2 * Copyright (c) 2005 Voltaire Inc. All rights reserved.
3 * Copyright (c) 2002-2005, Network Appliance, Inc. All rights reserved.
4 * Copyright (c) 1999-2005, Mellanox Technologies, Inc. All rights reserved.
5 * Copyright (c) 2005-2006 Intel Corporation. All rights reserved.
7 * This software is available to you under a choice of one of two
8 * licenses. You may choose to be licensed under the terms of the GNU
9 * General Public License (GPL) Version 2, available from the file
10 * COPYING in the main directory of this source tree, or the
11 * OpenIB.org BSD license below:
13 * Redistribution and use in source and binary forms, with or
14 * without modification, are permitted provided that the following
17 * - Redistributions of source code must retain the above
18 * copyright notice, this list of conditions and the following
21 * - Redistributions in binary form must reproduce the above
22 * copyright notice, this list of conditions and the following
23 * disclaimer in the documentation and/or other materials
24 * provided with the distribution.
26 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
27 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
28 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
29 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
30 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
31 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
32 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
36 #include <linux/completion.h>
38 #include <linux/in6.h>
39 #include <linux/mutex.h>
40 #include <linux/random.h>
41 #include <linux/igmp.h>
42 #include <linux/idr.h>
43 #include <linux/inetdevice.h>
44 #include <linux/slab.h>
45 #include <linux/module.h>
46 #include <net/route.h>
48 #include <net/net_namespace.h>
49 #include <net/netns/generic.h>
52 #include <net/ip_fib.h>
53 #include <net/ip6_route.h>
55 #include <rdma/rdma_cm.h>
56 #include <rdma/rdma_cm_ib.h>
57 #include <rdma/rdma_netlink.h>
59 #include <rdma/ib_cache.h>
60 #include <rdma/ib_cm.h>
61 #include <rdma/ib_sa.h>
62 #include <rdma/iw_cm.h>
64 #include "core_priv.h"
66 MODULE_AUTHOR("Sean Hefty");
67 MODULE_DESCRIPTION("Generic RDMA CM Agent");
68 MODULE_LICENSE("Dual BSD/GPL");
70 #define CMA_CM_RESPONSE_TIMEOUT 20
71 #define CMA_QUERY_CLASSPORT_INFO_TIMEOUT 3000
72 #define CMA_MAX_CM_RETRIES 15
73 #define CMA_CM_MRA_SETTING (IB_CM_MRA_FLAG_DELAY | 24)
74 #define CMA_IBOE_PACKET_LIFETIME 18
76 static const char * const cma_events[] = {
77 [RDMA_CM_EVENT_ADDR_RESOLVED] = "address resolved",
78 [RDMA_CM_EVENT_ADDR_ERROR] = "address error",
79 [RDMA_CM_EVENT_ROUTE_RESOLVED] = "route resolved ",
80 [RDMA_CM_EVENT_ROUTE_ERROR] = "route error",
81 [RDMA_CM_EVENT_CONNECT_REQUEST] = "connect request",
82 [RDMA_CM_EVENT_CONNECT_RESPONSE] = "connect response",
83 [RDMA_CM_EVENT_CONNECT_ERROR] = "connect error",
84 [RDMA_CM_EVENT_UNREACHABLE] = "unreachable",
85 [RDMA_CM_EVENT_REJECTED] = "rejected",
86 [RDMA_CM_EVENT_ESTABLISHED] = "established",
87 [RDMA_CM_EVENT_DISCONNECTED] = "disconnected",
88 [RDMA_CM_EVENT_DEVICE_REMOVAL] = "device removal",
89 [RDMA_CM_EVENT_MULTICAST_JOIN] = "multicast join",
90 [RDMA_CM_EVENT_MULTICAST_ERROR] = "multicast error",
91 [RDMA_CM_EVENT_ADDR_CHANGE] = "address change",
92 [RDMA_CM_EVENT_TIMEWAIT_EXIT] = "timewait exit",
95 const char *__attribute_const__ rdma_event_msg(enum rdma_cm_event_type event)
99 return (index < ARRAY_SIZE(cma_events) && cma_events[index]) ?
100 cma_events[index] : "unrecognized event";
102 EXPORT_SYMBOL(rdma_event_msg);
104 static void cma_add_one(struct ib_device *device);
105 static void cma_remove_one(struct ib_device *device, void *client_data);
107 static struct ib_client cma_client = {
110 .remove = cma_remove_one
113 static struct ib_sa_client sa_client;
114 static struct rdma_addr_client addr_client;
115 static LIST_HEAD(dev_list);
116 static LIST_HEAD(listen_any_list);
117 static DEFINE_MUTEX(lock);
118 static struct workqueue_struct *cma_wq;
119 static int cma_pernet_id;
128 static struct cma_pernet *cma_pernet(struct net *net)
130 return net_generic(net, cma_pernet_id);
133 static struct idr *cma_pernet_idr(struct net *net, enum rdma_port_space ps)
135 struct cma_pernet *pernet = cma_pernet(net);
139 return &pernet->tcp_ps;
141 return &pernet->udp_ps;
143 return &pernet->ipoib_ps;
145 return &pernet->ib_ps;
152 struct list_head list;
153 struct ib_device *device;
154 struct completion comp;
156 struct list_head id_list;
157 enum ib_gid_type *default_gid_type;
160 struct rdma_bind_list {
161 enum rdma_port_space ps;
162 struct hlist_head owners;
166 struct class_port_info_context {
167 struct ib_class_port_info *class_port_info;
168 struct ib_device *device;
169 struct completion done;
170 struct ib_sa_query *sa_query;
174 static int cma_ps_alloc(struct net *net, enum rdma_port_space ps,
175 struct rdma_bind_list *bind_list, int snum)
177 struct idr *idr = cma_pernet_idr(net, ps);
179 return idr_alloc(idr, bind_list, snum, snum + 1, GFP_KERNEL);
182 static struct rdma_bind_list *cma_ps_find(struct net *net,
183 enum rdma_port_space ps, int snum)
185 struct idr *idr = cma_pernet_idr(net, ps);
187 return idr_find(idr, snum);
190 static void cma_ps_remove(struct net *net, enum rdma_port_space ps, int snum)
192 struct idr *idr = cma_pernet_idr(net, ps);
194 idr_remove(idr, snum);
201 void cma_ref_dev(struct cma_device *cma_dev)
203 atomic_inc(&cma_dev->refcount);
206 struct cma_device *cma_enum_devices_by_ibdev(cma_device_filter filter,
209 struct cma_device *cma_dev;
210 struct cma_device *found_cma_dev = NULL;
214 list_for_each_entry(cma_dev, &dev_list, list)
215 if (filter(cma_dev->device, cookie)) {
216 found_cma_dev = cma_dev;
221 cma_ref_dev(found_cma_dev);
223 return found_cma_dev;
226 int cma_get_default_gid_type(struct cma_device *cma_dev,
229 if (port < rdma_start_port(cma_dev->device) ||
230 port > rdma_end_port(cma_dev->device))
233 return cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)];
236 int cma_set_default_gid_type(struct cma_device *cma_dev,
238 enum ib_gid_type default_gid_type)
240 unsigned long supported_gids;
242 if (port < rdma_start_port(cma_dev->device) ||
243 port > rdma_end_port(cma_dev->device))
246 supported_gids = roce_gid_type_mask_support(cma_dev->device, port);
248 if (!(supported_gids & 1 << default_gid_type))
251 cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)] =
257 struct ib_device *cma_get_ib_dev(struct cma_device *cma_dev)
259 return cma_dev->device;
263 * Device removal can occur at anytime, so we need extra handling to
264 * serialize notifying the user of device removal with other callbacks.
265 * We do this by disabling removal notification while a callback is in process,
266 * and reporting it after the callback completes.
268 struct rdma_id_private {
269 struct rdma_cm_id id;
271 struct rdma_bind_list *bind_list;
272 struct hlist_node node;
273 struct list_head list; /* listen_any_list or cma_device.list */
274 struct list_head listen_list; /* per device listens */
275 struct cma_device *cma_dev;
276 struct list_head mc_list;
279 enum rdma_cm_state state;
281 struct mutex qp_mutex;
283 struct completion comp;
285 struct mutex handler_mutex;
289 struct ib_sa_query *query;
305 enum ib_gid_type gid_type;
308 struct cma_multicast {
309 struct rdma_id_private *id_priv;
311 struct ib_sa_multicast *ib;
313 struct list_head list;
315 struct sockaddr_storage addr;
322 struct work_struct work;
323 struct rdma_id_private *id;
324 enum rdma_cm_state old_state;
325 enum rdma_cm_state new_state;
326 struct rdma_cm_event event;
329 struct cma_ndev_work {
330 struct work_struct work;
331 struct rdma_id_private *id;
332 struct rdma_cm_event event;
335 struct iboe_mcast_work {
336 struct work_struct work;
337 struct rdma_id_private *id;
338 struct cma_multicast *mc;
351 u8 ip_version; /* IP version: 7:4 */
353 union cma_ip_addr src_addr;
354 union cma_ip_addr dst_addr;
357 #define CMA_VERSION 0x00
359 struct cma_req_info {
360 struct ib_device *device;
362 union ib_gid local_gid;
368 static int cma_comp(struct rdma_id_private *id_priv, enum rdma_cm_state comp)
373 spin_lock_irqsave(&id_priv->lock, flags);
374 ret = (id_priv->state == comp);
375 spin_unlock_irqrestore(&id_priv->lock, flags);
379 static int cma_comp_exch(struct rdma_id_private *id_priv,
380 enum rdma_cm_state comp, enum rdma_cm_state exch)
385 spin_lock_irqsave(&id_priv->lock, flags);
386 if ((ret = (id_priv->state == comp)))
387 id_priv->state = exch;
388 spin_unlock_irqrestore(&id_priv->lock, flags);
392 static enum rdma_cm_state cma_exch(struct rdma_id_private *id_priv,
393 enum rdma_cm_state exch)
396 enum rdma_cm_state old;
398 spin_lock_irqsave(&id_priv->lock, flags);
399 old = id_priv->state;
400 id_priv->state = exch;
401 spin_unlock_irqrestore(&id_priv->lock, flags);
405 static inline u8 cma_get_ip_ver(const struct cma_hdr *hdr)
407 return hdr->ip_version >> 4;
410 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
412 hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
415 static int cma_igmp_send(struct net_device *ndev, union ib_gid *mgid, bool join)
417 struct in_device *in_dev = NULL;
421 in_dev = __in_dev_get_rtnl(ndev);
424 ip_mc_inc_group(in_dev,
425 *(__be32 *)(mgid->raw + 12));
427 ip_mc_dec_group(in_dev,
428 *(__be32 *)(mgid->raw + 12));
432 return (in_dev) ? 0 : -ENODEV;
435 static void _cma_attach_to_dev(struct rdma_id_private *id_priv,
436 struct cma_device *cma_dev)
438 cma_ref_dev(cma_dev);
439 id_priv->cma_dev = cma_dev;
440 id_priv->gid_type = 0;
441 id_priv->id.device = cma_dev->device;
442 id_priv->id.route.addr.dev_addr.transport =
443 rdma_node_get_transport(cma_dev->device->node_type);
444 list_add_tail(&id_priv->list, &cma_dev->id_list);
447 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
448 struct cma_device *cma_dev)
450 _cma_attach_to_dev(id_priv, cma_dev);
452 cma_dev->default_gid_type[id_priv->id.port_num -
453 rdma_start_port(cma_dev->device)];
456 void cma_deref_dev(struct cma_device *cma_dev)
458 if (atomic_dec_and_test(&cma_dev->refcount))
459 complete(&cma_dev->comp);
462 static inline void release_mc(struct kref *kref)
464 struct cma_multicast *mc = container_of(kref, struct cma_multicast, mcref);
466 kfree(mc->multicast.ib);
470 static void cma_release_dev(struct rdma_id_private *id_priv)
473 list_del(&id_priv->list);
474 cma_deref_dev(id_priv->cma_dev);
475 id_priv->cma_dev = NULL;
479 static inline struct sockaddr *cma_src_addr(struct rdma_id_private *id_priv)
481 return (struct sockaddr *) &id_priv->id.route.addr.src_addr;
484 static inline struct sockaddr *cma_dst_addr(struct rdma_id_private *id_priv)
486 return (struct sockaddr *) &id_priv->id.route.addr.dst_addr;
489 static inline unsigned short cma_family(struct rdma_id_private *id_priv)
491 return id_priv->id.route.addr.src_addr.ss_family;
494 static int cma_set_qkey(struct rdma_id_private *id_priv, u32 qkey)
496 struct ib_sa_mcmember_rec rec;
500 if (qkey && id_priv->qkey != qkey)
506 id_priv->qkey = qkey;
510 switch (id_priv->id.ps) {
513 id_priv->qkey = RDMA_UDP_QKEY;
516 ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid);
517 ret = ib_sa_get_mcmember_rec(id_priv->id.device,
518 id_priv->id.port_num, &rec.mgid,
521 id_priv->qkey = be32_to_cpu(rec.qkey);
529 static void cma_translate_ib(struct sockaddr_ib *sib, struct rdma_dev_addr *dev_addr)
531 dev_addr->dev_type = ARPHRD_INFINIBAND;
532 rdma_addr_set_sgid(dev_addr, (union ib_gid *) &sib->sib_addr);
533 ib_addr_set_pkey(dev_addr, ntohs(sib->sib_pkey));
536 static int cma_translate_addr(struct sockaddr *addr, struct rdma_dev_addr *dev_addr)
540 if (addr->sa_family != AF_IB) {
541 ret = rdma_translate_ip(addr, dev_addr, NULL);
543 cma_translate_ib((struct sockaddr_ib *) addr, dev_addr);
550 static inline int cma_validate_port(struct ib_device *device, u8 port,
551 enum ib_gid_type gid_type,
552 union ib_gid *gid, int dev_type,
556 struct net_device *ndev = NULL;
558 if ((dev_type == ARPHRD_INFINIBAND) && !rdma_protocol_ib(device, port))
561 if ((dev_type != ARPHRD_INFINIBAND) && rdma_protocol_ib(device, port))
564 if (dev_type == ARPHRD_ETHER && rdma_protocol_roce(device, port)) {
565 ndev = dev_get_by_index(&init_net, bound_if_index);
566 if (ndev && ndev->flags & IFF_LOOPBACK) {
567 pr_info("detected loopback device\n");
570 if (!device->get_netdev)
573 ndev = device->get_netdev(device, port);
578 gid_type = IB_GID_TYPE_IB;
581 ret = ib_find_cached_gid_by_port(device, gid, gid_type, port,
590 static int cma_acquire_dev(struct rdma_id_private *id_priv,
591 struct rdma_id_private *listen_id_priv)
593 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
594 struct cma_device *cma_dev;
595 union ib_gid gid, iboe_gid, *gidp;
599 if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
600 id_priv->id.ps == RDMA_PS_IPOIB)
604 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
607 memcpy(&gid, dev_addr->src_dev_addr +
608 rdma_addr_gid_offset(dev_addr), sizeof gid);
610 if (listen_id_priv) {
611 cma_dev = listen_id_priv->cma_dev;
612 port = listen_id_priv->id.port_num;
613 gidp = rdma_protocol_roce(cma_dev->device, port) ?
