xref: /linux/drivers/infiniband/core/ucma.c (revision d89dffa976bcd13fd87eb76e02e3b71c3a7868e3)
1 /*
2  * Copyright (c) 2005-2006 Intel Corporation.  All rights reserved.
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the
8  * OpenIB.org BSD license below:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *	copyright notice, this list of conditions and the following
16  *	disclaimer.
17  *
18  *      - Redistributions in binary form must reproduce the above
19  *	copyright notice, this list of conditions and the following
20  *	disclaimer in the documentation and/or other materials
21  *	provided with the distribution.
22  *
23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30  * SOFTWARE.
31  */
32 
33 #include <linux/completion.h>
34 #include <linux/file.h>
35 #include <linux/mutex.h>
36 #include <linux/poll.h>
37 #include <linux/sched.h>
38 #include <linux/idr.h>
39 #include <linux/in.h>
40 #include <linux/in6.h>
41 #include <linux/miscdevice.h>
42 #include <linux/slab.h>
43 #include <linux/sysctl.h>
44 #include <linux/module.h>
45 
46 #include <rdma/rdma_user_cm.h>
47 #include <rdma/ib_marshall.h>
48 #include <rdma/rdma_cm.h>
49 #include <rdma/rdma_cm_ib.h>
50 
51 MODULE_AUTHOR("Sean Hefty");
52 MODULE_DESCRIPTION("RDMA Userspace Connection Manager Access");
53 MODULE_LICENSE("Dual BSD/GPL");
54 
55 static unsigned int max_backlog = 1024;
56 
57 static struct ctl_table_header *ucma_ctl_table_hdr;
58 static ctl_table ucma_ctl_table[] = {
59 	{
60 		.procname	= "max_backlog",
61 		.data		= &max_backlog,
62 		.maxlen		= sizeof max_backlog,
63 		.mode		= 0644,
64 		.proc_handler	= proc_dointvec,
65 	},
66 	{ }
67 };
68 
69 struct ucma_file {
70 	struct mutex		mut;
71 	struct file		*filp;
72 	struct list_head	ctx_list;
73 	struct list_head	event_list;
74 	wait_queue_head_t	poll_wait;
75 };
76 
77 struct ucma_context {
78 	int			id;
79 	struct completion	comp;
80 	atomic_t		ref;
81 	int			events_reported;
82 	int			backlog;
83 
84 	struct ucma_file	*file;
85 	struct rdma_cm_id	*cm_id;
86 	u64			uid;
87 
88 	struct list_head	list;
89 	struct list_head	mc_list;
90 };
91 
92 struct ucma_multicast {
93 	struct ucma_context	*ctx;
94 	int			id;
95 	int			events_reported;
96 
97 	u64			uid;
98 	struct list_head	list;
99 	struct sockaddr_storage	addr;
100 };
101 
102 struct ucma_event {
103 	struct ucma_context	*ctx;
104 	struct ucma_multicast	*mc;
105 	struct list_head	list;
106 	struct rdma_cm_id	*cm_id;
107 	struct rdma_ucm_event_resp resp;
108 };
109 
110 static DEFINE_MUTEX(mut);
111 static DEFINE_IDR(ctx_idr);
112 static DEFINE_IDR(multicast_idr);
113 
114 static inline struct ucma_context *_ucma_find_context(int id,
115 						      struct ucma_file *file)
116 {
117 	struct ucma_context *ctx;
118 
119 	ctx = idr_find(&ctx_idr, id);
120 	if (!ctx)
121 		ctx = ERR_PTR(-ENOENT);
122 	else if (ctx->file != file)
123 		ctx = ERR_PTR(-EINVAL);
124 	return ctx;
125 }
126 
127 static struct ucma_context *ucma_get_ctx(struct ucma_file *file, int id)
128 {
129 	struct ucma_context *ctx;
130 
131 	mutex_lock(&mut);
132 	ctx = _ucma_find_context(id, file);
133 	if (!IS_ERR(ctx))
134 		atomic_inc(&ctx->ref);
135 	mutex_unlock(&mut);
136 	return ctx;
137 }
138 
139 static void ucma_put_ctx(struct ucma_context *ctx)
140 {
141 	if (atomic_dec_and_test(&ctx->ref))
142 		complete(&ctx->comp);
143 }
144 
145 static struct ucma_context *ucma_alloc_ctx(struct ucma_file *file)
146 {
147 	struct ucma_context *ctx;
148 	int ret;
149 
150 	ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
151 	if (!ctx)
152 		return NULL;
153 
154 	atomic_set(&ctx->ref, 1);
155 	init_completion(&ctx->comp);
156 	INIT_LIST_HEAD(&ctx->mc_list);
157 	ctx->file = file;
158 
159 	do {
160 		ret = idr_pre_get(&ctx_idr, GFP_KERNEL);
161 		if (!ret)
162 			goto error;
163 
164 		mutex_lock(&mut);
165 		ret = idr_get_new(&ctx_idr, ctx, &ctx->id);
166 		mutex_unlock(&mut);
167 	} while (ret == -EAGAIN);
168 
169 	if (ret)
170 		goto error;
171 
172 	list_add_tail(&ctx->list, &file->ctx_list);
173 	return ctx;
174 
175 error:
176 	kfree(ctx);
177 	return NULL;
178 }
179 
180 static struct ucma_multicast* ucma_alloc_multicast(struct ucma_context *ctx)
181 {
182 	struct ucma_multicast *mc;
183 	int ret;
184 
185 	mc = kzalloc(sizeof(*mc), GFP_KERNEL);
186 	if (!mc)
187 		return NULL;
188 
189 	do {
190 		ret = idr_pre_get(&multicast_idr, GFP_KERNEL);
191 		if (!ret)
192 			goto error;
193 
194 		mutex_lock(&mut);
195 		ret = idr_get_new(&multicast_idr, mc, &mc->id);
196 		mutex_unlock(&mut);
197 	} while (ret == -EAGAIN);
198 
199 	if (ret)
200 		goto error;
201 
202 	mc->ctx = ctx;
203 	list_add_tail(&mc->list, &ctx->mc_list);
204 	return mc;
205 
206 error:
207 	kfree(mc);
208 	return NULL;
209 }
210 
211 static void ucma_copy_conn_event(struct rdma_ucm_conn_param *dst,
212 				 struct rdma_conn_param *src)
213 {
214 	if (src->private_data_len)
215 		memcpy(dst->private_data, src->private_data,
216 		       src->private_data_len);
217 	dst->private_data_len = src->private_data_len;
218 	dst->responder_resources =src->responder_resources;
219 	dst->initiator_depth = src->initiator_depth;
220 	dst->flow_control = src->flow_control;
221 	dst->retry_count = src->retry_count;
222 	dst->rnr_retry_count = src->rnr_retry_count;
