Loading...
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/rbtree.h>
15#include <linux/igmp.h>
16#include <linux/xarray.h>
17#include <linux/inetdevice.h>
18#include <linux/slab.h>
19#include <linux/module.h>
20#include <net/route.h>
21
22#include <net/net_namespace.h>
23#include <net/netns/generic.h>
24#include <net/netevent.h>
25#include <net/tcp.h>
26#include <net/ipv6.h>
27#include <net/ip_fib.h>
28#include <net/ip6_route.h>
29
30#include <rdma/rdma_cm.h>
31#include <rdma/rdma_cm_ib.h>
32#include <rdma/rdma_netlink.h>
33#include <rdma/ib.h>
34#include <rdma/ib_cache.h>
35#include <rdma/ib_cm.h>
36#include <rdma/ib_sa.h>
37#include <rdma/iw_cm.h>
38
39#include "core_priv.h"
40#include "cma_priv.h"
41#include "cma_trace.h"
42
43MODULE_AUTHOR("Sean Hefty");
44MODULE_DESCRIPTION("Generic RDMA CM Agent");
45MODULE_LICENSE("Dual BSD/GPL");
46
47#define CMA_CM_RESPONSE_TIMEOUT 20
48#define CMA_MAX_CM_RETRIES 15
49#define CMA_CM_MRA_SETTING (IB_CM_MRA_FLAG_DELAY | 24)
50#define CMA_IBOE_PACKET_LIFETIME 16
51#define CMA_PREFERRED_ROCE_GID_TYPE IB_GID_TYPE_ROCE_UDP_ENCAP
52
53static const char * const cma_events[] = {
54 [RDMA_CM_EVENT_ADDR_RESOLVED] = "address resolved",
55 [RDMA_CM_EVENT_ADDR_ERROR] = "address error",
56 [RDMA_CM_EVENT_ROUTE_RESOLVED] = "route resolved ",
57 [RDMA_CM_EVENT_ROUTE_ERROR] = "route error",
58 [RDMA_CM_EVENT_CONNECT_REQUEST] = "connect request",
59 [RDMA_CM_EVENT_CONNECT_RESPONSE] = "connect response",
60 [RDMA_CM_EVENT_CONNECT_ERROR] = "connect error",
61 [RDMA_CM_EVENT_UNREACHABLE] = "unreachable",
62 [RDMA_CM_EVENT_REJECTED] = "rejected",
63 [RDMA_CM_EVENT_ESTABLISHED] = "established",
64 [RDMA_CM_EVENT_DISCONNECTED] = "disconnected",
65 [RDMA_CM_EVENT_DEVICE_REMOVAL] = "device removal",
66 [RDMA_CM_EVENT_MULTICAST_JOIN] = "multicast join",
67 [RDMA_CM_EVENT_MULTICAST_ERROR] = "multicast error",
68 [RDMA_CM_EVENT_ADDR_CHANGE] = "address change",
69 [RDMA_CM_EVENT_TIMEWAIT_EXIT] = "timewait exit",
70};
71
72static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid,
73 enum ib_gid_type gid_type);
74
75const char *__attribute_const__ rdma_event_msg(enum rdma_cm_event_type event)
76{
77 size_t index = event;
78
79 return (index < ARRAY_SIZE(cma_events) && cma_events[index]) ?
80 cma_events[index] : "unrecognized event";
81}
82EXPORT_SYMBOL(rdma_event_msg);
83
84const char *__attribute_const__ rdma_reject_msg(struct rdma_cm_id *id,
85 int reason)
86{
87 if (rdma_ib_or_roce(id->device, id->port_num))
88 return ibcm_reject_msg(reason);
89
90 if (rdma_protocol_iwarp(id->device, id->port_num))
91 return iwcm_reject_msg(reason);
92
93 WARN_ON_ONCE(1);
94 return "unrecognized transport";
95}
96EXPORT_SYMBOL(rdma_reject_msg);
97
98/**
99 * rdma_is_consumer_reject - return true if the consumer rejected the connect
100 * request.
101 * @id: Communication identifier that received the REJECT event.
102 * @reason: Value returned in the REJECT event status field.
103 */
104static bool rdma_is_consumer_reject(struct rdma_cm_id *id, int reason)
105{
106 if (rdma_ib_or_roce(id->device, id->port_num))
107 return reason == IB_CM_REJ_CONSUMER_DEFINED;
108
109 if (rdma_protocol_iwarp(id->device, id->port_num))
110 return reason == -ECONNREFUSED;
111
112 WARN_ON_ONCE(1);
113 return false;
114}
115
116const void *rdma_consumer_reject_data(struct rdma_cm_id *id,
117 struct rdma_cm_event *ev, u8 *data_len)
118{
119 const void *p;
120
121 if (rdma_is_consumer_reject(id, ev->status)) {
122 *data_len = ev->param.conn.private_data_len;
123 p = ev->param.conn.private_data;
124 } else {
125 *data_len = 0;
126 p = NULL;
127 }
128 return p;
129}
130EXPORT_SYMBOL(rdma_consumer_reject_data);
131
132/**
133 * rdma_iw_cm_id() - return the iw_cm_id pointer for this cm_id.
134 * @id: Communication Identifier
135 */
136struct iw_cm_id *rdma_iw_cm_id(struct rdma_cm_id *id)
137{
138 struct rdma_id_private *id_priv;
139
140 id_priv = container_of(id, struct rdma_id_private, id);
141 if (id->device->node_type == RDMA_NODE_RNIC)
142 return id_priv->cm_id.iw;
143 return NULL;
144}
145EXPORT_SYMBOL(rdma_iw_cm_id);
146
147/**
148 * rdma_res_to_id() - return the rdma_cm_id pointer for this restrack.
149 * @res: rdma resource tracking entry pointer
150 */
151struct rdma_cm_id *rdma_res_to_id(struct rdma_restrack_entry *res)
152{
153 struct rdma_id_private *id_priv =
154 container_of(res, struct rdma_id_private, res);
155
156 return &id_priv->id;
157}
158EXPORT_SYMBOL(rdma_res_to_id);
159
160static int cma_add_one(struct ib_device *device);
161static void cma_remove_one(struct ib_device *device, void *client_data);
162
163static struct ib_client cma_client = {
164 .name = "cma",
165 .add = cma_add_one,
166 .remove = cma_remove_one
167};
168
169static struct ib_sa_client sa_client;
170static LIST_HEAD(dev_list);
171static LIST_HEAD(listen_any_list);
172static DEFINE_MUTEX(lock);
173static struct rb_root id_table = RB_ROOT;
174/* Serialize operations of id_table tree */
175static DEFINE_SPINLOCK(id_table_lock);
176static struct workqueue_struct *cma_wq;
177static unsigned int cma_pernet_id;
178
179struct cma_pernet {
180 struct xarray tcp_ps;
181 struct xarray udp_ps;
182 struct xarray ipoib_ps;
183 struct xarray ib_ps;
184};
185
186static struct cma_pernet *cma_pernet(struct net *net)
187{
188 return net_generic(net, cma_pernet_id);
189}
190
191static
192struct xarray *cma_pernet_xa(struct net *net, enum rdma_ucm_port_space ps)
193{
194 struct cma_pernet *pernet = cma_pernet(net);
195
196 switch (ps) {
197 case RDMA_PS_TCP:
198 return &pernet->tcp_ps;
199 case RDMA_PS_UDP:
200 return &pernet->udp_ps;
201 case RDMA_PS_IPOIB:
202 return &pernet->ipoib_ps;
203 case RDMA_PS_IB:
204 return &pernet->ib_ps;
205 default:
206 return NULL;
207 }
208}
209
210struct id_table_entry {
211 struct list_head id_list;
212 struct rb_node rb_node;
213};
214
215struct cma_device {
216 struct list_head list;
217 struct ib_device *device;
218 struct completion comp;
219 refcount_t refcount;
220 struct list_head id_list;
221 enum ib_gid_type *default_gid_type;
222 u8 *default_roce_tos;
223};
224
225struct rdma_bind_list {
226 enum rdma_ucm_port_space ps;
227 struct hlist_head owners;
228 unsigned short port;
229};
230
231static int cma_ps_alloc(struct net *net, enum rdma_ucm_port_space ps,
232 struct rdma_bind_list *bind_list, int snum)
233{
234 struct xarray *xa = cma_pernet_xa(net, ps);
235
236 return xa_insert(xa, snum, bind_list, GFP_KERNEL);
237}
238
239static struct rdma_bind_list *cma_ps_find(struct net *net,
240 enum rdma_ucm_port_space ps, int snum)
241{
242 struct xarray *xa = cma_pernet_xa(net, ps);
243
244 return xa_load(xa, snum);
245}
246
247static void cma_ps_remove(struct net *net, enum rdma_ucm_port_space ps,
248 int snum)
249{
250 struct xarray *xa = cma_pernet_xa(net, ps);
251
252 xa_erase(xa, snum);
253}
254
255enum {
256 CMA_OPTION_AFONLY,
257};
258
259void cma_dev_get(struct cma_device *cma_dev)
260{
261 refcount_inc(&cma_dev->refcount);
262}
263
264void cma_dev_put(struct cma_device *cma_dev)
265{
266 if (refcount_dec_and_test(&cma_dev->refcount))
267 complete(&cma_dev->comp);
268}
269
270struct cma_device *cma_enum_devices_by_ibdev(cma_device_filter filter,
271 void *cookie)
272{
273 struct cma_device *cma_dev;
274 struct cma_device *found_cma_dev = NULL;
275
276 mutex_lock(&lock);
277
278 list_for_each_entry(cma_dev, &dev_list, list)
279 if (filter(cma_dev->device, cookie)) {
280 found_cma_dev = cma_dev;
281 break;
282 }
283
284 if (found_cma_dev)
285 cma_dev_get(found_cma_dev);
286 mutex_unlock(&lock);
287 return found_cma_dev;
288}
289
290int cma_get_default_gid_type(struct cma_device *cma_dev,
291 u32 port)
292{
293 if (!rdma_is_port_valid(cma_dev->device, port))
294 return -EINVAL;
295
296 return cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)];
297}
298
299int cma_set_default_gid_type(struct cma_device *cma_dev,
300 u32 port,
301 enum ib_gid_type default_gid_type)
302{
303 unsigned long supported_gids;
304
305 if (!rdma_is_port_valid(cma_dev->device, port))
306 return -EINVAL;
307
308 if (default_gid_type == IB_GID_TYPE_IB &&
309 rdma_protocol_roce_eth_encap(cma_dev->device, port))
310 default_gid_type = IB_GID_TYPE_ROCE;
311
312 supported_gids = roce_gid_type_mask_support(cma_dev->device, port);
313
314 if (!(supported_gids & 1 << default_gid_type))
315 return -EINVAL;
316
317 cma_dev->default_gid_type[port - rdma_start_port(cma_dev->device)] =
318 default_gid_type;
319
320 return 0;
321}
322
323int cma_get_default_roce_tos(struct cma_device *cma_dev, u32 port)
324{
325 if (!rdma_is_port_valid(cma_dev->device, port))
326 return -EINVAL;
327
328 return cma_dev->default_roce_tos[port - rdma_start_port(cma_dev->device)];
329}
330
331int cma_set_default_roce_tos(struct cma_device *cma_dev, u32 port,
332 u8 default_roce_tos)
333{
334 if (!rdma_is_port_valid(cma_dev->device, port))
335 return -EINVAL;
336
337 cma_dev->default_roce_tos[port - rdma_start_port(cma_dev->device)] =
338 default_roce_tos;
339
340 return 0;
341}
342struct ib_device *cma_get_ib_dev(struct cma_device *cma_dev)
343{
344 return cma_dev->device;
345}
346
347/*
348 * Device removal can occur at anytime, so we need extra handling to
349 * serialize notifying the user of device removal with other callbacks.
350 * We do this by disabling removal notification while a callback is in process,
351 * and reporting it after the callback completes.
352 */
353
354struct cma_multicast {
355 struct rdma_id_private *id_priv;
356 union {
357 struct ib_sa_multicast *sa_mc;
358 struct {
359 struct work_struct work;
360 struct rdma_cm_event event;
361 } iboe_join;
362 };
363 struct list_head list;
364 void *context;
365 struct sockaddr_storage addr;
366 u8 join_state;
367};
368
369struct cma_work {
370 struct work_struct work;
371 struct rdma_id_private *id;
372 enum rdma_cm_state old_state;
373 enum rdma_cm_state new_state;
374 struct rdma_cm_event event;
375};
376
377union cma_ip_addr {
378 struct in6_addr ip6;
379 struct {
380 __be32 pad[3];
381 __be32 addr;
382 } ip4;
383};
384
385struct cma_hdr {
386 u8 cma_version;
387 u8 ip_version; /* IP version: 7:4 */
388 __be16 port;
389 union cma_ip_addr src_addr;
390 union cma_ip_addr dst_addr;
391};
392
393#define CMA_VERSION 0x00
394
395struct cma_req_info {
396 struct sockaddr_storage listen_addr_storage;
397 struct sockaddr_storage src_addr_storage;
398 struct ib_device *device;
399 union ib_gid local_gid;
400 __be64 service_id;
401 int port;
402 bool has_gid;
403 u16 pkey;
404};
405
406static int cma_comp_exch(struct rdma_id_private *id_priv,
407 enum rdma_cm_state comp, enum rdma_cm_state exch)
408{
409 unsigned long flags;
410 int ret;
411
412 /*
413 * The FSM uses a funny double locking where state is protected by both
414 * the handler_mutex and the spinlock. State is not allowed to change
415 * to/from a handler_mutex protected value without also holding
416 * handler_mutex.
417 */
418 if (comp == RDMA_CM_CONNECT || exch == RDMA_CM_CONNECT)
419 lockdep_assert_held(&id_priv->handler_mutex);
420
421 spin_lock_irqsave(&id_priv->lock, flags);
422 if ((ret = (id_priv->state == comp)))
423 id_priv->state = exch;
424 spin_unlock_irqrestore(&id_priv->lock, flags);
425 return ret;
426}
427
428static inline u8 cma_get_ip_ver(const struct cma_hdr *hdr)
429{
430 return hdr->ip_version >> 4;
431}
432
433static void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
434{
435 hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
436}
437
438static struct sockaddr *cma_src_addr(struct rdma_id_private *id_priv)
439{
440 return (struct sockaddr *)&id_priv->id.route.addr.src_addr;
441}
442
443static inline struct sockaddr *cma_dst_addr(struct rdma_id_private *id_priv)
444{
445 return (struct sockaddr *)&id_priv->id.route.addr.dst_addr;
446}
447
448static int cma_igmp_send(struct net_device *ndev, union ib_gid *mgid, bool join)
449{
450 struct in_device *in_dev = NULL;
451
452 if (ndev) {
453 rtnl_lock();
454 in_dev = __in_dev_get_rtnl(ndev);
455 if (in_dev) {
456 if (join)
457 ip_mc_inc_group(in_dev,
458 *(__be32 *)(mgid->raw + 12));
459 else
460 ip_mc_dec_group(in_dev,
461 *(__be32 *)(mgid->raw + 12));
462 }
463 rtnl_unlock();
464 }
465 return (in_dev) ? 0 : -ENODEV;
466}
467
468static int compare_netdev_and_ip(int ifindex_a, struct sockaddr *sa,
469 struct id_table_entry *entry_b)
470{
471 struct rdma_id_private *id_priv = list_first_entry(
472 &entry_b->id_list, struct rdma_id_private, id_list_entry);
473 int ifindex_b = id_priv->id.route.addr.dev_addr.bound_dev_if;
474 struct sockaddr *sb = cma_dst_addr(id_priv);
475
476 if (ifindex_a != ifindex_b)
477 return (ifindex_a > ifindex_b) ? 1 : -1;
478
479 if (sa->sa_family != sb->sa_family)
480 return sa->sa_family - sb->sa_family;
481
482 if (sa->sa_family == AF_INET &&
483 __builtin_object_size(sa, 0) >= sizeof(struct sockaddr_in)) {
484 return memcmp(&((struct sockaddr_in *)sa)->sin_addr,
485 &((struct sockaddr_in *)sb)->sin_addr,
486 sizeof(((struct sockaddr_in *)sa)->sin_addr));
487 }
488
489 if (sa->sa_family == AF_INET6 &&
490 __builtin_object_size(sa, 0) >= sizeof(struct sockaddr_in6)) {
491 return ipv6_addr_cmp(&((struct sockaddr_in6 *)sa)->sin6_addr,
492 &((struct sockaddr_in6 *)sb)->sin6_addr);
493 }
494
495 return -1;
496}
497
498static int cma_add_id_to_tree(struct rdma_id_private *node_id_priv)
499{
500 struct rb_node **new, *parent = NULL;
501 struct id_table_entry *this, *node;
502 unsigned long flags;
503 int result;
504
505 node = kzalloc(sizeof(*node), GFP_KERNEL);
506 if (!node)
507 return -ENOMEM;
508
509 spin_lock_irqsave(&id_table_lock, flags);
510 new = &id_table.rb_node;
511 while (*new) {
512 this = container_of(*new, struct id_table_entry, rb_node);
513 result = compare_netdev_and_ip(
514 node_id_priv->id.route.addr.dev_addr.bound_dev_if,
515 cma_dst_addr(node_id_priv), this);
516
517 parent = *new;
518 if (result < 0)
519 new = &((*new)->rb_left);
520 else if (result > 0)
521 new = &((*new)->rb_right);
522 else {
523 list_add_tail(&node_id_priv->id_list_entry,
524 &this->id_list);
525 kfree(node);
526 goto unlock;
527 }
528 }
529
530 INIT_LIST_HEAD(&node->id_list);
531 list_add_tail(&node_id_priv->id_list_entry, &node->id_list);
532
533 rb_link_node(&node->rb_node, parent, new);
534 rb_insert_color(&node->rb_node, &id_table);
535
536unlock:
537 spin_unlock_irqrestore(&id_table_lock, flags);
538 return 0;
539}
540
541static struct id_table_entry *
542node_from_ndev_ip(struct rb_root *root, int ifindex, struct sockaddr *sa)
543{
544 struct rb_node *node = root->rb_node;
545 struct id_table_entry *data;
546 int result;
547
548 while (node) {
549 data = container_of(node, struct id_table_entry, rb_node);
550 result = compare_netdev_and_ip(ifindex, sa, data);
551 if (result < 0)
552 node = node->rb_left;
553 else if (result > 0)
554 node = node->rb_right;
555 else
556 return data;
557 }
558
559 return NULL;
560}
561
562static void cma_remove_id_from_tree(struct rdma_id_private *id_priv)
563{
564 struct id_table_entry *data;
565 unsigned long flags;
566
567 spin_lock_irqsave(&id_table_lock, flags);
568 if (list_empty(&id_priv->id_list_entry))
569 goto out;
570
571 data = node_from_ndev_ip(&id_table,
572 id_priv->id.route.addr.dev_addr.bound_dev_if,
573 cma_dst_addr(id_priv));
574 if (!data)
575 goto out;
576
577 list_del_init(&id_priv->id_list_entry);
578 if (list_empty(&data->id_list)) {
579 rb_erase(&data->rb_node, &id_table);
580 kfree(data);
581 }
582out:
583 spin_unlock_irqrestore(&id_table_lock, flags);
584}
585
586static void _cma_attach_to_dev(struct rdma_id_private *id_priv,
587 struct cma_device *cma_dev)
588{
589 cma_dev_get(cma_dev);
590 id_priv->cma_dev = cma_dev;
591 id_priv->id.device = cma_dev->device;
592 id_priv->id.route.addr.dev_addr.transport =
593 rdma_node_get_transport(cma_dev->device->node_type);
594 list_add_tail(&id_priv->device_item, &cma_dev->id_list);
595
596 trace_cm_id_attach(id_priv, cma_dev->device);
597}
598
599static void cma_attach_to_dev(struct rdma_id_private *id_priv,
600 struct cma_device *cma_dev)
601{
602 _cma_attach_to_dev(id_priv, cma_dev);
603 id_priv->gid_type =
604 cma_dev->default_gid_type[id_priv->id.port_num -
605 rdma_start_port(cma_dev->device)];
606}
607
608static void cma_release_dev(struct rdma_id_private *id_priv)
609{
610 mutex_lock(&lock);
611 list_del_init(&id_priv->device_item);
612 cma_dev_put(id_priv->cma_dev);
613 id_priv->cma_dev = NULL;
614 id_priv->id.device = NULL;
615 if (id_priv->id.route.addr.dev_addr.sgid_attr) {
616 rdma_put_gid_attr(id_priv->id.route.addr.dev_addr.sgid_attr);
617 id_priv->id.route.addr.dev_addr.sgid_attr = NULL;
618 }
619 mutex_unlock(&lock);
620}
621
622static inline unsigned short cma_family(struct rdma_id_private *id_priv)
623{
624 return id_priv->id.route.addr.src_addr.ss_family;
625}
626
627static int cma_set_default_qkey(struct rdma_id_private *id_priv)
628{
629 struct ib_sa_mcmember_rec rec;
630 int ret = 0;
631
632 switch (id_priv->id.ps) {
633 case RDMA_PS_UDP:
634 case RDMA_PS_IB:
635 id_priv->qkey = RDMA_UDP_QKEY;
636 break;
637 case RDMA_PS_IPOIB:
638 ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid);
639 ret = ib_sa_get_mcmember_rec(id_priv->id.device,
640 id_priv->id.port_num, &rec.mgid,
641 &rec);
642 if (!ret)
643 id_priv->qkey = be32_to_cpu(rec.qkey);
644 break;
645 default:
646 break;
647 }
648 return ret;
649}
650
651static int cma_set_qkey(struct rdma_id_private *id_priv, u32 qkey)
652{
653 if (!qkey ||
654 (id_priv->qkey && (id_priv->qkey != qkey)))
655 return -EINVAL;
656
657 id_priv->qkey = qkey;
658 return 0;
659}
660
661static void cma_translate_ib(struct sockaddr_ib *sib, struct rdma_dev_addr *dev_addr)
662{
663 dev_addr->dev_type = ARPHRD_INFINIBAND;
664 rdma_addr_set_sgid(dev_addr, (union ib_gid *) &sib->sib_addr);
665 ib_addr_set_pkey(dev_addr, ntohs(sib->sib_pkey));
666}
667
668static int cma_translate_addr(struct sockaddr *addr, struct rdma_dev_addr *dev_addr)
669{
670 int ret;
671
672 if (addr->sa_family != AF_IB) {
673 ret = rdma_translate_ip(addr, dev_addr);
674 } else {
675 cma_translate_ib((struct sockaddr_ib *) addr, dev_addr);
676 ret = 0;
677 }
678
679 return ret;
680}
681
682static const struct ib_gid_attr *
683cma_validate_port(struct ib_device *device, u32 port,
684 enum ib_gid_type gid_type,
685 union ib_gid *gid,
686 struct rdma_id_private *id_priv)
687{
688 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
689 const struct ib_gid_attr *sgid_attr = ERR_PTR(-ENODEV);
690 int bound_if_index = dev_addr->bound_dev_if;
691 int dev_type = dev_addr->dev_type;
692 struct net_device *ndev = NULL;
693
694 if (!rdma_dev_access_netns(device, id_priv->id.route.addr.dev_addr.net))
695 goto out;
696
697 if ((dev_type == ARPHRD_INFINIBAND) && !rdma_protocol_ib(device, port))
698 goto out;
699
700 if ((dev_type != ARPHRD_INFINIBAND) && rdma_protocol_ib(device, port))
701 goto out;
702
703 /*
704 * For drivers that do not associate more than one net device with
705 * their gid tables, such as iWARP drivers, it is sufficient to
706 * return the first table entry.
707 *
708 * Other driver classes might be included in the future.
709 */
710 if (rdma_protocol_iwarp(device, port)) {
711 sgid_attr = rdma_get_gid_attr(device, port, 0);
712 if (IS_ERR(sgid_attr))
713 goto out;
714
715 rcu_read_lock();
716 ndev = rcu_dereference(sgid_attr->ndev);
717 if (!net_eq(dev_net(ndev), dev_addr->net) ||
718 ndev->ifindex != bound_if_index)
719 sgid_attr = ERR_PTR(-ENODEV);
720 rcu_read_unlock();
721 goto out;
722 }
723
724 if (dev_type == ARPHRD_ETHER && rdma_protocol_roce(device, port)) {
725 ndev = dev_get_by_index(dev_addr->net, bound_if_index);
726 if (!ndev)
727 goto out;
728 } else {
729 gid_type = IB_GID_TYPE_IB;
730 }
731
732 sgid_attr = rdma_find_gid_by_port(device, gid, gid_type, port, ndev);
733 dev_put(ndev);
734out:
735 return sgid_attr;
736}
737
738static void cma_bind_sgid_attr(struct rdma_id_private *id_priv,
739 const struct ib_gid_attr *sgid_attr)
740{
741 WARN_ON(id_priv->id.route.addr.dev_addr.sgid_attr);
742 id_priv->id.route.addr.dev_addr.sgid_attr = sgid_attr;
743}
744
745/**
746 * cma_acquire_dev_by_src_ip - Acquire cma device, port, gid attribute
747 * based on source ip address.
748 * @id_priv: cm_id which should be bound to cma device
749 *
750 * cma_acquire_dev_by_src_ip() binds cm id to cma device, port and GID attribute
751 * based on source IP address. It returns 0 on success or error code otherwise.
752 * It is applicable to active and passive side cm_id.
753 */
754static int cma_acquire_dev_by_src_ip(struct rdma_id_private *id_priv)
755{
756 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
757 const struct ib_gid_attr *sgid_attr;
758 union ib_gid gid, iboe_gid, *gidp;
759 struct cma_device *cma_dev;
760 enum ib_gid_type gid_type;
761 int ret = -ENODEV;
762 u32 port;
763
764 if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
765 id_priv->id.ps == RDMA_PS_IPOIB)
766 return -EINVAL;
767
768 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
769 &iboe_gid);
770
771 memcpy(&gid, dev_addr->src_dev_addr +
772 rdma_addr_gid_offset(dev_addr), sizeof(gid));
773
774 mutex_lock(&lock);
775 list_for_each_entry(cma_dev, &dev_list, list) {
776 rdma_for_each_port (cma_dev->device, port) {
777 gidp = rdma_protocol_roce(cma_dev->device, port) ?
778 &iboe_gid : &gid;
779 gid_type = cma_dev->default_gid_type[port - 1];
780 sgid_attr = cma_validate_port(cma_dev->device, port,
781 gid_type, gidp, id_priv);
782 if (!IS_ERR(sgid_attr)) {
783 id_priv->id.port_num = port;
784 cma_bind_sgid_attr(id_priv, sgid_attr);
785 cma_attach_to_dev(id_priv, cma_dev);
786 ret = 0;
787 goto out;
788 }
789 }
790 }
791out:
792 mutex_unlock(&lock);
793 return ret;
794}
795
796/**
797 * cma_ib_acquire_dev - Acquire cma device, port and SGID attribute
798 * @id_priv: cm id to bind to cma device
799 * @listen_id_priv: listener cm id to match against
800 * @req: Pointer to req structure containaining incoming
801 * request information
802 * cma_ib_acquire_dev() acquires cma device, port and SGID attribute when
803 * rdma device matches for listen_id and incoming request. It also verifies
804 * that a GID table entry is present for the source address.
805 * Returns 0 on success, or returns error code otherwise.
806 */
807static int cma_ib_acquire_dev(struct rdma_id_private *id_priv,
808 const struct rdma_id_private *listen_id_priv,
809 struct cma_req_info *req)
810{
811 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
812 const struct ib_gid_attr *sgid_attr;
813 enum ib_gid_type gid_type;
814 union ib_gid gid;
815
816 if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
817 id_priv->id.ps == RDMA_PS_IPOIB)
818 return -EINVAL;
819
820 if (rdma_protocol_roce(req->device, req->port))
821 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
822 &gid);
823 else
824 memcpy(&gid, dev_addr->src_dev_addr +
825 rdma_addr_gid_offset(dev_addr), sizeof(gid));
826
827 gid_type = listen_id_priv->cma_dev->default_gid_type[req->port - 1];
828 sgid_attr = cma_validate_port(req->device, req->port,
829 gid_type, &gid, id_priv);
830 if (IS_ERR(sgid_attr))
831 return PTR_ERR(sgid_attr);
832
833 id_priv->id.port_num = req->port;
834 cma_bind_sgid_attr(id_priv, sgid_attr);
835 /* Need to acquire lock to protect against reader
836 * of cma_dev->id_list such as cma_netdev_callback() and
837 * cma_process_remove().
838 */
839 mutex_lock(&lock);
840 cma_attach_to_dev(id_priv, listen_id_priv->cma_dev);
841 mutex_unlock(&lock);
842 rdma_restrack_add(&id_priv->res);
843 return 0;
844}
845
846static int cma_iw_acquire_dev(struct rdma_id_private *id_priv,
847 const struct rdma_id_private *listen_id_priv)
848{
849 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
850 const struct ib_gid_attr *sgid_attr;
851 struct cma_device *cma_dev;
852 enum ib_gid_type gid_type;
853 int ret = -ENODEV;
854 union ib_gid gid;
855 u32 port;
856
857 if (dev_addr->dev_type != ARPHRD_INFINIBAND &&
858 id_priv->id.ps == RDMA_PS_IPOIB)
859 return -EINVAL;
860
861 memcpy(&gid, dev_addr->src_dev_addr +
862 rdma_addr_gid_offset(dev_addr), sizeof(gid));
863
864 mutex_lock(&lock);
865
866 cma_dev = listen_id_priv->cma_dev;
867 port = listen_id_priv->id.port_num;
868 gid_type = listen_id_priv->gid_type;
869 sgid_attr = cma_validate_port(cma_dev->device, port,
870 gid_type, &gid, id_priv);
871 if (!IS_ERR(sgid_attr)) {
872 id_priv->id.port_num = port;
873 cma_bind_sgid_attr(id_priv, sgid_attr);
874 ret = 0;
875 goto out;
876 }
877
878 list_for_each_entry(cma_dev, &dev_list, list) {
879 rdma_for_each_port (cma_dev->device, port) {
880 if (listen_id_priv->cma_dev == cma_dev &&
881 listen_id_priv->id.port_num == port)
882 continue;
883
884 gid_type = cma_dev->default_gid_type[port - 1];
885 sgid_attr = cma_validate_port(cma_dev->device, port,
886 gid_type, &gid, id_priv);
887 if (!IS_ERR(sgid_attr)) {
888 id_priv->id.port_num = port;
889 cma_bind_sgid_attr(id_priv, sgid_attr);
890 ret = 0;
891 goto out;
892 }
893 }
894 }
895
896out:
897 if (!ret) {
898 cma_attach_to_dev(id_priv, cma_dev);
899 rdma_restrack_add(&id_priv->res);
900 }
901
902 mutex_unlock(&lock);
903 return ret;
904}
905
906/*
907 * Select the source IB device and address to reach the destination IB address.
