ARM: OMAP: Add 24xx GPIO debounce support
[linux-ginger.git] / drivers / infiniband / core / cma.c
blob1eff1b2c0e08511d6de424381eb33fc5eb668b00
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.
7 * This Software is licensed under one of the following licenses:
9 * 1) under the terms of the "Common Public License 1.0" a copy of which is
10 * available from the Open Source Initiative, see
11 * http://www.opensource.org/licenses/cpl.php.
13 * 2) under the terms of the "The BSD License" a copy of which is
14 * available from the Open Source Initiative, see
15 * http://www.opensource.org/licenses/bsd-license.php.
17 * 3) under the terms of the "GNU General Public License (GPL) Version 2" a
18 * copy of which is available from the Open Source Initiative, see
19 * http://www.opensource.org/licenses/gpl-license.php.
21 * Licensee has the right to choose one of the above licenses.
23 * Redistributions of source code must retain the above copyright
24 * notice and one of the license notices.
26 * Redistributions in binary form must reproduce both the above copyright
27 * notice, one of the license notices in the documentation
28 * and/or other materials provided with the distribution.
32 #include <linux/completion.h>
33 #include <linux/in.h>
34 #include <linux/in6.h>
35 #include <linux/mutex.h>
36 #include <linux/random.h>
37 #include <linux/idr.h>
38 #include <linux/inetdevice.h>
40 #include <net/tcp.h>
42 #include <rdma/rdma_cm.h>
43 #include <rdma/rdma_cm_ib.h>
44 #include <rdma/ib_cache.h>
45 #include <rdma/ib_cm.h>
46 #include <rdma/ib_sa.h>
47 #include <rdma/iw_cm.h>
49 MODULE_AUTHOR("Sean Hefty");
50 MODULE_DESCRIPTION("Generic RDMA CM Agent");
51 MODULE_LICENSE("Dual BSD/GPL");
53 #define CMA_CM_RESPONSE_TIMEOUT 20
54 #define CMA_MAX_CM_RETRIES 15
55 #define CMA_CM_MRA_SETTING (IB_CM_MRA_FLAG_DELAY | 24)
57 static void cma_add_one(struct ib_device *device);
58 static void cma_remove_one(struct ib_device *device);
60 static struct ib_client cma_client = {
61 .name = "cma",
62 .add = cma_add_one,
63 .remove = cma_remove_one
66 static struct ib_sa_client sa_client;
67 static struct rdma_addr_client addr_client;
68 static LIST_HEAD(dev_list);
69 static LIST_HEAD(listen_any_list);
70 static DEFINE_MUTEX(lock);
71 static struct workqueue_struct *cma_wq;
72 static DEFINE_IDR(sdp_ps);
73 static DEFINE_IDR(tcp_ps);
74 static DEFINE_IDR(udp_ps);
75 static DEFINE_IDR(ipoib_ps);
76 static int next_port;
78 struct cma_device {
79 struct list_head list;
80 struct ib_device *device;
81 struct completion comp;
82 atomic_t refcount;
83 struct list_head id_list;
86 enum cma_state {
87 CMA_IDLE,
88 CMA_ADDR_QUERY,
89 CMA_ADDR_RESOLVED,
90 CMA_ROUTE_QUERY,
91 CMA_ROUTE_RESOLVED,
92 CMA_CONNECT,
93 CMA_DISCONNECT,
94 CMA_ADDR_BOUND,
95 CMA_LISTEN,
96 CMA_DEVICE_REMOVAL,
97 CMA_DESTROYING
100 struct rdma_bind_list {
101 struct idr *ps;
102 struct hlist_head owners;
103 unsigned short port;
107 * Device removal can occur at anytime, so we need extra handling to
108 * serialize notifying the user of device removal with other callbacks.
109 * We do this by disabling removal notification while a callback is in process,
110 * and reporting it after the callback completes.
112 struct rdma_id_private {
113 struct rdma_cm_id id;
115 struct rdma_bind_list *bind_list;
116 struct hlist_node node;
117 struct list_head list; /* listen_any_list or cma_device.list */
118 struct list_head listen_list; /* per device listens */
119 struct cma_device *cma_dev;
120 struct list_head mc_list;
122 int internal_id;
123 enum cma_state state;
124 spinlock_t lock;
125 struct mutex qp_mutex;
127 struct completion comp;
128 atomic_t refcount;
129 wait_queue_head_t wait_remove;
130 atomic_t dev_remove;
132 int backlog;
133 int timeout_ms;
134 struct ib_sa_query *query;
135 int query_id;
136 union {
137 struct ib_cm_id *ib;
138 struct iw_cm_id *iw;
139 } cm_id;
141 u32 seq_num;
142 u32 qkey;
143 u32 qp_num;
144 u8 srq;
145 u8 tos;
148 struct cma_multicast {
149 struct rdma_id_private *id_priv;
150 union {
151 struct ib_sa_multicast *ib;
152 } multicast;
153 struct list_head list;
154 void *context;
155 struct sockaddr addr;
156 u8 pad[sizeof(struct sockaddr_in6) -
157 sizeof(struct sockaddr)];
160 struct cma_work {
161 struct work_struct work;
162 struct rdma_id_private *id;
163 enum cma_state old_state;
164 enum cma_state new_state;
165 struct rdma_cm_event event;
168 union cma_ip_addr {
169 struct in6_addr ip6;
170 struct {
171 __u32 pad[3];
172 __u32 addr;
173 } ip4;
176 struct cma_hdr {
177 u8 cma_version;
178 u8 ip_version; /* IP version: 7:4 */
179 __u16 port;
180 union cma_ip_addr src_addr;
181 union cma_ip_addr dst_addr;
184 struct sdp_hh {
185 u8 bsdh[16];
186 u8 sdp_version; /* Major version: 7:4 */
187 u8 ip_version; /* IP version: 7:4 */
188 u8 sdp_specific1[10];
189 __u16 port;
190 __u16 sdp_specific2;
191 union cma_ip_addr src_addr;
192 union cma_ip_addr dst_addr;
195 struct sdp_hah {
196 u8 bsdh[16];
197 u8 sdp_version;
200 #define CMA_VERSION 0x00
201 #define SDP_MAJ_VERSION 0x2
203 static int cma_comp(struct rdma_id_private *id_priv, enum cma_state comp)
205 unsigned long flags;
206 int ret;
208 spin_lock_irqsave(&id_priv->lock, flags);
209 ret = (id_priv->state == comp);
210 spin_unlock_irqrestore(&id_priv->lock, flags);
211 return ret;
214 static int cma_comp_exch(struct rdma_id_private *id_priv,
215 enum cma_state comp, enum cma_state exch)
217 unsigned long flags;
218 int ret;
220 spin_lock_irqsave(&id_priv->lock, flags);
221 if ((ret = (id_priv->state == comp)))
222 id_priv->state = exch;
223 spin_unlock_irqrestore(&id_priv->lock, flags);
224 return ret;
227 static enum cma_state cma_exch(struct rdma_id_private *id_priv,
228 enum cma_state exch)
230 unsigned long flags;
231 enum cma_state old;
233 spin_lock_irqsave(&id_priv->lock, flags);
234 old = id_priv->state;
235 id_priv->state = exch;
236 spin_unlock_irqrestore(&id_priv->lock, flags);
237 return old;
240 static inline u8 cma_get_ip_ver(struct cma_hdr *hdr)
242 return hdr->ip_version >> 4;
245 static inline void cma_set_ip_ver(struct cma_hdr *hdr, u8 ip_ver)
247 hdr->ip_version = (ip_ver << 4) | (hdr->ip_version & 0xF);
250 static inline u8 sdp_get_majv(u8 sdp_version)
252 return sdp_version >> 4;
255 static inline u8 sdp_get_ip_ver(struct sdp_hh *hh)
257 return hh->ip_version >> 4;
260 static inline void sdp_set_ip_ver(struct sdp_hh *hh, u8 ip_ver)
262 hh->ip_version = (ip_ver << 4) | (hh->ip_version & 0xF);
265 static inline int cma_is_ud_ps(enum rdma_port_space ps)
267 return (ps == RDMA_PS_UDP || ps == RDMA_PS_IPOIB);
270 static void cma_attach_to_dev(struct rdma_id_private *id_priv,
271 struct cma_device *cma_dev)
273 atomic_inc(&cma_dev->refcount);
274 id_priv->cma_dev = cma_dev;
275 id_priv->id.device = cma_dev->device;
276 list_add_tail(&id_priv->list, &cma_dev->id_list);
279 static inline void cma_deref_dev(struct cma_device *cma_dev)
281 if (atomic_dec_and_test(&cma_dev->refcount))
282 complete(&cma_dev->comp);
285 static void cma_detach_from_dev(struct rdma_id_private *id_priv)
287 list_del(&id_priv->list);
288 cma_deref_dev(id_priv->cma_dev);
289 id_priv->cma_dev = NULL;
292 static int cma_set_qkey(struct ib_device *device, u8 port_num,
293 enum rdma_port_space ps,
294 struct rdma_dev_addr *dev_addr, u32 *qkey)
296 struct ib_sa_mcmember_rec rec;
297 int ret = 0;
299 switch (ps) {
300 case RDMA_PS_UDP:
301 *qkey = RDMA_UDP_QKEY;
302 break;
303 case RDMA_PS_IPOIB:
304 ib_addr_get_mgid(dev_addr, &rec.mgid);
305 ret = ib_sa_get_mcmember_rec(device, port_num, &rec.mgid, &rec);
306 *qkey = be32_to_cpu(rec.qkey);
307 break;
308 default:
309 break;
311 return ret;
314 static int cma_acquire_dev(struct rdma_id_private *id_priv)
316 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
317 struct cma_device *cma_dev;
318 union ib_gid gid;
319 int ret = -ENODEV;
321 switch (rdma_node_get_transport(dev_addr->dev_type)) {
322 case RDMA_TRANSPORT_IB:
323 ib_addr_get_sgid(dev_addr, &gid);
324 break;
325 case RDMA_TRANSPORT_IWARP:
326 iw_addr_get_sgid(dev_addr, &gid);
327 break;
328 default:
329 return -ENODEV;
332 list_for_each_entry(cma_dev, &dev_list, list) {
333 ret = ib_find_cached_gid(cma_dev->device, &gid,
334 &id_priv->id.port_num, NULL);
335 if (!ret) {
336 ret = cma_set_qkey(cma_dev->device,
337 id_priv->id.port_num,
338 id_priv->id.ps, dev_addr,
339 &id_priv->qkey);
340 if (!ret)
341 cma_attach_to_dev(id_priv, cma_dev);
342 break;
345 return ret;
348 static void cma_deref_id(struct rdma_id_private *id_priv)
350 if (atomic_dec_and_test(&id_priv->refcount))
351 complete(&id_priv->comp);
354 static int cma_disable_remove(struct rdma_id_private *id_priv,
355 enum cma_state state)
357 unsigned long flags;
358 int ret;
360 spin_lock_irqsave(&id_priv->lock, flags);
361 if (id_priv->state == state) {
362 atomic_inc(&id_priv->dev_remove);
363 ret = 0;
364 } else
365 ret = -EINVAL;
366 spin_unlock_irqrestore(&id_priv->lock, flags);
367 return ret;
370 static void cma_enable_remove(struct rdma_id_private *id_priv)
372 if (atomic_dec_and_test(&id_priv->dev_remove))
373 wake_up(&id_priv->wait_remove);
376 static int cma_has_cm_dev(struct rdma_id_private *id_priv)
378 return (id_priv->id.device && id_priv->cm_id.ib);
381 struct rdma_cm_id *rdma_create_id(rdma_cm_event_handler event_handler,
382 void *context, enum rdma_port_space ps)
384 struct rdma_id_private *id_priv;
386 id_priv = kzalloc(sizeof *id_priv, GFP_KERNEL);
387 if (!id_priv)
388 return ERR_PTR(-ENOMEM);
390 id_priv->state = CMA_IDLE;
