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_QUERY_CLASSPORT_INFO_TIMEOUT 3000
47 #define CMA_MAX_CM_RETRIES 15
48 #define CMA_CM_MRA_SETTING (IB_CM_MRA_FLAG_DELAY | 24)
49 #define CMA_IBOE_PACKET_LIFETIME 18
50 #define CMA_PREFERRED_ROCE_GID_TYPE IB_GID_TYPE_ROCE_UDP_ENCAP
52 static const char * const cma_events[] = {
53 [RDMA_CM_EVENT_ADDR_RESOLVED] = "address resolved",
54 [RDMA_CM_EVENT_ADDR_ERROR] = "address error",
55 [RDMA_CM_EVENT_ROUTE_RESOLVED] = "route resolved ",
56 [RDMA_CM_EVENT_ROUTE_ERROR] = "route error",
57 [RDMA_CM_EVENT_CONNECT_REQUEST] = "connect request",
58 [RDMA_CM_EVENT_CONNECT_RESPONSE] = "connect response",
59 [RDMA_CM_EVENT_CONNECT_ERROR] = "connect error",
60 [RDMA_CM_EVENT_UNREACHABLE] = "unreachable",
61 [RDMA_CM_EVENT_REJECTED] = "rejected",
62 [RDMA_CM_EVENT_ESTABLISHED] = "established",
63 [RDMA_CM_EVENT_DISCONNECTED] = "disconnected",
64 [RDMA_CM_EVENT_DEVICE_REMOVAL] = "device removal",
65 [RDMA_CM_EVENT_MULTICAST_JOIN] = "multicast join",
66 [RDMA_CM_EVENT_MULTICAST_ERROR] = "multicast error",
67 [RDMA_CM_EVENT_ADDR_CHANGE] = "address change",
68 [RDMA_CM_EVENT_TIMEWAIT_EXIT] = "timewait exit",
71 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid,
72 enum ib_gid_type gid_type);
74 const char *__attribute_const__ rdma_event_msg(enum rdma_cm_event_type event)
78 return (index < ARRAY_SIZE(cma_events) && cma_events[index]) ?
79 cma_events[index] : "unrecognized event";
81 EXPORT_SYMBOL(rdma_event_msg);
83 const char *__attribute_const__ rdma_reject_msg(struct rdma_cm_id *id,
86 if (rdma_ib_or_roce(id->device, id->port_num))
87 return ibcm_reject_msg(reason);
89 if (rdma_protocol_iwarp(id->device, id->port_num))
90 return iwcm_reject_msg(reason);
93 return "unrecognized transport";
95 EXPORT_SYMBOL(rdma_reject_msg);
98 * rdma_is_consumer_reject - return true if the consumer rejected the connect
100 * @id: Communication identifier that received the REJECT event.
101 * @reason: Value returned in the REJECT event status field.
103 static bool rdma_is_consumer_reject(struct rdma_cm_id *id, int reason)
105 if (rdma_ib_or_roce(id->device, id->port_num))
106 return reason == IB_CM_REJ_CONSUMER_DEFINED;
108 if (rdma_protocol_iwarp(id->device, id->port_num))
109 return reason == -ECONNREFUSED;
115 const void *rdma_consumer_reject_data(struct rdma_cm_id *id,
116 struct rdma_cm_event *ev, u8 *data_len)
120 if (rdma_is_consumer_reject(id, ev->status)) {
121 *data_len = ev->param.conn.private_data_len;
122 p = ev->param.conn.private_data;
129 EXPORT_SYMBOL(rdma_consumer_reject_data);
132 * rdma_iw_cm_id() - return the iw_cm_id pointer for this cm_id.
133 * @id: Communication Identifier
135 struct iw_cm_id *rdma_iw_cm_id(struct rdma_cm_id *id)
137 struct rdma_id_private *id_priv;
139 id_priv = container_of(id, struct rdma_id_private, id);
140 if (id->device->node_type == RDMA_NODE_RNIC)
141 return id_priv->cm_id.iw;
144 EXPORT_SYMBOL(rdma_iw_cm_id);
147 * rdma_res_to_id() - return the rdma_cm_id pointer for this restrack.
148 * @res: rdma resource tracking entry pointer
150 struct rdma_cm_id *rdma_res_to_id(struct rdma_restrack_entry *res)
152 struct rdma_id_private *id_priv =
153 container_of(res, struct rdma_id_private, res);
157 EXPORT_SYMBOL(rdma_res_to_id);
159 static int cma_add_one(struct ib_device *device);
160 static void cma_remove_one(struct ib_device *device, void *client_data);
162 static struct ib_client cma_client = {
165 .remove = cma_remove_one
168 static struct ib_sa_client sa_client;
169 static LIST_HEAD(dev_list);
170 static LIST_HEAD(listen_any_list);
171 static DEFINE_MUTEX(lock);
172 static struct workqueue_struct *cma_wq;
173 static unsigned int cma_pernet_id;
176 struct xarray tcp_ps;
177 struct xarray udp_ps;
178 struct xarray ipoib_ps;
182 static struct cma_pernet *cma_pernet(struct net *net)
184 return net_generic(net, cma_pernet_id);
188 struct xarray *cma_pernet_xa(struct net *net, enum rdma_ucm_port_space ps)
190 struct cma_pernet *pernet = cma_pernet(net);
194 return &pernet->tcp_ps;
196 return &pernet->udp_ps;
198 return &pernet->ipoib_ps;
200 return &pernet->ib_ps;
207 struct list_head list;
208 struct ib_device *device;
209 struct completion comp;
211 struct list_head id_list;
212 enum ib_gid_type *default_gid_type;
213 u8 *default_roce_tos;
216 struct rdma_bind_list {
217 enum rdma_ucm_port_space ps;
218 struct hlist_head owners;
222 struct class_port_info_context {
223 struct ib_class_port_info *class_port_info;
224 struct ib_device *device;
225 struct completion done;
226 struct ib_sa_query *sa_query;
230 static int cma_ps_alloc(struct net *net, enum rdma_ucm_port_space ps,
231 struct rdma_bind_list *bind_list, int snum)
233 struct xarray *xa = cma_pernet_xa(net, ps);
235 return xa_insert(xa, snum, bind_list, GFP_KERNEL);
238 static struct rdma_bind_list *cma_ps_find(struct net *net,
239 enum rdma_ucm_port_space ps, int snum)
241 struct xarray *xa = cma_pernet_xa(net, ps);
243 return xa_load(xa, snum);
246 static void cma_ps_remove(struct net *net, enum rdma_ucm_port_space ps,
249 struct xarray *xa = cma_pernet_xa(net, ps);
258 void cma_dev_get(struct cma_device *cma_dev)
260 refcount_inc(&cma_dev->refcount);
263 void cma_dev_put(struct cma_device *cma_dev)
265 if (refcount_dec_and_test(&cma_dev->refcount))
266 complete(&cma_dev->comp);
269 struct cma_device *cma_enum_devices_by_ibdev(cma_device_filter filter,
272 struct cma_device *cma_dev;
273 struct cma_device *found_cma_dev = NULL;
277 list_for_each_entry(cma_dev, &dev_list, list)
278 if (filter(cma_dev->device, cookie)) {
279 found_cma_dev = cma_dev;
284 cma_dev_get(found_cma_dev);
286 return found_cma_dev;
289 int cma_get_default_gid_type(struct cma_device *cma_dev,
292 if (!rdma_is_port_valid(cma_dev->device, port))
295 return cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)];
298 int cma_set_default_gid_type(struct cma_device *cma_dev,
300 enum ib_gid_type default_gid_type)
302 unsigned long supported_gids;
304 if (!rdma_is_port_valid(cma_dev->device, port))
307 if (default_gid_type == IB_GID_TYPE_IB &&
308 rdma_protocol_roce_eth_encap(cma_dev->device, port))
309 default_gid_type = IB_GID_TYPE_ROCE;
311 supported_gids = roce_gid_type_mask_support(cma_dev->device, port);
313 if (!(supported_gids & 1 << default_gid_type))
316 cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)] =
322 int cma_get_default_roce_tos(struct cma_device *cma_dev, unsigned int port)
324 if (!rdma_is_port_valid(cma_dev->device, port))
327 return cma_dev->default_roce_tos[port - rdma_start_port(cma_dev->device)];
330 int cma_set_default_roce_tos(struct cma_device *cma_dev, unsigned int port,
333 if (!rdma_is_port_valid(cma_dev->device, port))
336 cma_dev->default_roce_tos[port - rdma_start_port(cma_dev->device)] =
341 struct ib_device *cma_get_ib_dev(struct cma_device *cma_dev)
343 return cma_dev->device;
347 * Device removal can occur at anytime, so we need extra handling to
348 * serialize notifying the user of device removal with other callbacks.
349 * We do this by disabling removal notification while a callback is in process,
350 * and reporting it after the callback completes.
353 struct cma_multicast {
354 struct rdma_id_private *id_priv;
356 struct ib_sa_multicast *sa_mc;
358 struct work_struct work;
359 struct rdma_cm_event event;
362 struct list_head list;
364 struct sockaddr_storage addr;
369 struct work_struct work;
370 struct rdma_id_private *id;
371 enum rdma_cm_state old_state;
372 enum rdma_cm_state new_state;
373 struct rdma_cm_event event;
386 u8 ip_version; /* IP version: 7:4 */
388 union cma_ip_addr src_addr;
389 union cma_ip_addr dst_addr;
392 #define CMA_VERSION 0x00
394 struct cma_req_info {
395 struct sockaddr_storage listen_addr_storage;
396 struct sockaddr_storage src_addr_storage;
397 struct ib_device *device;
398 union ib_gid local_gid;
405 static int cma_comp_exch(struct rdma_id_private *id_priv,
406 enum rdma_cm_state comp, enum rdma_cm_state exch)
412 * The FSM uses a funny double locking where state is protected by both
413 * the handler_mutex and the spinlock. State is not allowed to change
414 * to/from a handler_mutex protected value without also holding
417 if (comp == RDMA_CM_CONNECT || exch == RDMA_CM_CONNECT)
418 lockdep_assert_held(&id_priv->handler_mutex);
420 spin_lock_irqsave(&id_priv->lock, flags);
421 if ((ret = (id_priv->state == comp)))
422 id_priv->state = exch;
423 spin_unlock_irqrestore(&id_priv->lock, flags);
427 static inline u8 cma_get_ip_ver(const struct cma_hdr *hdr)
429 return hdr->ip_version >> 4;
432 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
434 hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
437 static int cma_igmp_send(struct net_device *ndev, union ib_gid *mgid, bool join)
439 struct in_device *in_dev = NULL;
443 in_dev = __in_dev_get_rtnl(ndev);
446 ip_mc_inc_group(in_dev,
447 *(__be32 *)(mgid->raw + 12));
449 ip_mc_dec_group(in_dev,
450 *(__be32 *)(mgid->raw + 12));
454 return (in_dev) ? 0 : -ENODEV;
457 static void _cma_attach_to_dev(struct rdma_id_private *id_priv,
458 struct cma_device *cma_dev)
460 cma_dev_get(cma_dev);
461 id_priv->cma_dev = cma_dev;
462 id_priv->id.device = cma_dev->device;
463 id_priv->id.route.addr.dev_addr.transport =
464 rdma_node_get_transport(cma_dev->device->node_type);
465 list_add_tail(&id_priv->list, &cma_dev->id_list);
467 trace_cm_id_attach(id_priv, cma_dev->device);
470 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
471 struct cma_device *cma_dev)
473 _cma_attach_to_dev(id_priv, cma_dev);
475 cma_dev->default_gid_type[id_priv->id.port_num -
476 rdma_start_port(cma_dev->device)];
479 static void cma_release_dev(struct rdma_id_private *id_priv)
482 list_del(&id_priv->list);
483 cma_dev_put(id_priv->cma_dev);
484 id_priv->cma_dev = NULL;
485 id_priv->id.device = NULL;
486 if (id_priv->id.route.addr.dev_addr.sgid_attr) {
487 rdma_put_gid_attr(id_priv->id.route.addr.dev_addr.sgid_attr);
488 id_priv->id.route.addr.dev_addr.sgid_attr = NULL;
493 static inline struct sockaddr *cma_src_addr(struct rdma_id_private *id_priv)
495 return (struct sockaddr *) &id_priv->id.route.addr.src_addr;
498 static inline struct sockaddr *cma_dst_addr(struct rdma_id_private *id_priv)
500 return (struct sockaddr *) &id_priv->id.route.addr.dst_addr;
503 static inline unsigned short cma_family(struct rdma_id_private *id_priv)
505 return id_priv->id.route.addr.src_addr.ss_family;
508 static int cma_set_qkey(struct rdma_id_private *id_priv, u32 qkey)
510 struct ib_sa_mcmember_rec rec;
514 if (qkey && id_priv->qkey != qkey)
520 id_priv->qkey = qkey;
524 switch (id_priv->id.ps) {
527 id_priv->qkey = RDMA_UDP_QKEY;
530 ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid);
531 ret = ib_sa_get_mcmember_rec(id_priv->id.device,
532 id_priv->id.port_num, &rec.mgid,
535 id_priv->qkey = be32_to_cpu(rec.qkey);
543 static void cma_translate_ib(struct sockaddr_ib *sib, struct rdma_dev_addr *dev_addr)
545 dev_addr->dev_type = ARPHRD_INFINIBAND;
546 rdma_addr_set_sgid(dev_addr, (union ib_gid *) &sib->sib_addr);
547 ib_addr_set_pkey(dev_addr, ntohs(sib->sib_pkey));
550 static int cma_translate_addr(struct sockaddr *addr, struct rdma_dev_addr *dev_addr)
554 if (addr->sa_family != AF_IB) {
555 ret = rdma_translate_ip(addr, dev_addr);
557 cma_translate_ib((struct sockaddr_ib *) addr, dev_addr);
564 static const struct ib_gid_attr *
565 cma_validate_port(struct ib_device *device, u8 port,
566 enum ib_gid_type gid_type,
568 struct rdma_id_private *id_priv)
570 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
571 int bound_if_index = dev_addr->bound_dev_if;
572 const struct ib_gid_attr *sgid_attr;
573 int dev_type = dev_addr->dev_type;
574 struct net_device *ndev = NULL;
576 if (!rdma_dev_access_netns(device, id_priv->id.route.addr.dev_addr.net))
577 return ERR_PTR(-ENODEV);
579 if ((dev_type == ARPHRD_INFINIBAND) && !rdma_protocol_ib(device, port))
580 return ERR_PTR(-ENODEV);
582 if ((dev_type != ARPHRD_INFINIBAND) && rdma_protocol_ib(device, port))
583 return ERR_PTR(-ENODEV);
585 if (dev_type == ARPHRD_ETHER && rdma_protocol_roce(device, port)) {
586 ndev = dev_get_by_index(dev_addr->net, bound_if_index);
588 return ERR_PTR(-ENODEV);
590 gid_type = IB_GID_TYPE_IB;
593 sgid_attr = rdma_find_gid_by_port(device, gid, gid_type, port, ndev);
599 static void cma_bind_sgid_attr(struct rdma_id_private *id_priv,
600 const struct ib_gid_attr *sgid_attr)
602 WARN_ON(id_priv->id.route.addr.dev_addr.sgid_attr);
603 id_priv->id.route.addr.dev_addr.sgid_attr = sgid_attr;
607 * cma_acquire_dev_by_src_ip - Acquire cma device, port, gid attribute
608 * based on source ip address.
609 * @id_priv: cm_id which should be bound to cma device
611 * cma_acquire_dev_by_src_ip() binds cm id to cma device, port and GID attribute
612 * based on source IP address. It returns 0 on success or error code otherwise.
613 * It is applicable to active and passive side cm_id.
615 static int cma_acquire_dev_by_src_ip(struct rdma_id_private *id_priv)
617 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
618 const struct ib_gid_attr *sgid_attr;
619 union ib_gid gid, iboe_gid, *gidp;
620 struct cma_device *cma_dev;
621 enum ib_gid_type gid_type;
625 if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
626 id_priv->id.ps == RDMA_PS_IPOIB)
629 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
632 memcpy(&gid, dev_addr->src_dev_addr +
633 rdma_addr_gid_offset(dev_addr), sizeof(gid));
636 list_for_each_entry(cma_dev, &dev_list, list) {
637 rdma_for_each_port (cma_dev->device, port) {
638 gidp = rdma_protocol_roce(cma_dev->device, port) ?
640 gid_type = cma_dev->default_gid_type[port - 1];
641 sgid_attr = cma_validate_port(cma_dev->device, port,
642 gid_type, gidp, id_priv);
643 if (!IS_ERR(sgid_attr)) {
644 id_priv->id.port_num = port;
645 cma_bind_sgid_attr(id_priv, sgid_attr);
646 cma_attach_to_dev(id_priv, cma_dev);
658 * cma_ib_acquire_dev - Acquire cma device, port and SGID attribute
659 * @id_priv: cm id to bind to cma device
660 * @listen_id_priv: listener cm id to match against
661 * @req: Pointer to req structure containaining incoming
662 * request information
663 * cma_ib_acquire_dev() acquires cma device, port and SGID attribute when
664 * rdma device matches for listen_id and incoming request. It also verifies
665 * that a GID table entry is present for the source address.
666 * Returns 0 on success, or returns error code otherwise.
668 static int cma_ib_acquire_dev(struct rdma_id_private *id_priv,
669 const struct rdma_id_private *listen_id_priv,
670 struct cma_req_info *req)
672 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
673 const struct ib_gid_attr *sgid_attr;
674 enum ib_gid_type gid_type;
677 if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
678 id_priv->id.ps == RDMA_PS_IPOIB)
681 if (rdma_protocol_roce(req->device, req->port))
682 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
685 memcpy(&gid, dev_addr->src_dev_addr +
686 rdma_addr_gid_offset(dev_addr), sizeof(gid));
688 gid_type = listen_id_priv->cma_dev->default_gid_type[req->port - 1];
689 sgid_attr = cma_validate_port(req->device, req->port,
690 gid_type, &gid, id_priv);
691 if (IS_ERR(sgid_attr))
692 return PTR_ERR(sgid_attr);
694 id_priv->id.port_num = req->port;
695 cma_bind_sgid_attr(id_priv, sgid_attr);
696 /* Need to acquire lock to protect against reader
697 * of cma_dev->id_list such as cma_netdev_callback() and
698 * cma_process_remove().