908 */
909static int cma_resolve_ib_dev(struct rdma_id_private *id_priv)
910{
911 struct cma_device *cma_dev, *cur_dev;
912 struct sockaddr_ib *addr;
913 union ib_gid gid, sgid, *dgid;
914 unsigned int p;
915 u16 pkey, index;
916 enum ib_port_state port_state;
917 int ret;
918 int i;
919
920 cma_dev = NULL;
921 addr = (struct sockaddr_ib *) cma_dst_addr(id_priv);
922 dgid = (union ib_gid *) &addr->sib_addr;
923 pkey = ntohs(addr->sib_pkey);
924
925 mutex_lock(&lock);
926 list_for_each_entry(cur_dev, &dev_list, list) {
927 rdma_for_each_port (cur_dev->device, p) {
928 if (!rdma_cap_af_ib(cur_dev->device, p))
929 continue;
930
931 if (ib_find_cached_pkey(cur_dev->device, p, pkey, &index))
932 continue;
933
934 if (ib_get_cached_port_state(cur_dev->device, p, &port_state))
935 continue;
936
937 for (i = 0; i < cur_dev->device->port_data[p].immutable.gid_tbl_len;
938 ++i) {
939 ret = rdma_query_gid(cur_dev->device, p, i,
940 &gid);
941 if (ret)
942 continue;
943
944 if (!memcmp(&gid, dgid, sizeof(gid))) {
945 cma_dev = cur_dev;
946 sgid = gid;
947 id_priv->id.port_num = p;
948 goto found;
949 }
950
951 if (!cma_dev && (gid.global.subnet_prefix ==
952 dgid->global.subnet_prefix) &&
953 port_state == IB_PORT_ACTIVE) {
954 cma_dev = cur_dev;
955 sgid = gid;
956 id_priv->id.port_num = p;
957 goto found;
958 }
959 }
960 }
961 }
962 mutex_unlock(&lock);
963 return -ENODEV;
964
965found:
966 cma_attach_to_dev(id_priv, cma_dev);
967 rdma_restrack_add(&id_priv->res);
968 mutex_unlock(&lock);
969 addr = (struct sockaddr_ib *)cma_src_addr(id_priv);
970 memcpy(&addr->sib_addr, &sgid, sizeof(sgid));
971 cma_translate_ib(addr, &id_priv->id.route.addr.dev_addr);
972 return 0;
973}
974
975static void cma_id_get(struct rdma_id_private *id_priv)
976{
977 refcount_inc(&id_priv->refcount);
978}
979
980static void cma_id_put(struct rdma_id_private *id_priv)
981{
982 if (refcount_dec_and_test(&id_priv->refcount))
983 complete(&id_priv->comp);
984}
985
986static struct rdma_id_private *
987__rdma_create_id(struct net *net, rdma_cm_event_handler event_handler,
988 void *context, enum rdma_ucm_port_space ps,
989 enum ib_qp_type qp_type, const struct rdma_id_private *parent)
990{
991 struct rdma_id_private *id_priv;
992
993 id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
994 if (!id_priv)
995 return ERR_PTR(-ENOMEM);
996
997 id_priv->state = RDMA_CM_IDLE;
998 id_priv->id.context = context;
999 id_priv->id.event_handler = event_handler;
1000 id_priv->id.ps = ps;
1001 id_priv->id.qp_type = qp_type;
1002 id_priv->tos_set = false;
1003 id_priv->timeout_set = false;
1004 id_priv->min_rnr_timer_set = false;
1005 id_priv->gid_type = IB_GID_TYPE_IB;
1006 spin_lock_init(&id_priv->lock);
1007 mutex_init(&id_priv->qp_mutex);
1008 init_completion(&id_priv->comp);
1009 refcount_set(&id_priv->refcount, 1);
1010 mutex_init(&id_priv->handler_mutex);
1011 INIT_LIST_HEAD(&id_priv->device_item);
1012 INIT_LIST_HEAD(&id_priv->id_list_entry);
1013 INIT_LIST_HEAD(&id_priv->listen_list);
1014 INIT_LIST_HEAD(&id_priv->mc_list);
1015 get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
1016 id_priv->id.route.addr.dev_addr.net = get_net(net);
1017 id_priv->seq_num &= 0x00ffffff;
1018
1019 rdma_restrack_new(&id_priv->res, RDMA_RESTRACK_CM_ID);
1020 if (parent)
1021 rdma_restrack_parent_name(&id_priv->res, &parent->res);
1022
1023 return id_priv;
1024}
1025
1026struct rdma_cm_id *
1027__rdma_create_kernel_id(struct net *net, rdma_cm_event_handler event_handler,
1028 void *context, enum rdma_ucm_port_space ps,
1029 enum ib_qp_type qp_type, const char *caller)
1030{
1031 struct rdma_id_private *ret;
1032
1033 ret = __rdma_create_id(net, event_handler, context, ps, qp_type, NULL);
1034 if (IS_ERR(ret))
1035 return ERR_CAST(ret);
1036
1037 rdma_restrack_set_name(&ret->res, caller);
1038 return &ret->id;
1039}
1040EXPORT_SYMBOL(__rdma_create_kernel_id);
1041
1042struct rdma_cm_id *rdma_create_user_id(rdma_cm_event_handler event_handler,
1043 void *context,
1044 enum rdma_ucm_port_space ps,
1045 enum ib_qp_type qp_type)
1046{
1047 struct rdma_id_private *ret;
1048
1049 ret = __rdma_create_id(current->nsproxy->net_ns, event_handler, context,
1050 ps, qp_type, NULL);
1051 if (IS_ERR(ret))
1052 return ERR_CAST(ret);
1053
1054 rdma_restrack_set_name(&ret->res, NULL);
1055 return &ret->id;
1056}
1057EXPORT_SYMBOL(rdma_create_user_id);
1058
1059static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
1060{
1061 struct ib_qp_attr qp_attr;
1062 int qp_attr_mask, ret;
1063
1064 qp_attr.qp_state = IB_QPS_INIT;
1065 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1066 if (ret)
1067 return ret;
1068
1069 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
1070 if (ret)
1071 return ret;
1072
1073 qp_attr.qp_state = IB_QPS_RTR;
1074 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
1075 if (ret)
1076 return ret;
1077
1078 qp_attr.qp_state = IB_QPS_RTS;
1079 qp_attr.sq_psn = 0;
1080 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
1081
1082 return ret;
1083}
1084
1085static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
1086{
1087 struct ib_qp_attr qp_attr;
1088 int qp_attr_mask, ret;
1089
1090 qp_attr.qp_state = IB_QPS_INIT;
1091 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1092 if (ret)
1093 return ret;
1094
1095 return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
1096}
1097
1098int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
1099 struct ib_qp_init_attr *qp_init_attr)
1100{
1101 struct rdma_id_private *id_priv;
1102 struct ib_qp *qp;
1103 int ret;
1104
1105 id_priv = container_of(id, struct rdma_id_private, id);
1106 if (id->device != pd->device) {
1107 ret = -EINVAL;
1108 goto out_err;
1109 }
1110
1111 qp_init_attr->port_num = id->port_num;
1112 qp = ib_create_qp(pd, qp_init_attr);
1113 if (IS_ERR(qp)) {
1114 ret = PTR_ERR(qp);
1115 goto out_err;
1116 }
1117
1118 if (id->qp_type == IB_QPT_UD)
1119 ret = cma_init_ud_qp(id_priv, qp);
1120 else
1121 ret = cma_init_conn_qp(id_priv, qp);
1122 if (ret)
1123 goto out_destroy;
1124
1125 id->qp = qp;
1126 id_priv->qp_num = qp->qp_num;
1127 id_priv->srq = (qp->srq != NULL);
1128 trace_cm_qp_create(id_priv, pd, qp_init_attr, 0);
1129 return 0;
1130out_destroy:
1131 ib_destroy_qp(qp);
1132out_err:
1133 trace_cm_qp_create(id_priv, pd, qp_init_attr, ret);
1134 return ret;
1135}
1136EXPORT_SYMBOL(rdma_create_qp);
1137
1138void rdma_destroy_qp(struct rdma_cm_id *id)
1139{
1140 struct rdma_id_private *id_priv;
1141
1142 id_priv = container_of(id, struct rdma_id_private, id);
1143 trace_cm_qp_destroy(id_priv);
1144 mutex_lock(&id_priv->qp_mutex);
1145 ib_destroy_qp(id_priv->id.qp);
1146 id_priv->id.qp = NULL;
1147 mutex_unlock(&id_priv->qp_mutex);
1148}
1149EXPORT_SYMBOL(rdma_destroy_qp);
1150
1151static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
1152 struct rdma_conn_param *conn_param)
1153{
1154 struct ib_qp_attr qp_attr;
1155 int qp_attr_mask, ret;
1156
1157 mutex_lock(&id_priv->qp_mutex);
1158 if (!id_priv->id.qp) {
1159 ret = 0;
1160 goto out;
1161 }
1162
1163 /* Need to update QP attributes from default values. */
1164 qp_attr.qp_state = IB_QPS_INIT;
1165 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1166 if (ret)
1167 goto out;
1168
1169 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
1170 if (ret)
1171 goto out;
1172
1173 qp_attr.qp_state = IB_QPS_RTR;
1174 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1175 if (ret)
1176 goto out;
1177
1178 BUG_ON(id_priv->cma_dev->device != id_priv->id.device);
1179
1180 if (conn_param)
1181 qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
1182 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
1183out:
1184 mutex_unlock(&id_priv->qp_mutex);
1185 return ret;
1186}
1187
1188static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
1189 struct rdma_conn_param *conn_param)
1190{
1191 struct ib_qp_attr qp_attr;
1192 int qp_attr_mask, ret;
1193
1194 mutex_lock(&id_priv->qp_mutex);
1195 if (!id_priv->id.qp) {
1196 ret = 0;
1197 goto out;
1198 }
1199
1200 qp_attr.qp_state = IB_QPS_RTS;
1201 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
1202 if (ret)
1203 goto out;
1204
1205 if (conn_param)
1206 qp_attr.max_rd_atomic = conn_param->initiator_depth;
1207 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
1208out:
1209 mutex_unlock(&id_priv->qp_mutex);
1210 return ret;
1211}
1212
1213static int cma_modify_qp_err(struct rdma_id_private *id_priv)
1214{
1215 struct ib_qp_attr qp_attr;
1216 int ret;
1217
1218 mutex_lock(&id_priv->qp_mutex);
1219 if (!id_priv->id.qp) {
1220 ret = 0;
1221 goto out;
1222 }
1223
1224 qp_attr.qp_state = IB_QPS_ERR;
1225 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
1226out:
1227 mutex_unlock(&id_priv->qp_mutex);
1228 return ret;
1229}
1230
1231static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
1232 struct ib_qp_attr *qp_attr, int *qp_attr_mask)
1233{
1234 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
1235 int ret;
1236 u16 pkey;
1237
1238 if (rdma_cap_eth_ah(id_priv->id.device, id_priv->id.port_num))
1239 pkey = 0xffff;
1240 else
1241 pkey = ib_addr_get_pkey(dev_addr);
1242
1243 ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
1244 pkey, &qp_attr->pkey_index);
1245 if (ret)
1246 return ret;
1247
1248 qp_attr->port_num = id_priv->id.port_num;
1249 *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
1250
1251 if (id_priv->id.qp_type == IB_QPT_UD) {
1252 ret = cma_set_default_qkey(id_priv);
1253 if (ret)
1254 return ret;
1255
1256 qp_attr->qkey = id_priv->qkey;
1257 *qp_attr_mask |= IB_QP_QKEY;
1258 } else {
1259 qp_attr->qp_access_flags = 0;
1260 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
1261 }
1262 return 0;
1263}
1264
1265int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
1266 int *qp_attr_mask)
1267{
1268 struct rdma_id_private *id_priv;
1269 int ret = 0;
1270
1271 id_priv = container_of(id, struct rdma_id_private, id);
1272 if (rdma_cap_ib_cm(id->device, id->port_num)) {
1273 if (!id_priv->cm_id.ib || (id_priv->id.qp_type == IB_QPT_UD))
1274 ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
1275 else
1276 ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
1277 qp_attr_mask);
1278
1279 if (qp_attr->qp_state == IB_QPS_RTR)
1280 qp_attr->rq_psn = id_priv->seq_num;
1281 } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
1282 if (!id_priv->cm_id.iw) {
1283 qp_attr->qp_access_flags = 0;
1284 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
1285 } else
1286 ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
1287 qp_attr_mask);
1288 qp_attr->port_num = id_priv->id.port_num;
1289 *qp_attr_mask |= IB_QP_PORT;
1290 } else {
1291 ret = -ENOSYS;
1292 }
1293
1294 if ((*qp_attr_mask & IB_QP_TIMEOUT) && id_priv->timeout_set)
1295 qp_attr->timeout = id_priv->timeout;
1296
1297 if ((*qp_attr_mask & IB_QP_MIN_RNR_TIMER) && id_priv->min_rnr_timer_set)
1298 qp_attr->min_rnr_timer = id_priv->min_rnr_timer;
1299
1300 return ret;
1301}
1302EXPORT_SYMBOL(rdma_init_qp_attr);
1303
1304static inline bool cma_zero_addr(const struct sockaddr *addr)
1305{
1306 switch (addr->sa_family) {
1307 case AF_INET:
1308 return ipv4_is_zeronet(((struct sockaddr_in *)addr)->sin_addr.s_addr);
1309 case AF_INET6:
1310 return ipv6_addr_any(&((struct sockaddr_in6 *)addr)->sin6_addr);
1311 case AF_IB:
1312 return ib_addr_any(&((struct sockaddr_ib *)addr)->sib_addr);
1313 default:
1314 return false;
1315 }
1316}
1317
1318static inline bool cma_loopback_addr(const struct sockaddr *addr)
1319{
1320 switch (addr->sa_family) {
1321 case AF_INET:
1322 return ipv4_is_loopback(
1323 ((struct sockaddr_in *)addr)->sin_addr.s_addr);
1324 case AF_INET6:
1325 return ipv6_addr_loopback(
1326 &((struct sockaddr_in6 *)addr)->sin6_addr);
1327 case AF_IB:
1328 return ib_addr_loopback(
1329 &((struct sockaddr_ib *)addr)->sib_addr);
1330 default:
1331 return false;
1332 }
1333}
1334
1335static inline bool cma_any_addr(const struct sockaddr *addr)
1336{
1337 return cma_zero_addr(addr) || cma_loopback_addr(addr);
1338}
1339
1340static int cma_addr_cmp(const struct sockaddr *src, const struct sockaddr *dst)
1341{
1342 if (src->sa_family != dst->sa_family)
1343 return -1;
1344
1345 switch (src->sa_family) {
1346 case AF_INET:
1347 return ((struct sockaddr_in *)src)->sin_addr.s_addr !=
1348 ((struct sockaddr_in *)dst)->sin_addr.s_addr;
1349 case AF_INET6: {
1350 struct sockaddr_in6 *src_addr6 = (struct sockaddr_in6 *)src;
1351 struct sockaddr_in6 *dst_addr6 = (struct sockaddr_in6 *)dst;
1352 bool link_local;
1353
1354 if (ipv6_addr_cmp(&src_addr6->sin6_addr,
1355 &dst_addr6->sin6_addr))
1356 return 1;
1357 link_local = ipv6_addr_type(&dst_addr6->sin6_addr) &
1358 IPV6_ADDR_LINKLOCAL;
1359 /* Link local must match their scope_ids */
1360 return link_local ? (src_addr6->sin6_scope_id !=
1361 dst_addr6->sin6_scope_id) :
1362 0;
1363 }
1364
1365 default:
1366 return ib_addr_cmp(&((struct sockaddr_ib *) src)->sib_addr,
1367 &((struct sockaddr_ib *) dst)->sib_addr);
1368 }
1369}
1370
1371static __be16 cma_port(const struct sockaddr *addr)
1372{
1373 struct sockaddr_ib *sib;
1374
1375 switch (addr->sa_family) {
1376 case AF_INET:
1377 return ((struct sockaddr_in *) addr)->sin_port;
1378 case AF_INET6:
1379 return ((struct sockaddr_in6 *) addr)->sin6_port;
1380 case AF_IB:
1381 sib = (struct sockaddr_ib *) addr;
1382 return htons((u16) (be64_to_cpu(sib->sib_sid) &
1383 be64_to_cpu(sib->sib_sid_mask)));
1384 default:
1385 return 0;
1386 }
1387}
1388
1389static inline int cma_any_port(const struct sockaddr *addr)
1390{
1391 return !cma_port(addr);
1392}
1393
1394static void cma_save_ib_info(struct sockaddr *src_addr,
1395 struct sockaddr *dst_addr,
1396 const struct rdma_cm_id *listen_id,
1397 const struct sa_path_rec *path)
1398{
1399 struct sockaddr_ib *listen_ib, *ib;
1400
1401 listen_ib = (struct sockaddr_ib *) &listen_id->route.addr.src_addr;
1402 if (src_addr) {
1403 ib = (struct sockaddr_ib *)src_addr;
1404 ib->sib_family = AF_IB;
1405 if (path) {
1406 ib->sib_pkey = path->pkey;
1407 ib->sib_flowinfo = path->flow_label;
1408 memcpy(&ib->sib_addr, &path->sgid, 16);
1409 ib->sib_sid = path->service_id;
1410 ib->sib_scope_id = 0;
1411 } else {
1412 ib->sib_pkey = listen_ib->sib_pkey;
1413 ib->sib_flowinfo = listen_ib->sib_flowinfo;
1414 ib->sib_addr = listen_ib->sib_addr;
1415 ib->sib_sid = listen_ib->sib_sid;
1416 ib->sib_scope_id = listen_ib->sib_scope_id;
1417 }
1418 ib->sib_sid_mask = cpu_to_be64(0xffffffffffffffffULL);
1419 }
1420 if (dst_addr) {
1421 ib = (struct sockaddr_ib *)dst_addr;
1422 ib->sib_family = AF_IB;
1423 if (path) {
1424 ib->sib_pkey = path->pkey;
1425 ib->sib_flowinfo = path->flow_label;
1426 memcpy(&ib->sib_addr, &path->dgid, 16);
1427 }
1428 }
1429}
1430
1431static void cma_save_ip4_info(struct sockaddr_in *src_addr,
1432 struct sockaddr_in *dst_addr,
1433 struct cma_hdr *hdr,
1434 __be16 local_port)
1435{
1436 if (src_addr) {
1437 *src_addr = (struct sockaddr_in) {
1438 .sin_family = AF_INET,
1439 .sin_addr.s_addr = hdr->dst_addr.ip4.addr,
1440 .sin_port = local_port,
1441 };
1442 }
1443
1444 if (dst_addr) {
1445 *dst_addr = (struct sockaddr_in) {
1446 .sin_family = AF_INET,
1447 .sin_addr.s_addr = hdr->src_addr.ip4.addr,
1448 .sin_port = hdr->port,
1449 };
1450 }
1451}
1452
1453static void cma_save_ip6_info(struct sockaddr_in6 *src_addr,
1454 struct sockaddr_in6 *dst_addr,
1455 struct cma_hdr *hdr,
1456 __be16 local_port)
1457{
1458 if (src_addr) {
1459 *src_addr = (struct sockaddr_in6) {
1460 .sin6_family = AF_INET6,
1461 .sin6_addr = hdr->dst_addr.ip6,
1462 .sin6_port = local_port,
1463 };
1464 }
1465
1466 if (dst_addr) {
1467 *dst_addr = (struct sockaddr_in6) {
1468 .sin6_family = AF_INET6,
1469 .sin6_addr = hdr->src_addr.ip6,
1470 .sin6_port = hdr->port,
1471 };
1472 }
1473}
1474
1475static u16 cma_port_from_service_id(__be64 service_id)
1476{
1477 return (u16)be64_to_cpu(service_id);
1478}
1479
1480static int cma_save_ip_info(struct sockaddr *src_addr,
1481 struct sockaddr *dst_addr,
1482 const struct ib_cm_event *ib_event,
1483 __be64 service_id)
1484{
1485 struct cma_hdr *hdr;
1486 __be16 port;
1487
1488 hdr = ib_event->private_data;
1489 if (hdr->cma_version != CMA_VERSION)
1490 return -EINVAL;
1491
1492 port = htons(cma_port_from_service_id(service_id));
1493
1494 switch (cma_get_ip_ver(hdr)) {
1495 case 4:
1496 cma_save_ip4_info((struct sockaddr_in *)src_addr,
1497 (struct sockaddr_in *)dst_addr, hdr, port);
1498 break;
1499 case 6:
1500 cma_save_ip6_info((struct sockaddr_in6 *)src_addr,
1501 (struct sockaddr_in6 *)dst_addr, hdr, port);
1502 break;
1503 default:
1504 return -EAFNOSUPPORT;
1505 }
1506
1507 return 0;
1508}
1509
1510static int cma_save_net_info(struct sockaddr *src_addr,
1511 struct sockaddr *dst_addr,
1512 const struct rdma_cm_id *listen_id,
1513 const struct ib_cm_event *ib_event,
1514 sa_family_t sa_family, __be64 service_id)
1515{
1516 if (sa_family == AF_IB) {
1517 if (ib_event->event == IB_CM_REQ_RECEIVED)
1518 cma_save_ib_info(src_addr, dst_addr, listen_id,
1519 ib_event->param.req_rcvd.primary_path);
1520 else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED)
1521 cma_save_ib_info(src_addr, dst_addr, listen_id, NULL);
1522 return 0;
1523 }
1524
1525 return cma_save_ip_info(src_addr, dst_addr, ib_event, service_id);
1526}
1527
1528static int cma_save_req_info(const struct ib_cm_event *ib_event,
1529 struct cma_req_info *req)
1530{
1531 const struct ib_cm_req_event_param *req_param =
1532 &ib_event->param.req_rcvd;
1533 const struct ib_cm_sidr_req_event_param *sidr_param =
1534 &ib_event->param.sidr_req_rcvd;
1535
1536 switch (ib_event->event) {
1537 case IB_CM_REQ_RECEIVED:
1538 req->device = req_param->listen_id->device;
1539 req->port = req_param->port;
1540 memcpy(&req->local_gid, &req_param->primary_path->sgid,
1541 sizeof(req->local_gid));
1542 req->has_gid = true;
1543 req->service_id = req_param->primary_path->service_id;
1544 req->pkey = be16_to_cpu(req_param->primary_path->pkey);
1545 if (req->pkey != req_param->bth_pkey)
1546 pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and primary path P_Key (0x%x)\n"
1547 "RDMA CMA: in the future this may cause the request to be dropped\n",
1548 req_param->bth_pkey, req->pkey);
1549 break;
1550 case IB_CM_SIDR_REQ_RECEIVED:
1551 req->device = sidr_param->listen_id->device;
1552 req->port = sidr_param->port;
1553 req->has_gid = false;
1554 req->service_id = sidr_param->service_id;
1555 req->pkey = sidr_param->pkey;
1556 if (req->pkey != sidr_param->bth_pkey)
1557 pr_warn_ratelimited("RDMA CMA: got different BTH P_Key (0x%x) and SIDR request payload P_Key (0x%x)\n"
1558 "RDMA CMA: in the future this may cause the request to be dropped\n",
1559 sidr_param->bth_pkey, req->pkey);
1560 break;
1561 default:
1562 return -EINVAL;
1563 }
1564
1565 return 0;
1566}
1567
1568static bool validate_ipv4_net_dev(struct net_device *net_dev,
1569 const struct sockaddr_in *dst_addr,
1570 const struct sockaddr_in *src_addr)
1571{
1572 __be32 daddr = dst_addr->sin_addr.s_addr,
1573 saddr = src_addr->sin_addr.s_addr;
1574 struct fib_result res;
1575 struct flowi4 fl4;
1576 int err;
1577 bool ret;
1578
1579 if (ipv4_is_multicast(saddr) || ipv4_is_lbcast(saddr) ||
1580 ipv4_is_lbcast(daddr) || ipv4_is_zeronet(saddr) ||
1581 ipv4_is_zeronet(daddr) || ipv4_is_loopback(daddr) ||
1582 ipv4_is_loopback(saddr))
1583 return false;
1584
1585 memset(&fl4, 0, sizeof(fl4));
1586 fl4.flowi4_oif = net_dev->ifindex;
1587 fl4.daddr = daddr;
1588 fl4.saddr = saddr;
1589
1590 rcu_read_lock();
1591 err = fib_lookup(dev_net(net_dev), &fl4, &res, 0);
1592 ret = err == 0 && FIB_RES_DEV(res) == net_dev;
1593 rcu_read_unlock();
1594
1595 return ret;
1596}
1597
1598static bool validate_ipv6_net_dev(struct net_device *net_dev,
1599 const struct sockaddr_in6 *dst_addr,
1600 const struct sockaddr_in6 *src_addr)
1601{
1602#if IS_ENABLED(CONFIG_IPV6)
1603 const int strict = ipv6_addr_type(&dst_addr->sin6_addr) &
1604 IPV6_ADDR_LINKLOCAL;
1605 struct rt6_info *rt = rt6_lookup(dev_net(net_dev), &dst_addr->sin6_addr,
1606 &src_addr->sin6_addr, net_dev->ifindex,
1607 NULL, strict);
1608 bool ret;
1609
1610 if (!rt)
1611 return false;
1612
1613 ret = rt->rt6i_idev->dev == net_dev;
1614 ip6_rt_put(rt);
1615
1616 return ret;
1617#else
1618 return false;
1619#endif
1620}
1621
1622static bool validate_net_dev(struct net_device *net_dev,
1623 const struct sockaddr *daddr,
1624 const struct sockaddr *saddr)
1625{
1626 const struct sockaddr_in *daddr4 = (const struct sockaddr_in *)daddr;
1627 const struct sockaddr_in *saddr4 = (const struct sockaddr_in *)saddr;
1628 const struct sockaddr_in6 *daddr6 = (const struct sockaddr_in6 *)daddr;
1629 const struct sockaddr_in6 *saddr6 = (const struct sockaddr_in6 *)saddr;
1630
1631 switch (daddr->sa_family) {
1632 case AF_INET:
1633 return saddr->sa_family == AF_INET &&
1634 validate_ipv4_net_dev(net_dev, daddr4, saddr4);
1635
1636 case AF_INET6:
1637 return saddr->sa_family == AF_INET6 &&
1638 validate_ipv6_net_dev(net_dev, daddr6, saddr6);
1639
1640 default:
1641 return false;
1642 }
1643}
1644
1645static struct net_device *
1646roce_get_net_dev_by_cm_event(const struct ib_cm_event *ib_event)
1647{
1648 const struct ib_gid_attr *sgid_attr = NULL;
1649 struct net_device *ndev;
1650
1651 if (ib_event->event == IB_CM_REQ_RECEIVED)
1652 sgid_attr = ib_event->param.req_rcvd.ppath_sgid_attr;
1653 else if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED)
1654 sgid_attr = ib_event->param.sidr_req_rcvd.sgid_attr;
1655
1656 if (!sgid_attr)
1657 return NULL;
1658
1659 rcu_read_lock();
1660 ndev = rdma_read_gid_attr_ndev_rcu(sgid_attr);
1661 if (IS_ERR(ndev))
1662 ndev = NULL;
1663 else
1664 dev_hold(ndev);
1665 rcu_read_unlock();
1666 return ndev;
1667}
1668
1669static struct net_device *cma_get_net_dev(const struct ib_cm_event *ib_event,
1670 struct cma_req_info *req)
1671{
1672 struct sockaddr *listen_addr =
1673 (struct sockaddr *)&req->listen_addr_storage;
1674 struct sockaddr *src_addr = (struct sockaddr *)&req->src_addr_storage;
1675 struct net_device *net_dev;
1676 const union ib_gid *gid = req->has_gid ? &req->local_gid : NULL;
1677 int err;
1678
1679 err = cma_save_ip_info(listen_addr, src_addr, ib_event,
1680 req->service_id);
1681 if (err)
1682 return ERR_PTR(err);
1683
1684 if (rdma_protocol_roce(req->device, req->port))
1685 net_dev = roce_get_net_dev_by_cm_event(ib_event);
1686 else
1687 net_dev = ib_get_net_dev_by_params(req->device, req->port,
1688 req->pkey,
1689 gid, listen_addr);
1690 if (!net_dev)
1691 return ERR_PTR(-ENODEV);
1692
1693 return net_dev;
1694}
1695
1696static enum rdma_ucm_port_space rdma_ps_from_service_id(__be64 service_id)
1697{
1698 return (be64_to_cpu(service_id) >> 16) & 0xffff;
1699}
1700
1701static bool cma_match_private_data(struct rdma_id_private *id_priv,
1702 const struct cma_hdr *hdr)
1703{
1704 struct sockaddr *addr = cma_src_addr(id_priv);
1705 __be32 ip4_addr;
1706 struct in6_addr ip6_addr;
1707
1708 if (cma_any_addr(addr) && !id_priv->afonly)
1709 return true;
1710
1711 switch (addr->sa_family) {
1712 case AF_INET:
1713 ip4_addr = ((struct sockaddr_in *)addr)->sin_addr.s_addr;
1714 if (cma_get_ip_ver(hdr) != 4)
1715 return false;
1716 if (!cma_any_addr(addr) &&
1717 hdr->dst_addr.ip4.addr != ip4_addr)
1718 return false;
1719 break;
1720 case AF_INET6:
1721 ip6_addr = ((struct sockaddr_in6 *)addr)->sin6_addr;
1722 if (cma_get_ip_ver(hdr) != 6)
1723 return false;
1724 if (!cma_any_addr(addr) &&
1725 memcmp(&hdr->dst_addr.ip6, &ip6_addr, sizeof(ip6_addr)))
1726 return false;
1727 break;
1728 case AF_IB:
1729 return true;
1730 default:
1731 return false;
1732 }
1733
1734 return true;
1735}
1736
1737static bool cma_protocol_roce(const struct rdma_cm_id *id)
1738{
1739 struct ib_device *device = id->device;
1740 const u32 port_num = id->port_num ?: rdma_start_port(device);
1741
1742 return rdma_protocol_roce(device, port_num);
1743}
1744
1745static bool cma_is_req_ipv6_ll(const struct cma_req_info *req)
1746{
1747 const struct sockaddr *daddr =
1748 (const struct sockaddr *)&req->listen_addr_storage;
1749 const struct sockaddr_in6 *daddr6 = (const struct sockaddr_in6 *)daddr;
1750
1751 /* Returns true if the req is for IPv6 link local */
1752 return (daddr->sa_family == AF_INET6 &&
1753 (ipv6_addr_type(&daddr6->sin6_addr) & IPV6_ADDR_LINKLOCAL));
1754}
1755
1756static bool cma_match_net_dev(const struct rdma_cm_id *id,
1757 const struct net_device *net_dev,
1758 const struct cma_req_info *req)
1759{
1760 const struct rdma_addr *addr = &id->route.addr;
1761
1762 if (!net_dev)
1763 /* This request is an AF_IB request */
1764 return (!id->port_num || id->port_num == req->port) &&
1765 (addr->src_addr.ss_family == AF_IB);
1766
1767 /*
1768 * If the request is not for IPv6 link local, allow matching
1769 * request to any netdevice of the one or multiport rdma device.
1770 */
1771 if (!cma_is_req_ipv6_ll(req))
1772 return true;
1773 /*
1774 * Net namespaces must match, and if the listner is listening
1775 * on a specific netdevice than netdevice must match as well.
1776 */
1777 if (net_eq(dev_net(net_dev), addr->dev_addr.net) &&
1778 (!!addr->dev_addr.bound_dev_if ==
1779 (addr->dev_addr.bound_dev_if == net_dev->ifindex)))
1780 return true;
1781 else
1782 return false;
1783}
1784
1785static struct rdma_id_private *cma_find_listener(
1786 const struct rdma_bind_list *bind_list,
1787 const struct ib_cm_id *cm_id,
1788 const struct ib_cm_event *ib_event,
1789 const struct cma_req_info *req,
1790 const struct net_device *net_dev)
1791{
1792 struct rdma_id_private *id_priv, *id_priv_dev;
1793
1794 lockdep_assert_held(&lock);
1795
1796 if (!bind_list)
1797 return ERR_PTR(-EINVAL);
1798
1799 hlist_for_each_entry(id_priv, &bind_list->owners, node) {
1800 if (cma_match_private_data(id_priv, ib_event->private_data)) {
1801 if (id_priv->id.device == cm_id->device &&
1802 cma_match_net_dev(&id_priv->id, net_dev, req))
1803 return id_priv;
1804 list_for_each_entry(id_priv_dev,
1805 &id_priv->listen_list,
1806 listen_item) {
1807 if (id_priv_dev->id.device == cm_id->device &&
1808 cma_match_net_dev(&id_priv_dev->id,
1809 net_dev, req))
1810 return id_priv_dev;
1811 }
1812 }
1813 }
1814
1815 return ERR_PTR(-EINVAL);
1816}
1817
1818static struct rdma_id_private *
1819cma_ib_id_from_event(struct ib_cm_id *cm_id,
1820 const struct ib_cm_event *ib_event,
1821 struct cma_req_info *req,
1822 struct net_device **net_dev)
1823{
1824 struct rdma_bind_list *bind_list;
1825 struct rdma_id_private *id_priv;
1826 int err;
1827
1828 err = cma_save_req_info(ib_event, req);
1829 if (err)
1830 return ERR_PTR(err);
1831
1832 *net_dev = cma_get_net_dev(ib_event, req);
1833 if (IS_ERR(*net_dev)) {
1834 if (PTR_ERR(*net_dev) == -EAFNOSUPPORT) {
1835 /* Assuming the protocol is AF_IB */
1836 *net_dev = NULL;
1837 } else {
1838 return ERR_CAST(*net_dev);
1839 }
1840 }
1841
1842 mutex_lock(&lock);
1843 /*
1844 * Net namespace might be getting deleted while route lookup,
1845 * cm_id lookup is in progress. Therefore, perform netdevice
1846 * validation, cm_id lookup under rcu lock.
1847 * RCU lock along with netdevice state check, synchronizes with
1848 * netdevice migrating to different net namespace and also avoids
1849 * case where net namespace doesn't get deleted while lookup is in
1850 * progress.
1851 * If the device state is not IFF_UP, its properties such as ifindex
1852 * and nd_net cannot be trusted to remain valid without rcu lock.
1853 * net/core/dev.c change_net_namespace() ensures to synchronize with
1854 * ongoing operations on net device after device is closed using
1855 * synchronize_net().
1856 */
1857 rcu_read_lock();
1858 if (*net_dev) {
1859 /*
1860 * If netdevice is down, it is likely that it is administratively
1861 * down or it might be migrating to different namespace.
1862 * In that case avoid further processing, as the net namespace
1863 * or ifindex may change.
1864 */
1865 if (((*net_dev)->flags & IFF_UP) == 0) {
1866 id_priv = ERR_PTR(-EHOSTUNREACH);
1867 goto err;
1868 }
1869
1870 if (!validate_net_dev(*net_dev,
1871 (struct sockaddr *)&req->src_addr_storage,
1872 (struct sockaddr *)&req->listen_addr_storage)) {
1873 id_priv = ERR_PTR(-EHOSTUNREACH);
1874 goto err;
1875 }
1876 }
1877
1878 bind_list = cma_ps_find(*net_dev ? dev_net(*net_dev) : &init_net,
1879 rdma_ps_from_service_id(req->service_id),
1880 cma_port_from_service_id(req->service_id));
1881 id_priv = cma_find_listener(bind_list, cm_id, ib_event, req, *net_dev);
1882err:
1883 rcu_read_unlock();
1884 mutex_unlock(&lock);
1885 if (IS_ERR(id_priv) && *net_dev) {
1886 dev_put(*net_dev);
1887 *net_dev = NULL;
1888 }
1889 return id_priv;
1890}
1891
1892static inline u8 cma_user_data_offset(struct rdma_id_private *id_priv)
1893{
1894 return cma_family(id_priv) == AF_IB ? 0 : sizeof(struct cma_hdr);
1895}
1896
1897static void cma_cancel_route(struct rdma_id_private *id_priv)
1898{
1899 if (rdma_cap_ib_sa(id_priv->id.device, id_priv->id.port_num)) {
1900 if (id_priv->query)
1901 ib_sa_cancel_query(id_priv->query_id, id_priv->query);
1902 }
1903}
1904
1905static void _cma_cancel_listens(struct rdma_id_private *id_priv)
1906{
1907 struct rdma_id_private *dev_id_priv;
1908
1909 lockdep_assert_held(&lock);
1910
1911 /*
1912 * Remove from listen_any_list to prevent added devices from spawning
1913 * additional listen requests.
1914 */
1915 list_del_init(&id_priv->listen_any_item);
1916
1917 while (!list_empty(&id_priv->listen_list)) {
1918 dev_id_priv =
1919 list_first_entry(&id_priv->listen_list,
1920 struct rdma_id_private, listen_item);
1921 /* sync with device removal to avoid duplicate destruction */
1922 list_del_init(&dev_id_priv->device_item);
1923 list_del_init(&dev_id_priv->listen_item);
1924 mutex_unlock(&lock);
1925
1926 rdma_destroy_id(&dev_id_priv->id);
1927 mutex_lock(&lock);
1928 }
1929}
1930
1931static void cma_cancel_listens(struct rdma_id_private *id_priv)
1932{
1933 mutex_lock(&lock);
1934 _cma_cancel_listens(id_priv);
1935 mutex_unlock(&lock);
1936}
1937
1938static void cma_cancel_operation(struct rdma_id_private *id_priv,
1939 enum rdma_cm_state state)
1940{
1941 switch (state) {
1942 case RDMA_CM_ADDR_QUERY:
1943 /*
1944 * We can avoid doing the rdma_addr_cancel() based on state,
1945 * only RDMA_CM_ADDR_QUERY has a work that could still execute.
1946 * Notice that the addr_handler work could still be exiting
1947 * outside this state, however due to the interaction with the
1948 * handler_mutex the work is guaranteed not to touch id_priv
1949 * during exit.
1950 */
1951 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
1952 break;
1953 case RDMA_CM_ROUTE_QUERY:
1954 cma_cancel_route(id_priv);
1955 break;
1956 case RDMA_CM_LISTEN:
1957 if (cma_any_addr(cma_src_addr(id_priv)) && !id_priv->cma_dev)
1958 cma_cancel_listens(id_priv);
1959 break;
1960 default:
1961 break;
1962 }
1963}
1964
1965static void cma_release_port(struct rdma_id_private *id_priv)
1966{
1967 struct rdma_bind_list *bind_list = id_priv->bind_list;
1968 struct net *net = id_priv->id.route.addr.dev_addr.net;
1969
1970 if (!bind_list)
1971 return;
1972
1973 mutex_lock(&lock);
1974 hlist_del(&id_priv->node);
1975 if (hlist_empty(&bind_list->owners)) {
1976 cma_ps_remove(net, bind_list->ps, bind_list->port);
1977 kfree(bind_list);
1978 }
1979 mutex_unlock(&lock);
1980}
1981
1982static void destroy_mc(struct rdma_id_private *id_priv,
1983 struct cma_multicast *mc)
1984{
1985 bool send_only = mc->join_state == BIT(SENDONLY_FULLMEMBER_JOIN);
1986
1987 if (rdma_cap_ib_mcast(id_priv->id.device, id_priv->id.port_num))
1988 ib_sa_free_multicast(mc->sa_mc);
1989
1990 if (rdma_protocol_roce(id_priv->id.device, id_priv->id.port_num)) {
1991 struct rdma_dev_addr *dev_addr =
1992 &id_priv->id.route.addr.dev_addr;
1993 struct net_device *ndev = NULL;
1994
1995 if (dev_addr->bound_dev_if)
1996 ndev = dev_get_by_index(dev_addr->net,
1997 dev_addr->bound_dev_if);
1998 if (ndev && !send_only) {
1999 enum ib_gid_type gid_type;
2000 union ib_gid mgid;
2001
2002 gid_type = id_priv->cma_dev->default_gid_type
2003 [id_priv->id.port_num -
2004 rdma_start_port(
2005 id_priv->cma_dev->device)];
2006 cma_iboe_set_mgid((struct sockaddr *)&mc->addr, &mgid,
2007 gid_type);
2008 cma_igmp_send(ndev, &mgid, false);
2009 }
2010 dev_put(ndev);
2011
2012 cancel_work_sync(&mc->iboe_join.work);
2013 }
2014 kfree(mc);
2015}
2016
2017static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
2018{
2019 struct cma_multicast *mc;
2020
2021 while (!list_empty(&id_priv->mc_list)) {
2022 mc = list_first_entry(&id_priv->mc_list, struct cma_multicast,
2023 list);
2024 list_del(&mc->list);
2025 destroy_mc(id_priv, mc);
2026 }
2027}
2028
2029static void _destroy_id(struct rdma_id_private *id_priv,
2030 enum rdma_cm_state state)
2031{
2032 cma_cancel_operation(id_priv, state);
2033
2034 rdma_restrack_del(&id_priv->res);
2035 cma_remove_id_from_tree(id_priv);
2036 if (id_priv->cma_dev) {
2037 if (rdma_cap_ib_cm(id_priv->id.device, 1)) {
2038 if (id_priv->cm_id.ib)
2039 ib_destroy_cm_id(id_priv->cm_id.ib);
2040 } else if (rdma_cap_iw_cm(id_priv->id.device, 1)) {
2041 if (id_priv->cm_id.iw)
2042 iw_destroy_cm_id(id_priv->cm_id.iw);
2043 }
2044 cma_leave_mc_groups(id_priv);
2045 cma_release_dev(id_priv);
2046 }
2047
2048 cma_release_port(id_priv);
2049 cma_id_put(id_priv);
2050 wait_for_completion(&id_priv->comp);
2051
2052 if (id_priv->internal_id)
2053 cma_id_put(id_priv->id.context);
2054
2055 kfree(id_priv->id.route.path_rec);
2056 kfree(id_priv->id.route.path_rec_inbound);
2057 kfree(id_priv->id.route.path_rec_outbound);
2058
2059 put_net(id_priv->id.route.addr.dev_addr.net);
2060 kfree(id_priv);
2061}
2062
2063/*
2064 * destroy an ID from within the handler_mutex. This ensures that no other
2065 * handlers can start running concurrently.
2066 */
2067static void destroy_id_handler_unlock(struct rdma_id_private *id_priv)
2068 __releases(&idprv->handler_mutex)
2069{
2070 enum rdma_cm_state state;
2071 unsigned long flags;
2072
2073 trace_cm_id_destroy(id_priv);
2074
2075 /*
2076 * Setting the state to destroyed under the handler mutex provides a
2077 * fence against calling handler callbacks. If this is invoked due to
2078 * the failure of a handler callback then it guarentees that no future
2079 * handlers will be called.
