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))) {
446 id_priv->id.port_num = p;
450 if (!cma_dev && (gid.global.subnet_prefix ==
451 dgid->global.subnet_prefix)) {
454 id_priv->id.port_num = p;
464 cma_attach_to_dev(id_priv, cma_dev);
465 addr = (struct sockaddr_ib *) cma_src_addr(id_priv);
466 memcpy(&addr->sib_addr, &sgid, sizeof sgid);
467 cma_translate_ib(addr, &id_priv->id.route.addr.dev_addr);
471 static void cma_deref_id(struct rdma_id_private *id_priv)
473 if (atomic_dec_and_test(&id_priv->refcount))
474 complete(&id_priv->comp);
477 static int cma_disable_callback(struct rdma_id_private *id_priv,
478 enum rdma_cm_state state)
480 mutex_lock(&id_priv->handler_mutex);
481 if (id_priv->state != state) {
482 mutex_unlock(&id_priv->handler_mutex);
488 struct rdma_cm_id *rdma_create_id(rdma_cm_event_handler event_handler,
489 void *context, enum rdma_port_space ps,
490 enum ib_qp_type qp_type)
492 struct rdma_id_private *id_priv;
494 id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
496 return ERR_PTR(-ENOMEM);
498 id_priv->owner = task_pid_nr(current);
499 id_priv->state = RDMA_CM_IDLE;
500 id_priv->id.context = context;
501 id_priv->id.event_handler = event_handler;
503 id_priv->id.qp_type = qp_type;
504 spin_lock_init(&id_priv->lock);
505 mutex_init(&id_priv->qp_mutex);
506 init_completion(&id_priv->comp);
507 atomic_set(&id_priv->refcount, 1);
508 mutex_init(&id_priv->handler_mutex);
509 INIT_LIST_HEAD(&id_priv->listen_list);
510 INIT_LIST_HEAD(&id_priv->mc_list);
511 get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
515 EXPORT_SYMBOL(rdma_create_id);
517 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
519 struct ib_qp_attr qp_attr;
520 int qp_attr_mask, ret;
522 qp_attr.qp_state = IB_QPS_INIT;
523 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
527 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
531 qp_attr.qp_state = IB_QPS_RTR;
532 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
536 qp_attr.qp_state = IB_QPS_RTS;
538 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
543 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
545 struct ib_qp_attr qp_attr;
546 int qp_attr_mask, ret;
548 qp_attr.qp_state = IB_QPS_INIT;
549 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
553 return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
556 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
557 struct ib_qp_init_attr *qp_init_attr)
559 struct rdma_id_private *id_priv;
563 id_priv = container_of(id, struct rdma_id_private, id);
564 if (id->device != pd->device)
567 qp = ib_create_qp(pd, qp_init_attr);
571 if (id->qp_type == IB_QPT_UD)
572 ret = cma_init_ud_qp(id_priv, qp);
574 ret = cma_init_conn_qp(id_priv, qp);
579 id_priv->qp_num = qp->qp_num;
580 id_priv->srq = (qp->srq != NULL);
586 EXPORT_SYMBOL(rdma_create_qp);
588 void rdma_destroy_qp(struct rdma_cm_id *id)
590 struct rdma_id_private *id_priv;
592 id_priv = container_of(id, struct rdma_id_private, id);
593 mutex_lock(&id_priv->qp_mutex);
594 ib_destroy_qp(id_priv->id.qp);
595 id_priv->id.qp = NULL;
596 mutex_unlock(&id_priv->qp_mutex);
598 EXPORT_SYMBOL(rdma_destroy_qp);
600 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
601 struct rdma_conn_param *conn_param)
603 struct ib_qp_attr qp_attr;
604 int qp_attr_mask, ret;
606 mutex_lock(&id_priv->qp_mutex);
607 if (!id_priv->id.qp) {
612 /* Need to update QP attributes from default values. */
613 qp_attr.qp_state = IB_QPS_INIT;
614 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
618 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
622 qp_attr.qp_state = IB_QPS_RTR;
623 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
628 qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
629 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
631 mutex_unlock(&id_priv->qp_mutex);
635 static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
636 struct rdma_conn_param *conn_param)
638 struct ib_qp_attr qp_attr;
639 int qp_attr_mask, ret;
641 mutex_lock(&id_priv->qp_mutex);
642 if (!id_priv->id.qp) {
647 qp_attr.qp_state = IB_QPS_RTS;
648 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
653 qp_attr.max_rd_atomic = conn_param->initiator_depth;
654 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
656 mutex_unlock(&id_priv->qp_mutex);
660 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
662 struct ib_qp_attr qp_attr;
665 mutex_lock(&id_priv->qp_mutex);
666 if (!id_priv->id.qp) {
671 qp_attr.qp_state = IB_QPS_ERR;
672 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
674 mutex_unlock(&id_priv->qp_mutex);
678 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
679 struct ib_qp_attr *qp_attr, int *qp_attr_mask)
681 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
685 if (rdma_port_get_link_layer(id_priv->id.device, id_priv->id.port_num) ==
686 IB_LINK_LAYER_INFINIBAND)
687 pkey = ib_addr_get_pkey(dev_addr);
691 ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
692 pkey, &qp_attr->pkey_index);
696 qp_attr->port_num = id_priv->id.port_num;
697 *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
699 if (id_priv->id.qp_type == IB_QPT_UD) {
700 ret = cma_set_qkey(id_priv, 0);
704 qp_attr->qkey = id_priv->qkey;
705 *qp_attr_mask |= IB_QP_QKEY;
707 qp_attr->qp_access_flags = 0;
708 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
713 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
716 struct rdma_id_private *id_priv;
719 id_priv = container_of(id, struct rdma_id_private, id);
720 switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
721 case RDMA_TRANSPORT_IB:
722 if (!id_priv->cm_id.ib || (id_priv->id.qp_type == IB_QPT_UD))
723 ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
725 ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
727 if (qp_attr->qp_state == IB_QPS_RTR)
728 qp_attr->rq_psn = id_priv->seq_num;
730 case RDMA_TRANSPORT_IWARP:
731 if (!id_priv->cm_id.iw) {
732 qp_attr->qp_access_flags = 0;
733 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
735 ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
745 EXPORT_SYMBOL(rdma_init_qp_attr);
747 static inline int cma_zero_addr(struct sockaddr *addr)
749 switch (addr->sa_family) {
751 return ipv4_is_zeronet(((struct sockaddr_in *)addr)->sin_addr.s_addr);
753 return ipv6_addr_any(&((struct sockaddr_in6 *) addr)->sin6_addr);
755 return ib_addr_any(&((struct sockaddr_ib *) addr)->sib_addr);
761 static inline int cma_loopback_addr(struct sockaddr *addr)
763 switch (addr->sa_family) {
765 return ipv4_is_loopback(((struct sockaddr_in *) addr)->sin_addr.s_addr);
767 return ipv6_addr_loopback(&((struct sockaddr_in6 *) addr)->sin6_addr);
769 return ib_addr_loopback(&((struct sockaddr_ib *) addr)->sib_addr);
