1 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
3 * Copyright (c) 2005 Voltaire Inc. All rights reserved.
4 * Copyright (c) 2002-2005, Network Appliance, Inc. All rights reserved.
5 * Copyright (c) 1999-2019, Mellanox Technologies, Inc. All rights reserved.
6 * Copyright (c) 2005-2006 Intel Corporation. All rights reserved.
9 #include <linux/completion.h>
11 #include <linux/in6.h>
12 #include <linux/mutex.h>
13 #include <linux/random.h>
14 #include <linux/igmp.h>
15 #include <linux/xarray.h>
16 #include <linux/inetdevice.h>
17 #include <linux/slab.h>
18 #include <linux/module.h>
19 #include <net/route.h>
21 #include <net/net_namespace.h>
22 #include <net/netns/generic.h>
25 #include <net/ip_fib.h>
26 #include <net/ip6_route.h>
28 #include <rdma/rdma_cm.h>
29 #include <rdma/rdma_cm_ib.h>
30 #include <rdma/rdma_netlink.h>
32 #include <rdma/ib_cache.h>
33 #include <rdma/ib_cm.h>
34 #include <rdma/ib_sa.h>
35 #include <rdma/iw_cm.h>
37 #include "core_priv.h"
39 #include "cma_trace.h"
41 MODULE_AUTHOR("Sean Hefty");
42 MODULE_DESCRIPTION("Generic RDMA CM Agent");
43 MODULE_LICENSE("Dual BSD/GPL");
45 #define CMA_CM_RESPONSE_TIMEOUT 20
46 #define CMA_MAX_CM_RETRIES 15
47 #define CMA_CM_MRA_SETTING (IB_CM_MRA_FLAG_DELAY | 24)
48 #define CMA_IBOE_PACKET_LIFETIME 18
49 #define CMA_PREFERRED_ROCE_GID_TYPE IB_GID_TYPE_ROCE_UDP_ENCAP
51 static const char * const cma_events[] = {
52 [RDMA_CM_EVENT_ADDR_RESOLVED] = "address resolved",
53 [RDMA_CM_EVENT_ADDR_ERROR] = "address error",
54 [RDMA_CM_EVENT_ROUTE_RESOLVED] = "route resolved ",
55 [RDMA_CM_EVENT_ROUTE_ERROR] = "route error",
56 [RDMA_CM_EVENT_CONNECT_REQUEST] = "connect request",
57 [RDMA_CM_EVENT_CONNECT_RESPONSE] = "connect response",
58 [RDMA_CM_EVENT_CONNECT_ERROR] = "connect error",
59 [RDMA_CM_EVENT_UNREACHABLE] = "unreachable",
60 [RDMA_CM_EVENT_REJECTED] = "rejected",
61 [RDMA_CM_EVENT_ESTABLISHED] = "established",
62 [RDMA_CM_EVENT_DISCONNECTED] = "disconnected",
63 [RDMA_CM_EVENT_DEVICE_REMOVAL] = "device removal",
64 [RDMA_CM_EVENT_MULTICAST_JOIN] = "multicast join",
65 [RDMA_CM_EVENT_MULTICAST_ERROR] = "multicast error",
66 [RDMA_CM_EVENT_ADDR_CHANGE] = "address change",
67 [RDMA_CM_EVENT_TIMEWAIT_EXIT] = "timewait exit",
70 static void cma_set_mgid(struct rdma_id_private *id_priv, struct sockaddr *addr,
73 const char *__attribute_const__ rdma_event_msg(enum rdma_cm_event_type event)
77 return (index < ARRAY_SIZE(cma_events) && cma_events[index]) ?
78 cma_events[index] : "unrecognized event";
80 EXPORT_SYMBOL(rdma_event_msg);
82 const char *__attribute_const__ rdma_reject_msg(struct rdma_cm_id *id,
85 if (rdma_ib_or_roce(id->device, id->port_num))
86 return ibcm_reject_msg(reason);
88 if (rdma_protocol_iwarp(id->device, id->port_num))
89 return iwcm_reject_msg(reason);
92 return "unrecognized transport";
94 EXPORT_SYMBOL(rdma_reject_msg);
97 * rdma_is_consumer_reject - return true if the consumer rejected the connect
99 * @id: Communication identifier that received the REJECT event.
100 * @reason: Value returned in the REJECT event status field.
102 static bool rdma_is_consumer_reject(struct rdma_cm_id *id, int reason)
104 if (rdma_ib_or_roce(id->device, id->port_num))
105 return reason == IB_CM_REJ_CONSUMER_DEFINED;
107 if (rdma_protocol_iwarp(id->device, id->port_num))
108 return reason == -ECONNREFUSED;
114 const void *rdma_consumer_reject_data(struct rdma_cm_id *id,
115 struct rdma_cm_event *ev, u8 *data_len)
119 if (rdma_is_consumer_reject(id, ev->status)) {
120 *data_len = ev->param.conn.private_data_len;
121 p = ev->param.conn.private_data;
128 EXPORT_SYMBOL(rdma_consumer_reject_data);
131 * rdma_iw_cm_id() - return the iw_cm_id pointer for this cm_id.
132 * @id: Communication Identifier
134 struct iw_cm_id *rdma_iw_cm_id(struct rdma_cm_id *id)
136 struct rdma_id_private *id_priv;
138 id_priv = container_of(id, struct rdma_id_private, id);
139 if (id->device->node_type == RDMA_NODE_RNIC)
140 return id_priv->cm_id.iw;
143 EXPORT_SYMBOL(rdma_iw_cm_id);
146 * rdma_res_to_id() - return the rdma_cm_id pointer for this restrack.
147 * @res: rdma resource tracking entry pointer
149 struct rdma_cm_id *rdma_res_to_id(struct rdma_restrack_entry *res)
151 struct rdma_id_private *id_priv =
152 container_of(res, struct rdma_id_private, res);
156 EXPORT_SYMBOL(rdma_res_to_id);
158 static int cma_add_one(struct ib_device *device);
159 static void cma_remove_one(struct ib_device *device, void *client_data);
161 static struct ib_client cma_client = {
164 .remove = cma_remove_one
167 static struct ib_sa_client sa_client;
168 static LIST_HEAD(dev_list);
169 static LIST_HEAD(listen_any_list);
170 static DEFINE_MUTEX(lock);
171 static struct workqueue_struct *cma_wq;
172 static unsigned int cma_pernet_id;
175 struct xarray tcp_ps;
176 struct xarray udp_ps;
177 struct xarray ipoib_ps;
181 static struct cma_pernet *cma_pernet(struct net *net)
183 return net_generic(net, cma_pernet_id);
187 struct xarray *cma_pernet_xa(struct net *net, enum rdma_ucm_port_space ps)
189 struct cma_pernet *pernet = cma_pernet(net);
193 return &pernet->tcp_ps;
195 return &pernet->udp_ps;
197 return &pernet->ipoib_ps;
199 return &pernet->ib_ps;
206 struct list_head list;
207 struct ib_device *device;
208 struct completion comp;
210 struct list_head id_list;
211 enum ib_gid_type *default_gid_type;
212 u8 *default_roce_tos;
215 struct rdma_bind_list {
216 enum rdma_ucm_port_space ps;
217 struct hlist_head owners;
221 static int cma_ps_alloc(struct net *net, enum rdma_ucm_port_space ps,
222 struct rdma_bind_list *bind_list, int snum)
224 struct xarray *xa = cma_pernet_xa(net, ps);
226 return xa_insert(xa, snum, bind_list, GFP_KERNEL);
229 static struct rdma_bind_list *cma_ps_find(struct net *net,
230 enum rdma_ucm_port_space ps, int snum)
232 struct xarray *xa = cma_pernet_xa(net, ps);
234 return xa_load(xa, snum);
237 static void cma_ps_remove(struct net *net, enum rdma_ucm_port_space ps,
240 struct xarray *xa = cma_pernet_xa(net, ps);
249 void cma_dev_get(struct cma_device *cma_dev)
251 refcount_inc(&cma_dev->refcount);
254 void cma_dev_put(struct cma_device *cma_dev)
256 if (refcount_dec_and_test(&cma_dev->refcount))
257 complete(&cma_dev->comp);
260 struct cma_device *cma_enum_devices_by_ibdev(cma_device_filter filter,
263 struct cma_device *cma_dev;
264 struct cma_device *found_cma_dev = NULL;
268 list_for_each_entry(cma_dev, &dev_list, list)
269 if (filter(cma_dev->device, cookie)) {
270 found_cma_dev = cma_dev;
275 cma_dev_get(found_cma_dev);
277 return found_cma_dev;
280 int cma_get_default_gid_type(struct cma_device *cma_dev,
283 if (!rdma_is_port_valid(cma_dev->device, port))
286 return cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)];
289 int cma_set_default_gid_type(struct cma_device *cma_dev,
291 enum ib_gid_type default_gid_type)
293 unsigned long supported_gids;
295 if (!rdma_is_port_valid(cma_dev->device, port))
298 if (default_gid_type == IB_GID_TYPE_IB &&
299 rdma_protocol_roce_eth_encap(cma_dev->device, port))
300 default_gid_type = IB_GID_TYPE_ROCE;
302 supported_gids = roce_gid_type_mask_support(cma_dev->device, port);
304 if (!(supported_gids & 1 << default_gid_type))
307 cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)] =
313 int cma_get_default_roce_tos(struct cma_device *cma_dev, u32 port)
315 if (!rdma_is_port_valid(cma_dev->device, port))
318 return cma_dev->default_roce_tos[port - rdma_start_port(cma_dev->device)];
321 int cma_set_default_roce_tos(struct cma_device *cma_dev, u32 port,
324 if (!rdma_is_port_valid(cma_dev->device, port))
327 cma_dev->default_roce_tos[port - rdma_start_port(cma_dev->device)] =
332 struct ib_device *cma_get_ib_dev(struct cma_device *cma_dev)
334 return cma_dev->device;
338 * Device removal can occur at anytime, so we need extra handling to
339 * serialize notifying the user of device removal with other callbacks.
340 * We do this by disabling removal notification while a callback is in process,
341 * and reporting it after the callback completes.
344 struct cma_multicast {
345 struct rdma_id_private *id_priv;
347 struct ib_sa_multicast *sa_mc;
349 struct work_struct work;
350 struct rdma_cm_event event;
353 struct list_head list;
355 struct sockaddr_storage addr;
360 struct work_struct work;
361 struct rdma_id_private *id;
362 enum rdma_cm_state old_state;
363 enum rdma_cm_state new_state;
364 struct rdma_cm_event event;
377 u8 ip_version; /* IP version: 7:4 */
379 union cma_ip_addr src_addr;
380 union cma_ip_addr dst_addr;
383 #define CMA_VERSION 0x00
385 struct cma_req_info {
386 struct sockaddr_storage listen_addr_storage;
387 struct sockaddr_storage src_addr_storage;
388 struct ib_device *device;
389 union ib_gid local_gid;
396 static int cma_comp_exch(struct rdma_id_private *id_priv,
397 enum rdma_cm_state comp, enum rdma_cm_state exch)
403 * The FSM uses a funny double locking where state is protected by both
404 * the handler_mutex and the spinlock. State is not allowed to change
405 * to/from a handler_mutex protected value without also holding
408 if (comp == RDMA_CM_CONNECT || exch == RDMA_CM_CONNECT)
409 lockdep_assert_held(&id_priv->handler_mutex);
411 spin_lock_irqsave(&id_priv->lock, flags);
412 if ((ret = (id_priv->state == comp)))
413 id_priv->state = exch;
414 spin_unlock_irqrestore(&id_priv->lock, flags);
418 static inline u8 cma_get_ip_ver(const struct cma_hdr *hdr)
420 return hdr->ip_version >> 4;
423 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
425 hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
428 static int cma_igmp_send(struct net_device *ndev, union ib_gid *mgid, bool join)
430 struct in_device *in_dev = NULL;
434 in_dev = __in_dev_get_rtnl(ndev);
437 ip_mc_inc_group(in_dev,
438 *(__be32 *)(mgid->raw + 12));
440 ip_mc_dec_group(in_dev,
441 *(__be32 *)(mgid->raw + 12));
445 return (in_dev) ? 0 : -ENODEV;
448 static void _cma_attach_to_dev(struct rdma_id_private *id_priv,
449 struct cma_device *cma_dev)
451 cma_dev_get(cma_dev);
452 id_priv->cma_dev = cma_dev;
453 id_priv->id.device = cma_dev->device;
454 id_priv->id.route.addr.dev_addr.transport =
455 rdma_node_get_transport(cma_dev->device->node_type);
456 list_add_tail(&id_priv->list, &cma_dev->id_list);
458 trace_cm_id_attach(id_priv, cma_dev->device);
461 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
462 struct cma_device *cma_dev)
464 _cma_attach_to_dev(id_priv, cma_dev);
466 cma_dev->default_gid_type[id_priv->id.port_num -
467 rdma_start_port(cma_dev->device)];
470 static void cma_release_dev(struct rdma_id_private *id_priv)
473 list_del(&id_priv->list);
474 cma_dev_put(id_priv->cma_dev);
475 id_priv->cma_dev = NULL;
476 id_priv->id.device = NULL;
477 if (id_priv->id.route.addr.dev_addr.sgid_attr) {
478 rdma_put_gid_attr(id_priv->id.route.addr.dev_addr.sgid_attr);
479 id_priv->id.route.addr.dev_addr.sgid_attr = NULL;
484 static inline struct sockaddr *cma_src_addr(struct rdma_id_private *id_priv)
486 return (struct sockaddr *) &id_priv->id.route.addr.src_addr;
489 static inline struct sockaddr *cma_dst_addr(struct rdma_id_private *id_priv)
491 return (struct sockaddr *) &id_priv->id.route.addr.dst_addr;
494 static inline unsigned short cma_family(struct rdma_id_private *id_priv)
496 return id_priv->id.route.addr.src_addr.ss_family;
499 static int cma_set_qkey(struct rdma_id_private *id_priv, u32 qkey)
501 struct ib_sa_mcmember_rec rec;
505 if (qkey && id_priv->qkey != qkey)
511 id_priv->qkey = qkey;
515 switch (id_priv->id.ps) {
518 id_priv->qkey = RDMA_UDP_QKEY;
521 ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid);
522 ret = ib_sa_get_mcmember_rec(id_priv->id.device,
523 id_priv->id.port_num, &rec.mgid,
526 id_priv->qkey = be32_to_cpu(rec.qkey);
534 static void cma_translate_ib(struct sockaddr_ib *sib, struct rdma_dev_addr *dev_addr)
536 dev_addr->dev_type = ARPHRD_INFINIBAND;
537 rdma_addr_set_sgid(dev_addr, (union ib_gid *) &sib->sib_addr);
538 ib_addr_set_pkey(dev_addr, ntohs(sib->sib_pkey));
541 static int cma_translate_addr(struct sockaddr *addr, struct rdma_dev_addr *dev_addr)
545 if (addr->sa_family != AF_IB) {
546 ret = rdma_translate_ip(addr, dev_addr);
548 cma_translate_ib((struct sockaddr_ib *) addr, dev_addr);
555 static const struct ib_gid_attr *
556 cma_validate_port(struct ib_device *device, u32 port,
557 enum ib_gid_type gid_type,
559 struct rdma_id_private *id_priv)
561 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
562 int bound_if_index = dev_addr->bound_dev_if;
563 const struct ib_gid_attr *sgid_attr;
564 int dev_type = dev_addr->dev_type;
565 struct net_device *ndev = NULL;
567 if (!rdma_dev_access_netns(device, id_priv->id.route.addr.dev_addr.net))
568 return ERR_PTR(-ENODEV);
570 if ((dev_type == ARPHRD_INFINIBAND) && !rdma_protocol_ib(device, port))
571 return ERR_PTR(-ENODEV);
573 if ((dev_type != ARPHRD_INFINIBAND) && rdma_protocol_ib(device, port))
574 return ERR_PTR(-ENODEV);
576 if (dev_type == ARPHRD_ETHER && rdma_protocol_roce(device, port)) {
577 ndev = dev_get_by_index(dev_addr->net, bound_if_index);
579 return ERR_PTR(-ENODEV);
581 gid_type = IB_GID_TYPE_IB;
584 sgid_attr = rdma_find_gid_by_port(device, gid, gid_type, port, ndev);
590 static void cma_bind_sgid_attr(struct rdma_id_private *id_priv,
591 const struct ib_gid_attr *sgid_attr)
593 WARN_ON(id_priv->id.route.addr.dev_addr.sgid_attr);
594 id_priv->id.route.addr.dev_addr.sgid_attr = sgid_attr;
598 * cma_acquire_dev_by_src_ip - Acquire cma device, port, gid attribute
599 * based on source ip address.
600 * @id_priv: cm_id which should be bound to cma device
602 * cma_acquire_dev_by_src_ip() binds cm id to cma device, port and GID attribute
603 * based on source IP address. It returns 0 on success or error code otherwise.
604 * It is applicable to active and passive side cm_id.
606 static int cma_acquire_dev_by_src_ip(struct rdma_id_private *id_priv)
608 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
609 const struct ib_gid_attr *sgid_attr;
610 union ib_gid gid, iboe_gid, *gidp;
611 struct cma_device *cma_dev;
612 enum ib_gid_type gid_type;
616 if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
617 id_priv->id.ps == RDMA_PS_IPOIB)
620 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
623 memcpy(&gid, dev_addr->src_dev_addr +
624 rdma_addr_gid_offset(dev_addr), sizeof(gid));
627 list_for_each_entry(cma_dev, &dev_list, list) {
628 rdma_for_each_port (cma_dev->device, port) {
629 gidp = rdma_protocol_roce(cma_dev->device, port) ?
631 gid_type = cma_dev->default_gid_type[port - 1];
632 sgid_attr = cma_validate_port(cma_dev->device, port,
633 gid_type, gidp, id_priv);
634 if (!IS_ERR(sgid_attr)) {
635 id_priv->id.port_num = port;
636 cma_bind_sgid_attr(id_priv, sgid_attr);
637 cma_attach_to_dev(id_priv, cma_dev);
649 * cma_ib_acquire_dev - Acquire cma device, port and SGID attribute
650 * @id_priv: cm id to bind to cma device
651 * @listen_id_priv: listener cm id to match against
652 * @req: Pointer to req structure containaining incoming
653 * request information
654 * cma_ib_acquire_dev() acquires cma device, port and SGID attribute when
655 * rdma device matches for listen_id and incoming request. It also verifies
656 * that a GID table entry is present for the source address.
