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_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 int 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_XARRAY_FLAGS(queries, XA_FLAGS_ALLOC | XA_FLAGS_LOCK_IRQ);
188 static DEFINE_SPINLOCK(tid_lock);
191 #define PATH_REC_FIELD(field) \
192 .struct_offset_bytes = offsetof(struct sa_path_rec, field), \
193 .struct_size_bytes = sizeof_field(struct sa_path_rec, field), \
194 .field_name = "sa_path_rec:" #field
196 static const struct ib_field path_rec_table[] = {
197 { PATH_REC_FIELD(service_id),
201 { PATH_REC_FIELD(dgid),
205 { PATH_REC_FIELD(sgid),
209 { PATH_REC_FIELD(ib.dlid),
213 { PATH_REC_FIELD(ib.slid),
217 { PATH_REC_FIELD(ib.raw_traffic),
225 { PATH_REC_FIELD(flow_label),
229 { PATH_REC_FIELD(hop_limit),
233 { PATH_REC_FIELD(traffic_class),
237 { PATH_REC_FIELD(reversible),
241 { PATH_REC_FIELD(numb_path),
245 { PATH_REC_FIELD(pkey),
249 { PATH_REC_FIELD(qos_class),
253 { PATH_REC_FIELD(sl),
257 { PATH_REC_FIELD(mtu_selector),
261 { PATH_REC_FIELD(mtu),
265 { PATH_REC_FIELD(rate_selector),
269 { PATH_REC_FIELD(rate),
273 { PATH_REC_FIELD(packet_life_time_selector),
277 { PATH_REC_FIELD(packet_life_time),
281 { PATH_REC_FIELD(preference),
291 #define OPA_PATH_REC_FIELD(field) \
292 .struct_offset_bytes = \
293 offsetof(struct sa_path_rec, field), \
294 .struct_size_bytes = \
295 sizeof_field(struct sa_path_rec, field), \
296 .field_name = "sa_path_rec:" #field
298 static const struct ib_field opa_path_rec_table[] = {
299 { OPA_PATH_REC_FIELD(service_id),
303 { OPA_PATH_REC_FIELD(dgid),
307 { OPA_PATH_REC_FIELD(sgid),
311 { OPA_PATH_REC_FIELD(opa.dlid),
315 { OPA_PATH_REC_FIELD(opa.slid),
319 { OPA_PATH_REC_FIELD(opa.raw_traffic),
327 { OPA_PATH_REC_FIELD(flow_label),
331 { OPA_PATH_REC_FIELD(hop_limit),
335 { OPA_PATH_REC_FIELD(traffic_class),
339 { OPA_PATH_REC_FIELD(reversible),
343 { OPA_PATH_REC_FIELD(numb_path),
347 { OPA_PATH_REC_FIELD(pkey),
351 { OPA_PATH_REC_FIELD(opa.l2_8B),
355 { OPA_PATH_REC_FIELD(opa.l2_10B),
359 { OPA_PATH_REC_FIELD(opa.l2_9B),
363 { OPA_PATH_REC_FIELD(opa.l2_16B),
371 { OPA_PATH_REC_FIELD(opa.qos_type),
375 { OPA_PATH_REC_FIELD(opa.qos_priority),
383 { OPA_PATH_REC_FIELD(sl),
391 { OPA_PATH_REC_FIELD(mtu_selector),
395 { OPA_PATH_REC_FIELD(mtu),
399 { OPA_PATH_REC_FIELD(rate_selector),
403 { OPA_PATH_REC_FIELD(rate),
407 { OPA_PATH_REC_FIELD(packet_life_time_selector),
411 { OPA_PATH_REC_FIELD(packet_life_time),
415 { OPA_PATH_REC_FIELD(preference),
421 #define MCMEMBER_REC_FIELD(field) \
422 .struct_offset_bytes = offsetof(struct ib_sa_mcmember_rec, field), \
423 .struct_size_bytes = sizeof_field(struct ib_sa_mcmember_rec, field), \
424 .field_name = "sa_mcmember_rec:" #field
426 static const struct ib_field mcmember_rec_table[] = {
427 { MCMEMBER_REC_FIELD(mgid),
431 { MCMEMBER_REC_FIELD(port_gid),
435 { MCMEMBER_REC_FIELD(qkey),
439 { MCMEMBER_REC_FIELD(mlid),
443 { MCMEMBER_REC_FIELD(mtu_selector),
447 { MCMEMBER_REC_FIELD(mtu),
451 { MCMEMBER_REC_FIELD(traffic_class),
455 { MCMEMBER_REC_FIELD(pkey),
459 { MCMEMBER_REC_FIELD(rate_selector),
463 { MCMEMBER_REC_FIELD(rate),
467 { MCMEMBER_REC_FIELD(packet_life_time_selector),
471 { MCMEMBER_REC_FIELD(packet_life_time),
475 { MCMEMBER_REC_FIELD(sl),
479 { MCMEMBER_REC_FIELD(flow_label),
483 { MCMEMBER_REC_FIELD(hop_limit),
487 { MCMEMBER_REC_FIELD(scope),
491 { MCMEMBER_REC_FIELD(join_state),
495 { MCMEMBER_REC_FIELD(proxy_join),
505 #define SERVICE_REC_FIELD(field) \
506 .struct_offset_bytes = offsetof(struct ib_sa_service_rec, field), \
507 .struct_size_bytes = sizeof_field(struct ib_sa_service_rec, field), \
508 .field_name = "sa_service_rec:" #field
510 static const struct ib_field service_rec_table[] = {
511 { SERVICE_REC_FIELD(id),
515 { SERVICE_REC_FIELD(gid),
519 { SERVICE_REC_FIELD(pkey),
523 { SERVICE_REC_FIELD(lease),
527 { SERVICE_REC_FIELD(key),
531 { SERVICE_REC_FIELD(name),
535 { SERVICE_REC_FIELD(data8),
539 { SERVICE_REC_FIELD(data16),
543 { SERVICE_REC_FIELD(data32),
547 { SERVICE_REC_FIELD(data64),
553 #define CLASSPORTINFO_REC_FIELD(field) \
554 .struct_offset_bytes = offsetof(struct ib_class_port_info, field), \
555 .struct_size_bytes = sizeof_field(struct ib_class_port_info, field), \
556 .field_name = "ib_class_port_info:" #field
558 static const struct ib_field ib_classport_info_rec_table[] = {
559 { CLASSPORTINFO_REC_FIELD(base_version),
563 { CLASSPORTINFO_REC_FIELD(class_version),
567 { CLASSPORTINFO_REC_FIELD(capability_mask),
571 { CLASSPORTINFO_REC_FIELD(cap_mask2_resp_time),
575 { CLASSPORTINFO_REC_FIELD(redirect_gid),
579 { CLASSPORTINFO_REC_FIELD(redirect_tcslfl),
583 { CLASSPORTINFO_REC_FIELD(redirect_lid),
587 { CLASSPORTINFO_REC_FIELD(redirect_pkey),
592 { CLASSPORTINFO_REC_FIELD(redirect_qp),
596 { CLASSPORTINFO_REC_FIELD(redirect_qkey),
601 { CLASSPORTINFO_REC_FIELD(trap_gid),
605 { CLASSPORTINFO_REC_FIELD(trap_tcslfl),
610 { CLASSPORTINFO_REC_FIELD(trap_lid),
614 { CLASSPORTINFO_REC_FIELD(trap_pkey),
619 { CLASSPORTINFO_REC_FIELD(trap_hlqp),
623 { CLASSPORTINFO_REC_FIELD(trap_qkey),