391 id_priv->id.context = context;
392 id_priv->id.event_handler = event_handler;
393 id_priv->id.ps = ps;
394 spin_lock_init(&id_priv->lock);
395 mutex_init(&id_priv->qp_mutex);
396 init_completion(&id_priv->comp);
397 atomic_set(&id_priv->refcount, 1);
398 init_waitqueue_head(&id_priv->wait_remove);
399 atomic_set(&id_priv->dev_remove, 0);
400 INIT_LIST_HEAD(&id_priv->listen_list);
401 INIT_LIST_HEAD(&id_priv->mc_list);
402 get_random_bytes(&id_priv->seq_num, sizeof id_priv->seq_num);
404 return &id_priv->id;
406 EXPORT_SYMBOL(rdma_create_id);
408 static int cma_init_ud_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
410 struct ib_qp_attr qp_attr;
411 int qp_attr_mask, ret;
413 qp_attr.qp_state = IB_QPS_INIT;
414 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
415 if (ret)
416 return ret;
418 ret = ib_modify_qp(qp, &qp_attr, qp_attr_mask);
419 if (ret)
420 return ret;
422 qp_attr.qp_state = IB_QPS_RTR;
423 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
424 if (ret)
425 return ret;
427 qp_attr.qp_state = IB_QPS_RTS;
428 qp_attr.sq_psn = 0;
429 ret = ib_modify_qp(qp, &qp_attr, IB_QP_STATE | IB_QP_SQ_PSN);
431 return ret;
434 static int cma_init_conn_qp(struct rdma_id_private *id_priv, struct ib_qp *qp)
436 struct ib_qp_attr qp_attr;
437 int qp_attr_mask, ret;
439 qp_attr.qp_state = IB_QPS_INIT;
440 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
441 if (ret)
442 return ret;
444 return ib_modify_qp(qp, &qp_attr, qp_attr_mask);
447 int rdma_create_qp(struct rdma_cm_id *id, struct ib_pd *pd,
448 struct ib_qp_init_attr *qp_init_attr)
450 struct rdma_id_private *id_priv;
451 struct ib_qp *qp;
452 int ret;
454 id_priv = container_of(id, struct rdma_id_private, id);
455 if (id->device != pd->device)
456 return -EINVAL;
458 qp = ib_create_qp(pd, qp_init_attr);
459 if (IS_ERR(qp))
460 return PTR_ERR(qp);
462 if (cma_is_ud_ps(id_priv->id.ps))
463 ret = cma_init_ud_qp(id_priv, qp);
464 else
465 ret = cma_init_conn_qp(id_priv, qp);
466 if (ret)
467 goto err;
469 id->qp = qp;
470 id_priv->qp_num = qp->qp_num;
471 id_priv->srq = (qp->srq != NULL);
472 return 0;
473 err:
474 ib_destroy_qp(qp);
475 return ret;
477 EXPORT_SYMBOL(rdma_create_qp);
479 void rdma_destroy_qp(struct rdma_cm_id *id)
481 struct rdma_id_private *id_priv;
483 id_priv = container_of(id, struct rdma_id_private, id);
484 mutex_lock(&id_priv->qp_mutex);
485 ib_destroy_qp(id_priv->id.qp);
486 id_priv->id.qp = NULL;
487 mutex_unlock(&id_priv->qp_mutex);
489 EXPORT_SYMBOL(rdma_destroy_qp);
491 static int cma_modify_qp_rtr(struct rdma_id_private *id_priv,
492 struct rdma_conn_param *conn_param)
494 struct ib_qp_attr qp_attr;
495 int qp_attr_mask, ret;
497 mutex_lock(&id_priv->qp_mutex);
498 if (!id_priv->id.qp) {
499 ret = 0;
500 goto out;
503 /* Need to update QP attributes from default values. */
504 qp_attr.qp_state = IB_QPS_INIT;
505 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
506 if (ret)
507 goto out;
509 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
510 if (ret)
511 goto out;
513 qp_attr.qp_state = IB_QPS_RTR;
514 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
515 if (ret)
516 goto out;
518 if (conn_param)
519 qp_attr.max_dest_rd_atomic = conn_param->responder_resources;
520 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
521 out:
522 mutex_unlock(&id_priv->qp_mutex);
523 return ret;
526 static int cma_modify_qp_rts(struct rdma_id_private *id_priv,
527 struct rdma_conn_param *conn_param)
529 struct ib_qp_attr qp_attr;
530 int qp_attr_mask, ret;
532 mutex_lock(&id_priv->qp_mutex);
533 if (!id_priv->id.qp) {
534 ret = 0;
535 goto out;
538 qp_attr.qp_state = IB_QPS_RTS;
539 ret = rdma_init_qp_attr(&id_priv->id, &qp_attr, &qp_attr_mask);
540 if (ret)
541 goto out;
543 if (conn_param)
544 qp_attr.max_rd_atomic = conn_param->initiator_depth;
545 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, qp_attr_mask);
546 out:
547 mutex_unlock(&id_priv->qp_mutex);
548 return ret;
551 static int cma_modify_qp_err(struct rdma_id_private *id_priv)
553 struct ib_qp_attr qp_attr;
554 int ret;
556 mutex_lock(&id_priv->qp_mutex);
557 if (!id_priv->id.qp) {
558 ret = 0;
559 goto out;
562 qp_attr.qp_state = IB_QPS_ERR;
563 ret = ib_modify_qp(id_priv->id.qp, &qp_attr, IB_QP_STATE);
564 out:
565 mutex_unlock(&id_priv->qp_mutex);
566 return ret;
569 static int cma_ib_init_qp_attr(struct rdma_id_private *id_priv,
570 struct ib_qp_attr *qp_attr, int *qp_attr_mask)
572 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
573 int ret;
575 ret = ib_find_cached_pkey(id_priv->id.device, id_priv->id.port_num,
576 ib_addr_get_pkey(dev_addr),
577 &qp_attr->pkey_index);
578 if (ret)
579 return ret;
581 qp_attr->port_num = id_priv->id.port_num;
582 *qp_attr_mask = IB_QP_STATE | IB_QP_PKEY_INDEX | IB_QP_PORT;
584 if (cma_is_ud_ps(id_priv->id.ps)) {
585 qp_attr->qkey = id_priv->qkey;
586 *qp_attr_mask |= IB_QP_QKEY;
587 } else {
588 qp_attr->qp_access_flags = 0;
589 *qp_attr_mask |= IB_QP_ACCESS_FLAGS;
591 return 0;
594 int rdma_init_qp_attr(struct rdma_cm_id *id, struct ib_qp_attr *qp_attr,
595 int *qp_attr_mask)
597 struct rdma_id_private *id_priv;
598 int ret = 0;
600 id_priv = container_of(id, struct rdma_id_private, id);
601 switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
602 case RDMA_TRANSPORT_IB:
603 if (!id_priv->cm_id.ib || cma_is_ud_ps(id_priv->id.ps))
604 ret = cma_ib_init_qp_attr(id_priv, qp_attr, qp_attr_mask);
605 else
606 ret = ib_cm_init_qp_attr(id_priv->cm_id.ib, qp_attr,
607 qp_attr_mask);
608 if (qp_attr->qp_state == IB_QPS_RTR)
609 qp_attr->rq_psn = id_priv->seq_num;
610 break;
611 case RDMA_TRANSPORT_IWARP:
612 if (!id_priv->cm_id.iw) {
613 qp_attr->qp_access_flags = 0;
614 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
615 } else
616 ret = iw_cm_init_qp_attr(id_priv->cm_id.iw, qp_attr,
617 qp_attr_mask);
618 break;
619 default:
620 ret = -ENOSYS;
621 break;
624 return ret;
626 EXPORT_SYMBOL(rdma_init_qp_attr);
628 static inline int cma_zero_addr(struct sockaddr *addr)
630 struct in6_addr *ip6;
632 if (addr->sa_family == AF_INET)
633 return ipv4_is_zeronet(
634 ((struct sockaddr_in *)addr)->sin_addr.s_addr);
635 else {
636 ip6 = &((struct sockaddr_in6 *) addr)->sin6_addr;
637 return (ip6->s6_addr32[0] | ip6->s6_addr32[1] |
638 ip6->s6_addr32[2] | ip6->s6_addr32[3]) == 0;
642 static inline int cma_loopback_addr(struct sockaddr *addr)
644 return ipv4_is_loopback(((struct sockaddr_in *) addr)->sin_addr.s_addr);
647 static inline int cma_any_addr(struct sockaddr *addr)
649 return cma_zero_addr(addr) || cma_loopback_addr(addr);
652 static inline __be16 cma_port(struct sockaddr *addr)
654 if (addr->sa_family == AF_INET)
655 return ((struct sockaddr_in *) addr)->sin_port;
656 else
657 return ((struct sockaddr_in6 *) addr)->sin6_port;
660 static inline int cma_any_port(struct sockaddr *addr)
662 return !cma_port(addr);
665 static int cma_get_net_info(void *hdr, enum rdma_port_space ps,
666 u8 *ip_ver, __u16 *port,
667 union cma_ip_addr **src, union cma_ip_addr **dst)
669 switch (ps) {
670 case RDMA_PS_SDP:
671 if (sdp_get_majv(((struct sdp_hh *) hdr)->sdp_version) !=
672 SDP_MAJ_VERSION)
673 return -EINVAL;
675 *ip_ver = sdp_get_ip_ver(hdr);
676 *port = ((struct sdp_hh *) hdr)->port;
677 *src = &((struct sdp_hh *) hdr)->src_addr;
678 *dst = &((struct sdp_hh *) hdr)->dst_addr;
679 break;
680 default:
681 if (((struct cma_hdr *) hdr)->cma_version != CMA_VERSION)
682 return -EINVAL;
684 *ip_ver = cma_get_ip_ver(hdr);
685 *port = ((struct cma_hdr *) hdr)->port;
686 *src = &((struct cma_hdr *) hdr)->src_addr;
687 *dst = &((struct cma_hdr *) hdr)->dst_addr;
688 break;
691 if (*ip_ver != 4 && *ip_ver != 6)
692 return -EINVAL;
693 return 0;
696 static void cma_save_net_info(struct rdma_addr *addr,
697 struct rdma_addr *listen_addr,
698 u8 ip_ver, __u16 port,
699 union cma_ip_addr *src, union cma_ip_addr *dst)
701 struct sockaddr_in *listen4, *ip4;
702 struct sockaddr_in6 *listen6, *ip6;
704 switch (ip_ver) {
705 case 4:
706 listen4 = (struct sockaddr_in *) &listen_addr->src_addr;
707 ip4 = (struct sockaddr_in *) &addr->src_addr;
708 ip4->sin_family = listen4->sin_family;
709 ip4->sin_addr.s_addr = dst->ip4.addr;
710 ip4->sin_port = listen4->sin_port;
712 ip4 = (struct sockaddr_in *) &addr->dst_addr;
713 ip4->sin_family = listen4->sin_family;
714 ip4->sin_addr.s_addr = src->ip4.addr;
715 ip4->sin_port = port;
716 break;
717 case 6:
718 listen6 = (struct sockaddr_in6 *) &listen_addr->src_addr;
719 ip6 = (struct sockaddr_in6 *) &addr->src_addr;
720 ip6->sin6_family = listen6->sin6_family;
721 ip6->sin6_addr = dst->ip6;
722 ip6->sin6_port = listen6->sin6_port;
724 ip6 = (struct sockaddr_in6 *) &addr->dst_addr;
725 ip6->sin6_family = listen6->sin6_family;
726 ip6->sin6_addr = src->ip6;
727 ip6->sin6_port = port;
728 break;
729 default:
730 break;
734 static inline int cma_user_data_offset(enum rdma_port_space ps)
736 switch (ps) {
737 case RDMA_PS_SDP:
738 return 0;
739 default:
740 return sizeof(struct cma_hdr);
744 static void cma_cancel_route(struct rdma_id_private *id_priv)
746 switch (rdma_node_get_transport(id_priv->id.device->node_type)) {
747 case RDMA_TRANSPORT_IB:
748 if (id_priv->query)
749 ib_sa_cancel_query(id_priv->query_id, id_priv->query);
750 break;
751 default:
752 break;
756 static void cma_cancel_listens(struct rdma_id_private *id_priv)
758 struct rdma_id_private *dev_id_priv;
761 * Remove from listen_any_list to prevent added devices from spawning
762 * additional listen requests.
