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 __be16 ss_get_port(const struct sockaddr_storage *ss)
864 if (ss->ss_family == AF_INET)
865 return ((struct sockaddr_in *)ss)->sin_port;
866 else if (ss->ss_family == AF_INET6)
867 return ((struct sockaddr_in6 *)ss)->sin6_port;
871 static void cma_save_ip4_info(struct rdma_cm_id *id, struct rdma_cm_id *listen_id,
874 struct sockaddr_in *ip4;
876 ip4 = (struct sockaddr_in *) &id->route.addr.src_addr;
877 ip4->sin_family = AF_INET;
878 ip4->sin_addr.s_addr = hdr->dst_addr.ip4.addr;
879 ip4->sin_port = ss_get_port(&listen_id->route.addr.src_addr);
881 ip4 = (struct sockaddr_in *) &id->route.addr.dst_addr;
882 ip4->sin_family = AF_INET;
883 ip4->sin_addr.s_addr = hdr->src_addr.ip4.addr;
884 ip4->sin_port = hdr->port;
887 static void cma_save_ip6_info(struct rdma_cm_id *id, struct rdma_cm_id *listen_id,
890 struct sockaddr_in6 *ip6;
892 ip6 = (struct sockaddr_in6 *) &id->route.addr.src_addr;
893 ip6->sin6_family = AF_INET6;
894 ip6->sin6_addr = hdr->dst_addr.ip6;
895 ip6->sin6_port = ss_get_port(&listen_id->route.addr.src_addr);
897 ip6 = (struct sockaddr_in6 *) &id->route.addr.dst_addr;
898 ip6->sin6_family = AF_INET6;
899 ip6->sin6_addr = hdr->src_addr.ip6;
900 ip6->sin6_port = hdr->port;
903 static int cma_save_net_info(struct rdma_cm_id *id, struct rdma_cm_id *listen_id,
904 struct ib_cm_event *ib_event)
908 if ((listen_id->route.addr.src_addr.ss_family == AF_IB) &&
909 (ib_event->event == IB_CM_REQ_RECEIVED)) {
910 cma_save_ib_info(id, listen_id, ib_event->param.req_rcvd.primary_path);
914 hdr = ib_event->private_data;
915 if (hdr->cma_version != CMA_VERSION)
918 switch (cma_get_ip_ver(hdr)) {
920 cma_save_ip4_info(id, listen_id, hdr);
923 cma_save_ip6_info(id, listen_id, hdr);
931 static inline int cma_user_data_offset(struct rdma_id_private *id_priv)
933 return cma_family(id_priv) == AF_IB ? 0 : sizeof(struct cma_hdr);
936 static void cma_cancel_route(struct rdma_id_private *id_priv)
938 switch (rdma_port_get_link_layer(id_priv->id.device, id_priv->id.port_num)) {
939 case IB_LINK_LAYER_INFINIBAND:
941 ib_sa_cancel_query(id_priv->query_id, id_priv->query);
948 static void cma_cancel_listens(struct rdma_id_private *id_priv)
950 struct rdma_id_private *dev_id_priv;
953 * Remove from listen_any_list to prevent added devices from spawning
954 * additional listen requests.
957 list_del(&id_priv->list);
959 while (!list_empty(&id_priv->listen_list)) {
960 dev_id_priv = list_entry(id_priv->listen_list.next,
961 struct rdma_id_private, listen_list);
962 /* sync with device removal to avoid duplicate destruction */
963 list_del_init(&dev_id_priv->list);
964 list_del(&dev_id_priv->listen_list);
967 rdma_destroy_id(&dev_id_priv->id);
973 static void cma_cancel_operation(struct rdma_id_private *id_priv,
974 enum rdma_cm_state state)
977 case RDMA_CM_ADDR_QUERY:
978 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
980 case RDMA_CM_ROUTE_QUERY:
981 cma_cancel_route(id_priv);
984 if (cma_any_addr(cma_src_addr(id_priv)) && !id_priv->cma_dev)
985 cma_cancel_listens(id_priv);
992 static void cma_release_port(struct rdma_id_private *id_priv)
994 struct rdma_bind_list *bind_list = id_priv->bind_list;
1000 hlist_del(&id_priv->node);
1001 if (hlist_empty(&bind_list->owners)) {
1002 idr_remove(bind_list->ps, bind_list->port);
1005 mutex_unlock(&lock);
1008 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
1010 struct cma_multicast *mc;
1012 while (!list_empty(&id_priv->mc_list)) {
1013 mc = container_of(id_priv->mc_list.next,
1014 struct cma_multicast, list);
1015 list_del(&mc->list);
1016 switch (rdma_port_get_link_layer(id_priv->cma_dev->device, id_priv->id.port_num)) {
1017 case IB_LINK_LAYER_INFINIBAND:
1018 ib_sa_free_multicast(mc->multicast.ib);
1021 case IB_LINK_LAYER_ETHERNET:
1022 kref_put(&mc->mcref, release_mc);
1030 void rdma_destroy_id(struct rdma_cm_id *id)
1032 struct rdma_id_private *id_priv;
1033 enum rdma_cm_state state;
1035 id_priv = container_of(id, struct rdma_id_private, id);
1036 state = cma_exch(id_priv, RDMA_CM_DESTROYING);
1037 cma_cancel_operation(id_priv, state);
1040 * Wait for any active callback to finish. New callbacks will find
1041 * the id_priv state set to destroying and abort.
1043 mutex_lock(&id_priv->handler_mutex);
1044 mutex_unlock(&id_priv->handler_mutex);
1046 if (id_priv->cma_dev) {
1047 switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
1048 case RDMA_TRANSPORT_IB:
1049 if (id_priv->cm_id.ib)
1050 ib_destroy_cm_id(id_priv->cm_id.ib);
1052 case RDMA_TRANSPORT_IWARP:
1053 if (id_priv->cm_id.iw)
1054 iw_destroy_cm_id(id_priv->cm_id.iw);
1059 cma_leave_mc_groups(id_priv);
1060 cma_release_dev(id_priv);
1063 cma_release_port(id_priv);
1064 cma_deref_id(id_priv);
1065 wait_for_completion(&id_priv->comp);
1067 if (id_priv->internal_id)
1068 cma_deref_id(id_priv->id.context);
1070 kfree(id_priv->id.route.path_rec);
1073 EXPORT_SYMBOL(rdma_destroy_id);
1075 static int cma_rep_recv(struct rdma_id_private *id_priv)
1079 ret = cma_modify_qp_rtr(id_priv, NULL);
1083 ret = cma_modify_qp_rts(id_priv, NULL);
1087 ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
1093 cma_modify_qp_err(id_priv);
1094 ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
1099 static void cma_set_rep_event_data(struct rdma_cm_event *event,
1100 struct ib_cm_rep_event_param *rep_data,
1103 event->param.conn.private_data = private_data;
1104 event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
1105 event->param.conn.responder_resources = rep_data->responder_resources;
1106 event->param.conn.initiator_depth = rep_data->initiator_depth;
1107 event->param.conn.flow_control = rep_data->flow_control;
1108 event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
1109 event->param.conn.srq = rep_data->srq;
1110 event->param.conn.qp_num = rep_data->remote_qpn;
1113 static int cma_ib_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1115 struct rdma_id_private *id_priv = cm_id->context;
1116 struct rdma_cm_event event;
1119 if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
1120 cma_disable_callback(id_priv, RDMA_CM_CONNECT)) ||
1121 (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
1122 cma_disable_callback(id_priv, RDMA_CM_DISCONNECT)))
1125 memset(&event, 0, sizeof event);
1126 switch (ib_event->event) {
1127 case IB_CM_REQ_ERROR:
1128 case IB_CM_REP_ERROR:
1129 event.event = RDMA_CM_EVENT_UNREACHABLE;
1130 event.status = -ETIMEDOUT;
1132 case IB_CM_REP_RECEIVED:
1133 if (id_priv->id.qp) {
1134 event.status = cma_rep_recv(id_priv);
1135 event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
1136 RDMA_CM_EVENT_ESTABLISHED;
1138 event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
1140 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
1141 ib_event->private_data);
1143 case IB_CM_RTU_RECEIVED:
1144 case IB_CM_USER_ESTABLISHED:
1145 event.event = RDMA_CM_EVENT_ESTABLISHED;
1147 case IB_CM_DREQ_ERROR:
1148 event.status = -ETIMEDOUT; /* fall through */
1149 case IB_CM_DREQ_RECEIVED:
1150 case IB_CM_DREP_RECEIVED:
1151 if (!cma_comp_exch(id_priv, RDMA_CM_CONNECT,
1152 RDMA_CM_DISCONNECT))
1154 event.event = RDMA_CM_EVENT_DISCONNECTED;
1156 case IB_CM_TIMEWAIT_EXIT:
1157 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
1159 case IB_CM_MRA_RECEIVED:
1162 case IB_CM_REJ_RECEIVED:
1163 cma_modify_qp_err(id_priv);
1164 event.status = ib_event->param.rej_rcvd.reason;
1165 event.event = RDMA_CM_EVENT_REJECTED;
1166 event.param.conn.private_data = ib_event->private_data;
1167 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
1170 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
1175 ret = id_priv->id.event_handler(&id_priv->id, &event);
1177 /* Destroy the CM ID by returning a non-zero value. */
1178 id_priv->cm_id.ib = NULL;
1179 cma_exch(id_priv, RDMA_CM_DESTROYING);
1180 mutex_unlock(&id_priv->handler_mutex);
1181 rdma_destroy_id(&id_priv->id);
1185 mutex_unlock(&id_priv->handler_mutex);
1189 static struct rdma_id_private *cma_new_conn_id(struct rdma_cm_id *listen_id,
1190 struct ib_cm_event *ib_event)
1192 struct rdma_id_private *id_priv;
1193 struct rdma_cm_id *id;
1194 struct rdma_route *rt;
1197 id = rdma_create_id(listen_id->event_handler, listen_id->context,
1198 listen_id->ps, ib_event->param.req_rcvd.qp_type);
1202 id_priv = container_of(id, struct rdma_id_private, id);
1203 if (cma_save_net_info(id, listen_id, ib_event))
1207 rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
1208 rt->path_rec = kmalloc(sizeof *rt->path_rec * rt->num_paths,
1213 rt->path_rec[0] = *ib_event->param.req_rcvd.primary_path;
1214 if (rt->num_paths == 2)
1215 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
1217 if (cma_any_addr(cma_src_addr(id_priv))) {
1218 rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