775 static inline int cma_any_addr(struct sockaddr *addr)
777 return cma_zero_addr(addr) || cma_loopback_addr(addr);
780 static int cma_addr_cmp(struct sockaddr *src, struct sockaddr *dst)
782 if (src->sa_family != dst->sa_family)
785 switch (src->sa_family) {
787 return ((struct sockaddr_in *) src)->sin_addr.s_addr !=
788 ((struct sockaddr_in *) dst)->sin_addr.s_addr;
790 return ipv6_addr_cmp(&((struct sockaddr_in6 *) src)->sin6_addr,
791 &((struct sockaddr_in6 *) dst)->sin6_addr);
793 return ib_addr_cmp(&((struct sockaddr_ib *) src)->sib_addr,
794 &((struct sockaddr_ib *) dst)->sib_addr);
798 static __be16 cma_port(struct sockaddr *addr)
800 struct sockaddr_ib *sib;
802 switch (addr->sa_family) {
804 return ((struct sockaddr_in *) addr)->sin_port;
806 return ((struct sockaddr_in6 *) addr)->sin6_port;
808 sib = (struct sockaddr_ib *) addr;
809 return htons((u16) (be64_to_cpu(sib->sib_sid) &
810 be64_to_cpu(sib->sib_sid_mask)));
816 static inline int cma_any_port(struct sockaddr *addr)
818 return !cma_port(addr);
821 static void cma_save_ib_info(struct rdma_cm_id *id, struct rdma_cm_id *listen_id,
822 struct ib_sa_path_rec *path)
824 struct sockaddr_ib *listen_ib, *ib;
826 listen_ib = (struct sockaddr_ib *) &listen_id->route.addr.src_addr;
827 ib = (struct sockaddr_ib *) &id->route.addr.src_addr;
828 ib->sib_family = listen_ib->sib_family;
829 ib->sib_pkey = path->pkey;
830 ib->sib_flowinfo = path->flow_label;
831 memcpy(&ib->sib_addr, &path->sgid, 16);
832 ib->sib_sid = listen_ib->sib_sid;
833 ib->sib_sid_mask = cpu_to_be64(0xffffffffffffffffULL);
834 ib->sib_scope_id = listen_ib->sib_scope_id;
836 ib = (struct sockaddr_ib *) &id->route.addr.dst_addr;
837 ib->sib_family = listen_ib->sib_family;
838 ib->sib_pkey = path->pkey;
839 ib->sib_flowinfo = path->flow_label;
840 memcpy(&ib->sib_addr, &path->dgid, 16);
843 static void cma_save_ip4_info(struct rdma_cm_id *id, struct rdma_cm_id *listen_id,
846 struct sockaddr_in *listen4, *ip4;
848 listen4 = (struct sockaddr_in *) &listen_id->route.addr.src_addr;
849 ip4 = (struct sockaddr_in *) &id->route.addr.src_addr;
850 ip4->sin_family = listen4->sin_family;
851 ip4->sin_addr.s_addr = hdr->dst_addr.ip4.addr;
852 ip4->sin_port = listen4->sin_port;
854 ip4 = (struct sockaddr_in *) &id->route.addr.dst_addr;
855 ip4->sin_family = listen4->sin_family;
856 ip4->sin_addr.s_addr = hdr->src_addr.ip4.addr;
857 ip4->sin_port = hdr->port;
860 static void cma_save_ip6_info(struct rdma_cm_id *id, struct rdma_cm_id *listen_id,
863 struct sockaddr_in6 *listen6, *ip6;
865 listen6 = (struct sockaddr_in6 *) &listen_id->route.addr.src_addr;
866 ip6 = (struct sockaddr_in6 *) &id->route.addr.src_addr;
867 ip6->sin6_family = listen6->sin6_family;
868 ip6->sin6_addr = hdr->dst_addr.ip6;
869 ip6->sin6_port = listen6->sin6_port;
871 ip6 = (struct sockaddr_in6 *) &id->route.addr.dst_addr;
872 ip6->sin6_family = listen6->sin6_family;
873 ip6->sin6_addr = hdr->src_addr.ip6;
874 ip6->sin6_port = hdr->port;
877 static int cma_save_net_info(struct rdma_cm_id *id, struct rdma_cm_id *listen_id,
878 struct ib_cm_event *ib_event)
882 if ((listen_id->route.addr.src_addr.ss_family == AF_IB) &&
883 (ib_event->event == IB_CM_REQ_RECEIVED)) {
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;
1389 struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
1390 struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
1392 if (cma_disable_callback(id_priv, RDMA_CM_CONNECT))
1395 memset(&event, 0, sizeof event);
1396 switch (iw_event->event) {
1397 case IW_CM_EVENT_CLOSE:
1398 event.event = RDMA_CM_EVENT_DISCONNECTED;
1400 case IW_CM_EVENT_CONNECT_REPLY:
1401 memcpy(cma_src_addr(id_priv), laddr,
1402 rdma_addr_size(laddr));
1403 memcpy(cma_dst_addr(id_priv), raddr,
1404 rdma_addr_size(raddr));
1405 switch (iw_event->status) {
1407 event.event = RDMA_CM_EVENT_ESTABLISHED;
1408 event.param.conn.initiator_depth = iw_event->ird;
1409 event.param.conn.responder_resources = iw_event->ord;
1413 event.event = RDMA_CM_EVENT_REJECTED;
1416 event.event = RDMA_CM_EVENT_UNREACHABLE;
1419 event.event = RDMA_CM_EVENT_CONNECT_ERROR;
1423 case IW_CM_EVENT_ESTABLISHED:
1424 event.event = RDMA_CM_EVENT_ESTABLISHED;
1425 event.param.conn.initiator_depth = iw_event->ird;
1426 event.param.conn.responder_resources = iw_event->ord;
1432 event.status = iw_event->status;
1433 event.param.conn.private_data = iw_event->private_data;
1434 event.param.conn.private_data_len = iw_event->private_data_len;
1435 ret = id_priv->id.event_handler(&id_priv->id, &event);
1437 /* Destroy the CM ID by returning a non-zero value. */
1438 id_priv->cm_id.iw = NULL;
1439 cma_exch(id_priv, RDMA_CM_DESTROYING);
1440 mutex_unlock(&id_priv->handler_mutex);
1441 rdma_destroy_id(&id_priv->id);
1445 mutex_unlock(&id_priv->handler_mutex);
1449 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
1450 struct iw_cm_event *iw_event)
1452 struct rdma_cm_id *new_cm_id;
1453 struct rdma_id_private *listen_id, *conn_id;
1454 struct net_device *dev = NULL;
1455 struct rdma_cm_event event;
1457 struct ib_device_attr attr;
1458 struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
1459 struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
1461 listen_id = cm_id->context;
1462 if (cma_disable_callback(listen_id, RDMA_CM_LISTEN))
1463 return -ECONNABORTED;
1465 /* Create a new RDMA id for the new IW CM ID */
1466 new_cm_id = rdma_create_id(listen_id->id.event_handler,
1467 listen_id->id.context,
1468 RDMA_PS_TCP, IB_QPT_RC);
1469 if (IS_ERR(new_cm_id)) {
1473 conn_id = container_of(new_cm_id, struct rdma_id_private, id);
1474 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1475 conn_id->state = RDMA_CM_CONNECT;
1477 ret = rdma_translate_ip(laddr, &conn_id->id.route.addr.dev_addr);
1479 mutex_unlock(&conn_id->handler_mutex);
1480 rdma_destroy_id(new_cm_id);
1484 ret = cma_acquire_dev(conn_id);
1486 mutex_unlock(&conn_id->handler_mutex);
1487 rdma_destroy_id(new_cm_id);
1491 conn_id->cm_id.iw = cm_id;
1492 cm_id->context = conn_id;
1493 cm_id->cm_handler = cma_iw_handler;
1495 memcpy(cma_src_addr(conn_id), laddr, rdma_addr_size(laddr));
1496 memcpy(cma_dst_addr(conn_id), raddr, rdma_addr_size(raddr));
1498 ret = ib_query_device(conn_id->id.device, &attr);
1500 mutex_unlock(&conn_id->handler_mutex);
1501 rdma_destroy_id(new_cm_id);
1505 memset(&event, 0, sizeof event);
1506 event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1507 event.param.conn.private_data = iw_event->private_data;
1508 event.param.conn.private_data_len = iw_event->private_data_len;
1509 event.param.conn.initiator_depth = iw_event->ird;
1510 event.param.conn.responder_resources = iw_event->ord;
1513 * Protect against the user destroying conn_id from another thread
1514 * until we're done accessing it.
