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/idr.h>
42 #include <linux/inetdevice.h>
43 #include <linux/slab.h>
44 #include <linux/module.h>
45 #include <net/route.h>
47 #include <net/net_namespace.h>
48 #include <net/netns/generic.h>
51 #include <net/ip_fib.h>
52 #include <net/ip6_route.h>
54 #include <rdma/rdma_cm.h>
55 #include <rdma/rdma_cm_ib.h>
56 #include <rdma/rdma_netlink.h>
58 #include <rdma/ib_cache.h>
59 #include <rdma/ib_cm.h>
60 #include <rdma/ib_sa.h>
61 #include <rdma/iw_cm.h>
63 MODULE_AUTHOR("Sean Hefty");
64 MODULE_DESCRIPTION("Generic RDMA CM Agent");
65 MODULE_LICENSE("Dual BSD/GPL");
67 #define CMA_CM_RESPONSE_TIMEOUT 20
68 #define CMA_MAX_CM_RETRIES 15
69 #define CMA_CM_MRA_SETTING (IB_CM_MRA_FLAG_DELAY | 24)
70 #define CMA_IBOE_PACKET_LIFETIME 18
72 static const char * const cma_events[] = {
73 [RDMA_CM_EVENT_ADDR_RESOLVED] = "address resolved",
74 [RDMA_CM_EVENT_ADDR_ERROR] = "address error",
75 [RDMA_CM_EVENT_ROUTE_RESOLVED] = "route resolved ",
76 [RDMA_CM_EVENT_ROUTE_ERROR] = "route error",
77 [RDMA_CM_EVENT_CONNECT_REQUEST] = "connect request",
78 [RDMA_CM_EVENT_CONNECT_RESPONSE] = "connect response",
79 [RDMA_CM_EVENT_CONNECT_ERROR] = "connect error",
80 [RDMA_CM_EVENT_UNREACHABLE] = "unreachable",
81 [RDMA_CM_EVENT_REJECTED] = "rejected",
82 [RDMA_CM_EVENT_ESTABLISHED] = "established",
83 [RDMA_CM_EVENT_DISCONNECTED] = "disconnected",
84 [RDMA_CM_EVENT_DEVICE_REMOVAL] = "device removal",
85 [RDMA_CM_EVENT_MULTICAST_JOIN] = "multicast join",
86 [RDMA_CM_EVENT_MULTICAST_ERROR] = "multicast error",
87 [RDMA_CM_EVENT_ADDR_CHANGE] = "address change",
88 [RDMA_CM_EVENT_TIMEWAIT_EXIT] = "timewait exit",
91 const char *__attribute_const__ rdma_event_msg(enum rdma_cm_event_type event)
95 return (index < ARRAY_SIZE(cma_events) && cma_events[index]) ?
96 cma_events[index] : "unrecognized event";
98 EXPORT_SYMBOL(rdma_event_msg);
100 static void cma_add_one(struct ib_device *device);
101 static void cma_remove_one(struct ib_device *device, void *client_data);
103 static struct ib_client cma_client = {
106 .remove = cma_remove_one
109 static struct ib_sa_client sa_client;
110 static struct rdma_addr_client addr_client;
111 static LIST_HEAD(dev_list);
112 static LIST_HEAD(listen_any_list);
113 static DEFINE_MUTEX(lock);
114 static struct workqueue_struct *cma_wq;
115 static int cma_pernet_id;
124 static struct cma_pernet *cma_pernet(struct net *net)
126 return net_generic(net, cma_pernet_id);
129 static struct idr *cma_pernet_idr(struct net *net, enum rdma_port_space ps)
131 struct cma_pernet *pernet = cma_pernet(net);
135 return &pernet->tcp_ps;
137 return &pernet->udp_ps;
139 return &pernet->ipoib_ps;
141 return &pernet->ib_ps;
148 struct list_head list;
149 struct ib_device *device;
150 struct completion comp;
152 struct list_head id_list;
155 struct rdma_bind_list {
156 enum rdma_port_space ps;
157 struct hlist_head owners;
161 static int cma_ps_alloc(struct net *net, enum rdma_port_space ps,
162 struct rdma_bind_list *bind_list, int snum)
164 struct idr *idr = cma_pernet_idr(net, ps);
166 return idr_alloc(idr, bind_list, snum, snum + 1, GFP_KERNEL);
169 static struct rdma_bind_list *cma_ps_find(struct net *net,
170 enum rdma_port_space ps, int snum)
172 struct idr *idr = cma_pernet_idr(net, ps);
174 return idr_find(idr, snum);
177 static void cma_ps_remove(struct net *net, enum rdma_port_space ps, int snum)
179 struct idr *idr = cma_pernet_idr(net, ps);
181 idr_remove(idr, snum);
189 * Device removal can occur at anytime, so we need extra handling to
190 * serialize notifying the user of device removal with other callbacks.
191 * We do this by disabling removal notification while a callback is in process,
192 * and reporting it after the callback completes.
194 struct rdma_id_private {
195 struct rdma_cm_id id;
197 struct rdma_bind_list *bind_list;
198 struct hlist_node node;
199 struct list_head list; /* listen_any_list or cma_device.list */
200 struct list_head listen_list; /* per device listens */
201 struct cma_device *cma_dev;
202 struct list_head mc_list;
205 enum rdma_cm_state state;
207 struct mutex qp_mutex;
209 struct completion comp;
211 struct mutex handler_mutex;
215 struct ib_sa_query *query;
233 struct cma_multicast {
234 struct rdma_id_private *id_priv;
236 struct ib_sa_multicast *ib;
238 struct list_head list;
240 struct sockaddr_storage addr;
245 struct work_struct work;
246 struct rdma_id_private *id;
247 enum rdma_cm_state old_state;
248 enum rdma_cm_state new_state;
249 struct rdma_cm_event event;
252 struct cma_ndev_work {
253 struct work_struct work;
254 struct rdma_id_private *id;
255 struct rdma_cm_event event;
258 struct iboe_mcast_work {
259 struct work_struct work;
260 struct rdma_id_private *id;
261 struct cma_multicast *mc;
274 u8 ip_version; /* IP version: 7:4 */
276 union cma_ip_addr src_addr;
277 union cma_ip_addr dst_addr;
280 #define CMA_VERSION 0x00
282 struct cma_req_info {
283 struct ib_device *device;
285 union ib_gid local_gid;
291 static int cma_comp(struct rdma_id_private *id_priv, enum rdma_cm_state comp)
296 spin_lock_irqsave(&id_priv->lock, flags);
297 ret = (id_priv->state == comp);
298 spin_unlock_irqrestore(&id_priv->lock, flags);
302 static int cma_comp_exch(struct rdma_id_private *id_priv,
303 enum rdma_cm_state comp, enum rdma_cm_state exch)
308 spin_lock_irqsave(&id_priv->lock, flags);
309 if ((ret = (id_priv->state == comp)))
310 id_priv->state = exch;
311 spin_unlock_irqrestore(&id_priv->lock, flags);
315 static enum rdma_cm_state cma_exch(struct rdma_id_private *id_priv,
316 enum rdma_cm_state exch)
319 enum rdma_cm_state old;
321 spin_lock_irqsave(&id_priv->lock, flags);
322 old = id_priv->state;
323 id_priv->state = exch;
324 spin_unlock_irqrestore(&id_priv->lock, flags);
328 static inline u8 cma_get_ip_ver(const struct cma_hdr *hdr)
330 return hdr->ip_version >> 4;
333 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
335 hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
338 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
339 struct cma_device *cma_dev)
341 atomic_inc(&cma_dev->refcount);
342 id_priv->cma_dev = cma_dev;
343 id_priv->id.device = cma_dev->device;
344 id_priv->id.route.addr.dev_addr.transport =
345 rdma_node_get_transport(cma_dev->device->node_type);
346 list_add_tail(&id_priv->list, &cma_dev->id_list);
349 static inline void cma_deref_dev(struct cma_device *cma_dev)
351 if (atomic_dec_and_test(&cma_dev->refcount))
352 complete(&cma_dev->comp);
355 static inline void release_mc(struct kref *kref)
357 struct cma_multicast *mc = container_of(kref, struct cma_multicast, mcref);
359 kfree(mc->multicast.ib);
363 static void cma_release_dev(struct rdma_id_private *id_priv)
366 list_del(&id_priv->list);
367 cma_deref_dev(id_priv->cma_dev);
368 id_priv->cma_dev = NULL;
372 static inline struct sockaddr *cma_src_addr(struct rdma_id_private *id_priv)
374 return (struct sockaddr *) &id_priv->id.route.addr.src_addr;
377 static inline struct sockaddr *cma_dst_addr(struct rdma_id_private *id_priv)
379 return (struct sockaddr *) &id_priv->id.route.addr.dst_addr;
382 static inline unsigned short cma_family(struct rdma_id_private *id_priv)
384 return id_priv->id.route.addr.src_addr.ss_family;
387 static int cma_set_qkey(struct rdma_id_private *id_priv, u32 qkey)
389 struct ib_sa_mcmember_rec rec;
393 if (qkey && id_priv->qkey != qkey)
399 id_priv->qkey = qkey;
403 switch (id_priv->id.ps) {
406 id_priv->qkey = RDMA_UDP_QKEY;
409 ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid);
410 ret = ib_sa_get_mcmember_rec(id_priv->id.device,
411 id_priv->id.port_num, &rec.mgid,
414 id_priv->qkey = be32_to_cpu(rec.qkey);
422 static void cma_translate_ib(struct sockaddr_ib *sib, struct rdma_dev_addr *dev_addr)
424 dev_addr->dev_type = ARPHRD_INFINIBAND;
425 rdma_addr_set_sgid(dev_addr, (union ib_gid *) &sib->sib_addr);
426 ib_addr_set_pkey(dev_addr, ntohs(sib->sib_pkey));
429 static int cma_translate_addr(struct sockaddr *addr, struct rdma_dev_addr *dev_addr)
433 if (addr->sa_family != AF_IB) {
434 ret = rdma_translate_ip(addr, dev_addr, NULL);
436 cma_translate_ib((struct sockaddr_ib *) addr, dev_addr);
443 static inline int cma_validate_port(struct ib_device *device, u8 port,
444 union ib_gid *gid, int dev_type,
448 struct net_device *ndev = NULL;
450 if ((dev_type == ARPHRD_INFINIBAND) && !rdma_protocol_ib(device, port))
453 if ((dev_type != ARPHRD_INFINIBAND) && rdma_protocol_ib(device, port))
456 if (dev_type == ARPHRD_ETHER)
457 ndev = dev_get_by_index(&init_net, bound_if_index);
459 ret = ib_find_cached_gid_by_port(device, gid, port, ndev, NULL);
467 static int cma_acquire_dev(struct rdma_id_private *id_priv,
468 struct rdma_id_private *listen_id_priv)
470 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
471 struct cma_device *cma_dev;
472 union ib_gid gid, iboe_gid, *gidp;
476 if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
477 id_priv->id.ps == RDMA_PS_IPOIB)
481 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
484 memcpy(&gid, dev_addr->src_dev_addr +
485 rdma_addr_gid_offset(dev_addr), sizeof gid);
487 if (listen_id_priv) {
488 cma_dev = listen_id_priv->cma_dev;
489 port = listen_id_priv->id.port_num;
490 gidp = rdma_protocol_roce(cma_dev->device, port) ?
493 ret = cma_validate_port(cma_dev->device, port, gidp,
495 dev_addr->bound_dev_if);
497 id_priv->id.port_num = port;
502 list_for_each_entry(cma_dev, &dev_list, list) {
503 for (port = 1; port <= cma_dev->device->phys_port_cnt; ++port) {
504 if (listen_id_priv &&
505 listen_id_priv->cma_dev == cma_dev &&
506 listen_id_priv->id.port_num == port)
509 gidp = rdma_protocol_roce(cma_dev->device, port) ?
512 ret = cma_validate_port(cma_dev->device, port, gidp,
514 dev_addr->bound_dev_if);
516 id_priv->id.port_num = port;
524 cma_attach_to_dev(id_priv, cma_dev);
531 * Select the source IB device and address to reach the destination IB address.