1219 rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
1220 ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
1222 ret = cma_translate_addr(cma_src_addr(id_priv), &rt->addr.dev_addr);
1226 rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
1228 id_priv->state = RDMA_CM_CONNECT;
1232 rdma_destroy_id(id);
1236 static struct rdma_id_private *cma_new_udp_id(struct rdma_cm_id *listen_id,
1237 struct ib_cm_event *ib_event)
1239 struct rdma_id_private *id_priv;
1240 struct rdma_cm_id *id;
1243 id = rdma_create_id(listen_id->event_handler, listen_id->context,
1244 listen_id->ps, IB_QPT_UD);
1248 id_priv = container_of(id, struct rdma_id_private, id);
1249 if (cma_save_net_info(id, listen_id, ib_event))
1252 if (!cma_any_addr((struct sockaddr *) &id->route.addr.src_addr)) {
1253 ret = cma_translate_addr(cma_src_addr(id_priv), &id->route.addr.dev_addr);
1258 id_priv->state = RDMA_CM_CONNECT;
1261 rdma_destroy_id(id);
1265 static void cma_set_req_event_data(struct rdma_cm_event *event,
1266 struct ib_cm_req_event_param *req_data,
1267 void *private_data, int offset)
1269 event->param.conn.private_data = private_data + offset;
1270 event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
1271 event->param.conn.responder_resources = req_data->responder_resources;
1272 event->param.conn.initiator_depth = req_data->initiator_depth;
1273 event->param.conn.flow_control = req_data->flow_control;
1274 event->param.conn.retry_count = req_data->retry_count;
1275 event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
1276 event->param.conn.srq = req_data->srq;
1277 event->param.conn.qp_num = req_data->remote_qpn;
1280 static int cma_check_req_qp_type(struct rdma_cm_id *id, struct ib_cm_event *ib_event)
1282 return (((ib_event->event == IB_CM_REQ_RECEIVED) &&
1283 (ib_event->param.req_rcvd.qp_type == id->qp_type)) ||
1284 ((ib_event->event == IB_CM_SIDR_REQ_RECEIVED) &&
1285 (id->qp_type == IB_QPT_UD)) ||
1289 static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1291 struct rdma_id_private *listen_id, *conn_id;
1292 struct rdma_cm_event event;
1295 listen_id = cm_id->context;
1296 if (!cma_check_req_qp_type(&listen_id->id, ib_event))
1299 if (cma_disable_callback(listen_id, RDMA_CM_LISTEN))
1300 return -ECONNABORTED;
1302 memset(&event, 0, sizeof event);
1303 offset = cma_user_data_offset(listen_id);
1304 event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1305 if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) {
1306 conn_id = cma_new_udp_id(&listen_id->id, ib_event);
1307 event.param.ud.private_data = ib_event->private_data + offset;
1308 event.param.ud.private_data_len =
1309 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
1311 conn_id = cma_new_conn_id(&listen_id->id, ib_event);
1312 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
1313 ib_event->private_data, offset);
1320 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1321 ret = cma_acquire_dev(conn_id, listen_id);
1325 conn_id->cm_id.ib = cm_id;
1326 cm_id->context = conn_id;
1327 cm_id->cm_handler = cma_ib_handler;
1330 * Protect against the user destroying conn_id from another thread
1331 * until we're done accessing it.
1333 atomic_inc(&conn_id->refcount);
1334 ret = conn_id->id.event_handler(&conn_id->id, &event);
1338 * Acquire mutex to prevent user executing rdma_destroy_id()
1339 * while we're accessing the cm_id.
1342 if (cma_comp(conn_id, RDMA_CM_CONNECT) &&
1343 (conn_id->id.qp_type != IB_QPT_UD))
1344 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
1345 mutex_unlock(&lock);
1346 mutex_unlock(&conn_id->handler_mutex);
1347 mutex_unlock(&listen_id->handler_mutex);
1348 cma_deref_id(conn_id);
1352 cma_deref_id(conn_id);
1353 /* Destroy the CM ID by returning a non-zero value. */
1354 conn_id->cm_id.ib = NULL;
1356 cma_exch(conn_id, RDMA_CM_DESTROYING);
1357 mutex_unlock(&conn_id->handler_mutex);
1359 mutex_unlock(&listen_id->handler_mutex);
1361 rdma_destroy_id(&conn_id->id);
1365 __be64 rdma_get_service_id(struct rdma_cm_id *id, struct sockaddr *addr)
1367 if (addr->sa_family == AF_IB)
1368 return ((struct sockaddr_ib *) addr)->sib_sid;
1370 return cpu_to_be64(((u64)id->ps << 16) + be16_to_cpu(cma_port(addr)));
1372 EXPORT_SYMBOL(rdma_get_service_id);
1374 static void cma_set_compare_data(enum rdma_port_space ps, struct sockaddr *addr,
1375 struct ib_cm_compare_data *compare)
1377 struct cma_hdr *cma_data, *cma_mask;
1379 struct in6_addr ip6_addr;
1381 memset(compare, 0, sizeof *compare);
1382 cma_data = (void *) compare->data;
1383 cma_mask = (void *) compare->mask;
1385 switch (addr->sa_family) {
1387 ip4_addr = ((struct sockaddr_in *) addr)->sin_addr.s_addr;
1388 cma_set_ip_ver(cma_data, 4);
1389 cma_set_ip_ver(cma_mask, 0xF);
1390 if (!cma_any_addr(addr)) {
1391 cma_data->dst_addr.ip4.addr = ip4_addr;
1392 cma_mask->dst_addr.ip4.addr = htonl(~0);
1396 ip6_addr = ((struct sockaddr_in6 *) addr)->sin6_addr;
1397 cma_set_ip_ver(cma_data, 6);
1398 cma_set_ip_ver(cma_mask, 0xF);
1399 if (!cma_any_addr(addr)) {
1400 cma_data->dst_addr.ip6 = ip6_addr;
1401 memset(&cma_mask->dst_addr.ip6, 0xFF,
1402 sizeof cma_mask->dst_addr.ip6);
1410 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
1412 struct rdma_id_private *id_priv = iw_id->context;
1413 struct rdma_cm_event event;
1415 struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
1416 struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
1418 if (cma_disable_callback(id_priv, RDMA_CM_CONNECT))
1421 memset(&event, 0, sizeof event);
1422 switch (iw_event->event) {
1423 case IW_CM_EVENT_CLOSE:
1424 event.event = RDMA_CM_EVENT_DISCONNECTED;
1426 case IW_CM_EVENT_CONNECT_REPLY:
1427 memcpy(cma_src_addr(id_priv), laddr,
1428 rdma_addr_size(laddr));
1429 memcpy(cma_dst_addr(id_priv), raddr,
1430 rdma_addr_size(raddr));
1431 switch (iw_event->status) {
1433 event.event = RDMA_CM_EVENT_ESTABLISHED;
1434 event.param.conn.initiator_depth = iw_event->ird;
1435 event.param.conn.responder_resources = iw_event->ord;
1439 event.event = RDMA_CM_EVENT_REJECTED;
1442 event.event = RDMA_CM_EVENT_UNREACHABLE;
1445 event.event = RDMA_CM_EVENT_CONNECT_ERROR;
1449 case IW_CM_EVENT_ESTABLISHED:
1450 event.event = RDMA_CM_EVENT_ESTABLISHED;
1451 event.param.conn.initiator_depth = iw_event->ird;
1452 event.param.conn.responder_resources = iw_event->ord;
1458 event.status = iw_event->status;
1459 event.param.conn.private_data = iw_event->private_data;
1460 event.param.conn.private_data_len = iw_event->private_data_len;
1461 ret = id_priv->id.event_handler(&id_priv->id, &event);
1463 /* Destroy the CM ID by returning a non-zero value. */
1464 id_priv->cm_id.iw = NULL;
1465 cma_exch(id_priv, RDMA_CM_DESTROYING);
1466 mutex_unlock(&id_priv->handler_mutex);
1467 rdma_destroy_id(&id_priv->id);
1471 mutex_unlock(&id_priv->handler_mutex);
1475 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
1476 struct iw_cm_event *iw_event)
1478 struct rdma_cm_id *new_cm_id;
1479 struct rdma_id_private *listen_id, *conn_id;
1480 struct rdma_cm_event event;
1482 struct ib_device_attr attr;
1483 struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
1484 struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
1486 listen_id = cm_id->context;
1487 if (cma_disable_callback(listen_id, RDMA_CM_LISTEN))
1488 return -ECONNABORTED;
1490 /* Create a new RDMA id for the new IW CM ID */
1491 new_cm_id = rdma_create_id(listen_id->id.event_handler,
1492 listen_id->id.context,
1493 RDMA_PS_TCP, IB_QPT_RC);
1494 if (IS_ERR(new_cm_id)) {
1498 conn_id = container_of(new_cm_id, struct rdma_id_private, id);
1499 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1500 conn_id->state = RDMA_CM_CONNECT;
1502 ret = rdma_translate_ip(laddr, &conn_id->id.route.addr.dev_addr, NULL);
1504 mutex_unlock(&conn_id->handler_mutex);
1505 rdma_destroy_id(new_cm_id);
1509 ret = cma_acquire_dev(conn_id, listen_id);
1511 mutex_unlock(&conn_id->handler_mutex);
1512 rdma_destroy_id(new_cm_id);
1516 conn_id->cm_id.iw = cm_id;
1517 cm_id->context = conn_id;
1518 cm_id->cm_handler = cma_iw_handler;
1520 memcpy(cma_src_addr(conn_id), laddr, rdma_addr_size(laddr));
1521 memcpy(cma_dst_addr(conn_id), raddr, rdma_addr_size(raddr));
1523 ret = ib_query_device(conn_id->id.device, &attr);
1525 mutex_unlock(&conn_id->handler_mutex);
1526 rdma_destroy_id(new_cm_id);
1530 memset(&event, 0, sizeof event);
1531 event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1532 event.param.conn.private_data = iw_event->private_data;
1533 event.param.conn.private_data_len = iw_event->private_data_len;
1534 event.param.conn.initiator_depth = iw_event->ird;
1535 event.param.conn.responder_resources = iw_event->ord;
1538 * Protect against the user destroying conn_id from another thread
1539 * until we're done accessing it.
