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);
150 * rdma_iw_cm_id() - return the iw_cm_id pointer for this cm_id.
151 * @id: Communication Identifier
153 struct iw_cm_id *rdma_iw_cm_id(struct rdma_cm_id *id)
155 struct rdma_id_private *id_priv;
157 id_priv = container_of(id, struct rdma_id_private, id);
158 if (id->device->node_type == RDMA_NODE_RNIC)
159 return id_priv->cm_id.iw;
162 EXPORT_SYMBOL(rdma_iw_cm_id);
165 * rdma_res_to_id() - return the rdma_cm_id pointer for this restrack.
166 * @res: rdma resource tracking entry pointer
168 struct rdma_cm_id *rdma_res_to_id(struct rdma_restrack_entry *res)
170 struct rdma_id_private *id_priv =
171 container_of(res, struct rdma_id_private, res);
175 EXPORT_SYMBOL(rdma_res_to_id);
177 static void cma_add_one(struct ib_device *device);
178 static void cma_remove_one(struct ib_device *device, void *client_data);
180 static struct ib_client cma_client = {
183 .remove = cma_remove_one
186 static struct ib_sa_client sa_client;
187 static LIST_HEAD(dev_list);
188 static LIST_HEAD(listen_any_list);
189 static DEFINE_MUTEX(lock);
190 static struct workqueue_struct *cma_wq;
191 static unsigned int cma_pernet_id;
200 static struct cma_pernet *cma_pernet(struct net *net)
202 return net_generic(net, cma_pernet_id);
205 static struct idr *cma_pernet_idr(struct net *net, enum rdma_ucm_port_space ps)
207 struct cma_pernet *pernet = cma_pernet(net);
211 return &pernet->tcp_ps;
213 return &pernet->udp_ps;
215 return &pernet->ipoib_ps;
217 return &pernet->ib_ps;
224 struct list_head list;
225 struct ib_device *device;
226 struct completion comp;
228 struct list_head id_list;
229 enum ib_gid_type *default_gid_type;
230 u8 *default_roce_tos;
233 struct rdma_bind_list {
234 enum rdma_ucm_port_space ps;
235 struct hlist_head owners;
239 struct class_port_info_context {
240 struct ib_class_port_info *class_port_info;
241 struct ib_device *device;
242 struct completion done;
243 struct ib_sa_query *sa_query;
247 static int cma_ps_alloc(struct net *net, enum rdma_ucm_port_space ps,
248 struct rdma_bind_list *bind_list, int snum)
250 struct idr *idr = cma_pernet_idr(net, ps);
252 return idr_alloc(idr, bind_list, snum, snum + 1, GFP_KERNEL);
255 static struct rdma_bind_list *cma_ps_find(struct net *net,
256 enum rdma_ucm_port_space ps, int snum)
258 struct idr *idr = cma_pernet_idr(net, ps);
260 return idr_find(idr, snum);
263 static void cma_ps_remove(struct net *net, enum rdma_ucm_port_space ps,
266 struct idr *idr = cma_pernet_idr(net, ps);
268 idr_remove(idr, snum);
275 void cma_ref_dev(struct cma_device *cma_dev)
277 atomic_inc(&cma_dev->refcount);
280 struct cma_device *cma_enum_devices_by_ibdev(cma_device_filter filter,
283 struct cma_device *cma_dev;
284 struct cma_device *found_cma_dev = NULL;
288 list_for_each_entry(cma_dev, &dev_list, list)
289 if (filter(cma_dev->device, cookie)) {
290 found_cma_dev = cma_dev;
295 cma_ref_dev(found_cma_dev);
297 return found_cma_dev;
300 int cma_get_default_gid_type(struct cma_device *cma_dev,
303 if (!rdma_is_port_valid(cma_dev->device, port))
306 return cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)];
309 int cma_set_default_gid_type(struct cma_device *cma_dev,
311 enum ib_gid_type default_gid_type)
313 unsigned long supported_gids;
315 if (!rdma_is_port_valid(cma_dev->device, port))
318 supported_gids = roce_gid_type_mask_support(cma_dev->device, port);
320 if (!(supported_gids & 1 << default_gid_type))
323 cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)] =
329 int cma_get_default_roce_tos(struct cma_device *cma_dev, unsigned int port)
331 if (!rdma_is_port_valid(cma_dev->device, port))
334 return cma_dev->default_roce_tos[port - rdma_start_port(cma_dev->device)];
337 int cma_set_default_roce_tos(struct cma_device *cma_dev, unsigned int port,
340 if (!rdma_is_port_valid(cma_dev->device, port))
343 cma_dev->default_roce_tos[port - rdma_start_port(cma_dev->device)] =
348 struct ib_device *cma_get_ib_dev(struct cma_device *cma_dev)
350 return cma_dev->device;
354 * Device removal can occur at anytime, so we need extra handling to
355 * serialize notifying the user of device removal with other callbacks.
356 * We do this by disabling removal notification while a callback is in process,
357 * and reporting it after the callback completes.
360 struct cma_multicast {
361 struct rdma_id_private *id_priv;
363 struct ib_sa_multicast *ib;
365 struct list_head list;
367 struct sockaddr_storage addr;
373 struct work_struct work;
374 struct rdma_id_private *id;
375 enum rdma_cm_state old_state;
376 enum rdma_cm_state new_state;
377 struct rdma_cm_event event;
380 struct cma_ndev_work {
381 struct work_struct work;
382 struct rdma_id_private *id;
383 struct rdma_cm_event event;
386 struct iboe_mcast_work {
387 struct work_struct work;
388 struct rdma_id_private *id;
389 struct cma_multicast *mc;
402 u8 ip_version; /* IP version: 7:4 */
404 union cma_ip_addr src_addr;
405 union cma_ip_addr dst_addr;
408 #define CMA_VERSION 0x00
410 struct cma_req_info {
411 struct sockaddr_storage listen_addr_storage;
412 struct sockaddr_storage src_addr_storage;
413 struct ib_device *device;
414 union ib_gid local_gid;
421 static int cma_comp(struct rdma_id_private *id_priv, enum rdma_cm_state comp)
426 spin_lock_irqsave(&id_priv->lock, flags);
427 ret = (id_priv->state == comp);
428 spin_unlock_irqrestore(&id_priv->lock, flags);
432 static int cma_comp_exch(struct rdma_id_private *id_priv,
433 enum rdma_cm_state comp, enum rdma_cm_state exch)
438 spin_lock_irqsave(&id_priv->lock, flags);
439 if ((ret = (id_priv->state == comp)))
440 id_priv->state = exch;
441 spin_unlock_irqrestore(&id_priv->lock, flags);
445 static enum rdma_cm_state cma_exch(struct rdma_id_private *id_priv,
446 enum rdma_cm_state exch)
449 enum rdma_cm_state old;
451 spin_lock_irqsave(&id_priv->lock, flags);
452 old = id_priv->state;
453 id_priv->state = exch;
454 spin_unlock_irqrestore(&id_priv->lock, flags);
458 static inline u8 cma_get_ip_ver(const struct cma_hdr *hdr)
460 return hdr->ip_version >> 4;
463 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
465 hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
468 static int cma_igmp_send(struct net_device *ndev, union ib_gid *mgid, bool join)
470 struct in_device *in_dev = NULL;
474 in_dev = __in_dev_get_rtnl(ndev);
477 ip_mc_inc_group(in_dev,
478 *(__be32 *)(mgid->raw + 12));
480 ip_mc_dec_group(in_dev,
481 *(__be32 *)(mgid->raw + 12));
485 return (in_dev) ? 0 : -ENODEV;
488 static void _cma_attach_to_dev(struct rdma_id_private *id_priv,
489 struct cma_device *cma_dev)
491 cma_ref_dev(cma_dev);
492 id_priv->cma_dev = cma_dev;
493 id_priv->id.device = cma_dev->device;
494 id_priv->id.route.addr.dev_addr.transport =
495 rdma_node_get_transport(cma_dev->device->node_type);
496 list_add_tail(&id_priv->list, &cma_dev->id_list);
497 rdma_restrack_add(&id_priv->res);
500 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
501 struct cma_device *cma_dev)
503 _cma_attach_to_dev(id_priv, cma_dev);
505 cma_dev->default_gid_type[id_priv->id.port_num -
506 rdma_start_port(cma_dev->device)];
509 void cma_deref_dev(struct cma_device *cma_dev)
511 if (atomic_dec_and_test(&cma_dev->refcount))
512 complete(&cma_dev->comp);
515 static inline void release_mc(struct kref *kref)
517 struct cma_multicast *mc = container_of(kref, struct cma_multicast, mcref);
519 kfree(mc->multicast.ib);
523 static void cma_release_dev(struct rdma_id_private *id_priv)
526 list_del(&id_priv->list);
527 cma_deref_dev(id_priv->cma_dev);
528 id_priv->cma_dev = NULL;
532 static inline struct sockaddr *cma_src_addr(struct rdma_id_private *id_priv)
534 return (struct sockaddr *) &id_priv->id.route.addr.src_addr;
537 static inline struct sockaddr *cma_dst_addr(struct rdma_id_private *id_priv)
539 return (struct sockaddr *) &id_priv->id.route.addr.dst_addr;
542 static inline unsigned short cma_family(struct rdma_id_private *id_priv)
544 return id_priv->id.route.addr.src_addr.ss_family;
547 static int cma_set_qkey(struct rdma_id_private *id_priv, u32 qkey)
549 struct ib_sa_mcmember_rec rec;
553 if (qkey && id_priv->qkey != qkey)
559 id_priv->qkey = qkey;
563 switch (id_priv->id.ps) {
566 id_priv->qkey = RDMA_UDP_QKEY;
569 ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid);
570 ret = ib_sa_get_mcmember_rec(id_priv->id.device,
571 id_priv->id.port_num, &rec.mgid,
574 id_priv->qkey = be32_to_cpu(rec.qkey);
582 static void cma_translate_ib(struct sockaddr_ib *sib, struct rdma_dev_addr *dev_addr)
584 dev_addr->dev_type = ARPHRD_INFINIBAND;
585 rdma_addr_set_sgid(dev_addr, (union ib_gid *) &sib->sib_addr);
586 ib_addr_set_pkey(dev_addr, ntohs(sib->sib_pkey));
589 static int cma_translate_addr(struct sockaddr *addr, struct rdma_dev_addr *dev_addr)
593 if (addr->sa_family != AF_IB) {
594 ret = rdma_translate_ip(addr, dev_addr);
596 cma_translate_ib((struct sockaddr_ib *) addr, dev_addr);
603 static const struct ib_gid_attr *
604 cma_validate_port(struct ib_device *device, u8 port,
605 enum ib_gid_type gid_type,
607 struct rdma_id_private *id_priv)
609 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
610 int bound_if_index = dev_addr->bound_dev_if;
611 const struct ib_gid_attr *sgid_attr;
612 int dev_type = dev_addr->dev_type;
613 struct net_device *ndev = NULL;
615 if ((dev_type == ARPHRD_INFINIBAND) && !rdma_protocol_ib(device, port))
616 return ERR_PTR(-ENODEV);
618 if ((dev_type != ARPHRD_INFINIBAND) && rdma_protocol_ib(device, port))
619 return ERR_PTR(-ENODEV);
621 if (dev_type == ARPHRD_ETHER && rdma_protocol_roce(device, port)) {
622 ndev = dev_get_by_index(dev_addr->net, bound_if_index);
624 return ERR_PTR(-ENODEV);
626 gid_type = IB_GID_TYPE_IB;
629 sgid_attr = rdma_find_gid_by_port(device, gid, gid_type, port, ndev);
635 static void cma_bind_sgid_attr(struct rdma_id_private *id_priv,
636 const struct ib_gid_attr *sgid_attr)
638 WARN_ON(id_priv->id.route.addr.dev_addr.sgid_attr);
639 id_priv->id.route.addr.dev_addr.sgid_attr = sgid_attr;
642 static int cma_acquire_dev(struct rdma_id_private *id_priv,
643 const struct rdma_id_private *listen_id_priv)
645 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
646 const struct ib_gid_attr *sgid_attr;
647 struct cma_device *cma_dev;
648 union ib_gid gid, iboe_gid, *gidp;
649 enum ib_gid_type gid_type;
653 if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
654 id_priv->id.ps == RDMA_PS_IPOIB)
658 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
661 memcpy(&gid, dev_addr->src_dev_addr +
662 rdma_addr_gid_offset(dev_addr), sizeof gid);
664 if (listen_id_priv) {
665 cma_dev = listen_id_priv->cma_dev;
666 port = listen_id_priv->id.port_num;
667 gidp = rdma_protocol_roce(cma_dev->device, port) ?
669 gid_type = listen_id_priv->gid_type;
670 sgid_attr = cma_validate_port(cma_dev->device, port,
671 gid_type, gidp, id_priv);
672 if (!IS_ERR(sgid_attr)) {
673 id_priv->id.port_num = port;
674 cma_bind_sgid_attr(id_priv, sgid_attr);
680 list_for_each_entry(cma_dev, &dev_list, list) {
681 for (port = 1; port <= cma_dev->device->phys_port_cnt; ++port) {
682 if (listen_id_priv &&
683 listen_id_priv->cma_dev == cma_dev &&
684 listen_id_priv->id.port_num == port)
687 gidp = rdma_protocol_roce(cma_dev->device, port) ?
689 gid_type = cma_dev->default_gid_type[port - 1];
690 sgid_attr = cma_validate_port(cma_dev->device, port,
691 gid_type, gidp, id_priv);
692 if (!IS_ERR(sgid_attr)) {
693 id_priv->id.port_num = port;
694 cma_bind_sgid_attr(id_priv, sgid_attr);
703 cma_attach_to_dev(id_priv, cma_dev);
710 * Select the source IB device and address to reach the destination IB address.
