2 * Copyright (c) 2004 Topspin Communications. All rights reserved.
3 * Copyright (c) 2005 Voltaire, Inc. All rights reserved.
4 * Copyright (c) 2006 Intel Corporation. All rights reserved.
6 * This software is available to you under a choice of one of two
7 * licenses. You may choose to be licensed under the terms of the GNU
8 * General Public License (GPL) Version 2, available from the file
9 * COPYING in the main directory of this source tree, or the
10 * OpenIB.org BSD license below:
12 * Redistribution and use in source and binary forms, with or
13 * without modification, are permitted provided that the following
16 * - Redistributions of source code must retain the above
17 * copyright notice, this list of conditions and the following
20 * - Redistributions in binary form must reproduce the above
21 * copyright notice, this list of conditions and the following
22 * disclaimer in the documentation and/or other materials
23 * provided with the distribution.
25 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
26 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
27 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
28 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
29 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
30 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
31 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
35 #include <linux/module.h>
36 #include <linux/init.h>
37 #include <linux/err.h>
38 #include <linux/random.h>
39 #include <linux/spinlock.h>
40 #include <linux/slab.h>
41 #include <linux/dma-mapping.h>
42 #include <linux/kref.h>
43 #include <linux/xarray.h>
44 #include <linux/workqueue.h>
45 #include <uapi/linux/if_ether.h>
46 #include <rdma/ib_pack.h>
47 #include <rdma/ib_cache.h>
48 #include <rdma/rdma_netlink.h>
49 #include <net/netlink.h>
50 #include <uapi/rdma/ib_user_sa.h>
51 #include <rdma/ib_marshall.h>
52 #include <rdma/ib_addr.h>
53 #include <rdma/opa_addr.h>
55 #include "core_priv.h"
57 #define IB_SA_LOCAL_SVC_TIMEOUT_MIN 100
58 #define IB_SA_LOCAL_SVC_TIMEOUT_DEFAULT 2000
59 #define IB_SA_LOCAL_SVC_TIMEOUT_MAX 200000
60 #define IB_SA_CPI_MAX_RETRY_CNT 3
61 #define IB_SA_CPI_RETRY_WAIT 1000 /*msecs */
62 static int sa_local_svc_timeout_ms = IB_SA_LOCAL_SVC_TIMEOUT_DEFAULT;
71 enum rdma_class_port_info_type {
72 RDMA_CLASS_PORT_INFO_IB,
73 RDMA_CLASS_PORT_INFO_OPA
76 struct rdma_class_port_info {
77 enum rdma_class_port_info_type type;
79 struct ib_class_port_info ib;
80 struct opa_class_port_info opa;
84 struct ib_sa_classport_cache {
87 struct rdma_class_port_info data;
91 struct ib_mad_agent *agent;
92 struct ib_sa_sm_ah *sm_ah;
93 struct work_struct update_task;
94 struct ib_sa_classport_cache classport_info;
95 struct delayed_work ib_cpi_work;
96 spinlock_t classport_lock; /* protects class port info set */
101 struct ib_sa_device {
102 int start_port, end_port;
103 struct ib_event_handler event_handler;
104 struct ib_sa_port port[];
108 void (*callback)(struct ib_sa_query *, int, struct ib_sa_mad *);
109 void (*release)(struct ib_sa_query *);
110 struct ib_sa_client *client;
111 struct ib_sa_port *port;
112 struct ib_mad_send_buf *mad_buf;
113 struct ib_sa_sm_ah *sm_ah;
116 struct list_head list; /* Local svc request list */
117 u32 seq; /* Local svc request sequence number */
118 unsigned long timeout; /* Local svc timeout */
119 u8 path_use; /* How will the pathrecord be used */
122 #define IB_SA_ENABLE_LOCAL_SERVICE 0x00000001
123 #define IB_SA_CANCEL 0x00000002
124 #define IB_SA_QUERY_OPA 0x00000004
126 struct ib_sa_path_query {
127 void (*callback)(int, struct sa_path_rec *, void *);
129 struct ib_sa_query sa_query;
130 struct sa_path_rec *conv_pr;
133 struct ib_sa_guidinfo_query {
134 void (*callback)(int, struct ib_sa_guidinfo_rec *, void *);
136 struct ib_sa_query sa_query;
139 struct ib_sa_classport_info_query {
140 void (*callback)(void *);
142 struct ib_sa_query sa_query;
145 struct ib_sa_mcmember_query {
146 void (*callback)(int, struct ib_sa_mcmember_rec *, void *);
148 struct ib_sa_query sa_query;
151 static LIST_HEAD(ib_nl_request_list);
152 static DEFINE_SPINLOCK(ib_nl_request_lock);
153 static atomic_t ib_nl_sa_request_seq;
154 static struct workqueue_struct *ib_nl_wq;
155 static struct delayed_work ib_nl_timed_work;
156 static const struct nla_policy ib_nl_policy[LS_NLA_TYPE_MAX] = {
157 [LS_NLA_TYPE_PATH_RECORD] = {.type = NLA_BINARY,
158 .len = sizeof(struct ib_path_rec_data)},
159 [LS_NLA_TYPE_TIMEOUT] = {.type = NLA_U32},
160 [LS_NLA_TYPE_SERVICE_ID] = {.type = NLA_U64},
161 [LS_NLA_TYPE_DGID] = {.type = NLA_BINARY,
162 .len = sizeof(struct rdma_nla_ls_gid)},
163 [LS_NLA_TYPE_SGID] = {.type = NLA_BINARY,
164 .len = sizeof(struct rdma_nla_ls_gid)},
165 [LS_NLA_TYPE_TCLASS] = {.type = NLA_U8},
166 [LS_NLA_TYPE_PKEY] = {.type = NLA_U16},
167 [LS_NLA_TYPE_QOS_CLASS] = {.type = NLA_U16},
171 static int ib_sa_add_one(struct ib_device *device);
172 static void ib_sa_remove_one(struct ib_device *device, void *client_data);
174 static struct ib_client sa_client = {
176 .add = ib_sa_add_one,
177 .remove = ib_sa_remove_one
180 static DEFINE_XARRAY_FLAGS(queries, XA_FLAGS_ALLOC | XA_FLAGS_LOCK_IRQ);
182 static DEFINE_SPINLOCK(tid_lock);
185 #define PATH_REC_FIELD(field) \
186 .struct_offset_bytes = offsetof(struct sa_path_rec, field), \
187 .struct_size_bytes = sizeof_field(struct sa_path_rec, field), \
188 .field_name = "sa_path_rec:" #field
190 static const struct ib_field path_rec_table[] = {
191 { PATH_REC_FIELD(service_id),
