2 * Copyright (c) 2005 Voltaire Inc. All rights reserved.
3 * Copyright (c) 2002-2005, Network Appliance, Inc. All rights reserved.
4 * Copyright (c) 1999-2005, Mellanox Technologies, Inc. All rights reserved.
5 * Copyright (c) 2005-2006 Intel Corporation. All rights reserved.
7 * This software is available to you under a choice of one of two
8 * licenses. You may choose to be licensed under the terms of the GNU
9 * General Public License (GPL) Version 2, available from the file
10 * COPYING in the main directory of this source tree, or the
11 * OpenIB.org BSD license below:
13 * Redistribution and use in source and binary forms, with or
14 * without modification, are permitted provided that the following
17 * - Redistributions of source code must retain the above
18 * copyright notice, this list of conditions and the following
21 * - Redistributions in binary form must reproduce the above
22 * copyright notice, this list of conditions and the following
23 * disclaimer in the documentation and/or other materials
24 * provided with the distribution.
26 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
27 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
28 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
29 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
30 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
31 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
32 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
36 #include <linux/completion.h>
38 #include <linux/in6.h>
39 #include <linux/mutex.h>
40 #include <linux/random.h>
41 #include <linux/idr.h>
42 #include <linux/inetdevice.h>
43 #include <linux/slab.h>
44 #include <linux/module.h>
45 #include <net/route.h>
50 #include <rdma/rdma_cm.h>
51 #include <rdma/rdma_cm_ib.h>
52 #include <rdma/rdma_netlink.h>
54 #include <rdma/ib_cache.h>
55 #include <rdma/ib_cm.h>
56 #include <rdma/ib_sa.h>
57 #include <rdma/iw_cm.h>
59 MODULE_AUTHOR("Sean Hefty");
60 MODULE_DESCRIPTION("Generic RDMA CM Agent");
61 MODULE_LICENSE("Dual BSD/GPL");
63 #define CMA_CM_RESPONSE_TIMEOUT 20
64 #define CMA_MAX_CM_RETRIES 15
65 #define CMA_CM_MRA_SETTING (IB_CM_MRA_FLAG_DELAY | 24)
66 #define CMA_IBOE_PACKET_LIFETIME 18
68 static void cma_add_one(struct ib_device *device);
69 static void cma_remove_one(struct ib_device *device);
71 static struct ib_client cma_client = {
74 .remove = cma_remove_one
77 static struct ib_sa_client sa_client;
78 static struct rdma_addr_client addr_client;
79 static LIST_HEAD(dev_list);
80 static LIST_HEAD(listen_any_list);
81 static DEFINE_MUTEX(lock);
82 static struct workqueue_struct *cma_wq;
83 static DEFINE_IDR(tcp_ps);
84 static DEFINE_IDR(udp_ps);
85 static DEFINE_IDR(ipoib_ps);
86 static DEFINE_IDR(ib_ps);
89 struct list_head list;
90 struct ib_device *device;
91 struct completion comp;
93 struct list_head id_list;
96 struct rdma_bind_list {
98 struct hlist_head owners;
107 * Device removal can occur at anytime, so we need extra handling to
108 * serialize notifying the user of device removal with other callbacks.
109 * We do this by disabling removal notification while a callback is in process,
110 * and reporting it after the callback completes.
112 struct rdma_id_private {
113 struct rdma_cm_id id;
115 struct rdma_bind_list *bind_list;
116 struct hlist_node node;
117 struct list_head list; /* listen_any_list or cma_device.list */
118 struct list_head listen_list; /* per device listens */
119 struct cma_device *cma_dev;
120 struct list_head mc_list;
123 enum rdma_cm_state state;
125 struct mutex qp_mutex;
127 struct completion comp;
129 struct mutex handler_mutex;
133 struct ib_sa_query *query;
151 struct cma_multicast {
152 struct rdma_id_private *id_priv;
154 struct ib_sa_multicast *ib;
156 struct list_head list;
158 struct sockaddr_storage addr;
163 struct work_struct work;
164 struct rdma_id_private *id;
165 enum rdma_cm_state old_state;
166 enum rdma_cm_state new_state;
167 struct rdma_cm_event event;
170 struct cma_ndev_work {
171 struct work_struct work;
172 struct rdma_id_private *id;
173 struct rdma_cm_event event;
176 struct iboe_mcast_work {
177 struct work_struct work;
178 struct rdma_id_private *id;
179 struct cma_multicast *mc;
192 u8 ip_version; /* IP version: 7:4 */
194 union cma_ip_addr src_addr;
195 union cma_ip_addr dst_addr;
198 #define CMA_VERSION 0x00
200 static int cma_comp(struct rdma_id_private *id_priv, enum rdma_cm_state comp)
205 spin_lock_irqsave(&id_priv->lock, flags);
206 ret = (id_priv->state == comp);
207 spin_unlock_irqrestore(&id_priv->lock, flags);
211 static int cma_comp_exch(struct rdma_id_private *id_priv,
212 enum rdma_cm_state comp, enum rdma_cm_state exch)
217 spin_lock_irqsave(&id_priv->lock, flags);
218 if ((ret = (id_priv->state == comp)))
219 id_priv->state = exch;
220 spin_unlock_irqrestore(&id_priv->lock, flags);
224 static enum rdma_cm_state cma_exch(struct rdma_id_private *id_priv,
225 enum rdma_cm_state exch)
228 enum rdma_cm_state old;
230 spin_lock_irqsave(&id_priv->lock, flags);
231 old = id_priv->state;
232 id_priv->state = exch;
233 spin_unlock_irqrestore(&id_priv->lock, flags);
237 static inline u8 cma_get_ip_ver(struct cma_hdr *hdr)
239 return hdr->ip_version >> 4;
242 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
244 hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
247 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
248 struct cma_device *cma_dev)
250 atomic_inc(&cma_dev->refcount);
251 id_priv->cma_dev = cma_dev;
252 id_priv->id.device = cma_dev->device;
253 id_priv->id.route.addr.dev_addr.transport =
254 rdma_node_get_transport(cma_dev->device->node_type);
255 list_add_tail(&id_priv->list, &cma_dev->id_list);
258 static inline void cma_deref_dev(struct cma_device *cma_dev)
260 if (atomic_dec_and_test(&cma_dev->refcount))
261 complete(&cma_dev->comp);
264 static inline void release_mc(struct kref *kref)
266 struct cma_multicast *mc = container_of(kref, struct cma_multicast, mcref);
268 kfree(mc->multicast.ib);
272 static void cma_release_dev(struct rdma_id_private *id_priv)
275 list_del(&id_priv->list);
276 cma_deref_dev(id_priv->cma_dev);
277 id_priv->cma_dev = NULL;
281 static inline struct sockaddr *cma_src_addr(struct rdma_id_private *id_priv)
283 return (struct sockaddr *) &id_priv->id.route.addr.src_addr;
286 static inline struct sockaddr *cma_dst_addr(struct rdma_id_private *id_priv)
288 return (struct sockaddr *) &id_priv->id.route.addr.dst_addr;
291 static inline unsigned short cma_family(struct rdma_id_private *id_priv)
293 return id_priv->id.route.addr.src_addr.ss_family;
296 static int cma_set_qkey(struct rdma_id_private *id_priv, u32 qkey)
298 struct ib_sa_mcmember_rec rec;
302 if (qkey && id_priv->qkey != qkey)
308 id_priv->qkey = qkey;
312 switch (id_priv->id.ps) {
315 id_priv->qkey = RDMA_UDP_QKEY;
318 ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid);
319 ret = ib_sa_get_mcmember_rec(id_priv->id.device,
320 id_priv->id.port_num, &rec.mgid,
323 id_priv->qkey = be32_to_cpu(rec.qkey);
331 static void cma_translate_ib(struct sockaddr_ib *sib, struct rdma_dev_addr *dev_addr)
333 dev_addr->dev_type = ARPHRD_INFINIBAND;
334 rdma_addr_set_sgid(dev_addr, (union ib_gid *) &sib->sib_addr);
335 ib_addr_set_pkey(dev_addr, ntohs(sib->sib_pkey));
338 static int cma_translate_addr(struct sockaddr *addr, struct rdma_dev_addr *dev_addr)
342 if (addr->sa_family != AF_IB) {
343 ret = rdma_translate_ip(addr, dev_addr, NULL);
345 cma_translate_ib((struct sockaddr_ib *) addr, dev_addr);
352 static int cma_acquire_dev(struct rdma_id_private *id_priv,
353 struct rdma_id_private *listen_id_priv)
355 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
356 struct cma_device *cma_dev;
357 union ib_gid gid, iboe_gid;
360 enum rdma_link_layer dev_ll = dev_addr->dev_type == ARPHRD_INFINIBAND ?
361 IB_LINK_LAYER_INFINIBAND : IB_LINK_LAYER_ETHERNET;
363 if (dev_ll != IB_LINK_LAYER_INFINIBAND &&
364 id_priv->id.ps == RDMA_PS_IPOIB)
368 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
371 memcpy(&gid, dev_addr->src_dev_addr +
372 rdma_addr_gid_offset(dev_addr), sizeof gid);
373 if (listen_id_priv &&
374 rdma_port_get_link_layer(listen_id_priv->id.device,
375 listen_id_priv->id.port_num) == dev_ll) {
376 cma_dev = listen_id_priv->cma_dev;
377 port = listen_id_priv->id.port_num;
378 if (rdma_node_get_transport(cma_dev->device->node_type) == RDMA_TRANSPORT_IB &&
379 rdma_port_get_link_layer(cma_dev->device, port) == IB_LINK_LAYER_ETHERNET)
380 ret = ib_find_cached_gid(cma_dev->device, &iboe_gid,
383 ret = ib_find_cached_gid(cma_dev->device, &gid,
386 if (!ret && (port == found_port)) {
387 id_priv->id.port_num = found_port;
391 list_for_each_entry(cma_dev, &dev_list, list) {
392 for (port = 1; port <= cma_dev->device->phys_port_cnt; ++port) {
393 if (listen_id_priv &&
394 listen_id_priv->cma_dev == cma_dev &&
395 listen_id_priv->id.port_num == port)
397 if (rdma_port_get_link_layer(cma_dev->device, port) == dev_ll) {
398 if (rdma_node_get_transport(cma_dev->device->node_type) == RDMA_TRANSPORT_IB &&
399 rdma_port_get_link_layer(cma_dev->device, port) == IB_LINK_LAYER_ETHERNET)
400 ret = ib_find_cached_gid(cma_dev->device, &iboe_gid, &found_port, NULL);
402 ret = ib_find_cached_gid(cma_dev->device, &gid, &found_port, NULL);
404 if (!ret && (port == found_port)) {
405 id_priv->id.port_num = found_port;
414 cma_attach_to_dev(id_priv, cma_dev);
421 * Select the source IB device and address to reach the destination IB address.
