2 * Copyright (c) 2005 Voltaire Inc. All rights reserved.
3 * Copyright (c) 2002-2005, Network Appliance, Inc. All rights reserved.
4 * Copyright (c) 1999-2005, Mellanox Technologies, Inc. All rights reserved.
5 * Copyright (c) 2005-2006 Intel Corporation. All rights reserved.
7 * This software is available to you under a choice of one of two
8 * licenses. You may choose to be licensed under the terms of the GNU
9 * General Public License (GPL) Version 2, available from the file
10 * COPYING in the main directory of this source tree, or the
11 * OpenIB.org BSD license below:
13 * Redistribution and use in source and binary forms, with or
14 * without modification, are permitted provided that the following
17 * - Redistributions of source code must retain the above
18 * copyright notice, this list of conditions and the following
21 * - Redistributions in binary form must reproduce the above
22 * copyright notice, this list of conditions and the following
23 * disclaimer in the documentation and/or other materials
24 * provided with the distribution.
26 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
27 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
28 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
29 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
30 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
31 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
32 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
36 #include <linux/completion.h>
38 #include <linux/in6.h>
39 #include <linux/mutex.h>
40 #include <linux/random.h>
41 #include <linux/igmp.h>
42 #include <linux/idr.h>
43 #include <linux/inetdevice.h>
44 #include <linux/slab.h>
45 #include <linux/module.h>
46 #include <net/route.h>
48 #include <net/net_namespace.h>
49 #include <net/netns/generic.h>
52 #include <net/ip_fib.h>
53 #include <net/ip6_route.h>
55 #include <rdma/rdma_cm.h>
56 #include <rdma/rdma_cm_ib.h>
57 #include <rdma/rdma_netlink.h>
59 #include <rdma/ib_cache.h>
60 #include <rdma/ib_cm.h>
61 #include <rdma/ib_sa.h>
62 #include <rdma/iw_cm.h>
64 #include "core_priv.h"
67 MODULE_AUTHOR("Sean Hefty");
68 MODULE_DESCRIPTION("Generic RDMA CM Agent");
69 MODULE_LICENSE("Dual BSD/GPL");
71 #define CMA_CM_RESPONSE_TIMEOUT 20
72 #define CMA_QUERY_CLASSPORT_INFO_TIMEOUT 3000
73 #define CMA_MAX_CM_RETRIES 15
74 #define CMA_CM_MRA_SETTING (IB_CM_MRA_FLAG_DELAY | 24)
75 #define CMA_IBOE_PACKET_LIFETIME 18
76 #define CMA_PREFERRED_ROCE_GID_TYPE IB_GID_TYPE_ROCE_UDP_ENCAP
78 static const char * const cma_events[] = {
79 [RDMA_CM_EVENT_ADDR_RESOLVED] = "address resolved",
80 [RDMA_CM_EVENT_ADDR_ERROR] = "address error",
81 [RDMA_CM_EVENT_ROUTE_RESOLVED] = "route resolved ",
82 [RDMA_CM_EVENT_ROUTE_ERROR] = "route error",
83 [RDMA_CM_EVENT_CONNECT_REQUEST] = "connect request",
84 [RDMA_CM_EVENT_CONNECT_RESPONSE] = "connect response",
85 [RDMA_CM_EVENT_CONNECT_ERROR] = "connect error",
86 [RDMA_CM_EVENT_UNREACHABLE] = "unreachable",
87 [RDMA_CM_EVENT_REJECTED] = "rejected",
88 [RDMA_CM_EVENT_ESTABLISHED] = "established",
89 [RDMA_CM_EVENT_DISCONNECTED] = "disconnected",
90 [RDMA_CM_EVENT_DEVICE_REMOVAL] = "device removal",
91 [RDMA_CM_EVENT_MULTICAST_JOIN] = "multicast join",
92 [RDMA_CM_EVENT_MULTICAST_ERROR] = "multicast error",
93 [RDMA_CM_EVENT_ADDR_CHANGE] = "address change",
94 [RDMA_CM_EVENT_TIMEWAIT_EXIT] = "timewait exit",
97 const char *__attribute_const__ rdma_event_msg(enum rdma_cm_event_type event)
101 return (index < ARRAY_SIZE(cma_events) && cma_events[index]) ?
102 cma_events[index] : "unrecognized event";
104 EXPORT_SYMBOL(rdma_event_msg);
106 const char *__attribute_const__ rdma_reject_msg(struct rdma_cm_id *id,
109 if (rdma_ib_or_roce(id->device, id->port_num))
110 return ibcm_reject_msg(reason);
112 if (rdma_protocol_iwarp(id->device, id->port_num))
113 return iwcm_reject_msg(reason);
116 return "unrecognized transport";
118 EXPORT_SYMBOL(rdma_reject_msg);
120 bool rdma_is_consumer_reject(struct rdma_cm_id *id, int reason)
122 if (rdma_ib_or_roce(id->device, id->port_num))
123 return reason == IB_CM_REJ_CONSUMER_DEFINED;
125 if (rdma_protocol_iwarp(id->device, id->port_num))
126 return reason == -ECONNREFUSED;
131 EXPORT_SYMBOL(rdma_is_consumer_reject);
133 const void *rdma_consumer_reject_data(struct rdma_cm_id *id,
134 struct rdma_cm_event *ev, u8 *data_len)
138 if (rdma_is_consumer_reject(id, ev->status)) {
139 *data_len = ev->param.conn.private_data_len;
140 p = ev->param.conn.private_data;
147 EXPORT_SYMBOL(rdma_consumer_reject_data);
149 static void cma_add_one(struct ib_device *device);
150 static void cma_remove_one(struct ib_device *device, void *client_data);
152 static struct ib_client cma_client = {
155 .remove = cma_remove_one
158 static struct ib_sa_client sa_client;
159 static struct rdma_addr_client addr_client;
160 static LIST_HEAD(dev_list);
161 static LIST_HEAD(listen_any_list);
162 static DEFINE_MUTEX(lock);
163 static struct workqueue_struct *cma_wq;
164 static unsigned int cma_pernet_id;
173 static struct cma_pernet *cma_pernet(struct net *net)
175 return net_generic(net, cma_pernet_id);
178 static struct idr *cma_pernet_idr(struct net *net, enum rdma_ucm_port_space ps)
180 struct cma_pernet *pernet = cma_pernet(net);
184 return &pernet->tcp_ps;
186 return &pernet->udp_ps;
188 return &pernet->ipoib_ps;
190 return &pernet->ib_ps;
197 struct list_head list;
198 struct ib_device *device;
199 struct completion comp;
201 struct list_head id_list;
202 enum ib_gid_type *default_gid_type;
203 u8 *default_roce_tos;
206 struct rdma_bind_list {
207 enum rdma_ucm_port_space ps;
208 struct hlist_head owners;
212 struct class_port_info_context {
213 struct ib_class_port_info *class_port_info;
214 struct ib_device *device;
215 struct completion done;
216 struct ib_sa_query *sa_query;
220 static int cma_ps_alloc(struct net *net, enum rdma_ucm_port_space ps,
221 struct rdma_bind_list *bind_list, int snum)
223 struct idr *idr = cma_pernet_idr(net, ps);
225 return idr_alloc(idr, bind_list, snum, snum + 1, GFP_KERNEL);
228 static struct rdma_bind_list *cma_ps_find(struct net *net,
229 enum rdma_ucm_port_space ps, int snum)
231 struct idr *idr = cma_pernet_idr(net, ps);
233 return idr_find(idr, snum);
236 static void cma_ps_remove(struct net *net, enum rdma_ucm_port_space ps,
239 struct idr *idr = cma_pernet_idr(net, ps);
241 idr_remove(idr, snum);
248 void cma_ref_dev(struct cma_device *cma_dev)
250 atomic_inc(&cma_dev->refcount);
253 struct cma_device *cma_enum_devices_by_ibdev(cma_device_filter filter,
256 struct cma_device *cma_dev;
257 struct cma_device *found_cma_dev = NULL;
261 list_for_each_entry(cma_dev, &dev_list, list)
262 if (filter(cma_dev->device, cookie)) {
263 found_cma_dev = cma_dev;
268 cma_ref_dev(found_cma_dev);
270 return found_cma_dev;
273 int cma_get_default_gid_type(struct cma_device *cma_dev,
276 if (!rdma_is_port_valid(cma_dev->device, port))
279 return cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)];
282 int cma_set_default_gid_type(struct cma_device *cma_dev,
284 enum ib_gid_type default_gid_type)
286 unsigned long supported_gids;
288 if (!rdma_is_port_valid(cma_dev->device, port))
291 supported_gids = roce_gid_type_mask_support(cma_dev->device, port);
293 if (!(supported_gids & 1 << default_gid_type))
296 cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)] =
302 int cma_get_default_roce_tos(struct cma_device *cma_dev, unsigned int port)
304 if (!rdma_is_port_valid(cma_dev->device, port))
307 return cma_dev->default_roce_tos[port - rdma_start_port(cma_dev->device)];
310 int cma_set_default_roce_tos(struct cma_device *cma_dev, unsigned int port,
313 if (!rdma_is_port_valid(cma_dev->device, port))
316 cma_dev->default_roce_tos[port - rdma_start_port(cma_dev->device)] =
321 struct ib_device *cma_get_ib_dev(struct cma_device *cma_dev)
323 return cma_dev->device;
327 * Device removal can occur at anytime, so we need extra handling to
328 * serialize notifying the user of device removal with other callbacks.
329 * We do this by disabling removal notification while a callback is in process,
330 * and reporting it after the callback completes.
333 struct cma_multicast {
334 struct rdma_id_private *id_priv;
336 struct ib_sa_multicast *ib;
338 struct list_head list;
340 struct sockaddr_storage addr;
347 struct work_struct work;
348 struct rdma_id_private *id;
349 enum rdma_cm_state old_state;
350 enum rdma_cm_state new_state;
351 struct rdma_cm_event event;
354 struct cma_ndev_work {
355 struct work_struct work;
356 struct rdma_id_private *id;
357 struct rdma_cm_event event;
360 struct iboe_mcast_work {
361 struct work_struct work;
362 struct rdma_id_private *id;
363 struct cma_multicast *mc;
376 u8 ip_version; /* IP version: 7:4 */
378 union cma_ip_addr src_addr;
379 union cma_ip_addr dst_addr;
382 #define CMA_VERSION 0x00
384 struct cma_req_info {
385 struct ib_device *device;
387 union ib_gid local_gid;
393 static int cma_comp(struct rdma_id_private *id_priv, enum rdma_cm_state comp)
398 spin_lock_irqsave(&id_priv->lock, flags);
399 ret = (id_priv->state == comp);
400 spin_unlock_irqrestore(&id_priv->lock, flags);
404 static int cma_comp_exch(struct rdma_id_private *id_priv,
405 enum rdma_cm_state comp, enum rdma_cm_state exch)
410 spin_lock_irqsave(&id_priv->lock, flags);
411 if ((ret = (id_priv->state == comp)))
412 id_priv->state = exch;
413 spin_unlock_irqrestore(&id_priv->lock, flags);
417 static enum rdma_cm_state cma_exch(struct rdma_id_private *id_priv,
418 enum rdma_cm_state exch)
421 enum rdma_cm_state old;
423 spin_lock_irqsave(&id_priv->lock, flags);
424 old = id_priv->state;
425 id_priv->state = exch;
426 spin_unlock_irqrestore(&id_priv->lock, flags);
430 static inline u8 cma_get_ip_ver(const struct cma_hdr *hdr)
432 return hdr->ip_version >> 4;
435 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
437 hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
440 static int cma_igmp_send(struct net_device *ndev, union ib_gid *mgid, bool join)
442 struct in_device *in_dev = NULL;
446 in_dev = __in_dev_get_rtnl(ndev);
449 ip_mc_inc_group(in_dev,
450 *(__be32 *)(mgid->raw + 12));
452 ip_mc_dec_group(in_dev,
453 *(__be32 *)(mgid->raw + 12));
457 return (in_dev) ? 0 : -ENODEV;
460 static void _cma_attach_to_dev(struct rdma_id_private *id_priv,
461 struct cma_device *cma_dev)
463 cma_ref_dev(cma_dev);
464 id_priv->cma_dev = cma_dev;
465 id_priv->gid_type = 0;
466 id_priv->id.device = cma_dev->device;
467 id_priv->id.route.addr.dev_addr.transport =
468 rdma_node_get_transport(cma_dev->device->node_type);
469 list_add_tail(&id_priv->list, &cma_dev->id_list);
470 id_priv->res.type = RDMA_RESTRACK_CM_ID;
471 rdma_restrack_add(&id_priv->res);
474 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
475 struct cma_device *cma_dev)
477 _cma_attach_to_dev(id_priv, cma_dev);
479 cma_dev->default_gid_type[id_priv->id.port_num -
480 rdma_start_port(cma_dev->device)];
483 void cma_deref_dev(struct cma_device *cma_dev)
485 if (atomic_dec_and_test(&cma_dev->refcount))
486 complete(&cma_dev->comp);
489 static inline void release_mc(struct kref *kref)
491 struct cma_multicast *mc = container_of(kref, struct cma_multicast, mcref);
493 kfree(mc->multicast.ib);
497 static void cma_release_dev(struct rdma_id_private *id_priv)
500 list_del(&id_priv->list);
501 cma_deref_dev(id_priv->cma_dev);
502 id_priv->cma_dev = NULL;
506 static inline struct sockaddr *cma_src_addr(struct rdma_id_private *id_priv)
508 return (struct sockaddr *) &id_priv->id.route.addr.src_addr;
511 static inline struct sockaddr *cma_dst_addr(struct rdma_id_private *id_priv)
513 return (struct sockaddr *) &id_priv->id.route.addr.dst_addr;
516 static inline unsigned short cma_family(struct rdma_id_private *id_priv)
518 return id_priv->id.route.addr.src_addr.ss_family;
521 static int cma_set_qkey(struct rdma_id_private *id_priv, u32 qkey)
523 struct ib_sa_mcmember_rec rec;
527 if (qkey && id_priv->qkey != qkey)
533 id_priv->qkey = qkey;
537 switch (id_priv->id.ps) {
540 id_priv->qkey = RDMA_UDP_QKEY;
543 ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid);
544 ret = ib_sa_get_mcmember_rec(id_priv->id.device,
545 id_priv->id.port_num, &rec.mgid,
548 id_priv->qkey = be32_to_cpu(rec.qkey);
556 static void cma_translate_ib(struct sockaddr_ib *sib, struct rdma_dev_addr *dev_addr)
558 dev_addr->dev_type = ARPHRD_INFINIBAND;
559 rdma_addr_set_sgid(dev_addr, (union ib_gid *) &sib->sib_addr);
560 ib_addr_set_pkey(dev_addr, ntohs(sib->sib_pkey));
563 static int cma_translate_addr(struct sockaddr *addr, struct rdma_dev_addr *dev_addr)
567 if (addr->sa_family != AF_IB) {
568 ret = rdma_translate_ip(addr, dev_addr);
570 cma_translate_ib((struct sockaddr_ib *) addr, dev_addr);
577 static inline int cma_validate_port(struct ib_device *device, u8 port,
578 enum ib_gid_type gid_type,
580 struct rdma_id_private *id_priv)
582 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
583 int bound_if_index = dev_addr->bound_dev_if;
584 int dev_type = dev_addr->dev_type;
585 struct net_device *ndev = NULL;
588 if ((dev_type == ARPHRD_INFINIBAND) && !rdma_protocol_ib(device, port))
591 if ((dev_type != ARPHRD_INFINIBAND) && rdma_protocol_ib(device, port))
594 if (dev_type == ARPHRD_ETHER && rdma_protocol_roce(device, port)) {
595 ndev = dev_get_by_index(dev_addr->net, bound_if_index);
599 gid_type = IB_GID_TYPE_IB;
602 ret = ib_find_cached_gid_by_port(device, gid, gid_type, port,
611 static int cma_acquire_dev(struct rdma_id_private *id_priv,
612 struct rdma_id_private *listen_id_priv)
614 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
615 struct cma_device *cma_dev;
616 union ib_gid gid, iboe_gid, *gidp;
620 if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
621 id_priv->id.ps == RDMA_PS_IPOIB)
625 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
628 memcpy(&gid, dev_addr->src_dev_addr +
629 rdma_addr_gid_offset(dev_addr), sizeof gid);
631 if (listen_id_priv) {
632 cma_dev = listen_id_priv->cma_dev;
633 port = listen_id_priv->id.port_num;
634 gidp = rdma_protocol_roce(cma_dev->device, port) ?
