2 * Copyright (c) 2005 Voltaire Inc. All rights reserved.
3 * Copyright (c) 2002-2005, Network Appliance, Inc. All rights reserved.
4 * Copyright (c) 1999-2005, Mellanox Technologies, Inc. All rights reserved.
5 * Copyright (c) 2005-2006 Intel Corporation. All rights reserved.
7 * This software is available to you under a choice of one of two
8 * licenses. You may choose to be licensed under the terms of the GNU
9 * General Public License (GPL) Version 2, available from the file
10 * COPYING in the main directory of this source tree, or the
11 * OpenIB.org BSD license below:
13 * Redistribution and use in source and binary forms, with or
14 * without modification, are permitted provided that the following
17 * - Redistributions of source code must retain the above
18 * copyright notice, this list of conditions and the following
21 * - Redistributions in binary form must reproduce the above
22 * copyright notice, this list of conditions and the following
23 * disclaimer in the documentation and/or other materials
24 * provided with the distribution.
26 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
27 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
28 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
29 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
30 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
31 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
32 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
36 #include <linux/completion.h>
38 #include <linux/in6.h>
39 #include <linux/mutex.h>
40 #include <linux/random.h>
41 #include <linux/igmp.h>
42 #include <linux/idr.h>
43 #include <linux/inetdevice.h>
44 #include <linux/slab.h>
45 #include <linux/module.h>
46 #include <net/route.h>
48 #include <net/net_namespace.h>
49 #include <net/netns/generic.h>
52 #include <net/ip_fib.h>
53 #include <net/ip6_route.h>
55 #include <rdma/rdma_cm.h>
56 #include <rdma/rdma_cm_ib.h>
57 #include <rdma/rdma_netlink.h>
59 #include <rdma/ib_cache.h>
60 #include <rdma/ib_cm.h>
61 #include <rdma/ib_sa.h>
62 #include <rdma/iw_cm.h>
64 #include "core_priv.h"
66 MODULE_AUTHOR("Sean Hefty");
67 MODULE_DESCRIPTION("Generic RDMA CM Agent");
68 MODULE_LICENSE("Dual BSD/GPL");
70 #define CMA_CM_RESPONSE_TIMEOUT 20
71 #define CMA_QUERY_CLASSPORT_INFO_TIMEOUT 3000
72 #define CMA_MAX_CM_RETRIES 15
73 #define CMA_CM_MRA_SETTING (IB_CM_MRA_FLAG_DELAY | 24)
74 #define CMA_IBOE_PACKET_LIFETIME 18
76 static const char * const cma_events[] = {
77 [RDMA_CM_EVENT_ADDR_RESOLVED] = "address resolved",
78 [RDMA_CM_EVENT_ADDR_ERROR] = "address error",
79 [RDMA_CM_EVENT_ROUTE_RESOLVED] = "route resolved ",
80 [RDMA_CM_EVENT_ROUTE_ERROR] = "route error",
81 [RDMA_CM_EVENT_CONNECT_REQUEST] = "connect request",
82 [RDMA_CM_EVENT_CONNECT_RESPONSE] = "connect response",
83 [RDMA_CM_EVENT_CONNECT_ERROR] = "connect error",
84 [RDMA_CM_EVENT_UNREACHABLE] = "unreachable",
85 [RDMA_CM_EVENT_REJECTED] = "rejected",
86 [RDMA_CM_EVENT_ESTABLISHED] = "established",
87 [RDMA_CM_EVENT_DISCONNECTED] = "disconnected",
88 [RDMA_CM_EVENT_DEVICE_REMOVAL] = "device removal",
89 [RDMA_CM_EVENT_MULTICAST_JOIN] = "multicast join",
90 [RDMA_CM_EVENT_MULTICAST_ERROR] = "multicast error",
91 [RDMA_CM_EVENT_ADDR_CHANGE] = "address change",
92 [RDMA_CM_EVENT_TIMEWAIT_EXIT] = "timewait exit",
95 const char *__attribute_const__ rdma_event_msg(enum rdma_cm_event_type event)
99 return (index < ARRAY_SIZE(cma_events) && cma_events[index]) ?
100 cma_events[index] : "unrecognized event";
102 EXPORT_SYMBOL(rdma_event_msg);
104 const char *__attribute_const__ rdma_reject_msg(struct rdma_cm_id *id,
107 if (rdma_ib_or_roce(id->device, id->port_num))
108 return ibcm_reject_msg(reason);
110 if (rdma_protocol_iwarp(id->device, id->port_num))
111 return iwcm_reject_msg(reason);
114 return "unrecognized transport";
116 EXPORT_SYMBOL(rdma_reject_msg);
118 bool rdma_is_consumer_reject(struct rdma_cm_id *id, int reason)
120 if (rdma_ib_or_roce(id->device, id->port_num))
121 return reason == IB_CM_REJ_CONSUMER_DEFINED;
123 if (rdma_protocol_iwarp(id->device, id->port_num))
124 return reason == -ECONNREFUSED;
129 EXPORT_SYMBOL(rdma_is_consumer_reject);
131 const void *rdma_consumer_reject_data(struct rdma_cm_id *id,
132 struct rdma_cm_event *ev, u8 *data_len)
136 if (rdma_is_consumer_reject(id, ev->status)) {
137 *data_len = ev->param.conn.private_data_len;
138 p = ev->param.conn.private_data;
145 EXPORT_SYMBOL(rdma_consumer_reject_data);
147 static void cma_add_one(struct ib_device *device);
148 static void cma_remove_one(struct ib_device *device, void *client_data);
150 static struct ib_client cma_client = {
153 .remove = cma_remove_one
156 static struct ib_sa_client sa_client;
157 static struct rdma_addr_client addr_client;
158 static LIST_HEAD(dev_list);
159 static LIST_HEAD(listen_any_list);
160 static DEFINE_MUTEX(lock);
161 static struct workqueue_struct *cma_wq;
162 static unsigned int cma_pernet_id;
171 static struct cma_pernet *cma_pernet(struct net *net)
173 return net_generic(net, cma_pernet_id);
176 static struct idr *cma_pernet_idr(struct net *net, enum rdma_port_space ps)
178 struct cma_pernet *pernet = cma_pernet(net);
182 return &pernet->tcp_ps;
184 return &pernet->udp_ps;
186 return &pernet->ipoib_ps;
188 return &pernet->ib_ps;
195 struct list_head list;
196 struct ib_device *device;
197 struct completion comp;
199 struct list_head id_list;
200 enum ib_gid_type *default_gid_type;
203 struct rdma_bind_list {
204 enum rdma_port_space ps;
205 struct hlist_head owners;
209 struct class_port_info_context {
210 struct ib_class_port_info *class_port_info;
211 struct ib_device *device;
212 struct completion done;
213 struct ib_sa_query *sa_query;
217 static int cma_ps_alloc(struct net *net, enum rdma_port_space ps,
218 struct rdma_bind_list *bind_list, int snum)
220 struct idr *idr = cma_pernet_idr(net, ps);
222 return idr_alloc(idr, bind_list, snum, snum + 1, GFP_KERNEL);
225 static struct rdma_bind_list *cma_ps_find(struct net *net,
226 enum rdma_port_space ps, int snum)
228 struct idr *idr = cma_pernet_idr(net, ps);
230 return idr_find(idr, snum);
233 static void cma_ps_remove(struct net *net, enum rdma_port_space ps, int snum)
235 struct idr *idr = cma_pernet_idr(net, ps);
237 idr_remove(idr, snum);
244 void cma_ref_dev(struct cma_device *cma_dev)
246 atomic_inc(&cma_dev->refcount);
249 struct cma_device *cma_enum_devices_by_ibdev(cma_device_filter filter,
252 struct cma_device *cma_dev;
253 struct cma_device *found_cma_dev = NULL;
257 list_for_each_entry(cma_dev, &dev_list, list)
258 if (filter(cma_dev->device, cookie)) {
259 found_cma_dev = cma_dev;
264 cma_ref_dev(found_cma_dev);
266 return found_cma_dev;
269 int cma_get_default_gid_type(struct cma_device *cma_dev,
272 if (port < rdma_start_port(cma_dev->device) ||
273 port > rdma_end_port(cma_dev->device))
276 return cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)];
279 int cma_set_default_gid_type(struct cma_device *cma_dev,
281 enum ib_gid_type default_gid_type)
283 unsigned long supported_gids;
285 if (port < rdma_start_port(cma_dev->device) ||
286 port > rdma_end_port(cma_dev->device))
289 supported_gids = roce_gid_type_mask_support(cma_dev->device, port);
291 if (!(supported_gids & 1 << default_gid_type))
294 cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)] =
300 struct ib_device *cma_get_ib_dev(struct cma_device *cma_dev)
302 return cma_dev->device;
306 * Device removal can occur at anytime, so we need extra handling to
307 * serialize notifying the user of device removal with other callbacks.
308 * We do this by disabling removal notification while a callback is in process,
309 * and reporting it after the callback completes.
311 struct rdma_id_private {
312 struct rdma_cm_id id;
314 struct rdma_bind_list *bind_list;
315 struct hlist_node node;
316 struct list_head list; /* listen_any_list or cma_device.list */
317 struct list_head listen_list; /* per device listens */
318 struct cma_device *cma_dev;
319 struct list_head mc_list;
322 enum rdma_cm_state state;
324 struct mutex qp_mutex;
326 struct completion comp;
328 struct mutex handler_mutex;
332 struct ib_sa_query *query;
348 enum ib_gid_type gid_type;
351 struct cma_multicast {
352 struct rdma_id_private *id_priv;
354 struct ib_sa_multicast *ib;
356 struct list_head list;
358 struct sockaddr_storage addr;
365 struct work_struct work;
366 struct rdma_id_private *id;
367 enum rdma_cm_state old_state;
368 enum rdma_cm_state new_state;
369 struct rdma_cm_event event;
372 struct cma_ndev_work {
373 struct work_struct work;
374 struct rdma_id_private *id;
375 struct rdma_cm_event event;
378 struct iboe_mcast_work {
379 struct work_struct work;
380 struct rdma_id_private *id;
381 struct cma_multicast *mc;
394 u8 ip_version; /* IP version: 7:4 */
396 union cma_ip_addr src_addr;
397 union cma_ip_addr dst_addr;
400 #define CMA_VERSION 0x00
402 struct cma_req_info {
403 struct ib_device *device;
405 union ib_gid local_gid;
411 static int cma_comp(struct rdma_id_private *id_priv, enum rdma_cm_state comp)
416 spin_lock_irqsave(&id_priv->lock, flags);
417 ret = (id_priv->state == comp);
418 spin_unlock_irqrestore(&id_priv->lock, flags);
422 static int cma_comp_exch(struct rdma_id_private *id_priv,
423 enum rdma_cm_state comp, enum rdma_cm_state exch)
428 spin_lock_irqsave(&id_priv->lock, flags);
429 if ((ret = (id_priv->state == comp)))
430 id_priv->state = exch;
431 spin_unlock_irqrestore(&id_priv->lock, flags);
435 static enum rdma_cm_state cma_exch(struct rdma_id_private *id_priv,
436 enum rdma_cm_state exch)
439 enum rdma_cm_state old;
441 spin_lock_irqsave(&id_priv->lock, flags);
442 old = id_priv->state;
443 id_priv->state = exch;
444 spin_unlock_irqrestore(&id_priv->lock, flags);
448 static inline u8 cma_get_ip_ver(const struct cma_hdr *hdr)
450 return hdr->ip_version >> 4;
453 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
455 hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
458 static int cma_igmp_send(struct net_device *ndev, union ib_gid *mgid, bool join)
460 struct in_device *in_dev = NULL;
464 in_dev = __in_dev_get_rtnl(ndev);
467 ip_mc_inc_group(in_dev,
468 *(__be32 *)(mgid->raw + 12));
470 ip_mc_dec_group(in_dev,
471 *(__be32 *)(mgid->raw + 12));
475 return (in_dev) ? 0 : -ENODEV;
478 static void _cma_attach_to_dev(struct rdma_id_private *id_priv,
479 struct cma_device *cma_dev)
481 cma_ref_dev(cma_dev);
482 id_priv->cma_dev = cma_dev;
483 id_priv->gid_type = 0;
484 id_priv->id.device = cma_dev->device;
485 id_priv->id.route.addr.dev_addr.transport =
486 rdma_node_get_transport(cma_dev->device->node_type);
487 list_add_tail(&id_priv->list, &cma_dev->id_list);
490 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
491 struct cma_device *cma_dev)
493 _cma_attach_to_dev(id_priv, cma_dev);
495 cma_dev->default_gid_type[id_priv->id.port_num -
496 rdma_start_port(cma_dev->device)];
499 void cma_deref_dev(struct cma_device *cma_dev)
501 if (atomic_dec_and_test(&cma_dev->refcount))
502 complete(&cma_dev->comp);
505 static inline void release_mc(struct kref *kref)
507 struct cma_multicast *mc = container_of(kref, struct cma_multicast, mcref);
509 kfree(mc->multicast.ib);
513 static void cma_release_dev(struct rdma_id_private *id_priv)
516 list_del(&id_priv->list);
517 cma_deref_dev(id_priv->cma_dev);
518 id_priv->cma_dev = NULL;
522 static inline struct sockaddr *cma_src_addr(struct rdma_id_private *id_priv)
524 return (struct sockaddr *) &id_priv->id.route.addr.src_addr;
527 static inline struct sockaddr *cma_dst_addr(struct rdma_id_private *id_priv)
529 return (struct sockaddr *) &id_priv->id.route.addr.dst_addr;
532 static inline unsigned short cma_family(struct rdma_id_private *id_priv)
534 return id_priv->id.route.addr.src_addr.ss_family;
537 static int cma_set_qkey(struct rdma_id_private *id_priv, u32 qkey)
539 struct ib_sa_mcmember_rec rec;
543 if (qkey && id_priv->qkey != qkey)
549 id_priv->qkey = qkey;
553 switch (id_priv->id.ps) {
556 id_priv->qkey = RDMA_UDP_QKEY;
559 ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid);
560 ret = ib_sa_get_mcmember_rec(id_priv->id.device,
561 id_priv->id.port_num, &rec.mgid,
564 id_priv->qkey = be32_to_cpu(rec.qkey);
572 static void cma_translate_ib(struct sockaddr_ib *sib, struct rdma_dev_addr *dev_addr)
574 dev_addr->dev_type = ARPHRD_INFINIBAND;
575 rdma_addr_set_sgid(dev_addr, (union ib_gid *) &sib->sib_addr);
576 ib_addr_set_pkey(dev_addr, ntohs(sib->sib_pkey));
579 static int cma_translate_addr(struct sockaddr *addr, struct rdma_dev_addr *dev_addr)
583 if (addr->sa_family != AF_IB) {
584 ret = rdma_translate_ip(addr, dev_addr, NULL);
586 cma_translate_ib((struct sockaddr_ib *) addr, dev_addr);
593 static inline int cma_validate_port(struct ib_device *device, u8 port,
594 enum ib_gid_type gid_type,
595 union ib_gid *gid, int dev_type,
599 struct net_device *ndev = NULL;
601 if ((dev_type == ARPHRD_INFINIBAND) && !rdma_protocol_ib(device, port))
604 if ((dev_type != ARPHRD_INFINIBAND) && rdma_protocol_ib(device, port))
607 if (dev_type == ARPHRD_ETHER && rdma_protocol_roce(device, port)) {
608 ndev = dev_get_by_index(&init_net, bound_if_index);
609 if (ndev && ndev->flags & IFF_LOOPBACK) {
610 pr_info("detected loopback device\n");
613 if (!device->get_netdev)
616 ndev = device->get_netdev(device, port);
621 gid_type = IB_GID_TYPE_IB;
624 ret = ib_find_cached_gid_by_port(device, gid, gid_type, port,
633 static int cma_acquire_dev(struct rdma_id_private *id_priv,
634 struct rdma_id_private *listen_id_priv)
636 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
637 struct cma_device *cma_dev;
638 union ib_gid gid, iboe_gid, *gidp;
642 if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
643 id_priv->id.ps == RDMA_PS_IPOIB)
647 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
650 memcpy(&gid, dev_addr->src_dev_addr +
651 rdma_addr_gid_offset(dev_addr), sizeof gid);
653 if (listen_id_priv) {
654 cma_dev = listen_id_priv->cma_dev;
655 port = listen_id_priv->id.port_num;
656 gidp = rdma_protocol_roce(cma_dev->device, port) ?
