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>
48 #include <rdma/rdma_cm.h>
49 #include <rdma/rdma_cm_ib.h>
50 #include <rdma/ib_cache.h>
51 #include <rdma/ib_cm.h>
52 #include <rdma/ib_sa.h>
53 #include <rdma/iw_cm.h>
55 MODULE_AUTHOR("Sean Hefty");
56 MODULE_DESCRIPTION("Generic RDMA CM Agent");
57 MODULE_LICENSE("Dual BSD/GPL");
59 #define CMA_CM_RESPONSE_TIMEOUT 20
60 #define CMA_MAX_CM_RETRIES 15
61 #define CMA_CM_MRA_SETTING (IB_CM_MRA_FLAG_DELAY | 24)
63 static void cma_add_one(struct ib_device *device);
64 static void cma_remove_one(struct ib_device *device);
66 static struct ib_client cma_client = {
69 .remove = cma_remove_one
72 static struct ib_sa_client sa_client;
73 static struct rdma_addr_client addr_client;
74 static LIST_HEAD(dev_list);
75 static LIST_HEAD(listen_any_list);
76 static DEFINE_MUTEX(lock);
77 static struct workqueue_struct *cma_wq;
78 static DEFINE_IDR(sdp_ps);
79 static DEFINE_IDR(tcp_ps);
80 static DEFINE_IDR(udp_ps);
81 static DEFINE_IDR(ipoib_ps);
84 struct list_head list;
85 struct ib_device *device;
86 struct completion comp;
88 struct list_head id_list;
105 struct rdma_bind_list {
107 struct hlist_head owners;
112 * Device removal can occur at anytime, so we need extra handling to
113 * serialize notifying the user of device removal with other callbacks.
114 * We do this by disabling removal notification while a callback is in process,
115 * and reporting it after the callback completes.
117 struct rdma_id_private {
118 struct rdma_cm_id id;
120 struct rdma_bind_list *bind_list;
121 struct hlist_node node;
122 struct list_head list; /* listen_any_list or cma_device.list */
123 struct list_head listen_list; /* per device listens */
124 struct cma_device *cma_dev;
125 struct list_head mc_list;
128 enum cma_state state;
130 struct mutex qp_mutex;
132 struct completion comp;
134 struct mutex handler_mutex;
138 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;
163 struct work_struct work;
164 struct rdma_id_private *id;
165 enum cma_state old_state;
166 enum cma_state new_state;
167 struct rdma_cm_event event;
170 struct cma_ndev_work {
171 struct work_struct work;
172 struct rdma_id_private *id;
173 struct rdma_cm_event event;
186 u8 ip_version; /* IP version: 7:4 */
188 union cma_ip_addr src_addr;
189 union cma_ip_addr dst_addr;
194 u8 sdp_version; /* Major version: 7:4 */
195 u8 ip_version; /* IP version: 7:4 */
196 u8 sdp_specific1[10];
198 __be16 sdp_specific2;
199 union cma_ip_addr src_addr;
200 union cma_ip_addr dst_addr;
208 #define CMA_VERSION 0x00
209 #define SDP_MAJ_VERSION 0x2
211 static int cma_comp(struct rdma_id_private *id_priv, enum cma_state comp)
216 spin_lock_irqsave(&id_priv->lock, flags);
217 ret = (id_priv->state == comp);
218 spin_unlock_irqrestore(&id_priv->lock, flags);
222 static int cma_comp_exch(struct rdma_id_private *id_priv,
223 enum cma_state comp, enum cma_state exch)
228 spin_lock_irqsave(&id_priv->lock, flags);
229 if ((ret = (id_priv->state == comp)))
230 id_priv->state = exch;
231 spin_unlock_irqrestore(&id_priv->lock, flags);
235 static enum cma_state cma_exch(struct rdma_id_private *id_priv,
241 spin_lock_irqsave(&id_priv->lock, flags);
242 old = id_priv->state;
243 id_priv->state = exch;
244 spin_unlock_irqrestore(&id_priv->lock, flags);
248 static inline u8 cma_get_ip_ver(struct cma_hdr *hdr)
250 return hdr->ip_version >> 4;
253 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
255 hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
258 static inline u8 sdp_get_majv(u8 sdp_version)
260 return sdp_version >> 4;
263 static inline u8 sdp_get_ip_ver(struct sdp_hh *hh)
265 return hh->ip_version >> 4;
268 static inline void sdp_set_ip_ver(struct sdp_hh *hh, u8 ip_ver)
270 hh->ip_version = (ip_ver << 4) | (hh->ip_version & 0xF);
273 static inline int cma_is_ud_ps(enum rdma_port_space ps)
275 return (ps == RDMA_PS_UDP || ps == RDMA_PS_IPOIB);
278 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
279 struct cma_device *cma_dev)
281 atomic_inc(&cma_dev->refcount);
282 id_priv->cma_dev = cma_dev;
283 id_priv->id.device = cma_dev->device;
284 list_add_tail(&id_priv->list, &cma_dev->id_list);
287 static inline void cma_deref_dev(struct cma_device *cma_dev)
289 if (atomic_dec_and_test(&cma_dev->refcount))
290 complete(&cma_dev->comp);
293 static void cma_detach_from_dev(struct rdma_id_private *id_priv)
295 list_del(&id_priv->list);
296 cma_deref_dev(id_priv->cma_dev);
297 id_priv->cma_dev = NULL;
300 static int cma_set_qkey(struct rdma_id_private *id_priv)
302 struct ib_sa_mcmember_rec rec;
308 switch (id_priv->id.ps) {
310 id_priv->qkey = RDMA_UDP_QKEY;
313 ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid);
314 ret = ib_sa_get_mcmember_rec(id_priv->id.device,
315 id_priv->id.port_num, &rec.mgid,
318 id_priv->qkey = be32_to_cpu(rec.qkey);
326 static int cma_acquire_dev(struct rdma_id_private *id_priv)
328 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
329 struct cma_device *cma_dev;
333 rdma_addr_get_sgid(dev_addr, &gid);
334 list_for_each_entry(cma_dev, &dev_list, list) {
335 ret = ib_find_cached_gid(cma_dev->device, &gid,
336 &id_priv->id.port_num, NULL);
338 cma_attach_to_dev(id_priv, cma_dev);
345 static void cma_deref_id(struct rdma_id_private *id_priv)
347 if (atomic_dec_and_test(&id_priv->refcount))
348 complete(&id_priv->comp);
351 static int cma_disable_callback(struct rdma_id_private *id_priv,
352 enum cma_state state)
354 mutex_lock(&id_priv->handler_mutex);
355 if (id_priv->state != state) {
356 mutex_unlock(&id_priv->handler_mutex);
362 static int cma_has_cm_dev(struct rdma_id_private *id_priv)
364 return (id_priv->id.device && id_priv->cm_id.ib);
367 struct rdma_cm_id *rdma_create_id(rdma_cm_event_handler event_handler,
368 void *context, enum rdma_port_space ps)
370 struct rdma_id_private *id_priv;
372 id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
374 return ERR_PTR(-ENOMEM);
376 id_priv->state = CMA_IDLE;
377 id_priv->id.context = context;
378 id_priv->id.event_handler = event_handler;
380 spin_lock_init(&id_priv->lock);
381 mutex_init(&id_priv->qp_mutex);
382 init_completion(&id_priv->comp);
383 atomic_set(&id_priv->refcount, 1);
384 mutex_init(&id_priv->handler_mutex);
385 INIT_LIST_HEAD(&id_priv->listen_list);
386 INIT_LIST_HEAD(&id_priv->mc_list);
387 get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
391 EXPORT_SYMBOL(rdma_create_id);
393 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
395 struct ib_qp_attr qp_attr;
396 int qp_attr_mask, ret;
398 qp_attr.qp_state = IB_QPS_INIT;
399 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
403 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
407 qp_attr.qp_state = IB_QPS_RTR;
408 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
412 qp_attr.qp_state = IB_QPS_RTS;
414 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
419 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
421 struct ib_qp_attr qp_attr;
422 int qp_attr_mask, ret;
424 qp_attr.qp_state = IB_QPS_INIT;
425 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
429 return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
432 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
433 struct ib_qp_init_attr *qp_init_attr)
435 struct rdma_id_private *id_priv;
439 id_priv = container_of(id, struct rdma_id_private, id);
440 if (id->device != pd->device)
443 qp = ib_create_qp(pd, qp_init_attr);
447 if (cma_is_ud_ps(id_priv->id.ps))
448 ret = cma_init_ud_qp(id_priv, qp);
450 ret = cma_init_conn_qp(id_priv, qp);
455 id_priv->qp_num = qp->qp_num;
456 id_priv->srq = (qp->srq != NULL);
462 EXPORT_SYMBOL(rdma_create_qp);
464 void rdma_destroy_qp(struct rdma_cm_id *id)
466 struct rdma_id_private *id_priv;
468 id_priv = container_of(id, struct rdma_id_private, id);
469 mutex_lock(&id_priv->qp_mutex);
470 ib_destroy_qp(id_priv->id.qp);
471 id_priv->id.qp = NULL;
472 mutex_unlock(&id_priv->qp_mutex);
474 EXPORT_SYMBOL(rdma_destroy_qp);
476 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
477 struct rdma_conn_param *conn_param)
479 struct ib_qp_attr qp_attr;
480 int qp_attr_mask, ret;
482 mutex_lock(&id_priv->qp_mutex);
483 if (!id_priv->id.qp) {
488 /* Need to update QP attributes from default values. */
489 qp_attr.qp_state = IB_QPS_INIT;
490 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
494 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
498 qp_attr.qp_state = IB_QPS_RTR;
499 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
504 qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
505 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
507 mutex_unlock(&id_priv->qp_mutex);
511 static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
512 struct rdma_conn_param *conn_param)
