2 * Copyright (c) 2004, 2005 Intel Corporation. All rights reserved.
3 * Copyright (c) 2004 Topspin Corporation. All rights reserved.
4 * Copyright (c) 2004, 2005 Voltaire Corporation. All rights reserved.
5 * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
6 * Copyright (c) 2005 Open Grid Computing, Inc. All rights reserved.
7 * Copyright (c) 2005 Network Appliance, Inc. All rights reserved.
9 * This software is available to you under a choice of one of two
10 * licenses. You may choose to be licensed under the terms of the GNU
11 * General Public License (GPL) Version 2, available from the file
12 * COPYING in the main directory of this source tree, or the
13 * OpenIB.org BSD license below:
15 * Redistribution and use in source and binary forms, with or
16 * without modification, are permitted provided that the following
19 * - Redistributions of source code must retain the above
20 * copyright notice, this list of conditions and the following
23 * - Redistributions in binary form must reproduce the above
24 * copyright notice, this list of conditions and the following
25 * disclaimer in the documentation and/or other materials
26 * provided with the distribution.
28 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
29 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
30 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
31 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
32 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
33 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
34 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
38 #include <linux/dma-mapping.h>
39 #include <linux/err.h>
40 #include <linux/idr.h>
41 #include <linux/interrupt.h>
42 #include <linux/rbtree.h>
43 #include <linux/sched.h>
44 #include <linux/spinlock.h>
45 #include <linux/workqueue.h>
46 #include <linux/completion.h>
47 #include <linux/slab.h>
48 #include <linux/module.h>
49 #include <linux/sysctl.h>
51 #include <rdma/iw_cm.h>
52 #include <rdma/ib_addr.h>
53 #include <rdma/iw_portmap.h>
54 #include <rdma/rdma_netlink.h>
58 MODULE_AUTHOR("Tom Tucker");
59 MODULE_DESCRIPTION("iWARP CM");
60 MODULE_LICENSE("Dual BSD/GPL");
62 static const char * const iwcm_rej_reason_strs[] = {
63 [ECONNRESET] = "reset by remote host",
64 [ECONNREFUSED] = "refused by remote application",
65 [ETIMEDOUT] = "setup timeout",
68 const char *__attribute_const__ iwcm_reject_msg(int reason)
72 /* iWARP uses negative errnos */
75 if (index < ARRAY_SIZE(iwcm_rej_reason_strs) &&
76 iwcm_rej_reason_strs[index])
77 return iwcm_rej_reason_strs[index];
79 return "unrecognized reason";
81 EXPORT_SYMBOL(iwcm_reject_msg);
83 static struct rdma_nl_cbs iwcm_nl_cb_table[RDMA_NL_IWPM_NUM_OPS] = {
84 [RDMA_NL_IWPM_REG_PID] = {.dump = iwpm_register_pid_cb},
85 [RDMA_NL_IWPM_ADD_MAPPING] = {.dump = iwpm_add_mapping_cb},
86 [RDMA_NL_IWPM_QUERY_MAPPING] = {.dump = iwpm_add_and_query_mapping_cb},
87 [RDMA_NL_IWPM_REMOTE_INFO] = {.dump = iwpm_remote_info_cb},
88 [RDMA_NL_IWPM_HANDLE_ERR] = {.dump = iwpm_mapping_error_cb},
89 [RDMA_NL_IWPM_MAPINFO] = {.dump = iwpm_mapping_info_cb},
90 [RDMA_NL_IWPM_MAPINFO_NUM] = {.dump = iwpm_ack_mapping_info_cb}
93 static struct workqueue_struct *iwcm_wq;
95 struct work_struct work;
96 struct iwcm_id_private *cm_id;
97 struct list_head list;
98 struct iw_cm_event event;
99 struct list_head free_list;
102 static unsigned int default_backlog = 256;
104 static struct ctl_table_header *iwcm_ctl_table_hdr;
105 static struct ctl_table iwcm_ctl_table[] = {
107 .procname = "default_backlog",
108 .data = &default_backlog,
109 .maxlen = sizeof(default_backlog),
111 .proc_handler = proc_dointvec,
117 * The following services provide a mechanism for pre-allocating iwcm_work
118 * elements. The design pre-allocates them based on the cm_id type:
119 * LISTENING IDS: Get enough elements preallocated to handle the
121 * ACTIVE IDS: 4: CONNECT_REPLY, ESTABLISHED, DISCONNECT, CLOSE
122 * PASSIVE IDS: 3: ESTABLISHED, DISCONNECT, CLOSE
124 * Allocating them in connect and listen avoids having to deal
125 * with allocation failures on the event upcall from the provider (which
126 * is called in the interrupt context).
