]> git.proxmox.com Git - mirror_ubuntu-hirsute-kernel.git/blob - drivers/infiniband/core/cma.c
projects / mirror_ubuntu-hirsute-kernel.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
RDMA/cma: Remove cma_comp()
[mirror_ubuntu-hirsute-kernel.git] / drivers / infiniband / core / cma.c
1 // SPDX-License-Identifier: GPL-2.0 OR Linux-OpenIB
2 /*
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.
7 */
8
9 #include <linux/completion.h>
10 #include <linux/in.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>
20
21 #include <net/net_namespace.h>
22 #include <net/netns/generic.h>
23 #include <net/tcp.h>
24 #include <net/ipv6.h>
25 #include <net/ip_fib.h>
26 #include <net/ip6_route.h>
27
28 #include <rdma/rdma_cm.h>
29 #include <rdma/rdma_cm_ib.h>
30 #include <rdma/rdma_netlink.h>
31 #include <rdma/ib.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>
36
37 #include "core_priv.h"
38 #include "cma_priv.h"
39 #include "cma_trace.h"
40
41 MODULE_AUTHOR("Sean Hefty");
42 MODULE_DESCRIPTION("Generic RDMA CM Agent");
43 MODULE_LICENSE("Dual BSD/GPL");
44
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
51
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",
69 };
70
71 const char *__attribute_const__ rdma_event_msg(enum rdma_cm_event_type event)
72 {
73 size_t index = event;
74
75 return (index < ARRAY_SIZE(cma_events) && cma_events[index]) ?
76 cma_events[index] : "unrecognized event";
77 }
78 EXPORT_SYMBOL(rdma_event_msg);
79
80 const char *__attribute_const__ rdma_reject_msg(struct rdma_cm_id *id,
81 int reason)
82 {
83 if (rdma_ib_or_roce(id->device, id->port_num))
84 return ibcm_reject_msg(reason);
85
86 if (rdma_protocol_iwarp(id->device, id->port_num))
87 return iwcm_reject_msg(reason);
88
89 WARN_ON_ONCE(1);
90 return "unrecognized transport";
91 }
92 EXPORT_SYMBOL(rdma_reject_msg);
93
94 /**
95 * rdma_is_consumer_reject - return true if the consumer rejected the connect
96 * request.
97 * @id: Communication identifier that received the REJECT event.
98 * @reason: Value returned in the REJECT event status field.
99 */
100 static bool rdma_is_consumer_reject(struct rdma_cm_id *id, int reason)
101 {
102 if (rdma_ib_or_roce(id->device, id->port_num))
103 return reason == IB_CM_REJ_CONSUMER_DEFINED;
104
105 if (rdma_protocol_iwarp(id->device, id->port_num))
106 return reason == -ECONNREFUSED;
107
108 WARN_ON_ONCE(1);
109 return false;
110 }
111
112 const void *rdma_consumer_reject_data(struct rdma_cm_id *id,
113 struct rdma_cm_event *ev, u8 *data_len)
114 {
115 const void *p;
116
117 if (rdma_is_consumer_reject(id, ev->status)) {
118 *data_len = ev->param.conn.private_data_len;
119 p = ev->param.conn.private_data;
120 } else {
121 *data_len = 0;
122 p = NULL;
123 }
124 return p;
125 }
126 EXPORT_SYMBOL(rdma_consumer_reject_data);
127
128 /**
129 * rdma_iw_cm_id() - return the iw_cm_id pointer for this cm_id.
130 * @id: Communication Identifier
131 */
132 struct iw_cm_id *rdma_iw_cm_id(struct rdma_cm_id *id)
133 {
134 struct rdma_id_private *id_priv;
135
136 id_priv = container_of(id, struct rdma_id_private, id);
137 if (id->device->node_type == RDMA_NODE_RNIC)
138 return id_priv->cm_id.iw;
139 return NULL;
140 }
141 EXPORT_SYMBOL(rdma_iw_cm_id);
142
143 /**
144 * rdma_res_to_id() - return the rdma_cm_id pointer for this restrack.
145 * @res: rdma resource tracking entry pointer
146 */
147 struct rdma_cm_id *rdma_res_to_id(struct rdma_restrack_entry *res)
148 {
149 struct rdma_id_private *id_priv =
150 container_of(res, struct rdma_id_private, res);
151
152 return &id_priv->id;
153 }
154 EXPORT_SYMBOL(rdma_res_to_id);
155
156 static int cma_add_one(struct ib_device *device);
157 static void cma_remove_one(struct ib_device *device, void *client_data);
158
159 static struct ib_client cma_client = {
160 .name = "cma",
161 .add = cma_add_one,
162 .remove = cma_remove_one
163 };
164
165 static struct ib_sa_client sa_client;
166 static LIST_HEAD(dev_list);
167 static LIST_HEAD(listen_any_list);
168 static DEFINE_MUTEX(lock);
169 static struct workqueue_struct *cma_wq;
170 static unsigned int cma_pernet_id;
171
172 struct cma_pernet {
173 struct xarray tcp_ps;
174 struct xarray udp_ps;
175 struct xarray ipoib_ps;
176 struct xarray ib_ps;
177 };
178
179 static struct cma_pernet *cma_pernet(struct net *net)
180 {
181 return net_generic(net, cma_pernet_id);
182 }
183
184 static
185 struct xarray *cma_pernet_xa(struct net *net, enum rdma_ucm_port_space ps)
186 {
187 struct cma_pernet *pernet = cma_pernet(net);
188
189 switch (ps) {
190 case RDMA_PS_TCP:
191 return &pernet->tcp_ps;
192 case RDMA_PS_UDP:
193 return &pernet->udp_ps;
194 case RDMA_PS_IPOIB:
195 return &pernet->ipoib_ps;
196 case RDMA_PS_IB:
197 return &pernet->ib_ps;
198 default:
199 return NULL;
200 }
201 }
202
203 struct cma_device {
204 struct list_head list;
205 struct ib_device *device;
206 struct completion comp;
207 refcount_t refcount;
208 struct list_head id_list;
209 enum ib_gid_type *default_gid_type;
210 u8 *default_roce_tos;
211 };
212
213 struct rdma_bind_list {
214 enum rdma_ucm_port_space ps;
215 struct hlist_head owners;
216 unsigned short port;
217 };
218
219 struct class_port_info_context {
220 struct ib_class_port_info *class_port_info;
221 struct ib_device *device;
222 struct completion done;
223 struct ib_sa_query *sa_query;
224 u8 port_num;
225 };
226
227 static int cma_ps_alloc(struct net *net, enum rdma_ucm_port_space ps,
228 struct rdma_bind_list *bind_list, int snum)
229 {
230 struct xarray *xa = cma_pernet_xa(net, ps);
231
232 return xa_insert(xa, snum, bind_list, GFP_KERNEL);
233 }
234
235 static struct rdma_bind_list *cma_ps_find(struct net *net,
236 enum rdma_ucm_port_space ps, int snum)
237 {
238 struct xarray *xa = cma_pernet_xa(net, ps);
239
240 return xa_load(xa, snum);
241 }
242
243 static void cma_ps_remove(struct net *net, enum rdma_ucm_port_space ps,
244 int snum)
245 {
246 struct xarray *xa = cma_pernet_xa(net, ps);
247
248 xa_erase(xa, snum);
249 }
250
251 enum {
252 CMA_OPTION_AFONLY,
253 };
254
255 void cma_dev_get(struct cma_device *cma_dev)
256 {
257 refcount_inc(&cma_dev->refcount);
258 }
259
260 void cma_dev_put(struct cma_device *cma_dev)
261 {
262 if (refcount_dec_and_test(&cma_dev->refcount))
263 complete(&cma_dev->comp);
264 }
265
266 struct cma_device *cma_enum_devices_by_ibdev(cma_device_filter filter,
267 void *cookie)
268 {
269 struct cma_device *cma_dev;
270 struct cma_device *found_cma_dev = NULL;
271
272 mutex_lock(&lock);
273
274 list_for_each_entry(cma_dev, &dev_list, list)
275 if (filter(cma_dev->device, cookie)) {
276 found_cma_dev = cma_dev;
277 break;
278 }
279
280 if (found_cma_dev)
281 cma_dev_get(found_cma_dev);
282 mutex_unlock(&lock);
283 return found_cma_dev;
284 }
285
286 int cma_get_default_gid_type(struct cma_device *cma_dev,
287 unsigned int port)
288 {
289 if (!rdma_is_port_valid(cma_dev->device, port))
290 return -EINVAL;
291
292 return cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)];
293 }
294
295 int cma_set_default_gid_type(struct cma_device *cma_dev,
296 unsigned int port,
297 enum ib_gid_type default_gid_type)
298 {
299 unsigned long supported_gids;
300
301 if (!rdma_is_port_valid(cma_dev->device, port))
302 return -EINVAL;
303
304 supported_gids = roce_gid_type_mask_support(cma_dev->device, port);
305
306 if (!(supported_gids & 1 << default_gid_type))
307 return -EINVAL;
308
309 cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)] =
310 default_gid_type;
311
312 return 0;
313 }
314
315 int cma_get_default_roce_tos(struct cma_device *cma_dev, unsigned int port)
316 {
317 if (!rdma_is_port_valid(cma_dev->device, port))
318 return -EINVAL;
319
320 return cma_dev->default_roce_tos[port - rdma_start_port(cma_dev->device)];
321 }
322
323 int cma_set_default_roce_tos(struct cma_device *cma_dev, unsigned int port,
324 u8 default_roce_tos)
325 {
326 if (!rdma_is_port_valid(cma_dev->device, port))
327 return -EINVAL;
328
329 cma_dev->default_roce_tos[port - rdma_start_port(cma_dev->device)] =
330 default_roce_tos;
331
332 return 0;
333 }
334 struct ib_device *cma_get_ib_dev(struct cma_device *cma_dev)
335 {
336 return cma_dev->device;
337 }
338
339 /*
340 * Device removal can occur at anytime, so we need extra handling to
341 * serialize notifying the user of device removal with other callbacks.
342 * We do this by disabling removal notification while a callback is in process,
343 * and reporting it after the callback completes.
344 */
345
346 struct cma_multicast {
347 struct rdma_id_private *id_priv;
348 union {
349 struct ib_sa_multicast *ib;
350 } multicast;
351 struct list_head list;
352 void *context;
353 struct sockaddr_storage addr;
354 struct kref mcref;
355 u8 join_state;
356 };
357
358 struct cma_work {
359 struct work_struct work;
360 struct rdma_id_private *id;
361 enum rdma_cm_state old_state;
362 enum rdma_cm_state new_state;
363 struct rdma_cm_event event;
364 };
365
366 struct cma_ndev_work {
367 struct work_struct work;
368 struct rdma_id_private *id;
369 struct rdma_cm_event event;
370 };
371
372 struct iboe_mcast_work {
373 struct work_struct work;
374 struct rdma_id_private *id;
375 struct cma_multicast *mc;
376 };
377
378 union cma_ip_addr {
379 struct in6_addr ip6;
380 struct {
381 __be32 pad[3];
382 __be32 addr;
383 } ip4;
384 };
385
386 struct cma_hdr {
387 u8 cma_version;
388 u8 ip_version; /* IP version: 7:4 */
389 __be16 port;
390 union cma_ip_addr src_addr;
391 union cma_ip_addr dst_addr;
392 };
393
394 #define CMA_VERSION 0x00
395
396 struct cma_req_info {
397 struct sockaddr_storage listen_addr_storage;
398 struct sockaddr_storage src_addr_storage;
399 struct ib_device *device;
400 union ib_gid local_gid;
401 __be64 service_id;
402 int port;
403 bool has_gid;
404 u16 pkey;
405 };
406
407 static int cma_comp_exch(struct rdma_id_private *id_priv,
408 enum rdma_cm_state comp, enum rdma_cm_state exch)
409 {
410 unsigned long flags;
411 int ret;
412
413 /*
414 * The FSM uses a funny double locking where state is protected by both
415 * the handler_mutex and the spinlock. State is not allowed to change
416 * away from a handler_mutex protected value without also holding
417 * handler_mutex.
418 */
419 if (comp == RDMA_CM_CONNECT)
420 lockdep_assert_held(&id_priv->handler_mutex);
421
422 spin_lock_irqsave(&id_priv->lock, flags);
423 if ((ret = (id_priv->state == comp)))
424 id_priv->state = exch;
425 spin_unlock_irqrestore(&id_priv->lock, flags);
426 return ret;
427 }
428
429 static inline u8 cma_get_ip_ver(const struct cma_hdr *hdr)
430 {
431 return hdr->ip_version >> 4;
432 }
433
434 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
435 {
436 hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
437 }
438
439 static int cma_igmp_send(struct net_device *ndev, union ib_gid *mgid, bool join)
440 {
441 struct in_device *in_dev = NULL;
442
443 if (ndev) {
444 rtnl_lock();
445 in_dev = __in_dev_get_rtnl(ndev);
446 if (in_dev) {
447 if (join)
448 ip_mc_inc_group(in_dev,
449 *(__be32 *)(mgid->raw + 12));
450 else
451 ip_mc_dec_group(in_dev,
452 *(__be32 *)(mgid->raw + 12));
453 }
454 rtnl_unlock();
455 }
456 return (in_dev) ? 0 : -ENODEV;
457 }
458
459 static void _cma_attach_to_dev(struct rdma_id_private *id_priv,
460 struct cma_device *cma_dev)
461 {
462 cma_dev_get(cma_dev);
463 id_priv->cma_dev = cma_dev;
464 id_priv->id.device = cma_dev->device;
465 id_priv->id.route.addr.dev_addr.transport =
466 rdma_node_get_transport(cma_dev->device->node_type);
467 list_add_tail(&id_priv->list, &cma_dev->id_list);
468 if (id_priv->res.kern_name)
469 rdma_restrack_kadd(&id_priv->res);
470 else
471 rdma_restrack_uadd(&id_priv->res);
472 trace_cm_id_attach(id_priv, cma_dev->device);
473 }
474
475 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
476 struct cma_device *cma_dev)
477 {
478 _cma_attach_to_dev(id_priv, cma_dev);
479 id_priv->gid_type =
480 cma_dev->default_gid_type[id_priv->id.port_num -
481 rdma_start_port(cma_dev->device)];
482 }
483
484 static inline void release_mc(struct kref *kref)
485 {
486 struct cma_multicast *mc = container_of(kref, struct cma_multicast, mcref);
487
488 kfree(mc->multicast.ib);
489 kfree(mc);
490 }
491
492 static void cma_release_dev(struct rdma_id_private *id_priv)
493 {
494 mutex_lock(&lock);
495 list_del(&id_priv->list);
496 cma_dev_put(id_priv->cma_dev);
497 id_priv->cma_dev = NULL;
498 mutex_unlock(&lock);
499 }
500
501 static inline struct sockaddr *cma_src_addr(struct rdma_id_private *id_priv)
502 {
503 return (struct sockaddr *) &id_priv->id.route.addr.src_addr;
504 }
505
506 static inline struct sockaddr *cma_dst_addr(struct rdma_id_private *id_priv)
507 {
508 return (struct sockaddr *) &id_priv->id.route.addr.dst_addr;
509 }
510
511 static inline unsigned short cma_family(struct rdma_id_private *id_priv)
512 {
513 return id_priv->id.route.addr.src_addr.ss_family;
514 }
515
516 static int cma_set_qkey(struct rdma_id_private *id_priv, u32 qkey)
517 {
518 struct ib_sa_mcmember_rec rec;
519 int ret = 0;
520
521 if (id_priv->qkey) {
522 if (qkey && id_priv->qkey != qkey)
523 return -EINVAL;
524 return 0;
525 }
526
527 if (qkey) {
528 id_priv->qkey = qkey;
529 return 0;
530 }
531
532 switch (id_priv->id.ps) {
533 case RDMA_PS_UDP:
534 case RDMA_PS_IB:
535 id_priv->qkey = RDMA_UDP_QKEY;
536 break;
537 case RDMA_PS_IPOIB:
538 ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid);
539 ret = ib_sa_get_mcmember_rec(id_priv->id.device,
540 id_priv->id.port_num, &rec.mgid,
541 &rec);
542 if (!ret)
543 id_priv->qkey = be32_to_cpu(rec.qkey);
544 break;
545 default:
546 break;
547 }
548 return ret;
549 }
550
551 static void cma_translate_ib(struct sockaddr_ib *sib, struct rdma_dev_addr *dev_addr)
552 {
553 dev_addr->dev_type = ARPHRD_INFINIBAND;
554 rdma_addr_set_sgid(dev_addr, (union ib_gid *) &sib->sib_addr);
555 ib_addr_set_pkey(dev_addr, ntohs(sib->sib_pkey));
556 }
557
558 static int cma_translate_addr(struct sockaddr *addr, struct rdma_dev_addr *dev_addr)
559 {
560 int ret;
561
562 if (addr->sa_family != AF_IB) {
563 ret = rdma_translate_ip(addr, dev_addr);
564 } else {
565 cma_translate_ib((struct sockaddr_ib *) addr, dev_addr);
566 ret = 0;
567 }
568
569 return ret;
570 }
571
572 static const struct ib_gid_attr *
573 cma_validate_port(struct ib_device *device, u8 port,
574 enum ib_gid_type gid_type,
575 union ib_gid *gid,
576 struct rdma_id_private *id_priv)
577 {
578 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
579 int bound_if_index = dev_addr->bound_dev_if;
580 const struct ib_gid_attr *sgid_attr;
581 int dev_type = dev_addr->dev_type;
582 struct net_device *ndev = NULL;
583
584 if (!rdma_dev_access_netns(device, id_priv->id.route.addr.dev_addr.net))
585 return ERR_PTR(-ENODEV);
586
587 if ((dev_type == ARPHRD_INFINIBAND) && !rdma_protocol_ib(device, port))
588 return ERR_PTR(-ENODEV);
589
590 if ((dev_type != ARPHRD_INFINIBAND) && rdma_protocol_ib(device, port))
591 return ERR_PTR(-ENODEV);
592
593 if (dev_type == ARPHRD_ETHER && rdma_protocol_roce(device, port)) {
594 ndev = dev_get_by_index(dev_addr->net, bound_if_index);
595 if (!ndev)
596 return ERR_PTR(-ENODEV);
597 } else {
598 gid_type = IB_GID_TYPE_IB;
599 }
600
601 sgid_attr = rdma_find_gid_by_port(device, gid, gid_type, port, ndev);
602 if (ndev)
603 dev_put(ndev);
604 return sgid_attr;
605 }
606
607 static void cma_bind_sgid_attr(struct rdma_id_private *id_priv,
608 const struct ib_gid_attr *sgid_attr)
609 {
610 WARN_ON(id_priv->id.route.addr.dev_addr.sgid_attr);
611 id_priv->id.route.addr.dev_addr.sgid_attr = sgid_attr;
612 }
613
614 /**
615 * cma_acquire_dev_by_src_ip - Acquire cma device, port, gid attribute
616 * based on source ip address.
617 * @id_priv: cm_id which should be bound to cma device
618 *
619 * cma_acquire_dev_by_src_ip() binds cm id to cma device, port and GID attribute
620 * based on source IP address. It returns 0 on success or error code otherwise.
621 * It is applicable to active and passive side cm_id.
622 */
623 static int cma_acquire_dev_by_src_ip(struct rdma_id_private *id_priv)
624 {
625 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
626 const struct ib_gid_attr *sgid_attr;
627 union ib_gid gid, iboe_gid, *gidp;
628 struct cma_device *cma_dev;
629 enum ib_gid_type gid_type;
630 int ret = -ENODEV;
631 unsigned int port;
632
633 if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
634 id_priv->id.ps == RDMA_PS_IPOIB)
635 return -EINVAL;
636
637 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
638 &iboe_gid);
639
640 memcpy(&gid, dev_addr->src_dev_addr +
641 rdma_addr_gid_offset(dev_addr), sizeof(gid));
642
643 mutex_lock(&lock);
644 list_for_each_entry(cma_dev, &dev_list, list) {
645 rdma_for_each_port (cma_dev->device, port) {
646 gidp = rdma_protocol_roce(cma_dev->device, port) ?
647 &iboe_gid : &gid;
648 gid_type = cma_dev->default_gid_type[port - 1];
649 sgid_attr = cma_validate_port(cma_dev->device, port,
650 gid_type, gidp, id_priv);
651 if (!IS_ERR(sgid_attr)) {
652 id_priv->id.port_num = port;
653 cma_bind_sgid_attr(id_priv, sgid_attr);
654 cma_attach_to_dev(id_priv, cma_dev);
655 ret = 0;
656 goto out;
657 }
658 }
659 }
660 out:
661 mutex_unlock(&lock);
662 return ret;
663 }
664
665 /**
666 * cma_ib_acquire_dev - Acquire cma device, port and SGID attribute
667 * @id_priv: cm id to bind to cma device
668 * @listen_id_priv: listener cm id to match against
669 * @req: Pointer to req structure containaining incoming
670 * request information
671 * cma_ib_acquire_dev() acquires cma device, port and SGID attribute when
672 * rdma device matches for listen_id and incoming request. It also verifies
673 * that a GID table entry is present for the source address.
674 * Returns 0 on success, or returns error code otherwise.
675 */
676 static int cma_ib_acquire_dev(struct rdma_id_private *id_priv,
677 const struct rdma_id_private *listen_id_priv,
678 struct cma_req_info *req)
679 {
680 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
681 const struct ib_gid_attr *sgid_attr;
682 enum ib_gid_type gid_type;
683 union ib_gid gid;
684
685 if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
686 id_priv->id.ps == RDMA_PS_IPOIB)
687 return -EINVAL;
688
689 if (rdma_protocol_roce(req->device, req->port))
690 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
691 &gid);
692 else
693 memcpy(&gid, dev_addr->src_dev_addr +
694 rdma_addr_gid_offset(dev_addr), sizeof(gid));
695
696 gid_type = listen_id_priv->cma_dev->default_gid_type[req->port - 1];
697 sgid_attr = cma_validate_port(req->device, req->port,
698 gid_type, &gid, id_priv);
699 if (IS_ERR(sgid_attr))
700 return PTR_ERR(sgid_attr);
701
702 id_priv->id.port_num = req->port;
703 cma_bind_sgid_attr(id_priv, sgid_attr);
704 /* Need to acquire lock to protect against reader
705 * of cma_dev->id_list such as cma_netdev_callback() and
706 * cma_process_remove().
