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(unsigned int 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;
221 if (node_type == RDMA_NODE_UNSPECIFIED)
222 return RDMA_TRANSPORT_UNSPECIFIED;
224 return RDMA_TRANSPORT_IB;
226 EXPORT_SYMBOL(rdma_node_get_transport);
228 enum rdma_link_layer rdma_port_get_link_layer(struct ib_device *device, u8 port_num)
230 enum rdma_transport_type lt;
231 if (device->ops.get_link_layer)
232 return device->ops.get_link_layer(device, port_num);
234 lt = rdma_node_get_transport(device->node_type);
235 if (lt == RDMA_TRANSPORT_IB)
236 return IB_LINK_LAYER_INFINIBAND;
238 return IB_LINK_LAYER_ETHERNET;
240 EXPORT_SYMBOL(rdma_port_get_link_layer);
242 /* Protection domains */
245 * ib_alloc_pd - Allocates an unused protection domain.
246 * @device: The device on which to allocate the protection domain.
248 * A protection domain object provides an association between QPs, shared
249 * receive queues, address handles, memory regions, and memory windows.
251 * Every PD has a local_dma_lkey which can be used as the lkey value for local
254 struct ib_pd *__ib_alloc_pd(struct ib_device *device, unsigned int flags,
258 int mr_access_flags = 0;
261 pd = rdma_zalloc_drv_obj(device, ib_pd);
263 return ERR_PTR(-ENOMEM);
267 pd->__internal_mr = NULL;
268 atomic_set(&pd->usecnt, 0);
271 pd->res.type = RDMA_RESTRACK_PD;
272 rdma_restrack_set_task(&pd->res, caller);
274 ret = device->ops.alloc_pd(pd, NULL);
279 rdma_restrack_kadd(&pd->res);
281 if (device->attrs.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY)
282 pd->local_dma_lkey = device->local_dma_lkey;
284 mr_access_flags |= IB_ACCESS_LOCAL_WRITE;
286 if (flags & IB_PD_UNSAFE_GLOBAL_RKEY) {
287 pr_warn("%s: enabling unsafe global rkey\n", caller);
288 mr_access_flags |= IB_ACCESS_REMOTE_READ | IB_ACCESS_REMOTE_WRITE;
291 if (mr_access_flags) {
294 mr = pd->device->ops.get_dma_mr(pd, mr_access_flags);
300 mr->device = pd->device;
302 mr->type = IB_MR_TYPE_DMA;
304 mr->need_inval = false;
306 pd->__internal_mr = mr;
308 if (!(device->attrs.device_cap_flags & IB_DEVICE_LOCAL_DMA_LKEY))
309 pd->local_dma_lkey = pd->__internal_mr->lkey;
311 if (flags & IB_PD_UNSAFE_GLOBAL_RKEY)
312 pd->unsafe_global_rkey = pd->__internal_mr->rkey;
317 EXPORT_SYMBOL(__ib_alloc_pd);
320 * ib_dealloc_pd_user - Deallocates a protection domain.
321 * @pd: The protection domain to deallocate.
322 * @udata: Valid user data or NULL for kernel object
324 * It is an error to call this function while any resources in the pd still
325 * exist. The caller is responsible to synchronously destroy them and
326 * guarantee no new allocations will happen.
328 void ib_dealloc_pd_user(struct ib_pd *pd, struct ib_udata *udata)
332 if (pd->__internal_mr) {
333 ret = pd->device->ops.dereg_mr(pd->__internal_mr, NULL);
335 pd->__internal_mr = NULL;
338 /* uverbs manipulates usecnt with proper locking, while the kabi
339 requires the caller to guarantee we can't race here. */
340 WARN_ON(atomic_read(&pd->usecnt));
342 rdma_restrack_del(&pd->res);
343 pd->device->ops.dealloc_pd(pd, udata);
346 EXPORT_SYMBOL(ib_dealloc_pd_user);
348 /* Address handles */
351 * rdma_copy_ah_attr - Copy rdma ah attribute from source to destination.
352 * @dest: Pointer to destination ah_attr. Contents of the destination
353 * pointer is assumed to be invalid and attribute are overwritten.
354 * @src: Pointer to source ah_attr.
356 void rdma_copy_ah_attr(struct rdma_ah_attr *dest,
357 const struct rdma_ah_attr *src)
360 if (dest->grh.sgid_attr)
361 rdma_hold_gid_attr(dest->grh.sgid_attr);
363 EXPORT_SYMBOL(rdma_copy_ah_attr);
366 * rdma_replace_ah_attr - Replace valid ah_attr with new new one.
367 * @old: Pointer to existing ah_attr which needs to be replaced.
368 * old is assumed to be valid or zero'd
369 * @new: Pointer to the new ah_attr.
371 * rdma_replace_ah_attr() first releases any reference in the old ah_attr if
372 * old the ah_attr is valid; after that it copies the new attribute and holds
373 * the reference to the replaced ah_attr.
375 void rdma_replace_ah_attr(struct rdma_ah_attr *old,
376 const struct rdma_ah_attr *new)
378 rdma_destroy_ah_attr(old);
380 if (old->grh.sgid_attr)
381 rdma_hold_gid_attr(old->grh.sgid_attr);
383 EXPORT_SYMBOL(rdma_replace_ah_attr);
386 * rdma_move_ah_attr - Move ah_attr pointed by source to destination.
387 * @dest: Pointer to destination ah_attr to copy to.
388 * dest is assumed to be valid or zero'd
389 * @src: Pointer to the new ah_attr.
391 * rdma_move_ah_attr() first releases any reference in the destination ah_attr
392 * if it is valid. This also transfers ownership of internal references from
393 * src to dest, making src invalid in the process. No new reference of the src
396 void rdma_move_ah_attr(struct rdma_ah_attr *dest, struct rdma_ah_attr *src)
398 rdma_destroy_ah_attr(dest);
400 src->grh.sgid_attr = NULL;
402 EXPORT_SYMBOL(rdma_move_ah_attr);
405 * Validate that the rdma_ah_attr is valid for the device before passing it
408 static int rdma_check_ah_attr(struct ib_device *device,
409 struct rdma_ah_attr *ah_attr)
411 if (!rdma_is_port_valid(device, ah_attr->port_num))
414 if ((rdma_is_grh_required(device, ah_attr->port_num) ||
415 ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE) &&
416 !(ah_attr->ah_flags & IB_AH_GRH))
419 if (ah_attr->grh.sgid_attr) {
421 * Make sure the passed sgid_attr is consistent with the
424 if (ah_attr->grh.sgid_attr->index != ah_attr->grh.sgid_index ||
425 ah_attr->grh.sgid_attr->port_num != ah_attr->port_num)
432 * If the ah requires a GRH then ensure that sgid_attr pointer is filled in.
433 * On success the caller is responsible to call rdma_unfill_sgid_attr().
435 static int rdma_fill_sgid_attr(struct ib_device *device,
436 struct rdma_ah_attr *ah_attr,
437 const struct ib_gid_attr **old_sgid_attr)
439 const struct ib_gid_attr *sgid_attr;
440 struct ib_global_route *grh;
443 *old_sgid_attr = ah_attr->grh.sgid_attr;
445 ret = rdma_check_ah_attr(device, ah_attr);
449 if (!(ah_attr->ah_flags & IB_AH_GRH))
452 grh = rdma_ah_retrieve_grh(ah_attr);
457 rdma_get_gid_attr(device, ah_attr->port_num, grh->sgid_index);
458 if (IS_ERR(sgid_attr))
459 return PTR_ERR(sgid_attr);
461 /* Move ownerhip of the kref into the ah_attr */
462 grh->sgid_attr = sgid_attr;
466 static void rdma_unfill_sgid_attr(struct rdma_ah_attr *ah_attr,
467 const struct ib_gid_attr *old_sgid_attr)
470 * Fill didn't change anything, the caller retains ownership of
473 if (ah_attr->grh.sgid_attr == old_sgid_attr)
477 * Otherwise, we need to undo what rdma_fill_sgid_attr so the caller
478 * doesn't see any change in the rdma_ah_attr. If we get here
479 * old_sgid_attr is NULL.
481 rdma_destroy_ah_attr(ah_attr);
484 static const struct ib_gid_attr *
485 rdma_update_sgid_attr(struct rdma_ah_attr *ah_attr,
486 const struct ib_gid_attr *old_attr)
489 rdma_put_gid_attr(old_attr);
490 if (ah_attr->ah_flags & IB_AH_GRH) {
491 rdma_hold_gid_attr(ah_attr->grh.sgid_attr);
492 return ah_attr->grh.sgid_attr;
497 static struct ib_ah *_rdma_create_ah(struct ib_pd *pd,
498 struct rdma_ah_attr *ah_attr,
500 struct ib_udata *udata)
502 struct ib_device *device = pd->device;
506 might_sleep_if(flags & RDMA_CREATE_AH_SLEEPABLE);
508 if (!device->ops.create_ah)
509 return ERR_PTR(-EOPNOTSUPP);
511 ah = rdma_zalloc_drv_obj_gfp(
513 (flags & RDMA_CREATE_AH_SLEEPABLE) ? GFP_KERNEL : GFP_ATOMIC);
515 return ERR_PTR(-ENOMEM);
519 ah->type = ah_attr->type;
520 ah->sgid_attr = rdma_update_sgid_attr(ah_attr, NULL);
522 ret = device->ops.create_ah(ah, ah_attr, flags, udata);
528 atomic_inc(&pd->usecnt);
533 * rdma_create_ah - Creates an address handle for the
534 * given address vector.
535 * @pd: The protection domain associated with the address handle.
536 * @ah_attr: The attributes of the address vector.
537 * @flags: Create address handle flags (see enum rdma_create_ah_flags).
539 * It returns 0 on success and returns appropriate error code on error.
540 * The address handle is used to reference a local or global destination
541 * in all UD QP post sends.
543 struct ib_ah *rdma_create_ah(struct ib_pd *pd, struct rdma_ah_attr *ah_attr,
546 const struct ib_gid_attr *old_sgid_attr;
550 ret = rdma_fill_sgid_attr(pd->device, ah_attr, &old_sgid_attr);
554 ah = _rdma_create_ah(pd, ah_attr, flags, NULL);
556 rdma_unfill_sgid_attr(ah_attr, old_sgid_attr);
559 EXPORT_SYMBOL(rdma_create_ah);
562 * rdma_create_user_ah - Creates an address handle for the
563 * given address vector.
564 * It resolves destination mac address for ah attribute of RoCE type.
565 * @pd: The protection domain associated with the address handle.
566 * @ah_attr: The attributes of the address vector.
567 * @udata: pointer to user's input output buffer information need by
570 * It returns 0 on success and returns appropriate error code on error.