223 	dst->srq = src->srq;
224 	dst->qp_num = src->qp_num;
225 }
226 
227 static void ucma_copy_ud_event(struct rdma_ucm_ud_param *dst,
228 			       struct rdma_ud_param *src)
229 {
230 	if (src->private_data_len)
231 		memcpy(dst->private_data, src->private_data,
232 		       src->private_data_len);
233 	dst->private_data_len = src->private_data_len;
234 	ib_copy_ah_attr_to_user(&dst->ah_attr, &src->ah_attr);
235 	dst->qp_num = src->qp_num;
236 	dst->qkey = src->qkey;
237 }
238 
239 static void ucma_set_event_context(struct ucma_context *ctx,
240 				   struct rdma_cm_event *event,
241 				   struct ucma_event *uevent)
242 {
243 	uevent->ctx = ctx;
244 	switch (event->event) {
245 	case RDMA_CM_EVENT_MULTICAST_JOIN:
246 	case RDMA_CM_EVENT_MULTICAST_ERROR:
247 		uevent->mc = (struct ucma_multicast *)
248 			     event->param.ud.private_data;
249 		uevent->resp.uid = uevent->mc->uid;
250 		uevent->resp.id = uevent->mc->id;
251 		break;
252 	default:
253 		uevent->resp.uid = ctx->uid;
254 		uevent->resp.id = ctx->id;
255 		break;
256 	}
257 }
258 
259 static int ucma_event_handler(struct rdma_cm_id *cm_id,
260 			      struct rdma_cm_event *event)
261 {
262 	struct ucma_event *uevent;
263 	struct ucma_context *ctx = cm_id->context;
264 	int ret = 0;
265 
266 	uevent = kzalloc(sizeof(*uevent), GFP_KERNEL);
267 	if (!uevent)
268 		return event->event == RDMA_CM_EVENT_CONNECT_REQUEST;
269 
270 	uevent->cm_id = cm_id;
271 	ucma_set_event_context(ctx, event, uevent);
272 	uevent->resp.event = event->event;
273 	uevent->resp.status = event->status;
274 	if (cm_id->qp_type == IB_QPT_UD)
275 		ucma_copy_ud_event(&uevent->resp.param.ud, &event->param.ud);
276 	else
277 		ucma_copy_conn_event(&uevent->resp.param.conn,
278 				     &event->param.conn);
279 
280 	mutex_lock(&ctx->file->mut);
281 	if (event->event == RDMA_CM_EVENT_CONNECT_REQUEST) {
282 		if (!ctx->backlog) {
283 			ret = -ENOMEM;
284 			kfree(uevent);
285 			goto out;
286 		}
287 		ctx->backlog--;
288 	} else if (!ctx->uid) {
289 		/*
290 		 * We ignore events for new connections until userspace has set
291 		 * their context.  This can only happen if an error occurs on a
292 		 * new connection before the user accepts it.  This is okay,
293 		 * since the accept will just fail later.
294 		 */
295 		kfree(uevent);
296 		goto out;
297 	}
298 
299 	list_add_tail(&uevent->list, &ctx->file->event_list);
300 	wake_up_interruptible(&ctx->file->poll_wait);
301 out:
302 	mutex_unlock(&ctx->file->mut);
303 	return ret;
304 }
305 
306 static ssize_t ucma_get_event(struct ucma_file *file, const char __user *inbuf,
307 			      int in_len, int out_len)
308 {
309 	struct ucma_context *ctx;
310 	struct rdma_ucm_get_event cmd;
311 	struct ucma_event *uevent;
312 	int ret = 0;
313 	DEFINE_WAIT(wait);
314 
315 	if (out_len < sizeof uevent->resp)
316 		return -ENOSPC;
317 
318 	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
319 		return -EFAULT;
320 
321 	mutex_lock(&file->mut);
322 	while (list_empty(&file->event_list)) {
323 		mutex_unlock(&file->mut);
324 
325 		if (file->filp->f_flags & O_NONBLOCK)
326 			return -EAGAIN;
327 
328 		if (wait_event_interruptible(file->poll_wait,
329 					     !list_empty(&file->event_list)))
330 			return -ERESTARTSYS;
331 
332 		mutex_lock(&file->mut);
333 	}
334 
335 	uevent = list_entry(file->event_list.next, struct ucma_event, list);
336 
337 	if (uevent->resp.event == RDMA_CM_EVENT_CONNECT_REQUEST) {
338 		ctx = ucma_alloc_ctx(file);
339 		if (!ctx) {
340 			ret = -ENOMEM;
341 			goto done;
342 		}
343 		uevent->ctx->backlog++;
344 		ctx->cm_id = uevent->cm_id;
345 		ctx->cm_id->context = ctx;
346 		uevent->resp.id = ctx->id;
347 	}
348 
349 	if (copy_to_user((void __user *)(unsigned long)cmd.response,
350 			 &uevent->resp, sizeof uevent->resp)) {
351 		ret = -EFAULT;
352 		goto done;
353 	}
354 
355 	list_del(&uevent->list);
356 	uevent->ctx->events_reported++;
357 	if (uevent->mc)
358 		uevent->mc->events_reported++;
359 	kfree(uevent);
360 done:
361 	mutex_unlock(&file->mut);
362 	return ret;
363 }
364 
365 static int ucma_get_qp_type(struct rdma_ucm_create_id *cmd, enum ib_qp_type *qp_type)
366 {
367 	switch (cmd->ps) {
368 	case RDMA_PS_TCP:
369 		*qp_type = IB_QPT_RC;
370 		return 0;
371 	case RDMA_PS_UDP:
372 	case RDMA_PS_IPOIB:
373 		*qp_type = IB_QPT_UD;
374 		return 0;
375 	case RDMA_PS_IB:
376 		*qp_type = cmd->qp_type;
377 		return 0;
378 	default:
379 		return -EINVAL;
380 	}
381 }
382 
383 static ssize_t ucma_create_id(struct ucma_file *file, const char __user *inbuf,
384 			      int in_len, int out_len)
385 {
386 	struct rdma_ucm_create_id cmd;
387 	struct rdma_ucm_create_id_resp resp;
388 	struct ucma_context *ctx;
389 	enum ib_qp_type qp_type;
390 	int ret;
391 
392 	if (out_len < sizeof(resp))
393 		return -ENOSPC;
394 
395 	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
396 		return -EFAULT;
397 
398 	ret = ucma_get_qp_type(&cmd, &qp_type);
399 	if (ret)
400 		return ret;
401 
402 	mutex_lock(&file->mut);
403 	ctx = ucma_alloc_ctx(file);
404 	mutex_unlock(&file->mut);
405 	if (!ctx)
406 		return -ENOMEM;
407 
408 	ctx->uid = cmd.uid;
409 	ctx->cm_id = rdma_create_id(ucma_event_handler, ctx, cmd.ps, qp_type);
410 	if (IS_ERR(ctx->cm_id)) {