616 ret = cma_validate_port(cma_dev->device, port,
617 rdma_protocol_ib(cma_dev->device, port) ?
619 listen_id_priv->gid_type, gidp,
621 dev_addr->bound_dev_if);
623 id_priv->id.port_num = port;
628 list_for_each_entry(cma_dev, &dev_list, list) {
629 for (port = 1; port <= cma_dev->device->phys_port_cnt; ++port) {
630 if (listen_id_priv &&
631 listen_id_priv->cma_dev == cma_dev &&
632 listen_id_priv->id.port_num == port)
635 gidp = rdma_protocol_roce(cma_dev->device, port) ?
638 ret = cma_validate_port(cma_dev->device, port,
639 rdma_protocol_ib(cma_dev->device, port) ?
641 cma_dev->default_gid_type[port - 1],
642 gidp, dev_addr->dev_type,
643 dev_addr->bound_dev_if);
645 id_priv->id.port_num = port;
653 cma_attach_to_dev(id_priv, cma_dev);
660 * Select the source IB device and address to reach the destination IB address.
662 static int cma_resolve_ib_dev(struct rdma_id_private *id_priv)
664 struct cma_device *cma_dev, *cur_dev;
665 struct sockaddr_ib *addr;
666 union ib_gid gid, sgid, *dgid;
672 addr = (struct sockaddr_ib *) cma_dst_addr(id_priv);
673 dgid = (union ib_gid *) &addr->sib_addr;
674 pkey = ntohs(addr->sib_pkey);
677 list_for_each_entry(cur_dev, &dev_list, list) {
678 for (p = 1; p <= cur_dev->device->phys_port_cnt; ++p) {
679 if (!rdma_cap_af_ib(cur_dev->device, p))
682 if (ib_find_cached_pkey(cur_dev->device, p, pkey, &index))
685 for (i = 0; !ib_get_cached_gid(cur_dev->device, p, i,
688 if (!memcmp(&gid, dgid, sizeof(gid))) {
691 id_priv->id.port_num = p;
695 if (!cma_dev && (gid.global.subnet_prefix ==
696 dgid->global.subnet_prefix)) {
699 id_priv->id.port_num = p;
709 cma_attach_to_dev(id_priv, cma_dev);
711 addr = (struct sockaddr_ib *)cma_src_addr(id_priv);
712 memcpy(&addr->sib_addr, &sgid, sizeof(sgid));
713 cma_translate_ib(addr, &id_priv->id.route.addr.dev_addr);
717 static void cma_deref_id(struct rdma_id_private *id_priv)
719 if (atomic_dec_and_test(&id_priv->refcount))
720 complete(&id_priv->comp);
723 struct rdma_cm_id *rdma_create_id(struct net *net,
724 rdma_cm_event_handler event_handler,
725 void *context, enum rdma_port_space ps,
726 enum ib_qp_type qp_type)
728 struct rdma_id_private *id_priv;
730 id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
732 return ERR_PTR(-ENOMEM);
734 id_priv->owner = task_pid_nr(current);
735 id_priv->state = RDMA_CM_IDLE;
736 id_priv->id.context = context;
737 id_priv->id.event_handler = event_handler;
739 id_priv->id.qp_type = qp_type;
740 spin_lock_init(&id_priv->lock);
741 mutex_init(&id_priv->qp_mutex);
742 init_completion(&id_priv->comp);
743 atomic_set(&id_priv->refcount, 1);
744 mutex_init(&id_priv->handler_mutex);
745 INIT_LIST_HEAD(&id_priv->listen_list);
746 INIT_LIST_HEAD(&id_priv->mc_list);
747 get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
748 id_priv->id.route.addr.dev_addr.net = get_net(net);
749 id_priv->seq_num &= 0x00ffffff;
753 EXPORT_SYMBOL(rdma_create_id);
755 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
757 struct ib_qp_attr qp_attr;
758 int qp_attr_mask, ret;
760 qp_attr.qp_state = IB_QPS_INIT;
761 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
765 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
769 qp_attr.qp_state = IB_QPS_RTR;
770 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
774 qp_attr.qp_state = IB_QPS_RTS;
776 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
781 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
783 struct ib_qp_attr qp_attr;
784 int qp_attr_mask, ret;
786 qp_attr.qp_state = IB_QPS_INIT;
787 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
791 return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
794 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
795 struct ib_qp_init_attr *qp_init_attr)
797 struct rdma_id_private *id_priv;
801 id_priv = container_of(id, struct rdma_id_private, id);
802 if (id->device != pd->device)
805 qp_init_attr->port_num = id->port_num;
806 qp = ib_create_qp(pd, qp_init_attr);
810 if (id->qp_type == IB_QPT_UD)
811 ret = cma_init_ud_qp(id_priv, qp);
813 ret = cma_init_conn_qp(id_priv, qp);
818 id_priv->qp_num = qp->qp_num;
819 id_priv->srq = (qp->srq != NULL);
825 EXPORT_SYMBOL(rdma_create_qp);
827 void rdma_destroy_qp(struct rdma_cm_id *id)
829 struct rdma_id_private *id_priv;
831 id_priv = container_of(id, struct rdma_id_private, id);
832 mutex_lock(&id_priv->qp_mutex);
833 ib_destroy_qp(id_priv->id.qp);
834 id_priv->id.qp = NULL;
835 mutex_unlock(&id_priv->qp_mutex);
837 EXPORT_SYMBOL(rdma_destroy_qp);
839 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
840 struct rdma_conn_param *conn_param)
842 struct ib_qp_attr qp_attr;
843 int qp_attr_mask, ret;
846 mutex_lock(&id_priv->qp_mutex);
847 if (!id_priv->id.qp) {
852 /* Need to update QP attributes from default values. */
853 qp_attr.qp_state = IB_QPS_INIT;
854 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
858 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
862 qp_attr.qp_state = IB_QPS_RTR;
863 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
867 ret = ib_query_gid(id_priv->id.device, id_priv->id.port_num,
868 qp_attr.ah_attr.grh.sgid_index, &sgid, NULL);
872 BUG_ON(id_priv->cma_dev->device != id_priv->id.device);
875 qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
876 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
878 mutex_unlock(&id_priv->qp_mutex);
882 static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
883 struct rdma_conn_param *conn_param)
885 struct ib_qp_attr qp_attr;
886 int qp_attr_mask, ret;
888 mutex_lock(&id_priv->qp_mutex);
889 if (!id_priv->id.qp) {
894 qp_attr.qp_state = IB_QPS_RTS;
895 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
900 qp_attr.max_rd_atomic = conn_param->initiator_depth;
901 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
903 mutex_unlock(&id_priv->qp_mutex);
907 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
909 struct ib_qp_attr qp_attr;
912 mutex_lock(&id_priv->qp_mutex);
913 if (!id_priv->id.qp) {
918 qp_attr.qp_state = IB_QPS_ERR;
919 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
921 mutex_unlock(&id_priv->qp_mutex);
925 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
926 struct ib_qp_attr *qp_attr, int *qp_attr_mask)
928 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
932 if (rdma_cap_eth_ah(id_priv->id.device, id_priv->id.port_num))
935 pkey = ib_addr_get_pkey(dev_addr);
937 ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
938 pkey, &qp_attr->pkey_index);
942 qp_attr->port_num = id_priv->id.port_num;
943 *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
945 if (id_priv->id.qp_type == IB_QPT_UD) {
946 ret = cma_set_qkey(id_priv, 0);
950 qp_attr->qkey = id_priv->qkey;
951 *qp_attr_mask |= IB_QP_QKEY;
953 qp_attr->qp_access_flags = 0;
954 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
959 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
962 struct rdma_id_private *id_priv;
965 id_priv = container_of(id, struct rdma_id_private, id);
966 if (rdma_cap_ib_cm(id->device, id->port_num)) {
967 if (!id_priv->cm_id.ib || (id_priv->id.qp_type == IB_QPT_UD))
968 ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
970 ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
973 if (qp_attr->qp_state == IB_QPS_RTR)
974 qp_attr->rq_psn = id_priv->seq_num;
975 } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
976 if (!id_priv->cm_id.iw) {
977 qp_attr->qp_access_flags = 0;
978 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
980 ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
982 qp_attr->port_num = id_priv->id.port_num;
983 *qp_attr_mask |= IB_QP_PORT;
989 EXPORT_SYMBOL(rdma_init_qp_attr);
991 static inline int cma_zero_addr(struct sockaddr *addr)
993 switch (addr->sa_family) {
995 return ipv4_is_zeronet(((struct sockaddr_in *)addr)->sin_addr.s_addr);
997 return ipv6_addr_any(&((struct sockaddr_in6 *) addr)->sin6_addr);
999 return ib_addr_any(&((struct sockaddr_ib *) addr)->sib_addr);
1005 static inline int cma_loopback_addr(struct sockaddr *addr)
1007 switch (addr->sa_family) {
1009 return ipv4_is_loopback(((struct sockaddr_in *) addr)->sin_addr.s_addr);
1011 return ipv6_addr_loopback(&((struct sockaddr_in6 *) addr)->sin6_addr);
1013 return ib_addr_loopback(&((struct sockaddr_ib *) addr)->sib_addr);
1019 static inline int cma_any_addr(struct sockaddr *addr)
1021 return cma_zero_addr(addr) || cma_loopback_addr(addr);
1024 static int cma_addr_cmp(struct sockaddr *src, struct sockaddr *dst)
1026 if (src->sa_family != dst->sa_family)
1029 switch (src->sa_family) {
1031 return ((struct sockaddr_in *) src)->sin_addr.s_addr !=
1032 ((struct sockaddr_in *) dst)->sin_addr.s_addr;
1034 return ipv6_addr_cmp(&((struct sockaddr_in6 *) src)->sin6_addr,
1035 &((struct sockaddr_in6 *) dst)->sin6_addr);
1037 return ib_addr_cmp(&((struct sockaddr_ib *) src)->sib_addr,
1038 &((struct sockaddr_ib *) dst)->sib_addr);
1042 static __be16 cma_port(struct sockaddr *addr)
1044 struct sockaddr_ib *sib;
1046 switch (addr->sa_family) {
1048 return ((struct sockaddr_in *) addr)->sin_port;
1050 return ((struct sockaddr_in6 *) addr)->sin6_port;
1052 sib = (struct sockaddr_ib *) addr;
1053 return htons((u16) (be64_to_cpu(sib->sib_sid) &
1054 be64_to_cpu(sib->sib_sid_mask)));
1060 static inline int cma_any_port(struct sockaddr *addr)
1062 return !cma_port(addr);
1065 static void cma_save_ib_info(struct sockaddr *src_addr,
1066 struct sockaddr *dst_addr,
1067 struct rdma_cm_id *listen_id,
1068 struct ib_sa_path_rec *path)
1070 struct sockaddr_ib *listen_ib, *ib;
1072 listen_ib = (struct sockaddr_ib *) &listen_id->route.addr.src_addr;
1074 ib = (struct sockaddr_ib *)src_addr;
1075 ib->sib_family = AF_IB;
1077 ib->sib_pkey = path->pkey;
1078 ib->sib_flowinfo = path->flow_label;
1079 memcpy(&ib->sib_addr, &path->sgid, 16);
1080 ib->sib_sid = path->service_id;
1081 ib->sib_scope_id = 0;
1083 ib->sib_pkey = listen_ib->sib_pkey;
1084 ib->sib_flowinfo = listen_ib->sib_flowinfo;
1085 ib->sib_addr = listen_ib->sib_addr;
1086 ib->sib_sid = listen_ib->sib_sid;
1087 ib->sib_scope_id = listen_ib->sib_scope_id;
1089 ib->sib_sid_mask = cpu_to_be64(0xffffffffffffffffULL);
1092 ib = (struct sockaddr_ib *)dst_addr;
1093 ib->sib_family = AF_IB;
1095 ib->sib_pkey = path->pkey;
1096 ib->sib_flowinfo = path->flow_label;
1097 memcpy(&ib->sib_addr, &path->dgid, 16);
1102 static void cma_save_ip4_info(struct sockaddr_in *src_addr,
1103 struct sockaddr_in *dst_addr,
1104 struct cma_hdr *hdr,
1108 *src_addr = (struct sockaddr_in) {
1109 .sin_family = AF_INET,
1110 .sin_addr.s_addr = hdr->dst_addr.ip4.addr,
1111 .sin_port = local_port,
1116 *dst_addr = (struct sockaddr_in) {
1117 .sin_family = AF_INET,
1118 .sin_addr.s_addr = hdr->src_addr.ip4.addr,
1119 .sin_port = hdr->port,
1124 static void cma_save_ip6_info(struct sockaddr_in6 *src_addr,
1125 struct sockaddr_in6 *dst_addr,
1126 struct cma_hdr *hdr,
1130 *src_addr = (struct sockaddr_in6) {
1131 .sin6_family = AF_INET6,
1132 .sin6_addr = hdr->dst_addr.ip6,
1133 .sin6_port = local_port,
1138 *dst_addr = (struct sockaddr_in6) {
1139 .sin6_family = AF_INET6,
1140 .sin6_addr = hdr->src_addr.ip6,
1141 .sin6_port = hdr->port,
1146 static u16 cma_port_from_service_id(__be64 service_id)
1148 return (u16)be64_to_cpu(service_id);
1151 static int cma_save_ip_info(struct sockaddr *src_addr,
1152 struct sockaddr *dst_addr,
1153 struct ib_cm_event *ib_event,
1156 struct cma_hdr *hdr;
