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
75 #define CMA_PREFERRED_ROCE_GID_TYPE IB_GID_TYPE_ROCE_UDP_ENCAP
77 static const char * const cma_events[] = {
78 [RDMA_CM_EVENT_ADDR_RESOLVED] = "address resolved",
79 [RDMA_CM_EVENT_ADDR_ERROR] = "address error",
80 [RDMA_CM_EVENT_ROUTE_RESOLVED] = "route resolved ",
81 [RDMA_CM_EVENT_ROUTE_ERROR] = "route error",
82 [RDMA_CM_EVENT_CONNECT_REQUEST] = "connect request",
83 [RDMA_CM_EVENT_CONNECT_RESPONSE] = "connect response",
84 [RDMA_CM_EVENT_CONNECT_ERROR] = "connect error",
85 [RDMA_CM_EVENT_UNREACHABLE] = "unreachable",
86 [RDMA_CM_EVENT_REJECTED] = "rejected",
87 [RDMA_CM_EVENT_ESTABLISHED] = "established",
88 [RDMA_CM_EVENT_DISCONNECTED] = "disconnected",
89 [RDMA_CM_EVENT_DEVICE_REMOVAL] = "device removal",
90 [RDMA_CM_EVENT_MULTICAST_JOIN] = "multicast join",
91 [RDMA_CM_EVENT_MULTICAST_ERROR] = "multicast error",
92 [RDMA_CM_EVENT_ADDR_CHANGE] = "address change",
93 [RDMA_CM_EVENT_TIMEWAIT_EXIT] = "timewait exit",
96 const char *__attribute_const__ rdma_event_msg(enum rdma_cm_event_type event)
100 return (index < ARRAY_SIZE(cma_events) && cma_events[index]) ?
101 cma_events[index] : "unrecognized event";
103 EXPORT_SYMBOL(rdma_event_msg);
105 const char *__attribute_const__ rdma_reject_msg(struct rdma_cm_id *id,
108 if (rdma_ib_or_roce(id->device, id->port_num))
109 return ibcm_reject_msg(reason);
111 if (rdma_protocol_iwarp(id->device, id->port_num))
112 return iwcm_reject_msg(reason);
115 return "unrecognized transport";
117 EXPORT_SYMBOL(rdma_reject_msg);
119 bool rdma_is_consumer_reject(struct rdma_cm_id *id, int reason)
121 if (rdma_ib_or_roce(id->device, id->port_num))
122 return reason == IB_CM_REJ_CONSUMER_DEFINED;
124 if (rdma_protocol_iwarp(id->device, id->port_num))
125 return reason == -ECONNREFUSED;
130 EXPORT_SYMBOL(rdma_is_consumer_reject);
132 const void *rdma_consumer_reject_data(struct rdma_cm_id *id,
133 struct rdma_cm_event *ev, u8 *data_len)
137 if (rdma_is_consumer_reject(id, ev->status)) {
138 *data_len = ev->param.conn.private_data_len;
139 p = ev->param.conn.private_data;
146 EXPORT_SYMBOL(rdma_consumer_reject_data);
148 static void cma_add_one(struct ib_device *device);
149 static void cma_remove_one(struct ib_device *device, void *client_data);
151 static struct ib_client cma_client = {
154 .remove = cma_remove_one
157 static struct ib_sa_client sa_client;
158 static struct rdma_addr_client addr_client;
159 static LIST_HEAD(dev_list);
160 static LIST_HEAD(listen_any_list);
161 static DEFINE_MUTEX(lock);
162 static struct workqueue_struct *cma_wq;
163 static unsigned int cma_pernet_id;
172 static struct cma_pernet *cma_pernet(struct net *net)
174 return net_generic(net, cma_pernet_id);
177 static struct idr *cma_pernet_idr(struct net *net, enum rdma_port_space ps)
179 struct cma_pernet *pernet = cma_pernet(net);
183 return &pernet->tcp_ps;
185 return &pernet->udp_ps;
187 return &pernet->ipoib_ps;
189 return &pernet->ib_ps;
196 struct list_head list;
197 struct ib_device *device;
198 struct completion comp;
200 struct list_head id_list;
201 enum ib_gid_type *default_gid_type;
202 u8 *default_roce_tos;
205 struct rdma_bind_list {
206 enum rdma_port_space ps;
207 struct hlist_head owners;
211 struct class_port_info_context {
212 struct ib_class_port_info *class_port_info;
213 struct ib_device *device;
214 struct completion done;
215 struct ib_sa_query *sa_query;
219 static int cma_ps_alloc(struct net *net, enum rdma_port_space ps,
220 struct rdma_bind_list *bind_list, int snum)
222 struct idr *idr = cma_pernet_idr(net, ps);
224 return idr_alloc(idr, bind_list, snum, snum + 1, GFP_KERNEL);
227 static struct rdma_bind_list *cma_ps_find(struct net *net,
228 enum rdma_port_space ps, int snum)
230 struct idr *idr = cma_pernet_idr(net, ps);
232 return idr_find(idr, snum);
235 static void cma_ps_remove(struct net *net, enum rdma_port_space ps, int snum)
237 struct idr *idr = cma_pernet_idr(net, ps);
239 idr_remove(idr, snum);
246 void cma_ref_dev(struct cma_device *cma_dev)
248 atomic_inc(&cma_dev->refcount);
251 struct cma_device *cma_enum_devices_by_ibdev(cma_device_filter filter,
254 struct cma_device *cma_dev;
255 struct cma_device *found_cma_dev = NULL;
259 list_for_each_entry(cma_dev, &dev_list, list)
260 if (filter(cma_dev->device, cookie)) {
261 found_cma_dev = cma_dev;
266 cma_ref_dev(found_cma_dev);
268 return found_cma_dev;
271 int cma_get_default_gid_type(struct cma_device *cma_dev,
274 if (!rdma_is_port_valid(cma_dev->device, port))
277 return cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)];
280 int cma_set_default_gid_type(struct cma_device *cma_dev,
282 enum ib_gid_type default_gid_type)
284 unsigned long supported_gids;
286 if (!rdma_is_port_valid(cma_dev->device, port))
289 supported_gids = roce_gid_type_mask_support(cma_dev->device, port);
291 if (!(supported_gids & 1 << default_gid_type))
294 cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)] =
300 int cma_get_default_roce_tos(struct cma_device *cma_dev, unsigned int port)
302 if (!rdma_is_port_valid(cma_dev->device, port))
305 return cma_dev->default_roce_tos[port - rdma_start_port(cma_dev->device)];
308 int cma_set_default_roce_tos(struct cma_device *cma_dev, unsigned int port,
311 if (!rdma_is_port_valid(cma_dev->device, port))
314 cma_dev->default_roce_tos[port - rdma_start_port(cma_dev->device)] =
319 struct ib_device *cma_get_ib_dev(struct cma_device *cma_dev)
321 return cma_dev->device;
325 * Device removal can occur at anytime, so we need extra handling to
326 * serialize notifying the user of device removal with other callbacks.
327 * We do this by disabling removal notification while a callback is in process,
328 * and reporting it after the callback completes.
330 struct rdma_id_private {
331 struct rdma_cm_id id;
333 struct rdma_bind_list *bind_list;
334 struct hlist_node node;
335 struct list_head list; /* listen_any_list or cma_device.list */
336 struct list_head listen_list; /* per device listens */
337 struct cma_device *cma_dev;
338 struct list_head mc_list;
341 enum rdma_cm_state state;
343 struct mutex qp_mutex;
345 struct completion comp;
347 struct mutex handler_mutex;
351 struct ib_sa_query *query;
368 enum ib_gid_type gid_type;
371 struct cma_multicast {
372 struct rdma_id_private *id_priv;
374 struct ib_sa_multicast *ib;
376 struct list_head list;
378 struct sockaddr_storage addr;
385 struct work_struct work;
386 struct rdma_id_private *id;
387 enum rdma_cm_state old_state;
388 enum rdma_cm_state new_state;
389 struct rdma_cm_event event;
392 struct cma_ndev_work {
393 struct work_struct work;
394 struct rdma_id_private *id;
395 struct rdma_cm_event event;
398 struct iboe_mcast_work {
399 struct work_struct work;
400 struct rdma_id_private *id;
401 struct cma_multicast *mc;
414 u8 ip_version; /* IP version: 7:4 */
416 union cma_ip_addr src_addr;
417 union cma_ip_addr dst_addr;
420 #define CMA_VERSION 0x00
422 struct cma_req_info {
423 struct sockaddr_storage listen_addr_storage;
424 struct sockaddr_storage src_addr_storage;
425 struct ib_device *device;
427 union ib_gid local_gid;
433 static int cma_comp(struct rdma_id_private *id_priv, enum rdma_cm_state comp)
438 spin_lock_irqsave(&id_priv->lock, flags);
439 ret = (id_priv->state == comp);
440 spin_unlock_irqrestore(&id_priv->lock, flags);
444 static int cma_comp_exch(struct rdma_id_private *id_priv,
445 enum rdma_cm_state comp, enum rdma_cm_state exch)
450 spin_lock_irqsave(&id_priv->lock, flags);
451 if ((ret = (id_priv->state == comp)))
452 id_priv->state = exch;
453 spin_unlock_irqrestore(&id_priv->lock, flags);
457 static enum rdma_cm_state cma_exch(struct rdma_id_private *id_priv,
458 enum rdma_cm_state exch)
461 enum rdma_cm_state old;
463 spin_lock_irqsave(&id_priv->lock, flags);
464 old = id_priv->state;
465 id_priv->state = exch;
466 spin_unlock_irqrestore(&id_priv->lock, flags);
470 static inline u8 cma_get_ip_ver(const struct cma_hdr *hdr)
472 return hdr->ip_version >> 4;
475 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
477 hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
480 static int cma_igmp_send(struct net_device *ndev, union ib_gid *mgid, bool join)
482 struct in_device *in_dev = NULL;
486 in_dev = __in_dev_get_rtnl(ndev);
489 ip_mc_inc_group(in_dev,
490 *(__be32 *)(mgid->raw + 12));
492 ip_mc_dec_group(in_dev,
493 *(__be32 *)(mgid->raw + 12));
497 return (in_dev) ? 0 : -ENODEV;
500 static void _cma_attach_to_dev(struct rdma_id_private *id_priv,
501 struct cma_device *cma_dev)
503 cma_ref_dev(cma_dev);
504 id_priv->cma_dev = cma_dev;
505 id_priv->gid_type = 0;
506 id_priv->id.device = cma_dev->device;
507 id_priv->id.route.addr.dev_addr.transport =
508 rdma_node_get_transport(cma_dev->device->node_type);
509 list_add_tail(&id_priv->list, &cma_dev->id_list);
512 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
513 struct cma_device *cma_dev)
515 _cma_attach_to_dev(id_priv, cma_dev);
517 cma_dev->default_gid_type[id_priv->id.port_num -
518 rdma_start_port(cma_dev->device)];
521 void cma_deref_dev(struct cma_device *cma_dev)
523 if (atomic_dec_and_test(&cma_dev->refcount))
524 complete(&cma_dev->comp);
527 static inline void release_mc(struct kref *kref)
529 struct cma_multicast *mc = container_of(kref, struct cma_multicast, mcref);
531 kfree(mc->multicast.ib);
535 static void cma_release_dev(struct rdma_id_private *id_priv)
538 list_del(&id_priv->list);
539 cma_deref_dev(id_priv->cma_dev);
540 id_priv->cma_dev = NULL;
544 static inline struct sockaddr *cma_src_addr(struct rdma_id_private *id_priv)
546 return (struct sockaddr *) &id_priv->id.route.addr.src_addr;
549 static inline struct sockaddr *cma_dst_addr(struct rdma_id_private *id_priv)
551 return (struct sockaddr *) &id_priv->id.route.addr.dst_addr;
554 static inline unsigned short cma_family(struct rdma_id_private *id_priv)
556 return id_priv->id.route.addr.src_addr.ss_family;
559 static int cma_set_qkey(struct rdma_id_private *id_priv, u32 qkey)
561 struct ib_sa_mcmember_rec rec;
565 if (qkey && id_priv->qkey != qkey)
571 id_priv->qkey = qkey;
575 switch (id_priv->id.ps) {
578 id_priv->qkey = RDMA_UDP_QKEY;
581 ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid);
582 ret = ib_sa_get_mcmember_rec(id_priv->id.device,
583 id_priv->id.port_num, &rec.mgid,
586 id_priv->qkey = be32_to_cpu(rec.qkey);
594 static void cma_translate_ib(struct sockaddr_ib *sib, struct rdma_dev_addr *dev_addr)
596 dev_addr->dev_type = ARPHRD_INFINIBAND;
597 rdma_addr_set_sgid(dev_addr, (union ib_gid *) &sib->sib_addr);
598 ib_addr_set_pkey(dev_addr, ntohs(sib->sib_pkey));
601 static int cma_translate_addr(struct sockaddr *addr, struct rdma_dev_addr *dev_addr)
605 if (addr->sa_family != AF_IB) {
606 ret = rdma_translate_ip(addr, dev_addr, NULL);
608 cma_translate_ib((struct sockaddr_ib *) addr, dev_addr);
615 static inline int cma_validate_port(struct ib_device *device, u8 port,
616 enum ib_gid_type gid_type,
617 union ib_gid *gid, int dev_type,
621 struct net_device *ndev = NULL;
623 if ((dev_type == ARPHRD_INFINIBAND) && !rdma_protocol_ib(device, port))
626 if ((dev_type != ARPHRD_INFINIBAND) && rdma_protocol_ib(device, port))
629 if (dev_type == ARPHRD_ETHER && rdma_protocol_roce(device, port)) {
630 ndev = dev_get_by_index(&init_net, bound_if_index);
634 gid_type = IB_GID_TYPE_IB;
637 ret = ib_find_cached_gid_by_port(device, gid, gid_type, port,
646 static int cma_acquire_dev(struct rdma_id_private *id_priv,
647 struct rdma_id_private *listen_id_priv)
649 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
650 struct cma_device *cma_dev;
651 union ib_gid gid, iboe_gid, *gidp;
655 if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
656 id_priv->id.ps == RDMA_PS_IPOIB)
660 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
663 memcpy(&gid, dev_addr->src_dev_addr +
664 rdma_addr_gid_offset(dev_addr), sizeof gid);
666 if (listen_id_priv) {
667 cma_dev = listen_id_priv->cma_dev;
668 port = listen_id_priv->id.port_num;
669 gidp = rdma_protocol_roce(cma_dev->device, port) ?
672 ret = cma_validate_port(cma_dev->device, port,
673 rdma_protocol_ib(cma_dev->device, port) ?
675 listen_id_priv->gid_type, gidp,
677 dev_addr->bound_dev_if);
679 id_priv->id.port_num = port;
684 list_for_each_entry(cma_dev, &dev_list, list) {
685 for (port = 1; port <= cma_dev->device->phys_port_cnt; ++port) {
686 if (listen_id_priv &&
687 listen_id_priv->cma_dev == cma_dev &&
688 listen_id_priv->id.port_num == port)
691 gidp = rdma_protocol_roce(cma_dev->device, port) ?
694 ret = cma_validate_port(cma_dev->device, port,
695 rdma_protocol_ib(cma_dev->device, port) ?
697 cma_dev->default_gid_type[port - 1],
698 gidp, dev_addr->dev_type,
699 dev_addr->bound_dev_if);
701 id_priv->id.port_num = port;
709 cma_attach_to_dev(id_priv, cma_dev);
716 * Select the source IB device and address to reach the destination IB address.