712 static int cma_resolve_ib_dev(struct rdma_id_private *id_priv)
714 struct cma_device *cma_dev, *cur_dev;
715 struct sockaddr_ib *addr;
716 union ib_gid gid, sgid, *dgid;
719 enum ib_port_state port_state;
723 addr = (struct sockaddr_ib *) cma_dst_addr(id_priv);
724 dgid = (union ib_gid *) &addr->sib_addr;
725 pkey = ntohs(addr->sib_pkey);
728 list_for_each_entry(cur_dev, &dev_list, list) {
729 for (p = 1; p <= cur_dev->device->phys_port_cnt; ++p) {
730 if (!rdma_cap_af_ib(cur_dev->device, p))
733 if (ib_find_cached_pkey(cur_dev->device, p, pkey, &index))
736 if (ib_get_cached_port_state(cur_dev->device, p, &port_state))
738 for (i = 0; !rdma_query_gid(cur_dev->device,
741 if (!memcmp(&gid, dgid, sizeof(gid))) {
744 id_priv->id.port_num = p;
748 if (!cma_dev && (gid.global.subnet_prefix ==
749 dgid->global.subnet_prefix) &&
750 port_state == IB_PORT_ACTIVE) {
753 id_priv->id.port_num = p;
763 cma_attach_to_dev(id_priv, cma_dev);
765 addr = (struct sockaddr_ib *)cma_src_addr(id_priv);
766 memcpy(&addr->sib_addr, &sgid, sizeof(sgid));
767 cma_translate_ib(addr, &id_priv->id.route.addr.dev_addr);
771 static void cma_deref_id(struct rdma_id_private *id_priv)
773 if (atomic_dec_and_test(&id_priv->refcount))
774 complete(&id_priv->comp);
777 struct rdma_cm_id *__rdma_create_id(struct net *net,
778 rdma_cm_event_handler event_handler,
779 void *context, enum rdma_ucm_port_space ps,
780 enum ib_qp_type qp_type, const char *caller)
782 struct rdma_id_private *id_priv;
784 id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
786 return ERR_PTR(-ENOMEM);
789 id_priv->res.kern_name = caller;
791 rdma_restrack_set_task(&id_priv->res, current);
792 id_priv->res.type = RDMA_RESTRACK_CM_ID;
793 id_priv->state = RDMA_CM_IDLE;
794 id_priv->id.context = context;
795 id_priv->id.event_handler = event_handler;
797 id_priv->id.qp_type = qp_type;
798 id_priv->tos_set = false;
799 id_priv->gid_type = IB_GID_TYPE_IB;
800 spin_lock_init(&id_priv->lock);
801 mutex_init(&id_priv->qp_mutex);
802 init_completion(&id_priv->comp);
803 atomic_set(&id_priv->refcount, 1);
804 mutex_init(&id_priv->handler_mutex);
805 INIT_LIST_HEAD(&id_priv->listen_list);
806 INIT_LIST_HEAD(&id_priv->mc_list);
807 get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
808 id_priv->id.route.addr.dev_addr.net = get_net(net);
809 id_priv->seq_num &= 0x00ffffff;
813 EXPORT_SYMBOL(__rdma_create_id);
815 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
817 struct ib_qp_attr qp_attr;
818 int qp_attr_mask, ret;
820 qp_attr.qp_state = IB_QPS_INIT;
821 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
825 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
829 qp_attr.qp_state = IB_QPS_RTR;
830 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
834 qp_attr.qp_state = IB_QPS_RTS;
836 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
841 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
843 struct ib_qp_attr qp_attr;
844 int qp_attr_mask, ret;
846 qp_attr.qp_state = IB_QPS_INIT;
847 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
851 return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
854 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
855 struct ib_qp_init_attr *qp_init_attr)
857 struct rdma_id_private *id_priv;
861 id_priv = container_of(id, struct rdma_id_private, id);
862 if (id->device != pd->device)
865 qp_init_attr->port_num = id->port_num;
866 qp = ib_create_qp(pd, qp_init_attr);
870 if (id->qp_type == IB_QPT_UD)
871 ret = cma_init_ud_qp(id_priv, qp);
873 ret = cma_init_conn_qp(id_priv, qp);
878 id_priv->qp_num = qp->qp_num;
879 id_priv->srq = (qp->srq != NULL);
885 EXPORT_SYMBOL(rdma_create_qp);
887 void rdma_destroy_qp(struct rdma_cm_id *id)
889 struct rdma_id_private *id_priv;
891 id_priv = container_of(id, struct rdma_id_private, id);
892 mutex_lock(&id_priv->qp_mutex);
893 ib_destroy_qp(id_priv->id.qp);
894 id_priv->id.qp = NULL;
895 mutex_unlock(&id_priv->qp_mutex);
897 EXPORT_SYMBOL(rdma_destroy_qp);
899 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
900 struct rdma_conn_param *conn_param)
902 struct ib_qp_attr qp_attr;
903 int qp_attr_mask, ret;
905 mutex_lock(&id_priv->qp_mutex);
906 if (!id_priv->id.qp) {
911 /* Need to update QP attributes from default values. */
912 qp_attr.qp_state = IB_QPS_INIT;
913 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
917 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
921 qp_attr.qp_state = IB_QPS_RTR;
922 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
926 BUG_ON(id_priv->cma_dev->device != id_priv->id.device);
929 qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
930 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
932 mutex_unlock(&id_priv->qp_mutex);
936 static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
937 struct rdma_conn_param *conn_param)
939 struct ib_qp_attr qp_attr;
940 int qp_attr_mask, ret;
942 mutex_lock(&id_priv->qp_mutex);
943 if (!id_priv->id.qp) {
948 qp_attr.qp_state = IB_QPS_RTS;
949 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
954 qp_attr.max_rd_atomic = conn_param->initiator_depth;
955 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
957 mutex_unlock(&id_priv->qp_mutex);
961 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
963 struct ib_qp_attr qp_attr;
966 mutex_lock(&id_priv->qp_mutex);
967 if (!id_priv->id.qp) {
972 qp_attr.qp_state = IB_QPS_ERR;
973 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
975 mutex_unlock(&id_priv->qp_mutex);
979 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
980 struct ib_qp_attr *qp_attr, int *qp_attr_mask)
982 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
986 if (rdma_cap_eth_ah(id_priv->id.device, id_priv->id.port_num))
989 pkey = ib_addr_get_pkey(dev_addr);
991 ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
992 pkey, &qp_attr->pkey_index);
996 qp_attr->port_num = id_priv->id.port_num;
997 *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
999 if (id_priv->id.qp_type == IB_QPT_UD) {
1000 ret = cma_set_qkey(id_priv, 0);
1004 qp_attr->qkey = id_priv->qkey;
1005 *qp_attr_mask |= IB_QP_QKEY;
1007 qp_attr->qp_access_flags = 0;
1008 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
1013 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
1016 struct rdma_id_private *id_priv;
1019 id_priv = container_of(id, struct rdma_id_private, id);
1020 if (rdma_cap_ib_cm(id->device, id->port_num)) {
1021 if (!id_priv->cm_id.ib || (id_priv->id.qp_type == IB_QPT_UD))
1022 ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
1024 ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
1027 if (qp_attr->qp_state == IB_QPS_RTR)
1028 qp_attr->rq_psn = id_priv->seq_num;
1029 } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
1030 if (!id_priv->cm_id.iw) {
1031 qp_attr->qp_access_flags = 0;
1032 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
1034 ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
1036 qp_attr->port_num = id_priv->id.port_num;
1037 *qp_attr_mask |= IB_QP_PORT;
1043 EXPORT_SYMBOL(rdma_init_qp_attr);
1045 static inline bool cma_zero_addr(const struct sockaddr *addr)
1047 switch (addr->sa_family) {
1049 return ipv4_is_zeronet(((struct sockaddr_in *)addr)->sin_addr.s_addr);
1051 return ipv6_addr_any(&((struct sockaddr_in6 *)addr)->sin6_addr);
1053 return ib_addr_any(&((struct sockaddr_ib *)addr)->sib_addr);
1059 static inline bool cma_loopback_addr(const struct sockaddr *addr)
1061 switch (addr->sa_family) {
1063 return ipv4_is_loopback(
1064 ((struct sockaddr_in *)addr)->sin_addr.s_addr);
1066 return ipv6_addr_loopback(
1067 &((struct sockaddr_in6 *)addr)->sin6_addr);
1069 return ib_addr_loopback(
1070 &((struct sockaddr_ib *)addr)->sib_addr);
1076 static inline bool cma_any_addr(const struct sockaddr *addr)
1078 return cma_zero_addr(addr) || cma_loopback_addr(addr);
1081 static int cma_addr_cmp(const struct sockaddr *src, const struct sockaddr *dst)
1083 if (src->sa_family != dst->sa_family)
1086 switch (src->sa_family) {
1088 return ((struct sockaddr_in *)src)->sin_addr.s_addr !=
1089 ((struct sockaddr_in *)dst)->sin_addr.s_addr;
1091 struct sockaddr_in6 *src_addr6 = (struct sockaddr_in6 *)src;
1092 struct sockaddr_in6 *dst_addr6 = (struct sockaddr_in6 *)dst;
1095 if (ipv6_addr_cmp(&src_addr6->sin6_addr,
1096 &dst_addr6->sin6_addr))
1098 link_local = ipv6_addr_type(&dst_addr6->sin6_addr) &
1099 IPV6_ADDR_LINKLOCAL;
1100 /* Link local must match their scope_ids */
1101 return link_local ? (src_addr6->sin6_scope_id !=
1102 dst_addr6->sin6_scope_id) :
1107 return ib_addr_cmp(&((struct sockaddr_ib *) src)->sib_addr,
1108 &((struct sockaddr_ib *) dst)->sib_addr);
1112 static __be16 cma_port(const struct sockaddr *addr)
1114 struct sockaddr_ib *sib;
1116 switch (addr->sa_family) {
1118 return ((struct sockaddr_in *) addr)->sin_port;
1120 return ((struct sockaddr_in6 *) addr)->sin6_port;
1122 sib = (struct sockaddr_ib *) addr;
1123 return htons((u16) (be64_to_cpu(sib->sib_sid) &
1124 be64_to_cpu(sib->sib_sid_mask)));
1130 static inline int cma_any_port(const struct sockaddr *addr)
1132 return !cma_port(addr);
1135 static void cma_save_ib_info(struct sockaddr *src_addr,
1136 struct sockaddr *dst_addr,
1137 const struct rdma_cm_id *listen_id,
1138 const struct sa_path_rec *path)
1140 struct sockaddr_ib *listen_ib, *ib;
1142 listen_ib = (struct sockaddr_ib *) &listen_id->route.addr.src_addr;
1144 ib = (struct sockaddr_ib *)src_addr;
1145 ib->sib_family = AF_IB;
1147 ib->sib_pkey = path->pkey;
1148 ib->sib_flowinfo = path->flow_label;
1149 memcpy(&ib->sib_addr, &path->sgid, 16);
1150 ib->sib_sid = path->service_id;
1151 ib->sib_scope_id = 0;
1153 ib->sib_pkey = listen_ib->sib_pkey;
1154 ib->sib_flowinfo = listen_ib->sib_flowinfo;
1155 ib->sib_addr = listen_ib->sib_addr;
1156 ib->sib_sid = listen_ib->sib_sid;
1157 ib->sib_scope_id = listen_ib->sib_scope_id;
1159 ib->sib_sid_mask = cpu_to_be64(0xffffffffffffffffULL);
1162 ib = (struct sockaddr_ib *)dst_addr;
1163 ib->sib_family = AF_IB;
1165 ib->sib_pkey = path->pkey;
1166 ib->sib_flowinfo = path->flow_label;
1167 memcpy(&ib->sib_addr, &path->dgid, 16);
1172 static void cma_save_ip4_info(struct sockaddr_in *src_addr,
1173 struct sockaddr_in *dst_addr,
1174 struct cma_hdr *hdr,
1178 *src_addr = (struct sockaddr_in) {
1179 .sin_family = AF_INET,
1180 .sin_addr.s_addr = hdr->dst_addr.ip4.addr,
1181 .sin_port = local_port,
1186 *dst_addr = (struct sockaddr_in) {
1187 .sin_family = AF_INET,
1188 .sin_addr.s_addr = hdr->src_addr.ip4.addr,
1189 .sin_port = hdr->port,
1194 static void cma_save_ip6_info(struct sockaddr_in6 *src_addr,
1195 struct sockaddr_in6 *dst_addr,
1196 struct cma_hdr *hdr,
1200 *src_addr = (struct sockaddr_in6) {
1201 .sin6_family = AF_INET6,
1202 .sin6_addr = hdr->dst_addr.ip6,
1203 .sin6_port = local_port,
1208 *dst_addr = (struct sockaddr_in6) {
1209 .sin6_family = AF_INET6,
1210 .sin6_addr = hdr->src_addr.ip6,
1211 .sin6_port = hdr->port,
1216 static u16 cma_port_from_service_id(__be64 service_id)
1218 return (u16)be64_to_cpu(service_id);
1221 static int cma_save_ip_info(struct sockaddr *src_addr,
1222 struct sockaddr *dst_addr,
1223 const struct ib_cm_event *ib_event,
1226 struct cma_hdr *hdr;
1229 hdr = ib_event->private_data;
1230 if (hdr->cma_version != CMA_VERSION)
1233 port = htons(cma_port_from_service_id(service_id));
1235 switch (cma_get_ip_ver(hdr)) {
1237 cma_save_ip4_info((struct sockaddr_in *)src_addr,
1238 (struct sockaddr_in *)dst_addr, hdr, port);
1241 cma_save_ip6_info((struct sockaddr_in6 *)src_addr,
1242 (struct sockaddr_in6 *)dst_addr, hdr, port);
1245 return -EAFNOSUPPORT;
1251 static int cma_save_net_info(struct sockaddr *src_addr,
1252 struct sockaddr *dst_addr,
1253 const struct rdma_cm_id *listen_id,
1254 const struct ib_cm_event *ib_event,
1255 sa_family_t sa_family, __be64 service_id)
1257 if (sa_family == AF_IB) {
1258 if (ib_event->event == IB_CM_REQ_RECEIVED)
1259 cma_save_ib_info(src_addr, dst_addr, listen_id,
1260 ib_event->param.req_rcvd.primary_path);
1261 else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED)
1262 cma_save_ib_info(src_addr, dst_addr, listen_id, NULL);
1266 return cma_save_ip_info(src_addr, dst_addr, ib_event, service_id);
1269 static int cma_save_req_info(const struct ib_cm_event *ib_event,
1270 struct cma_req_info *req)
1272 const struct ib_cm_req_event_param *req_param =
1273 &ib_event->param.req_rcvd;
1274 const struct ib_cm_sidr_req_event_param *sidr_param =
1275 &ib_event->param.sidr_req_rcvd;
1277 switch (ib_event->event) {
1278 case IB_CM_REQ_RECEIVED:
1279 req->device = req_param->listen_id->device;
1280 req->port = req_param->port;
1281 memcpy(&req->local_gid, &req_param->primary_path->sgid,
1282 sizeof(req->local_gid));
1283 req->has_gid = true;
1284 req->service_id = req_param->primary_path->service_id;
1285 req->pkey = be16_to_cpu(req_param->primary_path->pkey);
1286 if (req->pkey != req_param->bth_pkey)
1287 pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and primary path P_Key (0x%x)\n"
1288 "RDMA CMA: in the future this may cause the request to be dropped\n",
1289 req_param->bth_pkey, req->pkey);
1291 case IB_CM_SIDR_REQ_RECEIVED:
1292 req->device = sidr_param->listen_id->device;
1293 req->port = sidr_param->port;
1294 req->has_gid = false;
1295 req->service_id = sidr_param->service_id;
1296 req->pkey = sidr_param->pkey;
1297 if (req->pkey != sidr_param->bth_pkey)
1298 pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and SIDR request payload P_Key (0x%x)\n"
1299 "RDMA CMA: in the future this may cause the request to be dropped\n",
1300 sidr_param->bth_pkey, req->pkey);
1309 static bool validate_ipv4_net_dev(struct net_device *net_dev,
1310 const struct sockaddr_in *dst_addr,
1311 const struct sockaddr_in *src_addr)
1313 __be32 daddr = dst_addr->sin_addr.s_addr,
1314 saddr = src_addr->sin_addr.s_addr;
1315 struct fib_result res;
1320 if (ipv4_is_multicast(saddr) || ipv4_is_lbcast(saddr) ||
1321 ipv4_is_lbcast(daddr) || ipv4_is_zeronet(saddr) ||
1322 ipv4_is_zeronet(daddr) || ipv4_is_loopback(daddr) ||
1323 ipv4_is_loopback(saddr))
1326 memset(&fl4, 0, sizeof(fl4));
1327 fl4.flowi4_iif = net_dev->ifindex;
1332 err = fib_lookup(dev_net(net_dev), &fl4, &res, 0);
1333 ret = err == 0 && FIB_RES_DEV(res) == net_dev;
1339 static bool validate_ipv6_net_dev(struct net_device *net_dev,
1340 const struct sockaddr_in6 *dst_addr,
1341 const struct sockaddr_in6 *src_addr)
1343 #if IS_ENABLED(CONFIG_IPV6)
1344 const int strict = ipv6_addr_type(&dst_addr->sin6_addr) &
1345 IPV6_ADDR_LINKLOCAL;
1346 struct rt6_info *rt = rt6_lookup(dev_net(net_dev), &dst_addr->sin6_addr,
1347 &src_addr->sin6_addr, net_dev->ifindex,
1354 ret = rt->rt6i_idev->dev == net_dev;
1363 static bool validate_net_dev(struct net_device *net_dev,
1364 const struct sockaddr *daddr,
1365 const struct sockaddr *saddr)
1367 const struct sockaddr_in *daddr4 = (const struct sockaddr_in *)daddr;
1368 const struct sockaddr_in *saddr4 = (const struct sockaddr_in *)saddr;
1369 const struct sockaddr_in6 *daddr6 = (const struct sockaddr_in6 *)daddr;
1370 const struct sockaddr_in6 *saddr6 = (const struct sockaddr_in6 *)saddr;
1372 switch (daddr->sa_family) {
1374 return saddr->sa_family == AF_INET &&
1375 validate_ipv4_net_dev(net_dev, daddr4, saddr4);
1378 return saddr->sa_family == AF_INET6 &&
1379 validate_ipv6_net_dev(net_dev, daddr6, saddr6);
1386 static struct net_device *
1387 roce_get_net_dev_by_cm_event(const struct ib_cm_event *ib_event)
1389 const struct ib_gid_attr *sgid_attr = NULL;
1391 if (ib_event->event == IB_CM_REQ_RECEIVED)
1392 sgid_attr = ib_event->param.req_rcvd.ppath_sgid_attr;
1393 else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED)
1394 sgid_attr = ib_event->param.sidr_req_rcvd.sgid_attr;
1398 dev_hold(sgid_attr->ndev);
1399 return sgid_attr->ndev;
1402 static struct net_device *cma_get_net_dev(const struct ib_cm_event *ib_event,
1403 struct cma_req_info *req)
1405 struct sockaddr *listen_addr =
1406 (struct sockaddr *)&req->listen_addr_storage;
1407 struct sockaddr *src_addr = (struct sockaddr *)&req->src_addr_storage;
1408 struct net_device *net_dev;
1409 const union ib_gid *gid = req->has_gid ? &req->local_gid : NULL;
1412 err = cma_save_ip_info(listen_addr, src_addr, ib_event,
1415 return ERR_PTR(err);
1417 if (rdma_protocol_roce(req->device, req->port))
1418 net_dev = roce_get_net_dev_by_cm_event(ib_event);
1420 net_dev = ib_get_net_dev_by_params(req->device, req->port,
1424 return ERR_PTR(-ENODEV);
1429 static enum rdma_ucm_port_space rdma_ps_from_service_id(__be64 service_id)
1431 return (be64_to_cpu(service_id) >> 16) & 0xffff;
1434 static bool cma_match_private_data(struct rdma_id_private *id_priv,
1435 const struct cma_hdr *hdr)
1437 struct sockaddr *addr = cma_src_addr(id_priv);
1439 struct in6_addr ip6_addr;
1441 if (cma_any_addr(addr) && !id_priv->afonly)
1444 switch (addr->sa_family) {
1446 ip4_addr = ((struct sockaddr_in *)addr)->sin_addr.s_addr;
1447 if (cma_get_ip_ver(hdr) != 4)
1449 if (!cma_any_addr(addr) &&
1450 hdr->dst_addr.ip4.addr != ip4_addr)
1454 ip6_addr = ((struct sockaddr_in6 *)addr)->sin6_addr;
1455 if (cma_get_ip_ver(hdr) != 6)
1457 if (!cma_any_addr(addr) &&
1458 memcmp(&hdr->dst_addr.ip6, &ip6_addr, sizeof(ip6_addr)))
1470 static bool cma_protocol_roce(const struct rdma_cm_id *id)
1472 struct ib_device *device = id->device;
1473 const int port_num = id->port_num ?: rdma_start_port(device);
1475 return rdma_protocol_roce(device, port_num);
1478 static bool cma_match_net_dev(const struct rdma_cm_id *id,
1479 const struct net_device *net_dev,
1482 const struct rdma_addr *addr = &id->route.addr;
1485 /* This request is an AF_IB request */
1486 return (!id->port_num || id->port_num == port_num) &&
1487 (addr->src_addr.ss_family == AF_IB);
1490 * Net namespaces must match, and if the listner is listening
1491 * on a specific netdevice than netdevice must match as well.
