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/idr.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[0];
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_service_query {
127 void (*callback)(int, struct ib_sa_service_rec *, void *);
129 struct ib_sa_query sa_query;
132 struct ib_sa_path_query {
133 void (*callback)(int, struct sa_path_rec *, void *);
135 struct ib_sa_query sa_query;
136 struct sa_path_rec *conv_pr;
139 struct ib_sa_guidinfo_query {
140 void (*callback)(int, struct ib_sa_guidinfo_rec *, void *);
142 struct ib_sa_query sa_query;
145 struct ib_sa_classport_info_query {
146 void (*callback)(void *);
148 struct ib_sa_query sa_query;
151 struct ib_sa_mcmember_query {
152 void (*callback)(int, struct ib_sa_mcmember_rec *, void *);
154 struct ib_sa_query sa_query;
157 static LIST_HEAD(ib_nl_request_list);
158 static DEFINE_SPINLOCK(ib_nl_request_lock);
159 static atomic_t ib_nl_sa_request_seq;
160 static struct workqueue_struct *ib_nl_wq;
161 static struct delayed_work ib_nl_timed_work;
162 static const struct nla_policy ib_nl_policy[LS_NLA_TYPE_MAX] = {
163 [LS_NLA_TYPE_PATH_RECORD] = {.type = NLA_BINARY,
164 .len = sizeof(struct ib_path_rec_data)},
165 [LS_NLA_TYPE_TIMEOUT] = {.type = NLA_U32},
166 [LS_NLA_TYPE_SERVICE_ID] = {.type = NLA_U64},
167 [LS_NLA_TYPE_DGID] = {.type = NLA_BINARY,
168 .len = sizeof(struct rdma_nla_ls_gid)},
169 [LS_NLA_TYPE_SGID] = {.type = NLA_BINARY,
170 .len = sizeof(struct rdma_nla_ls_gid)},
171 [LS_NLA_TYPE_TCLASS] = {.type = NLA_U8},
172 [LS_NLA_TYPE_PKEY] = {.type = NLA_U16},
173 [LS_NLA_TYPE_QOS_CLASS] = {.type = NLA_U16},
177 static void ib_sa_add_one(struct ib_device *device);
178 static void ib_sa_remove_one(struct ib_device *device, void *client_data);
180 static struct ib_client sa_client = {
182 .add = ib_sa_add_one,
183 .remove = ib_sa_remove_one
186 static DEFINE_SPINLOCK(idr_lock);
187 static DEFINE_IDR(query_idr);
189 static DEFINE_SPINLOCK(tid_lock);
192 #define PATH_REC_FIELD(field) \
193 .struct_offset_bytes = offsetof(struct sa_path_rec, field), \
194 .struct_size_bytes = sizeof((struct sa_path_rec *)0)->field, \
195 .field_name = "sa_path_rec:" #field
197 static const struct ib_field path_rec_table[] = {
198 { PATH_REC_FIELD(service_id),
202 { PATH_REC_FIELD(dgid),
206 { PATH_REC_FIELD(sgid),
210 { PATH_REC_FIELD(ib.dlid),
214 { PATH_REC_FIELD(ib.slid),
218 { PATH_REC_FIELD(ib.raw_traffic),
226 { PATH_REC_FIELD(flow_label),
230 { PATH_REC_FIELD(hop_limit),
234 { PATH_REC_FIELD(traffic_class),
238 { PATH_REC_FIELD(reversible),
242 { PATH_REC_FIELD(numb_path),
246 { PATH_REC_FIELD(pkey),
250 { PATH_REC_FIELD(qos_class),
254 { PATH_REC_FIELD(sl),
258 { PATH_REC_FIELD(mtu_selector),
262 { PATH_REC_FIELD(mtu),
266 { PATH_REC_FIELD(rate_selector),
270 { PATH_REC_FIELD(rate),
274 { PATH_REC_FIELD(packet_life_time_selector),
278 { PATH_REC_FIELD(packet_life_time),
282 { PATH_REC_FIELD(preference),
292 #define OPA_PATH_REC_FIELD(field) \
293 .struct_offset_bytes = \
294 offsetof(struct sa_path_rec, field), \
295 .struct_size_bytes = \
296 sizeof((struct sa_path_rec *)0)->field, \
297 .field_name = "sa_path_rec:" #field
299 static const struct ib_field opa_path_rec_table[] = {
300 { OPA_PATH_REC_FIELD(service_id),
304 { OPA_PATH_REC_FIELD(dgid),
308 { OPA_PATH_REC_FIELD(sgid),
312 { OPA_PATH_REC_FIELD(opa.dlid),
316 { OPA_PATH_REC_FIELD(opa.slid),
320 { OPA_PATH_REC_FIELD(opa.raw_traffic),
328 { OPA_PATH_REC_FIELD(flow_label),
332 { OPA_PATH_REC_FIELD(hop_limit),
336 { OPA_PATH_REC_FIELD(traffic_class),
340 { OPA_PATH_REC_FIELD(reversible),
344 { OPA_PATH_REC_FIELD(numb_path),
348 { OPA_PATH_REC_FIELD(pkey),
352 { OPA_PATH_REC_FIELD(opa.l2_8B),
356 { OPA_PATH_REC_FIELD(opa.l2_10B),
360 { OPA_PATH_REC_FIELD(opa.l2_9B),
364 { OPA_PATH_REC_FIELD(opa.l2_16B),
372 { OPA_PATH_REC_FIELD(opa.qos_type),
376 { OPA_PATH_REC_FIELD(opa.qos_priority),
384 { OPA_PATH_REC_FIELD(sl),
392 { OPA_PATH_REC_FIELD(mtu_selector),
396 { OPA_PATH_REC_FIELD(mtu),
400 { OPA_PATH_REC_FIELD(rate_selector),
404 { OPA_PATH_REC_FIELD(rate),
408 { OPA_PATH_REC_FIELD(packet_life_time_selector),
412 { OPA_PATH_REC_FIELD(packet_life_time),
416 { OPA_PATH_REC_FIELD(preference),
422 #define MCMEMBER_REC_FIELD(field) \
423 .struct_offset_bytes = offsetof(struct ib_sa_mcmember_rec, field), \
424 .struct_size_bytes = sizeof ((struct ib_sa_mcmember_rec *) 0)->field, \
425 .field_name = "sa_mcmember_rec:" #field
427 static const struct ib_field mcmember_rec_table[] = {
428 { MCMEMBER_REC_FIELD(mgid),
432 { MCMEMBER_REC_FIELD(port_gid),
436 { MCMEMBER_REC_FIELD(qkey),
440 { MCMEMBER_REC_FIELD(mlid),
444 { MCMEMBER_REC_FIELD(mtu_selector),
448 { MCMEMBER_REC_FIELD(mtu),
452 { MCMEMBER_REC_FIELD(traffic_class),
456 { MCMEMBER_REC_FIELD(pkey),
460 { MCMEMBER_REC_FIELD(rate_selector),
464 { MCMEMBER_REC_FIELD(rate),
468 { MCMEMBER_REC_FIELD(packet_life_time_selector),
472 { MCMEMBER_REC_FIELD(packet_life_time),
476 { MCMEMBER_REC_FIELD(sl),
480 { MCMEMBER_REC_FIELD(flow_label),
484 { MCMEMBER_REC_FIELD(hop_limit),
488 { MCMEMBER_REC_FIELD(scope),
492 { MCMEMBER_REC_FIELD(join_state),
496 { MCMEMBER_REC_FIELD(proxy_join),
506 #define SERVICE_REC_FIELD(field) \
507 .struct_offset_bytes = offsetof(struct ib_sa_service_rec, field), \
508 .struct_size_bytes = sizeof ((struct ib_sa_service_rec *) 0)->field, \
509 .field_name = "sa_service_rec:" #field
511 static const struct ib_field service_rec_table[] = {
512 { SERVICE_REC_FIELD(id),
516 { SERVICE_REC_FIELD(gid),
520 { SERVICE_REC_FIELD(pkey),
524 { SERVICE_REC_FIELD(lease),
528 { SERVICE_REC_FIELD(key),
532 { SERVICE_REC_FIELD(name),
536 { SERVICE_REC_FIELD(data8),
540 { SERVICE_REC_FIELD(data16),
544 { SERVICE_REC_FIELD(data32),
548 { SERVICE_REC_FIELD(data64),
554 #define CLASSPORTINFO_REC_FIELD(field) \
555 .struct_offset_bytes = offsetof(struct ib_class_port_info, field), \
556 .struct_size_bytes = sizeof((struct ib_class_port_info *)0)->field, \
557 .field_name = "ib_class_port_info:" #field
559 static const struct ib_field ib_classport_info_rec_table[] = {
560 { CLASSPORTINFO_REC_FIELD(base_version),
564 { CLASSPORTINFO_REC_FIELD(class_version),
568 { CLASSPORTINFO_REC_FIELD(capability_mask),
572 { CLASSPORTINFO_REC_FIELD(cap_mask2_resp_time),
576 { CLASSPORTINFO_REC_FIELD(redirect_gid),
580 { CLASSPORTINFO_REC_FIELD(redirect_tcslfl),
584 { CLASSPORTINFO_REC_FIELD(redirect_lid),