195 { PATH_REC_FIELD(dgid),
199 { PATH_REC_FIELD(sgid),
203 { PATH_REC_FIELD(ib.dlid),
207 { PATH_REC_FIELD(ib.slid),
211 { PATH_REC_FIELD(ib.raw_traffic),
219 { PATH_REC_FIELD(flow_label),
223 { PATH_REC_FIELD(hop_limit),
227 { PATH_REC_FIELD(traffic_class),
231 { PATH_REC_FIELD(reversible),
235 { PATH_REC_FIELD(numb_path),
239 { PATH_REC_FIELD(pkey),
243 { PATH_REC_FIELD(qos_class),
247 { PATH_REC_FIELD(sl),
251 { PATH_REC_FIELD(mtu_selector),
255 { PATH_REC_FIELD(mtu),
259 { PATH_REC_FIELD(rate_selector),
263 { PATH_REC_FIELD(rate),
267 { PATH_REC_FIELD(packet_life_time_selector),
271 { PATH_REC_FIELD(packet_life_time),
275 { PATH_REC_FIELD(preference),
285 #define OPA_PATH_REC_FIELD(field) \
286 .struct_offset_bytes = \
287 offsetof(struct sa_path_rec, field), \
288 .struct_size_bytes = \
289 sizeof_field(struct sa_path_rec, field), \
290 .field_name = "sa_path_rec:" #field
292 static const struct ib_field opa_path_rec_table[] = {
293 { OPA_PATH_REC_FIELD(service_id),
297 { OPA_PATH_REC_FIELD(dgid),
301 { OPA_PATH_REC_FIELD(sgid),
305 { OPA_PATH_REC_FIELD(opa.dlid),
309 { OPA_PATH_REC_FIELD(opa.slid),
313 { OPA_PATH_REC_FIELD(opa.raw_traffic),
321 { OPA_PATH_REC_FIELD(flow_label),
325 { OPA_PATH_REC_FIELD(hop_limit),
329 { OPA_PATH_REC_FIELD(traffic_class),
333 { OPA_PATH_REC_FIELD(reversible),
337 { OPA_PATH_REC_FIELD(numb_path),
341 { OPA_PATH_REC_FIELD(pkey),
345 { OPA_PATH_REC_FIELD(opa.l2_8B),
349 { OPA_PATH_REC_FIELD(opa.l2_10B),
353 { OPA_PATH_REC_FIELD(opa.l2_9B),
357 { OPA_PATH_REC_FIELD(opa.l2_16B),
365 { OPA_PATH_REC_FIELD(opa.qos_type),
369 { OPA_PATH_REC_FIELD(opa.qos_priority),
377 { OPA_PATH_REC_FIELD(sl),
385 { OPA_PATH_REC_FIELD(mtu_selector),
389 { OPA_PATH_REC_FIELD(mtu),
393 { OPA_PATH_REC_FIELD(rate_selector),
397 { OPA_PATH_REC_FIELD(rate),
401 { OPA_PATH_REC_FIELD(packet_life_time_selector),
405 { OPA_PATH_REC_FIELD(packet_life_time),
409 { OPA_PATH_REC_FIELD(preference),
415 #define MCMEMBER_REC_FIELD(field) \
416 .struct_offset_bytes = offsetof(struct ib_sa_mcmember_rec, field), \
417 .struct_size_bytes = sizeof_field(struct ib_sa_mcmember_rec, field), \
418 .field_name = "sa_mcmember_rec:" #field
420 static const struct ib_field mcmember_rec_table[] = {
421 { MCMEMBER_REC_FIELD(mgid),
425 { MCMEMBER_REC_FIELD(port_gid),
429 { MCMEMBER_REC_FIELD(qkey),
433 { MCMEMBER_REC_FIELD(mlid),
437 { MCMEMBER_REC_FIELD(mtu_selector),
441 { MCMEMBER_REC_FIELD(mtu),
445 { MCMEMBER_REC_FIELD(traffic_class),
449 { MCMEMBER_REC_FIELD(pkey),
453 { MCMEMBER_REC_FIELD(rate_selector),
457 { MCMEMBER_REC_FIELD(rate),
461 { MCMEMBER_REC_FIELD(packet_life_time_selector),
465 { MCMEMBER_REC_FIELD(packet_life_time),
469 { MCMEMBER_REC_FIELD(sl),
473 { MCMEMBER_REC_FIELD(flow_label),
477 { MCMEMBER_REC_FIELD(hop_limit),
481 { MCMEMBER_REC_FIELD(scope),
485 { MCMEMBER_REC_FIELD(join_state),
489 { MCMEMBER_REC_FIELD(proxy_join),
499 #define CLASSPORTINFO_REC_FIELD(field) \
500 .struct_offset_bytes = offsetof(struct ib_class_port_info, field), \
501 .struct_size_bytes = sizeof_field(struct ib_class_port_info, field), \
502 .field_name = "ib_class_port_info:" #field
504 static const struct ib_field ib_classport_info_rec_table[] = {
505 { CLASSPORTINFO_REC_FIELD(base_version),
509 { CLASSPORTINFO_REC_FIELD(class_version),
513 { CLASSPORTINFO_REC_FIELD(capability_mask),
517 { CLASSPORTINFO_REC_FIELD(cap_mask2_resp_time),
521 { CLASSPORTINFO_REC_FIELD(redirect_gid),
525 { CLASSPORTINFO_REC_FIELD(redirect_tcslfl),
529 { CLASSPORTINFO_REC_FIELD(redirect_lid),
533 { CLASSPORTINFO_REC_FIELD(redirect_pkey),
538 { CLASSPORTINFO_REC_FIELD(redirect_qp),
542 { CLASSPORTINFO_REC_FIELD(redirect_qkey),
547 { CLASSPORTINFO_REC_FIELD(trap_gid),
551 { CLASSPORTINFO_REC_FIELD(trap_tcslfl),
556 { CLASSPORTINFO_REC_FIELD(trap_lid),
560 { CLASSPORTINFO_REC_FIELD(trap_pkey),
565 { CLASSPORTINFO_REC_FIELD(trap_hlqp),
569 { CLASSPORTINFO_REC_FIELD(trap_qkey),
575 #define OPA_CLASSPORTINFO_REC_FIELD(field) \
576 .struct_offset_bytes =\
577 offsetof(struct opa_class_port_info, field), \
578 .struct_size_bytes = \
579 sizeof_field(struct opa_class_port_info, field), \
580 .field_name = "opa_class_port_info:" #field
582 static const struct ib_field opa_classport_info_rec_table[] = {
583 { OPA_CLASSPORTINFO_REC_FIELD(base_version),
587 { OPA_CLASSPORTINFO_REC_FIELD(class_version),
591 { OPA_CLASSPORTINFO_REC_FIELD(cap_mask),
595 { OPA_CLASSPORTINFO_REC_FIELD(cap_mask2_resp_time),
599 { OPA_CLASSPORTINFO_REC_FIELD(redirect_gid),
603 { OPA_CLASSPORTINFO_REC_FIELD(redirect_tc_fl),
607 { OPA_CLASSPORTINFO_REC_FIELD(redirect_lid),
611 { OPA_CLASSPORTINFO_REC_FIELD(redirect_sl_qp),
615 { OPA_CLASSPORTINFO_REC_FIELD(redirect_qkey),
619 { OPA_CLASSPORTINFO_REC_FIELD(trap_gid),
623 { OPA_CLASSPORTINFO_REC_FIELD(trap_tc_fl),
627 { OPA_CLASSPORTINFO_REC_FIELD(trap_lid),
631 { OPA_CLASSPORTINFO_REC_FIELD(trap_hl_qp),
635 { OPA_CLASSPORTINFO_REC_FIELD(trap_qkey),
639 { OPA_CLASSPORTINFO_REC_FIELD(trap_pkey),
643 { OPA_CLASSPORTINFO_REC_FIELD(redirect_pkey),
647 { OPA_CLASSPORTINFO_REC_FIELD(trap_sl_rsvd),
657 #define GUIDINFO_REC_FIELD(field) \
658 .struct_offset_bytes = offsetof(struct ib_sa_guidinfo_rec, field), \
659 .struct_size_bytes = sizeof_field(struct ib_sa_guidinfo_rec, field), \
660 .field_name = "sa_guidinfo_rec:" #field
662 static const struct ib_field guidinfo_rec_table[] = {
663 { GUIDINFO_REC_FIELD(lid),
667 { GUIDINFO_REC_FIELD(block_num),
671 { GUIDINFO_REC_FIELD(res1),
675 { GUIDINFO_REC_FIELD(res2),
679 { GUIDINFO_REC_FIELD(guid_info_list),
