xref: /linux/drivers/infiniband/core/ucma.c (revision f2ee442115c9b6219083c019939a9cc0c9abb2f8)
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 static struct ctl_path ucma_ctl_path[] = {
70 	{ .procname = "net" },
71 	{ .procname = "rdma_ucm" },
72 	{ }
73 };
74 
75 struct ucma_file {
76 	struct mutex		mut;
77 	struct file		*filp;
78 	struct list_head	ctx_list;
79 	struct list_head	event_list;
80 	wait_queue_head_t	poll_wait;
81 };
82 
83 struct ucma_context {
84 	int			id;
85 	struct completion	comp;
86 	atomic_t		ref;
87 	int			events_reported;
88 	int			backlog;
89 
90 	struct ucma_file	*file;
91 	struct rdma_cm_id	*cm_id;
92 	u64			uid;
93 
94 	struct list_head	list;
95 	struct list_head	mc_list;
96 };
97 
98 struct ucma_multicast {
99 	struct ucma_context	*ctx;
100 	int			id;
101 	int			events_reported;
102 
103 	u64			uid;
104 	struct list_head	list;
105 	struct sockaddr_storage	addr;
106 };
107 
108 struct ucma_event {
109 	struct ucma_context	*ctx;
110 	struct ucma_multicast	*mc;
111 	struct list_head	list;
112 	struct rdma_cm_id	*cm_id;
113 	struct rdma_ucm_event_resp resp;
114 };
115 
116 static DEFINE_MUTEX(mut);
117 static DEFINE_IDR(ctx_idr);
118 static DEFINE_IDR(multicast_idr);
119 
120 static inline struct ucma_context *_ucma_find_context(int id,
121 						      struct ucma_file *file)
122 {
123 	struct ucma_context *ctx;
124 
125 	ctx = idr_find(&ctx_idr, id);
126 	if (!ctx)
127 		ctx = ERR_PTR(-ENOENT);
128 	else if (ctx->file != file)
129 		ctx = ERR_PTR(-EINVAL);
130 	return ctx;
131 }
132 
133 static struct ucma_context *ucma_get_ctx(struct ucma_file *file, int id)
134 {
135 	struct ucma_context *ctx;
136 
137 	mutex_lock(&mut);
138 	ctx = _ucma_find_context(id, file);
139 	if (!IS_ERR(ctx))
140 		atomic_inc(&ctx->ref);
141 	mutex_unlock(&mut);
142 	return ctx;
143 }
144 
145 static void ucma_put_ctx(struct ucma_context *ctx)
146 {
147 	if (atomic_dec_and_test(&ctx->ref))
148 		complete(&ctx->comp);
149 }
150 
151 static struct ucma_context *ucma_alloc_ctx(struct ucma_file *file)
152 {
153 	struct ucma_context *ctx;
154 	int ret;
155 
156 	ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
157 	if (!ctx)
158 		return NULL;
159 
160 	atomic_set(&ctx->ref, 1);
161 	init_completion(&ctx->comp);
162 	INIT_LIST_HEAD(&ctx->mc_list);
163 	ctx->file = file;
164 
165 	do {
166 		ret = idr_pre_get(&ctx_idr, GFP_KERNEL);
167 		if (!ret)
168 			goto error;
169 
170 		mutex_lock(&mut);
171 		ret = idr_get_new(&ctx_idr, ctx, &ctx->id);
172 		mutex_unlock(&mut);
173 	} while (ret == -EAGAIN);
174 
175 	if (ret)
176 		goto error;
177 
178 	list_add_tail(&ctx->list, &file->ctx_list);
179 	return ctx;
180 
181 error:
182 	kfree(ctx);
183 	return NULL;
184 }
185 
186 static struct ucma_multicast* ucma_alloc_multicast(struct ucma_context *ctx)
187 {
188 	struct ucma_multicast *mc;
189 	int ret;
190 
191 	mc = kzalloc(sizeof(*mc), GFP_KERNEL);
192 	if (!mc)
193 		return NULL;
194 
195 	do {
196 		ret = idr_pre_get(&multicast_idr, GFP_KERNEL);
197 		if (!ret)
198 			goto error;
199 
200 		mutex_lock(&mut);
201 		ret = idr_get_new(&multicast_idr, mc, &mc->id);
202 		mutex_unlock(&mut);
203 	} while (ret == -EAGAIN);
204 
205 	if (ret)
206 		goto error;
207 
208 	mc->ctx = ctx;
209 	list_add_tail(&mc->list, &ctx->mc_list);
210 	return mc;
211 
212 error:
213 	kfree(mc);
214 	return NULL;
215 }
216 
217 static void ucma_copy_conn_event(struct rdma_ucm_conn_param *dst,
218 				 struct rdma_conn_param *src)
219 {
220 	if (src->private_data_len)
221 		memcpy(dst->private_data, src->private_data,
222 		       src->private_data_len);
223 	dst->private_data_len = src->private_data_len;
224 	dst->responder_resources =src->responder_resources;
225 	dst->initiator_depth = src->initiator_depth;
226 	dst->flow_control = src->flow_control;
227 	dst->retry_count = src->retry_count;
228 	dst->rnr_retry_count = src->rnr_retry_count;
229 	dst->srq = src->srq;
230 	dst->qp_num = src->qp_num;
231 }
232 
233 static void ucma_copy_ud_event(struct rdma_ucm_ud_param *dst,
234 			       struct rdma_ud_param *src)
235 {
236 	if (src->private_data_len)
237 		memcpy(dst->private_data, src->private_data,
238 		       src->private_data_len);
239 	dst->private_data_len = src->private_data_len;
240 	ib_copy_ah_attr_to_user(&dst->ah_attr, &src->ah_attr);
241 	dst->qp_num = src->qp_num;
242 	dst->qkey = src->qkey;
243 }
244 
245 static void ucma_set_event_context(struct ucma_context *ctx,
246 				   struct rdma_cm_event *event,
247 				   struct ucma_event *uevent)
248 {
249 	uevent->ctx = ctx;
250 	switch (event->event) {
251 	case RDMA_CM_EVENT_MULTICAST_JOIN:
252 	case RDMA_CM_EVENT_MULTICAST_ERROR:
253 		uevent->mc = (struct ucma_multicast *)
254 			     event->param.ud.private_data;
255 		uevent->resp.uid = uevent->mc->uid;
256 		uevent->resp.id = uevent->mc->id;
257 		break;
258 	default:
259 		uevent->resp.uid = ctx->uid;