1516 atomic_inc(&conn_id->refcount);
1517 ret = conn_id->id.event_handler(&conn_id->id, &event);
1519 /* User wants to destroy the CM ID */
1520 conn_id->cm_id.iw = NULL;
1521 cma_exch(conn_id, RDMA_CM_DESTROYING);
1522 mutex_unlock(&conn_id->handler_mutex);
1523 cma_deref_id(conn_id);
1524 rdma_destroy_id(&conn_id->id);
1528 mutex_unlock(&conn_id->handler_mutex);
1529 cma_deref_id(conn_id);
1534 mutex_unlock(&listen_id->handler_mutex);
1538 static int cma_ib_listen(struct rdma_id_private *id_priv)
1540 struct ib_cm_compare_data compare_data;
1541 struct sockaddr *addr;
1542 struct ib_cm_id *id;
1546 id = ib_create_cm_id(id_priv->id.device, cma_req_handler, id_priv);
1550 id_priv->cm_id.ib = id;
1552 addr = cma_src_addr(id_priv);
1553 svc_id = rdma_get_service_id(&id_priv->id, addr);
1554 if (cma_any_addr(addr) && !id_priv->afonly)
1555 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, NULL);
1557 cma_set_compare_data(id_priv->id.ps, addr, &compare_data);
1558 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, &compare_data);
1562 ib_destroy_cm_id(id_priv->cm_id.ib);
1563 id_priv->cm_id.ib = NULL;
1569 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
1572 struct iw_cm_id *id;
1574 id = iw_create_cm_id(id_priv->id.device,
1575 iw_conn_req_handler,
1580 id_priv->cm_id.iw = id;
1582 memcpy(&id_priv->cm_id.iw->local_addr, cma_src_addr(id_priv),
1583 rdma_addr_size(cma_src_addr(id_priv)));
1585 ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
1588 iw_destroy_cm_id(id_priv->cm_id.iw);
1589 id_priv->cm_id.iw = NULL;
1595 static int cma_listen_handler(struct rdma_cm_id *id,
1596 struct rdma_cm_event *event)
1598 struct rdma_id_private *id_priv = id->context;
1600 id->context = id_priv->id.context;
1601 id->event_handler = id_priv->id.event_handler;
1602 return id_priv->id.event_handler(id, event);
1605 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
1606 struct cma_device *cma_dev)
1608 struct rdma_id_private *dev_id_priv;
1609 struct rdma_cm_id *id;
1612 if (cma_family(id_priv) == AF_IB &&
1613 rdma_node_get_transport(cma_dev->device->node_type) != RDMA_TRANSPORT_IB)
1616 id = rdma_create_id(cma_listen_handler, id_priv, id_priv->id.ps,
1617 id_priv->id.qp_type);
1621 dev_id_priv = container_of(id, struct rdma_id_private, id);
1623 dev_id_priv->state = RDMA_CM_ADDR_BOUND;
1624 memcpy(cma_src_addr(dev_id_priv), cma_src_addr(id_priv),
1625 rdma_addr_size(cma_src_addr(id_priv)));
1627 cma_attach_to_dev(dev_id_priv, cma_dev);
1628 list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
1629 atomic_inc(&id_priv->refcount);
1630 dev_id_priv->internal_id = 1;
1631 dev_id_priv->afonly = id_priv->afonly;
1633 ret = rdma_listen(id, id_priv->backlog);
1635 printk(KERN_WARNING "RDMA CMA: cma_listen_on_dev, error %d, "
1636 "listening on device %s\n", ret, cma_dev->device->name);
1639 static void cma_listen_on_all(struct rdma_id_private *id_priv)
1641 struct cma_device *cma_dev;
1644 list_add_tail(&id_priv->list, &listen_any_list);
1645 list_for_each_entry(cma_dev, &dev_list, list)
1646 cma_listen_on_dev(id_priv, cma_dev);
1647 mutex_unlock(&lock);
1650 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
1652 struct rdma_id_private *id_priv;
1654 id_priv = container_of(id, struct rdma_id_private, id);
1655 id_priv->tos = (u8) tos;
1657 EXPORT_SYMBOL(rdma_set_service_type);
1659 static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec,
1662 struct cma_work *work = context;
1663 struct rdma_route *route;
1665 route = &work->id->id.route;
1668 route->num_paths = 1;
1669 *route->path_rec = *path_rec;
1671 work->old_state = RDMA_CM_ROUTE_QUERY;
1672 work->new_state = RDMA_CM_ADDR_RESOLVED;
1673 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
1674 work->event.status = status;
1677 queue_work(cma_wq, &work->work);
1680 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms,
1681 struct cma_work *work)
1683 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
1684 struct ib_sa_path_rec path_rec;
1685 ib_sa_comp_mask comp_mask;
1686 struct sockaddr_in6 *sin6;
1687 struct sockaddr_ib *sib;
1689 memset(&path_rec, 0, sizeof path_rec);
1690 rdma_addr_get_sgid(dev_addr, &path_rec.sgid);
1691 rdma_addr_get_dgid(dev_addr, &path_rec.dgid);
1692 path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
1693 path_rec.numb_path = 1;
1694 path_rec.reversible = 1;
1695 path_rec.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
1697 comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
1698 IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
1699 IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
1701 switch (cma_family(id_priv)) {
1703 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
1704 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
1707 sin6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
1708 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
1709 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
1712 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
1713 path_rec.traffic_class = (u8) (be32_to_cpu(sib->sib_flowinfo) >> 20);
1714 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
1718 id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
1719 id_priv->id.port_num, &path_rec,
1720 comp_mask, timeout_ms,
1721 GFP_KERNEL, cma_query_handler,
1722 work, &id_priv->query);
1724 return (id_priv->query_id < 0) ? id_priv->query_id : 0;
1727 static void cma_work_handler(struct work_struct *_work)
1729 struct cma_work *work = container_of(_work, struct cma_work, work);
1730 struct rdma_id_private *id_priv = work->id;
1733 mutex_lock(&id_priv->handler_mutex);
1734 if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
1737 if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1738 cma_exch(id_priv, RDMA_CM_DESTROYING);
1742 mutex_unlock(&id_priv->handler_mutex);
1743 cma_deref_id(id_priv);
1745 rdma_destroy_id(&id_priv->id);
1749 static void cma_ndev_work_handler(struct work_struct *_work)
1751 struct cma_ndev_work *work = container_of(_work, struct cma_ndev_work, work);
1752 struct rdma_id_private *id_priv = work->id;
1755 mutex_lock(&id_priv->handler_mutex);
1756 if (id_priv->state == RDMA_CM_DESTROYING ||
1757 id_priv->state == RDMA_CM_DEVICE_REMOVAL)
1760 if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1761 cma_exch(id_priv, RDMA_CM_DESTROYING);
1766 mutex_unlock(&id_priv->handler_mutex);
1767 cma_deref_id(id_priv);
1769 rdma_destroy_id(&id_priv->id);
1773 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms)
1775 struct rdma_route *route = &id_priv->id.route;
1776 struct cma_work *work;
1779 work = kzalloc(sizeof *work, GFP_KERNEL);
1784 INIT_WORK(&work->work, cma_work_handler);
1785 work->old_state = RDMA_CM_ROUTE_QUERY;
1786 work->new_state = RDMA_CM_ROUTE_RESOLVED;
1787 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1789 route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
1790 if (!route->path_rec) {
1795 ret = cma_query_ib_route(id_priv, timeout_ms, work);
1801 kfree(route->path_rec);
1802 route->path_rec = NULL;
1808 int rdma_set_ib_paths(struct rdma_cm_id *id,
1809 struct ib_sa_path_rec *path_rec, int num_paths)
1811 struct rdma_id_private *id_priv;
1814 id_priv = container_of(id, struct rdma_id_private, id);
1815 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
1816 RDMA_CM_ROUTE_RESOLVED))
1819 id->route.path_rec = kmemdup(path_rec, sizeof *path_rec * num_paths,
1821 if (!id->route.path_rec) {
1826 id->route.num_paths = num_paths;
1829 cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED);
1832 EXPORT_SYMBOL(rdma_set_ib_paths);
1834 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms)
1836 struct cma_work *work;
1838 work = kzalloc(sizeof *work, GFP_KERNEL);
1843 INIT_WORK(&work->work, cma_work_handler);
1844 work->old_state = RDMA_CM_ROUTE_QUERY;
1845 work->new_state = RDMA_CM_ROUTE_RESOLVED;
1846 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1847 queue_work(cma_wq, &work->work);
1851 static int iboe_tos_to_sl(struct net_device *ndev, int tos)
1854 struct net_device *dev;
1856 prio = rt_tos2priority(tos);
1857 dev = ndev->priv_flags & IFF_802_1Q_VLAN ?
1858 vlan_dev_real_dev(ndev) : ndev;
1861 return netdev_get_prio_tc_map(dev, prio);
1863 #if IS_ENABLED(CONFIG_VLAN_8021Q)
1864 if (ndev->priv_flags & IFF_802_1Q_VLAN)
1865 return (vlan_dev_get_egress_qos_mask(ndev, prio) &
1866 VLAN_PRIO_MASK) >> VLAN_PRIO_SHIFT;
1871 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
1873 struct rdma_route *route = &id_priv->id.route;
1874 struct rdma_addr *addr = &route->addr;
1875 struct cma_work *work;
1877 struct net_device *ndev = NULL;
1880 work = kzalloc(sizeof *work, GFP_KERNEL);
1885 INIT_WORK(&work->work, cma_work_handler);
1887 route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
1888 if (!route->path_rec) {
1893 route->num_paths = 1;
1895 if (addr->dev_addr.bound_dev_if)
1896 ndev = dev_get_by_index(&init_net, addr->dev_addr.bound_dev_if);
1902 vid = rdma_vlan_dev_vlan_id(ndev);
1904 iboe_mac_vlan_to_ll(&route->path_rec->sgid, addr->dev_addr.src_dev_addr, vid);
1905 iboe_mac_vlan_to_ll(&route->path_rec->dgid, addr->dev_addr.dst_dev_addr, vid);
1907 route->path_rec->hop_limit = 1;
1908 route->path_rec->reversible = 1;
1909 route->path_rec->pkey = cpu_to_be16(0xffff);
1910 route->path_rec->mtu_selector = IB_SA_EQ;
1911 route->path_rec->sl = iboe_tos_to_sl(ndev, id_priv->tos);
1912 route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
1913 route->path_rec->rate_selector = IB_SA_EQ;
1914 route->path_rec->rate = iboe_get_rate(ndev);
1916 route->path_rec->packet_life_time_selector = IB_SA_EQ;
1917 route->path_rec->packet_life_time = CMA_IBOE_PACKET_LIFETIME;
1918 if (!route->path_rec->mtu) {
1923 work->old_state = RDMA_CM_ROUTE_QUERY;
1924 work->new_state = RDMA_CM_ROUTE_RESOLVED;
1925 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1926 work->event.status = 0;
1928 queue_work(cma_wq, &work->work);
1933 kfree(route->path_rec);
1934 route->path_rec = NULL;
1940 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
1942 struct rdma_id_private *id_priv;
1945 id_priv = container_of(id, struct rdma_id_private, id);
1946 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY))
1949 atomic_inc(&id_priv->refcount);
1950 switch (rdma_node_get_transport(id->device->node_type)) {
1951 case RDMA_TRANSPORT_IB:
1952 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
1953 case IB_LINK_LAYER_INFINIBAND:
1954 ret = cma_resolve_ib_route(id_priv, timeout_ms);
1956 case IB_LINK_LAYER_ETHERNET:
1957 ret = cma_resolve_iboe_route(id_priv);
1963 case RDMA_TRANSPORT_IWARP:
1964 ret = cma_resolve_iw_route(id_priv, timeout_ms);
1975 cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED);
1976 cma_deref_id(id_priv);
1979 EXPORT_SYMBOL(rdma_resolve_route);
1981 static void cma_set_loopback(struct sockaddr *addr)
1983 switch (addr->sa_family) {
1985 ((struct sockaddr_in *) addr)->sin_addr.s_addr = htonl(INADDR_LOOPBACK);
1988 ipv6_addr_set(&((struct sockaddr_in6 *) addr)->sin6_addr,
1992 ib_addr_set(&((struct sockaddr_ib *) addr)->sib_addr,
1998 static int cma_bind_loopback(struct rdma_id_private *id_priv)
2000 struct cma_device *cma_dev, *cur_dev;
2001 struct ib_port_attr port_attr;
2009 list_for_each_entry(cur_dev, &dev_list, list) {
2010 if (cma_family(id_priv) == AF_IB &&
2011 rdma_node_get_transport(cur_dev->device->node_type) != RDMA_TRANSPORT_IB)
2017 for (p = 1; p <= cur_dev->device->phys_port_cnt; ++p) {
2018 if (!ib_query_port(cur_dev->device, p, &port_attr) &&
2019 port_attr.state == IB_PORT_ACTIVE) {
2034 ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid);
2038 ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
2042 id_priv->id.route.addr.dev_addr.dev_type =
2043 (rdma_port_get_link_layer(cma_dev->device, p) == IB_LINK_LAYER_INFINIBAND) ?