1493 if (net_eq(dev_net(net_dev), addr->dev_addr.net) &&
1494 (!!addr->dev_addr.bound_dev_if ==
1495 (addr->dev_addr.bound_dev_if == net_dev->ifindex)))
1501 static struct rdma_id_private *cma_find_listener(
1502 const struct rdma_bind_list *bind_list,
1503 const struct ib_cm_id *cm_id,
1504 const struct ib_cm_event *ib_event,
1505 const struct cma_req_info *req,
1506 const struct net_device *net_dev)
1508 struct rdma_id_private *id_priv, *id_priv_dev;
1510 lockdep_assert_held(&lock);
1513 return ERR_PTR(-EINVAL);
1515 hlist_for_each_entry(id_priv, &bind_list->owners, node) {
1516 if (cma_match_private_data(id_priv, ib_event->private_data)) {
1517 if (id_priv->id.device == cm_id->device &&
1518 cma_match_net_dev(&id_priv->id, net_dev, req->port))
1520 list_for_each_entry(id_priv_dev,
1521 &id_priv->listen_list,
1523 if (id_priv_dev->id.device == cm_id->device &&
1524 cma_match_net_dev(&id_priv_dev->id, net_dev, req->port))
1530 return ERR_PTR(-EINVAL);
1533 static struct rdma_id_private *
1534 cma_ib_id_from_event(struct ib_cm_id *cm_id,
1535 const struct ib_cm_event *ib_event,
1536 struct net_device **net_dev)
1538 struct cma_req_info req;
1539 struct rdma_bind_list *bind_list;
1540 struct rdma_id_private *id_priv;
1543 err = cma_save_req_info(ib_event, &req);
1545 return ERR_PTR(err);
1547 *net_dev = cma_get_net_dev(ib_event, &req);
1548 if (IS_ERR(*net_dev)) {
1549 if (PTR_ERR(*net_dev) == -EAFNOSUPPORT) {
1550 /* Assuming the protocol is AF_IB */
1553 return ERR_CAST(*net_dev);
1559 * Net namespace might be getting deleted while route lookup,
1560 * cm_id lookup is in progress. Therefore, perform netdevice
1561 * validation, cm_id lookup under rcu lock.
1562 * RCU lock along with netdevice state check, synchronizes with
1563 * netdevice migrating to different net namespace and also avoids
1564 * case where net namespace doesn't get deleted while lookup is in
1566 * If the device state is not IFF_UP, its properties such as ifindex
1567 * and nd_net cannot be trusted to remain valid without rcu lock.
1568 * net/core/dev.c change_net_namespace() ensures to synchronize with
1569 * ongoing operations on net device after device is closed using
1570 * synchronize_net().
1575 * If netdevice is down, it is likely that it is administratively
1576 * down or it might be migrating to different namespace.
1577 * In that case avoid further processing, as the net namespace
1578 * or ifindex may change.
1580 if (((*net_dev)->flags & IFF_UP) == 0) {
1581 id_priv = ERR_PTR(-EHOSTUNREACH);
1585 if (!validate_net_dev(*net_dev,
1586 (struct sockaddr *)&req.listen_addr_storage,
1587 (struct sockaddr *)&req.src_addr_storage)) {
1588 id_priv = ERR_PTR(-EHOSTUNREACH);
1593 bind_list = cma_ps_find(*net_dev ? dev_net(*net_dev) : &init_net,
1594 rdma_ps_from_service_id(req.service_id),
1595 cma_port_from_service_id(req.service_id));
1596 id_priv = cma_find_listener(bind_list, cm_id, ib_event, &req, *net_dev);
1599 mutex_unlock(&lock);
1600 if (IS_ERR(id_priv) && *net_dev) {
1607 static inline u8 cma_user_data_offset(struct rdma_id_private *id_priv)
1609 return cma_family(id_priv) == AF_IB ? 0 : sizeof(struct cma_hdr);
1612 static void cma_cancel_route(struct rdma_id_private *id_priv)
1614 if (rdma_cap_ib_sa(id_priv->id.device, id_priv->id.port_num)) {
1616 ib_sa_cancel_query(id_priv->query_id, id_priv->query);
1620 static void cma_cancel_listens(struct rdma_id_private *id_priv)
1622 struct rdma_id_private *dev_id_priv;
1625 * Remove from listen_any_list to prevent added devices from spawning
1626 * additional listen requests.
1629 list_del(&id_priv->list);
1631 while (!list_empty(&id_priv->listen_list)) {
1632 dev_id_priv = list_entry(id_priv->listen_list.next,
1633 struct rdma_id_private, listen_list);
1634 /* sync with device removal to avoid duplicate destruction */
1635 list_del_init(&dev_id_priv->list);
1636 list_del(&dev_id_priv->listen_list);
1637 mutex_unlock(&lock);
1639 rdma_destroy_id(&dev_id_priv->id);
1642 mutex_unlock(&lock);
1645 static void cma_cancel_operation(struct rdma_id_private *id_priv,
1646 enum rdma_cm_state state)
1649 case RDMA_CM_ADDR_QUERY:
1650 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
1652 case RDMA_CM_ROUTE_QUERY:
1653 cma_cancel_route(id_priv);
1655 case RDMA_CM_LISTEN:
1656 if (cma_any_addr(cma_src_addr(id_priv)) && !id_priv->cma_dev)
1657 cma_cancel_listens(id_priv);
1664 static void cma_release_port(struct rdma_id_private *id_priv)
1666 struct rdma_bind_list *bind_list = id_priv->bind_list;
1667 struct net *net = id_priv->id.route.addr.dev_addr.net;
1673 hlist_del(&id_priv->node);
1674 if (hlist_empty(&bind_list->owners)) {
1675 cma_ps_remove(net, bind_list->ps, bind_list->port);
1678 mutex_unlock(&lock);
1681 static void destroy_mc(struct rdma_id_private *id_priv,
1682 struct cma_multicast *mc)
1684 if (rdma_cap_ib_mcast(id_priv->id.device, id_priv->id.port_num)) {
1685 ib_sa_free_multicast(mc->multicast.ib);
1690 if (rdma_protocol_roce(id_priv->id.device,
1691 id_priv->id.port_num)) {
1692 struct rdma_dev_addr *dev_addr =
1693 &id_priv->id.route.addr.dev_addr;
1694 struct net_device *ndev = NULL;
1696 if (dev_addr->bound_dev_if)
1697 ndev = dev_get_by_index(dev_addr->net,
1698 dev_addr->bound_dev_if);
1700 cma_igmp_send(ndev, &mc->multicast.ib->rec.mgid, false);
1703 kref_put(&mc->mcref, release_mc);
1707 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
1709 struct cma_multicast *mc;
1711 while (!list_empty(&id_priv->mc_list)) {
1712 mc = list_first_entry(&id_priv->mc_list, struct cma_multicast,
1714 list_del(&mc->list);
1715 destroy_mc(id_priv, mc);
1719 void rdma_destroy_id(struct rdma_cm_id *id)
1721 struct rdma_id_private *id_priv;
1722 enum rdma_cm_state state;
1724 id_priv = container_of(id, struct rdma_id_private, id);
1725 state = cma_exch(id_priv, RDMA_CM_DESTROYING);
1726 cma_cancel_operation(id_priv, state);
1729 * Wait for any active callback to finish. New callbacks will find
1730 * the id_priv state set to destroying and abort.
1732 mutex_lock(&id_priv->handler_mutex);
1733 mutex_unlock(&id_priv->handler_mutex);
1735 rdma_restrack_del(&id_priv->res);
1736 if (id_priv->cma_dev) {
1737 if (rdma_cap_ib_cm(id_priv->id.device, 1)) {
1738 if (id_priv->cm_id.ib)
1739 ib_destroy_cm_id(id_priv->cm_id.ib);
1740 } else if (rdma_cap_iw_cm(id_priv->id.device, 1)) {
1741 if (id_priv->cm_id.iw)
1742 iw_destroy_cm_id(id_priv->cm_id.iw);
1744 cma_leave_mc_groups(id_priv);
1745 cma_release_dev(id_priv);
1748 cma_release_port(id_priv);
1749 cma_deref_id(id_priv);
1750 wait_for_completion(&id_priv->comp);
1752 if (id_priv->internal_id)
1753 cma_deref_id(id_priv->id.context);
1755 kfree(id_priv->id.route.path_rec);
1757 if (id_priv->id.route.addr.dev_addr.sgid_attr)
1758 rdma_put_gid_attr(id_priv->id.route.addr.dev_addr.sgid_attr);
1760 put_net(id_priv->id.route.addr.dev_addr.net);
1763 EXPORT_SYMBOL(rdma_destroy_id);
1765 static int cma_rep_recv(struct rdma_id_private *id_priv)
1769 ret = cma_modify_qp_rtr(id_priv, NULL);
1773 ret = cma_modify_qp_rts(id_priv, NULL);
1777 ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
1783 pr_debug_ratelimited("RDMA CM: CONNECT_ERROR: failed to handle reply. status %d\n", ret);
1784 cma_modify_qp_err(id_priv);
1785 ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
1790 static void cma_set_rep_event_data(struct rdma_cm_event *event,
1791 const struct ib_cm_rep_event_param *rep_data,
1794 event->param.conn.private_data = private_data;
1795 event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
1796 event->param.conn.responder_resources = rep_data->responder_resources;
1797 event->param.conn.initiator_depth = rep_data->initiator_depth;
1798 event->param.conn.flow_control = rep_data->flow_control;
1799 event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
1800 event->param.conn.srq = rep_data->srq;
1801 event->param.conn.qp_num = rep_data->remote_qpn;
1804 static int cma_ib_handler(struct ib_cm_id *cm_id,
1805 const struct ib_cm_event *ib_event)
1807 struct rdma_id_private *id_priv = cm_id->context;
1808 struct rdma_cm_event event = {};
1811 mutex_lock(&id_priv->handler_mutex);
1812 if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
1813 id_priv->state != RDMA_CM_CONNECT) ||
1814 (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
1815 id_priv->state != RDMA_CM_DISCONNECT))
1818 switch (ib_event->event) {
1819 case IB_CM_REQ_ERROR:
1820 case IB_CM_REP_ERROR:
1821 event.event = RDMA_CM_EVENT_UNREACHABLE;
1822 event.status = -ETIMEDOUT;
1824 case IB_CM_REP_RECEIVED:
1825 if (cma_comp(id_priv, RDMA_CM_CONNECT) &&
1826 (id_priv->id.qp_type != IB_QPT_UD))
1827 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
1828 if (id_priv->id.qp) {
1829 event.status = cma_rep_recv(id_priv);
1830 event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
1831 RDMA_CM_EVENT_ESTABLISHED;
1833 event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
1835 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
1836 ib_event->private_data);
1838 case IB_CM_RTU_RECEIVED:
1839 case IB_CM_USER_ESTABLISHED:
1840 event.event = RDMA_CM_EVENT_ESTABLISHED;
1842 case IB_CM_DREQ_ERROR:
1843 event.status = -ETIMEDOUT; /* fall through */
1844 case IB_CM_DREQ_RECEIVED:
1845 case IB_CM_DREP_RECEIVED:
1846 if (!cma_comp_exch(id_priv, RDMA_CM_CONNECT,
1847 RDMA_CM_DISCONNECT))
1849 event.event = RDMA_CM_EVENT_DISCONNECTED;
1851 case IB_CM_TIMEWAIT_EXIT:
1852 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
1854 case IB_CM_MRA_RECEIVED:
1857 case IB_CM_REJ_RECEIVED:
1858 pr_debug_ratelimited("RDMA CM: REJECTED: %s\n", rdma_reject_msg(&id_priv->id,
1859 ib_event->param.rej_rcvd.reason));
1860 cma_modify_qp_err(id_priv);
1861 event.status = ib_event->param.rej_rcvd.reason;
1862 event.event = RDMA_CM_EVENT_REJECTED;
1863 event.param.conn.private_data = ib_event->private_data;
1864 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
1867 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
1872 ret = id_priv->id.event_handler(&id_priv->id, &event);
1874 /* Destroy the CM ID by returning a non-zero value. */
1875 id_priv->cm_id.ib = NULL;
1876 cma_exch(id_priv, RDMA_CM_DESTROYING);
1877 mutex_unlock(&id_priv->handler_mutex);
1878 rdma_destroy_id(&id_priv->id);
1882 mutex_unlock(&id_priv->handler_mutex);
1886 static struct rdma_id_private *
1887 cma_ib_new_conn_id(const struct rdma_cm_id *listen_id,
1888 const struct ib_cm_event *ib_event,
1889 struct net_device *net_dev)
