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;
144 case IB_RATE_28_GBPS: return 11;
145 case IB_RATE_50_GBPS: return 20;
146 case IB_RATE_400_GBPS: return 160;
147 case IB_RATE_600_GBPS: return 240;
151 EXPORT_SYMBOL(ib_rate_to_mult);
153 __attribute_const__ enum ib_rate mult_to_ib_rate(int mult)
156 case 1: return IB_RATE_2_5_GBPS;
157 case 2: return IB_RATE_5_GBPS;
158 case 4: return IB_RATE_10_GBPS;
159 case 8: return IB_RATE_20_GBPS;
160 case 12: return IB_RATE_30_GBPS;
161 case 16: return IB_RATE_40_GBPS;
162 case 24: return IB_RATE_60_GBPS;
163 case 32: return IB_RATE_80_GBPS;
164 case 48: return IB_RATE_120_GBPS;
165 case 6: return IB_RATE_14_GBPS;
166 case 22: return IB_RATE_56_GBPS;
167 case 45: return IB_RATE_112_GBPS;
168 case 67: return IB_RATE_168_GBPS;
169 case 10: return IB_RATE_25_GBPS;
170 case 40: return IB_RATE_100_GBPS;
171 case 80: return IB_RATE_200_GBPS;
172 case 120: return IB_RATE_300_GBPS;
173 case 11: return IB_RATE_28_GBPS;
174 case 20: return IB_RATE_50_GBPS;
175 case 160: return IB_RATE_400_GBPS;
176 case 240: return IB_RATE_600_GBPS;
177 default: return IB_RATE_PORT_CURRENT;
180 EXPORT_SYMBOL(mult_to_ib_rate);
182 __attribute_const__ int ib_rate_to_mbps(enum ib_rate rate)
185 case IB_RATE_2_5_GBPS: return 2500;
186 case IB_RATE_5_GBPS: return 5000;
187 case IB_RATE_10_GBPS: return 10000;
188 case IB_RATE_20_GBPS: return 20000;
189 case IB_RATE_30_GBPS: return 30000;
190 case IB_RATE_40_GBPS: return 40000;
191 case IB_RATE_60_GBPS: return 60000;
192 case IB_RATE_80_GBPS: return 80000;
193 case IB_RATE_120_GBPS: return 120000;
194 case IB_RATE_14_GBPS: return 14062;
195 case IB_RATE_56_GBPS: return 56250;
196 case IB_RATE_112_GBPS: return 112500;
197 case IB_RATE_168_GBPS: return 168750;
198 case IB_RATE_25_GBPS: return 25781;
199 case IB_RATE_100_GBPS: return 103125;
200 case IB_RATE_200_GBPS: return 206250;
201 case IB_RATE_300_GBPS: return 309375;
202 case IB_RATE_28_GBPS: return 28125;
203 case IB_RATE_50_GBPS: return 53125;
204 case IB_RATE_400_GBPS: return 425000;
205 case IB_RATE_600_GBPS: return 637500;
209 EXPORT_SYMBOL(ib_rate_to_mbps);
211 __attribute_const__ enum rdma_transport_type
212 rdma_node_get_transport(enum rdma_node_type node_type)
215 if (node_type == RDMA_NODE_USNIC)
216 return RDMA_TRANSPORT_USNIC;
217 if (node_type == RDMA_NODE_USNIC_UDP)
218 return RDMA_TRANSPORT_USNIC_UDP;
219 if (node_type == RDMA_NODE_RNIC)
220 return RDMA_TRANSPORT_IWARP;
222 return RDMA_TRANSPORT_IB;
224 EXPORT_SYMBOL(rdma_node_get_transport);
226 enum rdma_link_layer rdma_port_get_link_layer(struct ib_device *device, u8 port_num)
228 enum rdma_transport_type lt;
229 if (device->ops.get_link_layer)
230 return device->ops.get_link_layer(device, port_num);
232 lt = rdma_node_get_transport(device->node_type);
233 if (lt == RDMA_TRANSPORT_IB)
234 return IB_LINK_LAYER_INFINIBAND;
236 return IB_LINK_LAYER_ETHERNET;
238 EXPORT_SYMBOL(rdma_port_get_link_layer);
240 /* Protection domains */
243 * ib_alloc_pd - Allocates an unused protection domain.
244 * @device: The device on which to allocate the protection domain.
246 * A protection domain object provides an association between QPs, shared
247 * receive queues, address handles, memory regions, and memory windows.
249 * Every PD has a local_dma_lkey which can be used as the lkey value for local
252 struct ib_pd *__ib_alloc_pd(struct ib_device *device, unsigned int flags,
256 int mr_access_flags = 0;
258 pd = device->ops.alloc_pd(device, NULL, NULL);
264 pd->__internal_mr = NULL;
265 atomic_set(&pd->usecnt, 0);
268 if (device->attrs.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY)
269 pd->local_dma_lkey = device->local_dma_lkey;
271 mr_access_flags |= IB_ACCESS_LOCAL_WRITE;
273 if (flags & IB_PD_UNSAFE_GLOBAL_RKEY) {
274 pr_warn("%s: enabling unsafe global rkey\n", caller);
275 mr_access_flags |= IB_ACCESS_REMOTE_READ | IB_ACCESS_REMOTE_WRITE;
278 pd->res.type = RDMA_RESTRACK_PD;
279 rdma_restrack_set_task(&pd->res, caller);
280 rdma_restrack_kadd(&pd->res);
282 if (mr_access_flags) {
285 mr = pd->device->ops.get_dma_mr(pd, mr_access_flags);
291 mr->device = pd->device;
294 mr->need_inval = false;
296 pd->__internal_mr = mr;
298 if (!(device->attrs.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY))
299 pd->local_dma_lkey = pd->__internal_mr->lkey;
301 if (flags & IB_PD_UNSAFE_GLOBAL_RKEY)
302 pd->unsafe_global_rkey = pd->__internal_mr->rkey;
307 EXPORT_SYMBOL(__ib_alloc_pd);
310 * ib_dealloc_pd - Deallocates a protection domain.
311 * @pd: The protection domain to deallocate.
313 * It is an error to call this function while any resources in the pd still
314 * exist. The caller is responsible to synchronously destroy them and
315 * guarantee no new allocations will happen.
317 void ib_dealloc_pd(struct ib_pd *pd)
321 if (pd->__internal_mr) {
322 ret = pd->device->ops.dereg_mr(pd->__internal_mr);
324 pd->__internal_mr = NULL;
327 /* uverbs manipulates usecnt with proper locking, while the kabi
328 requires the caller to guarantee we can't race here. */
329 WARN_ON(atomic_read(&pd->usecnt));
331 rdma_restrack_del(&pd->res);
332 /* Making delalloc_pd a void return is a WIP, no driver should return
334 ret = pd->device->ops.dealloc_pd(pd);
335 WARN_ONCE(ret, "Infiniband HW driver failed dealloc_pd");
337 EXPORT_SYMBOL(ib_dealloc_pd);
339 /* Address handles */
342 * rdma_copy_ah_attr - Copy rdma ah attribute from source to destination.
343 * @dest: Pointer to destination ah_attr. Contents of the destination
344 * pointer is assumed to be invalid and attribute are overwritten.
345 * @src: Pointer to source ah_attr.
347 void rdma_copy_ah_attr(struct rdma_ah_attr *dest,
348 const struct rdma_ah_attr *src)
351 if (dest->grh.sgid_attr)
352 rdma_hold_gid_attr(dest->grh.sgid_attr);
354 EXPORT_SYMBOL(rdma_copy_ah_attr);
357 * rdma_replace_ah_attr - Replace valid ah_attr with new new one.
358 * @old: Pointer to existing ah_attr which needs to be replaced.
359 * old is assumed to be valid or zero'd
360 * @new: Pointer to the new ah_attr.
362 * rdma_replace_ah_attr() first releases any reference in the old ah_attr if
363 * old the ah_attr is valid; after that it copies the new attribute and holds
364 * the reference to the replaced ah_attr.
366 void rdma_replace_ah_attr(struct rdma_ah_attr *old,
367 const struct rdma_ah_attr *new)
369 rdma_destroy_ah_attr(old);
371 if (old->grh.sgid_attr)
372 rdma_hold_gid_attr(old->grh.sgid_attr);
374 EXPORT_SYMBOL(rdma_replace_ah_attr);
377 * rdma_move_ah_attr - Move ah_attr pointed by source to destination.
378 * @dest: Pointer to destination ah_attr to copy to.
379 * dest is assumed to be valid or zero'd
380 * @src: Pointer to the new ah_attr.
382 * rdma_move_ah_attr() first releases any reference in the destination ah_attr
383 * if it is valid. This also transfers ownership of internal references from
384 * src to dest, making src invalid in the process. No new reference of the src
387 void rdma_move_ah_attr(struct rdma_ah_attr *dest, struct rdma_ah_attr *src)
389 rdma_destroy_ah_attr(dest);
391 src->grh.sgid_attr = NULL;
393 EXPORT_SYMBOL(rdma_move_ah_attr);
396 * Validate that the rdma_ah_attr is valid for the device before passing it
399 static int rdma_check_ah_attr(struct ib_device *device,
400 struct rdma_ah_attr *ah_attr)
402 if (!rdma_is_port_valid(device, ah_attr->port_num))
405 if ((rdma_is_grh_required(device, ah_attr->port_num) ||
406 ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE) &&
407 !(ah_attr->ah_flags & IB_AH_GRH))
410 if (ah_attr->grh.sgid_attr) {
412 * Make sure the passed sgid_attr is consistent with the
415 if (ah_attr->grh.sgid_attr->index != ah_attr->grh.sgid_index ||
416 ah_attr->grh.sgid_attr->port_num != ah_attr->port_num)
423 * If the ah requires a GRH then ensure that sgid_attr pointer is filled in.
424 * On success the caller is responsible to call rdma_unfill_sgid_attr().
