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>
47 #include <rdma/rdma_cm.h>
48 #include <rdma/rdma_cm_ib.h>
49 #include <rdma/ib_cache.h>
50 #include <rdma/ib_cm.h>
51 #include <rdma/ib_sa.h>
52 #include <rdma/iw_cm.h>
54 MODULE_AUTHOR("Sean Hefty");
55 MODULE_DESCRIPTION("Generic RDMA CM Agent");
56 MODULE_LICENSE("Dual BSD/GPL");
58 static int tavor_quirk = 0;
59 module_param_named(tavor_quirk, tavor_quirk, int, 0644);
60 MODULE_PARM_DESC(tavor_quirk, "Tavor performance quirk: limit MTU to 1K if > 0");
62 int unify_tcp_port_space = 1;
63 module_param(unify_tcp_port_space, int, 0644);
64 MODULE_PARM_DESC(unify_tcp_port_space, "Unify the host TCP and RDMA port "
65 "space allocation (default=1)");
67 #define CMA_CM_RESPONSE_TIMEOUT 20
68 #define CMA_MAX_CM_RETRIES 15
69 #define CMA_CM_MRA_SETTING (IB_CM_MRA_FLAG_DELAY | 24)
70 #define IBOE_PACKET_LIFETIME 18
72 static int cma_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
73 module_param_named(cma_response_timeout, cma_response_timeout, int, 0644);
74 MODULE_PARM_DESC(cma_response_timeout, "CMA_CM_RESPONSE_TIMEOUT default=20");
76 static int def_prec2sl = 3;
77 module_param_named(def_prec2sl, def_prec2sl, int, 0644);
78 MODULE_PARM_DESC(def_prec2sl, "Default value for SL priority with RoCE. Valid values 0 - 7");
80 static void cma_add_one(struct ib_device *device);
81 static void cma_remove_one(struct ib_device *device);
83 static struct ib_client cma_client = {
86 .remove = cma_remove_one
89 static struct ib_sa_client sa_client;
90 static struct rdma_addr_client addr_client;
91 static LIST_HEAD(dev_list);
92 static LIST_HEAD(listen_any_list);
93 static DEFINE_MUTEX(lock);
94 static struct workqueue_struct *cma_wq;
95 static DEFINE_IDR(sdp_ps);
96 static DEFINE_IDR(tcp_ps);
97 static DEFINE_IDR(udp_ps);
98 static DEFINE_IDR(ipoib_ps);
100 static int next_port;
104 struct list_head list;
105 struct ib_device *device;
106 struct completion comp;
108 struct list_head id_list;
125 struct rdma_bind_list {
127 struct hlist_head owners;
132 * Device removal can occur at anytime, so we need extra handling to
133 * serialize notifying the user of device removal with other callbacks.
134 * We do this by disabling removal notification while a callback is in process,
135 * and reporting it after the callback completes.
137 struct rdma_id_private {
138 struct rdma_cm_id id;
140 struct rdma_bind_list *bind_list;
142 struct hlist_node node;
143 struct list_head list; /* listen_any_list or cma_device.list */
144 struct list_head listen_list; /* per device listens */
145 struct cma_device *cma_dev;
146 struct list_head mc_list;
149 enum cma_state state;
151 struct mutex qp_mutex;
153 struct completion comp;
155 struct mutex handler_mutex;
159 struct ib_sa_query *query;
173 struct cma_multicast {
174 struct rdma_id_private *id_priv;
176 struct ib_sa_multicast *ib;
178 struct list_head list;
180 struct sockaddr_storage addr;
185 struct work_struct work;
186 struct rdma_id_private *id;
187 enum cma_state old_state;
188 enum cma_state new_state;
189 struct rdma_cm_event event;
192 struct cma_ndev_work {
193 struct work_struct work;
194 struct rdma_id_private *id;
195 struct rdma_cm_event event;
198 struct iboe_mcast_work {
199 struct work_struct work;
200 struct rdma_id_private *id;
201 struct cma_multicast *mc;
214 u8 ip_version; /* IP version: 7:4 */
216 union cma_ip_addr src_addr;
217 union cma_ip_addr dst_addr;
222 u8 sdp_version; /* Major version: 7:4 */
223 u8 ip_version; /* IP version: 7:4 */
224 u8 sdp_specific1[10];
226 __be16 sdp_specific2;
227 union cma_ip_addr src_addr;
228 union cma_ip_addr dst_addr;
236 #define CMA_VERSION 0x00
237 #define SDP_MAJ_VERSION 0x2
239 static int cma_comp(struct rdma_id_private *id_priv, enum cma_state comp)
244 spin_lock_irqsave(&id_priv->lock, flags);
245 ret = (id_priv->state == comp);
246 spin_unlock_irqrestore(&id_priv->lock, flags);
250 static int cma_comp_exch(struct rdma_id_private *id_priv,
251 enum cma_state comp, enum cma_state exch)
256 spin_lock_irqsave(&id_priv->lock, flags);
257 if ((ret = (id_priv->state == comp)))
258 id_priv->state = exch;
259 spin_unlock_irqrestore(&id_priv->lock, flags);
263 static enum cma_state cma_exch(struct rdma_id_private *id_priv,
269 spin_lock_irqsave(&id_priv->lock, flags);
270 old = id_priv->state;
271 id_priv->state = exch;
272 spin_unlock_irqrestore(&id_priv->lock, flags);
276 static inline u8 cma_get_ip_ver(struct cma_hdr *hdr)
278 return hdr->ip_version >> 4;
281 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
283 hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
286 static inline u8 sdp_get_majv(u8 sdp_version)
288 return sdp_version >> 4;
291 static inline u8 sdp_get_ip_ver(struct sdp_hh *hh)
293 return hh->ip_version >> 4;
296 static inline void sdp_set_ip_ver(struct sdp_hh *hh, u8 ip_ver)
298 hh->ip_version = (ip_ver << 4) | (hh->ip_version & 0xF);
301 static inline int cma_is_ud_ps(enum rdma_port_space ps)
303 return (ps == RDMA_PS_UDP || ps == RDMA_PS_IPOIB);
306 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
307 struct cma_device *cma_dev)
309 atomic_inc(&cma_dev->refcount);
310 id_priv->cma_dev = cma_dev;
311 id_priv->id.device = cma_dev->device;
312 id_priv->id.route.addr.dev_addr.transport =
313 rdma_node_get_transport(cma_dev->device->node_type);
314 list_add_tail(&id_priv->list, &cma_dev->id_list);
317 static inline void cma_deref_dev(struct cma_device *cma_dev)
319 if (atomic_dec_and_test(&cma_dev->refcount))
320 complete(&cma_dev->comp);
323 static inline void release_mc(struct kref *kref)
325 struct cma_multicast *mc = container_of(kref, struct cma_multicast, mcref);
327 kfree(mc->multicast.ib);
331 static void cma_detach_from_dev(struct rdma_id_private *id_priv)
333 list_del(&id_priv->list);
334 cma_deref_dev(id_priv->cma_dev);
335 id_priv->cma_dev = NULL;
338 static int cma_set_qkey(struct rdma_id_private *id_priv)
340 struct ib_sa_mcmember_rec rec;
346 switch (id_priv->id.ps) {
348 id_priv->qkey = RDMA_UDP_QKEY;
351 ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid);
352 ret = ib_sa_get_mcmember_rec(id_priv->id.device,
353 id_priv->id.port_num, &rec.mgid,
356 id_priv->qkey = be32_to_cpu(rec.qkey);
364 static int cma_acquire_dev(struct rdma_id_private *id_priv)
366 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
367 struct cma_device *cma_dev;
371 if (dev_addr->dev_type != ARPHRD_INFINIBAND) {
372 iboe_addr_get_sgid(dev_addr, &gid);
373 list_for_each_entry(cma_dev, &dev_list, list) {
374 ret = ib_find_cached_gid(cma_dev->device, &gid,
375 &id_priv->id.port_num, NULL);
381 memcpy(&gid, dev_addr->src_dev_addr +
382 rdma_addr_gid_offset(dev_addr), sizeof gid);
383 list_for_each_entry(cma_dev, &dev_list, list) {
384 ret = ib_find_cached_gid(cma_dev->device, &gid,
385 &id_priv->id.port_num, NULL);
392 cma_attach_to_dev(id_priv, cma_dev);
397 static void cma_deref_id(struct rdma_id_private *id_priv)
399 if (atomic_dec_and_test(&id_priv->refcount))
400 complete(&id_priv->comp);
403 static int cma_disable_callback(struct rdma_id_private *id_priv,
404 enum cma_state state)
406 mutex_lock(&id_priv->handler_mutex);
407 if (id_priv->state != state) {
408 mutex_unlock(&id_priv->handler_mutex);
414 static int cma_has_cm_dev(struct rdma_id_private *id_priv)
416 return (id_priv->id.device && id_priv->cm_id.ib);
419 struct rdma_cm_id *rdma_create_id(rdma_cm_event_handler event_handler,
420 void *context, enum rdma_port_space ps)
422 struct rdma_id_private *id_priv;
424 id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
426 return ERR_PTR(-ENOMEM);
428 id_priv->state = CMA_IDLE;
429 id_priv->id.context = context;
430 id_priv->id.event_handler = event_handler;
432 spin_lock_init(&id_priv->lock);
433 mutex_init(&id_priv->qp_mutex);
434 init_completion(&id_priv->comp);
435 atomic_set(&id_priv->refcount, 1);
436 mutex_init(&id_priv->handler_mutex);
437 INIT_LIST_HEAD(&id_priv->listen_list);
438 INIT_LIST_HEAD(&id_priv->mc_list);
439 get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
443 EXPORT_SYMBOL(rdma_create_id);
445 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
447 struct ib_qp_attr qp_attr;
448 int qp_attr_mask, ret;
450 qp_attr.qp_state = IB_QPS_INIT;
451 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
455 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
459 qp_attr.qp_state = IB_QPS_RTR;
460 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
464 qp_attr.qp_state = IB_QPS_RTS;
466 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
471 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
473 struct ib_qp_attr qp_attr;
474 int qp_attr_mask, ret;
476 qp_attr.qp_state = IB_QPS_INIT;
477 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
481 return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
484 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
485 struct ib_qp_init_attr *qp_init_attr)
487 struct rdma_id_private *id_priv;
491 id_priv = container_of(id, struct rdma_id_private, id);
492 if (id->device != pd->device)
495 qp = ib_create_qp(pd, qp_init_attr);
499 if (cma_is_ud_ps(id_priv->id.ps))
500 ret = cma_init_ud_qp(id_priv, qp);
502 ret = cma_init_conn_qp(id_priv, qp);
507 id_priv->qp_num = qp->qp_num;
508 id_priv->srq = (qp->srq != NULL);
514 EXPORT_SYMBOL(rdma_create_qp);
516 void rdma_destroy_qp(struct rdma_cm_id *id)
518 struct rdma_id_private *id_priv;
520 id_priv = container_of(id, struct rdma_id_private, id);
521 mutex_lock(&id_priv->qp_mutex);
522 ib_destroy_qp(id_priv->id.qp);
523 id_priv->id.qp = NULL;
524 mutex_unlock(&id_priv->qp_mutex);
526 EXPORT_SYMBOL(rdma_destroy_qp);
528 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
529 struct rdma_conn_param *conn_param)
531 struct ib_qp_attr qp_attr;
532 int qp_attr_mask, ret;
534 mutex_lock(&id_priv->qp_mutex);
