const struct ib_gid_attr *sgid_attr)
837 rdma_ah_set_grh(attr, dgid, flow_label, sgid_attr->index, hop_limit,
839 attr->grh.sgid_attr = sgid_attr;
841 EXPORT_SYMBOL(rdma_move_grh_sgid_attr);
844 * rdma_destroy_ah_attr - Release reference to SGID attribute of
846 * @ah_attr: Pointer to ah attribute
848 * Release reference to the SGID attribute of the ah attribute if it is
849 * non NULL. It is safe to call this multiple times, and safe to call it on
850 * a zero initialized ah_attr.
852 void rdma_destroy_ah_attr(struct rdma_ah_attr *ah_attr)
854 if (ah_attr->grh.sgid_attr) {
855 rdma_put_gid_attr(ah_attr->grh.sgid_attr);
856 ah_attr->grh.sgid_attr = NULL;
859 EXPORT_SYMBOL(rdma_destroy_ah_attr);
861 struct ib_ah *ib_create_ah_from_wc(struct ib_pd *pd, const struct ib_wc *wc,
862 const struct ib_grh *grh, u8 port_num)
864 struct rdma_ah_attr ah_attr;
868 ret = ib_init_ah_attr_from_wc(pd->device, port_num, wc, grh, &ah_attr);
872 ah = rdma_create_ah(pd, &ah_attr);
874 rdma_destroy_ah_attr(&ah_attr);
877 EXPORT_SYMBOL(ib_create_ah_from_wc);
879 int rdma_modify_ah(struct ib_ah *ah, struct rdma_ah_attr *ah_attr)
881 const struct ib_gid_attr *old_sgid_attr;
884 if (ah->type != ah_attr->type)
887 ret = rdma_fill_sgid_attr(ah->device, ah_attr, &old_sgid_attr);
891 ret = ah->device->modify_ah ?
892 ah->device->modify_ah(ah, ah_attr) :
895 ah->sgid_attr = rdma_update_sgid_attr(ah_attr, ah->sgid_attr);
896 rdma_unfill_sgid_attr(ah_attr, old_sgid_attr);
899 EXPORT_SYMBOL(rdma_modify_ah);
901 int rdma_query_ah(struct ib_ah *ah, struct rdma_ah_attr *ah_attr)
903 ah_attr->grh.sgid_attr = NULL;
905 return ah->device->query_ah ?
906 ah->device->query_ah(ah, ah_attr) :
909 EXPORT_SYMBOL(rdma_query_ah);
911 int rdma_destroy_ah(struct ib_ah *ah)
913 const struct ib_gid_attr *sgid_attr = ah->sgid_attr;
918 ret = ah->device->destroy_ah(ah);
920 atomic_dec(&pd->usecnt);
922 rdma_put_gid_attr(sgid_attr);
927 EXPORT_SYMBOL(rdma_destroy_ah);
929 /* Shared receive queues */
931 struct ib_srq *ib_create_srq(struct ib_pd *pd,
932 struct ib_srq_init_attr *srq_init_attr)
936 if (!pd->device->create_srq)
937 return ERR_PTR(-EOPNOTSUPP);
939 srq = pd->device->create_srq(pd, srq_init_attr, NULL);
942 srq->device = pd->device;
945 srq->event_handler = srq_init_attr->event_handler;
946 srq->srq_context = srq_init_attr->srq_context;
947 srq->srq_type = srq_init_attr->srq_type;
948 if (ib_srq_has_cq(srq->srq_type)) {
949 srq->ext.cq = srq_init_attr->ext.cq;
950 atomic_inc(&srq->ext.cq->usecnt);
952 if (srq->srq_type == IB_SRQT_XRC) {
953 srq->ext.xrc.xrcd = srq_init_attr->ext.xrc.xrcd;
954 atomic_inc(&srq->ext.xrc.xrcd->usecnt);
956 atomic_inc(&pd->usecnt);
957 atomic_set(&srq->usecnt, 0);
962 EXPORT_SYMBOL(ib_create_srq);
964 int ib_modify_srq(struct ib_srq *srq,
965 struct ib_srq_attr *srq_attr,
966 enum ib_srq_attr_mask srq_attr_mask)
968 return srq->device->modify_srq ?
969 srq->device->modify_srq(srq, srq_attr, srq_attr_mask, NULL) :
972 EXPORT_SYMBOL(ib_modify_srq);
974 int ib_query_srq(struct ib_srq *srq,
975 struct ib_srq_attr *srq_attr)
977 return srq->device->query_srq ?
978 srq->device->query_srq(srq, srq_attr) : -EOPNOTSUPP;
980 EXPORT_SYMBOL(ib_query_srq);
982 int ib_destroy_srq(struct ib_srq *srq)
985 enum ib_srq_type srq_type;
986 struct ib_xrcd *uninitialized_var(xrcd);
987 struct ib_cq *uninitialized_var(cq);
990 if (atomic_read(&srq->usecnt))
994 srq_type = srq->srq_type;
995 if (ib_srq_has_cq(srq_type))
997 if (srq_type == IB_SRQT_XRC)
998 xrcd = srq->ext.xrc.xrcd;
1000 ret = srq->device->destroy_srq(srq);
1002 atomic_dec(&pd->usecnt);
1003 if (srq_type == IB_SRQT_XRC)
1004 atomic_dec(&xrcd->usecnt);
1005 if (ib_srq_has_cq(srq_type))
1006 atomic_dec(&cq->usecnt);
1011 EXPORT_SYMBOL(ib_destroy_srq);
1015 static void __ib_shared_qp_event_handler(struct ib_event *event, void *context)
1017 struct ib_qp *qp = context;
1018 unsigned long flags;
1020 spin_lock_irqsave(&qp->device->event_handler_lock, flags);
1021 list_for_each_entry(event->element.qp, &qp->open_list, open_list)
1022 if (event->element.qp->event_handler)
1023 event->element.qp->event_handler(event, event->element.qp->qp_context);