707 */
708 mutex_lock(&lock);
709 cma_attach_to_dev(id_priv, listen_id_priv->cma_dev);
710 mutex_unlock(&lock);
711 return 0;
712 }
713
714 static int cma_iw_acquire_dev(struct rdma_id_private *id_priv,
715 const struct rdma_id_private *listen_id_priv)
716 {
717 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
718 const struct ib_gid_attr *sgid_attr;
719 struct cma_device *cma_dev;
720 enum ib_gid_type gid_type;
721 int ret = -ENODEV;
722 unsigned int port;
723 union ib_gid gid;
724
725 if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
726 id_priv->id.ps == RDMA_PS_IPOIB)
727 return -EINVAL;
728
729 memcpy(&gid, dev_addr->src_dev_addr +
730 rdma_addr_gid_offset(dev_addr), sizeof(gid));
731
732 mutex_lock(&lock);
733
734 cma_dev = listen_id_priv->cma_dev;
735 port = listen_id_priv->id.port_num;
736 gid_type = listen_id_priv->gid_type;
737 sgid_attr = cma_validate_port(cma_dev->device, port,
738 gid_type, &gid, id_priv);
739 if (!IS_ERR(sgid_attr)) {
740 id_priv->id.port_num = port;
741 cma_bind_sgid_attr(id_priv, sgid_attr);
742 ret = 0;
743 goto out;
744 }
745
746 list_for_each_entry(cma_dev, &dev_list, list) {
747 rdma_for_each_port (cma_dev->device, port) {
748 if (listen_id_priv->cma_dev == cma_dev &&
749 listen_id_priv->id.port_num == port)
750 continue;
751
752 gid_type = cma_dev->default_gid_type[port - 1];
753 sgid_attr = cma_validate_port(cma_dev->device, port,
754 gid_type, &gid, id_priv);
755 if (!IS_ERR(sgid_attr)) {
756 id_priv->id.port_num = port;
757 cma_bind_sgid_attr(id_priv, sgid_attr);
758 ret = 0;
759 goto out;
760 }
761 }
762 }
763
764 out:
765 if (!ret)
766 cma_attach_to_dev(id_priv, cma_dev);
767
768 mutex_unlock(&lock);
769 return ret;
770 }
771
772 /*
773 * Select the source IB device and address to reach the destination IB address.
774 */
775 static int cma_resolve_ib_dev(struct rdma_id_private *id_priv)
776 {
777 struct cma_device *cma_dev, *cur_dev;
778 struct sockaddr_ib *addr;
779 union ib_gid gid, sgid, *dgid;
780 unsigned int p;
781 u16 pkey, index;
782 enum ib_port_state port_state;
783 int i;
784
785 cma_dev = NULL;
786 addr = (struct sockaddr_ib *) cma_dst_addr(id_priv);
787 dgid = (union ib_gid *) &addr->sib_addr;
788 pkey = ntohs(addr->sib_pkey);
789
790 mutex_lock(&lock);
791 list_for_each_entry(cur_dev, &dev_list, list) {
792 rdma_for_each_port (cur_dev->device, p) {
793 if (!rdma_cap_af_ib(cur_dev->device, p))
794 continue;
795
796 if (ib_find_cached_pkey(cur_dev->device, p, pkey, &index))
797 continue;
798
799 if (ib_get_cached_port_state(cur_dev->device, p, &port_state))
800 continue;
801 for (i = 0; !rdma_query_gid(cur_dev->device,
802 p, i, &gid);
803 i++) {
804 if (!memcmp(&gid, dgid, sizeof(gid))) {
805 cma_dev = cur_dev;
806 sgid = gid;
807 id_priv->id.port_num = p;
808 goto found;
809 }
810
811 if (!cma_dev && (gid.global.subnet_prefix ==
812 dgid->global.subnet_prefix) &&
813 port_state == IB_PORT_ACTIVE) {
814 cma_dev = cur_dev;
815 sgid = gid;
816 id_priv->id.port_num = p;
817 goto found;
818 }
819 }
820 }
821 }
822 mutex_unlock(&lock);
823 return -ENODEV;
824
825 found:
826 cma_attach_to_dev(id_priv, cma_dev);
827 mutex_unlock(&lock);
828 addr = (struct sockaddr_ib *)cma_src_addr(id_priv);
829 memcpy(&addr->sib_addr, &sgid, sizeof(sgid));
830 cma_translate_ib(addr, &id_priv->id.route.addr.dev_addr);
831 return 0;
832 }
833
834 static void cma_id_get(struct rdma_id_private *id_priv)
835 {
836 refcount_inc(&id_priv->refcount);
837 }
838
839 static void cma_id_put(struct rdma_id_private *id_priv)
840 {
841 if (refcount_dec_and_test(&id_priv->refcount))
842 complete(&id_priv->comp);
843 }
844
845 struct rdma_cm_id *__rdma_create_id(struct net *net,
846 rdma_cm_event_handler event_handler,
847 void *context, enum rdma_ucm_port_space ps,
848 enum ib_qp_type qp_type, const char *caller)
849 {
850 struct rdma_id_private *id_priv;
851
852 id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
853 if (!id_priv)
854 return ERR_PTR(-ENOMEM);
855
856 rdma_restrack_set_task(&id_priv->res, caller);
857 id_priv->res.type = RDMA_RESTRACK_CM_ID;
858 id_priv->state = RDMA_CM_IDLE;
859 id_priv->id.context = context;
860 id_priv->id.event_handler = event_handler;
861 id_priv->id.ps = ps;
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;
876
877 return &id_priv->id;
878 }
879 EXPORT_SYMBOL(__rdma_create_id);
880
881 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
882 {
883 struct ib_qp_attr qp_attr;
884 int qp_attr_mask, ret;
885
886 qp_attr.qp_state = IB_QPS_INIT;
887 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
888 if (ret)
889 return ret;
890
891 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
892 if (ret)
893 return ret;
894
895 qp_attr.qp_state = IB_QPS_RTR;
896 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
897 if (ret)
898 return ret;
899
900 qp_attr.qp_state = IB_QPS_RTS;
901 qp_attr.sq_psn = 0;
902 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
903
904 return ret;
905 }
906
907 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
908 {
909 struct ib_qp_attr qp_attr;
910 int qp_attr_mask, ret;
911
912 qp_attr.qp_state = IB_QPS_INIT;
913 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
914 if (ret)
915 return ret;
916
917 return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
918 }
919
920 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
921 struct ib_qp_init_attr *qp_init_attr)
922 {
923 struct rdma_id_private *id_priv;
924 struct ib_qp *qp;
925 int ret;
926
927 id_priv = container_of(id, struct rdma_id_private, id);
928 if (id->device != pd->device) {
929 ret = -EINVAL;
930 goto out_err;
931 }
932
933 qp_init_attr->port_num = id->port_num;
934 qp = ib_create_qp(pd, qp_init_attr);
935 if (IS_ERR(qp)) {
936 ret = PTR_ERR(qp);
937 goto out_err;
938 }
939
940 if (id->qp_type == IB_QPT_UD)
941 ret = cma_init_ud_qp(id_priv, qp);
942 else
943 ret = cma_init_conn_qp(id_priv, qp);
944 if (ret)
945 goto out_destroy;
946
947 id->qp = qp;
948 id_priv->qp_num = qp->qp_num;
949 id_priv->srq = (qp->srq != NULL);
950 trace_cm_qp_create(id_priv, pd, qp_init_attr, 0);
951 return 0;
952 out_destroy:
953 ib_destroy_qp(qp);
954 out_err:
955 trace_cm_qp_create(id_priv, pd, qp_init_attr, ret);
956 return ret;
957 }
958 EXPORT_SYMBOL(rdma_create_qp);
959
960 void rdma_destroy_qp(struct rdma_cm_id *id)
961 {
962 struct rdma_id_private *id_priv;
963
964 id_priv = container_of(id, struct rdma_id_private, id);
965 trace_cm_qp_destroy(id_priv);
966 mutex_lock(&id_priv->qp_mutex);
967 ib_destroy_qp(id_priv->id.qp);
968 id_priv->id.qp = NULL;
969 mutex_unlock(&id_priv->qp_mutex);
970 }
971 EXPORT_SYMBOL(rdma_destroy_qp);
972
973 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
974 struct rdma_conn_param *conn_param)
975 {
976 struct ib_qp_attr qp_attr;
977 int qp_attr_mask, ret;
978
979 mutex_lock(&id_priv->qp_mutex);
980 if (!id_priv->id.qp) {
981 ret = 0;
982 goto out;
983 }
984
985 /* Need to update QP attributes from default values. */
986 qp_attr.qp_state = IB_QPS_INIT;
987 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
988 if (ret)
989 goto out;
990
991 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
992 if (ret)
993 goto out;
994
995 qp_attr.qp_state = IB_QPS_RTR;
996 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
997 if (ret)
998 goto out;
999
1000 BUG_ON(id_priv->cma_dev->device != id_priv->id.device);
1001
1002 if (conn_param)
1003 qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
1004 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
1005 out:
1006 mutex_unlock(&id_priv->qp_mutex);
1007 return ret;
1008 }
1009
1010 static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
1011 struct rdma_conn_param *conn_param)
1012 {
1013 struct ib_qp_attr qp_attr;
1014 int qp_attr_mask, ret;
1015
1016 mutex_lock(&id_priv->qp_mutex);
1017 if (!id_priv->id.qp) {
1018 ret = 0;
1019 goto out;
1020 }
1021
1022 qp_attr.qp_state = IB_QPS_RTS;
1023 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1024 if (ret)
1025 goto out;
1026
1027 if (conn_param)
1028 qp_attr.max_rd_atomic = conn_param->initiator_depth;
1029 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
1030 out:
1031 mutex_unlock(&id_priv->qp_mutex);
1032 return ret;
1033 }
1034
1035 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
1036 {
1037 struct ib_qp_attr qp_attr;
1038 int ret;
1039
1040 mutex_lock(&id_priv->qp_mutex);
1041 if (!id_priv->id.qp) {
1042 ret = 0;
1043 goto out;
1044 }
1045
1046 qp_attr.qp_state = IB_QPS_ERR;
1047 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
1048 out:
1049 mutex_unlock(&id_priv->qp_mutex);
1050 return ret;
1051 }
1052
1053 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
1054 struct ib_qp_attr *qp_attr, int *qp_attr_mask)
1055 {
1056 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
1057 int ret;
1058 u16 pkey;
1059
1060 if (rdma_cap_eth_ah(id_priv->id.device, id_priv->id.port_num))
1061 pkey = 0xffff;
1062 else
1063 pkey = ib_addr_get_pkey(dev_addr);
1064
1065 ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
1066 pkey, &qp_attr->pkey_index);
1067 if (ret)
1068 return ret;
1069
1070 qp_attr->port_num = id_priv->id.port_num;
1071 *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
1072
1073 if (id_priv->id.qp_type == IB_QPT_UD) {
1074 ret = cma_set_qkey(id_priv, 0);
1075 if (ret)
1076 return ret;
1077
1078 qp_attr->qkey = id_priv->qkey;
1079 *qp_attr_mask |= IB_QP_QKEY;
1080 } else {
1081 qp_attr->qp_access_flags = 0;
1082 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
1083 }
1084 return 0;
1085 }
1086
1087 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
1088 int *qp_attr_mask)
1089 {
1090 struct rdma_id_private *id_priv;
1091 int ret = 0;
1092
1093 id_priv = container_of(id, struct rdma_id_private, id);
1094 if (rdma_cap_ib_cm(id->device, id->port_num)) {
1095 if (!id_priv->cm_id.ib || (id_priv->id.qp_type == IB_QPT_UD))
1096 ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
1097 else
1098 ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
1099 qp_attr_mask);
1100
1101 if (qp_attr->qp_state == IB_QPS_RTR)
1102 qp_attr->rq_psn = id_priv->seq_num;
1103 } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
1104 if (!id_priv->cm_id.iw) {
1105 qp_attr->qp_access_flags = 0;
1106 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
1107 } else
1108 ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
1109 qp_attr_mask);
1110 qp_attr->port_num = id_priv->id.port_num;
1111 *qp_attr_mask |= IB_QP_PORT;
1112 } else
1113 ret = -ENOSYS;
1114
1115 if ((*qp_attr_mask & IB_QP_TIMEOUT) && id_priv->timeout_set)
1116 qp_attr->timeout = id_priv->timeout;
1117
1118 return ret;
1119 }
1120 EXPORT_SYMBOL(rdma_init_qp_attr);
1121
1122 static inline bool cma_zero_addr(const struct sockaddr *addr)
1123 {
1124 switch (addr->sa_family) {
1125 case AF_INET:
1126 return ipv4_is_zeronet(((struct sockaddr_in *)addr)->sin_addr.s_addr);
1127 case AF_INET6:
1128 return ipv6_addr_any(&((struct sockaddr_in6 *)addr)->sin6_addr);
1129 case AF_IB:
1130 return ib_addr_any(&((struct sockaddr_ib *)addr)->sib_addr);
1131 default:
1132 return false;
1133 }
1134 }
1135
1136 static inline bool cma_loopback_addr(const struct sockaddr *addr)
1137 {
1138 switch (addr->sa_family) {
1139 case AF_INET:
1140 return ipv4_is_loopback(
1141 ((struct sockaddr_in *)addr)->sin_addr.s_addr);
1142 case AF_INET6:
1143 return ipv6_addr_loopback(
1144 &((struct sockaddr_in6 *)addr)->sin6_addr);
1145 case AF_IB:
1146 return ib_addr_loopback(
1147 &((struct sockaddr_ib *)addr)->sib_addr);
1148 default:
1149 return false;
1150 }
1151 }
1152
1153 static inline bool cma_any_addr(const struct sockaddr *addr)
1154 {
1155 return cma_zero_addr(addr) || cma_loopback_addr(addr);
1156 }
1157
1158 static int cma_addr_cmp(const struct sockaddr *src, const struct sockaddr *dst)
1159 {
1160 if (src->sa_family != dst->sa_family)
1161 return -1;
1162
1163 switch (src->sa_family) {
1164 case AF_INET:
1165 return ((struct sockaddr_in *)src)->sin_addr.s_addr !=
1166 ((struct sockaddr_in *)dst)->sin_addr.s_addr;
1167 case AF_INET6: {
1168 struct sockaddr_in6 *src_addr6 = (struct sockaddr_in6 *)src;
1169 struct sockaddr_in6 *dst_addr6 = (struct sockaddr_in6 *)dst;
1170 bool link_local;
1171
1172 if (ipv6_addr_cmp(&src_addr6->sin6_addr,
1173 &dst_addr6->sin6_addr))
1174 return 1;
1175 link_local = ipv6_addr_type(&dst_addr6->sin6_addr) &
1176 IPV6_ADDR_LINKLOCAL;
1177 /* Link local must match their scope_ids */
1178 return link_local ? (src_addr6->sin6_scope_id !=
1179 dst_addr6->sin6_scope_id) :
1180 0;
1181 }
1182
1183 default:
1184 return ib_addr_cmp(&((struct sockaddr_ib *) src)->sib_addr,
1185 &((struct sockaddr_ib *) dst)->sib_addr);
1186 }
1187 }
1188
1189 static __be16 cma_port(const struct sockaddr *addr)
1190 {
1191 struct sockaddr_ib *sib;
1192
1193 switch (addr->sa_family) {
1194 case AF_INET:
1195 return ((struct sockaddr_in *) addr)->sin_port;
1196 case AF_INET6:
1197 return ((struct sockaddr_in6 *) addr)->sin6_port;
1198 case AF_IB:
1199 sib = (struct sockaddr_ib *) addr;
1200 return htons((u16) (be64_to_cpu(sib->sib_sid) &
1201 be64_to_cpu(sib->sib_sid_mask)));
1202 default:
1203 return 0;
1204 }
1205 }
1206
1207 static inline int cma_any_port(const struct sockaddr *addr)
1208 {
1209 return !cma_port(addr);
1210 }
1211
1212 static void cma_save_ib_info(struct sockaddr *src_addr,
1213 struct sockaddr *dst_addr,
1214 const struct rdma_cm_id *listen_id,
1215 const struct sa_path_rec *path)
1216 {
1217 struct sockaddr_ib *listen_ib, *ib;
1218
1219 listen_ib = (struct sockaddr_ib *) &listen_id->route.addr.src_addr;
1220 if (src_addr) {
1221 ib = (struct sockaddr_ib *)src_addr;
1222 ib->sib_family = AF_IB;
1223 if (path) {
1224 ib->sib_pkey = path->pkey;
1225 ib->sib_flowinfo = path->flow_label;
1226 memcpy(&ib->sib_addr, &path->sgid, 16);
1227 ib->sib_sid = path->service_id;
1228 ib->sib_scope_id = 0;
1229 } else {
1230 ib->sib_pkey = listen_ib->sib_pkey;
1231 ib->sib_flowinfo = listen_ib->sib_flowinfo;
1232 ib->sib_addr = listen_ib->sib_addr;
1233 ib->sib_sid = listen_ib->sib_sid;
1234 ib->sib_scope_id = listen_ib->sib_scope_id;
1235 }
1236 ib->sib_sid_mask = cpu_to_be64(0xffffffffffffffffULL);
1237 }
1238 if (dst_addr) {
1239 ib = (struct sockaddr_ib *)dst_addr;
1240 ib->sib_family = AF_IB;
1241 if (path) {
1242 ib->sib_pkey = path->pkey;
1243 ib->sib_flowinfo = path->flow_label;
1244 memcpy(&ib->sib_addr, &path->dgid, 16);
1245 }
1246 }
1247 }
1248
1249 static void cma_save_ip4_info(struct sockaddr_in *src_addr,
1250 struct sockaddr_in *dst_addr,
1251 struct cma_hdr *hdr,
1252 __be16 local_port)
1253 {
1254 if (src_addr) {
1255 *src_addr = (struct sockaddr_in) {
1256 .sin_family = AF_INET,
1257 .sin_addr.s_addr = hdr->dst_addr.ip4.addr,
1258 .sin_port = local_port,
1259 };
1260 }
1261
1262 if (dst_addr) {
1263 *dst_addr = (struct sockaddr_in) {
1264 .sin_family = AF_INET,
1265 .sin_addr.s_addr = hdr->src_addr.ip4.addr,
1266 .sin_port = hdr->port,
1267 };
1268 }
1269 }
1270
1271 static void cma_save_ip6_info(struct sockaddr_in6 *src_addr,
1272 struct sockaddr_in6 *dst_addr,
1273 struct cma_hdr *hdr,
1274 __be16 local_port)
1275 {
1276 if (src_addr) {
1277 *src_addr = (struct sockaddr_in6) {
1278 .sin6_family = AF_INET6,
1279 .sin6_addr = hdr->dst_addr.ip6,
1280 .sin6_port = local_port,
1281 };
1282 }
1283
1284 if (dst_addr) {
1285 *dst_addr = (struct sockaddr_in6) {
1286 .sin6_family = AF_INET6,
1287 .sin6_addr = hdr->src_addr.ip6,
1288 .sin6_port = hdr->port,
1289 };
1290 }
1291 }
1292
1293 static u16 cma_port_from_service_id(__be64 service_id)
1294 {
1295 return (u16)be64_to_cpu(service_id);
1296 }
1297
1298 static int cma_save_ip_info(struct sockaddr *src_addr,
1299 struct sockaddr *dst_addr,
1300 const struct ib_cm_event *ib_event,
1301 __be64 service_id)
1302 {
1303 struct cma_hdr *hdr;
1304 __be16 port;
1305
1306 hdr = ib_event->private_data;
1307 if (hdr->cma_version != CMA_VERSION)
1308 return -EINVAL;
1309
1310 port = htons(cma_port_from_service_id(service_id));
1311
1312 switch (cma_get_ip_ver(hdr)) {
1313 case 4:
1314 cma_save_ip4_info((struct sockaddr_in *)src_addr,
1315 (struct sockaddr_in *)dst_addr, hdr, port);
1316 break;
1317 case 6:
1318 cma_save_ip6_info((struct sockaddr_in6 *)src_addr,
1319 (struct sockaddr_in6 *)dst_addr, hdr, port);
1320 break;
1321 default:
1322 return -EAFNOSUPPORT;
1323 }
1324
1325 return 0;
1326 }
1327
1328 static int cma_save_net_info(struct sockaddr *src_addr,
1329 struct sockaddr *dst_addr,
1330 const struct rdma_cm_id *listen_id,
1331 const struct ib_cm_event *ib_event,