685 static inline void ib_sa_disable_local_svc(struct ib_sa_query *query)
687 query->flags &= ~IB_SA_ENABLE_LOCAL_SERVICE;
690 static inline int ib_sa_query_cancelled(struct ib_sa_query *query)
692 return (query->flags & IB_SA_CANCEL);
695 static void ib_nl_set_path_rec_attrs(struct sk_buff *skb,
696 struct ib_sa_query *query)
698 struct sa_path_rec *sa_rec = query->mad_buf->context[1];
699 struct ib_sa_mad *mad = query->mad_buf->mad;
700 ib_sa_comp_mask comp_mask = mad->sa_hdr.comp_mask;
703 struct rdma_ls_resolve_header *header;
705 query->mad_buf->context[1] = NULL;
707 /* Construct the family header first */
708 header = skb_put(skb, NLMSG_ALIGN(sizeof(*header)));
709 strscpy_pad(header->device_name,
710 dev_name(&query->port->agent->device->dev),
712 header->port_num = query->port->port_num;
714 if ((comp_mask & IB_SA_PATH_REC_REVERSIBLE) &&
715 sa_rec->reversible != 0)
716 query->path_use = LS_RESOLVE_PATH_USE_GMP;
718 query->path_use = LS_RESOLVE_PATH_USE_UNIDIRECTIONAL;
719 header->path_use = query->path_use;
721 /* Now build the attributes */
722 if (comp_mask & IB_SA_PATH_REC_SERVICE_ID) {
723 val64 = be64_to_cpu(sa_rec->service_id);
724 nla_put(skb, RDMA_NLA_F_MANDATORY | LS_NLA_TYPE_SERVICE_ID,
725 sizeof(val64), &val64);
727 if (comp_mask & IB_SA_PATH_REC_DGID)
728 nla_put(skb, RDMA_NLA_F_MANDATORY | LS_NLA_TYPE_DGID,
729 sizeof(sa_rec->dgid), &sa_rec->dgid);
730 if (comp_mask & IB_SA_PATH_REC_SGID)
731 nla_put(skb, RDMA_NLA_F_MANDATORY | LS_NLA_TYPE_SGID,
732 sizeof(sa_rec->sgid), &sa_rec->sgid);
733 if (comp_mask & IB_SA_PATH_REC_TRAFFIC_CLASS)
734 nla_put(skb, RDMA_NLA_F_MANDATORY | LS_NLA_TYPE_TCLASS,
735 sizeof(sa_rec->traffic_class), &sa_rec->traffic_class);
737 if (comp_mask & IB_SA_PATH_REC_PKEY) {
738 val16 = be16_to_cpu(sa_rec->pkey);
739 nla_put(skb, RDMA_NLA_F_MANDATORY | LS_NLA_TYPE_PKEY,
740 sizeof(val16), &val16);
742 if (comp_mask & IB_SA_PATH_REC_QOS_CLASS) {
743 val16 = be16_to_cpu(sa_rec->qos_class);
744 nla_put(skb, RDMA_NLA_F_MANDATORY | LS_NLA_TYPE_QOS_CLASS,
745 sizeof(val16), &val16);
749 static int ib_nl_get_path_rec_attrs_len(ib_sa_comp_mask comp_mask)
753 if (comp_mask & IB_SA_PATH_REC_SERVICE_ID)
754 len += nla_total_size(sizeof(u64));
755 if (comp_mask & IB_SA_PATH_REC_DGID)
756 len += nla_total_size(sizeof(struct rdma_nla_ls_gid));
757 if (comp_mask & IB_SA_PATH_REC_SGID)
758 len += nla_total_size(sizeof(struct rdma_nla_ls_gid));
759 if (comp_mask & IB_SA_PATH_REC_TRAFFIC_CLASS)
760 len += nla_total_size(sizeof(u8));
761 if (comp_mask & IB_SA_PATH_REC_PKEY)
762 len += nla_total_size(sizeof(u16));
763 if (comp_mask & IB_SA_PATH_REC_QOS_CLASS)
764 len += nla_total_size(sizeof(u16));
767 * Make sure that at least some of the required comp_mask bits are
770 if (WARN_ON(len == 0))
773 /* Add the family header */
774 len += NLMSG_ALIGN(sizeof(struct rdma_ls_resolve_header));
779 static int ib_nl_make_request(struct ib_sa_query *query, gfp_t gfp_mask)
781 struct sk_buff *skb = NULL;
782 struct nlmsghdr *nlh;
784 struct ib_sa_mad *mad;
791 INIT_LIST_HEAD(&query->list);
792 query->seq = (u32)atomic_inc_return(&ib_nl_sa_request_seq);
794 mad = query->mad_buf->mad;
795 len = ib_nl_get_path_rec_attrs_len(mad->sa_hdr.comp_mask);
799 skb = nlmsg_new(len, gfp_mask);
803 /* Put nlmsg header only for now */
804 data = ibnl_put_msg(skb, &nlh, query->seq, 0, RDMA_NL_LS,
805 RDMA_NL_LS_OP_RESOLVE, NLM_F_REQUEST);
812 ib_nl_set_path_rec_attrs(skb, query);
814 /* Repair the nlmsg header length */
817 gfp_flag = ((gfp_mask & GFP_ATOMIC) == GFP_ATOMIC) ? GFP_ATOMIC :
820 spin_lock_irqsave(&ib_nl_request_lock, flags);
821 ret = rdma_nl_multicast(&init_net, skb, RDMA_NL_GROUP_LS, gfp_flag);
826 /* Put the request on the list.*/
827 delay = msecs_to_jiffies(sa_local_svc_timeout_ms);
828 query->timeout = delay + jiffies;
829 list_add_tail(&query->list, &ib_nl_request_list);
830 /* Start the timeout if this is the only request */
831 if (ib_nl_request_list.next == &query->list)
832 queue_delayed_work(ib_nl_wq, &ib_nl_timed_work, delay);
835 spin_unlock_irqrestore(&ib_nl_request_lock, flags);
840 static int ib_nl_cancel_request(struct ib_sa_query *query)
843 struct ib_sa_query *wait_query;
846 spin_lock_irqsave(&ib_nl_request_lock, flags);
847 list_for_each_entry(wait_query, &ib_nl_request_list, list) {
848 /* Let the timeout to take care of the callback */
849 if (query == wait_query) {
850 query->flags |= IB_SA_CANCEL;
851 query->timeout = jiffies;
852 list_move(&query->list, &ib_nl_request_list);
854 mod_delayed_work(ib_nl_wq, &ib_nl_timed_work, 1);
858 spin_unlock_irqrestore(&ib_nl_request_lock, flags);
863 static void send_handler(struct ib_mad_agent *agent,
864 struct ib_mad_send_wc *mad_send_wc);
866 static void ib_nl_process_good_resolve_rsp(struct ib_sa_query *query,
867 const struct nlmsghdr *nlh)
869 struct ib_mad_send_wc mad_send_wc;
870 struct ib_sa_mad *mad = NULL;
871 const struct nlattr *head, *curr;
872 struct ib_path_rec_data *rec;
877 if (query->callback) {
878 head = (const struct nlattr *) nlmsg_data(nlh);
879 len = nlmsg_len(nlh);
880 switch (query->path_use) {
881 case LS_RESOLVE_PATH_USE_UNIDIRECTIONAL:
882 mask = IB_PATH_PRIMARY | IB_PATH_OUTBOUND;
885 case LS_RESOLVE_PATH_USE_ALL:
886 case LS_RESOLVE_PATH_USE_GMP:
888 mask = IB_PATH_PRIMARY | IB_PATH_GMP |
889 IB_PATH_BIDIRECTIONAL;
892 nla_for_each_attr(curr, head, len, rem) {
893 if (curr->nla_type == LS_NLA_TYPE_PATH_RECORD) {
894 rec = nla_data(curr);
896 * Get the first one. In the future, we may
897 * need to get up to 6 pathrecords.
