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 static void cma_add_one(struct ib_device *device);
105 static void cma_remove_one(struct ib_device *device, void *client_data);
107 static struct ib_client cma_client = {
110 .remove = cma_remove_one
113 static struct ib_sa_client sa_client;
114 static struct rdma_addr_client addr_client;
115 static LIST_HEAD(dev_list);
116 static LIST_HEAD(listen_any_list);
117 static DEFINE_MUTEX(lock);
118 static struct workqueue_struct *cma_wq;
119 static int cma_pernet_id;
128 static struct cma_pernet *cma_pernet(struct net *net)
130 return net_generic(net, cma_pernet_id);
133 static struct idr *cma_pernet_idr(struct net *net, enum rdma_port_space ps)
135 struct cma_pernet *pernet = cma_pernet(net);
139 return &pernet->tcp_ps;
141 return &pernet->udp_ps;
143 return &pernet->ipoib_ps;
145 return &pernet->ib_ps;
152 struct list_head list;
153 struct ib_device *device;
154 struct completion comp;
156 struct list_head id_list;
157 enum ib_gid_type *default_gid_type;
160 struct rdma_bind_list {
161 enum rdma_port_space ps;
162 struct hlist_head owners;
166 struct class_port_info_context {
167 struct ib_class_port_info *class_port_info;
168 struct ib_device *device;
169 struct completion done;
170 struct ib_sa_query *sa_query;
174 static int cma_ps_alloc(struct net *net, enum rdma_port_space ps,
175 struct rdma_bind_list *bind_list, int snum)
177 struct idr *idr = cma_pernet_idr(net, ps);
179 return idr_alloc(idr, bind_list, snum, snum + 1, GFP_KERNEL);
182 static struct rdma_bind_list *cma_ps_find(struct net *net,
183 enum rdma_port_space ps, int snum)
185 struct idr *idr = cma_pernet_idr(net, ps);
187 return idr_find(idr, snum);
190 static void cma_ps_remove(struct net *net, enum rdma_port_space ps, int snum)
192 struct idr *idr = cma_pernet_idr(net, ps);
194 idr_remove(idr, snum);
201 void cma_ref_dev(struct cma_device *cma_dev)
203 atomic_inc(&cma_dev->refcount);
206 struct cma_device *cma_enum_devices_by_ibdev(cma_device_filter filter,
209 struct cma_device *cma_dev;
210 struct cma_device *found_cma_dev = NULL;
214 list_for_each_entry(cma_dev, &dev_list, list)
215 if (filter(cma_dev->device, cookie)) {
216 found_cma_dev = cma_dev;
221 cma_ref_dev(found_cma_dev);
223 return found_cma_dev;
226 int cma_get_default_gid_type(struct cma_device *cma_dev,
229 if (port < rdma_start_port(cma_dev->device) ||
230 port > rdma_end_port(cma_dev->device))
233 return cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)];
236 int cma_set_default_gid_type(struct cma_device *cma_dev,
238 enum ib_gid_type default_gid_type)
240 unsigned long supported_gids;
242 if (port < rdma_start_port(cma_dev->device) ||
243 port > rdma_end_port(cma_dev->device))
246 supported_gids = roce_gid_type_mask_support(cma_dev->device, port);
248 if (!(supported_gids & 1 << default_gid_type))
251 cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)] =
257 struct ib_device *cma_get_ib_dev(struct cma_device *cma_dev)
259 return cma_dev->device;
263 * Device removal can occur at anytime, so we need extra handling to
264 * serialize notifying the user of device removal with other callbacks.
265 * We do this by disabling removal notification while a callback is in process,
266 * and reporting it after the callback completes.
268 struct rdma_id_private {
269 struct rdma_cm_id id;
271 struct rdma_bind_list *bind_list;
272 struct hlist_node node;
273 struct list_head list; /* listen_any_list or cma_device.list */
274 struct list_head listen_list; /* per device listens */
275 struct cma_device *cma_dev;
276 struct list_head mc_list;
279 enum rdma_cm_state state;
281 struct mutex qp_mutex;
283 struct completion comp;
285 struct mutex handler_mutex;
289 struct ib_sa_query *query;
305 enum ib_gid_type gid_type;
308 struct cma_multicast {
309 struct rdma_id_private *id_priv;
311 struct ib_sa_multicast *ib;
313 struct list_head list;
315 struct sockaddr_storage addr;
322 struct work_struct work;
323 struct rdma_id_private *id;
324 enum rdma_cm_state old_state;
325 enum rdma_cm_state new_state;
326 struct rdma_cm_event event;
329 struct cma_ndev_work {
330 struct work_struct work;
331 struct rdma_id_private *id;
332 struct rdma_cm_event event;
335 struct iboe_mcast_work {
336 struct work_struct work;
337 struct rdma_id_private *id;
338 struct cma_multicast *mc;
351 u8 ip_version; /* IP version: 7:4 */
353 union cma_ip_addr src_addr;
354 union cma_ip_addr dst_addr;
357 #define CMA_VERSION 0x00
359 struct cma_req_info {
360 struct ib_device *device;
362 union ib_gid local_gid;
368 static int cma_comp(struct rdma_id_private *id_priv, enum rdma_cm_state comp)
373 spin_lock_irqsave(&id_priv->lock, flags);
374 ret = (id_priv->state == comp);
375 spin_unlock_irqrestore(&id_priv->lock, flags);
379 static int cma_comp_exch(struct rdma_id_private *id_priv,
380 enum rdma_cm_state comp, enum rdma_cm_state exch)
385 spin_lock_irqsave(&id_priv->lock, flags);
386 if ((ret = (id_priv->state == comp)))
387 id_priv->state = exch;
388 spin_unlock_irqrestore(&id_priv->lock, flags);
392 static enum rdma_cm_state cma_exch(struct rdma_id_private *id_priv,
393 enum rdma_cm_state exch)
396 enum rdma_cm_state old;
398 spin_lock_irqsave(&id_priv->lock, flags);
399 old = id_priv->state;
400 id_priv->state = exch;
401 spin_unlock_irqrestore(&id_priv->lock, flags);
405 static inline u8 cma_get_ip_ver(const struct cma_hdr *hdr)
407 return hdr->ip_version >> 4;
410 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
412 hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
415 static int cma_igmp_send(struct net_device *ndev, union ib_gid *mgid, bool join)
417 struct in_device *in_dev = NULL;
421 in_dev = __in_dev_get_rtnl(ndev);
424 ip_mc_inc_group(in_dev,
425 *(__be32 *)(mgid->raw + 12));
427 ip_mc_dec_group(in_dev,
428 *(__be32 *)(mgid->raw + 12));
432 return (in_dev) ? 0 : -ENODEV;
435 static void _cma_attach_to_dev(struct rdma_id_private *id_priv,
436 struct cma_device *cma_dev)
438 cma_ref_dev(cma_dev);
439 id_priv->cma_dev = cma_dev;
440 id_priv->gid_type = 0;
441 id_priv->id.device = cma_dev->device;
442 id_priv->id.route.addr.dev_addr.transport =
443 rdma_node_get_transport(cma_dev->device->node_type);
444 list_add_tail(&id_priv->list, &cma_dev->id_list);
447 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
448 struct cma_device *cma_dev)
450 _cma_attach_to_dev(id_priv, cma_dev);
452 cma_dev->default_gid_type[id_priv->id.port_num -
453 rdma_start_port(cma_dev->device)];
456 void cma_deref_dev(struct cma_device *cma_dev)
458 if (atomic_dec_and_test(&cma_dev->refcount))
459 complete(&cma_dev->comp);
462 static inline void release_mc(struct kref *kref)
464 struct cma_multicast *mc = container_of(kref, struct cma_multicast, mcref);
466 kfree(mc->multicast.ib);
470 static void cma_release_dev(struct rdma_id_private *id_priv)
473 list_del(&id_priv->list);
474 cma_deref_dev(id_priv->cma_dev);
475 id_priv->cma_dev = NULL;
479 static inline struct sockaddr *cma_src_addr(struct rdma_id_private *id_priv)
481 return (struct sockaddr *) &id_priv->id.route.addr.src_addr;
484 static inline struct sockaddr *cma_dst_addr(struct rdma_id_private *id_priv)
486 return (struct sockaddr *) &id_priv->id.route.addr.dst_addr;
489 static inline unsigned short cma_family(struct rdma_id_private *id_priv)
491 return id_priv->id.route.addr.src_addr.ss_family;
494 static int cma_set_qkey(struct rdma_id_private *id_priv, u32 qkey)
496 struct ib_sa_mcmember_rec rec;
500 if (qkey && id_priv->qkey != qkey)
506 id_priv->qkey = qkey;
510 switch (id_priv->id.ps) {
513 id_priv->qkey = RDMA_UDP_QKEY;
516 ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid);
517 ret = ib_sa_get_mcmember_rec(id_priv->id.device,
518 id_priv->id.port_num, &rec.mgid,
521 id_priv->qkey = be32_to_cpu(rec.qkey);
529 static void cma_translate_ib(struct sockaddr_ib *sib, struct rdma_dev_addr *dev_addr)
531 dev_addr->dev_type = ARPHRD_INFINIBAND;
532 rdma_addr_set_sgid(dev_addr, (union ib_gid *) &sib->sib_addr);
533 ib_addr_set_pkey(dev_addr, ntohs(sib->sib_pkey));
536 static int cma_translate_addr(struct sockaddr *addr, struct rdma_dev_addr *dev_addr)
540 if (addr->sa_family != AF_IB) {
541 ret = rdma_translate_ip(addr, dev_addr, NULL);
543 cma_translate_ib((struct sockaddr_ib *) addr, dev_addr);
550 static inline int cma_validate_port(struct ib_device *device, u8 port,
551 enum ib_gid_type gid_type,
552 union ib_gid *gid, int dev_type,
556 struct net_device *ndev = NULL;
558 if ((dev_type == ARPHRD_INFINIBAND) && !rdma_protocol_ib(device, port))
561 if ((dev_type != ARPHRD_INFINIBAND) && rdma_protocol_ib(device, port))
564 if (dev_type == ARPHRD_ETHER && rdma_protocol_roce(device, port)) {
565 ndev = dev_get_by_index(&init_net, bound_if_index);
566 if (ndev && ndev->flags & IFF_LOOPBACK) {
567 pr_info("detected loopback device\n");
570 if (!device->get_netdev)
573 ndev = device->get_netdev(device, port);
578 gid_type = IB_GID_TYPE_IB;
581 ret = ib_find_cached_gid_by_port(device, gid, gid_type, port,
590 static int cma_acquire_dev(struct rdma_id_private *id_priv,
591 struct rdma_id_private *listen_id_priv)
593 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
594 struct cma_device *cma_dev;
595 union ib_gid gid, iboe_gid, *gidp;
599 if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
600 id_priv->id.ps == RDMA_PS_IPOIB)
604 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
607 memcpy(&gid, dev_addr->src_dev_addr +
608 rdma_addr_gid_offset(dev_addr), sizeof gid);
610 if (listen_id_priv) {
611 cma_dev = listen_id_priv->cma_dev;
612 port = listen_id_priv->id.port_num;
613 gidp = rdma_protocol_roce(cma_dev->device, port) ?
616 ret = cma_validate_port(cma_dev->device, port,
617 rdma_protocol_ib(cma_dev->device, port) ?
619 listen_id_priv->gid_type, gidp,
621 dev_addr->bound_dev_if);
623 id_priv->id.port_num = port;
628 list_for_each_entry(cma_dev, &dev_list, list) {
629 for (port = 1; port <= cma_dev->device->phys_port_cnt; ++port) {
630 if (listen_id_priv &&
631 listen_id_priv->cma_dev == cma_dev &&
632 listen_id_priv->id.port_num == port)
635 gidp = rdma_protocol_roce(cma_dev->device, port) ?
638 ret = cma_validate_port(cma_dev->device, port,
639 rdma_protocol_ib(cma_dev->device, port) ?
641 cma_dev->default_gid_type[port - 1],
642 gidp, dev_addr->dev_type,
643 dev_addr->bound_dev_if);
645 id_priv->id.port_num = port;
653 cma_attach_to_dev(id_priv, cma_dev);
660 * Select the source IB device and address to reach the destination IB address.
