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;
849 ib->sib_pkey = path->pkey;
850 ib->sib_flowinfo = path->flow_label;
851 memcpy(&ib->sib_addr, &path->sgid, 16);
853 ib->sib_pkey = listen_ib->sib_pkey;
854 ib->sib_flowinfo = listen_ib->sib_flowinfo;
855 ib->sib_addr = listen_ib->sib_addr;
857 ib->sib_sid = listen_ib->sib_sid;
858 ib->sib_sid_mask = cpu_to_be64(0xffffffffffffffffULL);
859 ib->sib_scope_id = listen_ib->sib_scope_id;
862 ib = (struct sockaddr_ib *) &id->route.addr.dst_addr;
863 ib->sib_family = listen_ib->sib_family;
864 ib->sib_pkey = path->pkey;
865 ib->sib_flowinfo = path->flow_label;
866 memcpy(&ib->sib_addr, &path->dgid, 16);
870 static __be16 ss_get_port(const struct sockaddr_storage *ss)
872 if (ss->ss_family == AF_INET)
873 return ((struct sockaddr_in *)ss)->sin_port;
874 else if (ss->ss_family == AF_INET6)
875 return ((struct sockaddr_in6 *)ss)->sin6_port;
879 static void cma_save_ip4_info(struct rdma_cm_id *id, struct rdma_cm_id *listen_id,
882 struct sockaddr_in *ip4;
884 ip4 = (struct sockaddr_in *) &id->route.addr.src_addr;
885 ip4->sin_family = AF_INET;
886 ip4->sin_addr.s_addr = hdr->dst_addr.ip4.addr;
887 ip4->sin_port = ss_get_port(&listen_id->route.addr.src_addr);
889 ip4 = (struct sockaddr_in *) &id->route.addr.dst_addr;
890 ip4->sin_family = AF_INET;
891 ip4->sin_addr.s_addr = hdr->src_addr.ip4.addr;
892 ip4->sin_port = hdr->port;
895 static void cma_save_ip6_info(struct rdma_cm_id *id, struct rdma_cm_id *listen_id,
898 struct sockaddr_in6 *ip6;
900 ip6 = (struct sockaddr_in6 *) &id->route.addr.src_addr;
901 ip6->sin6_family = AF_INET6;
902 ip6->sin6_addr = hdr->dst_addr.ip6;
903 ip6->sin6_port = ss_get_port(&listen_id->route.addr.src_addr);
905 ip6 = (struct sockaddr_in6 *) &id->route.addr.dst_addr;
906 ip6->sin6_family = AF_INET6;
907 ip6->sin6_addr = hdr->src_addr.ip6;
908 ip6->sin6_port = hdr->port;
911 static int cma_save_net_info(struct rdma_cm_id *id, struct rdma_cm_id *listen_id,
912 struct ib_cm_event *ib_event)
916 if (listen_id->route.addr.src_addr.ss_family == AF_IB) {
917 if (ib_event->event == IB_CM_REQ_RECEIVED)
918 cma_save_ib_info(id, listen_id, ib_event->param.req_rcvd.primary_path);
919 else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED)
920 cma_save_ib_info(id, listen_id, NULL);
924 hdr = ib_event->private_data;
925 if (hdr->cma_version != CMA_VERSION)
928 switch (cma_get_ip_ver(hdr)) {
930 cma_save_ip4_info(id, listen_id, hdr);
933 cma_save_ip6_info(id, listen_id, hdr);
941 static inline int cma_user_data_offset(struct rdma_id_private *id_priv)
943 return cma_family(id_priv) == AF_IB ? 0 : sizeof(struct cma_hdr);
946 static void cma_cancel_route(struct rdma_id_private *id_priv)
948 switch (rdma_port_get_link_layer(id_priv->id.device, id_priv->id.port_num)) {
949 case IB_LINK_LAYER_INFINIBAND:
951 ib_sa_cancel_query(id_priv->query_id, id_priv->query);
958 static void cma_cancel_listens(struct rdma_id_private *id_priv)
960 struct rdma_id_private *dev_id_priv;
963 * Remove from listen_any_list to prevent added devices from spawning
964 * additional listen requests.
967 list_del(&id_priv->list);
969 while (!list_empty(&id_priv->listen_list)) {
970 dev_id_priv = list_entry(id_priv->listen_list.next,
971 struct rdma_id_private, listen_list);
972 /* sync with device removal to avoid duplicate destruction */
973 list_del_init(&dev_id_priv->list);
974 list_del(&dev_id_priv->listen_list);
977 rdma_destroy_id(&dev_id_priv->id);
983 static void cma_cancel_operation(struct rdma_id_private *id_priv,
984 enum rdma_cm_state state)
987 case RDMA_CM_ADDR_QUERY:
988 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
990 case RDMA_CM_ROUTE_QUERY:
991 cma_cancel_route(id_priv);
994 if (cma_any_addr(cma_src_addr(id_priv)) && !id_priv->cma_dev)
995 cma_cancel_listens(id_priv);
1002 static void cma_release_port(struct rdma_id_private *id_priv)
1004 struct rdma_bind_list *bind_list = id_priv->bind_list;
1010 hlist_del(&id_priv->node);
1011 if (hlist_empty(&bind_list->owners)) {
1012 idr_remove(bind_list->ps, bind_list->port);
1015 mutex_unlock(&lock);
1018 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
1020 struct cma_multicast *mc;
1022 while (!list_empty(&id_priv->mc_list)) {
1023 mc = container_of(id_priv->mc_list.next,
1024 struct cma_multicast, list);
1025 list_del(&mc->list);
1026 switch (rdma_port_get_link_layer(id_priv->cma_dev->device, id_priv->id.port_num)) {
1027 case IB_LINK_LAYER_INFINIBAND:
1028 ib_sa_free_multicast(mc->multicast.ib);
1031 case IB_LINK_LAYER_ETHERNET:
1032 kref_put(&mc->mcref, release_mc);
1040 void rdma_destroy_id(struct rdma_cm_id *id)
1042 struct rdma_id_private *id_priv;
1043 enum rdma_cm_state state;
1045 id_priv = container_of(id, struct rdma_id_private, id);
1046 state = cma_exch(id_priv, RDMA_CM_DESTROYING);
1047 cma_cancel_operation(id_priv, state);
1050 * Wait for any active callback to finish. New callbacks will find
1051 * the id_priv state set to destroying and abort.
1053 mutex_lock(&id_priv->handler_mutex);
1054 mutex_unlock(&id_priv->handler_mutex);
1056 if (id_priv->cma_dev) {
1057 switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
1058 case RDMA_TRANSPORT_IB:
1059 if (id_priv->cm_id.ib)
1060 ib_destroy_cm_id(id_priv->cm_id.ib);
1062 case RDMA_TRANSPORT_IWARP:
1063 if (id_priv->cm_id.iw)
1064 iw_destroy_cm_id(id_priv->cm_id.iw);
1069 cma_leave_mc_groups(id_priv);
1070 cma_release_dev(id_priv);
1073 cma_release_port(id_priv);
1074 cma_deref_id(id_priv);
1075 wait_for_completion(&id_priv->comp);
1077 if (id_priv->internal_id)
1078 cma_deref_id(id_priv->id.context);
1080 kfree(id_priv->id.route.path_rec);
1083 EXPORT_SYMBOL(rdma_destroy_id);
1085 static int cma_rep_recv(struct rdma_id_private *id_priv)
1089 ret = cma_modify_qp_rtr(id_priv, NULL);
1093 ret = cma_modify_qp_rts(id_priv, NULL);
1097 ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
1103 cma_modify_qp_err(id_priv);
1104 ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
1109 static void cma_set_rep_event_data(struct rdma_cm_event *event,
1110 struct ib_cm_rep_event_param *rep_data,
1113 event->param.conn.private_data = private_data;
1114 event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
1115 event->param.conn.responder_resources = rep_data->responder_resources;
1116 event->param.conn.initiator_depth = rep_data->initiator_depth;
1117 event->param.conn.flow_control = rep_data->flow_control;
1118 event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
1119 event->param.conn.srq = rep_data->srq;
1120 event->param.conn.qp_num = rep_data->remote_qpn;
1123 static int cma_ib_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1125 struct rdma_id_private *id_priv = cm_id->context;
1126 struct rdma_cm_event event;
1129 if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
1130 cma_disable_callback(id_priv, RDMA_CM_CONNECT)) ||
1131 (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
1132 cma_disable_callback(id_priv, RDMA_CM_DISCONNECT)))
1135 memset(&event, 0, sizeof event);
1136 switch (ib_event->event) {
1137 case IB_CM_REQ_ERROR:
1138 case IB_CM_REP_ERROR:
1139 event.event = RDMA_CM_EVENT_UNREACHABLE;
1140 event.status = -ETIMEDOUT;
1142 case IB_CM_REP_RECEIVED:
1143 if (id_priv->id.qp) {
1144 event.status = cma_rep_recv(id_priv);
1145 event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
1146 RDMA_CM_EVENT_ESTABLISHED;
1148 event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
1150 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
1151 ib_event->private_data);
1153 case IB_CM_RTU_RECEIVED:
1154 case IB_CM_USER_ESTABLISHED:
1155 event.event = RDMA_CM_EVENT_ESTABLISHED;
1157 case IB_CM_DREQ_ERROR:
1158 event.status = -ETIMEDOUT; /* fall through */
1159 case IB_CM_DREQ_RECEIVED:
1160 case IB_CM_DREP_RECEIVED:
1161 if (!cma_comp_exch(id_priv, RDMA_CM_CONNECT,
1162 RDMA_CM_DISCONNECT))
1164 event.event = RDMA_CM_EVENT_DISCONNECTED;
1166 case IB_CM_TIMEWAIT_EXIT:
1167 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
1169 case IB_CM_MRA_RECEIVED:
1172 case IB_CM_REJ_RECEIVED:
1173 cma_modify_qp_err(id_priv);
1174 event.status = ib_event->param.rej_rcvd.reason;
1175 event.event = RDMA_CM_EVENT_REJECTED;
1176 event.param.conn.private_data = ib_event->private_data;
1177 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
1180 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
1185 ret = id_priv->id.event_handler(&id_priv->id, &event);
1187 /* Destroy the CM ID by returning a non-zero value. */
1188 id_priv->cm_id.ib = NULL;
1189 cma_exch(id_priv, RDMA_CM_DESTROYING);
1190 mutex_unlock(&id_priv->handler_mutex);
1191 rdma_destroy_id(&id_priv->id);
1195 mutex_unlock(&id_priv->handler_mutex);
1199 static struct rdma_id_private *cma_new_conn_id(struct rdma_cm_id *listen_id,
1200 struct ib_cm_event *ib_event)
1202 struct rdma_id_private *id_priv;
1203 struct rdma_cm_id *id;
1204 struct rdma_route *rt;
1207 id = rdma_create_id(listen_id->event_handler, listen_id->context,
1208 listen_id->ps, ib_event->param.req_rcvd.qp_type);
1212 id_priv = container_of(id, struct rdma_id_private, id);