1024 spin_unlock_irqrestore(&qp->device->event_handler_lock, flags);
1027 static void __ib_insert_xrcd_qp(struct ib_xrcd *xrcd, struct ib_qp *qp)
1029 mutex_lock(&xrcd->tgt_qp_mutex);
1030 list_add(&qp->xrcd_list, &xrcd->tgt_qp_list);
1031 mutex_unlock(&xrcd->tgt_qp_mutex);
1034 static struct ib_qp *__ib_open_qp(struct ib_qp *real_qp,
1035 void (*event_handler)(struct ib_event *, void *),
1039 unsigned long flags;
1042 qp = kzalloc(sizeof *qp, GFP_KERNEL);
1044 return ERR_PTR(-ENOMEM);
1046 qp->real_qp = real_qp;
1047 err = ib_open_shared_qp_security(qp, real_qp->device);
1050 return ERR_PTR(err);
1053 qp->real_qp = real_qp;
1054 atomic_inc(&real_qp->usecnt);
1055 qp->device = real_qp->device;
1056 qp->event_handler = event_handler;
1057 qp->qp_context = qp_context;
1058 qp->qp_num = real_qp->qp_num;
1059 qp->qp_type = real_qp->qp_type;
1061 spin_lock_irqsave(&real_qp->device->event_handler_lock, flags);
1062 list_add(&qp->open_list, &real_qp->open_list);
1063 spin_unlock_irqrestore(&real_qp->device->event_handler_lock, flags);
1068 struct ib_qp *ib_open_qp(struct ib_xrcd *xrcd,
1069 struct ib_qp_open_attr *qp_open_attr)
1071 struct ib_qp *qp, *real_qp;
1073 if (qp_open_attr->qp_type != IB_QPT_XRC_TGT)
1074 return ERR_PTR(-EINVAL);
1076 qp = ERR_PTR(-EINVAL);
1077 mutex_lock(&xrcd->tgt_qp_mutex);
1078 list_for_each_entry(real_qp, &xrcd->tgt_qp_list, xrcd_list) {
1079 if (real_qp->qp_num == qp_open_attr->qp_num) {
1080 qp = __ib_open_qp(real_qp, qp_open_attr->event_handler,
1081 qp_open_attr->qp_context);
1085 mutex_unlock(&xrcd->tgt_qp_mutex);
1088 EXPORT_SYMBOL(ib_open_qp);
1090 static struct ib_qp *create_xrc_qp(struct ib_qp *qp,
1091 struct ib_qp_init_attr *qp_init_attr)
1093 struct ib_qp *real_qp = qp;
1095 qp->event_handler = __ib_shared_qp_event_handler;
1096 qp->qp_context = qp;
1098 qp->send_cq = qp->recv_cq = NULL;
1100 qp->xrcd = qp_init_attr->xrcd;
1101 atomic_inc(&qp_init_attr->xrcd->usecnt);
1102 INIT_LIST_HEAD(&qp->open_list);
1104 qp = __ib_open_qp(real_qp, qp_init_attr->event_handler,
1105 qp_init_attr->qp_context);
1109 __ib_insert_xrcd_qp(qp_init_attr->xrcd, real_qp);
1113 struct ib_qp *ib_create_qp(struct ib_pd *pd,
1114 struct ib_qp_init_attr *qp_init_attr)
1116 struct ib_device *device = pd ? pd->device : qp_init_attr->xrcd->device;
1120 if (qp_init_attr->rwq_ind_tbl &&
1121 (qp_init_attr->recv_cq ||
1122 qp_init_attr->srq || qp_init_attr->cap.max_recv_wr ||
1123 qp_init_attr->cap.max_recv_sge))
1124 return ERR_PTR(-EINVAL);
1127 * If the callers is using the RDMA API calculate the resources
1128 * needed for the RDMA READ/WRITE operations.
1130 * Note that these callers need to pass in a port number.
1132 if (qp_init_attr->cap.max_rdma_ctxs)
1133 rdma_rw_init_qp(device, qp_init_attr);
1135 qp = _ib_create_qp(device, pd, qp_init_attr, NULL, NULL);
1139 ret = ib_create_qp_security(qp, device);
1144 qp->qp_type = qp_init_attr->qp_type;
1145 qp->rwq_ind_tbl = qp_init_attr->rwq_ind_tbl;
1147 atomic_set(&qp->usecnt, 0);
1149 spin_lock_init(&qp->mr_lock);
1150 INIT_LIST_HEAD(&qp->rdma_mrs);
1151 INIT_LIST_HEAD(&qp->sig_mrs);
1154 if (qp_init_attr->qp_type == IB_QPT_XRC_TGT) {
1155 struct ib_qp *xrc_qp = create_xrc_qp(qp, qp_init_attr);
1157 if (IS_ERR(xrc_qp)) {
1158 ret = PTR_ERR(xrc_qp);
1164 qp->event_handler = qp_init_attr->event_handler;
1165 qp->qp_context = qp_init_attr->qp_context;
1166 if (qp_init_attr->qp_type == IB_QPT_XRC_INI) {
1170 qp->recv_cq = qp_init_attr->recv_cq;
1171 if (qp_init_attr->recv_cq)
1172 atomic_inc(&qp_init_attr->recv_cq->usecnt);
1173 qp->srq = qp_init_attr->srq;
1175 atomic_inc(&qp_init_attr->srq->usecnt);
1178 qp->send_cq = qp_init_attr->send_cq;
1181 atomic_inc(&pd->usecnt);
1182 if (qp_init_attr->send_cq)
1183 atomic_inc(&qp_init_attr->send_cq->usecnt);
1184 if (qp_init_attr->rwq_ind_tbl)
1185 atomic_inc(&qp->rwq_ind_tbl->usecnt);
1187 if (qp_init_attr->cap.max_rdma_ctxs) {
1188 ret = rdma_rw_init_mrs(qp, qp_init_attr);
1194 * Note: all hw drivers guarantee that max_send_sge is lower than
1195 * the device RDMA WRITE SGE limit but not all hw drivers ensure that
1196 * max_send_sge <= max_sge_rd.