657 * Returns 0 on success, or returns error code otherwise.
659 static int cma_ib_acquire_dev(struct rdma_id_private *id_priv,
660 const struct rdma_id_private *listen_id_priv,
661 struct cma_req_info *req)
663 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
664 const struct ib_gid_attr *sgid_attr;
665 enum ib_gid_type gid_type;
668 if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
669 id_priv->id.ps == RDMA_PS_IPOIB)
672 if (rdma_protocol_roce(req->device, req->port))
673 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
676 memcpy(&gid, dev_addr->src_dev_addr +
677 rdma_addr_gid_offset(dev_addr), sizeof(gid));
679 gid_type = listen_id_priv->cma_dev->default_gid_type[req->port - 1];
680 sgid_attr = cma_validate_port(req->device, req->port,
681 gid_type, &gid, id_priv);
682 if (IS_ERR(sgid_attr))
683 return PTR_ERR(sgid_attr);
685 id_priv->id.port_num = req->port;
686 cma_bind_sgid_attr(id_priv, sgid_attr);
687 /* Need to acquire lock to protect against reader
688 * of cma_dev->id_list such as cma_netdev_callback() and
689 * cma_process_remove().
692 cma_attach_to_dev(id_priv, listen_id_priv->cma_dev);
694 rdma_restrack_add(&id_priv->res);
698 static int cma_iw_acquire_dev(struct rdma_id_private *id_priv,
699 const struct rdma_id_private *listen_id_priv)
701 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
702 const struct ib_gid_attr *sgid_attr;
703 struct cma_device *cma_dev;
704 enum ib_gid_type gid_type;
709 if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
710 id_priv->id.ps == RDMA_PS_IPOIB)
713 memcpy(&gid, dev_addr->src_dev_addr +
714 rdma_addr_gid_offset(dev_addr), sizeof(gid));
718 cma_dev = listen_id_priv->cma_dev;
719 port = listen_id_priv->id.port_num;
720 gid_type = listen_id_priv->gid_type;
721 sgid_attr = cma_validate_port(cma_dev->device, port,
722 gid_type, &gid, id_priv);
723 if (!IS_ERR(sgid_attr)) {
724 id_priv->id.port_num = port;
725 cma_bind_sgid_attr(id_priv, sgid_attr);
730 list_for_each_entry(cma_dev, &dev_list, list) {
731 rdma_for_each_port (cma_dev->device, port) {
732 if (listen_id_priv->cma_dev == cma_dev &&
733 listen_id_priv->id.port_num == port)
736 gid_type = cma_dev->default_gid_type[port - 1];
737 sgid_attr = cma_validate_port(cma_dev->device, port,
738 gid_type, &gid, id_priv);
739 if (!IS_ERR(sgid_attr)) {
740 id_priv->id.port_num = port;
741 cma_bind_sgid_attr(id_priv, sgid_attr);
750 cma_attach_to_dev(id_priv, cma_dev);
751 rdma_restrack_add(&id_priv->res);
759 * Select the source IB device and address to reach the destination IB address.
761 static int cma_resolve_ib_dev(struct rdma_id_private *id_priv)
763 struct cma_device *cma_dev, *cur_dev;
764 struct sockaddr_ib *addr;
765 union ib_gid gid, sgid, *dgid;
768 enum ib_port_state port_state;
772 addr = (struct sockaddr_ib *) cma_dst_addr(id_priv);
773 dgid = (union ib_gid *) &addr->sib_addr;
774 pkey = ntohs(addr->sib_pkey);
777 list_for_each_entry(cur_dev, &dev_list, list) {
778 rdma_for_each_port (cur_dev->device, p) {
779 if (!rdma_cap_af_ib(cur_dev->device, p))
782 if (ib_find_cached_pkey(cur_dev->device, p, pkey, &index))
785 if (ib_get_cached_port_state(cur_dev->device, p, &port_state))
787 for (i = 0; !rdma_query_gid(cur_dev->device,
790 if (!memcmp(&gid, dgid, sizeof(gid))) {
793 id_priv->id.port_num = p;
797 if (!cma_dev && (gid.global.subnet_prefix ==
798 dgid->global.subnet_prefix) &&
799 port_state == IB_PORT_ACTIVE) {
802 id_priv->id.port_num = p;
812 cma_attach_to_dev(id_priv, cma_dev);
813 rdma_restrack_add(&id_priv->res);
815 addr = (struct sockaddr_ib *)cma_src_addr(id_priv);
816 memcpy(&addr->sib_addr, &sgid, sizeof(sgid));
817 cma_translate_ib(addr, &id_priv->id.route.addr.dev_addr);
821 static void cma_id_get(struct rdma_id_private *id_priv)
823 refcount_inc(&id_priv->refcount);
826 static void cma_id_put(struct rdma_id_private *id_priv)
828 if (refcount_dec_and_test(&id_priv->refcount))
829 complete(&id_priv->comp);
832 static struct rdma_id_private *
833 __rdma_create_id(struct net *net, rdma_cm_event_handler event_handler,
834 void *context, enum rdma_ucm_port_space ps,
835 enum ib_qp_type qp_type, const struct rdma_id_private *parent)
837 struct rdma_id_private *id_priv;
839 id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
841 return ERR_PTR(-ENOMEM);
843 id_priv->state = RDMA_CM_IDLE;
844 id_priv->id.context = context;
845 id_priv->id.event_handler = event_handler;
847 id_priv->id.qp_type = qp_type;
848 id_priv->tos_set = false;
849 id_priv->timeout_set = false;
850 id_priv->min_rnr_timer_set = false;
851 id_priv->gid_type = IB_GID_TYPE_IB;
852 spin_lock_init(&id_priv->lock);
853 mutex_init(&id_priv->qp_mutex);
854 init_completion(&id_priv->comp);
855 refcount_set(&id_priv->refcount, 1);
856 mutex_init(&id_priv->handler_mutex);
857 INIT_LIST_HEAD(&id_priv->listen_list);
858 INIT_LIST_HEAD(&id_priv->mc_list);
859 get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
860 id_priv->id.route.addr.dev_addr.net = get_net(net);
861 id_priv->seq_num &= 0x00ffffff;
863 rdma_restrack_new(&id_priv->res, RDMA_RESTRACK_CM_ID);
865 rdma_restrack_parent_name(&id_priv->res, &parent->res);
871 __rdma_create_kernel_id(struct net *net, rdma_cm_event_handler event_handler,
872 void *context, enum rdma_ucm_port_space ps,
873 enum ib_qp_type qp_type, const char *caller)
875 struct rdma_id_private *ret;
877 ret = __rdma_create_id(net, event_handler, context, ps, qp_type, NULL);
879 return ERR_CAST(ret);
881 rdma_restrack_set_name(&ret->res, caller);
884 EXPORT_SYMBOL(__rdma_create_kernel_id);
886 struct rdma_cm_id *rdma_create_user_id(rdma_cm_event_handler event_handler,
888 enum rdma_ucm_port_space ps,
889 enum ib_qp_type qp_type)
891 struct rdma_id_private *ret;
893 ret = __rdma_create_id(current->nsproxy->net_ns, event_handler, context,
896 return ERR_CAST(ret);
898 rdma_restrack_set_name(&ret->res, NULL);
901 EXPORT_SYMBOL(rdma_create_user_id);
903 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
905 struct ib_qp_attr qp_attr;
906 int qp_attr_mask, ret;
908 qp_attr.qp_state = IB_QPS_INIT;
909 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
913 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
917 qp_attr.qp_state = IB_QPS_RTR;
918 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
922 qp_attr.qp_state = IB_QPS_RTS;
924 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
929 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
931 struct ib_qp_attr qp_attr;
932 int qp_attr_mask, ret;
934 qp_attr.qp_state = IB_QPS_INIT;
935 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
939 return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
942 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
943 struct ib_qp_init_attr *qp_init_attr)
945 struct rdma_id_private *id_priv;
949 id_priv = container_of(id, struct rdma_id_private, id);
950 if (id->device != pd->device) {
955 qp_init_attr->port_num = id->port_num;
956 qp = ib_create_qp(pd, qp_init_attr);
962 if (id->qp_type == IB_QPT_UD)
963 ret = cma_init_ud_qp(id_priv, qp);
965 ret = cma_init_conn_qp(id_priv, qp);
970 id_priv->qp_num = qp->qp_num;
971 id_priv->srq = (qp->srq != NULL);
972 trace_cm_qp_create(id_priv, pd, qp_init_attr, 0);
977 trace_cm_qp_create(id_priv, pd, qp_init_attr, ret);
980 EXPORT_SYMBOL(rdma_create_qp);
982 void rdma_destroy_qp(struct rdma_cm_id *id)
984 struct rdma_id_private *id_priv;
986 id_priv = container_of(id, struct rdma_id_private, id);
987 trace_cm_qp_destroy(id_priv);
988 mutex_lock(&id_priv->qp_mutex);
989 ib_destroy_qp(id_priv->id.qp);
990 id_priv->id.qp = NULL;
991 mutex_unlock(&id_priv->qp_mutex);
993 EXPORT_SYMBOL(rdma_destroy_qp);
995 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
996 struct rdma_conn_param *conn_param)
998 struct ib_qp_attr qp_attr;
999 int qp_attr_mask, ret;
1001 mutex_lock(&id_priv->qp_mutex);
1002 if (!id_priv->id.qp) {
1007 /* Need to update QP attributes from default values. */
1008 qp_attr.qp_state = IB_QPS_INIT;
1009 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1013 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
1017 qp_attr.qp_state = IB_QPS_RTR;
1018 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1022 BUG_ON(id_priv->cma_dev->device != id_priv->id.device);
1025 qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
1026 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
1028 mutex_unlock(&id_priv->qp_mutex);
1032 static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
1033 struct rdma_conn_param *conn_param)
1035 struct ib_qp_attr qp_attr;
1036 int qp_attr_mask, ret;
1038 mutex_lock(&id_priv->qp_mutex);
1039 if (!id_priv->id.qp) {
1044 qp_attr.qp_state = IB_QPS_RTS;
1045 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1050 qp_attr.max_rd_atomic = conn_param->initiator_depth;
1051 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
1053 mutex_unlock(&id_priv->qp_mutex);
1057 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
1059 struct ib_qp_attr qp_attr;
1062 mutex_lock(&id_priv->qp_mutex);
1063 if (!id_priv->id.qp) {
1068 qp_attr.qp_state = IB_QPS_ERR;
1069 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
1071 mutex_unlock(&id_priv->qp_mutex);
1075 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
1076 struct ib_qp_attr *qp_attr, int *qp_attr_mask)
1078 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
1082 if (rdma_cap_eth_ah(id_priv->id.device, id_priv->id.port_num))
1085 pkey = ib_addr_get_pkey(dev_addr);
1087 ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
1088 pkey, &qp_attr->pkey_index);
1092 qp_attr->port_num = id_priv->id.port_num;
1093 *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
1095 if (id_priv->id.qp_type == IB_QPT_UD) {
1096 ret = cma_set_qkey(id_priv, 0);
1100 qp_attr->qkey = id_priv->qkey;
1101 *qp_attr_mask |= IB_QP_QKEY;
1103 qp_attr->qp_access_flags = 0;
1104 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
1109 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
1112 struct rdma_id_private *id_priv;
1115 id_priv = container_of(id, struct rdma_id_private, id);
1116 if (rdma_cap_ib_cm(id->device, id->port_num)) {
1117 if (!id_priv->cm_id.ib || (id_priv->id.qp_type == IB_QPT_UD))
1118 ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
1120 ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
1123 if (qp_attr->qp_state == IB_QPS_RTR)
1124 qp_attr->rq_psn = id_priv->seq_num;
1125 } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
1126 if (!id_priv->cm_id.iw) {
1127 qp_attr->qp_access_flags = 0;
1128 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
1130 ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
1132 qp_attr->port_num = id_priv->id.port_num;
1133 *qp_attr_mask |= IB_QP_PORT;
1138 if ((*qp_attr_mask & IB_QP_TIMEOUT) && id_priv->timeout_set)
1139 qp_attr->timeout = id_priv->timeout;
1141 if ((*qp_attr_mask & IB_QP_MIN_RNR_TIMER) && id_priv->min_rnr_timer_set)
1142 qp_attr->min_rnr_timer = id_priv->min_rnr_timer;
1146 EXPORT_SYMBOL(rdma_init_qp_attr);
1148 static inline bool cma_zero_addr(const struct sockaddr *addr)
1150 switch (addr->sa_family) {
1152 return ipv4_is_zeronet(((struct sockaddr_in *)addr)->sin_addr.s_addr);
1154 return ipv6_addr_any(&((struct sockaddr_in6 *)addr)->sin6_addr);
1156 return ib_addr_any(&((struct sockaddr_ib *)addr)->sib_addr);
1162 static inline bool cma_loopback_addr(const struct sockaddr *addr)
1164 switch (addr->sa_family) {
1166 return ipv4_is_loopback(
1167 ((struct sockaddr_in *)addr)->sin_addr.s_addr);
1169 return ipv6_addr_loopback(
1170 &((struct sockaddr_in6 *)addr)->sin6_addr);
1172 return ib_addr_loopback(
1173 &((struct sockaddr_ib *)addr)->sib_addr);
1179 static inline bool cma_any_addr(const struct sockaddr *addr)
1181 return cma_zero_addr(addr) || cma_loopback_addr(addr);
1184 static int cma_addr_cmp(const struct sockaddr *src, const struct sockaddr *dst)
1186 if (src->sa_family != dst->sa_family)
1189 switch (src->sa_family) {
1191 return ((struct sockaddr_in *)src)->sin_addr.s_addr !=
1192 ((struct sockaddr_in *)dst)->sin_addr.s_addr;
1194 struct sockaddr_in6 *src_addr6 = (struct sockaddr_in6 *)src;
1195 struct sockaddr_in6 *dst_addr6 = (struct sockaddr_in6 *)dst;
1198 if (ipv6_addr_cmp(&src_addr6->sin6_addr,
1199 &dst_addr6->sin6_addr))
1201 link_local = ipv6_addr_type(&dst_addr6->sin6_addr) &
1202 IPV6_ADDR_LINKLOCAL;
1203 /* Link local must match their scope_ids */
1204 return link_local ? (src_addr6->sin6_scope_id !=
1205 dst_addr6->sin6_scope_id) :
1210 return ib_addr_cmp(&((struct sockaddr_ib *) src)->sib_addr,
1211 &((struct sockaddr_ib *) dst)->sib_addr);
1215 static __be16 cma_port(const struct sockaddr *addr)
1217 struct sockaddr_ib *sib;
1219 switch (addr->sa_family) {
1221 return ((struct sockaddr_in *) addr)->sin_port;
1223 return ((struct sockaddr_in6 *) addr)->sin6_port;
1225 sib = (struct sockaddr_ib *) addr;
1226 return htons((u16) (be64_to_cpu(sib->sib_sid) &
1227 be64_to_cpu(sib->sib_sid_mask)));
1233 static inline int cma_any_port(const struct sockaddr *addr)
1235 return !cma_port(addr);
1238 static void cma_save_ib_info(struct sockaddr *src_addr,
1239 struct sockaddr *dst_addr,
1240 const struct rdma_cm_id *listen_id,
1241 const struct sa_path_rec *path)
1243 struct sockaddr_ib *listen_ib, *ib;
1245 listen_ib = (struct sockaddr_ib *) &listen_id->route.addr.src_addr;
1247 ib = (struct sockaddr_ib *)src_addr;
1248 ib->sib_family = AF_IB;
1250 ib->sib_pkey = path->pkey;
1251 ib->sib_flowinfo = path->flow_label;
1252 memcpy(&ib->sib_addr, &path->sgid, 16);
1253 ib->sib_sid = path->service_id;
1254 ib->sib_scope_id = 0;
1256 ib->sib_pkey = listen_ib->sib_pkey;
1257 ib->sib_flowinfo = listen_ib->sib_flowinfo;
1258 ib->sib_addr = listen_ib->sib_addr;
1259 ib->sib_sid = listen_ib->sib_sid;
1260 ib->sib_scope_id = listen_ib->sib_scope_id;
1262 ib->sib_sid_mask = cpu_to_be64(0xffffffffffffffffULL);
1265 ib = (struct sockaddr_ib *)dst_addr;
1266 ib->sib_family = AF_IB;
1268 ib->sib_pkey = path->pkey;
1269 ib->sib_flowinfo = path->flow_label;
1270 memcpy(&ib->sib_addr, &path->dgid, 16);
1275 static void cma_save_ip4_info(struct sockaddr_in *src_addr,
1276 struct sockaddr_in *dst_addr,
1277 struct cma_hdr *hdr,
1281 *src_addr = (struct sockaddr_in) {
1282 .sin_family = AF_INET,
1283 .sin_addr.s_addr = hdr->dst_addr.ip4.addr,
1284 .sin_port = local_port,
1289 *dst_addr = (struct sockaddr_in) {
1290 .sin_family = AF_INET,
1291 .sin_addr.s_addr = hdr->src_addr.ip4.addr,
1292 .sin_port = hdr->port,
1297 static void cma_save_ip6_info(struct sockaddr_in6 *src_addr,
1298 struct sockaddr_in6 *dst_addr,
1299 struct cma_hdr *hdr,
1303 *src_addr = (struct sockaddr_in6) {
1304 .sin6_family = AF_INET6,
1305 .sin6_addr = hdr->dst_addr.ip6,
1306 .sin6_port = local_port,