411 		ret = PTR_ERR(ctx->cm_id);
412 		goto err1;
413 	}
414 
415 	resp.id = ctx->id;
416 	if (copy_to_user((void __user *)(unsigned long)cmd.response,
417 			 &resp, sizeof(resp))) {
418 		ret = -EFAULT;
419 		goto err2;
420 	}
421 	return 0;
422 
423 err2:
424 	rdma_destroy_id(ctx->cm_id);
425 err1:
426 	mutex_lock(&mut);
427 	idr_remove(&ctx_idr, ctx->id);
428 	mutex_unlock(&mut);
429 	kfree(ctx);
430 	return ret;
431 }
432 
433 static void ucma_cleanup_multicast(struct ucma_context *ctx)
434 {
435 	struct ucma_multicast *mc, *tmp;
436 
437 	mutex_lock(&mut);
438 	list_for_each_entry_safe(mc, tmp, &ctx->mc_list, list) {
439 		list_del(&mc->list);
440 		idr_remove(&multicast_idr, mc->id);
441 		kfree(mc);
442 	}
443 	mutex_unlock(&mut);
444 }
445 
446 static void ucma_cleanup_mc_events(struct ucma_multicast *mc)
447 {
448 	struct ucma_event *uevent, *tmp;
449 
450 	list_for_each_entry_safe(uevent, tmp, &mc->ctx->file->event_list, list) {
451 		if (uevent->mc != mc)
452 			continue;
453 
454 		list_del(&uevent->list);
455 		kfree(uevent);
456 	}
457 }
458 
459 /*
460  * We cannot hold file->mut when calling rdma_destroy_id() or we can
461  * deadlock.  We also acquire file->mut in ucma_event_handler(), and
462  * rdma_destroy_id() will wait until all callbacks have completed.
463  */
464 static int ucma_free_ctx(struct ucma_context *ctx)
465 {
466 	int events_reported;
467 	struct ucma_event *uevent, *tmp;
468 	LIST_HEAD(list);
469 
470 	/* No new events will be generated after destroying the id. */
471 	rdma_destroy_id(ctx->cm_id);
472 
473 	ucma_cleanup_multicast(ctx);
474 
475 	/* Cleanup events not yet reported to the user. */
476 	mutex_lock(&ctx->file->mut);
477 	list_for_each_entry_safe(uevent, tmp, &ctx->file->event_list, list) {
478 		if (uevent->ctx == ctx)
479 			list_move_tail(&uevent->list, &list);
480 	}
481 	list_del(&ctx->list);
482 	mutex_unlock(&ctx->file->mut);
483 
484 	list_for_each_entry_safe(uevent, tmp, &list, list) {
485 		list_del(&uevent->list);
486 		if (uevent->resp.event == RDMA_CM_EVENT_CONNECT_REQUEST)
487 			rdma_destroy_id(uevent->cm_id);
488 		kfree(uevent);
489 	}
490 
491 	events_reported = ctx->events_reported;
492 	kfree(ctx);
493 	return events_reported;
494 }
495 
496 static ssize_t ucma_destroy_id(struct ucma_file *file, const char __user *inbuf,
497 			       int in_len, int out_len)
498 {
499 	struct rdma_ucm_destroy_id cmd;
500 	struct rdma_ucm_destroy_id_resp resp;
501 	struct ucma_context *ctx;
502 	int ret = 0;
503 
504 	if (out_len < sizeof(resp))
505 		return -ENOSPC;
506 
507 	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
508 		return -EFAULT;
509 
510 	mutex_lock(&mut);
511 	ctx = _ucma_find_context(cmd.id, file);
512 	if (!IS_ERR(ctx))
513 		idr_remove(&ctx_idr, ctx->id);
514 	mutex_unlock(&mut);
515 
516 	if (IS_ERR(ctx))
517 		return PTR_ERR(ctx);
518 
519 	ucma_put_ctx(ctx);
520 	wait_for_completion(&ctx->comp);
521 	resp.events_reported = ucma_free_ctx(ctx);
522 
523 	if (copy_to_user((void __user *)(unsigned long)cmd.response,
524 			 &resp, sizeof(resp)))
525 		ret = -EFAULT;
526 
527 	return ret;
528 }
529 
530 static ssize_t ucma_bind_addr(struct ucma_file *file, const char __user *inbuf,
531 			      int in_len, int out_len)
532 {
533 	struct rdma_ucm_bind_addr cmd;
534 	struct ucma_context *ctx;
535 	int ret;
536 
537 	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
538 		return -EFAULT;
539 
540 	ctx = ucma_get_ctx(file, cmd.id);
541 	if (IS_ERR(ctx))
542 		return PTR_ERR(ctx);
543 
544 	ret = rdma_bind_addr(ctx->cm_id, (struct sockaddr *) &cmd.addr);
545 	ucma_put_ctx(ctx);
546 	return ret;
547 }
548 
549 static ssize_t ucma_resolve_addr(struct ucma_file *file,
550 				 const char __user *inbuf,
551 				 int in_len, int out_len)
552 {
553 	struct rdma_ucm_resolve_addr cmd;
554 	struct ucma_context *ctx;
555 	int ret;
556 
557 	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
558 		return -EFAULT;
559 
560 	ctx = ucma_get_ctx(file, cmd.id);
561 	if (IS_ERR(ctx))
562 		return PTR_ERR(ctx);
563 
564 	ret = rdma_resolve_addr(ctx->cm_id, (struct sockaddr *) &cmd.src_addr,
565 				(struct sockaddr *) &cmd.dst_addr,
566 				cmd.timeout_ms);
567 	ucma_put_ctx(ctx);
568 	return ret;
569 }
570 
571 static ssize_t ucma_resolve_route(struct ucma_file *file,
572 				  const char __user *inbuf,
573 				  int in_len, int out_len)
574 {
575 	struct rdma_ucm_resolve_route cmd;
576 	struct ucma_context *ctx;
577 	int ret;
578 
579 	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
580 		return -EFAULT;
581 
582 	ctx = ucma_get_ctx(file, cmd.id);
583 	if (IS_ERR(ctx))
584 		return PTR_ERR(ctx);
585 
586 	ret = rdma_resolve_route(ctx->cm_id, cmd.timeout_ms);
587 	ucma_put_ctx(ctx);
588 	return ret;
589 }
590 
591 static void ucma_copy_ib_route(struct rdma_ucm_query_route_resp *resp,
592 			       struct rdma_route *route)
593 {
594 	struct rdma_dev_addr *dev_addr;
595 
596 	resp->num_paths = route->num_paths;
597 	switch (route->num_paths) {
598 	case 0:
599 		dev_addr = &route->addr.dev_addr;
600 		rdma_addr_get_dgid(dev_addr,
601 				   (union ib_gid *) &resp->ib_route[0].dgid);
602 		rdma_addr_get_sgid(dev_addr,
603 				   (union ib_gid *) &resp->ib_route[0].sgid);
604 		resp->ib_route[0].pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