1159 hdr = ib_event->private_data;
1160 if (hdr->cma_version != CMA_VERSION)
1163 port = htons(cma_port_from_service_id(service_id));
1165 switch (cma_get_ip_ver(hdr)) {
1167 cma_save_ip4_info((struct sockaddr_in *)src_addr,
1168 (struct sockaddr_in *)dst_addr, hdr, port);
1171 cma_save_ip6_info((struct sockaddr_in6 *)src_addr,
1172 (struct sockaddr_in6 *)dst_addr, hdr, port);
1175 return -EAFNOSUPPORT;
1181 static int cma_save_net_info(struct sockaddr *src_addr,
1182 struct sockaddr *dst_addr,
1183 struct rdma_cm_id *listen_id,
1184 struct ib_cm_event *ib_event,
1185 sa_family_t sa_family, __be64 service_id)
1187 if (sa_family == AF_IB) {
1188 if (ib_event->event == IB_CM_REQ_RECEIVED)
1189 cma_save_ib_info(src_addr, dst_addr, listen_id,
1190 ib_event->param.req_rcvd.primary_path);
1191 else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED)
1192 cma_save_ib_info(src_addr, dst_addr, listen_id, NULL);
1196 return cma_save_ip_info(src_addr, dst_addr, ib_event, service_id);
1199 static int cma_save_req_info(const struct ib_cm_event *ib_event,
1200 struct cma_req_info *req)
1202 const struct ib_cm_req_event_param *req_param =
1203 &ib_event->param.req_rcvd;
1204 const struct ib_cm_sidr_req_event_param *sidr_param =
1205 &ib_event->param.sidr_req_rcvd;
1207 switch (ib_event->event) {
1208 case IB_CM_REQ_RECEIVED:
1209 req->device = req_param->listen_id->device;
1210 req->port = req_param->port;
1211 memcpy(&req->local_gid, &req_param->primary_path->sgid,
1212 sizeof(req->local_gid));
1213 req->has_gid = true;
1214 req->service_id = req_param->primary_path->service_id;
1215 req->pkey = be16_to_cpu(req_param->primary_path->pkey);
1216 if (req->pkey != req_param->bth_pkey)
1217 pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and primary path P_Key (0x%x)\n"
1218 "RDMA CMA: in the future this may cause the request to be dropped\n",
1219 req_param->bth_pkey, req->pkey);
1221 case IB_CM_SIDR_REQ_RECEIVED:
1222 req->device = sidr_param->listen_id->device;
1223 req->port = sidr_param->port;
1224 req->has_gid = false;
1225 req->service_id = sidr_param->service_id;
1226 req->pkey = sidr_param->pkey;
1227 if (req->pkey != sidr_param->bth_pkey)
1228 pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and SIDR request payload P_Key (0x%x)\n"
1229 "RDMA CMA: in the future this may cause the request to be dropped\n",
1230 sidr_param->bth_pkey, req->pkey);
1239 static bool validate_ipv4_net_dev(struct net_device *net_dev,
1240 const struct sockaddr_in *dst_addr,
1241 const struct sockaddr_in *src_addr)
1243 __be32 daddr = dst_addr->sin_addr.s_addr,
1244 saddr = src_addr->sin_addr.s_addr;
1245 struct fib_result res;
1250 if (ipv4_is_multicast(saddr) || ipv4_is_lbcast(saddr) ||
1251 ipv4_is_lbcast(daddr) || ipv4_is_zeronet(saddr) ||
1252 ipv4_is_zeronet(daddr) || ipv4_is_loopback(daddr) ||
1253 ipv4_is_loopback(saddr))
1256 memset(&fl4, 0, sizeof(fl4));
1257 fl4.flowi4_iif = net_dev->ifindex;
1262 err = fib_lookup(dev_net(net_dev), &fl4, &res, 0);
1263 ret = err == 0 && FIB_RES_DEV(res) == net_dev;
1269 static bool validate_ipv6_net_dev(struct net_device *net_dev,
1270 const struct sockaddr_in6 *dst_addr,
1271 const struct sockaddr_in6 *src_addr)
1273 #if IS_ENABLED(CONFIG_IPV6)
1274 const int strict = ipv6_addr_type(&dst_addr->sin6_addr) &
1275 IPV6_ADDR_LINKLOCAL;
1276 struct rt6_info *rt = rt6_lookup(dev_net(net_dev), &dst_addr->sin6_addr,
1277 &src_addr->sin6_addr, net_dev->ifindex,
1284 ret = rt->rt6i_idev->dev == net_dev;
1293 static bool validate_net_dev(struct net_device *net_dev,
1294 const struct sockaddr *daddr,
1295 const struct sockaddr *saddr)
1297 const struct sockaddr_in *daddr4 = (const struct sockaddr_in *)daddr;
1298 const struct sockaddr_in *saddr4 = (const struct sockaddr_in *)saddr;
1299 const struct sockaddr_in6 *daddr6 = (const struct sockaddr_in6 *)daddr;
1300 const struct sockaddr_in6 *saddr6 = (const struct sockaddr_in6 *)saddr;
1302 switch (daddr->sa_family) {
1304 return saddr->sa_family == AF_INET &&
1305 validate_ipv4_net_dev(net_dev, daddr4, saddr4);
1308 return saddr->sa_family == AF_INET6 &&
1309 validate_ipv6_net_dev(net_dev, daddr6, saddr6);
1316 static struct net_device *cma_get_net_dev(struct ib_cm_event *ib_event,
1317 const struct cma_req_info *req)
1319 struct sockaddr_storage listen_addr_storage, src_addr_storage;
1320 struct sockaddr *listen_addr = (struct sockaddr *)&listen_addr_storage,
1321 *src_addr = (struct sockaddr *)&src_addr_storage;
1322 struct net_device *net_dev;
1323 const union ib_gid *gid = req->has_gid ? &req->local_gid : NULL;
1326 err = cma_save_ip_info(listen_addr, src_addr, ib_event,
1329 return ERR_PTR(err);
1331 net_dev = ib_get_net_dev_by_params(req->device, req->port, req->pkey,
1334 return ERR_PTR(-ENODEV);
1336 if (!validate_net_dev(net_dev, listen_addr, src_addr)) {
1338 return ERR_PTR(-EHOSTUNREACH);
1344 static enum rdma_port_space rdma_ps_from_service_id(__be64 service_id)
1346 return (be64_to_cpu(service_id) >> 16) & 0xffff;
1349 static bool cma_match_private_data(struct rdma_id_private *id_priv,
1350 const struct cma_hdr *hdr)
1352 struct sockaddr *addr = cma_src_addr(id_priv);
1354 struct in6_addr ip6_addr;
1356 if (cma_any_addr(addr) && !id_priv->afonly)
1359 switch (addr->sa_family) {
1361 ip4_addr = ((struct sockaddr_in *)addr)->sin_addr.s_addr;
1362 if (cma_get_ip_ver(hdr) != 4)
1364 if (!cma_any_addr(addr) &&
1365 hdr->dst_addr.ip4.addr != ip4_addr)
1369 ip6_addr = ((struct sockaddr_in6 *)addr)->sin6_addr;
1370 if (cma_get_ip_ver(hdr) != 6)
1372 if (!cma_any_addr(addr) &&
1373 memcmp(&hdr->dst_addr.ip6, &ip6_addr, sizeof(ip6_addr)))
1385 static bool cma_protocol_roce_dev_port(struct ib_device *device, int port_num)
1387 enum rdma_link_layer ll = rdma_port_get_link_layer(device, port_num);
1388 enum rdma_transport_type transport =
1389 rdma_node_get_transport(device->node_type);
1391 return ll == IB_LINK_LAYER_ETHERNET && transport == RDMA_TRANSPORT_IB;
1394 static bool cma_protocol_roce(const struct rdma_cm_id *id)
1396 struct ib_device *device = id->device;
1397 const int port_num = id->port_num ?: rdma_start_port(device);
1399 return cma_protocol_roce_dev_port(device, port_num);
1402 static bool cma_match_net_dev(const struct rdma_cm_id *id,
1403 const struct net_device *net_dev,
1406 const struct rdma_addr *addr = &id->route.addr;
1409 /* This request is an AF_IB request or a RoCE request */
1410 return (!id->port_num || id->port_num == port_num) &&
1411 (addr->src_addr.ss_family == AF_IB ||
1412 cma_protocol_roce_dev_port(id->device, port_num));
1415 * Net namespaces must match, and if the listner is listening
1416 * on a specific netdevice than netdevice must match as well.
1418 if (net_eq(dev_net(net_dev), addr->dev_addr.net) &&
1419 (!!addr->dev_addr.bound_dev_if ==
1420 (addr->dev_addr.bound_dev_if == net_dev->ifindex)))
1426 static struct rdma_id_private *cma_find_listener(
1427 const struct rdma_bind_list *bind_list,
1428 const struct ib_cm_id *cm_id,
1429 const struct ib_cm_event *ib_event,
1430 const struct cma_req_info *req,
1431 const struct net_device *net_dev)
1433 struct rdma_id_private *id_priv, *id_priv_dev;
1436 return ERR_PTR(-EINVAL);
1438 hlist_for_each_entry(id_priv, &bind_list->owners, node) {
1439 if (cma_match_private_data(id_priv, ib_event->private_data)) {
1440 if (id_priv->id.device == cm_id->device &&
1441 cma_match_net_dev(&id_priv->id, net_dev, req->port))
1443 list_for_each_entry(id_priv_dev,
1444 &id_priv->listen_list,
1446 if (id_priv_dev->id.device == cm_id->device &&
1447 cma_match_net_dev(&id_priv_dev->id, net_dev, req->port))
1453 return ERR_PTR(-EINVAL);
1456 static struct rdma_id_private *cma_id_from_event(struct ib_cm_id *cm_id,
1457 struct ib_cm_event *ib_event,
1458 struct net_device **net_dev)
1460 struct cma_req_info req;
1461 struct rdma_bind_list *bind_list;
1462 struct rdma_id_private *id_priv;
1465 err = cma_save_req_info(ib_event, &req);
1467 return ERR_PTR(err);
1469 *net_dev = cma_get_net_dev(ib_event, &req);
1470 if (IS_ERR(*net_dev)) {
1471 if (PTR_ERR(*net_dev) == -EAFNOSUPPORT) {
1472 /* Assuming the protocol is AF_IB */
1474 } else if (cma_protocol_roce_dev_port(req.device, req.port)) {
1475 /* TODO find the net dev matching the request parameters
1476 * through the RoCE GID table */
1479 return ERR_CAST(*net_dev);
1483 bind_list = cma_ps_find(*net_dev ? dev_net(*net_dev) : &init_net,
1484 rdma_ps_from_service_id(req.service_id),
1485 cma_port_from_service_id(req.service_id));
1486 id_priv = cma_find_listener(bind_list, cm_id, ib_event, &req, *net_dev);
1487 if (IS_ERR(id_priv) && *net_dev) {
1495 static inline u8 cma_user_data_offset(struct rdma_id_private *id_priv)
1497 return cma_family(id_priv) == AF_IB ? 0 : sizeof(struct cma_hdr);
1500 static void cma_cancel_route(struct rdma_id_private *id_priv)
1502 if (rdma_cap_ib_sa(id_priv->id.device, id_priv->id.port_num)) {
1504 ib_sa_cancel_query(id_priv->query_id, id_priv->query);
1508 static void cma_cancel_listens(struct rdma_id_private *id_priv)
1510 struct rdma_id_private *dev_id_priv;
1513 * Remove from listen_any_list to prevent added devices from spawning
1514 * additional listen requests.
1517 list_del(&id_priv->list);
1519 while (!list_empty(&id_priv->listen_list)) {
1520 dev_id_priv = list_entry(id_priv->listen_list.next,
1521 struct rdma_id_private, listen_list);
1522 /* sync with device removal to avoid duplicate destruction */
1523 list_del_init(&dev_id_priv->list);
1524 list_del(&dev_id_priv->listen_list);
1525 mutex_unlock(&lock);
1527 rdma_destroy_id(&dev_id_priv->id);
1530 mutex_unlock(&lock);
1533 static void cma_cancel_operation(struct rdma_id_private *id_priv,
1534 enum rdma_cm_state state)
1537 case RDMA_CM_ADDR_QUERY:
1538 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
1540 case RDMA_CM_ROUTE_QUERY:
1541 cma_cancel_route(id_priv);
1543 case RDMA_CM_LISTEN:
1544 if (cma_any_addr(cma_src_addr(id_priv)) && !id_priv->cma_dev)
1545 cma_cancel_listens(id_priv);
1552 static void cma_release_port(struct rdma_id_private *id_priv)
1554 struct rdma_bind_list *bind_list = id_priv->bind_list;
1555 struct net *net = id_priv->id.route.addr.dev_addr.net;
1561 hlist_del(&id_priv->node);
1562 if (hlist_empty(&bind_list->owners)) {
1563 cma_ps_remove(net, bind_list->ps, bind_list->port);
1566 mutex_unlock(&lock);
1569 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
1571 struct cma_multicast *mc;
1573 while (!list_empty(&id_priv->mc_list)) {
1574 mc = container_of(id_priv->mc_list.next,
1575 struct cma_multicast, list);
1576 list_del(&mc->list);
1577 if (rdma_cap_ib_mcast(id_priv->cma_dev->device,
1578 id_priv->id.port_num)) {
1579 ib_sa_free_multicast(mc->multicast.ib);
1582 if (mc->igmp_joined) {
1583 struct rdma_dev_addr *dev_addr =
1584 &id_priv->id.route.addr.dev_addr;
1585 struct net_device *ndev = NULL;
1587 if (dev_addr->bound_dev_if)
1588 ndev = dev_get_by_index(&init_net,
1589 dev_addr->bound_dev_if);
1592 &mc->multicast.ib->rec.mgid,
1597 kref_put(&mc->mcref, release_mc);
1602 void rdma_destroy_id(struct rdma_cm_id *id)
1604 struct rdma_id_private *id_priv;
1605 enum rdma_cm_state state;
1607 id_priv = container_of(id, struct rdma_id_private, id);
1608 state = cma_exch(id_priv, RDMA_CM_DESTROYING);
1609 cma_cancel_operation(id_priv, state);
1612 * Wait for any active callback to finish. New callbacks will find
1613 * the id_priv state set to destroying and abort.