588 { CLASSPORTINFO_REC_FIELD(redirect_pkey),
593 { CLASSPORTINFO_REC_FIELD(redirect_qp),
597 { CLASSPORTINFO_REC_FIELD(redirect_qkey),
602 { CLASSPORTINFO_REC_FIELD(trap_gid),
606 { CLASSPORTINFO_REC_FIELD(trap_tcslfl),
611 { CLASSPORTINFO_REC_FIELD(trap_lid),
615 { CLASSPORTINFO_REC_FIELD(trap_pkey),
620 { CLASSPORTINFO_REC_FIELD(trap_hlqp),
624 { CLASSPORTINFO_REC_FIELD(trap_qkey),
630 #define OPA_CLASSPORTINFO_REC_FIELD(field) \
631 .struct_offset_bytes =\
632 offsetof(struct opa_class_port_info, field), \
633 .struct_size_bytes = \
634 sizeof((struct opa_class_port_info *)0)->field, \
635 .field_name = "opa_class_port_info:" #field
637 static const struct ib_field opa_classport_info_rec_table[] = {
638 { OPA_CLASSPORTINFO_REC_FIELD(base_version),
642 { OPA_CLASSPORTINFO_REC_FIELD(class_version),
646 { OPA_CLASSPORTINFO_REC_FIELD(cap_mask),
650 { OPA_CLASSPORTINFO_REC_FIELD(cap_mask2_resp_time),
654 { OPA_CLASSPORTINFO_REC_FIELD(redirect_gid),
658 { OPA_CLASSPORTINFO_REC_FIELD(redirect_tc_fl),
662 { OPA_CLASSPORTINFO_REC_FIELD(redirect_lid),
666 { OPA_CLASSPORTINFO_REC_FIELD(redirect_sl_qp),
670 { OPA_CLASSPORTINFO_REC_FIELD(redirect_qkey),
674 { OPA_CLASSPORTINFO_REC_FIELD(trap_gid),
678 { OPA_CLASSPORTINFO_REC_FIELD(trap_tc_fl),
682 { OPA_CLASSPORTINFO_REC_FIELD(trap_lid),
686 { OPA_CLASSPORTINFO_REC_FIELD(trap_hl_qp),
690 { OPA_CLASSPORTINFO_REC_FIELD(trap_qkey),
694 { OPA_CLASSPORTINFO_REC_FIELD(trap_pkey),
698 { OPA_CLASSPORTINFO_REC_FIELD(redirect_pkey),
702 { OPA_CLASSPORTINFO_REC_FIELD(trap_sl_rsvd),
712 #define GUIDINFO_REC_FIELD(field) \
713 .struct_offset_bytes = offsetof(struct ib_sa_guidinfo_rec, field), \
714 .struct_size_bytes = sizeof((struct ib_sa_guidinfo_rec *) 0)->field, \
715 .field_name = "sa_guidinfo_rec:" #field
717 static const struct ib_field guidinfo_rec_table[] = {
718 { GUIDINFO_REC_FIELD(lid),
722 { GUIDINFO_REC_FIELD(block_num),
726 { GUIDINFO_REC_FIELD(res1),
730 { GUIDINFO_REC_FIELD(res2),
734 { GUIDINFO_REC_FIELD(guid_info_list),
740 static inline void ib_sa_disable_local_svc(struct ib_sa_query *query)
742 query->flags &= ~IB_SA_ENABLE_LOCAL_SERVICE;
745 static inline int ib_sa_query_cancelled(struct ib_sa_query *query)
747 return (query->flags & IB_SA_CANCEL);
750 static void ib_nl_set_path_rec_attrs(struct sk_buff *skb,
751 struct ib_sa_query *query)
753 struct sa_path_rec *sa_rec = query->mad_buf->context[1];
754 struct ib_sa_mad *mad = query->mad_buf->mad;
755 ib_sa_comp_mask comp_mask = mad->sa_hdr.comp_mask;
758 struct rdma_ls_resolve_header *header;
760 query->mad_buf->context[1] = NULL;
762 /* Construct the family header first */
763 header = skb_put(skb, NLMSG_ALIGN(sizeof(*header)));
764 memcpy(header->device_name, query->port->agent->device->name,
766 header->port_num = query->port->port_num;
768 if ((comp_mask & IB_SA_PATH_REC_REVERSIBLE) &&
769 sa_rec->reversible != 0)
770 query->path_use = LS_RESOLVE_PATH_USE_GMP;
772 query->path_use = LS_RESOLVE_PATH_USE_UNIDIRECTIONAL;
773 header->path_use = query->path_use;
775 /* Now build the attributes */
776 if (comp_mask & IB_SA_PATH_REC_SERVICE_ID) {
777 val64 = be64_to_cpu(sa_rec->service_id);
778 nla_put(skb, RDMA_NLA_F_MANDATORY | LS_NLA_TYPE_SERVICE_ID,
779 sizeof(val64), &val64);
781 if (comp_mask & IB_SA_PATH_REC_DGID)
782 nla_put(skb, RDMA_NLA_F_MANDATORY | LS_NLA_TYPE_DGID,
783 sizeof(sa_rec->dgid), &sa_rec->dgid);
784 if (comp_mask & IB_SA_PATH_REC_SGID)
785 nla_put(skb, RDMA_NLA_F_MANDATORY | LS_NLA_TYPE_SGID,
786 sizeof(sa_rec->sgid), &sa_rec->sgid);
787 if (comp_mask & IB_SA_PATH_REC_TRAFFIC_CLASS)
788 nla_put(skb, RDMA_NLA_F_MANDATORY | LS_NLA_TYPE_TCLASS,
789 sizeof(sa_rec->traffic_class), &sa_rec->traffic_class);
791 if (comp_mask & IB_SA_PATH_REC_PKEY) {
792 val16 = be16_to_cpu(sa_rec->pkey);
793 nla_put(skb, RDMA_NLA_F_MANDATORY | LS_NLA_TYPE_PKEY,
794 sizeof(val16), &val16);
796 if (comp_mask & IB_SA_PATH_REC_QOS_CLASS) {
797 val16 = be16_to_cpu(sa_rec->qos_class);
798 nla_put(skb, RDMA_NLA_F_MANDATORY | LS_NLA_TYPE_QOS_CLASS,
799 sizeof(val16), &val16);
803 static int ib_nl_get_path_rec_attrs_len(ib_sa_comp_mask comp_mask)
807 if (comp_mask & IB_SA_PATH_REC_SERVICE_ID)
808 len += nla_total_size(sizeof(u64));
809 if (comp_mask & IB_SA_PATH_REC_DGID)
810 len += nla_total_size(sizeof(struct rdma_nla_ls_gid));
811 if (comp_mask & IB_SA_PATH_REC_SGID)
812 len += nla_total_size(sizeof(struct rdma_nla_ls_gid));
813 if (comp_mask & IB_SA_PATH_REC_TRAFFIC_CLASS)
814 len += nla_total_size(sizeof(u8));
815 if (comp_mask & IB_SA_PATH_REC_PKEY)
816 len += nla_total_size(sizeof(u16));
817 if (comp_mask & IB_SA_PATH_REC_QOS_CLASS)
818 len += nla_total_size(sizeof(u16));
821 * Make sure that at least some of the required comp_mask bits are
824 if (WARN_ON(len == 0))
827 /* Add the family header */
828 len += NLMSG_ALIGN(sizeof(struct rdma_ls_resolve_header));
833 static int ib_nl_send_msg(struct ib_sa_query *query, gfp_t gfp_mask)
835 struct sk_buff *skb = NULL;
836 struct nlmsghdr *nlh;
839 struct ib_sa_mad *mad;
842 mad = query->mad_buf->mad;
843 len = ib_nl_get_path_rec_attrs_len(mad->sa_hdr.comp_mask);
847 skb = nlmsg_new(len, gfp_mask);
851 /* Put nlmsg header only for now */
852 data = ibnl_put_msg(skb, &nlh, query->seq, 0, RDMA_NL_LS,
853 RDMA_NL_LS_OP_RESOLVE, NLM_F_REQUEST);
860 ib_nl_set_path_rec_attrs(skb, query);
862 /* Repair the nlmsg header length */
865 ret = rdma_nl_multicast(skb, RDMA_NL_GROUP_LS, gfp_mask);
874 static int ib_nl_make_request(struct ib_sa_query *query, gfp_t gfp_mask)
880 INIT_LIST_HEAD(&query->list);
881 query->seq = (u32)atomic_inc_return(&ib_nl_sa_request_seq);
883 /* Put the request on the list first.*/
884 spin_lock_irqsave(&ib_nl_request_lock, flags);
885 delay = msecs_to_jiffies(sa_local_svc_timeout_ms);
886 query->timeout = delay + jiffies;
887 list_add_tail(&query->list, &ib_nl_request_list);
888 /* Start the timeout if this is the only request */
889 if (ib_nl_request_list.next == &query->list)
890 queue_delayed_work(ib_nl_wq, &ib_nl_timed_work, delay);
891 spin_unlock_irqrestore(&ib_nl_request_lock, flags);
893 ret = ib_nl_send_msg(query, gfp_mask);
896 /* Remove the request */
897 spin_lock_irqsave(&ib_nl_request_lock, flags);
898 list_del(&query->list);
899 spin_unlock_irqrestore(&ib_nl_request_lock, flags);
907 static int ib_nl_cancel_request(struct ib_sa_query *query)
