?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
dm writecache: add cond_resched to loop in persistent_memory_claim()
[linux/fpc-iii.git] / drivers / infiniband / core / cma.c
blob12ada58c96a906253ce77b4622e02d7369859baa
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 bool rdma_is_consumer_reject(struct rdma_cm_id *id, int reason)
95 {
96 if (rdma_ib_or_roce(id->device, id->port_num))
97 return reason == IB_CM_REJ_CONSUMER_DEFINED;
98
99 if (rdma_protocol_iwarp(id->device, id->port_num))
100 return reason == -ECONNREFUSED;
101
102 WARN_ON_ONCE(1);
103 return false;
104 }
105 EXPORT_SYMBOL(rdma_is_consumer_reject);
106
107 const void *rdma_consumer_reject_data(struct rdma_cm_id *id,
108 struct rdma_cm_event *ev, u8 *data_len)
109 {
110 const void *p;
111
112 if (rdma_is_consumer_reject(id, ev->status)) {
113 *data_len = ev->param.conn.private_data_len;
114 p = ev->param.conn.private_data;
115 } else {
116 *data_len = 0;
117 p = NULL;
118 }
119 return p;
120 }
121 EXPORT_SYMBOL(rdma_consumer_reject_data);
122
123 /**
124 * rdma_iw_cm_id() - return the iw_cm_id pointer for this cm_id.
125 * @id: Communication Identifier
126 */
127 struct iw_cm_id *rdma_iw_cm_id(struct rdma_cm_id *id)
128 {
129 struct rdma_id_private *id_priv;
130
131 id_priv = container_of(id, struct rdma_id_private, id);
132 if (id->device->node_type == RDMA_NODE_RNIC)
133 return id_priv->cm_id.iw;
134 return NULL;
135 }
136 EXPORT_SYMBOL(rdma_iw_cm_id);
137
138 /**
139 * rdma_res_to_id() - return the rdma_cm_id pointer for this restrack.
140 * @res: rdma resource tracking entry pointer
141 */
142 struct rdma_cm_id *rdma_res_to_id(struct rdma_restrack_entry *res)
143 {
144 struct rdma_id_private *id_priv =
145 container_of(res, struct rdma_id_private, res);
146
147 return &id_priv->id;
148 }
149 EXPORT_SYMBOL(rdma_res_to_id);
150
151 static void cma_add_one(struct ib_device *device);
152 static void cma_remove_one(struct ib_device *device, void *client_data);
153
154 static struct ib_client cma_client = {
155 .name = "cma",
156 .add = cma_add_one,
157 .remove = cma_remove_one
158 };
159
160 static struct ib_sa_client sa_client;
161 static LIST_HEAD(dev_list);
162 static LIST_HEAD(listen_any_list);
163 static DEFINE_MUTEX(lock);
164 static struct workqueue_struct *cma_wq;
165 static unsigned int cma_pernet_id;
166
167 struct cma_pernet {
168 struct xarray tcp_ps;
169 struct xarray udp_ps;
170 struct xarray ipoib_ps;
171 struct xarray ib_ps;
172 };
173
174 static struct cma_pernet *cma_pernet(struct net *net)
175 {
176 return net_generic(net, cma_pernet_id);
177 }
178
179 static
180 struct xarray *cma_pernet_xa(struct net *net, enum rdma_ucm_port_space ps)
181 {
182 struct cma_pernet *pernet = cma_pernet(net);
183
184 switch (ps) {
185 case RDMA_PS_TCP:
186 return &pernet->tcp_ps;
187 case RDMA_PS_UDP:
188 return &pernet->udp_ps;
189 case RDMA_PS_IPOIB:
190 return &pernet->ipoib_ps;
191 case RDMA_PS_IB:
192 return &pernet->ib_ps;
193 default:
194 return NULL;
195 }
196 }
197
198 struct cma_device {
199 struct list_head list;
200 struct ib_device *device;
201 struct completion comp;
202 refcount_t refcount;
203 struct list_head id_list;
204 enum ib_gid_type *default_gid_type;
205 u8 *default_roce_tos;
206 };
207
208 struct rdma_bind_list {
209 enum rdma_ucm_port_space ps;
210 struct hlist_head owners;
211 unsigned short port;
212 };
213
214 struct class_port_info_context {
215 struct ib_class_port_info *class_port_info;
216 struct ib_device *device;
217 struct completion done;
218 struct ib_sa_query *sa_query;
219 u8 port_num;
220 };
221
222 static int cma_ps_alloc(struct net *net, enum rdma_ucm_port_space ps,
223 struct rdma_bind_list *bind_list, int snum)
224 {
225 struct xarray *xa = cma_pernet_xa(net, ps);
226
227 return xa_insert(xa, snum, bind_list, GFP_KERNEL);
228 }
229
230 static struct rdma_bind_list *cma_ps_find(struct net *net,
231 enum rdma_ucm_port_space ps, int snum)
232 {
233 struct xarray *xa = cma_pernet_xa(net, ps);
234
235 return xa_load(xa, snum);
236 }
237
238 static void cma_ps_remove(struct net *net, enum rdma_ucm_port_space ps,
239 int snum)
240 {
241 struct xarray *xa = cma_pernet_xa(net, ps);
242
243 xa_erase(xa, snum);
244 }
245
246 enum {
247 CMA_OPTION_AFONLY,
248 };
249
250 void cma_dev_get(struct cma_device *cma_dev)
251 {
252 refcount_inc(&cma_dev->refcount);
253 }
254
255 void cma_dev_put(struct cma_device *cma_dev)
256 {
257 if (refcount_dec_and_test(&cma_dev->refcount))
258 complete(&cma_dev->comp);
259 }
260
261 struct cma_device *cma_enum_devices_by_ibdev(cma_device_filter filter,
262 void *cookie)
263 {
264 struct cma_device *cma_dev;
265 struct cma_device *found_cma_dev = NULL;
266
267 mutex_lock(&lock);
268
269 list_for_each_entry(cma_dev, &dev_list, list)
270 if (filter(cma_dev->device, cookie)) {
271 found_cma_dev = cma_dev;
272 break;
273 }
274
275 if (found_cma_dev)
276 cma_dev_get(found_cma_dev);
277 mutex_unlock(&lock);
278 return found_cma_dev;
279 }
280
281 int cma_get_default_gid_type(struct cma_device *cma_dev,
282 unsigned int port)
283 {
284 if (!rdma_is_port_valid(cma_dev->device, port))
285 return -EINVAL;
286
287 return cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)];
288 }
289
290 int cma_set_default_gid_type(struct cma_device *cma_dev,
291 unsigned int port,
292 enum ib_gid_type default_gid_type)
293 {
294 unsigned long supported_gids;
295
296 if (!rdma_is_port_valid(cma_dev->device, port))
297 return -EINVAL;
298
299 supported_gids = roce_gid_type_mask_support(cma_dev->device, port);
300
301 if (!(supported_gids & 1 << default_gid_type))
302 return -EINVAL;
303
304 cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)] =
305 default_gid_type;
306
307 return 0;
308 }
309
310 int cma_get_default_roce_tos(struct cma_device *cma_dev, unsigned int port)
311 {
312 if (!rdma_is_port_valid(cma_dev->device, port))
313 return -EINVAL;
314
315 return cma_dev->default_roce_tos[port - rdma_start_port(cma_dev->device)];
316 }
317
318 int cma_set_default_roce_tos(struct cma_device *cma_dev, unsigned int port,
319 u8 default_roce_tos)
320 {
321 if (!rdma_is_port_valid(cma_dev->device, port))
322 return -EINVAL;
323
324 cma_dev->default_roce_tos[port - rdma_start_port(cma_dev->device)] =
325 default_roce_tos;
326
327 return 0;
328 }
329 struct ib_device *cma_get_ib_dev(struct cma_device *cma_dev)
330 {
331 return cma_dev->device;
332 }
333
334 /*
335 * Device removal can occur at anytime, so we need extra handling to
336 * serialize notifying the user of device removal with other callbacks.
337 * We do this by disabling removal notification while a callback is in process,
338 * and reporting it after the callback completes.
339 */
340
341 struct cma_multicast {
342 struct rdma_id_private *id_priv;
343 union {
344 struct ib_sa_multicast *ib;
345 } multicast;
346 struct list_head list;
347 void *context;
348 struct sockaddr_storage addr;
349 struct kref mcref;
350 u8 join_state;
351 };
352
353 struct cma_work {
354 struct work_struct work;
355 struct rdma_id_private *id;
356 enum rdma_cm_state old_state;
357 enum rdma_cm_state new_state;
358 struct rdma_cm_event event;
359 };
360
361 struct cma_ndev_work {
362 struct work_struct work;
363 struct rdma_id_private *id;
364 struct rdma_cm_event event;
365 };
366
367 struct iboe_mcast_work {
368 struct work_struct work;
369 struct rdma_id_private *id;
370 struct cma_multicast *mc;
371 };
372
373 union cma_ip_addr {
374 struct in6_addr ip6;
375 struct {
376 __be32 pad[3];
377 __be32 addr;
378 } ip4;
379 };
380
381 struct cma_hdr {
382 u8 cma_version;
383 u8 ip_version; /* IP version: 7:4 */
384 __be16 port;
385 union cma_ip_addr src_addr;
386 union cma_ip_addr dst_addr;
387 };
388
389 #define CMA_VERSION 0x00
390
391 struct cma_req_info {
392 struct sockaddr_storage listen_addr_storage;
393 struct sockaddr_storage src_addr_storage;
394 struct ib_device *device;
395 union ib_gid local_gid;
396 __be64 service_id;
397 int port;
398 bool has_gid;
399 u16 pkey;
400 };
401
402 static int cma_comp(struct rdma_id_private *id_priv, enum rdma_cm_state comp)
403 {
404 unsigned long flags;
405 int ret;
406
407 spin_lock_irqsave(&id_priv->lock, flags);
408 ret = (id_priv->state == comp);
409 spin_unlock_irqrestore(&id_priv->lock, flags);
410 return ret;
411 }
412
413 static int cma_comp_exch(struct rdma_id_private *id_priv,
414 enum rdma_cm_state comp, enum rdma_cm_state exch)
415 {
416 unsigned long flags;
417 int ret;
418
419 spin_lock_irqsave(&id_priv->lock, flags);
420 if ((ret = (id_priv->state == comp)))
421 id_priv->state = exch;
422 spin_unlock_irqrestore(&id_priv->lock, flags);
423 return ret;
424 }
425
426 static enum rdma_cm_state cma_exch(struct rdma_id_private *id_priv,
427 enum rdma_cm_state exch)
428 {
429 unsigned long flags;
430 enum rdma_cm_state old;
431
432 spin_lock_irqsave(&id_priv->lock, flags);
433 old = id_priv->state;
434 id_priv->state = exch;
435 spin_unlock_irqrestore(&id_priv->lock, flags);
436 return old;
437 }
438
439 static inline u8 cma_get_ip_ver(const struct cma_hdr *hdr)
440 {
441 return hdr->ip_version >> 4;
442 }
443
444 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
445 {
446 hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
447 }
448
449 static int cma_igmp_send(struct net_device *ndev, union ib_gid *mgid, bool join)
450 {
451 struct in_device *in_dev = NULL;
452
453 if (ndev) {
454 rtnl_lock();
455 in_dev = __in_dev_get_rtnl(ndev);
456 if (in_dev) {
457 if (join)
458 ip_mc_inc_group(in_dev,
459 *(__be32 *)(mgid->raw + 12));
460 else
461 ip_mc_dec_group(in_dev,
462 *(__be32 *)(mgid->raw + 12));
463 }
464 rtnl_unlock();
465 }
466 return (in_dev) ? 0 : -ENODEV;
467 }
468
469 static void _cma_attach_to_dev(struct rdma_id_private *id_priv,
470 struct cma_device *cma_dev)
471 {
472 cma_dev_get(cma_dev);
473 id_priv->cma_dev = cma_dev;
474 id_priv->id.device = cma_dev->device;
475 id_priv->id.route.addr.dev_addr.transport =
476 rdma_node_get_transport(cma_dev->device->node_type);
477 list_add_tail(&id_priv->list, &cma_dev->id_list);
478 if (id_priv->res.kern_name)
479 rdma_restrack_kadd(&id_priv->res);
480 else
481 rdma_restrack_uadd(&id_priv->res);
482 }
483
484 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
485 struct cma_device *cma_dev)
486 {
487 _cma_attach_to_dev(id_priv, cma_dev);
488 id_priv->gid_type =
489 cma_dev->default_gid_type[id_priv->id.port_num -
490 rdma_start_port(cma_dev->device)];
491 }
492
493 static inline void release_mc(struct kref *kref)
494 {
495 struct cma_multicast *mc = container_of(kref, struct cma_multicast, mcref);
496
497 kfree(mc->multicast.ib);
498 kfree(mc);
499 }
500
501 static void cma_release_dev(struct rdma_id_private *id_priv)
502 {
503 mutex_lock(&lock);
504 list_del(&id_priv->list);
505 cma_dev_put(id_priv->cma_dev);
506 id_priv->cma_dev = NULL;
507 mutex_unlock(&lock);
508 }
509
510 static inline struct sockaddr *cma_src_addr(struct rdma_id_private *id_priv)
511 {
512 return (struct sockaddr *) &id_priv->id.route.addr.src_addr;
513 }
514
515 static inline struct sockaddr *cma_dst_addr(struct rdma_id_private *id_priv)
516 {
517 return (struct sockaddr *) &id_priv->id.route.addr.dst_addr;
518 }
519
520 static inline unsigned short cma_family(struct rdma_id_private *id_priv)
521 {
522 return id_priv->id.route.addr.src_addr.ss_family;
523 }
524
525 static int cma_set_qkey(struct rdma_id_private *id_priv, u32 qkey)
526 {
527 struct ib_sa_mcmember_rec rec;
528 int ret = 0;
529
530 if (id_priv->qkey) {
531 if (qkey && id_priv->qkey != qkey)
532 return -EINVAL;
533 return 0;
534 }
535
536 if (qkey) {
537 id_priv->qkey = qkey;
538 return 0;
539 }
540
541 switch (id_priv->id.ps) {
542 case RDMA_PS_UDP:
543 case RDMA_PS_IB:
544 id_priv->qkey = RDMA_UDP_QKEY;
545 break;
546 case RDMA_PS_IPOIB:
547 ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid);
548 ret = ib_sa_get_mcmember_rec(id_priv->id.device,
549 id_priv->id.port_num, &rec.mgid,
550 &rec);
551 if (!ret)
552 id_priv->qkey = be32_to_cpu(rec.qkey);
553 break;
554 default:
555 break;
556 }
557 return ret;
558 }
559
560 static void cma_translate_ib(struct sockaddr_ib *sib, struct rdma_dev_addr *dev_addr)
561 {
562 dev_addr->dev_type = ARPHRD_INFINIBAND;
563 rdma_addr_set_sgid(dev_addr, (union ib_gid *) &sib->sib_addr);
564 ib_addr_set_pkey(dev_addr, ntohs(sib->sib_pkey));
565 }
566
567 static int cma_translate_addr(struct sockaddr *addr, struct rdma_dev_addr *dev_addr)
568 {
569 int ret;
570
571 if (addr->sa_family != AF_IB) {
572 ret = rdma_translate_ip(addr, dev_addr);
573 } else {
574 cma_translate_ib((struct sockaddr_ib *) addr, dev_addr);
575 ret = 0;
576 }
577
578 return ret;
579 }
580
581 static const struct ib_gid_attr *
582 cma_validate_port(struct ib_device *device, u8 port,
583 enum ib_gid_type gid_type,
584 union ib_gid *gid,
585 struct rdma_id_private *id_priv)
586 {
587 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
588 int bound_if_index = dev_addr->bound_dev_if;
589 const struct ib_gid_attr *sgid_attr;
590 int dev_type = dev_addr->dev_type;
591 struct net_device *ndev = NULL;
592
593 if (!rdma_dev_access_netns(device, id_priv->id.route.addr.dev_addr.net))
594 return ERR_PTR(-ENODEV);
595
596 if ((dev_type == ARPHRD_INFINIBAND) && !rdma_protocol_ib(device, port))
597 return ERR_PTR(-ENODEV);
598
599 if ((dev_type != ARPHRD_INFINIBAND) && rdma_protocol_ib(device, port))
600 return ERR_PTR(-ENODEV);
601
602 if (dev_type == ARPHRD_ETHER && rdma_protocol_roce(device, port)) {
603 ndev = dev_get_by_index(dev_addr->net, bound_if_index);
604 if (!ndev)
605 return ERR_PTR(-ENODEV);
606 } else {
607 gid_type = IB_GID_TYPE_IB;
608 }
609
610 sgid_attr = rdma_find_gid_by_port(device, gid, gid_type, port, ndev);
611 if (ndev)
612 dev_put(ndev);
613 return sgid_attr;
614 }
615
616 static void cma_bind_sgid_attr(struct rdma_id_private *id_priv,
617 const struct ib_gid_attr *sgid_attr)
618 {
619 WARN_ON(id_priv->id.route.addr.dev_addr.sgid_attr);
620 id_priv->id.route.addr.dev_addr.sgid_attr = sgid_attr;
621 }
622
623 /**
624 * cma_acquire_dev_by_src_ip - Acquire cma device, port, gid attribute
625 * based on source ip address.
626 * @id_priv: cm_id which should be bound to cma device
627 *
628 * cma_acquire_dev_by_src_ip() binds cm id to cma device, port and GID attribute
629 * based on source IP address. It returns 0 on success or error code otherwise.
630 * It is applicable to active and passive side cm_id.
631 */
632 static int cma_acquire_dev_by_src_ip(struct rdma_id_private *id_priv)
633 {
634 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
635 const struct ib_gid_attr *sgid_attr;
636 union ib_gid gid, iboe_gid, *gidp;
637 struct cma_device *cma_dev;
638 enum ib_gid_type gid_type;
639 int ret = -ENODEV;
640 unsigned int port;
641
642 if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
643 id_priv->id.ps == RDMA_PS_IPOIB)
644 return -EINVAL;
645
646 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
647 &iboe_gid);
648
649 memcpy(&gid, dev_addr->src_dev_addr +
650 rdma_addr_gid_offset(dev_addr), sizeof(gid));
651
652 mutex_lock(&lock);
653 list_for_each_entry(cma_dev, &dev_list, list) {
654 rdma_for_each_port (cma_dev->device, port) {
655 gidp = rdma_protocol_roce(cma_dev->device, port) ?
656 &iboe_gid : &gid;
657 gid_type = cma_dev->default_gid_type[port - 1];
658 sgid_attr = cma_validate_port(cma_dev->device, port,
659 gid_type, gidp, id_priv);
660 if (!IS_ERR(sgid_attr)) {
661 id_priv->id.port_num = port;
662 cma_bind_sgid_attr(id_priv, sgid_attr);
663 cma_attach_to_dev(id_priv, cma_dev);
664 ret = 0;
665 goto out;
666 }
667 }
668 }
669 out:
670 mutex_unlock(&lock);
671 return ret;
672 }
673
674 /**
675 * cma_ib_acquire_dev - Acquire cma device, port and SGID attribute
676 * @id_priv: cm id to bind to cma device
677 * @listen_id_priv: listener cm id to match against
678 * @req: Pointer to req structure containaining incoming
679 * request information
680 * cma_ib_acquire_dev() acquires cma device, port and SGID attribute when
681 * rdma device matches for listen_id and incoming request. It also verifies
682 * that a GID table entry is present for the source address.
683 * Returns 0 on success, or returns error code otherwise.
684 */
685 static int cma_ib_acquire_dev(struct rdma_id_private *id_priv,
686 const struct rdma_id_private *listen_id_priv,
687 struct cma_req_info *req)
688 {
689 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
690 const struct ib_gid_attr *sgid_attr;
691 enum ib_gid_type gid_type;
692 union ib_gid gid;
693
694 if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
695 id_priv->id.ps == RDMA_PS_IPOIB)
696 return -EINVAL;
697
698 if (rdma_protocol_roce(req->device, req->port))
699 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
700 &gid);
701 else
702 memcpy(&gid, dev_addr->src_dev_addr +
703 rdma_addr_gid_offset(dev_addr), sizeof(gid));
704
705 gid_type = listen_id_priv->cma_dev->default_gid_type[req->port - 1];
706 sgid_attr = cma_validate_port(req->device, req->port,
707 gid_type, &gid, id_priv);
708 if (IS_ERR(sgid_attr))
709 return PTR_ERR(sgid_attr);
710
711 id_priv->id.port_num = req->port;
712 cma_bind_sgid_attr(id_priv, sgid_attr);
713 /* Need to acquire lock to protect against reader
714 * of cma_dev->id_list such as cma_netdev_callback() and
715 * cma_process_remove().