637 ret = cma_validate_port(cma_dev->device, port,
638 rdma_protocol_ib(cma_dev->device, port) ?
640 listen_id_priv->gid_type, gidp,
643 id_priv->id.port_num = port;
648 list_for_each_entry(cma_dev, &dev_list, list) {
649 for (port = 1; port <= cma_dev->device->phys_port_cnt; ++port) {
650 if (listen_id_priv &&
651 listen_id_priv->cma_dev == cma_dev &&
652 listen_id_priv->id.port_num == port)
655 gidp = rdma_protocol_roce(cma_dev->device, port) ?
658 ret = cma_validate_port(cma_dev->device, port,
659 rdma_protocol_ib(cma_dev->device, port) ?
661 cma_dev->default_gid_type[port - 1],
664 id_priv->id.port_num = port;
672 cma_attach_to_dev(id_priv, cma_dev);
679 * Select the source IB device and address to reach the destination IB address.
681 static int cma_resolve_ib_dev(struct rdma_id_private *id_priv)
683 struct cma_device *cma_dev, *cur_dev;
684 struct sockaddr_ib *addr;
685 union ib_gid gid, sgid, *dgid;
688 enum ib_port_state port_state;
692 addr = (struct sockaddr_ib *) cma_dst_addr(id_priv);
693 dgid = (union ib_gid *) &addr->sib_addr;
694 pkey = ntohs(addr->sib_pkey);
696 list_for_each_entry(cur_dev, &dev_list, list) {
697 for (p = 1; p <= cur_dev->device->phys_port_cnt; ++p) {
698 if (!rdma_cap_af_ib(cur_dev->device, p))
701 if (ib_find_cached_pkey(cur_dev->device, p, pkey, &index))
704 if (ib_get_cached_port_state(cur_dev->device, p, &port_state))
706 for (i = 0; !ib_get_cached_gid(cur_dev->device, p, i,
709 if (!memcmp(&gid, dgid, sizeof(gid))) {
712 id_priv->id.port_num = p;
716 if (!cma_dev && (gid.global.subnet_prefix ==
717 dgid->global.subnet_prefix) &&
718 port_state == IB_PORT_ACTIVE) {
721 id_priv->id.port_num = p;
731 cma_attach_to_dev(id_priv, cma_dev);
732 addr = (struct sockaddr_ib *) cma_src_addr(id_priv);
733 memcpy(&addr->sib_addr, &sgid, sizeof sgid);
734 cma_translate_ib(addr, &id_priv->id.route.addr.dev_addr);
738 static void cma_deref_id(struct rdma_id_private *id_priv)
740 if (atomic_dec_and_test(&id_priv->refcount))
741 complete(&id_priv->comp);
744 struct rdma_cm_id *__rdma_create_id(struct net *net,
745 rdma_cm_event_handler event_handler,
746 void *context, enum rdma_ucm_port_space ps,
747 enum ib_qp_type qp_type, const char *caller)
749 struct rdma_id_private *id_priv;
751 id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
753 return ERR_PTR(-ENOMEM);
756 id_priv->res.kern_name = caller;
758 rdma_restrack_set_task(&id_priv->res, current);
759 id_priv->state = RDMA_CM_IDLE;
760 id_priv->id.context = context;
761 id_priv->id.event_handler = event_handler;
763 id_priv->id.qp_type = qp_type;
764 id_priv->tos_set = false;
765 spin_lock_init(&id_priv->lock);
766 mutex_init(&id_priv->qp_mutex);
767 init_completion(&id_priv->comp);
768 atomic_set(&id_priv->refcount, 1);
769 mutex_init(&id_priv->handler_mutex);
770 INIT_LIST_HEAD(&id_priv->listen_list);
771 INIT_LIST_HEAD(&id_priv->mc_list);
772 get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
773 id_priv->id.route.addr.dev_addr.net = get_net(net);
774 id_priv->seq_num &= 0x00ffffff;
778 EXPORT_SYMBOL(__rdma_create_id);
780 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
782 struct ib_qp_attr qp_attr;
783 int qp_attr_mask, ret;
785 qp_attr.qp_state = IB_QPS_INIT;
786 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
790 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
794 qp_attr.qp_state = IB_QPS_RTR;
795 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
799 qp_attr.qp_state = IB_QPS_RTS;
801 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
806 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
808 struct ib_qp_attr qp_attr;
809 int qp_attr_mask, ret;
811 qp_attr.qp_state = IB_QPS_INIT;
812 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
816 return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
819 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
820 struct ib_qp_init_attr *qp_init_attr)
822 struct rdma_id_private *id_priv;
826 id_priv = container_of(id, struct rdma_id_private, id);
827 if (id->device != pd->device)
830 qp_init_attr->port_num = id->port_num;
831 qp = ib_create_qp(pd, qp_init_attr);
835 if (id->qp_type == IB_QPT_UD)
836 ret = cma_init_ud_qp(id_priv, qp);
838 ret = cma_init_conn_qp(id_priv, qp);
843 id_priv->qp_num = qp->qp_num;
844 id_priv->srq = (qp->srq != NULL);
850 EXPORT_SYMBOL(rdma_create_qp);
852 void rdma_destroy_qp(struct rdma_cm_id *id)
854 struct rdma_id_private *id_priv;
856 id_priv = container_of(id, struct rdma_id_private, id);
857 mutex_lock(&id_priv->qp_mutex);
858 ib_destroy_qp(id_priv->id.qp);
859 id_priv->id.qp = NULL;
860 mutex_unlock(&id_priv->qp_mutex);
862 EXPORT_SYMBOL(rdma_destroy_qp);
864 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
865 struct rdma_conn_param *conn_param)
867 struct ib_qp_attr qp_attr;
868 int qp_attr_mask, ret;
871 mutex_lock(&id_priv->qp_mutex);
872 if (!id_priv->id.qp) {
877 /* Need to update QP attributes from default values. */
878 qp_attr.qp_state = IB_QPS_INIT;
879 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
883 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
887 qp_attr.qp_state = IB_QPS_RTR;
888 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
892 ret = ib_query_gid(id_priv->id.device, id_priv->id.port_num,
893 rdma_ah_read_grh(&qp_attr.ah_attr)->sgid_index,
898 BUG_ON(id_priv->cma_dev->device != id_priv->id.device);
901 qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
902 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
904 mutex_unlock(&id_priv->qp_mutex);
908 static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
909 struct rdma_conn_param *conn_param)
911 struct ib_qp_attr qp_attr;
912 int qp_attr_mask, ret;
914 mutex_lock(&id_priv->qp_mutex);
915 if (!id_priv->id.qp) {
920 qp_attr.qp_state = IB_QPS_RTS;
921 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
926 qp_attr.max_rd_atomic = conn_param->initiator_depth;
927 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
929 mutex_unlock(&id_priv->qp_mutex);
933 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
935 struct ib_qp_attr qp_attr;
938 mutex_lock(&id_priv->qp_mutex);
939 if (!id_priv->id.qp) {
944 qp_attr.qp_state = IB_QPS_ERR;
945 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
947 mutex_unlock(&id_priv->qp_mutex);
951 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
952 struct ib_qp_attr *qp_attr, int *qp_attr_mask)
954 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
958 if (rdma_cap_eth_ah(id_priv->id.device, id_priv->id.port_num))
961 pkey = ib_addr_get_pkey(dev_addr);
963 ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
964 pkey, &qp_attr->pkey_index);
968 qp_attr->port_num = id_priv->id.port_num;
969 *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
971 if (id_priv->id.qp_type == IB_QPT_UD) {
972 ret = cma_set_qkey(id_priv, 0);
976 qp_attr->qkey = id_priv->qkey;
977 *qp_attr_mask |= IB_QP_QKEY;
979 qp_attr->qp_access_flags = 0;
980 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
985 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
988 struct rdma_id_private *id_priv;
991 id_priv = container_of(id, struct rdma_id_private, id);
992 if (rdma_cap_ib_cm(id->device, id->port_num)) {
993 if (!id_priv->cm_id.ib || (id_priv->id.qp_type == IB_QPT_UD))
994 ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
996 ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
999 if (qp_attr->qp_state == IB_QPS_RTR)
1000 qp_attr->rq_psn = id_priv->seq_num;
1001 } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
1002 if (!id_priv->cm_id.iw) {
1003 qp_attr->qp_access_flags = 0;
1004 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
1006 ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
1008 qp_attr->port_num = id_priv->id.port_num;
1009 *qp_attr_mask |= IB_QP_PORT;
1015 EXPORT_SYMBOL(rdma_init_qp_attr);
1017 static inline int cma_zero_addr(struct sockaddr *addr)
1019 switch (addr->sa_family) {
1021 return ipv4_is_zeronet(((struct sockaddr_in *)addr)->sin_addr.s_addr);
1023 return ipv6_addr_any(&((struct sockaddr_in6 *) addr)->sin6_addr);
1025 return ib_addr_any(&((struct sockaddr_ib *) addr)->sib_addr);
1031 static inline int cma_loopback_addr(struct sockaddr *addr)
1033 switch (addr->sa_family) {
1035 return ipv4_is_loopback(((struct sockaddr_in *) addr)->sin_addr.s_addr);
1037 return ipv6_addr_loopback(&((struct sockaddr_in6 *) addr)->sin6_addr);
1039 return ib_addr_loopback(&((struct sockaddr_ib *) addr)->sib_addr);
1045 static inline int cma_any_addr(struct sockaddr *addr)
1047 return cma_zero_addr(addr) || cma_loopback_addr(addr);
1050 static int cma_addr_cmp(struct sockaddr *src, struct sockaddr *dst)
1052 if (src->sa_family != dst->sa_family)
1055 switch (src->sa_family) {
1057 return ((struct sockaddr_in *) src)->sin_addr.s_addr !=
1058 ((struct sockaddr_in *) dst)->sin_addr.s_addr;
1060 return ipv6_addr_cmp(&((struct sockaddr_in6 *) src)->sin6_addr,
1061 &((struct sockaddr_in6 *) dst)->sin6_addr);
1063 return ib_addr_cmp(&((struct sockaddr_ib *) src)->sib_addr,
1064 &((struct sockaddr_ib *) dst)->sib_addr);
1068 static __be16 cma_port(struct sockaddr *addr)
1070 struct sockaddr_ib *sib;
1072 switch (addr->sa_family) {
1074 return ((struct sockaddr_in *) addr)->sin_port;
1076 return ((struct sockaddr_in6 *) addr)->sin6_port;
1078 sib = (struct sockaddr_ib *) addr;
1079 return htons((u16) (be64_to_cpu(sib->sib_sid) &
1080 be64_to_cpu(sib->sib_sid_mask)));
1086 static inline int cma_any_port(struct sockaddr *addr)
1088 return !cma_port(addr);
1091 static void cma_save_ib_info(struct sockaddr *src_addr,
1092 struct sockaddr *dst_addr,
1093 struct rdma_cm_id *listen_id,
1094 struct sa_path_rec *path)
1096 struct sockaddr_ib *listen_ib, *ib;
1098 listen_ib = (struct sockaddr_ib *) &listen_id->route.addr.src_addr;
1100 ib = (struct sockaddr_ib *)src_addr;
1101 ib->sib_family = AF_IB;
1103 ib->sib_pkey = path->pkey;
1104 ib->sib_flowinfo = path->flow_label;
1105 memcpy(&ib->sib_addr, &path->sgid, 16);
1106 ib->sib_sid = path->service_id;
1107 ib->sib_scope_id = 0;
1109 ib->sib_pkey = listen_ib->sib_pkey;
1110 ib->sib_flowinfo = listen_ib->sib_flowinfo;
1111 ib->sib_addr = listen_ib->sib_addr;
1112 ib->sib_sid = listen_ib->sib_sid;
1113 ib->sib_scope_id = listen_ib->sib_scope_id;
1115 ib->sib_sid_mask = cpu_to_be64(0xffffffffffffffffULL);
1118 ib = (struct sockaddr_ib *)dst_addr;
1119 ib->sib_family = AF_IB;
1121 ib->sib_pkey = path->pkey;
1122 ib->sib_flowinfo = path->flow_label;
1123 memcpy(&ib->sib_addr, &path->dgid, 16);
1128 static void cma_save_ip4_info(struct sockaddr_in *src_addr,
1129 struct sockaddr_in *dst_addr,
1130 struct cma_hdr *hdr,
1134 *src_addr = (struct sockaddr_in) {
1135 .sin_family = AF_INET,
1136 .sin_addr.s_addr = hdr->dst_addr.ip4.addr,
1137 .sin_port = local_port,
1142 *dst_addr = (struct sockaddr_in) {
1143 .sin_family = AF_INET,
1144 .sin_addr.s_addr = hdr->src_addr.ip4.addr,
1145 .sin_port = hdr->port,
1150 static void cma_save_ip6_info(struct sockaddr_in6 *src_addr,
1151 struct sockaddr_in6 *dst_addr,
1152 struct cma_hdr *hdr,
1156 *src_addr = (struct sockaddr_in6) {
1157 .sin6_family = AF_INET6,
1158 .sin6_addr = hdr->dst_addr.ip6,
1159 .sin6_port = local_port,
1164 *dst_addr = (struct sockaddr_in6) {
1165 .sin6_family = AF_INET6,