910 struct ib_sa_query *wait_query;
913 spin_lock_irqsave(&ib_nl_request_lock, flags);
914 list_for_each_entry(wait_query, &ib_nl_request_list, list) {
915 /* Let the timeout to take care of the callback */
916 if (query == wait_query) {
917 query->flags |= IB_SA_CANCEL;
918 query->timeout = jiffies;
919 list_move(&query->list, &ib_nl_request_list);
921 mod_delayed_work(ib_nl_wq, &ib_nl_timed_work, 1);
925 spin_unlock_irqrestore(&ib_nl_request_lock, flags);
930 static void send_handler(struct ib_mad_agent *agent,
931 struct ib_mad_send_wc *mad_send_wc);
933 static void ib_nl_process_good_resolve_rsp(struct ib_sa_query *query,
934 const struct nlmsghdr *nlh)
936 struct ib_mad_send_wc mad_send_wc;
937 struct ib_sa_mad *mad = NULL;
938 const struct nlattr *head, *curr;
939 struct ib_path_rec_data *rec;
944 if (query->callback) {
945 head = (const struct nlattr *) nlmsg_data(nlh);
946 len = nlmsg_len(nlh);
947 switch (query->path_use) {
948 case LS_RESOLVE_PATH_USE_UNIDIRECTIONAL:
949 mask = IB_PATH_PRIMARY | IB_PATH_OUTBOUND;
952 case LS_RESOLVE_PATH_USE_ALL:
953 case LS_RESOLVE_PATH_USE_GMP:
955 mask = IB_PATH_PRIMARY | IB_PATH_GMP |
956 IB_PATH_BIDIRECTIONAL;
959 nla_for_each_attr(curr, head, len, rem) {
960 if (curr->nla_type == LS_NLA_TYPE_PATH_RECORD) {
961 rec = nla_data(curr);
963 * Get the first one. In the future, we may
964 * need to get up to 6 pathrecords.
966 if ((rec->flags & mask) == mask) {
967 mad = query->mad_buf->mad;
968 mad->mad_hdr.method |=
970 memcpy(mad->data, rec->path_rec,
971 sizeof(rec->path_rec));
977 query->callback(query, status, mad);
980 mad_send_wc.send_buf = query->mad_buf;
981 mad_send_wc.status = IB_WC_SUCCESS;
982 send_handler(query->mad_buf->mad_agent, &mad_send_wc);
985 static void ib_nl_request_timeout(struct work_struct *work)
988 struct ib_sa_query *query;
990 struct ib_mad_send_wc mad_send_wc;
993 spin_lock_irqsave(&ib_nl_request_lock, flags);
994 while (!list_empty(&ib_nl_request_list)) {
995 query = list_entry(ib_nl_request_list.next,
996 struct ib_sa_query, list);
998 if (time_after(query->timeout, jiffies)) {
999 delay = query->timeout - jiffies;
1000 if ((long)delay <= 0)
1002 queue_delayed_work(ib_nl_wq, &ib_nl_timed_work, delay);
1006 list_del(&query->list);
1007 ib_sa_disable_local_svc(query);
1008 /* Hold the lock to protect against query cancellation */
1009 if (ib_sa_query_cancelled(query))
1012 ret = ib_post_send_mad(query->mad_buf, NULL);
1014 mad_send_wc.send_buf = query->mad_buf;
1015 mad_send_wc.status = IB_WC_WR_FLUSH_ERR;
1016 spin_unlock_irqrestore(&ib_nl_request_lock, flags);
1017 send_handler(query->port->agent, &mad_send_wc);
1018 spin_lock_irqsave(&ib_nl_request_lock, flags);
1021 spin_unlock_irqrestore(&ib_nl_request_lock, flags);
1024 int ib_nl_handle_set_timeout(struct sk_buff *skb,
1025 struct nlmsghdr *nlh,
1026 struct netlink_ext_ack *extack)
1028 int timeout, delta, abs_delta;
1029 const struct nlattr *attr;
1030 unsigned long flags;
1031 struct ib_sa_query *query;
1033 struct nlattr *tb[LS_NLA_TYPE_MAX];
1036 if (!(nlh->nlmsg_flags & NLM_F_REQUEST) ||
1037 !(NETLINK_CB(skb).sk))
1040 ret = nla_parse(tb, LS_NLA_TYPE_MAX - 1, nlmsg_data(nlh),
1041 nlmsg_len(nlh), ib_nl_policy, NULL);
1042 attr = (const struct nlattr *)tb[LS_NLA_TYPE_TIMEOUT];
1044 goto settimeout_out;
1046 timeout = *(int *) nla_data(attr);
1047 if (timeout < IB_SA_LOCAL_SVC_TIMEOUT_MIN)
1048 timeout = IB_SA_LOCAL_SVC_TIMEOUT_MIN;
1049 if (timeout > IB_SA_LOCAL_SVC_TIMEOUT_MAX)
1050 timeout = IB_SA_LOCAL_SVC_TIMEOUT_MAX;
1052 delta = timeout - sa_local_svc_timeout_ms;
1059 spin_lock_irqsave(&ib_nl_request_lock, flags);
1060 sa_local_svc_timeout_ms = timeout;
1061 list_for_each_entry(query, &ib_nl_request_list, list) {
1062 if (delta < 0 && abs_delta > query->timeout)
1065 query->timeout += delta;
1067 /* Get the new delay from the first entry */
1069 delay = query->timeout - jiffies;
1075 mod_delayed_work(ib_nl_wq, &ib_nl_timed_work,
1076 (unsigned long)delay);
1077 spin_unlock_irqrestore(&ib_nl_request_lock, flags);
1084 static inline int ib_nl_is_good_resolve_resp(const struct nlmsghdr *nlh)
1086 struct nlattr *tb[LS_NLA_TYPE_MAX];
1089 if (nlh->nlmsg_flags & RDMA_NL_LS_F_ERR)
1092 ret = nla_parse(tb, LS_NLA_TYPE_MAX - 1, nlmsg_data(nlh),
1093 nlmsg_len(nlh), ib_nl_policy, NULL);
1100 int ib_nl_handle_resolve_resp(struct sk_buff *skb,
1101 struct nlmsghdr *nlh,
1102 struct netlink_ext_ack *extack)
1104 unsigned long flags;
1105 struct ib_sa_query *query;
1106 struct ib_mad_send_buf *send_buf;
1107 struct ib_mad_send_wc mad_send_wc;
1111 if ((nlh->nlmsg_flags & NLM_F_REQUEST) ||
1112 !(NETLINK_CB(skb).sk))
1115 spin_lock_irqsave(&ib_nl_request_lock, flags);
1116 list_for_each_entry(query, &ib_nl_request_list, list) {
1118 * If the query is cancelled, let the timeout routine
1121 if (nlh->nlmsg_seq == query->seq) {
1122 found = !ib_sa_query_cancelled(query);
1124 list_del(&query->list);
1130 spin_unlock_irqrestore(&ib_nl_request_lock, flags);
1134 send_buf = query->mad_buf;
1136 if (!ib_nl_is_good_resolve_resp(nlh)) {
1137 /* if the result is a failure, send out the packet via IB */
1138 ib_sa_disable_local_svc(query);
1139 ret = ib_post_send_mad(query->mad_buf, NULL);
1140 spin_unlock_irqrestore(&ib_nl_request_lock, flags);
1142 mad_send_wc.send_buf = send_buf;
1143 mad_send_wc.status = IB_WC_GENERAL_ERR;
1144 send_handler(query->port->agent, &mad_send_wc);
1147 spin_unlock_irqrestore(&ib_nl_request_lock, flags);
1148 ib_nl_process_good_resolve_rsp(query, nlh);
1155 static void free_sm_ah(struct kref *kref)
1157 struct ib_sa_sm_ah *sm_ah = container_of(kref, struct ib_sa_sm_ah, ref);
1159 rdma_destroy_ah(sm_ah->ah);
1163 void ib_sa_register_client(struct ib_sa_client *client)
1165 atomic_set(&client->users, 1);
1166 init_completion(&client->comp);
1168 EXPORT_SYMBOL(ib_sa_register_client);
1170 void ib_sa_unregister_client(struct ib_sa_client *client)
1172 ib_sa_client_put(client);
1173 wait_for_completion(&client->comp);
1175 EXPORT_SYMBOL(ib_sa_unregister_client);
1178 * ib_sa_cancel_query - try to cancel an SA query
1179 * @id:ID of query to cancel
1180 * @query:query pointer to cancel
1182 * Try to cancel an SA query. If the id and query don't match up or
1183 * the query has already completed, nothing is done. Otherwise the
1184 * query is canceled and will complete with a status of -EINTR.