662 static int cma_resolve_ib_dev(struct rdma_id_private *id_priv)
664 struct cma_device *cma_dev, *cur_dev;
665 struct sockaddr_ib *addr;
666 union ib_gid gid, sgid, *dgid;
672 addr = (struct sockaddr_ib *) cma_dst_addr(id_priv);
673 dgid = (union ib_gid *) &addr->sib_addr;
674 pkey = ntohs(addr->sib_pkey);
676 list_for_each_entry(cur_dev, &dev_list, list) {
677 for (p = 1; p <= cur_dev->device->phys_port_cnt; ++p) {
678 if (!rdma_cap_af_ib(cur_dev->device, p))
681 if (ib_find_cached_pkey(cur_dev->device, p, pkey, &index))
684 for (i = 0; !ib_get_cached_gid(cur_dev->device, p, i,
687 if (!memcmp(&gid, dgid, sizeof(gid))) {
690 id_priv->id.port_num = p;
694 if (!cma_dev && (gid.global.subnet_prefix ==
695 dgid->global.subnet_prefix)) {
698 id_priv->id.port_num = p;
708 cma_attach_to_dev(id_priv, cma_dev);
709 addr = (struct sockaddr_ib *) cma_src_addr(id_priv);
710 memcpy(&addr->sib_addr, &sgid, sizeof sgid);
711 cma_translate_ib(addr, &id_priv->id.route.addr.dev_addr);
715 static void cma_deref_id(struct rdma_id_private *id_priv)
717 if (atomic_dec_and_test(&id_priv->refcount))
718 complete(&id_priv->comp);
721 struct rdma_cm_id *rdma_create_id(struct net *net,
722 rdma_cm_event_handler event_handler,
723 void *context, enum rdma_port_space ps,
724 enum ib_qp_type qp_type)
726 struct rdma_id_private *id_priv;
728 id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
730 return ERR_PTR(-ENOMEM);
732 id_priv->owner = task_pid_nr(current);
733 id_priv->state = RDMA_CM_IDLE;
734 id_priv->id.context = context;
735 id_priv->id.event_handler = event_handler;
737 id_priv->id.qp_type = qp_type;
738 spin_lock_init(&id_priv->lock);
739 mutex_init(&id_priv->qp_mutex);
740 init_completion(&id_priv->comp);
741 atomic_set(&id_priv->refcount, 1);
742 mutex_init(&id_priv->handler_mutex);
743 INIT_LIST_HEAD(&id_priv->listen_list);
744 INIT_LIST_HEAD(&id_priv->mc_list);
745 get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
746 id_priv->id.route.addr.dev_addr.net = get_net(net);
750 EXPORT_SYMBOL(rdma_create_id);
752 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
754 struct ib_qp_attr qp_attr;
755 int qp_attr_mask, ret;
757 qp_attr.qp_state = IB_QPS_INIT;
758 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
762 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
766 qp_attr.qp_state = IB_QPS_RTR;
767 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
771 qp_attr.qp_state = IB_QPS_RTS;
773 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
778 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
780 struct ib_qp_attr qp_attr;
781 int qp_attr_mask, ret;
783 qp_attr.qp_state = IB_QPS_INIT;
784 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
788 return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
791 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
792 struct ib_qp_init_attr *qp_init_attr)
794 struct rdma_id_private *id_priv;
798 id_priv = container_of(id, struct rdma_id_private, id);
799 if (id->device != pd->device)
802 qp_init_attr->port_num = id->port_num;
803 qp = ib_create_qp(pd, qp_init_attr);
807 if (id->qp_type == IB_QPT_UD)
808 ret = cma_init_ud_qp(id_priv, qp);
810 ret = cma_init_conn_qp(id_priv, qp);
815 id_priv->qp_num = qp->qp_num;
816 id_priv->srq = (qp->srq != NULL);
822 EXPORT_SYMBOL(rdma_create_qp);
824 void rdma_destroy_qp(struct rdma_cm_id *id)
826 struct rdma_id_private *id_priv;
828 id_priv = container_of(id, struct rdma_id_private, id);
829 mutex_lock(&id_priv->qp_mutex);
830 ib_destroy_qp(id_priv->id.qp);
831 id_priv->id.qp = NULL;
832 mutex_unlock(&id_priv->qp_mutex);
834 EXPORT_SYMBOL(rdma_destroy_qp);
836 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
837 struct rdma_conn_param *conn_param)
839 struct ib_qp_attr qp_attr;
840 int qp_attr_mask, ret;
843 mutex_lock(&id_priv->qp_mutex);
844 if (!id_priv->id.qp) {
849 /* Need to update QP attributes from default values. */
850 qp_attr.qp_state = IB_QPS_INIT;
851 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
855 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
859 qp_attr.qp_state = IB_QPS_RTR;
860 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
864 ret = ib_query_gid(id_priv->id.device, id_priv->id.port_num,
865 qp_attr.ah_attr.grh.sgid_index, &sgid, NULL);
869 BUG_ON(id_priv->cma_dev->device != id_priv->id.device);
872 qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
873 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
875 mutex_unlock(&id_priv->qp_mutex);
879 static int cma_modify_qp_rts(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;
885 mutex_lock(&id_priv->qp_mutex);
886 if (!id_priv->id.qp) {
891 qp_attr.qp_state = IB_QPS_RTS;
892 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
897 qp_attr.max_rd_atomic = conn_param->initiator_depth;
898 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
900 mutex_unlock(&id_priv->qp_mutex);
904 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
906 struct ib_qp_attr qp_attr;
909 mutex_lock(&id_priv->qp_mutex);
910 if (!id_priv->id.qp) {
915 qp_attr.qp_state = IB_QPS_ERR;
916 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
918 mutex_unlock(&id_priv->qp_mutex);
922 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
923 struct ib_qp_attr *qp_attr, int *qp_attr_mask)
925 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
929 if (rdma_cap_eth_ah(id_priv->id.device, id_priv->id.port_num))
932 pkey = ib_addr_get_pkey(dev_addr);
934 ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
935 pkey, &qp_attr->pkey_index);
939 qp_attr->port_num = id_priv->id.port_num;
940 *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
942 if (id_priv->id.qp_type == IB_QPT_UD) {
943 ret = cma_set_qkey(id_priv, 0);
947 qp_attr->qkey = id_priv->qkey;
948 *qp_attr_mask |= IB_QP_QKEY;
950 qp_attr->qp_access_flags = 0;
951 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
956 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
959 struct rdma_id_private *id_priv;
962 id_priv = container_of(id, struct rdma_id_private, id);
963 if (rdma_cap_ib_cm(id->device, id->port_num)) {
964 if (!id_priv->cm_id.ib || (id_priv->id.qp_type == IB_QPT_UD))
965 ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
967 ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
970 if (qp_attr->qp_state == IB_QPS_RTR)
971 qp_attr->rq_psn = id_priv->seq_num;
972 } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
973 if (!id_priv->cm_id.iw) {
974 qp_attr->qp_access_flags = 0;
975 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
977 ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
984 EXPORT_SYMBOL(rdma_init_qp_attr);
986 static inline int cma_zero_addr(struct sockaddr *addr)
988 switch (addr->sa_family) {
990 return ipv4_is_zeronet(((struct sockaddr_in *)addr)->sin_addr.s_addr);
992 return ipv6_addr_any(&((struct sockaddr_in6 *) addr)->sin6_addr);
994 return ib_addr_any(&((struct sockaddr_ib *) addr)->sib_addr);
1000 static inline int cma_loopback_addr(struct sockaddr *addr)
1002 switch (addr->sa_family) {
1004 return ipv4_is_loopback(((struct sockaddr_in *) addr)->sin_addr.s_addr);
1006 return ipv6_addr_loopback(&((struct sockaddr_in6 *) addr)->sin6_addr);
1008 return ib_addr_loopback(&((struct sockaddr_ib *) addr)->sib_addr);
1014 static inline int cma_any_addr(struct sockaddr *addr)
1016 return cma_zero_addr(addr) || cma_loopback_addr(addr);
1019 static int cma_addr_cmp(struct sockaddr *src, struct sockaddr *dst)
1021 if (src->sa_family != dst->sa_family)
1024 switch (src->sa_family) {
1026 return ((struct sockaddr_in *) src)->sin_addr.s_addr !=
1027 ((struct sockaddr_in *) dst)->sin_addr.s_addr;
1029 return ipv6_addr_cmp(&((struct sockaddr_in6 *) src)->sin6_addr,
1030 &((struct sockaddr_in6 *) dst)->sin6_addr);
1032 return ib_addr_cmp(&((struct sockaddr_ib *) src)->sib_addr,
1033 &((struct sockaddr_ib *) dst)->sib_addr);
1037 static __be16 cma_port(struct sockaddr *addr)
1039 struct sockaddr_ib *sib;
1041 switch (addr->sa_family) {
1043 return ((struct sockaddr_in *) addr)->sin_port;
1045 return ((struct sockaddr_in6 *) addr)->sin6_port;
1047 sib = (struct sockaddr_ib *) addr;
1048 return htons((u16) (be64_to_cpu(sib->sib_sid) &
1049 be64_to_cpu(sib->sib_sid_mask)));
1055 static inline int cma_any_port(struct sockaddr *addr)
1057 return !cma_port(addr);
1060 static void cma_save_ib_info(struct sockaddr *src_addr,
1061 struct sockaddr *dst_addr,
1062 struct rdma_cm_id *listen_id,
1063 struct ib_sa_path_rec *path)
1065 struct sockaddr_ib *listen_ib, *ib;
1067 listen_ib = (struct sockaddr_ib *) &listen_id->route.addr.src_addr;
1069 ib = (struct sockaddr_ib *)src_addr;
1070 ib->sib_family = AF_IB;
1072 ib->sib_pkey = path->pkey;
1073 ib->sib_flowinfo = path->flow_label;
1074 memcpy(&ib->sib_addr, &path->sgid, 16);
1075 ib->sib_sid = path->service_id;
1076 ib->sib_scope_id = 0;
1078 ib->sib_pkey = listen_ib->sib_pkey;
1079 ib->sib_flowinfo = listen_ib->sib_flowinfo;
1080 ib->sib_addr = listen_ib->sib_addr;
1081 ib->sib_sid = listen_ib->sib_sid;
1082 ib->sib_scope_id = listen_ib->sib_scope_id;
1084 ib->sib_sid_mask = cpu_to_be64(0xffffffffffffffffULL);
1087 ib = (struct sockaddr_ib *)dst_addr;
1088 ib->sib_family = AF_IB;
1090 ib->sib_pkey = path->pkey;
1091 ib->sib_flowinfo = path->flow_label;
1092 memcpy(&ib->sib_addr, &path->dgid, 16);
1097 static void cma_save_ip4_info(struct sockaddr *src_addr,
1098 struct sockaddr *dst_addr,
1099 struct cma_hdr *hdr,
1102 struct sockaddr_in *ip4;
1105 ip4 = (struct sockaddr_in *)src_addr;
1106 ip4->sin_family = AF_INET;
1107 ip4->sin_addr.s_addr = hdr->dst_addr.ip4.addr;
1108 ip4->sin_port = local_port;
1112 ip4 = (struct sockaddr_in *)dst_addr;
1113 ip4->sin_family = AF_INET;
1114 ip4->sin_addr.s_addr = hdr->src_addr.ip4.addr;
1115 ip4->sin_port = hdr->port;
1119 static void cma_save_ip6_info(struct sockaddr *src_addr,
1120 struct sockaddr *dst_addr,
1121 struct cma_hdr *hdr,
1124 struct sockaddr_in6 *ip6;
1127 ip6 = (struct sockaddr_in6 *)src_addr;
1128 ip6->sin6_family = AF_INET6;
1129 ip6->sin6_addr = hdr->dst_addr.ip6;
1130 ip6->sin6_port = local_port;
1134 ip6 = (struct sockaddr_in6 *)dst_addr;
1135 ip6->sin6_family = AF_INET6;
1136 ip6->sin6_addr = hdr->src_addr.ip6;
1137 ip6->sin6_port = hdr->port;
1141 static u16 cma_port_from_service_id(__be64 service_id)
1143 return (u16)be64_to_cpu(service_id);
1146 static int cma_save_ip_info(struct sockaddr *src_addr,
1147 struct sockaddr *dst_addr,
1148 struct ib_cm_event *ib_event,
1151 struct cma_hdr *hdr;
1154 hdr = ib_event->private_data;
1155 if (hdr->cma_version != CMA_VERSION)
1158 port = htons(cma_port_from_service_id(service_id));
1160 switch (cma_get_ip_ver(hdr)) {
1162 cma_save_ip4_info(src_addr, dst_addr, hdr, port);
1165 cma_save_ip6_info(src_addr, dst_addr, hdr, port);
1168 return -EAFNOSUPPORT;
1174 static int cma_save_net_info(struct sockaddr *src_addr,
1175 struct sockaddr *dst_addr,
1176 struct rdma_cm_id *listen_id,
1177 struct ib_cm_event *ib_event,
1178 sa_family_t sa_family, __be64 service_id)
1180 if (sa_family == AF_IB) {
1181 if (ib_event->event == IB_CM_REQ_RECEIVED)
1182 cma_save_ib_info(src_addr, dst_addr, listen_id,
1183 ib_event->param.req_rcvd.primary_path);
1184 else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED)
1185 cma_save_ib_info(src_addr, dst_addr, listen_id, NULL);
1189 return cma_save_ip_info(src_addr, dst_addr, ib_event, service_id);
1192 static int cma_save_req_info(const struct ib_cm_event *ib_event,
1193 struct cma_req_info *req)
1195 const struct ib_cm_req_event_param *req_param =
1196 &ib_event->param.req_rcvd;
1197 const struct ib_cm_sidr_req_event_param *sidr_param =
1198 &ib_event->param.sidr_req_rcvd;
1200 switch (ib_event->event) {
1201 case IB_CM_REQ_RECEIVED:
1202 req->device = req_param->listen_id->device;
1203 req->port = req_param->port;
1204 memcpy(&req->local_gid, &req_param->primary_path->sgid,
1205 sizeof(req->local_gid));
1206 req->has_gid = true;
1207 req->service_id = req_param->primary_path->service_id;
1208 req->pkey = be16_to_cpu(req_param->primary_path->pkey);
1209 if (req->pkey != req_param->bth_pkey)
1210 pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and primary path P_Key (0x%x)\n"
1211 "RDMA CMA: in the future this may cause the request to be dropped\n",
1212 req_param->bth_pkey, req->pkey);
1214 case IB_CM_SIDR_REQ_RECEIVED:
1215 req->device = sidr_param->listen_id->device;
1216 req->port = sidr_param->port;
1217 req->has_gid = false;
1218 req->service_id = sidr_param->service_id;
1219 req->pkey = sidr_param->pkey;
1220 if (req->pkey != sidr_param->bth_pkey)
1221 pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and SIDR request payload P_Key (0x%x)\n"
1222 "RDMA CMA: in the future this may cause the request to be dropped\n",
1223 sidr_param->bth_pkey, req->pkey);
1232 static bool validate_ipv4_net_dev(struct net_device *net_dev,
1233 const struct sockaddr_in *dst_addr,
1234 const struct sockaddr_in *src_addr)
1236 __be32 daddr = dst_addr->sin_addr.s_addr,
1237 saddr = src_addr->sin_addr.s_addr;
1238 struct fib_result res;
1243 if (ipv4_is_multicast(saddr) || ipv4_is_lbcast(saddr) ||
1244 ipv4_is_lbcast(daddr) || ipv4_is_zeronet(saddr) ||
1245 ipv4_is_zeronet(daddr) || ipv4_is_loopback(daddr) ||
1246 ipv4_is_loopback(saddr))
1249 memset(&fl4, 0, sizeof(fl4));
1250 fl4.flowi4_iif = net_dev->ifindex;
1255 err = fib_lookup(dev_net(net_dev), &fl4, &res, 0);
1256 ret = err == 0 && FIB_RES_DEV(res) == net_dev;
1262 static bool validate_ipv6_net_dev(struct net_device *net_dev,
1263 const struct sockaddr_in6 *dst_addr,
1264 const struct sockaddr_in6 *src_addr)
1266 #if IS_ENABLED(CONFIG_IPV6)
1267 const int strict = ipv6_addr_type(&dst_addr->sin6_addr) &
1268 IPV6_ADDR_LINKLOCAL;
1269 struct rt6_info *rt = rt6_lookup(dev_net(net_dev), &dst_addr->sin6_addr,
1270 &src_addr->sin6_addr, net_dev->ifindex,
1277 ret = rt->rt6i_idev->dev == net_dev;
1286 static bool validate_net_dev(struct net_device *net_dev,
1287 const struct sockaddr *daddr,
1288 const struct sockaddr *saddr)
1290 const struct sockaddr_in *daddr4 = (const struct sockaddr_in *)daddr;
1291 const struct sockaddr_in *saddr4 = (const struct sockaddr_in *)saddr;
1292 const struct sockaddr_in6 *daddr6 = (const struct sockaddr_in6 *)daddr;
1293 const struct sockaddr_in6 *saddr6 = (const struct sockaddr_in6 *)saddr;
1295 switch (daddr->sa_family) {
1297 return saddr->sa_family == AF_INET &&
1298 validate_ipv4_net_dev(net_dev, daddr4, saddr4);
1301 return saddr->sa_family == AF_INET6 &&
1302 validate_ipv6_net_dev(net_dev, daddr6, saddr6);
1309 static struct net_device *cma_get_net_dev(struct ib_cm_event *ib_event,
1310 const struct cma_req_info *req)
1312 struct sockaddr_storage listen_addr_storage, src_addr_storage;
1313 struct sockaddr *listen_addr = (struct sockaddr *)&listen_addr_storage,
1314 *src_addr = (struct sockaddr *)&src_addr_storage;
1315 struct net_device *net_dev;
1316 const union ib_gid *gid = req->has_gid ? &req->local_gid : NULL;
1319 err = cma_save_ip_info(listen_addr, src_addr, ib_event,
1322 return ERR_PTR(err);
1324 net_dev = ib_get_net_dev_by_params(req->device, req->port, req->pkey,
1327 return ERR_PTR(-ENODEV);
1329 if (!validate_net_dev(net_dev, listen_addr, src_addr)) {
