2 * SPDX-License-Identifier: BSD-2-Clause OR GPL-2.0
4 * Copyright (c) 2005 Voltaire Inc. All rights reserved.
5 * Copyright (c) 2002-2005, Network Appliance, Inc. All rights reserved.
6 * Copyright (c) 1999-2005, Mellanox Technologies, Inc. All rights reserved.
7 * Copyright (c) 2005-2006 Intel Corporation. All rights reserved.
9 * This software is available to you under a choice of one of two
10 * licenses. You may choose to be licensed under the terms of the GNU
11 * General Public License (GPL) Version 2, available from the file
12 * COPYING in the main directory of this source tree, or the
13 * OpenIB.org BSD license below:
15 * Redistribution and use in source and binary forms, with or
16 * without modification, are permitted provided that the following
19 * - Redistributions of source code must retain the above
20 * copyright notice, this list of conditions and the following
23 * - Redistributions in binary form must reproduce the above
24 * copyright notice, this list of conditions and the following
25 * disclaimer in the documentation and/or other materials
26 * provided with the distribution.
28 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
29 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
30 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
31 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
32 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
33 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
34 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
40 #define LINUXKPI_PARAM_PREFIX ibcore_
42 #include <linux/completion.h>
44 #include <linux/in6.h>
45 #include <linux/mutex.h>
46 #include <linux/random.h>
47 #include <linux/idr.h>
48 #include <linux/inetdevice.h>
49 #include <linux/slab.h>
50 #include <linux/module.h>
51 #include <net/route.h>
56 #include <netinet6/scope6_var.h>
57 #include <netinet6/ip6_var.h>
59 #include <rdma/rdma_cm.h>
60 #include <rdma/rdma_cm_ib.h>
62 #include <rdma/ib_addr.h>
63 #include <rdma/ib_cache.h>
64 #include <rdma/ib_cm.h>
65 #include <rdma/ib_sa.h>
66 #include <rdma/iw_cm.h>
70 #include "core_priv.h"
72 MODULE_AUTHOR("Sean Hefty");
73 MODULE_DESCRIPTION("Generic RDMA CM Agent");
74 MODULE_LICENSE("Dual BSD/GPL");
76 #define CMA_CM_RESPONSE_TIMEOUT 20
77 #define CMA_QUERY_CLASSPORT_INFO_TIMEOUT 3000
78 #define CMA_MAX_CM_RETRIES 15
79 #define CMA_CM_MRA_SETTING (IB_CM_MRA_FLAG_DELAY | 24)
80 #define CMA_IBOE_PACKET_LIFETIME 18
82 static const char * const cma_events[] = {
83 [RDMA_CM_EVENT_ADDR_RESOLVED] = "address resolved",
84 [RDMA_CM_EVENT_ADDR_ERROR] = "address error",
85 [RDMA_CM_EVENT_ROUTE_RESOLVED] = "route resolved ",
86 [RDMA_CM_EVENT_ROUTE_ERROR] = "route error",
87 [RDMA_CM_EVENT_CONNECT_REQUEST] = "connect request",
88 [RDMA_CM_EVENT_CONNECT_RESPONSE] = "connect response",
89 [RDMA_CM_EVENT_CONNECT_ERROR] = "connect error",
90 [RDMA_CM_EVENT_UNREACHABLE] = "unreachable",
91 [RDMA_CM_EVENT_REJECTED] = "rejected",
92 [RDMA_CM_EVENT_ESTABLISHED] = "established",
93 [RDMA_CM_EVENT_DISCONNECTED] = "disconnected",
94 [RDMA_CM_EVENT_DEVICE_REMOVAL] = "device removal",
95 [RDMA_CM_EVENT_MULTICAST_JOIN] = "multicast join",
96 [RDMA_CM_EVENT_MULTICAST_ERROR] = "multicast error",
97 [RDMA_CM_EVENT_ADDR_CHANGE] = "address change",
98 [RDMA_CM_EVENT_TIMEWAIT_EXIT] = "timewait exit",
101 const char *__attribute_const__ rdma_event_msg(enum rdma_cm_event_type event)
103 size_t index = event;
105 return (index < ARRAY_SIZE(cma_events) && cma_events[index]) ?
106 cma_events[index] : "unrecognized event";
108 EXPORT_SYMBOL(rdma_event_msg);
110 static int cma_check_linklocal(struct rdma_dev_addr *, struct sockaddr *);
111 static void cma_add_one(struct ib_device *device);
112 static void cma_remove_one(struct ib_device *device, void *client_data);
114 static struct ib_client cma_client = {
117 .remove = cma_remove_one
120 static struct ib_sa_client sa_client;
121 static struct rdma_addr_client addr_client;
122 static LIST_HEAD(dev_list);
123 static LIST_HEAD(listen_any_list);
124 static DEFINE_MUTEX(lock);
125 static struct workqueue_struct *cma_wq;
134 VNET_DEFINE(struct cma_pernet, cma_pernet);
136 static struct cma_pernet *cma_pernet_ptr(struct vnet *vnet)
138 struct cma_pernet *retval;
140 CURVNET_SET_QUIET(vnet);
141 retval = &VNET(cma_pernet);
147 static struct idr *cma_pernet_idr(struct vnet *net, enum rdma_port_space ps)
149 struct cma_pernet *pernet = cma_pernet_ptr(net);
153 return &pernet->tcp_ps;
155 return &pernet->udp_ps;
157 return &pernet->ipoib_ps;
159 return &pernet->ib_ps;
166 struct list_head list;
167 struct ib_device *device;
168 struct completion comp;
170 struct list_head id_list;
171 struct sysctl_ctx_list sysctl_ctx;
172 enum ib_gid_type *default_gid_type;
175 struct rdma_bind_list {
176 enum rdma_port_space ps;
177 struct hlist_head owners;
181 struct class_port_info_context {
182 struct ib_class_port_info *class_port_info;
183 struct ib_device *device;
184 struct completion done;
185 struct ib_sa_query *sa_query;
189 static int cma_ps_alloc(struct vnet *vnet, enum rdma_port_space ps,
190 struct rdma_bind_list *bind_list, int snum)
192 struct idr *idr = cma_pernet_idr(vnet, ps);
194 return idr_alloc(idr, bind_list, snum, snum + 1, GFP_KERNEL);
197 static struct rdma_bind_list *cma_ps_find(struct vnet *net,
198 enum rdma_port_space ps, int snum)
200 struct idr *idr = cma_pernet_idr(net, ps);
202 return idr_find(idr, snum);
205 static void cma_ps_remove(struct vnet *net, enum rdma_port_space ps, int snum)
207 struct idr *idr = cma_pernet_idr(net, ps);
209 idr_remove(idr, snum);
216 void cma_ref_dev(struct cma_device *cma_dev)
218 atomic_inc(&cma_dev->refcount);
221 struct cma_device *cma_enum_devices_by_ibdev(cma_device_filter filter,
224 struct cma_device *cma_dev;
225 struct cma_device *found_cma_dev = NULL;
229 list_for_each_entry(cma_dev, &dev_list, list)
230 if (filter(cma_dev->device, cookie)) {
231 found_cma_dev = cma_dev;
236 cma_ref_dev(found_cma_dev);
238 return found_cma_dev;
241 int cma_get_default_gid_type(struct cma_device *cma_dev,
244 if (port < rdma_start_port(cma_dev->device) ||
245 port > rdma_end_port(cma_dev->device))
248 return cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)];
251 int cma_set_default_gid_type(struct cma_device *cma_dev,
253 enum ib_gid_type default_gid_type)
255 unsigned long supported_gids;
257 if (port < rdma_start_port(cma_dev->device) ||
258 port > rdma_end_port(cma_dev->device))
261 supported_gids = roce_gid_type_mask_support(cma_dev->device, port);
263 if (!(supported_gids & 1 << default_gid_type))
266 cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)] =
272 struct ib_device *cma_get_ib_dev(struct cma_device *cma_dev)
274 return cma_dev->device;
278 * Device removal can occur at anytime, so we need extra handling to
279 * serialize notifying the user of device removal with other callbacks.
280 * We do this by disabling removal notification while a callback is in process,
281 * and reporting it after the callback completes.
283 struct rdma_id_private {
284 struct rdma_cm_id id;
286 struct rdma_bind_list *bind_list;
287 struct hlist_node node;
288 struct list_head list; /* listen_any_list or cma_device.list */
289 struct list_head listen_list; /* per device listens */
290 struct cma_device *cma_dev;
291 struct list_head mc_list;
294 enum rdma_cm_state state;
296 struct mutex qp_mutex;
298 struct completion comp;
300 struct mutex handler_mutex;
304 struct ib_sa_query *query;
320 enum ib_gid_type gid_type;
323 struct cma_multicast {
324 struct rdma_id_private *id_priv;
326 struct ib_sa_multicast *ib;
328 struct list_head list;
330 struct sockaddr_storage addr;
337 struct work_struct work;
338 struct rdma_id_private *id;
339 enum rdma_cm_state old_state;
340 enum rdma_cm_state new_state;
341 struct rdma_cm_event event;
344 struct cma_ndev_work {
345 struct work_struct work;
346 struct rdma_id_private *id;
347 struct rdma_cm_event event;
350 struct iboe_mcast_work {
351 struct work_struct work;
352 struct rdma_id_private *id;
353 struct cma_multicast *mc;
366 u8 ip_version; /* IP version: 7:4 */
368 union cma_ip_addr src_addr;
369 union cma_ip_addr dst_addr;
372 #define CMA_VERSION 0x00
374 struct cma_req_info {
375 struct ib_device *device;
377 union ib_gid local_gid;
383 static int cma_comp(struct rdma_id_private *id_priv, enum rdma_cm_state comp)
388 spin_lock_irqsave(&id_priv->lock, flags);
389 ret = (id_priv->state == comp);
390 spin_unlock_irqrestore(&id_priv->lock, flags);
394 static int cma_comp_exch(struct rdma_id_private *id_priv,
395 enum rdma_cm_state comp, enum rdma_cm_state exch)
400 spin_lock_irqsave(&id_priv->lock, flags);
401 if ((ret = (id_priv->state == comp)))
402 id_priv->state = exch;
403 spin_unlock_irqrestore(&id_priv->lock, flags);
407 static enum rdma_cm_state cma_exch(struct rdma_id_private *id_priv,
408 enum rdma_cm_state exch)
411 enum rdma_cm_state old;
413 spin_lock_irqsave(&id_priv->lock, flags);
414 old = id_priv->state;
415 id_priv->state = exch;
416 spin_unlock_irqrestore(&id_priv->lock, flags);
420 static inline u8 cma_get_ip_ver(const struct cma_hdr *hdr)
422 return hdr->ip_version >> 4;
425 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
427 hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
430 static void _cma_attach_to_dev(struct rdma_id_private *id_priv,
431 struct cma_device *cma_dev)
433 cma_ref_dev(cma_dev);
434 id_priv->cma_dev = cma_dev;
435 id_priv->gid_type = 0;
436 id_priv->id.device = cma_dev->device;
437 id_priv->id.route.addr.dev_addr.transport =
438 rdma_node_get_transport(cma_dev->device->node_type);
439 list_add_tail(&id_priv->list, &cma_dev->id_list);
442 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
443 struct cma_device *cma_dev)
445 _cma_attach_to_dev(id_priv, cma_dev);
447 cma_dev->default_gid_type[id_priv->id.port_num -
448 rdma_start_port(cma_dev->device)];
451 void cma_deref_dev(struct cma_device *cma_dev)
453 if (atomic_dec_and_test(&cma_dev->refcount))
454 complete(&cma_dev->comp);
457 static inline void release_mc(struct kref *kref)
459 struct cma_multicast *mc = container_of(kref, struct cma_multicast, mcref);
461 kfree(mc->multicast.ib);
465 static void cma_release_dev(struct rdma_id_private *id_priv)
468 list_del(&id_priv->list);
469 cma_deref_dev(id_priv->cma_dev);
470 id_priv->cma_dev = NULL;
474 static inline struct sockaddr *cma_src_addr(struct rdma_id_private *id_priv)
476 return (struct sockaddr *) &id_priv->id.route.addr.src_addr;
479 static inline struct sockaddr *cma_dst_addr(struct rdma_id_private *id_priv)
481 return (struct sockaddr *) &id_priv->id.route.addr.dst_addr;
484 static inline unsigned short cma_family(struct rdma_id_private *id_priv)
486 return id_priv->id.route.addr.src_addr.ss_family;
489 static int cma_set_qkey(struct rdma_id_private *id_priv, u32 qkey)
491 struct ib_sa_mcmember_rec rec;
495 if (qkey && id_priv->qkey != qkey)
501 id_priv->qkey = qkey;
505 switch (id_priv->id.ps) {
508 id_priv->qkey = RDMA_UDP_QKEY;
511 ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid);
512 ret = ib_sa_get_mcmember_rec(id_priv->id.device,
513 id_priv->id.port_num, &rec.mgid,
516 id_priv->qkey = be32_to_cpu(rec.qkey);
524 static void cma_translate_ib(struct sockaddr_ib *sib, struct rdma_dev_addr *dev_addr)
526 dev_addr->dev_type = ARPHRD_INFINIBAND;
527 rdma_addr_set_sgid(dev_addr, (union ib_gid *) &sib->sib_addr);
528 ib_addr_set_pkey(dev_addr, ntohs(sib->sib_pkey));
531 static int cma_translate_addr(struct sockaddr *addr, struct rdma_dev_addr *dev_addr)
535 if (addr->sa_family != AF_IB) {
536 ret = rdma_translate_ip(addr, dev_addr);
538 cma_translate_ib((struct sockaddr_ib *) addr, dev_addr);
545 static inline int cma_validate_port(struct ib_device *device, u8 port,
546 enum ib_gid_type gid_type,
547 union ib_gid *gid, int dev_type,
552 struct net_device *ndev = NULL;
554 if ((dev_type == ARPHRD_INFINIBAND) && !rdma_protocol_ib(device, port))
557 if ((dev_type != ARPHRD_INFINIBAND) && rdma_protocol_ib(device, port))
560 if (dev_type == ARPHRD_ETHER && rdma_protocol_roce(device, port)) {
561 ndev = dev_get_by_index(net, bound_if_index);
562 if (ndev && ndev->if_flags & IFF_LOOPBACK) {
563 pr_info("detected loopback device\n");
566 if (!device->get_netdev)
569 ndev = device->get_netdev(device, port);
574 gid_type = IB_GID_TYPE_IB;
577 ret = ib_find_cached_gid_by_port(device, gid, gid_type, port,
586 static int cma_acquire_dev(struct rdma_id_private *id_priv,
587 struct rdma_id_private *listen_id_priv)
589 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
590 struct cma_device *cma_dev;
591 union ib_gid gid, iboe_gid, *gidp;
595 if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
596 id_priv->id.ps == RDMA_PS_IPOIB)
600 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
603 memcpy(&gid, dev_addr->src_dev_addr +
604 rdma_addr_gid_offset(dev_addr), sizeof gid);
606 if (listen_id_priv) {
607 cma_dev = listen_id_priv->cma_dev;
608 port = listen_id_priv->id.port_num;
609 gidp = rdma_protocol_roce(cma_dev->device, port) ?
