1/*
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.
8 *
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:
14 *
15 * Redistribution and use in source and binary forms, with or
16 * without modification, are permitted provided that the following
17 * conditions are met:
18 *
19 * - Redistributions of source code must retain the above
20 * copyright notice, this list of conditions and the following
21 * disclaimer.
22 *
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.
27 *
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
35 * SOFTWARE.
36 *
37 */
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>
50
51#include <rdma/iw_cm.h>
52#include <rdma/ib_addr.h>
53#include <rdma/iw_portmap.h>
54#include <rdma/rdma_netlink.h>
55
56#include "iwcm.h"
57
58MODULE_AUTHOR("Tom Tucker");
59MODULE_DESCRIPTION("iWARP CM");
60MODULE_LICENSE("Dual BSD/GPL");
61
62static const char * const iwcm_rej_reason_strs[] = {
63 [ECONNRESET] = "reset by remote host",
64 [ECONNREFUSED] = "refused by remote application",
65 [ETIMEDOUT] = "setup timeout",
66};
67
68const char *__attribute_const__ iwcm_reject_msg(int reason)
69{
70 size_t index;
71
72 /* iWARP uses negative errnos */
73 index = -reason;
74
75 if (index < ARRAY_SIZE(iwcm_rej_reason_strs) &&
76 iwcm_rej_reason_strs[index])
77 return iwcm_rej_reason_strs[index];
78 else
79 return "unrecognized reason";
80}
81EXPORT_SYMBOL(iwcm_reject_msg);
82
83static 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},
91 [RDMA_NL_IWPM_HELLO] = {.dump = iwpm_hello_cb}
92};
93
94static struct workqueue_struct *iwcm_wq;
95struct iwcm_work {
96 struct work_struct work;
97 struct iwcm_id_private *cm_id;
98 struct list_head list;
99 struct iw_cm_event event;
100 struct list_head free_list;
101};
102
103static unsigned int default_backlog = 256;
104
105static struct ctl_table_header *iwcm_ctl_table_hdr;
106static struct ctl_table iwcm_ctl_table[] = {
107 {
108 .procname = "default_backlog",
109 .data = &default_backlog,
110 .maxlen = sizeof(default_backlog),
111 .mode = 0644,
112 .proc_handler = proc_dointvec,
113 },
114};
115
116/*
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
120 * listen backlog.
121 * ACTIVE IDS: 4: CONNECT_REPLY, ESTABLISHED, DISCONNECT, CLOSE
122 * PASSIVE IDS: 3: ESTABLISHED, DISCONNECT, CLOSE
123 *
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).
127 *
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.
138 */
139
140static struct iwcm_work *get_work(struct iwcm_id_private *cm_id_priv)
141{
142 struct iwcm_work *work;
143
144 if (list_empty(head: &cm_id_priv->work_free_list))
145 return NULL;
146 work = list_entry(cm_id_priv->work_free_list.next, struct iwcm_work,
147 free_list);
148 list_del_init(entry: &work->free_list);
149 return work;
150}
151
152static void put_work(struct iwcm_work *work)
153{
154 list_add(new: &work->free_list, head: &work->cm_id->work_free_list);
155}
156
157static void dealloc_work_entries(struct iwcm_id_private *cm_id_priv)
158{
159 struct list_head *e, *tmp;
160
161 list_for_each_safe(e, tmp, &cm_id_priv->work_free_list) {
162 list_del(entry: e);
163 kfree(list_entry(e, struct iwcm_work, free_list));
164 }
165}
166
167static int alloc_work_entries(struct iwcm_id_private *cm_id_priv, int count)
168{
169 struct iwcm_work *work;
170
171 BUG_ON(!list_empty(&cm_id_priv->work_free_list));
172 while (count--) {
173 work = kmalloc(size: sizeof(struct iwcm_work), GFP_KERNEL);
174 if (!work) {
175 dealloc_work_entries(cm_id_priv);
176 return -ENOMEM;
177 }
178 work->cm_id = cm_id_priv;
179 INIT_LIST_HEAD(list: &work->list);
180 put_work(work);
181 }
182 return 0;
183}
184
185/*
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.
189 */
190static int copy_private_data(struct iw_cm_event *event)
191{
192 void *p;
193
194 p = kmemdup(p: event->private_data, size: event->private_data_len, GFP_ATOMIC);
195 if (!p)
196 return -ENOMEM;
197 event->private_data = p;
198 return 0;
199}
200
201static void free_cm_id(struct iwcm_id_private *cm_id_priv)
202{
203 dealloc_work_entries(cm_id_priv);
204 kfree(objp: cm_id_priv);
205}
206
207/*
208 * Release a reference on cm_id. If the last reference is being
209 * released, free the cm_id and return 1.