1186 void ib_sa_cancel_query(int id, struct ib_sa_query *query)
1188 unsigned long flags;
1189 struct ib_mad_agent *agent;
1190 struct ib_mad_send_buf *mad_buf;
1192 spin_lock_irqsave(&idr_lock, flags);
1193 if (idr_find(&query_idr, id) != query) {
1194 spin_unlock_irqrestore(&idr_lock, flags);
1197 agent = query->port->agent;
1198 mad_buf = query->mad_buf;
1199 spin_unlock_irqrestore(&idr_lock, flags);
1202 * If the query is still on the netlink request list, schedule
1203 * it to be cancelled by the timeout routine. Otherwise, it has been
1204 * sent to the MAD layer and has to be cancelled from there.
1206 if (!ib_nl_cancel_request(query))
1207 ib_cancel_mad(agent, mad_buf);
1209 EXPORT_SYMBOL(ib_sa_cancel_query);
1211 static u8 get_src_path_mask(struct ib_device *device, u8 port_num)
1213 struct ib_sa_device *sa_dev;
1214 struct ib_sa_port *port;
1215 unsigned long flags;
1218 sa_dev = ib_get_client_data(device, &sa_client);
1222 port = &sa_dev->port[port_num - sa_dev->start_port];
1223 spin_lock_irqsave(&port->ah_lock, flags);
1224 src_path_mask = port->sm_ah ? port->sm_ah->src_path_mask : 0x7f;
1225 spin_unlock_irqrestore(&port->ah_lock, flags);
1227 return src_path_mask;
1230 int ib_init_ah_from_path(struct ib_device *device, u8 port_num,
1231 struct sa_path_rec *rec,
1232 struct rdma_ah_attr *ah_attr)
1237 struct net_device *ndev = NULL;
1239 memset(ah_attr, 0, sizeof *ah_attr);
1240 ah_attr->type = rdma_ah_find_type(device, port_num);
1242 rdma_ah_set_dlid(ah_attr, be32_to_cpu(sa_path_get_dlid(rec)));
1244 if ((ah_attr->type == RDMA_AH_ATTR_TYPE_OPA) &&
1245 (rdma_ah_get_dlid(ah_attr) == be16_to_cpu(IB_LID_PERMISSIVE)))
1246 rdma_ah_set_make_grd(ah_attr, true);
1248 rdma_ah_set_sl(ah_attr, rec->sl);
1249 rdma_ah_set_path_bits(ah_attr, be32_to_cpu(sa_path_get_slid(rec)) &
1250 get_src_path_mask(device, port_num));
1251 rdma_ah_set_port_num(ah_attr, port_num);
1252 rdma_ah_set_static_rate(ah_attr, rec->rate);
1253 use_roce = rdma_cap_eth_ah(device, port_num);
1256 struct net_device *idev;
1257 struct net_device *resolved_dev;
1258 struct rdma_dev_addr dev_addr = {
1259 .bound_dev_if = ((sa_path_get_ifindex(rec) >= 0) ?
1260 sa_path_get_ifindex(rec) : 0),
1261 .net = sa_path_get_ndev(rec) ?
1262 sa_path_get_ndev(rec) :
1266 struct sockaddr_in _sockaddr_in;
1267 struct sockaddr_in6 _sockaddr_in6;
1268 } sgid_addr, dgid_addr;
1270 if (!device->get_netdev)
1273 rdma_gid2ip((struct sockaddr *)&sgid_addr, &rec->sgid);
1274 rdma_gid2ip((struct sockaddr *)&dgid_addr, &rec->dgid);
1276 /* validate the route */
1277 ret = rdma_resolve_ip_route((struct sockaddr *)&sgid_addr,
1278 (struct sockaddr *)&dgid_addr,
1283 if ((dev_addr.network == RDMA_NETWORK_IPV4 ||
1284 dev_addr.network == RDMA_NETWORK_IPV6) &&
1285 rec->rec_type != SA_PATH_REC_TYPE_ROCE_V2)
1288 idev = device->get_netdev(device, port_num);
1292 resolved_dev = dev_get_by_index(dev_addr.net,
1293 dev_addr.bound_dev_if);
1294 if (!resolved_dev) {
1298 ndev = ib_get_ndev_from_path(rec);
1300 if ((ndev && ndev != resolved_dev) ||
1301 (resolved_dev != idev &&
1302 !rdma_is_upper_dev_rcu(idev, resolved_dev)))
1303 ret = -EHOSTUNREACH;
1306 dev_put(resolved_dev);
1314 if (rec->hop_limit > 0 || use_roce) {
1315 enum ib_gid_type type = sa_conv_pathrec_to_gid_type(rec);
1317 ret = ib_find_cached_gid_by_port(device, &rec->sgid, type,
1318 port_num, ndev, &gid_index);
1325 rdma_ah_set_grh(ah_attr, &rec->dgid,
1326 be32_to_cpu(rec->flow_label),
1327 gid_index, rec->hop_limit,
1328 rec->traffic_class);
1334 u8 *dmac = sa_path_get_dmac(rec);
1338 memcpy(ah_attr->roce.dmac, dmac, ETH_ALEN);
1343 EXPORT_SYMBOL(ib_init_ah_from_path);
1345 static int alloc_mad(struct ib_sa_query *query, gfp_t gfp_mask)
1347 unsigned long flags;
1349 spin_lock_irqsave(&query->port->ah_lock, flags);
1350 if (!query->port->sm_ah) {
1351 spin_unlock_irqrestore(&query->port->ah_lock, flags);
1354 kref_get(&query->port->sm_ah->ref);
1355 query->sm_ah = query->port->sm_ah;
1356 spin_unlock_irqrestore(&query->port->ah_lock, flags);
1358 query->mad_buf = ib_create_send_mad(query->port->agent, 1,
1359 query->sm_ah->pkey_index,
1360 0, IB_MGMT_SA_HDR, IB_MGMT_SA_DATA,
1362 ((query->flags & IB_SA_QUERY_OPA) ?