612 ret = cma_validate_port(cma_dev->device, port,
613 rdma_protocol_ib(cma_dev->device, port) ?
615 listen_id_priv->gid_type, gidp,
618 dev_addr->bound_dev_if);
620 id_priv->id.port_num = port;
625 list_for_each_entry(cma_dev, &dev_list, list) {
626 for (port = 1; port <= cma_dev->device->phys_port_cnt; ++port) {
627 if (listen_id_priv &&
628 listen_id_priv->cma_dev == cma_dev &&
629 listen_id_priv->id.port_num == port)
632 gidp = rdma_protocol_roce(cma_dev->device, port) ?
635 ret = cma_validate_port(cma_dev->device, port,
636 rdma_protocol_ib(cma_dev->device, port) ?
638 cma_dev->default_gid_type[port - 1],
639 gidp, dev_addr->dev_type,
641 dev_addr->bound_dev_if);
643 id_priv->id.port_num = port;
651 cma_attach_to_dev(id_priv, cma_dev);
658 * Select the source IB device and address to reach the destination IB address.
660 static int cma_resolve_ib_dev(struct rdma_id_private *id_priv)
662 struct cma_device *cma_dev, *cur_dev;
663 struct sockaddr_ib *addr;
664 union ib_gid gid, sgid, *dgid;
670 addr = (struct sockaddr_ib *) cma_dst_addr(id_priv);
671 dgid = (union ib_gid *) &addr->sib_addr;
672 pkey = ntohs(addr->sib_pkey);
674 list_for_each_entry(cur_dev, &dev_list, list) {
675 for (p = 1; p <= cur_dev->device->phys_port_cnt; ++p) {
676 if (!rdma_cap_af_ib(cur_dev->device, p))
679 if (ib_find_cached_pkey(cur_dev->device, p, pkey, &index))
682 for (i = 0; !ib_get_cached_gid(cur_dev->device, p, i,
685 if (!memcmp(&gid, dgid, sizeof(gid))) {
688 id_priv->id.port_num = p;
692 if (!cma_dev && (gid.global.subnet_prefix ==
693 dgid->global.subnet_prefix)) {
696 id_priv->id.port_num = p;
706 cma_attach_to_dev(id_priv, cma_dev);
707 addr = (struct sockaddr_ib *) cma_src_addr(id_priv);
708 memcpy(&addr->sib_addr, &sgid, sizeof sgid);
709 cma_translate_ib(addr, &id_priv->id.route.addr.dev_addr);
713 static void cma_deref_id(struct rdma_id_private *id_priv)
715 if (atomic_dec_and_test(&id_priv->refcount))
716 complete(&id_priv->comp);
719 struct rdma_cm_id *rdma_create_id(struct vnet *net,
720 rdma_cm_event_handler event_handler,
721 void *context, enum rdma_port_space ps,
722 enum ib_qp_type qp_type)
724 struct rdma_id_private *id_priv;
726 id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
728 return ERR_PTR(-ENOMEM);
730 id_priv->owner = task_pid_nr(current);
731 id_priv->state = RDMA_CM_IDLE;
732 id_priv->id.context = context;
733 id_priv->id.event_handler = event_handler;
735 id_priv->id.qp_type = qp_type;
736 spin_lock_init(&id_priv->lock);
737 mutex_init(&id_priv->qp_mutex);
738 init_completion(&id_priv->comp);
739 atomic_set(&id_priv->refcount, 1);
740 mutex_init(&id_priv->handler_mutex);
741 INIT_LIST_HEAD(&id_priv->listen_list);
742 INIT_LIST_HEAD(&id_priv->mc_list);
743 get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
744 id_priv->id.route.addr.dev_addr.net = TD_TO_VNET(curthread);
748 EXPORT_SYMBOL(rdma_create_id);
750 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
752 struct ib_qp_attr qp_attr;
753 int qp_attr_mask, ret;
755 qp_attr.qp_state = IB_QPS_INIT;
756 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
760 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
764 qp_attr.qp_state = IB_QPS_RTR;
765 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
769 qp_attr.qp_state = IB_QPS_RTS;
771 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
776 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
778 struct ib_qp_attr qp_attr;
779 int qp_attr_mask, ret;
781 qp_attr.qp_state = IB_QPS_INIT;
782 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
786 return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
789 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
790 struct ib_qp_init_attr *qp_init_attr)
792 struct rdma_id_private *id_priv;
796 id_priv = container_of(id, struct rdma_id_private, id);
797 if (id->device != pd->device)
800 qp_init_attr->port_num = id->port_num;
801 qp = ib_create_qp(pd, qp_init_attr);
805 if (id->qp_type == IB_QPT_UD)
806 ret = cma_init_ud_qp(id_priv, qp);
808 ret = cma_init_conn_qp(id_priv, qp);
813 id_priv->qp_num = qp->qp_num;
814 id_priv->srq = (qp->srq != NULL);
820 EXPORT_SYMBOL(rdma_create_qp);
822 void rdma_destroy_qp(struct rdma_cm_id *id)
824 struct rdma_id_private *id_priv;
826 id_priv = container_of(id, struct rdma_id_private, id);
827 mutex_lock(&id_priv->qp_mutex);
828 ib_destroy_qp(id_priv->id.qp);
829 id_priv->id.qp = NULL;
830 mutex_unlock(&id_priv->qp_mutex);
832 EXPORT_SYMBOL(rdma_destroy_qp);
834 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
835 struct rdma_conn_param *conn_param)
837 struct ib_qp_attr qp_attr;
838 int qp_attr_mask, ret;
841 mutex_lock(&id_priv->qp_mutex);
842 if (!id_priv->id.qp) {
847 /* Need to update QP attributes from default values. */
848 qp_attr.qp_state = IB_QPS_INIT;
849 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
853 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
857 qp_attr.qp_state = IB_QPS_RTR;
858 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
862 ret = ib_query_gid(id_priv->id.device, id_priv->id.port_num,
863 qp_attr.ah_attr.grh.sgid_index, &sgid, NULL);
867 BUG_ON(id_priv->cma_dev->device != id_priv->id.device);
870 qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
871 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
873 mutex_unlock(&id_priv->qp_mutex);
877 static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
878 struct rdma_conn_param *conn_param)
880 struct ib_qp_attr qp_attr;
881 int qp_attr_mask, ret;
883 mutex_lock(&id_priv->qp_mutex);
884 if (!id_priv->id.qp) {
889 qp_attr.qp_state = IB_QPS_RTS;
890 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
895 qp_attr.max_rd_atomic = conn_param->initiator_depth;
896 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
898 mutex_unlock(&id_priv->qp_mutex);
902 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
904 struct ib_qp_attr qp_attr;
907 mutex_lock(&id_priv->qp_mutex);
908 if (!id_priv->id.qp) {
913 qp_attr.qp_state = IB_QPS_ERR;
914 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
916 mutex_unlock(&id_priv->qp_mutex);
920 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
921 struct ib_qp_attr *qp_attr, int *qp_attr_mask)
923 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
927 if (rdma_cap_eth_ah(id_priv->id.device, id_priv->id.port_num))
930 pkey = ib_addr_get_pkey(dev_addr);
932 ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
933 pkey, &qp_attr->pkey_index);
937 qp_attr->port_num = id_priv->id.port_num;
938 *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
940 if (id_priv->id.qp_type == IB_QPT_UD) {
941 ret = cma_set_qkey(id_priv, 0);
945 qp_attr->qkey = id_priv->qkey;
946 *qp_attr_mask |= IB_QP_QKEY;
948 qp_attr->qp_access_flags = 0;
949 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
954 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
957 struct rdma_id_private *id_priv;
960 id_priv = container_of(id, struct rdma_id_private, id);
961 if (rdma_cap_ib_cm(id->device, id->port_num)) {
962 if (!id_priv->cm_id.ib || (id_priv->id.qp_type == IB_QPT_UD))
963 ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
965 ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
968 if (qp_attr->qp_state == IB_QPS_RTR)
969 qp_attr->rq_psn = id_priv->seq_num;
970 } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
971 if (!id_priv->cm_id.iw) {
972 qp_attr->qp_access_flags = 0;
973 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
975 ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
982 EXPORT_SYMBOL(rdma_init_qp_attr);
984 static inline int cma_zero_addr(struct sockaddr *addr)
986 switch (addr->sa_family) {
988 return ipv4_is_zeronet(((struct sockaddr_in *)addr)->sin_addr.s_addr);
990 return ipv6_addr_any(&((struct sockaddr_in6 *) addr)->sin6_addr);
992 return ib_addr_any(&((struct sockaddr_ib *) addr)->sib_addr);
998 static inline int cma_loopback_addr(struct sockaddr *addr)
1000 switch (addr->sa_family) {
1002 return ipv4_is_loopback(((struct sockaddr_in *) addr)->sin_addr.s_addr);
1004 return ipv6_addr_loopback(&((struct sockaddr_in6 *) addr)->sin6_addr);
1006 return ib_addr_loopback(&((struct sockaddr_ib *) addr)->sib_addr);
1012 static inline int cma_any_addr(struct sockaddr *addr)
1014 return cma_zero_addr(addr) || cma_loopback_addr(addr);
1017 static int cma_addr_cmp(struct sockaddr *src, struct sockaddr *dst)
1019 if (src->sa_family != dst->sa_family)
1022 switch (src->sa_family) {
1024 return ((struct sockaddr_in *) src)->sin_addr.s_addr !=
1025 ((struct sockaddr_in *) dst)->sin_addr.s_addr;
1027 return ipv6_addr_cmp(&((struct sockaddr_in6 *) src)->sin6_addr,
1028 &((struct sockaddr_in6 *) dst)->sin6_addr);
1030 return ib_addr_cmp(&((struct sockaddr_ib *) src)->sib_addr,
1031 &((struct sockaddr_ib *) dst)->sib_addr);
1035 static __be16 cma_port(struct sockaddr *addr)
1037 struct sockaddr_ib *sib;
1039 switch (addr->sa_family) {
1041 return ((struct sockaddr_in *) addr)->sin_port;
1043 return ((struct sockaddr_in6 *) addr)->sin6_port;
1045 sib = (struct sockaddr_ib *) addr;
1046 return htons((u16) (be64_to_cpu(sib->sib_sid) &
1047 be64_to_cpu(sib->sib_sid_mask)));
1053 static inline int cma_any_port(struct sockaddr *addr)
1055 return !cma_port(addr);
1058 static void cma_save_ib_info(struct sockaddr *src_addr,
1059 struct sockaddr *dst_addr,
1060 struct rdma_cm_id *listen_id,
1061 struct ib_sa_path_rec *path)
1063 struct sockaddr_ib *listen_ib, *ib;
1065 listen_ib = (struct sockaddr_ib *) &listen_id->route.addr.src_addr;
1067 ib = (struct sockaddr_ib *)src_addr;
1068 ib->sib_family = AF_IB;
1070 ib->sib_pkey = path->pkey;
1071 ib->sib_flowinfo = path->flow_label;
1072 memcpy(&ib->sib_addr, &path->sgid, 16);
1073 ib->sib_sid = path->service_id;
1074 ib->sib_scope_id = 0;
1076 ib->sib_pkey = listen_ib->sib_pkey;
1077 ib->sib_flowinfo = listen_ib->sib_flowinfo;
1078 ib->sib_addr = listen_ib->sib_addr;
1079 ib->sib_sid = listen_ib->sib_sid;
1080 ib->sib_scope_id = listen_ib->sib_scope_id;
1082 ib->sib_sid_mask = cpu_to_be64(0xffffffffffffffffULL);
1085 ib = (struct sockaddr_ib *)dst_addr;
1086 ib->sib_family = AF_IB;
1088 ib->sib_pkey = path->pkey;
1089 ib->sib_flowinfo = path->flow_label;
1090 memcpy(&ib->sib_addr, &path->dgid, 16);
1095 static void cma_save_ip4_info(struct sockaddr_in *src_addr,
1096 struct sockaddr_in *dst_addr,
1097 struct cma_hdr *hdr,
1101 *src_addr = (struct sockaddr_in) {
1102 .sin_len = sizeof(struct sockaddr_in),
1103 .sin_family = AF_INET,
1104 .sin_addr.s_addr = hdr->dst_addr.ip4.addr,
1105 .sin_port = local_port,
1110 *dst_addr = (struct sockaddr_in) {
1111 .sin_len = sizeof(struct sockaddr_in),
1112 .sin_family = AF_INET,
1113 .sin_addr.s_addr = hdr->src_addr.ip4.addr,
1114 .sin_port = hdr->port,
1119 static void cma_ip6_clear_scope_id(struct in6_addr *addr)
1121 /* make sure link local scope ID gets zeroed */
1122 if (IN6_IS_SCOPE_LINKLOCAL(addr) ||
1123 IN6_IS_ADDR_MC_INTFACELOCAL(addr)) {
1124 /* use byte-access to be alignment safe */
1125 addr->s6_addr[2] = 0;
1126 addr->s6_addr[3] = 0;
1130 static void cma_save_ip6_info(struct sockaddr_in6 *src_addr,
1131 struct sockaddr_in6 *dst_addr,
1132 struct cma_hdr *hdr,
1136 *src_addr = (struct sockaddr_in6) {
1137 .sin6_len = sizeof(struct sockaddr_in6),
1138 .sin6_family = AF_INET6,
1139 .sin6_addr = hdr->dst_addr.ip6,
1140 .sin6_port = local_port,
1142 cma_ip6_clear_scope_id(&src_addr->sin6_addr);
1146 *dst_addr = (struct sockaddr_in6) {
1147 .sin6_len = sizeof(struct sockaddr_in6),
1148 .sin6_family = AF_INET6,
1149 .sin6_addr = hdr->src_addr.ip6,
1150 .sin6_port = hdr->port,
1152 cma_ip6_clear_scope_id(&dst_addr->sin6_addr);
1156 static u16 cma_port_from_service_id(__be64 service_id)
1158 return (u16)be64_to_cpu(service_id);
1161 static int cma_save_ip_info(struct sockaddr *src_addr,
1162 struct sockaddr *dst_addr,
1163 struct ib_cm_event *ib_event,
1166 struct cma_hdr *hdr;
1169 hdr = ib_event->private_data;
1170 if (hdr->cma_version != CMA_VERSION)
1173 port = htons(cma_port_from_service_id(service_id));
1175 switch (cma_get_ip_ver(hdr)) {
1177 cma_save_ip4_info((struct sockaddr_in *)src_addr,
1178 (struct sockaddr_in *)dst_addr, hdr, port);
1181 cma_save_ip6_info((struct sockaddr_in6 *)src_addr,
1182 (struct sockaddr_in6 *)dst_addr, hdr, port);
1185 return -EAFNOSUPPORT;
1191 static int cma_save_net_info(struct sockaddr *src_addr,
1192 struct sockaddr *dst_addr,
1193 struct rdma_cm_id *listen_id,
1194 struct ib_cm_event *ib_event,
1195 sa_family_t sa_family, __be64 service_id)
1197 if (sa_family == AF_IB) {
1198 if (ib_event->event == IB_CM_REQ_RECEIVED)
1199 cma_save_ib_info(src_addr, dst_addr, listen_id,
1200 ib_event->param.req_rcvd.primary_path);
1201 else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED)
1202 cma_save_ib_info(src_addr, dst_addr, listen_id, NULL);
1206 return cma_save_ip_info(src_addr, dst_addr, ib_event, service_id);
1209 static int cma_save_req_info(const struct ib_cm_event *ib_event,
1210 struct cma_req_info *req)
1212 const struct ib_cm_req_event_param *req_param =
1213 &ib_event->param.req_rcvd;
1214 const struct ib_cm_sidr_req_event_param *sidr_param =
1215 &ib_event->param.sidr_req_rcvd;
1217 switch (ib_event->event) {
1218 case IB_CM_REQ_RECEIVED:
1219 req->device = req_param->listen_id->device;
1220 req->port = req_param->port;
1221 memcpy(&req->local_gid, &req_param->primary_path->sgid,
1222 sizeof(req->local_gid));
1223 req->has_gid = true;
1224 req->service_id = req_param->primary_path->service_id;
1225 req->pkey = be16_to_cpu(req_param->primary_path->pkey);
1226 if (req->pkey != req_param->bth_pkey)
1227 pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and primary path P_Key (0x%x)\n"
1228 "RDMA CMA: in the future this may cause the request to be dropped\n",
1229 req_param->bth_pkey, req->pkey);
1231 case IB_CM_SIDR_REQ_RECEIVED:
1232 req->device = sidr_param->listen_id->device;
1233 req->port = sidr_param->port;
1234 req->has_gid = false;
1235 req->service_id = sidr_param->service_id;
1236 req->pkey = sidr_param->pkey;
1237 if (req->pkey != sidr_param->bth_pkey)
1238 pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and SIDR request payload P_Key (0x%x)\n"
1239 "RDMA CMA: in the future this may cause the request to be dropped\n",
1240 sidr_param->bth_pkey, req->pkey);
1249 static bool validate_ipv4_net_dev(struct net_device *net_dev,
1250 const struct sockaddr_in *dst_addr,
1251 const struct sockaddr_in *src_addr)
1254 struct sockaddr_in dst_tmp = *dst_addr;
1255 __be32 daddr = dst_addr->sin_addr.s_addr,
1256 saddr = src_addr->sin_addr.s_addr;
1257 struct net_device *src_dev;
1258 struct rtentry *rte;
1261 if (ipv4_is_multicast(saddr) || ipv4_is_lbcast(saddr) ||
1262 ipv4_is_lbcast(daddr) || ipv4_is_zeronet(saddr) ||
1263 ipv4_is_zeronet(daddr) || ipv4_is_loopback(daddr) ||
1264 ipv4_is_loopback(saddr))
1267 src_dev = ip_dev_find(net_dev->if_vnet, saddr);
1268 if (src_dev != net_dev) {
1269 if (src_dev != NULL)
1276 * Make sure the socket address length field
1277 * is set, else rtalloc1() will fail.