533 static int cma_resolve_ib_dev(struct rdma_id_private *id_priv)
535 struct cma_device *cma_dev, *cur_dev;
536 struct sockaddr_ib *addr;
537 union ib_gid gid, sgid, *dgid;
543 addr = (struct sockaddr_ib *) cma_dst_addr(id_priv);
544 dgid = (union ib_gid *) &addr->sib_addr;
545 pkey = ntohs(addr->sib_pkey);
547 list_for_each_entry(cur_dev, &dev_list, list) {
548 for (p = 1; p <= cur_dev->device->phys_port_cnt; ++p) {
549 if (!rdma_cap_af_ib(cur_dev->device, p))
552 if (ib_find_cached_pkey(cur_dev->device, p, pkey, &index))
555 for (i = 0; !ib_get_cached_gid(cur_dev->device, p, i,
558 if (!memcmp(&gid, dgid, sizeof(gid))) {
561 id_priv->id.port_num = p;
565 if (!cma_dev && (gid.global.subnet_prefix ==
566 dgid->global.subnet_prefix)) {
569 id_priv->id.port_num = p;
579 cma_attach_to_dev(id_priv, cma_dev);
580 addr = (struct sockaddr_ib *) cma_src_addr(id_priv);
581 memcpy(&addr->sib_addr, &sgid, sizeof sgid);
582 cma_translate_ib(addr, &id_priv->id.route.addr.dev_addr);
586 static void cma_deref_id(struct rdma_id_private *id_priv)
588 if (atomic_dec_and_test(&id_priv->refcount))
589 complete(&id_priv->comp);
592 static int cma_disable_callback(struct rdma_id_private *id_priv,
593 enum rdma_cm_state state)
595 mutex_lock(&id_priv->handler_mutex);
596 if (id_priv->state != state) {
597 mutex_unlock(&id_priv->handler_mutex);
603 struct rdma_cm_id *rdma_create_id(struct net *net,
604 rdma_cm_event_handler event_handler,
605 void *context, enum rdma_port_space ps,
606 enum ib_qp_type qp_type)
608 struct rdma_id_private *id_priv;
610 id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
612 return ERR_PTR(-ENOMEM);
614 id_priv->owner = task_pid_nr(current);
615 id_priv->state = RDMA_CM_IDLE;
616 id_priv->id.context = context;
617 id_priv->id.event_handler = event_handler;
619 id_priv->id.qp_type = qp_type;
620 spin_lock_init(&id_priv->lock);
621 mutex_init(&id_priv->qp_mutex);
622 init_completion(&id_priv->comp);
623 atomic_set(&id_priv->refcount, 1);
624 mutex_init(&id_priv->handler_mutex);
625 INIT_LIST_HEAD(&id_priv->listen_list);
626 INIT_LIST_HEAD(&id_priv->mc_list);
627 get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
628 id_priv->id.route.addr.dev_addr.net = get_net(net);
632 EXPORT_SYMBOL(rdma_create_id);
634 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
636 struct ib_qp_attr qp_attr;
637 int qp_attr_mask, ret;
639 qp_attr.qp_state = IB_QPS_INIT;
640 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
644 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
648 qp_attr.qp_state = IB_QPS_RTR;
649 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
653 qp_attr.qp_state = IB_QPS_RTS;
655 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
660 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
662 struct ib_qp_attr qp_attr;
663 int qp_attr_mask, ret;
665 qp_attr.qp_state = IB_QPS_INIT;
666 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
670 return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
673 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
674 struct ib_qp_init_attr *qp_init_attr)
676 struct rdma_id_private *id_priv;
680 id_priv = container_of(id, struct rdma_id_private, id);
681 if (id->device != pd->device)
684 qp = ib_create_qp(pd, qp_init_attr);
688 if (id->qp_type == IB_QPT_UD)
689 ret = cma_init_ud_qp(id_priv, qp);
691 ret = cma_init_conn_qp(id_priv, qp);
696 id_priv->qp_num = qp->qp_num;
697 id_priv->srq = (qp->srq != NULL);
703 EXPORT_SYMBOL(rdma_create_qp);
705 void rdma_destroy_qp(struct rdma_cm_id *id)
707 struct rdma_id_private *id_priv;
709 id_priv = container_of(id, struct rdma_id_private, id);
710 mutex_lock(&id_priv->qp_mutex);
711 ib_destroy_qp(id_priv->id.qp);
712 id_priv->id.qp = NULL;
713 mutex_unlock(&id_priv->qp_mutex);
715 EXPORT_SYMBOL(rdma_destroy_qp);
717 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
718 struct rdma_conn_param *conn_param)
720 struct ib_qp_attr qp_attr;
721 int qp_attr_mask, ret;
724 mutex_lock(&id_priv->qp_mutex);
725 if (!id_priv->id.qp) {
730 /* Need to update QP attributes from default values. */
731 qp_attr.qp_state = IB_QPS_INIT;
732 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
736 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
740 qp_attr.qp_state = IB_QPS_RTR;
741 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
745 ret = ib_query_gid(id_priv->id.device, id_priv->id.port_num,
746 qp_attr.ah_attr.grh.sgid_index, &sgid, NULL);
750 BUG_ON(id_priv->cma_dev->device != id_priv->id.device);
753 qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
754 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
756 mutex_unlock(&id_priv->qp_mutex);
760 static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
761 struct rdma_conn_param *conn_param)
763 struct ib_qp_attr qp_attr;
764 int qp_attr_mask, ret;
766 mutex_lock(&id_priv->qp_mutex);
767 if (!id_priv->id.qp) {
772 qp_attr.qp_state = IB_QPS_RTS;
773 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
778 qp_attr.max_rd_atomic = conn_param->initiator_depth;
779 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
781 mutex_unlock(&id_priv->qp_mutex);
785 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
787 struct ib_qp_attr qp_attr;
790 mutex_lock(&id_priv->qp_mutex);
791 if (!id_priv->id.qp) {
796 qp_attr.qp_state = IB_QPS_ERR;
797 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
799 mutex_unlock(&id_priv->qp_mutex);
803 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
804 struct ib_qp_attr *qp_attr, int *qp_attr_mask)
806 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
810 if (rdma_cap_eth_ah(id_priv->id.device, id_priv->id.port_num))
813 pkey = ib_addr_get_pkey(dev_addr);
815 ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
816 pkey, &qp_attr->pkey_index);
820 qp_attr->port_num = id_priv->id.port_num;
821 *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
823 if (id_priv->id.qp_type == IB_QPT_UD) {
824 ret = cma_set_qkey(id_priv, 0);
828 qp_attr->qkey = id_priv->qkey;
829 *qp_attr_mask |= IB_QP_QKEY;
831 qp_attr->qp_access_flags = 0;
832 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
837 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
840 struct rdma_id_private *id_priv;
843 id_priv = container_of(id, struct rdma_id_private, id);
844 if (rdma_cap_ib_cm(id->device, id->port_num)) {
845 if (!id_priv->cm_id.ib || (id_priv->id.qp_type == IB_QPT_UD))
846 ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
848 ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
851 if (qp_attr->qp_state == IB_QPS_RTR)
852 qp_attr->rq_psn = id_priv->seq_num;
853 } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
854 if (!id_priv->cm_id.iw) {
855 qp_attr->qp_access_flags = 0;
856 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
858 ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
865 EXPORT_SYMBOL(rdma_init_qp_attr);
867 static inline int cma_zero_addr(struct sockaddr *addr)
869 switch (addr->sa_family) {
871 return ipv4_is_zeronet(((struct sockaddr_in *)addr)->sin_addr.s_addr);
873 return ipv6_addr_any(&((struct sockaddr_in6 *) addr)->sin6_addr);
875 return ib_addr_any(&((struct sockaddr_ib *) addr)->sib_addr);
881 static inline int cma_loopback_addr(struct sockaddr *addr)
883 switch (addr->sa_family) {
885 return ipv4_is_loopback(((struct sockaddr_in *) addr)->sin_addr.s_addr);
887 return ipv6_addr_loopback(&((struct sockaddr_in6 *) addr)->sin6_addr);
889 return ib_addr_loopback(&((struct sockaddr_ib *) addr)->sib_addr);
895 static inline int cma_any_addr(struct sockaddr *addr)
897 return cma_zero_addr(addr) || cma_loopback_addr(addr);
900 static int cma_addr_cmp(struct sockaddr *src, struct sockaddr *dst)
902 if (src->sa_family != dst->sa_family)
905 switch (src->sa_family) {
907 return ((struct sockaddr_in *) src)->sin_addr.s_addr !=
908 ((struct sockaddr_in *) dst)->sin_addr.s_addr;
910 return ipv6_addr_cmp(&((struct sockaddr_in6 *) src)->sin6_addr,
911 &((struct sockaddr_in6 *) dst)->sin6_addr);
913 return ib_addr_cmp(&((struct sockaddr_ib *) src)->sib_addr,
914 &((struct sockaddr_ib *) dst)->sib_addr);
918 static __be16 cma_port(struct sockaddr *addr)
920 struct sockaddr_ib *sib;
922 switch (addr->sa_family) {
924 return ((struct sockaddr_in *) addr)->sin_port;
926 return ((struct sockaddr_in6 *) addr)->sin6_port;
928 sib = (struct sockaddr_ib *) addr;
929 return htons((u16) (be64_to_cpu(sib->sib_sid) &
930 be64_to_cpu(sib->sib_sid_mask)));
936 static inline int cma_any_port(struct sockaddr *addr)
938 return !cma_port(addr);
941 static void cma_save_ib_info(struct sockaddr *src_addr,
942 struct sockaddr *dst_addr,
943 struct rdma_cm_id *listen_id,
944 struct ib_sa_path_rec *path)
946 struct sockaddr_ib *listen_ib, *ib;
948 listen_ib = (struct sockaddr_ib *) &listen_id->route.addr.src_addr;
950 ib = (struct sockaddr_ib *)src_addr;
951 ib->sib_family = AF_IB;
953 ib->sib_pkey = path->pkey;
954 ib->sib_flowinfo = path->flow_label;
955 memcpy(&ib->sib_addr, &path->sgid, 16);
956 ib->sib_sid = path->service_id;
957 ib->sib_scope_id = 0;
959 ib->sib_pkey = listen_ib->sib_pkey;
960 ib->sib_flowinfo = listen_ib->sib_flowinfo;
961 ib->sib_addr = listen_ib->sib_addr;
962 ib->sib_sid = listen_ib->sib_sid;
963 ib->sib_scope_id = listen_ib->sib_scope_id;
965 ib->sib_sid_mask = cpu_to_be64(0xffffffffffffffffULL);
968 ib = (struct sockaddr_ib *)dst_addr;
969 ib->sib_family = AF_IB;
971 ib->sib_pkey = path->pkey;
972 ib->sib_flowinfo = path->flow_label;
973 memcpy(&ib->sib_addr, &path->dgid, 16);
978 static void cma_save_ip4_info(struct sockaddr *src_addr,
979 struct sockaddr *dst_addr,
983 struct sockaddr_in *ip4;
986 ip4 = (struct sockaddr_in *)src_addr;
987 ip4->sin_family = AF_INET;
988 ip4->sin_addr.s_addr = hdr->dst_addr.ip4.addr;
989 ip4->sin_port = local_port;
993 ip4 = (struct sockaddr_in *)dst_addr;
994 ip4->sin_family = AF_INET;
995 ip4->sin_addr.s_addr = hdr->src_addr.ip4.addr;
996 ip4->sin_port = hdr->port;
1000 static void cma_save_ip6_info(struct sockaddr *src_addr,
1001 struct sockaddr *dst_addr,
1002 struct cma_hdr *hdr,
1005 struct sockaddr_in6 *ip6;
1008 ip6 = (struct sockaddr_in6 *)src_addr;
1009 ip6->sin6_family = AF_INET6;
1010 ip6->sin6_addr = hdr->dst_addr.ip6;
1011 ip6->sin6_port = local_port;
1015 ip6 = (struct sockaddr_in6 *)dst_addr;
1016 ip6->sin6_family = AF_INET6;
1017 ip6->sin6_addr = hdr->src_addr.ip6;
1018 ip6->sin6_port = hdr->port;
1022 static u16 cma_port_from_service_id(__be64 service_id)