210 */
211static int iwcm_deref_id(struct iwcm_id_private *cm_id_priv)
212{
213 if (refcount_dec_and_test(r: &cm_id_priv->refcount)) {
214 BUG_ON(!list_empty(&cm_id_priv->work_list));
215 free_cm_id(cm_id_priv);
216 return 1;
217 }
218
219 return 0;
220}
221
222static void add_ref(struct iw_cm_id *cm_id)
223{
224 struct iwcm_id_private *cm_id_priv;
225 cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
226 refcount_inc(r: &cm_id_priv->refcount);
227}
228
229static void rem_ref(struct iw_cm_id *cm_id)
230{
231 struct iwcm_id_private *cm_id_priv;
232
233 cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
234
235 (void)iwcm_deref_id(cm_id_priv);
236}
237
238static int cm_event_handler(struct iw_cm_id *cm_id, struct iw_cm_event *event);
239
240struct iw_cm_id *iw_create_cm_id(struct ib_device *device,
241 iw_cm_handler cm_handler,
242 void *context)
243{
244 struct iwcm_id_private *cm_id_priv;
245
246 cm_id_priv = kzalloc(size: sizeof(*cm_id_priv), GFP_KERNEL);
247 if (!cm_id_priv)
248 return ERR_PTR(error: -ENOMEM);
249
250 cm_id_priv->state = IW_CM_STATE_IDLE;
251 cm_id_priv->id.device = device;
252 cm_id_priv->id.cm_handler = cm_handler;
253 cm_id_priv->id.context = context;
254 cm_id_priv->id.event_handler = cm_event_handler;
255 cm_id_priv->id.add_ref = add_ref;
256 cm_id_priv->id.rem_ref = rem_ref;
257 spin_lock_init(&cm_id_priv->lock);
258 refcount_set(r: &cm_id_priv->refcount, n: 1);
259 init_waitqueue_head(&cm_id_priv->connect_wait);
260 init_completion(x: &cm_id_priv->destroy_comp);
261 INIT_LIST_HEAD(list: &cm_id_priv->work_list);
262 INIT_LIST_HEAD(list: &cm_id_priv->work_free_list);
263
264 return &cm_id_priv->id;
265}
266EXPORT_SYMBOL(iw_create_cm_id);
267
268
269static int iwcm_modify_qp_err(struct ib_qp *qp)
270{
271 struct ib_qp_attr qp_attr;
272
273 if (!qp)
274 return -EINVAL;
275
276 qp_attr.qp_state = IB_QPS_ERR;
277 return ib_modify_qp(qp, qp_attr: &qp_attr, qp_attr_mask: IB_QP_STATE);
278}
279
280/*
281 * This is really the RDMAC CLOSING state. It is most similar to the
282 * IB SQD QP state.
283 */
284static int iwcm_modify_qp_sqd(struct ib_qp *qp)
285{
286 struct ib_qp_attr qp_attr;
287
288 BUG_ON(qp == NULL);
289 qp_attr.qp_state = IB_QPS_SQD;
290 return ib_modify_qp(qp, qp_attr: &qp_attr, qp_attr_mask: IB_QP_STATE);
291}
292
293/*
294 * CM_ID <-- CLOSING
295 *
296 * Block if a passive or active connection is currently being processed. Then
297 * process the event as follows:
298 * - If we are ESTABLISHED, move to CLOSING and modify the QP state
299 * based on the abrupt flag
300 * - If the connection is already in the CLOSING or IDLE state, the peer is
301 * disconnecting concurrently with us and we've already seen the
302 * DISCONNECT event -- ignore the request and return 0
303 * - Disconnect on a listening endpoint returns -EINVAL
304 */
305int iw_cm_disconnect(struct iw_cm_id *cm_id, int abrupt)
306{
307 struct iwcm_id_private *cm_id_priv;
308 unsigned long flags;
309 int ret = 0;
310 struct ib_qp *qp = NULL;
311
312 cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
313 /* Wait if we're currently in a connect or accept downcall */
314 wait_event(cm_id_priv->connect_wait,
315 !test_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags));
316
317 spin_lock_irqsave(&cm_id_priv->lock, flags);
318 switch (cm_id_priv->state) {
319 case IW_CM_STATE_ESTABLISHED:
320 cm_id_priv->state = IW_CM_STATE_CLOSING;
321
322 /* QP could be <nul> for user-mode client */
323 if (cm_id_priv->qp)
324 qp = cm_id_priv->qp;
325 else
326 ret = -EINVAL;
327 break;
328 case IW_CM_STATE_LISTEN:
329 ret = -EINVAL;
330 break;
331 case IW_CM_STATE_CLOSING:
332 /* remote peer closed first */
333 case IW_CM_STATE_IDLE:
334 /* accept or connect returned !0 */
335 break;
336 case IW_CM_STATE_CONN_RECV:
337 /*
338 * App called disconnect before/without calling accept after
339 * connect_request event delivered.
340 */
341 break;
342 case IW_CM_STATE_CONN_SENT:
343 /* Can only get here if wait above fails */
344 default:
345 BUG();
346 }
347 spin_unlock_irqrestore(lock: &cm_id_priv->lock, flags);
348
349 if (qp) {
350 if (abrupt)
351 ret = iwcm_modify_qp_err(qp);
352 else
353 ret = iwcm_modify_qp_sqd(qp);
354
355 /*
356 * If both sides are disconnecting the QP could
357 * already be in ERR or SQD states
358 */
359 ret = 0;
360 }
361
362 return ret;
363}
364EXPORT_SYMBOL(iw_cm_disconnect);
365
366/*
367 * CM_ID <-- DESTROYING
368 *
369 * Clean up all resources associated with the connection and release
370 * the initial reference taken by iw_create_cm_id.
371 */
372static void destroy_cm_id(struct iw_cm_id *cm_id)
373{
374 struct iwcm_id_private *cm_id_priv;
375 struct ib_qp *qp;
376 unsigned long flags;
377
378 cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
379 /*
380 * Wait if we're currently in a connect or accept downcall. A
381 * listening endpoint should never block here.
382 */
383 wait_event(cm_id_priv->connect_wait,
384 !test_bit(IWCM_F_CONNECT_WAIT, &cm_id_priv->flags));
385
386 /*
387 * Since we're deleting the cm_id, drop any events that
388 * might arrive before the last dereference.