423 static int cma_resolve_ib_dev(struct rdma_id_private *id_priv)
425 struct cma_device *cma_dev, *cur_dev;
426 struct sockaddr_ib *addr;
427 union ib_gid gid, sgid, *dgid;
433 addr = (struct sockaddr_ib *) cma_dst_addr(id_priv);
434 dgid = (union ib_gid *) &addr->sib_addr;
435 pkey = ntohs(addr->sib_pkey);
437 list_for_each_entry(cur_dev, &dev_list, list) {
438 if (rdma_node_get_transport(cur_dev->device->node_type) != RDMA_TRANSPORT_IB)
441 for (p = 1; p <= cur_dev->device->phys_port_cnt; ++p) {
442 if (ib_find_cached_pkey(cur_dev->device, p, pkey, &index))
445 for (i = 0; !ib_get_cached_gid(cur_dev->device, p, i, &gid); i++) {
446 if (!memcmp(&gid, dgid, sizeof(gid))) {
449 id_priv->id.port_num = p;
453 if (!cma_dev && (gid.global.subnet_prefix ==
454 dgid->global.subnet_prefix)) {
457 id_priv->id.port_num = p;
467 cma_attach_to_dev(id_priv, cma_dev);
468 addr = (struct sockaddr_ib *) cma_src_addr(id_priv);
469 memcpy(&addr->sib_addr, &sgid, sizeof sgid);
470 cma_translate_ib(addr, &id_priv->id.route.addr.dev_addr);
474 static void cma_deref_id(struct rdma_id_private *id_priv)
476 if (atomic_dec_and_test(&id_priv->refcount))
477 complete(&id_priv->comp);
480 static int cma_disable_callback(struct rdma_id_private *id_priv,
481 enum rdma_cm_state state)
483 mutex_lock(&id_priv->handler_mutex);
484 if (id_priv->state != state) {
485 mutex_unlock(&id_priv->handler_mutex);
491 struct rdma_cm_id *rdma_create_id(rdma_cm_event_handler event_handler,
492 void *context, enum rdma_port_space ps,
493 enum ib_qp_type qp_type)
495 struct rdma_id_private *id_priv;
497 id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
499 return ERR_PTR(-ENOMEM);
501 id_priv->owner = task_pid_nr(current);
502 id_priv->state = RDMA_CM_IDLE;
503 id_priv->id.context = context;
504 id_priv->id.event_handler = event_handler;
506 id_priv->id.qp_type = qp_type;
507 spin_lock_init(&id_priv->lock);
508 mutex_init(&id_priv->qp_mutex);
509 init_completion(&id_priv->comp);
510 atomic_set(&id_priv->refcount, 1);
511 mutex_init(&id_priv->handler_mutex);
512 INIT_LIST_HEAD(&id_priv->listen_list);
513 INIT_LIST_HEAD(&id_priv->mc_list);
514 get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
518 EXPORT_SYMBOL(rdma_create_id);
520 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
522 struct ib_qp_attr qp_attr;
523 int qp_attr_mask, ret;
525 qp_attr.qp_state = IB_QPS_INIT;
526 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
530 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
534 qp_attr.qp_state = IB_QPS_RTR;
535 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
539 qp_attr.qp_state = IB_QPS_RTS;
541 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
546 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
548 struct ib_qp_attr qp_attr;
549 int qp_attr_mask, ret;
551 qp_attr.qp_state = IB_QPS_INIT;
552 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
556 return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
559 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
560 struct ib_qp_init_attr *qp_init_attr)
562 struct rdma_id_private *id_priv;
566 id_priv = container_of(id, struct rdma_id_private, id);
567 if (id->device != pd->device)
570 qp = ib_create_qp(pd, qp_init_attr);
574 if (id->qp_type == IB_QPT_UD)
575 ret = cma_init_ud_qp(id_priv, qp);
577 ret = cma_init_conn_qp(id_priv, qp);
582 id_priv->qp_num = qp->qp_num;
583 id_priv->srq = (qp->srq != NULL);
589 EXPORT_SYMBOL(rdma_create_qp);
591 void rdma_destroy_qp(struct rdma_cm_id *id)
593 struct rdma_id_private *id_priv;
595 id_priv = container_of(id, struct rdma_id_private, id);
596 mutex_lock(&id_priv->qp_mutex);
597 ib_destroy_qp(id_priv->id.qp);
598 id_priv->id.qp = NULL;
599 mutex_unlock(&id_priv->qp_mutex);
601 EXPORT_SYMBOL(rdma_destroy_qp);
603 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
604 struct rdma_conn_param *conn_param)
606 struct ib_qp_attr qp_attr;
607 int qp_attr_mask, ret;
610 mutex_lock(&id_priv->qp_mutex);
611 if (!id_priv->id.qp) {
616 /* Need to update QP attributes from default values. */
617 qp_attr.qp_state = IB_QPS_INIT;
618 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
622 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
626 qp_attr.qp_state = IB_QPS_RTR;
627 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
631 ret = ib_query_gid(id_priv->id.device, id_priv->id.port_num,
632 qp_attr.ah_attr.grh.sgid_index, &sgid);
636 if (rdma_node_get_transport(id_priv->cma_dev->device->node_type)
637 == RDMA_TRANSPORT_IB &&
638 rdma_port_get_link_layer(id_priv->id.device, id_priv->id.port_num)
639 == IB_LINK_LAYER_ETHERNET) {
640 ret = rdma_addr_find_smac_by_sgid(&sgid, qp_attr.smac, NULL);
646 qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
647 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
649 mutex_unlock(&id_priv->qp_mutex);
653 static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
654 struct rdma_conn_param *conn_param)
656 struct ib_qp_attr qp_attr;
657 int qp_attr_mask, ret;
659 mutex_lock(&id_priv->qp_mutex);
660 if (!id_priv->id.qp) {
665 qp_attr.qp_state = IB_QPS_RTS;
666 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
671 qp_attr.max_rd_atomic = conn_param->initiator_depth;
672 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
674 mutex_unlock(&id_priv->qp_mutex);
678 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
680 struct ib_qp_attr qp_attr;
683 mutex_lock(&id_priv->qp_mutex);
684 if (!id_priv->id.qp) {
689 qp_attr.qp_state = IB_QPS_ERR;
690 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
692 mutex_unlock(&id_priv->qp_mutex);
696 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
697 struct ib_qp_attr *qp_attr, int *qp_attr_mask)
699 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
703 if (rdma_port_get_link_layer(id_priv->id.device, id_priv->id.port_num) ==
704 IB_LINK_LAYER_INFINIBAND)
705 pkey = ib_addr_get_pkey(dev_addr);
709 ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
710 pkey, &qp_attr->pkey_index);
714 qp_attr->port_num = id_priv->id.port_num;
715 *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
717 if (id_priv->id.qp_type == IB_QPT_UD) {
718 ret = cma_set_qkey(id_priv, 0);
722 qp_attr->qkey = id_priv->qkey;
723 *qp_attr_mask |= IB_QP_QKEY;
725 qp_attr->qp_access_flags = 0;
726 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
731 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
734 struct rdma_id_private *id_priv;
737 id_priv = container_of(id, struct rdma_id_private, id);
738 switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
739 case RDMA_TRANSPORT_IB:
740 if (!id_priv->cm_id.ib || (id_priv->id.qp_type == IB_QPT_UD))
741 ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
743 ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
746 if (qp_attr->qp_state == IB_QPS_RTR)
747 qp_attr->rq_psn = id_priv->seq_num;
749 case RDMA_TRANSPORT_IWARP:
750 if (!id_priv->cm_id.iw) {
751 qp_attr->qp_access_flags = 0;
752 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
754 ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
764 EXPORT_SYMBOL(rdma_init_qp_attr);
766 static inline int cma_zero_addr(struct sockaddr *addr)
768 switch (addr->sa_family) {
770 return ipv4_is_zeronet(((struct sockaddr_in *)addr)->sin_addr.s_addr);
772 return ipv6_addr_any(&((struct sockaddr_in6 *) addr)->sin6_addr);
774 return ib_addr_any(&((struct sockaddr_ib *) addr)->sib_addr);
780 static inline int cma_loopback_addr(struct sockaddr *addr)
782 switch (addr->sa_family) {
784 return ipv4_is_loopback(((struct sockaddr_in *) addr)->sin_addr.s_addr);
786 return ipv6_addr_loopback(&((struct sockaddr_in6 *) addr)->sin6_addr);
788 return ib_addr_loopback(&((struct sockaddr_ib *) addr)->sib_addr);
794 static inline int cma_any_addr(struct sockaddr *addr)
796 return cma_zero_addr(addr) || cma_loopback_addr(addr);
799 static int cma_addr_cmp(struct sockaddr *src, struct sockaddr *dst)
801 if (src->sa_family != dst->sa_family)
804 switch (src->sa_family) {
806 return ((struct sockaddr_in *) src)->sin_addr.s_addr !=
807 ((struct sockaddr_in *) dst)->sin_addr.s_addr;
809 return ipv6_addr_cmp(&((struct sockaddr_in6 *) src)->sin6_addr,
810 &((struct sockaddr_in6 *) dst)->sin6_addr);
812 return ib_addr_cmp(&((struct sockaddr_ib *) src)->sib_addr,
813 &((struct sockaddr_ib *) dst)->sib_addr);
817 static __be16 cma_port(struct sockaddr *addr)
819 struct sockaddr_ib *sib;
821 switch (addr->sa_family) {
823 return ((struct sockaddr_in *) addr)->sin_port;
825 return ((struct sockaddr_in6 *) addr)->sin6_port;
827 sib = (struct sockaddr_ib *) addr;
828 return htons((u16) (be64_to_cpu(sib->sib_sid) &
829 be64_to_cpu(sib->sib_sid_mask)));
835 static inline int cma_any_port(struct sockaddr *addr)
837 return !cma_port(addr);
840 static void cma_save_ib_info(struct rdma_cm_id *id, struct rdma_cm_id *listen_id,
841 struct ib_sa_path_rec *path)
843 struct sockaddr_ib *listen_ib, *ib;
845 listen_ib = (struct sockaddr_ib *) &listen_id->route.addr.src_addr;
846 ib = (struct sockaddr_ib *) &id->route.addr.src_addr;
847 ib->sib_family = listen_ib->sib_family;
848 ib->sib_pkey = path->pkey;
849 ib->sib_flowinfo = path->flow_label;
850 memcpy(&ib->sib_addr, &path->sgid, 16);
851 ib->sib_sid = listen_ib->sib_sid;
852 ib->sib_sid_mask = cpu_to_be64(0xffffffffffffffffULL);
853 ib->sib_scope_id = listen_ib->sib_scope_id;
855 ib = (struct sockaddr_ib *) &id->route.addr.dst_addr;
856 ib->sib_family = listen_ib->sib_family;
857 ib->sib_pkey = path->pkey;
858 ib->sib_flowinfo = path->flow_label;
859 memcpy(&ib->sib_addr, &path->dgid, 16);
862 static void cma_save_ip4_info(struct rdma_cm_id *id, struct rdma_cm_id *listen_id,
865 struct sockaddr_in *listen4, *ip4;
867 listen4 = (struct sockaddr_in *) &listen_id->route.addr.src_addr;
868 ip4 = (struct sockaddr_in *) &id->route.addr.src_addr;
869 ip4->sin_family = listen4->sin_family;
870 ip4->sin_addr.s_addr = hdr->dst_addr.ip4.addr;
871 ip4->sin_port = listen4->sin_port;
873 ip4 = (struct sockaddr_in *) &id->route.addr.dst_addr;
874 ip4->sin_family = listen4->sin_family;
875 ip4->sin_addr.s_addr = hdr->src_addr.ip4.addr;
876 ip4->sin_port = hdr->port;
879 static void cma_save_ip6_info(struct rdma_cm_id *id, struct rdma_cm_id *listen_id,
882 struct sockaddr_in6 *listen6, *ip6;
884 listen6 = (struct sockaddr_in6 *) &listen_id->route.addr.src_addr;
885 ip6 = (struct sockaddr_in6 *) &id->route.addr.src_addr;
886 ip6->sin6_family = listen6->sin6_family;
887 ip6->sin6_addr = hdr->dst_addr.ip6;
888 ip6->sin6_port = listen6->sin6_port;
890 ip6 = (struct sockaddr_in6 *) &id->route.addr.dst_addr;
891 ip6->sin6_family = listen6->sin6_family;
892 ip6->sin6_addr = hdr->src_addr.ip6;
893 ip6->sin6_port = hdr->port;
896 static int cma_save_net_info(struct rdma_cm_id *id, struct rdma_cm_id *listen_id,
897 struct ib_cm_event *ib_event)
901 if ((listen_id->route.addr.src_addr.ss_family == AF_IB) &&
902 (ib_event->event == IB_CM_REQ_RECEIVED)) {
903 cma_save_ib_info(id, listen_id, ib_event->param.req_rcvd.primary_path);
907 hdr = ib_event->private_data;
908 if (hdr->cma_version != CMA_VERSION)
911 switch (cma_get_ip_ver(hdr)) {
913 cma_save_ip4_info(id, listen_id, hdr);
916 cma_save_ip6_info(id, listen_id, hdr);
924 static inline int cma_user_data_offset(struct rdma_id_private *id_priv)
926 return cma_family(id_priv) == AF_IB ? 0 : sizeof(struct cma_hdr);
929 static void cma_cancel_route(struct rdma_id_private *id_priv)
931 switch (rdma_port_get_link_layer(id_priv->id.device, id_priv->id.port_num)) {
932 case IB_LINK_LAYER_INFINIBAND:
934 ib_sa_cancel_query(id_priv->query_id, id_priv->query);
941 static void cma_cancel_listens(struct rdma_id_private *id_priv)
943 struct rdma_id_private *dev_id_priv;
946 * Remove from listen_any_list to prevent added devices from spawning
947 * additional listen requests.