1024 return (u16)be64_to_cpu(service_id);
1027 static int cma_save_ip_info(struct sockaddr *src_addr,
1028 struct sockaddr *dst_addr,
1029 struct ib_cm_event *ib_event,
1032 struct cma_hdr *hdr;
1035 hdr = ib_event->private_data;
1036 if (hdr->cma_version != CMA_VERSION)
1039 port = htons(cma_port_from_service_id(service_id));
1041 switch (cma_get_ip_ver(hdr)) {
1043 cma_save_ip4_info(src_addr, dst_addr, hdr, port);
1046 cma_save_ip6_info(src_addr, dst_addr, hdr, port);
1049 return -EAFNOSUPPORT;
1055 static int cma_save_net_info(struct sockaddr *src_addr,
1056 struct sockaddr *dst_addr,
1057 struct rdma_cm_id *listen_id,
1058 struct ib_cm_event *ib_event,
1059 sa_family_t sa_family, __be64 service_id)
1061 if (sa_family == AF_IB) {
1062 if (ib_event->event == IB_CM_REQ_RECEIVED)
1063 cma_save_ib_info(src_addr, dst_addr, listen_id,
1064 ib_event->param.req_rcvd.primary_path);
1065 else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED)
1066 cma_save_ib_info(src_addr, dst_addr, listen_id, NULL);
1070 return cma_save_ip_info(src_addr, dst_addr, ib_event, service_id);
1073 static int cma_save_req_info(const struct ib_cm_event *ib_event,
1074 struct cma_req_info *req)
1076 const struct ib_cm_req_event_param *req_param =
1077 &ib_event->param.req_rcvd;
1078 const struct ib_cm_sidr_req_event_param *sidr_param =
1079 &ib_event->param.sidr_req_rcvd;
1081 switch (ib_event->event) {
1082 case IB_CM_REQ_RECEIVED:
1083 req->device = req_param->listen_id->device;
1084 req->port = req_param->port;
1085 memcpy(&req->local_gid, &req_param->primary_path->sgid,
1086 sizeof(req->local_gid));
1087 req->has_gid = true;
1088 req->service_id = req_param->primary_path->service_id;
1089 req->pkey = be16_to_cpu(req_param->primary_path->pkey);
1091 case IB_CM_SIDR_REQ_RECEIVED:
1092 req->device = sidr_param->listen_id->device;
1093 req->port = sidr_param->port;
1094 req->has_gid = false;
1095 req->service_id = sidr_param->service_id;
1096 req->pkey = sidr_param->pkey;
1105 static bool validate_ipv4_net_dev(struct net_device *net_dev,
1106 const struct sockaddr_in *dst_addr,
1107 const struct sockaddr_in *src_addr)
1109 __be32 daddr = dst_addr->sin_addr.s_addr,
1110 saddr = src_addr->sin_addr.s_addr;
1111 struct fib_result res;
1116 if (ipv4_is_multicast(saddr) || ipv4_is_lbcast(saddr) ||
1117 ipv4_is_lbcast(daddr) || ipv4_is_zeronet(saddr) ||
1118 ipv4_is_zeronet(daddr) || ipv4_is_loopback(daddr) ||
1119 ipv4_is_loopback(saddr))
1122 memset(&fl4, 0, sizeof(fl4));
1123 fl4.flowi4_iif = net_dev->ifindex;
1128 err = fib_lookup(dev_net(net_dev), &fl4, &res, 0);
1132 ret = FIB_RES_DEV(res) == net_dev;
1138 static bool validate_ipv6_net_dev(struct net_device *net_dev,
1139 const struct sockaddr_in6 *dst_addr,
1140 const struct sockaddr_in6 *src_addr)
1142 #if IS_ENABLED(CONFIG_IPV6)
1143 const int strict = ipv6_addr_type(&dst_addr->sin6_addr) &
1144 IPV6_ADDR_LINKLOCAL;
1145 struct rt6_info *rt = rt6_lookup(dev_net(net_dev), &dst_addr->sin6_addr,
1146 &src_addr->sin6_addr, net_dev->ifindex,
1153 ret = rt->rt6i_idev->dev == net_dev;
1162 static bool validate_net_dev(struct net_device *net_dev,
1163 const struct sockaddr *daddr,
1164 const struct sockaddr *saddr)
1166 const struct sockaddr_in *daddr4 = (const struct sockaddr_in *)daddr;
1167 const struct sockaddr_in *saddr4 = (const struct sockaddr_in *)saddr;
1168 const struct sockaddr_in6 *daddr6 = (const struct sockaddr_in6 *)daddr;
1169 const struct sockaddr_in6 *saddr6 = (const struct sockaddr_in6 *)saddr;
1171 switch (daddr->sa_family) {
1173 return saddr->sa_family == AF_INET &&
1174 validate_ipv4_net_dev(net_dev, daddr4, saddr4);
1177 return saddr->sa_family == AF_INET6 &&
1178 validate_ipv6_net_dev(net_dev, daddr6, saddr6);
1185 static struct net_device *cma_get_net_dev(struct ib_cm_event *ib_event,
1186 const struct cma_req_info *req)
1188 struct sockaddr_storage listen_addr_storage, src_addr_storage;
1189 struct sockaddr *listen_addr = (struct sockaddr *)&listen_addr_storage,
1190 *src_addr = (struct sockaddr *)&src_addr_storage;
1191 struct net_device *net_dev;
1192 const union ib_gid *gid = req->has_gid ? &req->local_gid : NULL;
1195 err = cma_save_ip_info(listen_addr, src_addr, ib_event,
1198 return ERR_PTR(err);
1200 net_dev = ib_get_net_dev_by_params(req->device, req->port, req->pkey,
1203 return ERR_PTR(-ENODEV);
1205 if (!validate_net_dev(net_dev, listen_addr, src_addr)) {
1207 return ERR_PTR(-EHOSTUNREACH);
1213 static enum rdma_port_space rdma_ps_from_service_id(__be64 service_id)
1215 return (be64_to_cpu(service_id) >> 16) & 0xffff;
1218 static bool cma_match_private_data(struct rdma_id_private *id_priv,
1219 const struct cma_hdr *hdr)
1221 struct sockaddr *addr = cma_src_addr(id_priv);
1223 struct in6_addr ip6_addr;
1225 if (cma_any_addr(addr) && !id_priv->afonly)
1228 switch (addr->sa_family) {
1230 ip4_addr = ((struct sockaddr_in *)addr)->sin_addr.s_addr;
1231 if (cma_get_ip_ver(hdr) != 4)
1233 if (!cma_any_addr(addr) &&
1234 hdr->dst_addr.ip4.addr != ip4_addr)
1238 ip6_addr = ((struct sockaddr_in6 *)addr)->sin6_addr;
1239 if (cma_get_ip_ver(hdr) != 6)
1241 if (!cma_any_addr(addr) &&
1242 memcmp(&hdr->dst_addr.ip6, &ip6_addr, sizeof(ip6_addr)))
1254 static bool cma_protocol_roce_dev_port(struct ib_device *device, int port_num)
1256 enum rdma_link_layer ll = rdma_port_get_link_layer(device, port_num);
1257 enum rdma_transport_type transport =
1258 rdma_node_get_transport(device->node_type);
1260 return ll == IB_LINK_LAYER_ETHERNET && transport == RDMA_TRANSPORT_IB;
1263 static bool cma_protocol_roce(const struct rdma_cm_id *id)
1265 struct ib_device *device = id->device;
1266 const int port_num = id->port_num ?: rdma_start_port(device);
1268 return cma_protocol_roce_dev_port(device, port_num);
1271 static bool cma_match_net_dev(const struct rdma_id_private *id_priv,
1272 const struct net_device *net_dev)
1274 const struct rdma_addr *addr = &id_priv->id.route.addr;
1277 /* This request is an AF_IB request or a RoCE request */
1278 return addr->src_addr.ss_family == AF_IB ||
1279 cma_protocol_roce(&id_priv->id);
1281 return !addr->dev_addr.bound_dev_if ||
1282 (net_eq(dev_net(net_dev), addr->dev_addr.net) &&
1283 addr->dev_addr.bound_dev_if == net_dev->ifindex);
1286 static struct rdma_id_private *cma_find_listener(
1287 const struct rdma_bind_list *bind_list,
1288 const struct ib_cm_id *cm_id,
1289 const struct ib_cm_event *ib_event,
1290 const struct cma_req_info *req,
1291 const struct net_device *net_dev)
1293 struct rdma_id_private *id_priv, *id_priv_dev;
1296 return ERR_PTR(-EINVAL);
1298 hlist_for_each_entry(id_priv, &bind_list->owners, node) {
1299 if (cma_match_private_data(id_priv, ib_event->private_data)) {
1300 if (id_priv->id.device == cm_id->device &&
1301 cma_match_net_dev(id_priv, net_dev))
1303 list_for_each_entry(id_priv_dev,
1304 &id_priv->listen_list,
1306 if (id_priv_dev->id.device == cm_id->device &&
1307 cma_match_net_dev(id_priv_dev, net_dev))
1313 return ERR_PTR(-EINVAL);
1316 static struct rdma_id_private *cma_id_from_event(struct ib_cm_id *cm_id,
1317 struct ib_cm_event *ib_event,
1318 struct net_device **net_dev)
1320 struct cma_req_info req;
1321 struct rdma_bind_list *bind_list;
1322 struct rdma_id_private *id_priv;
1325 err = cma_save_req_info(ib_event, &req);
1327 return ERR_PTR(err);
1329 *net_dev = cma_get_net_dev(ib_event, &req);
1330 if (IS_ERR(*net_dev)) {
1331 if (PTR_ERR(*net_dev) == -EAFNOSUPPORT) {
1332 /* Assuming the protocol is AF_IB */
1334 } else if (cma_protocol_roce_dev_port(req.device, req.port)) {
1335 /* TODO find the net dev matching the request parameters
1336 * through the RoCE GID table */
1339 return ERR_CAST(*net_dev);
1343 bind_list = cma_ps_find(*net_dev ? dev_net(*net_dev) : &init_net,
1344 rdma_ps_from_service_id(req.service_id),
1345 cma_port_from_service_id(req.service_id));
1346 id_priv = cma_find_listener(bind_list, cm_id, ib_event, &req, *net_dev);
1347 if (IS_ERR(id_priv) && *net_dev) {
1355 static inline int cma_user_data_offset(struct rdma_id_private *id_priv)
1357 return cma_family(id_priv) == AF_IB ? 0 : sizeof(struct cma_hdr);
1360 static void cma_cancel_route(struct rdma_id_private *id_priv)
1362 if (rdma_cap_ib_sa(id_priv->id.device, id_priv->id.port_num)) {
1364 ib_sa_cancel_query(id_priv->query_id, id_priv->query);
1368 static void cma_cancel_listens(struct rdma_id_private *id_priv)
1370 struct rdma_id_private *dev_id_priv;
1373 * Remove from listen_any_list to prevent added devices from spawning
1374 * additional listen requests.
1377 list_del(&id_priv->list);
1379 while (!list_empty(&id_priv->listen_list)) {
1380 dev_id_priv = list_entry(id_priv->listen_list.next,
1381 struct rdma_id_private, listen_list);
1382 /* sync with device removal to avoid duplicate destruction */
1383 list_del_init(&dev_id_priv->list);
1384 list_del(&dev_id_priv->listen_list);
1385 mutex_unlock(&lock);
1387 rdma_destroy_id(&dev_id_priv->id);
1390 mutex_unlock(&lock);
1393 static void cma_cancel_operation(struct rdma_id_private *id_priv,
1394 enum rdma_cm_state state)
1397 case RDMA_CM_ADDR_QUERY:
1398 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
1400 case RDMA_CM_ROUTE_QUERY:
1401 cma_cancel_route(id_priv);
1403 case RDMA_CM_LISTEN:
1404 if (cma_any_addr(cma_src_addr(id_priv)) && !id_priv->cma_dev)
1405 cma_cancel_listens(id_priv);
1412 static void cma_release_port(struct rdma_id_private *id_priv)
1414 struct rdma_bind_list *bind_list = id_priv->bind_list;
1415 struct net *net = id_priv->id.route.addr.dev_addr.net;
1421 hlist_del(&id_priv->node);
1422 if (hlist_empty(&bind_list->owners)) {
1423 cma_ps_remove(net, bind_list->ps, bind_list->port);
1426 mutex_unlock(&lock);
1429 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
1431 struct cma_multicast *mc;
1433 while (!list_empty(&id_priv->mc_list)) {
1434 mc = container_of(id_priv->mc_list.next,
1435 struct cma_multicast, list);
1436 list_del(&mc->list);
1437 if (rdma_cap_ib_mcast(id_priv->cma_dev->device,
1438 id_priv->id.port_num)) {
1439 ib_sa_free_multicast(mc->multicast.ib);
1442 kref_put(&mc->mcref, release_mc);
1446 void rdma_destroy_id(struct rdma_cm_id *id)
1448 struct rdma_id_private *id_priv;
1449 enum rdma_cm_state state;
1451 id_priv = container_of(id, struct rdma_id_private, id);
1452 state = cma_exch(id_priv, RDMA_CM_DESTROYING);
1453 cma_cancel_operation(id_priv, state);
1456 * Wait for any active callback to finish. New callbacks will find
1457 * the id_priv state set to destroying and abort.