389 */
390 set_bit(IWCM_F_DROP_EVENTS, addr: &cm_id_priv->flags);
391
392 spin_lock_irqsave(&cm_id_priv->lock, flags);
393 qp = cm_id_priv->qp;
394 cm_id_priv->qp = NULL;
395
396 switch (cm_id_priv->state) {
397 case IW_CM_STATE_LISTEN:
398 cm_id_priv->state = IW_CM_STATE_DESTROYING;
399 spin_unlock_irqrestore(lock: &cm_id_priv->lock, flags);
400 /* destroy the listening endpoint */
401 cm_id->device->ops.iw_destroy_listen(cm_id);
402 spin_lock_irqsave(&cm_id_priv->lock, flags);
403 break;
404 case IW_CM_STATE_ESTABLISHED:
405 cm_id_priv->state = IW_CM_STATE_DESTROYING;
406 spin_unlock_irqrestore(lock: &cm_id_priv->lock, flags);
407 /* Abrupt close of the connection */
408 (void)iwcm_modify_qp_err(qp);
409 spin_lock_irqsave(&cm_id_priv->lock, flags);
410 break;
411 case IW_CM_STATE_IDLE:
412 case IW_CM_STATE_CLOSING:
413 cm_id_priv->state = IW_CM_STATE_DESTROYING;
414 break;
415 case IW_CM_STATE_CONN_RECV:
416 /*
417 * App called destroy before/without calling accept after
418 * receiving connection request event notification or
419 * returned non zero from the event callback function.
420 * In either case, must tell the provider to reject.
421 */
422 cm_id_priv->state = IW_CM_STATE_DESTROYING;
423 spin_unlock_irqrestore(lock: &cm_id_priv->lock, flags);
424 cm_id->device->ops.iw_reject(cm_id, NULL, 0);
425 spin_lock_irqsave(&cm_id_priv->lock, flags);
426 break;
427 case IW_CM_STATE_CONN_SENT:
428 case IW_CM_STATE_DESTROYING:
429 default:
430 BUG();
431 break;
432 }
433 spin_unlock_irqrestore(lock: &cm_id_priv->lock, flags);
434 if (qp)
435 cm_id_priv->id.device->ops.iw_rem_ref(qp);
436
437 if (cm_id->mapped) {
438 iwpm_remove_mapinfo(local_addr: &cm_id->local_addr, mapped_addr: &cm_id->m_local_addr);
439 iwpm_remove_mapping(local_addr: &cm_id->local_addr, nl_client: RDMA_NL_IWCM);
440 }
441
442 (void)iwcm_deref_id(cm_id_priv);
443}
444
445/*
446 * This function is only called by the application thread and cannot
447 * be called by the event thread. The function will wait for all
448 * references to be released on the cm_id and then kfree the cm_id
449 * object.
450 */
451void iw_destroy_cm_id(struct iw_cm_id *cm_id)
452{
453 destroy_cm_id(cm_id);
454}
455EXPORT_SYMBOL(iw_destroy_cm_id);
456
457/**
458 * iw_cm_check_wildcard - If IP address is 0 then use original
459 * @pm_addr: sockaddr containing the ip to check for wildcard
460 * @cm_addr: sockaddr containing the actual IP address
461 * @cm_outaddr: sockaddr to set IP addr which leaving port
462 *
463 * Checks the pm_addr for wildcard and then sets cm_outaddr's
464 * IP to the actual (cm_addr).
465 */
466static void iw_cm_check_wildcard(struct sockaddr_storage *pm_addr,
467 struct sockaddr_storage *cm_addr,
468 struct sockaddr_storage *cm_outaddr)
469{
470 if (pm_addr->ss_family == AF_INET) {
471 struct sockaddr_in *pm4_addr = (struct sockaddr_in *)pm_addr;
472
473 if (pm4_addr->sin_addr.s_addr == htonl(INADDR_ANY)) {
474 struct sockaddr_in *cm4_addr =
475 (struct sockaddr_in *)cm_addr;
476 struct sockaddr_in *cm4_outaddr =
477 (struct sockaddr_in *)cm_outaddr;
478
479 cm4_outaddr->sin_addr = cm4_addr->sin_addr;
480 }
481 } else {
482 struct sockaddr_in6 *pm6_addr = (struct sockaddr_in6 *)pm_addr;
483
484 if (ipv6_addr_type(addr: &pm6_addr->sin6_addr) == IPV6_ADDR_ANY) {
485 struct sockaddr_in6 *cm6_addr =
486 (struct sockaddr_in6 *)cm_addr;
487 struct sockaddr_in6 *cm6_outaddr =
488 (struct sockaddr_in6 *)cm_outaddr;
489
490 cm6_outaddr->sin6_addr = cm6_addr->sin6_addr;
491 }
492 }
493}
494
495/**
496 * iw_cm_map - Use portmapper to map the ports
497 * @cm_id: connection manager pointer
498 * @active: Indicates the active side when true
499 * returns nonzero for error only if iwpm_create_mapinfo() fails
500 *
501 * Tries to add a mapping for a port using the Portmapper. If
502 * successful in mapping the IP/Port it will check the remote
503 * mapped IP address for a wildcard IP address and replace the
504 * zero IP address with the remote_addr.
505 */
506static int iw_cm_map(struct iw_cm_id *cm_id, bool active)
507{
508 const char *devname = dev_name(dev: &cm_id->device->dev);
509 const char *ifname = cm_id->device->iw_ifname;
510 struct iwpm_dev_data pm_reg_msg = {};
511 struct iwpm_sa_data pm_msg;
512 int status;
513
514 if (strlen(devname) >= sizeof(pm_reg_msg.dev_name) ||
515 strlen(ifname) >= sizeof(pm_reg_msg.if_name))
516 return -EINVAL;
517
518 cm_id->m_local_addr = cm_id->local_addr;
519 cm_id->m_remote_addr = cm_id->remote_addr;
520
521 strcpy(p: pm_reg_msg.dev_name, q: devname);
522 strcpy(p: pm_reg_msg.if_name, q: ifname);
523
524 if (iwpm_register_pid(pm_msg: &pm_reg_msg, nl_client: RDMA_NL_IWCM) ||
525 !iwpm_valid_pid())
526 return 0;
527
528 cm_id->mapped = true;
529 pm_msg.loc_addr = cm_id->local_addr;
530 pm_msg.rem_addr = cm_id->remote_addr;
531 pm_msg.flags = (cm_id->device->iw_driver_flags & IW_F_NO_PORT_MAP) ?