1332 sa_family_t sa_family, __be64 service_id)
1333 {
1334 if (sa_family == AF_IB) {
1335 if (ib_event->event == IB_CM_REQ_RECEIVED)
1336 cma_save_ib_info(src_addr, dst_addr, listen_id,
1337 ib_event->param.req_rcvd.primary_path);
1338 else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED)
1339 cma_save_ib_info(src_addr, dst_addr, listen_id, NULL);
1340 return 0;
1341 }
1342
1343 return cma_save_ip_info(src_addr, dst_addr, ib_event, service_id);
1344 }
1345
1346 static int cma_save_req_info(const struct ib_cm_event *ib_event,
1347 struct cma_req_info *req)
1348 {
1349 const struct ib_cm_req_event_param *req_param =
1350 &ib_event->param.req_rcvd;
1351 const struct ib_cm_sidr_req_event_param *sidr_param =
1352 &ib_event->param.sidr_req_rcvd;
1353
1354 switch (ib_event->event) {
1355 case IB_CM_REQ_RECEIVED:
1356 req->device = req_param->listen_id->device;
1357 req->port = req_param->port;
1358 memcpy(&req->local_gid, &req_param->primary_path->sgid,
1359 sizeof(req->local_gid));
1360 req->has_gid = true;
1361 req->service_id = req_param->primary_path->service_id;
1362 req->pkey = be16_to_cpu(req_param->primary_path->pkey);
1363 if (req->pkey != req_param->bth_pkey)
1364 pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and primary path P_Key (0x%x)\n"
1365 "RDMA CMA: in the future this may cause the request to be dropped\n",
1366 req_param->bth_pkey, req->pkey);
1367 break;
1368 case IB_CM_SIDR_REQ_RECEIVED:
1369 req->device = sidr_param->listen_id->device;
1370 req->port = sidr_param->port;
1371 req->has_gid = false;
1372 req->service_id = sidr_param->service_id;
1373 req->pkey = sidr_param->pkey;
1374 if (req->pkey != sidr_param->bth_pkey)
1375 pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and SIDR request payload P_Key (0x%x)\n"
1376 "RDMA CMA: in the future this may cause the request to be dropped\n",
1377 sidr_param->bth_pkey, req->pkey);
1378 break;
1379 default:
1380 return -EINVAL;
1381 }
1382
1383 return 0;
1384 }
1385
1386 static bool validate_ipv4_net_dev(struct net_device *net_dev,
1387 const struct sockaddr_in *dst_addr,
1388 const struct sockaddr_in *src_addr)
1389 {
1390 __be32 daddr = dst_addr->sin_addr.s_addr,
1391 saddr = src_addr->sin_addr.s_addr;
1392 struct fib_result res;
1393 struct flowi4 fl4;
1394 int err;
1395 bool ret;
1396
1397 if (ipv4_is_multicast(saddr) || ipv4_is_lbcast(saddr) ||
1398 ipv4_is_lbcast(daddr) || ipv4_is_zeronet(saddr) ||
1399 ipv4_is_zeronet(daddr) || ipv4_is_loopback(daddr) ||
1400 ipv4_is_loopback(saddr))
1401 return false;
1402
1403 memset(&fl4, 0, sizeof(fl4));
1404 fl4.flowi4_iif = net_dev->ifindex;
1405 fl4.daddr = daddr;
1406 fl4.saddr = saddr;
1407
1408 rcu_read_lock();
1409 err = fib_lookup(dev_net(net_dev), &fl4, &res, 0);
1410 ret = err == 0 && FIB_RES_DEV(res) == net_dev;
1411 rcu_read_unlock();
1412
1413 return ret;
1414 }
1415
1416 static bool validate_ipv6_net_dev(struct net_device *net_dev,
1417 const struct sockaddr_in6 *dst_addr,
1418 const struct sockaddr_in6 *src_addr)
1419 {
1420 #if IS_ENABLED(CONFIG_IPV6)
1421 const int strict = ipv6_addr_type(&dst_addr->sin6_addr) &
1422 IPV6_ADDR_LINKLOCAL;
1423 struct rt6_info *rt = rt6_lookup(dev_net(net_dev), &dst_addr->sin6_addr,
1424 &src_addr->sin6_addr, net_dev->ifindex,
1425 NULL, strict);
1426 bool ret;
1427
1428 if (!rt)
1429 return false;
1430
1431 ret = rt->rt6i_idev->dev == net_dev;
1432 ip6_rt_put(rt);
1433
1434 return ret;
1435 #else
1436 return false;
1437 #endif
1438 }
1439
1440 static bool validate_net_dev(struct net_device *net_dev,
1441 const struct sockaddr *daddr,
1442 const struct sockaddr *saddr)
1443 {
1444 const struct sockaddr_in *daddr4 = (const struct sockaddr_in *)daddr;
1445 const struct sockaddr_in *saddr4 = (const struct sockaddr_in *)saddr;
1446 const struct sockaddr_in6 *daddr6 = (const struct sockaddr_in6 *)daddr;
1447 const struct sockaddr_in6 *saddr6 = (const struct sockaddr_in6 *)saddr;
1448
1449 switch (daddr->sa_family) {
1450 case AF_INET:
1451 return saddr->sa_family == AF_INET &&
1452 validate_ipv4_net_dev(net_dev, daddr4, saddr4);
1453
1454 case AF_INET6:
1455 return saddr->sa_family == AF_INET6 &&
1456 validate_ipv6_net_dev(net_dev, daddr6, saddr6);
1457
1458 default:
1459 return false;
1460 }
1461 }
1462
1463 static struct net_device *
1464 roce_get_net_dev_by_cm_event(const struct ib_cm_event *ib_event)
1465 {
1466 const struct ib_gid_attr *sgid_attr = NULL;
1467 struct net_device *ndev;
1468
1469 if (ib_event->event == IB_CM_REQ_RECEIVED)
1470 sgid_attr = ib_event->param.req_rcvd.ppath_sgid_attr;
1471 else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED)
1472 sgid_attr = ib_event->param.sidr_req_rcvd.sgid_attr;
1473
1474 if (!sgid_attr)
1475 return NULL;
1476
1477 rcu_read_lock();
1478 ndev = rdma_read_gid_attr_ndev_rcu(sgid_attr);
1479 if (IS_ERR(ndev))
1480 ndev = NULL;
1481 else
1482 dev_hold(ndev);
1483 rcu_read_unlock();
1484 return ndev;
1485 }
1486
1487 static struct net_device *cma_get_net_dev(const struct ib_cm_event *ib_event,
1488 struct cma_req_info *req)
1489 {
1490 struct sockaddr *listen_addr =
1491 (struct sockaddr *)&req->listen_addr_storage;
1492 struct sockaddr *src_addr = (struct sockaddr *)&req->src_addr_storage;
1493 struct net_device *net_dev;
1494 const union ib_gid *gid = req->has_gid ? &req->local_gid : NULL;
1495 int err;
1496
1497 err = cma_save_ip_info(listen_addr, src_addr, ib_event,
1498 req->service_id);
1499 if (err)
1500 return ERR_PTR(err);
1501
1502 if (rdma_protocol_roce(req->device, req->port))
1503 net_dev = roce_get_net_dev_by_cm_event(ib_event);
1504 else
1505 net_dev = ib_get_net_dev_by_params(req->device, req->port,
1506 req->pkey,
1507 gid, listen_addr);
1508 if (!net_dev)
1509 return ERR_PTR(-ENODEV);
1510
1511 return net_dev;
1512 }
1513
1514 static enum rdma_ucm_port_space rdma_ps_from_service_id(__be64 service_id)
1515 {
1516 return (be64_to_cpu(service_id) >> 16) & 0xffff;
1517 }
1518
1519 static bool cma_match_private_data(struct rdma_id_private *id_priv,
1520 const struct cma_hdr *hdr)
1521 {
1522 struct sockaddr *addr = cma_src_addr(id_priv);
1523 __be32 ip4_addr;
1524 struct in6_addr ip6_addr;
1525
1526 if (cma_any_addr(addr) && !id_priv->afonly)
1527 return true;
1528
1529 switch (addr->sa_family) {
1530 case AF_INET:
1531 ip4_addr = ((struct sockaddr_in *)addr)->sin_addr.s_addr;
1532 if (cma_get_ip_ver(hdr) != 4)
1533 return false;
1534 if (!cma_any_addr(addr) &&
1535 hdr->dst_addr.ip4.addr != ip4_addr)
1536 return false;
1537 break;
1538 case AF_INET6:
1539 ip6_addr = ((struct sockaddr_in6 *)addr)->sin6_addr;
1540 if (cma_get_ip_ver(hdr) != 6)
1541 return false;
1542 if (!cma_any_addr(addr) &&
1543 memcmp(&hdr->dst_addr.ip6, &ip6_addr, sizeof(ip6_addr)))
1544 return false;
1545 break;
1546 case AF_IB:
1547 return true;
1548 default:
1549 return false;
1550 }
1551
1552 return true;
1553 }
1554
1555 static bool cma_protocol_roce(const struct rdma_cm_id *id)
1556 {
1557 struct ib_device *device = id->device;
1558 const int port_num = id->port_num ?: rdma_start_port(device);
1559
1560 return rdma_protocol_roce(device, port_num);
1561 }
1562
1563 static bool cma_is_req_ipv6_ll(const struct cma_req_info *req)
1564 {
1565 const struct sockaddr *daddr =
1566 (const struct sockaddr *)&req->listen_addr_storage;
1567 const struct sockaddr_in6 *daddr6 = (const struct sockaddr_in6 *)daddr;
1568
1569 /* Returns true if the req is for IPv6 link local */
1570 return (daddr->sa_family == AF_INET6 &&
1571 (ipv6_addr_type(&daddr6->sin6_addr) & IPV6_ADDR_LINKLOCAL));
1572 }
1573
1574 static bool cma_match_net_dev(const struct rdma_cm_id *id,
1575 const struct net_device *net_dev,
1576 const struct cma_req_info *req)
1577 {
1578 const struct rdma_addr *addr = &id->route.addr;
1579
1580 if (!net_dev)
1581 /* This request is an AF_IB request */
1582 return (!id->port_num || id->port_num == req->port) &&
1583 (addr->src_addr.ss_family == AF_IB);
1584
1585 /*
1586 * If the request is not for IPv6 link local, allow matching
1587 * request to any netdevice of the one or multiport rdma device.
1588 */
1589 if (!cma_is_req_ipv6_ll(req))
1590 return true;
1591 /*
1592 * Net namespaces must match, and if the listner is listening
1593 * on a specific netdevice than netdevice must match as well.
1594 */
1595 if (net_eq(dev_net(net_dev), addr->dev_addr.net) &&
1596 (!!addr->dev_addr.bound_dev_if ==
1597 (addr->dev_addr.bound_dev_if == net_dev->ifindex)))
1598 return true;
1599 else
1600 return false;
1601 }
1602
1603 static struct rdma_id_private *cma_find_listener(
1604 const struct rdma_bind_list *bind_list,
1605 const struct ib_cm_id *cm_id,
1606 const struct ib_cm_event *ib_event,
1607 const struct cma_req_info *req,
1608 const struct net_device *net_dev)
1609 {
1610 struct rdma_id_private *id_priv, *id_priv_dev;
1611
1612 lockdep_assert_held(&lock);
1613
1614 if (!bind_list)
1615 return ERR_PTR(-EINVAL);
1616
1617 hlist_for_each_entry(id_priv, &bind_list->owners, node) {
1618 if (cma_match_private_data(id_priv, ib_event->private_data)) {
1619 if (id_priv->id.device == cm_id->device &&
1620 cma_match_net_dev(&id_priv->id, net_dev, req))
1621 return id_priv;
1622 list_for_each_entry(id_priv_dev,
1623 &id_priv->listen_list,
1624 listen_list) {
1625 if (id_priv_dev->id.device == cm_id->device &&
1626 cma_match_net_dev(&id_priv_dev->id,
1627 net_dev, req))
1628 return id_priv_dev;
1629 }
1630 }
1631 }
1632
1633 return ERR_PTR(-EINVAL);
1634 }
1635
1636 static struct rdma_id_private *
1637 cma_ib_id_from_event(struct ib_cm_id *cm_id,
1638 const struct ib_cm_event *ib_event,
1639 struct cma_req_info *req,
1640 struct net_device **net_dev)
1641 {
1642 struct rdma_bind_list *bind_list;
1643 struct rdma_id_private *id_priv;
1644 int err;
1645
1646 err = cma_save_req_info(ib_event, req);
1647 if (err)
1648 return ERR_PTR(err);
1649
1650 *net_dev = cma_get_net_dev(ib_event, req);
1651 if (IS_ERR(*net_dev)) {
1652 if (PTR_ERR(*net_dev) == -EAFNOSUPPORT) {
1653 /* Assuming the protocol is AF_IB */
1654 *net_dev = NULL;
1655 } else {
1656 return ERR_CAST(*net_dev);
1657 }
1658 }
1659
1660 mutex_lock(&lock);
1661 /*
1662 * Net namespace might be getting deleted while route lookup,
1663 * cm_id lookup is in progress. Therefore, perform netdevice
1664 * validation, cm_id lookup under rcu lock.
1665 * RCU lock along with netdevice state check, synchronizes with
1666 * netdevice migrating to different net namespace and also avoids
1667 * case where net namespace doesn't get deleted while lookup is in
1668 * progress.
1669 * If the device state is not IFF_UP, its properties such as ifindex
1670 * and nd_net cannot be trusted to remain valid without rcu lock.
1671 * net/core/dev.c change_net_namespace() ensures to synchronize with
1672 * ongoing operations on net device after device is closed using
1673 * synchronize_net().
1674 */
1675 rcu_read_lock();
1676 if (*net_dev) {
1677 /*
1678 * If netdevice is down, it is likely that it is administratively
1679 * down or it might be migrating to different namespace.
1680 * In that case avoid further processing, as the net namespace
1681 * or ifindex may change.
1682 */
1683 if (((*net_dev)->flags & IFF_UP) == 0) {
1684 id_priv = ERR_PTR(-EHOSTUNREACH);
1685 goto err;
1686 }
1687
1688 if (!validate_net_dev(*net_dev,
1689 (struct sockaddr *)&req->listen_addr_storage,
1690 (struct sockaddr *)&req->src_addr_storage)) {
1691 id_priv = ERR_PTR(-EHOSTUNREACH);
1692 goto err;
1693 }
1694 }
1695
1696 bind_list = cma_ps_find(*net_dev ? dev_net(*net_dev) : &init_net,
1697 rdma_ps_from_service_id(req->service_id),
1698 cma_port_from_service_id(req->service_id));
1699 id_priv = cma_find_listener(bind_list, cm_id, ib_event, req, *net_dev);
1700 err:
1701 rcu_read_unlock();
1702 mutex_unlock(&lock);
1703 if (IS_ERR(id_priv) && *net_dev) {
1704 dev_put(*net_dev);
1705 *net_dev = NULL;
1706 }
1707 return id_priv;
1708 }
1709
1710 static inline u8 cma_user_data_offset(struct rdma_id_private *id_priv)
1711 {
1712 return cma_family(id_priv) == AF_IB ? 0 : sizeof(struct cma_hdr);
1713 }
1714
1715 static void cma_cancel_route(struct rdma_id_private *id_priv)
1716 {
1717 if (rdma_cap_ib_sa(id_priv->id.device, id_priv->id.port_num)) {
1718 if (id_priv->query)
1719 ib_sa_cancel_query(id_priv->query_id, id_priv->query);
1720 }
1721 }
1722
1723 static void cma_cancel_listens(struct rdma_id_private *id_priv)
1724 {
1725 struct rdma_id_private *dev_id_priv;
1726
1727 /*
1728 * Remove from listen_any_list to prevent added devices from spawning
1729 * additional listen requests.
1730 */
1731 mutex_lock(&lock);
1732 list_del(&id_priv->list);
1733
1734 while (!list_empty(&id_priv->listen_list)) {
1735 dev_id_priv = list_entry(id_priv->listen_list.next,
1736 struct rdma_id_private, listen_list);
1737 /* sync with device removal to avoid duplicate destruction */
1738 list_del_init(&dev_id_priv->list);
1739 list_del(&dev_id_priv->listen_list);
1740 mutex_unlock(&lock);
1741
1742 rdma_destroy_id(&dev_id_priv->id);
1743 mutex_lock(&lock);
1744 }
1745 mutex_unlock(&lock);
1746 }
1747
1748 static void cma_cancel_operation(struct rdma_id_private *id_priv,
1749 enum rdma_cm_state state)
1750 {
1751 switch (state) {
1752 case RDMA_CM_ADDR_QUERY:
1753 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
1754 break;
1755 case RDMA_CM_ROUTE_QUERY:
1756 cma_cancel_route(id_priv);
1757 break;
1758 case RDMA_CM_LISTEN:
1759 if (cma_any_addr(cma_src_addr(id_priv)) && !id_priv->cma_dev)
1760 cma_cancel_listens(id_priv);
1761 break;
1762 default:
1763 break;
1764 }
1765 }
1766
1767 static void cma_release_port(struct rdma_id_private *id_priv)
1768 {
1769 struct rdma_bind_list *bind_list = id_priv->bind_list;
1770 struct net *net = id_priv->id.route.addr.dev_addr.net;
1771
1772 if (!bind_list)
1773 return;
1774
1775 mutex_lock(&lock);
1776 hlist_del(&id_priv->node);
1777 if (hlist_empty(&bind_list->owners)) {
1778 cma_ps_remove(net, bind_list->ps, bind_list->port);
1779 kfree(bind_list);
1780 }
1781 mutex_unlock(&lock);
1782 }
1783
1784 static void cma_leave_roce_mc_group(struct rdma_id_private *id_priv,
1785 struct cma_multicast *mc)
1786 {
1787 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
1788 struct net_device *ndev = NULL;
1789
1790 if (dev_addr->bound_dev_if)
1791 ndev = dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if);
1792 if (ndev) {
1793 cma_igmp_send(ndev, &mc->multicast.ib->rec.mgid, false);
1794 dev_put(ndev);
1795 }
1796 kref_put(&mc->mcref, release_mc);
1797 }
1798
1799 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
1800 {
1801 struct cma_multicast *mc;
1802
1803 while (!list_empty(&id_priv->mc_list)) {
1804 mc = container_of(id_priv->mc_list.next,
1805 struct cma_multicast, list);
1806 list_del(&mc->list);
1807 if (rdma_cap_ib_mcast(id_priv->cma_dev->device,
1808 id_priv->id.port_num)) {
1809 ib_sa_free_multicast(mc->multicast.ib);
1810 kfree(mc);
1811 } else {
1812 cma_leave_roce_mc_group(id_priv, mc);
1813 }
1814 }
1815 }
1816
1817 static void _destroy_id(struct rdma_id_private *id_priv,
1818 enum rdma_cm_state state)
1819 {
1820 cma_cancel_operation(id_priv, state);
1821
1822 rdma_restrack_del(&id_priv->res);
1823 if (id_priv->cma_dev) {
1824 if (rdma_cap_ib_cm(id_priv->id.device, 1)) {
1825 if (id_priv->cm_id.ib)
1826 ib_destroy_cm_id(id_priv->cm_id.ib);
1827 } else if (rdma_cap_iw_cm(id_priv->id.device, 1)) {
1828 if (id_priv->cm_id.iw)
1829 iw_destroy_cm_id(id_priv->cm_id.iw);
1830 }
1831 cma_leave_mc_groups(id_priv);
1832 cma_release_dev(id_priv);
1833 }
1834
1835 cma_release_port(id_priv);
1836 cma_id_put(id_priv);
1837 wait_for_completion(&id_priv->comp);
1838
1839 if (id_priv->internal_id)
1840 cma_id_put(id_priv->id.context);
1841
1842 kfree(id_priv->id.route.path_rec);
1843
1844 if (id_priv->id.route.addr.dev_addr.sgid_attr)
1845 rdma_put_gid_attr(id_priv->id.route.addr.dev_addr.sgid_attr);
1846
1847 put_net(id_priv->id.route.addr.dev_addr.net);
1848 kfree(id_priv);
1849 }
1850
1851 /*
1852 * destroy an ID from within the handler_mutex. This ensures that no other
1853 * handlers can start running concurrently.
1854 */
1855 static void destroy_id_handler_unlock(struct rdma_id_private *id_priv)
1856 __releases(&idprv->handler_mutex)
1857 {
1858 enum rdma_cm_state state;
1859 unsigned long flags;
1860
1861 trace_cm_id_destroy(id_priv);
1862
1863 /*
1864 * Setting the state to destroyed under the handler mutex provides a
1865 * fence against calling handler callbacks. If this is invoked due to
1866 * the failure of a handler callback then it guarentees that no future
1867 * handlers will be called.