1166 .sin6_addr = hdr->src_addr.ip6,
1167 .sin6_port = hdr->port,
1172 static u16 cma_port_from_service_id(__be64 service_id)
1174 return (u16)be64_to_cpu(service_id);
1177 static int cma_save_ip_info(struct sockaddr *src_addr,
1178 struct sockaddr *dst_addr,
1179 struct ib_cm_event *ib_event,
1182 struct cma_hdr *hdr;
1185 hdr = ib_event->private_data;
1186 if (hdr->cma_version != CMA_VERSION)
1189 port = htons(cma_port_from_service_id(service_id));
1191 switch (cma_get_ip_ver(hdr)) {
1193 cma_save_ip4_info((struct sockaddr_in *)src_addr,
1194 (struct sockaddr_in *)dst_addr, hdr, port);
1197 cma_save_ip6_info((struct sockaddr_in6 *)src_addr,
1198 (struct sockaddr_in6 *)dst_addr, hdr, port);
1201 return -EAFNOSUPPORT;
1207 static int cma_save_net_info(struct sockaddr *src_addr,
1208 struct sockaddr *dst_addr,
1209 struct rdma_cm_id *listen_id,
1210 struct ib_cm_event *ib_event,
1211 sa_family_t sa_family, __be64 service_id)
1213 if (sa_family == AF_IB) {
1214 if (ib_event->event == IB_CM_REQ_RECEIVED)
1215 cma_save_ib_info(src_addr, dst_addr, listen_id,
1216 ib_event->param.req_rcvd.primary_path);
1217 else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED)
1218 cma_save_ib_info(src_addr, dst_addr, listen_id, NULL);
1222 return cma_save_ip_info(src_addr, dst_addr, ib_event, service_id);
1225 static int cma_save_req_info(const struct ib_cm_event *ib_event,
1226 struct cma_req_info *req)
1228 const struct ib_cm_req_event_param *req_param =
1229 &ib_event->param.req_rcvd;
1230 const struct ib_cm_sidr_req_event_param *sidr_param =
1231 &ib_event->param.sidr_req_rcvd;
1233 switch (ib_event->event) {
1234 case IB_CM_REQ_RECEIVED:
1235 req->device = req_param->listen_id->device;
1236 req->port = req_param->port;
1237 memcpy(&req->local_gid, &req_param->primary_path->sgid,
1238 sizeof(req->local_gid));
1239 req->has_gid = true;
1240 req->service_id = req_param->primary_path->service_id;
1241 req->pkey = be16_to_cpu(req_param->primary_path->pkey);
1242 if (req->pkey != req_param->bth_pkey)
1243 pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and primary path P_Key (0x%x)\n"
1244 "RDMA CMA: in the future this may cause the request to be dropped\n",
1245 req_param->bth_pkey, req->pkey);
1247 case IB_CM_SIDR_REQ_RECEIVED:
1248 req->device = sidr_param->listen_id->device;
1249 req->port = sidr_param->port;
1250 req->has_gid = false;
1251 req->service_id = sidr_param->service_id;
1252 req->pkey = sidr_param->pkey;
1253 if (req->pkey != sidr_param->bth_pkey)
1254 pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and SIDR request payload P_Key (0x%x)\n"
1255 "RDMA CMA: in the future this may cause the request to be dropped\n",
1256 sidr_param->bth_pkey, req->pkey);
1265 static bool validate_ipv4_net_dev(struct net_device *net_dev,
1266 const struct sockaddr_in *dst_addr,
1267 const struct sockaddr_in *src_addr)
1269 __be32 daddr = dst_addr->sin_addr.s_addr,
1270 saddr = src_addr->sin_addr.s_addr;
1271 struct fib_result res;
1276 if (ipv4_is_multicast(saddr) || ipv4_is_lbcast(saddr) ||
1277 ipv4_is_lbcast(daddr) || ipv4_is_zeronet(saddr) ||
1278 ipv4_is_zeronet(daddr) || ipv4_is_loopback(daddr) ||
1279 ipv4_is_loopback(saddr))
1282 memset(&fl4, 0, sizeof(fl4));
1283 fl4.flowi4_iif = net_dev->ifindex;
1288 err = fib_lookup(dev_net(net_dev), &fl4, &res, 0);
1289 ret = err == 0 && FIB_RES_DEV(res) == net_dev;
1295 static bool validate_ipv6_net_dev(struct net_device *net_dev,
1296 const struct sockaddr_in6 *dst_addr,
1297 const struct sockaddr_in6 *src_addr)
1299 #if IS_ENABLED(CONFIG_IPV6)
1300 const int strict = ipv6_addr_type(&dst_addr->sin6_addr) &
1301 IPV6_ADDR_LINKLOCAL;
1302 struct rt6_info *rt = rt6_lookup(dev_net(net_dev), &dst_addr->sin6_addr,
1303 &src_addr->sin6_addr, net_dev->ifindex,
1310 ret = rt->rt6i_idev->dev == net_dev;
1319 static bool validate_net_dev(struct net_device *net_dev,
1320 const struct sockaddr *daddr,
1321 const struct sockaddr *saddr)
1323 const struct sockaddr_in *daddr4 = (const struct sockaddr_in *)daddr;
1324 const struct sockaddr_in *saddr4 = (const struct sockaddr_in *)saddr;
1325 const struct sockaddr_in6 *daddr6 = (const struct sockaddr_in6 *)daddr;
1326 const struct sockaddr_in6 *saddr6 = (const struct sockaddr_in6 *)saddr;
1328 switch (daddr->sa_family) {
1330 return saddr->sa_family == AF_INET &&
1331 validate_ipv4_net_dev(net_dev, daddr4, saddr4);
1334 return saddr->sa_family == AF_INET6 &&
1335 validate_ipv6_net_dev(net_dev, daddr6, saddr6);
1342 static struct net_device *cma_get_net_dev(struct ib_cm_event *ib_event,
1343 const struct cma_req_info *req)
1345 struct sockaddr_storage listen_addr_storage, src_addr_storage;
1346 struct sockaddr *listen_addr = (struct sockaddr *)&listen_addr_storage,
1347 *src_addr = (struct sockaddr *)&src_addr_storage;
1348 struct net_device *net_dev;
1349 const union ib_gid *gid = req->has_gid ? &req->local_gid : NULL;
1352 err = cma_save_ip_info(listen_addr, src_addr, ib_event,
1355 return ERR_PTR(err);
1357 net_dev = ib_get_net_dev_by_params(req->device, req->port, req->pkey,
1360 return ERR_PTR(-ENODEV);
1362 if (!validate_net_dev(net_dev, listen_addr, src_addr)) {
1364 return ERR_PTR(-EHOSTUNREACH);
1370 static enum rdma_ucm_port_space rdma_ps_from_service_id(__be64 service_id)
1372 return (be64_to_cpu(service_id) >> 16) & 0xffff;
1375 static bool cma_match_private_data(struct rdma_id_private *id_priv,
1376 const struct cma_hdr *hdr)
1378 struct sockaddr *addr = cma_src_addr(id_priv);
1380 struct in6_addr ip6_addr;
1382 if (cma_any_addr(addr) && !id_priv->afonly)
1385 switch (addr->sa_family) {
1387 ip4_addr = ((struct sockaddr_in *)addr)->sin_addr.s_addr;
1388 if (cma_get_ip_ver(hdr) != 4)
1390 if (!cma_any_addr(addr) &&
1391 hdr->dst_addr.ip4.addr != ip4_addr)
1395 ip6_addr = ((struct sockaddr_in6 *)addr)->sin6_addr;
1396 if (cma_get_ip_ver(hdr) != 6)
1398 if (!cma_any_addr(addr) &&
1399 memcmp(&hdr->dst_addr.ip6, &ip6_addr, sizeof(ip6_addr)))
1411 static bool cma_protocol_roce(const struct rdma_cm_id *id)
1413 struct ib_device *device = id->device;
1414 const int port_num = id->port_num ?: rdma_start_port(device);
1416 return rdma_protocol_roce(device, port_num);
1419 static bool cma_match_net_dev(const struct rdma_cm_id *id,
1420 const struct net_device *net_dev,
1423 const struct rdma_addr *addr = &id->route.addr;
1426 /* This request is an AF_IB request or a RoCE request */
1427 return (!id->port_num || id->port_num == port_num) &&
1428 (addr->src_addr.ss_family == AF_IB ||
1429 rdma_protocol_roce(id->device, port_num));
1431 return !addr->dev_addr.bound_dev_if ||
1432 (net_eq(dev_net(net_dev), addr->dev_addr.net) &&
1433 addr->dev_addr.bound_dev_if == net_dev->ifindex);
1436 static struct rdma_id_private *cma_find_listener(
1437 const struct rdma_bind_list *bind_list,
1438 const struct ib_cm_id *cm_id,
1439 const struct ib_cm_event *ib_event,
1440 const struct cma_req_info *req,
1441 const struct net_device *net_dev)
1443 struct rdma_id_private *id_priv, *id_priv_dev;
1446 return ERR_PTR(-EINVAL);
1448 hlist_for_each_entry(id_priv, &bind_list->owners, node) {
1449 if (cma_match_private_data(id_priv, ib_event->private_data)) {
1450 if (id_priv->id.device == cm_id->device &&
1451 cma_match_net_dev(&id_priv->id, net_dev, req->port))
1453 list_for_each_entry(id_priv_dev,
1454 &id_priv->listen_list,
1456 if (id_priv_dev->id.device == cm_id->device &&
1457 cma_match_net_dev(&id_priv_dev->id, net_dev, req->port))
1463 return ERR_PTR(-EINVAL);
1466 static struct rdma_id_private *cma_id_from_event(struct ib_cm_id *cm_id,
1467 struct ib_cm_event *ib_event,
1468 struct net_device **net_dev)
1470 struct cma_req_info req;
1471 struct rdma_bind_list *bind_list;
1472 struct rdma_id_private *id_priv;
1475 err = cma_save_req_info(ib_event, &req);
1477 return ERR_PTR(err);
1479 *net_dev = cma_get_net_dev(ib_event, &req);
1480 if (IS_ERR(*net_dev)) {
1481 if (PTR_ERR(*net_dev) == -EAFNOSUPPORT) {
1482 /* Assuming the protocol is AF_IB */
1484 } else if (rdma_protocol_roce(req.device, req.port)) {
1485 /* TODO find the net dev matching the request parameters
1486 * through the RoCE GID table */
1489 return ERR_CAST(*net_dev);
1493 bind_list = cma_ps_find(*net_dev ? dev_net(*net_dev) : &init_net,
1494 rdma_ps_from_service_id(req.service_id),
1495 cma_port_from_service_id(req.service_id));
1496 id_priv = cma_find_listener(bind_list, cm_id, ib_event, &req, *net_dev);
1497 if (IS_ERR(id_priv) && *net_dev) {
1505 static inline u8 cma_user_data_offset(struct rdma_id_private *id_priv)
1507 return cma_family(id_priv) == AF_IB ? 0 : sizeof(struct cma_hdr);
1510 static void cma_cancel_route(struct rdma_id_private *id_priv)
1512 if (rdma_cap_ib_sa(id_priv->id.device, id_priv->id.port_num)) {
1514 ib_sa_cancel_query(id_priv->query_id, id_priv->query);
1518 static void cma_cancel_listens(struct rdma_id_private *id_priv)
1520 struct rdma_id_private *dev_id_priv;
1523 * Remove from listen_any_list to prevent added devices from spawning
1524 * additional listen requests.
1527 list_del(&id_priv->list);
1529 while (!list_empty(&id_priv->listen_list)) {
1530 dev_id_priv = list_entry(id_priv->listen_list.next,
1531 struct rdma_id_private, listen_list);
1532 /* sync with device removal to avoid duplicate destruction */
1533 list_del_init(&dev_id_priv->list);
1534 list_del(&dev_id_priv->listen_list);
1535 mutex_unlock(&lock);
1537 rdma_destroy_id(&dev_id_priv->id);
1540 mutex_unlock(&lock);
1543 static void cma_cancel_operation(struct rdma_id_private *id_priv,
1544 enum rdma_cm_state state)
1547 case RDMA_CM_ADDR_QUERY:
1548 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
1550 case RDMA_CM_ROUTE_QUERY:
1551 cma_cancel_route(id_priv);
1553 case RDMA_CM_LISTEN:
1554 if (cma_any_addr(cma_src_addr(id_priv)) && !id_priv->cma_dev)
1555 cma_cancel_listens(id_priv);
1562 static void cma_release_port(struct rdma_id_private *id_priv)
1564 struct rdma_bind_list *bind_list = id_priv->bind_list;
1565 struct net *net = id_priv->id.route.addr.dev_addr.net;
1571 hlist_del(&id_priv->node);
1572 if (hlist_empty(&bind_list->owners)) {
1573 cma_ps_remove(net, bind_list->ps, bind_list->port);
1576 mutex_unlock(&lock);
1579 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
1581 struct cma_multicast *mc;
1583 while (!list_empty(&id_priv->mc_list)) {
1584 mc = container_of(id_priv->mc_list.next,
1585 struct cma_multicast, list);
1586 list_del(&mc->list);
1587 if (rdma_cap_ib_mcast(id_priv->cma_dev->device,
1588 id_priv->id.port_num)) {
1589 ib_sa_free_multicast(mc->multicast.ib);
1592 if (mc->igmp_joined) {
1593 struct rdma_dev_addr *dev_addr =
1594 &id_priv->id.route.addr.dev_addr;
1595 struct net_device *ndev = NULL;
1597 if (dev_addr->bound_dev_if)
1598 ndev = dev_get_by_index(&init_net,
1599 dev_addr->bound_dev_if);
1602 &mc->multicast.ib->rec.mgid,
1607 kref_put(&mc->mcref, release_mc);
1612 void rdma_destroy_id(struct rdma_cm_id *id)
1614 struct rdma_id_private *id_priv;
1615 enum rdma_cm_state state;
1617 id_priv = container_of(id, struct rdma_id_private, id);
1618 state = cma_exch(id_priv, RDMA_CM_DESTROYING);
1619 cma_cancel_operation(id_priv, state);
1622 * Wait for any active callback to finish. New callbacks will find
1623 * the id_priv state set to destroying and abort.