532 IWPM_FLAGS_NO_PORT_MAP : 0;
533 if (active)
534 status = iwpm_add_and_query_mapping(pm_msg: &pm_msg,
535 nl_client: RDMA_NL_IWCM);
536 else
537 status = iwpm_add_mapping(pm_msg: &pm_msg, nl_client: RDMA_NL_IWCM);
538
539 if (!status) {
540 cm_id->m_local_addr = pm_msg.mapped_loc_addr;
541 if (active) {
542 cm_id->m_remote_addr = pm_msg.mapped_rem_addr;
543 iw_cm_check_wildcard(pm_addr: &pm_msg.mapped_rem_addr,
544 cm_addr: &cm_id->remote_addr,
545 cm_outaddr: &cm_id->m_remote_addr);
546 }
547 }
548
549 return iwpm_create_mapinfo(local_addr: &cm_id->local_addr,
550 mapped_addr: &cm_id->m_local_addr,
551 nl_client: RDMA_NL_IWCM, map_flags: pm_msg.flags);
552}
553
554/*
555 * CM_ID <-- LISTEN
556 *
557 * Start listening for connect requests. Generates one CONNECT_REQUEST
558 * event for each inbound connect request.
559 */
560int iw_cm_listen(struct iw_cm_id *cm_id, int backlog)
561{
562 struct iwcm_id_private *cm_id_priv;
563 unsigned long flags;
564 int ret;
565
566 cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
567
568 if (!backlog)
569 backlog = default_backlog;
570
571 ret = alloc_work_entries(cm_id_priv, count: backlog);
572 if (ret)
573 return ret;
574
575 spin_lock_irqsave(&cm_id_priv->lock, flags);
576 switch (cm_id_priv->state) {
577 case IW_CM_STATE_IDLE:
578 cm_id_priv->state = IW_CM_STATE_LISTEN;
579 spin_unlock_irqrestore(lock: &cm_id_priv->lock, flags);
580 ret = iw_cm_map(cm_id, active: false);
581 if (!ret)
582 ret = cm_id->device->ops.iw_create_listen(cm_id,
583 backlog);
584 if (ret)
585 cm_id_priv->state = IW_CM_STATE_IDLE;
586 spin_lock_irqsave(&cm_id_priv->lock, flags);
587 break;
588 default:
589 ret = -EINVAL;
590 }
591 spin_unlock_irqrestore(lock: &cm_id_priv->lock, flags);
592
593 return ret;
594}
595EXPORT_SYMBOL(iw_cm_listen);
596
597/*
598 * CM_ID <-- IDLE
599 *
600 * Rejects an inbound connection request. No events are generated.
601 */
602int iw_cm_reject(struct iw_cm_id *cm_id,
603 const void *private_data,
604 u8 private_data_len)
605{
606 struct iwcm_id_private *cm_id_priv;
607 unsigned long flags;
608 int ret;
609
610 cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
611 set_bit(IWCM_F_CONNECT_WAIT, addr: &cm_id_priv->flags);
612
613 spin_lock_irqsave(&cm_id_priv->lock, flags);
614 if (cm_id_priv->state != IW_CM_STATE_CONN_RECV) {
615 spin_unlock_irqrestore(lock: &cm_id_priv->lock, flags);
616 clear_bit(IWCM_F_CONNECT_WAIT, addr: &cm_id_priv->flags);
617 wake_up_all(&cm_id_priv->connect_wait);
618 return -EINVAL;
619 }
620 cm_id_priv->state = IW_CM_STATE_IDLE;
621 spin_unlock_irqrestore(lock: &cm_id_priv->lock, flags);
622
623 ret = cm_id->device->ops.iw_reject(cm_id, private_data,
624 private_data_len);
625
626 clear_bit(IWCM_F_CONNECT_WAIT, addr: &cm_id_priv->flags);
627 wake_up_all(&cm_id_priv->connect_wait);
628
629 return ret;
630}
631EXPORT_SYMBOL(iw_cm_reject);
632
633/*
634 * CM_ID <-- ESTABLISHED
635 *
636 * Accepts an inbound connection request and generates an ESTABLISHED
637 * event. Callers of iw_cm_disconnect and iw_destroy_cm_id will block
638 * until the ESTABLISHED event is received from the provider.