426 static int rdma_fill_sgid_attr(struct ib_device *device,
427 struct rdma_ah_attr *ah_attr,
428 const struct ib_gid_attr **old_sgid_attr)
430 const struct ib_gid_attr *sgid_attr;
431 struct ib_global_route *grh;
434 *old_sgid_attr = ah_attr->grh.sgid_attr;
436 ret = rdma_check_ah_attr(device, ah_attr);
440 if (!(ah_attr->ah_flags & IB_AH_GRH))
443 grh = rdma_ah_retrieve_grh(ah_attr);
448 rdma_get_gid_attr(device, ah_attr->port_num, grh->sgid_index);
449 if (IS_ERR(sgid_attr))
450 return PTR_ERR(sgid_attr);
452 /* Move ownerhip of the kref into the ah_attr */
453 grh->sgid_attr = sgid_attr;
457 static void rdma_unfill_sgid_attr(struct rdma_ah_attr *ah_attr,
458 const struct ib_gid_attr *old_sgid_attr)
461 * Fill didn't change anything, the caller retains ownership of
464 if (ah_attr->grh.sgid_attr == old_sgid_attr)
468 * Otherwise, we need to undo what rdma_fill_sgid_attr so the caller
469 * doesn't see any change in the rdma_ah_attr. If we get here
470 * old_sgid_attr is NULL.
472 rdma_destroy_ah_attr(ah_attr);
475 static const struct ib_gid_attr *
476 rdma_update_sgid_attr(struct rdma_ah_attr *ah_attr,
477 const struct ib_gid_attr *old_attr)
480 rdma_put_gid_attr(old_attr);
481 if (ah_attr->ah_flags & IB_AH_GRH) {
482 rdma_hold_gid_attr(ah_attr->grh.sgid_attr);
483 return ah_attr->grh.sgid_attr;
488 static struct ib_ah *_rdma_create_ah(struct ib_pd *pd,
489 struct rdma_ah_attr *ah_attr,
490 struct ib_udata *udata)
494 if (!pd->device->ops.create_ah)
495 return ERR_PTR(-EOPNOTSUPP);
497 ah = pd->device->ops.create_ah(pd, ah_attr, udata);
500 ah->device = pd->device;
503 ah->type = ah_attr->type;
504 ah->sgid_attr = rdma_update_sgid_attr(ah_attr, NULL);
506 atomic_inc(&pd->usecnt);
513 * rdma_create_ah - Creates an address handle for the
514 * given address vector.
515 * @pd: The protection domain associated with the address handle.
516 * @ah_attr: The attributes of the address vector.
518 * It returns 0 on success and returns appropriate error code on error.
519 * The address handle is used to reference a local or global destination
520 * in all UD QP post sends.
522 struct ib_ah *rdma_create_ah(struct ib_pd *pd, struct rdma_ah_attr *ah_attr)
524 const struct ib_gid_attr *old_sgid_attr;
528 ret = rdma_fill_sgid_attr(pd->device, ah_attr, &old_sgid_attr);
532 ah = _rdma_create_ah(pd, ah_attr, NULL);
534 rdma_unfill_sgid_attr(ah_attr, old_sgid_attr);
537 EXPORT_SYMBOL(rdma_create_ah);
540 * rdma_create_user_ah - Creates an address handle for the
541 * given address vector.
542 * It resolves destination mac address for ah attribute of RoCE type.
543 * @pd: The protection domain associated with the address handle.
544 * @ah_attr: The attributes of the address vector.
545 * @udata: pointer to user's input output buffer information need by
548 * It returns 0 on success and returns appropriate error code on error.
549 * The address handle is used to reference a local or global destination
550 * in all UD QP post sends.
552 struct ib_ah *rdma_create_user_ah(struct ib_pd *pd,
553 struct rdma_ah_attr *ah_attr,
554 struct ib_udata *udata)
556 const struct ib_gid_attr *old_sgid_attr;
560 err = rdma_fill_sgid_attr(pd->device, ah_attr, &old_sgid_attr);
564 if (ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE) {
565 err = ib_resolve_eth_dmac(pd->device, ah_attr);
572 ah = _rdma_create_ah(pd, ah_attr, udata);
575 rdma_unfill_sgid_attr(ah_attr, old_sgid_attr);
578 EXPORT_SYMBOL(rdma_create_user_ah);
580 int ib_get_rdma_header_version(const union rdma_network_hdr *hdr)
582 const struct iphdr *ip4h = (struct iphdr *)&hdr->roce4grh;
583 struct iphdr ip4h_checked;
584 const struct ipv6hdr *ip6h = (struct ipv6hdr *)&hdr->ibgrh;
586 /* If it's IPv6, the version must be 6, otherwise, the first
587 * 20 bytes (before the IPv4 header) are garbled.
589 if (ip6h->version != 6)
590 return (ip4h->version == 4) ? 4 : 0;
591 /* version may be 6 or 4 because the first 20 bytes could be garbled */
593 /* RoCE v2 requires no options, thus header length
600 * We can't write on scattered buffers so we need to copy to
603 memcpy(&ip4h_checked, ip4h, sizeof(ip4h_checked));
604 ip4h_checked.check = 0;
605 ip4h_checked.check = ip_fast_csum((u8 *)&ip4h_checked, 5);
606 /* if IPv4 header checksum is OK, believe it */
607 if (ip4h->check == ip4h_checked.check)
611 EXPORT_SYMBOL(ib_get_rdma_header_version);
613 static enum rdma_network_type ib_get_net_type_by_grh(struct ib_device *device,
615 const struct ib_grh *grh)
619 if (rdma_protocol_ib(device, port_num))
620 return RDMA_NETWORK_IB;
622 grh_version = ib_get_rdma_header_version((union rdma_network_hdr *)grh);
624 if (grh_version == 4)
625 return RDMA_NETWORK_IPV4;
627 if (grh->next_hdr == IPPROTO_UDP)
628 return RDMA_NETWORK_IPV6;
630 return RDMA_NETWORK_ROCE_V1;
633 struct find_gid_index_context {
635 enum ib_gid_type gid_type;
638 static bool find_gid_index(const union ib_gid *gid,
639 const struct ib_gid_attr *gid_attr,
642 struct find_gid_index_context *ctx = context;
644 if (ctx->gid_type != gid_attr->gid_type)
647 if ((!!(ctx->vlan_id != 0xffff) == !is_vlan_dev(gid_attr->ndev)) ||
648 (is_vlan_dev(gid_attr->ndev) &&
649 vlan_dev_vlan_id(gid_attr->ndev) != ctx->vlan_id))
655 static const struct ib_gid_attr *
656 get_sgid_attr_from_eth(struct ib_device *device, u8 port_num,
657 u16 vlan_id, const union ib_gid *sgid,
658 enum ib_gid_type gid_type)
660 struct find_gid_index_context context = {.vlan_id = vlan_id,
661 .gid_type = gid_type};
663 return rdma_find_gid_by_filter(device, sgid, port_num, find_gid_index,
667 int ib_get_gids_from_rdma_hdr(const union rdma_network_hdr *hdr,
668 enum rdma_network_type net_type,
669 union ib_gid *sgid, union ib_gid *dgid)
671 struct sockaddr_in src_in;
672 struct sockaddr_in dst_in;
673 __be32 src_saddr, dst_saddr;
678 if (net_type == RDMA_NETWORK_IPV4) {
679 memcpy(&src_in.sin_addr.s_addr,
680 &hdr->roce4grh.saddr, 4);
681 memcpy(&dst_in.sin_addr.s_addr,
682 &hdr->roce4grh.daddr, 4);
683 src_saddr = src_in.sin_addr.s_addr;
684 dst_saddr = dst_in.sin_addr.s_addr;
685 ipv6_addr_set_v4mapped(src_saddr,
686 (struct in6_addr *)sgid);
687 ipv6_addr_set_v4mapped(dst_saddr,
688 (struct in6_addr *)dgid);
690 } else if (net_type == RDMA_NETWORK_IPV6 ||
691 net_type == RDMA_NETWORK_IB) {
692 *dgid = hdr->ibgrh.dgid;
693 *sgid = hdr->ibgrh.sgid;
699 EXPORT_SYMBOL(ib_get_gids_from_rdma_hdr);
701 /* Resolve destination mac address and hop limit for unicast destination
702 * GID entry, considering the source GID entry as well.
703 * ah_attribute must have have valid port_num, sgid_index.
705 static int ib_resolve_unicast_gid_dmac(struct ib_device *device,
706 struct rdma_ah_attr *ah_attr)
708 struct ib_global_route *grh = rdma_ah_retrieve_grh(ah_attr);
709 const struct ib_gid_attr *sgid_attr = grh->sgid_attr;
710 int hop_limit = 0xff;
713 /* If destination is link local and source GID is RoCEv1,
714 * IP stack is not used.
716 if (rdma_link_local_addr((struct in6_addr *)grh->dgid.raw) &&
717 sgid_attr->gid_type == IB_GID_TYPE_ROCE) {
718 rdma_get_ll_mac((struct in6_addr *)grh->dgid.raw,
723 ret = rdma_addr_find_l2_eth_by_grh(&sgid_attr->gid, &grh->dgid,
725 sgid_attr, &hop_limit);
727 grh->hop_limit = hop_limit;
732 * This function initializes address handle attributes from the incoming packet.
733 * Incoming packet has dgid of the receiver node on which this code is
734 * getting executed and, sgid contains the GID of the sender.
736 * When resolving mac address of destination, the arrived dgid is used
737 * as sgid and, sgid is used as dgid because sgid contains destinations
738 * GID whom to respond to.
740 * On success the caller is responsible to call rdma_destroy_ah_attr on the
743 int ib_init_ah_attr_from_wc(struct ib_device *device, u8 port_num,
744 const struct ib_wc *wc, const struct ib_grh *grh,
745 struct rdma_ah_attr *ah_attr)
749 enum rdma_network_type net_type = RDMA_NETWORK_IB;
750 enum ib_gid_type gid_type = IB_GID_TYPE_IB;
751 const struct ib_gid_attr *sgid_attr;
758 memset(ah_attr, 0, sizeof *ah_attr);
759 ah_attr->type = rdma_ah_find_type(device, port_num);
760 if (rdma_cap_eth_ah(device, port_num)) {
761 if (wc->wc_flags & IB_WC_WITH_NETWORK_HDR_TYPE)
762 net_type = wc->network_hdr_type;
764 net_type = ib_get_net_type_by_grh(device, port_num, grh);
765 gid_type = ib_network_to_gid_type(net_type);
767 ret = ib_get_gids_from_rdma_hdr((union rdma_network_hdr *)grh, net_type,
772 rdma_ah_set_sl(ah_attr, wc->sl);
773 rdma_ah_set_port_num(ah_attr, port_num);
775 if (rdma_protocol_roce(device, port_num)) {
776 u16 vlan_id = wc->wc_flags & IB_WC_WITH_VLAN ?