718 static int cma_resolve_ib_dev(struct rdma_id_private *id_priv)
720 struct cma_device *cma_dev, *cur_dev;
721 struct sockaddr_ib *addr;
722 union ib_gid gid, sgid, *dgid;
725 enum ib_port_state port_state;
729 addr = (struct sockaddr_ib *) cma_dst_addr(id_priv);
730 dgid = (union ib_gid *) &addr->sib_addr;
731 pkey = ntohs(addr->sib_pkey);
734 list_for_each_entry(cur_dev, &dev_list, list) {
735 for (p = 1; p <= cur_dev->device->phys_port_cnt; ++p) {
736 if (!rdma_cap_af_ib(cur_dev->device, p))
739 if (ib_find_cached_pkey(cur_dev->device, p, pkey, &index))
742 if (ib_get_cached_port_state(cur_dev->device, p, &port_state))
744 for (i = 0; !ib_get_cached_gid(cur_dev->device, p, i,
747 if (!memcmp(&gid, dgid, sizeof(gid))) {
750 id_priv->id.port_num = p;
754 if (!cma_dev && (gid.global.subnet_prefix ==
755 dgid->global.subnet_prefix) &&
756 port_state == IB_PORT_ACTIVE) {
759 id_priv->id.port_num = p;
769 cma_attach_to_dev(id_priv, cma_dev);
771 addr = (struct sockaddr_ib *)cma_src_addr(id_priv);
772 memcpy(&addr->sib_addr, &sgid, sizeof(sgid));
773 cma_translate_ib(addr, &id_priv->id.route.addr.dev_addr);
777 static void cma_deref_id(struct rdma_id_private *id_priv)
779 if (atomic_dec_and_test(&id_priv->refcount))
780 complete(&id_priv->comp);
783 struct rdma_cm_id *rdma_create_id(struct net *net,
784 rdma_cm_event_handler event_handler,
785 void *context, enum rdma_port_space ps,
786 enum ib_qp_type qp_type)
788 struct rdma_id_private *id_priv;
790 id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
792 return ERR_PTR(-ENOMEM);
794 id_priv->owner = task_pid_nr(current);
795 id_priv->state = RDMA_CM_IDLE;
796 id_priv->id.context = context;
797 id_priv->id.event_handler = event_handler;
799 id_priv->id.qp_type = qp_type;
800 id_priv->tos_set = false;
801 spin_lock_init(&id_priv->lock);
802 mutex_init(&id_priv->qp_mutex);
803 init_completion(&id_priv->comp);
804 atomic_set(&id_priv->refcount, 1);
805 mutex_init(&id_priv->handler_mutex);
806 INIT_LIST_HEAD(&id_priv->listen_list);
807 INIT_LIST_HEAD(&id_priv->mc_list);
808 get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
809 id_priv->id.route.addr.dev_addr.net = get_net(net);
810 id_priv->seq_num &= 0x00ffffff;
814 EXPORT_SYMBOL(rdma_create_id);
816 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
818 struct ib_qp_attr qp_attr;
819 int qp_attr_mask, ret;
821 qp_attr.qp_state = IB_QPS_INIT;
822 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
826 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
830 qp_attr.qp_state = IB_QPS_RTR;
831 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
835 qp_attr.qp_state = IB_QPS_RTS;
837 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
842 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
844 struct ib_qp_attr qp_attr;
845 int qp_attr_mask, ret;
847 qp_attr.qp_state = IB_QPS_INIT;
848 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
852 return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
855 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
856 struct ib_qp_init_attr *qp_init_attr)
858 struct rdma_id_private *id_priv;
862 id_priv = container_of(id, struct rdma_id_private, id);
863 if (id->device != pd->device)
866 qp_init_attr->port_num = id->port_num;
867 qp = ib_create_qp(pd, qp_init_attr);
871 if (id->qp_type == IB_QPT_UD)
872 ret = cma_init_ud_qp(id_priv, qp);
874 ret = cma_init_conn_qp(id_priv, qp);
879 id_priv->qp_num = qp->qp_num;
880 id_priv->srq = (qp->srq != NULL);
886 EXPORT_SYMBOL(rdma_create_qp);
888 void rdma_destroy_qp(struct rdma_cm_id *id)
890 struct rdma_id_private *id_priv;
892 id_priv = container_of(id, struct rdma_id_private, id);
893 mutex_lock(&id_priv->qp_mutex);
894 ib_destroy_qp(id_priv->id.qp);
895 id_priv->id.qp = NULL;
896 mutex_unlock(&id_priv->qp_mutex);
898 EXPORT_SYMBOL(rdma_destroy_qp);
900 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
901 struct rdma_conn_param *conn_param)
903 struct ib_qp_attr qp_attr;
904 int qp_attr_mask, ret;
906 mutex_lock(&id_priv->qp_mutex);
907 if (!id_priv->id.qp) {
912 /* Need to update QP attributes from default values. */
913 qp_attr.qp_state = IB_QPS_INIT;
914 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
918 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
922 qp_attr.qp_state = IB_QPS_RTR;
923 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
927 BUG_ON(id_priv->cma_dev->device != id_priv->id.device);
930 qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
931 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
933 mutex_unlock(&id_priv->qp_mutex);
937 static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
938 struct rdma_conn_param *conn_param)
940 struct ib_qp_attr qp_attr;
941 int qp_attr_mask, ret;
943 mutex_lock(&id_priv->qp_mutex);
944 if (!id_priv->id.qp) {
949 qp_attr.qp_state = IB_QPS_RTS;
950 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
955 qp_attr.max_rd_atomic = conn_param->initiator_depth;
956 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
958 mutex_unlock(&id_priv->qp_mutex);
962 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
964 struct ib_qp_attr qp_attr;
967 mutex_lock(&id_priv->qp_mutex);
968 if (!id_priv->id.qp) {
973 qp_attr.qp_state = IB_QPS_ERR;
974 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
976 mutex_unlock(&id_priv->qp_mutex);
980 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
981 struct ib_qp_attr *qp_attr, int *qp_attr_mask)
983 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
987 if (rdma_cap_eth_ah(id_priv->id.device, id_priv->id.port_num))
990 pkey = ib_addr_get_pkey(dev_addr);
992 ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
993 pkey, &qp_attr->pkey_index);
997 qp_attr->port_num = id_priv->id.port_num;
998 *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
1000 if (id_priv->id.qp_type == IB_QPT_UD) {
1001 ret = cma_set_qkey(id_priv, 0);
1005 qp_attr->qkey = id_priv->qkey;
1006 *qp_attr_mask |= IB_QP_QKEY;
1008 qp_attr->qp_access_flags = 0;
1009 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
1014 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
1017 struct rdma_id_private *id_priv;
1020 id_priv = container_of(id, struct rdma_id_private, id);
1021 if (rdma_cap_ib_cm(id->device, id->port_num)) {
1022 if (!id_priv->cm_id.ib || (id_priv->id.qp_type == IB_QPT_UD))
1023 ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
1025 ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
1028 if (qp_attr->qp_state == IB_QPS_RTR)
1029 qp_attr->rq_psn = id_priv->seq_num;
1030 } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
1031 if (!id_priv->cm_id.iw) {
1032 qp_attr->qp_access_flags = 0;
1033 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
1035 ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
1037 qp_attr->port_num = id_priv->id.port_num;
1038 *qp_attr_mask |= IB_QP_PORT;
1044 EXPORT_SYMBOL(rdma_init_qp_attr);
1046 static inline int cma_zero_addr(struct sockaddr *addr)
1048 switch (addr->sa_family) {
1050 return ipv4_is_zeronet(((struct sockaddr_in *)addr)->sin_addr.s_addr);
1052 return ipv6_addr_any(&((struct sockaddr_in6 *) addr)->sin6_addr);
1054 return ib_addr_any(&((struct sockaddr_ib *) addr)->sib_addr);
1060 static inline int cma_loopback_addr(struct sockaddr *addr)
1062 switch (addr->sa_family) {
1064 return ipv4_is_loopback(((struct sockaddr_in *) addr)->sin_addr.s_addr);
1066 return ipv6_addr_loopback(&((struct sockaddr_in6 *) addr)->sin6_addr);
1068 return ib_addr_loopback(&((struct sockaddr_ib *) addr)->sib_addr);
1074 static inline int cma_any_addr(struct sockaddr *addr)
1076 return cma_zero_addr(addr) || cma_loopback_addr(addr);
1079 static int cma_addr_cmp(struct sockaddr *src, struct sockaddr *dst)
1081 if (src->sa_family != dst->sa_family)
1084 switch (src->sa_family) {
1086 return ((struct sockaddr_in *) src)->sin_addr.s_addr !=
1087 ((struct sockaddr_in *) dst)->sin_addr.s_addr;
1089 return ipv6_addr_cmp(&((struct sockaddr_in6 *) src)->sin6_addr,
1090 &((struct sockaddr_in6 *) dst)->sin6_addr);
1092 return ib_addr_cmp(&((struct sockaddr_ib *) src)->sib_addr,
1093 &((struct sockaddr_ib *) dst)->sib_addr);
1097 static __be16 cma_port(struct sockaddr *addr)
1099 struct sockaddr_ib *sib;
1101 switch (addr->sa_family) {
1103 return ((struct sockaddr_in *) addr)->sin_port;
1105 return ((struct sockaddr_in6 *) addr)->sin6_port;
1107 sib = (struct sockaddr_ib *) addr;
1108 return htons((u16) (be64_to_cpu(sib->sib_sid) &
1109 be64_to_cpu(sib->sib_sid_mask)));
1115 static inline int cma_any_port(struct sockaddr *addr)
1117 return !cma_port(addr);
1120 static void cma_save_ib_info(struct sockaddr *src_addr,
1121 struct sockaddr *dst_addr,
1122 struct rdma_cm_id *listen_id,
1123 struct sa_path_rec *path)
1125 struct sockaddr_ib *listen_ib, *ib;
1127 listen_ib = (struct sockaddr_ib *) &listen_id->route.addr.src_addr;
1129 ib = (struct sockaddr_ib *)src_addr;
1130 ib->sib_family = AF_IB;
1132 ib->sib_pkey = path->pkey;
1133 ib->sib_flowinfo = path->flow_label;
1134 memcpy(&ib->sib_addr, &path->sgid, 16);
1135 ib->sib_sid = path->service_id;
1136 ib->sib_scope_id = 0;
1138 ib->sib_pkey = listen_ib->sib_pkey;
1139 ib->sib_flowinfo = listen_ib->sib_flowinfo;
1140 ib->sib_addr = listen_ib->sib_addr;
1141 ib->sib_sid = listen_ib->sib_sid;
1142 ib->sib_scope_id = listen_ib->sib_scope_id;
1144 ib->sib_sid_mask = cpu_to_be64(0xffffffffffffffffULL);
1147 ib = (struct sockaddr_ib *)dst_addr;
1148 ib->sib_family = AF_IB;
1150 ib->sib_pkey = path->pkey;
1151 ib->sib_flowinfo = path->flow_label;
1152 memcpy(&ib->sib_addr, &path->dgid, 16);
1157 static void cma_save_ip4_info(struct sockaddr_in *src_addr,
1158 struct sockaddr_in *dst_addr,
1159 struct cma_hdr *hdr,
1163 *src_addr = (struct sockaddr_in) {
1164 .sin_family = AF_INET,
1165 .sin_addr.s_addr = hdr->dst_addr.ip4.addr,
1166 .sin_port = local_port,
1171 *dst_addr = (struct sockaddr_in) {
1172 .sin_family = AF_INET,
1173 .sin_addr.s_addr = hdr->src_addr.ip4.addr,
1174 .sin_port = hdr->port,
1179 static void cma_save_ip6_info(struct sockaddr_in6 *src_addr,
1180 struct sockaddr_in6 *dst_addr,
1181 struct cma_hdr *hdr,
1185 *src_addr = (struct sockaddr_in6) {
1186 .sin6_family = AF_INET6,
1187 .sin6_addr = hdr->dst_addr.ip6,
1188 .sin6_port = local_port,
1193 *dst_addr = (struct sockaddr_in6) {
1194 .sin6_family = AF_INET6,
1195 .sin6_addr = hdr->src_addr.ip6,
1196 .sin6_port = hdr->port,
1201 static u16 cma_port_from_service_id(__be64 service_id)
1203 return (u16)be64_to_cpu(service_id);
1206 static int cma_save_ip_info(struct sockaddr *src_addr,
1207 struct sockaddr *dst_addr,
1208 struct ib_cm_event *ib_event,
1211 struct cma_hdr *hdr;
1214 hdr = ib_event->private_data;
1215 if (hdr->cma_version != CMA_VERSION)
1218 port = htons(cma_port_from_service_id(service_id));
1220 switch (cma_get_ip_ver(hdr)) {
1222 cma_save_ip4_info((struct sockaddr_in *)src_addr,
1223 (struct sockaddr_in *)dst_addr, hdr, port);
1226 cma_save_ip6_info((struct sockaddr_in6 *)src_addr,
1227 (struct sockaddr_in6 *)dst_addr, hdr, port);
1230 return -EAFNOSUPPORT;
1236 static int cma_save_net_info(struct sockaddr *src_addr,
1237 struct sockaddr *dst_addr,
1238 struct rdma_cm_id *listen_id,
1239 struct ib_cm_event *ib_event,
1240 sa_family_t sa_family, __be64 service_id)
1242 if (sa_family == AF_IB) {
1243 if (ib_event->event == IB_CM_REQ_RECEIVED)
1244 cma_save_ib_info(src_addr, dst_addr, listen_id,
1245 ib_event->param.req_rcvd.primary_path);
1246 else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED)
1247 cma_save_ib_info(src_addr, dst_addr, listen_id, NULL);
