2 * Copyright (c) 2005 Voltaire Inc. All rights reserved.
3 * Copyright (c) 2002-2005, Network Appliance, Inc. All rights reserved.
4 * Copyright (c) 1999-2005, Mellanox Technologies, Inc. All rights reserved.
5 * Copyright (c) 2005-2006 Intel Corporation. All rights reserved.
7 * This software is available to you under a choice of one of two
8 * licenses. You may choose to be licensed under the terms of the GNU
9 * General Public License (GPL) Version 2, available from the file
10 * COPYING in the main directory of this source tree, or the
11 * OpenIB.org BSD license below:
13 * Redistribution and use in source and binary forms, with or
14 * without modification, are permitted provided that the following
17 * - Redistributions of source code must retain the above
18 * copyright notice, this list of conditions and the following
21 * - Redistributions in binary form must reproduce the above
22 * copyright notice, this list of conditions and the following
23 * disclaimer in the documentation and/or other materials
24 * provided with the distribution.
26 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
27 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
28 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
29 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
30 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
31 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
32 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
36 #include <linux/completion.h>
38 #include <linux/in6.h>
39 #include <linux/mutex.h>
40 #include <linux/random.h>
41 #include <linux/idr.h>
42 #include <linux/inetdevice.h>
43 #include <linux/slab.h>
44 #include <linux/module.h>
45 #include <net/route.h>
50 #include <rdma/rdma_cm.h>
51 #include <rdma/rdma_cm_ib.h>
52 #include <rdma/rdma_netlink.h>
54 #include <rdma/ib_cache.h>
55 #include <rdma/ib_cm.h>
56 #include <rdma/ib_sa.h>
57 #include <rdma/iw_cm.h>
59 MODULE_AUTHOR("Sean Hefty");
60 MODULE_DESCRIPTION("Generic RDMA CM Agent");
61 MODULE_LICENSE("Dual BSD/GPL");
63 #define CMA_CM_RESPONSE_TIMEOUT 20
64 #define CMA_MAX_CM_RETRIES 15
65 #define CMA_CM_MRA_SETTING (IB_CM_MRA_FLAG_DELAY | 24)
66 #define CMA_IBOE_PACKET_LIFETIME 18
68 static void cma_add_one(struct ib_device *device);
69 static void cma_remove_one(struct ib_device *device);
71 static struct ib_client cma_client = {
74 .remove = cma_remove_one
77 static struct ib_sa_client sa_client;
78 static struct rdma_addr_client addr_client;
79 static LIST_HEAD(dev_list);
80 static LIST_HEAD(listen_any_list);
81 static DEFINE_MUTEX(lock);
82 static struct workqueue_struct *cma_wq;
83 static DEFINE_IDR(tcp_ps);
84 static DEFINE_IDR(udp_ps);
85 static DEFINE_IDR(ipoib_ps);
86 static DEFINE_IDR(ib_ps);
89 struct list_head list;
90 struct ib_device *device;
91 struct completion comp;
93 struct list_head id_list;
96 struct rdma_bind_list {
98 struct hlist_head owners;
107 * Device removal can occur at anytime, so we need extra handling to
108 * serialize notifying the user of device removal with other callbacks.
109 * We do this by disabling removal notification while a callback is in process,
110 * and reporting it after the callback completes.
112 struct rdma_id_private {
113 struct rdma_cm_id id;
115 struct rdma_bind_list *bind_list;
116 struct hlist_node node;
117 struct list_head list; /* listen_any_list or cma_device.list */
118 struct list_head listen_list; /* per device listens */
119 struct cma_device *cma_dev;
120 struct list_head mc_list;
123 enum rdma_cm_state state;
125 struct mutex qp_mutex;
127 struct completion comp;
129 struct mutex handler_mutex;
133 struct ib_sa_query *query;
151 struct cma_multicast {
152 struct rdma_id_private *id_priv;
154 struct ib_sa_multicast *ib;
156 struct list_head list;
158 struct sockaddr_storage addr;
163 struct work_struct work;
164 struct rdma_id_private *id;
165 enum rdma_cm_state old_state;
166 enum rdma_cm_state new_state;
167 struct rdma_cm_event event;
170 struct cma_ndev_work {
171 struct work_struct work;
172 struct rdma_id_private *id;
173 struct rdma_cm_event event;
176 struct iboe_mcast_work {
177 struct work_struct work;
178 struct rdma_id_private *id;
179 struct cma_multicast *mc;
192 u8 ip_version; /* IP version: 7:4 */
194 union cma_ip_addr src_addr;
195 union cma_ip_addr dst_addr;
198 #define CMA_VERSION 0x00
200 static int cma_comp(struct rdma_id_private *id_priv, enum rdma_cm_state comp)
205 spin_lock_irqsave(&id_priv->lock, flags);
206 ret = (id_priv->state == comp);
207 spin_unlock_irqrestore(&id_priv->lock, flags);
211 static int cma_comp_exch(struct rdma_id_private *id_priv,
212 enum rdma_cm_state comp, enum rdma_cm_state exch)
217 spin_lock_irqsave(&id_priv->lock, flags);
218 if ((ret = (id_priv->state == comp)))
219 id_priv->state = exch;
220 spin_unlock_irqrestore(&id_priv->lock, flags);
224 static enum rdma_cm_state cma_exch(struct rdma_id_private *id_priv,
225 enum rdma_cm_state exch)
228 enum rdma_cm_state old;
230 spin_lock_irqsave(&id_priv->lock, flags);
231 old = id_priv->state;
232 id_priv->state = exch;
233 spin_unlock_irqrestore(&id_priv->lock, flags);
237 static inline u8 cma_get_ip_ver(struct cma_hdr *hdr)
239 return hdr->ip_version >> 4;
242 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
244 hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
247 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
248 struct cma_device *cma_dev)
250 atomic_inc(&cma_dev->refcount);
251 id_priv->cma_dev = cma_dev;
252 id_priv->id.device = cma_dev->device;
253 id_priv->id.route.addr.dev_addr.transport =
254 rdma_node_get_transport(cma_dev->device->node_type);
255 list_add_tail(&id_priv->list, &cma_dev->id_list);
258 static inline void cma_deref_dev(struct cma_device *cma_dev)
260 if (atomic_dec_and_test(&cma_dev->refcount))
261 complete(&cma_dev->comp);
264 static inline void release_mc(struct kref *kref)
266 struct cma_multicast *mc = container_of(kref, struct cma_multicast, mcref);
268 kfree(mc->multicast.ib);
272 static void cma_release_dev(struct rdma_id_private *id_priv)
275 list_del(&id_priv->list);
276 cma_deref_dev(id_priv->cma_dev);
277 id_priv->cma_dev = NULL;
281 static inline struct sockaddr *cma_src_addr(struct rdma_id_private *id_priv)
283 return (struct sockaddr *) &id_priv->id.route.addr.src_addr;
286 static inline struct sockaddr *cma_dst_addr(struct rdma_id_private *id_priv)
288 return (struct sockaddr *) &id_priv->id.route.addr.dst_addr;
291 static inline unsigned short cma_family(struct rdma_id_private *id_priv)
293 return id_priv->id.route.addr.src_addr.ss_family;
296 static int cma_set_qkey(struct rdma_id_private *id_priv, u32 qkey)
298 struct ib_sa_mcmember_rec rec;
302 if (qkey && id_priv->qkey != qkey)
308 id_priv->qkey = qkey;
312 switch (id_priv->id.ps) {
315 id_priv->qkey = RDMA_UDP_QKEY;
318 ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid);
319 ret = ib_sa_get_mcmember_rec(id_priv->id.device,
320 id_priv->id.port_num, &rec.mgid,
323 id_priv->qkey = be32_to_cpu(rec.qkey);
331 static int find_gid_port(struct ib_device *device, union ib_gid *gid, u8 port_num)
335 struct ib_port_attr props;
338 err = ib_query_port(device, port_num, &props);
342 for (i = 0; i < props.gid_tbl_len; ++i) {
343 err = ib_query_gid(device, port_num, i, &tmp);
346 if (!memcmp(&tmp, gid, sizeof tmp))
350 return -EADDRNOTAVAIL;
353 static void cma_translate_ib(struct sockaddr_ib *sib, struct rdma_dev_addr *dev_addr)
355 dev_addr->dev_type = ARPHRD_INFINIBAND;
356 rdma_addr_set_sgid(dev_addr, (union ib_gid *) &sib->sib_addr);
357 ib_addr_set_pkey(dev_addr, ntohs(sib->sib_pkey));
360 static int cma_translate_addr(struct sockaddr *addr, struct rdma_dev_addr *dev_addr)
364 if (addr->sa_family != AF_IB) {
365 ret = rdma_translate_ip(addr, dev_addr);
367 cma_translate_ib((struct sockaddr_ib *) addr, dev_addr);
374 static int cma_acquire_dev(struct rdma_id_private *id_priv)
376 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
377 struct cma_device *cma_dev;
378 union ib_gid gid, iboe_gid;
381 enum rdma_link_layer dev_ll = dev_addr->dev_type == ARPHRD_INFINIBAND ?
382 IB_LINK_LAYER_INFINIBAND : IB_LINK_LAYER_ETHERNET;
384 if (dev_ll != IB_LINK_LAYER_INFINIBAND &&
385 id_priv->id.ps == RDMA_PS_IPOIB)
389 iboe_addr_get_sgid(dev_addr, &iboe_gid);
390 memcpy(&gid, dev_addr->src_dev_addr +
391 rdma_addr_gid_offset(dev_addr), sizeof gid);
392 list_for_each_entry(cma_dev, &dev_list, list) {
393 for (port = 1; port <= cma_dev->device->phys_port_cnt; ++port) {
394 if (rdma_port_get_link_layer(cma_dev->device, port) == dev_ll) {
395 if (rdma_node_get_transport(cma_dev->device->node_type) == RDMA_TRANSPORT_IB &&
396 rdma_port_get_link_layer(cma_dev->device, port) == IB_LINK_LAYER_ETHERNET)
397 ret = find_gid_port(cma_dev->device, &iboe_gid, port);
399 ret = find_gid_port(cma_dev->device, &gid, port);
402 id_priv->id.port_num = port;
411 cma_attach_to_dev(id_priv, cma_dev);
418 * Select the source IB device and address to reach the destination IB address.