1459 mutex_lock(&id_priv->handler_mutex);
1460 mutex_unlock(&id_priv->handler_mutex);
1462 if (id_priv->cma_dev) {
1463 if (rdma_cap_ib_cm(id_priv->id.device, 1)) {
1464 if (id_priv->cm_id.ib)
1465 ib_destroy_cm_id(id_priv->cm_id.ib);
1466 } else if (rdma_cap_iw_cm(id_priv->id.device, 1)) {
1467 if (id_priv->cm_id.iw)
1468 iw_destroy_cm_id(id_priv->cm_id.iw);
1470 cma_leave_mc_groups(id_priv);
1471 cma_release_dev(id_priv);
1474 cma_release_port(id_priv);
1475 cma_deref_id(id_priv);
1476 wait_for_completion(&id_priv->comp);
1478 if (id_priv->internal_id)
1479 cma_deref_id(id_priv->id.context);
1481 kfree(id_priv->id.route.path_rec);
1482 put_net(id_priv->id.route.addr.dev_addr.net);
1485 EXPORT_SYMBOL(rdma_destroy_id);
1487 static int cma_rep_recv(struct rdma_id_private *id_priv)
1491 ret = cma_modify_qp_rtr(id_priv, NULL);
1495 ret = cma_modify_qp_rts(id_priv, NULL);
1499 ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
1505 cma_modify_qp_err(id_priv);
1506 ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
1511 static void cma_set_rep_event_data(struct rdma_cm_event *event,
1512 struct ib_cm_rep_event_param *rep_data,
1515 event->param.conn.private_data = private_data;
1516 event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
1517 event->param.conn.responder_resources = rep_data->responder_resources;
1518 event->param.conn.initiator_depth = rep_data->initiator_depth;
1519 event->param.conn.flow_control = rep_data->flow_control;
1520 event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
1521 event->param.conn.srq = rep_data->srq;
1522 event->param.conn.qp_num = rep_data->remote_qpn;
1525 static int cma_ib_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1527 struct rdma_id_private *id_priv = cm_id->context;
1528 struct rdma_cm_event event;
1531 if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
1532 cma_disable_callback(id_priv, RDMA_CM_CONNECT)) ||
1533 (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
1534 cma_disable_callback(id_priv, RDMA_CM_DISCONNECT)))
1537 memset(&event, 0, sizeof event);
1538 switch (ib_event->event) {
1539 case IB_CM_REQ_ERROR:
1540 case IB_CM_REP_ERROR:
1541 event.event = RDMA_CM_EVENT_UNREACHABLE;
1542 event.status = -ETIMEDOUT;
1544 case IB_CM_REP_RECEIVED:
1545 if (id_priv->id.qp) {
1546 event.status = cma_rep_recv(id_priv);
1547 event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
1548 RDMA_CM_EVENT_ESTABLISHED;
1550 event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
1552 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
1553 ib_event->private_data);
1555 case IB_CM_RTU_RECEIVED:
1556 case IB_CM_USER_ESTABLISHED:
1557 event.event = RDMA_CM_EVENT_ESTABLISHED;
1559 case IB_CM_DREQ_ERROR:
1560 event.status = -ETIMEDOUT; /* fall through */
1561 case IB_CM_DREQ_RECEIVED:
1562 case IB_CM_DREP_RECEIVED:
1563 if (!cma_comp_exch(id_priv, RDMA_CM_CONNECT,
1564 RDMA_CM_DISCONNECT))
1566 event.event = RDMA_CM_EVENT_DISCONNECTED;
1568 case IB_CM_TIMEWAIT_EXIT:
1569 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
1571 case IB_CM_MRA_RECEIVED:
1574 case IB_CM_REJ_RECEIVED:
1575 cma_modify_qp_err(id_priv);
1576 event.status = ib_event->param.rej_rcvd.reason;
1577 event.event = RDMA_CM_EVENT_REJECTED;
1578 event.param.conn.private_data = ib_event->private_data;
1579 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
1582 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
1587 ret = id_priv->id.event_handler(&id_priv->id, &event);
1589 /* Destroy the CM ID by returning a non-zero value. */
1590 id_priv->cm_id.ib = NULL;
1591 cma_exch(id_priv, RDMA_CM_DESTROYING);
1592 mutex_unlock(&id_priv->handler_mutex);
1593 rdma_destroy_id(&id_priv->id);
1597 mutex_unlock(&id_priv->handler_mutex);
1601 static struct rdma_id_private *cma_new_conn_id(struct rdma_cm_id *listen_id,
1602 struct ib_cm_event *ib_event,
1603 struct net_device *net_dev)
1605 struct rdma_id_private *id_priv;
1606 struct rdma_cm_id *id;
1607 struct rdma_route *rt;
1608 const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
1609 const __be64 service_id =
1610 ib_event->param.req_rcvd.primary_path->service_id;
1613 id = rdma_create_id(listen_id->route.addr.dev_addr.net,
1614 listen_id->event_handler, listen_id->context,
1615 listen_id->ps, ib_event->param.req_rcvd.qp_type);
1619 id_priv = container_of(id, struct rdma_id_private, id);
1620 if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
1621 (struct sockaddr *)&id->route.addr.dst_addr,
1622 listen_id, ib_event, ss_family, service_id))
1626 rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
1627 rt->path_rec = kmalloc(sizeof *rt->path_rec * rt->num_paths,
1632 rt->path_rec[0] = *ib_event->param.req_rcvd.primary_path;
1633 if (rt->num_paths == 2)
1634 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
1637 ret = rdma_copy_addr(&rt->addr.dev_addr, net_dev, NULL);
1641 if (!cma_protocol_roce(listen_id) &&
1642 cma_any_addr(cma_src_addr(id_priv))) {
1643 rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
1644 rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
1645 ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
1646 } else if (!cma_any_addr(cma_src_addr(id_priv))) {
1647 ret = cma_translate_addr(cma_src_addr(id_priv), &rt->addr.dev_addr);
1652 rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
1654 id_priv->state = RDMA_CM_CONNECT;
1658 rdma_destroy_id(id);
1662 static struct rdma_id_private *cma_new_udp_id(struct rdma_cm_id *listen_id,
1663 struct ib_cm_event *ib_event,
1664 struct net_device *net_dev)
1666 struct rdma_id_private *id_priv;
1667 struct rdma_cm_id *id;
1668 const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
1669 struct net *net = listen_id->route.addr.dev_addr.net;
1672 id = rdma_create_id(net, listen_id->event_handler, listen_id->context,
1673 listen_id->ps, IB_QPT_UD);
1677 id_priv = container_of(id, struct rdma_id_private, id);
1678 if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
1679 (struct sockaddr *)&id->route.addr.dst_addr,
1680 listen_id, ib_event, ss_family,
1681 ib_event->param.sidr_req_rcvd.service_id))
1685 ret = rdma_copy_addr(&id->route.addr.dev_addr, net_dev, NULL);
1689 if (!cma_any_addr(cma_src_addr(id_priv))) {
1690 ret = cma_translate_addr(cma_src_addr(id_priv),
1691 &id->route.addr.dev_addr);
1697 id_priv->state = RDMA_CM_CONNECT;
1700 rdma_destroy_id(id);
1704 static void cma_set_req_event_data(struct rdma_cm_event *event,
1705 struct ib_cm_req_event_param *req_data,
1706 void *private_data, int offset)
1708 event->param.conn.private_data = private_data + offset;
1709 event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
1710 event->param.conn.responder_resources = req_data->responder_resources;
1711 event->param.conn.initiator_depth = req_data->initiator_depth;
1712 event->param.conn.flow_control = req_data->flow_control;
1713 event->param.conn.retry_count = req_data->retry_count;
1714 event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
1715 event->param.conn.srq = req_data->srq;
1716 event->param.conn.qp_num = req_data->remote_qpn;
1719 static int cma_check_req_qp_type(struct rdma_cm_id *id, struct ib_cm_event *ib_event)
1721 return (((ib_event->event == IB_CM_REQ_RECEIVED) &&
1722 (ib_event->param.req_rcvd.qp_type == id->qp_type)) ||
1723 ((ib_event->event == IB_CM_SIDR_REQ_RECEIVED) &&
1724 (id->qp_type == IB_QPT_UD)) ||
1728 static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1730 struct rdma_id_private *listen_id, *conn_id;
1731 struct rdma_cm_event event;
1732 struct net_device *net_dev;
1735 listen_id = cma_id_from_event(cm_id, ib_event, &net_dev);
1736 if (IS_ERR(listen_id))
1737 return PTR_ERR(listen_id);
1739 if (!cma_check_req_qp_type(&listen_id->id, ib_event)) {
1744 if (cma_disable_callback(listen_id, RDMA_CM_LISTEN)) {
1745 ret = -ECONNABORTED;
1749 memset(&event, 0, sizeof event);
1750 offset = cma_user_data_offset(listen_id);
1751 event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1752 if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) {
1753 conn_id = cma_new_udp_id(&listen_id->id, ib_event, net_dev);
1754 event.param.ud.private_data = ib_event->private_data + offset;
1755 event.param.ud.private_data_len =
1756 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
1758 conn_id = cma_new_conn_id(&listen_id->id, ib_event, net_dev);
1759 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
1760 ib_event->private_data, offset);
1767 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1768 ret = cma_acquire_dev(conn_id, listen_id);
1772 conn_id->cm_id.ib = cm_id;
1773 cm_id->context = conn_id;
1774 cm_id->cm_handler = cma_ib_handler;
1777 * Protect against the user destroying conn_id from another thread
1778 * until we're done accessing it.
1780 atomic_inc(&conn_id->refcount);
1781 ret = conn_id->id.event_handler(&conn_id->id, &event);
1785 * Acquire mutex to prevent user executing rdma_destroy_id()
1786 * while we're accessing the cm_id.
1789 if (cma_comp(conn_id, RDMA_CM_CONNECT) &&
1790 (conn_id->id.qp_type != IB_QPT_UD))
1791 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
1792 mutex_unlock(&lock);
1793 mutex_unlock(&conn_id->handler_mutex);
1794 mutex_unlock(&listen_id->handler_mutex);
1795 cma_deref_id(conn_id);
1801 cma_deref_id(conn_id);
1802 /* Destroy the CM ID by returning a non-zero value. */
1803 conn_id->cm_id.ib = NULL;
1805 cma_exch(conn_id, RDMA_CM_DESTROYING);
1806 mutex_unlock(&conn_id->handler_mutex);
1808 mutex_unlock(&listen_id->handler_mutex);
1810 rdma_destroy_id(&conn_id->id);
1819 __be64 rdma_get_service_id(struct rdma_cm_id *id, struct sockaddr *addr)
1821 if (addr->sa_family == AF_IB)
1822 return ((struct sockaddr_ib *) addr)->sib_sid;
1824 return cpu_to_be64(((u64)id->ps << 16) + be16_to_cpu(cma_port(addr)));
1826 EXPORT_SYMBOL(rdma_get_service_id);
1828 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
1830 struct rdma_id_private *id_priv = iw_id->context;
1831 struct rdma_cm_event event;
1833 struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
1834 struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
1836 if (cma_disable_callback(id_priv, RDMA_CM_CONNECT))
1839 memset(&event, 0, sizeof event);
1840 switch (iw_event->event) {
1841 case IW_CM_EVENT_CLOSE:
1842 event.event = RDMA_CM_EVENT_DISCONNECTED;
1844 case IW_CM_EVENT_CONNECT_REPLY:
1845 memcpy(cma_src_addr(id_priv), laddr,
1846 rdma_addr_size(laddr));
1847 memcpy(cma_dst_addr(id_priv), raddr,
1848 rdma_addr_size(raddr));
1849 switch (iw_event->status) {
1851 event.event = RDMA_CM_EVENT_ESTABLISHED;
1852 event.param.conn.initiator_depth = iw_event->ird;
1853 event.param.conn.responder_resources = iw_event->ord;
1857 event.event = RDMA_CM_EVENT_REJECTED;
1860 event.event = RDMA_CM_EVENT_UNREACHABLE;
1863 event.event = RDMA_CM_EVENT_CONNECT_ERROR;
1867 case IW_CM_EVENT_ESTABLISHED:
1868 event.event = RDMA_CM_EVENT_ESTABLISHED;
1869 event.param.conn.initiator_depth = iw_event->ird;
1870 event.param.conn.responder_resources = iw_event->ord;
1876 event.status = iw_event->status;
1877 event.param.conn.private_data = iw_event->private_data;
1878 event.param.conn.private_data_len = iw_event->private_data_len;
1879 ret = id_priv->id.event_handler(&id_priv->id, &event);
1881 /* Destroy the CM ID by returning a non-zero value. */
1882 id_priv->cm_id.iw = NULL;
1883 cma_exch(id_priv, RDMA_CM_DESTROYING);
1884 mutex_unlock(&id_priv->handler_mutex);
1885 rdma_destroy_id(&id_priv->id);
1889 mutex_unlock(&id_priv->handler_mutex);
1893 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
1894 struct iw_cm_event *iw_event)
1896 struct rdma_cm_id *new_cm_id;
1897 struct rdma_id_private *listen_id, *conn_id;
1898 struct rdma_cm_event event;
1900 struct ib_device_attr attr;
1901 struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
1902 struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
1904 listen_id = cm_id->context;
1905 if (cma_disable_callback(listen_id, RDMA_CM_LISTEN))
1906 return -ECONNABORTED;
1908 /* Create a new RDMA id for the new IW CM ID */
1909 new_cm_id = rdma_create_id(listen_id->id.route.addr.dev_addr.net,
1910 listen_id->id.event_handler,
1911 listen_id->id.context,
1912 RDMA_PS_TCP, IB_QPT_RC);
1913 if (IS_ERR(new_cm_id)) {
1917 conn_id = container_of(new_cm_id, struct rdma_id_private, id);
1918 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1919 conn_id->state = RDMA_CM_CONNECT;
1921 ret = rdma_translate_ip(laddr, &conn_id->id.route.addr.dev_addr, NULL);
1923 mutex_unlock(&conn_id->handler_mutex);
1924 rdma_destroy_id(new_cm_id);
1928 ret = cma_acquire_dev(conn_id, listen_id);
1930 mutex_unlock(&conn_id->handler_mutex);
1931 rdma_destroy_id(new_cm_id);
1935 conn_id->cm_id.iw = cm_id;
1936 cm_id->context = conn_id;
1937 cm_id->cm_handler = cma_iw_handler;
1939 memcpy(cma_src_addr(conn_id), laddr, rdma_addr_size(laddr));
1940 memcpy(cma_dst_addr(conn_id), raddr, rdma_addr_size(raddr));
1942 ret = ib_query_device(conn_id->id.device, &attr);
1944 mutex_unlock(&conn_id->handler_mutex);
1945 rdma_destroy_id(new_cm_id);
1949 memset(&event, 0, sizeof event);
1950 event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1951 event.param.conn.private_data = iw_event->private_data;
1952 event.param.conn.private_data_len = iw_event->private_data_len;
1953 event.param.conn.initiator_depth = iw_event->ird;
1954 event.param.conn.responder_resources = iw_event->ord;
1957 * Protect against the user destroying conn_id from another thread
1958 * until we're done accessing it.