1251 return cma_save_ip_info(src_addr, dst_addr, ib_event, service_id);
1254 static int cma_save_req_info(const struct ib_cm_event *ib_event,
1255 struct cma_req_info *req)
1257 const struct ib_cm_req_event_param *req_param =
1258 &ib_event->param.req_rcvd;
1259 const struct ib_cm_sidr_req_event_param *sidr_param =
1260 &ib_event->param.sidr_req_rcvd;
1262 switch (ib_event->event) {
1263 case IB_CM_REQ_RECEIVED:
1264 req->device = req_param->listen_id->device;
1265 req->port = req_param->port;
1266 memcpy(&req->local_gid, &req_param->primary_path->sgid,
1267 sizeof(req->local_gid));
1268 req->has_gid = true;
1269 req->service_id = req_param->primary_path->service_id;
1270 req->pkey = be16_to_cpu(req_param->primary_path->pkey);
1271 if (req->pkey != req_param->bth_pkey)
1272 pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and primary path P_Key (0x%x)\n"
1273 "RDMA CMA: in the future this may cause the request to be dropped\n",
1274 req_param->bth_pkey, req->pkey);
1276 case IB_CM_SIDR_REQ_RECEIVED:
1277 req->device = sidr_param->listen_id->device;
1278 req->port = sidr_param->port;
1279 req->has_gid = false;
1280 req->service_id = sidr_param->service_id;
1281 req->pkey = sidr_param->pkey;
1282 if (req->pkey != sidr_param->bth_pkey)
1283 pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and SIDR request payload P_Key (0x%x)\n"
1284 "RDMA CMA: in the future this may cause the request to be dropped\n",
1285 sidr_param->bth_pkey, req->pkey);
1294 static bool validate_ipv4_net_dev(struct net_device *net_dev,
1295 const struct sockaddr_in *dst_addr,
1296 const struct sockaddr_in *src_addr)
1298 __be32 daddr = dst_addr->sin_addr.s_addr,
1299 saddr = src_addr->sin_addr.s_addr;
1300 struct fib_result res;
1305 if (ipv4_is_multicast(saddr) || ipv4_is_lbcast(saddr) ||
1306 ipv4_is_lbcast(daddr) || ipv4_is_zeronet(saddr) ||
1307 ipv4_is_zeronet(daddr) || ipv4_is_loopback(daddr) ||
1308 ipv4_is_loopback(saddr))
1311 memset(&fl4, 0, sizeof(fl4));
1312 fl4.flowi4_iif = net_dev->ifindex;
1317 err = fib_lookup(dev_net(net_dev), &fl4, &res, 0);
1318 ret = err == 0 && FIB_RES_DEV(res) == net_dev;
1324 static bool validate_ipv6_net_dev(struct net_device *net_dev,
1325 const struct sockaddr_in6 *dst_addr,
1326 const struct sockaddr_in6 *src_addr)
1328 #if IS_ENABLED(CONFIG_IPV6)
1329 const int strict = ipv6_addr_type(&dst_addr->sin6_addr) &
1330 IPV6_ADDR_LINKLOCAL;
1331 struct rt6_info *rt = rt6_lookup(dev_net(net_dev), &dst_addr->sin6_addr,
1332 &src_addr->sin6_addr, net_dev->ifindex,
1339 ret = rt->rt6i_idev->dev == net_dev;
1348 static bool validate_net_dev(struct net_device *net_dev,
1349 const struct sockaddr *daddr,
1350 const struct sockaddr *saddr)
1352 const struct sockaddr_in *daddr4 = (const struct sockaddr_in *)daddr;
1353 const struct sockaddr_in *saddr4 = (const struct sockaddr_in *)saddr;
1354 const struct sockaddr_in6 *daddr6 = (const struct sockaddr_in6 *)daddr;
1355 const struct sockaddr_in6 *saddr6 = (const struct sockaddr_in6 *)saddr;
1357 switch (daddr->sa_family) {
1359 return saddr->sa_family == AF_INET &&
1360 validate_ipv4_net_dev(net_dev, daddr4, saddr4);
1363 return saddr->sa_family == AF_INET6 &&
1364 validate_ipv6_net_dev(net_dev, daddr6, saddr6);
1371 static struct net_device *cma_get_net_dev(struct ib_cm_event *ib_event,
1372 struct cma_req_info *req)
1374 struct sockaddr *listen_addr =
1375 (struct sockaddr *)&req->listen_addr_storage;
1376 struct sockaddr *src_addr = (struct sockaddr *)&req->src_addr_storage;
1377 struct net_device *net_dev;
1378 const union ib_gid *gid = req->has_gid ? &req->local_gid : NULL;
1381 err = cma_save_ip_info(listen_addr, src_addr, ib_event,
1384 return ERR_PTR(err);
1386 net_dev = ib_get_net_dev_by_params(req->device, req->port, req->pkey,
1389 return ERR_PTR(-ENODEV);
1394 static enum rdma_port_space rdma_ps_from_service_id(__be64 service_id)
1396 return (be64_to_cpu(service_id) >> 16) & 0xffff;
1399 static bool cma_match_private_data(struct rdma_id_private *id_priv,
1400 const struct cma_hdr *hdr)
1402 struct sockaddr *addr = cma_src_addr(id_priv);
1404 struct in6_addr ip6_addr;
1406 if (cma_any_addr(addr) && !id_priv->afonly)
1409 switch (addr->sa_family) {
1411 ip4_addr = ((struct sockaddr_in *)addr)->sin_addr.s_addr;
1412 if (cma_get_ip_ver(hdr) != 4)
1414 if (!cma_any_addr(addr) &&
1415 hdr->dst_addr.ip4.addr != ip4_addr)
1419 ip6_addr = ((struct sockaddr_in6 *)addr)->sin6_addr;
1420 if (cma_get_ip_ver(hdr) != 6)
1422 if (!cma_any_addr(addr) &&
1423 memcmp(&hdr->dst_addr.ip6, &ip6_addr, sizeof(ip6_addr)))
1435 static bool cma_protocol_roce_dev_port(struct ib_device *device, int port_num)
1437 enum rdma_link_layer ll = rdma_port_get_link_layer(device, port_num);
1438 enum rdma_transport_type transport =
1439 rdma_node_get_transport(device->node_type);
1441 return ll == IB_LINK_LAYER_ETHERNET && transport == RDMA_TRANSPORT_IB;
1444 static bool cma_protocol_roce(const struct rdma_cm_id *id)
1446 struct ib_device *device = id->device;
1447 const int port_num = id->port_num ?: rdma_start_port(device);
1449 return cma_protocol_roce_dev_port(device, port_num);
1452 static bool cma_match_net_dev(const struct rdma_cm_id *id,
1453 const struct net_device *net_dev,
1456 const struct rdma_addr *addr = &id->route.addr;
1459 /* This request is an AF_IB request or a RoCE request */
1460 return (!id->port_num || id->port_num == port_num) &&
1461 (addr->src_addr.ss_family == AF_IB ||
1462 cma_protocol_roce_dev_port(id->device, port_num));
1465 * Net namespaces must match, and if the listner is listening
1466 * on a specific netdevice than netdevice must match as well.
1468 if (net_eq(dev_net(net_dev), addr->dev_addr.net) &&
1469 (!!addr->dev_addr.bound_dev_if ==
1470 (addr->dev_addr.bound_dev_if == net_dev->ifindex)))
1476 static struct rdma_id_private *cma_find_listener(
1477 const struct rdma_bind_list *bind_list,
1478 const struct ib_cm_id *cm_id,
1479 const struct ib_cm_event *ib_event,
1480 const struct cma_req_info *req,
1481 const struct net_device *net_dev)
1483 struct rdma_id_private *id_priv, *id_priv_dev;
1485 lockdep_assert_held(&lock);
1488 return ERR_PTR(-EINVAL);
1490 hlist_for_each_entry(id_priv, &bind_list->owners, node) {
1491 if (cma_match_private_data(id_priv, ib_event->private_data)) {
1492 if (id_priv->id.device == cm_id->device &&
1493 cma_match_net_dev(&id_priv->id, net_dev, req->port))
1495 list_for_each_entry(id_priv_dev,
1496 &id_priv->listen_list,
1498 if (id_priv_dev->id.device == cm_id->device &&
1499 cma_match_net_dev(&id_priv_dev->id, net_dev, req->port))
1505 return ERR_PTR(-EINVAL);
1508 static struct rdma_id_private *cma_id_from_event(struct ib_cm_id *cm_id,
1509 struct ib_cm_event *ib_event,
1510 struct net_device **net_dev)
1512 struct cma_req_info req;
1513 struct rdma_bind_list *bind_list;
1514 struct rdma_id_private *id_priv;
1517 err = cma_save_req_info(ib_event, &req);
1519 return ERR_PTR(err);
1521 *net_dev = cma_get_net_dev(ib_event, &req);
1522 if (IS_ERR(*net_dev)) {
1523 if (PTR_ERR(*net_dev) == -EAFNOSUPPORT) {
1524 /* Assuming the protocol is AF_IB */
1526 } else if (cma_protocol_roce_dev_port(req.device, req.port)) {
1527 /* TODO find the net dev matching the request parameters
1528 * through the RoCE GID table */
1531 return ERR_CAST(*net_dev);
1537 * Net namespace might be getting deleted while route lookup,
1538 * cm_id lookup is in progress. Therefore, perform netdevice
1539 * validation, cm_id lookup under rcu lock.
1540 * RCU lock along with netdevice state check, synchronizes with
1541 * netdevice migrating to different net namespace and also avoids
1542 * case where net namespace doesn't get deleted while lookup is in
1544 * If the device state is not IFF_UP, its properties such as ifindex
1545 * and nd_net cannot be trusted to remain valid without rcu lock.
1546 * net/core/dev.c change_net_namespace() ensures to synchronize with
1547 * ongoing operations on net device after device is closed using
1548 * synchronize_net().
1553 * If netdevice is down, it is likely that it is administratively
1554 * down or it might be migrating to different namespace.
1555 * In that case avoid further processing, as the net namespace
1556 * or ifindex may change.
1558 if (((*net_dev)->flags & IFF_UP) == 0) {
1559 id_priv = ERR_PTR(-EHOSTUNREACH);
1563 if (!validate_net_dev(*net_dev,
1564 (struct sockaddr *)&req.listen_addr_storage,
1565 (struct sockaddr *)&req.src_addr_storage)) {
1566 id_priv = ERR_PTR(-EHOSTUNREACH);
1571 bind_list = cma_ps_find(*net_dev ? dev_net(*net_dev) : &init_net,
1572 rdma_ps_from_service_id(req.service_id),
1573 cma_port_from_service_id(req.service_id));
1574 id_priv = cma_find_listener(bind_list, cm_id, ib_event, &req, *net_dev);
1577 mutex_unlock(&lock);
1578 if (IS_ERR(id_priv) && *net_dev) {
1585 static inline u8 cma_user_data_offset(struct rdma_id_private *id_priv)
1587 return cma_family(id_priv) == AF_IB ? 0 : sizeof(struct cma_hdr);
1590 static void cma_cancel_route(struct rdma_id_private *id_priv)
1592 if (rdma_cap_ib_sa(id_priv->id.device, id_priv->id.port_num)) {
1594 ib_sa_cancel_query(id_priv->query_id, id_priv->query);
1598 static void cma_cancel_listens(struct rdma_id_private *id_priv)
1600 struct rdma_id_private *dev_id_priv;
1603 * Remove from listen_any_list to prevent added devices from spawning
1604 * additional listen requests.
1607 list_del(&id_priv->list);
1609 while (!list_empty(&id_priv->listen_list)) {
1610 dev_id_priv = list_entry(id_priv->listen_list.next,
1611 struct rdma_id_private, listen_list);
1612 /* sync with device removal to avoid duplicate destruction */
1613 list_del_init(&dev_id_priv->list);
1614 list_del(&dev_id_priv->listen_list);
1615 mutex_unlock(&lock);
1617 rdma_destroy_id(&dev_id_priv->id);
1620 mutex_unlock(&lock);
1623 static void cma_cancel_operation(struct rdma_id_private *id_priv,
1624 enum rdma_cm_state state)
1627 case RDMA_CM_ADDR_QUERY:
1628 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
1630 case RDMA_CM_ROUTE_QUERY:
1631 cma_cancel_route(id_priv);
1633 case RDMA_CM_LISTEN:
1634 if (cma_any_addr(cma_src_addr(id_priv)) && !id_priv->cma_dev)
1635 cma_cancel_listens(id_priv);
1642 static void cma_release_port(struct rdma_id_private *id_priv)
1644 struct rdma_bind_list *bind_list = id_priv->bind_list;
1645 struct net *net = id_priv->id.route.addr.dev_addr.net;
1651 hlist_del(&id_priv->node);
1652 if (hlist_empty(&bind_list->owners)) {
1653 cma_ps_remove(net, bind_list->ps, bind_list->port);
1656 mutex_unlock(&lock);
1659 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
1661 struct cma_multicast *mc;
1663 while (!list_empty(&id_priv->mc_list)) {
1664 mc = container_of(id_priv->mc_list.next,
1665 struct cma_multicast, list);
1666 list_del(&mc->list);
1667 if (rdma_cap_ib_mcast(id_priv->cma_dev->device,
1668 id_priv->id.port_num)) {
1669 ib_sa_free_multicast(mc->multicast.ib);
1672 if (mc->igmp_joined) {
1673 struct rdma_dev_addr *dev_addr =
1674 &id_priv->id.route.addr.dev_addr;
1675 struct net_device *ndev = NULL;
1677 if (dev_addr->bound_dev_if)
1678 ndev = dev_get_by_index(&init_net,
1679 dev_addr->bound_dev_if);
1682 &mc->multicast.ib->rec.mgid,
1687 kref_put(&mc->mcref, release_mc);
1692 void rdma_destroy_id(struct rdma_cm_id *id)
1694 struct rdma_id_private *id_priv;
1695 enum rdma_cm_state state;
1697 id_priv = container_of(id, struct rdma_id_private, id);
1698 state = cma_exch(id_priv, RDMA_CM_DESTROYING);
1699 cma_cancel_operation(id_priv, state);
1702 * Wait for any active callback to finish. New callbacks will find
1703 * the id_priv state set to destroying and abort.