629 #define OPA_CLASSPORTINFO_REC_FIELD(field) \
630 .struct_offset_bytes =\
631 offsetof(struct opa_class_port_info, field), \
632 .struct_size_bytes = \
633 sizeof_field(struct opa_class_port_info, field), \
634 .field_name = "opa_class_port_info:" #field
636 static const struct ib_field opa_classport_info_rec_table[] = {
637 { OPA_CLASSPORTINFO_REC_FIELD(base_version),
641 { OPA_CLASSPORTINFO_REC_FIELD(class_version),
645 { OPA_CLASSPORTINFO_REC_FIELD(cap_mask),
649 { OPA_CLASSPORTINFO_REC_FIELD(cap_mask2_resp_time),
653 { OPA_CLASSPORTINFO_REC_FIELD(redirect_gid),
657 { OPA_CLASSPORTINFO_REC_FIELD(redirect_tc_fl),
661 { OPA_CLASSPORTINFO_REC_FIELD(redirect_lid),
665 { OPA_CLASSPORTINFO_REC_FIELD(redirect_sl_qp),
669 { OPA_CLASSPORTINFO_REC_FIELD(redirect_qkey),
673 { OPA_CLASSPORTINFO_REC_FIELD(trap_gid),
677 { OPA_CLASSPORTINFO_REC_FIELD(trap_tc_fl),
681 { OPA_CLASSPORTINFO_REC_FIELD(trap_lid),
685 { OPA_CLASSPORTINFO_REC_FIELD(trap_hl_qp),
689 { OPA_CLASSPORTINFO_REC_FIELD(trap_qkey),
693 { OPA_CLASSPORTINFO_REC_FIELD(trap_pkey),
697 { OPA_CLASSPORTINFO_REC_FIELD(redirect_pkey),
701 { OPA_CLASSPORTINFO_REC_FIELD(trap_sl_rsvd),
711 #define GUIDINFO_REC_FIELD(field) \
712 .struct_offset_bytes = offsetof(struct ib_sa_guidinfo_rec, field), \
713 .struct_size_bytes = sizeof_field(struct ib_sa_guidinfo_rec, field), \
714 .field_name = "sa_guidinfo_rec:" #field
716 static const struct ib_field guidinfo_rec_table[] = {
717 { GUIDINFO_REC_FIELD(lid),
721 { GUIDINFO_REC_FIELD(block_num),
725 { GUIDINFO_REC_FIELD(res1),
729 { GUIDINFO_REC_FIELD(res2),
733 { GUIDINFO_REC_FIELD(guid_info_list),
739 static inline void ib_sa_disable_local_svc(struct ib_sa_query *query)
741 query->flags &= ~IB_SA_ENABLE_LOCAL_SERVICE;
744 static inline int ib_sa_query_cancelled(struct ib_sa_query *query)
746 return (query->flags & IB_SA_CANCEL);
749 static void ib_nl_set_path_rec_attrs(struct sk_buff *skb,
750 struct ib_sa_query *query)
752 struct sa_path_rec *sa_rec = query->mad_buf->context[1];
753 struct ib_sa_mad *mad = query->mad_buf->mad;
754 ib_sa_comp_mask comp_mask = mad->sa_hdr.comp_mask;
757 struct rdma_ls_resolve_header *header;
759 query->mad_buf->context[1] = NULL;
761 /* Construct the family header first */
762 header = skb_put(skb, NLMSG_ALIGN(sizeof(*header)));
763 strscpy_pad(header->device_name,
764 dev_name(&query->port->agent->device->dev),
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_make_request(struct ib_sa_query *query, gfp_t gfp_mask)
835 struct sk_buff *skb = NULL;
836 struct nlmsghdr *nlh;
838 struct ib_sa_mad *mad;
845 INIT_LIST_HEAD(&query->list);
846 query->seq = (u32)atomic_inc_return(&ib_nl_sa_request_seq);
848 mad = query->mad_buf->mad;
849 len = ib_nl_get_path_rec_attrs_len(mad->sa_hdr.comp_mask);
853 skb = nlmsg_new(len, gfp_mask);
857 /* Put nlmsg header only for now */
858 data = ibnl_put_msg(skb, &nlh, query->seq, 0, RDMA_NL_LS,
859 RDMA_NL_LS_OP_RESOLVE, NLM_F_REQUEST);
866 ib_nl_set_path_rec_attrs(skb, query);
868 /* Repair the nlmsg header length */
871 gfp_flag = ((gfp_mask & GFP_ATOMIC) == GFP_ATOMIC) ? GFP_ATOMIC :
874 spin_lock_irqsave(&ib_nl_request_lock, flags);
875 ret = rdma_nl_multicast(&init_net, skb, RDMA_NL_GROUP_LS, gfp_flag);
880 /* Put the request on the list.*/
881 delay = msecs_to_jiffies(sa_local_svc_timeout_ms);
882 query->timeout = delay + jiffies;
883 list_add_tail(&query->list, &ib_nl_request_list);
884 /* Start the timeout if this is the only request */
885 if (ib_nl_request_list.next == &query->list)
886 queue_delayed_work(ib_nl_wq, &ib_nl_timed_work, delay);
889 spin_unlock_irqrestore(&ib_nl_request_lock, flags);
894 static int ib_nl_cancel_request(struct ib_sa_query *query)
897 struct ib_sa_query *wait_query;
900 spin_lock_irqsave(&ib_nl_request_lock, flags);
901 list_for_each_entry(wait_query, &ib_nl_request_list, list) {
902 /* Let the timeout to take care of the callback */
903 if (query == wait_query) {
904 query->flags |= IB_SA_CANCEL;
905 query->timeout = jiffies;
906 list_move(&query->list, &ib_nl_request_list);
908 mod_delayed_work(ib_nl_wq, &ib_nl_timed_work, 1);
912 spin_unlock_irqrestore(&ib_nl_request_lock, flags);
917 static void send_handler(struct ib_mad_agent *agent,
918 struct ib_mad_send_wc *mad_send_wc);
920 static void ib_nl_process_good_resolve_rsp(struct ib_sa_query *query,
921 const struct nlmsghdr *nlh)
923 struct ib_mad_send_wc mad_send_wc;
924 struct ib_sa_mad *mad = NULL;
925 const struct nlattr *head, *curr;
926 struct ib_path_rec_data *rec;
931 if (query->callback) {
932 head = (const struct nlattr *) nlmsg_data(nlh);
933 len = nlmsg_len(nlh);
934 switch (query->path_use) {
935 case LS_RESOLVE_PATH_USE_UNIDIRECTIONAL:
936 mask = IB_PATH_PRIMARY | IB_PATH_OUTBOUND;
939 case LS_RESOLVE_PATH_USE_ALL:
940 case LS_RESOLVE_PATH_USE_GMP:
942 mask = IB_PATH_PRIMARY | IB_PATH_GMP |
943 IB_PATH_BIDIRECTIONAL;
946 nla_for_each_attr(curr, head, len, rem) {
947 if (curr->nla_type == LS_NLA_TYPE_PATH_RECORD) {
948 rec = nla_data(curr);
950 * Get the first one. In the future, we may
951 * need to get up to 6 pathrecords.