2080 */
2081 lockdep_assert_held(&id_priv->handler_mutex);
2082 spin_lock_irqsave(&id_priv->lock, flags);
2083 state = id_priv->state;
2084 id_priv->state = RDMA_CM_DESTROYING;
2085 spin_unlock_irqrestore(&id_priv->lock, flags);
2086 mutex_unlock(&id_priv->handler_mutex);
2087 _destroy_id(id_priv, state);
2088}
2089
2090void rdma_destroy_id(struct rdma_cm_id *id)
2091{
2092 struct rdma_id_private *id_priv =
2093 container_of(id, struct rdma_id_private, id);
2094
2095 mutex_lock(&id_priv->handler_mutex);
2096 destroy_id_handler_unlock(id_priv);
2097}
2098EXPORT_SYMBOL(rdma_destroy_id);
2099
2100static int cma_rep_recv(struct rdma_id_private *id_priv)
2101{
2102 int ret;
2103
2104 ret = cma_modify_qp_rtr(id_priv, NULL);
2105 if (ret)
2106 goto reject;
2107
2108 ret = cma_modify_qp_rts(id_priv, NULL);
2109 if (ret)
2110 goto reject;
2111
2112 trace_cm_send_rtu(id_priv);
2113 ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
2114 if (ret)
2115 goto reject;
2116
2117 return 0;
2118reject:
2119 pr_debug_ratelimited("RDMA CM: CONNECT_ERROR: failed to handle reply. status %d\n", ret);
2120 cma_modify_qp_err(id_priv);
2121 trace_cm_send_rej(id_priv);
2122 ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
2123 NULL, 0, NULL, 0);
2124 return ret;
2125}
2126
2127static void cma_set_rep_event_data(struct rdma_cm_event *event,
2128 const struct ib_cm_rep_event_param *rep_data,
2129 void *private_data)
2130{
2131 event->param.conn.private_data = private_data;
2132 event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
2133 event->param.conn.responder_resources = rep_data->responder_resources;
2134 event->param.conn.initiator_depth = rep_data->initiator_depth;
2135 event->param.conn.flow_control = rep_data->flow_control;
2136 event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
2137 event->param.conn.srq = rep_data->srq;
2138 event->param.conn.qp_num = rep_data->remote_qpn;
2139
2140 event->ece.vendor_id = rep_data->ece.vendor_id;
2141 event->ece.attr_mod = rep_data->ece.attr_mod;
2142}
2143
2144static int cma_cm_event_handler(struct rdma_id_private *id_priv,
2145 struct rdma_cm_event *event)
2146{
2147 int ret;
2148
2149 lockdep_assert_held(&id_priv->handler_mutex);
2150
2151 trace_cm_event_handler(id_priv, event);
2152 ret = id_priv->id.event_handler(&id_priv->id, event);
2153 trace_cm_event_done(id_priv, event, ret);
2154 return ret;
2155}
2156
2157static int cma_ib_handler(struct ib_cm_id *cm_id,
2158 const struct ib_cm_event *ib_event)
2159{
2160 struct rdma_id_private *id_priv = cm_id->context;
2161 struct rdma_cm_event event = {};
2162 enum rdma_cm_state state;
2163 int ret;
2164
2165 mutex_lock(&id_priv->handler_mutex);
2166 state = READ_ONCE(id_priv->state);
2167 if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
2168 state != RDMA_CM_CONNECT) ||
2169 (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
2170 state != RDMA_CM_DISCONNECT))
2171 goto out;
2172
2173 switch (ib_event->event) {
2174 case IB_CM_REQ_ERROR:
2175 case IB_CM_REP_ERROR:
2176 event.event = RDMA_CM_EVENT_UNREACHABLE;
2177 event.status = -ETIMEDOUT;
2178 break;
2179 case IB_CM_REP_RECEIVED:
2180 if (state == RDMA_CM_CONNECT &&
2181 (id_priv->id.qp_type != IB_QPT_UD)) {
2182 trace_cm_send_mra(id_priv);
2183 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
2184 }
2185 if (id_priv->id.qp) {
2186 event.status = cma_rep_recv(id_priv);
2187 event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
2188 RDMA_CM_EVENT_ESTABLISHED;
2189 } else {
2190 event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
2191 }
2192 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
2193 ib_event->private_data);
2194 break;
2195 case IB_CM_RTU_RECEIVED:
2196 case IB_CM_USER_ESTABLISHED:
2197 event.event = RDMA_CM_EVENT_ESTABLISHED;
2198 break;
2199 case IB_CM_DREQ_ERROR:
2200 event.status = -ETIMEDOUT;
2201 fallthrough;
2202 case IB_CM_DREQ_RECEIVED:
2203 case IB_CM_DREP_RECEIVED:
2204 if (!cma_comp_exch(id_priv, RDMA_CM_CONNECT,
2205 RDMA_CM_DISCONNECT))
2206 goto out;
2207 event.event = RDMA_CM_EVENT_DISCONNECTED;
2208 break;
2209 case IB_CM_TIMEWAIT_EXIT:
2210 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
2211 break;
2212 case IB_CM_MRA_RECEIVED:
2213 /* ignore event */
2214 goto out;
2215 case IB_CM_REJ_RECEIVED:
2216 pr_debug_ratelimited("RDMA CM: REJECTED: %s\n", rdma_reject_msg(&id_priv->id,
2217 ib_event->param.rej_rcvd.reason));
2218 cma_modify_qp_err(id_priv);
2219 event.status = ib_event->param.rej_rcvd.reason;
2220 event.event = RDMA_CM_EVENT_REJECTED;
2221 event.param.conn.private_data = ib_event->private_data;
2222 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
2223 break;
2224 default:
2225 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
2226 ib_event->event);
2227 goto out;
2228 }
2229
2230 ret = cma_cm_event_handler(id_priv, &event);
2231 if (ret) {
2232 /* Destroy the CM ID by returning a non-zero value. */
2233 id_priv->cm_id.ib = NULL;
2234 destroy_id_handler_unlock(id_priv);
2235 return ret;
2236 }
2237out:
2238 mutex_unlock(&id_priv->handler_mutex);
2239 return 0;
2240}
2241
2242static struct rdma_id_private *
2243cma_ib_new_conn_id(const struct rdma_cm_id *listen_id,
2244 const struct ib_cm_event *ib_event,
2245 struct net_device *net_dev)
2246{
2247 struct rdma_id_private *listen_id_priv;
2248 struct rdma_id_private *id_priv;
2249 struct rdma_cm_id *id;
2250 struct rdma_route *rt;
2251 const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
2252 struct sa_path_rec *path = ib_event->param.req_rcvd.primary_path;
2253 const __be64 service_id =
2254 ib_event->param.req_rcvd.primary_path->service_id;
2255 int ret;
2256
2257 listen_id_priv = container_of(listen_id, struct rdma_id_private, id);
2258 id_priv = __rdma_create_id(listen_id->route.addr.dev_addr.net,
2259 listen_id->event_handler, listen_id->context,
2260 listen_id->ps,
2261 ib_event->param.req_rcvd.qp_type,
2262 listen_id_priv);
2263 if (IS_ERR(id_priv))
2264 return NULL;
2265
2266 id = &id_priv->id;
2267 if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
2268 (struct sockaddr *)&id->route.addr.dst_addr,
2269 listen_id, ib_event, ss_family, service_id))
2270 goto err;
2271
2272 rt = &id->route;
2273 rt->num_pri_alt_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
2274 rt->path_rec = kmalloc_array(rt->num_pri_alt_paths,
2275 sizeof(*rt->path_rec), GFP_KERNEL);
2276 if (!rt->path_rec)
2277 goto err;
2278
2279 rt->path_rec[0] = *path;
2280 if (rt->num_pri_alt_paths == 2)
2281 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
2282
2283 if (net_dev) {
2284 rdma_copy_src_l2_addr(&rt->addr.dev_addr, net_dev);
2285 } else {
2286 if (!cma_protocol_roce(listen_id) &&
2287 cma_any_addr(cma_src_addr(id_priv))) {
2288 rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
2289 rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
2290 ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
2291 } else if (!cma_any_addr(cma_src_addr(id_priv))) {
2292 ret = cma_translate_addr(cma_src_addr(id_priv), &rt->addr.dev_addr);
2293 if (ret)
2294 goto err;
2295 }
2296 }
2297 rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
2298
2299 id_priv->state = RDMA_CM_CONNECT;
2300 return id_priv;
2301
2302err:
2303 rdma_destroy_id(id);
2304 return NULL;
2305}
2306
2307static struct rdma_id_private *
2308cma_ib_new_udp_id(const struct rdma_cm_id *listen_id,
2309 const struct ib_cm_event *ib_event,
2310 struct net_device *net_dev)
2311{
2312 const struct rdma_id_private *listen_id_priv;
2313 struct rdma_id_private *id_priv;
2314 struct rdma_cm_id *id;
2315 const sa_family_t ss_family = listen_id->route.addr.src_addr.ss_family;
2316 struct net *net = listen_id->route.addr.dev_addr.net;
2317 int ret;
2318
2319 listen_id_priv = container_of(listen_id, struct rdma_id_private, id);
2320 id_priv = __rdma_create_id(net, listen_id->event_handler,
2321 listen_id->context, listen_id->ps, IB_QPT_UD,
2322 listen_id_priv);
2323 if (IS_ERR(id_priv))
2324 return NULL;
2325
2326 id = &id_priv->id;
2327 if (cma_save_net_info((struct sockaddr *)&id->route.addr.src_addr,
2328 (struct sockaddr *)&id->route.addr.dst_addr,
2329 listen_id, ib_event, ss_family,
2330 ib_event->param.sidr_req_rcvd.service_id))
2331 goto err;
2332
2333 if (net_dev) {
2334 rdma_copy_src_l2_addr(&id->route.addr.dev_addr, net_dev);
2335 } else {
2336 if (!cma_any_addr(cma_src_addr(id_priv))) {
2337 ret = cma_translate_addr(cma_src_addr(id_priv),
2338 &id->route.addr.dev_addr);
2339 if (ret)
2340 goto err;
2341 }
2342 }
2343
2344 id_priv->state = RDMA_CM_CONNECT;
2345 return id_priv;
2346err:
2347 rdma_destroy_id(id);
2348 return NULL;
2349}
2350
2351static void cma_set_req_event_data(struct rdma_cm_event *event,
2352 const struct ib_cm_req_event_param *req_data,
2353 void *private_data, int offset)
2354{
2355 event->param.conn.private_data = private_data + offset;
2356 event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
2357 event->param.conn.responder_resources = req_data->responder_resources;
2358 event->param.conn.initiator_depth = req_data->initiator_depth;
2359 event->param.conn.flow_control = req_data->flow_control;
2360 event->param.conn.retry_count = req_data->retry_count;
2361 event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
2362 event->param.conn.srq = req_data->srq;
2363 event->param.conn.qp_num = req_data->remote_qpn;
2364
2365 event->ece.vendor_id = req_data->ece.vendor_id;
2366 event->ece.attr_mod = req_data->ece.attr_mod;
2367}
2368
2369static int cma_ib_check_req_qp_type(const struct rdma_cm_id *id,
2370 const struct ib_cm_event *ib_event)
2371{
2372 return (((ib_event->event == IB_CM_REQ_RECEIVED) &&
2373 (ib_event->param.req_rcvd.qp_type == id->qp_type)) ||
2374 ((ib_event->event == IB_CM_SIDR_REQ_RECEIVED) &&
2375 (id->qp_type == IB_QPT_UD)) ||
2376 (!id->qp_type));
2377}
2378
2379static int cma_ib_req_handler(struct ib_cm_id *cm_id,
2380 const struct ib_cm_event *ib_event)
2381{
2382 struct rdma_id_private *listen_id, *conn_id = NULL;
2383 struct rdma_cm_event event = {};
2384 struct cma_req_info req = {};
2385 struct net_device *net_dev;
2386 u8 offset;
2387 int ret;
2388
2389 listen_id = cma_ib_id_from_event(cm_id, ib_event, &req, &net_dev);
2390 if (IS_ERR(listen_id))
2391 return PTR_ERR(listen_id);
2392
2393 trace_cm_req_handler(listen_id, ib_event->event);
2394 if (!cma_ib_check_req_qp_type(&listen_id->id, ib_event)) {
2395 ret = -EINVAL;
2396 goto net_dev_put;
2397 }
2398
2399 mutex_lock(&listen_id->handler_mutex);
2400 if (READ_ONCE(listen_id->state) != RDMA_CM_LISTEN) {
2401 ret = -ECONNABORTED;
2402 goto err_unlock;
2403 }
2404
2405 offset = cma_user_data_offset(listen_id);
2406 event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
2407 if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) {
2408 conn_id = cma_ib_new_udp_id(&listen_id->id, ib_event, net_dev);
2409 event.param.ud.private_data = ib_event->private_data + offset;
2410 event.param.ud.private_data_len =
2411 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
2412 } else {
2413 conn_id = cma_ib_new_conn_id(&listen_id->id, ib_event, net_dev);
2414 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
2415 ib_event->private_data, offset);
2416 }
2417 if (!conn_id) {
2418 ret = -ENOMEM;
2419 goto err_unlock;
2420 }
2421
2422 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
2423 ret = cma_ib_acquire_dev(conn_id, listen_id, &req);
2424 if (ret) {
2425 destroy_id_handler_unlock(conn_id);
2426 goto err_unlock;
2427 }
2428
2429 conn_id->cm_id.ib = cm_id;
2430 cm_id->context = conn_id;
2431 cm_id->cm_handler = cma_ib_handler;
2432
2433 ret = cma_cm_event_handler(conn_id, &event);
2434 if (ret) {
2435 /* Destroy the CM ID by returning a non-zero value. */
2436 conn_id->cm_id.ib = NULL;
2437 mutex_unlock(&listen_id->handler_mutex);
2438 destroy_id_handler_unlock(conn_id);
2439 goto net_dev_put;
2440 }
2441
2442 if (READ_ONCE(conn_id->state) == RDMA_CM_CONNECT &&
2443 conn_id->id.qp_type != IB_QPT_UD) {
2444 trace_cm_send_mra(cm_id->context);
2445 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
2446 }
2447 mutex_unlock(&conn_id->handler_mutex);
2448
2449err_unlock:
2450 mutex_unlock(&listen_id->handler_mutex);
2451
2452net_dev_put:
2453 dev_put(net_dev);
2454
2455 return ret;
2456}
2457
2458__be64 rdma_get_service_id(struct rdma_cm_id *id, struct sockaddr *addr)
2459{
2460 if (addr->sa_family == AF_IB)
2461 return ((struct sockaddr_ib *) addr)->sib_sid;
2462
2463 return cpu_to_be64(((u64)id->ps << 16) + be16_to_cpu(cma_port(addr)));
2464}
2465EXPORT_SYMBOL(rdma_get_service_id);
2466
2467void rdma_read_gids(struct rdma_cm_id *cm_id, union ib_gid *sgid,
2468 union ib_gid *dgid)
2469{
2470 struct rdma_addr *addr = &cm_id->route.addr;
2471
2472 if (!cm_id->device) {
2473 if (sgid)
2474 memset(sgid, 0, sizeof(*sgid));
2475 if (dgid)
2476 memset(dgid, 0, sizeof(*dgid));
2477 return;
2478 }
2479
2480 if (rdma_protocol_roce(cm_id->device, cm_id->port_num)) {
2481 if (sgid)
2482 rdma_ip2gid((struct sockaddr *)&addr->src_addr, sgid);
2483 if (dgid)
2484 rdma_ip2gid((struct sockaddr *)&addr->dst_addr, dgid);
2485 } else {
2486 if (sgid)
2487 rdma_addr_get_sgid(&addr->dev_addr, sgid);
2488 if (dgid)
2489 rdma_addr_get_dgid(&addr->dev_addr, dgid);
2490 }
2491}
2492EXPORT_SYMBOL(rdma_read_gids);
2493
2494static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
2495{
2496 struct rdma_id_private *id_priv = iw_id->context;
2497 struct rdma_cm_event event = {};
2498 int ret = 0;
2499 struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
2500 struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
2501
2502 mutex_lock(&id_priv->handler_mutex);
2503 if (READ_ONCE(id_priv->state) != RDMA_CM_CONNECT)
2504 goto out;
2505
2506 switch (iw_event->event) {
2507 case IW_CM_EVENT_CLOSE:
2508 event.event = RDMA_CM_EVENT_DISCONNECTED;
2509 break;
2510 case IW_CM_EVENT_CONNECT_REPLY:
2511 memcpy(cma_src_addr(id_priv), laddr,
2512 rdma_addr_size(laddr));
2513 memcpy(cma_dst_addr(id_priv), raddr,
2514 rdma_addr_size(raddr));
2515 switch (iw_event->status) {
2516 case 0:
2517 event.event = RDMA_CM_EVENT_ESTABLISHED;
2518 event.param.conn.initiator_depth = iw_event->ird;
2519 event.param.conn.responder_resources = iw_event->ord;
2520 break;
2521 case -ECONNRESET:
2522 case -ECONNREFUSED:
2523 event.event = RDMA_CM_EVENT_REJECTED;
2524 break;
2525 case -ETIMEDOUT:
2526 event.event = RDMA_CM_EVENT_UNREACHABLE;
2527 break;
2528 default:
2529 event.event = RDMA_CM_EVENT_CONNECT_ERROR;
2530 break;
2531 }
2532 break;
2533 case IW_CM_EVENT_ESTABLISHED:
2534 event.event = RDMA_CM_EVENT_ESTABLISHED;
2535 event.param.conn.initiator_depth = iw_event->ird;
2536 event.param.conn.responder_resources = iw_event->ord;
2537 break;
2538 default:
2539 goto out;
2540 }
2541
2542 event.status = iw_event->status;
2543 event.param.conn.private_data = iw_event->private_data;
2544 event.param.conn.private_data_len = iw_event->private_data_len;
2545 ret = cma_cm_event_handler(id_priv, &event);
2546 if (ret) {
2547 /* Destroy the CM ID by returning a non-zero value. */
2548 id_priv->cm_id.iw = NULL;
2549 destroy_id_handler_unlock(id_priv);
2550 return ret;
2551 }
2552
2553out:
2554 mutex_unlock(&id_priv->handler_mutex);
2555 return ret;
2556}
2557
2558static int iw_conn_req_handler(struct iw_cm_id *cm_id,
2559 struct iw_cm_event *iw_event)
2560{
2561 struct rdma_id_private *listen_id, *conn_id;
2562 struct rdma_cm_event event = {};
2563 int ret = -ECONNABORTED;
2564 struct sockaddr *laddr = (struct sockaddr *)&iw_event->local_addr;
2565 struct sockaddr *raddr = (struct sockaddr *)&iw_event->remote_addr;
2566
2567 event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
2568 event.param.conn.private_data = iw_event->private_data;
2569 event.param.conn.private_data_len = iw_event->private_data_len;
2570 event.param.conn.initiator_depth = iw_event->ird;
2571 event.param.conn.responder_resources = iw_event->ord;
2572
2573 listen_id = cm_id->context;
2574
2575 mutex_lock(&listen_id->handler_mutex);
2576 if (READ_ONCE(listen_id->state) != RDMA_CM_LISTEN)
2577 goto out;
2578
2579 /* Create a new RDMA id for the new IW CM ID */
2580 conn_id = __rdma_create_id(listen_id->id.route.addr.dev_addr.net,
2581 listen_id->id.event_handler,
2582 listen_id->id.context, RDMA_PS_TCP,
2583 IB_QPT_RC, listen_id);
2584 if (IS_ERR(conn_id)) {
2585 ret = -ENOMEM;
2586 goto out;
2587 }
2588 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
2589 conn_id->state = RDMA_CM_CONNECT;
2590
2591 ret = rdma_translate_ip(laddr, &conn_id->id.route.addr.dev_addr);
2592 if (ret) {
2593 mutex_unlock(&listen_id->handler_mutex);
2594 destroy_id_handler_unlock(conn_id);
2595 return ret;
2596 }
2597
2598 ret = cma_iw_acquire_dev(conn_id, listen_id);
2599 if (ret) {
2600 mutex_unlock(&listen_id->handler_mutex);
2601 destroy_id_handler_unlock(conn_id);
2602 return ret;
2603 }
2604
2605 conn_id->cm_id.iw = cm_id;
2606 cm_id->context = conn_id;
2607 cm_id->cm_handler = cma_iw_handler;
2608
2609 memcpy(cma_src_addr(conn_id), laddr, rdma_addr_size(laddr));
2610 memcpy(cma_dst_addr(conn_id), raddr, rdma_addr_size(raddr));
2611
2612 ret = cma_cm_event_handler(conn_id, &event);
2613 if (ret) {
2614 /* User wants to destroy the CM ID */
2615 conn_id->cm_id.iw = NULL;
2616 mutex_unlock(&listen_id->handler_mutex);
2617 destroy_id_handler_unlock(conn_id);
2618 return ret;
2619 }
2620
2621 mutex_unlock(&conn_id->handler_mutex);
2622
2623out:
2624 mutex_unlock(&listen_id->handler_mutex);
2625 return ret;
2626}
2627
2628static int cma_ib_listen(struct rdma_id_private *id_priv)
2629{
2630 struct sockaddr *addr;
2631 struct ib_cm_id *id;
2632 __be64 svc_id;
2633
2634 addr = cma_src_addr(id_priv);
2635 svc_id = rdma_get_service_id(&id_priv->id, addr);
2636 id = ib_cm_insert_listen(id_priv->id.device,
2637 cma_ib_req_handler, svc_id);
2638 if (IS_ERR(id))
2639 return PTR_ERR(id);
2640 id_priv->cm_id.ib = id;
2641
2642 return 0;
2643}
2644
2645static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
2646{
2647 int ret;
2648 struct iw_cm_id *id;
2649
2650 id = iw_create_cm_id(id_priv->id.device,
2651 iw_conn_req_handler,
2652 id_priv);
2653 if (IS_ERR(id))
2654 return PTR_ERR(id);
2655
2656 mutex_lock(&id_priv->qp_mutex);
2657 id->tos = id_priv->tos;
2658 id->tos_set = id_priv->tos_set;
2659 mutex_unlock(&id_priv->qp_mutex);
2660 id->afonly = id_priv->afonly;
2661 id_priv->cm_id.iw = id;
2662
2663 memcpy(&id_priv->cm_id.iw->local_addr, cma_src_addr(id_priv),
2664 rdma_addr_size(cma_src_addr(id_priv)));
2665
2666 ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
2667
2668 if (ret) {
2669 iw_destroy_cm_id(id_priv->cm_id.iw);
2670 id_priv->cm_id.iw = NULL;
2671 }
2672
2673 return ret;
2674}
2675
2676static int cma_listen_handler(struct rdma_cm_id *id,
2677 struct rdma_cm_event *event)
2678{
2679 struct rdma_id_private *id_priv = id->context;
2680
2681 /* Listening IDs are always destroyed on removal */
2682 if (event->event == RDMA_CM_EVENT_DEVICE_REMOVAL)
2683 return -1;
2684
2685 id->context = id_priv->id.context;
2686 id->event_handler = id_priv->id.event_handler;
2687 trace_cm_event_handler(id_priv, event);
2688 return id_priv->id.event_handler(id, event);
2689}
2690
2691static int cma_listen_on_dev(struct rdma_id_private *id_priv,
2692 struct cma_device *cma_dev,
2693 struct rdma_id_private **to_destroy)
2694{
2695 struct rdma_id_private *dev_id_priv;
2696 struct net *net = id_priv->id.route.addr.dev_addr.net;
2697 int ret;
2698
2699 lockdep_assert_held(&lock);
2700
2701 *to_destroy = NULL;
2702 if (cma_family(id_priv) == AF_IB && !rdma_cap_ib_cm(cma_dev->device, 1))
2703 return 0;
2704
2705 dev_id_priv =
2706 __rdma_create_id(net, cma_listen_handler, id_priv,
2707 id_priv->id.ps, id_priv->id.qp_type, id_priv);
2708 if (IS_ERR(dev_id_priv))
2709 return PTR_ERR(dev_id_priv);
2710
2711 dev_id_priv->state = RDMA_CM_ADDR_BOUND;
2712 memcpy(cma_src_addr(dev_id_priv), cma_src_addr(id_priv),
2713 rdma_addr_size(cma_src_addr(id_priv)));
2714
2715 _cma_attach_to_dev(dev_id_priv, cma_dev);
2716 rdma_restrack_add(&dev_id_priv->res);
2717 cma_id_get(id_priv);
2718 dev_id_priv->internal_id = 1;
2719 dev_id_priv->afonly = id_priv->afonly;
2720 mutex_lock(&id_priv->qp_mutex);
2721 dev_id_priv->tos_set = id_priv->tos_set;
2722 dev_id_priv->tos = id_priv->tos;
2723 mutex_unlock(&id_priv->qp_mutex);
2724
2725 ret = rdma_listen(&dev_id_priv->id, id_priv->backlog);
2726 if (ret)
2727 goto err_listen;
2728 list_add_tail(&dev_id_priv->listen_item, &id_priv->listen_list);
2729 return 0;
2730err_listen:
2731 /* Caller must destroy this after releasing lock */
2732 *to_destroy = dev_id_priv;
2733 dev_warn(&cma_dev->device->dev, "RDMA CMA: %s, error %d\n", __func__, ret);
2734 return ret;
2735}
2736
2737static int cma_listen_on_all(struct rdma_id_private *id_priv)
2738{
2739 struct rdma_id_private *to_destroy;
2740 struct cma_device *cma_dev;
2741 int ret;
2742
2743 mutex_lock(&lock);
2744 list_add_tail(&id_priv->listen_any_item, &listen_any_list);
2745 list_for_each_entry(cma_dev, &dev_list, list) {
2746 ret = cma_listen_on_dev(id_priv, cma_dev, &to_destroy);
2747 if (ret) {
2748 /* Prevent racing with cma_process_remove() */
2749 if (to_destroy)
2750 list_del_init(&to_destroy->device_item);
2751 goto err_listen;
2752 }
2753 }
2754 mutex_unlock(&lock);
2755 return 0;
2756
2757err_listen:
2758 _cma_cancel_listens(id_priv);
2759 mutex_unlock(&lock);
2760 if (to_destroy)
2761 rdma_destroy_id(&to_destroy->id);
2762 return ret;
2763}
2764
2765void rdma_set_service_type(struct rdma_cm_id *id, int tos)
2766{
2767 struct rdma_id_private *id_priv;
2768
2769 id_priv = container_of(id, struct rdma_id_private, id);
2770 mutex_lock(&id_priv->qp_mutex);
2771 id_priv->tos = (u8) tos;
2772 id_priv->tos_set = true;
2773 mutex_unlock(&id_priv->qp_mutex);
2774}
2775EXPORT_SYMBOL(rdma_set_service_type);
2776
2777/**
2778 * rdma_set_ack_timeout() - Set the ack timeout of QP associated
2779 * with a connection identifier.
2780 * @id: Communication identifier to associated with service type.
2781 * @timeout: Ack timeout to set a QP, expressed as 4.096 * 2^(timeout) usec.
2782 *
2783 * This function should be called before rdma_connect() on active side,
2784 * and on passive side before rdma_accept(). It is applicable to primary
2785 * path only. The timeout will affect the local side of the QP, it is not
2786 * negotiated with remote side and zero disables the timer. In case it is
2787 * set before rdma_resolve_route, the value will also be used to determine
2788 * PacketLifeTime for RoCE.
2789 *
2790 * Return: 0 for success
2791 */
2792int rdma_set_ack_timeout(struct rdma_cm_id *id, u8 timeout)
2793{
2794 struct rdma_id_private *id_priv;
2795
2796 if (id->qp_type != IB_QPT_RC && id->qp_type != IB_QPT_XRC_INI)
2797 return -EINVAL;
2798
2799 id_priv = container_of(id, struct rdma_id_private, id);
2800 mutex_lock(&id_priv->qp_mutex);
2801 id_priv->timeout = timeout;
2802 id_priv->timeout_set = true;
2803 mutex_unlock(&id_priv->qp_mutex);
2804
2805 return 0;
2806}
2807EXPORT_SYMBOL(rdma_set_ack_timeout);
2808
2809/**
2810 * rdma_set_min_rnr_timer() - Set the minimum RNR Retry timer of the
2811 * QP associated with a connection identifier.
2812 * @id: Communication identifier to associated with service type.
2813 * @min_rnr_timer: 5-bit value encoded as Table 45: "Encoding for RNR NAK
2814 * Timer Field" in the IBTA specification.
2815 *
2816 * This function should be called before rdma_connect() on active
2817 * side, and on passive side before rdma_accept(). The timer value
2818 * will be associated with the local QP. When it receives a send it is
2819 * not read to handle, typically if the receive queue is empty, an RNR
2820 * Retry NAK is returned to the requester with the min_rnr_timer
2821 * encoded. The requester will then wait at least the time specified
2822 * in the NAK before retrying. The default is zero, which translates
2823 * to a minimum RNR Timer value of 655 ms.
2824 *
2825 * Return: 0 for success
2826 */
2827int rdma_set_min_rnr_timer(struct rdma_cm_id *id, u8 min_rnr_timer)
2828{
2829 struct rdma_id_private *id_priv;
2830
2831 /* It is a five-bit value */
2832 if (min_rnr_timer & 0xe0)
2833 return -EINVAL;
2834
2835 if (WARN_ON(id->qp_type != IB_QPT_RC && id->qp_type != IB_QPT_XRC_TGT))
2836 return -EINVAL;
2837
2838 id_priv = container_of(id, struct rdma_id_private, id);
2839 mutex_lock(&id_priv->qp_mutex);
2840 id_priv->min_rnr_timer = min_rnr_timer;
2841 id_priv->min_rnr_timer_set = true;
2842 mutex_unlock(&id_priv->qp_mutex);
2843
2844 return 0;
2845}
2846EXPORT_SYMBOL(rdma_set_min_rnr_timer);
2847
2848static int route_set_path_rec_inbound(struct cma_work *work,
2849 struct sa_path_rec *path_rec)
2850{
2851 struct rdma_route *route = &work->id->id.route;
2852
2853 if (!route->path_rec_inbound) {
2854 route->path_rec_inbound =
2855 kzalloc(sizeof(*route->path_rec_inbound), GFP_KERNEL);
2856 if (!route->path_rec_inbound)
2857 return -ENOMEM;
2858 }
2859
2860 *route->path_rec_inbound = *path_rec;
2861 return 0;
2862}
2863
2864static int route_set_path_rec_outbound(struct cma_work *work,
2865 struct sa_path_rec *path_rec)
2866{
2867 struct rdma_route *route = &work->id->id.route;
2868
2869 if (!route->path_rec_outbound) {
2870 route->path_rec_outbound =
2871 kzalloc(sizeof(*route->path_rec_outbound), GFP_KERNEL);
2872 if (!route->path_rec_outbound)
2873 return -ENOMEM;
2874 }
2875
2876 *route->path_rec_outbound = *path_rec;
2877 return 0;
2878}
2879
2880static void cma_query_handler(int status, struct sa_path_rec *path_rec,
2881 unsigned int num_prs, void *context)
2882{
2883 struct cma_work *work = context;
2884 struct rdma_route *route;
2885 int i;
2886
2887 route = &work->id->id.route;
2888
2889 if (status)
2890 goto fail;
2891
2892 for (i = 0; i < num_prs; i++) {
2893 if (!path_rec[i].flags || (path_rec[i].flags & IB_PATH_GMP))
2894 *route->path_rec = path_rec[i];
2895 else if (path_rec[i].flags & IB_PATH_INBOUND)
2896 status = route_set_path_rec_inbound(work, &path_rec[i]);
2897 else if (path_rec[i].flags & IB_PATH_OUTBOUND)
2898 status = route_set_path_rec_outbound(work,
2899 &path_rec[i]);
2900 else
2901 status = -EINVAL;
2902
2903 if (status)
2904 goto fail;
2905 }
2906
2907 route->num_pri_alt_paths = 1;
2908 queue_work(cma_wq, &work->work);
2909 return;
2910
2911fail:
2912 work->old_state = RDMA_CM_ROUTE_QUERY;
2913 work->new_state = RDMA_CM_ADDR_RESOLVED;
2914 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
2915 work->event.status = status;
2916 pr_debug_ratelimited("RDMA CM: ROUTE_ERROR: failed to query path. status %d\n",
2917 status);
2918 queue_work(cma_wq, &work->work);
2919}
2920
2921static int cma_query_ib_route(struct rdma_id_private *id_priv,
2922 unsigned long timeout_ms, struct cma_work *work)
2923{
2924 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2925 struct sa_path_rec path_rec;
2926 ib_sa_comp_mask comp_mask;
2927 struct sockaddr_in6 *sin6;
2928 struct sockaddr_ib *sib;
2929
2930 memset(&path_rec, 0, sizeof path_rec);
2931
2932 if (rdma_cap_opa_ah(id_priv->id.device, id_priv->id.port_num))
2933 path_rec.rec_type = SA_PATH_REC_TYPE_OPA;
2934 else
2935 path_rec.rec_type = SA_PATH_REC_TYPE_IB;
2936 rdma_addr_get_sgid(dev_addr, &path_rec.sgid);
2937 rdma_addr_get_dgid(dev_addr, &path_rec.dgid);
2938 path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
2939 path_rec.numb_path = 1;
2940 path_rec.reversible = 1;
2941 path_rec.service_id = rdma_get_service_id(&id_priv->id,
2942 cma_dst_addr(id_priv));
2943
2944 comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
2945 IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
2946 IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
2947
2948 switch (cma_family(id_priv)) {
2949 case AF_INET:
2950 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
2951 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
2952 break;
2953 case AF_INET6:
2954 sin6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
2955 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
2956 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2957 break;
2958 case AF_IB:
2959 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
2960 path_rec.traffic_class = (u8) (be32_to_cpu(sib->sib_flowinfo) >> 20);
2961 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
2962 break;
2963 }
2964
2965 id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
2966 id_priv->id.port_num, &path_rec,
2967 comp_mask, timeout_ms,
2968 GFP_KERNEL, cma_query_handler,
2969 work, &id_priv->query);
2970
2971 return (id_priv->query_id < 0) ? id_priv->query_id : 0;
2972}
2973
2974static void cma_iboe_join_work_handler(struct work_struct *work)
2975{
2976 struct cma_multicast *mc =
2977 container_of(work, struct cma_multicast, iboe_join.work);
2978 struct rdma_cm_event *event = &mc->iboe_join.event;
2979 struct rdma_id_private *id_priv = mc->id_priv;
2980 int ret;
2981
2982 mutex_lock(&id_priv->handler_mutex);
2983 if (READ_ONCE(id_priv->state) == RDMA_CM_DESTROYING ||
2984 READ_ONCE(id_priv->state) == RDMA_CM_DEVICE_REMOVAL)
2985 goto out_unlock;
2986
2987 ret = cma_cm_event_handler(id_priv, event);
2988 WARN_ON(ret);
2989
2990out_unlock:
2991 mutex_unlock(&id_priv->handler_mutex);
2992 if (event->event == RDMA_CM_EVENT_MULTICAST_JOIN)
2993 rdma_destroy_ah_attr(&event->param.ud.ah_attr);
2994}
2995
2996static void cma_work_handler(struct work_struct *_work)
2997{
2998 struct cma_work *work = container_of(_work, struct cma_work, work);
2999 struct rdma_id_private *id_priv = work->id;
3000
3001 mutex_lock(&id_priv->handler_mutex);
3002 if (READ_ONCE(id_priv->state) == RDMA_CM_DESTROYING ||
3003 READ_ONCE(id_priv->state) == RDMA_CM_DEVICE_REMOVAL)
3004 goto out_unlock;
3005 if (work->old_state != 0 || work->new_state != 0) {
3006 if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
3007 goto out_unlock;
3008 }
3009
3010 if (cma_cm_event_handler(id_priv, &work->event)) {
3011 cma_id_put(id_priv);
3012 destroy_id_handler_unlock(id_priv);
3013 goto out_free;
3014 }
3015
3016out_unlock:
3017 mutex_unlock(&id_priv->handler_mutex);
3018 cma_id_put(id_priv);
3019out_free:
3020 if (work->event.event == RDMA_CM_EVENT_MULTICAST_JOIN)
3021 rdma_destroy_ah_attr(&work->event.param.ud.ah_attr);
3022 kfree(work);
3023}
3024
3025static void cma_init_resolve_route_work(struct cma_work *work,
3026 struct rdma_id_private *id_priv)
3027{
3028 work->id = id_priv;
3029 INIT_WORK(&work->work, cma_work_handler);
3030 work->old_state = RDMA_CM_ROUTE_QUERY;
3031 work->new_state = RDMA_CM_ROUTE_RESOLVED;
3032 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
3033}
3034
3035static void enqueue_resolve_addr_work(struct cma_work *work,
3036 struct rdma_id_private *id_priv)
3037{
3038 /* Balances with cma_id_put() in cma_work_handler */
3039 cma_id_get(id_priv);
3040
3041 work->id = id_priv;
3042 INIT_WORK(&work->work, cma_work_handler);
3043 work->old_state = RDMA_CM_ADDR_QUERY;
3044 work->new_state = RDMA_CM_ADDR_RESOLVED;
3045 work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
3046
3047 queue_work(cma_wq, &work->work);
3048}
3049
3050static int cma_resolve_ib_route(struct rdma_id_private *id_priv,
3051 unsigned long timeout_ms)
3052{
3053 struct rdma_route *route = &id_priv->id.route;
3054 struct cma_work *work;
3055 int ret;
3056
3057 work = kzalloc(sizeof *work, GFP_KERNEL);
3058 if (!work)
3059 return -ENOMEM;
3060
3061 cma_init_resolve_route_work(work, id_priv);
3062
3063 if (!route->path_rec)
3064 route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
3065 if (!route->path_rec) {
3066 ret = -ENOMEM;
3067 goto err1;
3068 }
3069
3070 ret = cma_query_ib_route(id_priv, timeout_ms, work);
3071 if (ret)
3072 goto err2;
3073
3074 return 0;
3075err2:
3076 kfree(route->path_rec);
3077 route->path_rec = NULL;
3078err1:
3079 kfree(work);
3080 return ret;
3081}
3082
3083static enum ib_gid_type cma_route_gid_type(enum rdma_network_type network_type,
3084 unsigned long supported_gids,
3085 enum ib_gid_type default_gid)
3086{
3087 if ((network_type == RDMA_NETWORK_IPV4 ||
3088 network_type == RDMA_NETWORK_IPV6) &&
3089 test_bit(IB_GID_TYPE_ROCE_UDP_ENCAP, &supported_gids))
3090 return IB_GID_TYPE_ROCE_UDP_ENCAP;
3091
3092 return default_gid;
3093}
3094
3095/*
3096 * cma_iboe_set_path_rec_l2_fields() is helper function which sets
3097 * path record type based on GID type.
3098 * It also sets up other L2 fields which includes destination mac address
3099 * netdev ifindex, of the path record.
3100 * It returns the netdev of the bound interface for this path record entry.
3101 */
3102static struct net_device *
3103cma_iboe_set_path_rec_l2_fields(struct rdma_id_private *id_priv)
3104{
3105 struct rdma_route *route = &id_priv->id.route;
3106 enum ib_gid_type gid_type = IB_GID_TYPE_ROCE;
3107 struct rdma_addr *addr = &route->addr;
3108 unsigned long supported_gids;
3109 struct net_device *ndev;
3110
3111 if (!addr->dev_addr.bound_dev_if)
3112 return NULL;
3113
3114 ndev = dev_get_by_index(addr->dev_addr.net,
3115 addr->dev_addr.bound_dev_if);
3116 if (!ndev)
3117 return NULL;
3118
3119 supported_gids = roce_gid_type_mask_support(id_priv->id.device,
3120 id_priv->id.port_num);
3121 gid_type = cma_route_gid_type(addr->dev_addr.network,
3122 supported_gids,
3123 id_priv->gid_type);
3124 /* Use the hint from IP Stack to select GID Type */
3125 if (gid_type < ib_network_to_gid_type(addr->dev_addr.network))
3126 gid_type = ib_network_to_gid_type(addr->dev_addr.network);
3127 route->path_rec->rec_type = sa_conv_gid_to_pathrec_type(gid_type);
3128
3129 route->path_rec->roce.route_resolved = true;
3130 sa_path_set_dmac(route->path_rec, addr->dev_addr.dst_dev_addr);
3131 return ndev;
3132}
3133
3134int rdma_set_ib_path(struct rdma_cm_id *id,
3135 struct sa_path_rec *path_rec)
3136{
3137 struct rdma_id_private *id_priv;
3138 struct net_device *ndev;
3139 int ret;
3140
3141 id_priv = container_of(id, struct rdma_id_private, id);
3142 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
3143 RDMA_CM_ROUTE_RESOLVED))
3144 return -EINVAL;
3145
3146 id->route.path_rec = kmemdup(path_rec, sizeof(*path_rec),
3147 GFP_KERNEL);
3148 if (!id->route.path_rec) {
3149 ret = -ENOMEM;
3150 goto err;
3151 }
3152
3153 if (rdma_protocol_roce(id->device, id->port_num)) {
3154 ndev = cma_iboe_set_path_rec_l2_fields(id_priv);
3155 if (!ndev) {
3156 ret = -ENODEV;
3157 goto err_free;
3158 }
3159 dev_put(ndev);
3160 }
3161
3162 id->route.num_pri_alt_paths = 1;
3163 return 0;
3164
3165err_free:
3166 kfree(id->route.path_rec);
3167 id->route.path_rec = NULL;
3168err:
3169 cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED);
3170 return ret;
3171}
3172EXPORT_SYMBOL(rdma_set_ib_path);
3173
3174static int cma_resolve_iw_route(struct rdma_id_private *id_priv)
3175{
3176 struct cma_work *work;
3177
3178 work = kzalloc(sizeof *work, GFP_KERNEL);
3179 if (!work)
3180 return -ENOMEM;
3181
3182 cma_init_resolve_route_work(work, id_priv);
3183 queue_work(cma_wq, &work->work);
3184 return 0;
3185}
3186
3187static int get_vlan_ndev_tc(struct net_device *vlan_ndev, int prio)
3188{
3189 struct net_device *dev;
3190
3191 dev = vlan_dev_real_dev(vlan_ndev);
3192 if (dev->num_tc)
3193 return netdev_get_prio_tc_map(dev, prio);
3194
3195 return (vlan_dev_get_egress_qos_mask(vlan_ndev, prio) &
3196 VLAN_PRIO_MASK) >> VLAN_PRIO_SHIFT;
3197}
3198
3199struct iboe_prio_tc_map {
3200 int input_prio;
3201 int output_tc;
3202 bool found;
3203};
3204
3205static int get_lower_vlan_dev_tc(struct net_device *dev,
3206 struct netdev_nested_priv *priv)
3207{
3208 struct iboe_prio_tc_map *map = (struct iboe_prio_tc_map *)priv->data;
3209
3210 if (is_vlan_dev(dev))
3211 map->output_tc = get_vlan_ndev_tc(dev, map->input_prio);
3212 else if (dev->num_tc)
3213 map->output_tc = netdev_get_prio_tc_map(dev, map->input_prio);
3214 else
3215 map->output_tc = 0;
3216 /* We are interested only in first level VLAN device, so always
3217 * return 1 to stop iterating over next level devices.
3218 */
3219 map->found = true;
3220 return 1;
3221}
3222
3223static int iboe_tos_to_sl(struct net_device *ndev, int tos)
3224{
3225 struct iboe_prio_tc_map prio_tc_map = {};
3226 int prio = rt_tos2priority(tos);
3227 struct netdev_nested_priv priv;
3228
3229 /* If VLAN device, get it directly from the VLAN netdev */
3230 if (is_vlan_dev(ndev))
3231 return get_vlan_ndev_tc(ndev, prio);
3232
3233 prio_tc_map.input_prio = prio;
3234 priv.data = (void *)&prio_tc_map;
3235 rcu_read_lock();
3236 netdev_walk_all_lower_dev_rcu(ndev,
3237 get_lower_vlan_dev_tc,
3238 &priv);
3239 rcu_read_unlock();
3240 /* If map is found from lower device, use it; Otherwise
3241 * continue with the current netdevice to get priority to tc map.
3242 */
3243 if (prio_tc_map.found)
3244 return prio_tc_map.output_tc;
3245 else if (ndev->num_tc)
3246 return netdev_get_prio_tc_map(ndev, prio);
3247 else
3248 return 0;
3249}
3250
3251static __be32 cma_get_roce_udp_flow_label(struct rdma_id_private *id_priv)
3252{
3253 struct sockaddr_in6 *addr6;
3254 u16 dport, sport;
3255 u32 hash, fl;
3256
3257 addr6 = (struct sockaddr_in6 *)cma_src_addr(id_priv);
3258 fl = be32_to_cpu(addr6->sin6_flowinfo) & IB_GRH_FLOWLABEL_MASK;
3259 if ((cma_family(id_priv) != AF_INET6) || !fl) {
3260 dport = be16_to_cpu(cma_port(cma_dst_addr(id_priv)));
3261 sport = be16_to_cpu(cma_port(cma_src_addr(id_priv)));
3262 hash = (u32)sport * 31 + dport;
3263 fl = hash & IB_GRH_FLOWLABEL_MASK;
3264 }
3265
3266 return cpu_to_be32(fl);
3267}
3268
3269static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
3270{
3271 struct rdma_route *route = &id_priv->id.route;
3272 struct rdma_addr *addr = &route->addr;
3273 struct cma_work *work;
3274 int ret;
3275 struct net_device *ndev;
3276
3277 u8 default_roce_tos = id_priv->cma_dev->default_roce_tos[id_priv->id.port_num -
3278 rdma_start_port(id_priv->cma_dev->device)];
3279 u8 tos;
3280
3281 mutex_lock(&id_priv->qp_mutex);
3282 tos = id_priv->tos_set ? id_priv->tos : default_roce_tos;
3283 mutex_unlock(&id_priv->qp_mutex);
3284
3285 work = kzalloc(sizeof *work, GFP_KERNEL);
3286 if (!work)
3287 return -ENOMEM;
3288
3289 route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
3290 if (!route->path_rec) {
3291 ret = -ENOMEM;
3292 goto err1;
3293 }
3294
3295 route->num_pri_alt_paths = 1;
3296
3297 ndev = cma_iboe_set_path_rec_l2_fields(id_priv);
3298 if (!ndev) {
3299 ret = -ENODEV;
3300 goto err2;
3301 }
3302
3303 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
3304 &route->path_rec->sgid);
3305 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.dst_addr,
3306 &route->path_rec->dgid);
3307
3308 if (((struct sockaddr *)&id_priv->id.route.addr.dst_addr)->sa_family != AF_IB)
3309 /* TODO: get the hoplimit from the inet/inet6 device */
3310 route->path_rec->hop_limit = addr->dev_addr.hoplimit;
3311 else
3312 route->path_rec->hop_limit = 1;
3313 route->path_rec->reversible = 1;
3314 route->path_rec->pkey = cpu_to_be16(0xffff);
3315 route->path_rec->mtu_selector = IB_SA_EQ;
3316 route->path_rec->sl = iboe_tos_to_sl(ndev, tos);
3317 route->path_rec->traffic_class = tos;
3318 route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
3319 route->path_rec->rate_selector = IB_SA_EQ;
3320 route->path_rec->rate = IB_RATE_PORT_CURRENT;
3321 dev_put(ndev);
3322 route->path_rec->packet_life_time_selector = IB_SA_EQ;
3323 /* In case ACK timeout is set, use this value to calculate
3324 * PacketLifeTime. As per IBTA 12.7.34,
3325 * local ACK timeout = (2 * PacketLifeTime + Local CA’s ACK delay).
