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 ib_device *device;
425 union ib_gid local_gid;
431 static int cma_comp(struct rdma_id_private *id_priv, enum rdma_cm_state comp)
436 spin_lock_irqsave(&id_priv->lock, flags);
437 ret = (id_priv->state == comp);
438 spin_unlock_irqrestore(&id_priv->lock, flags);
442 static int cma_comp_exch(struct rdma_id_private *id_priv,
443 enum rdma_cm_state comp, enum rdma_cm_state exch)
448 spin_lock_irqsave(&id_priv->lock, flags);
449 if ((ret = (id_priv->state == comp)))
450 id_priv->state = exch;
451 spin_unlock_irqrestore(&id_priv->lock, flags);
455 static enum rdma_cm_state cma_exch(struct rdma_id_private *id_priv,
456 enum rdma_cm_state exch)
459 enum rdma_cm_state old;
461 spin_lock_irqsave(&id_priv->lock, flags);
462 old = id_priv->state;
463 id_priv->state = exch;
464 spin_unlock_irqrestore(&id_priv->lock, flags);
468 static inline u8 cma_get_ip_ver(const struct cma_hdr *hdr)
470 return hdr->ip_version >> 4;
473 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
475 hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
478 static int cma_igmp_send(struct net_device *ndev, union ib_gid *mgid, bool join)
480 struct in_device *in_dev = NULL;
484 in_dev = __in_dev_get_rtnl(ndev);
487 ip_mc_inc_group(in_dev,
488 *(__be32 *)(mgid->raw + 12));
490 ip_mc_dec_group(in_dev,
491 *(__be32 *)(mgid->raw + 12));
495 return (in_dev) ? 0 : -ENODEV;
498 static void _cma_attach_to_dev(struct rdma_id_private *id_priv,
499 struct cma_device *cma_dev)
501 cma_ref_dev(cma_dev);
502 id_priv->cma_dev = cma_dev;
503 id_priv->gid_type = 0;
504 id_priv->id.device = cma_dev->device;
505 id_priv->id.route.addr.dev_addr.transport =
506 rdma_node_get_transport(cma_dev->device->node_type);
507 list_add_tail(&id_priv->list, &cma_dev->id_list);
510 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
511 struct cma_device *cma_dev)
513 _cma_attach_to_dev(id_priv, cma_dev);
515 cma_dev->default_gid_type[id_priv->id.port_num -
516 rdma_start_port(cma_dev->device)];
519 void cma_deref_dev(struct cma_device *cma_dev)
521 if (atomic_dec_and_test(&cma_dev->refcount))
522 complete(&cma_dev->comp);
525 static inline void release_mc(struct kref *kref)
527 struct cma_multicast *mc = container_of(kref, struct cma_multicast, mcref);
529 kfree(mc->multicast.ib);
533 static void cma_release_dev(struct rdma_id_private *id_priv)
536 list_del(&id_priv->list);
537 cma_deref_dev(id_priv->cma_dev);
538 id_priv->cma_dev = NULL;
542 static inline struct sockaddr *cma_src_addr(struct rdma_id_private *id_priv)
544 return (struct sockaddr *) &id_priv->id.route.addr.src_addr;
547 static inline struct sockaddr *cma_dst_addr(struct rdma_id_private *id_priv)
549 return (struct sockaddr *) &id_priv->id.route.addr.dst_addr;
552 static inline unsigned short cma_family(struct rdma_id_private *id_priv)
554 return id_priv->id.route.addr.src_addr.ss_family;
557 static int cma_set_qkey(struct rdma_id_private *id_priv, u32 qkey)
559 struct ib_sa_mcmember_rec rec;
563 if (qkey && id_priv->qkey != qkey)
569 id_priv->qkey = qkey;
573 switch (id_priv->id.ps) {
576 id_priv->qkey = RDMA_UDP_QKEY;
579 ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid);
580 ret = ib_sa_get_mcmember_rec(id_priv->id.device,
581 id_priv->id.port_num, &rec.mgid,
584 id_priv->qkey = be32_to_cpu(rec.qkey);
592 static void cma_translate_ib(struct sockaddr_ib *sib, struct rdma_dev_addr *dev_addr)
594 dev_addr->dev_type = ARPHRD_INFINIBAND;
595 rdma_addr_set_sgid(dev_addr, (union ib_gid *) &sib->sib_addr);
596 ib_addr_set_pkey(dev_addr, ntohs(sib->sib_pkey));
599 static int cma_translate_addr(struct sockaddr *addr, struct rdma_dev_addr *dev_addr)
603 if (addr->sa_family != AF_IB) {
604 ret = rdma_translate_ip(addr, dev_addr, NULL);
606 cma_translate_ib((struct sockaddr_ib *) addr, dev_addr);
613 static inline int cma_validate_port(struct ib_device *device, u8 port,
614 enum ib_gid_type gid_type,
615 union ib_gid *gid, int dev_type,
619 struct net_device *ndev = NULL;
621 if ((dev_type == ARPHRD_INFINIBAND) && !rdma_protocol_ib(device, port))
624 if ((dev_type != ARPHRD_INFINIBAND) && rdma_protocol_ib(device, port))
627 if (dev_type == ARPHRD_ETHER && rdma_protocol_roce(device, port))
628 ndev = dev_get_by_index(&init_net, bound_if_index);
630 gid_type = IB_GID_TYPE_IB;
633 ret = ib_find_cached_gid_by_port(device, gid, gid_type, port,
642 static int cma_acquire_dev(struct rdma_id_private *id_priv,
643 struct rdma_id_private *listen_id_priv)
645 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
646 struct cma_device *cma_dev;
647 union ib_gid gid, iboe_gid, *gidp;
651 if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
652 id_priv->id.ps == RDMA_PS_IPOIB)
656 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
659 memcpy(&gid, dev_addr->src_dev_addr +
660 rdma_addr_gid_offset(dev_addr), sizeof gid);
662 if (listen_id_priv) {
663 cma_dev = listen_id_priv->cma_dev;
664 port = listen_id_priv->id.port_num;
665 gidp = rdma_protocol_roce(cma_dev->device, port) ?
668 ret = cma_validate_port(cma_dev->device, port,
669 rdma_protocol_ib(cma_dev->device, port) ?
671 listen_id_priv->gid_type, gidp,
673 dev_addr->bound_dev_if);
675 id_priv->id.port_num = port;
680 list_for_each_entry(cma_dev, &dev_list, list) {
681 for (port = 1; port <= cma_dev->device->phys_port_cnt; ++port) {
682 if (listen_id_priv &&
683 listen_id_priv->cma_dev == cma_dev &&
684 listen_id_priv->id.port_num == port)
687 gidp = rdma_protocol_roce(cma_dev->device, port) ?
690 ret = cma_validate_port(cma_dev->device, port,
691 rdma_protocol_ib(cma_dev->device, port) ?
693 cma_dev->default_gid_type[port - 1],
694 gidp, dev_addr->dev_type,
695 dev_addr->bound_dev_if);
697 id_priv->id.port_num = port;
705 cma_attach_to_dev(id_priv, cma_dev);
712 * Select the source IB device and address to reach the destination IB address.
714 static int cma_resolve_ib_dev(struct rdma_id_private *id_priv)
716 struct cma_device *cma_dev, *cur_dev;
717 struct sockaddr_ib *addr;
718 union ib_gid gid, sgid, *dgid;
721 enum ib_port_state port_state;
725 addr = (struct sockaddr_ib *) cma_dst_addr(id_priv);
726 dgid = (union ib_gid *) &addr->sib_addr;
727 pkey = ntohs(addr->sib_pkey);
729 list_for_each_entry(cur_dev, &dev_list, list) {
730 for (p = 1; p <= cur_dev->device->phys_port_cnt; ++p) {
731 if (!rdma_cap_af_ib(cur_dev->device, p))
734 if (ib_find_cached_pkey(cur_dev->device, p, pkey, &index))
737 if (ib_get_cached_port_state(cur_dev->device, p, &port_state))
739 for (i = 0; !ib_get_cached_gid(cur_dev->device, p, i,
742 if (!memcmp(&gid, dgid, sizeof(gid))) {
745 id_priv->id.port_num = p;
749 if (!cma_dev && (gid.global.subnet_prefix ==
750 dgid->global.subnet_prefix) &&
751 port_state == IB_PORT_ACTIVE) {
754 id_priv->id.port_num = p;
764 cma_attach_to_dev(id_priv, cma_dev);
765 addr = (struct sockaddr_ib *) cma_src_addr(id_priv);
766 memcpy(&addr->sib_addr, &sgid, sizeof sgid);
767 cma_translate_ib(addr, &id_priv->id.route.addr.dev_addr);
771 static void cma_deref_id(struct rdma_id_private *id_priv)
773 if (atomic_dec_and_test(&id_priv->refcount))
774 complete(&id_priv->comp);
777 struct rdma_cm_id *rdma_create_id(struct net *net,
778 rdma_cm_event_handler event_handler,
779 void *context, enum rdma_port_space ps,
780 enum ib_qp_type qp_type)
782 struct rdma_id_private *id_priv;
784 id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
786 return ERR_PTR(-ENOMEM);
788 id_priv->owner = task_pid_nr(current);
789 id_priv->state = RDMA_CM_IDLE;
790 id_priv->id.context = context;
791 id_priv->id.event_handler = event_handler;
793 id_priv->id.qp_type = qp_type;
794 id_priv->tos_set = false;
795 spin_lock_init(&id_priv->lock);
796 mutex_init(&id_priv->qp_mutex);
797 init_completion(&id_priv->comp);
798 atomic_set(&id_priv->refcount, 1);
799 mutex_init(&id_priv->handler_mutex);
800 INIT_LIST_HEAD(&id_priv->listen_list);
801 INIT_LIST_HEAD(&id_priv->mc_list);
802 get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
803 id_priv->id.route.addr.dev_addr.net = get_net(net);
804 id_priv->seq_num &= 0x00ffffff;
808 EXPORT_SYMBOL(rdma_create_id);
810 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
812 struct ib_qp_attr qp_attr;
813 int qp_attr_mask, ret;
815 qp_attr.qp_state = IB_QPS_INIT;
816 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
820 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
824 qp_attr.qp_state = IB_QPS_RTR;
825 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
829 qp_attr.qp_state = IB_QPS_RTS;
831 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
836 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
838 struct ib_qp_attr qp_attr;
839 int qp_attr_mask, ret;
841 qp_attr.qp_state = IB_QPS_INIT;
842 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
846 return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
849 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
850 struct ib_qp_init_attr *qp_init_attr)
852 struct rdma_id_private *id_priv;
856 id_priv = container_of(id, struct rdma_id_private, id);
857 if (id->device != pd->device)
860 qp_init_attr->port_num = id->port_num;
861 qp = ib_create_qp(pd, qp_init_attr);
865 if (id->qp_type == IB_QPT_UD)
866 ret = cma_init_ud_qp(id_priv, qp);
868 ret = cma_init_conn_qp(id_priv, qp);
873 id_priv->qp_num = qp->qp_num;
874 id_priv->srq = (qp->srq != NULL);
880 EXPORT_SYMBOL(rdma_create_qp);
882 void rdma_destroy_qp(struct rdma_cm_id *id)
884 struct rdma_id_private *id_priv;
886 id_priv = container_of(id, struct rdma_id_private, id);
887 mutex_lock(&id_priv->qp_mutex);
888 ib_destroy_qp(id_priv->id.qp);
889 id_priv->id.qp = NULL;
890 mutex_unlock(&id_priv->qp_mutex);
892 EXPORT_SYMBOL(rdma_destroy_qp);
894 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
895 struct rdma_conn_param *conn_param)
897 struct ib_qp_attr qp_attr;
898 int qp_attr_mask, ret;
901 mutex_lock(&id_priv->qp_mutex);
902 if (!id_priv->id.qp) {
907 /* Need to update QP attributes from default values. */
908 qp_attr.qp_state = IB_QPS_INIT;
909 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
913 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
917 qp_attr.qp_state = IB_QPS_RTR;
918 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
922 ret = ib_query_gid(id_priv->id.device, id_priv->id.port_num,
923 rdma_ah_read_grh(&qp_attr.ah_attr)->sgid_index,
928 BUG_ON(id_priv->cma_dev->device != id_priv->id.device);
931 qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
932 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
934 mutex_unlock(&id_priv->qp_mutex);
938 static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
939 struct rdma_conn_param *conn_param)
941 struct ib_qp_attr qp_attr;
942 int qp_attr_mask, ret;
944 mutex_lock(&id_priv->qp_mutex);
945 if (!id_priv->id.qp) {
950 qp_attr.qp_state = IB_QPS_RTS;
951 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
956 qp_attr.max_rd_atomic = conn_param->initiator_depth;
957 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
959 mutex_unlock(&id_priv->qp_mutex);
963 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
965 struct ib_qp_attr qp_attr;
968 mutex_lock(&id_priv->qp_mutex);
969 if (!id_priv->id.qp) {
974 qp_attr.qp_state = IB_QPS_ERR;
975 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
977 mutex_unlock(&id_priv->qp_mutex);
981 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
982 struct ib_qp_attr *qp_attr, int *qp_attr_mask)
984 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
988 if (rdma_cap_eth_ah(id_priv->id.device, id_priv->id.port_num))
991 pkey = ib_addr_get_pkey(dev_addr);
993 ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
994 pkey, &qp_attr->pkey_index);
998 qp_attr->port_num = id_priv->id.port_num;
999 *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
1001 if (id_priv->id.qp_type == IB_QPT_UD) {
1002 ret = cma_set_qkey(id_priv, 0);
1006 qp_attr->qkey = id_priv->qkey;
1007 *qp_attr_mask |= IB_QP_QKEY;
1009 qp_attr->qp_access_flags = 0;
1010 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
1015 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
1018 struct rdma_id_private *id_priv;
1021 id_priv = container_of(id, struct rdma_id_private, id);
1022 if (rdma_cap_ib_cm(id->device, id->port_num)) {
1023 if (!id_priv->cm_id.ib || (id_priv->id.qp_type == IB_QPT_UD))
1024 ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
1026 ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
1029 if (qp_attr->qp_state == IB_QPS_RTR)
1030 qp_attr->rq_psn = id_priv->seq_num;
1031 } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
1032 if (!id_priv->cm_id.iw) {
1033 qp_attr->qp_access_flags = 0;
1034 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
1036 ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
1038 qp_attr->port_num = id_priv->id.port_num;
1039 *qp_attr_mask |= IB_QP_PORT;