659 ret = cma_validate_port(cma_dev->device, port,
660 rdma_protocol_ib(cma_dev->device, port) ?
662 listen_id_priv->gid_type, gidp,
664 dev_addr->bound_dev_if);
666 id_priv->id.port_num = port;
671 list_for_each_entry(cma_dev, &dev_list, list) {
672 for (port = 1; port <= cma_dev->device->phys_port_cnt; ++port) {
673 if (listen_id_priv &&
674 listen_id_priv->cma_dev == cma_dev &&
675 listen_id_priv->id.port_num == port)
678 gidp = rdma_protocol_roce(cma_dev->device, port) ?
681 ret = cma_validate_port(cma_dev->device, port,
682 rdma_protocol_ib(cma_dev->device, port) ?
684 cma_dev->default_gid_type[port - 1],
685 gidp, dev_addr->dev_type,
686 dev_addr->bound_dev_if);
688 id_priv->id.port_num = port;
696 cma_attach_to_dev(id_priv, cma_dev);
703 * Select the source IB device and address to reach the destination IB address.
705 static int cma_resolve_ib_dev(struct rdma_id_private *id_priv)
707 struct cma_device *cma_dev, *cur_dev;
708 struct sockaddr_ib *addr;
709 union ib_gid gid, sgid, *dgid;
715 addr = (struct sockaddr_ib *) cma_dst_addr(id_priv);
716 dgid = (union ib_gid *) &addr->sib_addr;
717 pkey = ntohs(addr->sib_pkey);
719 list_for_each_entry(cur_dev, &dev_list, list) {
720 for (p = 1; p <= cur_dev->device->phys_port_cnt; ++p) {
721 if (!rdma_cap_af_ib(cur_dev->device, p))
724 if (ib_find_cached_pkey(cur_dev->device, p, pkey, &index))
727 for (i = 0; !ib_get_cached_gid(cur_dev->device, p, i,
730 if (!memcmp(&gid, dgid, sizeof(gid))) {
733 id_priv->id.port_num = p;
737 if (!cma_dev && (gid.global.subnet_prefix ==
738 dgid->global.subnet_prefix)) {
741 id_priv->id.port_num = p;
751 cma_attach_to_dev(id_priv, cma_dev);
752 addr = (struct sockaddr_ib *) cma_src_addr(id_priv);
753 memcpy(&addr->sib_addr, &sgid, sizeof sgid);
754 cma_translate_ib(addr, &id_priv->id.route.addr.dev_addr);
758 static void cma_deref_id(struct rdma_id_private *id_priv)
760 if (atomic_dec_and_test(&id_priv->refcount))
761 complete(&id_priv->comp);
764 struct rdma_cm_id *rdma_create_id(struct net *net,
765 rdma_cm_event_handler event_handler,
766 void *context, enum rdma_port_space ps,
767 enum ib_qp_type qp_type)
769 struct rdma_id_private *id_priv;
771 id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
773 return ERR_PTR(-ENOMEM);
775 id_priv->owner = task_pid_nr(current);
776 id_priv->state = RDMA_CM_IDLE;
777 id_priv->id.context = context;
778 id_priv->id.event_handler = event_handler;
780 id_priv->id.qp_type = qp_type;
781 spin_lock_init(&id_priv->lock);
782 mutex_init(&id_priv->qp_mutex);
783 init_completion(&id_priv->comp);
784 atomic_set(&id_priv->refcount, 1);
785 mutex_init(&id_priv->handler_mutex);
786 INIT_LIST_HEAD(&id_priv->listen_list);
787 INIT_LIST_HEAD(&id_priv->mc_list);
788 get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
789 id_priv->id.route.addr.dev_addr.net = get_net(net);
793 EXPORT_SYMBOL(rdma_create_id);
795 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
797 struct ib_qp_attr qp_attr;
798 int qp_attr_mask, ret;
800 qp_attr.qp_state = IB_QPS_INIT;
801 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
805 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
809 qp_attr.qp_state = IB_QPS_RTR;
810 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
814 qp_attr.qp_state = IB_QPS_RTS;
816 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
821 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
823 struct ib_qp_attr qp_attr;
824 int qp_attr_mask, ret;
826 qp_attr.qp_state = IB_QPS_INIT;
827 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
831 return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
834 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
835 struct ib_qp_init_attr *qp_init_attr)
837 struct rdma_id_private *id_priv;
841 id_priv = container_of(id, struct rdma_id_private, id);
842 if (id->device != pd->device)
845 qp_init_attr->port_num = id->port_num;
846 qp = ib_create_qp(pd, qp_init_attr);
850 if (id->qp_type == IB_QPT_UD)
851 ret = cma_init_ud_qp(id_priv, qp);
853 ret = cma_init_conn_qp(id_priv, qp);
858 id_priv->qp_num = qp->qp_num;
859 id_priv->srq = (qp->srq != NULL);
865 EXPORT_SYMBOL(rdma_create_qp);
867 void rdma_destroy_qp(struct rdma_cm_id *id)
869 struct rdma_id_private *id_priv;
871 id_priv = container_of(id, struct rdma_id_private, id);
872 mutex_lock(&id_priv->qp_mutex);
873 ib_destroy_qp(id_priv->id.qp);
874 id_priv->id.qp = NULL;
875 mutex_unlock(&id_priv->qp_mutex);
877 EXPORT_SYMBOL(rdma_destroy_qp);
879 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
880 struct rdma_conn_param *conn_param)
882 struct ib_qp_attr qp_attr;
883 int qp_attr_mask, ret;
886 mutex_lock(&id_priv->qp_mutex);
887 if (!id_priv->id.qp) {
892 /* Need to update QP attributes from default values. */
893 qp_attr.qp_state = IB_QPS_INIT;
894 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
898 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
902 qp_attr.qp_state = IB_QPS_RTR;
903 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
907 ret = ib_query_gid(id_priv->id.device, id_priv->id.port_num,
908 qp_attr.ah_attr.grh.sgid_index, &sgid, NULL);
912 BUG_ON(id_priv->cma_dev->device != id_priv->id.device);
915 qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
916 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
918 mutex_unlock(&id_priv->qp_mutex);
922 static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
923 struct rdma_conn_param *conn_param)
925 struct ib_qp_attr qp_attr;
926 int qp_attr_mask, ret;
928 mutex_lock(&id_priv->qp_mutex);
929 if (!id_priv->id.qp) {
934 qp_attr.qp_state = IB_QPS_RTS;
935 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
940 qp_attr.max_rd_atomic = conn_param->initiator_depth;
941 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
943 mutex_unlock(&id_priv->qp_mutex);
947 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
949 struct ib_qp_attr qp_attr;
952 mutex_lock(&id_priv->qp_mutex);
953 if (!id_priv->id.qp) {
958 qp_attr.qp_state = IB_QPS_ERR;
959 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
961 mutex_unlock(&id_priv->qp_mutex);
965 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
966 struct ib_qp_attr *qp_attr, int *qp_attr_mask)
968 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
972 if (rdma_cap_eth_ah(id_priv->id.device, id_priv->id.port_num))
975 pkey = ib_addr_get_pkey(dev_addr);
977 ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
978 pkey, &qp_attr->pkey_index);
982 qp_attr->port_num = id_priv->id.port_num;
983 *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
985 if (id_priv->id.qp_type == IB_QPT_UD) {
986 ret = cma_set_qkey(id_priv, 0);
990 qp_attr->qkey = id_priv->qkey;
991 *qp_attr_mask |= IB_QP_QKEY;
993 qp_attr->qp_access_flags = 0;
994 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
999 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
1002 struct rdma_id_private *id_priv;
1005 id_priv = container_of(id, struct rdma_id_private, id);
1006 if (rdma_cap_ib_cm(id->device, id->port_num)) {
1007 if (!id_priv->cm_id.ib || (id_priv->id.qp_type == IB_QPT_UD))
1008 ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
1010 ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
1013 if (qp_attr->qp_state == IB_QPS_RTR)
1014 qp_attr->rq_psn = id_priv->seq_num;
1015 } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
1016 if (!id_priv->cm_id.iw) {
1017 qp_attr->qp_access_flags = 0;
1018 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
1020 ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
1027 EXPORT_SYMBOL(rdma_init_qp_attr);
1029 static inline int cma_zero_addr(struct sockaddr *addr)
1031 switch (addr->sa_family) {
1033 return ipv4_is_zeronet(((struct sockaddr_in *)addr)->sin_addr.s_addr);
1035 return ipv6_addr_any(&((struct sockaddr_in6 *) addr)->sin6_addr);
1037 return ib_addr_any(&((struct sockaddr_ib *) addr)->sib_addr);
1043 static inline int cma_loopback_addr(struct sockaddr *addr)
1045 switch (addr->sa_family) {
1047 return ipv4_is_loopback(((struct sockaddr_in *) addr)->sin_addr.s_addr);
1049 return ipv6_addr_loopback(&((struct sockaddr_in6 *) addr)->sin6_addr);
1051 return ib_addr_loopback(&((struct sockaddr_ib *) addr)->sib_addr);
1057 static inline int cma_any_addr(struct sockaddr *addr)
1059 return cma_zero_addr(addr) || cma_loopback_addr(addr);
1062 static int cma_addr_cmp(struct sockaddr *src, struct sockaddr *dst)
1064 if (src->sa_family != dst->sa_family)
1067 switch (src->sa_family) {
1069 return ((struct sockaddr_in *) src)->sin_addr.s_addr !=
1070 ((struct sockaddr_in *) dst)->sin_addr.s_addr;
1072 return ipv6_addr_cmp(&((struct sockaddr_in6 *) src)->sin6_addr,
1073 &((struct sockaddr_in6 *) dst)->sin6_addr);
1075 return ib_addr_cmp(&((struct sockaddr_ib *) src)->sib_addr,
1076 &((struct sockaddr_ib *) dst)->sib_addr);
1080 static __be16 cma_port(struct sockaddr *addr)
1082 struct sockaddr_ib *sib;
1084 switch (addr->sa_family) {
1086 return ((struct sockaddr_in *) addr)->sin_port;
1088 return ((struct sockaddr_in6 *) addr)->sin6_port;
1090 sib = (struct sockaddr_ib *) addr;
1091 return htons((u16) (be64_to_cpu(sib->sib_sid) &
1092 be64_to_cpu(sib->sib_sid_mask)));
1098 static inline int cma_any_port(struct sockaddr *addr)
1100 return !cma_port(addr);
1103 static void cma_save_ib_info(struct sockaddr *src_addr,
1104 struct sockaddr *dst_addr,
1105 struct rdma_cm_id *listen_id,
1106 struct ib_sa_path_rec *path)
1108 struct sockaddr_ib *listen_ib, *ib;
1110 listen_ib = (struct sockaddr_ib *) &listen_id->route.addr.src_addr;
1112 ib = (struct sockaddr_ib *)src_addr;
1113 ib->sib_family = AF_IB;
1115 ib->sib_pkey = path->pkey;
1116 ib->sib_flowinfo = path->flow_label;
1117 memcpy(&ib->sib_addr, &path->sgid, 16);
1118 ib->sib_sid = path->service_id;
1119 ib->sib_scope_id = 0;
1121 ib->sib_pkey = listen_ib->sib_pkey;
1122 ib->sib_flowinfo = listen_ib->sib_flowinfo;
1123 ib->sib_addr = listen_ib->sib_addr;
1124 ib->sib_sid = listen_ib->sib_sid;
1125 ib->sib_scope_id = listen_ib->sib_scope_id;
1127 ib->sib_sid_mask = cpu_to_be64(0xffffffffffffffffULL);
1130 ib = (struct sockaddr_ib *)dst_addr;
1131 ib->sib_family = AF_IB;
1133 ib->sib_pkey = path->pkey;
1134 ib->sib_flowinfo = path->flow_label;
1135 memcpy(&ib->sib_addr, &path->dgid, 16);
1140 static void cma_save_ip4_info(struct sockaddr_in *src_addr,
1141 struct sockaddr_in *dst_addr,
1142 struct cma_hdr *hdr,
1146 *src_addr = (struct sockaddr_in) {
1147 .sin_family = AF_INET,
1148 .sin_addr.s_addr = hdr->dst_addr.ip4.addr,
1149 .sin_port = local_port,
1154 *dst_addr = (struct sockaddr_in) {
1155 .sin_family = AF_INET,
1156 .sin_addr.s_addr = hdr->src_addr.ip4.addr,
1157 .sin_port = hdr->port,
1162 static void cma_save_ip6_info(struct sockaddr_in6 *src_addr,
1163 struct sockaddr_in6 *dst_addr,
1164 struct cma_hdr *hdr,
1168 *src_addr = (struct sockaddr_in6) {
1169 .sin6_family = AF_INET6,
1170 .sin6_addr = hdr->dst_addr.ip6,
1171 .sin6_port = local_port,
1176 *dst_addr = (struct sockaddr_in6) {
1177 .sin6_family = AF_INET6,
1178 .sin6_addr = hdr->src_addr.ip6,
1179 .sin6_port = hdr->port,
1184 static u16 cma_port_from_service_id(__be64 service_id)
1186 return (u16)be64_to_cpu(service_id);
1189 static int cma_save_ip_info(struct sockaddr *src_addr,
1190 struct sockaddr *dst_addr,
1191 struct ib_cm_event *ib_event,