1198 qp->max_write_sge = qp_init_attr->cap.max_send_sge;
1199 qp->max_read_sge = min_t(u32, qp_init_attr->cap.max_send_sge,
1200 device->attrs.max_sge_rd);
1206 return ERR_PTR(ret);
1209 EXPORT_SYMBOL(ib_create_qp);
1211 static const struct {
1213 enum ib_qp_attr_mask req_param[IB_QPT_MAX];
1214 enum ib_qp_attr_mask opt_param[IB_QPT_MAX];
1215 } qp_state_table[IB_QPS_ERR + 1][IB_QPS_ERR + 1] = {
1217 [IB_QPS_RESET] = { .valid = 1 },
1221 [IB_QPT_UD] = (IB_QP_PKEY_INDEX |
1224 [IB_QPT_RAW_PACKET] = IB_QP_PORT,
1225 [IB_QPT_UC] = (IB_QP_PKEY_INDEX |
1227 IB_QP_ACCESS_FLAGS),
1228 [IB_QPT_RC] = (IB_QP_PKEY_INDEX |
1230 IB_QP_ACCESS_FLAGS),
1231 [IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX |
1233 IB_QP_ACCESS_FLAGS),
1234 [IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX |
1236 IB_QP_ACCESS_FLAGS),
1237 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX |
1239 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX |
1245 [IB_QPS_RESET] = { .valid = 1 },
1246 [IB_QPS_ERR] = { .valid = 1 },
1250 [IB_QPT_UD] = (IB_QP_PKEY_INDEX |
1253 [IB_QPT_UC] = (IB_QP_PKEY_INDEX |
1255 IB_QP_ACCESS_FLAGS),
1256 [IB_QPT_RC] = (IB_QP_PKEY_INDEX |
1258 IB_QP_ACCESS_FLAGS),
1259 [IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX |
1261 IB_QP_ACCESS_FLAGS),
1262 [IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX |
1264 IB_QP_ACCESS_FLAGS),
1265 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX |
1267 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX |
1274 [IB_QPT_UC] = (IB_QP_AV |
1278 [IB_QPT_RC] = (IB_QP_AV |
1282 IB_QP_MAX_DEST_RD_ATOMIC |
1283 IB_QP_MIN_RNR_TIMER),
1284 [IB_QPT_XRC_INI] = (IB_QP_AV |
1288 [IB_QPT_XRC_TGT] = (IB_QP_AV |
1292 IB_QP_MAX_DEST_RD_ATOMIC |
1293 IB_QP_MIN_RNR_TIMER),
1296 [IB_QPT_UD] = (IB_QP_PKEY_INDEX |
1298 [IB_QPT_UC] = (IB_QP_ALT_PATH |
1299 IB_QP_ACCESS_FLAGS |
1301 [IB_QPT_RC] = (IB_QP_ALT_PATH |
1302 IB_QP_ACCESS_FLAGS |
1304 [IB_QPT_XRC_INI] = (IB_QP_ALT_PATH |
1305 IB_QP_ACCESS_FLAGS |
1307 [IB_QPT_XRC_TGT] = (IB_QP_ALT_PATH |
1308 IB_QP_ACCESS_FLAGS |
1310 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX |
1312 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX |
1318 [IB_QPS_RESET] = { .valid = 1 },
1319 [IB_QPS_ERR] = { .valid = 1 },
1323 [IB_QPT_UD] = IB_QP_SQ_PSN,
1324 [IB_QPT_UC] = IB_QP_SQ_PSN,
1325 [IB_QPT_RC] = (IB_QP_TIMEOUT |
1329 IB_QP_MAX_QP_RD_ATOMIC),
1330 [IB_QPT_XRC_INI] = (IB_QP_TIMEOUT |
1334 IB_QP_MAX_QP_RD_ATOMIC),
1335 [IB_QPT_XRC_TGT] = (IB_QP_TIMEOUT |
1337 [IB_QPT_SMI] = IB_QP_SQ_PSN,
1338 [IB_QPT_GSI] = IB_QP_SQ_PSN,
1341 [IB_QPT_UD] = (IB_QP_CUR_STATE |
1343 [IB_QPT_UC] = (IB_QP_CUR_STATE |
1345 IB_QP_ACCESS_FLAGS |
1346 IB_QP_PATH_MIG_STATE),
1347 [IB_QPT_RC] = (IB_QP_CUR_STATE |
1349 IB_QP_ACCESS_FLAGS |
1350 IB_QP_MIN_RNR_TIMER |
1351 IB_QP_PATH_MIG_STATE),
1352 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE |
1354 IB_QP_ACCESS_FLAGS |
1355 IB_QP_PATH_MIG_STATE),
1356 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE |
1358 IB_QP_ACCESS_FLAGS |
1359 IB_QP_MIN_RNR_TIMER |
1360 IB_QP_PATH_MIG_STATE),
1361 [IB_QPT_SMI] = (IB_QP_CUR_STATE |
1363 [IB_QPT_GSI] = (IB_QP_CUR_STATE |
1365 [IB_QPT_RAW_PACKET] = IB_QP_RATE_LIMIT,
1370 [IB_QPS_RESET] = { .valid = 1 },
1371 [IB_QPS_ERR] = { .valid = 1 },
1375 [IB_QPT_UD] = (IB_QP_CUR_STATE |
1377 [IB_QPT_UC] = (IB_QP_CUR_STATE |
1378 IB_QP_ACCESS_FLAGS |
1380 IB_QP_PATH_MIG_STATE),
1381 [IB_QPT_RC] = (IB_QP_CUR_STATE |
1382 IB_QP_ACCESS_FLAGS |
1384 IB_QP_PATH_MIG_STATE |
1385 IB_QP_MIN_RNR_TIMER),
1386 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE |
1387 IB_QP_ACCESS_FLAGS |
1389 IB_QP_PATH_MIG_STATE),
1390 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE |
1391 IB_QP_ACCESS_FLAGS |
1393 IB_QP_PATH_MIG_STATE |
1394 IB_QP_MIN_RNR_TIMER),
1395 [IB_QPT_SMI] = (IB_QP_CUR_STATE |
1397 [IB_QPT_GSI] = (IB_QP_CUR_STATE |
1399 [IB_QPT_RAW_PACKET] = IB_QP_RATE_LIMIT,
1405 [IB_QPT_UD] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1406 [IB_QPT_UC] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1407 [IB_QPT_RC] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1408 [IB_QPT_XRC_INI] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1409 [IB_QPT_XRC_TGT] = IB_QP_EN_SQD_ASYNC_NOTIFY, /* ??? */
1410 [IB_QPT_SMI] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1411 [IB_QPT_GSI] = IB_QP_EN_SQD_ASYNC_NOTIFY
1416 [IB_QPS_RESET] = { .valid = 1 },
1417 [IB_QPS_ERR] = { .valid = 1 },
1421 [IB_QPT_UD] = (IB_QP_CUR_STATE |
1423 [IB_QPT_UC] = (IB_QP_CUR_STATE |
1425 IB_QP_ACCESS_FLAGS |
1426 IB_QP_PATH_MIG_STATE),
1427 [IB_QPT_RC] = (IB_QP_CUR_STATE |
1429 IB_QP_ACCESS_FLAGS |
1430 IB_QP_MIN_RNR_TIMER |
1431 IB_QP_PATH_MIG_STATE),
1432 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE |
1434 IB_QP_ACCESS_FLAGS |
1435 IB_QP_PATH_MIG_STATE),
1436 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE |
1438 IB_QP_ACCESS_FLAGS |
1439 IB_QP_MIN_RNR_TIMER |
1440 IB_QP_PATH_MIG_STATE),
1441 [IB_QPT_SMI] = (IB_QP_CUR_STATE |
1443 [IB_QPT_GSI] = (IB_QP_CUR_STATE |
1450 [IB_QPT_UD] = (IB_QP_PKEY_INDEX |
1452 [IB_QPT_UC] = (IB_QP_AV |
1454 IB_QP_ACCESS_FLAGS |
1456 IB_QP_PATH_MIG_STATE),
1457 [IB_QPT_RC] = (IB_QP_PORT |
1462 IB_QP_MAX_QP_RD_ATOMIC |
1463 IB_QP_MAX_DEST_RD_ATOMIC |
1465 IB_QP_ACCESS_FLAGS |
1467 IB_QP_MIN_RNR_TIMER |
1468 IB_QP_PATH_MIG_STATE),
1469 [IB_QPT_XRC_INI] = (IB_QP_PORT |
1474 IB_QP_MAX_QP_RD_ATOMIC |
1476 IB_QP_ACCESS_FLAGS |
1478 IB_QP_PATH_MIG_STATE),
1479 [IB_QPT_XRC_TGT] = (IB_QP_PORT |
1482 IB_QP_MAX_DEST_RD_ATOMIC |
1484 IB_QP_ACCESS_FLAGS |
1486 IB_QP_MIN_RNR_TIMER |
1487 IB_QP_PATH_MIG_STATE),
1488 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX |
1490 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX |
1496 [IB_QPS_RESET] = { .valid = 1 },
1497 [IB_QPS_ERR] = { .valid = 1 },
1501 [IB_QPT_UD] = (IB_QP_CUR_STATE |
1503 [IB_QPT_UC] = (IB_QP_CUR_STATE |
1504 IB_QP_ACCESS_FLAGS),
1505 [IB_QPT_SMI] = (IB_QP_CUR_STATE |
1507 [IB_QPT_GSI] = (IB_QP_CUR_STATE |
1513 [IB_QPS_RESET] = { .valid = 1 },
1514 [IB_QPS_ERR] = { .valid = 1 }
1518 bool ib_modify_qp_is_ok(enum ib_qp_state cur_state, enum ib_qp_state next_state,
1519 enum ib_qp_type type, enum ib_qp_attr_mask mask,
1520 enum rdma_link_layer ll)
1522 enum ib_qp_attr_mask req_param, opt_param;
1524 if (mask & IB_QP_CUR_STATE &&
1525 cur_state != IB_QPS_RTR && cur_state != IB_QPS_RTS &&
1526 cur_state != IB_QPS_SQD && cur_state != IB_QPS_SQE)
1529 if (!qp_state_table[cur_state][next_state].valid)
1532 req_param = qp_state_table[cur_state][next_state].req_param[type];
1533 opt_param = qp_state_table[cur_state][next_state].opt_param[type];
1535 if ((mask & req_param) != req_param)
1538 if (mask & ~(req_param | opt_param | IB_QP_STATE))
1543 EXPORT_SYMBOL(ib_modify_qp_is_ok);
1546 * ib_resolve_eth_dmac - Resolve destination mac address
1547 * @device: Device to consider
1548 * @ah_attr: address handle attribute which describes the
1549 * source and destination parameters
1550 * ib_resolve_eth_dmac() resolves destination mac address and L3 hop limit It
1551 * returns 0 on success or appropriate error code. It initializes the
1552 * necessary ah_attr fields when call is successful.