1331 return ERR_PTR(-EHOSTUNREACH);
1337 static enum rdma_port_space rdma_ps_from_service_id(__be64 service_id)
1339 return (be64_to_cpu(service_id) >> 16) & 0xffff;
1342 static bool cma_match_private_data(struct rdma_id_private *id_priv,
1343 const struct cma_hdr *hdr)
1345 struct sockaddr *addr = cma_src_addr(id_priv);
1347 struct in6_addr ip6_addr;
1349 if (cma_any_addr(addr) && !id_priv->afonly)
1352 switch (addr->sa_family) {
1354 ip4_addr = ((struct sockaddr_in *)addr)->sin_addr.s_addr;
1355 if (cma_get_ip_ver(hdr) != 4)
1357 if (!cma_any_addr(addr) &&
1358 hdr->dst_addr.ip4.addr != ip4_addr)
1362 ip6_addr = ((struct sockaddr_in6 *)addr)->sin6_addr;
1363 if (cma_get_ip_ver(hdr) != 6)
1365 if (!cma_any_addr(addr) &&
1366 memcmp(&hdr->dst_addr.ip6, &ip6_addr, sizeof(ip6_addr)))
1378 static bool cma_protocol_roce_dev_port(struct ib_device *device, int port_num)
1380 enum rdma_link_layer ll = rdma_port_get_link_layer(device, port_num);
1381 enum rdma_transport_type transport =
1382 rdma_node_get_transport(device->node_type);
1384 return ll == IB_LINK_LAYER_ETHERNET && transport == RDMA_TRANSPORT_IB;
1387 static bool cma_protocol_roce(const struct rdma_cm_id *id)
1389 struct ib_device *device = id->device;
1390 const int port_num = id->port_num ?: rdma_start_port(device);
1392 return cma_protocol_roce_dev_port(device, port_num);
1395 static bool cma_match_net_dev(const struct rdma_cm_id *id,
1396 const struct net_device *net_dev,
1399 const struct rdma_addr *addr = &id->route.addr;
1402 /* This request is an AF_IB request or a RoCE request */
1403 return (!id->port_num || id->port_num == port_num) &&
1404 (addr->src_addr.ss_family == AF_IB ||
1405 cma_protocol_roce_dev_port(id->device, port_num));
1407 return !addr->dev_addr.bound_dev_if ||
1408 (net_eq(dev_net(net_dev), addr->dev_addr.net) &&
1409 addr->dev_addr.bound_dev_if == net_dev->ifindex);
1412 static struct rdma_id_private *cma_find_listener(
1413 const struct rdma_bind_list *bind_list,
1414 const struct ib_cm_id *cm_id,
1415 const struct ib_cm_event *ib_event,
1416 const struct cma_req_info *req,
1417 const struct net_device *net_dev)
1419 struct rdma_id_private *id_priv, *id_priv_dev;
1422 return ERR_PTR(-EINVAL);
1424 hlist_for_each_entry(id_priv, &bind_list->owners, node) {
1425 if (cma_match_private_data(id_priv, ib_event->private_data)) {
1426 if (id_priv->id.device == cm_id->device &&
1427 cma_match_net_dev(&id_priv->id, net_dev, req->port))
1429 list_for_each_entry(id_priv_dev,
1430 &id_priv->listen_list,
1432 if (id_priv_dev->id.device == cm_id->device &&
1433 cma_match_net_dev(&id_priv_dev->id, net_dev, req->port))
1439 return ERR_PTR(-EINVAL);
1442 static struct rdma_id_private *cma_id_from_event(struct ib_cm_id *cm_id,
1443 struct ib_cm_event *ib_event,
1444 struct net_device **net_dev)
1446 struct cma_req_info req;
1447 struct rdma_bind_list *bind_list;
1448 struct rdma_id_private *id_priv;
1451 err = cma_save_req_info(ib_event, &req);
1453 return ERR_PTR(err);
1455 *net_dev = cma_get_net_dev(ib_event, &req);
1456 if (IS_ERR(*net_dev)) {
1457 if (PTR_ERR(*net_dev) == -EAFNOSUPPORT) {
1458 /* Assuming the protocol is AF_IB */
1460 } else if (cma_protocol_roce_dev_port(req.device, req.port)) {
1461 /* TODO find the net dev matching the request parameters
1462 * through the RoCE GID table */
1465 return ERR_CAST(*net_dev);
1469 bind_list = cma_ps_find(*net_dev ? dev_net(*net_dev) : &init_net,
1470 rdma_ps_from_service_id(req.service_id),
1471 cma_port_from_service_id(req.service_id));
1472 id_priv = cma_find_listener(bind_list, cm_id, ib_event, &req, *net_dev);
1473 if (IS_ERR(id_priv) && *net_dev) {
1481 static inline int cma_user_data_offset(struct rdma_id_private *id_priv)
1483 return cma_family(id_priv) == AF_IB ? 0 : sizeof(struct cma_hdr);
1486 static void cma_cancel_route(struct rdma_id_private *id_priv)
1488 if (rdma_cap_ib_sa(id_priv->id.device, id_priv->id.port_num)) {
1490 ib_sa_cancel_query(id_priv->query_id, id_priv->query);
1494 static void cma_cancel_listens(struct rdma_id_private *id_priv)
1496 struct rdma_id_private *dev_id_priv;
1499 * Remove from listen_any_list to prevent added devices from spawning
1500 * additional listen requests.
1503 list_del(&id_priv->list);
1505 while (!list_empty(&id_priv->listen_list)) {
1506 dev_id_priv = list_entry(id_priv->listen_list.next,
1507 struct rdma_id_private, listen_list);
1508 /* sync with device removal to avoid duplicate destruction */
1509 list_del_init(&dev_id_priv->list);
1510 list_del(&dev_id_priv->listen_list);
1511 mutex_unlock(&lock);
1513 rdma_destroy_id(&dev_id_priv->id);
1516 mutex_unlock(&lock);
1519 static void cma_cancel_operation(struct rdma_id_private *id_priv,
1520 enum rdma_cm_state state)
1523 case RDMA_CM_ADDR_QUERY:
1524 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
1526 case RDMA_CM_ROUTE_QUERY:
1527 cma_cancel_route(id_priv);
1529 case RDMA_CM_LISTEN:
1530 if (cma_any_addr(cma_src_addr(id_priv)) && !id_priv->cma_dev)
1531 cma_cancel_listens(id_priv);
1538 static void cma_release_port(struct rdma_id_private *id_priv)
1540 struct rdma_bind_list *bind_list = id_priv->bind_list;
1541 struct net *net = id_priv->id.route.addr.dev_addr.net;
1547 hlist_del(&id_priv->node);
1548 if (hlist_empty(&bind_list->owners)) {
1549 cma_ps_remove(net, bind_list->ps, bind_list->port);
1552 mutex_unlock(&lock);
1555 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
1557 struct cma_multicast *mc;
1559 while (!list_empty(&id_priv->mc_list)) {
1560 mc = container_of(id_priv->mc_list.next,
1561 struct cma_multicast, list);
1562 list_del(&mc->list);
1563 if (rdma_cap_ib_mcast(id_priv->cma_dev->device,
1564 id_priv->id.port_num)) {
1565 ib_sa_free_multicast(mc->multicast.ib);
1568 if (mc->igmp_joined) {
1569 struct rdma_dev_addr *dev_addr =
1570 &id_priv->id.route.addr.dev_addr;
1571 struct net_device *ndev = NULL;
1573 if (dev_addr->bound_dev_if)
1574 ndev = dev_get_by_index(&init_net,
1575 dev_addr->bound_dev_if);
1578 &mc->multicast.ib->rec.mgid,
1583 kref_put(&mc->mcref, release_mc);
1588 void rdma_destroy_id(struct rdma_cm_id *id)
1590 struct rdma_id_private *id_priv;
1591 enum rdma_cm_state state;
1593 id_priv = container_of(id, struct rdma_id_private, id);
1594 state = cma_exch(id_priv, RDMA_CM_DESTROYING);
1595 cma_cancel_operation(id_priv, state);
1598 * Wait for any active callback to finish. New callbacks will find
1599 * the id_priv state set to destroying and abort.
1601 mutex_lock(&id_priv->handler_mutex);
1602 mutex_unlock(&id_priv->handler_mutex);
1604 if (id_priv->cma_dev) {
1605 if (rdma_cap_ib_cm(id_priv->id.device, 1)) {
1606 if (id_priv->cm_id.ib)
1607 ib_destroy_cm_id(id_priv->cm_id.ib);
1608 } else if (rdma_cap_iw_cm(id_priv->id.device, 1)) {
1609 if (id_priv->cm_id.iw)
1610 iw_destroy_cm_id(id_priv->cm_id.iw);
1612 cma_leave_mc_groups(id_priv);
1613 cma_release_dev(id_priv);
1616 cma_release_port(id_priv);
1617 cma_deref_id(id_priv);
1618 wait_for_completion(&id_priv->comp);
1620 if (id_priv->internal_id)
1621 cma_deref_id(id_priv->id.context);
1623 kfree(id_priv->id.route.path_rec);
1624 put_net(id_priv->id.route.addr.dev_addr.net);
1627 EXPORT_SYMBOL(rdma_destroy_id);
1629 static int cma_rep_recv(struct rdma_id_private *id_priv)
1633 ret = cma_modify_qp_rtr(id_priv, NULL);
1637 ret = cma_modify_qp_rts(id_priv, NULL);
1641 ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
1647 cma_modify_qp_err(id_priv);
1648 ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
1653 static void cma_set_rep_event_data(struct rdma_cm_event *event,
1654 struct ib_cm_rep_event_param *rep_data,
1657 event->param.conn.private_data = private_data;
1658 event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
1659 event->param.conn.responder_resources = rep_data->responder_resources;
1660 event->param.conn.initiator_depth = rep_data->initiator_depth;
1661 event->param.conn.flow_control = rep_data->flow_control;
1662 event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
1663 event->param.conn.srq = rep_data->srq;
1664 event->param.conn.qp_num = rep_data->remote_qpn;
1667 static int cma_ib_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1669 struct rdma_id_private *id_priv = cm_id->context;
1670 struct rdma_cm_event event;
1673 mutex_lock(&id_priv->handler_mutex);
1674 if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
1675 id_priv->state != RDMA_CM_CONNECT) ||
1676 (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
1677 id_priv->state != RDMA_CM_DISCONNECT))
1680 memset(&event, 0, sizeof event);
1681 switch (ib_event->event) {
1682 case IB_CM_REQ_ERROR:
1683 case IB_CM_REP_ERROR:
1684 event.event = RDMA_CM_EVENT_UNREACHABLE;
1685 event.status = -ETIMEDOUT;
1687 case IB_CM_REP_RECEIVED:
1688 if (id_priv->id.qp) {
1689 event.status = cma_rep_recv(id_priv);
1690 event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
1691 RDMA_CM_EVENT_ESTABLISHED;
1693 event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
1695 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
1696 ib_event->private_data);
1698 case IB_CM_RTU_RECEIVED:
1699 case IB_CM_USER_ESTABLISHED:
1700 event.event = RDMA_CM_EVENT_ESTABLISHED;
1702 case IB_CM_DREQ_ERROR:
1703 event.status = -ETIMEDOUT; /* fall through */
1704 case IB_CM_DREQ_RECEIVED:
1705 case IB_CM_DREP_RECEIVED:
1706 if (!cma_comp_exch(id_priv, RDMA_CM_CONNECT,
1707 RDMA_CM_DISCONNECT))
1709 event.event = RDMA_CM_EVENT_DISCONNECTED;
1711 case IB_CM_TIMEWAIT_EXIT:
1712 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
1714 case IB_CM_MRA_RECEIVED:
1717 case IB_CM_REJ_RECEIVED:
1718 cma_modify_qp_err(id_priv);
1719 event.status = ib_event->param.rej_rcvd.reason;
1720 event.event = RDMA_CM_EVENT_REJECTED;
1721 event.param.conn.private_data = ib_event->private_data;
1722 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
1725 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
1730 ret = id_priv->id.event_handler(&id_priv->id, &event);
1732 /* Destroy the CM ID by returning a non-zero value. */
1733 id_priv->cm_id.ib = NULL;
1734 cma_exch(id_priv, RDMA_CM_DESTROYING);
1735 mutex_unlock(&id_priv->handler_mutex);
1736 rdma_destroy_id(&id_priv->id);
1740 mutex_unlock(&id_priv->handler_mutex);
1744 static struct rdma_id_private *cma_new_conn_id(struct rdma_cm_id *listen_id,
1745 struct ib_cm_event *ib_event,
1746 struct net_device *net_dev)
1748 struct rdma_id_private *id_priv;
1749 struct rdma_cm_id *id;
1750 struct rdma_route *rt;
1751 const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
1752 const __be64 service_id =
1753 ib_event->param.req_rcvd.primary_path->service_id;
1756 id = rdma_create_id(listen_id->route.addr.dev_addr.net,
1757 listen_id->event_handler, listen_id->context,
1758 listen_id->ps, ib_event->param.req_rcvd.qp_type);
1762 id_priv = container_of(id, struct rdma_id_private, id);
1763 if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
1764 (struct sockaddr *)&id->route.addr.dst_addr,
1765 listen_id, ib_event, ss_family, service_id))
1769 rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
1770 rt->path_rec = kmalloc(sizeof *rt->path_rec * rt->num_paths,
1775 rt->path_rec[0] = *ib_event->param.req_rcvd.primary_path;
1776 if (rt->num_paths == 2)
1777 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
1780 ret = rdma_copy_addr(&rt->addr.dev_addr, net_dev, NULL);
1784 if (!cma_protocol_roce(listen_id) &&
1785 cma_any_addr(cma_src_addr(id_priv))) {
1786 rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
1787 rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
1788 ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
1789 } else if (!cma_any_addr(cma_src_addr(id_priv))) {
1790 ret = cma_translate_addr(cma_src_addr(id_priv), &rt->addr.dev_addr);
1795 rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
1797 id_priv->state = RDMA_CM_CONNECT;
1801 rdma_destroy_id(id);
1805 static struct rdma_id_private *cma_new_udp_id(struct rdma_cm_id *listen_id,
1806 struct ib_cm_event *ib_event,
1807 struct net_device *net_dev)
1809 struct rdma_id_private *id_priv;
1810 struct rdma_cm_id *id;
1811 const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
1812 struct net *net = listen_id->route.addr.dev_addr.net;
1815 id = rdma_create_id(net, listen_id->event_handler, listen_id->context,
1816 listen_id->ps, IB_QPT_UD);
1820 id_priv = container_of(id, struct rdma_id_private, id);
1821 if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
1822 (struct sockaddr *)&id->route.addr.dst_addr,
1823 listen_id, ib_event, ss_family,
1824 ib_event->param.sidr_req_rcvd.service_id))
1828 ret = rdma_copy_addr(&id->route.addr.dev_addr, net_dev, NULL);
1832 if (!cma_any_addr(cma_src_addr(id_priv))) {
1833 ret = cma_translate_addr(cma_src_addr(id_priv),
1834 &id->route.addr.dev_addr);
1840 id_priv->state = RDMA_CM_CONNECT;
1843 rdma_destroy_id(id);
1847 static void cma_set_req_event_data(struct rdma_cm_event *event,
1848 struct ib_cm_req_event_param *req_data,
1849 void *private_data, int offset)
1851 event->param.conn.private_data = private_data + offset;
1852 event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
1853 event->param.conn.responder_resources = req_data->responder_resources;
1854 event->param.conn.initiator_depth = req_data->initiator_depth;
1855 event->param.conn.flow_control = req_data->flow_control;
1856 event->param.conn.retry_count = req_data->retry_count;
1857 event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
1858 event->param.conn.srq = req_data->srq;
1859 event->param.conn.qp_num = req_data->remote_qpn;
1862 static int cma_check_req_qp_type(struct rdma_cm_id *id, struct ib_cm_event *ib_event)
1864 return (((ib_event->event == IB_CM_REQ_RECEIVED) &&
1865 (ib_event->param.req_rcvd.qp_type == id->qp_type)) ||
1866 ((ib_event->event == IB_CM_SIDR_REQ_RECEIVED) &&
1867 (id->qp_type == IB_QPT_UD)) ||
1871 static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1873 struct rdma_id_private *listen_id, *conn_id = NULL;
1874 struct rdma_cm_event event;
1875 struct net_device *net_dev;
1878 listen_id = cma_id_from_event(cm_id, ib_event, &net_dev);
1879 if (IS_ERR(listen_id))
1880 return PTR_ERR(listen_id);
1882 if (!cma_check_req_qp_type(&listen_id->id, ib_event)) {
1887 mutex_lock(&listen_id->handler_mutex);
1888 if (listen_id->state != RDMA_CM_LISTEN) {
1889 ret = -ECONNABORTED;
1893 memset(&event, 0, sizeof event);
1894 offset = cma_user_data_offset(listen_id);
1895 event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1896 if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) {
1897 conn_id = cma_new_udp_id(&listen_id->id, ib_event, net_dev);
1898 event.param.ud.private_data = ib_event->private_data + offset;
1899 event.param.ud.private_data_len =
1900 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
1902 conn_id = cma_new_conn_id(&listen_id->id, ib_event, net_dev);
1903 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
1904 ib_event->private_data, offset);
1911 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1912 ret = cma_acquire_dev(conn_id, listen_id);
1916 conn_id->cm_id.ib = cm_id;
1917 cm_id->context = conn_id;
1918 cm_id->cm_handler = cma_ib_handler;
1921 * Protect against the user destroying conn_id from another thread
1922 * until we're done accessing it.