899 if ((rec->flags & mask) == mask) {
900 mad = query->mad_buf->mad;
901 mad->mad_hdr.method |=
903 memcpy(mad->data, rec->path_rec,
904 sizeof(rec->path_rec));
910 query->callback(query, status, mad);
913 mad_send_wc.send_buf = query->mad_buf;
914 mad_send_wc.status = IB_WC_SUCCESS;
915 send_handler(query->mad_buf->mad_agent, &mad_send_wc);
918 static void ib_nl_request_timeout(struct work_struct *work)
921 struct ib_sa_query *query;
923 struct ib_mad_send_wc mad_send_wc;
926 spin_lock_irqsave(&ib_nl_request_lock, flags);
927 while (!list_empty(&ib_nl_request_list)) {
928 query = list_entry(ib_nl_request_list.next,
929 struct ib_sa_query, list);
931 if (time_after(query->timeout, jiffies)) {
932 delay = query->timeout - jiffies;
933 if ((long)delay <= 0)
935 queue_delayed_work(ib_nl_wq, &ib_nl_timed_work, delay);
939 list_del(&query->list);
940 ib_sa_disable_local_svc(query);
941 /* Hold the lock to protect against query cancellation */
942 if (ib_sa_query_cancelled(query))
945 ret = ib_post_send_mad(query->mad_buf, NULL);
947 mad_send_wc.send_buf = query->mad_buf;
948 mad_send_wc.status = IB_WC_WR_FLUSH_ERR;
949 spin_unlock_irqrestore(&ib_nl_request_lock, flags);
950 send_handler(query->port->agent, &mad_send_wc);
951 spin_lock_irqsave(&ib_nl_request_lock, flags);
954 spin_unlock_irqrestore(&ib_nl_request_lock, flags);
957 int ib_nl_handle_set_timeout(struct sk_buff *skb,
958 struct nlmsghdr *nlh,
959 struct netlink_ext_ack *extack)
961 int timeout, delta, abs_delta;
962 const struct nlattr *attr;
964 struct ib_sa_query *query;
966 struct nlattr *tb[LS_NLA_TYPE_MAX];
969 if (!(nlh->nlmsg_flags & NLM_F_REQUEST) ||
970 !(NETLINK_CB(skb).sk))
973 ret = nla_parse_deprecated(tb, LS_NLA_TYPE_MAX - 1, nlmsg_data(nlh),
974 nlmsg_len(nlh), ib_nl_policy, NULL);
975 attr = (const struct nlattr *)tb[LS_NLA_TYPE_TIMEOUT];
979 timeout = *(int *) nla_data(attr);
980 if (timeout < IB_SA_LOCAL_SVC_TIMEOUT_MIN)
981 timeout = IB_SA_LOCAL_SVC_TIMEOUT_MIN;
982 if (timeout > IB_SA_LOCAL_SVC_TIMEOUT_MAX)
983 timeout = IB_SA_LOCAL_SVC_TIMEOUT_MAX;
985 delta = timeout - sa_local_svc_timeout_ms;
992 spin_lock_irqsave(&ib_nl_request_lock, flags);
993 sa_local_svc_timeout_ms = timeout;
994 list_for_each_entry(query, &ib_nl_request_list, list) {
995 if (delta < 0 && abs_delta > query->timeout)
998 query->timeout += delta;
1000 /* Get the new delay from the first entry */
1002 delay = query->timeout - jiffies;
1008 mod_delayed_work(ib_nl_wq, &ib_nl_timed_work,
1009 (unsigned long)delay);
1010 spin_unlock_irqrestore(&ib_nl_request_lock, flags);
1017 static inline int ib_nl_is_good_resolve_resp(const struct nlmsghdr *nlh)
1019 struct nlattr *tb[LS_NLA_TYPE_MAX];
1022 if (nlh->nlmsg_flags & RDMA_NL_LS_F_ERR)
1025 ret = nla_parse_deprecated(tb, LS_NLA_TYPE_MAX - 1, nlmsg_data(nlh),
1026 nlmsg_len(nlh), ib_nl_policy, NULL);
1033 int ib_nl_handle_resolve_resp(struct sk_buff *skb,
1034 struct nlmsghdr *nlh,
1035 struct netlink_ext_ack *extack)
1037 unsigned long flags;
1038 struct ib_sa_query *query;
1039 struct ib_mad_send_buf *send_buf;
1040 struct ib_mad_send_wc mad_send_wc;
1044 if ((nlh->nlmsg_flags & NLM_F_REQUEST) ||
1045 !(NETLINK_CB(skb).sk))
1048 spin_lock_irqsave(&ib_nl_request_lock, flags);
1049 list_for_each_entry(query, &ib_nl_request_list, list) {
1051 * If the query is cancelled, let the timeout routine
1054 if (nlh->nlmsg_seq == query->seq) {
1055 found = !ib_sa_query_cancelled(query);
1057 list_del(&query->list);
1063 spin_unlock_irqrestore(&ib_nl_request_lock, flags);
1067 send_buf = query->mad_buf;
1069 if (!ib_nl_is_good_resolve_resp(nlh)) {
1070 /* if the result is a failure, send out the packet via IB */
1071 ib_sa_disable_local_svc(query);
1072 ret = ib_post_send_mad(query->mad_buf, NULL);
1073 spin_unlock_irqrestore(&ib_nl_request_lock, flags);
1075 mad_send_wc.send_buf = send_buf;
1076 mad_send_wc.status = IB_WC_GENERAL_ERR;
1077 send_handler(query->port->agent, &mad_send_wc);
1080 spin_unlock_irqrestore(&ib_nl_request_lock, flags);
1081 ib_nl_process_good_resolve_rsp(query, nlh);
1088 static void free_sm_ah(struct kref *kref)
1090 struct ib_sa_sm_ah *sm_ah = container_of(kref, struct ib_sa_sm_ah, ref);
1092 rdma_destroy_ah(sm_ah->ah, 0);
1096 void ib_sa_register_client(struct ib_sa_client *client)
1098 atomic_set(&client->users, 1);
1099 init_completion(&client->comp);
1101 EXPORT_SYMBOL(ib_sa_register_client);
1103 void ib_sa_unregister_client(struct ib_sa_client *client)
1105 ib_sa_client_put(client);
1106 wait_for_completion(&client->comp);
1108 EXPORT_SYMBOL(ib_sa_unregister_client);
1111 * ib_sa_cancel_query - try to cancel an SA query
1112 * @id:ID of query to cancel
1113 * @query:query pointer to cancel
1115 * Try to cancel an SA query. If the id and query don't match up or
1116 * the query has already completed, nothing is done. Otherwise the
1117 * query is canceled and will complete with a status of -EINTR.
1119 void ib_sa_cancel_query(int id, struct ib_sa_query *query)
1121 unsigned long flags;
1122 struct ib_mad_send_buf *mad_buf;
1124 xa_lock_irqsave(&queries, flags);
1125 if (xa_load(&queries, id) != query) {
1126 xa_unlock_irqrestore(&queries, flags);
1129 mad_buf = query->mad_buf;
1130 xa_unlock_irqrestore(&queries, flags);
1133 * If the query is still on the netlink request list, schedule
1134 * it to be cancelled by the timeout routine. Otherwise, it has been
1135 * sent to the MAD layer and has to be cancelled from there.
1137 if (!ib_nl_cancel_request(query))
1138 ib_cancel_mad(mad_buf);
1140 EXPORT_SYMBOL(ib_sa_cancel_query);
1142 static u8 get_src_path_mask(struct ib_device *device, u32 port_num)
1144 struct ib_sa_device *sa_dev;
1145 struct ib_sa_port *port;
1146 unsigned long flags;
1149 sa_dev = ib_get_client_data(device, &sa_client);
1153 port = &sa_dev->port[port_num - sa_dev->start_port];
1154 spin_lock_irqsave(&port->ah_lock, flags);
1155 src_path_mask = port->sm_ah ? port->sm_ah->src_path_mask : 0x7f;
1156 spin_unlock_irqrestore(&port->ah_lock, flags);
1158 return src_path_mask;
1161 static int init_ah_attr_grh_fields(struct ib_device *device, u32 port_num,
1162 struct sa_path_rec *rec,
1163 struct rdma_ah_attr *ah_attr,
1164 const struct ib_gid_attr *gid_attr)
1166 enum ib_gid_type type = sa_conv_pathrec_to_gid_type(rec);
1169 gid_attr = rdma_find_gid_by_port(device, &rec->sgid, type,
1171 if (IS_ERR(gid_attr))
1172 return PTR_ERR(gid_attr);
1174 rdma_hold_gid_attr(gid_attr);
1176 rdma_move_grh_sgid_attr(ah_attr, &rec->dgid,
1177 be32_to_cpu(rec->flow_label),
1178 rec->hop_limit, rec->traffic_class,
1184 * ib_init_ah_attr_from_path - Initialize address handle attributes based on
1185 * an SA path record.
1186 * @device: Device associated ah attributes initialization.
1187 * @port_num: Port on the specified device.
1188 * @rec: path record entry to use for ah attributes initialization.
1189 * @ah_attr: address handle attributes to initialization from path record.
1190 * @gid_attr: SGID attribute to consider during initialization.
1192 * When ib_init_ah_attr_from_path() returns success,
1193 * (a) for IB link layer it optionally contains a reference to SGID attribute
1194 * when GRH is present for IB link layer.
1195 * (b) for RoCE link layer it contains a reference to SGID attribute.
1196 * User must invoke rdma_destroy_ah_attr() to release reference to SGID
1197 * attributes which are initialized using ib_init_ah_attr_from_path().