1960 atomic_inc(&conn_id->refcount);
1961 ret = conn_id->id.event_handler(&conn_id->id, &event);
1963 /* User wants to destroy the CM ID */
1964 conn_id->cm_id.iw = NULL;
1965 cma_exch(conn_id, RDMA_CM_DESTROYING);
1966 mutex_unlock(&conn_id->handler_mutex);
1967 cma_deref_id(conn_id);
1968 rdma_destroy_id(&conn_id->id);
1972 mutex_unlock(&conn_id->handler_mutex);
1973 cma_deref_id(conn_id);
1976 mutex_unlock(&listen_id->handler_mutex);
1980 static int cma_ib_listen(struct rdma_id_private *id_priv)
1982 struct sockaddr *addr;
1983 struct ib_cm_id *id;
1986 addr = cma_src_addr(id_priv);
1987 svc_id = rdma_get_service_id(&id_priv->id, addr);
1988 id = ib_cm_insert_listen(id_priv->id.device, cma_req_handler, svc_id);
1991 id_priv->cm_id.ib = id;
1996 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
1999 struct iw_cm_id *id;
2001 id = iw_create_cm_id(id_priv->id.device,
2002 iw_conn_req_handler,
2007 id->tos = id_priv->tos;
2008 id_priv->cm_id.iw = id;
2010 memcpy(&id_priv->cm_id.iw->local_addr, cma_src_addr(id_priv),
2011 rdma_addr_size(cma_src_addr(id_priv)));
2013 ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
2016 iw_destroy_cm_id(id_priv->cm_id.iw);
2017 id_priv->cm_id.iw = NULL;
2023 static int cma_listen_handler(struct rdma_cm_id *id,
2024 struct rdma_cm_event *event)
2026 struct rdma_id_private *id_priv = id->context;
2028 id->context = id_priv->id.context;
2029 id->event_handler = id_priv->id.event_handler;
2030 return id_priv->id.event_handler(id, event);
2033 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
2034 struct cma_device *cma_dev)
2036 struct rdma_id_private *dev_id_priv;
2037 struct rdma_cm_id *id;
2038 struct net *net = id_priv->id.route.addr.dev_addr.net;
2041 if (cma_family(id_priv) == AF_IB && !rdma_cap_ib_cm(cma_dev->device, 1))
2044 id = rdma_create_id(net, cma_listen_handler, id_priv, id_priv->id.ps,
2045 id_priv->id.qp_type);
2049 dev_id_priv = container_of(id, struct rdma_id_private, id);
2051 dev_id_priv->state = RDMA_CM_ADDR_BOUND;
2052 memcpy(cma_src_addr(dev_id_priv), cma_src_addr(id_priv),
2053 rdma_addr_size(cma_src_addr(id_priv)));
2055 cma_attach_to_dev(dev_id_priv, cma_dev);
2056 list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
2057 atomic_inc(&id_priv->refcount);
2058 dev_id_priv->internal_id = 1;
2059 dev_id_priv->afonly = id_priv->afonly;
2061 ret = rdma_listen(id, id_priv->backlog);
2063 printk(KERN_WARNING "RDMA CMA: cma_listen_on_dev, error %d, "
2064 "listening on device %s\n", ret, cma_dev->device->name);
2067 static void cma_listen_on_all(struct rdma_id_private *id_priv)
2069 struct cma_device *cma_dev;
2072 list_add_tail(&id_priv->list, &listen_any_list);
2073 list_for_each_entry(cma_dev, &dev_list, list)
2074 cma_listen_on_dev(id_priv, cma_dev);
2075 mutex_unlock(&lock);
2078 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
2080 struct rdma_id_private *id_priv;
2082 id_priv = container_of(id, struct rdma_id_private, id);
2083 id_priv->tos = (u8) tos;
2085 EXPORT_SYMBOL(rdma_set_service_type);
2087 static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec,
2090 struct cma_work *work = context;
2091 struct rdma_route *route;
2093 route = &work->id->id.route;
2096 route->num_paths = 1;
2097 *route->path_rec = *path_rec;
2099 work->old_state = RDMA_CM_ROUTE_QUERY;
2100 work->new_state = RDMA_CM_ADDR_RESOLVED;
2101 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
2102 work->event.status = status;
2105 queue_work(cma_wq, &work->work);
2108 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms,
2109 struct cma_work *work)
2111 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2112 struct ib_sa_path_rec path_rec;
2113 ib_sa_comp_mask comp_mask;
2114 struct sockaddr_in6 *sin6;
2115 struct sockaddr_ib *sib;
2117 memset(&path_rec, 0, sizeof path_rec);
2118 rdma_addr_get_sgid(dev_addr, &path_rec.sgid);
2119 rdma_addr_get_dgid(dev_addr, &path_rec.dgid);
2120 path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
2121 path_rec.numb_path = 1;
2122 path_rec.reversible = 1;
2123 path_rec.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
2125 comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
2126 IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
2127 IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
2129 switch (cma_family(id_priv)) {
2131 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
2132 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
2135 sin6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
2136 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
2137 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2140 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
2141 path_rec.traffic_class = (u8) (be32_to_cpu(sib->sib_flowinfo) >> 20);
2142 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2146 id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
2147 id_priv->id.port_num, &path_rec,
2148 comp_mask, timeout_ms,
2149 GFP_KERNEL, cma_query_handler,
2150 work, &id_priv->query);
2152 return (id_priv->query_id < 0) ? id_priv->query_id : 0;
2155 static void cma_work_handler(struct work_struct *_work)
2157 struct cma_work *work = container_of(_work, struct cma_work, work);
2158 struct rdma_id_private *id_priv = work->id;
2161 mutex_lock(&id_priv->handler_mutex);
2162 if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
2165 if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
2166 cma_exch(id_priv, RDMA_CM_DESTROYING);
2170 mutex_unlock(&id_priv->handler_mutex);
2171 cma_deref_id(id_priv);
2173 rdma_destroy_id(&id_priv->id);
2177 static void cma_ndev_work_handler(struct work_struct *_work)
2179 struct cma_ndev_work *work = container_of(_work, struct cma_ndev_work, work);
2180 struct rdma_id_private *id_priv = work->id;
2183 mutex_lock(&id_priv->handler_mutex);
2184 if (id_priv->state == RDMA_CM_DESTROYING ||
2185 id_priv->state == RDMA_CM_DEVICE_REMOVAL)
2188 if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
2189 cma_exch(id_priv, RDMA_CM_DESTROYING);
2194 mutex_unlock(&id_priv->handler_mutex);
2195 cma_deref_id(id_priv);
2197 rdma_destroy_id(&id_priv->id);
2201 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms)
2203 struct rdma_route *route = &id_priv->id.route;
2204 struct cma_work *work;
2207 work = kzalloc(sizeof *work, GFP_KERNEL);
2212 INIT_WORK(&work->work, cma_work_handler);
2213 work->old_state = RDMA_CM_ROUTE_QUERY;
2214 work->new_state = RDMA_CM_ROUTE_RESOLVED;
2215 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
2217 route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
2218 if (!route->path_rec) {
2223 ret = cma_query_ib_route(id_priv, timeout_ms, work);
2229 kfree(route->path_rec);
2230 route->path_rec = NULL;
2236 int rdma_set_ib_paths(struct rdma_cm_id *id,
2237 struct ib_sa_path_rec *path_rec, int num_paths)
2239 struct rdma_id_private *id_priv;
2242 id_priv = container_of(id, struct rdma_id_private, id);
2243 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
2244 RDMA_CM_ROUTE_RESOLVED))
2247 id->route.path_rec = kmemdup(path_rec, sizeof *path_rec * num_paths,
2249 if (!id->route.path_rec) {
2254 id->route.num_paths = num_paths;
2257 cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED);
2260 EXPORT_SYMBOL(rdma_set_ib_paths);
2262 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms)
2264 struct cma_work *work;
2266 work = kzalloc(sizeof *work, GFP_KERNEL);
2271 INIT_WORK(&work->work, cma_work_handler);
2272 work->old_state = RDMA_CM_ROUTE_QUERY;
2273 work->new_state = RDMA_CM_ROUTE_RESOLVED;
2274 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
2275 queue_work(cma_wq, &work->work);
2279 static int iboe_tos_to_sl(struct net_device *ndev, int tos)
2282 struct net_device *dev;
2284 prio = rt_tos2priority(tos);
2285 dev = ndev->priv_flags & IFF_802_1Q_VLAN ?
2286 vlan_dev_real_dev(ndev) : ndev;
2289 return netdev_get_prio_tc_map(dev, prio);
2291 #if IS_ENABLED(CONFIG_VLAN_8021Q)
2292 if (ndev->priv_flags & IFF_802_1Q_VLAN)
2293 return (vlan_dev_get_egress_qos_mask(ndev, prio) &
2294 VLAN_PRIO_MASK) >> VLAN_PRIO_SHIFT;
2299 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
2301 struct rdma_route *route = &id_priv->id.route;
2302 struct rdma_addr *addr = &route->addr;
2303 struct cma_work *work;
2305 struct net_device *ndev = NULL;
2308 work = kzalloc(sizeof *work, GFP_KERNEL);
2313 INIT_WORK(&work->work, cma_work_handler);
2315 route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
2316 if (!route->path_rec) {
2321 route->num_paths = 1;
2323 if (addr->dev_addr.bound_dev_if) {
2324 ndev = dev_get_by_index(&init_net, addr->dev_addr.bound_dev_if);
2325 route->path_rec->net = &init_net;
2326 route->path_rec->ifindex = addr->dev_addr.bound_dev_if;
2333 memcpy(route->path_rec->dmac, addr->dev_addr.dst_dev_addr, ETH_ALEN);
2335 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
2336 &route->path_rec->sgid);
2337 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.dst_addr,
2338 &route->path_rec->dgid);
2340 route->path_rec->hop_limit = 1;
2341 route->path_rec->reversible = 1;
2342 route->path_rec->pkey = cpu_to_be16(0xffff);
2343 route->path_rec->mtu_selector = IB_SA_EQ;
2344 route->path_rec->sl = iboe_tos_to_sl(ndev, id_priv->tos);
2345 route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
2346 route->path_rec->rate_selector = IB_SA_EQ;
2347 route->path_rec->rate = iboe_get_rate(ndev);
2349 route->path_rec->packet_life_time_selector = IB_SA_EQ;
2350 route->path_rec->packet_life_time = CMA_IBOE_PACKET_LIFETIME;
2351 if (!route->path_rec->mtu) {
2356 work->old_state = RDMA_CM_ROUTE_QUERY;
2357 work->new_state = RDMA_CM_ROUTE_RESOLVED;
2358 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
2359 work->event.status = 0;
2361 queue_work(cma_wq, &work->work);
2366 kfree(route->path_rec);
2367 route->path_rec = NULL;
2373 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
2375 struct rdma_id_private *id_priv;
2378 id_priv = container_of(id, struct rdma_id_private, id);
2379 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY))
2382 atomic_inc(&id_priv->refcount);
2383 if (rdma_cap_ib_sa(id->device, id->port_num))
2384 ret = cma_resolve_ib_route(id_priv, timeout_ms);
2385 else if (rdma_protocol_roce(id->device, id->port_num))
2386 ret = cma_resolve_iboe_route(id_priv);
2387 else if (rdma_protocol_iwarp(id->device, id->port_num))
2388 ret = cma_resolve_iw_route(id_priv, timeout_ms);
2397 cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED);
2398 cma_deref_id(id_priv);
2401 EXPORT_SYMBOL(rdma_resolve_route);
2403 static void cma_set_loopback(struct sockaddr *addr)
2405 switch (addr->sa_family) {
2407 ((struct sockaddr_in *) addr)->sin_addr.s_addr = htonl(INADDR_LOOPBACK);
2410 ipv6_addr_set(&((struct sockaddr_in6 *) addr)->sin6_addr,
2414 ib_addr_set(&((struct sockaddr_ib *) addr)->sib_addr,
2420 static int cma_bind_loopback(struct rdma_id_private *id_priv)
2422 struct cma_device *cma_dev, *cur_dev;
2423 struct ib_port_attr port_attr;
2431 list_for_each_entry(cur_dev, &dev_list, list) {
2432 if (cma_family(id_priv) == AF_IB &&
2433 !rdma_cap_ib_cm(cur_dev->device, 1))
2439 for (p = 1; p <= cur_dev->device->phys_port_cnt; ++p) {
2440 if (!ib_query_port(cur_dev->device, p, &port_attr) &&
2441 port_attr.state == IB_PORT_ACTIVE) {
2456 ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid, NULL);
2460 ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
2464 id_priv->id.route.addr.dev_addr.dev_type =
2465 (rdma_protocol_ib(cma_dev->device, p)) ?