1213 if (cma_save_net_info(id, listen_id, ib_event))
1217 rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
1218 rt->path_rec = kmalloc(sizeof *rt->path_rec * rt->num_paths,
1223 rt->path_rec[0] = *ib_event->param.req_rcvd.primary_path;
1224 if (rt->num_paths == 2)
1225 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
1227 if (cma_any_addr(cma_src_addr(id_priv))) {
1228 rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
1229 rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
1230 ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
1232 ret = cma_translate_addr(cma_src_addr(id_priv), &rt->addr.dev_addr);
1236 rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
1238 id_priv->state = RDMA_CM_CONNECT;
1242 rdma_destroy_id(id);
1246 static struct rdma_id_private *cma_new_udp_id(struct rdma_cm_id *listen_id,
1247 struct ib_cm_event *ib_event)
1249 struct rdma_id_private *id_priv;
1250 struct rdma_cm_id *id;
1253 id = rdma_create_id(listen_id->event_handler, listen_id->context,
1254 listen_id->ps, IB_QPT_UD);
1258 id_priv = container_of(id, struct rdma_id_private, id);
1259 if (cma_save_net_info(id, listen_id, ib_event))
1262 if (!cma_any_addr((struct sockaddr *) &id->route.addr.src_addr)) {
1263 ret = cma_translate_addr(cma_src_addr(id_priv), &id->route.addr.dev_addr);
1268 id_priv->state = RDMA_CM_CONNECT;
1271 rdma_destroy_id(id);
1275 static void cma_set_req_event_data(struct rdma_cm_event *event,
1276 struct ib_cm_req_event_param *req_data,
1277 void *private_data, int offset)
1279 event->param.conn.private_data = private_data + offset;
1280 event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
1281 event->param.conn.responder_resources = req_data->responder_resources;
1282 event->param.conn.initiator_depth = req_data->initiator_depth;
1283 event->param.conn.flow_control = req_data->flow_control;
1284 event->param.conn.retry_count = req_data->retry_count;
1285 event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
1286 event->param.conn.srq = req_data->srq;
1287 event->param.conn.qp_num = req_data->remote_qpn;
1290 static int cma_check_req_qp_type(struct rdma_cm_id *id, struct ib_cm_event *ib_event)
1292 return (((ib_event->event == IB_CM_REQ_RECEIVED) &&
1293 (ib_event->param.req_rcvd.qp_type == id->qp_type)) ||
1294 ((ib_event->event == IB_CM_SIDR_REQ_RECEIVED) &&
1295 (id->qp_type == IB_QPT_UD)) ||
1299 static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1301 struct rdma_id_private *listen_id, *conn_id;
1302 struct rdma_cm_event event;
1305 listen_id = cm_id->context;
1306 if (!cma_check_req_qp_type(&listen_id->id, ib_event))
1309 if (cma_disable_callback(listen_id, RDMA_CM_LISTEN))
1310 return -ECONNABORTED;
1312 memset(&event, 0, sizeof event);
1313 offset = cma_user_data_offset(listen_id);
1314 event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1315 if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) {
1316 conn_id = cma_new_udp_id(&listen_id->id, ib_event);
1317 event.param.ud.private_data = ib_event->private_data + offset;
1318 event.param.ud.private_data_len =
1319 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
1321 conn_id = cma_new_conn_id(&listen_id->id, ib_event);
1322 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
1323 ib_event->private_data, offset);
1330 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1331 ret = cma_acquire_dev(conn_id, listen_id);
1335 conn_id->cm_id.ib = cm_id;
1336 cm_id->context = conn_id;
1337 cm_id->cm_handler = cma_ib_handler;
1340 * Protect against the user destroying conn_id from another thread
1341 * until we're done accessing it.
1343 atomic_inc(&conn_id->refcount);
1344 ret = conn_id->id.event_handler(&conn_id->id, &event);
1348 * Acquire mutex to prevent user executing rdma_destroy_id()
1349 * while we're accessing the cm_id.
1352 if (cma_comp(conn_id, RDMA_CM_CONNECT) &&
1353 (conn_id->id.qp_type != IB_QPT_UD))
1354 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
1355 mutex_unlock(&lock);
1356 mutex_unlock(&conn_id->handler_mutex);
1357 mutex_unlock(&listen_id->handler_mutex);
1358 cma_deref_id(conn_id);
1362 cma_deref_id(conn_id);
1363 /* Destroy the CM ID by returning a non-zero value. */
1364 conn_id->cm_id.ib = NULL;
1366 cma_exch(conn_id, RDMA_CM_DESTROYING);
1367 mutex_unlock(&conn_id->handler_mutex);
1369 mutex_unlock(&listen_id->handler_mutex);
1371 rdma_destroy_id(&conn_id->id);
1375 __be64 rdma_get_service_id(struct rdma_cm_id *id, struct sockaddr *addr)
1377 if (addr->sa_family == AF_IB)
1378 return ((struct sockaddr_ib *) addr)->sib_sid;
1380 return cpu_to_be64(((u64)id->ps << 16) + be16_to_cpu(cma_port(addr)));
1382 EXPORT_SYMBOL(rdma_get_service_id);
1384 static void cma_set_compare_data(enum rdma_port_space ps, struct sockaddr *addr,
1385 struct ib_cm_compare_data *compare)
1387 struct cma_hdr *cma_data, *cma_mask;
1389 struct in6_addr ip6_addr;
1391 memset(compare, 0, sizeof *compare);
1392 cma_data = (void *) compare->data;
1393 cma_mask = (void *) compare->mask;
1395 switch (addr->sa_family) {
1397 ip4_addr = ((struct sockaddr_in *) addr)->sin_addr.s_addr;
1398 cma_set_ip_ver(cma_data, 4);
1399 cma_set_ip_ver(cma_mask, 0xF);
1400 if (!cma_any_addr(addr)) {
1401 cma_data->dst_addr.ip4.addr = ip4_addr;
1402 cma_mask->dst_addr.ip4.addr = htonl(~0);
1406 ip6_addr = ((struct sockaddr_in6 *) addr)->sin6_addr;
1407 cma_set_ip_ver(cma_data, 6);
1408 cma_set_ip_ver(cma_mask, 0xF);
1409 if (!cma_any_addr(addr)) {
1410 cma_data->dst_addr.ip6 = ip6_addr;
1411 memset(&cma_mask->dst_addr.ip6, 0xFF,
1412 sizeof cma_mask->dst_addr.ip6);
1420 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
1422 struct rdma_id_private *id_priv = iw_id->context;
1423 struct rdma_cm_event event;
1425 struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
1426 struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
1428 if (cma_disable_callback(id_priv, RDMA_CM_CONNECT))
1431 memset(&event, 0, sizeof event);
1432 switch (iw_event->event) {
1433 case IW_CM_EVENT_CLOSE:
1434 event.event = RDMA_CM_EVENT_DISCONNECTED;
1436 case IW_CM_EVENT_CONNECT_REPLY:
1437 memcpy(cma_src_addr(id_priv), laddr,
1438 rdma_addr_size(laddr));
1439 memcpy(cma_dst_addr(id_priv), raddr,
1440 rdma_addr_size(raddr));
1441 switch (iw_event->status) {
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;
1449 event.event = RDMA_CM_EVENT_REJECTED;
1452 event.event = RDMA_CM_EVENT_UNREACHABLE;
1455 event.event = RDMA_CM_EVENT_CONNECT_ERROR;
1459 case IW_CM_EVENT_ESTABLISHED:
1460 event.event = RDMA_CM_EVENT_ESTABLISHED;
1461 event.param.conn.initiator_depth = iw_event->ird;
1462 event.param.conn.responder_resources = iw_event->ord;
1468 event.status = iw_event->status;
1469 event.param.conn.private_data = iw_event->private_data;
1470 event.param.conn.private_data_len = iw_event->private_data_len;
1471 ret = id_priv->id.event_handler(&id_priv->id, &event);
1473 /* Destroy the CM ID by returning a non-zero value. */
1474 id_priv->cm_id.iw = NULL;
1475 cma_exch(id_priv, RDMA_CM_DESTROYING);
1476 mutex_unlock(&id_priv->handler_mutex);
1477 rdma_destroy_id(&id_priv->id);
1481 mutex_unlock(&id_priv->handler_mutex);
1485 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
1486 struct iw_cm_event *iw_event)
1488 struct rdma_cm_id *new_cm_id;
1489 struct rdma_id_private *listen_id, *conn_id;
1490 struct rdma_cm_event event;
1492 struct ib_device_attr attr;
1493 struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
1494 struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
1496 listen_id = cm_id->context;
1497 if (cma_disable_callback(listen_id, RDMA_CM_LISTEN))
1498 return -ECONNABORTED;
1500 /* Create a new RDMA id for the new IW CM ID */
1501 new_cm_id = rdma_create_id(listen_id->id.event_handler,
1502 listen_id->id.context,
1503 RDMA_PS_TCP, IB_QPT_RC);
1504 if (IS_ERR(new_cm_id)) {
1508 conn_id = container_of(new_cm_id, struct rdma_id_private, id);
1509 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1510 conn_id->state = RDMA_CM_CONNECT;
1512 ret = rdma_translate_ip(laddr, &conn_id->id.route.addr.dev_addr, NULL);
1514 mutex_unlock(&conn_id->handler_mutex);
1515 rdma_destroy_id(new_cm_id);
1519 ret = cma_acquire_dev(conn_id, listen_id);
1521 mutex_unlock(&conn_id->handler_mutex);
1522 rdma_destroy_id(new_cm_id);
1526 conn_id->cm_id.iw = cm_id;
1527 cm_id->context = conn_id;
1528 cm_id->cm_handler = cma_iw_handler;
1530 memcpy(cma_src_addr(conn_id), laddr, rdma_addr_size(laddr));
1531 memcpy(cma_dst_addr(conn_id), raddr, rdma_addr_size(raddr));
1533 ret = ib_query_device(conn_id->id.device, &attr);
1535 mutex_unlock(&conn_id->handler_mutex);
1536 rdma_destroy_id(new_cm_id);
1540 memset(&event, 0, sizeof event);
1541 event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1542 event.param.conn.private_data = iw_event->private_data;
1543 event.param.conn.private_data_len = iw_event->private_data_len;
1544 event.param.conn.initiator_depth = iw_event->ird;
1545 event.param.conn.responder_resources = iw_event->ord;
1548 * Protect against the user destroying conn_id from another thread
1549 * until we're done accessing it.