1194 struct cma_hdr *hdr;
1197 hdr = ib_event->private_data;
1198 if (hdr->cma_version != CMA_VERSION)
1201 port = htons(cma_port_from_service_id(service_id));
1203 switch (cma_get_ip_ver(hdr)) {
1205 cma_save_ip4_info((struct sockaddr_in *)src_addr,
1206 (struct sockaddr_in *)dst_addr, hdr, port);
1209 cma_save_ip6_info((struct sockaddr_in6 *)src_addr,
1210 (struct sockaddr_in6 *)dst_addr, hdr, port);
1213 return -EAFNOSUPPORT;
1219 static int cma_save_net_info(struct sockaddr *src_addr,
1220 struct sockaddr *dst_addr,
1221 struct rdma_cm_id *listen_id,
1222 struct ib_cm_event *ib_event,
1223 sa_family_t sa_family, __be64 service_id)
1225 if (sa_family == AF_IB) {
1226 if (ib_event->event == IB_CM_REQ_RECEIVED)
1227 cma_save_ib_info(src_addr, dst_addr, listen_id,
1228 ib_event->param.req_rcvd.primary_path);
1229 else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED)
1230 cma_save_ib_info(src_addr, dst_addr, listen_id, NULL);
1234 return cma_save_ip_info(src_addr, dst_addr, ib_event, service_id);
1237 static int cma_save_req_info(const struct ib_cm_event *ib_event,
1238 struct cma_req_info *req)
1240 const struct ib_cm_req_event_param *req_param =
1241 &ib_event->param.req_rcvd;
1242 const struct ib_cm_sidr_req_event_param *sidr_param =
1243 &ib_event->param.sidr_req_rcvd;
1245 switch (ib_event->event) {
1246 case IB_CM_REQ_RECEIVED:
1247 req->device = req_param->listen_id->device;
1248 req->port = req_param->port;
1249 memcpy(&req->local_gid, &req_param->primary_path->sgid,
1250 sizeof(req->local_gid));
1251 req->has_gid = true;
1252 req->service_id = req_param->primary_path->service_id;
1253 req->pkey = be16_to_cpu(req_param->primary_path->pkey);
1254 if (req->pkey != req_param->bth_pkey)
1255 pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and primary path P_Key (0x%x)\n"
1256 "RDMA CMA: in the future this may cause the request to be dropped\n",
1257 req_param->bth_pkey, req->pkey);
1259 case IB_CM_SIDR_REQ_RECEIVED:
1260 req->device = sidr_param->listen_id->device;
1261 req->port = sidr_param->port;
1262 req->has_gid = false;
1263 req->service_id = sidr_param->service_id;
1264 req->pkey = sidr_param->pkey;
1265 if (req->pkey != sidr_param->bth_pkey)
1266 pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and SIDR request payload P_Key (0x%x)\n"
1267 "RDMA CMA: in the future this may cause the request to be dropped\n",
1268 sidr_param->bth_pkey, req->pkey);
1277 static bool validate_ipv4_net_dev(struct net_device *net_dev,
1278 const struct sockaddr_in *dst_addr,
1279 const struct sockaddr_in *src_addr)
1281 __be32 daddr = dst_addr->sin_addr.s_addr,
1282 saddr = src_addr->sin_addr.s_addr;
1283 struct fib_result res;
1288 if (ipv4_is_multicast(saddr) || ipv4_is_lbcast(saddr) ||
1289 ipv4_is_lbcast(daddr) || ipv4_is_zeronet(saddr) ||
1290 ipv4_is_zeronet(daddr) || ipv4_is_loopback(daddr) ||
1291 ipv4_is_loopback(saddr))
1294 memset(&fl4, 0, sizeof(fl4));
1295 fl4.flowi4_iif = net_dev->ifindex;
1300 err = fib_lookup(dev_net(net_dev), &fl4, &res, 0);
1301 ret = err == 0 && FIB_RES_DEV(res) == net_dev;
1307 static bool validate_ipv6_net_dev(struct net_device *net_dev,
1308 const struct sockaddr_in6 *dst_addr,
1309 const struct sockaddr_in6 *src_addr)
1311 #if IS_ENABLED(CONFIG_IPV6)
1312 const int strict = ipv6_addr_type(&dst_addr->sin6_addr) &
1313 IPV6_ADDR_LINKLOCAL;
1314 struct rt6_info *rt = rt6_lookup(dev_net(net_dev), &dst_addr->sin6_addr,
1315 &src_addr->sin6_addr, net_dev->ifindex,
1322 ret = rt->rt6i_idev->dev == net_dev;
1331 static bool validate_net_dev(struct net_device *net_dev,
1332 const struct sockaddr *daddr,
1333 const struct sockaddr *saddr)
1335 const struct sockaddr_in *daddr4 = (const struct sockaddr_in *)daddr;
1336 const struct sockaddr_in *saddr4 = (const struct sockaddr_in *)saddr;
1337 const struct sockaddr_in6 *daddr6 = (const struct sockaddr_in6 *)daddr;
1338 const struct sockaddr_in6 *saddr6 = (const struct sockaddr_in6 *)saddr;
1340 switch (daddr->sa_family) {
1342 return saddr->sa_family == AF_INET &&
1343 validate_ipv4_net_dev(net_dev, daddr4, saddr4);
1346 return saddr->sa_family == AF_INET6 &&
1347 validate_ipv6_net_dev(net_dev, daddr6, saddr6);
1354 static struct net_device *cma_get_net_dev(struct ib_cm_event *ib_event,
1355 const struct cma_req_info *req)
1357 struct sockaddr_storage listen_addr_storage, src_addr_storage;
1358 struct sockaddr *listen_addr = (struct sockaddr *)&listen_addr_storage,
1359 *src_addr = (struct sockaddr *)&src_addr_storage;
1360 struct net_device *net_dev;
1361 const union ib_gid *gid = req->has_gid ? &req->local_gid : NULL;
1364 err = cma_save_ip_info(listen_addr, src_addr, ib_event,
1367 return ERR_PTR(err);
1369 net_dev = ib_get_net_dev_by_params(req->device, req->port, req->pkey,
1372 return ERR_PTR(-ENODEV);
1374 if (!validate_net_dev(net_dev, listen_addr, src_addr)) {
1376 return ERR_PTR(-EHOSTUNREACH);
1382 static enum rdma_port_space rdma_ps_from_service_id(__be64 service_id)
1384 return (be64_to_cpu(service_id) >> 16) & 0xffff;
1387 static bool cma_match_private_data(struct rdma_id_private *id_priv,
1388 const struct cma_hdr *hdr)
1390 struct sockaddr *addr = cma_src_addr(id_priv);
1392 struct in6_addr ip6_addr;
1394 if (cma_any_addr(addr) && !id_priv->afonly)
1397 switch (addr->sa_family) {
1399 ip4_addr = ((struct sockaddr_in *)addr)->sin_addr.s_addr;
1400 if (cma_get_ip_ver(hdr) != 4)
1402 if (!cma_any_addr(addr) &&
1403 hdr->dst_addr.ip4.addr != ip4_addr)
1407 ip6_addr = ((struct sockaddr_in6 *)addr)->sin6_addr;
1408 if (cma_get_ip_ver(hdr) != 6)
1410 if (!cma_any_addr(addr) &&
1411 memcmp(&hdr->dst_addr.ip6, &ip6_addr, sizeof(ip6_addr)))
1423 static bool cma_protocol_roce_dev_port(struct ib_device *device, int port_num)
1425 enum rdma_link_layer ll = rdma_port_get_link_layer(device, port_num);
1426 enum rdma_transport_type transport =
1427 rdma_node_get_transport(device->node_type);
1429 return ll == IB_LINK_LAYER_ETHERNET && transport == RDMA_TRANSPORT_IB;
1432 static bool cma_protocol_roce(const struct rdma_cm_id *id)
1434 struct ib_device *device = id->device;
1435 const int port_num = id->port_num ?: rdma_start_port(device);
1437 return cma_protocol_roce_dev_port(device, port_num);
1440 static bool cma_match_net_dev(const struct rdma_cm_id *id,
1441 const struct net_device *net_dev,
1444 const struct rdma_addr *addr = &id->route.addr;
1447 /* This request is an AF_IB request or a RoCE request */
1448 return (!id->port_num || id->port_num == port_num) &&
1449 (addr->src_addr.ss_family == AF_IB ||
1450 cma_protocol_roce_dev_port(id->device, port_num));
1452 return !addr->dev_addr.bound_dev_if ||
1453 (net_eq(dev_net(net_dev), addr->dev_addr.net) &&
1454 addr->dev_addr.bound_dev_if == net_dev->ifindex);
1457 static struct rdma_id_private *cma_find_listener(
1458 const struct rdma_bind_list *bind_list,
1459 const struct ib_cm_id *cm_id,
1460 const struct ib_cm_event *ib_event,
1461 const struct cma_req_info *req,
1462 const struct net_device *net_dev)
1464 struct rdma_id_private *id_priv, *id_priv_dev;
1467 return ERR_PTR(-EINVAL);
1469 hlist_for_each_entry(id_priv, &bind_list->owners, node) {
1470 if (cma_match_private_data(id_priv, ib_event->private_data)) {
1471 if (id_priv->id.device == cm_id->device &&
1472 cma_match_net_dev(&id_priv->id, net_dev, req->port))
1474 list_for_each_entry(id_priv_dev,
1475 &id_priv->listen_list,
1477 if (id_priv_dev->id.device == cm_id->device &&
1478 cma_match_net_dev(&id_priv_dev->id, net_dev, req->port))
1484 return ERR_PTR(-EINVAL);
1487 static struct rdma_id_private *cma_id_from_event(struct ib_cm_id *cm_id,
1488 struct ib_cm_event *ib_event,
1489 struct net_device **net_dev)
1491 struct cma_req_info req;
1492 struct rdma_bind_list *bind_list;
1493 struct rdma_id_private *id_priv;
1496 err = cma_save_req_info(ib_event, &req);
1498 return ERR_PTR(err);
1500 *net_dev = cma_get_net_dev(ib_event, &req);
1501 if (IS_ERR(*net_dev)) {
1502 if (PTR_ERR(*net_dev) == -EAFNOSUPPORT) {
1503 /* Assuming the protocol is AF_IB */
1505 } else if (cma_protocol_roce_dev_port(req.device, req.port)) {
1506 /* TODO find the net dev matching the request parameters
1507 * through the RoCE GID table */
1510 return ERR_CAST(*net_dev);
1514 bind_list = cma_ps_find(*net_dev ? dev_net(*net_dev) : &init_net,
1515 rdma_ps_from_service_id(req.service_id),
1516 cma_port_from_service_id(req.service_id));
1517 id_priv = cma_find_listener(bind_list, cm_id, ib_event, &req, *net_dev);
1518 if (IS_ERR(id_priv) && *net_dev) {
1526 static inline int cma_user_data_offset(struct rdma_id_private *id_priv)
1528 return cma_family(id_priv) == AF_IB ? 0 : sizeof(struct cma_hdr);
1531 static void cma_cancel_route(struct rdma_id_private *id_priv)
1533 if (rdma_cap_ib_sa(id_priv->id.device, id_priv->id.port_num)) {
1535 ib_sa_cancel_query(id_priv->query_id, id_priv->query);
1539 static void cma_cancel_listens(struct rdma_id_private *id_priv)
1541 struct rdma_id_private *dev_id_priv;
1544 * Remove from listen_any_list to prevent added devices from spawning
1545 * additional listen requests.
1548 list_del(&id_priv->list);
1550 while (!list_empty(&id_priv->listen_list)) {
1551 dev_id_priv = list_entry(id_priv->listen_list.next,
1552 struct rdma_id_private, listen_list);
1553 /* sync with device removal to avoid duplicate destruction */
1554 list_del_init(&dev_id_priv->list);
1555 list_del(&dev_id_priv->listen_list);
1556 mutex_unlock(&lock);
1558 rdma_destroy_id(&dev_id_priv->id);
1561 mutex_unlock(&lock);
1564 static void cma_cancel_operation(struct rdma_id_private *id_priv,
1565 enum rdma_cm_state state)
1568 case RDMA_CM_ADDR_QUERY:
1569 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
1571 case RDMA_CM_ROUTE_QUERY:
1572 cma_cancel_route(id_priv);
1574 case RDMA_CM_LISTEN:
1575 if (cma_any_addr(cma_src_addr(id_priv)) && !id_priv->cma_dev)
1576 cma_cancel_listens(id_priv);
1583 static void cma_release_port(struct rdma_id_private *id_priv)
1585 struct rdma_bind_list *bind_list = id_priv->bind_list;
1586 struct net *net = id_priv->id.route.addr.dev_addr.net;
1592 hlist_del(&id_priv->node);
1593 if (hlist_empty(&bind_list->owners)) {
1594 cma_ps_remove(net, bind_list->ps, bind_list->port);
1597 mutex_unlock(&lock);
1600 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
1602 struct cma_multicast *mc;
1604 while (!list_empty(&id_priv->mc_list)) {
1605 mc = container_of(id_priv->mc_list.next,
1606 struct cma_multicast, list);
1607 list_del(&mc->list);
1608 if (rdma_cap_ib_mcast(id_priv->cma_dev->device,
1609 id_priv->id.port_num)) {
1610 ib_sa_free_multicast(mc->multicast.ib);
1613 if (mc->igmp_joined) {
1614 struct rdma_dev_addr *dev_addr =
1615 &id_priv->id.route.addr.dev_addr;
1616 struct net_device *ndev = NULL;
1618 if (dev_addr->bound_dev_if)
1619 ndev = dev_get_by_index(&init_net,
1620 dev_addr->bound_dev_if);
1623 &mc->multicast.ib->rec.mgid,
1628 kref_put(&mc->mcref, release_mc);
1633 void rdma_destroy_id(struct rdma_cm_id *id)
1635 struct rdma_id_private *id_priv;
1636 enum rdma_cm_state state;
1638 id_priv = container_of(id, struct rdma_id_private, id);
1639 state = cma_exch(id_priv, RDMA_CM_DESTROYING);
1640 cma_cancel_operation(id_priv, state);
1643 * Wait for any active callback to finish. New callbacks will find
1644 * the id_priv state set to destroying and abort.