1615 mutex_lock(&id_priv->handler_mutex);
1616 mutex_unlock(&id_priv->handler_mutex);
1618 if (id_priv->cma_dev) {
1619 if (rdma_cap_ib_cm(id_priv->id.device, 1)) {
1620 if (id_priv->cm_id.ib)
1621 ib_destroy_cm_id(id_priv->cm_id.ib);
1622 } else if (rdma_cap_iw_cm(id_priv->id.device, 1)) {
1623 if (id_priv->cm_id.iw)
1624 iw_destroy_cm_id(id_priv->cm_id.iw);
1626 cma_leave_mc_groups(id_priv);
1627 cma_release_dev(id_priv);
1630 cma_release_port(id_priv);
1631 cma_deref_id(id_priv);
1632 wait_for_completion(&id_priv->comp);
1634 if (id_priv->internal_id)
1635 cma_deref_id(id_priv->id.context);
1637 kfree(id_priv->id.route.path_rec);
1638 put_net(id_priv->id.route.addr.dev_addr.net);
1641 EXPORT_SYMBOL(rdma_destroy_id);
1643 static int cma_rep_recv(struct rdma_id_private *id_priv)
1647 ret = cma_modify_qp_rtr(id_priv, NULL);
1651 ret = cma_modify_qp_rts(id_priv, NULL);
1655 ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
1661 cma_modify_qp_err(id_priv);
1662 ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
1667 static void cma_set_rep_event_data(struct rdma_cm_event *event,
1668 struct ib_cm_rep_event_param *rep_data,
1671 event->param.conn.private_data = private_data;
1672 event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
1673 event->param.conn.responder_resources = rep_data->responder_resources;
1674 event->param.conn.initiator_depth = rep_data->initiator_depth;
1675 event->param.conn.flow_control = rep_data->flow_control;
1676 event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
1677 event->param.conn.srq = rep_data->srq;
1678 event->param.conn.qp_num = rep_data->remote_qpn;
1681 static int cma_ib_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1683 struct rdma_id_private *id_priv = cm_id->context;
1684 struct rdma_cm_event event;
1687 mutex_lock(&id_priv->handler_mutex);
1688 if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
1689 id_priv->state != RDMA_CM_CONNECT) ||
1690 (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
1691 id_priv->state != RDMA_CM_DISCONNECT))
1694 memset(&event, 0, sizeof event);
1695 switch (ib_event->event) {
1696 case IB_CM_REQ_ERROR:
1697 case IB_CM_REP_ERROR:
1698 event.event = RDMA_CM_EVENT_UNREACHABLE;
1699 event.status = -ETIMEDOUT;
1701 case IB_CM_REP_RECEIVED:
1702 if (id_priv->id.qp) {
1703 event.status = cma_rep_recv(id_priv);
1704 event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
1705 RDMA_CM_EVENT_ESTABLISHED;
1707 event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
1709 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
1710 ib_event->private_data);
1712 case IB_CM_RTU_RECEIVED:
1713 case IB_CM_USER_ESTABLISHED:
1714 event.event = RDMA_CM_EVENT_ESTABLISHED;
1716 case IB_CM_DREQ_ERROR:
1717 event.status = -ETIMEDOUT; /* fall through */
1718 case IB_CM_DREQ_RECEIVED:
1719 case IB_CM_DREP_RECEIVED:
1720 if (!cma_comp_exch(id_priv, RDMA_CM_CONNECT,
1721 RDMA_CM_DISCONNECT))
1723 event.event = RDMA_CM_EVENT_DISCONNECTED;
1725 case IB_CM_TIMEWAIT_EXIT:
1726 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
1728 case IB_CM_MRA_RECEIVED:
1731 case IB_CM_REJ_RECEIVED:
1732 cma_modify_qp_err(id_priv);
1733 event.status = ib_event->param.rej_rcvd.reason;
1734 event.event = RDMA_CM_EVENT_REJECTED;
1735 event.param.conn.private_data = ib_event->private_data;
1736 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
1739 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
1744 ret = id_priv->id.event_handler(&id_priv->id, &event);
1746 /* Destroy the CM ID by returning a non-zero value. */
1747 id_priv->cm_id.ib = NULL;
1748 cma_exch(id_priv, RDMA_CM_DESTROYING);
1749 mutex_unlock(&id_priv->handler_mutex);
1750 rdma_destroy_id(&id_priv->id);
1754 mutex_unlock(&id_priv->handler_mutex);
1758 static struct rdma_id_private *cma_new_conn_id(struct rdma_cm_id *listen_id,
1759 struct ib_cm_event *ib_event,
1760 struct net_device *net_dev)
1762 struct rdma_id_private *id_priv;
1763 struct rdma_cm_id *id;
1764 struct rdma_route *rt;
1765 const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
1766 const __be64 service_id =
1767 ib_event->param.req_rcvd.primary_path->service_id;
1770 id = rdma_create_id(listen_id->route.addr.dev_addr.net,
1771 listen_id->event_handler, listen_id->context,
1772 listen_id->ps, ib_event->param.req_rcvd.qp_type);
1776 id_priv = container_of(id, struct rdma_id_private, id);
1777 if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
1778 (struct sockaddr *)&id->route.addr.dst_addr,
1779 listen_id, ib_event, ss_family, service_id))
1783 rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
1784 rt->path_rec = kmalloc(sizeof *rt->path_rec * rt->num_paths,
1789 rt->path_rec[0] = *ib_event->param.req_rcvd.primary_path;
1790 if (rt->num_paths == 2)
1791 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
1794 ret = rdma_copy_addr(&rt->addr.dev_addr, net_dev, NULL);
1798 if (!cma_protocol_roce(listen_id) &&
1799 cma_any_addr(cma_src_addr(id_priv))) {
1800 rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
1801 rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
1802 ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
1803 } else if (!cma_any_addr(cma_src_addr(id_priv))) {
1804 ret = cma_translate_addr(cma_src_addr(id_priv), &rt->addr.dev_addr);
1809 rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
1811 id_priv->state = RDMA_CM_CONNECT;
1815 rdma_destroy_id(id);
1819 static struct rdma_id_private *cma_new_udp_id(struct rdma_cm_id *listen_id,
1820 struct ib_cm_event *ib_event,
1821 struct net_device *net_dev)
1823 struct rdma_id_private *id_priv;
1824 struct rdma_cm_id *id;
1825 const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
1826 struct net *net = listen_id->route.addr.dev_addr.net;
1829 id = rdma_create_id(net, listen_id->event_handler, listen_id->context,
1830 listen_id->ps, IB_QPT_UD);
1834 id_priv = container_of(id, struct rdma_id_private, id);
1835 if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
1836 (struct sockaddr *)&id->route.addr.dst_addr,
1837 listen_id, ib_event, ss_family,
1838 ib_event->param.sidr_req_rcvd.service_id))
1842 ret = rdma_copy_addr(&id->route.addr.dev_addr, net_dev, NULL);
1846 if (!cma_any_addr(cma_src_addr(id_priv))) {
1847 ret = cma_translate_addr(cma_src_addr(id_priv),
1848 &id->route.addr.dev_addr);
1854 id_priv->state = RDMA_CM_CONNECT;
1857 rdma_destroy_id(id);
1861 static void cma_set_req_event_data(struct rdma_cm_event *event,
1862 struct ib_cm_req_event_param *req_data,
1863 void *private_data, int offset)
1865 event->param.conn.private_data = private_data + offset;
1866 event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
1867 event->param.conn.responder_resources = req_data->responder_resources;
1868 event->param.conn.initiator_depth = req_data->initiator_depth;
1869 event->param.conn.flow_control = req_data->flow_control;
1870 event->param.conn.retry_count = req_data->retry_count;
1871 event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
1872 event->param.conn.srq = req_data->srq;
1873 event->param.conn.qp_num = req_data->remote_qpn;
1876 static int cma_check_req_qp_type(struct rdma_cm_id *id, struct ib_cm_event *ib_event)
1878 return (((ib_event->event == IB_CM_REQ_RECEIVED) &&
1879 (ib_event->param.req_rcvd.qp_type == id->qp_type)) ||
1880 ((ib_event->event == IB_CM_SIDR_REQ_RECEIVED) &&
1881 (id->qp_type == IB_QPT_UD)) ||
1885 static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1887 struct rdma_id_private *listen_id, *conn_id = NULL;
1888 struct rdma_cm_event event;
1889 struct net_device *net_dev;
1893 listen_id = cma_id_from_event(cm_id, ib_event, &net_dev);
1894 if (IS_ERR(listen_id))
1895 return PTR_ERR(listen_id);
1897 if (!cma_check_req_qp_type(&listen_id->id, ib_event)) {
1902 mutex_lock(&listen_id->handler_mutex);
1903 if (listen_id->state != RDMA_CM_LISTEN) {
1904 ret = -ECONNABORTED;
1908 memset(&event, 0, sizeof event);
1909 offset = cma_user_data_offset(listen_id);
1910 event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1911 if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) {
1912 conn_id = cma_new_udp_id(&listen_id->id, ib_event, net_dev);
1913 event.param.ud.private_data = ib_event->private_data + offset;
1914 event.param.ud.private_data_len =
1915 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
1917 conn_id = cma_new_conn_id(&listen_id->id, ib_event, net_dev);
1918 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
1919 ib_event->private_data, offset);
1926 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1927 ret = cma_acquire_dev(conn_id, listen_id);
1931 conn_id->cm_id.ib = cm_id;
1932 cm_id->context = conn_id;
1933 cm_id->cm_handler = cma_ib_handler;
1936 * Protect against the user destroying conn_id from another thread
1937 * until we're done accessing it.
1939 atomic_inc(&conn_id->refcount);
1940 ret = conn_id->id.event_handler(&conn_id->id, &event);
1944 * Acquire mutex to prevent user executing rdma_destroy_id()
1945 * while we're accessing the cm_id.
1948 if (cma_comp(conn_id, RDMA_CM_CONNECT) &&
1949 (conn_id->id.qp_type != IB_QPT_UD))
1950 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
1951 mutex_unlock(&lock);
1952 mutex_unlock(&conn_id->handler_mutex);
1953 mutex_unlock(&listen_id->handler_mutex);
1954 cma_deref_id(conn_id);
1960 cma_deref_id(conn_id);
1961 /* Destroy the CM ID by returning a non-zero value. */
1962 conn_id->cm_id.ib = NULL;
1964 cma_exch(conn_id, RDMA_CM_DESTROYING);
1965 mutex_unlock(&conn_id->handler_mutex);
1967 mutex_unlock(&listen_id->handler_mutex);
1969 rdma_destroy_id(&conn_id->id);
1978 __be64 rdma_get_service_id(struct rdma_cm_id *id, struct sockaddr *addr)
1980 if (addr->sa_family == AF_IB)
1981 return ((struct sockaddr_ib *) addr)->sib_sid;
1983 return cpu_to_be64(((u64)id->ps << 16) + be16_to_cpu(cma_port(addr)));
1985 EXPORT_SYMBOL(rdma_get_service_id);
1987 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
1989 struct rdma_id_private *id_priv = iw_id->context;
1990 struct rdma_cm_event event;
1992 struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
1993 struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
1995 mutex_lock(&id_priv->handler_mutex);
1996 if (id_priv->state != RDMA_CM_CONNECT)
1999 memset(&event, 0, sizeof event);
2000 switch (iw_event->event) {
2001 case IW_CM_EVENT_CLOSE:
2002 event.event = RDMA_CM_EVENT_DISCONNECTED;
2004 case IW_CM_EVENT_CONNECT_REPLY:
2005 memcpy(cma_src_addr(id_priv), laddr,
2006 rdma_addr_size(laddr));
2007 memcpy(cma_dst_addr(id_priv), raddr,
2008 rdma_addr_size(raddr));
2009 switch (iw_event->status) {
2011 event.event = RDMA_CM_EVENT_ESTABLISHED;
2012 event.param.conn.initiator_depth = iw_event->ird;
2013 event.param.conn.responder_resources = iw_event->ord;
2017 event.event = RDMA_CM_EVENT_REJECTED;
2020 event.event = RDMA_CM_EVENT_UNREACHABLE;
2023 event.event = RDMA_CM_EVENT_CONNECT_ERROR;
2027 case IW_CM_EVENT_ESTABLISHED:
2028 event.event = RDMA_CM_EVENT_ESTABLISHED;
2029 event.param.conn.initiator_depth = iw_event->ird;
2030 event.param.conn.responder_resources = iw_event->ord;
2036 event.status = iw_event->status;
2037 event.param.conn.private_data = iw_event->private_data;
2038 event.param.conn.private_data_len = iw_event->private_data_len;
2039 ret = id_priv->id.event_handler(&id_priv->id, &event);
2041 /* Destroy the CM ID by returning a non-zero value. */
2042 id_priv->cm_id.iw = NULL;
2043 cma_exch(id_priv, RDMA_CM_DESTROYING);
2044 mutex_unlock(&id_priv->handler_mutex);
2045 rdma_destroy_id(&id_priv->id);
2050 mutex_unlock(&id_priv->handler_mutex);
2054 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
2055 struct iw_cm_event *iw_event)
2057 struct rdma_cm_id *new_cm_id;
2058 struct rdma_id_private *listen_id, *conn_id;
2059 struct rdma_cm_event event;
2060 int ret = -ECONNABORTED;
2061 struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
2062 struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
2064 listen_id = cm_id->context;
2066 mutex_lock(&listen_id->handler_mutex);
2067 if (listen_id->state != RDMA_CM_LISTEN)
2070 /* Create a new RDMA id for the new IW CM ID */
2071 new_cm_id = rdma_create_id(listen_id->id.route.addr.dev_addr.net,
2072 listen_id->id.event_handler,
2073 listen_id->id.context,
2074 RDMA_PS_TCP, IB_QPT_RC);
2075 if (IS_ERR(new_cm_id)) {
2079 conn_id = container_of(new_cm_id, struct rdma_id_private, id);
2080 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
2081 conn_id->state = RDMA_CM_CONNECT;
2083 ret = rdma_translate_ip(laddr, &conn_id->id.route.addr.dev_addr, NULL);
2085 mutex_unlock(&conn_id->handler_mutex);
2086 rdma_destroy_id(new_cm_id);
2090 ret = cma_acquire_dev(conn_id, listen_id);
2092 mutex_unlock(&conn_id->handler_mutex);
2093 rdma_destroy_id(new_cm_id);
2097 conn_id->cm_id.iw = cm_id;
2098 cm_id->context = conn_id;
2099 cm_id->cm_handler = cma_iw_handler;
2101 memcpy(cma_src_addr(conn_id), laddr, rdma_addr_size(laddr));
2102 memcpy(cma_dst_addr(conn_id), raddr, rdma_addr_size(raddr));
2104 memset(&event, 0, sizeof event);
2105 event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
2106 event.param.conn.private_data = iw_event->private_data;
2107 event.param.conn.private_data_len = iw_event->private_data_len;
2108 event.param.conn.initiator_depth = iw_event->ird;
2109 event.param.conn.responder_resources = iw_event->ord;
2112 * Protect against the user destroying conn_id from another thread
2113 * until we're done accessing it.