1891 struct rdma_id_private *listen_id_priv;
1892 struct rdma_id_private *id_priv;
1893 struct rdma_cm_id *id;
1894 struct rdma_route *rt;
1895 const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
1896 struct sa_path_rec *path = ib_event->param.req_rcvd.primary_path;
1897 const __be64 service_id =
1898 ib_event->param.req_rcvd.primary_path->service_id;
1901 listen_id_priv = container_of(listen_id, struct rdma_id_private, id);
1902 id = __rdma_create_id(listen_id->route.addr.dev_addr.net,
1903 listen_id->event_handler, listen_id->context,
1904 listen_id->ps, ib_event->param.req_rcvd.qp_type,
1905 listen_id_priv->res.kern_name);
1909 id_priv = container_of(id, struct rdma_id_private, id);
1910 if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
1911 (struct sockaddr *)&id->route.addr.dst_addr,
1912 listen_id, ib_event, ss_family, service_id))
1916 rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
1917 rt->path_rec = kmalloc_array(rt->num_paths, sizeof(*rt->path_rec),
1922 rt->path_rec[0] = *path;
1923 if (rt->num_paths == 2)
1924 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
1927 rdma_copy_addr(&rt->addr.dev_addr, net_dev, NULL);
1929 if (!cma_protocol_roce(listen_id) &&
1930 cma_any_addr(cma_src_addr(id_priv))) {
1931 rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
1932 rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
1933 ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
1934 } else if (!cma_any_addr(cma_src_addr(id_priv))) {
1935 ret = cma_translate_addr(cma_src_addr(id_priv), &rt->addr.dev_addr);
1940 rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
1942 id_priv->state = RDMA_CM_CONNECT;
1946 rdma_destroy_id(id);
1950 static struct rdma_id_private *
1951 cma_ib_new_udp_id(const struct rdma_cm_id *listen_id,
1952 const struct ib_cm_event *ib_event,
1953 struct net_device *net_dev)
1955 const struct rdma_id_private *listen_id_priv;
1956 struct rdma_id_private *id_priv;
1957 struct rdma_cm_id *id;
1958 const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
1959 struct net *net = listen_id->route.addr.dev_addr.net;
1962 listen_id_priv = container_of(listen_id, struct rdma_id_private, id);
1963 id = __rdma_create_id(net, listen_id->event_handler, listen_id->context,
1964 listen_id->ps, IB_QPT_UD,
1965 listen_id_priv->res.kern_name);
1969 id_priv = container_of(id, struct rdma_id_private, id);
1970 if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
1971 (struct sockaddr *)&id->route.addr.dst_addr,
1972 listen_id, ib_event, ss_family,
1973 ib_event->param.sidr_req_rcvd.service_id))
1977 rdma_copy_addr(&id->route.addr.dev_addr, net_dev, NULL);
1979 if (!cma_any_addr(cma_src_addr(id_priv))) {
1980 ret = cma_translate_addr(cma_src_addr(id_priv),
1981 &id->route.addr.dev_addr);
1987 id_priv->state = RDMA_CM_CONNECT;
1990 rdma_destroy_id(id);
1994 static void cma_set_req_event_data(struct rdma_cm_event *event,
1995 const struct ib_cm_req_event_param *req_data,
1996 void *private_data, int offset)
1998 event->param.conn.private_data = private_data + offset;
1999 event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
2000 event->param.conn.responder_resources = req_data->responder_resources;
2001 event->param.conn.initiator_depth = req_data->initiator_depth;
2002 event->param.conn.flow_control = req_data->flow_control;
2003 event->param.conn.retry_count = req_data->retry_count;
2004 event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
2005 event->param.conn.srq = req_data->srq;
2006 event->param.conn.qp_num = req_data->remote_qpn;
2009 static int cma_ib_check_req_qp_type(const struct rdma_cm_id *id,
2010 const struct ib_cm_event *ib_event)
2012 return (((ib_event->event == IB_CM_REQ_RECEIVED) &&
2013 (ib_event->param.req_rcvd.qp_type == id->qp_type)) ||
2014 ((ib_event->event == IB_CM_SIDR_REQ_RECEIVED) &&
2015 (id->qp_type == IB_QPT_UD)) ||
2019 static int cma_ib_req_handler(struct ib_cm_id *cm_id,
2020 const struct ib_cm_event *ib_event)
2022 struct rdma_id_private *listen_id, *conn_id = NULL;
2023 struct rdma_cm_event event = {};
2024 struct net_device *net_dev;
2028 listen_id = cma_ib_id_from_event(cm_id, ib_event, &net_dev);
2029 if (IS_ERR(listen_id))
2030 return PTR_ERR(listen_id);
2032 if (!cma_ib_check_req_qp_type(&listen_id->id, ib_event)) {
2037 mutex_lock(&listen_id->handler_mutex);
2038 if (listen_id->state != RDMA_CM_LISTEN) {
2039 ret = -ECONNABORTED;
2043 offset = cma_user_data_offset(listen_id);
2044 event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
2045 if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) {
2046 conn_id = cma_ib_new_udp_id(&listen_id->id, ib_event, net_dev);
2047 event.param.ud.private_data = ib_event->private_data + offset;
2048 event.param.ud.private_data_len =
2049 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
2051 conn_id = cma_ib_new_conn_id(&listen_id->id, ib_event, net_dev);
2052 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
2053 ib_event->private_data, offset);
2060 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
2061 ret = cma_acquire_dev(conn_id, listen_id);
2065 conn_id->cm_id.ib = cm_id;
2066 cm_id->context = conn_id;
2067 cm_id->cm_handler = cma_ib_handler;
2070 * Protect against the user destroying conn_id from another thread
2071 * until we're done accessing it.
2073 atomic_inc(&conn_id->refcount);
2074 ret = conn_id->id.event_handler(&conn_id->id, &event);
2078 * Acquire mutex to prevent user executing rdma_destroy_id()
2079 * while we're accessing the cm_id.
2082 if (cma_comp(conn_id, RDMA_CM_CONNECT) &&
2083 (conn_id->id.qp_type != IB_QPT_UD))
2084 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
2085 mutex_unlock(&lock);
2086 mutex_unlock(&conn_id->handler_mutex);
2087 mutex_unlock(&listen_id->handler_mutex);
2088 cma_deref_id(conn_id);
2094 cma_deref_id(conn_id);
2095 /* Destroy the CM ID by returning a non-zero value. */
2096 conn_id->cm_id.ib = NULL;
2098 cma_exch(conn_id, RDMA_CM_DESTROYING);
2099 mutex_unlock(&conn_id->handler_mutex);
2101 mutex_unlock(&listen_id->handler_mutex);
2103 rdma_destroy_id(&conn_id->id);
2112 __be64 rdma_get_service_id(struct rdma_cm_id *id, struct sockaddr *addr)
2114 if (addr->sa_family == AF_IB)
2115 return ((struct sockaddr_ib *) addr)->sib_sid;
2117 return cpu_to_be64(((u64)id->ps << 16) + be16_to_cpu(cma_port(addr)));
2119 EXPORT_SYMBOL(rdma_get_service_id);
2121 void rdma_read_gids(struct rdma_cm_id *cm_id, union ib_gid *sgid,
2124 struct rdma_addr *addr = &cm_id->route.addr;
2126 if (!cm_id->device) {
2128 memset(sgid, 0, sizeof(*sgid));
2130 memset(dgid, 0, sizeof(*dgid));
2134 if (rdma_protocol_roce(cm_id->device, cm_id->port_num)) {
2136 rdma_ip2gid((struct sockaddr *)&addr->src_addr, sgid);
2138 rdma_ip2gid((struct sockaddr *)&addr->dst_addr, dgid);
2141 rdma_addr_get_sgid(&addr->dev_addr, sgid);
2143 rdma_addr_get_dgid(&addr->dev_addr, dgid);
2146 EXPORT_SYMBOL(rdma_read_gids);
2148 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
2150 struct rdma_id_private *id_priv = iw_id->context;
2151 struct rdma_cm_event event = {};
2153 struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
2154 struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
2156 mutex_lock(&id_priv->handler_mutex);
2157 if (id_priv->state != RDMA_CM_CONNECT)
2160 switch (iw_event->event) {
2161 case IW_CM_EVENT_CLOSE:
2162 event.event = RDMA_CM_EVENT_DISCONNECTED;
2164 case IW_CM_EVENT_CONNECT_REPLY:
2165 memcpy(cma_src_addr(id_priv), laddr,
2166 rdma_addr_size(laddr));
2167 memcpy(cma_dst_addr(id_priv), raddr,
2168 rdma_addr_size(raddr));
2169 switch (iw_event->status) {
2171 event.event = RDMA_CM_EVENT_ESTABLISHED;
2172 event.param.conn.initiator_depth = iw_event->ird;
2173 event.param.conn.responder_resources = iw_event->ord;
2177 event.event = RDMA_CM_EVENT_REJECTED;
2180 event.event = RDMA_CM_EVENT_UNREACHABLE;
2183 event.event = RDMA_CM_EVENT_CONNECT_ERROR;
2187 case IW_CM_EVENT_ESTABLISHED:
2188 event.event = RDMA_CM_EVENT_ESTABLISHED;
2189 event.param.conn.initiator_depth = iw_event->ird;
2190 event.param.conn.responder_resources = iw_event->ord;
2196 event.status = iw_event->status;
2197 event.param.conn.private_data = iw_event->private_data;
2198 event.param.conn.private_data_len = iw_event->private_data_len;
2199 ret = id_priv->id.event_handler(&id_priv->id, &event);
2201 /* Destroy the CM ID by returning a non-zero value. */
2202 id_priv->cm_id.iw = NULL;
2203 cma_exch(id_priv, RDMA_CM_DESTROYING);
2204 mutex_unlock(&id_priv->handler_mutex);
2205 rdma_destroy_id(&id_priv->id);
2210 mutex_unlock(&id_priv->handler_mutex);
2214 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
2215 struct iw_cm_event *iw_event)
2217 struct rdma_cm_id *new_cm_id;
2218 struct rdma_id_private *listen_id, *conn_id;
2219 struct rdma_cm_event event = {};
2220 int ret = -ECONNABORTED;
2221 struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
2222 struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
2224 event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
2225 event.param.conn.private_data = iw_event->private_data;
2226 event.param.conn.private_data_len = iw_event->private_data_len;
2227 event.param.conn.initiator_depth = iw_event->ird;
2228 event.param.conn.responder_resources = iw_event->ord;
2230 listen_id = cm_id->context;
2232 mutex_lock(&listen_id->handler_mutex);
2233 if (listen_id->state != RDMA_CM_LISTEN)
2236 /* Create a new RDMA id for the new IW CM ID */
2237 new_cm_id = __rdma_create_id(listen_id->id.route.addr.dev_addr.net,
2238 listen_id->id.event_handler,
2239 listen_id->id.context,
2240 RDMA_PS_TCP, IB_QPT_RC,
2241 listen_id->res.kern_name);
2242 if (IS_ERR(new_cm_id)) {
2246 conn_id = container_of(new_cm_id, struct rdma_id_private, id);
2247 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
2248 conn_id->state = RDMA_CM_CONNECT;
2250 ret = rdma_translate_ip(laddr, &conn_id->id.route.addr.dev_addr);
2252 mutex_unlock(&conn_id->handler_mutex);
2253 rdma_destroy_id(new_cm_id);
2257 ret = cma_acquire_dev(conn_id, listen_id);
2259 mutex_unlock(&conn_id->handler_mutex);
2260 rdma_destroy_id(new_cm_id);
2264 conn_id->cm_id.iw = cm_id;
2265 cm_id->context = conn_id;
2266 cm_id->cm_handler = cma_iw_handler;
2268 memcpy(cma_src_addr(conn_id), laddr, rdma_addr_size(laddr));
2269 memcpy(cma_dst_addr(conn_id), raddr, rdma_addr_size(raddr));
2272 * Protect against the user destroying conn_id from another thread
2273 * until we're done accessing it.