260 		uevent->resp.id = ctx->id;
261 		break;
262 	}
263 }
264 
265 static int ucma_event_handler(struct rdma_cm_id *cm_id,
266 			      struct rdma_cm_event *event)
267 {
268 	struct ucma_event *uevent;
269 	struct ucma_context *ctx = cm_id->context;
270 	int ret = 0;
271 
272 	uevent = kzalloc(sizeof(*uevent), GFP_KERNEL);
273 	if (!uevent)
274 		return event->event == RDMA_CM_EVENT_CONNECT_REQUEST;
275 
276 	uevent->cm_id = cm_id;
277 	ucma_set_event_context(ctx, event, uevent);
278 	uevent->resp.event = event->event;
279 	uevent->resp.status = event->status;
280 	if (cm_id->qp_type == IB_QPT_UD)
281 		ucma_copy_ud_event(&uevent->resp.param.ud, &event->param.ud);
282 	else
283 		ucma_copy_conn_event(&uevent->resp.param.conn,
284 				     &event->param.conn);
285 
286 	mutex_lock(&ctx->file->mut);
287 	if (event->event == RDMA_CM_EVENT_CONNECT_REQUEST) {
288 		if (!ctx->backlog) {
289 			ret = -ENOMEM;
290 			kfree(uevent);
291 			goto out;
292 		}
293 		ctx->backlog--;
294 	} else if (!ctx->uid) {
295 		/*
296 		 * We ignore events for new connections until userspace has set
297 		 * their context.  This can only happen if an error occurs on a
298 		 * new connection before the user accepts it.  This is okay,
299 		 * since the accept will just fail later.
300 		 */
301 		kfree(uevent);
302 		goto out;
303 	}
304 
305 	list_add_tail(&uevent->list, &ctx->file->event_list);
306 	wake_up_interruptible(&ctx->file->poll_wait);
307 out:
308 	mutex_unlock(&ctx->file->mut);
309 	return ret;
310 }
311 
312 static ssize_t ucma_get_event(struct ucma_file *file, const char __user *inbuf,
313 			      int in_len, int out_len)
314 {
315 	struct ucma_context *ctx;
316 	struct rdma_ucm_get_event cmd;
317 	struct ucma_event *uevent;
318 	int ret = 0;
319 	DEFINE_WAIT(wait);
320 
321 	if (out_len < sizeof uevent->resp)
322 		return -ENOSPC;
323 
324 	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
325 		return -EFAULT;
326 
327 	mutex_lock(&file->mut);
328 	while (list_empty(&file->event_list)) {
329 		mutex_unlock(&file->mut);
330 
331 		if (file->filp->f_flags & O_NONBLOCK)
332 			return -EAGAIN;
333 
334 		if (wait_event_interruptible(file->poll_wait,
335 					     !list_empty(&file->event_list)))
336 			return -ERESTARTSYS;
337 
338 		mutex_lock(&file->mut);
339 	}
340 
341 	uevent = list_entry(file->event_list.next, struct ucma_event, list);
342 
343 	if (uevent->resp.event == RDMA_CM_EVENT_CONNECT_REQUEST) {
344 		ctx = ucma_alloc_ctx(file);
345 		if (!ctx) {
346 			ret = -ENOMEM;
347 			goto done;
348 		}
349 		uevent->ctx->backlog++;
350 		ctx->cm_id = uevent->cm_id;
351 		ctx->cm_id->context = ctx;
352 		uevent->resp.id = ctx->id;
353 	}
354 
355 	if (copy_to_user((void __user *)(unsigned long)cmd.response,
356 			 &uevent->resp, sizeof uevent->resp)) {
357 		ret = -EFAULT;
358 		goto done;
359 	}
360 
361 	list_del(&uevent->list);
362 	uevent->ctx->events_reported++;
363 	if (uevent->mc)
364 		uevent->mc->events_reported++;
365 	kfree(uevent);
366 done:
367 	mutex_unlock(&file->mut);
368 	return ret;
369 }
370 
371 static int ucma_get_qp_type(struct rdma_ucm_create_id *cmd, enum ib_qp_type *qp_type)
372 {
373 	switch (cmd->ps) {
374 	case RDMA_PS_TCP:
375 		*qp_type = IB_QPT_RC;
376 		return 0;
377 	case RDMA_PS_UDP:
378 	case RDMA_PS_IPOIB:
379 		*qp_type = IB_QPT_UD;
380 		return 0;
381 	case RDMA_PS_IB:
382 		*qp_type = cmd->qp_type;
383 		return 0;
384 	default:
385 		return -EINVAL;
386 	}
387 }
388 
389 static ssize_t ucma_create_id(struct ucma_file *file, const char __user *inbuf,
390 			      int in_len, int out_len)
391 {
392 	struct rdma_ucm_create_id cmd;
393 	struct rdma_ucm_create_id_resp resp;
394 	struct ucma_context *ctx;
395 	enum ib_qp_type qp_type;
396 	int ret;
397 
398 	if (out_len < sizeof(resp))
399 		return -ENOSPC;
400 
401 	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
402 		return -EFAULT;
403 
404 	ret = ucma_get_qp_type(&cmd, &qp_type);
405 	if (ret)
406 		return ret;
407 
408 	mutex_lock(&file->mut);
409 	ctx = ucma_alloc_ctx(file);
410 	mutex_unlock(&file->mut);
411 	if (!ctx)
412 		return -ENOMEM;
413 
414 	ctx->uid = cmd.uid;
415 	ctx->cm_id = rdma_create_id(ucma_event_handler, ctx, cmd.ps, qp_type);
416 	if (IS_ERR(ctx->cm_id)) {
417 		ret = PTR_ERR(ctx->cm_id);
418 		goto err1;
419 	}
420 
421 	resp.id = ctx->id;
422 	if (copy_to_user((void __user *)(unsigned long)cmd.response,
423 			 &resp, sizeof(resp))) {
424 		ret = -EFAULT;
425 		goto err2;
426 	}
427 	return 0;
428 
429 err2:
430 	rdma_destroy_id(ctx->cm_id);
431 err1:
432 	mutex_lock(&mut);
433 	idr_remove(&ctx_idr, ctx->id);
434 	mutex_unlock(&mut);
435 	kfree(ctx);
436 	return ret;
437 }
438 
439 static void ucma_cleanup_multicast(struct ucma_context *ctx)
440 {
441 	struct ucma_multicast *mc, *tmp;
442 
443 	mutex_lock(&mut);
444 	list_for_each_entry_safe(mc, tmp, &ctx->mc_list, list) {