1551 atomic_inc(&conn_id->refcount);
1552 ret = conn_id->id.event_handler(&conn_id->id, &event);
1554 /* User wants to destroy the CM ID */
1555 conn_id->cm_id.iw = NULL;
1556 cma_exch(conn_id, RDMA_CM_DESTROYING);
1557 mutex_unlock(&conn_id->handler_mutex);
1558 cma_deref_id(conn_id);
1559 rdma_destroy_id(&conn_id->id);
1563 mutex_unlock(&conn_id->handler_mutex);
1564 cma_deref_id(conn_id);
1567 mutex_unlock(&listen_id->handler_mutex);
1571 static int cma_ib_listen(struct rdma_id_private *id_priv)
1573 struct ib_cm_compare_data compare_data;
1574 struct sockaddr *addr;
1575 struct ib_cm_id *id;
1579 id = ib_create_cm_id(id_priv->id.device, cma_req_handler, id_priv);
1583 id_priv->cm_id.ib = id;
1585 addr = cma_src_addr(id_priv);
1586 svc_id = rdma_get_service_id(&id_priv->id, addr);
1587 if (cma_any_addr(addr) && !id_priv->afonly)
1588 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, NULL);
1590 cma_set_compare_data(id_priv->id.ps, addr, &compare_data);
1591 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, &compare_data);
1595 ib_destroy_cm_id(id_priv->cm_id.ib);
1596 id_priv->cm_id.ib = NULL;
1602 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
1605 struct iw_cm_id *id;
1607 id = iw_create_cm_id(id_priv->id.device,
1608 iw_conn_req_handler,
1613 id_priv->cm_id.iw = id;
1615 memcpy(&id_priv->cm_id.iw->local_addr, cma_src_addr(id_priv),
1616 rdma_addr_size(cma_src_addr(id_priv)));
1618 ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
1621 iw_destroy_cm_id(id_priv->cm_id.iw);
1622 id_priv->cm_id.iw = NULL;
1628 static int cma_listen_handler(struct rdma_cm_id *id,
1629 struct rdma_cm_event *event)
1631 struct rdma_id_private *id_priv = id->context;
1633 id->context = id_priv->id.context;
1634 id->event_handler = id_priv->id.event_handler;
1635 return id_priv->id.event_handler(id, event);
1638 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
1639 struct cma_device *cma_dev)
1641 struct rdma_id_private *dev_id_priv;
1642 struct rdma_cm_id *id;
1645 if (cma_family(id_priv) == AF_IB &&
1646 rdma_node_get_transport(cma_dev->device->node_type) != RDMA_TRANSPORT_IB)
1649 id = rdma_create_id(cma_listen_handler, id_priv, id_priv->id.ps,
1650 id_priv->id.qp_type);
1654 dev_id_priv = container_of(id, struct rdma_id_private, id);
1656 dev_id_priv->state = RDMA_CM_ADDR_BOUND;
1657 memcpy(cma_src_addr(dev_id_priv), cma_src_addr(id_priv),
1658 rdma_addr_size(cma_src_addr(id_priv)));
1660 cma_attach_to_dev(dev_id_priv, cma_dev);
1661 list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
1662 atomic_inc(&id_priv->refcount);
1663 dev_id_priv->internal_id = 1;
1664 dev_id_priv->afonly = id_priv->afonly;
1666 ret = rdma_listen(id, id_priv->backlog);
1668 printk(KERN_WARNING "RDMA CMA: cma_listen_on_dev, error %d, "
1669 "listening on device %s\n", ret, cma_dev->device->name);
1672 static void cma_listen_on_all(struct rdma_id_private *id_priv)
1674 struct cma_device *cma_dev;
1677 list_add_tail(&id_priv->list, &listen_any_list);
1678 list_for_each_entry(cma_dev, &dev_list, list)
1679 cma_listen_on_dev(id_priv, cma_dev);
1680 mutex_unlock(&lock);
1683 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
1685 struct rdma_id_private *id_priv;
1687 id_priv = container_of(id, struct rdma_id_private, id);
1688 id_priv->tos = (u8) tos;
1690 EXPORT_SYMBOL(rdma_set_service_type);
1692 static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec,
1695 struct cma_work *work = context;
1696 struct rdma_route *route;
1698 route = &work->id->id.route;
1701 route->num_paths = 1;
1702 *route->path_rec = *path_rec;
1704 work->old_state = RDMA_CM_ROUTE_QUERY;
1705 work->new_state = RDMA_CM_ADDR_RESOLVED;
1706 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
1707 work->event.status = status;
1710 queue_work(cma_wq, &work->work);
1713 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms,
1714 struct cma_work *work)
1716 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
1717 struct ib_sa_path_rec path_rec;
1718 ib_sa_comp_mask comp_mask;
1719 struct sockaddr_in6 *sin6;
1720 struct sockaddr_ib *sib;
1722 memset(&path_rec, 0, sizeof path_rec);
1723 rdma_addr_get_sgid(dev_addr, &path_rec.sgid);
1724 rdma_addr_get_dgid(dev_addr, &path_rec.dgid);
1725 path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
1726 path_rec.numb_path = 1;
1727 path_rec.reversible = 1;
1728 path_rec.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
1730 comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
1731 IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
1732 IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
1734 switch (cma_family(id_priv)) {
1736 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
1737 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
1740 sin6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
1741 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
1742 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
1745 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
1746 path_rec.traffic_class = (u8) (be32_to_cpu(sib->sib_flowinfo) >> 20);
1747 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
1751 id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
1752 id_priv->id.port_num, &path_rec,
1753 comp_mask, timeout_ms,
1754 GFP_KERNEL, cma_query_handler,
1755 work, &id_priv->query);
1757 return (id_priv->query_id < 0) ? id_priv->query_id : 0;
1760 static void cma_work_handler(struct work_struct *_work)
1762 struct cma_work *work = container_of(_work, struct cma_work, work);
1763 struct rdma_id_private *id_priv = work->id;
1766 mutex_lock(&id_priv->handler_mutex);
1767 if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
1770 if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1771 cma_exch(id_priv, RDMA_CM_DESTROYING);
1775 mutex_unlock(&id_priv->handler_mutex);
1776 cma_deref_id(id_priv);
1778 rdma_destroy_id(&id_priv->id);
1782 static void cma_ndev_work_handler(struct work_struct *_work)
1784 struct cma_ndev_work *work = container_of(_work, struct cma_ndev_work, work);
1785 struct rdma_id_private *id_priv = work->id;
1788 mutex_lock(&id_priv->handler_mutex);
1789 if (id_priv->state == RDMA_CM_DESTROYING ||
1790 id_priv->state == RDMA_CM_DEVICE_REMOVAL)
1793 if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1794 cma_exch(id_priv, RDMA_CM_DESTROYING);
1799 mutex_unlock(&id_priv->handler_mutex);
1800 cma_deref_id(id_priv);
1802 rdma_destroy_id(&id_priv->id);
1806 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms)
1808 struct rdma_route *route = &id_priv->id.route;
1809 struct cma_work *work;
1812 work = kzalloc(sizeof *work, GFP_KERNEL);
1817 INIT_WORK(&work->work, cma_work_handler);
1818 work->old_state = RDMA_CM_ROUTE_QUERY;
1819 work->new_state = RDMA_CM_ROUTE_RESOLVED;
1820 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1822 route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
1823 if (!route->path_rec) {
1828 ret = cma_query_ib_route(id_priv, timeout_ms, work);
1834 kfree(route->path_rec);
1835 route->path_rec = NULL;
1841 int rdma_set_ib_paths(struct rdma_cm_id *id,
1842 struct ib_sa_path_rec *path_rec, int num_paths)
1844 struct rdma_id_private *id_priv;
1847 id_priv = container_of(id, struct rdma_id_private, id);
1848 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
1849 RDMA_CM_ROUTE_RESOLVED))
1852 id->route.path_rec = kmemdup(path_rec, sizeof *path_rec * num_paths,
1854 if (!id->route.path_rec) {
1859 id->route.num_paths = num_paths;
1862 cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED);
1865 EXPORT_SYMBOL(rdma_set_ib_paths);
1867 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms)
1869 struct cma_work *work;
1871 work = kzalloc(sizeof *work, GFP_KERNEL);
1876 INIT_WORK(&work->work, cma_work_handler);
1877 work->old_state = RDMA_CM_ROUTE_QUERY;
1878 work->new_state = RDMA_CM_ROUTE_RESOLVED;
1879 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1880 queue_work(cma_wq, &work->work);
1884 static int iboe_tos_to_sl(struct net_device *ndev, int tos)
1887 struct net_device *dev;
1889 prio = rt_tos2priority(tos);
1890 dev = ndev->priv_flags & IFF_802_1Q_VLAN ?