535 if (!id_priv->id.qp) {
540 /* Need to update QP attributes from default values. */
541 qp_attr.qp_state = IB_QPS_INIT;
542 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
546 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
550 qp_attr.qp_state = IB_QPS_RTR;
551 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
556 qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
557 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
559 mutex_unlock(&id_priv->qp_mutex);
563 static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
564 struct rdma_conn_param *conn_param)
566 struct ib_qp_attr qp_attr;
567 int qp_attr_mask, ret;
569 mutex_lock(&id_priv->qp_mutex);
570 if (!id_priv->id.qp) {
575 qp_attr.qp_state = IB_QPS_RTS;
576 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
581 qp_attr.max_rd_atomic = conn_param->initiator_depth;
582 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
584 mutex_unlock(&id_priv->qp_mutex);
588 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
590 struct ib_qp_attr qp_attr;
593 mutex_lock(&id_priv->qp_mutex);
594 if (!id_priv->id.qp) {
599 qp_attr.qp_state = IB_QPS_ERR;
600 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
602 mutex_unlock(&id_priv->qp_mutex);
606 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
607 struct ib_qp_attr *qp_attr, int *qp_attr_mask)
609 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
613 if (rdma_port_get_link_layer(id_priv->id.device, id_priv->id.port_num) ==
614 IB_LINK_LAYER_INFINIBAND)
615 pkey = ib_addr_get_pkey(dev_addr);
619 ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
620 pkey, &qp_attr->pkey_index);
624 qp_attr->port_num = id_priv->id.port_num;
625 *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
627 if (cma_is_ud_ps(id_priv->id.ps)) {
628 ret = cma_set_qkey(id_priv);
632 qp_attr->qkey = id_priv->qkey;
633 *qp_attr_mask |= IB_QP_QKEY;
635 qp_attr->qp_access_flags = 0;
636 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
641 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
644 struct rdma_id_private *id_priv;
647 id_priv = container_of(id, struct rdma_id_private, id);
648 switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
649 case RDMA_TRANSPORT_IB:
650 if (!id_priv->cm_id.ib || cma_is_ud_ps(id_priv->id.ps))
651 ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
653 ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
655 if (qp_attr->qp_state == IB_QPS_RTR)
656 qp_attr->rq_psn = id_priv->seq_num;
658 case RDMA_TRANSPORT_IWARP:
659 if (!id_priv->cm_id.iw) {
660 qp_attr->qp_access_flags = 0;
661 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
663 ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
673 EXPORT_SYMBOL(rdma_init_qp_attr);
675 static inline int cma_zero_addr(struct sockaddr *addr)
677 struct in6_addr *ip6;
679 if (addr->sa_family == AF_INET)
680 return ipv4_is_zeronet(
681 ((struct sockaddr_in *)addr)->sin_addr.s_addr);
683 ip6 = &((struct sockaddr_in6 *) addr)->sin6_addr;
684 return (ip6->s6_addr32[0] | ip6->s6_addr32[1] |
685 ip6->s6_addr32[2] | ip6->s6_addr32[3]) == 0;
689 static inline int cma_loopback_addr(struct sockaddr *addr)
691 if (addr->sa_family == AF_INET)
692 return ipv4_is_loopback(
693 ((struct sockaddr_in *) addr)->sin_addr.s_addr);
695 return ipv6_addr_loopback(
696 &((struct sockaddr_in6 *) addr)->sin6_addr);
699 static inline int cma_any_addr(struct sockaddr *addr)
701 return cma_zero_addr(addr) || cma_loopback_addr(addr);
704 static inline __be16 cma_port(struct sockaddr *addr)
706 if (addr->sa_family == AF_INET)
707 return ((struct sockaddr_in *) addr)->sin_port;
709 return ((struct sockaddr_in6 *) addr)->sin6_port;
712 static inline int cma_any_port(struct sockaddr *addr)
714 return !cma_port(addr);
717 static int cma_get_net_info(void *hdr, enum rdma_port_space ps,
718 u8 *ip_ver, __be16 *port,
719 union cma_ip_addr **src, union cma_ip_addr **dst)
723 if (sdp_get_majv(((struct sdp_hh *) hdr)->sdp_version) !=
727 *ip_ver = sdp_get_ip_ver(hdr);
728 *port = ((struct sdp_hh *) hdr)->port;
729 *src = &((struct sdp_hh *) hdr)->src_addr;
730 *dst = &((struct sdp_hh *) hdr)->dst_addr;
733 if (((struct cma_hdr *) hdr)->cma_version != CMA_VERSION)
736 *ip_ver = cma_get_ip_ver(hdr);
737 *port = ((struct cma_hdr *) hdr)->port;
738 *src = &((struct cma_hdr *) hdr)->src_addr;
739 *dst = &((struct cma_hdr *) hdr)->dst_addr;
743 if (*ip_ver != 4 && *ip_ver != 6)
748 static void cma_save_net_info(struct rdma_addr *addr,
749 struct rdma_addr *listen_addr,
750 u8 ip_ver, __be16 port,
751 union cma_ip_addr *src, union cma_ip_addr *dst)
753 struct sockaddr_in *listen4, *ip4;
754 struct sockaddr_in6 *listen6, *ip6;
758 listen4 = (struct sockaddr_in *) &listen_addr->src_addr;
759 ip4 = (struct sockaddr_in *) &addr->src_addr;
760 ip4->sin_family = listen4->sin_family;
761 ip4->sin_addr.s_addr = dst->ip4.addr;
762 ip4->sin_port = listen4->sin_port;
764 ip4 = (struct sockaddr_in *) &addr->dst_addr;
765 ip4->sin_family = listen4->sin_family;
766 ip4->sin_addr.s_addr = src->ip4.addr;
767 ip4->sin_port = port;
770 listen6 = (struct sockaddr_in6 *) &listen_addr->src_addr;
771 ip6 = (struct sockaddr_in6 *) &addr->src_addr;
772 ip6->sin6_family = listen6->sin6_family;
773 ip6->sin6_addr = dst->ip6;
774 ip6->sin6_port = listen6->sin6_port;
776 ip6 = (struct sockaddr_in6 *) &addr->dst_addr;
777 ip6->sin6_family = listen6->sin6_family;
778 ip6->sin6_addr = src->ip6;
779 ip6->sin6_port = port;
786 static inline int cma_user_data_offset(enum rdma_port_space ps)
792 return sizeof(struct cma_hdr);
796 static void cma_cancel_route(struct rdma_id_private *id_priv)
798 switch (rdma_port_get_link_layer(id_priv->id.device, id_priv->id.port_num)) {
799 case IB_LINK_LAYER_INFINIBAND:
801 ib_sa_cancel_query(id_priv->query_id, id_priv->query);
808 static void cma_cancel_listens(struct rdma_id_private *id_priv)
810 struct rdma_id_private *dev_id_priv;
813 * Remove from listen_any_list to prevent added devices from spawning
814 * additional listen requests.
817 list_del(&id_priv->list);
819 while (!list_empty(&id_priv->listen_list)) {
820 dev_id_priv = list_entry(id_priv->listen_list.next,
821 struct rdma_id_private, listen_list);
822 /* sync with device removal to avoid duplicate destruction */
823 list_del_init(&dev_id_priv->list);
824 list_del(&dev_id_priv->listen_list);
827 rdma_destroy_id(&dev_id_priv->id);
833 static void cma_cancel_operation(struct rdma_id_private *id_priv,
834 enum cma_state state)
838 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
840 case CMA_ROUTE_QUERY:
841 cma_cancel_route(id_priv);
844 if (cma_any_addr((struct sockaddr *) &id_priv->id.route.addr.src_addr)
845 && !id_priv->cma_dev)
846 cma_cancel_listens(id_priv);
853 static void cma_release_port(struct rdma_id_private *id_priv)
855 struct rdma_bind_list *bind_list = id_priv->bind_list;
861 hlist_del(&id_priv->node);
862 if (hlist_empty(&bind_list->owners)) {
863 idr_remove(bind_list->ps, bind_list->port);
868 sock_release(id_priv->sock);
871 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
873 struct cma_multicast *mc;
875 while (!list_empty(&id_priv->mc_list)) {
876 mc = container_of(id_priv->mc_list.next,
877 struct cma_multicast, list);
879 switch (rdma_port_get_link_layer(id_priv->cma_dev->device, id_priv->id.port_num)) {
880 case IB_LINK_LAYER_INFINIBAND:
881 ib_sa_free_multicast(mc->multicast.ib);
884 case IB_LINK_LAYER_ETHERNET:
885 kref_put(&mc->mcref, release_mc);
893 void rdma_destroy_id(struct rdma_cm_id *id)
895 struct rdma_id_private *id_priv;
896 enum cma_state state;
898 id_priv = container_of(id, struct rdma_id_private, id);
899 state = cma_exch(id_priv, CMA_DESTROYING);
900 cma_cancel_operation(id_priv, state);
903 if (id_priv->cma_dev) {
905 switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
906 case RDMA_TRANSPORT_IB:
907 if (id_priv->cm_id.ib && !IS_ERR(id_priv->cm_id.ib))
908 ib_destroy_cm_id(id_priv->cm_id.ib);
910 case RDMA_TRANSPORT_IWARP:
911 if (id_priv->cm_id.iw && !IS_ERR(id_priv->cm_id.iw))
912 iw_destroy_cm_id(id_priv->cm_id.iw);
917 cma_leave_mc_groups(id_priv);
919 cma_detach_from_dev(id_priv);
923 cma_release_port(id_priv);
924 cma_deref_id(id_priv);
925 wait_for_completion(&id_priv->comp);
927 if (id_priv->internal_id)
928 cma_deref_id(id_priv->id.context);
930 kfree(id_priv->id.route.path_rec);
933 EXPORT_SYMBOL(rdma_destroy_id);
935 static int cma_rep_recv(struct rdma_id_private *id_priv)
939 ret = cma_modify_qp_rtr(id_priv, NULL);
943 ret = cma_modify_qp_rts(id_priv, NULL);
947 ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
953 cma_modify_qp_err(id_priv);
954 ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
959 static int cma_verify_rep(struct rdma_id_private *id_priv, void *data)
961 if (id_priv->id.ps == RDMA_PS_SDP &&
962 sdp_get_majv(((struct sdp_hah *) data)->sdp_version) !=
969 static void cma_set_rep_event_data(struct rdma_cm_event *event,
970 struct ib_cm_rep_event_param *rep_data,
973 event->param.conn.private_data = private_data;
974 event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
975 event->param.conn.responder_resources = rep_data->responder_resources;
976 event->param.conn.initiator_depth = rep_data->initiator_depth;
977 event->param.conn.flow_control = rep_data->flow_control;
978 event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
979 event->param.conn.srq = rep_data->srq;
980 event->param.conn.qp_num = rep_data->remote_qpn;
983 static int cma_ib_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
985 struct rdma_id_private *id_priv = cm_id->context;
986 struct rdma_cm_event event;
989 if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
990 cma_disable_callback(id_priv, CMA_CONNECT)) ||
991 (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