445 		list_del(&mc->list);
446 		idr_remove(&multicast_idr, mc->id);
447 		kfree(mc);
448 	}
449 	mutex_unlock(&mut);
450 }
451 
452 static void ucma_cleanup_events(struct ucma_context *ctx)
453 {
454 	struct ucma_event *uevent, *tmp;
455 
456 	list_for_each_entry_safe(uevent, tmp, &ctx->file->event_list, list) {
457 		if (uevent->ctx != ctx)
458 			continue;
459 
460 		list_del(&uevent->list);
461 
462 		/* clear incoming connections. */
463 		if (uevent->resp.event == RDMA_CM_EVENT_CONNECT_REQUEST)
464 			rdma_destroy_id(uevent->cm_id);
465 
466 		kfree(uevent);
467 	}
468 }
469 
470 static void ucma_cleanup_mc_events(struct ucma_multicast *mc)
471 {
472 	struct ucma_event *uevent, *tmp;
473 
474 	list_for_each_entry_safe(uevent, tmp, &mc->ctx->file->event_list, list) {
475 		if (uevent->mc != mc)
476 			continue;
477 
478 		list_del(&uevent->list);
479 		kfree(uevent);
480 	}
481 }
482 
483 static int ucma_free_ctx(struct ucma_context *ctx)
484 {
485 	int events_reported;
486 
487 	/* No new events will be generated after destroying the id. */
488 	rdma_destroy_id(ctx->cm_id);
489 
490 	ucma_cleanup_multicast(ctx);
491 
492 	/* Cleanup events not yet reported to the user. */
493 	mutex_lock(&ctx->file->mut);
494 	ucma_cleanup_events(ctx);
495 	list_del(&ctx->list);
496 	mutex_unlock(&ctx->file->mut);
497 
498 	events_reported = ctx->events_reported;
499 	kfree(ctx);
500 	return events_reported;
501 }
502 
503 static ssize_t ucma_destroy_id(struct ucma_file *file, const char __user *inbuf,
504 			       int in_len, int out_len)
505 {
506 	struct rdma_ucm_destroy_id cmd;
507 	struct rdma_ucm_destroy_id_resp resp;
508 	struct ucma_context *ctx;
509 	int ret = 0;
510 
511 	if (out_len < sizeof(resp))
512 		return -ENOSPC;
513 
514 	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
515 		return -EFAULT;
516 
517 	mutex_lock(&mut);
518 	ctx = _ucma_find_context(cmd.id, file);
519 	if (!IS_ERR(ctx))
520 		idr_remove(&ctx_idr, ctx->id);
521 	mutex_unlock(&mut);
522 
523 	if (IS_ERR(ctx))
524 		return PTR_ERR(ctx);
525 
526 	ucma_put_ctx(ctx);
527 	wait_for_completion(&ctx->comp);
528 	resp.events_reported = ucma_free_ctx(ctx);
529 
530 	if (copy_to_user((void __user *)(unsigned long)cmd.response,
531 			 &resp, sizeof(resp)))
532 		ret = -EFAULT;
533 
534 	return ret;
535 }
536 
537 static ssize_t ucma_bind_addr(struct ucma_file *file, const char __user *inbuf,
538 			      int in_len, int out_len)
539 {
540 	struct rdma_ucm_bind_addr cmd;
541 	struct ucma_context *ctx;
542 	int ret;
543 
544 	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
545 		return -EFAULT;
546 
547 	ctx = ucma_get_ctx(file, cmd.id);
548 	if (IS_ERR(ctx))
549 		return PTR_ERR(ctx);
550 
551 	ret = rdma_bind_addr(ctx->cm_id, (struct sockaddr *) &cmd.addr);
552 	ucma_put_ctx(ctx);
553 	return ret;
554 }
555 
556 static ssize_t ucma_resolve_addr(struct ucma_file *file,
557 				 const char __user *inbuf,
558 				 int in_len, int out_len)
559 {
560 	struct rdma_ucm_resolve_addr cmd;
561 	struct ucma_context *ctx;
562 	int ret;
563 
564 	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
565 		return -EFAULT;
566 
567 	ctx = ucma_get_ctx(file, cmd.id);
568 	if (IS_ERR(ctx))
569 		return PTR_ERR(ctx);
570 
571 	ret = rdma_resolve_addr(ctx->cm_id, (struct sockaddr *) &cmd.src_addr,
572 				(struct sockaddr *) &cmd.dst_addr,
573 				cmd.timeout_ms);
574 	ucma_put_ctx(ctx);
575 	return ret;
576 }
577 
578 static ssize_t ucma_resolve_route(struct ucma_file *file,
579 				  const char __user *inbuf,
580 				  int in_len, int out_len)
581 {
582 	struct rdma_ucm_resolve_route cmd;
583 	struct ucma_context *ctx;
584 	int ret;
585 
586 	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
587 		return -EFAULT;
588 
589 	ctx = ucma_get_ctx(file, cmd.id);
590 	if (IS_ERR(ctx))
591 		return PTR_ERR(ctx);
592 
593 	ret = rdma_resolve_route(ctx->cm_id, cmd.timeout_ms);
594 	ucma_put_ctx(ctx);
595 	return ret;
596 }
597 
598 static void ucma_copy_ib_route(struct rdma_ucm_query_route_resp *resp,
599 			       struct rdma_route *route)
600 {
601 	struct rdma_dev_addr *dev_addr;
602 
603 	resp->num_paths = route->num_paths;
604 	switch (route->num_paths) {
605 	case 0:
606 		dev_addr = &route->addr.dev_addr;
607 		rdma_addr_get_dgid(dev_addr,
608 				   (union ib_gid *) &resp->ib_route[0].dgid);
609 		rdma_addr_get_sgid(dev_addr,
610 				   (union ib_gid *) &resp->ib_route[0].sgid);
611 		resp->ib_route[0].pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
612 		break;
613 	case 2:
614 		ib_copy_path_rec_to_user(&resp->ib_route[1],
615 					 &route->path_rec[1]);
616 		/* fall through */
617 	case 1:
618 		ib_copy_path_rec_to_user(&resp->ib_route[0],
619 					 &route->path_rec[0]);
620 		break;
621 	default:
622 		break;
623 	}
624 }
625 
626 static void ucma_copy_iboe_route(struct rdma_ucm_query_route_resp *resp,
