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_in *src_addr,
1098 struct sockaddr_in *dst_addr,
1099 struct cma_hdr *hdr,
1103 *src_addr = (struct sockaddr_in) {
1104 .sin_family = AF_INET,
1105 .sin_addr.s_addr = hdr->dst_addr.ip4.addr,
1106 .sin_port = local_port,
1111 *dst_addr = (struct sockaddr_in) {
1112 .sin_family = AF_INET,
1113 .sin_addr.s_addr = hdr->src_addr.ip4.addr,
1114 .sin_port = hdr->port,
1119 static void cma_save_ip6_info(struct sockaddr_in6 *src_addr,
1120 struct sockaddr_in6 *dst_addr,
1121 struct cma_hdr *hdr,
1125 *src_addr = (struct sockaddr_in6) {
1126 .sin6_family = AF_INET6,
1127 .sin6_addr = hdr->dst_addr.ip6,
1128 .sin6_port = local_port,
1133 *dst_addr = (struct sockaddr_in6) {
1134 .sin6_family = AF_INET6,
1135 .sin6_addr = hdr->src_addr.ip6,
1136 .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((struct sockaddr_in *)src_addr,
1163 (struct sockaddr_in *)dst_addr, hdr, port);
1166 cma_save_ip6_info((struct sockaddr_in6 *)src_addr,
1167 (struct sockaddr_in6 *)dst_addr, hdr, port);
1170 return -EAFNOSUPPORT;
1176 static int cma_save_net_info(struct sockaddr *src_addr,
1177 struct sockaddr *dst_addr,
1178 struct rdma_cm_id *listen_id,
1179 struct ib_cm_event *ib_event,
1180 sa_family_t sa_family, __be64 service_id)
1182 if (sa_family == AF_IB) {
1183 if (ib_event->event == IB_CM_REQ_RECEIVED)
1184 cma_save_ib_info(src_addr, dst_addr, listen_id,
1185 ib_event->param.req_rcvd.primary_path);
1186 else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED)
1187 cma_save_ib_info(src_addr, dst_addr, listen_id, NULL);
1191 return cma_save_ip_info(src_addr, dst_addr, ib_event, service_id);
1194 static int cma_save_req_info(const struct ib_cm_event *ib_event,
1195 struct cma_req_info *req)
1197 const struct ib_cm_req_event_param *req_param =
1198 &ib_event->param.req_rcvd;
1199 const struct ib_cm_sidr_req_event_param *sidr_param =
1200 &ib_event->param.sidr_req_rcvd;
1202 switch (ib_event->event) {
1203 case IB_CM_REQ_RECEIVED:
1204 req->device = req_param->listen_id->device;
1205 req->port = req_param->port;
1206 memcpy(&req->local_gid, &req_param->primary_path->sgid,
1207 sizeof(req->local_gid));
1208 req->has_gid = true;
1209 req->service_id = req_param->primary_path->service_id;
1210 req->pkey = be16_to_cpu(req_param->primary_path->pkey);
1211 if (req->pkey != req_param->bth_pkey)
1212 pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and primary path P_Key (0x%x)\n"
1213 "RDMA CMA: in the future this may cause the request to be dropped\n",
1214 req_param->bth_pkey, req->pkey);
1216 case IB_CM_SIDR_REQ_RECEIVED:
1217 req->device = sidr_param->listen_id->device;
1218 req->port = sidr_param->port;
1219 req->has_gid = false;
1220 req->service_id = sidr_param->service_id;
1221 req->pkey = sidr_param->pkey;
1222 if (req->pkey != sidr_param->bth_pkey)
1223 pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and SIDR request payload P_Key (0x%x)\n"
1224 "RDMA CMA: in the future this may cause the request to be dropped\n",
1225 sidr_param->bth_pkey, req->pkey);
1234 static bool validate_ipv4_net_dev(struct net_device *net_dev,
1235 const struct sockaddr_in *dst_addr,
1236 const struct sockaddr_in *src_addr)
1238 __be32 daddr = dst_addr->sin_addr.s_addr,
1239 saddr = src_addr->sin_addr.s_addr;
1240 struct fib_result res;
1245 if (ipv4_is_multicast(saddr) || ipv4_is_lbcast(saddr) ||
1246 ipv4_is_lbcast(daddr) || ipv4_is_zeronet(saddr) ||
1247 ipv4_is_zeronet(daddr) || ipv4_is_loopback(daddr) ||
1248 ipv4_is_loopback(saddr))
1251 memset(&fl4, 0, sizeof(fl4));
1252 fl4.flowi4_iif = net_dev->ifindex;
1257 err = fib_lookup(dev_net(net_dev), &fl4, &res, 0);
1258 ret = err == 0 && FIB_RES_DEV(res) == net_dev;
1264 static bool validate_ipv6_net_dev(struct net_device *net_dev,
1265 const struct sockaddr_in6 *dst_addr,
1266 const struct sockaddr_in6 *src_addr)
1268 #if IS_ENABLED(CONFIG_IPV6)
1269 const int strict = ipv6_addr_type(&dst_addr->sin6_addr) &
1270 IPV6_ADDR_LINKLOCAL;
1271 struct rt6_info *rt = rt6_lookup(dev_net(net_dev), &dst_addr->sin6_addr,
1272 &src_addr->sin6_addr, net_dev->ifindex,
1279 ret = rt->rt6i_idev->dev == net_dev;
1288 static bool validate_net_dev(struct net_device *net_dev,
1289 const struct sockaddr *daddr,
1290 const struct sockaddr *saddr)
1292 const struct sockaddr_in *daddr4 = (const struct sockaddr_in *)daddr;
1293 const struct sockaddr_in *saddr4 = (const struct sockaddr_in *)saddr;
1294 const struct sockaddr_in6 *daddr6 = (const struct sockaddr_in6 *)daddr;
1295 const struct sockaddr_in6 *saddr6 = (const struct sockaddr_in6 *)saddr;
1297 switch (daddr->sa_family) {
1299 return saddr->sa_family == AF_INET &&
1300 validate_ipv4_net_dev(net_dev, daddr4, saddr4);
1303 return saddr->sa_family == AF_INET6 &&
1304 validate_ipv6_net_dev(net_dev, daddr6, saddr6);
1311 static struct net_device *cma_get_net_dev(struct ib_cm_event *ib_event,
1312 const struct cma_req_info *req)
1314 struct sockaddr_storage listen_addr_storage, src_addr_storage;
1315 struct sockaddr *listen_addr = (struct sockaddr *)&listen_addr_storage,
1316 *src_addr = (struct sockaddr *)&src_addr_storage;
1317 struct net_device *net_dev;
1318 const union ib_gid *gid = req->has_gid ? &req->local_gid : NULL;
1321 err = cma_save_ip_info(listen_addr, src_addr, ib_event,
1324 return ERR_PTR(err);
1326 net_dev = ib_get_net_dev_by_params(req->device, req->port, req->pkey,
1329 return ERR_PTR(-ENODEV);
1331 if (!validate_net_dev(net_dev, listen_addr, src_addr)) {
1333 return ERR_PTR(-EHOSTUNREACH);
1339 static enum rdma_port_space rdma_ps_from_service_id(__be64 service_id)
1341 return (be64_to_cpu(service_id) >> 16) & 0xffff;
1344 static bool cma_match_private_data(struct rdma_id_private *id_priv,
1345 const struct cma_hdr *hdr)
1347 struct sockaddr *addr = cma_src_addr(id_priv);
1349 struct in6_addr ip6_addr;
1351 if (cma_any_addr(addr) && !id_priv->afonly)
1354 switch (addr->sa_family) {
1356 ip4_addr = ((struct sockaddr_in *)addr)->sin_addr.s_addr;
1357 if (cma_get_ip_ver(hdr) != 4)
1359 if (!cma_any_addr(addr) &&
1360 hdr->dst_addr.ip4.addr != ip4_addr)
1364 ip6_addr = ((struct sockaddr_in6 *)addr)->sin6_addr;
1365 if (cma_get_ip_ver(hdr) != 6)
1367 if (!cma_any_addr(addr) &&
1368 memcmp(&hdr->dst_addr.ip6, &ip6_addr, sizeof(ip6_addr)))
1380 static bool cma_protocol_roce_dev_port(struct ib_device *device, int port_num)
1382 enum rdma_link_layer ll = rdma_port_get_link_layer(device, port_num);
1383 enum rdma_transport_type transport =
1384 rdma_node_get_transport(device->node_type);
1386 return ll == IB_LINK_LAYER_ETHERNET && transport == RDMA_TRANSPORT_IB;
1389 static bool cma_protocol_roce(const struct rdma_cm_id *id)
1391 struct ib_device *device = id->device;
1392 const int port_num = id->port_num ?: rdma_start_port(device);
1394 return cma_protocol_roce_dev_port(device, port_num);
1397 static bool cma_match_net_dev(const struct rdma_cm_id *id,
1398 const struct net_device *net_dev,
1401 const struct rdma_addr *addr = &id->route.addr;
1404 /* This request is an AF_IB request or a RoCE request */
1405 return (!id->port_num || id->port_num == port_num) &&
1406 (addr->src_addr.ss_family == AF_IB ||
1407 cma_protocol_roce_dev_port(id->device, port_num));
1409 return !addr->dev_addr.bound_dev_if ||
1410 (net_eq(dev_net(net_dev), addr->dev_addr.net) &&
1411 addr->dev_addr.bound_dev_if == net_dev->ifindex);
1414 static struct rdma_id_private *cma_find_listener(
1415 const struct rdma_bind_list *bind_list,
1416 const struct ib_cm_id *cm_id,
1417 const struct ib_cm_event *ib_event,
1418 const struct cma_req_info *req,
1419 const struct net_device *net_dev)
1421 struct rdma_id_private *id_priv, *id_priv_dev;
1424 return ERR_PTR(-EINVAL);
1426 hlist_for_each_entry(id_priv, &bind_list->owners, node) {
1427 if (cma_match_private_data(id_priv, ib_event->private_data)) {
1428 if (id_priv->id.device == cm_id->device &&
1429 cma_match_net_dev(&id_priv->id, net_dev, req->port))
1431 list_for_each_entry(id_priv_dev,
1432 &id_priv->listen_list,
1434 if (id_priv_dev->id.device == cm_id->device &&
1435 cma_match_net_dev(&id_priv_dev->id, net_dev, req->port))
1441 return ERR_PTR(-EINVAL);
1444 static struct rdma_id_private *cma_id_from_event(struct ib_cm_id *cm_id,
1445 struct ib_cm_event *ib_event,
1446 struct net_device **net_dev)
1448 struct cma_req_info req;
1449 struct rdma_bind_list *bind_list;
1450 struct rdma_id_private *id_priv;
1453 err = cma_save_req_info(ib_event, &req);
1455 return ERR_PTR(err);
1457 *net_dev = cma_get_net_dev(ib_event, &req);
1458 if (IS_ERR(*net_dev)) {
1459 if (PTR_ERR(*net_dev) == -EAFNOSUPPORT) {
1460 /* Assuming the protocol is AF_IB */
1462 } else if (cma_protocol_roce_dev_port(req.device, req.port)) {
1463 /* TODO find the net dev matching the request parameters
1464 * through the RoCE GID table */
1467 return ERR_CAST(*net_dev);
1471 bind_list = cma_ps_find(*net_dev ? dev_net(*net_dev) : &init_net,
1472 rdma_ps_from_service_id(req.service_id),
1473 cma_port_from_service_id(req.service_id));
1474 id_priv = cma_find_listener(bind_list, cm_id, ib_event, &req, *net_dev);
1475 if (IS_ERR(id_priv) && *net_dev) {
1483 static inline int cma_user_data_offset(struct rdma_id_private *id_priv)
1485 return cma_family(id_priv) == AF_IB ? 0 : sizeof(struct cma_hdr);
1488 static void cma_cancel_route(struct rdma_id_private *id_priv)
1490 if (rdma_cap_ib_sa(id_priv->id.device, id_priv->id.port_num)) {
1492 ib_sa_cancel_query(id_priv->query_id, id_priv->query);
1496 static void cma_cancel_listens(struct rdma_id_private *id_priv)
1498 struct rdma_id_private *dev_id_priv;
1501 * Remove from listen_any_list to prevent added devices from spawning
1502 * additional listen requests.