1868 */
1869 lockdep_assert_held(&id_priv->handler_mutex);
1870 spin_lock_irqsave(&id_priv->lock, flags);
1871 state = id_priv->state;
1872 id_priv->state = RDMA_CM_DESTROYING;
1873 spin_unlock_irqrestore(&id_priv->lock, flags);
1874 mutex_unlock(&id_priv->handler_mutex);
1875 _destroy_id(id_priv, state);
1876 }
1877
1878 void rdma_destroy_id(struct rdma_cm_id *id)
1879 {
1880 struct rdma_id_private *id_priv =
1881 container_of(id, struct rdma_id_private, id);
1882
1883 mutex_lock(&id_priv->handler_mutex);
1884 destroy_id_handler_unlock(id_priv);
1885 }
1886 EXPORT_SYMBOL(rdma_destroy_id);
1887
1888 static int cma_rep_recv(struct rdma_id_private *id_priv)
1889 {
1890 int ret;
1891
1892 ret = cma_modify_qp_rtr(id_priv, NULL);
1893 if (ret)
1894 goto reject;
1895
1896 ret = cma_modify_qp_rts(id_priv, NULL);
1897 if (ret)
1898 goto reject;
1899
1900 trace_cm_send_rtu(id_priv);
1901 ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
1902 if (ret)
1903 goto reject;
1904
1905 return 0;
1906 reject:
1907 pr_debug_ratelimited("RDMA CM: CONNECT_ERROR: failed to handle reply. status %d\n", ret);
1908 cma_modify_qp_err(id_priv);
1909 trace_cm_send_rej(id_priv);
1910 ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
1911 NULL, 0, NULL, 0);
1912 return ret;
1913 }
1914
1915 static void cma_set_rep_event_data(struct rdma_cm_event *event,
1916 const struct ib_cm_rep_event_param *rep_data,
1917 void *private_data)
1918 {
1919 event->param.conn.private_data = private_data;
1920 event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
1921 event->param.conn.responder_resources = rep_data->responder_resources;
1922 event->param.conn.initiator_depth = rep_data->initiator_depth;
1923 event->param.conn.flow_control = rep_data->flow_control;
1924 event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
1925 event->param.conn.srq = rep_data->srq;
1926 event->param.conn.qp_num = rep_data->remote_qpn;
1927
1928 event->ece.vendor_id = rep_data->ece.vendor_id;
1929 event->ece.attr_mod = rep_data->ece.attr_mod;
1930 }
1931
1932 static int cma_cm_event_handler(struct rdma_id_private *id_priv,
1933 struct rdma_cm_event *event)
1934 {
1935 int ret;
1936
1937 lockdep_assert_held(&id_priv->handler_mutex);
1938
1939 trace_cm_event_handler(id_priv, event);
1940 ret = id_priv->id.event_handler(&id_priv->id, event);
1941 trace_cm_event_done(id_priv, event, ret);
1942 return ret;
1943 }
1944
1945 static int cma_ib_handler(struct ib_cm_id *cm_id,
1946 const struct ib_cm_event *ib_event)
1947 {
1948 struct rdma_id_private *id_priv = cm_id->context;
1949 struct rdma_cm_event event = {};
1950 enum rdma_cm_state state;
1951 int ret;
1952
1953 mutex_lock(&id_priv->handler_mutex);
1954 state = READ_ONCE(id_priv->state);
1955 if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
1956 state != RDMA_CM_CONNECT) ||
1957 (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
1958 state != RDMA_CM_DISCONNECT))
1959 goto out;
1960
1961 switch (ib_event->event) {
1962 case IB_CM_REQ_ERROR:
1963 case IB_CM_REP_ERROR:
1964 event.event = RDMA_CM_EVENT_UNREACHABLE;
1965 event.status = -ETIMEDOUT;
1966 break;
1967 case IB_CM_REP_RECEIVED:
1968 if (state == RDMA_CM_CONNECT &&
1969 (id_priv->id.qp_type != IB_QPT_UD)) {
1970 trace_cm_send_mra(id_priv);
1971 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
1972 }
1973 if (id_priv->id.qp) {
1974 event.status = cma_rep_recv(id_priv);
1975 event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
1976 RDMA_CM_EVENT_ESTABLISHED;
1977 } else {
1978 event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
1979 }
1980 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
1981 ib_event->private_data);
1982 break;
1983 case IB_CM_RTU_RECEIVED:
1984 case IB_CM_USER_ESTABLISHED:
1985 event.event = RDMA_CM_EVENT_ESTABLISHED;
1986 break;
1987 case IB_CM_DREQ_ERROR:
1988 event.status = -ETIMEDOUT;
1989 fallthrough;
1990 case IB_CM_DREQ_RECEIVED:
1991 case IB_CM_DREP_RECEIVED:
1992 if (!cma_comp_exch(id_priv, RDMA_CM_CONNECT,
1993 RDMA_CM_DISCONNECT))
1994 goto out;
1995 event.event = RDMA_CM_EVENT_DISCONNECTED;
1996 break;
1997 case IB_CM_TIMEWAIT_EXIT:
1998 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
1999 break;
2000 case IB_CM_MRA_RECEIVED:
2001 /* ignore event */
2002 goto out;
2003 case IB_CM_REJ_RECEIVED:
2004 pr_debug_ratelimited("RDMA CM: REJECTED: %s\n", rdma_reject_msg(&id_priv->id,
2005 ib_event->param.rej_rcvd.reason));
2006 cma_modify_qp_err(id_priv);
2007 event.status = ib_event->param.rej_rcvd.reason;
2008 event.event = RDMA_CM_EVENT_REJECTED;
2009 event.param.conn.private_data = ib_event->private_data;
2010 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
2011 break;
2012 default:
2013 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
2014 ib_event->event);
2015 goto out;
2016 }
2017
2018 ret = cma_cm_event_handler(id_priv, &event);
2019 if (ret) {
2020 /* Destroy the CM ID by returning a non-zero value. */
2021 id_priv->cm_id.ib = NULL;
2022 destroy_id_handler_unlock(id_priv);
2023 return ret;
2024 }
2025 out:
2026 mutex_unlock(&id_priv->handler_mutex);
2027 return 0;
2028 }
2029
2030 static struct rdma_id_private *
2031 cma_ib_new_conn_id(const struct rdma_cm_id *listen_id,
2032 const struct ib_cm_event *ib_event,
2033 struct net_device *net_dev)
2034 {
2035 struct rdma_id_private *listen_id_priv;
2036 struct rdma_id_private *id_priv;
2037 struct rdma_cm_id *id;
2038 struct rdma_route *rt;
2039 const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
2040 struct sa_path_rec *path = ib_event->param.req_rcvd.primary_path;
2041 const __be64 service_id =
2042 ib_event->param.req_rcvd.primary_path->service_id;
2043 int ret;
2044
2045 listen_id_priv = container_of(listen_id, struct rdma_id_private, id);
2046 id = __rdma_create_id(listen_id->route.addr.dev_addr.net,
2047 listen_id->event_handler, listen_id->context,
2048 listen_id->ps, ib_event->param.req_rcvd.qp_type,
2049 listen_id_priv->res.kern_name);
2050 if (IS_ERR(id))
2051 return NULL;
2052
2053 id_priv = container_of(id, struct rdma_id_private, id);
2054 if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
2055 (struct sockaddr *)&id->route.addr.dst_addr,
2056 listen_id, ib_event, ss_family, service_id))
2057 goto err;
2058
2059 rt = &id->route;
2060 rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
2061 rt->path_rec = kmalloc_array(rt->num_paths, sizeof(*rt->path_rec),
2062 GFP_KERNEL);
2063 if (!rt->path_rec)
2064 goto err;
2065
2066 rt->path_rec[0] = *path;
2067 if (rt->num_paths == 2)
2068 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
2069
2070 if (net_dev) {
2071 rdma_copy_src_l2_addr(&rt->addr.dev_addr, net_dev);
2072 } else {
2073 if (!cma_protocol_roce(listen_id) &&
2074 cma_any_addr(cma_src_addr(id_priv))) {
2075 rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
2076 rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
2077 ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
2078 } else if (!cma_any_addr(cma_src_addr(id_priv))) {
2079 ret = cma_translate_addr(cma_src_addr(id_priv), &rt->addr.dev_addr);
2080 if (ret)
2081 goto err;
2082 }
2083 }
2084 rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
2085
2086 id_priv->state = RDMA_CM_CONNECT;
2087 return id_priv;
2088
2089 err:
2090 rdma_destroy_id(id);
2091 return NULL;
2092 }
2093
2094 static struct rdma_id_private *
2095 cma_ib_new_udp_id(const struct rdma_cm_id *listen_id,
2096 const struct ib_cm_event *ib_event,
2097 struct net_device *net_dev)
2098 {
2099 const struct rdma_id_private *listen_id_priv;
2100 struct rdma_id_private *id_priv;
2101 struct rdma_cm_id *id;
2102 const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
2103 struct net *net = listen_id->route.addr.dev_addr.net;
2104 int ret;
2105
2106 listen_id_priv = container_of(listen_id, struct rdma_id_private, id);
2107 id = __rdma_create_id(net, listen_id->event_handler, listen_id->context,
2108 listen_id->ps, IB_QPT_UD,
2109 listen_id_priv->res.kern_name);
2110 if (IS_ERR(id))
2111 return NULL;
2112
2113 id_priv = container_of(id, struct rdma_id_private, id);
2114 if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
2115 (struct sockaddr *)&id->route.addr.dst_addr,
2116 listen_id, ib_event, ss_family,
2117 ib_event->param.sidr_req_rcvd.service_id))
2118 goto err;
2119
2120 if (net_dev) {
2121 rdma_copy_src_l2_addr(&id->route.addr.dev_addr, net_dev);
2122 } else {
2123 if (!cma_any_addr(cma_src_addr(id_priv))) {
2124 ret = cma_translate_addr(cma_src_addr(id_priv),
2125 &id->route.addr.dev_addr);
2126 if (ret)
2127 goto err;
2128 }
2129 }
2130
2131 id_priv->state = RDMA_CM_CONNECT;
2132 return id_priv;
2133 err:
2134 rdma_destroy_id(id);
2135 return NULL;
2136 }
2137
2138 static void cma_set_req_event_data(struct rdma_cm_event *event,
2139 const struct ib_cm_req_event_param *req_data,
2140 void *private_data, int offset)
2141 {
2142 event->param.conn.private_data = private_data + offset;
2143 event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
2144 event->param.conn.responder_resources = req_data->responder_resources;
2145 event->param.conn.initiator_depth = req_data->initiator_depth;
2146 event->param.conn.flow_control = req_data->flow_control;
2147 event->param.conn.retry_count = req_data->retry_count;
2148 event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
2149 event->param.conn.srq = req_data->srq;
2150 event->param.conn.qp_num = req_data->remote_qpn;
2151
2152 event->ece.vendor_id = req_data->ece.vendor_id;
2153 event->ece.attr_mod = req_data->ece.attr_mod;
2154 }
2155
2156 static int cma_ib_check_req_qp_type(const struct rdma_cm_id *id,
2157 const struct ib_cm_event *ib_event)
2158 {
2159 return (((ib_event->event == IB_CM_REQ_RECEIVED) &&
2160 (ib_event->param.req_rcvd.qp_type == id->qp_type)) ||
2161 ((ib_event->event == IB_CM_SIDR_REQ_RECEIVED) &&
2162 (id->qp_type == IB_QPT_UD)) ||
2163 (!id->qp_type));
2164 }
2165
2166 static int cma_ib_req_handler(struct ib_cm_id *cm_id,
2167 const struct ib_cm_event *ib_event)
2168 {
2169 struct rdma_id_private *listen_id, *conn_id = NULL;
2170 struct rdma_cm_event event = {};
2171 struct cma_req_info req = {};
2172 struct net_device *net_dev;
2173 u8 offset;
2174 int ret;
2175
2176 listen_id = cma_ib_id_from_event(cm_id, ib_event, &req, &net_dev);
2177 if (IS_ERR(listen_id))
2178 return PTR_ERR(listen_id);
2179
2180 trace_cm_req_handler(listen_id, ib_event->event);
2181 if (!cma_ib_check_req_qp_type(&listen_id->id, ib_event)) {
2182 ret = -EINVAL;
2183 goto net_dev_put;
2184 }
2185
2186 mutex_lock(&listen_id->handler_mutex);
2187 if (READ_ONCE(listen_id->state) != RDMA_CM_LISTEN) {
2188 ret = -ECONNABORTED;
2189 goto err_unlock;
2190 }
2191
2192 offset = cma_user_data_offset(listen_id);
2193 event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
2194 if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) {
2195 conn_id = cma_ib_new_udp_id(&listen_id->id, ib_event, net_dev);
2196 event.param.ud.private_data = ib_event->private_data + offset;
2197 event.param.ud.private_data_len =
2198 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
2199 } else {
2200 conn_id = cma_ib_new_conn_id(&listen_id->id, ib_event, net_dev);
2201 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
2202 ib_event->private_data, offset);
2203 }
2204 if (!conn_id) {
2205 ret = -ENOMEM;
2206 goto err_unlock;
2207 }
2208
2209 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
2210 ret = cma_ib_acquire_dev(conn_id, listen_id, &req);
2211 if (ret) {
2212 destroy_id_handler_unlock(conn_id);
2213 goto err_unlock;
2214 }
2215
2216 conn_id->cm_id.ib = cm_id;
2217 cm_id->context = conn_id;
2218 cm_id->cm_handler = cma_ib_handler;
2219
2220 ret = cma_cm_event_handler(conn_id, &event);
2221 if (ret) {
2222 /* Destroy the CM ID by returning a non-zero value. */
2223 conn_id->cm_id.ib = NULL;
2224 mutex_unlock(&listen_id->handler_mutex);
2225 destroy_id_handler_unlock(conn_id);
2226 goto net_dev_put;
2227 }
2228
2229 if (READ_ONCE(conn_id->state) == RDMA_CM_CONNECT &&
2230 conn_id->id.qp_type != IB_QPT_UD) {
2231 trace_cm_send_mra(cm_id->context);
2232 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
2233 }
2234 mutex_unlock(&conn_id->handler_mutex);
2235
2236 err_unlock:
2237 mutex_unlock(&listen_id->handler_mutex);
2238
2239 net_dev_put:
2240 if (net_dev)
2241 dev_put(net_dev);
2242
2243 return ret;
2244 }
2245
2246 __be64 rdma_get_service_id(struct rdma_cm_id *id, struct sockaddr *addr)
2247 {
2248 if (addr->sa_family == AF_IB)
2249 return ((struct sockaddr_ib *) addr)->sib_sid;
2250
2251 return cpu_to_be64(((u64)id->ps << 16) + be16_to_cpu(cma_port(addr)));
2252 }
2253 EXPORT_SYMBOL(rdma_get_service_id);
2254
2255 void rdma_read_gids(struct rdma_cm_id *cm_id, union ib_gid *sgid,
2256 union ib_gid *dgid)
2257 {
2258 struct rdma_addr *addr = &cm_id->route.addr;
2259
2260 if (!cm_id->device) {
2261 if (sgid)
2262 memset(sgid, 0, sizeof(*sgid));
2263 if (dgid)
2264 memset(dgid, 0, sizeof(*dgid));
2265 return;
2266 }
2267
2268 if (rdma_protocol_roce(cm_id->device, cm_id->port_num)) {
2269 if (sgid)
2270 rdma_ip2gid((struct sockaddr *)&addr->src_addr, sgid);
2271 if (dgid)
2272 rdma_ip2gid((struct sockaddr *)&addr->dst_addr, dgid);
2273 } else {
2274 if (sgid)
2275 rdma_addr_get_sgid(&addr->dev_addr, sgid);
2276 if (dgid)
2277 rdma_addr_get_dgid(&addr->dev_addr, dgid);
2278 }
2279 }
2280 EXPORT_SYMBOL(rdma_read_gids);
2281
2282 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
2283 {
2284 struct rdma_id_private *id_priv = iw_id->context;
2285 struct rdma_cm_event event = {};
2286 int ret = 0;
2287 struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
2288 struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
2289
2290 mutex_lock(&id_priv->handler_mutex);
2291 if (READ_ONCE(id_priv->state) != RDMA_CM_CONNECT)
2292 goto out;
2293
2294 switch (iw_event->event) {
2295 case IW_CM_EVENT_CLOSE:
2296 event.event = RDMA_CM_EVENT_DISCONNECTED;
2297 break;
2298 case IW_CM_EVENT_CONNECT_REPLY:
2299 memcpy(cma_src_addr(id_priv), laddr,
2300 rdma_addr_size(laddr));
2301 memcpy(cma_dst_addr(id_priv), raddr,
2302 rdma_addr_size(raddr));
2303 switch (iw_event->status) {
2304 case 0:
2305 event.event = RDMA_CM_EVENT_ESTABLISHED;
2306 event.param.conn.initiator_depth = iw_event->ird;
2307 event.param.conn.responder_resources = iw_event->ord;
2308 break;
2309 case -ECONNRESET:
2310 case -ECONNREFUSED:
2311 event.event = RDMA_CM_EVENT_REJECTED;
2312 break;
2313 case -ETIMEDOUT:
2314 event.event = RDMA_CM_EVENT_UNREACHABLE;
2315 break;
2316 default:
2317 event.event = RDMA_CM_EVENT_CONNECT_ERROR;
2318 break;
2319 }
2320 break;
2321 case IW_CM_EVENT_ESTABLISHED:
2322 event.event = RDMA_CM_EVENT_ESTABLISHED;
2323 event.param.conn.initiator_depth = iw_event->ird;
2324 event.param.conn.responder_resources = iw_event->ord;
2325 break;
2326 default:
2327 goto out;
2328 }
2329
2330 event.status = iw_event->status;
2331 event.param.conn.private_data = iw_event->private_data;
2332 event.param.conn.private_data_len = iw_event->private_data_len;
2333 ret = cma_cm_event_handler(id_priv, &event);
2334 if (ret) {
2335 /* Destroy the CM ID by returning a non-zero value. */
2336 id_priv->cm_id.iw = NULL;
2337 destroy_id_handler_unlock(id_priv);
2338 return ret;
2339 }
2340
2341 out:
2342 mutex_unlock(&id_priv->handler_mutex);
2343 return ret;
2344 }
2345
2346 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
2347 struct iw_cm_event *iw_event)
2348 {
2349 struct rdma_cm_id *new_cm_id;
2350 struct rdma_id_private *listen_id, *conn_id;
2351 struct rdma_cm_event event = {};
2352 int ret = -ECONNABORTED;
2353 struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
2354 struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
2355
2356 event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
2357 event.param.conn.private_data = iw_event->private_data;
2358 event.param.conn.private_data_len = iw_event->private_data_len;
2359 event.param.conn.initiator_depth = iw_event->ird;
2360 event.param.conn.responder_resources = iw_event->ord;
2361
2362 listen_id = cm_id->context;
2363
2364 mutex_lock(&listen_id->handler_mutex);
2365 if (READ_ONCE(listen_id->state) != RDMA_CM_LISTEN)
2366 goto out;
2367
2368 /* Create a new RDMA id for the new IW CM ID */
2369 new_cm_id = __rdma_create_id(listen_id->id.route.addr.dev_addr.net,
2370 listen_id->id.event_handler,
2371 listen_id->id.context,
2372 RDMA_PS_TCP, IB_QPT_RC,
2373 listen_id->res.kern_name);
2374 if (IS_ERR(new_cm_id)) {
2375 ret = -ENOMEM;
2376 goto out;
2377 }
2378 conn_id = container_of(new_cm_id, struct rdma_id_private, id);
2379 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
2380 conn_id->state = RDMA_CM_CONNECT;
2381
2382 ret = rdma_translate_ip(laddr, &conn_id->id.route.addr.dev_addr);
2383 if (ret) {
2384 mutex_unlock(&listen_id->handler_mutex);
2385 destroy_id_handler_unlock(conn_id);
2386 return ret;
2387 }
2388
2389 ret = cma_iw_acquire_dev(conn_id, listen_id);
2390 if (ret) {
2391 mutex_unlock(&listen_id->handler_mutex);
2392 destroy_id_handler_unlock(conn_id);
2393 return ret;
2394 }
2395
2396 conn_id->cm_id.iw = cm_id;
2397 cm_id->context = conn_id;
2398 cm_id->cm_handler = cma_iw_handler;
2399
2400 memcpy(cma_src_addr(conn_id), laddr, rdma_addr_size(laddr));
2401 memcpy(cma_dst_addr(conn_id), raddr, rdma_addr_size(raddr));
2402
2403 ret = cma_cm_event_handler(conn_id, &event);
2404 if (ret) {
2405 /* User wants to destroy the CM ID */
2406 conn_id->cm_id.iw = NULL;
2407 mutex_unlock(&listen_id->handler_mutex);
2408 destroy_id_handler_unlock(conn_id);
2409 return ret;
2410 }
2411
2412 mutex_unlock(&conn_id->handler_mutex);
2413
2414 out:
2415 mutex_unlock(&listen_id->handler_mutex);
2416 return ret;
2417 }
2418
2419 static int cma_ib_listen(struct rdma_id_private *id_priv)
2420 {
2421 struct sockaddr *addr;
2422 struct ib_cm_id *id;
2423 __be64 svc_id;
2424
2425 addr = cma_src_addr(id_priv);
2426 svc_id = rdma_get_service_id(&id_priv->id, addr);
2427 id = ib_cm_insert_listen(id_priv->id.device,
2428 cma_ib_req_handler, svc_id);
2429 if (IS_ERR(id))
2430 return PTR_ERR(id);
2431 id_priv->cm_id.ib = id;
2432
2433 return 0;
2434 }
2435
2436 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
2437 {
2438 int ret;
2439 struct iw_cm_id *id;
2440
2441 id = iw_create_cm_id(id_priv->id.device,
2442 iw_conn_req_handler,
2443 id_priv);
2444 if (IS_ERR(id))
2445 return PTR_ERR(id);
2446
2447 id->tos = id_priv->tos;
2448 id->tos_set = id_priv->tos_set;
2449 id_priv->cm_id.iw = id;
2450
2451 memcpy(&id_priv->cm_id.iw->local_addr, cma_src_addr(id_priv),
2452 rdma_addr_size(cma_src_addr(id_priv)));
2453
2454 ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
2455
2456 if (ret) {
2457 iw_destroy_cm_id(id_priv->cm_id.iw);
2458 id_priv->cm_id.iw = NULL;
2459 }
2460
2461 return ret;
2462 }
2463
2464 static int cma_listen_handler(struct rdma_cm_id *id,
2465 struct rdma_cm_event *event)
2466 {
2467 struct rdma_id_private *id_priv = id->context;
2468
2469 /* Listening IDs are always destroyed on removal */
2470 if (event->event == RDMA_CM_EVENT_DEVICE_REMOVAL)
2471 return -1;
2472
2473 id->context = id_priv->id.context;
2474 id->event_handler = id_priv->id.event_handler;
2475 trace_cm_event_handler(id_priv, event);
2476 return id_priv->id.event_handler(id, event);
2477 }
2478
2479 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
2480 struct cma_device *cma_dev)
2481 {
2482 struct rdma_id_private *dev_id_priv;
2483 struct rdma_cm_id *id;
2484 struct net *net = id_priv->id.route.addr.dev_addr.net;
2485 int ret;
2486
2487 lockdep_assert_held(&lock);
2488
2489 if (cma_family(id_priv) == AF_IB && !rdma_cap_ib_cm(cma_dev->device, 1))
2490 return;
2491
2492 id = __rdma_create_id(net, cma_listen_handler, id_priv, id_priv->id.ps,
2493 id_priv->id.qp_type, id_priv->res.kern_name);
2494 if (IS_ERR(id))
2495 return;
2496
2497 dev_id_priv = container_of(id, struct rdma_id_private, id);
2498
2499 dev_id_priv->state = RDMA_CM_ADDR_BOUND;
2500 memcpy(cma_src_addr(dev_id_priv), cma_src_addr(id_priv),
2501 rdma_addr_size(cma_src_addr(id_priv)));
2502
2503 _cma_attach_to_dev(dev_id_priv, cma_dev);
2504 list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
2505 cma_id_get(id_priv);
2506 dev_id_priv->internal_id = 1;
2507 dev_id_priv->afonly = id_priv->afonly;
2508 dev_id_priv->tos_set = id_priv->tos_set;
2509 dev_id_priv->tos = id_priv->tos;
2510
2511 ret = rdma_listen(id, id_priv->backlog);
2512 if (ret)
2513 dev_warn(&cma_dev->device->dev,
2514 "RDMA CMA: cma_listen_on_dev, error %d\n", ret);
2515 }
2516
2517 static void cma_listen_on_all(struct rdma_id_private *id_priv)
2518 {
2519 struct cma_device *cma_dev;
2520
2521 mutex_lock(&lock);
2522 list_add_tail(&id_priv->list, &listen_any_list);
2523 list_for_each_entry(cma_dev, &dev_list, list)
2524 cma_listen_on_dev(id_priv, cma_dev);
2525 mutex_unlock(&lock);
2526 }
2527
2528 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
2529 {
2530 struct rdma_id_private *id_priv;
2531
2532 id_priv = container_of(id, struct rdma_id_private, id);
2533 id_priv->tos = (u8) tos;
2534 id_priv->tos_set = true;
2535 }
2536 EXPORT_SYMBOL(rdma_set_service_type);
2537
2538 /**
2539 * rdma_set_ack_timeout() - Set the ack timeout of QP associated
2540 * with a connection identifier.