777 wc->vlan_id : 0xffff;
779 if (!(wc->wc_flags & IB_WC_GRH))
782 sgid_attr = get_sgid_attr_from_eth(device, port_num,
785 if (IS_ERR(sgid_attr))
786 return PTR_ERR(sgid_attr);
788 flow_class = be32_to_cpu(grh->version_tclass_flow);
789 rdma_move_grh_sgid_attr(ah_attr,
791 flow_class & 0xFFFFF,
793 (flow_class >> 20) & 0xFF,
796 ret = ib_resolve_unicast_gid_dmac(device, ah_attr);
798 rdma_destroy_ah_attr(ah_attr);
802 rdma_ah_set_dlid(ah_attr, wc->slid);
803 rdma_ah_set_path_bits(ah_attr, wc->dlid_path_bits);
805 if ((wc->wc_flags & IB_WC_GRH) == 0)
808 if (dgid.global.interface_id !=
809 cpu_to_be64(IB_SA_WELL_KNOWN_GUID)) {
810 sgid_attr = rdma_find_gid_by_port(
811 device, &dgid, IB_GID_TYPE_IB, port_num, NULL);
813 sgid_attr = rdma_get_gid_attr(device, port_num, 0);
815 if (IS_ERR(sgid_attr))
816 return PTR_ERR(sgid_attr);
817 flow_class = be32_to_cpu(grh->version_tclass_flow);
818 rdma_move_grh_sgid_attr(ah_attr,
820 flow_class & 0xFFFFF,
822 (flow_class >> 20) & 0xFF,
828 EXPORT_SYMBOL(ib_init_ah_attr_from_wc);
831 * rdma_move_grh_sgid_attr - Sets the sgid attribute of GRH, taking ownership
834 * @attr: Pointer to AH attribute structure
835 * @dgid: Destination GID
836 * @flow_label: Flow label
837 * @hop_limit: Hop limit
838 * @traffic_class: traffic class
839 * @sgid_attr: Pointer to SGID attribute
841 * This takes ownership of the sgid_attr reference. The caller must ensure
842 * rdma_destroy_ah_attr() is called before destroying the rdma_ah_attr after
843 * calling this function.
845 void rdma_move_grh_sgid_attr(struct rdma_ah_attr *attr, union ib_gid *dgid,
846 u32 flow_label, u8 hop_limit, u8 traffic_class,
847 const struct ib_gid_attr *sgid_attr)
849 rdma_ah_set_grh(attr, dgid, flow_label, sgid_attr->index, hop_limit,
851 attr->grh.sgid_attr = sgid_attr;
853 EXPORT_SYMBOL(rdma_move_grh_sgid_attr);
856 * rdma_destroy_ah_attr - Release reference to SGID attribute of
858 * @ah_attr: Pointer to ah attribute
860 * Release reference to the SGID attribute of the ah attribute if it is
861 * non NULL. It is safe to call this multiple times, and safe to call it on
862 * a zero initialized ah_attr.
864 void rdma_destroy_ah_attr(struct rdma_ah_attr *ah_attr)
866 if (ah_attr->grh.sgid_attr) {
867 rdma_put_gid_attr(ah_attr->grh.sgid_attr);
868 ah_attr->grh.sgid_attr = NULL;
871 EXPORT_SYMBOL(rdma_destroy_ah_attr);
873 struct ib_ah *ib_create_ah_from_wc(struct ib_pd *pd, const struct ib_wc *wc,
874 const struct ib_grh *grh, u8 port_num)
876 struct rdma_ah_attr ah_attr;
880 ret = ib_init_ah_attr_from_wc(pd->device, port_num, wc, grh, &ah_attr);
884 ah = rdma_create_ah(pd, &ah_attr);
886 rdma_destroy_ah_attr(&ah_attr);
889 EXPORT_SYMBOL(ib_create_ah_from_wc);
891 int rdma_modify_ah(struct ib_ah *ah, struct rdma_ah_attr *ah_attr)
893 const struct ib_gid_attr *old_sgid_attr;
896 if (ah->type != ah_attr->type)
899 ret = rdma_fill_sgid_attr(ah->device, ah_attr, &old_sgid_attr);
903 ret = ah->device->ops.modify_ah ?
904 ah->device->ops.modify_ah(ah, ah_attr) :
907 ah->sgid_attr = rdma_update_sgid_attr(ah_attr, ah->sgid_attr);
908 rdma_unfill_sgid_attr(ah_attr, old_sgid_attr);
911 EXPORT_SYMBOL(rdma_modify_ah);
913 int rdma_query_ah(struct ib_ah *ah, struct rdma_ah_attr *ah_attr)
915 ah_attr->grh.sgid_attr = NULL;
917 return ah->device->ops.query_ah ?
918 ah->device->ops.query_ah(ah, ah_attr) :
921 EXPORT_SYMBOL(rdma_query_ah);
923 int rdma_destroy_ah(struct ib_ah *ah)
925 const struct ib_gid_attr *sgid_attr = ah->sgid_attr;
930 ret = ah->device->ops.destroy_ah(ah);
932 atomic_dec(&pd->usecnt);
934 rdma_put_gid_attr(sgid_attr);
939 EXPORT_SYMBOL(rdma_destroy_ah);
941 /* Shared receive queues */
943 struct ib_srq *ib_create_srq(struct ib_pd *pd,
944 struct ib_srq_init_attr *srq_init_attr)
948 if (!pd->device->ops.create_srq)
949 return ERR_PTR(-EOPNOTSUPP);
951 srq = pd->device->ops.create_srq(pd, srq_init_attr, NULL);
954 srq->device = pd->device;
957 srq->event_handler = srq_init_attr->event_handler;
958 srq->srq_context = srq_init_attr->srq_context;
959 srq->srq_type = srq_init_attr->srq_type;
960 if (ib_srq_has_cq(srq->srq_type)) {
961 srq->ext.cq = srq_init_attr->ext.cq;
962 atomic_inc(&srq->ext.cq->usecnt);
964 if (srq->srq_type == IB_SRQT_XRC) {
965 srq->ext.xrc.xrcd = srq_init_attr->ext.xrc.xrcd;
966 atomic_inc(&srq->ext.xrc.xrcd->usecnt);
968 atomic_inc(&pd->usecnt);
969 atomic_set(&srq->usecnt, 0);
974 EXPORT_SYMBOL(ib_create_srq);
976 int ib_modify_srq(struct ib_srq *srq,
977 struct ib_srq_attr *srq_attr,
978 enum ib_srq_attr_mask srq_attr_mask)
980 return srq->device->ops.modify_srq ?
981 srq->device->ops.modify_srq(srq, srq_attr, srq_attr_mask,
984 EXPORT_SYMBOL(ib_modify_srq);
986 int ib_query_srq(struct ib_srq *srq,
987 struct ib_srq_attr *srq_attr)
989 return srq->device->ops.query_srq ?