128 * One exception is when creating the cm_id for incoming connection requests.
129 * There are two cases:
130 * 1) in the event upcall, cm_event_handler(), for a listening cm_id. If
131 * the backlog is exceeded, then no more connection request events will
132 * be processed. cm_event_handler() returns -ENOMEM in this case. Its up
133 * to the provider to reject the connection request.
134 * 2) in the connection request workqueue handler, cm_conn_req_handler().
135 * If work elements cannot be allocated for the new connect request cm_id,
136 * then IWCM will call the provider reject method. This is ok since
137 * cm_conn_req_handler() runs in the workqueue thread context.
140 static struct iwcm_work *get_work(struct iwcm_id_private *cm_id_priv)
142 struct iwcm_work *work;
144 if (list_empty(&cm_id_priv->work_free_list))
146 work = list_entry(cm_id_priv->work_free_list.next, struct iwcm_work,
148 list_del_init(&work->free_list);
152 static void put_work(struct iwcm_work *work)
154 list_add(&work->free_list, &work->cm_id->work_free_list);
157 static void dealloc_work_entries(struct iwcm_id_private *cm_id_priv)
159 struct list_head *e, *tmp;
161 list_for_each_safe(e, tmp, &cm_id_priv->work_free_list) {
163 kfree(list_entry(e, struct iwcm_work, free_list));
167 static int alloc_work_entries(struct iwcm_id_private *cm_id_priv, int count)
169 struct iwcm_work *work;
171 BUG_ON(!list_empty(&cm_id_priv->work_free_list));
173 work = kmalloc(sizeof(struct iwcm_work), GFP_KERNEL);
175 dealloc_work_entries(cm_id_priv);
178 work->cm_id = cm_id_priv;
179 INIT_LIST_HEAD(&work->list);
186 * Save private data from incoming connection requests to
187 * iw_cm_event, so the low level driver doesn't have to. Adjust
188 * the event ptr to point to the local copy.
190 static int copy_private_data(struct iw_cm_event *event)
194 p = kmemdup(event->private_data, event->private_data_len, GFP_ATOMIC);
197 event->private_data = p;
201 static void free_cm_id(struct iwcm_id_private *cm_id_priv)
203 dealloc_work_entries(cm_id_priv);
208 * Release a reference on cm_id. If the last reference is being
209 * released, free the cm_id and return 1.
211 static int iwcm_deref_id(struct iwcm_id_private *cm_id_priv)
213 BUG_ON(atomic_read(&cm_id_priv->refcount)==0);
214 if (atomic_dec_and_test(&cm_id_priv->refcount)) {
215 BUG_ON(!list_empty(&cm_id_priv->work_list));
216 free_cm_id(cm_id_priv);
223 static void add_ref(struct iw_cm_id *cm_id)
225 struct iwcm_id_private *cm_id_priv;
226 cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
227 atomic_inc(&cm_id_priv->refcount);
230 static void rem_ref(struct iw_cm_id *cm_id)
232 struct iwcm_id_private *cm_id_priv;
234 cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
236 (void)iwcm_deref_id(cm_id_priv);
239 static int cm_event_handler(struct iw_cm_id *cm_id, struct iw_cm_event *event);
241 struct iw_cm_id *iw_create_cm_id(struct ib_device *device,
242 iw_cm_handler cm_handler,
245 struct iwcm_id_private *cm_id_priv;
247 cm_id_priv = kzalloc(sizeof(*cm_id_priv), GFP_KERNEL);
249 return ERR_PTR(-ENOMEM);
251 cm_id_priv->state = IW_CM_STATE_IDLE;
252 cm_id_priv->id.device = device;
253 cm_id_priv->id.cm_handler = cm_handler;
254 cm_id_priv->id.context = context;
255 cm_id_priv->id.event_handler = cm_event_handler;
256 cm_id_priv->id.add_ref = add_ref;
257 cm_id_priv->id.rem_ref = rem_ref;
258 spin_lock_init(&cm_id_priv->lock);
259 atomic_set(&cm_id_priv->refcount, 1);
260 init_waitqueue_head(&cm_id_priv->connect_wait);
261 init_completion(&cm_id_priv->destroy_comp);
262 INIT_LIST_HEAD(&cm_id_priv->work_list);
263 INIT_LIST_HEAD(&cm_id_priv->work_free_list);
265 return &cm_id_priv->id;
267 EXPORT_SYMBOL(iw_create_cm_id);
270 static int iwcm_modify_qp_err(struct ib_qp *qp)