1541 atomic_inc(&conn_id->refcount);
1542 ret = conn_id->id.event_handler(&conn_id->id, &event);
1544 /* User wants to destroy the CM ID */
1545 conn_id->cm_id.iw = NULL;
1546 cma_exch(conn_id, RDMA_CM_DESTROYING);
1547 mutex_unlock(&conn_id->handler_mutex);
1548 cma_deref_id(conn_id);
1549 rdma_destroy_id(&conn_id->id);
1553 mutex_unlock(&conn_id->handler_mutex);
1554 cma_deref_id(conn_id);
1557 mutex_unlock(&listen_id->handler_mutex);
1561 static int cma_ib_listen(struct rdma_id_private *id_priv)
1563 struct ib_cm_compare_data compare_data;
1564 struct sockaddr *addr;
1565 struct ib_cm_id *id;
1569 id = ib_create_cm_id(id_priv->id.device, cma_req_handler, id_priv);
1573 id_priv->cm_id.ib = id;
1575 addr = cma_src_addr(id_priv);
1576 svc_id = rdma_get_service_id(&id_priv->id, addr);
1577 if (cma_any_addr(addr) && !id_priv->afonly)
1578 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, NULL);
1580 cma_set_compare_data(id_priv->id.ps, addr, &compare_data);
1581 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, &compare_data);
1585 ib_destroy_cm_id(id_priv->cm_id.ib);
1586 id_priv->cm_id.ib = NULL;
1592 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
1595 struct iw_cm_id *id;
1597 id = iw_create_cm_id(id_priv->id.device,
1598 iw_conn_req_handler,
1603 id_priv->cm_id.iw = id;
1605 memcpy(&id_priv->cm_id.iw->local_addr, cma_src_addr(id_priv),
1606 rdma_addr_size(cma_src_addr(id_priv)));
1608 ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
1611 iw_destroy_cm_id(id_priv->cm_id.iw);
1612 id_priv->cm_id.iw = NULL;
1618 static int cma_listen_handler(struct rdma_cm_id *id,
1619 struct rdma_cm_event *event)
1621 struct rdma_id_private *id_priv = id->context;
1623 id->context = id_priv->id.context;
1624 id->event_handler = id_priv->id.event_handler;
1625 return id_priv->id.event_handler(id, event);
1628 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
1629 struct cma_device *cma_dev)
1631 struct rdma_id_private *dev_id_priv;
1632 struct rdma_cm_id *id;
1635 if (cma_family(id_priv) == AF_IB &&
1636 rdma_node_get_transport(cma_dev->device->node_type) != RDMA_TRANSPORT_IB)
1639 id = rdma_create_id(cma_listen_handler, id_priv, id_priv->id.ps,
1640 id_priv->id.qp_type);
1644 dev_id_priv = container_of(id, struct rdma_id_private, id);
1646 dev_id_priv->state = RDMA_CM_ADDR_BOUND;
1647 memcpy(cma_src_addr(dev_id_priv), cma_src_addr(id_priv),
1648 rdma_addr_size(cma_src_addr(id_priv)));
1650 cma_attach_to_dev(dev_id_priv, cma_dev);
1651 list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
1652 atomic_inc(&id_priv->refcount);
1653 dev_id_priv->internal_id = 1;
1654 dev_id_priv->afonly = id_priv->afonly;
1656 ret = rdma_listen(id, id_priv->backlog);
1658 printk(KERN_WARNING "RDMA CMA: cma_listen_on_dev, error %d, "
1659 "listening on device %s\n", ret, cma_dev->device->name);
1662 static void cma_listen_on_all(struct rdma_id_private *id_priv)
1664 struct cma_device *cma_dev;
1667 list_add_tail(&id_priv->list, &listen_any_list);
1668 list_for_each_entry(cma_dev, &dev_list, list)
1669 cma_listen_on_dev(id_priv, cma_dev);
1670 mutex_unlock(&lock);
1673 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
1675 struct rdma_id_private *id_priv;
1677 id_priv = container_of(id, struct rdma_id_private, id);
1678 id_priv->tos = (u8) tos;
1680 EXPORT_SYMBOL(rdma_set_service_type);
1682 static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec,
1685 struct cma_work *work = context;
1686 struct rdma_route *route;
1688 route = &work->id->id.route;
1691 route->num_paths = 1;
1692 *route->path_rec = *path_rec;
1694 work->old_state = RDMA_CM_ROUTE_QUERY;
1695 work->new_state = RDMA_CM_ADDR_RESOLVED;
1696 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
1697 work->event.status = status;
1700 queue_work(cma_wq, &work->work);
1703 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms,
1704 struct cma_work *work)
1706 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
1707 struct ib_sa_path_rec path_rec;
1708 ib_sa_comp_mask comp_mask;
1709 struct sockaddr_in6 *sin6;
1710 struct sockaddr_ib *sib;
1712 memset(&path_rec, 0, sizeof path_rec);
1713 rdma_addr_get_sgid(dev_addr, &path_rec.sgid);
1714 rdma_addr_get_dgid(dev_addr, &path_rec.dgid);
1715 path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
1716 path_rec.numb_path = 1;
1717 path_rec.reversible = 1;
1718 path_rec.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
1720 comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
1721 IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
1722 IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
1724 switch (cma_family(id_priv)) {
1726 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
1727 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
1730 sin6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
1731 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
1732 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
1735 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
1736 path_rec.traffic_class = (u8) (be32_to_cpu(sib->sib_flowinfo) >> 20);
1737 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
1741 id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
1742 id_priv->id.port_num, &path_rec,
1743 comp_mask, timeout_ms,
1744 GFP_KERNEL, cma_query_handler,
1745 work, &id_priv->query);
1747 return (id_priv->query_id < 0) ? id_priv->query_id : 0;
1750 static void cma_work_handler(struct work_struct *_work)
1752 struct cma_work *work = container_of(_work, struct cma_work, work);
1753 struct rdma_id_private *id_priv = work->id;
1756 mutex_lock(&id_priv->handler_mutex);
1757 if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
1760 if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1761 cma_exch(id_priv, RDMA_CM_DESTROYING);
1765 mutex_unlock(&id_priv->handler_mutex);
1766 cma_deref_id(id_priv);
1768 rdma_destroy_id(&id_priv->id);
1772 static void cma_ndev_work_handler(struct work_struct *_work)
1774 struct cma_ndev_work *work = container_of(_work, struct cma_ndev_work, work);
1775 struct rdma_id_private *id_priv = work->id;
1778 mutex_lock(&id_priv->handler_mutex);
1779 if (id_priv->state == RDMA_CM_DESTROYING ||
1780 id_priv->state == RDMA_CM_DEVICE_REMOVAL)
1783 if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1784 cma_exch(id_priv, RDMA_CM_DESTROYING);
1789 mutex_unlock(&id_priv->handler_mutex);
1790 cma_deref_id(id_priv);
1792 rdma_destroy_id(&id_priv->id);
1796 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms)
1798 struct rdma_route *route = &id_priv->id.route;
1799 struct cma_work *work;
1802 work = kzalloc(sizeof *work, GFP_KERNEL);
1807 INIT_WORK(&work->work, cma_work_handler);
1808 work->old_state = RDMA_CM_ROUTE_QUERY;
1809 work->new_state = RDMA_CM_ROUTE_RESOLVED;
1810 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1812 route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
1813 if (!route->path_rec) {
1818 ret = cma_query_ib_route(id_priv, timeout_ms, work);
1824 kfree(route->path_rec);
1825 route->path_rec = NULL;
1831 int rdma_set_ib_paths(struct rdma_cm_id *id,
1832 struct ib_sa_path_rec *path_rec, int num_paths)
1834 struct rdma_id_private *id_priv;
1837 id_priv = container_of(id, struct rdma_id_private, id);
1838 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
1839 RDMA_CM_ROUTE_RESOLVED))
1842 id->route.path_rec = kmemdup(path_rec, sizeof *path_rec * num_paths,
1844 if (!id->route.path_rec) {
1849 id->route.num_paths = num_paths;
1852 cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED);
1855 EXPORT_SYMBOL(rdma_set_ib_paths);
1857 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms)
1859 struct cma_work *work;
1861 work = kzalloc(sizeof *work, GFP_KERNEL);
1866 INIT_WORK(&work->work, cma_work_handler);
1867 work->old_state = RDMA_CM_ROUTE_QUERY;
1868 work->new_state = RDMA_CM_ROUTE_RESOLVED;
1869 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1870 queue_work(cma_wq, &work->work);
1874 static int iboe_tos_to_sl(struct net_device *ndev, int tos)
1877 struct net_device *dev;
1879 prio = rt_tos2priority(tos);
1880 dev = ndev->priv_flags & IFF_802_1Q_VLAN ?