1625 mutex_lock(&id_priv->handler_mutex);
1626 mutex_unlock(&id_priv->handler_mutex);
1628 if (id_priv->cma_dev) {
1629 rdma_restrack_del(&id_priv->res);
1630 if (rdma_cap_ib_cm(id_priv->id.device, 1)) {
1631 if (id_priv->cm_id.ib)
1632 ib_destroy_cm_id(id_priv->cm_id.ib);
1633 } else if (rdma_cap_iw_cm(id_priv->id.device, 1)) {
1634 if (id_priv->cm_id.iw)
1635 iw_destroy_cm_id(id_priv->cm_id.iw);
1637 cma_leave_mc_groups(id_priv);
1638 cma_release_dev(id_priv);
1641 cma_release_port(id_priv);
1642 cma_deref_id(id_priv);
1643 wait_for_completion(&id_priv->comp);
1645 if (id_priv->internal_id)
1646 cma_deref_id(id_priv->id.context);
1648 kfree(id_priv->id.route.path_rec);
1649 put_net(id_priv->id.route.addr.dev_addr.net);
1652 EXPORT_SYMBOL(rdma_destroy_id);
1654 static int cma_rep_recv(struct rdma_id_private *id_priv)
1658 ret = cma_modify_qp_rtr(id_priv, NULL);
1662 ret = cma_modify_qp_rts(id_priv, NULL);
1666 ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
1672 pr_debug_ratelimited("RDMA CM: CONNECT_ERROR: failed to handle reply. status %d\n", ret);
1673 cma_modify_qp_err(id_priv);
1674 ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
1679 static void cma_set_rep_event_data(struct rdma_cm_event *event,
1680 struct ib_cm_rep_event_param *rep_data,
1683 event->param.conn.private_data = private_data;
1684 event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
1685 event->param.conn.responder_resources = rep_data->responder_resources;
1686 event->param.conn.initiator_depth = rep_data->initiator_depth;
1687 event->param.conn.flow_control = rep_data->flow_control;
1688 event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
1689 event->param.conn.srq = rep_data->srq;
1690 event->param.conn.qp_num = rep_data->remote_qpn;
1693 static int cma_ib_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1695 struct rdma_id_private *id_priv = cm_id->context;
1696 struct rdma_cm_event event;
1699 mutex_lock(&id_priv->handler_mutex);
1700 if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
1701 id_priv->state != RDMA_CM_CONNECT) ||
1702 (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
1703 id_priv->state != RDMA_CM_DISCONNECT))
1706 memset(&event, 0, sizeof event);
1707 switch (ib_event->event) {
1708 case IB_CM_REQ_ERROR:
1709 case IB_CM_REP_ERROR:
1710 event.event = RDMA_CM_EVENT_UNREACHABLE;
1711 event.status = -ETIMEDOUT;
1713 case IB_CM_REP_RECEIVED:
1714 if (cma_comp(id_priv, RDMA_CM_CONNECT) &&
1715 (id_priv->id.qp_type != IB_QPT_UD))
1716 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
1717 if (id_priv->id.qp) {
1718 event.status = cma_rep_recv(id_priv);
1719 event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
1720 RDMA_CM_EVENT_ESTABLISHED;
1722 event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
1724 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
1725 ib_event->private_data);
1727 case IB_CM_RTU_RECEIVED:
1728 case IB_CM_USER_ESTABLISHED:
1729 event.event = RDMA_CM_EVENT_ESTABLISHED;
1731 case IB_CM_DREQ_ERROR:
1732 event.status = -ETIMEDOUT; /* fall through */
1733 case IB_CM_DREQ_RECEIVED:
1734 case IB_CM_DREP_RECEIVED:
1735 if (!cma_comp_exch(id_priv, RDMA_CM_CONNECT,
1736 RDMA_CM_DISCONNECT))
1738 event.event = RDMA_CM_EVENT_DISCONNECTED;
1740 case IB_CM_TIMEWAIT_EXIT:
1741 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
1743 case IB_CM_MRA_RECEIVED:
1746 case IB_CM_REJ_RECEIVED:
1747 pr_debug_ratelimited("RDMA CM: REJECTED: %s\n", rdma_reject_msg(&id_priv->id,
1748 ib_event->param.rej_rcvd.reason));
1749 cma_modify_qp_err(id_priv);
1750 event.status = ib_event->param.rej_rcvd.reason;
1751 event.event = RDMA_CM_EVENT_REJECTED;
1752 event.param.conn.private_data = ib_event->private_data;
1753 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
1756 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
1761 ret = id_priv->id.event_handler(&id_priv->id, &event);
1763 /* Destroy the CM ID by returning a non-zero value. */
1764 id_priv->cm_id.ib = NULL;
1765 cma_exch(id_priv, RDMA_CM_DESTROYING);
1766 mutex_unlock(&id_priv->handler_mutex);
1767 rdma_destroy_id(&id_priv->id);
1771 mutex_unlock(&id_priv->handler_mutex);
1775 static struct rdma_id_private *cma_new_conn_id(struct rdma_cm_id *listen_id,
1776 struct ib_cm_event *ib_event,
1777 struct net_device *net_dev)
1779 struct rdma_id_private *listen_id_priv;
1780 struct rdma_id_private *id_priv;
1781 struct rdma_cm_id *id;
1782 struct rdma_route *rt;
1783 const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
1784 struct sa_path_rec *path = ib_event->param.req_rcvd.primary_path;
1785 const __be64 service_id =
1786 ib_event->param.req_rcvd.primary_path->service_id;
1789 listen_id_priv = container_of(listen_id, struct rdma_id_private, id);
1790 id = __rdma_create_id(listen_id->route.addr.dev_addr.net,
1791 listen_id->event_handler, listen_id->context,
1792 listen_id->ps, ib_event->param.req_rcvd.qp_type,
1793 listen_id_priv->res.kern_name);
1797 id_priv = container_of(id, struct rdma_id_private, id);
1798 if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
1799 (struct sockaddr *)&id->route.addr.dst_addr,
1800 listen_id, ib_event, ss_family, service_id))
1804 rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
1805 rt->path_rec = kmalloc(sizeof *rt->path_rec * rt->num_paths,
1810 rt->path_rec[0] = *path;
1811 if (rt->num_paths == 2)
1812 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
1815 rdma_copy_addr(&rt->addr.dev_addr, net_dev, NULL);
1817 if (!cma_protocol_roce(listen_id) &&
1818 cma_any_addr(cma_src_addr(id_priv))) {
1819 rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
1820 rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
1821 ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
1822 } else if (!cma_any_addr(cma_src_addr(id_priv))) {
1823 ret = cma_translate_addr(cma_src_addr(id_priv), &rt->addr.dev_addr);
1828 rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
1830 id_priv->state = RDMA_CM_CONNECT;
1834 rdma_destroy_id(id);
1838 static struct rdma_id_private *cma_new_udp_id(struct rdma_cm_id *listen_id,
1839 struct ib_cm_event *ib_event,
1840 struct net_device *net_dev)
1842 struct rdma_id_private *listen_id_priv;
1843 struct rdma_id_private *id_priv;
1844 struct rdma_cm_id *id;
1845 const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
1846 struct net *net = listen_id->route.addr.dev_addr.net;
1849 listen_id_priv = container_of(listen_id, struct rdma_id_private, id);
1850 id = __rdma_create_id(net, listen_id->event_handler, listen_id->context,
1851 listen_id->ps, IB_QPT_UD,
1852 listen_id_priv->res.kern_name);
1856 id_priv = container_of(id, struct rdma_id_private, id);
1857 if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
1858 (struct sockaddr *)&id->route.addr.dst_addr,
1859 listen_id, ib_event, ss_family,
1860 ib_event->param.sidr_req_rcvd.service_id))
1864 rdma_copy_addr(&id->route.addr.dev_addr, net_dev, NULL);
1866 if (!cma_any_addr(cma_src_addr(id_priv))) {
1867 ret = cma_translate_addr(cma_src_addr(id_priv),
1868 &id->route.addr.dev_addr);
1874 id_priv->state = RDMA_CM_CONNECT;
1877 rdma_destroy_id(id);
1881 static void cma_set_req_event_data(struct rdma_cm_event *event,
1882 struct ib_cm_req_event_param *req_data,
1883 void *private_data, int offset)
1885 event->param.conn.private_data = private_data + offset;
1886 event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
1887 event->param.conn.responder_resources = req_data->responder_resources;
1888 event->param.conn.initiator_depth = req_data->initiator_depth;
1889 event->param.conn.flow_control = req_data->flow_control;
1890 event->param.conn.retry_count = req_data->retry_count;
1891 event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
1892 event->param.conn.srq = req_data->srq;
1893 event->param.conn.qp_num = req_data->remote_qpn;
1896 static int cma_check_req_qp_type(struct rdma_cm_id *id, struct ib_cm_event *ib_event)
1898 return (((ib_event->event == IB_CM_REQ_RECEIVED) &&
1899 (ib_event->param.req_rcvd.qp_type == id->qp_type)) ||
1900 ((ib_event->event == IB_CM_SIDR_REQ_RECEIVED) &&
1901 (id->qp_type == IB_QPT_UD)) ||
1905 static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1907 struct rdma_id_private *listen_id, *conn_id = NULL;
1908 struct rdma_cm_event event;
1909 struct net_device *net_dev;
1913 listen_id = cma_id_from_event(cm_id, ib_event, &net_dev);
1914 if (IS_ERR(listen_id))
1915 return PTR_ERR(listen_id);
1917 if (!cma_check_req_qp_type(&listen_id->id, ib_event)) {
1922 mutex_lock(&listen_id->handler_mutex);
1923 if (listen_id->state != RDMA_CM_LISTEN) {
1924 ret = -ECONNABORTED;
1928 memset(&event, 0, sizeof event);
1929 offset = cma_user_data_offset(listen_id);
1930 event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1931 if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) {
1932 conn_id = cma_new_udp_id(&listen_id->id, ib_event, net_dev);
1933 event.param.ud.private_data = ib_event->private_data + offset;
1934 event.param.ud.private_data_len =
1935 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
1937 conn_id = cma_new_conn_id(&listen_id->id, ib_event, net_dev);
1938 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
1939 ib_event->private_data, offset);
1946 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1947 ret = cma_acquire_dev(conn_id, listen_id);
1951 conn_id->cm_id.ib = cm_id;
1952 cm_id->context = conn_id;
1953 cm_id->cm_handler = cma_ib_handler;
1956 * Protect against the user destroying conn_id from another thread
1957 * until we're done accessing it.
1959 atomic_inc(&conn_id->refcount);
1960 ret = conn_id->id.event_handler(&conn_id->id, &event);
1964 * Acquire mutex to prevent user executing rdma_destroy_id()
1965 * while we're accessing the cm_id.
1968 if (cma_comp(conn_id, RDMA_CM_CONNECT) &&
1969 (conn_id->id.qp_type != IB_QPT_UD))
1970 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
1971 mutex_unlock(&lock);
1972 mutex_unlock(&conn_id->handler_mutex);
1973 mutex_unlock(&listen_id->handler_mutex);
1974 cma_deref_id(conn_id);
1980 cma_deref_id(conn_id);
1981 /* Destroy the CM ID by returning a non-zero value. */
1982 conn_id->cm_id.ib = NULL;
1984 cma_exch(conn_id, RDMA_CM_DESTROYING);
1985 mutex_unlock(&conn_id->handler_mutex);
1987 mutex_unlock(&listen_id->handler_mutex);
1989 rdma_destroy_id(&conn_id->id);
1998 __be64 rdma_get_service_id(struct rdma_cm_id *id, struct sockaddr *addr)
2000 if (addr->sa_family == AF_IB)
2001 return ((struct sockaddr_ib *) addr)->sib_sid;
2003 return cpu_to_be64(((u64)id->ps << 16) + be16_to_cpu(cma_port(addr)));
2005 EXPORT_SYMBOL(rdma_get_service_id);
2007 void rdma_read_gids(struct rdma_cm_id *cm_id, union ib_gid *sgid,
2010 struct rdma_addr *addr = &cm_id->route.addr;
2012 if (!cm_id->device) {
2014 memset(sgid, 0, sizeof(*sgid));
2016 memset(dgid, 0, sizeof(*dgid));
2020 if (rdma_protocol_roce(cm_id->device, cm_id->port_num)) {
2022 rdma_ip2gid((struct sockaddr *)&addr->src_addr, sgid);
2024 rdma_ip2gid((struct sockaddr *)&addr->dst_addr, dgid);
2027 rdma_addr_get_sgid(&addr->dev_addr, sgid);
2029 rdma_addr_get_dgid(&addr->dev_addr, dgid);
2032 EXPORT_SYMBOL(rdma_read_gids);
2034 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
2036 struct rdma_id_private *id_priv = iw_id->context;
2037 struct rdma_cm_event event;
2039 struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
2040 struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
2042 mutex_lock(&id_priv->handler_mutex);
2043 if (id_priv->state != RDMA_CM_CONNECT)
2046 memset(&event, 0, sizeof event);
2047 switch (iw_event->event) {
2048 case IW_CM_EVENT_CLOSE:
2049 event.event = RDMA_CM_EVENT_DISCONNECTED;
2051 case IW_CM_EVENT_CONNECT_REPLY:
2052 memcpy(cma_src_addr(id_priv), laddr,
2053 rdma_addr_size(laddr));
2054 memcpy(cma_dst_addr(id_priv), raddr,
2055 rdma_addr_size(raddr));
2056 switch (iw_event->status) {
2058 event.event = RDMA_CM_EVENT_ESTABLISHED;
2059 event.param.conn.initiator_depth = iw_event->ird;
2060 event.param.conn.responder_resources = iw_event->ord;
2064 event.event = RDMA_CM_EVENT_REJECTED;
2067 event.event = RDMA_CM_EVENT_UNREACHABLE;
2070 event.event = RDMA_CM_EVENT_CONNECT_ERROR;
2074 case IW_CM_EVENT_ESTABLISHED:
2075 event.event = RDMA_CM_EVENT_ESTABLISHED;
2076 event.param.conn.initiator_depth = iw_event->ird;
2077 event.param.conn.responder_resources = iw_event->ord;
2083 event.status = iw_event->status;
2084 event.param.conn.private_data = iw_event->private_data;
2085 event.param.conn.private_data_len = iw_event->private_data_len;
2086 ret = id_priv->id.event_handler(&id_priv->id, &event);
2088 /* Destroy the CM ID by returning a non-zero value. */
2089 id_priv->cm_id.iw = NULL;
2090 cma_exch(id_priv, RDMA_CM_DESTROYING);
2091 mutex_unlock(&id_priv->handler_mutex);
2092 rdma_destroy_id(&id_priv->id);
2097 mutex_unlock(&id_priv->handler_mutex);
2101 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
2102 struct iw_cm_event *iw_event)
2104 struct rdma_cm_id *new_cm_id;
2105 struct rdma_id_private *listen_id, *conn_id;
2106 struct rdma_cm_event event;
2107 int ret = -ECONNABORTED;
2108 struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
2109 struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
2111 listen_id = cm_id->context;
2113 mutex_lock(&listen_id->handler_mutex);
2114 if (listen_id->state != RDMA_CM_LISTEN)
2117 /* Create a new RDMA id for the new IW CM ID */
2118 new_cm_id = __rdma_create_id(listen_id->id.route.addr.dev_addr.net,
2119 listen_id->id.event_handler,
2120 listen_id->id.context,
2121 RDMA_PS_TCP, IB_QPT_RC,
2122 listen_id->res.kern_name);
2123 if (IS_ERR(new_cm_id)) {
2127 conn_id = container_of(new_cm_id, struct rdma_id_private, id);
2128 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
2129 conn_id->state = RDMA_CM_CONNECT;
2131 ret = rdma_translate_ip(laddr, &conn_id->id.route.addr.dev_addr);
2133 mutex_unlock(&conn_id->handler_mutex);
2134 rdma_destroy_id(new_cm_id);
2138 ret = cma_acquire_dev(conn_id, listen_id);
2140 mutex_unlock(&conn_id->handler_mutex);
2141 rdma_destroy_id(new_cm_id);
2145 conn_id->cm_id.iw = cm_id;
2146 cm_id->context = conn_id;
2147 cm_id->cm_handler = cma_iw_handler;
2149 memcpy(cma_src_addr(conn_id), laddr, rdma_addr_size(laddr));
2150 memcpy(cma_dst_addr(conn_id), raddr, rdma_addr_size(raddr));
2152 memset(&event, 0, sizeof event);
2153 event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
2154 event.param.conn.private_data = iw_event->private_data;
2155 event.param.conn.private_data_len = iw_event->private_data_len;
2156 event.param.conn.initiator_depth = iw_event->ird;
2157 event.param.conn.responder_resources = iw_event->ord;
2160 * Protect against the user destroying conn_id from another thread
2161 * until we're done accessing it.