514 struct ib_qp_attr qp_attr;
515 int qp_attr_mask, ret;
517 mutex_lock(&id_priv->qp_mutex);
518 if (!id_priv->id.qp) {
523 qp_attr.qp_state = IB_QPS_RTS;
524 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
529 qp_attr.max_rd_atomic = conn_param->initiator_depth;
530 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
532 mutex_unlock(&id_priv->qp_mutex);
536 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
538 struct ib_qp_attr qp_attr;
541 mutex_lock(&id_priv->qp_mutex);
542 if (!id_priv->id.qp) {
547 qp_attr.qp_state = IB_QPS_ERR;
548 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
550 mutex_unlock(&id_priv->qp_mutex);
554 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
555 struct ib_qp_attr *qp_attr, int *qp_attr_mask)
557 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
560 ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
561 ib_addr_get_pkey(dev_addr),
562 &qp_attr->pkey_index);
566 qp_attr->port_num = id_priv->id.port_num;
567 *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
569 if (cma_is_ud_ps(id_priv->id.ps)) {
570 ret = cma_set_qkey(id_priv);
574 qp_attr->qkey = id_priv->qkey;
575 *qp_attr_mask |= IB_QP_QKEY;
577 qp_attr->qp_access_flags = 0;
578 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
583 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
586 struct rdma_id_private *id_priv;
589 id_priv = container_of(id, struct rdma_id_private, id);
590 switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
591 case RDMA_TRANSPORT_IB:
592 if (!id_priv->cm_id.ib || cma_is_ud_ps(id_priv->id.ps))
593 ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
595 ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
597 if (qp_attr->qp_state == IB_QPS_RTR)
598 qp_attr->rq_psn = id_priv->seq_num;
600 case RDMA_TRANSPORT_IWARP:
601 if (!id_priv->cm_id.iw) {
602 qp_attr->qp_access_flags = 0;
603 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
605 ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
615 EXPORT_SYMBOL(rdma_init_qp_attr);
617 static inline int cma_zero_addr(struct sockaddr *addr)
619 struct in6_addr *ip6;
621 if (addr->sa_family == AF_INET)
622 return ipv4_is_zeronet(
623 ((struct sockaddr_in *)addr)->sin_addr.s_addr);
625 ip6 = &((struct sockaddr_in6 *) addr)->sin6_addr;
626 return (ip6->s6_addr32[0] | ip6->s6_addr32[1] |
627 ip6->s6_addr32[2] | ip6->s6_addr32[3]) == 0;
631 static inline int cma_loopback_addr(struct sockaddr *addr)
633 if (addr->sa_family == AF_INET)
634 return ipv4_is_loopback(
635 ((struct sockaddr_in *) addr)->sin_addr.s_addr);
637 return ipv6_addr_loopback(
638 &((struct sockaddr_in6 *) addr)->sin6_addr);
641 static inline int cma_any_addr(struct sockaddr *addr)
643 return cma_zero_addr(addr) || cma_loopback_addr(addr);
646 static inline __be16 cma_port(struct sockaddr *addr)
648 if (addr->sa_family == AF_INET)
649 return ((struct sockaddr_in *) addr)->sin_port;
651 return ((struct sockaddr_in6 *) addr)->sin6_port;
654 static inline int cma_any_port(struct sockaddr *addr)
656 return !cma_port(addr);
659 static int cma_get_net_info(void *hdr, enum rdma_port_space ps,
660 u8 *ip_ver, __be16 *port,
661 union cma_ip_addr **src, union cma_ip_addr **dst)
665 if (sdp_get_majv(((struct sdp_hh *) hdr)->sdp_version) !=
669 *ip_ver = sdp_get_ip_ver(hdr);
670 *port = ((struct sdp_hh *) hdr)->port;
671 *src = &((struct sdp_hh *) hdr)->src_addr;
672 *dst = &((struct sdp_hh *) hdr)->dst_addr;
675 if (((struct cma_hdr *) hdr)->cma_version != CMA_VERSION)
678 *ip_ver = cma_get_ip_ver(hdr);
679 *port = ((struct cma_hdr *) hdr)->port;
680 *src = &((struct cma_hdr *) hdr)->src_addr;
681 *dst = &((struct cma_hdr *) hdr)->dst_addr;
685 if (*ip_ver != 4 && *ip_ver != 6)
690 static void cma_save_net_info(struct rdma_addr *addr,
691 struct rdma_addr *listen_addr,
692 u8 ip_ver, __be16 port,
693 union cma_ip_addr *src, union cma_ip_addr *dst)
695 struct sockaddr_in *listen4, *ip4;
696 struct sockaddr_in6 *listen6, *ip6;
700 listen4 = (struct sockaddr_in *) &listen_addr->src_addr;
701 ip4 = (struct sockaddr_in *) &addr->src_addr;
702 ip4->sin_family = listen4->sin_family;
703 ip4->sin_addr.s_addr = dst->ip4.addr;
704 ip4->sin_port = listen4->sin_port;
706 ip4 = (struct sockaddr_in *) &addr->dst_addr;
707 ip4->sin_family = listen4->sin_family;
708 ip4->sin_addr.s_addr = src->ip4.addr;
709 ip4->sin_port = port;
712 listen6 = (struct sockaddr_in6 *) &listen_addr->src_addr;
713 ip6 = (struct sockaddr_in6 *) &addr->src_addr;
714 ip6->sin6_family = listen6->sin6_family;
715 ip6->sin6_addr = dst->ip6;
716 ip6->sin6_port = listen6->sin6_port;
718 ip6 = (struct sockaddr_in6 *) &addr->dst_addr;
719 ip6->sin6_family = listen6->sin6_family;
720 ip6->sin6_addr = src->ip6;
721 ip6->sin6_port = port;
728 static inline int cma_user_data_offset(enum rdma_port_space ps)
734 return sizeof(struct cma_hdr);
738 static void cma_cancel_route(struct rdma_id_private *id_priv)
740 switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
741 case RDMA_TRANSPORT_IB:
743 ib_sa_cancel_query(id_priv->query_id, id_priv->query);
750 static void cma_cancel_listens(struct rdma_id_private *id_priv)
752 struct rdma_id_private *dev_id_priv;
755 * Remove from listen_any_list to prevent added devices from spawning
756 * additional listen requests.
759 list_del(&id_priv->list);
761 while (!list_empty(&id_priv->listen_list)) {
762 dev_id_priv = list_entry(id_priv->listen_list.next,
763 struct rdma_id_private, listen_list);
764 /* sync with device removal to avoid duplicate destruction */
765 list_del_init(&dev_id_priv->list);
766 list_del(&dev_id_priv->listen_list);
769 rdma_destroy_id(&dev_id_priv->id);
775 static void cma_cancel_operation(struct rdma_id_private *id_priv,
776 enum cma_state state)
780 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
782 case CMA_ROUTE_QUERY:
783 cma_cancel_route(id_priv);
786 if (cma_any_addr((struct sockaddr *) &id_priv->id.route.addr.src_addr)
787 && !id_priv->cma_dev)
788 cma_cancel_listens(id_priv);
795 static void cma_release_port(struct rdma_id_private *id_priv)
797 struct rdma_bind_list *bind_list = id_priv->bind_list;
803 hlist_del(&id_priv->node);
804 if (hlist_empty(&bind_list->owners)) {
805 idr_remove(bind_list->ps, bind_list->port);
811 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
813 struct cma_multicast *mc;
815 while (!list_empty(&id_priv->mc_list)) {
816 mc = container_of(id_priv->mc_list.next,
817 struct cma_multicast, list);
819 ib_sa_free_multicast(mc->multicast.ib);
824 void rdma_destroy_id(struct rdma_cm_id *id)
826 struct rdma_id_private *id_priv;
827 enum cma_state state;
829 id_priv = container_of(id, struct rdma_id_private, id);
830 state = cma_exch(id_priv, CMA_DESTROYING);
831 cma_cancel_operation(id_priv, state);
834 if (id_priv->cma_dev) {
836 switch (rdma_node_get_transport(id->device->node_type)) {
837 case RDMA_TRANSPORT_IB:
838 if (id_priv->cm_id.ib && !IS_ERR(id_priv->cm_id.ib))
839 ib_destroy_cm_id(id_priv->cm_id.ib);
841 case RDMA_TRANSPORT_IWARP:
842 if (id_priv->cm_id.iw && !IS_ERR(id_priv->cm_id.iw))
843 iw_destroy_cm_id(id_priv->cm_id.iw);
848 cma_leave_mc_groups(id_priv);
850 cma_detach_from_dev(id_priv);
854 cma_release_port(id_priv);
855 cma_deref_id(id_priv);
856 wait_for_completion(&id_priv->comp);
858 if (id_priv->internal_id)
859 cma_deref_id(id_priv->id.context);
861 kfree(id_priv->id.route.path_rec);
864 EXPORT_SYMBOL(rdma_destroy_id);
866 static int cma_rep_recv(struct rdma_id_private *id_priv)
870 ret = cma_modify_qp_rtr(id_priv, NULL);
874 ret = cma_modify_qp_rts(id_priv, NULL);
878 ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
884 cma_modify_qp_err(id_priv);
885 ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
890 static int cma_verify_rep(struct rdma_id_private *id_priv, void *data)
892 if (id_priv->id.ps == RDMA_PS_SDP &&
893 sdp_get_majv(((struct sdp_hah *) data)->sdp_version) !=
900 static void cma_set_rep_event_data(struct rdma_cm_event *event,
901 struct ib_cm_rep_event_param *rep_data,
904 event->param.conn.private_data = private_data;
905 event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
906 event->param.conn.responder_resources = rep_data->responder_resources;
907 event->param.conn.initiator_depth = rep_data->initiator_depth;
908 event->param.conn.flow_control = rep_data->flow_control;
909 event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
910 event->param.conn.srq = rep_data->srq;
911 event->param.conn.qp_num = rep_data->remote_qpn;
914 static int cma_ib_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
916 struct rdma_id_private *id_priv = cm_id->context;
917 struct rdma_cm_event event;
920 if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
921 cma_disable_callback(id_priv, CMA_CONNECT)) ||