1311 *dst_addr = (struct sockaddr_in6) {
1312 .sin6_family = AF_INET6,
1313 .sin6_addr = hdr->src_addr.ip6,
1314 .sin6_port = hdr->port,
1319 static u16 cma_port_from_service_id(__be64 service_id)
1321 return (u16)be64_to_cpu(service_id);
1324 static int cma_save_ip_info(struct sockaddr *src_addr,
1325 struct sockaddr *dst_addr,
1326 const struct ib_cm_event *ib_event,
1329 struct cma_hdr *hdr;
1332 hdr = ib_event->private_data;
1333 if (hdr->cma_version != CMA_VERSION)
1336 port = htons(cma_port_from_service_id(service_id));
1338 switch (cma_get_ip_ver(hdr)) {
1340 cma_save_ip4_info((struct sockaddr_in *)src_addr,
1341 (struct sockaddr_in *)dst_addr, hdr, port);
1344 cma_save_ip6_info((struct sockaddr_in6 *)src_addr,
1345 (struct sockaddr_in6 *)dst_addr, hdr, port);
1348 return -EAFNOSUPPORT;
1354 static int cma_save_net_info(struct sockaddr *src_addr,
1355 struct sockaddr *dst_addr,
1356 const struct rdma_cm_id *listen_id,
1357 const struct ib_cm_event *ib_event,
1358 sa_family_t sa_family, __be64 service_id)
1360 if (sa_family == AF_IB) {
1361 if (ib_event->event == IB_CM_REQ_RECEIVED)
1362 cma_save_ib_info(src_addr, dst_addr, listen_id,
1363 ib_event->param.req_rcvd.primary_path);
1364 else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED)
1365 cma_save_ib_info(src_addr, dst_addr, listen_id, NULL);
1369 return cma_save_ip_info(src_addr, dst_addr, ib_event, service_id);
1372 static int cma_save_req_info(const struct ib_cm_event *ib_event,
1373 struct cma_req_info *req)
1375 const struct ib_cm_req_event_param *req_param =
1376 &ib_event->param.req_rcvd;
1377 const struct ib_cm_sidr_req_event_param *sidr_param =
1378 &ib_event->param.sidr_req_rcvd;
1380 switch (ib_event->event) {
1381 case IB_CM_REQ_RECEIVED:
1382 req->device = req_param->listen_id->device;
1383 req->port = req_param->port;
1384 memcpy(&req->local_gid, &req_param->primary_path->sgid,
1385 sizeof(req->local_gid));
1386 req->has_gid = true;
1387 req->service_id = req_param->primary_path->service_id;
1388 req->pkey = be16_to_cpu(req_param->primary_path->pkey);
1389 if (req->pkey != req_param->bth_pkey)
1390 pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and primary path P_Key (0x%x)\n"
1391 "RDMA CMA: in the future this may cause the request to be dropped\n",
1392 req_param->bth_pkey, req->pkey);
1394 case IB_CM_SIDR_REQ_RECEIVED:
1395 req->device = sidr_param->listen_id->device;
1396 req->port = sidr_param->port;
1397 req->has_gid = false;
1398 req->service_id = sidr_param->service_id;
1399 req->pkey = sidr_param->pkey;
1400 if (req->pkey != sidr_param->bth_pkey)
1401 pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and SIDR request payload P_Key (0x%x)\n"
1402 "RDMA CMA: in the future this may cause the request to be dropped\n",
1403 sidr_param->bth_pkey, req->pkey);
1412 static bool validate_ipv4_net_dev(struct net_device *net_dev,
1413 const struct sockaddr_in *dst_addr,
1414 const struct sockaddr_in *src_addr)
1416 __be32 daddr = dst_addr->sin_addr.s_addr,
1417 saddr = src_addr->sin_addr.s_addr;
1418 struct fib_result res;
1423 if (ipv4_is_multicast(saddr) || ipv4_is_lbcast(saddr) ||
1424 ipv4_is_lbcast(daddr) || ipv4_is_zeronet(saddr) ||
1425 ipv4_is_zeronet(daddr) || ipv4_is_loopback(daddr) ||
1426 ipv4_is_loopback(saddr))
1429 memset(&fl4, 0, sizeof(fl4));
1430 fl4.flowi4_iif = net_dev->ifindex;
1435 err = fib_lookup(dev_net(net_dev), &fl4, &res, 0);
1436 ret = err == 0 && FIB_RES_DEV(res) == net_dev;
1442 static bool validate_ipv6_net_dev(struct net_device *net_dev,
1443 const struct sockaddr_in6 *dst_addr,
1444 const struct sockaddr_in6 *src_addr)
1446 #if IS_ENABLED(CONFIG_IPV6)
1447 const int strict = ipv6_addr_type(&dst_addr->sin6_addr) &
1448 IPV6_ADDR_LINKLOCAL;
1449 struct rt6_info *rt = rt6_lookup(dev_net(net_dev), &dst_addr->sin6_addr,
1450 &src_addr->sin6_addr, net_dev->ifindex,
1457 ret = rt->rt6i_idev->dev == net_dev;
1466 static bool validate_net_dev(struct net_device *net_dev,
1467 const struct sockaddr *daddr,
1468 const struct sockaddr *saddr)
1470 const struct sockaddr_in *daddr4 = (const struct sockaddr_in *)daddr;
1471 const struct sockaddr_in *saddr4 = (const struct sockaddr_in *)saddr;
1472 const struct sockaddr_in6 *daddr6 = (const struct sockaddr_in6 *)daddr;
1473 const struct sockaddr_in6 *saddr6 = (const struct sockaddr_in6 *)saddr;
1475 switch (daddr->sa_family) {
1477 return saddr->sa_family == AF_INET &&
1478 validate_ipv4_net_dev(net_dev, daddr4, saddr4);
1481 return saddr->sa_family == AF_INET6 &&
1482 validate_ipv6_net_dev(net_dev, daddr6, saddr6);
1489 static struct net_device *
1490 roce_get_net_dev_by_cm_event(const struct ib_cm_event *ib_event)
1492 const struct ib_gid_attr *sgid_attr = NULL;
1493 struct net_device *ndev;
1495 if (ib_event->event == IB_CM_REQ_RECEIVED)
1496 sgid_attr = ib_event->param.req_rcvd.ppath_sgid_attr;
1497 else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED)
1498 sgid_attr = ib_event->param.sidr_req_rcvd.sgid_attr;
1504 ndev = rdma_read_gid_attr_ndev_rcu(sgid_attr);
1513 static struct net_device *cma_get_net_dev(const struct ib_cm_event *ib_event,
1514 struct cma_req_info *req)
1516 struct sockaddr *listen_addr =
1517 (struct sockaddr *)&req->listen_addr_storage;
1518 struct sockaddr *src_addr = (struct sockaddr *)&req->src_addr_storage;
1519 struct net_device *net_dev;
1520 const union ib_gid *gid = req->has_gid ? &req->local_gid : NULL;
1523 err = cma_save_ip_info(listen_addr, src_addr, ib_event,
1526 return ERR_PTR(err);
1528 if (rdma_protocol_roce(req->device, req->port))
1529 net_dev = roce_get_net_dev_by_cm_event(ib_event);
1531 net_dev = ib_get_net_dev_by_params(req->device, req->port,
1535 return ERR_PTR(-ENODEV);
1540 static enum rdma_ucm_port_space rdma_ps_from_service_id(__be64 service_id)
1542 return (be64_to_cpu(service_id) >> 16) & 0xffff;
1545 static bool cma_match_private_data(struct rdma_id_private *id_priv,
1546 const struct cma_hdr *hdr)
1548 struct sockaddr *addr = cma_src_addr(id_priv);
1550 struct in6_addr ip6_addr;
1552 if (cma_any_addr(addr) && !id_priv->afonly)
1555 switch (addr->sa_family) {
1557 ip4_addr = ((struct sockaddr_in *)addr)->sin_addr.s_addr;
1558 if (cma_get_ip_ver(hdr) != 4)
1560 if (!cma_any_addr(addr) &&
1561 hdr->dst_addr.ip4.addr != ip4_addr)
1565 ip6_addr = ((struct sockaddr_in6 *)addr)->sin6_addr;
1566 if (cma_get_ip_ver(hdr) != 6)
1568 if (!cma_any_addr(addr) &&
1569 memcmp(&hdr->dst_addr.ip6, &ip6_addr, sizeof(ip6_addr)))
1581 static bool cma_protocol_roce(const struct rdma_cm_id *id)
1583 struct ib_device *device = id->device;
1584 const u32 port_num = id->port_num ?: rdma_start_port(device);
1586 return rdma_protocol_roce(device, port_num);
1589 static bool cma_is_req_ipv6_ll(const struct cma_req_info *req)
1591 const struct sockaddr *daddr =
1592 (const struct sockaddr *)&req->listen_addr_storage;
1593 const struct sockaddr_in6 *daddr6 = (const struct sockaddr_in6 *)daddr;
1595 /* Returns true if the req is for IPv6 link local */
1596 return (daddr->sa_family == AF_INET6 &&
1597 (ipv6_addr_type(&daddr6->sin6_addr) & IPV6_ADDR_LINKLOCAL));
1600 static bool cma_match_net_dev(const struct rdma_cm_id *id,
1601 const struct net_device *net_dev,
1602 const struct cma_req_info *req)
1604 const struct rdma_addr *addr = &id->route.addr;
1607 /* This request is an AF_IB request */
1608 return (!id->port_num || id->port_num == req->port) &&
1609 (addr->src_addr.ss_family == AF_IB);
1612 * If the request is not for IPv6 link local, allow matching
1613 * request to any netdevice of the one or multiport rdma device.
1615 if (!cma_is_req_ipv6_ll(req))
1618 * Net namespaces must match, and if the listner is listening
1619 * on a specific netdevice than netdevice must match as well.
1621 if (net_eq(dev_net(net_dev), addr->dev_addr.net) &&
1622 (!!addr->dev_addr.bound_dev_if ==
1623 (addr->dev_addr.bound_dev_if == net_dev->ifindex)))
1629 static struct rdma_id_private *cma_find_listener(
1630 const struct rdma_bind_list *bind_list,
1631 const struct ib_cm_id *cm_id,
1632 const struct ib_cm_event *ib_event,
1633 const struct cma_req_info *req,
1634 const struct net_device *net_dev)
1636 struct rdma_id_private *id_priv, *id_priv_dev;
1638 lockdep_assert_held(&lock);
1641 return ERR_PTR(-EINVAL);
1643 hlist_for_each_entry(id_priv, &bind_list->owners, node) {
1644 if (cma_match_private_data(id_priv, ib_event->private_data)) {
1645 if (id_priv->id.device == cm_id->device &&
1646 cma_match_net_dev(&id_priv->id, net_dev, req))
1648 list_for_each_entry(id_priv_dev,
1649 &id_priv->listen_list,
1651 if (id_priv_dev->id.device == cm_id->device &&
1652 cma_match_net_dev(&id_priv_dev->id,
1659 return ERR_PTR(-EINVAL);
1662 static struct rdma_id_private *
1663 cma_ib_id_from_event(struct ib_cm_id *cm_id,
1664 const struct ib_cm_event *ib_event,
1665 struct cma_req_info *req,
1666 struct net_device **net_dev)
1668 struct rdma_bind_list *bind_list;
1669 struct rdma_id_private *id_priv;
1672 err = cma_save_req_info(ib_event, req);
1674 return ERR_PTR(err);
1676 *net_dev = cma_get_net_dev(ib_event, req);
1677 if (IS_ERR(*net_dev)) {
1678 if (PTR_ERR(*net_dev) == -EAFNOSUPPORT) {
1679 /* Assuming the protocol is AF_IB */
1682 return ERR_CAST(*net_dev);
1688 * Net namespace might be getting deleted while route lookup,
1689 * cm_id lookup is in progress. Therefore, perform netdevice
1690 * validation, cm_id lookup under rcu lock.
1691 * RCU lock along with netdevice state check, synchronizes with
1692 * netdevice migrating to different net namespace and also avoids
1693 * case where net namespace doesn't get deleted while lookup is in
1695 * If the device state is not IFF_UP, its properties such as ifindex
1696 * and nd_net cannot be trusted to remain valid without rcu lock.
1697 * net/core/dev.c change_net_namespace() ensures to synchronize with
1698 * ongoing operations on net device after device is closed using
1699 * synchronize_net().
1704 * If netdevice is down, it is likely that it is administratively
1705 * down or it might be migrating to different namespace.
1706 * In that case avoid further processing, as the net namespace
1707 * or ifindex may change.
1709 if (((*net_dev)->flags & IFF_UP) == 0) {
1710 id_priv = ERR_PTR(-EHOSTUNREACH);
1714 if (!validate_net_dev(*net_dev,
1715 (struct sockaddr *)&req->listen_addr_storage,
1716 (struct sockaddr *)&req->src_addr_storage)) {
1717 id_priv = ERR_PTR(-EHOSTUNREACH);
1722 bind_list = cma_ps_find(*net_dev ? dev_net(*net_dev) : &init_net,
1723 rdma_ps_from_service_id(req->service_id),
1724 cma_port_from_service_id(req->service_id));
1725 id_priv = cma_find_listener(bind_list, cm_id, ib_event, req, *net_dev);
1728 mutex_unlock(&lock);
1729 if (IS_ERR(id_priv) && *net_dev) {
1736 static inline u8 cma_user_data_offset(struct rdma_id_private *id_priv)
1738 return cma_family(id_priv) == AF_IB ? 0 : sizeof(struct cma_hdr);
1741 static void cma_cancel_route(struct rdma_id_private *id_priv)
1743 if (rdma_cap_ib_sa(id_priv->id.device, id_priv->id.port_num)) {
1745 ib_sa_cancel_query(id_priv->query_id, id_priv->query);
1749 static void cma_cancel_listens(struct rdma_id_private *id_priv)
1751 struct rdma_id_private *dev_id_priv;
1754 * Remove from listen_any_list to prevent added devices from spawning
1755 * additional listen requests.
1758 list_del(&id_priv->list);
1760 while (!list_empty(&id_priv->listen_list)) {
1761 dev_id_priv = list_entry(id_priv->listen_list.next,
1762 struct rdma_id_private, listen_list);
1763 /* sync with device removal to avoid duplicate destruction */
1764 list_del_init(&dev_id_priv->list);
1765 list_del(&dev_id_priv->listen_list);
1766 mutex_unlock(&lock);
1768 rdma_destroy_id(&dev_id_priv->id);
1771 mutex_unlock(&lock);
1774 static void cma_cancel_operation(struct rdma_id_private *id_priv,
1775 enum rdma_cm_state state)
1778 case RDMA_CM_ADDR_QUERY:
1779 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
1781 case RDMA_CM_ROUTE_QUERY:
1782 cma_cancel_route(id_priv);
1784 case RDMA_CM_LISTEN:
1785 if (cma_any_addr(cma_src_addr(id_priv)) && !id_priv->cma_dev)
1786 cma_cancel_listens(id_priv);
1793 static void cma_release_port(struct rdma_id_private *id_priv)
1795 struct rdma_bind_list *bind_list = id_priv->bind_list;
1796 struct net *net = id_priv->id.route.addr.dev_addr.net;
1802 hlist_del(&id_priv->node);
1803 if (hlist_empty(&bind_list->owners)) {
1804 cma_ps_remove(net, bind_list->ps, bind_list->port);
1807 mutex_unlock(&lock);
1810 static void destroy_mc(struct rdma_id_private *id_priv,
1811 struct cma_multicast *mc)
1813 if (rdma_cap_ib_mcast(id_priv->id.device, id_priv->id.port_num))
1814 ib_sa_free_multicast(mc->sa_mc);
1816 if (rdma_protocol_roce(id_priv->id.device, id_priv->id.port_num)) {
1817 struct rdma_dev_addr *dev_addr =
1818 &id_priv->id.route.addr.dev_addr;
1819 struct net_device *ndev = NULL;
1821 if (dev_addr->bound_dev_if)
1822 ndev = dev_get_by_index(dev_addr->net,
1823 dev_addr->bound_dev_if);
1827 cma_set_mgid(id_priv, (struct sockaddr *)&mc->addr,
1829 cma_igmp_send(ndev, &mgid, false);
1833 cancel_work_sync(&mc->iboe_join.work);
1838 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
1840 struct cma_multicast *mc;
1842 while (!list_empty(&id_priv->mc_list)) {
1843 mc = list_first_entry(&id_priv->mc_list, struct cma_multicast,
1845 list_del(&mc->list);
1846 destroy_mc(id_priv, mc);
1850 static void _destroy_id(struct rdma_id_private *id_priv,
1851 enum rdma_cm_state state)
1853 cma_cancel_operation(id_priv, state);
1855 if (id_priv->cma_dev) {
1856 if (rdma_cap_ib_cm(id_priv->id.device, 1)) {
1857 if (id_priv->cm_id.ib)
1858 ib_destroy_cm_id(id_priv->cm_id.ib);
1859 } else if (rdma_cap_iw_cm(id_priv->id.device, 1)) {
1860 if (id_priv->cm_id.iw)
1861 iw_destroy_cm_id(id_priv->cm_id.iw);
1863 cma_leave_mc_groups(id_priv);
1864 rdma_restrack_del(&id_priv->res);
1865 cma_release_dev(id_priv);
1868 cma_release_port(id_priv);
1869 cma_id_put(id_priv);
1870 wait_for_completion(&id_priv->comp);
1872 if (id_priv->internal_id)
1873 cma_id_put(id_priv->id.context);
1875 kfree(id_priv->id.route.path_rec);
1877 put_net(id_priv->id.route.addr.dev_addr.net);
1882 * destroy an ID from within the handler_mutex. This ensures that no other
1883 * handlers can start running concurrently.
1885 static void destroy_id_handler_unlock(struct rdma_id_private *id_priv)
1886 __releases(&idprv->handler_mutex)
1888 enum rdma_cm_state state;
1889 unsigned long flags;
1891 trace_cm_id_destroy(id_priv);
1894 * Setting the state to destroyed under the handler mutex provides a
1895 * fence against calling handler callbacks. If this is invoked due to
1896 * the failure of a handler callback then it guarentees that no future
1897 * handlers will be called.