992 cma_disable_callback(id_priv, CMA_DISCONNECT)))
995 memset(&event, 0, sizeof event);
996 switch (ib_event->event) {
997 case IB_CM_REQ_ERROR:
998 case IB_CM_REP_ERROR:
999 event.event = RDMA_CM_EVENT_UNREACHABLE;
1000 event.status = -ETIMEDOUT;
1002 case IB_CM_REP_RECEIVED:
1003 event.status = cma_verify_rep(id_priv, ib_event->private_data);
1005 event.event = RDMA_CM_EVENT_CONNECT_ERROR;
1006 else if (id_priv->id.qp && id_priv->id.ps != RDMA_PS_SDP) {
1007 event.status = cma_rep_recv(id_priv);
1008 event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
1009 RDMA_CM_EVENT_ESTABLISHED;
1011 event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
1012 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
1013 ib_event->private_data);
1015 case IB_CM_RTU_RECEIVED:
1016 case IB_CM_USER_ESTABLISHED:
1017 event.event = RDMA_CM_EVENT_ESTABLISHED;
1019 case IB_CM_DREQ_ERROR:
1020 event.status = -ETIMEDOUT; /* fall through */
1021 case IB_CM_DREQ_RECEIVED:
1022 case IB_CM_DREP_RECEIVED:
1023 if (!cma_comp_exch(id_priv, CMA_CONNECT, CMA_DISCONNECT))
1025 event.event = RDMA_CM_EVENT_DISCONNECTED;
1027 case IB_CM_TIMEWAIT_EXIT:
1028 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
1030 case IB_CM_MRA_RECEIVED:
1033 case IB_CM_REJ_RECEIVED:
1034 cma_modify_qp_err(id_priv);
1035 event.status = ib_event->param.rej_rcvd.reason;
1036 event.event = RDMA_CM_EVENT_REJECTED;
1037 event.param.conn.private_data = ib_event->private_data;
1038 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
1041 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
1046 ret = id_priv->id.event_handler(&id_priv->id, &event);
1048 /* Destroy the CM ID by returning a non-zero value. */
1049 id_priv->cm_id.ib = NULL;
1050 cma_exch(id_priv, CMA_DESTROYING);
1051 mutex_unlock(&id_priv->handler_mutex);
1052 rdma_destroy_id(&id_priv->id);
1056 mutex_unlock(&id_priv->handler_mutex);
1060 static struct rdma_id_private *cma_new_conn_id(struct rdma_cm_id *listen_id,
1061 struct ib_cm_event *ib_event)
1063 struct rdma_id_private *id_priv;
1064 struct rdma_cm_id *id;
1065 struct rdma_route *rt;
1066 union cma_ip_addr *src, *dst;
1071 if (cma_get_net_info(ib_event->private_data, listen_id->ps,
1072 &ip_ver, &port, &src, &dst))
1075 id = rdma_create_id(listen_id->event_handler, listen_id->context,
1080 cma_save_net_info(&id->route.addr, &listen_id->route.addr,
1081 ip_ver, port, src, dst);
1084 rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
1085 rt->path_rec = kmalloc(sizeof *rt->path_rec * rt->num_paths,
1090 rt->path_rec[0] = *ib_event->param.req_rcvd.primary_path;
1091 if (rt->num_paths == 2)
1092 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
1094 if (cma_any_addr((struct sockaddr *) &rt->addr.src_addr)) {
1095 rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
1096 rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
1097 ib_addr_set_pkey(&rt->addr.dev_addr, rt->path_rec[0].pkey);
1099 ret = rdma_translate_ip((struct sockaddr *) &rt->addr.src_addr,
1100 &rt->addr.dev_addr);
1104 rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
1106 id_priv = container_of(id, struct rdma_id_private, id);
1107 id_priv->state = CMA_CONNECT;
1111 rdma_destroy_id(id);
1116 static struct rdma_id_private *cma_new_udp_id(struct rdma_cm_id *listen_id,
1117 struct ib_cm_event *ib_event)
1119 struct rdma_id_private *id_priv;
1120 struct rdma_cm_id *id;
1121 union cma_ip_addr *src, *dst;
1126 id = rdma_create_id(listen_id->event_handler, listen_id->context,
1132 if (cma_get_net_info(ib_event->private_data, listen_id->ps,
1133 &ip_ver, &port, &src, &dst))
1136 cma_save_net_info(&id->route.addr, &listen_id->route.addr,
1137 ip_ver, port, src, dst);
1139 if (!cma_any_addr((struct sockaddr *) &id->route.addr.src_addr)) {
1140 ret = rdma_translate_ip((struct sockaddr *) &id->route.addr.src_addr,
1141 &id->route.addr.dev_addr);
1146 id_priv = container_of(id, struct rdma_id_private, id);
1147 id_priv->state = CMA_CONNECT;
1150 rdma_destroy_id(id);
1154 static void cma_set_req_event_data(struct rdma_cm_event *event,
1155 struct ib_cm_req_event_param *req_data,
1156 void *private_data, int offset)
1158 event->param.conn.private_data = private_data + offset;
1159 event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
1160 event->param.conn.responder_resources = req_data->responder_resources;
1161 event->param.conn.initiator_depth = req_data->initiator_depth;
1162 event->param.conn.flow_control = req_data->flow_control;
1163 event->param.conn.retry_count = req_data->retry_count;
1164 event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
1165 event->param.conn.srq = req_data->srq;
1166 event->param.conn.qp_num = req_data->remote_qpn;
1169 static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1171 struct rdma_id_private *listen_id, *conn_id;
1172 struct rdma_cm_event event;
1175 listen_id = cm_id->context;
1176 if (cma_disable_callback(listen_id, CMA_LISTEN))
1177 return -ECONNABORTED;
1179 memset(&event, 0, sizeof event);
1180 offset = cma_user_data_offset(listen_id->id.ps);
1181 event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1182 if (cma_is_ud_ps(listen_id->id.ps)) {
1183 conn_id = cma_new_udp_id(&listen_id->id, ib_event);
1184 event.param.ud.private_data = ib_event->private_data + offset;
1185 event.param.ud.private_data_len =
1186 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
1188 conn_id = cma_new_conn_id(&listen_id->id, ib_event);
1189 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
1190 ib_event->private_data, offset);
1197 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1199 ret = cma_acquire_dev(conn_id);
1200 mutex_unlock(&lock);
1202 goto release_conn_id;
1204 conn_id->cm_id.ib = cm_id;
1205 cm_id->context = conn_id;
1206 cm_id->cm_handler = cma_ib_handler;
1208 ret = conn_id->id.event_handler(&conn_id->id, &event);
1211 * Acquire mutex to prevent user executing rdma_destroy_id()
1212 * while we're accessing the cm_id.
1215 if (cma_comp(conn_id, CMA_CONNECT) &&
1216 !cma_is_ud_ps(conn_id->id.ps))
1217 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
1218 mutex_unlock(&lock);
1219 mutex_unlock(&conn_id->handler_mutex);
1223 /* Destroy the CM ID by returning a non-zero value. */
1224 conn_id->cm_id.ib = NULL;
1227 cma_exch(conn_id, CMA_DESTROYING);
1228 mutex_unlock(&conn_id->handler_mutex);
1229 rdma_destroy_id(&conn_id->id);
1232 mutex_unlock(&listen_id->handler_mutex);
1236 static __be64 cma_get_service_id(enum rdma_port_space ps, struct sockaddr *addr)
1238 return cpu_to_be64(((u64)ps << 16) + be16_to_cpu(cma_port(addr)));
1241 static void cma_set_compare_data(enum rdma_port_space ps, struct sockaddr *addr,
1242 struct ib_cm_compare_data *compare)
1244 struct cma_hdr *cma_data, *cma_mask;
1245 struct sdp_hh *sdp_data, *sdp_mask;
1248 struct in6_addr ip6_addr;
1251 memset(compare, 0, sizeof *compare);
1252 cma_data = (void *) compare->data;
1253 cma_mask = (void *) compare->mask;
1254 sdp_data = (void *) compare->data;
1255 sdp_mask = (void *) compare->mask;
1257 switch (addr->sa_family) {
1259 ip4_addr = ((struct sockaddr_in *) addr)->sin_addr.s_addr;
1260 if (ps == RDMA_PS_SDP) {
1261 sdp_set_ip_ver(sdp_data, 4);
1262 sdp_set_ip_ver(sdp_mask, 0xF);
1263 sdp_data->dst_addr.ip4.addr = ip4_addr;
1264 sdp_mask->dst_addr.ip4.addr = htonl(~0);
1266 cma_set_ip_ver(cma_data, 4);
1267 cma_set_ip_ver(cma_mask, 0xF);
1268 cma_data->dst_addr.ip4.addr = ip4_addr;
1269 cma_mask->dst_addr.ip4.addr = htonl(~0);
1274 ip6_addr = ((struct sockaddr_in6 *) addr)->sin6_addr;
1275 if (ps == RDMA_PS_SDP) {
1276 sdp_set_ip_ver(sdp_data, 6);
1277 sdp_set_ip_ver(sdp_mask, 0xF);
1278 sdp_data->dst_addr.ip6 = ip6_addr;
1279 memset(&sdp_mask->dst_addr.ip6, 0xFF,
1280 sizeof sdp_mask->dst_addr.ip6);
1282 cma_set_ip_ver(cma_data, 6);
1283 cma_set_ip_ver(cma_mask, 0xF);
1284 cma_data->dst_addr.ip6 = ip6_addr;
1285 memset(&cma_mask->dst_addr.ip6, 0xFF,
1286 sizeof cma_mask->dst_addr.ip6);
1295 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
1297 struct rdma_id_private *id_priv = iw_id->context;
1298 struct rdma_cm_event event;
1299 struct sockaddr_in *sin;
1302 if (cma_disable_callback(id_priv, CMA_CONNECT))
1305 memset(&event, 0, sizeof event);
1306 switch (iw_event->event) {
1307 case IW_CM_EVENT_CLOSE:
1308 event.event = RDMA_CM_EVENT_DISCONNECTED;
1310 case IW_CM_EVENT_CONNECT_REPLY:
1311 sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1312 *sin = iw_event->local_addr;
1313 sin = (struct sockaddr_in *) &id_priv->id.route.addr.dst_addr;
1314 *sin = iw_event->remote_addr;
1315 switch ((int)iw_event->status) {
1317 event.event = RDMA_CM_EVENT_ESTABLISHED;
1321 event.event = RDMA_CM_EVENT_REJECTED;
1324 event.event = RDMA_CM_EVENT_UNREACHABLE;
1327 event.event = RDMA_CM_EVENT_CONNECT_ERROR;
1331 case IW_CM_EVENT_ESTABLISHED:
1332 event.event = RDMA_CM_EVENT_ESTABLISHED;
1338 event.status = iw_event->status;
1339 event.param.conn.private_data = iw_event->private_data;
1340 event.param.conn.private_data_len = iw_event->private_data_len;
1341 ret = id_priv->id.event_handler(&id_priv->id, &event);
1343 /* Destroy the CM ID by returning a non-zero value. */
1344 id_priv->cm_id.iw = NULL;
1345 cma_exch(id_priv, CMA_DESTROYING);
1346 mutex_unlock(&id_priv->handler_mutex);
1347 rdma_destroy_id(&id_priv->id);
1351 mutex_unlock(&id_priv->handler_mutex);
1355 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
1356 struct iw_cm_event *iw_event)
1358 struct rdma_cm_id *new_cm_id;
1359 struct rdma_id_private *listen_id, *conn_id;
1360 struct sockaddr_in *sin;
1361 struct net_device *dev = NULL;
1362 struct rdma_cm_event event;
1364 struct ib_device_attr attr;
1366 listen_id = cm_id->context;
1367 if (cma_disable_callback(listen_id, CMA_LISTEN))
1368 return -ECONNABORTED;