571 * The address handle is used to reference a local or global destination
572 * in all UD QP post sends.
574 struct ib_ah *rdma_create_user_ah(struct ib_pd *pd,
575 struct rdma_ah_attr *ah_attr,
576 struct ib_udata *udata)
578 const struct ib_gid_attr *old_sgid_attr;
582 err = rdma_fill_sgid_attr(pd->device, ah_attr, &old_sgid_attr);
586 if (ah_attr->type == RDMA_AH_ATTR_TYPE_ROCE) {
587 err = ib_resolve_eth_dmac(pd->device, ah_attr);
594 ah = _rdma_create_ah(pd, ah_attr, RDMA_CREATE_AH_SLEEPABLE, udata);
597 rdma_unfill_sgid_attr(ah_attr, old_sgid_attr);
600 EXPORT_SYMBOL(rdma_create_user_ah);
602 int ib_get_rdma_header_version(const union rdma_network_hdr *hdr)
604 const struct iphdr *ip4h = (struct iphdr *)&hdr->roce4grh;
605 struct iphdr ip4h_checked;
606 const struct ipv6hdr *ip6h = (struct ipv6hdr *)&hdr->ibgrh;
608 /* If it's IPv6, the version must be 6, otherwise, the first
609 * 20 bytes (before the IPv4 header) are garbled.
611 if (ip6h->version != 6)
612 return (ip4h->version == 4) ? 4 : 0;
613 /* version may be 6 or 4 because the first 20 bytes could be garbled */
615 /* RoCE v2 requires no options, thus header length
622 * We can't write on scattered buffers so we need to copy to
625 memcpy(&ip4h_checked, ip4h, sizeof(ip4h_checked));
626 ip4h_checked.check = 0;
627 ip4h_checked.check = ip_fast_csum((u8 *)&ip4h_checked, 5);
628 /* if IPv4 header checksum is OK, believe it */
629 if (ip4h->check == ip4h_checked.check)
633 EXPORT_SYMBOL(ib_get_rdma_header_version);
635 static enum rdma_network_type ib_get_net_type_by_grh(struct ib_device *device,
637 const struct ib_grh *grh)
641 if (rdma_protocol_ib(device, port_num))
642 return RDMA_NETWORK_IB;
644 grh_version = ib_get_rdma_header_version((union rdma_network_hdr *)grh);
646 if (grh_version == 4)
647 return RDMA_NETWORK_IPV4;
649 if (grh->next_hdr == IPPROTO_UDP)
650 return RDMA_NETWORK_IPV6;
652 return RDMA_NETWORK_ROCE_V1;
655 struct find_gid_index_context {
657 enum ib_gid_type gid_type;
660 static bool find_gid_index(const union ib_gid *gid,
661 const struct ib_gid_attr *gid_attr,
664 struct find_gid_index_context *ctx = context;
666 if (ctx->gid_type != gid_attr->gid_type)
669 if ((!!(ctx->vlan_id != 0xffff) == !is_vlan_dev(gid_attr->ndev)) ||
670 (is_vlan_dev(gid_attr->ndev) &&
671 vlan_dev_vlan_id(gid_attr->ndev) != ctx->vlan_id))
677 static const struct ib_gid_attr *
678 get_sgid_attr_from_eth(struct ib_device *device, u8 port_num,
679 u16 vlan_id, const union ib_gid *sgid,
680 enum ib_gid_type gid_type)
682 struct find_gid_index_context context = {.vlan_id = vlan_id,
683 .gid_type = gid_type};
685 return rdma_find_gid_by_filter(device, sgid, port_num, find_gid_index,
689 int ib_get_gids_from_rdma_hdr(const union rdma_network_hdr *hdr,
690 enum rdma_network_type net_type,
691 union ib_gid *sgid, union ib_gid *dgid)
693 struct sockaddr_in src_in;
694 struct sockaddr_in dst_in;
695 __be32 src_saddr, dst_saddr;
700 if (net_type == RDMA_NETWORK_IPV4) {
701 memcpy(&src_in.sin_addr.s_addr,
702 &hdr->roce4grh.saddr, 4);
703 memcpy(&dst_in.sin_addr.s_addr,
704 &hdr->roce4grh.daddr, 4);
705 src_saddr = src_in.sin_addr.s_addr;
706 dst_saddr = dst_in.sin_addr.s_addr;
707 ipv6_addr_set_v4mapped(src_saddr,
708 (struct in6_addr *)sgid);
709 ipv6_addr_set_v4mapped(dst_saddr,
710 (struct in6_addr *)dgid);
712 } else if (net_type == RDMA_NETWORK_IPV6 ||
713 net_type == RDMA_NETWORK_IB) {
714 *dgid = hdr->ibgrh.dgid;
715 *sgid = hdr->ibgrh.sgid;
721 EXPORT_SYMBOL(ib_get_gids_from_rdma_hdr);
723 /* Resolve destination mac address and hop limit for unicast destination
724 * GID entry, considering the source GID entry as well.
725 * ah_attribute must have have valid port_num, sgid_index.
727 static int ib_resolve_unicast_gid_dmac(struct ib_device *device,
728 struct rdma_ah_attr *ah_attr)
730 struct ib_global_route *grh = rdma_ah_retrieve_grh(ah_attr);
731 const struct ib_gid_attr *sgid_attr = grh->sgid_attr;
732 int hop_limit = 0xff;
735 /* If destination is link local and source GID is RoCEv1,
736 * IP stack is not used.
738 if (rdma_link_local_addr((struct in6_addr *)grh->dgid.raw) &&
739 sgid_attr->gid_type == IB_GID_TYPE_ROCE) {
740 rdma_get_ll_mac((struct in6_addr *)grh->dgid.raw,
745 ret = rdma_addr_find_l2_eth_by_grh(&sgid_attr->gid, &grh->dgid,
747 sgid_attr, &hop_limit);
749 grh->hop_limit = hop_limit;
754 * This function initializes address handle attributes from the incoming packet.
755 * Incoming packet has dgid of the receiver node on which this code is
756 * getting executed and, sgid contains the GID of the sender.
758 * When resolving mac address of destination, the arrived dgid is used
759 * as sgid and, sgid is used as dgid because sgid contains destinations
760 * GID whom to respond to.
762 * On success the caller is responsible to call rdma_destroy_ah_attr on the
765 int ib_init_ah_attr_from_wc(struct ib_device *device, u8 port_num,
766 const struct ib_wc *wc, const struct ib_grh *grh,
767 struct rdma_ah_attr *ah_attr)
771 enum rdma_network_type net_type = RDMA_NETWORK_IB;
772 enum ib_gid_type gid_type = IB_GID_TYPE_IB;
773 const struct ib_gid_attr *sgid_attr;
780 memset(ah_attr, 0, sizeof *ah_attr);
781 ah_attr->type = rdma_ah_find_type(device, port_num);
782 if (rdma_cap_eth_ah(device, port_num)) {
783 if (wc->wc_flags & IB_WC_WITH_NETWORK_HDR_TYPE)
784 net_type = wc->network_hdr_type;
786 net_type = ib_get_net_type_by_grh(device, port_num, grh);
787 gid_type = ib_network_to_gid_type(net_type);
789 ret = ib_get_gids_from_rdma_hdr((union rdma_network_hdr *)grh, net_type,
794 rdma_ah_set_sl(ah_attr, wc->sl);
795 rdma_ah_set_port_num(ah_attr, port_num);
797 if (rdma_protocol_roce(device, port_num)) {
798 u16 vlan_id = wc->wc_flags & IB_WC_WITH_VLAN ?
799 wc->vlan_id : 0xffff;
801 if (!(wc->wc_flags & IB_WC_GRH))
804 sgid_attr = get_sgid_attr_from_eth(device, port_num,
807 if (IS_ERR(sgid_attr))
808 return PTR_ERR(sgid_attr);
810 flow_class = be32_to_cpu(grh->version_tclass_flow);
811 rdma_move_grh_sgid_attr(ah_attr,
813 flow_class & 0xFFFFF,
815 (flow_class >> 20) & 0xFF,
818 ret = ib_resolve_unicast_gid_dmac(device, ah_attr);
820 rdma_destroy_ah_attr(ah_attr);
824 rdma_ah_set_dlid(ah_attr, wc->slid);
825 rdma_ah_set_path_bits(ah_attr, wc->dlid_path_bits);
827 if ((wc->wc_flags & IB_WC_GRH) == 0)
830 if (dgid.global.interface_id !=
831 cpu_to_be64(IB_SA_WELL_KNOWN_GUID)) {
832 sgid_attr = rdma_find_gid_by_port(
833 device, &dgid, IB_GID_TYPE_IB, port_num, NULL);
835 sgid_attr = rdma_get_gid_attr(device, port_num, 0);
837 if (IS_ERR(sgid_attr))
838 return PTR_ERR(sgid_attr);
839 flow_class = be32_to_cpu(grh->version_tclass_flow);
840 rdma_move_grh_sgid_attr(ah_attr,
842 flow_class & 0xFFFFF,
844 (flow_class >> 20) & 0xFF,
850 EXPORT_SYMBOL(ib_init_ah_attr_from_wc);
853 * rdma_move_grh_sgid_attr - Sets the sgid attribute of GRH, taking ownership
856 * @attr: Pointer to AH attribute structure
857 * @dgid: Destination GID
858 * @flow_label: Flow label
859 * @hop_limit: Hop limit
860 * @traffic_class: traffic class
861 * @sgid_attr: Pointer to SGID attribute
863 * This takes ownership of the sgid_attr reference. The caller must ensure
864 * rdma_destroy_ah_attr() is called before destroying the rdma_ah_attr after
865 * calling this function.
867 void rdma_move_grh_sgid_attr(struct rdma_ah_attr *attr, union ib_gid *dgid,
868 u32 flow_label, u8 hop_limit, u8 traffic_class,
869 const struct ib_gid_attr *sgid_attr)
871 rdma_ah_set_grh(attr, dgid, flow_label, sgid_attr->index, hop_limit,
873 attr->grh.sgid_attr = sgid_attr;
875 EXPORT_SYMBOL(rdma_move_grh_sgid_attr);
878 * rdma_destroy_ah_attr - Release reference to SGID attribute of
880 * @ah_attr: Pointer to ah attribute
882 * Release reference to the SGID attribute of the ah attribute if it is
883 * non NULL. It is safe to call this multiple times, and safe to call it on
884 * a zero initialized ah_attr.