2466 ARPHRD_INFINIBAND : ARPHRD_ETHER;
2468 rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2469 ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
2470 id_priv->id.port_num = p;
2471 cma_attach_to_dev(id_priv, cma_dev);
2472 cma_set_loopback(cma_src_addr(id_priv));
2474 mutex_unlock(&lock);
2478 static void addr_handler(int status, struct sockaddr *src_addr,
2479 struct rdma_dev_addr *dev_addr, void *context)
2481 struct rdma_id_private *id_priv = context;
2482 struct rdma_cm_event event;
2484 memset(&event, 0, sizeof event);
2485 mutex_lock(&id_priv->handler_mutex);
2486 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY,
2487 RDMA_CM_ADDR_RESOLVED))
2490 memcpy(cma_src_addr(id_priv), src_addr, rdma_addr_size(src_addr));
2491 if (!status && !id_priv->cma_dev)
2492 status = cma_acquire_dev(id_priv, NULL);
2495 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
2496 RDMA_CM_ADDR_BOUND))
2498 event.event = RDMA_CM_EVENT_ADDR_ERROR;
2499 event.status = status;
2501 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2503 if (id_priv->id.event_handler(&id_priv->id, &event)) {
2504 cma_exch(id_priv, RDMA_CM_DESTROYING);
2505 mutex_unlock(&id_priv->handler_mutex);
2506 cma_deref_id(id_priv);
2507 rdma_destroy_id(&id_priv->id);
2511 mutex_unlock(&id_priv->handler_mutex);
2512 cma_deref_id(id_priv);
2515 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
2517 struct cma_work *work;
2521 work = kzalloc(sizeof *work, GFP_KERNEL);
2525 if (!id_priv->cma_dev) {
2526 ret = cma_bind_loopback(id_priv);
2531 rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2532 rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
2535 INIT_WORK(&work->work, cma_work_handler);
2536 work->old_state = RDMA_CM_ADDR_QUERY;
2537 work->new_state = RDMA_CM_ADDR_RESOLVED;
2538 work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2539 queue_work(cma_wq, &work->work);
2546 static int cma_resolve_ib_addr(struct rdma_id_private *id_priv)
2548 struct cma_work *work;
2551 work = kzalloc(sizeof *work, GFP_KERNEL);
2555 if (!id_priv->cma_dev) {
2556 ret = cma_resolve_ib_dev(id_priv);
2561 rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, (union ib_gid *)
2562 &(((struct sockaddr_ib *) &id_priv->id.route.addr.dst_addr)->sib_addr));
2565 INIT_WORK(&work->work, cma_work_handler);
2566 work->old_state = RDMA_CM_ADDR_QUERY;
2567 work->new_state = RDMA_CM_ADDR_RESOLVED;
2568 work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2569 queue_work(cma_wq, &work->work);
2576 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2577 struct sockaddr *dst_addr)
2579 if (!src_addr || !src_addr->sa_family) {
2580 src_addr = (struct sockaddr *) &id->route.addr.src_addr;
2581 src_addr->sa_family = dst_addr->sa_family;
2582 if (dst_addr->sa_family == AF_INET6) {
2583 struct sockaddr_in6 *src_addr6 = (struct sockaddr_in6 *) src_addr;
2584 struct sockaddr_in6 *dst_addr6 = (struct sockaddr_in6 *) dst_addr;
2585 src_addr6->sin6_scope_id = dst_addr6->sin6_scope_id;
2586 if (ipv6_addr_type(&dst_addr6->sin6_addr) & IPV6_ADDR_LINKLOCAL)
2587 id->route.addr.dev_addr.bound_dev_if = dst_addr6->sin6_scope_id;
2588 } else if (dst_addr->sa_family == AF_IB) {
2589 ((struct sockaddr_ib *) src_addr)->sib_pkey =
2590 ((struct sockaddr_ib *) dst_addr)->sib_pkey;
2593 return rdma_bind_addr(id, src_addr);
2596 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2597 struct sockaddr *dst_addr, int timeout_ms)
2599 struct rdma_id_private *id_priv;
2602 id_priv = container_of(id, struct rdma_id_private, id);
2603 if (id_priv->state == RDMA_CM_IDLE) {
2604 ret = cma_bind_addr(id, src_addr, dst_addr);
2609 if (cma_family(id_priv) != dst_addr->sa_family)
2612 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY))
2615 atomic_inc(&id_priv->refcount);
2616 memcpy(cma_dst_addr(id_priv), dst_addr, rdma_addr_size(dst_addr));
2617 if (cma_any_addr(dst_addr)) {
2618 ret = cma_resolve_loopback(id_priv);
2620 if (dst_addr->sa_family == AF_IB) {
2621 ret = cma_resolve_ib_addr(id_priv);
2623 ret = rdma_resolve_ip(&addr_client, cma_src_addr(id_priv),
2624 dst_addr, &id->route.addr.dev_addr,
2625 timeout_ms, addr_handler, id_priv);
2633 cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
2634 cma_deref_id(id_priv);
2637 EXPORT_SYMBOL(rdma_resolve_addr);
2639 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse)
2641 struct rdma_id_private *id_priv;
2642 unsigned long flags;
2645 id_priv = container_of(id, struct rdma_id_private, id);
2646 spin_lock_irqsave(&id_priv->lock, flags);
2647 if (reuse || id_priv->state == RDMA_CM_IDLE) {
2648 id_priv->reuseaddr = reuse;
2653 spin_unlock_irqrestore(&id_priv->lock, flags);
2656 EXPORT_SYMBOL(rdma_set_reuseaddr);
2658 int rdma_set_afonly(struct rdma_cm_id *id, int afonly)
2660 struct rdma_id_private *id_priv;
2661 unsigned long flags;
2664 id_priv = container_of(id, struct rdma_id_private, id);
2665 spin_lock_irqsave(&id_priv->lock, flags);
2666 if (id_priv->state == RDMA_CM_IDLE || id_priv->state == RDMA_CM_ADDR_BOUND) {
2667 id_priv->options |= (1 << CMA_OPTION_AFONLY);
2668 id_priv->afonly = afonly;
2673 spin_unlock_irqrestore(&id_priv->lock, flags);
2676 EXPORT_SYMBOL(rdma_set_afonly);
2678 static void cma_bind_port(struct rdma_bind_list *bind_list,
2679 struct rdma_id_private *id_priv)
2681 struct sockaddr *addr;
2682 struct sockaddr_ib *sib;
2686 addr = cma_src_addr(id_priv);
2687 port = htons(bind_list->port);
2689 switch (addr->sa_family) {
2691 ((struct sockaddr_in *) addr)->sin_port = port;
2694 ((struct sockaddr_in6 *) addr)->sin6_port = port;
2697 sib = (struct sockaddr_ib *) addr;
2698 sid = be64_to_cpu(sib->sib_sid);
2699 mask = be64_to_cpu(sib->sib_sid_mask);
2700 sib->sib_sid = cpu_to_be64((sid & mask) | (u64) ntohs(port));
2701 sib->sib_sid_mask = cpu_to_be64(~0ULL);
2704 id_priv->bind_list = bind_list;
2705 hlist_add_head(&id_priv->node, &bind_list->owners);
2708 static int cma_alloc_port(enum rdma_port_space ps,
2709 struct rdma_id_private *id_priv, unsigned short snum)
2711 struct rdma_bind_list *bind_list;
2714 bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
2718 ret = cma_ps_alloc(id_priv->id.route.addr.dev_addr.net, ps, bind_list,
2724 bind_list->port = (unsigned short)ret;
2725 cma_bind_port(bind_list, id_priv);
2729 return ret == -ENOSPC ? -EADDRNOTAVAIL : ret;
2732 static int cma_alloc_any_port(enum rdma_port_space ps,
2733 struct rdma_id_private *id_priv)
2735 static unsigned int last_used_port;
2736 int low, high, remaining;
2738 struct net *net = id_priv->id.route.addr.dev_addr.net;
2740 inet_get_local_port_range(net, &low, &high);
2741 remaining = (high - low) + 1;
2742 rover = prandom_u32() % remaining + low;
2744 if (last_used_port != rover &&
2745 !cma_ps_find(net, ps, (unsigned short)rover)) {
2746 int ret = cma_alloc_port(ps, id_priv, rover);
2748 * Remember previously used port number in order to avoid
2749 * re-using same port immediately after it is closed.
2752 last_used_port = rover;
2753 if (ret != -EADDRNOTAVAIL)
2758 if ((rover < low) || (rover > high))
2762 return -EADDRNOTAVAIL;
2766 * Check that the requested port is available. This is called when trying to
2767 * bind to a specific port, or when trying to listen on a bound port. In
2768 * the latter case, the provided id_priv may already be on the bind_list, but
2769 * we still need to check that it's okay to start listening.
2771 static int cma_check_port(struct rdma_bind_list *bind_list,
2772 struct rdma_id_private *id_priv, uint8_t reuseaddr)
2774 struct rdma_id_private *cur_id;
2775 struct sockaddr *addr, *cur_addr;
2777 addr = cma_src_addr(id_priv);
2778 hlist_for_each_entry(cur_id, &bind_list->owners, node) {
2779 if (id_priv == cur_id)
2782 if ((cur_id->state != RDMA_CM_LISTEN) && reuseaddr &&
2786 cur_addr = cma_src_addr(cur_id);
2787 if (id_priv->afonly && cur_id->afonly &&
2788 (addr->sa_family != cur_addr->sa_family))
2791 if (cma_any_addr(addr) || cma_any_addr(cur_addr))
2792 return -EADDRNOTAVAIL;
2794 if (!cma_addr_cmp(addr, cur_addr))
2800 static int cma_use_port(enum rdma_port_space ps,
2801 struct rdma_id_private *id_priv)
2803 struct rdma_bind_list *bind_list;
2804 unsigned short snum;
2807 snum = ntohs(cma_port(cma_src_addr(id_priv)));
2808 if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
2811 bind_list = cma_ps_find(id_priv->id.route.addr.dev_addr.net, ps, snum);
2813 ret = cma_alloc_port(ps, id_priv, snum);
2815 ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr);
2817 cma_bind_port(bind_list, id_priv);
2822 static int cma_bind_listen(struct rdma_id_private *id_priv)
2824 struct rdma_bind_list *bind_list = id_priv->bind_list;
2828 if (bind_list->owners.first->next)
2829 ret = cma_check_port(bind_list, id_priv, 0);
2830 mutex_unlock(&lock);
2834 static enum rdma_port_space cma_select_inet_ps(
2835 struct rdma_id_private *id_priv)
2837 switch (id_priv->id.ps) {
2842 return id_priv->id.ps;
2849 static enum rdma_port_space cma_select_ib_ps(struct rdma_id_private *id_priv)
2851 enum rdma_port_space ps = 0;
2852 struct sockaddr_ib *sib;
2853 u64 sid_ps, mask, sid;
2855 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
2856 mask = be64_to_cpu(sib->sib_sid_mask) & RDMA_IB_IP_PS_MASK;
2857 sid = be64_to_cpu(sib->sib_sid) & mask;
2859 if ((id_priv->id.ps == RDMA_PS_IB) && (sid == (RDMA_IB_IP_PS_IB & mask))) {
2860 sid_ps = RDMA_IB_IP_PS_IB;
2862 } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_TCP)) &&
2863 (sid == (RDMA_IB_IP_PS_TCP & mask))) {
2864 sid_ps = RDMA_IB_IP_PS_TCP;
2866 } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_UDP)) &&
2867 (sid == (RDMA_IB_IP_PS_UDP & mask))) {
2868 sid_ps = RDMA_IB_IP_PS_UDP;
2873 sib->sib_sid = cpu_to_be64(sid_ps | ntohs(cma_port((struct sockaddr *) sib)));
2874 sib->sib_sid_mask = cpu_to_be64(RDMA_IB_IP_PS_MASK |
2875 be64_to_cpu(sib->sib_sid_mask));
2880 static int cma_get_port(struct rdma_id_private *id_priv)
2882 enum rdma_port_space ps;
2885 if (cma_family(id_priv) != AF_IB)
2886 ps = cma_select_inet_ps(id_priv);
2888 ps = cma_select_ib_ps(id_priv);
2890 return -EPROTONOSUPPORT;
2893 if (cma_any_port(cma_src_addr(id_priv)))
2894 ret = cma_alloc_any_port(ps, id_priv);
2896 ret = cma_use_port(ps, id_priv);
2897 mutex_unlock(&lock);
2902 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
2903 struct sockaddr *addr)
2905 #if IS_ENABLED(CONFIG_IPV6)
2906 struct sockaddr_in6 *sin6;
2908 if (addr->sa_family != AF_INET6)
2911 sin6 = (struct sockaddr_in6 *) addr;
2913 if (!(ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL))
2916 if (!sin6->sin6_scope_id)
2919 dev_addr->bound_dev_if = sin6->sin6_scope_id;
2924 int rdma_listen(struct rdma_cm_id *id, int backlog)
2926 struct rdma_id_private *id_priv;
2929 id_priv = container_of(id, struct rdma_id_private, id);
2930 if (id_priv->state == RDMA_CM_IDLE) {
2931 id->route.addr.src_addr.ss_family = AF_INET;
2932 ret = rdma_bind_addr(id, cma_src_addr(id_priv));
2937 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN))
2940 if (id_priv->reuseaddr) {
2941 ret = cma_bind_listen(id_priv);
2946 id_priv->backlog = backlog;
2948 if (rdma_cap_ib_cm(id->device, 1)) {
2949 ret = cma_ib_listen(id_priv);
2952 } else if (rdma_cap_iw_cm(id->device, 1)) {
2953 ret = cma_iw_listen(id_priv, backlog);
2961 cma_listen_on_all(id_priv);
2965 id_priv->backlog = 0;
2966 cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND);
2969 EXPORT_SYMBOL(rdma_listen);
2971 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
2973 struct rdma_id_private *id_priv;
2976 if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6 &&
2977 addr->sa_family != AF_IB)
2978 return -EAFNOSUPPORT;
2980 id_priv = container_of(id, struct rdma_id_private, id);
2981 if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND))
2984 ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
2988 memcpy(cma_src_addr(id_priv), addr, rdma_addr_size(addr));
2989 if (!cma_any_addr(addr)) {
2990 ret = cma_translate_addr(addr, &id->route.addr.dev_addr);
2994 ret = cma_acquire_dev(id_priv, NULL);
2999 if (!(id_priv->options & (1 << CMA_OPTION_AFONLY))) {
3000 if (addr->sa_family == AF_INET)
3001 id_priv->afonly = 1;
3002 #if IS_ENABLED(CONFIG_IPV6)
3003 else if (addr->sa_family == AF_INET6) {
3004 struct net *net = id_priv->id.route.addr.dev_addr.net;
3006 id_priv->afonly = net->ipv6.sysctl.bindv6only;
3010 ret = cma_get_port(id_priv);
3016 if (id_priv->cma_dev)
3017 cma_release_dev(id_priv);
3019 cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE);
3022 EXPORT_SYMBOL(rdma_bind_addr);
3024 static int cma_format_hdr(void *hdr, struct rdma_id_private *id_priv)
3026 struct cma_hdr *cma_hdr;
3029 cma_hdr->cma_version = CMA_VERSION;
3030 if (cma_family(id_priv) == AF_INET) {
3031 struct sockaddr_in *src4, *dst4;
3033 src4 = (struct sockaddr_in *) cma_src_addr(id_priv);
3034 dst4 = (struct sockaddr_in *) cma_dst_addr(id_priv);
3036 cma_set_ip_ver(cma_hdr, 4);