272 struct ib_qp_attr qp_attr;
277 qp_attr.qp_state = IB_QPS_ERR;
278 return ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
282 * This is really the RDMAC CLOSING state. It is most similar to the
285 static int iwcm_modify_qp_sqd(struct ib_qp *qp)
287 struct ib_qp_attr qp_attr;
290 qp_attr.qp_state = IB_QPS_SQD;
291 return ib_modify_qp(qp, &qp_attr, IB_QP_STATE);
297 * Block if a passive or active connection is currently being processed. Then
298 * process the event as follows:
299 * - If we are ESTABLISHED, move to CLOSING and modify the QP state
300 * based on the abrupt flag
301 * - If the connection is already in the CLOSING or IDLE state, the peer is
302 * disconnecting concurrently with us and we've already seen the
303 * DISCONNECT event -- ignore the request and return 0
304 * - Disconnect on a listening endpoint returns -EINVAL
306 int iw_cm_disconnect(struct iw_cm_id *cm_id, int abrupt)
308 struct iwcm_id_private *cm_id_priv;
311 struct ib_qp *qp = NULL;
313 cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
314 /* Wait if we're currently in a connect or accept downcall */
315 wait_event(cm_id_priv->connect_wait,
316 !test_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags));
318 spin_lock_irqsave(&cm_id_priv->lock, flags);
319 switch (cm_id_priv->state) {
320 case IW_CM_STATE_ESTABLISHED:
321 cm_id_priv->state = IW_CM_STATE_CLOSING;
323 /* QP could be <nul> for user-mode client */
329 case IW_CM_STATE_LISTEN:
332 case IW_CM_STATE_CLOSING:
333 /* remote peer closed first */
334 case IW_CM_STATE_IDLE:
335 /* accept or connect returned !0 */
337 case IW_CM_STATE_CONN_RECV:
339 * App called disconnect before/without calling accept after
340 * connect_request event delivered.
343 case IW_CM_STATE_CONN_SENT:
344 /* Can only get here if wait above fails */
348 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
352 ret = iwcm_modify_qp_err(qp);
354 ret = iwcm_modify_qp_sqd(qp);
357 * If both sides are disconnecting the QP could
358 * already be in ERR or SQD states
365 EXPORT_SYMBOL(iw_cm_disconnect);
368 * CM_ID <-- DESTROYING
370 * Clean up all resources associated with the connection and release
371 * the initial reference taken by iw_create_cm_id.
373 static void destroy_cm_id(struct iw_cm_id *cm_id)
375 struct iwcm_id_private *cm_id_priv;
378 cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
380 * Wait if we're currently in a connect or accept downcall. A
381 * listening endpoint should never block here.
383 wait_event(cm_id_priv->connect_wait,
384 !test_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags));
387 * Since we're deleting the cm_id, drop any events that
388 * might arrive before the last dereference.
390 set_bit(IWCM_F_DROP_EVENTS, &cm_id_priv->flags);
392 spin_lock_irqsave(&cm_id_priv->lock, flags);
393 switch (cm_id_priv->state) {
394 case IW_CM_STATE_LISTEN:
395 cm_id_priv->state = IW_CM_STATE_DESTROYING;
396 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
397 /* destroy the listening endpoint */
398 cm_id->device->iwcm->destroy_listen(cm_id);
399 spin_lock_irqsave(&cm_id_priv->lock, flags);
401 case IW_CM_STATE_ESTABLISHED:
402 cm_id_priv->state = IW_CM_STATE_DESTROYING;
403 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
404 /* Abrupt close of the connection */
405 (void)iwcm_modify_qp_err(cm_id_priv->qp);
406 spin_lock_irqsave(&cm_id_priv->lock, flags);
408 case IW_CM_STATE_IDLE:
409 case IW_CM_STATE_CLOSING:
410 cm_id_priv->state = IW_CM_STATE_DESTROYING;
412 case IW_CM_STATE_CONN_RECV:
414 * App called destroy before/without calling accept after
415 * receiving connection request event notification or
416 * returned non zero from the event callback function.
417 * In either case, must tell the provider to reject.