1363 OPA_MGMT_BASE_VERSION :
1364 IB_MGMT_BASE_VERSION));
1365 if (IS_ERR(query->mad_buf)) {
1366 kref_put(&query->sm_ah->ref, free_sm_ah);
1370 query->mad_buf->ah = query->sm_ah->ah;
1375 static void free_mad(struct ib_sa_query *query)
1377 ib_free_send_mad(query->mad_buf);
1378 kref_put(&query->sm_ah->ref, free_sm_ah);
1381 static void init_mad(struct ib_sa_query *query, struct ib_mad_agent *agent)
1383 struct ib_sa_mad *mad = query->mad_buf->mad;
1384 unsigned long flags;
1386 memset(mad, 0, sizeof *mad);
1388 if (query->flags & IB_SA_QUERY_OPA) {
1389 mad->mad_hdr.base_version = OPA_MGMT_BASE_VERSION;
1390 mad->mad_hdr.class_version = OPA_SA_CLASS_VERSION;
1392 mad->mad_hdr.base_version = IB_MGMT_BASE_VERSION;
1393 mad->mad_hdr.class_version = IB_SA_CLASS_VERSION;
1395 mad->mad_hdr.mgmt_class = IB_MGMT_CLASS_SUBN_ADM;
1396 spin_lock_irqsave(&tid_lock, flags);
1398 cpu_to_be64(((u64) agent->hi_tid) << 32 | tid++);
1399 spin_unlock_irqrestore(&tid_lock, flags);
1402 static int send_mad(struct ib_sa_query *query, int timeout_ms, gfp_t gfp_mask)
1404 bool preload = gfpflags_allow_blocking(gfp_mask);
1405 unsigned long flags;
1409 idr_preload(gfp_mask);
1410 spin_lock_irqsave(&idr_lock, flags);
1412 id = idr_alloc(&query_idr, query, 0, 0, GFP_NOWAIT);
1414 spin_unlock_irqrestore(&idr_lock, flags);
1420 query->mad_buf->timeout_ms = timeout_ms;
1421 query->mad_buf->context[0] = query;
1424 if ((query->flags & IB_SA_ENABLE_LOCAL_SERVICE) &&
1425 (!(query->flags & IB_SA_QUERY_OPA))) {
1426 if (!rdma_nl_chk_listeners(RDMA_NL_GROUP_LS)) {
1427 if (!ib_nl_make_request(query, gfp_mask))
1430 ib_sa_disable_local_svc(query);
1433 ret = ib_post_send_mad(query->mad_buf, NULL);
1435 spin_lock_irqsave(&idr_lock, flags);
1436 idr_remove(&query_idr, id);
1437 spin_unlock_irqrestore(&idr_lock, flags);
1441 * It's not safe to dereference query any more, because the
1442 * send may already have completed and freed the query in
1445 return ret ? ret : id;
1448 void ib_sa_unpack_path(void *attribute, struct sa_path_rec *rec)
1450 ib_unpack(path_rec_table, ARRAY_SIZE(path_rec_table), attribute, rec);
1452 EXPORT_SYMBOL(ib_sa_unpack_path);
1454 void ib_sa_pack_path(struct sa_path_rec *rec, void *attribute)
1456 ib_pack(path_rec_table, ARRAY_SIZE(path_rec_table), rec, attribute);
1458 EXPORT_SYMBOL(ib_sa_pack_path);
1460 static bool ib_sa_opa_pathrecord_support(struct ib_sa_client *client,
1461 struct ib_device *device,
1464 struct ib_sa_device *sa_dev = ib_get_client_data(device, &sa_client);
1465 struct ib_sa_port *port;
1466 unsigned long flags;
1472 port = &sa_dev->port[port_num - sa_dev->start_port];
1473 spin_lock_irqsave(&port->classport_lock, flags);
1474 if (!port->classport_info.valid)
1477 if (port->classport_info.data.type == RDMA_CLASS_PORT_INFO_OPA)
1478 ret = opa_get_cpi_capmask2(&port->classport_info.data.opa) &
1479 OPA_CLASS_PORT_INFO_PR_SUPPORT;
1481 spin_unlock_irqrestore(&port->classport_lock, flags);
1485 enum opa_pr_supported {
1492 * Check if current PR query can be an OPA query.
1493 * Retuns PR_NOT_SUPPORTED if a path record query is not
1494 * possible, PR_OPA_SUPPORTED if an OPA path record query
1495 * is possible and PR_IB_SUPPORTED if an IB path record
1496 * query is possible.
1498 static int opa_pr_query_possible(struct ib_sa_client *client,
1499 struct ib_device *device,
1501 struct sa_path_rec *rec)
1503 struct ib_port_attr port_attr;
1505 if (ib_query_port(device, port_num, &port_attr))
1506 return PR_NOT_SUPPORTED;
1508 if (ib_sa_opa_pathrecord_support(client, device, port_num))
1509 return PR_OPA_SUPPORTED;
1511 if (port_attr.lid >= be16_to_cpu(IB_MULTICAST_LID_BASE))
1512 return PR_NOT_SUPPORTED;
1514 return PR_IB_SUPPORTED;
1517 static void ib_sa_path_rec_callback(struct ib_sa_query *sa_query,
1519 struct ib_sa_mad *mad)
1521 struct ib_sa_path_query *query =
1522 container_of(sa_query, struct ib_sa_path_query, sa_query);
1525 struct sa_path_rec rec;
1527 if (sa_query->flags & IB_SA_QUERY_OPA) {
1528 ib_unpack(opa_path_rec_table,
1529 ARRAY_SIZE(opa_path_rec_table),
1531 rec.rec_type = SA_PATH_REC_TYPE_OPA;
1532 query->callback(status, &rec, query->context);
1534 ib_unpack(path_rec_table,
1535 ARRAY_SIZE(path_rec_table),
1537 rec.rec_type = SA_PATH_REC_TYPE_IB;
1538 sa_path_set_ndev(&rec, NULL);
1539 sa_path_set_ifindex(&rec, 0);
1540 sa_path_set_dmac_zero(&rec);
1542 if (query->conv_pr) {
1543 struct sa_path_rec opa;
1545 memset(&opa, 0, sizeof(struct sa_path_rec));
1546 sa_convert_path_ib_to_opa(&opa, &rec);
1547 query->callback(status, &opa, query->context);
1549 query->callback(status, &rec, query->context);
1553 query->callback(status, NULL, query->context);
1556 static void ib_sa_path_rec_release(struct ib_sa_query *sa_query)
1558 struct ib_sa_path_query *query =
1559 container_of(sa_query, struct ib_sa_path_query, sa_query);
1561 kfree(query->conv_pr);
1566 * ib_sa_path_rec_get - Start a Path get query
1568 * @device:device to send query on
1569 * @port_num: port number to send query on
1570 * @rec:Path Record to send in query
1571 * @comp_mask:component mask to send in query
1572 * @timeout_ms:time to wait for response
1573 * @gfp_mask:GFP mask to use for internal allocations
1574 * @callback:function called when query completes, times out or is
1576 * @context:opaque user context passed to callback
1577 * @sa_query:query context, used to cancel query
1579 * Send a Path Record Get query to the SA to look up a path. The
1580 * callback function will be called when the query completes (or
1581 * fails); status is 0 for a successful response, -EINTR if the query
1582 * is canceled, -ETIMEDOUT is the query timed out, or -EIO if an error
1583 * occurred sending the query. The resp parameter of the callback is
1584 * only valid if status is 0.
1586 * If the return value of ib_sa_path_rec_get() is negative, it is an
1587 * error code. Otherwise it is a query ID that can be used to cancel
1590 int ib_sa_path_rec_get(struct ib_sa_client *client,
1591 struct ib_device *device, u8 port_num,
1592 struct sa_path_rec *rec,
1593 ib_sa_comp_mask comp_mask,
1594 int timeout_ms, gfp_t gfp_mask,
1595 void (*callback)(int status,
1596 struct sa_path_rec *resp,
1599 struct ib_sa_query **sa_query)
1601 struct ib_sa_path_query *query;
1602 struct ib_sa_device *sa_dev = ib_get_client_data(device, &sa_client);
1603 struct ib_sa_port *port;
1604 struct ib_mad_agent *agent;
1605 struct ib_sa_mad *mad;
1606 enum opa_pr_supported status;
1612 if ((rec->rec_type != SA_PATH_REC_TYPE_IB) &&
1613 (rec->rec_type != SA_PATH_REC_TYPE_OPA))
1616 port = &sa_dev->port[port_num - sa_dev->start_port];
1617 agent = port->agent;
1619 query = kzalloc(sizeof(*query), gfp_mask);
1623 query->sa_query.port = port;
1624 if (rec->rec_type == SA_PATH_REC_TYPE_OPA) {
1625 status = opa_pr_query_possible(client, device, port_num, rec);
1626 if (status == PR_NOT_SUPPORTED) {
1629 } else if (status == PR_OPA_SUPPORTED) {
1630 query->sa_query.flags |= IB_SA_QUERY_OPA;
1633 kmalloc(sizeof(*query->conv_pr), gfp_mask);
1634 if (!query->conv_pr) {
1641 ret = alloc_mad(&query->sa_query, gfp_mask);
1645 ib_sa_client_get(client);
1646 query->sa_query.client = client;
1647 query->callback = callback;
1648 query->context = context;
1650 mad = query->sa_query.mad_buf->mad;
1651 init_mad(&query->sa_query, agent);
1653 query->sa_query.callback = callback ? ib_sa_path_rec_callback : NULL;
1654 query->sa_query.release = ib_sa_path_rec_release;
1655 mad->mad_hdr.method = IB_MGMT_METHOD_GET;
1656 mad->mad_hdr.attr_id = cpu_to_be16(IB_SA_ATTR_PATH_REC);
1657 mad->sa_hdr.comp_mask = comp_mask;
1659 if (query->sa_query.flags & IB_SA_QUERY_OPA) {
1660 ib_pack(opa_path_rec_table, ARRAY_SIZE(opa_path_rec_table),
1662 } else if (query->conv_pr) {
1663 sa_convert_path_opa_to_ib(query->conv_pr, rec);
1664 ib_pack(path_rec_table, ARRAY_SIZE(path_rec_table),
1665 query->conv_pr, mad->data);
1667 ib_pack(path_rec_table, ARRAY_SIZE(path_rec_table),
1671 *sa_query = &query->sa_query;
1673 query->sa_query.flags |= IB_SA_ENABLE_LOCAL_SERVICE;
1674 query->sa_query.mad_buf->context[1] = (query->conv_pr) ?