3037 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
3038 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
3039 cma_hdr->port = src4->sin_port;
3040 } else if (cma_family(id_priv) == AF_INET6) {
3041 struct sockaddr_in6 *src6, *dst6;
3043 src6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
3044 dst6 = (struct sockaddr_in6 *) cma_dst_addr(id_priv);
3046 cma_set_ip_ver(cma_hdr, 6);
3047 cma_hdr->src_addr.ip6 = src6->sin6_addr;
3048 cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
3049 cma_hdr->port = src6->sin6_port;
3054 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
3055 struct ib_cm_event *ib_event)
3057 struct rdma_id_private *id_priv = cm_id->context;
3058 struct rdma_cm_event event;
3059 struct ib_cm_sidr_rep_event_param *rep = &ib_event->param.sidr_rep_rcvd;
3062 if (cma_disable_callback(id_priv, RDMA_CM_CONNECT))
3065 memset(&event, 0, sizeof event);
3066 switch (ib_event->event) {
3067 case IB_CM_SIDR_REQ_ERROR:
3068 event.event = RDMA_CM_EVENT_UNREACHABLE;
3069 event.status = -ETIMEDOUT;
3071 case IB_CM_SIDR_REP_RECEIVED:
3072 event.param.ud.private_data = ib_event->private_data;
3073 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
3074 if (rep->status != IB_SIDR_SUCCESS) {
3075 event.event = RDMA_CM_EVENT_UNREACHABLE;
3076 event.status = ib_event->param.sidr_rep_rcvd.status;
3079 ret = cma_set_qkey(id_priv, rep->qkey);
3081 event.event = RDMA_CM_EVENT_ADDR_ERROR;
3085 ib_init_ah_from_path(id_priv->id.device, id_priv->id.port_num,
3086 id_priv->id.route.path_rec,
3087 &event.param.ud.ah_attr);
3088 event.param.ud.qp_num = rep->qpn;
3089 event.param.ud.qkey = rep->qkey;
3090 event.event = RDMA_CM_EVENT_ESTABLISHED;
3094 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
3099 ret = id_priv->id.event_handler(&id_priv->id, &event);
3101 /* Destroy the CM ID by returning a non-zero value. */
3102 id_priv->cm_id.ib = NULL;
3103 cma_exch(id_priv, RDMA_CM_DESTROYING);
3104 mutex_unlock(&id_priv->handler_mutex);
3105 rdma_destroy_id(&id_priv->id);
3109 mutex_unlock(&id_priv->handler_mutex);
3113 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
3114 struct rdma_conn_param *conn_param)
3116 struct ib_cm_sidr_req_param req;
3117 struct ib_cm_id *id;
3121 memset(&req, 0, sizeof req);
3122 offset = cma_user_data_offset(id_priv);
3123 req.private_data_len = offset + conn_param->private_data_len;
3124 if (req.private_data_len < conn_param->private_data_len)
3127 if (req.private_data_len) {
3128 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
3132 private_data = NULL;
3135 if (conn_param->private_data && conn_param->private_data_len)
3136 memcpy(private_data + offset, conn_param->private_data,
3137 conn_param->private_data_len);
3140 ret = cma_format_hdr(private_data, id_priv);
3143 req.private_data = private_data;
3146 id = ib_create_cm_id(id_priv->id.device, cma_sidr_rep_handler,
3152 id_priv->cm_id.ib = id;
3154 req.path = id_priv->id.route.path_rec;
3155 req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
3156 req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
3157 req.max_cm_retries = CMA_MAX_CM_RETRIES;
3159 ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
3161 ib_destroy_cm_id(id_priv->cm_id.ib);
3162 id_priv->cm_id.ib = NULL;
3165 kfree(private_data);
3169 static int cma_connect_ib(struct rdma_id_private *id_priv,
3170 struct rdma_conn_param *conn_param)
3172 struct ib_cm_req_param req;
3173 struct rdma_route *route;
3175 struct ib_cm_id *id;
3178 memset(&req, 0, sizeof req);
3179 offset = cma_user_data_offset(id_priv);
3180 req.private_data_len = offset + conn_param->private_data_len;
3181 if (req.private_data_len < conn_param->private_data_len)
3184 if (req.private_data_len) {
3185 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
3189 private_data = NULL;
3192 if (conn_param->private_data && conn_param->private_data_len)
3193 memcpy(private_data + offset, conn_param->private_data,
3194 conn_param->private_data_len);
3196 id = ib_create_cm_id(id_priv->id.device, cma_ib_handler, id_priv);
3201 id_priv->cm_id.ib = id;
3203 route = &id_priv->id.route;
3205 ret = cma_format_hdr(private_data, id_priv);
3208 req.private_data = private_data;
3211 req.primary_path = &route->path_rec[0];
3212 if (route->num_paths == 2)
3213 req.alternate_path = &route->path_rec[1];
3215 req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
3216 req.qp_num = id_priv->qp_num;
3217 req.qp_type = id_priv->id.qp_type;
3218 req.starting_psn = id_priv->seq_num;
3219 req.responder_resources = conn_param->responder_resources;
3220 req.initiator_depth = conn_param->initiator_depth;
3221 req.flow_control = conn_param->flow_control;
3222 req.retry_count = min_t(u8, 7, conn_param->retry_count);
3223 req.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
3224 req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
3225 req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
3226 req.max_cm_retries = CMA_MAX_CM_RETRIES;
3227 req.srq = id_priv->srq ? 1 : 0;
3229 ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
3231 if (ret && !IS_ERR(id)) {
3232 ib_destroy_cm_id(id);
3233 id_priv->cm_id.ib = NULL;
3236 kfree(private_data);
3240 static int cma_connect_iw(struct rdma_id_private *id_priv,
3241 struct rdma_conn_param *conn_param)
3243 struct iw_cm_id *cm_id;
3245 struct iw_cm_conn_param iw_param;
3247 cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
3249 return PTR_ERR(cm_id);
3251 cm_id->tos = id_priv->tos;
3252 id_priv->cm_id.iw = cm_id;
3254 memcpy(&cm_id->local_addr, cma_src_addr(id_priv),
3255 rdma_addr_size(cma_src_addr(id_priv)));
3256 memcpy(&cm_id->remote_addr, cma_dst_addr(id_priv),
3257 rdma_addr_size(cma_dst_addr(id_priv)));
3259 ret = cma_modify_qp_rtr(id_priv, conn_param);
3264 iw_param.ord = conn_param->initiator_depth;
3265 iw_param.ird = conn_param->responder_resources;
3266 iw_param.private_data = conn_param->private_data;
3267 iw_param.private_data_len = conn_param->private_data_len;
3268 iw_param.qpn = id_priv->id.qp ? id_priv->qp_num : conn_param->qp_num;
3270 memset(&iw_param, 0, sizeof iw_param);
3271 iw_param.qpn = id_priv->qp_num;
3273 ret = iw_cm_connect(cm_id, &iw_param);
3276 iw_destroy_cm_id(cm_id);
3277 id_priv->cm_id.iw = NULL;
3282 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
3284 struct rdma_id_private *id_priv;
3287 id_priv = container_of(id, struct rdma_id_private, id);
3288 if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT))
3292 id_priv->qp_num = conn_param->qp_num;
3293 id_priv->srq = conn_param->srq;
3296 if (rdma_cap_ib_cm(id->device, id->port_num)) {
3297 if (id->qp_type == IB_QPT_UD)
3298 ret = cma_resolve_ib_udp(id_priv, conn_param);
3300 ret = cma_connect_ib(id_priv, conn_param);
3301 } else if (rdma_cap_iw_cm(id->device, id->port_num))
3302 ret = cma_connect_iw(id_priv, conn_param);
3310 cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED);
3313 EXPORT_SYMBOL(rdma_connect);
3315 static int cma_accept_ib(struct rdma_id_private *id_priv,
3316 struct rdma_conn_param *conn_param)
3318 struct ib_cm_rep_param rep;
3321 ret = cma_modify_qp_rtr(id_priv, conn_param);
3325 ret = cma_modify_qp_rts(id_priv, conn_param);
3329 memset(&rep, 0, sizeof rep);
3330 rep.qp_num = id_priv->qp_num;
3331 rep.starting_psn = id_priv->seq_num;
3332 rep.private_data = conn_param->private_data;
3333 rep.private_data_len = conn_param->private_data_len;
3334 rep.responder_resources = conn_param->responder_resources;
3335 rep.initiator_depth = conn_param->initiator_depth;
3336 rep.failover_accepted = 0;
3337 rep.flow_control = conn_param->flow_control;
3338 rep.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
3339 rep.srq = id_priv->srq ? 1 : 0;
3341 ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
3346 static int cma_accept_iw(struct rdma_id_private *id_priv,
3347 struct rdma_conn_param *conn_param)
3349 struct iw_cm_conn_param iw_param;
3352 ret = cma_modify_qp_rtr(id_priv, conn_param);
3356 iw_param.ord = conn_param->initiator_depth;
3357 iw_param.ird = conn_param->responder_resources;
3358 iw_param.private_data = conn_param->private_data;
3359 iw_param.private_data_len = conn_param->private_data_len;
3360 if (id_priv->id.qp) {
3361 iw_param.qpn = id_priv->qp_num;
3363 iw_param.qpn = conn_param->qp_num;
3365 return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
3368 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
3369 enum ib_cm_sidr_status status, u32 qkey,
3370 const void *private_data, int private_data_len)
3372 struct ib_cm_sidr_rep_param rep;
3375 memset(&rep, 0, sizeof rep);
3376 rep.status = status;
3377 if (status == IB_SIDR_SUCCESS) {
3378 ret = cma_set_qkey(id_priv, qkey);
3381 rep.qp_num = id_priv->qp_num;
3382 rep.qkey = id_priv->qkey;
3384 rep.private_data = private_data;
3385 rep.private_data_len = private_data_len;
3387 return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
3390 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
3392 struct rdma_id_private *id_priv;
3395 id_priv = container_of(id, struct rdma_id_private, id);
3397 id_priv->owner = task_pid_nr(current);
3399 if (!cma_comp(id_priv, RDMA_CM_CONNECT))
3402 if (!id->qp && conn_param) {
3403 id_priv->qp_num = conn_param->qp_num;
3404 id_priv->srq = conn_param->srq;
3407 if (rdma_cap_ib_cm(id->device, id->port_num)) {
3408 if (id->qp_type == IB_QPT_UD) {
3410 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
3412 conn_param->private_data,
3413 conn_param->private_data_len);
3415 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
3419 ret = cma_accept_ib(id_priv, conn_param);
3421 ret = cma_rep_recv(id_priv);
3423 } else if (rdma_cap_iw_cm(id->device, id->port_num))
3424 ret = cma_accept_iw(id_priv, conn_param);
3433 cma_modify_qp_err(id_priv);
3434 rdma_reject(id, NULL, 0);
3437 EXPORT_SYMBOL(rdma_accept);
3439 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
3441 struct rdma_id_private *id_priv;
3444 id_priv = container_of(id, struct rdma_id_private, id);
3445 if (!id_priv->cm_id.ib)
3448 switch (id->device->node_type) {
3449 case RDMA_NODE_IB_CA:
3450 ret = ib_cm_notify(id_priv->cm_id.ib, event);
3458 EXPORT_SYMBOL(rdma_notify);
3460 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
3461 u8 private_data_len)
3463 struct rdma_id_private *id_priv;
3466 id_priv = container_of(id, struct rdma_id_private, id);
3467 if (!id_priv->cm_id.ib)
3470 if (rdma_cap_ib_cm(id->device, id->port_num)) {
3471 if (id->qp_type == IB_QPT_UD)
3472 ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT, 0,
3473 private_data, private_data_len);
3475 ret = ib_send_cm_rej(id_priv->cm_id.ib,
3476 IB_CM_REJ_CONSUMER_DEFINED, NULL,
3477 0, private_data, private_data_len);
3478 } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
3479 ret = iw_cm_reject(id_priv->cm_id.iw,
3480 private_data, private_data_len);
3486 EXPORT_SYMBOL(rdma_reject);
3488 int rdma_disconnect(struct rdma_cm_id *id)
3490 struct rdma_id_private *id_priv;
3493 id_priv = container_of(id, struct rdma_id_private, id);
3494 if (!id_priv->cm_id.ib)
3497 if (rdma_cap_ib_cm(id->device, id->port_num)) {
3498 ret = cma_modify_qp_err(id_priv);
3501 /* Initiate or respond to a disconnect. */
3502 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
3503 ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
3504 } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
3505 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
3512 EXPORT_SYMBOL(rdma_disconnect);
3514 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
3516 struct rdma_id_private *id_priv;
3517 struct cma_multicast *mc = multicast->context;
3518 struct rdma_cm_event event;
3521 id_priv = mc->id_priv;
3522 if (cma_disable_callback(id_priv, RDMA_CM_ADDR_BOUND) &&
3523 cma_disable_callback(id_priv, RDMA_CM_ADDR_RESOLVED))
3527 status = cma_set_qkey(id_priv, be32_to_cpu(multicast->rec.qkey));
3528 mutex_lock(&id_priv->qp_mutex);
3529 if (!status && id_priv->id.qp)
3530 status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid,
3531 be16_to_cpu(multicast->rec.mlid));
3532 mutex_unlock(&id_priv->qp_mutex);
3534 memset(&event, 0, sizeof event);
3535 event.status = status;
3536 event.param.ud.private_data = mc->context;
3538 event.event = RDMA_CM_EVENT_MULTICAST_JOIN;
3539 ib_init_ah_from_mcmember(id_priv->id.device,
3540 id_priv->id.port_num, &multicast->rec,
3541 &event.param.ud.ah_attr);
3542 event.param.ud.qp_num = 0xFFFFFF;
3543 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
3545 event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
3547 ret = id_priv->id.event_handler(&id_priv->id, &event);
3549 cma_exch(id_priv, RDMA_CM_DESTROYING);
3550 mutex_unlock(&id_priv->handler_mutex);
3551 rdma_destroy_id(&id_priv->id);
3555 mutex_unlock(&id_priv->handler_mutex);
3559 static void cma_set_mgid(struct rdma_id_private *id_priv,
3560 struct sockaddr *addr, union ib_gid *mgid)
3562 unsigned char mc_map[MAX_ADDR_LEN];