1705 mutex_lock(&id_priv->handler_mutex);
1706 mutex_unlock(&id_priv->handler_mutex);
1708 if (id_priv->cma_dev) {
1709 if (rdma_cap_ib_cm(id_priv->id.device, 1)) {
1710 if (id_priv->cm_id.ib)
1711 ib_destroy_cm_id(id_priv->cm_id.ib);
1712 } else if (rdma_cap_iw_cm(id_priv->id.device, 1)) {
1713 if (id_priv->cm_id.iw)
1714 iw_destroy_cm_id(id_priv->cm_id.iw);
1716 cma_leave_mc_groups(id_priv);
1717 cma_release_dev(id_priv);
1720 cma_release_port(id_priv);
1721 cma_deref_id(id_priv);
1722 wait_for_completion(&id_priv->comp);
1724 if (id_priv->internal_id)
1725 cma_deref_id(id_priv->id.context);
1727 kfree(id_priv->id.route.path_rec);
1728 put_net(id_priv->id.route.addr.dev_addr.net);
1731 EXPORT_SYMBOL(rdma_destroy_id);
1733 static int cma_rep_recv(struct rdma_id_private *id_priv)
1737 ret = cma_modify_qp_rtr(id_priv, NULL);
1741 ret = cma_modify_qp_rts(id_priv, NULL);
1745 ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
1751 pr_debug_ratelimited("RDMA CM: CONNECT_ERROR: failed to handle reply. status %d\n", ret);
1752 cma_modify_qp_err(id_priv);
1753 ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
1758 static void cma_set_rep_event_data(struct rdma_cm_event *event,
1759 struct ib_cm_rep_event_param *rep_data,
1762 event->param.conn.private_data = private_data;
1763 event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
1764 event->param.conn.responder_resources = rep_data->responder_resources;
1765 event->param.conn.initiator_depth = rep_data->initiator_depth;
1766 event->param.conn.flow_control = rep_data->flow_control;
1767 event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
1768 event->param.conn.srq = rep_data->srq;
1769 event->param.conn.qp_num = rep_data->remote_qpn;
1772 static int cma_ib_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1774 struct rdma_id_private *id_priv = cm_id->context;
1775 struct rdma_cm_event event;
1778 mutex_lock(&id_priv->handler_mutex);
1779 if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
1780 id_priv->state != RDMA_CM_CONNECT) ||
1781 (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
1782 id_priv->state != RDMA_CM_DISCONNECT))
1785 memset(&event, 0, sizeof event);
1786 switch (ib_event->event) {
1787 case IB_CM_REQ_ERROR:
1788 case IB_CM_REP_ERROR:
1789 event.event = RDMA_CM_EVENT_UNREACHABLE;
1790 event.status = -ETIMEDOUT;
1792 case IB_CM_REP_RECEIVED:
1793 if (cma_comp(id_priv, RDMA_CM_CONNECT) &&
1794 (id_priv->id.qp_type != IB_QPT_UD))
1795 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
1796 if (id_priv->id.qp) {
1797 event.status = cma_rep_recv(id_priv);
1798 event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
1799 RDMA_CM_EVENT_ESTABLISHED;
1801 event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
1803 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
1804 ib_event->private_data);
1806 case IB_CM_RTU_RECEIVED:
1807 case IB_CM_USER_ESTABLISHED:
1808 event.event = RDMA_CM_EVENT_ESTABLISHED;
1810 case IB_CM_DREQ_ERROR:
1811 event.status = -ETIMEDOUT; /* fall through */
1812 case IB_CM_DREQ_RECEIVED:
1813 case IB_CM_DREP_RECEIVED:
1814 if (!cma_comp_exch(id_priv, RDMA_CM_CONNECT,
1815 RDMA_CM_DISCONNECT))
1817 event.event = RDMA_CM_EVENT_DISCONNECTED;
1819 case IB_CM_TIMEWAIT_EXIT:
1820 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
1822 case IB_CM_MRA_RECEIVED:
1825 case IB_CM_REJ_RECEIVED:
1826 pr_debug_ratelimited("RDMA CM: REJECTED: %s\n", rdma_reject_msg(&id_priv->id,
1827 ib_event->param.rej_rcvd.reason));
1828 cma_modify_qp_err(id_priv);
1829 event.status = ib_event->param.rej_rcvd.reason;
1830 event.event = RDMA_CM_EVENT_REJECTED;
1831 event.param.conn.private_data = ib_event->private_data;
1832 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
1835 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
1840 ret = id_priv->id.event_handler(&id_priv->id, &event);
1842 /* Destroy the CM ID by returning a non-zero value. */
1843 id_priv->cm_id.ib = NULL;
1844 cma_exch(id_priv, RDMA_CM_DESTROYING);
1845 mutex_unlock(&id_priv->handler_mutex);
1846 rdma_destroy_id(&id_priv->id);
1850 mutex_unlock(&id_priv->handler_mutex);
1854 static struct rdma_id_private *cma_new_conn_id(struct rdma_cm_id *listen_id,
1855 struct ib_cm_event *ib_event,
1856 struct net_device *net_dev)
1858 struct rdma_id_private *id_priv;
1859 struct rdma_cm_id *id;
1860 struct rdma_route *rt;
1861 const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
1862 struct sa_path_rec *path = ib_event->param.req_rcvd.primary_path;
1863 const __be64 service_id =
1864 ib_event->param.req_rcvd.primary_path->service_id;
1867 id = rdma_create_id(listen_id->route.addr.dev_addr.net,
1868 listen_id->event_handler, listen_id->context,
1869 listen_id->ps, ib_event->param.req_rcvd.qp_type);
1873 id_priv = container_of(id, struct rdma_id_private, id);
1874 if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
1875 (struct sockaddr *)&id->route.addr.dst_addr,
1876 listen_id, ib_event, ss_family, service_id))
1880 rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
1881 rt->path_rec = kmalloc(sizeof *rt->path_rec * rt->num_paths,
1886 rt->path_rec[0] = *path;
1887 if (rt->num_paths == 2)
1888 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
1891 ret = rdma_copy_addr(&rt->addr.dev_addr, net_dev, NULL);
1895 if (!cma_protocol_roce(listen_id) &&
1896 cma_any_addr(cma_src_addr(id_priv))) {
1897 rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
1898 rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
1899 ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
1900 } else if (!cma_any_addr(cma_src_addr(id_priv))) {
1901 ret = cma_translate_addr(cma_src_addr(id_priv), &rt->addr.dev_addr);
1906 rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
1908 id_priv->state = RDMA_CM_CONNECT;
1912 rdma_destroy_id(id);
1916 static struct rdma_id_private *cma_new_udp_id(struct rdma_cm_id *listen_id,
1917 struct ib_cm_event *ib_event,
1918 struct net_device *net_dev)
1920 struct rdma_id_private *id_priv;
1921 struct rdma_cm_id *id;
1922 const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
1923 struct net *net = listen_id->route.addr.dev_addr.net;
1926 id = rdma_create_id(net, listen_id->event_handler, listen_id->context,
1927 listen_id->ps, IB_QPT_UD);
1931 id_priv = container_of(id, struct rdma_id_private, id);
1932 if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
1933 (struct sockaddr *)&id->route.addr.dst_addr,
1934 listen_id, ib_event, ss_family,
1935 ib_event->param.sidr_req_rcvd.service_id))
1939 ret = rdma_copy_addr(&id->route.addr.dev_addr, net_dev, NULL);
1943 if (!cma_any_addr(cma_src_addr(id_priv))) {
1944 ret = cma_translate_addr(cma_src_addr(id_priv),
1945 &id->route.addr.dev_addr);
1951 id_priv->state = RDMA_CM_CONNECT;
1954 rdma_destroy_id(id);
1958 static void cma_set_req_event_data(struct rdma_cm_event *event,
1959 struct ib_cm_req_event_param *req_data,
1960 void *private_data, int offset)
1962 event->param.conn.private_data = private_data + offset;
1963 event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
1964 event->param.conn.responder_resources = req_data->responder_resources;
1965 event->param.conn.initiator_depth = req_data->initiator_depth;
1966 event->param.conn.flow_control = req_data->flow_control;
1967 event->param.conn.retry_count = req_data->retry_count;
1968 event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
1969 event->param.conn.srq = req_data->srq;
1970 event->param.conn.qp_num = req_data->remote_qpn;
1973 static int cma_check_req_qp_type(struct rdma_cm_id *id, struct ib_cm_event *ib_event)
1975 return (((ib_event->event == IB_CM_REQ_RECEIVED) &&
1976 (ib_event->param.req_rcvd.qp_type == id->qp_type)) ||
1977 ((ib_event->event == IB_CM_SIDR_REQ_RECEIVED) &&
1978 (id->qp_type == IB_QPT_UD)) ||
1982 static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1984 struct rdma_id_private *listen_id, *conn_id = NULL;
1985 struct rdma_cm_event event;
1986 struct net_device *net_dev;
1990 listen_id = cma_id_from_event(cm_id, ib_event, &net_dev);
1991 if (IS_ERR(listen_id))
1992 return PTR_ERR(listen_id);
1994 if (!cma_check_req_qp_type(&listen_id->id, ib_event)) {
1999 mutex_lock(&listen_id->handler_mutex);
2000 if (listen_id->state != RDMA_CM_LISTEN) {
2001 ret = -ECONNABORTED;
2005 memset(&event, 0, sizeof event);
2006 offset = cma_user_data_offset(listen_id);
2007 event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
2008 if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) {
2009 conn_id = cma_new_udp_id(&listen_id->id, ib_event, net_dev);
2010 event.param.ud.private_data = ib_event->private_data + offset;
2011 event.param.ud.private_data_len =
2012 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
2014 conn_id = cma_new_conn_id(&listen_id->id, ib_event, net_dev);
2015 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
2016 ib_event->private_data, offset);
2023 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
2024 ret = cma_acquire_dev(conn_id, listen_id);
2028 conn_id->cm_id.ib = cm_id;
2029 cm_id->context = conn_id;
2030 cm_id->cm_handler = cma_ib_handler;
2033 * Protect against the user destroying conn_id from another thread
2034 * until we're done accessing it.
2036 atomic_inc(&conn_id->refcount);
2037 ret = conn_id->id.event_handler(&conn_id->id, &event);
2041 * Acquire mutex to prevent user executing rdma_destroy_id()
2042 * while we're accessing the cm_id.
2045 if (cma_comp(conn_id, RDMA_CM_CONNECT) &&
2046 (conn_id->id.qp_type != IB_QPT_UD))
2047 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
2048 mutex_unlock(&lock);
2049 mutex_unlock(&conn_id->handler_mutex);
2050 mutex_unlock(&listen_id->handler_mutex);
2051 cma_deref_id(conn_id);
2057 cma_deref_id(conn_id);
2058 /* Destroy the CM ID by returning a non-zero value. */
2059 conn_id->cm_id.ib = NULL;
2061 cma_exch(conn_id, RDMA_CM_DESTROYING);
2062 mutex_unlock(&conn_id->handler_mutex);
2064 mutex_unlock(&listen_id->handler_mutex);
2066 rdma_destroy_id(&conn_id->id);
2075 __be64 rdma_get_service_id(struct rdma_cm_id *id, struct sockaddr *addr)
2077 if (addr->sa_family == AF_IB)
2078 return ((struct sockaddr_ib *) addr)->sib_sid;
2080 return cpu_to_be64(((u64)id->ps << 16) + be16_to_cpu(cma_port(addr)));
2082 EXPORT_SYMBOL(rdma_get_service_id);
2084 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
2086 struct rdma_id_private *id_priv = iw_id->context;
2087 struct rdma_cm_event event;
2089 struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
2090 struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
2092 mutex_lock(&id_priv->handler_mutex);
2093 if (id_priv->state != RDMA_CM_CONNECT)
2096 memset(&event, 0, sizeof event);
2097 switch (iw_event->event) {
2098 case IW_CM_EVENT_CLOSE:
2099 event.event = RDMA_CM_EVENT_DISCONNECTED;
2101 case IW_CM_EVENT_CONNECT_REPLY:
2102 memcpy(cma_src_addr(id_priv), laddr,
2103 rdma_addr_size(laddr));
2104 memcpy(cma_dst_addr(id_priv), raddr,
2105 rdma_addr_size(raddr));
2106 switch (iw_event->status) {
2108 event.event = RDMA_CM_EVENT_ESTABLISHED;
2109 event.param.conn.initiator_depth = iw_event->ird;
2110 event.param.conn.responder_resources = iw_event->ord;
2114 event.event = RDMA_CM_EVENT_REJECTED;
2117 event.event = RDMA_CM_EVENT_UNREACHABLE;
2120 event.event = RDMA_CM_EVENT_CONNECT_ERROR;
2124 case IW_CM_EVENT_ESTABLISHED:
2125 event.event = RDMA_CM_EVENT_ESTABLISHED;
2126 event.param.conn.initiator_depth = iw_event->ird;
2127 event.param.conn.responder_resources = iw_event->ord;
2133 event.status = iw_event->status;
2134 event.param.conn.private_data = iw_event->private_data;
2135 event.param.conn.private_data_len = iw_event->private_data_len;
2136 ret = id_priv->id.event_handler(&id_priv->id, &event);
2138 /* Destroy the CM ID by returning a non-zero value. */
2139 id_priv->cm_id.iw = NULL;
2140 cma_exch(id_priv, RDMA_CM_DESTROYING);
2141 mutex_unlock(&id_priv->handler_mutex);
2142 rdma_destroy_id(&id_priv->id);
2147 mutex_unlock(&id_priv->handler_mutex);
2151 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
2152 struct iw_cm_event *iw_event)
2154 struct rdma_cm_id *new_cm_id;
2155 struct rdma_id_private *listen_id, *conn_id;
2156 struct rdma_cm_event event;
2157 int ret = -ECONNABORTED;
2158 struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
2159 struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
2161 listen_id = cm_id->context;
2163 mutex_lock(&listen_id->handler_mutex);
2164 if (listen_id->state != RDMA_CM_LISTEN)
2167 /* Create a new RDMA id for the new IW CM ID */
2168 new_cm_id = rdma_create_id(listen_id->id.route.addr.dev_addr.net,
2169 listen_id->id.event_handler,
2170 listen_id->id.context,
2171 RDMA_PS_TCP, IB_QPT_RC);
2172 if (IS_ERR(new_cm_id)) {
2176 conn_id = container_of(new_cm_id, struct rdma_id_private, id);
2177 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
2178 conn_id->state = RDMA_CM_CONNECT;
2180 ret = rdma_translate_ip(laddr, &conn_id->id.route.addr.dev_addr, NULL);
2182 mutex_unlock(&conn_id->handler_mutex);
2183 rdma_destroy_id(new_cm_id);
2187 ret = cma_acquire_dev(conn_id, listen_id);
2189 mutex_unlock(&conn_id->handler_mutex);
2190 rdma_destroy_id(new_cm_id);
2194 conn_id->cm_id.iw = cm_id;
2195 cm_id->context = conn_id;
2196 cm_id->cm_handler = cma_iw_handler;
2198 memcpy(cma_src_addr(conn_id), laddr, rdma_addr_size(laddr));
2199 memcpy(cma_dst_addr(conn_id), raddr, rdma_addr_size(raddr));
2201 memset(&event, 0, sizeof event);
2202 event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
2203 event.param.conn.private_data = iw_event->private_data;
2204 event.param.conn.private_data_len = iw_event->private_data_len;
2205 event.param.conn.initiator_depth = iw_event->ird;
2206 event.param.conn.responder_resources = iw_event->ord;
2209 * Protect against the user destroying conn_id from another thread
2210 * until we're done accessing it.