716 */
717 mutex_lock(&lock);
718 cma_attach_to_dev(id_priv, listen_id_priv->cma_dev);
719 mutex_unlock(&lock);
720 return 0;
721 }
722
723 static int cma_iw_acquire_dev(struct rdma_id_private *id_priv,
724 const struct rdma_id_private *listen_id_priv)
725 {
726 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
727 const struct ib_gid_attr *sgid_attr;
728 struct cma_device *cma_dev;
729 enum ib_gid_type gid_type;
730 int ret = -ENODEV;
731 unsigned int port;
732 union ib_gid gid;
733
734 if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
735 id_priv->id.ps == RDMA_PS_IPOIB)
736 return -EINVAL;
737
738 memcpy(&gid, dev_addr->src_dev_addr +
739 rdma_addr_gid_offset(dev_addr), sizeof(gid));
740
741 mutex_lock(&lock);
742
743 cma_dev = listen_id_priv->cma_dev;
744 port = listen_id_priv->id.port_num;
745 gid_type = listen_id_priv->gid_type;
746 sgid_attr = cma_validate_port(cma_dev->device, port,
747 gid_type, &gid, id_priv);
748 if (!IS_ERR(sgid_attr)) {
749 id_priv->id.port_num = port;
750 cma_bind_sgid_attr(id_priv, sgid_attr);
751 ret = 0;
752 goto out;
753 }
754
755 list_for_each_entry(cma_dev, &dev_list, list) {
756 rdma_for_each_port (cma_dev->device, port) {
757 if (listen_id_priv->cma_dev == cma_dev &&
758 listen_id_priv->id.port_num == port)
759 continue;
760
761 gid_type = cma_dev->default_gid_type[port - 1];
762 sgid_attr = cma_validate_port(cma_dev->device, port,
763 gid_type, &gid, id_priv);
764 if (!IS_ERR(sgid_attr)) {
765 id_priv->id.port_num = port;
766 cma_bind_sgid_attr(id_priv, sgid_attr);
767 ret = 0;
768 goto out;
769 }
770 }
771 }
772
773 out:
774 if (!ret)
775 cma_attach_to_dev(id_priv, cma_dev);
776
777 mutex_unlock(&lock);
778 return ret;
779 }
780
781 /*
782 * Select the source IB device and address to reach the destination IB address.
783 */
784 static int cma_resolve_ib_dev(struct rdma_id_private *id_priv)
785 {
786 struct cma_device *cma_dev, *cur_dev;
787 struct sockaddr_ib *addr;
788 union ib_gid gid, sgid, *dgid;
789 unsigned int p;
790 u16 pkey, index;
791 enum ib_port_state port_state;
792 int i;
793
794 cma_dev = NULL;
795 addr = (struct sockaddr_ib *) cma_dst_addr(id_priv);
796 dgid = (union ib_gid *) &addr->sib_addr;
797 pkey = ntohs(addr->sib_pkey);
798
799 mutex_lock(&lock);
800 list_for_each_entry(cur_dev, &dev_list, list) {
801 rdma_for_each_port (cur_dev->device, p) {
802 if (!rdma_cap_af_ib(cur_dev->device, p))
803 continue;
804
805 if (ib_find_cached_pkey(cur_dev->device, p, pkey, &index))
806 continue;
807
808 if (ib_get_cached_port_state(cur_dev->device, p, &port_state))
809 continue;
810 for (i = 0; !rdma_query_gid(cur_dev->device,
811 p, i, &gid);
812 i++) {
813 if (!memcmp(&gid, dgid, sizeof(gid))) {
814 cma_dev = cur_dev;
815 sgid = gid;
816 id_priv->id.port_num = p;
817 goto found;
818 }
819
820 if (!cma_dev && (gid.global.subnet_prefix ==
821 dgid->global.subnet_prefix) &&
822 port_state == IB_PORT_ACTIVE) {
823 cma_dev = cur_dev;
824 sgid = gid;
825 id_priv->id.port_num = p;
826 goto found;
827 }
828 }
829 }
830 }
831 mutex_unlock(&lock);
832 return -ENODEV;
833
834 found:
835 cma_attach_to_dev(id_priv, cma_dev);
836 mutex_unlock(&lock);
837 addr = (struct sockaddr_ib *)cma_src_addr(id_priv);
838 memcpy(&addr->sib_addr, &sgid, sizeof(sgid));
839 cma_translate_ib(addr, &id_priv->id.route.addr.dev_addr);
840 return 0;
841 }
842
843 static void cma_id_get(struct rdma_id_private *id_priv)
844 {
845 refcount_inc(&id_priv->refcount);
846 }
847
848 static void cma_id_put(struct rdma_id_private *id_priv)
849 {
850 if (refcount_dec_and_test(&id_priv->refcount))
851 complete(&id_priv->comp);
852 }
853
854 struct rdma_cm_id *__rdma_create_id(struct net *net,
855 rdma_cm_event_handler event_handler,
856 void *context, enum rdma_ucm_port_space ps,
857 enum ib_qp_type qp_type, const char *caller)
858 {
859 struct rdma_id_private *id_priv;
860
861 id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
862 if (!id_priv)
863 return ERR_PTR(-ENOMEM);
864
865 rdma_restrack_set_task(&id_priv->res, caller);
866 id_priv->res.type = RDMA_RESTRACK_CM_ID;
867 id_priv->state = RDMA_CM_IDLE;
868 id_priv->id.context = context;
869 id_priv->id.event_handler = event_handler;
870 id_priv->id.ps = ps;
871 id_priv->id.qp_type = qp_type;
872 id_priv->tos_set = false;
873 id_priv->timeout_set = false;
874 id_priv->gid_type = IB_GID_TYPE_IB;
875 spin_lock_init(&id_priv->lock);
876 mutex_init(&id_priv->qp_mutex);
877 init_completion(&id_priv->comp);
878 refcount_set(&id_priv->refcount, 1);
879 mutex_init(&id_priv->handler_mutex);
880 INIT_LIST_HEAD(&id_priv->listen_list);
881 INIT_LIST_HEAD(&id_priv->mc_list);
882 get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
883 id_priv->id.route.addr.dev_addr.net = get_net(net);
884 id_priv->seq_num &= 0x00ffffff;
885
886 trace_cm_id_create(id_priv);
887 return &id_priv->id;
888 }
889 EXPORT_SYMBOL(__rdma_create_id);
890
891 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
892 {
893 struct ib_qp_attr qp_attr;
894 int qp_attr_mask, ret;
895
896 qp_attr.qp_state = IB_QPS_INIT;
897 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
898 if (ret)
899 return ret;
900
901 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
902 if (ret)
903 return ret;
904
905 qp_attr.qp_state = IB_QPS_RTR;
906 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
907 if (ret)
908 return ret;
909
910 qp_attr.qp_state = IB_QPS_RTS;
911 qp_attr.sq_psn = 0;
912 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
913
914 return ret;
915 }
916
917 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
918 {
919 struct ib_qp_attr qp_attr;
920 int qp_attr_mask, ret;
921
922 qp_attr.qp_state = IB_QPS_INIT;
923 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
924 if (ret)
925 return ret;
926
927 return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
928 }
929
930 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
931 struct ib_qp_init_attr *qp_init_attr)
932 {
933 struct rdma_id_private *id_priv;
934 struct ib_qp *qp;
935 int ret;
936
937 id_priv = container_of(id, struct rdma_id_private, id);
938 if (id->device != pd->device) {
939 ret = -EINVAL;
940 goto out_err;
941 }
942
943 qp_init_attr->port_num = id->port_num;
944 qp = ib_create_qp(pd, qp_init_attr);
945 if (IS_ERR(qp)) {
946 ret = PTR_ERR(qp);
947 goto out_err;
948 }
949
950 if (id->qp_type == IB_QPT_UD)
951 ret = cma_init_ud_qp(id_priv, qp);
952 else
953 ret = cma_init_conn_qp(id_priv, qp);
954 if (ret)
955 goto out_destroy;
956
957 id->qp = qp;
958 id_priv->qp_num = qp->qp_num;
959 id_priv->srq = (qp->srq != NULL);
960 trace_cm_qp_create(id_priv, pd, qp_init_attr, 0);
961 return 0;
962 out_destroy:
963 ib_destroy_qp(qp);
964 out_err:
965 trace_cm_qp_create(id_priv, pd, qp_init_attr, ret);
966 return ret;
967 }
968 EXPORT_SYMBOL(rdma_create_qp);
969
970 void rdma_destroy_qp(struct rdma_cm_id *id)
971 {
972 struct rdma_id_private *id_priv;
973
974 id_priv = container_of(id, struct rdma_id_private, id);
975 trace_cm_qp_destroy(id_priv);
976 mutex_lock(&id_priv->qp_mutex);
977 ib_destroy_qp(id_priv->id.qp);
978 id_priv->id.qp = NULL;
979 mutex_unlock(&id_priv->qp_mutex);
980 }
981 EXPORT_SYMBOL(rdma_destroy_qp);
982
983 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
984 struct rdma_conn_param *conn_param)
985 {
986 struct ib_qp_attr qp_attr;
987 int qp_attr_mask, ret;
988
989 mutex_lock(&id_priv->qp_mutex);
990 if (!id_priv->id.qp) {
991 ret = 0;
992 goto out;
993 }
994
995 /* Need to update QP attributes from default values. */
996 qp_attr.qp_state = IB_QPS_INIT;
997 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
998 if (ret)
999 goto out;
1000
1001 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
1002 if (ret)
1003 goto out;
1004
1005 qp_attr.qp_state = IB_QPS_RTR;
1006 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1007 if (ret)
1008 goto out;
1009
1010 BUG_ON(id_priv->cma_dev->device != id_priv->id.device);
1011
1012 if (conn_param)
1013 qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
1014 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
1015 out:
1016 mutex_unlock(&id_priv->qp_mutex);
1017 return ret;
1018 }
1019
1020 static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
1021 struct rdma_conn_param *conn_param)
1022 {
1023 struct ib_qp_attr qp_attr;
1024 int qp_attr_mask, ret;
1025
1026 mutex_lock(&id_priv->qp_mutex);
1027 if (!id_priv->id.qp) {
1028 ret = 0;
1029 goto out;
1030 }
1031
1032 qp_attr.qp_state = IB_QPS_RTS;
1033 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1034 if (ret)
1035 goto out;
1036
1037 if (conn_param)
1038 qp_attr.max_rd_atomic = conn_param->initiator_depth;
1039 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
1040 out:
1041 mutex_unlock(&id_priv->qp_mutex);
1042 return ret;
1043 }
1044
1045 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
1046 {
1047 struct ib_qp_attr qp_attr;
1048 int ret;
1049
1050 mutex_lock(&id_priv->qp_mutex);
1051 if (!id_priv->id.qp) {
1052 ret = 0;
1053 goto out;
1054 }
1055
1056 qp_attr.qp_state = IB_QPS_ERR;
1057 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
1058 out:
1059 mutex_unlock(&id_priv->qp_mutex);
1060 return ret;
1061 }
1062
1063 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
1064 struct ib_qp_attr *qp_attr, int *qp_attr_mask)
1065 {
1066 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
1067 int ret;
1068 u16 pkey;
1069
1070 if (rdma_cap_eth_ah(id_priv->id.device, id_priv->id.port_num))
1071 pkey = 0xffff;
1072 else
1073 pkey = ib_addr_get_pkey(dev_addr);
1074
1075 ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
1076 pkey, &qp_attr->pkey_index);
1077 if (ret)
1078 return ret;
1079
1080 qp_attr->port_num = id_priv->id.port_num;
1081 *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
1082
1083 if (id_priv->id.qp_type == IB_QPT_UD) {
1084 ret = cma_set_qkey(id_priv, 0);
1085 if (ret)
1086 return ret;
1087
1088 qp_attr->qkey = id_priv->qkey;
1089 *qp_attr_mask |= IB_QP_QKEY;
1090 } else {
1091 qp_attr->qp_access_flags = 0;
1092 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
1093 }
1094 return 0;
1095 }
1096
1097 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
1098 int *qp_attr_mask)
1099 {
1100 struct rdma_id_private *id_priv;
1101 int ret = 0;
1102
1103 id_priv = container_of(id, struct rdma_id_private, id);
1104 if (rdma_cap_ib_cm(id->device, id->port_num)) {
1105 if (!id_priv->cm_id.ib || (id_priv->id.qp_type == IB_QPT_UD))
1106 ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
1107 else
1108 ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
1109 qp_attr_mask);
1110
1111 if (qp_attr->qp_state == IB_QPS_RTR)
1112 qp_attr->rq_psn = id_priv->seq_num;
1113 } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
1114 if (!id_priv->cm_id.iw) {
1115 qp_attr->qp_access_flags = 0;
1116 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
1117 } else
1118 ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
1119 qp_attr_mask);
1120 qp_attr->port_num = id_priv->id.port_num;
1121 *qp_attr_mask |= IB_QP_PORT;
1122 } else
1123 ret = -ENOSYS;
1124
1125 if ((*qp_attr_mask & IB_QP_TIMEOUT) && id_priv->timeout_set)
1126 qp_attr->timeout = id_priv->timeout;
1127
1128 return ret;
1129 }
1130 EXPORT_SYMBOL(rdma_init_qp_attr);
1131
1132 static inline bool cma_zero_addr(const struct sockaddr *addr)
1133 {
1134 switch (addr->sa_family) {
1135 case AF_INET:
1136 return ipv4_is_zeronet(((struct sockaddr_in *)addr)->sin_addr.s_addr);
1137 case AF_INET6:
1138 return ipv6_addr_any(&((struct sockaddr_in6 *)addr)->sin6_addr);
1139 case AF_IB:
1140 return ib_addr_any(&((struct sockaddr_ib *)addr)->sib_addr);
1141 default:
1142 return false;
1143 }
1144 }
1145
1146 static inline bool cma_loopback_addr(const struct sockaddr *addr)
1147 {
1148 switch (addr->sa_family) {
1149 case AF_INET:
1150 return ipv4_is_loopback(
1151 ((struct sockaddr_in *)addr)->sin_addr.s_addr);
1152 case AF_INET6:
1153 return ipv6_addr_loopback(
1154 &((struct sockaddr_in6 *)addr)->sin6_addr);
1155 case AF_IB:
1156 return ib_addr_loopback(
1157 &((struct sockaddr_ib *)addr)->sib_addr);
1158 default:
1159 return false;
1160 }
1161 }
1162
1163 static inline bool cma_any_addr(const struct sockaddr *addr)
1164 {
1165 return cma_zero_addr(addr) || cma_loopback_addr(addr);
1166 }
1167
1168 static int cma_addr_cmp(const struct sockaddr *src, const struct sockaddr *dst)
1169 {
1170 if (src->sa_family != dst->sa_family)
1171 return -1;
1172
1173 switch (src->sa_family) {
1174 case AF_INET:
1175 return ((struct sockaddr_in *)src)->sin_addr.s_addr !=
1176 ((struct sockaddr_in *)dst)->sin_addr.s_addr;
1177 case AF_INET6: {
1178 struct sockaddr_in6 *src_addr6 = (struct sockaddr_in6 *)src;
1179 struct sockaddr_in6 *dst_addr6 = (struct sockaddr_in6 *)dst;
1180 bool link_local;
1181
1182 if (ipv6_addr_cmp(&src_addr6->sin6_addr,
1183 &dst_addr6->sin6_addr))
1184 return 1;
1185 link_local = ipv6_addr_type(&dst_addr6->sin6_addr) &
1186 IPV6_ADDR_LINKLOCAL;
1187 /* Link local must match their scope_ids */
1188 return link_local ? (src_addr6->sin6_scope_id !=
1189 dst_addr6->sin6_scope_id) :
1190 0;
1191 }
1192
1193 default:
1194 return ib_addr_cmp(&((struct sockaddr_ib *) src)->sib_addr,
1195 &((struct sockaddr_ib *) dst)->sib_addr);
1196 }
1197 }
1198
1199 static __be16 cma_port(const struct sockaddr *addr)
1200 {
1201 struct sockaddr_ib *sib;
1202
1203 switch (addr->sa_family) {
1204 case AF_INET:
1205 return ((struct sockaddr_in *) addr)->sin_port;
1206 case AF_INET6:
1207 return ((struct sockaddr_in6 *) addr)->sin6_port;
1208 case AF_IB:
1209 sib = (struct sockaddr_ib *) addr;
1210 return htons((u16) (be64_to_cpu(sib->sib_sid) &
1211 be64_to_cpu(sib->sib_sid_mask)));
1212 default:
1213 return 0;
1214 }
1215 }
1216
1217 static inline int cma_any_port(const struct sockaddr *addr)
1218 {
1219 return !cma_port(addr);
1220 }
1221
1222 static void cma_save_ib_info(struct sockaddr *src_addr,
1223 struct sockaddr *dst_addr,
1224 const struct rdma_cm_id *listen_id,
1225 const struct sa_path_rec *path)
1226 {
1227 struct sockaddr_ib *listen_ib, *ib;
1228
1229 listen_ib = (struct sockaddr_ib *) &listen_id->route.addr.src_addr;
1230 if (src_addr) {
1231 ib = (struct sockaddr_ib *)src_addr;
1232 ib->sib_family = AF_IB;
1233 if (path) {
1234 ib->sib_pkey = path->pkey;
1235 ib->sib_flowinfo = path->flow_label;
1236 memcpy(&ib->sib_addr, &path->sgid, 16);
1237 ib->sib_sid = path->service_id;
1238 ib->sib_scope_id = 0;
1239 } else {
1240 ib->sib_pkey = listen_ib->sib_pkey;
1241 ib->sib_flowinfo = listen_ib->sib_flowinfo;
1242 ib->sib_addr = listen_ib->sib_addr;
1243 ib->sib_sid = listen_ib->sib_sid;
1244 ib->sib_scope_id = listen_ib->sib_scope_id;
1245 }
1246 ib->sib_sid_mask = cpu_to_be64(0xffffffffffffffffULL);
1247 }
1248 if (dst_addr) {
1249 ib = (struct sockaddr_ib *)dst_addr;
1250 ib->sib_family = AF_IB;
1251 if (path) {
1252 ib->sib_pkey = path->pkey;
1253 ib->sib_flowinfo = path->flow_label;
1254 memcpy(&ib->sib_addr, &path->dgid, 16);
1255 }
1256 }
1257 }
1258
1259 static void cma_save_ip4_info(struct sockaddr_in *src_addr,
1260 struct sockaddr_in *dst_addr,
1261 struct cma_hdr *hdr,
1262 __be16 local_port)
1263 {
1264 if (src_addr) {
1265 *src_addr = (struct sockaddr_in) {
1266 .sin_family = AF_INET,
1267 .sin_addr.s_addr = hdr->dst_addr.ip4.addr,
1268 .sin_port = local_port,
1269 };
1270 }
1271
1272 if (dst_addr) {
1273 *dst_addr = (struct sockaddr_in) {
1274 .sin_family = AF_INET,
1275 .sin_addr.s_addr = hdr->src_addr.ip4.addr,
1276 .sin_port = hdr->port,
1277 };
1278 }
1279 }
1280
1281 static void cma_save_ip6_info(struct sockaddr_in6 *src_addr,
1282 struct sockaddr_in6 *dst_addr,
1283 struct cma_hdr *hdr,
1284 __be16 local_port)
1285 {
1286 if (src_addr) {
1287 *src_addr = (struct sockaddr_in6) {
1288 .sin6_family = AF_INET6,
1289 .sin6_addr = hdr->dst_addr.ip6,
1290 .sin6_port = local_port,
1291 };
1292 }
1293
1294 if (dst_addr) {
1295 *dst_addr = (struct sockaddr_in6) {
1296 .sin6_family = AF_INET6,
1297 .sin6_addr = hdr->src_addr.ip6,
1298 .sin6_port = hdr->port,
1299 };
1300 }
1301 }
1302
1303 static u16 cma_port_from_service_id(__be64 service_id)
1304 {
1305 return (u16)be64_to_cpu(service_id);
1306 }
1307
1308 static int cma_save_ip_info(struct sockaddr *src_addr,
1309 struct sockaddr *dst_addr,