605 		break;
606 	case 2:
607 		ib_copy_path_rec_to_user(&resp->ib_route[1],
608 					 &route->path_rec[1]);
609 		/* fall through */
610 	case 1:
611 		ib_copy_path_rec_to_user(&resp->ib_route[0],
612 					 &route->path_rec[0]);
613 		break;
614 	default:
615 		break;
616 	}
617 }
618 
619 static void ucma_copy_iboe_route(struct rdma_ucm_query_route_resp *resp,
620 				 struct rdma_route *route)
621 {
622 	struct rdma_dev_addr *dev_addr;
623 	struct net_device *dev;
624 	u16 vid = 0;
625 
626 	resp->num_paths = route->num_paths;
627 	switch (route->num_paths) {
628 	case 0:
629 		dev_addr = &route->addr.dev_addr;
630 		dev = dev_get_by_index(&init_net, dev_addr->bound_dev_if);
631 			if (dev) {
632 				vid = rdma_vlan_dev_vlan_id(dev);
633 				dev_put(dev);
634 			}
635 
636 		iboe_mac_vlan_to_ll((union ib_gid *) &resp->ib_route[0].dgid,
637 				    dev_addr->dst_dev_addr, vid);
638 		iboe_addr_get_sgid(dev_addr,
639 				   (union ib_gid *) &resp->ib_route[0].sgid);
640 		resp->ib_route[0].pkey = cpu_to_be16(0xffff);
641 		break;
642 	case 2:
643 		ib_copy_path_rec_to_user(&resp->ib_route[1],
644 					 &route->path_rec[1]);
645 		/* fall through */
646 	case 1:
647 		ib_copy_path_rec_to_user(&resp->ib_route[0],
648 					 &route->path_rec[0]);
649 		break;
650 	default:
651 		break;
652 	}
653 }
654 
655 static void ucma_copy_iw_route(struct rdma_ucm_query_route_resp *resp,
656 			       struct rdma_route *route)
657 {
658 	struct rdma_dev_addr *dev_addr;
659 
660 	dev_addr = &route->addr.dev_addr;
661 	rdma_addr_get_dgid(dev_addr, (union ib_gid *) &resp->ib_route[0].dgid);
662 	rdma_addr_get_sgid(dev_addr, (union ib_gid *) &resp->ib_route[0].sgid);
663 }
664 
665 static ssize_t ucma_query_route(struct ucma_file *file,
666 				const char __user *inbuf,
667 				int in_len, int out_len)
668 {
669 	struct rdma_ucm_query_route cmd;
670 	struct rdma_ucm_query_route_resp resp;
671 	struct ucma_context *ctx;
672 	struct sockaddr *addr;
673 	int ret = 0;
674 
675 	if (out_len < sizeof(resp))
676 		return -ENOSPC;
677 
678 	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
679 		return -EFAULT;
680 
681 	ctx = ucma_get_ctx(file, cmd.id);
682 	if (IS_ERR(ctx))
683 		return PTR_ERR(ctx);
684 
685 	memset(&resp, 0, sizeof resp);
686 	addr = (struct sockaddr *) &ctx->cm_id->route.addr.src_addr;
687 	memcpy(&resp.src_addr, addr, addr->sa_family == AF_INET ?
688 				     sizeof(struct sockaddr_in) :
689 				     sizeof(struct sockaddr_in6));
690 	addr = (struct sockaddr *) &ctx->cm_id->route.addr.dst_addr;
691 	memcpy(&resp.dst_addr, addr, addr->sa_family == AF_INET ?
692 				     sizeof(struct sockaddr_in) :
693 				     sizeof(struct sockaddr_in6));
694 	if (!ctx->cm_id->device)
695 		goto out;
696 
697 	resp.node_guid = (__force __u64) ctx->cm_id->device->node_guid;
698 	resp.port_num = ctx->cm_id->port_num;
699 	switch (rdma_node_get_transport(ctx->cm_id->device->node_type)) {
700 	case RDMA_TRANSPORT_IB:
701 		switch (rdma_port_get_link_layer(ctx->cm_id->device,
702 			ctx->cm_id->port_num)) {
703 		case IB_LINK_LAYER_INFINIBAND:
704 			ucma_copy_ib_route(&resp, &ctx->cm_id->route);
705 			break;
706 		case IB_LINK_LAYER_ETHERNET:
707 			ucma_copy_iboe_route(&resp, &ctx->cm_id->route);
708 			break;
709 		default:
710 			break;
711 		}
712 		break;
713 	case RDMA_TRANSPORT_IWARP:
714 		ucma_copy_iw_route(&resp, &ctx->cm_id->route);
715 		break;
716 	default:
717 		break;
718 	}
719 
720 out:
721 	if (copy_to_user((void __user *)(unsigned long)cmd.response,
722 			 &resp, sizeof(resp)))
723 		ret = -EFAULT;
724 
725 	ucma_put_ctx(ctx);
726 	return ret;
727 }
728 
729 static void ucma_copy_conn_param(struct rdma_conn_param *dst,
730 				 struct rdma_ucm_conn_param *src)
731 {
732 	dst->private_data = src->private_data;
733 	dst->private_data_len = src->private_data_len;
734 	dst->responder_resources =src->responder_resources;
735 	dst->initiator_depth = src->initiator_depth;
736 	dst->flow_control = src->flow_control;
737 	dst->retry_count = src->retry_count;
738 	dst->rnr_retry_count = src->rnr_retry_count;
739 	dst->srq = src->srq;
740 	dst->qp_num = src->qp_num;
741 }
742 
743 static ssize_t ucma_connect(struct ucma_file *file, const char __user *inbuf,
744 			    int in_len, int out_len)
745 {
746 	struct rdma_ucm_connect cmd;
747 	struct rdma_conn_param conn_param;
748 	struct ucma_context *ctx;
749 	int ret;
750 
751 	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
752 		return -EFAULT;
753 
754 	if (!cmd.conn_param.valid)
755 		return -EINVAL;
756 
757 	ctx = ucma_get_ctx(file, cmd.id);
758 	if (IS_ERR(ctx))
759 		return PTR_ERR(ctx);
760 
761 	ucma_copy_conn_param(&conn_param, &cmd.conn_param);
762 	ret = rdma_connect(ctx->cm_id, &conn_param);
763 	ucma_put_ctx(ctx);
764 	return ret;
765 }
766 
767 static ssize_t ucma_listen(struct ucma_file *file, const char __user *inbuf,
768 			   int in_len, int out_len)
769 {
770 	struct rdma_ucm_listen cmd;
771 	struct ucma_context *ctx;
772 	int ret;
773 
774 	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
775 		return -EFAULT;
776 
777 	ctx = ucma_get_ctx(file, cmd.id);
778 	if (IS_ERR(ctx))
779 		return PTR_ERR(ctx);
780 
781 	ctx->backlog = cmd.backlog > 0 && cmd.backlog < max_backlog ?