1554 static int ib_resolve_eth_dmac(struct ib_device *device,
1555 struct rdma_ah_attr *ah_attr)
1559 if (rdma_is_multicast_addr((struct in6_addr *)ah_attr->grh.dgid.raw)) {
1560 if (ipv6_addr_v4mapped((struct in6_addr *)ah_attr->grh.dgid.raw)) {
1563 memcpy(&addr, ah_attr->grh.dgid.raw + 12, 4);
1564 ip_eth_mc_map(addr, (char *)ah_attr->roce.dmac);
1566 ipv6_eth_mc_map((struct in6_addr *)ah_attr->grh.dgid.raw,
1567 (char *)ah_attr->roce.dmac);
1570 ret = ib_resolve_unicast_gid_dmac(device, ah_attr);
1575 static bool is_qp_type_connected(const struct ib_qp *qp)
1577 return (qp->qp_type == IB_QPT_UC ||
1578 qp->qp_type == IB_QPT_RC ||
1579 qp->qp_type == IB_QPT_XRC_INI ||
1580 qp->qp_type == IB_QPT_XRC_TGT);
1584 * IB core internal function to perform QP attributes modification.
1586 static int _ib_modify_qp(struct ib_qp *qp, struct ib_qp_attr *attr,
1587 int attr_mask, struct ib_udata *udata)
1589 u8 port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
1590 const struct ib_gid_attr *old_sgid_attr_av;
1591 const struct ib_gid_attr *old_sgid_attr_alt_av;
1594 if (attr_mask & IB_QP_AV) {
1595 ret = rdma_fill_sgid_attr(qp->device, &attr->ah_attr,
1600 if (attr_mask & IB_QP_ALT_PATH) {
1602 * FIXME: This does not track the migration state, so if the
1603 * user loads a new alternate path after the HW has migrated
1604 * from primary->alternate we will keep the wrong
1605 * references. This is OK for IB because the reference
1606 * counting does not serve any functional purpose.
1608 ret = rdma_fill_sgid_attr(qp->device, &attr->alt_ah_attr,
1609 &old_sgid_attr_alt_av);
1614 * Today the core code can only handle alternate paths and APM
1615 * for IB. Ban them in roce mode.
1617 if (!(rdma_protocol_ib(qp->device,
1618 attr->alt_ah_attr.port_num) &&
1619 rdma_protocol_ib(qp->device, port))) {
1626 * If the user provided the qp_attr then we have to resolve it. Kernel
1627 * users have to provide already resolved rdma_ah_attr's
1629 if (udata && (attr_mask & IB_QP_AV) &&
1630 attr->ah_attr.type == RDMA_AH_ATTR_TYPE_ROCE &&
1631 is_qp_type_connected(qp)) {
1632 ret = ib_resolve_eth_dmac(qp->device, &attr->ah_attr);
1637 if (rdma_ib_or_roce(qp->device, port)) {
1638 if (attr_mask & IB_QP_RQ_PSN && attr->rq_psn & ~0xffffff) {
1639 pr_warn("%s: %s rq_psn overflow, masking to 24 bits\n",
1640 __func__, qp->device->name);
1641 attr->rq_psn &= 0xffffff;
1644 if (attr_mask & IB_QP_SQ_PSN && attr->sq_psn & ~0xffffff) {
1645 pr_warn("%s: %s sq_psn overflow, masking to 24 bits\n",
1646 __func__, qp->device->name);
1647 attr->sq_psn &= 0xffffff;
1651 ret = ib_security_modify_qp(qp, attr, attr_mask, udata);
1655 if (attr_mask & IB_QP_PORT)
1656 qp->port = attr->port_num;
1657 if (attr_mask & IB_QP_AV)
1659 rdma_update_sgid_attr(&attr->ah_attr, qp->av_sgid_attr);
1660 if (attr_mask & IB_QP_ALT_PATH)
1661 qp->alt_path_sgid_attr = rdma_update_sgid_attr(
1662 &attr->alt_ah_attr, qp->alt_path_sgid_attr);
1665 if (attr_mask & IB_QP_ALT_PATH)
1666 rdma_unfill_sgid_attr(&attr->alt_ah_attr, old_sgid_attr_alt_av);
1668 if (attr_mask & IB_QP_AV)
1669 rdma_unfill_sgid_attr(&attr->ah_attr, old_sgid_attr_av);
1674 * ib_modify_qp_with_udata - Modifies the attributes for the specified QP.
1675 * @ib_qp: The QP to modify.
1676 * @attr: On input, specifies the QP attributes to modify. On output,
1677 * the current values of selected QP attributes are returned.
1678 * @attr_mask: A bit-mask used to specify which attributes of the QP
1679 * are being modified.
1680 * @udata: pointer to user's input output buffer information
1681 * are being modified.
1682 * It returns 0 on success and returns appropriate error code on error.
