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(sdp_ps);
84 static DEFINE_IDR(tcp_ps);
85 static DEFINE_IDR(udp_ps);
86 static DEFINE_IDR(ipoib_ps);
87 static DEFINE_IDR(ib_ps);
90 struct list_head list;
91 struct ib_device *device;
92 struct completion comp;
94 struct list_head id_list;
97 struct rdma_bind_list {
99 struct hlist_head owners;
108 * Device removal can occur at anytime, so we need extra handling to
109 * serialize notifying the user of device removal with other callbacks.
110 * We do this by disabling removal notification while a callback is in process,
111 * and reporting it after the callback completes.
113 struct rdma_id_private {
114 struct rdma_cm_id id;
116 struct rdma_bind_list *bind_list;
117 struct hlist_node node;
118 struct list_head list; /* listen_any_list or cma_device.list */
119 struct list_head listen_list; /* per device listens */
120 struct cma_device *cma_dev;
121 struct list_head mc_list;
124 enum rdma_cm_state state;
126 struct mutex qp_mutex;
128 struct completion comp;
130 struct mutex handler_mutex;
134 struct ib_sa_query *query;
152 struct cma_multicast {
153 struct rdma_id_private *id_priv;
155 struct ib_sa_multicast *ib;
157 struct list_head list;
159 struct sockaddr_storage addr;
164 struct work_struct work;
165 struct rdma_id_private *id;
166 enum rdma_cm_state old_state;
167 enum rdma_cm_state new_state;
168 struct rdma_cm_event event;
171 struct cma_ndev_work {
172 struct work_struct work;
173 struct rdma_id_private *id;
174 struct rdma_cm_event event;
177 struct iboe_mcast_work {
178 struct work_struct work;
179 struct rdma_id_private *id;
180 struct cma_multicast *mc;
193 u8 ip_version; /* IP version: 7:4 */
195 union cma_ip_addr src_addr;
196 union cma_ip_addr dst_addr;
201 u8 sdp_version; /* Major version: 7:4 */
202 u8 ip_version; /* IP version: 7:4 */
203 u8 sdp_specific1[10];
205 __be16 sdp_specific2;
206 union cma_ip_addr src_addr;
207 union cma_ip_addr dst_addr;
215 #define CMA_VERSION 0x00
216 #define SDP_MAJ_VERSION 0x2
218 static int cma_comp(struct rdma_id_private *id_priv, enum rdma_cm_state comp)
223 spin_lock_irqsave(&id_priv->lock, flags);
224 ret = (id_priv->state == comp);
225 spin_unlock_irqrestore(&id_priv->lock, flags);
229 static int cma_comp_exch(struct rdma_id_private *id_priv,
230 enum rdma_cm_state comp, enum rdma_cm_state exch)
235 spin_lock_irqsave(&id_priv->lock, flags);
236 if ((ret = (id_priv->state == comp)))
237 id_priv->state = exch;
238 spin_unlock_irqrestore(&id_priv->lock, flags);
242 static enum rdma_cm_state cma_exch(struct rdma_id_private *id_priv,
243 enum rdma_cm_state exch)
246 enum rdma_cm_state old;
248 spin_lock_irqsave(&id_priv->lock, flags);
249 old = id_priv->state;
250 id_priv->state = exch;
251 spin_unlock_irqrestore(&id_priv->lock, flags);
255 static inline u8 cma_get_ip_ver(struct cma_hdr *hdr)
257 return hdr->ip_version >> 4;
260 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
262 hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
265 static inline u8 sdp_get_majv(u8 sdp_version)
267 return sdp_version >> 4;
270 static inline u8 sdp_get_ip_ver(struct sdp_hh *hh)
272 return hh->ip_version >> 4;
275 static inline void sdp_set_ip_ver(struct sdp_hh *hh, u8 ip_ver)
277 hh->ip_version = (ip_ver << 4) | (hh->ip_version & 0xF);
280 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
281 struct cma_device *cma_dev)
283 atomic_inc(&cma_dev->refcount);
284 id_priv->cma_dev = cma_dev;
285 id_priv->id.device = cma_dev->device;
286 id_priv->id.route.addr.dev_addr.transport =
287 rdma_node_get_transport(cma_dev->device->node_type);
288 list_add_tail(&id_priv->list, &cma_dev->id_list);
291 static inline void cma_deref_dev(struct cma_device *cma_dev)
293 if (atomic_dec_and_test(&cma_dev->refcount))
294 complete(&cma_dev->comp);
297 static inline void release_mc(struct kref *kref)
299 struct cma_multicast *mc = container_of(kref, struct cma_multicast, mcref);
301 kfree(mc->multicast.ib);
305 static void cma_release_dev(struct rdma_id_private *id_priv)
308 list_del(&id_priv->list);
309 cma_deref_dev(id_priv->cma_dev);
310 id_priv->cma_dev = NULL;
314 static inline struct sockaddr *cma_src_addr(struct rdma_id_private *id_priv)
316 return (struct sockaddr *) &id_priv->id.route.addr.src_addr;
319 static inline struct sockaddr *cma_dst_addr(struct rdma_id_private *id_priv)
321 return (struct sockaddr *) &id_priv->id.route.addr.dst_addr;
324 static inline unsigned short cma_family(struct rdma_id_private *id_priv)
326 return id_priv->id.route.addr.src_addr.ss_family;
329 static int cma_set_qkey(struct rdma_id_private *id_priv)
331 struct ib_sa_mcmember_rec rec;
337 switch (id_priv->id.ps) {
339 id_priv->qkey = RDMA_UDP_QKEY;
342 ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid);
343 ret = ib_sa_get_mcmember_rec(id_priv->id.device,
344 id_priv->id.port_num, &rec.mgid,
347 id_priv->qkey = be32_to_cpu(rec.qkey);
355 static int find_gid_port(struct ib_device *device, union ib_gid *gid, u8 port_num)
359 struct ib_port_attr props;
362 err = ib_query_port(device, port_num, &props);
366 for (i = 0; i < props.gid_tbl_len; ++i) {
367 err = ib_query_gid(device, port_num, i, &tmp);
370 if (!memcmp(&tmp, gid, sizeof tmp))
374 return -EADDRNOTAVAIL;
377 static void cma_translate_ib(struct sockaddr_ib *sib, struct rdma_dev_addr *dev_addr)
379 dev_addr->dev_type = ARPHRD_INFINIBAND;
380 rdma_addr_set_sgid(dev_addr, (union ib_gid *) &sib->sib_addr);
381 ib_addr_set_pkey(dev_addr, ntohs(sib->sib_pkey));
384 static int cma_translate_addr(struct sockaddr *addr, struct rdma_dev_addr *dev_addr)
388 if (addr->sa_family != AF_IB) {
389 ret = rdma_translate_ip(addr, dev_addr);
391 cma_translate_ib((struct sockaddr_ib *) addr, dev_addr);
398 static int cma_acquire_dev(struct rdma_id_private *id_priv)
400 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
401 struct cma_device *cma_dev;
402 union ib_gid gid, iboe_gid;
405 enum rdma_link_layer dev_ll = dev_addr->dev_type == ARPHRD_INFINIBAND ?
406 IB_LINK_LAYER_INFINIBAND : IB_LINK_LAYER_ETHERNET;
408 if (dev_ll != IB_LINK_LAYER_INFINIBAND &&
409 id_priv->id.ps == RDMA_PS_IPOIB)
413 iboe_addr_get_sgid(dev_addr, &iboe_gid);
414 memcpy(&gid, dev_addr->src_dev_addr +
415 rdma_addr_gid_offset(dev_addr), sizeof gid);
416 list_for_each_entry(cma_dev, &dev_list, list) {
417 for (port = 1; port <= cma_dev->device->phys_port_cnt; ++port) {
418 if (rdma_port_get_link_layer(cma_dev->device, port) == dev_ll) {
419 if (rdma_node_get_transport(cma_dev->device->node_type) == RDMA_TRANSPORT_IB &&
420 rdma_port_get_link_layer(cma_dev->device, port) == IB_LINK_LAYER_ETHERNET)
421 ret = find_gid_port(cma_dev->device, &iboe_gid, port);
423 ret = find_gid_port(cma_dev->device, &gid, port);
426 id_priv->id.port_num = port;
435 cma_attach_to_dev(id_priv, cma_dev);
442 * Select the source IB device and address to reach the destination IB address.
444 static int cma_resolve_ib_dev(struct rdma_id_private *id_priv)
446 struct cma_device *cma_dev, *cur_dev;
447 struct sockaddr_ib *addr;
448 union ib_gid gid, sgid, *dgid;
454 addr = (struct sockaddr_ib *) cma_dst_addr(id_priv);
455 dgid = (union ib_gid *) &addr->sib_addr;
456 pkey = ntohs(addr->sib_pkey);
458 list_for_each_entry(cur_dev, &dev_list, list) {
459 if (rdma_node_get_transport(cur_dev->device->node_type) != RDMA_TRANSPORT_IB)
462 for (p = 1; p <= cur_dev->device->phys_port_cnt; ++p) {
463 if (ib_find_cached_pkey(cur_dev->device, p, pkey, &index))
466 for (i = 0; !ib_get_cached_gid(cur_dev->device, p, i, &gid); i++) {
467 if (!memcmp(&gid, dgid, sizeof(gid))) {
474 if (!cma_dev && (gid.global.subnet_prefix ==
475 dgid->global.subnet_prefix)) {
488 cma_attach_to_dev(id_priv, cma_dev);
489 id_priv->id.port_num = port;
490 addr = (struct sockaddr_ib *) cma_src_addr(id_priv);
491 memcpy(&addr->sib_addr, &sgid, sizeof sgid);
492 cma_translate_ib(addr, &id_priv->id.route.addr.dev_addr);
496 static void cma_deref_id(struct rdma_id_private *id_priv)
498 if (atomic_dec_and_test(&id_priv->refcount))
499 complete(&id_priv->comp);
502 static int cma_disable_callback(struct rdma_id_private *id_priv,
503 enum rdma_cm_state state)
505 mutex_lock(&id_priv->handler_mutex);
506 if (id_priv->state != state) {
507 mutex_unlock(&id_priv->handler_mutex);
513 struct rdma_cm_id *rdma_create_id(rdma_cm_event_handler event_handler,
514 void *context, enum rdma_port_space ps,
515 enum ib_qp_type qp_type)
517 struct rdma_id_private *id_priv;
519 id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
521 return ERR_PTR(-ENOMEM);
523 id_priv->owner = task_pid_nr(current);
524 id_priv->state = RDMA_CM_IDLE;
525 id_priv->id.context = context;
526 id_priv->id.event_handler = event_handler;
528 id_priv->id.qp_type = qp_type;
529 spin_lock_init(&id_priv->lock);
530 mutex_init(&id_priv->qp_mutex);
531 init_completion(&id_priv->comp);
532 atomic_set(&id_priv->refcount, 1);
533 mutex_init(&id_priv->handler_mutex);
534 INIT_LIST_HEAD(&id_priv->listen_list);
535 INIT_LIST_HEAD(&id_priv->mc_list);
536 get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
540 EXPORT_SYMBOL(rdma_create_id);
542 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
544 struct ib_qp_attr qp_attr;
545 int qp_attr_mask, ret;
547 qp_attr.qp_state = IB_QPS_INIT;
548 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
552 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
556 qp_attr.qp_state = IB_QPS_RTR;
557 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
561 qp_attr.qp_state = IB_QPS_RTS;
563 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
568 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
570 struct ib_qp_attr qp_attr;