782 		       cmd.backlog : max_backlog;
783 	ret = rdma_listen(ctx->cm_id, ctx->backlog);
784 	ucma_put_ctx(ctx);
785 	return ret;
786 }
787 
788 static ssize_t ucma_accept(struct ucma_file *file, const char __user *inbuf,
789 			   int in_len, int out_len)
790 {
791 	struct rdma_ucm_accept cmd;
792 	struct rdma_conn_param conn_param;
793 	struct ucma_context *ctx;
794 	int ret;
795 
796 	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
797 		return -EFAULT;
798 
799 	ctx = ucma_get_ctx(file, cmd.id);
800 	if (IS_ERR(ctx))
801 		return PTR_ERR(ctx);
802 
803 	if (cmd.conn_param.valid) {
804 		ucma_copy_conn_param(&conn_param, &cmd.conn_param);
805 		mutex_lock(&file->mut);
806 		ret = rdma_accept(ctx->cm_id, &conn_param);
807 		if (!ret)
808 			ctx->uid = cmd.uid;
809 		mutex_unlock(&file->mut);
810 	} else
811 		ret = rdma_accept(ctx->cm_id, NULL);
812 
813 	ucma_put_ctx(ctx);
814 	return ret;
815 }
816 
817 static ssize_t ucma_reject(struct ucma_file *file, const char __user *inbuf,
818 			   int in_len, int out_len)
819 {
820 	struct rdma_ucm_reject cmd;
821 	struct ucma_context *ctx;
822 	int ret;
823 
824 	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
825 		return -EFAULT;
826 
827 	ctx = ucma_get_ctx(file, cmd.id);
828 	if (IS_ERR(ctx))
829 		return PTR_ERR(ctx);
830 
831 	ret = rdma_reject(ctx->cm_id, cmd.private_data, cmd.private_data_len);
832 	ucma_put_ctx(ctx);
833 	return ret;
834 }
835 
836 static ssize_t ucma_disconnect(struct ucma_file *file, const char __user *inbuf,
837 			       int in_len, int out_len)
838 {
839 	struct rdma_ucm_disconnect cmd;
840 	struct ucma_context *ctx;
841 	int ret;
842 
843 	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
844 		return -EFAULT;
845 
846 	ctx = ucma_get_ctx(file, cmd.id);
847 	if (IS_ERR(ctx))
848 		return PTR_ERR(ctx);
849 
850 	ret = rdma_disconnect(ctx->cm_id);
851 	ucma_put_ctx(ctx);
852 	return ret;
853 }
854 
855 static ssize_t ucma_init_qp_attr(struct ucma_file *file,
856 				 const char __user *inbuf,
857 				 int in_len, int out_len)
858 {
859 	struct rdma_ucm_init_qp_attr cmd;
860 	struct ib_uverbs_qp_attr resp;
861 	struct ucma_context *ctx;
862 	struct ib_qp_attr qp_attr;
863 	int ret;
864 
865 	if (out_len < sizeof(resp))
866 		return -ENOSPC;
867 
868 	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
869 		return -EFAULT;
870 
871 	ctx = ucma_get_ctx(file, cmd.id);
872 	if (IS_ERR(ctx))
873 		return PTR_ERR(ctx);
874 
875 	resp.qp_attr_mask = 0;
876 	memset(&qp_attr, 0, sizeof qp_attr);
877 	qp_attr.qp_state = cmd.qp_state;
878 	ret = rdma_init_qp_attr(ctx->cm_id, &qp_attr, &resp.qp_attr_mask);
879 	if (ret)
880 		goto out;
881 
882 	ib_copy_qp_attr_to_user(&resp, &qp_attr);
883 	if (copy_to_user((void __user *)(unsigned long)cmd.response,
884 			 &resp, sizeof(resp)))
885 		ret = -EFAULT;
886 
887 out:
888 	ucma_put_ctx(ctx);
889 	return ret;
890 }
891 
892 static int ucma_set_option_id(struct ucma_context *ctx, int optname,
893 			      void *optval, size_t optlen)
894 {
895 	int ret = 0;
896 
897 	switch (optname) {
898 	case RDMA_OPTION_ID_TOS:
899 		if (optlen != sizeof(u8)) {
900 			ret = -EINVAL;
901 			break;
902 		}
903 		rdma_set_service_type(ctx->cm_id, *((u8 *) optval));
904 		break;
905 	case RDMA_OPTION_ID_REUSEADDR:
906 		if (optlen != sizeof(int)) {
907 			ret = -EINVAL;
908 			break;
909 		}
910 		ret = rdma_set_reuseaddr(ctx->cm_id, *((int *) optval) ? 1 : 0);
911 		break;
912 	case RDMA_OPTION_ID_AFONLY:
913 		if (optlen != sizeof(int)) {
914 			ret = -EINVAL;
915 			break;
916 		}
917 		ret = rdma_set_afonly(ctx->cm_id, *((int *) optval) ? 1 : 0);
918 		break;
919 	default:
920 		ret = -ENOSYS;
921 	}
922 
923 	return ret;
924 }
925 
926 static int ucma_set_ib_path(struct ucma_context *ctx,
927 			    struct ib_path_rec_data *path_data, size_t optlen)
928 {
929 	struct ib_sa_path_rec sa_path;
930 	struct rdma_cm_event event;
931 	int ret;
932 
933 	if (optlen % sizeof(*path_data))
934 		return -EINVAL;
935 
936 	for (; optlen; optlen -= sizeof(*path_data), path_data++) {
937 		if (path_data->flags == (IB_PATH_GMP | IB_PATH_PRIMARY |
938 					 IB_PATH_BIDIRECTIONAL))
939 			break;
940 	}
941 
942 	if (!optlen)
943 		return -EINVAL;
944 
945 	ib_sa_unpack_path(path_data->path_rec, &sa_path);
946 	ret = rdma_set_ib_paths(ctx->cm_id, &sa_path, 1);
947 	if (ret)
948 		return ret;
949 