1684 int ib_modify_qp_with_udata(struct ib_qp *ib_qp, struct ib_qp_attr *attr,
1685 int attr_mask, struct ib_udata *udata)
1687 return _ib_modify_qp(ib_qp->real_qp, attr, attr_mask, udata);
1689 EXPORT_SYMBOL(ib_modify_qp_with_udata);
1691 int ib_get_eth_speed(struct ib_device *dev, u8 port_num, u8 *speed, u8 *width)
1695 struct net_device *netdev;
1696 struct ethtool_link_ksettings lksettings;
1698 if (rdma_port_get_link_layer(dev, port_num) != IB_LINK_LAYER_ETHERNET)
1701 if (!dev->get_netdev)
1704 netdev = dev->get_netdev(dev, port_num);
1709 rc = __ethtool_get_link_ksettings(netdev, &lksettings);
1714 if (!rc && lksettings.base.speed != (u32)SPEED_UNKNOWN) {
1715 netdev_speed = lksettings.base.speed;
1717 netdev_speed = SPEED_1000;
1718 pr_warn("%s speed is unknown, defaulting to %d\n", netdev->name,
1722 if (netdev_speed <= SPEED_1000) {
1723 *width = IB_WIDTH_1X;
1724 *speed = IB_SPEED_SDR;
1725 } else if (netdev_speed <= SPEED_10000) {
1726 *width = IB_WIDTH_1X;
1727 *speed = IB_SPEED_FDR10;
1728 } else if (netdev_speed <= SPEED_20000) {
1729 *width = IB_WIDTH_4X;
1730 *speed = IB_SPEED_DDR;
1731 } else if (netdev_speed <= SPEED_25000) {
1732 *width = IB_WIDTH_1X;
1733 *speed = IB_SPEED_EDR;
1734 } else if (netdev_speed <= SPEED_40000) {
1735 *width = IB_WIDTH_4X;
1736 *speed = IB_SPEED_FDR10;
1738 *width = IB_WIDTH_4X;
1739 *speed = IB_SPEED_EDR;
1744 EXPORT_SYMBOL(ib_get_eth_speed);
1746 int ib_modify_qp(struct ib_qp *qp,
1747 struct ib_qp_attr *qp_attr,
1750 return _ib_modify_qp(qp->real_qp, qp_attr, qp_attr_mask, NULL);
1752 EXPORT_SYMBOL(ib_modify_qp);
1754 int ib_query_qp(struct ib_qp *qp,
1755 struct ib_qp_attr *qp_attr,
1757 struct ib_qp_init_attr *qp_init_attr)
1759 qp_attr->ah_attr.grh.sgid_attr = NULL;
1760 qp_attr->alt_ah_attr.grh.sgid_attr = NULL;
1762 return qp->device->query_qp ?
1763 qp->device->query_qp(qp->real_qp, qp_attr, qp_attr_mask, qp_init_attr) :
1766 EXPORT_SYMBOL(ib_query_qp);
1768 int ib_close_qp(struct ib_qp *qp)
1770 struct ib_qp *real_qp;
1771 unsigned long flags;
1773 real_qp = qp->real_qp;
1777 spin_lock_irqsave(&real_qp->device->event_handler_lock, flags);
1778 list_del(&qp->open_list);
1779 spin_unlock_irqrestore(&real_qp->device->event_handler_lock, flags);
1781 atomic_dec(&real_qp->usecnt);
1783 ib_close_shared_qp_security(qp->qp_sec);
1788 EXPORT_SYMBOL(ib_close_qp);
1790 static int __ib_destroy_shared_qp(struct ib_qp *qp)
1792 struct ib_xrcd *xrcd;
1793 struct ib_qp *real_qp;
1796 real_qp = qp->real_qp;
1797 xrcd = real_qp->xrcd;
1799 mutex_lock(&xrcd->tgt_qp_mutex);
1801 if (atomic_read(&real_qp->usecnt) == 0)
1802 list_del(&real_qp->xrcd_list);
1805 mutex_unlock(&xrcd->tgt_qp_mutex);
1808 ret = ib_destroy_qp(real_qp);
1810 atomic_dec(&xrcd->usecnt);
1812 __ib_insert_xrcd_qp(xrcd, real_qp);
1818 int ib_destroy_qp(struct ib_qp *qp)
1820 const struct ib_gid_attr *alt_path_sgid_attr = qp->alt_path_sgid_attr;
1821 const struct ib_gid_attr *av_sgid_attr = qp->av_sgid_attr;
1823 struct ib_cq *scq, *rcq;
1825 struct ib_rwq_ind_table *ind_tbl;
1826 struct ib_qp_security *sec;
1829 WARN_ON_ONCE(qp->mrs_used > 0);
1831 if (atomic_read(&qp->usecnt))
1834 if (qp->real_qp != qp)
1835 return __ib_destroy_shared_qp(qp);
1841 ind_tbl = qp->rwq_ind_tbl;
1844 ib_destroy_qp_security_begin(sec);
1847 rdma_rw_cleanup_mrs(qp);
1849 rdma_restrack_del(&qp->res);
1850 ret = qp->device->destroy_qp(qp);
1852 if (alt_path_sgid_attr)
1853 rdma_put_gid_attr(alt_path_sgid_attr);
1855 rdma_put_gid_attr(av_sgid_attr);
1857 atomic_dec(&pd->usecnt);
1859 atomic_dec(&scq->usecnt);
1861 atomic_dec(&rcq->usecnt);
1863 atomic_dec(&srq->usecnt);
1865 atomic_dec(&ind_tbl->usecnt);
1867 ib_destroy_qp_security_end(sec);
1870 ib_destroy_qp_security_abort(sec);
1875 EXPORT_SYMBOL(ib_destroy_qp);
1877 /* Completion queues */
1879 struct ib_cq *__ib_create_cq(struct ib_device *device,
1880 ib_comp_handler comp_handler,
1881 void (*event_handler)(struct ib_event *, void *),
1883 const struct ib_cq_init_attr *cq_attr,
1888 cq = device->create_cq(device, cq_attr, NULL, NULL);
1891 cq->device = device;
1893 cq->comp_handler = comp_handler;
1894 cq->event_handler = event_handler;
1895 cq->cq_context = cq_context;
1896 atomic_set(&cq->usecnt, 0);
1897 cq->res.type = RDMA_RESTRACK_CQ;
1898 cq->res.kern_name = caller;
1899 rdma_restrack_add(&cq->res);
1904 EXPORT_SYMBOL(__ib_create_cq);
1906 int rdma_set_cq_moderation(struct ib_cq *cq, u16 cq_count, u16 cq_period)
1908 return cq->device->modify_cq ?
1909 cq->device->modify_cq(cq, cq_count, cq_period) : -EOPNOTSUPP;
1911 EXPORT_SYMBOL(rdma_set_cq_moderation);
1913 int ib_destroy_cq(struct ib_cq *cq)
1915 if (atomic_read(&cq->usecnt))
1918 rdma_restrack_del(&cq->res);
1919 return cq->device->destroy_cq(cq);
1921 EXPORT_SYMBOL(ib_destroy_cq);
1923 int ib_resize_cq(struct ib_cq *cq, int cqe)
1925 return cq->device->resize_cq ?