2275 atomic_inc(&conn_id->refcount);
2276 ret = conn_id->id.event_handler(&conn_id->id, &event);
2278 /* User wants to destroy the CM ID */
2279 conn_id->cm_id.iw = NULL;
2280 cma_exch(conn_id, RDMA_CM_DESTROYING);
2281 mutex_unlock(&conn_id->handler_mutex);
2282 mutex_unlock(&listen_id->handler_mutex);
2283 cma_deref_id(conn_id);
2284 rdma_destroy_id(&conn_id->id);
2288 mutex_unlock(&conn_id->handler_mutex);
2289 cma_deref_id(conn_id);
2292 mutex_unlock(&listen_id->handler_mutex);
2296 static int cma_ib_listen(struct rdma_id_private *id_priv)
2298 struct sockaddr *addr;
2299 struct ib_cm_id *id;
2302 addr = cma_src_addr(id_priv);
2303 svc_id = rdma_get_service_id(&id_priv->id, addr);
2304 id = ib_cm_insert_listen(id_priv->id.device,
2305 cma_ib_req_handler, svc_id);
2308 id_priv->cm_id.ib = id;
2313 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
2316 struct iw_cm_id *id;
2318 id = iw_create_cm_id(id_priv->id.device,
2319 iw_conn_req_handler,
2324 id->tos = id_priv->tos;
2325 id_priv->cm_id.iw = id;
2327 memcpy(&id_priv->cm_id.iw->local_addr, cma_src_addr(id_priv),
2328 rdma_addr_size(cma_src_addr(id_priv)));
2330 ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
2333 iw_destroy_cm_id(id_priv->cm_id.iw);
2334 id_priv->cm_id.iw = NULL;
2340 static int cma_listen_handler(struct rdma_cm_id *id,
2341 struct rdma_cm_event *event)
2343 struct rdma_id_private *id_priv = id->context;
2345 id->context = id_priv->id.context;
2346 id->event_handler = id_priv->id.event_handler;
2347 return id_priv->id.event_handler(id, event);
2350 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
2351 struct cma_device *cma_dev)
2353 struct rdma_id_private *dev_id_priv;
2354 struct rdma_cm_id *id;
2355 struct net *net = id_priv->id.route.addr.dev_addr.net;
2358 lockdep_assert_held(&lock);
2360 if (cma_family(id_priv) == AF_IB && !rdma_cap_ib_cm(cma_dev->device, 1))
2363 id = __rdma_create_id(net, cma_listen_handler, id_priv, id_priv->id.ps,
2364 id_priv->id.qp_type, id_priv->res.kern_name);
2368 dev_id_priv = container_of(id, struct rdma_id_private, id);
2370 dev_id_priv->state = RDMA_CM_ADDR_BOUND;
2371 memcpy(cma_src_addr(dev_id_priv), cma_src_addr(id_priv),
2372 rdma_addr_size(cma_src_addr(id_priv)));
2374 _cma_attach_to_dev(dev_id_priv, cma_dev);
2375 list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
2376 atomic_inc(&id_priv->refcount);
2377 dev_id_priv->internal_id = 1;
2378 dev_id_priv->afonly = id_priv->afonly;
2380 ret = rdma_listen(id, id_priv->backlog);
2382 pr_warn("RDMA CMA: cma_listen_on_dev, error %d, listening on device %s\n",
2383 ret, cma_dev->device->name);
2386 static void cma_listen_on_all(struct rdma_id_private *id_priv)
2388 struct cma_device *cma_dev;
2391 list_add_tail(&id_priv->list, &listen_any_list);
2392 list_for_each_entry(cma_dev, &dev_list, list)
2393 cma_listen_on_dev(id_priv, cma_dev);
2394 mutex_unlock(&lock);
2397 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
2399 struct rdma_id_private *id_priv;
2401 id_priv = container_of(id, struct rdma_id_private, id);
2402 id_priv->tos = (u8) tos;
2403 id_priv->tos_set = true;
2405 EXPORT_SYMBOL(rdma_set_service_type);
2407 static void cma_query_handler(int status, struct sa_path_rec *path_rec,
2410 struct cma_work *work = context;
2411 struct rdma_route *route;
2413 route = &work->id->id.route;
2416 route->num_paths = 1;
2417 *route->path_rec = *path_rec;
2419 work->old_state = RDMA_CM_ROUTE_QUERY;
2420 work->new_state = RDMA_CM_ADDR_RESOLVED;
2421 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
2422 work->event.status = status;
2423 pr_debug_ratelimited("RDMA CM: ROUTE_ERROR: failed to query path. status %d\n",
2427 queue_work(cma_wq, &work->work);
2430 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms,
2431 struct cma_work *work)
2433 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2434 struct sa_path_rec path_rec;
2435 ib_sa_comp_mask comp_mask;
2436 struct sockaddr_in6 *sin6;
2437 struct sockaddr_ib *sib;
2439 memset(&path_rec, 0, sizeof path_rec);
2441 if (rdma_cap_opa_ah(id_priv->id.device, id_priv->id.port_num))
2442 path_rec.rec_type = SA_PATH_REC_TYPE_OPA;
2444 path_rec.rec_type = SA_PATH_REC_TYPE_IB;
2445 rdma_addr_get_sgid(dev_addr, &path_rec.sgid);
2446 rdma_addr_get_dgid(dev_addr, &path_rec.dgid);
2447 path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
2448 path_rec.numb_path = 1;
2449 path_rec.reversible = 1;
2450 path_rec.service_id = rdma_get_service_id(&id_priv->id,
2451 cma_dst_addr(id_priv));
2453 comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
2454 IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
2455 IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
2457 switch (cma_family(id_priv)) {
2459 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
2460 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
2463 sin6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
2464 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
2465 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2468 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
2469 path_rec.traffic_class = (u8) (be32_to_cpu(sib->sib_flowinfo) >> 20);
2470 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2474 id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
2475 id_priv->id.port_num, &path_rec,
2476 comp_mask, timeout_ms,
2477 GFP_KERNEL, cma_query_handler,
2478 work, &id_priv->query);
2480 return (id_priv->query_id < 0) ? id_priv->query_id : 0;
2483 static void cma_work_handler(struct work_struct *_work)
2485 struct cma_work *work = container_of(_work, struct cma_work, work);
2486 struct rdma_id_private *id_priv = work->id;
2489 mutex_lock(&id_priv->handler_mutex);
2490 if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
2493 if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
2494 cma_exch(id_priv, RDMA_CM_DESTROYING);
2498 mutex_unlock(&id_priv->handler_mutex);
2499 cma_deref_id(id_priv);
2501 rdma_destroy_id(&id_priv->id);
2505 static void cma_ndev_work_handler(struct work_struct *_work)
2507 struct cma_ndev_work *work = container_of(_work, struct cma_ndev_work, work);
2508 struct rdma_id_private *id_priv = work->id;
2511 mutex_lock(&id_priv->handler_mutex);
2512 if (id_priv->state == RDMA_CM_DESTROYING ||
2513 id_priv->state == RDMA_CM_DEVICE_REMOVAL)
2516 if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
2517 cma_exch(id_priv, RDMA_CM_DESTROYING);
2522 mutex_unlock(&id_priv->handler_mutex);
2523 cma_deref_id(id_priv);
2525 rdma_destroy_id(&id_priv->id);
2529 static void cma_init_resolve_route_work(struct cma_work *work,
2530 struct rdma_id_private *id_priv)
2533 INIT_WORK(&work->work, cma_work_handler);
2534 work->old_state = RDMA_CM_ROUTE_QUERY;
2535 work->new_state = RDMA_CM_ROUTE_RESOLVED;
2536 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
2539 static void cma_init_resolve_addr_work(struct cma_work *work,
2540 struct rdma_id_private *id_priv)
2543 INIT_WORK(&work->work, cma_work_handler);
2544 work->old_state = RDMA_CM_ADDR_QUERY;
2545 work->new_state = RDMA_CM_ADDR_RESOLVED;
2546 work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2549 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms)
2551 struct rdma_route *route = &id_priv->id.route;
2552 struct cma_work *work;
2555 work = kzalloc(sizeof *work, GFP_KERNEL);
2559 cma_init_resolve_route_work(work, id_priv);
2561 if (!route->path_rec)
2562 route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
2563 if (!route->path_rec) {
2568 ret = cma_query_ib_route(id_priv, timeout_ms, work);
2574 kfree(route->path_rec);
2575 route->path_rec = NULL;
2581 static enum ib_gid_type cma_route_gid_type(enum rdma_network_type network_type,
2582 unsigned long supported_gids,
2583 enum ib_gid_type default_gid)
2585 if ((network_type == RDMA_NETWORK_IPV4 ||
2586 network_type == RDMA_NETWORK_IPV6) &&
2587 test_bit(IB_GID_TYPE_ROCE_UDP_ENCAP, &supported_gids))
2588 return IB_GID_TYPE_ROCE_UDP_ENCAP;
2594 * cma_iboe_set_path_rec_l2_fields() is helper function which sets
2595 * path record type based on GID type.
2596 * It also sets up other L2 fields which includes destination mac address
2597 * netdev ifindex, of the path record.
2598 * It returns the netdev of the bound interface for this path record entry.
2600 static struct net_device *
2601 cma_iboe_set_path_rec_l2_fields(struct rdma_id_private *id_priv)
2603 struct rdma_route *route = &id_priv->id.route;
2604 enum ib_gid_type gid_type = IB_GID_TYPE_ROCE;
2605 struct rdma_addr *addr = &route->addr;
2606 unsigned long supported_gids;
2607 struct net_device *ndev;
2609 if (!addr->dev_addr.bound_dev_if)
2612 ndev = dev_get_by_index(addr->dev_addr.net,
2613 addr->dev_addr.bound_dev_if);
2617 supported_gids = roce_gid_type_mask_support(id_priv->id.device,
2618 id_priv->id.port_num);
2619 gid_type = cma_route_gid_type(addr->dev_addr.network,
2622 /* Use the hint from IP Stack to select GID Type */
2623 if (gid_type < ib_network_to_gid_type(addr->dev_addr.network))
2624 gid_type = ib_network_to_gid_type(addr->dev_addr.network);
2625 route->path_rec->rec_type = sa_conv_gid_to_pathrec_type(gid_type);
2627 route->path_rec->roce.route_resolved = true;
2628 sa_path_set_dmac(route->path_rec, addr->dev_addr.dst_dev_addr);
2632 int rdma_set_ib_path(struct rdma_cm_id *id,
2633 struct sa_path_rec *path_rec)
2635 struct rdma_id_private *id_priv;
2636 struct net_device *ndev;
2639 id_priv = container_of(id, struct rdma_id_private, id);
2640 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
2641 RDMA_CM_ROUTE_RESOLVED))
2644 id->route.path_rec = kmemdup(path_rec, sizeof(*path_rec),
2646 if (!id->route.path_rec) {
2651 if (rdma_protocol_roce(id->device, id->port_num)) {
2652 ndev = cma_iboe_set_path_rec_l2_fields(id_priv);
2660 id->route.num_paths = 1;
2664 kfree(id->route.path_rec);
2665 id->route.path_rec = NULL;
2667 cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED);
2670 EXPORT_SYMBOL(rdma_set_ib_path);
2672 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms)
2674 struct cma_work *work;
2676 work = kzalloc(sizeof *work, GFP_KERNEL);
2680 cma_init_resolve_route_work(work, id_priv);
2681 queue_work(cma_wq, &work->work);
2685 static int iboe_tos_to_sl(struct net_device *ndev, int tos)
2688 struct net_device *dev;
2690 prio = rt_tos2priority(tos);
2691 dev = is_vlan_dev(ndev) ? vlan_dev_real_dev(ndev) : ndev;
2693 return netdev_get_prio_tc_map(dev, prio);
2695 #if IS_ENABLED(CONFIG_VLAN_8021Q)
2696 if (is_vlan_dev(ndev))
2697 return (vlan_dev_get_egress_qos_mask(ndev, prio) &
2698 VLAN_PRIO_MASK) >> VLAN_PRIO_SHIFT;
2703 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
2705 struct rdma_route *route = &id_priv->id.route;
2706 struct rdma_addr *addr = &route->addr;
2707 struct cma_work *work;
2709 struct net_device *ndev;
2711 u8 default_roce_tos = id_priv->cma_dev->default_roce_tos[id_priv->id.port_num -
2712 rdma_start_port(id_priv->cma_dev->device)];
2713 u8 tos = id_priv->tos_set ? id_priv->tos : default_roce_tos;
2716 work = kzalloc(sizeof *work, GFP_KERNEL);
2720 route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
2721 if (!route->path_rec) {
2726 route->num_paths = 1;
2728 ndev = cma_iboe_set_path_rec_l2_fields(id_priv);
2734 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
2735 &route->path_rec->sgid);
2736 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.dst_addr,
2737 &route->path_rec->dgid);
2739 if (((struct sockaddr *)&id_priv->id.route.addr.dst_addr)->sa_family != AF_IB)
2740 /* TODO: get the hoplimit from the inet/inet6 device */
2741 route->path_rec->hop_limit = addr->dev_addr.hoplimit;
2743 route->path_rec->hop_limit = 1;
2744 route->path_rec->reversible = 1;
2745 route->path_rec->pkey = cpu_to_be16(0xffff);
2746 route->path_rec->mtu_selector = IB_SA_EQ;
2747 route->path_rec->sl = iboe_tos_to_sl(ndev, tos);
2748 route->path_rec->traffic_class = tos;
2749 route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
2750 route->path_rec->rate_selector = IB_SA_EQ;
2751 route->path_rec->rate = iboe_get_rate(ndev);
2753 route->path_rec->packet_life_time_selector = IB_SA_EQ;
2754 route->path_rec->packet_life_time = CMA_IBOE_PACKET_LIFETIME;
2755 if (!route->path_rec->mtu) {
2760 cma_init_resolve_route_work(work, id_priv);
2761 queue_work(cma_wq, &work->work);
2766 kfree(route->path_rec);
2767 route->path_rec = NULL;
2768 route->num_paths = 0;
2774 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
2776 struct rdma_id_private *id_priv;
2779 id_priv = container_of(id, struct rdma_id_private, id);
2780 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY))
2783 atomic_inc(&id_priv->refcount);
2784 if (rdma_cap_ib_sa(id->device, id->port_num))
2785 ret = cma_resolve_ib_route(id_priv, timeout_ms);
2786 else if (rdma_protocol_roce(id->device, id->port_num))
2787 ret = cma_resolve_iboe_route(id_priv);
2788 else if (rdma_protocol_iwarp(id->device, id->port_num))
2789 ret = cma_resolve_iw_route(id_priv, timeout_ms);
2798 cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED);
2799 cma_deref_id(id_priv);
2802 EXPORT_SYMBOL(rdma_resolve_route);
2804 static void cma_set_loopback(struct sockaddr *addr)
2806 switch (addr->sa_family) {
2808 ((struct sockaddr_in *) addr)->sin_addr.s_addr = htonl(INADDR_LOOPBACK);
2811 ipv6_addr_set(&((struct sockaddr_in6 *) addr)->sin6_addr,
2815 ib_addr_set(&((struct sockaddr_ib *) addr)->sib_addr,
2821 static int cma_bind_loopback(struct rdma_id_private *id_priv)
2823 struct cma_device *cma_dev, *cur_dev;
2825 enum ib_port_state port_state;
2832 list_for_each_entry(cur_dev, &dev_list, list) {
2833 if (cma_family(id_priv) == AF_IB &&
2834 !rdma_cap_ib_cm(cur_dev->device, 1))
2840 for (p = 1; p <= cur_dev->device->phys_port_cnt; ++p) {
2841 if (!ib_get_cached_port_state(cur_dev->device, p, &port_state) &&
2842 port_state == IB_PORT_ACTIVE) {
2857 ret = rdma_query_gid(cma_dev->device, p, 0, &gid);
2861 ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
2865 id_priv->id.route.addr.dev_addr.dev_type =
2866 (rdma_protocol_ib(cma_dev->device, p)) ?
2867 ARPHRD_INFINIBAND : ARPHRD_ETHER;
2869 rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2870 ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
2871 id_priv->id.port_num = p;
2872 cma_attach_to_dev(id_priv, cma_dev);
2873 cma_set_loopback(cma_src_addr(id_priv));
2875 mutex_unlock(&lock);
2879 static void addr_handler(int status, struct sockaddr *src_addr,
2880 struct rdma_dev_addr *dev_addr, void *context)
2882 struct rdma_id_private *id_priv = context;
2883 struct rdma_cm_event event = {};
2884 struct sockaddr *addr;
2885 struct sockaddr_storage old_addr;
2887 mutex_lock(&id_priv->handler_mutex);
2888 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY,
2889 RDMA_CM_ADDR_RESOLVED))
2893 * Store the previous src address, so that if we fail to acquire
2894 * matching rdma device, old address can be restored back, which helps
2895 * to cancel the cma listen operation correctly.