1045 EXPORT_SYMBOL(rdma_init_qp_attr);
1047 static inline int cma_zero_addr(struct sockaddr *addr)
1049 switch (addr->sa_family) {
1051 return ipv4_is_zeronet(((struct sockaddr_in *)addr)->sin_addr.s_addr);
1053 return ipv6_addr_any(&((struct sockaddr_in6 *) addr)->sin6_addr);
1055 return ib_addr_any(&((struct sockaddr_ib *) addr)->sib_addr);
1061 static inline int cma_loopback_addr(struct sockaddr *addr)
1063 switch (addr->sa_family) {
1065 return ipv4_is_loopback(((struct sockaddr_in *) addr)->sin_addr.s_addr);
1067 return ipv6_addr_loopback(&((struct sockaddr_in6 *) addr)->sin6_addr);
1069 return ib_addr_loopback(&((struct sockaddr_ib *) addr)->sib_addr);
1075 static inline int cma_any_addr(struct sockaddr *addr)
1077 return cma_zero_addr(addr) || cma_loopback_addr(addr);
1080 static int cma_addr_cmp(struct sockaddr *src, struct sockaddr *dst)
1082 if (src->sa_family != dst->sa_family)
1085 switch (src->sa_family) {
1087 return ((struct sockaddr_in *) src)->sin_addr.s_addr !=
1088 ((struct sockaddr_in *) dst)->sin_addr.s_addr;
1090 return ipv6_addr_cmp(&((struct sockaddr_in6 *) src)->sin6_addr,
1091 &((struct sockaddr_in6 *) dst)->sin6_addr);
1093 return ib_addr_cmp(&((struct sockaddr_ib *) src)->sib_addr,
1094 &((struct sockaddr_ib *) dst)->sib_addr);
1098 static __be16 cma_port(struct sockaddr *addr)
1100 struct sockaddr_ib *sib;
1102 switch (addr->sa_family) {
1104 return ((struct sockaddr_in *) addr)->sin_port;
1106 return ((struct sockaddr_in6 *) addr)->sin6_port;
1108 sib = (struct sockaddr_ib *) addr;
1109 return htons((u16) (be64_to_cpu(sib->sib_sid) &
1110 be64_to_cpu(sib->sib_sid_mask)));
1116 static inline int cma_any_port(struct sockaddr *addr)
1118 return !cma_port(addr);
1121 static void cma_save_ib_info(struct sockaddr *src_addr,
1122 struct sockaddr *dst_addr,
1123 struct rdma_cm_id *listen_id,
1124 struct sa_path_rec *path)
1126 struct sockaddr_ib *listen_ib, *ib;
1128 listen_ib = (struct sockaddr_ib *) &listen_id->route.addr.src_addr;
1130 ib = (struct sockaddr_ib *)src_addr;
1131 ib->sib_family = AF_IB;
1133 ib->sib_pkey = path->pkey;
1134 ib->sib_flowinfo = path->flow_label;
1135 memcpy(&ib->sib_addr, &path->sgid, 16);
1136 ib->sib_sid = path->service_id;
1137 ib->sib_scope_id = 0;
1139 ib->sib_pkey = listen_ib->sib_pkey;
1140 ib->sib_flowinfo = listen_ib->sib_flowinfo;
1141 ib->sib_addr = listen_ib->sib_addr;
1142 ib->sib_sid = listen_ib->sib_sid;
1143 ib->sib_scope_id = listen_ib->sib_scope_id;
1145 ib->sib_sid_mask = cpu_to_be64(0xffffffffffffffffULL);
1148 ib = (struct sockaddr_ib *)dst_addr;
1149 ib->sib_family = AF_IB;
1151 ib->sib_pkey = path->pkey;
1152 ib->sib_flowinfo = path->flow_label;
1153 memcpy(&ib->sib_addr, &path->dgid, 16);
1158 static void cma_save_ip4_info(struct sockaddr_in *src_addr,
1159 struct sockaddr_in *dst_addr,
1160 struct cma_hdr *hdr,
1164 *src_addr = (struct sockaddr_in) {
1165 .sin_family = AF_INET,
1166 .sin_addr.s_addr = hdr->dst_addr.ip4.addr,
1167 .sin_port = local_port,
1172 *dst_addr = (struct sockaddr_in) {
1173 .sin_family = AF_INET,
1174 .sin_addr.s_addr = hdr->src_addr.ip4.addr,
1175 .sin_port = hdr->port,
1180 static void cma_save_ip6_info(struct sockaddr_in6 *src_addr,
1181 struct sockaddr_in6 *dst_addr,
1182 struct cma_hdr *hdr,
1186 *src_addr = (struct sockaddr_in6) {
1187 .sin6_family = AF_INET6,
1188 .sin6_addr = hdr->dst_addr.ip6,
1189 .sin6_port = local_port,
1194 *dst_addr = (struct sockaddr_in6) {
1195 .sin6_family = AF_INET6,
1196 .sin6_addr = hdr->src_addr.ip6,
1197 .sin6_port = hdr->port,
1202 static u16 cma_port_from_service_id(__be64 service_id)
1204 return (u16)be64_to_cpu(service_id);
1207 static int cma_save_ip_info(struct sockaddr *src_addr,
1208 struct sockaddr *dst_addr,
1209 struct ib_cm_event *ib_event,
1212 struct cma_hdr *hdr;
1215 hdr = ib_event->private_data;
1216 if (hdr->cma_version != CMA_VERSION)
1219 port = htons(cma_port_from_service_id(service_id));
1221 switch (cma_get_ip_ver(hdr)) {
1223 cma_save_ip4_info((struct sockaddr_in *)src_addr,
1224 (struct sockaddr_in *)dst_addr, hdr, port);
1227 cma_save_ip6_info((struct sockaddr_in6 *)src_addr,
1228 (struct sockaddr_in6 *)dst_addr, hdr, port);
1231 return -EAFNOSUPPORT;
1237 static int cma_save_net_info(struct sockaddr *src_addr,
1238 struct sockaddr *dst_addr,
1239 struct rdma_cm_id *listen_id,
1240 struct ib_cm_event *ib_event,
1241 sa_family_t sa_family, __be64 service_id)
1243 if (sa_family == AF_IB) {
1244 if (ib_event->event == IB_CM_REQ_RECEIVED)
1245 cma_save_ib_info(src_addr, dst_addr, listen_id,
1246 ib_event->param.req_rcvd.primary_path);
1247 else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED)
1248 cma_save_ib_info(src_addr, dst_addr, listen_id, NULL);
1252 return cma_save_ip_info(src_addr, dst_addr, ib_event, service_id);
1255 static int cma_save_req_info(const struct ib_cm_event *ib_event,
1256 struct cma_req_info *req)
1258 const struct ib_cm_req_event_param *req_param =
1259 &ib_event->param.req_rcvd;
1260 const struct ib_cm_sidr_req_event_param *sidr_param =
1261 &ib_event->param.sidr_req_rcvd;
1263 switch (ib_event->event) {
1264 case IB_CM_REQ_RECEIVED:
1265 req->device = req_param->listen_id->device;
1266 req->port = req_param->port;
1267 memcpy(&req->local_gid, &req_param->primary_path->sgid,
1268 sizeof(req->local_gid));
1269 req->has_gid = true;
1270 req->service_id = req_param->primary_path->service_id;
1271 req->pkey = be16_to_cpu(req_param->primary_path->pkey);
1272 if (req->pkey != req_param->bth_pkey)
1273 pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and primary path P_Key (0x%x)\n"
1274 "RDMA CMA: in the future this may cause the request to be dropped\n",
1275 req_param->bth_pkey, req->pkey);
1277 case IB_CM_SIDR_REQ_RECEIVED:
1278 req->device = sidr_param->listen_id->device;
1279 req->port = sidr_param->port;
1280 req->has_gid = false;
1281 req->service_id = sidr_param->service_id;
1282 req->pkey = sidr_param->pkey;
1283 if (req->pkey != sidr_param->bth_pkey)
1284 pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and SIDR request payload P_Key (0x%x)\n"
1285 "RDMA CMA: in the future this may cause the request to be dropped\n",
1286 sidr_param->bth_pkey, req->pkey);
1295 static bool validate_ipv4_net_dev(struct net_device *net_dev,
1296 const struct sockaddr_in *dst_addr,
1297 const struct sockaddr_in *src_addr)
1299 __be32 daddr = dst_addr->sin_addr.s_addr,
1300 saddr = src_addr->sin_addr.s_addr;
1301 struct fib_result res;
1306 if (ipv4_is_multicast(saddr) || ipv4_is_lbcast(saddr) ||
1307 ipv4_is_lbcast(daddr) || ipv4_is_zeronet(saddr) ||
1308 ipv4_is_zeronet(daddr) || ipv4_is_loopback(daddr) ||
1309 ipv4_is_loopback(saddr))
1312 memset(&fl4, 0, sizeof(fl4));
1313 fl4.flowi4_iif = net_dev->ifindex;
1318 err = fib_lookup(dev_net(net_dev), &fl4, &res, 0);
1319 ret = err == 0 && FIB_RES_DEV(res) == net_dev;
1325 static bool validate_ipv6_net_dev(struct net_device *net_dev,
1326 const struct sockaddr_in6 *dst_addr,
1327 const struct sockaddr_in6 *src_addr)
1329 #if IS_ENABLED(CONFIG_IPV6)
1330 const int strict = ipv6_addr_type(&dst_addr->sin6_addr) &
1331 IPV6_ADDR_LINKLOCAL;
1332 struct rt6_info *rt = rt6_lookup(dev_net(net_dev), &dst_addr->sin6_addr,
1333 &src_addr->sin6_addr, net_dev->ifindex,
1340 ret = rt->rt6i_idev->dev == net_dev;
1349 static bool validate_net_dev(struct net_device *net_dev,
1350 const struct sockaddr *daddr,
1351 const struct sockaddr *saddr)
1353 const struct sockaddr_in *daddr4 = (const struct sockaddr_in *)daddr;
1354 const struct sockaddr_in *saddr4 = (const struct sockaddr_in *)saddr;
1355 const struct sockaddr_in6 *daddr6 = (const struct sockaddr_in6 *)daddr;
1356 const struct sockaddr_in6 *saddr6 = (const struct sockaddr_in6 *)saddr;
1358 switch (daddr->sa_family) {
1360 return saddr->sa_family == AF_INET &&
1361 validate_ipv4_net_dev(net_dev, daddr4, saddr4);
1364 return saddr->sa_family == AF_INET6 &&
1365 validate_ipv6_net_dev(net_dev, daddr6, saddr6);
1372 static struct net_device *cma_get_net_dev(struct ib_cm_event *ib_event,
1373 const struct cma_req_info *req)
1375 struct sockaddr_storage listen_addr_storage, src_addr_storage;
1376 struct sockaddr *listen_addr = (struct sockaddr *)&listen_addr_storage,
1377 *src_addr = (struct sockaddr *)&src_addr_storage;
1378 struct net_device *net_dev;
1379 const union ib_gid *gid = req->has_gid ? &req->local_gid : NULL;
1382 err = cma_save_ip_info(listen_addr, src_addr, ib_event,
1385 return ERR_PTR(err);
1387 net_dev = ib_get_net_dev_by_params(req->device, req->port, req->pkey,
1390 return ERR_PTR(-ENODEV);
1392 if (!validate_net_dev(net_dev, listen_addr, src_addr)) {
1394 return ERR_PTR(-EHOSTUNREACH);
1400 static enum rdma_port_space rdma_ps_from_service_id(__be64 service_id)
1402 return (be64_to_cpu(service_id) >> 16) & 0xffff;
1405 static bool cma_match_private_data(struct rdma_id_private *id_priv,
1406 const struct cma_hdr *hdr)
1408 struct sockaddr *addr = cma_src_addr(id_priv);
1410 struct in6_addr ip6_addr;
1412 if (cma_any_addr(addr) && !id_priv->afonly)
1415 switch (addr->sa_family) {
1417 ip4_addr = ((struct sockaddr_in *)addr)->sin_addr.s_addr;
1418 if (cma_get_ip_ver(hdr) != 4)
1420 if (!cma_any_addr(addr) &&
1421 hdr->dst_addr.ip4.addr != ip4_addr)
1425 ip6_addr = ((struct sockaddr_in6 *)addr)->sin6_addr;
1426 if (cma_get_ip_ver(hdr) != 6)
1428 if (!cma_any_addr(addr) &&
1429 memcmp(&hdr->dst_addr.ip6, &ip6_addr, sizeof(ip6_addr)))
1441 static bool cma_protocol_roce_dev_port(struct ib_device *device, int port_num)
1443 enum rdma_link_layer ll = rdma_port_get_link_layer(device, port_num);
1444 enum rdma_transport_type transport =
1445 rdma_node_get_transport(device->node_type);
1447 return ll == IB_LINK_LAYER_ETHERNET && transport == RDMA_TRANSPORT_IB;
1450 static bool cma_protocol_roce(const struct rdma_cm_id *id)
1452 struct ib_device *device = id->device;
1453 const int port_num = id->port_num ?: rdma_start_port(device);
1455 return cma_protocol_roce_dev_port(device, port_num);
1458 static bool cma_match_net_dev(const struct rdma_cm_id *id,
1459 const struct net_device *net_dev,
1462 const struct rdma_addr *addr = &id->route.addr;
1465 /* This request is an AF_IB request or a RoCE request */
1466 return (!id->port_num || id->port_num == port_num) &&
1467 (addr->src_addr.ss_family == AF_IB ||
1468 cma_protocol_roce_dev_port(id->device, port_num));
1470 return !addr->dev_addr.bound_dev_if ||
1471 (net_eq(dev_net(net_dev), addr->dev_addr.net) &&
1472 addr->dev_addr.bound_dev_if == net_dev->ifindex);
1475 static struct rdma_id_private *cma_find_listener(
1476 const struct rdma_bind_list *bind_list,
1477 const struct ib_cm_id *cm_id,
1478 const struct ib_cm_event *ib_event,
1479 const struct cma_req_info *req,
1480 const struct net_device *net_dev)
1482 struct rdma_id_private *id_priv, *id_priv_dev;
1485 return ERR_PTR(-EINVAL);
1487 hlist_for_each_entry(id_priv, &bind_list->owners, node) {
1488 if (cma_match_private_data(id_priv, ib_event->private_data)) {
1489 if (id_priv->id.device == cm_id->device &&
1490 cma_match_net_dev(&id_priv->id, net_dev, req->port))
1492 list_for_each_entry(id_priv_dev,
1493 &id_priv->listen_list,
1495 if (id_priv_dev->id.device == cm_id->device &&
1496 cma_match_net_dev(&id_priv_dev->id, net_dev, req->port))
1502 return ERR_PTR(-EINVAL);
1505 static struct rdma_id_private *cma_id_from_event(struct ib_cm_id *cm_id,
1506 struct ib_cm_event *ib_event,
1507 struct net_device **net_dev)
1509 struct cma_req_info req;
1510 struct rdma_bind_list *bind_list;
1511 struct rdma_id_private *id_priv;
1514 err = cma_save_req_info(ib_event, &req);
1516 return ERR_PTR(err);
1518 *net_dev = cma_get_net_dev(ib_event, &req);
1519 if (IS_ERR(*net_dev)) {
1520 if (PTR_ERR(*net_dev) == -EAFNOSUPPORT) {
1521 /* Assuming the protocol is AF_IB */
1523 } else if (cma_protocol_roce_dev_port(req.device, req.port)) {
1524 /* TODO find the net dev matching the request parameters
1525 * through the RoCE GID table */
1528 return ERR_CAST(*net_dev);
1532 bind_list = cma_ps_find(*net_dev ? dev_net(*net_dev) : &init_net,
1533 rdma_ps_from_service_id(req.service_id),
1534 cma_port_from_service_id(req.service_id));
1535 id_priv = cma_find_listener(bind_list, cm_id, ib_event, &req, *net_dev);
1536 if (IS_ERR(id_priv) && *net_dev) {
1544 static inline u8 cma_user_data_offset(struct rdma_id_private *id_priv)
1546 return cma_family(id_priv) == AF_IB ? 0 : sizeof(struct cma_hdr);
1549 static void cma_cancel_route(struct rdma_id_private *id_priv)
1551 if (rdma_cap_ib_sa(id_priv->id.device, id_priv->id.port_num)) {
1553 ib_sa_cancel_query(id_priv->query_id, id_priv->query);
1557 static void cma_cancel_listens(struct rdma_id_private *id_priv)
1559 struct rdma_id_private *dev_id_priv;
1562 * Remove from listen_any_list to prevent added devices from spawning
1563 * additional listen requests.