701 cma_attach_to_dev(id_priv, listen_id_priv->cma_dev);
703 rdma_restrack_add(&id_priv->res);
707 static int cma_iw_acquire_dev(struct rdma_id_private *id_priv,
708 const struct rdma_id_private *listen_id_priv)
710 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
711 const struct ib_gid_attr *sgid_attr;
712 struct cma_device *cma_dev;
713 enum ib_gid_type gid_type;
718 if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
719 id_priv->id.ps == RDMA_PS_IPOIB)
722 memcpy(&gid, dev_addr->src_dev_addr +
723 rdma_addr_gid_offset(dev_addr), sizeof(gid));
727 cma_dev = listen_id_priv->cma_dev;
728 port = listen_id_priv->id.port_num;
729 gid_type = listen_id_priv->gid_type;
730 sgid_attr = cma_validate_port(cma_dev->device, port,
731 gid_type, &gid, id_priv);
732 if (!IS_ERR(sgid_attr)) {
733 id_priv->id.port_num = port;
734 cma_bind_sgid_attr(id_priv, sgid_attr);
739 list_for_each_entry(cma_dev, &dev_list, list) {
740 rdma_for_each_port (cma_dev->device, port) {
741 if (listen_id_priv->cma_dev == cma_dev &&
742 listen_id_priv->id.port_num == port)
745 gid_type = cma_dev->default_gid_type[port - 1];
746 sgid_attr = cma_validate_port(cma_dev->device, port,
747 gid_type, &gid, id_priv);
748 if (!IS_ERR(sgid_attr)) {
749 id_priv->id.port_num = port;
750 cma_bind_sgid_attr(id_priv, sgid_attr);
759 cma_attach_to_dev(id_priv, cma_dev);
760 rdma_restrack_add(&id_priv->res);
768 * Select the source IB device and address to reach the destination IB address.
770 static int cma_resolve_ib_dev(struct rdma_id_private *id_priv)
772 struct cma_device *cma_dev, *cur_dev;
773 struct sockaddr_ib *addr;
774 union ib_gid gid, sgid, *dgid;
777 enum ib_port_state port_state;
782 addr = (struct sockaddr_ib *) cma_dst_addr(id_priv);
783 dgid = (union ib_gid *) &addr->sib_addr;
784 pkey = ntohs(addr->sib_pkey);
787 list_for_each_entry(cur_dev, &dev_list, list) {
788 rdma_for_each_port (cur_dev->device, p) {
789 if (!rdma_cap_af_ib(cur_dev->device, p))
792 if (ib_find_cached_pkey(cur_dev->device, p, pkey, &index))
795 if (ib_get_cached_port_state(cur_dev->device, p, &port_state))
798 for (i = 0; i < cur_dev->device->port_data[p].immutable.gid_tbl_len;
800 ret = rdma_query_gid(cur_dev->device, p, i,
805 if (!memcmp(&gid, dgid, sizeof(gid))) {
808 id_priv->id.port_num = p;
812 if (!cma_dev && (gid.global.subnet_prefix ==
813 dgid->global.subnet_prefix) &&
814 port_state == IB_PORT_ACTIVE) {
817 id_priv->id.port_num = p;
827 cma_attach_to_dev(id_priv, cma_dev);
828 rdma_restrack_add(&id_priv->res);
830 addr = (struct sockaddr_ib *)cma_src_addr(id_priv);
831 memcpy(&addr->sib_addr, &sgid, sizeof(sgid));
832 cma_translate_ib(addr, &id_priv->id.route.addr.dev_addr);
836 static void cma_id_get(struct rdma_id_private *id_priv)
838 refcount_inc(&id_priv->refcount);
841 static void cma_id_put(struct rdma_id_private *id_priv)
843 if (refcount_dec_and_test(&id_priv->refcount))
844 complete(&id_priv->comp);
847 static struct rdma_id_private *
848 __rdma_create_id(struct net *net, rdma_cm_event_handler event_handler,
849 void *context, enum rdma_ucm_port_space ps,
850 enum ib_qp_type qp_type, const struct rdma_id_private *parent)
852 struct rdma_id_private *id_priv;
854 id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
856 return ERR_PTR(-ENOMEM);
858 id_priv->state = RDMA_CM_IDLE;
859 id_priv->id.context = context;
860 id_priv->id.event_handler = event_handler;
862 id_priv->id.qp_type = qp_type;
863 id_priv->tos_set = false;
864 id_priv->timeout_set = false;
865 id_priv->gid_type = IB_GID_TYPE_IB;
866 spin_lock_init(&id_priv->lock);
867 mutex_init(&id_priv->qp_mutex);
868 init_completion(&id_priv->comp);
869 refcount_set(&id_priv->refcount, 1);
870 mutex_init(&id_priv->handler_mutex);
871 INIT_LIST_HEAD(&id_priv->listen_list);
872 INIT_LIST_HEAD(&id_priv->mc_list);
873 get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
874 id_priv->id.route.addr.dev_addr.net = get_net(net);
875 id_priv->seq_num &= 0x00ffffff;
877 rdma_restrack_new(&id_priv->res, RDMA_RESTRACK_CM_ID);
879 rdma_restrack_parent_name(&id_priv->res, &parent->res);
885 __rdma_create_kernel_id(struct net *net, rdma_cm_event_handler event_handler,
886 void *context, enum rdma_ucm_port_space ps,
887 enum ib_qp_type qp_type, const char *caller)
889 struct rdma_id_private *ret;
891 ret = __rdma_create_id(net, event_handler, context, ps, qp_type, NULL);
893 return ERR_CAST(ret);
895 rdma_restrack_set_name(&ret->res, caller);
898 EXPORT_SYMBOL(__rdma_create_kernel_id);
900 struct rdma_cm_id *rdma_create_user_id(rdma_cm_event_handler event_handler,
902 enum rdma_ucm_port_space ps,
903 enum ib_qp_type qp_type)
905 struct rdma_id_private *ret;
907 ret = __rdma_create_id(current->nsproxy->net_ns, event_handler, context,
910 return ERR_CAST(ret);
912 rdma_restrack_set_name(&ret->res, NULL);
915 EXPORT_SYMBOL(rdma_create_user_id);
917 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
919 struct ib_qp_attr qp_attr;
920 int qp_attr_mask, ret;
922 qp_attr.qp_state = IB_QPS_INIT;
923 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
927 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
931 qp_attr.qp_state = IB_QPS_RTR;
932 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
936 qp_attr.qp_state = IB_QPS_RTS;
938 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
943 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
945 struct ib_qp_attr qp_attr;
946 int qp_attr_mask, ret;
948 qp_attr.qp_state = IB_QPS_INIT;
949 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
953 return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
956 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
957 struct ib_qp_init_attr *qp_init_attr)
959 struct rdma_id_private *id_priv;
963 id_priv = container_of(id, struct rdma_id_private, id);
964 if (id->device != pd->device) {
969 qp_init_attr->port_num = id->port_num;
970 qp = ib_create_qp(pd, qp_init_attr);
976 if (id->qp_type == IB_QPT_UD)
977 ret = cma_init_ud_qp(id_priv, qp);
979 ret = cma_init_conn_qp(id_priv, qp);
984 id_priv->qp_num = qp->qp_num;
985 id_priv->srq = (qp->srq != NULL);
986 trace_cm_qp_create(id_priv, pd, qp_init_attr, 0);
991 trace_cm_qp_create(id_priv, pd, qp_init_attr, ret);
994 EXPORT_SYMBOL(rdma_create_qp);
996 void rdma_destroy_qp(struct rdma_cm_id *id)
998 struct rdma_id_private *id_priv;
1000 id_priv = container_of(id, struct rdma_id_private, id);
1001 trace_cm_qp_destroy(id_priv);
1002 mutex_lock(&id_priv->qp_mutex);
1003 ib_destroy_qp(id_priv->id.qp);
1004 id_priv->id.qp = NULL;
1005 mutex_unlock(&id_priv->qp_mutex);
1007 EXPORT_SYMBOL(rdma_destroy_qp);
1009 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
1010 struct rdma_conn_param *conn_param)
1012 struct ib_qp_attr qp_attr;
1013 int qp_attr_mask, ret;
1015 mutex_lock(&id_priv->qp_mutex);
1016 if (!id_priv->id.qp) {
1021 /* Need to update QP attributes from default values. */
1022 qp_attr.qp_state = IB_QPS_INIT;
1023 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1027 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
1031 qp_attr.qp_state = IB_QPS_RTR;
1032 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1036 BUG_ON(id_priv->cma_dev->device != id_priv->id.device);
1039 qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
1040 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
1042 mutex_unlock(&id_priv->qp_mutex);
1046 static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
1047 struct rdma_conn_param *conn_param)
1049 struct ib_qp_attr qp_attr;
1050 int qp_attr_mask, ret;
1052 mutex_lock(&id_priv->qp_mutex);
1053 if (!id_priv->id.qp) {
1058 qp_attr.qp_state = IB_QPS_RTS;
1059 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1064 qp_attr.max_rd_atomic = conn_param->initiator_depth;
1065 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
1067 mutex_unlock(&id_priv->qp_mutex);
1071 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
1073 struct ib_qp_attr qp_attr;
1076 mutex_lock(&id_priv->qp_mutex);
1077 if (!id_priv->id.qp) {
1082 qp_attr.qp_state = IB_QPS_ERR;
1083 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
1085 mutex_unlock(&id_priv->qp_mutex);
1089 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
1090 struct ib_qp_attr *qp_attr, int *qp_attr_mask)
1092 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
1096 if (rdma_cap_eth_ah(id_priv->id.device, id_priv->id.port_num))
1099 pkey = ib_addr_get_pkey(dev_addr);
1101 ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
1102 pkey, &qp_attr->pkey_index);
1106 qp_attr->port_num = id_priv->id.port_num;
1107 *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
1109 if (id_priv->id.qp_type == IB_QPT_UD) {
1110 ret = cma_set_qkey(id_priv, 0);
1114 qp_attr->qkey = id_priv->qkey;
1115 *qp_attr_mask |= IB_QP_QKEY;
1117 qp_attr->qp_access_flags = 0;
1118 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
1123 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
1126 struct rdma_id_private *id_priv;
1129 id_priv = container_of(id, struct rdma_id_private, id);
1130 if (rdma_cap_ib_cm(id->device, id->port_num)) {
1131 if (!id_priv->cm_id.ib || (id_priv->id.qp_type == IB_QPT_UD))
1132 ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
1134 ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
1137 if (qp_attr->qp_state == IB_QPS_RTR)
1138 qp_attr->rq_psn = id_priv->seq_num;
1139 } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
1140 if (!id_priv->cm_id.iw) {
1141 qp_attr->qp_access_flags = 0;
1142 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
1144 ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
1146 qp_attr->port_num = id_priv->id.port_num;
1147 *qp_attr_mask |= IB_QP_PORT;
1151 if ((*qp_attr_mask & IB_QP_TIMEOUT) && id_priv->timeout_set)
1152 qp_attr->timeout = id_priv->timeout;
1156 EXPORT_SYMBOL(rdma_init_qp_attr);
1158 static inline bool cma_zero_addr(const struct sockaddr *addr)
1160 switch (addr->sa_family) {
1162 return ipv4_is_zeronet(((struct sockaddr_in *)addr)->sin_addr.s_addr);
1164 return ipv6_addr_any(&((struct sockaddr_in6 *)addr)->sin6_addr);
1166 return ib_addr_any(&((struct sockaddr_ib *)addr)->sib_addr);
1172 static inline bool cma_loopback_addr(const struct sockaddr *addr)
1174 switch (addr->sa_family) {
1176 return ipv4_is_loopback(
1177 ((struct sockaddr_in *)addr)->sin_addr.s_addr);
1179 return ipv6_addr_loopback(
1180 &((struct sockaddr_in6 *)addr)->sin6_addr);
1182 return ib_addr_loopback(
1183 &((struct sockaddr_ib *)addr)->sib_addr);
1189 static inline bool cma_any_addr(const struct sockaddr *addr)
1191 return cma_zero_addr(addr) || cma_loopback_addr(addr);
1194 static int cma_addr_cmp(const struct sockaddr *src, const struct sockaddr *dst)
1196 if (src->sa_family != dst->sa_family)
1199 switch (src->sa_family) {
1201 return ((struct sockaddr_in *)src)->sin_addr.s_addr !=
1202 ((struct sockaddr_in *)dst)->sin_addr.s_addr;
1204 struct sockaddr_in6 *src_addr6 = (struct sockaddr_in6 *)src;
1205 struct sockaddr_in6 *dst_addr6 = (struct sockaddr_in6 *)dst;
1208 if (ipv6_addr_cmp(&src_addr6->sin6_addr,
1209 &dst_addr6->sin6_addr))
1211 link_local = ipv6_addr_type(&dst_addr6->sin6_addr) &
1212 IPV6_ADDR_LINKLOCAL;
1213 /* Link local must match their scope_ids */
1214 return link_local ? (src_addr6->sin6_scope_id !=
1215 dst_addr6->sin6_scope_id) :
1220 return ib_addr_cmp(&((struct sockaddr_ib *) src)->sib_addr,
1221 &((struct sockaddr_ib *) dst)->sib_addr);
1225 static __be16 cma_port(const struct sockaddr *addr)
1227 struct sockaddr_ib *sib;
1229 switch (addr->sa_family) {
1231 return ((struct sockaddr_in *) addr)->sin_port;
1233 return ((struct sockaddr_in6 *) addr)->sin6_port;
1235 sib = (struct sockaddr_ib *) addr;
1236 return htons((u16) (be64_to_cpu(sib->sib_sid) &
1237 be64_to_cpu(sib->sib_sid_mask)));
1243 static inline int cma_any_port(const struct sockaddr *addr)
1245 return !cma_port(addr);
1248 static void cma_save_ib_info(struct sockaddr *src_addr,
1249 struct sockaddr *dst_addr,
1250 const struct rdma_cm_id *listen_id,
1251 const struct sa_path_rec *path)
1253 struct sockaddr_ib *listen_ib, *ib;
1255 listen_ib = (struct sockaddr_ib *) &listen_id->route.addr.src_addr;
1257 ib = (struct sockaddr_ib *)src_addr;
1258 ib->sib_family = AF_IB;
1260 ib->sib_pkey = path->pkey;
1261 ib->sib_flowinfo = path->flow_label;
1262 memcpy(&ib->sib_addr, &path->sgid, 16);
1263 ib->sib_sid = path->service_id;
1264 ib->sib_scope_id = 0;
1266 ib->sib_pkey = listen_ib->sib_pkey;
1267 ib->sib_flowinfo = listen_ib->sib_flowinfo;
1268 ib->sib_addr = listen_ib->sib_addr;
1269 ib->sib_sid = listen_ib->sib_sid;
1270 ib->sib_scope_id = listen_ib->sib_scope_id;
1272 ib->sib_sid_mask = cpu_to_be64(0xffffffffffffffffULL);
1275 ib = (struct sockaddr_ib *)dst_addr;
1276 ib->sib_family = AF_IB;
1278 ib->sib_pkey = path->pkey;
1279 ib->sib_flowinfo = path->flow_label;
1280 memcpy(&ib->sib_addr, &path->dgid, 16);
1285 static void cma_save_ip4_info(struct sockaddr_in *src_addr,
1286 struct sockaddr_in *dst_addr,
1287 struct cma_hdr *hdr,
1291 *src_addr = (struct sockaddr_in) {
1292 .sin_family = AF_INET,
1293 .sin_addr.s_addr = hdr->dst_addr.ip4.addr,
1294 .sin_port = local_port,
1299 *dst_addr = (struct sockaddr_in) {
1300 .sin_family = AF_INET,
1301 .sin_addr.s_addr = hdr->src_addr.ip4.addr,
1302 .sin_port = hdr->port,
1307 static void cma_save_ip6_info(struct sockaddr_in6 *src_addr,
1308 struct sockaddr_in6 *dst_addr,
1309 struct cma_hdr *hdr,
1313 *src_addr = (struct sockaddr_in6) {
1314 .sin6_family = AF_INET6,
1315 .sin6_addr = hdr->dst_addr.ip6,
1316 .sin6_port = local_port,
1321 *dst_addr = (struct sockaddr_in6) {
1322 .sin6_family = AF_INET6,
1323 .sin6_addr = hdr->src_addr.ip6,
1324 .sin6_port = hdr->port,
1329 static u16 cma_port_from_service_id(__be64 service_id)
1331 return (u16)be64_to_cpu(service_id);
1334 static int cma_save_ip_info(struct sockaddr *src_addr,
1335 struct sockaddr *dst_addr,
1336 const struct ib_cm_event *ib_event,
1339 struct cma_hdr *hdr;
1342 hdr = ib_event->private_data;
1343 if (hdr->cma_version != CMA_VERSION)
1346 port = htons(cma_port_from_service_id(service_id));
1348 switch (cma_get_ip_ver(hdr)) {
1350 cma_save_ip4_info((struct sockaddr_in *)src_addr,
1351 (struct sockaddr_in *)dst_addr, hdr, port);
1354 cma_save_ip6_info((struct sockaddr_in6 *)src_addr,
1355 (struct sockaddr_in6 *)dst_addr, hdr, port);