419 cm_id_priv->state = IW_CM_STATE_DESTROYING;
420 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
421 cm_id->device->iwcm->reject(cm_id, NULL, 0);
422 spin_lock_irqsave(&cm_id_priv->lock, flags);
424 case IW_CM_STATE_CONN_SENT:
425 case IW_CM_STATE_DESTROYING:
430 if (cm_id_priv->qp) {
431 cm_id_priv->id.device->iwcm->rem_ref(cm_id_priv->qp);
432 cm_id_priv->qp = NULL;
434 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
437 iwpm_remove_mapinfo(&cm_id->local_addr, &cm_id->m_local_addr);
438 iwpm_remove_mapping(&cm_id->local_addr, RDMA_NL_IWCM);
441 (void)iwcm_deref_id(cm_id_priv);
445 * This function is only called by the application thread and cannot
446 * be called by the event thread. The function will wait for all
447 * references to be released on the cm_id and then kfree the cm_id
450 void iw_destroy_cm_id(struct iw_cm_id *cm_id)
452 struct iwcm_id_private *cm_id_priv;
454 cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
455 destroy_cm_id(cm_id);
457 EXPORT_SYMBOL(iw_destroy_cm_id);
460 * iw_cm_check_wildcard - If IP address is 0 then use original
461 * @pm_addr: sockaddr containing the ip to check for wildcard
462 * @cm_addr: sockaddr containing the actual IP address
463 * @cm_outaddr: sockaddr to set IP addr which leaving port
465 * Checks the pm_addr for wildcard and then sets cm_outaddr's
466 * IP to the actual (cm_addr).
468 static void iw_cm_check_wildcard(struct sockaddr_storage *pm_addr,
469 struct sockaddr_storage *cm_addr,
470 struct sockaddr_storage *cm_outaddr)
472 if (pm_addr->ss_family == AF_INET) {
473 struct sockaddr_in *pm4_addr = (struct sockaddr_in *)pm_addr;
475 if (pm4_addr->sin_addr.s_addr == htonl(INADDR_ANY)) {
476 struct sockaddr_in *cm4_addr =
477 (struct sockaddr_in *)cm_addr;
478 struct sockaddr_in *cm4_outaddr =
479 (struct sockaddr_in *)cm_outaddr;
481 cm4_outaddr->sin_addr = cm4_addr->sin_addr;
484 struct sockaddr_in6 *pm6_addr = (struct sockaddr_in6 *)pm_addr;
486 if (ipv6_addr_type(&pm6_addr->sin6_addr) == IPV6_ADDR_ANY) {
487 struct sockaddr_in6 *cm6_addr =
488 (struct sockaddr_in6 *)cm_addr;
489 struct sockaddr_in6 *cm6_outaddr =
490 (struct sockaddr_in6 *)cm_outaddr;
492 cm6_outaddr->sin6_addr = cm6_addr->sin6_addr;
498 * iw_cm_map - Use portmapper to map the ports
499 * @cm_id: connection manager pointer
500 * @active: Indicates the active side when true
501 * returns nonzero for error only if iwpm_create_mapinfo() fails
503 * Tries to add a mapping for a port using the Portmapper. If
504 * successful in mapping the IP/Port it will check the remote
505 * mapped IP address for a wildcard IP address and replace the
506 * zero IP address with the remote_addr.
508 static int iw_cm_map(struct iw_cm_id *cm_id, bool active)
510 struct iwpm_dev_data pm_reg_msg;
511 struct iwpm_sa_data pm_msg;
514 cm_id->m_local_addr = cm_id->local_addr;
515 cm_id->m_remote_addr = cm_id->remote_addr;
517 memcpy(pm_reg_msg.dev_name, cm_id->device->name,
518 sizeof(pm_reg_msg.dev_name));
519 memcpy(pm_reg_msg.if_name, cm_id->device->iwcm->ifname,
520 sizeof(pm_reg_msg.if_name));
522 if (iwpm_register_pid(&pm_reg_msg, RDMA_NL_IWCM) ||
526 cm_id->mapped = true;
527 pm_msg.loc_addr = cm_id->local_addr;
528 pm_msg.rem_addr = cm_id->remote_addr;
530 status = iwpm_add_and_query_mapping(&pm_msg,
533 status = iwpm_add_mapping(&pm_msg, RDMA_NL_IWCM);
536 cm_id->m_local_addr = pm_msg.mapped_loc_addr;
538 cm_id->m_remote_addr = pm_msg.mapped_rem_addr;
539 iw_cm_check_wildcard(&pm_msg.mapped_rem_addr,
541 &cm_id->m_remote_addr);
545 return iwpm_create_mapinfo(&cm_id->local_addr,
546 &cm_id->m_local_addr,
553 * Start listening for connect requests. Generates one CONNECT_REQUEST
554 * event for each inbound connect request.