1675 query->conv_pr : rec;
1677 ret = send_mad(&query->sa_query, timeout_ms, gfp_mask);
1685 ib_sa_client_put(query->sa_query.client);
1686 free_mad(&query->sa_query);
1688 kfree(query->conv_pr);
1693 EXPORT_SYMBOL(ib_sa_path_rec_get);
1695 static void ib_sa_service_rec_callback(struct ib_sa_query *sa_query,
1697 struct ib_sa_mad *mad)
1699 struct ib_sa_service_query *query =
1700 container_of(sa_query, struct ib_sa_service_query, sa_query);
1703 struct ib_sa_service_rec rec;
1705 ib_unpack(service_rec_table, ARRAY_SIZE(service_rec_table),
1707 query->callback(status, &rec, query->context);
1709 query->callback(status, NULL, query->context);
1712 static void ib_sa_service_rec_release(struct ib_sa_query *sa_query)
1714 kfree(container_of(sa_query, struct ib_sa_service_query, sa_query));
1718 * ib_sa_service_rec_query - Start Service Record operation
1720 * @device:device to send request on
1721 * @port_num: port number to send request on
1722 * @method:SA method - should be get, set, or delete
1723 * @rec:Service Record to send in request
1724 * @comp_mask:component mask to send in request
1725 * @timeout_ms:time to wait for response
1726 * @gfp_mask:GFP mask to use for internal allocations
1727 * @callback:function called when request completes, times out or is
1729 * @context:opaque user context passed to callback
1730 * @sa_query:request context, used to cancel request
1732 * Send a Service Record set/get/delete to the SA to register,
1733 * unregister or query a service record.
1734 * The callback function will be called when the request completes (or
1735 * fails); status is 0 for a successful response, -EINTR if the query
1736 * is canceled, -ETIMEDOUT is the query timed out, or -EIO if an error
1737 * occurred sending the query. The resp parameter of the callback is
1738 * only valid if status is 0.
1740 * If the return value of ib_sa_service_rec_query() is negative, it is an
1741 * error code. Otherwise it is a request ID that can be used to cancel
1744 int ib_sa_service_rec_query(struct ib_sa_client *client,
1745 struct ib_device *device, u8 port_num, u8 method,
1746 struct ib_sa_service_rec *rec,
1747 ib_sa_comp_mask comp_mask,
1748 int timeout_ms, gfp_t gfp_mask,
1749 void (*callback)(int status,
1750 struct ib_sa_service_rec *resp,
1753 struct ib_sa_query **sa_query)
1755 struct ib_sa_service_query *query;
1756 struct ib_sa_device *sa_dev = ib_get_client_data(device, &sa_client);
1757 struct ib_sa_port *port;
1758 struct ib_mad_agent *agent;
1759 struct ib_sa_mad *mad;
1765 port = &sa_dev->port[port_num - sa_dev->start_port];
1766 agent = port->agent;
1768 if (method != IB_MGMT_METHOD_GET &&
1769 method != IB_MGMT_METHOD_SET &&
1770 method != IB_SA_METHOD_DELETE)
1773 query = kzalloc(sizeof(*query), gfp_mask);
1777 query->sa_query.port = port;
1778 ret = alloc_mad(&query->sa_query, gfp_mask);
1782 ib_sa_client_get(client);
1783 query->sa_query.client = client;
1784 query->callback = callback;
1785 query->context = context;
1787 mad = query->sa_query.mad_buf->mad;
1788 init_mad(&query->sa_query, agent);
1790 query->sa_query.callback = callback ? ib_sa_service_rec_callback : NULL;
1791 query->sa_query.release = ib_sa_service_rec_release;
1792 mad->mad_hdr.method = method;
1793 mad->mad_hdr.attr_id = cpu_to_be16(IB_SA_ATTR_SERVICE_REC);
1794 mad->sa_hdr.comp_mask = comp_mask;
1796 ib_pack(service_rec_table, ARRAY_SIZE(service_rec_table),
1799 *sa_query = &query->sa_query;
1801 ret = send_mad(&query->sa_query, timeout_ms, gfp_mask);
1809 ib_sa_client_put(query->sa_query.client);
1810 free_mad(&query->sa_query);
1816 EXPORT_SYMBOL(ib_sa_service_rec_query);
1818 static void ib_sa_mcmember_rec_callback(struct ib_sa_query *sa_query,
1820 struct ib_sa_mad *mad)
1822 struct ib_sa_mcmember_query *query =
1823 container_of(sa_query, struct ib_sa_mcmember_query, sa_query);
1826 struct ib_sa_mcmember_rec rec;
1828 ib_unpack(mcmember_rec_table, ARRAY_SIZE(mcmember_rec_table),
1830 query->callback(status, &rec, query->context);
1832 query->callback(status, NULL, query->context);
1835 static void ib_sa_mcmember_rec_release(struct ib_sa_query *sa_query)
1837 kfree(container_of(sa_query, struct ib_sa_mcmember_query, sa_query));
1840 int ib_sa_mcmember_rec_query(struct ib_sa_client *client,
1841 struct ib_device *device, u8 port_num,
1843 struct ib_sa_mcmember_rec *rec,
1844 ib_sa_comp_mask comp_mask,
1845 int timeout_ms, gfp_t gfp_mask,
1846 void (*callback)(int status,
1847 struct ib_sa_mcmember_rec *resp,
1850 struct ib_sa_query **sa_query)
1852 struct ib_sa_mcmember_query *query;
1853 struct ib_sa_device *sa_dev = ib_get_client_data(device, &sa_client);
1854 struct ib_sa_port *port;
1855 struct ib_mad_agent *agent;
1856 struct ib_sa_mad *mad;
1862 port = &sa_dev->port[port_num - sa_dev->start_port];
1863 agent = port->agent;
1865 query = kzalloc(sizeof(*query), gfp_mask);
1869 query->sa_query.port = port;
1870 ret = alloc_mad(&query->sa_query, gfp_mask);
1874 ib_sa_client_get(client);
1875 query->sa_query.client = client;
1876 query->callback = callback;
1877 query->context = context;
1879 mad = query->sa_query.mad_buf->mad;
1880 init_mad(&query->sa_query, agent);
1882 query->sa_query.callback = callback ? ib_sa_mcmember_rec_callback : NULL;
1883 query->sa_query.release = ib_sa_mcmember_rec_release;
1884 mad->mad_hdr.method = method;
1885 mad->mad_hdr.attr_id = cpu_to_be16(IB_SA_ATTR_MC_MEMBER_REC);
1886 mad->sa_hdr.comp_mask = comp_mask;
1888 ib_pack(mcmember_rec_table, ARRAY_SIZE(mcmember_rec_table),
1891 *sa_query = &query->sa_query;
1893 ret = send_mad(&query->sa_query, timeout_ms, gfp_mask);
1901 ib_sa_client_put(query->sa_query.client);
1902 free_mad(&query->sa_query);
1909 /* Support GuidInfoRecord */
1910 static void ib_sa_guidinfo_rec_callback(struct ib_sa_query *sa_query,
1912 struct ib_sa_mad *mad)
1914 struct ib_sa_guidinfo_query *query =
1915 container_of(sa_query, struct ib_sa_guidinfo_query, sa_query);
1918 struct ib_sa_guidinfo_rec rec;
1920 ib_unpack(guidinfo_rec_table, ARRAY_SIZE(guidinfo_rec_table),
1922 query->callback(status, &rec, query->context);
1924 query->callback(status, NULL, query->context);
1927 static void ib_sa_guidinfo_rec_release(struct ib_sa_query *sa_query)
1929 kfree(container_of(sa_query, struct ib_sa_guidinfo_query, sa_query));
1932 int ib_sa_guid_info_rec_query(struct ib_sa_client *client,
1933 struct ib_device *device, u8 port_num,
1934 struct ib_sa_guidinfo_rec *rec,
1935 ib_sa_comp_mask comp_mask, u8 method,
1936 int timeout_ms, gfp_t gfp_mask,
1937 void (*callback)(int status,
1938 struct ib_sa_guidinfo_rec *resp,
1941 struct ib_sa_query **sa_query)
1943 struct ib_sa_guidinfo_query *query;
1944 struct ib_sa_device *sa_dev = ib_get_client_data(device, &sa_client);
1945 struct ib_sa_port *port;
1946 struct ib_mad_agent *agent;
1947 struct ib_sa_mad *mad;