764 mutex_lock(&lock);
765 list_del(&id_priv->list);
767 while (!list_empty(&id_priv->listen_list)) {
768 dev_id_priv = list_entry(id_priv->listen_list.next,
769 struct rdma_id_private, listen_list);
770 /* sync with device removal to avoid duplicate destruction */
771 list_del_init(&dev_id_priv->list);
772 list_del(&dev_id_priv->listen_list);
773 mutex_unlock(&lock);
775 rdma_destroy_id(&dev_id_priv->id);
776 mutex_lock(&lock);
778 mutex_unlock(&lock);
781 static void cma_cancel_operation(struct rdma_id_private *id_priv,
782 enum cma_state state)
784 switch (state) {
785 case CMA_ADDR_QUERY:
786 rdma_addr_cancel(&id_priv->id.route.addr.dev_addr);
787 break;
788 case CMA_ROUTE_QUERY:
789 cma_cancel_route(id_priv);
790 break;
791 case CMA_LISTEN:
792 if (cma_any_addr(&id_priv->id.route.addr.src_addr) &&
793 !id_priv->cma_dev)
794 cma_cancel_listens(id_priv);
795 break;
796 default:
797 break;
801 static void cma_release_port(struct rdma_id_private *id_priv)
803 struct rdma_bind_list *bind_list = id_priv->bind_list;
805 if (!bind_list)
806 return;
808 mutex_lock(&lock);
809 hlist_del(&id_priv->node);
810 if (hlist_empty(&bind_list->owners)) {
811 idr_remove(bind_list->ps, bind_list->port);
812 kfree(bind_list);
814 mutex_unlock(&lock);
817 static void cma_leave_mc_groups(struct rdma_id_private *id_priv)
819 struct cma_multicast *mc;
821 while (!list_empty(&id_priv->mc_list)) {
822 mc = container_of(id_priv->mc_list.next,
823 struct cma_multicast, list);
824 list_del(&mc->list);
825 ib_sa_free_multicast(mc->multicast.ib);
826 kfree(mc);
830 void rdma_destroy_id(struct rdma_cm_id *id)
832 struct rdma_id_private *id_priv;
833 enum cma_state state;
835 id_priv = container_of(id, struct rdma_id_private, id);
836 state = cma_exch(id_priv, CMA_DESTROYING);
837 cma_cancel_operation(id_priv, state);
839 mutex_lock(&lock);
840 if (id_priv->cma_dev) {
841 mutex_unlock(&lock);
842 switch (rdma_node_get_transport(id->device->node_type)) {
843 case RDMA_TRANSPORT_IB:
844 if (id_priv->cm_id.ib && !IS_ERR(id_priv->cm_id.ib))
845 ib_destroy_cm_id(id_priv->cm_id.ib);
846 break;
847 case RDMA_TRANSPORT_IWARP:
848 if (id_priv->cm_id.iw && !IS_ERR(id_priv->cm_id.iw))
849 iw_destroy_cm_id(id_priv->cm_id.iw);
850 break;
851 default:
852 break;
854 cma_leave_mc_groups(id_priv);
855 mutex_lock(&lock);
856 cma_detach_from_dev(id_priv);
858 mutex_unlock(&lock);
860 cma_release_port(id_priv);
861 cma_deref_id(id_priv);
862 wait_for_completion(&id_priv->comp);
864 if (id_priv->internal_id)
865 cma_deref_id(id_priv->id.context);
867 kfree(id_priv->id.route.path_rec);
868 kfree(id_priv);
870 EXPORT_SYMBOL(rdma_destroy_id);
872 static int cma_rep_recv(struct rdma_id_private *id_priv)
874 int ret;
876 ret = cma_modify_qp_rtr(id_priv, NULL);
877 if (ret)
878 goto reject;
880 ret = cma_modify_qp_rts(id_priv, NULL);
881 if (ret)
882 goto reject;
884 ret = ib_send_cm_rtu(id_priv->cm_id.ib, NULL, 0);
885 if (ret)
886 goto reject;
888 return 0;
889 reject:
890 cma_modify_qp_err(id_priv);
891 ib_send_cm_rej(id_priv->cm_id.ib, IB_CM_REJ_CONSUMER_DEFINED,
892 NULL, 0, NULL, 0);
893 return ret;
896 static int cma_verify_rep(struct rdma_id_private *id_priv, void *data)
898 if (id_priv->id.ps == RDMA_PS_SDP &&
899 sdp_get_majv(((struct sdp_hah *) data)->sdp_version) !=
900 SDP_MAJ_VERSION)
901 return -EINVAL;
903 return 0;
906 static void cma_set_rep_event_data(struct rdma_cm_event *event,
907 struct ib_cm_rep_event_param *rep_data,
908 void *private_data)
910 event->param.conn.private_data = private_data;
911 event->param.conn.private_data_len = IB_CM_REP_PRIVATE_DATA_SIZE;
912 event->param.conn.responder_resources = rep_data->responder_resources;
913 event->param.conn.initiator_depth = rep_data->initiator_depth;
914 event->param.conn.flow_control = rep_data->flow_control;
915 event->param.conn.rnr_retry_count = rep_data->rnr_retry_count;
916 event->param.conn.srq = rep_data->srq;
917 event->param.conn.qp_num = rep_data->remote_qpn;
920 static int cma_ib_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
922 struct rdma_id_private *id_priv = cm_id->context;
923 struct rdma_cm_event event;
924 int ret = 0;
926 if (cma_disable_remove(id_priv, CMA_CONNECT))
927 return 0;
929 memset(&event, 0, sizeof event);
930 switch (ib_event->event) {
931 case IB_CM_REQ_ERROR:
932 case IB_CM_REP_ERROR:
933 event.event = RDMA_CM_EVENT_UNREACHABLE;
934 event.status = -ETIMEDOUT;
935 break;
936 case IB_CM_REP_RECEIVED:
937 event.status = cma_verify_rep(id_priv, ib_event->private_data);
938 if (event.status)
939 event.event = RDMA_CM_EVENT_CONNECT_ERROR;
940 else if (id_priv->id.qp && id_priv->id.ps != RDMA_PS_SDP) {
941 event.status = cma_rep_recv(id_priv);
942 event.event = event.status ? RDMA_CM_EVENT_CONNECT_ERROR :
943 RDMA_CM_EVENT_ESTABLISHED;
944 } else
945 event.event = RDMA_CM_EVENT_CONNECT_RESPONSE;
946 cma_set_rep_event_data(&event, &ib_event->param.rep_rcvd,
947 ib_event->private_data);
948 break;
949 case IB_CM_RTU_RECEIVED:
950 case IB_CM_USER_ESTABLISHED:
951 event.event = RDMA_CM_EVENT_ESTABLISHED;
952 break;
953 case IB_CM_DREQ_ERROR:
954 event.status = -ETIMEDOUT; /* fall through */
955 case IB_CM_DREQ_RECEIVED:
956 case IB_CM_DREP_RECEIVED:
957 if (!cma_comp_exch(id_priv, CMA_CONNECT, CMA_DISCONNECT))
958 goto out;
959 event.event = RDMA_CM_EVENT_DISCONNECTED;
960 break;
961 case IB_CM_TIMEWAIT_EXIT:
962 case IB_CM_MRA_RECEIVED:
963 /* ignore event */
964 goto out;
965 case IB_CM_REJ_RECEIVED:
966 cma_modify_qp_err(id_priv);
967 event.status = ib_event->param.rej_rcvd.reason;
968 event.event = RDMA_CM_EVENT_REJECTED;
969 event.param.conn.private_data = ib_event->private_data;
970 event.param.conn.private_data_len = IB_CM_REJ_PRIVATE_DATA_SIZE;
971 break;
972 default:
973 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d",
974 ib_event->event);
975 goto out;
978 ret = id_priv->id.event_handler(&id_priv->id, &event);
979 if (ret) {
980 /* Destroy the CM ID by returning a non-zero value. */
981 id_priv->cm_id.ib = NULL;
982 cma_exch(id_priv, CMA_DESTROYING);
983 cma_enable_remove(id_priv);
984 rdma_destroy_id(&id_priv->id);
985 return ret;
987 out:
988 cma_enable_remove(id_priv);
989 return ret;
992 static struct rdma_id_private *cma_new_conn_id(struct rdma_cm_id *listen_id,
993 struct ib_cm_event *ib_event)
995 struct rdma_id_private *id_priv;
996 struct rdma_cm_id *id;
997 struct rdma_route *rt;
998 union cma_ip_addr *src, *dst;
999 __u16 port;
1000 u8 ip_ver;
1002 if (cma_get_net_info(ib_event->private_data, listen_id->ps,
1003 &ip_ver, &port, &src, &dst))
1004 goto err;
1006 id = rdma_create_id(listen_id->event_handler, listen_id->context,
1007 listen_id->ps);
1008 if (IS_ERR(id))
1009 goto err;
1011 cma_save_net_info(&id->route.addr, &listen_id->route.addr,
1012 ip_ver, port, src, dst);
1014 rt = &id->route;
1015 rt->num_paths = ib_event->param.req_rcvd.alternate_path ? 2 : 1;
1016 rt->path_rec = kmalloc(sizeof *rt->path_rec * rt->num_paths,
1017 GFP_KERNEL);
1018 if (!rt->path_rec)
1019 goto destroy_id;
1021 rt->path_rec[0] = *ib_event->param.req_rcvd.primary_path;
1022 if (rt->num_paths == 2)
1023 rt->path_rec[1] = *ib_event->param.req_rcvd.alternate_path;
1025 ib_addr_set_sgid(&rt->addr.dev_addr, &rt->path_rec[0].sgid);
1026 ib_addr_set_dgid(&rt->addr.dev_addr, &rt->path_rec[0].dgid);
1027 ib_addr_set_pkey(&rt->addr.dev_addr, be16_to_cpu(rt->path_rec[0].pkey));
1028 rt->addr.dev_addr.dev_type = RDMA_NODE_IB_CA;
1030 id_priv = container_of(id, struct rdma_id_private, id);
1031 id_priv->state = CMA_CONNECT;
1032 return id_priv;
1034 destroy_id:
1035 rdma_destroy_id(id);
1036 err:
1037 return NULL;
1040 static struct rdma_id_private *cma_new_udp_id(struct rdma_cm_id *listen_id,
1041 struct ib_cm_event *ib_event)
1043 struct rdma_id_private *id_priv;
1044 struct rdma_cm_id *id;
1045 union cma_ip_addr *src, *dst;
1046 __u16 port;
1047 u8 ip_ver;
1048 int ret;
1050 id = rdma_create_id(listen_id->event_handler, listen_id->context,
1051 listen_id->ps);
1052 if (IS_ERR(id))
1053 return NULL;
1056 if (cma_get_net_info(ib_event->private_data, listen_id->ps,
1057 &ip_ver, &port, &src, &dst))
1058 goto err;
1060 cma_save_net_info(&id->route.addr, &listen_id->route.addr,
1061 ip_ver, port, src, dst);
1063 ret = rdma_translate_ip(&id->route.addr.src_addr,
1064 &id->route.addr.dev_addr);
1065 if (ret)
1066 goto err;
1068 id_priv = container_of(id, struct rdma_id_private, id);
1069 id_priv->state = CMA_CONNECT;
1070 return id_priv;
1071 err:
1072 rdma_destroy_id(id);
1073 return NULL;
1076 static void cma_set_req_event_data(struct rdma_cm_event *event,
1077 struct ib_cm_req_event_param *req_data,
1078 void *private_data, int offset)
1080 event->param.conn.private_data = private_data + offset;
1081 event->param.conn.private_data_len = IB_CM_REQ_PRIVATE_DATA_SIZE - offset;
1082 event->param.conn.responder_resources = req_data->responder_resources;
1083 event->param.conn.initiator_depth = req_data->initiator_depth;
1084 event->param.conn.flow_control = req_data->flow_control;
1085 event->param.conn.retry_count = req_data->retry_count;
1086 event->param.conn.rnr_retry_count = req_data->rnr_retry_count;
1087 event->param.conn.srq = req_data->srq;
1088 event->param.conn.qp_num = req_data->remote_qpn;