990 srq->device->ops.query_srq(srq, srq_attr) : -EOPNOTSUPP;
992 EXPORT_SYMBOL(ib_query_srq);
994 int ib_destroy_srq(struct ib_srq *srq)
997 enum ib_srq_type srq_type;
998 struct ib_xrcd *uninitialized_var(xrcd);
999 struct ib_cq *uninitialized_var(cq);
1002 if (atomic_read(&srq->usecnt))
1006 srq_type = srq->srq_type;
1007 if (ib_srq_has_cq(srq_type))
1009 if (srq_type == IB_SRQT_XRC)
1010 xrcd = srq->ext.xrc.xrcd;
1012 ret = srq->device->ops.destroy_srq(srq);
1014 atomic_dec(&pd->usecnt);
1015 if (srq_type == IB_SRQT_XRC)
1016 atomic_dec(&xrcd->usecnt);
1017 if (ib_srq_has_cq(srq_type))
1018 atomic_dec(&cq->usecnt);
1023 EXPORT_SYMBOL(ib_destroy_srq);
1027 static void __ib_shared_qp_event_handler(struct ib_event *event, void *context)
1029 struct ib_qp *qp = context;
1030 unsigned long flags;
1032 spin_lock_irqsave(&qp->device->event_handler_lock, flags);
1033 list_for_each_entry(event->element.qp, &qp->open_list, open_list)
1034 if (event->element.qp->event_handler)
1035 event->element.qp->event_handler(event, event->element.qp->qp_context);
1036 spin_unlock_irqrestore(&qp->device->event_handler_lock, flags);
1039 static void __ib_insert_xrcd_qp(struct ib_xrcd *xrcd, struct ib_qp *qp)
1041 mutex_lock(&xrcd->tgt_qp_mutex);
1042 list_add(&qp->xrcd_list, &xrcd->tgt_qp_list);
1043 mutex_unlock(&xrcd->tgt_qp_mutex);
1046 static struct ib_qp *__ib_open_qp(struct ib_qp *real_qp,
1047 void (*event_handler)(struct ib_event *, void *),
1051 unsigned long flags;
1054 qp = kzalloc(sizeof *qp, GFP_KERNEL);
1056 return ERR_PTR(-ENOMEM);
1058 qp->real_qp = real_qp;
1059 err = ib_open_shared_qp_security(qp, real_qp->device);
1062 return ERR_PTR(err);
1065 qp->real_qp = real_qp;
1066 atomic_inc(&real_qp->usecnt);
1067 qp->device = real_qp->device;
1068 qp->event_handler = event_handler;
1069 qp->qp_context = qp_context;
1070 qp->qp_num = real_qp->qp_num;
1071 qp->qp_type = real_qp->qp_type;
1073 spin_lock_irqsave(&real_qp->device->event_handler_lock, flags);
1074 list_add(&qp->open_list, &real_qp->open_list);
1075 spin_unlock_irqrestore(&real_qp->device->event_handler_lock, flags);
1080 struct ib_qp *ib_open_qp(struct ib_xrcd *xrcd,
1081 struct ib_qp_open_attr *qp_open_attr)
1083 struct ib_qp *qp, *real_qp;
1085 if (qp_open_attr->qp_type != IB_QPT_XRC_TGT)
1086 return ERR_PTR(-EINVAL);
1088 qp = ERR_PTR(-EINVAL);
1089 mutex_lock(&xrcd->tgt_qp_mutex);
1090 list_for_each_entry(real_qp, &xrcd->tgt_qp_list, xrcd_list) {
1091 if (real_qp->qp_num == qp_open_attr->qp_num) {
1092 qp = __ib_open_qp(real_qp, qp_open_attr->event_handler,
1093 qp_open_attr->qp_context);
1097 mutex_unlock(&xrcd->tgt_qp_mutex);
1100 EXPORT_SYMBOL(ib_open_qp);
1102 static struct ib_qp *ib_create_xrc_qp(struct ib_qp *qp,
1103 struct ib_qp_init_attr *qp_init_attr)
1105 struct ib_qp *real_qp = qp;
1107 qp->event_handler = __ib_shared_qp_event_handler;
1108 qp->qp_context = qp;
1110 qp->send_cq = qp->recv_cq = NULL;
1112 qp->xrcd = qp_init_attr->xrcd;
1113 atomic_inc(&qp_init_attr->xrcd->usecnt);
1114 INIT_LIST_HEAD(&qp->open_list);
1116 qp = __ib_open_qp(real_qp, qp_init_attr->event_handler,
1117 qp_init_attr->qp_context);
1119 __ib_insert_xrcd_qp(qp_init_attr->xrcd, real_qp);
1121 real_qp->device->ops.destroy_qp(real_qp);
1125 struct ib_qp *ib_create_qp(struct ib_pd *pd,
1126 struct ib_qp_init_attr *qp_init_attr)
1128 struct ib_device *device = pd ? pd->device : qp_init_attr->xrcd->device;
1132 if (qp_init_attr->rwq_ind_tbl &&
1133 (qp_init_attr->recv_cq ||
1134 qp_init_attr->srq || qp_init_attr->cap.max_recv_wr ||
1135 qp_init_attr->cap.max_recv_sge))
1136 return ERR_PTR(-EINVAL);
1139 * If the callers is using the RDMA API calculate the resources
1140 * needed for the RDMA READ/WRITE operations.
1142 * Note that these callers need to pass in a port number.
1144 if (qp_init_attr->cap.max_rdma_ctxs)
1145 rdma_rw_init_qp(device, qp_init_attr);
1147 qp = _ib_create_qp(device, pd, qp_init_attr, NULL, NULL);
1151 ret = ib_create_qp_security(qp, device);
1154 return ERR_PTR(ret);
1158 qp->qp_type = qp_init_attr->qp_type;
1159 qp->rwq_ind_tbl = qp_init_attr->rwq_ind_tbl;
1161 atomic_set(&qp->usecnt, 0);
1163 spin_lock_init(&qp->mr_lock);
1164 INIT_LIST_HEAD(&qp->rdma_mrs);
1165 INIT_LIST_HEAD(&qp->sig_mrs);
1168 if (qp_init_attr->qp_type == IB_QPT_XRC_TGT)
1169 return ib_create_xrc_qp(qp, qp_init_attr);
1171 qp->event_handler = qp_init_attr->event_handler;
1172 qp->qp_context = qp_init_attr->qp_context;
1173 if (qp_init_attr->qp_type == IB_QPT_XRC_INI) {
1177 qp->recv_cq = qp_init_attr->recv_cq;
1178 if (qp_init_attr->recv_cq)
1179 atomic_inc(&qp_init_attr->recv_cq->usecnt);
1180 qp->srq = qp_init_attr->srq;
1182 atomic_inc(&qp_init_attr->srq->usecnt);
1185 qp->send_cq = qp_init_attr->send_cq;
1188 atomic_inc(&pd->usecnt);
1189 if (qp_init_attr->send_cq)
1190 atomic_inc(&qp_init_attr->send_cq->usecnt);
1191 if (qp_init_attr->rwq_ind_tbl)
1192 atomic_inc(&qp->rwq_ind_tbl->usecnt);
1194 if (qp_init_attr->cap.max_rdma_ctxs) {
1195 ret = rdma_rw_init_mrs(qp, qp_init_attr);
1197 pr_err("failed to init MR pool ret= %d\n", ret);
1199 return ERR_PTR(ret);
1204 * Note: all hw drivers guarantee that max_send_sge is lower than
1205 * the device RDMA WRITE SGE limit but not all hw drivers ensure that
1206 * max_send_sge <= max_sge_rd.
1208 qp->max_write_sge = qp_init_attr->cap.max_send_sge;
1209 qp->max_read_sge = min_t(u32, qp_init_attr->cap.max_send_sge,
1210 device->attrs.max_sge_rd);
1214 EXPORT_SYMBOL(ib_create_qp);
1216 static const struct {
1218 enum ib_qp_attr_mask req_param[IB_QPT_MAX];
1219 enum ib_qp_attr_mask opt_param[IB_QPT_MAX];
1220 } qp_state_table[IB_QPS_ERR + 1][IB_QPS_ERR + 1] = {
1222 [IB_QPS_RESET] = { .valid = 1 },
1226 [IB_QPT_UD] = (IB_QP_PKEY_INDEX |
1229 [IB_QPT_RAW_PACKET] = IB_QP_PORT,
1230 [IB_QPT_UC] = (IB_QP_PKEY_INDEX |
1232 IB_QP_ACCESS_FLAGS),
1233 [IB_QPT_RC] = (IB_QP_PKEY_INDEX |
1235 IB_QP_ACCESS_FLAGS),
1236 [IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX |
1238 IB_QP_ACCESS_FLAGS),
1239 [IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX |
1241 IB_QP_ACCESS_FLAGS),
1242 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX |
1244 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX |
1250 [IB_QPS_RESET] = { .valid = 1 },
1251 [IB_QPS_ERR] = { .valid = 1 },
1255 [IB_QPT_UD] = (IB_QP_PKEY_INDEX |
1258 [IB_QPT_UC] = (IB_QP_PKEY_INDEX |
1260 IB_QP_ACCESS_FLAGS),
1261 [IB_QPT_RC] = (IB_QP_PKEY_INDEX |
1263 IB_QP_ACCESS_FLAGS),
1264 [IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX |
1266 IB_QP_ACCESS_FLAGS),
1267 [IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX |
1269 IB_QP_ACCESS_FLAGS),
1270 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX |
1272 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX |
1279 [IB_QPT_UC] = (IB_QP_AV |
1283 [IB_QPT_RC] = (IB_QP_AV |
1287 IB_QP_MAX_DEST_RD_ATOMIC |
1288 IB_QP_MIN_RNR_TIMER),
1289 [IB_QPT_XRC_INI] = (IB_QP_AV |
1293 [IB_QPT_XRC_TGT] = (IB_QP_AV |
1297 IB_QP_MAX_DEST_RD_ATOMIC |
1298 IB_QP_MIN_RNR_TIMER),
1301 [IB_QPT_UD] = (IB_QP_PKEY_INDEX |
1303 [IB_QPT_UC] = (IB_QP_ALT_PATH |
1304 IB_QP_ACCESS_FLAGS |
1306 [IB_QPT_RC] = (IB_QP_ALT_PATH |
1307 IB_QP_ACCESS_FLAGS |
1309 [IB_QPT_XRC_INI] = (IB_QP_ALT_PATH |
1310 IB_QP_ACCESS_FLAGS |
1312 [IB_QPT_XRC_TGT] = (IB_QP_ALT_PATH |
1313 IB_QP_ACCESS_FLAGS |
1315 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX |
1317 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX |
1323 [IB_QPS_RESET] = { .valid = 1 },
1324 [IB_QPS_ERR] = { .valid = 1 },
1328 [IB_QPT_UD] = IB_QP_SQ_PSN,
1329 [IB_QPT_UC] = IB_QP_SQ_PSN,
1330 [IB_QPT_RC] = (IB_QP_TIMEOUT |
1334 IB_QP_MAX_QP_RD_ATOMIC),
1335 [IB_QPT_XRC_INI] = (IB_QP_TIMEOUT |
1339 IB_QP_MAX_QP_RD_ATOMIC),
1340 [IB_QPT_XRC_TGT] = (IB_QP_TIMEOUT |
1342 [IB_QPT_SMI] = IB_QP_SQ_PSN,