922 (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
923 cma_disable_callback(id_priv, CMA_DISCONNECT)))
926 memset(&event, 0, sizeof event);
927 switch (ib_event->event) {
928 case IB_CM_REQ_ERROR:
929 case IB_CM_REP_ERROR:
930 event.event = RDMA_CM_EVENT_UNREACHABLE;
931 event.status = -ETIMEDOUT;
933 case IB_CM_REP_RECEIVED:
934 event.status = cma_verify_rep(id_priv, ib_event->private_data);
936 event.event = RDMA_CM_EVENT_CONNECT_ERROR;
937 else if (id_priv->id.qp && id_priv->id.ps != RDMA_PS_SDP) {
938 event.status = cma_rep_recv(id_priv);
939 event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
940 RDMA_CM_EVENT_ESTABLISHED;
942 event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
943 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
944 ib_event->private_data);
946 case IB_CM_RTU_RECEIVED:
947 case IB_CM_USER_ESTABLISHED:
948 event.event = RDMA_CM_EVENT_ESTABLISHED;
950 case IB_CM_DREQ_ERROR:
951 event.status = -ETIMEDOUT; /* fall through */
952 case IB_CM_DREQ_RECEIVED:
953 case IB_CM_DREP_RECEIVED:
954 if (!cma_comp_exch(id_priv, CMA_CONNECT, CMA_DISCONNECT))
956 event.event = RDMA_CM_EVENT_DISCONNECTED;
958 case IB_CM_TIMEWAIT_EXIT:
959 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
961 case IB_CM_MRA_RECEIVED:
964 case IB_CM_REJ_RECEIVED:
965 cma_modify_qp_err(id_priv);
966 event.status = ib_event->param.rej_rcvd.reason;
967 event.event = RDMA_CM_EVENT_REJECTED;
968 event.param.conn.private_data = ib_event->private_data;
969 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
972 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
977 ret = id_priv->id.event_handler(&id_priv->id, &event);
979 /* Destroy the CM ID by returning a non-zero value. */
980 id_priv->cm_id.ib = NULL;
981 cma_exch(id_priv, CMA_DESTROYING);
982 mutex_unlock(&id_priv->handler_mutex);
983 rdma_destroy_id(&id_priv->id);
987 mutex_unlock(&id_priv->handler_mutex);
991 static struct rdma_id_private *cma_new_conn_id(struct rdma_cm_id *listen_id,
992 struct ib_cm_event *ib_event)
994 struct rdma_id_private *id_priv;
995 struct rdma_cm_id *id;
996 struct rdma_route *rt;
997 union cma_ip_addr *src, *dst;
1002 if (cma_get_net_info(ib_event->private_data, listen_id->ps,
1003 &ip_ver, &port, &src, &dst))
1006 id = rdma_create_id(listen_id->event_handler, listen_id->context,
1011 cma_save_net_info(&id->route.addr, &listen_id->route.addr,
1012 ip_ver, port, src, dst);
1015 rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
1016 rt->path_rec = kmalloc(sizeof *rt->path_rec * rt->num_paths,
1021 rt->path_rec[0] = *ib_event->param.req_rcvd.primary_path;
1022 if (rt->num_paths == 2)
1023 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
1025 if (cma_any_addr((struct sockaddr *) &rt->addr.src_addr)) {
1026 rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
1027 rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
1028 ib_addr_set_pkey(&rt->addr.dev_addr, rt->path_rec[0].pkey);
1030 ret = rdma_translate_ip((struct sockaddr *) &rt->addr.src_addr,
1031 &rt->addr.dev_addr);
1035 rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
1037 id_priv = container_of(id, struct rdma_id_private, id);
1038 id_priv->state = CMA_CONNECT;
1042 rdma_destroy_id(id);
1047 static struct rdma_id_private *cma_new_udp_id(struct rdma_cm_id *listen_id,
1048 struct ib_cm_event *ib_event)
1050 struct rdma_id_private *id_priv;
1051 struct rdma_cm_id *id;
1052 union cma_ip_addr *src, *dst;
1057 id = rdma_create_id(listen_id->event_handler, listen_id->context,
1063 if (cma_get_net_info(ib_event->private_data, listen_id->ps,
1064 &ip_ver, &port, &src, &dst))
1067 cma_save_net_info(&id->route.addr, &listen_id->route.addr,
1068 ip_ver, port, src, dst);
1070 if (!cma_any_addr((struct sockaddr *) &id->route.addr.src_addr)) {
1071 ret = rdma_translate_ip((struct sockaddr *) &id->route.addr.src_addr,
1072 &id->route.addr.dev_addr);
1077 id_priv = container_of(id, struct rdma_id_private, id);
1078 id_priv->state = CMA_CONNECT;
1081 rdma_destroy_id(id);
1085 static void cma_set_req_event_data(struct rdma_cm_event *event,
1086 struct ib_cm_req_event_param *req_data,
1087 void *private_data, int offset)
1089 event->param.conn.private_data = private_data + offset;
1090 event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
1091 event->param.conn.responder_resources = req_data->responder_resources;
1092 event->param.conn.initiator_depth = req_data->initiator_depth;
1093 event->param.conn.flow_control = req_data->flow_control;
1094 event->param.conn.retry_count = req_data->retry_count;
1095 event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
1096 event->param.conn.srq = req_data->srq;
1097 event->param.conn.qp_num = req_data->remote_qpn;
1100 static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1102 struct rdma_id_private *listen_id, *conn_id;
1103 struct rdma_cm_event event;
1106 listen_id = cm_id->context;
1107 if (cma_disable_callback(listen_id, CMA_LISTEN))
1108 return -ECONNABORTED;
1110 memset(&event, 0, sizeof event);
1111 offset = cma_user_data_offset(listen_id->id.ps);
1112 event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1113 if (cma_is_ud_ps(listen_id->id.ps)) {
1114 conn_id = cma_new_udp_id(&listen_id->id, ib_event);
1115 event.param.ud.private_data = ib_event->private_data + offset;
1116 event.param.ud.private_data_len =
1117 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
1119 conn_id = cma_new_conn_id(&listen_id->id, ib_event);
1120 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
1121 ib_event->private_data, offset);
1128 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1130 ret = cma_acquire_dev(conn_id);
1131 mutex_unlock(&lock);
1133 goto release_conn_id;
1135 conn_id->cm_id.ib = cm_id;
1136 cm_id->context = conn_id;
1137 cm_id->cm_handler = cma_ib_handler;
1139 ret = conn_id->id.event_handler(&conn_id->id, &event);
1142 * Acquire mutex to prevent user executing rdma_destroy_id()
1143 * while we're accessing the cm_id.
1146 if (cma_comp(conn_id, CMA_CONNECT) &&
1147 !cma_is_ud_ps(conn_id->id.ps))
1148 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
1149 mutex_unlock(&lock);
1150 mutex_unlock(&conn_id->handler_mutex);
1154 /* Destroy the CM ID by returning a non-zero value. */
1155 conn_id->cm_id.ib = NULL;
1158 cma_exch(conn_id, CMA_DESTROYING);
1159 mutex_unlock(&conn_id->handler_mutex);
1160 rdma_destroy_id(&conn_id->id);
1163 mutex_unlock(&listen_id->handler_mutex);
1167 static __be64 cma_get_service_id(enum rdma_port_space ps, struct sockaddr *addr)
1169 return cpu_to_be64(((u64)ps << 16) + be16_to_cpu(cma_port(addr)));
1172 static void cma_set_compare_data(enum rdma_port_space ps, struct sockaddr *addr,
1173 struct ib_cm_compare_data *compare)
1175 struct cma_hdr *cma_data, *cma_mask;
1176 struct sdp_hh *sdp_data, *sdp_mask;
1178 struct in6_addr ip6_addr;
1180 memset(compare, 0, sizeof *compare);
1181 cma_data = (void *) compare->data;
1182 cma_mask = (void *) compare->mask;
1183 sdp_data = (void *) compare->data;
1184 sdp_mask = (void *) compare->mask;
1186 switch (addr->sa_family) {
1188 ip4_addr = ((struct sockaddr_in *) addr)->sin_addr.s_addr;
1189 if (ps == RDMA_PS_SDP) {
1190 sdp_set_ip_ver(sdp_data, 4);
1191 sdp_set_ip_ver(sdp_mask, 0xF);
1192 sdp_data->dst_addr.ip4.addr = ip4_addr;
1193 sdp_mask->dst_addr.ip4.addr = htonl(~0);
1195 cma_set_ip_ver(cma_data, 4);
1196 cma_set_ip_ver(cma_mask, 0xF);
1197 cma_data->dst_addr.ip4.addr = ip4_addr;
1198 cma_mask->dst_addr.ip4.addr = htonl(~0);
1202 ip6_addr = ((struct sockaddr_in6 *) addr)->sin6_addr;
1203 if (ps == RDMA_PS_SDP) {
1204 sdp_set_ip_ver(sdp_data, 6);
1205 sdp_set_ip_ver(sdp_mask, 0xF);
1206 sdp_data->dst_addr.ip6 = ip6_addr;
1207 memset(&sdp_mask->dst_addr.ip6, 0xFF,
1208 sizeof sdp_mask->dst_addr.ip6);
1210 cma_set_ip_ver(cma_data, 6);
1211 cma_set_ip_ver(cma_mask, 0xF);
1212 cma_data->dst_addr.ip6 = ip6_addr;
1213 memset(&cma_mask->dst_addr.ip6, 0xFF,
1214 sizeof cma_mask->dst_addr.ip6);
1222 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
1224 struct rdma_id_private *id_priv = iw_id->context;
1225 struct rdma_cm_event event;
1226 struct sockaddr_in *sin;
1229 if (cma_disable_callback(id_priv, CMA_CONNECT))
1232 memset(&event, 0, sizeof event);
1233 switch (iw_event->event) {
1234 case IW_CM_EVENT_CLOSE:
1235 event.event = RDMA_CM_EVENT_DISCONNECTED;
1237 case IW_CM_EVENT_CONNECT_REPLY:
1238 sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1239 *sin = iw_event->local_addr;
1240 sin = (struct sockaddr_in *) &id_priv->id.route.addr.dst_addr;
1241 *sin = iw_event->remote_addr;
1242 switch (iw_event->status) {
1244 event.event = RDMA_CM_EVENT_ESTABLISHED;
1248 event.event = RDMA_CM_EVENT_REJECTED;