886 void rdma_destroy_ah_attr(struct rdma_ah_attr *ah_attr)
888 if (ah_attr->grh.sgid_attr) {
889 rdma_put_gid_attr(ah_attr->grh.sgid_attr);
890 ah_attr->grh.sgid_attr = NULL;
893 EXPORT_SYMBOL(rdma_destroy_ah_attr);
895 struct ib_ah *ib_create_ah_from_wc(struct ib_pd *pd, const struct ib_wc *wc,
896 const struct ib_grh *grh, u8 port_num)
898 struct rdma_ah_attr ah_attr;
902 ret = ib_init_ah_attr_from_wc(pd->device, port_num, wc, grh, &ah_attr);
906 ah = rdma_create_ah(pd, &ah_attr, RDMA_CREATE_AH_SLEEPABLE);
908 rdma_destroy_ah_attr(&ah_attr);
911 EXPORT_SYMBOL(ib_create_ah_from_wc);
913 int rdma_modify_ah(struct ib_ah *ah, struct rdma_ah_attr *ah_attr)
915 const struct ib_gid_attr *old_sgid_attr;
918 if (ah->type != ah_attr->type)
921 ret = rdma_fill_sgid_attr(ah->device, ah_attr, &old_sgid_attr);
925 ret = ah->device->ops.modify_ah ?
926 ah->device->ops.modify_ah(ah, ah_attr) :
929 ah->sgid_attr = rdma_update_sgid_attr(ah_attr, ah->sgid_attr);
930 rdma_unfill_sgid_attr(ah_attr, old_sgid_attr);
933 EXPORT_SYMBOL(rdma_modify_ah);
935 int rdma_query_ah(struct ib_ah *ah, struct rdma_ah_attr *ah_attr)
937 ah_attr->grh.sgid_attr = NULL;
939 return ah->device->ops.query_ah ?
940 ah->device->ops.query_ah(ah, ah_attr) :
943 EXPORT_SYMBOL(rdma_query_ah);
945 int rdma_destroy_ah_user(struct ib_ah *ah, u32 flags, struct ib_udata *udata)
947 const struct ib_gid_attr *sgid_attr = ah->sgid_attr;
950 might_sleep_if(flags & RDMA_DESTROY_AH_SLEEPABLE);
954 ah->device->ops.destroy_ah(ah, flags);
955 atomic_dec(&pd->usecnt);
957 rdma_put_gid_attr(sgid_attr);
962 EXPORT_SYMBOL(rdma_destroy_ah_user);
964 /* Shared receive queues */
966 struct ib_srq *ib_create_srq(struct ib_pd *pd,
967 struct ib_srq_init_attr *srq_init_attr)
972 if (!pd->device->ops.create_srq)
973 return ERR_PTR(-EOPNOTSUPP);
975 srq = rdma_zalloc_drv_obj(pd->device, ib_srq);
977 return ERR_PTR(-ENOMEM);
979 srq->device = pd->device;
981 srq->event_handler = srq_init_attr->event_handler;
982 srq->srq_context = srq_init_attr->srq_context;
983 srq->srq_type = srq_init_attr->srq_type;
985 if (ib_srq_has_cq(srq->srq_type)) {
986 srq->ext.cq = srq_init_attr->ext.cq;
987 atomic_inc(&srq->ext.cq->usecnt);
989 if (srq->srq_type == IB_SRQT_XRC) {
990 srq->ext.xrc.xrcd = srq_init_attr->ext.xrc.xrcd;
991 atomic_inc(&srq->ext.xrc.xrcd->usecnt);
993 atomic_inc(&pd->usecnt);
995 ret = pd->device->ops.create_srq(srq, srq_init_attr, NULL);
997 atomic_dec(&srq->pd->usecnt);
998 if (srq->srq_type == IB_SRQT_XRC)
999 atomic_dec(&srq->ext.xrc.xrcd->usecnt);
1000 if (ib_srq_has_cq(srq->srq_type))
1001 atomic_dec(&srq->ext.cq->usecnt);
1003 return ERR_PTR(ret);
1008 EXPORT_SYMBOL(ib_create_srq);
1010 int ib_modify_srq(struct ib_srq *srq,
1011 struct ib_srq_attr *srq_attr,
1012 enum ib_srq_attr_mask srq_attr_mask)
1014 return srq->device->ops.modify_srq ?
1015 srq->device->ops.modify_srq(srq, srq_attr, srq_attr_mask,
1016 NULL) : -EOPNOTSUPP;
1018 EXPORT_SYMBOL(ib_modify_srq);
1020 int ib_query_srq(struct ib_srq *srq,
1021 struct ib_srq_attr *srq_attr)
1023 return srq->device->ops.query_srq ?
1024 srq->device->ops.query_srq(srq, srq_attr) : -EOPNOTSUPP;
1026 EXPORT_SYMBOL(ib_query_srq);
1028 int ib_destroy_srq_user(struct ib_srq *srq, struct ib_udata *udata)
1030 if (atomic_read(&srq->usecnt))
1033 srq->device->ops.destroy_srq(srq, udata);
1035 atomic_dec(&srq->pd->usecnt);
1036 if (srq->srq_type == IB_SRQT_XRC)
1037 atomic_dec(&srq->ext.xrc.xrcd->usecnt);
1038 if (ib_srq_has_cq(srq->srq_type))
1039 atomic_dec(&srq->ext.cq->usecnt);
1044 EXPORT_SYMBOL(ib_destroy_srq_user);
1048 static void __ib_shared_qp_event_handler(struct ib_event *event, void *context)
1050 struct ib_qp *qp = context;
1051 unsigned long flags;
1053 spin_lock_irqsave(&qp->device->event_handler_lock, flags);
1054 list_for_each_entry(event->element.qp, &qp->open_list, open_list)
1055 if (event->element.qp->event_handler)
1056 event->element.qp->event_handler(event, event->element.qp->qp_context);
1057 spin_unlock_irqrestore(&qp->device->event_handler_lock, flags);
1060 static void __ib_insert_xrcd_qp(struct ib_xrcd *xrcd, struct ib_qp *qp)
1062 mutex_lock(&xrcd->tgt_qp_mutex);
1063 list_add(&qp->xrcd_list, &xrcd->tgt_qp_list);
1064 mutex_unlock(&xrcd->tgt_qp_mutex);
1067 static struct ib_qp *__ib_open_qp(struct ib_qp *real_qp,
1068 void (*event_handler)(struct ib_event *, void *),
1072 unsigned long flags;
1075 qp = kzalloc(sizeof *qp, GFP_KERNEL);
1077 return ERR_PTR(-ENOMEM);
1079 qp->real_qp = real_qp;
1080 err = ib_open_shared_qp_security(qp, real_qp->device);
1083 return ERR_PTR(err);
1086 qp->real_qp = real_qp;
1087 atomic_inc(&real_qp->usecnt);
1088 qp->device = real_qp->device;
1089 qp->event_handler = event_handler;
1090 qp->qp_context = qp_context;
1091 qp->qp_num = real_qp->qp_num;
1092 qp->qp_type = real_qp->qp_type;
1094 spin_lock_irqsave(&real_qp->device->event_handler_lock, flags);
1095 list_add(&qp->open_list, &real_qp->open_list);
1096 spin_unlock_irqrestore(&real_qp->device->event_handler_lock, flags);
1101 struct ib_qp *ib_open_qp(struct ib_xrcd *xrcd,
1102 struct ib_qp_open_attr *qp_open_attr)
1104 struct ib_qp *qp, *real_qp;
1106 if (qp_open_attr->qp_type != IB_QPT_XRC_TGT)
1107 return ERR_PTR(-EINVAL);
1109 qp = ERR_PTR(-EINVAL);
1110 mutex_lock(&xrcd->tgt_qp_mutex);
1111 list_for_each_entry(real_qp, &xrcd->tgt_qp_list, xrcd_list) {
1112 if (real_qp->qp_num == qp_open_attr->qp_num) {
1113 qp = __ib_open_qp(real_qp, qp_open_attr->event_handler,
1114 qp_open_attr->qp_context);
1118 mutex_unlock(&xrcd->tgt_qp_mutex);
1121 EXPORT_SYMBOL(ib_open_qp);
1123 static struct ib_qp *create_xrc_qp_user(struct ib_qp *qp,
1124 struct ib_qp_init_attr *qp_init_attr,
1125 struct ib_udata *udata)
1127 struct ib_qp *real_qp = qp;
1129 qp->event_handler = __ib_shared_qp_event_handler;
1130 qp->qp_context = qp;
1132 qp->send_cq = qp->recv_cq = NULL;
1134 qp->xrcd = qp_init_attr->xrcd;
1135 atomic_inc(&qp_init_attr->xrcd->usecnt);
1136 INIT_LIST_HEAD(&qp->open_list);
1138 qp = __ib_open_qp(real_qp, qp_init_attr->event_handler,
1139 qp_init_attr->qp_context);
1143 __ib_insert_xrcd_qp(qp_init_attr->xrcd, real_qp);
1147 struct ib_qp *ib_create_qp_user(struct ib_pd *pd,
1148 struct ib_qp_init_attr *qp_init_attr,
1149 struct ib_udata *udata)
1151 struct ib_device *device = pd ? pd->device : qp_init_attr->xrcd->device;
1155 if (qp_init_attr->rwq_ind_tbl &&
1156 (qp_init_attr->recv_cq ||
1157 qp_init_attr->srq || qp_init_attr->cap.max_recv_wr ||
1158 qp_init_attr->cap.max_recv_sge))
1159 return ERR_PTR(-EINVAL);
1161 if ((qp_init_attr->create_flags & IB_QP_CREATE_INTEGRITY_EN) &&
1162 !(device->attrs.device_cap_flags & IB_DEVICE_INTEGRITY_HANDOVER))
1163 return ERR_PTR(-EINVAL);
1166 * If the callers is using the RDMA API calculate the resources
1167 * needed for the RDMA READ/WRITE operations.
1169 * Note that these callers need to pass in a port number.