2212 atomic_inc(&conn_id->refcount);
2213 ret = conn_id->id.event_handler(&conn_id->id, &event);
2215 /* User wants to destroy the CM ID */
2216 conn_id->cm_id.iw = NULL;
2217 cma_exch(conn_id, RDMA_CM_DESTROYING);
2218 mutex_unlock(&conn_id->handler_mutex);
2219 mutex_unlock(&listen_id->handler_mutex);
2220 cma_deref_id(conn_id);
2221 rdma_destroy_id(&conn_id->id);
2225 mutex_unlock(&conn_id->handler_mutex);
2226 cma_deref_id(conn_id);
2229 mutex_unlock(&listen_id->handler_mutex);
2233 static int cma_ib_listen(struct rdma_id_private *id_priv)
2235 struct sockaddr *addr;
2236 struct ib_cm_id *id;
2239 addr = cma_src_addr(id_priv);
2240 svc_id = rdma_get_service_id(&id_priv->id, addr);
2241 id = ib_cm_insert_listen(id_priv->id.device, cma_req_handler, svc_id);
2244 id_priv->cm_id.ib = id;
2249 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
2252 struct iw_cm_id *id;
2254 id = iw_create_cm_id(id_priv->id.device,
2255 iw_conn_req_handler,
2260 id->tos = id_priv->tos;
2261 id_priv->cm_id.iw = id;
2263 memcpy(&id_priv->cm_id.iw->local_addr, cma_src_addr(id_priv),
2264 rdma_addr_size(cma_src_addr(id_priv)));
2266 ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
2269 iw_destroy_cm_id(id_priv->cm_id.iw);
2270 id_priv->cm_id.iw = NULL;
2276 static int cma_listen_handler(struct rdma_cm_id *id,
2277 struct rdma_cm_event *event)
2279 struct rdma_id_private *id_priv = id->context;
2281 id->context = id_priv->id.context;
2282 id->event_handler = id_priv->id.event_handler;
2283 return id_priv->id.event_handler(id, event);
2286 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
2287 struct cma_device *cma_dev)
2289 struct rdma_id_private *dev_id_priv;
2290 struct rdma_cm_id *id;
2291 struct net *net = id_priv->id.route.addr.dev_addr.net;
2294 lockdep_assert_held(&lock);
2296 if (cma_family(id_priv) == AF_IB && !rdma_cap_ib_cm(cma_dev->device, 1))
2299 id = rdma_create_id(net, cma_listen_handler, id_priv, id_priv->id.ps,
2300 id_priv->id.qp_type);
2304 dev_id_priv = container_of(id, struct rdma_id_private, id);
2306 dev_id_priv->state = RDMA_CM_ADDR_BOUND;
2307 memcpy(cma_src_addr(dev_id_priv), cma_src_addr(id_priv),
2308 rdma_addr_size(cma_src_addr(id_priv)));
2310 _cma_attach_to_dev(dev_id_priv, cma_dev);
2311 list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
2312 atomic_inc(&id_priv->refcount);
2313 dev_id_priv->internal_id = 1;
2314 dev_id_priv->afonly = id_priv->afonly;
2316 ret = rdma_listen(id, id_priv->backlog);
2318 pr_warn("RDMA CMA: cma_listen_on_dev, error %d, listening on device %s\n",
2319 ret, cma_dev->device->name);
2322 static void cma_listen_on_all(struct rdma_id_private *id_priv)
2324 struct cma_device *cma_dev;
2327 list_add_tail(&id_priv->list, &listen_any_list);
2328 list_for_each_entry(cma_dev, &dev_list, list)
2329 cma_listen_on_dev(id_priv, cma_dev);
2330 mutex_unlock(&lock);
2333 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
2335 struct rdma_id_private *id_priv;
2337 id_priv = container_of(id, struct rdma_id_private, id);
2338 id_priv->tos = (u8) tos;
2339 id_priv->tos_set = true;
2341 EXPORT_SYMBOL(rdma_set_service_type);
2343 static void cma_query_handler(int status, struct sa_path_rec *path_rec,
2346 struct cma_work *work = context;
2347 struct rdma_route *route;
2349 route = &work->id->id.route;
2352 route->num_paths = 1;
2353 *route->path_rec = *path_rec;
2355 work->old_state = RDMA_CM_ROUTE_QUERY;
2356 work->new_state = RDMA_CM_ADDR_RESOLVED;
2357 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
2358 work->event.status = status;
2359 pr_debug_ratelimited("RDMA CM: ROUTE_ERROR: failed to query path. status %d\n",
2363 queue_work(cma_wq, &work->work);
2366 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms,
2367 struct cma_work *work)
2369 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2370 struct sa_path_rec path_rec;
2371 ib_sa_comp_mask comp_mask;
2372 struct sockaddr_in6 *sin6;
2373 struct sockaddr_ib *sib;
2375 memset(&path_rec, 0, sizeof path_rec);
2377 if (rdma_cap_opa_ah(id_priv->id.device, id_priv->id.port_num))
2378 path_rec.rec_type = SA_PATH_REC_TYPE_OPA;
2380 path_rec.rec_type = SA_PATH_REC_TYPE_IB;
2381 rdma_addr_get_sgid(dev_addr, &path_rec.sgid);
2382 rdma_addr_get_dgid(dev_addr, &path_rec.dgid);
2383 path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
2384 path_rec.numb_path = 1;
2385 path_rec.reversible = 1;
2386 path_rec.service_id = rdma_get_service_id(&id_priv->id,
2387 cma_dst_addr(id_priv));
2389 comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
2390 IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
2391 IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
2393 switch (cma_family(id_priv)) {
2395 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
2396 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
2399 sin6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
2400 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
2401 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2404 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
2405 path_rec.traffic_class = (u8) (be32_to_cpu(sib->sib_flowinfo) >> 20);
2406 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2410 id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
2411 id_priv->id.port_num, &path_rec,
2412 comp_mask, timeout_ms,
2413 GFP_KERNEL, cma_query_handler,
2414 work, &id_priv->query);
2416 return (id_priv->query_id < 0) ? id_priv->query_id : 0;
2419 static void cma_work_handler(struct work_struct *_work)
2421 struct cma_work *work = container_of(_work, struct cma_work, work);
2422 struct rdma_id_private *id_priv = work->id;
2425 mutex_lock(&id_priv->handler_mutex);
2426 if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
2429 if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
2430 cma_exch(id_priv, RDMA_CM_DESTROYING);
2434 mutex_unlock(&id_priv->handler_mutex);
2435 cma_deref_id(id_priv);
2437 rdma_destroy_id(&id_priv->id);
2441 static void cma_ndev_work_handler(struct work_struct *_work)
2443 struct cma_ndev_work *work = container_of(_work, struct cma_ndev_work, work);
2444 struct rdma_id_private *id_priv = work->id;
2447 mutex_lock(&id_priv->handler_mutex);
2448 if (id_priv->state == RDMA_CM_DESTROYING ||
2449 id_priv->state == RDMA_CM_DEVICE_REMOVAL)
2452 if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
2453 cma_exch(id_priv, RDMA_CM_DESTROYING);
2458 mutex_unlock(&id_priv->handler_mutex);
2459 cma_deref_id(id_priv);
2461 rdma_destroy_id(&id_priv->id);
2465 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms)
2467 struct rdma_route *route = &id_priv->id.route;
2468 struct cma_work *work;
2471 work = kzalloc(sizeof *work, GFP_KERNEL);
2476 INIT_WORK(&work->work, cma_work_handler);
2477 work->old_state = RDMA_CM_ROUTE_QUERY;
2478 work->new_state = RDMA_CM_ROUTE_RESOLVED;
2479 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
2481 if (!route->path_rec)
2482 route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
2483 if (!route->path_rec) {
2488 ret = cma_query_ib_route(id_priv, timeout_ms, work);
2494 kfree(route->path_rec);
2495 route->path_rec = NULL;
2501 int rdma_set_ib_paths(struct rdma_cm_id *id,
2502 struct sa_path_rec *path_rec, int num_paths)
2504 struct rdma_id_private *id_priv;
2507 id_priv = container_of(id, struct rdma_id_private, id);
2508 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
2509 RDMA_CM_ROUTE_RESOLVED))
2512 id->route.path_rec = kmemdup(path_rec, sizeof *path_rec * num_paths,
2514 if (!id->route.path_rec) {
2519 id->route.num_paths = num_paths;
2522 cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED);
2525 EXPORT_SYMBOL(rdma_set_ib_paths);
2527 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms)
2529 struct cma_work *work;
2531 work = kzalloc(sizeof *work, GFP_KERNEL);
2536 INIT_WORK(&work->work, cma_work_handler);
2537 work->old_state = RDMA_CM_ROUTE_QUERY;
2538 work->new_state = RDMA_CM_ROUTE_RESOLVED;
2539 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
2540 queue_work(cma_wq, &work->work);
2544 static int iboe_tos_to_sl(struct net_device *ndev, int tos)
2547 struct net_device *dev;
2549 prio = rt_tos2priority(tos);
2550 dev = is_vlan_dev(ndev) ? vlan_dev_real_dev(ndev) : ndev;
2552 return netdev_get_prio_tc_map(dev, prio);
2554 #if IS_ENABLED(CONFIG_VLAN_8021Q)
2555 if (is_vlan_dev(ndev))
2556 return (vlan_dev_get_egress_qos_mask(ndev, prio) &
2557 VLAN_PRIO_MASK) >> VLAN_PRIO_SHIFT;
2562 static enum ib_gid_type cma_route_gid_type(enum rdma_network_type network_type,
2563 unsigned long supported_gids,
2564 enum ib_gid_type default_gid)
2566 if ((network_type == RDMA_NETWORK_IPV4 ||
2567 network_type == RDMA_NETWORK_IPV6) &&
2568 test_bit(IB_GID_TYPE_ROCE_UDP_ENCAP, &supported_gids))
2569 return IB_GID_TYPE_ROCE_UDP_ENCAP;
2574 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
2576 struct rdma_route *route = &id_priv->id.route;
2577 struct rdma_addr *addr = &route->addr;
2578 struct cma_work *work;
2580 struct net_device *ndev = NULL;
2581 enum ib_gid_type gid_type = IB_GID_TYPE_IB;
2582 u8 default_roce_tos = id_priv->cma_dev->default_roce_tos[id_priv->id.port_num -
2583 rdma_start_port(id_priv->cma_dev->device)];
2584 u8 tos = id_priv->tos_set ? id_priv->tos : default_roce_tos;
2587 work = kzalloc(sizeof *work, GFP_KERNEL);
2592 INIT_WORK(&work->work, cma_work_handler);
2594 route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
2595 if (!route->path_rec) {
2600 route->num_paths = 1;
2602 if (addr->dev_addr.bound_dev_if) {
2603 unsigned long supported_gids;
2605 ndev = dev_get_by_index(&init_net, addr->dev_addr.bound_dev_if);
2611 supported_gids = roce_gid_type_mask_support(id_priv->id.device,
2612 id_priv->id.port_num);
2613 gid_type = cma_route_gid_type(addr->dev_addr.network,
2616 route->path_rec->rec_type =
2617 sa_conv_gid_to_pathrec_type(gid_type);
2618 sa_path_set_ndev(route->path_rec, &init_net);
2619 sa_path_set_ifindex(route->path_rec, ndev->ifindex);
2626 sa_path_set_dmac(route->path_rec, addr->dev_addr.dst_dev_addr);
2628 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
2629 &route->path_rec->sgid);
2630 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.dst_addr,
2631 &route->path_rec->dgid);
2633 /* Use the hint from IP Stack to select GID Type */
2634 if (gid_type < ib_network_to_gid_type(addr->dev_addr.network))
2635 gid_type = ib_network_to_gid_type(addr->dev_addr.network);
2636 route->path_rec->rec_type = sa_conv_gid_to_pathrec_type(gid_type);
2638 if (((struct sockaddr *)&id_priv->id.route.addr.dst_addr)->sa_family != AF_IB)
2639 /* TODO: get the hoplimit from the inet/inet6 device */
2640 route->path_rec->hop_limit = addr->dev_addr.hoplimit;
2642 route->path_rec->hop_limit = 1;
2643 route->path_rec->reversible = 1;
2644 route->path_rec->pkey = cpu_to_be16(0xffff);
2645 route->path_rec->mtu_selector = IB_SA_EQ;
2646 route->path_rec->sl = iboe_tos_to_sl(ndev, tos);
2647 route->path_rec->traffic_class = tos;
2648 route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
2649 route->path_rec->rate_selector = IB_SA_EQ;
2650 route->path_rec->rate = iboe_get_rate(ndev);
2652 route->path_rec->packet_life_time_selector = IB_SA_EQ;
2653 route->path_rec->packet_life_time = CMA_IBOE_PACKET_LIFETIME;
2654 if (!route->path_rec->mtu) {
2659 work->old_state = RDMA_CM_ROUTE_QUERY;
2660 work->new_state = RDMA_CM_ROUTE_RESOLVED;
2661 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
2662 work->event.status = 0;
2664 queue_work(cma_wq, &work->work);
2669 kfree(route->path_rec);
2670 route->path_rec = NULL;
2671 route->num_paths = 0;
2677 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
2679 struct rdma_id_private *id_priv;
2682 id_priv = container_of(id, struct rdma_id_private, id);
2683 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY))
2686 atomic_inc(&id_priv->refcount);
2687 if (rdma_cap_ib_sa(id->device, id->port_num))
2688 ret = cma_resolve_ib_route(id_priv, timeout_ms);
2689 else if (rdma_protocol_roce(id->device, id->port_num))
2690 ret = cma_resolve_iboe_route(id_priv);
2691 else if (rdma_protocol_iwarp(id->device, id->port_num))
2692 ret = cma_resolve_iw_route(id_priv, timeout_ms);
2701 cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED);
2702 cma_deref_id(id_priv);
2705 EXPORT_SYMBOL(rdma_resolve_route);
2707 static void cma_set_loopback(struct sockaddr *addr)
2709 switch (addr->sa_family) {
2711 ((struct sockaddr_in *) addr)->sin_addr.s_addr = htonl(INADDR_LOOPBACK);
2714 ipv6_addr_set(&((struct sockaddr_in6 *) addr)->sin6_addr,
2718 ib_addr_set(&((struct sockaddr_ib *) addr)->sib_addr,
2724 static int cma_bind_loopback(struct rdma_id_private *id_priv)
2726 struct cma_device *cma_dev, *cur_dev;
2728 enum ib_port_state port_state;
2735 list_for_each_entry(cur_dev, &dev_list, list) {
2736 if (cma_family(id_priv) == AF_IB &&
2737 !rdma_cap_ib_cm(cur_dev->device, 1))
2743 for (p = 1; p <= cur_dev->device->phys_port_cnt; ++p) {
2744 if (!ib_get_cached_port_state(cur_dev->device, p, &port_state) &&
2745 port_state == IB_PORT_ACTIVE) {
2760 ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid, NULL);
2764 ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
2768 id_priv->id.route.addr.dev_addr.dev_type =
2769 (rdma_protocol_ib(cma_dev->device, p)) ?