420 static int cma_resolve_ib_dev(struct rdma_id_private *id_priv)
422 struct cma_device *cma_dev, *cur_dev;
423 struct sockaddr_ib *addr;
424 union ib_gid gid, sgid, *dgid;
430 addr = (struct sockaddr_ib *) cma_dst_addr(id_priv);
431 dgid = (union ib_gid *) &addr->sib_addr;
432 pkey = ntohs(addr->sib_pkey);
434 list_for_each_entry(cur_dev, &dev_list, list) {
435 if (rdma_node_get_transport(cur_dev->device->node_type) != RDMA_TRANSPORT_IB)
438 for (p = 1; p <= cur_dev->device->phys_port_cnt; ++p) {
439 if (ib_find_cached_pkey(cur_dev->device, p, pkey, &index))
442 for (i = 0; !ib_get_cached_gid(cur_dev->device, p, i, &gid); i++) {
443 if (!memcmp(&gid, dgid, sizeof(gid))) {
450 if (!cma_dev && (gid.global.subnet_prefix ==
451 dgid->global.subnet_prefix)) {
464 cma_attach_to_dev(id_priv, cma_dev);
465 id_priv->id.port_num = port;
466 addr = (struct sockaddr_ib *) cma_src_addr(id_priv);
467 memcpy(&addr->sib_addr, &sgid, sizeof sgid);
468 cma_translate_ib(addr, &id_priv->id.route.addr.dev_addr);
472 static void cma_deref_id(struct rdma_id_private *id_priv)
474 if (atomic_dec_and_test(&id_priv->refcount))
475 complete(&id_priv->comp);
478 static int cma_disable_callback(struct rdma_id_private *id_priv,
479 enum rdma_cm_state state)
481 mutex_lock(&id_priv->handler_mutex);
482 if (id_priv->state != state) {
483 mutex_unlock(&id_priv->handler_mutex);
489 struct rdma_cm_id *rdma_create_id(rdma_cm_event_handler event_handler,
490 void *context, enum rdma_port_space ps,
491 enum ib_qp_type qp_type)
493 struct rdma_id_private *id_priv;
495 id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
497 return ERR_PTR(-ENOMEM);
499 id_priv->owner = task_pid_nr(current);
500 id_priv->state = RDMA_CM_IDLE;
501 id_priv->id.context = context;
502 id_priv->id.event_handler = event_handler;
504 id_priv->id.qp_type = qp_type;
505 spin_lock_init(&id_priv->lock);
506 mutex_init(&id_priv->qp_mutex);
507 init_completion(&id_priv->comp);
508 atomic_set(&id_priv->refcount, 1);
509 mutex_init(&id_priv->handler_mutex);
510 INIT_LIST_HEAD(&id_priv->listen_list);
511 INIT_LIST_HEAD(&id_priv->mc_list);
512 get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
516 EXPORT_SYMBOL(rdma_create_id);
518 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
520 struct ib_qp_attr qp_attr;
521 int qp_attr_mask, ret;
523 qp_attr.qp_state = IB_QPS_INIT;
524 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
528 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
532 qp_attr.qp_state = IB_QPS_RTR;
533 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
537 qp_attr.qp_state = IB_QPS_RTS;
539 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
544 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
546 struct ib_qp_attr qp_attr;
547 int qp_attr_mask, ret;
549 qp_attr.qp_state = IB_QPS_INIT;
550 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
554 return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
557 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
558 struct ib_qp_init_attr *qp_init_attr)
560 struct rdma_id_private *id_priv;
564 id_priv = container_of(id, struct rdma_id_private, id);
565 if (id->device != pd->device)
568 qp = ib_create_qp(pd, qp_init_attr);
572 if (id->qp_type == IB_QPT_UD)
573 ret = cma_init_ud_qp(id_priv, qp);
575 ret = cma_init_conn_qp(id_priv, qp);
580 id_priv->qp_num = qp->qp_num;
581 id_priv->srq = (qp->srq != NULL);
587 EXPORT_SYMBOL(rdma_create_qp);
589 void rdma_destroy_qp(struct rdma_cm_id *id)
591 struct rdma_id_private *id_priv;
593 id_priv = container_of(id, struct rdma_id_private, id);
594 mutex_lock(&id_priv->qp_mutex);
595 ib_destroy_qp(id_priv->id.qp);
596 id_priv->id.qp = NULL;
597 mutex_unlock(&id_priv->qp_mutex);
599 EXPORT_SYMBOL(rdma_destroy_qp);
601 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
602 struct rdma_conn_param *conn_param)
604 struct ib_qp_attr qp_attr;
605 int qp_attr_mask, ret;
607 mutex_lock(&id_priv->qp_mutex);
608 if (!id_priv->id.qp) {
613 /* Need to update QP attributes from default values. */
614 qp_attr.qp_state = IB_QPS_INIT;
615 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
619 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
623 qp_attr.qp_state = IB_QPS_RTR;
624 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
629 qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
630 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
632 mutex_unlock(&id_priv->qp_mutex);
636 static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
637 struct rdma_conn_param *conn_param)
639 struct ib_qp_attr qp_attr;
640 int qp_attr_mask, ret;
642 mutex_lock(&id_priv->qp_mutex);
643 if (!id_priv->id.qp) {
648 qp_attr.qp_state = IB_QPS_RTS;
649 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
654 qp_attr.max_rd_atomic = conn_param->initiator_depth;
655 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
657 mutex_unlock(&id_priv->qp_mutex);
661 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
663 struct ib_qp_attr qp_attr;
666 mutex_lock(&id_priv->qp_mutex);
667 if (!id_priv->id.qp) {
672 qp_attr.qp_state = IB_QPS_ERR;
673 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
675 mutex_unlock(&id_priv->qp_mutex);
679 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
680 struct ib_qp_attr *qp_attr, int *qp_attr_mask)
682 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
686 if (rdma_port_get_link_layer(id_priv->id.device, id_priv->id.port_num) ==
687 IB_LINK_LAYER_INFINIBAND)
688 pkey = ib_addr_get_pkey(dev_addr);
692 ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
693 pkey, &qp_attr->pkey_index);
697 qp_attr->port_num = id_priv->id.port_num;
698 *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
700 if (id_priv->id.qp_type == IB_QPT_UD) {
701 ret = cma_set_qkey(id_priv, 0);
705 qp_attr->qkey = id_priv->qkey;
706 *qp_attr_mask |= IB_QP_QKEY;
708 qp_attr->qp_access_flags = 0;
709 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
714 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
717 struct rdma_id_private *id_priv;
720 id_priv = container_of(id, struct rdma_id_private, id);
721 switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
722 case RDMA_TRANSPORT_IB:
723 if (!id_priv->cm_id.ib || (id_priv->id.qp_type == IB_QPT_UD))
724 ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
726 ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
728 if (qp_attr->qp_state == IB_QPS_RTR)
729 qp_attr->rq_psn = id_priv->seq_num;
731 case RDMA_TRANSPORT_IWARP:
732 if (!id_priv->cm_id.iw) {
733 qp_attr->qp_access_flags = 0;
734 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
736 ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
746 EXPORT_SYMBOL(rdma_init_qp_attr);
748 static inline int cma_zero_addr(struct sockaddr *addr)
750 switch (addr->sa_family) {
752 return ipv4_is_zeronet(((struct sockaddr_in *)addr)->sin_addr.s_addr);
754 return ipv6_addr_any(&((struct sockaddr_in6 *) addr)->sin6_addr);
756 return ib_addr_any(&((struct sockaddr_ib *) addr)->sib_addr);
762 static inline int cma_loopback_addr(struct sockaddr *addr)
764 switch (addr->sa_family) {
766 return ipv4_is_loopback(((struct sockaddr_in *) addr)->sin_addr.s_addr);
768 return ipv6_addr_loopback(&((struct sockaddr_in6 *) addr)->sin6_addr);
770 return ib_addr_loopback(&((struct sockaddr_ib *) addr)->sib_addr);
776 static inline int cma_any_addr(struct sockaddr *addr)
778 return cma_zero_addr(addr) || cma_loopback_addr(addr);
781 static int cma_addr_cmp(struct sockaddr *src, struct sockaddr *dst)
783 if (src->sa_family != dst->sa_family)
786 switch (src->sa_family) {
788 return ((struct sockaddr_in *) src)->sin_addr.s_addr !=
789 ((struct sockaddr_in *) dst)->sin_addr.s_addr;
791 return ipv6_addr_cmp(&((struct sockaddr_in6 *) src)->sin6_addr,
792 &((struct sockaddr_in6 *) dst)->sin6_addr);
794 return ib_addr_cmp(&((struct sockaddr_ib *) src)->sib_addr,
795 &((struct sockaddr_ib *) dst)->sib_addr);
799 static __be16 cma_port(struct sockaddr *addr)
801 struct sockaddr_ib *sib;
803 switch (addr->sa_family) {
805 return ((struct sockaddr_in *) addr)->sin_port;
807 return ((struct sockaddr_in6 *) addr)->sin6_port;
809 sib = (struct sockaddr_ib *) addr;
810 return htons((u16) (be64_to_cpu(sib->sib_sid) &
811 be64_to_cpu(sib->sib_sid_mask)));
817 static inline int cma_any_port(struct sockaddr *addr)
819 return !cma_port(addr);
822 static void cma_save_ib_info(struct rdma_cm_id *id, struct rdma_cm_id *listen_id,
823 struct ib_sa_path_rec *path)
825 struct sockaddr_ib *listen_ib, *ib;
827 listen_ib = (struct sockaddr_ib *) &listen_id->route.addr.src_addr;
828 ib = (struct sockaddr_ib *) &id->route.addr.src_addr;
829 ib->sib_family = listen_ib->sib_family;
830 ib->sib_pkey = path->pkey;
831 ib->sib_flowinfo = path->flow_label;
832 memcpy(&ib->sib_addr, &path->sgid, 16);
833 ib->sib_sid = listen_ib->sib_sid;
834 ib->sib_sid_mask = cpu_to_be64(0xffffffffffffffffULL);
835 ib->sib_scope_id = listen_ib->sib_scope_id;
837 ib = (struct sockaddr_ib *) &id->route.addr.dst_addr;
838 ib->sib_family = listen_ib->sib_family;
839 ib->sib_pkey = path->pkey;
840 ib->sib_flowinfo = path->flow_label;
841 memcpy(&ib->sib_addr, &path->dgid, 16);
844 static void cma_save_ip4_info(struct rdma_cm_id *id, struct rdma_cm_id *listen_id,
847 struct sockaddr_in *listen4, *ip4;
849 listen4 = (struct sockaddr_in *) &listen_id->route.addr.src_addr;
850 ip4 = (struct sockaddr_in *) &id->route.addr.src_addr;
851 ip4->sin_family = listen4->sin_family;
852 ip4->sin_addr.s_addr = hdr->dst_addr.ip4.addr;
853 ip4->sin_port = listen4->sin_port;
855 ip4 = (struct sockaddr_in *) &id->route.addr.dst_addr;
856 ip4->sin_family = listen4->sin_family;
857 ip4->sin_addr.s_addr = hdr->src_addr.ip4.addr;
858 ip4->sin_port = hdr->port;
861 static void cma_save_ip6_info(struct rdma_cm_id *id, struct rdma_cm_id *listen_id,
864 struct sockaddr_in6 *listen6, *ip6;
866 listen6 = (struct sockaddr_in6 *) &listen_id->route.addr.src_addr;
867 ip6 = (struct sockaddr_in6 *) &id->route.addr.src_addr;
868 ip6->sin6_family = listen6->sin6_family;
869 ip6->sin6_addr = hdr->dst_addr.ip6;
870 ip6->sin6_port = listen6->sin6_port;
872 ip6 = (struct sockaddr_in6 *) &id->route.addr.dst_addr;
873 ip6->sin6_family = listen6->sin6_family;
874 ip6->sin6_addr = hdr->src_addr.ip6;
875 ip6->sin6_port = hdr->port;
878 static int cma_save_net_info(struct rdma_cm_id *id, struct rdma_cm_id *listen_id,
879 struct ib_cm_event *ib_event)
883 if (listen_id->route.addr.src_addr.ss_family == AF_IB) {
884 cma_save_ib_info(id, listen_id, ib_event->param.req_rcvd.primary_path);
888 hdr = ib_event->private_data;
889 if (hdr->cma_version != CMA_VERSION)
892 switch (cma_get_ip_ver(hdr)) {
894 cma_save_ip4_info(id, listen_id, hdr);
897 cma_save_ip6_info(id, listen_id, hdr);
905 static inline int cma_user_data_offset(struct rdma_id_private *id_priv)
907 return cma_family(id_priv) == AF_IB ? 0 : sizeof(struct cma_hdr);
910 static void cma_cancel_route(struct rdma_id_private *id_priv)
912 switch (rdma_port_get_link_layer(id_priv->id.device, id_priv->id.port_num)) {
913 case IB_LINK_LAYER_INFINIBAND:
915 ib_sa_cancel_query(id_priv->query_id, id_priv->query);
922 static void cma_cancel_listens(struct rdma_id_private *id_priv)
924 struct rdma_id_private *dev_id_priv;
927 * Remove from listen_any_list to prevent added devices from spawning
928 * additional listen requests.