1251 event.event = RDMA_CM_EVENT_UNREACHABLE;
1254 event.event = RDMA_CM_EVENT_CONNECT_ERROR;
1258 case IW_CM_EVENT_ESTABLISHED:
1259 event.event = RDMA_CM_EVENT_ESTABLISHED;
1265 event.status = iw_event->status;
1266 event.param.conn.private_data = iw_event->private_data;
1267 event.param.conn.private_data_len = iw_event->private_data_len;
1268 ret = id_priv->id.event_handler(&id_priv->id, &event);
1270 /* Destroy the CM ID by returning a non-zero value. */
1271 id_priv->cm_id.iw = NULL;
1272 cma_exch(id_priv, CMA_DESTROYING);
1273 mutex_unlock(&id_priv->handler_mutex);
1274 rdma_destroy_id(&id_priv->id);
1278 mutex_unlock(&id_priv->handler_mutex);
1282 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
1283 struct iw_cm_event *iw_event)
1285 struct rdma_cm_id *new_cm_id;
1286 struct rdma_id_private *listen_id, *conn_id;
1287 struct sockaddr_in *sin;
1288 struct net_device *dev = NULL;
1289 struct rdma_cm_event event;
1291 struct ib_device_attr attr;
1293 listen_id = cm_id->context;
1294 if (cma_disable_callback(listen_id, CMA_LISTEN))
1295 return -ECONNABORTED;
1297 /* Create a new RDMA id for the new IW CM ID */
1298 new_cm_id = rdma_create_id(listen_id->id.event_handler,
1299 listen_id->id.context,
1301 if (IS_ERR(new_cm_id)) {
1305 conn_id = container_of(new_cm_id, struct rdma_id_private, id);
1306 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1307 conn_id->state = CMA_CONNECT;
1309 dev = ip_dev_find(&init_net, iw_event->local_addr.sin_addr.s_addr);
1311 ret = -EADDRNOTAVAIL;
1312 mutex_unlock(&conn_id->handler_mutex);
1313 rdma_destroy_id(new_cm_id);
1316 ret = rdma_copy_addr(&conn_id->id.route.addr.dev_addr, dev, NULL);
1318 mutex_unlock(&conn_id->handler_mutex);
1319 rdma_destroy_id(new_cm_id);
1324 ret = cma_acquire_dev(conn_id);
1325 mutex_unlock(&lock);
1327 mutex_unlock(&conn_id->handler_mutex);
1328 rdma_destroy_id(new_cm_id);
1332 conn_id->cm_id.iw = cm_id;
1333 cm_id->context = conn_id;
1334 cm_id->cm_handler = cma_iw_handler;
1336 sin = (struct sockaddr_in *) &new_cm_id->route.addr.src_addr;
1337 *sin = iw_event->local_addr;
1338 sin = (struct sockaddr_in *) &new_cm_id->route.addr.dst_addr;
1339 *sin = iw_event->remote_addr;
1341 ret = ib_query_device(conn_id->id.device, &attr);
1343 mutex_unlock(&conn_id->handler_mutex);
1344 rdma_destroy_id(new_cm_id);
1348 memset(&event, 0, sizeof event);
1349 event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1350 event.param.conn.private_data = iw_event->private_data;
1351 event.param.conn.private_data_len = iw_event->private_data_len;
1352 event.param.conn.initiator_depth = attr.max_qp_init_rd_atom;
1353 event.param.conn.responder_resources = attr.max_qp_rd_atom;
1354 ret = conn_id->id.event_handler(&conn_id->id, &event);
1356 /* User wants to destroy the CM ID */
1357 conn_id->cm_id.iw = NULL;
1358 cma_exch(conn_id, CMA_DESTROYING);
1359 mutex_unlock(&conn_id->handler_mutex);
1360 rdma_destroy_id(&conn_id->id);
1364 mutex_unlock(&conn_id->handler_mutex);
1369 mutex_unlock(&listen_id->handler_mutex);
1373 static int cma_ib_listen(struct rdma_id_private *id_priv)
1375 struct ib_cm_compare_data compare_data;
1376 struct sockaddr *addr;
1380 id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, cma_req_handler,
1382 if (IS_ERR(id_priv->cm_id.ib))
1383 return PTR_ERR(id_priv->cm_id.ib);
1385 addr = (struct sockaddr *) &id_priv->id.route.addr.src_addr;
1386 svc_id = cma_get_service_id(id_priv->id.ps, addr);
1387 if (cma_any_addr(addr))
1388 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, NULL);
1390 cma_set_compare_data(id_priv->id.ps, addr, &compare_data);
1391 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, &compare_data);
1395 ib_destroy_cm_id(id_priv->cm_id.ib);
1396 id_priv->cm_id.ib = NULL;
1402 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
1405 struct sockaddr_in *sin;
1407 id_priv->cm_id.iw = iw_create_cm_id(id_priv->id.device,
1408 iw_conn_req_handler,
1410 if (IS_ERR(id_priv->cm_id.iw))
1411 return PTR_ERR(id_priv->cm_id.iw);
1413 sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1414 id_priv->cm_id.iw->local_addr = *sin;
1416 ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
1419 iw_destroy_cm_id(id_priv->cm_id.iw);
1420 id_priv->cm_id.iw = NULL;
1426 static int cma_listen_handler(struct rdma_cm_id *id,
1427 struct rdma_cm_event *event)
1429 struct rdma_id_private *id_priv = id->context;
1431 id->context = id_priv->id.context;
1432 id->event_handler = id_priv->id.event_handler;
1433 return id_priv->id.event_handler(id, event);
1436 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
1437 struct cma_device *cma_dev)
1439 struct rdma_id_private *dev_id_priv;
1440 struct rdma_cm_id *id;
1443 id = rdma_create_id(cma_listen_handler, id_priv, id_priv->id.ps);
1447 dev_id_priv = container_of(id, struct rdma_id_private, id);
1449 dev_id_priv->state = CMA_ADDR_BOUND;
1450 memcpy(&id->route.addr.src_addr, &id_priv->id.route.addr.src_addr,
1451 ip_addr_size((struct sockaddr *) &id_priv->id.route.addr.src_addr));
1453 cma_attach_to_dev(dev_id_priv, cma_dev);
1454 list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
1455 atomic_inc(&id_priv->refcount);
1456 dev_id_priv->internal_id = 1;
1458 ret = rdma_listen(id, id_priv->backlog);
1460 printk(KERN_WARNING "RDMA CMA: cma_listen_on_dev, error %d, "
1461 "listening on device %s\n", ret, cma_dev->device->name);
1464 static void cma_listen_on_all(struct rdma_id_private *id_priv)
1466 struct cma_device *cma_dev;
1469 list_add_tail(&id_priv->list, &listen_any_list);
1470 list_for_each_entry(cma_dev, &dev_list, list)
1471 cma_listen_on_dev(id_priv, cma_dev);
1472 mutex_unlock(&lock);
1475 int rdma_listen(struct rdma_cm_id *id, int backlog)
1477 struct rdma_id_private *id_priv;
1480 id_priv = container_of(id, struct rdma_id_private, id);
1481 if (id_priv->state == CMA_IDLE) {
1482 ((struct sockaddr *) &id->route.addr.src_addr)->sa_family = AF_INET;
1483 ret = rdma_bind_addr(id, (struct sockaddr *) &id->route.addr.src_addr);
1488 if (!cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_LISTEN))
1491 id_priv->backlog = backlog;
1493 switch (rdma_node_get_transport(id->device->node_type)) {
1494 case RDMA_TRANSPORT_IB:
1495 ret = cma_ib_listen(id_priv);
1499 case RDMA_TRANSPORT_IWARP:
1500 ret = cma_iw_listen(id_priv, backlog);
1509 cma_listen_on_all(id_priv);
1513 id_priv->backlog = 0;
1514 cma_comp_exch(id_priv, CMA_LISTEN, CMA_ADDR_BOUND);
1517 EXPORT_SYMBOL(rdma_listen);
1519 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
1521 struct rdma_id_private *id_priv;
1523 id_priv = container_of(id, struct rdma_id_private, id);
1524 id_priv->tos = (u8) tos;
1526 EXPORT_SYMBOL(rdma_set_service_type);
1528 static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec,
1531 struct cma_work *work = context;
1532 struct rdma_route *route;
1534 route = &work->id->id.route;
1537 route->num_paths = 1;
1538 *route->path_rec = *path_rec;
1540 work->old_state = CMA_ROUTE_QUERY;
1541 work->new_state = CMA_ADDR_RESOLVED;
1542 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
1543 work->event.status = status;
1546 queue_work(cma_wq, &work->work);
1549 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms,
1550 struct cma_work *work)
1552 struct rdma_addr *addr = &id_priv->id.route.addr;
1553 struct ib_sa_path_rec path_rec;
1554 ib_sa_comp_mask comp_mask;
1555 struct sockaddr_in6 *sin6;
1557 memset(&path_rec, 0, sizeof path_rec);
1558 rdma_addr_get_sgid(&addr->dev_addr, &path_rec.sgid);
1559 rdma_addr_get_dgid(&addr->dev_addr, &path_rec.dgid);
1560 path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(&addr->dev_addr));
1561 path_rec.numb_path = 1;
1562 path_rec.reversible = 1;
1563 path_rec.service_id = cma_get_service_id(id_priv->id.ps,
1564 (struct sockaddr *) &addr->dst_addr);
1566 comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
1567 IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
1568 IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
1570 if (addr->src_addr.ss_family == AF_INET) {
1571 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
1572 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
1574 sin6 = (struct sockaddr_in6 *) &addr->src_addr;
1575 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
1576 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
1579 id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
1580 id_priv->id.port_num, &path_rec,
1581 comp_mask, timeout_ms,
1582 GFP_KERNEL, cma_query_handler,
1583 work, &id_priv->query);
1585 return (id_priv->query_id < 0) ? id_priv->query_id : 0;
1588 static void cma_work_handler(struct work_struct *_work)
1590 struct cma_work *work = container_of(_work, struct cma_work, work);
1591 struct rdma_id_private *id_priv = work->id;