2163 atomic_inc(&conn_id->refcount);
2164 ret = conn_id->id.event_handler(&conn_id->id, &event);
2166 /* User wants to destroy the CM ID */
2167 conn_id->cm_id.iw = NULL;
2168 cma_exch(conn_id, RDMA_CM_DESTROYING);
2169 mutex_unlock(&conn_id->handler_mutex);
2170 cma_deref_id(conn_id);
2171 rdma_destroy_id(&conn_id->id);
2175 mutex_unlock(&conn_id->handler_mutex);
2176 cma_deref_id(conn_id);
2179 mutex_unlock(&listen_id->handler_mutex);
2183 static int cma_ib_listen(struct rdma_id_private *id_priv)
2185 struct sockaddr *addr;
2186 struct ib_cm_id *id;
2189 addr = cma_src_addr(id_priv);
2190 svc_id = rdma_get_service_id(&id_priv->id, addr);
2191 id = ib_cm_insert_listen(id_priv->id.device, cma_req_handler, svc_id);
2194 id_priv->cm_id.ib = id;
2199 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
2202 struct iw_cm_id *id;
2204 id = iw_create_cm_id(id_priv->id.device,
2205 iw_conn_req_handler,
2210 id->tos = id_priv->tos;
2211 id_priv->cm_id.iw = id;
2213 memcpy(&id_priv->cm_id.iw->local_addr, cma_src_addr(id_priv),
2214 rdma_addr_size(cma_src_addr(id_priv)));
2216 ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
2219 iw_destroy_cm_id(id_priv->cm_id.iw);
2220 id_priv->cm_id.iw = NULL;
2226 static int cma_listen_handler(struct rdma_cm_id *id,
2227 struct rdma_cm_event *event)
2229 struct rdma_id_private *id_priv = id->context;
2231 id->context = id_priv->id.context;
2232 id->event_handler = id_priv->id.event_handler;
2233 return id_priv->id.event_handler(id, event);
2236 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
2237 struct cma_device *cma_dev)
2239 struct rdma_id_private *dev_id_priv;
2240 struct rdma_cm_id *id;
2241 struct net *net = id_priv->id.route.addr.dev_addr.net;
2244 if (cma_family(id_priv) == AF_IB && !rdma_cap_ib_cm(cma_dev->device, 1))
2247 id = __rdma_create_id(net, cma_listen_handler, id_priv, id_priv->id.ps,
2248 id_priv->id.qp_type, id_priv->res.kern_name);
2252 dev_id_priv = container_of(id, struct rdma_id_private, id);
2254 dev_id_priv->state = RDMA_CM_ADDR_BOUND;
2255 memcpy(cma_src_addr(dev_id_priv), cma_src_addr(id_priv),
2256 rdma_addr_size(cma_src_addr(id_priv)));
2258 _cma_attach_to_dev(dev_id_priv, cma_dev);
2259 list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
2260 atomic_inc(&id_priv->refcount);
2261 dev_id_priv->internal_id = 1;
2262 dev_id_priv->afonly = id_priv->afonly;
2264 ret = rdma_listen(id, id_priv->backlog);
2266 pr_warn("RDMA CMA: cma_listen_on_dev, error %d, listening on device %s\n",
2267 ret, cma_dev->device->name);
2270 static void cma_listen_on_all(struct rdma_id_private *id_priv)
2272 struct cma_device *cma_dev;
2275 list_add_tail(&id_priv->list, &listen_any_list);
2276 list_for_each_entry(cma_dev, &dev_list, list)
2277 cma_listen_on_dev(id_priv, cma_dev);
2278 mutex_unlock(&lock);
2281 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
2283 struct rdma_id_private *id_priv;
2285 id_priv = container_of(id, struct rdma_id_private, id);
2286 id_priv->tos = (u8) tos;
2287 id_priv->tos_set = true;
2289 EXPORT_SYMBOL(rdma_set_service_type);
2291 static void cma_query_handler(int status, struct sa_path_rec *path_rec,
2294 struct cma_work *work = context;
2295 struct rdma_route *route;
2297 route = &work->id->id.route;
2300 route->num_paths = 1;
2301 *route->path_rec = *path_rec;
2303 work->old_state = RDMA_CM_ROUTE_QUERY;
2304 work->new_state = RDMA_CM_ADDR_RESOLVED;
2305 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
2306 work->event.status = status;
2307 pr_debug_ratelimited("RDMA CM: ROUTE_ERROR: failed to query path. status %d\n",
2311 queue_work(cma_wq, &work->work);
2314 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms,
2315 struct cma_work *work)
2317 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2318 struct sa_path_rec path_rec;
2319 ib_sa_comp_mask comp_mask;
2320 struct sockaddr_in6 *sin6;
2321 struct sockaddr_ib *sib;
2323 memset(&path_rec, 0, sizeof path_rec);
2325 if (rdma_cap_opa_ah(id_priv->id.device, id_priv->id.port_num))
2326 path_rec.rec_type = SA_PATH_REC_TYPE_OPA;
2328 path_rec.rec_type = SA_PATH_REC_TYPE_IB;
2329 rdma_addr_get_sgid(dev_addr, &path_rec.sgid);
2330 rdma_addr_get_dgid(dev_addr, &path_rec.dgid);
2331 path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
2332 path_rec.numb_path = 1;
2333 path_rec.reversible = 1;
2334 path_rec.service_id = rdma_get_service_id(&id_priv->id,
2335 cma_dst_addr(id_priv));
2337 comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
2338 IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
2339 IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
2341 switch (cma_family(id_priv)) {
2343 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
2344 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
2347 sin6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
2348 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
2349 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2352 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
2353 path_rec.traffic_class = (u8) (be32_to_cpu(sib->sib_flowinfo) >> 20);
2354 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2358 id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
2359 id_priv->id.port_num, &path_rec,
2360 comp_mask, timeout_ms,
2361 GFP_KERNEL, cma_query_handler,
2362 work, &id_priv->query);
2364 return (id_priv->query_id < 0) ? id_priv->query_id : 0;
2367 static void cma_work_handler(struct work_struct *_work)
2369 struct cma_work *work = container_of(_work, struct cma_work, work);
2370 struct rdma_id_private *id_priv = work->id;
2373 mutex_lock(&id_priv->handler_mutex);
2374 if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
2377 if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
2378 cma_exch(id_priv, RDMA_CM_DESTROYING);
2382 mutex_unlock(&id_priv->handler_mutex);
2383 cma_deref_id(id_priv);
2385 rdma_destroy_id(&id_priv->id);
2389 static void cma_ndev_work_handler(struct work_struct *_work)
2391 struct cma_ndev_work *work = container_of(_work, struct cma_ndev_work, work);
2392 struct rdma_id_private *id_priv = work->id;
2395 mutex_lock(&id_priv->handler_mutex);
2396 if (id_priv->state == RDMA_CM_DESTROYING ||
2397 id_priv->state == RDMA_CM_DEVICE_REMOVAL)
2400 if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
2401 cma_exch(id_priv, RDMA_CM_DESTROYING);
2406 mutex_unlock(&id_priv->handler_mutex);
2407 cma_deref_id(id_priv);
2409 rdma_destroy_id(&id_priv->id);
2413 static void cma_init_resolve_route_work(struct cma_work *work,
2414 struct rdma_id_private *id_priv)
2417 INIT_WORK(&work->work, cma_work_handler);
2418 work->old_state = RDMA_CM_ROUTE_QUERY;
2419 work->new_state = RDMA_CM_ROUTE_RESOLVED;
2420 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
2423 static void cma_init_resolve_addr_work(struct cma_work *work,
2424 struct rdma_id_private *id_priv)
2427 INIT_WORK(&work->work, cma_work_handler);
2428 work->old_state = RDMA_CM_ADDR_QUERY;
2429 work->new_state = RDMA_CM_ADDR_RESOLVED;
2430 work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2433 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms)
2435 struct rdma_route *route = &id_priv->id.route;
2436 struct cma_work *work;
2439 work = kzalloc(sizeof *work, GFP_KERNEL);
2443 cma_init_resolve_route_work(work, id_priv);
2445 route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
2446 if (!route->path_rec) {
2451 ret = cma_query_ib_route(id_priv, timeout_ms, work);
2457 kfree(route->path_rec);
2458 route->path_rec = NULL;
2464 static enum ib_gid_type cma_route_gid_type(enum rdma_network_type network_type,
2465 unsigned long supported_gids,
2466 enum ib_gid_type default_gid)
2468 if ((network_type == RDMA_NETWORK_IPV4 ||
2469 network_type == RDMA_NETWORK_IPV6) &&
2470 test_bit(IB_GID_TYPE_ROCE_UDP_ENCAP, &supported_gids))
2471 return IB_GID_TYPE_ROCE_UDP_ENCAP;
2477 * cma_iboe_set_path_rec_l2_fields() is helper function which sets
2478 * path record type based on GID type.
2479 * It also sets up other L2 fields which includes destination mac address
2480 * netdev ifindex, of the path record.
2481 * It returns the netdev of the bound interface for this path record entry.
2483 static struct net_device *
2484 cma_iboe_set_path_rec_l2_fields(struct rdma_id_private *id_priv)
2486 struct rdma_route *route = &id_priv->id.route;
2487 enum ib_gid_type gid_type = IB_GID_TYPE_ROCE;
2488 struct rdma_addr *addr = &route->addr;
2489 unsigned long supported_gids;
2490 struct net_device *ndev;
2492 if (!addr->dev_addr.bound_dev_if)
2495 ndev = dev_get_by_index(addr->dev_addr.net,
2496 addr->dev_addr.bound_dev_if);
2500 supported_gids = roce_gid_type_mask_support(id_priv->id.device,
2501 id_priv->id.port_num);
2502 gid_type = cma_route_gid_type(addr->dev_addr.network,
2505 /* Use the hint from IP Stack to select GID Type */
2506 if (gid_type < ib_network_to_gid_type(addr->dev_addr.network))
2507 gid_type = ib_network_to_gid_type(addr->dev_addr.network);
2508 route->path_rec->rec_type = sa_conv_gid_to_pathrec_type(gid_type);
2510 route->path_rec->roce.route_resolved = true;
2511 sa_path_set_ndev(route->path_rec, addr->dev_addr.net);
2512 sa_path_set_ifindex(route->path_rec, ndev->ifindex);
2513 sa_path_set_dmac(route->path_rec, addr->dev_addr.dst_dev_addr);
2517 int rdma_set_ib_path(struct rdma_cm_id *id,
2518 struct sa_path_rec *path_rec)
2520 struct rdma_id_private *id_priv;
2521 struct net_device *ndev;
2524 id_priv = container_of(id, struct rdma_id_private, id);
2525 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
2526 RDMA_CM_ROUTE_RESOLVED))
2529 id->route.path_rec = kmemdup(path_rec, sizeof(*path_rec),
2531 if (!id->route.path_rec) {
2536 if (rdma_protocol_roce(id->device, id->port_num)) {
2537 ndev = cma_iboe_set_path_rec_l2_fields(id_priv);
2545 id->route.num_paths = 1;
2549 kfree(id->route.path_rec);
2550 id->route.path_rec = NULL;
2552 cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED);
2555 EXPORT_SYMBOL(rdma_set_ib_path);
2557 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms)
2559 struct cma_work *work;
2561 work = kzalloc(sizeof *work, GFP_KERNEL);
2565 cma_init_resolve_route_work(work, id_priv);
2566 queue_work(cma_wq, &work->work);
2570 static int iboe_tos_to_sl(struct net_device *ndev, int tos)
2573 struct net_device *dev;
2575 prio = rt_tos2priority(tos);
2576 dev = is_vlan_dev(ndev) ? vlan_dev_real_dev(ndev) : ndev;
2578 return netdev_get_prio_tc_map(dev, prio);
2580 #if IS_ENABLED(CONFIG_VLAN_8021Q)
2581 if (is_vlan_dev(ndev))
2582 return (vlan_dev_get_egress_qos_mask(ndev, prio) &
2583 VLAN_PRIO_MASK) >> VLAN_PRIO_SHIFT;
2588 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
2590 struct rdma_route *route = &id_priv->id.route;
2591 struct rdma_addr *addr = &route->addr;
2592 struct cma_work *work;
2594 struct net_device *ndev;
2596 u8 default_roce_tos = id_priv->cma_dev->default_roce_tos[id_priv->id.port_num -
2597 rdma_start_port(id_priv->cma_dev->device)];
2598 u8 tos = id_priv->tos_set ? id_priv->tos : default_roce_tos;
2601 work = kzalloc(sizeof *work, GFP_KERNEL);
2605 route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
2606 if (!route->path_rec) {
2611 route->num_paths = 1;
2613 ndev = cma_iboe_set_path_rec_l2_fields(id_priv);
2619 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
2620 &route->path_rec->sgid);
2621 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.dst_addr,
2622 &route->path_rec->dgid);
2624 if (((struct sockaddr *)&id_priv->id.route.addr.dst_addr)->sa_family != AF_IB)
2625 /* TODO: get the hoplimit from the inet/inet6 device */
2626 route->path_rec->hop_limit = addr->dev_addr.hoplimit;
2628 route->path_rec->hop_limit = 1;
2629 route->path_rec->reversible = 1;
2630 route->path_rec->pkey = cpu_to_be16(0xffff);
2631 route->path_rec->mtu_selector = IB_SA_EQ;
2632 route->path_rec->sl = iboe_tos_to_sl(ndev, tos);
2633 route->path_rec->traffic_class = tos;
2634 route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
2635 route->path_rec->rate_selector = IB_SA_EQ;
2636 route->path_rec->rate = iboe_get_rate(ndev);
2638 route->path_rec->packet_life_time_selector = IB_SA_EQ;
2639 route->path_rec->packet_life_time = CMA_IBOE_PACKET_LIFETIME;
2640 if (!route->path_rec->mtu) {
2645 cma_init_resolve_route_work(work, id_priv);
2646 queue_work(cma_wq, &work->work);
2651 kfree(route->path_rec);
2652 route->path_rec = NULL;
2658 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
2660 struct rdma_id_private *id_priv;
2663 id_priv = container_of(id, struct rdma_id_private, id);
2664 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY))
2667 atomic_inc(&id_priv->refcount);
2668 if (rdma_cap_ib_sa(id->device, id->port_num))
2669 ret = cma_resolve_ib_route(id_priv, timeout_ms);
2670 else if (rdma_protocol_roce(id->device, id->port_num))
2671 ret = cma_resolve_iboe_route(id_priv);
2672 else if (rdma_protocol_iwarp(id->device, id->port_num))
2673 ret = cma_resolve_iw_route(id_priv, timeout_ms);
2682 cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED);
2683 cma_deref_id(id_priv);
2686 EXPORT_SYMBOL(rdma_resolve_route);
2688 static void cma_set_loopback(struct sockaddr *addr)
2690 switch (addr->sa_family) {
2692 ((struct sockaddr_in *) addr)->sin_addr.s_addr = htonl(INADDR_LOOPBACK);
2695 ipv6_addr_set(&((struct sockaddr_in6 *) addr)->sin6_addr,
2699 ib_addr_set(&((struct sockaddr_ib *) addr)->sib_addr,
2705 static int cma_bind_loopback(struct rdma_id_private *id_priv)
2707 struct cma_device *cma_dev, *cur_dev;
2709 enum ib_port_state port_state;
2716 list_for_each_entry(cur_dev, &dev_list, list) {
2717 if (cma_family(id_priv) == AF_IB &&
2718 !rdma_cap_ib_cm(cur_dev->device, 1))
2724 for (p = 1; p <= cur_dev->device->phys_port_cnt; ++p) {
2725 if (!ib_get_cached_port_state(cur_dev->device, p, &port_state) &&
2726 port_state == IB_PORT_ACTIVE) {
2741 ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid, NULL);
2745 ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
2749 id_priv->id.route.addr.dev_addr.dev_type =
2750 (rdma_protocol_ib(cma_dev->device, p)) ?