2897 addr = cma_src_addr(id_priv);
2898 memcpy(&old_addr, addr, rdma_addr_size(addr));
2899 memcpy(addr, src_addr, rdma_addr_size(src_addr));
2900 if (!status && !id_priv->cma_dev) {
2901 status = cma_acquire_dev(id_priv, NULL);
2903 pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to acquire device. status %d\n",
2905 } else if (status) {
2906 pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to resolve IP. status %d\n", status);
2910 memcpy(addr, &old_addr,
2911 rdma_addr_size((struct sockaddr *)&old_addr));
2912 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
2913 RDMA_CM_ADDR_BOUND))
2915 event.event = RDMA_CM_EVENT_ADDR_ERROR;
2916 event.status = status;
2918 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2920 if (id_priv->id.event_handler(&id_priv->id, &event)) {
2921 cma_exch(id_priv, RDMA_CM_DESTROYING);
2922 mutex_unlock(&id_priv->handler_mutex);
2923 cma_deref_id(id_priv);
2924 rdma_destroy_id(&id_priv->id);
2928 mutex_unlock(&id_priv->handler_mutex);
2929 cma_deref_id(id_priv);
2932 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
2934 struct cma_work *work;
2938 work = kzalloc(sizeof *work, GFP_KERNEL);
2942 if (!id_priv->cma_dev) {
2943 ret = cma_bind_loopback(id_priv);
2948 rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2949 rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
2951 cma_init_resolve_addr_work(work, id_priv);
2952 queue_work(cma_wq, &work->work);
2959 static int cma_resolve_ib_addr(struct rdma_id_private *id_priv)
2961 struct cma_work *work;
2964 work = kzalloc(sizeof *work, GFP_KERNEL);
2968 if (!id_priv->cma_dev) {
2969 ret = cma_resolve_ib_dev(id_priv);
2974 rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, (union ib_gid *)
2975 &(((struct sockaddr_ib *) &id_priv->id.route.addr.dst_addr)->sib_addr));
2977 cma_init_resolve_addr_work(work, id_priv);
2978 queue_work(cma_wq, &work->work);
2985 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2986 const struct sockaddr *dst_addr)
2988 if (!src_addr || !src_addr->sa_family) {
2989 src_addr = (struct sockaddr *) &id->route.addr.src_addr;
2990 src_addr->sa_family = dst_addr->sa_family;
2991 if (IS_ENABLED(CONFIG_IPV6) &&
2992 dst_addr->sa_family == AF_INET6) {
2993 struct sockaddr_in6 *src_addr6 = (struct sockaddr_in6 *) src_addr;
2994 struct sockaddr_in6 *dst_addr6 = (struct sockaddr_in6 *) dst_addr;
2995 src_addr6->sin6_scope_id = dst_addr6->sin6_scope_id;
2996 if (ipv6_addr_type(&dst_addr6->sin6_addr) & IPV6_ADDR_LINKLOCAL)
2997 id->route.addr.dev_addr.bound_dev_if = dst_addr6->sin6_scope_id;
2998 } else if (dst_addr->sa_family == AF_IB) {
2999 ((struct sockaddr_ib *) src_addr)->sib_pkey =
3000 ((struct sockaddr_ib *) dst_addr)->sib_pkey;
3003 return rdma_bind_addr(id, src_addr);
3006 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
3007 const struct sockaddr *dst_addr, int timeout_ms)
3009 struct rdma_id_private *id_priv;
3012 id_priv = container_of(id, struct rdma_id_private, id);
3013 if (id_priv->state == RDMA_CM_IDLE) {
3014 ret = cma_bind_addr(id, src_addr, dst_addr);
3019 if (cma_family(id_priv) != dst_addr->sa_family)
3022 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY))
3025 memcpy(cma_dst_addr(id_priv), dst_addr, rdma_addr_size(dst_addr));
3026 atomic_inc(&id_priv->refcount);
3027 if (cma_any_addr(dst_addr)) {
3028 ret = cma_resolve_loopback(id_priv);
3030 if (dst_addr->sa_family == AF_IB) {
3031 ret = cma_resolve_ib_addr(id_priv);
3033 ret = rdma_resolve_ip(cma_src_addr(id_priv),
3034 dst_addr, &id->route.addr.dev_addr,
3035 timeout_ms, addr_handler, id_priv);
3043 cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
3044 cma_deref_id(id_priv);
3047 EXPORT_SYMBOL(rdma_resolve_addr);
3049 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse)
3051 struct rdma_id_private *id_priv;
3052 unsigned long flags;
3055 id_priv = container_of(id, struct rdma_id_private, id);
3056 spin_lock_irqsave(&id_priv->lock, flags);
3057 if (reuse || id_priv->state == RDMA_CM_IDLE) {
3058 id_priv->reuseaddr = reuse;
3063 spin_unlock_irqrestore(&id_priv->lock, flags);
3066 EXPORT_SYMBOL(rdma_set_reuseaddr);
3068 int rdma_set_afonly(struct rdma_cm_id *id, int afonly)
3070 struct rdma_id_private *id_priv;
3071 unsigned long flags;
3074 id_priv = container_of(id, struct rdma_id_private, id);
3075 spin_lock_irqsave(&id_priv->lock, flags);
3076 if (id_priv->state == RDMA_CM_IDLE || id_priv->state == RDMA_CM_ADDR_BOUND) {
3077 id_priv->options |= (1 << CMA_OPTION_AFONLY);
3078 id_priv->afonly = afonly;
3083 spin_unlock_irqrestore(&id_priv->lock, flags);
3086 EXPORT_SYMBOL(rdma_set_afonly);
3088 static void cma_bind_port(struct rdma_bind_list *bind_list,
3089 struct rdma_id_private *id_priv)
3091 struct sockaddr *addr;
3092 struct sockaddr_ib *sib;
3096 lockdep_assert_held(&lock);
3098 addr = cma_src_addr(id_priv);
3099 port = htons(bind_list->port);
3101 switch (addr->sa_family) {
3103 ((struct sockaddr_in *) addr)->sin_port = port;
3106 ((struct sockaddr_in6 *) addr)->sin6_port = port;
3109 sib = (struct sockaddr_ib *) addr;
3110 sid = be64_to_cpu(sib->sib_sid);
3111 mask = be64_to_cpu(sib->sib_sid_mask);
3112 sib->sib_sid = cpu_to_be64((sid & mask) | (u64) ntohs(port));
3113 sib->sib_sid_mask = cpu_to_be64(~0ULL);
3116 id_priv->bind_list = bind_list;
3117 hlist_add_head(&id_priv->node, &bind_list->owners);
3120 static int cma_alloc_port(enum rdma_ucm_port_space ps,
3121 struct rdma_id_private *id_priv, unsigned short snum)
3123 struct rdma_bind_list *bind_list;
3126 lockdep_assert_held(&lock);
3128 bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
3132 ret = cma_ps_alloc(id_priv->id.route.addr.dev_addr.net, ps, bind_list,
3138 bind_list->port = (unsigned short)ret;
3139 cma_bind_port(bind_list, id_priv);
3143 return ret == -ENOSPC ? -EADDRNOTAVAIL : ret;
3146 static int cma_port_is_unique(struct rdma_bind_list *bind_list,
3147 struct rdma_id_private *id_priv)
3149 struct rdma_id_private *cur_id;
3150 struct sockaddr *daddr = cma_dst_addr(id_priv);
3151 struct sockaddr *saddr = cma_src_addr(id_priv);
3152 __be16 dport = cma_port(daddr);
3154 lockdep_assert_held(&lock);
3156 hlist_for_each_entry(cur_id, &bind_list->owners, node) {
3157 struct sockaddr *cur_daddr = cma_dst_addr(cur_id);
3158 struct sockaddr *cur_saddr = cma_src_addr(cur_id);
3159 __be16 cur_dport = cma_port(cur_daddr);
3161 if (id_priv == cur_id)
3164 /* different dest port -> unique */
3165 if (!cma_any_port(daddr) &&
3166 !cma_any_port(cur_daddr) &&
3167 (dport != cur_dport))
3170 /* different src address -> unique */
3171 if (!cma_any_addr(saddr) &&
3172 !cma_any_addr(cur_saddr) &&
3173 cma_addr_cmp(saddr, cur_saddr))
3176 /* different dst address -> unique */
3177 if (!cma_any_addr(daddr) &&
3178 !cma_any_addr(cur_daddr) &&
3179 cma_addr_cmp(daddr, cur_daddr))
3182 return -EADDRNOTAVAIL;
3187 static int cma_alloc_any_port(enum rdma_ucm_port_space ps,
3188 struct rdma_id_private *id_priv)
3190 static unsigned int last_used_port;
3191 int low, high, remaining;
3193 struct net *net = id_priv->id.route.addr.dev_addr.net;
3195 lockdep_assert_held(&lock);
3197 inet_get_local_port_range(net, &low, &high);
3198 remaining = (high - low) + 1;
3199 rover = prandom_u32() % remaining + low;
3201 if (last_used_port != rover) {
3202 struct rdma_bind_list *bind_list;
3205 bind_list = cma_ps_find(net, ps, (unsigned short)rover);
3208 ret = cma_alloc_port(ps, id_priv, rover);
3210 ret = cma_port_is_unique(bind_list, id_priv);
3212 cma_bind_port(bind_list, id_priv);
3215 * Remember previously used port number in order to avoid
3216 * re-using same port immediately after it is closed.
3219 last_used_port = rover;
3220 if (ret != -EADDRNOTAVAIL)
3225 if ((rover < low) || (rover > high))
3229 return -EADDRNOTAVAIL;
3233 * Check that the requested port is available. This is called when trying to
3234 * bind to a specific port, or when trying to listen on a bound port. In
3235 * the latter case, the provided id_priv may already be on the bind_list, but
3236 * we still need to check that it's okay to start listening.
3238 static int cma_check_port(struct rdma_bind_list *bind_list,
3239 struct rdma_id_private *id_priv, uint8_t reuseaddr)
3241 struct rdma_id_private *cur_id;
3242 struct sockaddr *addr, *cur_addr;
3244 lockdep_assert_held(&lock);
3246 addr = cma_src_addr(id_priv);
3247 hlist_for_each_entry(cur_id, &bind_list->owners, node) {
3248 if (id_priv == cur_id)
3251 if ((cur_id->state != RDMA_CM_LISTEN) && reuseaddr &&
3255 cur_addr = cma_src_addr(cur_id);
3256 if (id_priv->afonly && cur_id->afonly &&
3257 (addr->sa_family != cur_addr->sa_family))
3260 if (cma_any_addr(addr) || cma_any_addr(cur_addr))
3261 return -EADDRNOTAVAIL;
3263 if (!cma_addr_cmp(addr, cur_addr))
3269 static int cma_use_port(enum rdma_ucm_port_space ps,
3270 struct rdma_id_private *id_priv)
3272 struct rdma_bind_list *bind_list;
3273 unsigned short snum;
3276 lockdep_assert_held(&lock);
3278 snum = ntohs(cma_port(cma_src_addr(id_priv)));
3279 if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
3282 bind_list = cma_ps_find(id_priv->id.route.addr.dev_addr.net, ps, snum);
3284 ret = cma_alloc_port(ps, id_priv, snum);
3286 ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr);
3288 cma_bind_port(bind_list, id_priv);
3293 static int cma_bind_listen(struct rdma_id_private *id_priv)
3295 struct rdma_bind_list *bind_list = id_priv->bind_list;
3299 if (bind_list->owners.first->next)
3300 ret = cma_check_port(bind_list, id_priv, 0);
3301 mutex_unlock(&lock);
3305 static enum rdma_ucm_port_space
3306 cma_select_inet_ps(struct rdma_id_private *id_priv)
3308 switch (id_priv->id.ps) {
3313 return id_priv->id.ps;
3320 static enum rdma_ucm_port_space
3321 cma_select_ib_ps(struct rdma_id_private *id_priv)
3323 enum rdma_ucm_port_space ps = 0;
3324 struct sockaddr_ib *sib;
3325 u64 sid_ps, mask, sid;
3327 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
3328 mask = be64_to_cpu(sib->sib_sid_mask) & RDMA_IB_IP_PS_MASK;
3329 sid = be64_to_cpu(sib->sib_sid) & mask;
3331 if ((id_priv->id.ps == RDMA_PS_IB) && (sid == (RDMA_IB_IP_PS_IB & mask))) {
3332 sid_ps = RDMA_IB_IP_PS_IB;
3334 } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_TCP)) &&
3335 (sid == (RDMA_IB_IP_PS_TCP & mask))) {
3336 sid_ps = RDMA_IB_IP_PS_TCP;
3338 } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_UDP)) &&
3339 (sid == (RDMA_IB_IP_PS_UDP & mask))) {
3340 sid_ps = RDMA_IB_IP_PS_UDP;
3345 sib->sib_sid = cpu_to_be64(sid_ps | ntohs(cma_port((struct sockaddr *) sib)));
3346 sib->sib_sid_mask = cpu_to_be64(RDMA_IB_IP_PS_MASK |
3347 be64_to_cpu(sib->sib_sid_mask));
3352 static int cma_get_port(struct rdma_id_private *id_priv)
3354 enum rdma_ucm_port_space ps;
3357 if (cma_family(id_priv) != AF_IB)
3358 ps = cma_select_inet_ps(id_priv);
3360 ps = cma_select_ib_ps(id_priv);
3362 return -EPROTONOSUPPORT;
3365 if (cma_any_port(cma_src_addr(id_priv)))
3366 ret = cma_alloc_any_port(ps, id_priv);
3368 ret = cma_use_port(ps, id_priv);
3369 mutex_unlock(&lock);
3374 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
3375 struct sockaddr *addr)
3377 #if IS_ENABLED(CONFIG_IPV6)
3378 struct sockaddr_in6 *sin6;
3380 if (addr->sa_family != AF_INET6)
3383 sin6 = (struct sockaddr_in6 *) addr;
3385 if (!(ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL))
3388 if (!sin6->sin6_scope_id)
3391 dev_addr->bound_dev_if = sin6->sin6_scope_id;
3396 int rdma_listen(struct rdma_cm_id *id, int backlog)
3398 struct rdma_id_private *id_priv;
3401 id_priv = container_of(id, struct rdma_id_private, id);
3402 if (id_priv->state == RDMA_CM_IDLE) {
3403 id->route.addr.src_addr.ss_family = AF_INET;
3404 ret = rdma_bind_addr(id, cma_src_addr(id_priv));
3409 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN))
3412 if (id_priv->reuseaddr) {
3413 ret = cma_bind_listen(id_priv);
3418 id_priv->backlog = backlog;
3420 if (rdma_cap_ib_cm(id->device, 1)) {
3421 ret = cma_ib_listen(id_priv);
3424 } else if (rdma_cap_iw_cm(id->device, 1)) {
3425 ret = cma_iw_listen(id_priv, backlog);
3433 cma_listen_on_all(id_priv);
3437 id_priv->backlog = 0;
3438 cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND);
3441 EXPORT_SYMBOL(rdma_listen);
3443 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
3445 struct rdma_id_private *id_priv;
3447 struct sockaddr *daddr;
3449 if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6 &&
3450 addr->sa_family != AF_IB)
3451 return -EAFNOSUPPORT;
3453 id_priv = container_of(id, struct rdma_id_private, id);
3454 if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND))
3457 ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
3461 memcpy(cma_src_addr(id_priv), addr, rdma_addr_size(addr));
3462 if (!cma_any_addr(addr)) {
3463 ret = cma_translate_addr(addr, &id->route.addr.dev_addr);
3467 ret = cma_acquire_dev(id_priv, NULL);
3472 if (!(id_priv->options & (1 << CMA_OPTION_AFONLY))) {
3473 if (addr->sa_family == AF_INET)
3474 id_priv->afonly = 1;
3475 #if IS_ENABLED(CONFIG_IPV6)
3476 else if (addr->sa_family == AF_INET6) {
3477 struct net *net = id_priv->id.route.addr.dev_addr.net;
3479 id_priv->afonly = net->ipv6.sysctl.bindv6only;
3483 daddr = cma_dst_addr(id_priv);
3484 daddr->sa_family = addr->sa_family;
3486 ret = cma_get_port(id_priv);
3492 rdma_restrack_del(&id_priv->res);
3493 if (id_priv->cma_dev)
3494 cma_release_dev(id_priv);
3496 cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE);
3499 EXPORT_SYMBOL(rdma_bind_addr);
3501 static int cma_format_hdr(void *hdr, struct rdma_id_private *id_priv)
3503 struct cma_hdr *cma_hdr;
3506 cma_hdr->cma_version = CMA_VERSION;
3507 if (cma_family(id_priv) == AF_INET) {
3508 struct sockaddr_in *src4, *dst4;
3510 src4 = (struct sockaddr_in *) cma_src_addr(id_priv);
3511 dst4 = (struct sockaddr_in *) cma_dst_addr(id_priv);
3513 cma_set_ip_ver(cma_hdr, 4);
3514 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
3515 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
3516 cma_hdr->port = src4->sin_port;
3517 } else if (cma_family(id_priv) == AF_INET6) {
3518 struct sockaddr_in6 *src6, *dst6;
3520 src6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