1091 static int cma_req_handler(struct ib_cm_id *cm_id, struct ib_cm_event *ib_event)
1093 struct rdma_id_private *listen_id, *conn_id;
1094 struct rdma_cm_event event;
1095 int offset, ret;
1097 listen_id = cm_id->context;
1098 if (cma_disable_remove(listen_id, CMA_LISTEN))
1099 return -ECONNABORTED;
1101 memset(&event, 0, sizeof event);
1102 offset = cma_user_data_offset(listen_id->id.ps);
1103 event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1104 if (cma_is_ud_ps(listen_id->id.ps)) {
1105 conn_id = cma_new_udp_id(&listen_id->id, ib_event);
1106 event.param.ud.private_data = ib_event->private_data + offset;
1107 event.param.ud.private_data_len =
1108 IB_CM_SIDR_REQ_PRIVATE_DATA_SIZE - offset;
1109 } else {
1110 ib_send_cm_mra(cm_id, CMA_CM_MRA_SETTING, NULL, 0);
1111 conn_id = cma_new_conn_id(&listen_id->id, ib_event);
1112 cma_set_req_event_data(&event, &ib_event->param.req_rcvd,
1113 ib_event->private_data, offset);
1115 if (!conn_id) {
1116 ret = -ENOMEM;
1117 goto out;
1120 atomic_inc(&conn_id->dev_remove);
1121 mutex_lock(&lock);
1122 ret = cma_acquire_dev(conn_id);
1123 mutex_unlock(&lock);
1124 if (ret)
1125 goto release_conn_id;
1127 conn_id->cm_id.ib = cm_id;
1128 cm_id->context = conn_id;
1129 cm_id->cm_handler = cma_ib_handler;
1131 ret = conn_id->id.event_handler(&conn_id->id, &event);
1132 if (!ret) {
1133 cma_enable_remove(conn_id);
1134 goto out;
1137 /* Destroy the CM ID by returning a non-zero value. */
1138 conn_id->cm_id.ib = NULL;
1140 release_conn_id:
1141 cma_exch(conn_id, CMA_DESTROYING);
1142 cma_enable_remove(conn_id);
1143 rdma_destroy_id(&conn_id->id);
1145 out:
1146 cma_enable_remove(listen_id);
1147 return ret;
1150 static __be64 cma_get_service_id(enum rdma_port_space ps, struct sockaddr *addr)
1152 return cpu_to_be64(((u64)ps << 16) + be16_to_cpu(cma_port(addr)));
1155 static void cma_set_compare_data(enum rdma_port_space ps, struct sockaddr *addr,
1156 struct ib_cm_compare_data *compare)
1158 struct cma_hdr *cma_data, *cma_mask;
1159 struct sdp_hh *sdp_data, *sdp_mask;
1160 __u32 ip4_addr;
1161 struct in6_addr ip6_addr;
1163 memset(compare, 0, sizeof *compare);
1164 cma_data = (void *) compare->data;
1165 cma_mask = (void *) compare->mask;
1166 sdp_data = (void *) compare->data;
1167 sdp_mask = (void *) compare->mask;
1169 switch (addr->sa_family) {
1170 case AF_INET:
1171 ip4_addr = ((struct sockaddr_in *) addr)->sin_addr.s_addr;
1172 if (ps == RDMA_PS_SDP) {
1173 sdp_set_ip_ver(sdp_data, 4);
1174 sdp_set_ip_ver(sdp_mask, 0xF);
1175 sdp_data->dst_addr.ip4.addr = ip4_addr;
1176 sdp_mask->dst_addr.ip4.addr = ~0;
1177 } else {
1178 cma_set_ip_ver(cma_data, 4);
1179 cma_set_ip_ver(cma_mask, 0xF);
1180 cma_data->dst_addr.ip4.addr = ip4_addr;
1181 cma_mask->dst_addr.ip4.addr = ~0;
1183 break;
1184 case AF_INET6:
1185 ip6_addr = ((struct sockaddr_in6 *) addr)->sin6_addr;
1186 if (ps == RDMA_PS_SDP) {
1187 sdp_set_ip_ver(sdp_data, 6);
1188 sdp_set_ip_ver(sdp_mask, 0xF);
1189 sdp_data->dst_addr.ip6 = ip6_addr;
1190 memset(&sdp_mask->dst_addr.ip6, 0xFF,
1191 sizeof sdp_mask->dst_addr.ip6);
1192 } else {
1193 cma_set_ip_ver(cma_data, 6);
1194 cma_set_ip_ver(cma_mask, 0xF);
1195 cma_data->dst_addr.ip6 = ip6_addr;
1196 memset(&cma_mask->dst_addr.ip6, 0xFF,
1197 sizeof cma_mask->dst_addr.ip6);
1199 break;
1200 default:
1201 break;
1205 static int cma_iw_handler(struct iw_cm_id *iw_id, struct iw_cm_event *iw_event)
1207 struct rdma_id_private *id_priv = iw_id->context;
1208 struct rdma_cm_event event;
1209 struct sockaddr_in *sin;
1210 int ret = 0;
1212 if (cma_disable_remove(id_priv, CMA_CONNECT))
1213 return 0;
1215 memset(&event, 0, sizeof event);
1216 switch (iw_event->event) {
1217 case IW_CM_EVENT_CLOSE:
1218 event.event = RDMA_CM_EVENT_DISCONNECTED;
1219 break;
1220 case IW_CM_EVENT_CONNECT_REPLY:
1221 sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1222 *sin = iw_event->local_addr;
1223 sin = (struct sockaddr_in *) &id_priv->id.route.addr.dst_addr;
1224 *sin = iw_event->remote_addr;
1225 switch (iw_event->status) {
1226 case 0:
1227 event.event = RDMA_CM_EVENT_ESTABLISHED;
1228 break;
1229 case -ECONNRESET:
1230 case -ECONNREFUSED:
1231 event.event = RDMA_CM_EVENT_REJECTED;
1232 break;
1233 case -ETIMEDOUT:
1234 event.event = RDMA_CM_EVENT_UNREACHABLE;
1235 break;
1236 default:
1237 event.event = RDMA_CM_EVENT_CONNECT_ERROR;
1238 break;
1240 break;
1241 case IW_CM_EVENT_ESTABLISHED:
1242 event.event = RDMA_CM_EVENT_ESTABLISHED;
1243 break;
1244 default:
1245 BUG_ON(1);
1248 event.status = iw_event->status;
1249 event.param.conn.private_data = iw_event->private_data;
1250 event.param.conn.private_data_len = iw_event->private_data_len;
1251 ret = id_priv->id.event_handler(&id_priv->id, &event);
1252 if (ret) {
1253 /* Destroy the CM ID by returning a non-zero value. */
1254 id_priv->cm_id.iw = NULL;
1255 cma_exch(id_priv, CMA_DESTROYING);
1256 cma_enable_remove(id_priv);
1257 rdma_destroy_id(&id_priv->id);
1258 return ret;
1261 cma_enable_remove(id_priv);
1262 return ret;
1265 static int iw_conn_req_handler(struct iw_cm_id *cm_id,
1266 struct iw_cm_event *iw_event)
1268 struct rdma_cm_id *new_cm_id;
1269 struct rdma_id_private *listen_id, *conn_id;
1270 struct sockaddr_in *sin;
1271 struct net_device *dev = NULL;
1272 struct rdma_cm_event event;
1273 int ret;
1274 struct ib_device_attr attr;
1276 listen_id = cm_id->context;
1277 if (cma_disable_remove(listen_id, CMA_LISTEN))
1278 return -ECONNABORTED;
1280 /* Create a new RDMA id for the new IW CM ID */
1281 new_cm_id = rdma_create_id(listen_id->id.event_handler,
1282 listen_id->id.context,
1283 RDMA_PS_TCP);
1284 if (!new_cm_id) {
1285 ret = -ENOMEM;
1286 goto out;
1288 conn_id = container_of(new_cm_id, struct rdma_id_private, id);
1289 atomic_inc(&conn_id->dev_remove);
1290 conn_id->state = CMA_CONNECT;
1292 dev = ip_dev_find(&init_net, iw_event->local_addr.sin_addr.s_addr);
1293 if (!dev) {
1294 ret = -EADDRNOTAVAIL;
1295 cma_enable_remove(conn_id);
1296 rdma_destroy_id(new_cm_id);
1297 goto out;
1299 ret = rdma_copy_addr(&conn_id->id.route.addr.dev_addr, dev, NULL);
1300 if (ret) {
1301 cma_enable_remove(conn_id);
1302 rdma_destroy_id(new_cm_id);
1303 goto out;
1306 mutex_lock(&lock);
1307 ret = cma_acquire_dev(conn_id);
1308 mutex_unlock(&lock);
1309 if (ret) {
1310 cma_enable_remove(conn_id);
1311 rdma_destroy_id(new_cm_id);
1312 goto out;
1315 conn_id->cm_id.iw = cm_id;
1316 cm_id->context = conn_id;
1317 cm_id->cm_handler = cma_iw_handler;
1319 sin = (struct sockaddr_in *) &new_cm_id->route.addr.src_addr;
1320 *sin = iw_event->local_addr;
1321 sin = (struct sockaddr_in *) &new_cm_id->route.addr.dst_addr;
1322 *sin = iw_event->remote_addr;
1324 ret = ib_query_device(conn_id->id.device, &attr);
1325 if (ret) {
1326 cma_enable_remove(conn_id);
1327 rdma_destroy_id(new_cm_id);
1328 goto out;
1331 memset(&event, 0, sizeof event);
1332 event.event = RDMA_CM_EVENT_CONNECT_REQUEST;
1333 event.param.conn.private_data = iw_event->private_data;
1334 event.param.conn.private_data_len = iw_event->private_data_len;
1335 event.param.conn.initiator_depth = attr.max_qp_init_rd_atom;
1336 event.param.conn.responder_resources = attr.max_qp_rd_atom;
1337 ret = conn_id->id.event_handler(&conn_id->id, &event);
1338 if (ret) {
1339 /* User wants to destroy the CM ID */
1340 conn_id->cm_id.iw = NULL;
1341 cma_exch(conn_id, CMA_DESTROYING);
1342 cma_enable_remove(conn_id);
1343 rdma_destroy_id(&conn_id->id);
1346 out:
1347 if (dev)
1348 dev_put(dev);
1349 cma_enable_remove(listen_id);
1350 return ret;
1353 static int cma_ib_listen(struct rdma_id_private *id_priv)
1355 struct ib_cm_compare_data compare_data;
1356 struct sockaddr *addr;
1357 __be64 svc_id;
1358 int ret;
1360 id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, cma_req_handler,
1361 id_priv);
1362 if (IS_ERR(id_priv->cm_id.ib))
1363 return PTR_ERR(id_priv->cm_id.ib);
1365 addr = &id_priv->id.route.addr.src_addr;
1366 svc_id = cma_get_service_id(id_priv->id.ps, addr);
1367 if (cma_any_addr(addr))
1368 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, NULL);
1369 else {
1370 cma_set_compare_data(id_priv->id.ps, addr, &compare_data);
1371 ret = ib_cm_listen(id_priv->cm_id.ib, svc_id, 0, &compare_data);
1374 if (ret) {
1375 ib_destroy_cm_id(id_priv->cm_id.ib);
1376 id_priv->cm_id.ib = NULL;
1379 return ret;
1382 static int cma_iw_listen(struct rdma_id_private *id_priv, int backlog)
1384 int ret;
1385 struct sockaddr_in *sin;
1387 id_priv->cm_id.iw = iw_create_cm_id(id_priv->id.device,
1388 iw_conn_req_handler,
1389 id_priv);
1390 if (IS_ERR(id_priv->cm_id.iw))
1391 return PTR_ERR(id_priv->cm_id.iw);
1393 sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1394 id_priv->cm_id.iw->local_addr = *sin;
1396 ret = iw_cm_listen(id_priv->cm_id.iw, backlog);
1398 if (ret) {
1399 iw_destroy_cm_id(id_priv->cm_id.iw);
1400 id_priv->cm_id.iw = NULL;
1403 return ret;
1406 static int cma_listen_handler(struct rdma_cm_id *id,
1407 struct rdma_cm_event *event)
1409 struct rdma_id_private *id_priv = id->context;
1411 id->context = id_priv->id.context;
1412 id->event_handler = id_priv->id.event_handler;
1413 return id_priv->id.event_handler(id, event);
1416 static void cma_listen_on_dev(struct rdma_id_private *id_priv,
1417 struct cma_device *cma_dev)
1419 struct rdma_id_private *dev_id_priv;