2751 ARPHRD_INFINIBAND : ARPHRD_ETHER;
2753 rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2754 ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
2755 id_priv->id.port_num = p;
2756 cma_attach_to_dev(id_priv, cma_dev);
2757 cma_set_loopback(cma_src_addr(id_priv));
2759 mutex_unlock(&lock);
2763 static void addr_handler(int status, struct sockaddr *src_addr,
2764 struct rdma_dev_addr *dev_addr, void *context)
2766 struct rdma_id_private *id_priv = context;
2767 struct rdma_cm_event event;
2769 memset(&event, 0, sizeof event);
2770 mutex_lock(&id_priv->handler_mutex);
2771 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY,
2772 RDMA_CM_ADDR_RESOLVED))
2775 memcpy(cma_src_addr(id_priv), src_addr, rdma_addr_size(src_addr));
2776 if (!status && !id_priv->cma_dev) {
2777 status = cma_acquire_dev(id_priv, NULL);
2779 pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to acquire device. status %d\n",
2782 pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to resolve IP. status %d\n", status);
2786 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
2787 RDMA_CM_ADDR_BOUND))
2789 event.event = RDMA_CM_EVENT_ADDR_ERROR;
2790 event.status = status;
2792 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2794 if (id_priv->id.event_handler(&id_priv->id, &event)) {
2795 cma_exch(id_priv, RDMA_CM_DESTROYING);
2796 mutex_unlock(&id_priv->handler_mutex);
2797 cma_deref_id(id_priv);
2798 rdma_destroy_id(&id_priv->id);
2802 mutex_unlock(&id_priv->handler_mutex);
2803 cma_deref_id(id_priv);
2806 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
2808 struct cma_work *work;
2812 work = kzalloc(sizeof *work, GFP_KERNEL);
2816 if (!id_priv->cma_dev) {
2817 ret = cma_bind_loopback(id_priv);
2822 rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2823 rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
2825 cma_init_resolve_addr_work(work, id_priv);
2826 queue_work(cma_wq, &work->work);
2833 static int cma_resolve_ib_addr(struct rdma_id_private *id_priv)
2835 struct cma_work *work;
2838 work = kzalloc(sizeof *work, GFP_KERNEL);
2842 if (!id_priv->cma_dev) {
2843 ret = cma_resolve_ib_dev(id_priv);
2848 rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, (union ib_gid *)
2849 &(((struct sockaddr_ib *) &id_priv->id.route.addr.dst_addr)->sib_addr));
2851 cma_init_resolve_addr_work(work, id_priv);
2852 queue_work(cma_wq, &work->work);
2859 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2860 struct sockaddr *dst_addr)
2862 if (!src_addr || !src_addr->sa_family) {
2863 src_addr = (struct sockaddr *) &id->route.addr.src_addr;
2864 src_addr->sa_family = dst_addr->sa_family;
2865 if (IS_ENABLED(CONFIG_IPV6) &&
2866 dst_addr->sa_family == AF_INET6) {
2867 struct sockaddr_in6 *src_addr6 = (struct sockaddr_in6 *) src_addr;
2868 struct sockaddr_in6 *dst_addr6 = (struct sockaddr_in6 *) dst_addr;
2869 src_addr6->sin6_scope_id = dst_addr6->sin6_scope_id;
2870 if (ipv6_addr_type(&dst_addr6->sin6_addr) & IPV6_ADDR_LINKLOCAL)
2871 id->route.addr.dev_addr.bound_dev_if = dst_addr6->sin6_scope_id;
2872 } else if (dst_addr->sa_family == AF_IB) {
2873 ((struct sockaddr_ib *) src_addr)->sib_pkey =
2874 ((struct sockaddr_ib *) dst_addr)->sib_pkey;
2877 return rdma_bind_addr(id, src_addr);
2880 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2881 struct sockaddr *dst_addr, int timeout_ms)
2883 struct rdma_id_private *id_priv;
2886 id_priv = container_of(id, struct rdma_id_private, id);
2887 memcpy(cma_dst_addr(id_priv), dst_addr, rdma_addr_size(dst_addr));
2888 if (id_priv->state == RDMA_CM_IDLE) {
2889 ret = cma_bind_addr(id, src_addr, dst_addr);
2891 memset(cma_dst_addr(id_priv), 0, rdma_addr_size(dst_addr));
2896 if (cma_family(id_priv) != dst_addr->sa_family) {
2897 memset(cma_dst_addr(id_priv), 0, rdma_addr_size(dst_addr));
2901 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY)) {
2902 memset(cma_dst_addr(id_priv), 0, rdma_addr_size(dst_addr));
2906 atomic_inc(&id_priv->refcount);
2907 if (cma_any_addr(dst_addr)) {
2908 ret = cma_resolve_loopback(id_priv);
2910 if (dst_addr->sa_family == AF_IB) {
2911 ret = cma_resolve_ib_addr(id_priv);
2913 ret = rdma_resolve_ip(&addr_client, cma_src_addr(id_priv),
2914 dst_addr, &id->route.addr.dev_addr,
2915 timeout_ms, addr_handler, id_priv);
2923 cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
2924 cma_deref_id(id_priv);
2927 EXPORT_SYMBOL(rdma_resolve_addr);
2929 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse)
2931 struct rdma_id_private *id_priv;
2932 unsigned long flags;
2935 id_priv = container_of(id, struct rdma_id_private, id);
2936 spin_lock_irqsave(&id_priv->lock, flags);
2937 if (reuse || id_priv->state == RDMA_CM_IDLE) {
2938 id_priv->reuseaddr = reuse;
2943 spin_unlock_irqrestore(&id_priv->lock, flags);
2946 EXPORT_SYMBOL(rdma_set_reuseaddr);
2948 int rdma_set_afonly(struct rdma_cm_id *id, int afonly)
2950 struct rdma_id_private *id_priv;
2951 unsigned long flags;
2954 id_priv = container_of(id, struct rdma_id_private, id);
2955 spin_lock_irqsave(&id_priv->lock, flags);
2956 if (id_priv->state == RDMA_CM_IDLE || id_priv->state == RDMA_CM_ADDR_BOUND) {
2957 id_priv->options |= (1 << CMA_OPTION_AFONLY);
2958 id_priv->afonly = afonly;
2963 spin_unlock_irqrestore(&id_priv->lock, flags);
2966 EXPORT_SYMBOL(rdma_set_afonly);
2968 static void cma_bind_port(struct rdma_bind_list *bind_list,
2969 struct rdma_id_private *id_priv)
2971 struct sockaddr *addr;
2972 struct sockaddr_ib *sib;
2976 addr = cma_src_addr(id_priv);
2977 port = htons(bind_list->port);
2979 switch (addr->sa_family) {
2981 ((struct sockaddr_in *) addr)->sin_port = port;
2984 ((struct sockaddr_in6 *) addr)->sin6_port = port;
2987 sib = (struct sockaddr_ib *) addr;
2988 sid = be64_to_cpu(sib->sib_sid);
2989 mask = be64_to_cpu(sib->sib_sid_mask);
2990 sib->sib_sid = cpu_to_be64((sid & mask) | (u64) ntohs(port));
2991 sib->sib_sid_mask = cpu_to_be64(~0ULL);
2994 id_priv->bind_list = bind_list;
2995 hlist_add_head(&id_priv->node, &bind_list->owners);
2998 static int cma_alloc_port(enum rdma_ucm_port_space ps,
2999 struct rdma_id_private *id_priv, unsigned short snum)
3001 struct rdma_bind_list *bind_list;
3004 bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
3008 ret = cma_ps_alloc(id_priv->id.route.addr.dev_addr.net, ps, bind_list,
3014 bind_list->port = (unsigned short)ret;
3015 cma_bind_port(bind_list, id_priv);
3019 return ret == -ENOSPC ? -EADDRNOTAVAIL : ret;
3022 static int cma_port_is_unique(struct rdma_bind_list *bind_list,
3023 struct rdma_id_private *id_priv)
3025 struct rdma_id_private *cur_id;
3026 struct sockaddr *daddr = cma_dst_addr(id_priv);
3027 struct sockaddr *saddr = cma_src_addr(id_priv);
3028 __be16 dport = cma_port(daddr);
3030 hlist_for_each_entry(cur_id, &bind_list->owners, node) {
3031 struct sockaddr *cur_daddr = cma_dst_addr(cur_id);
3032 struct sockaddr *cur_saddr = cma_src_addr(cur_id);
3033 __be16 cur_dport = cma_port(cur_daddr);
3035 if (id_priv == cur_id)
3038 /* different dest port -> unique */
3039 if (!cma_any_port(daddr) &&
3040 !cma_any_port(cur_daddr) &&
3041 (dport != cur_dport))
3044 /* different src address -> unique */
3045 if (!cma_any_addr(saddr) &&
3046 !cma_any_addr(cur_saddr) &&
3047 cma_addr_cmp(saddr, cur_saddr))
3050 /* different dst address -> unique */
3051 if (!cma_any_addr(daddr) &&
3052 !cma_any_addr(cur_daddr) &&
3053 cma_addr_cmp(daddr, cur_daddr))
3056 return -EADDRNOTAVAIL;
3061 static int cma_alloc_any_port(enum rdma_ucm_port_space ps,
3062 struct rdma_id_private *id_priv)
3064 static unsigned int last_used_port;
3065 int low, high, remaining;
3067 struct net *net = id_priv->id.route.addr.dev_addr.net;
3069 inet_get_local_port_range(net, &low, &high);
3070 remaining = (high - low) + 1;
3071 rover = prandom_u32() % remaining + low;
3073 if (last_used_port != rover) {
3074 struct rdma_bind_list *bind_list;
3077 bind_list = cma_ps_find(net, ps, (unsigned short)rover);
3080 ret = cma_alloc_port(ps, id_priv, rover);
3082 ret = cma_port_is_unique(bind_list, id_priv);
3084 cma_bind_port(bind_list, id_priv);
3087 * Remember previously used port number in order to avoid
3088 * re-using same port immediately after it is closed.
3091 last_used_port = rover;
3092 if (ret != -EADDRNOTAVAIL)
3097 if ((rover < low) || (rover > high))
3101 return -EADDRNOTAVAIL;
3105 * Check that the requested port is available. This is called when trying to
3106 * bind to a specific port, or when trying to listen on a bound port. In
3107 * the latter case, the provided id_priv may already be on the bind_list, but
3108 * we still need to check that it's okay to start listening.
3110 static int cma_check_port(struct rdma_bind_list *bind_list,
3111 struct rdma_id_private *id_priv, uint8_t reuseaddr)
3113 struct rdma_id_private *cur_id;
3114 struct sockaddr *addr, *cur_addr;
3116 addr = cma_src_addr(id_priv);
3117 hlist_for_each_entry(cur_id, &bind_list->owners, node) {
3118 if (id_priv == cur_id)
3121 if ((cur_id->state != RDMA_CM_LISTEN) && reuseaddr &&
3125 cur_addr = cma_src_addr(cur_id);
3126 if (id_priv->afonly && cur_id->afonly &&
3127 (addr->sa_family != cur_addr->sa_family))
3130 if (cma_any_addr(addr) || cma_any_addr(cur_addr))
3131 return -EADDRNOTAVAIL;
3133 if (!cma_addr_cmp(addr, cur_addr))
3139 static int cma_use_port(enum rdma_ucm_port_space ps,
3140 struct rdma_id_private *id_priv)
3142 struct rdma_bind_list *bind_list;
3143 unsigned short snum;
3146 snum = ntohs(cma_port(cma_src_addr(id_priv)));
3147 if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
3150 bind_list = cma_ps_find(id_priv->id.route.addr.dev_addr.net, ps, snum);
3152 ret = cma_alloc_port(ps, id_priv, snum);
3154 ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr);
3156 cma_bind_port(bind_list, id_priv);
3161 static int cma_bind_listen(struct rdma_id_private *id_priv)
3163 struct rdma_bind_list *bind_list = id_priv->bind_list;
3167 if (bind_list->owners.first->next)
3168 ret = cma_check_port(bind_list, id_priv, 0);
3169 mutex_unlock(&lock);
3173 static enum rdma_ucm_port_space
3174 cma_select_inet_ps(struct rdma_id_private *id_priv)
3176 switch (id_priv->id.ps) {
3181 return id_priv->id.ps;
3188 static enum rdma_ucm_port_space
3189 cma_select_ib_ps(struct rdma_id_private *id_priv)
3191 enum rdma_ucm_port_space ps = 0;
3192 struct sockaddr_ib *sib;
3193 u64 sid_ps, mask, sid;
3195 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
3196 mask = be64_to_cpu(sib->sib_sid_mask) & RDMA_IB_IP_PS_MASK;
3197 sid = be64_to_cpu(sib->sib_sid) & mask;
3199 if ((id_priv->id.ps == RDMA_PS_IB) && (sid == (RDMA_IB_IP_PS_IB & mask))) {
3200 sid_ps = RDMA_IB_IP_PS_IB;
3202 } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_TCP)) &&
3203 (sid == (RDMA_IB_IP_PS_TCP & mask))) {
3204 sid_ps = RDMA_IB_IP_PS_TCP;
3206 } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_UDP)) &&
3207 (sid == (RDMA_IB_IP_PS_UDP & mask))) {
3208 sid_ps = RDMA_IB_IP_PS_UDP;
3213 sib->sib_sid = cpu_to_be64(sid_ps | ntohs(cma_port((struct sockaddr *) sib)));
3214 sib->sib_sid_mask = cpu_to_be64(RDMA_IB_IP_PS_MASK |
3215 be64_to_cpu(sib->sib_sid_mask));
3220 static int cma_get_port(struct rdma_id_private *id_priv)
3222 enum rdma_ucm_port_space ps;
3225 if (cma_family(id_priv) != AF_IB)
3226 ps = cma_select_inet_ps(id_priv);
3228 ps = cma_select_ib_ps(id_priv);
3230 return -EPROTONOSUPPORT;
3233 if (cma_any_port(cma_src_addr(id_priv)))
3234 ret = cma_alloc_any_port(ps, id_priv);
3236 ret = cma_use_port(ps, id_priv);
3237 mutex_unlock(&lock);
3242 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
3243 struct sockaddr *addr)
3245 #if IS_ENABLED(CONFIG_IPV6)
3246 struct sockaddr_in6 *sin6;
3248 if (addr->sa_family != AF_INET6)
3251 sin6 = (struct sockaddr_in6 *) addr;
3253 if (!(ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL))
3256 if (!sin6->sin6_scope_id)
3259 dev_addr->bound_dev_if = sin6->sin6_scope_id;
3264 int rdma_listen(struct rdma_cm_id *id, int backlog)
3266 struct rdma_id_private *id_priv;
3269 id_priv = container_of(id, struct rdma_id_private, id);
3270 if (id_priv->state == RDMA_CM_IDLE) {
3271 id->route.addr.src_addr.ss_family = AF_INET;
3272 ret = rdma_bind_addr(id, cma_src_addr(id_priv));
3277 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN))
3280 if (id_priv->reuseaddr) {
3281 ret = cma_bind_listen(id_priv);
3286 id_priv->backlog = backlog;
3288 if (rdma_cap_ib_cm(id->device, 1)) {
3289 ret = cma_ib_listen(id_priv);
3292 } else if (rdma_cap_iw_cm(id->device, 1)) {
3293 ret = cma_iw_listen(id_priv, backlog);
3301 cma_listen_on_all(id_priv);
3305 id_priv->backlog = 0;
3306 cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND);
3309 EXPORT_SYMBOL(rdma_listen);
3311 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
3313 struct rdma_id_private *id_priv;
3315 struct sockaddr *daddr;
3317 if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6 &&
3318 addr->sa_family != AF_IB)
3319 return -EAFNOSUPPORT;
3321 id_priv = container_of(id, struct rdma_id_private, id);
3322 if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND))
3325 ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
3329 memcpy(cma_src_addr(id_priv), addr, rdma_addr_size(addr));
3330 if (!cma_any_addr(addr)) {
3331 ret = cma_translate_addr(addr, &id->route.addr.dev_addr);
3335 ret = cma_acquire_dev(id_priv, NULL);
3340 if (!(id_priv->options & (1 << CMA_OPTION_AFONLY))) {
3341 if (addr->sa_family == AF_INET)
3342 id_priv->afonly = 1;
3343 #if IS_ENABLED(CONFIG_IPV6)
3344 else if (addr->sa_family == AF_INET6) {
3345 struct net *net = id_priv->id.route.addr.dev_addr.net;
3347 id_priv->afonly = net->ipv6.sysctl.bindv6only;
3351 daddr = cma_dst_addr(id_priv);
3352 daddr->sa_family = addr->sa_family;
3354 ret = cma_get_port(id_priv);
3360 if (id_priv->cma_dev) {
3361 rdma_restrack_del(&id_priv->res);
3362 cma_release_dev(id_priv);
3365 cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE);
3368 EXPORT_SYMBOL(rdma_bind_addr);
3370 static int cma_format_hdr(void *hdr, struct rdma_id_private *id_priv)
3372 struct cma_hdr *cma_hdr;
3375 cma_hdr->cma_version = CMA_VERSION;
3376 if (cma_family(id_priv) == AF_INET) {
3377 struct sockaddr_in *src4, *dst4;
3379 src4 = (struct sockaddr_in *) cma_src_addr(id_priv);
3380 dst4 = (struct sockaddr_in *) cma_dst_addr(id_priv);
3382 cma_set_ip_ver(cma_hdr, 4);
3383 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
3384 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