1505 list_del(&id_priv->list);
1507 while (!list_empty(&id_priv->listen_list)) {
1508 dev_id_priv = list_entry(id_priv->listen_list.next,
1509 struct rdma_id_private, listen_list);
1510 /* sync with device removal to avoid duplicate destruction */
1511 list_del_init(&dev_id_priv->list);
1512 list_del(&dev_id_priv->listen_list);
1513 mutex_unlock(&lock);
1515 rdma_destroy_id(&dev_id_priv->id);
1518 mutex_unlock(&lock);
1521 static void cma_cancel_operation(struct rdma_id_private *id_priv,
1522 enum rdma_cm_state state)
1525 case RDMA_CM_ADDR_QUERY:
1526 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
1528 case RDMA_CM_ROUTE_QUERY:
1529 cma_cancel_route(id_priv);
1531 case RDMA_CM_LISTEN:
1532 if (cma_any_addr(cma_src_addr(id_priv)) && !id_priv->cma_dev)
1533 cma_cancel_listens(id_priv);
1540 static void cma_release_port(struct rdma_id_private *id_priv)
1542 struct rdma_bind_list *bind_list = id_priv->bind_list;
1543 struct net *net = id_priv->id.route.addr.dev_addr.net;
1549 hlist_del(&id_priv->node);
1550 if (hlist_empty(&bind_list->owners)) {
1551 cma_ps_remove(net, bind_list->ps, bind_list->port);
1554 mutex_unlock(&lock);
1557 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
1559 struct cma_multicast *mc;
1561 while (!list_empty(&id_priv->mc_list)) {
1562 mc = container_of(id_priv->mc_list.next,
1563 struct cma_multicast, list);
1564 list_del(&mc->list);
1565 if (rdma_cap_ib_mcast(id_priv->cma_dev->device,
1566 id_priv->id.port_num)) {
1567 ib_sa_free_multicast(mc->multicast.ib);
1570 if (mc->igmp_joined) {
1571 struct rdma_dev_addr *dev_addr =
1572 &id_priv->id.route.addr.dev_addr;
1573 struct net_device *ndev = NULL;
1575 if (dev_addr->bound_dev_if)
1576 ndev = dev_get_by_index(&init_net,
1577 dev_addr->bound_dev_if);
1580 &mc->multicast.ib->rec.mgid,
1585 kref_put(&mc->mcref, release_mc);
1590 void rdma_destroy_id(struct rdma_cm_id *id)
1592 struct rdma_id_private *id_priv;
1593 enum rdma_cm_state state;
1595 id_priv = container_of(id, struct rdma_id_private, id);
1596 state = cma_exch(id_priv, RDMA_CM_DESTROYING);
1597 cma_cancel_operation(id_priv, state);
1600 * Wait for any active callback to finish. New callbacks will find
1601 * the id_priv state set to destroying and abort.
1603 mutex_lock(&id_priv->handler_mutex);
1604 mutex_unlock(&id_priv->handler_mutex);
1606 if (id_priv->cma_dev) {
1607 if (rdma_cap_ib_cm(id_priv->id.device, 1)) {
1608 if (id_priv->cm_id.ib)
1609 ib_destroy_cm_id(id_priv->cm_id.ib);
1610 } else if (rdma_cap_iw_cm(id_priv->id.device, 1)) {
1611 if (id_priv->cm_id.iw)
1612 iw_destroy_cm_id(id_priv->cm_id.iw);
1614 cma_leave_mc_groups(id_priv);
1615 cma_release_dev(id_priv);
1618 cma_release_port(id_priv);
1619 cma_deref_id(id_priv);
1620 wait_for_completion(&id_priv->comp);
1622 if (id_priv->internal_id)
1623 cma_deref_id(id_priv->id.context);
1625 kfree(id_priv->id.route.path_rec);
1626 put_net(id_priv->id.route.addr.dev_addr.net);
1629 EXPORT_SYMBOL(rdma_destroy_id);
1631 static int cma_rep_recv(struct rdma_id_private *id_priv)
1635 ret = cma_modify_qp_rtr(id_priv, NULL);
1639 ret = cma_modify_qp_rts(id_priv, NULL);
1643 ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
1649 cma_modify_qp_err(id_priv);
1650 ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
1655 static void cma_set_rep_event_data(struct rdma_cm_event *event,
1656 struct ib_cm_rep_event_param *rep_data,
1659 event->param.conn.private_data = private_data;
1660 event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
1661 event->param.conn.responder_resources = rep_data->responder_resources;
1662 event->param.conn.initiator_depth = rep_data->initiator_depth;
1663 event->param.conn.flow_control = rep_data->flow_control;
1664 event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
1665 event->param.conn.srq = rep_data->srq;
1666 event->param.conn.qp_num = rep_data->remote_qpn;
1669 static int cma_ib_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1671 struct rdma_id_private *id_priv = cm_id->context;
1672 struct rdma_cm_event event;
1675 mutex_lock(&id_priv->handler_mutex);
1676 if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
1677 id_priv->state != RDMA_CM_CONNECT) ||
1678 (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
1679 id_priv->state != RDMA_CM_DISCONNECT))
1682 memset(&event, 0, sizeof event);
1683 switch (ib_event->event) {
1684 case IB_CM_REQ_ERROR:
1685 case IB_CM_REP_ERROR:
1686 event.event = RDMA_CM_EVENT_UNREACHABLE;
1687 event.status = -ETIMEDOUT;
1689 case IB_CM_REP_RECEIVED:
1690 if (id_priv->id.qp) {
1691 event.status = cma_rep_recv(id_priv);
1692 event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
1693 RDMA_CM_EVENT_ESTABLISHED;
1695 event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
1697 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
1698 ib_event->private_data);
1700 case IB_CM_RTU_RECEIVED:
1701 case IB_CM_USER_ESTABLISHED:
1702 event.event = RDMA_CM_EVENT_ESTABLISHED;
1704 case IB_CM_DREQ_ERROR:
1705 event.status = -ETIMEDOUT; /* fall through */
1706 case IB_CM_DREQ_RECEIVED:
1707 case IB_CM_DREP_RECEIVED:
1708 if (!cma_comp_exch(id_priv, RDMA_CM_CONNECT,
1709 RDMA_CM_DISCONNECT))
1711 event.event = RDMA_CM_EVENT_DISCONNECTED;
1713 case IB_CM_TIMEWAIT_EXIT:
1714 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
1716 case IB_CM_MRA_RECEIVED:
1719 case IB_CM_REJ_RECEIVED:
1720 cma_modify_qp_err(id_priv);
1721 event.status = ib_event->param.rej_rcvd.reason;
1722 event.event = RDMA_CM_EVENT_REJECTED;
1723 event.param.conn.private_data = ib_event->private_data;
1724 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
1727 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
1732 ret = id_priv->id.event_handler(&id_priv->id, &event);
1734 /* Destroy the CM ID by returning a non-zero value. */
1735 id_priv->cm_id.ib = NULL;
1736 cma_exch(id_priv, RDMA_CM_DESTROYING);
1737 mutex_unlock(&id_priv->handler_mutex);
1738 rdma_destroy_id(&id_priv->id);
1742 mutex_unlock(&id_priv->handler_mutex);
1746 static struct rdma_id_private *cma_new_conn_id(struct rdma_cm_id *listen_id,
1747 struct ib_cm_event *ib_event,
1748 struct net_device *net_dev)
1750 struct rdma_id_private *id_priv;
1751 struct rdma_cm_id *id;
1752 struct rdma_route *rt;
1753 const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
1754 const __be64 service_id =
1755 ib_event->param.req_rcvd.primary_path->service_id;
1758 id = rdma_create_id(listen_id->route.addr.dev_addr.net,
1759 listen_id->event_handler, listen_id->context,
1760 listen_id->ps, ib_event->param.req_rcvd.qp_type);
1764 id_priv = container_of(id, struct rdma_id_private, id);
1765 if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
1766 (struct sockaddr *)&id->route.addr.dst_addr,
1767 listen_id, ib_event, ss_family, service_id))
1771 rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
1772 rt->path_rec = kmalloc(sizeof *rt->path_rec * rt->num_paths,
1777 rt->path_rec[0] = *ib_event->param.req_rcvd.primary_path;
1778 if (rt->num_paths == 2)
1779 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
1782 ret = rdma_copy_addr(&rt->addr.dev_addr, net_dev, NULL);
1786 if (!cma_protocol_roce(listen_id) &&
1787 cma_any_addr(cma_src_addr(id_priv))) {
1788 rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
1789 rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
1790 ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
1791 } else if (!cma_any_addr(cma_src_addr(id_priv))) {
1792 ret = cma_translate_addr(cma_src_addr(id_priv), &rt->addr.dev_addr);
1797 rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
1799 id_priv->state = RDMA_CM_CONNECT;
1803 rdma_destroy_id(id);
1807 static struct rdma_id_private *cma_new_udp_id(struct rdma_cm_id *listen_id,
1808 struct ib_cm_event *ib_event,
1809 struct net_device *net_dev)
1811 struct rdma_id_private *id_priv;
1812 struct rdma_cm_id *id;
1813 const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
1814 struct net *net = listen_id->route.addr.dev_addr.net;
1817 id = rdma_create_id(net, listen_id->event_handler, listen_id->context,
1818 listen_id->ps, IB_QPT_UD);
1822 id_priv = container_of(id, struct rdma_id_private, id);
1823 if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
1824 (struct sockaddr *)&id->route.addr.dst_addr,
1825 listen_id, ib_event, ss_family,
1826 ib_event->param.sidr_req_rcvd.service_id))
1830 ret = rdma_copy_addr(&id->route.addr.dev_addr, net_dev, NULL);
1834 if (!cma_any_addr(cma_src_addr(id_priv))) {
1835 ret = cma_translate_addr(cma_src_addr(id_priv),
1836 &id->route.addr.dev_addr);
1842 id_priv->state = RDMA_CM_CONNECT;
1845 rdma_destroy_id(id);
1849 static void cma_set_req_event_data(struct rdma_cm_event *event,
1850 struct ib_cm_req_event_param *req_data,
1851 void *private_data, int offset)
1853 event->param.conn.private_data = private_data + offset;
1854 event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
1855 event->param.conn.responder_resources = req_data->responder_resources;
1856 event->param.conn.initiator_depth = req_data->initiator_depth;
1857 event->param.conn.flow_control = req_data->flow_control;
1858 event->param.conn.retry_count = req_data->retry_count;
1859 event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
1860 event->param.conn.srq = req_data->srq;
1861 event->param.conn.qp_num = req_data->remote_qpn;
1864 static int cma_check_req_qp_type(struct rdma_cm_id *id, struct ib_cm_event *ib_event)
1866 return (((ib_event->event == IB_CM_REQ_RECEIVED) &&
1867 (ib_event->param.req_rcvd.qp_type == id->qp_type)) ||
1868 ((ib_event->event == IB_CM_SIDR_REQ_RECEIVED) &&
1869 (id->qp_type == IB_QPT_UD)) ||
1873 static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1875 struct rdma_id_private *listen_id, *conn_id = NULL;
1876 struct rdma_cm_event event;
1877 struct net_device *net_dev;
1880 listen_id = cma_id_from_event(cm_id, ib_event, &net_dev);
1881 if (IS_ERR(listen_id))
1882 return PTR_ERR(listen_id);
1884 if (!cma_check_req_qp_type(&listen_id->id, ib_event)) {
1889 mutex_lock(&listen_id->handler_mutex);
1890 if (listen_id->state != RDMA_CM_LISTEN) {
1891 ret = -ECONNABORTED;
1895 memset(&event, 0, sizeof event);
1896 offset = cma_user_data_offset(listen_id);
1897 event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1898 if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) {
1899 conn_id = cma_new_udp_id(&listen_id->id, ib_event, net_dev);
1900 event.param.ud.private_data = ib_event->private_data + offset;
1901 event.param.ud.private_data_len =
1902 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
1904 conn_id = cma_new_conn_id(&listen_id->id, ib_event, net_dev);
1905 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
1906 ib_event->private_data, offset);
1913 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1914 ret = cma_acquire_dev(conn_id, listen_id);
1918 conn_id->cm_id.ib = cm_id;
1919 cm_id->context = conn_id;
1920 cm_id->cm_handler = cma_ib_handler;
1923 * Protect against the user destroying conn_id from another thread
1924 * until we're done accessing it.