1358 return -EAFNOSUPPORT;
1364 static int cma_save_net_info(struct sockaddr *src_addr,
1365 struct sockaddr *dst_addr,
1366 const struct rdma_cm_id *listen_id,
1367 const struct ib_cm_event *ib_event,
1368 sa_family_t sa_family, __be64 service_id)
1370 if (sa_family == AF_IB) {
1371 if (ib_event->event == IB_CM_REQ_RECEIVED)
1372 cma_save_ib_info(src_addr, dst_addr, listen_id,
1373 ib_event->param.req_rcvd.primary_path);
1374 else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED)
1375 cma_save_ib_info(src_addr, dst_addr, listen_id, NULL);
1379 return cma_save_ip_info(src_addr, dst_addr, ib_event, service_id);
1382 static int cma_save_req_info(const struct ib_cm_event *ib_event,
1383 struct cma_req_info *req)
1385 const struct ib_cm_req_event_param *req_param =
1386 &ib_event->param.req_rcvd;
1387 const struct ib_cm_sidr_req_event_param *sidr_param =
1388 &ib_event->param.sidr_req_rcvd;
1390 switch (ib_event->event) {
1391 case IB_CM_REQ_RECEIVED:
1392 req->device = req_param->listen_id->device;
1393 req->port = req_param->port;
1394 memcpy(&req->local_gid, &req_param->primary_path->sgid,
1395 sizeof(req->local_gid));
1396 req->has_gid = true;
1397 req->service_id = req_param->primary_path->service_id;
1398 req->pkey = be16_to_cpu(req_param->primary_path->pkey);
1399 if (req->pkey != req_param->bth_pkey)
1400 pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and primary path P_Key (0x%x)\n"
1401 "RDMA CMA: in the future this may cause the request to be dropped\n",
1402 req_param->bth_pkey, req->pkey);
1404 case IB_CM_SIDR_REQ_RECEIVED:
1405 req->device = sidr_param->listen_id->device;
1406 req->port = sidr_param->port;
1407 req->has_gid = false;
1408 req->service_id = sidr_param->service_id;
1409 req->pkey = sidr_param->pkey;
1410 if (req->pkey != sidr_param->bth_pkey)
1411 pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and SIDR request payload P_Key (0x%x)\n"
1412 "RDMA CMA: in the future this may cause the request to be dropped\n",
1413 sidr_param->bth_pkey, req->pkey);
1422 static bool validate_ipv4_net_dev(struct net_device *net_dev,
1423 const struct sockaddr_in *dst_addr,
1424 const struct sockaddr_in *src_addr)
1426 __be32 daddr = dst_addr->sin_addr.s_addr,
1427 saddr = src_addr->sin_addr.s_addr;
1428 struct fib_result res;
1433 if (ipv4_is_multicast(saddr) || ipv4_is_lbcast(saddr) ||
1434 ipv4_is_lbcast(daddr) || ipv4_is_zeronet(saddr) ||
1435 ipv4_is_zeronet(daddr) || ipv4_is_loopback(daddr) ||
1436 ipv4_is_loopback(saddr))
1439 memset(&fl4, 0, sizeof(fl4));
1440 fl4.flowi4_iif = net_dev->ifindex;
1445 err = fib_lookup(dev_net(net_dev), &fl4, &res, 0);
1446 ret = err == 0 && FIB_RES_DEV(res) == net_dev;
1452 static bool validate_ipv6_net_dev(struct net_device *net_dev,
1453 const struct sockaddr_in6 *dst_addr,
1454 const struct sockaddr_in6 *src_addr)
1456 #if IS_ENABLED(CONFIG_IPV6)
1457 const int strict = ipv6_addr_type(&dst_addr->sin6_addr) &
1458 IPV6_ADDR_LINKLOCAL;
1459 struct rt6_info *rt = rt6_lookup(dev_net(net_dev), &dst_addr->sin6_addr,
1460 &src_addr->sin6_addr, net_dev->ifindex,
1467 ret = rt->rt6i_idev->dev == net_dev;
1476 static bool validate_net_dev(struct net_device *net_dev,
1477 const struct sockaddr *daddr,
1478 const struct sockaddr *saddr)
1480 const struct sockaddr_in *daddr4 = (const struct sockaddr_in *)daddr;
1481 const struct sockaddr_in *saddr4 = (const struct sockaddr_in *)saddr;
1482 const struct sockaddr_in6 *daddr6 = (const struct sockaddr_in6 *)daddr;
1483 const struct sockaddr_in6 *saddr6 = (const struct sockaddr_in6 *)saddr;
1485 switch (daddr->sa_family) {
1487 return saddr->sa_family == AF_INET &&
1488 validate_ipv4_net_dev(net_dev, daddr4, saddr4);
1491 return saddr->sa_family == AF_INET6 &&
1492 validate_ipv6_net_dev(net_dev, daddr6, saddr6);
1499 static struct net_device *
1500 roce_get_net_dev_by_cm_event(const struct ib_cm_event *ib_event)
1502 const struct ib_gid_attr *sgid_attr = NULL;
1503 struct net_device *ndev;
1505 if (ib_event->event == IB_CM_REQ_RECEIVED)
1506 sgid_attr = ib_event->param.req_rcvd.ppath_sgid_attr;
1507 else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED)
1508 sgid_attr = ib_event->param.sidr_req_rcvd.sgid_attr;
1514 ndev = rdma_read_gid_attr_ndev_rcu(sgid_attr);
1523 static struct net_device *cma_get_net_dev(const struct ib_cm_event *ib_event,
1524 struct cma_req_info *req)
1526 struct sockaddr *listen_addr =
1527 (struct sockaddr *)&req->listen_addr_storage;
1528 struct sockaddr *src_addr = (struct sockaddr *)&req->src_addr_storage;
1529 struct net_device *net_dev;
1530 const union ib_gid *gid = req->has_gid ? &req->local_gid : NULL;
1533 err = cma_save_ip_info(listen_addr, src_addr, ib_event,
1536 return ERR_PTR(err);
1538 if (rdma_protocol_roce(req->device, req->port))
1539 net_dev = roce_get_net_dev_by_cm_event(ib_event);
1541 net_dev = ib_get_net_dev_by_params(req->device, req->port,
1545 return ERR_PTR(-ENODEV);
1550 static enum rdma_ucm_port_space rdma_ps_from_service_id(__be64 service_id)
1552 return (be64_to_cpu(service_id) >> 16) & 0xffff;
1555 static bool cma_match_private_data(struct rdma_id_private *id_priv,
1556 const struct cma_hdr *hdr)
1558 struct sockaddr *addr = cma_src_addr(id_priv);
1560 struct in6_addr ip6_addr;
1562 if (cma_any_addr(addr) && !id_priv->afonly)
1565 switch (addr->sa_family) {
1567 ip4_addr = ((struct sockaddr_in *)addr)->sin_addr.s_addr;
1568 if (cma_get_ip_ver(hdr) != 4)
1570 if (!cma_any_addr(addr) &&
1571 hdr->dst_addr.ip4.addr != ip4_addr)
1575 ip6_addr = ((struct sockaddr_in6 *)addr)->sin6_addr;
1576 if (cma_get_ip_ver(hdr) != 6)
1578 if (!cma_any_addr(addr) &&
1579 memcmp(&hdr->dst_addr.ip6, &ip6_addr, sizeof(ip6_addr)))
1591 static bool cma_protocol_roce(const struct rdma_cm_id *id)
1593 struct ib_device *device = id->device;
1594 const int port_num = id->port_num ?: rdma_start_port(device);
1596 return rdma_protocol_roce(device, port_num);
1599 static bool cma_is_req_ipv6_ll(const struct cma_req_info *req)
1601 const struct sockaddr *daddr =
1602 (const struct sockaddr *)&req->listen_addr_storage;
1603 const struct sockaddr_in6 *daddr6 = (const struct sockaddr_in6 *)daddr;
1605 /* Returns true if the req is for IPv6 link local */
1606 return (daddr->sa_family == AF_INET6 &&
1607 (ipv6_addr_type(&daddr6->sin6_addr) & IPV6_ADDR_LINKLOCAL));
1610 static bool cma_match_net_dev(const struct rdma_cm_id *id,
1611 const struct net_device *net_dev,
1612 const struct cma_req_info *req)
1614 const struct rdma_addr *addr = &id->route.addr;
1617 /* This request is an AF_IB request */
1618 return (!id->port_num || id->port_num == req->port) &&
1619 (addr->src_addr.ss_family == AF_IB);
1622 * If the request is not for IPv6 link local, allow matching
1623 * request to any netdevice of the one or multiport rdma device.
1625 if (!cma_is_req_ipv6_ll(req))
1628 * Net namespaces must match, and if the listner is listening
1629 * on a specific netdevice than netdevice must match as well.
1631 if (net_eq(dev_net(net_dev), addr->dev_addr.net) &&
1632 (!!addr->dev_addr.bound_dev_if ==
1633 (addr->dev_addr.bound_dev_if == net_dev->ifindex)))
1639 static struct rdma_id_private *cma_find_listener(
1640 const struct rdma_bind_list *bind_list,
1641 const struct ib_cm_id *cm_id,
1642 const struct ib_cm_event *ib_event,
1643 const struct cma_req_info *req,
1644 const struct net_device *net_dev)
1646 struct rdma_id_private *id_priv, *id_priv_dev;
1648 lockdep_assert_held(&lock);
1651 return ERR_PTR(-EINVAL);
1653 hlist_for_each_entry(id_priv, &bind_list->owners, node) {
1654 if (cma_match_private_data(id_priv, ib_event->private_data)) {
1655 if (id_priv->id.device == cm_id->device &&
1656 cma_match_net_dev(&id_priv->id, net_dev, req))
1658 list_for_each_entry(id_priv_dev,
1659 &id_priv->listen_list,
1661 if (id_priv_dev->id.device == cm_id->device &&
1662 cma_match_net_dev(&id_priv_dev->id,
1669 return ERR_PTR(-EINVAL);
1672 static struct rdma_id_private *
1673 cma_ib_id_from_event(struct ib_cm_id *cm_id,
1674 const struct ib_cm_event *ib_event,
1675 struct cma_req_info *req,
1676 struct net_device **net_dev)
1678 struct rdma_bind_list *bind_list;
1679 struct rdma_id_private *id_priv;
1682 err = cma_save_req_info(ib_event, req);
1684 return ERR_PTR(err);
1686 *net_dev = cma_get_net_dev(ib_event, req);
1687 if (IS_ERR(*net_dev)) {
1688 if (PTR_ERR(*net_dev) == -EAFNOSUPPORT) {
1689 /* Assuming the protocol is AF_IB */
1692 return ERR_CAST(*net_dev);
1698 * Net namespace might be getting deleted while route lookup,
1699 * cm_id lookup is in progress. Therefore, perform netdevice
1700 * validation, cm_id lookup under rcu lock.
1701 * RCU lock along with netdevice state check, synchronizes with
1702 * netdevice migrating to different net namespace and also avoids
1703 * case where net namespace doesn't get deleted while lookup is in
1705 * If the device state is not IFF_UP, its properties such as ifindex
1706 * and nd_net cannot be trusted to remain valid without rcu lock.
1707 * net/core/dev.c change_net_namespace() ensures to synchronize with
1708 * ongoing operations on net device after device is closed using
1709 * synchronize_net().
1714 * If netdevice is down, it is likely that it is administratively
1715 * down or it might be migrating to different namespace.
1716 * In that case avoid further processing, as the net namespace
1717 * or ifindex may change.
1719 if (((*net_dev)->flags & IFF_UP) == 0) {
1720 id_priv = ERR_PTR(-EHOSTUNREACH);
1724 if (!validate_net_dev(*net_dev,
1725 (struct sockaddr *)&req->src_addr_storage,
1726 (struct sockaddr *)&req->listen_addr_storage)) {
1727 id_priv = ERR_PTR(-EHOSTUNREACH);
1732 bind_list = cma_ps_find(*net_dev ? dev_net(*net_dev) : &init_net,
1733 rdma_ps_from_service_id(req->service_id),
1734 cma_port_from_service_id(req->service_id));
1735 id_priv = cma_find_listener(bind_list, cm_id, ib_event, req, *net_dev);
1738 mutex_unlock(&lock);
1739 if (IS_ERR(id_priv) && *net_dev) {
1746 static inline u8 cma_user_data_offset(struct rdma_id_private *id_priv)
1748 return cma_family(id_priv) == AF_IB ? 0 : sizeof(struct cma_hdr);
1751 static void cma_cancel_route(struct rdma_id_private *id_priv)
1753 if (rdma_cap_ib_sa(id_priv->id.device, id_priv->id.port_num)) {
1755 ib_sa_cancel_query(id_priv->query_id, id_priv->query);
1759 static void _cma_cancel_listens(struct rdma_id_private *id_priv)
1761 struct rdma_id_private *dev_id_priv;
1763 lockdep_assert_held(&lock);
1766 * Remove from listen_any_list to prevent added devices from spawning
1767 * additional listen requests.
1769 list_del(&id_priv->list);
1771 while (!list_empty(&id_priv->listen_list)) {
1772 dev_id_priv = list_entry(id_priv->listen_list.next,
1773 struct rdma_id_private, listen_list);
1774 /* sync with device removal to avoid duplicate destruction */
1775 list_del_init(&dev_id_priv->list);
1776 list_del(&dev_id_priv->listen_list);
1777 mutex_unlock(&lock);
1779 rdma_destroy_id(&dev_id_priv->id);
1784 static void cma_cancel_listens(struct rdma_id_private *id_priv)
1787 _cma_cancel_listens(id_priv);
1788 mutex_unlock(&lock);
1791 static void cma_cancel_operation(struct rdma_id_private *id_priv,
1792 enum rdma_cm_state state)
1795 case RDMA_CM_ADDR_QUERY:
1796 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
1798 case RDMA_CM_ROUTE_QUERY:
1799 cma_cancel_route(id_priv);
1801 case RDMA_CM_LISTEN:
1802 if (cma_any_addr(cma_src_addr(id_priv)) && !id_priv->cma_dev)
1803 cma_cancel_listens(id_priv);
1810 static void cma_release_port(struct rdma_id_private *id_priv)
1812 struct rdma_bind_list *bind_list = id_priv->bind_list;
1813 struct net *net = id_priv->id.route.addr.dev_addr.net;
1819 hlist_del(&id_priv->node);
1820 if (hlist_empty(&bind_list->owners)) {
1821 cma_ps_remove(net, bind_list->ps, bind_list->port);
1824 mutex_unlock(&lock);
1827 static void destroy_mc(struct rdma_id_private *id_priv,
1828 struct cma_multicast *mc)
1830 bool send_only = mc->join_state == BIT(SENDONLY_FULLMEMBER_JOIN);
1832 if (rdma_cap_ib_mcast(id_priv->id.device, id_priv->id.port_num))
1833 ib_sa_free_multicast(mc->sa_mc);
1835 if (rdma_protocol_roce(id_priv->id.device, id_priv->id.port_num)) {
1836 struct rdma_dev_addr *dev_addr =
1837 &id_priv->id.route.addr.dev_addr;
1838 struct net_device *ndev = NULL;
1840 if (dev_addr->bound_dev_if)
1841 ndev = dev_get_by_index(dev_addr->net,
1842 dev_addr->bound_dev_if);
1843 if (ndev && !send_only) {
1844 enum ib_gid_type gid_type;
1847 gid_type = id_priv->cma_dev->default_gid_type
1848 [id_priv->id.port_num -
1850 id_priv->cma_dev->device)];
1851 cma_iboe_set_mgid((struct sockaddr *)&mc->addr, &mgid,
1853 cma_igmp_send(ndev, &mgid, false);
1857 cancel_work_sync(&mc->iboe_join.work);
1862 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
1864 struct cma_multicast *mc;
1866 while (!list_empty(&id_priv->mc_list)) {
1867 mc = list_first_entry(&id_priv->mc_list, struct cma_multicast,
1869 list_del(&mc->list);
1870 destroy_mc(id_priv, mc);
1874 static void _destroy_id(struct rdma_id_private *id_priv,
1875 enum rdma_cm_state state)
1877 cma_cancel_operation(id_priv, state);
1879 rdma_restrack_del(&id_priv->res);
1880 if (id_priv->cma_dev) {
1881 if (rdma_cap_ib_cm(id_priv->id.device, 1)) {
1882 if (id_priv->cm_id.ib)
1883 ib_destroy_cm_id(id_priv->cm_id.ib);
1884 } else if (rdma_cap_iw_cm(id_priv->id.device, 1)) {
1885 if (id_priv->cm_id.iw)
1886 iw_destroy_cm_id(id_priv->cm_id.iw);
1888 cma_leave_mc_groups(id_priv);
1889 cma_release_dev(id_priv);
1892 cma_release_port(id_priv);
1893 cma_id_put(id_priv);
1894 wait_for_completion(&id_priv->comp);
1896 if (id_priv->internal_id)
1897 cma_id_put(id_priv->id.context);
1899 kfree(id_priv->id.route.path_rec);
1901 put_net(id_priv->id.route.addr.dev_addr.net);
1906 * destroy an ID from within the handler_mutex. This ensures that no other
1907 * handlers can start running concurrently.
1909 static void destroy_id_handler_unlock(struct rdma_id_private *id_priv)
1910 __releases(&idprv->handler_mutex)
1912 enum rdma_cm_state state;
1913 unsigned long flags;
1915 trace_cm_id_destroy(id_priv);
1918 * Setting the state to destroyed under the handler mutex provides a
1919 * fence against calling handler callbacks. If this is invoked due to
1920 * the failure of a handler callback then it guarentees that no future
1921 * handlers will be called.