2115 atomic_inc(&conn_id->refcount);
2116 ret = conn_id->id.event_handler(&conn_id->id, &event);
2118 /* User wants to destroy the CM ID */
2119 conn_id->cm_id.iw = NULL;
2120 cma_exch(conn_id, RDMA_CM_DESTROYING);
2121 mutex_unlock(&conn_id->handler_mutex);
2122 mutex_unlock(&listen_id->handler_mutex);
2123 cma_deref_id(conn_id);
2124 rdma_destroy_id(&conn_id->id);
2128 mutex_unlock(&conn_id->handler_mutex);
2129 cma_deref_id(conn_id);
2132 mutex_unlock(&listen_id->handler_mutex);
2136 static int cma_ib_listen(struct rdma_id_private *id_priv)
2138 struct sockaddr *addr;
2139 struct ib_cm_id *id;
2142 addr = cma_src_addr(id_priv);
2143 svc_id = rdma_get_service_id(&id_priv->id, addr);
2144 id = ib_cm_insert_listen(id_priv->id.device, cma_req_handler, svc_id);
2147 id_priv->cm_id.ib = id;
2152 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
2155 struct iw_cm_id *id;
2157 id = iw_create_cm_id(id_priv->id.device,
2158 iw_conn_req_handler,
2163 id->tos = id_priv->tos;
2164 id_priv->cm_id.iw = id;
2166 memcpy(&id_priv->cm_id.iw->local_addr, cma_src_addr(id_priv),
2167 rdma_addr_size(cma_src_addr(id_priv)));
2169 ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
2172 iw_destroy_cm_id(id_priv->cm_id.iw);
2173 id_priv->cm_id.iw = NULL;
2179 static int cma_listen_handler(struct rdma_cm_id *id,
2180 struct rdma_cm_event *event)
2182 struct rdma_id_private *id_priv = id->context;
2184 id->context = id_priv->id.context;
2185 id->event_handler = id_priv->id.event_handler;
2186 return id_priv->id.event_handler(id, event);
2189 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
2190 struct cma_device *cma_dev)
2192 struct rdma_id_private *dev_id_priv;
2193 struct rdma_cm_id *id;
2194 struct net *net = id_priv->id.route.addr.dev_addr.net;
2197 if (cma_family(id_priv) == AF_IB && !rdma_cap_ib_cm(cma_dev->device, 1))
2200 id = rdma_create_id(net, cma_listen_handler, id_priv, id_priv->id.ps,
2201 id_priv->id.qp_type);
2205 dev_id_priv = container_of(id, struct rdma_id_private, id);
2207 dev_id_priv->state = RDMA_CM_ADDR_BOUND;
2208 memcpy(cma_src_addr(dev_id_priv), cma_src_addr(id_priv),
2209 rdma_addr_size(cma_src_addr(id_priv)));
2211 _cma_attach_to_dev(dev_id_priv, cma_dev);
2212 list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
2213 atomic_inc(&id_priv->refcount);
2214 dev_id_priv->internal_id = 1;
2215 dev_id_priv->afonly = id_priv->afonly;
2217 ret = rdma_listen(id, id_priv->backlog);
2219 pr_warn("RDMA CMA: cma_listen_on_dev, error %d, listening on device %s\n",
2220 ret, cma_dev->device->name);
2223 static void cma_listen_on_all(struct rdma_id_private *id_priv)
2225 struct cma_device *cma_dev;
2228 list_add_tail(&id_priv->list, &listen_any_list);
2229 list_for_each_entry(cma_dev, &dev_list, list)
2230 cma_listen_on_dev(id_priv, cma_dev);
2231 mutex_unlock(&lock);
2234 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
2236 struct rdma_id_private *id_priv;
2238 id_priv = container_of(id, struct rdma_id_private, id);
2239 id_priv->tos = (u8) tos;
2241 EXPORT_SYMBOL(rdma_set_service_type);
2243 static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec,
2246 struct cma_work *work = context;
2247 struct rdma_route *route;
2249 route = &work->id->id.route;
2252 route->num_paths = 1;
2253 *route->path_rec = *path_rec;
2255 work->old_state = RDMA_CM_ROUTE_QUERY;
2256 work->new_state = RDMA_CM_ADDR_RESOLVED;
2257 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
2258 work->event.status = status;
2261 queue_work(cma_wq, &work->work);
2264 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms,
2265 struct cma_work *work)
2267 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2268 struct ib_sa_path_rec path_rec;
2269 ib_sa_comp_mask comp_mask;
2270 struct sockaddr_in6 *sin6;
2271 struct sockaddr_ib *sib;
2273 memset(&path_rec, 0, sizeof path_rec);
2274 rdma_addr_get_sgid(dev_addr, &path_rec.sgid);
2275 rdma_addr_get_dgid(dev_addr, &path_rec.dgid);
2276 path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
2277 path_rec.numb_path = 1;
2278 path_rec.reversible = 1;
2279 path_rec.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
2281 comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
2282 IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
2283 IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
2285 switch (cma_family(id_priv)) {
2287 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
2288 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
2291 sin6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
2292 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
2293 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2296 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
2297 path_rec.traffic_class = (u8) (be32_to_cpu(sib->sib_flowinfo) >> 20);
2298 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2302 id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
2303 id_priv->id.port_num, &path_rec,
2304 comp_mask, timeout_ms,
2305 GFP_KERNEL, cma_query_handler,
2306 work, &id_priv->query);
2308 return (id_priv->query_id < 0) ? id_priv->query_id : 0;
2311 static void cma_work_handler(struct work_struct *_work)
2313 struct cma_work *work = container_of(_work, struct cma_work, work);
2314 struct rdma_id_private *id_priv = work->id;
2317 mutex_lock(&id_priv->handler_mutex);
2318 if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
2321 if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
2322 cma_exch(id_priv, RDMA_CM_DESTROYING);
2326 mutex_unlock(&id_priv->handler_mutex);
2327 cma_deref_id(id_priv);
2329 rdma_destroy_id(&id_priv->id);
2333 static void cma_ndev_work_handler(struct work_struct *_work)
2335 struct cma_ndev_work *work = container_of(_work, struct cma_ndev_work, work);
2336 struct rdma_id_private *id_priv = work->id;
2339 mutex_lock(&id_priv->handler_mutex);
2340 if (id_priv->state == RDMA_CM_DESTROYING ||
2341 id_priv->state == RDMA_CM_DEVICE_REMOVAL)
2344 if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
2345 cma_exch(id_priv, RDMA_CM_DESTROYING);
2350 mutex_unlock(&id_priv->handler_mutex);
2351 cma_deref_id(id_priv);
2353 rdma_destroy_id(&id_priv->id);
2357 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms)
2359 struct rdma_route *route = &id_priv->id.route;
2360 struct cma_work *work;
2363 work = kzalloc(sizeof *work, GFP_KERNEL);
2368 INIT_WORK(&work->work, cma_work_handler);
2369 work->old_state = RDMA_CM_ROUTE_QUERY;
2370 work->new_state = RDMA_CM_ROUTE_RESOLVED;
2371 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
2373 if (!route->path_rec)
2374 route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
2375 if (!route->path_rec) {
2380 ret = cma_query_ib_route(id_priv, timeout_ms, work);
2386 kfree(route->path_rec);
2387 route->path_rec = NULL;
2393 int rdma_set_ib_paths(struct rdma_cm_id *id,
2394 struct ib_sa_path_rec *path_rec, int num_paths)
2396 struct rdma_id_private *id_priv;
2399 id_priv = container_of(id, struct rdma_id_private, id);
2400 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
2401 RDMA_CM_ROUTE_RESOLVED))
2404 id->route.path_rec = kmemdup(path_rec, sizeof *path_rec * num_paths,
2406 if (!id->route.path_rec) {
2411 id->route.num_paths = num_paths;
2414 cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED);
2417 EXPORT_SYMBOL(rdma_set_ib_paths);
2419 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms)
2421 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;
2432 queue_work(cma_wq, &work->work);
2436 static int iboe_tos_to_sl(struct net_device *ndev, int tos)
2439 struct net_device *dev;
2441 prio = rt_tos2priority(tos);
2442 dev = ndev->priv_flags & IFF_802_1Q_VLAN ?
2443 vlan_dev_real_dev(ndev) : ndev;
2446 return netdev_get_prio_tc_map(dev, prio);
2448 #if IS_ENABLED(CONFIG_VLAN_8021Q)
2449 if (ndev->priv_flags & IFF_802_1Q_VLAN)
2450 return (vlan_dev_get_egress_qos_mask(ndev, prio) &
2451 VLAN_PRIO_MASK) >> VLAN_PRIO_SHIFT;
2456 static enum ib_gid_type cma_route_gid_type(enum rdma_network_type network_type,
2457 unsigned long supported_gids,
2458 enum ib_gid_type default_gid)
2460 if ((network_type == RDMA_NETWORK_IPV4 ||
2461 network_type == RDMA_NETWORK_IPV6) &&
2462 test_bit(IB_GID_TYPE_ROCE_UDP_ENCAP, &supported_gids))
2463 return IB_GID_TYPE_ROCE_UDP_ENCAP;
2468 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
2470 struct rdma_route *route = &id_priv->id.route;
2471 struct rdma_addr *addr = &route->addr;
2472 struct cma_work *work;
2474 struct net_device *ndev = NULL;
2477 work = kzalloc(sizeof *work, GFP_KERNEL);
2482 INIT_WORK(&work->work, cma_work_handler);
2484 route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
2485 if (!route->path_rec) {
2490 route->num_paths = 1;
2492 if (addr->dev_addr.bound_dev_if) {
2493 unsigned long supported_gids;
2495 ndev = dev_get_by_index(&init_net, addr->dev_addr.bound_dev_if);
2501 if (ndev->flags & IFF_LOOPBACK) {
2503 if (!id_priv->id.device->get_netdev) {
2508 ndev = id_priv->id.device->get_netdev(id_priv->id.device,
2509 id_priv->id.port_num);
2516 route->path_rec->net = &init_net;
2517 route->path_rec->ifindex = ndev->ifindex;
2518 supported_gids = roce_gid_type_mask_support(id_priv->id.device,
2519 id_priv->id.port_num);
2520 route->path_rec->gid_type =
2521 cma_route_gid_type(addr->dev_addr.network,
2530 memcpy(route->path_rec->dmac, addr->dev_addr.dst_dev_addr, ETH_ALEN);
2532 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
2533 &route->path_rec->sgid);
2534 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.dst_addr,
2535 &route->path_rec->dgid);
2537 /* Use the hint from IP Stack to select GID Type */
2538 if (route->path_rec->gid_type < ib_network_to_gid_type(addr->dev_addr.network))
2539 route->path_rec->gid_type = ib_network_to_gid_type(addr->dev_addr.network);
2540 if (((struct sockaddr *)&id_priv->id.route.addr.dst_addr)->sa_family != AF_IB)
2541 /* TODO: get the hoplimit from the inet/inet6 device */
2542 route->path_rec->hop_limit = addr->dev_addr.hoplimit;
2544 route->path_rec->hop_limit = 1;
2545 route->path_rec->reversible = 1;
2546 route->path_rec->pkey = cpu_to_be16(0xffff);
2547 route->path_rec->mtu_selector = IB_SA_EQ;
2548 route->path_rec->sl = iboe_tos_to_sl(ndev, id_priv->tos);
2549 route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
2550 route->path_rec->rate_selector = IB_SA_EQ;
2551 route->path_rec->rate = iboe_get_rate(ndev);
2553 route->path_rec->packet_life_time_selector = IB_SA_EQ;
2554 route->path_rec->packet_life_time = CMA_IBOE_PACKET_LIFETIME;
2555 if (!route->path_rec->mtu) {
2560 work->old_state = RDMA_CM_ROUTE_QUERY;
2561 work->new_state = RDMA_CM_ROUTE_RESOLVED;
2562 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
2563 work->event.status = 0;
2565 queue_work(cma_wq, &work->work);
2570 kfree(route->path_rec);
2571 route->path_rec = NULL;
2572 route->num_paths = 0;
2578 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
2580 struct rdma_id_private *id_priv;
2583 id_priv = container_of(id, struct rdma_id_private, id);
2584 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY))
2587 atomic_inc(&id_priv->refcount);
2588 if (rdma_cap_ib_sa(id->device, id->port_num))
2589 ret = cma_resolve_ib_route(id_priv, timeout_ms);
2590 else if (rdma_protocol_roce(id->device, id->port_num))
2591 ret = cma_resolve_iboe_route(id_priv);
2592 else if (rdma_protocol_iwarp(id->device, id->port_num))
2593 ret = cma_resolve_iw_route(id_priv, timeout_ms);
2602 cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED);
2603 cma_deref_id(id_priv);
2606 EXPORT_SYMBOL(rdma_resolve_route);
2608 static void cma_set_loopback(struct sockaddr *addr)
2610 switch (addr->sa_family) {
2612 ((struct sockaddr_in *) addr)->sin_addr.s_addr = htonl(INADDR_LOOPBACK);
2615 ipv6_addr_set(&((struct sockaddr_in6 *) addr)->sin6_addr,
2619 ib_addr_set(&((struct sockaddr_ib *) addr)->sib_addr,
2625 static int cma_bind_loopback(struct rdma_id_private *id_priv)
2627 struct cma_device *cma_dev, *cur_dev;
2628 struct ib_port_attr port_attr;
2636 list_for_each_entry(cur_dev, &dev_list, list) {
2637 if (cma_family(id_priv) == AF_IB &&
2638 !rdma_cap_ib_cm(cur_dev->device, 1))
2644 for (p = 1; p <= cur_dev->device->phys_port_cnt; ++p) {
2645 if (!ib_query_port(cur_dev->device, p, &port_attr) &&
2646 port_attr.state == IB_PORT_ACTIVE) {
2661 ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid, NULL);
2665 ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
2669 id_priv->id.route.addr.dev_addr.dev_type =
2670 (rdma_protocol_ib(cma_dev->device, p)) ?