931 list_del(&id_priv->list);
933 while (!list_empty(&id_priv->listen_list)) {
934 dev_id_priv = list_entry(id_priv->listen_list.next,
935 struct rdma_id_private, listen_list);
936 /* sync with device removal to avoid duplicate destruction */
937 list_del_init(&dev_id_priv->list);
938 list_del(&dev_id_priv->listen_list);
941 rdma_destroy_id(&dev_id_priv->id);
947 static void cma_cancel_operation(struct rdma_id_private *id_priv,
948 enum rdma_cm_state state)
951 case RDMA_CM_ADDR_QUERY:
952 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
954 case RDMA_CM_ROUTE_QUERY:
955 cma_cancel_route(id_priv);
958 if (cma_any_addr(cma_src_addr(id_priv)) && !id_priv->cma_dev)
959 cma_cancel_listens(id_priv);
966 static void cma_release_port(struct rdma_id_private *id_priv)
968 struct rdma_bind_list *bind_list = id_priv->bind_list;
974 hlist_del(&id_priv->node);
975 if (hlist_empty(&bind_list->owners)) {
976 idr_remove(bind_list->ps, bind_list->port);
982 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
984 struct cma_multicast *mc;
986 while (!list_empty(&id_priv->mc_list)) {
987 mc = container_of(id_priv->mc_list.next,
988 struct cma_multicast, list);
990 switch (rdma_port_get_link_layer(id_priv->cma_dev->device, id_priv->id.port_num)) {
991 case IB_LINK_LAYER_INFINIBAND:
992 ib_sa_free_multicast(mc->multicast.ib);
995 case IB_LINK_LAYER_ETHERNET:
996 kref_put(&mc->mcref, release_mc);
1004 void rdma_destroy_id(struct rdma_cm_id *id)
1006 struct rdma_id_private *id_priv;
1007 enum rdma_cm_state state;
1009 id_priv = container_of(id, struct rdma_id_private, id);
1010 state = cma_exch(id_priv, RDMA_CM_DESTROYING);
1011 cma_cancel_operation(id_priv, state);
1014 * Wait for any active callback to finish. New callbacks will find
1015 * the id_priv state set to destroying and abort.
1017 mutex_lock(&id_priv->handler_mutex);
1018 mutex_unlock(&id_priv->handler_mutex);
1020 if (id_priv->cma_dev) {
1021 switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
1022 case RDMA_TRANSPORT_IB:
1023 if (id_priv->cm_id.ib)
1024 ib_destroy_cm_id(id_priv->cm_id.ib);
1026 case RDMA_TRANSPORT_IWARP:
1027 if (id_priv->cm_id.iw)
1028 iw_destroy_cm_id(id_priv->cm_id.iw);
1033 cma_leave_mc_groups(id_priv);
1034 cma_release_dev(id_priv);
1037 cma_release_port(id_priv);
1038 cma_deref_id(id_priv);
1039 wait_for_completion(&id_priv->comp);
1041 if (id_priv->internal_id)
1042 cma_deref_id(id_priv->id.context);
1044 kfree(id_priv->id.route.path_rec);
1047 EXPORT_SYMBOL(rdma_destroy_id);
1049 static int cma_rep_recv(struct rdma_id_private *id_priv)
1053 ret = cma_modify_qp_rtr(id_priv, NULL);
1057 ret = cma_modify_qp_rts(id_priv, NULL);
1061 ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
1067 cma_modify_qp_err(id_priv);
1068 ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
1073 static void cma_set_rep_event_data(struct rdma_cm_event *event,
1074 struct ib_cm_rep_event_param *rep_data,
1077 event->param.conn.private_data = private_data;
1078 event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
1079 event->param.conn.responder_resources = rep_data->responder_resources;
1080 event->param.conn.initiator_depth = rep_data->initiator_depth;
1081 event->param.conn.flow_control = rep_data->flow_control;
1082 event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
1083 event->param.conn.srq = rep_data->srq;
1084 event->param.conn.qp_num = rep_data->remote_qpn;
1087 static int cma_ib_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1089 struct rdma_id_private *id_priv = cm_id->context;
1090 struct rdma_cm_event event;
1093 if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
1094 cma_disable_callback(id_priv, RDMA_CM_CONNECT)) ||
1095 (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
1096 cma_disable_callback(id_priv, RDMA_CM_DISCONNECT)))
1099 memset(&event, 0, sizeof event);
1100 switch (ib_event->event) {
1101 case IB_CM_REQ_ERROR:
1102 case IB_CM_REP_ERROR:
1103 event.event = RDMA_CM_EVENT_UNREACHABLE;
1104 event.status = -ETIMEDOUT;
1106 case IB_CM_REP_RECEIVED:
1107 if (id_priv->id.qp) {
1108 event.status = cma_rep_recv(id_priv);
1109 event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
1110 RDMA_CM_EVENT_ESTABLISHED;
1112 event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
1114 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
1115 ib_event->private_data);
1117 case IB_CM_RTU_RECEIVED:
1118 case IB_CM_USER_ESTABLISHED:
1119 event.event = RDMA_CM_EVENT_ESTABLISHED;
1121 case IB_CM_DREQ_ERROR:
1122 event.status = -ETIMEDOUT; /* fall through */
1123 case IB_CM_DREQ_RECEIVED:
1124 case IB_CM_DREP_RECEIVED:
1125 if (!cma_comp_exch(id_priv, RDMA_CM_CONNECT,
1126 RDMA_CM_DISCONNECT))
1128 event.event = RDMA_CM_EVENT_DISCONNECTED;
1130 case IB_CM_TIMEWAIT_EXIT:
1131 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
1133 case IB_CM_MRA_RECEIVED:
1136 case IB_CM_REJ_RECEIVED:
1137 cma_modify_qp_err(id_priv);
1138 event.status = ib_event->param.rej_rcvd.reason;
1139 event.event = RDMA_CM_EVENT_REJECTED;
1140 event.param.conn.private_data = ib_event->private_data;
1141 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
1144 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
1149 ret = id_priv->id.event_handler(&id_priv->id, &event);
1151 /* Destroy the CM ID by returning a non-zero value. */
1152 id_priv->cm_id.ib = NULL;
1153 cma_exch(id_priv, RDMA_CM_DESTROYING);
1154 mutex_unlock(&id_priv->handler_mutex);
1155 rdma_destroy_id(&id_priv->id);
1159 mutex_unlock(&id_priv->handler_mutex);
1163 static struct rdma_id_private *cma_new_conn_id(struct rdma_cm_id *listen_id,
1164 struct ib_cm_event *ib_event)
1166 struct rdma_id_private *id_priv;
1167 struct rdma_cm_id *id;
1168 struct rdma_route *rt;
1171 id = rdma_create_id(listen_id->event_handler, listen_id->context,
1172 listen_id->ps, ib_event->param.req_rcvd.qp_type);
1176 id_priv = container_of(id, struct rdma_id_private, id);
1177 if (cma_save_net_info(id, listen_id, ib_event))
1181 rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
1182 rt->path_rec = kmalloc(sizeof *rt->path_rec * rt->num_paths,
1187 rt->path_rec[0] = *ib_event->param.req_rcvd.primary_path;
1188 if (rt->num_paths == 2)
1189 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
1191 if (cma_any_addr(cma_src_addr(id_priv))) {
1192 rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
1193 rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
1194 ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
1196 ret = cma_translate_addr(cma_src_addr(id_priv), &rt->addr.dev_addr);
1200 rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
1202 id_priv->state = RDMA_CM_CONNECT;
1206 rdma_destroy_id(id);
1210 static struct rdma_id_private *cma_new_udp_id(struct rdma_cm_id *listen_id,
1211 struct ib_cm_event *ib_event)
1213 struct rdma_id_private *id_priv;
1214 struct rdma_cm_id *id;
1217 id = rdma_create_id(listen_id->event_handler, listen_id->context,
1218 listen_id->ps, IB_QPT_UD);
1222 id_priv = container_of(id, struct rdma_id_private, id);
1223 if (cma_save_net_info(id, listen_id, ib_event))
1226 if (!cma_any_addr((struct sockaddr *) &id->route.addr.src_addr)) {
1227 ret = cma_translate_addr(cma_src_addr(id_priv), &id->route.addr.dev_addr);
1232 id_priv->state = RDMA_CM_CONNECT;
1235 rdma_destroy_id(id);
1239 static void cma_set_req_event_data(struct rdma_cm_event *event,
1240 struct ib_cm_req_event_param *req_data,
1241 void *private_data, int offset)
1243 event->param.conn.private_data = private_data + offset;
1244 event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
1245 event->param.conn.responder_resources = req_data->responder_resources;
1246 event->param.conn.initiator_depth = req_data->initiator_depth;
1247 event->param.conn.flow_control = req_data->flow_control;
1248 event->param.conn.retry_count = req_data->retry_count;
1249 event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
1250 event->param.conn.srq = req_data->srq;
1251 event->param.conn.qp_num = req_data->remote_qpn;
1254 static int cma_check_req_qp_type(struct rdma_cm_id *id, struct ib_cm_event *ib_event)
1256 return (((ib_event->event == IB_CM_REQ_RECEIVED) &&
1257 (ib_event->param.req_rcvd.qp_type == id->qp_type)) ||
1258 ((ib_event->event == IB_CM_SIDR_REQ_RECEIVED) &&
1259 (id->qp_type == IB_QPT_UD)) ||
1263 static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1265 struct rdma_id_private *listen_id, *conn_id;
1266 struct rdma_cm_event event;
1269 listen_id = cm_id->context;
1270 if (!cma_check_req_qp_type(&listen_id->id, ib_event))
1273 if (cma_disable_callback(listen_id, RDMA_CM_LISTEN))
1274 return -ECONNABORTED;
1276 memset(&event, 0, sizeof event);
1277 offset = cma_user_data_offset(listen_id);
1278 event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1279 if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) {
1280 conn_id = cma_new_udp_id(&listen_id->id, ib_event);
1281 event.param.ud.private_data = ib_event->private_data + offset;
1282 event.param.ud.private_data_len =
1283 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
1285 conn_id = cma_new_conn_id(&listen_id->id, ib_event);
1286 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
1287 ib_event->private_data, offset);
1294 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1295 ret = cma_acquire_dev(conn_id);
1299 conn_id->cm_id.ib = cm_id;
1300 cm_id->context = conn_id;
1301 cm_id->cm_handler = cma_ib_handler;
1304 * Protect against the user destroying conn_id from another thread
1305 * until we're done accessing it.
1307 atomic_inc(&conn_id->refcount);
1308 ret = conn_id->id.event_handler(&conn_id->id, &event);
1313 * Acquire mutex to prevent user executing rdma_destroy_id()
1314 * while we're accessing the cm_id.