3326 * Assuming a negligible local ACK delay, we can use
3327 * PacketLifeTime = local ACK timeout/2
3328 * as a reasonable approximation for RoCE networks.
3329 */
3330 mutex_lock(&id_priv->qp_mutex);
3331 if (id_priv->timeout_set && id_priv->timeout)
3332 route->path_rec->packet_life_time = id_priv->timeout - 1;
3333 else
3334 route->path_rec->packet_life_time = CMA_IBOE_PACKET_LIFETIME;
3335 mutex_unlock(&id_priv->qp_mutex);
3336
3337 if (!route->path_rec->mtu) {
3338 ret = -EINVAL;
3339 goto err2;
3340 }
3341
3342 if (rdma_protocol_roce_udp_encap(id_priv->id.device,
3343 id_priv->id.port_num))
3344 route->path_rec->flow_label =
3345 cma_get_roce_udp_flow_label(id_priv);
3346
3347 cma_init_resolve_route_work(work, id_priv);
3348 queue_work(cma_wq, &work->work);
3349
3350 return 0;
3351
3352err2:
3353 kfree(route->path_rec);
3354 route->path_rec = NULL;
3355 route->num_pri_alt_paths = 0;
3356err1:
3357 kfree(work);
3358 return ret;
3359}
3360
3361int rdma_resolve_route(struct rdma_cm_id *id, unsigned long timeout_ms)
3362{
3363 struct rdma_id_private *id_priv;
3364 int ret;
3365
3366 if (!timeout_ms)
3367 return -EINVAL;
3368
3369 id_priv = container_of(id, struct rdma_id_private, id);
3370 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY))
3371 return -EINVAL;
3372
3373 cma_id_get(id_priv);
3374 if (rdma_cap_ib_sa(id->device, id->port_num))
3375 ret = cma_resolve_ib_route(id_priv, timeout_ms);
3376 else if (rdma_protocol_roce(id->device, id->port_num)) {
3377 ret = cma_resolve_iboe_route(id_priv);
3378 if (!ret)
3379 cma_add_id_to_tree(id_priv);
3380 }
3381 else if (rdma_protocol_iwarp(id->device, id->port_num))
3382 ret = cma_resolve_iw_route(id_priv);
3383 else
3384 ret = -ENOSYS;
3385
3386 if (ret)
3387 goto err;
3388
3389 return 0;
3390err:
3391 cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED);
3392 cma_id_put(id_priv);
3393 return ret;
3394}
3395EXPORT_SYMBOL(rdma_resolve_route);
3396
3397static void cma_set_loopback(struct sockaddr *addr)
3398{
3399 switch (addr->sa_family) {
3400 case AF_INET:
3401 ((struct sockaddr_in *) addr)->sin_addr.s_addr = htonl(INADDR_LOOPBACK);
3402 break;
3403 case AF_INET6:
3404 ipv6_addr_set(&((struct sockaddr_in6 *) addr)->sin6_addr,
3405 0, 0, 0, htonl(1));
3406 break;
3407 default:
3408 ib_addr_set(&((struct sockaddr_ib *) addr)->sib_addr,
3409 0, 0, 0, htonl(1));
3410 break;
3411 }
3412}
3413
3414static int cma_bind_loopback(struct rdma_id_private *id_priv)
3415{
3416 struct cma_device *cma_dev, *cur_dev;
3417 union ib_gid gid;
3418 enum ib_port_state port_state;
3419 unsigned int p;
3420 u16 pkey;
3421 int ret;
3422
3423 cma_dev = NULL;
3424 mutex_lock(&lock);
3425 list_for_each_entry(cur_dev, &dev_list, list) {
3426 if (cma_family(id_priv) == AF_IB &&
3427 !rdma_cap_ib_cm(cur_dev->device, 1))
3428 continue;
3429
3430 if (!cma_dev)
3431 cma_dev = cur_dev;
3432
3433 rdma_for_each_port (cur_dev->device, p) {
3434 if (!ib_get_cached_port_state(cur_dev->device, p, &port_state) &&
3435 port_state == IB_PORT_ACTIVE) {
3436 cma_dev = cur_dev;
3437 goto port_found;
3438 }
3439 }
3440 }
3441
3442 if (!cma_dev) {
3443 ret = -ENODEV;
3444 goto out;
3445 }
3446
3447 p = 1;
3448
3449port_found:
3450 ret = rdma_query_gid(cma_dev->device, p, 0, &gid);
3451 if (ret)
3452 goto out;
3453
3454 ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
3455 if (ret)
3456 goto out;
3457
3458 id_priv->id.route.addr.dev_addr.dev_type =
3459 (rdma_protocol_ib(cma_dev->device, p)) ?
3460 ARPHRD_INFINIBAND : ARPHRD_ETHER;
3461
3462 rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
3463 ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
3464 id_priv->id.port_num = p;
3465 cma_attach_to_dev(id_priv, cma_dev);
3466 rdma_restrack_add(&id_priv->res);
3467 cma_set_loopback(cma_src_addr(id_priv));
3468out:
3469 mutex_unlock(&lock);
3470 return ret;
3471}
3472
3473static void addr_handler(int status, struct sockaddr *src_addr,
3474 struct rdma_dev_addr *dev_addr, void *context)
3475{
3476 struct rdma_id_private *id_priv = context;
3477 struct rdma_cm_event event = {};
3478 struct sockaddr *addr;
3479 struct sockaddr_storage old_addr;
3480
3481 mutex_lock(&id_priv->handler_mutex);
3482 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY,
3483 RDMA_CM_ADDR_RESOLVED))
3484 goto out;
3485
3486 /*
3487 * Store the previous src address, so that if we fail to acquire
3488 * matching rdma device, old address can be restored back, which helps
3489 * to cancel the cma listen operation correctly.
3490 */
3491 addr = cma_src_addr(id_priv);
3492 memcpy(&old_addr, addr, rdma_addr_size(addr));
3493 memcpy(addr, src_addr, rdma_addr_size(src_addr));
3494 if (!status && !id_priv->cma_dev) {
3495 status = cma_acquire_dev_by_src_ip(id_priv);
3496 if (status)
3497 pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to acquire device. status %d\n",
3498 status);
3499 rdma_restrack_add(&id_priv->res);
3500 } else if (status) {
3501 pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to resolve IP. status %d\n", status);
3502 }
3503
3504 if (status) {
3505 memcpy(addr, &old_addr,
3506 rdma_addr_size((struct sockaddr *)&old_addr));
3507 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
3508 RDMA_CM_ADDR_BOUND))
3509 goto out;
3510 event.event = RDMA_CM_EVENT_ADDR_ERROR;
3511 event.status = status;
3512 } else
3513 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
3514
3515 if (cma_cm_event_handler(id_priv, &event)) {
3516 destroy_id_handler_unlock(id_priv);
3517 return;
3518 }
3519out:
3520 mutex_unlock(&id_priv->handler_mutex);
3521}
3522
3523static int cma_resolve_loopback(struct rdma_id_private *id_priv)
3524{
3525 struct cma_work *work;
3526 union ib_gid gid;
3527 int ret;
3528
3529 work = kzalloc(sizeof *work, GFP_KERNEL);
3530 if (!work)
3531 return -ENOMEM;
3532
3533 if (!id_priv->cma_dev) {
3534 ret = cma_bind_loopback(id_priv);
3535 if (ret)
3536 goto err;
3537 }
3538
3539 rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
3540 rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
3541
3542 enqueue_resolve_addr_work(work, id_priv);
3543 return 0;
3544err:
3545 kfree(work);
3546 return ret;
3547}
3548
3549static int cma_resolve_ib_addr(struct rdma_id_private *id_priv)
3550{
3551 struct cma_work *work;
3552 int ret;
3553
3554 work = kzalloc(sizeof *work, GFP_KERNEL);
3555 if (!work)
3556 return -ENOMEM;
3557
3558 if (!id_priv->cma_dev) {
3559 ret = cma_resolve_ib_dev(id_priv);
3560 if (ret)
3561 goto err;
3562 }
3563
3564 rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, (union ib_gid *)
3565 &(((struct sockaddr_ib *) &id_priv->id.route.addr.dst_addr)->sib_addr));
3566
3567 enqueue_resolve_addr_work(work, id_priv);
3568 return 0;
3569err:
3570 kfree(work);
3571 return ret;
3572}
3573
3574int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse)
3575{
3576 struct rdma_id_private *id_priv;
3577 unsigned long flags;
3578 int ret;
3579
3580 id_priv = container_of(id, struct rdma_id_private, id);
3581 spin_lock_irqsave(&id_priv->lock, flags);
3582 if ((reuse && id_priv->state != RDMA_CM_LISTEN) ||
3583 id_priv->state == RDMA_CM_IDLE) {
3584 id_priv->reuseaddr = reuse;
3585 ret = 0;
3586 } else {
3587 ret = -EINVAL;
3588 }
3589 spin_unlock_irqrestore(&id_priv->lock, flags);
3590 return ret;
3591}
3592EXPORT_SYMBOL(rdma_set_reuseaddr);
3593
3594int rdma_set_afonly(struct rdma_cm_id *id, int afonly)
3595{
3596 struct rdma_id_private *id_priv;
3597 unsigned long flags;
3598 int ret;
3599
3600 id_priv = container_of(id, struct rdma_id_private, id);
3601 spin_lock_irqsave(&id_priv->lock, flags);
3602 if (id_priv->state == RDMA_CM_IDLE || id_priv->state == RDMA_CM_ADDR_BOUND) {
3603 id_priv->options |= (1 << CMA_OPTION_AFONLY);
3604 id_priv->afonly = afonly;
3605 ret = 0;
3606 } else {
3607 ret = -EINVAL;
3608 }
3609 spin_unlock_irqrestore(&id_priv->lock, flags);
3610 return ret;
3611}
3612EXPORT_SYMBOL(rdma_set_afonly);
3613
3614static void cma_bind_port(struct rdma_bind_list *bind_list,
3615 struct rdma_id_private *id_priv)
3616{
3617 struct sockaddr *addr;
3618 struct sockaddr_ib *sib;
3619 u64 sid, mask;
3620 __be16 port;
3621
3622 lockdep_assert_held(&lock);
3623
3624 addr = cma_src_addr(id_priv);
3625 port = htons(bind_list->port);
3626
3627 switch (addr->sa_family) {
3628 case AF_INET:
3629 ((struct sockaddr_in *) addr)->sin_port = port;
3630 break;
3631 case AF_INET6:
3632 ((struct sockaddr_in6 *) addr)->sin6_port = port;
3633 break;
3634 case AF_IB:
3635 sib = (struct sockaddr_ib *) addr;
3636 sid = be64_to_cpu(sib->sib_sid);
3637 mask = be64_to_cpu(sib->sib_sid_mask);
3638 sib->sib_sid = cpu_to_be64((sid & mask) | (u64) ntohs(port));
3639 sib->sib_sid_mask = cpu_to_be64(~0ULL);
3640 break;
3641 }
3642 id_priv->bind_list = bind_list;
3643 hlist_add_head(&id_priv->node, &bind_list->owners);
3644}
3645
3646static int cma_alloc_port(enum rdma_ucm_port_space ps,
3647 struct rdma_id_private *id_priv, unsigned short snum)
3648{
3649 struct rdma_bind_list *bind_list;
3650 int ret;
3651
3652 lockdep_assert_held(&lock);
3653
3654 bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
3655 if (!bind_list)
3656 return -ENOMEM;
3657
3658 ret = cma_ps_alloc(id_priv->id.route.addr.dev_addr.net, ps, bind_list,
3659 snum);
3660 if (ret < 0)
3661 goto err;
3662
3663 bind_list->ps = ps;
3664 bind_list->port = snum;
3665 cma_bind_port(bind_list, id_priv);
3666 return 0;
3667err:
3668 kfree(bind_list);
3669 return ret == -ENOSPC ? -EADDRNOTAVAIL : ret;
3670}
3671
3672static int cma_port_is_unique(struct rdma_bind_list *bind_list,
3673 struct rdma_id_private *id_priv)
3674{
3675 struct rdma_id_private *cur_id;
3676 struct sockaddr *daddr = cma_dst_addr(id_priv);
3677 struct sockaddr *saddr = cma_src_addr(id_priv);
3678 __be16 dport = cma_port(daddr);
3679
3680 lockdep_assert_held(&lock);
3681
3682 hlist_for_each_entry(cur_id, &bind_list->owners, node) {
3683 struct sockaddr *cur_daddr = cma_dst_addr(cur_id);
3684 struct sockaddr *cur_saddr = cma_src_addr(cur_id);
3685 __be16 cur_dport = cma_port(cur_daddr);
3686
3687 if (id_priv == cur_id)
3688 continue;
3689
3690 /* different dest port -> unique */
3691 if (!cma_any_port(daddr) &&
3692 !cma_any_port(cur_daddr) &&
3693 (dport != cur_dport))
3694 continue;
3695
3696 /* different src address -> unique */
3697 if (!cma_any_addr(saddr) &&
3698 !cma_any_addr(cur_saddr) &&
3699 cma_addr_cmp(saddr, cur_saddr))
3700 continue;
3701
3702 /* different dst address -> unique */
3703 if (!cma_any_addr(daddr) &&
3704 !cma_any_addr(cur_daddr) &&
3705 cma_addr_cmp(daddr, cur_daddr))
3706 continue;
3707
3708 return -EADDRNOTAVAIL;
3709 }
3710 return 0;
3711}
3712
3713static int cma_alloc_any_port(enum rdma_ucm_port_space ps,
3714 struct rdma_id_private *id_priv)
3715{
3716 static unsigned int last_used_port;
3717 int low, high, remaining;
3718 unsigned int rover;
3719 struct net *net = id_priv->id.route.addr.dev_addr.net;
3720
3721 lockdep_assert_held(&lock);
3722
3723 inet_get_local_port_range(net, &low, &high);
3724 remaining = (high - low) + 1;
3725 rover = get_random_u32_inclusive(low, remaining + low - 1);
3726retry:
3727 if (last_used_port != rover) {
3728 struct rdma_bind_list *bind_list;
3729 int ret;
3730
3731 bind_list = cma_ps_find(net, ps, (unsigned short)rover);
3732
3733 if (!bind_list) {
3734 ret = cma_alloc_port(ps, id_priv, rover);
3735 } else {
3736 ret = cma_port_is_unique(bind_list, id_priv);
3737 if (!ret)
3738 cma_bind_port(bind_list, id_priv);
3739 }
3740 /*
3741 * Remember previously used port number in order to avoid
3742 * re-using same port immediately after it is closed.
3743 */
3744 if (!ret)
3745 last_used_port = rover;
3746 if (ret != -EADDRNOTAVAIL)
3747 return ret;
3748 }
3749 if (--remaining) {
3750 rover++;
3751 if ((rover < low) || (rover > high))
3752 rover = low;
3753 goto retry;
3754 }
3755 return -EADDRNOTAVAIL;
3756}
3757
3758/*
3759 * Check that the requested port is available. This is called when trying to
3760 * bind to a specific port, or when trying to listen on a bound port. In
3761 * the latter case, the provided id_priv may already be on the bind_list, but
3762 * we still need to check that it's okay to start listening.
3763 */
3764static int cma_check_port(struct rdma_bind_list *bind_list,
3765 struct rdma_id_private *id_priv, uint8_t reuseaddr)
3766{
3767 struct rdma_id_private *cur_id;
3768 struct sockaddr *addr, *cur_addr;
3769
3770 lockdep_assert_held(&lock);
3771
3772 addr = cma_src_addr(id_priv);
3773 hlist_for_each_entry(cur_id, &bind_list->owners, node) {
3774 if (id_priv == cur_id)
3775 continue;
3776
3777 if (reuseaddr && cur_id->reuseaddr)
3778 continue;
3779
3780 cur_addr = cma_src_addr(cur_id);
3781 if (id_priv->afonly && cur_id->afonly &&
3782 (addr->sa_family != cur_addr->sa_family))
3783 continue;
3784
3785 if (cma_any_addr(addr) || cma_any_addr(cur_addr))
3786 return -EADDRNOTAVAIL;
3787
3788 if (!cma_addr_cmp(addr, cur_addr))
3789 return -EADDRINUSE;
3790 }
3791 return 0;
3792}
3793
3794static int cma_use_port(enum rdma_ucm_port_space ps,
3795 struct rdma_id_private *id_priv)
3796{
3797 struct rdma_bind_list *bind_list;
3798 unsigned short snum;
3799 int ret;
3800
3801 lockdep_assert_held(&lock);
3802
3803 snum = ntohs(cma_port(cma_src_addr(id_priv)));
3804 if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
3805 return -EACCES;
3806
3807 bind_list = cma_ps_find(id_priv->id.route.addr.dev_addr.net, ps, snum);
3808 if (!bind_list) {
3809 ret = cma_alloc_port(ps, id_priv, snum);
3810 } else {
3811 ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr);
3812 if (!ret)
3813 cma_bind_port(bind_list, id_priv);
3814 }
3815 return ret;
3816}
3817
3818static enum rdma_ucm_port_space
3819cma_select_inet_ps(struct rdma_id_private *id_priv)
3820{
3821 switch (id_priv->id.ps) {
3822 case RDMA_PS_TCP:
3823 case RDMA_PS_UDP:
3824 case RDMA_PS_IPOIB:
3825 case RDMA_PS_IB:
3826 return id_priv->id.ps;
3827 default:
3828
3829 return 0;
3830 }
3831}
3832
3833static enum rdma_ucm_port_space
3834cma_select_ib_ps(struct rdma_id_private *id_priv)
3835{
3836 enum rdma_ucm_port_space ps = 0;
3837 struct sockaddr_ib *sib;
3838 u64 sid_ps, mask, sid;
3839
3840 sib = (struct sockaddr_ib *) cma_src_addr(id_priv);
3841 mask = be64_to_cpu(sib->sib_sid_mask) & RDMA_IB_IP_PS_MASK;
3842 sid = be64_to_cpu(sib->sib_sid) & mask;
3843
3844 if ((id_priv->id.ps == RDMA_PS_IB) && (sid == (RDMA_IB_IP_PS_IB & mask))) {
3845 sid_ps = RDMA_IB_IP_PS_IB;
3846 ps = RDMA_PS_IB;
3847 } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_TCP)) &&
3848 (sid == (RDMA_IB_IP_PS_TCP & mask))) {
3849 sid_ps = RDMA_IB_IP_PS_TCP;
3850 ps = RDMA_PS_TCP;
3851 } else if (((id_priv->id.ps == RDMA_PS_IB) || (id_priv->id.ps == RDMA_PS_UDP)) &&
3852 (sid == (RDMA_IB_IP_PS_UDP & mask))) {
3853 sid_ps = RDMA_IB_IP_PS_UDP;
3854 ps = RDMA_PS_UDP;
3855 }
3856
3857 if (ps) {
3858 sib->sib_sid = cpu_to_be64(sid_ps | ntohs(cma_port((struct sockaddr *) sib)));
3859 sib->sib_sid_mask = cpu_to_be64(RDMA_IB_IP_PS_MASK |
3860 be64_to_cpu(sib->sib_sid_mask));
3861 }
3862 return ps;
3863}
3864
3865static int cma_get_port(struct rdma_id_private *id_priv)
3866{
3867 enum rdma_ucm_port_space ps;
3868 int ret;
3869
3870 if (cma_family(id_priv) != AF_IB)
3871 ps = cma_select_inet_ps(id_priv);
3872 else
3873 ps = cma_select_ib_ps(id_priv);
3874 if (!ps)
3875 return -EPROTONOSUPPORT;
3876
3877 mutex_lock(&lock);
3878 if (cma_any_port(cma_src_addr(id_priv)))
3879 ret = cma_alloc_any_port(ps, id_priv);
3880 else
3881 ret = cma_use_port(ps, id_priv);
3882 mutex_unlock(&lock);
3883
3884 return ret;
3885}
3886
3887static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
3888 struct sockaddr *addr)
3889{
3890#if IS_ENABLED(CONFIG_IPV6)
3891 struct sockaddr_in6 *sin6;
3892
3893 if (addr->sa_family != AF_INET6)
3894 return 0;
3895
3896 sin6 = (struct sockaddr_in6 *) addr;
3897
3898 if (!(ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL))
3899 return 0;
3900
3901 if (!sin6->sin6_scope_id)
3902 return -EINVAL;
3903
3904 dev_addr->bound_dev_if = sin6->sin6_scope_id;
3905#endif
3906 return 0;
3907}
3908
3909int rdma_listen(struct rdma_cm_id *id, int backlog)
3910{
3911 struct rdma_id_private *id_priv =
3912 container_of(id, struct rdma_id_private, id);
3913 int ret;
3914
3915 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN)) {
3916 struct sockaddr_in any_in = {
3917 .sin_family = AF_INET,
3918 .sin_addr.s_addr = htonl(INADDR_ANY),
3919 };
3920
3921 /* For a well behaved ULP state will be RDMA_CM_IDLE */
3922 ret = rdma_bind_addr(id, (struct sockaddr *)&any_in);
3923 if (ret)
3924 return ret;
3925 if (WARN_ON(!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND,
3926 RDMA_CM_LISTEN)))
3927 return -EINVAL;
3928 }
3929
3930 /*
3931 * Once the ID reaches RDMA_CM_LISTEN it is not allowed to be reusable
3932 * any more, and has to be unique in the bind list.
3933 */
3934 if (id_priv->reuseaddr) {
3935 mutex_lock(&lock);
3936 ret = cma_check_port(id_priv->bind_list, id_priv, 0);
3937 if (!ret)
3938 id_priv->reuseaddr = 0;
3939 mutex_unlock(&lock);
3940 if (ret)
3941 goto err;
3942 }
3943
3944 id_priv->backlog = backlog;
3945 if (id_priv->cma_dev) {
3946 if (rdma_cap_ib_cm(id->device, 1)) {
3947 ret = cma_ib_listen(id_priv);
3948 if (ret)
3949 goto err;
3950 } else if (rdma_cap_iw_cm(id->device, 1)) {
3951 ret = cma_iw_listen(id_priv, backlog);
3952 if (ret)
3953 goto err;
3954 } else {
3955 ret = -ENOSYS;
3956 goto err;
3957 }
3958 } else {
3959 ret = cma_listen_on_all(id_priv);
3960 if (ret)
3961 goto err;
3962 }
3963
3964 return 0;
3965err:
3966 id_priv->backlog = 0;
3967 /*
3968 * All the failure paths that lead here will not allow the req_handler's
3969 * to have run.
3970 */
3971 cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND);
3972 return ret;
3973}
3974EXPORT_SYMBOL(rdma_listen);
3975
3976static int rdma_bind_addr_dst(struct rdma_id_private *id_priv,
3977 struct sockaddr *addr, const struct sockaddr *daddr)
3978{
3979 struct sockaddr *id_daddr;
3980 int ret;
3981
3982 if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6 &&
3983 addr->sa_family != AF_IB)
3984 return -EAFNOSUPPORT;
3985
3986 if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND))
3987 return -EINVAL;
3988
3989 ret = cma_check_linklocal(&id_priv->id.route.addr.dev_addr, addr);
3990 if (ret)
3991 goto err1;
3992
3993 memcpy(cma_src_addr(id_priv), addr, rdma_addr_size(addr));
3994 if (!cma_any_addr(addr)) {
3995 ret = cma_translate_addr(addr, &id_priv->id.route.addr.dev_addr);
3996 if (ret)
3997 goto err1;
3998
3999 ret = cma_acquire_dev_by_src_ip(id_priv);
4000 if (ret)
4001 goto err1;
4002 }
4003
4004 if (!(id_priv->options & (1 << CMA_OPTION_AFONLY))) {
4005 if (addr->sa_family == AF_INET)
4006 id_priv->afonly = 1;
4007#if IS_ENABLED(CONFIG_IPV6)
4008 else if (addr->sa_family == AF_INET6) {
4009 struct net *net = id_priv->id.route.addr.dev_addr.net;
4010
4011 id_priv->afonly = net->ipv6.sysctl.bindv6only;
4012 }
4013#endif
4014 }
4015 id_daddr = cma_dst_addr(id_priv);
4016 if (daddr != id_daddr)
4017 memcpy(id_daddr, daddr, rdma_addr_size(addr));
4018 id_daddr->sa_family = addr->sa_family;
4019
4020 ret = cma_get_port(id_priv);
4021 if (ret)
4022 goto err2;
4023
4024 if (!cma_any_addr(addr))
4025 rdma_restrack_add(&id_priv->res);
4026 return 0;
4027err2:
4028 if (id_priv->cma_dev)
4029 cma_release_dev(id_priv);
4030err1:
4031 cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE);
4032 return ret;
4033}
4034
4035static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
4036 const struct sockaddr *dst_addr)
4037{
4038 struct rdma_id_private *id_priv =
4039 container_of(id, struct rdma_id_private, id);
4040 struct sockaddr_storage zero_sock = {};
4041
4042 if (src_addr && src_addr->sa_family)
4043 return rdma_bind_addr_dst(id_priv, src_addr, dst_addr);
4044
4045 /*
4046 * When the src_addr is not specified, automatically supply an any addr
4047 */
4048 zero_sock.ss_family = dst_addr->sa_family;
4049 if (IS_ENABLED(CONFIG_IPV6) && dst_addr->sa_family == AF_INET6) {
4050 struct sockaddr_in6 *src_addr6 =
4051 (struct sockaddr_in6 *)&zero_sock;
4052 struct sockaddr_in6 *dst_addr6 =
4053 (struct sockaddr_in6 *)dst_addr;
4054
4055 src_addr6->sin6_scope_id = dst_addr6->sin6_scope_id;
4056 if (ipv6_addr_type(&dst_addr6->sin6_addr) & IPV6_ADDR_LINKLOCAL)
4057 id->route.addr.dev_addr.bound_dev_if =
4058 dst_addr6->sin6_scope_id;
4059 } else if (dst_addr->sa_family == AF_IB) {
4060 ((struct sockaddr_ib *)&zero_sock)->sib_pkey =
4061 ((struct sockaddr_ib *)dst_addr)->sib_pkey;
4062 }
4063 return rdma_bind_addr_dst(id_priv, (struct sockaddr *)&zero_sock, dst_addr);
4064}
4065
4066/*
4067 * If required, resolve the source address for bind and leave the id_priv in
4068 * state RDMA_CM_ADDR_BOUND. This oddly uses the state to determine the prior
4069 * calls made by ULP, a previously bound ID will not be re-bound and src_addr is
4070 * ignored.
4071 */
4072static int resolve_prepare_src(struct rdma_id_private *id_priv,
4073 struct sockaddr *src_addr,
4074 const struct sockaddr *dst_addr)
4075{
4076 int ret;
4077
4078 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY)) {
4079 /* For a well behaved ULP state will be RDMA_CM_IDLE */
4080 ret = cma_bind_addr(&id_priv->id, src_addr, dst_addr);
4081 if (ret)
4082 return ret;
4083 if (WARN_ON(!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND,
4084 RDMA_CM_ADDR_QUERY)))
4085 return -EINVAL;
4086
4087 } else {
4088 memcpy(cma_dst_addr(id_priv), dst_addr, rdma_addr_size(dst_addr));
4089 }
4090
4091 if (cma_family(id_priv) != dst_addr->sa_family) {
4092 ret = -EINVAL;
4093 goto err_state;
4094 }
4095 return 0;
4096
4097err_state:
4098 cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
4099 return ret;
4100}
4101
4102int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
4103 const struct sockaddr *dst_addr, unsigned long timeout_ms)
4104{
4105 struct rdma_id_private *id_priv =
4106 container_of(id, struct rdma_id_private, id);
4107 int ret;
4108
4109 ret = resolve_prepare_src(id_priv, src_addr, dst_addr);
4110 if (ret)
4111 return ret;
4112
4113 if (cma_any_addr(dst_addr)) {
4114 ret = cma_resolve_loopback(id_priv);
4115 } else {
4116 if (dst_addr->sa_family == AF_IB) {
4117 ret = cma_resolve_ib_addr(id_priv);
4118 } else {
4119 /*
4120 * The FSM can return back to RDMA_CM_ADDR_BOUND after
4121 * rdma_resolve_ip() is called, eg through the error
4122 * path in addr_handler(). If this happens the existing
4123 * request must be canceled before issuing a new one.
4124 * Since canceling a request is a bit slow and this
4125 * oddball path is rare, keep track once a request has
4126 * been issued. The track turns out to be a permanent
4127 * state since this is the only cancel as it is
4128 * immediately before rdma_resolve_ip().
4129 */
4130 if (id_priv->used_resolve_ip)
4131 rdma_addr_cancel(&id->route.addr.dev_addr);
4132 else
4133 id_priv->used_resolve_ip = 1;
4134 ret = rdma_resolve_ip(cma_src_addr(id_priv), dst_addr,
4135 &id->route.addr.dev_addr,
4136 timeout_ms, addr_handler,
4137 false, id_priv);
4138 }
4139 }
4140 if (ret)
4141 goto err;
4142
4143 return 0;
4144err:
4145 cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
4146 return ret;
4147}
4148EXPORT_SYMBOL(rdma_resolve_addr);
4149
4150int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
4151{
4152 struct rdma_id_private *id_priv =
4153 container_of(id, struct rdma_id_private, id);
4154
4155 return rdma_bind_addr_dst(id_priv, addr, cma_dst_addr(id_priv));
4156}
4157EXPORT_SYMBOL(rdma_bind_addr);
4158
4159static int cma_format_hdr(void *hdr, struct rdma_id_private *id_priv)
4160{
4161 struct cma_hdr *cma_hdr;
4162
4163 cma_hdr = hdr;
4164 cma_hdr->cma_version = CMA_VERSION;
4165 if (cma_family(id_priv) == AF_INET) {
4166 struct sockaddr_in *src4, *dst4;
4167
4168 src4 = (struct sockaddr_in *) cma_src_addr(id_priv);
4169 dst4 = (struct sockaddr_in *) cma_dst_addr(id_priv);
4170
4171 cma_set_ip_ver(cma_hdr, 4);
4172 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
4173 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
4174 cma_hdr->port = src4->sin_port;
4175 } else if (cma_family(id_priv) == AF_INET6) {
4176 struct sockaddr_in6 *src6, *dst6;
4177
4178 src6 = (struct sockaddr_in6 *) cma_src_addr(id_priv);
4179 dst6 = (struct sockaddr_in6 *) cma_dst_addr(id_priv);
4180
4181 cma_set_ip_ver(cma_hdr, 6);
4182 cma_hdr->src_addr.ip6 = src6->sin6_addr;
4183 cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
4184 cma_hdr->port = src6->sin6_port;
4185 }
4186 return 0;
4187}
4188
4189static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
4190 const struct ib_cm_event *ib_event)
4191{
4192 struct rdma_id_private *id_priv = cm_id->context;
4193 struct rdma_cm_event event = {};
4194 const struct ib_cm_sidr_rep_event_param *rep =
4195 &ib_event->param.sidr_rep_rcvd;
4196 int ret;
4197
4198 mutex_lock(&id_priv->handler_mutex);
4199 if (READ_ONCE(id_priv->state) != RDMA_CM_CONNECT)
4200 goto out;
4201
4202 switch (ib_event->event) {
4203 case IB_CM_SIDR_REQ_ERROR:
4204 event.event = RDMA_CM_EVENT_UNREACHABLE;
4205 event.status = -ETIMEDOUT;
4206 break;
4207 case IB_CM_SIDR_REP_RECEIVED:
4208 event.param.ud.private_data = ib_event->private_data;
4209 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
4210 if (rep->status != IB_SIDR_SUCCESS) {
4211 event.event = RDMA_CM_EVENT_UNREACHABLE;
4212 event.status = ib_event->param.sidr_rep_rcvd.status;
4213 pr_debug_ratelimited("RDMA CM: UNREACHABLE: bad SIDR reply. status %d\n",
4214 event.status);
4215 break;
4216 }
4217 ret = cma_set_qkey(id_priv, rep->qkey);
4218 if (ret) {
4219 pr_debug_ratelimited("RDMA CM: ADDR_ERROR: failed to set qkey. status %d\n", ret);
4220 event.event = RDMA_CM_EVENT_ADDR_ERROR;
4221 event.status = ret;
4222 break;
4223 }
4224 ib_init_ah_attr_from_path(id_priv->id.device,
4225 id_priv->id.port_num,
4226 id_priv->id.route.path_rec,
4227 &event.param.ud.ah_attr,
4228 rep->sgid_attr);
4229 event.param.ud.qp_num = rep->qpn;
4230 event.param.ud.qkey = rep->qkey;
4231 event.event = RDMA_CM_EVENT_ESTABLISHED;
4232 event.status = 0;
4233 break;
4234 default:
4235 pr_err("RDMA CMA: unexpected IB CM event: %d\n",
4236 ib_event->event);
4237 goto out;
4238 }
4239
4240 ret = cma_cm_event_handler(id_priv, &event);
4241
4242 rdma_destroy_ah_attr(&event.param.ud.ah_attr);
4243 if (ret) {
4244 /* Destroy the CM ID by returning a non-zero value. */
4245 id_priv->cm_id.ib = NULL;
4246 destroy_id_handler_unlock(id_priv);
4247 return ret;
4248 }
4249out:
4250 mutex_unlock(&id_priv->handler_mutex);
4251 return 0;
4252}
4253
4254static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
4255 struct rdma_conn_param *conn_param)
4256{
4257 struct ib_cm_sidr_req_param req;
4258 struct ib_cm_id *id;
4259 void *private_data;
4260 u8 offset;
4261 int ret;
4262
4263 memset(&req, 0, sizeof req);
4264 offset = cma_user_data_offset(id_priv);
4265 if (check_add_overflow(offset, conn_param->private_data_len, &req.private_data_len))
4266 return -EINVAL;
4267
4268 if (req.private_data_len) {
4269 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
4270 if (!private_data)
4271 return -ENOMEM;
4272 } else {
4273 private_data = NULL;
4274 }
4275
4276 if (conn_param->private_data && conn_param->private_data_len)
4277 memcpy(private_data + offset, conn_param->private_data,
4278 conn_param->private_data_len);
4279
4280 if (private_data) {
4281 ret = cma_format_hdr(private_data, id_priv);
4282 if (ret)
4283 goto out;
4284 req.private_data = private_data;
4285 }
4286
4287 id = ib_create_cm_id(id_priv->id.device, cma_sidr_rep_handler,
4288 id_priv);
4289 if (IS_ERR(id)) {
4290 ret = PTR_ERR(id);
4291 goto out;
4292 }
4293 id_priv->cm_id.ib = id;
4294
4295 req.path = id_priv->id.route.path_rec;
4296 req.sgid_attr = id_priv->id.route.addr.dev_addr.sgid_attr;
4297 req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
4298 req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
4299 req.max_cm_retries = CMA_MAX_CM_RETRIES;
4300
4301 trace_cm_send_sidr_req(id_priv);
4302 ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
4303 if (ret) {
4304 ib_destroy_cm_id(id_priv->cm_id.ib);
4305 id_priv->cm_id.ib = NULL;
4306 }
4307out:
4308 kfree(private_data);
4309 return ret;
4310}
4311
4312static int cma_connect_ib(struct rdma_id_private *id_priv,
4313 struct rdma_conn_param *conn_param)
4314{
4315 struct ib_cm_req_param req;
4316 struct rdma_route *route;
4317 void *private_data;
4318 struct ib_cm_id *id;
4319 u8 offset;
4320 int ret;
4321
4322 memset(&req, 0, sizeof req);
4323 offset = cma_user_data_offset(id_priv);
4324 if (check_add_overflow(offset, conn_param->private_data_len, &req.private_data_len))
4325 return -EINVAL;
4326
4327 if (req.private_data_len) {
4328 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
4329 if (!private_data)
4330 return -ENOMEM;
4331 } else {
4332 private_data = NULL;
4333 }
4334
4335 if (conn_param->private_data && conn_param->private_data_len)
4336 memcpy(private_data + offset, conn_param->private_data,
4337 conn_param->private_data_len);
4338
4339 id = ib_create_cm_id(id_priv->id.device, cma_ib_handler, id_priv);
4340 if (IS_ERR(id)) {
4341 ret = PTR_ERR(id);
4342 goto out;
4343 }
4344 id_priv->cm_id.ib = id;
4345
4346 route = &id_priv->id.route;
4347 if (private_data) {
4348 ret = cma_format_hdr(private_data, id_priv);
4349 if (ret)
4350 goto out;
4351 req.private_data = private_data;
4352 }
4353
4354 req.primary_path = &route->path_rec[0];
4355 req.primary_path_inbound = route->path_rec_inbound;
4356 req.primary_path_outbound = route->path_rec_outbound;
4357 if (route->num_pri_alt_paths == 2)
4358 req.alternate_path = &route->path_rec[1];
4359
4360 req.ppath_sgid_attr = id_priv->id.route.addr.dev_addr.sgid_attr;
4361 /* Alternate path SGID attribute currently unsupported */
4362 req.service_id = rdma_get_service_id(&id_priv->id, cma_dst_addr(id_priv));
4363 req.qp_num = id_priv->qp_num;
4364 req.qp_type = id_priv->id.qp_type;
4365 req.starting_psn = id_priv->seq_num;
4366 req.responder_resources = conn_param->responder_resources;
4367 req.initiator_depth = conn_param->initiator_depth;
4368 req.flow_control = conn_param->flow_control;
4369 req.retry_count = min_t(u8, 7, conn_param->retry_count);
4370 req.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
4371 req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
4372 req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
4373 req.max_cm_retries = CMA_MAX_CM_RETRIES;
4374 req.srq = id_priv->srq ? 1 : 0;
4375 req.ece.vendor_id = id_priv->ece.vendor_id;
4376 req.ece.attr_mod = id_priv->ece.attr_mod;
4377
4378 trace_cm_send_req(id_priv);
4379 ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
4380out:
4381 if (ret && !IS_ERR(id)) {
4382 ib_destroy_cm_id(id);
4383 id_priv->cm_id.ib = NULL;
4384 }
4385
4386 kfree(private_data);
4387 return ret;
4388}
4389
4390static int cma_connect_iw(struct rdma_id_private *id_priv,
4391 struct rdma_conn_param *conn_param)
4392{
4393 struct iw_cm_id *cm_id;
4394 int ret;
4395 struct iw_cm_conn_param iw_param;
4396
4397 cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
4398 if (IS_ERR(cm_id))
4399 return PTR_ERR(cm_id);
4400
4401 mutex_lock(&id_priv->qp_mutex);
4402 cm_id->tos = id_priv->tos;
4403 cm_id->tos_set = id_priv->tos_set;
4404 mutex_unlock(&id_priv->qp_mutex);
4405
4406 id_priv->cm_id.iw = cm_id;
4407
4408 memcpy(&cm_id->local_addr, cma_src_addr(id_priv),
4409 rdma_addr_size(cma_src_addr(id_priv)));
4410 memcpy(&cm_id->remote_addr, cma_dst_addr(id_priv),
4411 rdma_addr_size(cma_dst_addr(id_priv)));
4412
4413 ret = cma_modify_qp_rtr(id_priv, conn_param);
4414 if (ret)
4415 goto out;
4416
4417 if (conn_param) {
4418 iw_param.ord = conn_param->initiator_depth;
4419 iw_param.ird = conn_param->responder_resources;
4420 iw_param.private_data = conn_param->private_data;
4421 iw_param.private_data_len = conn_param->private_data_len;
4422 iw_param.qpn = id_priv->id.qp ? id_priv->qp_num : conn_param->qp_num;
4423 } else {
4424 memset(&iw_param, 0, sizeof iw_param);
4425 iw_param.qpn = id_priv->qp_num;
4426 }
4427 ret = iw_cm_connect(cm_id, &iw_param);
4428out:
4429 if (ret) {
4430 iw_destroy_cm_id(cm_id);
4431 id_priv->cm_id.iw = NULL;
4432 }
4433 return ret;
4434}
4435
4436/**
4437 * rdma_connect_locked - Initiate an active connection request.