571 int qp_attr_mask, ret;
573 qp_attr.qp_state = IB_QPS_INIT;
574 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
578 return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
581 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
582 struct ib_qp_init_attr *qp_init_attr)
584 struct rdma_id_private *id_priv;
588 id_priv = container_of(id, struct rdma_id_private, id);
589 if (id->device != pd->device)
592 qp = ib_create_qp(pd, qp_init_attr);
596 if (id->qp_type == IB_QPT_UD)
597 ret = cma_init_ud_qp(id_priv, qp);
599 ret = cma_init_conn_qp(id_priv, qp);
604 id_priv->qp_num = qp->qp_num;
605 id_priv->srq = (qp->srq != NULL);
611 EXPORT_SYMBOL(rdma_create_qp);
613 void rdma_destroy_qp(struct rdma_cm_id *id)
615 struct rdma_id_private *id_priv;
617 id_priv = container_of(id, struct rdma_id_private, id);
618 mutex_lock(&id_priv->qp_mutex);
619 ib_destroy_qp(id_priv->id.qp);
620 id_priv->id.qp = NULL;
621 mutex_unlock(&id_priv->qp_mutex);
623 EXPORT_SYMBOL(rdma_destroy_qp);
625 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
626 struct rdma_conn_param *conn_param)
628 struct ib_qp_attr qp_attr;
629 int qp_attr_mask, ret;
631 mutex_lock(&id_priv->qp_mutex);
632 if (!id_priv->id.qp) {
637 /* Need to update QP attributes from default values. */
638 qp_attr.qp_state = IB_QPS_INIT;
639 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
643 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
647 qp_attr.qp_state = IB_QPS_RTR;
648 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
653 qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
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_rts(struct rdma_id_private *id_priv,
661 struct rdma_conn_param *conn_param)
663 struct ib_qp_attr qp_attr;
664 int qp_attr_mask, ret;
666 mutex_lock(&id_priv->qp_mutex);
667 if (!id_priv->id.qp) {
672 qp_attr.qp_state = IB_QPS_RTS;
673 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
678 qp_attr.max_rd_atomic = conn_param->initiator_depth;
679 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
681 mutex_unlock(&id_priv->qp_mutex);
685 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
687 struct ib_qp_attr qp_attr;
690 mutex_lock(&id_priv->qp_mutex);
691 if (!id_priv->id.qp) {
696 qp_attr.qp_state = IB_QPS_ERR;
697 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
699 mutex_unlock(&id_priv->qp_mutex);
703 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
704 struct ib_qp_attr *qp_attr, int *qp_attr_mask)
706 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
710 if (rdma_port_get_link_layer(id_priv->id.device, id_priv->id.port_num) ==
711 IB_LINK_LAYER_INFINIBAND)
712 pkey = ib_addr_get_pkey(dev_addr);
716 ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
717 pkey, &qp_attr->pkey_index);
721 qp_attr->port_num = id_priv->id.port_num;
722 *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
724 if (id_priv->id.qp_type == IB_QPT_UD) {
725 ret = cma_set_qkey(id_priv);
729 qp_attr->qkey = id_priv->qkey;
730 *qp_attr_mask |= IB_QP_QKEY;
732 qp_attr->qp_access_flags = 0;
733 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
738 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
741 struct rdma_id_private *id_priv;
744 id_priv = container_of(id, struct rdma_id_private, id);
745 switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
746 case RDMA_TRANSPORT_IB:
747 if (!id_priv->cm_id.ib || (id_priv->id.qp_type == IB_QPT_UD))
748 ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
750 ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
752 if (qp_attr->qp_state == IB_QPS_RTR)
753 qp_attr->rq_psn = id_priv->seq_num;
755 case RDMA_TRANSPORT_IWARP:
756 if (!id_priv->cm_id.iw) {
757 qp_attr->qp_access_flags = 0;
758 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
760 ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
770 EXPORT_SYMBOL(rdma_init_qp_attr);
772 static inline int cma_zero_addr(struct sockaddr *addr)
774 switch (addr->sa_family) {
776 return ipv4_is_zeronet(((struct sockaddr_in *)addr)->sin_addr.s_addr);
778 return ipv6_addr_any(&((struct sockaddr_in6 *) addr)->sin6_addr);
780 return ib_addr_any(&((struct sockaddr_ib *) addr)->sib_addr);
786 static inline int cma_loopback_addr(struct sockaddr *addr)
788 switch (addr->sa_family) {
790 return ipv4_is_loopback(((struct sockaddr_in *) addr)->sin_addr.s_addr);
792 return ipv6_addr_loopback(&((struct sockaddr_in6 *) addr)->sin6_addr);
794 return ib_addr_loopback(&((struct sockaddr_ib *) addr)->sib_addr);
800 static inline int cma_any_addr(struct sockaddr *addr)
802 return cma_zero_addr(addr) || cma_loopback_addr(addr);
805 static int cma_addr_cmp(struct sockaddr *src, struct sockaddr *dst)
807 if (src->sa_family != dst->sa_family)
810 switch (src->sa_family) {
812 return ((struct sockaddr_in *) src)->sin_addr.s_addr !=
813 ((struct sockaddr_in *) dst)->sin_addr.s_addr;
815 return ipv6_addr_cmp(&((struct sockaddr_in6 *) src)->sin6_addr,
816 &((struct sockaddr_in6 *) dst)->sin6_addr);
818 return ib_addr_cmp(&((struct sockaddr_ib *) src)->sib_addr,
819 &((struct sockaddr_ib *) dst)->sib_addr);
823 static __be16 cma_port(struct sockaddr *addr)
825 struct sockaddr_ib *sib;
827 switch (addr->sa_family) {
829 return ((struct sockaddr_in *) addr)->sin_port;
831 return ((struct sockaddr_in6 *) addr)->sin6_port;
833 sib = (struct sockaddr_ib *) addr;
834 return htons((u16) (be64_to_cpu(sib->sib_sid) &
835 be64_to_cpu(sib->sib_sid_mask)));
841 static inline int cma_any_port(struct sockaddr *addr)
843 return !cma_port(addr);
846 static int cma_get_net_info(void *hdr, enum rdma_port_space ps,
847 u8 *ip_ver, __be16 *port,
848 union cma_ip_addr **src, union cma_ip_addr **dst)
852 if (sdp_get_majv(((struct sdp_hh *) hdr)->sdp_version) !=
856 *ip_ver = sdp_get_ip_ver(hdr);
857 *port = ((struct sdp_hh *) hdr)->port;
858 *src = &((struct sdp_hh *) hdr)->src_addr;
859 *dst = &((struct sdp_hh *) hdr)->dst_addr;
862 if (((struct cma_hdr *) hdr)->cma_version != CMA_VERSION)
865 *ip_ver = cma_get_ip_ver(hdr);
866 *port = ((struct cma_hdr *) hdr)->port;
867 *src = &((struct cma_hdr *) hdr)->src_addr;
868 *dst = &((struct cma_hdr *) hdr)->dst_addr;
872 if (*ip_ver != 4 && *ip_ver != 6)
877 static void cma_save_net_info(struct rdma_addr *addr,
878 struct rdma_addr *listen_addr,
879 u8 ip_ver, __be16 port,
880 union cma_ip_addr *src, union cma_ip_addr *dst)
882 struct sockaddr_in *listen4, *ip4;
883 struct sockaddr_in6 *listen6, *ip6;
887 listen4 = (struct sockaddr_in *) &listen_addr->src_addr;
888 ip4 = (struct sockaddr_in *) &addr->src_addr;
889 ip4->sin_family = listen4->sin_family;
890 ip4->sin_addr.s_addr = dst->ip4.addr;
891 ip4->sin_port = listen4->sin_port;
893 ip4 = (struct sockaddr_in *) &addr->dst_addr;
894 ip4->sin_family = listen4->sin_family;
895 ip4->sin_addr.s_addr = src->ip4.addr;
896 ip4->sin_port = port;
899 listen6 = (struct sockaddr_in6 *) &listen_addr->src_addr;
900 ip6 = (struct sockaddr_in6 *) &addr->src_addr;
901 ip6->sin6_family = listen6->sin6_family;
902 ip6->sin6_addr = dst->ip6;
903 ip6->sin6_port = listen6->sin6_port;
905 ip6 = (struct sockaddr_in6 *) &addr->dst_addr;
906 ip6->sin6_family = listen6->sin6_family;
907 ip6->sin6_addr = src->ip6;
908 ip6->sin6_port = port;
915 static inline int cma_user_data_offset(enum rdma_port_space ps)
921 return sizeof(struct cma_hdr);
925 static void cma_cancel_route(struct rdma_id_private *id_priv)
927 switch (rdma_port_get_link_layer(id_priv->id.device, id_priv->id.port_num)) {
928 case IB_LINK_LAYER_INFINIBAND:
930 ib_sa_cancel_query(id_priv->query_id, id_priv->query);
937 static void cma_cancel_listens(struct rdma_id_private *id_priv)
939 struct rdma_id_private *dev_id_priv;
942 * Remove from listen_any_list to prevent added devices from spawning
943 * additional listen requests.
946 list_del(&id_priv->list);
948 while (!list_empty(&id_priv->listen_list)) {
949 dev_id_priv = list_entry(id_priv->listen_list.next,
950 struct rdma_id_private, listen_list);
951 /* sync with device removal to avoid duplicate destruction */
952 list_del_init(&dev_id_priv->list);
953 list_del(&dev_id_priv->listen_list);
956 rdma_destroy_id(&dev_id_priv->id);
962 static void cma_cancel_operation(struct rdma_id_private *id_priv,
963 enum rdma_cm_state state)
966 case RDMA_CM_ADDR_QUERY:
967 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
969 case RDMA_CM_ROUTE_QUERY:
970 cma_cancel_route(id_priv);
973 if (cma_any_addr(cma_src_addr(id_priv)) && !id_priv->cma_dev)
974 cma_cancel_listens(id_priv);
981 static void cma_release_port(struct rdma_id_private *id_priv)
983 struct rdma_bind_list *bind_list = id_priv->bind_list;
989 hlist_del(&id_priv->node);
990 if (hlist_empty(&bind_list->owners)) {
991 idr_remove(bind_list->ps, bind_list->port);
997 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
999 struct cma_multicast *mc;
1001 while (!list_empty(&id_priv->mc_list)) {
1002 mc = container_of(id_priv->mc_list.next,
1003 struct cma_multicast, list);
1004 list_del(&mc->list);
1005 switch (rdma_port_get_link_layer(id_priv->cma_dev->device, id_priv->id.port_num)) {
1006 case IB_LINK_LAYER_INFINIBAND:
1007 ib_sa_free_multicast(mc->multicast.ib);
1010 case IB_LINK_LAYER_ETHERNET:
1011 kref_put(&mc->mcref, release_mc);
1019 void rdma_destroy_id(struct rdma_cm_id *id)
1021 struct rdma_id_private *id_priv;
1022 enum rdma_cm_state state;
1024 id_priv = container_of(id, struct rdma_id_private, id);
1025 state = cma_exch(id_priv, RDMA_CM_DESTROYING);
1026 cma_cancel_operation(id_priv, state);
1029 * Wait for any active callback to finish. New callbacks will find
1030 * the id_priv state set to destroying and abort.