627 				 struct rdma_route *route)
628 {
629 	struct rdma_dev_addr *dev_addr;
630 	struct net_device *dev;
631 	u16 vid = 0;
632 
633 	resp->num_paths = route->num_paths;
634 	switch (route->num_paths) {
635 	case 0:
636 		dev_addr = &route->addr.dev_addr;
637 		dev = dev_get_by_index(&init_net, dev_addr->bound_dev_if);
638 			if (dev) {
639 				vid = rdma_vlan_dev_vlan_id(dev);
640 				dev_put(dev);
641 			}
642 
643 		iboe_mac_vlan_to_ll((union ib_gid *) &resp->ib_route[0].dgid,
644 				    dev_addr->dst_dev_addr, vid);
645 		iboe_addr_get_sgid(dev_addr,
646 				   (union ib_gid *) &resp->ib_route[0].sgid);
647 		resp->ib_route[0].pkey = cpu_to_be16(0xffff);
648 		break;
649 	case 2:
650 		ib_copy_path_rec_to_user(&resp->ib_route[1],
651 					 &route->path_rec[1]);
652 		/* fall through */
653 	case 1:
654 		ib_copy_path_rec_to_user(&resp->ib_route[0],
655 					 &route->path_rec[0]);
656 		break;
657 	default:
658 		break;
659 	}
660 }
661 
662 static void ucma_copy_iw_route(struct rdma_ucm_query_route_resp *resp,
663 			       struct rdma_route *route)
664 {
665 	struct rdma_dev_addr *dev_addr;
666 
667 	dev_addr = &route->addr.dev_addr;
668 	rdma_addr_get_dgid(dev_addr, (union ib_gid *) &resp->ib_route[0].dgid);
669 	rdma_addr_get_sgid(dev_addr, (union ib_gid *) &resp->ib_route[0].sgid);
670 }
671 
672 static ssize_t ucma_query_route(struct ucma_file *file,
673 				const char __user *inbuf,
674 				int in_len, int out_len)
675 {
676 	struct rdma_ucm_query_route cmd;
677 	struct rdma_ucm_query_route_resp resp;
678 	struct ucma_context *ctx;
679 	struct sockaddr *addr;
680 	int ret = 0;
681 
682 	if (out_len < sizeof(resp))
683 		return -ENOSPC;
684 
685 	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
686 		return -EFAULT;
687 
688 	ctx = ucma_get_ctx(file, cmd.id);
689 	if (IS_ERR(ctx))
690 		return PTR_ERR(ctx);
691 
692 	memset(&resp, 0, sizeof resp);
693 	addr = (struct sockaddr *) &ctx->cm_id->route.addr.src_addr;
694 	memcpy(&resp.src_addr, addr, addr->sa_family == AF_INET ?
695 				     sizeof(struct sockaddr_in) :
696 				     sizeof(struct sockaddr_in6));
697 	addr = (struct sockaddr *) &ctx->cm_id->route.addr.dst_addr;
698 	memcpy(&resp.dst_addr, addr, addr->sa_family == AF_INET ?
699 				     sizeof(struct sockaddr_in) :
700 				     sizeof(struct sockaddr_in6));
701 	if (!ctx->cm_id->device)
702 		goto out;
703 
704 	resp.node_guid = (__force __u64) ctx->cm_id->device->node_guid;
705 	resp.port_num = ctx->cm_id->port_num;
706 	switch (rdma_node_get_transport(ctx->cm_id->device->node_type)) {
707 	case RDMA_TRANSPORT_IB:
708 		switch (rdma_port_get_link_layer(ctx->cm_id->device,
709 			ctx->cm_id->port_num)) {
710 		case IB_LINK_LAYER_INFINIBAND:
711 			ucma_copy_ib_route(&resp, &ctx->cm_id->route);
712 			break;
713 		case IB_LINK_LAYER_ETHERNET:
714 			ucma_copy_iboe_route(&resp, &ctx->cm_id->route);
715 			break;
716 		default:
717 			break;
718 		}
719 		break;
720 	case RDMA_TRANSPORT_IWARP:
721 		ucma_copy_iw_route(&resp, &ctx->cm_id->route);
722 		break;
723 	default:
724 		break;
725 	}
726 
727 out:
728 	if (copy_to_user((void __user *)(unsigned long)cmd.response,
729 			 &resp, sizeof(resp)))
730 		ret = -EFAULT;
731 
732 	ucma_put_ctx(ctx);
733 	return ret;
734 }
735 
736 static void ucma_copy_conn_param(struct rdma_conn_param *dst,
737 				 struct rdma_ucm_conn_param *src)
738 {
739 	dst->private_data = src->private_data;
740 	dst->private_data_len = src->private_data_len;
741 	dst->responder_resources =src->responder_resources;
742 	dst->initiator_depth = src->initiator_depth;
743 	dst->flow_control = src->flow_control;
744 	dst->retry_count = src->retry_count;
745 	dst->rnr_retry_count = src->rnr_retry_count;
746 	dst->srq = src->srq;
747 	dst->qp_num = src->qp_num;
748 }
749 
750 static ssize_t ucma_connect(struct ucma_file *file, const char __user *inbuf,
751 			    int in_len, int out_len)
752 {
753 	struct rdma_ucm_connect cmd;
754 	struct rdma_conn_param conn_param;
755 	struct ucma_context *ctx;
756 	int ret;
757 
758 	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
759 		return -EFAULT;
760 
761 	if (!cmd.conn_param.valid)
762 		return -EINVAL;
763 
764 	ctx = ucma_get_ctx(file, cmd.id);
765 	if (IS_ERR(ctx))
766 		return PTR_ERR(ctx);
767 
768 	ucma_copy_conn_param(&conn_param, &cmd.conn_param);
769 	ret = rdma_connect(ctx->cm_id, &conn_param);
770 	ucma_put_ctx(ctx);
771 	return ret;
772 }
773 
774 static ssize_t ucma_listen(struct ucma_file *file, const char __user *inbuf,
775 			   int in_len, int out_len)
776 {
777 	struct rdma_ucm_listen cmd;
778 	struct ucma_context *ctx;
779 	int ret;
780 
781 	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
782 		return -EFAULT;
783 
784 	ctx = ucma_get_ctx(file, cmd.id);
785 	if (IS_ERR(ctx))
786 		return PTR_ERR(ctx);
787 
788 	ctx->backlog = cmd.backlog > 0 && cmd.backlog < max_backlog ?