4438 * @id: Connection identifier to connect.
4439 * @conn_param: Connection information used for connected QPs.
4440 *
4441 * Same as rdma_connect() but can only be called from the
4442 * RDMA_CM_EVENT_ROUTE_RESOLVED handler callback.
4443 */
4444int rdma_connect_locked(struct rdma_cm_id *id,
4445 struct rdma_conn_param *conn_param)
4446{
4447 struct rdma_id_private *id_priv =
4448 container_of(id, struct rdma_id_private, id);
4449 int ret;
4450
4451 if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT))
4452 return -EINVAL;
4453
4454 if (!id->qp) {
4455 id_priv->qp_num = conn_param->qp_num;
4456 id_priv->srq = conn_param->srq;
4457 }
4458
4459 if (rdma_cap_ib_cm(id->device, id->port_num)) {
4460 if (id->qp_type == IB_QPT_UD)
4461 ret = cma_resolve_ib_udp(id_priv, conn_param);
4462 else
4463 ret = cma_connect_ib(id_priv, conn_param);
4464 } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4465 ret = cma_connect_iw(id_priv, conn_param);
4466 } else {
4467 ret = -ENOSYS;
4468 }
4469 if (ret)
4470 goto err_state;
4471 return 0;
4472err_state:
4473 cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED);
4474 return ret;
4475}
4476EXPORT_SYMBOL(rdma_connect_locked);
4477
4478/**
4479 * rdma_connect - Initiate an active connection request.
4480 * @id: Connection identifier to connect.
4481 * @conn_param: Connection information used for connected QPs.
4482 *
4483 * Users must have resolved a route for the rdma_cm_id to connect with by having
4484 * called rdma_resolve_route before calling this routine.
4485 *
4486 * This call will either connect to a remote QP or obtain remote QP information
4487 * for unconnected rdma_cm_id's. The actual operation is based on the
4488 * rdma_cm_id's port space.
4489 */
4490int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
4491{
4492 struct rdma_id_private *id_priv =
4493 container_of(id, struct rdma_id_private, id);
4494 int ret;
4495
4496 mutex_lock(&id_priv->handler_mutex);
4497 ret = rdma_connect_locked(id, conn_param);
4498 mutex_unlock(&id_priv->handler_mutex);
4499 return ret;
4500}
4501EXPORT_SYMBOL(rdma_connect);
4502
4503/**
4504 * rdma_connect_ece - Initiate an active connection request with ECE data.
4505 * @id: Connection identifier to connect.
4506 * @conn_param: Connection information used for connected QPs.
4507 * @ece: ECE parameters
4508 *
4509 * See rdma_connect() explanation.
4510 */
4511int rdma_connect_ece(struct rdma_cm_id *id, struct rdma_conn_param *conn_param,
4512 struct rdma_ucm_ece *ece)
4513{
4514 struct rdma_id_private *id_priv =
4515 container_of(id, struct rdma_id_private, id);
4516
4517 id_priv->ece.vendor_id = ece->vendor_id;
4518 id_priv->ece.attr_mod = ece->attr_mod;
4519
4520 return rdma_connect(id, conn_param);
4521}
4522EXPORT_SYMBOL(rdma_connect_ece);
4523
4524static int cma_accept_ib(struct rdma_id_private *id_priv,
4525 struct rdma_conn_param *conn_param)
4526{
4527 struct ib_cm_rep_param rep;
4528 int ret;
4529
4530 ret = cma_modify_qp_rtr(id_priv, conn_param);
4531 if (ret)
4532 goto out;
4533
4534 ret = cma_modify_qp_rts(id_priv, conn_param);
4535 if (ret)
4536 goto out;
4537
4538 memset(&rep, 0, sizeof rep);
4539 rep.qp_num = id_priv->qp_num;
4540 rep.starting_psn = id_priv->seq_num;
4541 rep.private_data = conn_param->private_data;
4542 rep.private_data_len = conn_param->private_data_len;
4543 rep.responder_resources = conn_param->responder_resources;
4544 rep.initiator_depth = conn_param->initiator_depth;
4545 rep.failover_accepted = 0;
4546 rep.flow_control = conn_param->flow_control;
4547 rep.rnr_retry_count = min_t(u8, 7, conn_param->rnr_retry_count);
4548 rep.srq = id_priv->srq ? 1 : 0;
4549 rep.ece.vendor_id = id_priv->ece.vendor_id;
4550 rep.ece.attr_mod = id_priv->ece.attr_mod;
4551
4552 trace_cm_send_rep(id_priv);
4553 ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
4554out:
4555 return ret;
4556}
4557
4558static int cma_accept_iw(struct rdma_id_private *id_priv,
4559 struct rdma_conn_param *conn_param)
4560{
4561 struct iw_cm_conn_param iw_param;
4562 int ret;
4563
4564 if (!conn_param)
4565 return -EINVAL;
4566
4567 ret = cma_modify_qp_rtr(id_priv, conn_param);
4568 if (ret)
4569 return ret;
4570
4571 iw_param.ord = conn_param->initiator_depth;
4572 iw_param.ird = conn_param->responder_resources;
4573 iw_param.private_data = conn_param->private_data;
4574 iw_param.private_data_len = conn_param->private_data_len;
4575 if (id_priv->id.qp)
4576 iw_param.qpn = id_priv->qp_num;
4577 else
4578 iw_param.qpn = conn_param->qp_num;
4579
4580 return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
4581}
4582
4583static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
4584 enum ib_cm_sidr_status status, u32 qkey,
4585 const void *private_data, int private_data_len)
4586{
4587 struct ib_cm_sidr_rep_param rep;
4588 int ret;
4589
4590 memset(&rep, 0, sizeof rep);
4591 rep.status = status;
4592 if (status == IB_SIDR_SUCCESS) {
4593 if (qkey)
4594 ret = cma_set_qkey(id_priv, qkey);
4595 else
4596 ret = cma_set_default_qkey(id_priv);
4597 if (ret)
4598 return ret;
4599 rep.qp_num = id_priv->qp_num;
4600 rep.qkey = id_priv->qkey;
4601
4602 rep.ece.vendor_id = id_priv->ece.vendor_id;
4603 rep.ece.attr_mod = id_priv->ece.attr_mod;
4604 }
4605
4606 rep.private_data = private_data;
4607 rep.private_data_len = private_data_len;
4608
4609 trace_cm_send_sidr_rep(id_priv);
4610 return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
4611}
4612
4613/**
4614 * rdma_accept - Called to accept a connection request or response.
4615 * @id: Connection identifier associated with the request.
4616 * @conn_param: Information needed to establish the connection. This must be
4617 * provided if accepting a connection request. If accepting a connection
4618 * response, this parameter must be NULL.
4619 *
4620 * Typically, this routine is only called by the listener to accept a connection
4621 * request. It must also be called on the active side of a connection if the
4622 * user is performing their own QP transitions.
4623 *
4624 * In the case of error, a reject message is sent to the remote side and the
4625 * state of the qp associated with the id is modified to error, such that any
4626 * previously posted receive buffers would be flushed.
4627 *
4628 * This function is for use by kernel ULPs and must be called from under the
4629 * handler callback.
4630 */
4631int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
4632{
4633 struct rdma_id_private *id_priv =
4634 container_of(id, struct rdma_id_private, id);
4635 int ret;
4636
4637 lockdep_assert_held(&id_priv->handler_mutex);
4638
4639 if (READ_ONCE(id_priv->state) != RDMA_CM_CONNECT)
4640 return -EINVAL;
4641
4642 if (!id->qp && conn_param) {
4643 id_priv->qp_num = conn_param->qp_num;
4644 id_priv->srq = conn_param->srq;
4645 }
4646
4647 if (rdma_cap_ib_cm(id->device, id->port_num)) {
4648 if (id->qp_type == IB_QPT_UD) {
4649 if (conn_param)
4650 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
4651 conn_param->qkey,
4652 conn_param->private_data,
4653 conn_param->private_data_len);
4654 else
4655 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
4656 0, NULL, 0);
4657 } else {
4658 if (conn_param)
4659 ret = cma_accept_ib(id_priv, conn_param);
4660 else
4661 ret = cma_rep_recv(id_priv);
4662 }
4663 } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4664 ret = cma_accept_iw(id_priv, conn_param);
4665 } else {
4666 ret = -ENOSYS;
4667 }
4668 if (ret)
4669 goto reject;
4670
4671 return 0;
4672reject:
4673 cma_modify_qp_err(id_priv);
4674 rdma_reject(id, NULL, 0, IB_CM_REJ_CONSUMER_DEFINED);
4675 return ret;
4676}
4677EXPORT_SYMBOL(rdma_accept);
4678
4679int rdma_accept_ece(struct rdma_cm_id *id, struct rdma_conn_param *conn_param,
4680 struct rdma_ucm_ece *ece)
4681{
4682 struct rdma_id_private *id_priv =
4683 container_of(id, struct rdma_id_private, id);
4684
4685 id_priv->ece.vendor_id = ece->vendor_id;
4686 id_priv->ece.attr_mod = ece->attr_mod;
4687
4688 return rdma_accept(id, conn_param);
4689}
4690EXPORT_SYMBOL(rdma_accept_ece);
4691
4692void rdma_lock_handler(struct rdma_cm_id *id)
4693{
4694 struct rdma_id_private *id_priv =
4695 container_of(id, struct rdma_id_private, id);
4696
4697 mutex_lock(&id_priv->handler_mutex);
4698}
4699EXPORT_SYMBOL(rdma_lock_handler);
4700
4701void rdma_unlock_handler(struct rdma_cm_id *id)
4702{
4703 struct rdma_id_private *id_priv =
4704 container_of(id, struct rdma_id_private, id);
4705
4706 mutex_unlock(&id_priv->handler_mutex);
4707}
4708EXPORT_SYMBOL(rdma_unlock_handler);
4709
4710int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
4711{
4712 struct rdma_id_private *id_priv;
4713 int ret;
4714
4715 id_priv = container_of(id, struct rdma_id_private, id);
4716 if (!id_priv->cm_id.ib)
4717 return -EINVAL;
4718
4719 switch (id->device->node_type) {
4720 case RDMA_NODE_IB_CA:
4721 ret = ib_cm_notify(id_priv->cm_id.ib, event);
4722 break;
4723 default:
4724 ret = 0;
4725 break;
4726 }
4727 return ret;
4728}
4729EXPORT_SYMBOL(rdma_notify);
4730
4731int rdma_reject(struct rdma_cm_id *id, const void *private_data,
4732 u8 private_data_len, u8 reason)
4733{
4734 struct rdma_id_private *id_priv;
4735 int ret;
4736
4737 id_priv = container_of(id, struct rdma_id_private, id);
4738 if (!id_priv->cm_id.ib)
4739 return -EINVAL;
4740
4741 if (rdma_cap_ib_cm(id->device, id->port_num)) {
4742 if (id->qp_type == IB_QPT_UD) {
4743 ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT, 0,
4744 private_data, private_data_len);
4745 } else {
4746 trace_cm_send_rej(id_priv);
4747 ret = ib_send_cm_rej(id_priv->cm_id.ib, reason, NULL, 0,
4748 private_data, private_data_len);
4749 }
4750 } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4751 ret = iw_cm_reject(id_priv->cm_id.iw,
4752 private_data, private_data_len);
4753 } else {
4754 ret = -ENOSYS;
4755 }
4756
4757 return ret;
4758}
4759EXPORT_SYMBOL(rdma_reject);
4760
4761int rdma_disconnect(struct rdma_cm_id *id)
4762{
4763 struct rdma_id_private *id_priv;
4764 int ret;
4765
4766 id_priv = container_of(id, struct rdma_id_private, id);
4767 if (!id_priv->cm_id.ib)
4768 return -EINVAL;
4769
4770 if (rdma_cap_ib_cm(id->device, id->port_num)) {
4771 ret = cma_modify_qp_err(id_priv);
4772 if (ret)
4773 goto out;
4774 /* Initiate or respond to a disconnect. */
4775 trace_cm_disconnect(id_priv);
4776 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0)) {
4777 if (!ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0))
4778 trace_cm_sent_drep(id_priv);
4779 } else {
4780 trace_cm_sent_dreq(id_priv);
4781 }
4782 } else if (rdma_cap_iw_cm(id->device, id->port_num)) {
4783 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
4784 } else
4785 ret = -EINVAL;
4786
4787out:
4788 return ret;
4789}
4790EXPORT_SYMBOL(rdma_disconnect);
4791
4792static void cma_make_mc_event(int status, struct rdma_id_private *id_priv,
4793 struct ib_sa_multicast *multicast,
4794 struct rdma_cm_event *event,
4795 struct cma_multicast *mc)
4796{
4797 struct rdma_dev_addr *dev_addr;
4798 enum ib_gid_type gid_type;
4799 struct net_device *ndev;
4800
4801 if (status)
4802 pr_debug_ratelimited("RDMA CM: MULTICAST_ERROR: failed to join multicast. status %d\n",
4803 status);
4804
4805 event->status = status;
4806 event->param.ud.private_data = mc->context;
4807 if (status) {
4808 event->event = RDMA_CM_EVENT_MULTICAST_ERROR;
4809 return;
4810 }
4811
4812 dev_addr = &id_priv->id.route.addr.dev_addr;
4813 ndev = dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if);
4814 gid_type =
4815 id_priv->cma_dev
4816 ->default_gid_type[id_priv->id.port_num -
4817 rdma_start_port(
4818 id_priv->cma_dev->device)];
4819
4820 event->event = RDMA_CM_EVENT_MULTICAST_JOIN;
4821 if (ib_init_ah_from_mcmember(id_priv->id.device, id_priv->id.port_num,
4822 &multicast->rec, ndev, gid_type,
4823 &event->param.ud.ah_attr)) {
4824 event->event = RDMA_CM_EVENT_MULTICAST_ERROR;
4825 goto out;
4826 }
4827
4828 event->param.ud.qp_num = 0xFFFFFF;
4829 event->param.ud.qkey = id_priv->qkey;
4830
4831out:
4832 dev_put(ndev);
4833}
4834
4835static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
4836{
4837 struct cma_multicast *mc = multicast->context;
4838 struct rdma_id_private *id_priv = mc->id_priv;
4839 struct rdma_cm_event event = {};
4840 int ret = 0;
4841
4842 mutex_lock(&id_priv->handler_mutex);
4843 if (READ_ONCE(id_priv->state) == RDMA_CM_DEVICE_REMOVAL ||
4844 READ_ONCE(id_priv->state) == RDMA_CM_DESTROYING)
4845 goto out;
4846
4847 ret = cma_set_qkey(id_priv, be32_to_cpu(multicast->rec.qkey));
4848 if (!ret) {
4849 cma_make_mc_event(status, id_priv, multicast, &event, mc);
4850 ret = cma_cm_event_handler(id_priv, &event);
4851 }
4852 rdma_destroy_ah_attr(&event.param.ud.ah_attr);
4853 WARN_ON(ret);
4854
4855out:
4856 mutex_unlock(&id_priv->handler_mutex);
4857 return 0;
4858}
4859
4860static void cma_set_mgid(struct rdma_id_private *id_priv,
4861 struct sockaddr *addr, union ib_gid *mgid)
4862{
4863 unsigned char mc_map[MAX_ADDR_LEN];
4864 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4865 struct sockaddr_in *sin = (struct sockaddr_in *) addr;
4866 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
4867
4868 if (cma_any_addr(addr)) {
4869 memset(mgid, 0, sizeof *mgid);
4870 } else if ((addr->sa_family == AF_INET6) &&
4871 ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
4872 0xFF10A01B)) {
4873 /* IPv6 address is an SA assigned MGID. */
4874 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
4875 } else if (addr->sa_family == AF_IB) {
4876 memcpy(mgid, &((struct sockaddr_ib *) addr)->sib_addr, sizeof *mgid);
4877 } else if (addr->sa_family == AF_INET6) {
4878 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
4879 if (id_priv->id.ps == RDMA_PS_UDP)
4880 mc_map[7] = 0x01; /* Use RDMA CM signature */
4881 *mgid = *(union ib_gid *) (mc_map + 4);
4882 } else {
4883 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
4884 if (id_priv->id.ps == RDMA_PS_UDP)
4885 mc_map[7] = 0x01; /* Use RDMA CM signature */
4886 *mgid = *(union ib_gid *) (mc_map + 4);
4887 }
4888}
4889
4890static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
4891 struct cma_multicast *mc)
4892{
4893 struct ib_sa_mcmember_rec rec;
4894 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4895 ib_sa_comp_mask comp_mask;
4896 int ret;
4897
4898 ib_addr_get_mgid(dev_addr, &rec.mgid);
4899 ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
4900 &rec.mgid, &rec);
4901 if (ret)
4902 return ret;
4903
4904 if (!id_priv->qkey) {
4905 ret = cma_set_default_qkey(id_priv);
4906 if (ret)
4907 return ret;
4908 }
4909
4910 cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
4911 rec.qkey = cpu_to_be32(id_priv->qkey);
4912 rdma_addr_get_sgid(dev_addr, &rec.port_gid);
4913 rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
4914 rec.join_state = mc->join_state;
4915
4916 comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
4917 IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
4918 IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
4919 IB_SA_MCMEMBER_REC_FLOW_LABEL |
4920 IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
4921
4922 if (id_priv->id.ps == RDMA_PS_IPOIB)
4923 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
4924 IB_SA_MCMEMBER_REC_RATE_SELECTOR |
4925 IB_SA_MCMEMBER_REC_MTU_SELECTOR |
4926 IB_SA_MCMEMBER_REC_MTU |
4927 IB_SA_MCMEMBER_REC_HOP_LIMIT;
4928
4929 mc->sa_mc = ib_sa_join_multicast(&sa_client, id_priv->id.device,
4930 id_priv->id.port_num, &rec, comp_mask,
4931 GFP_KERNEL, cma_ib_mc_handler, mc);
4932 return PTR_ERR_OR_ZERO(mc->sa_mc);
4933}
4934
4935static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid,
4936 enum ib_gid_type gid_type)
4937{
4938 struct sockaddr_in *sin = (struct sockaddr_in *)addr;
4939 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
4940
4941 if (cma_any_addr(addr)) {
4942 memset(mgid, 0, sizeof *mgid);
4943 } else if (addr->sa_family == AF_INET6) {
4944 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
4945 } else {
4946 mgid->raw[0] =
4947 (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ? 0 : 0xff;
4948 mgid->raw[1] =
4949 (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) ? 0 : 0x0e;
4950 mgid->raw[2] = 0;
4951 mgid->raw[3] = 0;
4952 mgid->raw[4] = 0;
4953 mgid->raw[5] = 0;
4954 mgid->raw[6] = 0;
4955 mgid->raw[7] = 0;
4956 mgid->raw[8] = 0;
4957 mgid->raw[9] = 0;
4958 mgid->raw[10] = 0xff;
4959 mgid->raw[11] = 0xff;
4960 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
4961 }
4962}
4963
4964static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
4965 struct cma_multicast *mc)
4966{
4967 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
4968 int err = 0;
4969 struct sockaddr *addr = (struct sockaddr *)&mc->addr;
4970 struct net_device *ndev = NULL;
4971 struct ib_sa_multicast ib = {};
4972 enum ib_gid_type gid_type;
4973 bool send_only;
4974
4975 send_only = mc->join_state == BIT(SENDONLY_FULLMEMBER_JOIN);
4976
4977 if (cma_zero_addr(addr))
4978 return -EINVAL;
4979
4980 gid_type = id_priv->cma_dev->default_gid_type[id_priv->id.port_num -
4981 rdma_start_port(id_priv->cma_dev->device)];
4982 cma_iboe_set_mgid(addr, &ib.rec.mgid, gid_type);
4983
4984 ib.rec.pkey = cpu_to_be16(0xffff);
4985 if (dev_addr->bound_dev_if)
4986 ndev = dev_get_by_index(dev_addr->net, dev_addr->bound_dev_if);
4987 if (!ndev)
4988 return -ENODEV;
4989
4990 ib.rec.rate = IB_RATE_PORT_CURRENT;
4991 ib.rec.hop_limit = 1;
4992 ib.rec.mtu = iboe_get_mtu(ndev->mtu);
4993
4994 if (addr->sa_family == AF_INET) {
4995 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP) {
4996 ib.rec.hop_limit = IPV6_DEFAULT_HOPLIMIT;
4997 if (!send_only) {
4998 err = cma_igmp_send(ndev, &ib.rec.mgid,
4999 true);
5000 }
5001 }
5002 } else {
5003 if (gid_type == IB_GID_TYPE_ROCE_UDP_ENCAP)
5004 err = -ENOTSUPP;
5005 }
5006 dev_put(ndev);
5007 if (err || !ib.rec.mtu)
5008 return err ?: -EINVAL;
5009
5010 if (!id_priv->qkey)
5011 cma_set_default_qkey(id_priv);
5012
5013 rdma_ip2gid((struct sockaddr *)&id_priv->id.route.addr.src_addr,
5014 &ib.rec.port_gid);
5015 INIT_WORK(&mc->iboe_join.work, cma_iboe_join_work_handler);
5016 cma_make_mc_event(0, id_priv, &ib, &mc->iboe_join.event, mc);
5017 queue_work(cma_wq, &mc->iboe_join.work);
5018 return 0;
5019}
5020
5021int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
5022 u8 join_state, void *context)
5023{
5024 struct rdma_id_private *id_priv =
5025 container_of(id, struct rdma_id_private, id);
5026 struct cma_multicast *mc;
5027 int ret;
5028
5029 /* Not supported for kernel QPs */
5030 if (WARN_ON(id->qp))
5031 return -EINVAL;
5032
5033 /* ULP is calling this wrong. */
5034 if (!id->device || (READ_ONCE(id_priv->state) != RDMA_CM_ADDR_BOUND &&
5035 READ_ONCE(id_priv->state) != RDMA_CM_ADDR_RESOLVED))
5036 return -EINVAL;
5037
5038 if (id_priv->id.qp_type != IB_QPT_UD)
5039 return -EINVAL;
5040
5041 mc = kzalloc(sizeof(*mc), GFP_KERNEL);
5042 if (!mc)
5043 return -ENOMEM;
5044
5045 memcpy(&mc->addr, addr, rdma_addr_size(addr));
5046 mc->context = context;
5047 mc->id_priv = id_priv;
5048 mc->join_state = join_state;
5049
5050 if (rdma_protocol_roce(id->device, id->port_num)) {
5051 ret = cma_iboe_join_multicast(id_priv, mc);
5052 if (ret)
5053 goto out_err;
5054 } else if (rdma_cap_ib_mcast(id->device, id->port_num)) {
5055 ret = cma_join_ib_multicast(id_priv, mc);
5056 if (ret)
5057 goto out_err;
5058 } else {
5059 ret = -ENOSYS;
5060 goto out_err;
5061 }
5062
5063 spin_lock(&id_priv->lock);
5064 list_add(&mc->list, &id_priv->mc_list);
5065 spin_unlock(&id_priv->lock);
5066
5067 return 0;
5068out_err:
5069 kfree(mc);
5070 return ret;
5071}
5072EXPORT_SYMBOL(rdma_join_multicast);
5073
5074void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
5075{
5076 struct rdma_id_private *id_priv;
5077 struct cma_multicast *mc;
5078
5079 id_priv = container_of(id, struct rdma_id_private, id);
5080 spin_lock_irq(&id_priv->lock);
5081 list_for_each_entry(mc, &id_priv->mc_list, list) {
5082 if (memcmp(&mc->addr, addr, rdma_addr_size(addr)) != 0)
5083 continue;
5084 list_del(&mc->list);
5085 spin_unlock_irq(&id_priv->lock);
5086
5087 WARN_ON(id_priv->cma_dev->device != id->device);
5088 destroy_mc(id_priv, mc);
5089 return;
5090 }
5091 spin_unlock_irq(&id_priv->lock);
5092}
5093EXPORT_SYMBOL(rdma_leave_multicast);
5094
5095static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
5096{
5097 struct rdma_dev_addr *dev_addr;
5098 struct cma_work *work;
5099
5100 dev_addr = &id_priv->id.route.addr.dev_addr;
5101
5102 if ((dev_addr->bound_dev_if == ndev->ifindex) &&
5103 (net_eq(dev_net(ndev), dev_addr->net)) &&
5104 memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
5105 pr_info("RDMA CM addr change for ndev %s used by id %p\n",
5106 ndev->name, &id_priv->id);
5107 work = kzalloc(sizeof *work, GFP_KERNEL);
5108 if (!work)
5109 return -ENOMEM;
5110
5111 INIT_WORK(&work->work, cma_work_handler);
5112 work->id = id_priv;
5113 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
5114 cma_id_get(id_priv);
5115 queue_work(cma_wq, &work->work);
5116 }
5117
5118 return 0;
5119}
5120
5121static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
5122 void *ptr)
5123{
5124 struct net_device *ndev = netdev_notifier_info_to_dev(ptr);
5125 struct cma_device *cma_dev;
5126 struct rdma_id_private *id_priv;
5127 int ret = NOTIFY_DONE;
5128
5129 if (event != NETDEV_BONDING_FAILOVER)
5130 return NOTIFY_DONE;
5131
5132 if (!netif_is_bond_master(ndev))
5133 return NOTIFY_DONE;
5134
5135 mutex_lock(&lock);
5136 list_for_each_entry(cma_dev, &dev_list, list)
5137 list_for_each_entry(id_priv, &cma_dev->id_list, device_item) {
5138 ret = cma_netdev_change(ndev, id_priv);
5139 if (ret)
5140 goto out;
5141 }
5142
5143out:
5144 mutex_unlock(&lock);
5145 return ret;
5146}
5147
5148static void cma_netevent_work_handler(struct work_struct *_work)
5149{
5150 struct rdma_id_private *id_priv =
5151 container_of(_work, struct rdma_id_private, id.net_work);
5152 struct rdma_cm_event event = {};
5153
5154 mutex_lock(&id_priv->handler_mutex);
5155
5156 if (READ_ONCE(id_priv->state) == RDMA_CM_DESTROYING ||
5157 READ_ONCE(id_priv->state) == RDMA_CM_DEVICE_REMOVAL)
5158 goto out_unlock;
5159
5160 event.event = RDMA_CM_EVENT_UNREACHABLE;
5161 event.status = -ETIMEDOUT;
5162
5163 if (cma_cm_event_handler(id_priv, &event)) {
5164 __acquire(&id_priv->handler_mutex);
5165 id_priv->cm_id.ib = NULL;
5166 cma_id_put(id_priv);
5167 destroy_id_handler_unlock(id_priv);
5168 return;
5169 }
5170
5171out_unlock:
5172 mutex_unlock(&id_priv->handler_mutex);
5173 cma_id_put(id_priv);
5174}
5175
5176static int cma_netevent_callback(struct notifier_block *self,
5177 unsigned long event, void *ctx)
5178{
5179 struct id_table_entry *ips_node = NULL;
5180 struct rdma_id_private *current_id;
5181 struct neighbour *neigh = ctx;
5182 unsigned long flags;
5183
5184 if (event != NETEVENT_NEIGH_UPDATE)
5185 return NOTIFY_DONE;
5186
5187 spin_lock_irqsave(&id_table_lock, flags);
5188 if (neigh->tbl->family == AF_INET6) {
5189 struct sockaddr_in6 neigh_sock_6;
5190
5191 neigh_sock_6.sin6_family = AF_INET6;
5192 neigh_sock_6.sin6_addr = *(struct in6_addr *)neigh->primary_key;
5193 ips_node = node_from_ndev_ip(&id_table, neigh->dev->ifindex,
5194 (struct sockaddr *)&neigh_sock_6);
5195 } else if (neigh->tbl->family == AF_INET) {
5196 struct sockaddr_in neigh_sock_4;
5197
5198 neigh_sock_4.sin_family = AF_INET;
5199 neigh_sock_4.sin_addr.s_addr = *(__be32 *)(neigh->primary_key);
5200 ips_node = node_from_ndev_ip(&id_table, neigh->dev->ifindex,
5201 (struct sockaddr *)&neigh_sock_4);
5202 } else
5203 goto out;
5204
5205 if (!ips_node)
5206 goto out;
5207
5208 list_for_each_entry(current_id, &ips_node->id_list, id_list_entry) {
5209 if (!memcmp(current_id->id.route.addr.dev_addr.dst_dev_addr,
5210 neigh->ha, ETH_ALEN))
5211 continue;
5212 INIT_WORK(¤t_id->id.net_work, cma_netevent_work_handler);
5213 cma_id_get(current_id);
5214 queue_work(cma_wq, ¤t_id->id.net_work);
5215 }
5216out:
5217 spin_unlock_irqrestore(&id_table_lock, flags);
5218 return NOTIFY_DONE;
5219}
5220
5221static struct notifier_block cma_nb = {
5222 .notifier_call = cma_netdev_callback
5223};
5224
5225static struct notifier_block cma_netevent_cb = {
5226 .notifier_call = cma_netevent_callback
5227};
5228
5229static void cma_send_device_removal_put(struct rdma_id_private *id_priv)
5230{
5231 struct rdma_cm_event event = { .event = RDMA_CM_EVENT_DEVICE_REMOVAL };
5232 enum rdma_cm_state state;
5233 unsigned long flags;
5234
5235 mutex_lock(&id_priv->handler_mutex);
5236 /* Record that we want to remove the device */
5237 spin_lock_irqsave(&id_priv->lock, flags);
5238 state = id_priv->state;
5239 if (state == RDMA_CM_DESTROYING || state == RDMA_CM_DEVICE_REMOVAL) {
5240 spin_unlock_irqrestore(&id_priv->lock, flags);
5241 mutex_unlock(&id_priv->handler_mutex);
5242 cma_id_put(id_priv);
5243 return;
5244 }
5245 id_priv->state = RDMA_CM_DEVICE_REMOVAL;
5246 spin_unlock_irqrestore(&id_priv->lock, flags);
5247
5248 if (cma_cm_event_handler(id_priv, &event)) {
5249 /*
5250 * At this point the ULP promises it won't call
5251 * rdma_destroy_id() concurrently
5252 */
5253 cma_id_put(id_priv);
5254 mutex_unlock(&id_priv->handler_mutex);
5255 trace_cm_id_destroy(id_priv);
5256 _destroy_id(id_priv, state);
5257 return;
5258 }
5259 mutex_unlock(&id_priv->handler_mutex);
5260
5261 /*
5262 * If this races with destroy then the thread that first assigns state
5263 * to a destroying does the cancel.
5264 */
5265 cma_cancel_operation(id_priv, state);
5266 cma_id_put(id_priv);
5267}
5268
5269static void cma_process_remove(struct cma_device *cma_dev)
5270{
5271 mutex_lock(&lock);
5272 while (!list_empty(&cma_dev->id_list)) {
5273 struct rdma_id_private *id_priv = list_first_entry(
5274 &cma_dev->id_list, struct rdma_id_private, device_item);
5275
5276 list_del_init(&id_priv->listen_item);
5277 list_del_init(&id_priv->device_item);
5278 cma_id_get(id_priv);
5279 mutex_unlock(&lock);
5280
5281 cma_send_device_removal_put(id_priv);
5282
5283 mutex_lock(&lock);
5284 }
5285 mutex_unlock(&lock);
5286
5287 cma_dev_put(cma_dev);
5288 wait_for_completion(&cma_dev->comp);
5289}
5290
5291static bool cma_supported(struct ib_device *device)
5292{
5293 u32 i;
5294
5295 rdma_for_each_port(device, i) {
5296 if (rdma_cap_ib_cm(device, i) || rdma_cap_iw_cm(device, i))
5297 return true;
5298 }
5299 return false;
5300}
5301
5302static int cma_add_one(struct ib_device *device)
5303{
5304 struct rdma_id_private *to_destroy;
5305 struct cma_device *cma_dev;
5306 struct rdma_id_private *id_priv;
5307 unsigned long supported_gids = 0;
5308 int ret;
5309 u32 i;
5310
5311 if (!cma_supported(device))
5312 return -EOPNOTSUPP;
5313
5314 cma_dev = kmalloc(sizeof(*cma_dev), GFP_KERNEL);
5315 if (!cma_dev)
5316 return -ENOMEM;
5317
5318 cma_dev->device = device;
5319 cma_dev->default_gid_type = kcalloc(device->phys_port_cnt,
5320 sizeof(*cma_dev->default_gid_type),
5321 GFP_KERNEL);
5322 if (!cma_dev->default_gid_type) {
5323 ret = -ENOMEM;
5324 goto free_cma_dev;
5325 }
5326
5327 cma_dev->default_roce_tos = kcalloc(device->phys_port_cnt,
5328 sizeof(*cma_dev->default_roce_tos),
5329 GFP_KERNEL);
5330 if (!cma_dev->default_roce_tos) {
5331 ret = -ENOMEM;
5332 goto free_gid_type;
5333 }
5334
5335 rdma_for_each_port (device, i) {
5336 supported_gids = roce_gid_type_mask_support(device, i);
5337 WARN_ON(!supported_gids);
5338 if (supported_gids & (1 << CMA_PREFERRED_ROCE_GID_TYPE))
5339 cma_dev->default_gid_type[i - rdma_start_port(device)] =
5340 CMA_PREFERRED_ROCE_GID_TYPE;
5341 else
5342 cma_dev->default_gid_type[i - rdma_start_port(device)] =
5343 find_first_bit(&supported_gids, BITS_PER_LONG);
5344 cma_dev->default_roce_tos[i - rdma_start_port(device)] = 0;
5345 }
5346
5347 init_completion(&cma_dev->comp);
5348 refcount_set(&cma_dev->refcount, 1);
5349 INIT_LIST_HEAD(&cma_dev->id_list);
5350 ib_set_client_data(device, &cma_client, cma_dev);
5351
5352 mutex_lock(&lock);
5353 list_add_tail(&cma_dev->list, &dev_list);
5354 list_for_each_entry(id_priv, &listen_any_list, listen_any_item) {
5355 ret = cma_listen_on_dev(id_priv, cma_dev, &to_destroy);
5356 if (ret)
5357 goto free_listen;
5358 }
5359 mutex_unlock(&lock);
5360
5361 trace_cm_add_one(device);
5362 return 0;
5363
5364free_listen:
5365 list_del(&cma_dev->list);
5366 mutex_unlock(&lock);
5367
5368 /* cma_process_remove() will delete to_destroy */
5369 cma_process_remove(cma_dev);
5370 kfree(cma_dev->default_roce_tos);
5371free_gid_type:
5372 kfree(cma_dev->default_gid_type);
5373
5374free_cma_dev:
5375 kfree(cma_dev);
5376 return ret;
5377}
5378
5379static void cma_remove_one(struct ib_device *device, void *client_data)
5380{
5381 struct cma_device *cma_dev = client_data;
5382
5383 trace_cm_remove_one(device);
5384
5385 mutex_lock(&lock);
5386 list_del(&cma_dev->list);
5387 mutex_unlock(&lock);
5388
5389 cma_process_remove(cma_dev);
5390 kfree(cma_dev->default_roce_tos);
5391 kfree(cma_dev->default_gid_type);
5392 kfree(cma_dev);
5393}
5394
5395static int cma_init_net(struct net *net)
5396{
5397 struct cma_pernet *pernet = cma_pernet(net);
5398
5399 xa_init(&pernet->tcp_ps);
5400 xa_init(&pernet->udp_ps);
5401 xa_init(&pernet->ipoib_ps);
5402 xa_init(&pernet->ib_ps);
5403
5404 return 0;
5405}
5406
5407static void cma_exit_net(struct net *net)
5408{
5409 struct cma_pernet *pernet = cma_pernet(net);
5410
5411 WARN_ON(!xa_empty(&pernet->tcp_ps));
5412 WARN_ON(!xa_empty(&pernet->udp_ps));
5413 WARN_ON(!xa_empty(&pernet->ipoib_ps));
5414 WARN_ON(!xa_empty(&pernet->ib_ps));
5415}
5416
5417static struct pernet_operations cma_pernet_operations = {
5418 .init = cma_init_net,
5419 .exit = cma_exit_net,
5420 .id = &cma_pernet_id,
5421 .size = sizeof(struct cma_pernet),
5422};
5423
5424static int __init cma_init(void)
5425{
5426 int ret;
5427
5428 /*
5429 * There is a rare lock ordering dependency in cma_netdev_callback()
5430 * that only happens when bonding is enabled. Teach lockdep that rtnl
5431 * must never be nested under lock so it can find these without having
5432 * to test with bonding.
5433 */
5434 if (IS_ENABLED(CONFIG_LOCKDEP)) {
5435 rtnl_lock();
5436 mutex_lock(&lock);
5437 mutex_unlock(&lock);
5438 rtnl_unlock();
5439 }
5440
5441 cma_wq = alloc_ordered_workqueue("rdma_cm", WQ_MEM_RECLAIM);
5442 if (!cma_wq)
5443 return -ENOMEM;
5444
5445 ret = register_pernet_subsys(&cma_pernet_operations);
5446 if (ret)
5447 goto err_wq;
5448
5449 ib_sa_register_client(&sa_client);
5450 register_netdevice_notifier(&cma_nb);
5451 register_netevent_notifier(&cma_netevent_cb);
5452
5453 ret = ib_register_client(&cma_client);
5454 if (ret)
5455 goto err;
5456
5457 ret = cma_configfs_init();
5458 if (ret)
5459 goto err_ib;
5460
5461 return 0;
5462
5463err_ib:
5464 ib_unregister_client(&cma_client);
5465err:
5466 unregister_netevent_notifier(&cma_netevent_cb);
5467 unregister_netdevice_notifier(&cma_nb);
5468 ib_sa_unregister_client(&sa_client);
5469 unregister_pernet_subsys(&cma_pernet_operations);
5470err_wq:
5471 destroy_workqueue(cma_wq);
5472 return ret;
5473}
5474
5475static void __exit cma_cleanup(void)
5476{
5477 cma_configfs_exit();
5478 ib_unregister_client(&cma_client);
5479 unregister_netevent_notifier(&cma_netevent_cb);
5480 unregister_netdevice_notifier(&cma_nb);
5481 ib_sa_unregister_client(&sa_client);
5482 unregister_pernet_subsys(&cma_pernet_operations);
5483 destroy_workqueue(cma_wq);
5484}
5485
5486module_init(cma_init);
5487module_exit(cma_cleanup);
1/*
2 * Copyright (c) 2005 Voltaire Inc. All rights reserved.
3 * Copyright (c) 2002-2005, Network Appliance, Inc. All rights reserved.
4 * Copyright (c) 1999-2005, Mellanox Technologies, Inc. All rights reserved.
5 * Copyright (c) 2005-2006 Intel Corporation. All rights reserved.
6 *
7 * This software is available to you under a choice of one of two
8 * licenses. You may choose to be licensed under the terms of the GNU
9 * General Public License (GPL) Version 2, available from the file
10 * COPYING in the main directory of this source tree, or the
11 * OpenIB.org BSD license below:
12 *
13 * Redistribution and use in source and binary forms, with or
14 * without modification, are permitted provided that the following
15 * conditions are met:
16 *
17 * - Redistributions of source code must retain the above
18 * copyright notice, this list of conditions and the following
19 * disclaimer.
20 *
21 * - Redistributions in binary form must reproduce the above
22 * copyright notice, this list of conditions and the following
23 * disclaimer in the documentation and/or other materials
24 * provided with the distribution.
25 *
26 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
27 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
28 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
29 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
30 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
31 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
32 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
33 * SOFTWARE.