1370 /* Create a new RDMA id for the new IW CM ID */
1371 new_cm_id = rdma_create_id(listen_id->id.event_handler,
1372 listen_id->id.context,
1374 if (IS_ERR(new_cm_id)) {
1378 conn_id = container_of(new_cm_id, struct rdma_id_private, id);
1379 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1380 conn_id->state = CMA_CONNECT;
1382 dev = ip_dev_find(NULL, iw_event->local_addr.sin_addr.s_addr);
1384 ret = -EADDRNOTAVAIL;
1385 mutex_unlock(&conn_id->handler_mutex);
1386 rdma_destroy_id(new_cm_id);
1389 ret = rdma_copy_addr(&conn_id->id.route.addr.dev_addr, dev, NULL);
1391 mutex_unlock(&conn_id->handler_mutex);
1392 rdma_destroy_id(new_cm_id);
1397 ret = cma_acquire_dev(conn_id);
1398 mutex_unlock(&lock);
1400 mutex_unlock(&conn_id->handler_mutex);
1401 rdma_destroy_id(new_cm_id);
1405 conn_id->cm_id.iw = cm_id;
1406 cm_id->context = conn_id;
1407 cm_id->cm_handler = cma_iw_handler;
1409 sin = (struct sockaddr_in *) &new_cm_id->route.addr.src_addr;
1410 *sin = iw_event->local_addr;
1411 sin = (struct sockaddr_in *) &new_cm_id->route.addr.dst_addr;
1412 *sin = iw_event->remote_addr;
1414 ret = ib_query_device(conn_id->id.device, &attr);
1416 mutex_unlock(&conn_id->handler_mutex);
1417 rdma_destroy_id(new_cm_id);
1421 memset(&event, 0, sizeof event);
1422 event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1423 event.param.conn.private_data = iw_event->private_data;
1424 event.param.conn.private_data_len = iw_event->private_data_len;
1425 event.param.conn.initiator_depth = attr.max_qp_init_rd_atom;
1426 event.param.conn.responder_resources = attr.max_qp_rd_atom;
1427 ret = conn_id->id.event_handler(&conn_id->id, &event);
1429 /* User wants to destroy the CM ID */
1430 conn_id->cm_id.iw = NULL;
1431 cma_exch(conn_id, CMA_DESTROYING);
1432 mutex_unlock(&conn_id->handler_mutex);
1433 rdma_destroy_id(&conn_id->id);
1437 mutex_unlock(&conn_id->handler_mutex);
1442 mutex_unlock(&listen_id->handler_mutex);
1446 static int cma_ib_listen(struct rdma_id_private *id_priv)
1448 struct ib_cm_compare_data compare_data;
1449 struct sockaddr *addr;
1453 id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, cma_req_handler,
1455 if (IS_ERR(id_priv->cm_id.ib))
1456 return PTR_ERR(id_priv->cm_id.ib);
1458 addr = (struct sockaddr *) &id_priv->id.route.addr.src_addr;
1459 svc_id = cma_get_service_id(id_priv->id.ps, addr);
1460 if (cma_any_addr(addr))
1461 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, NULL);
1463 cma_set_compare_data(id_priv->id.ps, addr, &compare_data);
1464 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, &compare_data);
1468 ib_destroy_cm_id(id_priv->cm_id.ib);
1469 id_priv->cm_id.ib = NULL;
1475 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
1478 struct sockaddr_in *sin;
1480 id_priv->cm_id.iw = iw_create_cm_id(id_priv->id.device,
1482 iw_conn_req_handler,
1484 if (IS_ERR(id_priv->cm_id.iw))
1485 return PTR_ERR(id_priv->cm_id.iw);
1487 sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1488 id_priv->cm_id.iw->local_addr = *sin;
1490 ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
1493 iw_destroy_cm_id(id_priv->cm_id.iw);
1494 id_priv->cm_id.iw = NULL;
1500 static int cma_listen_handler(struct rdma_cm_id *id,
1501 struct rdma_cm_event *event)
1503 struct rdma_id_private *id_priv = id->context;
1505 id->context = id_priv->id.context;
1506 id->event_handler = id_priv->id.event_handler;
1507 return id_priv->id.event_handler(id, event);
1510 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
1511 struct cma_device *cma_dev)
1513 struct rdma_id_private *dev_id_priv;
1514 struct rdma_cm_id *id;
1517 id = rdma_create_id(cma_listen_handler, id_priv, id_priv->id.ps);
1521 dev_id_priv = container_of(id, struct rdma_id_private, id);
1523 dev_id_priv->state = CMA_ADDR_BOUND;
1524 memcpy(&id->route.addr.src_addr, &id_priv->id.route.addr.src_addr,
1525 ip_addr_size((struct sockaddr *) &id_priv->id.route.addr.src_addr));
1527 cma_attach_to_dev(dev_id_priv, cma_dev);
1528 list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
1529 atomic_inc(&id_priv->refcount);
1530 dev_id_priv->internal_id = 1;
1532 ret = rdma_listen(id, id_priv->backlog);
1534 printk(KERN_WARNING "RDMA CMA: cma_listen_on_dev, error %d, "
1535 "listening on device %s\n", ret, cma_dev->device->name);
1538 static void cma_listen_on_all(struct rdma_id_private *id_priv)
1540 struct cma_device *cma_dev;
1543 list_add_tail(&id_priv->list, &listen_any_list);
1544 list_for_each_entry(cma_dev, &dev_list, list)
1545 cma_listen_on_dev(id_priv, cma_dev);
1546 mutex_unlock(&lock);
1549 int rdma_listen(struct rdma_cm_id *id, int backlog)
1551 struct rdma_id_private *id_priv;
1554 id_priv = container_of(id, struct rdma_id_private, id);
1555 if (id_priv->state == CMA_IDLE) {
1556 ((struct sockaddr *) &id->route.addr.src_addr)->sa_family = AF_INET;
1557 ret = rdma_bind_addr(id, (struct sockaddr *) &id->route.addr.src_addr);
1562 if (!cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_LISTEN))
1565 id_priv->backlog = backlog;
1567 switch (rdma_node_get_transport(id->device->node_type)) {
1568 case RDMA_TRANSPORT_IB:
1569 ret = cma_ib_listen(id_priv);
1573 case RDMA_TRANSPORT_IWARP:
1574 ret = cma_iw_listen(id_priv, backlog);
1583 cma_listen_on_all(id_priv);
1587 id_priv->backlog = 0;
1588 cma_comp_exch(id_priv, CMA_LISTEN, CMA_ADDR_BOUND);
1591 EXPORT_SYMBOL(rdma_listen);
1593 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
1595 struct rdma_id_private *id_priv;
1597 id_priv = container_of(id, struct rdma_id_private, id);
1598 id_priv->tos = (u8) tos;
1600 EXPORT_SYMBOL(rdma_set_service_type);
1602 static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec,
1605 struct cma_work *work = context;
1606 struct rdma_route *route;
1608 route = &work->id->id.route;
1611 route->num_paths = 1;
1612 *route->path_rec = *path_rec;
1614 work->old_state = CMA_ROUTE_QUERY;
1615 work->new_state = CMA_ADDR_RESOLVED;
1616 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
1617 work->event.status = status;
1620 queue_work(cma_wq, &work->work);
1623 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms,
1624 struct cma_work *work)
1626 struct rdma_addr *addr = &id_priv->id.route.addr;
1627 struct ib_sa_path_rec path_rec;
1628 ib_sa_comp_mask comp_mask;
1629 struct sockaddr_in6 *sin6;
1631 memset(&path_rec, 0, sizeof path_rec);
1632 rdma_addr_get_sgid(&addr->dev_addr, &path_rec.sgid);
1633 rdma_addr_get_dgid(&addr->dev_addr, &path_rec.dgid);
1634 path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(&addr->dev_addr));
1635 path_rec.numb_path = 1;
1636 path_rec.reversible = 1;
1637 path_rec.service_id = cma_get_service_id(id_priv->id.ps,
1638 (struct sockaddr *) &addr->dst_addr);
1640 comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
1641 IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
1642 IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
1644 if (addr->src_addr.ss_family == AF_INET) {
1645 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
1646 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
1648 sin6 = (struct sockaddr_in6 *) &addr->src_addr;
1649 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
1650 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
1654 path_rec.mtu_selector = IB_SA_LT;
1655 path_rec.mtu = IB_MTU_2048;
1658 id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
1659 id_priv->id.port_num, &path_rec,
1660 comp_mask, timeout_ms,
1661 GFP_KERNEL, cma_query_handler,
1662 work, &id_priv->query);
1664 return (id_priv->query_id < 0) ? id_priv->query_id : 0;
1667 static void cma_work_handler(struct work_struct *_work)
1669 struct cma_work *work = container_of(_work, struct cma_work, work);
1670 struct rdma_id_private *id_priv = work->id;
1673 mutex_lock(&id_priv->handler_mutex);
1674 if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
1677 if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1678 cma_exch(id_priv, CMA_DESTROYING);
1682 mutex_unlock(&id_priv->handler_mutex);
1683 cma_deref_id(id_priv);
1685 rdma_destroy_id(&id_priv->id);
1689 static void cma_ndev_work_handler(struct work_struct *_work)
1691 struct cma_ndev_work *work = container_of(_work, struct cma_ndev_work, work);
1692 struct rdma_id_private *id_priv = work->id;
1695 mutex_lock(&id_priv->handler_mutex);
1696 if (id_priv->state == CMA_DESTROYING ||
1697 id_priv->state == CMA_DEVICE_REMOVAL)
1700 if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1701 cma_exch(id_priv, CMA_DESTROYING);
1706 mutex_unlock(&id_priv->handler_mutex);
1707 cma_deref_id(id_priv);
1709 rdma_destroy_id(&id_priv->id);
1713 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms)
1715 struct rdma_route *route = &id_priv->id.route;
1716 struct cma_work *work;
1719 work = kzalloc(sizeof *work, GFP_KERNEL);
1724 INIT_WORK(&work->work, cma_work_handler);
1725 work->old_state = CMA_ROUTE_QUERY;
1726 work->new_state = CMA_ROUTE_RESOLVED;
1727 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1729 route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
1730 if (!route->path_rec) {
1735 ret = cma_query_ib_route(id_priv, timeout_ms, work);
1741 kfree(route->path_rec);
1742 route->path_rec = NULL;
1748 int rdma_set_ib_paths(struct rdma_cm_id *id,
1749 struct ib_sa_path_rec *path_rec, int num_paths)
1751 struct rdma_id_private *id_priv;
1754 id_priv = container_of(id, struct rdma_id_private, id);
1755 if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ROUTE_RESOLVED))
1758 id->route.path_rec = kmalloc(sizeof *path_rec * num_paths, GFP_KERNEL);
1759 if (!id->route.path_rec) {
1764 memcpy(id->route.path_rec, path_rec, sizeof *path_rec * num_paths);
1767 cma_comp_exch(id_priv, CMA_ROUTE_RESOLVED, CMA_ADDR_RESOLVED);
1770 EXPORT_SYMBOL(rdma_set_ib_paths);
1772 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms)