1171 if (qp_init_attr->cap.max_rdma_ctxs)
1172 rdma_rw_init_qp(device, qp_init_attr);
1174 qp = _ib_create_qp(device, pd, qp_init_attr, NULL, NULL);
1178 ret = ib_create_qp_security(qp, device);
1182 qp->qp_type = qp_init_attr->qp_type;
1183 qp->rwq_ind_tbl = qp_init_attr->rwq_ind_tbl;
1185 atomic_set(&qp->usecnt, 0);
1187 spin_lock_init(&qp->mr_lock);
1188 INIT_LIST_HEAD(&qp->rdma_mrs);
1189 INIT_LIST_HEAD(&qp->sig_mrs);
1192 if (qp_init_attr->qp_type == IB_QPT_XRC_TGT) {
1193 struct ib_qp *xrc_qp =
1194 create_xrc_qp_user(qp, qp_init_attr, udata);
1196 if (IS_ERR(xrc_qp)) {
1197 ret = PTR_ERR(xrc_qp);
1203 qp->event_handler = qp_init_attr->event_handler;
1204 qp->qp_context = qp_init_attr->qp_context;
1205 if (qp_init_attr->qp_type == IB_QPT_XRC_INI) {
1209 qp->recv_cq = qp_init_attr->recv_cq;
1210 if (qp_init_attr->recv_cq)
1211 atomic_inc(&qp_init_attr->recv_cq->usecnt);
1212 qp->srq = qp_init_attr->srq;
1214 atomic_inc(&qp_init_attr->srq->usecnt);
1217 qp->send_cq = qp_init_attr->send_cq;
1220 atomic_inc(&pd->usecnt);
1221 if (qp_init_attr->send_cq)
1222 atomic_inc(&qp_init_attr->send_cq->usecnt);
1223 if (qp_init_attr->rwq_ind_tbl)
1224 atomic_inc(&qp->rwq_ind_tbl->usecnt);
1226 if (qp_init_attr->cap.max_rdma_ctxs) {
1227 ret = rdma_rw_init_mrs(qp, qp_init_attr);
1233 * Note: all hw drivers guarantee that max_send_sge is lower than
1234 * the device RDMA WRITE SGE limit but not all hw drivers ensure that
1235 * max_send_sge <= max_sge_rd.
1237 qp->max_write_sge = qp_init_attr->cap.max_send_sge;
1238 qp->max_read_sge = min_t(u32, qp_init_attr->cap.max_send_sge,
1239 device->attrs.max_sge_rd);
1240 if (qp_init_attr->create_flags & IB_QP_CREATE_INTEGRITY_EN)
1241 qp->integrity_en = true;
1247 return ERR_PTR(ret);
1250 EXPORT_SYMBOL(ib_create_qp_user);
1252 static const struct {
1254 enum ib_qp_attr_mask req_param[IB_QPT_MAX];
1255 enum ib_qp_attr_mask opt_param[IB_QPT_MAX];
1256 } qp_state_table[IB_QPS_ERR + 1][IB_QPS_ERR + 1] = {
1258 [IB_QPS_RESET] = { .valid = 1 },
1262 [IB_QPT_UD] = (IB_QP_PKEY_INDEX |
1265 [IB_QPT_RAW_PACKET] = IB_QP_PORT,
1266 [IB_QPT_UC] = (IB_QP_PKEY_INDEX |
1268 IB_QP_ACCESS_FLAGS),
1269 [IB_QPT_RC] = (IB_QP_PKEY_INDEX |
1271 IB_QP_ACCESS_FLAGS),
1272 [IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX |
1274 IB_QP_ACCESS_FLAGS),
1275 [IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX |
1277 IB_QP_ACCESS_FLAGS),
1278 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX |
1280 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX |
1286 [IB_QPS_RESET] = { .valid = 1 },
1287 [IB_QPS_ERR] = { .valid = 1 },
1291 [IB_QPT_UD] = (IB_QP_PKEY_INDEX |
1294 [IB_QPT_UC] = (IB_QP_PKEY_INDEX |
1296 IB_QP_ACCESS_FLAGS),
1297 [IB_QPT_RC] = (IB_QP_PKEY_INDEX |
1299 IB_QP_ACCESS_FLAGS),
1300 [IB_QPT_XRC_INI] = (IB_QP_PKEY_INDEX |
1302 IB_QP_ACCESS_FLAGS),
1303 [IB_QPT_XRC_TGT] = (IB_QP_PKEY_INDEX |
1305 IB_QP_ACCESS_FLAGS),
1306 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX |
1308 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX |
1315 [IB_QPT_UC] = (IB_QP_AV |
1319 [IB_QPT_RC] = (IB_QP_AV |
1323 IB_QP_MAX_DEST_RD_ATOMIC |
1324 IB_QP_MIN_RNR_TIMER),
1325 [IB_QPT_XRC_INI] = (IB_QP_AV |
1329 [IB_QPT_XRC_TGT] = (IB_QP_AV |
1333 IB_QP_MAX_DEST_RD_ATOMIC |
1334 IB_QP_MIN_RNR_TIMER),
1337 [IB_QPT_UD] = (IB_QP_PKEY_INDEX |
1339 [IB_QPT_UC] = (IB_QP_ALT_PATH |
1340 IB_QP_ACCESS_FLAGS |
1342 [IB_QPT_RC] = (IB_QP_ALT_PATH |
1343 IB_QP_ACCESS_FLAGS |
1345 [IB_QPT_XRC_INI] = (IB_QP_ALT_PATH |
1346 IB_QP_ACCESS_FLAGS |
1348 [IB_QPT_XRC_TGT] = (IB_QP_ALT_PATH |
1349 IB_QP_ACCESS_FLAGS |
1351 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX |
1353 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX |
1359 [IB_QPS_RESET] = { .valid = 1 },
1360 [IB_QPS_ERR] = { .valid = 1 },
1364 [IB_QPT_UD] = IB_QP_SQ_PSN,
1365 [IB_QPT_UC] = IB_QP_SQ_PSN,
1366 [IB_QPT_RC] = (IB_QP_TIMEOUT |
1370 IB_QP_MAX_QP_RD_ATOMIC),
1371 [IB_QPT_XRC_INI] = (IB_QP_TIMEOUT |
1375 IB_QP_MAX_QP_RD_ATOMIC),
1376 [IB_QPT_XRC_TGT] = (IB_QP_TIMEOUT |
1378 [IB_QPT_SMI] = IB_QP_SQ_PSN,
1379 [IB_QPT_GSI] = IB_QP_SQ_PSN,
1382 [IB_QPT_UD] = (IB_QP_CUR_STATE |
1384 [IB_QPT_UC] = (IB_QP_CUR_STATE |
1386 IB_QP_ACCESS_FLAGS |
1387 IB_QP_PATH_MIG_STATE),
1388 [IB_QPT_RC] = (IB_QP_CUR_STATE |
1390 IB_QP_ACCESS_FLAGS |
1391 IB_QP_MIN_RNR_TIMER |
1392 IB_QP_PATH_MIG_STATE),
1393 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE |
1395 IB_QP_ACCESS_FLAGS |
1396 IB_QP_PATH_MIG_STATE),
1397 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE |
1399 IB_QP_ACCESS_FLAGS |
1400 IB_QP_MIN_RNR_TIMER |
1401 IB_QP_PATH_MIG_STATE),
1402 [IB_QPT_SMI] = (IB_QP_CUR_STATE |
1404 [IB_QPT_GSI] = (IB_QP_CUR_STATE |
1406 [IB_QPT_RAW_PACKET] = IB_QP_RATE_LIMIT,
1411 [IB_QPS_RESET] = { .valid = 1 },
1412 [IB_QPS_ERR] = { .valid = 1 },
1416 [IB_QPT_UD] = (IB_QP_CUR_STATE |
1418 [IB_QPT_UC] = (IB_QP_CUR_STATE |
1419 IB_QP_ACCESS_FLAGS |
1421 IB_QP_PATH_MIG_STATE),
1422 [IB_QPT_RC] = (IB_QP_CUR_STATE |
1423 IB_QP_ACCESS_FLAGS |
1425 IB_QP_PATH_MIG_STATE |
1426 IB_QP_MIN_RNR_TIMER),
1427 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE |
1428 IB_QP_ACCESS_FLAGS |
1430 IB_QP_PATH_MIG_STATE),
1431 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE |
1432 IB_QP_ACCESS_FLAGS |
1434 IB_QP_PATH_MIG_STATE |
1435 IB_QP_MIN_RNR_TIMER),
1436 [IB_QPT_SMI] = (IB_QP_CUR_STATE |
1438 [IB_QPT_GSI] = (IB_QP_CUR_STATE |
1440 [IB_QPT_RAW_PACKET] = IB_QP_RATE_LIMIT,
1446 [IB_QPT_UD] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1447 [IB_QPT_UC] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1448 [IB_QPT_RC] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1449 [IB_QPT_XRC_INI] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1450 [IB_QPT_XRC_TGT] = IB_QP_EN_SQD_ASYNC_NOTIFY, /* ??? */
1451 [IB_QPT_SMI] = IB_QP_EN_SQD_ASYNC_NOTIFY,
1452 [IB_QPT_GSI] = IB_QP_EN_SQD_ASYNC_NOTIFY
1457 [IB_QPS_RESET] = { .valid = 1 },
1458 [IB_QPS_ERR] = { .valid = 1 },
1462 [IB_QPT_UD] = (IB_QP_CUR_STATE |
1464 [IB_QPT_UC] = (IB_QP_CUR_STATE |
1466 IB_QP_ACCESS_FLAGS |
1467 IB_QP_PATH_MIG_STATE),
1468 [IB_QPT_RC] = (IB_QP_CUR_STATE |
1470 IB_QP_ACCESS_FLAGS |
1471 IB_QP_MIN_RNR_TIMER |
1472 IB_QP_PATH_MIG_STATE),
1473 [IB_QPT_XRC_INI] = (IB_QP_CUR_STATE |
1475 IB_QP_ACCESS_FLAGS |
1476 IB_QP_PATH_MIG_STATE),
1477 [IB_QPT_XRC_TGT] = (IB_QP_CUR_STATE |
1479 IB_QP_ACCESS_FLAGS |
1480 IB_QP_MIN_RNR_TIMER |
1481 IB_QP_PATH_MIG_STATE),
1482 [IB_QPT_SMI] = (IB_QP_CUR_STATE |
1484 [IB_QPT_GSI] = (IB_QP_CUR_STATE |
1491 [IB_QPT_UD] = (IB_QP_PKEY_INDEX |
1493 [IB_QPT_UC] = (IB_QP_AV |
1495 IB_QP_ACCESS_FLAGS |
1497 IB_QP_PATH_MIG_STATE),
1498 [IB_QPT_RC] = (IB_QP_PORT |
1503 IB_QP_MAX_QP_RD_ATOMIC |
1504 IB_QP_MAX_DEST_RD_ATOMIC |
1506 IB_QP_ACCESS_FLAGS |
1508 IB_QP_MIN_RNR_TIMER |
1509 IB_QP_PATH_MIG_STATE),
1510 [IB_QPT_XRC_INI] = (IB_QP_PORT |
1515 IB_QP_MAX_QP_RD_ATOMIC |
1517 IB_QP_ACCESS_FLAGS |
1519 IB_QP_PATH_MIG_STATE),
1520 [IB_QPT_XRC_TGT] = (IB_QP_PORT |
1523 IB_QP_MAX_DEST_RD_ATOMIC |
1525 IB_QP_ACCESS_FLAGS |
1527 IB_QP_MIN_RNR_TIMER |
1528 IB_QP_PATH_MIG_STATE),
1529 [IB_QPT_SMI] = (IB_QP_PKEY_INDEX |
1531 [IB_QPT_GSI] = (IB_QP_PKEY_INDEX |
1537 [IB_QPS_RESET] = { .valid = 1 },
1538 [IB_QPS_ERR] = { .valid = 1 },
1542 [IB_QPT_UD] = (IB_QP_CUR_STATE |
1544 [IB_QPT_UC] = (IB_QP_CUR_STATE |
1545 IB_QP_ACCESS_FLAGS),
1546 [IB_QPT_SMI] = (IB_QP_CUR_STATE |
1548 [IB_QPT_GSI] = (IB_QP_CUR_STATE |
1554 [IB_QPS_RESET] = { .valid = 1 },
1555 [IB_QPS_ERR] = { .valid = 1 }
1559 bool ib_modify_qp_is_ok(enum ib_qp_state cur_state, enum ib_qp_state next_state,
1560 enum ib_qp_type type, enum ib_qp_attr_mask mask)
1562 enum ib_qp_attr_mask req_param, opt_param;
1564 if (mask & IB_QP_CUR_STATE &&
1565 cur_state != IB_QPS_RTR && cur_state != IB_QPS_RTS &&
1566 cur_state != IB_QPS_SQD && cur_state != IB_QPS_SQE)
1569 if (!qp_state_table[cur_state][next_state].valid)
1572 req_param = qp_state_table[cur_state][next_state].req_param[type];
1573 opt_param = qp_state_table[cur_state][next_state].opt_param[type];
1575 if ((mask & req_param) != req_param)
1578 if (mask & ~(req_param | opt_param | IB_QP_STATE))
1583 EXPORT_SYMBOL(ib_modify_qp_is_ok);
1586 * ib_resolve_eth_dmac - Resolve destination mac address
1587 * @device: Device to consider
1588 * @ah_attr: address handle attribute which describes the
1589 * source and destination parameters
1590 * ib_resolve_eth_dmac() resolves destination mac address and L3 hop limit It
1591 * returns 0 on success or appropriate error code. It initializes the
1592 * necessary ah_attr fields when call is successful.