1317 if (cma_comp(conn_id, RDMA_CM_CONNECT) && (conn_id->id.qp_type != IB_QPT_UD))
1318 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
1319 mutex_unlock(&lock);
1320 mutex_unlock(&conn_id->handler_mutex);
1321 mutex_unlock(&listen_id->handler_mutex);
1322 cma_deref_id(conn_id);
1326 cma_deref_id(conn_id);
1327 /* Destroy the CM ID by returning a non-zero value. */
1328 conn_id->cm_id.ib = NULL;
1330 cma_exch(conn_id, RDMA_CM_DESTROYING);
1331 mutex_unlock(&conn_id->handler_mutex);
1333 mutex_unlock(&listen_id->handler_mutex);
1335 rdma_destroy_id(&conn_id->id);
1339 __be64 rdma_get_service_id(struct rdma_cm_id *id, struct sockaddr *addr)
1341 if (addr->sa_family == AF_IB)
1342 return ((struct sockaddr_ib *) addr)->sib_sid;
1344 return cpu_to_be64(((u64)id->ps << 16) + be16_to_cpu(cma_port(addr)));
1346 EXPORT_SYMBOL(rdma_get_service_id);
1348 static void cma_set_compare_data(enum rdma_port_space ps, struct sockaddr *addr,
1349 struct ib_cm_compare_data *compare)
1351 struct cma_hdr *cma_data, *cma_mask;
1353 struct in6_addr ip6_addr;
1355 memset(compare, 0, sizeof *compare);
1356 cma_data = (void *) compare->data;
1357 cma_mask = (void *) compare->mask;
1359 switch (addr->sa_family) {
1361 ip4_addr = ((struct sockaddr_in *) addr)->sin_addr.s_addr;
1362 cma_set_ip_ver(cma_data, 4);
1363 cma_set_ip_ver(cma_mask, 0xF);
1364 if (!cma_any_addr(addr)) {
1365 cma_data->dst_addr.ip4.addr = ip4_addr;
1366 cma_mask->dst_addr.ip4.addr = htonl(~0);
1370 ip6_addr = ((struct sockaddr_in6 *) addr)->sin6_addr;
1371 cma_set_ip_ver(cma_data, 6);
1372 cma_set_ip_ver(cma_mask, 0xF);
1373 if (!cma_any_addr(addr)) {
1374 cma_data->dst_addr.ip6 = ip6_addr;
1375 memset(&cma_mask->dst_addr.ip6, 0xFF,
1376 sizeof cma_mask->dst_addr.ip6);
1384 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
1386 struct rdma_id_private *id_priv = iw_id->context;
1387 struct rdma_cm_event event;
1388 struct sockaddr_in *sin;
1391 if (cma_disable_callback(id_priv, RDMA_CM_CONNECT))
1394 memset(&event, 0, sizeof event);
1395 switch (iw_event->event) {
1396 case IW_CM_EVENT_CLOSE:
1397 event.event = RDMA_CM_EVENT_DISCONNECTED;
1399 case IW_CM_EVENT_CONNECT_REPLY:
1400 sin = (struct sockaddr_in *) cma_src_addr(id_priv);
1401 *sin = iw_event->local_addr;
1402 sin = (struct sockaddr_in *) cma_dst_addr(id_priv);
1403 *sin = iw_event->remote_addr;
1404 switch (iw_event->status) {
1406 event.event = RDMA_CM_EVENT_ESTABLISHED;
1407 event.param.conn.initiator_depth = iw_event->ird;
1408 event.param.conn.responder_resources = iw_event->ord;
1412 event.event = RDMA_CM_EVENT_REJECTED;
1415 event.event = RDMA_CM_EVENT_UNREACHABLE;
1418 event.event = RDMA_CM_EVENT_CONNECT_ERROR;
1422 case IW_CM_EVENT_ESTABLISHED:
1423 event.event = RDMA_CM_EVENT_ESTABLISHED;
1424 event.param.conn.initiator_depth = iw_event->ird;
1425 event.param.conn.responder_resources = iw_event->ord;
1431 event.status = iw_event->status;
1432 event.param.conn.private_data = iw_event->private_data;
1433 event.param.conn.private_data_len = iw_event->private_data_len;
1434 ret = id_priv->id.event_handler(&id_priv->id, &event);
1436 /* Destroy the CM ID by returning a non-zero value. */
1437 id_priv->cm_id.iw = NULL;
1438 cma_exch(id_priv, RDMA_CM_DESTROYING);
1439 mutex_unlock(&id_priv->handler_mutex);
1440 rdma_destroy_id(&id_priv->id);
1444 mutex_unlock(&id_priv->handler_mutex);
1448 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
1449 struct iw_cm_event *iw_event)
1451 struct rdma_cm_id *new_cm_id;
1452 struct rdma_id_private *listen_id, *conn_id;
1453 struct sockaddr_in *sin;
1454 struct net_device *dev = NULL;
1455 struct rdma_cm_event event;
1457 struct ib_device_attr attr;
1459 listen_id = cm_id->context;
1460 if (cma_disable_callback(listen_id, RDMA_CM_LISTEN))
1461 return -ECONNABORTED;
1463 /* Create a new RDMA id for the new IW CM ID */
1464 new_cm_id = rdma_create_id(listen_id->id.event_handler,
1465 listen_id->id.context,
1466 RDMA_PS_TCP, IB_QPT_RC);
1467 if (IS_ERR(new_cm_id)) {
1471 conn_id = container_of(new_cm_id, struct rdma_id_private, id);
1472 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1473 conn_id->state = RDMA_CM_CONNECT;
1475 dev = ip_dev_find(&init_net, iw_event->local_addr.sin_addr.s_addr);
1477 ret = -EADDRNOTAVAIL;
1478 mutex_unlock(&conn_id->handler_mutex);
1479 rdma_destroy_id(new_cm_id);
1482 ret = rdma_copy_addr(&conn_id->id.route.addr.dev_addr, dev, NULL);
1484 mutex_unlock(&conn_id->handler_mutex);
1485 rdma_destroy_id(new_cm_id);
1489 ret = cma_acquire_dev(conn_id);
1491 mutex_unlock(&conn_id->handler_mutex);
1492 rdma_destroy_id(new_cm_id);
1496 conn_id->cm_id.iw = cm_id;
1497 cm_id->context = conn_id;
1498 cm_id->cm_handler = cma_iw_handler;
1500 sin = (struct sockaddr_in *) cma_src_addr(conn_id);
1501 *sin = iw_event->local_addr;
1502 sin = (struct sockaddr_in *) cma_dst_addr(conn_id);
1503 *sin = iw_event->remote_addr;
1505 ret = ib_query_device(conn_id->id.device, &attr);
1507 mutex_unlock(&conn_id->handler_mutex);
1508 rdma_destroy_id(new_cm_id);
1512 memset(&event, 0, sizeof event);
1513 event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1514 event.param.conn.private_data = iw_event->private_data;
1515 event.param.conn.private_data_len = iw_event->private_data_len;
1516 event.param.conn.initiator_depth = iw_event->ird;
1517 event.param.conn.responder_resources = iw_event->ord;
1520 * Protect against the user destroying conn_id from another thread
1521 * until we're done accessing it.
1523 atomic_inc(&conn_id->refcount);
1524 ret = conn_id->id.event_handler(&conn_id->id, &event);
1526 /* User wants to destroy the CM ID */
1527 conn_id->cm_id.iw = NULL;
1528 cma_exch(conn_id, RDMA_CM_DESTROYING);
1529 mutex_unlock(&conn_id->handler_mutex);
1530 cma_deref_id(conn_id);
1531 rdma_destroy_id(&conn_id->id);
1535 mutex_unlock(&conn_id->handler_mutex);
1536 cma_deref_id(conn_id);
1541 mutex_unlock(&listen_id->handler_mutex);
1545 static int cma_ib_listen(struct rdma_id_private *id_priv)
1547 struct ib_cm_compare_data compare_data;
1548 struct sockaddr *addr;
1549 struct ib_cm_id *id;
1553 id = ib_create_cm_id(id_priv->id.device, cma_req_handler, id_priv);
1557 id_priv->cm_id.ib = id;
1559 addr = cma_src_addr(id_priv);
1560 svc_id = rdma_get_service_id(&id_priv->id, addr);
1561 if (cma_any_addr(addr) && !id_priv->afonly)
1562 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, NULL);
1564 cma_set_compare_data(id_priv->id.ps, addr, &compare_data);
1565 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, &compare_data);
1569 ib_destroy_cm_id(id_priv->cm_id.ib);
1570 id_priv->cm_id.ib = NULL;
1576 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
1579 struct sockaddr_in *sin;
1580 struct iw_cm_id *id;
1582 id = iw_create_cm_id(id_priv->id.device,
1583 iw_conn_req_handler,
1588 id_priv->cm_id.iw = id;
1590 sin = (struct sockaddr_in *) cma_src_addr(id_priv);
1591 id_priv->cm_id.iw->local_addr = *sin;
1593 ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
1596 iw_destroy_cm_id(id_priv->cm_id.iw);
1597 id_priv->cm_id.iw = NULL;
1603 static int cma_listen_handler(struct rdma_cm_id *id,
1604 struct rdma_cm_event *event)
1606 struct rdma_id_private *id_priv = id->context;
1608 id->context = id_priv->id.context;
1609 id->event_handler = id_priv->id.event_handler;
1610 return id_priv->id.event_handler(id, event);
1613 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
1614 struct cma_device *cma_dev)
1616 struct rdma_id_private *dev_id_priv;
1617 struct rdma_cm_id *id;
1620 if (cma_family(id_priv) == AF_IB &&
1621 rdma_node_get_transport(cma_dev->device->node_type) != RDMA_TRANSPORT_IB)
1624 id = rdma_create_id(cma_listen_handler, id_priv, id_priv->id.ps,
1625 id_priv->id.qp_type);
1629 dev_id_priv = container_of(id, struct rdma_id_private, id);
1631 dev_id_priv->state = RDMA_CM_ADDR_BOUND;
1632 memcpy(cma_src_addr(dev_id_priv), cma_src_addr(id_priv),
1633 rdma_addr_size(cma_src_addr(id_priv)));
1635 cma_attach_to_dev(dev_id_priv, cma_dev);
1636 list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
1637 atomic_inc(&id_priv->refcount);
1638 dev_id_priv->internal_id = 1;
1639 dev_id_priv->afonly = id_priv->afonly;
1641 ret = rdma_listen(id, id_priv->backlog);
1643 printk(KERN_WARNING "RDMA CMA: cma_listen_on_dev, error %d, "
1644 "listening on device %s\n", ret, cma_dev->device->name);
1647 static void cma_listen_on_all(struct rdma_id_private *id_priv)
1649 struct cma_device *cma_dev;
1652 list_add_tail(&id_priv->list, &listen_any_list);
1653 list_for_each_entry(cma_dev, &dev_list, list)
1654 cma_listen_on_dev(id_priv, cma_dev);
1655 mutex_unlock(&lock);
1658 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
1660 struct rdma_id_private *id_priv;
1662 id_priv = container_of(id, struct rdma_id_private, id);
1663 id_priv->tos = (u8) tos;
1665 EXPORT_SYMBOL(rdma_set_service_type);
1667 static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec,
1670 struct cma_work *work = context;
1671 struct rdma_route *route;
1673 route = &work->id->id.route;
1676 route->num_paths = 1;
1677 *route->path_rec = *path_rec;
1679 work->old_state = RDMA_CM_ROUTE_QUERY;
1680 work->new_state = RDMA_CM_ADDR_RESOLVED;
1681 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
1682 work->event.status = status;
1685 queue_work(cma_wq, &work->work);
1688 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms,
1689 struct cma_work *work)
1691 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
1692 struct ib_sa_path_rec path_rec;
1693 ib_sa_comp_mask comp_mask;
1694 struct sockaddr_in6 *sin6;
1695 struct sockaddr_ib *sib;
1697 memset(&path_rec, 0, sizeof path_rec);
1698 rdma_addr_get_sgid(dev_addr, &path_rec.sgid);
1699 rdma_addr_get_dgid(dev_addr, &path_rec.dgid);
1700 path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
1701 path_rec.numb_path = 1;
1702 path_rec.reversible = 1;
1703 path_rec.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
1705 comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
1706 IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
1707 IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
1709 switch (cma_family(id_priv)) {
1711 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
1712 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
1715 sin6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
1716 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
1717 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
1720 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
1721 path_rec.traffic_class = (u8) (be32_to_cpu(sib->sib_flowinfo) >> 20);
1722 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
1726 id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
1727 id_priv->id.port_num, &path_rec,
1728 comp_mask, timeout_ms,
1729 GFP_KERNEL, cma_query_handler,
1730 work, &id_priv->query);
1732 return (id_priv->query_id < 0) ? id_priv->query_id : 0;
1735 static void cma_work_handler(struct work_struct *_work)
1737 struct cma_work *work = container_of(_work, struct cma_work, work);
1738 struct rdma_id_private *id_priv = work->id;
1741 mutex_lock(&id_priv->handler_mutex);
1742 if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
1745 if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1746 cma_exch(id_priv, RDMA_CM_DESTROYING);
1750 mutex_unlock(&id_priv->handler_mutex);
1751 cma_deref_id(id_priv);
1753 rdma_destroy_id(&id_priv->id);
1757 static void cma_ndev_work_handler(struct work_struct *_work)
1759 struct cma_ndev_work *work = container_of(_work, struct cma_ndev_work, work);
1760 struct rdma_id_private *id_priv = work->id;
1763 mutex_lock(&id_priv->handler_mutex);
1764 if (id_priv->state == RDMA_CM_DESTROYING ||
1765 id_priv->state == RDMA_CM_DEVICE_REMOVAL)
1768 if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1769 cma_exch(id_priv, RDMA_CM_DESTROYING);
1774 mutex_unlock(&id_priv->handler_mutex);
1775 cma_deref_id(id_priv);
1777 rdma_destroy_id(&id_priv->id);
1781 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms)
1783 struct rdma_route *route = &id_priv->id.route;
1784 struct cma_work *work;
1787 work = kzalloc(sizeof *work, GFP_KERNEL);
1792 INIT_WORK(&work->work, cma_work_handler);
1793 work->old_state = RDMA_CM_ROUTE_QUERY;
1794 work->new_state = RDMA_CM_ROUTE_RESOLVED;
1795 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1797 route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
1798 if (!route->path_rec) {
1803 ret = cma_query_ib_route(id_priv, timeout_ms, work);
1809 kfree(route->path_rec);
1810 route->path_rec = NULL;
1816 int rdma_set_ib_paths(struct rdma_cm_id *id,
1817 struct ib_sa_path_rec *path_rec, int num_paths)
1819 struct rdma_id_private *id_priv;
1822 id_priv = container_of(id, struct rdma_id_private, id);
1823 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
1824 RDMA_CM_ROUTE_RESOLVED))
1827 id->route.path_rec = kmemdup(path_rec, sizeof *path_rec * num_paths,
1829 if (!id->route.path_rec) {
1834 id->route.num_paths = num_paths;
1837 cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED);
1840 EXPORT_SYMBOL(rdma_set_ib_paths);
1842 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms)
1844 struct cma_work *work;
1846 work = kzalloc(sizeof *work, GFP_KERNEL);
1851 INIT_WORK(&work->work, cma_work_handler);
1852 work->old_state = RDMA_CM_ROUTE_QUERY;
1853 work->new_state = RDMA_CM_ROUTE_RESOLVED;
1854 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1855 queue_work(cma_wq, &work->work);
1859 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
1861 struct rdma_route *route = &id_priv->id.route;
1862 struct rdma_addr *addr = &route->addr;
1863 struct cma_work *work;
1865 struct net_device *ndev = NULL;
1868 work = kzalloc(sizeof *work, GFP_KERNEL);
1873 INIT_WORK(&work->work, cma_work_handler);
1875 route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
1876 if (!route->path_rec) {
1881 route->num_paths = 1;
1883 if (addr->dev_addr.bound_dev_if)
1884 ndev = dev_get_by_index(&init_net, addr->dev_addr.bound_dev_if);
1890 vid = rdma_vlan_dev_vlan_id(ndev);
1892 iboe_mac_vlan_to_ll(&route->path_rec->sgid, addr->dev_addr.src_dev_addr, vid);
1893 iboe_mac_vlan_to_ll(&route->path_rec->dgid, addr->dev_addr.dst_dev_addr, vid);
1895 route->path_rec->hop_limit = 1;
1896 route->path_rec->reversible = 1;
1897 route->path_rec->pkey = cpu_to_be16(0xffff);
1898 route->path_rec->mtu_selector = IB_SA_EQ;
1899 route->path_rec->sl = netdev_get_prio_tc_map(
1900 ndev->priv_flags & IFF_802_1Q_VLAN ?