1881 vlan_dev_real_dev(ndev) : ndev;
1884 return netdev_get_prio_tc_map(dev, prio);
1886 #if IS_ENABLED(CONFIG_VLAN_8021Q)
1887 if (ndev->priv_flags & IFF_802_1Q_VLAN)
1888 return (vlan_dev_get_egress_qos_mask(ndev, prio) &
1889 VLAN_PRIO_MASK) >> VLAN_PRIO_SHIFT;
1894 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
1896 struct rdma_route *route = &id_priv->id.route;
1897 struct rdma_addr *addr = &route->addr;
1898 struct cma_work *work;
1900 struct net_device *ndev = NULL;
1903 work = kzalloc(sizeof *work, GFP_KERNEL);
1908 INIT_WORK(&work->work, cma_work_handler);
1910 route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
1911 if (!route->path_rec) {
1916 route->num_paths = 1;
1918 if (addr->dev_addr.bound_dev_if)
1919 ndev = dev_get_by_index(&init_net, addr->dev_addr.bound_dev_if);
1925 route->path_rec->vlan_id = rdma_vlan_dev_vlan_id(ndev);
1926 memcpy(route->path_rec->dmac, addr->dev_addr.dst_dev_addr, ETH_ALEN);
1927 memcpy(route->path_rec->smac, ndev->dev_addr, ndev->addr_len);
1929 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
1930 &route->path_rec->sgid);
1931 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.dst_addr,
1932 &route->path_rec->dgid);
1934 route->path_rec->hop_limit = 1;
1935 route->path_rec->reversible = 1;
1936 route->path_rec->pkey = cpu_to_be16(0xffff);
1937 route->path_rec->mtu_selector = IB_SA_EQ;
1938 route->path_rec->sl = iboe_tos_to_sl(ndev, id_priv->tos);
1939 route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
1940 route->path_rec->rate_selector = IB_SA_EQ;
1941 route->path_rec->rate = iboe_get_rate(ndev);
1943 route->path_rec->packet_life_time_selector = IB_SA_EQ;
1944 route->path_rec->packet_life_time = CMA_IBOE_PACKET_LIFETIME;
1945 if (!route->path_rec->mtu) {
1950 work->old_state = RDMA_CM_ROUTE_QUERY;
1951 work->new_state = RDMA_CM_ROUTE_RESOLVED;
1952 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1953 work->event.status = 0;
1955 queue_work(cma_wq, &work->work);
1960 kfree(route->path_rec);
1961 route->path_rec = NULL;
1967 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
1969 struct rdma_id_private *id_priv;
1972 id_priv = container_of(id, struct rdma_id_private, id);
1973 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY))
1976 atomic_inc(&id_priv->refcount);
1977 switch (rdma_node_get_transport(id->device->node_type)) {
1978 case RDMA_TRANSPORT_IB:
1979 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
1980 case IB_LINK_LAYER_INFINIBAND:
1981 ret = cma_resolve_ib_route(id_priv, timeout_ms);
1983 case IB_LINK_LAYER_ETHERNET:
1984 ret = cma_resolve_iboe_route(id_priv);
1990 case RDMA_TRANSPORT_IWARP:
1991 ret = cma_resolve_iw_route(id_priv, timeout_ms);
2002 cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED);
2003 cma_deref_id(id_priv);
2006 EXPORT_SYMBOL(rdma_resolve_route);
2008 static void cma_set_loopback(struct sockaddr *addr)
2010 switch (addr->sa_family) {
2012 ((struct sockaddr_in *) addr)->sin_addr.s_addr = htonl(INADDR_LOOPBACK);
2015 ipv6_addr_set(&((struct sockaddr_in6 *) addr)->sin6_addr,
2019 ib_addr_set(&((struct sockaddr_ib *) addr)->sib_addr,
2025 static int cma_bind_loopback(struct rdma_id_private *id_priv)
2027 struct cma_device *cma_dev, *cur_dev;
2028 struct ib_port_attr port_attr;
2036 list_for_each_entry(cur_dev, &dev_list, list) {
2037 if (cma_family(id_priv) == AF_IB &&
2038 rdma_node_get_transport(cur_dev->device->node_type) != RDMA_TRANSPORT_IB)
2044 for (p = 1; p <= cur_dev->device->phys_port_cnt; ++p) {
2045 if (!ib_query_port(cur_dev->device, p, &port_attr) &&
2046 port_attr.state == IB_PORT_ACTIVE) {
2061 ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid);
2065 ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
2069 id_priv->id.route.addr.dev_addr.dev_type =
2070 (rdma_port_get_link_layer(cma_dev->device, p) == IB_LINK_LAYER_INFINIBAND) ?
2071 ARPHRD_INFINIBAND : ARPHRD_ETHER;
2073 rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2074 ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
2075 id_priv->id.port_num = p;
2076 cma_attach_to_dev(id_priv, cma_dev);
2077 cma_set_loopback(cma_src_addr(id_priv));
2079 mutex_unlock(&lock);
2083 static void addr_handler(int status, struct sockaddr *src_addr,
2084 struct rdma_dev_addr *dev_addr, void *context)
2086 struct rdma_id_private *id_priv = context;
2087 struct rdma_cm_event event;
2089 memset(&event, 0, sizeof event);
2090 mutex_lock(&id_priv->handler_mutex);
2091 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY,
2092 RDMA_CM_ADDR_RESOLVED))
2095 memcpy(cma_src_addr(id_priv), src_addr, rdma_addr_size(src_addr));
2096 if (!status && !id_priv->cma_dev)
2097 status = cma_acquire_dev(id_priv, NULL);
2100 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
2101 RDMA_CM_ADDR_BOUND))
2103 event.event = RDMA_CM_EVENT_ADDR_ERROR;
2104 event.status = status;
2106 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2108 if (id_priv->id.event_handler(&id_priv->id, &event)) {
2109 cma_exch(id_priv, RDMA_CM_DESTROYING);
2110 mutex_unlock(&id_priv->handler_mutex);
2111 cma_deref_id(id_priv);
2112 rdma_destroy_id(&id_priv->id);
2116 mutex_unlock(&id_priv->handler_mutex);
2117 cma_deref_id(id_priv);
2120 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
2122 struct cma_work *work;
2126 work = kzalloc(sizeof *work, GFP_KERNEL);
2130 if (!id_priv->cma_dev) {
2131 ret = cma_bind_loopback(id_priv);
2136 rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2137 rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
2140 INIT_WORK(&work->work, cma_work_handler);
2141 work->old_state = RDMA_CM_ADDR_QUERY;
2142 work->new_state = RDMA_CM_ADDR_RESOLVED;
2143 work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2144 queue_work(cma_wq, &work->work);
2151 static int cma_resolve_ib_addr(struct rdma_id_private *id_priv)
2153 struct cma_work *work;
2156 work = kzalloc(sizeof *work, GFP_KERNEL);
2160 if (!id_priv->cma_dev) {
2161 ret = cma_resolve_ib_dev(id_priv);
2166 rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, (union ib_gid *)
2167 &(((struct sockaddr_ib *) &id_priv->id.route.addr.dst_addr)->sib_addr));
2170 INIT_WORK(&work->work, cma_work_handler);
2171 work->old_state = RDMA_CM_ADDR_QUERY;
2172 work->new_state = RDMA_CM_ADDR_RESOLVED;
2173 work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2174 queue_work(cma_wq, &work->work);
2181 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2182 struct sockaddr *dst_addr)
2184 if (!src_addr || !src_addr->sa_family) {
2185 src_addr = (struct sockaddr *) &id->route.addr.src_addr;
2186 src_addr->sa_family = dst_addr->sa_family;
2187 if (dst_addr->sa_family == AF_INET6) {
2188 ((struct sockaddr_in6 *) src_addr)->sin6_scope_id =
2189 ((struct sockaddr_in6 *) dst_addr)->sin6_scope_id;
2190 } else if (dst_addr->sa_family == AF_IB) {
2191 ((struct sockaddr_ib *) src_addr)->sib_pkey =
2192 ((struct sockaddr_ib *) dst_addr)->sib_pkey;
2195 return rdma_bind_addr(id, src_addr);
2198 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2199 struct sockaddr *dst_addr, int timeout_ms)
2201 struct rdma_id_private *id_priv;
2204 id_priv = container_of(id, struct rdma_id_private, id);
2205 if (id_priv->state == RDMA_CM_IDLE) {
2206 ret = cma_bind_addr(id, src_addr, dst_addr);
2211 if (cma_family(id_priv) != dst_addr->sa_family)
2214 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY))
2217 atomic_inc(&id_priv->refcount);
2218 memcpy(cma_dst_addr(id_priv), dst_addr, rdma_addr_size(dst_addr));
2219 if (cma_any_addr(dst_addr)) {
2220 ret = cma_resolve_loopback(id_priv);
2222 if (dst_addr->sa_family == AF_IB) {
2223 ret = cma_resolve_ib_addr(id_priv);
2225 ret = rdma_resolve_ip(&addr_client, cma_src_addr(id_priv),
2226 dst_addr, &id->route.addr.dev_addr,
2227 timeout_ms, addr_handler, id_priv);
2235 cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
2236 cma_deref_id(id_priv);
2239 EXPORT_SYMBOL(rdma_resolve_addr);
2241 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse)
2243 struct rdma_id_private *id_priv;
2244 unsigned long flags;
2247 id_priv = container_of(id, struct rdma_id_private, id);
2248 spin_lock_irqsave(&id_priv->lock, flags);
2249 if (reuse || id_priv->state == RDMA_CM_IDLE) {
2250 id_priv->reuseaddr = reuse;
2255 spin_unlock_irqrestore(&id_priv->lock, flags);
2258 EXPORT_SYMBOL(rdma_set_reuseaddr);
2260 int rdma_set_afonly(struct rdma_cm_id *id, int afonly)
2262 struct rdma_id_private *id_priv;
2263 unsigned long flags;
2266 id_priv = container_of(id, struct rdma_id_private, id);
2267 spin_lock_irqsave(&id_priv->lock, flags);
2268 if (id_priv->state == RDMA_CM_IDLE || id_priv->state == RDMA_CM_ADDR_BOUND) {
2269 id_priv->options |= (1 << CMA_OPTION_AFONLY);
2270 id_priv->afonly = afonly;
2275 spin_unlock_irqrestore(&id_priv->lock, flags);
2278 EXPORT_SYMBOL(rdma_set_afonly);
2280 static void cma_bind_port(struct rdma_bind_list *bind_list,
2281 struct rdma_id_private *id_priv)
2283 struct sockaddr *addr;
2284 struct sockaddr_ib *sib;
2288 addr = cma_src_addr(id_priv);
2289 port = htons(bind_list->port);
2291 switch (addr->sa_family) {
2293 ((struct sockaddr_in *) addr)->sin_port = port;
2296 ((struct sockaddr_in6 *) addr)->sin6_port = port;
2299 sib = (struct sockaddr_ib *) addr;
2300 sid = be64_to_cpu(sib->sib_sid);
2301 mask = be64_to_cpu(sib->sib_sid_mask);
2302 sib->sib_sid = cpu_to_be64((sid & mask) | (u64) ntohs(port));
2303 sib->sib_sid_mask = cpu_to_be64(~0ULL);
2306 id_priv->bind_list = bind_list;
2307 hlist_add_head(&id_priv->node, &bind_list->owners);
2310 static int cma_alloc_port(struct idr *ps, struct rdma_id_private *id_priv,
2311 unsigned short snum)
2313 struct rdma_bind_list *bind_list;
2316 bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
2320 ret = idr_alloc(ps, bind_list, snum, snum + 1, GFP_KERNEL);
2325 bind_list->port = (unsigned short)ret;
2326 cma_bind_port(bind_list, id_priv);
2330 return ret == -ENOSPC ? -EADDRNOTAVAIL : ret;
2333 static int cma_alloc_any_port(struct idr *ps, struct rdma_id_private *id_priv)
2335 static unsigned int last_used_port;
2336 int low, high, remaining;
2339 inet_get_local_port_range(&init_net, &low, &high);
2340 remaining = (high - low) + 1;
2341 rover = prandom_u32() % remaining + low;
2343 if (last_used_port != rover &&
2344 !idr_find(ps, (unsigned short) rover)) {
2345 int ret = cma_alloc_port(ps, id_priv, rover);
2347 * Remember previously used port number in order to avoid
2348 * re-using same port immediately after it is closed.