2044 ARPHRD_INFINIBAND : ARPHRD_ETHER;
2046 rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2047 ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
2048 id_priv->id.port_num = p;
2049 cma_attach_to_dev(id_priv, cma_dev);
2050 cma_set_loopback(cma_src_addr(id_priv));
2052 mutex_unlock(&lock);
2056 static void addr_handler(int status, struct sockaddr *src_addr,
2057 struct rdma_dev_addr *dev_addr, void *context)
2059 struct rdma_id_private *id_priv = context;
2060 struct rdma_cm_event event;
2062 memset(&event, 0, sizeof event);
2063 mutex_lock(&id_priv->handler_mutex);
2064 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY,
2065 RDMA_CM_ADDR_RESOLVED))
2068 if (!status && !id_priv->cma_dev)
2069 status = cma_acquire_dev(id_priv);
2072 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
2073 RDMA_CM_ADDR_BOUND))
2075 event.event = RDMA_CM_EVENT_ADDR_ERROR;
2076 event.status = status;
2078 memcpy(cma_src_addr(id_priv), src_addr, rdma_addr_size(src_addr));
2079 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2082 if (id_priv->id.event_handler(&id_priv->id, &event)) {
2083 cma_exch(id_priv, RDMA_CM_DESTROYING);
2084 mutex_unlock(&id_priv->handler_mutex);
2085 cma_deref_id(id_priv);
2086 rdma_destroy_id(&id_priv->id);
2090 mutex_unlock(&id_priv->handler_mutex);
2091 cma_deref_id(id_priv);
2094 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
2096 struct cma_work *work;
2100 work = kzalloc(sizeof *work, GFP_KERNEL);
2104 if (!id_priv->cma_dev) {
2105 ret = cma_bind_loopback(id_priv);
2110 rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2111 rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
2114 INIT_WORK(&work->work, cma_work_handler);
2115 work->old_state = RDMA_CM_ADDR_QUERY;
2116 work->new_state = RDMA_CM_ADDR_RESOLVED;
2117 work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2118 queue_work(cma_wq, &work->work);
2125 static int cma_resolve_ib_addr(struct rdma_id_private *id_priv)
2127 struct cma_work *work;
2130 work = kzalloc(sizeof *work, GFP_KERNEL);
2134 if (!id_priv->cma_dev) {
2135 ret = cma_resolve_ib_dev(id_priv);
2140 rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, (union ib_gid *)
2141 &(((struct sockaddr_ib *) &id_priv->id.route.addr.dst_addr)->sib_addr));
2144 INIT_WORK(&work->work, cma_work_handler);
2145 work->old_state = RDMA_CM_ADDR_QUERY;
2146 work->new_state = RDMA_CM_ADDR_RESOLVED;
2147 work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2148 queue_work(cma_wq, &work->work);
2155 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2156 struct sockaddr *dst_addr)
2158 if (!src_addr || !src_addr->sa_family) {
2159 src_addr = (struct sockaddr *) &id->route.addr.src_addr;
2160 src_addr->sa_family = dst_addr->sa_family;
2161 if (dst_addr->sa_family == AF_INET6) {
2162 ((struct sockaddr_in6 *) src_addr)->sin6_scope_id =
2163 ((struct sockaddr_in6 *) dst_addr)->sin6_scope_id;
2164 } else if (dst_addr->sa_family == AF_IB) {
2165 ((struct sockaddr_ib *) src_addr)->sib_pkey =
2166 ((struct sockaddr_ib *) dst_addr)->sib_pkey;
2169 return rdma_bind_addr(id, src_addr);
2172 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2173 struct sockaddr *dst_addr, int timeout_ms)
2175 struct rdma_id_private *id_priv;
2178 id_priv = container_of(id, struct rdma_id_private, id);
2179 if (id_priv->state == RDMA_CM_IDLE) {
2180 ret = cma_bind_addr(id, src_addr, dst_addr);
2185 if (cma_family(id_priv) != dst_addr->sa_family)
2188 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY))
2191 atomic_inc(&id_priv->refcount);
2192 memcpy(cma_dst_addr(id_priv), dst_addr, rdma_addr_size(dst_addr));
2193 if (cma_any_addr(dst_addr)) {
2194 ret = cma_resolve_loopback(id_priv);
2196 if (dst_addr->sa_family == AF_IB) {
2197 ret = cma_resolve_ib_addr(id_priv);
2199 ret = rdma_resolve_ip(&addr_client, cma_src_addr(id_priv),
2200 dst_addr, &id->route.addr.dev_addr,
2201 timeout_ms, addr_handler, id_priv);
2209 cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
2210 cma_deref_id(id_priv);
2213 EXPORT_SYMBOL(rdma_resolve_addr);
2215 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse)
2217 struct rdma_id_private *id_priv;
2218 unsigned long flags;
2221 id_priv = container_of(id, struct rdma_id_private, id);
2222 spin_lock_irqsave(&id_priv->lock, flags);
2223 if (reuse || id_priv->state == RDMA_CM_IDLE) {
2224 id_priv->reuseaddr = reuse;
2229 spin_unlock_irqrestore(&id_priv->lock, flags);
2232 EXPORT_SYMBOL(rdma_set_reuseaddr);
2234 int rdma_set_afonly(struct rdma_cm_id *id, int afonly)
2236 struct rdma_id_private *id_priv;
2237 unsigned long flags;
2240 id_priv = container_of(id, struct rdma_id_private, id);
2241 spin_lock_irqsave(&id_priv->lock, flags);
2242 if (id_priv->state == RDMA_CM_IDLE || id_priv->state == RDMA_CM_ADDR_BOUND) {
2243 id_priv->options |= (1 << CMA_OPTION_AFONLY);
2244 id_priv->afonly = afonly;
2249 spin_unlock_irqrestore(&id_priv->lock, flags);
2252 EXPORT_SYMBOL(rdma_set_afonly);
2254 static void cma_bind_port(struct rdma_bind_list *bind_list,
2255 struct rdma_id_private *id_priv)
2257 struct sockaddr *addr;
2258 struct sockaddr_ib *sib;
2262 addr = cma_src_addr(id_priv);
2263 port = htons(bind_list->port);
2265 switch (addr->sa_family) {
2267 ((struct sockaddr_in *) addr)->sin_port = port;
2270 ((struct sockaddr_in6 *) addr)->sin6_port = port;
2273 sib = (struct sockaddr_ib *) addr;
2274 sid = be64_to_cpu(sib->sib_sid);
2275 mask = be64_to_cpu(sib->sib_sid_mask);
2276 sib->sib_sid = cpu_to_be64((sid & mask) | (u64) ntohs(port));
2277 sib->sib_sid_mask = cpu_to_be64(~0ULL);
2280 id_priv->bind_list = bind_list;
2281 hlist_add_head(&id_priv->node, &bind_list->owners);
2284 static int cma_alloc_port(struct idr *ps, struct rdma_id_private *id_priv,
2285 unsigned short snum)
2287 struct rdma_bind_list *bind_list;
2290 bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
2294 ret = idr_alloc(ps, bind_list, snum, snum + 1, GFP_KERNEL);
2299 bind_list->port = (unsigned short)ret;
2300 cma_bind_port(bind_list, id_priv);
2304 return ret == -ENOSPC ? -EADDRNOTAVAIL : ret;
2307 static int cma_alloc_any_port(struct idr *ps, struct rdma_id_private *id_priv)
2309 static unsigned int last_used_port;
2310 int low, high, remaining;
2313 inet_get_local_port_range(&init_net, &low, &high);
2314 remaining = (high - low) + 1;
2315 rover = net_random() % remaining + low;
2317 if (last_used_port != rover &&
2318 !idr_find(ps, (unsigned short) rover)) {
2319 int ret = cma_alloc_port(ps, id_priv, rover);
2321 * Remember previously used port number in order to avoid
2322 * re-using same port immediately after it is closed.