1566 list_del(&id_priv->list);
1568 while (!list_empty(&id_priv->listen_list)) {
1569 dev_id_priv = list_entry(id_priv->listen_list.next,
1570 struct rdma_id_private, listen_list);
1571 /* sync with device removal to avoid duplicate destruction */
1572 list_del_init(&dev_id_priv->list);
1573 list_del(&dev_id_priv->listen_list);
1574 mutex_unlock(&lock);
1576 rdma_destroy_id(&dev_id_priv->id);
1579 mutex_unlock(&lock);
1582 static void cma_cancel_operation(struct rdma_id_private *id_priv,
1583 enum rdma_cm_state state)
1586 case RDMA_CM_ADDR_QUERY:
1587 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
1589 case RDMA_CM_ROUTE_QUERY:
1590 cma_cancel_route(id_priv);
1592 case RDMA_CM_LISTEN:
1593 if (cma_any_addr(cma_src_addr(id_priv)) && !id_priv->cma_dev)
1594 cma_cancel_listens(id_priv);
1601 static void cma_release_port(struct rdma_id_private *id_priv)
1603 struct rdma_bind_list *bind_list = id_priv->bind_list;
1604 struct net *net = id_priv->id.route.addr.dev_addr.net;
1610 hlist_del(&id_priv->node);
1611 if (hlist_empty(&bind_list->owners)) {
1612 cma_ps_remove(net, bind_list->ps, bind_list->port);
1615 mutex_unlock(&lock);
1618 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
1620 struct cma_multicast *mc;
1622 while (!list_empty(&id_priv->mc_list)) {
1623 mc = container_of(id_priv->mc_list.next,
1624 struct cma_multicast, list);
1625 list_del(&mc->list);
1626 if (rdma_cap_ib_mcast(id_priv->cma_dev->device,
1627 id_priv->id.port_num)) {
1628 ib_sa_free_multicast(mc->multicast.ib);
1631 if (mc->igmp_joined) {
1632 struct rdma_dev_addr *dev_addr =
1633 &id_priv->id.route.addr.dev_addr;
1634 struct net_device *ndev = NULL;
1636 if (dev_addr->bound_dev_if)
1637 ndev = dev_get_by_index(&init_net,
1638 dev_addr->bound_dev_if);
1641 &mc->multicast.ib->rec.mgid,
1646 kref_put(&mc->mcref, release_mc);
1651 void rdma_destroy_id(struct rdma_cm_id *id)
1653 struct rdma_id_private *id_priv;
1654 enum rdma_cm_state state;
1656 id_priv = container_of(id, struct rdma_id_private, id);
1657 state = cma_exch(id_priv, RDMA_CM_DESTROYING);
1658 cma_cancel_operation(id_priv, state);
1661 * Wait for any active callback to finish. New callbacks will find
1662 * the id_priv state set to destroying and abort.
1664 mutex_lock(&id_priv->handler_mutex);
1665 mutex_unlock(&id_priv->handler_mutex);
1667 if (id_priv->cma_dev) {
1668 if (rdma_cap_ib_cm(id_priv->id.device, 1)) {
1669 if (id_priv->cm_id.ib)
1670 ib_destroy_cm_id(id_priv->cm_id.ib);
1671 } else if (rdma_cap_iw_cm(id_priv->id.device, 1)) {
1672 if (id_priv->cm_id.iw)
1673 iw_destroy_cm_id(id_priv->cm_id.iw);
1675 cma_leave_mc_groups(id_priv);
1676 cma_release_dev(id_priv);
1679 cma_release_port(id_priv);
1680 cma_deref_id(id_priv);
1681 wait_for_completion(&id_priv->comp);
1683 if (id_priv->internal_id)
1684 cma_deref_id(id_priv->id.context);
1686 kfree(id_priv->id.route.path_rec);
1687 put_net(id_priv->id.route.addr.dev_addr.net);
1690 EXPORT_SYMBOL(rdma_destroy_id);
1692 static int cma_rep_recv(struct rdma_id_private *id_priv)
1696 ret = cma_modify_qp_rtr(id_priv, NULL);
1700 ret = cma_modify_qp_rts(id_priv, NULL);
1704 ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
1710 pr_debug_ratelimited("RDMA CM: CONNECT_ERROR: failed to handle reply. status %d\n", ret);
1711 cma_modify_qp_err(id_priv);
1712 ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
1717 static void cma_set_rep_event_data(struct rdma_cm_event *event,
1718 struct ib_cm_rep_event_param *rep_data,
1721 event->param.conn.private_data = private_data;
1722 event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
1723 event->param.conn.responder_resources = rep_data->responder_resources;
1724 event->param.conn.initiator_depth = rep_data->initiator_depth;
1725 event->param.conn.flow_control = rep_data->flow_control;
1726 event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
1727 event->param.conn.srq = rep_data->srq;
1728 event->param.conn.qp_num = rep_data->remote_qpn;
1731 static int cma_ib_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1733 struct rdma_id_private *id_priv = cm_id->context;
1734 struct rdma_cm_event event;
1737 mutex_lock(&id_priv->handler_mutex);
1738 if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
1739 id_priv->state != RDMA_CM_CONNECT) ||
1740 (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
1741 id_priv->state != RDMA_CM_DISCONNECT))
1744 memset(&event, 0, sizeof event);
1745 switch (ib_event->event) {
1746 case IB_CM_REQ_ERROR:
1747 case IB_CM_REP_ERROR:
1748 event.event = RDMA_CM_EVENT_UNREACHABLE;
1749 event.status = -ETIMEDOUT;
1751 case IB_CM_REP_RECEIVED:
1752 if (cma_comp(id_priv, RDMA_CM_CONNECT) &&
1753 (id_priv->id.qp_type != IB_QPT_UD))
1754 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
1755 if (id_priv->id.qp) {
1756 event.status = cma_rep_recv(id_priv);
1757 event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
1758 RDMA_CM_EVENT_ESTABLISHED;
1760 event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
1762 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
1763 ib_event->private_data);
1765 case IB_CM_RTU_RECEIVED:
1766 case IB_CM_USER_ESTABLISHED:
1767 event.event = RDMA_CM_EVENT_ESTABLISHED;
1769 case IB_CM_DREQ_ERROR:
1770 event.status = -ETIMEDOUT; /* fall through */
1771 case IB_CM_DREQ_RECEIVED:
1772 case IB_CM_DREP_RECEIVED:
1773 if (!cma_comp_exch(id_priv, RDMA_CM_CONNECT,
1774 RDMA_CM_DISCONNECT))
1776 event.event = RDMA_CM_EVENT_DISCONNECTED;
1778 case IB_CM_TIMEWAIT_EXIT:
1779 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
1781 case IB_CM_MRA_RECEIVED:
1784 case IB_CM_REJ_RECEIVED:
1785 pr_debug_ratelimited("RDMA CM: REJECTED: %s\n", rdma_reject_msg(&id_priv->id,
1786 ib_event->param.rej_rcvd.reason));
1787 cma_modify_qp_err(id_priv);
1788 event.status = ib_event->param.rej_rcvd.reason;
1789 event.event = RDMA_CM_EVENT_REJECTED;
1790 event.param.conn.private_data = ib_event->private_data;
1791 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
1794 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
1799 ret = id_priv->id.event_handler(&id_priv->id, &event);
1801 /* Destroy the CM ID by returning a non-zero value. */
1802 id_priv->cm_id.ib = NULL;
1803 cma_exch(id_priv, RDMA_CM_DESTROYING);
1804 mutex_unlock(&id_priv->handler_mutex);
1805 rdma_destroy_id(&id_priv->id);
1809 mutex_unlock(&id_priv->handler_mutex);
1813 static struct rdma_id_private *cma_new_conn_id(struct rdma_cm_id *listen_id,
1814 struct ib_cm_event *ib_event,
1815 struct net_device *net_dev)
1817 struct rdma_id_private *id_priv;
1818 struct rdma_cm_id *id;
1819 struct rdma_route *rt;
1820 const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
1821 struct sa_path_rec *path = ib_event->param.req_rcvd.primary_path;
1822 const __be64 service_id =
1823 ib_event->param.req_rcvd.primary_path->service_id;
1826 id = rdma_create_id(listen_id->route.addr.dev_addr.net,
1827 listen_id->event_handler, listen_id->context,
1828 listen_id->ps, ib_event->param.req_rcvd.qp_type);
1832 id_priv = container_of(id, struct rdma_id_private, id);
1833 if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
1834 (struct sockaddr *)&id->route.addr.dst_addr,
1835 listen_id, ib_event, ss_family, service_id))
1839 rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
1840 rt->path_rec = kmalloc(sizeof *rt->path_rec * rt->num_paths,
1845 rt->path_rec[0] = *path;
1846 if (rt->num_paths == 2)
1847 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
1850 rdma_copy_addr(&rt->addr.dev_addr, net_dev, NULL);
1852 if (!cma_protocol_roce(listen_id) &&
1853 cma_any_addr(cma_src_addr(id_priv))) {
1854 rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
1855 rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
1856 ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
1857 } else if (!cma_any_addr(cma_src_addr(id_priv))) {
1858 ret = cma_translate_addr(cma_src_addr(id_priv), &rt->addr.dev_addr);
1863 rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
1865 id_priv->state = RDMA_CM_CONNECT;
1869 rdma_destroy_id(id);
1873 static struct rdma_id_private *cma_new_udp_id(struct rdma_cm_id *listen_id,
1874 struct ib_cm_event *ib_event,
1875 struct net_device *net_dev)
1877 struct rdma_id_private *id_priv;
1878 struct rdma_cm_id *id;
1879 const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
1880 struct net *net = listen_id->route.addr.dev_addr.net;
1883 id = rdma_create_id(net, listen_id->event_handler, listen_id->context,
1884 listen_id->ps, IB_QPT_UD);
1888 id_priv = container_of(id, struct rdma_id_private, id);
1889 if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
1890 (struct sockaddr *)&id->route.addr.dst_addr,
1891 listen_id, ib_event, ss_family,
1892 ib_event->param.sidr_req_rcvd.service_id))
1896 rdma_copy_addr(&id->route.addr.dev_addr, net_dev, NULL);
1898 if (!cma_any_addr(cma_src_addr(id_priv))) {
1899 ret = cma_translate_addr(cma_src_addr(id_priv),
1900 &id->route.addr.dev_addr);
1906 id_priv->state = RDMA_CM_CONNECT;
1909 rdma_destroy_id(id);
1913 static void cma_set_req_event_data(struct rdma_cm_event *event,
1914 struct ib_cm_req_event_param *req_data,
1915 void *private_data, int offset)
1917 event->param.conn.private_data = private_data + offset;
1918 event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
1919 event->param.conn.responder_resources = req_data->responder_resources;
1920 event->param.conn.initiator_depth = req_data->initiator_depth;
1921 event->param.conn.flow_control = req_data->flow_control;
1922 event->param.conn.retry_count = req_data->retry_count;
1923 event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
1924 event->param.conn.srq = req_data->srq;
1925 event->param.conn.qp_num = req_data->remote_qpn;
1928 static int cma_check_req_qp_type(struct rdma_cm_id *id, struct ib_cm_event *ib_event)
1930 return (((ib_event->event == IB_CM_REQ_RECEIVED) &&
1931 (ib_event->param.req_rcvd.qp_type == id->qp_type)) ||
1932 ((ib_event->event == IB_CM_SIDR_REQ_RECEIVED) &&
1933 (id->qp_type == IB_QPT_UD)) ||
1937 static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1939 struct rdma_id_private *listen_id, *conn_id = NULL;
1940 struct rdma_cm_event event;
1941 struct net_device *net_dev;
1945 listen_id = cma_id_from_event(cm_id, ib_event, &net_dev);
1946 if (IS_ERR(listen_id))
1947 return PTR_ERR(listen_id);
1949 if (!cma_check_req_qp_type(&listen_id->id, ib_event)) {
1954 mutex_lock(&listen_id->handler_mutex);
1955 if (listen_id->state != RDMA_CM_LISTEN) {
1956 ret = -ECONNABORTED;
1960 memset(&event, 0, sizeof event);
1961 offset = cma_user_data_offset(listen_id);
1962 event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1963 if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) {
1964 conn_id = cma_new_udp_id(&listen_id->id, ib_event, net_dev);
1965 event.param.ud.private_data = ib_event->private_data + offset;
1966 event.param.ud.private_data_len =
1967 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
1969 conn_id = cma_new_conn_id(&listen_id->id, ib_event, net_dev);
1970 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
1971 ib_event->private_data, offset);
1978 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1979 ret = cma_acquire_dev(conn_id, listen_id);
1983 conn_id->cm_id.ib = cm_id;
1984 cm_id->context = conn_id;
1985 cm_id->cm_handler = cma_ib_handler;
1988 * Protect against the user destroying conn_id from another thread
1989 * until we're done accessing it.