1199 int ib_init_ah_attr_from_path(struct ib_device *device, u32 port_num,
1200 struct sa_path_rec *rec,
1201 struct rdma_ah_attr *ah_attr,
1202 const struct ib_gid_attr *gid_attr)
1206 memset(ah_attr, 0, sizeof(*ah_attr));
1207 ah_attr->type = rdma_ah_find_type(device, port_num);
1208 rdma_ah_set_sl(ah_attr, rec->sl);
1209 rdma_ah_set_port_num(ah_attr, port_num);
1210 rdma_ah_set_static_rate(ah_attr, rec->rate);
1212 if (sa_path_is_roce(rec)) {
1213 ret = roce_resolve_route_from_path(rec, gid_attr);
1217 memcpy(ah_attr->roce.dmac, sa_path_get_dmac(rec), ETH_ALEN);
1219 rdma_ah_set_dlid(ah_attr, be32_to_cpu(sa_path_get_dlid(rec)));
1220 if (sa_path_is_opa(rec) &&
1221 rdma_ah_get_dlid(ah_attr) == be16_to_cpu(IB_LID_PERMISSIVE))
1222 rdma_ah_set_make_grd(ah_attr, true);
1224 rdma_ah_set_path_bits(ah_attr,
1225 be32_to_cpu(sa_path_get_slid(rec)) &
1226 get_src_path_mask(device, port_num));
1229 if (rec->hop_limit > 0 || sa_path_is_roce(rec))
1230 ret = init_ah_attr_grh_fields(device, port_num,
1231 rec, ah_attr, gid_attr);
1234 EXPORT_SYMBOL(ib_init_ah_attr_from_path);
1236 static int alloc_mad(struct ib_sa_query *query, gfp_t gfp_mask)
1238 struct rdma_ah_attr ah_attr;
1239 unsigned long flags;
1241 spin_lock_irqsave(&query->port->ah_lock, flags);
1242 if (!query->port->sm_ah) {
1243 spin_unlock_irqrestore(&query->port->ah_lock, flags);
1246 kref_get(&query->port->sm_ah->ref);
1247 query->sm_ah = query->port->sm_ah;
1248 spin_unlock_irqrestore(&query->port->ah_lock, flags);
1251 * Always check if sm_ah has valid dlid assigned,
1252 * before querying for class port info
1254 if ((rdma_query_ah(query->sm_ah->ah, &ah_attr) < 0) ||
1255 !rdma_is_valid_unicast_lid(&ah_attr)) {
1256 kref_put(&query->sm_ah->ref, free_sm_ah);
1259 query->mad_buf = ib_create_send_mad(query->port->agent, 1,
1260 query->sm_ah->pkey_index,
1261 0, IB_MGMT_SA_HDR, IB_MGMT_SA_DATA,
1263 ((query->flags & IB_SA_QUERY_OPA) ?
1264 OPA_MGMT_BASE_VERSION :
1265 IB_MGMT_BASE_VERSION));
1266 if (IS_ERR(query->mad_buf)) {
1267 kref_put(&query->sm_ah->ref, free_sm_ah);
1271 query->mad_buf->ah = query->sm_ah->ah;
1276 static void free_mad(struct ib_sa_query *query)
1278 ib_free_send_mad(query->mad_buf);
1279 kref_put(&query->sm_ah->ref, free_sm_ah);
1282 static void init_mad(struct ib_sa_query *query, struct ib_mad_agent *agent)
1284 struct ib_sa_mad *mad = query->mad_buf->mad;
1285 unsigned long flags;
1287 memset(mad, 0, sizeof *mad);
1289 if (query->flags & IB_SA_QUERY_OPA) {
1290 mad->mad_hdr.base_version = OPA_MGMT_BASE_VERSION;
1291 mad->mad_hdr.class_version = OPA_SA_CLASS_VERSION;
1293 mad->mad_hdr.base_version = IB_MGMT_BASE_VERSION;
1294 mad->mad_hdr.class_version = IB_SA_CLASS_VERSION;
1296 mad->mad_hdr.mgmt_class = IB_MGMT_CLASS_SUBN_ADM;
1297 spin_lock_irqsave(&tid_lock, flags);
1299 cpu_to_be64(((u64) agent->hi_tid) << 32 | tid++);
1300 spin_unlock_irqrestore(&tid_lock, flags);
1303 static int send_mad(struct ib_sa_query *query, unsigned long timeout_ms,
1306 unsigned long flags;
1308 const int nmbr_sa_query_retries = 10;
1310 xa_lock_irqsave(&queries, flags);
1311 ret = __xa_alloc(&queries, &id, query, xa_limit_32b, gfp_mask);
1312 xa_unlock_irqrestore(&queries, flags);
1316 query->mad_buf->timeout_ms = timeout_ms / nmbr_sa_query_retries;
1317 query->mad_buf->retries = nmbr_sa_query_retries;
1318 if (!query->mad_buf->timeout_ms) {
1319 /* Special case, very small timeout_ms */
1320 query->mad_buf->timeout_ms = 1;
1321 query->mad_buf->retries = timeout_ms;
1323 query->mad_buf->context[0] = query;
1326 if ((query->flags & IB_SA_ENABLE_LOCAL_SERVICE) &&
1327 (!(query->flags & IB_SA_QUERY_OPA))) {
1328 if (rdma_nl_chk_listeners(RDMA_NL_GROUP_LS)) {
1329 if (!ib_nl_make_request(query, gfp_mask))
1332 ib_sa_disable_local_svc(query);
1335 ret = ib_post_send_mad(query->mad_buf, NULL);
1337 xa_lock_irqsave(&queries, flags);
1338 __xa_erase(&queries, id);
1339 xa_unlock_irqrestore(&queries, flags);
1343 * It's not safe to dereference query any more, because the
1344 * send may already have completed and freed the query in
1347 return ret ? ret : id;
1350 void ib_sa_unpack_path(void *attribute, struct sa_path_rec *rec)
1352 ib_unpack(path_rec_table, ARRAY_SIZE(path_rec_table), attribute, rec);
1354 EXPORT_SYMBOL(ib_sa_unpack_path);
1356 void ib_sa_pack_path(struct sa_path_rec *rec, void *attribute)
1358 ib_pack(path_rec_table, ARRAY_SIZE(path_rec_table), rec, attribute);
1360 EXPORT_SYMBOL(ib_sa_pack_path);
1362 static bool ib_sa_opa_pathrecord_support(struct ib_sa_client *client,
1363 struct ib_sa_device *sa_dev,
1366 struct ib_sa_port *port;
1367 unsigned long flags;
1370 port = &sa_dev->port[port_num - sa_dev->start_port];
1371 spin_lock_irqsave(&port->classport_lock, flags);
1372 if (!port->classport_info.valid)
1375 if (port->classport_info.data.type == RDMA_CLASS_PORT_INFO_OPA)
1376 ret = opa_get_cpi_capmask2(&port->classport_info.data.opa) &
1377 OPA_CLASS_PORT_INFO_PR_SUPPORT;
1379 spin_unlock_irqrestore(&port->classport_lock, flags);
1383 enum opa_pr_supported {
1390 * opa_pr_query_possible - Check if current PR query can be an OPA query.
1392 * Retuns PR_NOT_SUPPORTED if a path record query is not
1393 * possible, PR_OPA_SUPPORTED if an OPA path record query
1394 * is possible and PR_IB_SUPPORTED if an IB path record
1395 * query is possible.
1397 static int opa_pr_query_possible(struct ib_sa_client *client,
1398 struct ib_sa_device *sa_dev,
1399 struct ib_device *device, u32 port_num)
1401 struct ib_port_attr port_attr;
1403 if (ib_query_port(device, port_num, &port_attr))
1404 return PR_NOT_SUPPORTED;
1406 if (ib_sa_opa_pathrecord_support(client, sa_dev, port_num))
1407 return PR_OPA_SUPPORTED;
1409 if (port_attr.lid >= be16_to_cpu(IB_MULTICAST_LID_BASE))
1410 return PR_NOT_SUPPORTED;
1412 return PR_IB_SUPPORTED;
1415 static void ib_sa_path_rec_callback(struct ib_sa_query *sa_query,
1417 struct ib_sa_mad *mad)
1419 struct ib_sa_path_query *query =
1420 container_of(sa_query, struct ib_sa_path_query, sa_query);
1423 struct sa_path_rec rec;
1425 if (sa_query->flags & IB_SA_QUERY_OPA) {
1426 ib_unpack(opa_path_rec_table,
1427 ARRAY_SIZE(opa_path_rec_table),
1429 rec.rec_type = SA_PATH_REC_TYPE_OPA;
1430 query->callback(status, &rec, query->context);
1432 ib_unpack(path_rec_table,
1433 ARRAY_SIZE(path_rec_table),
1435 rec.rec_type = SA_PATH_REC_TYPE_IB;
1436 sa_path_set_dmac_zero(&rec);
1438 if (query->conv_pr) {
1439 struct sa_path_rec opa;
1441 memset(&opa, 0, sizeof(struct sa_path_rec));
1442 sa_convert_path_ib_to_opa(&opa, &rec);
1443 query->callback(status, &opa, query->context);
1445 query->callback(status, &rec, query->context);
1449 query->callback(status, NULL, query->context);
1452 static void ib_sa_path_rec_release(struct ib_sa_query *sa_query)
1454 struct ib_sa_path_query *query =
1455 container_of(sa_query, struct ib_sa_path_query, sa_query);
1457 kfree(query->conv_pr);
1462 * ib_sa_path_rec_get - Start a Path get query
1464 * @device:device to send query on
1465 * @port_num: port number to send query on
1466 * @rec:Path Record to send in query
1467 * @comp_mask:component mask to send in query
1468 * @timeout_ms:time to wait for response
1469 * @gfp_mask:GFP mask to use for internal allocations
1470 * @callback:function called when query completes, times out or is
1472 * @context:opaque user context passed to callback
1473 * @sa_query:query context, used to cancel query
1475 * Send a Path Record Get query to the SA to look up a path. The
1476 * callback function will be called when the query completes (or
1477 * fails); status is 0 for a successful response, -EINTR if the query
1478 * is canceled, -ETIMEDOUT is the query timed out, or -EIO if an error
1479 * occurred sending the query. The resp parameter of the callback is
1480 * only valid if status is 0.