2325 last_used_port = rover;
2326 if (ret != -EADDRNOTAVAIL)
2331 if ((rover < low) || (rover > high))
2335 return -EADDRNOTAVAIL;
2339 * Check that the requested port is available. This is called when trying to
2340 * bind to a specific port, or when trying to listen on a bound port. In
2341 * the latter case, the provided id_priv may already be on the bind_list, but
2342 * we still need to check that it's okay to start listening.
2344 static int cma_check_port(struct rdma_bind_list *bind_list,
2345 struct rdma_id_private *id_priv, uint8_t reuseaddr)
2347 struct rdma_id_private *cur_id;
2348 struct sockaddr *addr, *cur_addr;
2350 addr = cma_src_addr(id_priv);
2351 hlist_for_each_entry(cur_id, &bind_list->owners, node) {
2352 if (id_priv == cur_id)
2355 if ((cur_id->state != RDMA_CM_LISTEN) && reuseaddr &&
2359 cur_addr = cma_src_addr(cur_id);
2360 if (id_priv->afonly && cur_id->afonly &&
2361 (addr->sa_family != cur_addr->sa_family))
2364 if (cma_any_addr(addr) || cma_any_addr(cur_addr))
2365 return -EADDRNOTAVAIL;
2367 if (!cma_addr_cmp(addr, cur_addr))
2373 static int cma_use_port(struct idr *ps, struct rdma_id_private *id_priv)
2375 struct rdma_bind_list *bind_list;
2376 unsigned short snum;
2379 snum = ntohs(cma_port(cma_src_addr(id_priv)));
2380 if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
2383 bind_list = idr_find(ps, snum);
2385 ret = cma_alloc_port(ps, id_priv, snum);
2387 ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr);
2389 cma_bind_port(bind_list, id_priv);
2394 static int cma_bind_listen(struct rdma_id_private *id_priv)
2396 struct rdma_bind_list *bind_list = id_priv->bind_list;
2400 if (bind_list->owners.first->next)
2401 ret = cma_check_port(bind_list, id_priv, 0);
2402 mutex_unlock(&lock);
2406 static struct idr *cma_select_inet_ps(struct rdma_id_private *id_priv)
2408 switch (id_priv->id.ps) {
2422 static struct idr *cma_select_ib_ps(struct rdma_id_private *id_priv)
2424 struct idr *ps = NULL;
2425 struct sockaddr_ib *sib;
2426 u64 sid_ps, mask, sid;
2428 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
2429 mask = be64_to_cpu(sib->sib_sid_mask) & RDMA_IB_IP_PS_MASK;
2430 sid = be64_to_cpu(sib->sib_sid) & mask;
2432 if ((id_priv->id.ps == RDMA_PS_IB) && (sid == (RDMA_IB_IP_PS_IB & mask))) {
2433 sid_ps = RDMA_IB_IP_PS_IB;
2435 } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_TCP)) &&
2436 (sid == (RDMA_IB_IP_PS_TCP & mask))) {
2437 sid_ps = RDMA_IB_IP_PS_TCP;
2439 } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_UDP)) &&
2440 (sid == (RDMA_IB_IP_PS_UDP & mask))) {
2441 sid_ps = RDMA_IB_IP_PS_UDP;
2446 sib->sib_sid = cpu_to_be64(sid_ps | ntohs(cma_port((struct sockaddr *) sib)));
2447 sib->sib_sid_mask = cpu_to_be64(RDMA_IB_IP_PS_MASK |
2448 be64_to_cpu(sib->sib_sid_mask));
2453 static int cma_get_port(struct rdma_id_private *id_priv)
2458 if (cma_family(id_priv) != AF_IB)
2459 ps = cma_select_inet_ps(id_priv);
2461 ps = cma_select_ib_ps(id_priv);
2463 return -EPROTONOSUPPORT;
2466 if (cma_any_port(cma_src_addr(id_priv)))
2467 ret = cma_alloc_any_port(ps, id_priv);
2469 ret = cma_use_port(ps, id_priv);
2470 mutex_unlock(&lock);
2475 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
2476 struct sockaddr *addr)
2478 #if IS_ENABLED(CONFIG_IPV6)
2479 struct sockaddr_in6 *sin6;
2481 if (addr->sa_family != AF_INET6)
2484 sin6 = (struct sockaddr_in6 *) addr;
2485 if ((ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL) &&
2486 !sin6->sin6_scope_id)
2489 dev_addr->bound_dev_if = sin6->sin6_scope_id;
2494 int rdma_listen(struct rdma_cm_id *id, int backlog)
2496 struct rdma_id_private *id_priv;
2499 id_priv = container_of(id, struct rdma_id_private, id);
2500 if (id_priv->state == RDMA_CM_IDLE) {
2501 id->route.addr.src_addr.ss_family = AF_INET;
2502 ret = rdma_bind_addr(id, cma_src_addr(id_priv));
2507 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN))
2510 if (id_priv->reuseaddr) {
2511 ret = cma_bind_listen(id_priv);
2516 id_priv->backlog = backlog;
2518 switch (rdma_node_get_transport(id->device->node_type)) {
2519 case RDMA_TRANSPORT_IB:
2520 ret = cma_ib_listen(id_priv);
2524 case RDMA_TRANSPORT_IWARP:
2525 ret = cma_iw_listen(id_priv, backlog);
2534 cma_listen_on_all(id_priv);
2538 id_priv->backlog = 0;
2539 cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND);
2542 EXPORT_SYMBOL(rdma_listen);
2544 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
2546 struct rdma_id_private *id_priv;
2549 if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6 &&
2550 addr->sa_family != AF_IB)
2551 return -EAFNOSUPPORT;
2553 id_priv = container_of(id, struct rdma_id_private, id);
2554 if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND))
2557 ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
2561 if (!cma_any_addr(addr)) {
2562 ret = cma_translate_addr(addr, &id->route.addr.dev_addr);
2566 ret = cma_acquire_dev(id_priv);
2571 memcpy(cma_src_addr(id_priv), addr, rdma_addr_size(addr));
2572 if (!(id_priv->options & (1 << CMA_OPTION_AFONLY))) {
2573 if (addr->sa_family == AF_INET)
2574 id_priv->afonly = 1;
2575 #if IS_ENABLED(CONFIG_IPV6)
2576 else if (addr->sa_family == AF_INET6)
2577 id_priv->afonly = init_net.ipv6.sysctl.bindv6only;
2580 ret = cma_get_port(id_priv);
2586 if (id_priv->cma_dev)
2587 cma_release_dev(id_priv);
2589 cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE);
2592 EXPORT_SYMBOL(rdma_bind_addr);
2594 static int cma_format_hdr(void *hdr, struct rdma_id_private *id_priv)
2596 struct cma_hdr *cma_hdr;
2599 cma_hdr->cma_version = CMA_VERSION;
2600 if (cma_family(id_priv) == AF_INET) {
2601 struct sockaddr_in *src4, *dst4;
2603 src4 = (struct sockaddr_in *) cma_src_addr(id_priv);
2604 dst4 = (struct sockaddr_in *) cma_dst_addr(id_priv);
2606 cma_set_ip_ver(cma_hdr, 4);
2607 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2608 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2609 cma_hdr->port = src4->sin_port;
2610 } else if (cma_family(id_priv) == AF_INET6) {
2611 struct sockaddr_in6 *src6, *dst6;
2613 src6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
2614 dst6 = (struct sockaddr_in6 *) cma_dst_addr(id_priv);
2616 cma_set_ip_ver(cma_hdr, 6);
2617 cma_hdr->src_addr.ip6 = src6->sin6_addr;
2618 cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
2619 cma_hdr->port = src6->sin6_port;
2624 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
2625 struct ib_cm_event *ib_event)
2627 struct rdma_id_private *id_priv = cm_id->context;
2628 struct rdma_cm_event event;
2629 struct ib_cm_sidr_rep_event_param *rep = &ib_event->param.sidr_rep_rcvd;
2632 if (cma_disable_callback(id_priv, RDMA_CM_CONNECT))
2635 memset(&event, 0, sizeof event);
2636 switch (ib_event->event) {
2637 case IB_CM_SIDR_REQ_ERROR:
2638 event.event = RDMA_CM_EVENT_UNREACHABLE;
2639 event.status = -ETIMEDOUT;
2641 case IB_CM_SIDR_REP_RECEIVED:
2642 event.param.ud.private_data = ib_event->private_data;
2643 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
2644 if (rep->status != IB_SIDR_SUCCESS) {
2645 event.event = RDMA_CM_EVENT_UNREACHABLE;
2646 event.status = ib_event->param.sidr_rep_rcvd.status;
2649 ret = cma_set_qkey(id_priv, rep->qkey);
2651 event.event = RDMA_CM_EVENT_ADDR_ERROR;
2655 ib_init_ah_from_path(id_priv->id.device, id_priv->id.port_num,
2656 id_priv->id.route.path_rec,
2657 &event.param.ud.ah_attr);
2658 event.param.ud.qp_num = rep->qpn;
2659 event.param.ud.qkey = rep->qkey;
2660 event.event = RDMA_CM_EVENT_ESTABLISHED;
2664 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
2669 ret = id_priv->id.event_handler(&id_priv->id, &event);
2671 /* Destroy the CM ID by returning a non-zero value. */
2672 id_priv->cm_id.ib = NULL;
2673 cma_exch(id_priv, RDMA_CM_DESTROYING);
2674 mutex_unlock(&id_priv->handler_mutex);