2351 last_used_port = rover;
2352 if (ret != -EADDRNOTAVAIL)
2357 if ((rover < low) || (rover > high))
2361 return -EADDRNOTAVAIL;
2365 * Check that the requested port is available. This is called when trying to
2366 * bind to a specific port, or when trying to listen on a bound port. In
2367 * the latter case, the provided id_priv may already be on the bind_list, but
2368 * we still need to check that it's okay to start listening.
2370 static int cma_check_port(struct rdma_bind_list *bind_list,
2371 struct rdma_id_private *id_priv, uint8_t reuseaddr)
2373 struct rdma_id_private *cur_id;
2374 struct sockaddr *addr, *cur_addr;
2376 addr = cma_src_addr(id_priv);
2377 hlist_for_each_entry(cur_id, &bind_list->owners, node) {
2378 if (id_priv == cur_id)
2381 if ((cur_id->state != RDMA_CM_LISTEN) && reuseaddr &&
2385 cur_addr = cma_src_addr(cur_id);
2386 if (id_priv->afonly && cur_id->afonly &&
2387 (addr->sa_family != cur_addr->sa_family))
2390 if (cma_any_addr(addr) || cma_any_addr(cur_addr))
2391 return -EADDRNOTAVAIL;
2393 if (!cma_addr_cmp(addr, cur_addr))
2399 static int cma_use_port(struct idr *ps, struct rdma_id_private *id_priv)
2401 struct rdma_bind_list *bind_list;
2402 unsigned short snum;
2405 snum = ntohs(cma_port(cma_src_addr(id_priv)));
2406 if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
2409 bind_list = idr_find(ps, snum);
2411 ret = cma_alloc_port(ps, id_priv, snum);
2413 ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr);
2415 cma_bind_port(bind_list, id_priv);
2420 static int cma_bind_listen(struct rdma_id_private *id_priv)
2422 struct rdma_bind_list *bind_list = id_priv->bind_list;
2426 if (bind_list->owners.first->next)
2427 ret = cma_check_port(bind_list, id_priv, 0);
2428 mutex_unlock(&lock);
2432 static struct idr *cma_select_inet_ps(struct rdma_id_private *id_priv)
2434 switch (id_priv->id.ps) {
2448 static struct idr *cma_select_ib_ps(struct rdma_id_private *id_priv)
2450 struct idr *ps = NULL;
2451 struct sockaddr_ib *sib;
2452 u64 sid_ps, mask, sid;
2454 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
2455 mask = be64_to_cpu(sib->sib_sid_mask) & RDMA_IB_IP_PS_MASK;
2456 sid = be64_to_cpu(sib->sib_sid) & mask;
2458 if ((id_priv->id.ps == RDMA_PS_IB) && (sid == (RDMA_IB_IP_PS_IB & mask))) {
2459 sid_ps = RDMA_IB_IP_PS_IB;
2461 } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_TCP)) &&
2462 (sid == (RDMA_IB_IP_PS_TCP & mask))) {
2463 sid_ps = RDMA_IB_IP_PS_TCP;
2465 } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_UDP)) &&
2466 (sid == (RDMA_IB_IP_PS_UDP & mask))) {
2467 sid_ps = RDMA_IB_IP_PS_UDP;
2472 sib->sib_sid = cpu_to_be64(sid_ps | ntohs(cma_port((struct sockaddr *) sib)));
2473 sib->sib_sid_mask = cpu_to_be64(RDMA_IB_IP_PS_MASK |
2474 be64_to_cpu(sib->sib_sid_mask));
2479 static int cma_get_port(struct rdma_id_private *id_priv)
2484 if (cma_family(id_priv) != AF_IB)
2485 ps = cma_select_inet_ps(id_priv);
2487 ps = cma_select_ib_ps(id_priv);
2489 return -EPROTONOSUPPORT;
2492 if (cma_any_port(cma_src_addr(id_priv)))
2493 ret = cma_alloc_any_port(ps, id_priv);
2495 ret = cma_use_port(ps, id_priv);
2496 mutex_unlock(&lock);
2501 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
2502 struct sockaddr *addr)
2504 #if IS_ENABLED(CONFIG_IPV6)
2505 struct sockaddr_in6 *sin6;
2507 if (addr->sa_family != AF_INET6)
2510 sin6 = (struct sockaddr_in6 *) addr;
2512 if (!(ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL))
2515 if (!sin6->sin6_scope_id)
2518 dev_addr->bound_dev_if = sin6->sin6_scope_id;
2523 int rdma_listen(struct rdma_cm_id *id, int backlog)
2525 struct rdma_id_private *id_priv;
2528 id_priv = container_of(id, struct rdma_id_private, id);
2529 if (id_priv->state == RDMA_CM_IDLE) {
2530 id->route.addr.src_addr.ss_family = AF_INET;
2531 ret = rdma_bind_addr(id, cma_src_addr(id_priv));
2536 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN))
2539 if (id_priv->reuseaddr) {
2540 ret = cma_bind_listen(id_priv);
2545 id_priv->backlog = backlog;
2547 switch (rdma_node_get_transport(id->device->node_type)) {
2548 case RDMA_TRANSPORT_IB:
2549 ret = cma_ib_listen(id_priv);
2553 case RDMA_TRANSPORT_IWARP:
2554 ret = cma_iw_listen(id_priv, backlog);
2563 cma_listen_on_all(id_priv);
2567 id_priv->backlog = 0;
2568 cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND);
2571 EXPORT_SYMBOL(rdma_listen);
2573 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
2575 struct rdma_id_private *id_priv;
2578 if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6 &&
2579 addr->sa_family != AF_IB)
2580 return -EAFNOSUPPORT;
2582 id_priv = container_of(id, struct rdma_id_private, id);
2583 if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND))
2586 ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
2590 memcpy(cma_src_addr(id_priv), addr, rdma_addr_size(addr));
2591 if (!cma_any_addr(addr)) {
2592 ret = cma_translate_addr(addr, &id->route.addr.dev_addr);
2596 ret = cma_acquire_dev(id_priv, NULL);
2601 if (!(id_priv->options & (1 << CMA_OPTION_AFONLY))) {
2602 if (addr->sa_family == AF_INET)
2603 id_priv->afonly = 1;
2604 #if IS_ENABLED(CONFIG_IPV6)
2605 else if (addr->sa_family == AF_INET6)
2606 id_priv->afonly = init_net.ipv6.sysctl.bindv6only;
2609 ret = cma_get_port(id_priv);
2615 if (id_priv->cma_dev)
2616 cma_release_dev(id_priv);
2618 cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE);
2621 EXPORT_SYMBOL(rdma_bind_addr);
2623 static int cma_format_hdr(void *hdr, struct rdma_id_private *id_priv)
2625 struct cma_hdr *cma_hdr;
2628 cma_hdr->cma_version = CMA_VERSION;
2629 if (cma_family(id_priv) == AF_INET) {
2630 struct sockaddr_in *src4, *dst4;
2632 src4 = (struct sockaddr_in *) cma_src_addr(id_priv);
2633 dst4 = (struct sockaddr_in *) cma_dst_addr(id_priv);
2635 cma_set_ip_ver(cma_hdr, 4);
2636 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2637 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2638 cma_hdr->port = src4->sin_port;
2639 } else if (cma_family(id_priv) == AF_INET6) {
2640 struct sockaddr_in6 *src6, *dst6;
2642 src6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
2643 dst6 = (struct sockaddr_in6 *) cma_dst_addr(id_priv);
2645 cma_set_ip_ver(cma_hdr, 6);
2646 cma_hdr->src_addr.ip6 = src6->sin6_addr;
2647 cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
2648 cma_hdr->port = src6->sin6_port;
2653 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
2654 struct ib_cm_event *ib_event)
2656 struct rdma_id_private *id_priv = cm_id->context;
2657 struct rdma_cm_event event;
2658 struct ib_cm_sidr_rep_event_param *rep = &ib_event->param.sidr_rep_rcvd;
2661 if (cma_disable_callback(id_priv, RDMA_CM_CONNECT))
2664 memset(&event, 0, sizeof event);
2665 switch (ib_event->event) {
2666 case IB_CM_SIDR_REQ_ERROR:
2667 event.event = RDMA_CM_EVENT_UNREACHABLE;
2668 event.status = -ETIMEDOUT;
2670 case IB_CM_SIDR_REP_RECEIVED:
2671 event.param.ud.private_data = ib_event->private_data;
2672 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
2673 if (rep->status != IB_SIDR_SUCCESS) {
2674 event.event = RDMA_CM_EVENT_UNREACHABLE;
2675 event.status = ib_event->param.sidr_rep_rcvd.status;
2678 ret = cma_set_qkey(id_priv, rep->qkey);
2680 event.event = RDMA_CM_EVENT_ADDR_ERROR;
2684 ib_init_ah_from_path(id_priv->id.device, id_priv->id.port_num,
2685 id_priv->id.route.path_rec,
2686 &event.param.ud.ah_attr);
2687 event.param.ud.qp_num = rep->qpn;
2688 event.param.ud.qkey = rep->qkey;
2689 event.event = RDMA_CM_EVENT_ESTABLISHED;
2693 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
2698 ret = id_priv->id.event_handler(&id_priv->id, &event);
2700 /* Destroy the CM ID by returning a non-zero value. */
2701 id_priv->cm_id.ib = NULL;
2702 cma_exch(id_priv, RDMA_CM_DESTROYING);
2703 mutex_unlock(&id_priv->handler_mutex);