1926 cq->device->resize_cq(cq, cqe, NULL) : -EOPNOTSUPP;
1928 EXPORT_SYMBOL(ib_resize_cq);
1930 /* Memory regions */
1932 int ib_dereg_mr(struct ib_mr *mr)
1934 struct ib_pd *pd = mr->pd;
1935 struct ib_dm *dm = mr->dm;
1938 rdma_restrack_del(&mr->res);
1939 ret = mr->device->dereg_mr(mr);
1941 atomic_dec(&pd->usecnt);
1943 atomic_dec(&dm->usecnt);
1948 EXPORT_SYMBOL(ib_dereg_mr);
1951 * ib_alloc_mr() - Allocates a memory region
1952 * @pd: protection domain associated with the region
1953 * @mr_type: memory region type
1954 * @max_num_sg: maximum sg entries available for registration.
1957 * Memory registeration page/sg lists must not exceed max_num_sg.
1958 * For mr_type IB_MR_TYPE_MEM_REG, the total length cannot exceed
1959 * max_num_sg * used_page_size.
1962 struct ib_mr *ib_alloc_mr(struct ib_pd *pd,
1963 enum ib_mr_type mr_type,
1968 if (!pd->device->alloc_mr)
1969 return ERR_PTR(-EOPNOTSUPP);
1971 mr = pd->device->alloc_mr(pd, mr_type, max_num_sg);
1973 mr->device = pd->device;
1977 atomic_inc(&pd->usecnt);
1978 mr->need_inval = false;
1979 mr->res.type = RDMA_RESTRACK_MR;
1980 rdma_restrack_add(&mr->res);
1985 EXPORT_SYMBOL(ib_alloc_mr);
1987 /* "Fast" memory regions */
1989 struct ib_fmr *ib_alloc_fmr(struct ib_pd *pd,
1990 int mr_access_flags,
1991 struct ib_fmr_attr *fmr_attr)
1995 if (!pd->device->alloc_fmr)
1996 return ERR_PTR(-EOPNOTSUPP);
1998 fmr = pd->device->alloc_fmr(pd, mr_access_flags, fmr_attr);
2000 fmr->device = pd->device;
2002 atomic_inc(&pd->usecnt);
2007 EXPORT_SYMBOL(ib_alloc_fmr);
2009 int ib_unmap_fmr(struct list_head *fmr_list)
2013 if (list_empty(fmr_list))
2016 fmr = list_entry(fmr_list->next, struct ib_fmr, list);
2017 return fmr->device->unmap_fmr(fmr_list);
2019 EXPORT_SYMBOL(ib_unmap_fmr);
2021 int ib_dealloc_fmr(struct ib_fmr *fmr)
2027 ret = fmr->device->dealloc_fmr(fmr);
2029 atomic_dec(&pd->usecnt);
2033 EXPORT_SYMBOL(ib_dealloc_fmr);
2035 /* Multicast groups */
2037 static bool is_valid_mcast_lid(struct ib_qp *qp, u16 lid)
2039 struct ib_qp_init_attr init_attr = {};
2040 struct ib_qp_attr attr = {};
2041 int num_eth_ports = 0;
2044 /* If QP state >= init, it is assigned to a port and we can check this
2047 if (!ib_query_qp(qp, &attr, IB_QP_STATE | IB_QP_PORT, &init_attr)) {
2048 if (attr.qp_state >= IB_QPS_INIT) {
2049 if (rdma_port_get_link_layer(qp->device, attr.port_num) !=
2050 IB_LINK_LAYER_INFINIBAND)
2056 /* Can't get a quick answer, iterate over all ports */
2057 for (port = 0; port < qp->device->phys_port_cnt; port++)
2058 if (rdma_port_get_link_layer(qp->device, port) !=
2059 IB_LINK_LAYER_INFINIBAND)
2062 /* If we have at lease one Ethernet port, RoCE annex declares that
2063 * multicast LID should be ignored. We can't tell at this step if the
2064 * QP belongs to an IB or Ethernet port.
2069 /* If all the ports are IB, we can check according to IB spec. */
2071 return !(lid < be16_to_cpu(IB_MULTICAST_LID_BASE) ||
2072 lid == be16_to_cpu(IB_LID_PERMISSIVE));
2075 int ib_attach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid)
2079 if (!qp->device->attach_mcast)
2082 if (!rdma_is_multicast_addr((struct in6_addr *)gid->raw) ||
2083 qp->qp_type != IB_QPT_UD || !is_valid_mcast_lid(qp, lid))
2086 ret = qp->device->attach_mcast(qp, gid, lid);
2088 atomic_inc(&qp->usecnt);
2091 EXPORT_SYMBOL(ib_attach_mcast);
2093 int ib_detach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid)
2097 if (!qp->device->detach_mcast)
2100 if (!rdma_is_multicast_addr((struct in6_addr *)gid->raw) ||
2101 qp->qp_type != IB_QPT_UD || !is_valid_mcast_lid(qp, lid))
2104 ret = qp->device->detach_mcast(qp, gid, lid);
2106 atomic_dec(&qp->usecnt);
2109 EXPORT_SYMBOL(ib_detach_mcast);
2111 struct ib_xrcd *__ib_alloc_xrcd(struct ib_device *device, const char *caller)
2113 struct ib_xrcd *xrcd;
2115 if (!device->alloc_xrcd)
2116 return ERR_PTR(-EOPNOTSUPP);
2118 xrcd = device->alloc_xrcd(device, NULL, NULL);
2119 if (!IS_ERR(xrcd)) {
2120 xrcd->device = device;
2122 atomic_set(&xrcd->usecnt, 0);
2123 mutex_init(&xrcd->tgt_qp_mutex);
2124 INIT_LIST_HEAD(&xrcd->tgt_qp_list);
2129 EXPORT_SYMBOL(__ib_alloc_xrcd);
2131 int ib_dealloc_xrcd(struct ib_xrcd *xrcd)
2136 if (atomic_read(&xrcd->usecnt))
2139 while (!list_empty(&xrcd->tgt_qp_list)) {
2140 qp = list_entry(xrcd->tgt_qp_list.next, struct ib_qp, xrcd_list);
2141 ret = ib_destroy_qp(qp);
2146 return xrcd->device->dealloc_xrcd(xrcd);
2148 EXPORT_SYMBOL(ib_dealloc_xrcd);
2151 * ib_create_wq - Creates a WQ associated with the specified protection
2153 * @pd: The protection domain associated with the WQ.
2154 * @wq_attr: A list of initial attributes required to create the
2155 * WQ. If WQ creation succeeds, then the attributes are updated to
2156 * the actual capabilities of the created WQ.
2158 * wq_attr->max_wr and wq_attr->max_sge determine
2159 * the requested size of the WQ, and set to the actual values allocated
2161 * If ib_create_wq() succeeds, then max_wr and max_sge will always be
2162 * at least as large as the requested values.