1310 const struct ib_cm_event *ib_event,
1311 __be64 service_id)
1312 {
1313 struct cma_hdr *hdr;
1314 __be16 port;
1315
1316 hdr = ib_event->private_data;
1317 if (hdr->cma_version != CMA_VERSION)
1318 return -EINVAL;
1319
1320 port = htons(cma_port_from_service_id(service_id));
1321
1322 switch (cma_get_ip_ver(hdr)) {
1323 case 4:
1324 cma_save_ip4_info((struct sockaddr_in *)src_addr,
1325 (struct sockaddr_in *)dst_addr, hdr, port);
1326 break;
1327 case 6:
1328 cma_save_ip6_info((struct sockaddr_in6 *)src_addr,
1329 (struct sockaddr_in6 *)dst_addr, hdr, port);
1330 break;
1331 default:
1332 return -EAFNOSUPPORT;
1333 }
1334
1335 return 0;
1336 }
1337
1338 static int cma_save_net_info(struct sockaddr *src_addr,
1339 struct sockaddr *dst_addr,
1340 const struct rdma_cm_id *listen_id,
1341 const struct ib_cm_event *ib_event,
1342 sa_family_t sa_family, __be64 service_id)
1343 {
1344 if (sa_family == AF_IB) {
1345 if (ib_event->event == IB_CM_REQ_RECEIVED)
1346 cma_save_ib_info(src_addr, dst_addr, listen_id,
1347 ib_event->param.req_rcvd.primary_path);
1348 else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED)
1349 cma_save_ib_info(src_addr, dst_addr, listen_id, NULL);
1350 return 0;
1351 }
1352
1353 return cma_save_ip_info(src_addr, dst_addr, ib_event, service_id);
1354 }
1355
1356 static int cma_save_req_info(const struct ib_cm_event *ib_event,
1357 struct cma_req_info *req)
1358 {
1359 const struct ib_cm_req_event_param *req_param =
1360 &ib_event->param.req_rcvd;
1361 const struct ib_cm_sidr_req_event_param *sidr_param =
1362 &ib_event->param.sidr_req_rcvd;
1363
1364 switch (ib_event->event) {
1365 case IB_CM_REQ_RECEIVED:
1366 req->device = req_param->listen_id->device;
1367 req->port = req_param->port;
1368 memcpy(&req->local_gid, &req_param->primary_path->sgid,
1369 sizeof(req->local_gid));
1370 req->has_gid = true;
1371 req->service_id = req_param->primary_path->service_id;
1372 req->pkey = be16_to_cpu(req_param->primary_path->pkey);
1373 if (req->pkey != req_param->bth_pkey)
1374 pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and primary path P_Key (0x%x)\n"
1375 "RDMA CMA: in the future this may cause the request to be dropped\n",
1376 req_param->bth_pkey, req->pkey);
1377 break;
1378 case IB_CM_SIDR_REQ_RECEIVED:
1379 req->device = sidr_param->listen_id->device;
1380 req->port = sidr_param->port;
1381 req->has_gid = false;
1382 req->service_id = sidr_param->service_id;
1383 req->pkey = sidr_param->pkey;
1384 if (req->pkey != sidr_param->bth_pkey)
1385 pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and SIDR request payload P_Key (0x%x)\n"
1386 "RDMA CMA: in the future this may cause the request to be dropped\n",
1387 sidr_param->bth_pkey, req->pkey);
1388 break;
1389 default:
1390 return -EINVAL;
1391 }
1392
1393 return 0;
1394 }
1395
1396 static bool validate_ipv4_net_dev(struct net_device *net_dev,
1397 const struct sockaddr_in *dst_addr,
1398 const struct sockaddr_in *src_addr)
1399 {
1400 __be32 daddr = dst_addr->sin_addr.s_addr,
1401 saddr = src_addr->sin_addr.s_addr;
1402 struct fib_result res;
1403 struct flowi4 fl4;
1404 int err;
1405 bool ret;
1406
1407 if (ipv4_is_multicast(saddr) || ipv4_is_lbcast(saddr) ||
1408 ipv4_is_lbcast(daddr) || ipv4_is_zeronet(saddr) ||
1409 ipv4_is_zeronet(daddr) || ipv4_is_loopback(daddr) ||
1410 ipv4_is_loopback(saddr))
1411 return false;
1412
1413 memset(&fl4, 0, sizeof(fl4));
1414 fl4.flowi4_iif = net_dev->ifindex;
1415 fl4.daddr = daddr;
1416 fl4.saddr = saddr;
1417
1418 rcu_read_lock();
1419 err = fib_lookup(dev_net(net_dev), &fl4, &res, 0);
1420 ret = err == 0 && FIB_RES_DEV(res) == net_dev;
1421 rcu_read_unlock();
1422
1423 return ret;
1424 }
1425
1426 static bool validate_ipv6_net_dev(struct net_device *net_dev,
1427 const struct sockaddr_in6 *dst_addr,
1428 const struct sockaddr_in6 *src_addr)
1429 {
1430 #if IS_ENABLED(CONFIG_IPV6)
1431 const int strict = ipv6_addr_type(&dst_addr->sin6_addr) &
1432 IPV6_ADDR_LINKLOCAL;
1433 struct rt6_info *rt = rt6_lookup(dev_net(net_dev), &dst_addr->sin6_addr,
1434 &src_addr->sin6_addr, net_dev->ifindex,
1435 NULL, strict);
1436 bool ret;
1437
1438 if (!rt)
1439 return false;
1440
1441 ret = rt->rt6i_idev->dev == net_dev;
1442 ip6_rt_put(rt);
1443
1444 return ret;
1445 #else
1446 return false;
1447 #endif
1448 }
1449
1450 static bool validate_net_dev(struct net_device *net_dev,
1451 const struct sockaddr *daddr,
1452 const struct sockaddr *saddr)
1453 {
1454 const struct sockaddr_in *daddr4 = (const struct sockaddr_in *)daddr;
1455 const struct sockaddr_in *saddr4 = (const struct sockaddr_in *)saddr;
1456 const struct sockaddr_in6 *daddr6 = (const struct sockaddr_in6 *)daddr;
1457 const struct sockaddr_in6 *saddr6 = (const struct sockaddr_in6 *)saddr;
1458
1459 switch (daddr->sa_family) {
1460 case AF_INET:
1461 return saddr->sa_family == AF_INET &&
1462 validate_ipv4_net_dev(net_dev, daddr4, saddr4);
1463
1464 case AF_INET6:
1465 return saddr->sa_family == AF_INET6 &&
1466 validate_ipv6_net_dev(net_dev, daddr6, saddr6);
1467
1468 default:
1469 return false;
1470 }
1471 }
1472
1473 static struct net_device *
1474 roce_get_net_dev_by_cm_event(const struct ib_cm_event *ib_event)
1475 {
1476 const struct ib_gid_attr *sgid_attr = NULL;
1477 struct net_device *ndev;
1478
1479 if (ib_event->event == IB_CM_REQ_RECEIVED)
1480 sgid_attr = ib_event->param.req_rcvd.ppath_sgid_attr;
1481 else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED)
1482 sgid_attr = ib_event->param.sidr_req_rcvd.sgid_attr;
1483
1484 if (!sgid_attr)
1485 return NULL;
1486
1487 rcu_read_lock();
1488 ndev = rdma_read_gid_attr_ndev_rcu(sgid_attr);
1489 if (IS_ERR(ndev))
1490 ndev = NULL;
1491 else
1492 dev_hold(ndev);
1493 rcu_read_unlock();
1494 return ndev;
1495 }
1496
1497 static struct net_device *cma_get_net_dev(const struct ib_cm_event *ib_event,
1498 struct cma_req_info *req)
1499 {
1500 struct sockaddr *listen_addr =
1501 (struct sockaddr *)&req->listen_addr_storage;
1502 struct sockaddr *src_addr = (struct sockaddr *)&req->src_addr_storage;
1503 struct net_device *net_dev;
1504 const union ib_gid *gid = req->has_gid ? &req->local_gid : NULL;
1505 int err;
1506
1507 err = cma_save_ip_info(listen_addr, src_addr, ib_event,
1508 req->service_id);
1509 if (err)
1510 return ERR_PTR(err);
1511
1512 if (rdma_protocol_roce(req->device, req->port))
1513 net_dev = roce_get_net_dev_by_cm_event(ib_event);
1514 else
1515 net_dev = ib_get_net_dev_by_params(req->device, req->port,
1516 req->pkey,
1517 gid, listen_addr);
1518 if (!net_dev)
1519 return ERR_PTR(-ENODEV);
1520
1521 return net_dev;
1522 }
1523
1524 static enum rdma_ucm_port_space rdma_ps_from_service_id(__be64 service_id)
1525 {
1526 return (be64_to_cpu(service_id) >> 16) & 0xffff;
1527 }
1528
1529 static bool cma_match_private_data(struct rdma_id_private *id_priv,
1530 const struct cma_hdr *hdr)
1531 {
1532 struct sockaddr *addr = cma_src_addr(id_priv);
1533 __be32 ip4_addr;
1534 struct in6_addr ip6_addr;
1535
1536 if (cma_any_addr(addr) && !id_priv->afonly)
1537 return true;
1538
1539 switch (addr->sa_family) {
1540 case AF_INET:
1541 ip4_addr = ((struct sockaddr_in *)addr)->sin_addr.s_addr;
1542 if (cma_get_ip_ver(hdr) != 4)
1543 return false;
1544 if (!cma_any_addr(addr) &&
1545 hdr->dst_addr.ip4.addr != ip4_addr)
1546 return false;
1547 break;
1548 case AF_INET6:
1549 ip6_addr = ((struct sockaddr_in6 *)addr)->sin6_addr;
1550 if (cma_get_ip_ver(hdr) != 6)
1551 return false;
1552 if (!cma_any_addr(addr) &&
1553 memcmp(&hdr->dst_addr.ip6, &ip6_addr, sizeof(ip6_addr)))
1554 return false;
1555 break;
1556 case AF_IB:
1557 return true;
1558 default:
1559 return false;
1560 }
1561
1562 return true;
1563 }
1564
1565 static bool cma_protocol_roce(const struct rdma_cm_id *id)
1566 {
1567 struct ib_device *device = id->device;
1568 const int port_num = id->port_num ?: rdma_start_port(device);
1569
1570 return rdma_protocol_roce(device, port_num);
1571 }
1572
1573 static bool cma_is_req_ipv6_ll(const struct cma_req_info *req)
1574 {
1575 const struct sockaddr *daddr =
1576 (const struct sockaddr *)&req->listen_addr_storage;
1577 const struct sockaddr_in6 *daddr6 = (const struct sockaddr_in6 *)daddr;
1578
1579 /* Returns true if the req is for IPv6 link local */
1580 return (daddr->sa_family == AF_INET6 &&
1581 (ipv6_addr_type(&daddr6->sin6_addr) & IPV6_ADDR_LINKLOCAL));
1582 }
1583
1584 static bool cma_match_net_dev(const struct rdma_cm_id *id,
1585 const struct net_device *net_dev,
1586 const struct cma_req_info *req)
1587 {
1588 const struct rdma_addr *addr = &id->route.addr;
1589
1590 if (!net_dev)
1591 /* This request is an AF_IB request */
1592 return (!id->port_num || id->port_num == req->port) &&
1593 (addr->src_addr.ss_family == AF_IB);
1594
1595 /*
1596 * If the request is not for IPv6 link local, allow matching
1597 * request to any netdevice of the one or multiport rdma device.
1598 */
1599 if (!cma_is_req_ipv6_ll(req))
1600 return true;
1601 /*
1602 * Net namespaces must match, and if the listner is listening
1603 * on a specific netdevice than netdevice must match as well.
1604 */
1605 if (net_eq(dev_net(net_dev), addr->dev_addr.net) &&
1606 (!!addr->dev_addr.bound_dev_if ==
1607 (addr->dev_addr.bound_dev_if == net_dev->ifindex)))
1608 return true;
1609 else
1610 return false;
1611 }
1612
1613 static struct rdma_id_private *cma_find_listener(
1614 const struct rdma_bind_list *bind_list,
1615 const struct ib_cm_id *cm_id,
1616 const struct ib_cm_event *ib_event,
1617 const struct cma_req_info *req,
1618 const struct net_device *net_dev)
1619 {
1620 struct rdma_id_private *id_priv, *id_priv_dev;
1621
1622 lockdep_assert_held(&lock);
1623
1624 if (!bind_list)
1625 return ERR_PTR(-EINVAL);
1626
1627 hlist_for_each_entry(id_priv, &bind_list->owners, node) {
1628 if (cma_match_private_data(id_priv, ib_event->private_data)) {
1629 if (id_priv->id.device == cm_id->device &&
1630 cma_match_net_dev(&id_priv->id, net_dev, req))
1631 return id_priv;
1632 list_for_each_entry(id_priv_dev,
1633 &id_priv->listen_list,
1634 listen_list) {
1635 if (id_priv_dev->id.device == cm_id->device &&
1636 cma_match_net_dev(&id_priv_dev->id,
1637 net_dev, req))
1638 return id_priv_dev;
1639 }
1640 }
1641 }
1642
1643 return ERR_PTR(-EINVAL);
1644 }
1645
1646 static struct rdma_id_private *
1647 cma_ib_id_from_event(struct ib_cm_id *cm_id,
1648 const struct ib_cm_event *ib_event,
1649 struct cma_req_info *req,
1650 struct net_device **net_dev)
1651 {
1652 struct rdma_bind_list *bind_list;
1653 struct rdma_id_private *id_priv;
1654 int err;
1655
1656 err = cma_save_req_info(ib_event, req);
1657 if (err)
1658 return ERR_PTR(err);
1659
1660 *net_dev = cma_get_net_dev(ib_event, req);
1661 if (IS_ERR(*net_dev)) {
1662 if (PTR_ERR(*net_dev) == -EAFNOSUPPORT) {
1663 /* Assuming the protocol is AF_IB */
1664 *net_dev = NULL;
1665 } else {
1666 return ERR_CAST(*net_dev);
1667 }
1668 }
1669
1670 mutex_lock(&lock);
1671 /*
1672 * Net namespace might be getting deleted while route lookup,
1673 * cm_id lookup is in progress. Therefore, perform netdevice
1674 * validation, cm_id lookup under rcu lock.
1675 * RCU lock along with netdevice state check, synchronizes with
1676 * netdevice migrating to different net namespace and also avoids
1677 * case where net namespace doesn't get deleted while lookup is in
1678 * progress.
1679 * If the device state is not IFF_UP, its properties such as ifindex
1680 * and nd_net cannot be trusted to remain valid without rcu lock.
1681 * net/core/dev.c change_net_namespace() ensures to synchronize with
1682 * ongoing operations on net device after device is closed using
1683 * synchronize_net().
1684 */
1685 rcu_read_lock();
1686 if (*net_dev) {
1687 /*
1688 * If netdevice is down, it is likely that it is administratively
1689 * down or it might be migrating to different namespace.
1690 * In that case avoid further processing, as the net namespace
1691 * or ifindex may change.
1692 */
1693 if (((*net_dev)->flags & IFF_UP) == 0) {
1694 id_priv = ERR_PTR(-EHOSTUNREACH);
1695 goto err;
1696 }
1697
1698 if (!validate_net_dev(*net_dev,
1699 (struct sockaddr *)&req->listen_addr_storage,
1700 (struct sockaddr *)&req->src_addr_storage)) {
1701 id_priv = ERR_PTR(-EHOSTUNREACH);
1702 goto err;
1703 }
1704 }
1705
1706 bind_list = cma_ps_find(*net_dev ? dev_net(*net_dev) : &init_net,
1707 rdma_ps_from_service_id(req->service_id),
1708 cma_port_from_service_id(req->service_id));
1709 id_priv = cma_find_listener(bind_list, cm_id, ib_event, req, *net_dev);
1710 err:
1711 rcu_read_unlock();
1712 mutex_unlock(&lock);
1713 if (IS_ERR(id_priv) && *net_dev) {
1714 dev_put(*net_dev);
1715 *net_dev = NULL;
1716 }
1717 return id_priv;
1718 }
1719
1720 static inline u8 cma_user_data_offset(struct rdma_id_private *id_priv)
1721 {
1722 return cma_family(id_priv) == AF_IB ? 0 : sizeof(struct cma_hdr);
1723 }
1724
1725 static void cma_cancel_route(struct rdma_id_private *id_priv)
1726 {
1727 if (rdma_cap_ib_sa(id_priv->id.device, id_priv->id.port_num)) {
1728 if (id_priv->query)
1729 ib_sa_cancel_query(id_priv->query_id, id_priv->query);
1730 }
1731 }
1732
1733 static void cma_cancel_listens(struct rdma_id_private *id_priv)
1734 {
1735 struct rdma_id_private *dev_id_priv;
1736
1737 /*
1738 * Remove from listen_any_list to prevent added devices from spawning
1739 * additional listen requests.
1740 */
1741 mutex_lock(&lock);
1742 list_del(&id_priv->list);
1743
1744 while (!list_empty(&id_priv->listen_list)) {
1745 dev_id_priv = list_entry(id_priv->listen_list.next,
1746 struct rdma_id_private, listen_list);
1747 /* sync with device removal to avoid duplicate destruction */
1748 list_del_init(&dev_id_priv->list);
1749 list_del(&dev_id_priv->listen_list);
1750 mutex_unlock(&lock);
1751
1752 rdma_destroy_id(&dev_id_priv->id);
1753 mutex_lock(&lock);
1754 }
1755 mutex_unlock(&lock);
1756 }
1757
1758 static void cma_cancel_operation(struct rdma_id_private *id_priv,
1759 enum rdma_cm_state state)
1760 {
1761 switch (state) {
1762 case RDMA_CM_ADDR_QUERY:
1763 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
1764 break;
1765 case RDMA_CM_ROUTE_QUERY:
1766 cma_cancel_route(id_priv);
1767 break;
1768 case RDMA_CM_LISTEN:
1769 if (cma_any_addr(cma_src_addr(id_priv)) && !id_priv->cma_dev)
1770 cma_cancel_listens(id_priv);
1771 break;
1772 default:
1773 break;
1774 }
1775 }
1776
1777 static void cma_release_port(struct rdma_id_private *id_priv)
1778 {
1779 struct rdma_bind_list *bind_list = id_priv->bind_list;
1780 struct net *net = id_priv->id.route.addr.dev_addr.net;
1781
1782 if (!bind_list)
1783 return;
1784
1785 mutex_lock(&lock);
1786 hlist_del(&id_priv->node);
1787 if (hlist_empty(&bind_list->owners)) {
1788 cma_ps_remove(net, bind_list->ps, bind_list->port);
1789 kfree(bind_list);
1790 }
1791 mutex_unlock(&lock);
1792 }
1793
1794 static void cma_leave_roce_mc_group(struct rdma_id_private *id_priv,
1795 struct cma_multicast *mc)
1796 {
1797 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
1798 struct net_device *ndev = NULL;
1799
1800 if (dev_addr->bound_dev_if)
1801 ndev = dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if);
1802 if (ndev) {
1803 cma_igmp_send(ndev, &mc->multicast.ib->rec.mgid, false);
1804 dev_put(ndev);
1805 }
1806 kref_put(&mc->mcref, release_mc);
1807 }
1808
1809 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
1810 {
1811 struct cma_multicast *mc;
1812
1813 while (!list_empty(&id_priv->mc_list)) {
1814 mc = container_of(id_priv->mc_list.next,
1815 struct cma_multicast, list);
1816 list_del(&mc->list);
1817 if (rdma_cap_ib_mcast(id_priv->cma_dev->device,
1818 id_priv->id.port_num)) {
1819 ib_sa_free_multicast(mc->multicast.ib);
1820 kfree(mc);
1821 } else {
1822 cma_leave_roce_mc_group(id_priv, mc);
1823 }
1824 }
1825 }
1826
1827 void rdma_destroy_id(struct rdma_cm_id *id)
1828 {
1829 struct rdma_id_private *id_priv;
1830 enum rdma_cm_state state;
1831
1832 id_priv = container_of(id, struct rdma_id_private, id);
1833 trace_cm_id_destroy(id_priv);
1834 state = cma_exch(id_priv, RDMA_CM_DESTROYING);
1835 cma_cancel_operation(id_priv, state);
1836
1837 /*
1838 * Wait for any active callback to finish. New callbacks will find
1839 * the id_priv state set to destroying and abort.