1923 lockdep_assert_held(&id_priv->handler_mutex);
1924 spin_lock_irqsave(&id_priv->lock, flags);
1925 state = id_priv->state;
1926 id_priv->state = RDMA_CM_DESTROYING;
1927 spin_unlock_irqrestore(&id_priv->lock, flags);
1928 mutex_unlock(&id_priv->handler_mutex);
1929 _destroy_id(id_priv, state);
1932 void rdma_destroy_id(struct rdma_cm_id *id)
1934 struct rdma_id_private *id_priv =
1935 container_of(id, struct rdma_id_private, id);
1937 mutex_lock(&id_priv->handler_mutex);
1938 destroy_id_handler_unlock(id_priv);
1940 EXPORT_SYMBOL(rdma_destroy_id);
1942 static int cma_rep_recv(struct rdma_id_private *id_priv)
1946 ret = cma_modify_qp_rtr(id_priv, NULL);
1950 ret = cma_modify_qp_rts(id_priv, NULL);
1954 trace_cm_send_rtu(id_priv);
1955 ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
1961 pr_debug_ratelimited("RDMA CM: CONNECT_ERROR: failed to handle reply. status %d\n", ret);
1962 cma_modify_qp_err(id_priv);
1963 trace_cm_send_rej(id_priv);
1964 ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
1969 static void cma_set_rep_event_data(struct rdma_cm_event *event,
1970 const struct ib_cm_rep_event_param *rep_data,
1973 event->param.conn.private_data = private_data;
1974 event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
1975 event->param.conn.responder_resources = rep_data->responder_resources;
1976 event->param.conn.initiator_depth = rep_data->initiator_depth;
1977 event->param.conn.flow_control = rep_data->flow_control;
1978 event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
1979 event->param.conn.srq = rep_data->srq;
1980 event->param.conn.qp_num = rep_data->remote_qpn;
1982 event->ece.vendor_id = rep_data->ece.vendor_id;
1983 event->ece.attr_mod = rep_data->ece.attr_mod;
1986 static int cma_cm_event_handler(struct rdma_id_private *id_priv,
1987 struct rdma_cm_event *event)
1991 lockdep_assert_held(&id_priv->handler_mutex);
1993 trace_cm_event_handler(id_priv, event);
1994 ret = id_priv->id.event_handler(&id_priv->id, event);
1995 trace_cm_event_done(id_priv, event, ret);
1999 static int cma_ib_handler(struct ib_cm_id *cm_id,
2000 const struct ib_cm_event *ib_event)
2002 struct rdma_id_private *id_priv = cm_id->context;
2003 struct rdma_cm_event event = {};
2004 enum rdma_cm_state state;
2007 mutex_lock(&id_priv->handler_mutex);
2008 state = READ_ONCE(id_priv->state);
2009 if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
2010 state != RDMA_CM_CONNECT) ||
2011 (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
2012 state != RDMA_CM_DISCONNECT))
2015 switch (ib_event->event) {
2016 case IB_CM_REQ_ERROR:
2017 case IB_CM_REP_ERROR:
2018 event.event = RDMA_CM_EVENT_UNREACHABLE;
2019 event.status = -ETIMEDOUT;
2021 case IB_CM_REP_RECEIVED:
2022 if (state == RDMA_CM_CONNECT &&
2023 (id_priv->id.qp_type != IB_QPT_UD)) {
2024 trace_cm_send_mra(id_priv);
2025 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
2027 if (id_priv->id.qp) {
2028 event.status = cma_rep_recv(id_priv);
2029 event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
2030 RDMA_CM_EVENT_ESTABLISHED;
2032 event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
2034 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
2035 ib_event->private_data);
2037 case IB_CM_RTU_RECEIVED:
2038 case IB_CM_USER_ESTABLISHED:
2039 event.event = RDMA_CM_EVENT_ESTABLISHED;
2041 case IB_CM_DREQ_ERROR:
2042 event.status = -ETIMEDOUT;
2044 case IB_CM_DREQ_RECEIVED:
2045 case IB_CM_DREP_RECEIVED:
2046 if (!cma_comp_exch(id_priv, RDMA_CM_CONNECT,
2047 RDMA_CM_DISCONNECT))
2049 event.event = RDMA_CM_EVENT_DISCONNECTED;
2051 case IB_CM_TIMEWAIT_EXIT:
2052 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
2054 case IB_CM_MRA_RECEIVED:
2057 case IB_CM_REJ_RECEIVED:
2058 pr_debug_ratelimited("RDMA CM: REJECTED: %s\n", rdma_reject_msg(&id_priv->id,
2059 ib_event->param.rej_rcvd.reason));
2060 cma_modify_qp_err(id_priv);
2061 event.status = ib_event->param.rej_rcvd.reason;
2062 event.event = RDMA_CM_EVENT_REJECTED;
2063 event.param.conn.private_data = ib_event->private_data;
2064 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
2067 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
2072 ret = cma_cm_event_handler(id_priv, &event);
2074 /* Destroy the CM ID by returning a non-zero value. */
2075 id_priv->cm_id.ib = NULL;
2076 destroy_id_handler_unlock(id_priv);
2080 mutex_unlock(&id_priv->handler_mutex);
2084 static struct rdma_id_private *
2085 cma_ib_new_conn_id(const struct rdma_cm_id *listen_id,
2086 const struct ib_cm_event *ib_event,
2087 struct net_device *net_dev)
2089 struct rdma_id_private *listen_id_priv;
2090 struct rdma_id_private *id_priv;
2091 struct rdma_cm_id *id;
2092 struct rdma_route *rt;
2093 const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
2094 struct sa_path_rec *path = ib_event->param.req_rcvd.primary_path;
2095 const __be64 service_id =
2096 ib_event->param.req_rcvd.primary_path->service_id;
2099 listen_id_priv = container_of(listen_id, struct rdma_id_private, id);
2100 id_priv = __rdma_create_id(listen_id->route.addr.dev_addr.net,
2101 listen_id->event_handler, listen_id->context,
2103 ib_event->param.req_rcvd.qp_type,
2105 if (IS_ERR(id_priv))
2109 if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
2110 (struct sockaddr *)&id->route.addr.dst_addr,
2111 listen_id, ib_event, ss_family, service_id))
2115 rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
2116 rt->path_rec = kmalloc_array(rt->num_paths, sizeof(*rt->path_rec),
2121 rt->path_rec[0] = *path;
2122 if (rt->num_paths == 2)
2123 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
2126 rdma_copy_src_l2_addr(&rt->addr.dev_addr, net_dev);
2128 if (!cma_protocol_roce(listen_id) &&
2129 cma_any_addr(cma_src_addr(id_priv))) {
2130 rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
2131 rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
2132 ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
2133 } else if (!cma_any_addr(cma_src_addr(id_priv))) {
2134 ret = cma_translate_addr(cma_src_addr(id_priv), &rt->addr.dev_addr);
2139 rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
2141 id_priv->state = RDMA_CM_CONNECT;
2145 rdma_destroy_id(id);
2149 static struct rdma_id_private *
2150 cma_ib_new_udp_id(const struct rdma_cm_id *listen_id,
2151 const struct ib_cm_event *ib_event,
2152 struct net_device *net_dev)
2154 const struct rdma_id_private *listen_id_priv;
2155 struct rdma_id_private *id_priv;
2156 struct rdma_cm_id *id;
2157 const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
2158 struct net *net = listen_id->route.addr.dev_addr.net;
2161 listen_id_priv = container_of(listen_id, struct rdma_id_private, id);
2162 id_priv = __rdma_create_id(net, listen_id->event_handler,
2163 listen_id->context, listen_id->ps, IB_QPT_UD,
2165 if (IS_ERR(id_priv))
2169 if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
2170 (struct sockaddr *)&id->route.addr.dst_addr,
2171 listen_id, ib_event, ss_family,
2172 ib_event->param.sidr_req_rcvd.service_id))
2176 rdma_copy_src_l2_addr(&id->route.addr.dev_addr, net_dev);
2178 if (!cma_any_addr(cma_src_addr(id_priv))) {
2179 ret = cma_translate_addr(cma_src_addr(id_priv),
2180 &id->route.addr.dev_addr);
2186 id_priv->state = RDMA_CM_CONNECT;
2189 rdma_destroy_id(id);
2193 static void cma_set_req_event_data(struct rdma_cm_event *event,
2194 const struct ib_cm_req_event_param *req_data,
2195 void *private_data, int offset)
2197 event->param.conn.private_data = private_data + offset;
2198 event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
2199 event->param.conn.responder_resources = req_data->responder_resources;
2200 event->param.conn.initiator_depth = req_data->initiator_depth;
2201 event->param.conn.flow_control = req_data->flow_control;
2202 event->param.conn.retry_count = req_data->retry_count;
2203 event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
2204 event->param.conn.srq = req_data->srq;
2205 event->param.conn.qp_num = req_data->remote_qpn;
2207 event->ece.vendor_id = req_data->ece.vendor_id;
2208 event->ece.attr_mod = req_data->ece.attr_mod;
2211 static int cma_ib_check_req_qp_type(const struct rdma_cm_id *id,
2212 const struct ib_cm_event *ib_event)
2214 return (((ib_event->event == IB_CM_REQ_RECEIVED) &&
2215 (ib_event->param.req_rcvd.qp_type == id->qp_type)) ||
2216 ((ib_event->event == IB_CM_SIDR_REQ_RECEIVED) &&
2217 (id->qp_type == IB_QPT_UD)) ||
2221 static int cma_ib_req_handler(struct ib_cm_id *cm_id,
2222 const struct ib_cm_event *ib_event)
2224 struct rdma_id_private *listen_id, *conn_id = NULL;
2225 struct rdma_cm_event event = {};
2226 struct cma_req_info req = {};
2227 struct net_device *net_dev;
2231 listen_id = cma_ib_id_from_event(cm_id, ib_event, &req, &net_dev);
2232 if (IS_ERR(listen_id))
2233 return PTR_ERR(listen_id);
2235 trace_cm_req_handler(listen_id, ib_event->event);
2236 if (!cma_ib_check_req_qp_type(&listen_id->id, ib_event)) {
2241 mutex_lock(&listen_id->handler_mutex);
2242 if (READ_ONCE(listen_id->state) != RDMA_CM_LISTEN) {
2243 ret = -ECONNABORTED;
2247 offset = cma_user_data_offset(listen_id);
2248 event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
2249 if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) {
2250 conn_id = cma_ib_new_udp_id(&listen_id->id, ib_event, net_dev);
2251 event.param.ud.private_data = ib_event->private_data + offset;
2252 event.param.ud.private_data_len =
2253 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
2255 conn_id = cma_ib_new_conn_id(&listen_id->id, ib_event, net_dev);
2256 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
2257 ib_event->private_data, offset);
2264 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
2265 ret = cma_ib_acquire_dev(conn_id, listen_id, &req);
2267 destroy_id_handler_unlock(conn_id);
2271 conn_id->cm_id.ib = cm_id;
2272 cm_id->context = conn_id;
2273 cm_id->cm_handler = cma_ib_handler;
2275 ret = cma_cm_event_handler(conn_id, &event);
2277 /* Destroy the CM ID by returning a non-zero value. */
2278 conn_id->cm_id.ib = NULL;
2279 mutex_unlock(&listen_id->handler_mutex);
2280 destroy_id_handler_unlock(conn_id);
2284 if (READ_ONCE(conn_id->state) == RDMA_CM_CONNECT &&
2285 conn_id->id.qp_type != IB_QPT_UD) {
2286 trace_cm_send_mra(cm_id->context);
2287 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
2289 mutex_unlock(&conn_id->handler_mutex);
2292 mutex_unlock(&listen_id->handler_mutex);
2301 __be64 rdma_get_service_id(struct rdma_cm_id *id, struct sockaddr *addr)
2303 if (addr->sa_family == AF_IB)
2304 return ((struct sockaddr_ib *) addr)->sib_sid;
2306 return cpu_to_be64(((u64)id->ps << 16) + be16_to_cpu(cma_port(addr)));
2308 EXPORT_SYMBOL(rdma_get_service_id);
2310 void rdma_read_gids(struct rdma_cm_id *cm_id, union ib_gid *sgid,
2313 struct rdma_addr *addr = &cm_id->route.addr;
2315 if (!cm_id->device) {
2317 memset(sgid, 0, sizeof(*sgid));
2319 memset(dgid, 0, sizeof(*dgid));
2323 if (rdma_protocol_roce(cm_id->device, cm_id->port_num)) {
2325 rdma_ip2gid((struct sockaddr *)&addr->src_addr, sgid);
2327 rdma_ip2gid((struct sockaddr *)&addr->dst_addr, dgid);
2330 rdma_addr_get_sgid(&addr->dev_addr, sgid);
2332 rdma_addr_get_dgid(&addr->dev_addr, dgid);
2335 EXPORT_SYMBOL(rdma_read_gids);
2337 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
2339 struct rdma_id_private *id_priv = iw_id->context;
2340 struct rdma_cm_event event = {};
2342 struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
2343 struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
2345 mutex_lock(&id_priv->handler_mutex);
2346 if (READ_ONCE(id_priv->state) != RDMA_CM_CONNECT)
2349 switch (iw_event->event) {
2350 case IW_CM_EVENT_CLOSE:
2351 event.event = RDMA_CM_EVENT_DISCONNECTED;
2353 case IW_CM_EVENT_CONNECT_REPLY:
2354 memcpy(cma_src_addr(id_priv), laddr,
2355 rdma_addr_size(laddr));
2356 memcpy(cma_dst_addr(id_priv), raddr,
2357 rdma_addr_size(raddr));
2358 switch (iw_event->status) {
2360 event.event = RDMA_CM_EVENT_ESTABLISHED;
2361 event.param.conn.initiator_depth = iw_event->ird;
2362 event.param.conn.responder_resources = iw_event->ord;
2366 event.event = RDMA_CM_EVENT_REJECTED;
2369 event.event = RDMA_CM_EVENT_UNREACHABLE;
2372 event.event = RDMA_CM_EVENT_CONNECT_ERROR;
2376 case IW_CM_EVENT_ESTABLISHED:
2377 event.event = RDMA_CM_EVENT_ESTABLISHED;
2378 event.param.conn.initiator_depth = iw_event->ird;
2379 event.param.conn.responder_resources = iw_event->ord;
2385 event.status = iw_event->status;
2386 event.param.conn.private_data = iw_event->private_data;
2387 event.param.conn.private_data_len = iw_event->private_data_len;
2388 ret = cma_cm_event_handler(id_priv, &event);
2390 /* Destroy the CM ID by returning a non-zero value. */
2391 id_priv->cm_id.iw = NULL;
2392 destroy_id_handler_unlock(id_priv);
2397 mutex_unlock(&id_priv->handler_mutex);
2401 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
2402 struct iw_cm_event *iw_event)
2404 struct rdma_id_private *listen_id, *conn_id;
2405 struct rdma_cm_event event = {};
2406 int ret = -ECONNABORTED;
2407 struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
2408 struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
2410 event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
2411 event.param.conn.private_data = iw_event->private_data;
2412 event.param.conn.private_data_len = iw_event->private_data_len;
2413 event.param.conn.initiator_depth = iw_event->ird;
2414 event.param.conn.responder_resources = iw_event->ord;
2416 listen_id = cm_id->context;
2418 mutex_lock(&listen_id->handler_mutex);
2419 if (READ_ONCE(listen_id->state) != RDMA_CM_LISTEN)
2422 /* Create a new RDMA id for the new IW CM ID */
2423 conn_id = __rdma_create_id(listen_id->id.route.addr.dev_addr.net,
2424 listen_id->id.event_handler,
2425 listen_id->id.context, RDMA_PS_TCP,
2426 IB_QPT_RC, listen_id);
2427 if (IS_ERR(conn_id)) {
2431 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
2432 conn_id->state = RDMA_CM_CONNECT;
2434 ret = rdma_translate_ip(laddr, &conn_id->id.route.addr.dev_addr);
2436 mutex_unlock(&listen_id->handler_mutex);
2437 destroy_id_handler_unlock(conn_id);
2441 ret = cma_iw_acquire_dev(conn_id, listen_id);
2443 mutex_unlock(&listen_id->handler_mutex);
2444 destroy_id_handler_unlock(conn_id);
2448 conn_id->cm_id.iw = cm_id;
2449 cm_id->context = conn_id;
2450 cm_id->cm_handler = cma_iw_handler;
2452 memcpy(cma_src_addr(conn_id), laddr, rdma_addr_size(laddr));
2453 memcpy(cma_dst_addr(conn_id), raddr, rdma_addr_size(raddr));
2455 ret = cma_cm_event_handler(conn_id, &event);
2457 /* User wants to destroy the CM ID */
2458 conn_id->cm_id.iw = NULL;
2459 mutex_unlock(&listen_id->handler_mutex);
2460 destroy_id_handler_unlock(conn_id);
2464 mutex_unlock(&conn_id->handler_mutex);
2467 mutex_unlock(&listen_id->handler_mutex);
2471 static int cma_ib_listen(struct rdma_id_private *id_priv)
2473 struct sockaddr *addr;
2474 struct ib_cm_id *id;
2477 addr = cma_src_addr(id_priv);
2478 svc_id = rdma_get_service_id(&id_priv->id, addr);
2479 id = ib_cm_insert_listen(id_priv->id.device,
2480 cma_ib_req_handler, svc_id);
2483 id_priv->cm_id.ib = id;
2488 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
2491 struct iw_cm_id *id;
2493 id = iw_create_cm_id(id_priv->id.device,
2494 iw_conn_req_handler,
2499 mutex_lock(&id_priv->qp_mutex);
2500 id->tos = id_priv->tos;
2501 id->tos_set = id_priv->tos_set;
2502 mutex_unlock(&id_priv->qp_mutex);
2503 id_priv->cm_id.iw = id;
2505 memcpy(&id_priv->cm_id.iw->local_addr, cma_src_addr(id_priv),
2506 rdma_addr_size(cma_src_addr(id_priv)));
2508 ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
2511 iw_destroy_cm_id(id_priv->cm_id.iw);
2512 id_priv->cm_id.iw = NULL;
2518 static int cma_listen_handler(struct rdma_cm_id *id,
2519 struct rdma_cm_event *event)
2521 struct rdma_id_private *id_priv = id->context;
2523 /* Listening IDs are always destroyed on removal */
2524 if (event->event == RDMA_CM_EVENT_DEVICE_REMOVAL)
2527 id->context = id_priv->id.context;
2528 id->event_handler = id_priv->id.event_handler;
2529 trace_cm_event_handler(id_priv, event);
2530 return id_priv->id.event_handler(id, event);
2533 static int cma_listen_on_dev(struct rdma_id_private *id_priv,
2534 struct cma_device *cma_dev,
2535 struct rdma_id_private **to_destroy)
2537 struct rdma_id_private *dev_id_priv;
2538 struct net *net = id_priv->id.route.addr.dev_addr.net;
2541 lockdep_assert_held(&lock);
2544 if (cma_family(id_priv) == AF_IB && !rdma_cap_ib_cm(cma_dev->device, 1))
2548 __rdma_create_id(net, cma_listen_handler, id_priv,
2549 id_priv->id.ps, id_priv->id.qp_type, id_priv);
2550 if (IS_ERR(dev_id_priv))
2551 return PTR_ERR(dev_id_priv);
2553 dev_id_priv->state = RDMA_CM_ADDR_BOUND;
2554 memcpy(cma_src_addr(dev_id_priv), cma_src_addr(id_priv),
2555 rdma_addr_size(cma_src_addr(id_priv)));
2557 _cma_attach_to_dev(dev_id_priv, cma_dev);
2558 rdma_restrack_add(&dev_id_priv->res);
2559 cma_id_get(id_priv);
2560 dev_id_priv->internal_id = 1;
2561 dev_id_priv->afonly = id_priv->afonly;
2562 mutex_lock(&id_priv->qp_mutex);
2563 dev_id_priv->tos_set = id_priv->tos_set;
2564 dev_id_priv->tos = id_priv->tos;
2565 mutex_unlock(&id_priv->qp_mutex);
2567 ret = rdma_listen(&dev_id_priv->id, id_priv->backlog);
2570 list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
2573 /* Caller must destroy this after releasing lock */
2574 *to_destroy = dev_id_priv;
2575 dev_warn(&cma_dev->device->dev, "RDMA CMA: %s, error %d\n", __func__, ret);
2579 static int cma_listen_on_all(struct rdma_id_private *id_priv)
2581 struct rdma_id_private *to_destroy;
2582 struct cma_device *cma_dev;
2586 list_add_tail(&id_priv->list, &listen_any_list);
2587 list_for_each_entry(cma_dev, &dev_list, list) {
2588 ret = cma_listen_on_dev(id_priv, cma_dev, &to_destroy);
2590 /* Prevent racing with cma_process_remove() */
2592 list_del_init(&to_destroy->list);
2596 mutex_unlock(&lock);
2600 _cma_cancel_listens(id_priv);
2601 mutex_unlock(&lock);
2603 rdma_destroy_id(&to_destroy->id);
2607 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
2609 struct rdma_id_private *id_priv;
2611 id_priv = container_of(id, struct rdma_id_private, id);
2612 mutex_lock(&id_priv->qp_mutex);
2613 id_priv->tos = (u8) tos;
2614 id_priv->tos_set = true;
2615 mutex_unlock(&id_priv->qp_mutex);
2617 EXPORT_SYMBOL(rdma_set_service_type);
2620 * rdma_set_ack_timeout() - Set the ack timeout of QP associated
2621 * with a connection identifier.