1991 atomic_inc(&conn_id->refcount);
1992 ret = conn_id->id.event_handler(&conn_id->id, &event);
1996 * Acquire mutex to prevent user executing rdma_destroy_id()
1997 * while we're accessing the cm_id.
2000 if (cma_comp(conn_id, RDMA_CM_CONNECT) &&
2001 (conn_id->id.qp_type != IB_QPT_UD))
2002 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
2003 mutex_unlock(&lock);
2004 mutex_unlock(&conn_id->handler_mutex);
2005 mutex_unlock(&listen_id->handler_mutex);
2006 cma_deref_id(conn_id);
2012 cma_deref_id(conn_id);
2013 /* Destroy the CM ID by returning a non-zero value. */
2014 conn_id->cm_id.ib = NULL;
2016 cma_exch(conn_id, RDMA_CM_DESTROYING);
2017 mutex_unlock(&conn_id->handler_mutex);
2019 mutex_unlock(&listen_id->handler_mutex);
2021 rdma_destroy_id(&conn_id->id);
2030 __be64 rdma_get_service_id(struct rdma_cm_id *id, struct sockaddr *addr)
2032 if (addr->sa_family == AF_IB)
2033 return ((struct sockaddr_ib *) addr)->sib_sid;
2035 return cpu_to_be64(((u64)id->ps << 16) + be16_to_cpu(cma_port(addr)));
2037 EXPORT_SYMBOL(rdma_get_service_id);
2039 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
2041 struct rdma_id_private *id_priv = iw_id->context;
2042 struct rdma_cm_event event;
2044 struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
2045 struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
2047 mutex_lock(&id_priv->handler_mutex);
2048 if (id_priv->state != RDMA_CM_CONNECT)
2051 memset(&event, 0, sizeof event);
2052 switch (iw_event->event) {
2053 case IW_CM_EVENT_CLOSE:
2054 event.event = RDMA_CM_EVENT_DISCONNECTED;
2056 case IW_CM_EVENT_CONNECT_REPLY:
2057 memcpy(cma_src_addr(id_priv), laddr,
2058 rdma_addr_size(laddr));
2059 memcpy(cma_dst_addr(id_priv), raddr,
2060 rdma_addr_size(raddr));
2061 switch (iw_event->status) {
2063 event.event = RDMA_CM_EVENT_ESTABLISHED;
2064 event.param.conn.initiator_depth = iw_event->ird;
2065 event.param.conn.responder_resources = iw_event->ord;
2069 event.event = RDMA_CM_EVENT_REJECTED;
2072 event.event = RDMA_CM_EVENT_UNREACHABLE;
2075 event.event = RDMA_CM_EVENT_CONNECT_ERROR;
2079 case IW_CM_EVENT_ESTABLISHED:
2080 event.event = RDMA_CM_EVENT_ESTABLISHED;
2081 event.param.conn.initiator_depth = iw_event->ird;
2082 event.param.conn.responder_resources = iw_event->ord;
2088 event.status = iw_event->status;
2089 event.param.conn.private_data = iw_event->private_data;
2090 event.param.conn.private_data_len = iw_event->private_data_len;
2091 ret = id_priv->id.event_handler(&id_priv->id, &event);
2093 /* Destroy the CM ID by returning a non-zero value. */
2094 id_priv->cm_id.iw = NULL;
2095 cma_exch(id_priv, RDMA_CM_DESTROYING);
2096 mutex_unlock(&id_priv->handler_mutex);
2097 rdma_destroy_id(&id_priv->id);
2102 mutex_unlock(&id_priv->handler_mutex);
2106 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
2107 struct iw_cm_event *iw_event)
2109 struct rdma_cm_id *new_cm_id;
2110 struct rdma_id_private *listen_id, *conn_id;
2111 struct rdma_cm_event event;
2112 int ret = -ECONNABORTED;
2113 struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
2114 struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
2116 listen_id = cm_id->context;
2118 mutex_lock(&listen_id->handler_mutex);
2119 if (listen_id->state != RDMA_CM_LISTEN)
2122 /* Create a new RDMA id for the new IW CM ID */
2123 new_cm_id = rdma_create_id(listen_id->id.route.addr.dev_addr.net,
2124 listen_id->id.event_handler,
2125 listen_id->id.context,
2126 RDMA_PS_TCP, IB_QPT_RC);
2127 if (IS_ERR(new_cm_id)) {
2131 conn_id = container_of(new_cm_id, struct rdma_id_private, id);
2132 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
2133 conn_id->state = RDMA_CM_CONNECT;
2135 ret = rdma_translate_ip(laddr, &conn_id->id.route.addr.dev_addr, NULL);
2137 mutex_unlock(&conn_id->handler_mutex);
2138 rdma_destroy_id(new_cm_id);
2142 ret = cma_acquire_dev(conn_id, listen_id);
2144 mutex_unlock(&conn_id->handler_mutex);
2145 rdma_destroy_id(new_cm_id);
2149 conn_id->cm_id.iw = cm_id;
2150 cm_id->context = conn_id;
2151 cm_id->cm_handler = cma_iw_handler;
2153 memcpy(cma_src_addr(conn_id), laddr, rdma_addr_size(laddr));
2154 memcpy(cma_dst_addr(conn_id), raddr, rdma_addr_size(raddr));
2156 memset(&event, 0, sizeof event);
2157 event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
2158 event.param.conn.private_data = iw_event->private_data;
2159 event.param.conn.private_data_len = iw_event->private_data_len;
2160 event.param.conn.initiator_depth = iw_event->ird;
2161 event.param.conn.responder_resources = iw_event->ord;
2164 * Protect against the user destroying conn_id from another thread
2165 * until we're done accessing it.
2167 atomic_inc(&conn_id->refcount);
2168 ret = conn_id->id.event_handler(&conn_id->id, &event);
2170 /* User wants to destroy the CM ID */
2171 conn_id->cm_id.iw = NULL;
2172 cma_exch(conn_id, RDMA_CM_DESTROYING);
2173 mutex_unlock(&conn_id->handler_mutex);
2174 cma_deref_id(conn_id);
2175 rdma_destroy_id(&conn_id->id);
2179 mutex_unlock(&conn_id->handler_mutex);
2180 cma_deref_id(conn_id);
2183 mutex_unlock(&listen_id->handler_mutex);
2187 static int cma_ib_listen(struct rdma_id_private *id_priv)
2189 struct sockaddr *addr;
2190 struct ib_cm_id *id;
2193 addr = cma_src_addr(id_priv);
2194 svc_id = rdma_get_service_id(&id_priv->id, addr);
2195 id = ib_cm_insert_listen(id_priv->id.device, cma_req_handler, svc_id);
2198 id_priv->cm_id.ib = id;
2203 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
2206 struct iw_cm_id *id;
2208 id = iw_create_cm_id(id_priv->id.device,
2209 iw_conn_req_handler,
2214 id->tos = id_priv->tos;
2215 id_priv->cm_id.iw = id;
2217 memcpy(&id_priv->cm_id.iw->local_addr, cma_src_addr(id_priv),
2218 rdma_addr_size(cma_src_addr(id_priv)));
2220 ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
2223 iw_destroy_cm_id(id_priv->cm_id.iw);
2224 id_priv->cm_id.iw = NULL;
2230 static int cma_listen_handler(struct rdma_cm_id *id,
2231 struct rdma_cm_event *event)
2233 struct rdma_id_private *id_priv = id->context;
2235 id->context = id_priv->id.context;
2236 id->event_handler = id_priv->id.event_handler;
2237 return id_priv->id.event_handler(id, event);
2240 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
2241 struct cma_device *cma_dev)
2243 struct rdma_id_private *dev_id_priv;
2244 struct rdma_cm_id *id;
2245 struct net *net = id_priv->id.route.addr.dev_addr.net;
2248 if (cma_family(id_priv) == AF_IB && !rdma_cap_ib_cm(cma_dev->device, 1))
2251 id = rdma_create_id(net, cma_listen_handler, id_priv, id_priv->id.ps,
2252 id_priv->id.qp_type);
2256 dev_id_priv = container_of(id, struct rdma_id_private, id);
2258 dev_id_priv->state = RDMA_CM_ADDR_BOUND;
2259 memcpy(cma_src_addr(dev_id_priv), cma_src_addr(id_priv),
2260 rdma_addr_size(cma_src_addr(id_priv)));
2262 _cma_attach_to_dev(dev_id_priv, cma_dev);
2263 list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
2264 atomic_inc(&id_priv->refcount);
2265 dev_id_priv->internal_id = 1;
2266 dev_id_priv->afonly = id_priv->afonly;
2268 ret = rdma_listen(id, id_priv->backlog);
2270 pr_warn("RDMA CMA: cma_listen_on_dev, error %d, listening on device %s\n",
2271 ret, cma_dev->device->name);
2274 static void cma_listen_on_all(struct rdma_id_private *id_priv)
2276 struct cma_device *cma_dev;
2279 list_add_tail(&id_priv->list, &listen_any_list);
2280 list_for_each_entry(cma_dev, &dev_list, list)
2281 cma_listen_on_dev(id_priv, cma_dev);
2282 mutex_unlock(&lock);
2285 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
2287 struct rdma_id_private *id_priv;
2289 id_priv = container_of(id, struct rdma_id_private, id);
2290 id_priv->tos = (u8) tos;
2291 id_priv->tos_set = true;
2293 EXPORT_SYMBOL(rdma_set_service_type);
2295 static void cma_query_handler(int status, struct sa_path_rec *path_rec,
2298 struct cma_work *work = context;
2299 struct rdma_route *route;
2301 route = &work->id->id.route;
2304 route->num_paths = 1;
2305 *route->path_rec = *path_rec;
2307 work->old_state = RDMA_CM_ROUTE_QUERY;
2308 work->new_state = RDMA_CM_ADDR_RESOLVED;
2309 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
2310 work->event.status = status;
2311 pr_debug_ratelimited("RDMA CM: ROUTE_ERROR: failed to query path. status %d\n",
2315 queue_work(cma_wq, &work->work);
2318 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms,
2319 struct cma_work *work)
2321 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2322 struct sa_path_rec path_rec;
2323 ib_sa_comp_mask comp_mask;
2324 struct sockaddr_in6 *sin6;
2325 struct sockaddr_ib *sib;
2327 memset(&path_rec, 0, sizeof path_rec);
2329 if (rdma_cap_opa_ah(id_priv->id.device, id_priv->id.port_num))
2330 path_rec.rec_type = SA_PATH_REC_TYPE_OPA;
2332 path_rec.rec_type = SA_PATH_REC_TYPE_IB;
2333 rdma_addr_get_sgid(dev_addr, &path_rec.sgid);
2334 rdma_addr_get_dgid(dev_addr, &path_rec.dgid);
2335 path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
2336 path_rec.numb_path = 1;
2337 path_rec.reversible = 1;
2338 path_rec.service_id = rdma_get_service_id(&id_priv->id,
2339 cma_dst_addr(id_priv));
2341 comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
2342 IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
2343 IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
2345 switch (cma_family(id_priv)) {
2347 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
2348 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
2351 sin6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
2352 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
2353 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2356 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
2357 path_rec.traffic_class = (u8) (be32_to_cpu(sib->sib_flowinfo) >> 20);
2358 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2362 id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
2363 id_priv->id.port_num, &path_rec,
2364 comp_mask, timeout_ms,
2365 GFP_KERNEL, cma_query_handler,
2366 work, &id_priv->query);
2368 return (id_priv->query_id < 0) ? id_priv->query_id : 0;
2371 static void cma_work_handler(struct work_struct *_work)
2373 struct cma_work *work = container_of(_work, struct cma_work, work);
2374 struct rdma_id_private *id_priv = work->id;
2377 mutex_lock(&id_priv->handler_mutex);
2378 if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
2381 if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
2382 cma_exch(id_priv, RDMA_CM_DESTROYING);
2386 mutex_unlock(&id_priv->handler_mutex);
2387 cma_deref_id(id_priv);
2389 rdma_destroy_id(&id_priv->id);
2393 static void cma_ndev_work_handler(struct work_struct *_work)
2395 struct cma_ndev_work *work = container_of(_work, struct cma_ndev_work, work);
2396 struct rdma_id_private *id_priv = work->id;
2399 mutex_lock(&id_priv->handler_mutex);
2400 if (id_priv->state == RDMA_CM_DESTROYING ||
2401 id_priv->state == RDMA_CM_DEVICE_REMOVAL)
2404 if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
2405 cma_exch(id_priv, RDMA_CM_DESTROYING);
2410 mutex_unlock(&id_priv->handler_mutex);
2411 cma_deref_id(id_priv);
2413 rdma_destroy_id(&id_priv->id);
2417 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms)
2419 struct rdma_route *route = &id_priv->id.route;
2420 struct cma_work *work;
2423 work = kzalloc(sizeof *work, GFP_KERNEL);
2428 INIT_WORK(&work->work, cma_work_handler);
2429 work->old_state = RDMA_CM_ROUTE_QUERY;
2430 work->new_state = RDMA_CM_ROUTE_RESOLVED;
2431 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
2433 route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
2434 if (!route->path_rec) {
2439 ret = cma_query_ib_route(id_priv, timeout_ms, work);
2445 kfree(route->path_rec);
2446 route->path_rec = NULL;
2452 int rdma_set_ib_paths(struct rdma_cm_id *id,
2453 struct sa_path_rec *path_rec, int num_paths)
2455 struct rdma_id_private *id_priv;
2458 id_priv = container_of(id, struct rdma_id_private, id);
2459 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
2460 RDMA_CM_ROUTE_RESOLVED))
2463 id->route.path_rec = kmemdup(path_rec, sizeof *path_rec * num_paths,
2465 if (!id->route.path_rec) {
2470 id->route.num_paths = num_paths;
2473 cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED);
2476 EXPORT_SYMBOL(rdma_set_ib_paths);
2478 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms)
2480 struct cma_work *work;
2482 work = kzalloc(sizeof *work, GFP_KERNEL);
2487 INIT_WORK(&work->work, cma_work_handler);
2488 work->old_state = RDMA_CM_ROUTE_QUERY;
2489 work->new_state = RDMA_CM_ROUTE_RESOLVED;
2490 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
2491 queue_work(cma_wq, &work->work);
2495 static int iboe_tos_to_sl(struct net_device *ndev, int tos)
2498 struct net_device *dev;
2500 prio = rt_tos2priority(tos);
2501 dev = is_vlan_dev(ndev) ? vlan_dev_real_dev(ndev) : ndev;
2503 return netdev_get_prio_tc_map(dev, prio);
2505 #if IS_ENABLED(CONFIG_VLAN_8021Q)
2506 if (is_vlan_dev(ndev))
2507 return (vlan_dev_get_egress_qos_mask(ndev, prio) &
2508 VLAN_PRIO_MASK) >> VLAN_PRIO_SHIFT;
2513 static enum ib_gid_type cma_route_gid_type(enum rdma_network_type network_type,
2514 unsigned long supported_gids,