1840 */
1841 mutex_lock(&id_priv->handler_mutex);
1842 mutex_unlock(&id_priv->handler_mutex);
1843
1844 rdma_restrack_del(&id_priv->res);
1845 if (id_priv->cma_dev) {
1846 if (rdma_cap_ib_cm(id_priv->id.device, 1)) {
1847 if (id_priv->cm_id.ib)
1848 ib_destroy_cm_id(id_priv->cm_id.ib);
1849 } else if (rdma_cap_iw_cm(id_priv->id.device, 1)) {
1850 if (id_priv->cm_id.iw)
1851 iw_destroy_cm_id(id_priv->cm_id.iw);
1852 }
1853 cma_leave_mc_groups(id_priv);
1854 cma_release_dev(id_priv);
1855 }
1856
1857 cma_release_port(id_priv);
1858 cma_id_put(id_priv);
1859 wait_for_completion(&id_priv->comp);
1860
1861 if (id_priv->internal_id)
1862 cma_id_put(id_priv->id.context);
1863
1864 kfree(id_priv->id.route.path_rec);
1865
1866 if (id_priv->id.route.addr.dev_addr.sgid_attr)
1867 rdma_put_gid_attr(id_priv->id.route.addr.dev_addr.sgid_attr);
1868
1869 put_net(id_priv->id.route.addr.dev_addr.net);
1870 kfree(id_priv);
1871 }
1872 EXPORT_SYMBOL(rdma_destroy_id);
1873
1874 static int cma_rep_recv(struct rdma_id_private *id_priv)
1875 {
1876 int ret;
1877
1878 ret = cma_modify_qp_rtr(id_priv, NULL);
1879 if (ret)
1880 goto reject;
1881
1882 ret = cma_modify_qp_rts(id_priv, NULL);
1883 if (ret)
1884 goto reject;
1885
1886 trace_cm_send_rtu(id_priv);
1887 ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
1888 if (ret)
1889 goto reject;
1890
1891 return 0;
1892 reject:
1893 pr_debug_ratelimited("RDMA CM: CONNECT_ERROR: failed to handle reply. status %d\n", ret);
1894 cma_modify_qp_err(id_priv);
1895 trace_cm_send_rej(id_priv);
1896 ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
1897 NULL, 0, NULL, 0);
1898 return ret;
1899 }
1900
1901 static void cma_set_rep_event_data(struct rdma_cm_event *event,
1902 const struct ib_cm_rep_event_param *rep_data,
1903 void *private_data)
1904 {
1905 event->param.conn.private_data = private_data;
1906 event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
1907 event->param.conn.responder_resources = rep_data->responder_resources;
1908 event->param.conn.initiator_depth = rep_data->initiator_depth;
1909 event->param.conn.flow_control = rep_data->flow_control;
1910 event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
1911 event->param.conn.srq = rep_data->srq;
1912 event->param.conn.qp_num = rep_data->remote_qpn;
1913 }
1914
1915 static int cma_cm_event_handler(struct rdma_id_private *id_priv,
1916 struct rdma_cm_event *event)
1917 {
1918 int ret;
1919
1920 trace_cm_event_handler(id_priv, event);
1921 ret = id_priv->id.event_handler(&id_priv->id, event);
1922 trace_cm_event_done(id_priv, event, ret);
1923 return ret;
1924 }
1925
1926 static int cma_ib_handler(struct ib_cm_id *cm_id,
1927 const struct ib_cm_event *ib_event)
1928 {
1929 struct rdma_id_private *id_priv = cm_id->context;
1930 struct rdma_cm_event event = {};
1931 int ret = 0;
1932
1933 mutex_lock(&id_priv->handler_mutex);
1934 if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
1935 id_priv->state != RDMA_CM_CONNECT) ||
1936 (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
1937 id_priv->state != RDMA_CM_DISCONNECT))
1938 goto out;
1939
1940 switch (ib_event->event) {
1941 case IB_CM_REQ_ERROR:
1942 case IB_CM_REP_ERROR:
1943 event.event = RDMA_CM_EVENT_UNREACHABLE;
1944 event.status = -ETIMEDOUT;
1945 break;
1946 case IB_CM_REP_RECEIVED:
1947 if (cma_comp(id_priv, RDMA_CM_CONNECT) &&
1948 (id_priv->id.qp_type != IB_QPT_UD)) {
1949 trace_cm_send_mra(id_priv);
1950 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
1951 }
1952 if (id_priv->id.qp) {
1953 event.status = cma_rep_recv(id_priv);
1954 event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
1955 RDMA_CM_EVENT_ESTABLISHED;
1956 } else {
1957 event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
1958 }
1959 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
1960 ib_event->private_data);
1961 break;
1962 case IB_CM_RTU_RECEIVED:
1963 case IB_CM_USER_ESTABLISHED:
1964 event.event = RDMA_CM_EVENT_ESTABLISHED;
1965 break;
1966 case IB_CM_DREQ_ERROR:
1967 event.status = -ETIMEDOUT; /* fall through */
1968 case IB_CM_DREQ_RECEIVED:
1969 case IB_CM_DREP_RECEIVED:
1970 if (!cma_comp_exch(id_priv, RDMA_CM_CONNECT,
1971 RDMA_CM_DISCONNECT))
1972 goto out;
1973 event.event = RDMA_CM_EVENT_DISCONNECTED;
1974 break;
1975 case IB_CM_TIMEWAIT_EXIT:
1976 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
1977 break;
1978 case IB_CM_MRA_RECEIVED:
1979 /* ignore event */
1980 goto out;
1981 case IB_CM_REJ_RECEIVED:
1982 pr_debug_ratelimited("RDMA CM: REJECTED: %s\n", rdma_reject_msg(&id_priv->id,
1983 ib_event->param.rej_rcvd.reason));
1984 cma_modify_qp_err(id_priv);
1985 event.status = ib_event->param.rej_rcvd.reason;
1986 event.event = RDMA_CM_EVENT_REJECTED;
1987 event.param.conn.private_data = ib_event->private_data;
1988 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
1989 break;
1990 default:
1991 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
1992 ib_event->event);
1993 goto out;
1994 }
1995
1996 ret = cma_cm_event_handler(id_priv, &event);
1997 if (ret) {
1998 /* Destroy the CM ID by returning a non-zero value. */
1999 id_priv->cm_id.ib = NULL;
2000 cma_exch(id_priv, RDMA_CM_DESTROYING);
2001 mutex_unlock(&id_priv->handler_mutex);
2002 rdma_destroy_id(&id_priv->id);
2003 return ret;
2004 }
2005 out:
2006 mutex_unlock(&id_priv->handler_mutex);
2007 return ret;
2008 }
2009
2010 static struct rdma_id_private *
2011 cma_ib_new_conn_id(const struct rdma_cm_id *listen_id,
2012 const struct ib_cm_event *ib_event,
2013 struct net_device *net_dev)
2014 {
2015 struct rdma_id_private *listen_id_priv;
2016 struct rdma_id_private *id_priv;
2017 struct rdma_cm_id *id;
2018 struct rdma_route *rt;
2019 const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
2020 struct sa_path_rec *path = ib_event->param.req_rcvd.primary_path;
2021 const __be64 service_id =
2022 ib_event->param.req_rcvd.primary_path->service_id;
2023 int ret;
2024
2025 listen_id_priv = container_of(listen_id, struct rdma_id_private, id);
2026 id = __rdma_create_id(listen_id->route.addr.dev_addr.net,
2027 listen_id->event_handler, listen_id->context,
2028 listen_id->ps, ib_event->param.req_rcvd.qp_type,
2029 listen_id_priv->res.kern_name);
2030 if (IS_ERR(id))
2031 return NULL;
2032
2033 id_priv = container_of(id, struct rdma_id_private, id);
2034 if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
2035 (struct sockaddr *)&id->route.addr.dst_addr,
2036 listen_id, ib_event, ss_family, service_id))
2037 goto err;
2038
2039 rt = &id->route;
2040 rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
2041 rt->path_rec = kmalloc_array(rt->num_paths, sizeof(*rt->path_rec),
2042 GFP_KERNEL);
2043 if (!rt->path_rec)
2044 goto err;
2045
2046 rt->path_rec[0] = *path;
2047 if (rt->num_paths == 2)
2048 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
2049
2050 if (net_dev) {
2051 rdma_copy_src_l2_addr(&rt->addr.dev_addr, net_dev);
2052 } else {
2053 if (!cma_protocol_roce(listen_id) &&
2054 cma_any_addr(cma_src_addr(id_priv))) {
2055 rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
2056 rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
2057 ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
2058 } else if (!cma_any_addr(cma_src_addr(id_priv))) {
2059 ret = cma_translate_addr(cma_src_addr(id_priv), &rt->addr.dev_addr);
2060 if (ret)
2061 goto err;
2062 }
2063 }
2064 rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
2065
2066 id_priv->state = RDMA_CM_CONNECT;
2067 return id_priv;
2068
2069 err:
2070 rdma_destroy_id(id);
2071 return NULL;
2072 }
2073
2074 static struct rdma_id_private *
2075 cma_ib_new_udp_id(const struct rdma_cm_id *listen_id,
2076 const struct ib_cm_event *ib_event,
2077 struct net_device *net_dev)
2078 {
2079 const struct rdma_id_private *listen_id_priv;
2080 struct rdma_id_private *id_priv;
2081 struct rdma_cm_id *id;
2082 const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
2083 struct net *net = listen_id->route.addr.dev_addr.net;
2084 int ret;
2085
2086 listen_id_priv = container_of(listen_id, struct rdma_id_private, id);
2087 id = __rdma_create_id(net, listen_id->event_handler, listen_id->context,
2088 listen_id->ps, IB_QPT_UD,
2089 listen_id_priv->res.kern_name);
2090 if (IS_ERR(id))
2091 return NULL;
2092
2093 id_priv = container_of(id, struct rdma_id_private, id);
2094 if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
2095 (struct sockaddr *)&id->route.addr.dst_addr,
2096 listen_id, ib_event, ss_family,
2097 ib_event->param.sidr_req_rcvd.service_id))
2098 goto err;
2099
2100 if (net_dev) {
2101 rdma_copy_src_l2_addr(&id->route.addr.dev_addr, net_dev);
2102 } else {
2103 if (!cma_any_addr(cma_src_addr(id_priv))) {
2104 ret = cma_translate_addr(cma_src_addr(id_priv),
2105 &id->route.addr.dev_addr);
2106 if (ret)
2107 goto err;
2108 }
2109 }
2110
2111 id_priv->state = RDMA_CM_CONNECT;
2112 return id_priv;
2113 err:
2114 rdma_destroy_id(id);
2115 return NULL;
2116 }
2117
2118 static void cma_set_req_event_data(struct rdma_cm_event *event,
2119 const struct ib_cm_req_event_param *req_data,
2120 void *private_data, int offset)
2121 {
2122 event->param.conn.private_data = private_data + offset;
2123 event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
2124 event->param.conn.responder_resources = req_data->responder_resources;
2125 event->param.conn.initiator_depth = req_data->initiator_depth;
2126 event->param.conn.flow_control = req_data->flow_control;
2127 event->param.conn.retry_count = req_data->retry_count;
2128 event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
2129 event->param.conn.srq = req_data->srq;
2130 event->param.conn.qp_num = req_data->remote_qpn;
2131 }
2132
2133 static int cma_ib_check_req_qp_type(const struct rdma_cm_id *id,
2134 const struct ib_cm_event *ib_event)
2135 {
2136 return (((ib_event->event == IB_CM_REQ_RECEIVED) &&
2137 (ib_event->param.req_rcvd.qp_type == id->qp_type)) ||
2138 ((ib_event->event == IB_CM_SIDR_REQ_RECEIVED) &&
2139 (id->qp_type == IB_QPT_UD)) ||
2140 (!id->qp_type));
2141 }
2142
2143 static int cma_ib_req_handler(struct ib_cm_id *cm_id,
2144 const struct ib_cm_event *ib_event)
2145 {
2146 struct rdma_id_private *listen_id, *conn_id = NULL;
2147 struct rdma_cm_event event = {};
2148 struct cma_req_info req = {};
2149 struct net_device *net_dev;
2150 u8 offset;
2151 int ret;
2152
2153 listen_id = cma_ib_id_from_event(cm_id, ib_event, &req, &net_dev);
2154 if (IS_ERR(listen_id))
2155 return PTR_ERR(listen_id);
2156
2157 trace_cm_req_handler(listen_id, ib_event->event);
2158 if (!cma_ib_check_req_qp_type(&listen_id->id, ib_event)) {
2159 ret = -EINVAL;
2160 goto net_dev_put;
2161 }
2162
2163 mutex_lock(&listen_id->handler_mutex);
2164 if (listen_id->state != RDMA_CM_LISTEN) {
2165 ret = -ECONNABORTED;
2166 goto err1;
2167 }
2168
2169 offset = cma_user_data_offset(listen_id);
2170 event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
2171 if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) {
2172 conn_id = cma_ib_new_udp_id(&listen_id->id, ib_event, net_dev);
2173 event.param.ud.private_data = ib_event->private_data + offset;
2174 event.param.ud.private_data_len =
2175 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
2176 } else {
2177 conn_id = cma_ib_new_conn_id(&listen_id->id, ib_event, net_dev);
2178 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
2179 ib_event->private_data, offset);
2180 }
2181 if (!conn_id) {
2182 ret = -ENOMEM;
2183 goto err1;
2184 }
2185
2186 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
2187 ret = cma_ib_acquire_dev(conn_id, listen_id, &req);
2188 if (ret)
2189 goto err2;
2190
2191 conn_id->cm_id.ib = cm_id;
2192 cm_id->context = conn_id;
2193 cm_id->cm_handler = cma_ib_handler;
2194
2195 /*
2196 * Protect against the user destroying conn_id from another thread
2197 * until we're done accessing it.
2198 */
2199 cma_id_get(conn_id);
2200 ret = cma_cm_event_handler(conn_id, &event);
2201 if (ret)
2202 goto err3;
2203 /*
2204 * Acquire mutex to prevent user executing rdma_destroy_id()
2205 * while we're accessing the cm_id.
2206 */
2207 mutex_lock(&lock);
2208 if (cma_comp(conn_id, RDMA_CM_CONNECT) &&
2209 (conn_id->id.qp_type != IB_QPT_UD)) {
2210 trace_cm_send_mra(cm_id->context);
2211 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
2212 }
2213 mutex_unlock(&lock);
2214 mutex_unlock(&conn_id->handler_mutex);
2215 mutex_unlock(&listen_id->handler_mutex);
2216 cma_id_put(conn_id);
2217 if (net_dev)
2218 dev_put(net_dev);
2219 return 0;
2220
2221 err3:
2222 cma_id_put(conn_id);
2223 /* Destroy the CM ID by returning a non-zero value. */
2224 conn_id->cm_id.ib = NULL;
2225 err2:
2226 cma_exch(conn_id, RDMA_CM_DESTROYING);
2227 mutex_unlock(&conn_id->handler_mutex);
2228 err1:
2229 mutex_unlock(&listen_id->handler_mutex);
2230 if (conn_id)
2231 rdma_destroy_id(&conn_id->id);
2232
2233 net_dev_put:
2234 if (net_dev)
2235 dev_put(net_dev);
2236
2237 return ret;
2238 }
2239
2240 __be64 rdma_get_service_id(struct rdma_cm_id *id, struct sockaddr *addr)
2241 {
2242 if (addr->sa_family == AF_IB)
2243 return ((struct sockaddr_ib *) addr)->sib_sid;
2244
2245 return cpu_to_be64(((u64)id->ps << 16) + be16_to_cpu(cma_port(addr)));
2246 }
2247 EXPORT_SYMBOL(rdma_get_service_id);
2248
2249 void rdma_read_gids(struct rdma_cm_id *cm_id, union ib_gid *sgid,
2250 union ib_gid *dgid)
2251 {
2252 struct rdma_addr *addr = &cm_id->route.addr;
2253
2254 if (!cm_id->device) {
2255 if (sgid)
2256 memset(sgid, 0, sizeof(*sgid));
2257 if (dgid)
2258 memset(dgid, 0, sizeof(*dgid));
2259 return;
2260 }
2261
2262 if (rdma_protocol_roce(cm_id->device, cm_id->port_num)) {
2263 if (sgid)
2264 rdma_ip2gid((struct sockaddr *)&addr->src_addr, sgid);
2265 if (dgid)
2266 rdma_ip2gid((struct sockaddr *)&addr->dst_addr, dgid);
2267 } else {
2268 if (sgid)
2269 rdma_addr_get_sgid(&addr->dev_addr, sgid);
2270 if (dgid)
2271 rdma_addr_get_dgid(&addr->dev_addr, dgid);
2272 }
2273 }
2274 EXPORT_SYMBOL(rdma_read_gids);
2275
2276 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
2277 {
2278 struct rdma_id_private *id_priv = iw_id->context;
2279 struct rdma_cm_event event = {};
2280 int ret = 0;
2281 struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
2282 struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
2283
2284 mutex_lock(&id_priv->handler_mutex);
2285 if (id_priv->state != RDMA_CM_CONNECT)
2286 goto out;
2287
2288 switch (iw_event->event) {
2289 case IW_CM_EVENT_CLOSE:
2290 event.event = RDMA_CM_EVENT_DISCONNECTED;
2291 break;
2292 case IW_CM_EVENT_CONNECT_REPLY:
2293 memcpy(cma_src_addr(id_priv), laddr,
2294 rdma_addr_size(laddr));
2295 memcpy(cma_dst_addr(id_priv), raddr,
2296 rdma_addr_size(raddr));
2297 switch (iw_event->status) {
2298 case 0:
2299 event.event = RDMA_CM_EVENT_ESTABLISHED;
2300 event.param.conn.initiator_depth = iw_event->ird;
2301 event.param.conn.responder_resources = iw_event->ord;
2302 break;
2303 case -ECONNRESET:
2304 case -ECONNREFUSED:
2305 event.event = RDMA_CM_EVENT_REJECTED;
2306 break;
2307 case -ETIMEDOUT:
2308 event.event = RDMA_CM_EVENT_UNREACHABLE;
2309 break;
2310 default:
2311 event.event = RDMA_CM_EVENT_CONNECT_ERROR;
2312 break;
2313 }
2314 break;
2315 case IW_CM_EVENT_ESTABLISHED:
2316 event.event = RDMA_CM_EVENT_ESTABLISHED;
2317 event.param.conn.initiator_depth = iw_event->ird;
2318 event.param.conn.responder_resources = iw_event->ord;
2319 break;
2320 default:
2321 goto out;
2322 }
2323
2324 event.status = iw_event->status;
2325 event.param.conn.private_data = iw_event->private_data;
2326 event.param.conn.private_data_len = iw_event->private_data_len;
2327 ret = cma_cm_event_handler(id_priv, &event);
2328 if (ret) {
2329 /* Destroy the CM ID by returning a non-zero value. */
2330 id_priv->cm_id.iw = NULL;
2331 cma_exch(id_priv, RDMA_CM_DESTROYING);
2332 mutex_unlock(&id_priv->handler_mutex);
2333 rdma_destroy_id(&id_priv->id);
2334 return ret;
2335 }
2336
2337 out:
2338 mutex_unlock(&id_priv->handler_mutex);
2339 return ret;
2340 }
2341
2342 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
2343 struct iw_cm_event *iw_event)
2344 {
2345 struct rdma_cm_id *new_cm_id;
2346 struct rdma_id_private *listen_id, *conn_id;
2347 struct rdma_cm_event event = {};
2348 int ret = -ECONNABORTED;
2349 struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
2350 struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
2351
2352 event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
2353 event.param.conn.private_data = iw_event->private_data;
2354 event.param.conn.private_data_len = iw_event->private_data_len;
2355 event.param.conn.initiator_depth = iw_event->ird;
2356 event.param.conn.responder_resources = iw_event->ord;
2357
2358 listen_id = cm_id->context;
2359
2360 mutex_lock(&listen_id->handler_mutex);
2361 if (listen_id->state != RDMA_CM_LISTEN)
2362 goto out;
2363
2364 /* Create a new RDMA id for the new IW CM ID */
2365 new_cm_id = __rdma_create_id(listen_id->id.route.addr.dev_addr.net,
2366 listen_id->id.event_handler,
2367 listen_id->id.context,
2368 RDMA_PS_TCP, IB_QPT_RC,
2369 listen_id->res.kern_name);
2370 if (IS_ERR(new_cm_id)) {
2371 ret = -ENOMEM;
2372 goto out;
2373 }
2374 conn_id = container_of(new_cm_id, struct rdma_id_private, id);
2375 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
2376 conn_id->state = RDMA_CM_CONNECT;
2377
2378 ret = rdma_translate_ip(laddr, &conn_id->id.route.addr.dev_addr);
2379 if (ret) {
2380 mutex_unlock(&conn_id->handler_mutex);
2381 rdma_destroy_id(new_cm_id);
2382 goto out;
2383 }
2384
2385 ret = cma_iw_acquire_dev(conn_id, listen_id);
2386 if (ret) {
2387 mutex_unlock(&conn_id->handler_mutex);
2388 rdma_destroy_id(new_cm_id);
2389 goto out;
2390 }
2391
2392 conn_id->cm_id.iw = cm_id;
2393 cm_id->context = conn_id;
2394 cm_id->cm_handler = cma_iw_handler;
2395
2396 memcpy(cma_src_addr(conn_id), laddr, rdma_addr_size(laddr));
2397 memcpy(cma_dst_addr(conn_id), raddr, rdma_addr_size(raddr));
2398
2399 /*
2400 * Protect against the user destroying conn_id from another thread
2401 * until we're done accessing it.
2402 */
2403 cma_id_get(conn_id);
2404 ret = cma_cm_event_handler(conn_id, &event);
2405 if (ret) {
2406 /* User wants to destroy the CM ID */
2407 conn_id->cm_id.iw = NULL;
2408 cma_exch(conn_id, RDMA_CM_DESTROYING);
2409 mutex_unlock(&conn_id->handler_mutex);
2410 mutex_unlock(&listen_id->handler_mutex);
2411 cma_id_put(conn_id);
2412 rdma_destroy_id(&conn_id->id);
2413 return ret;
2414 }
2415
2416 mutex_unlock(&conn_id->handler_mutex);
2417 cma_id_put(conn_id);
2418
2419 out:
2420 mutex_unlock(&listen_id->handler_mutex);
2421 return ret;
2422 }
2423
2424 static int cma_ib_listen(struct rdma_id_private *id_priv)
2425 {
2426 struct sockaddr *addr;
2427 struct ib_cm_id *id;
2428 __be64 svc_id;
2429
2430 addr = cma_src_addr(id_priv);
2431 svc_id = rdma_get_service_id(&id_priv->id, addr);
2432 id = ib_cm_insert_listen(id_priv->id.device,
2433 cma_ib_req_handler, svc_id);
2434 if (IS_ERR(id))
2435 return PTR_ERR(id);
2436 id_priv->cm_id.ib = id;
2437
2438 return 0;
2439 }
2440
2441 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
2442 {
2443 int ret;
2444 struct iw_cm_id *id;
2445
2446 id = iw_create_cm_id(id_priv->id.device,
2447 iw_conn_req_handler,
2448 id_priv);
2449 if (IS_ERR(id))
2450 return PTR_ERR(id);
2451
2452 id->tos = id_priv->tos;
2453 id->tos_set = id_priv->tos_set;
2454 id_priv->cm_id.iw = id;
2455
2456 memcpy(&id_priv->cm_id.iw->local_addr, cma_src_addr(id_priv),
2457 rdma_addr_size(cma_src_addr(id_priv)));
2458
2459 ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
2460
2461 if (ret) {
2462 iw_destroy_cm_id(id_priv->cm_id.iw);
2463 id_priv->cm_id.iw = NULL;
2464 }
2465
2466 return ret;
2467 }
2468
2469 static int cma_listen_handler(struct rdma_cm_id *id,
2470 struct rdma_cm_event *event)
2471 {
2472 struct rdma_id_private *id_priv = id->context;
2473
2474 id->context = id_priv->id.context;
2475 id->event_handler = id_priv->id.event_handler;
2476 trace_cm_event_handler(id_priv, event);
2477 return id_priv->id.event_handler(id, event);
2478 }
2479
2480 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
2481 struct cma_device *cma_dev)
2482 {
2483 struct rdma_id_private *dev_id_priv;
2484 struct rdma_cm_id *id;
2485 struct net *net = id_priv->id.route.addr.dev_addr.net;
2486 int ret;
2487
2488 lockdep_assert_held(&lock);
2489
2490 if (cma_family(id_priv) == AF_IB && !rdma_cap_ib_cm(cma_dev->device, 1))
2491 return;
2492
2493 id = __rdma_create_id(net, cma_listen_handler, id_priv, id_priv->id.ps,
2494 id_priv->id.qp_type, id_priv->res.kern_name);
2495 if (IS_ERR(id))
2496 return;
2497
2498 dev_id_priv = container_of(id, struct rdma_id_private, id);
2499
2500 dev_id_priv->state = RDMA_CM_ADDR_BOUND;
2501 memcpy(cma_src_addr(dev_id_priv), cma_src_addr(id_priv),
2502 rdma_addr_size(cma_src_addr(id_priv)));
2503
2504 _cma_attach_to_dev(dev_id_priv, cma_dev);
2505 list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
2506 cma_id_get(id_priv);
2507 dev_id_priv->internal_id = 1;
2508 dev_id_priv->afonly = id_priv->afonly;
2509 dev_id_priv->tos_set = id_priv->tos_set;
2510 dev_id_priv->tos = id_priv->tos;
2511
2512 ret = rdma_listen(id, id_priv->backlog);
2513 if (ret)
2514 dev_warn(&cma_dev->device->dev,
2515 "RDMA CMA: cma_listen_on_dev, error %d\n", ret);
2516 }
2517
2518 static void cma_listen_on_all(struct rdma_id_private *id_priv)
2519 {
2520 struct cma_device *cma_dev;
2521
2522 mutex_lock(&lock);
2523 list_add_tail(&id_priv->list, &listen_any_list);
2524 list_for_each_entry(cma_dev, &dev_list, list)
2525 cma_listen_on_dev(id_priv, cma_dev);
2526 mutex_unlock(&lock);
2527 }
2528
2529 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
2530 {
2531 struct rdma_id_private *id_priv;
2532
2533 id_priv = container_of(id, struct rdma_id_private, id);
2534 id_priv->tos = (u8) tos;
2535 id_priv->tos_set = true;
2536 }
2537 EXPORT_SYMBOL(rdma_set_service_type);
2538
2539 /**
2540 * rdma_set_ack_timeout() - Set the ack timeout of QP associated
2541 * with a connection identifier.