2541 * @id: Communication identifier to associated with service type.
2542 * @timeout: Ack timeout to set a QP, expressed as 4.096 * 2^(timeout) usec.
2543 *
2544 * This function should be called before rdma_connect() on active side,
2545 * and on passive side before rdma_accept(). It is applicable to primary
2546 * path only. The timeout will affect the local side of the QP, it is not
2547 * negotiated with remote side and zero disables the timer. In case it is
2548 * set before rdma_resolve_route, the value will also be used to determine
2549 * PacketLifeTime for RoCE.
2550 *
2551 * Return: 0 for success
2552 */
2553 int rdma_set_ack_timeout(struct rdma_cm_id *id, u8 timeout)
2554 {
2555 struct rdma_id_private *id_priv;
2556
2557 if (id->qp_type != IB_QPT_RC)
2558 return -EINVAL;
2559
2560 id_priv = container_of(id, struct rdma_id_private, id);
2561 id_priv->timeout = timeout;
2562 id_priv->timeout_set = true;
2563
2564 return 0;
2565 }
2566 EXPORT_SYMBOL(rdma_set_ack_timeout);
2567
2568 static void cma_query_handler(int status, struct sa_path_rec *path_rec,
2569 void *context)
2570 {
2571 struct cma_work *work = context;
2572 struct rdma_route *route;
2573
2574 route = &work->id->id.route;
2575
2576 if (!status) {
2577 route->num_paths = 1;
2578 *route->path_rec = *path_rec;
2579 } else {
2580 work->old_state = RDMA_CM_ROUTE_QUERY;
2581 work->new_state = RDMA_CM_ADDR_RESOLVED;
2582 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
2583 work->event.status = status;
2584 pr_debug_ratelimited("RDMA CM: ROUTE_ERROR: failed to query path. status %d\n",
2585 status);
2586 }
2587
2588 queue_work(cma_wq, &work->work);
2589 }
2590
2591 static int cma_query_ib_route(struct rdma_id_private *id_priv,
2592 unsigned long timeout_ms, struct cma_work *work)
2593 {
2594 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2595 struct sa_path_rec path_rec;
2596 ib_sa_comp_mask comp_mask;
2597 struct sockaddr_in6 *sin6;
2598 struct sockaddr_ib *sib;
2599
2600 memset(&path_rec, 0, sizeof path_rec);
2601
2602 if (rdma_cap_opa_ah(id_priv->id.device, id_priv->id.port_num))
2603 path_rec.rec_type = SA_PATH_REC_TYPE_OPA;
2604 else
2605 path_rec.rec_type = SA_PATH_REC_TYPE_IB;
2606 rdma_addr_get_sgid(dev_addr, &path_rec.sgid);
2607 rdma_addr_get_dgid(dev_addr, &path_rec.dgid);
2608 path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
2609 path_rec.numb_path = 1;
2610 path_rec.reversible = 1;
2611 path_rec.service_id = rdma_get_service_id(&id_priv->id,
2612 cma_dst_addr(id_priv));
2613
2614 comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
2615 IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
2616 IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
2617
2618 switch (cma_family(id_priv)) {
2619 case AF_INET:
2620 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
2621 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
2622 break;
2623 case AF_INET6:
2624 sin6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
2625 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
2626 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2627 break;
2628 case AF_IB:
2629 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
2630 path_rec.traffic_class = (u8) (be32_to_cpu(sib->sib_flowinfo) >> 20);
2631 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2632 break;
2633 }
2634
2635 id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
2636 id_priv->id.port_num, &path_rec,
2637 comp_mask, timeout_ms,
2638 GFP_KERNEL, cma_query_handler,
2639 work, &id_priv->query);
2640
2641 return (id_priv->query_id < 0) ? id_priv->query_id : 0;
2642 }
2643
2644 static void cma_work_handler(struct work_struct *_work)
2645 {
2646 struct cma_work *work = container_of(_work, struct cma_work, work);
2647 struct rdma_id_private *id_priv = work->id;
2648
2649 mutex_lock(&id_priv->handler_mutex);
2650 if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
2651 goto out_unlock;
2652
2653 if (cma_cm_event_handler(id_priv, &work->event)) {
2654 cma_id_put(id_priv);
2655 destroy_id_handler_unlock(id_priv);
2656 goto out_free;
2657 }
2658
2659 out_unlock:
2660 mutex_unlock(&id_priv->handler_mutex);
2661 cma_id_put(id_priv);
2662 out_free:
2663 kfree(work);
2664 }
2665
2666 static void cma_ndev_work_handler(struct work_struct *_work)
2667 {
2668 struct cma_ndev_work *work = container_of(_work, struct cma_ndev_work, work);
2669 struct rdma_id_private *id_priv = work->id;
2670
2671 mutex_lock(&id_priv->handler_mutex);
2672 if (id_priv->state == RDMA_CM_DESTROYING ||
2673 id_priv->state == RDMA_CM_DEVICE_REMOVAL)
2674 goto out_unlock;
2675
2676 if (cma_cm_event_handler(id_priv, &work->event)) {
2677 cma_id_put(id_priv);
2678 destroy_id_handler_unlock(id_priv);
2679 goto out_free;
2680 }
2681
2682 out_unlock:
2683 mutex_unlock(&id_priv->handler_mutex);
2684 cma_id_put(id_priv);
2685 out_free:
2686 kfree(work);
2687 }
2688
2689 static void cma_init_resolve_route_work(struct cma_work *work,
2690 struct rdma_id_private *id_priv)
2691 {
2692 work->id = id_priv;
2693 INIT_WORK(&work->work, cma_work_handler);
2694 work->old_state = RDMA_CM_ROUTE_QUERY;
2695 work->new_state = RDMA_CM_ROUTE_RESOLVED;
2696 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
2697 }
2698
2699 static void enqueue_resolve_addr_work(struct cma_work *work,
2700 struct rdma_id_private *id_priv)
2701 {
2702 /* Balances with cma_id_put() in cma_work_handler */
2703 cma_id_get(id_priv);
2704
2705 work->id = id_priv;
2706 INIT_WORK(&work->work, cma_work_handler);
2707 work->old_state = RDMA_CM_ADDR_QUERY;
2708 work->new_state = RDMA_CM_ADDR_RESOLVED;
2709 work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2710
2711 queue_work(cma_wq, &work->work);
2712 }
2713
2714 static int cma_resolve_ib_route(struct rdma_id_private *id_priv,
2715 unsigned long timeout_ms)
2716 {
2717 struct rdma_route *route = &id_priv->id.route;
2718 struct cma_work *work;
2719 int ret;
2720
2721 work = kzalloc(sizeof *work, GFP_KERNEL);
2722 if (!work)
2723 return -ENOMEM;
2724
2725 cma_init_resolve_route_work(work, id_priv);
2726
2727 route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
2728 if (!route->path_rec) {
2729 ret = -ENOMEM;
2730 goto err1;
2731 }
2732
2733 ret = cma_query_ib_route(id_priv, timeout_ms, work);
2734 if (ret)
2735 goto err2;
2736
2737 return 0;
2738 err2:
2739 kfree(route->path_rec);
2740 route->path_rec = NULL;
2741 err1:
2742 kfree(work);
2743 return ret;
2744 }
2745
2746 static enum ib_gid_type cma_route_gid_type(enum rdma_network_type network_type,
2747 unsigned long supported_gids,
2748 enum ib_gid_type default_gid)
2749 {
2750 if ((network_type == RDMA_NETWORK_IPV4 ||
2751 network_type == RDMA_NETWORK_IPV6) &&
2752 test_bit(IB_GID_TYPE_ROCE_UDP_ENCAP, &supported_gids))
2753 return IB_GID_TYPE_ROCE_UDP_ENCAP;
2754
2755 return default_gid;
2756 }
2757
2758 /*
2759 * cma_iboe_set_path_rec_l2_fields() is helper function which sets
2760 * path record type based on GID type.
2761 * It also sets up other L2 fields which includes destination mac address
2762 * netdev ifindex, of the path record.
2763 * It returns the netdev of the bound interface for this path record entry.
2764 */
2765 static struct net_device *
2766 cma_iboe_set_path_rec_l2_fields(struct rdma_id_private *id_priv)
2767 {
2768 struct rdma_route *route = &id_priv->id.route;
2769 enum ib_gid_type gid_type = IB_GID_TYPE_ROCE;
2770 struct rdma_addr *addr = &route->addr;
2771 unsigned long supported_gids;
2772 struct net_device *ndev;
2773
2774 if (!addr->dev_addr.bound_dev_if)
2775 return NULL;
2776
2777 ndev = dev_get_by_index(addr->dev_addr.net,
2778 addr->dev_addr.bound_dev_if);
2779 if (!ndev)
2780 return NULL;
2781
2782 supported_gids = roce_gid_type_mask_support(id_priv->id.device,
2783 id_priv->id.port_num);
2784 gid_type = cma_route_gid_type(addr->dev_addr.network,
2785 supported_gids,
2786 id_priv->gid_type);
2787 /* Use the hint from IP Stack to select GID Type */
2788 if (gid_type < ib_network_to_gid_type(addr->dev_addr.network))
2789 gid_type = ib_network_to_gid_type(addr->dev_addr.network);
2790 route->path_rec->rec_type = sa_conv_gid_to_pathrec_type(gid_type);
2791
2792 route->path_rec->roce.route_resolved = true;
2793 sa_path_set_dmac(route->path_rec, addr->dev_addr.dst_dev_addr);
2794 return ndev;
2795 }
2796
2797 int rdma_set_ib_path(struct rdma_cm_id *id,
2798 struct sa_path_rec *path_rec)
2799 {
2800 struct rdma_id_private *id_priv;
2801 struct net_device *ndev;
2802 int ret;
2803
2804 id_priv = container_of(id, struct rdma_id_private, id);
2805 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
2806 RDMA_CM_ROUTE_RESOLVED))
2807 return -EINVAL;
2808
2809 id->route.path_rec = kmemdup(path_rec, sizeof(*path_rec),
2810 GFP_KERNEL);
2811 if (!id->route.path_rec) {
2812 ret = -ENOMEM;
2813 goto err;
2814 }
2815
2816 if (rdma_protocol_roce(id->device, id->port_num)) {
2817 ndev = cma_iboe_set_path_rec_l2_fields(id_priv);
2818 if (!ndev) {
2819 ret = -ENODEV;
2820 goto err_free;
2821 }
2822 dev_put(ndev);
2823 }
2824
2825 id->route.num_paths = 1;
2826 return 0;
2827
2828 err_free:
2829 kfree(id->route.path_rec);
2830 id->route.path_rec = NULL;
2831 err:
2832 cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED);
2833 return ret;
2834 }
2835 EXPORT_SYMBOL(rdma_set_ib_path);
2836
2837 static int cma_resolve_iw_route(struct rdma_id_private *id_priv)
2838 {
2839 struct cma_work *work;
2840
2841 work = kzalloc(sizeof *work, GFP_KERNEL);
2842 if (!work)
2843 return -ENOMEM;
2844
2845 cma_init_resolve_route_work(work, id_priv);
2846 queue_work(cma_wq, &work->work);
2847 return 0;
2848 }
2849
2850 static int get_vlan_ndev_tc(struct net_device *vlan_ndev, int prio)
2851 {
2852 struct net_device *dev;
2853
2854 dev = vlan_dev_real_dev(vlan_ndev);
2855 if (dev->num_tc)
2856 return netdev_get_prio_tc_map(dev, prio);
2857
2858 return (vlan_dev_get_egress_qos_mask(vlan_ndev, prio) &
2859 VLAN_PRIO_MASK) >> VLAN_PRIO_SHIFT;
2860 }
2861
2862 struct iboe_prio_tc_map {
2863 int input_prio;
2864 int output_tc;
2865 bool found;
2866 };
2867
2868 static int get_lower_vlan_dev_tc(struct net_device *dev, void *data)
2869 {
2870 struct iboe_prio_tc_map *map = data;
2871
2872 if (is_vlan_dev(dev))
2873 map->output_tc = get_vlan_ndev_tc(dev, map->input_prio);
2874 else if (dev->num_tc)
2875 map->output_tc = netdev_get_prio_tc_map(dev, map->input_prio);
2876 else
2877 map->output_tc = 0;
2878 /* We are interested only in first level VLAN device, so always
2879 * return 1 to stop iterating over next level devices.
2880 */
2881 map->found = true;
2882 return 1;
2883 }
2884
2885 static int iboe_tos_to_sl(struct net_device *ndev, int tos)
2886 {
2887 struct iboe_prio_tc_map prio_tc_map = {};
2888 int prio = rt_tos2priority(tos);
2889
2890 /* If VLAN device, get it directly from the VLAN netdev */
2891 if (is_vlan_dev(ndev))
2892 return get_vlan_ndev_tc(ndev, prio);
2893
2894 prio_tc_map.input_prio = prio;
2895 rcu_read_lock();
2896 netdev_walk_all_lower_dev_rcu(ndev,
2897 get_lower_vlan_dev_tc,
2898 &prio_tc_map);
2899 rcu_read_unlock();
2900 /* If map is found from lower device, use it; Otherwise
2901 * continue with the current netdevice to get priority to tc map.
2902 */
2903 if (prio_tc_map.found)
2904 return prio_tc_map.output_tc;
2905 else if (ndev->num_tc)
2906 return netdev_get_prio_tc_map(ndev, prio);
2907 else
2908 return 0;
2909 }
2910
2911 static __be32 cma_get_roce_udp_flow_label(struct rdma_id_private *id_priv)
2912 {
2913 struct sockaddr_in6 *addr6;
2914 u16 dport, sport;
2915 u32 hash, fl;
2916
2917 addr6 = (struct sockaddr_in6 *)cma_src_addr(id_priv);
2918 fl = be32_to_cpu(addr6->sin6_flowinfo) & IB_GRH_FLOWLABEL_MASK;
2919 if ((cma_family(id_priv) != AF_INET6) || !fl) {
2920 dport = be16_to_cpu(cma_port(cma_dst_addr(id_priv)));
2921 sport = be16_to_cpu(cma_port(cma_src_addr(id_priv)));
2922 hash = (u32)sport * 31 + dport;
2923 fl = hash & IB_GRH_FLOWLABEL_MASK;
2924 }
2925
2926 return cpu_to_be32(fl);
2927 }
2928
2929 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
2930 {
2931 struct rdma_route *route = &id_priv->id.route;
2932 struct rdma_addr *addr = &route->addr;
2933 struct cma_work *work;
2934 int ret;
2935 struct net_device *ndev;
2936
2937 u8 default_roce_tos = id_priv->cma_dev->default_roce_tos[id_priv->id.port_num -
2938 rdma_start_port(id_priv->cma_dev->device)];
2939 u8 tos = id_priv->tos_set ? id_priv->tos : default_roce_tos;
2940
2941
2942 work = kzalloc(sizeof *work, GFP_KERNEL);
2943 if (!work)
2944 return -ENOMEM;
2945
2946 route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
2947 if (!route->path_rec) {
2948 ret = -ENOMEM;
2949 goto err1;
2950 }
2951
2952 route->num_paths = 1;
2953
2954 ndev = cma_iboe_set_path_rec_l2_fields(id_priv);
2955 if (!ndev) {
2956 ret = -ENODEV;
2957 goto err2;
2958 }
2959
2960 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
2961 &route->path_rec->sgid);
2962 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.dst_addr,
2963 &route->path_rec->dgid);
2964
2965 if (((struct sockaddr *)&id_priv->id.route.addr.dst_addr)->sa_family != AF_IB)
2966 /* TODO: get the hoplimit from the inet/inet6 device */
2967 route->path_rec->hop_limit = addr->dev_addr.hoplimit;
2968 else
2969 route->path_rec->hop_limit = 1;
2970 route->path_rec->reversible = 1;
2971 route->path_rec->pkey = cpu_to_be16(0xffff);
2972 route->path_rec->mtu_selector = IB_SA_EQ;
2973 route->path_rec->sl = iboe_tos_to_sl(ndev, tos);
2974 route->path_rec->traffic_class = tos;
2975 route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
2976 route->path_rec->rate_selector = IB_SA_EQ;
2977 route->path_rec->rate = iboe_get_rate(ndev);
2978 dev_put(ndev);
2979 route->path_rec->packet_life_time_selector = IB_SA_EQ;
2980 /* In case ACK timeout is set, use this value to calculate
2981 * PacketLifeTime. As per IBTA 12.7.34,
2982 * local ACK timeout = (2 * PacketLifeTime + Local CA’s ACK delay).
2983 * Assuming a negligible local ACK delay, we can use
2984 * PacketLifeTime = local ACK timeout/2
2985 * as a reasonable approximation for RoCE networks.
2986 */
2987 route->path_rec->packet_life_time = id_priv->timeout_set ?
2988 id_priv->timeout - 1 : CMA_IBOE_PACKET_LIFETIME;
2989
2990 if (!route->path_rec->mtu) {
2991 ret = -EINVAL;
2992 goto err2;
2993 }
2994
2995 if (rdma_protocol_roce_udp_encap(id_priv->id.device,
2996 id_priv->id.port_num))
2997 route->path_rec->flow_label =
2998 cma_get_roce_udp_flow_label(id_priv);
2999
3000 cma_init_resolve_route_work(work, id_priv);
3001 queue_work(cma_wq, &work->work);
3002
3003 return 0;
3004
3005 err2:
3006 kfree(route->path_rec);
3007 route->path_rec = NULL;
3008 route->num_paths = 0;
3009 err1:
3010 kfree(work);
3011 return ret;
3012 }
3013
3014 int rdma_resolve_route(struct rdma_cm_id *id, unsigned long timeout_ms)
3015 {
3016 struct rdma_id_private *id_priv;
3017 int ret;
3018
3019 id_priv = container_of(id, struct rdma_id_private, id);
3020 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY))
3021 return -EINVAL;
3022
3023 cma_id_get(id_priv);
3024 if (rdma_cap_ib_sa(id->device, id->port_num))
3025 ret = cma_resolve_ib_route(id_priv, timeout_ms);
3026 else if (rdma_protocol_roce(id->device, id->port_num))
3027 ret = cma_resolve_iboe_route(id_priv);
3028 else if (rdma_protocol_iwarp(id->device, id->port_num))
3029 ret = cma_resolve_iw_route(id_priv);
3030 else
3031 ret = -ENOSYS;
3032
3033 if (ret)
3034 goto err;
3035
3036 return 0;
3037 err:
3038 cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED);
3039 cma_id_put(id_priv);
3040 return ret;
3041 }
3042 EXPORT_SYMBOL(rdma_resolve_route);
3043
3044 static void cma_set_loopback(struct sockaddr *addr)
3045 {
3046 switch (addr->sa_family) {
3047 case AF_INET:
3048 ((struct sockaddr_in *) addr)->sin_addr.s_addr = htonl(INADDR_LOOPBACK);
3049 break;
3050 case AF_INET6:
3051 ipv6_addr_set(&((struct sockaddr_in6 *) addr)->sin6_addr,
3052 0, 0, 0, htonl(1));
3053 break;
3054 default:
3055 ib_addr_set(&((struct sockaddr_ib *) addr)->sib_addr,
3056 0, 0, 0, htonl(1));
3057 break;
3058 }
3059 }
3060
3061 static int cma_bind_loopback(struct rdma_id_private *id_priv)
3062 {
3063 struct cma_device *cma_dev, *cur_dev;
3064 union ib_gid gid;
3065 enum ib_port_state port_state;
3066 unsigned int p;
3067 u16 pkey;
3068 int ret;
3069
3070 cma_dev = NULL;
3071 mutex_lock(&lock);
3072 list_for_each_entry(cur_dev, &dev_list, list) {
3073 if (cma_family(id_priv) == AF_IB &&
3074 !rdma_cap_ib_cm(cur_dev->device, 1))
3075 continue;
3076
3077 if (!cma_dev)
3078 cma_dev = cur_dev;
3079
3080 rdma_for_each_port (cur_dev->device, p) {
3081 if (!ib_get_cached_port_state(cur_dev->device, p, &port_state) &&
3082 port_state == IB_PORT_ACTIVE) {
3083 cma_dev = cur_dev;
3084 goto port_found;
3085 }
3086 }
3087 }
3088
3089 if (!cma_dev) {
3090 ret = -ENODEV;
3091 goto out;
3092 }
3093
3094 p = 1;
3095
3096 port_found:
3097 ret = rdma_query_gid(cma_dev->device, p, 0, &gid);
3098 if (ret)
3099 goto out;
3100
3101 ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
3102 if (ret)
3103 goto out;
3104
3105 id_priv->id.route.addr.dev_addr.dev_type =
3106 (rdma_protocol_ib(cma_dev->device, p)) ?
3107 ARPHRD_INFINIBAND : ARPHRD_ETHER;
3108
3109 rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
3110 ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
3111 id_priv->id.port_num = p;
3112 cma_attach_to_dev(id_priv, cma_dev);
3113 cma_set_loopback(cma_src_addr(id_priv));
3114 out:
3115 mutex_unlock(&lock);
3116 return ret;
3117 }
3118
3119 static void addr_handler(int status, struct sockaddr *src_addr,
3120 struct rdma_dev_addr *dev_addr, void *context)
3121 {
3122 struct rdma_id_private *id_priv = context;
3123 struct rdma_cm_event event = {};
3124 struct sockaddr *addr;
3125 struct sockaddr_storage old_addr;
3126
3127 mutex_lock(&id_priv->handler_mutex);
3128 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY,
3129 RDMA_CM_ADDR_RESOLVED))
3130 goto out;
3131
3132 /*
3133 * Store the previous src address, so that if we fail to acquire
3134 * matching rdma device, old address can be restored back, which helps
3135 * to cancel the cma listen operation correctly.