556 int iw_cm_listen(struct iw_cm_id *cm_id, int backlog)
558 struct iwcm_id_private *cm_id_priv;
562 cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
565 backlog = default_backlog;
567 ret = alloc_work_entries(cm_id_priv, backlog);
571 spin_lock_irqsave(&cm_id_priv->lock, flags);
572 switch (cm_id_priv->state) {
573 case IW_CM_STATE_IDLE:
574 cm_id_priv->state = IW_CM_STATE_LISTEN;
575 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
576 ret = iw_cm_map(cm_id, false);
578 ret = cm_id->device->iwcm->create_listen(cm_id, backlog);
580 cm_id_priv->state = IW_CM_STATE_IDLE;
581 spin_lock_irqsave(&cm_id_priv->lock, flags);
586 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
590 EXPORT_SYMBOL(iw_cm_listen);
595 * Rejects an inbound connection request. No events are generated.
597 int iw_cm_reject(struct iw_cm_id *cm_id,
598 const void *private_data,
601 struct iwcm_id_private *cm_id_priv;
605 cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
606 set_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
608 spin_lock_irqsave(&cm_id_priv->lock, flags);
609 if (cm_id_priv->state != IW_CM_STATE_CONN_RECV) {
610 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
611 clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
612 wake_up_all(&cm_id_priv->connect_wait);
615 cm_id_priv->state = IW_CM_STATE_IDLE;
616 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
618 ret = cm_id->device->iwcm->reject(cm_id, private_data,
621 clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
622 wake_up_all(&cm_id_priv->connect_wait);
626 EXPORT_SYMBOL(iw_cm_reject);
629 * CM_ID <-- ESTABLISHED
631 * Accepts an inbound connection request and generates an ESTABLISHED
632 * event. Callers of iw_cm_disconnect and iw_destroy_cm_id will block
633 * until the ESTABLISHED event is received from the provider.
635 int iw_cm_accept(struct iw_cm_id *cm_id,
636 struct iw_cm_conn_param *iw_param)
638 struct iwcm_id_private *cm_id_priv;
643 cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
644 set_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
646 spin_lock_irqsave(&cm_id_priv->lock, flags);
647 if (cm_id_priv->state != IW_CM_STATE_CONN_RECV) {
648 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
649 clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
650 wake_up_all(&cm_id_priv->connect_wait);
653 /* Get the ib_qp given the QPN */
654 qp = cm_id->device->iwcm->get_qp(cm_id->device, iw_param->qpn);
656 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
657 clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
658 wake_up_all(&cm_id_priv->connect_wait);
661 cm_id->device->iwcm->add_ref(qp);
663 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
665 ret = cm_id->device->iwcm->accept(cm_id, iw_param);
667 /* An error on accept precludes provider events */
668 BUG_ON(cm_id_priv->state != IW_CM_STATE_CONN_RECV);
669 cm_id_priv->state = IW_CM_STATE_IDLE;
670 spin_lock_irqsave(&cm_id_priv->lock, flags);
671 if (cm_id_priv->qp) {
672 cm_id->device->iwcm->rem_ref(qp);
673 cm_id_priv->qp = NULL;
675 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
676 clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
677 wake_up_all(&cm_id_priv->connect_wait);
682 EXPORT_SYMBOL(iw_cm_accept);
685 * Active Side: CM_ID <-- CONN_SENT
687 * If successful, results in the generation of a CONNECT_REPLY
688 * event. iw_cm_disconnect and iw_cm_destroy will block until the
689 * CONNECT_REPLY event is received from the provider.
691 int iw_cm_connect(struct iw_cm_id *cm_id, struct iw_cm_conn_param *iw_param)
693 struct iwcm_id_private *cm_id_priv;
698 cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
700 ret = alloc_work_entries(cm_id_priv, 4);
704 set_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
705 spin_lock_irqsave(&cm_id_priv->lock, flags);
707 if (cm_id_priv->state != IW_CM_STATE_IDLE) {
712 /* Get the ib_qp given the QPN */
713 qp = cm_id->device->iwcm->get_qp(cm_id->device, iw_param->qpn);
718 cm_id->device->iwcm->add_ref(qp);
720 cm_id_priv->state = IW_CM_STATE_CONN_SENT;
721 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
723 ret = iw_cm_map(cm_id, true);
725 ret = cm_id->device->iwcm->connect(cm_id, iw_param);
727 return 0; /* success */
729 spin_lock_irqsave(&cm_id_priv->lock, flags);
730 if (cm_id_priv->qp) {
731 cm_id->device->iwcm->rem_ref(qp);
732 cm_id_priv->qp = NULL;
734 cm_id_priv->state = IW_CM_STATE_IDLE;
736 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
737 clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
738 wake_up_all(&cm_id_priv->connect_wait);
741 EXPORT_SYMBOL(iw_cm_connect);
744 * Passive Side: new CM_ID <-- CONN_RECV
746 * Handles an inbound connect request. The function creates a new
747 * iw_cm_id to represent the new connection and inherits the client
748 * callback function and other attributes from the listening parent.