2704 rdma_destroy_id(&id_priv->id);
2708 mutex_unlock(&id_priv->handler_mutex);
2712 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
2713 struct rdma_conn_param *conn_param)
2715 struct ib_cm_sidr_req_param req;
2716 struct ib_cm_id *id;
2720 memset(&req, 0, sizeof req);
2721 offset = cma_user_data_offset(id_priv);
2722 req.private_data_len = offset + conn_param->private_data_len;
2723 if (req.private_data_len < conn_param->private_data_len)
2726 if (req.private_data_len) {
2727 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2731 private_data = NULL;
2734 if (conn_param->private_data && conn_param->private_data_len)
2735 memcpy(private_data + offset, conn_param->private_data,
2736 conn_param->private_data_len);
2739 ret = cma_format_hdr(private_data, id_priv);
2742 req.private_data = private_data;
2745 id = ib_create_cm_id(id_priv->id.device, cma_sidr_rep_handler,
2751 id_priv->cm_id.ib = id;
2753 req.path = id_priv->id.route.path_rec;
2754 req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
2755 req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
2756 req.max_cm_retries = CMA_MAX_CM_RETRIES;
2758 ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
2760 ib_destroy_cm_id(id_priv->cm_id.ib);
2761 id_priv->cm_id.ib = NULL;
2764 kfree(private_data);
2768 static int cma_connect_ib(struct rdma_id_private *id_priv,
2769 struct rdma_conn_param *conn_param)
2771 struct ib_cm_req_param req;
2772 struct rdma_route *route;
2774 struct ib_cm_id *id;
2777 memset(&req, 0, sizeof req);
2778 offset = cma_user_data_offset(id_priv);
2779 req.private_data_len = offset + conn_param->private_data_len;
2780 if (req.private_data_len < conn_param->private_data_len)
2783 if (req.private_data_len) {
2784 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2788 private_data = NULL;
2791 if (conn_param->private_data && conn_param->private_data_len)
2792 memcpy(private_data + offset, conn_param->private_data,
2793 conn_param->private_data_len);
2795 id = ib_create_cm_id(id_priv->id.device, cma_ib_handler, id_priv);
2800 id_priv->cm_id.ib = id;
2802 route = &id_priv->id.route;
2804 ret = cma_format_hdr(private_data, id_priv);
2807 req.private_data = private_data;
2810 req.primary_path = &route->path_rec[0];
2811 if (route->num_paths == 2)
2812 req.alternate_path = &route->path_rec[1];
2814 req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
2815 req.qp_num = id_priv->qp_num;
2816 req.qp_type = id_priv->id.qp_type;
2817 req.starting_psn = id_priv->seq_num;
2818 req.responder_resources = conn_param->responder_resources;
2819 req.initiator_depth = conn_param->initiator_depth;
2820 req.flow_control = conn_param->flow_control;
2821 req.retry_count = min_t(u8, 7, conn_param->retry_count);
2822 req.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
2823 req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2824 req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2825 req.max_cm_retries = CMA_MAX_CM_RETRIES;
2826 req.srq = id_priv->srq ? 1 : 0;
2828 ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
2830 if (ret && !IS_ERR(id)) {
2831 ib_destroy_cm_id(id);
2832 id_priv->cm_id.ib = NULL;
2835 kfree(private_data);
2839 static int cma_connect_iw(struct rdma_id_private *id_priv,
2840 struct rdma_conn_param *conn_param)
2842 struct iw_cm_id *cm_id;
2844 struct iw_cm_conn_param iw_param;
2846 cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
2848 return PTR_ERR(cm_id);
2850 id_priv->cm_id.iw = cm_id;
2852 memcpy(&cm_id->local_addr, cma_src_addr(id_priv),
2853 rdma_addr_size(cma_src_addr(id_priv)));
2854 memcpy(&cm_id->remote_addr, cma_dst_addr(id_priv),
2855 rdma_addr_size(cma_dst_addr(id_priv)));
2857 ret = cma_modify_qp_rtr(id_priv, conn_param);
2862 iw_param.ord = conn_param->initiator_depth;
2863 iw_param.ird = conn_param->responder_resources;
2864 iw_param.private_data = conn_param->private_data;
2865 iw_param.private_data_len = conn_param->private_data_len;
2866 iw_param.qpn = id_priv->id.qp ? id_priv->qp_num : conn_param->qp_num;
2868 memset(&iw_param, 0, sizeof iw_param);
2869 iw_param.qpn = id_priv->qp_num;
2871 ret = iw_cm_connect(cm_id, &iw_param);
2874 iw_destroy_cm_id(cm_id);
2875 id_priv->cm_id.iw = NULL;
2880 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2882 struct rdma_id_private *id_priv;
2885 id_priv = container_of(id, struct rdma_id_private, id);
2886 if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT))
2890 id_priv->qp_num = conn_param->qp_num;
2891 id_priv->srq = conn_param->srq;
2894 switch (rdma_node_get_transport(id->device->node_type)) {
2895 case RDMA_TRANSPORT_IB:
2896 if (id->qp_type == IB_QPT_UD)
2897 ret = cma_resolve_ib_udp(id_priv, conn_param);
2899 ret = cma_connect_ib(id_priv, conn_param);
2901 case RDMA_TRANSPORT_IWARP:
2902 ret = cma_connect_iw(id_priv, conn_param);
2913 cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED);
2916 EXPORT_SYMBOL(rdma_connect);
2918 static int cma_accept_ib(struct rdma_id_private *id_priv,
2919 struct rdma_conn_param *conn_param)
2921 struct ib_cm_rep_param rep;
2924 ret = cma_modify_qp_rtr(id_priv, conn_param);
2928 ret = cma_modify_qp_rts(id_priv, conn_param);
2932 memset(&rep, 0, sizeof rep);
2933 rep.qp_num = id_priv->qp_num;
2934 rep.starting_psn = id_priv->seq_num;
2935 rep.private_data = conn_param->private_data;
2936 rep.private_data_len = conn_param->private_data_len;
2937 rep.responder_resources = conn_param->responder_resources;
2938 rep.initiator_depth = conn_param->initiator_depth;
2939 rep.failover_accepted = 0;
2940 rep.flow_control = conn_param->flow_control;
2941 rep.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
2942 rep.srq = id_priv->srq ? 1 : 0;
2944 ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
2949 static int cma_accept_iw(struct rdma_id_private *id_priv,
2950 struct rdma_conn_param *conn_param)
2952 struct iw_cm_conn_param iw_param;
2955 ret = cma_modify_qp_rtr(id_priv, conn_param);
2959 iw_param.ord = conn_param->initiator_depth;
2960 iw_param.ird = conn_param->responder_resources;
2961 iw_param.private_data = conn_param->private_data;
2962 iw_param.private_data_len = conn_param->private_data_len;
2963 if (id_priv->id.qp) {
2964 iw_param.qpn = id_priv->qp_num;
2966 iw_param.qpn = conn_param->qp_num;
2968 return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
2971 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
2972 enum ib_cm_sidr_status status, u32 qkey,
2973 const void *private_data, int private_data_len)
2975 struct ib_cm_sidr_rep_param rep;
2978 memset(&rep, 0, sizeof rep);
2979 rep.status = status;
2980 if (status == IB_SIDR_SUCCESS) {
2981 ret = cma_set_qkey(id_priv, qkey);
2984 rep.qp_num = id_priv->qp_num;
2985 rep.qkey = id_priv->qkey;
2987 rep.private_data = private_data;
2988 rep.private_data_len = private_data_len;
2990 return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
2993 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2995 struct rdma_id_private *id_priv;
2998 id_priv = container_of(id, struct rdma_id_private, id);
3000 id_priv->owner = task_pid_nr(current);
3002 if (!cma_comp(id_priv, RDMA_CM_CONNECT))
3005 if (!id->qp && conn_param) {
3006 id_priv->qp_num = conn_param->qp_num;
3007 id_priv->srq = conn_param->srq;
3010 switch (rdma_node_get_transport(id->device->node_type)) {
3011 case RDMA_TRANSPORT_IB:
3012 if (id->qp_type == IB_QPT_UD) {
3014 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
3016 conn_param->private_data,
3017 conn_param->private_data_len);
3019 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
3023 ret = cma_accept_ib(id_priv, conn_param);
3025 ret = cma_rep_recv(id_priv);
3028 case RDMA_TRANSPORT_IWARP:
3029 ret = cma_accept_iw(id_priv, conn_param);
3041 cma_modify_qp_err(id_priv);
3042 rdma_reject(id, NULL, 0);