1891 vlan_dev_real_dev(ndev) : ndev;
1894 return netdev_get_prio_tc_map(dev, prio);
1896 #if IS_ENABLED(CONFIG_VLAN_8021Q)
1897 if (ndev->priv_flags & IFF_802_1Q_VLAN)
1898 return (vlan_dev_get_egress_qos_mask(ndev, prio) &
1899 VLAN_PRIO_MASK) >> VLAN_PRIO_SHIFT;
1904 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
1906 struct rdma_route *route = &id_priv->id.route;
1907 struct rdma_addr *addr = &route->addr;
1908 struct cma_work *work;
1910 struct net_device *ndev = NULL;
1913 work = kzalloc(sizeof *work, GFP_KERNEL);
1918 INIT_WORK(&work->work, cma_work_handler);
1920 route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
1921 if (!route->path_rec) {
1926 route->num_paths = 1;
1928 if (addr->dev_addr.bound_dev_if)
1929 ndev = dev_get_by_index(&init_net, addr->dev_addr.bound_dev_if);
1935 route->path_rec->vlan_id = rdma_vlan_dev_vlan_id(ndev);
1936 memcpy(route->path_rec->dmac, addr->dev_addr.dst_dev_addr, ETH_ALEN);
1937 memcpy(route->path_rec->smac, ndev->dev_addr, ndev->addr_len);
1939 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
1940 &route->path_rec->sgid);
1941 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.dst_addr,
1942 &route->path_rec->dgid);
1944 route->path_rec->hop_limit = 1;
1945 route->path_rec->reversible = 1;
1946 route->path_rec->pkey = cpu_to_be16(0xffff);
1947 route->path_rec->mtu_selector = IB_SA_EQ;
1948 route->path_rec->sl = iboe_tos_to_sl(ndev, id_priv->tos);
1949 route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
1950 route->path_rec->rate_selector = IB_SA_EQ;
1951 route->path_rec->rate = iboe_get_rate(ndev);
1953 route->path_rec->packet_life_time_selector = IB_SA_EQ;
1954 route->path_rec->packet_life_time = CMA_IBOE_PACKET_LIFETIME;
1955 if (!route->path_rec->mtu) {
1960 work->old_state = RDMA_CM_ROUTE_QUERY;
1961 work->new_state = RDMA_CM_ROUTE_RESOLVED;
1962 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1963 work->event.status = 0;
1965 queue_work(cma_wq, &work->work);
1970 kfree(route->path_rec);
1971 route->path_rec = NULL;
1977 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
1979 struct rdma_id_private *id_priv;
1982 id_priv = container_of(id, struct rdma_id_private, id);
1983 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY))
1986 atomic_inc(&id_priv->refcount);
1987 switch (rdma_node_get_transport(id->device->node_type)) {
1988 case RDMA_TRANSPORT_IB:
1989 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
1990 case IB_LINK_LAYER_INFINIBAND:
1991 ret = cma_resolve_ib_route(id_priv, timeout_ms);
1993 case IB_LINK_LAYER_ETHERNET:
1994 ret = cma_resolve_iboe_route(id_priv);
2000 case RDMA_TRANSPORT_IWARP:
2001 ret = cma_resolve_iw_route(id_priv, timeout_ms);
2012 cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED);
2013 cma_deref_id(id_priv);
2016 EXPORT_SYMBOL(rdma_resolve_route);
2018 static void cma_set_loopback(struct sockaddr *addr)
2020 switch (addr->sa_family) {
2022 ((struct sockaddr_in *) addr)->sin_addr.s_addr = htonl(INADDR_LOOPBACK);
2025 ipv6_addr_set(&((struct sockaddr_in6 *) addr)->sin6_addr,
2029 ib_addr_set(&((struct sockaddr_ib *) addr)->sib_addr,
2035 static int cma_bind_loopback(struct rdma_id_private *id_priv)
2037 struct cma_device *cma_dev, *cur_dev;
2038 struct ib_port_attr port_attr;
2046 list_for_each_entry(cur_dev, &dev_list, list) {
2047 if (cma_family(id_priv) == AF_IB &&
2048 rdma_node_get_transport(cur_dev->device->node_type) != RDMA_TRANSPORT_IB)
2054 for (p = 1; p <= cur_dev->device->phys_port_cnt; ++p) {
2055 if (!ib_query_port(cur_dev->device, p, &port_attr) &&
2056 port_attr.state == IB_PORT_ACTIVE) {
2071 ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid);
2075 ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
2079 id_priv->id.route.addr.dev_addr.dev_type =
2080 (rdma_port_get_link_layer(cma_dev->device, p) == IB_LINK_LAYER_INFINIBAND) ?
2081 ARPHRD_INFINIBAND : ARPHRD_ETHER;
2083 rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2084 ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
2085 id_priv->id.port_num = p;
2086 cma_attach_to_dev(id_priv, cma_dev);
2087 cma_set_loopback(cma_src_addr(id_priv));
2089 mutex_unlock(&lock);
2093 static void addr_handler(int status, struct sockaddr *src_addr,
2094 struct rdma_dev_addr *dev_addr, void *context)
2096 struct rdma_id_private *id_priv = context;
2097 struct rdma_cm_event event;
2099 memset(&event, 0, sizeof event);
2100 mutex_lock(&id_priv->handler_mutex);
2101 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY,
2102 RDMA_CM_ADDR_RESOLVED))
2105 memcpy(cma_src_addr(id_priv), src_addr, rdma_addr_size(src_addr));
2106 if (!status && !id_priv->cma_dev)
2107 status = cma_acquire_dev(id_priv, NULL);
2110 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
2111 RDMA_CM_ADDR_BOUND))
2113 event.event = RDMA_CM_EVENT_ADDR_ERROR;
2114 event.status = status;
2116 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2118 if (id_priv->id.event_handler(&id_priv->id, &event)) {
2119 cma_exch(id_priv, RDMA_CM_DESTROYING);
2120 mutex_unlock(&id_priv->handler_mutex);
2121 cma_deref_id(id_priv);
2122 rdma_destroy_id(&id_priv->id);
2126 mutex_unlock(&id_priv->handler_mutex);
2127 cma_deref_id(id_priv);
2130 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
2132 struct cma_work *work;
2136 work = kzalloc(sizeof *work, GFP_KERNEL);
2140 if (!id_priv->cma_dev) {
2141 ret = cma_bind_loopback(id_priv);
2146 rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2147 rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
2150 INIT_WORK(&work->work, cma_work_handler);
2151 work->old_state = RDMA_CM_ADDR_QUERY;
2152 work->new_state = RDMA_CM_ADDR_RESOLVED;
2153 work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2154 queue_work(cma_wq, &work->work);
2161 static int cma_resolve_ib_addr(struct rdma_id_private *id_priv)
2163 struct cma_work *work;
2166 work = kzalloc(sizeof *work, GFP_KERNEL);
2170 if (!id_priv->cma_dev) {
2171 ret = cma_resolve_ib_dev(id_priv);
2176 rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, (union ib_gid *)
2177 &(((struct sockaddr_ib *) &id_priv->id.route.addr.dst_addr)->sib_addr));
2180 INIT_WORK(&work->work, cma_work_handler);
2181 work->old_state = RDMA_CM_ADDR_QUERY;
2182 work->new_state = RDMA_CM_ADDR_RESOLVED;
2183 work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2184 queue_work(cma_wq, &work->work);
2191 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2192 struct sockaddr *dst_addr)
2194 if (!src_addr || !src_addr->sa_family) {
2195 src_addr = (struct sockaddr *) &id->route.addr.src_addr;
2196 src_addr->sa_family = dst_addr->sa_family;
2197 if (dst_addr->sa_family == AF_INET6) {
2198 ((struct sockaddr_in6 *) src_addr)->sin6_scope_id =
2199 ((struct sockaddr_in6 *) dst_addr)->sin6_scope_id;
2200 } else if (dst_addr->sa_family == AF_IB) {
2201 ((struct sockaddr_ib *) src_addr)->sib_pkey =
2202 ((struct sockaddr_ib *) dst_addr)->sib_pkey;
2205 return rdma_bind_addr(id, src_addr);
2208 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2209 struct sockaddr *dst_addr, int timeout_ms)
2211 struct rdma_id_private *id_priv;
2214 id_priv = container_of(id, struct rdma_id_private, id);
2215 if (id_priv->state == RDMA_CM_IDLE) {
2216 ret = cma_bind_addr(id, src_addr, dst_addr);
2221 if (cma_family(id_priv) != dst_addr->sa_family)
2224 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY))
2227 atomic_inc(&id_priv->refcount);
2228 memcpy(cma_dst_addr(id_priv), dst_addr, rdma_addr_size(dst_addr));
2229 if (cma_any_addr(dst_addr)) {
2230 ret = cma_resolve_loopback(id_priv);
2232 if (dst_addr->sa_family == AF_IB) {
2233 ret = cma_resolve_ib_addr(id_priv);
2235 ret = rdma_resolve_ip(&addr_client, cma_src_addr(id_priv),
2236 dst_addr, &id->route.addr.dev_addr,
2237 timeout_ms, addr_handler, id_priv);
2245 cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
2246 cma_deref_id(id_priv);
2249 EXPORT_SYMBOL(rdma_resolve_addr);
2251 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse)
2253 struct rdma_id_private *id_priv;
2254 unsigned long flags;
2257 id_priv = container_of(id, struct rdma_id_private, id);
2258 spin_lock_irqsave(&id_priv->lock, flags);
2259 if (reuse || id_priv->state == RDMA_CM_IDLE) {
2260 id_priv->reuseaddr = reuse;
2265 spin_unlock_irqrestore(&id_priv->lock, flags);
2268 EXPORT_SYMBOL(rdma_set_reuseaddr);