1343 [IB_QPT_GSI] = IB_QP_SQ_PSN,
1346 [IB_QPT_UD] = (IB_QP_CUR_STATE |
1348 [IB_QPT_UC] = (IB_QP_CUR_STATE |
1350 IB_QP_ACCESS_FLAGS |
1351 IB_QP_PATH_MIG_STATE),
1352 [IB_QPT_RC] = (IB_QP_CUR_STATE |
1354 IB_QP_ACCESS_FLAGS |
1355 IB_QP_MIN_RNR_TIMER |
1356 IB_QP_PATH_MIG_STATE),
1357 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE |
1359 IB_QP_ACCESS_FLAGS |
1360 IB_QP_PATH_MIG_STATE),
1361 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE |
1363 IB_QP_ACCESS_FLAGS |
1364 IB_QP_MIN_RNR_TIMER |
1365 IB_QP_PATH_MIG_STATE),
1366 [IB_QPT_SMI] = (IB_QP_CUR_STATE |
1368 [IB_QPT_GSI] = (IB_QP_CUR_STATE |
1370 [IB_QPT_RAW_PACKET] = IB_QP_RATE_LIMIT,
1375 [IB_QPS_RESET] = { .valid = 1 },
1376 [IB_QPS_ERR] = { .valid = 1 },
1380 [IB_QPT_UD] = (IB_QP_CUR_STATE |
1382 [IB_QPT_UC] = (IB_QP_CUR_STATE |
1383 IB_QP_ACCESS_FLAGS |
1385 IB_QP_PATH_MIG_STATE),
1386 [IB_QPT_RC] = (IB_QP_CUR_STATE |
1387 IB_QP_ACCESS_FLAGS |
1389 IB_QP_PATH_MIG_STATE |
1390 IB_QP_MIN_RNR_TIMER),
1391 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE |
1392 IB_QP_ACCESS_FLAGS |
1394 IB_QP_PATH_MIG_STATE),
1395 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE |
1396 IB_QP_ACCESS_FLAGS |
1398 IB_QP_PATH_MIG_STATE |
1399 IB_QP_MIN_RNR_TIMER),
1400 [IB_QPT_SMI] = (IB_QP_CUR_STATE |
1402 [IB_QPT_GSI] = (IB_QP_CUR_STATE |
1404 [IB_QPT_RAW_PACKET] = IB_QP_RATE_LIMIT,
1410 [IB_QPT_UD] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1411 [IB_QPT_UC] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1412 [IB_QPT_RC] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1413 [IB_QPT_XRC_INI] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1414 [IB_QPT_XRC_TGT] = IB_QP_EN_SQD_ASYNC_NOTIFY, /* ??? */
1415 [IB_QPT_SMI] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1416 [IB_QPT_GSI] = IB_QP_EN_SQD_ASYNC_NOTIFY
1421 [IB_QPS_RESET] = { .valid = 1 },
1422 [IB_QPS_ERR] = { .valid = 1 },
1426 [IB_QPT_UD] = (IB_QP_CUR_STATE |
1428 [IB_QPT_UC] = (IB_QP_CUR_STATE |
1430 IB_QP_ACCESS_FLAGS |
1431 IB_QP_PATH_MIG_STATE),
1432 [IB_QPT_RC] = (IB_QP_CUR_STATE |
1434 IB_QP_ACCESS_FLAGS |
1435 IB_QP_MIN_RNR_TIMER |
1436 IB_QP_PATH_MIG_STATE),
1437 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE |
1439 IB_QP_ACCESS_FLAGS |
1440 IB_QP_PATH_MIG_STATE),
1441 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE |
1443 IB_QP_ACCESS_FLAGS |
1444 IB_QP_MIN_RNR_TIMER |
1445 IB_QP_PATH_MIG_STATE),
1446 [IB_QPT_SMI] = (IB_QP_CUR_STATE |
1448 [IB_QPT_GSI] = (IB_QP_CUR_STATE |
1455 [IB_QPT_UD] = (IB_QP_PKEY_INDEX |
1457 [IB_QPT_UC] = (IB_QP_AV |
1459 IB_QP_ACCESS_FLAGS |
1461 IB_QP_PATH_MIG_STATE),
1462 [IB_QPT_RC] = (IB_QP_PORT |
1467 IB_QP_MAX_QP_RD_ATOMIC |
1468 IB_QP_MAX_DEST_RD_ATOMIC |
1470 IB_QP_ACCESS_FLAGS |
1472 IB_QP_MIN_RNR_TIMER |
1473 IB_QP_PATH_MIG_STATE),
1474 [IB_QPT_XRC_INI] = (IB_QP_PORT |
1479 IB_QP_MAX_QP_RD_ATOMIC |
1481 IB_QP_ACCESS_FLAGS |
1483 IB_QP_PATH_MIG_STATE),
1484 [IB_QPT_XRC_TGT] = (IB_QP_PORT |
1487 IB_QP_MAX_DEST_RD_ATOMIC |
1489 IB_QP_ACCESS_FLAGS |
1491 IB_QP_MIN_RNR_TIMER |
1492 IB_QP_PATH_MIG_STATE),
1493 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX |
1495 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX |
1501 [IB_QPS_RESET] = { .valid = 1 },
1502 [IB_QPS_ERR] = { .valid = 1 },
1506 [IB_QPT_UD] = (IB_QP_CUR_STATE |
1508 [IB_QPT_UC] = (IB_QP_CUR_STATE |
1509 IB_QP_ACCESS_FLAGS),
1510 [IB_QPT_SMI] = (IB_QP_CUR_STATE |
1512 [IB_QPT_GSI] = (IB_QP_CUR_STATE |
1518 [IB_QPS_RESET] = { .valid = 1 },
1519 [IB_QPS_ERR] = { .valid = 1 }
1523 bool ib_modify_qp_is_ok(enum ib_qp_state cur_state, enum ib_qp_state next_state,
1524 enum ib_qp_type type, enum ib_qp_attr_mask mask)
1526 enum ib_qp_attr_mask req_param, opt_param;
1528 if (mask & IB_QP_CUR_STATE &&
1529 cur_state != IB_QPS_RTR && cur_state != IB_QPS_RTS &&
1530 cur_state != IB_QPS_SQD && cur_state != IB_QPS_SQE)
1533 if (!qp_state_table[cur_state][next_state].valid)
1536 req_param = qp_state_table[cur_state][next_state].req_param[type];
1537 opt_param = qp_state_table[cur_state][next_state].opt_param[type];
1539 if ((mask & req_param) != req_param)
1542 if (mask & ~(req_param | opt_param | IB_QP_STATE))
1547 EXPORT_SYMBOL(ib_modify_qp_is_ok);
1550 * ib_resolve_eth_dmac - Resolve destination mac address
1551 * @device: Device to consider
1552 * @ah_attr: address handle attribute which describes the
1553 * source and destination parameters
1554 * ib_resolve_eth_dmac() resolves destination mac address and L3 hop limit It
1555 * returns 0 on success or appropriate error code. It initializes the
1556 * necessary ah_attr fields when call is successful.
1558 static int ib_resolve_eth_dmac(struct ib_device *device,
1559 struct rdma_ah_attr *ah_attr)
1563 if (rdma_is_multicast_addr((struct in6_addr *)ah_attr->grh.dgid.raw)) {
1564 if (ipv6_addr_v4mapped((struct in6_addr *)ah_attr->grh.dgid.raw)) {
1567 memcpy(&addr, ah_attr->grh.dgid.raw + 12, 4);
1568 ip_eth_mc_map(addr, (char *)ah_attr->roce.dmac);
1570 ipv6_eth_mc_map((struct in6_addr *)ah_attr->grh.dgid.raw,
1571 (char *)ah_attr->roce.dmac);
1574 ret = ib_resolve_unicast_gid_dmac(device, ah_attr);
1579 static bool is_qp_type_connected(const struct ib_qp *qp)
1581 return (qp->qp_type == IB_QPT_UC ||
1582 qp->qp_type == IB_QPT_RC ||
1583 qp->qp_type == IB_QPT_XRC_INI ||
1584 qp->qp_type == IB_QPT_XRC_TGT);
1588 * IB core internal function to perform QP attributes modification.
1590 static int _ib_modify_qp(struct ib_qp *qp, struct ib_qp_attr *attr,
1591 int attr_mask, struct ib_udata *udata)
1593 u8 port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
1594 const struct ib_gid_attr *old_sgid_attr_av;
1595 const struct ib_gid_attr *old_sgid_attr_alt_av;
1598 if (attr_mask & IB_QP_AV) {
1599 ret = rdma_fill_sgid_attr(qp->device, &attr->ah_attr,
1604 if (attr_mask & IB_QP_ALT_PATH) {
1606 * FIXME: This does not track the migration state, so if the
1607 * user loads a new alternate path after the HW has migrated
1608 * from primary->alternate we will keep the wrong
1609 * references. This is OK for IB because the reference
1610 * counting does not serve any functional purpose.
1612 ret = rdma_fill_sgid_attr(qp->device, &attr->alt_ah_attr,
1613 &old_sgid_attr_alt_av);
1618 * Today the core code can only handle alternate paths and APM
1619 * for IB. Ban them in roce mode.
1621 if (!(rdma_protocol_ib(qp->device,
1622 attr->alt_ah_attr.port_num) &&
1623 rdma_protocol_ib(qp->device, port))) {
1630 * If the user provided the qp_attr then we have to resolve it. Kernel
1631 * users have to provide already resolved rdma_ah_attr's
1633 if (udata && (attr_mask & IB_QP_AV) &&
1634 attr->ah_attr.type == RDMA_AH_ATTR_TYPE_ROCE &&
1635 is_qp_type_connected(qp)) {
1636 ret = ib_resolve_eth_dmac(qp->device, &attr->ah_attr);
1641 if (rdma_ib_or_roce(qp->device, port)) {
1642 if (attr_mask & IB_QP_RQ_PSN && attr->rq_psn & ~0xffffff) {
1643 dev_warn(&qp->device->dev,
1644 "%s rq_psn overflow, masking to 24 bits\n",
1646 attr->rq_psn &= 0xffffff;
1649 if (attr_mask & IB_QP_SQ_PSN && attr->sq_psn & ~0xffffff) {
1650 dev_warn(&qp->device->dev,
1651 " %s sq_psn overflow, masking to 24 bits\n",
1653 attr->sq_psn &= 0xffffff;
1657 ret = ib_security_modify_qp(qp, attr, attr_mask, udata);
1661 if (attr_mask & IB_QP_PORT)
1662 qp->port = attr->port_num;
1663 if (attr_mask & IB_QP_AV)
1665 rdma_update_sgid_attr(&attr->ah_attr, qp->av_sgid_attr);
1666 if (attr_mask & IB_QP_ALT_PATH)
1667 qp->alt_path_sgid_attr = rdma_update_sgid_attr(
1668 &attr->alt_ah_attr, qp->alt_path_sgid_attr);
1671 if (attr_mask & IB_QP_ALT_PATH)
1672 rdma_unfill_sgid_attr(&attr->alt_ah_attr, old_sgid_attr_alt_av);
1674 if (attr_mask & IB_QP_AV)
1675 rdma_unfill_sgid_attr(&attr->ah_attr, old_sgid_attr_av);
1680 * ib_modify_qp_with_udata - Modifies the attributes for the specified QP.