1774 struct cma_work *work;
1776 work = kzalloc(sizeof *work, GFP_KERNEL);
1781 INIT_WORK(&work->work, cma_work_handler);
1782 work->old_state = CMA_ROUTE_QUERY;
1783 work->new_state = CMA_ROUTE_RESOLVED;
1784 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1785 queue_work(cma_wq, &work->work);
1789 static u8 tos_to_sl(u8 tos)
1791 return def_prec2sl & 7;
1794 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
1796 struct rdma_route *route = &id_priv->id.route;
1797 struct rdma_addr *addr = &route->addr;
1798 struct cma_work *work;
1800 struct sockaddr_in *src_addr = (struct sockaddr_in *)&route->addr.src_addr;
1801 struct sockaddr_in *dst_addr = (struct sockaddr_in *)&route->addr.dst_addr;
1802 struct net_device *ndev = NULL;
1805 if (src_addr->sin_family != dst_addr->sin_family)
1808 work = kzalloc(sizeof *work, GFP_KERNEL);
1813 INIT_WORK(&work->work, cma_work_handler);
1815 route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
1816 if (!route->path_rec) {
1821 route->num_paths = 1;
1823 if (addr->dev_addr.bound_dev_if)
1824 ndev = dev_get_by_index(&init_net, addr->dev_addr.bound_dev_if);
1830 vid = rdma_vlan_dev_vlan_id(ndev);
1832 iboe_mac_vlan_to_ll(&route->path_rec->sgid, addr->dev_addr.src_dev_addr, vid);
1833 iboe_mac_vlan_to_ll(&route->path_rec->dgid, addr->dev_addr.dst_dev_addr, vid);
1835 route->path_rec->hop_limit = 1;
1836 route->path_rec->reversible = 1;
1837 route->path_rec->pkey = cpu_to_be16(0xffff);
1838 route->path_rec->mtu_selector = IB_SA_EQ;
1839 route->path_rec->sl = tos_to_sl(id_priv->tos);
1842 route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
1844 route->path_rec->mtu = iboe_get_mtu(ndev->if_mtu);
1846 route->path_rec->rate_selector = IB_SA_EQ;
1847 route->path_rec->rate = iboe_get_rate(ndev);
1849 route->path_rec->packet_life_time_selector = IB_SA_EQ;
1850 route->path_rec->packet_life_time = IBOE_PACKET_LIFETIME;
1851 if (!route->path_rec->mtu) {
1856 work->old_state = CMA_ROUTE_QUERY;
1857 work->new_state = CMA_ROUTE_RESOLVED;
1858 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1859 work->event.status = 0;
1861 queue_work(cma_wq, &work->work);
1866 kfree(route->path_rec);
1867 route->path_rec = NULL;
1873 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
1875 struct rdma_id_private *id_priv;
1878 id_priv = container_of(id, struct rdma_id_private, id);
1879 if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ROUTE_QUERY))
1882 atomic_inc(&id_priv->refcount);
1883 switch (rdma_node_get_transport(id->device->node_type)) {
1884 case RDMA_TRANSPORT_IB:
1885 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
1886 case IB_LINK_LAYER_INFINIBAND:
1887 ret = cma_resolve_ib_route(id_priv, timeout_ms);
1889 case IB_LINK_LAYER_ETHERNET:
1890 ret = cma_resolve_iboe_route(id_priv);
1896 case RDMA_TRANSPORT_IWARP:
1897 ret = cma_resolve_iw_route(id_priv, timeout_ms);
1908 cma_comp_exch(id_priv, CMA_ROUTE_QUERY, CMA_ADDR_RESOLVED);
1909 cma_deref_id(id_priv);
1912 EXPORT_SYMBOL(rdma_resolve_route);
1914 static int cma_bind_loopback(struct rdma_id_private *id_priv)
1916 struct cma_device *cma_dev;
1917 struct ib_port_attr port_attr;
1924 if (list_empty(&dev_list)) {
1928 list_for_each_entry(cma_dev, &dev_list, list)
1929 for (p = 1; p <= cma_dev->device->phys_port_cnt; ++p)
1930 if (!ib_query_port(cma_dev->device, p, &port_attr) &&
1931 port_attr.state == IB_PORT_ACTIVE)
1935 cma_dev = list_entry(dev_list.next, struct cma_device, list);
1938 ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid);
1942 ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
1946 id_priv->id.route.addr.dev_addr.dev_type =
1947 (rdma_port_get_link_layer(cma_dev->device, p) == IB_LINK_LAYER_INFINIBAND) ?
1948 ARPHRD_INFINIBAND : ARPHRD_ETHER;
1950 rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
1951 ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
1952 id_priv->id.port_num = p;
1953 cma_attach_to_dev(id_priv, cma_dev);
1955 mutex_unlock(&lock);
1959 static void addr_handler(int status, struct sockaddr *src_addr,
1960 struct rdma_dev_addr *dev_addr, void *context)
1962 struct rdma_id_private *id_priv = context;
1963 struct rdma_cm_event event;
1965 memset(&event, 0, sizeof event);
1966 mutex_lock(&id_priv->handler_mutex);
1969 * Grab mutex to block rdma_destroy_id() from removing the device while
1970 * we're trying to acquire it.
1973 if (!cma_comp_exch(id_priv, CMA_ADDR_QUERY, CMA_ADDR_RESOLVED)) {
1974 mutex_unlock(&lock);
1978 if (!status && !id_priv->cma_dev)
1979 status = cma_acquire_dev(id_priv);
1980 mutex_unlock(&lock);
1983 if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ADDR_BOUND))
1985 event.event = RDMA_CM_EVENT_ADDR_ERROR;
1986 event.status = status;
1988 memcpy(&id_priv->id.route.addr.src_addr, src_addr,
1989 ip_addr_size(src_addr));
1990 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
1993 if (id_priv->id.event_handler(&id_priv->id, &event)) {
1994 cma_exch(id_priv, CMA_DESTROYING);
1995 mutex_unlock(&id_priv->handler_mutex);
1996 cma_deref_id(id_priv);
1997 rdma_destroy_id(&id_priv->id);
2001 mutex_unlock(&id_priv->handler_mutex);
2002 cma_deref_id(id_priv);
2005 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
2007 struct cma_work *work;
2008 struct sockaddr *src, *dst;
2012 work = kzalloc(sizeof *work, GFP_KERNEL);
2016 if (!id_priv->cma_dev) {
2017 ret = cma_bind_loopback(id_priv);
2022 rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2023 rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
2025 src = (struct sockaddr *) &id_priv->id.route.addr.src_addr;
2026 if (cma_zero_addr(src)) {
2027 dst = (struct sockaddr *) &id_priv->id.route.addr.dst_addr;
2028 if ((src->sa_family = dst->sa_family) == AF_INET) {
2029 ((struct sockaddr_in *) src)->sin_addr.s_addr =
2030 ((struct sockaddr_in *) dst)->sin_addr.s_addr;
2032 ipv6_addr_copy(&((struct sockaddr_in6 *) src)->sin6_addr,
2033 &((struct sockaddr_in6 *) dst)->sin6_addr);
2038 INIT_WORK(&work->work, cma_work_handler);
2039 work->old_state = CMA_ADDR_QUERY;
2040 work->new_state = CMA_ADDR_RESOLVED;
2041 work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2042 queue_work(cma_wq, &work->work);
2049 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2050 struct sockaddr *dst_addr)
2052 if (!src_addr || !src_addr->sa_family) {
2053 src_addr = (struct sockaddr *) &id->route.addr.src_addr;
2054 if ((src_addr->sa_family = dst_addr->sa_family) == AF_INET6) {
2055 ((struct sockaddr_in6 *) src_addr)->sin6_scope_id =
2056 ((struct sockaddr_in6 *) dst_addr)->sin6_scope_id;
2059 if (!cma_any_addr(src_addr))
2060 return rdma_bind_addr(id, src_addr);
2062 struct sockaddr_in addr_in;
2064 memset(&addr_in, 0, sizeof addr_in);
2065 addr_in.sin_family = dst_addr->sa_family;
2066 addr_in.sin_len = sizeof addr_in;
2067 return rdma_bind_addr(id, (struct sockaddr *) &addr_in);
2071 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2072 struct sockaddr *dst_addr, int timeout_ms)
2074 struct rdma_id_private *id_priv;
2077 id_priv = container_of(id, struct rdma_id_private, id);
2078 if (id_priv->state == CMA_IDLE) {
2079 ret = cma_bind_addr(id, src_addr, dst_addr);
2084 if (!cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_ADDR_QUERY))
2087 atomic_inc(&id_priv->refcount);
2088 memcpy(&id->route.addr.dst_addr, dst_addr, ip_addr_size(dst_addr));
2089 if (cma_any_addr(dst_addr))
2090 ret = cma_resolve_loopback(id_priv);
2092 ret = rdma_resolve_ip(&addr_client, (struct sockaddr *) &id->route.addr.src_addr,
2093 dst_addr, &id->route.addr.dev_addr,
2094 timeout_ms, addr_handler, id_priv);
2100 cma_comp_exch(id_priv, CMA_ADDR_QUERY, CMA_ADDR_BOUND);
2101 cma_deref_id(id_priv);
2104 EXPORT_SYMBOL(rdma_resolve_addr);
2106 static void cma_bind_port(struct rdma_bind_list *bind_list,
2107 struct rdma_id_private *id_priv)
2109 struct sockaddr_in *sin;
2111 sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
2112 sin->sin_port = htons(bind_list->port);
2113 id_priv->bind_list = bind_list;
2114 hlist_add_head(&id_priv->node, &bind_list->owners);
2117 static int cma_alloc_port(struct idr *ps, struct rdma_id_private *id_priv,
2118 unsigned short snum)
2120 struct rdma_bind_list *bind_list;
2123 bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
2128 ret = idr_get_new_above(ps, bind_list, snum, &port);
2129 } while ((ret == -EAGAIN) && idr_pre_get(ps, GFP_KERNEL));
2135 ret = -EADDRNOTAVAIL;
2140 bind_list->port = (unsigned short) port;
2141 cma_bind_port(bind_list, id_priv);
2144 idr_remove(ps, port);
2150 static int cma_alloc_any_port(struct idr *ps, struct rdma_id_private *id_priv)
2153 struct rdma_bind_list *bind_list;
2154 int port, ret, low, high;
2156 bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
2161 /* FIXME: add proper port randomization per like inet_csk_get_port */
2163 ret = idr_get_new_above(ps, bind_list, next_port, &port);
2164 } while ((ret == -EAGAIN) && idr_pre_get(ps, GFP_KERNEL));
2169 inet_get_local_port_range(&low, &high);
2171 if (next_port != low) {
2172 idr_remove(ps, port);
2176 ret = -EADDRNOTAVAIL;
2183 next_port = port + 1;
2186 bind_list->port = (unsigned short) port;
2187 cma_bind_port(bind_list, id_priv);
2190 idr_remove(ps, port);
2199 static int cma_use_port(struct idr *ps, struct rdma_id_private *id_priv)
2201 struct rdma_id_private *cur_id;
2202 struct sockaddr_in *sin, *cur_sin;
2203 struct rdma_bind_list *bind_list;
2204 struct hlist_node *node;
2205 unsigned short snum;
2207 sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
2208 snum = ntohs(sin->sin_port);
2210 if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
2214 bind_list = idr_find(ps, snum);
2216 return cma_alloc_port(ps, id_priv, snum);
2219 * We don't support binding to any address if anyone is bound to
2220 * a specific address on the same port.