1594 mutex_lock(&id_priv->handler_mutex);
1595 if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
1598 if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1599 cma_exch(id_priv, CMA_DESTROYING);
1603 mutex_unlock(&id_priv->handler_mutex);
1604 cma_deref_id(id_priv);
1606 rdma_destroy_id(&id_priv->id);
1610 static void cma_ndev_work_handler(struct work_struct *_work)
1612 struct cma_ndev_work *work = container_of(_work, struct cma_ndev_work, work);
1613 struct rdma_id_private *id_priv = work->id;
1616 mutex_lock(&id_priv->handler_mutex);
1617 if (id_priv->state == CMA_DESTROYING ||
1618 id_priv->state == CMA_DEVICE_REMOVAL)
1621 if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1622 cma_exch(id_priv, CMA_DESTROYING);
1627 mutex_unlock(&id_priv->handler_mutex);
1628 cma_deref_id(id_priv);
1630 rdma_destroy_id(&id_priv->id);
1634 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms)
1636 struct rdma_route *route = &id_priv->id.route;
1637 struct cma_work *work;
1640 work = kzalloc(sizeof *work, GFP_KERNEL);
1645 INIT_WORK(&work->work, cma_work_handler);
1646 work->old_state = CMA_ROUTE_QUERY;
1647 work->new_state = CMA_ROUTE_RESOLVED;
1648 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1650 route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
1651 if (!route->path_rec) {
1656 ret = cma_query_ib_route(id_priv, timeout_ms, work);
1662 kfree(route->path_rec);
1663 route->path_rec = NULL;
1669 int rdma_set_ib_paths(struct rdma_cm_id *id,
1670 struct ib_sa_path_rec *path_rec, int num_paths)
1672 struct rdma_id_private *id_priv;
1675 id_priv = container_of(id, struct rdma_id_private, id);
1676 if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ROUTE_RESOLVED))
1679 id->route.path_rec = kmemdup(path_rec, sizeof *path_rec * num_paths,
1681 if (!id->route.path_rec) {
1686 id->route.num_paths = num_paths;
1689 cma_comp_exch(id_priv, CMA_ROUTE_RESOLVED, CMA_ADDR_RESOLVED);
1692 EXPORT_SYMBOL(rdma_set_ib_paths);
1694 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms)
1696 struct cma_work *work;
1698 work = kzalloc(sizeof *work, GFP_KERNEL);
1703 INIT_WORK(&work->work, cma_work_handler);
1704 work->old_state = CMA_ROUTE_QUERY;
1705 work->new_state = CMA_ROUTE_RESOLVED;
1706 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1707 queue_work(cma_wq, &work->work);
1711 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
1713 struct rdma_id_private *id_priv;
1716 id_priv = container_of(id, struct rdma_id_private, id);
1717 if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ROUTE_QUERY))
1720 atomic_inc(&id_priv->refcount);
1721 switch (rdma_node_get_transport(id->device->node_type)) {
1722 case RDMA_TRANSPORT_IB:
1723 ret = cma_resolve_ib_route(id_priv, timeout_ms);
1725 case RDMA_TRANSPORT_IWARP:
1726 ret = cma_resolve_iw_route(id_priv, timeout_ms);
1737 cma_comp_exch(id_priv, CMA_ROUTE_QUERY, CMA_ADDR_RESOLVED);
1738 cma_deref_id(id_priv);
1741 EXPORT_SYMBOL(rdma_resolve_route);
1743 static int cma_bind_loopback(struct rdma_id_private *id_priv)
1745 struct cma_device *cma_dev;
1746 struct ib_port_attr port_attr;
1753 if (list_empty(&dev_list)) {
1757 list_for_each_entry(cma_dev, &dev_list, list)
1758 for (p = 1; p <= cma_dev->device->phys_port_cnt; ++p)
1759 if (!ib_query_port(cma_dev->device, p, &port_attr) &&
1760 port_attr.state == IB_PORT_ACTIVE)
1764 cma_dev = list_entry(dev_list.next, struct cma_device, list);
1767 ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid);
1771 ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
1775 id_priv->id.route.addr.dev_addr.dev_type =
1776 (rdma_node_get_transport(cma_dev->device->node_type) == RDMA_TRANSPORT_IB) ?
1777 ARPHRD_INFINIBAND : ARPHRD_ETHER;
1779 rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
1780 ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
1781 id_priv->id.port_num = p;
1782 cma_attach_to_dev(id_priv, cma_dev);
1784 mutex_unlock(&lock);
1788 static void addr_handler(int status, struct sockaddr *src_addr,
1789 struct rdma_dev_addr *dev_addr, void *context)
1791 struct rdma_id_private *id_priv = context;
1792 struct rdma_cm_event event;
1794 memset(&event, 0, sizeof event);
1795 mutex_lock(&id_priv->handler_mutex);
1798 * Grab mutex to block rdma_destroy_id() from removing the device while
1799 * we're trying to acquire it.
1802 if (!cma_comp_exch(id_priv, CMA_ADDR_QUERY, CMA_ADDR_RESOLVED)) {
1803 mutex_unlock(&lock);
1807 if (!status && !id_priv->cma_dev)
1808 status = cma_acquire_dev(id_priv);
1809 mutex_unlock(&lock);
1812 if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ADDR_BOUND))
1814 event.event = RDMA_CM_EVENT_ADDR_ERROR;
1815 event.status = status;
1817 memcpy(&id_priv->id.route.addr.src_addr, src_addr,
1818 ip_addr_size(src_addr));
1819 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
1822 if (id_priv->id.event_handler(&id_priv->id, &event)) {
1823 cma_exch(id_priv, CMA_DESTROYING);
1824 mutex_unlock(&id_priv->handler_mutex);
1825 cma_deref_id(id_priv);
1826 rdma_destroy_id(&id_priv->id);
1830 mutex_unlock(&id_priv->handler_mutex);
1831 cma_deref_id(id_priv);
1834 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
1836 struct cma_work *work;
1837 struct sockaddr *src, *dst;
1841 work = kzalloc(sizeof *work, GFP_KERNEL);
1845 if (!id_priv->cma_dev) {
1846 ret = cma_bind_loopback(id_priv);
1851 rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
1852 rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
1854 src = (struct sockaddr *) &id_priv->id.route.addr.src_addr;
1855 if (cma_zero_addr(src)) {
1856 dst = (struct sockaddr *) &id_priv->id.route.addr.dst_addr;
1857 if ((src->sa_family = dst->sa_family) == AF_INET) {
1858 ((struct sockaddr_in *) src)->sin_addr.s_addr =
1859 ((struct sockaddr_in *) dst)->sin_addr.s_addr;
1861 ipv6_addr_copy(&((struct sockaddr_in6 *) src)->sin6_addr,
1862 &((struct sockaddr_in6 *) dst)->sin6_addr);
1867 INIT_WORK(&work->work, cma_work_handler);
1868 work->old_state = CMA_ADDR_QUERY;
1869 work->new_state = CMA_ADDR_RESOLVED;
1870 work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
1871 queue_work(cma_wq, &work->work);
1878 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
1879 struct sockaddr *dst_addr)
1881 if (!src_addr || !src_addr->sa_family) {
1882 src_addr = (struct sockaddr *) &id->route.addr.src_addr;
1883 if ((src_addr->sa_family = dst_addr->sa_family) == AF_INET6) {
1884 ((struct sockaddr_in6 *) src_addr)->sin6_scope_id =
1885 ((struct sockaddr_in6 *) dst_addr)->sin6_scope_id;
1888 return rdma_bind_addr(id, src_addr);
1891 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
1892 struct sockaddr *dst_addr, int timeout_ms)
1894 struct rdma_id_private *id_priv;
1897 id_priv = container_of(id, struct rdma_id_private, id);
1898 if (id_priv->state == CMA_IDLE) {
1899 ret = cma_bind_addr(id, src_addr, dst_addr);
1904 if (!cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_ADDR_QUERY))
1907 atomic_inc(&id_priv->refcount);
1908 memcpy(&id->route.addr.dst_addr, dst_addr, ip_addr_size(dst_addr));
1909 if (cma_any_addr(dst_addr))
1910 ret = cma_resolve_loopback(id_priv);
1912 ret = rdma_resolve_ip(&addr_client, (struct sockaddr *) &id->route.addr.src_addr,
1913 dst_addr, &id->route.addr.dev_addr,
1914 timeout_ms, addr_handler, id_priv);
1920 cma_comp_exch(id_priv, CMA_ADDR_QUERY, CMA_ADDR_BOUND);
1921 cma_deref_id(id_priv);
1924 EXPORT_SYMBOL(rdma_resolve_addr);
1926 static void cma_bind_port(struct rdma_bind_list *bind_list,
1927 struct rdma_id_private *id_priv)
1929 struct sockaddr_in *sin;
1931 sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1932 sin->sin_port = htons(bind_list->port);
1933 id_priv->bind_list = bind_list;
1934 hlist_add_head(&id_priv->node, &bind_list->owners);
1937 static int cma_alloc_port(struct idr *ps, struct rdma_id_private *id_priv,
1938 unsigned short snum)
1940 struct rdma_bind_list *bind_list;
1943 bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
1948 ret = idr_get_new_above(ps, bind_list, snum, &port);
1949 } while ((ret == -EAGAIN) && idr_pre_get(ps, GFP_KERNEL));
1955 ret = -EADDRNOTAVAIL;
1960 bind_list->port = (unsigned short) port;
1961 cma_bind_port(bind_list, id_priv);
1964 idr_remove(ps, port);
1970 static int cma_alloc_any_port(struct idr *ps, struct rdma_id_private *id_priv)
1972 static unsigned int last_used_port;
1973 int low, high, remaining;
1976 inet_get_local_port_range(&low, &high);
1977 remaining = (high - low) + 1;
1978 rover = net_random() % remaining + low;
1980 if (last_used_port != rover &&
1981 !idr_find(ps, (unsigned short) rover)) {
1982 int ret = cma_alloc_port(ps, id_priv, rover);
1984 * Remember previously used port number in order to avoid
1985 * re-using same port immediately after it is closed.