1032 mutex_lock(&id_priv->handler_mutex);
1033 mutex_unlock(&id_priv->handler_mutex);
1035 if (id_priv->cma_dev) {
1036 switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
1037 case RDMA_TRANSPORT_IB:
1038 if (id_priv->cm_id.ib)
1039 ib_destroy_cm_id(id_priv->cm_id.ib);
1041 case RDMA_TRANSPORT_IWARP:
1042 if (id_priv->cm_id.iw)
1043 iw_destroy_cm_id(id_priv->cm_id.iw);
1048 cma_leave_mc_groups(id_priv);
1049 cma_release_dev(id_priv);
1052 cma_release_port(id_priv);
1053 cma_deref_id(id_priv);
1054 wait_for_completion(&id_priv->comp);
1056 if (id_priv->internal_id)
1057 cma_deref_id(id_priv->id.context);
1059 kfree(id_priv->id.route.path_rec);
1062 EXPORT_SYMBOL(rdma_destroy_id);
1064 static int cma_rep_recv(struct rdma_id_private *id_priv)
1068 ret = cma_modify_qp_rtr(id_priv, NULL);
1072 ret = cma_modify_qp_rts(id_priv, NULL);
1076 ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
1082 cma_modify_qp_err(id_priv);
1083 ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
1088 static int cma_verify_rep(struct rdma_id_private *id_priv, void *data)
1090 if (id_priv->id.ps == RDMA_PS_SDP &&
1091 sdp_get_majv(((struct sdp_hah *) data)->sdp_version) !=
1098 static void cma_set_rep_event_data(struct rdma_cm_event *event,
1099 struct ib_cm_rep_event_param *rep_data,
1102 event->param.conn.private_data = private_data;
1103 event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
1104 event->param.conn.responder_resources = rep_data->responder_resources;
1105 event->param.conn.initiator_depth = rep_data->initiator_depth;
1106 event->param.conn.flow_control = rep_data->flow_control;
1107 event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
1108 event->param.conn.srq = rep_data->srq;
1109 event->param.conn.qp_num = rep_data->remote_qpn;
1112 static int cma_ib_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1114 struct rdma_id_private *id_priv = cm_id->context;
1115 struct rdma_cm_event event;
1118 if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
1119 cma_disable_callback(id_priv, RDMA_CM_CONNECT)) ||
1120 (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
1121 cma_disable_callback(id_priv, RDMA_CM_DISCONNECT)))
1124 memset(&event, 0, sizeof event);
1125 switch (ib_event->event) {
1126 case IB_CM_REQ_ERROR:
1127 case IB_CM_REP_ERROR:
1128 event.event = RDMA_CM_EVENT_UNREACHABLE;
1129 event.status = -ETIMEDOUT;
1131 case IB_CM_REP_RECEIVED:
1132 event.status = cma_verify_rep(id_priv, ib_event->private_data);
1134 event.event = RDMA_CM_EVENT_CONNECT_ERROR;
1135 else if (id_priv->id.qp && id_priv->id.ps != RDMA_PS_SDP) {
1136 event.status = cma_rep_recv(id_priv);
1137 event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
1138 RDMA_CM_EVENT_ESTABLISHED;
1140 event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
1141 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
1142 ib_event->private_data);
1144 case IB_CM_RTU_RECEIVED:
1145 case IB_CM_USER_ESTABLISHED:
1146 event.event = RDMA_CM_EVENT_ESTABLISHED;
1148 case IB_CM_DREQ_ERROR:
1149 event.status = -ETIMEDOUT; /* fall through */
1150 case IB_CM_DREQ_RECEIVED:
1151 case IB_CM_DREP_RECEIVED:
1152 if (!cma_comp_exch(id_priv, RDMA_CM_CONNECT,
1153 RDMA_CM_DISCONNECT))
1155 event.event = RDMA_CM_EVENT_DISCONNECTED;
1157 case IB_CM_TIMEWAIT_EXIT:
1158 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
1160 case IB_CM_MRA_RECEIVED:
1163 case IB_CM_REJ_RECEIVED:
1164 cma_modify_qp_err(id_priv);
1165 event.status = ib_event->param.rej_rcvd.reason;
1166 event.event = RDMA_CM_EVENT_REJECTED;
1167 event.param.conn.private_data = ib_event->private_data;
1168 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
1171 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
1176 ret = id_priv->id.event_handler(&id_priv->id, &event);
1178 /* Destroy the CM ID by returning a non-zero value. */
1179 id_priv->cm_id.ib = NULL;
1180 cma_exch(id_priv, RDMA_CM_DESTROYING);
1181 mutex_unlock(&id_priv->handler_mutex);
1182 rdma_destroy_id(&id_priv->id);
1186 mutex_unlock(&id_priv->handler_mutex);
1190 static struct rdma_id_private *cma_new_conn_id(struct rdma_cm_id *listen_id,
1191 struct ib_cm_event *ib_event)
1193 struct rdma_id_private *id_priv;
1194 struct rdma_cm_id *id;
1195 struct rdma_route *rt;
1196 union cma_ip_addr *src, *dst;
1201 if (cma_get_net_info(ib_event->private_data, listen_id->ps,
1202 &ip_ver, &port, &src, &dst))
1205 id = rdma_create_id(listen_id->event_handler, listen_id->context,
1206 listen_id->ps, ib_event->param.req_rcvd.qp_type);
1210 id_priv = container_of(id, struct rdma_id_private, id);
1211 cma_save_net_info(&id->route.addr, &listen_id->route.addr,
1212 ip_ver, port, src, dst);
1215 rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
1216 rt->path_rec = kmalloc(sizeof *rt->path_rec * rt->num_paths,
1221 rt->path_rec[0] = *ib_event->param.req_rcvd.primary_path;
1222 if (rt->num_paths == 2)
1223 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
1225 if (cma_any_addr(cma_src_addr(id_priv))) {
1226 rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
1227 rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
1228 ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
1230 ret = cma_translate_addr(cma_src_addr(id_priv), &rt->addr.dev_addr);
1234 rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
1236 id_priv->state = RDMA_CM_CONNECT;
1240 rdma_destroy_id(id);
1244 static struct rdma_id_private *cma_new_udp_id(struct rdma_cm_id *listen_id,
1245 struct ib_cm_event *ib_event)
1247 struct rdma_id_private *id_priv;
1248 struct rdma_cm_id *id;
1249 union cma_ip_addr *src, *dst;
1254 id = rdma_create_id(listen_id->event_handler, listen_id->context,
1255 listen_id->ps, IB_QPT_UD);
1259 id_priv = container_of(id, struct rdma_id_private, id);
1260 if (cma_get_net_info(ib_event->private_data, listen_id->ps,
1261 &ip_ver, &port, &src, &dst))
1264 cma_save_net_info(&id->route.addr, &listen_id->route.addr,
1265 ip_ver, port, src, dst);
1267 if (!cma_any_addr((struct sockaddr *) &id->route.addr.src_addr)) {
1268 ret = cma_translate_addr(cma_src_addr(id_priv), &id->route.addr.dev_addr);
1273 id_priv->state = RDMA_CM_CONNECT;
1276 rdma_destroy_id(id);
1280 static void cma_set_req_event_data(struct rdma_cm_event *event,
1281 struct ib_cm_req_event_param *req_data,
1282 void *private_data, int offset)
1284 event->param.conn.private_data = private_data + offset;
1285 event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
1286 event->param.conn.responder_resources = req_data->responder_resources;
1287 event->param.conn.initiator_depth = req_data->initiator_depth;
1288 event->param.conn.flow_control = req_data->flow_control;
1289 event->param.conn.retry_count = req_data->retry_count;
1290 event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
1291 event->param.conn.srq = req_data->srq;
1292 event->param.conn.qp_num = req_data->remote_qpn;
1295 static int cma_check_req_qp_type(struct rdma_cm_id *id, struct ib_cm_event *ib_event)
1297 return (((ib_event->event == IB_CM_REQ_RECEIVED) &&
1298 (ib_event->param.req_rcvd.qp_type == id->qp_type)) ||
1299 ((ib_event->event == IB_CM_SIDR_REQ_RECEIVED) &&
1300 (id->qp_type == IB_QPT_UD)) ||
1304 static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1306 struct rdma_id_private *listen_id, *conn_id;
1307 struct rdma_cm_event event;
1310 listen_id = cm_id->context;
1311 if (!cma_check_req_qp_type(&listen_id->id, ib_event))
1314 if (cma_disable_callback(listen_id, RDMA_CM_LISTEN))
1315 return -ECONNABORTED;
1317 memset(&event, 0, sizeof event);
1318 offset = cma_user_data_offset(listen_id->id.ps);
1319 event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1320 if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) {
1321 conn_id = cma_new_udp_id(&listen_id->id, ib_event);
1322 event.param.ud.private_data = ib_event->private_data + offset;
1323 event.param.ud.private_data_len =
1324 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
1326 conn_id = cma_new_conn_id(&listen_id->id, ib_event);
1327 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
1328 ib_event->private_data, offset);
1335 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1336 ret = cma_acquire_dev(conn_id);
1340 conn_id->cm_id.ib = cm_id;
1341 cm_id->context = conn_id;
1342 cm_id->cm_handler = cma_ib_handler;
1345 * Protect against the user destroying conn_id from another thread
1346 * until we're done accessing it.
1348 atomic_inc(&conn_id->refcount);
1349 ret = conn_id->id.event_handler(&conn_id->id, &event);
1354 * Acquire mutex to prevent user executing rdma_destroy_id()
1355 * while we're accessing the cm_id.
1358 if (cma_comp(conn_id, RDMA_CM_CONNECT) && (conn_id->id.qp_type != IB_QPT_UD))
1359 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
1360 mutex_unlock(&lock);
1361 mutex_unlock(&conn_id->handler_mutex);
1362 mutex_unlock(&listen_id->handler_mutex);
1363 cma_deref_id(conn_id);
1367 cma_deref_id(conn_id);
1368 /* Destroy the CM ID by returning a non-zero value. */
1369 conn_id->cm_id.ib = NULL;
1371 cma_exch(conn_id, RDMA_CM_DESTROYING);
1372 mutex_unlock(&conn_id->handler_mutex);
1374 mutex_unlock(&listen_id->handler_mutex);
1376 rdma_destroy_id(&conn_id->id);
1380 static __be64 cma_get_service_id(enum rdma_port_space ps, struct sockaddr *addr)
1382 if (addr->sa_family == AF_IB)
1383 return ((struct sockaddr_ib *) addr)->sib_sid;
1385 return cpu_to_be64(((u64)ps << 16) + be16_to_cpu(cma_port(addr)));
1388 static void cma_set_compare_data(enum rdma_port_space ps, struct sockaddr *addr,
1389 struct ib_cm_compare_data *compare)
1391 struct cma_hdr *cma_data, *cma_mask;
1392 struct sdp_hh *sdp_data, *sdp_mask;
1394 struct in6_addr ip6_addr;
1396 memset(compare, 0, sizeof *compare);
1397 cma_data = (void *) compare->data;
1398 cma_mask = (void *) compare->mask;
1399 sdp_data = (void *) compare->data;
1400 sdp_mask = (void *) compare->mask;
1402 switch (addr->sa_family) {
1404 ip4_addr = ((struct sockaddr_in *) addr)->sin_addr.s_addr;
1405 if (ps == RDMA_PS_SDP) {
1406 sdp_set_ip_ver(sdp_data, 4);
1407 sdp_set_ip_ver(sdp_mask, 0xF);
1408 sdp_data->dst_addr.ip4.addr = ip4_addr;
1409 sdp_mask->dst_addr.ip4.addr = htonl(~0);
1411 cma_set_ip_ver(cma_data, 4);
1412 cma_set_ip_ver(cma_mask, 0xF);
1413 if (!cma_any_addr(addr)) {
1414 cma_data->dst_addr.ip4.addr = ip4_addr;
1415 cma_mask->dst_addr.ip4.addr = htonl(~0);
1420 ip6_addr = ((struct sockaddr_in6 *) addr)->sin6_addr;
1421 if (ps == RDMA_PS_SDP) {
1422 sdp_set_ip_ver(sdp_data, 6);
1423 sdp_set_ip_ver(sdp_mask, 0xF);
1424 sdp_data->dst_addr.ip6 = ip6_addr;
1425 memset(&sdp_mask->dst_addr.ip6, 0xFF,
1426 sizeof sdp_mask->dst_addr.ip6);
1428 cma_set_ip_ver(cma_data, 6);
1429 cma_set_ip_ver(cma_mask, 0xF);
1430 if (!cma_any_addr(addr)) {
1431 cma_data->dst_addr.ip6 = ip6_addr;
1432 memset(&cma_mask->dst_addr.ip6, 0xFF,
1433 sizeof cma_mask->dst_addr.ip6);
1442 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
1444 struct rdma_id_private *id_priv = iw_id->context;
1445 struct rdma_cm_event event;
1446 struct sockaddr_in *sin;
1449 if (cma_disable_callback(id_priv, RDMA_CM_CONNECT))
1452 memset(&event, 0, sizeof event);
1453 switch (iw_event->event) {
1454 case IW_CM_EVENT_CLOSE:
1455 event.event = RDMA_CM_EVENT_DISCONNECTED;
1457 case IW_CM_EVENT_CONNECT_REPLY:
1458 sin = (struct sockaddr_in *) cma_src_addr(id_priv);
1459 *sin = iw_event->local_addr;
1460 sin = (struct sockaddr_in *) cma_dst_addr(id_priv);
1461 *sin = iw_event->remote_addr;
1462 switch (iw_event->status) {
1464 event.event = RDMA_CM_EVENT_ESTABLISHED;
1465 event.param.conn.initiator_depth = iw_event->ird;
1466 event.param.conn.responder_resources = iw_event->ord;
1470 event.event = RDMA_CM_EVENT_REJECTED;
1473 event.event = RDMA_CM_EVENT_UNREACHABLE;
1476 event.event = RDMA_CM_EVENT_CONNECT_ERROR;
1480 case IW_CM_EVENT_ESTABLISHED:
1481 event.event = RDMA_CM_EVENT_ESTABLISHED;
1482 event.param.conn.initiator_depth = iw_event->ird;
1483 event.param.conn.responder_resources = iw_event->ord;
1489 event.status = iw_event->status;
1490 event.param.conn.private_data = iw_event->private_data;
1491 event.param.conn.private_data_len = iw_event->private_data_len;
1492 ret = id_priv->id.event_handler(&id_priv->id, &event);
1494 /* Destroy the CM ID by returning a non-zero value. */
1495 id_priv->cm_id.iw = NULL;
1496 cma_exch(id_priv, RDMA_CM_DESTROYING);
1497 mutex_unlock(&id_priv->handler_mutex);
1498 rdma_destroy_id(&id_priv->id);
1502 mutex_unlock(&id_priv->handler_mutex);
1506 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
1507 struct iw_cm_event *iw_event)
1509 struct rdma_cm_id *new_cm_id;
1510 struct rdma_id_private *listen_id, *conn_id;
1511 struct sockaddr_in *sin;
1512 struct net_device *dev = NULL;
1513 struct rdma_cm_event event;
1515 struct ib_device_attr attr;
1517 listen_id = cm_id->context;
1518 if (cma_disable_callback(listen_id, RDMA_CM_LISTEN))
1519 return -ECONNABORTED;
1521 /* Create a new RDMA id for the new IW CM ID */
1522 new_cm_id = rdma_create_id(listen_id->id.event_handler,
1523 listen_id->id.context,
1524 RDMA_PS_TCP, IB_QPT_RC);
1525 if (IS_ERR(new_cm_id)) {
1529 conn_id = container_of(new_cm_id, struct rdma_id_private, id);
1530 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1531 conn_id->state = RDMA_CM_CONNECT;
1533 dev = ip_dev_find(&init_net, iw_event->local_addr.sin_addr.s_addr);
1535 ret = -EADDRNOTAVAIL;
1536 mutex_unlock(&conn_id->handler_mutex);
1537 rdma_destroy_id(new_cm_id);
1540 ret = rdma_copy_addr(&conn_id->id.route.addr.dev_addr, dev, NULL);
1542 mutex_unlock(&conn_id->handler_mutex);
1543 rdma_destroy_id(new_cm_id);
1547 ret = cma_acquire_dev(conn_id);
1549 mutex_unlock(&conn_id->handler_mutex);
1550 rdma_destroy_id(new_cm_id);
1554 conn_id->cm_id.iw = cm_id;
1555 cm_id->context = conn_id;
1556 cm_id->cm_handler = cma_iw_handler;
1558 sin = (struct sockaddr_in *) cma_src_addr(conn_id);
1559 *sin = iw_event->local_addr;
1560 sin = (struct sockaddr_in *) cma_dst_addr(conn_id);
1561 *sin = iw_event->remote_addr;
1563 ret = ib_query_device(conn_id->id.device, &attr);
1565 mutex_unlock(&conn_id->handler_mutex);
1566 rdma_destroy_id(new_cm_id);
1570 memset(&event, 0, sizeof event);
1571 event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1572 event.param.conn.private_data = iw_event->private_data;
1573 event.param.conn.private_data_len = iw_event->private_data_len;
1574 event.param.conn.initiator_depth = iw_event->ird;
1575 event.param.conn.responder_resources = iw_event->ord;
1578 * Protect against the user destroying conn_id from another thread
1579 * until we're done accessing it.