1482 * If the return value of ib_sa_path_rec_get() is negative, it is an
1483 * error code. Otherwise it is a query ID that can be used to cancel
1486 int ib_sa_path_rec_get(struct ib_sa_client *client,
1487 struct ib_device *device, u32 port_num,
1488 struct sa_path_rec *rec,
1489 ib_sa_comp_mask comp_mask,
1490 unsigned long timeout_ms, gfp_t gfp_mask,
1491 void (*callback)(int status,
1492 struct sa_path_rec *resp,
1495 struct ib_sa_query **sa_query)
1497 struct ib_sa_path_query *query;
1498 struct ib_sa_device *sa_dev = ib_get_client_data(device, &sa_client);
1499 struct ib_sa_port *port;
1500 struct ib_mad_agent *agent;
1501 struct ib_sa_mad *mad;
1502 enum opa_pr_supported status;
1508 if ((rec->rec_type != SA_PATH_REC_TYPE_IB) &&
1509 (rec->rec_type != SA_PATH_REC_TYPE_OPA))
1512 port = &sa_dev->port[port_num - sa_dev->start_port];
1513 agent = port->agent;
1515 query = kzalloc(sizeof(*query), gfp_mask);
1519 query->sa_query.port = port;
1520 if (rec->rec_type == SA_PATH_REC_TYPE_OPA) {
1521 status = opa_pr_query_possible(client, sa_dev, device, port_num);
1522 if (status == PR_NOT_SUPPORTED) {
1525 } else if (status == PR_OPA_SUPPORTED) {
1526 query->sa_query.flags |= IB_SA_QUERY_OPA;
1529 kmalloc(sizeof(*query->conv_pr), gfp_mask);
1530 if (!query->conv_pr) {
1537 ret = alloc_mad(&query->sa_query, gfp_mask);
1541 ib_sa_client_get(client);
1542 query->sa_query.client = client;
1543 query->callback = callback;
1544 query->context = context;
1546 mad = query->sa_query.mad_buf->mad;
1547 init_mad(&query->sa_query, agent);
1549 query->sa_query.callback = callback ? ib_sa_path_rec_callback : NULL;
1550 query->sa_query.release = ib_sa_path_rec_release;
1551 mad->mad_hdr.method = IB_MGMT_METHOD_GET;
1552 mad->mad_hdr.attr_id = cpu_to_be16(IB_SA_ATTR_PATH_REC);
1553 mad->sa_hdr.comp_mask = comp_mask;
1555 if (query->sa_query.flags & IB_SA_QUERY_OPA) {
1556 ib_pack(opa_path_rec_table, ARRAY_SIZE(opa_path_rec_table),
1558 } else if (query->conv_pr) {
1559 sa_convert_path_opa_to_ib(query->conv_pr, rec);
1560 ib_pack(path_rec_table, ARRAY_SIZE(path_rec_table),
1561 query->conv_pr, mad->data);
1563 ib_pack(path_rec_table, ARRAY_SIZE(path_rec_table),
1567 *sa_query = &query->sa_query;
1569 query->sa_query.flags |= IB_SA_ENABLE_LOCAL_SERVICE;
1570 query->sa_query.mad_buf->context[1] = (query->conv_pr) ?
1571 query->conv_pr : rec;
1573 ret = send_mad(&query->sa_query, timeout_ms, gfp_mask);
1581 ib_sa_client_put(query->sa_query.client);
1582 free_mad(&query->sa_query);
1584 kfree(query->conv_pr);
1589 EXPORT_SYMBOL(ib_sa_path_rec_get);
1591 static void ib_sa_mcmember_rec_callback(struct ib_sa_query *sa_query,
1593 struct ib_sa_mad *mad)
1595 struct ib_sa_mcmember_query *query =
1596 container_of(sa_query, struct ib_sa_mcmember_query, sa_query);
1599 struct ib_sa_mcmember_rec rec;
1601 ib_unpack(mcmember_rec_table, ARRAY_SIZE(mcmember_rec_table),
1603 query->callback(status, &rec, query->context);
1605 query->callback(status, NULL, query->context);
1608 static void ib_sa_mcmember_rec_release(struct ib_sa_query *sa_query)
1610 kfree(container_of(sa_query, struct ib_sa_mcmember_query, sa_query));
1613 int ib_sa_mcmember_rec_query(struct ib_sa_client *client,
1614 struct ib_device *device, u32 port_num,
1616 struct ib_sa_mcmember_rec *rec,
1617 ib_sa_comp_mask comp_mask,
1618 unsigned long timeout_ms, gfp_t gfp_mask,
1619 void (*callback)(int status,
1620 struct ib_sa_mcmember_rec *resp,
1623 struct ib_sa_query **sa_query)
1625 struct ib_sa_mcmember_query *query;
1626 struct ib_sa_device *sa_dev = ib_get_client_data(device, &sa_client);
1627 struct ib_sa_port *port;
1628 struct ib_mad_agent *agent;
1629 struct ib_sa_mad *mad;
1635 port = &sa_dev->port[port_num - sa_dev->start_port];
1636 agent = port->agent;
1638 query = kzalloc(sizeof(*query), gfp_mask);
1642 query->sa_query.port = port;
1643 ret = alloc_mad(&query->sa_query, gfp_mask);
1647 ib_sa_client_get(client);
1648 query->sa_query.client = client;
1649 query->callback = callback;
1650 query->context = context;
1652 mad = query->sa_query.mad_buf->mad;
1653 init_mad(&query->sa_query, agent);
1655 query->sa_query.callback = callback ? ib_sa_mcmember_rec_callback : NULL;
1656 query->sa_query.release = ib_sa_mcmember_rec_release;
1657 mad->mad_hdr.method = method;
1658 mad->mad_hdr.attr_id = cpu_to_be16(IB_SA_ATTR_MC_MEMBER_REC);
1659 mad->sa_hdr.comp_mask = comp_mask;
1661 ib_pack(mcmember_rec_table, ARRAY_SIZE(mcmember_rec_table),
1664 *sa_query = &query->sa_query;
1666 ret = send_mad(&query->sa_query, timeout_ms, gfp_mask);
1674 ib_sa_client_put(query->sa_query.client);
1675 free_mad(&query->sa_query);
1682 /* Support GuidInfoRecord */
1683 static void ib_sa_guidinfo_rec_callback(struct ib_sa_query *sa_query,
1685 struct ib_sa_mad *mad)
1687 struct ib_sa_guidinfo_query *query =
1688 container_of(sa_query, struct ib_sa_guidinfo_query, sa_query);
1691 struct ib_sa_guidinfo_rec rec;
1693 ib_unpack(guidinfo_rec_table, ARRAY_SIZE(guidinfo_rec_table),
1695 query->callback(status, &rec, query->context);
1697 query->callback(status, NULL, query->context);
1700 static void ib_sa_guidinfo_rec_release(struct ib_sa_query *sa_query)
1702 kfree(container_of(sa_query, struct ib_sa_guidinfo_query, sa_query));
1705 int ib_sa_guid_info_rec_query(struct ib_sa_client *client,
1706 struct ib_device *device, u32 port_num,
1707 struct ib_sa_guidinfo_rec *rec,
1708 ib_sa_comp_mask comp_mask, u8 method,
1709 unsigned long timeout_ms, gfp_t gfp_mask,
1710 void (*callback)(int status,
1711 struct ib_sa_guidinfo_rec *resp,
1714 struct ib_sa_query **sa_query)
1716 struct ib_sa_guidinfo_query *query;
1717 struct ib_sa_device *sa_dev = ib_get_client_data(device, &sa_client);
1718 struct ib_sa_port *port;
1719 struct ib_mad_agent *agent;
1720 struct ib_sa_mad *mad;
1726 if (method != IB_MGMT_METHOD_GET &&
1727 method != IB_MGMT_METHOD_SET &&
1728 method != IB_SA_METHOD_DELETE) {
1732 port = &sa_dev->port[port_num - sa_dev->start_port];