2270 int rdma_set_afonly(struct rdma_cm_id *id, int afonly)
2272 struct rdma_id_private *id_priv;
2273 unsigned long flags;
2276 id_priv = container_of(id, struct rdma_id_private, id);
2277 spin_lock_irqsave(&id_priv->lock, flags);
2278 if (id_priv->state == RDMA_CM_IDLE || id_priv->state == RDMA_CM_ADDR_BOUND) {
2279 id_priv->options |= (1 << CMA_OPTION_AFONLY);
2280 id_priv->afonly = afonly;
2285 spin_unlock_irqrestore(&id_priv->lock, flags);
2288 EXPORT_SYMBOL(rdma_set_afonly);
2290 static void cma_bind_port(struct rdma_bind_list *bind_list,
2291 struct rdma_id_private *id_priv)
2293 struct sockaddr *addr;
2294 struct sockaddr_ib *sib;
2298 addr = cma_src_addr(id_priv);
2299 port = htons(bind_list->port);
2301 switch (addr->sa_family) {
2303 ((struct sockaddr_in *) addr)->sin_port = port;
2306 ((struct sockaddr_in6 *) addr)->sin6_port = port;
2309 sib = (struct sockaddr_ib *) addr;
2310 sid = be64_to_cpu(sib->sib_sid);
2311 mask = be64_to_cpu(sib->sib_sid_mask);
2312 sib->sib_sid = cpu_to_be64((sid & mask) | (u64) ntohs(port));
2313 sib->sib_sid_mask = cpu_to_be64(~0ULL);
2316 id_priv->bind_list = bind_list;
2317 hlist_add_head(&id_priv->node, &bind_list->owners);
2320 static int cma_alloc_port(struct idr *ps, struct rdma_id_private *id_priv,
2321 unsigned short snum)
2323 struct rdma_bind_list *bind_list;
2326 bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
2330 ret = idr_alloc(ps, bind_list, snum, snum + 1, GFP_KERNEL);
2335 bind_list->port = (unsigned short)ret;
2336 cma_bind_port(bind_list, id_priv);
2340 return ret == -ENOSPC ? -EADDRNOTAVAIL : ret;
2343 static int cma_alloc_any_port(struct idr *ps, struct rdma_id_private *id_priv)
2345 static unsigned int last_used_port;
2346 int low, high, remaining;
2349 inet_get_local_port_range(&init_net, &low, &high);
2350 remaining = (high - low) + 1;
2351 rover = prandom_u32() % remaining + low;
2353 if (last_used_port != rover &&
2354 !idr_find(ps, (unsigned short) rover)) {
2355 int ret = cma_alloc_port(ps, id_priv, rover);
2357 * Remember previously used port number in order to avoid
2358 * re-using same port immediately after it is closed.
2361 last_used_port = rover;
2362 if (ret != -EADDRNOTAVAIL)
2367 if ((rover < low) || (rover > high))
2371 return -EADDRNOTAVAIL;
2375 * Check that the requested port is available. This is called when trying to
2376 * bind to a specific port, or when trying to listen on a bound port. In
2377 * the latter case, the provided id_priv may already be on the bind_list, but
2378 * we still need to check that it's okay to start listening.
2380 static int cma_check_port(struct rdma_bind_list *bind_list,
2381 struct rdma_id_private *id_priv, uint8_t reuseaddr)
2383 struct rdma_id_private *cur_id;
2384 struct sockaddr *addr, *cur_addr;
2386 addr = cma_src_addr(id_priv);
2387 hlist_for_each_entry(cur_id, &bind_list->owners, node) {
2388 if (id_priv == cur_id)
2391 if ((cur_id->state != RDMA_CM_LISTEN) && reuseaddr &&
2395 cur_addr = cma_src_addr(cur_id);
2396 if (id_priv->afonly && cur_id->afonly &&
2397 (addr->sa_family != cur_addr->sa_family))
2400 if (cma_any_addr(addr) || cma_any_addr(cur_addr))
2401 return -EADDRNOTAVAIL;
2403 if (!cma_addr_cmp(addr, cur_addr))
2409 static int cma_use_port(struct idr *ps, struct rdma_id_private *id_priv)
2411 struct rdma_bind_list *bind_list;
2412 unsigned short snum;
2415 snum = ntohs(cma_port(cma_src_addr(id_priv)));
2416 if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
2419 bind_list = idr_find(ps, snum);
2421 ret = cma_alloc_port(ps, id_priv, snum);
2423 ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr);
2425 cma_bind_port(bind_list, id_priv);
2430 static int cma_bind_listen(struct rdma_id_private *id_priv)
2432 struct rdma_bind_list *bind_list = id_priv->bind_list;
2436 if (bind_list->owners.first->next)
2437 ret = cma_check_port(bind_list, id_priv, 0);
2438 mutex_unlock(&lock);
2442 static struct idr *cma_select_inet_ps(struct rdma_id_private *id_priv)
2444 switch (id_priv->id.ps) {
2458 static struct idr *cma_select_ib_ps(struct rdma_id_private *id_priv)
2460 struct idr *ps = NULL;
2461 struct sockaddr_ib *sib;
2462 u64 sid_ps, mask, sid;
2464 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
2465 mask = be64_to_cpu(sib->sib_sid_mask) & RDMA_IB_IP_PS_MASK;
2466 sid = be64_to_cpu(sib->sib_sid) & mask;
2468 if ((id_priv->id.ps == RDMA_PS_IB) && (sid == (RDMA_IB_IP_PS_IB & mask))) {
2469 sid_ps = RDMA_IB_IP_PS_IB;
2471 } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_TCP)) &&
2472 (sid == (RDMA_IB_IP_PS_TCP & mask))) {
2473 sid_ps = RDMA_IB_IP_PS_TCP;
2475 } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_UDP)) &&
2476 (sid == (RDMA_IB_IP_PS_UDP & mask))) {
2477 sid_ps = RDMA_IB_IP_PS_UDP;
2482 sib->sib_sid = cpu_to_be64(sid_ps | ntohs(cma_port((struct sockaddr *) sib)));
2483 sib->sib_sid_mask = cpu_to_be64(RDMA_IB_IP_PS_MASK |
2484 be64_to_cpu(sib->sib_sid_mask));
2489 static int cma_get_port(struct rdma_id_private *id_priv)
2494 if (cma_family(id_priv) != AF_IB)
2495 ps = cma_select_inet_ps(id_priv);
2497 ps = cma_select_ib_ps(id_priv);
2499 return -EPROTONOSUPPORT;
2502 if (cma_any_port(cma_src_addr(id_priv)))
2503 ret = cma_alloc_any_port(ps, id_priv);
2505 ret = cma_use_port(ps, id_priv);
2506 mutex_unlock(&lock);
2511 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
2512 struct sockaddr *addr)
2514 #if IS_ENABLED(CONFIG_IPV6)
2515 struct sockaddr_in6 *sin6;
2517 if (addr->sa_family != AF_INET6)
2520 sin6 = (struct sockaddr_in6 *) addr;
2522 if (!(ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL))
2525 if (!sin6->sin6_scope_id)
2528 dev_addr->bound_dev_if = sin6->sin6_scope_id;
2533 int rdma_listen(struct rdma_cm_id *id, int backlog)
2535 struct rdma_id_private *id_priv;
2538 id_priv = container_of(id, struct rdma_id_private, id);
2539 if (id_priv->state == RDMA_CM_IDLE) {
2540 id->route.addr.src_addr.ss_family = AF_INET;
2541 ret = rdma_bind_addr(id, cma_src_addr(id_priv));
2546 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN))
2549 if (id_priv->reuseaddr) {
2550 ret = cma_bind_listen(id_priv);
2555 id_priv->backlog = backlog;
2557 switch (rdma_node_get_transport(id->device->node_type)) {
2558 case RDMA_TRANSPORT_IB:
2559 ret = cma_ib_listen(id_priv);
2563 case RDMA_TRANSPORT_IWARP:
2564 ret = cma_iw_listen(id_priv, backlog);
2573 cma_listen_on_all(id_priv);
2577 id_priv->backlog = 0;
2578 cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND);
2581 EXPORT_SYMBOL(rdma_listen);
2583 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
2585 struct rdma_id_private *id_priv;
2588 if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6 &&
2589 addr->sa_family != AF_IB)
2590 return -EAFNOSUPPORT;
2592 id_priv = container_of(id, struct rdma_id_private, id);
2593 if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND))
2596 ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
2600 memcpy(cma_src_addr(id_priv), addr, rdma_addr_size(addr));
2601 if (!cma_any_addr(addr)) {
2602 ret = cma_translate_addr(addr, &id->route.addr.dev_addr);
2606 ret = cma_acquire_dev(id_priv, NULL);
2611 if (!(id_priv->options & (1 << CMA_OPTION_AFONLY))) {
2612 if (addr->sa_family == AF_INET)
2613 id_priv->afonly = 1;
2614 #if IS_ENABLED(CONFIG_IPV6)
2615 else if (addr->sa_family == AF_INET6)
2616 id_priv->afonly = init_net.ipv6.sysctl.bindv6only;
2619 ret = cma_get_port(id_priv);
2625 if (id_priv->cma_dev)
2626 cma_release_dev(id_priv);
2628 cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE);
2631 EXPORT_SYMBOL(rdma_bind_addr);
2633 static int cma_format_hdr(void *hdr, struct rdma_id_private *id_priv)
2635 struct cma_hdr *cma_hdr;
2638 cma_hdr->cma_version = CMA_VERSION;
2639 if (cma_family(id_priv) == AF_INET) {
2640 struct sockaddr_in *src4, *dst4;
2642 src4 = (struct sockaddr_in *) cma_src_addr(id_priv);
2643 dst4 = (struct sockaddr_in *) cma_dst_addr(id_priv);
2645 cma_set_ip_ver(cma_hdr, 4);
2646 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2647 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2648 cma_hdr->port = src4->sin_port;
2649 } else if (cma_family(id_priv) == AF_INET6) {
2650 struct sockaddr_in6 *src6, *dst6;
2652 src6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
2653 dst6 = (struct sockaddr_in6 *) cma_dst_addr(id_priv);