950 	memset(&event, 0, sizeof event);
951 	event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
952 	return ucma_event_handler(ctx->cm_id, &event);
953 }
954 
955 static int ucma_set_option_ib(struct ucma_context *ctx, int optname,
956 			      void *optval, size_t optlen)
957 {
958 	int ret;
959 
960 	switch (optname) {
961 	case RDMA_OPTION_IB_PATH:
962 		ret = ucma_set_ib_path(ctx, optval, optlen);
963 		break;
964 	default:
965 		ret = -ENOSYS;
966 	}
967 
968 	return ret;
969 }
970 
971 static int ucma_set_option_level(struct ucma_context *ctx, int level,
972 				 int optname, void *optval, size_t optlen)
973 {
974 	int ret;
975 
976 	switch (level) {
977 	case RDMA_OPTION_ID:
978 		ret = ucma_set_option_id(ctx, optname, optval, optlen);
979 		break;
980 	case RDMA_OPTION_IB:
981 		ret = ucma_set_option_ib(ctx, optname, optval, optlen);
982 		break;
983 	default:
984 		ret = -ENOSYS;
985 	}
986 
987 	return ret;
988 }
989 
990 static ssize_t ucma_set_option(struct ucma_file *file, const char __user *inbuf,
991 			       int in_len, int out_len)
992 {
993 	struct rdma_ucm_set_option cmd;
994 	struct ucma_context *ctx;
995 	void *optval;
996 	int ret;
997 
998 	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
999 		return -EFAULT;
1000 
1001 	ctx = ucma_get_ctx(file, cmd.id);
1002 	if (IS_ERR(ctx))
1003 		return PTR_ERR(ctx);
1004 
1005 	optval = kmalloc(cmd.optlen, GFP_KERNEL);
1006 	if (!optval) {
1007 		ret = -ENOMEM;
1008 		goto out1;
1009 	}
1010 
1011 	if (copy_from_user(optval, (void __user *) (unsigned long) cmd.optval,
1012 			   cmd.optlen)) {
1013 		ret = -EFAULT;
1014 		goto out2;
1015 	}
1016 
1017 	ret = ucma_set_option_level(ctx, cmd.level, cmd.optname, optval,
1018 				    cmd.optlen);
1019 out2:
1020 	kfree(optval);
1021 out1:
1022 	ucma_put_ctx(ctx);
1023 	return ret;
1024 }
1025 
1026 static ssize_t ucma_notify(struct ucma_file *file, const char __user *inbuf,
1027 			   int in_len, int out_len)
1028 {
1029 	struct rdma_ucm_notify cmd;
1030 	struct ucma_context *ctx;
1031 	int ret;
1032 
1033 	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1034 		return -EFAULT;
1035 
1036 	ctx = ucma_get_ctx(file, cmd.id);
1037 	if (IS_ERR(ctx))
1038 		return PTR_ERR(ctx);
1039 
1040 	ret = rdma_notify(ctx->cm_id, (enum ib_event_type) cmd.event);
1041 	ucma_put_ctx(ctx);
1042 	return ret;
1043 }
1044 
1045 static ssize_t ucma_join_multicast(struct ucma_file *file,
1046 				   const char __user *inbuf,
1047 				   int in_len, int out_len)
1048 {
1049 	struct rdma_ucm_join_mcast cmd;
1050 	struct rdma_ucm_create_id_resp resp;
1051 	struct ucma_context *ctx;
1052 	struct ucma_multicast *mc;
1053 	int ret;
1054 
1055 	if (out_len < sizeof(resp))
1056 		return -ENOSPC;
1057 
1058 	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1059 		return -EFAULT;
1060 
1061 	ctx = ucma_get_ctx(file, cmd.id);
1062 	if (IS_ERR(ctx))
1063 		return PTR_ERR(ctx);
1064 
1065 	mutex_lock(&file->mut);
1066 	mc = ucma_alloc_multicast(ctx);
1067 	if (!mc) {
1068 		ret = -ENOMEM;
1069 		goto err1;
1070 	}
1071 
1072 	mc->uid = cmd.uid;
1073 	memcpy(&mc->addr, &cmd.addr, sizeof cmd.addr);
1074 	ret = rdma_join_multicast(ctx->cm_id, (struct sockaddr *) &mc->addr, mc);
1075 	if (ret)
1076 		goto err2;
1077 
1078 	resp.id = mc->id;
1079 	if (copy_to_user((void __user *)(unsigned long)cmd.response,
1080 			 &resp, sizeof(resp))) {
1081 		ret = -EFAULT;
1082 		goto err3;
1083 	}
1084 
1085 	mutex_unlock(&file->mut);
1086 	ucma_put_ctx(ctx);
1087 	return 0;
1088 
1089 err3:
1090 	rdma_leave_multicast(ctx->cm_id, (struct sockaddr *) &mc->addr);
1091 	ucma_cleanup_mc_events(mc);
1092 err2:
1093 	mutex_lock(&mut);
1094 	idr_remove(&multicast_idr, mc->id);
1095 	mutex_unlock(&mut);
1096 	list_del(&mc->list);
1097 	kfree(mc);
1098 err1:
1099 	mutex_unlock(&file->mut);
1100 	ucma_put_ctx(ctx);
1101 	return ret;
1102 }
1103 
1104 static ssize_t ucma_leave_multicast(struct ucma_file *file,
1105 				    const char __user *inbuf,
1106 				    int in_len, int out_len)
1107 {
1108 	struct rdma_ucm_destroy_id cmd;
1109 	struct rdma_ucm_destroy_id_resp resp;
1110 	struct ucma_multicast *mc;
1111 	int ret = 0;
1112 
1113 	if (out_len < sizeof(resp))
1114 		return -ENOSPC;
1115 
1116 	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1117 		return -EFAULT;
1118 
1119 	mutex_lock(&mut);
1120 	mc = idr_find(&multicast_idr, cmd.id);