3045 EXPORT_SYMBOL(rdma_accept);
3047 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
3049 struct rdma_id_private *id_priv;
3052 id_priv = container_of(id, struct rdma_id_private, id);
3053 if (!id_priv->cm_id.ib)
3056 switch (id->device->node_type) {
3057 case RDMA_NODE_IB_CA:
3058 ret = ib_cm_notify(id_priv->cm_id.ib, event);
3066 EXPORT_SYMBOL(rdma_notify);
3068 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
3069 u8 private_data_len)
3071 struct rdma_id_private *id_priv;
3074 id_priv = container_of(id, struct rdma_id_private, id);
3075 if (!id_priv->cm_id.ib)
3078 switch (rdma_node_get_transport(id->device->node_type)) {
3079 case RDMA_TRANSPORT_IB:
3080 if (id->qp_type == IB_QPT_UD)
3081 ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT, 0,
3082 private_data, private_data_len);
3084 ret = ib_send_cm_rej(id_priv->cm_id.ib,
3085 IB_CM_REJ_CONSUMER_DEFINED, NULL,
3086 0, private_data, private_data_len);
3088 case RDMA_TRANSPORT_IWARP:
3089 ret = iw_cm_reject(id_priv->cm_id.iw,
3090 private_data, private_data_len);
3098 EXPORT_SYMBOL(rdma_reject);
3100 int rdma_disconnect(struct rdma_cm_id *id)
3102 struct rdma_id_private *id_priv;
3105 id_priv = container_of(id, struct rdma_id_private, id);
3106 if (!id_priv->cm_id.ib)
3109 switch (rdma_node_get_transport(id->device->node_type)) {
3110 case RDMA_TRANSPORT_IB:
3111 ret = cma_modify_qp_err(id_priv);
3114 /* Initiate or respond to a disconnect. */
3115 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
3116 ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
3118 case RDMA_TRANSPORT_IWARP:
3119 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
3128 EXPORT_SYMBOL(rdma_disconnect);
3130 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
3132 struct rdma_id_private *id_priv;
3133 struct cma_multicast *mc = multicast->context;
3134 struct rdma_cm_event event;
3137 id_priv = mc->id_priv;
3138 if (cma_disable_callback(id_priv, RDMA_CM_ADDR_BOUND) &&
3139 cma_disable_callback(id_priv, RDMA_CM_ADDR_RESOLVED))
3143 status = cma_set_qkey(id_priv, be32_to_cpu(multicast->rec.qkey));
3144 mutex_lock(&id_priv->qp_mutex);
3145 if (!status && id_priv->id.qp)
3146 status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid,
3147 be16_to_cpu(multicast->rec.mlid));
3148 mutex_unlock(&id_priv->qp_mutex);
3150 memset(&event, 0, sizeof event);
3151 event.status = status;
3152 event.param.ud.private_data = mc->context;
3154 event.event = RDMA_CM_EVENT_MULTICAST_JOIN;
3155 ib_init_ah_from_mcmember(id_priv->id.device,
3156 id_priv->id.port_num, &multicast->rec,
3157 &event.param.ud.ah_attr);
3158 event.param.ud.qp_num = 0xFFFFFF;
3159 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
3161 event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
3163 ret = id_priv->id.event_handler(&id_priv->id, &event);
3165 cma_exch(id_priv, RDMA_CM_DESTROYING);
3166 mutex_unlock(&id_priv->handler_mutex);
3167 rdma_destroy_id(&id_priv->id);
3171 mutex_unlock(&id_priv->handler_mutex);
3175 static void cma_set_mgid(struct rdma_id_private *id_priv,
3176 struct sockaddr *addr, union ib_gid *mgid)
3178 unsigned char mc_map[MAX_ADDR_LEN];
3179 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3180 struct sockaddr_in *sin = (struct sockaddr_in *) addr;
3181 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
3183 if (cma_any_addr(addr)) {
3184 memset(mgid, 0, sizeof *mgid);
3185 } else if ((addr->sa_family == AF_INET6) &&
3186 ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
3188 /* IPv6 address is an SA assigned MGID. */
3189 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3190 } else if (addr->sa_family == AF_IB) {
3191 memcpy(mgid, &((struct sockaddr_ib *) addr)->sib_addr, sizeof *mgid);
3192 } else if ((addr->sa_family == AF_INET6)) {
3193 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
3194 if (id_priv->id.ps == RDMA_PS_UDP)
3195 mc_map[7] = 0x01; /* Use RDMA CM signature */
3196 *mgid = *(union ib_gid *) (mc_map + 4);
3198 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
3199 if (id_priv->id.ps == RDMA_PS_UDP)
3200 mc_map[7] = 0x01; /* Use RDMA CM signature */
3201 *mgid = *(union ib_gid *) (mc_map + 4);
3205 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
3206 struct cma_multicast *mc)
3208 struct ib_sa_mcmember_rec rec;
3209 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3210 ib_sa_comp_mask comp_mask;
3213 ib_addr_get_mgid(dev_addr, &rec.mgid);
3214 ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
3219 ret = cma_set_qkey(id_priv, 0);
3223 cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
3224 rec.qkey = cpu_to_be32(id_priv->qkey);
3225 rdma_addr_get_sgid(dev_addr, &rec.port_gid);
3226 rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
3229 comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
3230 IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
3231 IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
3232 IB_SA_MCMEMBER_REC_FLOW_LABEL |
3233 IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
3235 if (id_priv->id.ps == RDMA_PS_IPOIB)
3236 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
3237 IB_SA_MCMEMBER_REC_RATE_SELECTOR |
3238 IB_SA_MCMEMBER_REC_MTU_SELECTOR |
3239 IB_SA_MCMEMBER_REC_MTU |
3240 IB_SA_MCMEMBER_REC_HOP_LIMIT;
3242 mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device,
3243 id_priv->id.port_num, &rec,
3244 comp_mask, GFP_KERNEL,
3245 cma_ib_mc_handler, mc);
3246 return PTR_ERR_OR_ZERO(mc->multicast.ib);
3249 static void iboe_mcast_work_handler(struct work_struct *work)
3251 struct iboe_mcast_work *mw = container_of(work, struct iboe_mcast_work, work);
3252 struct cma_multicast *mc = mw->mc;
3253 struct ib_sa_multicast *m = mc->multicast.ib;
3255 mc->multicast.ib->context = mc;
3256 cma_ib_mc_handler(0, m);
3257 kref_put(&mc->mcref, release_mc);
3261 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid)
3263 struct sockaddr_in *sin = (struct sockaddr_in *)addr;
3264 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
3266 if (cma_any_addr(addr)) {
3267 memset(mgid, 0, sizeof *mgid);
3268 } else if (addr->sa_family == AF_INET6) {
3269 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3271 mgid->raw[0] = 0xff;
3272 mgid->raw[1] = 0x0e;
3281 mgid->raw[10] = 0xff;
3282 mgid->raw[11] = 0xff;
3283 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
3287 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
3288 struct cma_multicast *mc)
3290 struct iboe_mcast_work *work;
3291 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3293 struct sockaddr *addr = (struct sockaddr *)&mc->addr;
3294 struct net_device *ndev = NULL;
3296 if (cma_zero_addr((struct sockaddr *)&mc->addr))
3299 work = kzalloc(sizeof *work, GFP_KERNEL);
3303 mc->multicast.ib = kzalloc(sizeof(struct ib_sa_multicast), GFP_KERNEL);
3304 if (!mc->multicast.ib) {
3309 cma_iboe_set_mgid(addr, &mc->multicast.ib->rec.mgid);
3311 mc->multicast.ib->rec.pkey = cpu_to_be16(0xffff);
3312 if (id_priv->id.ps == RDMA_PS_UDP)
3313 mc->multicast.ib->rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
3315 if (dev_addr->bound_dev_if)
3316 ndev = dev_get_by_index(&init_net, dev_addr->bound_dev_if);
3321 mc->multicast.ib->rec.rate = iboe_get_rate(ndev);
3322 mc->multicast.ib->rec.hop_limit = 1;
3323 mc->multicast.ib->rec.mtu = iboe_get_mtu(ndev->mtu);
3325 if (!mc->multicast.ib->rec.mtu) {
3329 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
3330 &mc->multicast.ib->rec.port_gid);
3333 INIT_WORK(&work->work, iboe_mcast_work_handler);
3334 kref_get(&mc->mcref);
3335 queue_work(cma_wq, &work->work);
3340 kfree(mc->multicast.ib);
3346 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
3349 struct rdma_id_private *id_priv;
3350 struct cma_multicast *mc;
3353 id_priv = container_of(id, struct rdma_id_private, id);