2542 * @id: Communication identifier to associated with service type.
2543 * @timeout: Ack timeout to set a QP, expressed as 4.096 * 2^(timeout) usec.
2544 *
2545 * This function should be called before rdma_connect() on active side,
2546 * and on passive side before rdma_accept(). It is applicable to primary
2547 * path only. The timeout will affect the local side of the QP, it is not
2548 * negotiated with remote side and zero disables the timer. In case it is
2549 * set before rdma_resolve_route, the value will also be used to determine
2550 * PacketLifeTime for RoCE.
2551 *
2552 * Return: 0 for success
2553 */
2554 int rdma_set_ack_timeout(struct rdma_cm_id *id, u8 timeout)
2555 {
2556 struct rdma_id_private *id_priv;
2557
2558 if (id->qp_type != IB_QPT_RC)
2559 return -EINVAL;
2560
2561 id_priv = container_of(id, struct rdma_id_private, id);
2562 id_priv->timeout = timeout;
2563 id_priv->timeout_set = true;
2564
2565 return 0;
2566 }
2567 EXPORT_SYMBOL(rdma_set_ack_timeout);
2568
2569 static void cma_query_handler(int status, struct sa_path_rec *path_rec,
2570 void *context)
2571 {
2572 struct cma_work *work = context;
2573 struct rdma_route *route;
2574
2575 route = &work->id->id.route;
2576
2577 if (!status) {
2578 route->num_paths = 1;
2579 *route->path_rec = *path_rec;
2580 } else {
2581 work->old_state = RDMA_CM_ROUTE_QUERY;
2582 work->new_state = RDMA_CM_ADDR_RESOLVED;
2583 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
2584 work->event.status = status;
2585 pr_debug_ratelimited("RDMA CM: ROUTE_ERROR: failed to query path. status %d\n",
2586 status);
2587 }
2588
2589 queue_work(cma_wq, &work->work);
2590 }
2591
2592 static int cma_query_ib_route(struct rdma_id_private *id_priv,
2593 unsigned long timeout_ms, struct cma_work *work)
2594 {
2595 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2596 struct sa_path_rec path_rec;
2597 ib_sa_comp_mask comp_mask;
2598 struct sockaddr_in6 *sin6;
2599 struct sockaddr_ib *sib;
2600
2601 memset(&path_rec, 0, sizeof path_rec);
2602
2603 if (rdma_cap_opa_ah(id_priv->id.device, id_priv->id.port_num))
2604 path_rec.rec_type = SA_PATH_REC_TYPE_OPA;
2605 else
2606 path_rec.rec_type = SA_PATH_REC_TYPE_IB;
2607 rdma_addr_get_sgid(dev_addr, &path_rec.sgid);
2608 rdma_addr_get_dgid(dev_addr, &path_rec.dgid);
2609 path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
2610 path_rec.numb_path = 1;
2611 path_rec.reversible = 1;
2612 path_rec.service_id = rdma_get_service_id(&id_priv->id,
2613 cma_dst_addr(id_priv));
2614
2615 comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
2616 IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
2617 IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
2618
2619 switch (cma_family(id_priv)) {
2620 case AF_INET:
2621 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
2622 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
2623 break;
2624 case AF_INET6:
2625 sin6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
2626 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
2627 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2628 break;
2629 case AF_IB:
2630 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
2631 path_rec.traffic_class = (u8) (be32_to_cpu(sib->sib_flowinfo) >> 20);
2632 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2633 break;
2634 }
2635
2636 id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
2637 id_priv->id.port_num, &path_rec,
2638 comp_mask, timeout_ms,
2639 GFP_KERNEL, cma_query_handler,
2640 work, &id_priv->query);
2641
2642 return (id_priv->query_id < 0) ? id_priv->query_id : 0;
2643 }
2644
2645 static void cma_work_handler(struct work_struct *_work)
2646 {
2647 struct cma_work *work = container_of(_work, struct cma_work, work);
2648 struct rdma_id_private *id_priv = work->id;
2649 int destroy = 0;
2650
2651 mutex_lock(&id_priv->handler_mutex);
2652 if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
2653 goto out;
2654
2655 if (cma_cm_event_handler(id_priv, &work->event)) {
2656 cma_exch(id_priv, RDMA_CM_DESTROYING);
2657 destroy = 1;
2658 }
2659 out:
2660 mutex_unlock(&id_priv->handler_mutex);
2661 cma_id_put(id_priv);
2662 if (destroy)
2663 rdma_destroy_id(&id_priv->id);
2664 kfree(work);
2665 }
2666
2667 static void cma_ndev_work_handler(struct work_struct *_work)
2668 {
2669 struct cma_ndev_work *work = container_of(_work, struct cma_ndev_work, work);
2670 struct rdma_id_private *id_priv = work->id;
2671 int destroy = 0;
2672
2673 mutex_lock(&id_priv->handler_mutex);
2674 if (id_priv->state == RDMA_CM_DESTROYING ||
2675 id_priv->state == RDMA_CM_DEVICE_REMOVAL)
2676 goto out;
2677
2678 if (cma_cm_event_handler(id_priv, &work->event)) {
2679 cma_exch(id_priv, RDMA_CM_DESTROYING);
2680 destroy = 1;
2681 }
2682
2683 out:
2684 mutex_unlock(&id_priv->handler_mutex);
2685 cma_id_put(id_priv);
2686 if (destroy)
2687 rdma_destroy_id(&id_priv->id);
2688 kfree(work);
2689 }
2690
2691 static void cma_init_resolve_route_work(struct cma_work *work,
2692 struct rdma_id_private *id_priv)
2693 {
2694 work->id = id_priv;
2695 INIT_WORK(&work->work, cma_work_handler);
2696 work->old_state = RDMA_CM_ROUTE_QUERY;
2697 work->new_state = RDMA_CM_ROUTE_RESOLVED;
2698 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
2699 }
2700
2701 static void enqueue_resolve_addr_work(struct cma_work *work,
2702 struct rdma_id_private *id_priv)
2703 {
2704 /* Balances with cma_id_put() in cma_work_handler */
2705 cma_id_get(id_priv);
2706
2707 work->id = id_priv;
2708 INIT_WORK(&work->work, cma_work_handler);
2709 work->old_state = RDMA_CM_ADDR_QUERY;
2710 work->new_state = RDMA_CM_ADDR_RESOLVED;
2711 work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2712
2713 queue_work(cma_wq, &work->work);
2714 }
2715
2716 static int cma_resolve_ib_route(struct rdma_id_private *id_priv,
2717 unsigned long timeout_ms)
2718 {
2719 struct rdma_route *route = &id_priv->id.route;
2720 struct cma_work *work;
2721 int ret;
2722
2723 work = kzalloc(sizeof *work, GFP_KERNEL);
2724 if (!work)
2725 return -ENOMEM;
2726
2727 cma_init_resolve_route_work(work, id_priv);
2728
2729 route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
2730 if (!route->path_rec) {
2731 ret = -ENOMEM;
2732 goto err1;
2733 }
2734
2735 ret = cma_query_ib_route(id_priv, timeout_ms, work);
2736 if (ret)
2737 goto err2;
2738
2739 return 0;
2740 err2:
2741 kfree(route->path_rec);
2742 route->path_rec = NULL;
2743 err1:
2744 kfree(work);
2745 return ret;
2746 }
2747
2748 static enum ib_gid_type cma_route_gid_type(enum rdma_network_type network_type,
2749 unsigned long supported_gids,
2750 enum ib_gid_type default_gid)
2751 {
2752 if ((network_type == RDMA_NETWORK_IPV4 ||
2753 network_type == RDMA_NETWORK_IPV6) &&
2754 test_bit(IB_GID_TYPE_ROCE_UDP_ENCAP, &supported_gids))
2755 return IB_GID_TYPE_ROCE_UDP_ENCAP;
2756
2757 return default_gid;
2758 }
2759
2760 /*
2761 * cma_iboe_set_path_rec_l2_fields() is helper function which sets
2762 * path record type based on GID type.
2763 * It also sets up other L2 fields which includes destination mac address
2764 * netdev ifindex, of the path record.
2765 * It returns the netdev of the bound interface for this path record entry.
2766 */
2767 static struct net_device *
2768 cma_iboe_set_path_rec_l2_fields(struct rdma_id_private *id_priv)
2769 {
2770 struct rdma_route *route = &id_priv->id.route;
2771 enum ib_gid_type gid_type = IB_GID_TYPE_ROCE;
2772 struct rdma_addr *addr = &route->addr;
2773 unsigned long supported_gids;
2774 struct net_device *ndev;
2775
2776 if (!addr->dev_addr.bound_dev_if)
2777 return NULL;
2778
2779 ndev = dev_get_by_index(addr->dev_addr.net,
2780 addr->dev_addr.bound_dev_if);
2781 if (!ndev)
2782 return NULL;
2783
2784 supported_gids = roce_gid_type_mask_support(id_priv->id.device,
2785 id_priv->id.port_num);
2786 gid_type = cma_route_gid_type(addr->dev_addr.network,
2787 supported_gids,
2788 id_priv->gid_type);
2789 /* Use the hint from IP Stack to select GID Type */
2790 if (gid_type < ib_network_to_gid_type(addr->dev_addr.network))
2791 gid_type = ib_network_to_gid_type(addr->dev_addr.network);
2792 route->path_rec->rec_type = sa_conv_gid_to_pathrec_type(gid_type);
2793
2794 route->path_rec->roce.route_resolved = true;
2795 sa_path_set_dmac(route->path_rec, addr->dev_addr.dst_dev_addr);
2796 return ndev;
2797 }
2798
2799 int rdma_set_ib_path(struct rdma_cm_id *id,
2800 struct sa_path_rec *path_rec)
2801 {
2802 struct rdma_id_private *id_priv;
2803 struct net_device *ndev;
2804 int ret;
2805
2806 id_priv = container_of(id, struct rdma_id_private, id);
2807 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
2808 RDMA_CM_ROUTE_RESOLVED))
2809 return -EINVAL;
2810
2811 id->route.path_rec = kmemdup(path_rec, sizeof(*path_rec),
2812 GFP_KERNEL);
2813 if (!id->route.path_rec) {
2814 ret = -ENOMEM;
2815 goto err;
2816 }
2817
2818 if (rdma_protocol_roce(id->device, id->port_num)) {
2819 ndev = cma_iboe_set_path_rec_l2_fields(id_priv);
2820 if (!ndev) {
2821 ret = -ENODEV;
2822 goto err_free;
2823 }
2824 dev_put(ndev);
2825 }
2826
2827 id->route.num_paths = 1;
2828 return 0;
2829
2830 err_free:
2831 kfree(id->route.path_rec);
2832 id->route.path_rec = NULL;
2833 err:
2834 cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED);
2835 return ret;
2836 }
2837 EXPORT_SYMBOL(rdma_set_ib_path);
2838
2839 static int cma_resolve_iw_route(struct rdma_id_private *id_priv)
2840 {
2841 struct cma_work *work;
2842
2843 work = kzalloc(sizeof *work, GFP_KERNEL);
2844 if (!work)
2845 return -ENOMEM;
2846
2847 cma_init_resolve_route_work(work, id_priv);
2848 queue_work(cma_wq, &work->work);
2849 return 0;
2850 }
2851
2852 static int get_vlan_ndev_tc(struct net_device *vlan_ndev, int prio)
2853 {
2854 struct net_device *dev;
2855
2856 dev = vlan_dev_real_dev(vlan_ndev);
2857 if (dev->num_tc)
2858 return netdev_get_prio_tc_map(dev, prio);
2859
2860 return (vlan_dev_get_egress_qos_mask(vlan_ndev, prio) &
2861 VLAN_PRIO_MASK) >> VLAN_PRIO_SHIFT;
2862 }
2863
2864 struct iboe_prio_tc_map {
2865 int input_prio;
2866 int output_tc;
2867 bool found;
2868 };
2869
2870 static int get_lower_vlan_dev_tc(struct net_device *dev, void *data)
2871 {
2872 struct iboe_prio_tc_map *map = data;
2873
2874 if (is_vlan_dev(dev))
2875 map->output_tc = get_vlan_ndev_tc(dev, map->input_prio);
2876 else if (dev->num_tc)
2877 map->output_tc = netdev_get_prio_tc_map(dev, map->input_prio);
2878 else
2879 map->output_tc = 0;
2880 /* We are interested only in first level VLAN device, so always
2881 * return 1 to stop iterating over next level devices.
2882 */
2883 map->found = true;
2884 return 1;
2885 }
2886
2887 static int iboe_tos_to_sl(struct net_device *ndev, int tos)
2888 {
2889 struct iboe_prio_tc_map prio_tc_map = {};
2890 int prio = rt_tos2priority(tos);
2891
2892 /* If VLAN device, get it directly from the VLAN netdev */
2893 if (is_vlan_dev(ndev))
2894 return get_vlan_ndev_tc(ndev, prio);
2895
2896 prio_tc_map.input_prio = prio;
2897 rcu_read_lock();
2898 netdev_walk_all_lower_dev_rcu(ndev,
2899 get_lower_vlan_dev_tc,
2900 &prio_tc_map);
2901 rcu_read_unlock();
2902 /* If map is found from lower device, use it; Otherwise
2903 * continue with the current netdevice to get priority to tc map.
2904 */
2905 if (prio_tc_map.found)
2906 return prio_tc_map.output_tc;
2907 else if (ndev->num_tc)
2908 return netdev_get_prio_tc_map(ndev, prio);
2909 else
2910 return 0;
2911 }
2912
2913 static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
2914 {
2915 struct rdma_route *route = &id_priv->id.route;
2916 struct rdma_addr *addr = &route->addr;
2917 struct cma_work *work;
2918 int ret;
2919 struct net_device *ndev;
2920
2921 u8 default_roce_tos = id_priv->cma_dev->default_roce_tos[id_priv->id.port_num -
2922 rdma_start_port(id_priv->cma_dev->device)];
2923 u8 tos = id_priv->tos_set ? id_priv->tos : default_roce_tos;
2924
2925
2926 work = kzalloc(sizeof *work, GFP_KERNEL);
2927 if (!work)
2928 return -ENOMEM;
2929
2930 route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
2931 if (!route->path_rec) {
2932 ret = -ENOMEM;
2933 goto err1;
2934 }
2935
2936 route->num_paths = 1;
2937
2938 ndev = cma_iboe_set_path_rec_l2_fields(id_priv);
2939 if (!ndev) {
2940 ret = -ENODEV;
2941 goto err2;
2942 }
2943
2944 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
2945 &route->path_rec->sgid);
2946 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.dst_addr,
2947 &route->path_rec->dgid);
2948
2949 if (((struct sockaddr *)&id_priv->id.route.addr.dst_addr)->sa_family != AF_IB)
2950 /* TODO: get the hoplimit from the inet/inet6 device */
2951 route->path_rec->hop_limit = addr->dev_addr.hoplimit;
2952 else
2953 route->path_rec->hop_limit = 1;
2954 route->path_rec->reversible = 1;
2955 route->path_rec->pkey = cpu_to_be16(0xffff);
2956 route->path_rec->mtu_selector = IB_SA_EQ;
2957 route->path_rec->sl = iboe_tos_to_sl(ndev, tos);
2958 route->path_rec->traffic_class = tos;
2959 route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
2960 route->path_rec->rate_selector = IB_SA_EQ;
2961 route->path_rec->rate = iboe_get_rate(ndev);
2962 dev_put(ndev);
2963 route->path_rec->packet_life_time_selector = IB_SA_EQ;
2964 /* In case ACK timeout is set, use this value to calculate
2965 * PacketLifeTime. As per IBTA 12.7.34,
2966 * local ACK timeout = (2 * PacketLifeTime + Local CA’s ACK delay).
2967 * Assuming a negligible local ACK delay, we can use
2968 * PacketLifeTime = local ACK timeout/2
2969 * as a reasonable approximation for RoCE networks.
2970 */
2971 route->path_rec->packet_life_time = id_priv->timeout_set ?
2972 id_priv->timeout - 1 : CMA_IBOE_PACKET_LIFETIME;
2973
2974 if (!route->path_rec->mtu) {
2975 ret = -EINVAL;
2976 goto err2;
2977 }
2978
2979 cma_init_resolve_route_work(work, id_priv);
2980 queue_work(cma_wq, &work->work);
2981
2982 return 0;
2983
2984 err2:
2985 kfree(route->path_rec);
2986 route->path_rec = NULL;
2987 route->num_paths = 0;
2988 err1:
2989 kfree(work);
2990 return ret;
2991 }
2992
2993 int rdma_resolve_route(struct rdma_cm_id *id, unsigned long timeout_ms)
2994 {
2995 struct rdma_id_private *id_priv;
2996 int ret;
2997
2998 id_priv = container_of(id, struct rdma_id_private, id);
2999 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY))
3000 return -EINVAL;
3001
3002 cma_id_get(id_priv);
3003 if (rdma_cap_ib_sa(id->device, id->port_num))
3004 ret = cma_resolve_ib_route(id_priv, timeout_ms);
3005 else if (rdma_protocol_roce(id->device, id->port_num))
3006 ret = cma_resolve_iboe_route(id_priv);
3007 else if (rdma_protocol_iwarp(id->device, id->port_num))
3008 ret = cma_resolve_iw_route(id_priv);
3009 else
3010 ret = -ENOSYS;
3011
3012 if (ret)
3013 goto err;
3014
3015 return 0;
3016 err:
3017 cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED);
3018 cma_id_put(id_priv);
3019 return ret;
3020 }
3021 EXPORT_SYMBOL(rdma_resolve_route);
3022
3023 static void cma_set_loopback(struct sockaddr *addr)
3024 {
3025 switch (addr->sa_family) {
3026 case AF_INET:
3027 ((struct sockaddr_in *) addr)->sin_addr.s_addr = htonl(INADDR_LOOPBACK);
3028 break;
3029 case AF_INET6:
3030 ipv6_addr_set(&((struct sockaddr_in6 *) addr)->sin6_addr,
3031 0, 0, 0, htonl(1));
3032 break;
3033 default:
3034 ib_addr_set(&((struct sockaddr_ib *) addr)->sib_addr,
3035 0, 0, 0, htonl(1));
3036 break;
3037 }
3038 }
3039
3040 static int cma_bind_loopback(struct rdma_id_private *id_priv)
3041 {
3042 struct cma_device *cma_dev, *cur_dev;
3043 union ib_gid gid;
3044 enum ib_port_state port_state;
3045 unsigned int p;
3046 u16 pkey;
3047 int ret;
3048
3049 cma_dev = NULL;
3050 mutex_lock(&lock);
3051 list_for_each_entry(cur_dev, &dev_list, list) {
3052 if (cma_family(id_priv) == AF_IB &&
3053 !rdma_cap_ib_cm(cur_dev->device, 1))
3054 continue;
3055
3056 if (!cma_dev)
3057 cma_dev = cur_dev;
3058
3059 rdma_for_each_port (cur_dev->device, p) {
3060 if (!ib_get_cached_port_state(cur_dev->device, p, &port_state) &&
3061 port_state == IB_PORT_ACTIVE) {
3062 cma_dev = cur_dev;
3063 goto port_found;
3064 }
3065 }
3066 }
3067
3068 if (!cma_dev) {
3069 ret = -ENODEV;
3070 goto out;
3071 }
3072
3073 p = 1;
3074
3075 port_found:
3076 ret = rdma_query_gid(cma_dev->device, p, 0, &gid);
3077 if (ret)
3078 goto out;
3079
3080 ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
3081 if (ret)
3082 goto out;
3083
3084 id_priv->id.route.addr.dev_addr.dev_type =
3085 (rdma_protocol_ib(cma_dev->device, p)) ?