1733 agent = port->agent;
1735 query = kzalloc(sizeof(*query), gfp_mask);
1739 query->sa_query.port = port;
1740 ret = alloc_mad(&query->sa_query, gfp_mask);
1744 ib_sa_client_get(client);
1745 query->sa_query.client = client;
1746 query->callback = callback;
1747 query->context = context;
1749 mad = query->sa_query.mad_buf->mad;
1750 init_mad(&query->sa_query, agent);
1752 query->sa_query.callback = callback ? ib_sa_guidinfo_rec_callback : NULL;
1753 query->sa_query.release = ib_sa_guidinfo_rec_release;
1755 mad->mad_hdr.method = method;
1756 mad->mad_hdr.attr_id = cpu_to_be16(IB_SA_ATTR_GUID_INFO_REC);
1757 mad->sa_hdr.comp_mask = comp_mask;
1759 ib_pack(guidinfo_rec_table, ARRAY_SIZE(guidinfo_rec_table), rec,
1762 *sa_query = &query->sa_query;
1764 ret = send_mad(&query->sa_query, timeout_ms, gfp_mask);
1772 ib_sa_client_put(query->sa_query.client);
1773 free_mad(&query->sa_query);
1779 EXPORT_SYMBOL(ib_sa_guid_info_rec_query);
1781 struct ib_classport_info_context {
1782 struct completion done;
1783 struct ib_sa_query *sa_query;
1786 static void ib_classportinfo_cb(void *context)
1788 struct ib_classport_info_context *cb_ctx = context;
1790 complete(&cb_ctx->done);
1793 static void ib_sa_classport_info_rec_callback(struct ib_sa_query *sa_query,
1795 struct ib_sa_mad *mad)
1797 unsigned long flags;
1798 struct ib_sa_classport_info_query *query =
1799 container_of(sa_query, struct ib_sa_classport_info_query, sa_query);
1800 struct ib_sa_classport_cache *info = &sa_query->port->classport_info;
1803 if (sa_query->flags & IB_SA_QUERY_OPA) {
1804 struct opa_class_port_info rec;
1806 ib_unpack(opa_classport_info_rec_table,
1807 ARRAY_SIZE(opa_classport_info_rec_table),
1810 spin_lock_irqsave(&sa_query->port->classport_lock,
1812 if (!status && !info->valid) {
1813 memcpy(&info->data.opa, &rec,
1814 sizeof(info->data.opa));
1817 info->data.type = RDMA_CLASS_PORT_INFO_OPA;
1819 spin_unlock_irqrestore(&sa_query->port->classport_lock,
1823 struct ib_class_port_info rec;
1825 ib_unpack(ib_classport_info_rec_table,
1826 ARRAY_SIZE(ib_classport_info_rec_table),
1829 spin_lock_irqsave(&sa_query->port->classport_lock,
1831 if (!status && !info->valid) {
1832 memcpy(&info->data.ib, &rec,
1833 sizeof(info->data.ib));
1836 info->data.type = RDMA_CLASS_PORT_INFO_IB;
1838 spin_unlock_irqrestore(&sa_query->port->classport_lock,
1842 query->callback(query->context);
1845 static void ib_sa_classport_info_rec_release(struct ib_sa_query *sa_query)
1847 kfree(container_of(sa_query, struct ib_sa_classport_info_query,
1851 static int ib_sa_classport_info_rec_query(struct ib_sa_port *port,
1852 unsigned long timeout_ms,
1853 void (*callback)(void *context),
1855 struct ib_sa_query **sa_query)
1857 struct ib_mad_agent *agent;
1858 struct ib_sa_classport_info_query *query;
1859 struct ib_sa_mad *mad;
1860 gfp_t gfp_mask = GFP_KERNEL;
1863 agent = port->agent;
1865 query = kzalloc(sizeof(*query), gfp_mask);
1869 query->sa_query.port = port;
1870 query->sa_query.flags |= rdma_cap_opa_ah(port->agent->device,
1872 IB_SA_QUERY_OPA : 0;
1873 ret = alloc_mad(&query->sa_query, gfp_mask);
1877 query->callback = callback;
1878 query->context = context;
1880 mad = query->sa_query.mad_buf->mad;
1881 init_mad(&query->sa_query, agent);
1883 query->sa_query.callback = ib_sa_classport_info_rec_callback;
1884 query->sa_query.release = ib_sa_classport_info_rec_release;
1885 mad->mad_hdr.method = IB_MGMT_METHOD_GET;
1886 mad->mad_hdr.attr_id = cpu_to_be16(IB_SA_ATTR_CLASS_PORTINFO);
1887 mad->sa_hdr.comp_mask = 0;
1888 *sa_query = &query->sa_query;
1890 ret = send_mad(&query->sa_query, timeout_ms, gfp_mask);
1898 free_mad(&query->sa_query);
1905 static void update_ib_cpi(struct work_struct *work)
1907 struct ib_sa_port *port =
1908 container_of(work, struct ib_sa_port, ib_cpi_work.work);
1909 struct ib_classport_info_context *cb_context;
1910 unsigned long flags;
1913 /* If the classport info is valid, nothing
1916 spin_lock_irqsave(&port->classport_lock, flags);
1917 if (port->classport_info.valid) {
1918 spin_unlock_irqrestore(&port->classport_lock, flags);
1921 spin_unlock_irqrestore(&port->classport_lock, flags);
1923 cb_context = kmalloc(sizeof(*cb_context), GFP_KERNEL);
1927 init_completion(&cb_context->done);
1929 ret = ib_sa_classport_info_rec_query(port, 3000,
1930 ib_classportinfo_cb, cb_context,
1931 &cb_context->sa_query);
1934 wait_for_completion(&cb_context->done);
1937 spin_lock_irqsave(&port->classport_lock, flags);
1939 /* If the classport info is still not valid, the query should have
1940 * failed for some reason. Retry issuing the query
1942 if (!port->classport_info.valid) {
1943 port->classport_info.retry_cnt++;
1944 if (port->classport_info.retry_cnt <=
1945 IB_SA_CPI_MAX_RETRY_CNT) {
1946 unsigned long delay =
1947 msecs_to_jiffies(IB_SA_CPI_RETRY_WAIT);
1949 queue_delayed_work(ib_wq, &port->ib_cpi_work, delay);
1952 spin_unlock_irqrestore(&port->classport_lock, flags);
1958 static void send_handler(struct ib_mad_agent *agent,
1959 struct ib_mad_send_wc *mad_send_wc)
1961 struct ib_sa_query *query = mad_send_wc->send_buf->context[0];
1962 unsigned long flags;
1964 if (query->callback)
1965 switch (mad_send_wc->status) {
1967 /* No callback -- already got recv */
1969 case IB_WC_RESP_TIMEOUT_ERR:
1970 query->callback(query, -ETIMEDOUT, NULL);
1972 case IB_WC_WR_FLUSH_ERR:
1973 query->callback(query, -EINTR, NULL);
1976 query->callback(query, -EIO, NULL);
1980 xa_lock_irqsave(&queries, flags);
1981 __xa_erase(&queries, query->id);
1982 xa_unlock_irqrestore(&queries, flags);
1986 ib_sa_client_put(query->client);
1987 query->release(query);
1990 static void recv_handler(struct ib_mad_agent *mad_agent,
1991 struct ib_mad_send_buf *send_buf,
1992 struct ib_mad_recv_wc *mad_recv_wc)
1994 struct ib_sa_query *query;
1999 query = send_buf->context[0];
2000 if (query->callback) {
2001 if (mad_recv_wc->wc->status == IB_WC_SUCCESS)
2002 query->callback(query,
2003 mad_recv_wc->recv_buf.mad->mad_hdr.status ?