639 */
640int iw_cm_accept(struct iw_cm_id *cm_id,
641 struct iw_cm_conn_param *iw_param)
642{
643 struct iwcm_id_private *cm_id_priv;
644 struct ib_qp *qp;
645 unsigned long flags;
646 int ret;
647
648 cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
649 set_bit(IWCM_F_CONNECT_WAIT, addr: &cm_id_priv->flags);
650
651 spin_lock_irqsave(&cm_id_priv->lock, flags);
652 if (cm_id_priv->state != IW_CM_STATE_CONN_RECV) {
653 spin_unlock_irqrestore(lock: &cm_id_priv->lock, flags);
654 clear_bit(IWCM_F_CONNECT_WAIT, addr: &cm_id_priv->flags);
655 wake_up_all(&cm_id_priv->connect_wait);
656 return -EINVAL;
657 }
658 /* Get the ib_qp given the QPN */
659 qp = cm_id->device->ops.iw_get_qp(cm_id->device, iw_param->qpn);
660 if (!qp) {
661 spin_unlock_irqrestore(lock: &cm_id_priv->lock, flags);
662 clear_bit(IWCM_F_CONNECT_WAIT, addr: &cm_id_priv->flags);
663 wake_up_all(&cm_id_priv->connect_wait);
664 return -EINVAL;
665 }
666 cm_id->device->ops.iw_add_ref(qp);
667 cm_id_priv->qp = qp;
668 spin_unlock_irqrestore(lock: &cm_id_priv->lock, flags);
669
670 ret = cm_id->device->ops.iw_accept(cm_id, iw_param);
671 if (ret) {
672 /* An error on accept precludes provider events */
673 BUG_ON(cm_id_priv->state != IW_CM_STATE_CONN_RECV);
674 cm_id_priv->state = IW_CM_STATE_IDLE;
675 spin_lock_irqsave(&cm_id_priv->lock, flags);
676 qp = cm_id_priv->qp;
677 cm_id_priv->qp = NULL;
678 spin_unlock_irqrestore(lock: &cm_id_priv->lock, flags);
679 if (qp)
680 cm_id->device->ops.iw_rem_ref(qp);
681 clear_bit(IWCM_F_CONNECT_WAIT, addr: &cm_id_priv->flags);
682 wake_up_all(&cm_id_priv->connect_wait);
683 }
684
685 return ret;
686}
687EXPORT_SYMBOL(iw_cm_accept);
688
689/*
690 * Active Side: CM_ID <-- CONN_SENT
691 *
692 * If successful, results in the generation of a CONNECT_REPLY
693 * event. iw_cm_disconnect and iw_cm_destroy will block until the
694 * CONNECT_REPLY event is received from the provider.
695 */
696int iw_cm_connect(struct iw_cm_id *cm_id, struct iw_cm_conn_param *iw_param)
697{
698 struct iwcm_id_private *cm_id_priv;
699 int ret;
700 unsigned long flags;
701 struct ib_qp *qp = NULL;
702
703 cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
704
705 ret = alloc_work_entries(cm_id_priv, count: 4);
706 if (ret)
707 return ret;
708
709 set_bit(IWCM_F_CONNECT_WAIT, addr: &cm_id_priv->flags);
710 spin_lock_irqsave(&cm_id_priv->lock, flags);
711
712 if (cm_id_priv->state != IW_CM_STATE_IDLE) {
713 ret = -EINVAL;
714 goto err;
715 }
716
717 /* Get the ib_qp given the QPN */
718 qp = cm_id->device->ops.iw_get_qp(cm_id->device, iw_param->qpn);
719 if (!qp) {
720 ret = -EINVAL;
721 goto err;
722 }
723 cm_id->device->ops.iw_add_ref(qp);
724 cm_id_priv->qp = qp;
725 cm_id_priv->state = IW_CM_STATE_CONN_SENT;
726 spin_unlock_irqrestore(lock: &cm_id_priv->lock, flags);
727
728 ret = iw_cm_map(cm_id, active: true);
729 if (!ret)
730 ret = cm_id->device->ops.iw_connect(cm_id, iw_param);
731 if (!ret)
732 return 0; /* success */
733
734 spin_lock_irqsave(&cm_id_priv->lock, flags);
735 qp = cm_id_priv->qp;
736 cm_id_priv->qp = NULL;
737 cm_id_priv->state = IW_CM_STATE_IDLE;
738err:
739 spin_unlock_irqrestore(lock: &cm_id_priv->lock, flags);
740 if (qp)
741 cm_id->device->ops.iw_rem_ref(qp);
742 clear_bit(IWCM_F_CONNECT_WAIT, addr: &cm_id_priv->flags);
743 wake_up_all(&cm_id_priv->connect_wait);
744 return ret;
745}
746EXPORT_SYMBOL(iw_cm_connect);
747
748/*
749 * Passive Side: new CM_ID <-- CONN_RECV
750 *
751 * Handles an inbound connect request. The function creates a new
752 * iw_cm_id to represent the new connection and inherits the client
753 * callback function and other attributes from the listening parent.
754 *
755 * The work item contains a pointer to the listen_cm_id and the event. The
756 * listen_cm_id contains the client cm_handler, context and
757 * device. These are copied when the device is cloned. The event
758 * contains the new four tuple.
759 *
760 * An error on the child should not affect the parent, so this
761 * function does not return a value.
762 */
763static void cm_conn_req_handler(struct iwcm_id_private *listen_id_priv,
764 struct iw_cm_event *iw_event)
765{
766 unsigned long flags;
767 struct iw_cm_id *cm_id;
768 struct iwcm_id_private *cm_id_priv;
769 int ret;
770
771 /*
772 * The provider should never generate a connection request
773 * event with a bad status.
774 */
775 BUG_ON(iw_event->status);
776
777 cm_id = iw_create_cm_id(listen_id_priv->id.device,
778 listen_id_priv->id.cm_handler,
779 listen_id_priv->id.context);
780 /* If the cm_id could not be created, ignore the request */
781 if (IS_ERR(ptr: cm_id))
782 goto out;
783
784 cm_id->provider_data = iw_event->provider_data;
785 cm_id->m_local_addr = iw_event->local_addr;
786 cm_id->m_remote_addr = iw_event->remote_addr;
787 cm_id->local_addr = listen_id_priv->id.local_addr;
788
789 ret = iwpm_get_remote_info(mapped_loc_addr: &listen_id_priv->id.m_local_addr,
790 mapped_rem_addr: &iw_event->remote_addr,
791 remote_addr: &cm_id->remote_addr,
792 nl_client: RDMA_NL_IWCM);
793 if (ret) {
794 cm_id->remote_addr = iw_event->remote_addr;
795 } else {
796 iw_cm_check_wildcard(pm_addr: &listen_id_priv->id.m_local_addr,
797 cm_addr: &iw_event->local_addr,
798 cm_outaddr: &cm_id->local_addr);
799 iw_event->local_addr = cm_id->local_addr;
800 iw_event->remote_addr = cm_id->remote_addr;
801 }
802
803 cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
804 cm_id_priv->state = IW_CM_STATE_CONN_RECV;
805
806 /*
807 * We could be destroying the listening id. If so, ignore this
808 * upcall.