1279 dst_tmp.sin_len = sizeof(dst_tmp);
1281 CURVNET_SET(net_dev->if_vnet);
1282 rte = rtalloc1((struct sockaddr *)&dst_tmp, 1, 0);
1285 ret = (rte->rt_ifp == net_dev);
1296 static bool validate_ipv6_net_dev(struct net_device *net_dev,
1297 const struct sockaddr_in6 *dst_addr,
1298 const struct sockaddr_in6 *src_addr)
1301 struct sockaddr_in6 dst_tmp = *dst_addr;
1302 struct in6_addr in6_addr = src_addr->sin6_addr;
1303 struct net_device *src_dev;
1304 struct rtentry *rte;
1307 src_dev = ip6_dev_find(net_dev->if_vnet, in6_addr);
1308 if (src_dev != net_dev)
1312 * Make sure the socket address length field
1313 * is set, else rtalloc1() will fail.
1315 dst_tmp.sin6_len = sizeof(dst_tmp);
1317 CURVNET_SET(net_dev->if_vnet);
1318 rte = rtalloc1((struct sockaddr *)&dst_tmp, 1, 0);
1321 ret = (rte->rt_ifp == net_dev);
1332 static bool validate_net_dev(struct net_device *net_dev,
1333 const struct sockaddr *daddr,
1334 const struct sockaddr *saddr)
1336 const struct sockaddr_in *daddr4 = (const struct sockaddr_in *)daddr;
1337 const struct sockaddr_in *saddr4 = (const struct sockaddr_in *)saddr;
1338 const struct sockaddr_in6 *daddr6 = (const struct sockaddr_in6 *)daddr;
1339 const struct sockaddr_in6 *saddr6 = (const struct sockaddr_in6 *)saddr;
1341 switch (daddr->sa_family) {
1343 return saddr->sa_family == AF_INET &&
1344 validate_ipv4_net_dev(net_dev, daddr4, saddr4);
1347 return saddr->sa_family == AF_INET6 &&
1348 validate_ipv6_net_dev(net_dev, daddr6, saddr6);
1355 static struct net_device *cma_get_net_dev(struct ib_cm_event *ib_event,
1356 const struct cma_req_info *req)
1358 struct sockaddr_storage listen_addr_storage, src_addr_storage;
1359 struct sockaddr *listen_addr = (struct sockaddr *)&listen_addr_storage,
1360 *src_addr = (struct sockaddr *)&src_addr_storage;
1361 struct net_device *net_dev;
1362 const union ib_gid *gid = req->has_gid ? &req->local_gid : NULL;
1365 err = cma_save_ip_info(listen_addr, src_addr, ib_event,
1368 return ERR_PTR(err);
1370 net_dev = ib_get_net_dev_by_params(req->device, req->port, req->pkey,
1373 return ERR_PTR(-ENODEV);
1375 if (!validate_net_dev(net_dev, listen_addr, src_addr)) {
1377 return ERR_PTR(-EHOSTUNREACH);
1383 static enum rdma_port_space rdma_ps_from_service_id(__be64 service_id)
1385 return (be64_to_cpu(service_id) >> 16) & 0xffff;
1388 static bool cma_match_private_data(struct rdma_id_private *id_priv,
1389 const struct cma_hdr *hdr)
1391 struct sockaddr *addr = cma_src_addr(id_priv);
1393 struct in6_addr ip6_addr;
1395 if (cma_any_addr(addr) && !id_priv->afonly)
1398 switch (addr->sa_family) {
1400 ip4_addr = ((struct sockaddr_in *)addr)->sin_addr.s_addr;
1401 if (cma_get_ip_ver(hdr) != 4)
1403 if (!cma_any_addr(addr) &&
1404 hdr->dst_addr.ip4.addr != ip4_addr)
1408 ip6_addr = ((struct sockaddr_in6 *)addr)->sin6_addr;
1409 if (cma_get_ip_ver(hdr) != 6)
1411 cma_ip6_clear_scope_id(&ip6_addr);
1412 if (!cma_any_addr(addr) &&
1413 memcmp(&hdr->dst_addr.ip6, &ip6_addr, sizeof(ip6_addr)))
1425 static bool cma_protocol_roce_dev_port(struct ib_device *device, int port_num)
1427 enum rdma_link_layer ll = rdma_port_get_link_layer(device, port_num);
1428 enum rdma_transport_type transport =
1429 rdma_node_get_transport(device->node_type);
1431 return ll == IB_LINK_LAYER_ETHERNET && transport == RDMA_TRANSPORT_IB;
1434 static bool cma_protocol_roce(const struct rdma_cm_id *id)
1436 struct ib_device *device = id->device;
1437 const int port_num = id->port_num ?: rdma_start_port(device);
1439 return cma_protocol_roce_dev_port(device, port_num);
1442 static bool cma_match_net_dev(const struct rdma_cm_id *id,
1443 const struct net_device *net_dev,
1446 const struct rdma_addr *addr = &id->route.addr;
1449 /* This request is an AF_IB request or a RoCE request */
1450 return (!id->port_num || id->port_num == port_num) &&
1451 (addr->src_addr.ss_family == AF_IB ||
1452 cma_protocol_roce_dev_port(id->device, port_num));
1454 return !addr->dev_addr.bound_dev_if ||
1455 (net_eq(dev_net(net_dev), addr->dev_addr.net) &&
1456 addr->dev_addr.bound_dev_if == net_dev->if_index);
1459 static struct rdma_id_private *cma_find_listener(
1460 const struct rdma_bind_list *bind_list,
1461 const struct ib_cm_id *cm_id,
1462 const struct ib_cm_event *ib_event,
1463 const struct cma_req_info *req,
1464 const struct net_device *net_dev)
1466 struct rdma_id_private *id_priv, *id_priv_dev;
1469 return ERR_PTR(-EINVAL);
1471 hlist_for_each_entry(id_priv, &bind_list->owners, node) {
1472 if (cma_match_private_data(id_priv, ib_event->private_data)) {
1473 if (id_priv->id.device == cm_id->device &&
1474 cma_match_net_dev(&id_priv->id, net_dev, req->port))
1476 list_for_each_entry(id_priv_dev,
1477 &id_priv->listen_list,
1479 if (id_priv_dev->id.device == cm_id->device &&
1480 cma_match_net_dev(&id_priv_dev->id, net_dev, req->port))
1486 return ERR_PTR(-EINVAL);
1489 static struct rdma_id_private *cma_id_from_event(struct ib_cm_id *cm_id,
1490 struct ib_cm_event *ib_event,
1491 struct net_device **net_dev)
1493 struct cma_req_info req;
1494 struct rdma_bind_list *bind_list;
1495 struct rdma_id_private *id_priv;
1498 err = cma_save_req_info(ib_event, &req);
1500 return ERR_PTR(err);
1502 *net_dev = cma_get_net_dev(ib_event, &req);
1503 if (IS_ERR(*net_dev)) {
1504 if (PTR_ERR(*net_dev) == -EAFNOSUPPORT) {
1505 /* Assuming the protocol is AF_IB */
1507 } else if (cma_protocol_roce_dev_port(req.device, req.port)) {
1508 /* TODO find the net dev matching the request parameters
1509 * through the RoCE GID table */
1512 return ERR_CAST(*net_dev);
1516 bind_list = cma_ps_find(*net_dev ? dev_net(*net_dev) : &init_net,
1517 rdma_ps_from_service_id(req.service_id),
1518 cma_port_from_service_id(req.service_id));
1519 id_priv = cma_find_listener(bind_list, cm_id, ib_event, &req, *net_dev);
1520 if (IS_ERR(id_priv) && *net_dev) {
1528 static inline int cma_user_data_offset(struct rdma_id_private *id_priv)
1530 return cma_family(id_priv) == AF_IB ? 0 : sizeof(struct cma_hdr);
1533 static void cma_cancel_route(struct rdma_id_private *id_priv)
1535 if (rdma_cap_ib_sa(id_priv->id.device, id_priv->id.port_num)) {
1537 ib_sa_cancel_query(id_priv->query_id, id_priv->query);
1541 static void cma_cancel_listens(struct rdma_id_private *id_priv)
1543 struct rdma_id_private *dev_id_priv;
1546 * Remove from listen_any_list to prevent added devices from spawning
1547 * additional listen requests.
1550 list_del(&id_priv->list);
1552 while (!list_empty(&id_priv->listen_list)) {
1553 dev_id_priv = list_entry(id_priv->listen_list.next,
1554 struct rdma_id_private, listen_list);
1555 /* sync with device removal to avoid duplicate destruction */
1556 list_del_init(&dev_id_priv->list);
1557 list_del(&dev_id_priv->listen_list);
1558 mutex_unlock(&lock);
1560 rdma_destroy_id(&dev_id_priv->id);
1563 mutex_unlock(&lock);
1566 static void cma_cancel_operation(struct rdma_id_private *id_priv,
1567 enum rdma_cm_state state)
1570 case RDMA_CM_ADDR_QUERY:
1571 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
1573 case RDMA_CM_ROUTE_QUERY:
1574 cma_cancel_route(id_priv);
1576 case RDMA_CM_LISTEN:
1577 if (cma_any_addr(cma_src_addr(id_priv)) && !id_priv->cma_dev)
1578 cma_cancel_listens(id_priv);
1585 static void cma_release_port(struct rdma_id_private *id_priv)
1587 struct rdma_bind_list *bind_list = id_priv->bind_list;
1588 struct vnet *net = id_priv->id.route.addr.dev_addr.net;
1594 hlist_del(&id_priv->node);
1595 if (hlist_empty(&bind_list->owners)) {
1596 cma_ps_remove(net, bind_list->ps, bind_list->port);
1599 mutex_unlock(&lock);
1602 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
1604 struct cma_multicast *mc;
1606 while (!list_empty(&id_priv->mc_list)) {
1607 mc = container_of(id_priv->mc_list.next,
1608 struct cma_multicast, list);
1609 list_del(&mc->list);
1610 if (rdma_cap_ib_mcast(id_priv->cma_dev->device,
1611 id_priv->id.port_num)) {
1612 ib_sa_free_multicast(mc->multicast.ib);
1615 if (mc->igmp_joined) {
1616 struct rdma_dev_addr *dev_addr =
1617 &id_priv->id.route.addr.dev_addr;
1618 struct net_device *ndev = NULL;
1620 if (dev_addr->bound_dev_if)
1621 ndev = dev_get_by_index(dev_addr->net,
1622 dev_addr->bound_dev_if);
1627 kref_put(&mc->mcref, release_mc);
1632 void rdma_destroy_id(struct rdma_cm_id *id)
1634 struct rdma_id_private *id_priv;
1635 enum rdma_cm_state state;
1637 id_priv = container_of(id, struct rdma_id_private, id);
1638 state = cma_exch(id_priv, RDMA_CM_DESTROYING);
1639 cma_cancel_operation(id_priv, state);
1642 * Wait for any active callback to finish. New callbacks will find
1643 * the id_priv state set to destroying and abort.