1581 atomic_inc(&conn_id->refcount);
1582 ret = conn_id->id.event_handler(&conn_id->id, &event);
1584 /* User wants to destroy the CM ID */
1585 conn_id->cm_id.iw = NULL;
1586 cma_exch(conn_id, RDMA_CM_DESTROYING);
1587 mutex_unlock(&conn_id->handler_mutex);
1588 cma_deref_id(conn_id);
1589 rdma_destroy_id(&conn_id->id);
1593 mutex_unlock(&conn_id->handler_mutex);
1594 cma_deref_id(conn_id);
1599 mutex_unlock(&listen_id->handler_mutex);
1603 static int cma_ib_listen(struct rdma_id_private *id_priv)
1605 struct ib_cm_compare_data compare_data;
1606 struct sockaddr *addr;
1607 struct ib_cm_id *id;
1611 id = ib_create_cm_id(id_priv->id.device, cma_req_handler, id_priv);
1615 id_priv->cm_id.ib = id;
1617 addr = cma_src_addr(id_priv);
1618 svc_id = cma_get_service_id(id_priv->id.ps, addr);
1619 if (cma_any_addr(addr) && !id_priv->afonly)
1620 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, NULL);
1622 cma_set_compare_data(id_priv->id.ps, addr, &compare_data);
1623 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, &compare_data);
1627 ib_destroy_cm_id(id_priv->cm_id.ib);
1628 id_priv->cm_id.ib = NULL;
1634 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
1637 struct sockaddr_in *sin;
1638 struct iw_cm_id *id;
1640 id = iw_create_cm_id(id_priv->id.device,
1641 iw_conn_req_handler,
1646 id_priv->cm_id.iw = id;
1648 sin = (struct sockaddr_in *) cma_src_addr(id_priv);
1649 id_priv->cm_id.iw->local_addr = *sin;
1651 ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
1654 iw_destroy_cm_id(id_priv->cm_id.iw);
1655 id_priv->cm_id.iw = NULL;
1661 static int cma_listen_handler(struct rdma_cm_id *id,
1662 struct rdma_cm_event *event)
1664 struct rdma_id_private *id_priv = id->context;
1666 id->context = id_priv->id.context;
1667 id->event_handler = id_priv->id.event_handler;
1668 return id_priv->id.event_handler(id, event);
1671 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
1672 struct cma_device *cma_dev)
1674 struct rdma_id_private *dev_id_priv;
1675 struct rdma_cm_id *id;
1678 id = rdma_create_id(cma_listen_handler, id_priv, id_priv->id.ps,
1679 id_priv->id.qp_type);
1683 dev_id_priv = container_of(id, struct rdma_id_private, id);
1685 dev_id_priv->state = RDMA_CM_ADDR_BOUND;
1686 memcpy(cma_src_addr(dev_id_priv), cma_src_addr(id_priv),
1687 rdma_addr_size(cma_src_addr(id_priv)));
1689 cma_attach_to_dev(dev_id_priv, cma_dev);
1690 list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
1691 atomic_inc(&id_priv->refcount);
1692 dev_id_priv->internal_id = 1;
1693 dev_id_priv->afonly = id_priv->afonly;
1695 ret = rdma_listen(id, id_priv->backlog);
1697 printk(KERN_WARNING "RDMA CMA: cma_listen_on_dev, error %d, "
1698 "listening on device %s\n", ret, cma_dev->device->name);
1701 static void cma_listen_on_all(struct rdma_id_private *id_priv)
1703 struct cma_device *cma_dev;
1706 list_add_tail(&id_priv->list, &listen_any_list);
1707 list_for_each_entry(cma_dev, &dev_list, list)
1708 cma_listen_on_dev(id_priv, cma_dev);
1709 mutex_unlock(&lock);
1712 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
1714 struct rdma_id_private *id_priv;
1716 id_priv = container_of(id, struct rdma_id_private, id);
1717 id_priv->tos = (u8) tos;
1719 EXPORT_SYMBOL(rdma_set_service_type);
1721 static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec,
1724 struct cma_work *work = context;
1725 struct rdma_route *route;
1727 route = &work->id->id.route;
1730 route->num_paths = 1;
1731 *route->path_rec = *path_rec;
1733 work->old_state = RDMA_CM_ROUTE_QUERY;
1734 work->new_state = RDMA_CM_ADDR_RESOLVED;
1735 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
1736 work->event.status = status;
1739 queue_work(cma_wq, &work->work);
1742 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms,
1743 struct cma_work *work)
1745 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
1746 struct ib_sa_path_rec path_rec;
1747 ib_sa_comp_mask comp_mask;
1748 struct sockaddr_in6 *sin6;
1749 struct sockaddr_ib *sib;
1751 memset(&path_rec, 0, sizeof path_rec);
1752 rdma_addr_get_sgid(dev_addr, &path_rec.sgid);
1753 rdma_addr_get_dgid(dev_addr, &path_rec.dgid);
1754 path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
1755 path_rec.numb_path = 1;
1756 path_rec.reversible = 1;
1757 path_rec.service_id = cma_get_service_id(id_priv->id.ps, cma_dst_addr(id_priv));
1759 comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
1760 IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
1761 IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
1763 switch (cma_family(id_priv)) {
1765 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
1766 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
1769 sin6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
1770 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
1771 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
1774 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
1775 path_rec.traffic_class = (u8) (be32_to_cpu(sib->sib_flowinfo) >> 20);
1776 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
1780 id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
1781 id_priv->id.port_num, &path_rec,
1782 comp_mask, timeout_ms,
1783 GFP_KERNEL, cma_query_handler,
1784 work, &id_priv->query);
1786 return (id_priv->query_id < 0) ? id_priv->query_id : 0;
1789 static void cma_work_handler(struct work_struct *_work)
1791 struct cma_work *work = container_of(_work, struct cma_work, work);
1792 struct rdma_id_private *id_priv = work->id;
1795 mutex_lock(&id_priv->handler_mutex);
1796 if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
1799 if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1800 cma_exch(id_priv, RDMA_CM_DESTROYING);
1804 mutex_unlock(&id_priv->handler_mutex);
1805 cma_deref_id(id_priv);
1807 rdma_destroy_id(&id_priv->id);
1811 static void cma_ndev_work_handler(struct work_struct *_work)
1813 struct cma_ndev_work *work = container_of(_work, struct cma_ndev_work, work);
1814 struct rdma_id_private *id_priv = work->id;
1817 mutex_lock(&id_priv->handler_mutex);
1818 if (id_priv->state == RDMA_CM_DESTROYING ||
1819 id_priv->state == RDMA_CM_DEVICE_REMOVAL)
1822 if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1823 cma_exch(id_priv, RDMA_CM_DESTROYING);
1828 mutex_unlock(&id_priv->handler_mutex);
1829 cma_deref_id(id_priv);
1831 rdma_destroy_id(&id_priv->id);
1835 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms)
1837 struct rdma_route *route = &id_priv->id.route;
1838 struct cma_work *work;
1841 work = kzalloc(sizeof *work, GFP_KERNEL);
1846 INIT_WORK(&work->work, cma_work_handler);
1847 work->old_state = RDMA_CM_ROUTE_QUERY;
1848 work->new_state = RDMA_CM_ROUTE_RESOLVED;
1849 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1851 route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
1852 if (!route->path_rec) {
1857 ret = cma_query_ib_route(id_priv, timeout_ms, work);
1863 kfree(route->path_rec);
1864 route->path_rec = NULL;
1870 int rdma_set_ib_paths(struct rdma_cm_id *id,
1871 struct ib_sa_path_rec *path_rec, int num_paths)
1873 struct rdma_id_private *id_priv;
1876 id_priv = container_of(id, struct rdma_id_private, id);
1877 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
1878 RDMA_CM_ROUTE_RESOLVED))
1881 id->route.path_rec = kmemdup(path_rec, sizeof *path_rec * num_paths,
1883 if (!id->route.path_rec) {
1888 id->route.num_paths = num_paths;
1891 cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED);
1894 EXPORT_SYMBOL(rdma_set_ib_paths);
1896 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms)
1898 struct cma_work *work;
1900 work = kzalloc(sizeof *work, GFP_KERNEL);
1905 INIT_WORK(&work->work, cma_work_handler);
1906 work->old_state = RDMA_CM_ROUTE_QUERY;
1907 work->new_state = RDMA_CM_ROUTE_RESOLVED;
1908 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1909 queue_work(cma_wq, &work->work);
1913 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
1915 struct rdma_route *route = &id_priv->id.route;
1916 struct rdma_addr *addr = &route->addr;
1917 struct cma_work *work;
1919 struct net_device *ndev = NULL;
1922 work = kzalloc(sizeof *work, GFP_KERNEL);
1927 INIT_WORK(&work->work, cma_work_handler);
1929 route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
1930 if (!route->path_rec) {
1935 route->num_paths = 1;
1937 if (addr->dev_addr.bound_dev_if)
1938 ndev = dev_get_by_index(&init_net, addr->dev_addr.bound_dev_if);
1944 vid = rdma_vlan_dev_vlan_id(ndev);
1946 iboe_mac_vlan_to_ll(&route->path_rec->sgid, addr->dev_addr.src_dev_addr, vid);
1947 iboe_mac_vlan_to_ll(&route->path_rec->dgid, addr->dev_addr.dst_dev_addr, vid);
1949 route->path_rec->hop_limit = 1;
1950 route->path_rec->reversible = 1;
1951 route->path_rec->pkey = cpu_to_be16(0xffff);
1952 route->path_rec->mtu_selector = IB_SA_EQ;
1953 route->path_rec->sl = netdev_get_prio_tc_map(
1954 ndev->priv_flags & IFF_802_1Q_VLAN ?