2655 cma_set_ip_ver(cma_hdr, 6);
2656 cma_hdr->src_addr.ip6 = src6->sin6_addr;
2657 cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
2658 cma_hdr->port = src6->sin6_port;
2663 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
2664 struct ib_cm_event *ib_event)
2666 struct rdma_id_private *id_priv = cm_id->context;
2667 struct rdma_cm_event event;
2668 struct ib_cm_sidr_rep_event_param *rep = &ib_event->param.sidr_rep_rcvd;
2671 if (cma_disable_callback(id_priv, RDMA_CM_CONNECT))
2674 memset(&event, 0, sizeof event);
2675 switch (ib_event->event) {
2676 case IB_CM_SIDR_REQ_ERROR:
2677 event.event = RDMA_CM_EVENT_UNREACHABLE;
2678 event.status = -ETIMEDOUT;
2680 case IB_CM_SIDR_REP_RECEIVED:
2681 event.param.ud.private_data = ib_event->private_data;
2682 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
2683 if (rep->status != IB_SIDR_SUCCESS) {
2684 event.event = RDMA_CM_EVENT_UNREACHABLE;
2685 event.status = ib_event->param.sidr_rep_rcvd.status;
2688 ret = cma_set_qkey(id_priv, rep->qkey);
2690 event.event = RDMA_CM_EVENT_ADDR_ERROR;
2694 ib_init_ah_from_path(id_priv->id.device, id_priv->id.port_num,
2695 id_priv->id.route.path_rec,
2696 &event.param.ud.ah_attr);
2697 event.param.ud.qp_num = rep->qpn;
2698 event.param.ud.qkey = rep->qkey;
2699 event.event = RDMA_CM_EVENT_ESTABLISHED;
2703 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
2708 ret = id_priv->id.event_handler(&id_priv->id, &event);
2710 /* Destroy the CM ID by returning a non-zero value. */
2711 id_priv->cm_id.ib = NULL;
2712 cma_exch(id_priv, RDMA_CM_DESTROYING);
2713 mutex_unlock(&id_priv->handler_mutex);
2714 rdma_destroy_id(&id_priv->id);
2718 mutex_unlock(&id_priv->handler_mutex);
2722 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
2723 struct rdma_conn_param *conn_param)
2725 struct ib_cm_sidr_req_param req;
2726 struct ib_cm_id *id;
2730 memset(&req, 0, sizeof req);
2731 offset = cma_user_data_offset(id_priv);
2732 req.private_data_len = offset + conn_param->private_data_len;
2733 if (req.private_data_len < conn_param->private_data_len)
2736 if (req.private_data_len) {
2737 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2741 private_data = NULL;
2744 if (conn_param->private_data && conn_param->private_data_len)
2745 memcpy(private_data + offset, conn_param->private_data,
2746 conn_param->private_data_len);
2749 ret = cma_format_hdr(private_data, id_priv);
2752 req.private_data = private_data;
2755 id = ib_create_cm_id(id_priv->id.device, cma_sidr_rep_handler,
2761 id_priv->cm_id.ib = id;
2763 req.path = id_priv->id.route.path_rec;
2764 req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
2765 req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
2766 req.max_cm_retries = CMA_MAX_CM_RETRIES;
2768 ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
2770 ib_destroy_cm_id(id_priv->cm_id.ib);
2771 id_priv->cm_id.ib = NULL;
2774 kfree(private_data);
2778 static int cma_connect_ib(struct rdma_id_private *id_priv,
2779 struct rdma_conn_param *conn_param)
2781 struct ib_cm_req_param req;
2782 struct rdma_route *route;
2784 struct ib_cm_id *id;
2787 memset(&req, 0, sizeof req);
2788 offset = cma_user_data_offset(id_priv);
2789 req.private_data_len = offset + conn_param->private_data_len;
2790 if (req.private_data_len < conn_param->private_data_len)
2793 if (req.private_data_len) {
2794 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2798 private_data = NULL;
2801 if (conn_param->private_data && conn_param->private_data_len)
2802 memcpy(private_data + offset, conn_param->private_data,
2803 conn_param->private_data_len);
2805 id = ib_create_cm_id(id_priv->id.device, cma_ib_handler, id_priv);
2810 id_priv->cm_id.ib = id;
2812 route = &id_priv->id.route;
2814 ret = cma_format_hdr(private_data, id_priv);
2817 req.private_data = private_data;
2820 req.primary_path = &route->path_rec[0];
2821 if (route->num_paths == 2)
2822 req.alternate_path = &route->path_rec[1];
2824 req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
2825 req.qp_num = id_priv->qp_num;
2826 req.qp_type = id_priv->id.qp_type;
2827 req.starting_psn = id_priv->seq_num;
2828 req.responder_resources = conn_param->responder_resources;
2829 req.initiator_depth = conn_param->initiator_depth;
2830 req.flow_control = conn_param->flow_control;
2831 req.retry_count = min_t(u8, 7, conn_param->retry_count);
2832 req.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
2833 req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2834 req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2835 req.max_cm_retries = CMA_MAX_CM_RETRIES;
2836 req.srq = id_priv->srq ? 1 : 0;
2838 ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
2840 if (ret && !IS_ERR(id)) {
2841 ib_destroy_cm_id(id);
2842 id_priv->cm_id.ib = NULL;
2845 kfree(private_data);
2849 static int cma_connect_iw(struct rdma_id_private *id_priv,
2850 struct rdma_conn_param *conn_param)
2852 struct iw_cm_id *cm_id;
2854 struct iw_cm_conn_param iw_param;
2856 cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
2858 return PTR_ERR(cm_id);
2860 id_priv->cm_id.iw = cm_id;
2862 memcpy(&cm_id->local_addr, cma_src_addr(id_priv),
2863 rdma_addr_size(cma_src_addr(id_priv)));
2864 memcpy(&cm_id->remote_addr, cma_dst_addr(id_priv),
2865 rdma_addr_size(cma_dst_addr(id_priv)));
2867 ret = cma_modify_qp_rtr(id_priv, conn_param);
2872 iw_param.ord = conn_param->initiator_depth;
2873 iw_param.ird = conn_param->responder_resources;
2874 iw_param.private_data = conn_param->private_data;
2875 iw_param.private_data_len = conn_param->private_data_len;
2876 iw_param.qpn = id_priv->id.qp ? id_priv->qp_num : conn_param->qp_num;
2878 memset(&iw_param, 0, sizeof iw_param);
2879 iw_param.qpn = id_priv->qp_num;
2881 ret = iw_cm_connect(cm_id, &iw_param);
2884 iw_destroy_cm_id(cm_id);
2885 id_priv->cm_id.iw = NULL;
2890 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2892 struct rdma_id_private *id_priv;
2895 id_priv = container_of(id, struct rdma_id_private, id);
2896 if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT))
2900 id_priv->qp_num = conn_param->qp_num;
2901 id_priv->srq = conn_param->srq;
2904 switch (rdma_node_get_transport(id->device->node_type)) {
2905 case RDMA_TRANSPORT_IB:
2906 if (id->qp_type == IB_QPT_UD)
2907 ret = cma_resolve_ib_udp(id_priv, conn_param);
2909 ret = cma_connect_ib(id_priv, conn_param);
2911 case RDMA_TRANSPORT_IWARP:
2912 ret = cma_connect_iw(id_priv, conn_param);
2923 cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED);
2926 EXPORT_SYMBOL(rdma_connect);
2928 static int cma_accept_ib(struct rdma_id_private *id_priv,
2929 struct rdma_conn_param *conn_param)
2931 struct ib_cm_rep_param rep;
2934 ret = cma_modify_qp_rtr(id_priv, conn_param);
2938 ret = cma_modify_qp_rts(id_priv, conn_param);
2942 memset(&rep, 0, sizeof rep);
2943 rep.qp_num = id_priv->qp_num;
2944 rep.starting_psn = id_priv->seq_num;
2945 rep.private_data = conn_param->private_data;
2946 rep.private_data_len = conn_param->private_data_len;
2947 rep.responder_resources = conn_param->responder_resources;
2948 rep.initiator_depth = conn_param->initiator_depth;
2949 rep.failover_accepted = 0;
2950 rep.flow_control = conn_param->flow_control;
2951 rep.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
2952 rep.srq = id_priv->srq ? 1 : 0;
2954 ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
2959 static int cma_accept_iw(struct rdma_id_private *id_priv,
2960 struct rdma_conn_param *conn_param)
2962 struct iw_cm_conn_param iw_param;
2965 ret = cma_modify_qp_rtr(id_priv, conn_param);
2969 iw_param.ord = conn_param->initiator_depth;
2970 iw_param.ird = conn_param->responder_resources;
2971 iw_param.private_data = conn_param->private_data;
2972 iw_param.private_data_len = conn_param->private_data_len;
2973 if (id_priv->id.qp) {
2974 iw_param.qpn = id_priv->qp_num;
2976 iw_param.qpn = conn_param->qp_num;
2978 return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
2981 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
2982 enum ib_cm_sidr_status status, u32 qkey,
2983 const void *private_data, int private_data_len)
2985 struct ib_cm_sidr_rep_param rep;
2988 memset(&rep, 0, sizeof rep);
2989 rep.status = status;
2990 if (status == IB_SIDR_SUCCESS) {
2991 ret = cma_set_qkey(id_priv, qkey);
2994 rep.qp_num = id_priv->qp_num;