2770 ARPHRD_INFINIBAND : ARPHRD_ETHER;
2772 rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2773 ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
2774 id_priv->id.port_num = p;
2775 cma_attach_to_dev(id_priv, cma_dev);
2776 cma_set_loopback(cma_src_addr(id_priv));
2778 mutex_unlock(&lock);
2782 static void addr_handler(int status, struct sockaddr *src_addr,
2783 struct rdma_dev_addr *dev_addr, void *context)
2785 struct rdma_id_private *id_priv = context;
2786 struct rdma_cm_event event;
2788 memset(&event, 0, sizeof event);
2789 mutex_lock(&id_priv->handler_mutex);
2790 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY,
2791 RDMA_CM_ADDR_RESOLVED))
2794 memcpy(cma_src_addr(id_priv), src_addr, rdma_addr_size(src_addr));
2795 if (!status && !id_priv->cma_dev) {
2796 status = cma_acquire_dev(id_priv, NULL);
2798 pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to acquire device. status %d\n",
2800 } else if (status) {
2801 pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to resolve IP. status %d\n", status);
2805 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
2806 RDMA_CM_ADDR_BOUND))
2808 event.event = RDMA_CM_EVENT_ADDR_ERROR;
2809 event.status = status;
2811 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2813 if (id_priv->id.event_handler(&id_priv->id, &event)) {
2814 cma_exch(id_priv, RDMA_CM_DESTROYING);
2815 mutex_unlock(&id_priv->handler_mutex);
2816 cma_deref_id(id_priv);
2817 rdma_destroy_id(&id_priv->id);
2821 mutex_unlock(&id_priv->handler_mutex);
2822 cma_deref_id(id_priv);
2825 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
2827 struct cma_work *work;
2831 work = kzalloc(sizeof *work, GFP_KERNEL);
2835 if (!id_priv->cma_dev) {
2836 ret = cma_bind_loopback(id_priv);
2841 rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2842 rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
2845 INIT_WORK(&work->work, cma_work_handler);
2846 work->old_state = RDMA_CM_ADDR_QUERY;
2847 work->new_state = RDMA_CM_ADDR_RESOLVED;
2848 work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2849 queue_work(cma_wq, &work->work);
2856 static int cma_resolve_ib_addr(struct rdma_id_private *id_priv)
2858 struct cma_work *work;
2861 work = kzalloc(sizeof *work, GFP_KERNEL);
2865 if (!id_priv->cma_dev) {
2866 ret = cma_resolve_ib_dev(id_priv);
2871 rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, (union ib_gid *)
2872 &(((struct sockaddr_ib *) &id_priv->id.route.addr.dst_addr)->sib_addr));
2875 INIT_WORK(&work->work, cma_work_handler);
2876 work->old_state = RDMA_CM_ADDR_QUERY;
2877 work->new_state = RDMA_CM_ADDR_RESOLVED;
2878 work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2879 queue_work(cma_wq, &work->work);
2886 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2887 struct sockaddr *dst_addr)
2889 if (!src_addr || !src_addr->sa_family) {
2890 src_addr = (struct sockaddr *) &id->route.addr.src_addr;
2891 src_addr->sa_family = dst_addr->sa_family;
2892 if (IS_ENABLED(CONFIG_IPV6) &&
2893 dst_addr->sa_family == AF_INET6) {
2894 struct sockaddr_in6 *src_addr6 = (struct sockaddr_in6 *) src_addr;
2895 struct sockaddr_in6 *dst_addr6 = (struct sockaddr_in6 *) dst_addr;
2896 src_addr6->sin6_scope_id = dst_addr6->sin6_scope_id;
2897 if (ipv6_addr_type(&dst_addr6->sin6_addr) & IPV6_ADDR_LINKLOCAL)
2898 id->route.addr.dev_addr.bound_dev_if = dst_addr6->sin6_scope_id;
2899 } else if (dst_addr->sa_family == AF_IB) {
2900 ((struct sockaddr_ib *) src_addr)->sib_pkey =
2901 ((struct sockaddr_ib *) dst_addr)->sib_pkey;
2904 return rdma_bind_addr(id, src_addr);
2907 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2908 struct sockaddr *dst_addr, int timeout_ms)
2910 struct rdma_id_private *id_priv;
2913 id_priv = container_of(id, struct rdma_id_private, id);
2914 memcpy(cma_dst_addr(id_priv), dst_addr, rdma_addr_size(dst_addr));
2915 if (id_priv->state == RDMA_CM_IDLE) {
2916 ret = cma_bind_addr(id, src_addr, dst_addr);
2918 memset(cma_dst_addr(id_priv), 0, rdma_addr_size(dst_addr));
2923 if (cma_family(id_priv) != dst_addr->sa_family) {
2924 memset(cma_dst_addr(id_priv), 0, rdma_addr_size(dst_addr));
2928 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY)) {
2929 memset(cma_dst_addr(id_priv), 0, rdma_addr_size(dst_addr));
2933 atomic_inc(&id_priv->refcount);
2934 if (cma_any_addr(dst_addr)) {
2935 ret = cma_resolve_loopback(id_priv);
2937 if (dst_addr->sa_family == AF_IB) {
2938 ret = cma_resolve_ib_addr(id_priv);
2940 ret = rdma_resolve_ip(&addr_client, cma_src_addr(id_priv),
2941 dst_addr, &id->route.addr.dev_addr,
2942 timeout_ms, addr_handler, id_priv);
2950 cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
2951 cma_deref_id(id_priv);
2954 EXPORT_SYMBOL(rdma_resolve_addr);
2956 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse)
2958 struct rdma_id_private *id_priv;
2959 unsigned long flags;
2962 id_priv = container_of(id, struct rdma_id_private, id);
2963 spin_lock_irqsave(&id_priv->lock, flags);
2964 if (reuse || id_priv->state == RDMA_CM_IDLE) {
2965 id_priv->reuseaddr = reuse;
2970 spin_unlock_irqrestore(&id_priv->lock, flags);
2973 EXPORT_SYMBOL(rdma_set_reuseaddr);
2975 int rdma_set_afonly(struct rdma_cm_id *id, int afonly)
2977 struct rdma_id_private *id_priv;
2978 unsigned long flags;
2981 id_priv = container_of(id, struct rdma_id_private, id);
2982 spin_lock_irqsave(&id_priv->lock, flags);
2983 if (id_priv->state == RDMA_CM_IDLE || id_priv->state == RDMA_CM_ADDR_BOUND) {
2984 id_priv->options |= (1 << CMA_OPTION_AFONLY);
2985 id_priv->afonly = afonly;
2990 spin_unlock_irqrestore(&id_priv->lock, flags);
2993 EXPORT_SYMBOL(rdma_set_afonly);
2995 static void cma_bind_port(struct rdma_bind_list *bind_list,
2996 struct rdma_id_private *id_priv)
2998 struct sockaddr *addr;
2999 struct sockaddr_ib *sib;
3003 lockdep_assert_held(&lock);
3005 addr = cma_src_addr(id_priv);
3006 port = htons(bind_list->port);
3008 switch (addr->sa_family) {
3010 ((struct sockaddr_in *) addr)->sin_port = port;
3013 ((struct sockaddr_in6 *) addr)->sin6_port = port;
3016 sib = (struct sockaddr_ib *) addr;
3017 sid = be64_to_cpu(sib->sib_sid);
3018 mask = be64_to_cpu(sib->sib_sid_mask);
3019 sib->sib_sid = cpu_to_be64((sid & mask) | (u64) ntohs(port));
3020 sib->sib_sid_mask = cpu_to_be64(~0ULL);
3023 id_priv->bind_list = bind_list;
3024 hlist_add_head(&id_priv->node, &bind_list->owners);
3027 static int cma_alloc_port(enum rdma_port_space ps,
3028 struct rdma_id_private *id_priv, unsigned short snum)
3030 struct rdma_bind_list *bind_list;
3033 lockdep_assert_held(&lock);
3035 bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
3039 ret = cma_ps_alloc(id_priv->id.route.addr.dev_addr.net, ps, bind_list,
3045 bind_list->port = (unsigned short)ret;
3046 cma_bind_port(bind_list, id_priv);
3050 return ret == -ENOSPC ? -EADDRNOTAVAIL : ret;
3053 static int cma_port_is_unique(struct rdma_bind_list *bind_list,
3054 struct rdma_id_private *id_priv)
3056 struct rdma_id_private *cur_id;
3057 struct sockaddr *daddr = cma_dst_addr(id_priv);
3058 struct sockaddr *saddr = cma_src_addr(id_priv);
3059 __be16 dport = cma_port(daddr);
3061 lockdep_assert_held(&lock);
3063 hlist_for_each_entry(cur_id, &bind_list->owners, node) {
3064 struct sockaddr *cur_daddr = cma_dst_addr(cur_id);
3065 struct sockaddr *cur_saddr = cma_src_addr(cur_id);
3066 __be16 cur_dport = cma_port(cur_daddr);
3068 if (id_priv == cur_id)
3071 /* different dest port -> unique */
3072 if (!cma_any_port(daddr) &&
3073 !cma_any_port(cur_daddr) &&
3074 (dport != cur_dport))
3077 /* different src address -> unique */
3078 if (!cma_any_addr(saddr) &&
3079 !cma_any_addr(cur_saddr) &&
3080 cma_addr_cmp(saddr, cur_saddr))
3083 /* different dst address -> unique */
3084 if (!cma_any_addr(daddr) &&
3085 !cma_any_addr(cur_daddr) &&
3086 cma_addr_cmp(daddr, cur_daddr))
3089 return -EADDRNOTAVAIL;
3094 static int cma_alloc_any_port(enum rdma_port_space ps,
3095 struct rdma_id_private *id_priv)
3097 static unsigned int last_used_port;
3098 int low, high, remaining;
3100 struct net *net = id_priv->id.route.addr.dev_addr.net;
3102 lockdep_assert_held(&lock);
3104 inet_get_local_port_range(net, &low, &high);
3105 remaining = (high - low) + 1;
3106 rover = prandom_u32() % remaining + low;
3108 if (last_used_port != rover) {
3109 struct rdma_bind_list *bind_list;
3112 bind_list = cma_ps_find(net, ps, (unsigned short)rover);
3115 ret = cma_alloc_port(ps, id_priv, rover);
3117 ret = cma_port_is_unique(bind_list, id_priv);
3119 cma_bind_port(bind_list, id_priv);
3122 * Remember previously used port number in order to avoid
3123 * re-using same port immediately after it is closed.
3126 last_used_port = rover;
3127 if (ret != -EADDRNOTAVAIL)
3132 if ((rover < low) || (rover > high))
3136 return -EADDRNOTAVAIL;
3140 * Check that the requested port is available. This is called when trying to
3141 * bind to a specific port, or when trying to listen on a bound port. In
3142 * the latter case, the provided id_priv may already be on the bind_list, but
3143 * we still need to check that it's okay to start listening.
3145 static int cma_check_port(struct rdma_bind_list *bind_list,
3146 struct rdma_id_private *id_priv, uint8_t reuseaddr)
3148 struct rdma_id_private *cur_id;
3149 struct sockaddr *addr, *cur_addr;
3151 lockdep_assert_held(&lock);
3153 addr = cma_src_addr(id_priv);
3154 hlist_for_each_entry(cur_id, &bind_list->owners, node) {
3155 if (id_priv == cur_id)
3158 if ((cur_id->state != RDMA_CM_LISTEN) && reuseaddr &&
3162 cur_addr = cma_src_addr(cur_id);
3163 if (id_priv->afonly && cur_id->afonly &&
3164 (addr->sa_family != cur_addr->sa_family))
3167 if (cma_any_addr(addr) || cma_any_addr(cur_addr))
3168 return -EADDRNOTAVAIL;
3170 if (!cma_addr_cmp(addr, cur_addr))
3176 static int cma_use_port(enum rdma_port_space ps,
3177 struct rdma_id_private *id_priv)
3179 struct rdma_bind_list *bind_list;
3180 unsigned short snum;
3183 lockdep_assert_held(&lock);
3185 snum = ntohs(cma_port(cma_src_addr(id_priv)));
3186 if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
3189 bind_list = cma_ps_find(id_priv->id.route.addr.dev_addr.net, ps, snum);
3191 ret = cma_alloc_port(ps, id_priv, snum);
3193 ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr);
3195 cma_bind_port(bind_list, id_priv);
3200 static int cma_bind_listen(struct rdma_id_private *id_priv)
3202 struct rdma_bind_list *bind_list = id_priv->bind_list;
3206 if (bind_list->owners.first->next)
3207 ret = cma_check_port(bind_list, id_priv, 0);
3208 mutex_unlock(&lock);
3212 static enum rdma_port_space cma_select_inet_ps(
3213 struct rdma_id_private *id_priv)
3215 switch (id_priv->id.ps) {
3220 return id_priv->id.ps;
3227 static enum rdma_port_space cma_select_ib_ps(struct rdma_id_private *id_priv)
3229 enum rdma_port_space ps = 0;
3230 struct sockaddr_ib *sib;
3231 u64 sid_ps, mask, sid;
3233 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
3234 mask = be64_to_cpu(sib->sib_sid_mask) & RDMA_IB_IP_PS_MASK;
3235 sid = be64_to_cpu(sib->sib_sid) & mask;
3237 if ((id_priv->id.ps == RDMA_PS_IB) && (sid == (RDMA_IB_IP_PS_IB & mask))) {
3238 sid_ps = RDMA_IB_IP_PS_IB;
3240 } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_TCP)) &&
3241 (sid == (RDMA_IB_IP_PS_TCP & mask))) {
3242 sid_ps = RDMA_IB_IP_PS_TCP;
3244 } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_UDP)) &&
3245 (sid == (RDMA_IB_IP_PS_UDP & mask))) {
3246 sid_ps = RDMA_IB_IP_PS_UDP;
3251 sib->sib_sid = cpu_to_be64(sid_ps | ntohs(cma_port((struct sockaddr *) sib)));
3252 sib->sib_sid_mask = cpu_to_be64(RDMA_IB_IP_PS_MASK |
3253 be64_to_cpu(sib->sib_sid_mask));
3258 static int cma_get_port(struct rdma_id_private *id_priv)
3260 enum rdma_port_space ps;
3263 if (cma_family(id_priv) != AF_IB)
3264 ps = cma_select_inet_ps(id_priv);
3266 ps = cma_select_ib_ps(id_priv);
3268 return -EPROTONOSUPPORT;
3271 if (cma_any_port(cma_src_addr(id_priv)))
3272 ret = cma_alloc_any_port(ps, id_priv);
3274 ret = cma_use_port(ps, id_priv);
3275 mutex_unlock(&lock);
3280 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
3281 struct sockaddr *addr)
3283 #if IS_ENABLED(CONFIG_IPV6)
3284 struct sockaddr_in6 *sin6;
3286 if (addr->sa_family != AF_INET6)
3289 sin6 = (struct sockaddr_in6 *) addr;
3291 if (!(ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL))
3294 if (!sin6->sin6_scope_id)
3297 dev_addr->bound_dev_if = sin6->sin6_scope_id;
3302 int rdma_listen(struct rdma_cm_id *id, int backlog)
3304 struct rdma_id_private *id_priv;
3307 id_priv = container_of(id, struct rdma_id_private, id);
3308 if (id_priv->state == RDMA_CM_IDLE) {
3309 id->route.addr.src_addr.ss_family = AF_INET;
3310 ret = rdma_bind_addr(id, cma_src_addr(id_priv));
3315 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN))
3318 if (id_priv->reuseaddr) {
3319 ret = cma_bind_listen(id_priv);
3324 id_priv->backlog = backlog;
3326 if (rdma_cap_ib_cm(id->device, 1)) {
3327 ret = cma_ib_listen(id_priv);
3330 } else if (rdma_cap_iw_cm(id->device, 1)) {
3331 ret = cma_iw_listen(id_priv, backlog);
3339 cma_listen_on_all(id_priv);
3343 id_priv->backlog = 0;
3344 cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND);
3347 EXPORT_SYMBOL(rdma_listen);
3349 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
3351 struct rdma_id_private *id_priv;
3353 struct sockaddr *daddr;
3355 if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6 &&
3356 addr->sa_family != AF_IB)
3357 return -EAFNOSUPPORT;
3359 id_priv = container_of(id, struct rdma_id_private, id);
3360 if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND))
3363 ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
3367 memcpy(cma_src_addr(id_priv), addr, rdma_addr_size(addr));
3368 if (!cma_any_addr(addr)) {
3369 ret = cma_translate_addr(addr, &id->route.addr.dev_addr);
3373 ret = cma_acquire_dev(id_priv, NULL);
3378 if (!(id_priv->options & (1 << CMA_OPTION_AFONLY))) {
3379 if (addr->sa_family == AF_INET)
3380 id_priv->afonly = 1;
3381 #if IS_ENABLED(CONFIG_IPV6)
3382 else if (addr->sa_family == AF_INET6) {
3383 struct net *net = id_priv->id.route.addr.dev_addr.net;
3385 id_priv->afonly = net->ipv6.sysctl.bindv6only;
3389 daddr = cma_dst_addr(id_priv);
3390 daddr->sa_family = addr->sa_family;
3392 ret = cma_get_port(id_priv);
3398 if (id_priv->cma_dev)
3399 cma_release_dev(id_priv);
3401 cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE);
3404 EXPORT_SYMBOL(rdma_bind_addr);
3406 static int cma_format_hdr(void *hdr, struct rdma_id_private *id_priv)
3408 struct cma_hdr *cma_hdr;
3411 cma_hdr->cma_version = CMA_VERSION;
3412 if (cma_family(id_priv) == AF_INET) {