1955 vlan_dev_real_dev(ndev) : ndev,
1956 rt_tos2priority(id_priv->tos));
1958 route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
1959 route->path_rec->rate_selector = IB_SA_EQ;
1960 route->path_rec->rate = iboe_get_rate(ndev);
1962 route->path_rec->packet_life_time_selector = IB_SA_EQ;
1963 route->path_rec->packet_life_time = CMA_IBOE_PACKET_LIFETIME;
1964 if (!route->path_rec->mtu) {
1969 work->old_state = RDMA_CM_ROUTE_QUERY;
1970 work->new_state = RDMA_CM_ROUTE_RESOLVED;
1971 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1972 work->event.status = 0;
1974 queue_work(cma_wq, &work->work);
1979 kfree(route->path_rec);
1980 route->path_rec = NULL;
1986 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
1988 struct rdma_id_private *id_priv;
1991 id_priv = container_of(id, struct rdma_id_private, id);
1992 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY))
1995 atomic_inc(&id_priv->refcount);
1996 switch (rdma_node_get_transport(id->device->node_type)) {
1997 case RDMA_TRANSPORT_IB:
1998 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
1999 case IB_LINK_LAYER_INFINIBAND:
2000 ret = cma_resolve_ib_route(id_priv, timeout_ms);
2002 case IB_LINK_LAYER_ETHERNET:
2003 ret = cma_resolve_iboe_route(id_priv);
2009 case RDMA_TRANSPORT_IWARP:
2010 ret = cma_resolve_iw_route(id_priv, timeout_ms);
2021 cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED);
2022 cma_deref_id(id_priv);
2025 EXPORT_SYMBOL(rdma_resolve_route);
2027 static void cma_set_loopback(struct sockaddr *addr)
2029 switch (addr->sa_family) {
2031 ((struct sockaddr_in *) addr)->sin_addr.s_addr = htonl(INADDR_LOOPBACK);
2034 ipv6_addr_set(&((struct sockaddr_in6 *) addr)->sin6_addr,
2038 ib_addr_set(&((struct sockaddr_ib *) addr)->sib_addr,
2044 static int cma_bind_loopback(struct rdma_id_private *id_priv)
2046 struct cma_device *cma_dev, *cur_dev;
2047 struct ib_port_attr port_attr;
2055 list_for_each_entry(cur_dev, &dev_list, list) {
2056 if (cma_family(id_priv) == AF_IB &&
2057 rdma_node_get_transport(cur_dev->device->node_type) != RDMA_TRANSPORT_IB)
2063 for (p = 1; p <= cur_dev->device->phys_port_cnt; ++p) {
2064 if (!ib_query_port(cur_dev->device, p, &port_attr) &&
2065 port_attr.state == IB_PORT_ACTIVE) {
2080 ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid);
2084 ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
2088 id_priv->id.route.addr.dev_addr.dev_type =
2089 (rdma_port_get_link_layer(cma_dev->device, p) == IB_LINK_LAYER_INFINIBAND) ?
2090 ARPHRD_INFINIBAND : ARPHRD_ETHER;
2092 rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2093 ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
2094 id_priv->id.port_num = p;
2095 cma_attach_to_dev(id_priv, cma_dev);
2096 cma_set_loopback(cma_src_addr(id_priv));
2098 mutex_unlock(&lock);
2102 static void addr_handler(int status, struct sockaddr *src_addr,
2103 struct rdma_dev_addr *dev_addr, void *context)
2105 struct rdma_id_private *id_priv = context;
2106 struct rdma_cm_event event;
2108 memset(&event, 0, sizeof event);
2109 mutex_lock(&id_priv->handler_mutex);
2110 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY,
2111 RDMA_CM_ADDR_RESOLVED))
2114 if (!status && !id_priv->cma_dev)
2115 status = cma_acquire_dev(id_priv);
2118 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
2119 RDMA_CM_ADDR_BOUND))
2121 event.event = RDMA_CM_EVENT_ADDR_ERROR;
2122 event.status = status;
2124 memcpy(cma_src_addr(id_priv), src_addr, rdma_addr_size(src_addr));
2125 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2128 if (id_priv->id.event_handler(&id_priv->id, &event)) {
2129 cma_exch(id_priv, RDMA_CM_DESTROYING);
2130 mutex_unlock(&id_priv->handler_mutex);
2131 cma_deref_id(id_priv);
2132 rdma_destroy_id(&id_priv->id);
2136 mutex_unlock(&id_priv->handler_mutex);
2137 cma_deref_id(id_priv);
2140 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
2142 struct cma_work *work;
2146 work = kzalloc(sizeof *work, GFP_KERNEL);
2150 if (!id_priv->cma_dev) {
2151 ret = cma_bind_loopback(id_priv);
2156 rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2157 rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
2160 INIT_WORK(&work->work, cma_work_handler);
2161 work->old_state = RDMA_CM_ADDR_QUERY;
2162 work->new_state = RDMA_CM_ADDR_RESOLVED;
2163 work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2164 queue_work(cma_wq, &work->work);
2171 static int cma_resolve_ib_addr(struct rdma_id_private *id_priv)
2173 struct cma_work *work;
2176 work = kzalloc(sizeof *work, GFP_KERNEL);
2180 if (!id_priv->cma_dev) {
2181 ret = cma_resolve_ib_dev(id_priv);
2186 rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, (union ib_gid *)
2187 &(((struct sockaddr_ib *) &id_priv->id.route.addr.dst_addr)->sib_addr));
2190 INIT_WORK(&work->work, cma_work_handler);
2191 work->old_state = RDMA_CM_ADDR_QUERY;
2192 work->new_state = RDMA_CM_ADDR_RESOLVED;
2193 work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2194 queue_work(cma_wq, &work->work);
2201 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2202 struct sockaddr *dst_addr)
2204 if (!src_addr || !src_addr->sa_family) {
2205 src_addr = (struct sockaddr *) &id->route.addr.src_addr;
2206 src_addr->sa_family = dst_addr->sa_family;
2207 if (dst_addr->sa_family == AF_INET6) {
2208 ((struct sockaddr_in6 *) src_addr)->sin6_scope_id =
2209 ((struct sockaddr_in6 *) dst_addr)->sin6_scope_id;
2210 } else if (dst_addr->sa_family == AF_IB) {
2211 ((struct sockaddr_ib *) src_addr)->sib_pkey =
2212 ((struct sockaddr_ib *) dst_addr)->sib_pkey;
2215 return rdma_bind_addr(id, src_addr);
2218 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2219 struct sockaddr *dst_addr, int timeout_ms)
2221 struct rdma_id_private *id_priv;
2224 id_priv = container_of(id, struct rdma_id_private, id);
2225 if (id_priv->state == RDMA_CM_IDLE) {
2226 ret = cma_bind_addr(id, src_addr, dst_addr);
2231 if (cma_family(id_priv) != dst_addr->sa_family)
2234 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY))
2237 atomic_inc(&id_priv->refcount);
2238 memcpy(cma_dst_addr(id_priv), dst_addr, rdma_addr_size(dst_addr));
2239 if (cma_any_addr(dst_addr)) {
2240 ret = cma_resolve_loopback(id_priv);
2242 if (dst_addr->sa_family == AF_IB) {
2243 ret = cma_resolve_ib_addr(id_priv);
2245 ret = rdma_resolve_ip(&addr_client, cma_src_addr(id_priv),
2246 dst_addr, &id->route.addr.dev_addr,
2247 timeout_ms, addr_handler, id_priv);
2255 cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
2256 cma_deref_id(id_priv);
2259 EXPORT_SYMBOL(rdma_resolve_addr);
2261 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse)
2263 struct rdma_id_private *id_priv;
2264 unsigned long flags;
2267 id_priv = container_of(id, struct rdma_id_private, id);
2268 spin_lock_irqsave(&id_priv->lock, flags);
2269 if (reuse || id_priv->state == RDMA_CM_IDLE) {
2270 id_priv->reuseaddr = reuse;
2275 spin_unlock_irqrestore(&id_priv->lock, flags);
2278 EXPORT_SYMBOL(rdma_set_reuseaddr);
2280 int rdma_set_afonly(struct rdma_cm_id *id, int afonly)
2282 struct rdma_id_private *id_priv;
2283 unsigned long flags;
2286 id_priv = container_of(id, struct rdma_id_private, id);
2287 spin_lock_irqsave(&id_priv->lock, flags);
2288 if (id_priv->state == RDMA_CM_IDLE || id_priv->state == RDMA_CM_ADDR_BOUND) {
2289 id_priv->options |= (1 << CMA_OPTION_AFONLY);
2290 id_priv->afonly = afonly;
2295 spin_unlock_irqrestore(&id_priv->lock, flags);
2298 EXPORT_SYMBOL(rdma_set_afonly);
2300 static void cma_bind_port(struct rdma_bind_list *bind_list,
2301 struct rdma_id_private *id_priv)
2303 struct sockaddr *addr;
2304 struct sockaddr_ib *sib;
2308 addr = cma_src_addr(id_priv);
2309 port = htons(bind_list->port);
2311 switch (addr->sa_family) {
2313 ((struct sockaddr_in *) addr)->sin_port = port;
2316 ((struct sockaddr_in6 *) addr)->sin6_port = port;
2319 sib = (struct sockaddr_ib *) addr;
2320 sid = be64_to_cpu(sib->sib_sid);
2321 mask = be64_to_cpu(sib->sib_sid_mask);
2322 sib->sib_sid = cpu_to_be64((sid & mask) | (u64) ntohs(port));
2323 sib->sib_sid_mask = cpu_to_be64(~0ULL);
2326 id_priv->bind_list = bind_list;
2327 hlist_add_head(&id_priv->node, &bind_list->owners);
2330 static int cma_alloc_port(struct idr *ps, struct rdma_id_private *id_priv,
2331 unsigned short snum)
2333 struct rdma_bind_list *bind_list;
2336 bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
2340 ret = idr_alloc(ps, bind_list, snum, snum + 1, GFP_KERNEL);
2345 bind_list->port = (unsigned short)ret;
2346 cma_bind_port(bind_list, id_priv);
2350 return ret == -ENOSPC ? -EADDRNOTAVAIL : ret;
2353 static int cma_alloc_any_port(struct idr *ps, struct rdma_id_private *id_priv)
2355 static unsigned int last_used_port;
2356 int low, high, remaining;
2359 inet_get_local_port_range(&low, &high);
2360 remaining = (high - low) + 1;
2361 rover = net_random() % remaining + low;
2363 if (last_used_port != rover &&
2364 !idr_find(ps, (unsigned short) rover)) {
2365 int ret = cma_alloc_port(ps, id_priv, rover);
2367 * Remember previously used port number in order to avoid
2368 * re-using same port immediately after it is closed.
2371 last_used_port = rover;
2372 if (ret != -EADDRNOTAVAIL)
2377 if ((rover < low) || (rover > high))
2381 return -EADDRNOTAVAIL;
2385 * Check that the requested port is available. This is called when trying to
2386 * bind to a specific port, or when trying to listen on a bound port. In
2387 * the latter case, the provided id_priv may already be on the bind_list, but
2388 * we still need to check that it's okay to start listening.