1901 vlan_dev_real_dev(ndev) : ndev,
1902 rt_tos2priority(id_priv->tos));
1904 route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
1905 route->path_rec->rate_selector = IB_SA_EQ;
1906 route->path_rec->rate = iboe_get_rate(ndev);
1908 route->path_rec->packet_life_time_selector = IB_SA_EQ;
1909 route->path_rec->packet_life_time = CMA_IBOE_PACKET_LIFETIME;
1910 if (!route->path_rec->mtu) {
1915 work->old_state = RDMA_CM_ROUTE_QUERY;
1916 work->new_state = RDMA_CM_ROUTE_RESOLVED;
1917 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1918 work->event.status = 0;
1920 queue_work(cma_wq, &work->work);
1925 kfree(route->path_rec);
1926 route->path_rec = NULL;
1932 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
1934 struct rdma_id_private *id_priv;
1937 id_priv = container_of(id, struct rdma_id_private, id);
1938 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY))
1941 atomic_inc(&id_priv->refcount);
1942 switch (rdma_node_get_transport(id->device->node_type)) {
1943 case RDMA_TRANSPORT_IB:
1944 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
1945 case IB_LINK_LAYER_INFINIBAND:
1946 ret = cma_resolve_ib_route(id_priv, timeout_ms);
1948 case IB_LINK_LAYER_ETHERNET:
1949 ret = cma_resolve_iboe_route(id_priv);
1955 case RDMA_TRANSPORT_IWARP:
1956 ret = cma_resolve_iw_route(id_priv, timeout_ms);
1967 cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED);
1968 cma_deref_id(id_priv);
1971 EXPORT_SYMBOL(rdma_resolve_route);
1973 static void cma_set_loopback(struct sockaddr *addr)
1975 switch (addr->sa_family) {
1977 ((struct sockaddr_in *) addr)->sin_addr.s_addr = htonl(INADDR_LOOPBACK);
1980 ipv6_addr_set(&((struct sockaddr_in6 *) addr)->sin6_addr,
1984 ib_addr_set(&((struct sockaddr_ib *) addr)->sib_addr,
1990 static int cma_bind_loopback(struct rdma_id_private *id_priv)
1992 struct cma_device *cma_dev, *cur_dev;
1993 struct ib_port_attr port_attr;
2001 list_for_each_entry(cur_dev, &dev_list, list) {
2002 if (cma_family(id_priv) == AF_IB &&
2003 rdma_node_get_transport(cur_dev->device->node_type) != RDMA_TRANSPORT_IB)
2009 for (p = 1; p <= cur_dev->device->phys_port_cnt; ++p) {
2010 if (!ib_query_port(cur_dev->device, p, &port_attr) &&
2011 port_attr.state == IB_PORT_ACTIVE) {
2026 ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid);
2030 ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
2034 id_priv->id.route.addr.dev_addr.dev_type =
2035 (rdma_port_get_link_layer(cma_dev->device, p) == IB_LINK_LAYER_INFINIBAND) ?
2036 ARPHRD_INFINIBAND : ARPHRD_ETHER;
2038 rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2039 ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
2040 id_priv->id.port_num = p;
2041 cma_attach_to_dev(id_priv, cma_dev);
2042 cma_set_loopback(cma_src_addr(id_priv));
2044 mutex_unlock(&lock);
2048 static void addr_handler(int status, struct sockaddr *src_addr,
2049 struct rdma_dev_addr *dev_addr, void *context)
2051 struct rdma_id_private *id_priv = context;
2052 struct rdma_cm_event event;
2054 memset(&event, 0, sizeof event);
2055 mutex_lock(&id_priv->handler_mutex);
2056 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY,
2057 RDMA_CM_ADDR_RESOLVED))
2060 if (!status && !id_priv->cma_dev)
2061 status = cma_acquire_dev(id_priv);
2064 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
2065 RDMA_CM_ADDR_BOUND))
2067 event.event = RDMA_CM_EVENT_ADDR_ERROR;
2068 event.status = status;
2070 memcpy(cma_src_addr(id_priv), src_addr, rdma_addr_size(src_addr));
2071 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2074 if (id_priv->id.event_handler(&id_priv->id, &event)) {
2075 cma_exch(id_priv, RDMA_CM_DESTROYING);
2076 mutex_unlock(&id_priv->handler_mutex);
2077 cma_deref_id(id_priv);
2078 rdma_destroy_id(&id_priv->id);
2082 mutex_unlock(&id_priv->handler_mutex);
2083 cma_deref_id(id_priv);
2086 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
2088 struct cma_work *work;
2092 work = kzalloc(sizeof *work, GFP_KERNEL);
2096 if (!id_priv->cma_dev) {
2097 ret = cma_bind_loopback(id_priv);
2102 rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2103 rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
2106 INIT_WORK(&work->work, cma_work_handler);
2107 work->old_state = RDMA_CM_ADDR_QUERY;
2108 work->new_state = RDMA_CM_ADDR_RESOLVED;
2109 work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2110 queue_work(cma_wq, &work->work);
2117 static int cma_resolve_ib_addr(struct rdma_id_private *id_priv)
2119 struct cma_work *work;
2122 work = kzalloc(sizeof *work, GFP_KERNEL);
2126 if (!id_priv->cma_dev) {
2127 ret = cma_resolve_ib_dev(id_priv);
2132 rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, (union ib_gid *)
2133 &(((struct sockaddr_ib *) &id_priv->id.route.addr.dst_addr)->sib_addr));
2136 INIT_WORK(&work->work, cma_work_handler);
2137 work->old_state = RDMA_CM_ADDR_QUERY;
2138 work->new_state = RDMA_CM_ADDR_RESOLVED;
2139 work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2140 queue_work(cma_wq, &work->work);
2147 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2148 struct sockaddr *dst_addr)
2150 if (!src_addr || !src_addr->sa_family) {
2151 src_addr = (struct sockaddr *) &id->route.addr.src_addr;
2152 src_addr->sa_family = dst_addr->sa_family;
2153 if (dst_addr->sa_family == AF_INET6) {
2154 ((struct sockaddr_in6 *) src_addr)->sin6_scope_id =
2155 ((struct sockaddr_in6 *) dst_addr)->sin6_scope_id;
2156 } else if (dst_addr->sa_family == AF_IB) {
2157 ((struct sockaddr_ib *) src_addr)->sib_pkey =
2158 ((struct sockaddr_ib *) dst_addr)->sib_pkey;
2161 return rdma_bind_addr(id, src_addr);
2164 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2165 struct sockaddr *dst_addr, int timeout_ms)
2167 struct rdma_id_private *id_priv;
2170 id_priv = container_of(id, struct rdma_id_private, id);
2171 if (id_priv->state == RDMA_CM_IDLE) {
2172 ret = cma_bind_addr(id, src_addr, dst_addr);
2177 if (cma_family(id_priv) != dst_addr->sa_family)
2180 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY))
2183 atomic_inc(&id_priv->refcount);
2184 memcpy(cma_dst_addr(id_priv), dst_addr, rdma_addr_size(dst_addr));
2185 if (cma_any_addr(dst_addr)) {
2186 ret = cma_resolve_loopback(id_priv);
2188 if (dst_addr->sa_family == AF_IB) {
2189 ret = cma_resolve_ib_addr(id_priv);
2191 ret = rdma_resolve_ip(&addr_client, cma_src_addr(id_priv),
2192 dst_addr, &id->route.addr.dev_addr,
2193 timeout_ms, addr_handler, id_priv);
2201 cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
2202 cma_deref_id(id_priv);
2205 EXPORT_SYMBOL(rdma_resolve_addr);
2207 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse)
2209 struct rdma_id_private *id_priv;
2210 unsigned long flags;
2213 id_priv = container_of(id, struct rdma_id_private, id);
2214 spin_lock_irqsave(&id_priv->lock, flags);
2215 if (reuse || id_priv->state == RDMA_CM_IDLE) {
2216 id_priv->reuseaddr = reuse;
2221 spin_unlock_irqrestore(&id_priv->lock, flags);
2224 EXPORT_SYMBOL(rdma_set_reuseaddr);
2226 int rdma_set_afonly(struct rdma_cm_id *id, int afonly)
2228 struct rdma_id_private *id_priv;
2229 unsigned long flags;
2232 id_priv = container_of(id, struct rdma_id_private, id);
2233 spin_lock_irqsave(&id_priv->lock, flags);
2234 if (id_priv->state == RDMA_CM_IDLE || id_priv->state == RDMA_CM_ADDR_BOUND) {
2235 id_priv->options |= (1 << CMA_OPTION_AFONLY);
2236 id_priv->afonly = afonly;
2241 spin_unlock_irqrestore(&id_priv->lock, flags);
2244 EXPORT_SYMBOL(rdma_set_afonly);
2246 static void cma_bind_port(struct rdma_bind_list *bind_list,
2247 struct rdma_id_private *id_priv)
2249 struct sockaddr *addr;
2250 struct sockaddr_ib *sib;
2254 addr = cma_src_addr(id_priv);
2255 port = htons(bind_list->port);
2257 switch (addr->sa_family) {
2259 ((struct sockaddr_in *) addr)->sin_port = port;
2262 ((struct sockaddr_in6 *) addr)->sin6_port = port;
2265 sib = (struct sockaddr_ib *) addr;
2266 sid = be64_to_cpu(sib->sib_sid);
2267 mask = be64_to_cpu(sib->sib_sid_mask);
2268 sib->sib_sid = cpu_to_be64((sid & mask) | (u64) ntohs(port));
2269 sib->sib_sid_mask = cpu_to_be64(~0ULL);
2272 id_priv->bind_list = bind_list;
2273 hlist_add_head(&id_priv->node, &bind_list->owners);
2276 static int cma_alloc_port(struct idr *ps, struct rdma_id_private *id_priv,
2277 unsigned short snum)
2279 struct rdma_bind_list *bind_list;
2282 bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
2286 ret = idr_alloc(ps, bind_list, snum, snum + 1, GFP_KERNEL);
2291 bind_list->port = (unsigned short)ret;
2292 cma_bind_port(bind_list, id_priv);
2296 return ret == -ENOSPC ? -EADDRNOTAVAIL : ret;
2299 static int cma_alloc_any_port(struct idr *ps, struct rdma_id_private *id_priv)
2301 static unsigned int last_used_port;
2302 int low, high, remaining;
2305 inet_get_local_port_range(&low, &high);
2306 remaining = (high - low) + 1;
2307 rover = net_random() % remaining + low;
2309 if (last_used_port != rover &&
2310 !idr_find(ps, (unsigned short) rover)) {
2311 int ret = cma_alloc_port(ps, id_priv, rover);
2313 * Remember previously used port number in order to avoid
2314 * re-using same port immediately after it is closed.