3136 */
3137 addr = cma_src_addr(id_priv);
3138 memcpy(&old_addr, addr, rdma_addr_size(addr));
3139 memcpy(addr, src_addr, rdma_addr_size(src_addr));
3140 if (!status && !id_priv->cma_dev) {
3141 status = cma_acquire_dev_by_src_ip(id_priv);
3142 if (status)
3143 pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to acquire device. status %d\n",
3144 status);
3145 } else if (status) {
3146 pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to resolve IP. status %d\n", status);
3147 }
3148
3149 if (status) {
3150 memcpy(addr, &old_addr,
3151 rdma_addr_size((struct sockaddr *)&old_addr));
3152 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
3153 RDMA_CM_ADDR_BOUND))
3154 goto out;
3155 event.event = RDMA_CM_EVENT_ADDR_ERROR;
3156 event.status = status;
3157 } else
3158 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
3159
3160 if (cma_cm_event_handler(id_priv, &event)) {
3161 destroy_id_handler_unlock(id_priv);
3162 return;
3163 }
3164 out:
3165 mutex_unlock(&id_priv->handler_mutex);
3166 }
3167
3168 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
3169 {
3170 struct cma_work *work;
3171 union ib_gid gid;
3172 int ret;
3173
3174 work = kzalloc(sizeof *work, GFP_KERNEL);
3175 if (!work)
3176 return -ENOMEM;
3177
3178 if (!id_priv->cma_dev) {
3179 ret = cma_bind_loopback(id_priv);
3180 if (ret)
3181 goto err;
3182 }
3183
3184 rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
3185 rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
3186
3187 enqueue_resolve_addr_work(work, id_priv);
3188 return 0;
3189 err:
3190 kfree(work);
3191 return ret;
3192 }
3193
3194 static int cma_resolve_ib_addr(struct rdma_id_private *id_priv)
3195 {
3196 struct cma_work *work;
3197 int ret;
3198
3199 work = kzalloc(sizeof *work, GFP_KERNEL);
3200 if (!work)
3201 return -ENOMEM;
3202
3203 if (!id_priv->cma_dev) {
3204 ret = cma_resolve_ib_dev(id_priv);
3205 if (ret)
3206 goto err;
3207 }
3208
3209 rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, (union ib_gid *)
3210 &(((struct sockaddr_ib *) &id_priv->id.route.addr.dst_addr)->sib_addr));
3211
3212 enqueue_resolve_addr_work(work, id_priv);
3213 return 0;
3214 err:
3215 kfree(work);
3216 return ret;
3217 }
3218
3219 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
3220 const struct sockaddr *dst_addr)
3221 {
3222 if (!src_addr || !src_addr->sa_family) {
3223 src_addr = (struct sockaddr *) &id->route.addr.src_addr;
3224 src_addr->sa_family = dst_addr->sa_family;
3225 if (IS_ENABLED(CONFIG_IPV6) &&
3226 dst_addr->sa_family == AF_INET6) {
3227 struct sockaddr_in6 *src_addr6 = (struct sockaddr_in6 *) src_addr;
3228 struct sockaddr_in6 *dst_addr6 = (struct sockaddr_in6 *) dst_addr;
3229 src_addr6->sin6_scope_id = dst_addr6->sin6_scope_id;
3230 if (ipv6_addr_type(&dst_addr6->sin6_addr) & IPV6_ADDR_LINKLOCAL)
3231 id->route.addr.dev_addr.bound_dev_if = dst_addr6->sin6_scope_id;
3232 } else if (dst_addr->sa_family == AF_IB) {
3233 ((struct sockaddr_ib *) src_addr)->sib_pkey =
3234 ((struct sockaddr_ib *) dst_addr)->sib_pkey;
3235 }
3236 }
3237 return rdma_bind_addr(id, src_addr);
3238 }
3239
3240 /*
3241 * If required, resolve the source address for bind and leave the id_priv in
3242 * state RDMA_CM_ADDR_BOUND. This oddly uses the state to determine the prior
3243 * calls made by ULP, a previously bound ID will not be re-bound and src_addr is
3244 * ignored.
3245 */
3246 static int resolve_prepare_src(struct rdma_id_private *id_priv,
3247 struct sockaddr *src_addr,
3248 const struct sockaddr *dst_addr)
3249 {
3250 int ret;
3251
3252 memcpy(cma_dst_addr(id_priv), dst_addr, rdma_addr_size(dst_addr));
3253 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY)) {
3254 /* For a well behaved ULP state will be RDMA_CM_IDLE */
3255 ret = cma_bind_addr(&id_priv->id, src_addr, dst_addr);
3256 if (ret)
3257 goto err_dst;
3258 if (WARN_ON(!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND,
3259 RDMA_CM_ADDR_QUERY))) {
3260 ret = -EINVAL;
3261 goto err_dst;
3262 }
3263 }
3264
3265 if (cma_family(id_priv) != dst_addr->sa_family) {
3266 ret = -EINVAL;
3267 goto err_state;
3268 }
3269 return 0;
3270
3271 err_state:
3272 cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
3273 err_dst:
3274 memset(cma_dst_addr(id_priv), 0, rdma_addr_size(dst_addr));
3275 return ret;
3276 }
3277
3278 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
3279 const struct sockaddr *dst_addr, unsigned long timeout_ms)
3280 {
3281 struct rdma_id_private *id_priv =
3282 container_of(id, struct rdma_id_private, id);
3283 int ret;
3284
3285 ret = resolve_prepare_src(id_priv, src_addr, dst_addr);
3286 if (ret)
3287 return ret;
3288
3289 if (cma_any_addr(dst_addr)) {
3290 ret = cma_resolve_loopback(id_priv);
3291 } else {
3292 if (dst_addr->sa_family == AF_IB) {
3293 ret = cma_resolve_ib_addr(id_priv);
3294 } else {
3295 ret = rdma_resolve_ip(cma_src_addr(id_priv), dst_addr,
3296 &id->route.addr.dev_addr,
3297 timeout_ms, addr_handler,
3298 false, id_priv);
3299 }
3300 }
3301 if (ret)
3302 goto err;
3303
3304 return 0;
3305 err:
3306 cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
3307 return ret;
3308 }
3309 EXPORT_SYMBOL(rdma_resolve_addr);
3310
3311 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse)
3312 {
3313 struct rdma_id_private *id_priv;
3314 unsigned long flags;
3315 int ret;
3316
3317 id_priv = container_of(id, struct rdma_id_private, id);
3318 spin_lock_irqsave(&id_priv->lock, flags);
3319 if ((reuse && id_priv->state != RDMA_CM_LISTEN) ||
3320 id_priv->state == RDMA_CM_IDLE) {
3321 id_priv->reuseaddr = reuse;
3322 ret = 0;
3323 } else {
3324 ret = -EINVAL;
3325 }
3326 spin_unlock_irqrestore(&id_priv->lock, flags);
3327 return ret;
3328 }
3329 EXPORT_SYMBOL(rdma_set_reuseaddr);
3330
3331 int rdma_set_afonly(struct rdma_cm_id *id, int afonly)
3332 {
3333 struct rdma_id_private *id_priv;
3334 unsigned long flags;
3335 int ret;
3336
3337 id_priv = container_of(id, struct rdma_id_private, id);
3338 spin_lock_irqsave(&id_priv->lock, flags);
3339 if (id_priv->state == RDMA_CM_IDLE || id_priv->state == RDMA_CM_ADDR_BOUND) {
3340 id_priv->options |= (1 << CMA_OPTION_AFONLY);
3341 id_priv->afonly = afonly;
3342 ret = 0;
3343 } else {
3344 ret = -EINVAL;
3345 }
3346 spin_unlock_irqrestore(&id_priv->lock, flags);
3347 return ret;
3348 }
3349 EXPORT_SYMBOL(rdma_set_afonly);
3350
3351 static void cma_bind_port(struct rdma_bind_list *bind_list,
3352 struct rdma_id_private *id_priv)
3353 {
3354 struct sockaddr *addr;
3355 struct sockaddr_ib *sib;
3356 u64 sid, mask;
3357 __be16 port;
3358
3359 lockdep_assert_held(&lock);
3360
3361 addr = cma_src_addr(id_priv);
3362 port = htons(bind_list->port);
3363
3364 switch (addr->sa_family) {
3365 case AF_INET:
3366 ((struct sockaddr_in *) addr)->sin_port = port;
3367 break;
3368 case AF_INET6:
3369 ((struct sockaddr_in6 *) addr)->sin6_port = port;
3370 break;
3371 case AF_IB:
3372 sib = (struct sockaddr_ib *) addr;
3373 sid = be64_to_cpu(sib->sib_sid);
3374 mask = be64_to_cpu(sib->sib_sid_mask);
3375 sib->sib_sid = cpu_to_be64((sid & mask) | (u64) ntohs(port));
3376 sib->sib_sid_mask = cpu_to_be64(~0ULL);
3377 break;
3378 }
3379 id_priv->bind_list = bind_list;
3380 hlist_add_head(&id_priv->node, &bind_list->owners);
3381 }
3382
3383 static int cma_alloc_port(enum rdma_ucm_port_space ps,
3384 struct rdma_id_private *id_priv, unsigned short snum)
3385 {
3386 struct rdma_bind_list *bind_list;
3387 int ret;
3388
3389 lockdep_assert_held(&lock);
3390
3391 bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
3392 if (!bind_list)
3393 return -ENOMEM;
3394
3395 ret = cma_ps_alloc(id_priv->id.route.addr.dev_addr.net, ps, bind_list,
3396 snum);
3397 if (ret < 0)
3398 goto err;
3399
3400 bind_list->ps = ps;
3401 bind_list->port = snum;
3402 cma_bind_port(bind_list, id_priv);
3403 return 0;
3404 err:
3405 kfree(bind_list);
3406 return ret == -ENOSPC ? -EADDRNOTAVAIL : ret;
3407 }
3408
3409 static int cma_port_is_unique(struct rdma_bind_list *bind_list,
3410 struct rdma_id_private *id_priv)
3411 {
3412 struct rdma_id_private *cur_id;
3413 struct sockaddr *daddr = cma_dst_addr(id_priv);
3414 struct sockaddr *saddr = cma_src_addr(id_priv);
3415 __be16 dport = cma_port(daddr);
3416
3417 lockdep_assert_held(&lock);
3418
3419 hlist_for_each_entry(cur_id, &bind_list->owners, node) {
3420 struct sockaddr *cur_daddr = cma_dst_addr(cur_id);
3421 struct sockaddr *cur_saddr = cma_src_addr(cur_id);
3422 __be16 cur_dport = cma_port(cur_daddr);
3423
3424 if (id_priv == cur_id)
3425 continue;
3426
3427 /* different dest port -> unique */
3428 if (!cma_any_port(daddr) &&
3429 !cma_any_port(cur_daddr) &&
3430 (dport != cur_dport))
3431 continue;
3432
3433 /* different src address -> unique */
3434 if (!cma_any_addr(saddr) &&
3435 !cma_any_addr(cur_saddr) &&
3436 cma_addr_cmp(saddr, cur_saddr))
3437 continue;
3438
3439 /* different dst address -> unique */
3440 if (!cma_any_addr(daddr) &&
3441 !cma_any_addr(cur_daddr) &&
3442 cma_addr_cmp(daddr, cur_daddr))
3443 continue;
3444
3445 return -EADDRNOTAVAIL;
3446 }
3447 return 0;
3448 }
3449
3450 static int cma_alloc_any_port(enum rdma_ucm_port_space ps,
3451 struct rdma_id_private *id_priv)
3452 {
3453 static unsigned int last_used_port;
3454 int low, high, remaining;
3455 unsigned int rover;
3456 struct net *net = id_priv->id.route.addr.dev_addr.net;
3457
3458 lockdep_assert_held(&lock);
3459
3460 inet_get_local_port_range(net, &low, &high);
3461 remaining = (high - low) + 1;
3462 rover = prandom_u32() % remaining + low;
3463 retry:
3464 if (last_used_port != rover) {
3465 struct rdma_bind_list *bind_list;
3466 int ret;
3467
3468 bind_list = cma_ps_find(net, ps, (unsigned short)rover);
3469
3470 if (!bind_list) {
3471 ret = cma_alloc_port(ps, id_priv, rover);
3472 } else {
3473 ret = cma_port_is_unique(bind_list, id_priv);
3474 if (!ret)
3475 cma_bind_port(bind_list, id_priv);
3476 }
3477 /*
3478 * Remember previously used port number in order to avoid
3479 * re-using same port immediately after it is closed.
3480 */
3481 if (!ret)
3482 last_used_port = rover;
3483 if (ret != -EADDRNOTAVAIL)
3484 return ret;
3485 }
3486 if (--remaining) {
3487 rover++;
3488 if ((rover < low) || (rover > high))
3489 rover = low;
3490 goto retry;
3491 }
3492 return -EADDRNOTAVAIL;
3493 }
3494
3495 /*
3496 * Check that the requested port is available. This is called when trying to
3497 * bind to a specific port, or when trying to listen on a bound port. In
3498 * the latter case, the provided id_priv may already be on the bind_list, but
3499 * we still need to check that it's okay to start listening.
3500 */
3501 static int cma_check_port(struct rdma_bind_list *bind_list,
3502 struct rdma_id_private *id_priv, uint8_t reuseaddr)
3503 {
3504 struct rdma_id_private *cur_id;
3505 struct sockaddr *addr, *cur_addr;
3506
3507 lockdep_assert_held(&lock);
3508
3509 addr = cma_src_addr(id_priv);
3510 hlist_for_each_entry(cur_id, &bind_list->owners, node) {
3511 if (id_priv == cur_id)
3512 continue;
3513
3514 if (reuseaddr && cur_id->reuseaddr)
3515 continue;
3516
3517 cur_addr = cma_src_addr(cur_id);
3518 if (id_priv->afonly && cur_id->afonly &&
3519 (addr->sa_family != cur_addr->sa_family))
3520 continue;
3521
3522 if (cma_any_addr(addr) || cma_any_addr(cur_addr))
3523 return -EADDRNOTAVAIL;
3524
3525 if (!cma_addr_cmp(addr, cur_addr))
3526 return -EADDRINUSE;
3527 }
3528 return 0;
3529 }
3530
3531 static int cma_use_port(enum rdma_ucm_port_space ps,
3532 struct rdma_id_private *id_priv)
3533 {
3534 struct rdma_bind_list *bind_list;
3535 unsigned short snum;
3536 int ret;
3537
3538 lockdep_assert_held(&lock);
3539
3540 snum = ntohs(cma_port(cma_src_addr(id_priv)));
3541 if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
3542 return -EACCES;
3543
3544 bind_list = cma_ps_find(id_priv->id.route.addr.dev_addr.net, ps, snum);
3545 if (!bind_list) {
3546 ret = cma_alloc_port(ps, id_priv, snum);
3547 } else {
3548 ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr);
3549 if (!ret)
3550 cma_bind_port(bind_list, id_priv);
3551 }
3552 return ret;
3553 }
3554
3555 static enum rdma_ucm_port_space
3556 cma_select_inet_ps(struct rdma_id_private *id_priv)
3557 {
3558 switch (id_priv->id.ps) {
3559 case RDMA_PS_TCP:
3560 case RDMA_PS_UDP:
3561 case RDMA_PS_IPOIB:
3562 case RDMA_PS_IB:
3563 return id_priv->id.ps;
3564 default:
3565
3566 return 0;
3567 }
3568 }
3569
3570 static enum rdma_ucm_port_space
3571 cma_select_ib_ps(struct rdma_id_private *id_priv)
3572 {
3573 enum rdma_ucm_port_space ps = 0;
3574 struct sockaddr_ib *sib;
3575 u64 sid_ps, mask, sid;
3576
3577 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
3578 mask = be64_to_cpu(sib->sib_sid_mask) & RDMA_IB_IP_PS_MASK;
3579 sid = be64_to_cpu(sib->sib_sid) & mask;
3580
3581 if ((id_priv->id.ps == RDMA_PS_IB) && (sid == (RDMA_IB_IP_PS_IB & mask))) {
3582 sid_ps = RDMA_IB_IP_PS_IB;
3583 ps = RDMA_PS_IB;
3584 } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_TCP)) &&
3585 (sid == (RDMA_IB_IP_PS_TCP & mask))) {
3586 sid_ps = RDMA_IB_IP_PS_TCP;
3587 ps = RDMA_PS_TCP;
3588 } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_UDP)) &&
3589 (sid == (RDMA_IB_IP_PS_UDP & mask))) {
3590 sid_ps = RDMA_IB_IP_PS_UDP;
3591 ps = RDMA_PS_UDP;
3592 }
3593
3594 if (ps) {
3595 sib->sib_sid = cpu_to_be64(sid_ps | ntohs(cma_port((struct sockaddr *) sib)));
3596 sib->sib_sid_mask = cpu_to_be64(RDMA_IB_IP_PS_MASK |
3597 be64_to_cpu(sib->sib_sid_mask));
3598 }
3599 return ps;
3600 }
3601
3602 static int cma_get_port(struct rdma_id_private *id_priv)
3603 {
3604 enum rdma_ucm_port_space ps;
3605 int ret;
3606
3607 if (cma_family(id_priv) != AF_IB)
3608 ps = cma_select_inet_ps(id_priv);
3609 else
3610 ps = cma_select_ib_ps(id_priv);
3611 if (!ps)
3612 return -EPROTONOSUPPORT;
3613
3614 mutex_lock(&lock);
3615 if (cma_any_port(cma_src_addr(id_priv)))
3616 ret = cma_alloc_any_port(ps, id_priv);
3617 else
3618 ret = cma_use_port(ps, id_priv);
3619 mutex_unlock(&lock);
3620
3621 return ret;
3622 }
3623
3624 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
3625 struct sockaddr *addr)
3626 {
3627 #if IS_ENABLED(CONFIG_IPV6)
3628 struct sockaddr_in6 *sin6;
3629
3630 if (addr->sa_family != AF_INET6)
3631 return 0;
3632
3633 sin6 = (struct sockaddr_in6 *) addr;
3634
3635 if (!(ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL))
3636 return 0;
3637
3638 if (!sin6->sin6_scope_id)
3639 return -EINVAL;
3640
3641 dev_addr->bound_dev_if = sin6->sin6_scope_id;
3642 #endif
3643 return 0;
3644 }
3645
3646 int rdma_listen(struct rdma_cm_id *id, int backlog)
3647 {
3648 struct rdma_id_private *id_priv =
3649 container_of(id, struct rdma_id_private, id);
3650 int ret;
3651
3652 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN)) {
3653 /* For a well behaved ULP state will be RDMA_CM_IDLE */
3654 id->route.addr.src_addr.ss_family = AF_INET;
3655 ret = rdma_bind_addr(id, cma_src_addr(id_priv));
3656 if (ret)
3657 return ret;
3658 if (WARN_ON(!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND,
3659 RDMA_CM_LISTEN)))
3660 return -EINVAL;
3661 }
3662
3663 /*
3664 * Once the ID reaches RDMA_CM_LISTEN it is not allowed to be reusable
3665 * any more, and has to be unique in the bind list.
3666 */
3667 if (id_priv->reuseaddr) {
3668 mutex_lock(&lock);
3669 ret = cma_check_port(id_priv->bind_list, id_priv, 0);
3670 if (!ret)
3671 id_priv->reuseaddr = 0;
3672 mutex_unlock(&lock);
3673 if (ret)
3674 goto err;
3675 }
3676
3677 id_priv->backlog = backlog;
3678 if (id->device) {
3679 if (rdma_cap_ib_cm(id->device, 1)) {
3680 ret = cma_ib_listen(id_priv);
3681 if (ret)
3682 goto err;
3683 } else if (rdma_cap_iw_cm(id->device, 1)) {
3684 ret = cma_iw_listen(id_priv, backlog);
3685 if (ret)
3686 goto err;
3687 } else {
3688 ret = -ENOSYS;
3689 goto err;
3690 }
3691 } else
3692 cma_listen_on_all(id_priv);
3693
3694 return 0;
3695 err:
3696 id_priv->backlog = 0;
3697 /*
3698 * All the failure paths that lead here will not allow the req_handler's
3699 * to have run.