953 if ((rec->flags & mask) == mask) {
954 mad = query->mad_buf->mad;
955 mad->mad_hdr.method |=
957 memcpy(mad->data, rec->path_rec,
958 sizeof(rec->path_rec));
964 query->callback(query, status, mad);
967 mad_send_wc.send_buf = query->mad_buf;
968 mad_send_wc.status = IB_WC_SUCCESS;
969 send_handler(query->mad_buf->mad_agent, &mad_send_wc);
972 static void ib_nl_request_timeout(struct work_struct *work)
975 struct ib_sa_query *query;
977 struct ib_mad_send_wc mad_send_wc;
980 spin_lock_irqsave(&ib_nl_request_lock, flags);
981 while (!list_empty(&ib_nl_request_list)) {
982 query = list_entry(ib_nl_request_list.next,
983 struct ib_sa_query, list);
985 if (time_after(query->timeout, jiffies)) {
986 delay = query->timeout - jiffies;
987 if ((long)delay <= 0)
989 queue_delayed_work(ib_nl_wq, &ib_nl_timed_work, delay);
993 list_del(&query->list);
994 ib_sa_disable_local_svc(query);
995 /* Hold the lock to protect against query cancellation */
996 if (ib_sa_query_cancelled(query))
999 ret = ib_post_send_mad(query->mad_buf, NULL);
1001 mad_send_wc.send_buf = query->mad_buf;
1002 mad_send_wc.status = IB_WC_WR_FLUSH_ERR;
1003 spin_unlock_irqrestore(&ib_nl_request_lock, flags);
1004 send_handler(query->port->agent, &mad_send_wc);
1005 spin_lock_irqsave(&ib_nl_request_lock, flags);
1008 spin_unlock_irqrestore(&ib_nl_request_lock, flags);
1011 int ib_nl_handle_set_timeout(struct sk_buff *skb,
1012 struct nlmsghdr *nlh,
1013 struct netlink_ext_ack *extack)
1015 int timeout, delta, abs_delta;
1016 const struct nlattr *attr;
1017 unsigned long flags;
1018 struct ib_sa_query *query;
1020 struct nlattr *tb[LS_NLA_TYPE_MAX];
1023 if (!(nlh->nlmsg_flags & NLM_F_REQUEST) ||
1024 !(NETLINK_CB(skb).sk))
1027 ret = nla_parse_deprecated(tb, LS_NLA_TYPE_MAX - 1, nlmsg_data(nlh),
1028 nlmsg_len(nlh), ib_nl_policy, NULL);
1029 attr = (const struct nlattr *)tb[LS_NLA_TYPE_TIMEOUT];
1031 goto settimeout_out;
1033 timeout = *(int *) nla_data(attr);
1034 if (timeout < IB_SA_LOCAL_SVC_TIMEOUT_MIN)
1035 timeout = IB_SA_LOCAL_SVC_TIMEOUT_MIN;
1036 if (timeout > IB_SA_LOCAL_SVC_TIMEOUT_MAX)
1037 timeout = IB_SA_LOCAL_SVC_TIMEOUT_MAX;
1039 delta = timeout - sa_local_svc_timeout_ms;
1046 spin_lock_irqsave(&ib_nl_request_lock, flags);
1047 sa_local_svc_timeout_ms = timeout;
1048 list_for_each_entry(query, &ib_nl_request_list, list) {
1049 if (delta < 0 && abs_delta > query->timeout)
1052 query->timeout += delta;
1054 /* Get the new delay from the first entry */
1056 delay = query->timeout - jiffies;
1062 mod_delayed_work(ib_nl_wq, &ib_nl_timed_work,
1063 (unsigned long)delay);
1064 spin_unlock_irqrestore(&ib_nl_request_lock, flags);
1071 static inline int ib_nl_is_good_resolve_resp(const struct nlmsghdr *nlh)
1073 struct nlattr *tb[LS_NLA_TYPE_MAX];
1076 if (nlh->nlmsg_flags & RDMA_NL_LS_F_ERR)
1079 ret = nla_parse_deprecated(tb, LS_NLA_TYPE_MAX - 1, nlmsg_data(nlh),
1080 nlmsg_len(nlh), ib_nl_policy, NULL);
1087 int ib_nl_handle_resolve_resp(struct sk_buff *skb,
1088 struct nlmsghdr *nlh,
1089 struct netlink_ext_ack *extack)
1091 unsigned long flags;
1092 struct ib_sa_query *query;
1093 struct ib_mad_send_buf *send_buf;
1094 struct ib_mad_send_wc mad_send_wc;
1098 if ((nlh->nlmsg_flags & NLM_F_REQUEST) ||
1099 !(NETLINK_CB(skb).sk))
1102 spin_lock_irqsave(&ib_nl_request_lock, flags);
1103 list_for_each_entry(query, &ib_nl_request_list, list) {
1105 * If the query is cancelled, let the timeout routine
1108 if (nlh->nlmsg_seq == query->seq) {
1109 found = !ib_sa_query_cancelled(query);
1111 list_del(&query->list);
1117 spin_unlock_irqrestore(&ib_nl_request_lock, flags);
1121 send_buf = query->mad_buf;
1123 if (!ib_nl_is_good_resolve_resp(nlh)) {
1124 /* if the result is a failure, send out the packet via IB */
1125 ib_sa_disable_local_svc(query);
1126 ret = ib_post_send_mad(query->mad_buf, NULL);
1127 spin_unlock_irqrestore(&ib_nl_request_lock, flags);
1129 mad_send_wc.send_buf = send_buf;
1130 mad_send_wc.status = IB_WC_GENERAL_ERR;
1131 send_handler(query->port->agent, &mad_send_wc);
1134 spin_unlock_irqrestore(&ib_nl_request_lock, flags);
1135 ib_nl_process_good_resolve_rsp(query, nlh);
1142 static void free_sm_ah(struct kref *kref)
1144 struct ib_sa_sm_ah *sm_ah = container_of(kref, struct ib_sa_sm_ah, ref);
1146 rdma_destroy_ah(sm_ah->ah, 0);
1150 void ib_sa_register_client(struct ib_sa_client *client)
1152 atomic_set(&client->users, 1);
1153 init_completion(&client->comp);
1155 EXPORT_SYMBOL(ib_sa_register_client);
1157 void ib_sa_unregister_client(struct ib_sa_client *client)
1159 ib_sa_client_put(client);
1160 wait_for_completion(&client->comp);
1162 EXPORT_SYMBOL(ib_sa_unregister_client);
1165 * ib_sa_cancel_query - try to cancel an SA query
1166 * @id:ID of query to cancel
1167 * @query:query pointer to cancel
1169 * Try to cancel an SA query. If the id and query don't match up or
1170 * the query has already completed, nothing is done. Otherwise the
1171 * query is canceled and will complete with a status of -EINTR.
1173 void ib_sa_cancel_query(int id, struct ib_sa_query *query)
1175 unsigned long flags;
1176 struct ib_mad_agent *agent;
1177 struct ib_mad_send_buf *mad_buf;
1179 xa_lock_irqsave(&queries, flags);
1180 if (xa_load(&queries, id) != query) {
1181 xa_unlock_irqrestore(&queries, flags);
1184 agent = query->port->agent;
1185 mad_buf = query->mad_buf;
1186 xa_unlock_irqrestore(&queries, flags);
1189 * If the query is still on the netlink request list, schedule
1190 * it to be cancelled by the timeout routine. Otherwise, it has been
1191 * sent to the MAD layer and has to be cancelled from there.
1193 if (!ib_nl_cancel_request(query))
1194 ib_cancel_mad(agent, mad_buf);
1196 EXPORT_SYMBOL(ib_sa_cancel_query);
1198 static u8 get_src_path_mask(struct ib_device *device, u8 port_num)
1200 struct ib_sa_device *sa_dev;
1201 struct ib_sa_port *port;
1202 unsigned long flags;
1205 sa_dev = ib_get_client_data(device, &sa_client);
1209 port = &sa_dev->port[port_num - sa_dev->start_port];
1210 spin_lock_irqsave(&port->ah_lock, flags);
1211 src_path_mask = port->sm_ah ? port->sm_ah->src_path_mask : 0x7f;
1212 spin_unlock_irqrestore(&port->ah_lock, flags);
1214 return src_path_mask;
1217 static int init_ah_attr_grh_fields(struct ib_device *device, u8 port_num,
1218 struct sa_path_rec *rec,
1219 struct rdma_ah_attr *ah_attr,
1220 const struct ib_gid_attr *gid_attr)
1222 enum ib_gid_type type = sa_conv_pathrec_to_gid_type(rec);
1225 gid_attr = rdma_find_gid_by_port(device, &rec->sgid, type,
1227 if (IS_ERR(gid_attr))
1228 return PTR_ERR(gid_attr);
1230 rdma_hold_gid_attr(gid_attr);
1232 rdma_move_grh_sgid_attr(ah_attr, &rec->dgid,
1233 be32_to_cpu(rec->flow_label),
1234 rec->hop_limit, rec->traffic_class,
1240 * ib_init_ah_attr_from_path - Initialize address handle attributes based on
1241 * an SA path record.