3385 cma_hdr->port = src4->sin_port;
3386 } else if (cma_family(id_priv) == AF_INET6) {
3387 struct sockaddr_in6 *src6, *dst6;
3389 src6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
3390 dst6 = (struct sockaddr_in6 *) cma_dst_addr(id_priv);
3392 cma_set_ip_ver(cma_hdr, 6);
3393 cma_hdr->src_addr.ip6 = src6->sin6_addr;
3394 cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
3395 cma_hdr->port = src6->sin6_port;
3400 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
3401 struct ib_cm_event *ib_event)
3403 struct rdma_id_private *id_priv = cm_id->context;
3404 struct rdma_cm_event event;
3405 struct ib_cm_sidr_rep_event_param *rep = &ib_event->param.sidr_rep_rcvd;
3408 mutex_lock(&id_priv->handler_mutex);
3409 if (id_priv->state != RDMA_CM_CONNECT)
3412 memset(&event, 0, sizeof event);
3413 switch (ib_event->event) {
3414 case IB_CM_SIDR_REQ_ERROR:
3415 event.event = RDMA_CM_EVENT_UNREACHABLE;
3416 event.status = -ETIMEDOUT;
3418 case IB_CM_SIDR_REP_RECEIVED:
3419 event.param.ud.private_data = ib_event->private_data;
3420 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
3421 if (rep->status != IB_SIDR_SUCCESS) {
3422 event.event = RDMA_CM_EVENT_UNREACHABLE;
3423 event.status = ib_event->param.sidr_rep_rcvd.status;
3424 pr_debug_ratelimited("RDMA CM: UNREACHABLE: bad SIDR reply. status %d\n",
3428 ret = cma_set_qkey(id_priv, rep->qkey);
3430 pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to set qkey. status %d\n", ret);
3431 event.event = RDMA_CM_EVENT_ADDR_ERROR;
3435 ib_init_ah_attr_from_path(id_priv->id.device,
3436 id_priv->id.port_num,
3437 id_priv->id.route.path_rec,
3438 &event.param.ud.ah_attr);
3439 event.param.ud.qp_num = rep->qpn;
3440 event.param.ud.qkey = rep->qkey;
3441 event.event = RDMA_CM_EVENT_ESTABLISHED;
3445 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
3450 ret = id_priv->id.event_handler(&id_priv->id, &event);
3452 /* Destroy the CM ID by returning a non-zero value. */
3453 id_priv->cm_id.ib = NULL;
3454 cma_exch(id_priv, RDMA_CM_DESTROYING);
3455 mutex_unlock(&id_priv->handler_mutex);
3456 rdma_destroy_id(&id_priv->id);
3460 mutex_unlock(&id_priv->handler_mutex);
3464 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
3465 struct rdma_conn_param *conn_param)
3467 struct ib_cm_sidr_req_param req;
3468 struct ib_cm_id *id;
3473 memset(&req, 0, sizeof req);
3474 offset = cma_user_data_offset(id_priv);
3475 req.private_data_len = offset + conn_param->private_data_len;
3476 if (req.private_data_len < conn_param->private_data_len)
3479 if (req.private_data_len) {
3480 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
3484 private_data = NULL;
3487 if (conn_param->private_data && conn_param->private_data_len)
3488 memcpy(private_data + offset, conn_param->private_data,
3489 conn_param->private_data_len);
3492 ret = cma_format_hdr(private_data, id_priv);
3495 req.private_data = private_data;
3498 id = ib_create_cm_id(id_priv->id.device, cma_sidr_rep_handler,
3504 id_priv->cm_id.ib = id;
3506 req.path = id_priv->id.route.path_rec;
3507 req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
3508 req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
3509 req.max_cm_retries = CMA_MAX_CM_RETRIES;
3511 ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
3513 ib_destroy_cm_id(id_priv->cm_id.ib);
3514 id_priv->cm_id.ib = NULL;
3517 kfree(private_data);
3521 static int cma_connect_ib(struct rdma_id_private *id_priv,
3522 struct rdma_conn_param *conn_param)
3524 struct ib_cm_req_param req;
3525 struct rdma_route *route;
3527 struct ib_cm_id *id;
3531 memset(&req, 0, sizeof req);
3532 offset = cma_user_data_offset(id_priv);
3533 req.private_data_len = offset + conn_param->private_data_len;
3534 if (req.private_data_len < conn_param->private_data_len)
3537 if (req.private_data_len) {
3538 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
3542 private_data = NULL;
3545 if (conn_param->private_data && conn_param->private_data_len)
3546 memcpy(private_data + offset, conn_param->private_data,
3547 conn_param->private_data_len);
3549 id = ib_create_cm_id(id_priv->id.device, cma_ib_handler, id_priv);
3554 id_priv->cm_id.ib = id;
3556 route = &id_priv->id.route;
3558 ret = cma_format_hdr(private_data, id_priv);
3561 req.private_data = private_data;
3564 req.primary_path = &route->path_rec[0];
3565 if (route->num_paths == 2)
3566 req.alternate_path = &route->path_rec[1];
3568 req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
3569 req.qp_num = id_priv->qp_num;
3570 req.qp_type = id_priv->id.qp_type;
3571 req.starting_psn = id_priv->seq_num;
3572 req.responder_resources = conn_param->responder_resources;
3573 req.initiator_depth = conn_param->initiator_depth;
3574 req.flow_control = conn_param->flow_control;
3575 req.retry_count = min_t(u8, 7, conn_param->retry_count);
3576 req.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
3577 req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
3578 req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
3579 req.max_cm_retries = CMA_MAX_CM_RETRIES;
3580 req.srq = id_priv->srq ? 1 : 0;
3582 ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
3584 if (ret && !IS_ERR(id)) {
3585 ib_destroy_cm_id(id);
3586 id_priv->cm_id.ib = NULL;
3589 kfree(private_data);
3593 static int cma_connect_iw(struct rdma_id_private *id_priv,
3594 struct rdma_conn_param *conn_param)
3596 struct iw_cm_id *cm_id;
3598 struct iw_cm_conn_param iw_param;
3600 cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
3602 return PTR_ERR(cm_id);
3604 cm_id->tos = id_priv->tos;
3605 id_priv->cm_id.iw = cm_id;
3607 memcpy(&cm_id->local_addr, cma_src_addr(id_priv),
3608 rdma_addr_size(cma_src_addr(id_priv)));
3609 memcpy(&cm_id->remote_addr, cma_dst_addr(id_priv),
3610 rdma_addr_size(cma_dst_addr(id_priv)));
3612 ret = cma_modify_qp_rtr(id_priv, conn_param);
3617 iw_param.ord = conn_param->initiator_depth;
3618 iw_param.ird = conn_param->responder_resources;
3619 iw_param.private_data = conn_param->private_data;
3620 iw_param.private_data_len = conn_param->private_data_len;
3621 iw_param.qpn = id_priv->id.qp ? id_priv->qp_num : conn_param->qp_num;
3623 memset(&iw_param, 0, sizeof iw_param);
3624 iw_param.qpn = id_priv->qp_num;
3626 ret = iw_cm_connect(cm_id, &iw_param);
3629 iw_destroy_cm_id(cm_id);
3630 id_priv->cm_id.iw = NULL;
3635 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
3637 struct rdma_id_private *id_priv;
3640 id_priv = container_of(id, struct rdma_id_private, id);
3641 if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT))
3645 id_priv->qp_num = conn_param->qp_num;
3646 id_priv->srq = conn_param->srq;
3649 if (rdma_cap_ib_cm(id->device, id->port_num)) {
3650 if (id->qp_type == IB_QPT_UD)
3651 ret = cma_resolve_ib_udp(id_priv, conn_param);
3653 ret = cma_connect_ib(id_priv, conn_param);
3654 } else if (rdma_cap_iw_cm(id->device, id->port_num))
3655 ret = cma_connect_iw(id_priv, conn_param);
3663 cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED);
3666 EXPORT_SYMBOL(rdma_connect);
3668 static int cma_accept_ib(struct rdma_id_private *id_priv,
3669 struct rdma_conn_param *conn_param)
3671 struct ib_cm_rep_param rep;
3674 ret = cma_modify_qp_rtr(id_priv, conn_param);
3678 ret = cma_modify_qp_rts(id_priv, conn_param);
3682 memset(&rep, 0, sizeof rep);
3683 rep.qp_num = id_priv->qp_num;
3684 rep.starting_psn = id_priv->seq_num;
3685 rep.private_data = conn_param->private_data;
3686 rep.private_data_len = conn_param->private_data_len;
3687 rep.responder_resources = conn_param->responder_resources;
3688 rep.initiator_depth = conn_param->initiator_depth;
3689 rep.failover_accepted = 0;
3690 rep.flow_control = conn_param->flow_control;
3691 rep.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
3692 rep.srq = id_priv->srq ? 1 : 0;
3694 ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
3699 static int cma_accept_iw(struct rdma_id_private *id_priv,
3700 struct rdma_conn_param *conn_param)
3702 struct iw_cm_conn_param iw_param;
3708 ret = cma_modify_qp_rtr(id_priv, conn_param);
3712 iw_param.ord = conn_param->initiator_depth;
3713 iw_param.ird = conn_param->responder_resources;
3714 iw_param.private_data = conn_param->private_data;
3715 iw_param.private_data_len = conn_param->private_data_len;
3716 if (id_priv->id.qp) {
3717 iw_param.qpn = id_priv->qp_num;
3719 iw_param.qpn = conn_param->qp_num;
3721 return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
3724 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
3725 enum ib_cm_sidr_status status, u32 qkey,
3726 const void *private_data, int private_data_len)
3728 struct ib_cm_sidr_rep_param rep;
3731 memset(&rep, 0, sizeof rep);
3732 rep.status = status;
3733 if (status == IB_SIDR_SUCCESS) {
3734 ret = cma_set_qkey(id_priv, qkey);
3737 rep.qp_num = id_priv->qp_num;
3738 rep.qkey = id_priv->qkey;
3740 rep.private_data = private_data;
3741 rep.private_data_len = private_data_len;
3743 return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
3746 int __rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param,
3749 struct rdma_id_private *id_priv;
3752 id_priv = container_of(id, struct rdma_id_private, id);
3755 id_priv->res.kern_name = caller;
3757 rdma_restrack_set_task(&id_priv->res, current);
3759 if (!cma_comp(id_priv, RDMA_CM_CONNECT))
3762 if (!id->qp && conn_param) {
3763 id_priv->qp_num = conn_param->qp_num;
3764 id_priv->srq = conn_param->srq;
3767 if (rdma_cap_ib_cm(id->device, id->port_num)) {
3768 if (id->qp_type == IB_QPT_UD) {
3770 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
3772 conn_param->private_data,
3773 conn_param->private_data_len);
3775 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
3779 ret = cma_accept_ib(id_priv, conn_param);
3781 ret = cma_rep_recv(id_priv);
3783 } else if (rdma_cap_iw_cm(id->device, id->port_num))
3784 ret = cma_accept_iw(id_priv, conn_param);
3793 cma_modify_qp_err(id_priv);
3794 rdma_reject(id, NULL, 0);
3797 EXPORT_SYMBOL(__rdma_accept);
3799 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
3801 struct rdma_id_private *id_priv;
3804 id_priv = container_of(id, struct rdma_id_private, id);
3805 if (!id_priv->cm_id.ib)
3808 switch (id->device->node_type) {
3809 case RDMA_NODE_IB_CA:
3810 ret = ib_cm_notify(id_priv->cm_id.ib, event);
3818 EXPORT_SYMBOL(rdma_notify);
3820 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
3821 u8 private_data_len)
3823 struct rdma_id_private *id_priv;
3826 id_priv = container_of(id, struct rdma_id_private, id);
3827 if (!id_priv->cm_id.ib)
3830 if (rdma_cap_ib_cm(id->device, id->port_num)) {
3831 if (id->qp_type == IB_QPT_UD)
3832 ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT, 0,
3833 private_data, private_data_len);
3835 ret = ib_send_cm_rej(id_priv->cm_id.ib,
3836 IB_CM_REJ_CONSUMER_DEFINED, NULL,
3837 0, private_data, private_data_len);
3838 } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
3839 ret = iw_cm_reject(id_priv->cm_id.iw,
3840 private_data, private_data_len);
3846 EXPORT_SYMBOL(rdma_reject);
3848 int rdma_disconnect(struct rdma_cm_id *id)
3850 struct rdma_id_private *id_priv;
3853 id_priv = container_of(id, struct rdma_id_private, id);
3854 if (!id_priv->cm_id.ib)
3857 if (rdma_cap_ib_cm(id->device, id->port_num)) {
3858 ret = cma_modify_qp_err(id_priv);
3861 /* Initiate or respond to a disconnect. */
3862 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
3863 ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
3864 } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
3865 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
3872 EXPORT_SYMBOL(rdma_disconnect);
3874 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
3876 struct rdma_id_private *id_priv;
3877 struct cma_multicast *mc = multicast->context;
3878 struct rdma_cm_event event;
3881 id_priv = mc->id_priv;
3882 mutex_lock(&id_priv->handler_mutex);
3883 if (id_priv->state != RDMA_CM_ADDR_BOUND &&
3884 id_priv->state != RDMA_CM_ADDR_RESOLVED)
3888 status = cma_set_qkey(id_priv, be32_to_cpu(multicast->rec.qkey));
3890 pr_debug_ratelimited("RDMA CM: MULTICAST_ERROR: failed to join multicast. status %d\n",
3892 mutex_lock(&id_priv->qp_mutex);
3893 if (!status && id_priv->id.qp) {
3894 status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid,
3895 be16_to_cpu(multicast->rec.mlid));
3897 pr_debug_ratelimited("RDMA CM: MULTICAST_ERROR: failed to attach QP. status %d\n",
3900 mutex_unlock(&id_priv->qp_mutex);
3902 memset(&event, 0, sizeof event);
3903 event.status = status;
3904 event.param.ud.private_data = mc->context;
3906 struct rdma_dev_addr *dev_addr =
3907 &id_priv->id.route.addr.dev_addr;
3908 struct net_device *ndev =
3909 dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if);
3910 enum ib_gid_type gid_type =
3911 id_priv->cma_dev->default_gid_type[id_priv->id.port_num -
3912 rdma_start_port(id_priv->cma_dev->device)];
3914 event.event = RDMA_CM_EVENT_MULTICAST_JOIN;
3915 ret = ib_init_ah_from_mcmember(id_priv->id.device,
3916 id_priv->id.port_num,
3919 &event.param.ud.ah_attr);
3921 event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
3923 event.param.ud.qp_num = 0xFFFFFF;
3924 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
3928 event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
3930 ret = id_priv->id.event_handler(&id_priv->id, &event);
3932 cma_exch(id_priv, RDMA_CM_DESTROYING);
3933 mutex_unlock(&id_priv->handler_mutex);
3934 rdma_destroy_id(&id_priv->id);
3939 mutex_unlock(&id_priv->handler_mutex);
3943 static void cma_set_mgid(struct rdma_id_private *id_priv,
3944 struct sockaddr *addr, union ib_gid *mgid)
3946 unsigned char mc_map[MAX_ADDR_LEN];
3947 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3948 struct sockaddr_in *sin = (struct sockaddr_in *) addr;
3949 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
3951 if (cma_any_addr(addr)) {
3952 memset(mgid, 0, sizeof *mgid);
3953 } else if ((addr->sa_family == AF_INET6) &&
3954 ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
3956 /* IPv6 address is an SA assigned MGID. */
3957 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3958 } else if (addr->sa_family == AF_IB) {
3959 memcpy(mgid, &((struct sockaddr_ib *) addr)->sib_addr, sizeof *mgid);
3960 } else if ((addr->sa_family == AF_INET6)) {
3961 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
3962 if (id_priv->id.ps == RDMA_PS_UDP)
3963 mc_map[7] = 0x01; /* Use RDMA CM signature */
3964 *mgid = *(union ib_gid *) (mc_map + 4);