2515 enum ib_gid_type default_gid)
2517 if ((network_type == RDMA_NETWORK_IPV4 ||
2518 network_type == RDMA_NETWORK_IPV6) &&
2519 test_bit(IB_GID_TYPE_ROCE_UDP_ENCAP, &supported_gids))
2520 return IB_GID_TYPE_ROCE_UDP_ENCAP;
2525 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
2527 struct rdma_route *route = &id_priv->id.route;
2528 struct rdma_addr *addr = &route->addr;
2529 struct cma_work *work;
2531 struct net_device *ndev = NULL;
2532 enum ib_gid_type gid_type = IB_GID_TYPE_IB;
2533 u8 default_roce_tos = id_priv->cma_dev->default_roce_tos[id_priv->id.port_num -
2534 rdma_start_port(id_priv->cma_dev->device)];
2535 u8 tos = id_priv->tos_set ? id_priv->tos : default_roce_tos;
2538 work = kzalloc(sizeof *work, GFP_KERNEL);
2543 INIT_WORK(&work->work, cma_work_handler);
2545 route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
2546 if (!route->path_rec) {
2551 route->num_paths = 1;
2553 if (addr->dev_addr.bound_dev_if) {
2554 unsigned long supported_gids;
2556 ndev = dev_get_by_index(&init_net, addr->dev_addr.bound_dev_if);
2562 supported_gids = roce_gid_type_mask_support(id_priv->id.device,
2563 id_priv->id.port_num);
2564 gid_type = cma_route_gid_type(addr->dev_addr.network,
2567 route->path_rec->rec_type =
2568 sa_conv_gid_to_pathrec_type(gid_type);
2569 sa_path_set_ndev(route->path_rec, &init_net);
2570 sa_path_set_ifindex(route->path_rec, ndev->ifindex);
2577 sa_path_set_dmac(route->path_rec, addr->dev_addr.dst_dev_addr);
2579 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
2580 &route->path_rec->sgid);
2581 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.dst_addr,
2582 &route->path_rec->dgid);
2584 /* Use the hint from IP Stack to select GID Type */
2585 if (gid_type < ib_network_to_gid_type(addr->dev_addr.network))
2586 gid_type = ib_network_to_gid_type(addr->dev_addr.network);
2587 route->path_rec->rec_type = sa_conv_gid_to_pathrec_type(gid_type);
2589 if (((struct sockaddr *)&id_priv->id.route.addr.dst_addr)->sa_family != AF_IB)
2590 /* TODO: get the hoplimit from the inet/inet6 device */
2591 route->path_rec->hop_limit = addr->dev_addr.hoplimit;
2593 route->path_rec->hop_limit = 1;
2594 route->path_rec->reversible = 1;
2595 route->path_rec->pkey = cpu_to_be16(0xffff);
2596 route->path_rec->mtu_selector = IB_SA_EQ;
2597 route->path_rec->sl = iboe_tos_to_sl(ndev, tos);
2598 route->path_rec->traffic_class = tos;
2599 route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
2600 route->path_rec->rate_selector = IB_SA_EQ;
2601 route->path_rec->rate = iboe_get_rate(ndev);
2603 route->path_rec->packet_life_time_selector = IB_SA_EQ;
2604 route->path_rec->packet_life_time = CMA_IBOE_PACKET_LIFETIME;
2605 if (!route->path_rec->mtu) {
2610 work->old_state = RDMA_CM_ROUTE_QUERY;
2611 work->new_state = RDMA_CM_ROUTE_RESOLVED;
2612 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
2613 work->event.status = 0;
2615 queue_work(cma_wq, &work->work);
2620 kfree(route->path_rec);
2621 route->path_rec = NULL;
2627 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
2629 struct rdma_id_private *id_priv;
2632 id_priv = container_of(id, struct rdma_id_private, id);
2633 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY))
2636 atomic_inc(&id_priv->refcount);
2637 if (rdma_cap_ib_sa(id->device, id->port_num))
2638 ret = cma_resolve_ib_route(id_priv, timeout_ms);
2639 else if (rdma_protocol_roce(id->device, id->port_num))
2640 ret = cma_resolve_iboe_route(id_priv);
2641 else if (rdma_protocol_iwarp(id->device, id->port_num))
2642 ret = cma_resolve_iw_route(id_priv, timeout_ms);
2651 cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED);
2652 cma_deref_id(id_priv);
2655 EXPORT_SYMBOL(rdma_resolve_route);
2657 static void cma_set_loopback(struct sockaddr *addr)
2659 switch (addr->sa_family) {
2661 ((struct sockaddr_in *) addr)->sin_addr.s_addr = htonl(INADDR_LOOPBACK);
2664 ipv6_addr_set(&((struct sockaddr_in6 *) addr)->sin6_addr,
2668 ib_addr_set(&((struct sockaddr_ib *) addr)->sib_addr,
2674 static int cma_bind_loopback(struct rdma_id_private *id_priv)
2676 struct cma_device *cma_dev, *cur_dev;
2678 enum ib_port_state port_state;
2685 list_for_each_entry(cur_dev, &dev_list, list) {
2686 if (cma_family(id_priv) == AF_IB &&
2687 !rdma_cap_ib_cm(cur_dev->device, 1))
2693 for (p = 1; p <= cur_dev->device->phys_port_cnt; ++p) {
2694 if (!ib_get_cached_port_state(cur_dev->device, p, &port_state) &&
2695 port_state == IB_PORT_ACTIVE) {
2710 ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid, NULL);
2714 ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
2718 id_priv->id.route.addr.dev_addr.dev_type =
2719 (rdma_protocol_ib(cma_dev->device, p)) ?
2720 ARPHRD_INFINIBAND : ARPHRD_ETHER;
2722 rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2723 ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
2724 id_priv->id.port_num = p;
2725 cma_attach_to_dev(id_priv, cma_dev);
2726 cma_set_loopback(cma_src_addr(id_priv));
2728 mutex_unlock(&lock);
2732 static void addr_handler(int status, struct sockaddr *src_addr,
2733 struct rdma_dev_addr *dev_addr, void *context)
2735 struct rdma_id_private *id_priv = context;
2736 struct rdma_cm_event event;
2738 memset(&event, 0, sizeof event);
2739 mutex_lock(&id_priv->handler_mutex);
2740 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY,
2741 RDMA_CM_ADDR_RESOLVED))
2744 memcpy(cma_src_addr(id_priv), src_addr, rdma_addr_size(src_addr));
2745 if (!status && !id_priv->cma_dev) {
2746 status = cma_acquire_dev(id_priv, NULL);
2748 pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to acquire device. status %d\n",
2751 pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to resolve IP. status %d\n", status);
2755 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
2756 RDMA_CM_ADDR_BOUND))
2758 event.event = RDMA_CM_EVENT_ADDR_ERROR;
2759 event.status = status;
2761 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2763 if (id_priv->id.event_handler(&id_priv->id, &event)) {
2764 cma_exch(id_priv, RDMA_CM_DESTROYING);
2765 mutex_unlock(&id_priv->handler_mutex);
2766 cma_deref_id(id_priv);
2767 rdma_destroy_id(&id_priv->id);
2771 mutex_unlock(&id_priv->handler_mutex);
2772 cma_deref_id(id_priv);
2775 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
2777 struct cma_work *work;
2781 work = kzalloc(sizeof *work, GFP_KERNEL);
2785 if (!id_priv->cma_dev) {
2786 ret = cma_bind_loopback(id_priv);
2791 rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2792 rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
2795 INIT_WORK(&work->work, cma_work_handler);
2796 work->old_state = RDMA_CM_ADDR_QUERY;
2797 work->new_state = RDMA_CM_ADDR_RESOLVED;
2798 work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2799 queue_work(cma_wq, &work->work);
2806 static int cma_resolve_ib_addr(struct rdma_id_private *id_priv)
2808 struct cma_work *work;
2811 work = kzalloc(sizeof *work, GFP_KERNEL);
2815 if (!id_priv->cma_dev) {
2816 ret = cma_resolve_ib_dev(id_priv);
2821 rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, (union ib_gid *)
2822 &(((struct sockaddr_ib *) &id_priv->id.route.addr.dst_addr)->sib_addr));
2825 INIT_WORK(&work->work, cma_work_handler);
2826 work->old_state = RDMA_CM_ADDR_QUERY;
2827 work->new_state = RDMA_CM_ADDR_RESOLVED;
2828 work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2829 queue_work(cma_wq, &work->work);
2836 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2837 struct sockaddr *dst_addr)
2839 if (!src_addr || !src_addr->sa_family) {
2840 src_addr = (struct sockaddr *) &id->route.addr.src_addr;
2841 src_addr->sa_family = dst_addr->sa_family;
2842 if (IS_ENABLED(CONFIG_IPV6) &&
2843 dst_addr->sa_family == AF_INET6) {
2844 struct sockaddr_in6 *src_addr6 = (struct sockaddr_in6 *) src_addr;
2845 struct sockaddr_in6 *dst_addr6 = (struct sockaddr_in6 *) dst_addr;
2846 src_addr6->sin6_scope_id = dst_addr6->sin6_scope_id;
2847 if (ipv6_addr_type(&dst_addr6->sin6_addr) & IPV6_ADDR_LINKLOCAL)
2848 id->route.addr.dev_addr.bound_dev_if = dst_addr6->sin6_scope_id;
2849 } else if (dst_addr->sa_family == AF_IB) {
2850 ((struct sockaddr_ib *) src_addr)->sib_pkey =
2851 ((struct sockaddr_ib *) dst_addr)->sib_pkey;
2854 return rdma_bind_addr(id, src_addr);
2857 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2858 struct sockaddr *dst_addr, int timeout_ms)
2860 struct rdma_id_private *id_priv;
2863 id_priv = container_of(id, struct rdma_id_private, id);
2864 memcpy(cma_dst_addr(id_priv), dst_addr, rdma_addr_size(dst_addr));
2865 if (id_priv->state == RDMA_CM_IDLE) {
2866 ret = cma_bind_addr(id, src_addr, dst_addr);
2868 memset(cma_dst_addr(id_priv), 0, rdma_addr_size(dst_addr));
2873 if (cma_family(id_priv) != dst_addr->sa_family) {
2874 memset(cma_dst_addr(id_priv), 0, rdma_addr_size(dst_addr));
2878 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY)) {
2879 memset(cma_dst_addr(id_priv), 0, rdma_addr_size(dst_addr));
2883 atomic_inc(&id_priv->refcount);
2884 if (cma_any_addr(dst_addr)) {
2885 ret = cma_resolve_loopback(id_priv);
2887 if (dst_addr->sa_family == AF_IB) {
2888 ret = cma_resolve_ib_addr(id_priv);
2890 ret = rdma_resolve_ip(&addr_client, cma_src_addr(id_priv),
2891 dst_addr, &id->route.addr.dev_addr,
2892 timeout_ms, addr_handler, id_priv);
2900 cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
2901 cma_deref_id(id_priv);
2904 EXPORT_SYMBOL(rdma_resolve_addr);
2906 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse)
2908 struct rdma_id_private *id_priv;
2909 unsigned long flags;
2912 id_priv = container_of(id, struct rdma_id_private, id);
2913 spin_lock_irqsave(&id_priv->lock, flags);
2914 if (reuse || id_priv->state == RDMA_CM_IDLE) {
2915 id_priv->reuseaddr = reuse;
2920 spin_unlock_irqrestore(&id_priv->lock, flags);
2923 EXPORT_SYMBOL(rdma_set_reuseaddr);
2925 int rdma_set_afonly(struct rdma_cm_id *id, int afonly)
2927 struct rdma_id_private *id_priv;
2928 unsigned long flags;
2931 id_priv = container_of(id, struct rdma_id_private, id);
2932 spin_lock_irqsave(&id_priv->lock, flags);
2933 if (id_priv->state == RDMA_CM_IDLE || id_priv->state == RDMA_CM_ADDR_BOUND) {
2934 id_priv->options |= (1 << CMA_OPTION_AFONLY);
2935 id_priv->afonly = afonly;
2940 spin_unlock_irqrestore(&id_priv->lock, flags);
2943 EXPORT_SYMBOL(rdma_set_afonly);
2945 static void cma_bind_port(struct rdma_bind_list *bind_list,
2946 struct rdma_id_private *id_priv)
2948 struct sockaddr *addr;
2949 struct sockaddr_ib *sib;
2953 addr = cma_src_addr(id_priv);
2954 port = htons(bind_list->port);
2956 switch (addr->sa_family) {
2958 ((struct sockaddr_in *) addr)->sin_port = port;
2961 ((struct sockaddr_in6 *) addr)->sin6_port = port;
2964 sib = (struct sockaddr_ib *) addr;
2965 sid = be64_to_cpu(sib->sib_sid);
2966 mask = be64_to_cpu(sib->sib_sid_mask);
2967 sib->sib_sid = cpu_to_be64((sid & mask) | (u64) ntohs(port));
2968 sib->sib_sid_mask = cpu_to_be64(~0ULL);
2971 id_priv->bind_list = bind_list;
2972 hlist_add_head(&id_priv->node, &bind_list->owners);
2975 static int cma_alloc_port(enum rdma_port_space ps,
2976 struct rdma_id_private *id_priv, unsigned short snum)
2978 struct rdma_bind_list *bind_list;
2981 bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
2985 ret = cma_ps_alloc(id_priv->id.route.addr.dev_addr.net, ps, bind_list,
2991 bind_list->port = (unsigned short)ret;
2992 cma_bind_port(bind_list, id_priv);
2996 return ret == -ENOSPC ? -EADDRNOTAVAIL : ret;
2999 static int cma_port_is_unique(struct rdma_bind_list *bind_list,
3000 struct rdma_id_private *id_priv)
3002 struct rdma_id_private *cur_id;
3003 struct sockaddr *daddr = cma_dst_addr(id_priv);
3004 struct sockaddr *saddr = cma_src_addr(id_priv);
3005 __be16 dport = cma_port(daddr);
3007 hlist_for_each_entry(cur_id, &bind_list->owners, node) {
3008 struct sockaddr *cur_daddr = cma_dst_addr(cur_id);
3009 struct sockaddr *cur_saddr = cma_src_addr(cur_id);
3010 __be16 cur_dport = cma_port(cur_daddr);
3012 if (id_priv == cur_id)
3015 /* different dest port -> unique */
3016 if (!cma_any_port(cur_daddr) &&
3017 (dport != cur_dport))
3020 /* different src address -> unique */
3021 if (!cma_any_addr(saddr) &&
3022 !cma_any_addr(cur_saddr) &&
3023 cma_addr_cmp(saddr, cur_saddr))
3026 /* different dst address -> unique */
3027 if (!cma_any_addr(cur_daddr) &&
3028 cma_addr_cmp(daddr, cur_daddr))
3031 return -EADDRNOTAVAIL;
3036 static int cma_alloc_any_port(enum rdma_port_space ps,
3037 struct rdma_id_private *id_priv)
3039 static unsigned int last_used_port;
3040 int low, high, remaining;
3042 struct net *net = id_priv->id.route.addr.dev_addr.net;
3044 inet_get_local_port_range(net, &low, &high);
3045 remaining = (high - low) + 1;
3046 rover = prandom_u32() % remaining + low;
3048 if (last_used_port != rover) {
3049 struct rdma_bind_list *bind_list;
3052 bind_list = cma_ps_find(net, ps, (unsigned short)rover);
3055 ret = cma_alloc_port(ps, id_priv, rover);
3057 ret = cma_port_is_unique(bind_list, id_priv);
3059 cma_bind_port(bind_list, id_priv);
3062 * Remember previously used port number in order to avoid
3063 * re-using same port immediately after it is closed.