34 */
35
36#include <linux/completion.h>
37#include <linux/in.h>
38#include <linux/in6.h>
39#include <linux/mutex.h>
40#include <linux/random.h>
41#include <linux/idr.h>
42#include <linux/inetdevice.h>
43#include <linux/slab.h>
44#include <linux/module.h>
45#include <net/route.h>
46
47#include <net/tcp.h>
48#include <net/ipv6.h>
49
50#include <rdma/rdma_cm.h>
51#include <rdma/rdma_cm_ib.h>
52#include <rdma/rdma_netlink.h>
53#include <rdma/ib_cache.h>
54#include <rdma/ib_cm.h>
55#include <rdma/ib_sa.h>
56#include <rdma/iw_cm.h>
57
58MODULE_AUTHOR("Sean Hefty");
59MODULE_DESCRIPTION("Generic RDMA CM Agent");
60MODULE_LICENSE("Dual BSD/GPL");
61
62#define CMA_CM_RESPONSE_TIMEOUT 20
63#define CMA_MAX_CM_RETRIES 15
64#define CMA_CM_MRA_SETTING (IB_CM_MRA_FLAG_DELAY | 24)
65#define CMA_IBOE_PACKET_LIFETIME 18
66
67static void cma_add_one(struct ib_device *device);
68static void cma_remove_one(struct ib_device *device);
69
70static struct ib_client cma_client = {
71 .name = "cma",
72 .add = cma_add_one,
73 .remove = cma_remove_one
74};
75
76static struct ib_sa_client sa_client;
77static struct rdma_addr_client addr_client;
78static LIST_HEAD(dev_list);
79static LIST_HEAD(listen_any_list);
80static DEFINE_MUTEX(lock);
81static struct workqueue_struct *cma_wq;
82static DEFINE_IDR(sdp_ps);
83static DEFINE_IDR(tcp_ps);
84static DEFINE_IDR(udp_ps);
85static DEFINE_IDR(ipoib_ps);
86static DEFINE_IDR(ib_ps);
87
88struct cma_device {
89 struct list_head list;
90 struct ib_device *device;
91 struct completion comp;
92 atomic_t refcount;
93 struct list_head id_list;
94};
95
96struct rdma_bind_list {
97 struct idr *ps;
98 struct hlist_head owners;
99 unsigned short port;
100};
101
102/*
103 * Device removal can occur at anytime, so we need extra handling to
104 * serialize notifying the user of device removal with other callbacks.
105 * We do this by disabling removal notification while a callback is in process,
106 * and reporting it after the callback completes.
107 */
108struct rdma_id_private {
109 struct rdma_cm_id id;
110
111 struct rdma_bind_list *bind_list;
112 struct hlist_node node;
113 struct list_head list; /* listen_any_list or cma_device.list */
114 struct list_head listen_list; /* per device listens */
115 struct cma_device *cma_dev;
116 struct list_head mc_list;
117
118 int internal_id;
119 enum rdma_cm_state state;
120 spinlock_t lock;
121 struct mutex qp_mutex;
122
123 struct completion comp;
124 atomic_t refcount;
125 struct mutex handler_mutex;
126
127 int backlog;
128 int timeout_ms;
129 struct ib_sa_query *query;
130 int query_id;
131 union {
132 struct ib_cm_id *ib;
133 struct iw_cm_id *iw;
134 } cm_id;
135
136 u32 seq_num;
137 u32 qkey;
138 u32 qp_num;
139 pid_t owner;
140 u8 srq;
141 u8 tos;
142 u8 reuseaddr;
143};
144
145struct cma_multicast {
146 struct rdma_id_private *id_priv;
147 union {
148 struct ib_sa_multicast *ib;
149 } multicast;
150 struct list_head list;
151 void *context;
152 struct sockaddr_storage addr;
153 struct kref mcref;
154};
155
156struct cma_work {
157 struct work_struct work;
158 struct rdma_id_private *id;
159 enum rdma_cm_state old_state;
160 enum rdma_cm_state new_state;
161 struct rdma_cm_event event;
162};
163
164struct cma_ndev_work {
165 struct work_struct work;
166 struct rdma_id_private *id;
167 struct rdma_cm_event event;
168};
169
170struct iboe_mcast_work {
171 struct work_struct work;
172 struct rdma_id_private *id;
173 struct cma_multicast *mc;
174};
175
176union cma_ip_addr {
177 struct in6_addr ip6;
178 struct {
179 __be32 pad[3];
180 __be32 addr;
181 } ip4;
182};
183
184struct cma_hdr {
185 u8 cma_version;
186 u8 ip_version; /* IP version: 7:4 */
187 __be16 port;
188 union cma_ip_addr src_addr;
189 union cma_ip_addr dst_addr;
190};
191
192struct sdp_hh {
193 u8 bsdh[16];
194 u8 sdp_version; /* Major version: 7:4 */
195 u8 ip_version; /* IP version: 7:4 */
196 u8 sdp_specific1[10];
197 __be16 port;
198 __be16 sdp_specific2;
199 union cma_ip_addr src_addr;
200 union cma_ip_addr dst_addr;
201};
202
203struct sdp_hah {
204 u8 bsdh[16];
205 u8 sdp_version;
206};
207
208#define CMA_VERSION 0x00
209#define SDP_MAJ_VERSION 0x2
210
211static int cma_comp(struct rdma_id_private *id_priv, enum rdma_cm_state comp)
212{
213 unsigned long flags;
214 int ret;
215
216 spin_lock_irqsave(&id_priv->lock, flags);
217 ret = (id_priv->state == comp);
218 spin_unlock_irqrestore(&id_priv->lock, flags);
219 return ret;
220}
221
222static int cma_comp_exch(struct rdma_id_private *id_priv,
223 enum rdma_cm_state comp, enum rdma_cm_state exch)
224{
225 unsigned long flags;
226 int ret;
227
228 spin_lock_irqsave(&id_priv->lock, flags);
229 if ((ret = (id_priv->state == comp)))
230 id_priv->state = exch;
231 spin_unlock_irqrestore(&id_priv->lock, flags);
232 return ret;
233}
234
235static enum rdma_cm_state cma_exch(struct rdma_id_private *id_priv,
236 enum rdma_cm_state exch)
237{
238 unsigned long flags;
239 enum rdma_cm_state old;
240
241 spin_lock_irqsave(&id_priv->lock, flags);
242 old = id_priv->state;
243 id_priv->state = exch;
244 spin_unlock_irqrestore(&id_priv->lock, flags);
245 return old;
246}
247
248static inline u8 cma_get_ip_ver(struct cma_hdr *hdr)
249{
250 return hdr->ip_version >> 4;
251}
252
253static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
254{
255 hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
256}
257
258static inline u8 sdp_get_majv(u8 sdp_version)
259{
260 return sdp_version >> 4;
261}
262
263static inline u8 sdp_get_ip_ver(struct sdp_hh *hh)
264{
265 return hh->ip_version >> 4;
266}
267
268static inline void sdp_set_ip_ver(struct sdp_hh *hh, u8 ip_ver)
269{
270 hh->ip_version = (ip_ver << 4) | (hh->ip_version & 0xF);
271}
272
273static void cma_attach_to_dev(struct rdma_id_private *id_priv,
274 struct cma_device *cma_dev)
275{
276 atomic_inc(&cma_dev->refcount);
277 id_priv->cma_dev = cma_dev;
278 id_priv->id.device = cma_dev->device;
279 id_priv->id.route.addr.dev_addr.transport =
280 rdma_node_get_transport(cma_dev->device->node_type);
281 list_add_tail(&id_priv->list, &cma_dev->id_list);
282}
283
284static inline void cma_deref_dev(struct cma_device *cma_dev)
285{
286 if (atomic_dec_and_test(&cma_dev->refcount))
287 complete(&cma_dev->comp);
288}
289
290static inline void release_mc(struct kref *kref)
291{
292 struct cma_multicast *mc = container_of(kref, struct cma_multicast, mcref);
293
294 kfree(mc->multicast.ib);
295 kfree(mc);
296}
297
298static void cma_release_dev(struct rdma_id_private *id_priv)
299{
300 mutex_lock(&lock);
301 list_del(&id_priv->list);
302 cma_deref_dev(id_priv->cma_dev);
303 id_priv->cma_dev = NULL;
304 mutex_unlock(&lock);
305}
306
307static int cma_set_qkey(struct rdma_id_private *id_priv)
308{
309 struct ib_sa_mcmember_rec rec;
310 int ret = 0;
311
312 if (id_priv->qkey)
313 return 0;
314
315 switch (id_priv->id.ps) {
316 case RDMA_PS_UDP:
317 id_priv->qkey = RDMA_UDP_QKEY;
318 break;
319 case RDMA_PS_IPOIB:
320 ib_addr_get_mgid(&id_priv->id.route.addr.dev_addr, &rec.mgid);
321 ret = ib_sa_get_mcmember_rec(id_priv->id.device,
322 id_priv->id.port_num, &rec.mgid,
323 &rec);
324 if (!ret)
325 id_priv->qkey = be32_to_cpu(rec.qkey);
326 break;
327 default:
328 break;
329 }
330 return ret;
331}
332
333static int find_gid_port(struct ib_device *device, union ib_gid *gid, u8 port_num)
334{
335 int i;
336 int err;
337 struct ib_port_attr props;
338 union ib_gid tmp;
339
340 err = ib_query_port(device, port_num, &props);
341 if (err)
342 return 1;
343
344 for (i = 0; i < props.gid_tbl_len; ++i) {
345 err = ib_query_gid(device, port_num, i, &tmp);
346 if (err)
347 return 1;
348 if (!memcmp(&tmp, gid, sizeof tmp))
349 return 0;
350 }
351
352 return -EAGAIN;
353}
354
355static int cma_acquire_dev(struct rdma_id_private *id_priv)
356{
357 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
358 struct cma_device *cma_dev;
359 union ib_gid gid, iboe_gid;
360 int ret = -ENODEV;
361 u8 port;
362 enum rdma_link_layer dev_ll = dev_addr->dev_type == ARPHRD_INFINIBAND ?
363 IB_LINK_LAYER_INFINIBAND : IB_LINK_LAYER_ETHERNET;
364
365 if (dev_ll != IB_LINK_LAYER_INFINIBAND &&
366 id_priv->id.ps == RDMA_PS_IPOIB)
367 return -EINVAL;
368
369 mutex_lock(&lock);
370 iboe_addr_get_sgid(dev_addr, &iboe_gid);
371 memcpy(&gid, dev_addr->src_dev_addr +
372 rdma_addr_gid_offset(dev_addr), sizeof gid);
373 list_for_each_entry(cma_dev, &dev_list, list) {
374 for (port = 1; port <= cma_dev->device->phys_port_cnt; ++port) {
375 if (rdma_port_get_link_layer(cma_dev->device, port) == dev_ll) {
376 if (rdma_node_get_transport(cma_dev->device->node_type) == RDMA_TRANSPORT_IB &&
377 rdma_port_get_link_layer(cma_dev->device, port) == IB_LINK_LAYER_ETHERNET)
378 ret = find_gid_port(cma_dev->device, &iboe_gid, port);
379 else
380 ret = find_gid_port(cma_dev->device, &gid, port);
381
382 if (!ret) {
383 id_priv->id.port_num = port;
384 goto out;
385 } else if (ret == 1)
386 break;
387 }
388 }
389 }
390
391out:
392 if (!ret)
393 cma_attach_to_dev(id_priv, cma_dev);
394
395 mutex_unlock(&lock);
396 return ret;
397}
398
399static void cma_deref_id(struct rdma_id_private *id_priv)
400{
401 if (atomic_dec_and_test(&id_priv->refcount))
402 complete(&id_priv->comp);
403}
404
405static int cma_disable_callback(struct rdma_id_private *id_priv,
406 enum rdma_cm_state state)
407{
408 mutex_lock(&id_priv->handler_mutex);
409 if (id_priv->state != state) {
410 mutex_unlock(&id_priv->handler_mutex);
411 return -EINVAL;
412 }
413 return 0;
414}
415
416struct rdma_cm_id *rdma_create_id(rdma_cm_event_handler event_handler,
417 void *context, enum rdma_port_space ps,
418 enum ib_qp_type qp_type)
419{
420 struct rdma_id_private *id_priv;
421
422 id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
423 if (!id_priv)
424 return ERR_PTR(-ENOMEM);
425
426 id_priv->owner = task_pid_nr(current);
427 id_priv->state = RDMA_CM_IDLE;
428 id_priv->id.context = context;
429 id_priv->id.event_handler = event_handler;
430 id_priv->id.ps = ps;
431 id_priv->id.qp_type = qp_type;
432 spin_lock_init(&id_priv->lock);
433 mutex_init(&id_priv->qp_mutex);
434 init_completion(&id_priv->comp);
435 atomic_set(&id_priv->refcount, 1);
436 mutex_init(&id_priv->handler_mutex);
437 INIT_LIST_HEAD(&id_priv->listen_list);
438 INIT_LIST_HEAD(&id_priv->mc_list);
439 get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
440
441 return &id_priv->id;
442}
443EXPORT_SYMBOL(rdma_create_id);
444
445static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
446{
447 struct ib_qp_attr qp_attr;
448 int qp_attr_mask, ret;
449
450 qp_attr.qp_state = IB_QPS_INIT;
451 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
452 if (ret)
453 return ret;
454
455 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
456 if (ret)
457 return ret;
458
459 qp_attr.qp_state = IB_QPS_RTR;
460 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
461 if (ret)
462 return ret;
463
464 qp_attr.qp_state = IB_QPS_RTS;
465 qp_attr.sq_psn = 0;
466 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
467
468 return ret;
469}
470
471static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
472{
473 struct ib_qp_attr qp_attr;
474 int qp_attr_mask, ret;
475
476 qp_attr.qp_state = IB_QPS_INIT;
477 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
478 if (ret)
479 return ret;
480
481 return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
482}
483
484int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
485 struct ib_qp_init_attr *qp_init_attr)
486{
487 struct rdma_id_private *id_priv;
488 struct ib_qp *qp;
489 int ret;
490
491 id_priv = container_of(id, struct rdma_id_private, id);
492 if (id->device != pd->device)
493 return -EINVAL;
494
495 qp = ib_create_qp(pd, qp_init_attr);
496 if (IS_ERR(qp))
497 return PTR_ERR(qp);
498
499 if (id->qp_type == IB_QPT_UD)
500 ret = cma_init_ud_qp(id_priv, qp);
501 else
502 ret = cma_init_conn_qp(id_priv, qp);
503 if (ret)
504 goto err;
505
506 id->qp = qp;
507 id_priv->qp_num = qp->qp_num;
508 id_priv->srq = (qp->srq != NULL);
509 return 0;
510err:
511 ib_destroy_qp(qp);
512 return ret;
513}
514EXPORT_SYMBOL(rdma_create_qp);
515
516void rdma_destroy_qp(struct rdma_cm_id *id)
517{
518 struct rdma_id_private *id_priv;
519
520 id_priv = container_of(id, struct rdma_id_private, id);
521 mutex_lock(&id_priv->qp_mutex);
522 ib_destroy_qp(id_priv->id.qp);
523 id_priv->id.qp = NULL;
524 mutex_unlock(&id_priv->qp_mutex);
525}
526EXPORT_SYMBOL(rdma_destroy_qp);
527
528static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
529 struct rdma_conn_param *conn_param)
530{
531 struct ib_qp_attr qp_attr;
532 int qp_attr_mask, ret;
533
534 mutex_lock(&id_priv->qp_mutex);
535 if (!id_priv->id.qp) {
536 ret = 0;
537 goto out;
538 }
539
540 /* Need to update QP attributes from default values. */
541 qp_attr.qp_state = IB_QPS_INIT;
542 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
543 if (ret)
544 goto out;
545
546 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
547 if (ret)
548 goto out;
549
550 qp_attr.qp_state = IB_QPS_RTR;
551 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
552 if (ret)
553 goto out;
554
555 if (conn_param)
556 qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
557 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
558out:
559 mutex_unlock(&id_priv->qp_mutex);
560 return ret;
561}
562
563static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
564 struct rdma_conn_param *conn_param)
565{
566 struct ib_qp_attr qp_attr;
567 int qp_attr_mask, ret;
568
569 mutex_lock(&id_priv->qp_mutex);
570 if (!id_priv->id.qp) {
571 ret = 0;
572 goto out;
573 }
574
575 qp_attr.qp_state = IB_QPS_RTS;
576 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
577 if (ret)
578 goto out;
579
580 if (conn_param)
581 qp_attr.max_rd_atomic = conn_param->initiator_depth;
582 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
583out:
584 mutex_unlock(&id_priv->qp_mutex);
585 return ret;
586}
587
588static int cma_modify_qp_err(struct rdma_id_private *id_priv)
589{
590 struct ib_qp_attr qp_attr;
591 int ret;
592
593 mutex_lock(&id_priv->qp_mutex);
594 if (!id_priv->id.qp) {
595 ret = 0;
596 goto out;
597 }
598
599 qp_attr.qp_state = IB_QPS_ERR;
600 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
601out:
602 mutex_unlock(&id_priv->qp_mutex);
603 return ret;
604}
605
606static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
607 struct ib_qp_attr *qp_attr, int *qp_attr_mask)
608{
609 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
610 int ret;
611 u16 pkey;
612
613 if (rdma_port_get_link_layer(id_priv->id.device, id_priv->id.port_num) ==
614 IB_LINK_LAYER_INFINIBAND)
615 pkey = ib_addr_get_pkey(dev_addr);
616 else
617 pkey = 0xffff;
618
619 ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
620 pkey, &qp_attr->pkey_index);
621 if (ret)
622 return ret;
623
624 qp_attr->port_num = id_priv->id.port_num;
625 *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
626
627 if (id_priv->id.qp_type == IB_QPT_UD) {
628 ret = cma_set_qkey(id_priv);
629 if (ret)
630 return ret;
631
632 qp_attr->qkey = id_priv->qkey;
633 *qp_attr_mask |= IB_QP_QKEY;
634 } else {
635 qp_attr->qp_access_flags = 0;
636 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
637 }
638 return 0;
639}
640
641int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
642 int *qp_attr_mask)
643{
644 struct rdma_id_private *id_priv;
645 int ret = 0;
646
647 id_priv = container_of(id, struct rdma_id_private, id);
648 switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
649 case RDMA_TRANSPORT_IB:
650 if (!id_priv->cm_id.ib || (id_priv->id.qp_type == IB_QPT_UD))
651 ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
652 else
653 ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
654 qp_attr_mask);
655 if (qp_attr->qp_state == IB_QPS_RTR)
656 qp_attr->rq_psn = id_priv->seq_num;
657 break;
658 case RDMA_TRANSPORT_IWARP:
659 if (!id_priv->cm_id.iw) {
660 qp_attr->qp_access_flags = 0;
661 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
662 } else
663 ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
664 qp_attr_mask);
665 break;
666 default:
667 ret = -ENOSYS;
668 break;
669 }
670
671 return ret;
672}
673EXPORT_SYMBOL(rdma_init_qp_attr);
674
675static inline int cma_zero_addr(struct sockaddr *addr)
676{
677 struct in6_addr *ip6;
678
679 if (addr->sa_family == AF_INET)
680 return ipv4_is_zeronet(
681 ((struct sockaddr_in *)addr)->sin_addr.s_addr);
682 else {
683 ip6 = &((struct sockaddr_in6 *) addr)->sin6_addr;
684 return (ip6->s6_addr32[0] | ip6->s6_addr32[1] |
685 ip6->s6_addr32[2] | ip6->s6_addr32[3]) == 0;
686 }
687}
688
689static inline int cma_loopback_addr(struct sockaddr *addr)
690{
691 if (addr->sa_family == AF_INET)
692 return ipv4_is_loopback(
693 ((struct sockaddr_in *) addr)->sin_addr.s_addr);
694 else
695 return ipv6_addr_loopback(
696 &((struct sockaddr_in6 *) addr)->sin6_addr);
697}
698
699static inline int cma_any_addr(struct sockaddr *addr)
700{
701 return cma_zero_addr(addr) || cma_loopback_addr(addr);
702}
703
704static int cma_addr_cmp(struct sockaddr *src, struct sockaddr *dst)
705{
706 if (src->sa_family != dst->sa_family)
707 return -1;
708
709 switch (src->sa_family) {
710 case AF_INET:
711 return ((struct sockaddr_in *) src)->sin_addr.s_addr !=
712 ((struct sockaddr_in *) dst)->sin_addr.s_addr;
713 default:
714 return ipv6_addr_cmp(&((struct sockaddr_in6 *) src)->sin6_addr,
715 &((struct sockaddr_in6 *) dst)->sin6_addr);
716 }
717}
718
719static inline __be16 cma_port(struct sockaddr *addr)
720{
721 if (addr->sa_family == AF_INET)
722 return ((struct sockaddr_in *) addr)->sin_port;
723 else
724 return ((struct sockaddr_in6 *) addr)->sin6_port;
725}
726
727static inline int cma_any_port(struct sockaddr *addr)
728{
729 return !cma_port(addr);
730}
731
732static int cma_get_net_info(void *hdr, enum rdma_port_space ps,
733 u8 *ip_ver, __be16 *port,
734 union cma_ip_addr **src, union cma_ip_addr **dst)
735{
736 switch (ps) {
737 case RDMA_PS_SDP:
738 if (sdp_get_majv(((struct sdp_hh *) hdr)->sdp_version) !=
739 SDP_MAJ_VERSION)
740 return -EINVAL;
741
742 *ip_ver = sdp_get_ip_ver(hdr);
743 *port = ((struct sdp_hh *) hdr)->port;
744 *src = &((struct sdp_hh *) hdr)->src_addr;
745 *dst = &((struct sdp_hh *) hdr)->dst_addr;
746 break;
747 default:
748 if (((struct cma_hdr *) hdr)->cma_version != CMA_VERSION)
749 return -EINVAL;
750
751 *ip_ver = cma_get_ip_ver(hdr);
752 *port = ((struct cma_hdr *) hdr)->port;
753 *src = &((struct cma_hdr *) hdr)->src_addr;
754 *dst = &((struct cma_hdr *) hdr)->dst_addr;
755 break;
756 }
757
758 if (*ip_ver != 4 && *ip_ver != 6)
759 return -EINVAL;
760 return 0;
761}
762
763static void cma_save_net_info(struct rdma_addr *addr,
764 struct rdma_addr *listen_addr,
765 u8 ip_ver, __be16 port,
766 union cma_ip_addr *src, union cma_ip_addr *dst)
767{
768 struct sockaddr_in *listen4, *ip4;
769 struct sockaddr_in6 *listen6, *ip6;
770
771 switch (ip_ver) {
772 case 4:
773 listen4 = (struct sockaddr_in *) &listen_addr->src_addr;
774 ip4 = (struct sockaddr_in *) &addr->src_addr;
775 ip4->sin_family = listen4->sin_family;
776 ip4->sin_addr.s_addr = dst->ip4.addr;
777 ip4->sin_port = listen4->sin_port;
778
779 ip4 = (struct sockaddr_in *) &addr->dst_addr;
780 ip4->sin_family = listen4->sin_family;
781 ip4->sin_addr.s_addr = src->ip4.addr;
782 ip4->sin_port = port;
783 break;
784 case 6:
785 listen6 = (struct sockaddr_in6 *) &listen_addr->src_addr;
786 ip6 = (struct sockaddr_in6 *) &addr->src_addr;
787 ip6->sin6_family = listen6->sin6_family;
788 ip6->sin6_addr = dst->ip6;
789 ip6->sin6_port = listen6->sin6_port;
790
791 ip6 = (struct sockaddr_in6 *) &addr->dst_addr;
792 ip6->sin6_family = listen6->sin6_family;
793 ip6->sin6_addr = src->ip6;
794 ip6->sin6_port = port;
795 break;
796 default:
797 break;
798 }
799}
800
801static inline int cma_user_data_offset(enum rdma_port_space ps)
802{
803 switch (ps) {
804 case RDMA_PS_SDP:
805 return 0;
806 default:
807 return sizeof(struct cma_hdr);
808 }
809}
810
811static void cma_cancel_route(struct rdma_id_private *id_priv)
812{
813 switch (rdma_port_get_link_layer(id_priv->id.device, id_priv->id.port_num)) {
814 case IB_LINK_LAYER_INFINIBAND:
815 if (id_priv->query)
816 ib_sa_cancel_query(id_priv->query_id, id_priv->query);
817 break;
818 default:
819 break;
820 }
821}
822
823static void cma_cancel_listens(struct rdma_id_private *id_priv)
824{
825 struct rdma_id_private *dev_id_priv;
826
827 /*
828 * Remove from listen_any_list to prevent added devices from spawning
829 * additional listen requests.
830 */
831 mutex_lock(&lock);
832 list_del(&id_priv->list);
833
834 while (!list_empty(&id_priv->listen_list)) {
835 dev_id_priv = list_entry(id_priv->listen_list.next,
836 struct rdma_id_private, listen_list);
837 /* sync with device removal to avoid duplicate destruction */
838 list_del_init(&dev_id_priv->list);
839 list_del(&dev_id_priv->listen_list);
840 mutex_unlock(&lock);
841
842 rdma_destroy_id(&dev_id_priv->id);
843 mutex_lock(&lock);
844 }
845 mutex_unlock(&lock);
846}
847
848static void cma_cancel_operation(struct rdma_id_private *id_priv,
849 enum rdma_cm_state state)
850{
851 switch (state) {
852 case RDMA_CM_ADDR_QUERY:
853 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
854 break;
855 case RDMA_CM_ROUTE_QUERY:
856 cma_cancel_route(id_priv);
857 break;
858 case RDMA_CM_LISTEN:
859 if (cma_any_addr((struct sockaddr *) &id_priv->id.route.addr.src_addr)
860 && !id_priv->cma_dev)
861 cma_cancel_listens(id_priv);
862 break;
863 default:
864 break;
865 }
866}
867
868static void cma_release_port(struct rdma_id_private *id_priv)
869{
870 struct rdma_bind_list *bind_list = id_priv->bind_list;
871
872 if (!bind_list)
873 return;
874
875 mutex_lock(&lock);
876 hlist_del(&id_priv->node);
877 if (hlist_empty(&bind_list->owners)) {
878 idr_remove(bind_list->ps, bind_list->port);
879 kfree(bind_list);
880 }
881 mutex_unlock(&lock);
882}
883
884static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
885{
886 struct cma_multicast *mc;
887
888 while (!list_empty(&id_priv->mc_list)) {
889 mc = container_of(id_priv->mc_list.next,
890 struct cma_multicast, list);
891 list_del(&mc->list);
892 switch (rdma_port_get_link_layer(id_priv->cma_dev->device, id_priv->id.port_num)) {
893 case IB_LINK_LAYER_INFINIBAND:
894 ib_sa_free_multicast(mc->multicast.ib);
895 kfree(mc);
896 break;
897 case IB_LINK_LAYER_ETHERNET:
898 kref_put(&mc->mcref, release_mc);
899 break;
900 default:
901 break;
902 }
903 }
904}
905
906void rdma_destroy_id(struct rdma_cm_id *id)
907{
908 struct rdma_id_private *id_priv;
909 enum rdma_cm_state state;
910
911 id_priv = container_of(id, struct rdma_id_private, id);
912 state = cma_exch(id_priv, RDMA_CM_DESTROYING);
913 cma_cancel_operation(id_priv, state);
914
915 /*
916 * Wait for any active callback to finish. New callbacks will find
917 * the id_priv state set to destroying and abort.
918 */
919 mutex_lock(&id_priv->handler_mutex);
920 mutex_unlock(&id_priv->handler_mutex);
921
922 if (id_priv->cma_dev) {
923 switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
924 case RDMA_TRANSPORT_IB:
925 if (id_priv->cm_id.ib)
926 ib_destroy_cm_id(id_priv->cm_id.ib);
927 break;
928 case RDMA_TRANSPORT_IWARP:
929 if (id_priv->cm_id.iw)
930 iw_destroy_cm_id(id_priv->cm_id.iw);
931 break;
932 default:
933 break;
934 }
935 cma_leave_mc_groups(id_priv);
936 cma_release_dev(id_priv);
937 }
938
939 cma_release_port(id_priv);
940 cma_deref_id(id_priv);
941 wait_for_completion(&id_priv->comp);
942
943 if (id_priv->internal_id)
944 cma_deref_id(id_priv->id.context);
945
946 kfree(id_priv->id.route.path_rec);
947 kfree(id_priv);
948}
949EXPORT_SYMBOL(rdma_destroy_id);
950
951static int cma_rep_recv(struct rdma_id_private *id_priv)
952{
953 int ret;
954
955 ret = cma_modify_qp_rtr(id_priv, NULL);
956 if (ret)
957 goto reject;
958
959 ret = cma_modify_qp_rts(id_priv, NULL);
960 if (ret)
961 goto reject;
962
963 ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
964 if (ret)
965 goto reject;
966
967 return 0;
968reject:
969 cma_modify_qp_err(id_priv);
970 ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
971 NULL, 0, NULL, 0);
972 return ret;
973}
974
975static int cma_verify_rep(struct rdma_id_private *id_priv, void *data)
976{
977 if (id_priv->id.ps == RDMA_PS_SDP &&
978 sdp_get_majv(((struct sdp_hah *) data)->sdp_version) !=
979 SDP_MAJ_VERSION)
980 return -EINVAL;
981
982 return 0;
983}
984
985static void cma_set_rep_event_data(struct rdma_cm_event *event,
986 struct ib_cm_rep_event_param *rep_data,
987 void *private_data)
988{
989 event->param.conn.private_data = private_data;
990 event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
991 event->param.conn.responder_resources = rep_data->responder_resources;
992 event->param.conn.initiator_depth = rep_data->initiator_depth;
993 event->param.conn.flow_control = rep_data->flow_control;
994 event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
995 event->param.conn.srq = rep_data->srq;
996 event->param.conn.qp_num = rep_data->remote_qpn;
997}
998
999static int cma_ib_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1000{
1001 struct rdma_id_private *id_priv = cm_id->context;
1002 struct rdma_cm_event event;
1003 int ret = 0;
1004
1005 if ((ib_event->event != IB_CM_TIMEWAIT_EXIT &&
1006 cma_disable_callback(id_priv, RDMA_CM_CONNECT)) ||
1007 (ib_event->event == IB_CM_TIMEWAIT_EXIT &&
1008 cma_disable_callback(id_priv, RDMA_CM_DISCONNECT)))
1009 return 0;
1010
1011 memset(&event, 0, sizeof event);
1012 switch (ib_event->event) {
1013 case IB_CM_REQ_ERROR:
1014 case IB_CM_REP_ERROR:
1015 event.event = RDMA_CM_EVENT_UNREACHABLE;
1016 event.status = -ETIMEDOUT;
1017 break;
1018 case IB_CM_REP_RECEIVED:
1019 event.status = cma_verify_rep(id_priv, ib_event->private_data);
1020 if (event.status)
1021 event.event = RDMA_CM_EVENT_CONNECT_ERROR;
1022 else if (id_priv->id.qp && id_priv->id.ps != RDMA_PS_SDP) {
1023 event.status = cma_rep_recv(id_priv);
1024 event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
1025 RDMA_CM_EVENT_ESTABLISHED;
1026 } else
1027 event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
1028 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
1029 ib_event->private_data);
1030 break;
1031 case IB_CM_RTU_RECEIVED:
1032 case IB_CM_USER_ESTABLISHED:
1033 event.event = RDMA_CM_EVENT_ESTABLISHED;
1034 break;
1035 case IB_CM_DREQ_ERROR:
1036 event.status = -ETIMEDOUT; /* fall through */
1037 case IB_CM_DREQ_RECEIVED:
1038 case IB_CM_DREP_RECEIVED:
1039 if (!cma_comp_exch(id_priv, RDMA_CM_CONNECT,
1040 RDMA_CM_DISCONNECT))
1041 goto out;
1042 event.event = RDMA_CM_EVENT_DISCONNECTED;
1043 break;
1044 case IB_CM_TIMEWAIT_EXIT:
1045 event.event = RDMA_CM_EVENT_TIMEWAIT_EXIT;
1046 break;
1047 case IB_CM_MRA_RECEIVED:
1048 /* ignore event */
1049 goto out;
1050 case IB_CM_REJ_RECEIVED:
1051 cma_modify_qp_err(id_priv);
1052 event.status = ib_event->param.rej_rcvd.reason;
1053 event.event = RDMA_CM_EVENT_REJECTED;
1054 event.param.conn.private_data = ib_event->private_data;
1055 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
1056 break;
1057 default:
1058 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
1059 ib_event->event);
1060 goto out;
1061 }
1062
1063 ret = id_priv->id.event_handler(&id_priv->id, &event);
1064 if (ret) {
1065 /* Destroy the CM ID by returning a non-zero value. */
1066 id_priv->cm_id.ib = NULL;
1067 cma_exch(id_priv, RDMA_CM_DESTROYING);
1068 mutex_unlock(&id_priv->handler_mutex);
1069 rdma_destroy_id(&id_priv->id);
1070 return ret;
1071 }
1072out:
1073 mutex_unlock(&id_priv->handler_mutex);
1074 return ret;
1075}
1076
1077static struct rdma_id_private *cma_new_conn_id(struct rdma_cm_id *listen_id,
1078 struct ib_cm_event *ib_event)
1079{
1080 struct rdma_id_private *id_priv;
1081 struct rdma_cm_id *id;
1082 struct rdma_route *rt;
1083 union cma_ip_addr *src, *dst;
1084 __be16 port;
1085 u8 ip_ver;
1086 int ret;
1087
1088 if (cma_get_net_info(ib_event->private_data, listen_id->ps,
1089 &ip_ver, &port, &src, &dst))
1090 return NULL;
1091
1092 id = rdma_create_id(listen_id->event_handler, listen_id->context,
1093 listen_id->ps, ib_event->param.req_rcvd.qp_type);
1094 if (IS_ERR(id))
1095 return NULL;
1096
1097 cma_save_net_info(&id->route.addr, &listen_id->route.addr,
1098 ip_ver, port, src, dst);
1099
1100 rt = &id->route;
1101 rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
1102 rt->path_rec = kmalloc(sizeof *rt->path_rec * rt->num_paths,
1103 GFP_KERNEL);
1104 if (!rt->path_rec)
1105 goto err;
1106
1107 rt->path_rec[0] = *ib_event->param.req_rcvd.primary_path;
1108 if (rt->num_paths == 2)
1109 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
1110
1111 if (cma_any_addr((struct sockaddr *) &rt->addr.src_addr)) {
1112 rt->addr.dev_addr.dev_type = ARPHRD_INFINIBAND;
1113 rdma_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
1114 ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
1115 } else {
1116 ret = rdma_translate_ip((struct sockaddr *) &rt->addr.src_addr,
1117 &rt->addr.dev_addr);
1118 if (ret)
1119 goto err;
1120 }
1121 rdma_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
1122
1123 id_priv = container_of(id, struct rdma_id_private, id);
1124 id_priv->state = RDMA_CM_CONNECT;
1125 return id_priv;
1126
1127err:
1128 rdma_destroy_id(id);
1129 return NULL;
1130}
1131
1132static struct rdma_id_private *cma_new_udp_id(struct rdma_cm_id *listen_id,
1133 struct ib_cm_event *ib_event)
1134{
1135 struct rdma_id_private *id_priv;
1136 struct rdma_cm_id *id;
1137 union cma_ip_addr *src, *dst;
1138 __be16 port;
1139 u8 ip_ver;
1140 int ret;
1141
1142 id = rdma_create_id(listen_id->event_handler, listen_id->context,
1143 listen_id->ps, IB_QPT_UD);
1144 if (IS_ERR(id))
1145 return NULL;
1146
1147
1148 if (cma_get_net_info(ib_event->private_data, listen_id->ps,
1149 &ip_ver, &port, &src, &dst))
1150 goto err;
1151
1152 cma_save_net_info(&id->route.addr, &listen_id->route.addr,
1153 ip_ver, port, src, dst);
1154
1155 if (!cma_any_addr((struct sockaddr *) &id->route.addr.src_addr)) {
1156 ret = rdma_translate_ip((struct sockaddr *) &id->route.addr.src_addr,
1157 &id->route.addr.dev_addr);
1158 if (ret)
1159 goto err;
1160 }
1161
1162 id_priv = container_of(id, struct rdma_id_private, id);
1163 id_priv->state = RDMA_CM_CONNECT;
1164 return id_priv;
1165err:
1166 rdma_destroy_id(id);
1167 return NULL;
1168}
1169
1170static void cma_set_req_event_data(struct rdma_cm_event *event,
1171 struct ib_cm_req_event_param *req_data,
1172 void *private_data, int offset)
1173{
1174 event->param.conn.private_data = private_data + offset;
1175 event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
1176 event->param.conn.responder_resources = req_data->responder_resources;
1177 event->param.conn.initiator_depth = req_data->initiator_depth;
1178 event->param.conn.flow_control = req_data->flow_control;
1179 event->param.conn.retry_count = req_data->retry_count;
1180 event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
1181 event->param.conn.srq = req_data->srq;
1182 event->param.conn.qp_num = req_data->remote_qpn;
1183}
1184
1185static int cma_check_req_qp_type(struct rdma_cm_id *id, struct ib_cm_event *ib_event)
1186{
1187 return (((ib_event->event == IB_CM_REQ_RECEIVED) &&
1188 (ib_event->param.req_rcvd.qp_type == id->qp_type)) ||
1189 ((ib_event->event == IB_CM_SIDR_REQ_RECEIVED) &&
1190 (id->qp_type == IB_QPT_UD)) ||
1191 (!id->qp_type));
1192}
1193
1194static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1195{
1196 struct rdma_id_private *listen_id, *conn_id;
1197 struct rdma_cm_event event;
1198 int offset, ret;
1199
1200 listen_id = cm_id->context;
1201 if (!cma_check_req_qp_type(&listen_id->id, ib_event))
1202 return -EINVAL;
1203
1204 if (cma_disable_callback(listen_id, RDMA_CM_LISTEN))
1205 return -ECONNABORTED;
1206
1207 memset(&event, 0, sizeof event);
1208 offset = cma_user_data_offset(listen_id->id.ps);
1209 event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1210 if (ib_event->event == IB_CM_SIDR_REQ_RECEIVED) {
1211 conn_id = cma_new_udp_id(&listen_id->id, ib_event);
1212 event.param.ud.private_data = ib_event->private_data + offset;
1213 event.param.ud.private_data_len =
1214 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
1215 } else {
1216 conn_id = cma_new_conn_id(&listen_id->id, ib_event);
1217 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
1218 ib_event->private_data, offset);
1219 }
1220 if (!conn_id) {
1221 ret = -ENOMEM;
1222 goto err1;
1223 }
1224
1225 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1226 ret = cma_acquire_dev(conn_id);
1227 if (ret)
1228 goto err2;
1229
1230 conn_id->cm_id.ib = cm_id;
1231 cm_id->context = conn_id;
1232 cm_id->cm_handler = cma_ib_handler;
1233
1234 /*
1235 * Protect against the user destroying conn_id from another thread
1236 * until we're done accessing it.
1237 */
1238 atomic_inc(&conn_id->refcount);
1239 ret = conn_id->id.event_handler(&conn_id->id, &event);
1240 if (ret)
1241 goto err3;
1242
1243 /*
1244 * Acquire mutex to prevent user executing rdma_destroy_id()
1245 * while we're accessing the cm_id.