1988 last_used_port = rover;
1989 if (ret != -EADDRNOTAVAIL)
1994 if ((rover < low) || (rover > high))
1998 return -EADDRNOTAVAIL;
2001 static int cma_use_port(struct idr *ps, struct rdma_id_private *id_priv)
2003 struct rdma_id_private *cur_id;
2004 struct sockaddr_in *sin, *cur_sin;
2005 struct rdma_bind_list *bind_list;
2006 struct hlist_node *node;
2007 unsigned short snum;
2009 sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
2010 snum = ntohs(sin->sin_port);
2011 if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
2014 bind_list = idr_find(ps, snum);
2016 return cma_alloc_port(ps, id_priv, snum);
2019 * We don't support binding to any address if anyone is bound to
2020 * a specific address on the same port.
2022 if (cma_any_addr((struct sockaddr *) &id_priv->id.route.addr.src_addr))
2023 return -EADDRNOTAVAIL;
2025 hlist_for_each_entry(cur_id, node, &bind_list->owners, node) {
2026 if (cma_any_addr((struct sockaddr *) &cur_id->id.route.addr.src_addr))
2027 return -EADDRNOTAVAIL;
2029 cur_sin = (struct sockaddr_in *) &cur_id->id.route.addr.src_addr;
2030 if (sin->sin_addr.s_addr == cur_sin->sin_addr.s_addr)
2034 cma_bind_port(bind_list, id_priv);
2038 static int cma_get_port(struct rdma_id_private *id_priv)
2043 switch (id_priv->id.ps) {
2057 return -EPROTONOSUPPORT;
2061 if (cma_any_port((struct sockaddr *) &id_priv->id.route.addr.src_addr))
2062 ret = cma_alloc_any_port(ps, id_priv);
2064 ret = cma_use_port(ps, id_priv);
2065 mutex_unlock(&lock);
2070 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
2071 struct sockaddr *addr)
2073 #if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
2074 struct sockaddr_in6 *sin6;
2076 if (addr->sa_family != AF_INET6)
2079 sin6 = (struct sockaddr_in6 *) addr;
2080 if ((ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL) &&
2081 !sin6->sin6_scope_id)
2084 dev_addr->bound_dev_if = sin6->sin6_scope_id;
2089 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
2091 struct rdma_id_private *id_priv;
2094 if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6)
2095 return -EAFNOSUPPORT;
2097 id_priv = container_of(id, struct rdma_id_private, id);
2098 if (!cma_comp_exch(id_priv, CMA_IDLE, CMA_ADDR_BOUND))
2101 ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
2105 if (!cma_any_addr(addr)) {
2106 ret = rdma_translate_ip(addr, &id->route.addr.dev_addr);
2111 ret = cma_acquire_dev(id_priv);
2112 mutex_unlock(&lock);
2117 memcpy(&id->route.addr.src_addr, addr, ip_addr_size(addr));
2118 ret = cma_get_port(id_priv);
2124 if (id_priv->cma_dev) {
2126 cma_detach_from_dev(id_priv);
2127 mutex_unlock(&lock);
2130 cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_IDLE);
2133 EXPORT_SYMBOL(rdma_bind_addr);
2135 static int cma_format_hdr(void *hdr, enum rdma_port_space ps,
2136 struct rdma_route *route)
2138 struct cma_hdr *cma_hdr;
2139 struct sdp_hh *sdp_hdr;
2141 if (route->addr.src_addr.ss_family == AF_INET) {
2142 struct sockaddr_in *src4, *dst4;
2144 src4 = (struct sockaddr_in *) &route->addr.src_addr;
2145 dst4 = (struct sockaddr_in *) &route->addr.dst_addr;
2150 if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION)
2152 sdp_set_ip_ver(sdp_hdr, 4);
2153 sdp_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2154 sdp_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2155 sdp_hdr->port = src4->sin_port;
2159 cma_hdr->cma_version = CMA_VERSION;
2160 cma_set_ip_ver(cma_hdr, 4);
2161 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2162 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2163 cma_hdr->port = src4->sin_port;
2167 struct sockaddr_in6 *src6, *dst6;
2169 src6 = (struct sockaddr_in6 *) &route->addr.src_addr;
2170 dst6 = (struct sockaddr_in6 *) &route->addr.dst_addr;
2175 if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION)
2177 sdp_set_ip_ver(sdp_hdr, 6);
2178 sdp_hdr->src_addr.ip6 = src6->sin6_addr;
2179 sdp_hdr->dst_addr.ip6 = dst6->sin6_addr;
2180 sdp_hdr->port = src6->sin6_port;
2184 cma_hdr->cma_version = CMA_VERSION;
2185 cma_set_ip_ver(cma_hdr, 6);
2186 cma_hdr->src_addr.ip6 = src6->sin6_addr;
2187 cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
2188 cma_hdr->port = src6->sin6_port;
2195 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
2196 struct ib_cm_event *ib_event)
2198 struct rdma_id_private *id_priv = cm_id->context;
2199 struct rdma_cm_event event;
2200 struct ib_cm_sidr_rep_event_param *rep = &ib_event->param.sidr_rep_rcvd;
2203 if (cma_disable_callback(id_priv, CMA_CONNECT))
2206 memset(&event, 0, sizeof event);
2207 switch (ib_event->event) {
2208 case IB_CM_SIDR_REQ_ERROR:
2209 event.event = RDMA_CM_EVENT_UNREACHABLE;
2210 event.status = -ETIMEDOUT;
2212 case IB_CM_SIDR_REP_RECEIVED:
2213 event.param.ud.private_data = ib_event->private_data;
2214 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
2215 if (rep->status != IB_SIDR_SUCCESS) {
2216 event.event = RDMA_CM_EVENT_UNREACHABLE;
2217 event.status = ib_event->param.sidr_rep_rcvd.status;
2220 ret = cma_set_qkey(id_priv);
2222 event.event = RDMA_CM_EVENT_ADDR_ERROR;
2223 event.status = -EINVAL;
2226 if (id_priv->qkey != rep->qkey) {
2227 event.event = RDMA_CM_EVENT_UNREACHABLE;
2228 event.status = -EINVAL;
2231 ib_init_ah_from_path(id_priv->id.device, id_priv->id.port_num,
2232 id_priv->id.route.path_rec,
2233 &event.param.ud.ah_attr);
2234 event.param.ud.qp_num = rep->qpn;
2235 event.param.ud.qkey = rep->qkey;
2236 event.event = RDMA_CM_EVENT_ESTABLISHED;
2240 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
2245 ret = id_priv->id.event_handler(&id_priv->id, &event);
2247 /* Destroy the CM ID by returning a non-zero value. */
2248 id_priv->cm_id.ib = NULL;
2249 cma_exch(id_priv, CMA_DESTROYING);
2250 mutex_unlock(&id_priv->handler_mutex);
2251 rdma_destroy_id(&id_priv->id);
2255 mutex_unlock(&id_priv->handler_mutex);
2259 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
2260 struct rdma_conn_param *conn_param)
2262 struct ib_cm_sidr_req_param req;
2263 struct rdma_route *route;
2266 req.private_data_len = sizeof(struct cma_hdr) +
2267 conn_param->private_data_len;
2268 req.private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2269 if (!req.private_data)
2272 if (conn_param->private_data && conn_param->private_data_len)
2273 memcpy((void *) req.private_data + sizeof(struct cma_hdr),
2274 conn_param->private_data, conn_param->private_data_len);
2276 route = &id_priv->id.route;
2277 ret = cma_format_hdr((void *) req.private_data, id_priv->id.ps, route);
2281 id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device,
2282 cma_sidr_rep_handler, id_priv);
2283 if (IS_ERR(id_priv->cm_id.ib)) {
2284 ret = PTR_ERR(id_priv->cm_id.ib);
2288 req.path = route->path_rec;
2289 req.service_id = cma_get_service_id(id_priv->id.ps,
2290 (struct sockaddr *) &route->addr.dst_addr);
2291 req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
2292 req.max_cm_retries = CMA_MAX_CM_RETRIES;
2294 ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
2296 ib_destroy_cm_id(id_priv->cm_id.ib);
2297 id_priv->cm_id.ib = NULL;
2300 kfree(req.private_data);
2304 static int cma_connect_ib(struct rdma_id_private *id_priv,
2305 struct rdma_conn_param *conn_param)
2307 struct ib_cm_req_param req;
2308 struct rdma_route *route;
2312 memset(&req, 0, sizeof req);
2313 offset = cma_user_data_offset(id_priv->id.ps);
2314 req.private_data_len = offset + conn_param->private_data_len;
2315 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2319 if (conn_param->private_data && conn_param->private_data_len)
2320 memcpy(private_data + offset, conn_param->private_data,
2321 conn_param->private_data_len);
2323 id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, cma_ib_handler,
2325 if (IS_ERR(id_priv->cm_id.ib)) {
2326 ret = PTR_ERR(id_priv->cm_id.ib);
2330 route = &id_priv->id.route;
2331 ret = cma_format_hdr(private_data, id_priv->id.ps, route);
2334 req.private_data = private_data;
2336 req.primary_path = &route->path_rec[0];
2337 if (route->num_paths == 2)
2338 req.alternate_path = &route->path_rec[1];
2340 req.service_id = cma_get_service_id(id_priv->id.ps,