950 list_del(&id_priv->list);
952 while (!list_empty(&id_priv->listen_list)) {
953 dev_id_priv = list_entry(id_priv->listen_list.next,
954 struct rdma_id_private, listen_list);
955 /* sync with device removal to avoid duplicate destruction */
956 list_del_init(&dev_id_priv->list);
957 list_del(&dev_id_priv->listen_list);
960 rdma_destroy_id(&dev_id_priv->id);
966 static void cma_cancel_operation(struct rdma_id_private *id_priv,
967 enum rdma_cm_state state)
970 case RDMA_CM_ADDR_QUERY:
971 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
973 case RDMA_CM_ROUTE_QUERY:
974 cma_cancel_route(id_priv);
977 if (cma_any_addr(cma_src_addr(id_priv)) && !id_priv->cma_dev)
978 cma_cancel_listens(id_priv);
985 static void cma_release_port(struct rdma_id_private *id_priv)
987 struct rdma_bind_list *bind_list = id_priv->bind_list;
993 hlist_del(&id_priv->node);
994 if (hlist_empty(&bind_list->owners)) {
995 idr_remove(bind_list->ps, bind_list->port);
1001 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
1003 struct cma_multicast *mc;
1005 while (!list_empty(&id_priv->mc_list)) {
1006 mc = container_of(id_priv->mc_list.next,
1007 struct cma_multicast, list);
1008 list_del(&mc->list);
1009 switch (rdma_port_get_link_layer(id_priv->cma_dev->device, id_priv->id.port_num)) {
1010 case IB_LINK_LAYER_INFINIBAND:
1011 ib_sa_free_multicast(mc->multicast.ib);
1014 case IB_LINK_LAYER_ETHERNET:
1015 kref_put(&mc->mcref, release_mc);
1023 void rdma_destroy_id(struct rdma_cm_id *id)
1025 struct rdma_id_private *id_priv;
1026 enum rdma_cm_state state;
1028 id_priv = container_of(id, struct rdma_id_private, id);
1029 state = cma_exch(id_priv, RDMA_CM_DESTROYING);
1030 cma_cancel_operation(id_priv, state);
1033 * Wait for any active callback to finish. New callbacks will find
1034 * the id_priv state set to destroying and abort.
1036 mutex_lock(&id_priv->handler_mutex);
1037 mutex_unlock(&id_priv->handler_mutex);
1039 if (id_priv->cma_dev) {
1040 switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
1041 case RDMA_TRANSPORT_IB:
1042 if (id_priv->cm_id.ib)
1043 ib_destroy_cm_id(id_priv->cm_id.ib);
1045 case RDMA_TRANSPORT_IWARP:
1046 if (id_priv->cm_id.iw)
1047 iw_destroy_cm_id(id_priv->cm_id.iw);
1052 cma_leave_mc_groups(id_priv);
1053 cma_release_dev(id_priv);
1056 cma_release_port(id_priv);
1057 cma_deref_id(id_priv);
1058 wait_for_completion(&id_priv->comp);
1060 if (id_priv->internal_id)
1061 cma_deref_id(id_priv->id.context);
1063 kfree(id_priv->id.route.path_rec);
1066 EXPORT_SYMBOL(rdma_destroy_id);
1068 static int cma_rep_recv(struct rdma_id_private *id_priv)
1072 ret = cma_modify_qp_rtr(id_priv, NULL);
1076 ret = cma_modify_qp_rts(id_priv, NULL);
1080 ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
1086 cma_modify_qp_err(id_priv);
1087 ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
1092 static void cma_set_rep_event_data(struct rdma_cm_event *event,
1093 struct ib_cm_rep_event_param *rep_data,
1096 event->param.conn.private_data = private_data;
1097 event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
1098 event->param.conn.responder_resources = rep_data->responder_resources;
1099 event->param.conn.initiator_depth = rep_data->initiator_depth;
1100 event->param.conn.flow_control = rep_data->flow_control;
1101 event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
1102 event->param.conn.srq = rep_data->srq;
1103 event->param.conn.qp_num = rep_data->remote_qpn;
1106 static int cma_ib_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1108 struct rdma_id_private *id_priv = cm_id->context;
1109 struct rdma_cm_event event;
1112 if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
1113 cma_disable_callback(id_priv, RDMA_CM_CONNECT)) ||
1114 (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
1115 cma_disable_callback(id_priv, RDMA_CM_DISCONNECT)))
1118 memset(&event, 0, sizeof event);
1119 switch (ib_event->event) {
1120 case IB_CM_REQ_ERROR:
1121 case IB_CM_REP_ERROR:
1122 event.event = RDMA_CM_EVENT_UNREACHABLE;
1123 event.status = -ETIMEDOUT;
1125 case IB_CM_REP_RECEIVED:
1126 if (id_priv->id.qp) {
1127 event.status = cma_rep_recv(id_priv);
1128 event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
1129 RDMA_CM_EVENT_ESTABLISHED;
1131 event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
1133 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
1134 ib_event->private_data);
1136 case IB_CM_RTU_RECEIVED:
1137 case IB_CM_USER_ESTABLISHED:
1138 event.event = RDMA_CM_EVENT_ESTABLISHED;
1140 case IB_CM_DREQ_ERROR:
1141 event.status = -ETIMEDOUT; /* fall through */
1142 case IB_CM_DREQ_RECEIVED:
1143 case IB_CM_DREP_RECEIVED:
1144 if (!cma_comp_exch(id_priv, RDMA_CM_CONNECT,
1145 RDMA_CM_DISCONNECT))
1147 event.event = RDMA_CM_EVENT_DISCONNECTED;
1149 case IB_CM_TIMEWAIT_EXIT:
1150 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
1152 case IB_CM_MRA_RECEIVED:
1155 case IB_CM_REJ_RECEIVED:
1156 cma_modify_qp_err(id_priv);
1157 event.status = ib_event->param.rej_rcvd.reason;
1158 event.event = RDMA_CM_EVENT_REJECTED;
1159 event.param.conn.private_data = ib_event->private_data;
1160 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
1163 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
1168 ret = id_priv->id.event_handler(&id_priv->id, &event);
1170 /* Destroy the CM ID by returning a non-zero value. */
1171 id_priv->cm_id.ib = NULL;
1172 cma_exch(id_priv, RDMA_CM_DESTROYING);
1173 mutex_unlock(&id_priv->handler_mutex);
1174 rdma_destroy_id(&id_priv->id);
1178 mutex_unlock(&id_priv->handler_mutex);
1182 static struct rdma_id_private *cma_new_conn_id(struct rdma_cm_id *listen_id,
1183 struct ib_cm_event *ib_event)
1185 struct rdma_id_private *id_priv;
1186 struct rdma_cm_id *id;
1187 struct rdma_route *rt;
1190 id = rdma_create_id(listen_id->event_handler, listen_id->context,
1191 listen_id->ps, ib_event->param.req_rcvd.qp_type);
1195 id_priv = container_of(id, struct rdma_id_private, id);
1196 if (cma_save_net_info(id, listen_id, ib_event))
1200 rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
1201 rt->path_rec = kmalloc(sizeof *rt->path_rec * rt->num_paths,
1206 rt->path_rec[0] = *ib_event->param.req_rcvd.primary_path;
1207 if (rt->num_paths == 2)
1208 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
1210 if (cma_any_addr(cma_src_addr(id_priv))) {
1211 rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
1212 rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
1213 ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
1215 ret = cma_translate_addr(cma_src_addr(id_priv), &rt->addr.dev_addr);
1219 rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
1221 id_priv->state = RDMA_CM_CONNECT;
1225 rdma_destroy_id(id);
1229 static struct rdma_id_private *cma_new_udp_id(struct rdma_cm_id *listen_id,
1230 struct ib_cm_event *ib_event)
1232 struct rdma_id_private *id_priv;
1233 struct rdma_cm_id *id;
1236 id = rdma_create_id(listen_id->event_handler, listen_id->context,
1237 listen_id->ps, IB_QPT_UD);
1241 id_priv = container_of(id, struct rdma_id_private, id);
1242 if (cma_save_net_info(id, listen_id, ib_event))
1245 if (!cma_any_addr((struct sockaddr *) &id->route.addr.src_addr)) {
1246 ret = cma_translate_addr(cma_src_addr(id_priv), &id->route.addr.dev_addr);
1251 id_priv->state = RDMA_CM_CONNECT;
1254 rdma_destroy_id(id);
1258 static void cma_set_req_event_data(struct rdma_cm_event *event,
1259 struct ib_cm_req_event_param *req_data,
1260 void *private_data, int offset)
1262 event->param.conn.private_data = private_data + offset;
1263 event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
1264 event->param.conn.responder_resources = req_data->responder_resources;
1265 event->param.conn.initiator_depth = req_data->initiator_depth;
1266 event->param.conn.flow_control = req_data->flow_control;
1267 event->param.conn.retry_count = req_data->retry_count;
1268 event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
1269 event->param.conn.srq = req_data->srq;
1270 event->param.conn.qp_num = req_data->remote_qpn;
1273 static int cma_check_req_qp_type(struct rdma_cm_id *id, struct ib_cm_event *ib_event)
1275 return (((ib_event->event == IB_CM_REQ_RECEIVED) &&
1276 (ib_event->param.req_rcvd.qp_type == id->qp_type)) ||
1277 ((ib_event->event == IB_CM_SIDR_REQ_RECEIVED) &&
1278 (id->qp_type == IB_QPT_UD)) ||
1282 static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1284 struct rdma_id_private *listen_id, *conn_id;
1285 struct rdma_cm_event event;
1288 listen_id = cm_id->context;
1289 if (!cma_check_req_qp_type(&listen_id->id, ib_event))
1292 if (cma_disable_callback(listen_id, RDMA_CM_LISTEN))
1293 return -ECONNABORTED;
1295 memset(&event, 0, sizeof event);
1296 offset = cma_user_data_offset(listen_id);
1297 event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1298 if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) {
1299 conn_id = cma_new_udp_id(&listen_id->id, ib_event);
1300 event.param.ud.private_data = ib_event->private_data + offset;
1301 event.param.ud.private_data_len =
1302 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
1304 conn_id = cma_new_conn_id(&listen_id->id, ib_event);
1305 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
1306 ib_event->private_data, offset);
1313 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1314 ret = cma_acquire_dev(conn_id, listen_id);
1318 conn_id->cm_id.ib = cm_id;
1319 cm_id->context = conn_id;
1320 cm_id->cm_handler = cma_ib_handler;
1323 * Protect against the user destroying conn_id from another thread
1324 * until we're done accessing it.
1326 atomic_inc(&conn_id->refcount);
1327 ret = conn_id->id.event_handler(&conn_id->id, &event);
1331 * Acquire mutex to prevent user executing rdma_destroy_id()
1332 * while we're accessing the cm_id.