2222 if (cma_any_addr((struct sockaddr *) &id_priv->id.route.addr.src_addr))
2223 return -EADDRNOTAVAIL;
2225 hlist_for_each_entry(cur_id, node, &bind_list->owners, node) {
2226 if (cma_any_addr((struct sockaddr *) &cur_id->id.route.addr.src_addr))
2227 return -EADDRNOTAVAIL;
2229 cur_sin = (struct sockaddr_in *) &cur_id->id.route.addr.src_addr;
2230 if (sin->sin_addr.s_addr == cur_sin->sin_addr.s_addr)
2234 cma_bind_port(bind_list, id_priv);
2238 static int cma_get_tcp_port(struct rdma_id_private *id_priv)
2242 struct socket *sock;
2244 ret = sock_create_kern(AF_INET, SOCK_STREAM, IPPROTO_TCP, &sock);
2248 ret = sock->ops->bind(sock,
2249 (struct sockaddr *) &id_priv->id.route.addr.src_addr,
2250 ip_addr_size((struct sockaddr *) &id_priv->id.route.addr.src_addr));
2253 (struct sockaddr *)&id_priv->id.route.addr.src_addr,
2261 size = ip_addr_size((struct sockaddr *) &id_priv->id.route.addr.src_addr);
2262 ret = sock_getname(sock,
2263 (struct sockaddr *) &id_priv->id.route.addr.src_addr,
2270 id_priv->sock = sock;
2274 static int cma_get_port(struct rdma_id_private *id_priv)
2279 switch (id_priv->id.ps) {
2285 if (unify_tcp_port_space) {
2286 ret = cma_get_tcp_port(id_priv);
2298 return -EPROTONOSUPPORT;
2302 if (cma_any_port((struct sockaddr *) &id_priv->id.route.addr.src_addr))
2303 ret = cma_alloc_any_port(ps, id_priv);
2305 ret = cma_use_port(ps, id_priv);
2306 mutex_unlock(&lock);
2311 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
2312 struct sockaddr *addr)
2315 struct sockaddr_in6 *sin6;
2317 if (addr->sa_family != AF_INET6)
2320 sin6 = (struct sockaddr_in6 *) addr;
2322 if ((ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL) &&
2324 if (IN6_IS_SCOPE_LINKLOCAL(&sin6->sin6_addr) &&
2326 !sin6->sin6_scope_id)
2329 dev_addr->bound_dev_if = sin6->sin6_scope_id;
2334 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
2336 struct rdma_id_private *id_priv;
2339 if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6)
2340 return -EAFNOSUPPORT;
2342 id_priv = container_of(id, struct rdma_id_private, id);
2343 if (!cma_comp_exch(id_priv, CMA_IDLE, CMA_ADDR_BOUND))
2346 ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
2350 if (!cma_any_addr(addr)) {
2351 ret = rdma_translate_ip(addr, &id->route.addr.dev_addr);
2356 ret = cma_acquire_dev(id_priv);
2357 mutex_unlock(&lock);
2362 memcpy(&id->route.addr.src_addr, addr, ip_addr_size(addr));
2363 ret = cma_get_port(id_priv);
2369 if (id_priv->cma_dev) {
2371 cma_detach_from_dev(id_priv);
2372 mutex_unlock(&lock);
2375 cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_IDLE);
2378 EXPORT_SYMBOL(rdma_bind_addr);
2380 static int cma_format_hdr(void *hdr, enum rdma_port_space ps,
2381 struct rdma_route *route)
2383 struct cma_hdr *cma_hdr;
2384 struct sdp_hh *sdp_hdr;
2386 if (route->addr.src_addr.ss_family == AF_INET) {
2387 struct sockaddr_in *src4, *dst4;
2389 src4 = (struct sockaddr_in *) &route->addr.src_addr;
2390 dst4 = (struct sockaddr_in *) &route->addr.dst_addr;
2395 if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION)
2397 sdp_set_ip_ver(sdp_hdr, 4);
2398 sdp_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2399 sdp_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2400 sdp_hdr->port = src4->sin_port;
2404 cma_hdr->cma_version = CMA_VERSION;
2405 cma_set_ip_ver(cma_hdr, 4);
2406 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2407 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2408 cma_hdr->port = src4->sin_port;
2412 struct sockaddr_in6 *src6, *dst6;
2414 src6 = (struct sockaddr_in6 *) &route->addr.src_addr;
2415 dst6 = (struct sockaddr_in6 *) &route->addr.dst_addr;
2420 if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION)
2422 sdp_set_ip_ver(sdp_hdr, 6);
2423 sdp_hdr->src_addr.ip6 = src6->sin6_addr;
2424 sdp_hdr->dst_addr.ip6 = dst6->sin6_addr;
2425 sdp_hdr->port = src6->sin6_port;
2429 cma_hdr->cma_version = CMA_VERSION;
2430 cma_set_ip_ver(cma_hdr, 6);
2431 cma_hdr->src_addr.ip6 = src6->sin6_addr;
2432 cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
2433 cma_hdr->port = src6->sin6_port;
2440 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
2441 struct ib_cm_event *ib_event)
2443 struct rdma_id_private *id_priv = cm_id->context;
2444 struct rdma_cm_event event;
2445 struct ib_cm_sidr_rep_event_param *rep = &ib_event->param.sidr_rep_rcvd;
2448 if (cma_disable_callback(id_priv, CMA_CONNECT))
2451 memset(&event, 0, sizeof event);
2452 switch (ib_event->event) {
2453 case IB_CM_SIDR_REQ_ERROR:
2454 event.event = RDMA_CM_EVENT_UNREACHABLE;
2455 event.status = -ETIMEDOUT;
2457 case IB_CM_SIDR_REP_RECEIVED:
2458 event.param.ud.private_data = ib_event->private_data;
2459 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
2460 if (rep->status != IB_SIDR_SUCCESS) {
2461 event.event = RDMA_CM_EVENT_UNREACHABLE;
2462 event.status = ib_event->param.sidr_rep_rcvd.status;
2465 ret = cma_set_qkey(id_priv);
2467 event.event = RDMA_CM_EVENT_ADDR_ERROR;
2468 event.status = -EINVAL;
2471 if (id_priv->qkey != rep->qkey) {
2472 event.event = RDMA_CM_EVENT_UNREACHABLE;
2473 event.status = -EINVAL;
2476 ib_init_ah_from_path(id_priv->id.device, id_priv->id.port_num,
2477 id_priv->id.route.path_rec,
2478 &event.param.ud.ah_attr);
2479 event.param.ud.qp_num = rep->qpn;
2480 event.param.ud.qkey = rep->qkey;
2481 event.event = RDMA_CM_EVENT_ESTABLISHED;
2485 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
2490 ret = id_priv->id.event_handler(&id_priv->id, &event);
2492 /* Destroy the CM ID by returning a non-zero value. */
2493 id_priv->cm_id.ib = NULL;
2494 cma_exch(id_priv, CMA_DESTROYING);
2495 mutex_unlock(&id_priv->handler_mutex);
2496 rdma_destroy_id(&id_priv->id);
2500 mutex_unlock(&id_priv->handler_mutex);
2504 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
2505 struct rdma_conn_param *conn_param)
2507 struct ib_cm_sidr_req_param req;
2508 struct rdma_route *route;
2511 req.private_data_len = sizeof(struct cma_hdr) +
2512 conn_param->private_data_len;
2513 req.private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2514 if (!req.private_data)
2517 if (conn_param->private_data && conn_param->private_data_len)
2518 memcpy((void *) req.private_data + sizeof(struct cma_hdr),
2519 conn_param->private_data, conn_param->private_data_len);
2521 route = &id_priv->id.route;
2522 ret = cma_format_hdr((void *) req.private_data, id_priv->id.ps, route);
2526 id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device,
2527 cma_sidr_rep_handler, id_priv);
2528 if (IS_ERR(id_priv->cm_id.ib)) {
2529 ret = PTR_ERR(id_priv->cm_id.ib);
2533 req.path = route->path_rec;
2534 req.service_id = cma_get_service_id(id_priv->id.ps,
2535 (struct sockaddr *) &route->addr.dst_addr);
2536 req.timeout_ms = 1 << (cma_response_timeout - 8);
2537 req.max_cm_retries = CMA_MAX_CM_RETRIES;
2539 ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
2541 ib_destroy_cm_id(id_priv->cm_id.ib);
2542 id_priv->cm_id.ib = NULL;
2545 kfree(req.private_data);
2549 static int cma_connect_ib(struct rdma_id_private *id_priv,
2550 struct rdma_conn_param *conn_param)
2552 struct ib_cm_req_param req;
2553 struct rdma_route *route;
2557 memset(&req, 0, sizeof req);
2558 offset = cma_user_data_offset(id_priv->id.ps);
2559 req.private_data_len = offset + conn_param->private_data_len;
2560 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2564 if (conn_param->private_data && conn_param->private_data_len)
2565 memcpy(private_data + offset, conn_param->private_data,
2566 conn_param->private_data_len);
2568 id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, cma_ib_handler,
2570 if (IS_ERR(id_priv->cm_id.ib)) {
2571 ret = PTR_ERR(id_priv->cm_id.ib);
2575 route = &id_priv->id.route;
2576 ret = cma_format_hdr(private_data, id_priv->id.ps, route);
2579 req.private_data = private_data;
2581 req.primary_path = &route->path_rec[0];
2582 if (route->num_paths == 2)
2583 req.alternate_path = &route->path_rec[1];
2585 req.service_id = cma_get_service_id(id_priv->id.ps,
2586 (struct sockaddr *) &route->addr.dst_addr);
2587 req.qp_num = id_priv->qp_num;
2588 req.qp_type = IB_QPT_RC;
2589 req.starting_psn = id_priv->seq_num;
2590 req.responder_resources = conn_param->responder_resources;
2591 req.initiator_depth = conn_param->initiator_depth;
2592 req.flow_control = conn_param->flow_control;
2593 req.retry_count = conn_param->retry_count;
2594 req.rnr_retry_count = conn_param->rnr_retry_count;
2595 req.remote_cm_response_timeout = cma_response_timeout;
2596 req.local_cm_response_timeout = cma_response_timeout;
2597 req.max_cm_retries = CMA_MAX_CM_RETRIES;
2598 req.srq = id_priv->srq ? 1 : 0;
2600 ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
2602 if (ret && !IS_ERR(id_priv->cm_id.ib)) {
2603 ib_destroy_cm_id(id_priv->cm_id.ib);
2604 id_priv->cm_id.ib = NULL;
2607 kfree(private_data);
2611 static int cma_connect_iw(struct rdma_id_private *id_priv,
2612 struct rdma_conn_param *conn_param)
2614 struct iw_cm_id *cm_id;
2615 struct sockaddr_in* sin;
2617 struct iw_cm_conn_param iw_param;
2619 cm_id = iw_create_cm_id(id_priv->id.device, id_priv->sock,