1420 struct rdma_cm_id *id;
1421 int ret;
1423 id = rdma_create_id(cma_listen_handler, id_priv, id_priv->id.ps);
1424 if (IS_ERR(id))
1425 return;
1427 dev_id_priv = container_of(id, struct rdma_id_private, id);
1429 dev_id_priv->state = CMA_ADDR_BOUND;
1430 memcpy(&id->route.addr.src_addr, &id_priv->id.route.addr.src_addr,
1431 ip_addr_size(&id_priv->id.route.addr.src_addr));
1433 cma_attach_to_dev(dev_id_priv, cma_dev);
1434 list_add_tail(&dev_id_priv->listen_list, &id_priv->listen_list);
1435 atomic_inc(&id_priv->refcount);
1436 dev_id_priv->internal_id = 1;
1438 ret = rdma_listen(id, id_priv->backlog);
1439 if (ret)
1440 printk(KERN_WARNING "RDMA CMA: cma_listen_on_dev, error %d, "
1441 "listening on device %s", ret, cma_dev->device->name);
1444 static void cma_listen_on_all(struct rdma_id_private *id_priv)
1446 struct cma_device *cma_dev;
1448 mutex_lock(&lock);
1449 list_add_tail(&id_priv->list, &listen_any_list);
1450 list_for_each_entry(cma_dev, &dev_list, list)
1451 cma_listen_on_dev(id_priv, cma_dev);
1452 mutex_unlock(&lock);
1455 static int cma_bind_any(struct rdma_cm_id *id, sa_family_t af)
1457 struct sockaddr_in addr_in;
1459 memset(&addr_in, 0, sizeof addr_in);
1460 addr_in.sin_family = af;
1461 return rdma_bind_addr(id, (struct sockaddr *) &addr_in);
1464 int rdma_listen(struct rdma_cm_id *id, int backlog)
1466 struct rdma_id_private *id_priv;
1467 int ret;
1469 id_priv = container_of(id, struct rdma_id_private, id);
1470 if (id_priv->state == CMA_IDLE) {
1471 ret = cma_bind_any(id, AF_INET);
1472 if (ret)
1473 return ret;
1476 if (!cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_LISTEN))
1477 return -EINVAL;
1479 id_priv->backlog = backlog;
1480 if (id->device) {
1481 switch (rdma_node_get_transport(id->device->node_type)) {
1482 case RDMA_TRANSPORT_IB:
1483 ret = cma_ib_listen(id_priv);
1484 if (ret)
1485 goto err;
1486 break;
1487 case RDMA_TRANSPORT_IWARP:
1488 ret = cma_iw_listen(id_priv, backlog);
1489 if (ret)
1490 goto err;
1491 break;
1492 default:
1493 ret = -ENOSYS;
1494 goto err;
1496 } else
1497 cma_listen_on_all(id_priv);
1499 return 0;
1500 err:
1501 id_priv->backlog = 0;
1502 cma_comp_exch(id_priv, CMA_LISTEN, CMA_ADDR_BOUND);
1503 return ret;
1505 EXPORT_SYMBOL(rdma_listen);
1507 void rdma_set_service_type(struct rdma_cm_id *id, int tos)
1509 struct rdma_id_private *id_priv;
1511 id_priv = container_of(id, struct rdma_id_private, id);
1512 id_priv->tos = (u8) tos;
1514 EXPORT_SYMBOL(rdma_set_service_type);
1516 static void cma_query_handler(int status, struct ib_sa_path_rec *path_rec,
1517 void *context)
1519 struct cma_work *work = context;
1520 struct rdma_route *route;
1522 route = &work->id->id.route;
1524 if (!status) {
1525 route->num_paths = 1;
1526 *route->path_rec = *path_rec;
1527 } else {
1528 work->old_state = CMA_ROUTE_QUERY;
1529 work->new_state = CMA_ADDR_RESOLVED;
1530 work->event.event = RDMA_CM_EVENT_ROUTE_ERROR;
1531 work->event.status = status;
1534 queue_work(cma_wq, &work->work);
1537 static int cma_query_ib_route(struct rdma_id_private *id_priv, int timeout_ms,
1538 struct cma_work *work)
1540 struct rdma_addr *addr = &id_priv->id.route.addr;
1541 struct ib_sa_path_rec path_rec;
1542 ib_sa_comp_mask comp_mask;
1543 struct sockaddr_in6 *sin6;
1545 memset(&path_rec, 0, sizeof path_rec);
1546 ib_addr_get_sgid(&addr->dev_addr, &path_rec.sgid);
1547 ib_addr_get_dgid(&addr->dev_addr, &path_rec.dgid);
1548 path_rec.pkey = cpu_to_be16(ib_addr_get_pkey(&addr->dev_addr));
1549 path_rec.numb_path = 1;
1550 path_rec.reversible = 1;
1551 path_rec.service_id = cma_get_service_id(id_priv->id.ps, &addr->dst_addr);
1553 comp_mask = IB_SA_PATH_REC_DGID | IB_SA_PATH_REC_SGID |
1554 IB_SA_PATH_REC_PKEY | IB_SA_PATH_REC_NUMB_PATH |
1555 IB_SA_PATH_REC_REVERSIBLE | IB_SA_PATH_REC_SERVICE_ID;
1557 if (addr->src_addr.sa_family == AF_INET) {
1558 path_rec.qos_class = cpu_to_be16((u16) id_priv->tos);
1559 comp_mask |= IB_SA_PATH_REC_QOS_CLASS;
1560 } else {
1561 sin6 = (struct sockaddr_in6 *) &addr->src_addr;
1562 path_rec.traffic_class = (u8) (be32_to_cpu(sin6->sin6_flowinfo) >> 20);
1563 comp_mask |= IB_SA_PATH_REC_TRAFFIC_CLASS;
1566 id_priv->query_id = ib_sa_path_rec_get(&sa_client, id_priv->id.device,
1567 id_priv->id.port_num, &path_rec,
1568 comp_mask, timeout_ms,
1569 GFP_KERNEL, cma_query_handler,
1570 work, &id_priv->query);
1572 return (id_priv->query_id < 0) ? id_priv->query_id : 0;
1575 static void cma_work_handler(struct work_struct *_work)
1577 struct cma_work *work = container_of(_work, struct cma_work, work);
1578 struct rdma_id_private *id_priv = work->id;
1579 int destroy = 0;
1581 atomic_inc(&id_priv->dev_remove);
1582 if (!cma_comp_exch(id_priv, work->old_state, work->new_state))
1583 goto out;
1585 if (id_priv->id.event_handler(&id_priv->id, &work->event)) {
1586 cma_exch(id_priv, CMA_DESTROYING);
1587 destroy = 1;
1589 out:
1590 cma_enable_remove(id_priv);
1591 cma_deref_id(id_priv);
1592 if (destroy)
1593 rdma_destroy_id(&id_priv->id);
1594 kfree(work);
1597 static int cma_resolve_ib_route(struct rdma_id_private *id_priv, int timeout_ms)
1599 struct rdma_route *route = &id_priv->id.route;
1600 struct cma_work *work;
1601 int ret;
1603 work = kzalloc(sizeof *work, GFP_KERNEL);
1604 if (!work)
1605 return -ENOMEM;
1607 work->id = id_priv;
1608 INIT_WORK(&work->work, cma_work_handler);
1609 work->old_state = CMA_ROUTE_QUERY;
1610 work->new_state = CMA_ROUTE_RESOLVED;
1611 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1613 route->path_rec = kmalloc(sizeof *route->path_rec, GFP_KERNEL);
1614 if (!route->path_rec) {
1615 ret = -ENOMEM;
1616 goto err1;
1619 ret = cma_query_ib_route(id_priv, timeout_ms, work);
1620 if (ret)
1621 goto err2;
1623 return 0;
1624 err2:
1625 kfree(route->path_rec);
1626 route->path_rec = NULL;
1627 err1:
1628 kfree(work);
1629 return ret;
1632 int rdma_set_ib_paths(struct rdma_cm_id *id,
1633 struct ib_sa_path_rec *path_rec, int num_paths)
1635 struct rdma_id_private *id_priv;
1636 int ret;
1638 id_priv = container_of(id, struct rdma_id_private, id);
1639 if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ROUTE_RESOLVED))
1640 return -EINVAL;
1642 id->route.path_rec = kmalloc(sizeof *path_rec * num_paths, GFP_KERNEL);
1643 if (!id->route.path_rec) {
1644 ret = -ENOMEM;
1645 goto err;
1648 memcpy(id->route.path_rec, path_rec, sizeof *path_rec * num_paths);
1649 return 0;
1650 err:
1651 cma_comp_exch(id_priv, CMA_ROUTE_RESOLVED, CMA_ADDR_RESOLVED);
1652 return ret;
1654 EXPORT_SYMBOL(rdma_set_ib_paths);
1656 static int cma_resolve_iw_route(struct rdma_id_private *id_priv, int timeout_ms)
1658 struct cma_work *work;
1660 work = kzalloc(sizeof *work, GFP_KERNEL);
1661 if (!work)
1662 return -ENOMEM;
1664 work->id = id_priv;
1665 INIT_WORK(&work->work, cma_work_handler);
1666 work->old_state = CMA_ROUTE_QUERY;
1667 work->new_state = CMA_ROUTE_RESOLVED;
1668 work->event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
1669 queue_work(cma_wq, &work->work);
1670 return 0;
1673 int rdma_resolve_route(struct rdma_cm_id *id, int timeout_ms)
1675 struct rdma_id_private *id_priv;
1676 int ret;
1678 id_priv = container_of(id, struct rdma_id_private, id);
1679 if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ROUTE_QUERY))
1680 return -EINVAL;
1682 atomic_inc(&id_priv->refcount);
1683 switch (rdma_node_get_transport(id->device->node_type)) {
1684 case RDMA_TRANSPORT_IB:
1685 ret = cma_resolve_ib_route(id_priv, timeout_ms);
1686 break;
1687 case RDMA_TRANSPORT_IWARP:
1688 ret = cma_resolve_iw_route(id_priv, timeout_ms);
1689 break;
1690 default:
1691 ret = -ENOSYS;
1692 break;
1694 if (ret)
1695 goto err;
1697 return 0;
1698 err:
1699 cma_comp_exch(id_priv, CMA_ROUTE_QUERY, CMA_ADDR_RESOLVED);
1700 cma_deref_id(id_priv);
1701 return ret;
1703 EXPORT_SYMBOL(rdma_resolve_route);
1705 static int cma_bind_loopback(struct rdma_id_private *id_priv)
1707 struct cma_device *cma_dev;
1708 struct ib_port_attr port_attr;
1709 union ib_gid gid;
1710 u16 pkey;
1711 int ret;
1712 u8 p;
1714 mutex_lock(&lock);
1715 if (list_empty(&dev_list)) {
1716 ret = -ENODEV;
1717 goto out;
1719 list_for_each_entry(cma_dev, &dev_list, list)
1720 for (p = 1; p <= cma_dev->device->phys_port_cnt; ++p)
1721 if (!ib_query_port(cma_dev->device, p, &port_attr) &&
1722 port_attr.state == IB_PORT_ACTIVE)
1723 goto port_found;
1725 p = 1;
1726 cma_dev = list_entry(dev_list.next, struct cma_device, list);
1728 port_found:
1729 ret = ib_get_cached_gid(cma_dev->device, p, 0, &gid);
1730 if (ret)
1731 goto out;
1733 ret = ib_get_cached_pkey(cma_dev->device, p, 0, &pkey);
1734 if (ret)
1735 goto out;
1737 ib_addr_set_sgid(&id_priv->id.route.addr.dev_addr, &gid);
1738 ib_addr_set_pkey(&id_priv->id.route.addr.dev_addr, pkey);
1739 id_priv->id.port_num = p;
1740 cma_attach_to_dev(id_priv, cma_dev);
1741 out:
1742 mutex_unlock(&lock);
1743 return ret;
1746 static void addr_handler(int status, struct sockaddr *src_addr,
1747 struct rdma_dev_addr *dev_addr, void *context)
1749 struct rdma_id_private *id_priv = context;
1750 struct rdma_cm_event event;
1752 memset(&event, 0, sizeof event);
1753 atomic_inc(&id_priv->dev_remove);
1756 * Grab mutex to block rdma_destroy_id() from removing the device while
1757 * we're trying to acquire it.