2317 last_used_port = rover;
2318 if (ret != -EADDRNOTAVAIL)
2323 if ((rover < low) || (rover > high))
2327 return -EADDRNOTAVAIL;
2331 * Check that the requested port is available. This is called when trying to
2332 * bind to a specific port, or when trying to listen on a bound port. In
2333 * the latter case, the provided id_priv may already be on the bind_list, but
2334 * we still need to check that it's okay to start listening.
2336 static int cma_check_port(struct rdma_bind_list *bind_list,
2337 struct rdma_id_private *id_priv, uint8_t reuseaddr)
2339 struct rdma_id_private *cur_id;
2340 struct sockaddr *addr, *cur_addr;
2342 addr = cma_src_addr(id_priv);
2343 hlist_for_each_entry(cur_id, &bind_list->owners, node) {
2344 if (id_priv == cur_id)
2347 if ((cur_id->state != RDMA_CM_LISTEN) && reuseaddr &&
2351 cur_addr = cma_src_addr(cur_id);
2352 if (id_priv->afonly && cur_id->afonly &&
2353 (addr->sa_family != cur_addr->sa_family))
2356 if (cma_any_addr(addr) || cma_any_addr(cur_addr))
2357 return -EADDRNOTAVAIL;
2359 if (!cma_addr_cmp(addr, cur_addr))
2365 static int cma_use_port(struct idr *ps, struct rdma_id_private *id_priv)
2367 struct rdma_bind_list *bind_list;
2368 unsigned short snum;
2371 snum = ntohs(cma_port(cma_src_addr(id_priv)));
2372 if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
2375 bind_list = idr_find(ps, snum);
2377 ret = cma_alloc_port(ps, id_priv, snum);
2379 ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr);
2381 cma_bind_port(bind_list, id_priv);
2386 static int cma_bind_listen(struct rdma_id_private *id_priv)
2388 struct rdma_bind_list *bind_list = id_priv->bind_list;
2392 if (bind_list->owners.first->next)
2393 ret = cma_check_port(bind_list, id_priv, 0);
2394 mutex_unlock(&lock);
2398 static struct idr *cma_select_inet_ps(struct rdma_id_private *id_priv)
2400 switch (id_priv->id.ps) {
2414 static struct idr *cma_select_ib_ps(struct rdma_id_private *id_priv)
2416 struct idr *ps = NULL;
2417 struct sockaddr_ib *sib;
2418 u64 sid_ps, mask, sid;
2420 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
2421 mask = be64_to_cpu(sib->sib_sid_mask) & RDMA_IB_IP_PS_MASK;
2422 sid = be64_to_cpu(sib->sib_sid) & mask;
2424 if ((id_priv->id.ps == RDMA_PS_IB) && (sid == (RDMA_IB_IP_PS_IB & mask))) {
2425 sid_ps = RDMA_IB_IP_PS_IB;
2427 } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_TCP)) &&
2428 (sid == (RDMA_IB_IP_PS_TCP & mask))) {
2429 sid_ps = RDMA_IB_IP_PS_TCP;
2431 } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_UDP)) &&
2432 (sid == (RDMA_IB_IP_PS_UDP & mask))) {
2433 sid_ps = RDMA_IB_IP_PS_UDP;
2438 sib->sib_sid = cpu_to_be64(sid_ps | ntohs(cma_port((struct sockaddr *) sib)));
2439 sib->sib_sid_mask = cpu_to_be64(RDMA_IB_IP_PS_MASK |
2440 be64_to_cpu(sib->sib_sid_mask));
2445 static int cma_get_port(struct rdma_id_private *id_priv)
2450 if (cma_family(id_priv) != AF_IB)
2451 ps = cma_select_inet_ps(id_priv);
2453 ps = cma_select_ib_ps(id_priv);
2455 return -EPROTONOSUPPORT;
2458 if (cma_any_port(cma_src_addr(id_priv)))
2459 ret = cma_alloc_any_port(ps, id_priv);
2461 ret = cma_use_port(ps, id_priv);
2462 mutex_unlock(&lock);
2467 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
2468 struct sockaddr *addr)
2470 #if IS_ENABLED(CONFIG_IPV6)
2471 struct sockaddr_in6 *sin6;
2473 if (addr->sa_family != AF_INET6)
2476 sin6 = (struct sockaddr_in6 *) addr;
2477 if ((ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL) &&
2478 !sin6->sin6_scope_id)
2481 dev_addr->bound_dev_if = sin6->sin6_scope_id;
2486 int rdma_listen(struct rdma_cm_id *id, int backlog)
2488 struct rdma_id_private *id_priv;
2491 id_priv = container_of(id, struct rdma_id_private, id);
2492 if (id_priv->state == RDMA_CM_IDLE) {
2493 id->route.addr.src_addr.ss_family = AF_INET;
2494 ret = rdma_bind_addr(id, cma_src_addr(id_priv));
2499 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN))
2502 if (id_priv->reuseaddr) {
2503 ret = cma_bind_listen(id_priv);
2508 id_priv->backlog = backlog;
2510 switch (rdma_node_get_transport(id->device->node_type)) {
2511 case RDMA_TRANSPORT_IB:
2512 ret = cma_ib_listen(id_priv);
2516 case RDMA_TRANSPORT_IWARP:
2517 ret = cma_iw_listen(id_priv, backlog);
2526 cma_listen_on_all(id_priv);
2530 id_priv->backlog = 0;
2531 cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND);
2534 EXPORT_SYMBOL(rdma_listen);
2536 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
2538 struct rdma_id_private *id_priv;
2541 if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6 &&
2542 addr->sa_family != AF_IB)
2543 return -EAFNOSUPPORT;
2545 id_priv = container_of(id, struct rdma_id_private, id);
2546 if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND))
2549 ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
2553 if (!cma_any_addr(addr)) {
2554 ret = cma_translate_addr(addr, &id->route.addr.dev_addr);
2558 ret = cma_acquire_dev(id_priv);
2563 memcpy(cma_src_addr(id_priv), addr, rdma_addr_size(addr));
2564 if (!(id_priv->options & (1 << CMA_OPTION_AFONLY))) {
2565 if (addr->sa_family == AF_INET)
2566 id_priv->afonly = 1;
2567 #if IS_ENABLED(CONFIG_IPV6)
2568 else if (addr->sa_family == AF_INET6)
2569 id_priv->afonly = init_net.ipv6.sysctl.bindv6only;
2572 ret = cma_get_port(id_priv);
2578 if (id_priv->cma_dev)
2579 cma_release_dev(id_priv);
2581 cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE);
2584 EXPORT_SYMBOL(rdma_bind_addr);
2586 static int cma_format_hdr(void *hdr, struct rdma_id_private *id_priv)
2588 struct cma_hdr *cma_hdr;
2591 cma_hdr->cma_version = CMA_VERSION;
2592 if (cma_family(id_priv) == AF_INET) {
2593 struct sockaddr_in *src4, *dst4;
2595 src4 = (struct sockaddr_in *) cma_src_addr(id_priv);
2596 dst4 = (struct sockaddr_in *) cma_dst_addr(id_priv);
2598 cma_set_ip_ver(cma_hdr, 4);
2599 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2600 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2601 cma_hdr->port = src4->sin_port;
2602 } else if (cma_family(id_priv) == AF_INET6) {
2603 struct sockaddr_in6 *src6, *dst6;
2605 src6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
2606 dst6 = (struct sockaddr_in6 *) cma_dst_addr(id_priv);
2608 cma_set_ip_ver(cma_hdr, 6);
2609 cma_hdr->src_addr.ip6 = src6->sin6_addr;
2610 cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
2611 cma_hdr->port = src6->sin6_port;
2616 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
2617 struct ib_cm_event *ib_event)
2619 struct rdma_id_private *id_priv = cm_id->context;
2620 struct rdma_cm_event event;
2621 struct ib_cm_sidr_rep_event_param *rep = &ib_event->param.sidr_rep_rcvd;
2624 if (cma_disable_callback(id_priv, RDMA_CM_CONNECT))
2627 memset(&event, 0, sizeof event);
2628 switch (ib_event->event) {
2629 case IB_CM_SIDR_REQ_ERROR:
2630 event.event = RDMA_CM_EVENT_UNREACHABLE;
2631 event.status = -ETIMEDOUT;
2633 case IB_CM_SIDR_REP_RECEIVED:
2634 event.param.ud.private_data = ib_event->private_data;
2635 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
2636 if (rep->status != IB_SIDR_SUCCESS) {
2637 event.event = RDMA_CM_EVENT_UNREACHABLE;
2638 event.status = ib_event->param.sidr_rep_rcvd.status;
2641 ret = cma_set_qkey(id_priv, rep->qkey);
2643 event.event = RDMA_CM_EVENT_ADDR_ERROR;
2647 ib_init_ah_from_path(id_priv->id.device, id_priv->id.port_num,
2648 id_priv->id.route.path_rec,
2649 &event.param.ud.ah_attr);
2650 event.param.ud.qp_num = rep->qpn;
2651 event.param.ud.qkey = rep->qkey;
2652 event.event = RDMA_CM_EVENT_ESTABLISHED;
2656 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
2661 ret = id_priv->id.event_handler(&id_priv->id, &event);
2663 /* Destroy the CM ID by returning a non-zero value. */
2664 id_priv->cm_id.ib = NULL;
2665 cma_exch(id_priv, RDMA_CM_DESTROYING);
2666 mutex_unlock(&id_priv->handler_mutex);
2667 rdma_destroy_id(&id_priv->id);
2671 mutex_unlock(&id_priv->handler_mutex);
2675 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
2676 struct rdma_conn_param *conn_param)
2678 struct ib_cm_sidr_req_param req;
2679 struct ib_cm_id *id;
2682 offset = cma_user_data_offset(id_priv);
2683 req.private_data_len = offset + conn_param->private_data_len;
2684 if (req.private_data_len < conn_param->private_data_len)
2687 if (req.private_data_len) {
2688 req.private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2689 if (!req.private_data)
2692 req.private_data = NULL;
2695 if (conn_param->private_data && conn_param->private_data_len)
2696 memcpy((void *) req.private_data + offset,
2697 conn_param->private_data, conn_param->private_data_len);
2699 if (req.private_data) {
2700 ret = cma_format_hdr((void *) req.private_data, id_priv);
2705 id = ib_create_cm_id(id_priv->id.device, cma_sidr_rep_handler,
2711 id_priv->cm_id.ib = id;
2713 req.path = id_priv->id.route.path_rec;
2714 req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
2715 req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
2716 req.max_cm_retries = CMA_MAX_CM_RETRIES;
2718 ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
2720 ib_destroy_cm_id(id_priv->cm_id.ib);
2721 id_priv->cm_id.ib = NULL;
2724 kfree(req.private_data);
2728 static int cma_connect_ib(struct rdma_id_private *id_priv,
2729 struct rdma_conn_param *conn_param)
2731 struct ib_cm_req_param req;
2732 struct rdma_route *route;
2734 struct ib_cm_id *id;
2737 memset(&req, 0, sizeof req);
2738 offset = cma_user_data_offset(id_priv);
2739 req.private_data_len = offset + conn_param->private_data_len;
2740 if (req.private_data_len < conn_param->private_data_len)
2743 if (req.private_data_len) {
2744 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2748 private_data = NULL;
2751 if (conn_param->private_data && conn_param->private_data_len)
2752 memcpy(private_data + offset, conn_param->private_data,
2753 conn_param->private_data_len);
2755 id = ib_create_cm_id(id_priv->id.device, cma_ib_handler, id_priv);
2760 id_priv->cm_id.ib = id;
2762 route = &id_priv->id.route;
2764 ret = cma_format_hdr(private_data, id_priv);
2767 req.private_data = private_data;
2770 req.primary_path = &route->path_rec[0];
2771 if (route->num_paths == 2)
2772 req.alternate_path = &route->path_rec[1];
2774 req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
2775 req.qp_num = id_priv->qp_num;
2776 req.qp_type = id_priv->id.qp_type;
2777 req.starting_psn = id_priv->seq_num;
2778 req.responder_resources = conn_param->responder_resources;