1646 mutex_lock(&id_priv->handler_mutex);
1647 mutex_unlock(&id_priv->handler_mutex);
1649 if (id_priv->cma_dev) {
1650 if (rdma_cap_ib_cm(id_priv->id.device, 1)) {
1651 if (id_priv->cm_id.ib)
1652 ib_destroy_cm_id(id_priv->cm_id.ib);
1653 } else if (rdma_cap_iw_cm(id_priv->id.device, 1)) {
1654 if (id_priv->cm_id.iw)
1655 iw_destroy_cm_id(id_priv->cm_id.iw);
1657 cma_leave_mc_groups(id_priv);
1658 cma_release_dev(id_priv);
1661 cma_release_port(id_priv);
1662 cma_deref_id(id_priv);
1663 wait_for_completion(&id_priv->comp);
1665 if (id_priv->internal_id)
1666 cma_deref_id(id_priv->id.context);
1668 kfree(id_priv->id.route.path_rec);
1669 put_net(id_priv->id.route.addr.dev_addr.net);
1672 EXPORT_SYMBOL(rdma_destroy_id);
1674 static int cma_rep_recv(struct rdma_id_private *id_priv)
1678 ret = cma_modify_qp_rtr(id_priv, NULL);
1682 ret = cma_modify_qp_rts(id_priv, NULL);
1686 ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
1692 cma_modify_qp_err(id_priv);
1693 ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
1698 static void cma_set_rep_event_data(struct rdma_cm_event *event,
1699 struct ib_cm_rep_event_param *rep_data,
1702 event->param.conn.private_data = private_data;
1703 event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
1704 event->param.conn.responder_resources = rep_data->responder_resources;
1705 event->param.conn.initiator_depth = rep_data->initiator_depth;
1706 event->param.conn.flow_control = rep_data->flow_control;
1707 event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
1708 event->param.conn.srq = rep_data->srq;
1709 event->param.conn.qp_num = rep_data->remote_qpn;
1712 static int cma_ib_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1714 struct rdma_id_private *id_priv = cm_id->context;
1715 struct rdma_cm_event event;
1718 mutex_lock(&id_priv->handler_mutex);
1719 if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
1720 id_priv->state != RDMA_CM_CONNECT) ||
1721 (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
1722 id_priv->state != RDMA_CM_DISCONNECT))
1725 memset(&event, 0, sizeof event);
1726 switch (ib_event->event) {
1727 case IB_CM_REQ_ERROR:
1728 case IB_CM_REP_ERROR:
1729 event.event = RDMA_CM_EVENT_UNREACHABLE;
1730 event.status = -ETIMEDOUT;
1732 case IB_CM_REP_RECEIVED:
1733 if (id_priv->id.qp) {
1734 event.status = cma_rep_recv(id_priv);
1735 event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
1736 RDMA_CM_EVENT_ESTABLISHED;
1738 event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
1740 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
1741 ib_event->private_data);
1743 case IB_CM_RTU_RECEIVED:
1744 case IB_CM_USER_ESTABLISHED:
1745 event.event = RDMA_CM_EVENT_ESTABLISHED;
1747 case IB_CM_DREQ_ERROR:
1748 event.status = -ETIMEDOUT; /* fall through */
1749 case IB_CM_DREQ_RECEIVED:
1750 case IB_CM_DREP_RECEIVED:
1751 if (!cma_comp_exch(id_priv, RDMA_CM_CONNECT,
1752 RDMA_CM_DISCONNECT))
1754 event.event = RDMA_CM_EVENT_DISCONNECTED;
1756 case IB_CM_TIMEWAIT_EXIT:
1757 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
1759 case IB_CM_MRA_RECEIVED:
1762 case IB_CM_REJ_RECEIVED:
1763 cma_modify_qp_err(id_priv);
1764 event.status = ib_event->param.rej_rcvd.reason;
1765 event.event = RDMA_CM_EVENT_REJECTED;
1766 event.param.conn.private_data = ib_event->private_data;
1767 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
1770 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
1775 ret = id_priv->id.event_handler(&id_priv->id, &event);
1777 /* Destroy the CM ID by returning a non-zero value. */
1778 id_priv->cm_id.ib = NULL;
1779 cma_exch(id_priv, RDMA_CM_DESTROYING);
1780 mutex_unlock(&id_priv->handler_mutex);
1781 rdma_destroy_id(&id_priv->id);
1785 mutex_unlock(&id_priv->handler_mutex);
1789 static struct rdma_id_private *cma_new_conn_id(struct rdma_cm_id *listen_id,
1790 struct ib_cm_event *ib_event,
1791 struct net_device *net_dev)
1793 struct rdma_id_private *id_priv;
1794 struct rdma_cm_id *id;
1795 struct rdma_route *rt;
1796 const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
1797 const __be64 service_id =
1798 ib_event->param.req_rcvd.primary_path->service_id;
1801 id = rdma_create_id(listen_id->route.addr.dev_addr.net,
1802 listen_id->event_handler, listen_id->context,
1803 listen_id->ps, ib_event->param.req_rcvd.qp_type);
1807 id_priv = container_of(id, struct rdma_id_private, id);
1808 if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
1809 (struct sockaddr *)&id->route.addr.dst_addr,
1810 listen_id, ib_event, ss_family, service_id))
1814 rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
1815 rt->path_rec = kmalloc(sizeof *rt->path_rec * rt->num_paths,
1820 rt->path_rec[0] = *ib_event->param.req_rcvd.primary_path;
1821 if (rt->num_paths == 2)
1822 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
1825 ret = rdma_copy_addr(&rt->addr.dev_addr, net_dev, NULL);
1829 if (!cma_protocol_roce(listen_id) &&
1830 cma_any_addr(cma_src_addr(id_priv))) {
1831 rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
1832 rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
1833 ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
1834 } else if (!cma_any_addr(cma_src_addr(id_priv))) {
1835 ret = cma_translate_addr(cma_src_addr(id_priv), &rt->addr.dev_addr);
1840 rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
1842 id_priv->state = RDMA_CM_CONNECT;
1846 rdma_destroy_id(id);
1850 static struct rdma_id_private *cma_new_udp_id(struct rdma_cm_id *listen_id,
1851 struct ib_cm_event *ib_event,
1852 struct net_device *net_dev)
1854 struct rdma_id_private *id_priv;
1855 struct rdma_cm_id *id;
1856 const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
1857 struct net *net = listen_id->route.addr.dev_addr.net;
1860 id = rdma_create_id(net, listen_id->event_handler, listen_id->context,
1861 listen_id->ps, IB_QPT_UD);
1865 id_priv = container_of(id, struct rdma_id_private, id);
1866 if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
1867 (struct sockaddr *)&id->route.addr.dst_addr,
1868 listen_id, ib_event, ss_family,
1869 ib_event->param.sidr_req_rcvd.service_id))
1873 ret = rdma_copy_addr(&id->route.addr.dev_addr, net_dev, NULL);
1877 if (!cma_any_addr(cma_src_addr(id_priv))) {
1878 ret = cma_translate_addr(cma_src_addr(id_priv),
1879 &id->route.addr.dev_addr);
1885 id_priv->state = RDMA_CM_CONNECT;
1888 rdma_destroy_id(id);
1892 static void cma_set_req_event_data(struct rdma_cm_event *event,
1893 struct ib_cm_req_event_param *req_data,
1894 void *private_data, int offset)
1896 event->param.conn.private_data = private_data + offset;
1897 event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
1898 event->param.conn.responder_resources = req_data->responder_resources;
1899 event->param.conn.initiator_depth = req_data->initiator_depth;
1900 event->param.conn.flow_control = req_data->flow_control;
1901 event->param.conn.retry_count = req_data->retry_count;
1902 event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
1903 event->param.conn.srq = req_data->srq;
1904 event->param.conn.qp_num = req_data->remote_qpn;
1907 static int cma_check_req_qp_type(struct rdma_cm_id *id, struct ib_cm_event *ib_event)
1909 return (((ib_event->event == IB_CM_REQ_RECEIVED) &&
1910 (ib_event->param.req_rcvd.qp_type == id->qp_type)) ||
1911 ((ib_event->event == IB_CM_SIDR_REQ_RECEIVED) &&
1912 (id->qp_type == IB_QPT_UD)) ||
1916 static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1918 struct rdma_id_private *listen_id, *conn_id = NULL;
1919 struct rdma_cm_event event;
1920 struct net_device *net_dev;
1923 listen_id = cma_id_from_event(cm_id, ib_event, &net_dev);
1924 if (IS_ERR(listen_id))
1925 return PTR_ERR(listen_id);
1927 if (!cma_check_req_qp_type(&listen_id->id, ib_event)) {
1932 mutex_lock(&listen_id->handler_mutex);
1933 if (listen_id->state != RDMA_CM_LISTEN) {
1934 ret = -ECONNABORTED;
1938 memset(&event, 0, sizeof event);
1939 offset = cma_user_data_offset(listen_id);
1940 event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1941 if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) {
1942 conn_id = cma_new_udp_id(&listen_id->id, ib_event, net_dev);
1943 event.param.ud.private_data = ib_event->private_data + offset;
1944 event.param.ud.private_data_len =
1945 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
1947 conn_id = cma_new_conn_id(&listen_id->id, ib_event, net_dev);
1948 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
1949 ib_event->private_data, offset);
1956 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1957 ret = cma_acquire_dev(conn_id, listen_id);
1961 conn_id->cm_id.ib = cm_id;
1962 cm_id->context = conn_id;
1963 cm_id->cm_handler = cma_ib_handler;
1966 * Protect against the user destroying conn_id from another thread
1967 * until we're done accessing it.
1969 atomic_inc(&conn_id->refcount);
1970 ret = conn_id->id.event_handler(&conn_id->id, &event);
1974 * Acquire mutex to prevent user executing rdma_destroy_id()
1975 * while we're accessing the cm_id.
1978 if (cma_comp(conn_id, RDMA_CM_CONNECT) &&
1979 (conn_id->id.qp_type != IB_QPT_UD))
1980 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
1981 mutex_unlock(&lock);
1982 mutex_unlock(&conn_id->handler_mutex);
1983 mutex_unlock(&listen_id->handler_mutex);
1984 cma_deref_id(conn_id);
1990 cma_deref_id(conn_id);
1991 /* Destroy the CM ID by returning a non-zero value. */
1992 conn_id->cm_id.ib = NULL;
1994 cma_exch(conn_id, RDMA_CM_DESTROYING);
1995 mutex_unlock(&conn_id->handler_mutex);
1997 mutex_unlock(&listen_id->handler_mutex);
1999 rdma_destroy_id(&conn_id->id);
2008 __be64 rdma_get_service_id(struct rdma_cm_id *id, struct sockaddr *addr)
2010 if (addr->sa_family == AF_IB)
2011 return ((struct sockaddr_ib *) addr)->sib_sid;
2013 return cpu_to_be64(((u64)id->ps << 16) + be16_to_cpu(cma_port(addr)));
2015 EXPORT_SYMBOL(rdma_get_service_id);
2017 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
2019 struct rdma_id_private *id_priv = iw_id->context;
2020 struct rdma_cm_event event;
2022 struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
2023 struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
2025 mutex_lock(&id_priv->handler_mutex);
2026 if (id_priv->state != RDMA_CM_CONNECT)
2029 memset(&event, 0, sizeof event);
2030 switch (iw_event->event) {
2031 case IW_CM_EVENT_CLOSE:
2032 event.event = RDMA_CM_EVENT_DISCONNECTED;
2034 case IW_CM_EVENT_CONNECT_REPLY:
2035 memcpy(cma_src_addr(id_priv), laddr,
2036 rdma_addr_size(laddr));
2037 memcpy(cma_dst_addr(id_priv), raddr,
2038 rdma_addr_size(raddr));
2039 switch (iw_event->status) {
2041 event.event = RDMA_CM_EVENT_ESTABLISHED;
2042 event.param.conn.initiator_depth = iw_event->ird;
2043 event.param.conn.responder_resources = iw_event->ord;
2047 event.event = RDMA_CM_EVENT_REJECTED;
2050 event.event = RDMA_CM_EVENT_UNREACHABLE;
2053 event.event = RDMA_CM_EVENT_CONNECT_ERROR;
2057 case IW_CM_EVENT_ESTABLISHED:
2058 event.event = RDMA_CM_EVENT_ESTABLISHED;
2059 event.param.conn.initiator_depth = iw_event->ird;
2060 event.param.conn.responder_resources = iw_event->ord;
2066 event.status = iw_event->status;
2067 event.param.conn.private_data = iw_event->private_data;
2068 event.param.conn.private_data_len = iw_event->private_data_len;
2069 ret = id_priv->id.event_handler(&id_priv->id, &event);
2071 /* Destroy the CM ID by returning a non-zero value. */
2072 id_priv->cm_id.iw = NULL;
2073 cma_exch(id_priv, RDMA_CM_DESTROYING);
2074 mutex_unlock(&id_priv->handler_mutex);
2075 rdma_destroy_id(&id_priv->id);
2080 mutex_unlock(&id_priv->handler_mutex);
2084 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
2085 struct iw_cm_event *iw_event)
2087 struct rdma_cm_id *new_cm_id;
2088 struct rdma_id_private *listen_id, *conn_id;
2089 struct rdma_cm_event event;
2090 int ret = -ECONNABORTED;
2091 struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
2092 struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
2094 listen_id = cm_id->context;
2096 mutex_lock(&listen_id->handler_mutex);
2097 if (listen_id->state != RDMA_CM_LISTEN)
2100 /* Create a new RDMA id for the new IW CM ID */
2101 new_cm_id = rdma_create_id(listen_id->id.route.addr.dev_addr.net,
2102 listen_id->id.event_handler,
2103 listen_id->id.context,
2104 RDMA_PS_TCP, IB_QPT_RC);
2105 if (IS_ERR(new_cm_id)) {
2109 conn_id = container_of(new_cm_id, struct rdma_id_private, id);
2110 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
2111 conn_id->state = RDMA_CM_CONNECT;
2113 ret = rdma_translate_ip(laddr, &conn_id->id.route.addr.dev_addr, NULL);
2115 mutex_unlock(&conn_id->handler_mutex);
2116 rdma_destroy_id(new_cm_id);
2120 ret = cma_acquire_dev(conn_id, listen_id);
2122 mutex_unlock(&conn_id->handler_mutex);
2123 rdma_destroy_id(new_cm_id);
2127 conn_id->cm_id.iw = cm_id;
2128 cm_id->context = conn_id;
2129 cm_id->cm_handler = cma_iw_handler;
2131 memcpy(cma_src_addr(conn_id), laddr, rdma_addr_size(laddr));
2132 memcpy(cma_dst_addr(conn_id), raddr, rdma_addr_size(raddr));
2134 memset(&event, 0, sizeof event);
2135 event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
2136 event.param.conn.private_data = iw_event->private_data;
2137 event.param.conn.private_data_len = iw_event->private_data_len;
2138 event.param.conn.initiator_depth = iw_event->ird;
2139 event.param.conn.responder_resources = iw_event->ord;
2142 * Protect against the user destroying conn_id from another thread
2143 * until we're done accessing it.