1759 mutex_lock(&lock);
1760 if (!cma_comp_exch(id_priv, CMA_ADDR_QUERY, CMA_ADDR_RESOLVED)) {
1761 mutex_unlock(&lock);
1762 goto out;
1765 if (!status && !id_priv->cma_dev)
1766 status = cma_acquire_dev(id_priv);
1767 mutex_unlock(&lock);
1769 if (status) {
1770 if (!cma_comp_exch(id_priv, CMA_ADDR_RESOLVED, CMA_ADDR_BOUND))
1771 goto out;
1772 event.event = RDMA_CM_EVENT_ADDR_ERROR;
1773 event.status = status;
1774 } else {
1775 memcpy(&id_priv->id.route.addr.src_addr, src_addr,
1776 ip_addr_size(src_addr));
1777 event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
1780 if (id_priv->id.event_handler(&id_priv->id, &event)) {
1781 cma_exch(id_priv, CMA_DESTROYING);
1782 cma_enable_remove(id_priv);
1783 cma_deref_id(id_priv);
1784 rdma_destroy_id(&id_priv->id);
1785 return;
1787 out:
1788 cma_enable_remove(id_priv);
1789 cma_deref_id(id_priv);
1792 static int cma_resolve_loopback(struct rdma_id_private *id_priv)
1794 struct cma_work *work;
1795 struct sockaddr_in *src_in, *dst_in;
1796 union ib_gid gid;
1797 int ret;
1799 work = kzalloc(sizeof *work, GFP_KERNEL);
1800 if (!work)
1801 return -ENOMEM;
1803 if (!id_priv->cma_dev) {
1804 ret = cma_bind_loopback(id_priv);
1805 if (ret)
1806 goto err;
1809 ib_addr_get_sgid(&id_priv->id.route.addr.dev_addr, &gid);
1810 ib_addr_set_dgid(&id_priv->id.route.addr.dev_addr, &gid);
1812 if (cma_zero_addr(&id_priv->id.route.addr.src_addr)) {
1813 src_in = (struct sockaddr_in *)&id_priv->id.route.addr.src_addr;
1814 dst_in = (struct sockaddr_in *)&id_priv->id.route.addr.dst_addr;
1815 src_in->sin_family = dst_in->sin_family;
1816 src_in->sin_addr.s_addr = dst_in->sin_addr.s_addr;
1819 work->id = id_priv;
1820 INIT_WORK(&work->work, cma_work_handler);
1821 work->old_state = CMA_ADDR_QUERY;
1822 work->new_state = CMA_ADDR_RESOLVED;
1823 work->event.event = RDMA_CM_EVENT_ADDR_RESOLVED;
1824 queue_work(cma_wq, &work->work);
1825 return 0;
1826 err:
1827 kfree(work);
1828 return ret;
1831 static int cma_bind_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
1832 struct sockaddr *dst_addr)
1834 if (src_addr && src_addr->sa_family)
1835 return rdma_bind_addr(id, src_addr);
1836 else
1837 return cma_bind_any(id, dst_addr->sa_family);
1840 int rdma_resolve_addr(struct rdma_cm_id *id, struct sockaddr *src_addr,
1841 struct sockaddr *dst_addr, int timeout_ms)
1843 struct rdma_id_private *id_priv;
1844 int ret;
1846 id_priv = container_of(id, struct rdma_id_private, id);
1847 if (id_priv->state == CMA_IDLE) {
1848 ret = cma_bind_addr(id, src_addr, dst_addr);
1849 if (ret)
1850 return ret;
1853 if (!cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_ADDR_QUERY))
1854 return -EINVAL;
1856 atomic_inc(&id_priv->refcount);
1857 memcpy(&id->route.addr.dst_addr, dst_addr, ip_addr_size(dst_addr));
1858 if (cma_any_addr(dst_addr))
1859 ret = cma_resolve_loopback(id_priv);
1860 else
1861 ret = rdma_resolve_ip(&addr_client, &id->route.addr.src_addr,
1862 dst_addr, &id->route.addr.dev_addr,
1863 timeout_ms, addr_handler, id_priv);
1864 if (ret)
1865 goto err;
1867 return 0;
1868 err:
1869 cma_comp_exch(id_priv, CMA_ADDR_QUERY, CMA_ADDR_BOUND);
1870 cma_deref_id(id_priv);
1871 return ret;
1873 EXPORT_SYMBOL(rdma_resolve_addr);
1875 static void cma_bind_port(struct rdma_bind_list *bind_list,
1876 struct rdma_id_private *id_priv)
1878 struct sockaddr_in *sin;
1880 sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1881 sin->sin_port = htons(bind_list->port);
1882 id_priv->bind_list = bind_list;
1883 hlist_add_head(&id_priv->node, &bind_list->owners);
1886 static int cma_alloc_port(struct idr *ps, struct rdma_id_private *id_priv,
1887 unsigned short snum)
1889 struct rdma_bind_list *bind_list;
1890 int port, ret;
1892 bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
1893 if (!bind_list)
1894 return -ENOMEM;
1896 do {
1897 ret = idr_get_new_above(ps, bind_list, snum, &port);
1898 } while ((ret == -EAGAIN) && idr_pre_get(ps, GFP_KERNEL));
1900 if (ret)
1901 goto err1;
1903 if (port != snum) {
1904 ret = -EADDRNOTAVAIL;
1905 goto err2;
1908 bind_list->ps = ps;
1909 bind_list->port = (unsigned short) port;
1910 cma_bind_port(bind_list, id_priv);
1911 return 0;
1912 err2:
1913 idr_remove(ps, port);
1914 err1:
1915 kfree(bind_list);
1916 return ret;
1919 static int cma_alloc_any_port(struct idr *ps, struct rdma_id_private *id_priv)
1921 struct rdma_bind_list *bind_list;
1922 int port, ret, low, high;
1924 bind_list = kzalloc(sizeof *bind_list, GFP_KERNEL);
1925 if (!bind_list)
1926 return -ENOMEM;
1928 retry:
1929 /* FIXME: add proper port randomization per like inet_csk_get_port */
1930 do {
1931 ret = idr_get_new_above(ps, bind_list, next_port, &port);
1932 } while ((ret == -EAGAIN) && idr_pre_get(ps, GFP_KERNEL));
1934 if (ret)
1935 goto err1;
1937 inet_get_local_port_range(&low, &high);
1938 if (port > high) {
1939 if (next_port != low) {
1940 idr_remove(ps, port);
1941 next_port = low;
1942 goto retry;
1944 ret = -EADDRNOTAVAIL;
1945 goto err2;
1948 if (port == high)
1949 next_port = low;
1950 else
1951 next_port = port + 1;
1953 bind_list->ps = ps;
1954 bind_list->port = (unsigned short) port;
1955 cma_bind_port(bind_list, id_priv);
1956 return 0;
1957 err2:
1958 idr_remove(ps, port);
1959 err1:
1960 kfree(bind_list);
1961 return ret;
1964 static int cma_use_port(struct idr *ps, struct rdma_id_private *id_priv)
1966 struct rdma_id_private *cur_id;
1967 struct sockaddr_in *sin, *cur_sin;
1968 struct rdma_bind_list *bind_list;
1969 struct hlist_node *node;
1970 unsigned short snum;
1972 sin = (struct sockaddr_in *) &id_priv->id.route.addr.src_addr;
1973 snum = ntohs(sin->sin_port);
1974 if (snum < PROT_SOCK && !capable(CAP_NET_BIND_SERVICE))
1975 return -EACCES;
1977 bind_list = idr_find(ps, snum);
1978 if (!bind_list)
1979 return cma_alloc_port(ps, id_priv, snum);
1982 * We don't support binding to any address if anyone is bound to
1983 * a specific address on the same port.
1985 if (cma_any_addr(&id_priv->id.route.addr.src_addr))
1986 return -EADDRNOTAVAIL;
1988 hlist_for_each_entry(cur_id, node, &bind_list->owners, node) {
1989 if (cma_any_addr(&cur_id->id.route.addr.src_addr))
1990 return -EADDRNOTAVAIL;
1992 cur_sin = (struct sockaddr_in *) &cur_id->id.route.addr.src_addr;
1993 if (sin->sin_addr.s_addr == cur_sin->sin_addr.s_addr)
1994 return -EADDRINUSE;
1997 cma_bind_port(bind_list, id_priv);
1998 return 0;
2001 static int cma_get_port(struct rdma_id_private *id_priv)
2003 struct idr *ps;
2004 int ret;
2006 switch (id_priv->id.ps) {
2007 case RDMA_PS_SDP:
2008 ps = &sdp_ps;
2009 break;
2010 case RDMA_PS_TCP:
2011 ps = &tcp_ps;
2012 break;
2013 case RDMA_PS_UDP:
2014 ps = &udp_ps;
2015 break;
2016 case RDMA_PS_IPOIB:
2017 ps = &ipoib_ps;
2018 break;
2019 default:
2020 return -EPROTONOSUPPORT;
2023 mutex_lock(&lock);
2024 if (cma_any_port(&id_priv->id.route.addr.src_addr))
2025 ret = cma_alloc_any_port(ps, id_priv);
2026 else
2027 ret = cma_use_port(ps, id_priv);
2028 mutex_unlock(&lock);
2030 return ret;
2033 int rdma_bind_addr(struct rdma_cm_id *id, struct sockaddr *addr)
2035 struct rdma_id_private *id_priv;
2036 int ret;
2038 if (addr->sa_family != AF_INET)
2039 return -EAFNOSUPPORT;
2041 id_priv = container_of(id, struct rdma_id_private, id);
2042 if (!cma_comp_exch(id_priv, CMA_IDLE, CMA_ADDR_BOUND))
2043 return -EINVAL;
2045 if (!cma_any_addr(addr)) {
2046 ret = rdma_translate_ip(addr, &id->route.addr.dev_addr);
2047 if (ret)
2048 goto err1;
2050 mutex_lock(&lock);
2051 ret = cma_acquire_dev(id_priv);
2052 mutex_unlock(&lock);
2053 if (ret)
2054 goto err1;
2057 memcpy(&id->route.addr.src_addr, addr, ip_addr_size(addr));
2058 ret = cma_get_port(id_priv);
2059 if (ret)
2060 goto err2;
2062 return 0;
2063 err2:
2064 if (!cma_any_addr(addr)) {
2065 mutex_lock(&lock);
2066 cma_detach_from_dev(id_priv);
2067 mutex_unlock(&lock);
2069 err1:
2070 cma_comp_exch(id_priv, CMA_ADDR_BOUND, CMA_IDLE);
2071 return ret;
2073 EXPORT_SYMBOL(rdma_bind_addr);
2075 static int cma_format_hdr(void *hdr, enum rdma_port_space ps,
2076 struct rdma_route *route)
2078 struct sockaddr_in *src4, *dst4;
2079 struct cma_hdr *cma_hdr;
2080 struct sdp_hh *sdp_hdr;
2082 src4 = (struct sockaddr_in *) &route->addr.src_addr;
2083 dst4 = (struct sockaddr_in *) &route->addr.dst_addr;
2085 switch (ps) {
2086 case RDMA_PS_SDP:
2087 sdp_hdr = hdr;
2088 if (sdp_get_majv(sdp_hdr->sdp_version) != SDP_MAJ_VERSION)
2089 return -EINVAL;
2090 sdp_set_ip_ver(sdp_hdr, 4);
2091 sdp_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2092 sdp_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2093 sdp_hdr->port = src4->sin_port;
2094 break;
2095 default:
2096 cma_hdr = hdr;
2097 cma_hdr->cma_version = CMA_VERSION;
2098 cma_set_ip_ver(cma_hdr, 4);
2099 cma_hdr->src_addr.ip4.addr = src4->sin_addr.s_addr;
2100 cma_hdr->dst_addr.ip4.addr = dst4->sin_addr.s_addr;
2101 cma_hdr->port = src4->sin_port;
2102 break;
2104 return 0;
2107 static int cma_sidr_rep_handler(struct ib_cm_id *cm_id,
2108 struct ib_cm_event *ib_event)
2110 struct rdma_id_private *id_priv = cm_id->context;
2111 struct rdma_cm_event event;
2112 struct ib_cm_sidr_rep_event_param *rep = &ib_event->param.sidr_rep_rcvd;
2113 int ret = 0;
2115 if (cma_disable_remove(id_priv, CMA_CONNECT))
2116 return 0;
2118 memset(&event, 0, sizeof event);
2119 switch (ib_event->event) {
2120 case IB_CM_SIDR_REQ_ERROR:
2121 event.event = RDMA_CM_EVENT_UNREACHABLE;
2122 event.status = -ETIMEDOUT;
2123 break;
2124 case IB_CM_SIDR_REP_RECEIVED:
2125 event.param.ud.private_data = ib_event->private_data;
2126 event.param.ud.private_data_len = IB_CM_SIDR_REP_PRIVATE_DATA_SIZE;
2127 if (rep->status != IB_SIDR_SUCCESS) {
2128 event.event = RDMA_CM_EVENT_UNREACHABLE;
2129 event.status = ib_event->param.sidr_rep_rcvd.status;
2130 break;
2132 if (id_priv->qkey != rep->qkey) {
2133 event.event = RDMA_CM_EVENT_UNREACHABLE;
2134 event.status = -EINVAL;
2135 break;