789 		       cmd.backlog : max_backlog;
790 	ret = rdma_listen(ctx->cm_id, ctx->backlog);
791 	ucma_put_ctx(ctx);
792 	return ret;
793 }
794 
795 static ssize_t ucma_accept(struct ucma_file *file, const char __user *inbuf,
796 			   int in_len, int out_len)
797 {
798 	struct rdma_ucm_accept cmd;
799 	struct rdma_conn_param conn_param;
800 	struct ucma_context *ctx;
801 	int ret;
802 
803 	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
804 		return -EFAULT;
805 
806 	ctx = ucma_get_ctx(file, cmd.id);
807 	if (IS_ERR(ctx))
808 		return PTR_ERR(ctx);
809 
810 	if (cmd.conn_param.valid) {
811 		ctx->uid = cmd.uid;
812 		ucma_copy_conn_param(&conn_param, &cmd.conn_param);
813 		ret = rdma_accept(ctx->cm_id, &conn_param);
814 	} else
815 		ret = rdma_accept(ctx->cm_id, NULL);
816 
817 	ucma_put_ctx(ctx);
818 	return ret;
819 }
820 
821 static ssize_t ucma_reject(struct ucma_file *file, const char __user *inbuf,
822 			   int in_len, int out_len)
823 {
824 	struct rdma_ucm_reject cmd;
825 	struct ucma_context *ctx;
826 	int ret;
827 
828 	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
829 		return -EFAULT;
830 
831 	ctx = ucma_get_ctx(file, cmd.id);
832 	if (IS_ERR(ctx))
833 		return PTR_ERR(ctx);
834 
835 	ret = rdma_reject(ctx->cm_id, cmd.private_data, cmd.private_data_len);
836 	ucma_put_ctx(ctx);
837 	return ret;
838 }
839 
840 static ssize_t ucma_disconnect(struct ucma_file *file, const char __user *inbuf,
841 			       int in_len, int out_len)
842 {
843 	struct rdma_ucm_disconnect cmd;
844 	struct ucma_context *ctx;
845 	int ret;
846 
847 	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
848 		return -EFAULT;
849 
850 	ctx = ucma_get_ctx(file, cmd.id);
851 	if (IS_ERR(ctx))
852 		return PTR_ERR(ctx);
853 
854 	ret = rdma_disconnect(ctx->cm_id);
855 	ucma_put_ctx(ctx);
856 	return ret;
857 }
858 
859 static ssize_t ucma_init_qp_attr(struct ucma_file *file,
860 				 const char __user *inbuf,
861 				 int in_len, int out_len)
862 {
863 	struct rdma_ucm_init_qp_attr cmd;
864 	struct ib_uverbs_qp_attr resp;
865 	struct ucma_context *ctx;
866 	struct ib_qp_attr qp_attr;
867 	int ret;
868 
869 	if (out_len < sizeof(resp))
870 		return -ENOSPC;
871 
872 	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
873 		return -EFAULT;
874 
875 	ctx = ucma_get_ctx(file, cmd.id);
876 	if (IS_ERR(ctx))
877 		return PTR_ERR(ctx);
878 
879 	resp.qp_attr_mask = 0;
880 	memset(&qp_attr, 0, sizeof qp_attr);
881 	qp_attr.qp_state = cmd.qp_state;
882 	ret = rdma_init_qp_attr(ctx->cm_id, &qp_attr, &resp.qp_attr_mask);
883 	if (ret)
884 		goto out;
885 
886 	ib_copy_qp_attr_to_user(&resp, &qp_attr);
887 	if (copy_to_user((void __user *)(unsigned long)cmd.response,
888 			 &resp, sizeof(resp)))
889 		ret = -EFAULT;
890 
891 out:
892 	ucma_put_ctx(ctx);
893 	return ret;
894 }
895 
896 static int ucma_set_option_id(struct ucma_context *ctx, int optname,
897 			      void *optval, size_t optlen)
898 {
899 	int ret = 0;
900 
901 	switch (optname) {
902 	case RDMA_OPTION_ID_TOS:
903 		if (optlen != sizeof(u8)) {
904 			ret = -EINVAL;
905 			break;
906 		}
907 		rdma_set_service_type(ctx->cm_id, *((u8 *) optval));
908 		break;
909 	case RDMA_OPTION_ID_REUSEADDR:
910 		if (optlen != sizeof(int)) {
911 			ret = -EINVAL;
912 			break;
913 		}
914 		ret = rdma_set_reuseaddr(ctx->cm_id, *((int *) optval) ? 1 : 0);
915 		break;
916 	default:
917 		ret = -ENOSYS;
918 	}
919 
920 	return ret;
921 }
922 
923 static int ucma_set_ib_path(struct ucma_context *ctx,
924 			    struct ib_path_rec_data *path_data, size_t optlen)
925 {
926 	struct ib_sa_path_rec sa_path;
927 	struct rdma_cm_event event;
928 	int ret;
929 
930 	if (optlen % sizeof(*path_data))
931 		return -EINVAL;
932 
933 	for (; optlen; optlen -= sizeof(*path_data), path_data++) {
934 		if (path_data->flags == (IB_PATH_GMP | IB_PATH_PRIMARY |
935 					 IB_PATH_BIDIRECTIONAL))
936 			break;
937 	}
938 
939 	if (!optlen)
940 		return -EINVAL;
941 
942 	ib_sa_unpack_path(path_data->path_rec, &sa_path);
943 	ret = rdma_set_ib_paths(ctx->cm_id, &sa_path, 1);
944 	if (ret)
945 		return ret;
946 
947 	memset(&event, 0, sizeof event);
948 	event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
949 	return ucma_event_handler(ctx->cm_id, &event);
950 }
951 
952 static int ucma_set_option_ib(struct ucma_context *ctx, int optname,