1926 atomic_inc(&conn_id->refcount);
1927 ret = conn_id->id.event_handler(&conn_id->id, &event);
1931 * Acquire mutex to prevent user executing rdma_destroy_id()
1932 * while we're accessing the cm_id.
1935 if (cma_comp(conn_id, RDMA_CM_CONNECT) &&
1936 (conn_id->id.qp_type != IB_QPT_UD))
1937 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
1938 mutex_unlock(&lock);
1939 mutex_unlock(&conn_id->handler_mutex);
1940 mutex_unlock(&listen_id->handler_mutex);
1941 cma_deref_id(conn_id);
1947 cma_deref_id(conn_id);
1948 /* Destroy the CM ID by returning a non-zero value. */
1949 conn_id->cm_id.ib = NULL;
1951 cma_exch(conn_id, RDMA_CM_DESTROYING);
1952 mutex_unlock(&conn_id->handler_mutex);
1954 mutex_unlock(&listen_id->handler_mutex);
1956 rdma_destroy_id(&conn_id->id);
1965 __be64 rdma_get_service_id(struct rdma_cm_id *id, struct sockaddr *addr)
1967 if (addr->sa_family == AF_IB)
1968 return ((struct sockaddr_ib *) addr)->sib_sid;
1970 return cpu_to_be64(((u64)id->ps << 16) + be16_to_cpu(cma_port(addr)));
1972 EXPORT_SYMBOL(rdma_get_service_id);
1974 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
1976 struct rdma_id_private *id_priv = iw_id->context;
1977 struct rdma_cm_event event;
1979 struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
1980 struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
1982 mutex_lock(&id_priv->handler_mutex);
1983 if (id_priv->state != RDMA_CM_CONNECT)
1986 memset(&event, 0, sizeof event);
1987 switch (iw_event->event) {
1988 case IW_CM_EVENT_CLOSE:
1989 event.event = RDMA_CM_EVENT_DISCONNECTED;
1991 case IW_CM_EVENT_CONNECT_REPLY:
1992 memcpy(cma_src_addr(id_priv), laddr,
1993 rdma_addr_size(laddr));
1994 memcpy(cma_dst_addr(id_priv), raddr,
1995 rdma_addr_size(raddr));
1996 switch (iw_event->status) {
1998 event.event = RDMA_CM_EVENT_ESTABLISHED;
1999 event.param.conn.initiator_depth = iw_event->ird;
2000 event.param.conn.responder_resources = iw_event->ord;
2004 event.event = RDMA_CM_EVENT_REJECTED;
2007 event.event = RDMA_CM_EVENT_UNREACHABLE;
2010 event.event = RDMA_CM_EVENT_CONNECT_ERROR;
2014 case IW_CM_EVENT_ESTABLISHED:
2015 event.event = RDMA_CM_EVENT_ESTABLISHED;
2016 event.param.conn.initiator_depth = iw_event->ird;
2017 event.param.conn.responder_resources = iw_event->ord;
2023 event.status = iw_event->status;
2024 event.param.conn.private_data = iw_event->private_data;
2025 event.param.conn.private_data_len = iw_event->private_data_len;
2026 ret = id_priv->id.event_handler(&id_priv->id, &event);
2028 /* Destroy the CM ID by returning a non-zero value. */
2029 id_priv->cm_id.iw = NULL;
2030 cma_exch(id_priv, RDMA_CM_DESTROYING);
2031 mutex_unlock(&id_priv->handler_mutex);
2032 rdma_destroy_id(&id_priv->id);
2037 mutex_unlock(&id_priv->handler_mutex);
2041 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
2042 struct iw_cm_event *iw_event)
2044 struct rdma_cm_id *new_cm_id;
2045 struct rdma_id_private *listen_id, *conn_id;
2046 struct rdma_cm_event event;
2047 int ret = -ECONNABORTED;
2048 struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
2049 struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
2051 listen_id = cm_id->context;
2053 mutex_lock(&listen_id->handler_mutex);
2054 if (listen_id->state != RDMA_CM_LISTEN)
2057 /* Create a new RDMA id for the new IW CM ID */
2058 new_cm_id = rdma_create_id(listen_id->id.route.addr.dev_addr.net,
2059 listen_id->id.event_handler,
2060 listen_id->id.context,
2061 RDMA_PS_TCP, IB_QPT_RC);
2062 if (IS_ERR(new_cm_id)) {
2066 conn_id = container_of(new_cm_id, struct rdma_id_private, id);
2067 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
2068 conn_id->state = RDMA_CM_CONNECT;
2070 ret = rdma_translate_ip(laddr, &conn_id->id.route.addr.dev_addr, NULL);
2072 mutex_unlock(&conn_id->handler_mutex);
2073 rdma_destroy_id(new_cm_id);
2077 ret = cma_acquire_dev(conn_id, listen_id);
2079 mutex_unlock(&conn_id->handler_mutex);
2080 rdma_destroy_id(new_cm_id);
2084 conn_id->cm_id.iw = cm_id;
2085 cm_id->context = conn_id;
2086 cm_id->cm_handler = cma_iw_handler;
2088 memcpy(cma_src_addr(conn_id), laddr, rdma_addr_size(laddr));
2089 memcpy(cma_dst_addr(conn_id), raddr, rdma_addr_size(raddr));
2091 memset(&event, 0, sizeof event);
2092 event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
2093 event.param.conn.private_data = iw_event->private_data;
2094 event.param.conn.private_data_len = iw_event->private_data_len;
2095 event.param.conn.initiator_depth = iw_event->ird;
2096 event.param.conn.responder_resources = iw_event->ord;
2099 * Protect against the user destroying conn_id from another thread
2100 * until we're done accessing it.
2102 atomic_inc(&conn_id->refcount);
2103 ret = conn_id->id.event_handler(&conn_id->id, &event);
2105 /* User wants to destroy the CM ID */
2106 conn_id->cm_id.iw = NULL;
2107 cma_exch(conn_id, RDMA_CM_DESTROYING);
2108 mutex_unlock(&conn_id->handler_mutex);
2109 cma_deref_id(conn_id);
2110 rdma_destroy_id(&conn_id->id);
2114 mutex_unlock(&conn_id->handler_mutex);
2115 cma_deref_id(conn_id);
2118 mutex_unlock(&listen_id->handler_mutex);
2122 static int cma_ib_listen(struct rdma_id_private *id_priv)
2124 struct sockaddr *addr;
2125 struct ib_cm_id *id;
2128 addr = cma_src_addr(id_priv);
2129 svc_id = rdma_get_service_id(&id_priv->id, addr);
2130 id = ib_cm_insert_listen(id_priv->id.device, cma_req_handler, svc_id);
2133 id_priv->cm_id.ib = id;
2138 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
2141 struct iw_cm_id *id;
2143 id = iw_create_cm_id(id_priv->id.device,
2144 iw_conn_req_handler,
2149 id->tos = id_priv->tos;
2150 id_priv->cm_id.iw = id;
2152 memcpy(&id_priv->cm_id.iw->local_addr, cma_src_addr(id_priv),
2153 rdma_addr_size(cma_src_addr(id_priv)));
2155 ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
2158 iw_destroy_cm_id(id_priv->cm_id.iw);
2159 id_priv->cm_id.iw = NULL;
2165 static int cma_listen_handler(struct rdma_cm_id *id,
2166 struct rdma_cm_event *event)
2168 struct rdma_id_private *id_priv = id->context;
2170 id->context = id_priv->id.context;
2171 id->event_handler = id_priv->id.event_handler;
2172 return id_priv->id.event_handler(id, event);
2175 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
2176 struct cma_device *cma_dev)
2178 struct rdma_id_private *dev_id_priv;
2179 struct rdma_cm_id *id;
2180 struct net *net = id_priv->id.route.addr.dev_addr.net;
2183 if (cma_family(id_priv) == AF_IB && !rdma_cap_ib_cm(cma_dev->device, 1))
2186 id = rdma_create_id(net, cma_listen_handler, id_priv, id_priv->id.ps,
2187 id_priv->id.qp_type);
2191 dev_id_priv = container_of(id, struct rdma_id_private, id);
2193 dev_id_priv->state = RDMA_CM_ADDR_BOUND;
2194 memcpy(cma_src_addr(dev_id_priv), cma_src_addr(id_priv),
2195 rdma_addr_size(cma_src_addr(id_priv)));
2197 _cma_attach_to_dev(dev_id_priv, cma_dev);
2198 list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
2199 atomic_inc(&id_priv->refcount);
2200 dev_id_priv->internal_id = 1;
2201 dev_id_priv->afonly = id_priv->afonly;
2203 ret = rdma_listen(id, id_priv->backlog);
2205 pr_warn("RDMA CMA: cma_listen_on_dev, error %d, listening on device %s\n",
2206 ret, cma_dev->device->name);
2209 static void cma_listen_on_all(struct rdma_id_private *id_priv)
2211 struct cma_device *cma_dev;
2214 list_add_tail(&id_priv->list, &listen_any_list);
2215 list_for_each_entry(cma_dev, &dev_list, list)
2216 cma_listen_on_dev(id_priv, cma_dev);
2217 mutex_unlock(&lock);
2220 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
2222 struct rdma_id_private *id_priv;
2224 id_priv = container_of(id, struct rdma_id_private, id);
2225 id_priv->tos = (u8) tos;
2227 EXPORT_SYMBOL(rdma_set_service_type);
2229 static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec,
2232 struct cma_work *work = context;
2233 struct rdma_route *route;
2235 route = &work->id->id.route;
2238 route->num_paths = 1;
2239 *route->path_rec = *path_rec;
2241 work->old_state = RDMA_CM_ROUTE_QUERY;
2242 work->new_state = RDMA_CM_ADDR_RESOLVED;
2243 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
2244 work->event.status = status;
2247 queue_work(cma_wq, &work->work);
2250 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms,
2251 struct cma_work *work)
2253 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2254 struct ib_sa_path_rec path_rec;
2255 ib_sa_comp_mask comp_mask;
2256 struct sockaddr_in6 *sin6;