2137 ib_init_ah_from_path(id_priv->id.device, id_priv->id.port_num,
2138 id_priv->id.route.path_rec,
2139 &event.param.ud.ah_attr);
2140 event.param.ud.qp_num = rep->qpn;
2141 event.param.ud.qkey = rep->qkey;
2142 event.event = RDMA_CM_EVENT_ESTABLISHED;
2143 event.status = 0;
2144 break;
2145 default:
2146 printk(KERN_ERR "RDMA CMA: unexpected IB CM event: %d",
2147 ib_event->event);
2148 goto out;
2151 ret = id_priv->id.event_handler(&id_priv->id, &event);
2152 if (ret) {
2153 /* Destroy the CM ID by returning a non-zero value. */
2154 id_priv->cm_id.ib = NULL;
2155 cma_exch(id_priv, CMA_DESTROYING);
2156 cma_enable_remove(id_priv);
2157 rdma_destroy_id(&id_priv->id);
2158 return ret;
2160 out:
2161 cma_enable_remove(id_priv);
2162 return ret;
2165 static int cma_resolve_ib_udp(struct rdma_id_private *id_priv,
2166 struct rdma_conn_param *conn_param)
2168 struct ib_cm_sidr_req_param req;
2169 struct rdma_route *route;
2170 int ret;
2172 req.private_data_len = sizeof(struct cma_hdr) +
2173 conn_param->private_data_len;
2174 req.private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2175 if (!req.private_data)
2176 return -ENOMEM;
2178 if (conn_param->private_data && conn_param->private_data_len)
2179 memcpy((void *) req.private_data + sizeof(struct cma_hdr),
2180 conn_param->private_data, conn_param->private_data_len);
2182 route = &id_priv->id.route;
2183 ret = cma_format_hdr((void *) req.private_data, id_priv->id.ps, route);
2184 if (ret)
2185 goto out;
2187 id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device,
2188 cma_sidr_rep_handler, id_priv);
2189 if (IS_ERR(id_priv->cm_id.ib)) {
2190 ret = PTR_ERR(id_priv->cm_id.ib);
2191 goto out;
2194 req.path = route->path_rec;
2195 req.service_id = cma_get_service_id(id_priv->id.ps,
2196 &route->addr.dst_addr);
2197 req.timeout_ms = 1 << (CMA_CM_RESPONSE_TIMEOUT - 8);
2198 req.max_cm_retries = CMA_MAX_CM_RETRIES;
2200 ret = ib_send_cm_sidr_req(id_priv->cm_id.ib, &req);
2201 if (ret) {
2202 ib_destroy_cm_id(id_priv->cm_id.ib);
2203 id_priv->cm_id.ib = NULL;
2205 out:
2206 kfree(req.private_data);
2207 return ret;
2210 static int cma_connect_ib(struct rdma_id_private *id_priv,
2211 struct rdma_conn_param *conn_param)
2213 struct ib_cm_req_param req;
2214 struct rdma_route *route;
2215 void *private_data;
2216 int offset, ret;
2218 memset(&req, 0, sizeof req);
2219 offset = cma_user_data_offset(id_priv->id.ps);
2220 req.private_data_len = offset + conn_param->private_data_len;
2221 private_data = kzalloc(req.private_data_len, GFP_ATOMIC);
2222 if (!private_data)
2223 return -ENOMEM;
2225 if (conn_param->private_data && conn_param->private_data_len)
2226 memcpy(private_data + offset, conn_param->private_data,
2227 conn_param->private_data_len);
2229 id_priv->cm_id.ib = ib_create_cm_id(id_priv->id.device, cma_ib_handler,
2230 id_priv);
2231 if (IS_ERR(id_priv->cm_id.ib)) {
2232 ret = PTR_ERR(id_priv->cm_id.ib);
2233 goto out;
2236 route = &id_priv->id.route;
2237 ret = cma_format_hdr(private_data, id_priv->id.ps, route);
2238 if (ret)
2239 goto out;
2240 req.private_data = private_data;
2242 req.primary_path = &route->path_rec[0];
2243 if (route->num_paths == 2)
2244 req.alternate_path = &route->path_rec[1];
2246 req.service_id = cma_get_service_id(id_priv->id.ps,
2247 &route->addr.dst_addr);
2248 req.qp_num = id_priv->qp_num;
2249 req.qp_type = IB_QPT_RC;
2250 req.starting_psn = id_priv->seq_num;
2251 req.responder_resources = conn_param->responder_resources;
2252 req.initiator_depth = conn_param->initiator_depth;
2253 req.flow_control = conn_param->flow_control;
2254 req.retry_count = conn_param->retry_count;
2255 req.rnr_retry_count = conn_param->rnr_retry_count;
2256 req.remote_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2257 req.local_cm_response_timeout = CMA_CM_RESPONSE_TIMEOUT;
2258 req.max_cm_retries = CMA_MAX_CM_RETRIES;
2259 req.srq = id_priv->srq ? 1 : 0;
2261 ret = ib_send_cm_req(id_priv->cm_id.ib, &req);
2262 out:
2263 if (ret && !IS_ERR(id_priv->cm_id.ib)) {
2264 ib_destroy_cm_id(id_priv->cm_id.ib);
2265 id_priv->cm_id.ib = NULL;
2268 kfree(private_data);
2269 return ret;
2272 static int cma_connect_iw(struct rdma_id_private *id_priv,
2273 struct rdma_conn_param *conn_param)
2275 struct iw_cm_id *cm_id;
2276 struct sockaddr_in* sin;
2277 int ret;
2278 struct iw_cm_conn_param iw_param;
2280 cm_id = iw_create_cm_id(id_priv->id.device, cma_iw_handler, id_priv);
2281 if (IS_ERR(cm_id)) {
2282 ret = PTR_ERR(cm_id);
2283 goto out;
2286 id_priv->cm_id.iw = cm_id;
2288 sin = (struct sockaddr_in*) &id_priv->id.route.addr.src_addr;
2289 cm_id->local_addr = *sin;
2291 sin = (struct sockaddr_in*) &id_priv->id.route.addr.dst_addr;
2292 cm_id->remote_addr = *sin;
2294 ret = cma_modify_qp_rtr(id_priv, conn_param);
2295 if (ret)
2296 goto out;
2298 iw_param.ord = conn_param->initiator_depth;
2299 iw_param.ird = conn_param->responder_resources;
2300 iw_param.private_data = conn_param->private_data;
2301 iw_param.private_data_len = conn_param->private_data_len;
2302 if (id_priv->id.qp)
2303 iw_param.qpn = id_priv->qp_num;
2304 else
2305 iw_param.qpn = conn_param->qp_num;
2306 ret = iw_cm_connect(cm_id, &iw_param);
2307 out:
2308 if (ret && !IS_ERR(cm_id)) {
2309 iw_destroy_cm_id(cm_id);
2310 id_priv->cm_id.iw = NULL;
2312 return ret;
2315 int rdma_connect(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2317 struct rdma_id_private *id_priv;
2318 int ret;
2320 id_priv = container_of(id, struct rdma_id_private, id);
2321 if (!cma_comp_exch(id_priv, CMA_ROUTE_RESOLVED, CMA_CONNECT))
2322 return -EINVAL;
2324 if (!id->qp) {
2325 id_priv->qp_num = conn_param->qp_num;
2326 id_priv->srq = conn_param->srq;
2329 switch (rdma_node_get_transport(id->device->node_type)) {
2330 case RDMA_TRANSPORT_IB:
2331 if (cma_is_ud_ps(id->ps))
2332 ret = cma_resolve_ib_udp(id_priv, conn_param);
2333 else
2334 ret = cma_connect_ib(id_priv, conn_param);
2335 break;
2336 case RDMA_TRANSPORT_IWARP:
2337 ret = cma_connect_iw(id_priv, conn_param);
2338 break;
2339 default:
2340 ret = -ENOSYS;
2341 break;
2343 if (ret)
2344 goto err;
2346 return 0;
2347 err:
2348 cma_comp_exch(id_priv, CMA_CONNECT, CMA_ROUTE_RESOLVED);
2349 return ret;
2351 EXPORT_SYMBOL(rdma_connect);
2353 static int cma_accept_ib(struct rdma_id_private *id_priv,
2354 struct rdma_conn_param *conn_param)
2356 struct ib_cm_rep_param rep;
2357 int ret;
2359 ret = cma_modify_qp_rtr(id_priv, conn_param);
2360 if (ret)
2361 goto out;
2363 ret = cma_modify_qp_rts(id_priv, conn_param);
2364 if (ret)
2365 goto out;
2367 memset(&rep, 0, sizeof rep);
2368 rep.qp_num = id_priv->qp_num;
2369 rep.starting_psn = id_priv->seq_num;
2370 rep.private_data = conn_param->private_data;
2371 rep.private_data_len = conn_param->private_data_len;
2372 rep.responder_resources = conn_param->responder_resources;
2373 rep.initiator_depth = conn_param->initiator_depth;
2374 rep.failover_accepted = 0;
2375 rep.flow_control = conn_param->flow_control;
2376 rep.rnr_retry_count = conn_param->rnr_retry_count;
2377 rep.srq = id_priv->srq ? 1 : 0;
2379 ret = ib_send_cm_rep(id_priv->cm_id.ib, &rep);
2380 out:
2381 return ret;
2384 static int cma_accept_iw(struct rdma_id_private *id_priv,
2385 struct rdma_conn_param *conn_param)
2387 struct iw_cm_conn_param iw_param;
2388 int ret;
2390 ret = cma_modify_qp_rtr(id_priv, conn_param);
2391 if (ret)
2392 return ret;
2394 iw_param.ord = conn_param->initiator_depth;
2395 iw_param.ird = conn_param->responder_resources;
2396 iw_param.private_data = conn_param->private_data;
2397 iw_param.private_data_len = conn_param->private_data_len;
2398 if (id_priv->id.qp) {
2399 iw_param.qpn = id_priv->qp_num;
2400 } else
2401 iw_param.qpn = conn_param->qp_num;
2403 return iw_cm_accept(id_priv->cm_id.iw, &iw_param);
2406 static int cma_send_sidr_rep(struct rdma_id_private *id_priv,
2407 enum ib_cm_sidr_status status,
2408 const void *private_data, int private_data_len)
2410 struct ib_cm_sidr_rep_param rep;
2412 memset(&rep, 0, sizeof rep);
2413 rep.status = status;
2414 if (status == IB_SIDR_SUCCESS) {
2415 rep.qp_num = id_priv->qp_num;
2416 rep.qkey = id_priv->qkey;
2418 rep.private_data = private_data;
2419 rep.private_data_len = private_data_len;
2421 return ib_send_cm_sidr_rep(id_priv->cm_id.ib, &rep);
2424 int rdma_accept(struct rdma_cm_id *id, struct rdma_conn_param *conn_param)
2426 struct rdma_id_private *id_priv;
2427 int ret;
2429 id_priv = container_of(id, struct rdma_id_private, id);
2430 if (!cma_comp(id_priv, CMA_CONNECT))
2431 return -EINVAL;
2433 if (!id->qp && conn_param) {
2434 id_priv->qp_num = conn_param->qp_num;
2435 id_priv->srq = conn_param->srq;
2438 switch (rdma_node_get_transport(id->device->node_type)) {
2439 case RDMA_TRANSPORT_IB:
2440 if (cma_is_ud_ps(id->ps))
2441 ret = cma_send_sidr_rep(id_priv, IB_SIDR_SUCCESS,
2442 conn_param->private_data,
2443 conn_param->private_data_len);
2444 else if (conn_param)
2445 ret = cma_accept_ib(id_priv, conn_param);
2446 else
2447 ret = cma_rep_recv(id_priv);
2448 break;
2449 case RDMA_TRANSPORT_IWARP:
2450 ret = cma_accept_iw(id_priv, conn_param);
2451 break;
2452 default:
2453 ret = -ENOSYS;
2454 break;
2457 if (ret)
2458 goto reject;
2460 return 0;
2461 reject:
2462 cma_modify_qp_err(id_priv);
2463 rdma_reject(id, NULL, 0);
2464 return ret;
2466 EXPORT_SYMBOL(rdma_accept);
2468 int rdma_notify(struct rdma_cm_id *id, enum ib_event_type event)
2470 struct rdma_id_private *id_priv;
2471 int ret;
2473 id_priv = container_of(id, struct rdma_id_private, id);
2474 if (!cma_has_cm_dev(id_priv))
2475 return -EINVAL;
2477 switch (id->device->node_type) {
2478 case RDMA_NODE_IB_CA:
2479 ret = ib_cm_notify(id_priv->cm_id.ib, event);
2480 break;
2481 default:
2482 ret = 0;
2483 break;
2485 return ret;
2487 EXPORT_SYMBOL(rdma_notify);
2489 int rdma_reject(struct rdma_cm_id *id, const void *private_data,
2490 u8 private_data_len)
2492 struct rdma_id_private *id_priv;