2779 req.initiator_depth = conn_param->initiator_depth;
2780 req.flow_control = conn_param->flow_control;
2781 req.retry_count = min_t(u8, 7, conn_param->retry_count);
2782 req.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
2783 req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2784 req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2785 req.max_cm_retries = CMA_MAX_CM_RETRIES;
2786 req.srq = id_priv->srq ? 1 : 0;
2788 ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
2790 if (ret && !IS_ERR(id)) {
2791 ib_destroy_cm_id(id);
2792 id_priv->cm_id.ib = NULL;
2795 kfree(private_data);
2799 static int cma_connect_iw(struct rdma_id_private *id_priv,
2800 struct rdma_conn_param *conn_param)
2802 struct iw_cm_id *cm_id;
2803 struct sockaddr_in* sin;
2805 struct iw_cm_conn_param iw_param;
2807 cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
2809 return PTR_ERR(cm_id);
2811 id_priv->cm_id.iw = cm_id;
2813 sin = (struct sockaddr_in *) cma_src_addr(id_priv);
2814 cm_id->local_addr = *sin;
2816 sin = (struct sockaddr_in *) cma_dst_addr(id_priv);
2817 cm_id->remote_addr = *sin;
2819 ret = cma_modify_qp_rtr(id_priv, conn_param);
2824 iw_param.ord = conn_param->initiator_depth;
2825 iw_param.ird = conn_param->responder_resources;
2826 iw_param.private_data = conn_param->private_data;
2827 iw_param.private_data_len = conn_param->private_data_len;
2828 iw_param.qpn = id_priv->id.qp ? id_priv->qp_num : conn_param->qp_num;
2830 memset(&iw_param, 0, sizeof iw_param);
2831 iw_param.qpn = id_priv->qp_num;
2833 ret = iw_cm_connect(cm_id, &iw_param);
2836 iw_destroy_cm_id(cm_id);
2837 id_priv->cm_id.iw = NULL;
2842 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2844 struct rdma_id_private *id_priv;
2847 id_priv = container_of(id, struct rdma_id_private, id);
2848 if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT))
2852 id_priv->qp_num = conn_param->qp_num;
2853 id_priv->srq = conn_param->srq;
2856 switch (rdma_node_get_transport(id->device->node_type)) {
2857 case RDMA_TRANSPORT_IB:
2858 if (id->qp_type == IB_QPT_UD)
2859 ret = cma_resolve_ib_udp(id_priv, conn_param);
2861 ret = cma_connect_ib(id_priv, conn_param);
2863 case RDMA_TRANSPORT_IWARP:
2864 ret = cma_connect_iw(id_priv, conn_param);
2875 cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED);
2878 EXPORT_SYMBOL(rdma_connect);
2880 static int cma_accept_ib(struct rdma_id_private *id_priv,
2881 struct rdma_conn_param *conn_param)
2883 struct ib_cm_rep_param rep;
2886 ret = cma_modify_qp_rtr(id_priv, conn_param);
2890 ret = cma_modify_qp_rts(id_priv, conn_param);
2894 memset(&rep, 0, sizeof rep);
2895 rep.qp_num = id_priv->qp_num;
2896 rep.starting_psn = id_priv->seq_num;
2897 rep.private_data = conn_param->private_data;
2898 rep.private_data_len = conn_param->private_data_len;
2899 rep.responder_resources = conn_param->responder_resources;
2900 rep.initiator_depth = conn_param->initiator_depth;
2901 rep.failover_accepted = 0;
2902 rep.flow_control = conn_param->flow_control;
2903 rep.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
2904 rep.srq = id_priv->srq ? 1 : 0;
2906 ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
2911 static int cma_accept_iw(struct rdma_id_private *id_priv,
2912 struct rdma_conn_param *conn_param)
2914 struct iw_cm_conn_param iw_param;
2917 ret = cma_modify_qp_rtr(id_priv, conn_param);
2921 iw_param.ord = conn_param->initiator_depth;
2922 iw_param.ird = conn_param->responder_resources;
2923 iw_param.private_data = conn_param->private_data;
2924 iw_param.private_data_len = conn_param->private_data_len;
2925 if (id_priv->id.qp) {
2926 iw_param.qpn = id_priv->qp_num;
2928 iw_param.qpn = conn_param->qp_num;
2930 return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
2933 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
2934 enum ib_cm_sidr_status status, u32 qkey,
2935 const void *private_data, int private_data_len)
2937 struct ib_cm_sidr_rep_param rep;
2940 memset(&rep, 0, sizeof rep);
2941 rep.status = status;
2942 if (status == IB_SIDR_SUCCESS) {
2943 ret = cma_set_qkey(id_priv, qkey);
2946 rep.qp_num = id_priv->qp_num;
2947 rep.qkey = id_priv->qkey;
2949 rep.private_data = private_data;
2950 rep.private_data_len = private_data_len;
2952 return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
2955 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2957 struct rdma_id_private *id_priv;
2960 id_priv = container_of(id, struct rdma_id_private, id);
2962 id_priv->owner = task_pid_nr(current);
2964 if (!cma_comp(id_priv, RDMA_CM_CONNECT))
2967 if (!id->qp && conn_param) {
2968 id_priv->qp_num = conn_param->qp_num;
2969 id_priv->srq = conn_param->srq;
2972 switch (rdma_node_get_transport(id->device->node_type)) {
2973 case RDMA_TRANSPORT_IB:
2974 if (id->qp_type == IB_QPT_UD) {
2976 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
2978 conn_param->private_data,
2979 conn_param->private_data_len);
2981 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
2985 ret = cma_accept_ib(id_priv, conn_param);
2987 ret = cma_rep_recv(id_priv);
2990 case RDMA_TRANSPORT_IWARP:
2991 ret = cma_accept_iw(id_priv, conn_param);
3003 cma_modify_qp_err(id_priv);
3004 rdma_reject(id, NULL, 0);
3007 EXPORT_SYMBOL(rdma_accept);
3009 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
3011 struct rdma_id_private *id_priv;
3014 id_priv = container_of(id, struct rdma_id_private, id);
3015 if (!id_priv->cm_id.ib)
3018 switch (id->device->node_type) {
3019 case RDMA_NODE_IB_CA:
3020 ret = ib_cm_notify(id_priv->cm_id.ib, event);
3028 EXPORT_SYMBOL(rdma_notify);
3030 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
3031 u8 private_data_len)
3033 struct rdma_id_private *id_priv;
3036 id_priv = container_of(id, struct rdma_id_private, id);
3037 if (!id_priv->cm_id.ib)
3040 switch (rdma_node_get_transport(id->device->node_type)) {
3041 case RDMA_TRANSPORT_IB:
3042 if (id->qp_type == IB_QPT_UD)
3043 ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT, 0,
3044 private_data, private_data_len);
3046 ret = ib_send_cm_rej(id_priv->cm_id.ib,
3047 IB_CM_REJ_CONSUMER_DEFINED, NULL,
3048 0, private_data, private_data_len);
3050 case RDMA_TRANSPORT_IWARP:
3051 ret = iw_cm_reject(id_priv->cm_id.iw,
3052 private_data, private_data_len);
3060 EXPORT_SYMBOL(rdma_reject);
3062 int rdma_disconnect(struct rdma_cm_id *id)
3064 struct rdma_id_private *id_priv;
3067 id_priv = container_of(id, struct rdma_id_private, id);
3068 if (!id_priv->cm_id.ib)
3071 switch (rdma_node_get_transport(id->device->node_type)) {
3072 case RDMA_TRANSPORT_IB:
3073 ret = cma_modify_qp_err(id_priv);
3076 /* Initiate or respond to a disconnect. */
3077 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
3078 ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
3080 case RDMA_TRANSPORT_IWARP:
3081 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
3090 EXPORT_SYMBOL(rdma_disconnect);
3092 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
3094 struct rdma_id_private *id_priv;
3095 struct cma_multicast *mc = multicast->context;
3096 struct rdma_cm_event event;
3099 id_priv = mc->id_priv;
3100 if (cma_disable_callback(id_priv, RDMA_CM_ADDR_BOUND) &&
3101 cma_disable_callback(id_priv, RDMA_CM_ADDR_RESOLVED))
3105 status = cma_set_qkey(id_priv, be32_to_cpu(multicast->rec.qkey));
3106 mutex_lock(&id_priv->qp_mutex);
3107 if (!status && id_priv->id.qp)
3108 status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid,
3109 be16_to_cpu(multicast->rec.mlid));
3110 mutex_unlock(&id_priv->qp_mutex);
3112 memset(&event, 0, sizeof event);
3113 event.status = status;
3114 event.param.ud.private_data = mc->context;
3116 event.event = RDMA_CM_EVENT_MULTICAST_JOIN;
3117 ib_init_ah_from_mcmember(id_priv->id.device,
3118 id_priv->id.port_num, &multicast->rec,
3119 &event.param.ud.ah_attr);
3120 event.param.ud.qp_num = 0xFFFFFF;
3121 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
3123 event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
3125 ret = id_priv->id.event_handler(&id_priv->id, &event);
3127 cma_exch(id_priv, RDMA_CM_DESTROYING);
3128 mutex_unlock(&id_priv->handler_mutex);
3129 rdma_destroy_id(&id_priv->id);
3133 mutex_unlock(&id_priv->handler_mutex);
3137 static void cma_set_mgid(struct rdma_id_private *id_priv,
3138 struct sockaddr *addr, union ib_gid *mgid)
3140 unsigned char mc_map[MAX_ADDR_LEN];
3141 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3142 struct sockaddr_in *sin = (struct sockaddr_in *) addr;
3143 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
3145 if (cma_any_addr(addr)) {
3146 memset(mgid, 0, sizeof *mgid);
3147 } else if ((addr->sa_family == AF_INET6) &&
3148 ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
3150 /* IPv6 address is an SA assigned MGID. */
3151 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3152 } else if (addr->sa_family == AF_IB) {
3153 memcpy(mgid, &((struct sockaddr_ib *) addr)->sib_addr, sizeof *mgid);
3154 } else if ((addr->sa_family == AF_INET6)) {
3155 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
3156 if (id_priv->id.ps == RDMA_PS_UDP)
3157 mc_map[7] = 0x01; /* Use RDMA CM signature */
3158 *mgid = *(union ib_gid *) (mc_map + 4);
3160 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
3161 if (id_priv->id.ps == RDMA_PS_UDP)
3162 mc_map[7] = 0x01; /* Use RDMA CM signature */
3163 *mgid = *(union ib_gid *) (mc_map + 4);
3167 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
3168 struct cma_multicast *mc)
3170 struct ib_sa_mcmember_rec rec;
3171 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3172 ib_sa_comp_mask comp_mask;
3175 ib_addr_get_mgid(dev_addr, &rec.mgid);
3176 ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
3181 ret = cma_set_qkey(id_priv, 0);
3185 cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
3186 rec.qkey = cpu_to_be32(id_priv->qkey);
3187 rdma_addr_get_sgid(dev_addr, &rec.port_gid);
3188 rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
3191 comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
3192 IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
3193 IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
3194 IB_SA_MCMEMBER_REC_FLOW_LABEL |
3195 IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
3197 if (id_priv->id.ps == RDMA_PS_IPOIB)
3198 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