1924 atomic_inc(&conn_id->refcount);
1925 ret = conn_id->id.event_handler(&conn_id->id, &event);
1929 * Acquire mutex to prevent user executing rdma_destroy_id()
1930 * while we're accessing the cm_id.
1933 if (cma_comp(conn_id, RDMA_CM_CONNECT) &&
1934 (conn_id->id.qp_type != IB_QPT_UD))
1935 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
1936 mutex_unlock(&lock);
1937 mutex_unlock(&conn_id->handler_mutex);
1938 mutex_unlock(&listen_id->handler_mutex);
1939 cma_deref_id(conn_id);
1945 cma_deref_id(conn_id);
1946 /* Destroy the CM ID by returning a non-zero value. */
1947 conn_id->cm_id.ib = NULL;
1949 cma_exch(conn_id, RDMA_CM_DESTROYING);
1950 mutex_unlock(&conn_id->handler_mutex);
1952 mutex_unlock(&listen_id->handler_mutex);
1954 rdma_destroy_id(&conn_id->id);
1963 __be64 rdma_get_service_id(struct rdma_cm_id *id, struct sockaddr *addr)
1965 if (addr->sa_family == AF_IB)
1966 return ((struct sockaddr_ib *) addr)->sib_sid;
1968 return cpu_to_be64(((u64)id->ps << 16) + be16_to_cpu(cma_port(addr)));
1970 EXPORT_SYMBOL(rdma_get_service_id);
1972 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
1974 struct rdma_id_private *id_priv = iw_id->context;
1975 struct rdma_cm_event event;
1977 struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
1978 struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
1980 mutex_lock(&id_priv->handler_mutex);
1981 if (id_priv->state != RDMA_CM_CONNECT)
1984 memset(&event, 0, sizeof event);
1985 switch (iw_event->event) {
1986 case IW_CM_EVENT_CLOSE:
1987 event.event = RDMA_CM_EVENT_DISCONNECTED;
1989 case IW_CM_EVENT_CONNECT_REPLY:
1990 memcpy(cma_src_addr(id_priv), laddr,
1991 rdma_addr_size(laddr));
1992 memcpy(cma_dst_addr(id_priv), raddr,
1993 rdma_addr_size(raddr));
1994 switch (iw_event->status) {
1996 event.event = RDMA_CM_EVENT_ESTABLISHED;
1997 event.param.conn.initiator_depth = iw_event->ird;
1998 event.param.conn.responder_resources = iw_event->ord;
2002 event.event = RDMA_CM_EVENT_REJECTED;
2005 event.event = RDMA_CM_EVENT_UNREACHABLE;
2008 event.event = RDMA_CM_EVENT_CONNECT_ERROR;
2012 case IW_CM_EVENT_ESTABLISHED:
2013 event.event = RDMA_CM_EVENT_ESTABLISHED;
2014 event.param.conn.initiator_depth = iw_event->ird;
2015 event.param.conn.responder_resources = iw_event->ord;
2021 event.status = iw_event->status;
2022 event.param.conn.private_data = iw_event->private_data;
2023 event.param.conn.private_data_len = iw_event->private_data_len;
2024 ret = id_priv->id.event_handler(&id_priv->id, &event);
2026 /* Destroy the CM ID by returning a non-zero value. */
2027 id_priv->cm_id.iw = NULL;
2028 cma_exch(id_priv, RDMA_CM_DESTROYING);
2029 mutex_unlock(&id_priv->handler_mutex);
2030 rdma_destroy_id(&id_priv->id);
2035 mutex_unlock(&id_priv->handler_mutex);
2039 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
2040 struct iw_cm_event *iw_event)
2042 struct rdma_cm_id *new_cm_id;
2043 struct rdma_id_private *listen_id, *conn_id;
2044 struct rdma_cm_event event;
2045 int ret = -ECONNABORTED;
2046 struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
2047 struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
2049 listen_id = cm_id->context;
2051 mutex_lock(&listen_id->handler_mutex);
2052 if (listen_id->state != RDMA_CM_LISTEN)
2055 /* Create a new RDMA id for the new IW CM ID */
2056 new_cm_id = rdma_create_id(listen_id->id.route.addr.dev_addr.net,
2057 listen_id->id.event_handler,
2058 listen_id->id.context,
2059 RDMA_PS_TCP, IB_QPT_RC);
2060 if (IS_ERR(new_cm_id)) {
2064 conn_id = container_of(new_cm_id, struct rdma_id_private, id);
2065 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
2066 conn_id->state = RDMA_CM_CONNECT;
2068 ret = rdma_translate_ip(laddr, &conn_id->id.route.addr.dev_addr, NULL);
2070 mutex_unlock(&conn_id->handler_mutex);
2071 rdma_destroy_id(new_cm_id);
2075 ret = cma_acquire_dev(conn_id, listen_id);
2077 mutex_unlock(&conn_id->handler_mutex);
2078 rdma_destroy_id(new_cm_id);
2082 conn_id->cm_id.iw = cm_id;
2083 cm_id->context = conn_id;
2084 cm_id->cm_handler = cma_iw_handler;
2086 memcpy(cma_src_addr(conn_id), laddr, rdma_addr_size(laddr));
2087 memcpy(cma_dst_addr(conn_id), raddr, rdma_addr_size(raddr));
2089 memset(&event, 0, sizeof event);
2090 event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
2091 event.param.conn.private_data = iw_event->private_data;
2092 event.param.conn.private_data_len = iw_event->private_data_len;
2093 event.param.conn.initiator_depth = iw_event->ird;
2094 event.param.conn.responder_resources = iw_event->ord;
2097 * Protect against the user destroying conn_id from another thread
2098 * until we're done accessing it.
2100 atomic_inc(&conn_id->refcount);
2101 ret = conn_id->id.event_handler(&conn_id->id, &event);
2103 /* User wants to destroy the CM ID */
2104 conn_id->cm_id.iw = NULL;
2105 cma_exch(conn_id, RDMA_CM_DESTROYING);
2106 mutex_unlock(&conn_id->handler_mutex);
2107 cma_deref_id(conn_id);
2108 rdma_destroy_id(&conn_id->id);
2112 mutex_unlock(&conn_id->handler_mutex);
2113 cma_deref_id(conn_id);
2116 mutex_unlock(&listen_id->handler_mutex);
2120 static int cma_ib_listen(struct rdma_id_private *id_priv)
2122 struct sockaddr *addr;
2123 struct ib_cm_id *id;
2126 addr = cma_src_addr(id_priv);
2127 svc_id = rdma_get_service_id(&id_priv->id, addr);
2128 id = ib_cm_insert_listen(id_priv->id.device, cma_req_handler, svc_id);
2131 id_priv->cm_id.ib = id;
2136 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
2139 struct iw_cm_id *id;
2141 id = iw_create_cm_id(id_priv->id.device,
2142 iw_conn_req_handler,
2147 id->tos = id_priv->tos;
2148 id_priv->cm_id.iw = id;
2150 memcpy(&id_priv->cm_id.iw->local_addr, cma_src_addr(id_priv),
2151 rdma_addr_size(cma_src_addr(id_priv)));
2153 ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
2156 iw_destroy_cm_id(id_priv->cm_id.iw);
2157 id_priv->cm_id.iw = NULL;
2163 static int cma_listen_handler(struct rdma_cm_id *id,
2164 struct rdma_cm_event *event)
2166 struct rdma_id_private *id_priv = id->context;
2168 id->context = id_priv->id.context;
2169 id->event_handler = id_priv->id.event_handler;
2170 return id_priv->id.event_handler(id, event);
2173 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
2174 struct cma_device *cma_dev)
2176 struct rdma_id_private *dev_id_priv;
2177 struct rdma_cm_id *id;
2178 struct net *net = id_priv->id.route.addr.dev_addr.net;
2181 if (cma_family(id_priv) == AF_IB && !rdma_cap_ib_cm(cma_dev->device, 1))
2184 id = rdma_create_id(net, cma_listen_handler, id_priv, id_priv->id.ps,
2185 id_priv->id.qp_type);
2189 dev_id_priv = container_of(id, struct rdma_id_private, id);
2191 dev_id_priv->state = RDMA_CM_ADDR_BOUND;
2192 memcpy(cma_src_addr(dev_id_priv), cma_src_addr(id_priv),
2193 rdma_addr_size(cma_src_addr(id_priv)));
2195 _cma_attach_to_dev(dev_id_priv, cma_dev);
2196 list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
2197 atomic_inc(&id_priv->refcount);
2198 dev_id_priv->internal_id = 1;
2199 dev_id_priv->afonly = id_priv->afonly;
2201 ret = rdma_listen(id, id_priv->backlog);
2203 pr_warn("RDMA CMA: cma_listen_on_dev, error %d, listening on device %s\n",
2204 ret, cma_dev->device->name);
2207 static void cma_listen_on_all(struct rdma_id_private *id_priv)
2209 struct cma_device *cma_dev;
2212 list_add_tail(&id_priv->list, &listen_any_list);
2213 list_for_each_entry(cma_dev, &dev_list, list)
2214 cma_listen_on_dev(id_priv, cma_dev);
2215 mutex_unlock(&lock);
2218 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
2220 struct rdma_id_private *id_priv;
2222 id_priv = container_of(id, struct rdma_id_private, id);
2223 id_priv->tos = (u8) tos;
2225 EXPORT_SYMBOL(rdma_set_service_type);
2227 static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec,
2230 struct cma_work *work = context;
2231 struct rdma_route *route;
2233 route = &work->id->id.route;
2236 route->num_paths = 1;
2237 *route->path_rec = *path_rec;
2239 work->old_state = RDMA_CM_ROUTE_QUERY;
2240 work->new_state = RDMA_CM_ADDR_RESOLVED;
2241 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
2242 work->event.status = status;
2245 queue_work(cma_wq, &work->work);
2248 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms,
2249 struct cma_work *work)
2251 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2252 struct ib_sa_path_rec path_rec;
2253 ib_sa_comp_mask comp_mask;
2254 struct sockaddr_in6 *sin6;
2255 struct sockaddr_ib *sib;
2257 memset(&path_rec, 0, sizeof path_rec);
2258 rdma_addr_get_sgid(dev_addr, &path_rec.sgid);
2259 rdma_addr_get_dgid(dev_addr, &path_rec.dgid);
2260 path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
2261 path_rec.numb_path = 1;
2262 path_rec.reversible = 1;
2263 path_rec.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
2265 comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
2266 IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
2267 IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
2269 switch (cma_family(id_priv)) {
2271 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
2272 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
2275 sin6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
2276 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
2277 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2280 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
2281 path_rec.traffic_class = (u8) (be32_to_cpu(sib->sib_flowinfo) >> 20);
2282 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2286 id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
2287 id_priv->id.port_num, &path_rec,
2288 comp_mask, timeout_ms,
2289 GFP_KERNEL, cma_query_handler,
2290 work, &id_priv->query);
2292 return (id_priv->query_id < 0) ? id_priv->query_id : 0;
2295 static void cma_work_handler(struct work_struct *_work)
2297 struct cma_work *work = container_of(_work, struct cma_work, work);
2298 struct rdma_id_private *id_priv = work->id;
2301 mutex_lock(&id_priv->handler_mutex);
2302 if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
2305 if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
2306 cma_exch(id_priv, RDMA_CM_DESTROYING);
2310 mutex_unlock(&id_priv->handler_mutex);
2311 cma_deref_id(id_priv);
2313 rdma_destroy_id(&id_priv->id);
2317 static void cma_ndev_work_handler(struct work_struct *_work)
2319 struct cma_ndev_work *work = container_of(_work, struct cma_ndev_work, work);
2320 struct rdma_id_private *id_priv = work->id;
2323 mutex_lock(&id_priv->handler_mutex);
2324 if (id_priv->state == RDMA_CM_DESTROYING ||
2325 id_priv->state == RDMA_CM_DEVICE_REMOVAL)
2328 if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
2329 cma_exch(id_priv, RDMA_CM_DESTROYING);
2334 mutex_unlock(&id_priv->handler_mutex);
2335 cma_deref_id(id_priv);
2337 rdma_destroy_id(&id_priv->id);
2341 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms)
2343 struct rdma_route *route = &id_priv->id.route;
2344 struct cma_work *work;
2347 work = kzalloc(sizeof *work, GFP_KERNEL);
2352 INIT_WORK(&work->work, cma_work_handler);
2353 work->old_state = RDMA_CM_ROUTE_QUERY;
2354 work->new_state = RDMA_CM_ROUTE_RESOLVED;
2355 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
2357 route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
2358 if (!route->path_rec) {
2363 ret = cma_query_ib_route(id_priv, timeout_ms, work);
2369 kfree(route->path_rec);
2370 route->path_rec = NULL;
2376 int rdma_set_ib_paths(struct rdma_cm_id *id,
2377 struct ib_sa_path_rec *path_rec, int num_paths)
2379 struct rdma_id_private *id_priv;
2382 id_priv = container_of(id, struct rdma_id_private, id);
2383 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
2384 RDMA_CM_ROUTE_RESOLVED))
2387 id->route.path_rec = kmemdup(path_rec, sizeof *path_rec * num_paths,
2389 if (!id->route.path_rec) {
2394 id->route.num_paths = num_paths;
2397 cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED);
2400 EXPORT_SYMBOL(rdma_set_ib_paths);
2402 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms)
2404 struct cma_work *work;
2406 work = kzalloc(sizeof *work, GFP_KERNEL);
2411 INIT_WORK(&work->work, cma_work_handler);
2412 work->old_state = RDMA_CM_ROUTE_QUERY;
2413 work->new_state = RDMA_CM_ROUTE_RESOLVED;
2414 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
2415 queue_work(cma_wq, &work->work);
2419 static int iboe_tos_to_sl(struct net_device *ndev, int tos)
2422 struct net_device *dev;
2424 prio = rt_tos2priority(tos);
2425 dev = ndev->priv_flags & IFF_802_1Q_VLAN ?