2341 (struct sockaddr *) &route->addr.dst_addr);
2342 req.qp_num = id_priv->qp_num;
2343 req.qp_type = IB_QPT_RC;
2344 req.starting_psn = id_priv->seq_num;
2345 req.responder_resources = conn_param->responder_resources;
2346 req.initiator_depth = conn_param->initiator_depth;
2347 req.flow_control = conn_param->flow_control;
2348 req.retry_count = conn_param->retry_count;
2349 req.rnr_retry_count = conn_param->rnr_retry_count;
2350 req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2351 req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2352 req.max_cm_retries = CMA_MAX_CM_RETRIES;
2353 req.srq = id_priv->srq ? 1 : 0;
2355 ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
2357 if (ret && !IS_ERR(id_priv->cm_id.ib)) {
2358 ib_destroy_cm_id(id_priv->cm_id.ib);
2359 id_priv->cm_id.ib = NULL;
2362 kfree(private_data);
2366 static int cma_connect_iw(struct rdma_id_private *id_priv,
2367 struct rdma_conn_param *conn_param)
2369 struct iw_cm_id *cm_id;
2370 struct sockaddr_in* sin;
2372 struct iw_cm_conn_param iw_param;
2374 cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
2375 if (IS_ERR(cm_id)) {
2376 ret = PTR_ERR(cm_id);
2380 id_priv->cm_id.iw = cm_id;
2382 sin = (struct sockaddr_in*) &id_priv->id.route.addr.src_addr;
2383 cm_id->local_addr = *sin;
2385 sin = (struct sockaddr_in*) &id_priv->id.route.addr.dst_addr;
2386 cm_id->remote_addr = *sin;
2388 ret = cma_modify_qp_rtr(id_priv, conn_param);
2392 iw_param.ord = conn_param->initiator_depth;
2393 iw_param.ird = conn_param->responder_resources;
2394 iw_param.private_data = conn_param->private_data;
2395 iw_param.private_data_len = conn_param->private_data_len;
2397 iw_param.qpn = id_priv->qp_num;
2399 iw_param.qpn = conn_param->qp_num;
2400 ret = iw_cm_connect(cm_id, &iw_param);
2402 if (ret && !IS_ERR(cm_id)) {
2403 iw_destroy_cm_id(cm_id);
2404 id_priv->cm_id.iw = NULL;
2409 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2411 struct rdma_id_private *id_priv;
2414 id_priv = container_of(id, struct rdma_id_private, id);
2415 if (!cma_comp_exch(id_priv, CMA_ROUTE_RESOLVED, CMA_CONNECT))
2419 id_priv->qp_num = conn_param->qp_num;
2420 id_priv->srq = conn_param->srq;
2423 switch (rdma_node_get_transport(id->device->node_type)) {
2424 case RDMA_TRANSPORT_IB:
2425 if (cma_is_ud_ps(id->ps))
2426 ret = cma_resolve_ib_udp(id_priv, conn_param);
2428 ret = cma_connect_ib(id_priv, conn_param);
2430 case RDMA_TRANSPORT_IWARP:
2431 ret = cma_connect_iw(id_priv, conn_param);
2442 cma_comp_exch(id_priv, CMA_CONNECT, CMA_ROUTE_RESOLVED);
2445 EXPORT_SYMBOL(rdma_connect);
2447 static int cma_accept_ib(struct rdma_id_private *id_priv,
2448 struct rdma_conn_param *conn_param)
2450 struct ib_cm_rep_param rep;
2453 ret = cma_modify_qp_rtr(id_priv, conn_param);
2457 ret = cma_modify_qp_rts(id_priv, conn_param);
2461 memset(&rep, 0, sizeof rep);
2462 rep.qp_num = id_priv->qp_num;
2463 rep.starting_psn = id_priv->seq_num;
2464 rep.private_data = conn_param->private_data;
2465 rep.private_data_len = conn_param->private_data_len;
2466 rep.responder_resources = conn_param->responder_resources;
2467 rep.initiator_depth = conn_param->initiator_depth;
2468 rep.failover_accepted = 0;
2469 rep.flow_control = conn_param->flow_control;
2470 rep.rnr_retry_count = conn_param->rnr_retry_count;
2471 rep.srq = id_priv->srq ? 1 : 0;
2473 ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
2478 static int cma_accept_iw(struct rdma_id_private *id_priv,
2479 struct rdma_conn_param *conn_param)
2481 struct iw_cm_conn_param iw_param;
2484 ret = cma_modify_qp_rtr(id_priv, conn_param);
2488 iw_param.ord = conn_param->initiator_depth;
2489 iw_param.ird = conn_param->responder_resources;
2490 iw_param.private_data = conn_param->private_data;
2491 iw_param.private_data_len = conn_param->private_data_len;
2492 if (id_priv->id.qp) {
2493 iw_param.qpn = id_priv->qp_num;
2495 iw_param.qpn = conn_param->qp_num;
2497 return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
2500 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
2501 enum ib_cm_sidr_status status,
2502 const void *private_data, int private_data_len)
2504 struct ib_cm_sidr_rep_param rep;
2507 memset(&rep, 0, sizeof rep);
2508 rep.status = status;
2509 if (status == IB_SIDR_SUCCESS) {
2510 ret = cma_set_qkey(id_priv);
2513 rep.qp_num = id_priv->qp_num;
2514 rep.qkey = id_priv->qkey;
2516 rep.private_data = private_data;
2517 rep.private_data_len = private_data_len;
2519 return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
2522 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2524 struct rdma_id_private *id_priv;
2527 id_priv = container_of(id, struct rdma_id_private, id);
2528 if (!cma_comp(id_priv, CMA_CONNECT))
2531 if (!id->qp && conn_param) {
2532 id_priv->qp_num = conn_param->qp_num;
2533 id_priv->srq = conn_param->srq;
2536 switch (rdma_node_get_transport(id->device->node_type)) {
2537 case RDMA_TRANSPORT_IB:
2538 if (cma_is_ud_ps(id->ps))
2539 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
2540 conn_param->private_data,
2541 conn_param->private_data_len);
2542 else if (conn_param)
2543 ret = cma_accept_ib(id_priv, conn_param);
2545 ret = cma_rep_recv(id_priv);
2547 case RDMA_TRANSPORT_IWARP:
2548 ret = cma_accept_iw(id_priv, conn_param);
2560 cma_modify_qp_err(id_priv);
2561 rdma_reject(id, NULL, 0);
2564 EXPORT_SYMBOL(rdma_accept);
2566 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
2568 struct rdma_id_private *id_priv;
2571 id_priv = container_of(id, struct rdma_id_private, id);
2572 if (!cma_has_cm_dev(id_priv))
2575 switch (id->device->node_type) {
2576 case RDMA_NODE_IB_CA:
2577 ret = ib_cm_notify(id_priv->cm_id.ib, event);
2585 EXPORT_SYMBOL(rdma_notify);
2587 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
2588 u8 private_data_len)
2590 struct rdma_id_private *id_priv;
2593 id_priv = container_of(id, struct rdma_id_private, id);
2594 if (!cma_has_cm_dev(id_priv))
2597 switch (rdma_node_get_transport(id->device->node_type)) {
2598 case RDMA_TRANSPORT_IB:
2599 if (cma_is_ud_ps(id->ps))
2600 ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT,
2601 private_data, private_data_len);
2603 ret = ib_send_cm_rej(id_priv->cm_id.ib,
2604 IB_CM_REJ_CONSUMER_DEFINED, NULL,
2605 0, private_data, private_data_len);
2607 case RDMA_TRANSPORT_IWARP:
2608 ret = iw_cm_reject(id_priv->cm_id.iw,
2609 private_data, private_data_len);
2617 EXPORT_SYMBOL(rdma_reject);
2619 int rdma_disconnect(struct rdma_cm_id *id)
2621 struct rdma_id_private *id_priv;
2624 id_priv = container_of(id, struct rdma_id_private, id);
2625 if (!cma_has_cm_dev(id_priv))
2628 switch (rdma_node_get_transport(id->device->node_type)) {
2629 case RDMA_TRANSPORT_IB:
2630 ret = cma_modify_qp_err(id_priv);
2633 /* Initiate or respond to a disconnect. */
2634 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
2635 ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
2637 case RDMA_TRANSPORT_IWARP:
2638 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
2647 EXPORT_SYMBOL(rdma_disconnect);
2649 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
2651 struct rdma_id_private *id_priv;
2652 struct cma_multicast *mc = multicast->context;
2653 struct rdma_cm_event event;
2656 id_priv = mc->id_priv;
2657 if (cma_disable_callback(id_priv, CMA_ADDR_BOUND) &&
2658 cma_disable_callback(id_priv, CMA_ADDR_RESOLVED))
2661 mutex_lock(&id_priv->qp_mutex);
2662 if (!status && id_priv->id.qp)
2663 status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid,
2664 multicast->rec.mlid);
2665 mutex_unlock(&id_priv->qp_mutex);
2667 memset(&event, 0, sizeof event);
2668 event.status = status;
2669 event.param.ud.private_data = mc->context;
2671 event.event = RDMA_CM_EVENT_MULTICAST_JOIN;
2672 ib_init_ah_from_mcmember(id_priv->id.device,
2673 id_priv->id.port_num, &multicast->rec,
2674 &event.param.ud.ah_attr);
2675 event.param.ud.qp_num = 0xFFFFFF;
2676 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
2678 event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
2680 ret = id_priv->id.event_handler(&id_priv->id, &event);