2164 struct ib_wq *ib_create_wq(struct ib_pd *pd,
2165 struct ib_wq_init_attr *wq_attr)
2169 if (!pd->device->create_wq)
2170 return ERR_PTR(-EOPNOTSUPP);
2172 wq = pd->device->create_wq(pd, wq_attr, NULL);
2174 wq->event_handler = wq_attr->event_handler;
2175 wq->wq_context = wq_attr->wq_context;
2176 wq->wq_type = wq_attr->wq_type;
2177 wq->cq = wq_attr->cq;
2178 wq->device = pd->device;
2181 atomic_inc(&pd->usecnt);
2182 atomic_inc(&wq_attr->cq->usecnt);
2183 atomic_set(&wq->usecnt, 0);
2187 EXPORT_SYMBOL(ib_create_wq);
2190 * ib_destroy_wq - Destroys the specified WQ.
2191 * @wq: The WQ to destroy.
2193 int ib_destroy_wq(struct ib_wq *wq)
2196 struct ib_cq *cq = wq->cq;
2197 struct ib_pd *pd = wq->pd;
2199 if (atomic_read(&wq->usecnt))
2202 err = wq->device->destroy_wq(wq);
2204 atomic_dec(&pd->usecnt);
2205 atomic_dec(&cq->usecnt);
2209 EXPORT_SYMBOL(ib_destroy_wq);
2212 * ib_modify_wq - Modifies the specified WQ.
2213 * @wq: The WQ to modify.
2214 * @wq_attr: On input, specifies the WQ attributes to modify.
2215 * @wq_attr_mask: A bit-mask used to specify which attributes of the WQ
2216 * are being modified.
2217 * On output, the current values of selected WQ attributes are returned.
2219 int ib_modify_wq(struct ib_wq *wq, struct ib_wq_attr *wq_attr,
2224 if (!wq->device->modify_wq)
2227 err = wq->device->modify_wq(wq, wq_attr, wq_attr_mask, NULL);
2230 EXPORT_SYMBOL(ib_modify_wq);
2233 * ib_create_rwq_ind_table - Creates a RQ Indirection Table.
2234 * @device: The device on which to create the rwq indirection table.
2235 * @ib_rwq_ind_table_init_attr: A list of initial attributes required to
2236 * create the Indirection Table.
2238 * Note: The life time of ib_rwq_ind_table_init_attr->ind_tbl is not less
2239 * than the created ib_rwq_ind_table object and the caller is responsible
2240 * for its memory allocation/free.
2242 struct ib_rwq_ind_table *ib_create_rwq_ind_table(struct ib_device *device,
2243 struct ib_rwq_ind_table_init_attr *init_attr)
2245 struct ib_rwq_ind_table *rwq_ind_table;
2249 if (!device->create_rwq_ind_table)
2250 return ERR_PTR(-EOPNOTSUPP);
2252 table_size = (1 << init_attr->log_ind_tbl_size);
2253 rwq_ind_table = device->create_rwq_ind_table(device,
2255 if (IS_ERR(rwq_ind_table))
2256 return rwq_ind_table;
2258 rwq_ind_table->ind_tbl = init_attr->ind_tbl;
2259 rwq_ind_table->log_ind_tbl_size = init_attr->log_ind_tbl_size;
2260 rwq_ind_table->device = device;
2261 rwq_ind_table->uobject = NULL;
2262 atomic_set(&rwq_ind_table->usecnt, 0);
2264 for (i = 0; i < table_size; i++)
2265 atomic_inc(&rwq_ind_table->ind_tbl[i]->usecnt);
2267 return rwq_ind_table;
2269 EXPORT_SYMBOL(ib_create_rwq_ind_table);
2272 * ib_destroy_rwq_ind_table - Destroys the specified Indirection Table.
2273 * @wq_ind_table: The Indirection Table to destroy.
2275 int ib_destroy_rwq_ind_table(struct ib_rwq_ind_table *rwq_ind_table)
2278 u32 table_size = (1 << rwq_ind_table->log_ind_tbl_size);
2279 struct ib_wq **ind_tbl = rwq_ind_table->ind_tbl;
2281 if (atomic_read(&rwq_ind_table->usecnt))
2284 err = rwq_ind_table->device->destroy_rwq_ind_table(rwq_ind_table);
2286 for (i = 0; i < table_size; i++)
2287 atomic_dec(&ind_tbl[i]->usecnt);
2292 EXPORT_SYMBOL(ib_destroy_rwq_ind_table);
2294 int ib_check_mr_status(struct ib_mr *mr, u32 check_mask,
2295 struct ib_mr_status *mr_status)
2297 return mr->device->check_mr_status ?
2298 mr->device->check_mr_status(mr, check_mask, mr_status) : -EOPNOTSUPP;
2300 EXPORT_SYMBOL(ib_check_mr_status);
2302 int ib_set_vf_link_state(struct ib_device *device, int vf, u8 port,
2305 if (!device->set_vf_link_state)
2308 return device->set_vf_link_state(device, vf, port, state);
2310 EXPORT_SYMBOL(ib_set_vf_link_state);
2312 int ib_get_vf_config(struct ib_device *device, int vf, u8 port,
2313 struct ifla_vf_info *info)
2315 if (!device->get_vf_config)
2318 return device->get_vf_config(device, vf, port, info);
2320 EXPORT_SYMBOL(ib_get_vf_config);
2322 int ib_get_vf_stats(struct ib_device *device, int vf, u8 port,
2323 struct ifla_vf_stats *stats)
2325 if (!device->get_vf_stats)
2328 return device->get_vf_stats(device, vf, port, stats);
2330 EXPORT_SYMBOL(ib_get_vf_stats);
2332 int ib_set_vf_guid(struct ib_device *device, int vf, u8 port, u64 guid,
2335 if (!device->set_vf_guid)
2338 return device->set_vf_guid(device, vf, port, guid, type);
2340 EXPORT_SYMBOL(ib_set_vf_guid);
2343 * ib_map_mr_sg() - Map the largest prefix of a dma mapped SG list
2344 * and set it the memory region.
2345 * @mr: memory region
2346 * @sg: dma mapped scatterlist
2347 * @sg_nents: number of entries in sg
2348 * @sg_offset: offset in bytes into sg
2349 * @page_size: page vector desired page size
2352 * - The first sg element is allowed to have an offset.
2353 * - Each sg element must either be aligned to page_size or virtually
2354 * contiguous to the previous element. In case an sg element has a
2355 * non-contiguous offset, the mapping prefix will not include it.
2356 * - The last sg element is allowed to have length less than page_size.
2357 * - If sg_nents total byte length exceeds the mr max_num_sge * page_size
2358 * then only max_num_sg entries will be mapped.
2359 * - If the MR was allocated with type IB_MR_TYPE_SG_GAPS, none of these
2360 * constraints holds and the page_size argument is ignored.
2362 * Returns the number of sg elements that were mapped to the memory region.
2364 * After this completes successfully, the memory region
2365 * is ready for registration.