1242 * @device: Device associated ah attributes initialization.
1243 * @port_num: Port on the specified device.
1244 * @rec: path record entry to use for ah attributes initialization.
1245 * @ah_attr: address handle attributes to initialization from path record.
1246 * @gid_attr: SGID attribute to consider during initialization.
1248 * When ib_init_ah_attr_from_path() returns success,
1249 * (a) for IB link layer it optionally contains a reference to SGID attribute
1250 * when GRH is present for IB link layer.
1251 * (b) for RoCE link layer it contains a reference to SGID attribute.
1252 * User must invoke rdma_destroy_ah_attr() to release reference to SGID
1253 * attributes which are initialized using ib_init_ah_attr_from_path().
1255 int ib_init_ah_attr_from_path(struct ib_device *device, u8 port_num,
1256 struct sa_path_rec *rec,
1257 struct rdma_ah_attr *ah_attr,
1258 const struct ib_gid_attr *gid_attr)
1262 memset(ah_attr, 0, sizeof(*ah_attr));
1263 ah_attr->type = rdma_ah_find_type(device, port_num);
1264 rdma_ah_set_sl(ah_attr, rec->sl);
1265 rdma_ah_set_port_num(ah_attr, port_num);
1266 rdma_ah_set_static_rate(ah_attr, rec->rate);
1268 if (sa_path_is_roce(rec)) {
1269 ret = roce_resolve_route_from_path(rec, gid_attr);
1273 memcpy(ah_attr->roce.dmac, sa_path_get_dmac(rec), ETH_ALEN);
1275 rdma_ah_set_dlid(ah_attr, be32_to_cpu(sa_path_get_dlid(rec)));
1276 if (sa_path_is_opa(rec) &&
1277 rdma_ah_get_dlid(ah_attr) == be16_to_cpu(IB_LID_PERMISSIVE))
1278 rdma_ah_set_make_grd(ah_attr, true);
1280 rdma_ah_set_path_bits(ah_attr,
1281 be32_to_cpu(sa_path_get_slid(rec)) &
1282 get_src_path_mask(device, port_num));
1285 if (rec->hop_limit > 0 || sa_path_is_roce(rec))
1286 ret = init_ah_attr_grh_fields(device, port_num,
1287 rec, ah_attr, gid_attr);
1290 EXPORT_SYMBOL(ib_init_ah_attr_from_path);
1292 static int alloc_mad(struct ib_sa_query *query, gfp_t gfp_mask)
1294 struct rdma_ah_attr ah_attr;
1295 unsigned long flags;
1297 spin_lock_irqsave(&query->port->ah_lock, flags);
1298 if (!query->port->sm_ah) {
1299 spin_unlock_irqrestore(&query->port->ah_lock, flags);
1302 kref_get(&query->port->sm_ah->ref);
1303 query->sm_ah = query->port->sm_ah;
1304 spin_unlock_irqrestore(&query->port->ah_lock, flags);
1307 * Always check if sm_ah has valid dlid assigned,
1308 * before querying for class port info
1310 if ((rdma_query_ah(query->sm_ah->ah, &ah_attr) < 0) ||
1311 !rdma_is_valid_unicast_lid(&ah_attr)) {
1312 kref_put(&query->sm_ah->ref, free_sm_ah);
1315 query->mad_buf = ib_create_send_mad(query->port->agent, 1,
1316 query->sm_ah->pkey_index,
1317 0, IB_MGMT_SA_HDR, IB_MGMT_SA_DATA,
1319 ((query->flags & IB_SA_QUERY_OPA) ?
1320 OPA_MGMT_BASE_VERSION :
1321 IB_MGMT_BASE_VERSION));
1322 if (IS_ERR(query->mad_buf)) {
1323 kref_put(&query->sm_ah->ref, free_sm_ah);
1327 query->mad_buf->ah = query->sm_ah->ah;
1332 static void free_mad(struct ib_sa_query *query)
1334 ib_free_send_mad(query->mad_buf);
1335 kref_put(&query->sm_ah->ref, free_sm_ah);
1338 static void init_mad(struct ib_sa_query *query, struct ib_mad_agent *agent)
1340 struct ib_sa_mad *mad = query->mad_buf->mad;
1341 unsigned long flags;
1343 memset(mad, 0, sizeof *mad);
1345 if (query->flags & IB_SA_QUERY_OPA) {
1346 mad->mad_hdr.base_version = OPA_MGMT_BASE_VERSION;
1347 mad->mad_hdr.class_version = OPA_SA_CLASS_VERSION;
1349 mad->mad_hdr.base_version = IB_MGMT_BASE_VERSION;
1350 mad->mad_hdr.class_version = IB_SA_CLASS_VERSION;
1352 mad->mad_hdr.mgmt_class = IB_MGMT_CLASS_SUBN_ADM;
1353 spin_lock_irqsave(&tid_lock, flags);
1355 cpu_to_be64(((u64) agent->hi_tid) << 32 | tid++);
1356 spin_unlock_irqrestore(&tid_lock, flags);
1359 static int send_mad(struct ib_sa_query *query, unsigned long timeout_ms,
1362 unsigned long flags;
1365 xa_lock_irqsave(&queries, flags);
1366 ret = __xa_alloc(&queries, &id, query, xa_limit_32b, gfp_mask);
1367 xa_unlock_irqrestore(&queries, flags);
1371 query->mad_buf->timeout_ms = timeout_ms;
1372 query->mad_buf->context[0] = query;
1375 if ((query->flags & IB_SA_ENABLE_LOCAL_SERVICE) &&
1376 (!(query->flags & IB_SA_QUERY_OPA))) {
1377 if (rdma_nl_chk_listeners(RDMA_NL_GROUP_LS)) {
1378 if (!ib_nl_make_request(query, gfp_mask))
1381 ib_sa_disable_local_svc(query);
1384 ret = ib_post_send_mad(query->mad_buf, NULL);
1386 xa_lock_irqsave(&queries, flags);
1387 __xa_erase(&queries, id);
1388 xa_unlock_irqrestore(&queries, flags);
1392 * It's not safe to dereference query any more, because the
1393 * send may already have completed and freed the query in
1396 return ret ? ret : id;
1399 void ib_sa_unpack_path(void *attribute, struct sa_path_rec *rec)
1401 ib_unpack(path_rec_table, ARRAY_SIZE(path_rec_table), attribute, rec);
1403 EXPORT_SYMBOL(ib_sa_unpack_path);
1405 void ib_sa_pack_path(struct sa_path_rec *rec, void *attribute)
1407 ib_pack(path_rec_table, ARRAY_SIZE(path_rec_table), rec, attribute);
1409 EXPORT_SYMBOL(ib_sa_pack_path);
1411 static bool ib_sa_opa_pathrecord_support(struct ib_sa_client *client,
1412 struct ib_sa_device *sa_dev,
1415 struct ib_sa_port *port;
1416 unsigned long flags;
1419 port = &sa_dev->port[port_num - sa_dev->start_port];
1420 spin_lock_irqsave(&port->classport_lock, flags);
1421 if (!port->classport_info.valid)
1424 if (port->classport_info.data.type == RDMA_CLASS_PORT_INFO_OPA)
1425 ret = opa_get_cpi_capmask2(&port->classport_info.data.opa) &
1426 OPA_CLASS_PORT_INFO_PR_SUPPORT;
1428 spin_unlock_irqrestore(&port->classport_lock, flags);
1432 enum opa_pr_supported {
1439 * Check if current PR query can be an OPA query.
1440 * Retuns PR_NOT_SUPPORTED if a path record query is not
1441 * possible, PR_OPA_SUPPORTED if an OPA path record query
1442 * is possible and PR_IB_SUPPORTED if an IB path record
1443 * query is possible.