809 */
810 spin_lock_irqsave(&listen_id_priv->lock, flags);
811 if (listen_id_priv->state != IW_CM_STATE_LISTEN) {
812 spin_unlock_irqrestore(lock: &listen_id_priv->lock, flags);
813 iw_cm_reject(cm_id, NULL, 0);
814 iw_destroy_cm_id(cm_id);
815 goto out;
816 }
817 spin_unlock_irqrestore(lock: &listen_id_priv->lock, flags);
818
819 ret = alloc_work_entries(cm_id_priv, count: 3);
820 if (ret) {
821 iw_cm_reject(cm_id, NULL, 0);
822 iw_destroy_cm_id(cm_id);
823 goto out;
824 }
825
826 /* Call the client CM handler */
827 ret = cm_id->cm_handler(cm_id, iw_event);
828 if (ret) {
829 iw_cm_reject(cm_id, NULL, 0);
830 iw_destroy_cm_id(cm_id);
831 }
832
833out:
834 if (iw_event->private_data_len)
835 kfree(objp: iw_event->private_data);
836}
837
838/*
839 * Passive Side: CM_ID <-- ESTABLISHED
840 *
841 * The provider generated an ESTABLISHED event which means that
842 * the MPA negotion has completed successfully and we are now in MPA
843 * FPDU mode.
844 *
845 * This event can only be received in the CONN_RECV state. If the
846 * remote peer closed, the ESTABLISHED event would be received followed
847 * by the CLOSE event. If the app closes, it will block until we wake
848 * it up after processing this event.
849 */
850static int cm_conn_est_handler(struct iwcm_id_private *cm_id_priv,
851 struct iw_cm_event *iw_event)
852{
853 unsigned long flags;
854 int ret;
855
856 spin_lock_irqsave(&cm_id_priv->lock, flags);
857
858 /*
859 * We clear the CONNECT_WAIT bit here to allow the callback
860 * function to call iw_cm_disconnect. Calling iw_destroy_cm_id
861 * from a callback handler is not allowed.
862 */
863 clear_bit(IWCM_F_CONNECT_WAIT, addr: &cm_id_priv->flags);
864 BUG_ON(cm_id_priv->state != IW_CM_STATE_CONN_RECV);
865 cm_id_priv->state = IW_CM_STATE_ESTABLISHED;
866 spin_unlock_irqrestore(lock: &cm_id_priv->lock, flags);
867 ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, iw_event);
868 wake_up_all(&cm_id_priv->connect_wait);
869
870 return ret;
871}
872
873/*
874 * Active Side: CM_ID <-- ESTABLISHED
875 *
876 * The app has called connect and is waiting for the established event to
877 * post it's requests to the server. This event will wake up anyone
878 * blocked in iw_cm_disconnect or iw_destroy_id.
879 */
880static int cm_conn_rep_handler(struct iwcm_id_private *cm_id_priv,
881 struct iw_cm_event *iw_event)
882{
883 struct ib_qp *qp = NULL;
884 unsigned long flags;
885 int ret;
886
887 spin_lock_irqsave(&cm_id_priv->lock, flags);
888 /*
889 * Clear the connect wait bit so a callback function calling
890 * iw_cm_disconnect will not wait and deadlock this thread
891 */
892 clear_bit(IWCM_F_CONNECT_WAIT, addr: &cm_id_priv->flags);
893 BUG_ON(cm_id_priv->state != IW_CM_STATE_CONN_SENT);
894 if (iw_event->status == 0) {
895 cm_id_priv->id.m_local_addr = iw_event->local_addr;
896 cm_id_priv->id.m_remote_addr = iw_event->remote_addr;
897 iw_event->local_addr = cm_id_priv->id.local_addr;
898 iw_event->remote_addr = cm_id_priv->id.remote_addr;
899 cm_id_priv->state = IW_CM_STATE_ESTABLISHED;
900 } else {
901 /* REJECTED or RESET */
902 qp = cm_id_priv->qp;
903 cm_id_priv->qp = NULL;
904 cm_id_priv->state = IW_CM_STATE_IDLE;
905 }
906 spin_unlock_irqrestore(lock: &cm_id_priv->lock, flags);
907 if (qp)
908 cm_id_priv->id.device->ops.iw_rem_ref(qp);
909 ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, iw_event);
910
911 if (iw_event->private_data_len)
912 kfree(objp: iw_event->private_data);
913
914 /* Wake up waiters on connect complete */
915 wake_up_all(&cm_id_priv->connect_wait);
916
917 return ret;
918}
919
920/*
921 * CM_ID <-- CLOSING
922 *
923 * If in the ESTABLISHED state, move to CLOSING.
924 */
925static void cm_disconnect_handler(struct iwcm_id_private *cm_id_priv,
926 struct iw_cm_event *iw_event)
927{
928 unsigned long flags;
929
930 spin_lock_irqsave(&cm_id_priv->lock, flags);
931 if (cm_id_priv->state == IW_CM_STATE_ESTABLISHED)
932 cm_id_priv->state = IW_CM_STATE_CLOSING;
933 spin_unlock_irqrestore(lock: &cm_id_priv->lock, flags);
934}
935
936/*
937 * CM_ID <-- IDLE
938 *
939 * If in the ESTBLISHED or CLOSING states, the QP will have have been
940 * moved by the provider to the ERR state. Disassociate the CM_ID from
941 * the QP, move to IDLE, and remove the 'connected' reference.