2671 ARPHRD_INFINIBAND : ARPHRD_ETHER;
2673 rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2674 ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
2675 id_priv->id.port_num = p;
2676 cma_attach_to_dev(id_priv, cma_dev);
2677 cma_set_loopback(cma_src_addr(id_priv));
2679 mutex_unlock(&lock);
2683 static void addr_handler(int status, struct sockaddr *src_addr,
2684 struct rdma_dev_addr *dev_addr, void *context)
2686 struct rdma_id_private *id_priv = context;
2687 struct rdma_cm_event event;
2689 memset(&event, 0, sizeof event);
2690 mutex_lock(&id_priv->handler_mutex);
2691 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY,
2692 RDMA_CM_ADDR_RESOLVED))
2695 memcpy(cma_src_addr(id_priv), src_addr, rdma_addr_size(src_addr));
2696 if (!status && !id_priv->cma_dev)
2697 status = cma_acquire_dev(id_priv, NULL);
2700 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
2701 RDMA_CM_ADDR_BOUND))
2703 event.event = RDMA_CM_EVENT_ADDR_ERROR;
2704 event.status = status;
2706 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2708 if (id_priv->id.event_handler(&id_priv->id, &event)) {
2709 cma_exch(id_priv, RDMA_CM_DESTROYING);
2710 mutex_unlock(&id_priv->handler_mutex);
2711 cma_deref_id(id_priv);
2712 rdma_destroy_id(&id_priv->id);
2716 mutex_unlock(&id_priv->handler_mutex);
2717 cma_deref_id(id_priv);
2720 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
2722 struct cma_work *work;
2726 work = kzalloc(sizeof *work, GFP_KERNEL);
2730 if (!id_priv->cma_dev) {
2731 ret = cma_bind_loopback(id_priv);
2736 rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2737 rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
2740 INIT_WORK(&work->work, cma_work_handler);
2741 work->old_state = RDMA_CM_ADDR_QUERY;
2742 work->new_state = RDMA_CM_ADDR_RESOLVED;
2743 work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2744 queue_work(cma_wq, &work->work);
2751 static int cma_resolve_ib_addr(struct rdma_id_private *id_priv)
2753 struct cma_work *work;
2756 work = kzalloc(sizeof *work, GFP_KERNEL);
2760 if (!id_priv->cma_dev) {
2761 ret = cma_resolve_ib_dev(id_priv);
2766 rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, (union ib_gid *)
2767 &(((struct sockaddr_ib *) &id_priv->id.route.addr.dst_addr)->sib_addr));
2770 INIT_WORK(&work->work, cma_work_handler);
2771 work->old_state = RDMA_CM_ADDR_QUERY;
2772 work->new_state = RDMA_CM_ADDR_RESOLVED;
2773 work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2774 queue_work(cma_wq, &work->work);
2781 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2782 struct sockaddr *dst_addr)
2784 if (!src_addr || !src_addr->sa_family) {
2785 src_addr = (struct sockaddr *) &id->route.addr.src_addr;
2786 src_addr->sa_family = dst_addr->sa_family;
2787 if (IS_ENABLED(CONFIG_IPV6) &&
2788 dst_addr->sa_family == AF_INET6) {
2789 struct sockaddr_in6 *src_addr6 = (struct sockaddr_in6 *) src_addr;
2790 struct sockaddr_in6 *dst_addr6 = (struct sockaddr_in6 *) dst_addr;
2791 src_addr6->sin6_scope_id = dst_addr6->sin6_scope_id;
2792 if (ipv6_addr_type(&dst_addr6->sin6_addr) & IPV6_ADDR_LINKLOCAL)
2793 id->route.addr.dev_addr.bound_dev_if = dst_addr6->sin6_scope_id;
2794 } else if (dst_addr->sa_family == AF_IB) {
2795 ((struct sockaddr_ib *) src_addr)->sib_pkey =
2796 ((struct sockaddr_ib *) dst_addr)->sib_pkey;
2799 return rdma_bind_addr(id, src_addr);
2802 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2803 struct sockaddr *dst_addr, int timeout_ms)
2805 struct rdma_id_private *id_priv;
2808 id_priv = container_of(id, struct rdma_id_private, id);
2809 if (id_priv->state == RDMA_CM_IDLE) {
2810 ret = cma_bind_addr(id, src_addr, dst_addr);
2815 if (cma_family(id_priv) != dst_addr->sa_family)
2818 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY))
2821 atomic_inc(&id_priv->refcount);
2822 memcpy(cma_dst_addr(id_priv), dst_addr, rdma_addr_size(dst_addr));
2823 if (cma_any_addr(dst_addr)) {
2824 ret = cma_resolve_loopback(id_priv);
2826 if (dst_addr->sa_family == AF_IB) {
2827 ret = cma_resolve_ib_addr(id_priv);
2829 ret = rdma_resolve_ip(&addr_client, cma_src_addr(id_priv),
2830 dst_addr, &id->route.addr.dev_addr,
2831 timeout_ms, addr_handler, id_priv);
2839 cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
2840 cma_deref_id(id_priv);
2843 EXPORT_SYMBOL(rdma_resolve_addr);
2845 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse)
2847 struct rdma_id_private *id_priv;
2848 unsigned long flags;
2851 id_priv = container_of(id, struct rdma_id_private, id);
2852 spin_lock_irqsave(&id_priv->lock, flags);
2853 if (reuse || id_priv->state == RDMA_CM_IDLE) {
2854 id_priv->reuseaddr = reuse;
2859 spin_unlock_irqrestore(&id_priv->lock, flags);
2862 EXPORT_SYMBOL(rdma_set_reuseaddr);
2864 int rdma_set_afonly(struct rdma_cm_id *id, int afonly)
2866 struct rdma_id_private *id_priv;
2867 unsigned long flags;
2870 id_priv = container_of(id, struct rdma_id_private, id);
2871 spin_lock_irqsave(&id_priv->lock, flags);
2872 if (id_priv->state == RDMA_CM_IDLE || id_priv->state == RDMA_CM_ADDR_BOUND) {
2873 id_priv->options |= (1 << CMA_OPTION_AFONLY);
2874 id_priv->afonly = afonly;
2879 spin_unlock_irqrestore(&id_priv->lock, flags);
2882 EXPORT_SYMBOL(rdma_set_afonly);
2884 static void cma_bind_port(struct rdma_bind_list *bind_list,
2885 struct rdma_id_private *id_priv)
2887 struct sockaddr *addr;
2888 struct sockaddr_ib *sib;
2892 addr = cma_src_addr(id_priv);
2893 port = htons(bind_list->port);
2895 switch (addr->sa_family) {
2897 ((struct sockaddr_in *) addr)->sin_port = port;
2900 ((struct sockaddr_in6 *) addr)->sin6_port = port;
2903 sib = (struct sockaddr_ib *) addr;
2904 sid = be64_to_cpu(sib->sib_sid);
2905 mask = be64_to_cpu(sib->sib_sid_mask);
2906 sib->sib_sid = cpu_to_be64((sid & mask) | (u64) ntohs(port));
2907 sib->sib_sid_mask = cpu_to_be64(~0ULL);
2910 id_priv->bind_list = bind_list;
2911 hlist_add_head(&id_priv->node, &bind_list->owners);
2914 static int cma_alloc_port(enum rdma_port_space ps,
2915 struct rdma_id_private *id_priv, unsigned short snum)
2917 struct rdma_bind_list *bind_list;
2920 bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
2924 ret = cma_ps_alloc(id_priv->id.route.addr.dev_addr.net, ps, bind_list,
2930 bind_list->port = (unsigned short)ret;
2931 cma_bind_port(bind_list, id_priv);
2935 return ret == -ENOSPC ? -EADDRNOTAVAIL : ret;
2938 static int cma_alloc_any_port(enum rdma_port_space ps,
2939 struct rdma_id_private *id_priv)
2941 static unsigned int last_used_port;
2942 int low, high, remaining;
2944 struct net *net = id_priv->id.route.addr.dev_addr.net;
2946 inet_get_local_port_range(net, &low, &high);
2947 remaining = (high - low) + 1;
2948 rover = prandom_u32() % remaining + low;
2950 if (last_used_port != rover &&
2951 !cma_ps_find(net, ps, (unsigned short)rover)) {
2952 int ret = cma_alloc_port(ps, id_priv, rover);
2954 * Remember previously used port number in order to avoid
2955 * re-using same port immediately after it is closed.
2958 last_used_port = rover;
2959 if (ret != -EADDRNOTAVAIL)
2964 if ((rover < low) || (rover > high))
2968 return -EADDRNOTAVAIL;
2972 * Check that the requested port is available. This is called when trying to
2973 * bind to a specific port, or when trying to listen on a bound port. In
2974 * the latter case, the provided id_priv may already be on the bind_list, but
2975 * we still need to check that it's okay to start listening.
2977 static int cma_check_port(struct rdma_bind_list *bind_list,
2978 struct rdma_id_private *id_priv, uint8_t reuseaddr)
2980 struct rdma_id_private *cur_id;
2981 struct sockaddr *addr, *cur_addr;
2983 addr = cma_src_addr(id_priv);
2984 hlist_for_each_entry(cur_id, &bind_list->owners, node) {
2985 if (id_priv == cur_id)
2988 if ((cur_id->state != RDMA_CM_LISTEN) && reuseaddr &&
2992 cur_addr = cma_src_addr(cur_id);
2993 if (id_priv->afonly && cur_id->afonly &&
2994 (addr->sa_family != cur_addr->sa_family))
2997 if (cma_any_addr(addr) || cma_any_addr(cur_addr))
2998 return -EADDRNOTAVAIL;
3000 if (!cma_addr_cmp(addr, cur_addr))
3006 static int cma_use_port(enum rdma_port_space ps,
3007 struct rdma_id_private *id_priv)
3009 struct rdma_bind_list *bind_list;
3010 unsigned short snum;
3013 snum = ntohs(cma_port(cma_src_addr(id_priv)));
3014 if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
3017 bind_list = cma_ps_find(id_priv->id.route.addr.dev_addr.net, ps, snum);
3019 ret = cma_alloc_port(ps, id_priv, snum);
3021 ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr);
3023 cma_bind_port(bind_list, id_priv);
3028 static int cma_bind_listen(struct rdma_id_private *id_priv)
3030 struct rdma_bind_list *bind_list = id_priv->bind_list;
3034 if (bind_list->owners.first->next)
3035 ret = cma_check_port(bind_list, id_priv, 0);
3036 mutex_unlock(&lock);
3040 static enum rdma_port_space cma_select_inet_ps(
3041 struct rdma_id_private *id_priv)
3043 switch (id_priv->id.ps) {
3048 return id_priv->id.ps;
3055 static enum rdma_port_space cma_select_ib_ps(struct rdma_id_private *id_priv)
3057 enum rdma_port_space ps = 0;
3058 struct sockaddr_ib *sib;
3059 u64 sid_ps, mask, sid;
3061 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
3062 mask = be64_to_cpu(sib->sib_sid_mask) & RDMA_IB_IP_PS_MASK;
3063 sid = be64_to_cpu(sib->sib_sid) & mask;
3065 if ((id_priv->id.ps == RDMA_PS_IB) && (sid == (RDMA_IB_IP_PS_IB & mask))) {
3066 sid_ps = RDMA_IB_IP_PS_IB;
3068 } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_TCP)) &&
3069 (sid == (RDMA_IB_IP_PS_TCP & mask))) {
3070 sid_ps = RDMA_IB_IP_PS_TCP;
3072 } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_UDP)) &&
3073 (sid == (RDMA_IB_IP_PS_UDP & mask))) {
3074 sid_ps = RDMA_IB_IP_PS_UDP;
3079 sib->sib_sid = cpu_to_be64(sid_ps | ntohs(cma_port((struct sockaddr *) sib)));
3080 sib->sib_sid_mask = cpu_to_be64(RDMA_IB_IP_PS_MASK |
3081 be64_to_cpu(sib->sib_sid_mask));
3086 static int cma_get_port(struct rdma_id_private *id_priv)
3088 enum rdma_port_space ps;
3091 if (cma_family(id_priv) != AF_IB)
3092 ps = cma_select_inet_ps(id_priv);
3094 ps = cma_select_ib_ps(id_priv);
3096 return -EPROTONOSUPPORT;
3099 if (cma_any_port(cma_src_addr(id_priv)))
3100 ret = cma_alloc_any_port(ps, id_priv);
3102 ret = cma_use_port(ps, id_priv);
3103 mutex_unlock(&lock);
3108 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
3109 struct sockaddr *addr)
3111 #if IS_ENABLED(CONFIG_IPV6)
3112 struct sockaddr_in6 *sin6;
3114 if (addr->sa_family != AF_INET6)
3117 sin6 = (struct sockaddr_in6 *) addr;
3119 if (!(ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL))
3122 if (!sin6->sin6_scope_id)
3125 dev_addr->bound_dev_if = sin6->sin6_scope_id;
3130 int rdma_listen(struct rdma_cm_id *id, int backlog)
3132 struct rdma_id_private *id_priv;
3135 id_priv = container_of(id, struct rdma_id_private, id);
3136 if (id_priv->state == RDMA_CM_IDLE) {
3137 id->route.addr.src_addr.ss_family = AF_INET;
3138 ret = rdma_bind_addr(id, cma_src_addr(id_priv));
3143 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN))
3146 if (id_priv->reuseaddr) {
3147 ret = cma_bind_listen(id_priv);
3152 id_priv->backlog = backlog;
3154 if (rdma_cap_ib_cm(id->device, 1)) {
3155 ret = cma_ib_listen(id_priv);
3158 } else if (rdma_cap_iw_cm(id->device, 1)) {
3159 ret = cma_iw_listen(id_priv, backlog);
3167 cma_listen_on_all(id_priv);
3171 id_priv->backlog = 0;
3172 cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND);
3175 EXPORT_SYMBOL(rdma_listen);
3177 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
3179 struct rdma_id_private *id_priv;
3182 if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6 &&
3183 addr->sa_family != AF_IB)
3184 return -EAFNOSUPPORT;
3186 id_priv = container_of(id, struct rdma_id_private, id);
3187 if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND))
3190 ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
3194 memcpy(cma_src_addr(id_priv), addr, rdma_addr_size(addr));
3195 if (!cma_any_addr(addr)) {
3196 ret = cma_translate_addr(addr, &id->route.addr.dev_addr);
3200 ret = cma_acquire_dev(id_priv, NULL);
3205 if (!(id_priv->options & (1 << CMA_OPTION_AFONLY))) {
3206 if (addr->sa_family == AF_INET)
3207 id_priv->afonly = 1;
3208 #if IS_ENABLED(CONFIG_IPV6)
3209 else if (addr->sa_family == AF_INET6) {
3210 struct net *net = id_priv->id.route.addr.dev_addr.net;
3212 id_priv->afonly = net->ipv6.sysctl.bindv6only;
3216 ret = cma_get_port(id_priv);
3222 if (id_priv->cma_dev)
3223 cma_release_dev(id_priv);
3225 cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE);
3228 EXPORT_SYMBOL(rdma_bind_addr);
3230 static int cma_format_hdr(void *hdr, struct rdma_id_private *id_priv)
3232 struct cma_hdr *cma_hdr;
3235 cma_hdr->cma_version = CMA_VERSION;
3236 if (cma_family(id_priv) == AF_INET) {
3237 struct sockaddr_in *src4, *dst4;
3239 src4 = (struct sockaddr_in *) cma_src_addr(id_priv);
3240 dst4 = (struct sockaddr_in *) cma_dst_addr(id_priv);
3242 cma_set_ip_ver(cma_hdr, 4);
3243 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
3244 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
3245 cma_hdr->port = src4->sin_port;
3246 } else if (cma_family(id_priv) == AF_INET6) {
3247 struct sockaddr_in6 *src6, *dst6;
3249 src6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
3250 dst6 = (struct sockaddr_in6 *) cma_dst_addr(id_priv);
3252 cma_set_ip_ver(cma_hdr, 6);
3253 cma_hdr->src_addr.ip6 = src6->sin6_addr;
3254 cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
3255 cma_hdr->port = src6->sin6_port;
3260 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
3261 struct ib_cm_event *ib_event)
3263 struct rdma_id_private *id_priv = cm_id->context;
3264 struct rdma_cm_event event;
3265 struct ib_cm_sidr_rep_event_param *rep = &ib_event->param.sidr_rep_rcvd;
3268 mutex_lock(&id_priv->handler_mutex);
3269 if (id_priv->state != RDMA_CM_CONNECT)
3272 memset(&event, 0, sizeof event);
3273 switch (ib_event->event) {
3274 case IB_CM_SIDR_REQ_ERROR:
3275 event.event = RDMA_CM_EVENT_UNREACHABLE;
3276 event.status = -ETIMEDOUT;
3278 case IB_CM_SIDR_REP_RECEIVED:
3279 event.param.ud.private_data = ib_event->private_data;