1899 lockdep_assert_held(&id_priv->handler_mutex);
1900 spin_lock_irqsave(&id_priv->lock, flags);
1901 state = id_priv->state;
1902 id_priv->state = RDMA_CM_DESTROYING;
1903 spin_unlock_irqrestore(&id_priv->lock, flags);
1904 mutex_unlock(&id_priv->handler_mutex);
1905 _destroy_id(id_priv, state);
1908 void rdma_destroy_id(struct rdma_cm_id *id)
1910 struct rdma_id_private *id_priv =
1911 container_of(id, struct rdma_id_private, id);
1913 mutex_lock(&id_priv->handler_mutex);
1914 destroy_id_handler_unlock(id_priv);
1916 EXPORT_SYMBOL(rdma_destroy_id);
1918 static int cma_rep_recv(struct rdma_id_private *id_priv)
1922 ret = cma_modify_qp_rtr(id_priv, NULL);
1926 ret = cma_modify_qp_rts(id_priv, NULL);
1930 trace_cm_send_rtu(id_priv);
1931 ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
1937 pr_debug_ratelimited("RDMA CM: CONNECT_ERROR: failed to handle reply. status %d\n", ret);
1938 cma_modify_qp_err(id_priv);
1939 trace_cm_send_rej(id_priv);
1940 ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
1945 static void cma_set_rep_event_data(struct rdma_cm_event *event,
1946 const struct ib_cm_rep_event_param *rep_data,
1949 event->param.conn.private_data = private_data;
1950 event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
1951 event->param.conn.responder_resources = rep_data->responder_resources;
1952 event->param.conn.initiator_depth = rep_data->initiator_depth;
1953 event->param.conn.flow_control = rep_data->flow_control;
1954 event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
1955 event->param.conn.srq = rep_data->srq;
1956 event->param.conn.qp_num = rep_data->remote_qpn;
1958 event->ece.vendor_id = rep_data->ece.vendor_id;
1959 event->ece.attr_mod = rep_data->ece.attr_mod;
1962 static int cma_cm_event_handler(struct rdma_id_private *id_priv,
1963 struct rdma_cm_event *event)
1967 lockdep_assert_held(&id_priv->handler_mutex);
1969 trace_cm_event_handler(id_priv, event);
1970 ret = id_priv->id.event_handler(&id_priv->id, event);
1971 trace_cm_event_done(id_priv, event, ret);
1975 static int cma_ib_handler(struct ib_cm_id *cm_id,
1976 const struct ib_cm_event *ib_event)
1978 struct rdma_id_private *id_priv = cm_id->context;
1979 struct rdma_cm_event event = {};
1980 enum rdma_cm_state state;
1983 mutex_lock(&id_priv->handler_mutex);
1984 state = READ_ONCE(id_priv->state);
1985 if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
1986 state != RDMA_CM_CONNECT) ||
1987 (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
1988 state != RDMA_CM_DISCONNECT))
1991 switch (ib_event->event) {
1992 case IB_CM_REQ_ERROR:
1993 case IB_CM_REP_ERROR:
1994 event.event = RDMA_CM_EVENT_UNREACHABLE;
1995 event.status = -ETIMEDOUT;
1997 case IB_CM_REP_RECEIVED:
1998 if (state == RDMA_CM_CONNECT &&
1999 (id_priv->id.qp_type != IB_QPT_UD)) {
2000 trace_cm_send_mra(id_priv);
2001 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
2003 if (id_priv->id.qp) {
2004 event.status = cma_rep_recv(id_priv);
2005 event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
2006 RDMA_CM_EVENT_ESTABLISHED;
2008 event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
2010 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
2011 ib_event->private_data);
2013 case IB_CM_RTU_RECEIVED:
2014 case IB_CM_USER_ESTABLISHED:
2015 event.event = RDMA_CM_EVENT_ESTABLISHED;
2017 case IB_CM_DREQ_ERROR:
2018 event.status = -ETIMEDOUT;
2020 case IB_CM_DREQ_RECEIVED:
2021 case IB_CM_DREP_RECEIVED:
2022 if (!cma_comp_exch(id_priv, RDMA_CM_CONNECT,
2023 RDMA_CM_DISCONNECT))
2025 event.event = RDMA_CM_EVENT_DISCONNECTED;
2027 case IB_CM_TIMEWAIT_EXIT:
2028 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
2030 case IB_CM_MRA_RECEIVED:
2033 case IB_CM_REJ_RECEIVED:
2034 pr_debug_ratelimited("RDMA CM: REJECTED: %s\n", rdma_reject_msg(&id_priv->id,
2035 ib_event->param.rej_rcvd.reason));
2036 cma_modify_qp_err(id_priv);
2037 event.status = ib_event->param.rej_rcvd.reason;
2038 event.event = RDMA_CM_EVENT_REJECTED;
2039 event.param.conn.private_data = ib_event->private_data;
2040 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
2043 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
2048 ret = cma_cm_event_handler(id_priv, &event);
2050 /* Destroy the CM ID by returning a non-zero value. */
2051 id_priv->cm_id.ib = NULL;
2052 destroy_id_handler_unlock(id_priv);
2056 mutex_unlock(&id_priv->handler_mutex);
2060 static struct rdma_id_private *
2061 cma_ib_new_conn_id(const struct rdma_cm_id *listen_id,
2062 const struct ib_cm_event *ib_event,
2063 struct net_device *net_dev)
2065 struct rdma_id_private *listen_id_priv;
2066 struct rdma_id_private *id_priv;
2067 struct rdma_cm_id *id;
2068 struct rdma_route *rt;
2069 const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
2070 struct sa_path_rec *path = ib_event->param.req_rcvd.primary_path;
2071 const __be64 service_id =
2072 ib_event->param.req_rcvd.primary_path->service_id;
2075 listen_id_priv = container_of(listen_id, struct rdma_id_private, id);
2076 id_priv = __rdma_create_id(listen_id->route.addr.dev_addr.net,
2077 listen_id->event_handler, listen_id->context,
2079 ib_event->param.req_rcvd.qp_type,
2081 if (IS_ERR(id_priv))
2085 if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
2086 (struct sockaddr *)&id->route.addr.dst_addr,
2087 listen_id, ib_event, ss_family, service_id))
2091 rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
2092 rt->path_rec = kmalloc_array(rt->num_paths, sizeof(*rt->path_rec),
2097 rt->path_rec[0] = *path;
2098 if (rt->num_paths == 2)
2099 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
2102 rdma_copy_src_l2_addr(&rt->addr.dev_addr, net_dev);
2104 if (!cma_protocol_roce(listen_id) &&
2105 cma_any_addr(cma_src_addr(id_priv))) {
2106 rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
2107 rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
2108 ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
2109 } else if (!cma_any_addr(cma_src_addr(id_priv))) {
2110 ret = cma_translate_addr(cma_src_addr(id_priv), &rt->addr.dev_addr);
2115 rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
2117 id_priv->state = RDMA_CM_CONNECT;
2121 rdma_destroy_id(id);
2125 static struct rdma_id_private *
2126 cma_ib_new_udp_id(const struct rdma_cm_id *listen_id,
2127 const struct ib_cm_event *ib_event,
2128 struct net_device *net_dev)
2130 const struct rdma_id_private *listen_id_priv;
2131 struct rdma_id_private *id_priv;
2132 struct rdma_cm_id *id;
2133 const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
2134 struct net *net = listen_id->route.addr.dev_addr.net;
2137 listen_id_priv = container_of(listen_id, struct rdma_id_private, id);
2138 id_priv = __rdma_create_id(net, listen_id->event_handler,
2139 listen_id->context, listen_id->ps, IB_QPT_UD,
2141 if (IS_ERR(id_priv))
2145 if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
2146 (struct sockaddr *)&id->route.addr.dst_addr,
2147 listen_id, ib_event, ss_family,
2148 ib_event->param.sidr_req_rcvd.service_id))
2152 rdma_copy_src_l2_addr(&id->route.addr.dev_addr, net_dev);
2154 if (!cma_any_addr(cma_src_addr(id_priv))) {
2155 ret = cma_translate_addr(cma_src_addr(id_priv),
2156 &id->route.addr.dev_addr);
2162 id_priv->state = RDMA_CM_CONNECT;
2165 rdma_destroy_id(id);
2169 static void cma_set_req_event_data(struct rdma_cm_event *event,
2170 const struct ib_cm_req_event_param *req_data,
2171 void *private_data, int offset)
2173 event->param.conn.private_data = private_data + offset;
2174 event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
2175 event->param.conn.responder_resources = req_data->responder_resources;
2176 event->param.conn.initiator_depth = req_data->initiator_depth;
2177 event->param.conn.flow_control = req_data->flow_control;
2178 event->param.conn.retry_count = req_data->retry_count;
2179 event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
2180 event->param.conn.srq = req_data->srq;
2181 event->param.conn.qp_num = req_data->remote_qpn;
2183 event->ece.vendor_id = req_data->ece.vendor_id;
2184 event->ece.attr_mod = req_data->ece.attr_mod;
2187 static int cma_ib_check_req_qp_type(const struct rdma_cm_id *id,
2188 const struct ib_cm_event *ib_event)
2190 return (((ib_event->event == IB_CM_REQ_RECEIVED) &&
2191 (ib_event->param.req_rcvd.qp_type == id->qp_type)) ||
2192 ((ib_event->event == IB_CM_SIDR_REQ_RECEIVED) &&
2193 (id->qp_type == IB_QPT_UD)) ||
2197 static int cma_ib_req_handler(struct ib_cm_id *cm_id,
2198 const struct ib_cm_event *ib_event)
2200 struct rdma_id_private *listen_id, *conn_id = NULL;
2201 struct rdma_cm_event event = {};
2202 struct cma_req_info req = {};
2203 struct net_device *net_dev;
2207 listen_id = cma_ib_id_from_event(cm_id, ib_event, &req, &net_dev);
2208 if (IS_ERR(listen_id))
2209 return PTR_ERR(listen_id);
2211 trace_cm_req_handler(listen_id, ib_event->event);
2212 if (!cma_ib_check_req_qp_type(&listen_id->id, ib_event)) {
2217 mutex_lock(&listen_id->handler_mutex);
2218 if (READ_ONCE(listen_id->state) != RDMA_CM_LISTEN) {
2219 ret = -ECONNABORTED;
2223 offset = cma_user_data_offset(listen_id);
2224 event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
2225 if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) {
2226 conn_id = cma_ib_new_udp_id(&listen_id->id, ib_event, net_dev);
2227 event.param.ud.private_data = ib_event->private_data + offset;
2228 event.param.ud.private_data_len =
2229 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
2231 conn_id = cma_ib_new_conn_id(&listen_id->id, ib_event, net_dev);
2232 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
2233 ib_event->private_data, offset);
2240 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
2241 ret = cma_ib_acquire_dev(conn_id, listen_id, &req);
2243 destroy_id_handler_unlock(conn_id);
2247 conn_id->cm_id.ib = cm_id;
2248 cm_id->context = conn_id;
2249 cm_id->cm_handler = cma_ib_handler;
2251 ret = cma_cm_event_handler(conn_id, &event);
2253 /* Destroy the CM ID by returning a non-zero value. */
2254 conn_id->cm_id.ib = NULL;
2255 mutex_unlock(&listen_id->handler_mutex);
2256 destroy_id_handler_unlock(conn_id);
2260 if (READ_ONCE(conn_id->state) == RDMA_CM_CONNECT &&
2261 conn_id->id.qp_type != IB_QPT_UD) {
2262 trace_cm_send_mra(cm_id->context);
2263 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
2265 mutex_unlock(&conn_id->handler_mutex);
2268 mutex_unlock(&listen_id->handler_mutex);
2277 __be64 rdma_get_service_id(struct rdma_cm_id *id, struct sockaddr *addr)
2279 if (addr->sa_family == AF_IB)
2280 return ((struct sockaddr_ib *) addr)->sib_sid;
2282 return cpu_to_be64(((u64)id->ps << 16) + be16_to_cpu(cma_port(addr)));
2284 EXPORT_SYMBOL(rdma_get_service_id);
2286 void rdma_read_gids(struct rdma_cm_id *cm_id, union ib_gid *sgid,
2289 struct rdma_addr *addr = &cm_id->route.addr;
2291 if (!cm_id->device) {
2293 memset(sgid, 0, sizeof(*sgid));
2295 memset(dgid, 0, sizeof(*dgid));
2299 if (rdma_protocol_roce(cm_id->device, cm_id->port_num)) {
2301 rdma_ip2gid((struct sockaddr *)&addr->src_addr, sgid);
2303 rdma_ip2gid((struct sockaddr *)&addr->dst_addr, dgid);
2306 rdma_addr_get_sgid(&addr->dev_addr, sgid);
2308 rdma_addr_get_dgid(&addr->dev_addr, dgid);
2311 EXPORT_SYMBOL(rdma_read_gids);
2313 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
2315 struct rdma_id_private *id_priv = iw_id->context;
2316 struct rdma_cm_event event = {};
2318 struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
2319 struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
2321 mutex_lock(&id_priv->handler_mutex);
2322 if (READ_ONCE(id_priv->state) != RDMA_CM_CONNECT)
2325 switch (iw_event->event) {
2326 case IW_CM_EVENT_CLOSE:
2327 event.event = RDMA_CM_EVENT_DISCONNECTED;
2329 case IW_CM_EVENT_CONNECT_REPLY:
2330 memcpy(cma_src_addr(id_priv), laddr,
2331 rdma_addr_size(laddr));
2332 memcpy(cma_dst_addr(id_priv), raddr,
2333 rdma_addr_size(raddr));
2334 switch (iw_event->status) {
2336 event.event = RDMA_CM_EVENT_ESTABLISHED;
2337 event.param.conn.initiator_depth = iw_event->ird;
2338 event.param.conn.responder_resources = iw_event->ord;
2342 event.event = RDMA_CM_EVENT_REJECTED;
2345 event.event = RDMA_CM_EVENT_UNREACHABLE;
2348 event.event = RDMA_CM_EVENT_CONNECT_ERROR;
2352 case IW_CM_EVENT_ESTABLISHED:
2353 event.event = RDMA_CM_EVENT_ESTABLISHED;
2354 event.param.conn.initiator_depth = iw_event->ird;
2355 event.param.conn.responder_resources = iw_event->ord;
2361 event.status = iw_event->status;
2362 event.param.conn.private_data = iw_event->private_data;
2363 event.param.conn.private_data_len = iw_event->private_data_len;
2364 ret = cma_cm_event_handler(id_priv, &event);
2366 /* Destroy the CM ID by returning a non-zero value. */
2367 id_priv->cm_id.iw = NULL;
2368 destroy_id_handler_unlock(id_priv);
2373 mutex_unlock(&id_priv->handler_mutex);
2377 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
2378 struct iw_cm_event *iw_event)
2380 struct rdma_id_private *listen_id, *conn_id;
2381 struct rdma_cm_event event = {};
2382 int ret = -ECONNABORTED;
2383 struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
2384 struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
2386 event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
2387 event.param.conn.private_data = iw_event->private_data;
2388 event.param.conn.private_data_len = iw_event->private_data_len;
2389 event.param.conn.initiator_depth = iw_event->ird;
2390 event.param.conn.responder_resources = iw_event->ord;
2392 listen_id = cm_id->context;
2394 mutex_lock(&listen_id->handler_mutex);
2395 if (READ_ONCE(listen_id->state) != RDMA_CM_LISTEN)
2398 /* Create a new RDMA id for the new IW CM ID */
2399 conn_id = __rdma_create_id(listen_id->id.route.addr.dev_addr.net,
2400 listen_id->id.event_handler,
2401 listen_id->id.context, RDMA_PS_TCP,
2402 IB_QPT_RC, listen_id);
2403 if (IS_ERR(conn_id)) {
2407 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
2408 conn_id->state = RDMA_CM_CONNECT;
2410 ret = rdma_translate_ip(laddr, &conn_id->id.route.addr.dev_addr);
2412 mutex_unlock(&listen_id->handler_mutex);
2413 destroy_id_handler_unlock(conn_id);
2417 ret = cma_iw_acquire_dev(conn_id, listen_id);
2419 mutex_unlock(&listen_id->handler_mutex);
2420 destroy_id_handler_unlock(conn_id);
2424 conn_id->cm_id.iw = cm_id;
2425 cm_id->context = conn_id;
2426 cm_id->cm_handler = cma_iw_handler;
2428 memcpy(cma_src_addr(conn_id), laddr, rdma_addr_size(laddr));
2429 memcpy(cma_dst_addr(conn_id), raddr, rdma_addr_size(raddr));
2431 ret = cma_cm_event_handler(conn_id, &event);
2433 /* User wants to destroy the CM ID */
2434 conn_id->cm_id.iw = NULL;
2435 mutex_unlock(&listen_id->handler_mutex);
2436 destroy_id_handler_unlock(conn_id);
2440 mutex_unlock(&conn_id->handler_mutex);
2443 mutex_unlock(&listen_id->handler_mutex);
2447 static int cma_ib_listen(struct rdma_id_private *id_priv)
2449 struct sockaddr *addr;
2450 struct ib_cm_id *id;
2453 addr = cma_src_addr(id_priv);
2454 svc_id = rdma_get_service_id(&id_priv->id, addr);
2455 id = ib_cm_insert_listen(id_priv->id.device,
2456 cma_ib_req_handler, svc_id);
2459 id_priv->cm_id.ib = id;
2464 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
2467 struct iw_cm_id *id;
2469 id = iw_create_cm_id(id_priv->id.device,
2470 iw_conn_req_handler,
2475 id->tos = id_priv->tos;
2476 id->tos_set = id_priv->tos_set;
2477 id->afonly = id_priv->afonly;
2478 id_priv->cm_id.iw = id;
2480 memcpy(&id_priv->cm_id.iw->local_addr, cma_src_addr(id_priv),
2481 rdma_addr_size(cma_src_addr(id_priv)));
2483 ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
2486 iw_destroy_cm_id(id_priv->cm_id.iw);
2487 id_priv->cm_id.iw = NULL;
2493 static int cma_listen_handler(struct rdma_cm_id *id,
2494 struct rdma_cm_event *event)
2496 struct rdma_id_private *id_priv = id->context;
2498 /* Listening IDs are always destroyed on removal */
2499 if (event->event == RDMA_CM_EVENT_DEVICE_REMOVAL)
2502 id->context = id_priv->id.context;
2503 id->event_handler = id_priv->id.event_handler;
2504 trace_cm_event_handler(id_priv, event);
2505 return id_priv->id.event_handler(id, event);
2508 static int cma_listen_on_dev(struct rdma_id_private *id_priv,
2509 struct cma_device *cma_dev,
2510 struct rdma_id_private **to_destroy)
2512 struct rdma_id_private *dev_id_priv;
2513 struct net *net = id_priv->id.route.addr.dev_addr.net;
2516 lockdep_assert_held(&lock);
2519 if (cma_family(id_priv) == AF_IB && !rdma_cap_ib_cm(cma_dev->device, 1))
2523 __rdma_create_id(net, cma_listen_handler, id_priv,
2524 id_priv->id.ps, id_priv->id.qp_type, id_priv);
2525 if (IS_ERR(dev_id_priv))
2526 return PTR_ERR(dev_id_priv);
2528 dev_id_priv->state = RDMA_CM_ADDR_BOUND;
2529 memcpy(cma_src_addr(dev_id_priv), cma_src_addr(id_priv),
2530 rdma_addr_size(cma_src_addr(id_priv)));
2532 _cma_attach_to_dev(dev_id_priv, cma_dev);
2533 rdma_restrack_add(&dev_id_priv->res);
2534 cma_id_get(id_priv);
2535 dev_id_priv->internal_id = 1;
2536 dev_id_priv->afonly = id_priv->afonly;
2537 dev_id_priv->tos_set = id_priv->tos_set;
2538 dev_id_priv->tos = id_priv->tos;
2540 ret = rdma_listen(&dev_id_priv->id, id_priv->backlog);
2543 list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
2546 /* Caller must destroy this after releasing lock */
2547 *to_destroy = dev_id_priv;
2548 dev_warn(&cma_dev->device->dev, "RDMA CMA: %s, error %d\n", __func__, ret);
2552 static int cma_listen_on_all(struct rdma_id_private *id_priv)
2554 struct rdma_id_private *to_destroy;
2555 struct cma_device *cma_dev;
2559 list_add_tail(&id_priv->list, &listen_any_list);
2560 list_for_each_entry(cma_dev, &dev_list, list) {
2561 ret = cma_listen_on_dev(id_priv, cma_dev, &to_destroy);
2563 /* Prevent racing with cma_process_remove() */
2565 list_del_init(&to_destroy->list);
2569 mutex_unlock(&lock);
2573 list_del(&id_priv->list);
2574 mutex_unlock(&lock);
2576 rdma_destroy_id(&to_destroy->id);
2580 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
2582 struct rdma_id_private *id_priv;
2584 id_priv = container_of(id, struct rdma_id_private, id);
2585 id_priv->tos = (u8) tos;
2586 id_priv->tos_set = true;
2588 EXPORT_SYMBOL(rdma_set_service_type);
2591 * rdma_set_ack_timeout() - Set the ack timeout of QP associated
2592 * with a connection identifier.