2622 * @id: Communication identifier to associated with service type.
2623 * @timeout: Ack timeout to set a QP, expressed as 4.096 * 2^(timeout) usec.
2625 * This function should be called before rdma_connect() on active side,
2626 * and on passive side before rdma_accept(). It is applicable to primary
2627 * path only. The timeout will affect the local side of the QP, it is not
2628 * negotiated with remote side and zero disables the timer. In case it is
2629 * set before rdma_resolve_route, the value will also be used to determine
2630 * PacketLifeTime for RoCE.
2632 * Return: 0 for success
2634 int rdma_set_ack_timeout(struct rdma_cm_id *id, u8 timeout)
2636 struct rdma_id_private *id_priv;
2638 if (id->qp_type != IB_QPT_RC && id->qp_type != IB_QPT_XRC_INI)
2641 id_priv = container_of(id, struct rdma_id_private, id);
2642 mutex_lock(&id_priv->qp_mutex);
2643 id_priv->timeout = timeout;
2644 id_priv->timeout_set = true;
2645 mutex_unlock(&id_priv->qp_mutex);
2649 EXPORT_SYMBOL(rdma_set_ack_timeout);
2651 static void cma_query_handler(int status, struct sa_path_rec *path_rec,
2654 struct cma_work *work = context;
2655 struct rdma_route *route;
2657 route = &work->id->id.route;
2660 route->num_paths = 1;
2661 *route->path_rec = *path_rec;
2663 work->old_state = RDMA_CM_ROUTE_QUERY;
2664 work->new_state = RDMA_CM_ADDR_RESOLVED;
2665 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
2666 work->event.status = status;
2667 pr_debug_ratelimited("RDMA CM: ROUTE_ERROR: failed to query path. status %d\n",
2671 queue_work(cma_wq, &work->work);
2674 static int cma_query_ib_route(struct rdma_id_private *id_priv,
2675 unsigned long timeout_ms, struct cma_work *work)
2677 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2678 struct sa_path_rec path_rec;
2679 ib_sa_comp_mask comp_mask;
2680 struct sockaddr_in6 *sin6;
2681 struct sockaddr_ib *sib;
2683 memset(&path_rec, 0, sizeof path_rec);
2685 if (rdma_cap_opa_ah(id_priv->id.device, id_priv->id.port_num))
2686 path_rec.rec_type = SA_PATH_REC_TYPE_OPA;
2688 path_rec.rec_type = SA_PATH_REC_TYPE_IB;
2689 rdma_addr_get_sgid(dev_addr, &path_rec.sgid);
2690 rdma_addr_get_dgid(dev_addr, &path_rec.dgid);
2691 path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
2692 path_rec.numb_path = 1;
2693 path_rec.reversible = 1;
2694 path_rec.service_id = rdma_get_service_id(&id_priv->id,
2695 cma_dst_addr(id_priv));
2697 comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
2698 IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
2699 IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
2701 switch (cma_family(id_priv)) {
2703 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
2704 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
2707 sin6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
2708 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
2709 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2712 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
2713 path_rec.traffic_class = (u8) (be32_to_cpu(sib->sib_flowinfo) >> 20);
2714 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2718 id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
2719 id_priv->id.port_num, &path_rec,
2720 comp_mask, timeout_ms,
2721 GFP_KERNEL, cma_query_handler,
2722 work, &id_priv->query);
2724 return (id_priv->query_id < 0) ? id_priv->query_id : 0;
2727 static void cma_iboe_join_work_handler(struct work_struct *work)
2729 struct cma_multicast *mc =
2730 container_of(work, struct cma_multicast, iboe_join.work);
2731 struct rdma_cm_event *event = &mc->iboe_join.event;
2732 struct rdma_id_private *id_priv = mc->id_priv;
2735 mutex_lock(&id_priv->handler_mutex);
2736 if (READ_ONCE(id_priv->state) == RDMA_CM_DESTROYING ||
2737 READ_ONCE(id_priv->state) == RDMA_CM_DEVICE_REMOVAL)
2740 ret = cma_cm_event_handler(id_priv, event);
2744 mutex_unlock(&id_priv->handler_mutex);
2745 if (event->event == RDMA_CM_EVENT_MULTICAST_JOIN)
2746 rdma_destroy_ah_attr(&event->param.ud.ah_attr);
2749 static void cma_work_handler(struct work_struct *_work)
2751 struct cma_work *work = container_of(_work, struct cma_work, work);
2752 struct rdma_id_private *id_priv = work->id;
2754 mutex_lock(&id_priv->handler_mutex);
2755 if (READ_ONCE(id_priv->state) == RDMA_CM_DESTROYING ||
2756 READ_ONCE(id_priv->state) == RDMA_CM_DEVICE_REMOVAL)
2758 if (work->old_state != 0 || work->new_state != 0) {
2759 if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
2763 if (cma_cm_event_handler(id_priv, &work->event)) {
2764 cma_id_put(id_priv);
2765 destroy_id_handler_unlock(id_priv);
2770 mutex_unlock(&id_priv->handler_mutex);
2771 cma_id_put(id_priv);
2773 if (work->event.event == RDMA_CM_EVENT_MULTICAST_JOIN)
2774 rdma_destroy_ah_attr(&work->event.param.ud.ah_attr);
2778 static void cma_init_resolve_route_work(struct cma_work *work,
2779 struct rdma_id_private *id_priv)
2782 INIT_WORK(&work->work, cma_work_handler);
2783 work->old_state = RDMA_CM_ROUTE_QUERY;
2784 work->new_state = RDMA_CM_ROUTE_RESOLVED;
2785 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
2788 static void enqueue_resolve_addr_work(struct cma_work *work,
2789 struct rdma_id_private *id_priv)
2791 /* Balances with cma_id_put() in cma_work_handler */
2792 cma_id_get(id_priv);
2795 INIT_WORK(&work->work, cma_work_handler);
2796 work->old_state = RDMA_CM_ADDR_QUERY;
2797 work->new_state = RDMA_CM_ADDR_RESOLVED;
2798 work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2800 queue_work(cma_wq, &work->work);
2803 static int cma_resolve_ib_route(struct rdma_id_private *id_priv,
2804 unsigned long timeout_ms)
2806 struct rdma_route *route = &id_priv->id.route;
2807 struct cma_work *work;
2810 work = kzalloc(sizeof *work, GFP_KERNEL);
2814 cma_init_resolve_route_work(work, id_priv);
2816 if (!route->path_rec)
2817 route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
2818 if (!route->path_rec) {
2823 ret = cma_query_ib_route(id_priv, timeout_ms, work);
2829 kfree(route->path_rec);
2830 route->path_rec = NULL;
2836 static enum ib_gid_type cma_route_gid_type(enum rdma_network_type network_type,
2837 unsigned long supported_gids,
2838 enum ib_gid_type default_gid)
2840 if ((network_type == RDMA_NETWORK_IPV4 ||
2841 network_type == RDMA_NETWORK_IPV6) &&
2842 test_bit(IB_GID_TYPE_ROCE_UDP_ENCAP, &supported_gids))
2843 return IB_GID_TYPE_ROCE_UDP_ENCAP;
2849 * cma_iboe_set_path_rec_l2_fields() is helper function which sets
2850 * path record type based on GID type.
2851 * It also sets up other L2 fields which includes destination mac address
2852 * netdev ifindex, of the path record.
2853 * It returns the netdev of the bound interface for this path record entry.
2855 static struct net_device *
2856 cma_iboe_set_path_rec_l2_fields(struct rdma_id_private *id_priv)
2858 struct rdma_route *route = &id_priv->id.route;
2859 enum ib_gid_type gid_type = IB_GID_TYPE_ROCE;
2860 struct rdma_addr *addr = &route->addr;
2861 unsigned long supported_gids;
2862 struct net_device *ndev;
2864 if (!addr->dev_addr.bound_dev_if)
2867 ndev = dev_get_by_index(addr->dev_addr.net,
2868 addr->dev_addr.bound_dev_if);
2872 supported_gids = roce_gid_type_mask_support(id_priv->id.device,
2873 id_priv->id.port_num);
2874 gid_type = cma_route_gid_type(addr->dev_addr.network,
2877 /* Use the hint from IP Stack to select GID Type */
2878 if (gid_type < ib_network_to_gid_type(addr->dev_addr.network))
2879 gid_type = ib_network_to_gid_type(addr->dev_addr.network);
2880 route->path_rec->rec_type = sa_conv_gid_to_pathrec_type(gid_type);
2882 route->path_rec->roce.route_resolved = true;
2883 sa_path_set_dmac(route->path_rec, addr->dev_addr.dst_dev_addr);
2887 int rdma_set_ib_path(struct rdma_cm_id *id,
2888 struct sa_path_rec *path_rec)
2890 struct rdma_id_private *id_priv;
2891 struct net_device *ndev;
2894 id_priv = container_of(id, struct rdma_id_private, id);
2895 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
2896 RDMA_CM_ROUTE_RESOLVED))
2899 id->route.path_rec = kmemdup(path_rec, sizeof(*path_rec),
2901 if (!id->route.path_rec) {
2906 if (rdma_protocol_roce(id->device, id->port_num)) {
2907 ndev = cma_iboe_set_path_rec_l2_fields(id_priv);
2915 id->route.num_paths = 1;
2919 kfree(id->route.path_rec);
2920 id->route.path_rec = NULL;
2922 cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED);
2925 EXPORT_SYMBOL(rdma_set_ib_path);
2927 static int cma_resolve_iw_route(struct rdma_id_private *id_priv)
2929 struct cma_work *work;
2931 work = kzalloc(sizeof *work, GFP_KERNEL);
2935 cma_init_resolve_route_work(work, id_priv);
2936 queue_work(cma_wq, &work->work);
2940 static int get_vlan_ndev_tc(struct net_device *vlan_ndev, int prio)
2942 struct net_device *dev;
2944 dev = vlan_dev_real_dev(vlan_ndev);
2946 return netdev_get_prio_tc_map(dev, prio);
2948 return (vlan_dev_get_egress_qos_mask(vlan_ndev, prio) &
2949 VLAN_PRIO_MASK) >> VLAN_PRIO_SHIFT;
2952 struct iboe_prio_tc_map {
2958 static int get_lower_vlan_dev_tc(struct net_device *dev,
2959 struct netdev_nested_priv *priv)
2961 struct iboe_prio_tc_map *map = (struct iboe_prio_tc_map *)priv->data;
2963 if (is_vlan_dev(dev))
2964 map->output_tc = get_vlan_ndev_tc(dev, map->input_prio);
2965 else if (dev->num_tc)
2966 map->output_tc = netdev_get_prio_tc_map(dev, map->input_prio);
2969 /* We are interested only in first level VLAN device, so always
2970 * return 1 to stop iterating over next level devices.
2976 static int iboe_tos_to_sl(struct net_device *ndev, int tos)
2978 struct iboe_prio_tc_map prio_tc_map = {};
2979 int prio = rt_tos2priority(tos);
2980 struct netdev_nested_priv priv;
2982 /* If VLAN device, get it directly from the VLAN netdev */
2983 if (is_vlan_dev(ndev))
2984 return get_vlan_ndev_tc(ndev, prio);
2986 prio_tc_map.input_prio = prio;
2987 priv.data = (void *)&prio_tc_map;
2989 netdev_walk_all_lower_dev_rcu(ndev,
2990 get_lower_vlan_dev_tc,
2993 /* If map is found from lower device, use it; Otherwise
2994 * continue with the current netdevice to get priority to tc map.
2996 if (prio_tc_map.found)
2997 return prio_tc_map.output_tc;
2998 else if (ndev->num_tc)
2999 return netdev_get_prio_tc_map(ndev, prio);
3004 static __be32 cma_get_roce_udp_flow_label(struct rdma_id_private *id_priv)
3006 struct sockaddr_in6 *addr6;
3010 addr6 = (struct sockaddr_in6 *)cma_src_addr(id_priv);
3011 fl = be32_to_cpu(addr6->sin6_flowinfo) & IB_GRH_FLOWLABEL_MASK;
3012 if ((cma_family(id_priv) != AF_INET6) || !fl) {
3013 dport = be16_to_cpu(cma_port(cma_dst_addr(id_priv)));
3014 sport = be16_to_cpu(cma_port(cma_src_addr(id_priv)));
3015 hash = (u32)sport * 31 + dport;
3016 fl = hash & IB_GRH_FLOWLABEL_MASK;
3019 return cpu_to_be32(fl);
3022 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
3024 struct rdma_route *route = &id_priv->id.route;
3025 struct rdma_addr *addr = &route->addr;
3026 struct cma_work *work;
3028 struct net_device *ndev;
3030 u8 default_roce_tos = id_priv->cma_dev->default_roce_tos[id_priv->id.port_num -
3031 rdma_start_port(id_priv->cma_dev->device)];
3034 mutex_lock(&id_priv->qp_mutex);
3035 tos = id_priv->tos_set ? id_priv->tos : default_roce_tos;
3036 mutex_unlock(&id_priv->qp_mutex);
3038 work = kzalloc(sizeof *work, GFP_KERNEL);
3042 route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
3043 if (!route->path_rec) {
3048 route->num_paths = 1;
3050 ndev = cma_iboe_set_path_rec_l2_fields(id_priv);
3056 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
3057 &route->path_rec->sgid);
3058 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.dst_addr,
3059 &route->path_rec->dgid);
3061 if (((struct sockaddr *)&id_priv->id.route.addr.dst_addr)->sa_family != AF_IB)
3062 /* TODO: get the hoplimit from the inet/inet6 device */
3063 route->path_rec->hop_limit = addr->dev_addr.hoplimit;
3065 route->path_rec->hop_limit = 1;
3066 route->path_rec->reversible = 1;
3067 route->path_rec->pkey = cpu_to_be16(0xffff);
3068 route->path_rec->mtu_selector = IB_SA_EQ;
3069 route->path_rec->sl = iboe_tos_to_sl(ndev, tos);
3070 route->path_rec->traffic_class = tos;
3071 route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
3072 route->path_rec->rate_selector = IB_SA_EQ;
3073 route->path_rec->rate = iboe_get_rate(ndev);
3075 route->path_rec->packet_life_time_selector = IB_SA_EQ;
3076 /* In case ACK timeout is set, use this value to calculate
3077 * PacketLifeTime. As per IBTA 12.7.34,
3078 * local ACK timeout = (2 * PacketLifeTime + Local CA’s ACK delay).
3079 * Assuming a negligible local ACK delay, we can use
3080 * PacketLifeTime = local ACK timeout/2
3081 * as a reasonable approximation for RoCE networks.
3083 mutex_lock(&id_priv->qp_mutex);
3084 if (id_priv->timeout_set && id_priv->timeout)
3085 route->path_rec->packet_life_time = id_priv->timeout - 1;
3087 route->path_rec->packet_life_time = CMA_IBOE_PACKET_LIFETIME;
3088 mutex_unlock(&id_priv->qp_mutex);
3090 if (!route->path_rec->mtu) {
3095 if (rdma_protocol_roce_udp_encap(id_priv->id.device,
3096 id_priv->id.port_num))
3097 route->path_rec->flow_label =
3098 cma_get_roce_udp_flow_label(id_priv);
3100 cma_init_resolve_route_work(work, id_priv);
3101 queue_work(cma_wq, &work->work);
3106 kfree(route->path_rec);
3107 route->path_rec = NULL;
3108 route->num_paths = 0;
3114 int rdma_resolve_route(struct rdma_cm_id *id, unsigned long timeout_ms)
3116 struct rdma_id_private *id_priv;
3119 id_priv = container_of(id, struct rdma_id_private, id);
3120 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY))
3123 cma_id_get(id_priv);
3124 if (rdma_cap_ib_sa(id->device, id->port_num))
3125 ret = cma_resolve_ib_route(id_priv, timeout_ms);
3126 else if (rdma_protocol_roce(id->device, id->port_num))
3127 ret = cma_resolve_iboe_route(id_priv);
3128 else if (rdma_protocol_iwarp(id->device, id->port_num))
3129 ret = cma_resolve_iw_route(id_priv);
3138 cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED);
3139 cma_id_put(id_priv);
3142 EXPORT_SYMBOL(rdma_resolve_route);
3144 static void cma_set_loopback(struct sockaddr *addr)
3146 switch (addr->sa_family) {
3148 ((struct sockaddr_in *) addr)->sin_addr.s_addr = htonl(INADDR_LOOPBACK);
3151 ipv6_addr_set(&((struct sockaddr_in6 *) addr)->sin6_addr,
3155 ib_addr_set(&((struct sockaddr_ib *) addr)->sib_addr,
3161 static int cma_bind_loopback(struct rdma_id_private *id_priv)
3163 struct cma_device *cma_dev, *cur_dev;
3165 enum ib_port_state port_state;
3172 list_for_each_entry(cur_dev, &dev_list, list) {
3173 if (cma_family(id_priv) == AF_IB &&
3174 !rdma_cap_ib_cm(cur_dev->device, 1))
3180 rdma_for_each_port (cur_dev->device, p) {
3181 if (!ib_get_cached_port_state(cur_dev->device, p, &port_state) &&
3182 port_state == IB_PORT_ACTIVE) {
3197 ret = rdma_query_gid(cma_dev->device, p, 0, &gid);
3201 ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
3205 id_priv->id.route.addr.dev_addr.dev_type =
3206 (rdma_protocol_ib(cma_dev->device, p)) ?