1681 * @ib_qp: The QP to modify.
1682 * @attr: On input, specifies the QP attributes to modify. On output,
1683 * the current values of selected QP attributes are returned.
1684 * @attr_mask: A bit-mask used to specify which attributes of the QP
1685 * are being modified.
1686 * @udata: pointer to user's input output buffer information
1687 * are being modified.
1688 * It returns 0 on success and returns appropriate error code on error.
1690 int ib_modify_qp_with_udata(struct ib_qp *ib_qp, struct ib_qp_attr *attr,
1691 int attr_mask, struct ib_udata *udata)
1693 return _ib_modify_qp(ib_qp->real_qp, attr, attr_mask, udata);
1695 EXPORT_SYMBOL(ib_modify_qp_with_udata);
1697 int ib_get_eth_speed(struct ib_device *dev, u8 port_num, u8 *speed, u8 *width)
1701 struct net_device *netdev;
1702 struct ethtool_link_ksettings lksettings;
1704 if (rdma_port_get_link_layer(dev, port_num) != IB_LINK_LAYER_ETHERNET)
1707 if (!dev->ops.get_netdev)
1710 netdev = dev->ops.get_netdev(dev, port_num);
1715 rc = __ethtool_get_link_ksettings(netdev, &lksettings);
1721 netdev_speed = lksettings.base.speed;
1723 netdev_speed = SPEED_1000;
1724 pr_warn("%s speed is unknown, defaulting to %d\n", netdev->name,
1728 if (netdev_speed <= SPEED_1000) {
1729 *width = IB_WIDTH_1X;
1730 *speed = IB_SPEED_SDR;
1731 } else if (netdev_speed <= SPEED_10000) {
1732 *width = IB_WIDTH_1X;
1733 *speed = IB_SPEED_FDR10;
1734 } else if (netdev_speed <= SPEED_20000) {
1735 *width = IB_WIDTH_4X;
1736 *speed = IB_SPEED_DDR;
1737 } else if (netdev_speed <= SPEED_25000) {
1738 *width = IB_WIDTH_1X;
1739 *speed = IB_SPEED_EDR;
1740 } else if (netdev_speed <= SPEED_40000) {
1741 *width = IB_WIDTH_4X;
1742 *speed = IB_SPEED_FDR10;
1744 *width = IB_WIDTH_4X;
1745 *speed = IB_SPEED_EDR;
1750 EXPORT_SYMBOL(ib_get_eth_speed);
1752 int ib_modify_qp(struct ib_qp *qp,
1753 struct ib_qp_attr *qp_attr,
1756 return _ib_modify_qp(qp->real_qp, qp_attr, qp_attr_mask, NULL);
1758 EXPORT_SYMBOL(ib_modify_qp);
1760 int ib_query_qp(struct ib_qp *qp,
1761 struct ib_qp_attr *qp_attr,
1763 struct ib_qp_init_attr *qp_init_attr)
1765 qp_attr->ah_attr.grh.sgid_attr = NULL;
1766 qp_attr->alt_ah_attr.grh.sgid_attr = NULL;
1768 return qp->device->ops.query_qp ?
1769 qp->device->ops.query_qp(qp->real_qp, qp_attr, qp_attr_mask,
1770 qp_init_attr) : -EOPNOTSUPP;
1772 EXPORT_SYMBOL(ib_query_qp);
1774 int ib_close_qp(struct ib_qp *qp)
1776 struct ib_qp *real_qp;
1777 unsigned long flags;
1779 real_qp = qp->real_qp;
1783 spin_lock_irqsave(&real_qp->device->event_handler_lock, flags);
1784 list_del(&qp->open_list);
1785 spin_unlock_irqrestore(&real_qp->device->event_handler_lock, flags);
1787 atomic_dec(&real_qp->usecnt);
1789 ib_close_shared_qp_security(qp->qp_sec);
1794 EXPORT_SYMBOL(ib_close_qp);
1796 static int __ib_destroy_shared_qp(struct ib_qp *qp)
1798 struct ib_xrcd *xrcd;
1799 struct ib_qp *real_qp;
1802 real_qp = qp->real_qp;
1803 xrcd = real_qp->xrcd;
1805 mutex_lock(&xrcd->tgt_qp_mutex);
1807 if (atomic_read(&real_qp->usecnt) == 0)
1808 list_del(&real_qp->xrcd_list);
1811 mutex_unlock(&xrcd->tgt_qp_mutex);
1814 ret = ib_destroy_qp(real_qp);
1816 atomic_dec(&xrcd->usecnt);
1818 __ib_insert_xrcd_qp(xrcd, real_qp);
1824 int ib_destroy_qp(struct ib_qp *qp)
1826 const struct ib_gid_attr *alt_path_sgid_attr = qp->alt_path_sgid_attr;
1827 const struct ib_gid_attr *av_sgid_attr = qp->av_sgid_attr;
1829 struct ib_cq *scq, *rcq;
1831 struct ib_rwq_ind_table *ind_tbl;
1832 struct ib_qp_security *sec;
1835 WARN_ON_ONCE(qp->mrs_used > 0);
1837 if (atomic_read(&qp->usecnt))
1840 if (qp->real_qp != qp)
1841 return __ib_destroy_shared_qp(qp);
1847 ind_tbl = qp->rwq_ind_tbl;
1850 ib_destroy_qp_security_begin(sec);
1853 rdma_rw_cleanup_mrs(qp);
1855 rdma_restrack_del(&qp->res);
1856 ret = qp->device->ops.destroy_qp(qp);
1858 if (alt_path_sgid_attr)
1859 rdma_put_gid_attr(alt_path_sgid_attr);
1861 rdma_put_gid_attr(av_sgid_attr);
1863 atomic_dec(&pd->usecnt);
1865 atomic_dec(&scq->usecnt);
1867 atomic_dec(&rcq->usecnt);
1869 atomic_dec(&srq->usecnt);
1871 atomic_dec(&ind_tbl->usecnt);
1873 ib_destroy_qp_security_end(sec);
1876 ib_destroy_qp_security_abort(sec);
1881 EXPORT_SYMBOL(ib_destroy_qp);
1883 /* Completion queues */
1885 struct ib_cq *__ib_create_cq(struct ib_device *device,
1886 ib_comp_handler comp_handler,
1887 void (*event_handler)(struct ib_event *, void *),
1889 const struct ib_cq_init_attr *cq_attr,
1894 cq = device->ops.create_cq(device, cq_attr, NULL, NULL);
1897 cq->device = device;
1899 cq->comp_handler = comp_handler;
1900 cq->event_handler = event_handler;
1901 cq->cq_context = cq_context;
1902 atomic_set(&cq->usecnt, 0);
1903 cq->res.type = RDMA_RESTRACK_CQ;
1904 rdma_restrack_set_task(&cq->res, caller);
1905 rdma_restrack_kadd(&cq->res);
1910 EXPORT_SYMBOL(__ib_create_cq);
1912 int rdma_set_cq_moderation(struct ib_cq *cq, u16 cq_count, u16 cq_period)
1914 return cq->device->ops.modify_cq ?
1915 cq->device->ops.modify_cq(cq, cq_count,
1916 cq_period) : -EOPNOTSUPP;
1918 EXPORT_SYMBOL(rdma_set_cq_moderation);
1920 int ib_destroy_cq(struct ib_cq *cq)
1922 if (atomic_read(&cq->usecnt))
1925 rdma_restrack_del(&cq->res);
1926 return cq->device->ops.destroy_cq(cq);
1928 EXPORT_SYMBOL(ib_destroy_cq);
1930 int ib_resize_cq(struct ib_cq *cq, int cqe)
1932 return cq->device->ops.resize_cq ?
1933 cq->device->ops.resize_cq(cq, cqe, NULL) : -EOPNOTSUPP;
1935 EXPORT_SYMBOL(ib_resize_cq);
1937 /* Memory regions */
1939 int ib_dereg_mr(struct ib_mr *mr)
1941 struct ib_pd *pd = mr->pd;
1942 struct ib_dm *dm = mr->dm;
1945 rdma_restrack_del(&mr->res);
1946 ret = mr->device->ops.dereg_mr(mr);
1948 atomic_dec(&pd->usecnt);
1950 atomic_dec(&dm->usecnt);
1955 EXPORT_SYMBOL(ib_dereg_mr);
1958 * ib_alloc_mr() - Allocates a memory region
1959 * @pd: protection domain associated with the region
1960 * @mr_type: memory region type
1961 * @max_num_sg: maximum sg entries available for registration.
1964 * Memory registeration page/sg lists must not exceed max_num_sg.
1965 * For mr_type IB_MR_TYPE_MEM_REG, the total length cannot exceed
1966 * max_num_sg * used_page_size.