2675 rdma_destroy_id(&id_priv->id);
2679 mutex_unlock(&id_priv->handler_mutex);
2683 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
2684 struct rdma_conn_param *conn_param)
2686 struct ib_cm_sidr_req_param req;
2687 struct ib_cm_id *id;
2691 memset(&req, 0, sizeof req);
2692 offset = cma_user_data_offset(id_priv);
2693 req.private_data_len = offset + conn_param->private_data_len;
2694 if (req.private_data_len < conn_param->private_data_len)
2697 if (req.private_data_len) {
2698 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2702 private_data = NULL;
2705 if (conn_param->private_data && conn_param->private_data_len)
2706 memcpy(private_data + offset, conn_param->private_data,
2707 conn_param->private_data_len);
2710 ret = cma_format_hdr(private_data, id_priv);
2713 req.private_data = private_data;
2716 id = ib_create_cm_id(id_priv->id.device, cma_sidr_rep_handler,
2722 id_priv->cm_id.ib = id;
2724 req.path = id_priv->id.route.path_rec;
2725 req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
2726 req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
2727 req.max_cm_retries = CMA_MAX_CM_RETRIES;
2729 ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
2731 ib_destroy_cm_id(id_priv->cm_id.ib);
2732 id_priv->cm_id.ib = NULL;
2735 kfree(private_data);
2739 static int cma_connect_ib(struct rdma_id_private *id_priv,
2740 struct rdma_conn_param *conn_param)
2742 struct ib_cm_req_param req;
2743 struct rdma_route *route;
2745 struct ib_cm_id *id;
2748 memset(&req, 0, sizeof req);
2749 offset = cma_user_data_offset(id_priv);
2750 req.private_data_len = offset + conn_param->private_data_len;
2751 if (req.private_data_len < conn_param->private_data_len)
2754 if (req.private_data_len) {
2755 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2759 private_data = NULL;
2762 if (conn_param->private_data && conn_param->private_data_len)
2763 memcpy(private_data + offset, conn_param->private_data,
2764 conn_param->private_data_len);
2766 id = ib_create_cm_id(id_priv->id.device, cma_ib_handler, id_priv);
2771 id_priv->cm_id.ib = id;
2773 route = &id_priv->id.route;
2775 ret = cma_format_hdr(private_data, id_priv);
2778 req.private_data = private_data;
2781 req.primary_path = &route->path_rec[0];
2782 if (route->num_paths == 2)
2783 req.alternate_path = &route->path_rec[1];
2785 req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
2786 req.qp_num = id_priv->qp_num;
2787 req.qp_type = id_priv->id.qp_type;
2788 req.starting_psn = id_priv->seq_num;
2789 req.responder_resources = conn_param->responder_resources;
2790 req.initiator_depth = conn_param->initiator_depth;
2791 req.flow_control = conn_param->flow_control;
2792 req.retry_count = min_t(u8, 7, conn_param->retry_count);
2793 req.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
2794 req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2795 req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2796 req.max_cm_retries = CMA_MAX_CM_RETRIES;
2797 req.srq = id_priv->srq ? 1 : 0;
2799 ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
2801 if (ret && !IS_ERR(id)) {
2802 ib_destroy_cm_id(id);
2803 id_priv->cm_id.ib = NULL;
2806 kfree(private_data);
2810 static int cma_connect_iw(struct rdma_id_private *id_priv,
2811 struct rdma_conn_param *conn_param)
2813 struct iw_cm_id *cm_id;
2815 struct iw_cm_conn_param iw_param;
2817 cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
2819 return PTR_ERR(cm_id);
2821 id_priv->cm_id.iw = cm_id;
2823 memcpy(&cm_id->local_addr, cma_src_addr(id_priv),
2824 rdma_addr_size(cma_src_addr(id_priv)));
2825 memcpy(&cm_id->remote_addr, cma_dst_addr(id_priv),
2826 rdma_addr_size(cma_dst_addr(id_priv)));
2828 ret = cma_modify_qp_rtr(id_priv, conn_param);
2833 iw_param.ord = conn_param->initiator_depth;
2834 iw_param.ird = conn_param->responder_resources;
2835 iw_param.private_data = conn_param->private_data;
2836 iw_param.private_data_len = conn_param->private_data_len;
2837 iw_param.qpn = id_priv->id.qp ? id_priv->qp_num : conn_param->qp_num;
2839 memset(&iw_param, 0, sizeof iw_param);
2840 iw_param.qpn = id_priv->qp_num;
2842 ret = iw_cm_connect(cm_id, &iw_param);
2845 iw_destroy_cm_id(cm_id);
2846 id_priv->cm_id.iw = NULL;
2851 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2853 struct rdma_id_private *id_priv;
2856 id_priv = container_of(id, struct rdma_id_private, id);
2857 if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT))
2861 id_priv->qp_num = conn_param->qp_num;
2862 id_priv->srq = conn_param->srq;
2865 switch (rdma_node_get_transport(id->device->node_type)) {
2866 case RDMA_TRANSPORT_IB:
2867 if (id->qp_type == IB_QPT_UD)
2868 ret = cma_resolve_ib_udp(id_priv, conn_param);
2870 ret = cma_connect_ib(id_priv, conn_param);
2872 case RDMA_TRANSPORT_IWARP:
2873 ret = cma_connect_iw(id_priv, conn_param);
2884 cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED);
2887 EXPORT_SYMBOL(rdma_connect);
2889 static int cma_accept_ib(struct rdma_id_private *id_priv,
2890 struct rdma_conn_param *conn_param)
2892 struct ib_cm_rep_param rep;
2895 ret = cma_modify_qp_rtr(id_priv, conn_param);
2899 ret = cma_modify_qp_rts(id_priv, conn_param);
2903 memset(&rep, 0, sizeof rep);
2904 rep.qp_num = id_priv->qp_num;
2905 rep.starting_psn = id_priv->seq_num;
2906 rep.private_data = conn_param->private_data;
2907 rep.private_data_len = conn_param->private_data_len;
2908 rep.responder_resources = conn_param->responder_resources;
2909 rep.initiator_depth = conn_param->initiator_depth;
2910 rep.failover_accepted = 0;
2911 rep.flow_control = conn_param->flow_control;
2912 rep.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
2913 rep.srq = id_priv->srq ? 1 : 0;
2915 ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
2920 static int cma_accept_iw(struct rdma_id_private *id_priv,
2921 struct rdma_conn_param *conn_param)
2923 struct iw_cm_conn_param iw_param;
2926 ret = cma_modify_qp_rtr(id_priv, conn_param);
2930 iw_param.ord = conn_param->initiator_depth;
2931 iw_param.ird = conn_param->responder_resources;
2932 iw_param.private_data = conn_param->private_data;
2933 iw_param.private_data_len = conn_param->private_data_len;
2934 if (id_priv->id.qp) {
2935 iw_param.qpn = id_priv->qp_num;
2937 iw_param.qpn = conn_param->qp_num;
2939 return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
2942 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
2943 enum ib_cm_sidr_status status, u32 qkey,
2944 const void *private_data, int private_data_len)
2946 struct ib_cm_sidr_rep_param rep;
2949 memset(&rep, 0, sizeof rep);
2950 rep.status = status;
2951 if (status == IB_SIDR_SUCCESS) {
2952 ret = cma_set_qkey(id_priv, qkey);
2955 rep.qp_num = id_priv->qp_num;
2956 rep.qkey = id_priv->qkey;
2958 rep.private_data = private_data;
2959 rep.private_data_len = private_data_len;
2961 return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
2964 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2966 struct rdma_id_private *id_priv;
2969 id_priv = container_of(id, struct rdma_id_private, id);
2971 id_priv->owner = task_pid_nr(current);
2973 if (!cma_comp(id_priv, RDMA_CM_CONNECT))
2976 if (!id->qp && conn_param) {
2977 id_priv->qp_num = conn_param->qp_num;
2978 id_priv->srq = conn_param->srq;
2981 switch (rdma_node_get_transport(id->device->node_type)) {
2982 case RDMA_TRANSPORT_IB:
2983 if (id->qp_type == IB_QPT_UD) {
2985 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
2987 conn_param->private_data,
2988 conn_param->private_data_len);
2990 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
2994 ret = cma_accept_ib(id_priv, conn_param);
2996 ret = cma_rep_recv(id_priv);
2999 case RDMA_TRANSPORT_IWARP:
3000 ret = cma_accept_iw(id_priv, conn_param);
3012 cma_modify_qp_err(id_priv);