2426 vlan_dev_real_dev(ndev) : ndev;
2429 return netdev_get_prio_tc_map(dev, prio);
2431 #if IS_ENABLED(CONFIG_VLAN_8021Q)
2432 if (ndev->priv_flags & IFF_802_1Q_VLAN)
2433 return (vlan_dev_get_egress_qos_mask(ndev, prio) &
2434 VLAN_PRIO_MASK) >> VLAN_PRIO_SHIFT;
2439 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
2441 struct rdma_route *route = &id_priv->id.route;
2442 struct rdma_addr *addr = &route->addr;
2443 struct cma_work *work;
2445 struct net_device *ndev = NULL;
2448 work = kzalloc(sizeof *work, GFP_KERNEL);
2453 INIT_WORK(&work->work, cma_work_handler);
2455 route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
2456 if (!route->path_rec) {
2461 route->num_paths = 1;
2463 if (addr->dev_addr.bound_dev_if) {
2464 ndev = dev_get_by_index(&init_net, addr->dev_addr.bound_dev_if);
2470 if (ndev->flags & IFF_LOOPBACK) {
2472 if (!id_priv->id.device->get_netdev) {
2477 ndev = id_priv->id.device->get_netdev(id_priv->id.device,
2478 id_priv->id.port_num);
2485 route->path_rec->net = &init_net;
2486 route->path_rec->ifindex = ndev->ifindex;
2487 route->path_rec->gid_type = id_priv->gid_type;
2494 memcpy(route->path_rec->dmac, addr->dev_addr.dst_dev_addr, ETH_ALEN);
2496 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
2497 &route->path_rec->sgid);
2498 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.dst_addr,
2499 &route->path_rec->dgid);
2501 /* Use the hint from IP Stack to select GID Type */
2502 if (route->path_rec->gid_type < ib_network_to_gid_type(addr->dev_addr.network))
2503 route->path_rec->gid_type = ib_network_to_gid_type(addr->dev_addr.network);
2504 if (((struct sockaddr *)&id_priv->id.route.addr.dst_addr)->sa_family != AF_IB)
2505 /* TODO: get the hoplimit from the inet/inet6 device */
2506 route->path_rec->hop_limit = addr->dev_addr.hoplimit;
2508 route->path_rec->hop_limit = 1;
2509 route->path_rec->reversible = 1;
2510 route->path_rec->pkey = cpu_to_be16(0xffff);
2511 route->path_rec->mtu_selector = IB_SA_EQ;
2512 route->path_rec->sl = iboe_tos_to_sl(ndev, id_priv->tos);
2513 route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
2514 route->path_rec->rate_selector = IB_SA_EQ;
2515 route->path_rec->rate = iboe_get_rate(ndev);
2517 route->path_rec->packet_life_time_selector = IB_SA_EQ;
2518 route->path_rec->packet_life_time = CMA_IBOE_PACKET_LIFETIME;
2519 if (!route->path_rec->mtu) {
2524 work->old_state = RDMA_CM_ROUTE_QUERY;
2525 work->new_state = RDMA_CM_ROUTE_RESOLVED;
2526 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
2527 work->event.status = 0;
2529 queue_work(cma_wq, &work->work);
2534 kfree(route->path_rec);
2535 route->path_rec = NULL;
2541 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
2543 struct rdma_id_private *id_priv;
2546 id_priv = container_of(id, struct rdma_id_private, id);
2547 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY))
2550 atomic_inc(&id_priv->refcount);
2551 if (rdma_cap_ib_sa(id->device, id->port_num))
2552 ret = cma_resolve_ib_route(id_priv, timeout_ms);
2553 else if (rdma_protocol_roce(id->device, id->port_num))
2554 ret = cma_resolve_iboe_route(id_priv);
2555 else if (rdma_protocol_iwarp(id->device, id->port_num))
2556 ret = cma_resolve_iw_route(id_priv, timeout_ms);
2565 cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED);
2566 cma_deref_id(id_priv);
2569 EXPORT_SYMBOL(rdma_resolve_route);
2571 static void cma_set_loopback(struct sockaddr *addr)
2573 switch (addr->sa_family) {
2575 ((struct sockaddr_in *) addr)->sin_addr.s_addr = htonl(INADDR_LOOPBACK);
2578 ipv6_addr_set(&((struct sockaddr_in6 *) addr)->sin6_addr,
2582 ib_addr_set(&((struct sockaddr_ib *) addr)->sib_addr,
2588 static int cma_bind_loopback(struct rdma_id_private *id_priv)
2590 struct cma_device *cma_dev, *cur_dev;
2591 struct ib_port_attr port_attr;
2599 list_for_each_entry(cur_dev, &dev_list, list) {
2600 if (cma_family(id_priv) == AF_IB &&
2601 !rdma_cap_ib_cm(cur_dev->device, 1))
2607 for (p = 1; p <= cur_dev->device->phys_port_cnt; ++p) {
2608 if (!ib_query_port(cur_dev->device, p, &port_attr) &&
2609 port_attr.state == IB_PORT_ACTIVE) {
2624 ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid, NULL);
2628 ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
2632 id_priv->id.route.addr.dev_addr.dev_type =
2633 (rdma_protocol_ib(cma_dev->device, p)) ?
2634 ARPHRD_INFINIBAND : ARPHRD_ETHER;
2636 rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2637 ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
2638 id_priv->id.port_num = p;
2639 cma_attach_to_dev(id_priv, cma_dev);
2640 cma_set_loopback(cma_src_addr(id_priv));
2642 mutex_unlock(&lock);
2646 static void addr_handler(int status, struct sockaddr *src_addr,
2647 struct rdma_dev_addr *dev_addr, void *context)
2649 struct rdma_id_private *id_priv = context;
2650 struct rdma_cm_event event;
2652 memset(&event, 0, sizeof event);
2653 mutex_lock(&id_priv->handler_mutex);
2654 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY,
2655 RDMA_CM_ADDR_RESOLVED))
2658 memcpy(cma_src_addr(id_priv), src_addr, rdma_addr_size(src_addr));
2659 if (!status && !id_priv->cma_dev)
2660 status = cma_acquire_dev(id_priv, NULL);
2663 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
2664 RDMA_CM_ADDR_BOUND))
2666 event.event = RDMA_CM_EVENT_ADDR_ERROR;
2667 event.status = status;
2669 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2671 if (id_priv->id.event_handler(&id_priv->id, &event)) {
2672 cma_exch(id_priv, RDMA_CM_DESTROYING);
2673 mutex_unlock(&id_priv->handler_mutex);
2674 cma_deref_id(id_priv);
2675 rdma_destroy_id(&id_priv->id);
2679 mutex_unlock(&id_priv->handler_mutex);
2680 cma_deref_id(id_priv);
2683 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
2685 struct cma_work *work;
2689 work = kzalloc(sizeof *work, GFP_KERNEL);
2693 if (!id_priv->cma_dev) {
2694 ret = cma_bind_loopback(id_priv);
2699 rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2700 rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
2703 INIT_WORK(&work->work, cma_work_handler);
2704 work->old_state = RDMA_CM_ADDR_QUERY;
2705 work->new_state = RDMA_CM_ADDR_RESOLVED;
2706 work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2707 queue_work(cma_wq, &work->work);
2714 static int cma_resolve_ib_addr(struct rdma_id_private *id_priv)
2716 struct cma_work *work;
2719 work = kzalloc(sizeof *work, GFP_KERNEL);
2723 if (!id_priv->cma_dev) {
2724 ret = cma_resolve_ib_dev(id_priv);
2729 rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, (union ib_gid *)
2730 &(((struct sockaddr_ib *) &id_priv->id.route.addr.dst_addr)->sib_addr));
2733 INIT_WORK(&work->work, cma_work_handler);
2734 work->old_state = RDMA_CM_ADDR_QUERY;
2735 work->new_state = RDMA_CM_ADDR_RESOLVED;
2736 work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2737 queue_work(cma_wq, &work->work);
2744 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2745 struct sockaddr *dst_addr)
2747 if (!src_addr || !src_addr->sa_family) {
2748 src_addr = (struct sockaddr *) &id->route.addr.src_addr;
2749 src_addr->sa_family = dst_addr->sa_family;
2750 if (dst_addr->sa_family == AF_INET6) {
2751 struct sockaddr_in6 *src_addr6 = (struct sockaddr_in6 *) src_addr;
2752 struct sockaddr_in6 *dst_addr6 = (struct sockaddr_in6 *) dst_addr;
2753 src_addr6->sin6_scope_id = dst_addr6->sin6_scope_id;
2754 if (ipv6_addr_type(&dst_addr6->sin6_addr) & IPV6_ADDR_LINKLOCAL)
2755 id->route.addr.dev_addr.bound_dev_if = dst_addr6->sin6_scope_id;
2756 } else if (dst_addr->sa_family == AF_IB) {
2757 ((struct sockaddr_ib *) src_addr)->sib_pkey =
2758 ((struct sockaddr_ib *) dst_addr)->sib_pkey;
2761 return rdma_bind_addr(id, src_addr);
2764 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2765 struct sockaddr *dst_addr, int timeout_ms)
2767 struct rdma_id_private *id_priv;
2770 id_priv = container_of(id, struct rdma_id_private, id);
2771 if (id_priv->state == RDMA_CM_IDLE) {
2772 ret = cma_bind_addr(id, src_addr, dst_addr);
2777 if (cma_family(id_priv) != dst_addr->sa_family)
2780 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY))
2783 atomic_inc(&id_priv->refcount);
2784 memcpy(cma_dst_addr(id_priv), dst_addr, rdma_addr_size(dst_addr));
2785 if (cma_any_addr(dst_addr)) {
2786 ret = cma_resolve_loopback(id_priv);
2788 if (dst_addr->sa_family == AF_IB) {
2789 ret = cma_resolve_ib_addr(id_priv);
2791 ret = rdma_resolve_ip(&addr_client, cma_src_addr(id_priv),
2792 dst_addr, &id->route.addr.dev_addr,
2793 timeout_ms, addr_handler, id_priv);
2801 cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
2802 cma_deref_id(id_priv);
2805 EXPORT_SYMBOL(rdma_resolve_addr);
2807 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse)
2809 struct rdma_id_private *id_priv;
2810 unsigned long flags;
2813 id_priv = container_of(id, struct rdma_id_private, id);
2814 spin_lock_irqsave(&id_priv->lock, flags);
2815 if (reuse || id_priv->state == RDMA_CM_IDLE) {
2816 id_priv->reuseaddr = reuse;
2821 spin_unlock_irqrestore(&id_priv->lock, flags);
2824 EXPORT_SYMBOL(rdma_set_reuseaddr);
2826 int rdma_set_afonly(struct rdma_cm_id *id, int afonly)
2828 struct rdma_id_private *id_priv;
2829 unsigned long flags;
2832 id_priv = container_of(id, struct rdma_id_private, id);
2833 spin_lock_irqsave(&id_priv->lock, flags);
2834 if (id_priv->state == RDMA_CM_IDLE || id_priv->state == RDMA_CM_ADDR_BOUND) {
2835 id_priv->options |= (1 << CMA_OPTION_AFONLY);
2836 id_priv->afonly = afonly;
2841 spin_unlock_irqrestore(&id_priv->lock, flags);
2844 EXPORT_SYMBOL(rdma_set_afonly);
2846 static void cma_bind_port(struct rdma_bind_list *bind_list,
2847 struct rdma_id_private *id_priv)
2849 struct sockaddr *addr;
2850 struct sockaddr_ib *sib;
2854 addr = cma_src_addr(id_priv);
2855 port = htons(bind_list->port);
2857 switch (addr->sa_family) {
2859 ((struct sockaddr_in *) addr)->sin_port = port;
2862 ((struct sockaddr_in6 *) addr)->sin6_port = port;
2865 sib = (struct sockaddr_ib *) addr;
2866 sid = be64_to_cpu(sib->sib_sid);
2867 mask = be64_to_cpu(sib->sib_sid_mask);
2868 sib->sib_sid = cpu_to_be64((sid & mask) | (u64) ntohs(port));
2869 sib->sib_sid_mask = cpu_to_be64(~0ULL);
2872 id_priv->bind_list = bind_list;
2873 hlist_add_head(&id_priv->node, &bind_list->owners);
2876 static int cma_alloc_port(enum rdma_port_space ps,
2877 struct rdma_id_private *id_priv, unsigned short snum)
2879 struct rdma_bind_list *bind_list;
2882 bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
2886 ret = cma_ps_alloc(id_priv->id.route.addr.dev_addr.net, ps, bind_list,
2892 bind_list->port = (unsigned short)ret;
2893 cma_bind_port(bind_list, id_priv);
2897 return ret == -ENOSPC ? -EADDRNOTAVAIL : ret;
2900 static int cma_alloc_any_port(enum rdma_port_space ps,
2901 struct rdma_id_private *id_priv)
2903 static unsigned int last_used_port;
2904 int low, high, remaining;
2906 struct net *net = id_priv->id.route.addr.dev_addr.net;
2908 inet_get_local_port_range(net, &low, &high);
2909 remaining = (high - low) + 1;
2910 rover = prandom_u32() % remaining + low;
2912 if (last_used_port != rover &&
2913 !cma_ps_find(net, ps, (unsigned short)rover)) {
2914 int ret = cma_alloc_port(ps, id_priv, rover);
2916 * Remember previously used port number in order to avoid
2917 * re-using same port immediately after it is closed.