1645 mutex_lock(&id_priv->handler_mutex);
1646 mutex_unlock(&id_priv->handler_mutex);
1648 if (id_priv->cma_dev) {
1649 if (rdma_cap_ib_cm(id_priv->id.device, 1)) {
1650 if (id_priv->cm_id.ib)
1651 ib_destroy_cm_id(id_priv->cm_id.ib);
1652 } else if (rdma_cap_iw_cm(id_priv->id.device, 1)) {
1653 if (id_priv->cm_id.iw)
1654 iw_destroy_cm_id(id_priv->cm_id.iw);
1656 cma_leave_mc_groups(id_priv);
1657 cma_release_dev(id_priv);
1660 cma_release_port(id_priv);
1661 cma_deref_id(id_priv);
1662 wait_for_completion(&id_priv->comp);
1664 if (id_priv->internal_id)
1665 cma_deref_id(id_priv->id.context);
1667 kfree(id_priv->id.route.path_rec);
1670 EXPORT_SYMBOL(rdma_destroy_id);
1672 static int cma_rep_recv(struct rdma_id_private *id_priv)
1676 ret = cma_modify_qp_rtr(id_priv, NULL);
1680 ret = cma_modify_qp_rts(id_priv, NULL);
1684 ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
1690 cma_modify_qp_err(id_priv);
1691 ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
1696 static void cma_set_rep_event_data(struct rdma_cm_event *event,
1697 struct ib_cm_rep_event_param *rep_data,
1700 event->param.conn.private_data = private_data;
1701 event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
1702 event->param.conn.responder_resources = rep_data->responder_resources;
1703 event->param.conn.initiator_depth = rep_data->initiator_depth;
1704 event->param.conn.flow_control = rep_data->flow_control;
1705 event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
1706 event->param.conn.srq = rep_data->srq;
1707 event->param.conn.qp_num = rep_data->remote_qpn;
1710 static int cma_ib_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1712 struct rdma_id_private *id_priv = cm_id->context;
1713 struct rdma_cm_event event;
1716 mutex_lock(&id_priv->handler_mutex);
1717 if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
1718 id_priv->state != RDMA_CM_CONNECT) ||
1719 (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
1720 id_priv->state != RDMA_CM_DISCONNECT))
1723 memset(&event, 0, sizeof event);
1724 switch (ib_event->event) {
1725 case IB_CM_REQ_ERROR:
1726 case IB_CM_REP_ERROR:
1727 event.event = RDMA_CM_EVENT_UNREACHABLE;
1728 event.status = -ETIMEDOUT;
1730 case IB_CM_REP_RECEIVED:
1731 if (id_priv->id.qp) {
1732 event.status = cma_rep_recv(id_priv);
1733 event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
1734 RDMA_CM_EVENT_ESTABLISHED;
1736 event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
1738 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
1739 ib_event->private_data);
1741 case IB_CM_RTU_RECEIVED:
1742 case IB_CM_USER_ESTABLISHED:
1743 event.event = RDMA_CM_EVENT_ESTABLISHED;
1745 case IB_CM_DREQ_ERROR:
1746 event.status = -ETIMEDOUT; /* fall through */
1747 case IB_CM_DREQ_RECEIVED:
1748 case IB_CM_DREP_RECEIVED:
1749 if (!cma_comp_exch(id_priv, RDMA_CM_CONNECT,
1750 RDMA_CM_DISCONNECT))
1752 event.event = RDMA_CM_EVENT_DISCONNECTED;
1754 case IB_CM_TIMEWAIT_EXIT:
1755 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
1757 case IB_CM_MRA_RECEIVED:
1760 case IB_CM_REJ_RECEIVED:
1761 cma_modify_qp_err(id_priv);
1762 event.status = ib_event->param.rej_rcvd.reason;
1763 event.event = RDMA_CM_EVENT_REJECTED;
1764 event.param.conn.private_data = ib_event->private_data;
1765 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
1768 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
1773 ret = id_priv->id.event_handler(&id_priv->id, &event);
1775 /* Destroy the CM ID by returning a non-zero value. */
1776 id_priv->cm_id.ib = NULL;
1777 cma_exch(id_priv, RDMA_CM_DESTROYING);
1778 mutex_unlock(&id_priv->handler_mutex);
1779 rdma_destroy_id(&id_priv->id);
1783 mutex_unlock(&id_priv->handler_mutex);
1787 static struct rdma_id_private *cma_new_conn_id(struct rdma_cm_id *listen_id,
1788 struct ib_cm_event *ib_event,
1789 struct net_device *net_dev)
1791 struct rdma_id_private *id_priv;
1792 struct rdma_cm_id *id;
1793 struct rdma_route *rt;
1794 const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
1795 const __be64 service_id =
1796 ib_event->param.req_rcvd.primary_path->service_id;
1799 id = rdma_create_id(listen_id->route.addr.dev_addr.net,
1800 listen_id->event_handler, listen_id->context,
1801 listen_id->ps, ib_event->param.req_rcvd.qp_type);
1805 id_priv = container_of(id, struct rdma_id_private, id);
1806 if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
1807 (struct sockaddr *)&id->route.addr.dst_addr,
1808 listen_id, ib_event, ss_family, service_id))
1812 rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
1813 rt->path_rec = kmalloc(sizeof *rt->path_rec * rt->num_paths,
1818 rt->path_rec[0] = *ib_event->param.req_rcvd.primary_path;
1819 if (rt->num_paths == 2)
1820 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
1823 ret = rdma_copy_addr(&rt->addr.dev_addr, net_dev, NULL);
1827 if (!cma_protocol_roce(listen_id) &&
1828 cma_any_addr(cma_src_addr(id_priv))) {
1829 rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
1830 rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
1831 ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
1832 } else if (!cma_any_addr(cma_src_addr(id_priv))) {
1833 ret = cma_translate_addr(cma_src_addr(id_priv), &rt->addr.dev_addr);
1838 rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
1840 id_priv->state = RDMA_CM_CONNECT;
1844 rdma_destroy_id(id);
1848 static struct rdma_id_private *cma_new_udp_id(struct rdma_cm_id *listen_id,
1849 struct ib_cm_event *ib_event,
1850 struct net_device *net_dev)
1852 struct rdma_id_private *id_priv;
1853 struct rdma_cm_id *id;
1854 const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
1855 struct vnet *net = listen_id->route.addr.dev_addr.net;
1858 id = rdma_create_id(net, listen_id->event_handler, listen_id->context,
1859 listen_id->ps, IB_QPT_UD);
1863 id_priv = container_of(id, struct rdma_id_private, id);
1864 if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
1865 (struct sockaddr *)&id->route.addr.dst_addr,
1866 listen_id, ib_event, ss_family,
1867 ib_event->param.sidr_req_rcvd.service_id))
1871 ret = rdma_copy_addr(&id->route.addr.dev_addr, net_dev, NULL);
1875 if (!cma_any_addr(cma_src_addr(id_priv))) {
1876 ret = cma_translate_addr(cma_src_addr(id_priv),
1877 &id->route.addr.dev_addr);
1883 id_priv->state = RDMA_CM_CONNECT;
1886 rdma_destroy_id(id);
1890 static void cma_set_req_event_data(struct rdma_cm_event *event,
1891 struct ib_cm_req_event_param *req_data,
1892 void *private_data, int offset)
1894 event->param.conn.private_data = (char *)private_data + offset;
1895 event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
1896 event->param.conn.responder_resources = req_data->responder_resources;
1897 event->param.conn.initiator_depth = req_data->initiator_depth;
1898 event->param.conn.flow_control = req_data->flow_control;
1899 event->param.conn.retry_count = req_data->retry_count;
1900 event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
1901 event->param.conn.srq = req_data->srq;
1902 event->param.conn.qp_num = req_data->remote_qpn;
1905 static int cma_check_req_qp_type(struct rdma_cm_id *id, struct ib_cm_event *ib_event)
1907 return (((ib_event->event == IB_CM_REQ_RECEIVED) &&
1908 (ib_event->param.req_rcvd.qp_type == id->qp_type)) ||
1909 ((ib_event->event == IB_CM_SIDR_REQ_RECEIVED) &&
1910 (id->qp_type == IB_QPT_UD)) ||
1914 static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1916 struct rdma_id_private *listen_id, *conn_id = NULL;
1917 struct rdma_cm_event event;
1918 struct net_device *net_dev;
1921 listen_id = cma_id_from_event(cm_id, ib_event, &net_dev);
1922 if (IS_ERR(listen_id))
1923 return PTR_ERR(listen_id);
1925 if (!cma_check_req_qp_type(&listen_id->id, ib_event)) {
1930 mutex_lock(&listen_id->handler_mutex);
1931 if (listen_id->state != RDMA_CM_LISTEN) {
1932 ret = -ECONNABORTED;
1936 memset(&event, 0, sizeof event);
1937 offset = cma_user_data_offset(listen_id);
1938 event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1939 if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) {
1940 conn_id = cma_new_udp_id(&listen_id->id, ib_event, net_dev);
1941 event.param.ud.private_data = (char *)ib_event->private_data + offset;
1942 event.param.ud.private_data_len =
1943 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
1945 conn_id = cma_new_conn_id(&listen_id->id, ib_event, net_dev);
1946 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
1947 ib_event->private_data, offset);
1954 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1955 ret = cma_acquire_dev(conn_id, listen_id);
1959 conn_id->cm_id.ib = cm_id;
1960 cm_id->context = conn_id;
1961 cm_id->cm_handler = cma_ib_handler;
1964 * Protect against the user destroying conn_id from another thread
1965 * until we're done accessing it.
1967 atomic_inc(&conn_id->refcount);
1968 ret = conn_id->id.event_handler(&conn_id->id, &event);
1972 * Acquire mutex to prevent user executing rdma_destroy_id()
1973 * while we're accessing the cm_id.
1976 if (cma_comp(conn_id, RDMA_CM_CONNECT) &&
1977 (conn_id->id.qp_type != IB_QPT_UD))
1978 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
1979 mutex_unlock(&lock);
1980 mutex_unlock(&conn_id->handler_mutex);
1981 mutex_unlock(&listen_id->handler_mutex);
1982 cma_deref_id(conn_id);
1988 cma_deref_id(conn_id);
1989 /* Destroy the CM ID by returning a non-zero value. */
1990 conn_id->cm_id.ib = NULL;
1992 cma_exch(conn_id, RDMA_CM_DESTROYING);
1993 mutex_unlock(&conn_id->handler_mutex);
1995 mutex_unlock(&listen_id->handler_mutex);
1997 rdma_destroy_id(&conn_id->id);
2006 __be64 rdma_get_service_id(struct rdma_cm_id *id, struct sockaddr *addr)
2008 if (addr->sa_family == AF_IB)
2009 return ((struct sockaddr_ib *) addr)->sib_sid;
2011 return cpu_to_be64(((u64)id->ps << 16) + be16_to_cpu(cma_port(addr)));
2013 EXPORT_SYMBOL(rdma_get_service_id);
2015 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
2017 struct rdma_id_private *id_priv = iw_id->context;
2018 struct rdma_cm_event event;
2020 struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
2021 struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
2023 mutex_lock(&id_priv->handler_mutex);
2024 if (id_priv->state != RDMA_CM_CONNECT)
2027 memset(&event, 0, sizeof event);
2028 switch (iw_event->event) {
2029 case IW_CM_EVENT_CLOSE:
2030 event.event = RDMA_CM_EVENT_DISCONNECTED;
2032 case IW_CM_EVENT_CONNECT_REPLY:
2033 memcpy(cma_src_addr(id_priv), laddr,
2034 rdma_addr_size(laddr));
2035 memcpy(cma_dst_addr(id_priv), raddr,
2036 rdma_addr_size(raddr));
2037 switch (iw_event->status) {
2039 event.event = RDMA_CM_EVENT_ESTABLISHED;
2040 event.param.conn.initiator_depth = iw_event->ird;
2041 event.param.conn.responder_resources = iw_event->ord;
2045 event.event = RDMA_CM_EVENT_REJECTED;
2048 event.event = RDMA_CM_EVENT_UNREACHABLE;
2051 event.event = RDMA_CM_EVENT_CONNECT_ERROR;
2055 case IW_CM_EVENT_ESTABLISHED:
2056 event.event = RDMA_CM_EVENT_ESTABLISHED;
2057 event.param.conn.initiator_depth = iw_event->ird;
2058 event.param.conn.responder_resources = iw_event->ord;
2064 event.status = iw_event->status;
2065 event.param.conn.private_data = iw_event->private_data;
2066 event.param.conn.private_data_len = iw_event->private_data_len;
2067 ret = id_priv->id.event_handler(&id_priv->id, &event);
2069 /* Destroy the CM ID by returning a non-zero value. */
2070 id_priv->cm_id.iw = NULL;
2071 cma_exch(id_priv, RDMA_CM_DESTROYING);
2072 mutex_unlock(&id_priv->handler_mutex);
2073 rdma_destroy_id(&id_priv->id);
2078 mutex_unlock(&id_priv->handler_mutex);
2082 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
2083 struct iw_cm_event *iw_event)
2085 struct rdma_cm_id *new_cm_id;
2086 struct rdma_id_private *listen_id, *conn_id;
2087 struct rdma_cm_event event;
2088 int ret = -ECONNABORTED;
2089 struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
2090 struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
2092 listen_id = cm_id->context;
2094 mutex_lock(&listen_id->handler_mutex);
2095 if (listen_id->state != RDMA_CM_LISTEN)
2098 /* Create a new RDMA id for the new IW CM ID */
2099 new_cm_id = rdma_create_id(listen_id->id.route.addr.dev_addr.net,
2100 listen_id->id.event_handler,
2101 listen_id->id.context,
2102 RDMA_PS_TCP, IB_QPT_RC);
2103 if (IS_ERR(new_cm_id)) {
2107 conn_id = container_of(new_cm_id, struct rdma_id_private, id);
2108 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
2109 conn_id->state = RDMA_CM_CONNECT;
2111 ret = rdma_translate_ip(laddr, &conn_id->id.route.addr.dev_addr);
2113 mutex_unlock(&conn_id->handler_mutex);
2114 rdma_destroy_id(new_cm_id);
2118 ret = cma_acquire_dev(conn_id, listen_id);
2120 mutex_unlock(&conn_id->handler_mutex);
2121 rdma_destroy_id(new_cm_id);
2125 conn_id->cm_id.iw = cm_id;
2126 cm_id->context = conn_id;
2127 cm_id->cm_handler = cma_iw_handler;
2129 memcpy(cma_src_addr(conn_id), laddr, rdma_addr_size(laddr));
2130 memcpy(cma_dst_addr(conn_id), raddr, rdma_addr_size(raddr));
2132 memset(&event, 0, sizeof event);
2133 event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
2134 event.param.conn.private_data = iw_event->private_data;
2135 event.param.conn.private_data_len = iw_event->private_data_len;
2136 event.param.conn.initiator_depth = iw_event->ird;
2137 event.param.conn.responder_resources = iw_event->ord;
2140 * Protect against the user destroying conn_id from another thread
2141 * until we're done accessing it.