3207 ARPHRD_INFINIBAND : ARPHRD_ETHER;
3209 rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
3210 ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
3211 id_priv->id.port_num = p;
3212 cma_attach_to_dev(id_priv, cma_dev);
3213 rdma_restrack_add(&id_priv->res);
3214 cma_set_loopback(cma_src_addr(id_priv));
3216 mutex_unlock(&lock);
3220 static void addr_handler(int status, struct sockaddr *src_addr,
3221 struct rdma_dev_addr *dev_addr, void *context)
3223 struct rdma_id_private *id_priv = context;
3224 struct rdma_cm_event event = {};
3225 struct sockaddr *addr;
3226 struct sockaddr_storage old_addr;
3228 mutex_lock(&id_priv->handler_mutex);
3229 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY,
3230 RDMA_CM_ADDR_RESOLVED))
3234 * Store the previous src address, so that if we fail to acquire
3235 * matching rdma device, old address can be restored back, which helps
3236 * to cancel the cma listen operation correctly.
3238 addr = cma_src_addr(id_priv);
3239 memcpy(&old_addr, addr, rdma_addr_size(addr));
3240 memcpy(addr, src_addr, rdma_addr_size(src_addr));
3241 if (!status && !id_priv->cma_dev) {
3242 status = cma_acquire_dev_by_src_ip(id_priv);
3244 pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to acquire device. status %d\n",
3246 rdma_restrack_add(&id_priv->res);
3247 } else if (status) {
3248 pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to resolve IP. status %d\n", status);
3252 memcpy(addr, &old_addr,
3253 rdma_addr_size((struct sockaddr *)&old_addr));
3254 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
3255 RDMA_CM_ADDR_BOUND))
3257 event.event = RDMA_CM_EVENT_ADDR_ERROR;
3258 event.status = status;
3260 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
3262 if (cma_cm_event_handler(id_priv, &event)) {
3263 destroy_id_handler_unlock(id_priv);
3267 mutex_unlock(&id_priv->handler_mutex);
3270 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
3272 struct cma_work *work;
3276 work = kzalloc(sizeof *work, GFP_KERNEL);
3280 if (!id_priv->cma_dev) {
3281 ret = cma_bind_loopback(id_priv);
3286 rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
3287 rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
3289 enqueue_resolve_addr_work(work, id_priv);
3296 static int cma_resolve_ib_addr(struct rdma_id_private *id_priv)
3298 struct cma_work *work;
3301 work = kzalloc(sizeof *work, GFP_KERNEL);
3305 if (!id_priv->cma_dev) {
3306 ret = cma_resolve_ib_dev(id_priv);
3311 rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, (union ib_gid *)
3312 &(((struct sockaddr_ib *) &id_priv->id.route.addr.dst_addr)->sib_addr));
3314 enqueue_resolve_addr_work(work, id_priv);
3321 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
3322 const struct sockaddr *dst_addr)
3324 struct sockaddr_storage zero_sock = {};
3326 if (src_addr && src_addr->sa_family)
3327 return rdma_bind_addr(id, src_addr);
3330 * When the src_addr is not specified, automatically supply an any addr
3332 zero_sock.ss_family = dst_addr->sa_family;
3333 if (IS_ENABLED(CONFIG_IPV6) && dst_addr->sa_family == AF_INET6) {
3334 struct sockaddr_in6 *src_addr6 =
3335 (struct sockaddr_in6 *)&zero_sock;
3336 struct sockaddr_in6 *dst_addr6 =
3337 (struct sockaddr_in6 *)dst_addr;
3339 src_addr6->sin6_scope_id = dst_addr6->sin6_scope_id;
3340 if (ipv6_addr_type(&dst_addr6->sin6_addr) & IPV6_ADDR_LINKLOCAL)
3341 id->route.addr.dev_addr.bound_dev_if =
3342 dst_addr6->sin6_scope_id;
3343 } else if (dst_addr->sa_family == AF_IB) {
3344 ((struct sockaddr_ib *)&zero_sock)->sib_pkey =
3345 ((struct sockaddr_ib *)dst_addr)->sib_pkey;
3347 return rdma_bind_addr(id, (struct sockaddr *)&zero_sock);
3351 * If required, resolve the source address for bind and leave the id_priv in
3352 * state RDMA_CM_ADDR_BOUND. This oddly uses the state to determine the prior
3353 * calls made by ULP, a previously bound ID will not be re-bound and src_addr is
3356 static int resolve_prepare_src(struct rdma_id_private *id_priv,
3357 struct sockaddr *src_addr,
3358 const struct sockaddr *dst_addr)
3362 memcpy(cma_dst_addr(id_priv), dst_addr, rdma_addr_size(dst_addr));
3363 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY)) {
3364 /* For a well behaved ULP state will be RDMA_CM_IDLE */
3365 ret = cma_bind_addr(&id_priv->id, src_addr, dst_addr);
3368 if (WARN_ON(!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND,
3369 RDMA_CM_ADDR_QUERY))) {
3375 if (cma_family(id_priv) != dst_addr->sa_family) {
3382 cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
3384 memset(cma_dst_addr(id_priv), 0, rdma_addr_size(dst_addr));
3388 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
3389 const struct sockaddr *dst_addr, unsigned long timeout_ms)
3391 struct rdma_id_private *id_priv =
3392 container_of(id, struct rdma_id_private, id);
3395 ret = resolve_prepare_src(id_priv, src_addr, dst_addr);
3399 if (cma_any_addr(dst_addr)) {
3400 ret = cma_resolve_loopback(id_priv);
3402 if (dst_addr->sa_family == AF_IB) {
3403 ret = cma_resolve_ib_addr(id_priv);
3405 ret = rdma_resolve_ip(cma_src_addr(id_priv), dst_addr,
3406 &id->route.addr.dev_addr,
3407 timeout_ms, addr_handler,
3416 cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
3419 EXPORT_SYMBOL(rdma_resolve_addr);
3421 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse)
3423 struct rdma_id_private *id_priv;
3424 unsigned long flags;
3427 id_priv = container_of(id, struct rdma_id_private, id);
3428 spin_lock_irqsave(&id_priv->lock, flags);
3429 if ((reuse && id_priv->state != RDMA_CM_LISTEN) ||
3430 id_priv->state == RDMA_CM_IDLE) {
3431 id_priv->reuseaddr = reuse;
3436 spin_unlock_irqrestore(&id_priv->lock, flags);
3439 EXPORT_SYMBOL(rdma_set_reuseaddr);
3441 int rdma_set_afonly(struct rdma_cm_id *id, int afonly)
3443 struct rdma_id_private *id_priv;
3444 unsigned long flags;
3447 id_priv = container_of(id, struct rdma_id_private, id);
3448 spin_lock_irqsave(&id_priv->lock, flags);
3449 if (id_priv->state == RDMA_CM_IDLE || id_priv->state == RDMA_CM_ADDR_BOUND) {
3450 id_priv->options |= (1 << CMA_OPTION_AFONLY);
3451 id_priv->afonly = afonly;
3456 spin_unlock_irqrestore(&id_priv->lock, flags);
3459 EXPORT_SYMBOL(rdma_set_afonly);
3461 static void cma_bind_port(struct rdma_bind_list *bind_list,
3462 struct rdma_id_private *id_priv)
3464 struct sockaddr *addr;
3465 struct sockaddr_ib *sib;
3469 lockdep_assert_held(&lock);
3471 addr = cma_src_addr(id_priv);
3472 port = htons(bind_list->port);
3474 switch (addr->sa_family) {
3476 ((struct sockaddr_in *) addr)->sin_port = port;
3479 ((struct sockaddr_in6 *) addr)->sin6_port = port;
3482 sib = (struct sockaddr_ib *) addr;
3483 sid = be64_to_cpu(sib->sib_sid);
3484 mask = be64_to_cpu(sib->sib_sid_mask);
3485 sib->sib_sid = cpu_to_be64((sid & mask) | (u64) ntohs(port));
3486 sib->sib_sid_mask = cpu_to_be64(~0ULL);
3489 id_priv->bind_list = bind_list;
3490 hlist_add_head(&id_priv->node, &bind_list->owners);
3493 static int cma_alloc_port(enum rdma_ucm_port_space ps,
3494 struct rdma_id_private *id_priv, unsigned short snum)
3496 struct rdma_bind_list *bind_list;
3499 lockdep_assert_held(&lock);
3501 bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
3505 ret = cma_ps_alloc(id_priv->id.route.addr.dev_addr.net, ps, bind_list,
3511 bind_list->port = snum;
3512 cma_bind_port(bind_list, id_priv);
3516 return ret == -ENOSPC ? -EADDRNOTAVAIL : ret;
3519 static int cma_port_is_unique(struct rdma_bind_list *bind_list,
3520 struct rdma_id_private *id_priv)
3522 struct rdma_id_private *cur_id;
3523 struct sockaddr *daddr = cma_dst_addr(id_priv);
3524 struct sockaddr *saddr = cma_src_addr(id_priv);
3525 __be16 dport = cma_port(daddr);
3527 lockdep_assert_held(&lock);
3529 hlist_for_each_entry(cur_id, &bind_list->owners, node) {
3530 struct sockaddr *cur_daddr = cma_dst_addr(cur_id);
3531 struct sockaddr *cur_saddr = cma_src_addr(cur_id);
3532 __be16 cur_dport = cma_port(cur_daddr);
3534 if (id_priv == cur_id)
3537 /* different dest port -> unique */
3538 if (!cma_any_port(daddr) &&
3539 !cma_any_port(cur_daddr) &&
3540 (dport != cur_dport))
3543 /* different src address -> unique */
3544 if (!cma_any_addr(saddr) &&
3545 !cma_any_addr(cur_saddr) &&
3546 cma_addr_cmp(saddr, cur_saddr))
3549 /* different dst address -> unique */
3550 if (!cma_any_addr(daddr) &&
3551 !cma_any_addr(cur_daddr) &&
3552 cma_addr_cmp(daddr, cur_daddr))
3555 return -EADDRNOTAVAIL;
3560 static int cma_alloc_any_port(enum rdma_ucm_port_space ps,
3561 struct rdma_id_private *id_priv)
3563 static unsigned int last_used_port;
3564 int low, high, remaining;
3566 struct net *net = id_priv->id.route.addr.dev_addr.net;
3568 lockdep_assert_held(&lock);
3570 inet_get_local_port_range(net, &low, &high);
3571 remaining = (high - low) + 1;
3572 rover = prandom_u32() % remaining + low;
3574 if (last_used_port != rover) {
3575 struct rdma_bind_list *bind_list;
3578 bind_list = cma_ps_find(net, ps, (unsigned short)rover);
3581 ret = cma_alloc_port(ps, id_priv, rover);
3583 ret = cma_port_is_unique(bind_list, id_priv);
3585 cma_bind_port(bind_list, id_priv);
3588 * Remember previously used port number in order to avoid
3589 * re-using same port immediately after it is closed.
3592 last_used_port = rover;
3593 if (ret != -EADDRNOTAVAIL)
3598 if ((rover < low) || (rover > high))
3602 return -EADDRNOTAVAIL;
3606 * Check that the requested port is available. This is called when trying to
3607 * bind to a specific port, or when trying to listen on a bound port. In
3608 * the latter case, the provided id_priv may already be on the bind_list, but
3609 * we still need to check that it's okay to start listening.
3611 static int cma_check_port(struct rdma_bind_list *bind_list,
3612 struct rdma_id_private *id_priv, uint8_t reuseaddr)
3614 struct rdma_id_private *cur_id;
3615 struct sockaddr *addr, *cur_addr;
3617 lockdep_assert_held(&lock);
3619 addr = cma_src_addr(id_priv);
3620 hlist_for_each_entry(cur_id, &bind_list->owners, node) {
3621 if (id_priv == cur_id)
3624 if (reuseaddr && cur_id->reuseaddr)
3627 cur_addr = cma_src_addr(cur_id);
3628 if (id_priv->afonly && cur_id->afonly &&
3629 (addr->sa_family != cur_addr->sa_family))
3632 if (cma_any_addr(addr) || cma_any_addr(cur_addr))
3633 return -EADDRNOTAVAIL;
3635 if (!cma_addr_cmp(addr, cur_addr))
3641 static int cma_use_port(enum rdma_ucm_port_space ps,
3642 struct rdma_id_private *id_priv)
3644 struct rdma_bind_list *bind_list;
3645 unsigned short snum;
3648 lockdep_assert_held(&lock);
3650 snum = ntohs(cma_port(cma_src_addr(id_priv)));
3651 if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
3654 bind_list = cma_ps_find(id_priv->id.route.addr.dev_addr.net, ps, snum);
3656 ret = cma_alloc_port(ps, id_priv, snum);
3658 ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr);
3660 cma_bind_port(bind_list, id_priv);
3665 static enum rdma_ucm_port_space
3666 cma_select_inet_ps(struct rdma_id_private *id_priv)
3668 switch (id_priv->id.ps) {
3673 return id_priv->id.ps;
3680 static enum rdma_ucm_port_space
3681 cma_select_ib_ps(struct rdma_id_private *id_priv)
3683 enum rdma_ucm_port_space ps = 0;
3684 struct sockaddr_ib *sib;
3685 u64 sid_ps, mask, sid;
3687 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
3688 mask = be64_to_cpu(sib->sib_sid_mask) & RDMA_IB_IP_PS_MASK;
3689 sid = be64_to_cpu(sib->sib_sid) & mask;
3691 if ((id_priv->id.ps == RDMA_PS_IB) && (sid == (RDMA_IB_IP_PS_IB & mask))) {
3692 sid_ps = RDMA_IB_IP_PS_IB;
3694 } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_TCP)) &&
3695 (sid == (RDMA_IB_IP_PS_TCP & mask))) {
3696 sid_ps = RDMA_IB_IP_PS_TCP;
3698 } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_UDP)) &&
3699 (sid == (RDMA_IB_IP_PS_UDP & mask))) {
3700 sid_ps = RDMA_IB_IP_PS_UDP;
3705 sib->sib_sid = cpu_to_be64(sid_ps | ntohs(cma_port((struct sockaddr *) sib)));
3706 sib->sib_sid_mask = cpu_to_be64(RDMA_IB_IP_PS_MASK |
3707 be64_to_cpu(sib->sib_sid_mask));
3712 static int cma_get_port(struct rdma_id_private *id_priv)
3714 enum rdma_ucm_port_space ps;
3717 if (cma_family(id_priv) != AF_IB)
3718 ps = cma_select_inet_ps(id_priv);
3720 ps = cma_select_ib_ps(id_priv);
3722 return -EPROTONOSUPPORT;
3725 if (cma_any_port(cma_src_addr(id_priv)))
3726 ret = cma_alloc_any_port(ps, id_priv);
3728 ret = cma_use_port(ps, id_priv);
3729 mutex_unlock(&lock);
3734 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
3735 struct sockaddr *addr)
3737 #if IS_ENABLED(CONFIG_IPV6)
3738 struct sockaddr_in6 *sin6;
3740 if (addr->sa_family != AF_INET6)
3743 sin6 = (struct sockaddr_in6 *) addr;
3745 if (!(ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL))
3748 if (!sin6->sin6_scope_id)
3751 dev_addr->bound_dev_if = sin6->sin6_scope_id;
3756 int rdma_listen(struct rdma_cm_id *id, int backlog)
3758 struct rdma_id_private *id_priv =
3759 container_of(id, struct rdma_id_private, id);
3762 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN)) {
3763 struct sockaddr_in any_in = {
3764 .sin_family = AF_INET,
3765 .sin_addr.s_addr = htonl(INADDR_ANY),
3768 /* For a well behaved ULP state will be RDMA_CM_IDLE */
3769 ret = rdma_bind_addr(id, (struct sockaddr *)&any_in);
3772 if (WARN_ON(!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND,
3778 * Once the ID reaches RDMA_CM_LISTEN it is not allowed to be reusable
3779 * any more, and has to be unique in the bind list.