2493 int ret;
2495 id_priv = container_of(id, struct rdma_id_private, id);
2496 if (!cma_has_cm_dev(id_priv))
2497 return -EINVAL;
2499 switch (rdma_node_get_transport(id->device->node_type)) {
2500 case RDMA_TRANSPORT_IB:
2501 if (cma_is_ud_ps(id->ps))
2502 ret = cma_send_sidr_rep(id_priv, IB_SIDR_REJECT,
2503 private_data, private_data_len);
2504 else
2505 ret = ib_send_cm_rej(id_priv->cm_id.ib,
2506 IB_CM_REJ_CONSUMER_DEFINED, NULL,
2507 0, private_data, private_data_len);
2508 break;
2509 case RDMA_TRANSPORT_IWARP:
2510 ret = iw_cm_reject(id_priv->cm_id.iw,
2511 private_data, private_data_len);
2512 break;
2513 default:
2514 ret = -ENOSYS;
2515 break;
2517 return ret;
2519 EXPORT_SYMBOL(rdma_reject);
2521 int rdma_disconnect(struct rdma_cm_id *id)
2523 struct rdma_id_private *id_priv;
2524 int ret;
2526 id_priv = container_of(id, struct rdma_id_private, id);
2527 if (!cma_has_cm_dev(id_priv))
2528 return -EINVAL;
2530 switch (rdma_node_get_transport(id->device->node_type)) {
2531 case RDMA_TRANSPORT_IB:
2532 ret = cma_modify_qp_err(id_priv);
2533 if (ret)
2534 goto out;
2535 /* Initiate or respond to a disconnect. */
2536 if (ib_send_cm_dreq(id_priv->cm_id.ib, NULL, 0))
2537 ib_send_cm_drep(id_priv->cm_id.ib, NULL, 0);
2538 break;
2539 case RDMA_TRANSPORT_IWARP:
2540 ret = iw_cm_disconnect(id_priv->cm_id.iw, 0);
2541 break;
2542 default:
2543 ret = -EINVAL;
2544 break;
2546 out:
2547 return ret;
2549 EXPORT_SYMBOL(rdma_disconnect);
2551 static int cma_ib_mc_handler(int status, struct ib_sa_multicast *multicast)
2553 struct rdma_id_private *id_priv;
2554 struct cma_multicast *mc = multicast->context;
2555 struct rdma_cm_event event;
2556 int ret;
2558 id_priv = mc->id_priv;
2559 if (cma_disable_remove(id_priv, CMA_ADDR_BOUND) &&
2560 cma_disable_remove(id_priv, CMA_ADDR_RESOLVED))
2561 return 0;
2563 mutex_lock(&id_priv->qp_mutex);
2564 if (!status && id_priv->id.qp)
2565 status = ib_attach_mcast(id_priv->id.qp, &multicast->rec.mgid,
2566 multicast->rec.mlid);
2567 mutex_unlock(&id_priv->qp_mutex);
2569 memset(&event, 0, sizeof event);
2570 event.status = status;
2571 event.param.ud.private_data = mc->context;
2572 if (!status) {
2573 event.event = RDMA_CM_EVENT_MULTICAST_JOIN;
2574 ib_init_ah_from_mcmember(id_priv->id.device,
2575 id_priv->id.port_num, &multicast->rec,
2576 &event.param.ud.ah_attr);
2577 event.param.ud.qp_num = 0xFFFFFF;
2578 event.param.ud.qkey = be32_to_cpu(multicast->rec.qkey);
2579 } else
2580 event.event = RDMA_CM_EVENT_MULTICAST_ERROR;
2582 ret = id_priv->id.event_handler(&id_priv->id, &event);
2583 if (ret) {
2584 cma_exch(id_priv, CMA_DESTROYING);
2585 cma_enable_remove(id_priv);
2586 rdma_destroy_id(&id_priv->id);
2587 return 0;
2590 cma_enable_remove(id_priv);
2591 return 0;
2594 static void cma_set_mgid(struct rdma_id_private *id_priv,
2595 struct sockaddr *addr, union ib_gid *mgid)
2597 unsigned char mc_map[MAX_ADDR_LEN];
2598 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2599 struct sockaddr_in *sin = (struct sockaddr_in *) addr;
2600 struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *) addr;
2602 if (cma_any_addr(addr)) {
2603 memset(mgid, 0, sizeof *mgid);
2604 } else if ((addr->sa_family == AF_INET6) &&
2605 ((be32_to_cpu(sin6->sin6_addr.s6_addr32[0]) & 0xFF10A01B) ==
2606 0xFF10A01B)) {
2607 /* IPv6 address is an SA assigned MGID. */
2608 memcpy(mgid, &sin6->sin6_addr, sizeof *mgid);
2609 } else {
2610 ip_ib_mc_map(sin->sin_addr.s_addr, dev_addr->broadcast, mc_map);
2611 if (id_priv->id.ps == RDMA_PS_UDP)
2612 mc_map[7] = 0x01; /* Use RDMA CM signature */
2613 *mgid = *(union ib_gid *) (mc_map + 4);
2617 static int cma_join_ib_multicast(struct rdma_id_private *id_priv,
2618 struct cma_multicast *mc)
2620 struct ib_sa_mcmember_rec rec;
2621 struct rdma_dev_addr *dev_addr = &id_priv->id.route.addr.dev_addr;
2622 ib_sa_comp_mask comp_mask;
2623 int ret;
2625 ib_addr_get_mgid(dev_addr, &rec.mgid);
2626 ret = ib_sa_get_mcmember_rec(id_priv->id.device, id_priv->id.port_num,
2627 &rec.mgid, &rec);
2628 if (ret)
2629 return ret;
2631 cma_set_mgid(id_priv, &mc->addr, &rec.mgid);
2632 if (id_priv->id.ps == RDMA_PS_UDP)
2633 rec.qkey = cpu_to_be32(RDMA_UDP_QKEY);
2634 ib_addr_get_sgid(dev_addr, &rec.port_gid);
2635 rec.pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
2636 rec.join_state = 1;
2638 comp_mask = IB_SA_MCMEMBER_REC_MGID | IB_SA_MCMEMBER_REC_PORT_GID |
2639 IB_SA_MCMEMBER_REC_PKEY | IB_SA_MCMEMBER_REC_JOIN_STATE |
2640 IB_SA_MCMEMBER_REC_QKEY | IB_SA_MCMEMBER_REC_SL |
2641 IB_SA_MCMEMBER_REC_FLOW_LABEL |
2642 IB_SA_MCMEMBER_REC_TRAFFIC_CLASS;
2644 mc->multicast.ib = ib_sa_join_multicast(&sa_client, id_priv->id.device,
2645 id_priv->id.port_num, &rec,
2646 comp_mask, GFP_KERNEL,
2647 cma_ib_mc_handler, mc);
2648 if (IS_ERR(mc->multicast.ib))
2649 return PTR_ERR(mc->multicast.ib);
2651 return 0;
2654 int rdma_join_multicast(struct rdma_cm_id *id, struct sockaddr *addr,
2655 void *context)
2657 struct rdma_id_private *id_priv;
2658 struct cma_multicast *mc;
2659 int ret;
2661 id_priv = container_of(id, struct rdma_id_private, id);
2662 if (!cma_comp(id_priv, CMA_ADDR_BOUND) &&
2663 !cma_comp(id_priv, CMA_ADDR_RESOLVED))
2664 return -EINVAL;
2666 mc = kmalloc(sizeof *mc, GFP_KERNEL);
2667 if (!mc)
2668 return -ENOMEM;
2670 memcpy(&mc->addr, addr, ip_addr_size(addr));
2671 mc->context = context;
2672 mc->id_priv = id_priv;
2674 spin_lock(&id_priv->lock);
2675 list_add(&mc->list, &id_priv->mc_list);
2676 spin_unlock(&id_priv->lock);
2678 switch (rdma_node_get_transport(id->device->node_type)) {
2679 case RDMA_TRANSPORT_IB:
2680 ret = cma_join_ib_multicast(id_priv, mc);
2681 break;
2682 default:
2683 ret = -ENOSYS;
2684 break;
2687 if (ret) {
2688 spin_lock_irq(&id_priv->lock);
2689 list_del(&mc->list);
2690 spin_unlock_irq(&id_priv->lock);
2691 kfree(mc);
2693 return ret;
2695 EXPORT_SYMBOL(rdma_join_multicast);
2697 void rdma_leave_multicast(struct rdma_cm_id *id, struct sockaddr *addr)
2699 struct rdma_id_private *id_priv;
2700 struct cma_multicast *mc;
2702 id_priv = container_of(id, struct rdma_id_private, id);
2703 spin_lock_irq(&id_priv->lock);
2704 list_for_each_entry(mc, &id_priv->mc_list, list) {
2705 if (!memcmp(&mc->addr, addr, ip_addr_size(addr))) {
2706 list_del(&mc->list);
2707 spin_unlock_irq(&id_priv->lock);
2709 if (id->qp)
2710 ib_detach_mcast(id->qp,
2711 &mc->multicast.ib->rec.mgid,
2712 mc->multicast.ib->rec.mlid);
2713 ib_sa_free_multicast(mc->multicast.ib);
2714 kfree(mc);
2715 return;
2718 spin_unlock_irq(&id_priv->lock);
2720 EXPORT_SYMBOL(rdma_leave_multicast);
2722 static void cma_add_one(struct ib_device *device)
2724 struct cma_device *cma_dev;
2725 struct rdma_id_private *id_priv;
2727 cma_dev = kmalloc(sizeof *cma_dev, GFP_KERNEL);
2728 if (!cma_dev)
2729 return;
2731 cma_dev->device = device;
2733 init_completion(&cma_dev->comp);
2734 atomic_set(&cma_dev->refcount, 1);
2735 INIT_LIST_HEAD(&cma_dev->id_list);
2736 ib_set_client_data(device, &cma_client, cma_dev);
2738 mutex_lock(&lock);
2739 list_add_tail(&cma_dev->list, &dev_list);
2740 list_for_each_entry(id_priv, &listen_any_list, list)
2741 cma_listen_on_dev(id_priv, cma_dev);
2742 mutex_unlock(&lock);
2745 static int cma_remove_id_dev(struct rdma_id_private *id_priv)
2747 struct rdma_cm_event event;
2748 enum cma_state state;
2750 /* Record that we want to remove the device */
2751 state = cma_exch(id_priv, CMA_DEVICE_REMOVAL);
2752 if (state == CMA_DESTROYING)
2753 return 0;
2755 cma_cancel_operation(id_priv, state);
2756 wait_event(id_priv->wait_remove, !atomic_read(&id_priv->dev_remove));
2758 /* Check for destruction from another callback. */
2759 if (!cma_comp(id_priv, CMA_DEVICE_REMOVAL))
2760 return 0;
2762 memset(&event, 0, sizeof event);
2763 event.event = RDMA_CM_EVENT_DEVICE_REMOVAL;
2764 return id_priv->id.event_handler(&id_priv->id, &event);
2767 static void cma_process_remove(struct cma_device *cma_dev)
2769 struct rdma_id_private *id_priv;
2770 int ret;
2772 mutex_lock(&lock);
2773 while (!list_empty(&cma_dev->id_list)) {
2774 id_priv = list_entry(cma_dev->id_list.next,
2775 struct rdma_id_private, list);
2777 list_del(&id_priv->listen_list);
2778 list_del_init(&id_priv->list);
2779 atomic_inc(&id_priv->refcount);
2780 mutex_unlock(&lock);
2782 ret = id_priv->internal_id ? 1 : cma_remove_id_dev(id_priv);
2783 cma_deref_id(id_priv);
2784 if (ret)
2785 rdma_destroy_id(&id_priv->id);
2787 mutex_lock(&lock);
2789 mutex_unlock(&lock);
2791 cma_deref_dev(cma_dev);
2792 wait_for_completion(&cma_dev->comp);
2795 static void cma_remove_one(struct ib_device *device)
2797 struct cma_device *cma_dev;
2799 cma_dev = ib_get_client_data(device, &cma_client);
2800 if (!cma_dev)
2801 return;
2803 mutex_lock(&lock);
2804 list_del(&cma_dev->list);
2805 mutex_unlock(&lock);
2807 cma_process_remove(cma_dev);
2808 kfree(cma_dev);
2811 static int cma_init(void)
2813 int ret, low, high, remaining;
2815 get_random_bytes(&next_port, sizeof next_port);
2816 inet_get_local_port_range(&low, &high);
2817 remaining = (high - low) + 1;
2818 next_port = ((unsigned int) next_port % remaining) + low;
2820 cma_wq = create_singlethread_workqueue("rdma_cm");
2821 if (!cma_wq)
2822 return -ENOMEM;
2824 ib_sa_register_client(&sa_client);
2825 rdma_addr_register_client(&addr_client);
2827 ret = ib_register_client(&cma_client);
2828 if (ret)
2829 goto err;
2830 return 0;
2832 err:
2833 rdma_addr_unregister_client(&addr_client);
2834 ib_sa_unregister_client(&sa_client);
2835 destroy_workqueue(cma_wq);
2836 return ret;
2839 static void cma_cleanup(void)
2841 ib_unregister_client(&cma_client);
2842 rdma_addr_unregister_client(&addr_client);
2843 ib_sa_unregister_client(&sa_client);
2844 destroy_workqueue(cma_wq);
2845 idr_destroy(&sdp_ps);
2846 idr_destroy(&tcp_ps);
2847 idr_destroy(&udp_ps);
2848 idr_destroy(&ipoib_ps);
2851 module_init(cma_init);
2852 module_exit(cma_cleanup);