2920 last_used_port = rover;
2921 if (ret != -EADDRNOTAVAIL)
2926 if ((rover < low) || (rover > high))
2930 return -EADDRNOTAVAIL;
2934 * Check that the requested port is available. This is called when trying to
2935 * bind to a specific port, or when trying to listen on a bound port. In
2936 * the latter case, the provided id_priv may already be on the bind_list, but
2937 * we still need to check that it's okay to start listening.
2939 static int cma_check_port(struct rdma_bind_list *bind_list,
2940 struct rdma_id_private *id_priv, uint8_t reuseaddr)
2942 struct rdma_id_private *cur_id;
2943 struct sockaddr *addr, *cur_addr;
2945 addr = cma_src_addr(id_priv);
2946 hlist_for_each_entry(cur_id, &bind_list->owners, node) {
2947 if (id_priv == cur_id)
2950 if ((cur_id->state != RDMA_CM_LISTEN) && reuseaddr &&
2954 cur_addr = cma_src_addr(cur_id);
2955 if (id_priv->afonly && cur_id->afonly &&
2956 (addr->sa_family != cur_addr->sa_family))
2959 if (cma_any_addr(addr) || cma_any_addr(cur_addr))
2960 return -EADDRNOTAVAIL;
2962 if (!cma_addr_cmp(addr, cur_addr))
2968 static int cma_use_port(enum rdma_port_space ps,
2969 struct rdma_id_private *id_priv)
2971 struct rdma_bind_list *bind_list;
2972 unsigned short snum;
2975 snum = ntohs(cma_port(cma_src_addr(id_priv)));
2976 if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
2979 bind_list = cma_ps_find(id_priv->id.route.addr.dev_addr.net, ps, snum);
2981 ret = cma_alloc_port(ps, id_priv, snum);
2983 ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr);
2985 cma_bind_port(bind_list, id_priv);
2990 static int cma_bind_listen(struct rdma_id_private *id_priv)
2992 struct rdma_bind_list *bind_list = id_priv->bind_list;
2996 if (bind_list->owners.first->next)
2997 ret = cma_check_port(bind_list, id_priv, 0);
2998 mutex_unlock(&lock);
3002 static enum rdma_port_space cma_select_inet_ps(
3003 struct rdma_id_private *id_priv)
3005 switch (id_priv->id.ps) {
3010 return id_priv->id.ps;
3017 static enum rdma_port_space cma_select_ib_ps(struct rdma_id_private *id_priv)
3019 enum rdma_port_space ps = 0;
3020 struct sockaddr_ib *sib;
3021 u64 sid_ps, mask, sid;
3023 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
3024 mask = be64_to_cpu(sib->sib_sid_mask) & RDMA_IB_IP_PS_MASK;
3025 sid = be64_to_cpu(sib->sib_sid) & mask;
3027 if ((id_priv->id.ps == RDMA_PS_IB) && (sid == (RDMA_IB_IP_PS_IB & mask))) {
3028 sid_ps = RDMA_IB_IP_PS_IB;
3030 } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_TCP)) &&
3031 (sid == (RDMA_IB_IP_PS_TCP & mask))) {
3032 sid_ps = RDMA_IB_IP_PS_TCP;
3034 } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_UDP)) &&
3035 (sid == (RDMA_IB_IP_PS_UDP & mask))) {
3036 sid_ps = RDMA_IB_IP_PS_UDP;
3041 sib->sib_sid = cpu_to_be64(sid_ps | ntohs(cma_port((struct sockaddr *) sib)));
3042 sib->sib_sid_mask = cpu_to_be64(RDMA_IB_IP_PS_MASK |
3043 be64_to_cpu(sib->sib_sid_mask));
3048 static int cma_get_port(struct rdma_id_private *id_priv)
3050 enum rdma_port_space ps;
3053 if (cma_family(id_priv) != AF_IB)
3054 ps = cma_select_inet_ps(id_priv);
3056 ps = cma_select_ib_ps(id_priv);
3058 return -EPROTONOSUPPORT;
3061 if (cma_any_port(cma_src_addr(id_priv)))
3062 ret = cma_alloc_any_port(ps, id_priv);
3064 ret = cma_use_port(ps, id_priv);
3065 mutex_unlock(&lock);
3070 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
3071 struct sockaddr *addr)
3073 #if IS_ENABLED(CONFIG_IPV6)
3074 struct sockaddr_in6 *sin6;
3076 if (addr->sa_family != AF_INET6)
3079 sin6 = (struct sockaddr_in6 *) addr;
3081 if (!(ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL))
3084 if (!sin6->sin6_scope_id)
3087 dev_addr->bound_dev_if = sin6->sin6_scope_id;
3092 int rdma_listen(struct rdma_cm_id *id, int backlog)
3094 struct rdma_id_private *id_priv;
3097 id_priv = container_of(id, struct rdma_id_private, id);
3098 if (id_priv->state == RDMA_CM_IDLE) {
3099 id->route.addr.src_addr.ss_family = AF_INET;
3100 ret = rdma_bind_addr(id, cma_src_addr(id_priv));
3105 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN))
3108 if (id_priv->reuseaddr) {
3109 ret = cma_bind_listen(id_priv);
3114 id_priv->backlog = backlog;
3116 if (rdma_cap_ib_cm(id->device, 1)) {
3117 ret = cma_ib_listen(id_priv);
3120 } else if (rdma_cap_iw_cm(id->device, 1)) {
3121 ret = cma_iw_listen(id_priv, backlog);
3129 cma_listen_on_all(id_priv);
3133 id_priv->backlog = 0;
3134 cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND);
3137 EXPORT_SYMBOL(rdma_listen);
3139 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
3141 struct rdma_id_private *id_priv;
3144 if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6 &&
3145 addr->sa_family != AF_IB)
3146 return -EAFNOSUPPORT;
3148 id_priv = container_of(id, struct rdma_id_private, id);
3149 if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND))
3152 ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
3156 memcpy(cma_src_addr(id_priv), addr, rdma_addr_size(addr));
3157 if (!cma_any_addr(addr)) {
3158 ret = cma_translate_addr(addr, &id->route.addr.dev_addr);
3162 ret = cma_acquire_dev(id_priv, NULL);
3167 if (!(id_priv->options & (1 << CMA_OPTION_AFONLY))) {
3168 if (addr->sa_family == AF_INET)
3169 id_priv->afonly = 1;
3170 #if IS_ENABLED(CONFIG_IPV6)
3171 else if (addr->sa_family == AF_INET6) {
3172 struct net *net = id_priv->id.route.addr.dev_addr.net;
3174 id_priv->afonly = net->ipv6.sysctl.bindv6only;
3178 ret = cma_get_port(id_priv);
3184 if (id_priv->cma_dev)
3185 cma_release_dev(id_priv);
3187 cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE);
3190 EXPORT_SYMBOL(rdma_bind_addr);
3192 static int cma_format_hdr(void *hdr, struct rdma_id_private *id_priv)
3194 struct cma_hdr *cma_hdr;
3197 cma_hdr->cma_version = CMA_VERSION;
3198 if (cma_family(id_priv) == AF_INET) {
3199 struct sockaddr_in *src4, *dst4;
3201 src4 = (struct sockaddr_in *) cma_src_addr(id_priv);
3202 dst4 = (struct sockaddr_in *) cma_dst_addr(id_priv);
3204 cma_set_ip_ver(cma_hdr, 4);
3205 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
3206 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
3207 cma_hdr->port = src4->sin_port;
3208 } else if (cma_family(id_priv) == AF_INET6) {
3209 struct sockaddr_in6 *src6, *dst6;
3211 src6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
3212 dst6 = (struct sockaddr_in6 *) cma_dst_addr(id_priv);
3214 cma_set_ip_ver(cma_hdr, 6);
3215 cma_hdr->src_addr.ip6 = src6->sin6_addr;
3216 cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
3217 cma_hdr->port = src6->sin6_port;
3222 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
3223 struct ib_cm_event *ib_event)
3225 struct rdma_id_private *id_priv = cm_id->context;
3226 struct rdma_cm_event event;
3227 struct ib_cm_sidr_rep_event_param *rep = &ib_event->param.sidr_rep_rcvd;
3230 mutex_lock(&id_priv->handler_mutex);
3231 if (id_priv->state != RDMA_CM_CONNECT)
3234 memset(&event, 0, sizeof event);
3235 switch (ib_event->event) {
3236 case IB_CM_SIDR_REQ_ERROR:
3237 event.event = RDMA_CM_EVENT_UNREACHABLE;
3238 event.status = -ETIMEDOUT;
3240 case IB_CM_SIDR_REP_RECEIVED:
3241 event.param.ud.private_data = ib_event->private_data;
3242 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
3243 if (rep->status != IB_SIDR_SUCCESS) {
3244 event.event = RDMA_CM_EVENT_UNREACHABLE;
3245 event.status = ib_event->param.sidr_rep_rcvd.status;
3248 ret = cma_set_qkey(id_priv, rep->qkey);
3250 event.event = RDMA_CM_EVENT_ADDR_ERROR;
3254 ib_init_ah_from_path(id_priv->id.device, id_priv->id.port_num,
3255 id_priv->id.route.path_rec,
3256 &event.param.ud.ah_attr);
3257 event.param.ud.qp_num = rep->qpn;
3258 event.param.ud.qkey = rep->qkey;
3259 event.event = RDMA_CM_EVENT_ESTABLISHED;
3263 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
3268 ret = id_priv->id.event_handler(&id_priv->id, &event);
3270 /* Destroy the CM ID by returning a non-zero value. */
3271 id_priv->cm_id.ib = NULL;
3272 cma_exch(id_priv, RDMA_CM_DESTROYING);
3273 mutex_unlock(&id_priv->handler_mutex);
3274 rdma_destroy_id(&id_priv->id);
3278 mutex_unlock(&id_priv->handler_mutex);
3282 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
3283 struct rdma_conn_param *conn_param)
3285 struct ib_cm_sidr_req_param req;
3286 struct ib_cm_id *id;
3290 memset(&req, 0, sizeof req);
3291 offset = cma_user_data_offset(id_priv);
3292 req.private_data_len = offset + conn_param->private_data_len;
3293 if (req.private_data_len < conn_param->private_data_len)
3296 if (req.private_data_len) {
3297 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
3301 private_data = NULL;
3304 if (conn_param->private_data && conn_param->private_data_len)
3305 memcpy(private_data + offset, conn_param->private_data,
3306 conn_param->private_data_len);
3309 ret = cma_format_hdr(private_data, id_priv);
3312 req.private_data = private_data;
3315 id = ib_create_cm_id(id_priv->id.device, cma_sidr_rep_handler,
3321 id_priv->cm_id.ib = id;
3323 req.path = id_priv->id.route.path_rec;
3324 req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
3325 req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
3326 req.max_cm_retries = CMA_MAX_CM_RETRIES;
3328 ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
3330 ib_destroy_cm_id(id_priv->cm_id.ib);
3331 id_priv->cm_id.ib = NULL;
3334 kfree(private_data);
3338 static int cma_connect_ib(struct rdma_id_private *id_priv,
3339 struct rdma_conn_param *conn_param)
3341 struct ib_cm_req_param req;
3342 struct rdma_route *route;
3344 struct ib_cm_id *id;
3347 memset(&req, 0, sizeof req);
3348 offset = cma_user_data_offset(id_priv);
3349 req.private_data_len = offset + conn_param->private_data_len;
3350 if (req.private_data_len < conn_param->private_data_len)
3353 if (req.private_data_len) {
3354 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
3358 private_data = NULL;
3361 if (conn_param->private_data && conn_param->private_data_len)
3362 memcpy(private_data + offset, conn_param->private_data,
3363 conn_param->private_data_len);
3365 id = ib_create_cm_id(id_priv->id.device, cma_ib_handler, id_priv);
3370 id_priv->cm_id.ib = id;
3372 route = &id_priv->id.route;
3374 ret = cma_format_hdr(private_data, id_priv);
3377 req.private_data = private_data;
3380 req.primary_path = &route->path_rec[0];
3381 if (route->num_paths == 2)
3382 req.alternate_path = &route->path_rec[1];
3384 req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
3385 req.qp_num = id_priv->qp_num;
3386 req.qp_type = id_priv->id.qp_type;
3387 req.starting_psn = id_priv->seq_num;
3388 req.responder_resources = conn_param->responder_resources;
3389 req.initiator_depth = conn_param->initiator_depth;
3390 req.flow_control = conn_param->flow_control;
3391 req.retry_count = min_t(u8, 7, conn_param->retry_count);
3392 req.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
3393 req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
3394 req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
3395 req.max_cm_retries = CMA_MAX_CM_RETRIES;
3396 req.srq = id_priv->srq ? 1 : 0;
3398 ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
3400 if (ret && !IS_ERR(id)) {
3401 ib_destroy_cm_id(id);
3402 id_priv->cm_id.ib = NULL;
3405 kfree(private_data);
3409 static int cma_connect_iw(struct rdma_id_private *id_priv,
3410 struct rdma_conn_param *conn_param)
3412 struct iw_cm_id *cm_id;
3414 struct iw_cm_conn_param iw_param;
3416 cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
3418 return PTR_ERR(cm_id);
3420 cm_id->tos = id_priv->tos;
3421 id_priv->cm_id.iw = cm_id;
3423 memcpy(&cm_id->local_addr, cma_src_addr(id_priv),
3424 rdma_addr_size(cma_src_addr(id_priv)));
3425 memcpy(&cm_id->remote_addr, cma_dst_addr(id_priv),
3426 rdma_addr_size(cma_dst_addr(id_priv)));
3428 ret = cma_modify_qp_rtr(id_priv, conn_param);