1969 struct ib_mr *ib_alloc_mr(struct ib_pd *pd,
1970 enum ib_mr_type mr_type,
1975 if (!pd->device->ops.alloc_mr)
1976 return ERR_PTR(-EOPNOTSUPP);
1978 mr = pd->device->ops.alloc_mr(pd, mr_type, max_num_sg);
1980 mr->device = pd->device;
1984 atomic_inc(&pd->usecnt);
1985 mr->need_inval = false;
1986 mr->res.type = RDMA_RESTRACK_MR;
1987 rdma_restrack_kadd(&mr->res);
1992 EXPORT_SYMBOL(ib_alloc_mr);
1994 /* "Fast" memory regions */
1996 struct ib_fmr *ib_alloc_fmr(struct ib_pd *pd,
1997 int mr_access_flags,
1998 struct ib_fmr_attr *fmr_attr)
2002 if (!pd->device->ops.alloc_fmr)
2003 return ERR_PTR(-EOPNOTSUPP);
2005 fmr = pd->device->ops.alloc_fmr(pd, mr_access_flags, fmr_attr);
2007 fmr->device = pd->device;
2009 atomic_inc(&pd->usecnt);
2014 EXPORT_SYMBOL(ib_alloc_fmr);
2016 int ib_unmap_fmr(struct list_head *fmr_list)
2020 if (list_empty(fmr_list))
2023 fmr = list_entry(fmr_list->next, struct ib_fmr, list);
2024 return fmr->device->ops.unmap_fmr(fmr_list);
2026 EXPORT_SYMBOL(ib_unmap_fmr);
2028 int ib_dealloc_fmr(struct ib_fmr *fmr)
2034 ret = fmr->device->ops.dealloc_fmr(fmr);
2036 atomic_dec(&pd->usecnt);
2040 EXPORT_SYMBOL(ib_dealloc_fmr);
2042 /* Multicast groups */
2044 static bool is_valid_mcast_lid(struct ib_qp *qp, u16 lid)
2046 struct ib_qp_init_attr init_attr = {};
2047 struct ib_qp_attr attr = {};
2048 int num_eth_ports = 0;
2051 /* If QP state >= init, it is assigned to a port and we can check this
2054 if (!ib_query_qp(qp, &attr, IB_QP_STATE | IB_QP_PORT, &init_attr)) {
2055 if (attr.qp_state >= IB_QPS_INIT) {
2056 if (rdma_port_get_link_layer(qp->device, attr.port_num) !=
2057 IB_LINK_LAYER_INFINIBAND)
2063 /* Can't get a quick answer, iterate over all ports */
2064 for (port = 0; port < qp->device->phys_port_cnt; port++)
2065 if (rdma_port_get_link_layer(qp->device, port) !=
2066 IB_LINK_LAYER_INFINIBAND)
2069 /* If we have at lease one Ethernet port, RoCE annex declares that
2070 * multicast LID should be ignored. We can't tell at this step if the
2071 * QP belongs to an IB or Ethernet port.
2076 /* If all the ports are IB, we can check according to IB spec. */
2078 return !(lid < be16_to_cpu(IB_MULTICAST_LID_BASE) ||
2079 lid == be16_to_cpu(IB_LID_PERMISSIVE));
2082 int ib_attach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid)
2086 if (!qp->device->ops.attach_mcast)
2089 if (!rdma_is_multicast_addr((struct in6_addr *)gid->raw) ||
2090 qp->qp_type != IB_QPT_UD || !is_valid_mcast_lid(qp, lid))
2093 ret = qp->device->ops.attach_mcast(qp, gid, lid);
2095 atomic_inc(&qp->usecnt);
2098 EXPORT_SYMBOL(ib_attach_mcast);
2100 int ib_detach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid)
2104 if (!qp->device->ops.detach_mcast)
2107 if (!rdma_is_multicast_addr((struct in6_addr *)gid->raw) ||
2108 qp->qp_type != IB_QPT_UD || !is_valid_mcast_lid(qp, lid))
2111 ret = qp->device->ops.detach_mcast(qp, gid, lid);
2113 atomic_dec(&qp->usecnt);
2116 EXPORT_SYMBOL(ib_detach_mcast);
2118 struct ib_xrcd *__ib_alloc_xrcd(struct ib_device *device, const char *caller)
2120 struct ib_xrcd *xrcd;
2122 if (!device->ops.alloc_xrcd)
2123 return ERR_PTR(-EOPNOTSUPP);
2125 xrcd = device->ops.alloc_xrcd(device, NULL, NULL);
2126 if (!IS_ERR(xrcd)) {
2127 xrcd->device = device;
2129 atomic_set(&xrcd->usecnt, 0);
2130 mutex_init(&xrcd->tgt_qp_mutex);
2131 INIT_LIST_HEAD(&xrcd->tgt_qp_list);
2136 EXPORT_SYMBOL(__ib_alloc_xrcd);
2138 int ib_dealloc_xrcd(struct ib_xrcd *xrcd)
2143 if (atomic_read(&xrcd->usecnt))
2146 while (!list_empty(&xrcd->tgt_qp_list)) {
2147 qp = list_entry(xrcd->tgt_qp_list.next, struct ib_qp, xrcd_list);
2148 ret = ib_destroy_qp(qp);
2153 return xrcd->device->ops.dealloc_xrcd(xrcd);
2155 EXPORT_SYMBOL(ib_dealloc_xrcd);
2158 * ib_create_wq - Creates a WQ associated with the specified protection
2160 * @pd: The protection domain associated with the WQ.
2161 * @wq_attr: A list of initial attributes required to create the
2162 * WQ. If WQ creation succeeds, then the attributes are updated to
2163 * the actual capabilities of the created WQ.
2165 * wq_attr->max_wr and wq_attr->max_sge determine
2166 * the requested size of the WQ, and set to the actual values allocated
2168 * If ib_create_wq() succeeds, then max_wr and max_sge will always be
2169 * at least as large as the requested values.
2171 struct ib_wq *ib_create_wq(struct ib_pd *pd,
2172 struct ib_wq_init_attr *wq_attr)
2176 if (!pd->device->ops.create_wq)
2177 return ERR_PTR(-EOPNOTSUPP);
2179 wq = pd->device->ops.create_wq(pd, wq_attr, NULL);
2181 wq->event_handler = wq_attr->event_handler;
2182 wq->wq_context = wq_attr->wq_context;
2183 wq->wq_type = wq_attr->wq_type;
2184 wq->cq = wq_attr->cq;
2185 wq->device = pd->device;
2188 atomic_inc(&pd->usecnt);
2189 atomic_inc(&wq_attr->cq->usecnt);
2190 atomic_set(&wq->usecnt, 0);
2194 EXPORT_SYMBOL(ib_create_wq);
2197 * ib_destroy_wq - Destroys the specified WQ.
2198 * @wq: The WQ to destroy.
2200 int ib_destroy_wq(struct ib_wq *wq)
2203 struct ib_cq *cq = wq->cq;
2204 struct ib_pd *pd = wq->pd;
2206 if (atomic_read(&wq->usecnt))
2209 err = wq->device->ops.destroy_wq(wq);
2211 atomic_dec(&pd->usecnt);
2212 atomic_dec(&cq->usecnt);
2216 EXPORT_SYMBOL(ib_destroy_wq);
2219 * ib_modify_wq - Modifies the specified WQ.
2220 * @wq: The WQ to modify.
2221 * @wq_attr: On input, specifies the WQ attributes to modify.
2222 * @wq_attr_mask: A bit-mask used to specify which attributes of the WQ
2223 * are being modified.
2224 * On output, the current values of selected WQ attributes are returned.
2226 int ib_modify_wq(struct ib_wq *wq, struct ib_wq_attr *wq_attr,
2231 if (!wq->device->ops.modify_wq)
2234 err = wq->device->ops.modify_wq(wq, wq_attr, wq_attr_mask, NULL);
2237 EXPORT_SYMBOL(ib_modify_wq);
2240 * ib_create_rwq_ind_table - Creates a RQ Indirection Table.
2241 * @device: The device on which to create the rwq indirection table.
2242 * @ib_rwq_ind_table_init_attr: A list of initial attributes required to
2243 * create the Indirection Table.
2245 * Note: The life time of ib_rwq_ind_table_init_attr->ind_tbl is not less
2246 * than the created ib_rwq_ind_table object and the caller is responsible
2247 * for its memory allocation/free.
2249 struct ib_rwq_ind_table *ib_create_rwq_ind_table(struct ib_device *device,
2250 struct ib_rwq_ind_table_init_attr *init_attr)
2252 struct ib_rwq_ind_table *rwq_ind_table;
2256 if (!device->ops.create_rwq_ind_table)
2257 return ERR_PTR(-EOPNOTSUPP);
2259 table_size = (1 << init_attr->log_ind_tbl_size);
2260 rwq_ind_table = device->ops.create_rwq_ind_table(device,
2262 if (IS_ERR(rwq_ind_table))
2263 return rwq_ind_table;
2265 rwq_ind_table->ind_tbl = init_attr->ind_tbl;
2266 rwq_ind_table->log_ind_tbl_size = init_attr->log_ind_tbl_size;
2267 rwq_ind_table->device = device;
2268 rwq_ind_table->uobject = NULL;
2269 atomic_set(&rwq_ind_table->usecnt, 0);
2271 for (i = 0; i < table_size; i++)
2272 atomic_inc(&rwq_ind_table->ind_tbl[i]->usecnt);
2274 return rwq_ind_table;
2276 EXPORT_SYMBOL(ib_create_rwq_ind_table);
2279 * ib_destroy_rwq_ind_table - Destroys the specified Indirection Table.
2280 * @wq_ind_table: The Indirection Table to destroy.
2282 int ib_destroy_rwq_ind_table(struct ib_rwq_ind_table *rwq_ind_table)
2285 u32 table_size = (1 << rwq_ind_table->log_ind_tbl_size);
2286 struct ib_wq **ind_tbl = rwq_ind_table->ind_tbl;
2288 if (atomic_read(&rwq_ind_table->usecnt))
2291 err = rwq_ind_table->device->ops.destroy_rwq_ind_table(rwq_ind_table);
2293 for (i = 0; i < table_size; i++)
2294 atomic_dec(&ind_tbl[i]->usecnt);
2299 EXPORT_SYMBOL(ib_destroy_rwq_ind_table);
2301 int ib_check_mr_status(struct ib_mr *mr, u32 check_mask,
2302 struct ib_mr_status *mr_status)
2304 if (!mr->device->ops.check_mr_status)
2307 return mr->device->ops.check_mr_status(mr, check_mask, mr_status);
2309 EXPORT_SYMBOL(ib_check_mr_status);
2311 int ib_set_vf_link_state(struct ib_device *device, int vf, u8 port,
2314 if (!device->ops.set_vf_link_state)
2317 return device->ops.set_vf_link_state(device, vf, port, state);
2319 EXPORT_SYMBOL(ib_set_vf_link_state);
2321 int ib_get_vf_config(struct ib_device *device, int vf, u8 port,
2322 struct ifla_vf_info *info)
2324 if (!device->ops.get_vf_config)
2327 return device->ops.get_vf_config(device, vf, port, info);
2329 EXPORT_SYMBOL(ib_get_vf_config);
2331 int ib_get_vf_stats(struct ib_device *device, int vf, u8 port,
2332 struct ifla_vf_stats *stats)
2334 if (!device->ops.get_vf_stats)
2337 return device->ops.get_vf_stats(device, vf, port, stats);
2339 EXPORT_SYMBOL(ib_get_vf_stats);
2341 int ib_set_vf_guid(struct ib_device *device, int vf, u8 port, u64 guid,
2344 if (!device->ops.set_vf_guid)
2347 return device->ops.set_vf_guid(device, vf, port, guid, type);
2349 EXPORT_SYMBOL(ib_set_vf_guid);
2352 * ib_map_mr_sg() - Map the largest prefix of a dma mapped SG list
2353 * and set it the memory region.