2995 rep.qkey = id_priv->qkey;
2997 rep.private_data = private_data;
2998 rep.private_data_len = private_data_len;
3000 return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
3003 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
3005 struct rdma_id_private *id_priv;
3008 id_priv = container_of(id, struct rdma_id_private, id);
3010 id_priv->owner = task_pid_nr(current);
3012 if (!cma_comp(id_priv, RDMA_CM_CONNECT))
3015 if (!id->qp && conn_param) {
3016 id_priv->qp_num = conn_param->qp_num;
3017 id_priv->srq = conn_param->srq;
3020 switch (rdma_node_get_transport(id->device->node_type)) {
3021 case RDMA_TRANSPORT_IB:
3022 if (id->qp_type == IB_QPT_UD) {
3024 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
3026 conn_param->private_data,
3027 conn_param->private_data_len);
3029 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
3033 ret = cma_accept_ib(id_priv, conn_param);
3035 ret = cma_rep_recv(id_priv);
3038 case RDMA_TRANSPORT_IWARP:
3039 ret = cma_accept_iw(id_priv, conn_param);
3051 cma_modify_qp_err(id_priv);
3052 rdma_reject(id, NULL, 0);
3055 EXPORT_SYMBOL(rdma_accept);
3057 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
3059 struct rdma_id_private *id_priv;
3062 id_priv = container_of(id, struct rdma_id_private, id);
3063 if (!id_priv->cm_id.ib)
3066 switch (id->device->node_type) {
3067 case RDMA_NODE_IB_CA:
3068 ret = ib_cm_notify(id_priv->cm_id.ib, event);
3076 EXPORT_SYMBOL(rdma_notify);
3078 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
3079 u8 private_data_len)
3081 struct rdma_id_private *id_priv;
3084 id_priv = container_of(id, struct rdma_id_private, id);
3085 if (!id_priv->cm_id.ib)
3088 switch (rdma_node_get_transport(id->device->node_type)) {
3089 case RDMA_TRANSPORT_IB:
3090 if (id->qp_type == IB_QPT_UD)
3091 ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT, 0,
3092 private_data, private_data_len);
3094 ret = ib_send_cm_rej(id_priv->cm_id.ib,
3095 IB_CM_REJ_CONSUMER_DEFINED, NULL,
3096 0, private_data, private_data_len);
3098 case RDMA_TRANSPORT_IWARP:
3099 ret = iw_cm_reject(id_priv->cm_id.iw,
3100 private_data, private_data_len);
3108 EXPORT_SYMBOL(rdma_reject);
3110 int rdma_disconnect(struct rdma_cm_id *id)
3112 struct rdma_id_private *id_priv;
3115 id_priv = container_of(id, struct rdma_id_private, id);
3116 if (!id_priv->cm_id.ib)
3119 switch (rdma_node_get_transport(id->device->node_type)) {
3120 case RDMA_TRANSPORT_IB:
3121 ret = cma_modify_qp_err(id_priv);
3124 /* Initiate or respond to a disconnect. */
3125 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
3126 ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
3128 case RDMA_TRANSPORT_IWARP:
3129 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
3138 EXPORT_SYMBOL(rdma_disconnect);
3140 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
3142 struct rdma_id_private *id_priv;
3143 struct cma_multicast *mc = multicast->context;
3144 struct rdma_cm_event event;
3147 id_priv = mc->id_priv;
3148 if (cma_disable_callback(id_priv, RDMA_CM_ADDR_BOUND) &&
3149 cma_disable_callback(id_priv, RDMA_CM_ADDR_RESOLVED))
3153 status = cma_set_qkey(id_priv, be32_to_cpu(multicast->rec.qkey));
3154 mutex_lock(&id_priv->qp_mutex);
3155 if (!status && id_priv->id.qp)
3156 status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid,
3157 be16_to_cpu(multicast->rec.mlid));
3158 mutex_unlock(&id_priv->qp_mutex);
3160 memset(&event, 0, sizeof event);
3161 event.status = status;
3162 event.param.ud.private_data = mc->context;
3164 event.event = RDMA_CM_EVENT_MULTICAST_JOIN;
3165 ib_init_ah_from_mcmember(id_priv->id.device,
3166 id_priv->id.port_num, &multicast->rec,
3167 &event.param.ud.ah_attr);
3168 event.param.ud.qp_num = 0xFFFFFF;
3169 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
3171 event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
3173 ret = id_priv->id.event_handler(&id_priv->id, &event);
3175 cma_exch(id_priv, RDMA_CM_DESTROYING);
3176 mutex_unlock(&id_priv->handler_mutex);
3177 rdma_destroy_id(&id_priv->id);
3181 mutex_unlock(&id_priv->handler_mutex);
3185 static void cma_set_mgid(struct rdma_id_private *id_priv,
3186 struct sockaddr *addr, union ib_gid *mgid)
3188 unsigned char mc_map[MAX_ADDR_LEN];
3189 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3190 struct sockaddr_in *sin = (struct sockaddr_in *) addr;
3191 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
3193 if (cma_any_addr(addr)) {
3194 memset(mgid, 0, sizeof *mgid);
3195 } else if ((addr->sa_family == AF_INET6) &&
3196 ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
3198 /* IPv6 address is an SA assigned MGID. */
3199 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3200 } else if (addr->sa_family == AF_IB) {
3201 memcpy(mgid, &((struct sockaddr_ib *) addr)->sib_addr, sizeof *mgid);
3202 } else if ((addr->sa_family == AF_INET6)) {
3203 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
3204 if (id_priv->id.ps == RDMA_PS_UDP)
3205 mc_map[7] = 0x01; /* Use RDMA CM signature */
3206 *mgid = *(union ib_gid *) (mc_map + 4);
3208 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
3209 if (id_priv->id.ps == RDMA_PS_UDP)
3210 mc_map[7] = 0x01; /* Use RDMA CM signature */
3211 *mgid = *(union ib_gid *) (mc_map + 4);
3215 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
3216 struct cma_multicast *mc)
3218 struct ib_sa_mcmember_rec rec;
3219 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3220 ib_sa_comp_mask comp_mask;
3223 ib_addr_get_mgid(dev_addr, &rec.mgid);
3224 ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
3229 ret = cma_set_qkey(id_priv, 0);
3233 cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
3234 rec.qkey = cpu_to_be32(id_priv->qkey);
3235 rdma_addr_get_sgid(dev_addr, &rec.port_gid);
3236 rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
3239 comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
3240 IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
3241 IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
3242 IB_SA_MCMEMBER_REC_FLOW_LABEL |
3243 IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
3245 if (id_priv->id.ps == RDMA_PS_IPOIB)
3246 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
3247 IB_SA_MCMEMBER_REC_RATE_SELECTOR |
3248 IB_SA_MCMEMBER_REC_MTU_SELECTOR |
3249 IB_SA_MCMEMBER_REC_MTU |
3250 IB_SA_MCMEMBER_REC_HOP_LIMIT;
3252 mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device,
3253 id_priv->id.port_num, &rec,
3254 comp_mask, GFP_KERNEL,
3255 cma_ib_mc_handler, mc);
3256 return PTR_ERR_OR_ZERO(mc->multicast.ib);
3259 static void iboe_mcast_work_handler(struct work_struct *work)
3261 struct iboe_mcast_work *mw = container_of(work, struct iboe_mcast_work, work);
3262 struct cma_multicast *mc = mw->mc;
3263 struct ib_sa_multicast *m = mc->multicast.ib;
3265 mc->multicast.ib->context = mc;
3266 cma_ib_mc_handler(0, m);
3267 kref_put(&mc->mcref, release_mc);
3271 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid)
3273 struct sockaddr_in *sin = (struct sockaddr_in *)addr;
3274 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
3276 if (cma_any_addr(addr)) {
3277 memset(mgid, 0, sizeof *mgid);
3278 } else if (addr->sa_family == AF_INET6) {
3279 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3281 mgid->raw[0] = 0xff;
3282 mgid->raw[1] = 0x0e;
3291 mgid->raw[10] = 0xff;
3292 mgid->raw[11] = 0xff;
3293 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
3297 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
3298 struct cma_multicast *mc)
3300 struct iboe_mcast_work *work;
3301 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3303 struct sockaddr *addr = (struct sockaddr *)&mc->addr;
3304 struct net_device *ndev = NULL;
3306 if (cma_zero_addr((struct sockaddr *)&mc->addr))
3309 work = kzalloc(sizeof *work, GFP_KERNEL);
3313 mc->multicast.ib = kzalloc(sizeof(struct ib_sa_multicast), GFP_KERNEL);
3314 if (!mc->multicast.ib) {
3319 cma_iboe_set_mgid(addr, &mc->multicast.ib->rec.mgid);
3321 mc->multicast.ib->rec.pkey = cpu_to_be16(0xffff);
3322 if (id_priv->id.ps == RDMA_PS_UDP)
3323 mc->multicast.ib->rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
3325 if (dev_addr->bound_dev_if)
3326 ndev = dev_get_by_index(&init_net, dev_addr->bound_dev_if);
3331 mc->multicast.ib->rec.rate = iboe_get_rate(ndev);