2257 struct sockaddr_ib *sib;
2259 memset(&path_rec, 0, sizeof path_rec);
2260 rdma_addr_get_sgid(dev_addr, &path_rec.sgid);
2261 rdma_addr_get_dgid(dev_addr, &path_rec.dgid);
2262 path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
2263 path_rec.numb_path = 1;
2264 path_rec.reversible = 1;
2265 path_rec.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
2267 comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
2268 IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
2269 IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
2271 switch (cma_family(id_priv)) {
2273 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
2274 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
2277 sin6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
2278 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
2279 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2282 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
2283 path_rec.traffic_class = (u8) (be32_to_cpu(sib->sib_flowinfo) >> 20);
2284 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2288 id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
2289 id_priv->id.port_num, &path_rec,
2290 comp_mask, timeout_ms,
2291 GFP_KERNEL, cma_query_handler,
2292 work, &id_priv->query);
2294 return (id_priv->query_id < 0) ? id_priv->query_id : 0;
2297 static void cma_work_handler(struct work_struct *_work)
2299 struct cma_work *work = container_of(_work, struct cma_work, work);
2300 struct rdma_id_private *id_priv = work->id;
2303 mutex_lock(&id_priv->handler_mutex);
2304 if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
2307 if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
2308 cma_exch(id_priv, RDMA_CM_DESTROYING);
2312 mutex_unlock(&id_priv->handler_mutex);
2313 cma_deref_id(id_priv);
2315 rdma_destroy_id(&id_priv->id);
2319 static void cma_ndev_work_handler(struct work_struct *_work)
2321 struct cma_ndev_work *work = container_of(_work, struct cma_ndev_work, work);
2322 struct rdma_id_private *id_priv = work->id;
2325 mutex_lock(&id_priv->handler_mutex);
2326 if (id_priv->state == RDMA_CM_DESTROYING ||
2327 id_priv->state == RDMA_CM_DEVICE_REMOVAL)
2330 if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
2331 cma_exch(id_priv, RDMA_CM_DESTROYING);
2336 mutex_unlock(&id_priv->handler_mutex);
2337 cma_deref_id(id_priv);
2339 rdma_destroy_id(&id_priv->id);
2343 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms)
2345 struct rdma_route *route = &id_priv->id.route;
2346 struct cma_work *work;
2349 work = kzalloc(sizeof *work, GFP_KERNEL);
2354 INIT_WORK(&work->work, cma_work_handler);
2355 work->old_state = RDMA_CM_ROUTE_QUERY;
2356 work->new_state = RDMA_CM_ROUTE_RESOLVED;
2357 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
2359 route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
2360 if (!route->path_rec) {
2365 ret = cma_query_ib_route(id_priv, timeout_ms, work);
2371 kfree(route->path_rec);
2372 route->path_rec = NULL;
2378 int rdma_set_ib_paths(struct rdma_cm_id *id,
2379 struct ib_sa_path_rec *path_rec, int num_paths)
2381 struct rdma_id_private *id_priv;
2384 id_priv = container_of(id, struct rdma_id_private, id);
2385 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
2386 RDMA_CM_ROUTE_RESOLVED))
2389 id->route.path_rec = kmemdup(path_rec, sizeof *path_rec * num_paths,
2391 if (!id->route.path_rec) {
2396 id->route.num_paths = num_paths;
2399 cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED);
2402 EXPORT_SYMBOL(rdma_set_ib_paths);
2404 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms)
2406 struct cma_work *work;
2408 work = kzalloc(sizeof *work, GFP_KERNEL);
2413 INIT_WORK(&work->work, cma_work_handler);
2414 work->old_state = RDMA_CM_ROUTE_QUERY;
2415 work->new_state = RDMA_CM_ROUTE_RESOLVED;
2416 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
2417 queue_work(cma_wq, &work->work);
2421 static int iboe_tos_to_sl(struct net_device *ndev, int tos)
2424 struct net_device *dev;
2426 prio = rt_tos2priority(tos);
2427 dev = ndev->priv_flags & IFF_802_1Q_VLAN ?
2428 vlan_dev_real_dev(ndev) : ndev;
2431 return netdev_get_prio_tc_map(dev, prio);
2433 #if IS_ENABLED(CONFIG_VLAN_8021Q)
2434 if (ndev->priv_flags & IFF_802_1Q_VLAN)
2435 return (vlan_dev_get_egress_qos_mask(ndev, prio) &
2436 VLAN_PRIO_MASK) >> VLAN_PRIO_SHIFT;
2441 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
2443 struct rdma_route *route = &id_priv->id.route;
2444 struct rdma_addr *addr = &route->addr;
2445 struct cma_work *work;
2447 struct net_device *ndev = NULL;
2450 work = kzalloc(sizeof *work, GFP_KERNEL);
2455 INIT_WORK(&work->work, cma_work_handler);
2457 route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
2458 if (!route->path_rec) {
2463 route->num_paths = 1;
2465 if (addr->dev_addr.bound_dev_if) {
2466 ndev = dev_get_by_index(&init_net, addr->dev_addr.bound_dev_if);
2472 if (ndev->flags & IFF_LOOPBACK) {
2474 if (!id_priv->id.device->get_netdev) {
2479 ndev = id_priv->id.device->get_netdev(id_priv->id.device,
2480 id_priv->id.port_num);
2487 route->path_rec->net = &init_net;
2488 route->path_rec->ifindex = ndev->ifindex;
2489 route->path_rec->gid_type = id_priv->gid_type;
2496 memcpy(route->path_rec->dmac, addr->dev_addr.dst_dev_addr, ETH_ALEN);
2498 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
2499 &route->path_rec->sgid);
2500 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.dst_addr,
2501 &route->path_rec->dgid);
2503 /* Use the hint from IP Stack to select GID Type */
2504 if (route->path_rec->gid_type < ib_network_to_gid_type(addr->dev_addr.network))
2505 route->path_rec->gid_type = ib_network_to_gid_type(addr->dev_addr.network);
2506 if (((struct sockaddr *)&id_priv->id.route.addr.dst_addr)->sa_family != AF_IB)
2507 /* TODO: get the hoplimit from the inet/inet6 device */
2508 route->path_rec->hop_limit = addr->dev_addr.hoplimit;
2510 route->path_rec->hop_limit = 1;
2511 route->path_rec->reversible = 1;
2512 route->path_rec->pkey = cpu_to_be16(0xffff);
2513 route->path_rec->mtu_selector = IB_SA_EQ;
2514 route->path_rec->sl = iboe_tos_to_sl(ndev, id_priv->tos);
2515 route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
2516 route->path_rec->rate_selector = IB_SA_EQ;
2517 route->path_rec->rate = iboe_get_rate(ndev);
2519 route->path_rec->packet_life_time_selector = IB_SA_EQ;
2520 route->path_rec->packet_life_time = CMA_IBOE_PACKET_LIFETIME;
2521 if (!route->path_rec->mtu) {
2526 work->old_state = RDMA_CM_ROUTE_QUERY;
2527 work->new_state = RDMA_CM_ROUTE_RESOLVED;
2528 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
2529 work->event.status = 0;
2531 queue_work(cma_wq, &work->work);
2536 kfree(route->path_rec);
2537 route->path_rec = NULL;
2543 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
2545 struct rdma_id_private *id_priv;
2548 id_priv = container_of(id, struct rdma_id_private, id);
2549 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY))
2552 atomic_inc(&id_priv->refcount);
2553 if (rdma_cap_ib_sa(id->device, id->port_num))
2554 ret = cma_resolve_ib_route(id_priv, timeout_ms);
2555 else if (rdma_protocol_roce(id->device, id->port_num))
2556 ret = cma_resolve_iboe_route(id_priv);
2557 else if (rdma_protocol_iwarp(id->device, id->port_num))
2558 ret = cma_resolve_iw_route(id_priv, timeout_ms);
2567 cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED);
2568 cma_deref_id(id_priv);
2571 EXPORT_SYMBOL(rdma_resolve_route);
2573 static void cma_set_loopback(struct sockaddr *addr)
2575 switch (addr->sa_family) {
2577 ((struct sockaddr_in *) addr)->sin_addr.s_addr = htonl(INADDR_LOOPBACK);
2580 ipv6_addr_set(&((struct sockaddr_in6 *) addr)->sin6_addr,
2584 ib_addr_set(&((struct sockaddr_ib *) addr)->sib_addr,
2590 static int cma_bind_loopback(struct rdma_id_private *id_priv)
2592 struct cma_device *cma_dev, *cur_dev;
2593 struct ib_port_attr port_attr;
2601 list_for_each_entry(cur_dev, &dev_list, list) {
2602 if (cma_family(id_priv) == AF_IB &&
2603 !rdma_cap_ib_cm(cur_dev->device, 1))
2609 for (p = 1; p <= cur_dev->device->phys_port_cnt; ++p) {
2610 if (!ib_query_port(cur_dev->device, p, &port_attr) &&
2611 port_attr.state == IB_PORT_ACTIVE) {
2626 ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid, NULL);
2630 ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
2634 id_priv->id.route.addr.dev_addr.dev_type =
2635 (rdma_protocol_ib(cma_dev->device, p)) ?