1445 static int opa_pr_query_possible(struct ib_sa_client *client,
1446 struct ib_sa_device *sa_dev,
1447 struct ib_device *device, u8 port_num,
1448 struct sa_path_rec *rec)
1450 struct ib_port_attr port_attr;
1452 if (ib_query_port(device, port_num, &port_attr))
1453 return PR_NOT_SUPPORTED;
1455 if (ib_sa_opa_pathrecord_support(client, sa_dev, port_num))
1456 return PR_OPA_SUPPORTED;
1458 if (port_attr.lid >= be16_to_cpu(IB_MULTICAST_LID_BASE))
1459 return PR_NOT_SUPPORTED;
1461 return PR_IB_SUPPORTED;
1464 static void ib_sa_path_rec_callback(struct ib_sa_query *sa_query,
1466 struct ib_sa_mad *mad)
1468 struct ib_sa_path_query *query =
1469 container_of(sa_query, struct ib_sa_path_query, sa_query);
1472 struct sa_path_rec rec;
1474 if (sa_query->flags & IB_SA_QUERY_OPA) {
1475 ib_unpack(opa_path_rec_table,
1476 ARRAY_SIZE(opa_path_rec_table),
1478 rec.rec_type = SA_PATH_REC_TYPE_OPA;
1479 query->callback(status, &rec, query->context);
1481 ib_unpack(path_rec_table,
1482 ARRAY_SIZE(path_rec_table),
1484 rec.rec_type = SA_PATH_REC_TYPE_IB;
1485 sa_path_set_dmac_zero(&rec);
1487 if (query->conv_pr) {
1488 struct sa_path_rec opa;
1490 memset(&opa, 0, sizeof(struct sa_path_rec));
1491 sa_convert_path_ib_to_opa(&opa, &rec);
1492 query->callback(status, &opa, query->context);
1494 query->callback(status, &rec, query->context);
1498 query->callback(status, NULL, query->context);
1501 static void ib_sa_path_rec_release(struct ib_sa_query *sa_query)
1503 struct ib_sa_path_query *query =
1504 container_of(sa_query, struct ib_sa_path_query, sa_query);
1506 kfree(query->conv_pr);
1511 * ib_sa_path_rec_get - Start a Path get query
1513 * @device:device to send query on
1514 * @port_num: port number to send query on
1515 * @rec:Path Record to send in query
1516 * @comp_mask:component mask to send in query
1517 * @timeout_ms:time to wait for response
1518 * @gfp_mask:GFP mask to use for internal allocations
1519 * @callback:function called when query completes, times out or is
1521 * @context:opaque user context passed to callback
1522 * @sa_query:query context, used to cancel query
1524 * Send a Path Record Get query to the SA to look up a path. The
1525 * callback function will be called when the query completes (or
1526 * fails); status is 0 for a successful response, -EINTR if the query
1527 * is canceled, -ETIMEDOUT is the query timed out, or -EIO if an error
1528 * occurred sending the query. The resp parameter of the callback is
1529 * only valid if status is 0.
1531 * If the return value of ib_sa_path_rec_get() is negative, it is an
1532 * error code. Otherwise it is a query ID that can be used to cancel
1535 int ib_sa_path_rec_get(struct ib_sa_client *client,
1536 struct ib_device *device, u8 port_num,
1537 struct sa_path_rec *rec,
1538 ib_sa_comp_mask comp_mask,
1539 unsigned long timeout_ms, gfp_t gfp_mask,
1540 void (*callback)(int status,
1541 struct sa_path_rec *resp,
1544 struct ib_sa_query **sa_query)
1546 struct ib_sa_path_query *query;
1547 struct ib_sa_device *sa_dev = ib_get_client_data(device, &sa_client);
1548 struct ib_sa_port *port;
1549 struct ib_mad_agent *agent;
1550 struct ib_sa_mad *mad;
1551 enum opa_pr_supported status;
1557 if ((rec->rec_type != SA_PATH_REC_TYPE_IB) &&
1558 (rec->rec_type != SA_PATH_REC_TYPE_OPA))
1561 port = &sa_dev->port[port_num - sa_dev->start_port];
1562 agent = port->agent;
1564 query = kzalloc(sizeof(*query), gfp_mask);
1568 query->sa_query.port = port;
1569 if (rec->rec_type == SA_PATH_REC_TYPE_OPA) {
1570 status = opa_pr_query_possible(client, sa_dev, device, port_num,
1572 if (status == PR_NOT_SUPPORTED) {
1575 } else if (status == PR_OPA_SUPPORTED) {
1576 query->sa_query.flags |= IB_SA_QUERY_OPA;
1579 kmalloc(sizeof(*query->conv_pr), gfp_mask);
1580 if (!query->conv_pr) {
1587 ret = alloc_mad(&query->sa_query, gfp_mask);
1591 ib_sa_client_get(client);
1592 query->sa_query.client = client;
1593 query->callback = callback;
1594 query->context = context;
1596 mad = query->sa_query.mad_buf->mad;
1597 init_mad(&query->sa_query, agent);
1599 query->sa_query.callback = callback ? ib_sa_path_rec_callback : NULL;
1600 query->sa_query.release = ib_sa_path_rec_release;
1601 mad->mad_hdr.method = IB_MGMT_METHOD_GET;
1602 mad->mad_hdr.attr_id = cpu_to_be16(IB_SA_ATTR_PATH_REC);
1603 mad->sa_hdr.comp_mask = comp_mask;
1605 if (query->sa_query.flags & IB_SA_QUERY_OPA) {
1606 ib_pack(opa_path_rec_table, ARRAY_SIZE(opa_path_rec_table),
1608 } else if (query->conv_pr) {
1609 sa_convert_path_opa_to_ib(query->conv_pr, rec);
1610 ib_pack(path_rec_table, ARRAY_SIZE(path_rec_table),
1611 query->conv_pr, mad->data);
1613 ib_pack(path_rec_table, ARRAY_SIZE(path_rec_table),
1617 *sa_query = &query->sa_query;
1619 query->sa_query.flags |= IB_SA_ENABLE_LOCAL_SERVICE;
1620 query->sa_query.mad_buf->context[1] = (query->conv_pr) ?
1621 query->conv_pr : rec;
1623 ret = send_mad(&query->sa_query, timeout_ms, gfp_mask);
1631 ib_sa_client_put(query->sa_query.client);
1632 free_mad(&query->sa_query);
1634 kfree(query->conv_pr);
1639 EXPORT_SYMBOL(ib_sa_path_rec_get);
1641 static void ib_sa_service_rec_callback(struct ib_sa_query *sa_query,
1643 struct ib_sa_mad *mad)
1645 struct ib_sa_service_query *query =
1646 container_of(sa_query, struct ib_sa_service_query, sa_query);
1649 struct ib_sa_service_rec rec;
1651 ib_unpack(service_rec_table, ARRAY_SIZE(service_rec_table),
1653 query->callback(status, &rec, query->context);
1655 query->callback(status, NULL, query->context);
1658 static void ib_sa_service_rec_release(struct ib_sa_query *sa_query)
1660 kfree(container_of(sa_query, struct ib_sa_service_query, sa_query));
1664 * ib_sa_service_rec_query - Start Service Record operation
1666 * @device:device to send request on
1667 * @port_num: port number to send request on
1668 * @method:SA method - should be get, set, or delete
1669 * @rec:Service Record to send in request
1670 * @comp_mask:component mask to send in request
1671 * @timeout_ms:time to wait for response
1672 * @gfp_mask:GFP mask to use for internal allocations
1673 * @callback:function called when request completes, times out or is
1675 * @context:opaque user context passed to callback
1676 * @sa_query:request context, used to cancel request
1678 * Send a Service Record set/get/delete to the SA to register,
1679 * unregister or query a service record.
1680 * The callback function will be called when the request completes (or
1681 * fails); status is 0 for a successful response, -EINTR if the query
1682 * is canceled, -ETIMEDOUT is the query timed out, or -EIO if an error
1683 * occurred sending the query. The resp parameter of the callback is
1684 * only valid if status is 0.