2593 * @id: Communication identifier to associated with service type.
2594 * @timeout: Ack timeout to set a QP, expressed as 4.096 * 2^(timeout) usec.
2596 * This function should be called before rdma_connect() on active side,
2597 * and on passive side before rdma_accept(). It is applicable to primary
2598 * path only. The timeout will affect the local side of the QP, it is not
2599 * negotiated with remote side and zero disables the timer. In case it is
2600 * set before rdma_resolve_route, the value will also be used to determine
2601 * PacketLifeTime for RoCE.
2603 * Return: 0 for success
2605 int rdma_set_ack_timeout(struct rdma_cm_id *id, u8 timeout)
2607 struct rdma_id_private *id_priv;
2609 if (id->qp_type != IB_QPT_RC)
2612 id_priv = container_of(id, struct rdma_id_private, id);
2613 id_priv->timeout = timeout;
2614 id_priv->timeout_set = true;
2618 EXPORT_SYMBOL(rdma_set_ack_timeout);
2621 * rdma_set_min_rnr_timer() - Set the minimum RNR Retry timer of the
2622 * QP associated with a connection identifier.
2623 * @id: Communication identifier to associated with service type.
2624 * @min_rnr_timer: 5-bit value encoded as Table 45: "Encoding for RNR NAK
2625 * Timer Field" in the IBTA specification.
2627 * This function should be called before rdma_connect() on active
2628 * side, and on passive side before rdma_accept(). The timer value
2629 * will be associated with the local QP. When it receives a send it is
2630 * not read to handle, typically if the receive queue is empty, an RNR
2631 * Retry NAK is returned to the requester with the min_rnr_timer
2632 * encoded. The requester will then wait at least the time specified
2633 * in the NAK before retrying. The default is zero, which translates
2634 * to a minimum RNR Timer value of 655 ms.
2636 * Return: 0 for success
2638 int rdma_set_min_rnr_timer(struct rdma_cm_id *id, u8 min_rnr_timer)
2640 struct rdma_id_private *id_priv;
2642 /* It is a five-bit value */
2643 if (min_rnr_timer & 0xe0)
2646 if (WARN_ON(id->qp_type != IB_QPT_RC && id->qp_type != IB_QPT_XRC_TGT))
2649 id_priv = container_of(id, struct rdma_id_private, id);
2650 id_priv->min_rnr_timer = min_rnr_timer;
2651 id_priv->min_rnr_timer_set = true;
2655 EXPORT_SYMBOL(rdma_set_min_rnr_timer);
2657 static void cma_query_handler(int status, struct sa_path_rec *path_rec,
2660 struct cma_work *work = context;
2661 struct rdma_route *route;
2663 route = &work->id->id.route;
2666 route->num_paths = 1;
2667 *route->path_rec = *path_rec;
2669 work->old_state = RDMA_CM_ROUTE_QUERY;
2670 work->new_state = RDMA_CM_ADDR_RESOLVED;
2671 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
2672 work->event.status = status;
2673 pr_debug_ratelimited("RDMA CM: ROUTE_ERROR: failed to query path. status %d\n",
2677 queue_work(cma_wq, &work->work);
2680 static int cma_query_ib_route(struct rdma_id_private *id_priv,
2681 unsigned long timeout_ms, struct cma_work *work)
2683 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2684 struct sa_path_rec path_rec;
2685 ib_sa_comp_mask comp_mask;
2686 struct sockaddr_in6 *sin6;
2687 struct sockaddr_ib *sib;
2689 memset(&path_rec, 0, sizeof path_rec);
2691 if (rdma_cap_opa_ah(id_priv->id.device, id_priv->id.port_num))
2692 path_rec.rec_type = SA_PATH_REC_TYPE_OPA;
2694 path_rec.rec_type = SA_PATH_REC_TYPE_IB;
2695 rdma_addr_get_sgid(dev_addr, &path_rec.sgid);
2696 rdma_addr_get_dgid(dev_addr, &path_rec.dgid);
2697 path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
2698 path_rec.numb_path = 1;
2699 path_rec.reversible = 1;
2700 path_rec.service_id = rdma_get_service_id(&id_priv->id,
2701 cma_dst_addr(id_priv));
2703 comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
2704 IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
2705 IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
2707 switch (cma_family(id_priv)) {
2709 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
2710 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
2713 sin6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
2714 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
2715 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2718 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
2719 path_rec.traffic_class = (u8) (be32_to_cpu(sib->sib_flowinfo) >> 20);
2720 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2724 id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
2725 id_priv->id.port_num, &path_rec,
2726 comp_mask, timeout_ms,
2727 GFP_KERNEL, cma_query_handler,
2728 work, &id_priv->query);
2730 return (id_priv->query_id < 0) ? id_priv->query_id : 0;
2733 static void cma_iboe_join_work_handler(struct work_struct *work)
2735 struct cma_multicast *mc =
2736 container_of(work, struct cma_multicast, iboe_join.work);
2737 struct rdma_cm_event *event = &mc->iboe_join.event;
2738 struct rdma_id_private *id_priv = mc->id_priv;
2741 mutex_lock(&id_priv->handler_mutex);
2742 if (READ_ONCE(id_priv->state) == RDMA_CM_DESTROYING ||
2743 READ_ONCE(id_priv->state) == RDMA_CM_DEVICE_REMOVAL)
2746 ret = cma_cm_event_handler(id_priv, event);
2750 mutex_unlock(&id_priv->handler_mutex);
2751 if (event->event == RDMA_CM_EVENT_MULTICAST_JOIN)
2752 rdma_destroy_ah_attr(&event->param.ud.ah_attr);
2755 static void cma_work_handler(struct work_struct *_work)
2757 struct cma_work *work = container_of(_work, struct cma_work, work);
2758 struct rdma_id_private *id_priv = work->id;
2760 mutex_lock(&id_priv->handler_mutex);
2761 if (READ_ONCE(id_priv->state) == RDMA_CM_DESTROYING ||
2762 READ_ONCE(id_priv->state) == RDMA_CM_DEVICE_REMOVAL)
2764 if (work->old_state != 0 || work->new_state != 0) {
2765 if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
2769 if (cma_cm_event_handler(id_priv, &work->event)) {
2770 cma_id_put(id_priv);
2771 destroy_id_handler_unlock(id_priv);
2776 mutex_unlock(&id_priv->handler_mutex);
2777 cma_id_put(id_priv);
2779 if (work->event.event == RDMA_CM_EVENT_MULTICAST_JOIN)
2780 rdma_destroy_ah_attr(&work->event.param.ud.ah_attr);
2784 static void cma_init_resolve_route_work(struct cma_work *work,
2785 struct rdma_id_private *id_priv)
2788 INIT_WORK(&work->work, cma_work_handler);
2789 work->old_state = RDMA_CM_ROUTE_QUERY;
2790 work->new_state = RDMA_CM_ROUTE_RESOLVED;
2791 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
2794 static void enqueue_resolve_addr_work(struct cma_work *work,
2795 struct rdma_id_private *id_priv)
2797 /* Balances with cma_id_put() in cma_work_handler */
2798 cma_id_get(id_priv);
2801 INIT_WORK(&work->work, cma_work_handler);
2802 work->old_state = RDMA_CM_ADDR_QUERY;
2803 work->new_state = RDMA_CM_ADDR_RESOLVED;
2804 work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2806 queue_work(cma_wq, &work->work);
2809 static int cma_resolve_ib_route(struct rdma_id_private *id_priv,
2810 unsigned long timeout_ms)
2812 struct rdma_route *route = &id_priv->id.route;
2813 struct cma_work *work;
2816 work = kzalloc(sizeof *work, GFP_KERNEL);
2820 cma_init_resolve_route_work(work, id_priv);
2822 route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
2823 if (!route->path_rec) {
2828 ret = cma_query_ib_route(id_priv, timeout_ms, work);
2834 kfree(route->path_rec);
2835 route->path_rec = NULL;
2841 static enum ib_gid_type cma_route_gid_type(enum rdma_network_type network_type,
2842 unsigned long supported_gids,
2843 enum ib_gid_type default_gid)
2845 if ((network_type == RDMA_NETWORK_IPV4 ||
2846 network_type == RDMA_NETWORK_IPV6) &&
2847 test_bit(IB_GID_TYPE_ROCE_UDP_ENCAP, &supported_gids))
2848 return IB_GID_TYPE_ROCE_UDP_ENCAP;
2854 * cma_iboe_set_path_rec_l2_fields() is helper function which sets
2855 * path record type based on GID type.
2856 * It also sets up other L2 fields which includes destination mac address
2857 * netdev ifindex, of the path record.
2858 * It returns the netdev of the bound interface for this path record entry.
2860 static struct net_device *
2861 cma_iboe_set_path_rec_l2_fields(struct rdma_id_private *id_priv)
2863 struct rdma_route *route = &id_priv->id.route;
2864 enum ib_gid_type gid_type = IB_GID_TYPE_ROCE;
2865 struct rdma_addr *addr = &route->addr;
2866 unsigned long supported_gids;
2867 struct net_device *ndev;
2869 if (!addr->dev_addr.bound_dev_if)
2872 ndev = dev_get_by_index(addr->dev_addr.net,
2873 addr->dev_addr.bound_dev_if);
2877 supported_gids = roce_gid_type_mask_support(id_priv->id.device,
2878 id_priv->id.port_num);
2879 gid_type = cma_route_gid_type(addr->dev_addr.network,
2882 /* Use the hint from IP Stack to select GID Type */
2883 if (gid_type < ib_network_to_gid_type(addr->dev_addr.network))
2884 gid_type = ib_network_to_gid_type(addr->dev_addr.network);
2885 route->path_rec->rec_type = sa_conv_gid_to_pathrec_type(gid_type);
2887 route->path_rec->roce.route_resolved = true;
2888 sa_path_set_dmac(route->path_rec, addr->dev_addr.dst_dev_addr);
2892 int rdma_set_ib_path(struct rdma_cm_id *id,
2893 struct sa_path_rec *path_rec)
2895 struct rdma_id_private *id_priv;
2896 struct net_device *ndev;
2899 id_priv = container_of(id, struct rdma_id_private, id);
2900 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
2901 RDMA_CM_ROUTE_RESOLVED))
2904 id->route.path_rec = kmemdup(path_rec, sizeof(*path_rec),
2906 if (!id->route.path_rec) {
2911 if (rdma_protocol_roce(id->device, id->port_num)) {
2912 ndev = cma_iboe_set_path_rec_l2_fields(id_priv);
2920 id->route.num_paths = 1;
2924 kfree(id->route.path_rec);
2925 id->route.path_rec = NULL;
2927 cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED);
2930 EXPORT_SYMBOL(rdma_set_ib_path);
2932 static int cma_resolve_iw_route(struct rdma_id_private *id_priv)
2934 struct cma_work *work;
2936 work = kzalloc(sizeof *work, GFP_KERNEL);
2940 cma_init_resolve_route_work(work, id_priv);
2941 queue_work(cma_wq, &work->work);
2945 static int get_vlan_ndev_tc(struct net_device *vlan_ndev, int prio)
2947 struct net_device *dev;
2949 dev = vlan_dev_real_dev(vlan_ndev);
2951 return netdev_get_prio_tc_map(dev, prio);
2953 return (vlan_dev_get_egress_qos_mask(vlan_ndev, prio) &
2954 VLAN_PRIO_MASK) >> VLAN_PRIO_SHIFT;
2957 struct iboe_prio_tc_map {
2963 static int get_lower_vlan_dev_tc(struct net_device *dev,
2964 struct netdev_nested_priv *priv)
2966 struct iboe_prio_tc_map *map = (struct iboe_prio_tc_map *)priv->data;
2968 if (is_vlan_dev(dev))
2969 map->output_tc = get_vlan_ndev_tc(dev, map->input_prio);
2970 else if (dev->num_tc)
2971 map->output_tc = netdev_get_prio_tc_map(dev, map->input_prio);
2974 /* We are interested only in first level VLAN device, so always
2975 * return 1 to stop iterating over next level devices.
2981 static int iboe_tos_to_sl(struct net_device *ndev, int tos)
2983 struct iboe_prio_tc_map prio_tc_map = {};
2984 int prio = rt_tos2priority(tos);
2985 struct netdev_nested_priv priv;
2987 /* If VLAN device, get it directly from the VLAN netdev */
2988 if (is_vlan_dev(ndev))
2989 return get_vlan_ndev_tc(ndev, prio);
2991 prio_tc_map.input_prio = prio;
2992 priv.data = (void *)&prio_tc_map;
2994 netdev_walk_all_lower_dev_rcu(ndev,
2995 get_lower_vlan_dev_tc,
2998 /* If map is found from lower device, use it; Otherwise
2999 * continue with the current netdevice to get priority to tc map.
3001 if (prio_tc_map.found)
3002 return prio_tc_map.output_tc;
3003 else if (ndev->num_tc)
3004 return netdev_get_prio_tc_map(ndev, prio);
3009 static __be32 cma_get_roce_udp_flow_label(struct rdma_id_private *id_priv)
3011 struct sockaddr_in6 *addr6;
3015 addr6 = (struct sockaddr_in6 *)cma_src_addr(id_priv);
3016 fl = be32_to_cpu(addr6->sin6_flowinfo) & IB_GRH_FLOWLABEL_MASK;
3017 if ((cma_family(id_priv) != AF_INET6) || !fl) {
3018 dport = be16_to_cpu(cma_port(cma_dst_addr(id_priv)));
3019 sport = be16_to_cpu(cma_port(cma_src_addr(id_priv)));
3020 hash = (u32)sport * 31 + dport;
3021 fl = hash & IB_GRH_FLOWLABEL_MASK;
3024 return cpu_to_be32(fl);
3027 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
3029 struct rdma_route *route = &id_priv->id.route;
3030 struct rdma_addr *addr = &route->addr;
3031 struct cma_work *work;
3033 struct net_device *ndev;
3035 u8 default_roce_tos = id_priv->cma_dev->default_roce_tos[id_priv->id.port_num -
3036 rdma_start_port(id_priv->cma_dev->device)];
3037 u8 tos = id_priv->tos_set ? id_priv->tos : default_roce_tos;
3040 work = kzalloc(sizeof *work, GFP_KERNEL);
3044 route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
3045 if (!route->path_rec) {
3050 route->num_paths = 1;
3052 ndev = cma_iboe_set_path_rec_l2_fields(id_priv);
3058 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
3059 &route->path_rec->sgid);
3060 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.dst_addr,
3061 &route->path_rec->dgid);
3063 if (((struct sockaddr *)&id_priv->id.route.addr.dst_addr)->sa_family != AF_IB)
3064 /* TODO: get the hoplimit from the inet/inet6 device */
3065 route->path_rec->hop_limit = addr->dev_addr.hoplimit;
3067 route->path_rec->hop_limit = 1;
3068 route->path_rec->reversible = 1;
3069 route->path_rec->pkey = cpu_to_be16(0xffff);
3070 route->path_rec->mtu_selector = IB_SA_EQ;
3071 route->path_rec->sl = iboe_tos_to_sl(ndev, tos);
3072 route->path_rec->traffic_class = tos;
3073 route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
3074 route->path_rec->rate_selector = IB_SA_EQ;
3075 route->path_rec->rate = iboe_get_rate(ndev);
3077 route->path_rec->packet_life_time_selector = IB_SA_EQ;
3078 /* In case ACK timeout is set, use this value to calculate
3079 * PacketLifeTime. As per IBTA 12.7.34,
3080 * local ACK timeout = (2 * PacketLifeTime + Local CA’s ACK delay).
3081 * Assuming a negligible local ACK delay, we can use
3082 * PacketLifeTime = local ACK timeout/2
3083 * as a reasonable approximation for RoCE networks.
3085 route->path_rec->packet_life_time = id_priv->timeout_set ?