3413 struct sockaddr_in *src4, *dst4;
3415 src4 = (struct sockaddr_in *) cma_src_addr(id_priv);
3416 dst4 = (struct sockaddr_in *) cma_dst_addr(id_priv);
3418 cma_set_ip_ver(cma_hdr, 4);
3419 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
3420 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
3421 cma_hdr->port = src4->sin_port;
3422 } else if (cma_family(id_priv) == AF_INET6) {
3423 struct sockaddr_in6 *src6, *dst6;
3425 src6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
3426 dst6 = (struct sockaddr_in6 *) cma_dst_addr(id_priv);
3428 cma_set_ip_ver(cma_hdr, 6);
3429 cma_hdr->src_addr.ip6 = src6->sin6_addr;
3430 cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
3431 cma_hdr->port = src6->sin6_port;
3436 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
3437 struct ib_cm_event *ib_event)
3439 struct rdma_id_private *id_priv = cm_id->context;
3440 struct rdma_cm_event event;
3441 struct ib_cm_sidr_rep_event_param *rep = &ib_event->param.sidr_rep_rcvd;
3444 mutex_lock(&id_priv->handler_mutex);
3445 if (id_priv->state != RDMA_CM_CONNECT)
3448 memset(&event, 0, sizeof event);
3449 switch (ib_event->event) {
3450 case IB_CM_SIDR_REQ_ERROR:
3451 event.event = RDMA_CM_EVENT_UNREACHABLE;
3452 event.status = -ETIMEDOUT;
3454 case IB_CM_SIDR_REP_RECEIVED:
3455 event.param.ud.private_data = ib_event->private_data;
3456 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
3457 if (rep->status != IB_SIDR_SUCCESS) {
3458 event.event = RDMA_CM_EVENT_UNREACHABLE;
3459 event.status = ib_event->param.sidr_rep_rcvd.status;
3460 pr_debug_ratelimited("RDMA CM: UNREACHABLE: bad SIDR reply. status %d\n",
3464 ret = cma_set_qkey(id_priv, rep->qkey);
3466 pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to set qkey. status %d\n", ret);
3467 event.event = RDMA_CM_EVENT_ADDR_ERROR;
3471 ib_init_ah_from_path(id_priv->id.device, id_priv->id.port_num,
3472 id_priv->id.route.path_rec,
3473 &event.param.ud.ah_attr);
3474 event.param.ud.qp_num = rep->qpn;
3475 event.param.ud.qkey = rep->qkey;
3476 event.event = RDMA_CM_EVENT_ESTABLISHED;
3480 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
3485 ret = id_priv->id.event_handler(&id_priv->id, &event);
3487 /* Destroy the CM ID by returning a non-zero value. */
3488 id_priv->cm_id.ib = NULL;
3489 cma_exch(id_priv, RDMA_CM_DESTROYING);
3490 mutex_unlock(&id_priv->handler_mutex);
3491 rdma_destroy_id(&id_priv->id);
3495 mutex_unlock(&id_priv->handler_mutex);
3499 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
3500 struct rdma_conn_param *conn_param)
3502 struct ib_cm_sidr_req_param req;
3503 struct ib_cm_id *id;
3508 memset(&req, 0, sizeof req);
3509 offset = cma_user_data_offset(id_priv);
3510 req.private_data_len = offset + conn_param->private_data_len;
3511 if (req.private_data_len < conn_param->private_data_len)
3514 if (req.private_data_len) {
3515 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
3519 private_data = NULL;
3522 if (conn_param->private_data && conn_param->private_data_len)
3523 memcpy(private_data + offset, conn_param->private_data,
3524 conn_param->private_data_len);
3527 ret = cma_format_hdr(private_data, id_priv);
3530 req.private_data = private_data;
3533 id = ib_create_cm_id(id_priv->id.device, cma_sidr_rep_handler,
3539 id_priv->cm_id.ib = id;
3541 req.path = id_priv->id.route.path_rec;
3542 req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
3543 req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
3544 req.max_cm_retries = CMA_MAX_CM_RETRIES;
3546 ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
3548 ib_destroy_cm_id(id_priv->cm_id.ib);
3549 id_priv->cm_id.ib = NULL;
3552 kfree(private_data);
3556 static int cma_connect_ib(struct rdma_id_private *id_priv,
3557 struct rdma_conn_param *conn_param)
3559 struct ib_cm_req_param req;
3560 struct rdma_route *route;
3562 struct ib_cm_id *id;
3566 memset(&req, 0, sizeof req);
3567 offset = cma_user_data_offset(id_priv);
3568 req.private_data_len = offset + conn_param->private_data_len;
3569 if (req.private_data_len < conn_param->private_data_len)
3572 if (req.private_data_len) {
3573 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
3577 private_data = NULL;
3580 if (conn_param->private_data && conn_param->private_data_len)
3581 memcpy(private_data + offset, conn_param->private_data,
3582 conn_param->private_data_len);
3584 id = ib_create_cm_id(id_priv->id.device, cma_ib_handler, id_priv);
3589 id_priv->cm_id.ib = id;
3591 route = &id_priv->id.route;
3593 ret = cma_format_hdr(private_data, id_priv);
3596 req.private_data = private_data;
3599 req.primary_path = &route->path_rec[0];
3600 if (route->num_paths == 2)
3601 req.alternate_path = &route->path_rec[1];
3603 req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
3604 req.qp_num = id_priv->qp_num;
3605 req.qp_type = id_priv->id.qp_type;
3606 req.starting_psn = id_priv->seq_num;
3607 req.responder_resources = conn_param->responder_resources;
3608 req.initiator_depth = conn_param->initiator_depth;
3609 req.flow_control = conn_param->flow_control;
3610 req.retry_count = min_t(u8, 7, conn_param->retry_count);
3611 req.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
3612 req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
3613 req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
3614 req.max_cm_retries = CMA_MAX_CM_RETRIES;
3615 req.srq = id_priv->srq ? 1 : 0;
3617 ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
3619 if (ret && !IS_ERR(id)) {
3620 ib_destroy_cm_id(id);
3621 id_priv->cm_id.ib = NULL;
3624 kfree(private_data);
3628 static int cma_connect_iw(struct rdma_id_private *id_priv,
3629 struct rdma_conn_param *conn_param)
3631 struct iw_cm_id *cm_id;
3633 struct iw_cm_conn_param iw_param;
3635 cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
3637 return PTR_ERR(cm_id);
3639 cm_id->tos = id_priv->tos;
3640 id_priv->cm_id.iw = cm_id;
3642 memcpy(&cm_id->local_addr, cma_src_addr(id_priv),
3643 rdma_addr_size(cma_src_addr(id_priv)));
3644 memcpy(&cm_id->remote_addr, cma_dst_addr(id_priv),
3645 rdma_addr_size(cma_dst_addr(id_priv)));
3647 ret = cma_modify_qp_rtr(id_priv, conn_param);
3652 iw_param.ord = conn_param->initiator_depth;
3653 iw_param.ird = conn_param->responder_resources;
3654 iw_param.private_data = conn_param->private_data;
3655 iw_param.private_data_len = conn_param->private_data_len;
3656 iw_param.qpn = id_priv->id.qp ? id_priv->qp_num : conn_param->qp_num;
3658 memset(&iw_param, 0, sizeof iw_param);
3659 iw_param.qpn = id_priv->qp_num;
3661 ret = iw_cm_connect(cm_id, &iw_param);
3664 iw_destroy_cm_id(cm_id);
3665 id_priv->cm_id.iw = NULL;
3670 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
3672 struct rdma_id_private *id_priv;
3675 id_priv = container_of(id, struct rdma_id_private, id);
3676 if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT))
3680 id_priv->qp_num = conn_param->qp_num;
3681 id_priv->srq = conn_param->srq;
3684 if (rdma_cap_ib_cm(id->device, id->port_num)) {
3685 if (id->qp_type == IB_QPT_UD)
3686 ret = cma_resolve_ib_udp(id_priv, conn_param);
3688 ret = cma_connect_ib(id_priv, conn_param);
3689 } else if (rdma_cap_iw_cm(id->device, id->port_num))
3690 ret = cma_connect_iw(id_priv, conn_param);
3698 cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED);
3701 EXPORT_SYMBOL(rdma_connect);
3703 static int cma_accept_ib(struct rdma_id_private *id_priv,
3704 struct rdma_conn_param *conn_param)
3706 struct ib_cm_rep_param rep;
3709 ret = cma_modify_qp_rtr(id_priv, conn_param);
3713 ret = cma_modify_qp_rts(id_priv, conn_param);
3717 memset(&rep, 0, sizeof rep);
3718 rep.qp_num = id_priv->qp_num;
3719 rep.starting_psn = id_priv->seq_num;
3720 rep.private_data = conn_param->private_data;
3721 rep.private_data_len = conn_param->private_data_len;
3722 rep.responder_resources = conn_param->responder_resources;
3723 rep.initiator_depth = conn_param->initiator_depth;
3724 rep.failover_accepted = 0;
3725 rep.flow_control = conn_param->flow_control;
3726 rep.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
3727 rep.srq = id_priv->srq ? 1 : 0;
3729 ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
3734 static int cma_accept_iw(struct rdma_id_private *id_priv,
3735 struct rdma_conn_param *conn_param)
3737 struct iw_cm_conn_param iw_param;
3743 ret = cma_modify_qp_rtr(id_priv, conn_param);
3747 iw_param.ord = conn_param->initiator_depth;
3748 iw_param.ird = conn_param->responder_resources;
3749 iw_param.private_data = conn_param->private_data;
3750 iw_param.private_data_len = conn_param->private_data_len;
3751 if (id_priv->id.qp) {
3752 iw_param.qpn = id_priv->qp_num;
3754 iw_param.qpn = conn_param->qp_num;
3756 return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
3759 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
3760 enum ib_cm_sidr_status status, u32 qkey,
3761 const void *private_data, int private_data_len)
3763 struct ib_cm_sidr_rep_param rep;
3766 memset(&rep, 0, sizeof rep);
3767 rep.status = status;
3768 if (status == IB_SIDR_SUCCESS) {
3769 ret = cma_set_qkey(id_priv, qkey);
3772 rep.qp_num = id_priv->qp_num;
3773 rep.qkey = id_priv->qkey;
3775 rep.private_data = private_data;
3776 rep.private_data_len = private_data_len;
3778 return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
3781 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
3783 struct rdma_id_private *id_priv;
3786 id_priv = container_of(id, struct rdma_id_private, id);
3788 id_priv->owner = task_pid_nr(current);
3790 if (!cma_comp(id_priv, RDMA_CM_CONNECT))
3793 if (!id->qp && conn_param) {
3794 id_priv->qp_num = conn_param->qp_num;
3795 id_priv->srq = conn_param->srq;
3798 if (rdma_cap_ib_cm(id->device, id->port_num)) {
3799 if (id->qp_type == IB_QPT_UD) {
3801 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
3803 conn_param->private_data,
3804 conn_param->private_data_len);
3806 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
3810 ret = cma_accept_ib(id_priv, conn_param);
3812 ret = cma_rep_recv(id_priv);
3814 } else if (rdma_cap_iw_cm(id->device, id->port_num))
3815 ret = cma_accept_iw(id_priv, conn_param);
3824 cma_modify_qp_err(id_priv);
3825 rdma_reject(id, NULL, 0);
3828 EXPORT_SYMBOL(rdma_accept);
3830 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
3832 struct rdma_id_private *id_priv;
3835 id_priv = container_of(id, struct rdma_id_private, id);
3836 if (!id_priv->cm_id.ib)
3839 switch (id->device->node_type) {
3840 case RDMA_NODE_IB_CA:
3841 ret = ib_cm_notify(id_priv->cm_id.ib, event);
3849 EXPORT_SYMBOL(rdma_notify);
3851 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
3852 u8 private_data_len)
3854 struct rdma_id_private *id_priv;
3857 id_priv = container_of(id, struct rdma_id_private, id);
3858 if (!id_priv->cm_id.ib)
3861 if (rdma_cap_ib_cm(id->device, id->port_num)) {
3862 if (id->qp_type == IB_QPT_UD)
3863 ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT, 0,
3864 private_data, private_data_len);
3866 ret = ib_send_cm_rej(id_priv->cm_id.ib,
3867 IB_CM_REJ_CONSUMER_DEFINED, NULL,
3868 0, private_data, private_data_len);
3869 } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
3870 ret = iw_cm_reject(id_priv->cm_id.iw,
3871 private_data, private_data_len);
3877 EXPORT_SYMBOL(rdma_reject);
3879 int rdma_disconnect(struct rdma_cm_id *id)
3881 struct rdma_id_private *id_priv;
3884 id_priv = container_of(id, struct rdma_id_private, id);
3885 if (!id_priv->cm_id.ib)
3888 if (rdma_cap_ib_cm(id->device, id->port_num)) {
3889 ret = cma_modify_qp_err(id_priv);
3892 /* Initiate or respond to a disconnect. */
3893 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
3894 ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
3895 } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
3896 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
3903 EXPORT_SYMBOL(rdma_disconnect);
3905 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
3907 struct rdma_id_private *id_priv;
3908 struct cma_multicast *mc = multicast->context;
3909 struct rdma_cm_event event;
3912 id_priv = mc->id_priv;
3913 mutex_lock(&id_priv->handler_mutex);
3914 if (id_priv->state != RDMA_CM_ADDR_BOUND &&
3915 id_priv->state != RDMA_CM_ADDR_RESOLVED)
3919 status = cma_set_qkey(id_priv, be32_to_cpu(multicast->rec.qkey));
3921 pr_debug_ratelimited("RDMA CM: MULTICAST_ERROR: failed to join multicast. status %d\n",
3923 mutex_lock(&id_priv->qp_mutex);
3924 if (!status && id_priv->id.qp) {
3925 status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid,
3926 be16_to_cpu(multicast->rec.mlid));
3928 pr_debug_ratelimited("RDMA CM: MULTICAST_ERROR: failed to attach QP. status %d\n",
3931 mutex_unlock(&id_priv->qp_mutex);
3933 memset(&event, 0, sizeof event);
3934 event.status = status;
3935 event.param.ud.private_data = mc->context;
3937 struct rdma_dev_addr *dev_addr =
3938 &id_priv->id.route.addr.dev_addr;
3939 struct net_device *ndev =
3940 dev_get_by_index(&init_net, dev_addr->bound_dev_if);
3941 enum ib_gid_type gid_type =
3942 id_priv->cma_dev->default_gid_type[id_priv->id.port_num -
3943 rdma_start_port(id_priv->cma_dev->device)];
3945 event.event = RDMA_CM_EVENT_MULTICAST_JOIN;
3946 ib_init_ah_from_mcmember(id_priv->id.device,
3947 id_priv->id.port_num, &multicast->rec,
3949 &event.param.ud.ah_attr);
3950 event.param.ud.qp_num = 0xFFFFFF;
3951 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
3955 event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
3957 ret = id_priv->id.event_handler(&id_priv->id, &event);
3959 cma_exch(id_priv, RDMA_CM_DESTROYING);
3960 mutex_unlock(&id_priv->handler_mutex);
3961 rdma_destroy_id(&id_priv->id);
3966 mutex_unlock(&id_priv->handler_mutex);
3970 static void cma_set_mgid(struct rdma_id_private *id_priv,
3971 struct sockaddr *addr, union ib_gid *mgid)
3973 unsigned char mc_map[MAX_ADDR_LEN];
3974 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3975 struct sockaddr_in *sin = (struct sockaddr_in *) addr;
3976 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
3978 if (cma_any_addr(addr)) {
3979 memset(mgid, 0, sizeof *mgid);
3980 } else if ((addr->sa_family == AF_INET6) &&
3981 ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
3983 /* IPv6 address is an SA assigned MGID. */
3984 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3985 } else if (addr->sa_family == AF_IB) {
3986 memcpy(mgid, &((struct sockaddr_ib *) addr)->sib_addr, sizeof *mgid);
3987 } else if ((addr->sa_family == AF_INET6)) {
3988 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
3989 if (id_priv->id.ps == RDMA_PS_UDP)