1686 * If the return value of ib_sa_service_rec_query() is negative, it is an
1687 * error code. Otherwise it is a request ID that can be used to cancel
1690 int ib_sa_service_rec_query(struct ib_sa_client *client,
1691 struct ib_device *device, u8 port_num, u8 method,
1692 struct ib_sa_service_rec *rec,
1693 ib_sa_comp_mask comp_mask,
1694 unsigned long timeout_ms, gfp_t gfp_mask,
1695 void (*callback)(int status,
1696 struct ib_sa_service_rec *resp,
1699 struct ib_sa_query **sa_query)
1701 struct ib_sa_service_query *query;
1702 struct ib_sa_device *sa_dev = ib_get_client_data(device, &sa_client);
1703 struct ib_sa_port *port;
1704 struct ib_mad_agent *agent;
1705 struct ib_sa_mad *mad;
1711 port = &sa_dev->port[port_num - sa_dev->start_port];
1712 agent = port->agent;
1714 if (method != IB_MGMT_METHOD_GET &&
1715 method != IB_MGMT_METHOD_SET &&
1716 method != IB_SA_METHOD_DELETE)
1719 query = kzalloc(sizeof(*query), gfp_mask);
1723 query->sa_query.port = port;
1724 ret = alloc_mad(&query->sa_query, gfp_mask);
1728 ib_sa_client_get(client);
1729 query->sa_query.client = client;
1730 query->callback = callback;
1731 query->context = context;
1733 mad = query->sa_query.mad_buf->mad;
1734 init_mad(&query->sa_query, agent);
1736 query->sa_query.callback = callback ? ib_sa_service_rec_callback : NULL;
1737 query->sa_query.release = ib_sa_service_rec_release;
1738 mad->mad_hdr.method = method;
1739 mad->mad_hdr.attr_id = cpu_to_be16(IB_SA_ATTR_SERVICE_REC);
1740 mad->sa_hdr.comp_mask = comp_mask;
1742 ib_pack(service_rec_table, ARRAY_SIZE(service_rec_table),
1745 *sa_query = &query->sa_query;
1747 ret = send_mad(&query->sa_query, timeout_ms, gfp_mask);
1755 ib_sa_client_put(query->sa_query.client);
1756 free_mad(&query->sa_query);
1762 EXPORT_SYMBOL(ib_sa_service_rec_query);
1764 static void ib_sa_mcmember_rec_callback(struct ib_sa_query *sa_query,
1766 struct ib_sa_mad *mad)
1768 struct ib_sa_mcmember_query *query =
1769 container_of(sa_query, struct ib_sa_mcmember_query, sa_query);
1772 struct ib_sa_mcmember_rec rec;
1774 ib_unpack(mcmember_rec_table, ARRAY_SIZE(mcmember_rec_table),
1776 query->callback(status, &rec, query->context);
1778 query->callback(status, NULL, query->context);
1781 static void ib_sa_mcmember_rec_release(struct ib_sa_query *sa_query)
1783 kfree(container_of(sa_query, struct ib_sa_mcmember_query, sa_query));
1786 int ib_sa_mcmember_rec_query(struct ib_sa_client *client,
1787 struct ib_device *device, u8 port_num,
1789 struct ib_sa_mcmember_rec *rec,
1790 ib_sa_comp_mask comp_mask,
1791 unsigned long timeout_ms, gfp_t gfp_mask,
1792 void (*callback)(int status,
1793 struct ib_sa_mcmember_rec *resp,
1796 struct ib_sa_query **sa_query)
1798 struct ib_sa_mcmember_query *query;
1799 struct ib_sa_device *sa_dev = ib_get_client_data(device, &sa_client);
1800 struct ib_sa_port *port;
1801 struct ib_mad_agent *agent;
1802 struct ib_sa_mad *mad;
1808 port = &sa_dev->port[port_num - sa_dev->start_port];
1809 agent = port->agent;
1811 query = kzalloc(sizeof(*query), gfp_mask);
1815 query->sa_query.port = port;
1816 ret = alloc_mad(&query->sa_query, gfp_mask);
1820 ib_sa_client_get(client);
1821 query->sa_query.client = client;
1822 query->callback = callback;
1823 query->context = context;
1825 mad = query->sa_query.mad_buf->mad;
1826 init_mad(&query->sa_query, agent);
1828 query->sa_query.callback = callback ? ib_sa_mcmember_rec_callback : NULL;
1829 query->sa_query.release = ib_sa_mcmember_rec_release;
1830 mad->mad_hdr.method = method;
1831 mad->mad_hdr.attr_id = cpu_to_be16(IB_SA_ATTR_MC_MEMBER_REC);
1832 mad->sa_hdr.comp_mask = comp_mask;
1834 ib_pack(mcmember_rec_table, ARRAY_SIZE(mcmember_rec_table),
1837 *sa_query = &query->sa_query;
1839 ret = send_mad(&query->sa_query, timeout_ms, gfp_mask);
1847 ib_sa_client_put(query->sa_query.client);
1848 free_mad(&query->sa_query);
1855 /* Support GuidInfoRecord */
1856 static void ib_sa_guidinfo_rec_callback(struct ib_sa_query *sa_query,
1858 struct ib_sa_mad *mad)
1860 struct ib_sa_guidinfo_query *query =
1861 container_of(sa_query, struct ib_sa_guidinfo_query, sa_query);
1864 struct ib_sa_guidinfo_rec rec;
1866 ib_unpack(guidinfo_rec_table, ARRAY_SIZE(guidinfo_rec_table),
1868 query->callback(status, &rec, query->context);
1870 query->callback(status, NULL, query->context);
1873 static void ib_sa_guidinfo_rec_release(struct ib_sa_query *sa_query)
1875 kfree(container_of(sa_query, struct ib_sa_guidinfo_query, sa_query));
1878 int ib_sa_guid_info_rec_query(struct ib_sa_client *client,
1879 struct ib_device *device, u8 port_num,
1880 struct ib_sa_guidinfo_rec *rec,
1881 ib_sa_comp_mask comp_mask, u8 method,
1882 unsigned long timeout_ms, gfp_t gfp_mask,
1883 void (*callback)(int status,
1884 struct ib_sa_guidinfo_rec *resp,
1887 struct ib_sa_query **sa_query)
1889 struct ib_sa_guidinfo_query *query;
1890 struct ib_sa_device *sa_dev = ib_get_client_data(device, &sa_client);
1891 struct ib_sa_port *port;
1892 struct ib_mad_agent *agent;
1893 struct ib_sa_mad *mad;
1899 if (method != IB_MGMT_METHOD_GET &&
1900 method != IB_MGMT_METHOD_SET &&
1901 method != IB_SA_METHOD_DELETE) {
1905 port = &sa_dev->port[port_num - sa_dev->start_port];
1906 agent = port->agent;
1908 query = kzalloc(sizeof(*query), gfp_mask);
1912 query->sa_query.port = port;
1913 ret = alloc_mad(&query->sa_query, gfp_mask);
1917 ib_sa_client_get(client);
1918 query->sa_query.client = client;
1919 query->callback = callback;
1920 query->context = context;
1922 mad = query->sa_query.mad_buf->mad;
1923 init_mad(&query->sa_query, agent);
1925 query->sa_query.callback = callback ? ib_sa_guidinfo_rec_callback : NULL;
1926 query->sa_query.release = ib_sa_guidinfo_rec_release;
1928 mad->mad_hdr.method = method;
1929 mad->mad_hdr.attr_id = cpu_to_be16(IB_SA_ATTR_GUID_INFO_REC);
1930 mad->sa_hdr.comp_mask = comp_mask;
1932 ib_pack(guidinfo_rec_table, ARRAY_SIZE(guidinfo_rec_table), rec,