3066 last_used_port = rover;
3067 if (ret != -EADDRNOTAVAIL)
3072 if ((rover < low) || (rover > high))
3076 return -EADDRNOTAVAIL;
3080 * Check that the requested port is available. This is called when trying to
3081 * bind to a specific port, or when trying to listen on a bound port. In
3082 * the latter case, the provided id_priv may already be on the bind_list, but
3083 * we still need to check that it's okay to start listening.
3085 static int cma_check_port(struct rdma_bind_list *bind_list,
3086 struct rdma_id_private *id_priv, uint8_t reuseaddr)
3088 struct rdma_id_private *cur_id;
3089 struct sockaddr *addr, *cur_addr;
3091 addr = cma_src_addr(id_priv);
3092 hlist_for_each_entry(cur_id, &bind_list->owners, node) {
3093 if (id_priv == cur_id)
3096 if ((cur_id->state != RDMA_CM_LISTEN) && reuseaddr &&
3100 cur_addr = cma_src_addr(cur_id);
3101 if (id_priv->afonly && cur_id->afonly &&
3102 (addr->sa_family != cur_addr->sa_family))
3105 if (cma_any_addr(addr) || cma_any_addr(cur_addr))
3106 return -EADDRNOTAVAIL;
3108 if (!cma_addr_cmp(addr, cur_addr))
3114 static int cma_use_port(enum rdma_port_space ps,
3115 struct rdma_id_private *id_priv)
3117 struct rdma_bind_list *bind_list;
3118 unsigned short snum;
3121 snum = ntohs(cma_port(cma_src_addr(id_priv)));
3122 if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
3125 bind_list = cma_ps_find(id_priv->id.route.addr.dev_addr.net, ps, snum);
3127 ret = cma_alloc_port(ps, id_priv, snum);
3129 ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr);
3131 cma_bind_port(bind_list, id_priv);
3136 static int cma_bind_listen(struct rdma_id_private *id_priv)
3138 struct rdma_bind_list *bind_list = id_priv->bind_list;
3142 if (bind_list->owners.first->next)
3143 ret = cma_check_port(bind_list, id_priv, 0);
3144 mutex_unlock(&lock);
3148 static enum rdma_port_space cma_select_inet_ps(
3149 struct rdma_id_private *id_priv)
3151 switch (id_priv->id.ps) {
3156 return id_priv->id.ps;
3163 static enum rdma_port_space cma_select_ib_ps(struct rdma_id_private *id_priv)
3165 enum rdma_port_space ps = 0;
3166 struct sockaddr_ib *sib;
3167 u64 sid_ps, mask, sid;
3169 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
3170 mask = be64_to_cpu(sib->sib_sid_mask) & RDMA_IB_IP_PS_MASK;
3171 sid = be64_to_cpu(sib->sib_sid) & mask;
3173 if ((id_priv->id.ps == RDMA_PS_IB) && (sid == (RDMA_IB_IP_PS_IB & mask))) {
3174 sid_ps = RDMA_IB_IP_PS_IB;
3176 } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_TCP)) &&
3177 (sid == (RDMA_IB_IP_PS_TCP & mask))) {
3178 sid_ps = RDMA_IB_IP_PS_TCP;
3180 } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_UDP)) &&
3181 (sid == (RDMA_IB_IP_PS_UDP & mask))) {
3182 sid_ps = RDMA_IB_IP_PS_UDP;
3187 sib->sib_sid = cpu_to_be64(sid_ps | ntohs(cma_port((struct sockaddr *) sib)));
3188 sib->sib_sid_mask = cpu_to_be64(RDMA_IB_IP_PS_MASK |
3189 be64_to_cpu(sib->sib_sid_mask));
3194 static int cma_get_port(struct rdma_id_private *id_priv)
3196 enum rdma_port_space ps;
3199 if (cma_family(id_priv) != AF_IB)
3200 ps = cma_select_inet_ps(id_priv);
3202 ps = cma_select_ib_ps(id_priv);
3204 return -EPROTONOSUPPORT;
3207 if (cma_any_port(cma_src_addr(id_priv)))
3208 ret = cma_alloc_any_port(ps, id_priv);
3210 ret = cma_use_port(ps, id_priv);
3211 mutex_unlock(&lock);
3216 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
3217 struct sockaddr *addr)
3219 #if IS_ENABLED(CONFIG_IPV6)
3220 struct sockaddr_in6 *sin6;
3222 if (addr->sa_family != AF_INET6)
3225 sin6 = (struct sockaddr_in6 *) addr;
3227 if (!(ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL))
3230 if (!sin6->sin6_scope_id)
3233 dev_addr->bound_dev_if = sin6->sin6_scope_id;
3238 int rdma_listen(struct rdma_cm_id *id, int backlog)
3240 struct rdma_id_private *id_priv;
3243 id_priv = container_of(id, struct rdma_id_private, id);
3244 if (id_priv->state == RDMA_CM_IDLE) {
3245 id->route.addr.src_addr.ss_family = AF_INET;
3246 ret = rdma_bind_addr(id, cma_src_addr(id_priv));
3251 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN))
3254 if (id_priv->reuseaddr) {
3255 ret = cma_bind_listen(id_priv);
3260 id_priv->backlog = backlog;
3262 if (rdma_cap_ib_cm(id->device, 1)) {
3263 ret = cma_ib_listen(id_priv);
3266 } else if (rdma_cap_iw_cm(id->device, 1)) {
3267 ret = cma_iw_listen(id_priv, backlog);
3275 cma_listen_on_all(id_priv);
3279 id_priv->backlog = 0;
3280 cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND);
3283 EXPORT_SYMBOL(rdma_listen);
3285 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
3287 struct rdma_id_private *id_priv;
3289 struct sockaddr *daddr;
3291 if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6 &&
3292 addr->sa_family != AF_IB)
3293 return -EAFNOSUPPORT;
3295 id_priv = container_of(id, struct rdma_id_private, id);
3296 if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND))
3299 ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
3303 memcpy(cma_src_addr(id_priv), addr, rdma_addr_size(addr));
3304 if (!cma_any_addr(addr)) {
3305 ret = cma_translate_addr(addr, &id->route.addr.dev_addr);
3309 ret = cma_acquire_dev(id_priv, NULL);
3314 if (!(id_priv->options & (1 << CMA_OPTION_AFONLY))) {
3315 if (addr->sa_family == AF_INET)
3316 id_priv->afonly = 1;
3317 #if IS_ENABLED(CONFIG_IPV6)
3318 else if (addr->sa_family == AF_INET6) {
3319 struct net *net = id_priv->id.route.addr.dev_addr.net;
3321 id_priv->afonly = net->ipv6.sysctl.bindv6only;
3325 ret = cma_get_port(id_priv);
3329 daddr = cma_dst_addr(id_priv);
3330 daddr->sa_family = addr->sa_family;
3334 if (id_priv->cma_dev)
3335 cma_release_dev(id_priv);
3337 cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE);
3340 EXPORT_SYMBOL(rdma_bind_addr);
3342 static int cma_format_hdr(void *hdr, struct rdma_id_private *id_priv)
3344 struct cma_hdr *cma_hdr;
3347 cma_hdr->cma_version = CMA_VERSION;
3348 if (cma_family(id_priv) == AF_INET) {
3349 struct sockaddr_in *src4, *dst4;
3351 src4 = (struct sockaddr_in *) cma_src_addr(id_priv);
3352 dst4 = (struct sockaddr_in *) cma_dst_addr(id_priv);
3354 cma_set_ip_ver(cma_hdr, 4);
3355 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
3356 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
3357 cma_hdr->port = src4->sin_port;
3358 } else if (cma_family(id_priv) == AF_INET6) {
3359 struct sockaddr_in6 *src6, *dst6;
3361 src6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
3362 dst6 = (struct sockaddr_in6 *) cma_dst_addr(id_priv);
3364 cma_set_ip_ver(cma_hdr, 6);
3365 cma_hdr->src_addr.ip6 = src6->sin6_addr;
3366 cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
3367 cma_hdr->port = src6->sin6_port;
3372 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
3373 struct ib_cm_event *ib_event)
3375 struct rdma_id_private *id_priv = cm_id->context;
3376 struct rdma_cm_event event;
3377 struct ib_cm_sidr_rep_event_param *rep = &ib_event->param.sidr_rep_rcvd;
3380 mutex_lock(&id_priv->handler_mutex);
3381 if (id_priv->state != RDMA_CM_CONNECT)
3384 memset(&event, 0, sizeof event);
3385 switch (ib_event->event) {
3386 case IB_CM_SIDR_REQ_ERROR:
3387 event.event = RDMA_CM_EVENT_UNREACHABLE;
3388 event.status = -ETIMEDOUT;
3390 case IB_CM_SIDR_REP_RECEIVED:
3391 event.param.ud.private_data = ib_event->private_data;
3392 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
3393 if (rep->status != IB_SIDR_SUCCESS) {
3394 event.event = RDMA_CM_EVENT_UNREACHABLE;
3395 event.status = ib_event->param.sidr_rep_rcvd.status;
3396 pr_debug_ratelimited("RDMA CM: UNREACHABLE: bad SIDR reply. status %d\n",
3400 ret = cma_set_qkey(id_priv, rep->qkey);
3402 pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to set qkey. status %d\n", ret);
3403 event.event = RDMA_CM_EVENT_ADDR_ERROR;
3407 ib_init_ah_from_path(id_priv->id.device, id_priv->id.port_num,
3408 id_priv->id.route.path_rec,
3409 &event.param.ud.ah_attr);
3410 event.param.ud.qp_num = rep->qpn;
3411 event.param.ud.qkey = rep->qkey;
3412 event.event = RDMA_CM_EVENT_ESTABLISHED;
3416 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
3421 ret = id_priv->id.event_handler(&id_priv->id, &event);
3423 /* Destroy the CM ID by returning a non-zero value. */
3424 id_priv->cm_id.ib = NULL;
3425 cma_exch(id_priv, RDMA_CM_DESTROYING);
3426 mutex_unlock(&id_priv->handler_mutex);
3427 rdma_destroy_id(&id_priv->id);
3431 mutex_unlock(&id_priv->handler_mutex);
3435 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
3436 struct rdma_conn_param *conn_param)
3438 struct ib_cm_sidr_req_param req;
3439 struct ib_cm_id *id;
3444 memset(&req, 0, sizeof req);
3445 offset = cma_user_data_offset(id_priv);
3446 req.private_data_len = offset + conn_param->private_data_len;
3447 if (req.private_data_len < conn_param->private_data_len)
3450 if (req.private_data_len) {
3451 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
3455 private_data = NULL;
3458 if (conn_param->private_data && conn_param->private_data_len)
3459 memcpy(private_data + offset, conn_param->private_data,
3460 conn_param->private_data_len);
3463 ret = cma_format_hdr(private_data, id_priv);
3466 req.private_data = private_data;
3469 id = ib_create_cm_id(id_priv->id.device, cma_sidr_rep_handler,
3475 id_priv->cm_id.ib = id;
3477 req.path = id_priv->id.route.path_rec;
3478 req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
3479 req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
3480 req.max_cm_retries = CMA_MAX_CM_RETRIES;
3482 ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
3484 ib_destroy_cm_id(id_priv->cm_id.ib);
3485 id_priv->cm_id.ib = NULL;
3488 kfree(private_data);
3492 static int cma_connect_ib(struct rdma_id_private *id_priv,
3493 struct rdma_conn_param *conn_param)
3495 struct ib_cm_req_param req;
3496 struct rdma_route *route;
3498 struct ib_cm_id *id;
3502 memset(&req, 0, sizeof req);
3503 offset = cma_user_data_offset(id_priv);
3504 req.private_data_len = offset + conn_param->private_data_len;
3505 if (req.private_data_len < conn_param->private_data_len)
3508 if (req.private_data_len) {
3509 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
3513 private_data = NULL;
3516 if (conn_param->private_data && conn_param->private_data_len)
3517 memcpy(private_data + offset, conn_param->private_data,
3518 conn_param->private_data_len);
3520 id = ib_create_cm_id(id_priv->id.device, cma_ib_handler, id_priv);
3525 id_priv->cm_id.ib = id;
3527 route = &id_priv->id.route;
3529 ret = cma_format_hdr(private_data, id_priv);
3532 req.private_data = private_data;
3535 req.primary_path = &route->path_rec[0];
3536 if (route->num_paths == 2)
3537 req.alternate_path = &route->path_rec[1];
3539 req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
3540 req.qp_num = id_priv->qp_num;
3541 req.qp_type = id_priv->id.qp_type;
3542 req.starting_psn = id_priv->seq_num;
3543 req.responder_resources = conn_param->responder_resources;
3544 req.initiator_depth = conn_param->initiator_depth;
3545 req.flow_control = conn_param->flow_control;
3546 req.retry_count = min_t(u8, 7, conn_param->retry_count);
3547 req.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
3548 req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
3549 req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
3550 req.max_cm_retries = CMA_MAX_CM_RETRIES;
3551 req.srq = id_priv->srq ? 1 : 0;
3553 ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
3555 if (ret && !IS_ERR(id)) {
3556 ib_destroy_cm_id(id);
3557 id_priv->cm_id.ib = NULL;
3560 kfree(private_data);
3564 static int cma_connect_iw(struct rdma_id_private *id_priv,
3565 struct rdma_conn_param *conn_param)
3567 struct iw_cm_id *cm_id;
3569 struct iw_cm_conn_param iw_param;