3781 if (id_priv->reuseaddr) {
3783 ret = cma_check_port(id_priv->bind_list, id_priv, 0);
3785 id_priv->reuseaddr = 0;
3786 mutex_unlock(&lock);
3791 id_priv->backlog = backlog;
3792 if (id_priv->cma_dev) {
3793 if (rdma_cap_ib_cm(id->device, 1)) {
3794 ret = cma_ib_listen(id_priv);
3797 } else if (rdma_cap_iw_cm(id->device, 1)) {
3798 ret = cma_iw_listen(id_priv, backlog);
3806 ret = cma_listen_on_all(id_priv);
3813 id_priv->backlog = 0;
3815 * All the failure paths that lead here will not allow the req_handler's
3818 cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND);
3821 EXPORT_SYMBOL(rdma_listen);
3823 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
3825 struct rdma_id_private *id_priv;
3827 struct sockaddr *daddr;
3829 if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6 &&
3830 addr->sa_family != AF_IB)
3831 return -EAFNOSUPPORT;
3833 id_priv = container_of(id, struct rdma_id_private, id);
3834 if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND))
3837 ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
3841 memcpy(cma_src_addr(id_priv), addr, rdma_addr_size(addr));
3842 if (!cma_any_addr(addr)) {
3843 ret = cma_translate_addr(addr, &id->route.addr.dev_addr);
3847 ret = cma_acquire_dev_by_src_ip(id_priv);
3852 if (!(id_priv->options & (1 << CMA_OPTION_AFONLY))) {
3853 if (addr->sa_family == AF_INET)
3854 id_priv->afonly = 1;
3855 #if IS_ENABLED(CONFIG_IPV6)
3856 else if (addr->sa_family == AF_INET6) {
3857 struct net *net = id_priv->id.route.addr.dev_addr.net;
3859 id_priv->afonly = net->ipv6.sysctl.bindv6only;
3863 daddr = cma_dst_addr(id_priv);
3864 daddr->sa_family = addr->sa_family;
3866 ret = cma_get_port(id_priv);
3870 if (!cma_any_addr(addr))
3871 rdma_restrack_add(&id_priv->res);
3874 if (id_priv->cma_dev)
3875 cma_release_dev(id_priv);
3877 cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE);
3880 EXPORT_SYMBOL(rdma_bind_addr);
3882 static int cma_format_hdr(void *hdr, struct rdma_id_private *id_priv)
3884 struct cma_hdr *cma_hdr;
3887 cma_hdr->cma_version = CMA_VERSION;
3888 if (cma_family(id_priv) == AF_INET) {
3889 struct sockaddr_in *src4, *dst4;
3891 src4 = (struct sockaddr_in *) cma_src_addr(id_priv);
3892 dst4 = (struct sockaddr_in *) cma_dst_addr(id_priv);
3894 cma_set_ip_ver(cma_hdr, 4);
3895 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
3896 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
3897 cma_hdr->port = src4->sin_port;
3898 } else if (cma_family(id_priv) == AF_INET6) {
3899 struct sockaddr_in6 *src6, *dst6;
3901 src6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
3902 dst6 = (struct sockaddr_in6 *) cma_dst_addr(id_priv);
3904 cma_set_ip_ver(cma_hdr, 6);
3905 cma_hdr->src_addr.ip6 = src6->sin6_addr;
3906 cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
3907 cma_hdr->port = src6->sin6_port;
3912 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
3913 const struct ib_cm_event *ib_event)
3915 struct rdma_id_private *id_priv = cm_id->context;
3916 struct rdma_cm_event event = {};
3917 const struct ib_cm_sidr_rep_event_param *rep =
3918 &ib_event->param.sidr_rep_rcvd;
3921 mutex_lock(&id_priv->handler_mutex);
3922 if (READ_ONCE(id_priv->state) != RDMA_CM_CONNECT)
3925 switch (ib_event->event) {
3926 case IB_CM_SIDR_REQ_ERROR:
3927 event.event = RDMA_CM_EVENT_UNREACHABLE;
3928 event.status = -ETIMEDOUT;
3930 case IB_CM_SIDR_REP_RECEIVED:
3931 event.param.ud.private_data = ib_event->private_data;
3932 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
3933 if (rep->status != IB_SIDR_SUCCESS) {
3934 event.event = RDMA_CM_EVENT_UNREACHABLE;
3935 event.status = ib_event->param.sidr_rep_rcvd.status;
3936 pr_debug_ratelimited("RDMA CM: UNREACHABLE: bad SIDR reply. status %d\n",
3940 ret = cma_set_qkey(id_priv, rep->qkey);
3942 pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to set qkey. status %d\n", ret);
3943 event.event = RDMA_CM_EVENT_ADDR_ERROR;
3947 ib_init_ah_attr_from_path(id_priv->id.device,
3948 id_priv->id.port_num,
3949 id_priv->id.route.path_rec,
3950 &event.param.ud.ah_attr,
3952 event.param.ud.qp_num = rep->qpn;
3953 event.param.ud.qkey = rep->qkey;
3954 event.event = RDMA_CM_EVENT_ESTABLISHED;
3958 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
3963 ret = cma_cm_event_handler(id_priv, &event);
3965 rdma_destroy_ah_attr(&event.param.ud.ah_attr);
3967 /* Destroy the CM ID by returning a non-zero value. */
3968 id_priv->cm_id.ib = NULL;
3969 destroy_id_handler_unlock(id_priv);
3973 mutex_unlock(&id_priv->handler_mutex);
3977 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
3978 struct rdma_conn_param *conn_param)
3980 struct ib_cm_sidr_req_param req;
3981 struct ib_cm_id *id;
3986 memset(&req, 0, sizeof req);
3987 offset = cma_user_data_offset(id_priv);
3988 req.private_data_len = offset + conn_param->private_data_len;
3989 if (req.private_data_len < conn_param->private_data_len)
3992 if (req.private_data_len) {
3993 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
3997 private_data = NULL;
4000 if (conn_param->private_data && conn_param->private_data_len)
4001 memcpy(private_data + offset, conn_param->private_data,
4002 conn_param->private_data_len);
4005 ret = cma_format_hdr(private_data, id_priv);
4008 req.private_data = private_data;
4011 id = ib_create_cm_id(id_priv->id.device, cma_sidr_rep_handler,
4017 id_priv->cm_id.ib = id;
4019 req.path = id_priv->id.route.path_rec;
4020 req.sgid_attr = id_priv->id.route.addr.dev_addr.sgid_attr;
4021 req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
4022 req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
4023 req.max_cm_retries = CMA_MAX_CM_RETRIES;
4025 trace_cm_send_sidr_req(id_priv);
4026 ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
4028 ib_destroy_cm_id(id_priv->cm_id.ib);
4029 id_priv->cm_id.ib = NULL;
4032 kfree(private_data);
4036 static int cma_connect_ib(struct rdma_id_private *id_priv,
4037 struct rdma_conn_param *conn_param)
4039 struct ib_cm_req_param req;
4040 struct rdma_route *route;
4042 struct ib_cm_id *id;
4046 memset(&req, 0, sizeof req);
4047 offset = cma_user_data_offset(id_priv);
4048 req.private_data_len = offset + conn_param->private_data_len;
4049 if (req.private_data_len < conn_param->private_data_len)
4052 if (req.private_data_len) {
4053 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
4057 private_data = NULL;
4060 if (conn_param->private_data && conn_param->private_data_len)
4061 memcpy(private_data + offset, conn_param->private_data,
4062 conn_param->private_data_len);
4064 id = ib_create_cm_id(id_priv->id.device, cma_ib_handler, id_priv);
4069 id_priv->cm_id.ib = id;
4071 route = &id_priv->id.route;
4073 ret = cma_format_hdr(private_data, id_priv);
4076 req.private_data = private_data;
4079 req.primary_path = &route->path_rec[0];
4080 if (route->num_paths == 2)
4081 req.alternate_path = &route->path_rec[1];
4083 req.ppath_sgid_attr = id_priv->id.route.addr.dev_addr.sgid_attr;
4084 /* Alternate path SGID attribute currently unsupported */
4085 req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
4086 req.qp_num = id_priv->qp_num;
4087 req.qp_type = id_priv->id.qp_type;
4088 req.starting_psn = id_priv->seq_num;
4089 req.responder_resources = conn_param->responder_resources;
4090 req.initiator_depth = conn_param->initiator_depth;
4091 req.flow_control = conn_param->flow_control;
4092 req.retry_count = min_t(u8, 7, conn_param->retry_count);
4093 req.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
4094 req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
4095 req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
4096 req.max_cm_retries = CMA_MAX_CM_RETRIES;
4097 req.srq = id_priv->srq ? 1 : 0;
4098 req.ece.vendor_id = id_priv->ece.vendor_id;
4099 req.ece.attr_mod = id_priv->ece.attr_mod;
4101 trace_cm_send_req(id_priv);
4102 ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
4104 if (ret && !IS_ERR(id)) {
4105 ib_destroy_cm_id(id);
4106 id_priv->cm_id.ib = NULL;
4109 kfree(private_data);
4113 static int cma_connect_iw(struct rdma_id_private *id_priv,
4114 struct rdma_conn_param *conn_param)
4116 struct iw_cm_id *cm_id;
4118 struct iw_cm_conn_param iw_param;
4120 cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
4122 return PTR_ERR(cm_id);
4124 mutex_lock(&id_priv->qp_mutex);
4125 cm_id->tos = id_priv->tos;
4126 cm_id->tos_set = id_priv->tos_set;
4127 mutex_unlock(&id_priv->qp_mutex);
4129 id_priv->cm_id.iw = cm_id;
4131 memcpy(&cm_id->local_addr, cma_src_addr(id_priv),
4132 rdma_addr_size(cma_src_addr(id_priv)));
4133 memcpy(&cm_id->remote_addr, cma_dst_addr(id_priv),
4134 rdma_addr_size(cma_dst_addr(id_priv)));
4136 ret = cma_modify_qp_rtr(id_priv, conn_param);
4141 iw_param.ord = conn_param->initiator_depth;
4142 iw_param.ird = conn_param->responder_resources;
4143 iw_param.private_data = conn_param->private_data;
4144 iw_param.private_data_len = conn_param->private_data_len;
4145 iw_param.qpn = id_priv->id.qp ? id_priv->qp_num : conn_param->qp_num;
4147 memset(&iw_param, 0, sizeof iw_param);
4148 iw_param.qpn = id_priv->qp_num;
4150 ret = iw_cm_connect(cm_id, &iw_param);
4153 iw_destroy_cm_id(cm_id);
4154 id_priv->cm_id.iw = NULL;
4160 * rdma_connect_locked - Initiate an active connection request.
4161 * @id: Connection identifier to connect.
4162 * @conn_param: Connection information used for connected QPs.
4164 * Same as rdma_connect() but can only be called from the
4165 * RDMA_CM_EVENT_ROUTE_RESOLVED handler callback.
4167 int rdma_connect_locked(struct rdma_cm_id *id,
4168 struct rdma_conn_param *conn_param)
4170 struct rdma_id_private *id_priv =
4171 container_of(id, struct rdma_id_private, id);
4174 if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT))
4178 id_priv->qp_num = conn_param->qp_num;
4179 id_priv->srq = conn_param->srq;
4182 if (rdma_cap_ib_cm(id->device, id->port_num)) {
4183 if (id->qp_type == IB_QPT_UD)
4184 ret = cma_resolve_ib_udp(id_priv, conn_param);
4186 ret = cma_connect_ib(id_priv, conn_param);
4187 } else if (rdma_cap_iw_cm(id->device, id->port_num))
4188 ret = cma_connect_iw(id_priv, conn_param);
4195 cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED);
4198 EXPORT_SYMBOL(rdma_connect_locked);
4201 * rdma_connect - Initiate an active connection request.
4202 * @id: Connection identifier to connect.
4203 * @conn_param: Connection information used for connected QPs.
4205 * Users must have resolved a route for the rdma_cm_id to connect with by having
4206 * called rdma_resolve_route before calling this routine.
4208 * This call will either connect to a remote QP or obtain remote QP information
4209 * for unconnected rdma_cm_id's. The actual operation is based on the
4210 * rdma_cm_id's port space.
4212 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
4214 struct rdma_id_private *id_priv =
4215 container_of(id, struct rdma_id_private, id);
4218 mutex_lock(&id_priv->handler_mutex);
4219 ret = rdma_connect_locked(id, conn_param);
4220 mutex_unlock(&id_priv->handler_mutex);
4223 EXPORT_SYMBOL(rdma_connect);
4226 * rdma_connect_ece - Initiate an active connection request with ECE data.
4227 * @id: Connection identifier to connect.
4228 * @conn_param: Connection information used for connected QPs.
4229 * @ece: ECE parameters
4231 * See rdma_connect() explanation.
4233 int rdma_connect_ece(struct rdma_cm_id *id, struct rdma_conn_param *conn_param,
4234 struct rdma_ucm_ece *ece)
4236 struct rdma_id_private *id_priv =
4237 container_of(id, struct rdma_id_private, id);
4239 id_priv->ece.vendor_id = ece->vendor_id;
4240 id_priv->ece.attr_mod = ece->attr_mod;
4242 return rdma_connect(id, conn_param);
4244 EXPORT_SYMBOL(rdma_connect_ece);
4246 static int cma_accept_ib(struct rdma_id_private *id_priv,
4247 struct rdma_conn_param *conn_param)
4249 struct ib_cm_rep_param rep;
4252 ret = cma_modify_qp_rtr(id_priv, conn_param);
4256 ret = cma_modify_qp_rts(id_priv, conn_param);
4260 memset(&rep, 0, sizeof rep);
4261 rep.qp_num = id_priv->qp_num;
4262 rep.starting_psn = id_priv->seq_num;
4263 rep.private_data = conn_param->private_data;
4264 rep.private_data_len = conn_param->private_data_len;
4265 rep.responder_resources = conn_param->responder_resources;
4266 rep.initiator_depth = conn_param->initiator_depth;
4267 rep.failover_accepted = 0;
4268 rep.flow_control = conn_param->flow_control;
4269 rep.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
4270 rep.srq = id_priv->srq ? 1 : 0;
4271 rep.ece.vendor_id = id_priv->ece.vendor_id;
4272 rep.ece.attr_mod = id_priv->ece.attr_mod;
4274 trace_cm_send_rep(id_priv);
4275 ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
4280 static int cma_accept_iw(struct rdma_id_private *id_priv,
4281 struct rdma_conn_param *conn_param)
4283 struct iw_cm_conn_param iw_param;
4289 ret = cma_modify_qp_rtr(id_priv, conn_param);
4293 iw_param.ord = conn_param->initiator_depth;
4294 iw_param.ird = conn_param->responder_resources;
4295 iw_param.private_data = conn_param->private_data;
4296 iw_param.private_data_len = conn_param->private_data_len;
4297 if (id_priv->id.qp) {
4298 iw_param.qpn = id_priv->qp_num;
4300 iw_param.qpn = conn_param->qp_num;
4302 return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
4305 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
4306 enum ib_cm_sidr_status status, u32 qkey,
4307 const void *private_data, int private_data_len)
4309 struct ib_cm_sidr_rep_param rep;
4312 memset(&rep, 0, sizeof rep);
4313 rep.status = status;
4314 if (status == IB_SIDR_SUCCESS) {
4315 ret = cma_set_qkey(id_priv, qkey);
4318 rep.qp_num = id_priv->qp_num;
4319 rep.qkey = id_priv->qkey;
4321 rep.ece.vendor_id = id_priv->ece.vendor_id;
4322 rep.ece.attr_mod = id_priv->ece.attr_mod;
4325 rep.private_data = private_data;
4326 rep.private_data_len = private_data_len;
4328 trace_cm_send_sidr_rep(id_priv);
4329 return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
4333 * rdma_accept - Called to accept a connection request or response.
4334 * @id: Connection identifier associated with the request.
4335 * @conn_param: Information needed to establish the connection. This must be
4336 * provided if accepting a connection request. If accepting a connection
4337 * response, this parameter must be NULL.
4339 * Typically, this routine is only called by the listener to accept a connection
4340 * request. It must also be called on the active side of a connection if the
4341 * user is performing their own QP transitions.
4343 * In the case of error, a reject message is sent to the remote side and the
4344 * state of the qp associated with the id is modified to error, such that any
4345 * previously posted receive buffers would be flushed.