1935 *sa_query = &query->sa_query;
1937 ret = send_mad(&query->sa_query, timeout_ms, gfp_mask);
1945 ib_sa_client_put(query->sa_query.client);
1946 free_mad(&query->sa_query);
1952 EXPORT_SYMBOL(ib_sa_guid_info_rec_query);
1954 bool ib_sa_sendonly_fullmem_support(struct ib_sa_client *client,
1955 struct ib_device *device,
1958 struct ib_sa_device *sa_dev = ib_get_client_data(device, &sa_client);
1959 struct ib_sa_port *port;
1961 unsigned long flags;
1966 port = &sa_dev->port[port_num - sa_dev->start_port];
1968 spin_lock_irqsave(&port->classport_lock, flags);
1969 if ((port->classport_info.valid) &&
1970 (port->classport_info.data.type == RDMA_CLASS_PORT_INFO_IB))
1971 ret = ib_get_cpi_capmask2(&port->classport_info.data.ib)
1972 & IB_SA_CAP_MASK2_SENDONLY_FULL_MEM_SUPPORT;
1973 spin_unlock_irqrestore(&port->classport_lock, flags);
1976 EXPORT_SYMBOL(ib_sa_sendonly_fullmem_support);
1978 struct ib_classport_info_context {
1979 struct completion done;
1980 struct ib_sa_query *sa_query;
1983 static void ib_classportinfo_cb(void *context)
1985 struct ib_classport_info_context *cb_ctx = context;
1987 complete(&cb_ctx->done);
1990 static void ib_sa_classport_info_rec_callback(struct ib_sa_query *sa_query,
1992 struct ib_sa_mad *mad)
1994 unsigned long flags;
1995 struct ib_sa_classport_info_query *query =
1996 container_of(sa_query, struct ib_sa_classport_info_query, sa_query);
1997 struct ib_sa_classport_cache *info = &sa_query->port->classport_info;
2000 if (sa_query->flags & IB_SA_QUERY_OPA) {
2001 struct opa_class_port_info rec;
2003 ib_unpack(opa_classport_info_rec_table,
2004 ARRAY_SIZE(opa_classport_info_rec_table),
2007 spin_lock_irqsave(&sa_query->port->classport_lock,
2009 if (!status && !info->valid) {
2010 memcpy(&info->data.opa, &rec,
2011 sizeof(info->data.opa));
2014 info->data.type = RDMA_CLASS_PORT_INFO_OPA;
2016 spin_unlock_irqrestore(&sa_query->port->classport_lock,
2020 struct ib_class_port_info rec;
2022 ib_unpack(ib_classport_info_rec_table,
2023 ARRAY_SIZE(ib_classport_info_rec_table),
2026 spin_lock_irqsave(&sa_query->port->classport_lock,
2028 if (!status && !info->valid) {
2029 memcpy(&info->data.ib, &rec,
2030 sizeof(info->data.ib));
2033 info->data.type = RDMA_CLASS_PORT_INFO_IB;
2035 spin_unlock_irqrestore(&sa_query->port->classport_lock,
2039 query->callback(query->context);
2042 static void ib_sa_classport_info_rec_release(struct ib_sa_query *sa_query)
2044 kfree(container_of(sa_query, struct ib_sa_classport_info_query,
2048 static int ib_sa_classport_info_rec_query(struct ib_sa_port *port,
2049 unsigned long timeout_ms,
2050 void (*callback)(void *context),
2052 struct ib_sa_query **sa_query)
2054 struct ib_mad_agent *agent;
2055 struct ib_sa_classport_info_query *query;
2056 struct ib_sa_mad *mad;
2057 gfp_t gfp_mask = GFP_KERNEL;
2060 agent = port->agent;
2062 query = kzalloc(sizeof(*query), gfp_mask);
2066 query->sa_query.port = port;
2067 query->sa_query.flags |= rdma_cap_opa_ah(port->agent->device,
2069 IB_SA_QUERY_OPA : 0;
2070 ret = alloc_mad(&query->sa_query, gfp_mask);
2074 query->callback = callback;
2075 query->context = context;
2077 mad = query->sa_query.mad_buf->mad;
2078 init_mad(&query->sa_query, agent);
2080 query->sa_query.callback = ib_sa_classport_info_rec_callback;
2081 query->sa_query.release = ib_sa_classport_info_rec_release;
2082 mad->mad_hdr.method = IB_MGMT_METHOD_GET;
2083 mad->mad_hdr.attr_id = cpu_to_be16(IB_SA_ATTR_CLASS_PORTINFO);
2084 mad->sa_hdr.comp_mask = 0;
2085 *sa_query = &query->sa_query;
2087 ret = send_mad(&query->sa_query, timeout_ms, gfp_mask);
2095 free_mad(&query->sa_query);
2102 static void update_ib_cpi(struct work_struct *work)
2104 struct ib_sa_port *port =
2105 container_of(work, struct ib_sa_port, ib_cpi_work.work);
2106 struct ib_classport_info_context *cb_context;
2107 unsigned long flags;
2110 /* If the classport info is valid, nothing
2113 spin_lock_irqsave(&port->classport_lock, flags);
2114 if (port->classport_info.valid) {
2115 spin_unlock_irqrestore(&port->classport_lock, flags);
2118 spin_unlock_irqrestore(&port->classport_lock, flags);
2120 cb_context = kmalloc(sizeof(*cb_context), GFP_KERNEL);
2124 init_completion(&cb_context->done);
2126 ret = ib_sa_classport_info_rec_query(port, 3000,
2127 ib_classportinfo_cb, cb_context,
2128 &cb_context->sa_query);
2131 wait_for_completion(&cb_context->done);
2134 spin_lock_irqsave(&port->classport_lock, flags);
2136 /* If the classport info is still not valid, the query should have
2137 * failed for some reason. Retry issuing the query
2139 if (!port->classport_info.valid) {
2140 port->classport_info.retry_cnt++;
2141 if (port->classport_info.retry_cnt <=
2142 IB_SA_CPI_MAX_RETRY_CNT) {
2143 unsigned long delay =
2144 msecs_to_jiffies(IB_SA_CPI_RETRY_WAIT);
2146 queue_delayed_work(ib_wq, &port->ib_cpi_work, delay);
2149 spin_unlock_irqrestore(&port->classport_lock, flags);
2155 static void send_handler(struct ib_mad_agent *agent,
2156 struct ib_mad_send_wc *mad_send_wc)
2158 struct ib_sa_query *query = mad_send_wc->send_buf->context[0];
2159 unsigned long flags;
2161 if (query->callback)
2162 switch (mad_send_wc->status) {
2164 /* No callback -- already got recv */
2166 case IB_WC_RESP_TIMEOUT_ERR:
2167 query->callback(query, -ETIMEDOUT, NULL);
2169 case IB_WC_WR_FLUSH_ERR:
2170 query->callback(query, -EINTR, NULL);
2173 query->callback(query, -EIO, NULL);
2177 xa_lock_irqsave(&queries, flags);
2178 __xa_erase(&queries, query->id);
2179 xa_unlock_irqrestore(&queries, flags);
2183 ib_sa_client_put(query->client);
2184 query->release(query);
2187 static void recv_handler(struct ib_mad_agent *mad_agent,
2188 struct ib_mad_send_buf *send_buf,
2189 struct ib_mad_recv_wc *mad_recv_wc)
2191 struct ib_sa_query *query;
2196 query = send_buf->context[0];
2197 if (query->callback) {
2198 if (mad_recv_wc->wc->status == IB_WC_SUCCESS)
2199 query->callback(query,
2200 mad_recv_wc->recv_buf.mad->mad_hdr.status ?