942 *
943 * If in some other state, the cm_id was destroyed asynchronously.
944 * This is the last reference that will result in waking up
945 * the app thread blocked in iw_destroy_cm_id.
946 */
947static int cm_close_handler(struct iwcm_id_private *cm_id_priv,
948 struct iw_cm_event *iw_event)
949{
950 struct ib_qp *qp;
951 unsigned long flags;
952 int ret = 0, notify_event = 0;
953 spin_lock_irqsave(&cm_id_priv->lock, flags);
954 qp = cm_id_priv->qp;
955 cm_id_priv->qp = NULL;
956
957 switch (cm_id_priv->state) {
958 case IW_CM_STATE_ESTABLISHED:
959 case IW_CM_STATE_CLOSING:
960 cm_id_priv->state = IW_CM_STATE_IDLE;
961 notify_event = 1;
962 break;
963 case IW_CM_STATE_DESTROYING:
964 break;
965 default:
966 BUG();
967 }
968 spin_unlock_irqrestore(lock: &cm_id_priv->lock, flags);
969
970 if (qp)
971 cm_id_priv->id.device->ops.iw_rem_ref(qp);
972 if (notify_event)
973 ret = cm_id_priv->id.cm_handler(&cm_id_priv->id, iw_event);
974 return ret;
975}
976
977static int process_event(struct iwcm_id_private *cm_id_priv,
978 struct iw_cm_event *iw_event)
979{
980 int ret = 0;
981
982 switch (iw_event->event) {
983 case IW_CM_EVENT_CONNECT_REQUEST:
984 cm_conn_req_handler(listen_id_priv: cm_id_priv, iw_event);
985 break;
986 case IW_CM_EVENT_CONNECT_REPLY:
987 ret = cm_conn_rep_handler(cm_id_priv, iw_event);
988 break;
989 case IW_CM_EVENT_ESTABLISHED:
990 ret = cm_conn_est_handler(cm_id_priv, iw_event);
991 break;
992 case IW_CM_EVENT_DISCONNECT:
993 cm_disconnect_handler(cm_id_priv, iw_event);
994 break;
995 case IW_CM_EVENT_CLOSE:
996 ret = cm_close_handler(cm_id_priv, iw_event);
997 break;
998 default:
999 BUG();
1000 }
1001
1002 return ret;
1003}
1004
1005/*
1006 * Process events on the work_list for the cm_id. If the callback
1007 * function requests that the cm_id be deleted, a flag is set in the
1008 * cm_id flags to indicate that when the last reference is
1009 * removed, the cm_id is to be destroyed. This is necessary to
1010 * distinguish between an object that will be destroyed by the app
1011 * thread asleep on the destroy_comp list vs. an object destroyed
1012 * here synchronously when the last reference is removed.
1013 */
1014static void cm_work_handler(struct work_struct *_work)
1015{
1016 struct iwcm_work *work = container_of(_work, struct iwcm_work, work);
1017 struct iw_cm_event levent;
1018 struct iwcm_id_private *cm_id_priv = work->cm_id;
1019 unsigned long flags;
1020 int empty;
1021 int ret = 0;
1022
1023 spin_lock_irqsave(&cm_id_priv->lock, flags);
1024 empty = list_empty(head: &cm_id_priv->work_list);
1025 while (!empty) {
1026 work = list_entry(cm_id_priv->work_list.next,
1027 struct iwcm_work, list);
1028 list_del_init(entry: &work->list);
1029 empty = list_empty(head: &cm_id_priv->work_list);
1030 levent = work->event;
1031 put_work(work);
1032 spin_unlock_irqrestore(lock: &cm_id_priv->lock, flags);
1033
1034 if (!test_bit(IWCM_F_DROP_EVENTS, &cm_id_priv->flags)) {
1035 ret = process_event(cm_id_priv, iw_event: &levent);
1036 if (ret)
1037 destroy_cm_id(cm_id: &cm_id_priv->id);
1038 } else
1039 pr_debug("dropping event %d\n", levent.event);
1040 if (iwcm_deref_id(cm_id_priv))
1041 return;
1042 if (empty)
1043 return;
1044 spin_lock_irqsave(&cm_id_priv->lock, flags);
1045 }
1046 spin_unlock_irqrestore(lock: &cm_id_priv->lock, flags);
1047}
1048
1049/*
1050 * This function is called on interrupt context. Schedule events on
1051 * the iwcm_wq thread to allow callback functions to downcall into
1052 * the CM and/or block. Events are queued to a per-CM_ID
1053 * work_list. If this is the first event on the work_list, the work
1054 * element is also queued on the iwcm_wq thread.
1055 *
1056 * Each event holds a reference on the cm_id. Until the last posted
1057 * event has been delivered and processed, the cm_id cannot be
1058 * deleted.
1059 *
1060 * Returns:
1061 * 0 - the event was handled.
1062 * -ENOMEM - the event was not handled due to lack of resources.