3086 ARPHRD_INFINIBAND : ARPHRD_ETHER;
3087
3088 rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
3089 ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
3090 id_priv->id.port_num = p;
3091 cma_attach_to_dev(id_priv, cma_dev);
3092 cma_set_loopback(cma_src_addr(id_priv));
3093 out:
3094 mutex_unlock(&lock);
3095 return ret;
3096 }
3097
3098 static void addr_handler(int status, struct sockaddr *src_addr,
3099 struct rdma_dev_addr *dev_addr, void *context)
3100 {
3101 struct rdma_id_private *id_priv = context;
3102 struct rdma_cm_event event = {};
3103 struct sockaddr *addr;
3104 struct sockaddr_storage old_addr;
3105
3106 mutex_lock(&id_priv->handler_mutex);
3107 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY,
3108 RDMA_CM_ADDR_RESOLVED))
3109 goto out;
3110
3111 /*
3112 * Store the previous src address, so that if we fail to acquire
3113 * matching rdma device, old address can be restored back, which helps
3114 * to cancel the cma listen operation correctly.
3115 */
3116 addr = cma_src_addr(id_priv);
3117 memcpy(&old_addr, addr, rdma_addr_size(addr));
3118 memcpy(addr, src_addr, rdma_addr_size(src_addr));
3119 if (!status && !id_priv->cma_dev) {
3120 status = cma_acquire_dev_by_src_ip(id_priv);
3121 if (status)
3122 pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to acquire device. status %d\n",
3123 status);
3124 } else if (status) {
3125 pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to resolve IP. status %d\n", status);
3126 }
3127
3128 if (status) {
3129 memcpy(addr, &old_addr,
3130 rdma_addr_size((struct sockaddr *)&old_addr));
3131 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
3132 RDMA_CM_ADDR_BOUND))
3133 goto out;
3134 event.event = RDMA_CM_EVENT_ADDR_ERROR;
3135 event.status = status;
3136 } else
3137 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
3138
3139 if (cma_cm_event_handler(id_priv, &event)) {
3140 cma_exch(id_priv, RDMA_CM_DESTROYING);
3141 mutex_unlock(&id_priv->handler_mutex);
3142 rdma_destroy_id(&id_priv->id);
3143 return;
3144 }
3145 out:
3146 mutex_unlock(&id_priv->handler_mutex);
3147 }
3148
3149 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
3150 {
3151 struct cma_work *work;
3152 union ib_gid gid;
3153 int ret;
3154
3155 work = kzalloc(sizeof *work, GFP_KERNEL);
3156 if (!work)
3157 return -ENOMEM;
3158
3159 if (!id_priv->cma_dev) {
3160 ret = cma_bind_loopback(id_priv);
3161 if (ret)
3162 goto err;
3163 }
3164
3165 rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
3166 rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
3167
3168 enqueue_resolve_addr_work(work, id_priv);
3169 return 0;
3170 err:
3171 kfree(work);
3172 return ret;
3173 }
3174
3175 static int cma_resolve_ib_addr(struct rdma_id_private *id_priv)
3176 {
3177 struct cma_work *work;
3178 int ret;
3179
3180 work = kzalloc(sizeof *work, GFP_KERNEL);
3181 if (!work)
3182 return -ENOMEM;
3183
3184 if (!id_priv->cma_dev) {
3185 ret = cma_resolve_ib_dev(id_priv);
3186 if (ret)
3187 goto err;
3188 }
3189
3190 rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, (union ib_gid *)
3191 &(((struct sockaddr_ib *) &id_priv->id.route.addr.dst_addr)->sib_addr));
3192
3193 enqueue_resolve_addr_work(work, id_priv);
3194 return 0;
3195 err:
3196 kfree(work);
3197 return ret;
3198 }
3199
3200 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
3201 const struct sockaddr *dst_addr)
3202 {
3203 if (!src_addr || !src_addr->sa_family) {
3204 src_addr = (struct sockaddr *) &id->route.addr.src_addr;
3205 src_addr->sa_family = dst_addr->sa_family;
3206 if (IS_ENABLED(CONFIG_IPV6) &&
3207 dst_addr->sa_family == AF_INET6) {
3208 struct sockaddr_in6 *src_addr6 = (struct sockaddr_in6 *) src_addr;
3209 struct sockaddr_in6 *dst_addr6 = (struct sockaddr_in6 *) dst_addr;
3210 src_addr6->sin6_scope_id = dst_addr6->sin6_scope_id;
3211 if (ipv6_addr_type(&dst_addr6->sin6_addr) & IPV6_ADDR_LINKLOCAL)
3212 id->route.addr.dev_addr.bound_dev_if = dst_addr6->sin6_scope_id;
3213 } else if (dst_addr->sa_family == AF_IB) {
3214 ((struct sockaddr_ib *) src_addr)->sib_pkey =
3215 ((struct sockaddr_ib *) dst_addr)->sib_pkey;
3216 }
3217 }
3218 return rdma_bind_addr(id, src_addr);
3219 }
3220
3221 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
3222 const struct sockaddr *dst_addr, unsigned long timeout_ms)
3223 {
3224 struct rdma_id_private *id_priv;
3225 int ret;
3226
3227 id_priv = container_of(id, struct rdma_id_private, id);
3228 memcpy(cma_dst_addr(id_priv), dst_addr, rdma_addr_size(dst_addr));
3229 if (id_priv->state == RDMA_CM_IDLE) {
3230 ret = cma_bind_addr(id, src_addr, dst_addr);
3231 if (ret) {
3232 memset(cma_dst_addr(id_priv), 0,
3233 rdma_addr_size(dst_addr));
3234 return ret;
3235 }
3236 }
3237
3238 if (cma_family(id_priv) != dst_addr->sa_family) {
3239 memset(cma_dst_addr(id_priv), 0, rdma_addr_size(dst_addr));
3240 return -EINVAL;
3241 }
3242
3243 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY)) {
3244 memset(cma_dst_addr(id_priv), 0, rdma_addr_size(dst_addr));
3245 return -EINVAL;
3246 }
3247
3248 if (cma_any_addr(dst_addr)) {
3249 ret = cma_resolve_loopback(id_priv);
3250 } else {
3251 if (dst_addr->sa_family == AF_IB) {
3252 ret = cma_resolve_ib_addr(id_priv);
3253 } else {
3254 ret = rdma_resolve_ip(cma_src_addr(id_priv), dst_addr,
3255 &id->route.addr.dev_addr,
3256 timeout_ms, addr_handler,
3257 false, id_priv);
3258 }
3259 }
3260 if (ret)
3261 goto err;
3262
3263 return 0;
3264 err:
3265 cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
3266 return ret;
3267 }
3268 EXPORT_SYMBOL(rdma_resolve_addr);
3269
3270 int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse)
3271 {
3272 struct rdma_id_private *id_priv;
3273 unsigned long flags;
3274 int ret;
3275
3276 id_priv = container_of(id, struct rdma_id_private, id);
3277 spin_lock_irqsave(&id_priv->lock, flags);
3278 if (reuse || id_priv->state == RDMA_CM_IDLE) {
3279 id_priv->reuseaddr = reuse;
3280 ret = 0;
3281 } else {
3282 ret = -EINVAL;
3283 }
3284 spin_unlock_irqrestore(&id_priv->lock, flags);
3285 return ret;
3286 }
3287 EXPORT_SYMBOL(rdma_set_reuseaddr);
3288
3289 int rdma_set_afonly(struct rdma_cm_id *id, int afonly)
3290 {
3291 struct rdma_id_private *id_priv;
3292 unsigned long flags;
3293 int ret;
3294
3295 id_priv = container_of(id, struct rdma_id_private, id);
3296 spin_lock_irqsave(&id_priv->lock, flags);
3297 if (id_priv->state == RDMA_CM_IDLE || id_priv->state == RDMA_CM_ADDR_BOUND) {
3298 id_priv->options |= (1 << CMA_OPTION_AFONLY);
3299 id_priv->afonly = afonly;
3300 ret = 0;
3301 } else {
3302 ret = -EINVAL;
3303 }
3304 spin_unlock_irqrestore(&id_priv->lock, flags);
3305 return ret;
3306 }
3307 EXPORT_SYMBOL(rdma_set_afonly);
3308
3309 static void cma_bind_port(struct rdma_bind_list *bind_list,
3310 struct rdma_id_private *id_priv)
3311 {
3312 struct sockaddr *addr;
3313 struct sockaddr_ib *sib;
3314 u64 sid, mask;
3315 __be16 port;
3316
3317 lockdep_assert_held(&lock);
3318
3319 addr = cma_src_addr(id_priv);
3320 port = htons(bind_list->port);
3321
3322 switch (addr->sa_family) {
3323 case AF_INET:
3324 ((struct sockaddr_in *) addr)->sin_port = port;
3325 break;
3326 case AF_INET6:
3327 ((struct sockaddr_in6 *) addr)->sin6_port = port;
3328 break;
3329 case AF_IB:
3330 sib = (struct sockaddr_ib *) addr;
3331 sid = be64_to_cpu(sib->sib_sid);
3332 mask = be64_to_cpu(sib->sib_sid_mask);
3333 sib->sib_sid = cpu_to_be64((sid & mask) | (u64) ntohs(port));
3334 sib->sib_sid_mask = cpu_to_be64(~0ULL);
3335 break;
3336 }
3337 id_priv->bind_list = bind_list;
3338 hlist_add_head(&id_priv->node, &bind_list->owners);
3339 }
3340
3341 static int cma_alloc_port(enum rdma_ucm_port_space ps,
3342 struct rdma_id_private *id_priv, unsigned short snum)
3343 {
3344 struct rdma_bind_list *bind_list;
3345 int ret;
3346
3347 lockdep_assert_held(&lock);
3348
3349 bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
3350 if (!bind_list)
3351 return -ENOMEM;
3352
3353 ret = cma_ps_alloc(id_priv->id.route.addr.dev_addr.net, ps, bind_list,
3354 snum);
3355 if (ret < 0)
3356 goto err;
3357
3358 bind_list->ps = ps;
3359 bind_list->port = snum;
3360 cma_bind_port(bind_list, id_priv);
3361 return 0;
3362 err:
3363 kfree(bind_list);
3364 return ret == -ENOSPC ? -EADDRNOTAVAIL : ret;
3365 }
3366
3367 static int cma_port_is_unique(struct rdma_bind_list *bind_list,
3368 struct rdma_id_private *id_priv)
3369 {
3370 struct rdma_id_private *cur_id;
3371 struct sockaddr *daddr = cma_dst_addr(id_priv);
3372 struct sockaddr *saddr = cma_src_addr(id_priv);
3373 __be16 dport = cma_port(daddr);
3374
3375 lockdep_assert_held(&lock);
3376
3377 hlist_for_each_entry(cur_id, &bind_list->owners, node) {
3378 struct sockaddr *cur_daddr = cma_dst_addr(cur_id);
3379 struct sockaddr *cur_saddr = cma_src_addr(cur_id);
3380 __be16 cur_dport = cma_port(cur_daddr);
3381
3382 if (id_priv == cur_id)
3383 continue;
3384
3385 /* different dest port -> unique */
3386 if (!cma_any_port(daddr) &&
3387 !cma_any_port(cur_daddr) &&
3388 (dport != cur_dport))
3389 continue;
3390
3391 /* different src address -> unique */
3392 if (!cma_any_addr(saddr) &&
3393 !cma_any_addr(cur_saddr) &&
3394 cma_addr_cmp(saddr, cur_saddr))
3395 continue;
3396
3397 /* different dst address -> unique */
3398 if (!cma_any_addr(daddr) &&
3399 !cma_any_addr(cur_daddr) &&
3400 cma_addr_cmp(daddr, cur_daddr))
3401 continue;
3402
3403 return -EADDRNOTAVAIL;
3404 }
3405 return 0;
3406 }
3407
3408 static int cma_alloc_any_port(enum rdma_ucm_port_space ps,
3409 struct rdma_id_private *id_priv)
3410 {
3411 static unsigned int last_used_port;
3412 int low, high, remaining;
3413 unsigned int rover;
3414 struct net *net = id_priv->id.route.addr.dev_addr.net;
3415
3416 lockdep_assert_held(&lock);
3417
3418 inet_get_local_port_range(net, &low, &high);
3419 remaining = (high - low) + 1;
3420 rover = prandom_u32() % remaining + low;
3421 retry:
3422 if (last_used_port != rover) {
3423 struct rdma_bind_list *bind_list;
3424 int ret;
3425
3426 bind_list = cma_ps_find(net, ps, (unsigned short)rover);
3427
3428 if (!bind_list) {
3429 ret = cma_alloc_port(ps, id_priv, rover);
3430 } else {
3431 ret = cma_port_is_unique(bind_list, id_priv);
3432 if (!ret)
3433 cma_bind_port(bind_list, id_priv);
3434 }
3435 /*
3436 * Remember previously used port number in order to avoid
3437 * re-using same port immediately after it is closed.
3438 */
3439 if (!ret)
3440 last_used_port = rover;
3441 if (ret != -EADDRNOTAVAIL)
3442 return ret;
3443 }
3444 if (--remaining) {
3445 rover++;
3446 if ((rover < low) || (rover > high))
3447 rover = low;
3448 goto retry;
3449 }
3450 return -EADDRNOTAVAIL;
3451 }
3452
3453 /*
3454 * Check that the requested port is available. This is called when trying to
3455 * bind to a specific port, or when trying to listen on a bound port. In
3456 * the latter case, the provided id_priv may already be on the bind_list, but
3457 * we still need to check that it's okay to start listening.
3458 */
3459 static int cma_check_port(struct rdma_bind_list *bind_list,
3460 struct rdma_id_private *id_priv, uint8_t reuseaddr)
3461 {
3462 struct rdma_id_private *cur_id;
3463 struct sockaddr *addr, *cur_addr;
3464
3465 lockdep_assert_held(&lock);
3466
3467 addr = cma_src_addr(id_priv);
3468 hlist_for_each_entry(cur_id, &bind_list->owners, node) {
3469 if (id_priv == cur_id)
3470 continue;
3471
3472 if ((cur_id->state != RDMA_CM_LISTEN) && reuseaddr &&
3473 cur_id->reuseaddr)
3474 continue;
3475
3476 cur_addr = cma_src_addr(cur_id);
3477 if (id_priv->afonly && cur_id->afonly &&
3478 (addr->sa_family != cur_addr->sa_family))
3479 continue;
3480
3481 if (cma_any_addr(addr) || cma_any_addr(cur_addr))
3482 return -EADDRNOTAVAIL;
3483
3484 if (!cma_addr_cmp(addr, cur_addr))
3485 return -EADDRINUSE;
3486 }
3487 return 0;
3488 }
3489
3490 static int cma_use_port(enum rdma_ucm_port_space ps,
3491 struct rdma_id_private *id_priv)
3492 {
3493 struct rdma_bind_list *bind_list;
3494 unsigned short snum;
3495 int ret;
3496
3497 lockdep_assert_held(&lock);
3498
3499 snum = ntohs(cma_port(cma_src_addr(id_priv)));
3500 if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
3501 return -EACCES;
3502
3503 bind_list = cma_ps_find(id_priv->id.route.addr.dev_addr.net, ps, snum);
3504 if (!bind_list) {
3505 ret = cma_alloc_port(ps, id_priv, snum);
3506 } else {
3507 ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr);
3508 if (!ret)
3509 cma_bind_port(bind_list, id_priv);
3510 }
3511 return ret;
3512 }
3513
3514 static int cma_bind_listen(struct rdma_id_private *id_priv)
3515 {
3516 struct rdma_bind_list *bind_list = id_priv->bind_list;
3517 int ret = 0;
3518
3519 mutex_lock(&lock);
3520 if (bind_list->owners.first->next)
3521 ret = cma_check_port(bind_list, id_priv, 0);
3522 mutex_unlock(&lock);
3523 return ret;
3524 }
3525
3526 static enum rdma_ucm_port_space
3527 cma_select_inet_ps(struct rdma_id_private *id_priv)
3528 {
3529 switch (id_priv->id.ps) {
3530 case RDMA_PS_TCP:
3531 case RDMA_PS_UDP:
3532 case RDMA_PS_IPOIB:
3533 case RDMA_PS_IB:
3534 return id_priv->id.ps;
3535 default:
3536
3537 return 0;
3538 }
3539 }
3540
3541 static enum rdma_ucm_port_space
3542 cma_select_ib_ps(struct rdma_id_private *id_priv)
3543 {
3544 enum rdma_ucm_port_space ps = 0;
3545 struct sockaddr_ib *sib;
3546 u64 sid_ps, mask, sid;
3547
3548 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
3549 mask = be64_to_cpu(sib->sib_sid_mask) & RDMA_IB_IP_PS_MASK;
3550 sid = be64_to_cpu(sib->sib_sid) & mask;
3551
3552 if ((id_priv->id.ps == RDMA_PS_IB) && (sid == (RDMA_IB_IP_PS_IB & mask))) {
3553 sid_ps = RDMA_IB_IP_PS_IB;
3554 ps = RDMA_PS_IB;
3555 } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_TCP)) &&
3556 (sid == (RDMA_IB_IP_PS_TCP & mask))) {
3557 sid_ps = RDMA_IB_IP_PS_TCP;
3558 ps = RDMA_PS_TCP;
3559 } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_UDP)) &&
3560 (sid == (RDMA_IB_IP_PS_UDP & mask))) {
3561 sid_ps = RDMA_IB_IP_PS_UDP;
3562 ps = RDMA_PS_UDP;
3563 }
3564
3565 if (ps) {
3566 sib->sib_sid = cpu_to_be64(sid_ps | ntohs(cma_port((struct sockaddr *) sib)));
3567 sib->sib_sid_mask = cpu_to_be64(RDMA_IB_IP_PS_MASK |
3568 be64_to_cpu(sib->sib_sid_mask));
3569 }
3570 return ps;
3571 }
3572
3573 static int cma_get_port(struct rdma_id_private *id_priv)
3574 {
3575 enum rdma_ucm_port_space ps;
3576 int ret;
3577
3578 if (cma_family(id_priv) != AF_IB)
3579 ps = cma_select_inet_ps(id_priv);
3580 else
3581 ps = cma_select_ib_ps(id_priv);
3582 if (!ps)
3583 return -EPROTONOSUPPORT;
3584
3585 mutex_lock(&lock);
3586 if (cma_any_port(cma_src_addr(id_priv)))
3587 ret = cma_alloc_any_port(ps, id_priv);
3588 else
3589 ret = cma_use_port(ps, id_priv);
3590 mutex_unlock(&lock);
3591
3592 return ret;
3593 }
3594
3595 static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
3596 struct sockaddr *addr)
3597 {
3598 #if IS_ENABLED(CONFIG_IPV6)
3599 struct sockaddr_in6 *sin6;
3600
3601 if (addr->sa_family != AF_INET6)
3602 return 0;
3603
3604 sin6 = (struct sockaddr_in6 *) addr;
3605
3606 if (!(ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL))
3607 return 0;
3608
3609 if (!sin6->sin6_scope_id)
3610 return -EINVAL;
3611
3612 dev_addr->bound_dev_if = sin6->sin6_scope_id;
3613 #endif
3614 return 0;
3615 }
3616
3617 int rdma_listen(struct rdma_cm_id *id, int backlog)
3618 {
3619 struct rdma_id_private *id_priv;
3620 int ret;
3621
3622 id_priv = container_of(id, struct rdma_id_private, id);
3623 if (id_priv->state == RDMA_CM_IDLE) {
3624 id->route.addr.src_addr.ss_family = AF_INET;
3625 ret = rdma_bind_addr(id, cma_src_addr(id_priv));
3626 if (ret)
3627 return ret;
3628 }
3629
3630 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN))
3631 return -EINVAL;
3632
3633 if (id_priv->reuseaddr) {
3634 ret = cma_bind_listen(id_priv);
3635 if (ret)
3636 goto err;
3637 }
3638
3639 id_priv->backlog = backlog;
3640 if (id->device) {
3641 if (rdma_cap_ib_cm(id->device, 1)) {
3642 ret = cma_ib_listen(id_priv);
3643 if (ret)
3644 goto err;
3645 } else if (rdma_cap_iw_cm(id->device, 1)) {
3646 ret = cma_iw_listen(id_priv, backlog);
3647 if (ret)
3648 goto err;
3649 } else {
3650 ret = -ENOSYS;
3651 goto err;
3652 }
3653 } else
3654 cma_listen_on_all(id_priv);
3655
3656 return 0;
3657 err:
3658 id_priv->backlog = 0;
3659 cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND);
3660 return ret;
3661 }
3662 EXPORT_SYMBOL(rdma_listen);
3663
3664 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
3665 {
3666 struct rdma_id_private *id_priv;
3667 int ret;
3668 struct sockaddr *daddr;
3669
3670 if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6 &&
3671 addr->sa_family != AF_IB)
3672 return -EAFNOSUPPORT;
3673
3674 id_priv = container_of(id, struct rdma_id_private, id);
3675 if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND))
3676 return -EINVAL;
3677
3678 ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
3679 if (ret)
3680 goto err1;
3681
3682 memcpy(cma_src_addr(id_priv), addr, rdma_addr_size(addr));
3683 if (!cma_any_addr(addr)) {
3684 ret = cma_translate_addr(addr, &id->route.addr.dev_addr);
3685 if (ret)
3686 goto err1;
3687
3688 ret = cma_acquire_dev_by_src_ip(id_priv);
3689 if (ret)
3690 goto err1;
3691 }
3692
3693 if (!(id_priv->options & (1 << CMA_OPTION_AFONLY))) {
3694 if (addr->sa_family == AF_INET)
3695 id_priv->afonly = 1;
3696 #if IS_ENABLED(CONFIG_IPV6)
3697 else if (addr->sa_family == AF_INET6) {