3563 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3564 struct sockaddr_in *sin = (struct sockaddr_in *) addr;
3565 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
3567 if (cma_any_addr(addr)) {
3568 memset(mgid, 0, sizeof *mgid);
3569 } else if ((addr->sa_family == AF_INET6) &&
3570 ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
3572 /* IPv6 address is an SA assigned MGID. */
3573 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3574 } else if (addr->sa_family == AF_IB) {
3575 memcpy(mgid, &((struct sockaddr_ib *) addr)->sib_addr, sizeof *mgid);
3576 } else if ((addr->sa_family == AF_INET6)) {
3577 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
3578 if (id_priv->id.ps == RDMA_PS_UDP)
3579 mc_map[7] = 0x01; /* Use RDMA CM signature */
3580 *mgid = *(union ib_gid *) (mc_map + 4);
3582 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
3583 if (id_priv->id.ps == RDMA_PS_UDP)
3584 mc_map[7] = 0x01; /* Use RDMA CM signature */
3585 *mgid = *(union ib_gid *) (mc_map + 4);
3589 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
3590 struct cma_multicast *mc)
3592 struct ib_sa_mcmember_rec rec;
3593 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3594 ib_sa_comp_mask comp_mask;
3597 ib_addr_get_mgid(dev_addr, &rec.mgid);
3598 ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
3603 ret = cma_set_qkey(id_priv, 0);
3607 cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
3608 rec.qkey = cpu_to_be32(id_priv->qkey);
3609 rdma_addr_get_sgid(dev_addr, &rec.port_gid);
3610 rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
3613 comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
3614 IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
3615 IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
3616 IB_SA_MCMEMBER_REC_FLOW_LABEL |
3617 IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
3619 if (id_priv->id.ps == RDMA_PS_IPOIB)
3620 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
3621 IB_SA_MCMEMBER_REC_RATE_SELECTOR |
3622 IB_SA_MCMEMBER_REC_MTU_SELECTOR |
3623 IB_SA_MCMEMBER_REC_MTU |
3624 IB_SA_MCMEMBER_REC_HOP_LIMIT;
3626 mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device,
3627 id_priv->id.port_num, &rec,
3628 comp_mask, GFP_KERNEL,
3629 cma_ib_mc_handler, mc);
3630 return PTR_ERR_OR_ZERO(mc->multicast.ib);
3633 static void iboe_mcast_work_handler(struct work_struct *work)
3635 struct iboe_mcast_work *mw = container_of(work, struct iboe_mcast_work, work);
3636 struct cma_multicast *mc = mw->mc;
3637 struct ib_sa_multicast *m = mc->multicast.ib;
3639 mc->multicast.ib->context = mc;
3640 cma_ib_mc_handler(0, m);
3641 kref_put(&mc->mcref, release_mc);
3645 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid)
3647 struct sockaddr_in *sin = (struct sockaddr_in *)addr;
3648 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
3650 if (cma_any_addr(addr)) {
3651 memset(mgid, 0, sizeof *mgid);
3652 } else if (addr->sa_family == AF_INET6) {
3653 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3655 mgid->raw[0] = 0xff;
3656 mgid->raw[1] = 0x0e;
3665 mgid->raw[10] = 0xff;
3666 mgid->raw[11] = 0xff;
3667 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
3671 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
3672 struct cma_multicast *mc)
3674 struct iboe_mcast_work *work;
3675 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3677 struct sockaddr *addr = (struct sockaddr *)&mc->addr;
3678 struct net_device *ndev = NULL;
3680 if (cma_zero_addr((struct sockaddr *)&mc->addr))
3683 work = kzalloc(sizeof *work, GFP_KERNEL);
3687 mc->multicast.ib = kzalloc(sizeof(struct ib_sa_multicast), GFP_KERNEL);
3688 if (!mc->multicast.ib) {
3693 cma_iboe_set_mgid(addr, &mc->multicast.ib->rec.mgid);
3695 mc->multicast.ib->rec.pkey = cpu_to_be16(0xffff);
3696 if (id_priv->id.ps == RDMA_PS_UDP)
3697 mc->multicast.ib->rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
3699 if (dev_addr->bound_dev_if)
3700 ndev = dev_get_by_index(&init_net, dev_addr->bound_dev_if);
3705 mc->multicast.ib->rec.rate = iboe_get_rate(ndev);
3706 mc->multicast.ib->rec.hop_limit = 1;
3707 mc->multicast.ib->rec.mtu = iboe_get_mtu(ndev->mtu);
3709 if (!mc->multicast.ib->rec.mtu) {
3713 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
3714 &mc->multicast.ib->rec.port_gid);
3717 INIT_WORK(&work->work, iboe_mcast_work_handler);
3718 kref_get(&mc->mcref);
3719 queue_work(cma_wq, &work->work);
3724 kfree(mc->multicast.ib);
3730 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
3733 struct rdma_id_private *id_priv;
3734 struct cma_multicast *mc;
3737 id_priv = container_of(id, struct rdma_id_private, id);
3738 if (!cma_comp(id_priv, RDMA_CM_ADDR_BOUND) &&
3739 !cma_comp(id_priv, RDMA_CM_ADDR_RESOLVED))
3742 mc = kmalloc(sizeof *mc, GFP_KERNEL);
3746 memcpy(&mc->addr, addr, rdma_addr_size(addr));
3747 mc->context = context;
3748 mc->id_priv = id_priv;
3750 spin_lock(&id_priv->lock);
3751 list_add(&mc->list, &id_priv->mc_list);
3752 spin_unlock(&id_priv->lock);
3754 if (rdma_protocol_roce(id->device, id->port_num)) {
3755 kref_init(&mc->mcref);
3756 ret = cma_iboe_join_multicast(id_priv, mc);
3757 } else if (rdma_cap_ib_mcast(id->device, id->port_num))
3758 ret = cma_join_ib_multicast(id_priv, mc);
3763 spin_lock_irq(&id_priv->lock);
3764 list_del(&mc->list);
3765 spin_unlock_irq(&id_priv->lock);
3770 EXPORT_SYMBOL(rdma_join_multicast);
3772 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
3774 struct rdma_id_private *id_priv;
3775 struct cma_multicast *mc;
3777 id_priv = container_of(id, struct rdma_id_private, id);
3778 spin_lock_irq(&id_priv->lock);
3779 list_for_each_entry(mc, &id_priv->mc_list, list) {
3780 if (!memcmp(&mc->addr, addr, rdma_addr_size(addr))) {
3781 list_del(&mc->list);
3782 spin_unlock_irq(&id_priv->lock);
3785 ib_detach_mcast(id->qp,
3786 &mc->multicast.ib->rec.mgid,
3787 be16_to_cpu(mc->multicast.ib->rec.mlid));
3789 BUG_ON(id_priv->cma_dev->device != id->device);
3791 if (rdma_cap_ib_mcast(id->device, id->port_num)) {
3792 ib_sa_free_multicast(mc->multicast.ib);
3794 } else if (rdma_protocol_roce(id->device, id->port_num))
3795 kref_put(&mc->mcref, release_mc);
3800 spin_unlock_irq(&id_priv->lock);
3802 EXPORT_SYMBOL(rdma_leave_multicast);
3804 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
3806 struct rdma_dev_addr *dev_addr;
3807 struct cma_ndev_work *work;
3809 dev_addr = &id_priv->id.route.addr.dev_addr;
3811 if ((dev_addr->bound_dev_if == ndev->ifindex) &&
3812 (net_eq(dev_net(ndev), dev_addr->net)) &&
3813 memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
3814 printk(KERN_INFO "RDMA CM addr change for ndev %s used by id %p\n",
3815 ndev->name, &id_priv->id);
3816 work = kzalloc(sizeof *work, GFP_KERNEL);
3820 INIT_WORK(&work->work, cma_ndev_work_handler);
3822 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
3823 atomic_inc(&id_priv->refcount);
3824 queue_work(cma_wq, &work->work);
3830 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
3833 struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
3834 struct cma_device *cma_dev;
3835 struct rdma_id_private *id_priv;
3836 int ret = NOTIFY_DONE;
3838 if (event != NETDEV_BONDING_FAILOVER)
3841 if (!(ndev->flags & IFF_MASTER) || !(ndev->priv_flags & IFF_BONDING))
3845 list_for_each_entry(cma_dev, &dev_list, list)
3846 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
3847 ret = cma_netdev_change(ndev, id_priv);
3853 mutex_unlock(&lock);
3857 static struct notifier_block cma_nb = {
3858 .notifier_call = cma_netdev_callback
3861 static void cma_add_one(struct ib_device *device)
3863 struct cma_device *cma_dev;
3864 struct rdma_id_private *id_priv;
3866 cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
3870 cma_dev->device = device;
3872 init_completion(&cma_dev->comp);
3873 atomic_set(&cma_dev->refcount, 1);
3874 INIT_LIST_HEAD(&cma_dev->id_list);
3875 ib_set_client_data(device, &cma_client, cma_dev);
3878 list_add_tail(&cma_dev->list, &dev_list);
3879 list_for_each_entry(id_priv, &listen_any_list, list)
3880 cma_listen_on_dev(id_priv, cma_dev);
3881 mutex_unlock(&lock);
3884 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
3886 struct rdma_cm_event event;
3887 enum rdma_cm_state state;
3890 /* Record that we want to remove the device */
3891 state = cma_exch(id_priv, RDMA_CM_DEVICE_REMOVAL);
3892 if (state == RDMA_CM_DESTROYING)
3895 cma_cancel_operation(id_priv, state);
3896 mutex_lock(&id_priv->handler_mutex);
3898 /* Check for destruction from another callback. */
3899 if (!cma_comp(id_priv, RDMA_CM_DEVICE_REMOVAL))
3902 memset(&event, 0, sizeof event);
3903 event.event = RDMA_CM_EVENT_DEVICE_REMOVAL;
3904 ret = id_priv->id.event_handler(&id_priv->id, &event);
3906 mutex_unlock(&id_priv->handler_mutex);
3910 static void cma_process_remove(struct cma_device *cma_dev)
3912 struct rdma_id_private *id_priv;
3916 while (!list_empty(&cma_dev->id_list)) {
3917 id_priv = list_entry(cma_dev->id_list.next,
3918 struct rdma_id_private, list);
3920 list_del(&id_priv->listen_list);
3921 list_del_init(&id_priv->list);
3922 atomic_inc(&id_priv->refcount);
3923 mutex_unlock(&lock);
3925 ret = id_priv->internal_id ? 1 : cma_remove_id_dev(id_priv);
3926 cma_deref_id(id_priv);
3928 rdma_destroy_id(&id_priv->id);
3932 mutex_unlock(&lock);
3934 cma_deref_dev(cma_dev);
3935 wait_for_completion(&cma_dev->comp);
3938 static void cma_remove_one(struct ib_device *device, void *client_data)
3940 struct cma_device *cma_dev = client_data;
3946 list_del(&cma_dev->list);
3947 mutex_unlock(&lock);
3949 cma_process_remove(cma_dev);
3953 static int cma_get_id_stats(struct sk_buff *skb, struct netlink_callback *cb)
3955 struct nlmsghdr *nlh;
3956 struct rdma_cm_id_stats *id_stats;
3957 struct rdma_id_private *id_priv;
3958 struct rdma_cm_id *id = NULL;
3959 struct cma_device *cma_dev;
3960 int i_dev = 0, i_id = 0;
3963 * We export all of the IDs as a sequence of messages. Each
3964 * ID gets its own netlink message.
3968 list_for_each_entry(cma_dev, &dev_list, list) {
3969 if (i_dev < cb->args[0]) {
3975 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
3976 if (i_id < cb->args[1]) {
3981 id_stats = ibnl_put_msg(skb, &nlh, cb->nlh->nlmsg_seq,
3982 sizeof *id_stats, RDMA_NL_RDMA_CM,
3983 RDMA_NL_RDMA_CM_ID_STATS,
3988 memset(id_stats, 0, sizeof *id_stats);
3990 id_stats->node_type = id->route.addr.dev_addr.dev_type;
3991 id_stats->port_num = id->port_num;
3992 id_stats->bound_dev_if =
3993 id->route.addr.dev_addr.bound_dev_if;
3995 if (ibnl_put_attr(skb, nlh,
3996 rdma_addr_size(cma_src_addr(id_priv)),
3997 cma_src_addr(id_priv),
3998 RDMA_NL_RDMA_CM_ATTR_SRC_ADDR))
4000 if (ibnl_put_attr(skb, nlh,
4001 rdma_addr_size(cma_src_addr(id_priv)),
4002 cma_dst_addr(id_priv),
4003 RDMA_NL_RDMA_CM_ATTR_DST_ADDR))
4006 id_stats->pid = id_priv->owner;
4007 id_stats->port_space = id->ps;
4008 id_stats->cm_state = id_priv->state;
4009 id_stats->qp_num = id_priv->qp_num;
4010 id_stats->qp_type = id->qp_type;
4020 mutex_unlock(&lock);
4021 cb->args[0] = i_dev;
4027 static const struct ibnl_client_cbs cma_cb_table[] = {
4028 [RDMA_NL_RDMA_CM_ID_STATS] = { .dump = cma_get_id_stats,
4029 .module = THIS_MODULE },
4032 static int cma_init_net(struct net *net)
4034 struct cma_pernet *pernet = cma_pernet(net);
4036 idr_init(&pernet->tcp_ps);
4037 idr_init(&pernet->udp_ps);
4038 idr_init(&pernet->ipoib_ps);
4039 idr_init(&pernet->ib_ps);
4044 static void cma_exit_net(struct net *net)
4046 struct cma_pernet *pernet = cma_pernet(net);
4048 idr_destroy(&pernet->tcp_ps);
4049 idr_destroy(&pernet->udp_ps);
4050 idr_destroy(&pernet->ipoib_ps);
4051 idr_destroy(&pernet->ib_ps);
4054 static struct pernet_operations cma_pernet_operations = {
4055 .init = cma_init_net,
4056 .exit = cma_exit_net,
4057 .id = &cma_pernet_id,
4058 .size = sizeof(struct cma_pernet),
4061 static int __init cma_init(void)
4065 cma_wq = create_singlethread_workqueue("rdma_cm");
4069 ret = register_pernet_subsys(&cma_pernet_operations);
4073 ib_sa_register_client(&sa_client);
4074 rdma_addr_register_client(&addr_client);
4075 register_netdevice_notifier(&cma_nb);
4077 ret = ib_register_client(&cma_client);
4081 if (ibnl_add_client(RDMA_NL_RDMA_CM, RDMA_NL_RDMA_CM_NUM_OPS, cma_cb_table))
4082 printk(KERN_WARNING "RDMA CMA: failed to add netlink callback\n");
4087 unregister_netdevice_notifier(&cma_nb);
4088 rdma_addr_unregister_client(&addr_client);
4089 ib_sa_unregister_client(&sa_client);
4091 destroy_workqueue(cma_wq);
4095 static void __exit cma_cleanup(void)
4097 ibnl_remove_client(RDMA_NL_RDMA_CM);
4098 ib_unregister_client(&cma_client);
4099 unregister_netdevice_notifier(&cma_nb);
4100 rdma_addr_unregister_client(&addr_client);
4101 ib_sa_unregister_client(&sa_client);
4102 unregister_pernet_subsys(&cma_pernet_operations);
4103 destroy_workqueue(cma_wq);
4106 module_init(cma_init);
4107 module_exit(cma_cleanup);