1335 if (cma_comp(conn_id, RDMA_CM_CONNECT) &&
1336 (conn_id->id.qp_type != IB_QPT_UD))
1337 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
1338 mutex_unlock(&lock);
1339 mutex_unlock(&conn_id->handler_mutex);
1340 mutex_unlock(&listen_id->handler_mutex);
1341 cma_deref_id(conn_id);
1345 cma_deref_id(conn_id);
1346 /* Destroy the CM ID by returning a non-zero value. */
1347 conn_id->cm_id.ib = NULL;
1349 cma_exch(conn_id, RDMA_CM_DESTROYING);
1350 mutex_unlock(&conn_id->handler_mutex);
1352 mutex_unlock(&listen_id->handler_mutex);
1354 rdma_destroy_id(&conn_id->id);
1358 __be64 rdma_get_service_id(struct rdma_cm_id *id, struct sockaddr *addr)
1360 if (addr->sa_family == AF_IB)
1361 return ((struct sockaddr_ib *) addr)->sib_sid;
1363 return cpu_to_be64(((u64)id->ps << 16) + be16_to_cpu(cma_port(addr)));
1365 EXPORT_SYMBOL(rdma_get_service_id);
1367 static void cma_set_compare_data(enum rdma_port_space ps, struct sockaddr *addr,
1368 struct ib_cm_compare_data *compare)
1370 struct cma_hdr *cma_data, *cma_mask;
1372 struct in6_addr ip6_addr;
1374 memset(compare, 0, sizeof *compare);
1375 cma_data = (void *) compare->data;
1376 cma_mask = (void *) compare->mask;
1378 switch (addr->sa_family) {
1380 ip4_addr = ((struct sockaddr_in *) addr)->sin_addr.s_addr;
1381 cma_set_ip_ver(cma_data, 4);
1382 cma_set_ip_ver(cma_mask, 0xF);
1383 if (!cma_any_addr(addr)) {
1384 cma_data->dst_addr.ip4.addr = ip4_addr;
1385 cma_mask->dst_addr.ip4.addr = htonl(~0);
1389 ip6_addr = ((struct sockaddr_in6 *) addr)->sin6_addr;
1390 cma_set_ip_ver(cma_data, 6);
1391 cma_set_ip_ver(cma_mask, 0xF);
1392 if (!cma_any_addr(addr)) {
1393 cma_data->dst_addr.ip6 = ip6_addr;
1394 memset(&cma_mask->dst_addr.ip6, 0xFF,
1395 sizeof cma_mask->dst_addr.ip6);
1403 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
1405 struct rdma_id_private *id_priv = iw_id->context;
1406 struct rdma_cm_event event;
1408 struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
1409 struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
1411 if (cma_disable_callback(id_priv, RDMA_CM_CONNECT))
1414 memset(&event, 0, sizeof event);
1415 switch (iw_event->event) {
1416 case IW_CM_EVENT_CLOSE:
1417 event.event = RDMA_CM_EVENT_DISCONNECTED;
1419 case IW_CM_EVENT_CONNECT_REPLY:
1420 memcpy(cma_src_addr(id_priv), laddr,
1421 rdma_addr_size(laddr));
1422 memcpy(cma_dst_addr(id_priv), raddr,
1423 rdma_addr_size(raddr));
1424 switch (iw_event->status) {
1426 event.event = RDMA_CM_EVENT_ESTABLISHED;
1427 event.param.conn.initiator_depth = iw_event->ird;
1428 event.param.conn.responder_resources = iw_event->ord;
1432 event.event = RDMA_CM_EVENT_REJECTED;
1435 event.event = RDMA_CM_EVENT_UNREACHABLE;
1438 event.event = RDMA_CM_EVENT_CONNECT_ERROR;
1442 case IW_CM_EVENT_ESTABLISHED:
1443 event.event = RDMA_CM_EVENT_ESTABLISHED;
1444 event.param.conn.initiator_depth = iw_event->ird;
1445 event.param.conn.responder_resources = iw_event->ord;
1451 event.status = iw_event->status;
1452 event.param.conn.private_data = iw_event->private_data;
1453 event.param.conn.private_data_len = iw_event->private_data_len;
1454 ret = id_priv->id.event_handler(&id_priv->id, &event);
1456 /* Destroy the CM ID by returning a non-zero value. */
1457 id_priv->cm_id.iw = NULL;
1458 cma_exch(id_priv, RDMA_CM_DESTROYING);
1459 mutex_unlock(&id_priv->handler_mutex);
1460 rdma_destroy_id(&id_priv->id);
1464 mutex_unlock(&id_priv->handler_mutex);
1468 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
1469 struct iw_cm_event *iw_event)
1471 struct rdma_cm_id *new_cm_id;
1472 struct rdma_id_private *listen_id, *conn_id;
1473 struct rdma_cm_event event;
1475 struct ib_device_attr attr;
1476 struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
1477 struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
1479 listen_id = cm_id->context;
1480 if (cma_disable_callback(listen_id, RDMA_CM_LISTEN))
1481 return -ECONNABORTED;
1483 /* Create a new RDMA id for the new IW CM ID */
1484 new_cm_id = rdma_create_id(listen_id->id.event_handler,
1485 listen_id->id.context,
1486 RDMA_PS_TCP, IB_QPT_RC);
1487 if (IS_ERR(new_cm_id)) {
1491 conn_id = container_of(new_cm_id, struct rdma_id_private, id);
1492 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1493 conn_id->state = RDMA_CM_CONNECT;
1495 ret = rdma_translate_ip(laddr, &conn_id->id.route.addr.dev_addr, NULL);
1497 mutex_unlock(&conn_id->handler_mutex);
1498 rdma_destroy_id(new_cm_id);
1502 ret = cma_acquire_dev(conn_id, listen_id);
1504 mutex_unlock(&conn_id->handler_mutex);
1505 rdma_destroy_id(new_cm_id);
1509 conn_id->cm_id.iw = cm_id;
1510 cm_id->context = conn_id;
1511 cm_id->cm_handler = cma_iw_handler;
1513 memcpy(cma_src_addr(conn_id), laddr, rdma_addr_size(laddr));
1514 memcpy(cma_dst_addr(conn_id), raddr, rdma_addr_size(raddr));
1516 ret = ib_query_device(conn_id->id.device, &attr);
1518 mutex_unlock(&conn_id->handler_mutex);
1519 rdma_destroy_id(new_cm_id);
1523 memset(&event, 0, sizeof event);
1524 event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1525 event.param.conn.private_data = iw_event->private_data;
1526 event.param.conn.private_data_len = iw_event->private_data_len;
1527 event.param.conn.initiator_depth = iw_event->ird;
1528 event.param.conn.responder_resources = iw_event->ord;
1531 * Protect against the user destroying conn_id from another thread
1532 * until we're done accessing it.
1534 atomic_inc(&conn_id->refcount);
1535 ret = conn_id->id.event_handler(&conn_id->id, &event);
1537 /* User wants to destroy the CM ID */
1538 conn_id->cm_id.iw = NULL;
1539 cma_exch(conn_id, RDMA_CM_DESTROYING);
1540 mutex_unlock(&conn_id->handler_mutex);
1541 cma_deref_id(conn_id);
1542 rdma_destroy_id(&conn_id->id);
1546 mutex_unlock(&conn_id->handler_mutex);
1547 cma_deref_id(conn_id);
1550 mutex_unlock(&listen_id->handler_mutex);
1554 static int cma_ib_listen(struct rdma_id_private *id_priv)
1556 struct ib_cm_compare_data compare_data;
1557 struct sockaddr *addr;
1558 struct ib_cm_id *id;
1562 id = ib_create_cm_id(id_priv->id.device, cma_req_handler, id_priv);
1566 id_priv->cm_id.ib = id;
1568 addr = cma_src_addr(id_priv);
1569 svc_id = rdma_get_service_id(&id_priv->id, addr);
1570 if (cma_any_addr(addr) && !id_priv->afonly)
1571 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, NULL);
1573 cma_set_compare_data(id_priv->id.ps, addr, &compare_data);
1574 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, &compare_data);
1578 ib_destroy_cm_id(id_priv->cm_id.ib);
1579 id_priv->cm_id.ib = NULL;
1585 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
1588 struct iw_cm_id *id;
1590 id = iw_create_cm_id(id_priv->id.device,
1591 iw_conn_req_handler,
1596 id_priv->cm_id.iw = id;
1598 memcpy(&id_priv->cm_id.iw->local_addr, cma_src_addr(id_priv),
1599 rdma_addr_size(cma_src_addr(id_priv)));
1601 ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
1604 iw_destroy_cm_id(id_priv->cm_id.iw);
1605 id_priv->cm_id.iw = NULL;
1611 static int cma_listen_handler(struct rdma_cm_id *id,
1612 struct rdma_cm_event *event)
1614 struct rdma_id_private *id_priv = id->context;
1616 id->context = id_priv->id.context;
1617 id->event_handler = id_priv->id.event_handler;
1618 return id_priv->id.event_handler(id, event);
1621 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
1622 struct cma_device *cma_dev)
1624 struct rdma_id_private *dev_id_priv;
1625 struct rdma_cm_id *id;
1628 if (cma_family(id_priv) == AF_IB &&
1629 rdma_node_get_transport(cma_dev->device->node_type) != RDMA_TRANSPORT_IB)
1632 id = rdma_create_id(cma_listen_handler, id_priv, id_priv->id.ps,
1633 id_priv->id.qp_type);
1637 dev_id_priv = container_of(id, struct rdma_id_private, id);
1639 dev_id_priv->state = RDMA_CM_ADDR_BOUND;
1640 memcpy(cma_src_addr(dev_id_priv), cma_src_addr(id_priv),
1641 rdma_addr_size(cma_src_addr(id_priv)));
1643 cma_attach_to_dev(dev_id_priv, cma_dev);
1644 list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
1645 atomic_inc(&id_priv->refcount);
1646 dev_id_priv->internal_id = 1;
1647 dev_id_priv->afonly = id_priv->afonly;
1649 ret = rdma_listen(id, id_priv->backlog);
1651 printk(KERN_WARNING "RDMA CMA: cma_listen_on_dev, error %d, "
1652 "listening on device %s\n", ret, cma_dev->device->name);
1655 static void cma_listen_on_all(struct rdma_id_private *id_priv)
1657 struct cma_device *cma_dev;
1660 list_add_tail(&id_priv->list, &listen_any_list);
1661 list_for_each_entry(cma_dev, &dev_list, list)
1662 cma_listen_on_dev(id_priv, cma_dev);
1663 mutex_unlock(&lock);
1666 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
1668 struct rdma_id_private *id_priv;
1670 id_priv = container_of(id, struct rdma_id_private, id);
1671 id_priv->tos = (u8) tos;
1673 EXPORT_SYMBOL(rdma_set_service_type);
1675 static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec,
1678 struct cma_work *work = context;
1679 struct rdma_route *route;
1681 route = &work->id->id.route;
1684 route->num_paths = 1;
1685 *route->path_rec = *path_rec;
1687 work->old_state = RDMA_CM_ROUTE_QUERY;
1688 work->new_state = RDMA_CM_ADDR_RESOLVED;
1689 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
1690 work->event.status = status;
1693 queue_work(cma_wq, &work->work);
1696 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms,
1697 struct cma_work *work)
1699 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
1700 struct ib_sa_path_rec path_rec;
1701 ib_sa_comp_mask comp_mask;
1702 struct sockaddr_in6 *sin6;
1703 struct sockaddr_ib *sib;
1705 memset(&path_rec, 0, sizeof path_rec);
1706 rdma_addr_get_sgid(dev_addr, &path_rec.sgid);
1707 rdma_addr_get_dgid(dev_addr, &path_rec.dgid);
1708 path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
1709 path_rec.numb_path = 1;
1710 path_rec.reversible = 1;
1711 path_rec.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
1713 comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
1714 IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
1715 IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
1717 switch (cma_family(id_priv)) {
1719 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
1720 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
1723 sin6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
1724 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
1725 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
1728 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
1729 path_rec.traffic_class = (u8) (be32_to_cpu(sib->sib_flowinfo) >> 20);
1730 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
1734 id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
1735 id_priv->id.port_num, &path_rec,
1736 comp_mask, timeout_ms,
1737 GFP_KERNEL, cma_query_handler,
1738 work, &id_priv->query);
1740 return (id_priv->query_id < 0) ? id_priv->query_id : 0;
1743 static void cma_work_handler(struct work_struct *_work)
1745 struct cma_work *work = container_of(_work, struct cma_work, work);
1746 struct rdma_id_private *id_priv = work->id;
1749 mutex_lock(&id_priv->handler_mutex);
1750 if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
1753 if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1754 cma_exch(id_priv, RDMA_CM_DESTROYING);
1758 mutex_unlock(&id_priv->handler_mutex);
1759 cma_deref_id(id_priv);
1761 rdma_destroy_id(&id_priv->id);
1765 static void cma_ndev_work_handler(struct work_struct *_work)
1767 struct cma_ndev_work *work = container_of(_work, struct cma_ndev_work, work);
1768 struct rdma_id_private *id_priv = work->id;
1771 mutex_lock(&id_priv->handler_mutex);
1772 if (id_priv->state == RDMA_CM_DESTROYING ||
1773 id_priv->state == RDMA_CM_DEVICE_REMOVAL)
1776 if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1777 cma_exch(id_priv, RDMA_CM_DESTROYING);
1782 mutex_unlock(&id_priv->handler_mutex);
1783 cma_deref_id(id_priv);
1785 rdma_destroy_id(&id_priv->id);
1789 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms)
1791 struct rdma_route *route = &id_priv->id.route;
1792 struct cma_work *work;
1795 work = kzalloc(sizeof *work, GFP_KERNEL);
1800 INIT_WORK(&work->work, cma_work_handler);
1801 work->old_state = RDMA_CM_ROUTE_QUERY;
1802 work->new_state = RDMA_CM_ROUTE_RESOLVED;
1803 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1805 route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
1806 if (!route->path_rec) {
1811 ret = cma_query_ib_route(id_priv, timeout_ms, work);
1817 kfree(route->path_rec);
1818 route->path_rec = NULL;
1824 int rdma_set_ib_paths(struct rdma_cm_id *id,
1825 struct ib_sa_path_rec *path_rec, int num_paths)
1827 struct rdma_id_private *id_priv;
1830 id_priv = container_of(id, struct rdma_id_private, id);
1831 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
1832 RDMA_CM_ROUTE_RESOLVED))
1835 id->route.path_rec = kmemdup(path_rec, sizeof *path_rec * num_paths,
1837 if (!id->route.path_rec) {
1842 id->route.num_paths = num_paths;
1845 cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED);
1848 EXPORT_SYMBOL(rdma_set_ib_paths);
1850 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms)
1852 struct cma_work *work;
1854 work = kzalloc(sizeof *work, GFP_KERNEL);
1859 INIT_WORK(&work->work, cma_work_handler);
1860 work->old_state = RDMA_CM_ROUTE_QUERY;
1861 work->new_state = RDMA_CM_ROUTE_RESOLVED;
1862 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1863 queue_work(cma_wq, &work->work);
1867 static int iboe_tos_to_sl(struct net_device *ndev, int tos)
1870 struct net_device *dev;
1872 prio = rt_tos2priority(tos);
1873 dev = ndev->priv_flags & IFF_802_1Q_VLAN ?