2636 ARPHRD_INFINIBAND : ARPHRD_ETHER;
2638 rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2639 ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
2640 id_priv->id.port_num = p;
2641 cma_attach_to_dev(id_priv, cma_dev);
2642 cma_set_loopback(cma_src_addr(id_priv));
2644 mutex_unlock(&lock);
2648 static void addr_handler(int status, struct sockaddr *src_addr,
2649 struct rdma_dev_addr *dev_addr, void *context)
2651 struct rdma_id_private *id_priv = context;
2652 struct rdma_cm_event event;
2654 memset(&event, 0, sizeof event);
2655 mutex_lock(&id_priv->handler_mutex);
2656 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY,
2657 RDMA_CM_ADDR_RESOLVED))
2660 memcpy(cma_src_addr(id_priv), src_addr, rdma_addr_size(src_addr));
2661 if (!status && !id_priv->cma_dev)
2662 status = cma_acquire_dev(id_priv, NULL);
2665 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
2666 RDMA_CM_ADDR_BOUND))
2668 event.event = RDMA_CM_EVENT_ADDR_ERROR;
2669 event.status = status;
2671 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2673 if (id_priv->id.event_handler(&id_priv->id, &event)) {
2674 cma_exch(id_priv, RDMA_CM_DESTROYING);
2675 mutex_unlock(&id_priv->handler_mutex);
2676 cma_deref_id(id_priv);
2677 rdma_destroy_id(&id_priv->id);
2681 mutex_unlock(&id_priv->handler_mutex);
2682 cma_deref_id(id_priv);
2685 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
2687 struct cma_work *work;
2691 work = kzalloc(sizeof *work, GFP_KERNEL);
2695 if (!id_priv->cma_dev) {
2696 ret = cma_bind_loopback(id_priv);
2701 rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2702 rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
2705 INIT_WORK(&work->work, cma_work_handler);
2706 work->old_state = RDMA_CM_ADDR_QUERY;
2707 work->new_state = RDMA_CM_ADDR_RESOLVED;
2708 work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2709 queue_work(cma_wq, &work->work);
2716 static int cma_resolve_ib_addr(struct rdma_id_private *id_priv)
2718 struct cma_work *work;
2721 work = kzalloc(sizeof *work, GFP_KERNEL);
2725 if (!id_priv->cma_dev) {
2726 ret = cma_resolve_ib_dev(id_priv);
2731 rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, (union ib_gid *)
2732 &(((struct sockaddr_ib *) &id_priv->id.route.addr.dst_addr)->sib_addr));
2735 INIT_WORK(&work->work, cma_work_handler);
2736 work->old_state = RDMA_CM_ADDR_QUERY;
2737 work->new_state = RDMA_CM_ADDR_RESOLVED;
2738 work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2739 queue_work(cma_wq, &work->work);
2746 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2747 struct sockaddr *dst_addr)
2749 if (!src_addr || !src_addr->sa_family) {
2750 src_addr = (struct sockaddr *) &id->route.addr.src_addr;
2751 src_addr->sa_family = dst_addr->sa_family;
2752 if (dst_addr->sa_family == AF_INET6) {
2753 struct sockaddr_in6 *src_addr6 = (struct sockaddr_in6 *) src_addr;
2754 struct sockaddr_in6 *dst_addr6 = (struct sockaddr_in6 *) dst_addr;
2755 src_addr6->sin6_scope_id = dst_addr6->sin6_scope_id;
2756 if (ipv6_addr_type(&dst_addr6->sin6_addr) & IPV6_ADDR_LINKLOCAL)
2757 id->route.addr.dev_addr.bound_dev_if = dst_addr6->sin6_scope_id;
2758 } else if (dst_addr->sa_family == AF_IB) {
2759 ((struct sockaddr_ib *) src_addr)->sib_pkey =
2760 ((struct sockaddr_ib *) dst_addr)->sib_pkey;
2763 return rdma_bind_addr(id, src_addr);
2766 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2767 struct sockaddr *dst_addr, int timeout_ms)
2769 struct rdma_id_private *id_priv;
2772 id_priv = container_of(id, struct rdma_id_private, id);
2773 if (id_priv->state == RDMA_CM_IDLE) {
2774 ret = cma_bind_addr(id, src_addr, dst_addr);
2779 if (cma_family(id_priv) != dst_addr->sa_family)
2782 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY))
2785 atomic_inc(&id_priv->refcount);
2786 memcpy(cma_dst_addr(id_priv), dst_addr, rdma_addr_size(dst_addr));
2787 if (cma_any_addr(dst_addr)) {
2788 ret = cma_resolve_loopback(id_priv);
2790 if (dst_addr->sa_family == AF_IB) {
2791 ret = cma_resolve_ib_addr(id_priv);
2793 ret = rdma_resolve_ip(&addr_client, cma_src_addr(id_priv),
2794 dst_addr, &id->route.addr.dev_addr,
2795 timeout_ms, addr_handler, id_priv);
2803 cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
2804 cma_deref_id(id_priv);
2807 EXPORT_SYMBOL(rdma_resolve_addr);
2809 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse)
2811 struct rdma_id_private *id_priv;
2812 unsigned long flags;
2815 id_priv = container_of(id, struct rdma_id_private, id);
2816 spin_lock_irqsave(&id_priv->lock, flags);
2817 if (reuse || id_priv->state == RDMA_CM_IDLE) {
2818 id_priv->reuseaddr = reuse;
2823 spin_unlock_irqrestore(&id_priv->lock, flags);
2826 EXPORT_SYMBOL(rdma_set_reuseaddr);
2828 int rdma_set_afonly(struct rdma_cm_id *id, int afonly)
2830 struct rdma_id_private *id_priv;
2831 unsigned long flags;
2834 id_priv = container_of(id, struct rdma_id_private, id);
2835 spin_lock_irqsave(&id_priv->lock, flags);
2836 if (id_priv->state == RDMA_CM_IDLE || id_priv->state == RDMA_CM_ADDR_BOUND) {
2837 id_priv->options |= (1 << CMA_OPTION_AFONLY);
2838 id_priv->afonly = afonly;
2843 spin_unlock_irqrestore(&id_priv->lock, flags);
2846 EXPORT_SYMBOL(rdma_set_afonly);
2848 static void cma_bind_port(struct rdma_bind_list *bind_list,
2849 struct rdma_id_private *id_priv)
2851 struct sockaddr *addr;
2852 struct sockaddr_ib *sib;
2856 addr = cma_src_addr(id_priv);
2857 port = htons(bind_list->port);
2859 switch (addr->sa_family) {
2861 ((struct sockaddr_in *) addr)->sin_port = port;
2864 ((struct sockaddr_in6 *) addr)->sin6_port = port;
2867 sib = (struct sockaddr_ib *) addr;
2868 sid = be64_to_cpu(sib->sib_sid);
2869 mask = be64_to_cpu(sib->sib_sid_mask);
2870 sib->sib_sid = cpu_to_be64((sid & mask) | (u64) ntohs(port));
2871 sib->sib_sid_mask = cpu_to_be64(~0ULL);
2874 id_priv->bind_list = bind_list;
2875 hlist_add_head(&id_priv->node, &bind_list->owners);
2878 static int cma_alloc_port(enum rdma_port_space ps,
2879 struct rdma_id_private *id_priv, unsigned short snum)
2881 struct rdma_bind_list *bind_list;
2884 bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
2888 ret = cma_ps_alloc(id_priv->id.route.addr.dev_addr.net, ps, bind_list,
2894 bind_list->port = (unsigned short)ret;
2895 cma_bind_port(bind_list, id_priv);
2899 return ret == -ENOSPC ? -EADDRNOTAVAIL : ret;
2902 static int cma_alloc_any_port(enum rdma_port_space ps,
2903 struct rdma_id_private *id_priv)
2905 static unsigned int last_used_port;
2906 int low, high, remaining;
2908 struct net *net = id_priv->id.route.addr.dev_addr.net;
2910 inet_get_local_port_range(net, &low, &high);
2911 remaining = (high - low) + 1;
2912 rover = prandom_u32() % remaining + low;
2914 if (last_used_port != rover &&
2915 !cma_ps_find(net, ps, (unsigned short)rover)) {
2916 int ret = cma_alloc_port(ps, id_priv, rover);
2918 * Remember previously used port number in order to avoid
2919 * re-using same port immediately after it is closed.
2922 last_used_port = rover;
2923 if (ret != -EADDRNOTAVAIL)
2928 if ((rover < low) || (rover > high))
2932 return -EADDRNOTAVAIL;
2936 * Check that the requested port is available. This is called when trying to
2937 * bind to a specific port, or when trying to listen on a bound port. In
2938 * the latter case, the provided id_priv may already be on the bind_list, but
2939 * we still need to check that it's okay to start listening.
2941 static int cma_check_port(struct rdma_bind_list *bind_list,
2942 struct rdma_id_private *id_priv, uint8_t reuseaddr)
2944 struct rdma_id_private *cur_id;
2945 struct sockaddr *addr, *cur_addr;
2947 addr = cma_src_addr(id_priv);
2948 hlist_for_each_entry(cur_id, &bind_list->owners, node) {
2949 if (id_priv == cur_id)
2952 if ((cur_id->state != RDMA_CM_LISTEN) && reuseaddr &&
2956 cur_addr = cma_src_addr(cur_id);
2957 if (id_priv->afonly && cur_id->afonly &&
2958 (addr->sa_family != cur_addr->sa_family))
2961 if (cma_any_addr(addr) || cma_any_addr(cur_addr))
2962 return -EADDRNOTAVAIL;
2964 if (!cma_addr_cmp(addr, cur_addr))
2970 static int cma_use_port(enum rdma_port_space ps,
2971 struct rdma_id_private *id_priv)
2973 struct rdma_bind_list *bind_list;
2974 unsigned short snum;
2977 snum = ntohs(cma_port(cma_src_addr(id_priv)));
2978 if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
2981 bind_list = cma_ps_find(id_priv->id.route.addr.dev_addr.net, ps, snum);
2983 ret = cma_alloc_port(ps, id_priv, snum);
2985 ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr);
2987 cma_bind_port(bind_list, id_priv);
2992 static int cma_bind_listen(struct rdma_id_private *id_priv)
2994 struct rdma_bind_list *bind_list = id_priv->bind_list;
2998 if (bind_list->owners.first->next)
2999 ret = cma_check_port(bind_list, id_priv, 0);
3000 mutex_unlock(&lock);
3004 static enum rdma_port_space cma_select_inet_ps(
3005 struct rdma_id_private *id_priv)
3007 switch (id_priv->id.ps) {
3012 return id_priv->id.ps;
3019 static enum rdma_port_space cma_select_ib_ps(struct rdma_id_private *id_priv)
3021 enum rdma_port_space ps = 0;
3022 struct sockaddr_ib *sib;
3023 u64 sid_ps, mask, sid;
3025 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
3026 mask = be64_to_cpu(sib->sib_sid_mask) & RDMA_IB_IP_PS_MASK;
3027 sid = be64_to_cpu(sib->sib_sid) & mask;
3029 if ((id_priv->id.ps == RDMA_PS_IB) && (sid == (RDMA_IB_IP_PS_IB & mask))) {
3030 sid_ps = RDMA_IB_IP_PS_IB;
3032 } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_TCP)) &&
3033 (sid == (RDMA_IB_IP_PS_TCP & mask))) {
3034 sid_ps = RDMA_IB_IP_PS_TCP;
3036 } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_UDP)) &&
3037 (sid == (RDMA_IB_IP_PS_UDP & mask))) {
3038 sid_ps = RDMA_IB_IP_PS_UDP;
3043 sib->sib_sid = cpu_to_be64(sid_ps | ntohs(cma_port((struct sockaddr *) sib)));
3044 sib->sib_sid_mask = cpu_to_be64(RDMA_IB_IP_PS_MASK |
3045 be64_to_cpu(sib->sib_sid_mask));
3050 static int cma_get_port(struct rdma_id_private *id_priv)
3052 enum rdma_port_space ps;
3055 if (cma_family(id_priv) != AF_IB)
3056 ps = cma_select_inet_ps(id_priv);
3058 ps = cma_select_ib_ps(id_priv);
3060 return -EPROTONOSUPPORT;
3063 if (cma_any_port(cma_src_addr(id_priv)))
3064 ret = cma_alloc_any_port(ps, id_priv);
3066 ret = cma_use_port(ps, id_priv);
3067 mutex_unlock(&lock);
3072 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
3073 struct sockaddr *addr)
3075 #if IS_ENABLED(CONFIG_IPV6)
3076 struct sockaddr_in6 *sin6;
3078 if (addr->sa_family != AF_INET6)
3081 sin6 = (struct sockaddr_in6 *) addr;
3083 if (!(ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL))
3086 if (!sin6->sin6_scope_id)