3354 if (!cma_comp(id_priv, RDMA_CM_ADDR_BOUND) &&
3355 !cma_comp(id_priv, RDMA_CM_ADDR_RESOLVED))
3358 mc = kmalloc(sizeof *mc, GFP_KERNEL);
3362 memcpy(&mc->addr, addr, rdma_addr_size(addr));
3363 mc->context = context;
3364 mc->id_priv = id_priv;
3366 spin_lock(&id_priv->lock);
3367 list_add(&mc->list, &id_priv->mc_list);
3368 spin_unlock(&id_priv->lock);
3370 switch (rdma_node_get_transport(id->device->node_type)) {
3371 case RDMA_TRANSPORT_IB:
3372 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
3373 case IB_LINK_LAYER_INFINIBAND:
3374 ret = cma_join_ib_multicast(id_priv, mc);
3376 case IB_LINK_LAYER_ETHERNET:
3377 kref_init(&mc->mcref);
3378 ret = cma_iboe_join_multicast(id_priv, mc);
3390 spin_lock_irq(&id_priv->lock);
3391 list_del(&mc->list);
3392 spin_unlock_irq(&id_priv->lock);
3397 EXPORT_SYMBOL(rdma_join_multicast);
3399 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
3401 struct rdma_id_private *id_priv;
3402 struct cma_multicast *mc;
3404 id_priv = container_of(id, struct rdma_id_private, id);
3405 spin_lock_irq(&id_priv->lock);
3406 list_for_each_entry(mc, &id_priv->mc_list, list) {
3407 if (!memcmp(&mc->addr, addr, rdma_addr_size(addr))) {
3408 list_del(&mc->list);
3409 spin_unlock_irq(&id_priv->lock);
3412 ib_detach_mcast(id->qp,
3413 &mc->multicast.ib->rec.mgid,
3414 be16_to_cpu(mc->multicast.ib->rec.mlid));
3415 if (rdma_node_get_transport(id_priv->cma_dev->device->node_type) == RDMA_TRANSPORT_IB) {
3416 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
3417 case IB_LINK_LAYER_INFINIBAND:
3418 ib_sa_free_multicast(mc->multicast.ib);
3421 case IB_LINK_LAYER_ETHERNET:
3422 kref_put(&mc->mcref, release_mc);
3431 spin_unlock_irq(&id_priv->lock);
3433 EXPORT_SYMBOL(rdma_leave_multicast);
3435 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
3437 struct rdma_dev_addr *dev_addr;
3438 struct cma_ndev_work *work;
3440 dev_addr = &id_priv->id.route.addr.dev_addr;
3442 if ((dev_addr->bound_dev_if == ndev->ifindex) &&
3443 memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
3444 printk(KERN_INFO "RDMA CM addr change for ndev %s used by id %p\n",
3445 ndev->name, &id_priv->id);
3446 work = kzalloc(sizeof *work, GFP_KERNEL);
3450 INIT_WORK(&work->work, cma_ndev_work_handler);
3452 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
3453 atomic_inc(&id_priv->refcount);
3454 queue_work(cma_wq, &work->work);
3460 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
3463 struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
3464 struct cma_device *cma_dev;
3465 struct rdma_id_private *id_priv;
3466 int ret = NOTIFY_DONE;
3468 if (dev_net(ndev) != &init_net)
3471 if (event != NETDEV_BONDING_FAILOVER)
3474 if (!(ndev->flags & IFF_MASTER) || !(ndev->priv_flags & IFF_BONDING))
3478 list_for_each_entry(cma_dev, &dev_list, list)
3479 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
3480 ret = cma_netdev_change(ndev, id_priv);
3486 mutex_unlock(&lock);
3490 static struct notifier_block cma_nb = {
3491 .notifier_call = cma_netdev_callback
3494 static void cma_add_one(struct ib_device *device)
3496 struct cma_device *cma_dev;
3497 struct rdma_id_private *id_priv;
3499 cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
3503 cma_dev->device = device;
3505 init_completion(&cma_dev->comp);
3506 atomic_set(&cma_dev->refcount, 1);
3507 INIT_LIST_HEAD(&cma_dev->id_list);
3508 ib_set_client_data(device, &cma_client, cma_dev);
3511 list_add_tail(&cma_dev->list, &dev_list);
3512 list_for_each_entry(id_priv, &listen_any_list, list)
3513 cma_listen_on_dev(id_priv, cma_dev);
3514 mutex_unlock(&lock);
3517 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
3519 struct rdma_cm_event event;
3520 enum rdma_cm_state state;
3523 /* Record that we want to remove the device */
3524 state = cma_exch(id_priv, RDMA_CM_DEVICE_REMOVAL);
3525 if (state == RDMA_CM_DESTROYING)
3528 cma_cancel_operation(id_priv, state);
3529 mutex_lock(&id_priv->handler_mutex);
3531 /* Check for destruction from another callback. */
3532 if (!cma_comp(id_priv, RDMA_CM_DEVICE_REMOVAL))
3535 memset(&event, 0, sizeof event);
3536 event.event = RDMA_CM_EVENT_DEVICE_REMOVAL;
3537 ret = id_priv->id.event_handler(&id_priv->id, &event);
3539 mutex_unlock(&id_priv->handler_mutex);
3543 static void cma_process_remove(struct cma_device *cma_dev)
3545 struct rdma_id_private *id_priv;
3549 while (!list_empty(&cma_dev->id_list)) {
3550 id_priv = list_entry(cma_dev->id_list.next,
3551 struct rdma_id_private, list);
3553 list_del(&id_priv->listen_list);
3554 list_del_init(&id_priv->list);
3555 atomic_inc(&id_priv->refcount);
3556 mutex_unlock(&lock);
3558 ret = id_priv->internal_id ? 1 : cma_remove_id_dev(id_priv);
3559 cma_deref_id(id_priv);
3561 rdma_destroy_id(&id_priv->id);
3565 mutex_unlock(&lock);
3567 cma_deref_dev(cma_dev);
3568 wait_for_completion(&cma_dev->comp);
3571 static void cma_remove_one(struct ib_device *device)
3573 struct cma_device *cma_dev;
3575 cma_dev = ib_get_client_data(device, &cma_client);
3580 list_del(&cma_dev->list);
3581 mutex_unlock(&lock);
3583 cma_process_remove(cma_dev);
3587 static int cma_get_id_stats(struct sk_buff *skb, struct netlink_callback *cb)
3589 struct nlmsghdr *nlh;
3590 struct rdma_cm_id_stats *id_stats;
3591 struct rdma_id_private *id_priv;
3592 struct rdma_cm_id *id = NULL;
3593 struct cma_device *cma_dev;
3594 int i_dev = 0, i_id = 0;
3597 * We export all of the IDs as a sequence of messages. Each
3598 * ID gets its own netlink message.
3602 list_for_each_entry(cma_dev, &dev_list, list) {
3603 if (i_dev < cb->args[0]) {
3609 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
3610 if (i_id < cb->args[1]) {
3615 id_stats = ibnl_put_msg(skb, &nlh, cb->nlh->nlmsg_seq,
3616 sizeof *id_stats, RDMA_NL_RDMA_CM,
3617 RDMA_NL_RDMA_CM_ID_STATS,
3622 memset(id_stats, 0, sizeof *id_stats);
3624 id_stats->node_type = id->route.addr.dev_addr.dev_type;
3625 id_stats->port_num = id->port_num;
3626 id_stats->bound_dev_if =
3627 id->route.addr.dev_addr.bound_dev_if;
3629 if (ibnl_put_attr(skb, nlh,
3630 rdma_addr_size(cma_src_addr(id_priv)),
3631 cma_src_addr(id_priv),
3632 RDMA_NL_RDMA_CM_ATTR_SRC_ADDR))
3634 if (ibnl_put_attr(skb, nlh,
3635 rdma_addr_size(cma_src_addr(id_priv)),
3636 cma_dst_addr(id_priv),
3637 RDMA_NL_RDMA_CM_ATTR_DST_ADDR))
3640 id_stats->pid = id_priv->owner;
3641 id_stats->port_space = id->ps;
3642 id_stats->cm_state = id_priv->state;
3643 id_stats->qp_num = id_priv->qp_num;
3644 id_stats->qp_type = id->qp_type;
3654 mutex_unlock(&lock);
3655 cb->args[0] = i_dev;
3661 static const struct ibnl_client_cbs cma_cb_table[] = {
3662 [RDMA_NL_RDMA_CM_ID_STATS] = { .dump = cma_get_id_stats,
3663 .module = THIS_MODULE },
3666 static int __init cma_init(void)
3670 cma_wq = create_singlethread_workqueue("rdma_cm");
3674 ib_sa_register_client(&sa_client);
3675 rdma_addr_register_client(&addr_client);
3676 register_netdevice_notifier(&cma_nb);
3678 ret = ib_register_client(&cma_client);
3682 if (ibnl_add_client(RDMA_NL_RDMA_CM, RDMA_NL_RDMA_CM_NUM_OPS, cma_cb_table))
3683 printk(KERN_WARNING "RDMA CMA: failed to add netlink callback\n");
3688 unregister_netdevice_notifier(&cma_nb);
3689 rdma_addr_unregister_client(&addr_client);
3690 ib_sa_unregister_client(&sa_client);
3691 destroy_workqueue(cma_wq);
3695 static void __exit cma_cleanup(void)
3697 ibnl_remove_client(RDMA_NL_RDMA_CM);
3698 ib_unregister_client(&cma_client);
3699 unregister_netdevice_notifier(&cma_nb);
3700 rdma_addr_unregister_client(&addr_client);
3701 ib_sa_unregister_client(&sa_client);
3702 destroy_workqueue(cma_wq);
3703 idr_destroy(&tcp_ps);
3704 idr_destroy(&udp_ps);
3705 idr_destroy(&ipoib_ps);
3706 idr_destroy(&ib_ps);
3709 module_init(cma_init);
3710 module_exit(cma_cleanup);