2682 cma_exch(id_priv, CMA_DESTROYING);
2683 mutex_unlock(&id_priv->handler_mutex);
2684 rdma_destroy_id(&id_priv->id);
2688 mutex_unlock(&id_priv->handler_mutex);
2692 static void cma_set_mgid(struct rdma_id_private *id_priv,
2693 struct sockaddr *addr, union ib_gid *mgid)
2695 unsigned char mc_map[MAX_ADDR_LEN];
2696 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2697 struct sockaddr_in *sin = (struct sockaddr_in *) addr;
2698 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
2700 if (cma_any_addr(addr)) {
2701 memset(mgid, 0, sizeof *mgid);
2702 } else if ((addr->sa_family == AF_INET6) &&
2703 ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
2705 /* IPv6 address is an SA assigned MGID. */
2706 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
2707 } else if ((addr->sa_family == AF_INET6)) {
2708 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
2709 if (id_priv->id.ps == RDMA_PS_UDP)
2710 mc_map[7] = 0x01; /* Use RDMA CM signature */
2711 *mgid = *(union ib_gid *) (mc_map + 4);
2713 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
2714 if (id_priv->id.ps == RDMA_PS_UDP)
2715 mc_map[7] = 0x01; /* Use RDMA CM signature */
2716 *mgid = *(union ib_gid *) (mc_map + 4);
2720 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
2721 struct cma_multicast *mc)
2723 struct ib_sa_mcmember_rec rec;
2724 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2725 ib_sa_comp_mask comp_mask;
2728 ib_addr_get_mgid(dev_addr, &rec.mgid);
2729 ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
2734 cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
2735 if (id_priv->id.ps == RDMA_PS_UDP)
2736 rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
2737 rdma_addr_get_sgid(dev_addr, &rec.port_gid);
2738 rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
2741 comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
2742 IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
2743 IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
2744 IB_SA_MCMEMBER_REC_FLOW_LABEL |
2745 IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
2747 if (id_priv->id.ps == RDMA_PS_IPOIB)
2748 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
2749 IB_SA_MCMEMBER_REC_RATE_SELECTOR;
2751 mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device,
2752 id_priv->id.port_num, &rec,
2753 comp_mask, GFP_KERNEL,
2754 cma_ib_mc_handler, mc);
2755 if (IS_ERR(mc->multicast.ib))
2756 return PTR_ERR(mc->multicast.ib);
2761 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
2764 struct rdma_id_private *id_priv;
2765 struct cma_multicast *mc;
2768 id_priv = container_of(id, struct rdma_id_private, id);
2769 if (!cma_comp(id_priv, CMA_ADDR_BOUND) &&
2770 !cma_comp(id_priv, CMA_ADDR_RESOLVED))
2773 mc = kmalloc(sizeof *mc, GFP_KERNEL);
2777 memcpy(&mc->addr, addr, ip_addr_size(addr));
2778 mc->context = context;
2779 mc->id_priv = id_priv;
2781 spin_lock(&id_priv->lock);
2782 list_add(&mc->list, &id_priv->mc_list);
2783 spin_unlock(&id_priv->lock);
2785 switch (rdma_node_get_transport(id->device->node_type)) {
2786 case RDMA_TRANSPORT_IB:
2787 ret = cma_join_ib_multicast(id_priv, mc);
2795 spin_lock_irq(&id_priv->lock);
2796 list_del(&mc->list);
2797 spin_unlock_irq(&id_priv->lock);
2802 EXPORT_SYMBOL(rdma_join_multicast);
2804 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
2806 struct rdma_id_private *id_priv;
2807 struct cma_multicast *mc;
2809 id_priv = container_of(id, struct rdma_id_private, id);
2810 spin_lock_irq(&id_priv->lock);
2811 list_for_each_entry(mc, &id_priv->mc_list, list) {
2812 if (!memcmp(&mc->addr, addr, ip_addr_size(addr))) {
2813 list_del(&mc->list);
2814 spin_unlock_irq(&id_priv->lock);
2817 ib_detach_mcast(id->qp,
2818 &mc->multicast.ib->rec.mgid,
2819 mc->multicast.ib->rec.mlid);
2820 ib_sa_free_multicast(mc->multicast.ib);
2825 spin_unlock_irq(&id_priv->lock);
2827 EXPORT_SYMBOL(rdma_leave_multicast);
2829 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
2831 struct rdma_dev_addr *dev_addr;
2832 struct cma_ndev_work *work;
2834 dev_addr = &id_priv->id.route.addr.dev_addr;
2836 if ((dev_addr->bound_dev_if == ndev->ifindex) &&
2837 memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
2838 printk(KERN_INFO "RDMA CM addr change for ndev %s used by id %p\n",
2839 ndev->name, &id_priv->id);
2840 work = kzalloc(sizeof *work, GFP_KERNEL);
2844 INIT_WORK(&work->work, cma_ndev_work_handler);
2846 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
2847 atomic_inc(&id_priv->refcount);
2848 queue_work(cma_wq, &work->work);
2854 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
2857 struct net_device *ndev = (struct net_device *)ctx;
2858 struct cma_device *cma_dev;
2859 struct rdma_id_private *id_priv;
2860 int ret = NOTIFY_DONE;
2862 if (dev_net(ndev) != &init_net)
2865 if (event != NETDEV_BONDING_FAILOVER)
2868 if (!(ndev->flags & IFF_MASTER) || !(ndev->priv_flags & IFF_BONDING))
2872 list_for_each_entry(cma_dev, &dev_list, list)
2873 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
2874 ret = cma_netdev_change(ndev, id_priv);
2880 mutex_unlock(&lock);
2884 static struct notifier_block cma_nb = {
2885 .notifier_call = cma_netdev_callback
2888 static void cma_add_one(struct ib_device *device)
2890 struct cma_device *cma_dev;
2891 struct rdma_id_private *id_priv;
2893 cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
2897 cma_dev->device = device;
2899 init_completion(&cma_dev->comp);
2900 atomic_set(&cma_dev->refcount, 1);
2901 INIT_LIST_HEAD(&cma_dev->id_list);
2902 ib_set_client_data(device, &cma_client, cma_dev);
2905 list_add_tail(&cma_dev->list, &dev_list);
2906 list_for_each_entry(id_priv, &listen_any_list, list)
2907 cma_listen_on_dev(id_priv, cma_dev);
2908 mutex_unlock(&lock);
2911 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
2913 struct rdma_cm_event event;
2914 enum cma_state state;
2917 /* Record that we want to remove the device */
2918 state = cma_exch(id_priv, CMA_DEVICE_REMOVAL);
2919 if (state == CMA_DESTROYING)
2922 cma_cancel_operation(id_priv, state);
2923 mutex_lock(&id_priv->handler_mutex);
2925 /* Check for destruction from another callback. */
2926 if (!cma_comp(id_priv, CMA_DEVICE_REMOVAL))
2929 memset(&event, 0, sizeof event);
2930 event.event = RDMA_CM_EVENT_DEVICE_REMOVAL;
2931 ret = id_priv->id.event_handler(&id_priv->id, &event);
2933 mutex_unlock(&id_priv->handler_mutex);
2937 static void cma_process_remove(struct cma_device *cma_dev)
2939 struct rdma_id_private *id_priv;
2943 while (!list_empty(&cma_dev->id_list)) {
2944 id_priv = list_entry(cma_dev->id_list.next,
2945 struct rdma_id_private, list);
2947 list_del(&id_priv->listen_list);
2948 list_del_init(&id_priv->list);
2949 atomic_inc(&id_priv->refcount);
2950 mutex_unlock(&lock);
2952 ret = id_priv->internal_id ? 1 : cma_remove_id_dev(id_priv);
2953 cma_deref_id(id_priv);
2955 rdma_destroy_id(&id_priv->id);
2959 mutex_unlock(&lock);
2961 cma_deref_dev(cma_dev);
2962 wait_for_completion(&cma_dev->comp);
2965 static void cma_remove_one(struct ib_device *device)
2967 struct cma_device *cma_dev;
2969 cma_dev = ib_get_client_data(device, &cma_client);
2974 list_del(&cma_dev->list);
2975 mutex_unlock(&lock);
2977 cma_process_remove(cma_dev);
2981 static int __init cma_init(void)
2985 cma_wq = create_singlethread_workqueue("rdma_cm");
2989 ib_sa_register_client(&sa_client);
2990 rdma_addr_register_client(&addr_client);
2991 register_netdevice_notifier(&cma_nb);
2993 ret = ib_register_client(&cma_client);
2999 unregister_netdevice_notifier(&cma_nb);
3000 rdma_addr_unregister_client(&addr_client);
3001 ib_sa_unregister_client(&sa_client);
3002 destroy_workqueue(cma_wq);
3006 static void __exit cma_cleanup(void)
3008 ib_unregister_client(&cma_client);
3009 unregister_netdevice_notifier(&cma_nb);
3010 rdma_addr_unregister_client(&addr_client);
3011 ib_sa_unregister_client(&sa_client);
3012 destroy_workqueue(cma_wq);
3013 idr_destroy(&sdp_ps);
3014 idr_destroy(&tcp_ps);
3015 idr_destroy(&udp_ps);
3016 idr_destroy(&ipoib_ps);
3019 module_init(cma_init);
3020 module_exit(cma_cleanup);