3698 struct net *net = id_priv->id.route.addr.dev_addr.net;
3699
3700 id_priv->afonly = net->ipv6.sysctl.bindv6only;
3701 }
3702 #endif
3703 }
3704 daddr = cma_dst_addr(id_priv);
3705 daddr->sa_family = addr->sa_family;
3706
3707 ret = cma_get_port(id_priv);
3708 if (ret)
3709 goto err2;
3710
3711 return 0;
3712 err2:
3713 rdma_restrack_del(&id_priv->res);
3714 if (id_priv->cma_dev)
3715 cma_release_dev(id_priv);
3716 err1:
3717 cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE);
3718 return ret;
3719 }
3720 EXPORT_SYMBOL(rdma_bind_addr);
3721
3722 static int cma_format_hdr(void *hdr, struct rdma_id_private *id_priv)
3723 {
3724 struct cma_hdr *cma_hdr;
3725
3726 cma_hdr = hdr;
3727 cma_hdr->cma_version = CMA_VERSION;
3728 if (cma_family(id_priv) == AF_INET) {
3729 struct sockaddr_in *src4, *dst4;
3730
3731 src4 = (struct sockaddr_in *) cma_src_addr(id_priv);
3732 dst4 = (struct sockaddr_in *) cma_dst_addr(id_priv);
3733
3734 cma_set_ip_ver(cma_hdr, 4);
3735 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
3736 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
3737 cma_hdr->port = src4->sin_port;
3738 } else if (cma_family(id_priv) == AF_INET6) {
3739 struct sockaddr_in6 *src6, *dst6;
3740
3741 src6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
3742 dst6 = (struct sockaddr_in6 *) cma_dst_addr(id_priv);
3743
3744 cma_set_ip_ver(cma_hdr, 6);
3745 cma_hdr->src_addr.ip6 = src6->sin6_addr;
3746 cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
3747 cma_hdr->port = src6->sin6_port;
3748 }
3749 return 0;
3750 }
3751
3752 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
3753 const struct ib_cm_event *ib_event)
3754 {
3755 struct rdma_id_private *id_priv = cm_id->context;
3756 struct rdma_cm_event event = {};
3757 const struct ib_cm_sidr_rep_event_param *rep =
3758 &ib_event->param.sidr_rep_rcvd;
3759 int ret = 0;
3760
3761 mutex_lock(&id_priv->handler_mutex);
3762 if (id_priv->state != RDMA_CM_CONNECT)
3763 goto out;
3764
3765 switch (ib_event->event) {
3766 case IB_CM_SIDR_REQ_ERROR:
3767 event.event = RDMA_CM_EVENT_UNREACHABLE;
3768 event.status = -ETIMEDOUT;
3769 break;
3770 case IB_CM_SIDR_REP_RECEIVED:
3771 event.param.ud.private_data = ib_event->private_data;
3772 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
3773 if (rep->status != IB_SIDR_SUCCESS) {
3774 event.event = RDMA_CM_EVENT_UNREACHABLE;
3775 event.status = ib_event->param.sidr_rep_rcvd.status;
3776 pr_debug_ratelimited("RDMA CM: UNREACHABLE: bad SIDR reply. status %d\n",
3777 event.status);
3778 break;
3779 }
3780 ret = cma_set_qkey(id_priv, rep->qkey);
3781 if (ret) {
3782 pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to set qkey. status %d\n", ret);
3783 event.event = RDMA_CM_EVENT_ADDR_ERROR;
3784 event.status = ret;
3785 break;
3786 }
3787 ib_init_ah_attr_from_path(id_priv->id.device,
3788 id_priv->id.port_num,
3789 id_priv->id.route.path_rec,
3790 &event.param.ud.ah_attr,
3791 rep->sgid_attr);
3792 event.param.ud.qp_num = rep->qpn;
3793 event.param.ud.qkey = rep->qkey;
3794 event.event = RDMA_CM_EVENT_ESTABLISHED;
3795 event.status = 0;
3796 break;
3797 default:
3798 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
3799 ib_event->event);
3800 goto out;
3801 }
3802
3803 ret = cma_cm_event_handler(id_priv, &event);
3804
3805 rdma_destroy_ah_attr(&event.param.ud.ah_attr);
3806 if (ret) {
3807 /* Destroy the CM ID by returning a non-zero value. */
3808 id_priv->cm_id.ib = NULL;
3809 cma_exch(id_priv, RDMA_CM_DESTROYING);
3810 mutex_unlock(&id_priv->handler_mutex);
3811 rdma_destroy_id(&id_priv->id);
3812 return ret;
3813 }
3814 out:
3815 mutex_unlock(&id_priv->handler_mutex);
3816 return ret;
3817 }
3818
3819 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
3820 struct rdma_conn_param *conn_param)
3821 {
3822 struct ib_cm_sidr_req_param req;
3823 struct ib_cm_id *id;
3824 void *private_data;
3825 u8 offset;
3826 int ret;
3827
3828 memset(&req, 0, sizeof req);
3829 offset = cma_user_data_offset(id_priv);
3830 req.private_data_len = offset + conn_param->private_data_len;
3831 if (req.private_data_len < conn_param->private_data_len)
3832 return -EINVAL;
3833
3834 if (req.private_data_len) {
3835 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
3836 if (!private_data)
3837 return -ENOMEM;
3838 } else {
3839 private_data = NULL;
3840 }
3841
3842 if (conn_param->private_data && conn_param->private_data_len)
3843 memcpy(private_data + offset, conn_param->private_data,
3844 conn_param->private_data_len);
3845
3846 if (private_data) {
3847 ret = cma_format_hdr(private_data, id_priv);
3848 if (ret)
3849 goto out;
3850 req.private_data = private_data;
3851 }
3852
3853 id = ib_create_cm_id(id_priv->id.device, cma_sidr_rep_handler,
3854 id_priv);
3855 if (IS_ERR(id)) {
3856 ret = PTR_ERR(id);
3857 goto out;
3858 }
3859 id_priv->cm_id.ib = id;
3860
3861 req.path = id_priv->id.route.path_rec;
3862 req.sgid_attr = id_priv->id.route.addr.dev_addr.sgid_attr;
3863 req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
3864 req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
3865 req.max_cm_retries = CMA_MAX_CM_RETRIES;
3866
3867 trace_cm_send_sidr_req(id_priv);
3868 ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
3869 if (ret) {
3870 ib_destroy_cm_id(id_priv->cm_id.ib);
3871 id_priv->cm_id.ib = NULL;
3872 }
3873 out:
3874 kfree(private_data);
3875 return ret;
3876 }
3877
3878 static int cma_connect_ib(struct rdma_id_private *id_priv,
3879 struct rdma_conn_param *conn_param)
3880 {
3881 struct ib_cm_req_param req;
3882 struct rdma_route *route;
3883 void *private_data;
3884 struct ib_cm_id *id;
3885 u8 offset;
3886 int ret;
3887
3888 memset(&req, 0, sizeof req);
3889 offset = cma_user_data_offset(id_priv);
3890 req.private_data_len = offset + conn_param->private_data_len;
3891 if (req.private_data_len < conn_param->private_data_len)
3892 return -EINVAL;
3893
3894 if (req.private_data_len) {
3895 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
3896 if (!private_data)
3897 return -ENOMEM;
3898 } else {
3899 private_data = NULL;
3900 }
3901
3902 if (conn_param->private_data && conn_param->private_data_len)
3903 memcpy(private_data + offset, conn_param->private_data,
3904 conn_param->private_data_len);
3905
3906 id = ib_create_cm_id(id_priv->id.device, cma_ib_handler, id_priv);
3907 if (IS_ERR(id)) {
3908 ret = PTR_ERR(id);
3909 goto out;
3910 }
3911 id_priv->cm_id.ib = id;
3912
3913 route = &id_priv->id.route;
3914 if (private_data) {
3915 ret = cma_format_hdr(private_data, id_priv);
3916 if (ret)
3917 goto out;
3918 req.private_data = private_data;
3919 }
3920
3921 req.primary_path = &route->path_rec[0];
3922 if (route->num_paths == 2)
3923 req.alternate_path = &route->path_rec[1];
3924
3925 req.ppath_sgid_attr = id_priv->id.route.addr.dev_addr.sgid_attr;
3926 /* Alternate path SGID attribute currently unsupported */
3927 req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
3928 req.qp_num = id_priv->qp_num;
3929 req.qp_type = id_priv->id.qp_type;
3930 req.starting_psn = id_priv->seq_num;
3931 req.responder_resources = conn_param->responder_resources;
3932 req.initiator_depth = conn_param->initiator_depth;
3933 req.flow_control = conn_param->flow_control;
3934 req.retry_count = min_t(u8, 7, conn_param->retry_count);
3935 req.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
3936 req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
3937 req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
3938 req.max_cm_retries = CMA_MAX_CM_RETRIES;
3939 req.srq = id_priv->srq ? 1 : 0;
3940
3941 trace_cm_send_req(id_priv);
3942 ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
3943 out:
3944 if (ret && !IS_ERR(id)) {
3945 ib_destroy_cm_id(id);
3946 id_priv->cm_id.ib = NULL;
3947 }
3948
3949 kfree(private_data);
3950 return ret;
3951 }
3952
3953 static int cma_connect_iw(struct rdma_id_private *id_priv,
3954 struct rdma_conn_param *conn_param)
3955 {
3956 struct iw_cm_id *cm_id;
3957 int ret;
3958 struct iw_cm_conn_param iw_param;
3959
3960 cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
3961 if (IS_ERR(cm_id))
3962 return PTR_ERR(cm_id);
3963
3964 cm_id->tos = id_priv->tos;
3965 cm_id->tos_set = id_priv->tos_set;
3966 id_priv->cm_id.iw = cm_id;
3967
3968 memcpy(&cm_id->local_addr, cma_src_addr(id_priv),
3969 rdma_addr_size(cma_src_addr(id_priv)));
3970 memcpy(&cm_id->remote_addr, cma_dst_addr(id_priv),
3971 rdma_addr_size(cma_dst_addr(id_priv)));
3972
3973 ret = cma_modify_qp_rtr(id_priv, conn_param);
3974 if (ret)
3975 goto out;
3976
3977 if (conn_param) {
3978 iw_param.ord = conn_param->initiator_depth;
3979 iw_param.ird = conn_param->responder_resources;
3980 iw_param.private_data = conn_param->private_data;
3981 iw_param.private_data_len = conn_param->private_data_len;
3982 iw_param.qpn = id_priv->id.qp ? id_priv->qp_num : conn_param->qp_num;
3983 } else {
3984 memset(&iw_param, 0, sizeof iw_param);
3985 iw_param.qpn = id_priv->qp_num;
3986 }
3987 ret = iw_cm_connect(cm_id, &iw_param);
3988 out:
3989 if (ret) {
3990 iw_destroy_cm_id(cm_id);
3991 id_priv->cm_id.iw = NULL;
3992 }
3993 return ret;
3994 }
3995
3996 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
3997 {
3998 struct rdma_id_private *id_priv;
3999 int ret;
4000
4001 id_priv = container_of(id, struct rdma_id_private, id);
4002 if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT))
4003 return -EINVAL;
4004
4005 if (!id->qp) {
4006 id_priv->qp_num = conn_param->qp_num;
4007 id_priv->srq = conn_param->srq;
4008 }
4009
4010 if (rdma_cap_ib_cm(id->device, id->port_num)) {
4011 if (id->qp_type == IB_QPT_UD)
4012 ret = cma_resolve_ib_udp(id_priv, conn_param);
4013 else
4014 ret = cma_connect_ib(id_priv, conn_param);
4015 } else if (rdma_cap_iw_cm(id->device, id->port_num))
4016 ret = cma_connect_iw(id_priv, conn_param);
4017 else
4018 ret = -ENOSYS;
4019 if (ret)
4020 goto err;
4021
4022 return 0;
4023 err:
4024 cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED);
4025 return ret;
4026 }
4027 EXPORT_SYMBOL(rdma_connect);
4028
4029 static int cma_accept_ib(struct rdma_id_private *id_priv,
4030 struct rdma_conn_param *conn_param)
4031 {
4032 struct ib_cm_rep_param rep;
4033 int ret;
4034
4035 ret = cma_modify_qp_rtr(id_priv, conn_param);
4036 if (ret)
4037 goto out;
4038
4039 ret = cma_modify_qp_rts(id_priv, conn_param);
4040 if (ret)
4041 goto out;
4042
4043 memset(&rep, 0, sizeof rep);
4044 rep.qp_num = id_priv->qp_num;
4045 rep.starting_psn = id_priv->seq_num;
4046 rep.private_data = conn_param->private_data;
4047 rep.private_data_len = conn_param->private_data_len;
4048 rep.responder_resources = conn_param->responder_resources;
4049 rep.initiator_depth = conn_param->initiator_depth;
4050 rep.failover_accepted = 0;
4051 rep.flow_control = conn_param->flow_control;
4052 rep.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
4053 rep.srq = id_priv->srq ? 1 : 0;
4054
4055 trace_cm_send_rep(id_priv);
4056 ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
4057 out:
4058 return ret;
4059 }
4060
4061 static int cma_accept_iw(struct rdma_id_private *id_priv,
4062 struct rdma_conn_param *conn_param)
4063 {
4064 struct iw_cm_conn_param iw_param;
4065 int ret;
4066
4067 if (!conn_param)
4068 return -EINVAL;
4069
4070 ret = cma_modify_qp_rtr(id_priv, conn_param);
4071 if (ret)
4072 return ret;
4073
4074 iw_param.ord = conn_param->initiator_depth;
4075 iw_param.ird = conn_param->responder_resources;
4076 iw_param.private_data = conn_param->private_data;
4077 iw_param.private_data_len = conn_param->private_data_len;
4078 if (id_priv->id.qp) {
4079 iw_param.qpn = id_priv->qp_num;
4080 } else
4081 iw_param.qpn = conn_param->qp_num;
4082
4083 return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
4084 }
4085
4086 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
4087 enum ib_cm_sidr_status status, u32 qkey,
4088 const void *private_data, int private_data_len)
4089 {
4090 struct ib_cm_sidr_rep_param rep;
4091 int ret;
4092
4093 memset(&rep, 0, sizeof rep);
4094 rep.status = status;
4095 if (status == IB_SIDR_SUCCESS) {
4096 ret = cma_set_qkey(id_priv, qkey);
4097 if (ret)
4098 return ret;
4099 rep.qp_num = id_priv->qp_num;
4100 rep.qkey = id_priv->qkey;
4101 }
4102 rep.private_data = private_data;
4103 rep.private_data_len = private_data_len;
4104
4105 trace_cm_send_sidr_rep(id_priv);
4106 return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
4107 }
4108
4109 int __rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param,
4110 const char *caller)
4111 {
4112 struct rdma_id_private *id_priv;
4113 int ret;
4114
4115 id_priv = container_of(id, struct rdma_id_private, id);
4116
4117 rdma_restrack_set_task(&id_priv->res, caller);
4118
4119 if (!cma_comp(id_priv, RDMA_CM_CONNECT))
4120 return -EINVAL;
4121
4122 if (!id->qp && conn_param) {
4123 id_priv->qp_num = conn_param->qp_num;
4124 id_priv->srq = conn_param->srq;
4125 }
4126
4127 if (rdma_cap_ib_cm(id->device, id->port_num)) {
4128 if (id->qp_type == IB_QPT_UD) {
4129 if (conn_param)
4130 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
4131 conn_param->qkey,
4132 conn_param->private_data,
4133 conn_param->private_data_len);
4134 else
4135 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
4136 0, NULL, 0);
4137 } else {
4138 if (conn_param)
4139 ret = cma_accept_ib(id_priv, conn_param);
4140 else
4141 ret = cma_rep_recv(id_priv);
4142 }
4143 } else if (rdma_cap_iw_cm(id->device, id->port_num))
4144 ret = cma_accept_iw(id_priv, conn_param);
4145 else
4146 ret = -ENOSYS;
4147
4148 if (ret)
4149 goto reject;
4150
4151 return 0;
4152 reject:
4153 cma_modify_qp_err(id_priv);
4154 rdma_reject(id, NULL, 0);
4155 return ret;
4156 }
4157 EXPORT_SYMBOL(__rdma_accept);
4158
4159 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
4160 {
4161 struct rdma_id_private *id_priv;
4162 int ret;
4163
4164 id_priv = container_of(id, struct rdma_id_private, id);
4165 if (!id_priv->cm_id.ib)
4166 return -EINVAL;
4167
4168 switch (id->device->node_type) {
4169 case RDMA_NODE_IB_CA:
4170 ret = ib_cm_notify(id_priv->cm_id.ib, event);
4171 break;
4172 default:
4173 ret = 0;
4174 break;
4175 }
4176 return ret;
4177 }
4178 EXPORT_SYMBOL(rdma_notify);
4179
4180 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
4181 u8 private_data_len)
4182 {
4183 struct rdma_id_private *id_priv;
4184 int ret;
4185
4186 id_priv = container_of(id, struct rdma_id_private, id);
4187 if (!id_priv->cm_id.ib)
4188 return -EINVAL;
4189
4190 if (rdma_cap_ib_cm(id->device, id->port_num)) {
4191 if (id->qp_type == IB_QPT_UD) {
4192 ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT, 0,
4193 private_data, private_data_len);
4194 } else {
4195 trace_cm_send_rej(id_priv);
4196 ret = ib_send_cm_rej(id_priv->cm_id.ib,
4197 IB_CM_REJ_CONSUMER_DEFINED, NULL,
4198 0, private_data, private_data_len);
4199 }
4200 } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4201 ret = iw_cm_reject(id_priv->cm_id.iw,
4202 private_data, private_data_len);
4203 } else
4204 ret = -ENOSYS;
4205
4206 return ret;
4207 }
4208 EXPORT_SYMBOL(rdma_reject);
4209
4210 int rdma_disconnect(struct rdma_cm_id *id)
4211 {
4212 struct rdma_id_private *id_priv;
4213 int ret;
4214
4215 id_priv = container_of(id, struct rdma_id_private, id);
4216 if (!id_priv->cm_id.ib)
4217 return -EINVAL;
4218
4219 if (rdma_cap_ib_cm(id->device, id->port_num)) {
4220 ret = cma_modify_qp_err(id_priv);
4221 if (ret)
4222 goto out;
4223 /* Initiate or respond to a disconnect. */
4224 trace_cm_disconnect(id_priv);
4225 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0)) {
4226 if (!ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0))
4227 trace_cm_sent_drep(id_priv);
4228 } else {
4229 trace_cm_sent_dreq(id_priv);
4230 }
4231 } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4232 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
4233 } else
4234 ret = -EINVAL;
4235
4236 out:
4237 return ret;
4238 }
4239 EXPORT_SYMBOL(rdma_disconnect);
4240
4241 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
4242 {
4243 struct rdma_id_private *id_priv;
4244 struct cma_multicast *mc = multicast->context;
4245 struct rdma_cm_event event = {};
4246 int ret = 0;
4247
4248 id_priv = mc->id_priv;
4249 mutex_lock(&id_priv->handler_mutex);
4250 if (id_priv->state != RDMA_CM_ADDR_BOUND &&
4251 id_priv->state != RDMA_CM_ADDR_RESOLVED)
4252 goto out;
4253
4254 if (!status)
4255 status = cma_set_qkey(id_priv, be32_to_cpu(multicast->rec.qkey));
4256 else
4257 pr_debug_ratelimited("RDMA CM: MULTICAST_ERROR: failed to join multicast. status %d\n",
4258 status);
4259 mutex_lock(&id_priv->qp_mutex);
4260 if (!status && id_priv->id.qp) {
4261 status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid,
4262 be16_to_cpu(multicast->rec.mlid));
4263 if (status)
4264 pr_debug_ratelimited("RDMA CM: MULTICAST_ERROR: failed to attach QP. status %d\n",
4265 status);
4266 }
4267 mutex_unlock(&id_priv->qp_mutex);
4268
4269 event.status = status;
4270 event.param.ud.private_data = mc->context;
4271 if (!status) {
4272 struct rdma_dev_addr *dev_addr =
4273 &id_priv->id.route.addr.dev_addr;
4274 struct net_device *ndev =
4275 dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if);
4276 enum ib_gid_type gid_type =
4277 id_priv->cma_dev->default_gid_type[id_priv->id.port_num -
4278 rdma_start_port(id_priv->cma_dev->device)];
4279
4280 event.event = RDMA_CM_EVENT_MULTICAST_JOIN;