2143 atomic_inc(&conn_id->refcount);
2144 ret = conn_id->id.event_handler(&conn_id->id, &event);
2146 /* User wants to destroy the CM ID */
2147 conn_id->cm_id.iw = NULL;
2148 cma_exch(conn_id, RDMA_CM_DESTROYING);
2149 mutex_unlock(&conn_id->handler_mutex);
2150 cma_deref_id(conn_id);
2151 rdma_destroy_id(&conn_id->id);
2155 mutex_unlock(&conn_id->handler_mutex);
2156 cma_deref_id(conn_id);
2159 mutex_unlock(&listen_id->handler_mutex);
2163 static int cma_ib_listen(struct rdma_id_private *id_priv)
2165 struct sockaddr *addr;
2166 struct ib_cm_id *id;
2169 addr = cma_src_addr(id_priv);
2170 svc_id = rdma_get_service_id(&id_priv->id, addr);
2171 id = ib_cm_insert_listen(id_priv->id.device, cma_req_handler, svc_id);
2174 id_priv->cm_id.ib = id;
2179 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
2182 struct iw_cm_id *id;
2184 id = iw_create_cm_id(id_priv->id.device,
2185 iw_conn_req_handler,
2190 id->tos = id_priv->tos;
2191 id_priv->cm_id.iw = id;
2193 memcpy(&id_priv->cm_id.iw->local_addr, cma_src_addr(id_priv),
2194 rdma_addr_size(cma_src_addr(id_priv)));
2196 ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
2199 iw_destroy_cm_id(id_priv->cm_id.iw);
2200 id_priv->cm_id.iw = NULL;
2206 static int cma_listen_handler(struct rdma_cm_id *id,
2207 struct rdma_cm_event *event)
2209 struct rdma_id_private *id_priv = id->context;
2211 id->context = id_priv->id.context;
2212 id->event_handler = id_priv->id.event_handler;
2213 return id_priv->id.event_handler(id, event);
2216 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
2217 struct cma_device *cma_dev)
2219 struct rdma_id_private *dev_id_priv;
2220 struct rdma_cm_id *id;
2221 struct vnet *net = id_priv->id.route.addr.dev_addr.net;
2224 if (cma_family(id_priv) == AF_IB && !rdma_cap_ib_cm(cma_dev->device, 1))
2227 id = rdma_create_id(net, cma_listen_handler, id_priv, id_priv->id.ps,
2228 id_priv->id.qp_type);
2232 dev_id_priv = container_of(id, struct rdma_id_private, id);
2234 dev_id_priv->state = RDMA_CM_ADDR_BOUND;
2235 memcpy(cma_src_addr(dev_id_priv), cma_src_addr(id_priv),
2236 rdma_addr_size(cma_src_addr(id_priv)));
2238 _cma_attach_to_dev(dev_id_priv, cma_dev);
2239 list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
2240 atomic_inc(&id_priv->refcount);
2241 dev_id_priv->internal_id = 1;
2242 dev_id_priv->afonly = id_priv->afonly;
2244 ret = rdma_listen(id, id_priv->backlog);
2246 pr_warn("RDMA CMA: cma_listen_on_dev, error %d, listening on device %s\n",
2247 ret, cma_dev->device->name);
2250 static void cma_listen_on_all(struct rdma_id_private *id_priv)
2252 struct cma_device *cma_dev;
2255 list_add_tail(&id_priv->list, &listen_any_list);
2256 list_for_each_entry(cma_dev, &dev_list, list)
2257 cma_listen_on_dev(id_priv, cma_dev);
2258 mutex_unlock(&lock);
2261 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
2263 struct rdma_id_private *id_priv;
2265 id_priv = container_of(id, struct rdma_id_private, id);
2266 id_priv->tos = (u8) tos;
2268 EXPORT_SYMBOL(rdma_set_service_type);
2270 static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec,
2273 struct cma_work *work = context;
2274 struct rdma_route *route;
2276 route = &work->id->id.route;
2279 route->num_paths = 1;
2280 *route->path_rec = *path_rec;
2282 work->old_state = RDMA_CM_ROUTE_QUERY;
2283 work->new_state = RDMA_CM_ADDR_RESOLVED;
2284 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
2285 work->event.status = status;
2288 queue_work(cma_wq, &work->work);
2291 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms,
2292 struct cma_work *work)
2294 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2295 struct ib_sa_path_rec path_rec;
2296 ib_sa_comp_mask comp_mask;
2297 struct sockaddr_in6 *sin6;
2298 struct sockaddr_ib *sib;
2300 memset(&path_rec, 0, sizeof path_rec);
2301 rdma_addr_get_sgid(dev_addr, &path_rec.sgid);
2302 rdma_addr_get_dgid(dev_addr, &path_rec.dgid);
2303 path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
2304 path_rec.numb_path = 1;
2305 path_rec.reversible = 1;
2306 path_rec.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
2308 comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
2309 IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
2310 IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
2312 switch (cma_family(id_priv)) {
2314 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
2315 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
2318 sin6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
2319 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
2320 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2323 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
2324 path_rec.traffic_class = (u8) (be32_to_cpu(sib->sib_flowinfo) >> 20);
2325 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2329 id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
2330 id_priv->id.port_num, &path_rec,
2331 comp_mask, timeout_ms,
2332 GFP_KERNEL, cma_query_handler,
2333 work, &id_priv->query);
2335 return (id_priv->query_id < 0) ? id_priv->query_id : 0;
2338 static void cma_work_handler(struct work_struct *_work)
2340 struct cma_work *work = container_of(_work, struct cma_work, work);
2341 struct rdma_id_private *id_priv = work->id;
2344 mutex_lock(&id_priv->handler_mutex);
2345 if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
2348 if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
2349 cma_exch(id_priv, RDMA_CM_DESTROYING);
2353 mutex_unlock(&id_priv->handler_mutex);
2354 cma_deref_id(id_priv);
2356 rdma_destroy_id(&id_priv->id);
2360 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms)
2362 struct rdma_route *route = &id_priv->id.route;
2363 struct cma_work *work;
2366 work = kzalloc(sizeof *work, GFP_KERNEL);
2371 INIT_WORK(&work->work, cma_work_handler);
2372 work->old_state = RDMA_CM_ROUTE_QUERY;
2373 work->new_state = RDMA_CM_ROUTE_RESOLVED;
2374 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
2376 route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
2377 if (!route->path_rec) {
2382 ret = cma_query_ib_route(id_priv, timeout_ms, work);
2388 kfree(route->path_rec);
2389 route->path_rec = NULL;
2395 int rdma_set_ib_paths(struct rdma_cm_id *id,
2396 struct ib_sa_path_rec *path_rec, int num_paths)
2398 struct rdma_id_private *id_priv;
2401 id_priv = container_of(id, struct rdma_id_private, id);
2402 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
2403 RDMA_CM_ROUTE_RESOLVED))
2406 id->route.path_rec = kmemdup(path_rec, sizeof *path_rec * num_paths,
2408 if (!id->route.path_rec) {
2413 id->route.num_paths = num_paths;
2416 cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED);
2419 EXPORT_SYMBOL(rdma_set_ib_paths);
2421 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms)
2423 struct cma_work *work;
2425 work = kzalloc(sizeof *work, GFP_KERNEL);
2430 INIT_WORK(&work->work, cma_work_handler);
2431 work->old_state = RDMA_CM_ROUTE_QUERY;
2432 work->new_state = RDMA_CM_ROUTE_RESOLVED;
2433 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
2434 queue_work(cma_wq, &work->work);
2438 static int iboe_tos_to_sl(struct net_device *ndev, int tos)
2440 /* get service level, SL, from type of service, TOS */
2443 /* range check input argument and map 1:1 */
2449 /* final mappings are done by the vendor specific drivers */
2453 static enum ib_gid_type cma_route_gid_type(enum rdma_network_type network_type,
2454 unsigned long supported_gids,
2455 enum ib_gid_type default_gid)
2457 if ((network_type == RDMA_NETWORK_IPV4 ||
2458 network_type == RDMA_NETWORK_IPV6) &&
2459 test_bit(IB_GID_TYPE_ROCE_UDP_ENCAP, &supported_gids))
2460 return IB_GID_TYPE_ROCE_UDP_ENCAP;
2465 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
2467 struct rdma_route *route = &id_priv->id.route;
2468 struct rdma_addr *addr = &route->addr;
2469 struct cma_work *work;
2471 struct net_device *ndev = NULL;
2474 work = kzalloc(sizeof *work, GFP_KERNEL);
2479 INIT_WORK(&work->work, cma_work_handler);
2481 route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
2482 if (!route->path_rec) {
2487 route->num_paths = 1;
2489 if (addr->dev_addr.bound_dev_if) {
2490 unsigned long supported_gids;
2492 ndev = dev_get_by_index(addr->dev_addr.net,
2493 addr->dev_addr.bound_dev_if);
2499 if (ndev->if_flags & IFF_LOOPBACK) {
2501 if (!id_priv->id.device->get_netdev) {
2506 ndev = id_priv->id.device->get_netdev(id_priv->id.device,
2507 id_priv->id.port_num);
2514 route->path_rec->net = ndev->if_vnet;
2515 route->path_rec->ifindex = ndev->if_index;
2516 supported_gids = roce_gid_type_mask_support(id_priv->id.device,
2517 id_priv->id.port_num);
2518 route->path_rec->gid_type =
2519 cma_route_gid_type(addr->dev_addr.network,
2528 memcpy(route->path_rec->dmac, addr->dev_addr.dst_dev_addr, ETH_ALEN);
2530 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
2531 &route->path_rec->sgid);
2532 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.dst_addr,
2533 &route->path_rec->dgid);
2535 /* Use the hint from IP Stack to select GID Type */
2536 if (route->path_rec->gid_type < ib_network_to_gid_type(addr->dev_addr.network))
2537 route->path_rec->gid_type = ib_network_to_gid_type(addr->dev_addr.network);
2538 if (((struct sockaddr *)&id_priv->id.route.addr.dst_addr)->sa_family != AF_IB)
2539 /* TODO: get the hoplimit from the inet/inet6 device */
2540 route->path_rec->hop_limit = addr->dev_addr.hoplimit;
2542 route->path_rec->hop_limit = 1;
2543 route->path_rec->reversible = 1;
2544 route->path_rec->pkey = cpu_to_be16(0xffff);
2545 route->path_rec->mtu_selector = IB_SA_EQ;
2546 route->path_rec->sl = iboe_tos_to_sl(ndev, id_priv->tos);
2547 route->path_rec->mtu = iboe_get_mtu(ndev->if_mtu);
2548 route->path_rec->rate_selector = IB_SA_EQ;
2549 route->path_rec->rate = iboe_get_rate(ndev);
2551 route->path_rec->packet_life_time_selector = IB_SA_EQ;
2552 route->path_rec->packet_life_time = CMA_IBOE_PACKET_LIFETIME;
2553 if (!route->path_rec->mtu) {
2558 work->old_state = RDMA_CM_ROUTE_QUERY;
2559 work->new_state = RDMA_CM_ROUTE_RESOLVED;
2560 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
2561 work->event.status = 0;
2563 queue_work(cma_wq, &work->work);
2568 kfree(route->path_rec);
2569 route->path_rec = NULL;
2575 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
2577 struct rdma_id_private *id_priv;
2580 id_priv = container_of(id, struct rdma_id_private, id);
2581 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY))
2584 atomic_inc(&id_priv->refcount);
2585 if (rdma_cap_ib_sa(id->device, id->port_num))
2586 ret = cma_resolve_ib_route(id_priv, timeout_ms);
2587 else if (rdma_protocol_roce(id->device, id->port_num))
2588 ret = cma_resolve_iboe_route(id_priv);
2589 else if (rdma_protocol_iwarp(id->device, id->port_num))
2590 ret = cma_resolve_iw_route(id_priv, timeout_ms);
2599 cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED);
2600 cma_deref_id(id_priv);
2603 EXPORT_SYMBOL(rdma_resolve_route);
2605 static void cma_set_loopback(struct sockaddr *addr)
2607 switch (addr->sa_family) {
2609 ((struct sockaddr_in *) addr)->sin_addr.s_addr = htonl(INADDR_LOOPBACK);
2612 ipv6_addr_set(&((struct sockaddr_in6 *) addr)->sin6_addr,
2616 ib_addr_set(&((struct sockaddr_ib *) addr)->sib_addr,
2622 static int cma_bind_loopback(struct rdma_id_private *id_priv)
2624 struct cma_device *cma_dev, *cur_dev;
2625 struct ib_port_attr port_attr;
2633 list_for_each_entry(cur_dev, &dev_list, list) {
2634 if (cma_family(id_priv) == AF_IB &&
2635 !rdma_cap_ib_cm(cur_dev->device, 1))
2641 for (p = 1; p <= cur_dev->device->phys_port_cnt; ++p) {
2642 if (!ib_query_port(cur_dev->device, p, &port_attr) &&
2643 port_attr.state == IB_PORT_ACTIVE) {
2658 ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid, NULL);
2662 ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
2666 id_priv->id.route.addr.dev_addr.dev_type =
2667 (rdma_protocol_ib(cma_dev->device, p)) ?