2620 cma_iw_handler, id_priv);
2621 if (IS_ERR(cm_id)) {
2622 ret = PTR_ERR(cm_id);
2626 id_priv->cm_id.iw = cm_id;
2628 sin = (struct sockaddr_in*) &id_priv->id.route.addr.src_addr;
2629 cm_id->local_addr = *sin;
2631 sin = (struct sockaddr_in*) &id_priv->id.route.addr.dst_addr;
2632 cm_id->remote_addr = *sin;
2634 ret = cma_modify_qp_rtr(id_priv, conn_param);
2638 iw_param.ord = conn_param->initiator_depth;
2639 iw_param.ird = conn_param->responder_resources;
2640 iw_param.private_data = conn_param->private_data;
2641 iw_param.private_data_len = conn_param->private_data_len;
2643 iw_param.qpn = id_priv->qp_num;
2645 iw_param.qpn = conn_param->qp_num;
2646 ret = iw_cm_connect(cm_id, &iw_param);
2648 if (ret && !IS_ERR(cm_id)) {
2649 iw_destroy_cm_id(cm_id);
2650 id_priv->cm_id.iw = NULL;
2655 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2657 struct rdma_id_private *id_priv;
2660 id_priv = container_of(id, struct rdma_id_private, id);
2661 if (!cma_comp_exch(id_priv, CMA_ROUTE_RESOLVED, CMA_CONNECT))
2665 id_priv->qp_num = conn_param->qp_num;
2666 id_priv->srq = conn_param->srq;
2669 switch (rdma_node_get_transport(id->device->node_type)) {
2670 case RDMA_TRANSPORT_IB:
2671 if (cma_is_ud_ps(id->ps))
2672 ret = cma_resolve_ib_udp(id_priv, conn_param);
2674 ret = cma_connect_ib(id_priv, conn_param);
2676 case RDMA_TRANSPORT_IWARP:
2677 ret = cma_connect_iw(id_priv, conn_param);
2688 cma_comp_exch(id_priv, CMA_CONNECT, CMA_ROUTE_RESOLVED);
2691 EXPORT_SYMBOL(rdma_connect);
2693 static int cma_accept_ib(struct rdma_id_private *id_priv,
2694 struct rdma_conn_param *conn_param)
2696 struct ib_cm_rep_param rep;
2699 ret = cma_modify_qp_rtr(id_priv, conn_param);
2703 ret = cma_modify_qp_rts(id_priv, conn_param);
2707 memset(&rep, 0, sizeof rep);
2708 rep.qp_num = id_priv->qp_num;
2709 rep.starting_psn = id_priv->seq_num;
2710 rep.private_data = conn_param->private_data;
2711 rep.private_data_len = conn_param->private_data_len;
2712 rep.responder_resources = conn_param->responder_resources;
2713 rep.initiator_depth = conn_param->initiator_depth;
2714 rep.failover_accepted = 0;
2715 rep.flow_control = conn_param->flow_control;
2716 rep.rnr_retry_count = conn_param->rnr_retry_count;
2717 rep.srq = id_priv->srq ? 1 : 0;
2719 ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
2724 static int cma_accept_iw(struct rdma_id_private *id_priv,
2725 struct rdma_conn_param *conn_param)
2727 struct iw_cm_conn_param iw_param;
2730 ret = cma_modify_qp_rtr(id_priv, conn_param);
2734 iw_param.ord = conn_param->initiator_depth;
2735 iw_param.ird = conn_param->responder_resources;
2736 iw_param.private_data = conn_param->private_data;
2737 iw_param.private_data_len = conn_param->private_data_len;
2738 if (id_priv->id.qp) {
2739 iw_param.qpn = id_priv->qp_num;
2741 iw_param.qpn = conn_param->qp_num;
2743 return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
2746 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
2747 enum ib_cm_sidr_status status,
2748 const void *private_data, int private_data_len)
2750 struct ib_cm_sidr_rep_param rep;
2753 memset(&rep, 0, sizeof rep);
2754 rep.status = status;
2755 if (status == IB_SIDR_SUCCESS) {
2756 ret = cma_set_qkey(id_priv);
2759 rep.qp_num = id_priv->qp_num;
2760 rep.qkey = id_priv->qkey;
2762 rep.private_data = private_data;
2763 rep.private_data_len = private_data_len;
2765 return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
2768 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2770 struct rdma_id_private *id_priv;
2773 id_priv = container_of(id, struct rdma_id_private, id);
2774 if (!cma_comp(id_priv, CMA_CONNECT))
2777 if (!id->qp && conn_param) {
2778 id_priv->qp_num = conn_param->qp_num;
2779 id_priv->srq = conn_param->srq;
2782 switch (rdma_node_get_transport(id->device->node_type)) {
2783 case RDMA_TRANSPORT_IB:
2784 if (cma_is_ud_ps(id->ps))
2785 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
2786 conn_param->private_data,
2787 conn_param->private_data_len);
2788 else if (conn_param)
2789 ret = cma_accept_ib(id_priv, conn_param);
2791 ret = cma_rep_recv(id_priv);
2793 case RDMA_TRANSPORT_IWARP:
2794 ret = cma_accept_iw(id_priv, conn_param);
2806 cma_modify_qp_err(id_priv);
2807 rdma_reject(id, NULL, 0);
2810 EXPORT_SYMBOL(rdma_accept);
2812 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
2814 struct rdma_id_private *id_priv;
2817 id_priv = container_of(id, struct rdma_id_private, id);
2818 if (!cma_has_cm_dev(id_priv))
2821 switch (id->device->node_type) {
2822 case RDMA_NODE_IB_CA:
2823 ret = ib_cm_notify(id_priv->cm_id.ib, event);
2831 EXPORT_SYMBOL(rdma_notify);
2833 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
2834 u8 private_data_len)
2836 struct rdma_id_private *id_priv;
2839 id_priv = container_of(id, struct rdma_id_private, id);
2840 if (!cma_has_cm_dev(id_priv))
2843 switch (rdma_node_get_transport(id->device->node_type)) {
2844 case RDMA_TRANSPORT_IB:
2845 if (cma_is_ud_ps(id->ps))
2846 ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT,
2847 private_data, private_data_len);
2849 ret = ib_send_cm_rej(id_priv->cm_id.ib,
2850 IB_CM_REJ_CONSUMER_DEFINED, NULL,
2851 0, private_data, private_data_len);
2853 case RDMA_TRANSPORT_IWARP:
2854 ret = iw_cm_reject(id_priv->cm_id.iw,
2855 private_data, private_data_len);
2863 EXPORT_SYMBOL(rdma_reject);
2865 int rdma_disconnect(struct rdma_cm_id *id)
2867 struct rdma_id_private *id_priv;
2870 id_priv = container_of(id, struct rdma_id_private, id);
2871 if (!cma_has_cm_dev(id_priv))
2874 switch (rdma_node_get_transport(id->device->node_type)) {
2875 case RDMA_TRANSPORT_IB:
2876 ret = cma_modify_qp_err(id_priv);
2879 /* Initiate or respond to a disconnect. */
2880 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
2881 ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
2883 case RDMA_TRANSPORT_IWARP:
2884 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
2893 EXPORT_SYMBOL(rdma_disconnect);
2895 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
2897 struct rdma_id_private *id_priv;
2898 struct cma_multicast *mc = multicast->context;
2899 struct rdma_cm_event event;
2902 id_priv = mc->id_priv;
2903 if (cma_disable_callback(id_priv, CMA_ADDR_BOUND) &&
2904 cma_disable_callback(id_priv, CMA_ADDR_RESOLVED))
2907 mutex_lock(&id_priv->qp_mutex);
2908 if (!status && id_priv->id.qp)
2909 status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid,
2910 multicast->rec.mlid);
2911 mutex_unlock(&id_priv->qp_mutex);
2913 memset(&event, 0, sizeof event);
2914 event.status = status;
2915 event.param.ud.private_data = mc->context;
2917 event.event = RDMA_CM_EVENT_MULTICAST_JOIN;
2918 ib_init_ah_from_mcmember(id_priv->id.device,
2919 id_priv->id.port_num, &multicast->rec,
2920 &event.param.ud.ah_attr);
2921 event.param.ud.qp_num = 0xFFFFFF;
2922 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
2924 event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
2926 ret = id_priv->id.event_handler(&id_priv->id, &event);
2928 cma_exch(id_priv, CMA_DESTROYING);
2929 mutex_unlock(&id_priv->handler_mutex);
2930 rdma_destroy_id(&id_priv->id);
2934 mutex_unlock(&id_priv->handler_mutex);
2938 static void cma_set_mgid(struct rdma_id_private *id_priv,
2939 struct sockaddr *addr, union ib_gid *mgid)
2941 #if defined(INET) || defined(INET6)
2942 unsigned char mc_map[MAX_ADDR_LEN];
2943 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2946 struct sockaddr_in *sin = (struct sockaddr_in *) addr;
2949 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
2952 if (cma_any_addr(addr)) {
2953 memset(mgid, 0, sizeof *mgid);
2955 } else if ((addr->sa_family == AF_INET6) &&
2956 ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
2958 /* IPv6 address is an SA assigned MGID. */
2959 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
2960 } else if ((addr->sa_family == AF_INET6)) {
2961 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
2962 if (id_priv->id.ps == RDMA_PS_UDP)
2963 mc_map[7] = 0x01; /* Use RDMA CM signature */
2964 *mgid = *(union ib_gid *) (mc_map + 4);
2968 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
2969 if (id_priv->id.ps == RDMA_PS_UDP)
2970 mc_map[7] = 0x01; /* Use RDMA CM signature */
2971 *mgid = *(union ib_gid *) (mc_map + 4);
2976 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
2977 struct cma_multicast *mc)
2979 struct ib_sa_mcmember_rec rec;
2980 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2981 ib_sa_comp_mask comp_mask;
2984 ib_addr_get_mgid(dev_addr, &rec.mgid);
2985 ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
2990 cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
2991 if (id_priv->id.ps == RDMA_PS_UDP)
2992 rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
2993 rdma_addr_get_sgid(dev_addr, &rec.port_gid);
2994 rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
2997 comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
2998 IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
2999 IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
3000 IB_SA_MCMEMBER_REC_FLOW_LABEL |
3001 IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