4347 * This function is for use by kernel ULPs and must be called from under the
4350 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
4352 struct rdma_id_private *id_priv =
4353 container_of(id, struct rdma_id_private, id);
4356 lockdep_assert_held(&id_priv->handler_mutex);
4358 if (READ_ONCE(id_priv->state) != RDMA_CM_CONNECT)
4361 if (!id->qp && conn_param) {
4362 id_priv->qp_num = conn_param->qp_num;
4363 id_priv->srq = conn_param->srq;
4366 if (rdma_cap_ib_cm(id->device, id->port_num)) {
4367 if (id->qp_type == IB_QPT_UD) {
4369 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
4371 conn_param->private_data,
4372 conn_param->private_data_len);
4374 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
4378 ret = cma_accept_ib(id_priv, conn_param);
4380 ret = cma_rep_recv(id_priv);
4382 } else if (rdma_cap_iw_cm(id->device, id->port_num))
4383 ret = cma_accept_iw(id_priv, conn_param);
4392 cma_modify_qp_err(id_priv);
4393 rdma_reject(id, NULL, 0, IB_CM_REJ_CONSUMER_DEFINED);
4396 EXPORT_SYMBOL(rdma_accept);
4398 int rdma_accept_ece(struct rdma_cm_id *id, struct rdma_conn_param *conn_param,
4399 struct rdma_ucm_ece *ece)
4401 struct rdma_id_private *id_priv =
4402 container_of(id, struct rdma_id_private, id);
4404 id_priv->ece.vendor_id = ece->vendor_id;
4405 id_priv->ece.attr_mod = ece->attr_mod;
4407 return rdma_accept(id, conn_param);
4409 EXPORT_SYMBOL(rdma_accept_ece);
4411 void rdma_lock_handler(struct rdma_cm_id *id)
4413 struct rdma_id_private *id_priv =
4414 container_of(id, struct rdma_id_private, id);
4416 mutex_lock(&id_priv->handler_mutex);
4418 EXPORT_SYMBOL(rdma_lock_handler);
4420 void rdma_unlock_handler(struct rdma_cm_id *id)
4422 struct rdma_id_private *id_priv =
4423 container_of(id, struct rdma_id_private, id);
4425 mutex_unlock(&id_priv->handler_mutex);
4427 EXPORT_SYMBOL(rdma_unlock_handler);
4429 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
4431 struct rdma_id_private *id_priv;
4434 id_priv = container_of(id, struct rdma_id_private, id);
4435 if (!id_priv->cm_id.ib)
4438 switch (id->device->node_type) {
4439 case RDMA_NODE_IB_CA:
4440 ret = ib_cm_notify(id_priv->cm_id.ib, event);
4448 EXPORT_SYMBOL(rdma_notify);
4450 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
4451 u8 private_data_len, u8 reason)
4453 struct rdma_id_private *id_priv;
4456 id_priv = container_of(id, struct rdma_id_private, id);
4457 if (!id_priv->cm_id.ib)
4460 if (rdma_cap_ib_cm(id->device, id->port_num)) {
4461 if (id->qp_type == IB_QPT_UD) {
4462 ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT, 0,
4463 private_data, private_data_len);
4465 trace_cm_send_rej(id_priv);
4466 ret = ib_send_cm_rej(id_priv->cm_id.ib, reason, NULL, 0,
4467 private_data, private_data_len);
4469 } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4470 ret = iw_cm_reject(id_priv->cm_id.iw,
4471 private_data, private_data_len);
4477 EXPORT_SYMBOL(rdma_reject);
4479 int rdma_disconnect(struct rdma_cm_id *id)
4481 struct rdma_id_private *id_priv;
4484 id_priv = container_of(id, struct rdma_id_private, id);
4485 if (!id_priv->cm_id.ib)
4488 if (rdma_cap_ib_cm(id->device, id->port_num)) {
4489 ret = cma_modify_qp_err(id_priv);
4492 /* Initiate or respond to a disconnect. */
4493 trace_cm_disconnect(id_priv);
4494 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0)) {
4495 if (!ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0))
4496 trace_cm_sent_drep(id_priv);
4498 trace_cm_sent_dreq(id_priv);
4500 } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4501 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
4508 EXPORT_SYMBOL(rdma_disconnect);
4510 static void cma_make_mc_event(int status, struct rdma_id_private *id_priv,
4511 struct ib_sa_multicast *multicast,
4512 struct rdma_cm_event *event,
4513 struct cma_multicast *mc)
4515 struct rdma_dev_addr *dev_addr;
4516 enum ib_gid_type gid_type;
4517 struct net_device *ndev;
4520 status = cma_set_qkey(id_priv, be32_to_cpu(multicast->rec.qkey));
4522 pr_debug_ratelimited("RDMA CM: MULTICAST_ERROR: failed to join multicast. status %d\n",
4525 event->status = status;
4526 event->param.ud.private_data = mc->context;
4528 event->event = RDMA_CM_EVENT_MULTICAST_ERROR;
4532 dev_addr = &id_priv->id.route.addr.dev_addr;
4533 ndev = dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if);
4536 ->default_gid_type[id_priv->id.port_num -
4538 id_priv->cma_dev->device)];
4540 event->event = RDMA_CM_EVENT_MULTICAST_JOIN;
4541 if (ib_init_ah_from_mcmember(id_priv->id.device, id_priv->id.port_num,
4542 &multicast->rec, ndev, gid_type,
4543 &event->param.ud.ah_attr)) {
4544 event->event = RDMA_CM_EVENT_MULTICAST_ERROR;
4548 event->param.ud.qp_num = 0xFFFFFF;
4549 event->param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
4556 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
4558 struct cma_multicast *mc = multicast->context;
4559 struct rdma_id_private *id_priv = mc->id_priv;
4560 struct rdma_cm_event event = {};
4563 mutex_lock(&id_priv->handler_mutex);
4564 if (READ_ONCE(id_priv->state) == RDMA_CM_DEVICE_REMOVAL ||
4565 READ_ONCE(id_priv->state) == RDMA_CM_DESTROYING)
4568 cma_make_mc_event(status, id_priv, multicast, &event, mc);
4569 ret = cma_cm_event_handler(id_priv, &event);
4570 rdma_destroy_ah_attr(&event.param.ud.ah_attr);
4574 mutex_unlock(&id_priv->handler_mutex);
4578 static void cma_set_mgid(struct rdma_id_private *id_priv,
4579 struct sockaddr *addr, union ib_gid *mgid)
4581 unsigned char mc_map[MAX_ADDR_LEN];
4582 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4583 struct sockaddr_in *sin = (struct sockaddr_in *) addr;
4584 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
4586 if (cma_any_addr(addr)) {
4587 memset(mgid, 0, sizeof *mgid);
4588 } else if ((addr->sa_family == AF_INET6) &&
4589 ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
4591 /* IPv6 address is an SA assigned MGID. */
4592 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
4593 } else if (addr->sa_family == AF_IB) {
4594 memcpy(mgid, &((struct sockaddr_ib *) addr)->sib_addr, sizeof *mgid);
4595 } else if (addr->sa_family == AF_INET6) {
4596 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
4597 if (id_priv->id.ps == RDMA_PS_UDP)
4598 mc_map[7] = 0x01; /* Use RDMA CM signature */
4599 *mgid = *(union ib_gid *) (mc_map + 4);
4601 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
4602 if (id_priv->id.ps == RDMA_PS_UDP)
4603 mc_map[7] = 0x01; /* Use RDMA CM signature */
4604 *mgid = *(union ib_gid *) (mc_map + 4);
4608 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
4609 struct cma_multicast *mc)
4611 struct ib_sa_mcmember_rec rec;
4612 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4613 ib_sa_comp_mask comp_mask;
4616 ib_addr_get_mgid(dev_addr, &rec.mgid);
4617 ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
4622 ret = cma_set_qkey(id_priv, 0);
4626 cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
4627 rec.qkey = cpu_to_be32(id_priv->qkey);
4628 rdma_addr_get_sgid(dev_addr, &rec.port_gid);
4629 rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
4630 rec.join_state = mc->join_state;
4632 if ((rec.join_state == BIT(SENDONLY_FULLMEMBER_JOIN)) &&
4633 (!ib_sa_sendonly_fullmem_support(&sa_client,
4635 id_priv->id.port_num))) {
4637 &id_priv->id.device->dev,
4638 "RDMA CM: port %u Unable to multicast join: SM doesn't support Send Only Full Member option\n",
4639 id_priv->id.port_num);
4643 comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
4644 IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
4645 IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
4646 IB_SA_MCMEMBER_REC_FLOW_LABEL |
4647 IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
4649 if (id_priv->id.ps == RDMA_PS_IPOIB)
4650 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
4651 IB_SA_MCMEMBER_REC_RATE_SELECTOR |
4652 IB_SA_MCMEMBER_REC_MTU_SELECTOR |
4653 IB_SA_MCMEMBER_REC_MTU |
4654 IB_SA_MCMEMBER_REC_HOP_LIMIT;
4656 mc->sa_mc = ib_sa_join_multicast(&sa_client, id_priv->id.device,
4657 id_priv->id.port_num, &rec, comp_mask,
4658 GFP_KERNEL, cma_ib_mc_handler, mc);
4659 return PTR_ERR_OR_ZERO(mc->sa_mc);
4662 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid,
4663 enum ib_gid_type gid_type)
4665 struct sockaddr_in *sin = (struct sockaddr_in *)addr;
4666 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
4668 if (cma_any_addr(addr)) {
4669 memset(mgid, 0, sizeof *mgid);
4670 } else if (addr->sa_family == AF_INET6) {
4671 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
4674 (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ? 0 : 0xff;
4676 (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ? 0 : 0x0e;
4685 mgid->raw[10] = 0xff;
4686 mgid->raw[11] = 0xff;
4687 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
4691 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
4692 struct cma_multicast *mc)
4694 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4696 struct sockaddr *addr = (struct sockaddr *)&mc->addr;
4697 struct net_device *ndev = NULL;
4698 struct ib_sa_multicast ib;
4699 enum ib_gid_type gid_type;
4702 send_only = mc->join_state == BIT(SENDONLY_FULLMEMBER_JOIN);
4704 if (cma_zero_addr(addr))
4707 gid_type = id_priv->cma_dev->default_gid_type[id_priv->id.port_num -
4708 rdma_start_port(id_priv->cma_dev->device)];
4709 cma_iboe_set_mgid(addr, &ib.rec.mgid, gid_type);
4711 ib.rec.pkey = cpu_to_be16(0xffff);
4712 if (id_priv->id.ps == RDMA_PS_UDP)
4713 ib.rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
4715 if (dev_addr->bound_dev_if)
4716 ndev = dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if);
4720 ib.rec.rate = iboe_get_rate(ndev);
4721 ib.rec.hop_limit = 1;
4722 ib.rec.mtu = iboe_get_mtu(ndev->mtu);
4724 if (addr->sa_family == AF_INET) {
4725 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) {
4726 ib.rec.hop_limit = IPV6_DEFAULT_HOPLIMIT;
4728 err = cma_igmp_send(ndev, &ib.rec.mgid,
4733 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP)
4737 if (err || !ib.rec.mtu)
4738 return err ?: -EINVAL;
4740 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
4742 INIT_WORK(&mc->iboe_join.work, cma_iboe_join_work_handler);
4743 cma_make_mc_event(0, id_priv, &ib, &mc->iboe_join.event, mc);
4744 queue_work(cma_wq, &mc->iboe_join.work);
4748 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
4749 u8 join_state, void *context)
4751 struct rdma_id_private *id_priv =
4752 container_of(id, struct rdma_id_private, id);
4753 struct cma_multicast *mc;
4756 /* Not supported for kernel QPs */
4757 if (WARN_ON(id->qp))
4760 /* ULP is calling this wrong. */
4761 if (!id->device || (READ_ONCE(id_priv->state) != RDMA_CM_ADDR_BOUND &&
4762 READ_ONCE(id_priv->state) != RDMA_CM_ADDR_RESOLVED))
4765 mc = kzalloc(sizeof(*mc), GFP_KERNEL);
4769 memcpy(&mc->addr, addr, rdma_addr_size(addr));
4770 mc->context = context;
4771 mc->id_priv = id_priv;
4772 mc->join_state = join_state;
4774 if (rdma_protocol_roce(id->device, id->port_num)) {
4775 ret = cma_iboe_join_multicast(id_priv, mc);
4778 } else if (rdma_cap_ib_mcast(id->device, id->port_num)) {
4779 ret = cma_join_ib_multicast(id_priv, mc);
4787 spin_lock(&id_priv->lock);
4788 list_add(&mc->list, &id_priv->mc_list);
4789 spin_unlock(&id_priv->lock);
4796 EXPORT_SYMBOL(rdma_join_multicast);
4798 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
4800 struct rdma_id_private *id_priv;
4801 struct cma_multicast *mc;
4803 id_priv = container_of(id, struct rdma_id_private, id);
4804 spin_lock_irq(&id_priv->lock);
4805 list_for_each_entry(mc, &id_priv->mc_list, list) {
4806 if (memcmp(&mc->addr, addr, rdma_addr_size(addr)) != 0)
4808 list_del(&mc->list);
4809 spin_unlock_irq(&id_priv->lock);
4811 WARN_ON(id_priv->cma_dev->device != id->device);
4812 destroy_mc(id_priv, mc);
4815 spin_unlock_irq(&id_priv->lock);
4817 EXPORT_SYMBOL(rdma_leave_multicast);
4819 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
4821 struct rdma_dev_addr *dev_addr;
4822 struct cma_work *work;
4824 dev_addr = &id_priv->id.route.addr.dev_addr;
4826 if ((dev_addr->bound_dev_if == ndev->ifindex) &&
4827 (net_eq(dev_net(ndev), dev_addr->net)) &&
4828 memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
4829 pr_info("RDMA CM addr change for ndev %s used by id %p\n",
4830 ndev->name, &id_priv->id);
4831 work = kzalloc(sizeof *work, GFP_KERNEL);
4835 INIT_WORK(&work->work, cma_work_handler);
4837 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
4838 cma_id_get(id_priv);
4839 queue_work(cma_wq, &work->work);
4845 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
4848 struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
4849 struct cma_device *cma_dev;
4850 struct rdma_id_private *id_priv;
4851 int ret = NOTIFY_DONE;
4853 if (event != NETDEV_BONDING_FAILOVER)
4856 if (!netif_is_bond_master(ndev))
4860 list_for_each_entry(cma_dev, &dev_list, list)
4861 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
4862 ret = cma_netdev_change(ndev, id_priv);
4868 mutex_unlock(&lock);
4872 static struct notifier_block cma_nb = {
4873 .notifier_call = cma_netdev_callback
4876 static void cma_send_device_removal_put(struct rdma_id_private *id_priv)
4878 struct rdma_cm_event event = { .event = RDMA_CM_EVENT_DEVICE_REMOVAL };
4879 enum rdma_cm_state state;
4880 unsigned long flags;
4882 mutex_lock(&id_priv->handler_mutex);
4883 /* Record that we want to remove the device */
4884 spin_lock_irqsave(&id_priv->lock, flags);
4885 state = id_priv->state;
4886 if (state == RDMA_CM_DESTROYING || state == RDMA_CM_DEVICE_REMOVAL) {
4887 spin_unlock_irqrestore(&id_priv->lock, flags);
4888 mutex_unlock(&id_priv->handler_mutex);
4889 cma_id_put(id_priv);
4892 id_priv->state = RDMA_CM_DEVICE_REMOVAL;
4893 spin_unlock_irqrestore(&id_priv->lock, flags);
4895 if (cma_cm_event_handler(id_priv, &event)) {
4897 * At this point the ULP promises it won't call
4898 * rdma_destroy_id() concurrently
4900 cma_id_put(id_priv);
4901 mutex_unlock(&id_priv->handler_mutex);
4902 trace_cm_id_destroy(id_priv);
4903 _destroy_id(id_priv, state);
4906 mutex_unlock(&id_priv->handler_mutex);
4909 * If this races with destroy then the thread that first assigns state
4910 * to a destroying does the cancel.
4912 cma_cancel_operation(id_priv, state);
4913 cma_id_put(id_priv);
4916 static void cma_process_remove(struct cma_device *cma_dev)
4919 while (!list_empty(&cma_dev->id_list)) {
4920 struct rdma_id_private *id_priv = list_first_entry(
4921 &cma_dev->id_list, struct rdma_id_private, list);
4923 list_del(&id_priv->listen_list);
4924 list_del_init(&id_priv->list);
4925 cma_id_get(id_priv);
4926 mutex_unlock(&lock);
4928 cma_send_device_removal_put(id_priv);
4932 mutex_unlock(&lock);
4934 cma_dev_put(cma_dev);
4935 wait_for_completion(&cma_dev->comp);
4938 static int cma_add_one(struct ib_device *device)
4940 struct rdma_id_private *to_destroy;
4941 struct cma_device *cma_dev;
4942 struct rdma_id_private *id_priv;
4944 unsigned long supported_gids = 0;
4947 cma_dev = kmalloc(sizeof(*cma_dev), GFP_KERNEL);
4951 cma_dev->device = device;
4952 cma_dev->default_gid_type = kcalloc(device->phys_port_cnt,
4953 sizeof(*cma_dev->default_gid_type),
4955 if (!cma_dev->default_gid_type) {
4960 cma_dev->default_roce_tos = kcalloc(device->phys_port_cnt,
4961 sizeof(*cma_dev->default_roce_tos),
4963 if (!cma_dev->default_roce_tos) {
4968 rdma_for_each_port (device, i) {
4969 supported_gids = roce_gid_type_mask_support(device, i);
4970 WARN_ON(!supported_gids);
4971 if (supported_gids & (1 << CMA_PREFERRED_ROCE_GID_TYPE))
4972 cma_dev->default_gid_type[i - rdma_start_port(device)] =
4973 CMA_PREFERRED_ROCE_GID_TYPE;
4975 cma_dev->default_gid_type[i - rdma_start_port(device)] =
4976 find_first_bit(&supported_gids, BITS_PER_LONG);
4977 cma_dev->default_roce_tos[i - rdma_start_port(device)] = 0;
4980 init_completion(&cma_dev->comp);
4981 refcount_set(&cma_dev->refcount, 1);
4982 INIT_LIST_HEAD(&cma_dev->id_list);
4983 ib_set_client_data(device, &cma_client, cma_dev);
4986 list_add_tail(&cma_dev->list, &dev_list);
4987 list_for_each_entry(id_priv, &listen_any_list, list) {
4988 ret = cma_listen_on_dev(id_priv, cma_dev, &to_destroy);
4992 mutex_unlock(&lock);
4994 trace_cm_add_one(device);
4998 list_del(&cma_dev->list);
4999 mutex_unlock(&lock);
5001 /* cma_process_remove() will delete to_destroy */
5002 cma_process_remove(cma_dev);
5003 kfree(cma_dev->default_roce_tos);
5005 kfree(cma_dev->default_gid_type);
5012 static void cma_remove_one(struct ib_device *device, void *client_data)
5014 struct cma_device *cma_dev = client_data;
5016 trace_cm_remove_one(device);
5019 list_del(&cma_dev->list);
5020 mutex_unlock(&lock);
5022 cma_process_remove(cma_dev);
5023 kfree(cma_dev->default_roce_tos);
5024 kfree(cma_dev->default_gid_type);
5028 static int cma_init_net(struct net *net)
5030 struct cma_pernet *pernet = cma_pernet(net);
5032 xa_init(&pernet->tcp_ps);
5033 xa_init(&pernet->udp_ps);
5034 xa_init(&pernet->ipoib_ps);
5035 xa_init(&pernet->ib_ps);
5040 static void cma_exit_net(struct net *net)
5042 struct cma_pernet *pernet = cma_pernet(net);
5044 WARN_ON(!xa_empty(&pernet->tcp_ps));
5045 WARN_ON(!xa_empty(&pernet->udp_ps));
5046 WARN_ON(!xa_empty(&pernet->ipoib_ps));
5047 WARN_ON(!xa_empty(&pernet->ib_ps));
5050 static struct pernet_operations cma_pernet_operations = {
5051 .init = cma_init_net,
5052 .exit = cma_exit_net,
5053 .id = &cma_pernet_id,
5054 .size = sizeof(struct cma_pernet),
5057 static int __init cma_init(void)
5062 * There is a rare lock ordering dependency in cma_netdev_callback()
5063 * that only happens when bonding is enabled. Teach lockdep that rtnl
5064 * must never be nested under lock so it can find these without having
5065 * to test with bonding.
5067 if (IS_ENABLED(CONFIG_LOCKDEP)) {
5070 mutex_unlock(&lock);
5074 cma_wq = alloc_ordered_workqueue("rdma_cm", WQ_MEM_RECLAIM);
5078 ret = register_pernet_subsys(&cma_pernet_operations);
5082 ib_sa_register_client(&sa_client);
5083 register_netdevice_notifier(&cma_nb);
5085 ret = ib_register_client(&cma_client);
5089 ret = cma_configfs_init();
5096 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);
5102 destroy_workqueue(cma_wq);
5106 static void __exit cma_cleanup(void)
5108 cma_configfs_exit();
5109 ib_unregister_client(&cma_client);
5110 unregister_netdevice_notifier(&cma_nb);
5111 ib_sa_unregister_client(&sa_client);
5112 unregister_pernet_subsys(&cma_pernet_operations);
5113 destroy_workqueue(cma_wq);
5116 module_init(cma_init);
5117 module_exit(cma_cleanup);