2668 ARPHRD_INFINIBAND : ARPHRD_ETHER;
2670 rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2671 ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
2672 id_priv->id.port_num = p;
2673 cma_attach_to_dev(id_priv, cma_dev);
2674 cma_set_loopback(cma_src_addr(id_priv));
2676 mutex_unlock(&lock);
2680 static void addr_handler(int status, struct sockaddr *src_addr,
2681 struct rdma_dev_addr *dev_addr, void *context)
2683 struct rdma_id_private *id_priv = context;
2684 struct rdma_cm_event event;
2686 memset(&event, 0, sizeof event);
2687 mutex_lock(&id_priv->handler_mutex);
2688 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY,
2689 RDMA_CM_ADDR_RESOLVED))
2692 memcpy(cma_src_addr(id_priv), src_addr, rdma_addr_size(src_addr));
2693 if (!status && !id_priv->cma_dev)
2694 status = cma_acquire_dev(id_priv, NULL);
2697 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
2698 RDMA_CM_ADDR_BOUND))
2700 event.event = RDMA_CM_EVENT_ADDR_ERROR;
2701 event.status = status;
2703 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2705 if (id_priv->id.event_handler(&id_priv->id, &event)) {
2706 cma_exch(id_priv, RDMA_CM_DESTROYING);
2707 mutex_unlock(&id_priv->handler_mutex);
2708 cma_deref_id(id_priv);
2709 rdma_destroy_id(&id_priv->id);
2713 mutex_unlock(&id_priv->handler_mutex);
2714 cma_deref_id(id_priv);
2717 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
2719 struct cma_work *work;
2723 work = kzalloc(sizeof *work, GFP_KERNEL);
2727 if (!id_priv->cma_dev) {
2728 ret = cma_bind_loopback(id_priv);
2733 rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2734 rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
2737 INIT_WORK(&work->work, cma_work_handler);
2738 work->old_state = RDMA_CM_ADDR_QUERY;
2739 work->new_state = RDMA_CM_ADDR_RESOLVED;
2740 work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2741 queue_work(cma_wq, &work->work);
2748 static int cma_resolve_ib_addr(struct rdma_id_private *id_priv)
2750 struct cma_work *work;
2753 work = kzalloc(sizeof *work, GFP_KERNEL);
2757 if (!id_priv->cma_dev) {
2758 ret = cma_resolve_ib_dev(id_priv);
2763 rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, (union ib_gid *)
2764 &(((struct sockaddr_ib *) &id_priv->id.route.addr.dst_addr)->sib_addr));
2767 INIT_WORK(&work->work, cma_work_handler);
2768 work->old_state = RDMA_CM_ADDR_QUERY;
2769 work->new_state = RDMA_CM_ADDR_RESOLVED;
2770 work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2771 queue_work(cma_wq, &work->work);
2778 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2779 struct sockaddr *dst_addr)
2781 if (!src_addr || !src_addr->sa_family) {
2782 src_addr = (struct sockaddr *) &id->route.addr.src_addr;
2783 src_addr->sa_family = dst_addr->sa_family;
2784 if (dst_addr->sa_family == AF_INET6) {
2785 struct sockaddr_in6 *src_addr6 = (struct sockaddr_in6 *) src_addr;
2786 struct sockaddr_in6 *dst_addr6 = (struct sockaddr_in6 *) dst_addr;
2787 src_addr6->sin6_scope_id = dst_addr6->sin6_scope_id;
2788 if (IN6_IS_SCOPE_LINKLOCAL(&dst_addr6->sin6_addr) ||
2789 IN6_IS_ADDR_MC_INTFACELOCAL(&dst_addr6->sin6_addr))
2790 id->route.addr.dev_addr.bound_dev_if = dst_addr6->sin6_scope_id;
2791 } else if (dst_addr->sa_family == AF_IB) {
2792 ((struct sockaddr_ib *) src_addr)->sib_pkey =
2793 ((struct sockaddr_ib *) dst_addr)->sib_pkey;
2796 return rdma_bind_addr(id, src_addr);
2799 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2800 struct sockaddr *dst_addr, int timeout_ms)
2802 struct rdma_id_private *id_priv;
2805 id_priv = container_of(id, struct rdma_id_private, id);
2806 if (id_priv->state == RDMA_CM_IDLE) {
2807 ret = cma_bind_addr(id, src_addr, dst_addr);
2812 if (cma_family(id_priv) != dst_addr->sa_family)
2815 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY))
2818 atomic_inc(&id_priv->refcount);
2819 memcpy(cma_dst_addr(id_priv), dst_addr, rdma_addr_size(dst_addr));
2820 if (cma_any_addr(dst_addr)) {
2821 ret = cma_resolve_loopback(id_priv);
2823 if (dst_addr->sa_family == AF_IB) {
2824 ret = cma_resolve_ib_addr(id_priv);
2826 ret = cma_check_linklocal(&id->route.addr.dev_addr, dst_addr);
2830 ret = rdma_resolve_ip(&addr_client, cma_src_addr(id_priv),
2831 dst_addr, &id->route.addr.dev_addr,
2832 timeout_ms, addr_handler, id_priv);
2840 cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
2841 cma_deref_id(id_priv);
2844 EXPORT_SYMBOL(rdma_resolve_addr);
2846 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse)
2848 struct rdma_id_private *id_priv;
2849 unsigned long flags;
2852 id_priv = container_of(id, struct rdma_id_private, id);
2853 spin_lock_irqsave(&id_priv->lock, flags);
2854 if (reuse || id_priv->state == RDMA_CM_IDLE) {
2855 id_priv->reuseaddr = reuse;
2860 spin_unlock_irqrestore(&id_priv->lock, flags);
2863 EXPORT_SYMBOL(rdma_set_reuseaddr);
2865 int rdma_set_afonly(struct rdma_cm_id *id, int afonly)
2867 struct rdma_id_private *id_priv;
2868 unsigned long flags;
2871 id_priv = container_of(id, struct rdma_id_private, id);
2872 spin_lock_irqsave(&id_priv->lock, flags);
2873 if (id_priv->state == RDMA_CM_IDLE || id_priv->state == RDMA_CM_ADDR_BOUND) {
2874 id_priv->options |= (1 << CMA_OPTION_AFONLY);
2875 id_priv->afonly = afonly;
2880 spin_unlock_irqrestore(&id_priv->lock, flags);
2883 EXPORT_SYMBOL(rdma_set_afonly);
2885 static void cma_bind_port(struct rdma_bind_list *bind_list,
2886 struct rdma_id_private *id_priv)
2888 struct sockaddr *addr;
2889 struct sockaddr_ib *sib;
2893 addr = cma_src_addr(id_priv);
2894 port = htons(bind_list->port);
2896 switch (addr->sa_family) {
2898 ((struct sockaddr_in *) addr)->sin_port = port;
2901 ((struct sockaddr_in6 *) addr)->sin6_port = port;
2904 sib = (struct sockaddr_ib *) addr;
2905 sid = be64_to_cpu(sib->sib_sid);
2906 mask = be64_to_cpu(sib->sib_sid_mask);
2907 sib->sib_sid = cpu_to_be64((sid & mask) | (u64) ntohs(port));
2908 sib->sib_sid_mask = cpu_to_be64(~0ULL);
2911 id_priv->bind_list = bind_list;
2912 hlist_add_head(&id_priv->node, &bind_list->owners);
2915 static int cma_alloc_port(enum rdma_port_space ps,
2916 struct rdma_id_private *id_priv, unsigned short snum)
2918 struct rdma_bind_list *bind_list;
2921 bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
2925 ret = cma_ps_alloc(id_priv->id.route.addr.dev_addr.net, ps, bind_list,
2931 bind_list->port = (unsigned short)ret;
2932 cma_bind_port(bind_list, id_priv);
2936 return ret == -ENOSPC ? -EADDRNOTAVAIL : ret;
2939 static int cma_alloc_any_port(enum rdma_port_space ps,
2940 struct rdma_id_private *id_priv)
2942 static unsigned int last_used_port;
2943 int low, high, remaining;
2945 struct vnet *net = id_priv->id.route.addr.dev_addr.net;
2948 inet_get_local_port_range(net, &low, &high);
2949 remaining = (high - low) + 1;
2950 get_random_bytes(&rand, sizeof(rand));
2951 rover = rand % remaining + low;
2953 if (last_used_port != rover &&
2954 !cma_ps_find(net, ps, (unsigned short)rover)) {
2955 int ret = cma_alloc_port(ps, id_priv, rover);
2957 * Remember previously used port number in order to avoid
2958 * re-using same port immediately after it is closed.
2961 last_used_port = rover;
2962 if (ret != -EADDRNOTAVAIL)
2967 if ((rover < low) || (rover > high))
2971 return -EADDRNOTAVAIL;
2975 * Check that the requested port is available. This is called when trying to
2976 * bind to a specific port, or when trying to listen on a bound port. In
2977 * the latter case, the provided id_priv may already be on the bind_list, but
2978 * we still need to check that it's okay to start listening.
2980 static int cma_check_port(struct rdma_bind_list *bind_list,
2981 struct rdma_id_private *id_priv, uint8_t reuseaddr)
2983 struct rdma_id_private *cur_id;
2984 struct sockaddr *addr, *cur_addr;
2986 addr = cma_src_addr(id_priv);
2987 hlist_for_each_entry(cur_id, &bind_list->owners, node) {
2988 if (id_priv == cur_id)
2991 if ((cur_id->state != RDMA_CM_LISTEN) && reuseaddr &&
2995 cur_addr = cma_src_addr(cur_id);
2996 if (id_priv->afonly && cur_id->afonly &&
2997 (addr->sa_family != cur_addr->sa_family))
3000 if (cma_any_addr(addr) || cma_any_addr(cur_addr))
3001 return -EADDRNOTAVAIL;
3003 if (!cma_addr_cmp(addr, cur_addr))
3009 static int cma_use_port(enum rdma_port_space ps,
3010 struct rdma_id_private *id_priv)
3012 struct rdma_bind_list *bind_list;
3013 unsigned short snum;
3016 snum = ntohs(cma_port(cma_src_addr(id_priv)));
3017 if (snum < IPPORT_RESERVED &&
3018 priv_check(curthread, PRIV_NETINET_BINDANY) != 0)
3021 bind_list = cma_ps_find(id_priv->id.route.addr.dev_addr.net, ps, snum);
3023 ret = cma_alloc_port(ps, id_priv, snum);
3025 ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr);
3027 cma_bind_port(bind_list, id_priv);
3032 static int cma_bind_listen(struct rdma_id_private *id_priv)
3034 struct rdma_bind_list *bind_list = id_priv->bind_list;
3038 if (bind_list->owners.first->next)
3039 ret = cma_check_port(bind_list, id_priv, 0);
3040 mutex_unlock(&lock);
3044 static enum rdma_port_space cma_select_inet_ps(
3045 struct rdma_id_private *id_priv)
3047 switch (id_priv->id.ps) {
3052 return id_priv->id.ps;
3059 static enum rdma_port_space cma_select_ib_ps(struct rdma_id_private *id_priv)
3061 enum rdma_port_space ps = 0;
3062 struct sockaddr_ib *sib;
3063 u64 sid_ps, mask, sid;
3065 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
3066 mask = be64_to_cpu(sib->sib_sid_mask) & RDMA_IB_IP_PS_MASK;
3067 sid = be64_to_cpu(sib->sib_sid) & mask;
3069 if ((id_priv->id.ps == RDMA_PS_IB) && (sid == (RDMA_IB_IP_PS_IB & mask))) {
3070 sid_ps = RDMA_IB_IP_PS_IB;
3072 } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_TCP)) &&
3073 (sid == (RDMA_IB_IP_PS_TCP & mask))) {
3074 sid_ps = RDMA_IB_IP_PS_TCP;
3076 } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_UDP)) &&
3077 (sid == (RDMA_IB_IP_PS_UDP & mask))) {
3078 sid_ps = RDMA_IB_IP_PS_UDP;
3083 sib->sib_sid = cpu_to_be64(sid_ps | ntohs(cma_port((struct sockaddr *) sib)));
3084 sib->sib_sid_mask = cpu_to_be64(RDMA_IB_IP_PS_MASK |
3085 be64_to_cpu(sib->sib_sid_mask));
3090 static int cma_get_port(struct rdma_id_private *id_priv)
3092 enum rdma_port_space ps;
3095 if (cma_family(id_priv) != AF_IB)
3096 ps = cma_select_inet_ps(id_priv);
3098 ps = cma_select_ib_ps(id_priv);
3100 return -EPROTONOSUPPORT;
3103 if (cma_any_port(cma_src_addr(id_priv)))
3104 ret = cma_alloc_any_port(ps, id_priv);
3106 ret = cma_use_port(ps, id_priv);
3107 mutex_unlock(&lock);
3112 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
3113 struct sockaddr *addr)
3116 struct sockaddr_in6 sin6;
3118 if (addr->sa_family != AF_INET6)
3121 sin6 = *(struct sockaddr_in6 *)addr;
3123 if (IN6_IS_SCOPE_LINKLOCAL(&sin6.sin6_addr) ||
3124 IN6_IS_ADDR_MC_INTFACELOCAL(&sin6.sin6_addr)) {
3127 CURVNET_SET_QUIET(dev_addr->net);
3128 failure = sa6_recoverscope(&sin6) || sin6.sin6_scope_id == 0;
3131 /* check if IPv6 scope ID is not set */
3134 dev_addr->bound_dev_if = sin6.sin6_scope_id;
3140 int rdma_listen(struct rdma_cm_id *id, int backlog)
3142 struct rdma_id_private *id_priv;
3145 id_priv = container_of(id, struct rdma_id_private, id);
3146 if (id_priv->state == RDMA_CM_IDLE) {
3147 id->route.addr.src_addr.ss_family = AF_INET;
3148 ret = rdma_bind_addr(id, cma_src_addr(id_priv));
3153 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN))
3156 if (id_priv->reuseaddr) {
3157 ret = cma_bind_listen(id_priv);
3162 id_priv->backlog = backlog;
3164 if (rdma_cap_ib_cm(id->device, 1)) {
3165 ret = cma_ib_listen(id_priv);
3168 } else if (rdma_cap_iw_cm(id->device, 1)) {
3169 ret = cma_iw_listen(id_priv, backlog);
3177 cma_listen_on_all(id_priv);
3181 id_priv->backlog = 0;
3182 cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND);
3185 EXPORT_SYMBOL(rdma_listen);
3187 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
3189 struct rdma_id_private *id_priv;
3192 if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6 &&
3193 addr->sa_family != AF_IB)
3194 return -EAFNOSUPPORT;
3196 id_priv = container_of(id, struct rdma_id_private, id);
3197 if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND))
3200 ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
3204 memcpy(cma_src_addr(id_priv), addr, rdma_addr_size(addr));
3205 if (!cma_any_addr(addr)) {
3206 ret = cma_translate_addr(addr, &id->route.addr.dev_addr);
3210 ret = cma_acquire_dev(id_priv, NULL);
3215 if (!(id_priv->options & (1 << CMA_OPTION_AFONLY))) {
3216 if (addr->sa_family == AF_INET)
3217 id_priv->afonly = 1;
3219 else if (addr->sa_family == AF_INET6) {
3220 CURVNET_SET_QUIET(id_priv->id.route.addr.dev_addr.net);
3221 id_priv->afonly = V_ip6_v6only;
3226 ret = cma_get_port(id_priv);
3232 if (id_priv->cma_dev)
3233 cma_release_dev(id_priv);
3235 cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE);
3238 EXPORT_SYMBOL(rdma_bind_addr);
3240 static int cma_format_hdr(void *hdr, struct rdma_id_private *id_priv)
3242 struct cma_hdr *cma_hdr;
3245 cma_hdr->cma_version = CMA_VERSION;
3246 if (cma_family(id_priv) == AF_INET) {
3247 struct sockaddr_in *src4, *dst4;
3249 src4 = (struct sockaddr_in *) cma_src_addr(id_priv);
3250 dst4 = (struct sockaddr_in *) cma_dst_addr(id_priv);
3252 cma_set_ip_ver(cma_hdr, 4);
3253 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
3254 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
3255 cma_hdr->port = src4->sin_port;
3256 } else if (cma_family(id_priv) == AF_INET6) {
3257 struct sockaddr_in6 *src6, *dst6;
3259 src6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
3260 dst6 = (struct sockaddr_in6 *) cma_dst_addr(id_priv);