3433 iw_param.ord = conn_param->initiator_depth;
3434 iw_param.ird = conn_param->responder_resources;
3435 iw_param.private_data = conn_param->private_data;
3436 iw_param.private_data_len = conn_param->private_data_len;
3437 iw_param.qpn = id_priv->id.qp ? id_priv->qp_num : conn_param->qp_num;
3439 memset(&iw_param, 0, sizeof iw_param);
3440 iw_param.qpn = id_priv->qp_num;
3442 ret = iw_cm_connect(cm_id, &iw_param);
3445 iw_destroy_cm_id(cm_id);
3446 id_priv->cm_id.iw = NULL;
3451 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
3453 struct rdma_id_private *id_priv;
3456 id_priv = container_of(id, struct rdma_id_private, id);
3457 if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT))
3461 id_priv->qp_num = conn_param->qp_num;
3462 id_priv->srq = conn_param->srq;
3465 if (rdma_cap_ib_cm(id->device, id->port_num)) {
3466 if (id->qp_type == IB_QPT_UD)
3467 ret = cma_resolve_ib_udp(id_priv, conn_param);
3469 ret = cma_connect_ib(id_priv, conn_param);
3470 } else if (rdma_cap_iw_cm(id->device, id->port_num))
3471 ret = cma_connect_iw(id_priv, conn_param);
3479 cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED);
3482 EXPORT_SYMBOL(rdma_connect);
3484 static int cma_accept_ib(struct rdma_id_private *id_priv,
3485 struct rdma_conn_param *conn_param)
3487 struct ib_cm_rep_param rep;
3490 ret = cma_modify_qp_rtr(id_priv, conn_param);
3494 ret = cma_modify_qp_rts(id_priv, conn_param);
3498 memset(&rep, 0, sizeof rep);
3499 rep.qp_num = id_priv->qp_num;
3500 rep.starting_psn = id_priv->seq_num;
3501 rep.private_data = conn_param->private_data;
3502 rep.private_data_len = conn_param->private_data_len;
3503 rep.responder_resources = conn_param->responder_resources;
3504 rep.initiator_depth = conn_param->initiator_depth;
3505 rep.failover_accepted = 0;
3506 rep.flow_control = conn_param->flow_control;
3507 rep.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
3508 rep.srq = id_priv->srq ? 1 : 0;
3510 ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
3515 static int cma_accept_iw(struct rdma_id_private *id_priv,
3516 struct rdma_conn_param *conn_param)
3518 struct iw_cm_conn_param iw_param;
3521 ret = cma_modify_qp_rtr(id_priv, conn_param);
3525 iw_param.ord = conn_param->initiator_depth;
3526 iw_param.ird = conn_param->responder_resources;
3527 iw_param.private_data = conn_param->private_data;
3528 iw_param.private_data_len = conn_param->private_data_len;
3529 if (id_priv->id.qp) {
3530 iw_param.qpn = id_priv->qp_num;
3532 iw_param.qpn = conn_param->qp_num;
3534 return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
3537 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
3538 enum ib_cm_sidr_status status, u32 qkey,
3539 const void *private_data, int private_data_len)
3541 struct ib_cm_sidr_rep_param rep;
3544 memset(&rep, 0, sizeof rep);
3545 rep.status = status;
3546 if (status == IB_SIDR_SUCCESS) {
3547 ret = cma_set_qkey(id_priv, qkey);
3550 rep.qp_num = id_priv->qp_num;
3551 rep.qkey = id_priv->qkey;
3553 rep.private_data = private_data;
3554 rep.private_data_len = private_data_len;
3556 return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
3559 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
3561 struct rdma_id_private *id_priv;
3564 id_priv = container_of(id, struct rdma_id_private, id);
3566 id_priv->owner = task_pid_nr(current);
3568 if (!cma_comp(id_priv, RDMA_CM_CONNECT))
3571 if (!id->qp && conn_param) {
3572 id_priv->qp_num = conn_param->qp_num;
3573 id_priv->srq = conn_param->srq;
3576 if (rdma_cap_ib_cm(id->device, id->port_num)) {
3577 if (id->qp_type == IB_QPT_UD) {
3579 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
3581 conn_param->private_data,
3582 conn_param->private_data_len);
3584 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
3588 ret = cma_accept_ib(id_priv, conn_param);
3590 ret = cma_rep_recv(id_priv);
3592 } else if (rdma_cap_iw_cm(id->device, id->port_num))
3593 ret = cma_accept_iw(id_priv, conn_param);
3602 cma_modify_qp_err(id_priv);
3603 rdma_reject(id, NULL, 0);
3606 EXPORT_SYMBOL(rdma_accept);
3608 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
3610 struct rdma_id_private *id_priv;
3613 id_priv = container_of(id, struct rdma_id_private, id);
3614 if (!id_priv->cm_id.ib)
3617 switch (id->device->node_type) {
3618 case RDMA_NODE_IB_CA:
3619 ret = ib_cm_notify(id_priv->cm_id.ib, event);
3627 EXPORT_SYMBOL(rdma_notify);
3629 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
3630 u8 private_data_len)
3632 struct rdma_id_private *id_priv;
3635 id_priv = container_of(id, struct rdma_id_private, id);
3636 if (!id_priv->cm_id.ib)
3639 if (rdma_cap_ib_cm(id->device, id->port_num)) {
3640 if (id->qp_type == IB_QPT_UD)
3641 ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT, 0,
3642 private_data, private_data_len);
3644 ret = ib_send_cm_rej(id_priv->cm_id.ib,
3645 IB_CM_REJ_CONSUMER_DEFINED, NULL,
3646 0, private_data, private_data_len);
3647 } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
3648 ret = iw_cm_reject(id_priv->cm_id.iw,
3649 private_data, private_data_len);
3655 EXPORT_SYMBOL(rdma_reject);
3657 int rdma_disconnect(struct rdma_cm_id *id)
3659 struct rdma_id_private *id_priv;
3662 id_priv = container_of(id, struct rdma_id_private, id);
3663 if (!id_priv->cm_id.ib)
3666 if (rdma_cap_ib_cm(id->device, id->port_num)) {
3667 ret = cma_modify_qp_err(id_priv);
3670 /* Initiate or respond to a disconnect. */
3671 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
3672 ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
3673 } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
3674 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
3681 EXPORT_SYMBOL(rdma_disconnect);
3683 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
3685 struct rdma_id_private *id_priv;
3686 struct cma_multicast *mc = multicast->context;
3687 struct rdma_cm_event event;
3690 id_priv = mc->id_priv;
3691 mutex_lock(&id_priv->handler_mutex);
3692 if (id_priv->state != RDMA_CM_ADDR_BOUND &&
3693 id_priv->state != RDMA_CM_ADDR_RESOLVED)
3697 status = cma_set_qkey(id_priv, be32_to_cpu(multicast->rec.qkey));
3698 mutex_lock(&id_priv->qp_mutex);
3699 if (!status && id_priv->id.qp)
3700 status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid,
3701 be16_to_cpu(multicast->rec.mlid));
3702 mutex_unlock(&id_priv->qp_mutex);
3704 memset(&event, 0, sizeof event);
3705 event.status = status;
3706 event.param.ud.private_data = mc->context;
3708 struct rdma_dev_addr *dev_addr =
3709 &id_priv->id.route.addr.dev_addr;
3710 struct net_device *ndev =
3711 dev_get_by_index(&init_net, dev_addr->bound_dev_if);
3712 enum ib_gid_type gid_type =
3713 id_priv->cma_dev->default_gid_type[id_priv->id.port_num -
3714 rdma_start_port(id_priv->cma_dev->device)];
3716 event.event = RDMA_CM_EVENT_MULTICAST_JOIN;
3717 ib_init_ah_from_mcmember(id_priv->id.device,
3718 id_priv->id.port_num, &multicast->rec,
3720 &event.param.ud.ah_attr);
3721 event.param.ud.qp_num = 0xFFFFFF;
3722 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
3726 event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
3728 ret = id_priv->id.event_handler(&id_priv->id, &event);
3730 cma_exch(id_priv, RDMA_CM_DESTROYING);
3731 mutex_unlock(&id_priv->handler_mutex);
3732 rdma_destroy_id(&id_priv->id);
3737 mutex_unlock(&id_priv->handler_mutex);
3741 static void cma_set_mgid(struct rdma_id_private *id_priv,
3742 struct sockaddr *addr, union ib_gid *mgid)
3744 unsigned char mc_map[MAX_ADDR_LEN];
3745 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3746 struct sockaddr_in *sin = (struct sockaddr_in *) addr;
3747 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
3749 if (cma_any_addr(addr)) {
3750 memset(mgid, 0, sizeof *mgid);
3751 } else if ((addr->sa_family == AF_INET6) &&
3752 ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
3754 /* IPv6 address is an SA assigned MGID. */
3755 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3756 } else if (addr->sa_family == AF_IB) {
3757 memcpy(mgid, &((struct sockaddr_ib *) addr)->sib_addr, sizeof *mgid);
3758 } else if ((addr->sa_family == AF_INET6)) {
3759 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
3760 if (id_priv->id.ps == RDMA_PS_UDP)
3761 mc_map[7] = 0x01; /* Use RDMA CM signature */
3762 *mgid = *(union ib_gid *) (mc_map + 4);
3764 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
3765 if (id_priv->id.ps == RDMA_PS_UDP)
3766 mc_map[7] = 0x01; /* Use RDMA CM signature */
3767 *mgid = *(union ib_gid *) (mc_map + 4);
3771 static void cma_query_sa_classport_info_cb(int status,
3772 struct ib_class_port_info *rec,
3775 struct class_port_info_context *cb_ctx = context;
3779 if (status || !rec) {
3780 pr_debug("RDMA CM: %s port %u failed query ClassPortInfo status: %d\n",
3781 cb_ctx->device->name, cb_ctx->port_num, status);
3785 memcpy(cb_ctx->class_port_info, rec, sizeof(struct ib_class_port_info));
3788 complete(&cb_ctx->done);
3791 static int cma_query_sa_classport_info(struct ib_device *device, u8 port_num,
3792 struct ib_class_port_info *class_port_info)
3794 struct class_port_info_context *cb_ctx;
3797 cb_ctx = kmalloc(sizeof(*cb_ctx), GFP_KERNEL);
3801 cb_ctx->device = device;
3802 cb_ctx->class_port_info = class_port_info;
3803 cb_ctx->port_num = port_num;
3804 init_completion(&cb_ctx->done);
3806 ret = ib_sa_classport_info_rec_query(&sa_client, device, port_num,
3807 CMA_QUERY_CLASSPORT_INFO_TIMEOUT,
3808 GFP_KERNEL, cma_query_sa_classport_info_cb,
3809 cb_ctx, &cb_ctx->sa_query);
3811 pr_err("RDMA CM: %s port %u failed to send ClassPortInfo query, ret: %d\n",
3812 device->name, port_num, ret);
3816 wait_for_completion(&cb_ctx->done);
3823 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
3824 struct cma_multicast *mc)
3826 struct ib_sa_mcmember_rec rec;
3827 struct ib_class_port_info class_port_info;
3828 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3829 ib_sa_comp_mask comp_mask;
3832 ib_addr_get_mgid(dev_addr, &rec.mgid);
3833 ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
3838 ret = cma_set_qkey(id_priv, 0);
3842 cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
3843 rec.qkey = cpu_to_be32(id_priv->qkey);
3844 rdma_addr_get_sgid(dev_addr, &rec.port_gid);
3845 rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
3846 rec.join_state = mc->join_state;
3848 if (rec.join_state == BIT(SENDONLY_FULLMEMBER_JOIN)) {
3849 ret = cma_query_sa_classport_info(id_priv->id.device,
3850 id_priv->id.port_num,
3856 if (!(ib_get_cpi_capmask2(&class_port_info) &
3857 IB_SA_CAP_MASK2_SENDONLY_FULL_MEM_SUPPORT)) {
3858 pr_warn("RDMA CM: %s port %u Unable to multicast join\n"
3859 "RDMA CM: SM doesn't support Send Only Full Member option\n",
3860 id_priv->id.device->name, id_priv->id.port_num);
3865 comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
3866 IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
3867 IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
3868 IB_SA_MCMEMBER_REC_FLOW_LABEL |
3869 IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
3871 if (id_priv->id.ps == RDMA_PS_IPOIB)
3872 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
3873 IB_SA_MCMEMBER_REC_RATE_SELECTOR |
3874 IB_SA_MCMEMBER_REC_MTU_SELECTOR |
3875 IB_SA_MCMEMBER_REC_MTU |
3876 IB_SA_MCMEMBER_REC_HOP_LIMIT;
3878 mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device,
3879 id_priv->id.port_num, &rec,
3880 comp_mask, GFP_KERNEL,
3881 cma_ib_mc_handler, mc);
3882 return PTR_ERR_OR_ZERO(mc->multicast.ib);
3885 static void iboe_mcast_work_handler(struct work_struct *work)
3887 struct iboe_mcast_work *mw = container_of(work, struct iboe_mcast_work, work);
3888 struct cma_multicast *mc = mw->mc;
3889 struct ib_sa_multicast *m = mc->multicast.ib;
3891 mc->multicast.ib->context = mc;