3013 rdma_reject(id, NULL, 0);
3016 EXPORT_SYMBOL(rdma_accept);
3018 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
3020 struct rdma_id_private *id_priv;
3023 id_priv = container_of(id, struct rdma_id_private, id);
3024 if (!id_priv->cm_id.ib)
3027 switch (id->device->node_type) {
3028 case RDMA_NODE_IB_CA:
3029 ret = ib_cm_notify(id_priv->cm_id.ib, event);
3037 EXPORT_SYMBOL(rdma_notify);
3039 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
3040 u8 private_data_len)
3042 struct rdma_id_private *id_priv;
3045 id_priv = container_of(id, struct rdma_id_private, id);
3046 if (!id_priv->cm_id.ib)
3049 switch (rdma_node_get_transport(id->device->node_type)) {
3050 case RDMA_TRANSPORT_IB:
3051 if (id->qp_type == IB_QPT_UD)
3052 ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT, 0,
3053 private_data, private_data_len);
3055 ret = ib_send_cm_rej(id_priv->cm_id.ib,
3056 IB_CM_REJ_CONSUMER_DEFINED, NULL,
3057 0, private_data, private_data_len);
3059 case RDMA_TRANSPORT_IWARP:
3060 ret = iw_cm_reject(id_priv->cm_id.iw,
3061 private_data, private_data_len);
3069 EXPORT_SYMBOL(rdma_reject);
3071 int rdma_disconnect(struct rdma_cm_id *id)
3073 struct rdma_id_private *id_priv;
3076 id_priv = container_of(id, struct rdma_id_private, id);
3077 if (!id_priv->cm_id.ib)
3080 switch (rdma_node_get_transport(id->device->node_type)) {
3081 case RDMA_TRANSPORT_IB:
3082 ret = cma_modify_qp_err(id_priv);
3085 /* Initiate or respond to a disconnect. */
3086 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
3087 ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
3089 case RDMA_TRANSPORT_IWARP:
3090 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
3099 EXPORT_SYMBOL(rdma_disconnect);
3101 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
3103 struct rdma_id_private *id_priv;
3104 struct cma_multicast *mc = multicast->context;
3105 struct rdma_cm_event event;
3108 id_priv = mc->id_priv;
3109 if (cma_disable_callback(id_priv, RDMA_CM_ADDR_BOUND) &&
3110 cma_disable_callback(id_priv, RDMA_CM_ADDR_RESOLVED))
3114 status = cma_set_qkey(id_priv, be32_to_cpu(multicast->rec.qkey));
3115 mutex_lock(&id_priv->qp_mutex);
3116 if (!status && id_priv->id.qp)
3117 status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid,
3118 be16_to_cpu(multicast->rec.mlid));
3119 mutex_unlock(&id_priv->qp_mutex);
3121 memset(&event, 0, sizeof event);
3122 event.status = status;
3123 event.param.ud.private_data = mc->context;
3125 event.event = RDMA_CM_EVENT_MULTICAST_JOIN;
3126 ib_init_ah_from_mcmember(id_priv->id.device,
3127 id_priv->id.port_num, &multicast->rec,
3128 &event.param.ud.ah_attr);
3129 event.param.ud.qp_num = 0xFFFFFF;
3130 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
3132 event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
3134 ret = id_priv->id.event_handler(&id_priv->id, &event);
3136 cma_exch(id_priv, RDMA_CM_DESTROYING);
3137 mutex_unlock(&id_priv->handler_mutex);
3138 rdma_destroy_id(&id_priv->id);
3142 mutex_unlock(&id_priv->handler_mutex);
3146 static void cma_set_mgid(struct rdma_id_private *id_priv,
3147 struct sockaddr *addr, union ib_gid *mgid)
3149 unsigned char mc_map[MAX_ADDR_LEN];
3150 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3151 struct sockaddr_in *sin = (struct sockaddr_in *) addr;
3152 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
3154 if (cma_any_addr(addr)) {
3155 memset(mgid, 0, sizeof *mgid);
3156 } else if ((addr->sa_family == AF_INET6) &&
3157 ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
3159 /* IPv6 address is an SA assigned MGID. */
3160 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3161 } else if (addr->sa_family == AF_IB) {
3162 memcpy(mgid, &((struct sockaddr_ib *) addr)->sib_addr, sizeof *mgid);
3163 } else if ((addr->sa_family == AF_INET6)) {
3164 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
3165 if (id_priv->id.ps == RDMA_PS_UDP)
3166 mc_map[7] = 0x01; /* Use RDMA CM signature */
3167 *mgid = *(union ib_gid *) (mc_map + 4);
3169 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
3170 if (id_priv->id.ps == RDMA_PS_UDP)
3171 mc_map[7] = 0x01; /* Use RDMA CM signature */
3172 *mgid = *(union ib_gid *) (mc_map + 4);
3176 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
3177 struct cma_multicast *mc)
3179 struct ib_sa_mcmember_rec rec;
3180 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3181 ib_sa_comp_mask comp_mask;
3184 ib_addr_get_mgid(dev_addr, &rec.mgid);
3185 ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
3190 ret = cma_set_qkey(id_priv, 0);
3194 cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
3195 rec.qkey = cpu_to_be32(id_priv->qkey);
3196 rdma_addr_get_sgid(dev_addr, &rec.port_gid);
3197 rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
3200 comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
3201 IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
3202 IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
3203 IB_SA_MCMEMBER_REC_FLOW_LABEL |
3204 IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
3206 if (id_priv->id.ps == RDMA_PS_IPOIB)
3207 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
3208 IB_SA_MCMEMBER_REC_RATE_SELECTOR |
3209 IB_SA_MCMEMBER_REC_MTU_SELECTOR |
3210 IB_SA_MCMEMBER_REC_MTU |
3211 IB_SA_MCMEMBER_REC_HOP_LIMIT;
3213 mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device,
3214 id_priv->id.port_num, &rec,
3215 comp_mask, GFP_KERNEL,
3216 cma_ib_mc_handler, mc);
3217 return PTR_ERR_OR_ZERO(mc->multicast.ib);
3220 static void iboe_mcast_work_handler(struct work_struct *work)
3222 struct iboe_mcast_work *mw = container_of(work, struct iboe_mcast_work, work);
3223 struct cma_multicast *mc = mw->mc;
3224 struct ib_sa_multicast *m = mc->multicast.ib;
3226 mc->multicast.ib->context = mc;
3227 cma_ib_mc_handler(0, m);
3228 kref_put(&mc->mcref, release_mc);
3232 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid)
3234 struct sockaddr_in *sin = (struct sockaddr_in *)addr;
3235 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
3237 if (cma_any_addr(addr)) {
3238 memset(mgid, 0, sizeof *mgid);
3239 } else if (addr->sa_family == AF_INET6) {
3240 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3242 mgid->raw[0] = 0xff;
3243 mgid->raw[1] = 0x0e;
3252 mgid->raw[10] = 0xff;
3253 mgid->raw[11] = 0xff;
3254 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
3258 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
3259 struct cma_multicast *mc)
3261 struct iboe_mcast_work *work;
3262 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3264 struct sockaddr *addr = (struct sockaddr *)&mc->addr;
3265 struct net_device *ndev = NULL;
3267 if (cma_zero_addr((struct sockaddr *)&mc->addr))
3270 work = kzalloc(sizeof *work, GFP_KERNEL);
3274 mc->multicast.ib = kzalloc(sizeof(struct ib_sa_multicast), GFP_KERNEL);
3275 if (!mc->multicast.ib) {
3280 cma_iboe_set_mgid(addr, &mc->multicast.ib->rec.mgid);
3282 mc->multicast.ib->rec.pkey = cpu_to_be16(0xffff);
3283 if (id_priv->id.ps == RDMA_PS_UDP)
3284 mc->multicast.ib->rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
3286 if (dev_addr->bound_dev_if)
3287 ndev = dev_get_by_index(&init_net, dev_addr->bound_dev_if);
3292 mc->multicast.ib->rec.rate = iboe_get_rate(ndev);
3293 mc->multicast.ib->rec.hop_limit = 1;
3294 mc->multicast.ib->rec.mtu = iboe_get_mtu(ndev->mtu);
3296 if (!mc->multicast.ib->rec.mtu) {
3300 iboe_addr_get_sgid(dev_addr, &mc->multicast.ib->rec.port_gid);
3303 INIT_WORK(&work->work, iboe_mcast_work_handler);
3304 kref_get(&mc->mcref);
3305 queue_work(cma_wq, &work->work);
3310 kfree(mc->multicast.ib);
3316 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
3319 struct rdma_id_private *id_priv;
3320 struct cma_multicast *mc;
3323 id_priv = container_of(id, struct rdma_id_private, id);