3089 dev_addr->bound_dev_if = sin6->sin6_scope_id;
3094 int rdma_listen(struct rdma_cm_id *id, int backlog)
3096 struct rdma_id_private *id_priv;
3099 id_priv = container_of(id, struct rdma_id_private, id);
3100 if (id_priv->state == RDMA_CM_IDLE) {
3101 id->route.addr.src_addr.ss_family = AF_INET;
3102 ret = rdma_bind_addr(id, cma_src_addr(id_priv));
3107 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN))
3110 if (id_priv->reuseaddr) {
3111 ret = cma_bind_listen(id_priv);
3116 id_priv->backlog = backlog;
3118 if (rdma_cap_ib_cm(id->device, 1)) {
3119 ret = cma_ib_listen(id_priv);
3122 } else if (rdma_cap_iw_cm(id->device, 1)) {
3123 ret = cma_iw_listen(id_priv, backlog);
3131 cma_listen_on_all(id_priv);
3135 id_priv->backlog = 0;
3136 cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND);
3139 EXPORT_SYMBOL(rdma_listen);
3141 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
3143 struct rdma_id_private *id_priv;
3146 if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6 &&
3147 addr->sa_family != AF_IB)
3148 return -EAFNOSUPPORT;
3150 id_priv = container_of(id, struct rdma_id_private, id);
3151 if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND))
3154 ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
3158 memcpy(cma_src_addr(id_priv), addr, rdma_addr_size(addr));
3159 if (!cma_any_addr(addr)) {
3160 ret = cma_translate_addr(addr, &id->route.addr.dev_addr);
3164 ret = cma_acquire_dev(id_priv, NULL);
3169 if (!(id_priv->options & (1 << CMA_OPTION_AFONLY))) {
3170 if (addr->sa_family == AF_INET)
3171 id_priv->afonly = 1;
3172 #if IS_ENABLED(CONFIG_IPV6)
3173 else if (addr->sa_family == AF_INET6) {
3174 struct net *net = id_priv->id.route.addr.dev_addr.net;
3176 id_priv->afonly = net->ipv6.sysctl.bindv6only;
3180 ret = cma_get_port(id_priv);
3186 if (id_priv->cma_dev)
3187 cma_release_dev(id_priv);
3189 cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE);
3192 EXPORT_SYMBOL(rdma_bind_addr);
3194 static int cma_format_hdr(void *hdr, struct rdma_id_private *id_priv)
3196 struct cma_hdr *cma_hdr;
3199 cma_hdr->cma_version = CMA_VERSION;
3200 if (cma_family(id_priv) == AF_INET) {
3201 struct sockaddr_in *src4, *dst4;
3203 src4 = (struct sockaddr_in *) cma_src_addr(id_priv);
3204 dst4 = (struct sockaddr_in *) cma_dst_addr(id_priv);
3206 cma_set_ip_ver(cma_hdr, 4);
3207 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
3208 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
3209 cma_hdr->port = src4->sin_port;
3210 } else if (cma_family(id_priv) == AF_INET6) {
3211 struct sockaddr_in6 *src6, *dst6;
3213 src6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
3214 dst6 = (struct sockaddr_in6 *) cma_dst_addr(id_priv);
3216 cma_set_ip_ver(cma_hdr, 6);
3217 cma_hdr->src_addr.ip6 = src6->sin6_addr;
3218 cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
3219 cma_hdr->port = src6->sin6_port;
3224 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
3225 struct ib_cm_event *ib_event)
3227 struct rdma_id_private *id_priv = cm_id->context;
3228 struct rdma_cm_event event;
3229 struct ib_cm_sidr_rep_event_param *rep = &ib_event->param.sidr_rep_rcvd;
3232 mutex_lock(&id_priv->handler_mutex);
3233 if (id_priv->state != RDMA_CM_CONNECT)
3236 memset(&event, 0, sizeof event);
3237 switch (ib_event->event) {
3238 case IB_CM_SIDR_REQ_ERROR:
3239 event.event = RDMA_CM_EVENT_UNREACHABLE;
3240 event.status = -ETIMEDOUT;
3242 case IB_CM_SIDR_REP_RECEIVED:
3243 event.param.ud.private_data = ib_event->private_data;
3244 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
3245 if (rep->status != IB_SIDR_SUCCESS) {
3246 event.event = RDMA_CM_EVENT_UNREACHABLE;
3247 event.status = ib_event->param.sidr_rep_rcvd.status;
3250 ret = cma_set_qkey(id_priv, rep->qkey);
3252 event.event = RDMA_CM_EVENT_ADDR_ERROR;
3256 ib_init_ah_from_path(id_priv->id.device, id_priv->id.port_num,
3257 id_priv->id.route.path_rec,
3258 &event.param.ud.ah_attr);
3259 event.param.ud.qp_num = rep->qpn;
3260 event.param.ud.qkey = rep->qkey;
3261 event.event = RDMA_CM_EVENT_ESTABLISHED;
3265 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
3270 ret = id_priv->id.event_handler(&id_priv->id, &event);
3272 /* Destroy the CM ID by returning a non-zero value. */
3273 id_priv->cm_id.ib = NULL;
3274 cma_exch(id_priv, RDMA_CM_DESTROYING);
3275 mutex_unlock(&id_priv->handler_mutex);
3276 rdma_destroy_id(&id_priv->id);
3280 mutex_unlock(&id_priv->handler_mutex);
3284 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
3285 struct rdma_conn_param *conn_param)
3287 struct ib_cm_sidr_req_param req;
3288 struct ib_cm_id *id;
3292 memset(&req, 0, sizeof req);
3293 offset = cma_user_data_offset(id_priv);
3294 req.private_data_len = offset + conn_param->private_data_len;
3295 if (req.private_data_len < conn_param->private_data_len)
3298 if (req.private_data_len) {
3299 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
3303 private_data = NULL;
3306 if (conn_param->private_data && conn_param->private_data_len)
3307 memcpy(private_data + offset, conn_param->private_data,
3308 conn_param->private_data_len);
3311 ret = cma_format_hdr(private_data, id_priv);
3314 req.private_data = private_data;
3317 id = ib_create_cm_id(id_priv->id.device, cma_sidr_rep_handler,
3323 id_priv->cm_id.ib = id;
3325 req.path = id_priv->id.route.path_rec;
3326 req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
3327 req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
3328 req.max_cm_retries = CMA_MAX_CM_RETRIES;
3330 ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
3332 ib_destroy_cm_id(id_priv->cm_id.ib);
3333 id_priv->cm_id.ib = NULL;
3336 kfree(private_data);
3340 static int cma_connect_ib(struct rdma_id_private *id_priv,
3341 struct rdma_conn_param *conn_param)
3343 struct ib_cm_req_param req;
3344 struct rdma_route *route;
3346 struct ib_cm_id *id;
3349 memset(&req, 0, sizeof req);
3350 offset = cma_user_data_offset(id_priv);
3351 req.private_data_len = offset + conn_param->private_data_len;
3352 if (req.private_data_len < conn_param->private_data_len)
3355 if (req.private_data_len) {
3356 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
3360 private_data = NULL;
3363 if (conn_param->private_data && conn_param->private_data_len)
3364 memcpy(private_data + offset, conn_param->private_data,
3365 conn_param->private_data_len);
3367 id = ib_create_cm_id(id_priv->id.device, cma_ib_handler, id_priv);
3372 id_priv->cm_id.ib = id;
3374 route = &id_priv->id.route;
3376 ret = cma_format_hdr(private_data, id_priv);
3379 req.private_data = private_data;
3382 req.primary_path = &route->path_rec[0];
3383 if (route->num_paths == 2)
3384 req.alternate_path = &route->path_rec[1];
3386 req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
3387 req.qp_num = id_priv->qp_num;
3388 req.qp_type = id_priv->id.qp_type;
3389 req.starting_psn = id_priv->seq_num;
3390 req.responder_resources = conn_param->responder_resources;
3391 req.initiator_depth = conn_param->initiator_depth;
3392 req.flow_control = conn_param->flow_control;
3393 req.retry_count = min_t(u8, 7, conn_param->retry_count);
3394 req.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
3395 req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
3396 req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
3397 req.max_cm_retries = CMA_MAX_CM_RETRIES;
3398 req.srq = id_priv->srq ? 1 : 0;
3400 ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
3402 if (ret && !IS_ERR(id)) {
3403 ib_destroy_cm_id(id);
3404 id_priv->cm_id.ib = NULL;
3407 kfree(private_data);
3411 static int cma_connect_iw(struct rdma_id_private *id_priv,
3412 struct rdma_conn_param *conn_param)
3414 struct iw_cm_id *cm_id;
3416 struct iw_cm_conn_param iw_param;
3418 cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
3420 return PTR_ERR(cm_id);
3422 cm_id->tos = id_priv->tos;
3423 id_priv->cm_id.iw = cm_id;
3425 memcpy(&cm_id->local_addr, cma_src_addr(id_priv),
3426 rdma_addr_size(cma_src_addr(id_priv)));
3427 memcpy(&cm_id->remote_addr, cma_dst_addr(id_priv),
3428 rdma_addr_size(cma_dst_addr(id_priv)));
3430 ret = cma_modify_qp_rtr(id_priv, conn_param);
3435 iw_param.ord = conn_param->initiator_depth;
3436 iw_param.ird = conn_param->responder_resources;
3437 iw_param.private_data = conn_param->private_data;
3438 iw_param.private_data_len = conn_param->private_data_len;
3439 iw_param.qpn = id_priv->id.qp ? id_priv->qp_num : conn_param->qp_num;
3441 memset(&iw_param, 0, sizeof iw_param);
3442 iw_param.qpn = id_priv->qp_num;
3444 ret = iw_cm_connect(cm_id, &iw_param);
3447 iw_destroy_cm_id(cm_id);
3448 id_priv->cm_id.iw = NULL;
3453 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
3455 struct rdma_id_private *id_priv;
3458 id_priv = container_of(id, struct rdma_id_private, id);
3459 if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT))
3463 id_priv->qp_num = conn_param->qp_num;
3464 id_priv->srq = conn_param->srq;
3467 if (rdma_cap_ib_cm(id->device, id->port_num)) {
3468 if (id->qp_type == IB_QPT_UD)
3469 ret = cma_resolve_ib_udp(id_priv, conn_param);
3471 ret = cma_connect_ib(id_priv, conn_param);
3472 } else if (rdma_cap_iw_cm(id->device, id->port_num))
3473 ret = cma_connect_iw(id_priv, conn_param);
3481 cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED);
3484 EXPORT_SYMBOL(rdma_connect);
3486 static int cma_accept_ib(struct rdma_id_private *id_priv,
3487 struct rdma_conn_param *conn_param)
3489 struct ib_cm_rep_param rep;
3492 ret = cma_modify_qp_rtr(id_priv, conn_param);
3496 ret = cma_modify_qp_rts(id_priv, conn_param);
3500 memset(&rep, 0, sizeof rep);
3501 rep.qp_num = id_priv->qp_num;
3502 rep.starting_psn = id_priv->seq_num;
3503 rep.private_data = conn_param->private_data;
3504 rep.private_data_len = conn_param->private_data_len;
3505 rep.responder_resources = conn_param->responder_resources;
3506 rep.initiator_depth = conn_param->initiator_depth;
3507 rep.failover_accepted = 0;
3508 rep.flow_control = conn_param->flow_control;
3509 rep.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
3510 rep.srq = id_priv->srq ? 1 : 0;
3512 ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
3517 static int cma_accept_iw(struct rdma_id_private *id_priv,
3518 struct rdma_conn_param *conn_param)
3520 struct iw_cm_conn_param iw_param;
3523 ret = cma_modify_qp_rtr(id_priv, conn_param);
3527 iw_param.ord = conn_param->initiator_depth;
3528 iw_param.ird = conn_param->responder_resources;
3529 iw_param.private_data = conn_param->private_data;
3530 iw_param.private_data_len = conn_param->private_data_len;