4281 ret = ib_init_ah_from_mcmember(id_priv->id.device,
4282 id_priv->id.port_num,
4283 &multicast->rec,
4284 ndev, gid_type,
4285 &event.param.ud.ah_attr);
4286 if (ret)
4287 event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
4288
4289 event.param.ud.qp_num = 0xFFFFFF;
4290 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
4291 if (ndev)
4292 dev_put(ndev);
4293 } else
4294 event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
4295
4296 ret = cma_cm_event_handler(id_priv, &event);
4297
4298 rdma_destroy_ah_attr(&event.param.ud.ah_attr);
4299 if (ret) {
4300 cma_exch(id_priv, RDMA_CM_DESTROYING);
4301 mutex_unlock(&id_priv->handler_mutex);
4302 rdma_destroy_id(&id_priv->id);
4303 return 0;
4304 }
4305
4306 out:
4307 mutex_unlock(&id_priv->handler_mutex);
4308 return 0;
4309 }
4310
4311 static void cma_set_mgid(struct rdma_id_private *id_priv,
4312 struct sockaddr *addr, union ib_gid *mgid)
4313 {
4314 unsigned char mc_map[MAX_ADDR_LEN];
4315 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4316 struct sockaddr_in *sin = (struct sockaddr_in *) addr;
4317 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
4318
4319 if (cma_any_addr(addr)) {
4320 memset(mgid, 0, sizeof *mgid);
4321 } else if ((addr->sa_family == AF_INET6) &&
4322 ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
4323 0xFF10A01B)) {
4324 /* IPv6 address is an SA assigned MGID. */
4325 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
4326 } else if (addr->sa_family == AF_IB) {
4327 memcpy(mgid, &((struct sockaddr_ib *) addr)->sib_addr, sizeof *mgid);
4328 } else if (addr->sa_family == AF_INET6) {
4329 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
4330 if (id_priv->id.ps == RDMA_PS_UDP)
4331 mc_map[7] = 0x01; /* Use RDMA CM signature */
4332 *mgid = *(union ib_gid *) (mc_map + 4);
4333 } else {
4334 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
4335 if (id_priv->id.ps == RDMA_PS_UDP)
4336 mc_map[7] = 0x01; /* Use RDMA CM signature */
4337 *mgid = *(union ib_gid *) (mc_map + 4);
4338 }
4339 }
4340
4341 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
4342 struct cma_multicast *mc)
4343 {
4344 struct ib_sa_mcmember_rec rec;
4345 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4346 ib_sa_comp_mask comp_mask;
4347 int ret;
4348
4349 ib_addr_get_mgid(dev_addr, &rec.mgid);
4350 ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
4351 &rec.mgid, &rec);
4352 if (ret)
4353 return ret;
4354
4355 ret = cma_set_qkey(id_priv, 0);
4356 if (ret)
4357 return ret;
4358
4359 cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
4360 rec.qkey = cpu_to_be32(id_priv->qkey);
4361 rdma_addr_get_sgid(dev_addr, &rec.port_gid);
4362 rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
4363 rec.join_state = mc->join_state;
4364
4365 if ((rec.join_state == BIT(SENDONLY_FULLMEMBER_JOIN)) &&
4366 (!ib_sa_sendonly_fullmem_support(&sa_client,
4367 id_priv->id.device,
4368 id_priv->id.port_num))) {
4369 dev_warn(
4370 &id_priv->id.device->dev,
4371 "RDMA CM: port %u Unable to multicast join: SM doesn't support Send Only Full Member option\n",
4372 id_priv->id.port_num);
4373 return -EOPNOTSUPP;
4374 }
4375
4376 comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
4377 IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
4378 IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
4379 IB_SA_MCMEMBER_REC_FLOW_LABEL |
4380 IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
4381
4382 if (id_priv->id.ps == RDMA_PS_IPOIB)
4383 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
4384 IB_SA_MCMEMBER_REC_RATE_SELECTOR |
4385 IB_SA_MCMEMBER_REC_MTU_SELECTOR |
4386 IB_SA_MCMEMBER_REC_MTU |
4387 IB_SA_MCMEMBER_REC_HOP_LIMIT;
4388
4389 mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device,
4390 id_priv->id.port_num, &rec,
4391 comp_mask, GFP_KERNEL,
4392 cma_ib_mc_handler, mc);
4393 return PTR_ERR_OR_ZERO(mc->multicast.ib);
4394 }
4395
4396 static void iboe_mcast_work_handler(struct work_struct *work)
4397 {
4398 struct iboe_mcast_work *mw = container_of(work, struct iboe_mcast_work, work);
4399 struct cma_multicast *mc = mw->mc;
4400 struct ib_sa_multicast *m = mc->multicast.ib;
4401
4402 mc->multicast.ib->context = mc;
4403 cma_ib_mc_handler(0, m);
4404 kref_put(&mc->mcref, release_mc);
4405 kfree(mw);
4406 }
4407
4408 static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid,
4409 enum ib_gid_type gid_type)
4410 {
4411 struct sockaddr_in *sin = (struct sockaddr_in *)addr;
4412 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
4413
4414 if (cma_any_addr(addr)) {
4415 memset(mgid, 0, sizeof *mgid);
4416 } else if (addr->sa_family == AF_INET6) {
4417 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
4418 } else {
4419 mgid->raw[0] =
4420 (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ? 0 : 0xff;
4421 mgid->raw[1] =
4422 (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ? 0 : 0x0e;
4423 mgid->raw[2] = 0;
4424 mgid->raw[3] = 0;
4425 mgid->raw[4] = 0;
4426 mgid->raw[5] = 0;
4427 mgid->raw[6] = 0;
4428 mgid->raw[7] = 0;
4429 mgid->raw[8] = 0;
4430 mgid->raw[9] = 0;
4431 mgid->raw[10] = 0xff;
4432 mgid->raw[11] = 0xff;
4433 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
4434 }
4435 }
4436
4437 static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
4438 struct cma_multicast *mc)
4439 {
4440 struct iboe_mcast_work *work;
4441 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4442 int err = 0;
4443 struct sockaddr *addr = (struct sockaddr *)&mc->addr;
4444 struct net_device *ndev = NULL;
4445 enum ib_gid_type gid_type;
4446 bool send_only;
4447
4448 send_only = mc->join_state == BIT(SENDONLY_FULLMEMBER_JOIN);
4449
4450 if (cma_zero_addr((struct sockaddr *)&mc->addr))
4451 return -EINVAL;
4452
4453 work = kzalloc(sizeof *work, GFP_KERNEL);
4454 if (!work)
4455 return -ENOMEM;
4456
4457 mc->multicast.ib = kzalloc(sizeof(struct ib_sa_multicast), GFP_KERNEL);
4458 if (!mc->multicast.ib) {
4459 err = -ENOMEM;
4460 goto out1;
4461 }
4462
4463 gid_type = id_priv->cma_dev->default_gid_type[id_priv->id.port_num -
4464 rdma_start_port(id_priv->cma_dev->device)];
4465 cma_iboe_set_mgid(addr, &mc->multicast.ib->rec.mgid, gid_type);
4466
4467 mc->multicast.ib->rec.pkey = cpu_to_be16(0xffff);
4468 if (id_priv->id.ps == RDMA_PS_UDP)
4469 mc->multicast.ib->rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
4470
4471 if (dev_addr->bound_dev_if)
4472 ndev = dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if);
4473 if (!ndev) {
4474 err = -ENODEV;
4475 goto out2;
4476 }
4477 mc->multicast.ib->rec.rate = iboe_get_rate(ndev);
4478 mc->multicast.ib->rec.hop_limit = 1;
4479 mc->multicast.ib->rec.mtu = iboe_get_mtu(ndev->mtu);
4480
4481 if (addr->sa_family == AF_INET) {
4482 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) {
4483 mc->multicast.ib->rec.hop_limit = IPV6_DEFAULT_HOPLIMIT;
4484 if (!send_only) {
4485 err = cma_igmp_send(ndev, &mc->multicast.ib->rec.mgid,
4486 true);
4487 }
4488 }
4489 } else {
4490 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP)
4491 err = -ENOTSUPP;
4492 }
4493 dev_put(ndev);
4494 if (err || !mc->multicast.ib->rec.mtu) {
4495 if (!err)
4496 err = -EINVAL;
4497 goto out2;
4498 }
4499 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
4500 &mc->multicast.ib->rec.port_gid);
4501 work->id = id_priv;
4502 work->mc = mc;
4503 INIT_WORK(&work->work, iboe_mcast_work_handler);
4504 kref_get(&mc->mcref);
4505 queue_work(cma_wq, &work->work);
4506
4507 return 0;
4508
4509 out2:
4510 kfree(mc->multicast.ib);
4511 out1:
4512 kfree(work);
4513 return err;
4514 }
4515
4516 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
4517 u8 join_state, void *context)
4518 {
4519 struct rdma_id_private *id_priv;
4520 struct cma_multicast *mc;
4521 int ret;
4522
4523 if (!id->device)
4524 return -EINVAL;
4525
4526 id_priv = container_of(id, struct rdma_id_private, id);
4527 if (!cma_comp(id_priv, RDMA_CM_ADDR_BOUND) &&
4528 !cma_comp(id_priv, RDMA_CM_ADDR_RESOLVED))
4529 return -EINVAL;
4530
4531 mc = kmalloc(sizeof *mc, GFP_KERNEL);
4532 if (!mc)
4533 return -ENOMEM;
4534
4535 memcpy(&mc->addr, addr, rdma_addr_size(addr));
4536 mc->context = context;
4537 mc->id_priv = id_priv;
4538 mc->join_state = join_state;
4539
4540 if (rdma_protocol_roce(id->device, id->port_num)) {
4541 kref_init(&mc->mcref);
4542 ret = cma_iboe_join_multicast(id_priv, mc);
4543 if (ret)
4544 goto out_err;
4545 } else if (rdma_cap_ib_mcast(id->device, id->port_num)) {
4546 ret = cma_join_ib_multicast(id_priv, mc);
4547 if (ret)
4548 goto out_err;
4549 } else {
4550 ret = -ENOSYS;
4551 goto out_err;
4552 }
4553
4554 spin_lock(&id_priv->lock);
4555 list_add(&mc->list, &id_priv->mc_list);
4556 spin_unlock(&id_priv->lock);
4557
4558 return 0;
4559 out_err:
4560 kfree(mc);
4561 return ret;
4562 }
4563 EXPORT_SYMBOL(rdma_join_multicast);
4564
4565 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
4566 {
4567 struct rdma_id_private *id_priv;
4568 struct cma_multicast *mc;
4569
4570 id_priv = container_of(id, struct rdma_id_private, id);
4571 spin_lock_irq(&id_priv->lock);
4572 list_for_each_entry(mc, &id_priv->mc_list, list) {
4573 if (!memcmp(&mc->addr, addr, rdma_addr_size(addr))) {
4574 list_del(&mc->list);
4575 spin_unlock_irq(&id_priv->lock);
4576
4577 if (id->qp)
4578 ib_detach_mcast(id->qp,
4579 &mc->multicast.ib->rec.mgid,
4580 be16_to_cpu(mc->multicast.ib->rec.mlid));
4581
4582 BUG_ON(id_priv->cma_dev->device != id->device);
4583
4584 if (rdma_cap_ib_mcast(id->device, id->port_num)) {
4585 ib_sa_free_multicast(mc->multicast.ib);
4586 kfree(mc);
4587 } else if (rdma_protocol_roce(id->device, id->port_num)) {
4588 cma_leave_roce_mc_group(id_priv, mc);
4589 }
4590 return;
4591 }
4592 }
4593 spin_unlock_irq(&id_priv->lock);
4594 }
4595 EXPORT_SYMBOL(rdma_leave_multicast);
4596
4597 static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
4598 {
4599 struct rdma_dev_addr *dev_addr;
4600 struct cma_ndev_work *work;
4601
4602 dev_addr = &id_priv->id.route.addr.dev_addr;
4603
4604 if ((dev_addr->bound_dev_if == ndev->ifindex) &&
4605 (net_eq(dev_net(ndev), dev_addr->net)) &&
4606 memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
4607 pr_info("RDMA CM addr change for ndev %s used by id %p\n",
4608 ndev->name, &id_priv->id);
4609 work = kzalloc(sizeof *work, GFP_KERNEL);
4610 if (!work)
4611 return -ENOMEM;
4612
4613 INIT_WORK(&work->work, cma_ndev_work_handler);
4614 work->id = id_priv;
4615 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
4616 cma_id_get(id_priv);
4617 queue_work(cma_wq, &work->work);
4618 }
4619
4620 return 0;
4621 }
4622
4623 static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
4624 void *ptr)
4625 {
4626 struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
4627 struct cma_device *cma_dev;
4628 struct rdma_id_private *id_priv;
4629 int ret = NOTIFY_DONE;
4630
4631 if (event != NETDEV_BONDING_FAILOVER)
4632 return NOTIFY_DONE;
4633
4634 if (!netif_is_bond_master(ndev))
4635 return NOTIFY_DONE;
4636
4637 mutex_lock(&lock);
4638 list_for_each_entry(cma_dev, &dev_list, list)
4639 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
4640 ret = cma_netdev_change(ndev, id_priv);
4641 if (ret)
4642 goto out;
4643 }
4644
4645 out:
4646 mutex_unlock(&lock);
4647 return ret;
4648 }
4649
4650 static struct notifier_block cma_nb = {
4651 .notifier_call = cma_netdev_callback
4652 };
4653
4654 static void cma_add_one(struct ib_device *device)
4655 {
4656 struct cma_device *cma_dev;
4657 struct rdma_id_private *id_priv;
4658 unsigned int i;
4659 unsigned long supported_gids = 0;
4660
4661 cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
4662 if (!cma_dev)
4663 return;
4664
4665 cma_dev->device = device;
4666 cma_dev->default_gid_type = kcalloc(device->phys_port_cnt,
4667 sizeof(*cma_dev->default_gid_type),
4668 GFP_KERNEL);
4669 if (!cma_dev->default_gid_type)
4670 goto free_cma_dev;
4671
4672 cma_dev->default_roce_tos = kcalloc(device->phys_port_cnt,
4673 sizeof(*cma_dev->default_roce_tos),
4674 GFP_KERNEL);
4675 if (!cma_dev->default_roce_tos)
4676 goto free_gid_type;
4677
4678 rdma_for_each_port (device, i) {
4679 supported_gids = roce_gid_type_mask_support(device, i);
4680 WARN_ON(!supported_gids);
4681 if (supported_gids & (1 << CMA_PREFERRED_ROCE_GID_TYPE))
4682 cma_dev->default_gid_type[i - rdma_start_port(device)] =
4683 CMA_PREFERRED_ROCE_GID_TYPE;
4684 else
4685 cma_dev->default_gid_type[i - rdma_start_port(device)] =
4686 find_first_bit(&supported_gids, BITS_PER_LONG);
4687 cma_dev->default_roce_tos[i - rdma_start_port(device)] = 0;
4688 }
4689
4690 init_completion(&cma_dev->comp);
4691 refcount_set(&cma_dev->refcount, 1);
4692 INIT_LIST_HEAD(&cma_dev->id_list);
4693 ib_set_client_data(device, &cma_client, cma_dev);
4694
4695 mutex_lock(&lock);
4696 list_add_tail(&cma_dev->list, &dev_list);
4697 list_for_each_entry(id_priv, &listen_any_list, list)
4698 cma_listen_on_dev(id_priv, cma_dev);
4699 mutex_unlock(&lock);
4700
4701 trace_cm_add_one(device);
4702 return;
4703
4704 free_gid_type:
4705 kfree(cma_dev->default_gid_type);
4706
4707 free_cma_dev:
4708 kfree(cma_dev);
4709
4710 return;
4711 }
4712
4713 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
4714 {
4715 struct rdma_cm_event event = {};
4716 enum rdma_cm_state state;
4717 int ret = 0;
4718
4719 /* Record that we want to remove the device */
4720 state = cma_exch(id_priv, RDMA_CM_DEVICE_REMOVAL);
4721 if (state == RDMA_CM_DESTROYING)
4722 return 0;
4723
4724 cma_cancel_operation(id_priv, state);
4725 mutex_lock(&id_priv->handler_mutex);
4726
4727 /* Check for destruction from another callback. */
4728 if (!cma_comp(id_priv, RDMA_CM_DEVICE_REMOVAL))
4729 goto out;
4730
4731 event.event = RDMA_CM_EVENT_DEVICE_REMOVAL;
4732 ret = cma_cm_event_handler(id_priv, &event);
4733 out:
4734 mutex_unlock(&id_priv->handler_mutex);
4735 return ret;
4736 }
4737
4738 static void cma_process_remove(struct cma_device *cma_dev)
4739 {
4740 struct rdma_id_private *id_priv;
4741 int ret;
4742
4743 mutex_lock(&lock);
4744 while (!list_empty(&cma_dev->id_list)) {
4745 id_priv = list_entry(cma_dev->id_list.next,
4746 struct rdma_id_private, list);
4747
4748 list_del(&id_priv->listen_list);
4749 list_del_init(&id_priv->list);
4750 cma_id_get(id_priv);
4751 mutex_unlock(&lock);
4752
4753 ret = id_priv->internal_id ? 1 : cma_remove_id_dev(id_priv);
4754 cma_id_put(id_priv);
4755 if (ret)
4756 rdma_destroy_id(&id_priv->id);
4757
4758 mutex_lock(&lock);
4759 }
4760 mutex_unlock(&lock);
4761
4762 cma_dev_put(cma_dev);
4763 wait_for_completion(&cma_dev->comp);
4764 }
4765
4766 static void cma_remove_one(struct ib_device *device, void *client_data)
4767 {
4768 struct cma_device *cma_dev = client_data;
4769
4770 trace_cm_remove_one(device);
4771
4772 if (!cma_dev)
4773 return;
4774
4775 mutex_lock(&lock);
4776 list_del(&cma_dev->list);
4777 mutex_unlock(&lock);
4778
4779 cma_process_remove(cma_dev);
4780 kfree(cma_dev->default_roce_tos);
4781 kfree(cma_dev->default_gid_type);
4782 kfree(cma_dev);
4783 }
4784
4785 static int cma_init_net(struct net *net)
4786 {
4787 struct cma_pernet *pernet = cma_pernet(net);
4788
4789 xa_init(&pernet->tcp_ps);
4790 xa_init(&pernet->udp_ps);
4791 xa_init(&pernet->ipoib_ps);
4792 xa_init(&pernet->ib_ps);
4793
4794 return 0;
4795 }
4796
4797 static void cma_exit_net(struct net *net)
4798 {
4799 struct cma_pernet *pernet = cma_pernet(net);
4800
4801 WARN_ON(!xa_empty(&pernet->tcp_ps));
4802 WARN_ON(!xa_empty(&pernet->udp_ps));
4803 WARN_ON(!xa_empty(&pernet->ipoib_ps));
4804 WARN_ON(!xa_empty(&pernet->ib_ps));
4805 }
4806
4807 static struct pernet_operations cma_pernet_operations = {
4808 .init = cma_init_net,
4809 .exit = cma_exit_net,
4810 .id = &cma_pernet_id,
4811 .size = sizeof(struct cma_pernet),
4812 };
4813
4814 static int __init cma_init(void)
4815 {
4816 int ret;
4817
4818 /*
4819 * There is a rare lock ordering dependency in cma_netdev_callback()
4820 * that only happens when bonding is enabled. Teach lockdep that rtnl
4821 * must never be nested under lock so it can find these without having
4822 * to test with bonding.
4823 */
4824 if (IS_ENABLED(CONFIG_LOCKDEP)) {
4825 rtnl_lock();
4826 mutex_lock(&lock);
4827 mutex_unlock(&lock);
4828 rtnl_unlock();
4829 }
4830
4831 cma_wq = alloc_ordered_workqueue("rdma_cm", WQ_MEM_RECLAIM);
4832 if (!cma_wq)
4833 return -ENOMEM;
4834
4835 ret = register_pernet_subsys(&cma_pernet_operations);
4836 if (ret)
4837 goto err_wq;
4838
4839 ib_sa_register_client(&sa_client);
4840 register_netdevice_notifier(&cma_nb);
4841
4842 ret = ib_register_client(&cma_client);
4843 if (ret)
4844 goto err;
4845
4846 ret = cma_configfs_init();
4847 if (ret)
4848 goto err_ib;
4849
4850 return 0;
4851
4852 err_ib:
4853 ib_unregister_client(&cma_client);
4854 err:
4855 unregister_netdevice_notifier(&cma_nb);
4856 ib_sa_unregister_client(&sa_client);
4857 unregister_pernet_subsys(&cma_pernet_operations);
4858 err_wq:
4859 destroy_workqueue(cma_wq);
4860 return ret;
4861 }
4862
4863 static void __exit cma_cleanup(void)
4864 {
4865 cma_configfs_exit();
4866 ib_unregister_client(&cma_client);
4867 unregister_netdevice_notifier(&cma_nb);
4868 ib_sa_unregister_client(&sa_client);
4869 unregister_pernet_subsys(&cma_pernet_operations);
4870 destroy_workqueue(cma_wq);
4871 }
4872
4873 module_init(cma_init);
4874 module_exit(cma_cleanup);
Linux with added FPC-III support