750 * The work item contains a pointer to the listen_cm_id and the event. The
751 * listen_cm_id contains the client cm_handler, context and
752 * device. These are copied when the device is cloned. The event
753 * contains the new four tuple.
755 * An error on the child should not affect the parent, so this
756 * function does not return a value.
758 static void cm_conn_req_handler(struct iwcm_id_private *listen_id_priv,
759 struct iw_cm_event *iw_event)
762 struct iw_cm_id *cm_id;
763 struct iwcm_id_private *cm_id_priv;
767 * The provider should never generate a connection request
768 * event with a bad status.
770 BUG_ON(iw_event->status);
772 cm_id = iw_create_cm_id(listen_id_priv->id.device,
773 listen_id_priv->id.cm_handler,
774 listen_id_priv->id.context);
775 /* If the cm_id could not be created, ignore the request */
779 cm_id->provider_data = iw_event->provider_data;
780 cm_id->m_local_addr = iw_event->local_addr;
781 cm_id->m_remote_addr = iw_event->remote_addr;
782 cm_id->local_addr = listen_id_priv->id.local_addr;
784 ret = iwpm_get_remote_info(&listen_id_priv->id.m_local_addr,
785 &iw_event->remote_addr,
789 cm_id->remote_addr = iw_event->remote_addr;
791 iw_cm_check_wildcard(&listen_id_priv->id.m_local_addr,
792 &iw_event->local_addr,
794 iw_event->local_addr = cm_id->local_addr;
795 iw_event->remote_addr = cm_id->remote_addr;
798 cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
799 cm_id_priv->state = IW_CM_STATE_CONN_RECV;
802 * We could be destroying the listening id. If so, ignore this
805 spin_lock_irqsave(&listen_id_priv->lock, flags);
806 if (listen_id_priv->state != IW_CM_STATE_LISTEN) {
807 spin_unlock_irqrestore(&listen_id_priv->lock, flags);
808 iw_cm_reject(cm_id, NULL, 0);
809 iw_destroy_cm_id(cm_id);
812 spin_unlock_irqrestore(&listen_id_priv->lock, flags);
814 ret = alloc_work_entries(cm_id_priv, 3);
816 iw_cm_reject(cm_id, NULL, 0);
817 iw_destroy_cm_id(cm_id);
821 /* Call the client CM handler */
822 ret = cm_id->cm_handler(cm_id, iw_event);
824 iw_cm_reject(cm_id, NULL, 0);
825 iw_destroy_cm_id(cm_id);
829 if (iw_event->private_data_len)
830 kfree(iw_event->private_data);
834 * Passive Side: CM_ID <-- ESTABLISHED
836 * The provider generated an ESTABLISHED event which means that
837 * the MPA negotion has completed successfully and we are now in MPA
840 * This event can only be received in the CONN_RECV state. If the
841 * remote peer closed, the ESTABLISHED event would be received followed
842 * by the CLOSE event. If the app closes, it will block until we wake
843 * it up after processing this event.
845 static int cm_conn_est_handler(struct iwcm_id_private *cm_id_priv,
846 struct iw_cm_event *iw_event)
851 spin_lock_irqsave(&cm_id_priv->lock, flags);
854 * We clear the CONNECT_WAIT bit here to allow the callback
855 * function to call iw_cm_disconnect. Calling iw_destroy_cm_id
856 * from a callback handler is not allowed.
858 clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
859 BUG_ON(cm_id_priv->state != IW_CM_STATE_CONN_RECV);
860 cm_id_priv->state = IW_CM_STATE_ESTABLISHED;
861 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
862 ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, iw_event);
863 wake_up_all(&cm_id_priv->connect_wait);
869 * Active Side: CM_ID <-- ESTABLISHED
871 * The app has called connect and is waiting for the established event to
872 * post it's requests to the server. This event will wake up anyone
873 * blocked in iw_cm_disconnect or iw_destroy_id.
875 static int cm_conn_rep_handler(struct iwcm_id_private *cm_id_priv,
876 struct iw_cm_event *iw_event)
881 spin_lock_irqsave(&cm_id_priv->lock, flags);
883 * Clear the connect wait bit so a callback function calling
884 * iw_cm_disconnect will not wait and deadlock this thread
886 clear_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags);
887 BUG_ON(cm_id_priv->state != IW_CM_STATE_CONN_SENT);
888 if (iw_event->status == 0) {
889 cm_id_priv->id.m_local_addr = iw_event->local_addr;
890 cm_id_priv->id.m_remote_addr = iw_event->remote_addr;
891 iw_event->local_addr = cm_id_priv->id.local_addr;
892 iw_event->remote_addr = cm_id_priv->id.remote_addr;
893 cm_id_priv->state = IW_CM_STATE_ESTABLISHED;
895 /* REJECTED or RESET */
896 cm_id_priv->id.device->iwcm->rem_ref(cm_id_priv->qp);
897 cm_id_priv->qp = NULL;
898 cm_id_priv->state = IW_CM_STATE_IDLE;
900 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
901 ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, iw_event);
903 if (iw_event->private_data_len)
904 kfree(iw_event->private_data);
906 /* Wake up waiters on connect complete */
907 wake_up_all(&cm_id_priv->connect_wait);
915 * If in the ESTABLISHED state, move to CLOSING.