953 			      void *optval, size_t optlen)
954 {
955 	int ret;
956 
957 	switch (optname) {
958 	case RDMA_OPTION_IB_PATH:
959 		ret = ucma_set_ib_path(ctx, optval, optlen);
960 		break;
961 	default:
962 		ret = -ENOSYS;
963 	}
964 
965 	return ret;
966 }
967 
968 static int ucma_set_option_level(struct ucma_context *ctx, int level,
969 				 int optname, void *optval, size_t optlen)
970 {
971 	int ret;
972 
973 	switch (level) {
974 	case RDMA_OPTION_ID:
975 		ret = ucma_set_option_id(ctx, optname, optval, optlen);
976 		break;
977 	case RDMA_OPTION_IB:
978 		ret = ucma_set_option_ib(ctx, optname, optval, optlen);
979 		break;
980 	default:
981 		ret = -ENOSYS;
982 	}
983 
984 	return ret;
985 }
986 
987 static ssize_t ucma_set_option(struct ucma_file *file, const char __user *inbuf,
988 			       int in_len, int out_len)
989 {
990 	struct rdma_ucm_set_option cmd;
991 	struct ucma_context *ctx;
992 	void *optval;
993 	int ret;
994 
995 	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
996 		return -EFAULT;
997 
998 	ctx = ucma_get_ctx(file, cmd.id);
999 	if (IS_ERR(ctx))
1000 		return PTR_ERR(ctx);
1001 
1002 	optval = kmalloc(cmd.optlen, GFP_KERNEL);
1003 	if (!optval) {
1004 		ret = -ENOMEM;
1005 		goto out1;
1006 	}
1007 
1008 	if (copy_from_user(optval, (void __user *) (unsigned long) cmd.optval,
1009 			   cmd.optlen)) {
1010 		ret = -EFAULT;
1011 		goto out2;
1012 	}
1013 
1014 	ret = ucma_set_option_level(ctx, cmd.level, cmd.optname, optval,
1015 				    cmd.optlen);
1016 out2:
1017 	kfree(optval);
1018 out1:
1019 	ucma_put_ctx(ctx);
1020 	return ret;
1021 }
1022 
1023 static ssize_t ucma_notify(struct ucma_file *file, const char __user *inbuf,
1024 			   int in_len, int out_len)
1025 {
1026 	struct rdma_ucm_notify cmd;
1027 	struct ucma_context *ctx;
1028 	int ret;
1029 
1030 	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1031 		return -EFAULT;
1032 
1033 	ctx = ucma_get_ctx(file, cmd.id);
1034 	if (IS_ERR(ctx))
1035 		return PTR_ERR(ctx);
1036 
1037 	ret = rdma_notify(ctx->cm_id, (enum ib_event_type) cmd.event);
1038 	ucma_put_ctx(ctx);
1039 	return ret;
1040 }
1041 
1042 static ssize_t ucma_join_multicast(struct ucma_file *file,
1043 				   const char __user *inbuf,
1044 				   int in_len, int out_len)
1045 {
1046 	struct rdma_ucm_join_mcast cmd;
1047 	struct rdma_ucm_create_id_resp resp;
1048 	struct ucma_context *ctx;
1049 	struct ucma_multicast *mc;
1050 	int ret;
1051 
1052 	if (out_len < sizeof(resp))
1053 		return -ENOSPC;
1054 
1055 	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1056 		return -EFAULT;
1057 
1058 	ctx = ucma_get_ctx(file, cmd.id);
1059 	if (IS_ERR(ctx))
1060 		return PTR_ERR(ctx);
1061 
1062 	mutex_lock(&file->mut);
1063 	mc = ucma_alloc_multicast(ctx);
1064 	if (!mc) {
1065 		ret = -ENOMEM;
1066 		goto err1;
1067 	}
1068 
1069 	mc->uid = cmd.uid;
1070 	memcpy(&mc->addr, &cmd.addr, sizeof cmd.addr);
1071 	ret = rdma_join_multicast(ctx->cm_id, (struct sockaddr *) &mc->addr, mc);
1072 	if (ret)
1073 		goto err2;
1074 
1075 	resp.id = mc->id;
1076 	if (copy_to_user((void __user *)(unsigned long)cmd.response,
1077 			 &resp, sizeof(resp))) {
1078 		ret = -EFAULT;
1079 		goto err3;
1080 	}
1081 
1082 	mutex_unlock(&file->mut);
1083 	ucma_put_ctx(ctx);
1084 	return 0;
1085 
1086 err3:
1087 	rdma_leave_multicast(ctx->cm_id, (struct sockaddr *) &mc->addr);
1088 	ucma_cleanup_mc_events(mc);
1089 err2:
1090 	mutex_lock(&mut);
1091 	idr_remove(&multicast_idr, mc->id);
1092 	mutex_unlock(&mut);
1093 	list_del(&mc->list);
1094 	kfree(mc);
1095 err1:
1096 	mutex_unlock(&file->mut);
1097 	ucma_put_ctx(ctx);
1098 	return ret;
1099 }
1100 
1101 static ssize_t ucma_leave_multicast(struct ucma_file *file,
1102 				    const char __user *inbuf,
1103 				    int in_len, int out_len)
1104 {
1105 	struct rdma_ucm_destroy_id cmd;
1106 	struct rdma_ucm_destroy_id_resp resp;
1107 	struct ucma_multicast *mc;
1108 	int ret = 0;
1109 
1110 	if (out_len < sizeof(resp))
1111 		return -ENOSPC;
1112 
1113 	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1114 		return -EFAULT;
1115 
1116 	mutex_lock(&mut);
1117 	mc = idr_find(&multicast_idr, cmd.id);
1118 	if (!mc)
1119 		mc = ERR_PTR(-ENOENT);
1120 	else if (mc->ctx->file != file)
1121 		mc = ERR_PTR(-EINVAL);
1122 	else {