2390 static int cma_check_port(struct rdma_bind_list *bind_list,
2391 struct rdma_id_private *id_priv, uint8_t reuseaddr)
2393 struct rdma_id_private *cur_id;
2394 struct sockaddr *addr, *cur_addr;
2396 addr = cma_src_addr(id_priv);
2397 hlist_for_each_entry(cur_id, &bind_list->owners, node) {
2398 if (id_priv == cur_id)
2401 if ((cur_id->state != RDMA_CM_LISTEN) && reuseaddr &&
2405 cur_addr = cma_src_addr(cur_id);
2406 if (id_priv->afonly && cur_id->afonly &&
2407 (addr->sa_family != cur_addr->sa_family))
2410 if (cma_any_addr(addr) || cma_any_addr(cur_addr))
2411 return -EADDRNOTAVAIL;
2413 if (!cma_addr_cmp(addr, cur_addr))
2419 static int cma_use_port(struct idr *ps, struct rdma_id_private *id_priv)
2421 struct rdma_bind_list *bind_list;
2422 unsigned short snum;
2425 snum = ntohs(cma_port(cma_src_addr(id_priv)));
2426 if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
2429 bind_list = idr_find(ps, snum);
2431 ret = cma_alloc_port(ps, id_priv, snum);
2433 ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr);
2435 cma_bind_port(bind_list, id_priv);
2440 static int cma_bind_listen(struct rdma_id_private *id_priv)
2442 struct rdma_bind_list *bind_list = id_priv->bind_list;
2446 if (bind_list->owners.first->next)
2447 ret = cma_check_port(bind_list, id_priv, 0);
2448 mutex_unlock(&lock);
2452 static struct idr *cma_select_inet_ps(struct rdma_id_private *id_priv)
2454 switch (id_priv->id.ps) {
2470 static struct idr *cma_select_ib_ps(struct rdma_id_private *id_priv)
2472 struct idr *ps = NULL;
2473 struct sockaddr_ib *sib;
2474 u64 sid_ps, mask, sid;
2476 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
2477 mask = be64_to_cpu(sib->sib_sid_mask) & RDMA_IB_IP_PS_MASK;
2478 sid = be64_to_cpu(sib->sib_sid) & mask;
2480 if ((id_priv->id.ps == RDMA_PS_IB) && (sid == (RDMA_IB_IP_PS_IB & mask))) {
2481 sid_ps = RDMA_IB_IP_PS_IB;
2483 } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_TCP)) &&
2484 (sid == (RDMA_IB_IP_PS_TCP & mask))) {
2485 sid_ps = RDMA_IB_IP_PS_TCP;
2487 } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_UDP)) &&
2488 (sid == (RDMA_IB_IP_PS_UDP & mask))) {
2489 sid_ps = RDMA_IB_IP_PS_UDP;
2494 sib->sib_sid = cpu_to_be64(sid_ps | ntohs(cma_port((struct sockaddr *) sib)));
2495 sib->sib_sid_mask = cpu_to_be64(RDMA_IB_IP_PS_MASK |
2496 be64_to_cpu(sib->sib_sid_mask));
2501 static int cma_get_port(struct rdma_id_private *id_priv)
2506 if (cma_family(id_priv) != AF_IB)
2507 ps = cma_select_inet_ps(id_priv);
2509 ps = cma_select_ib_ps(id_priv);
2511 return -EPROTONOSUPPORT;
2514 if (cma_any_port(cma_src_addr(id_priv)))
2515 ret = cma_alloc_any_port(ps, id_priv);
2517 ret = cma_use_port(ps, id_priv);
2518 mutex_unlock(&lock);
2523 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
2524 struct sockaddr *addr)
2526 #if IS_ENABLED(CONFIG_IPV6)
2527 struct sockaddr_in6 *sin6;
2529 if (addr->sa_family != AF_INET6)
2532 sin6 = (struct sockaddr_in6 *) addr;
2533 if ((ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL) &&
2534 !sin6->sin6_scope_id)
2537 dev_addr->bound_dev_if = sin6->sin6_scope_id;
2542 int rdma_listen(struct rdma_cm_id *id, int backlog)
2544 struct rdma_id_private *id_priv;
2547 id_priv = container_of(id, struct rdma_id_private, id);
2548 if (id_priv->state == RDMA_CM_IDLE) {
2549 id->route.addr.src_addr.ss_family = AF_INET;
2550 ret = rdma_bind_addr(id, cma_src_addr(id_priv));
2555 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN))
2558 if (id_priv->reuseaddr) {
2559 ret = cma_bind_listen(id_priv);
2564 id_priv->backlog = backlog;
2566 switch (rdma_node_get_transport(id->device->node_type)) {
2567 case RDMA_TRANSPORT_IB:
2568 ret = cma_ib_listen(id_priv);
2572 case RDMA_TRANSPORT_IWARP:
2573 ret = cma_iw_listen(id_priv, backlog);
2582 cma_listen_on_all(id_priv);
2586 id_priv->backlog = 0;
2587 cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND);
2590 EXPORT_SYMBOL(rdma_listen);
2592 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
2594 struct rdma_id_private *id_priv;
2597 if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6 &&
2598 addr->sa_family != AF_IB)
2599 return -EAFNOSUPPORT;
2601 id_priv = container_of(id, struct rdma_id_private, id);
2602 if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND))
2605 ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
2609 if (!cma_any_addr(addr)) {
2610 ret = cma_translate_addr(addr, &id->route.addr.dev_addr);
2614 ret = cma_acquire_dev(id_priv);
2619 memcpy(cma_src_addr(id_priv), addr, rdma_addr_size(addr));
2620 if (!(id_priv->options & (1 << CMA_OPTION_AFONLY))) {
2621 if (addr->sa_family == AF_INET)
2622 id_priv->afonly = 1;
2623 #if IS_ENABLED(CONFIG_IPV6)
2624 else if (addr->sa_family == AF_INET6)
2625 id_priv->afonly = init_net.ipv6.sysctl.bindv6only;
2628 ret = cma_get_port(id_priv);
2634 if (id_priv->cma_dev)
2635 cma_release_dev(id_priv);
2637 cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE);
2640 EXPORT_SYMBOL(rdma_bind_addr);
2642 static int cma_format_hdr(void *hdr, struct rdma_id_private *id_priv)
2644 struct cma_hdr *cma_hdr;
2645 struct sdp_hh *sdp_hdr;
2647 if (cma_family(id_priv) == AF_INET) {
2648 struct sockaddr_in *src4, *dst4;
2650 src4 = (struct sockaddr_in *) cma_src_addr(id_priv);
2651 dst4 = (struct sockaddr_in *) cma_dst_addr(id_priv);
2653 switch (id_priv->id.ps) {
2656 if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION)
2658 sdp_set_ip_ver(sdp_hdr, 4);
2659 sdp_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2660 sdp_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2661 sdp_hdr->port = src4->sin_port;
2665 cma_hdr->cma_version = CMA_VERSION;
2666 cma_set_ip_ver(cma_hdr, 4);
2667 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2668 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2669 cma_hdr->port = src4->sin_port;
2673 struct sockaddr_in6 *src6, *dst6;
2675 src6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
2676 dst6 = (struct sockaddr_in6 *) cma_dst_addr(id_priv);
2678 switch (id_priv->id.ps) {
2681 if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION)
2683 sdp_set_ip_ver(sdp_hdr, 6);
2684 sdp_hdr->src_addr.ip6 = src6->sin6_addr;
2685 sdp_hdr->dst_addr.ip6 = dst6->sin6_addr;
2686 sdp_hdr->port = src6->sin6_port;
2690 cma_hdr->cma_version = CMA_VERSION;
2691 cma_set_ip_ver(cma_hdr, 6);
2692 cma_hdr->src_addr.ip6 = src6->sin6_addr;
2693 cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
2694 cma_hdr->port = src6->sin6_port;
2701 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
2702 struct ib_cm_event *ib_event)
2704 struct rdma_id_private *id_priv = cm_id->context;
2705 struct rdma_cm_event event;
2706 struct ib_cm_sidr_rep_event_param *rep = &ib_event->param.sidr_rep_rcvd;
2709 if (cma_disable_callback(id_priv, RDMA_CM_CONNECT))
2712 memset(&event, 0, sizeof event);
2713 switch (ib_event->event) {
2714 case IB_CM_SIDR_REQ_ERROR:
2715 event.event = RDMA_CM_EVENT_UNREACHABLE;
2716 event.status = -ETIMEDOUT;
2718 case IB_CM_SIDR_REP_RECEIVED:
2719 event.param.ud.private_data = ib_event->private_data;
2720 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
2721 if (rep->status != IB_SIDR_SUCCESS) {
2722 event.event = RDMA_CM_EVENT_UNREACHABLE;
2723 event.status = ib_event->param.sidr_rep_rcvd.status;
2726 ret = cma_set_qkey(id_priv);
2728 event.event = RDMA_CM_EVENT_ADDR_ERROR;
2729 event.status = -EINVAL;
2732 if (id_priv->qkey != rep->qkey) {
2733 event.event = RDMA_CM_EVENT_UNREACHABLE;
2734 event.status = -EINVAL;
2737 ib_init_ah_from_path(id_priv->id.device, id_priv->id.port_num,
2738 id_priv->id.route.path_rec,
2739 &event.param.ud.ah_attr);
2740 event.param.ud.qp_num = rep->qpn;
2741 event.param.ud.qkey = rep->qkey;
2742 event.event = RDMA_CM_EVENT_ESTABLISHED;
2746 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
2751 ret = id_priv->id.event_handler(&id_priv->id, &event);
2753 /* Destroy the CM ID by returning a non-zero value. */
2754 id_priv->cm_id.ib = NULL;
2755 cma_exch(id_priv, RDMA_CM_DESTROYING);
2756 mutex_unlock(&id_priv->handler_mutex);
2757 rdma_destroy_id(&id_priv->id);
2761 mutex_unlock(&id_priv->handler_mutex);
2765 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
2766 struct rdma_conn_param *conn_param)
2768 struct ib_cm_sidr_req_param req;
2769 struct ib_cm_id *id;
2772 req.private_data_len = sizeof(struct cma_hdr) +
2773 conn_param->private_data_len;
2774 if (req.private_data_len < conn_param->private_data_len)
2777 req.private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2778 if (!req.private_data)
2781 if (conn_param->private_data && conn_param->private_data_len)
2782 memcpy((void *) req.private_data + sizeof(struct cma_hdr),
2783 conn_param->private_data, conn_param->private_data_len);
2785 ret = cma_format_hdr((void *) req.private_data, id_priv);
2789 id = ib_create_cm_id(id_priv->id.device, cma_sidr_rep_handler,
2795 id_priv->cm_id.ib = id;
2797 req.path = id_priv->id.route.path_rec;
2798 req.service_id = cma_get_service_id(id_priv->id.ps, cma_dst_addr(id_priv));
2799 req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
2800 req.max_cm_retries = CMA_MAX_CM_RETRIES;
2802 ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
2804 ib_destroy_cm_id(id_priv->cm_id.ib);
2805 id_priv->cm_id.ib = NULL;
2808 kfree(req.private_data);
2812 static int cma_connect_ib(struct rdma_id_private *id_priv,
2813 struct rdma_conn_param *conn_param)
2815 struct ib_cm_req_param req;
2816 struct rdma_route *route;
2818 struct ib_cm_id *id;
2821 memset(&req, 0, sizeof req);
2822 offset = cma_user_data_offset(id_priv->id.ps);
2823 req.private_data_len = offset + conn_param->private_data_len;
2824 if (req.private_data_len < conn_param->private_data_len)
2827 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2831 if (conn_param->private_data && conn_param->private_data_len)
2832 memcpy(private_data + offset, conn_param->private_data,
2833 conn_param->private_data_len);
2835 id = ib_create_cm_id(id_priv->id.device, cma_ib_handler, id_priv);
2840 id_priv->cm_id.ib = id;
2842 route = &id_priv->id.route;
2843 ret = cma_format_hdr(private_data, id_priv);
2846 req.private_data = private_data;
2848 req.primary_path = &route->path_rec[0];
2849 if (route->num_paths == 2)
2850 req.alternate_path = &route->path_rec[1];
2852 req.service_id = cma_get_service_id(id_priv->id.ps, cma_dst_addr(id_priv));
2853 req.qp_num = id_priv->qp_num;
2854 req.qp_type = id_priv->id.qp_type;
2855 req.starting_psn = id_priv->seq_num;