1121 	if (!mc)
1122 		mc = ERR_PTR(-ENOENT);
1123 	else if (mc->ctx->file != file)
1124 		mc = ERR_PTR(-EINVAL);
1125 	else {
1126 		idr_remove(&multicast_idr, mc->id);
1127 		atomic_inc(&mc->ctx->ref);
1128 	}
1129 	mutex_unlock(&mut);
1130 
1131 	if (IS_ERR(mc)) {
1132 		ret = PTR_ERR(mc);
1133 		goto out;
1134 	}
1135 
1136 	rdma_leave_multicast(mc->ctx->cm_id, (struct sockaddr *) &mc->addr);
1137 	mutex_lock(&mc->ctx->file->mut);
1138 	ucma_cleanup_mc_events(mc);
1139 	list_del(&mc->list);
1140 	mutex_unlock(&mc->ctx->file->mut);
1141 
1142 	ucma_put_ctx(mc->ctx);
1143 	resp.events_reported = mc->events_reported;
1144 	kfree(mc);
1145 
1146 	if (copy_to_user((void __user *)(unsigned long)cmd.response,
1147 			 &resp, sizeof(resp)))
1148 		ret = -EFAULT;
1149 out:
1150 	return ret;
1151 }
1152 
1153 static void ucma_lock_files(struct ucma_file *file1, struct ucma_file *file2)
1154 {
1155 	/* Acquire mutex's based on pointer comparison to prevent deadlock. */
1156 	if (file1 < file2) {
1157 		mutex_lock(&file1->mut);
1158 		mutex_lock(&file2->mut);
1159 	} else {
1160 		mutex_lock(&file2->mut);
1161 		mutex_lock(&file1->mut);
1162 	}
1163 }
1164 
1165 static void ucma_unlock_files(struct ucma_file *file1, struct ucma_file *file2)
1166 {
1167 	if (file1 < file2) {
1168 		mutex_unlock(&file2->mut);
1169 		mutex_unlock(&file1->mut);
1170 	} else {
1171 		mutex_unlock(&file1->mut);
1172 		mutex_unlock(&file2->mut);
1173 	}
1174 }
1175 
1176 static void ucma_move_events(struct ucma_context *ctx, struct ucma_file *file)
1177 {
1178 	struct ucma_event *uevent, *tmp;
1179 
1180 	list_for_each_entry_safe(uevent, tmp, &ctx->file->event_list, list)
1181 		if (uevent->ctx == ctx)
1182 			list_move_tail(&uevent->list, &file->event_list);
1183 }
1184 
1185 static ssize_t ucma_migrate_id(struct ucma_file *new_file,
1186 			       const char __user *inbuf,
1187 			       int in_len, int out_len)
1188 {
1189 	struct rdma_ucm_migrate_id cmd;
1190 	struct rdma_ucm_migrate_resp resp;
1191 	struct ucma_context *ctx;
1192 	struct file *filp;
1193 	struct ucma_file *cur_file;
1194 	int ret = 0;
1195 
1196 	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1197 		return -EFAULT;
1198 
1199 	/* Get current fd to protect against it being closed */
1200 	filp = fget(cmd.fd);
1201 	if (!filp)
1202 		return -ENOENT;
1203 
1204 	/* Validate current fd and prevent destruction of id. */
1205 	ctx = ucma_get_ctx(filp->private_data, cmd.id);
1206 	if (IS_ERR(ctx)) {
1207 		ret = PTR_ERR(ctx);
1208 		goto file_put;
1209 	}
1210 
1211 	cur_file = ctx->file;
1212 	if (cur_file == new_file) {
1213 		resp.events_reported = ctx->events_reported;
1214 		goto response;
1215 	}
1216 
1217 	/*
1218 	 * Migrate events between fd's, maintaining order, and avoiding new
1219 	 * events being added before existing events.
1220 	 */
1221 	ucma_lock_files(cur_file, new_file);
1222 	mutex_lock(&mut);
1223 
1224 	list_move_tail(&ctx->list, &new_file->ctx_list);
1225 	ucma_move_events(ctx, new_file);
1226 	ctx->file = new_file;
1227 	resp.events_reported = ctx->events_reported;
1228 
1229 	mutex_unlock(&mut);
1230 	ucma_unlock_files(cur_file, new_file);
1231 
1232 response:
1233 	if (copy_to_user((void __user *)(unsigned long)cmd.response,
1234 			 &resp, sizeof(resp)))
1235 		ret = -EFAULT;
1236 
1237 	ucma_put_ctx(ctx);
1238 file_put:
1239 	fput(filp);
1240 	return ret;
1241 }
1242 
1243 static ssize_t (*ucma_cmd_table[])(struct ucma_file *file,
1244 				   const char __user *inbuf,
1245 				   int in_len, int out_len) = {
1246 	[RDMA_USER_CM_CMD_CREATE_ID]	= ucma_create_id,
1247 	[RDMA_USER_CM_CMD_DESTROY_ID]	= ucma_destroy_id,
1248 	[RDMA_USER_CM_CMD_BIND_ADDR]	= ucma_bind_addr,
1249 	[RDMA_USER_CM_CMD_RESOLVE_ADDR]	= ucma_resolve_addr,
1250 	[RDMA_USER_CM_CMD_RESOLVE_ROUTE]= ucma_resolve_route,
1251 	[RDMA_USER_CM_CMD_QUERY_ROUTE]	= ucma_query_route,
1252 	[RDMA_USER_CM_CMD_CONNECT]	= ucma_connect,
1253 	[RDMA_USER_CM_CMD_LISTEN]	= ucma_listen,
1254 	[RDMA_USER_CM_CMD_ACCEPT]	= ucma_accept,
1255 	[RDMA_USER_CM_CMD_REJECT]	= ucma_reject,
1256 	[RDMA_USER_CM_CMD_DISCONNECT]	= ucma_disconnect,
1257 	[RDMA_USER_CM_CMD_INIT_QP_ATTR]	= ucma_init_qp_attr,
1258 	[RDMA_USER_CM_CMD_GET_EVENT]	= ucma_get_event,
1259 	[RDMA_USER_CM_CMD_GET_OPTION]	= NULL,
1260 	[RDMA_USER_CM_CMD_SET_OPTION]	= ucma_set_option,
1261 	[RDMA_USER_CM_CMD_NOTIFY]	= ucma_notify,