3086 id_priv->timeout - 1 : CMA_IBOE_PACKET_LIFETIME;
3088 if (!route->path_rec->mtu) {
3093 if (rdma_protocol_roce_udp_encap(id_priv->id.device,
3094 id_priv->id.port_num))
3095 route->path_rec->flow_label =
3096 cma_get_roce_udp_flow_label(id_priv);
3098 cma_init_resolve_route_work(work, id_priv);
3099 queue_work(cma_wq, &work->work);
3104 kfree(route->path_rec);
3105 route->path_rec = NULL;
3106 route->num_paths = 0;
3112 int rdma_resolve_route(struct rdma_cm_id *id, unsigned long timeout_ms)
3114 struct rdma_id_private *id_priv;
3117 id_priv = container_of(id, struct rdma_id_private, id);
3118 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY))
3121 cma_id_get(id_priv);
3122 if (rdma_cap_ib_sa(id->device, id->port_num))
3123 ret = cma_resolve_ib_route(id_priv, timeout_ms);
3124 else if (rdma_protocol_roce(id->device, id->port_num))
3125 ret = cma_resolve_iboe_route(id_priv);
3126 else if (rdma_protocol_iwarp(id->device, id->port_num))
3127 ret = cma_resolve_iw_route(id_priv);
3136 cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED);
3137 cma_id_put(id_priv);
3140 EXPORT_SYMBOL(rdma_resolve_route);
3142 static void cma_set_loopback(struct sockaddr *addr)
3144 switch (addr->sa_family) {
3146 ((struct sockaddr_in *) addr)->sin_addr.s_addr = htonl(INADDR_LOOPBACK);
3149 ipv6_addr_set(&((struct sockaddr_in6 *) addr)->sin6_addr,
3153 ib_addr_set(&((struct sockaddr_ib *) addr)->sib_addr,
3159 static int cma_bind_loopback(struct rdma_id_private *id_priv)
3161 struct cma_device *cma_dev, *cur_dev;
3163 enum ib_port_state port_state;
3170 list_for_each_entry(cur_dev, &dev_list, list) {
3171 if (cma_family(id_priv) == AF_IB &&
3172 !rdma_cap_ib_cm(cur_dev->device, 1))
3178 rdma_for_each_port (cur_dev->device, p) {
3179 if (!ib_get_cached_port_state(cur_dev->device, p, &port_state) &&
3180 port_state == IB_PORT_ACTIVE) {
3195 ret = rdma_query_gid(cma_dev->device, p, 0, &gid);
3199 ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
3203 id_priv->id.route.addr.dev_addr.dev_type =
3204 (rdma_protocol_ib(cma_dev->device, p)) ?
3205 ARPHRD_INFINIBAND : ARPHRD_ETHER;
3207 rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
3208 ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
3209 id_priv->id.port_num = p;
3210 cma_attach_to_dev(id_priv, cma_dev);
3211 rdma_restrack_add(&id_priv->res);
3212 cma_set_loopback(cma_src_addr(id_priv));
3214 mutex_unlock(&lock);
3218 static void addr_handler(int status, struct sockaddr *src_addr,
3219 struct rdma_dev_addr *dev_addr, void *context)
3221 struct rdma_id_private *id_priv = context;
3222 struct rdma_cm_event event = {};
3223 struct sockaddr *addr;
3224 struct sockaddr_storage old_addr;
3226 mutex_lock(&id_priv->handler_mutex);
3227 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY,
3228 RDMA_CM_ADDR_RESOLVED))
3232 * Store the previous src address, so that if we fail to acquire
3233 * matching rdma device, old address can be restored back, which helps
3234 * to cancel the cma listen operation correctly.
3236 addr = cma_src_addr(id_priv);
3237 memcpy(&old_addr, addr, rdma_addr_size(addr));
3238 memcpy(addr, src_addr, rdma_addr_size(src_addr));
3239 if (!status && !id_priv->cma_dev) {
3240 status = cma_acquire_dev_by_src_ip(id_priv);
3242 pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to acquire device. status %d\n",
3244 rdma_restrack_add(&id_priv->res);
3245 } else if (status) {
3246 pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to resolve IP. status %d\n", status);
3250 memcpy(addr, &old_addr,
3251 rdma_addr_size((struct sockaddr *)&old_addr));
3252 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
3253 RDMA_CM_ADDR_BOUND))
3255 event.event = RDMA_CM_EVENT_ADDR_ERROR;
3256 event.status = status;
3258 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
3260 if (cma_cm_event_handler(id_priv, &event)) {
3261 destroy_id_handler_unlock(id_priv);
3265 mutex_unlock(&id_priv->handler_mutex);
3268 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
3270 struct cma_work *work;
3274 work = kzalloc(sizeof *work, GFP_KERNEL);
3278 if (!id_priv->cma_dev) {
3279 ret = cma_bind_loopback(id_priv);
3284 rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
3285 rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
3287 enqueue_resolve_addr_work(work, id_priv);
3294 static int cma_resolve_ib_addr(struct rdma_id_private *id_priv)
3296 struct cma_work *work;
3299 work = kzalloc(sizeof *work, GFP_KERNEL);
3303 if (!id_priv->cma_dev) {
3304 ret = cma_resolve_ib_dev(id_priv);
3309 rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, (union ib_gid *)
3310 &(((struct sockaddr_ib *) &id_priv->id.route.addr.dst_addr)->sib_addr));
3312 enqueue_resolve_addr_work(work, id_priv);
3319 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
3320 const struct sockaddr *dst_addr)
3322 if (!src_addr || !src_addr->sa_family) {
3323 src_addr = (struct sockaddr *) &id->route.addr.src_addr;
3324 src_addr->sa_family = dst_addr->sa_family;
3325 if (IS_ENABLED(CONFIG_IPV6) &&
3326 dst_addr->sa_family == AF_INET6) {
3327 struct sockaddr_in6 *src_addr6 = (struct sockaddr_in6 *) src_addr;
3328 struct sockaddr_in6 *dst_addr6 = (struct sockaddr_in6 *) dst_addr;
3329 src_addr6->sin6_scope_id = dst_addr6->sin6_scope_id;
3330 if (ipv6_addr_type(&dst_addr6->sin6_addr) & IPV6_ADDR_LINKLOCAL)
3331 id->route.addr.dev_addr.bound_dev_if = dst_addr6->sin6_scope_id;
3332 } else if (dst_addr->sa_family == AF_IB) {
3333 ((struct sockaddr_ib *) src_addr)->sib_pkey =
3334 ((struct sockaddr_ib *) dst_addr)->sib_pkey;
3337 return rdma_bind_addr(id, src_addr);
3341 * If required, resolve the source address for bind and leave the id_priv in
3342 * state RDMA_CM_ADDR_BOUND. This oddly uses the state to determine the prior
3343 * calls made by ULP, a previously bound ID will not be re-bound and src_addr is
3346 static int resolve_prepare_src(struct rdma_id_private *id_priv,
3347 struct sockaddr *src_addr,
3348 const struct sockaddr *dst_addr)
3352 memcpy(cma_dst_addr(id_priv), dst_addr, rdma_addr_size(dst_addr));
3353 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY)) {
3354 /* For a well behaved ULP state will be RDMA_CM_IDLE */
3355 ret = cma_bind_addr(&id_priv->id, src_addr, dst_addr);
3358 if (WARN_ON(!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND,
3359 RDMA_CM_ADDR_QUERY))) {
3365 if (cma_family(id_priv) != dst_addr->sa_family) {
3372 cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
3374 memset(cma_dst_addr(id_priv), 0, rdma_addr_size(dst_addr));
3378 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
3379 const struct sockaddr *dst_addr, unsigned long timeout_ms)
3381 struct rdma_id_private *id_priv =
3382 container_of(id, struct rdma_id_private, id);
3385 ret = resolve_prepare_src(id_priv, src_addr, dst_addr);
3389 if (cma_any_addr(dst_addr)) {
3390 ret = cma_resolve_loopback(id_priv);
3392 if (dst_addr->sa_family == AF_IB) {
3393 ret = cma_resolve_ib_addr(id_priv);
3395 ret = rdma_resolve_ip(cma_src_addr(id_priv), dst_addr,
3396 &id->route.addr.dev_addr,
3397 timeout_ms, addr_handler,
3406 cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
3409 EXPORT_SYMBOL(rdma_resolve_addr);
3411 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse)
3413 struct rdma_id_private *id_priv;
3414 unsigned long flags;
3417 id_priv = container_of(id, struct rdma_id_private, id);
3418 spin_lock_irqsave(&id_priv->lock, flags);
3419 if ((reuse && id_priv->state != RDMA_CM_LISTEN) ||
3420 id_priv->state == RDMA_CM_IDLE) {
3421 id_priv->reuseaddr = reuse;
3426 spin_unlock_irqrestore(&id_priv->lock, flags);
3429 EXPORT_SYMBOL(rdma_set_reuseaddr);
3431 int rdma_set_afonly(struct rdma_cm_id *id, int afonly)
3433 struct rdma_id_private *id_priv;
3434 unsigned long flags;
3437 id_priv = container_of(id, struct rdma_id_private, id);
3438 spin_lock_irqsave(&id_priv->lock, flags);
3439 if (id_priv->state == RDMA_CM_IDLE || id_priv->state == RDMA_CM_ADDR_BOUND) {
3440 id_priv->options |= (1 << CMA_OPTION_AFONLY);
3441 id_priv->afonly = afonly;
3446 spin_unlock_irqrestore(&id_priv->lock, flags);
3449 EXPORT_SYMBOL(rdma_set_afonly);
3451 static void cma_bind_port(struct rdma_bind_list *bind_list,
3452 struct rdma_id_private *id_priv)
3454 struct sockaddr *addr;
3455 struct sockaddr_ib *sib;
3459 lockdep_assert_held(&lock);
3461 addr = cma_src_addr(id_priv);
3462 port = htons(bind_list->port);
3464 switch (addr->sa_family) {
3466 ((struct sockaddr_in *) addr)->sin_port = port;
3469 ((struct sockaddr_in6 *) addr)->sin6_port = port;
3472 sib = (struct sockaddr_ib *) addr;
3473 sid = be64_to_cpu(sib->sib_sid);
3474 mask = be64_to_cpu(sib->sib_sid_mask);
3475 sib->sib_sid = cpu_to_be64((sid & mask) | (u64) ntohs(port));
3476 sib->sib_sid_mask = cpu_to_be64(~0ULL);
3479 id_priv->bind_list = bind_list;
3480 hlist_add_head(&id_priv->node, &bind_list->owners);
3483 static int cma_alloc_port(enum rdma_ucm_port_space ps,
3484 struct rdma_id_private *id_priv, unsigned short snum)
3486 struct rdma_bind_list *bind_list;
3489 lockdep_assert_held(&lock);
3491 bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
3495 ret = cma_ps_alloc(id_priv->id.route.addr.dev_addr.net, ps, bind_list,
3501 bind_list->port = snum;
3502 cma_bind_port(bind_list, id_priv);
3506 return ret == -ENOSPC ? -EADDRNOTAVAIL : ret;
3509 static int cma_port_is_unique(struct rdma_bind_list *bind_list,
3510 struct rdma_id_private *id_priv)
3512 struct rdma_id_private *cur_id;
3513 struct sockaddr *daddr = cma_dst_addr(id_priv);
3514 struct sockaddr *saddr = cma_src_addr(id_priv);
3515 __be16 dport = cma_port(daddr);
3517 lockdep_assert_held(&lock);
3519 hlist_for_each_entry(cur_id, &bind_list->owners, node) {
3520 struct sockaddr *cur_daddr = cma_dst_addr(cur_id);
3521 struct sockaddr *cur_saddr = cma_src_addr(cur_id);
3522 __be16 cur_dport = cma_port(cur_daddr);
3524 if (id_priv == cur_id)
3527 /* different dest port -> unique */
3528 if (!cma_any_port(daddr) &&
3529 !cma_any_port(cur_daddr) &&
3530 (dport != cur_dport))
3533 /* different src address -> unique */
3534 if (!cma_any_addr(saddr) &&
3535 !cma_any_addr(cur_saddr) &&
3536 cma_addr_cmp(saddr, cur_saddr))
3539 /* different dst address -> unique */
3540 if (!cma_any_addr(daddr) &&
3541 !cma_any_addr(cur_daddr) &&
3542 cma_addr_cmp(daddr, cur_daddr))
3545 return -EADDRNOTAVAIL;
3550 static int cma_alloc_any_port(enum rdma_ucm_port_space ps,
3551 struct rdma_id_private *id_priv)
3553 static unsigned int last_used_port;
3554 int low, high, remaining;
3556 struct net *net = id_priv->id.route.addr.dev_addr.net;
3558 lockdep_assert_held(&lock);
3560 inet_get_local_port_range(net, &low, &high);
3561 remaining = (high - low) + 1;
3562 rover = prandom_u32() % remaining + low;
3564 if (last_used_port != rover) {
3565 struct rdma_bind_list *bind_list;
3568 bind_list = cma_ps_find(net, ps, (unsigned short)rover);
3571 ret = cma_alloc_port(ps, id_priv, rover);
3573 ret = cma_port_is_unique(bind_list, id_priv);
3575 cma_bind_port(bind_list, id_priv);
3578 * Remember previously used port number in order to avoid
3579 * re-using same port immediately after it is closed.
3582 last_used_port = rover;
3583 if (ret != -EADDRNOTAVAIL)
3588 if ((rover < low) || (rover > high))
3592 return -EADDRNOTAVAIL;
3596 * Check that the requested port is available. This is called when trying to
3597 * bind to a specific port, or when trying to listen on a bound port. In
3598 * the latter case, the provided id_priv may already be on the bind_list, but
3599 * we still need to check that it's okay to start listening.
3601 static int cma_check_port(struct rdma_bind_list *bind_list,
3602 struct rdma_id_private *id_priv, uint8_t reuseaddr)
3604 struct rdma_id_private *cur_id;
3605 struct sockaddr *addr, *cur_addr;
3607 lockdep_assert_held(&lock);
3609 addr = cma_src_addr(id_priv);
3610 hlist_for_each_entry(cur_id, &bind_list->owners, node) {
3611 if (id_priv == cur_id)
3614 if (reuseaddr && cur_id->reuseaddr)
3617 cur_addr = cma_src_addr(cur_id);
3618 if (id_priv->afonly && cur_id->afonly &&
3619 (addr->sa_family != cur_addr->sa_family))
3622 if (cma_any_addr(addr) || cma_any_addr(cur_addr))
3623 return -EADDRNOTAVAIL;
3625 if (!cma_addr_cmp(addr, cur_addr))
3631 static int cma_use_port(enum rdma_ucm_port_space ps,
3632 struct rdma_id_private *id_priv)
3634 struct rdma_bind_list *bind_list;
3635 unsigned short snum;
3638 lockdep_assert_held(&lock);
3640 snum = ntohs(cma_port(cma_src_addr(id_priv)));
3641 if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
3644 bind_list = cma_ps_find(id_priv->id.route.addr.dev_addr.net, ps, snum);
3646 ret = cma_alloc_port(ps, id_priv, snum);
3648 ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr);
3650 cma_bind_port(bind_list, id_priv);
3655 static enum rdma_ucm_port_space
3656 cma_select_inet_ps(struct rdma_id_private *id_priv)
3658 switch (id_priv->id.ps) {
3663 return id_priv->id.ps;
3670 static enum rdma_ucm_port_space
3671 cma_select_ib_ps(struct rdma_id_private *id_priv)
3673 enum rdma_ucm_port_space ps = 0;
3674 struct sockaddr_ib *sib;
3675 u64 sid_ps, mask, sid;
3677 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
3678 mask = be64_to_cpu(sib->sib_sid_mask) & RDMA_IB_IP_PS_MASK;
3679 sid = be64_to_cpu(sib->sib_sid) & mask;
3681 if ((id_priv->id.ps == RDMA_PS_IB) && (sid == (RDMA_IB_IP_PS_IB & mask))) {
3682 sid_ps = RDMA_IB_IP_PS_IB;
3684 } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_TCP)) &&
3685 (sid == (RDMA_IB_IP_PS_TCP & mask))) {
3686 sid_ps = RDMA_IB_IP_PS_TCP;
3688 } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_UDP)) &&
3689 (sid == (RDMA_IB_IP_PS_UDP & mask))) {
3690 sid_ps = RDMA_IB_IP_PS_UDP;
3695 sib->sib_sid = cpu_to_be64(sid_ps | ntohs(cma_port((struct sockaddr *) sib)));
3696 sib->sib_sid_mask = cpu_to_be64(RDMA_IB_IP_PS_MASK |
3697 be64_to_cpu(sib->sib_sid_mask));
3702 static int cma_get_port(struct rdma_id_private *id_priv)
3704 enum rdma_ucm_port_space ps;
3707 if (cma_family(id_priv) != AF_IB)
3708 ps = cma_select_inet_ps(id_priv);
3710 ps = cma_select_ib_ps(id_priv);
3712 return -EPROTONOSUPPORT;
3715 if (cma_any_port(cma_src_addr(id_priv)))
3716 ret = cma_alloc_any_port(ps, id_priv);
3718 ret = cma_use_port(ps, id_priv);
3719 mutex_unlock(&lock);
3724 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
3725 struct sockaddr *addr)
3727 #if IS_ENABLED(CONFIG_IPV6)
3728 struct sockaddr_in6 *sin6;
3730 if (addr->sa_family != AF_INET6)
3733 sin6 = (struct sockaddr_in6 *) addr;
3735 if (!(ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL))
3738 if (!sin6->sin6_scope_id)
3741 dev_addr->bound_dev_if = sin6->sin6_scope_id;
3746 int rdma_listen(struct rdma_cm_id *id, int backlog)
3748 struct rdma_id_private *id_priv =
3749 container_of(id, struct rdma_id_private, id);
3752 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN)) {
3753 /* For a well behaved ULP state will be RDMA_CM_IDLE */
3754 id->route.addr.src_addr.ss_family = AF_INET;
3755 ret = rdma_bind_addr(id, cma_src_addr(id_priv));
3758 if (WARN_ON(!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND,
3764 * Once the ID reaches RDMA_CM_LISTEN it is not allowed to be reusable
3765 * any more, and has to be unique in the bind list.