1594 static int ib_resolve_eth_dmac(struct ib_device *device,
1595 struct rdma_ah_attr *ah_attr)
1599 if (rdma_is_multicast_addr((struct in6_addr *)ah_attr->grh.dgid.raw)) {
1600 if (ipv6_addr_v4mapped((struct in6_addr *)ah_attr->grh.dgid.raw)) {
1603 memcpy(&addr, ah_attr->grh.dgid.raw + 12, 4);
1604 ip_eth_mc_map(addr, (char *)ah_attr->roce.dmac);
1606 ipv6_eth_mc_map((struct in6_addr *)ah_attr->grh.dgid.raw,
1607 (char *)ah_attr->roce.dmac);
1610 ret = ib_resolve_unicast_gid_dmac(device, ah_attr);
1615 static bool is_qp_type_connected(const struct ib_qp *qp)
1617 return (qp->qp_type == IB_QPT_UC ||
1618 qp->qp_type == IB_QPT_RC ||
1619 qp->qp_type == IB_QPT_XRC_INI ||
1620 qp->qp_type == IB_QPT_XRC_TGT);
1624 * IB core internal function to perform QP attributes modification.
1626 static int _ib_modify_qp(struct ib_qp *qp, struct ib_qp_attr *attr,
1627 int attr_mask, struct ib_udata *udata)
1629 u8 port = attr_mask & IB_QP_PORT ? attr->port_num : qp->port;
1630 const struct ib_gid_attr *old_sgid_attr_av;
1631 const struct ib_gid_attr *old_sgid_attr_alt_av;
1634 if (attr_mask & IB_QP_AV) {
1635 ret = rdma_fill_sgid_attr(qp->device, &attr->ah_attr,
1640 if (attr_mask & IB_QP_ALT_PATH) {
1642 * FIXME: This does not track the migration state, so if the
1643 * user loads a new alternate path after the HW has migrated
1644 * from primary->alternate we will keep the wrong
1645 * references. This is OK for IB because the reference
1646 * counting does not serve any functional purpose.
1648 ret = rdma_fill_sgid_attr(qp->device, &attr->alt_ah_attr,
1649 &old_sgid_attr_alt_av);
1654 * Today the core code can only handle alternate paths and APM
1655 * for IB. Ban them in roce mode.
1657 if (!(rdma_protocol_ib(qp->device,
1658 attr->alt_ah_attr.port_num) &&
1659 rdma_protocol_ib(qp->device, port))) {
1666 * If the user provided the qp_attr then we have to resolve it. Kernel
1667 * users have to provide already resolved rdma_ah_attr's
1669 if (udata && (attr_mask & IB_QP_AV) &&
1670 attr->ah_attr.type == RDMA_AH_ATTR_TYPE_ROCE &&
1671 is_qp_type_connected(qp)) {
1672 ret = ib_resolve_eth_dmac(qp->device, &attr->ah_attr);
1677 if (rdma_ib_or_roce(qp->device, port)) {
1678 if (attr_mask & IB_QP_RQ_PSN && attr->rq_psn & ~0xffffff) {
1679 dev_warn(&qp->device->dev,
1680 "%s rq_psn overflow, masking to 24 bits\n",
1682 attr->rq_psn &= 0xffffff;
1685 if (attr_mask & IB_QP_SQ_PSN && attr->sq_psn & ~0xffffff) {
1686 dev_warn(&qp->device->dev,
1687 " %s sq_psn overflow, masking to 24 bits\n",
1689 attr->sq_psn &= 0xffffff;
1694 * Bind this qp to a counter automatically based on the rdma counter
1695 * rules. This only set in RST2INIT with port specified
1697 if (!qp->counter && (attr_mask & IB_QP_PORT) &&
1698 ((attr_mask & IB_QP_STATE) && attr->qp_state == IB_QPS_INIT))
1699 rdma_counter_bind_qp_auto(qp, attr->port_num);
1701 ret = ib_security_modify_qp(qp, attr, attr_mask, udata);
1705 if (attr_mask & IB_QP_PORT)
1706 qp->port = attr->port_num;
1707 if (attr_mask & IB_QP_AV)
1709 rdma_update_sgid_attr(&attr->ah_attr, qp->av_sgid_attr);
1710 if (attr_mask & IB_QP_ALT_PATH)
1711 qp->alt_path_sgid_attr = rdma_update_sgid_attr(
1712 &attr->alt_ah_attr, qp->alt_path_sgid_attr);
1715 if (attr_mask & IB_QP_ALT_PATH)
1716 rdma_unfill_sgid_attr(&attr->alt_ah_attr, old_sgid_attr_alt_av);
1718 if (attr_mask & IB_QP_AV)
1719 rdma_unfill_sgid_attr(&attr->ah_attr, old_sgid_attr_av);
1724 * ib_modify_qp_with_udata - Modifies the attributes for the specified QP.
1725 * @ib_qp: The QP to modify.
1726 * @attr: On input, specifies the QP attributes to modify. On output,
1727 * the current values of selected QP attributes are returned.
1728 * @attr_mask: A bit-mask used to specify which attributes of the QP
1729 * are being modified.
1730 * @udata: pointer to user's input output buffer information
1731 * are being modified.
1732 * It returns 0 on success and returns appropriate error code on error.
1734 int ib_modify_qp_with_udata(struct ib_qp *ib_qp, struct ib_qp_attr *attr,
1735 int attr_mask, struct ib_udata *udata)
1737 return _ib_modify_qp(ib_qp->real_qp, attr, attr_mask, udata);
1739 EXPORT_SYMBOL(ib_modify_qp_with_udata);
1741 int ib_get_eth_speed(struct ib_device *dev, u8 port_num, u8 *speed, u8 *width)
1745 struct net_device *netdev;
1746 struct ethtool_link_ksettings lksettings;
1748 if (rdma_port_get_link_layer(dev, port_num) != IB_LINK_LAYER_ETHERNET)
1751 netdev = ib_device_get_netdev(dev, port_num);
1756 rc = __ethtool_get_link_ksettings(netdev, &lksettings);
1762 netdev_speed = lksettings.base.speed;
1764 netdev_speed = SPEED_1000;
1765 pr_warn("%s speed is unknown, defaulting to %d\n", netdev->name,
1769 if (netdev_speed <= SPEED_1000) {
1770 *width = IB_WIDTH_1X;
1771 *speed = IB_SPEED_SDR;
1772 } else if (netdev_speed <= SPEED_10000) {
1773 *width = IB_WIDTH_1X;
1774 *speed = IB_SPEED_FDR10;
1775 } else if (netdev_speed <= SPEED_20000) {
1776 *width = IB_WIDTH_4X;
1777 *speed = IB_SPEED_DDR;
1778 } else if (netdev_speed <= SPEED_25000) {
1779 *width = IB_WIDTH_1X;
1780 *speed = IB_SPEED_EDR;
1781 } else if (netdev_speed <= SPEED_40000) {
1782 *width = IB_WIDTH_4X;
1783 *speed = IB_SPEED_FDR10;
1785 *width = IB_WIDTH_4X;
1786 *speed = IB_SPEED_EDR;
1791 EXPORT_SYMBOL(ib_get_eth_speed);
1793 int ib_modify_qp(struct ib_qp *qp,
1794 struct ib_qp_attr *qp_attr,
1797 return _ib_modify_qp(qp->real_qp, qp_attr, qp_attr_mask, NULL);
1799 EXPORT_SYMBOL(ib_modify_qp);
1801 int ib_query_qp(struct ib_qp *qp,
1802 struct ib_qp_attr *qp_attr,
1804 struct ib_qp_init_attr *qp_init_attr)
1806 qp_attr->ah_attr.grh.sgid_attr = NULL;
1807 qp_attr->alt_ah_attr.grh.sgid_attr = NULL;
1809 return qp->device->ops.query_qp ?