3003 if (id_priv->id.ps == RDMA_PS_IPOIB)
3004 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
3005 IB_SA_MCMEMBER_REC_RATE_SELECTOR;
3007 mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device,
3008 id_priv->id.port_num, &rec,
3009 comp_mask, GFP_KERNEL,
3010 cma_ib_mc_handler, mc);
3011 if (IS_ERR(mc->multicast.ib))
3012 return PTR_ERR(mc->multicast.ib);
3018 static void iboe_mcast_work_handler(struct work_struct *work)
3020 struct iboe_mcast_work *mw = container_of(work, struct iboe_mcast_work, work);
3021 struct cma_multicast *mc = mw->mc;
3022 struct ib_sa_multicast *m = mc->multicast.ib;
3024 mc->multicast.ib->context = mc;
3025 cma_ib_mc_handler(0, m);
3026 kref_put(&mc->mcref, release_mc);
3030 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid)
3032 struct sockaddr_in *sin = (struct sockaddr_in *)addr;
3033 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
3035 if (cma_any_addr(addr)) {
3036 memset(mgid, 0, sizeof *mgid);
3037 } else if (addr->sa_family == AF_INET6)
3038 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3040 mgid->raw[0] = 0xff;
3041 mgid->raw[1] = 0x0e;
3050 mgid->raw[10] = 0xff;
3051 mgid->raw[11] = 0xff;
3052 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
3056 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
3057 struct cma_multicast *mc)
3059 struct iboe_mcast_work *work;
3060 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3062 struct sockaddr *addr = (struct sockaddr *)&mc->addr;
3063 struct net_device *ndev = NULL;
3065 if (cma_zero_addr((struct sockaddr *)&mc->addr))
3068 work = kzalloc(sizeof *work, GFP_KERNEL);
3072 mc->multicast.ib = kzalloc(sizeof(struct ib_sa_multicast), GFP_KERNEL);
3073 if (!mc->multicast.ib) {
3078 cma_iboe_set_mgid(addr, &mc->multicast.ib->rec.mgid);
3080 mc->multicast.ib->rec.pkey = cpu_to_be16(0xffff);
3081 if (id_priv->id.ps == RDMA_PS_UDP)
3082 mc->multicast.ib->rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
3084 if (dev_addr->bound_dev_if)
3085 ndev = dev_get_by_index(&init_net, dev_addr->bound_dev_if);
3091 mc->multicast.ib->rec.rate = iboe_get_rate(ndev);
3092 mc->multicast.ib->rec.hop_limit = 1;
3094 mc->multicast.ib->rec.mtu = iboe_get_mtu(ndev->mtu);
3096 mc->multicast.ib->rec.mtu = iboe_get_mtu(ndev->if_mtu);
3099 if (!mc->multicast.ib->rec.mtu) {
3103 iboe_addr_get_sgid(dev_addr, &mc->multicast.ib->rec.port_gid);
3106 INIT_WORK(&work->work, iboe_mcast_work_handler);
3107 kref_get(&mc->mcref);
3108 queue_work(cma_wq, &work->work);
3113 kfree(mc->multicast.ib);
3119 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
3122 struct rdma_id_private *id_priv;
3123 struct cma_multicast *mc;
3126 id_priv = container_of(id, struct rdma_id_private, id);
3127 if (!cma_comp(id_priv, CMA_ADDR_BOUND) &&
3128 !cma_comp(id_priv, CMA_ADDR_RESOLVED))
3131 mc = kmalloc(sizeof *mc, GFP_KERNEL);
3135 memcpy(&mc->addr, addr, ip_addr_size(addr));
3136 mc->context = context;
3137 mc->id_priv = id_priv;
3139 spin_lock(&id_priv->lock);
3140 list_add(&mc->list, &id_priv->mc_list);
3141 spin_unlock(&id_priv->lock);
3143 switch (rdma_node_get_transport(id->device->node_type)) {
3144 case RDMA_TRANSPORT_IB:
3145 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
3146 case IB_LINK_LAYER_INFINIBAND:
3147 ret = cma_join_ib_multicast(id_priv, mc);
3149 case IB_LINK_LAYER_ETHERNET:
3150 kref_init(&mc->mcref);
3151 ret = cma_iboe_join_multicast(id_priv, mc);
3163 spin_lock_irq(&id_priv->lock);
3164 list_del(&mc->list);
3165 spin_unlock_irq(&id_priv->lock);
3171 EXPORT_SYMBOL(rdma_join_multicast);
3173 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
3175 struct rdma_id_private *id_priv;
3176 struct cma_multicast *mc;
3178 id_priv = container_of(id, struct rdma_id_private, id);
3179 spin_lock_irq(&id_priv->lock);
3180 list_for_each_entry(mc, &id_priv->mc_list, list) {
3181 if (!memcmp(&mc->addr, addr, ip_addr_size(addr))) {
3182 list_del(&mc->list);
3183 spin_unlock_irq(&id_priv->lock);
3186 ib_detach_mcast(id->qp,
3187 &mc->multicast.ib->rec.mgid,
3188 mc->multicast.ib->rec.mlid);
3189 if (rdma_node_get_transport(id_priv->cma_dev->device->node_type) == RDMA_TRANSPORT_IB) {
3190 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
3191 case IB_LINK_LAYER_INFINIBAND:
3192 ib_sa_free_multicast(mc->multicast.ib);
3195 case IB_LINK_LAYER_ETHERNET:
3196 kref_put(&mc->mcref, release_mc);
3205 spin_unlock_irq(&id_priv->lock);
3207 EXPORT_SYMBOL(rdma_leave_multicast);
3209 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
3211 struct rdma_dev_addr *dev_addr;
3212 struct cma_ndev_work *work;
3214 dev_addr = &id_priv->id.route.addr.dev_addr;
3217 if ((dev_addr->bound_dev_if == ndev->ifindex) &&
3218 memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
3219 printk(KERN_INFO "RDMA CM addr change for ndev %s used by id %p\n",
3220 ndev->name, &id_priv->id);
3222 if ((dev_addr->bound_dev_if == ndev->if_index) &&
3223 memcmp(dev_addr->src_dev_addr, IF_LLADDR(ndev), ndev->if_addrlen)) {
3224 printk(KERN_INFO "RDMA CM addr change for ndev %s used by id %p\n",
3225 ndev->if_xname, &id_priv->id);
3227 work = kzalloc(sizeof *work, GFP_KERNEL);
3231 INIT_WORK(&work->work, cma_ndev_work_handler);
3233 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
3234 atomic_inc(&id_priv->refcount);
3235 queue_work(cma_wq, &work->work);
3241 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
3244 struct net_device *ndev = (struct net_device *)ctx;
3245 struct cma_device *cma_dev;
3246 struct rdma_id_private *id_priv;
3247 int ret = NOTIFY_DONE;
3250 if (dev_net(ndev) != &init_net)
3253 if (event != NETDEV_BONDING_FAILOVER)
3256 if (!(ndev->flags & IFF_MASTER) || !(ndev->priv_flags & IFF_BONDING))
3259 if (event != NETDEV_DOWN && event != NETDEV_UNREGISTER)
3264 list_for_each_entry(cma_dev, &dev_list, list)
3265 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
3266 ret = cma_netdev_change(ndev, id_priv);
3272 mutex_unlock(&lock);
3276 static struct notifier_block cma_nb = {
3277 .notifier_call = cma_netdev_callback
3280 static void cma_add_one(struct ib_device *device)
3282 struct cma_device *cma_dev;
3283 struct rdma_id_private *id_priv;
3285 cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
3289 cma_dev->device = device;
3291 init_completion(&cma_dev->comp);
3292 atomic_set(&cma_dev->refcount, 1);
3293 INIT_LIST_HEAD(&cma_dev->id_list);
3294 ib_set_client_data(device, &cma_client, cma_dev);
3297 list_add_tail(&cma_dev->list, &dev_list);
3298 list_for_each_entry(id_priv, &listen_any_list, list)
3299 cma_listen_on_dev(id_priv, cma_dev);
3300 mutex_unlock(&lock);
3303 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
3305 struct rdma_cm_event event;
3306 enum cma_state state;
3309 /* Record that we want to remove the device */
3310 state = cma_exch(id_priv, CMA_DEVICE_REMOVAL);
3311 if (state == CMA_DESTROYING)
3314 cma_cancel_operation(id_priv, state);
3315 mutex_lock(&id_priv->handler_mutex);
3317 /* Check for destruction from another callback. */
3318 if (!cma_comp(id_priv, CMA_DEVICE_REMOVAL))
3321 memset(&event, 0, sizeof event);
3322 event.event = RDMA_CM_EVENT_DEVICE_REMOVAL;
3323 ret = id_priv->id.event_handler(&id_priv->id, &event);
3325 mutex_unlock(&id_priv->handler_mutex);
3329 static void cma_process_remove(struct cma_device *cma_dev)
3331 struct rdma_id_private *id_priv;
3335 while (!list_empty(&cma_dev->id_list)) {
3336 id_priv = list_entry(cma_dev->id_list.next,
3337 struct rdma_id_private, list);
3339 list_del(&id_priv->listen_list);
3340 list_del_init(&id_priv->list);
3341 atomic_inc(&id_priv->refcount);
3342 mutex_unlock(&lock);
3344 ret = id_priv->internal_id ? 1 : cma_remove_id_dev(id_priv);
3345 cma_deref_id(id_priv);
3347 rdma_destroy_id(&id_priv->id);
3351 mutex_unlock(&lock);
3353 cma_deref_dev(cma_dev);
3354 wait_for_completion(&cma_dev->comp);
3357 static void cma_remove_one(struct ib_device *device)
3359 struct cma_device *cma_dev;
3361 cma_dev = ib_get_client_data(device, &cma_client);
3366 list_del(&cma_dev->list);
3367 mutex_unlock(&lock);
3369 cma_process_remove(cma_dev);
3373 static int cma_init(void)
3377 int low, high, remaining;
3379 get_random_bytes(&next_port, sizeof next_port);
3380 inet_get_local_port_range(&low, &high);
3381 remaining = (high - low) + 1;
3382 next_port = ((unsigned int) next_port % remaining) + low;
3385 cma_wq = create_singlethread_workqueue("rdma_cm");
3389 ib_sa_register_client(&sa_client);
3390 rdma_addr_register_client(&addr_client);
3391 register_netdevice_notifier(&cma_nb);
3393 ret = ib_register_client(&cma_client);
3399 unregister_netdevice_notifier(&cma_nb);
3400 rdma_addr_unregister_client(&addr_client);
3401 ib_sa_unregister_client(&sa_client);
3402 destroy_workqueue(cma_wq);
3406 static void cma_cleanup(void)
3408 ib_unregister_client(&cma_client);
3409 unregister_netdevice_notifier(&cma_nb);
3410 rdma_addr_unregister_client(&addr_client);
3411 ib_sa_unregister_client(&sa_client);
3412 destroy_workqueue(cma_wq);
3413 idr_destroy(&sdp_ps);
3414 idr_destroy(&tcp_ps);
3415 idr_destroy(&udp_ps);
3416 idr_destroy(&ipoib_ps);
3419 module_init(cma_init);
3420 module_exit(cma_cleanup);