2145 atomic_inc(&conn_id->refcount);
2146 ret = conn_id->id.event_handler(&conn_id->id, &event);
2148 /* User wants to destroy the CM ID */
2149 conn_id->cm_id.iw = NULL;
2150 cma_exch(conn_id, RDMA_CM_DESTROYING);
2151 mutex_unlock(&conn_id->handler_mutex);
2152 cma_deref_id(conn_id);
2153 rdma_destroy_id(&conn_id->id);
2157 mutex_unlock(&conn_id->handler_mutex);
2158 cma_deref_id(conn_id);
2161 mutex_unlock(&listen_id->handler_mutex);
2165 static int cma_ib_listen(struct rdma_id_private *id_priv)
2167 struct sockaddr *addr;
2168 struct ib_cm_id *id;
2171 addr = cma_src_addr(id_priv);
2172 svc_id = rdma_get_service_id(&id_priv->id, addr);
2173 id = ib_cm_insert_listen(id_priv->id.device, cma_req_handler, svc_id);
2176 id_priv->cm_id.ib = id;
2181 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
2184 struct iw_cm_id *id;
2186 id = iw_create_cm_id(id_priv->id.device,
2187 iw_conn_req_handler,
2192 id->tos = id_priv->tos;
2193 id_priv->cm_id.iw = id;
2195 memcpy(&id_priv->cm_id.iw->local_addr, cma_src_addr(id_priv),
2196 rdma_addr_size(cma_src_addr(id_priv)));
2198 ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
2201 iw_destroy_cm_id(id_priv->cm_id.iw);
2202 id_priv->cm_id.iw = NULL;
2208 static int cma_listen_handler(struct rdma_cm_id *id,
2209 struct rdma_cm_event *event)
2211 struct rdma_id_private *id_priv = id->context;
2213 id->context = id_priv->id.context;
2214 id->event_handler = id_priv->id.event_handler;
2215 return id_priv->id.event_handler(id, event);
2218 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
2219 struct cma_device *cma_dev)
2221 struct rdma_id_private *dev_id_priv;
2222 struct rdma_cm_id *id;
2223 struct net *net = id_priv->id.route.addr.dev_addr.net;
2226 if (cma_family(id_priv) == AF_IB && !rdma_cap_ib_cm(cma_dev->device, 1))
2229 id = rdma_create_id(net, cma_listen_handler, id_priv, id_priv->id.ps,
2230 id_priv->id.qp_type);
2234 dev_id_priv = container_of(id, struct rdma_id_private, id);
2236 dev_id_priv->state = RDMA_CM_ADDR_BOUND;
2237 memcpy(cma_src_addr(dev_id_priv), cma_src_addr(id_priv),
2238 rdma_addr_size(cma_src_addr(id_priv)));
2240 _cma_attach_to_dev(dev_id_priv, cma_dev);
2241 list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
2242 atomic_inc(&id_priv->refcount);
2243 dev_id_priv->internal_id = 1;
2244 dev_id_priv->afonly = id_priv->afonly;
2246 ret = rdma_listen(id, id_priv->backlog);
2248 pr_warn("RDMA CMA: cma_listen_on_dev, error %d, listening on device %s\n",
2249 ret, cma_dev->device->name);
2252 static void cma_listen_on_all(struct rdma_id_private *id_priv)
2254 struct cma_device *cma_dev;
2257 list_add_tail(&id_priv->list, &listen_any_list);
2258 list_for_each_entry(cma_dev, &dev_list, list)
2259 cma_listen_on_dev(id_priv, cma_dev);
2260 mutex_unlock(&lock);
2263 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
2265 struct rdma_id_private *id_priv;
2267 id_priv = container_of(id, struct rdma_id_private, id);
2268 id_priv->tos = (u8) tos;
2270 EXPORT_SYMBOL(rdma_set_service_type);
2272 static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec,
2275 struct cma_work *work = context;
2276 struct rdma_route *route;
2278 route = &work->id->id.route;
2281 route->num_paths = 1;
2282 *route->path_rec = *path_rec;
2284 work->old_state = RDMA_CM_ROUTE_QUERY;
2285 work->new_state = RDMA_CM_ADDR_RESOLVED;
2286 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
2287 work->event.status = status;
2290 queue_work(cma_wq, &work->work);
2293 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms,
2294 struct cma_work *work)
2296 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2297 struct ib_sa_path_rec path_rec;
2298 ib_sa_comp_mask comp_mask;
2299 struct sockaddr_in6 *sin6;
2300 struct sockaddr_ib *sib;
2302 memset(&path_rec, 0, sizeof path_rec);
2303 rdma_addr_get_sgid(dev_addr, &path_rec.sgid);
2304 rdma_addr_get_dgid(dev_addr, &path_rec.dgid);
2305 path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
2306 path_rec.numb_path = 1;
2307 path_rec.reversible = 1;
2308 path_rec.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
2310 comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
2311 IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
2312 IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
2314 switch (cma_family(id_priv)) {
2316 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
2317 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
2320 sin6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
2321 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
2322 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2325 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
2326 path_rec.traffic_class = (u8) (be32_to_cpu(sib->sib_flowinfo) >> 20);
2327 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2331 id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
2332 id_priv->id.port_num, &path_rec,
2333 comp_mask, timeout_ms,
2334 GFP_KERNEL, cma_query_handler,
2335 work, &id_priv->query);
2337 return (id_priv->query_id < 0) ? id_priv->query_id : 0;
2340 static void cma_work_handler(struct work_struct *_work)
2342 struct cma_work *work = container_of(_work, struct cma_work, work);
2343 struct rdma_id_private *id_priv = work->id;
2346 mutex_lock(&id_priv->handler_mutex);
2347 if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
2350 if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
2351 cma_exch(id_priv, RDMA_CM_DESTROYING);
2355 mutex_unlock(&id_priv->handler_mutex);
2356 cma_deref_id(id_priv);
2358 rdma_destroy_id(&id_priv->id);
2362 static void cma_ndev_work_handler(struct work_struct *_work)
2364 struct cma_ndev_work *work = container_of(_work, struct cma_ndev_work, work);
2365 struct rdma_id_private *id_priv = work->id;
2368 mutex_lock(&id_priv->handler_mutex);
2369 if (id_priv->state == RDMA_CM_DESTROYING ||
2370 id_priv->state == RDMA_CM_DEVICE_REMOVAL)
2373 if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
2374 cma_exch(id_priv, RDMA_CM_DESTROYING);
2379 mutex_unlock(&id_priv->handler_mutex);
2380 cma_deref_id(id_priv);
2382 rdma_destroy_id(&id_priv->id);
2386 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms)
2388 struct rdma_route *route = &id_priv->id.route;
2389 struct cma_work *work;
2392 work = kzalloc(sizeof *work, GFP_KERNEL);
2397 INIT_WORK(&work->work, cma_work_handler);
2398 work->old_state = RDMA_CM_ROUTE_QUERY;
2399 work->new_state = RDMA_CM_ROUTE_RESOLVED;
2400 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
2402 route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
2403 if (!route->path_rec) {
2408 ret = cma_query_ib_route(id_priv, timeout_ms, work);
2414 kfree(route->path_rec);
2415 route->path_rec = NULL;
2421 int rdma_set_ib_paths(struct rdma_cm_id *id,
2422 struct ib_sa_path_rec *path_rec, int num_paths)
2424 struct rdma_id_private *id_priv;
2427 id_priv = container_of(id, struct rdma_id_private, id);
2428 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
2429 RDMA_CM_ROUTE_RESOLVED))
2432 id->route.path_rec = kmemdup(path_rec, sizeof *path_rec * num_paths,
2434 if (!id->route.path_rec) {
2439 id->route.num_paths = num_paths;
2442 cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED);
2445 EXPORT_SYMBOL(rdma_set_ib_paths);
2447 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms)
2449 struct cma_work *work;
2451 work = kzalloc(sizeof *work, GFP_KERNEL);
2456 INIT_WORK(&work->work, cma_work_handler);
2457 work->old_state = RDMA_CM_ROUTE_QUERY;
2458 work->new_state = RDMA_CM_ROUTE_RESOLVED;
2459 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
2460 queue_work(cma_wq, &work->work);
2464 static int iboe_tos_to_sl(struct net_device *ndev, int tos)
2467 struct net_device *dev;
2469 prio = rt_tos2priority(tos);
2470 dev = ndev->priv_flags & IFF_802_1Q_VLAN ?
2471 vlan_dev_real_dev(ndev) : ndev;
2474 return netdev_get_prio_tc_map(dev, prio);
2476 #if IS_ENABLED(CONFIG_VLAN_8021Q)
2477 if (ndev->priv_flags & IFF_802_1Q_VLAN)
2478 return (vlan_dev_get_egress_qos_mask(ndev, prio) &
2479 VLAN_PRIO_MASK) >> VLAN_PRIO_SHIFT;
2484 static enum ib_gid_type cma_route_gid_type(enum rdma_network_type network_type,
2485 unsigned long supported_gids,
2486 enum ib_gid_type default_gid)
2488 if ((network_type == RDMA_NETWORK_IPV4 ||
2489 network_type == RDMA_NETWORK_IPV6) &&
2490 test_bit(IB_GID_TYPE_ROCE_UDP_ENCAP, &supported_gids))
2491 return IB_GID_TYPE_ROCE_UDP_ENCAP;
2496 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
2498 struct rdma_route *route = &id_priv->id.route;
2499 struct rdma_addr *addr = &route->addr;
2500 struct cma_work *work;
2502 struct net_device *ndev = NULL;
2505 work = kzalloc(sizeof *work, GFP_KERNEL);
2510 INIT_WORK(&work->work, cma_work_handler);
2512 route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
2513 if (!route->path_rec) {
2518 route->num_paths = 1;
2520 if (addr->dev_addr.bound_dev_if) {
2521 unsigned long supported_gids;
2523 ndev = dev_get_by_index(&init_net, addr->dev_addr.bound_dev_if);
2529 if (ndev->flags & IFF_LOOPBACK) {
2531 if (!id_priv->id.device->get_netdev) {
2536 ndev = id_priv->id.device->get_netdev(id_priv->id.device,
2537 id_priv->id.port_num);
2544 route->path_rec->net = &init_net;
2545 route->path_rec->ifindex = ndev->ifindex;
2546 supported_gids = roce_gid_type_mask_support(id_priv->id.device,
2547 id_priv->id.port_num);
2548 route->path_rec->gid_type =
2549 cma_route_gid_type(addr->dev_addr.network,
2558 memcpy(route->path_rec->dmac, addr->dev_addr.dst_dev_addr, ETH_ALEN);
2560 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
2561 &route->path_rec->sgid);
2562 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.dst_addr,
2563 &route->path_rec->dgid);
2565 /* Use the hint from IP Stack to select GID Type */
2566 if (route->path_rec->gid_type < ib_network_to_gid_type(addr->dev_addr.network))
2567 route->path_rec->gid_type = ib_network_to_gid_type(addr->dev_addr.network);
2568 if (((struct sockaddr *)&id_priv->id.route.addr.dst_addr)->sa_family != AF_IB)
2569 /* TODO: get the hoplimit from the inet/inet6 device */
2570 route->path_rec->hop_limit = addr->dev_addr.hoplimit;
2572 route->path_rec->hop_limit = 1;
2573 route->path_rec->reversible = 1;
2574 route->path_rec->pkey = cpu_to_be16(0xffff);
2575 route->path_rec->mtu_selector = IB_SA_EQ;
2576 route->path_rec->sl = iboe_tos_to_sl(ndev, id_priv->tos);
2577 route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
2578 route->path_rec->rate_selector = IB_SA_EQ;
2579 route->path_rec->rate = iboe_get_rate(ndev);
2581 route->path_rec->packet_life_time_selector = IB_SA_EQ;
2582 route->path_rec->packet_life_time = CMA_IBOE_PACKET_LIFETIME;
2583 if (!route->path_rec->mtu) {
2588 work->old_state = RDMA_CM_ROUTE_QUERY;
2589 work->new_state = RDMA_CM_ROUTE_RESOLVED;
2590 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
2591 work->event.status = 0;
2593 queue_work(cma_wq, &work->work);
2598 kfree(route->path_rec);
2599 route->path_rec = NULL;
2605 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
2607 struct rdma_id_private *id_priv;
2610 id_priv = container_of(id, struct rdma_id_private, id);
2611 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY))
2614 atomic_inc(&id_priv->refcount);
2615 if (rdma_cap_ib_sa(id->device, id->port_num))
2616 ret = cma_resolve_ib_route(id_priv, timeout_ms);
2617 else if (rdma_protocol_roce(id->device, id->port_num))
2618 ret = cma_resolve_iboe_route(id_priv);
2619 else if (rdma_protocol_iwarp(id->device, id->port_num))
2620 ret = cma_resolve_iw_route(id_priv, timeout_ms);
2629 cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED);
2630 cma_deref_id(id_priv);
2633 EXPORT_SYMBOL(rdma_resolve_route);
2635 static void cma_set_loopback(struct sockaddr *addr)
2637 switch (addr->sa_family) {
2639 ((struct sockaddr_in *) addr)->sin_addr.s_addr = htonl(INADDR_LOOPBACK);
2642 ipv6_addr_set(&((struct sockaddr_in6 *) addr)->sin6_addr,
2646 ib_addr_set(&((struct sockaddr_ib *) addr)->sib_addr,
2652 static int cma_bind_loopback(struct rdma_id_private *id_priv)
2654 struct cma_device *cma_dev, *cur_dev;
2655 struct ib_port_attr port_attr;
2663 list_for_each_entry(cur_dev, &dev_list, list) {
2664 if (cma_family(id_priv) == AF_IB &&
2665 !rdma_cap_ib_cm(cur_dev->device, 1))
2671 for (p = 1; p <= cur_dev->device->phys_port_cnt; ++p) {
2672 if (!ib_query_port(cur_dev->device, p, &port_attr) &&
2673 port_attr.state == IB_PORT_ACTIVE) {
2688 ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid, NULL);
2692 ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
2696 id_priv->id.route.addr.dev_addr.dev_type =
2697 (rdma_protocol_ib(cma_dev->device, p)) ?