3767 if (id_priv->reuseaddr) {
3769 ret = cma_check_port(id_priv->bind_list, id_priv, 0);
3771 id_priv->reuseaddr = 0;
3772 mutex_unlock(&lock);
3777 id_priv->backlog = backlog;
3778 if (id_priv->cma_dev) {
3779 if (rdma_cap_ib_cm(id->device, 1)) {
3780 ret = cma_ib_listen(id_priv);
3783 } else if (rdma_cap_iw_cm(id->device, 1)) {
3784 ret = cma_iw_listen(id_priv, backlog);
3792 ret = cma_listen_on_all(id_priv);
3799 id_priv->backlog = 0;
3801 * All the failure paths that lead here will not allow the req_handler's
3804 cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND);
3807 EXPORT_SYMBOL(rdma_listen);
3809 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
3811 struct rdma_id_private *id_priv;
3813 struct sockaddr *daddr;
3815 if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6 &&
3816 addr->sa_family != AF_IB)
3817 return -EAFNOSUPPORT;
3819 id_priv = container_of(id, struct rdma_id_private, id);
3820 if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND))
3823 ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
3827 memcpy(cma_src_addr(id_priv), addr, rdma_addr_size(addr));
3828 if (!cma_any_addr(addr)) {
3829 ret = cma_translate_addr(addr, &id->route.addr.dev_addr);
3833 ret = cma_acquire_dev_by_src_ip(id_priv);
3838 if (!(id_priv->options & (1 << CMA_OPTION_AFONLY))) {
3839 if (addr->sa_family == AF_INET)
3840 id_priv->afonly = 1;
3841 #if IS_ENABLED(CONFIG_IPV6)
3842 else if (addr->sa_family == AF_INET6) {
3843 struct net *net = id_priv->id.route.addr.dev_addr.net;
3845 id_priv->afonly = net->ipv6.sysctl.bindv6only;
3849 daddr = cma_dst_addr(id_priv);
3850 daddr->sa_family = addr->sa_family;
3852 ret = cma_get_port(id_priv);
3856 if (!cma_any_addr(addr))
3857 rdma_restrack_add(&id_priv->res);
3860 if (id_priv->cma_dev)
3861 cma_release_dev(id_priv);
3863 cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE);
3866 EXPORT_SYMBOL(rdma_bind_addr);
3868 static int cma_format_hdr(void *hdr, struct rdma_id_private *id_priv)
3870 struct cma_hdr *cma_hdr;
3873 cma_hdr->cma_version = CMA_VERSION;
3874 if (cma_family(id_priv) == AF_INET) {
3875 struct sockaddr_in *src4, *dst4;
3877 src4 = (struct sockaddr_in *) cma_src_addr(id_priv);
3878 dst4 = (struct sockaddr_in *) cma_dst_addr(id_priv);
3880 cma_set_ip_ver(cma_hdr, 4);
3881 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
3882 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
3883 cma_hdr->port = src4->sin_port;
3884 } else if (cma_family(id_priv) == AF_INET6) {
3885 struct sockaddr_in6 *src6, *dst6;
3887 src6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
3888 dst6 = (struct sockaddr_in6 *) cma_dst_addr(id_priv);
3890 cma_set_ip_ver(cma_hdr, 6);
3891 cma_hdr->src_addr.ip6 = src6->sin6_addr;
3892 cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
3893 cma_hdr->port = src6->sin6_port;
3898 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
3899 const struct ib_cm_event *ib_event)
3901 struct rdma_id_private *id_priv = cm_id->context;
3902 struct rdma_cm_event event = {};
3903 const struct ib_cm_sidr_rep_event_param *rep =
3904 &ib_event->param.sidr_rep_rcvd;
3907 mutex_lock(&id_priv->handler_mutex);
3908 if (READ_ONCE(id_priv->state) != RDMA_CM_CONNECT)
3911 switch (ib_event->event) {
3912 case IB_CM_SIDR_REQ_ERROR:
3913 event.event = RDMA_CM_EVENT_UNREACHABLE;
3914 event.status = -ETIMEDOUT;
3916 case IB_CM_SIDR_REP_RECEIVED:
3917 event.param.ud.private_data = ib_event->private_data;
3918 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
3919 if (rep->status != IB_SIDR_SUCCESS) {
3920 event.event = RDMA_CM_EVENT_UNREACHABLE;
3921 event.status = ib_event->param.sidr_rep_rcvd.status;
3922 pr_debug_ratelimited("RDMA CM: UNREACHABLE: bad SIDR reply. status %d\n",
3926 ret = cma_set_qkey(id_priv, rep->qkey);
3928 pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to set qkey. status %d\n", ret);
3929 event.event = RDMA_CM_EVENT_ADDR_ERROR;
3933 ib_init_ah_attr_from_path(id_priv->id.device,
3934 id_priv->id.port_num,
3935 id_priv->id.route.path_rec,
3936 &event.param.ud.ah_attr,
3938 event.param.ud.qp_num = rep->qpn;
3939 event.param.ud.qkey = rep->qkey;
3940 event.event = RDMA_CM_EVENT_ESTABLISHED;
3944 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
3949 ret = cma_cm_event_handler(id_priv, &event);
3951 rdma_destroy_ah_attr(&event.param.ud.ah_attr);
3953 /* Destroy the CM ID by returning a non-zero value. */
3954 id_priv->cm_id.ib = NULL;
3955 destroy_id_handler_unlock(id_priv);
3959 mutex_unlock(&id_priv->handler_mutex);
3963 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
3964 struct rdma_conn_param *conn_param)
3966 struct ib_cm_sidr_req_param req;
3967 struct ib_cm_id *id;
3972 memset(&req, 0, sizeof req);
3973 offset = cma_user_data_offset(id_priv);
3974 req.private_data_len = offset + conn_param->private_data_len;
3975 if (req.private_data_len < conn_param->private_data_len)
3978 if (req.private_data_len) {
3979 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
3983 private_data = NULL;
3986 if (conn_param->private_data && conn_param->private_data_len)
3987 memcpy(private_data + offset, conn_param->private_data,
3988 conn_param->private_data_len);
3991 ret = cma_format_hdr(private_data, id_priv);
3994 req.private_data = private_data;
3997 id = ib_create_cm_id(id_priv->id.device, cma_sidr_rep_handler,
4003 id_priv->cm_id.ib = id;
4005 req.path = id_priv->id.route.path_rec;
4006 req.sgid_attr = id_priv->id.route.addr.dev_addr.sgid_attr;
4007 req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
4008 req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
4009 req.max_cm_retries = CMA_MAX_CM_RETRIES;
4011 trace_cm_send_sidr_req(id_priv);
4012 ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
4014 ib_destroy_cm_id(id_priv->cm_id.ib);
4015 id_priv->cm_id.ib = NULL;
4018 kfree(private_data);
4022 static int cma_connect_ib(struct rdma_id_private *id_priv,
4023 struct rdma_conn_param *conn_param)
4025 struct ib_cm_req_param req;
4026 struct rdma_route *route;
4028 struct ib_cm_id *id;
4032 memset(&req, 0, sizeof req);
4033 offset = cma_user_data_offset(id_priv);
4034 req.private_data_len = offset + conn_param->private_data_len;
4035 if (req.private_data_len < conn_param->private_data_len)
4038 if (req.private_data_len) {
4039 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
4043 private_data = NULL;
4046 if (conn_param->private_data && conn_param->private_data_len)
4047 memcpy(private_data + offset, conn_param->private_data,
4048 conn_param->private_data_len);
4050 id = ib_create_cm_id(id_priv->id.device, cma_ib_handler, id_priv);
4055 id_priv->cm_id.ib = id;
4057 route = &id_priv->id.route;
4059 ret = cma_format_hdr(private_data, id_priv);
4062 req.private_data = private_data;
4065 req.primary_path = &route->path_rec[0];
4066 if (route->num_paths == 2)
4067 req.alternate_path = &route->path_rec[1];
4069 req.ppath_sgid_attr = id_priv->id.route.addr.dev_addr.sgid_attr;
4070 /* Alternate path SGID attribute currently unsupported */
4071 req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
4072 req.qp_num = id_priv->qp_num;
4073 req.qp_type = id_priv->id.qp_type;
4074 req.starting_psn = id_priv->seq_num;
4075 req.responder_resources = conn_param->responder_resources;
4076 req.initiator_depth = conn_param->initiator_depth;
4077 req.flow_control = conn_param->flow_control;
4078 req.retry_count = min_t(u8, 7, conn_param->retry_count);
4079 req.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
4080 req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
4081 req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
4082 req.max_cm_retries = CMA_MAX_CM_RETRIES;
4083 req.srq = id_priv->srq ? 1 : 0;
4084 req.ece.vendor_id = id_priv->ece.vendor_id;
4085 req.ece.attr_mod = id_priv->ece.attr_mod;
4087 trace_cm_send_req(id_priv);
4088 ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
4090 if (ret && !IS_ERR(id)) {
4091 ib_destroy_cm_id(id);
4092 id_priv->cm_id.ib = NULL;
4095 kfree(private_data);
4099 static int cma_connect_iw(struct rdma_id_private *id_priv,
4100 struct rdma_conn_param *conn_param)
4102 struct iw_cm_id *cm_id;
4104 struct iw_cm_conn_param iw_param;
4106 cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
4108 return PTR_ERR(cm_id);
4110 cm_id->tos = id_priv->tos;
4111 cm_id->tos_set = id_priv->tos_set;
4112 id_priv->cm_id.iw = cm_id;
4114 memcpy(&cm_id->local_addr, cma_src_addr(id_priv),
4115 rdma_addr_size(cma_src_addr(id_priv)));
4116 memcpy(&cm_id->remote_addr, cma_dst_addr(id_priv),
4117 rdma_addr_size(cma_dst_addr(id_priv)));
4119 ret = cma_modify_qp_rtr(id_priv, conn_param);
4124 iw_param.ord = conn_param->initiator_depth;
4125 iw_param.ird = conn_param->responder_resources;
4126 iw_param.private_data = conn_param->private_data;
4127 iw_param.private_data_len = conn_param->private_data_len;
4128 iw_param.qpn = id_priv->id.qp ? id_priv->qp_num : conn_param->qp_num;
4130 memset(&iw_param, 0, sizeof iw_param);
4131 iw_param.qpn = id_priv->qp_num;
4133 ret = iw_cm_connect(cm_id, &iw_param);
4136 iw_destroy_cm_id(cm_id);
4137 id_priv->cm_id.iw = NULL;
4143 * rdma_connect_locked - Initiate an active connection request.
4144 * @id: Connection identifier to connect.
4145 * @conn_param: Connection information used for connected QPs.
4147 * Same as rdma_connect() but can only be called from the
4148 * RDMA_CM_EVENT_ROUTE_RESOLVED handler callback.
4150 int rdma_connect_locked(struct rdma_cm_id *id,
4151 struct rdma_conn_param *conn_param)
4153 struct rdma_id_private *id_priv =
4154 container_of(id, struct rdma_id_private, id);
4157 if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT))
4161 id_priv->qp_num = conn_param->qp_num;
4162 id_priv->srq = conn_param->srq;
4165 if (rdma_cap_ib_cm(id->device, id->port_num)) {
4166 if (id->qp_type == IB_QPT_UD)
4167 ret = cma_resolve_ib_udp(id_priv, conn_param);
4169 ret = cma_connect_ib(id_priv, conn_param);
4170 } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4171 ret = cma_connect_iw(id_priv, conn_param);
4179 cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED);
4182 EXPORT_SYMBOL(rdma_connect_locked);
4185 * rdma_connect - Initiate an active connection request.
4186 * @id: Connection identifier to connect.
4187 * @conn_param: Connection information used for connected QPs.
4189 * Users must have resolved a route for the rdma_cm_id to connect with by having
4190 * called rdma_resolve_route before calling this routine.
4192 * This call will either connect to a remote QP or obtain remote QP information
4193 * for unconnected rdma_cm_id's. The actual operation is based on the
4194 * rdma_cm_id's port space.
4196 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
4198 struct rdma_id_private *id_priv =
4199 container_of(id, struct rdma_id_private, id);
4202 mutex_lock(&id_priv->handler_mutex);
4203 ret = rdma_connect_locked(id, conn_param);
4204 mutex_unlock(&id_priv->handler_mutex);
4207 EXPORT_SYMBOL(rdma_connect);
4210 * rdma_connect_ece - Initiate an active connection request with ECE data.
4211 * @id: Connection identifier to connect.
4212 * @conn_param: Connection information used for connected QPs.
4213 * @ece: ECE parameters
4215 * See rdma_connect() explanation.
4217 int rdma_connect_ece(struct rdma_cm_id *id, struct rdma_conn_param *conn_param,
4218 struct rdma_ucm_ece *ece)
4220 struct rdma_id_private *id_priv =
4221 container_of(id, struct rdma_id_private, id);
4223 id_priv->ece.vendor_id = ece->vendor_id;
4224 id_priv->ece.attr_mod = ece->attr_mod;
4226 return rdma_connect(id, conn_param);
4228 EXPORT_SYMBOL(rdma_connect_ece);
4230 static int cma_accept_ib(struct rdma_id_private *id_priv,
4231 struct rdma_conn_param *conn_param)
4233 struct ib_cm_rep_param rep;
4236 ret = cma_modify_qp_rtr(id_priv, conn_param);
4240 ret = cma_modify_qp_rts(id_priv, conn_param);
4244 memset(&rep, 0, sizeof rep);
4245 rep.qp_num = id_priv->qp_num;
4246 rep.starting_psn = id_priv->seq_num;
4247 rep.private_data = conn_param->private_data;
4248 rep.private_data_len = conn_param->private_data_len;
4249 rep.responder_resources = conn_param->responder_resources;
4250 rep.initiator_depth = conn_param->initiator_depth;
4251 rep.failover_accepted = 0;
4252 rep.flow_control = conn_param->flow_control;
4253 rep.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
4254 rep.srq = id_priv->srq ? 1 : 0;
4255 rep.ece.vendor_id = id_priv->ece.vendor_id;
4256 rep.ece.attr_mod = id_priv->ece.attr_mod;
4258 trace_cm_send_rep(id_priv);
4259 ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
4264 static int cma_accept_iw(struct rdma_id_private *id_priv,
4265 struct rdma_conn_param *conn_param)
4267 struct iw_cm_conn_param iw_param;
4273 ret = cma_modify_qp_rtr(id_priv, conn_param);
4277 iw_param.ord = conn_param->initiator_depth;
4278 iw_param.ird = conn_param->responder_resources;
4279 iw_param.private_data = conn_param->private_data;
4280 iw_param.private_data_len = conn_param->private_data_len;
4282 iw_param.qpn = id_priv->qp_num;
4284 iw_param.qpn = conn_param->qp_num;
4286 return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
4289 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
4290 enum ib_cm_sidr_status status, u32 qkey,
4291 const void *private_data, int private_data_len)
4293 struct ib_cm_sidr_rep_param rep;
4296 memset(&rep, 0, sizeof rep);
4297 rep.status = status;
4298 if (status == IB_SIDR_SUCCESS) {
4299 ret = cma_set_qkey(id_priv, qkey);
4302 rep.qp_num = id_priv->qp_num;
4303 rep.qkey = id_priv->qkey;
4305 rep.ece.vendor_id = id_priv->ece.vendor_id;
4306 rep.ece.attr_mod = id_priv->ece.attr_mod;
4309 rep.private_data = private_data;
4310 rep.private_data_len = private_data_len;
4312 trace_cm_send_sidr_rep(id_priv);
4313 return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
4317 * rdma_accept - Called to accept a connection request or response.
4318 * @id: Connection identifier associated with the request.
4319 * @conn_param: Information needed to establish the connection. This must be
4320 * provided if accepting a connection request. If accepting a connection
4321 * response, this parameter must be NULL.
4323 * Typically, this routine is only called by the listener to accept a connection
4324 * request. It must also be called on the active side of a connection if the
4325 * user is performing their own QP transitions.
4327 * In the case of error, a reject message is sent to the remote side and the
4328 * state of the qp associated with the id is modified to error, such that any
4329 * previously posted receive buffers would be flushed.