1953 if (method != IB_MGMT_METHOD_GET &&
1954 method != IB_MGMT_METHOD_SET &&
1955 method != IB_SA_METHOD_DELETE) {
1959 port = &sa_dev->port[port_num - sa_dev->start_port];
1960 agent = port->agent;
1962 query = kzalloc(sizeof(*query), gfp_mask);
1966 query->sa_query.port = port;
1967 ret = alloc_mad(&query->sa_query, gfp_mask);
1971 ib_sa_client_get(client);
1972 query->sa_query.client = client;
1973 query->callback = callback;
1974 query->context = context;
1976 mad = query->sa_query.mad_buf->mad;
1977 init_mad(&query->sa_query, agent);
1979 query->sa_query.callback = callback ? ib_sa_guidinfo_rec_callback : NULL;
1980 query->sa_query.release = ib_sa_guidinfo_rec_release;
1982 mad->mad_hdr.method = method;
1983 mad->mad_hdr.attr_id = cpu_to_be16(IB_SA_ATTR_GUID_INFO_REC);
1984 mad->sa_hdr.comp_mask = comp_mask;
1986 ib_pack(guidinfo_rec_table, ARRAY_SIZE(guidinfo_rec_table), rec,
1989 *sa_query = &query->sa_query;
1991 ret = send_mad(&query->sa_query, timeout_ms, gfp_mask);
1999 ib_sa_client_put(query->sa_query.client);
2000 free_mad(&query->sa_query);
2006 EXPORT_SYMBOL(ib_sa_guid_info_rec_query);
2008 bool ib_sa_sendonly_fullmem_support(struct ib_sa_client *client,
2009 struct ib_device *device,
2012 struct ib_sa_device *sa_dev = ib_get_client_data(device, &sa_client);
2013 struct ib_sa_port *port;
2015 unsigned long flags;
2020 port = &sa_dev->port[port_num - sa_dev->start_port];
2022 spin_lock_irqsave(&port->classport_lock, flags);
2023 if ((port->classport_info.valid) &&
2024 (port->classport_info.data.type == RDMA_CLASS_PORT_INFO_IB))
2025 ret = ib_get_cpi_capmask2(&port->classport_info.data.ib)
2026 & IB_SA_CAP_MASK2_SENDONLY_FULL_MEM_SUPPORT;
2027 spin_unlock_irqrestore(&port->classport_lock, flags);
2030 EXPORT_SYMBOL(ib_sa_sendonly_fullmem_support);
2032 struct ib_classport_info_context {
2033 struct completion done;
2034 struct ib_sa_query *sa_query;
2037 static void ib_classportinfo_cb(void *context)
2039 struct ib_classport_info_context *cb_ctx = context;
2041 complete(&cb_ctx->done);
2044 static void ib_sa_classport_info_rec_callback(struct ib_sa_query *sa_query,
2046 struct ib_sa_mad *mad)
2048 unsigned long flags;
2049 struct ib_sa_classport_info_query *query =
2050 container_of(sa_query, struct ib_sa_classport_info_query, sa_query);
2051 struct ib_sa_classport_cache *info = &sa_query->port->classport_info;
2054 if (sa_query->flags & IB_SA_QUERY_OPA) {
2055 struct opa_class_port_info rec;
2057 ib_unpack(opa_classport_info_rec_table,
2058 ARRAY_SIZE(opa_classport_info_rec_table),
2061 spin_lock_irqsave(&sa_query->port->classport_lock,
2063 if (!status && !info->valid) {
2064 memcpy(&info->data.opa, &rec,
2065 sizeof(info->data.opa));
2068 info->data.type = RDMA_CLASS_PORT_INFO_OPA;
2070 spin_unlock_irqrestore(&sa_query->port->classport_lock,
2074 struct ib_class_port_info rec;
2076 ib_unpack(ib_classport_info_rec_table,
2077 ARRAY_SIZE(ib_classport_info_rec_table),
2080 spin_lock_irqsave(&sa_query->port->classport_lock,
2082 if (!status && !info->valid) {
2083 memcpy(&info->data.ib, &rec,
2084 sizeof(info->data.ib));
2087 info->data.type = RDMA_CLASS_PORT_INFO_IB;
2089 spin_unlock_irqrestore(&sa_query->port->classport_lock,
2093 query->callback(query->context);
2096 static void ib_sa_classport_info_rec_release(struct ib_sa_query *sa_query)
2098 kfree(container_of(sa_query, struct ib_sa_classport_info_query,
2102 static int ib_sa_classport_info_rec_query(struct ib_sa_port *port,
2104 void (*callback)(void *context),
2106 struct ib_sa_query **sa_query)
2108 struct ib_mad_agent *agent;
2109 struct ib_sa_classport_info_query *query;
2110 struct ib_sa_mad *mad;
2111 gfp_t gfp_mask = GFP_KERNEL;
2114 agent = port->agent;
2116 query = kzalloc(sizeof(*query), gfp_mask);
2120 query->sa_query.port = port;
2121 query->sa_query.flags |= rdma_cap_opa_ah(port->agent->device,
2123 IB_SA_QUERY_OPA : 0;
2124 ret = alloc_mad(&query->sa_query, gfp_mask);
2128 query->callback = callback;
2129 query->context = context;
2131 mad = query->sa_query.mad_buf->mad;
2132 init_mad(&query->sa_query, agent);
2134 query->sa_query.callback = ib_sa_classport_info_rec_callback;
2135 query->sa_query.release = ib_sa_classport_info_rec_release;
2136 mad->mad_hdr.method = IB_MGMT_METHOD_GET;
2137 mad->mad_hdr.attr_id = cpu_to_be16(IB_SA_ATTR_CLASS_PORTINFO);
2138 mad->sa_hdr.comp_mask = 0;
2139 *sa_query = &query->sa_query;
2141 ret = send_mad(&query->sa_query, timeout_ms, gfp_mask);
2149 free_mad(&query->sa_query);
2156 static void update_ib_cpi(struct work_struct *work)
2158 struct ib_sa_port *port =
2159 container_of(work, struct ib_sa_port, ib_cpi_work.work);
2160 struct ib_classport_info_context *cb_context;
2161 unsigned long flags;
2164 /* If the classport info is valid, nothing
2167 spin_lock_irqsave(&port->classport_lock, flags);
2168 if (port->classport_info.valid) {
2169 spin_unlock_irqrestore(&port->classport_lock, flags);
2172 spin_unlock_irqrestore(&port->classport_lock, flags);
2174 cb_context = kmalloc(sizeof(*cb_context), GFP_KERNEL);
2178 init_completion(&cb_context->done);
2180 ret = ib_sa_classport_info_rec_query(port, 3000,
2181 ib_classportinfo_cb, cb_context,
2182 &cb_context->sa_query);
2185 wait_for_completion(&cb_context->done);
2188 spin_lock_irqsave(&port->classport_lock, flags);
2190 /* If the classport info is still not valid, the query should have
2191 * failed for some reason. Retry issuing the query
2193 if (!port->classport_info.valid) {
2194 port->classport_info.retry_cnt++;
2195 if (port->classport_info.retry_cnt <=
2196 IB_SA_CPI_MAX_RETRY_CNT) {
2197 unsigned long delay =
2198 msecs_to_jiffies(IB_SA_CPI_RETRY_WAIT);
2200 queue_delayed_work(ib_wq, &port->ib_cpi_work, delay);
2203 spin_unlock_irqrestore(&port->classport_lock, flags);
2209 static void send_handler(struct ib_mad_agent *agent,
2210 struct ib_mad_send_wc *mad_send_wc)
2212 struct ib_sa_query *query = mad_send_wc->send_buf->context[0];
2213 unsigned long flags;
2215 if (query->callback)
2216 switch (mad_send_wc->status) {
2218 /* No callback -- already got recv */
2220 case IB_WC_RESP_TIMEOUT_ERR:
2221 query->callback(query, -ETIMEDOUT, NULL);
2223 case IB_WC_WR_FLUSH_ERR:
2224 query->callback(query, -EINTR, NULL);
2227 query->callback(query, -EIO, NULL);
2231 spin_lock_irqsave(&idr_lock, flags);
2232 idr_remove(&query_idr, query->id);
2233 spin_unlock_irqrestore(&idr_lock, flags);
2237 ib_sa_client_put(query->client);
2238 query->release(query);
2241 static void recv_handler(struct ib_mad_agent *mad_agent,
2242 struct ib_mad_send_buf *send_buf,
2243 struct ib_mad_recv_wc *mad_recv_wc)
2245 struct ib_sa_query *query;
2250 query = send_buf->context[0];
2251 if (query->callback) {
2252 if (mad_recv_wc->wc->status == IB_WC_SUCCESS)
2253 query->callback(query,
2254 mad_recv_wc->recv_buf.mad->mad_hdr.status ?