3324 if (!cma_comp(id_priv, RDMA_CM_ADDR_BOUND) &&
3325 !cma_comp(id_priv, RDMA_CM_ADDR_RESOLVED))
3328 mc = kmalloc(sizeof *mc, GFP_KERNEL);
3332 memcpy(&mc->addr, addr, rdma_addr_size(addr));
3333 mc->context = context;
3334 mc->id_priv = id_priv;
3336 spin_lock(&id_priv->lock);
3337 list_add(&mc->list, &id_priv->mc_list);
3338 spin_unlock(&id_priv->lock);
3340 switch (rdma_node_get_transport(id->device->node_type)) {
3341 case RDMA_TRANSPORT_IB:
3342 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
3343 case IB_LINK_LAYER_INFINIBAND:
3344 ret = cma_join_ib_multicast(id_priv, mc);
3346 case IB_LINK_LAYER_ETHERNET:
3347 kref_init(&mc->mcref);
3348 ret = cma_iboe_join_multicast(id_priv, mc);
3360 spin_lock_irq(&id_priv->lock);
3361 list_del(&mc->list);
3362 spin_unlock_irq(&id_priv->lock);
3367 EXPORT_SYMBOL(rdma_join_multicast);
3369 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
3371 struct rdma_id_private *id_priv;
3372 struct cma_multicast *mc;
3374 id_priv = container_of(id, struct rdma_id_private, id);
3375 spin_lock_irq(&id_priv->lock);
3376 list_for_each_entry(mc, &id_priv->mc_list, list) {
3377 if (!memcmp(&mc->addr, addr, rdma_addr_size(addr))) {
3378 list_del(&mc->list);
3379 spin_unlock_irq(&id_priv->lock);
3382 ib_detach_mcast(id->qp,
3383 &mc->multicast.ib->rec.mgid,
3384 be16_to_cpu(mc->multicast.ib->rec.mlid));
3385 if (rdma_node_get_transport(id_priv->cma_dev->device->node_type) == RDMA_TRANSPORT_IB) {
3386 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
3387 case IB_LINK_LAYER_INFINIBAND:
3388 ib_sa_free_multicast(mc->multicast.ib);
3391 case IB_LINK_LAYER_ETHERNET:
3392 kref_put(&mc->mcref, release_mc);
3401 spin_unlock_irq(&id_priv->lock);
3403 EXPORT_SYMBOL(rdma_leave_multicast);
3405 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
3407 struct rdma_dev_addr *dev_addr;
3408 struct cma_ndev_work *work;
3410 dev_addr = &id_priv->id.route.addr.dev_addr;
3412 if ((dev_addr->bound_dev_if == ndev->ifindex) &&
3413 memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
3414 printk(KERN_INFO "RDMA CM addr change for ndev %s used by id %p\n",
3415 ndev->name, &id_priv->id);
3416 work = kzalloc(sizeof *work, GFP_KERNEL);
3420 INIT_WORK(&work->work, cma_ndev_work_handler);
3422 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
3423 atomic_inc(&id_priv->refcount);
3424 queue_work(cma_wq, &work->work);
3430 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
3433 struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
3434 struct cma_device *cma_dev;
3435 struct rdma_id_private *id_priv;
3436 int ret = NOTIFY_DONE;
3438 if (dev_net(ndev) != &init_net)
3441 if (event != NETDEV_BONDING_FAILOVER)
3444 if (!(ndev->flags & IFF_MASTER) || !(ndev->priv_flags & IFF_BONDING))
3448 list_for_each_entry(cma_dev, &dev_list, list)
3449 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
3450 ret = cma_netdev_change(ndev, id_priv);
3456 mutex_unlock(&lock);
3460 static struct notifier_block cma_nb = {
3461 .notifier_call = cma_netdev_callback
3464 static void cma_add_one(struct ib_device *device)
3466 struct cma_device *cma_dev;
3467 struct rdma_id_private *id_priv;
3469 cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
3473 cma_dev->device = device;
3475 init_completion(&cma_dev->comp);
3476 atomic_set(&cma_dev->refcount, 1);
3477 INIT_LIST_HEAD(&cma_dev->id_list);
3478 ib_set_client_data(device, &cma_client, cma_dev);
3481 list_add_tail(&cma_dev->list, &dev_list);
3482 list_for_each_entry(id_priv, &listen_any_list, list)
3483 cma_listen_on_dev(id_priv, cma_dev);
3484 mutex_unlock(&lock);
3487 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
3489 struct rdma_cm_event event;
3490 enum rdma_cm_state state;
3493 /* Record that we want to remove the device */
3494 state = cma_exch(id_priv, RDMA_CM_DEVICE_REMOVAL);
3495 if (state == RDMA_CM_DESTROYING)
3498 cma_cancel_operation(id_priv, state);
3499 mutex_lock(&id_priv->handler_mutex);
3501 /* Check for destruction from another callback. */
3502 if (!cma_comp(id_priv, RDMA_CM_DEVICE_REMOVAL))
3505 memset(&event, 0, sizeof event);
3506 event.event = RDMA_CM_EVENT_DEVICE_REMOVAL;
3507 ret = id_priv->id.event_handler(&id_priv->id, &event);
3509 mutex_unlock(&id_priv->handler_mutex);
3513 static void cma_process_remove(struct cma_device *cma_dev)
3515 struct rdma_id_private *id_priv;
3519 while (!list_empty(&cma_dev->id_list)) {
3520 id_priv = list_entry(cma_dev->id_list.next,
3521 struct rdma_id_private, list);
3523 list_del(&id_priv->listen_list);
3524 list_del_init(&id_priv->list);
3525 atomic_inc(&id_priv->refcount);
3526 mutex_unlock(&lock);
3528 ret = id_priv->internal_id ? 1 : cma_remove_id_dev(id_priv);
3529 cma_deref_id(id_priv);
3531 rdma_destroy_id(&id_priv->id);
3535 mutex_unlock(&lock);
3537 cma_deref_dev(cma_dev);
3538 wait_for_completion(&cma_dev->comp);
3541 static void cma_remove_one(struct ib_device *device)
3543 struct cma_device *cma_dev;
3545 cma_dev = ib_get_client_data(device, &cma_client);
3550 list_del(&cma_dev->list);
3551 mutex_unlock(&lock);
3553 cma_process_remove(cma_dev);
3557 static int cma_get_id_stats(struct sk_buff *skb, struct netlink_callback *cb)
3559 struct nlmsghdr *nlh;
3560 struct rdma_cm_id_stats *id_stats;
3561 struct rdma_id_private *id_priv;
3562 struct rdma_cm_id *id = NULL;
3563 struct cma_device *cma_dev;
3564 int i_dev = 0, i_id = 0;
3567 * We export all of the IDs as a sequence of messages. Each
3568 * ID gets its own netlink message.
3572 list_for_each_entry(cma_dev, &dev_list, list) {
3573 if (i_dev < cb->args[0]) {
3579 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
3580 if (i_id < cb->args[1]) {
3585 id_stats = ibnl_put_msg(skb, &nlh, cb->nlh->nlmsg_seq,
3586 sizeof *id_stats, RDMA_NL_RDMA_CM,
3587 RDMA_NL_RDMA_CM_ID_STATS);
3591 memset(id_stats, 0, sizeof *id_stats);
3593 id_stats->node_type = id->route.addr.dev_addr.dev_type;
3594 id_stats->port_num = id->port_num;
3595 id_stats->bound_dev_if =
3596 id->route.addr.dev_addr.bound_dev_if;
3598 if (ibnl_put_attr(skb, nlh,
3599 rdma_addr_size(cma_src_addr(id_priv)),
3600 cma_src_addr(id_priv),
3601 RDMA_NL_RDMA_CM_ATTR_SRC_ADDR))
3603 if (ibnl_put_attr(skb, nlh,
3604 rdma_addr_size(cma_src_addr(id_priv)),
3605 cma_dst_addr(id_priv),
3606 RDMA_NL_RDMA_CM_ATTR_DST_ADDR))
3609 id_stats->pid = id_priv->owner;
3610 id_stats->port_space = id->ps;
3611 id_stats->cm_state = id_priv->state;
3612 id_stats->qp_num = id_priv->qp_num;
3613 id_stats->qp_type = id->qp_type;
3623 mutex_unlock(&lock);
3624 cb->args[0] = i_dev;
3630 static const struct ibnl_client_cbs cma_cb_table[] = {
3631 [RDMA_NL_RDMA_CM_ID_STATS] = { .dump = cma_get_id_stats,
3632 .module = THIS_MODULE },
3635 static int __init cma_init(void)
3639 cma_wq = create_singlethread_workqueue("rdma_cm");
3643 ib_sa_register_client(&sa_client);
3644 rdma_addr_register_client(&addr_client);
3645 register_netdevice_notifier(&cma_nb);
3647 ret = ib_register_client(&cma_client);
3651 if (ibnl_add_client(RDMA_NL_RDMA_CM, RDMA_NL_RDMA_CM_NUM_OPS, cma_cb_table))
3652 printk(KERN_WARNING "RDMA CMA: failed to add netlink callback\n");
3657 unregister_netdevice_notifier(&cma_nb);
3658 rdma_addr_unregister_client(&addr_client);
3659 ib_sa_unregister_client(&sa_client);
3660 destroy_workqueue(cma_wq);
3664 static void __exit cma_cleanup(void)
3666 ibnl_remove_client(RDMA_NL_RDMA_CM);
3667 ib_unregister_client(&cma_client);
3668 unregister_netdevice_notifier(&cma_nb);
3669 rdma_addr_unregister_client(&addr_client);
3670 ib_sa_unregister_client(&sa_client);
3671 destroy_workqueue(cma_wq);
3672 idr_destroy(&tcp_ps);
3673 idr_destroy(&udp_ps);
3674 idr_destroy(&ipoib_ps);
3675 idr_destroy(&ib_ps);
3678 module_init(cma_init);
3679 module_exit(cma_cleanup);