3262 cma_set_ip_ver(cma_hdr, 6);
3263 cma_hdr->src_addr.ip6 = src6->sin6_addr;
3264 cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
3265 cma_hdr->port = src6->sin6_port;
3266 cma_ip6_clear_scope_id(&cma_hdr->src_addr.ip6);
3267 cma_ip6_clear_scope_id(&cma_hdr->dst_addr.ip6);
3272 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
3273 struct ib_cm_event *ib_event)
3275 struct rdma_id_private *id_priv = cm_id->context;
3276 struct rdma_cm_event event;
3277 struct ib_cm_sidr_rep_event_param *rep = &ib_event->param.sidr_rep_rcvd;
3280 mutex_lock(&id_priv->handler_mutex);
3281 if (id_priv->state != RDMA_CM_CONNECT)
3284 memset(&event, 0, sizeof event);
3285 switch (ib_event->event) {
3286 case IB_CM_SIDR_REQ_ERROR:
3287 event.event = RDMA_CM_EVENT_UNREACHABLE;
3288 event.status = -ETIMEDOUT;
3290 case IB_CM_SIDR_REP_RECEIVED:
3291 event.param.ud.private_data = ib_event->private_data;
3292 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
3293 if (rep->status != IB_SIDR_SUCCESS) {
3294 event.event = RDMA_CM_EVENT_UNREACHABLE;
3295 event.status = ib_event->param.sidr_rep_rcvd.status;
3298 ret = cma_set_qkey(id_priv, rep->qkey);
3300 event.event = RDMA_CM_EVENT_ADDR_ERROR;
3304 ret = ib_init_ah_from_path(id_priv->id.device,
3305 id_priv->id.port_num,
3306 id_priv->id.route.path_rec,
3307 &event.param.ud.ah_attr);
3309 event.event = RDMA_CM_EVENT_ADDR_ERROR;
3313 event.param.ud.qp_num = rep->qpn;
3314 event.param.ud.qkey = rep->qkey;
3315 event.event = RDMA_CM_EVENT_ESTABLISHED;
3319 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
3324 ret = id_priv->id.event_handler(&id_priv->id, &event);
3326 /* Destroy the CM ID by returning a non-zero value. */
3327 id_priv->cm_id.ib = NULL;
3328 cma_exch(id_priv, RDMA_CM_DESTROYING);
3329 mutex_unlock(&id_priv->handler_mutex);
3330 rdma_destroy_id(&id_priv->id);
3334 mutex_unlock(&id_priv->handler_mutex);
3338 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
3339 struct rdma_conn_param *conn_param)
3341 struct ib_cm_sidr_req_param req;
3342 struct ib_cm_id *id;
3346 memset(&req, 0, sizeof req);
3347 offset = cma_user_data_offset(id_priv);
3348 req.private_data_len = offset + conn_param->private_data_len;
3349 if (req.private_data_len < conn_param->private_data_len)
3352 if (req.private_data_len) {
3353 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
3357 private_data = NULL;
3360 if (conn_param->private_data && conn_param->private_data_len)
3361 memcpy((char *)private_data + offset, conn_param->private_data,
3362 conn_param->private_data_len);
3365 ret = cma_format_hdr(private_data, id_priv);
3368 req.private_data = private_data;
3371 id = ib_create_cm_id(id_priv->id.device, cma_sidr_rep_handler,
3377 id_priv->cm_id.ib = id;
3379 req.path = id_priv->id.route.path_rec;
3380 req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
3381 req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
3382 req.max_cm_retries = CMA_MAX_CM_RETRIES;
3384 ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
3386 ib_destroy_cm_id(id_priv->cm_id.ib);
3387 id_priv->cm_id.ib = NULL;
3390 kfree(private_data);
3394 static int cma_connect_ib(struct rdma_id_private *id_priv,
3395 struct rdma_conn_param *conn_param)
3397 struct ib_cm_req_param req;
3398 struct rdma_route *route;
3400 struct ib_cm_id *id;
3403 memset(&req, 0, sizeof req);
3404 offset = cma_user_data_offset(id_priv);
3405 req.private_data_len = offset + conn_param->private_data_len;
3406 if (req.private_data_len < conn_param->private_data_len)
3409 if (req.private_data_len) {
3410 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
3414 private_data = NULL;
3417 if (conn_param->private_data && conn_param->private_data_len)
3418 memcpy((char *)private_data + offset, conn_param->private_data,
3419 conn_param->private_data_len);
3421 id = ib_create_cm_id(id_priv->id.device, cma_ib_handler, id_priv);
3426 id_priv->cm_id.ib = id;
3428 route = &id_priv->id.route;
3430 ret = cma_format_hdr(private_data, id_priv);
3433 req.private_data = private_data;
3436 req.primary_path = &route->path_rec[0];
3437 if (route->num_paths == 2)
3438 req.alternate_path = &route->path_rec[1];
3440 req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
3441 req.qp_num = id_priv->qp_num;
3442 req.qp_type = id_priv->id.qp_type;
3443 req.starting_psn = id_priv->seq_num;
3444 req.responder_resources = conn_param->responder_resources;
3445 req.initiator_depth = conn_param->initiator_depth;
3446 req.flow_control = conn_param->flow_control;
3447 req.retry_count = min_t(u8, 7, conn_param->retry_count);
3448 req.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
3449 req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
3450 req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
3451 req.max_cm_retries = CMA_MAX_CM_RETRIES;
3452 req.srq = id_priv->srq ? 1 : 0;
3454 ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
3456 if (ret && !IS_ERR(id)) {
3457 ib_destroy_cm_id(id);
3458 id_priv->cm_id.ib = NULL;
3461 kfree(private_data);
3465 static int cma_connect_iw(struct rdma_id_private *id_priv,
3466 struct rdma_conn_param *conn_param)
3468 struct iw_cm_id *cm_id;
3470 struct iw_cm_conn_param iw_param;
3472 cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
3474 return PTR_ERR(cm_id);
3476 cm_id->tos = id_priv->tos;
3477 id_priv->cm_id.iw = cm_id;
3479 memcpy(&cm_id->local_addr, cma_src_addr(id_priv),
3480 rdma_addr_size(cma_src_addr(id_priv)));
3481 memcpy(&cm_id->remote_addr, cma_dst_addr(id_priv),
3482 rdma_addr_size(cma_dst_addr(id_priv)));
3484 ret = cma_modify_qp_rtr(id_priv, conn_param);
3489 iw_param.ord = conn_param->initiator_depth;
3490 iw_param.ird = conn_param->responder_resources;
3491 iw_param.private_data = conn_param->private_data;
3492 iw_param.private_data_len = conn_param->private_data_len;
3493 iw_param.qpn = id_priv->id.qp ? id_priv->qp_num : conn_param->qp_num;
3495 memset(&iw_param, 0, sizeof iw_param);
3496 iw_param.qpn = id_priv->qp_num;
3498 ret = iw_cm_connect(cm_id, &iw_param);
3501 iw_destroy_cm_id(cm_id);
3502 id_priv->cm_id.iw = NULL;
3507 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
3509 struct rdma_id_private *id_priv;
3512 id_priv = container_of(id, struct rdma_id_private, id);
3513 if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT))
3517 id_priv->qp_num = conn_param->qp_num;
3518 id_priv->srq = conn_param->srq;
3521 if (rdma_cap_ib_cm(id->device, id->port_num)) {
3522 if (id->qp_type == IB_QPT_UD)
3523 ret = cma_resolve_ib_udp(id_priv, conn_param);
3525 ret = cma_connect_ib(id_priv, conn_param);
3526 } else if (rdma_cap_iw_cm(id->device, id->port_num))
3527 ret = cma_connect_iw(id_priv, conn_param);
3535 cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED);
3538 EXPORT_SYMBOL(rdma_connect);
3540 static int cma_accept_ib(struct rdma_id_private *id_priv,
3541 struct rdma_conn_param *conn_param)
3543 struct ib_cm_rep_param rep;
3546 ret = cma_modify_qp_rtr(id_priv, conn_param);
3550 ret = cma_modify_qp_rts(id_priv, conn_param);
3554 memset(&rep, 0, sizeof rep);
3555 rep.qp_num = id_priv->qp_num;
3556 rep.starting_psn = id_priv->seq_num;
3557 rep.private_data = conn_param->private_data;
3558 rep.private_data_len = conn_param->private_data_len;
3559 rep.responder_resources = conn_param->responder_resources;
3560 rep.initiator_depth = conn_param->initiator_depth;
3561 rep.failover_accepted = 0;
3562 rep.flow_control = conn_param->flow_control;
3563 rep.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
3564 rep.srq = id_priv->srq ? 1 : 0;
3566 ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
3571 static int cma_accept_iw(struct rdma_id_private *id_priv,
3572 struct rdma_conn_param *conn_param)
3574 struct iw_cm_conn_param iw_param;
3577 ret = cma_modify_qp_rtr(id_priv, conn_param);
3581 iw_param.ord = conn_param->initiator_depth;
3582 iw_param.ird = conn_param->responder_resources;
3583 iw_param.private_data = conn_param->private_data;
3584 iw_param.private_data_len = conn_param->private_data_len;
3585 if (id_priv->id.qp) {
3586 iw_param.qpn = id_priv->qp_num;
3588 iw_param.qpn = conn_param->qp_num;
3590 return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
3593 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
3594 enum ib_cm_sidr_status status, u32 qkey,
3595 const void *private_data, int private_data_len)
3597 struct ib_cm_sidr_rep_param rep;
3600 memset(&rep, 0, sizeof rep);
3601 rep.status = status;
3602 if (status == IB_SIDR_SUCCESS) {
3603 ret = cma_set_qkey(id_priv, qkey);
3606 rep.qp_num = id_priv->qp_num;
3607 rep.qkey = id_priv->qkey;
3609 rep.private_data = private_data;
3610 rep.private_data_len = private_data_len;
3612 return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
3615 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
3617 struct rdma_id_private *id_priv;
3620 id_priv = container_of(id, struct rdma_id_private, id);
3622 id_priv->owner = task_pid_nr(current);
3624 if (!cma_comp(id_priv, RDMA_CM_CONNECT))
3627 if (!id->qp && conn_param) {
3628 id_priv->qp_num = conn_param->qp_num;
3629 id_priv->srq = conn_param->srq;
3632 if (rdma_cap_ib_cm(id->device, id->port_num)) {
3633 if (id->qp_type == IB_QPT_UD) {
3635 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
3637 conn_param->private_data,
3638 conn_param->private_data_len);
3640 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
3644 ret = cma_accept_ib(id_priv, conn_param);
3646 ret = cma_rep_recv(id_priv);
3648 } else if (rdma_cap_iw_cm(id->device, id->port_num))
3649 ret = cma_accept_iw(id_priv, conn_param);
3658 cma_modify_qp_err(id_priv);
3659 rdma_reject(id, NULL, 0);
3662 EXPORT_SYMBOL(rdma_accept);
3664 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
3666 struct rdma_id_private *id_priv;
3669 id_priv = container_of(id, struct rdma_id_private, id);
3670 if (!id_priv->cm_id.ib)
3673 switch (id->device->node_type) {
3674 case RDMA_NODE_IB_CA:
3675 ret = ib_cm_notify(id_priv->cm_id.ib, event);
3683 EXPORT_SYMBOL(rdma_notify);
3685 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
3686 u8 private_data_len)
3688 struct rdma_id_private *id_priv;
3691 id_priv = container_of(id, struct rdma_id_private, id);
3692 if (!id_priv->cm_id.ib)
3695 if (rdma_cap_ib_cm(id->device, id->port_num)) {
3696 if (id->qp_type == IB_QPT_UD)
3697 ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT, 0,
3698 private_data, private_data_len);
3700 ret = ib_send_cm_rej(id_priv->cm_id.ib,
3701 IB_CM_REJ_CONSUMER_DEFINED, NULL,
3702 0, private_data, private_data_len);
3703 } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
3704 ret = iw_cm_reject(id_priv->cm_id.iw,
3705 private_data, private_data_len);
3711 EXPORT_SYMBOL(rdma_reject);
3713 int rdma_disconnect(struct rdma_cm_id *id)
3715 struct rdma_id_private *id_priv;
3718 id_priv = container_of(id, struct rdma_id_private, id);
3719 if (!id_priv->cm_id.ib)
3722 if (rdma_cap_ib_cm(id->device, id->port_num)) {
3723 ret = cma_modify_qp_err(id_priv);
3726 /* Initiate or respond to a disconnect. */
3727 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
3728 ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
3729 } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
3730 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
3737 EXPORT_SYMBOL(rdma_disconnect);
3739 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
3741 struct rdma_id_private *id_priv;
3742 struct cma_multicast *mc = multicast->context;
3743 struct rdma_cm_event event;
3746 id_priv = mc->id_priv;
3747 mutex_lock(&id_priv->handler_mutex);
3748 if (id_priv->state != RDMA_CM_ADDR_BOUND &&
3749 id_priv->state != RDMA_CM_ADDR_RESOLVED)
3753 status = cma_set_qkey(id_priv, be32_to_cpu(multicast->rec.qkey));
3754 mutex_lock(&id_priv->qp_mutex);
3755 if (!status && id_priv->id.qp)
3756 status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid,
3757 be16_to_cpu(multicast->rec.mlid));
3758 mutex_unlock(&id_priv->qp_mutex);
3760 memset(&event, 0, sizeof event);
3761 event.status = status;
3762 event.param.ud.private_data = mc->context;
3764 struct rdma_dev_addr *dev_addr =
3765 &id_priv->id.route.addr.dev_addr;
3766 struct net_device *ndev =
3767 dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if);
3768 enum ib_gid_type gid_type =
3769 id_priv->cma_dev->default_gid_type[id_priv->id.port_num -
3770 rdma_start_port(id_priv->cma_dev->device)];
3772 event.event = RDMA_CM_EVENT_MULTICAST_JOIN;
3773 ib_init_ah_from_mcmember(id_priv->id.device,
3774 id_priv->id.port_num, &multicast->rec,
3776 &event.param.ud.ah_attr);
3777 event.param.ud.qp_num = 0xFFFFFF;
3778 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
3782 event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
3784 ret = id_priv->id.event_handler(&id_priv->id, &event);
3786 cma_exch(id_priv, RDMA_CM_DESTROYING);
3787 mutex_unlock(&id_priv->handler_mutex);
3788 rdma_destroy_id(&id_priv->id);
3793 mutex_unlock(&id_priv->handler_mutex);
3797 static void cma_set_mgid(struct rdma_id_private *id_priv,
3798 struct sockaddr *addr, union ib_gid *mgid)
3800 unsigned char mc_map[MAX_ADDR_LEN];
3801 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3802 struct sockaddr_in *sin = (struct sockaddr_in *) addr;
3803 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
3805 if (cma_any_addr(addr)) {
3806 memset(mgid, 0, sizeof *mgid);
3807 } else if ((addr->sa_family == AF_INET6) &&
3808 ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
3810 /* IPv6 address is an SA assigned MGID. */