1874 vlan_dev_real_dev(ndev) : ndev;
1877 return netdev_get_prio_tc_map(dev, prio);
1879 #if IS_ENABLED(CONFIG_VLAN_8021Q)
1880 if (ndev->priv_flags & IFF_802_1Q_VLAN)
1881 return (vlan_dev_get_egress_qos_mask(ndev, prio) &
1882 VLAN_PRIO_MASK) >> VLAN_PRIO_SHIFT;
1887 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
1889 struct rdma_route *route = &id_priv->id.route;
1890 struct rdma_addr *addr = &route->addr;
1891 struct cma_work *work;
1893 struct net_device *ndev = NULL;
1896 work = kzalloc(sizeof *work, GFP_KERNEL);
1901 INIT_WORK(&work->work, cma_work_handler);
1903 route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
1904 if (!route->path_rec) {
1909 route->num_paths = 1;
1911 if (addr->dev_addr.bound_dev_if)
1912 ndev = dev_get_by_index(&init_net, addr->dev_addr.bound_dev_if);
1918 route->path_rec->vlan_id = rdma_vlan_dev_vlan_id(ndev);
1919 memcpy(route->path_rec->dmac, addr->dev_addr.dst_dev_addr, ETH_ALEN);
1920 memcpy(route->path_rec->smac, ndev->dev_addr, ndev->addr_len);
1922 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
1923 &route->path_rec->sgid);
1924 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.dst_addr,
1925 &route->path_rec->dgid);
1927 route->path_rec->hop_limit = 1;
1928 route->path_rec->reversible = 1;
1929 route->path_rec->pkey = cpu_to_be16(0xffff);
1930 route->path_rec->mtu_selector = IB_SA_EQ;
1931 route->path_rec->sl = iboe_tos_to_sl(ndev, id_priv->tos);
1932 route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
1933 route->path_rec->rate_selector = IB_SA_EQ;
1934 route->path_rec->rate = iboe_get_rate(ndev);
1936 route->path_rec->packet_life_time_selector = IB_SA_EQ;
1937 route->path_rec->packet_life_time = CMA_IBOE_PACKET_LIFETIME;
1938 if (!route->path_rec->mtu) {
1943 work->old_state = RDMA_CM_ROUTE_QUERY;
1944 work->new_state = RDMA_CM_ROUTE_RESOLVED;
1945 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1946 work->event.status = 0;
1948 queue_work(cma_wq, &work->work);
1953 kfree(route->path_rec);
1954 route->path_rec = NULL;
1960 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
1962 struct rdma_id_private *id_priv;
1965 id_priv = container_of(id, struct rdma_id_private, id);
1966 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY))
1969 atomic_inc(&id_priv->refcount);
1970 switch (rdma_node_get_transport(id->device->node_type)) {
1971 case RDMA_TRANSPORT_IB:
1972 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
1973 case IB_LINK_LAYER_INFINIBAND:
1974 ret = cma_resolve_ib_route(id_priv, timeout_ms);
1976 case IB_LINK_LAYER_ETHERNET:
1977 ret = cma_resolve_iboe_route(id_priv);
1983 case RDMA_TRANSPORT_IWARP:
1984 ret = cma_resolve_iw_route(id_priv, timeout_ms);
1995 cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED);
1996 cma_deref_id(id_priv);
1999 EXPORT_SYMBOL(rdma_resolve_route);
2001 static void cma_set_loopback(struct sockaddr *addr)
2003 switch (addr->sa_family) {
2005 ((struct sockaddr_in *) addr)->sin_addr.s_addr = htonl(INADDR_LOOPBACK);
2008 ipv6_addr_set(&((struct sockaddr_in6 *) addr)->sin6_addr,
2012 ib_addr_set(&((struct sockaddr_ib *) addr)->sib_addr,
2018 static int cma_bind_loopback(struct rdma_id_private *id_priv)
2020 struct cma_device *cma_dev, *cur_dev;
2021 struct ib_port_attr port_attr;
2029 list_for_each_entry(cur_dev, &dev_list, list) {
2030 if (cma_family(id_priv) == AF_IB &&
2031 rdma_node_get_transport(cur_dev->device->node_type) != RDMA_TRANSPORT_IB)
2037 for (p = 1; p <= cur_dev->device->phys_port_cnt; ++p) {
2038 if (!ib_query_port(cur_dev->device, p, &port_attr) &&
2039 port_attr.state == IB_PORT_ACTIVE) {
2054 ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid);
2058 ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
2062 id_priv->id.route.addr.dev_addr.dev_type =
2063 (rdma_port_get_link_layer(cma_dev->device, p) == IB_LINK_LAYER_INFINIBAND) ?
2064 ARPHRD_INFINIBAND : ARPHRD_ETHER;
2066 rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2067 ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
2068 id_priv->id.port_num = p;
2069 cma_attach_to_dev(id_priv, cma_dev);
2070 cma_set_loopback(cma_src_addr(id_priv));
2072 mutex_unlock(&lock);
2076 static void addr_handler(int status, struct sockaddr *src_addr,
2077 struct rdma_dev_addr *dev_addr, void *context)
2079 struct rdma_id_private *id_priv = context;
2080 struct rdma_cm_event event;
2082 memset(&event, 0, sizeof event);
2083 mutex_lock(&id_priv->handler_mutex);
2084 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY,
2085 RDMA_CM_ADDR_RESOLVED))
2088 memcpy(cma_src_addr(id_priv), src_addr, rdma_addr_size(src_addr));
2089 if (!status && !id_priv->cma_dev)
2090 status = cma_acquire_dev(id_priv, NULL);
2093 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
2094 RDMA_CM_ADDR_BOUND))
2096 event.event = RDMA_CM_EVENT_ADDR_ERROR;
2097 event.status = status;
2099 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2101 if (id_priv->id.event_handler(&id_priv->id, &event)) {
2102 cma_exch(id_priv, RDMA_CM_DESTROYING);
2103 mutex_unlock(&id_priv->handler_mutex);
2104 cma_deref_id(id_priv);
2105 rdma_destroy_id(&id_priv->id);
2109 mutex_unlock(&id_priv->handler_mutex);
2110 cma_deref_id(id_priv);
2113 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
2115 struct cma_work *work;
2119 work = kzalloc(sizeof *work, GFP_KERNEL);
2123 if (!id_priv->cma_dev) {
2124 ret = cma_bind_loopback(id_priv);
2129 rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2130 rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
2133 INIT_WORK(&work->work, cma_work_handler);
2134 work->old_state = RDMA_CM_ADDR_QUERY;
2135 work->new_state = RDMA_CM_ADDR_RESOLVED;
2136 work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2137 queue_work(cma_wq, &work->work);
2144 static int cma_resolve_ib_addr(struct rdma_id_private *id_priv)
2146 struct cma_work *work;
2149 work = kzalloc(sizeof *work, GFP_KERNEL);
2153 if (!id_priv->cma_dev) {
2154 ret = cma_resolve_ib_dev(id_priv);
2159 rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, (union ib_gid *)
2160 &(((struct sockaddr_ib *) &id_priv->id.route.addr.dst_addr)->sib_addr));
2163 INIT_WORK(&work->work, cma_work_handler);
2164 work->old_state = RDMA_CM_ADDR_QUERY;
2165 work->new_state = RDMA_CM_ADDR_RESOLVED;
2166 work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2167 queue_work(cma_wq, &work->work);
2174 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2175 struct sockaddr *dst_addr)
2177 if (!src_addr || !src_addr->sa_family) {
2178 src_addr = (struct sockaddr *) &id->route.addr.src_addr;
2179 src_addr->sa_family = dst_addr->sa_family;
2180 if (dst_addr->sa_family == AF_INET6) {
2181 ((struct sockaddr_in6 *) src_addr)->sin6_scope_id =
2182 ((struct sockaddr_in6 *) dst_addr)->sin6_scope_id;
2183 } else if (dst_addr->sa_family == AF_IB) {
2184 ((struct sockaddr_ib *) src_addr)->sib_pkey =
2185 ((struct sockaddr_ib *) dst_addr)->sib_pkey;
2188 return rdma_bind_addr(id, src_addr);
2191 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2192 struct sockaddr *dst_addr, int timeout_ms)
2194 struct rdma_id_private *id_priv;
2197 id_priv = container_of(id, struct rdma_id_private, id);
2198 if (id_priv->state == RDMA_CM_IDLE) {
2199 ret = cma_bind_addr(id, src_addr, dst_addr);
2204 if (cma_family(id_priv) != dst_addr->sa_family)
2207 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY))
2210 atomic_inc(&id_priv->refcount);
2211 memcpy(cma_dst_addr(id_priv), dst_addr, rdma_addr_size(dst_addr));
2212 if (cma_any_addr(dst_addr)) {
2213 ret = cma_resolve_loopback(id_priv);
2215 if (dst_addr->sa_family == AF_IB) {
2216 ret = cma_resolve_ib_addr(id_priv);
2218 ret = rdma_resolve_ip(&addr_client, cma_src_addr(id_priv),
2219 dst_addr, &id->route.addr.dev_addr,
2220 timeout_ms, addr_handler, id_priv);
2228 cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
2229 cma_deref_id(id_priv);
2232 EXPORT_SYMBOL(rdma_resolve_addr);
2234 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse)
2236 struct rdma_id_private *id_priv;
2237 unsigned long flags;
2240 id_priv = container_of(id, struct rdma_id_private, id);
2241 spin_lock_irqsave(&id_priv->lock, flags);
2242 if (reuse || id_priv->state == RDMA_CM_IDLE) {
2243 id_priv->reuseaddr = reuse;
2248 spin_unlock_irqrestore(&id_priv->lock, flags);
2251 EXPORT_SYMBOL(rdma_set_reuseaddr);
2253 int rdma_set_afonly(struct rdma_cm_id *id, int afonly)
2255 struct rdma_id_private *id_priv;
2256 unsigned long flags;
2259 id_priv = container_of(id, struct rdma_id_private, id);
2260 spin_lock_irqsave(&id_priv->lock, flags);
2261 if (id_priv->state == RDMA_CM_IDLE || id_priv->state == RDMA_CM_ADDR_BOUND) {
2262 id_priv->options |= (1 << CMA_OPTION_AFONLY);
2263 id_priv->afonly = afonly;
2268 spin_unlock_irqrestore(&id_priv->lock, flags);
2271 EXPORT_SYMBOL(rdma_set_afonly);
2273 static void cma_bind_port(struct rdma_bind_list *bind_list,
2274 struct rdma_id_private *id_priv)
2276 struct sockaddr *addr;
2277 struct sockaddr_ib *sib;
2281 addr = cma_src_addr(id_priv);
2282 port = htons(bind_list->port);
2284 switch (addr->sa_family) {
2286 ((struct sockaddr_in *) addr)->sin_port = port;
2289 ((struct sockaddr_in6 *) addr)->sin6_port = port;
2292 sib = (struct sockaddr_ib *) addr;
2293 sid = be64_to_cpu(sib->sib_sid);
2294 mask = be64_to_cpu(sib->sib_sid_mask);
2295 sib->sib_sid = cpu_to_be64((sid & mask) | (u64) ntohs(port));
2296 sib->sib_sid_mask = cpu_to_be64(~0ULL);
2299 id_priv->bind_list = bind_list;
2300 hlist_add_head(&id_priv->node, &bind_list->owners);
2303 static int cma_alloc_port(struct idr *ps, struct rdma_id_private *id_priv,
2304 unsigned short snum)
2306 struct rdma_bind_list *bind_list;
2309 bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
2313 ret = idr_alloc(ps, bind_list, snum, snum + 1, GFP_KERNEL);
2318 bind_list->port = (unsigned short)ret;
2319 cma_bind_port(bind_list, id_priv);
2323 return ret == -ENOSPC ? -EADDRNOTAVAIL : ret;
2326 static int cma_alloc_any_port(struct idr *ps, struct rdma_id_private *id_priv)
2328 static unsigned int last_used_port;
2329 int low, high, remaining;
2332 inet_get_local_port_range(&init_net, &low, &high);
2333 remaining = (high - low) + 1;
2334 rover = prandom_u32() % remaining + low;
2336 if (last_used_port != rover &&
2337 !idr_find(ps, (unsigned short) rover)) {
2338 int ret = cma_alloc_port(ps, id_priv, rover);
2340 * Remember previously used port number in order to avoid
2341 * re-using same port immediately after it is closed.