3280 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
3281 if (rep->status != IB_SIDR_SUCCESS) {
3282 event.event = RDMA_CM_EVENT_UNREACHABLE;
3283 event.status = ib_event->param.sidr_rep_rcvd.status;
3286 ret = cma_set_qkey(id_priv, rep->qkey);
3288 event.event = RDMA_CM_EVENT_ADDR_ERROR;
3292 ib_init_ah_from_path(id_priv->id.device, id_priv->id.port_num,
3293 id_priv->id.route.path_rec,
3294 &event.param.ud.ah_attr);
3295 event.param.ud.qp_num = rep->qpn;
3296 event.param.ud.qkey = rep->qkey;
3297 event.event = RDMA_CM_EVENT_ESTABLISHED;
3301 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
3306 ret = id_priv->id.event_handler(&id_priv->id, &event);
3308 /* Destroy the CM ID by returning a non-zero value. */
3309 id_priv->cm_id.ib = NULL;
3310 cma_exch(id_priv, RDMA_CM_DESTROYING);
3311 mutex_unlock(&id_priv->handler_mutex);
3312 rdma_destroy_id(&id_priv->id);
3316 mutex_unlock(&id_priv->handler_mutex);
3320 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
3321 struct rdma_conn_param *conn_param)
3323 struct ib_cm_sidr_req_param req;
3324 struct ib_cm_id *id;
3329 memset(&req, 0, sizeof req);
3330 offset = cma_user_data_offset(id_priv);
3331 req.private_data_len = offset + conn_param->private_data_len;
3332 if (req.private_data_len < conn_param->private_data_len)
3335 if (req.private_data_len) {
3336 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
3340 private_data = NULL;
3343 if (conn_param->private_data && conn_param->private_data_len)
3344 memcpy(private_data + offset, conn_param->private_data,
3345 conn_param->private_data_len);
3348 ret = cma_format_hdr(private_data, id_priv);
3351 req.private_data = private_data;
3354 id = ib_create_cm_id(id_priv->id.device, cma_sidr_rep_handler,
3360 id_priv->cm_id.ib = id;
3362 req.path = id_priv->id.route.path_rec;
3363 req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
3364 req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
3365 req.max_cm_retries = CMA_MAX_CM_RETRIES;
3367 ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
3369 ib_destroy_cm_id(id_priv->cm_id.ib);
3370 id_priv->cm_id.ib = NULL;
3373 kfree(private_data);
3377 static int cma_connect_ib(struct rdma_id_private *id_priv,
3378 struct rdma_conn_param *conn_param)
3380 struct ib_cm_req_param req;
3381 struct rdma_route *route;
3383 struct ib_cm_id *id;
3387 memset(&req, 0, sizeof req);
3388 offset = cma_user_data_offset(id_priv);
3389 req.private_data_len = offset + conn_param->private_data_len;
3390 if (req.private_data_len < conn_param->private_data_len)
3393 if (req.private_data_len) {
3394 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
3398 private_data = NULL;
3401 if (conn_param->private_data && conn_param->private_data_len)
3402 memcpy(private_data + offset, conn_param->private_data,
3403 conn_param->private_data_len);
3405 id = ib_create_cm_id(id_priv->id.device, cma_ib_handler, id_priv);
3410 id_priv->cm_id.ib = id;
3412 route = &id_priv->id.route;
3414 ret = cma_format_hdr(private_data, id_priv);
3417 req.private_data = private_data;
3420 req.primary_path = &route->path_rec[0];
3421 if (route->num_paths == 2)
3422 req.alternate_path = &route->path_rec[1];
3424 req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
3425 req.qp_num = id_priv->qp_num;
3426 req.qp_type = id_priv->id.qp_type;
3427 req.starting_psn = id_priv->seq_num;
3428 req.responder_resources = conn_param->responder_resources;
3429 req.initiator_depth = conn_param->initiator_depth;
3430 req.flow_control = conn_param->flow_control;
3431 req.retry_count = min_t(u8, 7, conn_param->retry_count);
3432 req.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
3433 req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
3434 req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
3435 req.max_cm_retries = CMA_MAX_CM_RETRIES;
3436 req.srq = id_priv->srq ? 1 : 0;
3438 ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
3440 if (ret && !IS_ERR(id)) {
3441 ib_destroy_cm_id(id);
3442 id_priv->cm_id.ib = NULL;
3445 kfree(private_data);
3449 static int cma_connect_iw(struct rdma_id_private *id_priv,
3450 struct rdma_conn_param *conn_param)
3452 struct iw_cm_id *cm_id;
3454 struct iw_cm_conn_param iw_param;
3456 cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
3458 return PTR_ERR(cm_id);
3460 cm_id->tos = id_priv->tos;
3461 id_priv->cm_id.iw = cm_id;
3463 memcpy(&cm_id->local_addr, cma_src_addr(id_priv),
3464 rdma_addr_size(cma_src_addr(id_priv)));
3465 memcpy(&cm_id->remote_addr, cma_dst_addr(id_priv),
3466 rdma_addr_size(cma_dst_addr(id_priv)));
3468 ret = cma_modify_qp_rtr(id_priv, conn_param);
3473 iw_param.ord = conn_param->initiator_depth;
3474 iw_param.ird = conn_param->responder_resources;
3475 iw_param.private_data = conn_param->private_data;
3476 iw_param.private_data_len = conn_param->private_data_len;
3477 iw_param.qpn = id_priv->id.qp ? id_priv->qp_num : conn_param->qp_num;
3479 memset(&iw_param, 0, sizeof iw_param);
3480 iw_param.qpn = id_priv->qp_num;
3482 ret = iw_cm_connect(cm_id, &iw_param);
3485 iw_destroy_cm_id(cm_id);
3486 id_priv->cm_id.iw = NULL;
3491 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
3493 struct rdma_id_private *id_priv;
3496 id_priv = container_of(id, struct rdma_id_private, id);
3497 if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT))
3501 id_priv->qp_num = conn_param->qp_num;
3502 id_priv->srq = conn_param->srq;
3505 if (rdma_cap_ib_cm(id->device, id->port_num)) {
3506 if (id->qp_type == IB_QPT_UD)
3507 ret = cma_resolve_ib_udp(id_priv, conn_param);
3509 ret = cma_connect_ib(id_priv, conn_param);
3510 } else if (rdma_cap_iw_cm(id->device, id->port_num))
3511 ret = cma_connect_iw(id_priv, conn_param);
3519 cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED);
3522 EXPORT_SYMBOL(rdma_connect);
3524 static int cma_accept_ib(struct rdma_id_private *id_priv,
3525 struct rdma_conn_param *conn_param)
3527 struct ib_cm_rep_param rep;
3530 ret = cma_modify_qp_rtr(id_priv, conn_param);
3534 ret = cma_modify_qp_rts(id_priv, conn_param);
3538 memset(&rep, 0, sizeof rep);
3539 rep.qp_num = id_priv->qp_num;
3540 rep.starting_psn = id_priv->seq_num;
3541 rep.private_data = conn_param->private_data;
3542 rep.private_data_len = conn_param->private_data_len;
3543 rep.responder_resources = conn_param->responder_resources;
3544 rep.initiator_depth = conn_param->initiator_depth;
3545 rep.failover_accepted = 0;
3546 rep.flow_control = conn_param->flow_control;
3547 rep.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
3548 rep.srq = id_priv->srq ? 1 : 0;
3550 ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
3555 static int cma_accept_iw(struct rdma_id_private *id_priv,
3556 struct rdma_conn_param *conn_param)
3558 struct iw_cm_conn_param iw_param;
3564 ret = cma_modify_qp_rtr(id_priv, conn_param);
3568 iw_param.ord = conn_param->initiator_depth;
3569 iw_param.ird = conn_param->responder_resources;
3570 iw_param.private_data = conn_param->private_data;
3571 iw_param.private_data_len = conn_param->private_data_len;
3572 if (id_priv->id.qp) {
3573 iw_param.qpn = id_priv->qp_num;
3575 iw_param.qpn = conn_param->qp_num;
3577 return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
3580 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
3581 enum ib_cm_sidr_status status, u32 qkey,
3582 const void *private_data, int private_data_len)
3584 struct ib_cm_sidr_rep_param rep;
3587 memset(&rep, 0, sizeof rep);
3588 rep.status = status;
3589 if (status == IB_SIDR_SUCCESS) {
3590 ret = cma_set_qkey(id_priv, qkey);
3593 rep.qp_num = id_priv->qp_num;
3594 rep.qkey = id_priv->qkey;
3596 rep.private_data = private_data;
3597 rep.private_data_len = private_data_len;
3599 return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
3602 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
3604 struct rdma_id_private *id_priv;
3607 id_priv = container_of(id, struct rdma_id_private, id);
3609 id_priv->owner = task_pid_nr(current);
3611 if (!cma_comp(id_priv, RDMA_CM_CONNECT))
3614 if (!id->qp && conn_param) {
3615 id_priv->qp_num = conn_param->qp_num;
3616 id_priv->srq = conn_param->srq;
3619 if (rdma_cap_ib_cm(id->device, id->port_num)) {
3620 if (id->qp_type == IB_QPT_UD) {
3622 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
3624 conn_param->private_data,
3625 conn_param->private_data_len);
3627 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
3631 ret = cma_accept_ib(id_priv, conn_param);
3633 ret = cma_rep_recv(id_priv);
3635 } else if (rdma_cap_iw_cm(id->device, id->port_num))
3636 ret = cma_accept_iw(id_priv, conn_param);
3645 cma_modify_qp_err(id_priv);
3646 rdma_reject(id, NULL, 0);
3649 EXPORT_SYMBOL(rdma_accept);
3651 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
3653 struct rdma_id_private *id_priv;
3656 id_priv = container_of(id, struct rdma_id_private, id);
3657 if (!id_priv->cm_id.ib)
3660 switch (id->device->node_type) {
3661 case RDMA_NODE_IB_CA:
3662 ret = ib_cm_notify(id_priv->cm_id.ib, event);
3670 EXPORT_SYMBOL(rdma_notify);
3672 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
3673 u8 private_data_len)
3675 struct rdma_id_private *id_priv;
3678 id_priv = container_of(id, struct rdma_id_private, id);
3679 if (!id_priv->cm_id.ib)
3682 if (rdma_cap_ib_cm(id->device, id->port_num)) {
3683 if (id->qp_type == IB_QPT_UD)
3684 ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT, 0,
3685 private_data, private_data_len);
3687 ret = ib_send_cm_rej(id_priv->cm_id.ib,
3688 IB_CM_REJ_CONSUMER_DEFINED, NULL,
3689 0, private_data, private_data_len);
3690 } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
3691 ret = iw_cm_reject(id_priv->cm_id.iw,
3692 private_data, private_data_len);
3698 EXPORT_SYMBOL(rdma_reject);
3700 int rdma_disconnect(struct rdma_cm_id *id)
3702 struct rdma_id_private *id_priv;
3705 id_priv = container_of(id, struct rdma_id_private, id);
3706 if (!id_priv->cm_id.ib)
3709 if (rdma_cap_ib_cm(id->device, id->port_num)) {
3710 ret = cma_modify_qp_err(id_priv);
3713 /* Initiate or respond to a disconnect. */
3714 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
3715 ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
3716 } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
3717 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
3724 EXPORT_SYMBOL(rdma_disconnect);
3726 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
3728 struct rdma_id_private *id_priv;
3729 struct cma_multicast *mc = multicast->context;
3730 struct rdma_cm_event event;
3733 id_priv = mc->id_priv;
3734 mutex_lock(&id_priv->handler_mutex);
3735 if (id_priv->state != RDMA_CM_ADDR_BOUND &&
3736 id_priv->state != RDMA_CM_ADDR_RESOLVED)
3740 status = cma_set_qkey(id_priv, be32_to_cpu(multicast->rec.qkey));
3741 mutex_lock(&id_priv->qp_mutex);
3742 if (!status && id_priv->id.qp)
3743 status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid,
3744 be16_to_cpu(multicast->rec.mlid));
3745 mutex_unlock(&id_priv->qp_mutex);
3747 memset(&event, 0, sizeof event);
3748 event.status = status;
3749 event.param.ud.private_data = mc->context;
3751 struct rdma_dev_addr *dev_addr =
3752 &id_priv->id.route.addr.dev_addr;
3753 struct net_device *ndev =
3754 dev_get_by_index(&init_net, dev_addr->bound_dev_if);
3755 enum ib_gid_type gid_type =
3756 id_priv->cma_dev->default_gid_type[id_priv->id.port_num -
3757 rdma_start_port(id_priv->cma_dev->device)];
3759 event.event = RDMA_CM_EVENT_MULTICAST_JOIN;
3760 ib_init_ah_from_mcmember(id_priv->id.device,
3761 id_priv->id.port_num, &multicast->rec,
3763 &event.param.ud.ah_attr);
3764 event.param.ud.qp_num = 0xFFFFFF;
3765 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
3769 event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
3771 ret = id_priv->id.event_handler(&id_priv->id, &event);
3773 cma_exch(id_priv, RDMA_CM_DESTROYING);
3774 mutex_unlock(&id_priv->handler_mutex);
3775 rdma_destroy_id(&id_priv->id);
3780 mutex_unlock(&id_priv->handler_mutex);
3784 static void cma_set_mgid(struct rdma_id_private *id_priv,
3785 struct sockaddr *addr, union ib_gid *mgid)
3787 unsigned char mc_map[MAX_ADDR_LEN];
3788 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3789 struct sockaddr_in *sin = (struct sockaddr_in *) addr;
3790 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
3792 if (cma_any_addr(addr)) {
3793 memset(mgid, 0, sizeof *mgid);
3794 } else if ((addr->sa_family == AF_INET6) &&
3795 ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
3797 /* IPv6 address is an SA assigned MGID. */
3798 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3799 } else if (addr->sa_family == AF_IB) {
3800 memcpy(mgid, &((struct sockaddr_ib *) addr)->sib_addr, sizeof *mgid);
3801 } else if ((addr->sa_family == AF_INET6)) {
3802 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
3803 if (id_priv->id.ps == RDMA_PS_UDP)
3804 mc_map[7] = 0x01; /* Use RDMA CM signature */
3805 *mgid = *(union ib_gid *) (mc_map + 4);
3807 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
3808 if (id_priv->id.ps == RDMA_PS_UDP)
3809 mc_map[7] = 0x01; /* Use RDMA CM signature */
3810 *mgid = *(union ib_gid *) (mc_map + 4);
3814 static void cma_query_sa_classport_info_cb(int status,
3815 struct ib_class_port_info *rec,
3818 struct class_port_info_context *cb_ctx = context;
3822 if (status || !rec) {
3823 pr_debug("RDMA CM: %s port %u failed query ClassPortInfo status: %d\n",
3824 cb_ctx->device->name, cb_ctx->port_num, status);
3828 memcpy(cb_ctx->class_port_info, rec, sizeof(struct ib_class_port_info));
3831 complete(&cb_ctx->done);
3834 static int cma_query_sa_classport_info(struct ib_device *device, u8 port_num,
3835 struct ib_class_port_info *class_port_info)
3837 struct class_port_info_context *cb_ctx;
3840 cb_ctx = kmalloc(sizeof(*cb_ctx), GFP_KERNEL);
3844 cb_ctx->device = device;
3845 cb_ctx->class_port_info = class_port_info;
3846 cb_ctx->port_num = port_num;
3847 init_completion(&cb_ctx->done);
3849 ret = ib_sa_classport_info_rec_query(&sa_client, device, port_num,
3850 CMA_QUERY_CLASSPORT_INFO_TIMEOUT,
3851 GFP_KERNEL, cma_query_sa_classport_info_cb,