1246 */
1247 mutex_lock(&lock);
1248 if (cma_comp(conn_id, RDMA_CM_CONNECT) && (conn_id->id.qp_type != IB_QPT_UD))
1249 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
1250 mutex_unlock(&lock);
1251 mutex_unlock(&conn_id->handler_mutex);
1252 mutex_unlock(&listen_id->handler_mutex);
1253 cma_deref_id(conn_id);
1254 return 0;
1255
1256err3:
1257 cma_deref_id(conn_id);
1258 /* Destroy the CM ID by returning a non-zero value. */
1259 conn_id->cm_id.ib = NULL;
1260err2:
1261 cma_exch(conn_id, RDMA_CM_DESTROYING);
1262 mutex_unlock(&conn_id->handler_mutex);
1263err1:
1264 mutex_unlock(&listen_id->handler_mutex);
1265 if (conn_id)
1266 rdma_destroy_id(&conn_id->id);
1267 return ret;
1268}
1269
1270static __be64 cma_get_service_id(enum rdma_port_space ps, struct sockaddr *addr)
1271{
1272 return cpu_to_be64(((u64)ps << 16) + be16_to_cpu(cma_port(addr)));
1273}
1274
1275static void cma_set_compare_data(enum rdma_port_space ps, struct sockaddr *addr,
1276 struct ib_cm_compare_data *compare)
1277{
1278 struct cma_hdr *cma_data, *cma_mask;
1279 struct sdp_hh *sdp_data, *sdp_mask;
1280 __be32 ip4_addr;
1281 struct in6_addr ip6_addr;
1282
1283 memset(compare, 0, sizeof *compare);
1284 cma_data = (void *) compare->data;
1285 cma_mask = (void *) compare->mask;
1286 sdp_data = (void *) compare->data;
1287 sdp_mask = (void *) compare->mask;
1288
1289 switch (addr->sa_family) {
1290 case AF_INET:
1291 ip4_addr = ((struct sockaddr_in *) addr)->sin_addr.s_addr;
1292 if (ps == RDMA_PS_SDP) {
1293 sdp_set_ip_ver(sdp_data, 4);
1294 sdp_set_ip_ver(sdp_mask, 0xF);
1295 sdp_data->dst_addr.ip4.addr = ip4_addr;
1296 sdp_mask->dst_addr.ip4.addr = htonl(~0);
1297 } else {
1298 cma_set_ip_ver(cma_data, 4);
1299 cma_set_ip_ver(cma_mask, 0xF);
1300 cma_data->dst_addr.ip4.addr = ip4_addr;
1301 cma_mask->dst_addr.ip4.addr = htonl(~0);
1302 }
1303 break;
1304 case AF_INET6:
1305 ip6_addr = ((struct sockaddr_in6 *) addr)->sin6_addr;
1306 if (ps == RDMA_PS_SDP) {
1307 sdp_set_ip_ver(sdp_data, 6);
1308 sdp_set_ip_ver(sdp_mask, 0xF);
1309 sdp_data->dst_addr.ip6 = ip6_addr;
1310 memset(&sdp_mask->dst_addr.ip6, 0xFF,
1311 sizeof sdp_mask->dst_addr.ip6);
1312 } else {
1313 cma_set_ip_ver(cma_data, 6);
1314 cma_set_ip_ver(cma_mask, 0xF);
1315 cma_data->dst_addr.ip6 = ip6_addr;
1316 memset(&cma_mask->dst_addr.ip6, 0xFF,
1317 sizeof cma_mask->dst_addr.ip6);
1318 }
1319 break;
1320 default:
1321 break;
1322 }
1323}
1324
1325static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
1326{
1327 struct rdma_id_private *id_priv = iw_id->context;
1328 struct rdma_cm_event event;
1329 struct sockaddr_in *sin;
1330 int ret = 0;
1331
1332 if (cma_disable_callback(id_priv, RDMA_CM_CONNECT))
1333 return 0;
1334
1335 memset(&event, 0, sizeof event);
1336 switch (iw_event->event) {
1337 case IW_CM_EVENT_CLOSE:
1338 event.event = RDMA_CM_EVENT_DISCONNECTED;
1339 break;
1340 case IW_CM_EVENT_CONNECT_REPLY:
1341 sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1342 *sin = iw_event->local_addr;
1343 sin = (struct sockaddr_in *) &id_priv->id.route.addr.dst_addr;
1344 *sin = iw_event->remote_addr;
1345 switch (iw_event->status) {
1346 case 0:
1347 event.event = RDMA_CM_EVENT_ESTABLISHED;
1348 event.param.conn.initiator_depth = iw_event->ird;
1349 event.param.conn.responder_resources = iw_event->ord;
1350 break;
1351 case -ECONNRESET:
1352 case -ECONNREFUSED:
1353 event.event = RDMA_CM_EVENT_REJECTED;
1354 break;
1355 case -ETIMEDOUT:
1356 event.event = RDMA_CM_EVENT_UNREACHABLE;
1357 break;
1358 default:
1359 event.event = RDMA_CM_EVENT_CONNECT_ERROR;
1360 break;
1361 }
1362 break;
1363 case IW_CM_EVENT_ESTABLISHED:
1364 event.event = RDMA_CM_EVENT_ESTABLISHED;
1365 event.param.conn.initiator_depth = iw_event->ird;
1366 event.param.conn.responder_resources = iw_event->ord;
1367 break;
1368 default:
1369 BUG_ON(1);
1370 }
1371
1372 event.status = iw_event->status;
1373 event.param.conn.private_data = iw_event->private_data;
1374 event.param.conn.private_data_len = iw_event->private_data_len;
1375 ret = id_priv->id.event_handler(&id_priv->id, &event);
1376 if (ret) {
1377 /* Destroy the CM ID by returning a non-zero value. */
1378 id_priv->cm_id.iw = NULL;
1379 cma_exch(id_priv, RDMA_CM_DESTROYING);
1380 mutex_unlock(&id_priv->handler_mutex);
1381 rdma_destroy_id(&id_priv->id);
1382 return ret;
1383 }
1384
1385 mutex_unlock(&id_priv->handler_mutex);
1386 return ret;
1387}
1388
1389static int iw_conn_req_handler(struct iw_cm_id *cm_id,
1390 struct iw_cm_event *iw_event)
1391{
1392 struct rdma_cm_id *new_cm_id;
1393 struct rdma_id_private *listen_id, *conn_id;
1394 struct sockaddr_in *sin;
1395 struct net_device *dev = NULL;
1396 struct rdma_cm_event event;
1397 int ret;
1398 struct ib_device_attr attr;
1399
1400 listen_id = cm_id->context;
1401 if (cma_disable_callback(listen_id, RDMA_CM_LISTEN))
1402 return -ECONNABORTED;
1403
1404 /* Create a new RDMA id for the new IW CM ID */
1405 new_cm_id = rdma_create_id(listen_id->id.event_handler,
1406 listen_id->id.context,
1407 RDMA_PS_TCP, IB_QPT_RC);
1408 if (IS_ERR(new_cm_id)) {
1409 ret = -ENOMEM;
1410 goto out;
1411 }
1412 conn_id = container_of(new_cm_id, struct rdma_id_private, id);
1413 mutex_lock_nested(&conn_id->handler_mutex, SINGLE_DEPTH_NESTING);
1414 conn_id->state = RDMA_CM_CONNECT;
1415
1416 dev = ip_dev_find(&init_net, iw_event->local_addr.sin_addr.s_addr);
1417 if (!dev) {
1418 ret = -EADDRNOTAVAIL;
1419 mutex_unlock(&conn_id->handler_mutex);
1420 rdma_destroy_id(new_cm_id);
1421 goto out;
1422 }
1423 ret = rdma_copy_addr(&conn_id->id.route.addr.dev_addr, dev, NULL);
1424 if (ret) {
1425 mutex_unlock(&conn_id->handler_mutex);
1426 rdma_destroy_id(new_cm_id);
1427 goto out;
1428 }
1429
1430 ret = cma_acquire_dev(conn_id);
1431 if (ret) {
1432 mutex_unlock(&conn_id->handler_mutex);
1433 rdma_destroy_id(new_cm_id);
1434 goto out;
1435 }
1436
1437 conn_id->cm_id.iw = cm_id;
1438 cm_id->context = conn_id;
1439 cm_id->cm_handler = cma_iw_handler;
1440
1441 sin = (struct sockaddr_in *) &new_cm_id->route.addr.src_addr;
1442 *sin = iw_event->local_addr;
1443 sin = (struct sockaddr_in *) &new_cm_id->route.addr.dst_addr;
1444 *sin = iw_event->remote_addr;
1445
1446 ret = ib_query_device(conn_id->id.device, &attr);
1447 if (ret) {
1448 mutex_unlock(&conn_id->handler_mutex);
1449 rdma_destroy_id(new_cm_id);
1450 goto out;
1451 }
1452
1453 memset(&event, 0, sizeof event);
1454 event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1455 event.param.conn.private_data = iw_event->private_data;
1456 event.param.conn.private_data_len = iw_event->private_data_len;
1457 event.param.conn.initiator_depth = iw_event->ird;
1458 event.param.conn.responder_resources = iw_event->ord;
1459
1460 /*
1461 * Protect against the user destroying conn_id from another thread
1462 * until we're done accessing it.
1463 */
1464 atomic_inc(&conn_id->refcount);
1465 ret = conn_id->id.event_handler(&conn_id->id, &event);
1466 if (ret) {
1467 /* User wants to destroy the CM ID */
1468 conn_id->cm_id.iw = NULL;
1469 cma_exch(conn_id, RDMA_CM_DESTROYING);
1470 mutex_unlock(&conn_id->handler_mutex);
1471 cma_deref_id(conn_id);
1472 rdma_destroy_id(&conn_id->id);
1473 goto out;
1474 }
1475
1476 mutex_unlock(&conn_id->handler_mutex);
1477 cma_deref_id(conn_id);
1478
1479out:
1480 if (dev)
1481 dev_put(dev);
1482 mutex_unlock(&listen_id->handler_mutex);
1483 return ret;
1484}
1485
1486static int cma_ib_listen(struct rdma_id_private *id_priv)
1487{
1488 struct ib_cm_compare_data compare_data;
1489 struct sockaddr *addr;
1490 struct ib_cm_id *id;
1491 __be64 svc_id;
1492 int ret;
1493
1494 id = ib_create_cm_id(id_priv->id.device, cma_req_handler, id_priv);
1495 if (IS_ERR(id))
1496 return PTR_ERR(id);
1497
1498 id_priv->cm_id.ib = id;
1499
1500 addr = (struct sockaddr *) &id_priv->id.route.addr.src_addr;
1501 svc_id = cma_get_service_id(id_priv->id.ps, addr);
1502 if (cma_any_addr(addr))
1503 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, NULL);
1504 else {
1505 cma_set_compare_data(id_priv->id.ps, addr, &compare_data);
1506 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, &compare_data);
1507 }
1508
1509 if (ret) {
1510 ib_destroy_cm_id(id_priv->cm_id.ib);
1511 id_priv->cm_id.ib = NULL;
1512 }
1513
1514 return ret;
1515}
1516
1517static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
1518{
1519 int ret;
1520 struct sockaddr_in *sin;
1521 struct iw_cm_id *id;
1522
1523 id = iw_create_cm_id(id_priv->id.device,
1524 iw_conn_req_handler,
1525 id_priv);
1526 if (IS_ERR(id))
1527 return PTR_ERR(id);
1528
1529 id_priv->cm_id.iw = id;
1530
1531 sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1532 id_priv->cm_id.iw->local_addr = *sin;
1533
1534 ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
1535
1536 if (ret) {
1537 iw_destroy_cm_id(id_priv->cm_id.iw);
1538 id_priv->cm_id.iw = NULL;
1539 }
1540
1541 return ret;
1542}
1543
1544static int cma_listen_handler(struct rdma_cm_id *id,
1545 struct rdma_cm_event *event)
1546{
1547 struct rdma_id_private *id_priv = id->context;
1548
1549 id->context = id_priv->id.context;
1550 id->event_handler = id_priv->id.event_handler;
1551 return id_priv->id.event_handler(id, event);
1552}
1553
1554static void cma_listen_on_dev(struct rdma_id_private *id_priv,
1555 struct cma_device *cma_dev)
1556{
1557 struct rdma_id_private *dev_id_priv;
1558 struct rdma_cm_id *id;
1559 int ret;
1560
1561 id = rdma_create_id(cma_listen_handler, id_priv, id_priv->id.ps,
1562 id_priv->id.qp_type);
1563 if (IS_ERR(id))
1564 return;
1565
1566 dev_id_priv = container_of(id, struct rdma_id_private, id);
1567
1568 dev_id_priv->state = RDMA_CM_ADDR_BOUND;
1569 memcpy(&id->route.addr.src_addr, &id_priv->id.route.addr.src_addr,
1570 ip_addr_size((struct sockaddr *) &id_priv->id.route.addr.src_addr));
1571
1572 cma_attach_to_dev(dev_id_priv, cma_dev);
1573 list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
1574 atomic_inc(&id_priv->refcount);
1575 dev_id_priv->internal_id = 1;
1576
1577 ret = rdma_listen(id, id_priv->backlog);
1578 if (ret)
1579 printk(KERN_WARNING "RDMA CMA: cma_listen_on_dev, error %d, "
1580 "listening on device %s\n", ret, cma_dev->device->name);
1581}
1582
1583static void cma_listen_on_all(struct rdma_id_private *id_priv)
1584{
1585 struct cma_device *cma_dev;
1586
1587 mutex_lock(&lock);
1588 list_add_tail(&id_priv->list, &listen_any_list);
1589 list_for_each_entry(cma_dev, &dev_list, list)
1590 cma_listen_on_dev(id_priv, cma_dev);
1591 mutex_unlock(&lock);
1592}
1593
1594void rdma_set_service_type(struct rdma_cm_id *id, int tos)
1595{
1596 struct rdma_id_private *id_priv;
1597
1598 id_priv = container_of(id, struct rdma_id_private, id);
1599 id_priv->tos = (u8) tos;
1600}
1601EXPORT_SYMBOL(rdma_set_service_type);
1602
1603static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec,
1604 void *context)
1605{
1606 struct cma_work *work = context;
1607 struct rdma_route *route;
1608
1609 route = &work->id->id.route;
1610
1611 if (!status) {
1612 route->num_paths = 1;
1613 *route->path_rec = *path_rec;
1614 } else {
1615 work->old_state = RDMA_CM_ROUTE_QUERY;
1616 work->new_state = RDMA_CM_ADDR_RESOLVED;
1617 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
1618 work->event.status = status;
1619 }
1620
1621 queue_work(cma_wq, &work->work);
1622}
1623
1624static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms,
1625 struct cma_work *work)
1626{
1627 struct rdma_addr *addr = &id_priv->id.route.addr;
1628 struct ib_sa_path_rec path_rec;
1629 ib_sa_comp_mask comp_mask;
1630 struct sockaddr_in6 *sin6;
1631
1632 memset(&path_rec, 0, sizeof path_rec);
1633 rdma_addr_get_sgid(&addr->dev_addr, &path_rec.sgid);
1634 rdma_addr_get_dgid(&addr->dev_addr, &path_rec.dgid);
1635 path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(&addr->dev_addr));
1636 path_rec.numb_path = 1;
1637 path_rec.reversible = 1;
1638 path_rec.service_id = cma_get_service_id(id_priv->id.ps,
1639 (struct sockaddr *) &addr->dst_addr);
1640
1641 comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
1642 IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
1643 IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
1644
1645 if (addr->src_addr.ss_family == AF_INET) {
1646 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
1647 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
1648 } else {
1649 sin6 = (struct sockaddr_in6 *) &addr->src_addr;
1650 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
1651 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
1652 }
1653
1654 id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
1655 id_priv->id.port_num, &path_rec,
1656 comp_mask, timeout_ms,
1657 GFP_KERNEL, cma_query_handler,
1658 work, &id_priv->query);
1659
1660 return (id_priv->query_id < 0) ? id_priv->query_id : 0;
1661}
1662
1663static void cma_work_handler(struct work_struct *_work)
1664{
1665 struct cma_work *work = container_of(_work, struct cma_work, work);
1666 struct rdma_id_private *id_priv = work->id;
1667 int destroy = 0;
1668
1669 mutex_lock(&id_priv->handler_mutex);
1670 if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
1671 goto out;
1672
1673 if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1674 cma_exch(id_priv, RDMA_CM_DESTROYING);
1675 destroy = 1;
1676 }
1677out:
1678 mutex_unlock(&id_priv->handler_mutex);
1679 cma_deref_id(id_priv);
1680 if (destroy)
1681 rdma_destroy_id(&id_priv->id);
1682 kfree(work);
1683}
1684
1685static void cma_ndev_work_handler(struct work_struct *_work)
1686{
1687 struct cma_ndev_work *work = container_of(_work, struct cma_ndev_work, work);
1688 struct rdma_id_private *id_priv = work->id;
1689 int destroy = 0;
1690
1691 mutex_lock(&id_priv->handler_mutex);
1692 if (id_priv->state == RDMA_CM_DESTROYING ||
1693 id_priv->state == RDMA_CM_DEVICE_REMOVAL)
1694 goto out;
1695
1696 if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1697 cma_exch(id_priv, RDMA_CM_DESTROYING);
1698 destroy = 1;
1699 }
1700
1701out:
1702 mutex_unlock(&id_priv->handler_mutex);
1703 cma_deref_id(id_priv);
1704 if (destroy)
1705 rdma_destroy_id(&id_priv->id);
1706 kfree(work);
1707}
1708
1709static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms)
1710{
1711 struct rdma_route *route = &id_priv->id.route;
1712 struct cma_work *work;
1713 int ret;
1714
1715 work = kzalloc(sizeof *work, GFP_KERNEL);
1716 if (!work)
1717 return -ENOMEM;
1718
1719 work->id = id_priv;
1720 INIT_WORK(&work->work, cma_work_handler);
1721 work->old_state = RDMA_CM_ROUTE_QUERY;
1722 work->new_state = RDMA_CM_ROUTE_RESOLVED;
1723 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1724
1725 route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
1726 if (!route->path_rec) {
1727 ret = -ENOMEM;
1728 goto err1;
1729 }
1730
1731 ret = cma_query_ib_route(id_priv, timeout_ms, work);
1732 if (ret)
1733 goto err2;
1734
1735 return 0;
1736err2:
1737 kfree(route->path_rec);
1738 route->path_rec = NULL;
1739err1:
1740 kfree(work);
1741 return ret;
1742}
1743
1744int rdma_set_ib_paths(struct rdma_cm_id *id,
1745 struct ib_sa_path_rec *path_rec, int num_paths)
1746{
1747 struct rdma_id_private *id_priv;
1748 int ret;
1749
1750 id_priv = container_of(id, struct rdma_id_private, id);
1751 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
1752 RDMA_CM_ROUTE_RESOLVED))
1753 return -EINVAL;
1754
1755 id->route.path_rec = kmemdup(path_rec, sizeof *path_rec * num_paths,
1756 GFP_KERNEL);
1757 if (!id->route.path_rec) {
1758 ret = -ENOMEM;
1759 goto err;
1760 }
1761
1762 id->route.num_paths = num_paths;
1763 return 0;
1764err:
1765 cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_ADDR_RESOLVED);
1766 return ret;
1767}
1768EXPORT_SYMBOL(rdma_set_ib_paths);
1769
1770static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms)
1771{
1772 struct cma_work *work;
1773
1774 work = kzalloc(sizeof *work, GFP_KERNEL);
1775 if (!work)
1776 return -ENOMEM;
1777
1778 work->id = id_priv;
1779 INIT_WORK(&work->work, cma_work_handler);
1780 work->old_state = RDMA_CM_ROUTE_QUERY;
1781 work->new_state = RDMA_CM_ROUTE_RESOLVED;
1782 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1783 queue_work(cma_wq, &work->work);
1784 return 0;
1785}
1786
1787static int cma_resolve_iboe_route(struct rdma_id_private *id_priv)
1788{
1789 struct rdma_route *route = &id_priv->id.route;
1790 struct rdma_addr *addr = &route->addr;
1791 struct cma_work *work;
1792 int ret;
1793 struct sockaddr_in *src_addr = (struct sockaddr_in *)&route->addr.src_addr;
1794 struct sockaddr_in *dst_addr = (struct sockaddr_in *)&route->addr.dst_addr;
1795 struct net_device *ndev = NULL;
1796 u16 vid;
1797
1798 if (src_addr->sin_family != dst_addr->sin_family)
1799 return -EINVAL;
1800
1801 work = kzalloc(sizeof *work, GFP_KERNEL);
1802 if (!work)
1803 return -ENOMEM;
1804
1805 work->id = id_priv;
1806 INIT_WORK(&work->work, cma_work_handler);
1807
1808 route->path_rec = kzalloc(sizeof *route->path_rec, GFP_KERNEL);
1809 if (!route->path_rec) {
1810 ret = -ENOMEM;
1811 goto err1;
1812 }
1813
1814 route->num_paths = 1;
1815
1816 if (addr->dev_addr.bound_dev_if)
1817 ndev = dev_get_by_index(&init_net, addr->dev_addr.bound_dev_if);
1818 if (!ndev) {
1819 ret = -ENODEV;
1820 goto err2;
1821 }
1822
1823 vid = rdma_vlan_dev_vlan_id(ndev);
1824
1825 iboe_mac_vlan_to_ll(&route->path_rec->sgid, addr->dev_addr.src_dev_addr, vid);
1826 iboe_mac_vlan_to_ll(&route->path_rec->dgid, addr->dev_addr.dst_dev_addr, vid);
1827
1828 route->path_rec->hop_limit = 1;
1829 route->path_rec->reversible = 1;
1830 route->path_rec->pkey = cpu_to_be16(0xffff);
1831 route->path_rec->mtu_selector = IB_SA_EQ;
1832 route->path_rec->sl = netdev_get_prio_tc_map(
1833 ndev->priv_flags & IFF_802_1Q_VLAN ?
1834 vlan_dev_real_dev(ndev) : ndev,
1835 rt_tos2priority(id_priv->tos));
1836
1837 route->path_rec->mtu = iboe_get_mtu(ndev->mtu);
1838 route->path_rec->rate_selector = IB_SA_EQ;
1839 route->path_rec->rate = iboe_get_rate(ndev);
1840 dev_put(ndev);
1841 route->path_rec->packet_life_time_selector = IB_SA_EQ;
1842 route->path_rec->packet_life_time = CMA_IBOE_PACKET_LIFETIME;
1843 if (!route->path_rec->mtu) {
1844 ret = -EINVAL;
1845 goto err2;
1846 }
1847
1848 work->old_state = RDMA_CM_ROUTE_QUERY;
1849 work->new_state = RDMA_CM_ROUTE_RESOLVED;
1850 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1851 work->event.status = 0;
1852
1853 queue_work(cma_wq, &work->work);
1854
1855 return 0;
1856
1857err2:
1858 kfree(route->path_rec);
1859 route->path_rec = NULL;
1860err1:
1861 kfree(work);
1862 return ret;
1863}
1864
1865int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
1866{
1867 struct rdma_id_private *id_priv;
1868 int ret;
1869
1870 id_priv = container_of(id, struct rdma_id_private, id);
1871 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED, RDMA_CM_ROUTE_QUERY))
1872 return -EINVAL;
1873
1874 atomic_inc(&id_priv->refcount);
1875 switch (rdma_node_get_transport(id->device->node_type)) {
1876 case RDMA_TRANSPORT_IB:
1877 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
1878 case IB_LINK_LAYER_INFINIBAND:
1879 ret = cma_resolve_ib_route(id_priv, timeout_ms);
1880 break;
1881 case IB_LINK_LAYER_ETHERNET:
1882 ret = cma_resolve_iboe_route(id_priv);
1883 break;
1884 default:
1885 ret = -ENOSYS;
1886 }
1887 break;
1888 case RDMA_TRANSPORT_IWARP:
1889 ret = cma_resolve_iw_route(id_priv, timeout_ms);
1890 break;
1891 default:
1892 ret = -ENOSYS;
1893 break;
1894 }
1895 if (ret)
1896 goto err;
1897
1898 return 0;
1899err:
1900 cma_comp_exch(id_priv, RDMA_CM_ROUTE_QUERY, RDMA_CM_ADDR_RESOLVED);
1901 cma_deref_id(id_priv);
1902 return ret;
1903}
1904EXPORT_SYMBOL(rdma_resolve_route);
1905
1906static int cma_bind_loopback(struct rdma_id_private *id_priv)
1907{
1908 struct cma_device *cma_dev;
1909 struct ib_port_attr port_attr;
1910 union ib_gid gid;
1911 u16 pkey;
1912 int ret;
1913 u8 p;
1914
1915 mutex_lock(&lock);
1916 if (list_empty(&dev_list)) {
1917 ret = -ENODEV;
1918 goto out;
1919 }
1920 list_for_each_entry(cma_dev, &dev_list, list)
1921 for (p = 1; p <= cma_dev->device->phys_port_cnt; ++p)
1922 if (!ib_query_port(cma_dev->device, p, &port_attr) &&
1923 port_attr.state == IB_PORT_ACTIVE)
1924 goto port_found;
1925
1926 p = 1;
1927 cma_dev = list_entry(dev_list.next, struct cma_device, list);
1928
1929port_found:
1930 ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid);
1931 if (ret)
1932 goto out;
1933
1934 ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
1935 if (ret)
1936 goto out;
1937
1938 id_priv->id.route.addr.dev_addr.dev_type =
1939 (rdma_port_get_link_layer(cma_dev->device, p) == IB_LINK_LAYER_INFINIBAND) ?
1940 ARPHRD_INFINIBAND : ARPHRD_ETHER;
1941
1942 rdma_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
1943 ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
1944 id_priv->id.port_num = p;
1945 cma_attach_to_dev(id_priv, cma_dev);
1946out:
1947 mutex_unlock(&lock);
1948 return ret;
1949}
1950
1951static void addr_handler(int status, struct sockaddr *src_addr,
1952 struct rdma_dev_addr *dev_addr, void *context)
1953{
1954 struct rdma_id_private *id_priv = context;
1955 struct rdma_cm_event event;
1956
1957 memset(&event, 0, sizeof event);
1958 mutex_lock(&id_priv->handler_mutex);
1959 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY,
1960 RDMA_CM_ADDR_RESOLVED))
1961 goto out;
1962
1963 if (!status && !id_priv->cma_dev)
1964 status = cma_acquire_dev(id_priv);
1965
1966 if (status) {
1967 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_RESOLVED,
1968 RDMA_CM_ADDR_BOUND))
1969 goto out;
1970 event.event = RDMA_CM_EVENT_ADDR_ERROR;
1971 event.status = status;
1972 } else {
1973 memcpy(&id_priv->id.route.addr.src_addr, src_addr,
1974 ip_addr_size(src_addr));
1975 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
1976 }
1977
1978 if (id_priv->id.event_handler(&id_priv->id, &event)) {
1979 cma_exch(id_priv, RDMA_CM_DESTROYING);
1980 mutex_unlock(&id_priv->handler_mutex);
1981 cma_deref_id(id_priv);
1982 rdma_destroy_id(&id_priv->id);
1983 return;
1984 }
1985out:
1986 mutex_unlock(&id_priv->handler_mutex);
1987 cma_deref_id(id_priv);
1988}
1989
1990static int cma_resolve_loopback(struct rdma_id_private *id_priv)
1991{
1992 struct cma_work *work;
1993 struct sockaddr *src, *dst;
1994 union ib_gid gid;
1995 int ret;
1996
1997 work = kzalloc(sizeof *work, GFP_KERNEL);
1998 if (!work)
1999 return -ENOMEM;
2000
2001 if (!id_priv->cma_dev) {
2002 ret = cma_bind_loopback(id_priv);
2003 if (ret)
2004 goto err;
2005 }
2006
2007 rdma_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
2008 rdma_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
2009
2010 src = (struct sockaddr *) &id_priv->id.route.addr.src_addr;
2011 if (cma_zero_addr(src)) {
2012 dst = (struct sockaddr *) &id_priv->id.route.addr.dst_addr;
2013 if ((src->sa_family = dst->sa_family) == AF_INET) {
2014 ((struct sockaddr_in *)src)->sin_addr =
2015 ((struct sockaddr_in *)dst)->sin_addr;
2016 } else {
2017 ((struct sockaddr_in6 *)src)->sin6_addr =
2018 ((struct sockaddr_in6 *)dst)->sin6_addr;
2019 }
2020 }
2021
2022 work->id = id_priv;
2023 INIT_WORK(&work->work, cma_work_handler);
2024 work->old_state = RDMA_CM_ADDR_QUERY;
2025 work->new_state = RDMA_CM_ADDR_RESOLVED;
2026 work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
2027 queue_work(cma_wq, &work->work);
2028 return 0;
2029err:
2030 kfree(work);
2031 return ret;
2032}
2033
2034static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2035 struct sockaddr *dst_addr)
2036{
2037 if (!src_addr || !src_addr->sa_family) {
2038 src_addr = (struct sockaddr *) &id->route.addr.src_addr;
2039 if ((src_addr->sa_family = dst_addr->sa_family) == AF_INET6) {
2040 ((struct sockaddr_in6 *) src_addr)->sin6_scope_id =
2041 ((struct sockaddr_in6 *) dst_addr)->sin6_scope_id;
2042 }
2043 }
2044 return rdma_bind_addr(id, src_addr);
2045}
2046
2047int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
2048 struct sockaddr *dst_addr, int timeout_ms)
2049{
2050 struct rdma_id_private *id_priv;
2051 int ret;
2052
2053 id_priv = container_of(id, struct rdma_id_private, id);
2054 if (id_priv->state == RDMA_CM_IDLE) {
2055 ret = cma_bind_addr(id, src_addr, dst_addr);
2056 if (ret)
2057 return ret;
2058 }
2059
2060 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_ADDR_QUERY))
2061 return -EINVAL;
2062
2063 atomic_inc(&id_priv->refcount);
2064 memcpy(&id->route.addr.dst_addr, dst_addr, ip_addr_size(dst_addr));
2065 if (cma_any_addr(dst_addr))
2066 ret = cma_resolve_loopback(id_priv);
2067 else
2068 ret = rdma_resolve_ip(&addr_client, (struct sockaddr *) &id->route.addr.src_addr,
2069 dst_addr, &id->route.addr.dev_addr,
2070 timeout_ms, addr_handler, id_priv);
2071 if (ret)
2072 goto err;
2073
2074 return 0;
2075err:
2076 cma_comp_exch(id_priv, RDMA_CM_ADDR_QUERY, RDMA_CM_ADDR_BOUND);
2077 cma_deref_id(id_priv);
2078 return ret;
2079}
2080EXPORT_SYMBOL(rdma_resolve_addr);
2081
2082int rdma_set_reuseaddr(struct rdma_cm_id *id, int reuse)
2083{
2084 struct rdma_id_private *id_priv;
2085 unsigned long flags;
2086 int ret;
2087
2088 id_priv = container_of(id, struct rdma_id_private, id);
2089 spin_lock_irqsave(&id_priv->lock, flags);
2090 if (id_priv->state == RDMA_CM_IDLE) {
2091 id_priv->reuseaddr = reuse;
2092 ret = 0;
2093 } else {
2094 ret = -EINVAL;
2095 }
2096 spin_unlock_irqrestore(&id_priv->lock, flags);
2097 return ret;
2098}
2099EXPORT_SYMBOL(rdma_set_reuseaddr);
2100
2101static void cma_bind_port(struct rdma_bind_list *bind_list,
2102 struct rdma_id_private *id_priv)
2103{
2104 struct sockaddr_in *sin;
2105
2106 sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
2107 sin->sin_port = htons(bind_list->port);
2108 id_priv->bind_list = bind_list;
2109 hlist_add_head(&id_priv->node, &bind_list->owners);
2110}
2111
2112static int cma_alloc_port(struct idr *ps, struct rdma_id_private *id_priv,
2113 unsigned short snum)
2114{
2115 struct rdma_bind_list *bind_list;
2116 int port, ret;
2117
2118 bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
2119 if (!bind_list)
2120 return -ENOMEM;
2121
2122 do {
2123 ret = idr_get_new_above(ps, bind_list, snum, &port);
2124 } while ((ret == -EAGAIN) && idr_pre_get(ps, GFP_KERNEL));
2125
2126 if (ret)
2127 goto err1;
2128
2129 if (port != snum) {
2130 ret = -EADDRNOTAVAIL;
2131 goto err2;
2132 }
2133
2134 bind_list->ps = ps;
2135 bind_list->port = (unsigned short) port;
2136 cma_bind_port(bind_list, id_priv);
2137 return 0;
2138err2:
2139 idr_remove(ps, port);
2140err1:
2141 kfree(bind_list);
2142 return ret;
2143}
2144
2145static int cma_alloc_any_port(struct idr *ps, struct rdma_id_private *id_priv)
2146{
2147 static unsigned int last_used_port;
2148 int low, high, remaining;
2149 unsigned int rover;
2150
2151 inet_get_local_port_range(&low, &high);
2152 remaining = (high - low) + 1;
2153 rover = net_random() % remaining + low;
2154retry:
2155 if (last_used_port != rover &&
2156 !idr_find(ps, (unsigned short) rover)) {
2157 int ret = cma_alloc_port(ps, id_priv, rover);
2158 /*
2159 * Remember previously used port number in order to avoid
2160 * re-using same port immediately after it is closed.
2161 */
2162 if (!ret)
2163 last_used_port = rover;
2164 if (ret != -EADDRNOTAVAIL)
2165 return ret;
2166 }
2167 if (--remaining) {
2168 rover++;
2169 if ((rover < low) || (rover > high))
2170 rover = low;
2171 goto retry;
2172 }
2173 return -EADDRNOTAVAIL;
2174}
2175
2176/*
2177 * Check that the requested port is available. This is called when trying to
2178 * bind to a specific port, or when trying to listen on a bound port. In
2179 * the latter case, the provided id_priv may already be on the bind_list, but
2180 * we still need to check that it's okay to start listening.