1123 		idr_remove(&multicast_idr, mc->id);
1124 		atomic_inc(&mc->ctx->ref);
1125 	}
1126 	mutex_unlock(&mut);
1127 
1128 	if (IS_ERR(mc)) {
1129 		ret = PTR_ERR(mc);
1130 		goto out;
1131 	}
1132 
1133 	rdma_leave_multicast(mc->ctx->cm_id, (struct sockaddr *) &mc->addr);
1134 	mutex_lock(&mc->ctx->file->mut);
1135 	ucma_cleanup_mc_events(mc);
1136 	list_del(&mc->list);
1137 	mutex_unlock(&mc->ctx->file->mut);
1138 
1139 	ucma_put_ctx(mc->ctx);
1140 	resp.events_reported = mc->events_reported;
1141 	kfree(mc);
1142 
1143 	if (copy_to_user((void __user *)(unsigned long)cmd.response,
1144 			 &resp, sizeof(resp)))
1145 		ret = -EFAULT;
1146 out:
1147 	return ret;
1148 }
1149 
1150 static void ucma_lock_files(struct ucma_file *file1, struct ucma_file *file2)
1151 {
1152 	/* Acquire mutex's based on pointer comparison to prevent deadlock. */
1153 	if (file1 < file2) {
1154 		mutex_lock(&file1->mut);
1155 		mutex_lock(&file2->mut);
1156 	} else {
1157 		mutex_lock(&file2->mut);
1158 		mutex_lock(&file1->mut);
1159 	}
1160 }
1161 
1162 static void ucma_unlock_files(struct ucma_file *file1, struct ucma_file *file2)
1163 {
1164 	if (file1 < file2) {
1165 		mutex_unlock(&file2->mut);
1166 		mutex_unlock(&file1->mut);
1167 	} else {
1168 		mutex_unlock(&file1->mut);
1169 		mutex_unlock(&file2->mut);
1170 	}
1171 }
1172 
1173 static void ucma_move_events(struct ucma_context *ctx, struct ucma_file *file)
1174 {
1175 	struct ucma_event *uevent, *tmp;
1176 
1177 	list_for_each_entry_safe(uevent, tmp, &ctx->file->event_list, list)
1178 		if (uevent->ctx == ctx)
1179 			list_move_tail(&uevent->list, &file->event_list);
1180 }
1181 
1182 static ssize_t ucma_migrate_id(struct ucma_file *new_file,
1183 			       const char __user *inbuf,
1184 			       int in_len, int out_len)
1185 {
1186 	struct rdma_ucm_migrate_id cmd;
1187 	struct rdma_ucm_migrate_resp resp;
1188 	struct ucma_context *ctx;
1189 	struct file *filp;
1190 	struct ucma_file *cur_file;
1191 	int ret = 0;
1192 
1193 	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1194 		return -EFAULT;
1195 
1196 	/* Get current fd to protect against it being closed */
1197 	filp = fget(cmd.fd);
1198 	if (!filp)
1199 		return -ENOENT;
1200 
1201 	/* Validate current fd and prevent destruction of id. */
1202 	ctx = ucma_get_ctx(filp->private_data, cmd.id);
1203 	if (IS_ERR(ctx)) {
1204 		ret = PTR_ERR(ctx);
1205 		goto file_put;
1206 	}
1207 
1208 	cur_file = ctx->file;
1209 	if (cur_file == new_file) {
1210 		resp.events_reported = ctx->events_reported;
1211 		goto response;
1212 	}
1213 
1214 	/*
1215 	 * Migrate events between fd's, maintaining order, and avoiding new
1216 	 * events being added before existing events.
1217 	 */
1218 	ucma_lock_files(cur_file, new_file);
1219 	mutex_lock(&mut);
1220 
1221 	list_move_tail(&ctx->list, &new_file->ctx_list);
1222 	ucma_move_events(ctx, new_file);
1223 	ctx->file = new_file;
1224 	resp.events_reported = ctx->events_reported;
1225 
1226 	mutex_unlock(&mut);
1227 	ucma_unlock_files(cur_file, new_file);
1228 
1229 response:
1230 	if (copy_to_user((void __user *)(unsigned long)cmd.response,
1231 			 &resp, sizeof(resp)))
1232 		ret = -EFAULT;
1233 
1234 	ucma_put_ctx(ctx);
1235 file_put:
1236 	fput(filp);
1237 	return ret;
1238 }
1239 
1240 static ssize_t (*ucma_cmd_table[])(struct ucma_file *file,
1241 				   const char __user *inbuf,
1242 				   int in_len, int out_len) = {
1243 	[RDMA_USER_CM_CMD_CREATE_ID]	= ucma_create_id,
1244 	[RDMA_USER_CM_CMD_DESTROY_ID]	= ucma_destroy_id,
1245 	[RDMA_USER_CM_CMD_BIND_ADDR]	= ucma_bind_addr,
1246 	[RDMA_USER_CM_CMD_RESOLVE_ADDR]	= ucma_resolve_addr,
1247 	[RDMA_USER_CM_CMD_RESOLVE_ROUTE]= ucma_resolve_route,
1248 	[RDMA_USER_CM_CMD_QUERY_ROUTE]	= ucma_query_route,
1249 	[RDMA_USER_CM_CMD_CONNECT]	= ucma_connect,
1250 	[RDMA_USER_CM_CMD_LISTEN]	= ucma_listen,
1251 	[RDMA_USER_CM_CMD_ACCEPT]	= ucma_accept,
1252 	[RDMA_USER_CM_CMD_REJECT]	= ucma_reject,
1253 	[RDMA_USER_CM_CMD_DISCONNECT]	= ucma_disconnect,
1254 	[RDMA_USER_CM_CMD_INIT_QP_ATTR]	= ucma_init_qp_attr,
1255 	[RDMA_USER_CM_CMD_GET_EVENT]	= ucma_get_event,
1256 	[RDMA_USER_CM_CMD_GET_OPTION]	= NULL,
1257 	[RDMA_USER_CM_CMD_SET_OPTION]	= ucma_set_option,
1258 	[RDMA_USER_CM_CMD_NOTIFY]	= ucma_notify,
1259 	[RDMA_USER_CM_CMD_JOIN_MCAST]	= ucma_join_multicast,