2698 ARPHRD_INFINIBAND : ARPHRD_ETHER;
2700 rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2701 ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
2702 id_priv->id.port_num = p;
2703 cma_attach_to_dev(id_priv, cma_dev);
2704 cma_set_loopback(cma_src_addr(id_priv));
2706 mutex_unlock(&lock);
2710 static void addr_handler(int status, struct sockaddr *src_addr,
2711 struct rdma_dev_addr *dev_addr, void *context)
2713 struct rdma_id_private *id_priv = context;
2714 struct rdma_cm_event event;
2716 memset(&event, 0, sizeof event);
2717 mutex_lock(&id_priv->handler_mutex);
2718 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY,
2719 RDMA_CM_ADDR_RESOLVED))
2722 memcpy(cma_src_addr(id_priv), src_addr, rdma_addr_size(src_addr));
2723 if (!status && !id_priv->cma_dev)
2724 status = cma_acquire_dev(id_priv, NULL);
2727 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
2728 RDMA_CM_ADDR_BOUND))
2730 event.event = RDMA_CM_EVENT_ADDR_ERROR;
2731 event.status = status;
2733 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2735 if (id_priv->id.event_handler(&id_priv->id, &event)) {
2736 cma_exch(id_priv, RDMA_CM_DESTROYING);
2737 mutex_unlock(&id_priv->handler_mutex);
2738 cma_deref_id(id_priv);
2739 rdma_destroy_id(&id_priv->id);
2743 mutex_unlock(&id_priv->handler_mutex);
2744 cma_deref_id(id_priv);
2747 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
2749 struct cma_work *work;
2753 work = kzalloc(sizeof *work, GFP_KERNEL);
2757 if (!id_priv->cma_dev) {
2758 ret = cma_bind_loopback(id_priv);
2763 rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2764 rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
2767 INIT_WORK(&work->work, cma_work_handler);
2768 work->old_state = RDMA_CM_ADDR_QUERY;
2769 work->new_state = RDMA_CM_ADDR_RESOLVED;
2770 work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2771 queue_work(cma_wq, &work->work);
2778 static int cma_resolve_ib_addr(struct rdma_id_private *id_priv)
2780 struct cma_work *work;
2783 work = kzalloc(sizeof *work, GFP_KERNEL);
2787 if (!id_priv->cma_dev) {
2788 ret = cma_resolve_ib_dev(id_priv);
2793 rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, (union ib_gid *)
2794 &(((struct sockaddr_ib *) &id_priv->id.route.addr.dst_addr)->sib_addr));
2797 INIT_WORK(&work->work, cma_work_handler);
2798 work->old_state = RDMA_CM_ADDR_QUERY;
2799 work->new_state = RDMA_CM_ADDR_RESOLVED;
2800 work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2801 queue_work(cma_wq, &work->work);
2808 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2809 struct sockaddr *dst_addr)
2811 if (!src_addr || !src_addr->sa_family) {
2812 src_addr = (struct sockaddr *) &id->route.addr.src_addr;
2813 src_addr->sa_family = dst_addr->sa_family;
2814 if (IS_ENABLED(CONFIG_IPV6) &&
2815 dst_addr->sa_family == AF_INET6) {
2816 struct sockaddr_in6 *src_addr6 = (struct sockaddr_in6 *) src_addr;
2817 struct sockaddr_in6 *dst_addr6 = (struct sockaddr_in6 *) dst_addr;
2818 src_addr6->sin6_scope_id = dst_addr6->sin6_scope_id;
2819 if (ipv6_addr_type(&dst_addr6->sin6_addr) & IPV6_ADDR_LINKLOCAL)
2820 id->route.addr.dev_addr.bound_dev_if = dst_addr6->sin6_scope_id;
2821 } else if (dst_addr->sa_family == AF_IB) {
2822 ((struct sockaddr_ib *) src_addr)->sib_pkey =
2823 ((struct sockaddr_ib *) dst_addr)->sib_pkey;
2826 return rdma_bind_addr(id, src_addr);
2829 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2830 struct sockaddr *dst_addr, int timeout_ms)
2832 struct rdma_id_private *id_priv;
2835 id_priv = container_of(id, struct rdma_id_private, id);
2836 if (id_priv->state == RDMA_CM_IDLE) {
2837 ret = cma_bind_addr(id, src_addr, dst_addr);
2842 if (cma_family(id_priv) != dst_addr->sa_family)
2845 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY))
2848 atomic_inc(&id_priv->refcount);
2849 memcpy(cma_dst_addr(id_priv), dst_addr, rdma_addr_size(dst_addr));
2850 if (cma_any_addr(dst_addr)) {
2851 ret = cma_resolve_loopback(id_priv);
2853 if (dst_addr->sa_family == AF_IB) {
2854 ret = cma_resolve_ib_addr(id_priv);
2856 ret = rdma_resolve_ip(&addr_client, cma_src_addr(id_priv),
2857 dst_addr, &id->route.addr.dev_addr,
2858 timeout_ms, addr_handler, id_priv);
2866 cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
2867 cma_deref_id(id_priv);
2870 EXPORT_SYMBOL(rdma_resolve_addr);
2872 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse)
2874 struct rdma_id_private *id_priv;
2875 unsigned long flags;
2878 id_priv = container_of(id, struct rdma_id_private, id);
2879 spin_lock_irqsave(&id_priv->lock, flags);
2880 if (reuse || id_priv->state == RDMA_CM_IDLE) {
2881 id_priv->reuseaddr = reuse;
2886 spin_unlock_irqrestore(&id_priv->lock, flags);
2889 EXPORT_SYMBOL(rdma_set_reuseaddr);
2891 int rdma_set_afonly(struct rdma_cm_id *id, int afonly)
2893 struct rdma_id_private *id_priv;
2894 unsigned long flags;
2897 id_priv = container_of(id, struct rdma_id_private, id);
2898 spin_lock_irqsave(&id_priv->lock, flags);
2899 if (id_priv->state == RDMA_CM_IDLE || id_priv->state == RDMA_CM_ADDR_BOUND) {
2900 id_priv->options |= (1 << CMA_OPTION_AFONLY);
2901 id_priv->afonly = afonly;
2906 spin_unlock_irqrestore(&id_priv->lock, flags);
2909 EXPORT_SYMBOL(rdma_set_afonly);
2911 static void cma_bind_port(struct rdma_bind_list *bind_list,
2912 struct rdma_id_private *id_priv)
2914 struct sockaddr *addr;
2915 struct sockaddr_ib *sib;
2919 addr = cma_src_addr(id_priv);
2920 port = htons(bind_list->port);
2922 switch (addr->sa_family) {
2924 ((struct sockaddr_in *) addr)->sin_port = port;
2927 ((struct sockaddr_in6 *) addr)->sin6_port = port;
2930 sib = (struct sockaddr_ib *) addr;
2931 sid = be64_to_cpu(sib->sib_sid);
2932 mask = be64_to_cpu(sib->sib_sid_mask);
2933 sib->sib_sid = cpu_to_be64((sid & mask) | (u64) ntohs(port));
2934 sib->sib_sid_mask = cpu_to_be64(~0ULL);
2937 id_priv->bind_list = bind_list;
2938 hlist_add_head(&id_priv->node, &bind_list->owners);
2941 static int cma_alloc_port(enum rdma_port_space ps,
2942 struct rdma_id_private *id_priv, unsigned short snum)
2944 struct rdma_bind_list *bind_list;
2947 bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
2951 ret = cma_ps_alloc(id_priv->id.route.addr.dev_addr.net, ps, bind_list,
2957 bind_list->port = (unsigned short)ret;
2958 cma_bind_port(bind_list, id_priv);
2962 return ret == -ENOSPC ? -EADDRNOTAVAIL : ret;
2965 static int cma_alloc_any_port(enum rdma_port_space ps,
2966 struct rdma_id_private *id_priv)
2968 static unsigned int last_used_port;
2969 int low, high, remaining;
2971 struct net *net = id_priv->id.route.addr.dev_addr.net;
2973 inet_get_local_port_range(net, &low, &high);
2974 remaining = (high - low) + 1;
2975 rover = prandom_u32() % remaining + low;
2977 if (last_used_port != rover &&
2978 !cma_ps_find(net, ps, (unsigned short)rover)) {
2979 int ret = cma_alloc_port(ps, id_priv, rover);
2981 * Remember previously used port number in order to avoid
2982 * re-using same port immediately after it is closed.
2985 last_used_port = rover;
2986 if (ret != -EADDRNOTAVAIL)
2991 if ((rover < low) || (rover > high))
2995 return -EADDRNOTAVAIL;
2999 * Check that the requested port is available. This is called when trying to
3000 * bind to a specific port, or when trying to listen on a bound port. In
3001 * the latter case, the provided id_priv may already be on the bind_list, but
3002 * we still need to check that it's okay to start listening.
3004 static int cma_check_port(struct rdma_bind_list *bind_list,
3005 struct rdma_id_private *id_priv, uint8_t reuseaddr)
3007 struct rdma_id_private *cur_id;
3008 struct sockaddr *addr, *cur_addr;
3010 addr = cma_src_addr(id_priv);
3011 hlist_for_each_entry(cur_id, &bind_list->owners, node) {
3012 if (id_priv == cur_id)
3015 if ((cur_id->state != RDMA_CM_LISTEN) && reuseaddr &&
3019 cur_addr = cma_src_addr(cur_id);
3020 if (id_priv->afonly && cur_id->afonly &&
3021 (addr->sa_family != cur_addr->sa_family))
3024 if (cma_any_addr(addr) || cma_any_addr(cur_addr))
3025 return -EADDRNOTAVAIL;
3027 if (!cma_addr_cmp(addr, cur_addr))
3033 static int cma_use_port(enum rdma_port_space ps,
3034 struct rdma_id_private *id_priv)
3036 struct rdma_bind_list *bind_list;
3037 unsigned short snum;
3040 snum = ntohs(cma_port(cma_src_addr(id_priv)));
3041 if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
3044 bind_list = cma_ps_find(id_priv->id.route.addr.dev_addr.net, ps, snum);
3046 ret = cma_alloc_port(ps, id_priv, snum);
3048 ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr);
3050 cma_bind_port(bind_list, id_priv);
3055 static int cma_bind_listen(struct rdma_id_private *id_priv)
3057 struct rdma_bind_list *bind_list = id_priv->bind_list;
3061 if (bind_list->owners.first->next)
3062 ret = cma_check_port(bind_list, id_priv, 0);
3063 mutex_unlock(&lock);
3067 static enum rdma_port_space cma_select_inet_ps(
3068 struct rdma_id_private *id_priv)
3070 switch (id_priv->id.ps) {
3075 return id_priv->id.ps;
3082 static enum rdma_port_space cma_select_ib_ps(struct rdma_id_private *id_priv)
3084 enum rdma_port_space ps = 0;
3085 struct sockaddr_ib *sib;
3086 u64 sid_ps, mask, sid;
3088 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
3089 mask = be64_to_cpu(sib->sib_sid_mask) & RDMA_IB_IP_PS_MASK;
3090 sid = be64_to_cpu(sib->sib_sid) & mask;
3092 if ((id_priv->id.ps == RDMA_PS_IB) && (sid == (RDMA_IB_IP_PS_IB & mask))) {
3093 sid_ps = RDMA_IB_IP_PS_IB;
3095 } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_TCP)) &&
3096 (sid == (RDMA_IB_IP_PS_TCP & mask))) {
3097 sid_ps = RDMA_IB_IP_PS_TCP;
3099 } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_UDP)) &&
3100 (sid == (RDMA_IB_IP_PS_UDP & mask))) {
3101 sid_ps = RDMA_IB_IP_PS_UDP;
3106 sib->sib_sid = cpu_to_be64(sid_ps | ntohs(cma_port((struct sockaddr *) sib)));
3107 sib->sib_sid_mask = cpu_to_be64(RDMA_IB_IP_PS_MASK |
3108 be64_to_cpu(sib->sib_sid_mask));
3113 static int cma_get_port(struct rdma_id_private *id_priv)
3115 enum rdma_port_space ps;
3118 if (cma_family(id_priv) != AF_IB)
3119 ps = cma_select_inet_ps(id_priv);
3121 ps = cma_select_ib_ps(id_priv);
3123 return -EPROTONOSUPPORT;
3126 if (cma_any_port(cma_src_addr(id_priv)))
3127 ret = cma_alloc_any_port(ps, id_priv);
3129 ret = cma_use_port(ps, id_priv);
3130 mutex_unlock(&lock);
3135 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
3136 struct sockaddr *addr)
3138 #if IS_ENABLED(CONFIG_IPV6)
3139 struct sockaddr_in6 *sin6;
3141 if (addr->sa_family != AF_INET6)
3144 sin6 = (struct sockaddr_in6 *) addr;
3146 if (!(ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL))
3149 if (!sin6->sin6_scope_id)
3152 dev_addr->bound_dev_if = sin6->sin6_scope_id;
3157 int rdma_listen(struct rdma_cm_id *id, int backlog)
3159 struct rdma_id_private *id_priv;
3162 id_priv = container_of(id, struct rdma_id_private, id);
3163 if (id_priv->state == RDMA_CM_IDLE) {
3164 id->route.addr.src_addr.ss_family = AF_INET;
3165 ret = rdma_bind_addr(id, cma_src_addr(id_priv));
3170 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN))
3173 if (id_priv->reuseaddr) {
3174 ret = cma_bind_listen(id_priv);
3179 id_priv->backlog = backlog;
3181 if (rdma_cap_ib_cm(id->device, 1)) {
3182 ret = cma_ib_listen(id_priv);
3185 } else if (rdma_cap_iw_cm(id->device, 1)) {
3186 ret = cma_iw_listen(id_priv, backlog);
3194 cma_listen_on_all(id_priv);
3198 id_priv->backlog = 0;
3199 cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND);
3202 EXPORT_SYMBOL(rdma_listen);
3204 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
3206 struct rdma_id_private *id_priv;
3209 if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6 &&
3210 addr->sa_family != AF_IB)
3211 return -EAFNOSUPPORT;
3213 id_priv = container_of(id, struct rdma_id_private, id);
3214 if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND))
3217 ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
3221 memcpy(cma_src_addr(id_priv), addr, rdma_addr_size(addr));
3222 if (!cma_any_addr(addr)) {
3223 ret = cma_translate_addr(addr, &id->route.addr.dev_addr);
3227 ret = cma_acquire_dev(id_priv, NULL);
3232 if (!(id_priv->options & (1 << CMA_OPTION_AFONLY))) {
3233 if (addr->sa_family == AF_INET)
3234 id_priv->afonly = 1;
3235 #if IS_ENABLED(CONFIG_IPV6)
3236 else if (addr->sa_family == AF_INET6) {
3237 struct net *net = id_priv->id.route.addr.dev_addr.net;
3239 id_priv->afonly = net->ipv6.sysctl.bindv6only;
3243 ret = cma_get_port(id_priv);
3249 if (id_priv->cma_dev)
3250 cma_release_dev(id_priv);
3252 cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE);
3255 EXPORT_SYMBOL(rdma_bind_addr);
3257 static int cma_format_hdr(void *hdr, struct rdma_id_private *id_priv)
3259 struct cma_hdr *cma_hdr;
3262 cma_hdr->cma_version = CMA_VERSION;
3263 if (cma_family(id_priv) == AF_INET) {
3264 struct sockaddr_in *src4, *dst4;
3266 src4 = (struct sockaddr_in *) cma_src_addr(id_priv);
3267 dst4 = (struct sockaddr_in *) cma_dst_addr(id_priv);
3269 cma_set_ip_ver(cma_hdr, 4);
3270 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
3271 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
3272 cma_hdr->port = src4->sin_port;
3273 } else if (cma_family(id_priv) == AF_INET6) {
3274 struct sockaddr_in6 *src6, *dst6;
3276 src6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
3277 dst6 = (struct sockaddr_in6 *) cma_dst_addr(id_priv);
3279 cma_set_ip_ver(cma_hdr, 6);
3280 cma_hdr->src_addr.ip6 = src6->sin6_addr;
3281 cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
3282 cma_hdr->port = src6->sin6_port;
3287 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
3288 struct ib_cm_event *ib_event)
3290 struct rdma_id_private *id_priv = cm_id->context;
3291 struct rdma_cm_event event;
3292 struct ib_cm_sidr_rep_event_param *rep = &ib_event->param.sidr_rep_rcvd;
3295 mutex_lock(&id_priv->handler_mutex);
3296 if (id_priv->state != RDMA_CM_CONNECT)
3299 memset(&event, 0, sizeof event);
3300 switch (ib_event->event) {
3301 case IB_CM_SIDR_REQ_ERROR:
3302 event.event = RDMA_CM_EVENT_UNREACHABLE;
3303 event.status = -ETIMEDOUT;
3305 case IB_CM_SIDR_REP_RECEIVED:
3306 event.param.ud.private_data = ib_event->private_data;