2256 (struct ib_sa_mad *) mad_recv_wc->recv_buf.mad);
2258 query->callback(query, -EIO, NULL);
2261 ib_free_recv_mad(mad_recv_wc);
2264 static void update_sm_ah(struct work_struct *work)
2266 struct ib_sa_port *port =
2267 container_of(work, struct ib_sa_port, update_task);
2268 struct ib_sa_sm_ah *new_ah;
2269 struct ib_port_attr port_attr;
2270 struct rdma_ah_attr ah_attr;
2272 if (ib_query_port(port->agent->device, port->port_num, &port_attr)) {
2273 pr_warn("Couldn't query port\n");
2277 new_ah = kmalloc(sizeof(*new_ah), GFP_KERNEL);
2281 kref_init(&new_ah->ref);
2282 new_ah->src_path_mask = (1 << port_attr.lmc) - 1;
2284 new_ah->pkey_index = 0;
2285 if (ib_find_pkey(port->agent->device, port->port_num,
2286 IB_DEFAULT_PKEY_FULL, &new_ah->pkey_index))
2287 pr_err("Couldn't find index for default PKey\n");
2289 memset(&ah_attr, 0, sizeof(ah_attr));
2290 ah_attr.type = rdma_ah_find_type(port->agent->device,
2292 rdma_ah_set_dlid(&ah_attr, port_attr.sm_lid);
2293 rdma_ah_set_sl(&ah_attr, port_attr.sm_sl);
2294 rdma_ah_set_port_num(&ah_attr, port->port_num);
2295 if (port_attr.grh_required) {
2296 if (ah_attr.type == RDMA_AH_ATTR_TYPE_OPA) {
2297 rdma_ah_set_make_grd(&ah_attr, true);
2299 rdma_ah_set_ah_flags(&ah_attr, IB_AH_GRH);
2300 rdma_ah_set_subnet_prefix(&ah_attr,
2301 cpu_to_be64(port_attr.subnet_prefix));
2302 rdma_ah_set_interface_id(&ah_attr,
2303 cpu_to_be64(IB_SA_WELL_KNOWN_GUID));
2307 new_ah->ah = rdma_create_ah(port->agent->qp->pd, &ah_attr);
2308 if (IS_ERR(new_ah->ah)) {
2309 pr_warn("Couldn't create new SM AH\n");
2314 spin_lock_irq(&port->ah_lock);
2316 kref_put(&port->sm_ah->ref, free_sm_ah);
2317 port->sm_ah = new_ah;
2318 spin_unlock_irq(&port->ah_lock);
2321 static void ib_sa_event(struct ib_event_handler *handler,
2322 struct ib_event *event)
2324 if (event->event == IB_EVENT_PORT_ERR ||
2325 event->event == IB_EVENT_PORT_ACTIVE ||
2326 event->event == IB_EVENT_LID_CHANGE ||
2327 event->event == IB_EVENT_PKEY_CHANGE ||
2328 event->event == IB_EVENT_SM_CHANGE ||
2329 event->event == IB_EVENT_CLIENT_REREGISTER) {
2330 unsigned long flags;
2331 struct ib_sa_device *sa_dev =
2332 container_of(handler, typeof(*sa_dev), event_handler);
2333 u8 port_num = event->element.port_num - sa_dev->start_port;
2334 struct ib_sa_port *port = &sa_dev->port[port_num];
2336 if (!rdma_cap_ib_sa(handler->device, port->port_num))
2339 spin_lock_irqsave(&port->ah_lock, flags);
2341 kref_put(&port->sm_ah->ref, free_sm_ah);
2343 spin_unlock_irqrestore(&port->ah_lock, flags);
2345 if (event->event == IB_EVENT_SM_CHANGE ||
2346 event->event == IB_EVENT_CLIENT_REREGISTER ||
2347 event->event == IB_EVENT_LID_CHANGE ||
2348 event->event == IB_EVENT_PORT_ACTIVE) {
2349 unsigned long delay =
2350 msecs_to_jiffies(IB_SA_CPI_RETRY_WAIT);
2352 spin_lock_irqsave(&port->classport_lock, flags);
2353 port->classport_info.valid = false;
2354 port->classport_info.retry_cnt = 0;
2355 spin_unlock_irqrestore(&port->classport_lock, flags);
2356 queue_delayed_work(ib_wq,
2357 &port->ib_cpi_work, delay);
2359 queue_work(ib_wq, &sa_dev->port[port_num].update_task);
2363 static void ib_sa_add_one(struct ib_device *device)
2365 struct ib_sa_device *sa_dev;
2369 s = rdma_start_port(device);
2370 e = rdma_end_port(device);
2372 sa_dev = kzalloc(sizeof *sa_dev +
2373 (e - s + 1) * sizeof (struct ib_sa_port),
2378 sa_dev->start_port = s;
2379 sa_dev->end_port = e;
2381 for (i = 0; i <= e - s; ++i) {
2382 spin_lock_init(&sa_dev->port[i].ah_lock);
2383 if (!rdma_cap_ib_sa(device, i + 1))
2386 sa_dev->port[i].sm_ah = NULL;
2387 sa_dev->port[i].port_num = i + s;
2389 spin_lock_init(&sa_dev->port[i].classport_lock);
2390 sa_dev->port[i].classport_info.valid = false;
2392 sa_dev->port[i].agent =
2393 ib_register_mad_agent(device, i + s, IB_QPT_GSI,
2394 NULL, 0, send_handler,
2395 recv_handler, sa_dev, 0);
2396 if (IS_ERR(sa_dev->port[i].agent))
2399 INIT_WORK(&sa_dev->port[i].update_task, update_sm_ah);
2400 INIT_DELAYED_WORK(&sa_dev->port[i].ib_cpi_work,
2409 ib_set_client_data(device, &sa_client, sa_dev);
2412 * We register our event handler after everything is set up,
2413 * and then update our cached info after the event handler is
2414 * registered to avoid any problems if a port changes state
2415 * during our initialization.
2418 INIT_IB_EVENT_HANDLER(&sa_dev->event_handler, device, ib_sa_event);
2419 ib_register_event_handler(&sa_dev->event_handler);
2421 for (i = 0; i <= e - s; ++i) {
2422 if (rdma_cap_ib_sa(device, i + 1))
2423 update_sm_ah(&sa_dev->port[i].update_task);
2430 if (rdma_cap_ib_sa(device, i + 1))
2431 ib_unregister_mad_agent(sa_dev->port[i].agent);
2438 static void ib_sa_remove_one(struct ib_device *device, void *client_data)
2440 struct ib_sa_device *sa_dev = client_data;
2446 ib_unregister_event_handler(&sa_dev->event_handler);
2447 flush_workqueue(ib_wq);
2449 for (i = 0; i <= sa_dev->end_port - sa_dev->start_port; ++i) {
2450 if (rdma_cap_ib_sa(device, i + 1)) {
2451 cancel_delayed_work_sync(&sa_dev->port[i].ib_cpi_work);
2452 ib_unregister_mad_agent(sa_dev->port[i].agent);
2453 if (sa_dev->port[i].sm_ah)
2454 kref_put(&sa_dev->port[i].sm_ah->ref, free_sm_ah);
2462 int ib_sa_init(void)
2466 get_random_bytes(&tid, sizeof tid);
2468 atomic_set(&ib_nl_sa_request_seq, 0);
2470 ret = ib_register_client(&sa_client);
2472 pr_err("Couldn't register ib_sa client\n");
2478 pr_err("Couldn't initialize multicast handling\n");
2482 ib_nl_wq = alloc_ordered_workqueue("ib_nl_sa_wq", WQ_MEM_RECLAIM);
2488 INIT_DELAYED_WORK(&ib_nl_timed_work, ib_nl_request_timeout);
2495 ib_unregister_client(&sa_client);
2500 void ib_sa_cleanup(void)
2502 cancel_delayed_work(&ib_nl_timed_work);
2503 flush_workqueue(ib_nl_wq);
2504 destroy_workqueue(ib_nl_wq);
2506 ib_unregister_client(&sa_client);
2507 idr_destroy(&query_idr);