1063 */
1064static int cm_event_handler(struct iw_cm_id *cm_id,
1065 struct iw_cm_event *iw_event)
1066{
1067 struct iwcm_work *work;
1068 struct iwcm_id_private *cm_id_priv;
1069 unsigned long flags;
1070 int ret = 0;
1071
1072 cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
1073
1074 spin_lock_irqsave(&cm_id_priv->lock, flags);
1075 work = get_work(cm_id_priv);
1076 if (!work) {
1077 ret = -ENOMEM;
1078 goto out;
1079 }
1080
1081 INIT_WORK(&work->work, cm_work_handler);
1082 work->cm_id = cm_id_priv;
1083 work->event = *iw_event;
1084
1085 if ((work->event.event == IW_CM_EVENT_CONNECT_REQUEST ||
1086 work->event.event == IW_CM_EVENT_CONNECT_REPLY) &&
1087 work->event.private_data_len) {
1088 ret = copy_private_data(event: &work->event);
1089 if (ret) {
1090 put_work(work);
1091 goto out;
1092 }
1093 }
1094
1095 refcount_inc(r: &cm_id_priv->refcount);
1096 if (list_empty(head: &cm_id_priv->work_list)) {
1097 list_add_tail(new: &work->list, head: &cm_id_priv->work_list);
1098 queue_work(wq: iwcm_wq, work: &work->work);
1099 } else
1100 list_add_tail(new: &work->list, head: &cm_id_priv->work_list);
1101out:
1102 spin_unlock_irqrestore(lock: &cm_id_priv->lock, flags);
1103 return ret;
1104}
1105
1106static int iwcm_init_qp_init_attr(struct iwcm_id_private *cm_id_priv,
1107 struct ib_qp_attr *qp_attr,
1108 int *qp_attr_mask)
1109{
1110 unsigned long flags;
1111 int ret;
1112
1113 spin_lock_irqsave(&cm_id_priv->lock, flags);
1114 switch (cm_id_priv->state) {
1115 case IW_CM_STATE_IDLE:
1116 case IW_CM_STATE_CONN_SENT:
1117 case IW_CM_STATE_CONN_RECV:
1118 case IW_CM_STATE_ESTABLISHED:
1119 *qp_attr_mask = IB_QP_STATE | IB_QP_ACCESS_FLAGS;
1120 qp_attr->qp_access_flags = IB_ACCESS_REMOTE_WRITE|
1121 IB_ACCESS_REMOTE_READ;
1122 ret = 0;
1123 break;
1124 default:
1125 ret = -EINVAL;
1126 break;
1127 }
1128 spin_unlock_irqrestore(lock: &cm_id_priv->lock, flags);
1129 return ret;
1130}
1131
1132static int iwcm_init_qp_rts_attr(struct iwcm_id_private *cm_id_priv,
1133 struct ib_qp_attr *qp_attr,
1134 int *qp_attr_mask)
1135{
1136 unsigned long flags;
1137 int ret;
1138
1139 spin_lock_irqsave(&cm_id_priv->lock, flags);
1140 switch (cm_id_priv->state) {
1141 case IW_CM_STATE_IDLE:
1142 case IW_CM_STATE_CONN_SENT:
1143 case IW_CM_STATE_CONN_RECV:
1144 case IW_CM_STATE_ESTABLISHED:
1145 *qp_attr_mask = 0;
1146 ret = 0;
1147 break;
1148 default:
1149 ret = -EINVAL;
1150 break;
1151 }
1152 spin_unlock_irqrestore(lock: &cm_id_priv->lock, flags);
1153 return ret;
1154}
1155
1156int iw_cm_init_qp_attr(struct iw_cm_id *cm_id,
1157 struct ib_qp_attr *qp_attr,
1158 int *qp_attr_mask)
1159{
1160 struct iwcm_id_private *cm_id_priv;
1161 int ret;
1162
1163 cm_id_priv = container_of(cm_id, struct iwcm_id_private, id);
1164 switch (qp_attr->qp_state) {
1165 case IB_QPS_INIT:
1166 case IB_QPS_RTR:
1167 ret = iwcm_init_qp_init_attr(cm_id_priv,
1168 qp_attr, qp_attr_mask);
1169 break;
1170 case IB_QPS_RTS:
1171 ret = iwcm_init_qp_rts_attr(cm_id_priv,
1172 qp_attr, qp_attr_mask);
1173 break;
1174 default:
1175 ret = -EINVAL;
1176 break;
1177 }
1178 return ret;
1179}
1180EXPORT_SYMBOL(iw_cm_init_qp_attr);
1181
1182static int __init iw_cm_init(void)
1183{
1184 int ret;
1185
1186 ret = iwpm_init(RDMA_NL_IWCM);
1187 if (ret)
1188 return ret;
1189
1190 iwcm_wq = alloc_ordered_workqueue("iw_cm_wq", 0);
1191 if (!iwcm_wq)
1192 goto err_alloc;
1193
1194 iwcm_ctl_table_hdr = register_net_sysctl(&init_net, "net/iw_cm",
1195 iwcm_ctl_table);
1196 if (!iwcm_ctl_table_hdr) {
1197 pr_err("iw_cm: couldn't register sysctl paths\n");
1198 goto err_sysctl;
1199 }
1200
1201 rdma_nl_register(index: RDMA_NL_IWCM, cb_table: iwcm_nl_cb_table);
1202 return 0;
1203
1204err_sysctl:
1205 destroy_workqueue(wq: iwcm_wq);
1206err_alloc:
1207 iwpm_exit(RDMA_NL_IWCM);
1208 return -ENOMEM;
1209}
1210
1211static void __exit iw_cm_cleanup(void)
1212{
1213 rdma_nl_unregister(index: RDMA_NL_IWCM);
1214 unregister_net_sysctl_table(header: iwcm_ctl_table_hdr);
1215 destroy_workqueue(wq: iwcm_wq);
1216 iwpm_exit(RDMA_NL_IWCM);
1217}
1218
1219MODULE_ALIAS_RDMA_NETLINK(RDMA_NL_IWCM, 2);
1220
1221module_init(iw_cm_init);
1222module_exit(iw_cm_cleanup);
1223

source code of linux/drivers/infiniband/core/iwcm.c