3892 cma_ib_mc_handler(0, m);
3893 kref_put(&mc->mcref, release_mc);
3897 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid)
3899 struct sockaddr_in *sin = (struct sockaddr_in *)addr;
3900 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
3902 if (cma_any_addr(addr)) {
3903 memset(mgid, 0, sizeof *mgid);
3904 } else if (addr->sa_family == AF_INET6) {
3905 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3907 mgid->raw[0] = 0xff;
3908 mgid->raw[1] = 0x0e;
3917 mgid->raw[10] = 0xff;
3918 mgid->raw[11] = 0xff;
3919 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
3923 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
3924 struct cma_multicast *mc)
3926 struct iboe_mcast_work *work;
3927 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3929 struct sockaddr *addr = (struct sockaddr *)&mc->addr;
3930 struct net_device *ndev = NULL;
3931 enum ib_gid_type gid_type;
3934 send_only = mc->join_state == BIT(SENDONLY_FULLMEMBER_JOIN);
3936 if (cma_zero_addr((struct sockaddr *)&mc->addr))
3939 work = kzalloc(sizeof *work, GFP_KERNEL);
3943 mc->multicast.ib = kzalloc(sizeof(struct ib_sa_multicast), GFP_KERNEL);
3944 if (!mc->multicast.ib) {
3949 cma_iboe_set_mgid(addr, &mc->multicast.ib->rec.mgid);
3951 mc->multicast.ib->rec.pkey = cpu_to_be16(0xffff);
3952 if (id_priv->id.ps == RDMA_PS_UDP)
3953 mc->multicast.ib->rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
3955 if (dev_addr->bound_dev_if)
3956 ndev = dev_get_by_index(&init_net, dev_addr->bound_dev_if);
3961 mc->multicast.ib->rec.rate = iboe_get_rate(ndev);
3962 mc->multicast.ib->rec.hop_limit = 1;
3963 mc->multicast.ib->rec.mtu = iboe_get_mtu(ndev->mtu);
3965 gid_type = id_priv->cma_dev->default_gid_type[id_priv->id.port_num -
3966 rdma_start_port(id_priv->cma_dev->device)];
3967 if (addr->sa_family == AF_INET) {
3968 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) {
3969 mc->multicast.ib->rec.hop_limit = IPV6_DEFAULT_HOPLIMIT;
3971 err = cma_igmp_send(ndev, &mc->multicast.ib->rec.mgid,
3974 mc->igmp_joined = true;
3978 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP)
3982 if (err || !mc->multicast.ib->rec.mtu) {
3987 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
3988 &mc->multicast.ib->rec.port_gid);
3991 INIT_WORK(&work->work, iboe_mcast_work_handler);
3992 kref_get(&mc->mcref);
3993 queue_work(cma_wq, &work->work);
3998 kfree(mc->multicast.ib);
4004 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
4005 u8 join_state, void *context)
4007 struct rdma_id_private *id_priv;
4008 struct cma_multicast *mc;
4011 id_priv = container_of(id, struct rdma_id_private, id);
4012 if (!cma_comp(id_priv, RDMA_CM_ADDR_BOUND) &&
4013 !cma_comp(id_priv, RDMA_CM_ADDR_RESOLVED))
4016 mc = kmalloc(sizeof *mc, GFP_KERNEL);
4020 memcpy(&mc->addr, addr, rdma_addr_size(addr));
4021 mc->context = context;
4022 mc->id_priv = id_priv;
4023 mc->igmp_joined = false;
4024 mc->join_state = join_state;
4025 spin_lock(&id_priv->lock);
4026 list_add(&mc->list, &id_priv->mc_list);
4027 spin_unlock(&id_priv->lock);
4029 if (rdma_protocol_roce(id->device, id->port_num)) {
4030 kref_init(&mc->mcref);
4031 ret = cma_iboe_join_multicast(id_priv, mc);
4032 } else if (rdma_cap_ib_mcast(id->device, id->port_num))
4033 ret = cma_join_ib_multicast(id_priv, mc);
4038 spin_lock_irq(&id_priv->lock);
4039 list_del(&mc->list);
4040 spin_unlock_irq(&id_priv->lock);
4045 EXPORT_SYMBOL(rdma_join_multicast);
4047 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
4049 struct rdma_id_private *id_priv;
4050 struct cma_multicast *mc;
4052 id_priv = container_of(id, struct rdma_id_private, id);
4053 spin_lock_irq(&id_priv->lock);
4054 list_for_each_entry(mc, &id_priv->mc_list, list) {
4055 if (!memcmp(&mc->addr, addr, rdma_addr_size(addr))) {
4056 list_del(&mc->list);
4057 spin_unlock_irq(&id_priv->lock);
4060 ib_detach_mcast(id->qp,
4061 &mc->multicast.ib->rec.mgid,
4062 be16_to_cpu(mc->multicast.ib->rec.mlid));
4064 BUG_ON(id_priv->cma_dev->device != id->device);
4066 if (rdma_cap_ib_mcast(id->device, id->port_num)) {
4067 ib_sa_free_multicast(mc->multicast.ib);
4069 } else if (rdma_protocol_roce(id->device, id->port_num)) {
4070 if (mc->igmp_joined) {
4071 struct rdma_dev_addr *dev_addr =
4072 &id->route.addr.dev_addr;
4073 struct net_device *ndev = NULL;
4075 if (dev_addr->bound_dev_if)
4076 ndev = dev_get_by_index(&init_net,
4077 dev_addr->bound_dev_if);
4080 &mc->multicast.ib->rec.mgid,
4084 mc->igmp_joined = false;
4086 kref_put(&mc->mcref, release_mc);
4091 spin_unlock_irq(&id_priv->lock);
4093 EXPORT_SYMBOL(rdma_leave_multicast);
4095 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
4097 struct rdma_dev_addr *dev_addr;
4098 struct cma_ndev_work *work;
4100 dev_addr = &id_priv->id.route.addr.dev_addr;
4102 if ((dev_addr->bound_dev_if == ndev->ifindex) &&
4103 (net_eq(dev_net(ndev), dev_addr->net)) &&
4104 memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
4105 pr_info("RDMA CM addr change for ndev %s used by id %p\n",
4106 ndev->name, &id_priv->id);
4107 work = kzalloc(sizeof *work, GFP_KERNEL);
4111 INIT_WORK(&work->work, cma_ndev_work_handler);
4113 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
4114 atomic_inc(&id_priv->refcount);
4115 queue_work(cma_wq, &work->work);
4121 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
4124 struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
4125 struct cma_device *cma_dev;
4126 struct rdma_id_private *id_priv;
4127 int ret = NOTIFY_DONE;
4129 if (event != NETDEV_BONDING_FAILOVER)
4132 if (!(ndev->flags & IFF_MASTER) || !(ndev->priv_flags & IFF_BONDING))
4136 list_for_each_entry(cma_dev, &dev_list, list)
4137 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
4138 ret = cma_netdev_change(ndev, id_priv);
4144 mutex_unlock(&lock);
4148 static struct notifier_block cma_nb = {
4149 .notifier_call = cma_netdev_callback
4152 static void cma_add_one(struct ib_device *device)
4154 struct cma_device *cma_dev;
4155 struct rdma_id_private *id_priv;
4157 unsigned long supported_gids = 0;
4159 cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
4163 cma_dev->device = device;
4164 cma_dev->default_gid_type = kcalloc(device->phys_port_cnt,
4165 sizeof(*cma_dev->default_gid_type),
4167 if (!cma_dev->default_gid_type) {
4171 for (i = rdma_start_port(device); i <= rdma_end_port(device); i++) {
4172 supported_gids = roce_gid_type_mask_support(device, i);
4173 WARN_ON(!supported_gids);
4174 cma_dev->default_gid_type[i - rdma_start_port(device)] =
4175 find_first_bit(&supported_gids, BITS_PER_LONG);
4178 init_completion(&cma_dev->comp);
4179 atomic_set(&cma_dev->refcount, 1);
4180 INIT_LIST_HEAD(&cma_dev->id_list);
4181 ib_set_client_data(device, &cma_client, cma_dev);
4184 list_add_tail(&cma_dev->list, &dev_list);
4185 list_for_each_entry(id_priv, &listen_any_list, list)
4186 cma_listen_on_dev(id_priv, cma_dev);
4187 mutex_unlock(&lock);
4190 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
4192 struct rdma_cm_event event;
4193 enum rdma_cm_state state;
4196 /* Record that we want to remove the device */
4197 state = cma_exch(id_priv, RDMA_CM_DEVICE_REMOVAL);
4198 if (state == RDMA_CM_DESTROYING)
4201 cma_cancel_operation(id_priv, state);
4202 mutex_lock(&id_priv->handler_mutex);
4204 /* Check for destruction from another callback. */
4205 if (!cma_comp(id_priv, RDMA_CM_DEVICE_REMOVAL))
4208 memset(&event, 0, sizeof event);
4209 event.event = RDMA_CM_EVENT_DEVICE_REMOVAL;
4210 ret = id_priv->id.event_handler(&id_priv->id, &event);
4212 mutex_unlock(&id_priv->handler_mutex);
4216 static void cma_process_remove(struct cma_device *cma_dev)
4218 struct rdma_id_private *id_priv;
4222 while (!list_empty(&cma_dev->id_list)) {
4223 id_priv = list_entry(cma_dev->id_list.next,
4224 struct rdma_id_private, list);
4226 list_del(&id_priv->listen_list);
4227 list_del_init(&id_priv->list);
4228 atomic_inc(&id_priv->refcount);
4229 mutex_unlock(&lock);
4231 ret = id_priv->internal_id ? 1 : cma_remove_id_dev(id_priv);
4232 cma_deref_id(id_priv);
4234 rdma_destroy_id(&id_priv->id);
4238 mutex_unlock(&lock);
4240 cma_deref_dev(cma_dev);
4241 wait_for_completion(&cma_dev->comp);
4244 static void cma_remove_one(struct ib_device *device, void *client_data)
4246 struct cma_device *cma_dev = client_data;
4252 list_del(&cma_dev->list);
4253 mutex_unlock(&lock);
4255 cma_process_remove(cma_dev);
4256 kfree(cma_dev->default_gid_type);
4260 static int cma_get_id_stats(struct sk_buff *skb, struct netlink_callback *cb)
4262 struct nlmsghdr *nlh;
4263 struct rdma_cm_id_stats *id_stats;
4264 struct rdma_id_private *id_priv;
4265 struct rdma_cm_id *id = NULL;
4266 struct cma_device *cma_dev;
4267 int i_dev = 0, i_id = 0;
4270 * We export all of the IDs as a sequence of messages. Each
4271 * ID gets its own netlink message.
4275 list_for_each_entry(cma_dev, &dev_list, list) {
4276 if (i_dev < cb->args[0]) {
4282 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
4283 if (i_id < cb->args[1]) {
4288 id_stats = ibnl_put_msg(skb, &nlh, cb->nlh->nlmsg_seq,
4289 sizeof *id_stats, RDMA_NL_RDMA_CM,
4290 RDMA_NL_RDMA_CM_ID_STATS,
4295 memset(id_stats, 0, sizeof *id_stats);
4297 id_stats->node_type = id->route.addr.dev_addr.dev_type;
4298 id_stats->port_num = id->port_num;
4299 id_stats->bound_dev_if =
4300 id->route.addr.dev_addr.bound_dev_if;
4302 if (ibnl_put_attr(skb, nlh,
4303 rdma_addr_size(cma_src_addr(id_priv)),
4304 cma_src_addr(id_priv),
4305 RDMA_NL_RDMA_CM_ATTR_SRC_ADDR))
4307 if (ibnl_put_attr(skb, nlh,
4308 rdma_addr_size(cma_src_addr(id_priv)),
4309 cma_dst_addr(id_priv),
4310 RDMA_NL_RDMA_CM_ATTR_DST_ADDR))
4313 id_stats->pid = id_priv->owner;
4314 id_stats->port_space = id->ps;
4315 id_stats->cm_state = id_priv->state;
4316 id_stats->qp_num = id_priv->qp_num;
4317 id_stats->qp_type = id->qp_type;
4327 mutex_unlock(&lock);
4328 cb->args[0] = i_dev;
4334 static const struct ibnl_client_cbs cma_cb_table[] = {
4335 [RDMA_NL_RDMA_CM_ID_STATS] = { .dump = cma_get_id_stats,
4336 .module = THIS_MODULE },
4339 static int cma_init_net(struct net *net)
4341 struct cma_pernet *pernet = cma_pernet(net);
4343 idr_init(&pernet->tcp_ps);
4344 idr_init(&pernet->udp_ps);
4345 idr_init(&pernet->ipoib_ps);
4346 idr_init(&pernet->ib_ps);
4351 static void cma_exit_net(struct net *net)
4353 struct cma_pernet *pernet = cma_pernet(net);
4355 idr_destroy(&pernet->tcp_ps);
4356 idr_destroy(&pernet->udp_ps);
4357 idr_destroy(&pernet->ipoib_ps);
4358 idr_destroy(&pernet->ib_ps);
4361 static struct pernet_operations cma_pernet_operations = {
4362 .init = cma_init_net,
4363 .exit = cma_exit_net,
4364 .id = &cma_pernet_id,
4365 .size = sizeof(struct cma_pernet),
4368 static int __init cma_init(void)
4372 cma_wq = create_singlethread_workqueue("rdma_cm");
4376 ret = register_pernet_subsys(&cma_pernet_operations);
4380 ib_sa_register_client(&sa_client);
4381 rdma_addr_register_client(&addr_client);
4382 register_netdevice_notifier(&cma_nb);
4384 ret = ib_register_client(&cma_client);
4388 if (ibnl_add_client(RDMA_NL_RDMA_CM, ARRAY_SIZE(cma_cb_table),
4390 pr_warn("RDMA CMA: failed to add netlink callback\n");
4391 cma_configfs_init();
4396 unregister_netdevice_notifier(&cma_nb);
4397 rdma_addr_unregister_client(&addr_client);
4398 ib_sa_unregister_client(&sa_client);
4400 destroy_workqueue(cma_wq);
4404 static void __exit cma_cleanup(void)
4406 cma_configfs_exit();
4407 ibnl_remove_client(RDMA_NL_RDMA_CM);
4408 ib_unregister_client(&cma_client);
4409 unregister_netdevice_notifier(&cma_nb);
4410 rdma_addr_unregister_client(&addr_client);
4411 ib_sa_unregister_client(&sa_client);
4412 unregister_pernet_subsys(&cma_pernet_operations);
4413 destroy_workqueue(cma_wq);
4416 module_init(cma_init);
4417 module_exit(cma_cleanup);