917 static void cm_disconnect_handler(struct iwcm_id_private *cm_id_priv,
918 struct iw_cm_event *iw_event)
922 spin_lock_irqsave(&cm_id_priv->lock, flags);
923 if (cm_id_priv->state == IW_CM_STATE_ESTABLISHED)
924 cm_id_priv->state = IW_CM_STATE_CLOSING;
925 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
931 * If in the ESTBLISHED or CLOSING states, the QP will have have been
932 * moved by the provider to the ERR state. Disassociate the CM_ID from
933 * the QP, move to IDLE, and remove the 'connected' reference.
935 * If in some other state, the cm_id was destroyed asynchronously.
936 * This is the last reference that will result in waking up
937 * the app thread blocked in iw_destroy_cm_id.
939 static int cm_close_handler(struct iwcm_id_private *cm_id_priv,
940 struct iw_cm_event *iw_event)
944 spin_lock_irqsave(&cm_id_priv->lock, flags);
946 if (cm_id_priv->qp) {
947 cm_id_priv->id.device->iwcm->rem_ref(cm_id_priv->qp);
948 cm_id_priv->qp = NULL;
950 switch (cm_id_priv->state) {
951 case IW_CM_STATE_ESTABLISHED:
952 case IW_CM_STATE_CLOSING:
953 cm_id_priv->state = IW_CM_STATE_IDLE;
954 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
955 ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, iw_event);
956 spin_lock_irqsave(&cm_id_priv->lock, flags);
958 case IW_CM_STATE_DESTROYING:
963 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
968 static int process_event(struct iwcm_id_private *cm_id_priv,
969 struct iw_cm_event *iw_event)
973 switch (iw_event->event) {
974 case IW_CM_EVENT_CONNECT_REQUEST:
975 cm_conn_req_handler(cm_id_priv, iw_event);
977 case IW_CM_EVENT_CONNECT_REPLY:
978 ret = cm_conn_rep_handler(cm_id_priv, iw_event);
980 case IW_CM_EVENT_ESTABLISHED:
981 ret = cm_conn_est_handler(cm_id_priv, iw_event);
983 case IW_CM_EVENT_DISCONNECT:
984 cm_disconnect_handler(cm_id_priv, iw_event);
986 case IW_CM_EVENT_CLOSE:
987 ret = cm_close_handler(cm_id_priv, iw_event);
997 * Process events on the work_list for the cm_id. If the callback
998 * function requests that the cm_id be deleted, a flag is set in the
999 * cm_id flags to indicate that when the last reference is
1000 * removed, the cm_id is to be destroyed. This is necessary to
1001 * distinguish between an object that will be destroyed by the app
1002 * thread asleep on the destroy_comp list vs. an object destroyed
1003 * here synchronously when the last reference is removed.
1005 static void cm_work_handler(struct work_struct *_work)
1007 struct iwcm_work *work = container_of(_work, struct iwcm_work, work);
1008 struct iw_cm_event levent;
1009 struct iwcm_id_private *cm_id_priv = work->cm_id;
1010 unsigned long flags;
1014 spin_lock_irqsave(&cm_id_priv->lock, flags);
1015 empty = list_empty(&cm_id_priv->work_list);
1017 work = list_entry(cm_id_priv->work_list.next,
1018 struct iwcm_work, list);
1019 list_del_init(&work->list);
1020 empty = list_empty(&cm_id_priv->work_list);
1021 levent = work->event;
1023 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1025 if (!test_bit(IWCM_F_DROP_EVENTS, &cm_id_priv->flags)) {
1026 ret = process_event(cm_id_priv, &levent);
1028 destroy_cm_id(&cm_id_priv->id);
1030 pr_debug("dropping event %d\n", levent.event);
1031 if (iwcm_deref_id(cm_id_priv))
1035 spin_lock_irqsave(&cm_id_priv->lock, flags);
1037 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1041 * This function is called on interrupt context. Schedule events on
1042 * the iwcm_wq thread to allow callback functions to downcall into
1043 * the CM and/or block. Events are queued to a per-CM_ID
1044 * work_list. If this is the first event on the work_list, the work
1045 * element is also queued on the iwcm_wq thread.