1810 qp->device->ops.query_qp(qp->real_qp, qp_attr, qp_attr_mask,
1811 qp_init_attr) : -EOPNOTSUPP;
1813 EXPORT_SYMBOL(ib_query_qp);
1815 int ib_close_qp(struct ib_qp *qp)
1817 struct ib_qp *real_qp;
1818 unsigned long flags;
1820 real_qp = qp->real_qp;
1824 spin_lock_irqsave(&real_qp->device->event_handler_lock, flags);
1825 list_del(&qp->open_list);
1826 spin_unlock_irqrestore(&real_qp->device->event_handler_lock, flags);
1828 atomic_dec(&real_qp->usecnt);
1830 ib_close_shared_qp_security(qp->qp_sec);
1835 EXPORT_SYMBOL(ib_close_qp);
1837 static int __ib_destroy_shared_qp(struct ib_qp *qp)
1839 struct ib_xrcd *xrcd;
1840 struct ib_qp *real_qp;
1843 real_qp = qp->real_qp;
1844 xrcd = real_qp->xrcd;
1846 mutex_lock(&xrcd->tgt_qp_mutex);
1848 if (atomic_read(&real_qp->usecnt) == 0)
1849 list_del(&real_qp->xrcd_list);
1852 mutex_unlock(&xrcd->tgt_qp_mutex);
1855 ret = ib_destroy_qp(real_qp);
1857 atomic_dec(&xrcd->usecnt);
1859 __ib_insert_xrcd_qp(xrcd, real_qp);
1865 int ib_destroy_qp_user(struct ib_qp *qp, struct ib_udata *udata)
1867 const struct ib_gid_attr *alt_path_sgid_attr = qp->alt_path_sgid_attr;
1868 const struct ib_gid_attr *av_sgid_attr = qp->av_sgid_attr;
1870 struct ib_cq *scq, *rcq;
1872 struct ib_rwq_ind_table *ind_tbl;
1873 struct ib_qp_security *sec;
1876 WARN_ON_ONCE(qp->mrs_used > 0);
1878 if (atomic_read(&qp->usecnt))
1881 if (qp->real_qp != qp)
1882 return __ib_destroy_shared_qp(qp);
1888 ind_tbl = qp->rwq_ind_tbl;
1891 ib_destroy_qp_security_begin(sec);
1894 rdma_rw_cleanup_mrs(qp);
1896 rdma_counter_unbind_qp(qp, true);
1897 rdma_restrack_del(&qp->res);
1898 ret = qp->device->ops.destroy_qp(qp, udata);
1900 if (alt_path_sgid_attr)
1901 rdma_put_gid_attr(alt_path_sgid_attr);
1903 rdma_put_gid_attr(av_sgid_attr);
1905 atomic_dec(&pd->usecnt);
1907 atomic_dec(&scq->usecnt);
1909 atomic_dec(&rcq->usecnt);
1911 atomic_dec(&srq->usecnt);
1913 atomic_dec(&ind_tbl->usecnt);
1915 ib_destroy_qp_security_end(sec);
1918 ib_destroy_qp_security_abort(sec);
1923 EXPORT_SYMBOL(ib_destroy_qp_user);
1925 /* Completion queues */
1927 struct ib_cq *__ib_create_cq(struct ib_device *device,
1928 ib_comp_handler comp_handler,
1929 void (*event_handler)(struct ib_event *, void *),
1931 const struct ib_cq_init_attr *cq_attr,
1937 cq = rdma_zalloc_drv_obj(device, ib_cq);
1939 return ERR_PTR(-ENOMEM);
1941 cq->device = device;
1943 cq->comp_handler = comp_handler;
1944 cq->event_handler = event_handler;
1945 cq->cq_context = cq_context;
1946 atomic_set(&cq->usecnt, 0);
1947 cq->res.type = RDMA_RESTRACK_CQ;
1948 rdma_restrack_set_task(&cq->res, caller);
1950 ret = device->ops.create_cq(cq, cq_attr, NULL);
1953 return ERR_PTR(ret);
1956 rdma_restrack_kadd(&cq->res);
1959 EXPORT_SYMBOL(__ib_create_cq);
1961 int rdma_set_cq_moderation(struct ib_cq *cq, u16 cq_count, u16 cq_period)
1963 return cq->device->ops.modify_cq ?
1964 cq->device->ops.modify_cq(cq, cq_count,
1965 cq_period) : -EOPNOTSUPP;
1967 EXPORT_SYMBOL(rdma_set_cq_moderation);
1969 int ib_destroy_cq_user(struct ib_cq *cq, struct ib_udata *udata)
1971 if (atomic_read(&cq->usecnt))
1974 rdma_restrack_del(&cq->res);
1975 cq->device->ops.destroy_cq(cq, udata);
1979 EXPORT_SYMBOL(ib_destroy_cq_user);
1981 int ib_resize_cq(struct ib_cq *cq, int cqe)
1983 return cq->device->ops.resize_cq ?
1984 cq->device->ops.resize_cq(cq, cqe, NULL) : -EOPNOTSUPP;
1986 EXPORT_SYMBOL(ib_resize_cq);
1988 /* Memory regions */
1990 int ib_dereg_mr_user(struct ib_mr *mr, struct ib_udata *udata)
1992 struct ib_pd *pd = mr->pd;
1993 struct ib_dm *dm = mr->dm;
1994 struct ib_sig_attrs *sig_attrs = mr->sig_attrs;
1997 rdma_restrack_del(&mr->res);
1998 ret = mr->device->ops.dereg_mr(mr, udata);
2000 atomic_dec(&pd->usecnt);
2002 atomic_dec(&dm->usecnt);
2008 EXPORT_SYMBOL(ib_dereg_mr_user);
2011 * ib_alloc_mr_user() - Allocates a memory region
2012 * @pd: protection domain associated with the region
2013 * @mr_type: memory region type
2014 * @max_num_sg: maximum sg entries available for registration.
2015 * @udata: user data or null for kernel objects
2018 * Memory registeration page/sg lists must not exceed max_num_sg.
2019 * For mr_type IB_MR_TYPE_MEM_REG, the total length cannot exceed
2020 * max_num_sg * used_page_size.
2023 struct ib_mr *ib_alloc_mr_user(struct ib_pd *pd, enum ib_mr_type mr_type,
2024 u32 max_num_sg, struct ib_udata *udata)
2028 if (!pd->device->ops.alloc_mr)
2029 return ERR_PTR(-EOPNOTSUPP);
2031 if (WARN_ON_ONCE(mr_type == IB_MR_TYPE_INTEGRITY))
2032 return ERR_PTR(-EINVAL);
2034 mr = pd->device->ops.alloc_mr(pd, mr_type, max_num_sg, udata);
2036 mr->device = pd->device;
2040 atomic_inc(&pd->usecnt);
2041 mr->need_inval = false;
2042 mr->res.type = RDMA_RESTRACK_MR;
2043 rdma_restrack_kadd(&mr->res);
2045 mr->sig_attrs = NULL;
2050 EXPORT_SYMBOL(ib_alloc_mr_user);
2053 * ib_alloc_mr_integrity() - Allocates an integrity memory region
2054 * @pd: protection domain associated with the region
2055 * @max_num_data_sg: maximum data sg entries available for registration
2056 * @max_num_meta_sg: maximum metadata sg entries available for
2060 * Memory registration page/sg lists must not exceed max_num_sg,
2061 * also the integrity page/sg lists must not exceed max_num_meta_sg.
2064 struct ib_mr *ib_alloc_mr_integrity(struct ib_pd *pd,
2065 u32 max_num_data_sg,
2066 u32 max_num_meta_sg)
2069 struct ib_sig_attrs *sig_attrs;
2071 if (!pd->device->ops.alloc_mr_integrity ||
2072 !pd->device->ops.map_mr_sg_pi)
2073 return ERR_PTR(-EOPNOTSUPP);
2075 if (!max_num_meta_sg)
2076 return ERR_PTR(-EINVAL);
2078 sig_attrs = kzalloc(sizeof(struct ib_sig_attrs), GFP_KERNEL);
2080 return ERR_PTR(-ENOMEM);
2082 mr = pd->device->ops.alloc_mr_integrity(pd, max_num_data_sg,
2089 mr->device = pd->device;
2093 atomic_inc(&pd->usecnt);
2094 mr->need_inval = false;
2095 mr->res.type = RDMA_RESTRACK_MR;
2096 rdma_restrack_kadd(&mr->res);
2097 mr->type = IB_MR_TYPE_INTEGRITY;
2098 mr->sig_attrs = sig_attrs;
2102 EXPORT_SYMBOL(ib_alloc_mr_integrity);
2104 /* "Fast" memory regions */
2106 struct ib_fmr *ib_alloc_fmr(struct ib_pd *pd,
2107 int mr_access_flags,
2108 struct ib_fmr_attr *fmr_attr)
2112 if (!pd->device->ops.alloc_fmr)
2113 return ERR_PTR(-EOPNOTSUPP);
2115 fmr = pd->device->ops.alloc_fmr(pd, mr_access_flags, fmr_attr);
2117 fmr->device = pd->device;
2119 atomic_inc(&pd->usecnt);
2124 EXPORT_SYMBOL(ib_alloc_fmr);
2126 int ib_unmap_fmr(struct list_head *fmr_list)
2130 if (list_empty(fmr_list))
2133 fmr = list_entry(fmr_list->next, struct ib_fmr, list);
2134 return fmr->device->ops.unmap_fmr(fmr_list);
2136 EXPORT_SYMBOL(ib_unmap_fmr);
2138 int ib_dealloc_fmr(struct ib_fmr *fmr)
2144 ret = fmr->device->ops.dealloc_fmr(fmr);
2146 atomic_dec(&pd->usecnt);
2150 EXPORT_SYMBOL(ib_dealloc_fmr);
2152 /* Multicast groups */
2154 static bool is_valid_mcast_lid(struct ib_qp *qp, u16 lid)
2156 struct ib_qp_init_attr init_attr = {};
2157 struct ib_qp_attr attr = {};
2158 int num_eth_ports = 0;
2161 /* If QP state >= init, it is assigned to a port and we can check this
2164 if (!ib_query_qp(qp, &attr, IB_QP_STATE | IB_QP_PORT, &init_attr)) {
2165 if (attr.qp_state >= IB_QPS_INIT) {
2166 if (rdma_port_get_link_layer(qp->device, attr.port_num) !=
2167 IB_LINK_LAYER_INFINIBAND)
2173 /* Can't get a quick answer, iterate over all ports */
2174 for (port = 0; port < qp->device->phys_port_cnt; port++)
2175 if (rdma_port_get_link_layer(qp->device, port) !=
2176 IB_LINK_LAYER_INFINIBAND)
2179 /* If we have at lease one Ethernet port, RoCE annex declares that
2180 * multicast LID should be ignored. We can't tell at this step if the
2181 * QP belongs to an IB or Ethernet port.