3531 if (id_priv->id.qp) {
3532 iw_param.qpn = id_priv->qp_num;
3534 iw_param.qpn = conn_param->qp_num;
3536 return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
3539 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
3540 enum ib_cm_sidr_status status, u32 qkey,
3541 const void *private_data, int private_data_len)
3543 struct ib_cm_sidr_rep_param rep;
3546 memset(&rep, 0, sizeof rep);
3547 rep.status = status;
3548 if (status == IB_SIDR_SUCCESS) {
3549 ret = cma_set_qkey(id_priv, qkey);
3552 rep.qp_num = id_priv->qp_num;
3553 rep.qkey = id_priv->qkey;
3555 rep.private_data = private_data;
3556 rep.private_data_len = private_data_len;
3558 return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
3561 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
3563 struct rdma_id_private *id_priv;
3566 id_priv = container_of(id, struct rdma_id_private, id);
3568 id_priv->owner = task_pid_nr(current);
3570 if (!cma_comp(id_priv, RDMA_CM_CONNECT))
3573 if (!id->qp && conn_param) {
3574 id_priv->qp_num = conn_param->qp_num;
3575 id_priv->srq = conn_param->srq;
3578 if (rdma_cap_ib_cm(id->device, id->port_num)) {
3579 if (id->qp_type == IB_QPT_UD) {
3581 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
3583 conn_param->private_data,
3584 conn_param->private_data_len);
3586 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
3590 ret = cma_accept_ib(id_priv, conn_param);
3592 ret = cma_rep_recv(id_priv);
3594 } else if (rdma_cap_iw_cm(id->device, id->port_num))
3595 ret = cma_accept_iw(id_priv, conn_param);
3604 cma_modify_qp_err(id_priv);
3605 rdma_reject(id, NULL, 0);
3608 EXPORT_SYMBOL(rdma_accept);
3610 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
3612 struct rdma_id_private *id_priv;
3615 id_priv = container_of(id, struct rdma_id_private, id);
3616 if (!id_priv->cm_id.ib)
3619 switch (id->device->node_type) {
3620 case RDMA_NODE_IB_CA:
3621 ret = ib_cm_notify(id_priv->cm_id.ib, event);
3629 EXPORT_SYMBOL(rdma_notify);
3631 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
3632 u8 private_data_len)
3634 struct rdma_id_private *id_priv;
3637 id_priv = container_of(id, struct rdma_id_private, id);
3638 if (!id_priv->cm_id.ib)
3641 if (rdma_cap_ib_cm(id->device, id->port_num)) {
3642 if (id->qp_type == IB_QPT_UD)
3643 ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT, 0,
3644 private_data, private_data_len);
3646 ret = ib_send_cm_rej(id_priv->cm_id.ib,
3647 IB_CM_REJ_CONSUMER_DEFINED, NULL,
3648 0, private_data, private_data_len);
3649 } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
3650 ret = iw_cm_reject(id_priv->cm_id.iw,
3651 private_data, private_data_len);
3657 EXPORT_SYMBOL(rdma_reject);
3659 int rdma_disconnect(struct rdma_cm_id *id)
3661 struct rdma_id_private *id_priv;
3664 id_priv = container_of(id, struct rdma_id_private, id);
3665 if (!id_priv->cm_id.ib)
3668 if (rdma_cap_ib_cm(id->device, id->port_num)) {
3669 ret = cma_modify_qp_err(id_priv);
3672 /* Initiate or respond to a disconnect. */
3673 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
3674 ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
3675 } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
3676 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
3683 EXPORT_SYMBOL(rdma_disconnect);
3685 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
3687 struct rdma_id_private *id_priv;
3688 struct cma_multicast *mc = multicast->context;
3689 struct rdma_cm_event event;
3692 id_priv = mc->id_priv;
3693 mutex_lock(&id_priv->handler_mutex);
3694 if (id_priv->state != RDMA_CM_ADDR_BOUND &&
3695 id_priv->state != RDMA_CM_ADDR_RESOLVED)
3699 status = cma_set_qkey(id_priv, be32_to_cpu(multicast->rec.qkey));
3700 mutex_lock(&id_priv->qp_mutex);
3701 if (!status && id_priv->id.qp)
3702 status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid,
3703 be16_to_cpu(multicast->rec.mlid));
3704 mutex_unlock(&id_priv->qp_mutex);
3706 memset(&event, 0, sizeof event);
3707 event.status = status;
3708 event.param.ud.private_data = mc->context;
3710 struct rdma_dev_addr *dev_addr =
3711 &id_priv->id.route.addr.dev_addr;
3712 struct net_device *ndev =
3713 dev_get_by_index(&init_net, dev_addr->bound_dev_if);
3714 enum ib_gid_type gid_type =
3715 id_priv->cma_dev->default_gid_type[id_priv->id.port_num -
3716 rdma_start_port(id_priv->cma_dev->device)];
3718 event.event = RDMA_CM_EVENT_MULTICAST_JOIN;
3719 ib_init_ah_from_mcmember(id_priv->id.device,
3720 id_priv->id.port_num, &multicast->rec,
3722 &event.param.ud.ah_attr);
3723 event.param.ud.qp_num = 0xFFFFFF;
3724 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
3728 event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
3730 ret = id_priv->id.event_handler(&id_priv->id, &event);
3732 cma_exch(id_priv, RDMA_CM_DESTROYING);
3733 mutex_unlock(&id_priv->handler_mutex);
3734 rdma_destroy_id(&id_priv->id);
3739 mutex_unlock(&id_priv->handler_mutex);
3743 static void cma_set_mgid(struct rdma_id_private *id_priv,
3744 struct sockaddr *addr, union ib_gid *mgid)
3746 unsigned char mc_map[MAX_ADDR_LEN];
3747 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3748 struct sockaddr_in *sin = (struct sockaddr_in *) addr;
3749 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
3751 if (cma_any_addr(addr)) {
3752 memset(mgid, 0, sizeof *mgid);
3753 } else if ((addr->sa_family == AF_INET6) &&
3754 ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
3756 /* IPv6 address is an SA assigned MGID. */
3757 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3758 } else if (addr->sa_family == AF_IB) {
3759 memcpy(mgid, &((struct sockaddr_ib *) addr)->sib_addr, sizeof *mgid);
3760 } else if ((addr->sa_family == AF_INET6)) {
3761 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
3762 if (id_priv->id.ps == RDMA_PS_UDP)
3763 mc_map[7] = 0x01; /* Use RDMA CM signature */
3764 *mgid = *(union ib_gid *) (mc_map + 4);
3766 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
3767 if (id_priv->id.ps == RDMA_PS_UDP)
3768 mc_map[7] = 0x01; /* Use RDMA CM signature */
3769 *mgid = *(union ib_gid *) (mc_map + 4);
3773 static void cma_query_sa_classport_info_cb(int status,
3774 struct ib_class_port_info *rec,
3777 struct class_port_info_context *cb_ctx = context;
3781 if (status || !rec) {
3782 pr_debug("RDMA CM: %s port %u failed query ClassPortInfo status: %d\n",
3783 cb_ctx->device->name, cb_ctx->port_num, status);
3787 memcpy(cb_ctx->class_port_info, rec, sizeof(struct ib_class_port_info));
3790 complete(&cb_ctx->done);
3793 static int cma_query_sa_classport_info(struct ib_device *device, u8 port_num,
3794 struct ib_class_port_info *class_port_info)
3796 struct class_port_info_context *cb_ctx;
3799 cb_ctx = kmalloc(sizeof(*cb_ctx), GFP_KERNEL);
3803 cb_ctx->device = device;
3804 cb_ctx->class_port_info = class_port_info;
3805 cb_ctx->port_num = port_num;
3806 init_completion(&cb_ctx->done);
3808 ret = ib_sa_classport_info_rec_query(&sa_client, device, port_num,
3809 CMA_QUERY_CLASSPORT_INFO_TIMEOUT,
3810 GFP_KERNEL, cma_query_sa_classport_info_cb,
3811 cb_ctx, &cb_ctx->sa_query);
3813 pr_err("RDMA CM: %s port %u failed to send ClassPortInfo query, ret: %d\n",
3814 device->name, port_num, ret);
3818 wait_for_completion(&cb_ctx->done);
3825 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
3826 struct cma_multicast *mc)
3828 struct ib_sa_mcmember_rec rec;
3829 struct ib_class_port_info class_port_info;
3830 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3831 ib_sa_comp_mask comp_mask;
3834 ib_addr_get_mgid(dev_addr, &rec.mgid);
3835 ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
3840 ret = cma_set_qkey(id_priv, 0);
3844 cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
3845 rec.qkey = cpu_to_be32(id_priv->qkey);
3846 rdma_addr_get_sgid(dev_addr, &rec.port_gid);
3847 rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
3848 rec.join_state = mc->join_state;
3850 if (rec.join_state == BIT(SENDONLY_FULLMEMBER_JOIN)) {
3851 ret = cma_query_sa_classport_info(id_priv->id.device,
3852 id_priv->id.port_num,
3858 if (!(ib_get_cpi_capmask2(&class_port_info) &
3859 IB_SA_CAP_MASK2_SENDONLY_FULL_MEM_SUPPORT)) {
3860 pr_warn("RDMA CM: %s port %u Unable to multicast join\n"
3861 "RDMA CM: SM doesn't support Send Only Full Member option\n",
3862 id_priv->id.device->name, id_priv->id.port_num);
3867 comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
3868 IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
3869 IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
3870 IB_SA_MCMEMBER_REC_FLOW_LABEL |
3871 IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
3873 if (id_priv->id.ps == RDMA_PS_IPOIB)
3874 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
3875 IB_SA_MCMEMBER_REC_RATE_SELECTOR |
3876 IB_SA_MCMEMBER_REC_MTU_SELECTOR |
3877 IB_SA_MCMEMBER_REC_MTU |
3878 IB_SA_MCMEMBER_REC_HOP_LIMIT;
3880 mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device,
3881 id_priv->id.port_num, &rec,
3882 comp_mask, GFP_KERNEL,
3883 cma_ib_mc_handler, mc);
3884 return PTR_ERR_OR_ZERO(mc->multicast.ib);
3887 static void iboe_mcast_work_handler(struct work_struct *work)
3889 struct iboe_mcast_work *mw = container_of(work, struct iboe_mcast_work, work);
3890 struct cma_multicast *mc = mw->mc;
3891 struct ib_sa_multicast *m = mc->multicast.ib;
3893 mc->multicast.ib->context = mc;
3894 cma_ib_mc_handler(0, m);
3895 kref_put(&mc->mcref, release_mc);
3899 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid)
3901 struct sockaddr_in *sin = (struct sockaddr_in *)addr;
3902 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
3904 if (cma_any_addr(addr)) {
3905 memset(mgid, 0, sizeof *mgid);
3906 } else if (addr->sa_family == AF_INET6) {
3907 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3909 mgid->raw[0] = 0xff;
3910 mgid->raw[1] = 0x0e;
3919 mgid->raw[10] = 0xff;
3920 mgid->raw[11] = 0xff;
3921 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
3925 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
3926 struct cma_multicast *mc)
3928 struct iboe_mcast_work *work;
3929 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3931 struct sockaddr *addr = (struct sockaddr *)&mc->addr;
3932 struct net_device *ndev = NULL;
3933 enum ib_gid_type gid_type;
3936 send_only = mc->join_state == BIT(SENDONLY_FULLMEMBER_JOIN);
3938 if (cma_zero_addr((struct sockaddr *)&mc->addr))
3941 work = kzalloc(sizeof *work, GFP_KERNEL);
3945 mc->multicast.ib = kzalloc(sizeof(struct ib_sa_multicast), GFP_KERNEL);
3946 if (!mc->multicast.ib) {