2344 last_used_port = rover;
2345 if (ret != -EADDRNOTAVAIL)
2350 if ((rover < low) || (rover > high))
2354 return -EADDRNOTAVAIL;
2358 * Check that the requested port is available. This is called when trying to
2359 * bind to a specific port, or when trying to listen on a bound port. In
2360 * the latter case, the provided id_priv may already be on the bind_list, but
2361 * we still need to check that it's okay to start listening.
2363 static int cma_check_port(struct rdma_bind_list *bind_list,
2364 struct rdma_id_private *id_priv, uint8_t reuseaddr)
2366 struct rdma_id_private *cur_id;
2367 struct sockaddr *addr, *cur_addr;
2369 addr = cma_src_addr(id_priv);
2370 hlist_for_each_entry(cur_id, &bind_list->owners, node) {
2371 if (id_priv == cur_id)
2374 if ((cur_id->state != RDMA_CM_LISTEN) && reuseaddr &&
2378 cur_addr = cma_src_addr(cur_id);
2379 if (id_priv->afonly && cur_id->afonly &&
2380 (addr->sa_family != cur_addr->sa_family))
2383 if (cma_any_addr(addr) || cma_any_addr(cur_addr))
2384 return -EADDRNOTAVAIL;
2386 if (!cma_addr_cmp(addr, cur_addr))
2392 static int cma_use_port(struct idr *ps, struct rdma_id_private *id_priv)
2394 struct rdma_bind_list *bind_list;
2395 unsigned short snum;
2398 snum = ntohs(cma_port(cma_src_addr(id_priv)));
2399 if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
2402 bind_list = idr_find(ps, snum);
2404 ret = cma_alloc_port(ps, id_priv, snum);
2406 ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr);
2408 cma_bind_port(bind_list, id_priv);
2413 static int cma_bind_listen(struct rdma_id_private *id_priv)
2415 struct rdma_bind_list *bind_list = id_priv->bind_list;
2419 if (bind_list->owners.first->next)
2420 ret = cma_check_port(bind_list, id_priv, 0);
2421 mutex_unlock(&lock);
2425 static struct idr *cma_select_inet_ps(struct rdma_id_private *id_priv)
2427 switch (id_priv->id.ps) {
2441 static struct idr *cma_select_ib_ps(struct rdma_id_private *id_priv)
2443 struct idr *ps = NULL;
2444 struct sockaddr_ib *sib;
2445 u64 sid_ps, mask, sid;
2447 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
2448 mask = be64_to_cpu(sib->sib_sid_mask) & RDMA_IB_IP_PS_MASK;
2449 sid = be64_to_cpu(sib->sib_sid) & mask;
2451 if ((id_priv->id.ps == RDMA_PS_IB) && (sid == (RDMA_IB_IP_PS_IB & mask))) {
2452 sid_ps = RDMA_IB_IP_PS_IB;
2454 } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_TCP)) &&
2455 (sid == (RDMA_IB_IP_PS_TCP & mask))) {
2456 sid_ps = RDMA_IB_IP_PS_TCP;
2458 } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_UDP)) &&
2459 (sid == (RDMA_IB_IP_PS_UDP & mask))) {
2460 sid_ps = RDMA_IB_IP_PS_UDP;
2465 sib->sib_sid = cpu_to_be64(sid_ps | ntohs(cma_port((struct sockaddr *) sib)));
2466 sib->sib_sid_mask = cpu_to_be64(RDMA_IB_IP_PS_MASK |
2467 be64_to_cpu(sib->sib_sid_mask));
2472 static int cma_get_port(struct rdma_id_private *id_priv)
2477 if (cma_family(id_priv) != AF_IB)
2478 ps = cma_select_inet_ps(id_priv);
2480 ps = cma_select_ib_ps(id_priv);
2482 return -EPROTONOSUPPORT;
2485 if (cma_any_port(cma_src_addr(id_priv)))
2486 ret = cma_alloc_any_port(ps, id_priv);
2488 ret = cma_use_port(ps, id_priv);
2489 mutex_unlock(&lock);
2494 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
2495 struct sockaddr *addr)
2497 #if IS_ENABLED(CONFIG_IPV6)
2498 struct sockaddr_in6 *sin6;
2500 if (addr->sa_family != AF_INET6)
2503 sin6 = (struct sockaddr_in6 *) addr;
2505 if (!(ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL))
2508 if (!sin6->sin6_scope_id)
2511 dev_addr->bound_dev_if = sin6->sin6_scope_id;
2516 int rdma_listen(struct rdma_cm_id *id, int backlog)
2518 struct rdma_id_private *id_priv;
2521 id_priv = container_of(id, struct rdma_id_private, id);
2522 if (id_priv->state == RDMA_CM_IDLE) {
2523 id->route.addr.src_addr.ss_family = AF_INET;
2524 ret = rdma_bind_addr(id, cma_src_addr(id_priv));
2529 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN))
2532 if (id_priv->reuseaddr) {
2533 ret = cma_bind_listen(id_priv);
2538 id_priv->backlog = backlog;
2540 switch (rdma_node_get_transport(id->device->node_type)) {
2541 case RDMA_TRANSPORT_IB:
2542 ret = cma_ib_listen(id_priv);
2546 case RDMA_TRANSPORT_IWARP:
2547 ret = cma_iw_listen(id_priv, backlog);
2556 cma_listen_on_all(id_priv);
2560 id_priv->backlog = 0;
2561 cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND);
2564 EXPORT_SYMBOL(rdma_listen);
2566 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
2568 struct rdma_id_private *id_priv;
2571 if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6 &&
2572 addr->sa_family != AF_IB)
2573 return -EAFNOSUPPORT;
2575 id_priv = container_of(id, struct rdma_id_private, id);
2576 if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND))
2579 ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
2583 memcpy(cma_src_addr(id_priv), addr, rdma_addr_size(addr));
2584 if (!cma_any_addr(addr)) {
2585 ret = cma_translate_addr(addr, &id->route.addr.dev_addr);
2589 ret = cma_acquire_dev(id_priv, NULL);
2594 if (!(id_priv->options & (1 << CMA_OPTION_AFONLY))) {
2595 if (addr->sa_family == AF_INET)
2596 id_priv->afonly = 1;
2597 #if IS_ENABLED(CONFIG_IPV6)
2598 else if (addr->sa_family == AF_INET6)
2599 id_priv->afonly = init_net.ipv6.sysctl.bindv6only;
2602 ret = cma_get_port(id_priv);
2608 if (id_priv->cma_dev)
2609 cma_release_dev(id_priv);
2611 cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE);
2614 EXPORT_SYMBOL(rdma_bind_addr);
2616 static int cma_format_hdr(void *hdr, struct rdma_id_private *id_priv)
2618 struct cma_hdr *cma_hdr;
2621 cma_hdr->cma_version = CMA_VERSION;
2622 if (cma_family(id_priv) == AF_INET) {
2623 struct sockaddr_in *src4, *dst4;
2625 src4 = (struct sockaddr_in *) cma_src_addr(id_priv);
2626 dst4 = (struct sockaddr_in *) cma_dst_addr(id_priv);
2628 cma_set_ip_ver(cma_hdr, 4);
2629 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2630 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2631 cma_hdr->port = src4->sin_port;
2632 } else if (cma_family(id_priv) == AF_INET6) {
2633 struct sockaddr_in6 *src6, *dst6;
2635 src6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
2636 dst6 = (struct sockaddr_in6 *) cma_dst_addr(id_priv);
2638 cma_set_ip_ver(cma_hdr, 6);
2639 cma_hdr->src_addr.ip6 = src6->sin6_addr;
2640 cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
2641 cma_hdr->port = src6->sin6_port;
2646 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
2647 struct ib_cm_event *ib_event)
2649 struct rdma_id_private *id_priv = cm_id->context;
2650 struct rdma_cm_event event;
2651 struct ib_cm_sidr_rep_event_param *rep = &ib_event->param.sidr_rep_rcvd;
2654 if (cma_disable_callback(id_priv, RDMA_CM_CONNECT))
2657 memset(&event, 0, sizeof event);
2658 switch (ib_event->event) {
2659 case IB_CM_SIDR_REQ_ERROR:
2660 event.event = RDMA_CM_EVENT_UNREACHABLE;
2661 event.status = -ETIMEDOUT;
2663 case IB_CM_SIDR_REP_RECEIVED:
2664 event.param.ud.private_data = ib_event->private_data;
2665 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
2666 if (rep->status != IB_SIDR_SUCCESS) {
2667 event.event = RDMA_CM_EVENT_UNREACHABLE;
2668 event.status = ib_event->param.sidr_rep_rcvd.status;
2671 ret = cma_set_qkey(id_priv, rep->qkey);
2673 event.event = RDMA_CM_EVENT_ADDR_ERROR;
2677 ib_init_ah_from_path(id_priv->id.device, id_priv->id.port_num,
2678 id_priv->id.route.path_rec,
2679 &event.param.ud.ah_attr);
2680 event.param.ud.qp_num = rep->qpn;
2681 event.param.ud.qkey = rep->qkey;
2682 event.event = RDMA_CM_EVENT_ESTABLISHED;
2686 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
2691 ret = id_priv->id.event_handler(&id_priv->id, &event);
2693 /* Destroy the CM ID by returning a non-zero value. */
2694 id_priv->cm_id.ib = NULL;
2695 cma_exch(id_priv, RDMA_CM_DESTROYING);
2696 mutex_unlock(&id_priv->handler_mutex);
2697 rdma_destroy_id(&id_priv->id);
2701 mutex_unlock(&id_priv->handler_mutex);
2705 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
2706 struct rdma_conn_param *conn_param)
2708 struct ib_cm_sidr_req_param req;
2709 struct ib_cm_id *id;
2713 memset(&req, 0, sizeof req);
2714 offset = cma_user_data_offset(id_priv);
2715 req.private_data_len = offset + conn_param->private_data_len;
2716 if (req.private_data_len < conn_param->private_data_len)
2719 if (req.private_data_len) {
2720 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2724 private_data = NULL;
2727 if (conn_param->private_data && conn_param->private_data_len)
2728 memcpy(private_data + offset, conn_param->private_data,
2729 conn_param->private_data_len);
2732 ret = cma_format_hdr(private_data, id_priv);
2735 req.private_data = private_data;
2738 id = ib_create_cm_id(id_priv->id.device, cma_sidr_rep_handler,
2744 id_priv->cm_id.ib = id;
2746 req.path = id_priv->id.route.path_rec;
2747 req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
2748 req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
2749 req.max_cm_retries = CMA_MAX_CM_RETRIES;
2751 ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
2753 ib_destroy_cm_id(id_priv->cm_id.ib);
2754 id_priv->cm_id.ib = NULL;
2757 kfree(private_data);
2761 static int cma_connect_ib(struct rdma_id_private *id_priv,
2762 struct rdma_conn_param *conn_param)
2764 struct ib_cm_req_param req;
2765 struct rdma_route *route;
2767 struct ib_cm_id *id;
2770 memset(&req, 0, sizeof req);
2771 offset = cma_user_data_offset(id_priv);
2772 req.private_data_len = offset + conn_param->private_data_len;
2773 if (req.private_data_len < conn_param->private_data_len)
2776 if (req.private_data_len) {
2777 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2781 private_data = NULL;
2784 if (conn_param->private_data && conn_param->private_data_len)
2785 memcpy(private_data + offset, conn_param->private_data,
2786 conn_param->private_data_len);
2788 id = ib_create_cm_id(id_priv->id.device, cma_ib_handler, id_priv);
2793 id_priv->cm_id.ib = id;
2795 route = &id_priv->id.route;
2797 ret = cma_format_hdr(private_data, id_priv);
2800 req.private_data = private_data;
2803 req.primary_path = &route->path_rec[0];
2804 if (route->num_paths == 2)
2805 req.alternate_path = &route->path_rec[1];
2807 req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
2808 req.qp_num = id_priv->qp_num;
2809 req.qp_type = id_priv->id.qp_type;
2810 req.starting_psn = id_priv->seq_num;
2811 req.responder_resources = conn_param->responder_resources;