2202 (struct ib_sa_mad *) mad_recv_wc->recv_buf.mad);
2204 query->callback(query, -EIO, NULL);
2207 ib_free_recv_mad(mad_recv_wc);
2210 static void update_sm_ah(struct work_struct *work)
2212 struct ib_sa_port *port =
2213 container_of(work, struct ib_sa_port, update_task);
2214 struct ib_sa_sm_ah *new_ah;
2215 struct ib_port_attr port_attr;
2216 struct rdma_ah_attr ah_attr;
2219 if (ib_query_port(port->agent->device, port->port_num, &port_attr)) {
2220 pr_warn("Couldn't query port\n");
2224 new_ah = kmalloc(sizeof(*new_ah), GFP_KERNEL);
2228 kref_init(&new_ah->ref);
2229 new_ah->src_path_mask = (1 << port_attr.lmc) - 1;
2231 new_ah->pkey_index = 0;
2232 if (ib_find_pkey(port->agent->device, port->port_num,
2233 IB_DEFAULT_PKEY_FULL, &new_ah->pkey_index))
2234 pr_err("Couldn't find index for default PKey\n");
2236 memset(&ah_attr, 0, sizeof(ah_attr));
2237 ah_attr.type = rdma_ah_find_type(port->agent->device,
2239 rdma_ah_set_dlid(&ah_attr, port_attr.sm_lid);
2240 rdma_ah_set_sl(&ah_attr, port_attr.sm_sl);
2241 rdma_ah_set_port_num(&ah_attr, port->port_num);
2243 grh_required = rdma_is_grh_required(port->agent->device,
2247 * The OPA sm_lid of 0xFFFF needs special handling so that it can be
2248 * differentiated from a permissive LID of 0xFFFF. We set the
2249 * grh_required flag here so the SA can program the DGID in the
2250 * address handle appropriately
2252 if (ah_attr.type == RDMA_AH_ATTR_TYPE_OPA &&
2254 port_attr.sm_lid == be16_to_cpu(IB_LID_PERMISSIVE)))
2255 rdma_ah_set_make_grd(&ah_attr, true);
2257 if (ah_attr.type == RDMA_AH_ATTR_TYPE_IB && grh_required) {
2258 rdma_ah_set_ah_flags(&ah_attr, IB_AH_GRH);
2259 rdma_ah_set_subnet_prefix(&ah_attr,
2260 cpu_to_be64(port_attr.subnet_prefix));
2261 rdma_ah_set_interface_id(&ah_attr,
2262 cpu_to_be64(IB_SA_WELL_KNOWN_GUID));
2265 new_ah->ah = rdma_create_ah(port->agent->qp->pd, &ah_attr,
2266 RDMA_CREATE_AH_SLEEPABLE);
2267 if (IS_ERR(new_ah->ah)) {
2268 pr_warn("Couldn't create new SM AH\n");
2273 spin_lock_irq(&port->ah_lock);
2275 kref_put(&port->sm_ah->ref, free_sm_ah);
2276 port->sm_ah = new_ah;
2277 spin_unlock_irq(&port->ah_lock);
2280 static void ib_sa_event(struct ib_event_handler *handler,
2281 struct ib_event *event)
2283 if (event->event == IB_EVENT_PORT_ERR ||
2284 event->event == IB_EVENT_PORT_ACTIVE ||
2285 event->event == IB_EVENT_LID_CHANGE ||
2286 event->event == IB_EVENT_PKEY_CHANGE ||
2287 event->event == IB_EVENT_SM_CHANGE ||
2288 event->event == IB_EVENT_CLIENT_REREGISTER) {
2289 unsigned long flags;
2290 struct ib_sa_device *sa_dev =
2291 container_of(handler, typeof(*sa_dev), event_handler);
2292 u8 port_num = event->element.port_num - sa_dev->start_port;
2293 struct ib_sa_port *port = &sa_dev->port[port_num];
2295 if (!rdma_cap_ib_sa(handler->device, port->port_num))
2298 spin_lock_irqsave(&port->ah_lock, flags);
2300 kref_put(&port->sm_ah->ref, free_sm_ah);
2302 spin_unlock_irqrestore(&port->ah_lock, flags);
2304 if (event->event == IB_EVENT_SM_CHANGE ||
2305 event->event == IB_EVENT_CLIENT_REREGISTER ||
2306 event->event == IB_EVENT_LID_CHANGE ||
2307 event->event == IB_EVENT_PORT_ACTIVE) {
2308 unsigned long delay =
2309 msecs_to_jiffies(IB_SA_CPI_RETRY_WAIT);
2311 spin_lock_irqsave(&port->classport_lock, flags);
2312 port->classport_info.valid = false;
2313 port->classport_info.retry_cnt = 0;
2314 spin_unlock_irqrestore(&port->classport_lock, flags);
2315 queue_delayed_work(ib_wq,
2316 &port->ib_cpi_work, delay);
2318 queue_work(ib_wq, &sa_dev->port[port_num].update_task);
2322 static int ib_sa_add_one(struct ib_device *device)
2324 struct ib_sa_device *sa_dev;
2329 s = rdma_start_port(device);
2330 e = rdma_end_port(device);
2332 sa_dev = kzalloc(struct_size(sa_dev, port, e - s + 1), GFP_KERNEL);
2336 sa_dev->start_port = s;
2337 sa_dev->end_port = e;
2339 for (i = 0; i <= e - s; ++i) {
2340 spin_lock_init(&sa_dev->port[i].ah_lock);
2341 if (!rdma_cap_ib_sa(device, i + 1))
2344 sa_dev->port[i].sm_ah = NULL;
2345 sa_dev->port[i].port_num = i + s;
2347 spin_lock_init(&sa_dev->port[i].classport_lock);
2348 sa_dev->port[i].classport_info.valid = false;
2350 sa_dev->port[i].agent =
2351 ib_register_mad_agent(device, i + s, IB_QPT_GSI,
2352 NULL, 0, send_handler,
2353 recv_handler, sa_dev, 0);
2354 if (IS_ERR(sa_dev->port[i].agent)) {
2355 ret = PTR_ERR(sa_dev->port[i].agent);
2359 INIT_WORK(&sa_dev->port[i].update_task, update_sm_ah);
2360 INIT_DELAYED_WORK(&sa_dev->port[i].ib_cpi_work,
2371 ib_set_client_data(device, &sa_client, sa_dev);
2374 * We register our event handler after everything is set up,
2375 * and then update our cached info after the event handler is
2376 * registered to avoid any problems if a port changes state
2377 * during our initialization.
2380 INIT_IB_EVENT_HANDLER(&sa_dev->event_handler, device, ib_sa_event);
2381 ib_register_event_handler(&sa_dev->event_handler);
2383 for (i = 0; i <= e - s; ++i) {
2384 if (rdma_cap_ib_sa(device, i + 1))
2385 update_sm_ah(&sa_dev->port[i].update_task);
2392 if (rdma_cap_ib_sa(device, i + 1))
2393 ib_unregister_mad_agent(sa_dev->port[i].agent);
2400 static void ib_sa_remove_one(struct ib_device *device, void *client_data)
2402 struct ib_sa_device *sa_dev = client_data;
2405 ib_unregister_event_handler(&sa_dev->event_handler);
2406 flush_workqueue(ib_wq);
2408 for (i = 0; i <= sa_dev->end_port - sa_dev->start_port; ++i) {
2409 if (rdma_cap_ib_sa(device, i + 1)) {
2410 cancel_delayed_work_sync(&sa_dev->port[i].ib_cpi_work);
2411 ib_unregister_mad_agent(sa_dev->port[i].agent);
2412 if (sa_dev->port[i].sm_ah)
2413 kref_put(&sa_dev->port[i].sm_ah->ref, free_sm_ah);
2421 int ib_sa_init(void)
2425 get_random_bytes(&tid, sizeof tid);
2427 atomic_set(&ib_nl_sa_request_seq, 0);
2429 ret = ib_register_client(&sa_client);
2431 pr_err("Couldn't register ib_sa client\n");
2437 pr_err("Couldn't initialize multicast handling\n");
2441 ib_nl_wq = alloc_ordered_workqueue("ib_nl_sa_wq", WQ_MEM_RECLAIM);
2447 INIT_DELAYED_WORK(&ib_nl_timed_work, ib_nl_request_timeout);
2454 ib_unregister_client(&sa_client);
2459 void ib_sa_cleanup(void)
2461 cancel_delayed_work(&ib_nl_timed_work);
2462 flush_workqueue(ib_nl_wq);
2463 destroy_workqueue(ib_nl_wq);
2465 ib_unregister_client(&sa_client);
2466 WARN_ON(!xa_empty(&queries));