2367 int ib_map_mr_sg(struct ib_mr *mr, struct scatterlist *sg, int sg_nents,
2368 unsigned int *sg_offset, unsigned int page_size)
2370 if (unlikely(!mr->device->map_mr_sg))
2373 mr->page_size = page_size;
2375 return mr->device->map_mr_sg(mr, sg, sg_nents, sg_offset);
2377 EXPORT_SYMBOL(ib_map_mr_sg);
2380 * ib_sg_to_pages() - Convert the largest prefix of a sg list
2382 * @mr: memory region
2383 * @sgl: dma mapped scatterlist
2384 * @sg_nents: number of entries in sg
2385 * @sg_offset_p: IN: start offset in bytes into sg
2386 * OUT: offset in bytes for element n of the sg of the first
2387 * byte that has not been processed where n is the return
2388 * value of this function.
2389 * @set_page: driver page assignment function pointer
2391 * Core service helper for drivers to convert the largest
2392 * prefix of given sg list to a page vector. The sg list
2393 * prefix converted is the prefix that meet the requirements
2396 * Returns the number of sg elements that were assigned to
2399 int ib_sg_to_pages(struct ib_mr *mr, struct scatterlist *sgl, int sg_nents,
2400 unsigned int *sg_offset_p, int (*set_page)(struct ib_mr *, u64))
2402 struct scatterlist *sg;
2403 u64 last_end_dma_addr = 0;
2404 unsigned int sg_offset = sg_offset_p ? *sg_offset_p : 0;
2405 unsigned int last_page_off = 0;
2406 u64 page_mask = ~((u64)mr->page_size - 1);
2409 if (unlikely(sg_nents <= 0 || sg_offset > sg_dma_len(&sgl[0])))
2412 mr->iova = sg_dma_address(&sgl[0]) + sg_offset;
2415 for_each_sg(sgl, sg, sg_nents, i) {
2416 u64 dma_addr = sg_dma_address(sg) + sg_offset;
2417 u64 prev_addr = dma_addr;
2418 unsigned int dma_len = sg_dma_len(sg) - sg_offset;
2419 u64 end_dma_addr = dma_addr + dma_len;
2420 u64 page_addr = dma_addr & page_mask;
2423 * For the second and later elements, check whether either the
2424 * end of element i-1 or the start of element i is not aligned
2425 * on a page boundary.
2427 if (i && (last_page_off != 0 || page_addr != dma_addr)) {
2428 /* Stop mapping if there is a gap. */
2429 if (last_end_dma_addr != dma_addr)
2433 * Coalesce this element with the last. If it is small
2434 * enough just update mr->length. Otherwise start
2435 * mapping from the next page.
2441 ret = set_page(mr, page_addr);
2442 if (unlikely(ret < 0)) {
2443 sg_offset = prev_addr - sg_dma_address(sg);
2444 mr->length += prev_addr - dma_addr;
2446 *sg_offset_p = sg_offset;
2447 return i || sg_offset ? i : ret;
2449 prev_addr = page_addr;
2451 page_addr += mr->page_size;
2452 } while (page_addr < end_dma_addr);
2454 mr->length += dma_len;
2455 last_end_dma_addr = end_dma_addr;
2456 last_page_off = end_dma_addr & ~page_mask;
2465 EXPORT_SYMBOL(ib_sg_to_pages);
2467 struct ib_drain_cqe {
2469 struct completion done;
2472 static void ib_drain_qp_done(struct ib_cq *cq, struct ib_wc *wc)
2474 struct ib_drain_cqe *cqe = container_of(wc->wr_cqe, struct ib_drain_cqe,
2477 complete(&cqe->done);
2481 * Post a WR and block until its completion is reaped for the SQ.
2483 static void __ib_drain_sq(struct ib_qp *qp)
2485 struct ib_cq *cq = qp->send_cq;
2486 struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
2487 struct ib_drain_cqe sdrain;
2488 struct ib_rdma_wr swr = {
2491 { .wr_cqe = &sdrain.cqe, },
2492 .opcode = IB_WR_RDMA_WRITE,
2497 ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
2499 WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
2503 sdrain.cqe.done = ib_drain_qp_done;
2504 init_completion(&sdrain.done);
2506 ret = ib_post_send(qp, &swr.wr, NULL);
2508 WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
2512 if (cq->poll_ctx == IB_POLL_DIRECT)
2513 while (wait_for_completion_timeout(&sdrain.done, HZ / 10) <= 0)
2514 ib_process_cq_direct(cq, -1);
2516 wait_for_completion(&sdrain.done);
2520 * Post a WR and block until its completion is reaped for the RQ.
2522 static void __ib_drain_rq(struct ib_qp *qp)
2524 struct ib_cq *cq = qp->recv_cq;
2525 struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
2526 struct ib_drain_cqe rdrain;
2527 struct ib_recv_wr rwr = {};
2530 ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
2532 WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
2536 rwr.wr_cqe = &rdrain.cqe;
2537 rdrain.cqe.done = ib_drain_qp_done;
2538 init_completion(&rdrain.done);
2540 ret = ib_post_recv(qp, &rwr, NULL);
2542 WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
2546 if (cq->poll_ctx == IB_POLL_DIRECT)
2547 while (wait_for_completion_timeout(&rdrain.done, HZ / 10) <= 0)
2548 ib_process_cq_direct(cq, -1);
2550 wait_for_completion(&rdrain.done);
2554 * ib_drain_sq() - Block until all SQ CQEs have been consumed by the
2556 * @qp: queue pair to drain
2558 * If the device has a provider-specific drain function, then
2559 * call that. Otherwise call the generic drain function
2564 * ensure there is room in the CQ and SQ for the drain work request and
2567 * allocate the CQ using ib_alloc_cq().
2569 * ensure that there are no other contexts that are posting WRs concurrently.
2570 * Otherwise the drain is not guaranteed.
2572 void ib_drain_sq(struct ib_qp *qp)
2574 if (qp->device->drain_sq)
2575 qp->device->drain_sq(qp);
2579 EXPORT_SYMBOL(ib_drain_sq);
2582 * ib_drain_rq() - Block until all RQ CQEs have been consumed by the
2584 * @qp: queue pair to drain
2586 * If the device has a provider-specific drain function, then
2587 * call that. Otherwise call the generic drain function
2592 * ensure there is room in the CQ and RQ for the drain work request and
2595 * allocate the CQ using ib_alloc_cq().
2597 * ensure that there are no other contexts that are posting WRs concurrently.
2598 * Otherwise the drain is not guaranteed.
2600 void ib_drain_rq(struct ib_qp *qp)
2602 if (qp->device->drain_rq)
2603 qp->device->drain_rq(qp);
2607 EXPORT_SYMBOL(ib_drain_rq);
2610 * ib_drain_qp() - Block until all CQEs have been consumed by the
2611 * application on both the RQ and SQ.
2612 * @qp: queue pair to drain
2616 * ensure there is room in the CQ(s), SQ, and RQ for drain work requests
2619 * allocate the CQs using ib_alloc_cq().
2621 * ensure that there are no other contexts that are posting WRs concurrently.
2622 * Otherwise the drain is not guaranteed.
2624 void ib_drain_qp(struct ib_qp *qp)
2630 EXPORT_SYMBOL(ib_drain_qp);