3571 cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
3573 return PTR_ERR(cm_id);
3575 cm_id->tos = id_priv->tos;
3576 id_priv->cm_id.iw = cm_id;
3578 memcpy(&cm_id->local_addr, cma_src_addr(id_priv),
3579 rdma_addr_size(cma_src_addr(id_priv)));
3580 memcpy(&cm_id->remote_addr, cma_dst_addr(id_priv),
3581 rdma_addr_size(cma_dst_addr(id_priv)));
3583 ret = cma_modify_qp_rtr(id_priv, conn_param);
3588 iw_param.ord = conn_param->initiator_depth;
3589 iw_param.ird = conn_param->responder_resources;
3590 iw_param.private_data = conn_param->private_data;
3591 iw_param.private_data_len = conn_param->private_data_len;
3592 iw_param.qpn = id_priv->id.qp ? id_priv->qp_num : conn_param->qp_num;
3594 memset(&iw_param, 0, sizeof iw_param);
3595 iw_param.qpn = id_priv->qp_num;
3597 ret = iw_cm_connect(cm_id, &iw_param);
3600 iw_destroy_cm_id(cm_id);
3601 id_priv->cm_id.iw = NULL;
3606 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
3608 struct rdma_id_private *id_priv;
3611 id_priv = container_of(id, struct rdma_id_private, id);
3612 if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT))
3616 id_priv->qp_num = conn_param->qp_num;
3617 id_priv->srq = conn_param->srq;
3620 if (rdma_cap_ib_cm(id->device, id->port_num)) {
3621 if (id->qp_type == IB_QPT_UD)
3622 ret = cma_resolve_ib_udp(id_priv, conn_param);
3624 ret = cma_connect_ib(id_priv, conn_param);
3625 } else if (rdma_cap_iw_cm(id->device, id->port_num))
3626 ret = cma_connect_iw(id_priv, conn_param);
3634 cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED);
3637 EXPORT_SYMBOL(rdma_connect);
3639 static int cma_accept_ib(struct rdma_id_private *id_priv,
3640 struct rdma_conn_param *conn_param)
3642 struct ib_cm_rep_param rep;
3645 ret = cma_modify_qp_rtr(id_priv, conn_param);
3649 ret = cma_modify_qp_rts(id_priv, conn_param);
3653 memset(&rep, 0, sizeof rep);
3654 rep.qp_num = id_priv->qp_num;
3655 rep.starting_psn = id_priv->seq_num;
3656 rep.private_data = conn_param->private_data;
3657 rep.private_data_len = conn_param->private_data_len;
3658 rep.responder_resources = conn_param->responder_resources;
3659 rep.initiator_depth = conn_param->initiator_depth;
3660 rep.failover_accepted = 0;
3661 rep.flow_control = conn_param->flow_control;
3662 rep.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
3663 rep.srq = id_priv->srq ? 1 : 0;
3665 ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
3670 static int cma_accept_iw(struct rdma_id_private *id_priv,
3671 struct rdma_conn_param *conn_param)
3673 struct iw_cm_conn_param iw_param;
3679 ret = cma_modify_qp_rtr(id_priv, conn_param);
3683 iw_param.ord = conn_param->initiator_depth;
3684 iw_param.ird = conn_param->responder_resources;
3685 iw_param.private_data = conn_param->private_data;
3686 iw_param.private_data_len = conn_param->private_data_len;
3687 if (id_priv->id.qp) {
3688 iw_param.qpn = id_priv->qp_num;
3690 iw_param.qpn = conn_param->qp_num;
3692 return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
3695 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
3696 enum ib_cm_sidr_status status, u32 qkey,
3697 const void *private_data, int private_data_len)
3699 struct ib_cm_sidr_rep_param rep;
3702 memset(&rep, 0, sizeof rep);
3703 rep.status = status;
3704 if (status == IB_SIDR_SUCCESS) {
3705 ret = cma_set_qkey(id_priv, qkey);
3708 rep.qp_num = id_priv->qp_num;
3709 rep.qkey = id_priv->qkey;
3711 rep.private_data = private_data;
3712 rep.private_data_len = private_data_len;
3714 return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
3717 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
3719 struct rdma_id_private *id_priv;
3722 id_priv = container_of(id, struct rdma_id_private, id);
3724 id_priv->owner = task_pid_nr(current);
3726 if (!cma_comp(id_priv, RDMA_CM_CONNECT))
3729 if (!id->qp && conn_param) {
3730 id_priv->qp_num = conn_param->qp_num;
3731 id_priv->srq = conn_param->srq;
3734 if (rdma_cap_ib_cm(id->device, id->port_num)) {
3735 if (id->qp_type == IB_QPT_UD) {
3737 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
3739 conn_param->private_data,
3740 conn_param->private_data_len);
3742 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
3746 ret = cma_accept_ib(id_priv, conn_param);
3748 ret = cma_rep_recv(id_priv);
3750 } else if (rdma_cap_iw_cm(id->device, id->port_num))
3751 ret = cma_accept_iw(id_priv, conn_param);
3760 cma_modify_qp_err(id_priv);
3761 rdma_reject(id, NULL, 0);
3764 EXPORT_SYMBOL(rdma_accept);
3766 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
3768 struct rdma_id_private *id_priv;
3771 id_priv = container_of(id, struct rdma_id_private, id);
3772 if (!id_priv->cm_id.ib)
3775 switch (id->device->node_type) {
3776 case RDMA_NODE_IB_CA:
3777 ret = ib_cm_notify(id_priv->cm_id.ib, event);
3785 EXPORT_SYMBOL(rdma_notify);
3787 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
3788 u8 private_data_len)
3790 struct rdma_id_private *id_priv;
3793 id_priv = container_of(id, struct rdma_id_private, id);
3794 if (!id_priv->cm_id.ib)
3797 if (rdma_cap_ib_cm(id->device, id->port_num)) {
3798 if (id->qp_type == IB_QPT_UD)
3799 ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT, 0,
3800 private_data, private_data_len);
3802 ret = ib_send_cm_rej(id_priv->cm_id.ib,
3803 IB_CM_REJ_CONSUMER_DEFINED, NULL,
3804 0, private_data, private_data_len);
3805 } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
3806 ret = iw_cm_reject(id_priv->cm_id.iw,
3807 private_data, private_data_len);
3813 EXPORT_SYMBOL(rdma_reject);
3815 int rdma_disconnect(struct rdma_cm_id *id)
3817 struct rdma_id_private *id_priv;
3820 id_priv = container_of(id, struct rdma_id_private, id);
3821 if (!id_priv->cm_id.ib)
3824 if (rdma_cap_ib_cm(id->device, id->port_num)) {
3825 ret = cma_modify_qp_err(id_priv);
3828 /* Initiate or respond to a disconnect. */
3829 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
3830 ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
3831 } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
3832 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
3839 EXPORT_SYMBOL(rdma_disconnect);
3841 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
3843 struct rdma_id_private *id_priv;
3844 struct cma_multicast *mc = multicast->context;
3845 struct rdma_cm_event event;
3848 id_priv = mc->id_priv;
3849 mutex_lock(&id_priv->handler_mutex);
3850 if (id_priv->state != RDMA_CM_ADDR_BOUND &&
3851 id_priv->state != RDMA_CM_ADDR_RESOLVED)
3855 status = cma_set_qkey(id_priv, be32_to_cpu(multicast->rec.qkey));
3857 pr_debug_ratelimited("RDMA CM: MULTICAST_ERROR: failed to join multicast. status %d\n",
3859 mutex_lock(&id_priv->qp_mutex);
3860 if (!status && id_priv->id.qp) {
3861 status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid,
3862 be16_to_cpu(multicast->rec.mlid));
3864 pr_debug_ratelimited("RDMA CM: MULTICAST_ERROR: failed to attach QP. status %d\n",
3867 mutex_unlock(&id_priv->qp_mutex);
3869 memset(&event, 0, sizeof event);
3870 event.status = status;
3871 event.param.ud.private_data = mc->context;
3873 struct rdma_dev_addr *dev_addr =
3874 &id_priv->id.route.addr.dev_addr;
3875 struct net_device *ndev =
3876 dev_get_by_index(&init_net, dev_addr->bound_dev_if);
3877 enum ib_gid_type gid_type =
3878 id_priv->cma_dev->default_gid_type[id_priv->id.port_num -
3879 rdma_start_port(id_priv->cma_dev->device)];
3881 event.event = RDMA_CM_EVENT_MULTICAST_JOIN;
3882 ib_init_ah_from_mcmember(id_priv->id.device,
3883 id_priv->id.port_num, &multicast->rec,
3885 &event.param.ud.ah_attr);
3886 event.param.ud.qp_num = 0xFFFFFF;
3887 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
3891 event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
3893 ret = id_priv->id.event_handler(&id_priv->id, &event);
3895 cma_exch(id_priv, RDMA_CM_DESTROYING);
3896 mutex_unlock(&id_priv->handler_mutex);
3897 rdma_destroy_id(&id_priv->id);
3902 mutex_unlock(&id_priv->handler_mutex);
3906 static void cma_set_mgid(struct rdma_id_private *id_priv,
3907 struct sockaddr *addr, union ib_gid *mgid)
3909 unsigned char mc_map[MAX_ADDR_LEN];
3910 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3911 struct sockaddr_in *sin = (struct sockaddr_in *) addr;
3912 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
3914 if (cma_any_addr(addr)) {
3915 memset(mgid, 0, sizeof *mgid);
3916 } else if ((addr->sa_family == AF_INET6) &&
3917 ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
3919 /* IPv6 address is an SA assigned MGID. */
3920 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3921 } else if (addr->sa_family == AF_IB) {
3922 memcpy(mgid, &((struct sockaddr_ib *) addr)->sib_addr, sizeof *mgid);
3923 } else if ((addr->sa_family == AF_INET6)) {
3924 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
3925 if (id_priv->id.ps == RDMA_PS_UDP)
3926 mc_map[7] = 0x01; /* Use RDMA CM signature */
3927 *mgid = *(union ib_gid *) (mc_map + 4);
3929 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
3930 if (id_priv->id.ps == RDMA_PS_UDP)
3931 mc_map[7] = 0x01; /* Use RDMA CM signature */
3932 *mgid = *(union ib_gid *) (mc_map + 4);
3936 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
3937 struct cma_multicast *mc)
3939 struct ib_sa_mcmember_rec rec;
3940 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3941 ib_sa_comp_mask comp_mask;
3944 ib_addr_get_mgid(dev_addr, &rec.mgid);
3945 ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
3950 ret = cma_set_qkey(id_priv, 0);
3954 cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
3955 rec.qkey = cpu_to_be32(id_priv->qkey);
3956 rdma_addr_get_sgid(dev_addr, &rec.port_gid);
3957 rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
3958 rec.join_state = mc->join_state;
3960 if ((rec.join_state == BIT(SENDONLY_FULLMEMBER_JOIN)) &&
3961 (!ib_sa_sendonly_fullmem_support(&sa_client,
3963 id_priv->id.port_num))) {
3964 pr_warn("RDMA CM: %s port %u Unable to multicast join\n"
3965 "RDMA CM: SM doesn't support Send Only Full Member option\n",
3966 id_priv->id.device->name, id_priv->id.port_num);
3970 comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
3971 IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
3972 IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
3973 IB_SA_MCMEMBER_REC_FLOW_LABEL |
3974 IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
3976 if (id_priv->id.ps == RDMA_PS_IPOIB)
3977 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
3978 IB_SA_MCMEMBER_REC_RATE_SELECTOR |
3979 IB_SA_MCMEMBER_REC_MTU_SELECTOR |
3980 IB_SA_MCMEMBER_REC_MTU |
3981 IB_SA_MCMEMBER_REC_HOP_LIMIT;
3983 mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device,
3984 id_priv->id.port_num, &rec,
3985 comp_mask, GFP_KERNEL,
3986 cma_ib_mc_handler, mc);
3987 return PTR_ERR_OR_ZERO(mc->multicast.ib);
3990 static void iboe_mcast_work_handler(struct work_struct *work)
3992 struct iboe_mcast_work *mw = container_of(work, struct iboe_mcast_work, work);
3993 struct cma_multicast *mc = mw->mc;
3994 struct ib_sa_multicast *m = mc->multicast.ib;
3996 mc->multicast.ib->context = mc;
3997 cma_ib_mc_handler(0, m);
3998 kref_put(&mc->mcref, release_mc);
4002 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid,
4003 enum ib_gid_type gid_type)
4005 struct sockaddr_in *sin = (struct sockaddr_in *)addr;
4006 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
4008 if (cma_any_addr(addr)) {
4009 memset(mgid, 0, sizeof *mgid);
4010 } else if (addr->sa_family == AF_INET6) {
4011 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
4013 mgid->raw[0] = (gid_type == IB_GID_TYPE_IB) ? 0xff : 0;
4014 mgid->raw[1] = (gid_type == IB_GID_TYPE_IB) ? 0x0e : 0;
4023 mgid->raw[10] = 0xff;
4024 mgid->raw[11] = 0xff;