3700 */
3701 cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND);
3702 return ret;
3703 }
3704 EXPORT_SYMBOL(rdma_listen);
3705
3706 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
3707 {
3708 struct rdma_id_private *id_priv;
3709 int ret;
3710 struct sockaddr *daddr;
3711
3712 if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6 &&
3713 addr->sa_family != AF_IB)
3714 return -EAFNOSUPPORT;
3715
3716 id_priv = container_of(id, struct rdma_id_private, id);
3717 if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND))
3718 return -EINVAL;
3719
3720 ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
3721 if (ret)
3722 goto err1;
3723
3724 memcpy(cma_src_addr(id_priv), addr, rdma_addr_size(addr));
3725 if (!cma_any_addr(addr)) {
3726 ret = cma_translate_addr(addr, &id->route.addr.dev_addr);
3727 if (ret)
3728 goto err1;
3729
3730 ret = cma_acquire_dev_by_src_ip(id_priv);
3731 if (ret)
3732 goto err1;
3733 }
3734
3735 if (!(id_priv->options & (1 << CMA_OPTION_AFONLY))) {
3736 if (addr->sa_family == AF_INET)
3737 id_priv->afonly = 1;
3738 #if IS_ENABLED(CONFIG_IPV6)
3739 else if (addr->sa_family == AF_INET6) {
3740 struct net *net = id_priv->id.route.addr.dev_addr.net;
3741
3742 id_priv->afonly = net->ipv6.sysctl.bindv6only;
3743 }
3744 #endif
3745 }
3746 daddr = cma_dst_addr(id_priv);
3747 daddr->sa_family = addr->sa_family;
3748
3749 ret = cma_get_port(id_priv);
3750 if (ret)
3751 goto err2;
3752
3753 return 0;
3754 err2:
3755 rdma_restrack_del(&id_priv->res);
3756 if (id_priv->cma_dev)
3757 cma_release_dev(id_priv);
3758 err1:
3759 cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE);
3760 return ret;
3761 }
3762 EXPORT_SYMBOL(rdma_bind_addr);
3763
3764 static int cma_format_hdr(void *hdr, struct rdma_id_private *id_priv)
3765 {
3766 struct cma_hdr *cma_hdr;
3767
3768 cma_hdr = hdr;
3769 cma_hdr->cma_version = CMA_VERSION;
3770 if (cma_family(id_priv) == AF_INET) {
3771 struct sockaddr_in *src4, *dst4;
3772
3773 src4 = (struct sockaddr_in *) cma_src_addr(id_priv);
3774 dst4 = (struct sockaddr_in *) cma_dst_addr(id_priv);
3775
3776 cma_set_ip_ver(cma_hdr, 4);
3777 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
3778 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
3779 cma_hdr->port = src4->sin_port;
3780 } else if (cma_family(id_priv) == AF_INET6) {
3781 struct sockaddr_in6 *src6, *dst6;
3782
3783 src6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
3784 dst6 = (struct sockaddr_in6 *) cma_dst_addr(id_priv);
3785
3786 cma_set_ip_ver(cma_hdr, 6);
3787 cma_hdr->src_addr.ip6 = src6->sin6_addr;
3788 cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
3789 cma_hdr->port = src6->sin6_port;
3790 }
3791 return 0;
3792 }
3793
3794 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
3795 const struct ib_cm_event *ib_event)
3796 {
3797 struct rdma_id_private *id_priv = cm_id->context;
3798 struct rdma_cm_event event = {};
3799 const struct ib_cm_sidr_rep_event_param *rep =
3800 &ib_event->param.sidr_rep_rcvd;
3801 int ret;
3802
3803 mutex_lock(&id_priv->handler_mutex);
3804 if (READ_ONCE(id_priv->state) != RDMA_CM_CONNECT)
3805 goto out;
3806
3807 switch (ib_event->event) {
3808 case IB_CM_SIDR_REQ_ERROR:
3809 event.event = RDMA_CM_EVENT_UNREACHABLE;
3810 event.status = -ETIMEDOUT;
3811 break;
3812 case IB_CM_SIDR_REP_RECEIVED:
3813 event.param.ud.private_data = ib_event->private_data;
3814 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
3815 if (rep->status != IB_SIDR_SUCCESS) {
3816 event.event = RDMA_CM_EVENT_UNREACHABLE;
3817 event.status = ib_event->param.sidr_rep_rcvd.status;
3818 pr_debug_ratelimited("RDMA CM: UNREACHABLE: bad SIDR reply. status %d\n",
3819 event.status);
3820 break;
3821 }
3822 ret = cma_set_qkey(id_priv, rep->qkey);
3823 if (ret) {
3824 pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to set qkey. status %d\n", ret);
3825 event.event = RDMA_CM_EVENT_ADDR_ERROR;
3826 event.status = ret;
3827 break;
3828 }
3829 ib_init_ah_attr_from_path(id_priv->id.device,
3830 id_priv->id.port_num,
3831 id_priv->id.route.path_rec,
3832 &event.param.ud.ah_attr,
3833 rep->sgid_attr);
3834 event.param.ud.qp_num = rep->qpn;
3835 event.param.ud.qkey = rep->qkey;
3836 event.event = RDMA_CM_EVENT_ESTABLISHED;
3837 event.status = 0;
3838 break;
3839 default:
3840 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
3841 ib_event->event);
3842 goto out;
3843 }
3844
3845 ret = cma_cm_event_handler(id_priv, &event);
3846
3847 rdma_destroy_ah_attr(&event.param.ud.ah_attr);
3848 if (ret) {
3849 /* Destroy the CM ID by returning a non-zero value. */
3850 id_priv->cm_id.ib = NULL;
3851 destroy_id_handler_unlock(id_priv);
3852 return ret;
3853 }
3854 out:
3855 mutex_unlock(&id_priv->handler_mutex);
3856 return 0;
3857 }
3858
3859 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
3860 struct rdma_conn_param *conn_param)
3861 {
3862 struct ib_cm_sidr_req_param req;
3863 struct ib_cm_id *id;
3864 void *private_data;
3865 u8 offset;
3866 int ret;
3867
3868 memset(&req, 0, sizeof req);
3869 offset = cma_user_data_offset(id_priv);
3870 req.private_data_len = offset + conn_param->private_data_len;
3871 if (req.private_data_len < conn_param->private_data_len)
3872 return -EINVAL;
3873
3874 if (req.private_data_len) {
3875 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
3876 if (!private_data)
3877 return -ENOMEM;
3878 } else {
3879 private_data = NULL;
3880 }
3881
3882 if (conn_param->private_data && conn_param->private_data_len)
3883 memcpy(private_data + offset, conn_param->private_data,
3884 conn_param->private_data_len);
3885
3886 if (private_data) {
3887 ret = cma_format_hdr(private_data, id_priv);
3888 if (ret)
3889 goto out;
3890 req.private_data = private_data;
3891 }
3892
3893 id = ib_create_cm_id(id_priv->id.device, cma_sidr_rep_handler,
3894 id_priv);
3895 if (IS_ERR(id)) {
3896 ret = PTR_ERR(id);
3897 goto out;
3898 }
3899 id_priv->cm_id.ib = id;
3900
3901 req.path = id_priv->id.route.path_rec;
3902 req.sgid_attr = id_priv->id.route.addr.dev_addr.sgid_attr;
3903 req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
3904 req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
3905 req.max_cm_retries = CMA_MAX_CM_RETRIES;
3906
3907 trace_cm_send_sidr_req(id_priv);
3908 ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
3909 if (ret) {
3910 ib_destroy_cm_id(id_priv->cm_id.ib);
3911 id_priv->cm_id.ib = NULL;
3912 }
3913 out:
3914 kfree(private_data);
3915 return ret;
3916 }
3917
3918 static int cma_connect_ib(struct rdma_id_private *id_priv,
3919 struct rdma_conn_param *conn_param)
3920 {
3921 struct ib_cm_req_param req;
3922 struct rdma_route *route;
3923 void *private_data;
3924 struct ib_cm_id *id;
3925 u8 offset;
3926 int ret;
3927
3928 memset(&req, 0, sizeof req);
3929 offset = cma_user_data_offset(id_priv);
3930 req.private_data_len = offset + conn_param->private_data_len;
3931 if (req.private_data_len < conn_param->private_data_len)
3932 return -EINVAL;
3933
3934 if (req.private_data_len) {
3935 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
3936 if (!private_data)
3937 return -ENOMEM;
3938 } else {
3939 private_data = NULL;
3940 }
3941
3942 if (conn_param->private_data && conn_param->private_data_len)
3943 memcpy(private_data + offset, conn_param->private_data,
3944 conn_param->private_data_len);
3945
3946 id = ib_create_cm_id(id_priv->id.device, cma_ib_handler, id_priv);
3947 if (IS_ERR(id)) {
3948 ret = PTR_ERR(id);
3949 goto out;
3950 }
3951 id_priv->cm_id.ib = id;
3952
3953 route = &id_priv->id.route;
3954 if (private_data) {
3955 ret = cma_format_hdr(private_data, id_priv);
3956 if (ret)
3957 goto out;
3958 req.private_data = private_data;
3959 }
3960
3961 req.primary_path = &route->path_rec[0];
3962 if (route->num_paths == 2)
3963 req.alternate_path = &route->path_rec[1];
3964
3965 req.ppath_sgid_attr = id_priv->id.route.addr.dev_addr.sgid_attr;
3966 /* Alternate path SGID attribute currently unsupported */
3967 req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
3968 req.qp_num = id_priv->qp_num;
3969 req.qp_type = id_priv->id.qp_type;
3970 req.starting_psn = id_priv->seq_num;
3971 req.responder_resources = conn_param->responder_resources;
3972 req.initiator_depth = conn_param->initiator_depth;
3973 req.flow_control = conn_param->flow_control;
3974 req.retry_count = min_t(u8, 7, conn_param->retry_count);
3975 req.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
3976 req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
3977 req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
3978 req.max_cm_retries = CMA_MAX_CM_RETRIES;
3979 req.srq = id_priv->srq ? 1 : 0;
3980 req.ece.vendor_id = id_priv->ece.vendor_id;
3981 req.ece.attr_mod = id_priv->ece.attr_mod;
3982
3983 trace_cm_send_req(id_priv);
3984 ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
3985 out:
3986 if (ret && !IS_ERR(id)) {
3987 ib_destroy_cm_id(id);
3988 id_priv->cm_id.ib = NULL;
3989 }
3990
3991 kfree(private_data);
3992 return ret;
3993 }
3994
3995 static int cma_connect_iw(struct rdma_id_private *id_priv,
3996 struct rdma_conn_param *conn_param)
3997 {
3998 struct iw_cm_id *cm_id;
3999 int ret;
4000 struct iw_cm_conn_param iw_param;
4001
4002 cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
4003 if (IS_ERR(cm_id))
4004 return PTR_ERR(cm_id);
4005
4006 cm_id->tos = id_priv->tos;
4007 cm_id->tos_set = id_priv->tos_set;
4008 id_priv->cm_id.iw = cm_id;
4009
4010 memcpy(&cm_id->local_addr, cma_src_addr(id_priv),
4011 rdma_addr_size(cma_src_addr(id_priv)));
4012 memcpy(&cm_id->remote_addr, cma_dst_addr(id_priv),
4013 rdma_addr_size(cma_dst_addr(id_priv)));
4014
4015 ret = cma_modify_qp_rtr(id_priv, conn_param);
4016 if (ret)
4017 goto out;
4018
4019 if (conn_param) {
4020 iw_param.ord = conn_param->initiator_depth;
4021 iw_param.ird = conn_param->responder_resources;
4022 iw_param.private_data = conn_param->private_data;
4023 iw_param.private_data_len = conn_param->private_data_len;
4024 iw_param.qpn = id_priv->id.qp ? id_priv->qp_num : conn_param->qp_num;
4025 } else {
4026 memset(&iw_param, 0, sizeof iw_param);
4027 iw_param.qpn = id_priv->qp_num;
4028 }
4029 ret = iw_cm_connect(cm_id, &iw_param);
4030 out:
4031 if (ret) {
4032 iw_destroy_cm_id(cm_id);
4033 id_priv->cm_id.iw = NULL;
4034 }
4035 return ret;
4036 }
4037
4038 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
4039 {
4040 struct rdma_id_private *id_priv =
4041 container_of(id, struct rdma_id_private, id);
4042 int ret;
4043
4044 mutex_lock(&id_priv->handler_mutex);
4045 if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT)) {
4046 ret = -EINVAL;
4047 goto err_unlock;
4048 }
4049
4050 if (!id->qp) {
4051 id_priv->qp_num = conn_param->qp_num;
4052 id_priv->srq = conn_param->srq;
4053 }
4054
4055 if (rdma_cap_ib_cm(id->device, id->port_num)) {
4056 if (id->qp_type == IB_QPT_UD)
4057 ret = cma_resolve_ib_udp(id_priv, conn_param);
4058 else
4059 ret = cma_connect_ib(id_priv, conn_param);
4060 } else if (rdma_cap_iw_cm(id->device, id->port_num))
4061 ret = cma_connect_iw(id_priv, conn_param);
4062 else
4063 ret = -ENOSYS;
4064 if (ret)
4065 goto err_state;
4066 mutex_unlock(&id_priv->handler_mutex);
4067 return 0;
4068 err_state:
4069 cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED);
4070 err_unlock:
4071 mutex_unlock(&id_priv->handler_mutex);
4072 return ret;
4073 }
4074 EXPORT_SYMBOL(rdma_connect);
4075
4076 /**
4077 * rdma_connect_ece - Initiate an active connection request with ECE data.
4078 * @id: Connection identifier to connect.
4079 * @conn_param: Connection information used for connected QPs.
4080 * @ece: ECE parameters
4081 *
4082 * See rdma_connect() explanation.
4083 */
4084 int rdma_connect_ece(struct rdma_cm_id *id, struct rdma_conn_param *conn_param,
4085 struct rdma_ucm_ece *ece)
4086 {
4087 struct rdma_id_private *id_priv =
4088 container_of(id, struct rdma_id_private, id);
4089
4090 id_priv->ece.vendor_id = ece->vendor_id;
4091 id_priv->ece.attr_mod = ece->attr_mod;
4092
4093 return rdma_connect(id, conn_param);
4094 }
4095 EXPORT_SYMBOL(rdma_connect_ece);
4096
4097 static int cma_accept_ib(struct rdma_id_private *id_priv,
4098 struct rdma_conn_param *conn_param)
4099 {
4100 struct ib_cm_rep_param rep;
4101 int ret;
4102
4103 ret = cma_modify_qp_rtr(id_priv, conn_param);
4104 if (ret)
4105 goto out;
4106
4107 ret = cma_modify_qp_rts(id_priv, conn_param);
4108 if (ret)
4109 goto out;
4110
4111 memset(&rep, 0, sizeof rep);
4112 rep.qp_num = id_priv->qp_num;
4113 rep.starting_psn = id_priv->seq_num;
4114 rep.private_data = conn_param->private_data;
4115 rep.private_data_len = conn_param->private_data_len;
4116 rep.responder_resources = conn_param->responder_resources;
4117 rep.initiator_depth = conn_param->initiator_depth;
4118 rep.failover_accepted = 0;
4119 rep.flow_control = conn_param->flow_control;
4120 rep.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
4121 rep.srq = id_priv->srq ? 1 : 0;
4122 rep.ece.vendor_id = id_priv->ece.vendor_id;
4123 rep.ece.attr_mod = id_priv->ece.attr_mod;
4124
4125 trace_cm_send_rep(id_priv);
4126 ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
4127 out:
4128 return ret;
4129 }
4130
4131 static int cma_accept_iw(struct rdma_id_private *id_priv,
4132 struct rdma_conn_param *conn_param)
4133 {
4134 struct iw_cm_conn_param iw_param;
4135 int ret;
4136
4137 if (!conn_param)
4138 return -EINVAL;
4139
4140 ret = cma_modify_qp_rtr(id_priv, conn_param);
4141 if (ret)
4142 return ret;
4143
4144 iw_param.ord = conn_param->initiator_depth;
4145 iw_param.ird = conn_param->responder_resources;
4146 iw_param.private_data = conn_param->private_data;
4147 iw_param.private_data_len = conn_param->private_data_len;
4148 if (id_priv->id.qp) {
4149 iw_param.qpn = id_priv->qp_num;
4150 } else
4151 iw_param.qpn = conn_param->qp_num;
4152
4153 return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
4154 }
4155
4156 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
4157 enum ib_cm_sidr_status status, u32 qkey,
4158 const void *private_data, int private_data_len)
4159 {
4160 struct ib_cm_sidr_rep_param rep;
4161 int ret;
4162
4163 memset(&rep, 0, sizeof rep);
4164 rep.status = status;
4165 if (status == IB_SIDR_SUCCESS) {
4166 ret = cma_set_qkey(id_priv, qkey);
4167 if (ret)
4168 return ret;
4169 rep.qp_num = id_priv->qp_num;
4170 rep.qkey = id_priv->qkey;
4171
4172 rep.ece.vendor_id = id_priv->ece.vendor_id;
4173 rep.ece.attr_mod = id_priv->ece.attr_mod;
4174 }
4175
4176 rep.private_data = private_data;
4177 rep.private_data_len = private_data_len;
4178
4179 trace_cm_send_sidr_rep(id_priv);
4180 return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
4181 }
4182
4183 int __rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param,
4184 const char *caller)
4185 {
4186 struct rdma_id_private *id_priv =
4187 container_of(id, struct rdma_id_private, id);
4188 int ret;
4189
4190 lockdep_assert_held(&id_priv->handler_mutex);
4191
4192 rdma_restrack_set_task(&id_priv->res, caller);
4193
4194 if (READ_ONCE(id_priv->state) != RDMA_CM_CONNECT)
4195 return -EINVAL;
4196
4197 if (!id->qp && conn_param) {
4198 id_priv->qp_num = conn_param->qp_num;
4199 id_priv->srq = conn_param->srq;
4200 }
4201
4202 if (rdma_cap_ib_cm(id->device, id->port_num)) {
4203 if (id->qp_type == IB_QPT_UD) {
4204 if (conn_param)
4205 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
4206 conn_param->qkey,
4207 conn_param->private_data,
4208 conn_param->private_data_len);
4209 else
4210 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
4211 0, NULL, 0);
4212 } else {
4213 if (conn_param)
4214 ret = cma_accept_ib(id_priv, conn_param);
4215 else
4216 ret = cma_rep_recv(id_priv);
4217 }
4218 } else if (rdma_cap_iw_cm(id->device, id->port_num))
4219 ret = cma_accept_iw(id_priv, conn_param);
4220 else
4221 ret = -ENOSYS;
4222
4223 if (ret)
4224 goto reject;
4225
4226 return 0;
4227 reject:
4228 cma_modify_qp_err(id_priv);
4229 rdma_reject(id, NULL, 0, IB_CM_REJ_CONSUMER_DEFINED);
4230 return ret;
4231 }
4232 EXPORT_SYMBOL(__rdma_accept);
4233
4234 int __rdma_accept_ece(struct rdma_cm_id *id, struct rdma_conn_param *conn_param,
4235 const char *caller, struct rdma_ucm_ece *ece)
4236 {
4237 struct rdma_id_private *id_priv =
4238 container_of(id, struct rdma_id_private, id);
4239
4240 id_priv->ece.vendor_id = ece->vendor_id;
4241 id_priv->ece.attr_mod = ece->attr_mod;
4242
4243 return __rdma_accept(id, conn_param, caller);
4244 }
4245 EXPORT_SYMBOL(__rdma_accept_ece);
4246
4247 void rdma_lock_handler(struct rdma_cm_id *id)
4248 {
4249 struct rdma_id_private *id_priv =
4250 container_of(id, struct rdma_id_private, id);
4251
4252 mutex_lock(&id_priv->handler_mutex);
4253 }
4254 EXPORT_SYMBOL(rdma_lock_handler);
4255
4256 void rdma_unlock_handler(struct rdma_cm_id *id)
4257 {
4258 struct rdma_id_private *id_priv =
4259 container_of(id, struct rdma_id_private, id);
4260
4261 mutex_unlock(&id_priv->handler_mutex);
4262 }
4263 EXPORT_SYMBOL(rdma_unlock_handler);
4264
4265 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
4266 {
4267 struct rdma_id_private *id_priv;
4268 int ret;
4269
4270 id_priv = container_of(id, struct rdma_id_private, id);
4271 if (!id_priv->cm_id.ib)
4272 return -EINVAL;
4273
4274 switch (id->device->node_type) {
4275 case RDMA_NODE_IB_CA:
4276 ret = ib_cm_notify(id_priv->cm_id.ib, event);
4277 break;
4278 default:
4279 ret = 0;
4280 break;
4281 }
4282 return ret;
4283 }
4284 EXPORT_SYMBOL(rdma_notify);
4285
4286 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
4287 u8 private_data_len, u8 reason)
4288 {
4289 struct rdma_id_private *id_priv;
4290 int ret;
4291
4292 id_priv = container_of(id, struct rdma_id_private, id);
4293 if (!id_priv->cm_id.ib)
4294 return -EINVAL;
4295
4296 if (rdma_cap_ib_cm(id->device, id->port_num)) {
4297 if (id->qp_type == IB_QPT_UD) {
4298 ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT, 0,
4299 private_data, private_data_len);
4300 } else {
4301 trace_cm_send_rej(id_priv);
4302 ret = ib_send_cm_rej(id_priv->cm_id.ib, reason, NULL, 0,
4303 private_data, private_data_len);
4304 }
4305 } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4306 ret = iw_cm_reject(id_priv->cm_id.iw,
4307 private_data, private_data_len);
4308 } else
4309 ret = -ENOSYS;
4310
4311 return ret;
4312 }
4313 EXPORT_SYMBOL(rdma_reject);
4314
4315 int rdma_disconnect(struct rdma_cm_id *id)
4316 {
4317 struct rdma_id_private *id_priv;
4318 int ret;
4319
4320 id_priv = container_of(id, struct rdma_id_private, id);
4321 if (!id_priv->cm_id.ib)
4322 return -EINVAL;
4323
4324 if (rdma_cap_ib_cm(id->device, id->port_num)) {
4325 ret = cma_modify_qp_err(id_priv);
4326 if (ret)
4327 goto out;
4328 /* Initiate or respond to a disconnect. */
4329 trace_cm_disconnect(id_priv);
4330 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0)) {
4331 if (!ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0))
4332 trace_cm_sent_drep(id_priv);
4333 } else {
4334 trace_cm_sent_dreq(id_priv);
4335 }
4336 } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4337 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
4338 } else
4339 ret = -EINVAL;
4340
4341 out:
4342 return ret;
4343 }
4344 EXPORT_SYMBOL(rdma_disconnect);
4345