3811 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3812 } else if (addr->sa_family == AF_IB) {
3813 memcpy(mgid, &((struct sockaddr_ib *) addr)->sib_addr, sizeof *mgid);
3814 } else if (addr->sa_family == AF_INET6) {
3815 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
3816 if (id_priv->id.ps == RDMA_PS_UDP)
3817 mc_map[7] = 0x01; /* Use RDMA CM signature */
3818 *mgid = *(union ib_gid *) (mc_map + 4);
3820 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
3821 if (id_priv->id.ps == RDMA_PS_UDP)
3822 mc_map[7] = 0x01; /* Use RDMA CM signature */
3823 *mgid = *(union ib_gid *) (mc_map + 4);
3827 static void cma_query_sa_classport_info_cb(int status,
3828 struct ib_class_port_info *rec,
3831 struct class_port_info_context *cb_ctx = context;
3835 if (status || !rec) {
3836 pr_debug("RDMA CM: %s port %u failed query ClassPortInfo status: %d\n",
3837 cb_ctx->device->name, cb_ctx->port_num, status);
3841 memcpy(cb_ctx->class_port_info, rec, sizeof(struct ib_class_port_info));
3844 complete(&cb_ctx->done);
3847 static int cma_query_sa_classport_info(struct ib_device *device, u8 port_num,
3848 struct ib_class_port_info *class_port_info)
3850 struct class_port_info_context *cb_ctx;
3853 cb_ctx = kmalloc(sizeof(*cb_ctx), GFP_KERNEL);
3857 cb_ctx->device = device;
3858 cb_ctx->class_port_info = class_port_info;
3859 cb_ctx->port_num = port_num;
3860 init_completion(&cb_ctx->done);
3862 ret = ib_sa_classport_info_rec_query(&sa_client, device, port_num,
3863 CMA_QUERY_CLASSPORT_INFO_TIMEOUT,
3864 GFP_KERNEL, cma_query_sa_classport_info_cb,
3865 cb_ctx, &cb_ctx->sa_query);
3867 pr_err("RDMA CM: %s port %u failed to send ClassPortInfo query, ret: %d\n",
3868 device->name, port_num, ret);
3872 wait_for_completion(&cb_ctx->done);
3879 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
3880 struct cma_multicast *mc)
3882 struct ib_sa_mcmember_rec rec;
3883 struct ib_class_port_info class_port_info;
3884 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3885 ib_sa_comp_mask comp_mask;
3888 ib_addr_get_mgid(dev_addr, &rec.mgid);
3889 ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
3894 ret = cma_set_qkey(id_priv, 0);
3898 cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
3899 rec.qkey = cpu_to_be32(id_priv->qkey);
3900 rdma_addr_get_sgid(dev_addr, &rec.port_gid);
3901 rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
3902 rec.join_state = mc->join_state;
3904 if (rec.join_state == BIT(SENDONLY_FULLMEMBER_JOIN)) {
3905 ret = cma_query_sa_classport_info(id_priv->id.device,
3906 id_priv->id.port_num,
3912 if (!(ib_get_cpi_capmask2(&class_port_info) &
3913 IB_SA_CAP_MASK2_SENDONLY_FULL_MEM_SUPPORT)) {
3914 pr_warn("RDMA CM: %s port %u Unable to multicast join\n"
3915 "RDMA CM: SM doesn't support Send Only Full Member option\n",
3916 id_priv->id.device->name, id_priv->id.port_num);
3921 comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
3922 IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
3923 IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
3924 IB_SA_MCMEMBER_REC_FLOW_LABEL |
3925 IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
3927 if (id_priv->id.ps == RDMA_PS_IPOIB)
3928 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
3929 IB_SA_MCMEMBER_REC_RATE_SELECTOR |
3930 IB_SA_MCMEMBER_REC_MTU_SELECTOR |
3931 IB_SA_MCMEMBER_REC_MTU |
3932 IB_SA_MCMEMBER_REC_HOP_LIMIT;
3934 mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device,
3935 id_priv->id.port_num, &rec,
3936 comp_mask, GFP_KERNEL,
3937 cma_ib_mc_handler, mc);
3938 return PTR_ERR_OR_ZERO(mc->multicast.ib);
3941 static void iboe_mcast_work_handler(struct work_struct *work)
3943 struct iboe_mcast_work *mw = container_of(work, struct iboe_mcast_work, work);
3944 struct cma_multicast *mc = mw->mc;
3945 struct ib_sa_multicast *m = mc->multicast.ib;
3947 mc->multicast.ib->context = mc;
3948 cma_ib_mc_handler(0, m);
3949 kref_put(&mc->mcref, release_mc);
3953 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid)
3955 struct sockaddr_in *sin = (struct sockaddr_in *)addr;
3956 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
3958 if (cma_any_addr(addr)) {
3959 memset(mgid, 0, sizeof *mgid);
3960 } else if (addr->sa_family == AF_INET6) {
3961 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3963 mgid->raw[0] = 0xff;
3964 mgid->raw[1] = 0x0e;
3973 mgid->raw[10] = 0xff;
3974 mgid->raw[11] = 0xff;
3975 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
3979 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
3980 struct cma_multicast *mc)
3982 struct iboe_mcast_work *work;
3983 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3985 struct sockaddr *addr = (struct sockaddr *)&mc->addr;
3986 struct net_device *ndev = NULL;
3987 enum ib_gid_type gid_type;
3990 send_only = mc->join_state == BIT(SENDONLY_FULLMEMBER_JOIN);
3992 if (cma_zero_addr((struct sockaddr *)&mc->addr))
3995 work = kzalloc(sizeof *work, GFP_KERNEL);
3999 mc->multicast.ib = kzalloc(sizeof(struct ib_sa_multicast), GFP_KERNEL);
4000 if (!mc->multicast.ib) {
4005 cma_iboe_set_mgid(addr, &mc->multicast.ib->rec.mgid);
4007 mc->multicast.ib->rec.pkey = cpu_to_be16(0xffff);
4008 if (id_priv->id.ps == RDMA_PS_UDP)
4009 mc->multicast.ib->rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
4011 if (dev_addr->bound_dev_if)
4012 ndev = dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if);
4017 mc->multicast.ib->rec.rate = iboe_get_rate(ndev);
4018 mc->multicast.ib->rec.hop_limit = 1;
4019 mc->multicast.ib->rec.mtu = iboe_get_mtu(ndev->if_mtu);
4021 gid_type = id_priv->cma_dev->default_gid_type[id_priv->id.port_num -
4022 rdma_start_port(id_priv->cma_dev->device)];
4023 if (addr->sa_family == AF_INET) {
4024 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) {
4025 mc->multicast.ib->rec.hop_limit = IPV6_DEFAULT_HOPLIMIT;
4027 mc->igmp_joined = true;
4031 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP)
4035 if (err || !mc->multicast.ib->rec.mtu) {
4040 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
4041 &mc->multicast.ib->rec.port_gid);
4044 INIT_WORK(&work->work, iboe_mcast_work_handler);
4045 kref_get(&mc->mcref);
4046 queue_work(cma_wq, &work->work);
4051 kfree(mc->multicast.ib);
4057 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
4058 u8 join_state, void *context)
4060 struct rdma_id_private *id_priv;
4061 struct cma_multicast *mc;
4064 id_priv = container_of(id, struct rdma_id_private, id);
4065 if (!cma_comp(id_priv, RDMA_CM_ADDR_BOUND) &&
4066 !cma_comp(id_priv, RDMA_CM_ADDR_RESOLVED))
4069 mc = kmalloc(sizeof *mc, GFP_KERNEL);
4073 memcpy(&mc->addr, addr, rdma_addr_size(addr));
4074 mc->context = context;
4075 mc->id_priv = id_priv;
4076 mc->igmp_joined = false;
4077 mc->join_state = join_state;
4078 spin_lock(&id_priv->lock);
4079 list_add(&mc->list, &id_priv->mc_list);
4080 spin_unlock(&id_priv->lock);
4082 if (rdma_protocol_roce(id->device, id->port_num)) {
4083 kref_init(&mc->mcref);
4084 ret = cma_iboe_join_multicast(id_priv, mc);
4085 } else if (rdma_cap_ib_mcast(id->device, id->port_num))
4086 ret = cma_join_ib_multicast(id_priv, mc);
4091 spin_lock_irq(&id_priv->lock);
4092 list_del(&mc->list);
4093 spin_unlock_irq(&id_priv->lock);
4098 EXPORT_SYMBOL(rdma_join_multicast);
4100 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
4102 struct rdma_id_private *id_priv;
4103 struct cma_multicast *mc;
4105 id_priv = container_of(id, struct rdma_id_private, id);
4106 spin_lock_irq(&id_priv->lock);
4107 list_for_each_entry(mc, &id_priv->mc_list, list) {
4108 if (!memcmp(&mc->addr, addr, rdma_addr_size(addr))) {
4109 list_del(&mc->list);
4110 spin_unlock_irq(&id_priv->lock);
4113 ib_detach_mcast(id->qp,
4114 &mc->multicast.ib->rec.mgid,
4115 be16_to_cpu(mc->multicast.ib->rec.mlid));
4117 BUG_ON(id_priv->cma_dev->device != id->device);
4119 if (rdma_cap_ib_mcast(id->device, id->port_num)) {
4120 ib_sa_free_multicast(mc->multicast.ib);
4122 } else if (rdma_protocol_roce(id->device, id->port_num)) {
4123 if (mc->igmp_joined) {
4124 struct rdma_dev_addr *dev_addr =
4125 &id->route.addr.dev_addr;
4126 struct net_device *ndev = NULL;
4128 if (dev_addr->bound_dev_if)
4129 ndev = dev_get_by_index(dev_addr->net,
4130 dev_addr->bound_dev_if);
4134 mc->igmp_joined = false;
4136 kref_put(&mc->mcref, release_mc);
4141 spin_unlock_irq(&id_priv->lock);
4143 EXPORT_SYMBOL(rdma_leave_multicast);
4146 sysctl_cma_default_roce_mode(SYSCTL_HANDLER_ARGS)
4148 struct cma_device *cma_dev = arg1;
4149 const int port = arg2;
4153 strlcpy(buf, ib_cache_gid_type_str(
4154 cma_get_default_gid_type(cma_dev, port)), sizeof(buf));
4156 error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
4157 if (error != 0 || req->newptr == NULL)
4160 error = ib_cache_gid_parse_type_str(buf);
4166 cma_set_default_gid_type(cma_dev, port, error);
4172 static void cma_add_one(struct ib_device *device)
4174 struct cma_device *cma_dev;
4175 struct rdma_id_private *id_priv;
4178 cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
4182 sysctl_ctx_init(&cma_dev->sysctl_ctx);
4184 cma_dev->device = device;
4185 cma_dev->default_gid_type = kcalloc(device->phys_port_cnt,
4186 sizeof(*cma_dev->default_gid_type),
4188 if (!cma_dev->default_gid_type) {
4192 for (i = rdma_start_port(device); i <= rdma_end_port(device); i++) {
4193 unsigned long supported_gids;
4194 unsigned int default_gid_type;
4196 supported_gids = roce_gid_type_mask_support(device, i);
4198 if (WARN_ON(!supported_gids)) {
4199 /* set something valid */
4200 default_gid_type = 0;
4201 } else if (test_bit(IB_GID_TYPE_ROCE_UDP_ENCAP, &supported_gids)) {
4202 /* prefer RoCEv2, if supported */
4203 default_gid_type = IB_GID_TYPE_ROCE_UDP_ENCAP;
4205 default_gid_type = find_first_bit(&supported_gids,
4208 cma_dev->default_gid_type[i - rdma_start_port(device)] =
4212 init_completion(&cma_dev->comp);
4213 atomic_set(&cma_dev->refcount, 1);
4214 INIT_LIST_HEAD(&cma_dev->id_list);
4215 ib_set_client_data(device, &cma_client, cma_dev);
4218 list_add_tail(&cma_dev->list, &dev_list);
4219 list_for_each_entry(id_priv, &listen_any_list, list)
4220 cma_listen_on_dev(id_priv, cma_dev);
4221 mutex_unlock(&lock);
4223 for (i = rdma_start_port(device); i <= rdma_end_port(device); i++) {
4226 snprintf(buf, sizeof(buf), "default_roce_mode_port%d", i);
4228 (void) SYSCTL_ADD_PROC(&cma_dev->sysctl_ctx,
4229 SYSCTL_CHILDREN(device->ports_parent->parent->oidp),
4230 OID_AUTO, buf, CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_MPSAFE,
4231 cma_dev, i, &sysctl_cma_default_roce_mode, "A",
4232 "Default RoCE mode. Valid values: IB/RoCE v1 and RoCE v2");
4236 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
4238 struct rdma_cm_event event;
4239 enum rdma_cm_state state;
4242 /* Record that we want to remove the device */
4243 state = cma_exch(id_priv, RDMA_CM_DEVICE_REMOVAL);
4244 if (state == RDMA_CM_DESTROYING)
4247 cma_cancel_operation(id_priv, state);
4248 mutex_lock(&id_priv->handler_mutex);
4250 /* Check for destruction from another callback. */
4251 if (!cma_comp(id_priv, RDMA_CM_DEVICE_REMOVAL))
4254 memset(&event, 0, sizeof event);
4255 event.event = RDMA_CM_EVENT_DEVICE_REMOVAL;
4256 ret = id_priv->id.event_handler(&id_priv->id, &event);
4258 mutex_unlock(&id_priv->handler_mutex);
4262 static void cma_process_remove(struct cma_device *cma_dev)
4264 struct rdma_id_private *id_priv;
4268 while (!list_empty(&cma_dev->id_list)) {
4269 id_priv = list_entry(cma_dev->id_list.next,
4270 struct rdma_id_private, list);
4272 list_del(&id_priv->listen_list);
4273 list_del_init(&id_priv->list);
4274 atomic_inc(&id_priv->refcount);
4275 mutex_unlock(&lock);
4277 ret = id_priv->internal_id ? 1 : cma_remove_id_dev(id_priv);
4278 cma_deref_id(id_priv);
4280 rdma_destroy_id(&id_priv->id);
4284 mutex_unlock(&lock);
4286 cma_deref_dev(cma_dev);
4287 wait_for_completion(&cma_dev->comp);
4290 static void cma_remove_one(struct ib_device *device, void *client_data)
4292 struct cma_device *cma_dev = client_data;
4298 list_del(&cma_dev->list);
4299 mutex_unlock(&lock);
4301 cma_process_remove(cma_dev);
4302 sysctl_ctx_free(&cma_dev->sysctl_ctx);
4303 kfree(cma_dev->default_gid_type);
4307 static void cma_init_vnet(void *arg)
4309 struct cma_pernet *pernet = &VNET(cma_pernet);
4311 idr_init(&pernet->tcp_ps);
4312 idr_init(&pernet->udp_ps);
4313 idr_init(&pernet->ipoib_ps);
4314 idr_init(&pernet->ib_ps);
4316 VNET_SYSINIT(cma_init_vnet, SI_SUB_OFED_MODINIT - 1, SI_ORDER_FIRST, cma_init_vnet, NULL);
4318 static void cma_destroy_vnet(void *arg)
4320 struct cma_pernet *pernet = &VNET(cma_pernet);
4322 idr_destroy(&pernet->tcp_ps);
4323 idr_destroy(&pernet->udp_ps);
4324 idr_destroy(&pernet->ipoib_ps);
4325 idr_destroy(&pernet->ib_ps);
4327 VNET_SYSUNINIT(cma_destroy_vnet, SI_SUB_OFED_MODINIT - 1, SI_ORDER_SECOND, cma_destroy_vnet, NULL);
4329 static int __init cma_init(void)
4333 cma_wq = alloc_ordered_workqueue("rdma_cm", WQ_MEM_RECLAIM);
4337 ib_sa_register_client(&sa_client);
4338 rdma_addr_register_client(&addr_client);
4340 ret = ib_register_client(&cma_client);
4344 cma_configfs_init();
4349 rdma_addr_unregister_client(&addr_client);
4350 ib_sa_unregister_client(&sa_client);
4351 destroy_workqueue(cma_wq);
4355 static void __exit cma_cleanup(void)
4357 cma_configfs_exit();
4358 ib_unregister_client(&cma_client);
4359 rdma_addr_unregister_client(&addr_client);
4360 ib_sa_unregister_client(&sa_client);
4361 destroy_workqueue(cma_wq);
4364 module_init(cma_init);
4365 module_exit(cma_cleanup);