3307 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
3308 if (rep->status != IB_SIDR_SUCCESS) {
3309 event.event = RDMA_CM_EVENT_UNREACHABLE;
3310 event.status = ib_event->param.sidr_rep_rcvd.status;
3313 ret = cma_set_qkey(id_priv, rep->qkey);
3315 event.event = RDMA_CM_EVENT_ADDR_ERROR;
3319 ib_init_ah_from_path(id_priv->id.device, id_priv->id.port_num,
3320 id_priv->id.route.path_rec,
3321 &event.param.ud.ah_attr);
3322 event.param.ud.qp_num = rep->qpn;
3323 event.param.ud.qkey = rep->qkey;
3324 event.event = RDMA_CM_EVENT_ESTABLISHED;
3328 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
3333 ret = id_priv->id.event_handler(&id_priv->id, &event);
3335 /* Destroy the CM ID by returning a non-zero value. */
3336 id_priv->cm_id.ib = NULL;
3337 cma_exch(id_priv, RDMA_CM_DESTROYING);
3338 mutex_unlock(&id_priv->handler_mutex);
3339 rdma_destroy_id(&id_priv->id);
3343 mutex_unlock(&id_priv->handler_mutex);
3347 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
3348 struct rdma_conn_param *conn_param)
3350 struct ib_cm_sidr_req_param req;
3351 struct ib_cm_id *id;
3355 memset(&req, 0, sizeof req);
3356 offset = cma_user_data_offset(id_priv);
3357 req.private_data_len = offset + conn_param->private_data_len;
3358 if (req.private_data_len < conn_param->private_data_len)
3361 if (req.private_data_len) {
3362 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
3366 private_data = NULL;
3369 if (conn_param->private_data && conn_param->private_data_len)
3370 memcpy(private_data + offset, conn_param->private_data,
3371 conn_param->private_data_len);
3374 ret = cma_format_hdr(private_data, id_priv);
3377 req.private_data = private_data;
3380 id = ib_create_cm_id(id_priv->id.device, cma_sidr_rep_handler,
3386 id_priv->cm_id.ib = id;
3388 req.path = id_priv->id.route.path_rec;
3389 req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
3390 req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
3391 req.max_cm_retries = CMA_MAX_CM_RETRIES;
3393 ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
3395 ib_destroy_cm_id(id_priv->cm_id.ib);
3396 id_priv->cm_id.ib = NULL;
3399 kfree(private_data);
3403 static int cma_connect_ib(struct rdma_id_private *id_priv,
3404 struct rdma_conn_param *conn_param)
3406 struct ib_cm_req_param req;
3407 struct rdma_route *route;
3409 struct ib_cm_id *id;
3412 memset(&req, 0, sizeof req);
3413 offset = cma_user_data_offset(id_priv);
3414 req.private_data_len = offset + conn_param->private_data_len;
3415 if (req.private_data_len < conn_param->private_data_len)
3418 if (req.private_data_len) {
3419 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
3423 private_data = NULL;
3426 if (conn_param->private_data && conn_param->private_data_len)
3427 memcpy(private_data + offset, conn_param->private_data,
3428 conn_param->private_data_len);
3430 id = ib_create_cm_id(id_priv->id.device, cma_ib_handler, id_priv);
3435 id_priv->cm_id.ib = id;
3437 route = &id_priv->id.route;
3439 ret = cma_format_hdr(private_data, id_priv);
3442 req.private_data = private_data;
3445 req.primary_path = &route->path_rec[0];
3446 if (route->num_paths == 2)
3447 req.alternate_path = &route->path_rec[1];
3449 req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
3450 req.qp_num = id_priv->qp_num;
3451 req.qp_type = id_priv->id.qp_type;
3452 req.starting_psn = id_priv->seq_num;
3453 req.responder_resources = conn_param->responder_resources;
3454 req.initiator_depth = conn_param->initiator_depth;
3455 req.flow_control = conn_param->flow_control;
3456 req.retry_count = min_t(u8, 7, conn_param->retry_count);
3457 req.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
3458 req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
3459 req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
3460 req.max_cm_retries = CMA_MAX_CM_RETRIES;
3461 req.srq = id_priv->srq ? 1 : 0;
3463 ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
3465 if (ret && !IS_ERR(id)) {
3466 ib_destroy_cm_id(id);
3467 id_priv->cm_id.ib = NULL;
3470 kfree(private_data);
3474 static int cma_connect_iw(struct rdma_id_private *id_priv,
3475 struct rdma_conn_param *conn_param)
3477 struct iw_cm_id *cm_id;
3479 struct iw_cm_conn_param iw_param;
3481 cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
3483 return PTR_ERR(cm_id);
3485 cm_id->tos = id_priv->tos;
3486 id_priv->cm_id.iw = cm_id;
3488 memcpy(&cm_id->local_addr, cma_src_addr(id_priv),
3489 rdma_addr_size(cma_src_addr(id_priv)));
3490 memcpy(&cm_id->remote_addr, cma_dst_addr(id_priv),
3491 rdma_addr_size(cma_dst_addr(id_priv)));
3493 ret = cma_modify_qp_rtr(id_priv, conn_param);
3498 iw_param.ord = conn_param->initiator_depth;
3499 iw_param.ird = conn_param->responder_resources;
3500 iw_param.private_data = conn_param->private_data;
3501 iw_param.private_data_len = conn_param->private_data_len;
3502 iw_param.qpn = id_priv->id.qp ? id_priv->qp_num : conn_param->qp_num;
3504 memset(&iw_param, 0, sizeof iw_param);
3505 iw_param.qpn = id_priv->qp_num;
3507 ret = iw_cm_connect(cm_id, &iw_param);
3510 iw_destroy_cm_id(cm_id);
3511 id_priv->cm_id.iw = NULL;
3516 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
3518 struct rdma_id_private *id_priv;
3521 id_priv = container_of(id, struct rdma_id_private, id);
3522 if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT))
3526 id_priv->qp_num = conn_param->qp_num;
3527 id_priv->srq = conn_param->srq;
3530 if (rdma_cap_ib_cm(id->device, id->port_num)) {
3531 if (id->qp_type == IB_QPT_UD)
3532 ret = cma_resolve_ib_udp(id_priv, conn_param);
3534 ret = cma_connect_ib(id_priv, conn_param);
3535 } else if (rdma_cap_iw_cm(id->device, id->port_num))
3536 ret = cma_connect_iw(id_priv, conn_param);
3544 cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED);
3547 EXPORT_SYMBOL(rdma_connect);
3549 static int cma_accept_ib(struct rdma_id_private *id_priv,
3550 struct rdma_conn_param *conn_param)
3552 struct ib_cm_rep_param rep;
3555 ret = cma_modify_qp_rtr(id_priv, conn_param);
3559 ret = cma_modify_qp_rts(id_priv, conn_param);
3563 memset(&rep, 0, sizeof rep);
3564 rep.qp_num = id_priv->qp_num;
3565 rep.starting_psn = id_priv->seq_num;
3566 rep.private_data = conn_param->private_data;
3567 rep.private_data_len = conn_param->private_data_len;
3568 rep.responder_resources = conn_param->responder_resources;
3569 rep.initiator_depth = conn_param->initiator_depth;
3570 rep.failover_accepted = 0;
3571 rep.flow_control = conn_param->flow_control;
3572 rep.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
3573 rep.srq = id_priv->srq ? 1 : 0;
3575 ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
3580 static int cma_accept_iw(struct rdma_id_private *id_priv,
3581 struct rdma_conn_param *conn_param)
3583 struct iw_cm_conn_param iw_param;
3586 ret = cma_modify_qp_rtr(id_priv, conn_param);
3590 iw_param.ord = conn_param->initiator_depth;
3591 iw_param.ird = conn_param->responder_resources;
3592 iw_param.private_data = conn_param->private_data;
3593 iw_param.private_data_len = conn_param->private_data_len;
3594 if (id_priv->id.qp) {
3595 iw_param.qpn = id_priv->qp_num;
3597 iw_param.qpn = conn_param->qp_num;
3599 return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
3602 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
3603 enum ib_cm_sidr_status status, u32 qkey,
3604 const void *private_data, int private_data_len)
3606 struct ib_cm_sidr_rep_param rep;
3609 memset(&rep, 0, sizeof rep);
3610 rep.status = status;
3611 if (status == IB_SIDR_SUCCESS) {
3612 ret = cma_set_qkey(id_priv, qkey);
3615 rep.qp_num = id_priv->qp_num;
3616 rep.qkey = id_priv->qkey;
3618 rep.private_data = private_data;
3619 rep.private_data_len = private_data_len;
3621 return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
3624 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
3626 struct rdma_id_private *id_priv;
3629 id_priv = container_of(id, struct rdma_id_private, id);
3631 id_priv->owner = task_pid_nr(current);
3633 if (!cma_comp(id_priv, RDMA_CM_CONNECT))
3636 if (!id->qp && conn_param) {
3637 id_priv->qp_num = conn_param->qp_num;
3638 id_priv->srq = conn_param->srq;
3641 if (rdma_cap_ib_cm(id->device, id->port_num)) {
3642 if (id->qp_type == IB_QPT_UD) {
3644 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
3646 conn_param->private_data,
3647 conn_param->private_data_len);
3649 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
3653 ret = cma_accept_ib(id_priv, conn_param);
3655 ret = cma_rep_recv(id_priv);
3657 } else if (rdma_cap_iw_cm(id->device, id->port_num))
3658 ret = cma_accept_iw(id_priv, conn_param);
3667 cma_modify_qp_err(id_priv);
3668 rdma_reject(id, NULL, 0);
3671 EXPORT_SYMBOL(rdma_accept);
3673 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
3675 struct rdma_id_private *id_priv;
3678 id_priv = container_of(id, struct rdma_id_private, id);
3679 if (!id_priv->cm_id.ib)
3682 switch (id->device->node_type) {
3683 case RDMA_NODE_IB_CA:
3684 ret = ib_cm_notify(id_priv->cm_id.ib, event);
3692 EXPORT_SYMBOL(rdma_notify);
3694 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
3695 u8 private_data_len)
3697 struct rdma_id_private *id_priv;
3700 id_priv = container_of(id, struct rdma_id_private, id);
3701 if (!id_priv->cm_id.ib)
3704 if (rdma_cap_ib_cm(id->device, id->port_num)) {
3705 if (id->qp_type == IB_QPT_UD)
3706 ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT, 0,
3707 private_data, private_data_len);
3709 ret = ib_send_cm_rej(id_priv->cm_id.ib,
3710 IB_CM_REJ_CONSUMER_DEFINED, NULL,
3711 0, private_data, private_data_len);
3712 } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
3713 ret = iw_cm_reject(id_priv->cm_id.iw,
3714 private_data, private_data_len);
3720 EXPORT_SYMBOL(rdma_reject);
3722 int rdma_disconnect(struct rdma_cm_id *id)
3724 struct rdma_id_private *id_priv;
3727 id_priv = container_of(id, struct rdma_id_private, id);
3728 if (!id_priv->cm_id.ib)
3731 if (rdma_cap_ib_cm(id->device, id->port_num)) {
3732 ret = cma_modify_qp_err(id_priv);
3735 /* Initiate or respond to a disconnect. */
3736 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
3737 ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
3738 } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
3739 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
3746 EXPORT_SYMBOL(rdma_disconnect);
3748 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
3750 struct rdma_id_private *id_priv;
3751 struct cma_multicast *mc = multicast->context;
3752 struct rdma_cm_event event;
3755 id_priv = mc->id_priv;
3756 mutex_lock(&id_priv->handler_mutex);
3757 if (id_priv->state != RDMA_CM_ADDR_BOUND &&
3758 id_priv->state != RDMA_CM_ADDR_RESOLVED)
3762 status = cma_set_qkey(id_priv, be32_to_cpu(multicast->rec.qkey));
3763 mutex_lock(&id_priv->qp_mutex);
3764 if (!status && id_priv->id.qp)
3765 status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid,
3766 be16_to_cpu(multicast->rec.mlid));
3767 mutex_unlock(&id_priv->qp_mutex);
3769 memset(&event, 0, sizeof event);
3770 event.status = status;
3771 event.param.ud.private_data = mc->context;
3773 struct rdma_dev_addr *dev_addr =
3774 &id_priv->id.route.addr.dev_addr;
3775 struct net_device *ndev =
3776 dev_get_by_index(&init_net, dev_addr->bound_dev_if);
3777 enum ib_gid_type gid_type =
3778 id_priv->cma_dev->default_gid_type[id_priv->id.port_num -
3779 rdma_start_port(id_priv->cma_dev->device)];
3781 event.event = RDMA_CM_EVENT_MULTICAST_JOIN;
3782 ib_init_ah_from_mcmember(id_priv->id.device,
3783 id_priv->id.port_num, &multicast->rec,
3785 &event.param.ud.ah_attr);
3786 event.param.ud.qp_num = 0xFFFFFF;
3787 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
3791 event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
3793 ret = id_priv->id.event_handler(&id_priv->id, &event);
3795 cma_exch(id_priv, RDMA_CM_DESTROYING);
3796 mutex_unlock(&id_priv->handler_mutex);
3797 rdma_destroy_id(&id_priv->id);
3802 mutex_unlock(&id_priv->handler_mutex);
3806 static void cma_set_mgid(struct rdma_id_private *id_priv,
3807 struct sockaddr *addr, union ib_gid *mgid)
3809 unsigned char mc_map[MAX_ADDR_LEN];
3810 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3811 struct sockaddr_in *sin = (struct sockaddr_in *) addr;
3812 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
3814 if (cma_any_addr(addr)) {
3815 memset(mgid, 0, sizeof *mgid);
3816 } else if ((addr->sa_family == AF_INET6) &&
3817 ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
3819 /* IPv6 address is an SA assigned MGID. */
3820 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3821 } else if (addr->sa_family == AF_IB) {
3822 memcpy(mgid, &((struct sockaddr_ib *) addr)->sib_addr, sizeof *mgid);
3823 } else if ((addr->sa_family == AF_INET6)) {
3824 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
3825 if (id_priv->id.ps == RDMA_PS_UDP)
3826 mc_map[7] = 0x01; /* Use RDMA CM signature */
3827 *mgid = *(union ib_gid *) (mc_map + 4);
3829 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
3830 if (id_priv->id.ps == RDMA_PS_UDP)
3831 mc_map[7] = 0x01; /* Use RDMA CM signature */
3832 *mgid = *(union ib_gid *) (mc_map + 4);
3836 static void cma_query_sa_classport_info_cb(int status,
3837 struct ib_class_port_info *rec,
3840 struct class_port_info_context *cb_ctx = context;
3844 if (status || !rec) {
3845 pr_debug("RDMA CM: %s port %u failed query ClassPortInfo status: %d\n",
3846 cb_ctx->device->name, cb_ctx->port_num, status);
3850 memcpy(cb_ctx->class_port_info, rec, sizeof(struct ib_class_port_info));
3853 complete(&cb_ctx->done);
3856 static int cma_query_sa_classport_info(struct ib_device *device, u8 port_num,
3857 struct ib_class_port_info *class_port_info)
3859 struct class_port_info_context *cb_ctx;
3862 cb_ctx = kmalloc(sizeof(*cb_ctx), GFP_KERNEL);
3866 cb_ctx->device = device;
3867 cb_ctx->class_port_info = class_port_info;
3868 cb_ctx->port_num = port_num;
3869 init_completion(&cb_ctx->done);
3871 ret = ib_sa_classport_info_rec_query(&sa_client, device, port_num,
3872 CMA_QUERY_CLASSPORT_INFO_TIMEOUT,