3951 cma_iboe_set_mgid(addr, &mc->multicast.ib->rec.mgid);
3953 mc->multicast.ib->rec.pkey = cpu_to_be16(0xffff);
3954 if (id_priv->id.ps == RDMA_PS_UDP)
3955 mc->multicast.ib->rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
3957 if (dev_addr->bound_dev_if)
3958 ndev = dev_get_by_index(&init_net, dev_addr->bound_dev_if);
3963 mc->multicast.ib->rec.rate = iboe_get_rate(ndev);
3964 mc->multicast.ib->rec.hop_limit = 1;
3965 mc->multicast.ib->rec.mtu = iboe_get_mtu(ndev->mtu);
3967 gid_type = id_priv->cma_dev->default_gid_type[id_priv->id.port_num -
3968 rdma_start_port(id_priv->cma_dev->device)];
3969 if (addr->sa_family == AF_INET) {
3970 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) {
3971 mc->multicast.ib->rec.hop_limit = IPV6_DEFAULT_HOPLIMIT;
3973 err = cma_igmp_send(ndev, &mc->multicast.ib->rec.mgid,
3976 mc->igmp_joined = true;
3980 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP)
3984 if (err || !mc->multicast.ib->rec.mtu) {
3989 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
3990 &mc->multicast.ib->rec.port_gid);
3993 INIT_WORK(&work->work, iboe_mcast_work_handler);
3994 kref_get(&mc->mcref);
3995 queue_work(cma_wq, &work->work);
4000 kfree(mc->multicast.ib);
4006 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
4007 u8 join_state, void *context)
4009 struct rdma_id_private *id_priv;
4010 struct cma_multicast *mc;
4013 id_priv = container_of(id, struct rdma_id_private, id);
4014 if (!cma_comp(id_priv, RDMA_CM_ADDR_BOUND) &&
4015 !cma_comp(id_priv, RDMA_CM_ADDR_RESOLVED))
4018 mc = kmalloc(sizeof *mc, GFP_KERNEL);
4022 memcpy(&mc->addr, addr, rdma_addr_size(addr));
4023 mc->context = context;
4024 mc->id_priv = id_priv;
4025 mc->igmp_joined = false;
4026 mc->join_state = join_state;
4027 spin_lock(&id_priv->lock);
4028 list_add(&mc->list, &id_priv->mc_list);
4029 spin_unlock(&id_priv->lock);
4031 if (rdma_protocol_roce(id->device, id->port_num)) {
4032 kref_init(&mc->mcref);
4033 ret = cma_iboe_join_multicast(id_priv, mc);
4034 } else if (rdma_cap_ib_mcast(id->device, id->port_num))
4035 ret = cma_join_ib_multicast(id_priv, mc);
4040 spin_lock_irq(&id_priv->lock);
4041 list_del(&mc->list);
4042 spin_unlock_irq(&id_priv->lock);
4047 EXPORT_SYMBOL(rdma_join_multicast);
4049 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
4051 struct rdma_id_private *id_priv;
4052 struct cma_multicast *mc;
4054 id_priv = container_of(id, struct rdma_id_private, id);
4055 spin_lock_irq(&id_priv->lock);
4056 list_for_each_entry(mc, &id_priv->mc_list, list) {
4057 if (!memcmp(&mc->addr, addr, rdma_addr_size(addr))) {
4058 list_del(&mc->list);
4059 spin_unlock_irq(&id_priv->lock);
4062 ib_detach_mcast(id->qp,
4063 &mc->multicast.ib->rec.mgid,
4064 be16_to_cpu(mc->multicast.ib->rec.mlid));
4066 BUG_ON(id_priv->cma_dev->device != id->device);
4068 if (rdma_cap_ib_mcast(id->device, id->port_num)) {
4069 ib_sa_free_multicast(mc->multicast.ib);
4071 } else if (rdma_protocol_roce(id->device, id->port_num)) {
4072 if (mc->igmp_joined) {
4073 struct rdma_dev_addr *dev_addr =
4074 &id->route.addr.dev_addr;
4075 struct net_device *ndev = NULL;
4077 if (dev_addr->bound_dev_if)
4078 ndev = dev_get_by_index(&init_net,
4079 dev_addr->bound_dev_if);
4082 &mc->multicast.ib->rec.mgid,
4086 mc->igmp_joined = false;
4088 kref_put(&mc->mcref, release_mc);
4093 spin_unlock_irq(&id_priv->lock);
4095 EXPORT_SYMBOL(rdma_leave_multicast);
4097 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
4099 struct rdma_dev_addr *dev_addr;
4100 struct cma_ndev_work *work;
4102 dev_addr = &id_priv->id.route.addr.dev_addr;
4104 if ((dev_addr->bound_dev_if == ndev->ifindex) &&
4105 (net_eq(dev_net(ndev), dev_addr->net)) &&
4106 memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
4107 pr_info("RDMA CM addr change for ndev %s used by id %p\n",
4108 ndev->name, &id_priv->id);
4109 work = kzalloc(sizeof *work, GFP_KERNEL);
4113 INIT_WORK(&work->work, cma_ndev_work_handler);
4115 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
4116 atomic_inc(&id_priv->refcount);
4117 queue_work(cma_wq, &work->work);
4123 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
4126 struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
4127 struct cma_device *cma_dev;
4128 struct rdma_id_private *id_priv;
4129 int ret = NOTIFY_DONE;
4131 if (event != NETDEV_BONDING_FAILOVER)
4134 if (!(ndev->flags & IFF_MASTER) || !(ndev->priv_flags & IFF_BONDING))
4138 list_for_each_entry(cma_dev, &dev_list, list)
4139 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
4140 ret = cma_netdev_change(ndev, id_priv);
4146 mutex_unlock(&lock);
4150 static struct notifier_block cma_nb = {
4151 .notifier_call = cma_netdev_callback
4154 static void cma_add_one(struct ib_device *device)
4156 struct cma_device *cma_dev;
4157 struct rdma_id_private *id_priv;
4159 unsigned long supported_gids = 0;
4161 cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
4165 cma_dev->device = device;
4166 cma_dev->default_gid_type = kcalloc(device->phys_port_cnt,
4167 sizeof(*cma_dev->default_gid_type),
4169 if (!cma_dev->default_gid_type) {
4173 for (i = rdma_start_port(device); i <= rdma_end_port(device); i++) {
4174 supported_gids = roce_gid_type_mask_support(device, i);
4175 WARN_ON(!supported_gids);
4176 cma_dev->default_gid_type[i - rdma_start_port(device)] =
4177 find_first_bit(&supported_gids, BITS_PER_LONG);
4180 init_completion(&cma_dev->comp);
4181 atomic_set(&cma_dev->refcount, 1);
4182 INIT_LIST_HEAD(&cma_dev->id_list);
4183 ib_set_client_data(device, &cma_client, cma_dev);
4186 list_add_tail(&cma_dev->list, &dev_list);
4187 list_for_each_entry(id_priv, &listen_any_list, list)
4188 cma_listen_on_dev(id_priv, cma_dev);
4189 mutex_unlock(&lock);
4192 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
4194 struct rdma_cm_event event;
4195 enum rdma_cm_state state;
4198 /* Record that we want to remove the device */
4199 state = cma_exch(id_priv, RDMA_CM_DEVICE_REMOVAL);
4200 if (state == RDMA_CM_DESTROYING)
4203 cma_cancel_operation(id_priv, state);
4204 mutex_lock(&id_priv->handler_mutex);
4206 /* Check for destruction from another callback. */
4207 if (!cma_comp(id_priv, RDMA_CM_DEVICE_REMOVAL))
4210 memset(&event, 0, sizeof event);
4211 event.event = RDMA_CM_EVENT_DEVICE_REMOVAL;
4212 ret = id_priv->id.event_handler(&id_priv->id, &event);
4214 mutex_unlock(&id_priv->handler_mutex);
4218 static void cma_process_remove(struct cma_device *cma_dev)
4220 struct rdma_id_private *id_priv;
4224 while (!list_empty(&cma_dev->id_list)) {
4225 id_priv = list_entry(cma_dev->id_list.next,
4226 struct rdma_id_private, list);
4228 list_del(&id_priv->listen_list);
4229 list_del_init(&id_priv->list);
4230 atomic_inc(&id_priv->refcount);
4231 mutex_unlock(&lock);
4233 ret = id_priv->internal_id ? 1 : cma_remove_id_dev(id_priv);
4234 cma_deref_id(id_priv);
4236 rdma_destroy_id(&id_priv->id);
4240 mutex_unlock(&lock);
4242 cma_deref_dev(cma_dev);
4243 wait_for_completion(&cma_dev->comp);
4246 static void cma_remove_one(struct ib_device *device, void *client_data)
4248 struct cma_device *cma_dev = client_data;
4254 list_del(&cma_dev->list);
4255 mutex_unlock(&lock);
4257 cma_process_remove(cma_dev);
4258 kfree(cma_dev->default_gid_type);
4262 static int cma_get_id_stats(struct sk_buff *skb, struct netlink_callback *cb)
4264 struct nlmsghdr *nlh;
4265 struct rdma_cm_id_stats *id_stats;
4266 struct rdma_id_private *id_priv;
4267 struct rdma_cm_id *id = NULL;
4268 struct cma_device *cma_dev;
4269 int i_dev = 0, i_id = 0;
4272 * We export all of the IDs as a sequence of messages. Each
4273 * ID gets its own netlink message.
4277 list_for_each_entry(cma_dev, &dev_list, list) {
4278 if (i_dev < cb->args[0]) {
4284 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
4285 if (i_id < cb->args[1]) {
4290 id_stats = ibnl_put_msg(skb, &nlh, cb->nlh->nlmsg_seq,
4291 sizeof *id_stats, RDMA_NL_RDMA_CM,
4292 RDMA_NL_RDMA_CM_ID_STATS,
4297 memset(id_stats, 0, sizeof *id_stats);
4299 id_stats->node_type = id->route.addr.dev_addr.dev_type;
4300 id_stats->port_num = id->port_num;
4301 id_stats->bound_dev_if =
4302 id->route.addr.dev_addr.bound_dev_if;
4304 if (ibnl_put_attr(skb, nlh,
4305 rdma_addr_size(cma_src_addr(id_priv)),
4306 cma_src_addr(id_priv),
4307 RDMA_NL_RDMA_CM_ATTR_SRC_ADDR))
4309 if (ibnl_put_attr(skb, nlh,
4310 rdma_addr_size(cma_src_addr(id_priv)),
4311 cma_dst_addr(id_priv),
4312 RDMA_NL_RDMA_CM_ATTR_DST_ADDR))
4315 id_stats->pid = id_priv->owner;
4316 id_stats->port_space = id->ps;
4317 id_stats->cm_state = id_priv->state;
4318 id_stats->qp_num = id_priv->qp_num;
4319 id_stats->qp_type = id->qp_type;
4329 mutex_unlock(&lock);
4330 cb->args[0] = i_dev;
4336 static const struct ibnl_client_cbs cma_cb_table[] = {
4337 [RDMA_NL_RDMA_CM_ID_STATS] = { .dump = cma_get_id_stats,
4338 .module = THIS_MODULE },
4341 static int cma_init_net(struct net *net)
4343 struct cma_pernet *pernet = cma_pernet(net);
4345 idr_init(&pernet->tcp_ps);
4346 idr_init(&pernet->udp_ps);
4347 idr_init(&pernet->ipoib_ps);
4348 idr_init(&pernet->ib_ps);
4353 static void cma_exit_net(struct net *net)
4355 struct cma_pernet *pernet = cma_pernet(net);
4357 idr_destroy(&pernet->tcp_ps);
4358 idr_destroy(&pernet->udp_ps);
4359 idr_destroy(&pernet->ipoib_ps);
4360 idr_destroy(&pernet->ib_ps);
4363 static struct pernet_operations cma_pernet_operations = {
4364 .init = cma_init_net,
4365 .exit = cma_exit_net,
4366 .id = &cma_pernet_id,
4367 .size = sizeof(struct cma_pernet),
4370 static int __init cma_init(void)
4374 cma_wq = alloc_ordered_workqueue("rdma_cm", WQ_MEM_RECLAIM);
4378 ret = register_pernet_subsys(&cma_pernet_operations);
4382 ib_sa_register_client(&sa_client);
4383 rdma_addr_register_client(&addr_client);
4384 register_netdevice_notifier(&cma_nb);
4386 ret = ib_register_client(&cma_client);
4390 if (ibnl_add_client(RDMA_NL_RDMA_CM, ARRAY_SIZE(cma_cb_table),
4392 pr_warn("RDMA CMA: failed to add netlink callback\n");
4393 cma_configfs_init();
4398 unregister_netdevice_notifier(&cma_nb);
4399 rdma_addr_unregister_client(&addr_client);
4400 ib_sa_unregister_client(&sa_client);
4402 destroy_workqueue(cma_wq);
4406 static void __exit cma_cleanup(void)
4408 cma_configfs_exit();
4409 ibnl_remove_client(RDMA_NL_RDMA_CM);
4410 ib_unregister_client(&cma_client);
4411 unregister_netdevice_notifier(&cma_nb);
4412 rdma_addr_unregister_client(&addr_client);
4413 ib_sa_unregister_client(&sa_client);
4414 unregister_pernet_subsys(&cma_pernet_operations);
4415 destroy_workqueue(cma_wq);
4418 module_init(cma_init);
4419 module_exit(cma_cleanup);