4331 * This function is for use by kernel ULPs and must be called from under the
4334 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
4336 struct rdma_id_private *id_priv =
4337 container_of(id, struct rdma_id_private, id);
4340 lockdep_assert_held(&id_priv->handler_mutex);
4342 if (READ_ONCE(id_priv->state) != RDMA_CM_CONNECT)
4345 if (!id->qp && conn_param) {
4346 id_priv->qp_num = conn_param->qp_num;
4347 id_priv->srq = conn_param->srq;
4350 if (rdma_cap_ib_cm(id->device, id->port_num)) {
4351 if (id->qp_type == IB_QPT_UD) {
4353 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
4355 conn_param->private_data,
4356 conn_param->private_data_len);
4358 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
4362 ret = cma_accept_ib(id_priv, conn_param);
4364 ret = cma_rep_recv(id_priv);
4366 } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4367 ret = cma_accept_iw(id_priv, conn_param);
4376 cma_modify_qp_err(id_priv);
4377 rdma_reject(id, NULL, 0, IB_CM_REJ_CONSUMER_DEFINED);
4380 EXPORT_SYMBOL(rdma_accept);
4382 int rdma_accept_ece(struct rdma_cm_id *id, struct rdma_conn_param *conn_param,
4383 struct rdma_ucm_ece *ece)
4385 struct rdma_id_private *id_priv =
4386 container_of(id, struct rdma_id_private, id);
4388 id_priv->ece.vendor_id = ece->vendor_id;
4389 id_priv->ece.attr_mod = ece->attr_mod;
4391 return rdma_accept(id, conn_param);
4393 EXPORT_SYMBOL(rdma_accept_ece);
4395 void rdma_lock_handler(struct rdma_cm_id *id)
4397 struct rdma_id_private *id_priv =
4398 container_of(id, struct rdma_id_private, id);
4400 mutex_lock(&id_priv->handler_mutex);
4402 EXPORT_SYMBOL(rdma_lock_handler);
4404 void rdma_unlock_handler(struct rdma_cm_id *id)
4406 struct rdma_id_private *id_priv =
4407 container_of(id, struct rdma_id_private, id);
4409 mutex_unlock(&id_priv->handler_mutex);
4411 EXPORT_SYMBOL(rdma_unlock_handler);
4413 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
4415 struct rdma_id_private *id_priv;
4418 id_priv = container_of(id, struct rdma_id_private, id);
4419 if (!id_priv->cm_id.ib)
4422 switch (id->device->node_type) {
4423 case RDMA_NODE_IB_CA:
4424 ret = ib_cm_notify(id_priv->cm_id.ib, event);
4432 EXPORT_SYMBOL(rdma_notify);
4434 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
4435 u8 private_data_len, u8 reason)
4437 struct rdma_id_private *id_priv;
4440 id_priv = container_of(id, struct rdma_id_private, id);
4441 if (!id_priv->cm_id.ib)
4444 if (rdma_cap_ib_cm(id->device, id->port_num)) {
4445 if (id->qp_type == IB_QPT_UD) {
4446 ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT, 0,
4447 private_data, private_data_len);
4449 trace_cm_send_rej(id_priv);
4450 ret = ib_send_cm_rej(id_priv->cm_id.ib, reason, NULL, 0,
4451 private_data, private_data_len);
4453 } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4454 ret = iw_cm_reject(id_priv->cm_id.iw,
4455 private_data, private_data_len);
4462 EXPORT_SYMBOL(rdma_reject);
4464 int rdma_disconnect(struct rdma_cm_id *id)
4466 struct rdma_id_private *id_priv;
4469 id_priv = container_of(id, struct rdma_id_private, id);
4470 if (!id_priv->cm_id.ib)
4473 if (rdma_cap_ib_cm(id->device, id->port_num)) {
4474 ret = cma_modify_qp_err(id_priv);
4477 /* Initiate or respond to a disconnect. */
4478 trace_cm_disconnect(id_priv);
4479 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0)) {
4480 if (!ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0))
4481 trace_cm_sent_drep(id_priv);
4483 trace_cm_sent_dreq(id_priv);
4485 } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4486 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
4493 EXPORT_SYMBOL(rdma_disconnect);
4495 static void cma_make_mc_event(int status, struct rdma_id_private *id_priv,
4496 struct ib_sa_multicast *multicast,
4497 struct rdma_cm_event *event,
4498 struct cma_multicast *mc)
4500 struct rdma_dev_addr *dev_addr;
4501 enum ib_gid_type gid_type;
4502 struct net_device *ndev;
4505 status = cma_set_qkey(id_priv, be32_to_cpu(multicast->rec.qkey));
4507 pr_debug_ratelimited("RDMA CM: MULTICAST_ERROR: failed to join multicast. status %d\n",
4510 event->status = status;
4511 event->param.ud.private_data = mc->context;
4513 event->event = RDMA_CM_EVENT_MULTICAST_ERROR;
4517 dev_addr = &id_priv->id.route.addr.dev_addr;
4518 ndev = dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if);
4521 ->default_gid_type[id_priv->id.port_num -
4523 id_priv->cma_dev->device)];
4525 event->event = RDMA_CM_EVENT_MULTICAST_JOIN;
4526 if (ib_init_ah_from_mcmember(id_priv->id.device, id_priv->id.port_num,
4527 &multicast->rec, ndev, gid_type,
4528 &event->param.ud.ah_attr)) {
4529 event->event = RDMA_CM_EVENT_MULTICAST_ERROR;
4533 event->param.ud.qp_num = 0xFFFFFF;
4534 event->param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
4541 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
4543 struct cma_multicast *mc = multicast->context;
4544 struct rdma_id_private *id_priv = mc->id_priv;
4545 struct rdma_cm_event event = {};
4548 mutex_lock(&id_priv->handler_mutex);
4549 if (READ_ONCE(id_priv->state) == RDMA_CM_DEVICE_REMOVAL ||
4550 READ_ONCE(id_priv->state) == RDMA_CM_DESTROYING)
4553 cma_make_mc_event(status, id_priv, multicast, &event, mc);
4554 ret = cma_cm_event_handler(id_priv, &event);
4555 rdma_destroy_ah_attr(&event.param.ud.ah_attr);
4559 mutex_unlock(&id_priv->handler_mutex);
4563 static void cma_set_mgid(struct rdma_id_private *id_priv,
4564 struct sockaddr *addr, union ib_gid *mgid)
4566 unsigned char mc_map[MAX_ADDR_LEN];
4567 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4568 struct sockaddr_in *sin = (struct sockaddr_in *) addr;
4569 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
4571 if (cma_any_addr(addr)) {
4572 memset(mgid, 0, sizeof *mgid);
4573 } else if ((addr->sa_family == AF_INET6) &&
4574 ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
4576 /* IPv6 address is an SA assigned MGID. */
4577 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
4578 } else if (addr->sa_family == AF_IB) {
4579 memcpy(mgid, &((struct sockaddr_ib *) addr)->sib_addr, sizeof *mgid);
4580 } else if (addr->sa_family == AF_INET6) {
4581 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
4582 if (id_priv->id.ps == RDMA_PS_UDP)
4583 mc_map[7] = 0x01; /* Use RDMA CM signature */
4584 *mgid = *(union ib_gid *) (mc_map + 4);
4586 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
4587 if (id_priv->id.ps == RDMA_PS_UDP)
4588 mc_map[7] = 0x01; /* Use RDMA CM signature */
4589 *mgid = *(union ib_gid *) (mc_map + 4);
4593 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
4594 struct cma_multicast *mc)
4596 struct ib_sa_mcmember_rec rec;
4597 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4598 ib_sa_comp_mask comp_mask;
4601 ib_addr_get_mgid(dev_addr, &rec.mgid);
4602 ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
4607 ret = cma_set_qkey(id_priv, 0);
4611 cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
4612 rec.qkey = cpu_to_be32(id_priv->qkey);
4613 rdma_addr_get_sgid(dev_addr, &rec.port_gid);
4614 rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
4615 rec.join_state = mc->join_state;
4617 comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
4618 IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
4619 IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
4620 IB_SA_MCMEMBER_REC_FLOW_LABEL |
4621 IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
4623 if (id_priv->id.ps == RDMA_PS_IPOIB)
4624 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
4625 IB_SA_MCMEMBER_REC_RATE_SELECTOR |
4626 IB_SA_MCMEMBER_REC_MTU_SELECTOR |
4627 IB_SA_MCMEMBER_REC_MTU |
4628 IB_SA_MCMEMBER_REC_HOP_LIMIT;
4630 mc->sa_mc = ib_sa_join_multicast(&sa_client, id_priv->id.device,
4631 id_priv->id.port_num, &rec, comp_mask,
4632 GFP_KERNEL, cma_ib_mc_handler, mc);
4633 return PTR_ERR_OR_ZERO(mc->sa_mc);
4636 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid,
4637 enum ib_gid_type gid_type)
4639 struct sockaddr_in *sin = (struct sockaddr_in *)addr;
4640 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
4642 if (cma_any_addr(addr)) {
4643 memset(mgid, 0, sizeof *mgid);
4644 } else if (addr->sa_family == AF_INET6) {
4645 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
4648 (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ? 0 : 0xff;
4650 (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ? 0 : 0x0e;
4659 mgid->raw[10] = 0xff;
4660 mgid->raw[11] = 0xff;
4661 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
4665 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
4666 struct cma_multicast *mc)
4668 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4670 struct sockaddr *addr = (struct sockaddr *)&mc->addr;
4671 struct net_device *ndev = NULL;
4672 struct ib_sa_multicast ib;
4673 enum ib_gid_type gid_type;
4676 send_only = mc->join_state == BIT(SENDONLY_FULLMEMBER_JOIN);
4678 if (cma_zero_addr(addr))
4681 gid_type = id_priv->cma_dev->default_gid_type[id_priv->id.port_num -
4682 rdma_start_port(id_priv->cma_dev->device)];
4683 cma_iboe_set_mgid(addr, &ib.rec.mgid, gid_type);
4685 ib.rec.pkey = cpu_to_be16(0xffff);
4686 if (id_priv->id.ps == RDMA_PS_UDP)
4687 ib.rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
4689 if (dev_addr->bound_dev_if)
4690 ndev = dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if);
4694 ib.rec.rate = iboe_get_rate(ndev);
4695 ib.rec.hop_limit = 1;
4696 ib.rec.mtu = iboe_get_mtu(ndev->mtu);
4698 if (addr->sa_family == AF_INET) {
4699 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) {
4700 ib.rec.hop_limit = IPV6_DEFAULT_HOPLIMIT;
4702 err = cma_igmp_send(ndev, &ib.rec.mgid,
4707 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP)
4711 if (err || !ib.rec.mtu)
4712 return err ?: -EINVAL;
4714 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
4716 INIT_WORK(&mc->iboe_join.work, cma_iboe_join_work_handler);
4717 cma_make_mc_event(0, id_priv, &ib, &mc->iboe_join.event, mc);
4718 queue_work(cma_wq, &mc->iboe_join.work);
4722 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
4723 u8 join_state, void *context)
4725 struct rdma_id_private *id_priv =
4726 container_of(id, struct rdma_id_private, id);
4727 struct cma_multicast *mc;
4730 /* Not supported for kernel QPs */
4731 if (WARN_ON(id->qp))
4734 /* ULP is calling this wrong. */
4735 if (!id->device || (READ_ONCE(id_priv->state) != RDMA_CM_ADDR_BOUND &&
4736 READ_ONCE(id_priv->state) != RDMA_CM_ADDR_RESOLVED))
4739 mc = kzalloc(sizeof(*mc), GFP_KERNEL);
4743 memcpy(&mc->addr, addr, rdma_addr_size(addr));
4744 mc->context = context;
4745 mc->id_priv = id_priv;
4746 mc->join_state = join_state;
4748 if (rdma_protocol_roce(id->device, id->port_num)) {
4749 ret = cma_iboe_join_multicast(id_priv, mc);
4752 } else if (rdma_cap_ib_mcast(id->device, id->port_num)) {
4753 ret = cma_join_ib_multicast(id_priv, mc);
4761 spin_lock(&id_priv->lock);
4762 list_add(&mc->list, &id_priv->mc_list);
4763 spin_unlock(&id_priv->lock);
4770 EXPORT_SYMBOL(rdma_join_multicast);
4772 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
4774 struct rdma_id_private *id_priv;
4775 struct cma_multicast *mc;
4777 id_priv = container_of(id, struct rdma_id_private, id);
4778 spin_lock_irq(&id_priv->lock);
4779 list_for_each_entry(mc, &id_priv->mc_list, list) {
4780 if (memcmp(&mc->addr, addr, rdma_addr_size(addr)) != 0)
4782 list_del(&mc->list);
4783 spin_unlock_irq(&id_priv->lock);
4785 WARN_ON(id_priv->cma_dev->device != id->device);
4786 destroy_mc(id_priv, mc);
4789 spin_unlock_irq(&id_priv->lock);
4791 EXPORT_SYMBOL(rdma_leave_multicast);
4793 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
4795 struct rdma_dev_addr *dev_addr;
4796 struct cma_work *work;
4798 dev_addr = &id_priv->id.route.addr.dev_addr;
4800 if ((dev_addr->bound_dev_if == ndev->ifindex) &&
4801 (net_eq(dev_net(ndev), dev_addr->net)) &&
4802 memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
4803 pr_info("RDMA CM addr change for ndev %s used by id %p\n",
4804 ndev->name, &id_priv->id);
4805 work = kzalloc(sizeof *work, GFP_KERNEL);
4809 INIT_WORK(&work->work, cma_work_handler);
4811 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
4812 cma_id_get(id_priv);
4813 queue_work(cma_wq, &work->work);
4819 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
4822 struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
4823 struct cma_device *cma_dev;
4824 struct rdma_id_private *id_priv;
4825 int ret = NOTIFY_DONE;
4827 if (event != NETDEV_BONDING_FAILOVER)
4830 if (!netif_is_bond_master(ndev))
4834 list_for_each_entry(cma_dev, &dev_list, list)
4835 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
4836 ret = cma_netdev_change(ndev, id_priv);
4842 mutex_unlock(&lock);
4846 static struct notifier_block cma_nb = {
4847 .notifier_call = cma_netdev_callback
4850 static void cma_send_device_removal_put(struct rdma_id_private *id_priv)
4852 struct rdma_cm_event event = { .event = RDMA_CM_EVENT_DEVICE_REMOVAL };
4853 enum rdma_cm_state state;
4854 unsigned long flags;
4856 mutex_lock(&id_priv->handler_mutex);
4857 /* Record that we want to remove the device */
4858 spin_lock_irqsave(&id_priv->lock, flags);
4859 state = id_priv->state;
4860 if (state == RDMA_CM_DESTROYING || state == RDMA_CM_DEVICE_REMOVAL) {
4861 spin_unlock_irqrestore(&id_priv->lock, flags);
4862 mutex_unlock(&id_priv->handler_mutex);
4863 cma_id_put(id_priv);
4866 id_priv->state = RDMA_CM_DEVICE_REMOVAL;
4867 spin_unlock_irqrestore(&id_priv->lock, flags);
4869 if (cma_cm_event_handler(id_priv, &event)) {
4871 * At this point the ULP promises it won't call
4872 * rdma_destroy_id() concurrently
4874 cma_id_put(id_priv);
4875 mutex_unlock(&id_priv->handler_mutex);
4876 trace_cm_id_destroy(id_priv);
4877 _destroy_id(id_priv, state);
4880 mutex_unlock(&id_priv->handler_mutex);
4883 * If this races with destroy then the thread that first assigns state
4884 * to a destroying does the cancel.
4886 cma_cancel_operation(id_priv, state);
4887 cma_id_put(id_priv);
4890 static void cma_process_remove(struct cma_device *cma_dev)
4893 while (!list_empty(&cma_dev->id_list)) {
4894 struct rdma_id_private *id_priv = list_first_entry(
4895 &cma_dev->id_list, struct rdma_id_private, list);
4897 list_del(&id_priv->listen_list);
4898 list_del_init(&id_priv->list);
4899 cma_id_get(id_priv);
4900 mutex_unlock(&lock);
4902 cma_send_device_removal_put(id_priv);
4906 mutex_unlock(&lock);
4908 cma_dev_put(cma_dev);
4909 wait_for_completion(&cma_dev->comp);
4912 static bool cma_supported(struct ib_device *device)
4916 rdma_for_each_port(device, i) {
4917 if (rdma_cap_ib_cm(device, i) || rdma_cap_iw_cm(device, i))
4923 static int cma_add_one(struct ib_device *device)
4925 struct rdma_id_private *to_destroy;
4926 struct cma_device *cma_dev;
4927 struct rdma_id_private *id_priv;
4928 unsigned long supported_gids = 0;
4932 if (!cma_supported(device))
4935 cma_dev = kmalloc(sizeof(*cma_dev), GFP_KERNEL);
4939 cma_dev->device = device;
4940 cma_dev->default_gid_type = kcalloc(device->phys_port_cnt,
4941 sizeof(*cma_dev->default_gid_type),
4943 if (!cma_dev->default_gid_type) {
4948 cma_dev->default_roce_tos = kcalloc(device->phys_port_cnt,
4949 sizeof(*cma_dev->default_roce_tos),
4951 if (!cma_dev->default_roce_tos) {
4956 rdma_for_each_port (device, i) {
4957 supported_gids = roce_gid_type_mask_support(device, i);
4958 WARN_ON(!supported_gids);
4959 if (supported_gids & (1 << CMA_PREFERRED_ROCE_GID_TYPE))
4960 cma_dev->default_gid_type[i - rdma_start_port(device)] =
4961 CMA_PREFERRED_ROCE_GID_TYPE;
4963 cma_dev->default_gid_type[i - rdma_start_port(device)] =
4964 find_first_bit(&supported_gids, BITS_PER_LONG);
4965 cma_dev->default_roce_tos[i - rdma_start_port(device)] = 0;
4968 init_completion(&cma_dev->comp);
4969 refcount_set(&cma_dev->refcount, 1);
4970 INIT_LIST_HEAD(&cma_dev->id_list);
4971 ib_set_client_data(device, &cma_client, cma_dev);
4974 list_add_tail(&cma_dev->list, &dev_list);
4975 list_for_each_entry(id_priv, &listen_any_list, list) {
4976 ret = cma_listen_on_dev(id_priv, cma_dev, &to_destroy);
4980 mutex_unlock(&lock);
4982 trace_cm_add_one(device);
4986 list_del(&cma_dev->list);
4987 mutex_unlock(&lock);
4989 /* cma_process_remove() will delete to_destroy */
4990 cma_process_remove(cma_dev);
4991 kfree(cma_dev->default_roce_tos);
4993 kfree(cma_dev->default_gid_type);
5000 static void cma_remove_one(struct ib_device *device, void *client_data)
5002 struct cma_device *cma_dev = client_data;
5004 trace_cm_remove_one(device);
5007 list_del(&cma_dev->list);
5008 mutex_unlock(&lock);
5010 cma_process_remove(cma_dev);
5011 kfree(cma_dev->default_roce_tos);
5012 kfree(cma_dev->default_gid_type);
5016 static int cma_init_net(struct net *net)
5018 struct cma_pernet *pernet = cma_pernet(net);
5020 xa_init(&pernet->tcp_ps);
5021 xa_init(&pernet->udp_ps);
5022 xa_init(&pernet->ipoib_ps);
5023 xa_init(&pernet->ib_ps);
5028 static void cma_exit_net(struct net *net)
5030 struct cma_pernet *pernet = cma_pernet(net);
5032 WARN_ON(!xa_empty(&pernet->tcp_ps));
5033 WARN_ON(!xa_empty(&pernet->udp_ps));
5034 WARN_ON(!xa_empty(&pernet->ipoib_ps));
5035 WARN_ON(!xa_empty(&pernet->ib_ps));
5038 static struct pernet_operations cma_pernet_operations = {
5039 .init = cma_init_net,
5040 .exit = cma_exit_net,
5041 .id = &cma_pernet_id,
5042 .size = sizeof(struct cma_pernet),
5045 static int __init cma_init(void)
5050 * There is a rare lock ordering dependency in cma_netdev_callback()
5051 * that only happens when bonding is enabled. Teach lockdep that rtnl
5052 * must never be nested under lock so it can find these without having
5053 * to test with bonding.
5055 if (IS_ENABLED(CONFIG_LOCKDEP)) {
5058 mutex_unlock(&lock);
5062 cma_wq = alloc_ordered_workqueue("rdma_cm", WQ_MEM_RECLAIM);
5066 ret = register_pernet_subsys(&cma_pernet_operations);
5070 ib_sa_register_client(&sa_client);
5071 register_netdevice_notifier(&cma_nb);
5073 ret = ib_register_client(&cma_client);
5077 ret = cma_configfs_init();
5084 ib_unregister_client(&cma_client);
5086 unregister_netdevice_notifier(&cma_nb);
5087 ib_sa_unregister_client(&sa_client);
5088 unregister_pernet_subsys(&cma_pernet_operations);
5090 destroy_workqueue(cma_wq);
5094 static void __exit cma_cleanup(void)
5096 cma_configfs_exit();
5097 ib_unregister_client(&cma_client);
5098 unregister_netdevice_notifier(&cma_nb);
5099 ib_sa_unregister_client(&sa_client);
5100 unregister_pernet_subsys(&cma_pernet_operations);
5101 destroy_workqueue(cma_wq);
5104 module_init(cma_init);
5105 module_exit(cma_cleanup);