3332 mc->multicast.ib->rec.hop_limit = 1;
3333 mc->multicast.ib->rec.mtu = iboe_get_mtu(ndev->mtu);
3335 if (!mc->multicast.ib->rec.mtu) {
3339 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
3340 &mc->multicast.ib->rec.port_gid);
3343 INIT_WORK(&work->work, iboe_mcast_work_handler);
3344 kref_get(&mc->mcref);
3345 queue_work(cma_wq, &work->work);
3350 kfree(mc->multicast.ib);
3356 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
3359 struct rdma_id_private *id_priv;
3360 struct cma_multicast *mc;
3363 id_priv = container_of(id, struct rdma_id_private, id);
3364 if (!cma_comp(id_priv, RDMA_CM_ADDR_BOUND) &&
3365 !cma_comp(id_priv, RDMA_CM_ADDR_RESOLVED))
3368 mc = kmalloc(sizeof *mc, GFP_KERNEL);
3372 memcpy(&mc->addr, addr, rdma_addr_size(addr));
3373 mc->context = context;
3374 mc->id_priv = id_priv;
3376 spin_lock(&id_priv->lock);
3377 list_add(&mc->list, &id_priv->mc_list);
3378 spin_unlock(&id_priv->lock);
3380 switch (rdma_node_get_transport(id->device->node_type)) {
3381 case RDMA_TRANSPORT_IB:
3382 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
3383 case IB_LINK_LAYER_INFINIBAND:
3384 ret = cma_join_ib_multicast(id_priv, mc);
3386 case IB_LINK_LAYER_ETHERNET:
3387 kref_init(&mc->mcref);
3388 ret = cma_iboe_join_multicast(id_priv, mc);
3400 spin_lock_irq(&id_priv->lock);
3401 list_del(&mc->list);
3402 spin_unlock_irq(&id_priv->lock);
3407 EXPORT_SYMBOL(rdma_join_multicast);
3409 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
3411 struct rdma_id_private *id_priv;
3412 struct cma_multicast *mc;
3414 id_priv = container_of(id, struct rdma_id_private, id);
3415 spin_lock_irq(&id_priv->lock);
3416 list_for_each_entry(mc, &id_priv->mc_list, list) {
3417 if (!memcmp(&mc->addr, addr, rdma_addr_size(addr))) {
3418 list_del(&mc->list);
3419 spin_unlock_irq(&id_priv->lock);
3422 ib_detach_mcast(id->qp,
3423 &mc->multicast.ib->rec.mgid,
3424 be16_to_cpu(mc->multicast.ib->rec.mlid));
3425 if (rdma_node_get_transport(id_priv->cma_dev->device->node_type) == RDMA_TRANSPORT_IB) {
3426 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
3427 case IB_LINK_LAYER_INFINIBAND:
3428 ib_sa_free_multicast(mc->multicast.ib);
3431 case IB_LINK_LAYER_ETHERNET:
3432 kref_put(&mc->mcref, release_mc);
3441 spin_unlock_irq(&id_priv->lock);
3443 EXPORT_SYMBOL(rdma_leave_multicast);
3445 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
3447 struct rdma_dev_addr *dev_addr;
3448 struct cma_ndev_work *work;
3450 dev_addr = &id_priv->id.route.addr.dev_addr;
3452 if ((dev_addr->bound_dev_if == ndev->ifindex) &&
3453 memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
3454 printk(KERN_INFO "RDMA CM addr change for ndev %s used by id %p\n",
3455 ndev->name, &id_priv->id);
3456 work = kzalloc(sizeof *work, GFP_KERNEL);
3460 INIT_WORK(&work->work, cma_ndev_work_handler);
3462 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
3463 atomic_inc(&id_priv->refcount);
3464 queue_work(cma_wq, &work->work);
3470 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
3473 struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
3474 struct cma_device *cma_dev;
3475 struct rdma_id_private *id_priv;
3476 int ret = NOTIFY_DONE;
3478 if (dev_net(ndev) != &init_net)
3481 if (event != NETDEV_BONDING_FAILOVER)
3484 if (!(ndev->flags & IFF_MASTER) || !(ndev->priv_flags & IFF_BONDING))
3488 list_for_each_entry(cma_dev, &dev_list, list)
3489 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
3490 ret = cma_netdev_change(ndev, id_priv);
3496 mutex_unlock(&lock);
3500 static struct notifier_block cma_nb = {
3501 .notifier_call = cma_netdev_callback
3504 static void cma_add_one(struct ib_device *device)
3506 struct cma_device *cma_dev;
3507 struct rdma_id_private *id_priv;
3509 cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
3513 cma_dev->device = device;
3515 init_completion(&cma_dev->comp);
3516 atomic_set(&cma_dev->refcount, 1);
3517 INIT_LIST_HEAD(&cma_dev->id_list);
3518 ib_set_client_data(device, &cma_client, cma_dev);
3521 list_add_tail(&cma_dev->list, &dev_list);
3522 list_for_each_entry(id_priv, &listen_any_list, list)
3523 cma_listen_on_dev(id_priv, cma_dev);
3524 mutex_unlock(&lock);
3527 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
3529 struct rdma_cm_event event;
3530 enum rdma_cm_state state;
3533 /* Record that we want to remove the device */
3534 state = cma_exch(id_priv, RDMA_CM_DEVICE_REMOVAL);
3535 if (state == RDMA_CM_DESTROYING)
3538 cma_cancel_operation(id_priv, state);
3539 mutex_lock(&id_priv->handler_mutex);
3541 /* Check for destruction from another callback. */
3542 if (!cma_comp(id_priv, RDMA_CM_DEVICE_REMOVAL))
3545 memset(&event, 0, sizeof event);
3546 event.event = RDMA_CM_EVENT_DEVICE_REMOVAL;
3547 ret = id_priv->id.event_handler(&id_priv->id, &event);
3549 mutex_unlock(&id_priv->handler_mutex);
3553 static void cma_process_remove(struct cma_device *cma_dev)
3555 struct rdma_id_private *id_priv;
3559 while (!list_empty(&cma_dev->id_list)) {
3560 id_priv = list_entry(cma_dev->id_list.next,
3561 struct rdma_id_private, list);
3563 list_del(&id_priv->listen_list);
3564 list_del_init(&id_priv->list);
3565 atomic_inc(&id_priv->refcount);
3566 mutex_unlock(&lock);
3568 ret = id_priv->internal_id ? 1 : cma_remove_id_dev(id_priv);
3569 cma_deref_id(id_priv);
3571 rdma_destroy_id(&id_priv->id);
3575 mutex_unlock(&lock);
3577 cma_deref_dev(cma_dev);
3578 wait_for_completion(&cma_dev->comp);
3581 static void cma_remove_one(struct ib_device *device)
3583 struct cma_device *cma_dev;
3585 cma_dev = ib_get_client_data(device, &cma_client);
3590 list_del(&cma_dev->list);
3591 mutex_unlock(&lock);
3593 cma_process_remove(cma_dev);
3597 static int cma_get_id_stats(struct sk_buff *skb, struct netlink_callback *cb)
3599 struct nlmsghdr *nlh;
3600 struct rdma_cm_id_stats *id_stats;
3601 struct rdma_id_private *id_priv;
3602 struct rdma_cm_id *id = NULL;
3603 struct cma_device *cma_dev;
3604 int i_dev = 0, i_id = 0;
3607 * We export all of the IDs as a sequence of messages. Each
3608 * ID gets its own netlink message.
3612 list_for_each_entry(cma_dev, &dev_list, list) {
3613 if (i_dev < cb->args[0]) {
3619 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
3620 if (i_id < cb->args[1]) {
3625 id_stats = ibnl_put_msg(skb, &nlh, cb->nlh->nlmsg_seq,
3626 sizeof *id_stats, RDMA_NL_RDMA_CM,
3627 RDMA_NL_RDMA_CM_ID_STATS,
3632 memset(id_stats, 0, sizeof *id_stats);
3634 id_stats->node_type = id->route.addr.dev_addr.dev_type;
3635 id_stats->port_num = id->port_num;
3636 id_stats->bound_dev_if =
3637 id->route.addr.dev_addr.bound_dev_if;
3639 if (ibnl_put_attr(skb, nlh,
3640 rdma_addr_size(cma_src_addr(id_priv)),
3641 cma_src_addr(id_priv),
3642 RDMA_NL_RDMA_CM_ATTR_SRC_ADDR))
3644 if (ibnl_put_attr(skb, nlh,
3645 rdma_addr_size(cma_src_addr(id_priv)),
3646 cma_dst_addr(id_priv),
3647 RDMA_NL_RDMA_CM_ATTR_DST_ADDR))
3650 id_stats->pid = id_priv->owner;
3651 id_stats->port_space = id->ps;
3652 id_stats->cm_state = id_priv->state;
3653 id_stats->qp_num = id_priv->qp_num;
3654 id_stats->qp_type = id->qp_type;
3664 mutex_unlock(&lock);
3665 cb->args[0] = i_dev;
3671 static const struct ibnl_client_cbs cma_cb_table[] = {
3672 [RDMA_NL_RDMA_CM_ID_STATS] = { .dump = cma_get_id_stats,
3673 .module = THIS_MODULE },
3676 static int __init cma_init(void)
3680 cma_wq = create_singlethread_workqueue("rdma_cm");
3684 ib_sa_register_client(&sa_client);
3685 rdma_addr_register_client(&addr_client);
3686 register_netdevice_notifier(&cma_nb);
3688 ret = ib_register_client(&cma_client);
3692 if (ibnl_add_client(RDMA_NL_RDMA_CM, RDMA_NL_RDMA_CM_NUM_OPS, cma_cb_table))
3693 printk(KERN_WARNING "RDMA CMA: failed to add netlink callback\n");
3698 unregister_netdevice_notifier(&cma_nb);
3699 rdma_addr_unregister_client(&addr_client);
3700 ib_sa_unregister_client(&sa_client);
3701 destroy_workqueue(cma_wq);
3705 static void __exit cma_cleanup(void)
3707 ibnl_remove_client(RDMA_NL_RDMA_CM);
3708 ib_unregister_client(&cma_client);
3709 unregister_netdevice_notifier(&cma_nb);
3710 rdma_addr_unregister_client(&addr_client);
3711 ib_sa_unregister_client(&sa_client);
3712 destroy_workqueue(cma_wq);
3713 idr_destroy(&tcp_ps);
3714 idr_destroy(&udp_ps);
3715 idr_destroy(&ipoib_ps);
3716 idr_destroy(&ib_ps);
3719 module_init(cma_init);
3720 module_exit(cma_cleanup);