3966 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
3967 if (id_priv->id.ps == RDMA_PS_UDP)
3968 mc_map[7] = 0x01; /* Use RDMA CM signature */
3969 *mgid = *(union ib_gid *) (mc_map + 4);
3973 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
3974 struct cma_multicast *mc)
3976 struct ib_sa_mcmember_rec rec;
3977 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3978 ib_sa_comp_mask comp_mask;
3981 ib_addr_get_mgid(dev_addr, &rec.mgid);
3982 ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
3987 ret = cma_set_qkey(id_priv, 0);
3991 cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
3992 rec.qkey = cpu_to_be32(id_priv->qkey);
3993 rdma_addr_get_sgid(dev_addr, &rec.port_gid);
3994 rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
3995 rec.join_state = mc->join_state;
3997 if ((rec.join_state == BIT(SENDONLY_FULLMEMBER_JOIN)) &&
3998 (!ib_sa_sendonly_fullmem_support(&sa_client,
4000 id_priv->id.port_num))) {
4001 pr_warn("RDMA CM: %s port %u Unable to multicast join\n"
4002 "RDMA CM: SM doesn't support Send Only Full Member option\n",
4003 id_priv->id.device->name, id_priv->id.port_num);
4007 comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
4008 IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
4009 IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
4010 IB_SA_MCMEMBER_REC_FLOW_LABEL |
4011 IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
4013 if (id_priv->id.ps == RDMA_PS_IPOIB)
4014 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
4015 IB_SA_MCMEMBER_REC_RATE_SELECTOR |
4016 IB_SA_MCMEMBER_REC_MTU_SELECTOR |
4017 IB_SA_MCMEMBER_REC_MTU |
4018 IB_SA_MCMEMBER_REC_HOP_LIMIT;
4020 mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device,
4021 id_priv->id.port_num, &rec,
4022 comp_mask, GFP_KERNEL,
4023 cma_ib_mc_handler, mc);
4024 return PTR_ERR_OR_ZERO(mc->multicast.ib);
4027 static void iboe_mcast_work_handler(struct work_struct *work)
4029 struct iboe_mcast_work *mw = container_of(work, struct iboe_mcast_work, work);
4030 struct cma_multicast *mc = mw->mc;
4031 struct ib_sa_multicast *m = mc->multicast.ib;
4033 mc->multicast.ib->context = mc;
4034 cma_ib_mc_handler(0, m);
4035 kref_put(&mc->mcref, release_mc);
4039 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid,
4040 enum ib_gid_type gid_type)
4042 struct sockaddr_in *sin = (struct sockaddr_in *)addr;
4043 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
4045 if (cma_any_addr(addr)) {
4046 memset(mgid, 0, sizeof *mgid);
4047 } else if (addr->sa_family == AF_INET6) {
4048 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
4051 (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ? 0 : 0xff;
4053 (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ? 0 : 0x0e;
4062 mgid->raw[10] = 0xff;
4063 mgid->raw[11] = 0xff;
4064 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
4068 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
4069 struct cma_multicast *mc)
4071 struct iboe_mcast_work *work;
4072 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4074 struct sockaddr *addr = (struct sockaddr *)&mc->addr;
4075 struct net_device *ndev = NULL;
4076 enum ib_gid_type gid_type;
4079 send_only = mc->join_state == BIT(SENDONLY_FULLMEMBER_JOIN);
4081 if (cma_zero_addr((struct sockaddr *)&mc->addr))
4084 work = kzalloc(sizeof *work, GFP_KERNEL);
4088 mc->multicast.ib = kzalloc(sizeof(struct ib_sa_multicast), GFP_KERNEL);
4089 if (!mc->multicast.ib) {
4094 gid_type = id_priv->cma_dev->default_gid_type[id_priv->id.port_num -
4095 rdma_start_port(id_priv->cma_dev->device)];
4096 cma_iboe_set_mgid(addr, &mc->multicast.ib->rec.mgid, gid_type);
4098 mc->multicast.ib->rec.pkey = cpu_to_be16(0xffff);
4099 if (id_priv->id.ps == RDMA_PS_UDP)
4100 mc->multicast.ib->rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
4102 if (dev_addr->bound_dev_if)
4103 ndev = dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if);
4108 mc->multicast.ib->rec.rate = iboe_get_rate(ndev);
4109 mc->multicast.ib->rec.hop_limit = 1;
4110 mc->multicast.ib->rec.mtu = iboe_get_mtu(ndev->mtu);
4112 if (addr->sa_family == AF_INET) {
4113 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) {
4114 mc->multicast.ib->rec.hop_limit = IPV6_DEFAULT_HOPLIMIT;
4116 err = cma_igmp_send(ndev, &mc->multicast.ib->rec.mgid,
4119 mc->igmp_joined = true;
4123 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP)
4127 if (err || !mc->multicast.ib->rec.mtu) {
4132 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
4133 &mc->multicast.ib->rec.port_gid);
4136 INIT_WORK(&work->work, iboe_mcast_work_handler);
4137 kref_get(&mc->mcref);
4138 queue_work(cma_wq, &work->work);
4143 kfree(mc->multicast.ib);
4149 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
4150 u8 join_state, void *context)
4152 struct rdma_id_private *id_priv;
4153 struct cma_multicast *mc;
4159 id_priv = container_of(id, struct rdma_id_private, id);
4160 if (!cma_comp(id_priv, RDMA_CM_ADDR_BOUND) &&
4161 !cma_comp(id_priv, RDMA_CM_ADDR_RESOLVED))
4164 mc = kmalloc(sizeof *mc, GFP_KERNEL);
4168 memcpy(&mc->addr, addr, rdma_addr_size(addr));
4169 mc->context = context;
4170 mc->id_priv = id_priv;
4171 mc->igmp_joined = false;
4172 mc->join_state = join_state;
4173 spin_lock(&id_priv->lock);
4174 list_add(&mc->list, &id_priv->mc_list);
4175 spin_unlock(&id_priv->lock);
4177 if (rdma_protocol_roce(id->device, id->port_num)) {
4178 kref_init(&mc->mcref);
4179 ret = cma_iboe_join_multicast(id_priv, mc);
4180 } else if (rdma_cap_ib_mcast(id->device, id->port_num))
4181 ret = cma_join_ib_multicast(id_priv, mc);
4186 spin_lock_irq(&id_priv->lock);
4187 list_del(&mc->list);
4188 spin_unlock_irq(&id_priv->lock);
4193 EXPORT_SYMBOL(rdma_join_multicast);
4195 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
4197 struct rdma_id_private *id_priv;
4198 struct cma_multicast *mc;
4200 id_priv = container_of(id, struct rdma_id_private, id);
4201 spin_lock_irq(&id_priv->lock);
4202 list_for_each_entry(mc, &id_priv->mc_list, list) {
4203 if (!memcmp(&mc->addr, addr, rdma_addr_size(addr))) {
4204 list_del(&mc->list);
4205 spin_unlock_irq(&id_priv->lock);
4208 ib_detach_mcast(id->qp,
4209 &mc->multicast.ib->rec.mgid,
4210 be16_to_cpu(mc->multicast.ib->rec.mlid));
4212 BUG_ON(id_priv->cma_dev->device != id->device);
4214 if (rdma_cap_ib_mcast(id->device, id->port_num)) {
4215 ib_sa_free_multicast(mc->multicast.ib);
4217 } else if (rdma_protocol_roce(id->device, id->port_num)) {
4218 if (mc->igmp_joined) {
4219 struct rdma_dev_addr *dev_addr =
4220 &id->route.addr.dev_addr;
4221 struct net_device *ndev = NULL;
4223 if (dev_addr->bound_dev_if)
4224 ndev = dev_get_by_index(dev_addr->net,
4225 dev_addr->bound_dev_if);
4228 &mc->multicast.ib->rec.mgid,
4232 mc->igmp_joined = false;
4234 kref_put(&mc->mcref, release_mc);
4239 spin_unlock_irq(&id_priv->lock);
4241 EXPORT_SYMBOL(rdma_leave_multicast);
4243 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
4245 struct rdma_dev_addr *dev_addr;
4246 struct cma_ndev_work *work;
4248 dev_addr = &id_priv->id.route.addr.dev_addr;
4250 if ((dev_addr->bound_dev_if == ndev->ifindex) &&
4251 (net_eq(dev_net(ndev), dev_addr->net)) &&
4252 memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
4253 pr_info("RDMA CM addr change for ndev %s used by id %p\n",
4254 ndev->name, &id_priv->id);
4255 work = kzalloc(sizeof *work, GFP_KERNEL);
4259 INIT_WORK(&work->work, cma_ndev_work_handler);
4261 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
4262 atomic_inc(&id_priv->refcount);
4263 queue_work(cma_wq, &work->work);
4269 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
4272 struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
4273 struct cma_device *cma_dev;
4274 struct rdma_id_private *id_priv;
4275 int ret = NOTIFY_DONE;
4277 if (event != NETDEV_BONDING_FAILOVER)
4280 if (!netif_is_bond_master(ndev))
4284 list_for_each_entry(cma_dev, &dev_list, list)
4285 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
4286 ret = cma_netdev_change(ndev, id_priv);
4292 mutex_unlock(&lock);
4296 static struct notifier_block cma_nb = {
4297 .notifier_call = cma_netdev_callback
4300 static void cma_add_one(struct ib_device *device)
4302 struct cma_device *cma_dev;
4303 struct rdma_id_private *id_priv;
4305 unsigned long supported_gids = 0;
4307 cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
4311 cma_dev->device = device;
4312 cma_dev->default_gid_type = kcalloc(device->phys_port_cnt,
4313 sizeof(*cma_dev->default_gid_type),
4315 if (!cma_dev->default_gid_type)
4318 cma_dev->default_roce_tos = kcalloc(device->phys_port_cnt,
4319 sizeof(*cma_dev->default_roce_tos),
4321 if (!cma_dev->default_roce_tos)
4324 for (i = rdma_start_port(device); i <= rdma_end_port(device); i++) {
4325 supported_gids = roce_gid_type_mask_support(device, i);
4326 WARN_ON(!supported_gids);
4327 if (supported_gids & (1 << CMA_PREFERRED_ROCE_GID_TYPE))
4328 cma_dev->default_gid_type[i - rdma_start_port(device)] =
4329 CMA_PREFERRED_ROCE_GID_TYPE;
4331 cma_dev->default_gid_type[i - rdma_start_port(device)] =
4332 find_first_bit(&supported_gids, BITS_PER_LONG);
4333 cma_dev->default_roce_tos[i - rdma_start_port(device)] = 0;
4336 init_completion(&cma_dev->comp);
4337 atomic_set(&cma_dev->refcount, 1);
4338 INIT_LIST_HEAD(&cma_dev->id_list);
4339 ib_set_client_data(device, &cma_client, cma_dev);
4342 list_add_tail(&cma_dev->list, &dev_list);
4343 list_for_each_entry(id_priv, &listen_any_list, list)
4344 cma_listen_on_dev(id_priv, cma_dev);
4345 mutex_unlock(&lock);
4350 kfree(cma_dev->default_gid_type);
4358 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
4360 struct rdma_cm_event event;
4361 enum rdma_cm_state state;
4364 /* Record that we want to remove the device */
4365 state = cma_exch(id_priv, RDMA_CM_DEVICE_REMOVAL);
4366 if (state == RDMA_CM_DESTROYING)
4369 cma_cancel_operation(id_priv, state);
4370 mutex_lock(&id_priv->handler_mutex);
4372 /* Check for destruction from another callback. */
4373 if (!cma_comp(id_priv, RDMA_CM_DEVICE_REMOVAL))
4376 memset(&event, 0, sizeof event);
4377 event.event = RDMA_CM_EVENT_DEVICE_REMOVAL;
4378 ret = id_priv->id.event_handler(&id_priv->id, &event);
4380 mutex_unlock(&id_priv->handler_mutex);
4384 static void cma_process_remove(struct cma_device *cma_dev)
4386 struct rdma_id_private *id_priv;
4390 while (!list_empty(&cma_dev->id_list)) {
4391 id_priv = list_entry(cma_dev->id_list.next,
4392 struct rdma_id_private, list);
4394 list_del(&id_priv->listen_list);
4395 list_del_init(&id_priv->list);
4396 atomic_inc(&id_priv->refcount);
4397 mutex_unlock(&lock);
4399 ret = id_priv->internal_id ? 1 : cma_remove_id_dev(id_priv);
4400 cma_deref_id(id_priv);
4402 rdma_destroy_id(&id_priv->id);
4406 mutex_unlock(&lock);
4408 cma_deref_dev(cma_dev);
4409 wait_for_completion(&cma_dev->comp);
4412 static void cma_remove_one(struct ib_device *device, void *client_data)
4414 struct cma_device *cma_dev = client_data;
4420 list_del(&cma_dev->list);
4421 mutex_unlock(&lock);
4423 cma_process_remove(cma_dev);
4424 kfree(cma_dev->default_roce_tos);
4425 kfree(cma_dev->default_gid_type);
4429 static int cma_get_id_stats(struct sk_buff *skb, struct netlink_callback *cb)
4431 struct nlmsghdr *nlh;
4432 struct rdma_cm_id_stats *id_stats;
4433 struct rdma_id_private *id_priv;
4434 struct rdma_cm_id *id = NULL;
4435 struct cma_device *cma_dev;
4436 int i_dev = 0, i_id = 0;
4439 * We export all of the IDs as a sequence of messages. Each
4440 * ID gets its own netlink message.
4444 list_for_each_entry(cma_dev, &dev_list, list) {
4445 if (i_dev < cb->args[0]) {
4451 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
4452 if (i_id < cb->args[1]) {
4457 id_stats = ibnl_put_msg(skb, &nlh, cb->nlh->nlmsg_seq,
4458 sizeof *id_stats, RDMA_NL_RDMA_CM,
4459 RDMA_NL_RDMA_CM_ID_STATS,
4464 memset(id_stats, 0, sizeof *id_stats);
4466 id_stats->node_type = id->route.addr.dev_addr.dev_type;
4467 id_stats->port_num = id->port_num;
4468 id_stats->bound_dev_if =
4469 id->route.addr.dev_addr.bound_dev_if;
4471 if (ibnl_put_attr(skb, nlh,
4472 rdma_addr_size(cma_src_addr(id_priv)),
4473 cma_src_addr(id_priv),
4474 RDMA_NL_RDMA_CM_ATTR_SRC_ADDR))
4476 if (ibnl_put_attr(skb, nlh,
4477 rdma_addr_size(cma_dst_addr(id_priv)),
4478 cma_dst_addr(id_priv),
4479 RDMA_NL_RDMA_CM_ATTR_DST_ADDR))
4482 id_stats->pid = task_pid_vnr(id_priv->res.task);
4483 id_stats->port_space = id->ps;
4484 id_stats->cm_state = id_priv->state;
4485 id_stats->qp_num = id_priv->qp_num;
4486 id_stats->qp_type = id->qp_type;
4489 nlmsg_end(skb, nlh);
4497 mutex_unlock(&lock);
4498 cb->args[0] = i_dev;
4504 static const struct rdma_nl_cbs cma_cb_table[RDMA_NL_RDMA_CM_NUM_OPS] = {
4505 [RDMA_NL_RDMA_CM_ID_STATS] = { .dump = cma_get_id_stats},
4508 static int cma_init_net(struct net *net)
4510 struct cma_pernet *pernet = cma_pernet(net);
4512 idr_init(&pernet->tcp_ps);
4513 idr_init(&pernet->udp_ps);
4514 idr_init(&pernet->ipoib_ps);
4515 idr_init(&pernet->ib_ps);
4520 static void cma_exit_net(struct net *net)
4522 struct cma_pernet *pernet = cma_pernet(net);
4524 idr_destroy(&pernet->tcp_ps);
4525 idr_destroy(&pernet->udp_ps);
4526 idr_destroy(&pernet->ipoib_ps);
4527 idr_destroy(&pernet->ib_ps);
4530 static struct pernet_operations cma_pernet_operations = {
4531 .init = cma_init_net,
4532 .exit = cma_exit_net,
4533 .id = &cma_pernet_id,
4534 .size = sizeof(struct cma_pernet),
4537 static int __init cma_init(void)
4541 cma_wq = alloc_ordered_workqueue("rdma_cm", WQ_MEM_RECLAIM);
4545 ret = register_pernet_subsys(&cma_pernet_operations);
4549 ib_sa_register_client(&sa_client);
4550 rdma_addr_register_client(&addr_client);
4551 register_netdevice_notifier(&cma_nb);
4553 ret = ib_register_client(&cma_client);
4557 rdma_nl_register(RDMA_NL_RDMA_CM, cma_cb_table);
4558 cma_configfs_init();
4563 unregister_netdevice_notifier(&cma_nb);
4564 rdma_addr_unregister_client(&addr_client);
4565 ib_sa_unregister_client(&sa_client);
4567 destroy_workqueue(cma_wq);
4571 static void __exit cma_cleanup(void)
4573 cma_configfs_exit();
4574 rdma_nl_unregister(RDMA_NL_RDMA_CM);
4575 ib_unregister_client(&cma_client);
4576 unregister_netdevice_notifier(&cma_nb);
4577 rdma_addr_unregister_client(&addr_client);
4578 ib_sa_unregister_client(&sa_client);
4579 unregister_pernet_subsys(&cma_pernet_operations);
4580 destroy_workqueue(cma_wq);
4583 MODULE_ALIAS_RDMA_NETLINK(RDMA_NL_RDMA_CM, 1);
4585 module_init(cma_init);
4586 module_exit(cma_cleanup);