4025 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
4029 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
4030 struct cma_multicast *mc)
4032 struct iboe_mcast_work *work;
4033 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4035 struct sockaddr *addr = (struct sockaddr *)&mc->addr;
4036 struct net_device *ndev = NULL;
4037 enum ib_gid_type gid_type;
4040 send_only = mc->join_state == BIT(SENDONLY_FULLMEMBER_JOIN);
4042 if (cma_zero_addr((struct sockaddr *)&mc->addr))
4045 work = kzalloc(sizeof *work, GFP_KERNEL);
4049 mc->multicast.ib = kzalloc(sizeof(struct ib_sa_multicast), GFP_KERNEL);
4050 if (!mc->multicast.ib) {
4055 gid_type = id_priv->cma_dev->default_gid_type[id_priv->id.port_num -
4056 rdma_start_port(id_priv->cma_dev->device)];
4057 cma_iboe_set_mgid(addr, &mc->multicast.ib->rec.mgid, gid_type);
4059 mc->multicast.ib->rec.pkey = cpu_to_be16(0xffff);
4060 if (id_priv->id.ps == RDMA_PS_UDP)
4061 mc->multicast.ib->rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
4063 if (dev_addr->bound_dev_if)
4064 ndev = dev_get_by_index(&init_net, dev_addr->bound_dev_if);
4069 mc->multicast.ib->rec.rate = iboe_get_rate(ndev);
4070 mc->multicast.ib->rec.hop_limit = 1;
4071 mc->multicast.ib->rec.mtu = iboe_get_mtu(ndev->mtu);
4073 if (addr->sa_family == AF_INET) {
4074 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) {
4075 mc->multicast.ib->rec.hop_limit = IPV6_DEFAULT_HOPLIMIT;
4077 err = cma_igmp_send(ndev, &mc->multicast.ib->rec.mgid,
4080 mc->igmp_joined = true;
4084 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP)
4088 if (err || !mc->multicast.ib->rec.mtu) {
4093 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
4094 &mc->multicast.ib->rec.port_gid);
4097 INIT_WORK(&work->work, iboe_mcast_work_handler);
4098 kref_get(&mc->mcref);
4099 queue_work(cma_wq, &work->work);
4104 kfree(mc->multicast.ib);
4110 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
4111 u8 join_state, void *context)
4113 struct rdma_id_private *id_priv;
4114 struct cma_multicast *mc;
4117 id_priv = container_of(id, struct rdma_id_private, id);
4118 if (!cma_comp(id_priv, RDMA_CM_ADDR_BOUND) &&
4119 !cma_comp(id_priv, RDMA_CM_ADDR_RESOLVED))
4122 mc = kmalloc(sizeof *mc, GFP_KERNEL);
4126 memcpy(&mc->addr, addr, rdma_addr_size(addr));
4127 mc->context = context;
4128 mc->id_priv = id_priv;
4129 mc->igmp_joined = false;
4130 mc->join_state = join_state;
4131 spin_lock(&id_priv->lock);
4132 list_add(&mc->list, &id_priv->mc_list);
4133 spin_unlock(&id_priv->lock);
4135 if (rdma_protocol_roce(id->device, id->port_num)) {
4136 kref_init(&mc->mcref);
4137 ret = cma_iboe_join_multicast(id_priv, mc);
4138 } else if (rdma_cap_ib_mcast(id->device, id->port_num))
4139 ret = cma_join_ib_multicast(id_priv, mc);
4144 spin_lock_irq(&id_priv->lock);
4145 list_del(&mc->list);
4146 spin_unlock_irq(&id_priv->lock);
4151 EXPORT_SYMBOL(rdma_join_multicast);
4153 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
4155 struct rdma_id_private *id_priv;
4156 struct cma_multicast *mc;
4158 id_priv = container_of(id, struct rdma_id_private, id);
4159 spin_lock_irq(&id_priv->lock);
4160 list_for_each_entry(mc, &id_priv->mc_list, list) {
4161 if (!memcmp(&mc->addr, addr, rdma_addr_size(addr))) {
4162 list_del(&mc->list);
4163 spin_unlock_irq(&id_priv->lock);
4166 ib_detach_mcast(id->qp,
4167 &mc->multicast.ib->rec.mgid,
4168 be16_to_cpu(mc->multicast.ib->rec.mlid));
4170 BUG_ON(id_priv->cma_dev->device != id->device);
4172 if (rdma_cap_ib_mcast(id->device, id->port_num)) {
4173 ib_sa_free_multicast(mc->multicast.ib);
4175 } else if (rdma_protocol_roce(id->device, id->port_num)) {
4176 if (mc->igmp_joined) {
4177 struct rdma_dev_addr *dev_addr =
4178 &id->route.addr.dev_addr;
4179 struct net_device *ndev = NULL;
4181 if (dev_addr->bound_dev_if)
4182 ndev = dev_get_by_index(&init_net,
4183 dev_addr->bound_dev_if);
4186 &mc->multicast.ib->rec.mgid,
4190 mc->igmp_joined = false;
4192 kref_put(&mc->mcref, release_mc);
4197 spin_unlock_irq(&id_priv->lock);
4199 EXPORT_SYMBOL(rdma_leave_multicast);
4201 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
4203 struct rdma_dev_addr *dev_addr;
4204 struct cma_ndev_work *work;
4206 dev_addr = &id_priv->id.route.addr.dev_addr;
4208 if ((dev_addr->bound_dev_if == ndev->ifindex) &&
4209 (net_eq(dev_net(ndev), dev_addr->net)) &&
4210 memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
4211 pr_info("RDMA CM addr change for ndev %s used by id %p\n",
4212 ndev->name, &id_priv->id);
4213 work = kzalloc(sizeof *work, GFP_KERNEL);
4217 INIT_WORK(&work->work, cma_ndev_work_handler);
4219 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
4220 atomic_inc(&id_priv->refcount);
4221 queue_work(cma_wq, &work->work);
4227 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
4230 struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
4231 struct cma_device *cma_dev;
4232 struct rdma_id_private *id_priv;
4233 int ret = NOTIFY_DONE;
4235 if (event != NETDEV_BONDING_FAILOVER)
4238 if (!(ndev->flags & IFF_MASTER) || !(ndev->priv_flags & IFF_BONDING))
4242 list_for_each_entry(cma_dev, &dev_list, list)
4243 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
4244 ret = cma_netdev_change(ndev, id_priv);
4250 mutex_unlock(&lock);
4254 static struct notifier_block cma_nb = {
4255 .notifier_call = cma_netdev_callback
4258 static void cma_add_one(struct ib_device *device)
4260 struct cma_device *cma_dev;
4261 struct rdma_id_private *id_priv;
4263 unsigned long supported_gids = 0;
4265 cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
4269 cma_dev->device = device;
4270 cma_dev->default_gid_type = kcalloc(device->phys_port_cnt,
4271 sizeof(*cma_dev->default_gid_type),
4273 if (!cma_dev->default_gid_type)
4276 cma_dev->default_roce_tos = kcalloc(device->phys_port_cnt,
4277 sizeof(*cma_dev->default_roce_tos),
4279 if (!cma_dev->default_roce_tos)
4282 for (i = rdma_start_port(device); i <= rdma_end_port(device); i++) {
4283 supported_gids = roce_gid_type_mask_support(device, i);
4284 WARN_ON(!supported_gids);
4285 if (supported_gids & (1 << CMA_PREFERRED_ROCE_GID_TYPE))
4286 cma_dev->default_gid_type[i - rdma_start_port(device)] =
4287 CMA_PREFERRED_ROCE_GID_TYPE;
4289 cma_dev->default_gid_type[i - rdma_start_port(device)] =
4290 find_first_bit(&supported_gids, BITS_PER_LONG);
4291 cma_dev->default_roce_tos[i - rdma_start_port(device)] = 0;
4294 init_completion(&cma_dev->comp);
4295 atomic_set(&cma_dev->refcount, 1);
4296 INIT_LIST_HEAD(&cma_dev->id_list);
4297 ib_set_client_data(device, &cma_client, cma_dev);
4300 list_add_tail(&cma_dev->list, &dev_list);
4301 list_for_each_entry(id_priv, &listen_any_list, list)
4302 cma_listen_on_dev(id_priv, cma_dev);
4303 mutex_unlock(&lock);
4308 kfree(cma_dev->default_gid_type);
4316 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
4318 struct rdma_cm_event event;
4319 enum rdma_cm_state state;
4322 /* Record that we want to remove the device */
4323 state = cma_exch(id_priv, RDMA_CM_DEVICE_REMOVAL);
4324 if (state == RDMA_CM_DESTROYING)
4327 cma_cancel_operation(id_priv, state);
4328 mutex_lock(&id_priv->handler_mutex);
4330 /* Check for destruction from another callback. */
4331 if (!cma_comp(id_priv, RDMA_CM_DEVICE_REMOVAL))
4334 memset(&event, 0, sizeof event);
4335 event.event = RDMA_CM_EVENT_DEVICE_REMOVAL;
4336 ret = id_priv->id.event_handler(&id_priv->id, &event);
4338 mutex_unlock(&id_priv->handler_mutex);
4342 static void cma_process_remove(struct cma_device *cma_dev)
4344 struct rdma_id_private *id_priv;
4348 while (!list_empty(&cma_dev->id_list)) {
4349 id_priv = list_entry(cma_dev->id_list.next,
4350 struct rdma_id_private, list);
4352 list_del(&id_priv->listen_list);
4353 list_del_init(&id_priv->list);
4354 atomic_inc(&id_priv->refcount);
4355 mutex_unlock(&lock);
4357 ret = id_priv->internal_id ? 1 : cma_remove_id_dev(id_priv);
4358 cma_deref_id(id_priv);
4360 rdma_destroy_id(&id_priv->id);
4364 mutex_unlock(&lock);
4366 cma_deref_dev(cma_dev);
4367 wait_for_completion(&cma_dev->comp);
4370 static void cma_remove_one(struct ib_device *device, void *client_data)
4372 struct cma_device *cma_dev = client_data;
4378 list_del(&cma_dev->list);
4379 mutex_unlock(&lock);
4381 cma_process_remove(cma_dev);
4382 kfree(cma_dev->default_roce_tos);
4383 kfree(cma_dev->default_gid_type);
4387 static int cma_get_id_stats(struct sk_buff *skb, struct netlink_callback *cb)
4389 struct nlmsghdr *nlh;
4390 struct rdma_cm_id_stats *id_stats;
4391 struct rdma_id_private *id_priv;
4392 struct rdma_cm_id *id = NULL;
4393 struct cma_device *cma_dev;
4394 int i_dev = 0, i_id = 0;
4397 * We export all of the IDs as a sequence of messages. Each
4398 * ID gets its own netlink message.
4402 list_for_each_entry(cma_dev, &dev_list, list) {
4403 if (i_dev < cb->args[0]) {
4409 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
4410 if (i_id < cb->args[1]) {
4415 id_stats = ibnl_put_msg(skb, &nlh, cb->nlh->nlmsg_seq,
4416 sizeof *id_stats, RDMA_NL_RDMA_CM,
4417 RDMA_NL_RDMA_CM_ID_STATS,
4422 memset(id_stats, 0, sizeof *id_stats);
4424 id_stats->node_type = id->route.addr.dev_addr.dev_type;
4425 id_stats->port_num = id->port_num;
4426 id_stats->bound_dev_if =
4427 id->route.addr.dev_addr.bound_dev_if;
4429 if (ibnl_put_attr(skb, nlh,
4430 rdma_addr_size(cma_src_addr(id_priv)),
4431 cma_src_addr(id_priv),
4432 RDMA_NL_RDMA_CM_ATTR_SRC_ADDR))
4434 if (ibnl_put_attr(skb, nlh,
4435 rdma_addr_size(cma_src_addr(id_priv)),
4436 cma_dst_addr(id_priv),
4437 RDMA_NL_RDMA_CM_ATTR_DST_ADDR))
4440 id_stats->pid = id_priv->owner;
4441 id_stats->port_space = id->ps;
4442 id_stats->cm_state = id_priv->state;
4443 id_stats->qp_num = id_priv->qp_num;
4444 id_stats->qp_type = id->qp_type;
4454 mutex_unlock(&lock);
4455 cb->args[0] = i_dev;
4461 static const struct rdma_nl_cbs cma_cb_table[RDMA_NL_RDMA_CM_NUM_OPS] = {
4462 [RDMA_NL_RDMA_CM_ID_STATS] = { .dump = cma_get_id_stats},
4465 static int cma_init_net(struct net *net)
4467 struct cma_pernet *pernet = cma_pernet(net);
4469 idr_init(&pernet->tcp_ps);
4470 idr_init(&pernet->udp_ps);
4471 idr_init(&pernet->ipoib_ps);
4472 idr_init(&pernet->ib_ps);
4477 static void cma_exit_net(struct net *net)
4479 struct cma_pernet *pernet = cma_pernet(net);
4481 idr_destroy(&pernet->tcp_ps);
4482 idr_destroy(&pernet->udp_ps);
4483 idr_destroy(&pernet->ipoib_ps);
4484 idr_destroy(&pernet->ib_ps);
4487 static struct pernet_operations cma_pernet_operations = {
4488 .init = cma_init_net,
4489 .exit = cma_exit_net,
4490 .id = &cma_pernet_id,
4491 .size = sizeof(struct cma_pernet),
4494 static int __init cma_init(void)
4498 cma_wq = alloc_ordered_workqueue("rdma_cm", WQ_MEM_RECLAIM);
4502 ret = register_pernet_subsys(&cma_pernet_operations);
4506 ib_sa_register_client(&sa_client);
4507 rdma_addr_register_client(&addr_client);
4508 register_netdevice_notifier(&cma_nb);
4510 ret = ib_register_client(&cma_client);
4514 rdma_nl_register(RDMA_NL_RDMA_CM, cma_cb_table);
4515 cma_configfs_init();
4520 unregister_netdevice_notifier(&cma_nb);
4521 rdma_addr_unregister_client(&addr_client);
4522 ib_sa_unregister_client(&sa_client);
4524 destroy_workqueue(cma_wq);
4528 static void __exit cma_cleanup(void)
4530 cma_configfs_exit();
4531 rdma_nl_unregister(RDMA_NL_RDMA_CM);
4532 ib_unregister_client(&cma_client);
4533 unregister_netdevice_notifier(&cma_nb);
4534 rdma_addr_unregister_client(&addr_client);
4535 ib_sa_unregister_client(&sa_client);
4536 unregister_pernet_subsys(&cma_pernet_operations);
4537 destroy_workqueue(cma_wq);
4540 MODULE_ALIAS_RDMA_NETLINK(RDMA_NL_RDMA_CM, 1);
4542 module_init(cma_init);
4543 module_exit(cma_cleanup);