1262 	[RDMA_USER_CM_CMD_JOIN_MCAST]	= ucma_join_multicast,
1263 	[RDMA_USER_CM_CMD_LEAVE_MCAST]	= ucma_leave_multicast,
1264 	[RDMA_USER_CM_CMD_MIGRATE_ID]	= ucma_migrate_id
1265 };
1266 
1267 static ssize_t ucma_write(struct file *filp, const char __user *buf,
1268 			  size_t len, loff_t *pos)
1269 {
1270 	struct ucma_file *file = filp->private_data;
1271 	struct rdma_ucm_cmd_hdr hdr;
1272 	ssize_t ret;
1273 
1274 	if (len < sizeof(hdr))
1275 		return -EINVAL;
1276 
1277 	if (copy_from_user(&hdr, buf, sizeof(hdr)))
1278 		return -EFAULT;
1279 
1280 	if (hdr.cmd >= ARRAY_SIZE(ucma_cmd_table))
1281 		return -EINVAL;
1282 
1283 	if (hdr.in + sizeof(hdr) > len)
1284 		return -EINVAL;
1285 
1286 	if (!ucma_cmd_table[hdr.cmd])
1287 		return -ENOSYS;
1288 
1289 	ret = ucma_cmd_table[hdr.cmd](file, buf + sizeof(hdr), hdr.in, hdr.out);
1290 	if (!ret)
1291 		ret = len;
1292 
1293 	return ret;
1294 }
1295 
1296 static unsigned int ucma_poll(struct file *filp, struct poll_table_struct *wait)
1297 {
1298 	struct ucma_file *file = filp->private_data;
1299 	unsigned int mask = 0;
1300 
1301 	poll_wait(filp, &file->poll_wait, wait);
1302 
1303 	if (!list_empty(&file->event_list))
1304 		mask = POLLIN | POLLRDNORM;
1305 
1306 	return mask;
1307 }
1308 
1309 /*
1310  * ucma_open() does not need the BKL:
1311  *
1312  *  - no global state is referred to;
1313  *  - there is no ioctl method to race against;
1314  *  - no further module initialization is required for open to work
1315  *    after the device is registered.
1316  */
1317 static int ucma_open(struct inode *inode, struct file *filp)
1318 {
1319 	struct ucma_file *file;
1320 
1321 	file = kmalloc(sizeof *file, GFP_KERNEL);
1322 	if (!file)
1323 		return -ENOMEM;
1324 
1325 	INIT_LIST_HEAD(&file->event_list);
1326 	INIT_LIST_HEAD(&file->ctx_list);
1327 	init_waitqueue_head(&file->poll_wait);
1328 	mutex_init(&file->mut);
1329 
1330 	filp->private_data = file;
1331 	file->filp = filp;
1332 
1333 	return nonseekable_open(inode, filp);
1334 }
1335 
1336 static int ucma_close(struct inode *inode, struct file *filp)
1337 {
1338 	struct ucma_file *file = filp->private_data;
1339 	struct ucma_context *ctx, *tmp;
1340 
1341 	mutex_lock(&file->mut);
1342 	list_for_each_entry_safe(ctx, tmp, &file->ctx_list, list) {
1343 		mutex_unlock(&file->mut);
1344 
1345 		mutex_lock(&mut);
1346 		idr_remove(&ctx_idr, ctx->id);
1347 		mutex_unlock(&mut);
1348 
1349 		ucma_free_ctx(ctx);
1350 		mutex_lock(&file->mut);
1351 	}
1352 	mutex_unlock(&file->mut);
1353 	kfree(file);
1354 	return 0;
1355 }
1356 
1357 static const struct file_operations ucma_fops = {
1358 	.owner 	 = THIS_MODULE,
1359 	.open 	 = ucma_open,
1360 	.release = ucma_close,
1361 	.write	 = ucma_write,
1362 	.poll    = ucma_poll,
1363 	.llseek	 = no_llseek,
1364 };
1365 
1366 static struct miscdevice ucma_misc = {
1367 	.minor		= MISC_DYNAMIC_MINOR,
1368 	.name		= "rdma_cm",
1369 	.nodename	= "infiniband/rdma_cm",
1370 	.mode		= 0666,
1371 	.fops		= &ucma_fops,
1372 };
1373 
1374 static ssize_t show_abi_version(struct device *dev,
1375 				struct device_attribute *attr,
1376 				char *buf)
1377 {
1378 	return sprintf(buf, "%d\n", RDMA_USER_CM_ABI_VERSION);
1379 }
1380 static DEVICE_ATTR(abi_version, S_IRUGO, show_abi_version, NULL);
1381 
1382 static int __init ucma_init(void)
1383 {
1384 	int ret;
1385 
1386 	ret = misc_register(&ucma_misc);
1387 	if (ret)
1388 		return ret;
1389 
1390 	ret = device_create_file(ucma_misc.this_device, &dev_attr_abi_version);
1391 	if (ret) {
1392 		printk(KERN_ERR "rdma_ucm: couldn't create abi_version attr\n");
1393 		goto err1;
1394 	}
1395 
1396 	ucma_ctl_table_hdr = register_net_sysctl(&init_net, "net/rdma_ucm", ucma_ctl_table);
1397 	if (!ucma_ctl_table_hdr) {
1398 		printk(KERN_ERR "rdma_ucm: couldn't register sysctl paths\n");
1399 		ret = -ENOMEM;
1400 		goto err2;
1401 	}
1402 	return 0;
1403 err2:
1404 	device_remove_file(ucma_misc.this_device, &dev_attr_abi_version);
1405 err1:
1406 	misc_deregister(&ucma_misc);
1407 	return ret;
1408 }
1409 
1410 static void __exit ucma_cleanup(void)
1411 {
1412 	unregister_net_sysctl_table(ucma_ctl_table_hdr);
1413 	device_remove_file(ucma_misc.this_device, &dev_attr_abi_version);
1414 	misc_deregister(&ucma_misc);
1415 	idr_destroy(&ctx_idr);
1416 }
1417 
1418 module_init(ucma_init);
1419 module_exit(ucma_cleanup);
1420