1047 * Each event holds a reference on the cm_id. Until the last posted
1048 * event has been delivered and processed, the cm_id cannot be
1052 * 0 - the event was handled.
1053 * -ENOMEM - the event was not handled due to lack of resources.
1055 static int cm_event_handler(struct iw_cm_id *cm_id,
1056 struct iw_cm_event *iw_event)
1058 struct iwcm_work *work;
1059 struct iwcm_id_private *cm_id_priv;
1060 unsigned long flags;
1063 cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
1065 spin_lock_irqsave(&cm_id_priv->lock, flags);
1066 work = get_work(cm_id_priv);
1072 INIT_WORK(&work->work, cm_work_handler);
1073 work->cm_id = cm_id_priv;
1074 work->event = *iw_event;
1076 if ((work->event.event == IW_CM_EVENT_CONNECT_REQUEST ||
1077 work->event.event == IW_CM_EVENT_CONNECT_REPLY) &&
1078 work->event.private_data_len) {
1079 ret = copy_private_data(&work->event);
1086 atomic_inc(&cm_id_priv->refcount);
1087 if (list_empty(&cm_id_priv->work_list)) {
1088 list_add_tail(&work->list, &cm_id_priv->work_list);
1089 queue_work(iwcm_wq, &work->work);
1091 list_add_tail(&work->list, &cm_id_priv->work_list);
1093 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1097 static int iwcm_init_qp_init_attr(struct iwcm_id_private *cm_id_priv,
1098 struct ib_qp_attr *qp_attr,
1101 unsigned long flags;
1104 spin_lock_irqsave(&cm_id_priv->lock, flags);
1105 switch (cm_id_priv->state) {
1106 case IW_CM_STATE_IDLE:
1107 case IW_CM_STATE_CONN_SENT:
1108 case IW_CM_STATE_CONN_RECV:
1109 case IW_CM_STATE_ESTABLISHED:
1110 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
1111 qp_attr->qp_access_flags = IB_ACCESS_REMOTE_WRITE|
1112 IB_ACCESS_REMOTE_READ;
1119 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1123 static int iwcm_init_qp_rts_attr(struct iwcm_id_private *cm_id_priv,
1124 struct ib_qp_attr *qp_attr,
1127 unsigned long flags;
1130 spin_lock_irqsave(&cm_id_priv->lock, flags);
1131 switch (cm_id_priv->state) {
1132 case IW_CM_STATE_IDLE:
1133 case IW_CM_STATE_CONN_SENT:
1134 case IW_CM_STATE_CONN_RECV:
1135 case IW_CM_STATE_ESTABLISHED:
1143 spin_unlock_irqrestore(&cm_id_priv->lock, flags);
1147 int iw_cm_init_qp_attr(struct iw_cm_id *cm_id,
1148 struct ib_qp_attr *qp_attr,
1151 struct iwcm_id_private *cm_id_priv;
1154 cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
1155 switch (qp_attr->qp_state) {
1158 ret = iwcm_init_qp_init_attr(cm_id_priv,
1159 qp_attr, qp_attr_mask);
1162 ret = iwcm_init_qp_rts_attr(cm_id_priv,
1163 qp_attr, qp_attr_mask);
1171 EXPORT_SYMBOL(iw_cm_init_qp_attr);
1173 static int __init iw_cm_init(void)
1177 ret = iwpm_init(RDMA_NL_IWCM);
1181 iwcm_wq = alloc_ordered_workqueue("iw_cm_wq", 0);
1185 iwcm_ctl_table_hdr = register_net_sysctl(&init_net, "net/iw_cm",
1187 if (!iwcm_ctl_table_hdr) {
1188 pr_err("iw_cm: couldn't register sysctl paths\n");
1192 rdma_nl_register(RDMA_NL_IWCM, iwcm_nl_cb_table);
1196 destroy_workqueue(iwcm_wq);
1198 iwpm_exit(RDMA_NL_IWCM);
1202 static void __exit iw_cm_cleanup(void)
1204 rdma_nl_unregister(RDMA_NL_IWCM);
1205 unregister_net_sysctl_table(iwcm_ctl_table_hdr);
1206 destroy_workqueue(iwcm_wq);
1207 iwpm_exit(RDMA_NL_IWCM);
1210 MODULE_ALIAS_RDMA_NETLINK(RDMA_NL_IWCM, 2);
1212 module_init(iw_cm_init);
1213 module_exit(iw_cm_cleanup);