2181 */
2182static int cma_check_port(struct rdma_bind_list *bind_list,
2183 struct rdma_id_private *id_priv, uint8_t reuseaddr)
2184{
2185 struct rdma_id_private *cur_id;
2186 struct sockaddr *addr, *cur_addr;
2187 struct hlist_node *node;
2188
2189 addr = (struct sockaddr *) &id_priv->id.route.addr.src_addr;
2190 if (cma_any_addr(addr) && !reuseaddr)
2191 return -EADDRNOTAVAIL;
2192
2193 hlist_for_each_entry(cur_id, node, &bind_list->owners, node) {
2194 if (id_priv == cur_id)
2195 continue;
2196
2197 if ((cur_id->state == RDMA_CM_LISTEN) ||
2198 !reuseaddr || !cur_id->reuseaddr) {
2199 cur_addr = (struct sockaddr *) &cur_id->id.route.addr.src_addr;
2200 if (cma_any_addr(cur_addr))
2201 return -EADDRNOTAVAIL;
2202
2203 if (!cma_addr_cmp(addr, cur_addr))
2204 return -EADDRINUSE;
2205 }
2206 }
2207 return 0;
2208}
2209
2210static int cma_use_port(struct idr *ps, struct rdma_id_private *id_priv)
2211{
2212 struct rdma_bind_list *bind_list;
2213 unsigned short snum;
2214 int ret;
2215
2216 snum = ntohs(cma_port((struct sockaddr *) &id_priv->id.route.addr.src_addr));
2217 if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
2218 return -EACCES;
2219
2220 bind_list = idr_find(ps, snum);
2221 if (!bind_list) {
2222 ret = cma_alloc_port(ps, id_priv, snum);
2223 } else {
2224 ret = cma_check_port(bind_list, id_priv, id_priv->reuseaddr);
2225 if (!ret)
2226 cma_bind_port(bind_list, id_priv);
2227 }
2228 return ret;
2229}
2230
2231static int cma_bind_listen(struct rdma_id_private *id_priv)
2232{
2233 struct rdma_bind_list *bind_list = id_priv->bind_list;
2234 int ret = 0;
2235
2236 mutex_lock(&lock);
2237 if (bind_list->owners.first->next)
2238 ret = cma_check_port(bind_list, id_priv, 0);
2239 mutex_unlock(&lock);
2240 return ret;
2241}
2242
2243static int cma_get_port(struct rdma_id_private *id_priv)
2244{
2245 struct idr *ps;
2246 int ret;
2247
2248 switch (id_priv->id.ps) {
2249 case RDMA_PS_SDP:
2250 ps = &sdp_ps;
2251 break;
2252 case RDMA_PS_TCP:
2253 ps = &tcp_ps;
2254 break;
2255 case RDMA_PS_UDP:
2256 ps = &udp_ps;
2257 break;
2258 case RDMA_PS_IPOIB:
2259 ps = &ipoib_ps;
2260 break;
2261 case RDMA_PS_IB:
2262 ps = &ib_ps;
2263 break;
2264 default:
2265 return -EPROTONOSUPPORT;
2266 }
2267
2268 mutex_lock(&lock);
2269 if (cma_any_port((struct sockaddr *) &id_priv->id.route.addr.src_addr))
2270 ret = cma_alloc_any_port(ps, id_priv);
2271 else
2272 ret = cma_use_port(ps, id_priv);
2273 mutex_unlock(&lock);
2274
2275 return ret;
2276}
2277
2278static int cma_check_linklocal(struct rdma_dev_addr *dev_addr,
2279 struct sockaddr *addr)
2280{
2281#if defined(CONFIG_IPV6) || defined(CONFIG_IPV6_MODULE)
2282 struct sockaddr_in6 *sin6;
2283
2284 if (addr->sa_family != AF_INET6)
2285 return 0;
2286
2287 sin6 = (struct sockaddr_in6 *) addr;
2288 if ((ipv6_addr_type(&sin6->sin6_addr) & IPV6_ADDR_LINKLOCAL) &&
2289 !sin6->sin6_scope_id)
2290 return -EINVAL;
2291
2292 dev_addr->bound_dev_if = sin6->sin6_scope_id;
2293#endif
2294 return 0;
2295}
2296
2297int rdma_listen(struct rdma_cm_id *id, int backlog)
2298{
2299 struct rdma_id_private *id_priv;
2300 int ret;
2301
2302 id_priv = container_of(id, struct rdma_id_private, id);
2303 if (id_priv->state == RDMA_CM_IDLE) {
2304 ((struct sockaddr *) &id->route.addr.src_addr)->sa_family = AF_INET;
2305 ret = rdma_bind_addr(id, (struct sockaddr *) &id->route.addr.src_addr);
2306 if (ret)
2307 return ret;
2308 }
2309
2310 if (!cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_LISTEN))
2311 return -EINVAL;
2312
2313 if (id_priv->reuseaddr) {
2314 ret = cma_bind_listen(id_priv);
2315 if (ret)
2316 goto err;
2317 }
2318
2319 id_priv->backlog = backlog;
2320 if (id->device) {
2321 switch (rdma_node_get_transport(id->device->node_type)) {
2322 case RDMA_TRANSPORT_IB:
2323 ret = cma_ib_listen(id_priv);
2324 if (ret)
2325 goto err;
2326 break;
2327 case RDMA_TRANSPORT_IWARP:
2328 ret = cma_iw_listen(id_priv, backlog);
2329 if (ret)
2330 goto err;
2331 break;
2332 default:
2333 ret = -ENOSYS;
2334 goto err;
2335 }
2336 } else
2337 cma_listen_on_all(id_priv);
2338
2339 return 0;
2340err:
2341 id_priv->backlog = 0;
2342 cma_comp_exch(id_priv, RDMA_CM_LISTEN, RDMA_CM_ADDR_BOUND);
2343 return ret;
2344}
2345EXPORT_SYMBOL(rdma_listen);
2346
2347int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
2348{
2349 struct rdma_id_private *id_priv;
2350 int ret;
2351
2352 if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6)
2353 return -EAFNOSUPPORT;
2354
2355 id_priv = container_of(id, struct rdma_id_private, id);
2356 if (!cma_comp_exch(id_priv, RDMA_CM_IDLE, RDMA_CM_ADDR_BOUND))
2357 return -EINVAL;
2358
2359 ret = cma_check_linklocal(&id->route.addr.dev_addr, addr);
2360 if (ret)
2361 goto err1;
2362
2363 if (!cma_any_addr(addr)) {
2364 ret = rdma_translate_ip(addr, &id->route.addr.dev_addr);
2365 if (ret)
2366 goto err1;
2367
2368 ret = cma_acquire_dev(id_priv);
2369 if (ret)
2370 goto err1;
2371 }
2372
2373 memcpy(&id->route.addr.src_addr, addr, ip_addr_size(addr));
2374 ret = cma_get_port(id_priv);
2375 if (ret)
2376 goto err2;
2377
2378 return 0;
2379err2:
2380 if (id_priv->cma_dev)
2381 cma_release_dev(id_priv);
2382err1:
2383 cma_comp_exch(id_priv, RDMA_CM_ADDR_BOUND, RDMA_CM_IDLE);
2384 return ret;
2385}
2386EXPORT_SYMBOL(rdma_bind_addr);
2387
2388static int cma_format_hdr(void *hdr, enum rdma_port_space ps,
2389 struct rdma_route *route)
2390{
2391 struct cma_hdr *cma_hdr;
2392 struct sdp_hh *sdp_hdr;
2393
2394 if (route->addr.src_addr.ss_family == AF_INET) {
2395 struct sockaddr_in *src4, *dst4;
2396
2397 src4 = (struct sockaddr_in *) &route->addr.src_addr;
2398 dst4 = (struct sockaddr_in *) &route->addr.dst_addr;
2399
2400 switch (ps) {
2401 case RDMA_PS_SDP:
2402 sdp_hdr = hdr;
2403 if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION)
2404 return -EINVAL;
2405 sdp_set_ip_ver(sdp_hdr, 4);
2406 sdp_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2407 sdp_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2408 sdp_hdr->port = src4->sin_port;
2409 break;
2410 default:
2411 cma_hdr = hdr;
2412 cma_hdr->cma_version = CMA_VERSION;
2413 cma_set_ip_ver(cma_hdr, 4);
2414 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2415 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2416 cma_hdr->port = src4->sin_port;
2417 break;
2418 }
2419 } else {
2420 struct sockaddr_in6 *src6, *dst6;
2421
2422 src6 = (struct sockaddr_in6 *) &route->addr.src_addr;
2423 dst6 = (struct sockaddr_in6 *) &route->addr.dst_addr;
2424
2425 switch (ps) {
2426 case RDMA_PS_SDP:
2427 sdp_hdr = hdr;
2428 if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION)
2429 return -EINVAL;
2430 sdp_set_ip_ver(sdp_hdr, 6);
2431 sdp_hdr->src_addr.ip6 = src6->sin6_addr;
2432 sdp_hdr->dst_addr.ip6 = dst6->sin6_addr;
2433 sdp_hdr->port = src6->sin6_port;
2434 break;
2435 default:
2436 cma_hdr = hdr;
2437 cma_hdr->cma_version = CMA_VERSION;
2438 cma_set_ip_ver(cma_hdr, 6);
2439 cma_hdr->src_addr.ip6 = src6->sin6_addr;
2440 cma_hdr->dst_addr.ip6 = dst6->sin6_addr;
2441 cma_hdr->port = src6->sin6_port;
2442 break;
2443 }
2444 }
2445 return 0;
2446}
2447
2448static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
2449 struct ib_cm_event *ib_event)
2450{
2451 struct rdma_id_private *id_priv = cm_id->context;
2452 struct rdma_cm_event event;
2453 struct ib_cm_sidr_rep_event_param *rep = &ib_event->param.sidr_rep_rcvd;
2454 int ret = 0;
2455
2456 if (cma_disable_callback(id_priv, RDMA_CM_CONNECT))
2457 return 0;
2458
2459 memset(&event, 0, sizeof event);
2460 switch (ib_event->event) {
2461 case IB_CM_SIDR_REQ_ERROR:
2462 event.event = RDMA_CM_EVENT_UNREACHABLE;
2463 event.status = -ETIMEDOUT;
2464 break;
2465 case IB_CM_SIDR_REP_RECEIVED:
2466 event.param.ud.private_data = ib_event->private_data;
2467 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
2468 if (rep->status != IB_SIDR_SUCCESS) {
2469 event.event = RDMA_CM_EVENT_UNREACHABLE;
2470 event.status = ib_event->param.sidr_rep_rcvd.status;
2471 break;
2472 }
2473 ret = cma_set_qkey(id_priv);
2474 if (ret) {
2475 event.event = RDMA_CM_EVENT_ADDR_ERROR;
2476 event.status = -EINVAL;
2477 break;
2478 }
2479 if (id_priv->qkey != rep->qkey) {
2480 event.event = RDMA_CM_EVENT_UNREACHABLE;
2481 event.status = -EINVAL;
2482 break;
2483 }
2484 ib_init_ah_from_path(id_priv->id.device, id_priv->id.port_num,
2485 id_priv->id.route.path_rec,
2486 &event.param.ud.ah_attr);
2487 event.param.ud.qp_num = rep->qpn;
2488 event.param.ud.qkey = rep->qkey;
2489 event.event = RDMA_CM_EVENT_ESTABLISHED;
2490 event.status = 0;
2491 break;
2492 default:
2493 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d\n",
2494 ib_event->event);
2495 goto out;
2496 }
2497
2498 ret = id_priv->id.event_handler(&id_priv->id, &event);
2499 if (ret) {
2500 /* Destroy the CM ID by returning a non-zero value. */
2501 id_priv->cm_id.ib = NULL;
2502 cma_exch(id_priv, RDMA_CM_DESTROYING);
2503 mutex_unlock(&id_priv->handler_mutex);
2504 rdma_destroy_id(&id_priv->id);
2505 return ret;
2506 }
2507out:
2508 mutex_unlock(&id_priv->handler_mutex);
2509 return ret;
2510}
2511
2512static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
2513 struct rdma_conn_param *conn_param)
2514{
2515 struct ib_cm_sidr_req_param req;
2516 struct rdma_route *route;
2517 struct ib_cm_id *id;
2518 int ret;
2519
2520 req.private_data_len = sizeof(struct cma_hdr) +
2521 conn_param->private_data_len;
2522 if (req.private_data_len < conn_param->private_data_len)
2523 return -EINVAL;
2524
2525 req.private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2526 if (!req.private_data)
2527 return -ENOMEM;
2528
2529 if (conn_param->private_data && conn_param->private_data_len)
2530 memcpy((void *) req.private_data + sizeof(struct cma_hdr),
2531 conn_param->private_data, conn_param->private_data_len);
2532
2533 route = &id_priv->id.route;
2534 ret = cma_format_hdr((void *) req.private_data, id_priv->id.ps, route);
2535 if (ret)
2536 goto out;
2537
2538 id = ib_create_cm_id(id_priv->id.device, cma_sidr_rep_handler,
2539 id_priv);
2540 if (IS_ERR(id)) {
2541 ret = PTR_ERR(id);
2542 goto out;
2543 }
2544 id_priv->cm_id.ib = id;
2545
2546 req.path = route->path_rec;
2547 req.service_id = cma_get_service_id(id_priv->id.ps,
2548 (struct sockaddr *) &route->addr.dst_addr);
2549 req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
2550 req.max_cm_retries = CMA_MAX_CM_RETRIES;
2551
2552 ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
2553 if (ret) {
2554 ib_destroy_cm_id(id_priv->cm_id.ib);
2555 id_priv->cm_id.ib = NULL;
2556 }
2557out:
2558 kfree(req.private_data);
2559 return ret;
2560}
2561
2562static int cma_connect_ib(struct rdma_id_private *id_priv,
2563 struct rdma_conn_param *conn_param)
2564{
2565 struct ib_cm_req_param req;
2566 struct rdma_route *route;
2567 void *private_data;
2568 struct ib_cm_id *id;
2569 int offset, ret;
2570
2571 memset(&req, 0, sizeof req);
2572 offset = cma_user_data_offset(id_priv->id.ps);
2573 req.private_data_len = offset + conn_param->private_data_len;
2574 if (req.private_data_len < conn_param->private_data_len)
2575 return -EINVAL;
2576
2577 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2578 if (!private_data)
2579 return -ENOMEM;
2580
2581 if (conn_param->private_data && conn_param->private_data_len)
2582 memcpy(private_data + offset, conn_param->private_data,
2583 conn_param->private_data_len);
2584
2585 id = ib_create_cm_id(id_priv->id.device, cma_ib_handler, id_priv);
2586 if (IS_ERR(id)) {
2587 ret = PTR_ERR(id);
2588 goto out;
2589 }
2590 id_priv->cm_id.ib = id;
2591
2592 route = &id_priv->id.route;
2593 ret = cma_format_hdr(private_data, id_priv->id.ps, route);
2594 if (ret)
2595 goto out;
2596 req.private_data = private_data;
2597
2598 req.primary_path = &route->path_rec[0];
2599 if (route->num_paths == 2)
2600 req.alternate_path = &route->path_rec[1];
2601
2602 req.service_id = cma_get_service_id(id_priv->id.ps,
2603 (struct sockaddr *) &route->addr.dst_addr);
2604 req.qp_num = id_priv->qp_num;
2605 req.qp_type = id_priv->id.qp_type;
2606 req.starting_psn = id_priv->seq_num;
2607 req.responder_resources = conn_param->responder_resources;
2608 req.initiator_depth = conn_param->initiator_depth;
2609 req.flow_control = conn_param->flow_control;
2610 req.retry_count = conn_param->retry_count;
2611 req.rnr_retry_count = conn_param->rnr_retry_count;
2612 req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2613 req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2614 req.max_cm_retries = CMA_MAX_CM_RETRIES;
2615 req.srq = id_priv->srq ? 1 : 0;
2616
2617 ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
2618out:
2619 if (ret && !IS_ERR(id)) {
2620 ib_destroy_cm_id(id);
2621 id_priv->cm_id.ib = NULL;
2622 }
2623
2624 kfree(private_data);
2625 return ret;
2626}
2627
2628static int cma_connect_iw(struct rdma_id_private *id_priv,
2629 struct rdma_conn_param *conn_param)
2630{
2631 struct iw_cm_id *cm_id;
2632 struct sockaddr_in* sin;
2633 int ret;
2634 struct iw_cm_conn_param iw_param;
2635
2636 cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
2637 if (IS_ERR(cm_id))
2638 return PTR_ERR(cm_id);
2639
2640 id_priv->cm_id.iw = cm_id;
2641
2642 sin = (struct sockaddr_in*) &id_priv->id.route.addr.src_addr;
2643 cm_id->local_addr = *sin;
2644
2645 sin = (struct sockaddr_in*) &id_priv->id.route.addr.dst_addr;
2646 cm_id->remote_addr = *sin;
2647
2648 ret = cma_modify_qp_rtr(id_priv, conn_param);
2649 if (ret)
2650 goto out;
2651
2652 if (conn_param) {
2653 iw_param.ord = conn_param->initiator_depth;
2654 iw_param.ird = conn_param->responder_resources;
2655 iw_param.private_data = conn_param->private_data;
2656 iw_param.private_data_len = conn_param->private_data_len;
2657 iw_param.qpn = id_priv->id.qp ? id_priv->qp_num : conn_param->qp_num;
2658 } else {
2659 memset(&iw_param, 0, sizeof iw_param);
2660 iw_param.qpn = id_priv->qp_num;
2661 }
2662 ret = iw_cm_connect(cm_id, &iw_param);
2663out:
2664 if (ret) {
2665 iw_destroy_cm_id(cm_id);
2666 id_priv->cm_id.iw = NULL;
2667 }
2668 return ret;
2669}
2670
2671int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2672{
2673 struct rdma_id_private *id_priv;
2674 int ret;
2675
2676 id_priv = container_of(id, struct rdma_id_private, id);
2677 if (!cma_comp_exch(id_priv, RDMA_CM_ROUTE_RESOLVED, RDMA_CM_CONNECT))
2678 return -EINVAL;
2679
2680 if (!id->qp) {
2681 id_priv->qp_num = conn_param->qp_num;
2682 id_priv->srq = conn_param->srq;
2683 }
2684
2685 switch (rdma_node_get_transport(id->device->node_type)) {
2686 case RDMA_TRANSPORT_IB:
2687 if (id->qp_type == IB_QPT_UD)
2688 ret = cma_resolve_ib_udp(id_priv, conn_param);
2689 else
2690 ret = cma_connect_ib(id_priv, conn_param);
2691 break;
2692 case RDMA_TRANSPORT_IWARP:
2693 ret = cma_connect_iw(id_priv, conn_param);
2694 break;
2695 default:
2696 ret = -ENOSYS;
2697 break;
2698 }
2699 if (ret)
2700 goto err;
2701
2702 return 0;
2703err:
2704 cma_comp_exch(id_priv, RDMA_CM_CONNECT, RDMA_CM_ROUTE_RESOLVED);
2705 return ret;
2706}
2707EXPORT_SYMBOL(rdma_connect);
2708
2709static int cma_accept_ib(struct rdma_id_private *id_priv,
2710 struct rdma_conn_param *conn_param)
2711{
2712 struct ib_cm_rep_param rep;
2713 int ret;
2714
2715 ret = cma_modify_qp_rtr(id_priv, conn_param);
2716 if (ret)
2717 goto out;
2718
2719 ret = cma_modify_qp_rts(id_priv, conn_param);
2720 if (ret)
2721 goto out;
2722
2723 memset(&rep, 0, sizeof rep);
2724 rep.qp_num = id_priv->qp_num;
2725 rep.starting_psn = id_priv->seq_num;
2726 rep.private_data = conn_param->private_data;
2727 rep.private_data_len = conn_param->private_data_len;
2728 rep.responder_resources = conn_param->responder_resources;
2729 rep.initiator_depth = conn_param->initiator_depth;
2730 rep.failover_accepted = 0;
2731 rep.flow_control = conn_param->flow_control;
2732 rep.rnr_retry_count = conn_param->rnr_retry_count;
2733 rep.srq = id_priv->srq ? 1 : 0;
2734
2735 ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
2736out:
2737 return ret;
2738}
2739
2740static int cma_accept_iw(struct rdma_id_private *id_priv,
2741 struct rdma_conn_param *conn_param)
2742{
2743 struct iw_cm_conn_param iw_param;
2744 int ret;
2745
2746 ret = cma_modify_qp_rtr(id_priv, conn_param);
2747 if (ret)
2748 return ret;
2749
2750 iw_param.ord = conn_param->initiator_depth;
2751 iw_param.ird = conn_param->responder_resources;
2752 iw_param.private_data = conn_param->private_data;
2753 iw_param.private_data_len = conn_param->private_data_len;
2754 if (id_priv->id.qp) {
2755 iw_param.qpn = id_priv->qp_num;
2756 } else
2757 iw_param.qpn = conn_param->qp_num;
2758
2759 return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
2760}
2761
2762static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
2763 enum ib_cm_sidr_status status,
2764 const void *private_data, int private_data_len)
2765{
2766 struct ib_cm_sidr_rep_param rep;
2767 int ret;
2768
2769 memset(&rep, 0, sizeof rep);
2770 rep.status = status;
2771 if (status == IB_SIDR_SUCCESS) {
2772 ret = cma_set_qkey(id_priv);
2773 if (ret)
2774 return ret;
2775 rep.qp_num = id_priv->qp_num;
2776 rep.qkey = id_priv->qkey;
2777 }
2778 rep.private_data = private_data;
2779 rep.private_data_len = private_data_len;
2780
2781 return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
2782}
2783
2784int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2785{
2786 struct rdma_id_private *id_priv;
2787 int ret;
2788
2789 id_priv = container_of(id, struct rdma_id_private, id);
2790
2791 id_priv->owner = task_pid_nr(current);
2792
2793 if (!cma_comp(id_priv, RDMA_CM_CONNECT))
2794 return -EINVAL;
2795
2796 if (!id->qp && conn_param) {
2797 id_priv->qp_num = conn_param->qp_num;
2798 id_priv->srq = conn_param->srq;
2799 }
2800
2801 switch (rdma_node_get_transport(id->device->node_type)) {
2802 case RDMA_TRANSPORT_IB:
2803 if (id->qp_type == IB_QPT_UD) {
2804 if (conn_param)
2805 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
2806 conn_param->private_data,
2807 conn_param->private_data_len);
2808 else
2809 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
2810 NULL, 0);
2811 } else {
2812 if (conn_param)
2813 ret = cma_accept_ib(id_priv, conn_param);
2814 else
2815 ret = cma_rep_recv(id_priv);
2816 }
2817 break;
2818 case RDMA_TRANSPORT_IWARP:
2819 ret = cma_accept_iw(id_priv, conn_param);
2820 break;
2821 default:
2822 ret = -ENOSYS;
2823 break;
2824 }
2825
2826 if (ret)
2827 goto reject;
2828
2829 return 0;
2830reject:
2831 cma_modify_qp_err(id_priv);
2832 rdma_reject(id, NULL, 0);
2833 return ret;
2834}
2835EXPORT_SYMBOL(rdma_accept);
2836
2837int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
2838{
2839 struct rdma_id_private *id_priv;
2840 int ret;
2841
2842 id_priv = container_of(id, struct rdma_id_private, id);
2843 if (!id_priv->cm_id.ib)
2844 return -EINVAL;
2845
2846 switch (id->device->node_type) {
2847 case RDMA_NODE_IB_CA:
2848 ret = ib_cm_notify(id_priv->cm_id.ib, event);
2849 break;
2850 default:
2851 ret = 0;
2852 break;
2853 }
2854 return ret;
2855}
2856EXPORT_SYMBOL(rdma_notify);
2857
2858int rdma_reject(struct rdma_cm_id *id, const void *private_data,
2859 u8 private_data_len)
2860{
2861 struct rdma_id_private *id_priv;
2862 int ret;
2863
2864 id_priv = container_of(id, struct rdma_id_private, id);
2865 if (!id_priv->cm_id.ib)
2866 return -EINVAL;
2867
2868 switch (rdma_node_get_transport(id->device->node_type)) {
2869 case RDMA_TRANSPORT_IB:
2870 if (id->qp_type == IB_QPT_UD)
2871 ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT,
2872 private_data, private_data_len);
2873 else
2874 ret = ib_send_cm_rej(id_priv->cm_id.ib,
2875 IB_CM_REJ_CONSUMER_DEFINED, NULL,
2876 0, private_data, private_data_len);
2877 break;
2878 case RDMA_TRANSPORT_IWARP:
2879 ret = iw_cm_reject(id_priv->cm_id.iw,
2880 private_data, private_data_len);
2881 break;
2882 default:
2883 ret = -ENOSYS;
2884 break;
2885 }
2886 return ret;
2887}
2888EXPORT_SYMBOL(rdma_reject);
2889
2890int rdma_disconnect(struct rdma_cm_id *id)
2891{
2892 struct rdma_id_private *id_priv;
2893 int ret;
2894
2895 id_priv = container_of(id, struct rdma_id_private, id);
2896 if (!id_priv->cm_id.ib)
2897 return -EINVAL;
2898
2899 switch (rdma_node_get_transport(id->device->node_type)) {
2900 case RDMA_TRANSPORT_IB:
2901 ret = cma_modify_qp_err(id_priv);
2902 if (ret)
2903 goto out;
2904 /* Initiate or respond to a disconnect. */
2905 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
2906 ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
2907 break;
2908 case RDMA_TRANSPORT_IWARP:
2909 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
2910 break;
2911 default:
2912 ret = -EINVAL;
2913 break;
2914 }
2915out:
2916 return ret;
2917}
2918EXPORT_SYMBOL(rdma_disconnect);
2919
2920static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
2921{
2922 struct rdma_id_private *id_priv;
2923 struct cma_multicast *mc = multicast->context;
2924 struct rdma_cm_event event;
2925 int ret;
2926
2927 id_priv = mc->id_priv;
2928 if (cma_disable_callback(id_priv, RDMA_CM_ADDR_BOUND) &&
2929 cma_disable_callback(id_priv, RDMA_CM_ADDR_RESOLVED))
2930 return 0;
2931
2932 mutex_lock(&id_priv->qp_mutex);
2933 if (!status && id_priv->id.qp)
2934 status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid,
2935 be16_to_cpu(multicast->rec.mlid));
2936 mutex_unlock(&id_priv->qp_mutex);
2937
2938 memset(&event, 0, sizeof event);
2939 event.status = status;
2940 event.param.ud.private_data = mc->context;
2941 if (!status) {
2942 event.event = RDMA_CM_EVENT_MULTICAST_JOIN;
2943 ib_init_ah_from_mcmember(id_priv->id.device,
2944 id_priv->id.port_num, &multicast->rec,
2945 &event.param.ud.ah_attr);
2946 event.param.ud.qp_num = 0xFFFFFF;
2947 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
2948 } else
2949 event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
2950
2951 ret = id_priv->id.event_handler(&id_priv->id, &event);
2952 if (ret) {
2953 cma_exch(id_priv, RDMA_CM_DESTROYING);
2954 mutex_unlock(&id_priv->handler_mutex);
2955 rdma_destroy_id(&id_priv->id);
2956 return 0;
2957 }
2958
2959 mutex_unlock(&id_priv->handler_mutex);
2960 return 0;
2961}
2962
2963static void cma_set_mgid(struct rdma_id_private *id_priv,
2964 struct sockaddr *addr, union ib_gid *mgid)
2965{
2966 unsigned char mc_map[MAX_ADDR_LEN];
2967 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2968 struct sockaddr_in *sin = (struct sockaddr_in *) addr;
2969 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
2970
2971 if (cma_any_addr(addr)) {
2972 memset(mgid, 0, sizeof *mgid);
2973 } else if ((addr->sa_family == AF_INET6) &&
2974 ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFFF0FFFF) ==
2975 0xFF10A01B)) {
2976 /* IPv6 address is an SA assigned MGID. */
2977 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
2978 } else if ((addr->sa_family == AF_INET6)) {
2979 ipv6_ib_mc_map(&sin6->sin6_addr, dev_addr->broadcast, mc_map);
2980 if (id_priv->id.ps == RDMA_PS_UDP)
2981 mc_map[7] = 0x01; /* Use RDMA CM signature */
2982 *mgid = *(union ib_gid *) (mc_map + 4);
2983 } else {
2984 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
2985 if (id_priv->id.ps == RDMA_PS_UDP)
2986 mc_map[7] = 0x01; /* Use RDMA CM signature */
2987 *mgid = *(union ib_gid *) (mc_map + 4);
2988 }
2989}
2990
2991static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
2992 struct cma_multicast *mc)
2993{
2994 struct ib_sa_mcmember_rec rec;
2995 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2996 ib_sa_comp_mask comp_mask;
2997 int ret;
2998
2999 ib_addr_get_mgid(dev_addr, &rec.mgid);
3000 ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
3001 &rec.mgid, &rec);
3002 if (ret)
3003 return ret;
3004
3005 cma_set_mgid(id_priv, (struct sockaddr *) &mc->addr, &rec.mgid);
3006 if (id_priv->id.ps == RDMA_PS_UDP)
3007 rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
3008 rdma_addr_get_sgid(dev_addr, &rec.port_gid);
3009 rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
3010 rec.join_state = 1;
3011
3012 comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
3013 IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
3014 IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
3015 IB_SA_MCMEMBER_REC_FLOW_LABEL |
3016 IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
3017
3018 if (id_priv->id.ps == RDMA_PS_IPOIB)
3019 comp_mask |= IB_SA_MCMEMBER_REC_RATE |
3020 IB_SA_MCMEMBER_REC_RATE_SELECTOR;
3021
3022 mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device,
3023 id_priv->id.port_num, &rec,
3024 comp_mask, GFP_KERNEL,
3025 cma_ib_mc_handler, mc);
3026 if (IS_ERR(mc->multicast.ib))
3027 return PTR_ERR(mc->multicast.ib);
3028
3029 return 0;
3030}
3031
3032static void iboe_mcast_work_handler(struct work_struct *work)
3033{
3034 struct iboe_mcast_work *mw = container_of(work, struct iboe_mcast_work, work);
3035 struct cma_multicast *mc = mw->mc;
3036 struct ib_sa_multicast *m = mc->multicast.ib;
3037
3038 mc->multicast.ib->context = mc;
3039 cma_ib_mc_handler(0, m);
3040 kref_put(&mc->mcref, release_mc);
3041 kfree(mw);
3042}
3043
3044static void cma_iboe_set_mgid(struct sockaddr *addr, union ib_gid *mgid)
3045{
3046 struct sockaddr_in *sin = (struct sockaddr_in *)addr;
3047 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)addr;
3048
3049 if (cma_any_addr(addr)) {
3050 memset(mgid, 0, sizeof *mgid);
3051 } else if (addr->sa_family == AF_INET6) {
3052 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
3053 } else {
3054 mgid->raw[0] = 0xff;
3055 mgid->raw[1] = 0x0e;
3056 mgid->raw[2] = 0;
3057 mgid->raw[3] = 0;
3058 mgid->raw[4] = 0;
3059 mgid->raw[5] = 0;
3060 mgid->raw[6] = 0;
3061 mgid->raw[7] = 0;
3062 mgid->raw[8] = 0;
3063 mgid->raw[9] = 0;
3064 mgid->raw[10] = 0xff;
3065 mgid->raw[11] = 0xff;
3066 *(__be32 *)(&mgid->raw[12]) = sin->sin_addr.s_addr;
3067 }
3068}
3069
3070static int cma_iboe_join_multicast(struct rdma_id_private *id_priv,
3071 struct cma_multicast *mc)
3072{
3073 struct iboe_mcast_work *work;
3074 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
3075 int err;
3076 struct sockaddr *addr = (struct sockaddr *)&mc->addr;
3077 struct net_device *ndev = NULL;
3078
3079 if (cma_zero_addr((struct sockaddr *)&mc->addr))
3080 return -EINVAL;
3081
3082 work = kzalloc(sizeof *work, GFP_KERNEL);
3083 if (!work)
3084 return -ENOMEM;
3085
3086 mc->multicast.ib = kzalloc(sizeof(struct ib_sa_multicast), GFP_KERNEL);
3087 if (!mc->multicast.ib) {
3088 err = -ENOMEM;
3089 goto out1;
3090 }
3091
3092 cma_iboe_set_mgid(addr, &mc->multicast.ib->rec.mgid);
3093
3094 mc->multicast.ib->rec.pkey = cpu_to_be16(0xffff);
3095 if (id_priv->id.ps == RDMA_PS_UDP)
3096 mc->multicast.ib->rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
3097
3098 if (dev_addr->bound_dev_if)
3099 ndev = dev_get_by_index(&init_net, dev_addr->bound_dev_if);
3100 if (!ndev) {
3101 err = -ENODEV;
3102 goto out2;
3103 }
3104 mc->multicast.ib->rec.rate = iboe_get_rate(ndev);
3105 mc->multicast.ib->rec.hop_limit = 1;
3106 mc->multicast.ib->rec.mtu = iboe_get_mtu(ndev->mtu);
3107 dev_put(ndev);
3108 if (!mc->multicast.ib->rec.mtu) {
3109 err = -EINVAL;
3110 goto out2;
3111 }
3112 iboe_addr_get_sgid(dev_addr, &mc->multicast.ib->rec.port_gid);
3113 work->id = id_priv;
3114 work->mc = mc;
3115 INIT_WORK(&work->work, iboe_mcast_work_handler);
3116 kref_get(&mc->mcref);
3117 queue_work(cma_wq, &work->work);
3118
3119 return 0;
3120
3121out2:
3122 kfree(mc->multicast.ib);
3123out1:
3124 kfree(work);
3125 return err;
3126}
3127
3128int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
3129 void *context)
3130{
3131 struct rdma_id_private *id_priv;
3132 struct cma_multicast *mc;
3133 int ret;
3134
3135 id_priv = container_of(id, struct rdma_id_private, id);
3136 if (!cma_comp(id_priv, RDMA_CM_ADDR_BOUND) &&
3137 !cma_comp(id_priv, RDMA_CM_ADDR_RESOLVED))
3138 return -EINVAL;
3139
3140 mc = kmalloc(sizeof *mc, GFP_KERNEL);
3141 if (!mc)
3142 return -ENOMEM;
3143
3144 memcpy(&mc->addr, addr, ip_addr_size(addr));
3145 mc->context = context;
3146 mc->id_priv = id_priv;
3147
3148 spin_lock(&id_priv->lock);
3149 list_add(&mc->list, &id_priv->mc_list);
3150 spin_unlock(&id_priv->lock);
3151
3152 switch (rdma_node_get_transport(id->device->node_type)) {
3153 case RDMA_TRANSPORT_IB:
3154 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
3155 case IB_LINK_LAYER_INFINIBAND:
3156 ret = cma_join_ib_multicast(id_priv, mc);
3157 break;
3158 case IB_LINK_LAYER_ETHERNET:
3159 kref_init(&mc->mcref);
3160 ret = cma_iboe_join_multicast(id_priv, mc);
3161 break;
3162 default:
3163 ret = -EINVAL;
3164 }
3165 break;
3166 default:
3167 ret = -ENOSYS;
3168 break;
3169 }
3170
3171 if (ret) {
3172 spin_lock_irq(&id_priv->lock);
3173 list_del(&mc->list);
3174 spin_unlock_irq(&id_priv->lock);
3175 kfree(mc);
3176 }
3177 return ret;
3178}
3179EXPORT_SYMBOL(rdma_join_multicast);
3180
3181void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
3182{
3183 struct rdma_id_private *id_priv;
3184 struct cma_multicast *mc;
3185
3186 id_priv = container_of(id, struct rdma_id_private, id);
3187 spin_lock_irq(&id_priv->lock);
3188 list_for_each_entry(mc, &id_priv->mc_list, list) {
3189 if (!memcmp(&mc->addr, addr, ip_addr_size(addr))) {
3190 list_del(&mc->list);
3191 spin_unlock_irq(&id_priv->lock);
3192
3193 if (id->qp)
3194 ib_detach_mcast(id->qp,
3195 &mc->multicast.ib->rec.mgid,
3196 be16_to_cpu(mc->multicast.ib->rec.mlid));
3197 if (rdma_node_get_transport(id_priv->cma_dev->device->node_type) == RDMA_TRANSPORT_IB) {
3198 switch (rdma_port_get_link_layer(id->device, id->port_num)) {
3199 case IB_LINK_LAYER_INFINIBAND:
3200 ib_sa_free_multicast(mc->multicast.ib);
3201 kfree(mc);
3202 break;
3203 case IB_LINK_LAYER_ETHERNET:
3204 kref_put(&mc->mcref, release_mc);
3205 break;
3206 default:
3207 break;
3208 }
3209 }
3210 return;
3211 }
3212 }
3213 spin_unlock_irq(&id_priv->lock);
3214}
3215EXPORT_SYMBOL(rdma_leave_multicast);
3216
3217static int cma_netdev_change(struct net_device *ndev, struct rdma_id_private *id_priv)
3218{
3219 struct rdma_dev_addr *dev_addr;
3220 struct cma_ndev_work *work;
3221
3222 dev_addr = &id_priv->id.route.addr.dev_addr;
3223
3224 if ((dev_addr->bound_dev_if == ndev->ifindex) &&
3225 memcmp(dev_addr->src_dev_addr, ndev->dev_addr, ndev->addr_len)) {
3226 printk(KERN_INFO "RDMA CM addr change for ndev %s used by id %p\n",
3227 ndev->name, &id_priv->id);
3228 work = kzalloc(sizeof *work, GFP_KERNEL);
3229 if (!work)
3230 return -ENOMEM;
3231
3232 INIT_WORK(&work->work, cma_ndev_work_handler);
3233 work->id = id_priv;
3234 work->event.event = RDMA_CM_EVENT_ADDR_CHANGE;
3235 atomic_inc(&id_priv->refcount);
3236 queue_work(cma_wq, &work->work);
3237 }
3238
3239 return 0;
3240}
3241
3242static int cma_netdev_callback(struct notifier_block *self, unsigned long event,
3243 void *ctx)
3244{
3245 struct net_device *ndev = (struct net_device *)ctx;
3246 struct cma_device *cma_dev;
3247 struct rdma_id_private *id_priv;
3248 int ret = NOTIFY_DONE;
3249
3250 if (dev_net(ndev) != &init_net)
3251 return NOTIFY_DONE;
3252
3253 if (event != NETDEV_BONDING_FAILOVER)
3254 return NOTIFY_DONE;
3255
3256 if (!(ndev->flags & IFF_MASTER) || !(ndev->priv_flags & IFF_BONDING))
3257 return NOTIFY_DONE;
3258
3259 mutex_lock(&lock);
3260 list_for_each_entry(cma_dev, &dev_list, list)
3261 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
3262 ret = cma_netdev_change(ndev, id_priv);
3263 if (ret)
3264 goto out;
3265 }
3266
3267out:
3268 mutex_unlock(&lock);
3269 return ret;
3270}
3271
3272static struct notifier_block cma_nb = {
3273 .notifier_call = cma_netdev_callback
3274};
3275
3276static void cma_add_one(struct ib_device *device)
3277{
3278 struct cma_device *cma_dev;
3279 struct rdma_id_private *id_priv;
3280
3281 cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
3282 if (!cma_dev)
3283 return;
3284
3285 cma_dev->device = device;
3286
3287 init_completion(&cma_dev->comp);
3288 atomic_set(&cma_dev->refcount, 1);
3289 INIT_LIST_HEAD(&cma_dev->id_list);
3290 ib_set_client_data(device, &cma_client, cma_dev);
3291
3292 mutex_lock(&lock);
3293 list_add_tail(&cma_dev->list, &dev_list);
3294 list_for_each_entry(id_priv, &listen_any_list, list)
3295 cma_listen_on_dev(id_priv, cma_dev);
3296 mutex_unlock(&lock);
3297}
3298
3299static int cma_remove_id_dev(struct rdma_id_private *id_priv)
3300{
3301 struct rdma_cm_event event;
3302 enum rdma_cm_state state;
3303 int ret = 0;
3304
3305 /* Record that we want to remove the device */
3306 state = cma_exch(id_priv, RDMA_CM_DEVICE_REMOVAL);
3307 if (state == RDMA_CM_DESTROYING)
3308 return 0;
3309
3310 cma_cancel_operation(id_priv, state);
3311 mutex_lock(&id_priv->handler_mutex);
3312
3313 /* Check for destruction from another callback. */
3314 if (!cma_comp(id_priv, RDMA_CM_DEVICE_REMOVAL))
3315 goto out;
3316
3317 memset(&event, 0, sizeof event);
3318 event.event = RDMA_CM_EVENT_DEVICE_REMOVAL;
3319 ret = id_priv->id.event_handler(&id_priv->id, &event);
3320out:
3321 mutex_unlock(&id_priv->handler_mutex);
3322 return ret;
3323}
3324
3325static void cma_process_remove(struct cma_device *cma_dev)
3326{
3327 struct rdma_id_private *id_priv;
3328 int ret;
3329
3330 mutex_lock(&lock);
3331 while (!list_empty(&cma_dev->id_list)) {
3332 id_priv = list_entry(cma_dev->id_list.next,
3333 struct rdma_id_private, list);
3334
3335 list_del(&id_priv->listen_list);
3336 list_del_init(&id_priv->list);
3337 atomic_inc(&id_priv->refcount);
3338 mutex_unlock(&lock);
3339
3340 ret = id_priv->internal_id ? 1 : cma_remove_id_dev(id_priv);
3341 cma_deref_id(id_priv);
3342 if (ret)
3343 rdma_destroy_id(&id_priv->id);
3344
3345 mutex_lock(&lock);
3346 }
3347 mutex_unlock(&lock);
3348
3349 cma_deref_dev(cma_dev);
3350 wait_for_completion(&cma_dev->comp);
3351}
3352
3353static void cma_remove_one(struct ib_device *device)
3354{
3355 struct cma_device *cma_dev;
3356
3357 cma_dev = ib_get_client_data(device, &cma_client);
3358 if (!cma_dev)
3359 return;
3360
3361 mutex_lock(&lock);
3362 list_del(&cma_dev->list);
3363 mutex_unlock(&lock);
3364
3365 cma_process_remove(cma_dev);
3366 kfree(cma_dev);
3367}
3368
3369static int cma_get_id_stats(struct sk_buff *skb, struct netlink_callback *cb)
3370{
3371 struct nlmsghdr *nlh;
3372 struct rdma_cm_id_stats *id_stats;
3373 struct rdma_id_private *id_priv;
3374 struct rdma_cm_id *id = NULL;
3375 struct cma_device *cma_dev;
3376 int i_dev = 0, i_id = 0;
3377
3378 /*
3379 * We export all of the IDs as a sequence of messages. Each
3380 * ID gets its own netlink message.
3381 */
3382 mutex_lock(&lock);
3383
3384 list_for_each_entry(cma_dev, &dev_list, list) {
3385 if (i_dev < cb->args[0]) {
3386 i_dev++;
3387 continue;
3388 }
3389
3390 i_id = 0;
3391 list_for_each_entry(id_priv, &cma_dev->id_list, list) {
3392 if (i_id < cb->args[1]) {
3393 i_id++;
3394 continue;
3395 }
3396
3397 id_stats = ibnl_put_msg(skb, &nlh, cb->nlh->nlmsg_seq,
3398 sizeof *id_stats, RDMA_NL_RDMA_CM,
3399 RDMA_NL_RDMA_CM_ID_STATS);
3400 if (!id_stats)
3401 goto out;
3402
3403 memset(id_stats, 0, sizeof *id_stats);
3404 id = &id_priv->id;
3405 id_stats->node_type = id->route.addr.dev_addr.dev_type;
3406 id_stats->port_num = id->port_num;
3407 id_stats->bound_dev_if =
3408 id->route.addr.dev_addr.bound_dev_if;
3409
3410 if (id->route.addr.src_addr.ss_family == AF_INET) {
3411 if (ibnl_put_attr(skb, nlh,
3412 sizeof(struct sockaddr_in),
3413 &id->route.addr.src_addr,
3414 RDMA_NL_RDMA_CM_ATTR_SRC_ADDR)) {
3415 goto out;
3416 }
3417 if (ibnl_put_attr(skb, nlh,
3418 sizeof(struct sockaddr_in),
3419 &id->route.addr.dst_addr,
3420 RDMA_NL_RDMA_CM_ATTR_DST_ADDR)) {
3421 goto out;
3422 }
3423 } else if (id->route.addr.src_addr.ss_family == AF_INET6) {
3424 if (ibnl_put_attr(skb, nlh,
3425 sizeof(struct sockaddr_in6),
3426 &id->route.addr.src_addr,
3427 RDMA_NL_RDMA_CM_ATTR_SRC_ADDR)) {
3428 goto out;
3429 }
3430 if (ibnl_put_attr(skb, nlh,
3431 sizeof(struct sockaddr_in6),
3432 &id->route.addr.dst_addr,
3433 RDMA_NL_RDMA_CM_ATTR_DST_ADDR)) {
3434 goto out;
3435 }
3436 }
3437
3438 id_stats->pid = id_priv->owner;
3439 id_stats->port_space = id->ps;
3440 id_stats->cm_state = id_priv->state;
3441 id_stats->qp_num = id_priv->qp_num;
3442 id_stats->qp_type = id->qp_type;
3443
3444 i_id++;
3445 }
3446
3447 cb->args[1] = 0;
3448 i_dev++;
3449 }
3450
3451out:
3452 mutex_unlock(&lock);
3453 cb->args[0] = i_dev;
3454 cb->args[1] = i_id;
3455
3456 return skb->len;
3457}
3458
3459static const struct ibnl_client_cbs cma_cb_table[] = {
3460 [RDMA_NL_RDMA_CM_ID_STATS] = { .dump = cma_get_id_stats },
3461};
3462
3463static int __init cma_init(void)
3464{
3465 int ret;
3466
3467 cma_wq = create_singlethread_workqueue("rdma_cm");
3468 if (!cma_wq)
3469 return -ENOMEM;
3470
3471 ib_sa_register_client(&sa_client);
3472 rdma_addr_register_client(&addr_client);
3473 register_netdevice_notifier(&cma_nb);
3474
3475 ret = ib_register_client(&cma_client);
3476 if (ret)
3477 goto err;
3478
3479 if (ibnl_add_client(RDMA_NL_RDMA_CM, RDMA_NL_RDMA_CM_NUM_OPS, cma_cb_table))
3480 printk(KERN_WARNING "RDMA CMA: failed to add netlink callback\n");
3481
3482 return 0;
3483
3484err:
3485 unregister_netdevice_notifier(&cma_nb);
3486 rdma_addr_unregister_client(&addr_client);
3487 ib_sa_unregister_client(&sa_client);
3488 destroy_workqueue(cma_wq);
3489 return ret;
3490}
3491
3492static void __exit cma_cleanup(void)
3493{
3494 ibnl_remove_client(RDMA_NL_RDMA_CM);
3495 ib_unregister_client(&cma_client);
3496 unregister_netdevice_notifier(&cma_nb);
3497 rdma_addr_unregister_client(&addr_client);
3498 ib_sa_unregister_client(&sa_client);
3499 destroy_workqueue(cma_wq);
3500 idr_destroy(&sdp_ps);
3501 idr_destroy(&tcp_ps);
3502 idr_destroy(&udp_ps);
3503 idr_destroy(&ipoib_ps);
3504 idr_destroy(&ib_ps);
3505}
3506
3507module_init(cma_init);
3508module_exit(cma_cleanup);