2856 req.responder_resources = conn_param->responder_resources;
2857 req.initiator_depth = conn_param->initiator_depth;
2858 req.flow_control = conn_param->flow_control;
2859 req.retry_count = min_t(u8, 7, conn_param->retry_count);
2860 req.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
2861 req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2862 req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2863 req.max_cm_retries = CMA_MAX_CM_RETRIES;
2864 req.srq = id_priv->srq ? 1 : 0;
2866 ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
2868 if (ret && !IS_ERR(id)) {
2869 ib_destroy_cm_id(id);
2870 id_priv->cm_id.ib = NULL;
2873 kfree(private_data);
2877 static int cma_connect_iw(struct rdma_id_private *id_priv,
2878 struct rdma_conn_param *conn_param)
2880 struct iw_cm_id *cm_id;
2881 struct sockaddr_in* sin;
2883 struct iw_cm_conn_param iw_param;
2885 cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
2887 return PTR_ERR(cm_id);
2889 id_priv->cm_id.iw = cm_id;
2891 sin = (struct sockaddr_in *) cma_src_addr(id_priv);
2892 cm_id->local_addr = *sin;
2894 sin = (struct sockaddr_in *) cma_dst_addr(id_priv);
2895 cm_id->remote_addr = *sin;
2897 ret = cma_modify_qp_rtr(id_priv, conn_param);
2902 iw_param.ord = conn_param->initiator_depth;
2903 iw_param.ird = conn_param->responder_resources;
2904 iw_param.private_data = conn_param->private_data;
2905 iw_param.private_data_len = conn_param->private_data_len;
2906 iw_param.qpn = id_priv->id.qp ? id_priv->qp_num : conn_param->qp_num;
2908 memset(&iw_param, 0, sizeof iw_param);
2909 iw_param.qpn = id_priv->qp_num;
2911 ret = iw_cm_connect(cm_id, &iw_param);
2914 iw_destroy_cm_id(cm_id);
2915 id_priv->cm_id.iw = NULL;
2920 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2922 struct rdma_id_private *id_priv;
2925 id_priv = container_of(id, struct rdma_id_private, id);
2926 if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT))
2930 id_priv->qp_num = conn_param->qp_num;
2931 id_priv->srq = conn_param->srq;
2934 switch (rdma_node_get_transport(id->device->node_type)) {
2935 case RDMA_TRANSPORT_IB:
2936 if (id->qp_type == IB_QPT_UD)
2937 ret = cma_resolve_ib_udp(id_priv, conn_param);
2939 ret = cma_connect_ib(id_priv, conn_param);
2941 case RDMA_TRANSPORT_IWARP:
2942 ret = cma_connect_iw(id_priv, conn_param);
2953 cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED);
2956 EXPORT_SYMBOL(rdma_connect);
2958 static int cma_accept_ib(struct rdma_id_private *id_priv,
2959 struct rdma_conn_param *conn_param)
2961 struct ib_cm_rep_param rep;
2964 ret = cma_modify_qp_rtr(id_priv, conn_param);
2968 ret = cma_modify_qp_rts(id_priv, conn_param);
2972 memset(&rep, 0, sizeof rep);
2973 rep.qp_num = id_priv->qp_num;
2974 rep.starting_psn = id_priv->seq_num;
2975 rep.private_data = conn_param->private_data;
2976 rep.private_data_len = conn_param->private_data_len;
2977 rep.responder_resources = conn_param->responder_resources;
2978 rep.initiator_depth = conn_param->initiator_depth;
2979 rep.failover_accepted = 0;
2980 rep.flow_control = conn_param->flow_control;
2981 rep.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
2982 rep.srq = id_priv->srq ? 1 : 0;
2984 ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
2989 static int cma_accept_iw(struct rdma_id_private *id_priv,
2990 struct rdma_conn_param *conn_param)
2992 struct iw_cm_conn_param iw_param;
2995 ret = cma_modify_qp_rtr(id_priv, conn_param);
2999 iw_param.ord = conn_param->initiator_depth;
3000 iw_param.ird = conn_param->responder_resources;
3001 iw_param.private_data = conn_param->private_data;
3002 iw_param.private_data_len = conn_param->private_data_len;
3003 if (id_priv->id.qp) {
3004 iw_param.qpn = id_priv->qp_num;
3006 iw_param.qpn = conn_param->qp_num;
3008 return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
3011 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
3012 enum ib_cm_sidr_status status,
3013 const void *private_data, int private_data_len)
3015 struct ib_cm_sidr_rep_param rep;
3018 memset(&rep, 0, sizeof rep);
3019 rep.status = status;
3020 if (status == IB_SIDR_SUCCESS) {
3021 ret = cma_set_qkey(id_priv);
3024 rep.qp_num = id_priv->qp_num;
3025 rep.qkey = id_priv->qkey;
3027 rep.private_data = private_data;
3028 rep.private_data_len = private_data_len;
3030 return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
3033 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
3035 struct rdma_id_private *id_priv;
3038 id_priv = container_of(id, struct rdma_id_private, id);
3040 id_priv->owner = task_pid_nr(current);
3042 if (!cma_comp(id_priv, RDMA_CM_CONNECT))
3045 if (!id->qp && conn_param) {
3046 id_priv->qp_num = conn_param->qp_num;
3047 id_priv->srq = conn_param->srq;
3050 switch (rdma_node_get_transport(id->device->node_type)) {
3051 case RDMA_TRANSPORT_IB:
3052 if (id->qp_type == IB_QPT_UD) {
3054 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
3055 conn_param->private_data,
3056 conn_param->private_data_len);
3058 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
3062 ret = cma_accept_ib(id_priv, conn_param);
3064 ret = cma_rep_recv(id_priv);
3067 case RDMA_TRANSPORT_IWARP:
3068 ret = cma_accept_iw(id_priv, conn_param);
3080 cma_modify_qp_err(id_priv);
3081 rdma_reject(id, NULL, 0);
3084 EXPORT_SYMBOL(rdma_accept);
3086 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
3088 struct rdma_id_private *id_priv;
3091 id_priv = container_of(id, struct rdma_id_private, id);
3092 if (!id_priv->cm_id.ib)
3095 switch (id->device->node_type) {
3096 case RDMA_NODE_IB_CA:
3097 ret = ib_cm_notify(id_priv->cm_id.ib, event);
3105 EXPORT_SYMBOL(rdma_notify);
3107 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
3108 u8 private_data_len)
3110 struct rdma_id_private *id_priv;
3113 id_priv = container_of(id, struct rdma_id_private, id);
3114 if (!id_priv->cm_id.ib)
3117 switch (rdma_node_get_transport(id->device->node_type)) {
3118 case RDMA_TRANSPORT_IB:
3119 if (id->qp_type == IB_QPT_UD)
3120 ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT,
3121 private_data, private_data_len);
3123 ret = ib_send_cm_rej(id_priv->cm_id.ib,
3124 IB_CM_REJ_CONSUMER_DEFINED, NULL,
3125 0, private_data, private_data_len);
3127 case RDMA_TRANSPORT_IWARP:
3128 ret = iw_cm_reject(id_priv->cm_id.iw,
3129 private_data, private_data_len);
3137 EXPORT_SYMBOL(rdma_reject);
3139 int rdma_disconnect(struct rdma_cm_id *id)
3141 struct rdma_id_private *id_priv;
3144 id_priv = container_of(id, struct rdma_id_private, id);
3145 if (!id_priv->cm_id.ib)
3148 switch (rdma_node_get_transport(id->device->node_type)) {
3149 case RDMA_TRANSPORT_IB:
3150 ret = cma_modify_qp_err(id_priv);
3153 /* Initiate or respond to a disconnect. */
3154 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
3155 ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
3157 case RDMA_TRANSPORT_IWARP:
3158 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
3167 EXPORT_SYMBOL(rdma_disconnect);
3169 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
3171 struct rdma_id_private *id_priv;
3172 struct cma_multicast *mc = multicast->context;
3173 struct rdma_cm_event event;
3176 id_priv = mc->id_priv;
3177 if (cma_disable_callback(id_priv, RDMA_CM_ADDR_BOUND) &&
3178 cma_disable_callback(id_priv, RDMA_CM_ADDR_RESOLVED))
3181 mutex_lock(&id_priv->qp_mutex);
3182 if (!status && id_priv->id.qp)
3183 status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid,
3184 be16_to_cpu(multicast->rec.mlid));
3185 mutex_unlock(&id_priv->qp_mutex);
3187 memset(&event, 0, sizeof event);
3188 event.status = status;
3189 event.param.ud.private_data = mc->context;
3191 event.event = RDMA_CM_EVENT_MULTICAST_JOIN;
3192 ib_init_ah_from_mcmember(id_priv->id.device,
3193 id_priv->id.port_num, &multicast->rec,
3194 &event.param.ud.ah_attr);
3195 event.param.ud.qp_num = 0xFFFFFF;
3196 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
3198 event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
3200 ret = id_priv->id.event_handler(&id_priv->id, &event);
3202 cma_exch(id_priv, RDMA_CM_DESTROYING);
3203 mutex_unlock(&id_priv->handler_mutex);
3204 rdma_destroy_id(&id_priv->id);
3208 mutex_unlock(&id_priv->handler_mutex);
3212 static void cma_set_mgid(struct rdma_id_private *id_priv,
3213 struct sockaddr *addr, union ib_gid *mgid)
3215 unsigned char mc_map[MAX_ADDR_LEN];
3216 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3217 struct sockaddr_in *sin = (struct sockaddr_in *) addr;
3218 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
3220 if (cma_any_addr(addr)) {
3221 memset(mgid, 0, sizeof *mgid);
3222 } else if ((addr->sa_family == AF_INET6) &&
3223 ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
3225 /* IPv6 address is an SA assigned MGID. */
3226 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3227 } else if ((addr->sa_family == AF_INET6)) {
3228 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
3229 if (id_priv->id.ps == RDMA_PS_UDP)
3230 mc_map[7] = 0x01; /* Use RDMA CM signature */
3231 *mgid = *(union ib_gid *) (mc_map + 4);
3233 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
3234 if (id_priv->id.ps == RDMA_PS_UDP)
3235 mc_map[7] = 0x01; /* Use RDMA CM signature */
3236 *mgid = *(union ib_gid *) (mc_map + 4);
3240 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
3241 struct cma_multicast *mc)
3243 struct ib_sa_mcmember_rec rec;
3244 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3245 ib_sa_comp_mask comp_mask;
3248 ib_addr_get_mgid(dev_addr, &rec.mgid);
3249 ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
3254 cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
3255 if (id_priv->id.ps == RDMA_PS_UDP)
3256 rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
3257 rdma_addr_get_sgid(dev_addr, &rec.port_gid);
3258 rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
3261 comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
3262 IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
3263 IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
3264 IB_SA_MCMEMBER_REC_FLOW_LABEL |
3265 IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
3267 if (id_priv->id.ps == RDMA_PS_IPOIB)
3268 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
3269 IB_SA_MCMEMBER_REC_RATE_SELECTOR |
3270 IB_SA_MCMEMBER_REC_MTU_SELECTOR |
3271 IB_SA_MCMEMBER_REC_MTU |
3272 IB_SA_MCMEMBER_REC_HOP_LIMIT;
3274 mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device,
3275 id_priv->id.port_num, &rec,
3276 comp_mask, GFP_KERNEL,