3873 GFP_KERNEL, cma_query_sa_classport_info_cb,
3874 cb_ctx, &cb_ctx->sa_query);
3876 pr_err("RDMA CM: %s port %u failed to send ClassPortInfo query, ret: %d\n",
3877 device->name, port_num, ret);
3881 wait_for_completion(&cb_ctx->done);
3888 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
3889 struct cma_multicast *mc)
3891 struct ib_sa_mcmember_rec rec;
3892 struct ib_class_port_info class_port_info;
3893 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3894 ib_sa_comp_mask comp_mask;
3897 ib_addr_get_mgid(dev_addr, &rec.mgid);
3898 ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
3903 ret = cma_set_qkey(id_priv, 0);
3907 cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
3908 rec.qkey = cpu_to_be32(id_priv->qkey);
3909 rdma_addr_get_sgid(dev_addr, &rec.port_gid);
3910 rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
3911 rec.join_state = mc->join_state;
3913 if (rec.join_state == BIT(SENDONLY_FULLMEMBER_JOIN)) {
3914 ret = cma_query_sa_classport_info(id_priv->id.device,
3915 id_priv->id.port_num,
3921 if (!(ib_get_cpi_capmask2(&class_port_info) &
3922 IB_SA_CAP_MASK2_SENDONLY_FULL_MEM_SUPPORT)) {
3923 pr_warn("RDMA CM: %s port %u Unable to multicast join\n"
3924 "RDMA CM: SM doesn't support Send Only Full Member option\n",
3925 id_priv->id.device->name, id_priv->id.port_num);
3930 comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
3931 IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
3932 IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
3933 IB_SA_MCMEMBER_REC_FLOW_LABEL |
3934 IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
3936 if (id_priv->id.ps == RDMA_PS_IPOIB)
3937 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
3938 IB_SA_MCMEMBER_REC_RATE_SELECTOR |
3939 IB_SA_MCMEMBER_REC_MTU_SELECTOR |
3940 IB_SA_MCMEMBER_REC_MTU |
3941 IB_SA_MCMEMBER_REC_HOP_LIMIT;
3943 mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device,
3944 id_priv->id.port_num, &rec,
3945 comp_mask, GFP_KERNEL,
3946 cma_ib_mc_handler, mc);
3947 return PTR_ERR_OR_ZERO(mc->multicast.ib);
3950 static void iboe_mcast_work_handler(struct work_struct *work)
3952 struct iboe_mcast_work *mw = container_of(work, struct iboe_mcast_work, work);
3953 struct cma_multicast *mc = mw->mc;
3954 struct ib_sa_multicast *m = mc->multicast.ib;
3956 mc->multicast.ib->context = mc;
3957 cma_ib_mc_handler(0, m);
3958 kref_put(&mc->mcref, release_mc);
3962 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid)
3964 struct sockaddr_in *sin = (struct sockaddr_in *)addr;
3965 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
3967 if (cma_any_addr(addr)) {
3968 memset(mgid, 0, sizeof *mgid);
3969 } else if (addr->sa_family == AF_INET6) {
3970 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3972 mgid->raw[0] = 0xff;
3973 mgid->raw[1] = 0x0e;
3982 mgid->raw[10] = 0xff;
3983 mgid->raw[11] = 0xff;
3984 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
3988 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
3989 struct cma_multicast *mc)
3991 struct iboe_mcast_work *work;
3992 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3994 struct sockaddr *addr = (struct sockaddr *)&mc->addr;
3995 struct net_device *ndev = NULL;
3996 enum ib_gid_type gid_type;
3999 send_only = mc->join_state == BIT(SENDONLY_FULLMEMBER_JOIN);
4001 if (cma_zero_addr((struct sockaddr *)&mc->addr))
4004 work = kzalloc(sizeof *work, GFP_KERNEL);
4008 mc->multicast.ib = kzalloc(sizeof(struct ib_sa_multicast), GFP_KERNEL);
4009 if (!mc->multicast.ib) {
4014 cma_iboe_set_mgid(addr, &mc->multicast.ib->rec.mgid);
4016 mc->multicast.ib->rec.pkey = cpu_to_be16(0xffff);
4017 if (id_priv->id.ps == RDMA_PS_UDP)
4018 mc->multicast.ib->rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
4020 if (dev_addr->bound_dev_if)
4021 ndev = dev_get_by_index(&init_net, dev_addr->bound_dev_if);
4026 mc->multicast.ib->rec.rate = iboe_get_rate(ndev);
4027 mc->multicast.ib->rec.hop_limit = 1;
4028 mc->multicast.ib->rec.mtu = iboe_get_mtu(ndev->mtu);
4030 gid_type = id_priv->cma_dev->default_gid_type[id_priv->id.port_num -
4031 rdma_start_port(id_priv->cma_dev->device)];
4032 if (addr->sa_family == AF_INET) {
4033 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) {
4034 mc->multicast.ib->rec.hop_limit = IPV6_DEFAULT_HOPLIMIT;
4036 err = cma_igmp_send(ndev, &mc->multicast.ib->rec.mgid,
4039 mc->igmp_joined = true;
4043 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP)
4047 if (err || !mc->multicast.ib->rec.mtu) {
4052 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
4053 &mc->multicast.ib->rec.port_gid);
4056 INIT_WORK(&work->work, iboe_mcast_work_handler);
4057 kref_get(&mc->mcref);
4058 queue_work(cma_wq, &work->work);
4063 kfree(mc->multicast.ib);
4069 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
4070 u8 join_state, void *context)
4072 struct rdma_id_private *id_priv;
4073 struct cma_multicast *mc;
4076 id_priv = container_of(id, struct rdma_id_private, id);
4077 if (!cma_comp(id_priv, RDMA_CM_ADDR_BOUND) &&
4078 !cma_comp(id_priv, RDMA_CM_ADDR_RESOLVED))
4081 mc = kmalloc(sizeof *mc, GFP_KERNEL);
4085 memcpy(&mc->addr, addr, rdma_addr_size(addr));
4086 mc->context = context;
4087 mc->id_priv = id_priv;
4088 mc->igmp_joined = false;
4089 mc->join_state = join_state;
4090 spin_lock(&id_priv->lock);
4091 list_add(&mc->list, &id_priv->mc_list);
4092 spin_unlock(&id_priv->lock);
4094 if (rdma_protocol_roce(id->device, id->port_num)) {
4095 kref_init(&mc->mcref);
4096 ret = cma_iboe_join_multicast(id_priv, mc);
4097 } else if (rdma_cap_ib_mcast(id->device, id->port_num))
4098 ret = cma_join_ib_multicast(id_priv, mc);
4103 spin_lock_irq(&id_priv->lock);
4104 list_del(&mc->list);
4105 spin_unlock_irq(&id_priv->lock);
4110 EXPORT_SYMBOL(rdma_join_multicast);
4112 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
4114 struct rdma_id_private *id_priv;
4115 struct cma_multicast *mc;
4117 id_priv = container_of(id, struct rdma_id_private, id);
4118 spin_lock_irq(&id_priv->lock);
4119 list_for_each_entry(mc, &id_priv->mc_list, list) {
4120 if (!memcmp(&mc->addr, addr, rdma_addr_size(addr))) {
4121 list_del(&mc->list);
4122 spin_unlock_irq(&id_priv->lock);
4125 ib_detach_mcast(id->qp,
4126 &mc->multicast.ib->rec.mgid,
4127 be16_to_cpu(mc->multicast.ib->rec.mlid));
4129 BUG_ON(id_priv->cma_dev->device != id->device);
4131 if (rdma_cap_ib_mcast(id->device, id->port_num)) {
4132 ib_sa_free_multicast(mc->multicast.ib);
4134 } else if (rdma_protocol_roce(id->device, id->port_num)) {
4135 if (mc->igmp_joined) {
4136 struct rdma_dev_addr *dev_addr =
4137 &id->route.addr.dev_addr;
4138 struct net_device *ndev = NULL;
4140 if (dev_addr->bound_dev_if)
4141 ndev = dev_get_by_index(&init_net,
4142 dev_addr->bound_dev_if);
4145 &mc->multicast.ib->rec.mgid,
4149 mc->igmp_joined = false;
4151 kref_put(&mc->mcref, release_mc);
4156 spin_unlock_irq(&id_priv->lock);
4158 EXPORT_SYMBOL(rdma_leave_multicast);
4160 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
4162 struct rdma_dev_addr *dev_addr;
4163 struct cma_ndev_work *work;
4165 dev_addr = &id_priv->id.route.addr.dev_addr;
4167 if ((dev_addr->bound_dev_if == ndev->ifindex) &&
4168 (net_eq(dev_net(ndev), dev_addr->net)) &&
4169 memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
4170 pr_info("RDMA CM addr change for ndev %s used by id %p\n",
4171 ndev->name, &id_priv->id);
4172 work = kzalloc(sizeof *work, GFP_KERNEL);
4176 INIT_WORK(&work->work, cma_ndev_work_handler);
4178 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
4179 atomic_inc(&id_priv->refcount);
4180 queue_work(cma_wq, &work->work);
4186 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
4189 struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
4190 struct cma_device *cma_dev;
4191 struct rdma_id_private *id_priv;
4192 int ret = NOTIFY_DONE;
4194 if (event != NETDEV_BONDING_FAILOVER)
4197 if (!(ndev->flags & IFF_MASTER) || !(ndev->priv_flags & IFF_BONDING))
4201 list_for_each_entry(cma_dev, &dev_list, list)
4202 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
4203 ret = cma_netdev_change(ndev, id_priv);
4209 mutex_unlock(&lock);
4213 static struct notifier_block cma_nb = {
4214 .notifier_call = cma_netdev_callback
4217 static void cma_add_one(struct ib_device *device)
4219 struct cma_device *cma_dev;
4220 struct rdma_id_private *id_priv;
4222 unsigned long supported_gids = 0;
4224 cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
4228 cma_dev->device = device;
4229 cma_dev->default_gid_type = kcalloc(device->phys_port_cnt,
4230 sizeof(*cma_dev->default_gid_type),
4232 if (!cma_dev->default_gid_type) {
4236 for (i = rdma_start_port(device); i <= rdma_end_port(device); i++) {
4237 supported_gids = roce_gid_type_mask_support(device, i);
4238 WARN_ON(!supported_gids);
4239 cma_dev->default_gid_type[i - rdma_start_port(device)] =
4240 find_first_bit(&supported_gids, BITS_PER_LONG);
4243 init_completion(&cma_dev->comp);
4244 atomic_set(&cma_dev->refcount, 1);
4245 INIT_LIST_HEAD(&cma_dev->id_list);
4246 ib_set_client_data(device, &cma_client, cma_dev);
4249 list_add_tail(&cma_dev->list, &dev_list);
4250 list_for_each_entry(id_priv, &listen_any_list, list)
4251 cma_listen_on_dev(id_priv, cma_dev);
4252 mutex_unlock(&lock);
4255 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
4257 struct rdma_cm_event event;
4258 enum rdma_cm_state state;
4261 /* Record that we want to remove the device */
4262 state = cma_exch(id_priv, RDMA_CM_DEVICE_REMOVAL);
4263 if (state == RDMA_CM_DESTROYING)
4266 cma_cancel_operation(id_priv, state);
4267 mutex_lock(&id_priv->handler_mutex);
4269 /* Check for destruction from another callback. */
4270 if (!cma_comp(id_priv, RDMA_CM_DEVICE_REMOVAL))
4273 memset(&event, 0, sizeof event);
4274 event.event = RDMA_CM_EVENT_DEVICE_REMOVAL;
4275 ret = id_priv->id.event_handler(&id_priv->id, &event);
4277 mutex_unlock(&id_priv->handler_mutex);
4281 static void cma_process_remove(struct cma_device *cma_dev)
4283 struct rdma_id_private *id_priv;
4287 while (!list_empty(&cma_dev->id_list)) {
4288 id_priv = list_entry(cma_dev->id_list.next,
4289 struct rdma_id_private, list);
4291 list_del(&id_priv->listen_list);
4292 list_del_init(&id_priv->list);
4293 atomic_inc(&id_priv->refcount);
4294 mutex_unlock(&lock);
4296 ret = id_priv->internal_id ? 1 : cma_remove_id_dev(id_priv);
4297 cma_deref_id(id_priv);
4299 rdma_destroy_id(&id_priv->id);
4303 mutex_unlock(&lock);
4305 cma_deref_dev(cma_dev);
4306 wait_for_completion(&cma_dev->comp);
4309 static void cma_remove_one(struct ib_device *device, void *client_data)
4311 struct cma_device *cma_dev = client_data;
4317 list_del(&cma_dev->list);
4318 mutex_unlock(&lock);
4320 cma_process_remove(cma_dev);
4321 kfree(cma_dev->default_gid_type);
4325 static int cma_get_id_stats(struct sk_buff *skb, struct netlink_callback *cb)
4327 struct nlmsghdr *nlh;
4328 struct rdma_cm_id_stats *id_stats;
4329 struct rdma_id_private *id_priv;
4330 struct rdma_cm_id *id = NULL;
4331 struct cma_device *cma_dev;
4332 int i_dev = 0, i_id = 0;
4335 * We export all of the IDs as a sequence of messages. Each
4336 * ID gets its own netlink message.
4340 list_for_each_entry(cma_dev, &dev_list, list) {
4341 if (i_dev < cb->args[0]) {
4347 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
4348 if (i_id < cb->args[1]) {
4353 id_stats = ibnl_put_msg(skb, &nlh, cb->nlh->nlmsg_seq,
4354 sizeof *id_stats, RDMA_NL_RDMA_CM,
4355 RDMA_NL_RDMA_CM_ID_STATS,
4360 memset(id_stats, 0, sizeof *id_stats);
4362 id_stats->node_type = id->route.addr.dev_addr.dev_type;
4363 id_stats->port_num = id->port_num;
4364 id_stats->bound_dev_if =
4365 id->route.addr.dev_addr.bound_dev_if;
4367 if (ibnl_put_attr(skb, nlh,
4368 rdma_addr_size(cma_src_addr(id_priv)),
4369 cma_src_addr(id_priv),
4370 RDMA_NL_RDMA_CM_ATTR_SRC_ADDR))
4372 if (ibnl_put_attr(skb, nlh,
4373 rdma_addr_size(cma_src_addr(id_priv)),
4374 cma_dst_addr(id_priv),
4375 RDMA_NL_RDMA_CM_ATTR_DST_ADDR))
4378 id_stats->pid = id_priv->owner;
4379 id_stats->port_space = id->ps;
4380 id_stats->cm_state = id_priv->state;
4381 id_stats->qp_num = id_priv->qp_num;
4382 id_stats->qp_type = id->qp_type;
4392 mutex_unlock(&lock);
4393 cb->args[0] = i_dev;
4399 static const struct ibnl_client_cbs cma_cb_table[] = {
4400 [RDMA_NL_RDMA_CM_ID_STATS] = { .dump = cma_get_id_stats,
4401 .module = THIS_MODULE },
4404 static int cma_init_net(struct net *net)
4406 struct cma_pernet *pernet = cma_pernet(net);
4408 idr_init(&pernet->tcp_ps);
4409 idr_init(&pernet->udp_ps);
4410 idr_init(&pernet->ipoib_ps);
4411 idr_init(&pernet->ib_ps);
4416 static void cma_exit_net(struct net *net)
4418 struct cma_pernet *pernet = cma_pernet(net);
4420 idr_destroy(&pernet->tcp_ps);
4421 idr_destroy(&pernet->udp_ps);
4422 idr_destroy(&pernet->ipoib_ps);
4423 idr_destroy(&pernet->ib_ps);
4426 static struct pernet_operations cma_pernet_operations = {
4427 .init = cma_init_net,
4428 .exit = cma_exit_net,
4429 .id = &cma_pernet_id,
4430 .size = sizeof(struct cma_pernet),
4433 static int __init cma_init(void)
4437 cma_wq = alloc_ordered_workqueue("rdma_cm", WQ_MEM_RECLAIM);
4441 ret = register_pernet_subsys(&cma_pernet_operations);
4445 ib_sa_register_client(&sa_client);
4446 rdma_addr_register_client(&addr_client);
4447 register_netdevice_notifier(&cma_nb);
4449 ret = ib_register_client(&cma_client);
4453 if (ibnl_add_client(RDMA_NL_RDMA_CM, ARRAY_SIZE(cma_cb_table),
4455 pr_warn("RDMA CMA: failed to add netlink callback\n");
4456 cma_configfs_init();
4461 unregister_netdevice_notifier(&cma_nb);
4462 rdma_addr_unregister_client(&addr_client);
4463 ib_sa_unregister_client(&sa_client);
4465 destroy_workqueue(cma_wq);
4469 static void __exit cma_cleanup(void)
4471 cma_configfs_exit();
4472 ibnl_remove_client(RDMA_NL_RDMA_CM);
4473 ib_unregister_client(&cma_client);
4474 unregister_netdevice_notifier(&cma_nb);
4475 rdma_addr_unregister_client(&addr_client);
4476 ib_sa_unregister_client(&sa_client);
4477 unregister_pernet_subsys(&cma_pernet_operations);
4478 destroy_workqueue(cma_wq);
4481 module_init(cma_init);
4482 module_exit(cma_cleanup);