3521 dst6 = (struct sockaddr_in6 *) cma_dst_addr(id_priv);
3523 cma_set_ip_ver(cma_hdr, 6);
3524 cma_hdr->src_addr.ip6 = src6->sin6_addr;
3525 cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
3526 cma_hdr->port = src6->sin6_port;
3531 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
3532 const struct ib_cm_event *ib_event)
3534 struct rdma_id_private *id_priv = cm_id->context;
3535 struct rdma_cm_event event = {};
3536 const struct ib_cm_sidr_rep_event_param *rep =
3537 &ib_event->param.sidr_rep_rcvd;
3540 mutex_lock(&id_priv->handler_mutex);
3541 if (id_priv->state != RDMA_CM_CONNECT)
3544 switch (ib_event->event) {
3545 case IB_CM_SIDR_REQ_ERROR:
3546 event.event = RDMA_CM_EVENT_UNREACHABLE;
3547 event.status = -ETIMEDOUT;
3549 case IB_CM_SIDR_REP_RECEIVED:
3550 event.param.ud.private_data = ib_event->private_data;
3551 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
3552 if (rep->status != IB_SIDR_SUCCESS) {
3553 event.event = RDMA_CM_EVENT_UNREACHABLE;
3554 event.status = ib_event->param.sidr_rep_rcvd.status;
3555 pr_debug_ratelimited("RDMA CM: UNREACHABLE: bad SIDR reply. status %d\n",
3559 ret = cma_set_qkey(id_priv, rep->qkey);
3561 pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to set qkey. status %d\n", ret);
3562 event.event = RDMA_CM_EVENT_ADDR_ERROR;
3566 ib_init_ah_attr_from_path(id_priv->id.device,
3567 id_priv->id.port_num,
3568 id_priv->id.route.path_rec,
3569 &event.param.ud.ah_attr,
3571 event.param.ud.qp_num = rep->qpn;
3572 event.param.ud.qkey = rep->qkey;
3573 event.event = RDMA_CM_EVENT_ESTABLISHED;
3577 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
3582 ret = id_priv->id.event_handler(&id_priv->id, &event);
3584 rdma_destroy_ah_attr(&event.param.ud.ah_attr);
3586 /* Destroy the CM ID by returning a non-zero value. */
3587 id_priv->cm_id.ib = NULL;
3588 cma_exch(id_priv, RDMA_CM_DESTROYING);
3589 mutex_unlock(&id_priv->handler_mutex);
3590 rdma_destroy_id(&id_priv->id);
3594 mutex_unlock(&id_priv->handler_mutex);
3598 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
3599 struct rdma_conn_param *conn_param)
3601 struct ib_cm_sidr_req_param req;
3602 struct ib_cm_id *id;
3607 memset(&req, 0, sizeof req);
3608 offset = cma_user_data_offset(id_priv);
3609 req.private_data_len = offset + conn_param->private_data_len;
3610 if (req.private_data_len < conn_param->private_data_len)
3613 if (req.private_data_len) {
3614 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
3618 private_data = NULL;
3621 if (conn_param->private_data && conn_param->private_data_len)
3622 memcpy(private_data + offset, conn_param->private_data,
3623 conn_param->private_data_len);
3626 ret = cma_format_hdr(private_data, id_priv);
3629 req.private_data = private_data;
3632 id = ib_create_cm_id(id_priv->id.device, cma_sidr_rep_handler,
3638 id_priv->cm_id.ib = id;
3640 req.path = id_priv->id.route.path_rec;
3641 req.sgid_attr = id_priv->id.route.addr.dev_addr.sgid_attr;
3642 req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
3643 req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
3644 req.max_cm_retries = CMA_MAX_CM_RETRIES;
3646 ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
3648 ib_destroy_cm_id(id_priv->cm_id.ib);
3649 id_priv->cm_id.ib = NULL;
3652 kfree(private_data);
3656 static int cma_connect_ib(struct rdma_id_private *id_priv,
3657 struct rdma_conn_param *conn_param)
3659 struct ib_cm_req_param req;
3660 struct rdma_route *route;
3662 struct ib_cm_id *id;
3666 memset(&req, 0, sizeof req);
3667 offset = cma_user_data_offset(id_priv);
3668 req.private_data_len = offset + conn_param->private_data_len;
3669 if (req.private_data_len < conn_param->private_data_len)
3672 if (req.private_data_len) {
3673 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
3677 private_data = NULL;
3680 if (conn_param->private_data && conn_param->private_data_len)
3681 memcpy(private_data + offset, conn_param->private_data,
3682 conn_param->private_data_len);
3684 id = ib_create_cm_id(id_priv->id.device, cma_ib_handler, id_priv);
3689 id_priv->cm_id.ib = id;
3691 route = &id_priv->id.route;
3693 ret = cma_format_hdr(private_data, id_priv);
3696 req.private_data = private_data;
3699 req.primary_path = &route->path_rec[0];
3700 if (route->num_paths == 2)
3701 req.alternate_path = &route->path_rec[1];
3703 req.ppath_sgid_attr = id_priv->id.route.addr.dev_addr.sgid_attr;
3704 /* Alternate path SGID attribute currently unsupported */
3705 req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
3706 req.qp_num = id_priv->qp_num;
3707 req.qp_type = id_priv->id.qp_type;
3708 req.starting_psn = id_priv->seq_num;
3709 req.responder_resources = conn_param->responder_resources;
3710 req.initiator_depth = conn_param->initiator_depth;
3711 req.flow_control = conn_param->flow_control;
3712 req.retry_count = min_t(u8, 7, conn_param->retry_count);
3713 req.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
3714 req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
3715 req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
3716 req.max_cm_retries = CMA_MAX_CM_RETRIES;
3717 req.srq = id_priv->srq ? 1 : 0;
3719 ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
3721 if (ret && !IS_ERR(id)) {
3722 ib_destroy_cm_id(id);
3723 id_priv->cm_id.ib = NULL;
3726 kfree(private_data);
3730 static int cma_connect_iw(struct rdma_id_private *id_priv,
3731 struct rdma_conn_param *conn_param)
3733 struct iw_cm_id *cm_id;
3735 struct iw_cm_conn_param iw_param;
3737 cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
3739 return PTR_ERR(cm_id);
3741 cm_id->tos = id_priv->tos;
3742 id_priv->cm_id.iw = cm_id;
3744 memcpy(&cm_id->local_addr, cma_src_addr(id_priv),
3745 rdma_addr_size(cma_src_addr(id_priv)));
3746 memcpy(&cm_id->remote_addr, cma_dst_addr(id_priv),
3747 rdma_addr_size(cma_dst_addr(id_priv)));
3749 ret = cma_modify_qp_rtr(id_priv, conn_param);
3754 iw_param.ord = conn_param->initiator_depth;
3755 iw_param.ird = conn_param->responder_resources;
3756 iw_param.private_data = conn_param->private_data;
3757 iw_param.private_data_len = conn_param->private_data_len;
3758 iw_param.qpn = id_priv->id.qp ? id_priv->qp_num : conn_param->qp_num;
3760 memset(&iw_param, 0, sizeof iw_param);
3761 iw_param.qpn = id_priv->qp_num;
3763 ret = iw_cm_connect(cm_id, &iw_param);
3766 iw_destroy_cm_id(cm_id);
3767 id_priv->cm_id.iw = NULL;
3772 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
3774 struct rdma_id_private *id_priv;
3777 id_priv = container_of(id, struct rdma_id_private, id);
3778 if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT))
3782 id_priv->qp_num = conn_param->qp_num;
3783 id_priv->srq = conn_param->srq;
3786 if (rdma_cap_ib_cm(id->device, id->port_num)) {
3787 if (id->qp_type == IB_QPT_UD)
3788 ret = cma_resolve_ib_udp(id_priv, conn_param);
3790 ret = cma_connect_ib(id_priv, conn_param);
3791 } else if (rdma_cap_iw_cm(id->device, id->port_num))
3792 ret = cma_connect_iw(id_priv, conn_param);
3800 cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED);
3803 EXPORT_SYMBOL(rdma_connect);
3805 static int cma_accept_ib(struct rdma_id_private *id_priv,
3806 struct rdma_conn_param *conn_param)
3808 struct ib_cm_rep_param rep;
3811 ret = cma_modify_qp_rtr(id_priv, conn_param);
3815 ret = cma_modify_qp_rts(id_priv, conn_param);
3819 memset(&rep, 0, sizeof rep);
3820 rep.qp_num = id_priv->qp_num;
3821 rep.starting_psn = id_priv->seq_num;
3822 rep.private_data = conn_param->private_data;
3823 rep.private_data_len = conn_param->private_data_len;
3824 rep.responder_resources = conn_param->responder_resources;
3825 rep.initiator_depth = conn_param->initiator_depth;
3826 rep.failover_accepted = 0;
3827 rep.flow_control = conn_param->flow_control;
3828 rep.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
3829 rep.srq = id_priv->srq ? 1 : 0;
3831 ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
3836 static int cma_accept_iw(struct rdma_id_private *id_priv,
3837 struct rdma_conn_param *conn_param)
3839 struct iw_cm_conn_param iw_param;
3845 ret = cma_modify_qp_rtr(id_priv, conn_param);
3849 iw_param.ord = conn_param->initiator_depth;
3850 iw_param.ird = conn_param->responder_resources;
3851 iw_param.private_data = conn_param->private_data;
3852 iw_param.private_data_len = conn_param->private_data_len;
3853 if (id_priv->id.qp) {
3854 iw_param.qpn = id_priv->qp_num;
3856 iw_param.qpn = conn_param->qp_num;
3858 return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
3861 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
3862 enum ib_cm_sidr_status status, u32 qkey,
3863 const void *private_data, int private_data_len)
3865 struct ib_cm_sidr_rep_param rep;
3868 memset(&rep, 0, sizeof rep);
3869 rep.status = status;
3870 if (status == IB_SIDR_SUCCESS) {
3871 ret = cma_set_qkey(id_priv, qkey);
3874 rep.qp_num = id_priv->qp_num;
3875 rep.qkey = id_priv->qkey;
3877 rep.private_data = private_data;
3878 rep.private_data_len = private_data_len;
3880 return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
3883 int __rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param,
3886 struct rdma_id_private *id_priv;
3889 id_priv = container_of(id, struct rdma_id_private, id);
3892 id_priv->res.kern_name = caller;
3894 rdma_restrack_set_task(&id_priv->res, current);
3896 if (!cma_comp(id_priv, RDMA_CM_CONNECT))
3899 if (!id->qp && conn_param) {
3900 id_priv->qp_num = conn_param->qp_num;
3901 id_priv->srq = conn_param->srq;
3904 if (rdma_cap_ib_cm(id->device, id->port_num)) {
3905 if (id->qp_type == IB_QPT_UD) {
3907 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
3909 conn_param->private_data,
3910 conn_param->private_data_len);
3912 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
3916 ret = cma_accept_ib(id_priv, conn_param);
3918 ret = cma_rep_recv(id_priv);
3920 } else if (rdma_cap_iw_cm(id->device, id->port_num))
3921 ret = cma_accept_iw(id_priv, conn_param);
3930 cma_modify_qp_err(id_priv);
3931 rdma_reject(id, NULL, 0);
3934 EXPORT_SYMBOL(__rdma_accept);
3936 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
3938 struct rdma_id_private *id_priv;
3941 id_priv = container_of(id, struct rdma_id_private, id);
3942 if (!id_priv->cm_id.ib)
3945 switch (id->device->node_type) {
3946 case RDMA_NODE_IB_CA:
3947 ret = ib_cm_notify(id_priv->cm_id.ib, event);
3955 EXPORT_SYMBOL(rdma_notify);
3957 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
3958 u8 private_data_len)
3960 struct rdma_id_private *id_priv;
3963 id_priv = container_of(id, struct rdma_id_private, id);
3964 if (!id_priv->cm_id.ib)
3967 if (rdma_cap_ib_cm(id->device, id->port_num)) {
3968 if (id->qp_type == IB_QPT_UD)
3969 ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT, 0,
3970 private_data, private_data_len);
3972 ret = ib_send_cm_rej(id_priv->cm_id.ib,
3973 IB_CM_REJ_CONSUMER_DEFINED, NULL,
3974 0, private_data, private_data_len);
3975 } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
3976 ret = iw_cm_reject(id_priv->cm_id.iw,
3977 private_data, private_data_len);
3983 EXPORT_SYMBOL(rdma_reject);
3985 int rdma_disconnect(struct rdma_cm_id *id)
3987 struct rdma_id_private *id_priv;
3990 id_priv = container_of(id, struct rdma_id_private, id);
3991 if (!id_priv->cm_id.ib)
3994 if (rdma_cap_ib_cm(id->device, id->port_num)) {
3995 ret = cma_modify_qp_err(id_priv);
3998 /* Initiate or respond to a disconnect. */
3999 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
4000 ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
4001 } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4002 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
4009 EXPORT_SYMBOL(rdma_disconnect);
4011 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
4013 struct rdma_id_private *id_priv;
4014 struct cma_multicast *mc = multicast->context;
4015 struct rdma_cm_event event = {};
4018 id_priv = mc->id_priv;
4019 mutex_lock(&id_priv->handler_mutex);
4020 if (id_priv->state != RDMA_CM_ADDR_BOUND &&
4021 id_priv->state != RDMA_CM_ADDR_RESOLVED)
4025 status = cma_set_qkey(id_priv, be32_to_cpu(multicast->rec.qkey));
4027 pr_debug_ratelimited("RDMA CM: MULTICAST_ERROR: failed to join multicast. status %d\n",
4029 event.status = status;
4030 event.param.ud.private_data = mc->context;
4032 struct rdma_dev_addr *dev_addr =
4033 &id_priv->id.route.addr.dev_addr;
4034 struct net_device *ndev =
4035 dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if);
4036 enum ib_gid_type gid_type =
4037 id_priv->cma_dev->default_gid_type[id_priv->id.port_num -
4038 rdma_start_port(id_priv->cma_dev->device)];
4040 event.event = RDMA_CM_EVENT_MULTICAST_JOIN;
4041 ret = ib_init_ah_from_mcmember(id_priv->id.device,
4042 id_priv->id.port_num,
4045 &event.param.ud.ah_attr);
4047 event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
4049 event.param.ud.qp_num = 0xFFFFFF;
4050 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
4054 event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
4056 ret = id_priv->id.event_handler(&id_priv->id, &event);
4058 rdma_destroy_ah_attr(&event.param.ud.ah_attr);
4060 cma_exch(id_priv, RDMA_CM_DESTROYING);
4061 mutex_unlock(&id_priv->handler_mutex);
4062 rdma_destroy_id(&id_priv->id);
4067 mutex_unlock(&id_priv->handler_mutex);
4071 static void cma_set_mgid(struct rdma_id_private *id_priv,
4072 struct sockaddr *addr, union ib_gid *mgid)
4074 unsigned char mc_map[MAX_ADDR_LEN];
4075 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4076 struct sockaddr_in *sin = (struct sockaddr_in *) addr;
4077 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
4079 if (cma_any_addr(addr)) {
4080 memset(mgid, 0, sizeof *mgid);
4081 } else if ((addr->sa_family == AF_INET6) &&
4082 ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
4084 /* IPv6 address is an SA assigned MGID. */
4085 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
4086 } else if (addr->sa_family == AF_IB) {
4087 memcpy(mgid, &((struct sockaddr_ib *) addr)->sib_addr, sizeof *mgid);
4088 } else if (addr->sa_family == AF_INET6) {
4089 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