3277 cma_ib_mc_handler, mc);
3278 return PTR_RET(mc->multicast.ib);
3281 static void iboe_mcast_work_handler(struct work_struct *work)
3283 struct iboe_mcast_work *mw = container_of(work, struct iboe_mcast_work, work);
3284 struct cma_multicast *mc = mw->mc;
3285 struct ib_sa_multicast *m = mc->multicast.ib;
3287 mc->multicast.ib->context = mc;
3288 cma_ib_mc_handler(0, m);
3289 kref_put(&mc->mcref, release_mc);
3293 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid)
3295 struct sockaddr_in *sin = (struct sockaddr_in *)addr;
3296 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
3298 if (cma_any_addr(addr)) {
3299 memset(mgid, 0, sizeof *mgid);
3300 } else if (addr->sa_family == AF_INET6) {
3301 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3303 mgid->raw[0] = 0xff;
3304 mgid->raw[1] = 0x0e;
3313 mgid->raw[10] = 0xff;
3314 mgid->raw[11] = 0xff;
3315 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
3319 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
3320 struct cma_multicast *mc)
3322 struct iboe_mcast_work *work;
3323 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3325 struct sockaddr *addr = (struct sockaddr *)&mc->addr;
3326 struct net_device *ndev = NULL;
3328 if (cma_zero_addr((struct sockaddr *)&mc->addr))
3331 work = kzalloc(sizeof *work, GFP_KERNEL);
3335 mc->multicast.ib = kzalloc(sizeof(struct ib_sa_multicast), GFP_KERNEL);
3336 if (!mc->multicast.ib) {
3341 cma_iboe_set_mgid(addr, &mc->multicast.ib->rec.mgid);
3343 mc->multicast.ib->rec.pkey = cpu_to_be16(0xffff);
3344 if (id_priv->id.ps == RDMA_PS_UDP)
3345 mc->multicast.ib->rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
3347 if (dev_addr->bound_dev_if)
3348 ndev = dev_get_by_index(&init_net, dev_addr->bound_dev_if);
3353 mc->multicast.ib->rec.rate = iboe_get_rate(ndev);
3354 mc->multicast.ib->rec.hop_limit = 1;
3355 mc->multicast.ib->rec.mtu = iboe_get_mtu(ndev->mtu);
3357 if (!mc->multicast.ib->rec.mtu) {
3361 iboe_addr_get_sgid(dev_addr, &mc->multicast.ib->rec.port_gid);
3364 INIT_WORK(&work->work, iboe_mcast_work_handler);
3365 kref_get(&mc->mcref);
3366 queue_work(cma_wq, &work->work);
3371 kfree(mc->multicast.ib);
3377 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
3380 struct rdma_id_private *id_priv;
3381 struct cma_multicast *mc;
3384 id_priv = container_of(id, struct rdma_id_private, id);
3385 if (!cma_comp(id_priv, RDMA_CM_ADDR_BOUND) &&
3386 !cma_comp(id_priv, RDMA_CM_ADDR_RESOLVED))
3389 mc = kmalloc(sizeof *mc, GFP_KERNEL);
3393 memcpy(&mc->addr, addr, rdma_addr_size(addr));
3394 mc->context = context;
3395 mc->id_priv = id_priv;
3397 spin_lock(&id_priv->lock);
3398 list_add(&mc->list, &id_priv->mc_list);
3399 spin_unlock(&id_priv->lock);
3401 switch (rdma_node_get_transport(id->device->node_type)) {
3402 case RDMA_TRANSPORT_IB:
3403 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
3404 case IB_LINK_LAYER_INFINIBAND:
3405 ret = cma_join_ib_multicast(id_priv, mc);
3407 case IB_LINK_LAYER_ETHERNET:
3408 kref_init(&mc->mcref);
3409 ret = cma_iboe_join_multicast(id_priv, mc);
3421 spin_lock_irq(&id_priv->lock);
3422 list_del(&mc->list);
3423 spin_unlock_irq(&id_priv->lock);
3428 EXPORT_SYMBOL(rdma_join_multicast);
3430 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
3432 struct rdma_id_private *id_priv;
3433 struct cma_multicast *mc;
3435 id_priv = container_of(id, struct rdma_id_private, id);
3436 spin_lock_irq(&id_priv->lock);
3437 list_for_each_entry(mc, &id_priv->mc_list, list) {
3438 if (!memcmp(&mc->addr, addr, rdma_addr_size(addr))) {
3439 list_del(&mc->list);
3440 spin_unlock_irq(&id_priv->lock);
3443 ib_detach_mcast(id->qp,
3444 &mc->multicast.ib->rec.mgid,
3445 be16_to_cpu(mc->multicast.ib->rec.mlid));
3446 if (rdma_node_get_transport(id_priv->cma_dev->device->node_type) == RDMA_TRANSPORT_IB) {
3447 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
3448 case IB_LINK_LAYER_INFINIBAND:
3449 ib_sa_free_multicast(mc->multicast.ib);
3452 case IB_LINK_LAYER_ETHERNET:
3453 kref_put(&mc->mcref, release_mc);
3462 spin_unlock_irq(&id_priv->lock);
3464 EXPORT_SYMBOL(rdma_leave_multicast);
3466 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
3468 struct rdma_dev_addr *dev_addr;
3469 struct cma_ndev_work *work;
3471 dev_addr = &id_priv->id.route.addr.dev_addr;
3473 if ((dev_addr->bound_dev_if == ndev->ifindex) &&
3474 memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
3475 printk(KERN_INFO "RDMA CM addr change for ndev %s used by id %p\n",
3476 ndev->name, &id_priv->id);
3477 work = kzalloc(sizeof *work, GFP_KERNEL);
3481 INIT_WORK(&work->work, cma_ndev_work_handler);
3483 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
3484 atomic_inc(&id_priv->refcount);
3485 queue_work(cma_wq, &work->work);
3491 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
3494 struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
3495 struct cma_device *cma_dev;
3496 struct rdma_id_private *id_priv;
3497 int ret = NOTIFY_DONE;
3499 if (dev_net(ndev) != &init_net)
3502 if (event != NETDEV_BONDING_FAILOVER)
3505 if (!(ndev->flags & IFF_MASTER) || !(ndev->priv_flags & IFF_BONDING))
3509 list_for_each_entry(cma_dev, &dev_list, list)
3510 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
3511 ret = cma_netdev_change(ndev, id_priv);
3517 mutex_unlock(&lock);
3521 static struct notifier_block cma_nb = {
3522 .notifier_call = cma_netdev_callback
3525 static void cma_add_one(struct ib_device *device)
3527 struct cma_device *cma_dev;
3528 struct rdma_id_private *id_priv;
3530 cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
3534 cma_dev->device = device;
3536 init_completion(&cma_dev->comp);
3537 atomic_set(&cma_dev->refcount, 1);
3538 INIT_LIST_HEAD(&cma_dev->id_list);
3539 ib_set_client_data(device, &cma_client, cma_dev);
3542 list_add_tail(&cma_dev->list, &dev_list);
3543 list_for_each_entry(id_priv, &listen_any_list, list)
3544 cma_listen_on_dev(id_priv, cma_dev);
3545 mutex_unlock(&lock);
3548 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
3550 struct rdma_cm_event event;
3551 enum rdma_cm_state state;
3554 /* Record that we want to remove the device */
3555 state = cma_exch(id_priv, RDMA_CM_DEVICE_REMOVAL);
3556 if (state == RDMA_CM_DESTROYING)
3559 cma_cancel_operation(id_priv, state);
3560 mutex_lock(&id_priv->handler_mutex);
3562 /* Check for destruction from another callback. */
3563 if (!cma_comp(id_priv, RDMA_CM_DEVICE_REMOVAL))
3566 memset(&event, 0, sizeof event);
3567 event.event = RDMA_CM_EVENT_DEVICE_REMOVAL;
3568 ret = id_priv->id.event_handler(&id_priv->id, &event);
3570 mutex_unlock(&id_priv->handler_mutex);
3574 static void cma_process_remove(struct cma_device *cma_dev)
3576 struct rdma_id_private *id_priv;
3580 while (!list_empty(&cma_dev->id_list)) {
3581 id_priv = list_entry(cma_dev->id_list.next,
3582 struct rdma_id_private, list);
3584 list_del(&id_priv->listen_list);
3585 list_del_init(&id_priv->list);
3586 atomic_inc(&id_priv->refcount);
3587 mutex_unlock(&lock);
3589 ret = id_priv->internal_id ? 1 : cma_remove_id_dev(id_priv);
3590 cma_deref_id(id_priv);
3592 rdma_destroy_id(&id_priv->id);
3596 mutex_unlock(&lock);
3598 cma_deref_dev(cma_dev);
3599 wait_for_completion(&cma_dev->comp);
3602 static void cma_remove_one(struct ib_device *device)
3604 struct cma_device *cma_dev;
3606 cma_dev = ib_get_client_data(device, &cma_client);
3611 list_del(&cma_dev->list);
3612 mutex_unlock(&lock);
3614 cma_process_remove(cma_dev);
3618 static int cma_get_id_stats(struct sk_buff *skb, struct netlink_callback *cb)
3620 struct nlmsghdr *nlh;
3621 struct rdma_cm_id_stats *id_stats;
3622 struct rdma_id_private *id_priv;
3623 struct rdma_cm_id *id = NULL;
3624 struct cma_device *cma_dev;
3625 int i_dev = 0, i_id = 0;
3628 * We export all of the IDs as a sequence of messages. Each
3629 * ID gets its own netlink message.
3633 list_for_each_entry(cma_dev, &dev_list, list) {
3634 if (i_dev < cb->args[0]) {
3640 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
3641 if (i_id < cb->args[1]) {
3646 id_stats = ibnl_put_msg(skb, &nlh, cb->nlh->nlmsg_seq,
3647 sizeof *id_stats, RDMA_NL_RDMA_CM,
3648 RDMA_NL_RDMA_CM_ID_STATS);
3652 memset(id_stats, 0, sizeof *id_stats);
3654 id_stats->node_type = id->route.addr.dev_addr.dev_type;
3655 id_stats->port_num = id->port_num;
3656 id_stats->bound_dev_if =
3657 id->route.addr.dev_addr.bound_dev_if;
3659 if (cma_family(id_priv) == AF_INET) {
3660 if (ibnl_put_attr(skb, nlh,
3661 sizeof(struct sockaddr_in),
3662 cma_src_addr(id_priv),
3663 RDMA_NL_RDMA_CM_ATTR_SRC_ADDR)) {
3666 if (ibnl_put_attr(skb, nlh,
3667 sizeof(struct sockaddr_in),
3668 cma_dst_addr(id_priv),
3669 RDMA_NL_RDMA_CM_ATTR_DST_ADDR)) {
3672 } else if (cma_family(id_priv) == AF_INET6) {
3673 if (ibnl_put_attr(skb, nlh,
3674 sizeof(struct sockaddr_in6),
3675 cma_src_addr(id_priv),
3676 RDMA_NL_RDMA_CM_ATTR_SRC_ADDR)) {
3679 if (ibnl_put_attr(skb, nlh,
3680 sizeof(struct sockaddr_in6),
3681 cma_dst_addr(id_priv),
3682 RDMA_NL_RDMA_CM_ATTR_DST_ADDR)) {
3687 id_stats->pid = id_priv->owner;
3688 id_stats->port_space = id->ps;
3689 id_stats->cm_state = id_priv->state;
3690 id_stats->qp_num = id_priv->qp_num;
3691 id_stats->qp_type = id->qp_type;
3701 mutex_unlock(&lock);
3702 cb->args[0] = i_dev;
3708 static const struct ibnl_client_cbs cma_cb_table[] = {
3709 [RDMA_NL_RDMA_CM_ID_STATS] = { .dump = cma_get_id_stats,
3710 .module = THIS_MODULE },
3713 static int __init cma_init(void)
3717 cma_wq = create_singlethread_workqueue("rdma_cm");
3721 ib_sa_register_client(&sa_client);
3722 rdma_addr_register_client(&addr_client);
3723 register_netdevice_notifier(&cma_nb);
3725 ret = ib_register_client(&cma_client);
3729 if (ibnl_add_client(RDMA_NL_RDMA_CM, RDMA_NL_RDMA_CM_NUM_OPS, cma_cb_table))
3730 printk(KERN_WARNING "RDMA CMA: failed to add netlink callback\n");
3735 unregister_netdevice_notifier(&cma_nb);
3736 rdma_addr_unregister_client(&addr_client);
3737 ib_sa_unregister_client(&sa_client);
3738 destroy_workqueue(cma_wq);
3742 static void __exit cma_cleanup(void)
3744 ibnl_remove_client(RDMA_NL_RDMA_CM);
3745 ib_unregister_client(&cma_client);
3746 unregister_netdevice_notifier(&cma_nb);
3747 rdma_addr_unregister_client(&addr_client);
3748 ib_sa_unregister_client(&sa_client);
3749 destroy_workqueue(cma_wq);
3750 idr_destroy(&sdp_ps);
3751 idr_destroy(&tcp_ps);
3752 idr_destroy(&udp_ps);
3753 idr_destroy(&ipoib_ps);
3754 idr_destroy(&ib_ps);
3757 module_init(cma_init);
3758 module_exit(cma_cleanup);