3852 cb_ctx, &cb_ctx->sa_query);
3854 pr_err("RDMA CM: %s port %u failed to send ClassPortInfo query, ret: %d\n",
3855 device->name, port_num, ret);
3859 wait_for_completion(&cb_ctx->done);
3866 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
3867 struct cma_multicast *mc)
3869 struct ib_sa_mcmember_rec rec;
3870 struct ib_class_port_info class_port_info;
3871 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3872 ib_sa_comp_mask comp_mask;
3875 ib_addr_get_mgid(dev_addr, &rec.mgid);
3876 ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
3881 ret = cma_set_qkey(id_priv, 0);
3885 cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
3886 rec.qkey = cpu_to_be32(id_priv->qkey);
3887 rdma_addr_get_sgid(dev_addr, &rec.port_gid);
3888 rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
3889 rec.join_state = mc->join_state;
3891 if (rec.join_state == BIT(SENDONLY_FULLMEMBER_JOIN)) {
3892 ret = cma_query_sa_classport_info(id_priv->id.device,
3893 id_priv->id.port_num,
3899 if (!(ib_get_cpi_capmask2(&class_port_info) &
3900 IB_SA_CAP_MASK2_SENDONLY_FULL_MEM_SUPPORT)) {
3901 pr_warn("RDMA CM: %s port %u Unable to multicast join\n"
3902 "RDMA CM: SM doesn't support Send Only Full Member option\n",
3903 id_priv->id.device->name, id_priv->id.port_num);
3908 comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
3909 IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
3910 IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
3911 IB_SA_MCMEMBER_REC_FLOW_LABEL |
3912 IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
3914 if (id_priv->id.ps == RDMA_PS_IPOIB)
3915 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
3916 IB_SA_MCMEMBER_REC_RATE_SELECTOR |
3917 IB_SA_MCMEMBER_REC_MTU_SELECTOR |
3918 IB_SA_MCMEMBER_REC_MTU |
3919 IB_SA_MCMEMBER_REC_HOP_LIMIT;
3921 mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device,
3922 id_priv->id.port_num, &rec,
3923 comp_mask, GFP_KERNEL,
3924 cma_ib_mc_handler, mc);
3925 return PTR_ERR_OR_ZERO(mc->multicast.ib);
3928 static void iboe_mcast_work_handler(struct work_struct *work)
3930 struct iboe_mcast_work *mw = container_of(work, struct iboe_mcast_work, work);
3931 struct cma_multicast *mc = mw->mc;
3932 struct ib_sa_multicast *m = mc->multicast.ib;
3934 mc->multicast.ib->context = mc;
3935 cma_ib_mc_handler(0, m);
3936 kref_put(&mc->mcref, release_mc);
3940 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid)
3942 struct sockaddr_in *sin = (struct sockaddr_in *)addr;
3943 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
3945 if (cma_any_addr(addr)) {
3946 memset(mgid, 0, sizeof *mgid);
3947 } else if (addr->sa_family == AF_INET6) {
3948 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3950 mgid->raw[0] = 0xff;
3951 mgid->raw[1] = 0x0e;
3960 mgid->raw[10] = 0xff;
3961 mgid->raw[11] = 0xff;
3962 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
3966 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
3967 struct cma_multicast *mc)
3969 struct iboe_mcast_work *work;
3970 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3972 struct sockaddr *addr = (struct sockaddr *)&mc->addr;
3973 struct net_device *ndev = NULL;
3974 enum ib_gid_type gid_type;
3977 send_only = mc->join_state == BIT(SENDONLY_FULLMEMBER_JOIN);
3979 if (cma_zero_addr((struct sockaddr *)&mc->addr))
3982 work = kzalloc(sizeof *work, GFP_KERNEL);
3986 mc->multicast.ib = kzalloc(sizeof(struct ib_sa_multicast), GFP_KERNEL);
3987 if (!mc->multicast.ib) {
3992 cma_iboe_set_mgid(addr, &mc->multicast.ib->rec.mgid);
3994 mc->multicast.ib->rec.pkey = cpu_to_be16(0xffff);
3995 if (id_priv->id.ps == RDMA_PS_UDP)
3996 mc->multicast.ib->rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
3998 if (dev_addr->bound_dev_if)
3999 ndev = dev_get_by_index(&init_net, dev_addr->bound_dev_if);
4004 mc->multicast.ib->rec.rate = iboe_get_rate(ndev);
4005 mc->multicast.ib->rec.hop_limit = 1;
4006 mc->multicast.ib->rec.mtu = iboe_get_mtu(ndev->mtu);
4008 gid_type = id_priv->cma_dev->default_gid_type[id_priv->id.port_num -
4009 rdma_start_port(id_priv->cma_dev->device)];
4010 if (addr->sa_family == AF_INET) {
4011 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) {
4012 mc->multicast.ib->rec.hop_limit = IPV6_DEFAULT_HOPLIMIT;
4014 err = cma_igmp_send(ndev, &mc->multicast.ib->rec.mgid,
4017 mc->igmp_joined = true;
4021 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP)
4025 if (err || !mc->multicast.ib->rec.mtu) {
4030 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
4031 &mc->multicast.ib->rec.port_gid);
4034 INIT_WORK(&work->work, iboe_mcast_work_handler);
4035 kref_get(&mc->mcref);
4036 queue_work(cma_wq, &work->work);
4041 kfree(mc->multicast.ib);
4047 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
4048 u8 join_state, void *context)
4050 struct rdma_id_private *id_priv;
4051 struct cma_multicast *mc;
4057 id_priv = container_of(id, struct rdma_id_private, id);
4058 if (!cma_comp(id_priv, RDMA_CM_ADDR_BOUND) &&
4059 !cma_comp(id_priv, RDMA_CM_ADDR_RESOLVED))
4062 mc = kmalloc(sizeof *mc, GFP_KERNEL);
4066 memcpy(&mc->addr, addr, rdma_addr_size(addr));
4067 mc->context = context;
4068 mc->id_priv = id_priv;
4069 mc->igmp_joined = false;
4070 mc->join_state = join_state;
4071 spin_lock(&id_priv->lock);
4072 list_add(&mc->list, &id_priv->mc_list);
4073 spin_unlock(&id_priv->lock);
4075 if (rdma_protocol_roce(id->device, id->port_num)) {
4076 kref_init(&mc->mcref);
4077 ret = cma_iboe_join_multicast(id_priv, mc);
4078 } else if (rdma_cap_ib_mcast(id->device, id->port_num))
4079 ret = cma_join_ib_multicast(id_priv, mc);
4084 spin_lock_irq(&id_priv->lock);
4085 list_del(&mc->list);
4086 spin_unlock_irq(&id_priv->lock);
4091 EXPORT_SYMBOL(rdma_join_multicast);
4093 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
4095 struct rdma_id_private *id_priv;
4096 struct cma_multicast *mc;
4098 id_priv = container_of(id, struct rdma_id_private, id);
4099 spin_lock_irq(&id_priv->lock);
4100 list_for_each_entry(mc, &id_priv->mc_list, list) {
4101 if (!memcmp(&mc->addr, addr, rdma_addr_size(addr))) {
4102 list_del(&mc->list);
4103 spin_unlock_irq(&id_priv->lock);
4106 ib_detach_mcast(id->qp,
4107 &mc->multicast.ib->rec.mgid,
4108 be16_to_cpu(mc->multicast.ib->rec.mlid));
4110 BUG_ON(id_priv->cma_dev->device != id->device);
4112 if (rdma_cap_ib_mcast(id->device, id->port_num)) {
4113 ib_sa_free_multicast(mc->multicast.ib);
4115 } else if (rdma_protocol_roce(id->device, id->port_num)) {
4116 if (mc->igmp_joined) {
4117 struct rdma_dev_addr *dev_addr =
4118 &id->route.addr.dev_addr;
4119 struct net_device *ndev = NULL;
4121 if (dev_addr->bound_dev_if)
4122 ndev = dev_get_by_index(&init_net,
4123 dev_addr->bound_dev_if);
4126 &mc->multicast.ib->rec.mgid,
4130 mc->igmp_joined = false;
4132 kref_put(&mc->mcref, release_mc);
4137 spin_unlock_irq(&id_priv->lock);
4139 EXPORT_SYMBOL(rdma_leave_multicast);
4141 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
4143 struct rdma_dev_addr *dev_addr;
4144 struct cma_ndev_work *work;
4146 dev_addr = &id_priv->id.route.addr.dev_addr;
4148 if ((dev_addr->bound_dev_if == ndev->ifindex) &&
4149 (net_eq(dev_net(ndev), dev_addr->net)) &&
4150 memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
4151 pr_info("RDMA CM addr change for ndev %s used by id %p\n",
4152 ndev->name, &id_priv->id);
4153 work = kzalloc(sizeof *work, GFP_KERNEL);
4157 INIT_WORK(&work->work, cma_ndev_work_handler);
4159 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
4160 atomic_inc(&id_priv->refcount);
4161 queue_work(cma_wq, &work->work);
4167 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
4170 struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
4171 struct cma_device *cma_dev;
4172 struct rdma_id_private *id_priv;
4173 int ret = NOTIFY_DONE;
4175 if (event != NETDEV_BONDING_FAILOVER)
4178 if (!(ndev->flags & IFF_MASTER) || !(ndev->priv_flags & IFF_BONDING))
4182 list_for_each_entry(cma_dev, &dev_list, list)
4183 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
4184 ret = cma_netdev_change(ndev, id_priv);
4190 mutex_unlock(&lock);
4194 static struct notifier_block cma_nb = {
4195 .notifier_call = cma_netdev_callback
4198 static void cma_add_one(struct ib_device *device)
4200 struct cma_device *cma_dev;
4201 struct rdma_id_private *id_priv;
4203 unsigned long supported_gids = 0;
4205 cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
4209 cma_dev->device = device;
4210 cma_dev->default_gid_type = kcalloc(device->phys_port_cnt,
4211 sizeof(*cma_dev->default_gid_type),
4213 if (!cma_dev->default_gid_type) {
4217 for (i = rdma_start_port(device); i <= rdma_end_port(device); i++) {
4218 supported_gids = roce_gid_type_mask_support(device, i);
4219 WARN_ON(!supported_gids);
4220 cma_dev->default_gid_type[i - rdma_start_port(device)] =
4221 find_first_bit(&supported_gids, BITS_PER_LONG);
4224 init_completion(&cma_dev->comp);
4225 atomic_set(&cma_dev->refcount, 1);
4226 INIT_LIST_HEAD(&cma_dev->id_list);
4227 ib_set_client_data(device, &cma_client, cma_dev);
4230 list_add_tail(&cma_dev->list, &dev_list);
4231 list_for_each_entry(id_priv, &listen_any_list, list)
4232 cma_listen_on_dev(id_priv, cma_dev);
4233 mutex_unlock(&lock);
4236 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
4238 struct rdma_cm_event event;
4239 enum rdma_cm_state state;
4242 /* Record that we want to remove the device */
4243 state = cma_exch(id_priv, RDMA_CM_DEVICE_REMOVAL);
4244 if (state == RDMA_CM_DESTROYING)
4247 cma_cancel_operation(id_priv, state);
4248 mutex_lock(&id_priv->handler_mutex);
4250 /* Check for destruction from another callback. */
4251 if (!cma_comp(id_priv, RDMA_CM_DEVICE_REMOVAL))
4254 memset(&event, 0, sizeof event);
4255 event.event = RDMA_CM_EVENT_DEVICE_REMOVAL;
4256 ret = id_priv->id.event_handler(&id_priv->id, &event);
4258 mutex_unlock(&id_priv->handler_mutex);
4262 static void cma_process_remove(struct cma_device *cma_dev)
4264 struct rdma_id_private *id_priv;
4268 while (!list_empty(&cma_dev->id_list)) {
4269 id_priv = list_entry(cma_dev->id_list.next,
4270 struct rdma_id_private, list);
4272 list_del(&id_priv->listen_list);
4273 list_del_init(&id_priv->list);
4274 atomic_inc(&id_priv->refcount);
4275 mutex_unlock(&lock);
4277 ret = id_priv->internal_id ? 1 : cma_remove_id_dev(id_priv);
4278 cma_deref_id(id_priv);
4280 rdma_destroy_id(&id_priv->id);
4284 mutex_unlock(&lock);
4286 cma_deref_dev(cma_dev);
4287 wait_for_completion(&cma_dev->comp);
4290 static void cma_remove_one(struct ib_device *device, void *client_data)
4292 struct cma_device *cma_dev = client_data;
4298 list_del(&cma_dev->list);
4299 mutex_unlock(&lock);
4301 cma_process_remove(cma_dev);
4302 kfree(cma_dev->default_gid_type);
4306 static int cma_get_id_stats(struct sk_buff *skb, struct netlink_callback *cb)
4308 struct nlmsghdr *nlh;
4309 struct rdma_cm_id_stats *id_stats;
4310 struct rdma_id_private *id_priv;
4311 struct rdma_cm_id *id = NULL;
4312 struct cma_device *cma_dev;
4313 int i_dev = 0, i_id = 0;
4316 * We export all of the IDs as a sequence of messages. Each
4317 * ID gets its own netlink message.
4321 list_for_each_entry(cma_dev, &dev_list, list) {
4322 if (i_dev < cb->args[0]) {
4328 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
4329 if (i_id < cb->args[1]) {
4334 id_stats = ibnl_put_msg(skb, &nlh, cb->nlh->nlmsg_seq,
4335 sizeof *id_stats, RDMA_NL_RDMA_CM,
4336 RDMA_NL_RDMA_CM_ID_STATS,
4341 memset(id_stats, 0, sizeof *id_stats);
4343 id_stats->node_type = id->route.addr.dev_addr.dev_type;
4344 id_stats->port_num = id->port_num;
4345 id_stats->bound_dev_if =
4346 id->route.addr.dev_addr.bound_dev_if;
4348 if (ibnl_put_attr(skb, nlh,
4349 rdma_addr_size(cma_src_addr(id_priv)),
4350 cma_src_addr(id_priv),
4351 RDMA_NL_RDMA_CM_ATTR_SRC_ADDR))
4353 if (ibnl_put_attr(skb, nlh,
4354 rdma_addr_size(cma_dst_addr(id_priv)),
4355 cma_dst_addr(id_priv),
4356 RDMA_NL_RDMA_CM_ATTR_DST_ADDR))
4359 id_stats->pid = id_priv->owner;
4360 id_stats->port_space = id->ps;
4361 id_stats->cm_state = id_priv->state;
4362 id_stats->qp_num = id_priv->qp_num;
4363 id_stats->qp_type = id->qp_type;
4373 mutex_unlock(&lock);
4374 cb->args[0] = i_dev;
4380 static const struct ibnl_client_cbs cma_cb_table[] = {
4381 [RDMA_NL_RDMA_CM_ID_STATS] = { .dump = cma_get_id_stats,
4382 .module = THIS_MODULE },
4385 static int cma_init_net(struct net *net)
4387 struct cma_pernet *pernet = cma_pernet(net);
4389 idr_init(&pernet->tcp_ps);
4390 idr_init(&pernet->udp_ps);
4391 idr_init(&pernet->ipoib_ps);
4392 idr_init(&pernet->ib_ps);
4397 static void cma_exit_net(struct net *net)
4399 struct cma_pernet *pernet = cma_pernet(net);
4401 idr_destroy(&pernet->tcp_ps);
4402 idr_destroy(&pernet->udp_ps);
4403 idr_destroy(&pernet->ipoib_ps);
4404 idr_destroy(&pernet->ib_ps);
4407 static struct pernet_operations cma_pernet_operations = {
4408 .init = cma_init_net,
4409 .exit = cma_exit_net,
4410 .id = &cma_pernet_id,
4411 .size = sizeof(struct cma_pernet),
4414 static int __init cma_init(void)
4418 cma_wq = alloc_ordered_workqueue("rdma_cm", WQ_MEM_RECLAIM);
4422 ret = register_pernet_subsys(&cma_pernet_operations);
4426 ib_sa_register_client(&sa_client);
4427 rdma_addr_register_client(&addr_client);
4428 register_netdevice_notifier(&cma_nb);
4430 ret = ib_register_client(&cma_client);
4434 if (ibnl_add_client(RDMA_NL_RDMA_CM, ARRAY_SIZE(cma_cb_table),
4436 pr_warn("RDMA CMA: failed to add netlink callback\n");
4437 cma_configfs_init();
4442 unregister_netdevice_notifier(&cma_nb);
4443 rdma_addr_unregister_client(&addr_client);
4444 ib_sa_unregister_client(&sa_client);
4445 unregister_pernet_subsys(&cma_pernet_operations);
4447 destroy_workqueue(cma_wq);
4451 static void __exit cma_cleanup(void)
4453 cma_configfs_exit();
4454 ibnl_remove_client(RDMA_NL_RDMA_CM);
4455 ib_unregister_client(&cma_client);
4456 unregister_netdevice_notifier(&cma_nb);
4457 rdma_addr_unregister_client(&addr_client);
4458 ib_sa_unregister_client(&sa_client);
4459 unregister_pernet_subsys(&cma_pernet_operations);
4460 destroy_workqueue(cma_wq);
4463 module_init(cma_init);
4464 module_exit(cma_cleanup);