4346 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
4347 {
4348 struct rdma_id_private *id_priv;
4349 struct cma_multicast *mc = multicast->context;
4350 struct rdma_cm_event event = {};
4351 int ret = 0;
4352
4353 id_priv = mc->id_priv;
4354 mutex_lock(&id_priv->handler_mutex);
4355 if (READ_ONCE(id_priv->state) == RDMA_CM_DEVICE_REMOVAL ||
4356 READ_ONCE(id_priv->state) == RDMA_CM_DESTROYING)
4357 goto out;
4358
4359 if (!status)
4360 status = cma_set_qkey(id_priv, be32_to_cpu(multicast->rec.qkey));
4361 else
4362 pr_debug_ratelimited("RDMA CM: MULTICAST_ERROR: failed to join multicast. status %d\n",
4363 status);
4364 mutex_lock(&id_priv->qp_mutex);
4365 if (!status && id_priv->id.qp) {
4366 status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid,
4367 be16_to_cpu(multicast->rec.mlid));
4368 if (status)
4369 pr_debug_ratelimited("RDMA CM: MULTICAST_ERROR: failed to attach QP. status %d\n",
4370 status);
4371 }
4372 mutex_unlock(&id_priv->qp_mutex);
4373
4374 event.status = status;
4375 event.param.ud.private_data = mc->context;
4376 if (!status) {
4377 struct rdma_dev_addr *dev_addr =
4378 &id_priv->id.route.addr.dev_addr;
4379 struct net_device *ndev =
4380 dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if);
4381 enum ib_gid_type gid_type =
4382 id_priv->cma_dev->default_gid_type[id_priv->id.port_num -
4383 rdma_start_port(id_priv->cma_dev->device)];
4384
4385 event.event = RDMA_CM_EVENT_MULTICAST_JOIN;
4386 ret = ib_init_ah_from_mcmember(id_priv->id.device,
4387 id_priv->id.port_num,
4388 &multicast->rec,
4389 ndev, gid_type,
4390 &event.param.ud.ah_attr);
4391 if (ret)
4392 event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
4393
4394 event.param.ud.qp_num = 0xFFFFFF;
4395 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
4396 if (ndev)
4397 dev_put(ndev);
4398 } else
4399 event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
4400
4401 ret = cma_cm_event_handler(id_priv, &event);
4402
4403 rdma_destroy_ah_attr(&event.param.ud.ah_attr);
4404 if (ret) {
4405 destroy_id_handler_unlock(id_priv);
4406 return 0;
4407 }
4408
4409 out:
4410 mutex_unlock(&id_priv->handler_mutex);
4411 return 0;
4412 }
4413
4414 static void cma_set_mgid(struct rdma_id_private *id_priv,
4415 struct sockaddr *addr, union ib_gid *mgid)
4416 {
4417 unsigned char mc_map[MAX_ADDR_LEN];
4418 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4419 struct sockaddr_in *sin = (struct sockaddr_in *) addr;
4420 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
4421
4422 if (cma_any_addr(addr)) {
4423 memset(mgid, 0, sizeof *mgid);
4424 } else if ((addr->sa_family == AF_INET6) &&
4425 ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
4426 0xFF10A01B)) {
4427 /* IPv6 address is an SA assigned MGID. */
4428 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
4429 } else if (addr->sa_family == AF_IB) {
4430 memcpy(mgid, &((struct sockaddr_ib *) addr)->sib_addr, sizeof *mgid);
4431 } else if (addr->sa_family == AF_INET6) {
4432 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
4433 if (id_priv->id.ps == RDMA_PS_UDP)
4434 mc_map[7] = 0x01; /* Use RDMA CM signature */
4435 *mgid = *(union ib_gid *) (mc_map + 4);
4436 } else {
4437 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
4438 if (id_priv->id.ps == RDMA_PS_UDP)
4439 mc_map[7] = 0x01; /* Use RDMA CM signature */
4440 *mgid = *(union ib_gid *) (mc_map + 4);
4441 }
4442 }
4443
4444 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
4445 struct cma_multicast *mc)
4446 {
4447 struct ib_sa_mcmember_rec rec;
4448 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4449 ib_sa_comp_mask comp_mask;
4450 int ret;
4451
4452 ib_addr_get_mgid(dev_addr, &rec.mgid);
4453 ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
4454 &rec.mgid, &rec);
4455 if (ret)
4456 return ret;
4457
4458 ret = cma_set_qkey(id_priv, 0);
4459 if (ret)
4460 return ret;
4461
4462 cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
4463 rec.qkey = cpu_to_be32(id_priv->qkey);
4464 rdma_addr_get_sgid(dev_addr, &rec.port_gid);
4465 rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
4466 rec.join_state = mc->join_state;
4467
4468 if ((rec.join_state == BIT(SENDONLY_FULLMEMBER_JOIN)) &&
4469 (!ib_sa_sendonly_fullmem_support(&sa_client,
4470 id_priv->id.device,
4471 id_priv->id.port_num))) {
4472 dev_warn(
4473 &id_priv->id.device->dev,
4474 "RDMA CM: port %u Unable to multicast join: SM doesn't support Send Only Full Member option\n",
4475 id_priv->id.port_num);
4476 return -EOPNOTSUPP;
4477 }
4478
4479 comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
4480 IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
4481 IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
4482 IB_SA_MCMEMBER_REC_FLOW_LABEL |
4483 IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
4484
4485 if (id_priv->id.ps == RDMA_PS_IPOIB)
4486 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
4487 IB_SA_MCMEMBER_REC_RATE_SELECTOR |
4488 IB_SA_MCMEMBER_REC_MTU_SELECTOR |
4489 IB_SA_MCMEMBER_REC_MTU |
4490 IB_SA_MCMEMBER_REC_HOP_LIMIT;
4491
4492 mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device,
4493 id_priv->id.port_num, &rec,
4494 comp_mask, GFP_KERNEL,
4495 cma_ib_mc_handler, mc);
4496 return PTR_ERR_OR_ZERO(mc->multicast.ib);
4497 }
4498
4499 static void iboe_mcast_work_handler(struct work_struct *work)
4500 {
4501 struct iboe_mcast_work *mw = container_of(work, struct iboe_mcast_work, work);
4502 struct cma_multicast *mc = mw->mc;
4503 struct ib_sa_multicast *m = mc->multicast.ib;
4504
4505 mc->multicast.ib->context = mc;
4506 cma_ib_mc_handler(0, m);
4507 kref_put(&mc->mcref, release_mc);
4508 kfree(mw);
4509 }
4510
4511 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid,
4512 enum ib_gid_type gid_type)
4513 {
4514 struct sockaddr_in *sin = (struct sockaddr_in *)addr;
4515 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
4516
4517 if (cma_any_addr(addr)) {
4518 memset(mgid, 0, sizeof *mgid);
4519 } else if (addr->sa_family == AF_INET6) {
4520 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
4521 } else {
4522 mgid->raw[0] =
4523 (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ? 0 : 0xff;
4524 mgid->raw[1] =
4525 (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ? 0 : 0x0e;
4526 mgid->raw[2] = 0;
4527 mgid->raw[3] = 0;
4528 mgid->raw[4] = 0;
4529 mgid->raw[5] = 0;
4530 mgid->raw[6] = 0;
4531 mgid->raw[7] = 0;
4532 mgid->raw[8] = 0;
4533 mgid->raw[9] = 0;
4534 mgid->raw[10] = 0xff;
4535 mgid->raw[11] = 0xff;
4536 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
4537 }
4538 }
4539
4540 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
4541 struct cma_multicast *mc)
4542 {
4543 struct iboe_mcast_work *work;
4544 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4545 int err = 0;
4546 struct sockaddr *addr = (struct sockaddr *)&mc->addr;
4547 struct net_device *ndev = NULL;
4548 enum ib_gid_type gid_type;
4549 bool send_only;
4550
4551 send_only = mc->join_state == BIT(SENDONLY_FULLMEMBER_JOIN);
4552
4553 if (cma_zero_addr((struct sockaddr *)&mc->addr))
4554 return -EINVAL;
4555
4556 work = kzalloc(sizeof *work, GFP_KERNEL);
4557 if (!work)
4558 return -ENOMEM;
4559
4560 mc->multicast.ib = kzalloc(sizeof(struct ib_sa_multicast), GFP_KERNEL);
4561 if (!mc->multicast.ib) {
4562 err = -ENOMEM;
4563 goto out1;
4564 }
4565
4566 gid_type = id_priv->cma_dev->default_gid_type[id_priv->id.port_num -
4567 rdma_start_port(id_priv->cma_dev->device)];
4568 cma_iboe_set_mgid(addr, &mc->multicast.ib->rec.mgid, gid_type);
4569
4570 mc->multicast.ib->rec.pkey = cpu_to_be16(0xffff);
4571 if (id_priv->id.ps == RDMA_PS_UDP)
4572 mc->multicast.ib->rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
4573
4574 if (dev_addr->bound_dev_if)
4575 ndev = dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if);
4576 if (!ndev) {
4577 err = -ENODEV;
4578 goto out2;
4579 }
4580 mc->multicast.ib->rec.rate = iboe_get_rate(ndev);
4581 mc->multicast.ib->rec.hop_limit = 1;
4582 mc->multicast.ib->rec.mtu = iboe_get_mtu(ndev->mtu);
4583
4584 if (addr->sa_family == AF_INET) {
4585 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) {
4586 mc->multicast.ib->rec.hop_limit = IPV6_DEFAULT_HOPLIMIT;
4587 if (!send_only) {
4588 err = cma_igmp_send(ndev, &mc->multicast.ib->rec.mgid,
4589 true);
4590 }
4591 }
4592 } else {
4593 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP)
4594 err = -ENOTSUPP;
4595 }
4596 dev_put(ndev);
4597 if (err || !mc->multicast.ib->rec.mtu) {
4598 if (!err)
4599 err = -EINVAL;
4600 goto out2;
4601 }
4602 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
4603 &mc->multicast.ib->rec.port_gid);
4604 work->id = id_priv;
4605 work->mc = mc;
4606 INIT_WORK(&work->work, iboe_mcast_work_handler);
4607 kref_get(&mc->mcref);
4608 queue_work(cma_wq, &work->work);
4609
4610 return 0;
4611
4612 out2:
4613 kfree(mc->multicast.ib);
4614 out1:
4615 kfree(work);
4616 return err;
4617 }
4618
4619 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
4620 u8 join_state, void *context)
4621 {
4622 struct rdma_id_private *id_priv =
4623 container_of(id, struct rdma_id_private, id);
4624 struct cma_multicast *mc;
4625 int ret;
4626
4627 /* ULP is calling this wrong. */
4628 if (!id->device || (READ_ONCE(id_priv->state) != RDMA_CM_ADDR_BOUND &&
4629 READ_ONCE(id_priv->state) != RDMA_CM_ADDR_RESOLVED))
4630 return -EINVAL;
4631
4632 mc = kmalloc(sizeof *mc, GFP_KERNEL);
4633 if (!mc)
4634 return -ENOMEM;
4635
4636 memcpy(&mc->addr, addr, rdma_addr_size(addr));
4637 mc->context = context;
4638 mc->id_priv = id_priv;
4639 mc->join_state = join_state;
4640
4641 if (rdma_protocol_roce(id->device, id->port_num)) {
4642 kref_init(&mc->mcref);
4643 ret = cma_iboe_join_multicast(id_priv, mc);
4644 if (ret)
4645 goto out_err;
4646 } else if (rdma_cap_ib_mcast(id->device, id->port_num)) {
4647 ret = cma_join_ib_multicast(id_priv, mc);
4648 if (ret)
4649 goto out_err;
4650 } else {
4651 ret = -ENOSYS;
4652 goto out_err;
4653 }
4654
4655 spin_lock(&id_priv->lock);
4656 list_add(&mc->list, &id_priv->mc_list);
4657 spin_unlock(&id_priv->lock);
4658
4659 return 0;
4660 out_err:
4661 kfree(mc);
4662 return ret;
4663 }
4664 EXPORT_SYMBOL(rdma_join_multicast);
4665
4666 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
4667 {
4668 struct rdma_id_private *id_priv;
4669 struct cma_multicast *mc;
4670
4671 id_priv = container_of(id, struct rdma_id_private, id);
4672 spin_lock_irq(&id_priv->lock);
4673 list_for_each_entry(mc, &id_priv->mc_list, list) {
4674 if (!memcmp(&mc->addr, addr, rdma_addr_size(addr))) {
4675 list_del(&mc->list);
4676 spin_unlock_irq(&id_priv->lock);
4677
4678 if (id->qp)
4679 ib_detach_mcast(id->qp,
4680 &mc->multicast.ib->rec.mgid,
4681 be16_to_cpu(mc->multicast.ib->rec.mlid));
4682
4683 BUG_ON(id_priv->cma_dev->device != id->device);
4684
4685 if (rdma_cap_ib_mcast(id->device, id->port_num)) {
4686 ib_sa_free_multicast(mc->multicast.ib);
4687 kfree(mc);
4688 } else if (rdma_protocol_roce(id->device, id->port_num)) {
4689 cma_leave_roce_mc_group(id_priv, mc);
4690 }
4691 return;
4692 }
4693 }
4694 spin_unlock_irq(&id_priv->lock);
4695 }
4696 EXPORT_SYMBOL(rdma_leave_multicast);
4697
4698 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
4699 {
4700 struct rdma_dev_addr *dev_addr;
4701 struct cma_ndev_work *work;
4702
4703 dev_addr = &id_priv->id.route.addr.dev_addr;
4704
4705 if ((dev_addr->bound_dev_if == ndev->ifindex) &&
4706 (net_eq(dev_net(ndev), dev_addr->net)) &&
4707 memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
4708 pr_info("RDMA CM addr change for ndev %s used by id %p\n",
4709 ndev->name, &id_priv->id);
4710 work = kzalloc(sizeof *work, GFP_KERNEL);
4711 if (!work)
4712 return -ENOMEM;
4713
4714 INIT_WORK(&work->work, cma_ndev_work_handler);
4715 work->id = id_priv;
4716 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
4717 cma_id_get(id_priv);
4718 queue_work(cma_wq, &work->work);
4719 }
4720
4721 return 0;
4722 }
4723
4724 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
4725 void *ptr)
4726 {
4727 struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
4728 struct cma_device *cma_dev;
4729 struct rdma_id_private *id_priv;
4730 int ret = NOTIFY_DONE;
4731
4732 if (event != NETDEV_BONDING_FAILOVER)
4733 return NOTIFY_DONE;
4734
4735 if (!netif_is_bond_master(ndev))
4736 return NOTIFY_DONE;
4737
4738 mutex_lock(&lock);
4739 list_for_each_entry(cma_dev, &dev_list, list)
4740 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
4741 ret = cma_netdev_change(ndev, id_priv);
4742 if (ret)
4743 goto out;
4744 }
4745
4746 out:
4747 mutex_unlock(&lock);
4748 return ret;
4749 }
4750
4751 static struct notifier_block cma_nb = {
4752 .notifier_call = cma_netdev_callback
4753 };
4754
4755 static int cma_add_one(struct ib_device *device)
4756 {
4757 struct cma_device *cma_dev;
4758 struct rdma_id_private *id_priv;
4759 unsigned int i;
4760 unsigned long supported_gids = 0;
4761 int ret;
4762
4763 cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
4764 if (!cma_dev)
4765 return -ENOMEM;
4766
4767 cma_dev->device = device;
4768 cma_dev->default_gid_type = kcalloc(device->phys_port_cnt,
4769 sizeof(*cma_dev->default_gid_type),
4770 GFP_KERNEL);
4771 if (!cma_dev->default_gid_type) {
4772 ret = -ENOMEM;
4773 goto free_cma_dev;
4774 }
4775
4776 cma_dev->default_roce_tos = kcalloc(device->phys_port_cnt,
4777 sizeof(*cma_dev->default_roce_tos),
4778 GFP_KERNEL);
4779 if (!cma_dev->default_roce_tos) {
4780 ret = -ENOMEM;
4781 goto free_gid_type;
4782 }
4783
4784 rdma_for_each_port (device, i) {
4785 supported_gids = roce_gid_type_mask_support(device, i);
4786 WARN_ON(!supported_gids);
4787 if (supported_gids & (1 << CMA_PREFERRED_ROCE_GID_TYPE))
4788 cma_dev->default_gid_type[i - rdma_start_port(device)] =
4789 CMA_PREFERRED_ROCE_GID_TYPE;
4790 else
4791 cma_dev->default_gid_type[i - rdma_start_port(device)] =
4792 find_first_bit(&supported_gids, BITS_PER_LONG);
4793 cma_dev->default_roce_tos[i - rdma_start_port(device)] = 0;
4794 }
4795
4796 init_completion(&cma_dev->comp);
4797 refcount_set(&cma_dev->refcount, 1);
4798 INIT_LIST_HEAD(&cma_dev->id_list);
4799 ib_set_client_data(device, &cma_client, cma_dev);
4800
4801 mutex_lock(&lock);
4802 list_add_tail(&cma_dev->list, &dev_list);
4803 list_for_each_entry(id_priv, &listen_any_list, list)
4804 cma_listen_on_dev(id_priv, cma_dev);
4805 mutex_unlock(&lock);
4806
4807 trace_cm_add_one(device);
4808 return 0;
4809
4810 free_gid_type:
4811 kfree(cma_dev->default_gid_type);
4812
4813 free_cma_dev:
4814 kfree(cma_dev);
4815 return ret;
4816 }
4817
4818 static void cma_send_device_removal_put(struct rdma_id_private *id_priv)
4819 {
4820 struct rdma_cm_event event = { .event = RDMA_CM_EVENT_DEVICE_REMOVAL };
4821 enum rdma_cm_state state;
4822 unsigned long flags;
4823
4824 mutex_lock(&id_priv->handler_mutex);
4825 /* Record that we want to remove the device */
4826 spin_lock_irqsave(&id_priv->lock, flags);
4827 state = id_priv->state;
4828 if (state == RDMA_CM_DESTROYING || state == RDMA_CM_DEVICE_REMOVAL) {
4829 spin_unlock_irqrestore(&id_priv->lock, flags);
4830 mutex_unlock(&id_priv->handler_mutex);
4831 cma_id_put(id_priv);
4832 return;
4833 }
4834 id_priv->state = RDMA_CM_DEVICE_REMOVAL;
4835 spin_unlock_irqrestore(&id_priv->lock, flags);
4836
4837 if (cma_cm_event_handler(id_priv, &event)) {
4838 /*
4839 * At this point the ULP promises it won't call
4840 * rdma_destroy_id() concurrently
4841 */
4842 cma_id_put(id_priv);
4843 mutex_unlock(&id_priv->handler_mutex);
4844 trace_cm_id_destroy(id_priv);
4845 _destroy_id(id_priv, state);
4846 return;
4847 }
4848 mutex_unlock(&id_priv->handler_mutex);
4849
4850 /*
4851 * If this races with destroy then the thread that first assigns state
4852 * to a destroying does the cancel.
4853 */
4854 cma_cancel_operation(id_priv, state);
4855 cma_id_put(id_priv);
4856 }
4857
4858 static void cma_process_remove(struct cma_device *cma_dev)
4859 {
4860 mutex_lock(&lock);
4861 while (!list_empty(&cma_dev->id_list)) {
4862 struct rdma_id_private *id_priv = list_first_entry(
4863 &cma_dev->id_list, struct rdma_id_private, list);
4864
4865 list_del(&id_priv->listen_list);
4866 list_del_init(&id_priv->list);
4867 cma_id_get(id_priv);
4868 mutex_unlock(&lock);
4869
4870 cma_send_device_removal_put(id_priv);
4871
4872 mutex_lock(&lock);
4873 }
4874 mutex_unlock(&lock);
4875
4876 cma_dev_put(cma_dev);
4877 wait_for_completion(&cma_dev->comp);
4878 }
4879
4880 static void cma_remove_one(struct ib_device *device, void *client_data)
4881 {
4882 struct cma_device *cma_dev = client_data;
4883
4884 trace_cm_remove_one(device);
4885
4886 mutex_lock(&lock);
4887 list_del(&cma_dev->list);
4888 mutex_unlock(&lock);
4889
4890 cma_process_remove(cma_dev);
4891 kfree(cma_dev->default_roce_tos);
4892 kfree(cma_dev->default_gid_type);
4893 kfree(cma_dev);
4894 }
4895
4896 static int cma_init_net(struct net *net)
4897 {
4898 struct cma_pernet *pernet = cma_pernet(net);
4899
4900 xa_init(&pernet->tcp_ps);
4901 xa_init(&pernet->udp_ps);
4902 xa_init(&pernet->ipoib_ps);
4903 xa_init(&pernet->ib_ps);
4904
4905 return 0;
4906 }
4907
4908 static void cma_exit_net(struct net *net)
4909 {
4910 struct cma_pernet *pernet = cma_pernet(net);
4911
4912 WARN_ON(!xa_empty(&pernet->tcp_ps));
4913 WARN_ON(!xa_empty(&pernet->udp_ps));
4914 WARN_ON(!xa_empty(&pernet->ipoib_ps));
4915 WARN_ON(!xa_empty(&pernet->ib_ps));
4916 }
4917
4918 static struct pernet_operations cma_pernet_operations = {
4919 .init = cma_init_net,
4920 .exit = cma_exit_net,
4921 .id = &cma_pernet_id,
4922 .size = sizeof(struct cma_pernet),
4923 };
4924
4925 static int __init cma_init(void)
4926 {
4927 int ret;
4928
4929 /*
4930 * There is a rare lock ordering dependency in cma_netdev_callback()
4931 * that only happens when bonding is enabled. Teach lockdep that rtnl
4932 * must never be nested under lock so it can find these without having
4933 * to test with bonding.
4934 */
4935 if (IS_ENABLED(CONFIG_LOCKDEP)) {
4936 rtnl_lock();
4937 mutex_lock(&lock);
4938 mutex_unlock(&lock);
4939 rtnl_unlock();
4940 }
4941
4942 cma_wq = alloc_ordered_workqueue("rdma_cm", WQ_MEM_RECLAIM);
4943 if (!cma_wq)
4944 return -ENOMEM;
4945
4946 ret = register_pernet_subsys(&cma_pernet_operations);
4947 if (ret)
4948 goto err_wq;
4949
4950 ib_sa_register_client(&sa_client);
4951 register_netdevice_notifier(&cma_nb);
4952
4953 ret = ib_register_client(&cma_client);
4954 if (ret)
4955 goto err;
4956
4957 ret = cma_configfs_init();
4958 if (ret)
4959 goto err_ib;
4960
4961 return 0;
4962
4963 err_ib:
4964 ib_unregister_client(&cma_client);
4965 err:
4966 unregister_netdevice_notifier(&cma_nb);
4967 ib_sa_unregister_client(&sa_client);
4968 unregister_pernet_subsys(&cma_pernet_operations);
4969 err_wq:
4970 destroy_workqueue(cma_wq);
4971 return ret;
4972 }
4973
4974 static void __exit cma_cleanup(void)
4975 {
4976 cma_configfs_exit();
4977 ib_unregister_client(&cma_client);
4978 unregister_netdevice_notifier(&cma_nb);
4979 ib_sa_unregister_client(&sa_client);
4980 unregister_pernet_subsys(&cma_pernet_operations);
4981 destroy_workqueue(cma_wq);
4982 }
4983
4984 module_init(cma_init);
4985 module_exit(cma_cleanup);
Ubuntu Hirsute Linux kernel mirror