3199 IB_SA_MCMEMBER_REC_RATE_SELECTOR |
3200 IB_SA_MCMEMBER_REC_MTU_SELECTOR |
3201 IB_SA_MCMEMBER_REC_MTU |
3202 IB_SA_MCMEMBER_REC_HOP_LIMIT;
3204 mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device,
3205 id_priv->id.port_num, &rec,
3206 comp_mask, GFP_KERNEL,
3207 cma_ib_mc_handler, mc);
3208 return PTR_RET(mc->multicast.ib);
3211 static void iboe_mcast_work_handler(struct work_struct *work)
3213 struct iboe_mcast_work *mw = container_of(work, struct iboe_mcast_work, work);
3214 struct cma_multicast *mc = mw->mc;
3215 struct ib_sa_multicast *m = mc->multicast.ib;
3217 mc->multicast.ib->context = mc;
3218 cma_ib_mc_handler(0, m);
3219 kref_put(&mc->mcref, release_mc);
3223 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid)
3225 struct sockaddr_in *sin = (struct sockaddr_in *)addr;
3226 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
3228 if (cma_any_addr(addr)) {
3229 memset(mgid, 0, sizeof *mgid);
3230 } else if (addr->sa_family == AF_INET6) {
3231 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3233 mgid->raw[0] = 0xff;
3234 mgid->raw[1] = 0x0e;
3243 mgid->raw[10] = 0xff;
3244 mgid->raw[11] = 0xff;
3245 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
3249 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
3250 struct cma_multicast *mc)
3252 struct iboe_mcast_work *work;
3253 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3255 struct sockaddr *addr = (struct sockaddr *)&mc->addr;
3256 struct net_device *ndev = NULL;
3258 if (cma_zero_addr((struct sockaddr *)&mc->addr))
3261 work = kzalloc(sizeof *work, GFP_KERNEL);
3265 mc->multicast.ib = kzalloc(sizeof(struct ib_sa_multicast), GFP_KERNEL);
3266 if (!mc->multicast.ib) {
3271 cma_iboe_set_mgid(addr, &mc->multicast.ib->rec.mgid);
3273 mc->multicast.ib->rec.pkey = cpu_to_be16(0xffff);
3274 if (id_priv->id.ps == RDMA_PS_UDP)
3275 mc->multicast.ib->rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
3277 if (dev_addr->bound_dev_if)
3278 ndev = dev_get_by_index(&init_net, dev_addr->bound_dev_if);
3283 mc->multicast.ib->rec.rate = iboe_get_rate(ndev);
3284 mc->multicast.ib->rec.hop_limit = 1;
3285 mc->multicast.ib->rec.mtu = iboe_get_mtu(ndev->mtu);
3287 if (!mc->multicast.ib->rec.mtu) {
3291 iboe_addr_get_sgid(dev_addr, &mc->multicast.ib->rec.port_gid);
3294 INIT_WORK(&work->work, iboe_mcast_work_handler);
3295 kref_get(&mc->mcref);
3296 queue_work(cma_wq, &work->work);
3301 kfree(mc->multicast.ib);
3307 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
3310 struct rdma_id_private *id_priv;
3311 struct cma_multicast *mc;
3314 id_priv = container_of(id, struct rdma_id_private, id);
3315 if (!cma_comp(id_priv, RDMA_CM_ADDR_BOUND) &&
3316 !cma_comp(id_priv, RDMA_CM_ADDR_RESOLVED))
3319 mc = kmalloc(sizeof *mc, GFP_KERNEL);
3323 memcpy(&mc->addr, addr, rdma_addr_size(addr));
3324 mc->context = context;
3325 mc->id_priv = id_priv;
3327 spin_lock(&id_priv->lock);
3328 list_add(&mc->list, &id_priv->mc_list);
3329 spin_unlock(&id_priv->lock);
3331 switch (rdma_node_get_transport(id->device->node_type)) {
3332 case RDMA_TRANSPORT_IB:
3333 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
3334 case IB_LINK_LAYER_INFINIBAND:
3335 ret = cma_join_ib_multicast(id_priv, mc);
3337 case IB_LINK_LAYER_ETHERNET:
3338 kref_init(&mc->mcref);
3339 ret = cma_iboe_join_multicast(id_priv, mc);
3351 spin_lock_irq(&id_priv->lock);
3352 list_del(&mc->list);
3353 spin_unlock_irq(&id_priv->lock);
3358 EXPORT_SYMBOL(rdma_join_multicast);
3360 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
3362 struct rdma_id_private *id_priv;
3363 struct cma_multicast *mc;
3365 id_priv = container_of(id, struct rdma_id_private, id);
3366 spin_lock_irq(&id_priv->lock);
3367 list_for_each_entry(mc, &id_priv->mc_list, list) {
3368 if (!memcmp(&mc->addr, addr, rdma_addr_size(addr))) {
3369 list_del(&mc->list);
3370 spin_unlock_irq(&id_priv->lock);
3373 ib_detach_mcast(id->qp,
3374 &mc->multicast.ib->rec.mgid,
3375 be16_to_cpu(mc->multicast.ib->rec.mlid));
3376 if (rdma_node_get_transport(id_priv->cma_dev->device->node_type) == RDMA_TRANSPORT_IB) {
3377 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
3378 case IB_LINK_LAYER_INFINIBAND:
3379 ib_sa_free_multicast(mc->multicast.ib);
3382 case IB_LINK_LAYER_ETHERNET:
3383 kref_put(&mc->mcref, release_mc);
3392 spin_unlock_irq(&id_priv->lock);
3394 EXPORT_SYMBOL(rdma_leave_multicast);
3396 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
3398 struct rdma_dev_addr *dev_addr;
3399 struct cma_ndev_work *work;
3401 dev_addr = &id_priv->id.route.addr.dev_addr;
3403 if ((dev_addr->bound_dev_if == ndev->ifindex) &&
3404 memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
3405 printk(KERN_INFO "RDMA CM addr change for ndev %s used by id %p\n",
3406 ndev->name, &id_priv->id);
3407 work = kzalloc(sizeof *work, GFP_KERNEL);
3411 INIT_WORK(&work->work, cma_ndev_work_handler);
3413 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
3414 atomic_inc(&id_priv->refcount);
3415 queue_work(cma_wq, &work->work);
3421 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
3424 struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
3425 struct cma_device *cma_dev;
3426 struct rdma_id_private *id_priv;
3427 int ret = NOTIFY_DONE;
3429 if (dev_net(ndev) != &init_net)
3432 if (event != NETDEV_BONDING_FAILOVER)
3435 if (!(ndev->flags & IFF_MASTER) || !(ndev->priv_flags & IFF_BONDING))
3439 list_for_each_entry(cma_dev, &dev_list, list)
3440 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
3441 ret = cma_netdev_change(ndev, id_priv);
3447 mutex_unlock(&lock);
3451 static struct notifier_block cma_nb = {
3452 .notifier_call = cma_netdev_callback
3455 static void cma_add_one(struct ib_device *device)
3457 struct cma_device *cma_dev;
3458 struct rdma_id_private *id_priv;
3460 cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
3464 cma_dev->device = device;
3466 init_completion(&cma_dev->comp);
3467 atomic_set(&cma_dev->refcount, 1);
3468 INIT_LIST_HEAD(&cma_dev->id_list);
3469 ib_set_client_data(device, &cma_client, cma_dev);
3472 list_add_tail(&cma_dev->list, &dev_list);
3473 list_for_each_entry(id_priv, &listen_any_list, list)
3474 cma_listen_on_dev(id_priv, cma_dev);
3475 mutex_unlock(&lock);
3478 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
3480 struct rdma_cm_event event;
3481 enum rdma_cm_state state;
3484 /* Record that we want to remove the device */
3485 state = cma_exch(id_priv, RDMA_CM_DEVICE_REMOVAL);
3486 if (state == RDMA_CM_DESTROYING)
3489 cma_cancel_operation(id_priv, state);
3490 mutex_lock(&id_priv->handler_mutex);
3492 /* Check for destruction from another callback. */
3493 if (!cma_comp(id_priv, RDMA_CM_DEVICE_REMOVAL))
3496 memset(&event, 0, sizeof event);
3497 event.event = RDMA_CM_EVENT_DEVICE_REMOVAL;
3498 ret = id_priv->id.event_handler(&id_priv->id, &event);
3500 mutex_unlock(&id_priv->handler_mutex);
3504 static void cma_process_remove(struct cma_device *cma_dev)
3506 struct rdma_id_private *id_priv;
3510 while (!list_empty(&cma_dev->id_list)) {
3511 id_priv = list_entry(cma_dev->id_list.next,
3512 struct rdma_id_private, list);
3514 list_del(&id_priv->listen_list);
3515 list_del_init(&id_priv->list);
3516 atomic_inc(&id_priv->refcount);
3517 mutex_unlock(&lock);
3519 ret = id_priv->internal_id ? 1 : cma_remove_id_dev(id_priv);
3520 cma_deref_id(id_priv);
3522 rdma_destroy_id(&id_priv->id);
3526 mutex_unlock(&lock);
3528 cma_deref_dev(cma_dev);
3529 wait_for_completion(&cma_dev->comp);
3532 static void cma_remove_one(struct ib_device *device)
3534 struct cma_device *cma_dev;
3536 cma_dev = ib_get_client_data(device, &cma_client);
3541 list_del(&cma_dev->list);
3542 mutex_unlock(&lock);
3544 cma_process_remove(cma_dev);
3548 static int cma_get_id_stats(struct sk_buff *skb, struct netlink_callback *cb)
3550 struct nlmsghdr *nlh;
3551 struct rdma_cm_id_stats *id_stats;
3552 struct rdma_id_private *id_priv;
3553 struct rdma_cm_id *id = NULL;
3554 struct cma_device *cma_dev;
3555 int i_dev = 0, i_id = 0;
3558 * We export all of the IDs as a sequence of messages. Each
3559 * ID gets its own netlink message.
3563 list_for_each_entry(cma_dev, &dev_list, list) {
3564 if (i_dev < cb->args[0]) {
3570 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
3571 if (i_id < cb->args[1]) {
3576 id_stats = ibnl_put_msg(skb, &nlh, cb->nlh->nlmsg_seq,
3577 sizeof *id_stats, RDMA_NL_RDMA_CM,
3578 RDMA_NL_RDMA_CM_ID_STATS);
3582 memset(id_stats, 0, sizeof *id_stats);
3584 id_stats->node_type = id->route.addr.dev_addr.dev_type;
3585 id_stats->port_num = id->port_num;
3586 id_stats->bound_dev_if =
3587 id->route.addr.dev_addr.bound_dev_if;
3589 if (ibnl_put_attr(skb, nlh,
3590 rdma_addr_size(cma_src_addr(id_priv)),
3591 cma_src_addr(id_priv),
3592 RDMA_NL_RDMA_CM_ATTR_SRC_ADDR))
3594 if (ibnl_put_attr(skb, nlh,
3595 rdma_addr_size(cma_src_addr(id_priv)),
3596 cma_dst_addr(id_priv),
3597 RDMA_NL_RDMA_CM_ATTR_DST_ADDR))
3600 id_stats->pid = id_priv->owner;
3601 id_stats->port_space = id->ps;
3602 id_stats->cm_state = id_priv->state;
3603 id_stats->qp_num = id_priv->qp_num;
3604 id_stats->qp_type = id->qp_type;
3614 mutex_unlock(&lock);
3615 cb->args[0] = i_dev;
3621 static const struct ibnl_client_cbs cma_cb_table[] = {
3622 [RDMA_NL_RDMA_CM_ID_STATS] = { .dump = cma_get_id_stats,
3623 .module = THIS_MODULE },
3626 static int __init cma_init(void)
3630 cma_wq = create_singlethread_workqueue("rdma_cm");
3634 ib_sa_register_client(&sa_client);
3635 rdma_addr_register_client(&addr_client);
3636 register_netdevice_notifier(&cma_nb);
3638 ret = ib_register_client(&cma_client);
3642 if (ibnl_add_client(RDMA_NL_RDMA_CM, RDMA_NL_RDMA_CM_NUM_OPS, cma_cb_table))
3643 printk(KERN_WARNING "RDMA CMA: failed to add netlink callback\n");
3648 unregister_netdevice_notifier(&cma_nb);
3649 rdma_addr_unregister_client(&addr_client);
3650 ib_sa_unregister_client(&sa_client);
3651 destroy_workqueue(cma_wq);
3655 static void __exit cma_cleanup(void)
3657 ibnl_remove_client(RDMA_NL_RDMA_CM);
3658 ib_unregister_client(&cma_client);
3659 unregister_netdevice_notifier(&cma_nb);
3660 rdma_addr_unregister_client(&addr_client);
3661 ib_sa_unregister_client(&sa_client);
3662 destroy_workqueue(cma_wq);
3663 idr_destroy(&tcp_ps);
3664 idr_destroy(&udp_ps);
3665 idr_destroy(&ipoib_ps);
3666 idr_destroy(&ib_ps);
3669 module_init(cma_init);
3670 module_exit(cma_cleanup);