2354 * @mr: memory region
2355 * @sg: dma mapped scatterlist
2356 * @sg_nents: number of entries in sg
2357 * @sg_offset: offset in bytes into sg
2358 * @page_size: page vector desired page size
2361 * - The first sg element is allowed to have an offset.
2362 * - Each sg element must either be aligned to page_size or virtually
2363 * contiguous to the previous element. In case an sg element has a
2364 * non-contiguous offset, the mapping prefix will not include it.
2365 * - The last sg element is allowed to have length less than page_size.
2366 * - If sg_nents total byte length exceeds the mr max_num_sge * page_size
2367 * then only max_num_sg entries will be mapped.
2368 * - If the MR was allocated with type IB_MR_TYPE_SG_GAPS, none of these
2369 * constraints holds and the page_size argument is ignored.
2371 * Returns the number of sg elements that were mapped to the memory region.
2373 * After this completes successfully, the memory region
2374 * is ready for registration.
2376 int ib_map_mr_sg(struct ib_mr *mr, struct scatterlist *sg, int sg_nents,
2377 unsigned int *sg_offset, unsigned int page_size)
2379 if (unlikely(!mr->device->ops.map_mr_sg))
2382 mr->page_size = page_size;
2384 return mr->device->ops.map_mr_sg(mr, sg, sg_nents, sg_offset);
2386 EXPORT_SYMBOL(ib_map_mr_sg);
2389 * ib_sg_to_pages() - Convert the largest prefix of a sg list
2391 * @mr: memory region
2392 * @sgl: dma mapped scatterlist
2393 * @sg_nents: number of entries in sg
2394 * @sg_offset_p: IN: start offset in bytes into sg
2395 * OUT: offset in bytes for element n of the sg of the first
2396 * byte that has not been processed where n is the return
2397 * value of this function.
2398 * @set_page: driver page assignment function pointer
2400 * Core service helper for drivers to convert the largest
2401 * prefix of given sg list to a page vector. The sg list
2402 * prefix converted is the prefix that meet the requirements
2405 * Returns the number of sg elements that were assigned to
2408 int ib_sg_to_pages(struct ib_mr *mr, struct scatterlist *sgl, int sg_nents,
2409 unsigned int *sg_offset_p, int (*set_page)(struct ib_mr *, u64))
2411 struct scatterlist *sg;
2412 u64 last_end_dma_addr = 0;
2413 unsigned int sg_offset = sg_offset_p ? *sg_offset_p : 0;
2414 unsigned int last_page_off = 0;
2415 u64 page_mask = ~((u64)mr->page_size - 1);
2418 if (unlikely(sg_nents <= 0 || sg_offset > sg_dma_len(&sgl[0])))
2421 mr->iova = sg_dma_address(&sgl[0]) + sg_offset;
2424 for_each_sg(sgl, sg, sg_nents, i) {
2425 u64 dma_addr = sg_dma_address(sg) + sg_offset;
2426 u64 prev_addr = dma_addr;
2427 unsigned int dma_len = sg_dma_len(sg) - sg_offset;
2428 u64 end_dma_addr = dma_addr + dma_len;
2429 u64 page_addr = dma_addr & page_mask;
2432 * For the second and later elements, check whether either the
2433 * end of element i-1 or the start of element i is not aligned
2434 * on a page boundary.
2436 if (i && (last_page_off != 0 || page_addr != dma_addr)) {
2437 /* Stop mapping if there is a gap. */
2438 if (last_end_dma_addr != dma_addr)
2442 * Coalesce this element with the last. If it is small
2443 * enough just update mr->length. Otherwise start
2444 * mapping from the next page.
2450 ret = set_page(mr, page_addr);
2451 if (unlikely(ret < 0)) {
2452 sg_offset = prev_addr - sg_dma_address(sg);
2453 mr->length += prev_addr - dma_addr;
2455 *sg_offset_p = sg_offset;
2456 return i || sg_offset ? i : ret;
2458 prev_addr = page_addr;
2460 page_addr += mr->page_size;
2461 } while (page_addr < end_dma_addr);
2463 mr->length += dma_len;
2464 last_end_dma_addr = end_dma_addr;
2465 last_page_off = end_dma_addr & ~page_mask;
2474 EXPORT_SYMBOL(ib_sg_to_pages);
2476 struct ib_drain_cqe {
2478 struct completion done;
2481 static void ib_drain_qp_done(struct ib_cq *cq, struct ib_wc *wc)
2483 struct ib_drain_cqe *cqe = container_of(wc->wr_cqe, struct ib_drain_cqe,
2486 complete(&cqe->done);
2490 * Post a WR and block until its completion is reaped for the SQ.
2492 static void __ib_drain_sq(struct ib_qp *qp)
2494 struct ib_cq *cq = qp->send_cq;
2495 struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
2496 struct ib_drain_cqe sdrain;
2497 struct ib_rdma_wr swr = {
2500 { .wr_cqe = &sdrain.cqe, },
2501 .opcode = IB_WR_RDMA_WRITE,
2506 ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
2508 WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
2512 sdrain.cqe.done = ib_drain_qp_done;
2513 init_completion(&sdrain.done);
2515 ret = ib_post_send(qp, &swr.wr, NULL);
2517 WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
2521 if (cq->poll_ctx == IB_POLL_DIRECT)
2522 while (wait_for_completion_timeout(&sdrain.done, HZ / 10) <= 0)
2523 ib_process_cq_direct(cq, -1);
2525 wait_for_completion(&sdrain.done);
2529 * Post a WR and block until its completion is reaped for the RQ.
2531 static void __ib_drain_rq(struct ib_qp *qp)
2533 struct ib_cq *cq = qp->recv_cq;
2534 struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
2535 struct ib_drain_cqe rdrain;
2536 struct ib_recv_wr rwr = {};
2539 ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
2541 WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
2545 rwr.wr_cqe = &rdrain.cqe;
2546 rdrain.cqe.done = ib_drain_qp_done;
2547 init_completion(&rdrain.done);
2549 ret = ib_post_recv(qp, &rwr, NULL);
2551 WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
2555 if (cq->poll_ctx == IB_POLL_DIRECT)
2556 while (wait_for_completion_timeout(&rdrain.done, HZ / 10) <= 0)
2557 ib_process_cq_direct(cq, -1);
2559 wait_for_completion(&rdrain.done);
2563 * ib_drain_sq() - Block until all SQ CQEs have been consumed by the
2565 * @qp: queue pair to drain
2567 * If the device has a provider-specific drain function, then
2568 * call that. Otherwise call the generic drain function
2573 * ensure there is room in the CQ and SQ for the drain work request and
2576 * allocate the CQ using ib_alloc_cq().
2578 * ensure that there are no other contexts that are posting WRs concurrently.
2579 * Otherwise the drain is not guaranteed.
2581 void ib_drain_sq(struct ib_qp *qp)
2583 if (qp->device->ops.drain_sq)
2584 qp->device->ops.drain_sq(qp);
2588 EXPORT_SYMBOL(ib_drain_sq);
2591 * ib_drain_rq() - Block until all RQ CQEs have been consumed by the
2593 * @qp: queue pair to drain
2595 * If the device has a provider-specific drain function, then
2596 * call that. Otherwise call the generic drain function
2601 * ensure there is room in the CQ and RQ for the drain work request and
2604 * allocate the CQ using ib_alloc_cq().
2606 * ensure that there are no other contexts that are posting WRs concurrently.
2607 * Otherwise the drain is not guaranteed.
2609 void ib_drain_rq(struct ib_qp *qp)
2611 if (qp->device->ops.drain_rq)
2612 qp->device->ops.drain_rq(qp);
2616 EXPORT_SYMBOL(ib_drain_rq);
2619 * ib_drain_qp() - Block until all CQEs have been consumed by the
2620 * application on both the RQ and SQ.
2621 * @qp: queue pair to drain
2625 * ensure there is room in the CQ(s), SQ, and RQ for drain work requests
2628 * allocate the CQs using ib_alloc_cq().
2630 * ensure that there are no other contexts that are posting WRs concurrently.
2631 * Otherwise the drain is not guaranteed.
2633 void ib_drain_qp(struct ib_qp *qp)
2639 EXPORT_SYMBOL(ib_drain_qp);
2641 struct net_device *rdma_alloc_netdev(struct ib_device *device, u8 port_num,
2642 enum rdma_netdev_t type, const char *name,
2643 unsigned char name_assign_type,
2644 void (*setup)(struct net_device *))
2646 struct rdma_netdev_alloc_params params;
2647 struct net_device *netdev;
2650 if (!device->ops.rdma_netdev_get_params)
2651 return ERR_PTR(-EOPNOTSUPP);
2653 rc = device->ops.rdma_netdev_get_params(device, port_num, type,
2658 netdev = alloc_netdev_mqs(params.sizeof_priv, name, name_assign_type,
2659 setup, params.txqs, params.rxqs);
2661 return ERR_PTR(-ENOMEM);
2665 EXPORT_SYMBOL(rdma_alloc_netdev);
2667 int rdma_init_netdev(struct ib_device *device, u8 port_num,
2668 enum rdma_netdev_t type, const char *name,
2669 unsigned char name_assign_type,
2670 void (*setup)(struct net_device *),
2671 struct net_device *netdev)
2673 struct rdma_netdev_alloc_params params;
2676 if (!device->ops.rdma_netdev_get_params)
2679 rc = device->ops.rdma_netdev_get_params(device, port_num, type,
2684 return params.initialize_rdma_netdev(device, port_num,
2685 netdev, params.param);
2687 EXPORT_SYMBOL(rdma_init_netdev);