3990 mc_map[7] = 0x01; /* Use RDMA CM signature */
3991 *mgid = *(union ib_gid *) (mc_map + 4);
3993 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
3994 if (id_priv->id.ps == RDMA_PS_UDP)
3995 mc_map[7] = 0x01; /* Use RDMA CM signature */
3996 *mgid = *(union ib_gid *) (mc_map + 4);
4000 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
4001 struct cma_multicast *mc)
4003 struct ib_sa_mcmember_rec rec;
4004 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4005 ib_sa_comp_mask comp_mask;
4008 ib_addr_get_mgid(dev_addr, &rec.mgid);
4009 ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
4014 ret = cma_set_qkey(id_priv, 0);
4018 cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
4019 rec.qkey = cpu_to_be32(id_priv->qkey);
4020 rdma_addr_get_sgid(dev_addr, &rec.port_gid);
4021 rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
4022 rec.join_state = mc->join_state;
4024 if ((rec.join_state == BIT(SENDONLY_FULLMEMBER_JOIN)) &&
4025 (!ib_sa_sendonly_fullmem_support(&sa_client,
4027 id_priv->id.port_num))) {
4028 pr_warn("RDMA CM: %s port %u Unable to multicast join\n"
4029 "RDMA CM: SM doesn't support Send Only Full Member option\n",
4030 id_priv->id.device->name, id_priv->id.port_num);
4034 comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
4035 IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
4036 IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
4037 IB_SA_MCMEMBER_REC_FLOW_LABEL |
4038 IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
4040 if (id_priv->id.ps == RDMA_PS_IPOIB)
4041 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
4042 IB_SA_MCMEMBER_REC_RATE_SELECTOR |
4043 IB_SA_MCMEMBER_REC_MTU_SELECTOR |
4044 IB_SA_MCMEMBER_REC_MTU |
4045 IB_SA_MCMEMBER_REC_HOP_LIMIT;
4047 mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device,
4048 id_priv->id.port_num, &rec,
4049 comp_mask, GFP_KERNEL,
4050 cma_ib_mc_handler, mc);
4051 return PTR_ERR_OR_ZERO(mc->multicast.ib);
4054 static void iboe_mcast_work_handler(struct work_struct *work)
4056 struct iboe_mcast_work *mw = container_of(work, struct iboe_mcast_work, work);
4057 struct cma_multicast *mc = mw->mc;
4058 struct ib_sa_multicast *m = mc->multicast.ib;
4060 mc->multicast.ib->context = mc;
4061 cma_ib_mc_handler(0, m);
4062 kref_put(&mc->mcref, release_mc);
4066 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid,
4067 enum ib_gid_type gid_type)
4069 struct sockaddr_in *sin = (struct sockaddr_in *)addr;
4070 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
4072 if (cma_any_addr(addr)) {
4073 memset(mgid, 0, sizeof *mgid);
4074 } else if (addr->sa_family == AF_INET6) {
4075 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
4077 mgid->raw[0] = (gid_type == IB_GID_TYPE_IB) ? 0xff : 0;
4078 mgid->raw[1] = (gid_type == IB_GID_TYPE_IB) ? 0x0e : 0;
4087 mgid->raw[10] = 0xff;
4088 mgid->raw[11] = 0xff;
4089 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
4093 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
4094 struct cma_multicast *mc)
4096 struct iboe_mcast_work *work;
4097 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4099 struct sockaddr *addr = (struct sockaddr *)&mc->addr;
4100 struct net_device *ndev = NULL;
4101 enum ib_gid_type gid_type;
4104 send_only = mc->join_state == BIT(SENDONLY_FULLMEMBER_JOIN);
4106 if (cma_zero_addr((struct sockaddr *)&mc->addr))
4109 work = kzalloc(sizeof *work, GFP_KERNEL);
4113 mc->multicast.ib = kzalloc(sizeof(struct ib_sa_multicast), GFP_KERNEL);
4114 if (!mc->multicast.ib) {
4119 gid_type = id_priv->cma_dev->default_gid_type[id_priv->id.port_num -
4120 rdma_start_port(id_priv->cma_dev->device)];
4121 cma_iboe_set_mgid(addr, &mc->multicast.ib->rec.mgid, gid_type);
4123 mc->multicast.ib->rec.pkey = cpu_to_be16(0xffff);
4124 if (id_priv->id.ps == RDMA_PS_UDP)
4125 mc->multicast.ib->rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
4127 if (dev_addr->bound_dev_if)
4128 ndev = dev_get_by_index(&init_net, dev_addr->bound_dev_if);
4133 mc->multicast.ib->rec.rate = iboe_get_rate(ndev);
4134 mc->multicast.ib->rec.hop_limit = 1;
4135 mc->multicast.ib->rec.mtu = iboe_get_mtu(ndev->mtu);
4137 if (addr->sa_family == AF_INET) {
4138 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) {
4139 mc->multicast.ib->rec.hop_limit = IPV6_DEFAULT_HOPLIMIT;
4141 err = cma_igmp_send(ndev, &mc->multicast.ib->rec.mgid,
4144 mc->igmp_joined = true;
4148 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP)
4152 if (err || !mc->multicast.ib->rec.mtu) {
4157 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
4158 &mc->multicast.ib->rec.port_gid);
4161 INIT_WORK(&work->work, iboe_mcast_work_handler);
4162 kref_get(&mc->mcref);
4163 queue_work(cma_wq, &work->work);
4168 kfree(mc->multicast.ib);
4174 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
4175 u8 join_state, void *context)
4177 struct rdma_id_private *id_priv;
4178 struct cma_multicast *mc;
4184 id_priv = container_of(id, struct rdma_id_private, id);
4185 if (!cma_comp(id_priv, RDMA_CM_ADDR_BOUND) &&
4186 !cma_comp(id_priv, RDMA_CM_ADDR_RESOLVED))
4189 mc = kmalloc(sizeof *mc, GFP_KERNEL);
4193 memcpy(&mc->addr, addr, rdma_addr_size(addr));
4194 mc->context = context;
4195 mc->id_priv = id_priv;
4196 mc->igmp_joined = false;
4197 mc->join_state = join_state;
4198 spin_lock(&id_priv->lock);
4199 list_add(&mc->list, &id_priv->mc_list);
4200 spin_unlock(&id_priv->lock);
4202 if (rdma_protocol_roce(id->device, id->port_num)) {
4203 kref_init(&mc->mcref);
4204 ret = cma_iboe_join_multicast(id_priv, mc);
4205 } else if (rdma_cap_ib_mcast(id->device, id->port_num))
4206 ret = cma_join_ib_multicast(id_priv, mc);
4211 spin_lock_irq(&id_priv->lock);
4212 list_del(&mc->list);
4213 spin_unlock_irq(&id_priv->lock);
4218 EXPORT_SYMBOL(rdma_join_multicast);
4220 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
4222 struct rdma_id_private *id_priv;
4223 struct cma_multicast *mc;
4225 id_priv = container_of(id, struct rdma_id_private, id);
4226 spin_lock_irq(&id_priv->lock);
4227 list_for_each_entry(mc, &id_priv->mc_list, list) {
4228 if (!memcmp(&mc->addr, addr, rdma_addr_size(addr))) {
4229 list_del(&mc->list);
4230 spin_unlock_irq(&id_priv->lock);
4233 ib_detach_mcast(id->qp,
4234 &mc->multicast.ib->rec.mgid,
4235 be16_to_cpu(mc->multicast.ib->rec.mlid));
4237 BUG_ON(id_priv->cma_dev->device != id->device);
4239 if (rdma_cap_ib_mcast(id->device, id->port_num)) {
4240 ib_sa_free_multicast(mc->multicast.ib);
4242 } else if (rdma_protocol_roce(id->device, id->port_num)) {
4243 if (mc->igmp_joined) {
4244 struct rdma_dev_addr *dev_addr =
4245 &id->route.addr.dev_addr;
4246 struct net_device *ndev = NULL;
4248 if (dev_addr->bound_dev_if)
4249 ndev = dev_get_by_index(&init_net,
4250 dev_addr->bound_dev_if);
4253 &mc->multicast.ib->rec.mgid,
4257 mc->igmp_joined = false;
4259 kref_put(&mc->mcref, release_mc);
4264 spin_unlock_irq(&id_priv->lock);
4266 EXPORT_SYMBOL(rdma_leave_multicast);
4268 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
4270 struct rdma_dev_addr *dev_addr;
4271 struct cma_ndev_work *work;
4273 dev_addr = &id_priv->id.route.addr.dev_addr;
4275 if ((dev_addr->bound_dev_if == ndev->ifindex) &&
4276 (net_eq(dev_net(ndev), dev_addr->net)) &&
4277 memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
4278 pr_info("RDMA CM addr change for ndev %s used by id %p\n",
4279 ndev->name, &id_priv->id);
4280 work = kzalloc(sizeof *work, GFP_KERNEL);
4284 INIT_WORK(&work->work, cma_ndev_work_handler);
4286 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
4287 atomic_inc(&id_priv->refcount);
4288 queue_work(cma_wq, &work->work);
4294 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
4297 struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
4298 struct cma_device *cma_dev;
4299 struct rdma_id_private *id_priv;
4300 int ret = NOTIFY_DONE;
4302 if (event != NETDEV_BONDING_FAILOVER)
4305 if (!(ndev->flags & IFF_MASTER) || !(ndev->priv_flags & IFF_BONDING))
4309 list_for_each_entry(cma_dev, &dev_list, list)
4310 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
4311 ret = cma_netdev_change(ndev, id_priv);
4317 mutex_unlock(&lock);
4321 static struct notifier_block cma_nb = {
4322 .notifier_call = cma_netdev_callback
4325 static void cma_add_one(struct ib_device *device)
4327 struct cma_device *cma_dev;
4328 struct rdma_id_private *id_priv;
4330 unsigned long supported_gids = 0;
4332 cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
4336 cma_dev->device = device;
4337 cma_dev->default_gid_type = kcalloc(device->phys_port_cnt,
4338 sizeof(*cma_dev->default_gid_type),
4340 if (!cma_dev->default_gid_type)
4343 cma_dev->default_roce_tos = kcalloc(device->phys_port_cnt,
4344 sizeof(*cma_dev->default_roce_tos),
4346 if (!cma_dev->default_roce_tos)
4349 for (i = rdma_start_port(device); i <= rdma_end_port(device); i++) {
4350 supported_gids = roce_gid_type_mask_support(device, i);
4351 WARN_ON(!supported_gids);
4352 if (supported_gids & (1 << CMA_PREFERRED_ROCE_GID_TYPE))
4353 cma_dev->default_gid_type[i - rdma_start_port(device)] =
4354 CMA_PREFERRED_ROCE_GID_TYPE;
4356 cma_dev->default_gid_type[i - rdma_start_port(device)] =
4357 find_first_bit(&supported_gids, BITS_PER_LONG);
4358 cma_dev->default_roce_tos[i - rdma_start_port(device)] = 0;
4361 init_completion(&cma_dev->comp);
4362 atomic_set(&cma_dev->refcount, 1);
4363 INIT_LIST_HEAD(&cma_dev->id_list);
4364 ib_set_client_data(device, &cma_client, cma_dev);
4367 list_add_tail(&cma_dev->list, &dev_list);
4368 list_for_each_entry(id_priv, &listen_any_list, list)
4369 cma_listen_on_dev(id_priv, cma_dev);
4370 mutex_unlock(&lock);
4375 kfree(cma_dev->default_gid_type);
4383 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
4385 struct rdma_cm_event event;
4386 enum rdma_cm_state state;
4389 /* Record that we want to remove the device */
4390 state = cma_exch(id_priv, RDMA_CM_DEVICE_REMOVAL);
4391 if (state == RDMA_CM_DESTROYING)
4394 cma_cancel_operation(id_priv, state);
4395 mutex_lock(&id_priv->handler_mutex);
4397 /* Check for destruction from another callback. */
4398 if (!cma_comp(id_priv, RDMA_CM_DEVICE_REMOVAL))
4401 memset(&event, 0, sizeof event);
4402 event.event = RDMA_CM_EVENT_DEVICE_REMOVAL;
4403 ret = id_priv->id.event_handler(&id_priv->id, &event);
4405 mutex_unlock(&id_priv->handler_mutex);
4409 static void cma_process_remove(struct cma_device *cma_dev)
4411 struct rdma_id_private *id_priv;
4415 while (!list_empty(&cma_dev->id_list)) {
4416 id_priv = list_entry(cma_dev->id_list.next,
4417 struct rdma_id_private, list);
4419 list_del(&id_priv->listen_list);
4420 list_del_init(&id_priv->list);
4421 atomic_inc(&id_priv->refcount);
4422 mutex_unlock(&lock);
4424 ret = id_priv->internal_id ? 1 : cma_remove_id_dev(id_priv);
4425 cma_deref_id(id_priv);
4427 rdma_destroy_id(&id_priv->id);
4431 mutex_unlock(&lock);
4433 cma_deref_dev(cma_dev);
4434 wait_for_completion(&cma_dev->comp);
4437 static void cma_remove_one(struct ib_device *device, void *client_data)
4439 struct cma_device *cma_dev = client_data;
4445 list_del(&cma_dev->list);
4446 mutex_unlock(&lock);
4448 cma_process_remove(cma_dev);
4449 kfree(cma_dev->default_roce_tos);
4450 kfree(cma_dev->default_gid_type);
4454 static int cma_get_id_stats(struct sk_buff *skb, struct netlink_callback *cb)
4456 struct nlmsghdr *nlh;
4457 struct rdma_cm_id_stats *id_stats;
4458 struct rdma_id_private *id_priv;
4459 struct rdma_cm_id *id = NULL;
4460 struct cma_device *cma_dev;
4461 int i_dev = 0, i_id = 0;
4464 * We export all of the IDs as a sequence of messages. Each
4465 * ID gets its own netlink message.
4469 list_for_each_entry(cma_dev, &dev_list, list) {
4470 if (i_dev < cb->args[0]) {
4476 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
4477 if (i_id < cb->args[1]) {
4482 id_stats = ibnl_put_msg(skb, &nlh, cb->nlh->nlmsg_seq,
4483 sizeof *id_stats, RDMA_NL_RDMA_CM,
4484 RDMA_NL_RDMA_CM_ID_STATS,
4489 memset(id_stats, 0, sizeof *id_stats);
4491 id_stats->node_type = id->route.addr.dev_addr.dev_type;
4492 id_stats->port_num = id->port_num;
4493 id_stats->bound_dev_if =
4494 id->route.addr.dev_addr.bound_dev_if;
4496 if (ibnl_put_attr(skb, nlh,
4497 rdma_addr_size(cma_src_addr(id_priv)),
4498 cma_src_addr(id_priv),
4499 RDMA_NL_RDMA_CM_ATTR_SRC_ADDR))
4501 if (ibnl_put_attr(skb, nlh,
4502 rdma_addr_size(cma_dst_addr(id_priv)),
4503 cma_dst_addr(id_priv),
4504 RDMA_NL_RDMA_CM_ATTR_DST_ADDR))
4507 id_stats->pid = id_priv->owner;
4508 id_stats->port_space = id->ps;
4509 id_stats->cm_state = id_priv->state;
4510 id_stats->qp_num = id_priv->qp_num;
4511 id_stats->qp_type = id->qp_type;
4514 nlmsg_end(skb, nlh);
4522 mutex_unlock(&lock);
4523 cb->args[0] = i_dev;
4529 static const struct rdma_nl_cbs cma_cb_table[RDMA_NL_RDMA_CM_NUM_OPS] = {
4530 [RDMA_NL_RDMA_CM_ID_STATS] = { .dump = cma_get_id_stats},
4533 static int cma_init_net(struct net *net)
4535 struct cma_pernet *pernet = cma_pernet(net);
4537 idr_init(&pernet->tcp_ps);
4538 idr_init(&pernet->udp_ps);
4539 idr_init(&pernet->ipoib_ps);
4540 idr_init(&pernet->ib_ps);
4545 static void cma_exit_net(struct net *net)
4547 struct cma_pernet *pernet = cma_pernet(net);
4549 idr_destroy(&pernet->tcp_ps);
4550 idr_destroy(&pernet->udp_ps);
4551 idr_destroy(&pernet->ipoib_ps);
4552 idr_destroy(&pernet->ib_ps);
4555 static struct pernet_operations cma_pernet_operations = {
4556 .init = cma_init_net,
4557 .exit = cma_exit_net,
4558 .id = &cma_pernet_id,
4559 .size = sizeof(struct cma_pernet),
4562 static int __init cma_init(void)
4566 cma_wq = alloc_ordered_workqueue("rdma_cm", WQ_MEM_RECLAIM);
4570 ret = register_pernet_subsys(&cma_pernet_operations);
4574 ib_sa_register_client(&sa_client);
4575 rdma_addr_register_client(&addr_client);
4576 register_netdevice_notifier(&cma_nb);
4578 ret = ib_register_client(&cma_client);
4582 rdma_nl_register(RDMA_NL_RDMA_CM, cma_cb_table);
4583 cma_configfs_init();
4588 unregister_netdevice_notifier(&cma_nb);
4589 rdma_addr_unregister_client(&addr_client);
4590 ib_sa_unregister_client(&sa_client);
4591 unregister_pernet_subsys(&cma_pernet_operations);
4593 destroy_workqueue(cma_wq);
4597 static void __exit cma_cleanup(void)
4599 cma_configfs_exit();
4600 rdma_nl_unregister(RDMA_NL_RDMA_CM);
4601 ib_unregister_client(&cma_client);
4602 unregister_netdevice_notifier(&cma_nb);
4603 rdma_addr_unregister_client(&addr_client);
4604 ib_sa_unregister_client(&sa_client);
4605 unregister_pernet_subsys(&cma_pernet_operations);
4606 destroy_workqueue(cma_wq);
4609 MODULE_ALIAS_RDMA_NETLINK(RDMA_NL_RDMA_CM, 1);
4611 module_init(cma_init);
4612 module_exit(cma_cleanup);