2812 req.initiator_depth = conn_param->initiator_depth;
2813 req.flow_control = conn_param->flow_control;
2814 req.retry_count = min_t(u8, 7, conn_param->retry_count);
2815 req.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
2816 req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2817 req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2818 req.max_cm_retries = CMA_MAX_CM_RETRIES;
2819 req.srq = id_priv->srq ? 1 : 0;
2821 ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
2823 if (ret && !IS_ERR(id)) {
2824 ib_destroy_cm_id(id);
2825 id_priv->cm_id.ib = NULL;
2828 kfree(private_data);
2832 static int cma_connect_iw(struct rdma_id_private *id_priv,
2833 struct rdma_conn_param *conn_param)
2835 struct iw_cm_id *cm_id;
2837 struct iw_cm_conn_param iw_param;
2839 cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
2841 return PTR_ERR(cm_id);
2843 id_priv->cm_id.iw = cm_id;
2845 memcpy(&cm_id->local_addr, cma_src_addr(id_priv),
2846 rdma_addr_size(cma_src_addr(id_priv)));
2847 memcpy(&cm_id->remote_addr, cma_dst_addr(id_priv),
2848 rdma_addr_size(cma_dst_addr(id_priv)));
2850 ret = cma_modify_qp_rtr(id_priv, conn_param);
2855 iw_param.ord = conn_param->initiator_depth;
2856 iw_param.ird = conn_param->responder_resources;
2857 iw_param.private_data = conn_param->private_data;
2858 iw_param.private_data_len = conn_param->private_data_len;
2859 iw_param.qpn = id_priv->id.qp ? id_priv->qp_num : conn_param->qp_num;
2861 memset(&iw_param, 0, sizeof iw_param);
2862 iw_param.qpn = id_priv->qp_num;
2864 ret = iw_cm_connect(cm_id, &iw_param);
2867 iw_destroy_cm_id(cm_id);
2868 id_priv->cm_id.iw = NULL;
2873 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2875 struct rdma_id_private *id_priv;
2878 id_priv = container_of(id, struct rdma_id_private, id);
2879 if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT))
2883 id_priv->qp_num = conn_param->qp_num;
2884 id_priv->srq = conn_param->srq;
2887 switch (rdma_node_get_transport(id->device->node_type)) {
2888 case RDMA_TRANSPORT_IB:
2889 if (id->qp_type == IB_QPT_UD)
2890 ret = cma_resolve_ib_udp(id_priv, conn_param);
2892 ret = cma_connect_ib(id_priv, conn_param);
2894 case RDMA_TRANSPORT_IWARP:
2895 ret = cma_connect_iw(id_priv, conn_param);
2906 cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED);
2909 EXPORT_SYMBOL(rdma_connect);
2911 static int cma_accept_ib(struct rdma_id_private *id_priv,
2912 struct rdma_conn_param *conn_param)
2914 struct ib_cm_rep_param rep;
2917 ret = cma_modify_qp_rtr(id_priv, conn_param);
2921 ret = cma_modify_qp_rts(id_priv, conn_param);
2925 memset(&rep, 0, sizeof rep);
2926 rep.qp_num = id_priv->qp_num;
2927 rep.starting_psn = id_priv->seq_num;
2928 rep.private_data = conn_param->private_data;
2929 rep.private_data_len = conn_param->private_data_len;
2930 rep.responder_resources = conn_param->responder_resources;
2931 rep.initiator_depth = conn_param->initiator_depth;
2932 rep.failover_accepted = 0;
2933 rep.flow_control = conn_param->flow_control;
2934 rep.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
2935 rep.srq = id_priv->srq ? 1 : 0;
2937 ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
2942 static int cma_accept_iw(struct rdma_id_private *id_priv,
2943 struct rdma_conn_param *conn_param)
2945 struct iw_cm_conn_param iw_param;
2948 ret = cma_modify_qp_rtr(id_priv, conn_param);
2952 iw_param.ord = conn_param->initiator_depth;
2953 iw_param.ird = conn_param->responder_resources;
2954 iw_param.private_data = conn_param->private_data;
2955 iw_param.private_data_len = conn_param->private_data_len;
2956 if (id_priv->id.qp) {
2957 iw_param.qpn = id_priv->qp_num;
2959 iw_param.qpn = conn_param->qp_num;
2961 return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
2964 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
2965 enum ib_cm_sidr_status status, u32 qkey,
2966 const void *private_data, int private_data_len)
2968 struct ib_cm_sidr_rep_param rep;
2971 memset(&rep, 0, sizeof rep);
2972 rep.status = status;
2973 if (status == IB_SIDR_SUCCESS) {
2974 ret = cma_set_qkey(id_priv, qkey);
2977 rep.qp_num = id_priv->qp_num;
2978 rep.qkey = id_priv->qkey;
2980 rep.private_data = private_data;
2981 rep.private_data_len = private_data_len;
2983 return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
2986 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2988 struct rdma_id_private *id_priv;
2991 id_priv = container_of(id, struct rdma_id_private, id);
2993 id_priv->owner = task_pid_nr(current);
2995 if (!cma_comp(id_priv, RDMA_CM_CONNECT))
2998 if (!id->qp && conn_param) {
2999 id_priv->qp_num = conn_param->qp_num;
3000 id_priv->srq = conn_param->srq;
3003 switch (rdma_node_get_transport(id->device->node_type)) {
3004 case RDMA_TRANSPORT_IB:
3005 if (id->qp_type == IB_QPT_UD) {
3007 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
3009 conn_param->private_data,
3010 conn_param->private_data_len);
3012 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
3016 ret = cma_accept_ib(id_priv, conn_param);
3018 ret = cma_rep_recv(id_priv);
3021 case RDMA_TRANSPORT_IWARP:
3022 ret = cma_accept_iw(id_priv, conn_param);
3034 cma_modify_qp_err(id_priv);
3035 rdma_reject(id, NULL, 0);
3038 EXPORT_SYMBOL(rdma_accept);
3040 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
3042 struct rdma_id_private *id_priv;
3045 id_priv = container_of(id, struct rdma_id_private, id);
3046 if (!id_priv->cm_id.ib)
3049 switch (id->device->node_type) {
3050 case RDMA_NODE_IB_CA:
3051 ret = ib_cm_notify(id_priv->cm_id.ib, event);
3059 EXPORT_SYMBOL(rdma_notify);
3061 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
3062 u8 private_data_len)
3064 struct rdma_id_private *id_priv;
3067 id_priv = container_of(id, struct rdma_id_private, id);
3068 if (!id_priv->cm_id.ib)
3071 switch (rdma_node_get_transport(id->device->node_type)) {
3072 case RDMA_TRANSPORT_IB:
3073 if (id->qp_type == IB_QPT_UD)
3074 ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT, 0,
3075 private_data, private_data_len);
3077 ret = ib_send_cm_rej(id_priv->cm_id.ib,
3078 IB_CM_REJ_CONSUMER_DEFINED, NULL,
3079 0, private_data, private_data_len);
3081 case RDMA_TRANSPORT_IWARP:
3082 ret = iw_cm_reject(id_priv->cm_id.iw,
3083 private_data, private_data_len);
3091 EXPORT_SYMBOL(rdma_reject);
3093 int rdma_disconnect(struct rdma_cm_id *id)
3095 struct rdma_id_private *id_priv;
3098 id_priv = container_of(id, struct rdma_id_private, id);
3099 if (!id_priv->cm_id.ib)
3102 switch (rdma_node_get_transport(id->device->node_type)) {
3103 case RDMA_TRANSPORT_IB:
3104 ret = cma_modify_qp_err(id_priv);
3107 /* Initiate or respond to a disconnect. */
3108 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
3109 ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
3111 case RDMA_TRANSPORT_IWARP:
3112 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
3121 EXPORT_SYMBOL(rdma_disconnect);
3123 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
3125 struct rdma_id_private *id_priv;
3126 struct cma_multicast *mc = multicast->context;
3127 struct rdma_cm_event event;
3130 id_priv = mc->id_priv;
3131 if (cma_disable_callback(id_priv, RDMA_CM_ADDR_BOUND) &&
3132 cma_disable_callback(id_priv, RDMA_CM_ADDR_RESOLVED))
3136 status = cma_set_qkey(id_priv, be32_to_cpu(multicast->rec.qkey));
3137 mutex_lock(&id_priv->qp_mutex);
3138 if (!status && id_priv->id.qp)
3139 status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid,
3140 be16_to_cpu(multicast->rec.mlid));
3141 mutex_unlock(&id_priv->qp_mutex);
3143 memset(&event, 0, sizeof event);
3144 event.status = status;
3145 event.param.ud.private_data = mc->context;
3147 event.event = RDMA_CM_EVENT_MULTICAST_JOIN;
3148 ib_init_ah_from_mcmember(id_priv->id.device,
3149 id_priv->id.port_num, &multicast->rec,
3150 &event.param.ud.ah_attr);
3151 event.param.ud.qp_num = 0xFFFFFF;
3152 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
3154 event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
3156 ret = id_priv->id.event_handler(&id_priv->id, &event);
3158 cma_exch(id_priv, RDMA_CM_DESTROYING);
3159 mutex_unlock(&id_priv->handler_mutex);
3160 rdma_destroy_id(&id_priv->id);
3164 mutex_unlock(&id_priv->handler_mutex);
3168 static void cma_set_mgid(struct rdma_id_private *id_priv,
3169 struct sockaddr *addr, union ib_gid *mgid)
3171 unsigned char mc_map[MAX_ADDR_LEN];
3172 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3173 struct sockaddr_in *sin = (struct sockaddr_in *) addr;
3174 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
3176 if (cma_any_addr(addr)) {
3177 memset(mgid, 0, sizeof *mgid);
3178 } else if ((addr->sa_family == AF_INET6) &&
3179 ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
3181 /* IPv6 address is an SA assigned MGID. */
3182 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3183 } else if (addr->sa_family == AF_IB) {
3184 memcpy(mgid, &((struct sockaddr_ib *) addr)->sib_addr, sizeof *mgid);
3185 } else if ((addr->sa_family == AF_INET6)) {
3186 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
3187 if (id_priv->id.ps == RDMA_PS_UDP)
3188 mc_map[7] = 0x01; /* Use RDMA CM signature */
3189 *mgid = *(union ib_gid *) (mc_map + 4);
3191 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
3192 if (id_priv->id.ps == RDMA_PS_UDP)
3193 mc_map[7] = 0x01; /* Use RDMA CM signature */
3194 *mgid = *(union ib_gid *) (mc_map + 4);
3198 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
3199 struct cma_multicast *mc)
3201 struct ib_sa_mcmember_rec rec;
3202 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3203 ib_sa_comp_mask comp_mask;
3206 ib_addr_get_mgid(dev_addr, &rec.mgid);
3207 ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
3212 ret = cma_set_qkey(id_priv, 0);
3216 cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
3217 rec.qkey = cpu_to_be32(id_priv->qkey);
3218 rdma_addr_get_sgid(dev_addr, &rec.port_gid);
3219 rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
3222 comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
3223 IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
3224 IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
3225 IB_SA_MCMEMBER_REC_FLOW_LABEL |
3226 IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
3228 if (id_priv->id.ps == RDMA_PS_IPOIB)
3229 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
3230 IB_SA_MCMEMBER_REC_RATE_SELECTOR |