1260 	[RDMA_USER_CM_CMD_LEAVE_MCAST]	= ucma_leave_multicast,
1261 	[RDMA_USER_CM_CMD_MIGRATE_ID]	= ucma_migrate_id
1262 };
1263 
1264 static ssize_t ucma_write(struct file *filp, const char __user *buf,
1265 			  size_t len, loff_t *pos)
1266 {
1267 	struct ucma_file *file = filp->private_data;
1268 	struct rdma_ucm_cmd_hdr hdr;
1269 	ssize_t ret;
1270 
1271 	if (len < sizeof(hdr))
1272 		return -EINVAL;
1273 
1274 	if (copy_from_user(&hdr, buf, sizeof(hdr)))
1275 		return -EFAULT;
1276 
1277 	if (hdr.cmd >= ARRAY_SIZE(ucma_cmd_table))
1278 		return -EINVAL;
1279 
1280 	if (hdr.in + sizeof(hdr) > len)
1281 		return -EINVAL;
1282 
1283 	if (!ucma_cmd_table[hdr.cmd])
1284 		return -ENOSYS;
1285 
1286 	ret = ucma_cmd_table[hdr.cmd](file, buf + sizeof(hdr), hdr.in, hdr.out);
1287 	if (!ret)
1288 		ret = len;
1289 
1290 	return ret;
1291 }
1292 
1293 static unsigned int ucma_poll(struct file *filp, struct poll_table_struct *wait)
1294 {
1295 	struct ucma_file *file = filp->private_data;
1296 	unsigned int mask = 0;
1297 
1298 	poll_wait(filp, &file->poll_wait, wait);
1299 
1300 	if (!list_empty(&file->event_list))
1301 		mask = POLLIN | POLLRDNORM;
1302 
1303 	return mask;
1304 }
1305 
1306 /*
1307  * ucma_open() does not need the BKL:
1308  *
1309  *  - no global state is referred to;
1310  *  - there is no ioctl method to race against;
1311  *  - no further module initialization is required for open to work
1312  *    after the device is registered.
1313  */
1314 static int ucma_open(struct inode *inode, struct file *filp)
1315 {
1316 	struct ucma_file *file;
1317 
1318 	file = kmalloc(sizeof *file, GFP_KERNEL);
1319 	if (!file)
1320 		return -ENOMEM;
1321 
1322 	INIT_LIST_HEAD(&file->event_list);
1323 	INIT_LIST_HEAD(&file->ctx_list);
1324 	init_waitqueue_head(&file->poll_wait);
1325 	mutex_init(&file->mut);
1326 
1327 	filp->private_data = file;
1328 	file->filp = filp;
1329 
1330 	return nonseekable_open(inode, filp);
1331 }
1332 
1333 static int ucma_close(struct inode *inode, struct file *filp)
1334 {
1335 	struct ucma_file *file = filp->private_data;
1336 	struct ucma_context *ctx, *tmp;
1337 
1338 	mutex_lock(&file->mut);
1339 	list_for_each_entry_safe(ctx, tmp, &file->ctx_list, list) {
1340 		mutex_unlock(&file->mut);
1341 
1342 		mutex_lock(&mut);
1343 		idr_remove(&ctx_idr, ctx->id);
1344 		mutex_unlock(&mut);
1345 
1346 		ucma_free_ctx(ctx);
1347 		mutex_lock(&file->mut);
1348 	}
1349 	mutex_unlock(&file->mut);
1350 	kfree(file);
1351 	return 0;
1352 }
1353 
1354 static const struct file_operations ucma_fops = {
1355 	.owner 	 = THIS_MODULE,
1356 	.open 	 = ucma_open,
1357 	.release = ucma_close,
1358 	.write	 = ucma_write,
1359 	.poll    = ucma_poll,
1360 	.llseek	 = no_llseek,
1361 };
1362 
1363 static struct miscdevice ucma_misc = {
1364 	.minor		= MISC_DYNAMIC_MINOR,
1365 	.name		= "rdma_cm",
1366 	.nodename	= "infiniband/rdma_cm",
1367 	.mode		= 0666,
1368 	.fops		= &ucma_fops,
1369 };
1370 
1371 static ssize_t show_abi_version(struct device *dev,
1372 				struct device_attribute *attr,
1373 				char *buf)
1374 {
1375 	return sprintf(buf, "%d\n", RDMA_USER_CM_ABI_VERSION);
1376 }
1377 static DEVICE_ATTR(abi_version, S_IRUGO, show_abi_version, NULL);
1378 
1379 static int __init ucma_init(void)
1380 {
1381 	int ret;
1382 
1383 	ret = misc_register(&ucma_misc);
1384 	if (ret)
1385 		return ret;
1386 
1387 	ret = device_create_file(ucma_misc.this_device, &dev_attr_abi_version);
1388 	if (ret) {
1389 		printk(KERN_ERR "rdma_ucm: couldn't create abi_version attr\n");
1390 		goto err1;
1391 	}
1392 
1393 	ucma_ctl_table_hdr = register_sysctl_paths(ucma_ctl_path, ucma_ctl_table);
1394 	if (!ucma_ctl_table_hdr) {
1395 		printk(KERN_ERR "rdma_ucm: couldn't register sysctl paths\n");
1396 		ret = -ENOMEM;
1397 		goto err2;
1398 	}
1399 	return 0;
1400 err2:
1401 	device_remove_file(ucma_misc.this_device, &dev_attr_abi_version);
1402 err1:
1403 	misc_deregister(&ucma_misc);
1404 	return ret;
1405 }
1406 
1407 static void __exit ucma_cleanup(void)
1408 {
1409 	unregister_sysctl_table(ucma_ctl_table_hdr);
1410 	device_remove_file(ucma_misc.this_device, &dev_attr_abi_version);
1411 	misc_deregister(&ucma_misc);
1412 	idr_destroy(&ctx_idr);
1413 }
1414 
1415 module_init(ucma_init);
1416 module_exit(ucma_cleanup);
1417