2186 /* If all the ports are IB, we can check according to IB spec. */
2188 return !(lid < be16_to_cpu(IB_MULTICAST_LID_BASE) ||
2189 lid == be16_to_cpu(IB_LID_PERMISSIVE));
2192 int ib_attach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid)
2196 if (!qp->device->ops.attach_mcast)
2199 if (!rdma_is_multicast_addr((struct in6_addr *)gid->raw) ||
2200 qp->qp_type != IB_QPT_UD || !is_valid_mcast_lid(qp, lid))
2203 ret = qp->device->ops.attach_mcast(qp, gid, lid);
2205 atomic_inc(&qp->usecnt);
2208 EXPORT_SYMBOL(ib_attach_mcast);
2210 int ib_detach_mcast(struct ib_qp *qp, union ib_gid *gid, u16 lid)
2214 if (!qp->device->ops.detach_mcast)
2217 if (!rdma_is_multicast_addr((struct in6_addr *)gid->raw) ||
2218 qp->qp_type != IB_QPT_UD || !is_valid_mcast_lid(qp, lid))
2221 ret = qp->device->ops.detach_mcast(qp, gid, lid);
2223 atomic_dec(&qp->usecnt);
2226 EXPORT_SYMBOL(ib_detach_mcast);
2228 struct ib_xrcd *__ib_alloc_xrcd(struct ib_device *device, const char *caller)
2230 struct ib_xrcd *xrcd;
2232 if (!device->ops.alloc_xrcd)
2233 return ERR_PTR(-EOPNOTSUPP);
2235 xrcd = device->ops.alloc_xrcd(device, NULL);
2236 if (!IS_ERR(xrcd)) {
2237 xrcd->device = device;
2239 atomic_set(&xrcd->usecnt, 0);
2240 mutex_init(&xrcd->tgt_qp_mutex);
2241 INIT_LIST_HEAD(&xrcd->tgt_qp_list);
2246 EXPORT_SYMBOL(__ib_alloc_xrcd);
2248 int ib_dealloc_xrcd(struct ib_xrcd *xrcd, struct ib_udata *udata)
2253 if (atomic_read(&xrcd->usecnt))
2256 while (!list_empty(&xrcd->tgt_qp_list)) {
2257 qp = list_entry(xrcd->tgt_qp_list.next, struct ib_qp, xrcd_list);
2258 ret = ib_destroy_qp(qp);
2262 mutex_destroy(&xrcd->tgt_qp_mutex);
2264 return xrcd->device->ops.dealloc_xrcd(xrcd, udata);
2266 EXPORT_SYMBOL(ib_dealloc_xrcd);
2269 * ib_create_wq - Creates a WQ associated with the specified protection
2271 * @pd: The protection domain associated with the WQ.
2272 * @wq_attr: A list of initial attributes required to create the
2273 * WQ. If WQ creation succeeds, then the attributes are updated to
2274 * the actual capabilities of the created WQ.
2276 * wq_attr->max_wr and wq_attr->max_sge determine
2277 * the requested size of the WQ, and set to the actual values allocated
2279 * If ib_create_wq() succeeds, then max_wr and max_sge will always be
2280 * at least as large as the requested values.
2282 struct ib_wq *ib_create_wq(struct ib_pd *pd,
2283 struct ib_wq_init_attr *wq_attr)
2287 if (!pd->device->ops.create_wq)
2288 return ERR_PTR(-EOPNOTSUPP);
2290 wq = pd->device->ops.create_wq(pd, wq_attr, NULL);
2292 wq->event_handler = wq_attr->event_handler;
2293 wq->wq_context = wq_attr->wq_context;
2294 wq->wq_type = wq_attr->wq_type;
2295 wq->cq = wq_attr->cq;
2296 wq->device = pd->device;
2299 atomic_inc(&pd->usecnt);
2300 atomic_inc(&wq_attr->cq->usecnt);
2301 atomic_set(&wq->usecnt, 0);
2305 EXPORT_SYMBOL(ib_create_wq);
2308 * ib_destroy_wq - Destroys the specified user WQ.
2309 * @wq: The WQ to destroy.
2310 * @udata: Valid user data
2312 int ib_destroy_wq(struct ib_wq *wq, struct ib_udata *udata)
2314 struct ib_cq *cq = wq->cq;
2315 struct ib_pd *pd = wq->pd;
2317 if (atomic_read(&wq->usecnt))
2320 wq->device->ops.destroy_wq(wq, udata);
2321 atomic_dec(&pd->usecnt);
2322 atomic_dec(&cq->usecnt);
2326 EXPORT_SYMBOL(ib_destroy_wq);
2329 * ib_modify_wq - Modifies the specified WQ.
2330 * @wq: The WQ to modify.
2331 * @wq_attr: On input, specifies the WQ attributes to modify.
2332 * @wq_attr_mask: A bit-mask used to specify which attributes of the WQ
2333 * are being modified.
2334 * On output, the current values of selected WQ attributes are returned.
2336 int ib_modify_wq(struct ib_wq *wq, struct ib_wq_attr *wq_attr,
2341 if (!wq->device->ops.modify_wq)
2344 err = wq->device->ops.modify_wq(wq, wq_attr, wq_attr_mask, NULL);
2347 EXPORT_SYMBOL(ib_modify_wq);
2350 * ib_create_rwq_ind_table - Creates a RQ Indirection Table.
2351 * @device: The device on which to create the rwq indirection table.
2352 * @ib_rwq_ind_table_init_attr: A list of initial attributes required to
2353 * create the Indirection Table.
2355 * Note: The life time of ib_rwq_ind_table_init_attr->ind_tbl is not less
2356 * than the created ib_rwq_ind_table object and the caller is responsible
2357 * for its memory allocation/free.
2359 struct ib_rwq_ind_table *ib_create_rwq_ind_table(struct ib_device *device,
2360 struct ib_rwq_ind_table_init_attr *init_attr)
2362 struct ib_rwq_ind_table *rwq_ind_table;
2366 if (!device->ops.create_rwq_ind_table)
2367 return ERR_PTR(-EOPNOTSUPP);
2369 table_size = (1 << init_attr->log_ind_tbl_size);
2370 rwq_ind_table = device->ops.create_rwq_ind_table(device,
2372 if (IS_ERR(rwq_ind_table))
2373 return rwq_ind_table;
2375 rwq_ind_table->ind_tbl = init_attr->ind_tbl;
2376 rwq_ind_table->log_ind_tbl_size = init_attr->log_ind_tbl_size;
2377 rwq_ind_table->device = device;
2378 rwq_ind_table->uobject = NULL;
2379 atomic_set(&rwq_ind_table->usecnt, 0);
2381 for (i = 0; i < table_size; i++)
2382 atomic_inc(&rwq_ind_table->ind_tbl[i]->usecnt);
2384 return rwq_ind_table;
2386 EXPORT_SYMBOL(ib_create_rwq_ind_table);
2389 * ib_destroy_rwq_ind_table - Destroys the specified Indirection Table.
2390 * @wq_ind_table: The Indirection Table to destroy.
2392 int ib_destroy_rwq_ind_table(struct ib_rwq_ind_table *rwq_ind_table)
2395 u32 table_size = (1 << rwq_ind_table->log_ind_tbl_size);
2396 struct ib_wq **ind_tbl = rwq_ind_table->ind_tbl;
2398 if (atomic_read(&rwq_ind_table->usecnt))
2401 err = rwq_ind_table->device->ops.destroy_rwq_ind_table(rwq_ind_table);
2403 for (i = 0; i < table_size; i++)
2404 atomic_dec(&ind_tbl[i]->usecnt);
2409 EXPORT_SYMBOL(ib_destroy_rwq_ind_table);
2411 int ib_check_mr_status(struct ib_mr *mr, u32 check_mask,
2412 struct ib_mr_status *mr_status)
2414 if (!mr->device->ops.check_mr_status)
2417 return mr->device->ops.check_mr_status(mr, check_mask, mr_status);
2419 EXPORT_SYMBOL(ib_check_mr_status);
2421 int ib_set_vf_link_state(struct ib_device *device, int vf, u8 port,
2424 if (!device->ops.set_vf_link_state)
2427 return device->ops.set_vf_link_state(device, vf, port, state);
2429 EXPORT_SYMBOL(ib_set_vf_link_state);
2431 int ib_get_vf_config(struct ib_device *device, int vf, u8 port,
2432 struct ifla_vf_info *info)
2434 if (!device->ops.get_vf_config)
2437 return device->ops.get_vf_config(device, vf, port, info);
2439 EXPORT_SYMBOL(ib_get_vf_config);
2441 int ib_get_vf_stats(struct ib_device *device, int vf, u8 port,
2442 struct ifla_vf_stats *stats)
2444 if (!device->ops.get_vf_stats)
2447 return device->ops.get_vf_stats(device, vf, port, stats);
2449 EXPORT_SYMBOL(ib_get_vf_stats);
2451 int ib_set_vf_guid(struct ib_device *device, int vf, u8 port, u64 guid,
2454 if (!device->ops.set_vf_guid)
2457 return device->ops.set_vf_guid(device, vf, port, guid, type);
2459 EXPORT_SYMBOL(ib_set_vf_guid);
2462 * ib_map_mr_sg_pi() - Map the dma mapped SG lists for PI (protection
2463 * information) and set an appropriate memory region for registration.
2464 * @mr: memory region
2465 * @data_sg: dma mapped scatterlist for data
2466 * @data_sg_nents: number of entries in data_sg
2467 * @data_sg_offset: offset in bytes into data_sg
2468 * @meta_sg: dma mapped scatterlist for metadata
2469 * @meta_sg_nents: number of entries in meta_sg
2470 * @meta_sg_offset: offset in bytes into meta_sg
2471 * @page_size: page vector desired page size
2474 * - The MR must be allocated with type IB_MR_TYPE_INTEGRITY.
2476 * Return: 0 on success.
2478 * After this completes successfully, the memory region
2479 * is ready for registration.
2481 int ib_map_mr_sg_pi(struct ib_mr *mr, struct scatterlist *data_sg,
2482 int data_sg_nents, unsigned int *data_sg_offset,
2483 struct scatterlist *meta_sg, int meta_sg_nents,
2484 unsigned int *meta_sg_offset, unsigned int page_size)
2486 if (unlikely(!mr->device->ops.map_mr_sg_pi ||
2487 WARN_ON_ONCE(mr->type != IB_MR_TYPE_INTEGRITY)))
2490 mr->page_size = page_size;
2492 return mr->device->ops.map_mr_sg_pi(mr, data_sg, data_sg_nents,
2493 data_sg_offset, meta_sg,
2494 meta_sg_nents, meta_sg_offset);
2496 EXPORT_SYMBOL(ib_map_mr_sg_pi);
2499 * ib_map_mr_sg() - Map the largest prefix of a dma mapped SG list
2500 * and set it the memory region.
2501 * @mr: memory region
2502 * @sg: dma mapped scatterlist
2503 * @sg_nents: number of entries in sg
2504 * @sg_offset: offset in bytes into sg
2505 * @page_size: page vector desired page size
2508 * - The first sg element is allowed to have an offset.