3231 IB_SA_MCMEMBER_REC_MTU_SELECTOR |
3232 IB_SA_MCMEMBER_REC_MTU |
3233 IB_SA_MCMEMBER_REC_HOP_LIMIT;
3235 mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device,
3236 id_priv->id.port_num, &rec,
3237 comp_mask, GFP_KERNEL,
3238 cma_ib_mc_handler, mc);
3239 return PTR_ERR_OR_ZERO(mc->multicast.ib);
3242 static void iboe_mcast_work_handler(struct work_struct *work)
3244 struct iboe_mcast_work *mw = container_of(work, struct iboe_mcast_work, work);
3245 struct cma_multicast *mc = mw->mc;
3246 struct ib_sa_multicast *m = mc->multicast.ib;
3248 mc->multicast.ib->context = mc;
3249 cma_ib_mc_handler(0, m);
3250 kref_put(&mc->mcref, release_mc);
3254 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid)
3256 struct sockaddr_in *sin = (struct sockaddr_in *)addr;
3257 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
3259 if (cma_any_addr(addr)) {
3260 memset(mgid, 0, sizeof *mgid);
3261 } else if (addr->sa_family == AF_INET6) {
3262 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3264 mgid->raw[0] = 0xff;
3265 mgid->raw[1] = 0x0e;
3274 mgid->raw[10] = 0xff;
3275 mgid->raw[11] = 0xff;
3276 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
3280 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
3281 struct cma_multicast *mc)
3283 struct iboe_mcast_work *work;
3284 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3286 struct sockaddr *addr = (struct sockaddr *)&mc->addr;
3287 struct net_device *ndev = NULL;
3289 if (cma_zero_addr((struct sockaddr *)&mc->addr))
3292 work = kzalloc(sizeof *work, GFP_KERNEL);
3296 mc->multicast.ib = kzalloc(sizeof(struct ib_sa_multicast), GFP_KERNEL);
3297 if (!mc->multicast.ib) {
3302 cma_iboe_set_mgid(addr, &mc->multicast.ib->rec.mgid);
3304 mc->multicast.ib->rec.pkey = cpu_to_be16(0xffff);
3305 if (id_priv->id.ps == RDMA_PS_UDP)
3306 mc->multicast.ib->rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
3308 if (dev_addr->bound_dev_if)
3309 ndev = dev_get_by_index(&init_net, dev_addr->bound_dev_if);
3314 mc->multicast.ib->rec.rate = iboe_get_rate(ndev);
3315 mc->multicast.ib->rec.hop_limit = 1;
3316 mc->multicast.ib->rec.mtu = iboe_get_mtu(ndev->mtu);
3318 if (!mc->multicast.ib->rec.mtu) {
3322 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
3323 &mc->multicast.ib->rec.port_gid);
3326 INIT_WORK(&work->work, iboe_mcast_work_handler);
3327 kref_get(&mc->mcref);
3328 queue_work(cma_wq, &work->work);
3333 kfree(mc->multicast.ib);
3339 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
3342 struct rdma_id_private *id_priv;
3343 struct cma_multicast *mc;
3346 id_priv = container_of(id, struct rdma_id_private, id);
3347 if (!cma_comp(id_priv, RDMA_CM_ADDR_BOUND) &&
3348 !cma_comp(id_priv, RDMA_CM_ADDR_RESOLVED))
3351 mc = kmalloc(sizeof *mc, GFP_KERNEL);
3355 memcpy(&mc->addr, addr, rdma_addr_size(addr));
3356 mc->context = context;
3357 mc->id_priv = id_priv;
3359 spin_lock(&id_priv->lock);
3360 list_add(&mc->list, &id_priv->mc_list);
3361 spin_unlock(&id_priv->lock);
3363 switch (rdma_node_get_transport(id->device->node_type)) {
3364 case RDMA_TRANSPORT_IB:
3365 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
3366 case IB_LINK_LAYER_INFINIBAND:
3367 ret = cma_join_ib_multicast(id_priv, mc);
3369 case IB_LINK_LAYER_ETHERNET:
3370 kref_init(&mc->mcref);
3371 ret = cma_iboe_join_multicast(id_priv, mc);
3383 spin_lock_irq(&id_priv->lock);
3384 list_del(&mc->list);
3385 spin_unlock_irq(&id_priv->lock);
3390 EXPORT_SYMBOL(rdma_join_multicast);
3392 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
3394 struct rdma_id_private *id_priv;
3395 struct cma_multicast *mc;
3397 id_priv = container_of(id, struct rdma_id_private, id);
3398 spin_lock_irq(&id_priv->lock);
3399 list_for_each_entry(mc, &id_priv->mc_list, list) {
3400 if (!memcmp(&mc->addr, addr, rdma_addr_size(addr))) {
3401 list_del(&mc->list);
3402 spin_unlock_irq(&id_priv->lock);
3405 ib_detach_mcast(id->qp,
3406 &mc->multicast.ib->rec.mgid,
3407 be16_to_cpu(mc->multicast.ib->rec.mlid));
3408 if (rdma_node_get_transport(id_priv->cma_dev->device->node_type) == RDMA_TRANSPORT_IB) {
3409 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
3410 case IB_LINK_LAYER_INFINIBAND:
3411 ib_sa_free_multicast(mc->multicast.ib);
3414 case IB_LINK_LAYER_ETHERNET:
3415 kref_put(&mc->mcref, release_mc);
3424 spin_unlock_irq(&id_priv->lock);
3426 EXPORT_SYMBOL(rdma_leave_multicast);
3428 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
3430 struct rdma_dev_addr *dev_addr;
3431 struct cma_ndev_work *work;
3433 dev_addr = &id_priv->id.route.addr.dev_addr;
3435 if ((dev_addr->bound_dev_if == ndev->ifindex) &&
3436 memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
3437 printk(KERN_INFO "RDMA CM addr change for ndev %s used by id %p\n",
3438 ndev->name, &id_priv->id);
3439 work = kzalloc(sizeof *work, GFP_KERNEL);
3443 INIT_WORK(&work->work, cma_ndev_work_handler);
3445 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
3446 atomic_inc(&id_priv->refcount);
3447 queue_work(cma_wq, &work->work);
3453 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
3456 struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
3457 struct cma_device *cma_dev;
3458 struct rdma_id_private *id_priv;
3459 int ret = NOTIFY_DONE;
3461 if (dev_net(ndev) != &init_net)
3464 if (event != NETDEV_BONDING_FAILOVER)
3467 if (!(ndev->flags & IFF_MASTER) || !(ndev->priv_flags & IFF_BONDING))
3471 list_for_each_entry(cma_dev, &dev_list, list)
3472 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
3473 ret = cma_netdev_change(ndev, id_priv);
3479 mutex_unlock(&lock);
3483 static struct notifier_block cma_nb = {
3484 .notifier_call = cma_netdev_callback
3487 static void cma_add_one(struct ib_device *device)
3489 struct cma_device *cma_dev;
3490 struct rdma_id_private *id_priv;
3492 cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
3496 cma_dev->device = device;
3498 init_completion(&cma_dev->comp);
3499 atomic_set(&cma_dev->refcount, 1);
3500 INIT_LIST_HEAD(&cma_dev->id_list);
3501 ib_set_client_data(device, &cma_client, cma_dev);
3504 list_add_tail(&cma_dev->list, &dev_list);
3505 list_for_each_entry(id_priv, &listen_any_list, list)
3506 cma_listen_on_dev(id_priv, cma_dev);
3507 mutex_unlock(&lock);
3510 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
3512 struct rdma_cm_event event;
3513 enum rdma_cm_state state;
3516 /* Record that we want to remove the device */
3517 state = cma_exch(id_priv, RDMA_CM_DEVICE_REMOVAL);
3518 if (state == RDMA_CM_DESTROYING)
3521 cma_cancel_operation(id_priv, state);
3522 mutex_lock(&id_priv->handler_mutex);
3524 /* Check for destruction from another callback. */
3525 if (!cma_comp(id_priv, RDMA_CM_DEVICE_REMOVAL))
3528 memset(&event, 0, sizeof event);
3529 event.event = RDMA_CM_EVENT_DEVICE_REMOVAL;
3530 ret = id_priv->id.event_handler(&id_priv->id, &event);
3532 mutex_unlock(&id_priv->handler_mutex);
3536 static void cma_process_remove(struct cma_device *cma_dev)
3538 struct rdma_id_private *id_priv;
3542 while (!list_empty(&cma_dev->id_list)) {
3543 id_priv = list_entry(cma_dev->id_list.next,
3544 struct rdma_id_private, list);
3546 list_del(&id_priv->listen_list);
3547 list_del_init(&id_priv->list);
3548 atomic_inc(&id_priv->refcount);
3549 mutex_unlock(&lock);
3551 ret = id_priv->internal_id ? 1 : cma_remove_id_dev(id_priv);
3552 cma_deref_id(id_priv);
3554 rdma_destroy_id(&id_priv->id);
3558 mutex_unlock(&lock);
3560 cma_deref_dev(cma_dev);
3561 wait_for_completion(&cma_dev->comp);
3564 static void cma_remove_one(struct ib_device *device)
3566 struct cma_device *cma_dev;
3568 cma_dev = ib_get_client_data(device, &cma_client);
3573 list_del(&cma_dev->list);
3574 mutex_unlock(&lock);
3576 cma_process_remove(cma_dev);
3580 static int cma_get_id_stats(struct sk_buff *skb, struct netlink_callback *cb)
3582 struct nlmsghdr *nlh;
3583 struct rdma_cm_id_stats *id_stats;
3584 struct rdma_id_private *id_priv;
3585 struct rdma_cm_id *id = NULL;
3586 struct cma_device *cma_dev;
3587 int i_dev = 0, i_id = 0;
3590 * We export all of the IDs as a sequence of messages. Each
3591 * ID gets its own netlink message.
3595 list_for_each_entry(cma_dev, &dev_list, list) {
3596 if (i_dev < cb->args[0]) {
3602 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
3603 if (i_id < cb->args[1]) {
3608 id_stats = ibnl_put_msg(skb, &nlh, cb->nlh->nlmsg_seq,
3609 sizeof *id_stats, RDMA_NL_RDMA_CM,
3610 RDMA_NL_RDMA_CM_ID_STATS,
3615 memset(id_stats, 0, sizeof *id_stats);
3617 id_stats->node_type = id->route.addr.dev_addr.dev_type;
3618 id_stats->port_num = id->port_num;
3619 id_stats->bound_dev_if =
3620 id->route.addr.dev_addr.bound_dev_if;
3622 if (ibnl_put_attr(skb, nlh,
3623 rdma_addr_size(cma_src_addr(id_priv)),
3624 cma_src_addr(id_priv),
3625 RDMA_NL_RDMA_CM_ATTR_SRC_ADDR))
3627 if (ibnl_put_attr(skb, nlh,
3628 rdma_addr_size(cma_src_addr(id_priv)),
3629 cma_dst_addr(id_priv),
3630 RDMA_NL_RDMA_CM_ATTR_DST_ADDR))
3633 id_stats->pid = id_priv->owner;
3634 id_stats->port_space = id->ps;
3635 id_stats->cm_state = id_priv->state;
3636 id_stats->qp_num = id_priv->qp_num;
3637 id_stats->qp_type = id->qp_type;
3647 mutex_unlock(&lock);
3648 cb->args[0] = i_dev;
3654 static const struct ibnl_client_cbs cma_cb_table[] = {
3655 [RDMA_NL_RDMA_CM_ID_STATS] = { .dump = cma_get_id_stats,
3656 .module = THIS_MODULE },
3659 static int __init cma_init(void)
3663 cma_wq = create_singlethread_workqueue("rdma_cm");
3667 ib_sa_register_client(&sa_client);
3668 rdma_addr_register_client(&addr_client);
3669 register_netdevice_notifier(&cma_nb);
3671 ret = ib_register_client(&cma_client);
3675 if (ibnl_add_client(RDMA_NL_RDMA_CM, RDMA_NL_RDMA_CM_NUM_OPS, cma_cb_table))
3676 printk(KERN_WARNING "RDMA CMA: failed to add netlink callback\n");
3681 unregister_netdevice_notifier(&cma_nb);
3682 rdma_addr_unregister_client(&addr_client);
3683 ib_sa_unregister_client(&sa_client);
3684 destroy_workqueue(cma_wq);
3688 static void __exit cma_cleanup(void)
3690 ibnl_remove_client(RDMA_NL_RDMA_CM);
3691 ib_unregister_client(&cma_client);
3692 unregister_netdevice_notifier(&cma_nb);
3693 rdma_addr_unregister_client(&addr_client);
3694 ib_sa_unregister_client(&sa_client);
3695 destroy_workqueue(cma_wq);
3696 idr_destroy(&tcp_ps);
3697 idr_destroy(&udp_ps);
3698 idr_destroy(&ipoib_ps);
3699 idr_destroy(&ib_ps);
3702 module_init(cma_init);
3703 module_exit(cma_cleanup);