2005 (struct ib_sa_mad *) mad_recv_wc->recv_buf.mad);
2007 query->callback(query, -EIO, NULL);
2010 ib_free_recv_mad(mad_recv_wc);
2013 static void update_sm_ah(struct work_struct *work)
2015 struct ib_sa_port *port =
2016 container_of(work, struct ib_sa_port, update_task);
2017 struct ib_sa_sm_ah *new_ah;
2018 struct ib_port_attr port_attr;
2019 struct rdma_ah_attr ah_attr;
2022 if (ib_query_port(port->agent->device, port->port_num, &port_attr)) {
2023 pr_warn("Couldn't query port\n");
2027 new_ah = kmalloc(sizeof(*new_ah), GFP_KERNEL);
2031 kref_init(&new_ah->ref);
2032 new_ah->src_path_mask = (1 << port_attr.lmc) - 1;
2034 new_ah->pkey_index = 0;
2035 if (ib_find_pkey(port->agent->device, port->port_num,
2036 IB_DEFAULT_PKEY_FULL, &new_ah->pkey_index))
2037 pr_err("Couldn't find index for default PKey\n");
2039 memset(&ah_attr, 0, sizeof(ah_attr));
2040 ah_attr.type = rdma_ah_find_type(port->agent->device,
2042 rdma_ah_set_dlid(&ah_attr, port_attr.sm_lid);
2043 rdma_ah_set_sl(&ah_attr, port_attr.sm_sl);
2044 rdma_ah_set_port_num(&ah_attr, port->port_num);
2046 grh_required = rdma_is_grh_required(port->agent->device,
2050 * The OPA sm_lid of 0xFFFF needs special handling so that it can be
2051 * differentiated from a permissive LID of 0xFFFF. We set the
2052 * grh_required flag here so the SA can program the DGID in the
2053 * address handle appropriately
2055 if (ah_attr.type == RDMA_AH_ATTR_TYPE_OPA &&
2057 port_attr.sm_lid == be16_to_cpu(IB_LID_PERMISSIVE)))
2058 rdma_ah_set_make_grd(&ah_attr, true);
2060 if (ah_attr.type == RDMA_AH_ATTR_TYPE_IB && grh_required) {
2061 rdma_ah_set_ah_flags(&ah_attr, IB_AH_GRH);
2062 rdma_ah_set_subnet_prefix(&ah_attr,
2063 cpu_to_be64(port_attr.subnet_prefix));
2064 rdma_ah_set_interface_id(&ah_attr,
2065 cpu_to_be64(IB_SA_WELL_KNOWN_GUID));
2068 new_ah->ah = rdma_create_ah(port->agent->qp->pd, &ah_attr,
2069 RDMA_CREATE_AH_SLEEPABLE);
2070 if (IS_ERR(new_ah->ah)) {
2071 pr_warn("Couldn't create new SM AH\n");
2076 spin_lock_irq(&port->ah_lock);
2078 kref_put(&port->sm_ah->ref, free_sm_ah);
2079 port->sm_ah = new_ah;
2080 spin_unlock_irq(&port->ah_lock);
2083 static void ib_sa_event(struct ib_event_handler *handler,
2084 struct ib_event *event)
2086 if (event->event == IB_EVENT_PORT_ERR ||
2087 event->event == IB_EVENT_PORT_ACTIVE ||
2088 event->event == IB_EVENT_LID_CHANGE ||
2089 event->event == IB_EVENT_PKEY_CHANGE ||
2090 event->event == IB_EVENT_SM_CHANGE ||
2091 event->event == IB_EVENT_CLIENT_REREGISTER) {
2092 unsigned long flags;
2093 struct ib_sa_device *sa_dev =
2094 container_of(handler, typeof(*sa_dev), event_handler);
2095 u32 port_num = event->element.port_num - sa_dev->start_port;
2096 struct ib_sa_port *port = &sa_dev->port[port_num];
2098 if (!rdma_cap_ib_sa(handler->device, port->port_num))
2101 spin_lock_irqsave(&port->ah_lock, flags);
2103 kref_put(&port->sm_ah->ref, free_sm_ah);
2105 spin_unlock_irqrestore(&port->ah_lock, flags);
2107 if (event->event == IB_EVENT_SM_CHANGE ||
2108 event->event == IB_EVENT_CLIENT_REREGISTER ||
2109 event->event == IB_EVENT_LID_CHANGE ||
2110 event->event == IB_EVENT_PORT_ACTIVE) {
2111 unsigned long delay =
2112 msecs_to_jiffies(IB_SA_CPI_RETRY_WAIT);
2114 spin_lock_irqsave(&port->classport_lock, flags);
2115 port->classport_info.valid = false;
2116 port->classport_info.retry_cnt = 0;
2117 spin_unlock_irqrestore(&port->classport_lock, flags);
2118 queue_delayed_work(ib_wq,
2119 &port->ib_cpi_work, delay);
2121 queue_work(ib_wq, &sa_dev->port[port_num].update_task);
2125 static int ib_sa_add_one(struct ib_device *device)
2127 struct ib_sa_device *sa_dev;
2132 s = rdma_start_port(device);
2133 e = rdma_end_port(device);
2135 sa_dev = kzalloc(struct_size(sa_dev, port, e - s + 1), GFP_KERNEL);
2139 sa_dev->start_port = s;
2140 sa_dev->end_port = e;
2142 for (i = 0; i <= e - s; ++i) {
2143 spin_lock_init(&sa_dev->port[i].ah_lock);
2144 if (!rdma_cap_ib_sa(device, i + 1))
2147 sa_dev->port[i].sm_ah = NULL;
2148 sa_dev->port[i].port_num = i + s;
2150 spin_lock_init(&sa_dev->port[i].classport_lock);
2151 sa_dev->port[i].classport_info.valid = false;
2153 sa_dev->port[i].agent =
2154 ib_register_mad_agent(device, i + s, IB_QPT_GSI,
2155 NULL, 0, send_handler,
2156 recv_handler, sa_dev, 0);
2157 if (IS_ERR(sa_dev->port[i].agent)) {
2158 ret = PTR_ERR(sa_dev->port[i].agent);
2162 INIT_WORK(&sa_dev->port[i].update_task, update_sm_ah);
2163 INIT_DELAYED_WORK(&sa_dev->port[i].ib_cpi_work,
2174 ib_set_client_data(device, &sa_client, sa_dev);
2177 * We register our event handler after everything is set up,
2178 * and then update our cached info after the event handler is
2179 * registered to avoid any problems if a port changes state
2180 * during our initialization.
2183 INIT_IB_EVENT_HANDLER(&sa_dev->event_handler, device, ib_sa_event);
2184 ib_register_event_handler(&sa_dev->event_handler);
2186 for (i = 0; i <= e - s; ++i) {
2187 if (rdma_cap_ib_sa(device, i + 1))
2188 update_sm_ah(&sa_dev->port[i].update_task);
2195 if (rdma_cap_ib_sa(device, i + 1))
2196 ib_unregister_mad_agent(sa_dev->port[i].agent);
2203 static void ib_sa_remove_one(struct ib_device *device, void *client_data)
2205 struct ib_sa_device *sa_dev = client_data;
2208 ib_unregister_event_handler(&sa_dev->event_handler);
2209 flush_workqueue(ib_wq);
2211 for (i = 0; i <= sa_dev->end_port - sa_dev->start_port; ++i) {
2212 if (rdma_cap_ib_sa(device, i + 1)) {
2213 cancel_delayed_work_sync(&sa_dev->port[i].ib_cpi_work);
2214 ib_unregister_mad_agent(sa_dev->port[i].agent);
2215 if (sa_dev->port[i].sm_ah)
2216 kref_put(&sa_dev->port[i].sm_ah->ref, free_sm_ah);
2224 int ib_sa_init(void)
2228 get_random_bytes(&tid, sizeof tid);
2230 atomic_set(&ib_nl_sa_request_seq, 0);
2232 ret = ib_register_client(&sa_client);
2234 pr_err("Couldn't register ib_sa client\n");
2240 pr_err("Couldn't initialize multicast handling\n");
2244 ib_nl_wq = alloc_ordered_workqueue("ib_nl_sa_wq", WQ_MEM_RECLAIM);
2250 INIT_DELAYED_WORK(&ib_nl_timed_work, ib_nl_request_timeout);
2257 ib_unregister_client(&sa_client);
2262 void ib_sa_cleanup(void)
2264 cancel_delayed_work(&ib_nl_timed_work);
2265 flush_workqueue(ib_nl_wq);
2266 destroy_workqueue(ib_nl_wq);
2268 ib_unregister_client(&sa_client);
2269 WARN_ON(!xa_empty(&queries));