2509 * - Each sg element must either be aligned to page_size or virtually
2510 * contiguous to the previous element. In case an sg element has a
2511 * non-contiguous offset, the mapping prefix will not include it.
2512 * - The last sg element is allowed to have length less than page_size.
2513 * - If sg_nents total byte length exceeds the mr max_num_sge * page_size
2514 * then only max_num_sg entries will be mapped.
2515 * - If the MR was allocated with type IB_MR_TYPE_SG_GAPS, none of these
2516 * constraints holds and the page_size argument is ignored.
2518 * Returns the number of sg elements that were mapped to the memory region.
2520 * After this completes successfully, the memory region
2521 * is ready for registration.
2523 int ib_map_mr_sg(struct ib_mr *mr, struct scatterlist *sg, int sg_nents,
2524 unsigned int *sg_offset, unsigned int page_size)
2526 if (unlikely(!mr->device->ops.map_mr_sg))
2529 mr->page_size = page_size;
2531 return mr->device->ops.map_mr_sg(mr, sg, sg_nents, sg_offset);
2533 EXPORT_SYMBOL(ib_map_mr_sg);
2536 * ib_sg_to_pages() - Convert the largest prefix of a sg list
2538 * @mr: memory region
2539 * @sgl: dma mapped scatterlist
2540 * @sg_nents: number of entries in sg
2541 * @sg_offset_p: IN: start offset in bytes into sg
2542 * OUT: offset in bytes for element n of the sg of the first
2543 * byte that has not been processed where n is the return
2544 * value of this function.
2545 * @set_page: driver page assignment function pointer
2547 * Core service helper for drivers to convert the largest
2548 * prefix of given sg list to a page vector. The sg list
2549 * prefix converted is the prefix that meet the requirements
2552 * Returns the number of sg elements that were assigned to
2555 int ib_sg_to_pages(struct ib_mr *mr, struct scatterlist *sgl, int sg_nents,
2556 unsigned int *sg_offset_p, int (*set_page)(struct ib_mr *, u64))
2558 struct scatterlist *sg;
2559 u64 last_end_dma_addr = 0;
2560 unsigned int sg_offset = sg_offset_p ? *sg_offset_p : 0;
2561 unsigned int last_page_off = 0;
2562 u64 page_mask = ~((u64)mr->page_size - 1);
2565 if (unlikely(sg_nents <= 0 || sg_offset > sg_dma_len(&sgl[0])))
2568 mr->iova = sg_dma_address(&sgl[0]) + sg_offset;
2571 for_each_sg(sgl, sg, sg_nents, i) {
2572 u64 dma_addr = sg_dma_address(sg) + sg_offset;
2573 u64 prev_addr = dma_addr;
2574 unsigned int dma_len = sg_dma_len(sg) - sg_offset;
2575 u64 end_dma_addr = dma_addr + dma_len;
2576 u64 page_addr = dma_addr & page_mask;
2579 * For the second and later elements, check whether either the
2580 * end of element i-1 or the start of element i is not aligned
2581 * on a page boundary.
2583 if (i && (last_page_off != 0 || page_addr != dma_addr)) {
2584 /* Stop mapping if there is a gap. */
2585 if (last_end_dma_addr != dma_addr)
2589 * Coalesce this element with the last. If it is small
2590 * enough just update mr->length. Otherwise start
2591 * mapping from the next page.
2597 ret = set_page(mr, page_addr);
2598 if (unlikely(ret < 0)) {
2599 sg_offset = prev_addr - sg_dma_address(sg);
2600 mr->length += prev_addr - dma_addr;
2602 *sg_offset_p = sg_offset;
2603 return i || sg_offset ? i : ret;
2605 prev_addr = page_addr;
2607 page_addr += mr->page_size;
2608 } while (page_addr < end_dma_addr);
2610 mr->length += dma_len;
2611 last_end_dma_addr = end_dma_addr;
2612 last_page_off = end_dma_addr & ~page_mask;
2621 EXPORT_SYMBOL(ib_sg_to_pages);
2623 struct ib_drain_cqe {
2625 struct completion done;
2628 static void ib_drain_qp_done(struct ib_cq *cq, struct ib_wc *wc)
2630 struct ib_drain_cqe *cqe = container_of(wc->wr_cqe, struct ib_drain_cqe,
2633 complete(&cqe->done);
2637 * Post a WR and block until its completion is reaped for the SQ.
2639 static void __ib_drain_sq(struct ib_qp *qp)
2641 struct ib_cq *cq = qp->send_cq;
2642 struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
2643 struct ib_drain_cqe sdrain;
2644 struct ib_rdma_wr swr = {
2647 { .wr_cqe = &sdrain.cqe, },
2648 .opcode = IB_WR_RDMA_WRITE,
2653 ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
2655 WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
2659 sdrain.cqe.done = ib_drain_qp_done;
2660 init_completion(&sdrain.done);
2662 ret = ib_post_send(qp, &swr.wr, NULL);
2664 WARN_ONCE(ret, "failed to drain send queue: %d\n", ret);
2668 if (cq->poll_ctx == IB_POLL_DIRECT)
2669 while (wait_for_completion_timeout(&sdrain.done, HZ / 10) <= 0)
2670 ib_process_cq_direct(cq, -1);
2672 wait_for_completion(&sdrain.done);
2676 * Post a WR and block until its completion is reaped for the RQ.
2678 static void __ib_drain_rq(struct ib_qp *qp)
2680 struct ib_cq *cq = qp->recv_cq;
2681 struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
2682 struct ib_drain_cqe rdrain;
2683 struct ib_recv_wr rwr = {};
2686 ret = ib_modify_qp(qp, &attr, IB_QP_STATE);
2688 WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
2692 rwr.wr_cqe = &rdrain.cqe;
2693 rdrain.cqe.done = ib_drain_qp_done;
2694 init_completion(&rdrain.done);
2696 ret = ib_post_recv(qp, &rwr, NULL);
2698 WARN_ONCE(ret, "failed to drain recv queue: %d\n", ret);
2702 if (cq->poll_ctx == IB_POLL_DIRECT)
2703 while (wait_for_completion_timeout(&rdrain.done, HZ / 10) <= 0)
2704 ib_process_cq_direct(cq, -1);
2706 wait_for_completion(&rdrain.done);
2710 * ib_drain_sq() - Block until all SQ CQEs have been consumed by the
2712 * @qp: queue pair to drain
2714 * If the device has a provider-specific drain function, then
2715 * call that. Otherwise call the generic drain function
2720 * ensure there is room in the CQ and SQ for the drain work request and
2723 * allocate the CQ using ib_alloc_cq().
2725 * ensure that there are no other contexts that are posting WRs concurrently.
2726 * Otherwise the drain is not guaranteed.
2728 void ib_drain_sq(struct ib_qp *qp)
2730 if (qp->device->ops.drain_sq)
2731 qp->device->ops.drain_sq(qp);
2735 EXPORT_SYMBOL(ib_drain_sq);
2738 * ib_drain_rq() - Block until all RQ CQEs have been consumed by the
2740 * @qp: queue pair to drain
2742 * If the device has a provider-specific drain function, then
2743 * call that. Otherwise call the generic drain function
2748 * ensure there is room in the CQ and RQ for the drain work request and
2751 * allocate the CQ using ib_alloc_cq().
2753 * ensure that there are no other contexts that are posting WRs concurrently.
2754 * Otherwise the drain is not guaranteed.
2756 void ib_drain_rq(struct ib_qp *qp)
2758 if (qp->device->ops.drain_rq)
2759 qp->device->ops.drain_rq(qp);
2763 EXPORT_SYMBOL(ib_drain_rq);
2766 * ib_drain_qp() - Block until all CQEs have been consumed by the
2767 * application on both the RQ and SQ.
2768 * @qp: queue pair to drain
2772 * ensure there is room in the CQ(s), SQ, and RQ for drain work requests
2775 * allocate the CQs using ib_alloc_cq().
2777 * ensure that there are no other contexts that are posting WRs concurrently.
2778 * Otherwise the drain is not guaranteed.
2780 void ib_drain_qp(struct ib_qp *qp)
2786 EXPORT_SYMBOL(ib_drain_qp);
2788 struct net_device *rdma_alloc_netdev(struct ib_device *device, u8 port_num,
2789 enum rdma_netdev_t type, const char *name,
2790 unsigned char name_assign_type,
2791 void (*setup)(struct net_device *))
2793 struct rdma_netdev_alloc_params params;
2794 struct net_device *netdev;
2797 if (!device->ops.rdma_netdev_get_params)
2798 return ERR_PTR(-EOPNOTSUPP);
2800 rc = device->ops.rdma_netdev_get_params(device, port_num, type,
2805 netdev = alloc_netdev_mqs(params.sizeof_priv, name, name_assign_type,
2806 setup, params.txqs, params.rxqs);
2808 return ERR_PTR(-ENOMEM);
2812 EXPORT_SYMBOL(rdma_alloc_netdev);
2814 int rdma_init_netdev(struct ib_device *device, u8 port_num,
2815 enum rdma_netdev_t type, const char *name,
2816 unsigned char name_assign_type,
2817 void (*setup)(struct net_device *),
2818 struct net_device *netdev)
2820 struct rdma_netdev_alloc_params params;
2823 if (!device->ops.rdma_netdev_get_params)
2826 rc = device->ops.rdma_netdev_get_params(device, port_num, type,
2831 return params.initialize_rdma_netdev(device, port_num,
2832 netdev, params.param);
2834 EXPORT_SYMBOL(rdma_init_netdev);
2836 void __rdma_block_iter_start(struct ib_block_iter *biter,
2837 struct scatterlist *sglist, unsigned int nents,
2840 memset(biter, 0, sizeof(struct ib_block_iter));
2841 biter->__sg = sglist;
2842 biter->__sg_nents = nents;
2844 /* Driver provides best block size to use */
2845 biter->__pg_bit = __fls(pgsz);
2847 EXPORT_SYMBOL(__rdma_block_iter_start);
2849 bool __rdma_block_iter_next(struct ib_block_iter *biter)
2851 unsigned int block_offset;
2853 if (!biter->__sg_nents || !biter->__sg)
2856 biter->__dma_addr = sg_dma_address(biter->__sg) + biter->__sg_advance;
2857 block_offset = biter->__dma_addr & (BIT_ULL(biter->__pg_bit) - 1);
2858 biter->__sg_advance += BIT_ULL(biter->__pg_bit) - block_offset;
2860 if (biter->__sg_advance >= sg_dma_len(biter->__sg)) {
2861 biter->__sg_advance = 0;
2862 biter->__sg = sg_next(biter->__sg);
2863 biter->__sg_nents--;
2868 EXPORT_SYMBOL(__rdma_block_iter_next);