4090 if (id_priv->id.ps == RDMA_PS_UDP)
4091 mc_map[7] = 0x01; /* Use RDMA CM signature */
4092 *mgid = *(union ib_gid *) (mc_map + 4);
4094 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
4095 if (id_priv->id.ps == RDMA_PS_UDP)
4096 mc_map[7] = 0x01; /* Use RDMA CM signature */
4097 *mgid = *(union ib_gid *) (mc_map + 4);
4101 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
4102 struct cma_multicast *mc)
4104 struct ib_sa_mcmember_rec rec;
4105 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4106 ib_sa_comp_mask comp_mask;
4109 ib_addr_get_mgid(dev_addr, &rec.mgid);
4110 ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
4115 ret = cma_set_qkey(id_priv, 0);
4119 cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
4120 rec.qkey = cpu_to_be32(id_priv->qkey);
4121 rdma_addr_get_sgid(dev_addr, &rec.port_gid);
4122 rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
4123 rec.join_state = mc->join_state;
4125 if ((rec.join_state == BIT(SENDONLY_FULLMEMBER_JOIN)) &&
4126 (!ib_sa_sendonly_fullmem_support(&sa_client,
4128 id_priv->id.port_num))) {
4129 pr_warn("RDMA CM: %s port %u Unable to multicast join\n"
4130 "RDMA CM: SM doesn't support Send Only Full Member option\n",
4131 id_priv->id.device->name, id_priv->id.port_num);
4135 comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
4136 IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
4137 IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
4138 IB_SA_MCMEMBER_REC_FLOW_LABEL |
4139 IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
4141 if (id_priv->id.ps == RDMA_PS_IPOIB)
4142 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
4143 IB_SA_MCMEMBER_REC_RATE_SELECTOR |
4144 IB_SA_MCMEMBER_REC_MTU_SELECTOR |
4145 IB_SA_MCMEMBER_REC_MTU |
4146 IB_SA_MCMEMBER_REC_HOP_LIMIT;
4148 mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device,
4149 id_priv->id.port_num, &rec,
4150 comp_mask, GFP_KERNEL,
4151 cma_ib_mc_handler, mc);
4152 return PTR_ERR_OR_ZERO(mc->multicast.ib);
4155 static void iboe_mcast_work_handler(struct work_struct *work)
4157 struct iboe_mcast_work *mw = container_of(work, struct iboe_mcast_work, work);
4158 struct cma_multicast *mc = mw->mc;
4159 struct ib_sa_multicast *m = mc->multicast.ib;
4161 mc->multicast.ib->context = mc;
4162 cma_ib_mc_handler(0, m);
4163 kref_put(&mc->mcref, release_mc);
4167 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid,
4168 enum ib_gid_type gid_type)
4170 struct sockaddr_in *sin = (struct sockaddr_in *)addr;
4171 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
4173 if (cma_any_addr(addr)) {
4174 memset(mgid, 0, sizeof *mgid);
4175 } else if (addr->sa_family == AF_INET6) {
4176 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
4179 (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ? 0 : 0xff;
4181 (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ? 0 : 0x0e;
4190 mgid->raw[10] = 0xff;
4191 mgid->raw[11] = 0xff;
4192 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
4196 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
4197 struct cma_multicast *mc)
4199 struct iboe_mcast_work *work;
4200 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4202 struct sockaddr *addr = (struct sockaddr *)&mc->addr;
4203 struct net_device *ndev = NULL;
4204 enum ib_gid_type gid_type;
4207 send_only = mc->join_state == BIT(SENDONLY_FULLMEMBER_JOIN);
4209 if (cma_zero_addr((struct sockaddr *)&mc->addr))
4212 work = kzalloc(sizeof *work, GFP_KERNEL);
4216 mc->multicast.ib = kzalloc(sizeof(struct ib_sa_multicast), GFP_KERNEL);
4217 if (!mc->multicast.ib) {
4222 gid_type = id_priv->cma_dev->default_gid_type[id_priv->id.port_num -
4223 rdma_start_port(id_priv->cma_dev->device)];
4224 cma_iboe_set_mgid(addr, &mc->multicast.ib->rec.mgid, gid_type);
4226 mc->multicast.ib->rec.pkey = cpu_to_be16(0xffff);
4227 if (id_priv->id.ps == RDMA_PS_UDP)
4228 mc->multicast.ib->rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
4230 if (dev_addr->bound_dev_if)
4231 ndev = dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if);
4236 mc->multicast.ib->rec.rate = iboe_get_rate(ndev);
4237 mc->multicast.ib->rec.hop_limit = 1;
4238 mc->multicast.ib->rec.mtu = iboe_get_mtu(ndev->mtu);
4240 if (addr->sa_family == AF_INET) {
4241 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) {
4242 mc->multicast.ib->rec.hop_limit = IPV6_DEFAULT_HOPLIMIT;
4244 err = cma_igmp_send(ndev, &mc->multicast.ib->rec.mgid,
4249 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP)
4253 if (err || !mc->multicast.ib->rec.mtu) {
4258 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
4259 &mc->multicast.ib->rec.port_gid);
4262 INIT_WORK(&work->work, iboe_mcast_work_handler);
4263 kref_get(&mc->mcref);
4264 queue_work(cma_wq, &work->work);
4269 kfree(mc->multicast.ib);
4275 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
4276 u8 join_state, void *context)
4278 struct rdma_id_private *id_priv;
4279 struct cma_multicast *mc;
4282 /* Not supported for kernel QPs */
4283 if (WARN_ON(id->qp))
4289 id_priv = container_of(id, struct rdma_id_private, id);
4290 if (!cma_comp(id_priv, RDMA_CM_ADDR_BOUND) &&
4291 !cma_comp(id_priv, RDMA_CM_ADDR_RESOLVED))
4294 mc = kmalloc(sizeof *mc, GFP_KERNEL);
4298 memcpy(&mc->addr, addr, rdma_addr_size(addr));
4299 mc->context = context;
4300 mc->id_priv = id_priv;
4301 mc->join_state = join_state;
4303 if (rdma_protocol_roce(id->device, id->port_num)) {
4304 kref_init(&mc->mcref);
4305 ret = cma_iboe_join_multicast(id_priv, mc);
4308 } else if (rdma_cap_ib_mcast(id->device, id->port_num)) {
4309 ret = cma_join_ib_multicast(id_priv, mc);
4317 spin_lock(&id_priv->lock);
4318 list_add(&mc->list, &id_priv->mc_list);
4319 spin_unlock(&id_priv->lock);
4326 EXPORT_SYMBOL(rdma_join_multicast);
4328 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
4330 struct rdma_id_private *id_priv;
4331 struct cma_multicast *mc;
4333 id_priv = container_of(id, struct rdma_id_private, id);
4334 spin_lock_irq(&id_priv->lock);
4335 list_for_each_entry(mc, &id_priv->mc_list, list) {
4336 if (memcmp(&mc->addr, addr, rdma_addr_size(addr)) != 0)
4338 list_del(&mc->list);
4339 spin_unlock_irq(&id_priv->lock);
4341 WARN_ON(id_priv->cma_dev->device != id->device);
4342 destroy_mc(id_priv, mc);
4345 spin_unlock_irq(&id_priv->lock);
4347 EXPORT_SYMBOL(rdma_leave_multicast);
4349 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
4351 struct rdma_dev_addr *dev_addr;
4352 struct cma_ndev_work *work;
4354 dev_addr = &id_priv->id.route.addr.dev_addr;
4356 if ((dev_addr->bound_dev_if == ndev->ifindex) &&
4357 (net_eq(dev_net(ndev), dev_addr->net)) &&
4358 memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
4359 pr_info("RDMA CM addr change for ndev %s used by id %p\n",
4360 ndev->name, &id_priv->id);
4361 work = kzalloc(sizeof *work, GFP_KERNEL);
4365 INIT_WORK(&work->work, cma_ndev_work_handler);
4367 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
4368 atomic_inc(&id_priv->refcount);
4369 queue_work(cma_wq, &work->work);
4375 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
4378 struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
4379 struct cma_device *cma_dev;
4380 struct rdma_id_private *id_priv;
4381 int ret = NOTIFY_DONE;
4383 if (event != NETDEV_BONDING_FAILOVER)
4386 if (!netif_is_bond_master(ndev))
4390 list_for_each_entry(cma_dev, &dev_list, list)
4391 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
4392 ret = cma_netdev_change(ndev, id_priv);
4398 mutex_unlock(&lock);
4402 static struct notifier_block cma_nb = {
4403 .notifier_call = cma_netdev_callback
4406 static void cma_add_one(struct ib_device *device)
4408 struct cma_device *cma_dev;
4409 struct rdma_id_private *id_priv;
4411 unsigned long supported_gids = 0;
4413 cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
4417 cma_dev->device = device;
4418 cma_dev->default_gid_type = kcalloc(device->phys_port_cnt,
4419 sizeof(*cma_dev->default_gid_type),
4421 if (!cma_dev->default_gid_type)
4424 cma_dev->default_roce_tos = kcalloc(device->phys_port_cnt,
4425 sizeof(*cma_dev->default_roce_tos),
4427 if (!cma_dev->default_roce_tos)
4430 for (i = rdma_start_port(device); i <= rdma_end_port(device); i++) {
4431 supported_gids = roce_gid_type_mask_support(device, i);
4432 WARN_ON(!supported_gids);
4433 if (supported_gids & (1 << CMA_PREFERRED_ROCE_GID_TYPE))
4434 cma_dev->default_gid_type[i - rdma_start_port(device)] =
4435 CMA_PREFERRED_ROCE_GID_TYPE;
4437 cma_dev->default_gid_type[i - rdma_start_port(device)] =
4438 find_first_bit(&supported_gids, BITS_PER_LONG);
4439 cma_dev->default_roce_tos[i - rdma_start_port(device)] = 0;
4442 init_completion(&cma_dev->comp);
4443 atomic_set(&cma_dev->refcount, 1);
4444 INIT_LIST_HEAD(&cma_dev->id_list);
4445 ib_set_client_data(device, &cma_client, cma_dev);
4448 list_add_tail(&cma_dev->list, &dev_list);
4449 list_for_each_entry(id_priv, &listen_any_list, list)
4450 cma_listen_on_dev(id_priv, cma_dev);
4451 mutex_unlock(&lock);
4456 kfree(cma_dev->default_gid_type);
4464 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
4466 struct rdma_cm_event event = {};
4467 enum rdma_cm_state state;
4470 /* Record that we want to remove the device */
4471 state = cma_exch(id_priv, RDMA_CM_DEVICE_REMOVAL);
4472 if (state == RDMA_CM_DESTROYING)
4475 cma_cancel_operation(id_priv, state);
4476 mutex_lock(&id_priv->handler_mutex);
4478 /* Check for destruction from another callback. */
4479 if (!cma_comp(id_priv, RDMA_CM_DEVICE_REMOVAL))
4482 event.event = RDMA_CM_EVENT_DEVICE_REMOVAL;
4483 ret = id_priv->id.event_handler(&id_priv->id, &event);
4485 mutex_unlock(&id_priv->handler_mutex);
4489 static void cma_process_remove(struct cma_device *cma_dev)
4491 struct rdma_id_private *id_priv;
4495 while (!list_empty(&cma_dev->id_list)) {
4496 id_priv = list_entry(cma_dev->id_list.next,
4497 struct rdma_id_private, list);
4499 list_del(&id_priv->listen_list);
4500 list_del_init(&id_priv->list);
4501 atomic_inc(&id_priv->refcount);
4502 mutex_unlock(&lock);
4504 ret = id_priv->internal_id ? 1 : cma_remove_id_dev(id_priv);
4505 cma_deref_id(id_priv);
4507 rdma_destroy_id(&id_priv->id);
4511 mutex_unlock(&lock);
4513 cma_deref_dev(cma_dev);
4514 wait_for_completion(&cma_dev->comp);
4517 static void cma_remove_one(struct ib_device *device, void *client_data)
4519 struct cma_device *cma_dev = client_data;
4525 list_del(&cma_dev->list);
4526 mutex_unlock(&lock);
4528 cma_process_remove(cma_dev);
4529 kfree(cma_dev->default_roce_tos);
4530 kfree(cma_dev->default_gid_type);
4534 static int cma_get_id_stats(struct sk_buff *skb, struct netlink_callback *cb)
4536 struct nlmsghdr *nlh;
4537 struct rdma_cm_id_stats *id_stats;
4538 struct rdma_id_private *id_priv;
4539 struct rdma_cm_id *id = NULL;
4540 struct cma_device *cma_dev;
4541 int i_dev = 0, i_id = 0;
4544 * We export all of the IDs as a sequence of messages. Each
4545 * ID gets its own netlink message.
4549 list_for_each_entry(cma_dev, &dev_list, list) {
4550 if (i_dev < cb->args[0]) {
4556 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
4557 if (i_id < cb->args[1]) {
4562 id_stats = ibnl_put_msg(skb, &nlh, cb->nlh->nlmsg_seq,
4563 sizeof *id_stats, RDMA_NL_RDMA_CM,
4564 RDMA_NL_RDMA_CM_ID_STATS,
4569 memset(id_stats, 0, sizeof *id_stats);
4571 id_stats->node_type = id->route.addr.dev_addr.dev_type;
4572 id_stats->port_num = id->port_num;
4573 id_stats->bound_dev_if =
4574 id->route.addr.dev_addr.bound_dev_if;
4576 if (ibnl_put_attr(skb, nlh,
4577 rdma_addr_size(cma_src_addr(id_priv)),
4578 cma_src_addr(id_priv),
4579 RDMA_NL_RDMA_CM_ATTR_SRC_ADDR))
4581 if (ibnl_put_attr(skb, nlh,
4582 rdma_addr_size(cma_dst_addr(id_priv)),
4583 cma_dst_addr(id_priv),
4584 RDMA_NL_RDMA_CM_ATTR_DST_ADDR))
4587 id_stats->pid = task_pid_vnr(id_priv->res.task);
4588 id_stats->port_space = id->ps;
4589 id_stats->cm_state = id_priv->state;
4590 id_stats->qp_num = id_priv->qp_num;
4591 id_stats->qp_type = id->qp_type;
4594 nlmsg_end(skb, nlh);
4602 mutex_unlock(&lock);
4603 cb->args[0] = i_dev;
4609 static const struct rdma_nl_cbs cma_cb_table[RDMA_NL_RDMA_CM_NUM_OPS] = {
4610 [RDMA_NL_RDMA_CM_ID_STATS] = { .dump = cma_get_id_stats},
4613 static int cma_init_net(struct net *net)
4615 struct cma_pernet *pernet = cma_pernet(net);
4617 idr_init(&pernet->tcp_ps);
4618 idr_init(&pernet->udp_ps);
4619 idr_init(&pernet->ipoib_ps);
4620 idr_init(&pernet->ib_ps);
4625 static void cma_exit_net(struct net *net)
4627 struct cma_pernet *pernet = cma_pernet(net);
4629 idr_destroy(&pernet->tcp_ps);
4630 idr_destroy(&pernet->udp_ps);
4631 idr_destroy(&pernet->ipoib_ps);
4632 idr_destroy(&pernet->ib_ps);
4635 static struct pernet_operations cma_pernet_operations = {
4636 .init = cma_init_net,
4637 .exit = cma_exit_net,
4638 .id = &cma_pernet_id,
4639 .size = sizeof(struct cma_pernet),
4642 static int __init cma_init(void)
4647 * There is a rare lock ordering dependency in cma_netdev_callback()
4648 * that only happens when bonding is enabled. Teach lockdep that rtnl
4649 * must never be nested under lock so it can find these without having
4650 * to test with bonding.
4652 if (IS_ENABLED(CONFIG_LOCKDEP)) {
4655 mutex_unlock(&lock);
4659 cma_wq = alloc_ordered_workqueue("rdma_cm", WQ_MEM_RECLAIM);
4663 ret = register_pernet_subsys(&cma_pernet_operations);
4667 ib_sa_register_client(&sa_client);
4668 register_netdevice_notifier(&cma_nb);
4670 ret = ib_register_client(&cma_client);
4674 rdma_nl_register(RDMA_NL_RDMA_CM, cma_cb_table);
4675 cma_configfs_init();
4680 unregister_netdevice_notifier(&cma_nb);
4681 ib_sa_unregister_client(&sa_client);
4682 unregister_pernet_subsys(&cma_pernet_operations);
4684 destroy_workqueue(cma_wq);
4688 static void __exit cma_cleanup(void)
4690 cma_configfs_exit();
4691 rdma_nl_unregister(RDMA_NL_RDMA_CM);
4692 ib_unregister_client(&cma_client);
4693 unregister_netdevice_notifier(&cma_nb);
4694 ib_sa_unregister_client(&sa_client);
4695 unregister_pernet_subsys(&cma_pernet_operations);
4696 destroy_workqueue(cma_wq);
4699 MODULE_ALIAS_RDMA_NETLINK(RDMA_NL_RDMA_CM, 1);
4701 module_init(cma_init);
4702 module_exit(cma_cleanup);