xref: /openbmc/linux/drivers/infiniband/core/ucma.c (revision 7dd65feb)
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 
43 #include <rdma/rdma_user_cm.h>
44 #include <rdma/ib_marshall.h>
45 #include <rdma/rdma_cm.h>
46 #include <rdma/rdma_cm_ib.h>
47 
48 MODULE_AUTHOR("Sean Hefty");
49 MODULE_DESCRIPTION("RDMA Userspace Connection Manager Access");
50 MODULE_LICENSE("Dual BSD/GPL");
51 
52 enum {
53 	UCMA_MAX_BACKLOG	= 128
54 };
55 
56 struct ucma_file {
57 	struct mutex		mut;
58 	struct file		*filp;
59 	struct list_head	ctx_list;
60 	struct list_head	event_list;
61 	wait_queue_head_t	poll_wait;
62 };
63 
64 struct ucma_context {
65 	int			id;
66 	struct completion	comp;
67 	atomic_t		ref;
68 	int			events_reported;
69 	int			backlog;
70 
71 	struct ucma_file	*file;
72 	struct rdma_cm_id	*cm_id;
73 	u64			uid;
74 
75 	struct list_head	list;
76 	struct list_head	mc_list;
77 };
78 
79 struct ucma_multicast {
80 	struct ucma_context	*ctx;
81 	int			id;
82 	int			events_reported;
83 
84 	u64			uid;
85 	struct list_head	list;
86 	struct sockaddr_storage	addr;
87 };
88 
89 struct ucma_event {
90 	struct ucma_context	*ctx;
91 	struct ucma_multicast	*mc;
92 	struct list_head	list;
93 	struct rdma_cm_id	*cm_id;
94 	struct rdma_ucm_event_resp resp;
95 };
96 
97 static DEFINE_MUTEX(mut);
98 static DEFINE_IDR(ctx_idr);
99 static DEFINE_IDR(multicast_idr);
100 
101 static inline struct ucma_context *_ucma_find_context(int id,
102 						      struct ucma_file *file)
103 {
104 	struct ucma_context *ctx;
105 
106 	ctx = idr_find(&ctx_idr, id);
107 	if (!ctx)
108 		ctx = ERR_PTR(-ENOENT);
109 	else if (ctx->file != file)
110 		ctx = ERR_PTR(-EINVAL);
111 	return ctx;
112 }
113 
114 static struct ucma_context *ucma_get_ctx(struct ucma_file *file, int id)
115 {
116 	struct ucma_context *ctx;
117 
118 	mutex_lock(&mut);
119 	ctx = _ucma_find_context(id, file);
120 	if (!IS_ERR(ctx))
121 		atomic_inc(&ctx->ref);
122 	mutex_unlock(&mut);
123 	return ctx;
124 }
125 
126 static void ucma_put_ctx(struct ucma_context *ctx)
127 {
128 	if (atomic_dec_and_test(&ctx->ref))
129 		complete(&ctx->comp);
130 }
131 
132 static struct ucma_context *ucma_alloc_ctx(struct ucma_file *file)
133 {
134 	struct ucma_context *ctx;
135 	int ret;
136 
137 	ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
138 	if (!ctx)
139 		return NULL;
140 
141 	atomic_set(&ctx->ref, 1);
142 	init_completion(&ctx->comp);
143 	INIT_LIST_HEAD(&ctx->mc_list);
144 	ctx->file = file;
145 
146 	do {
147 		ret = idr_pre_get(&ctx_idr, GFP_KERNEL);
148 		if (!ret)
149 			goto error;
150 
151 		mutex_lock(&mut);
152 		ret = idr_get_new(&ctx_idr, ctx, &ctx->id);
153 		mutex_unlock(&mut);
154 	} while (ret == -EAGAIN);
155 
156 	if (ret)
157 		goto error;
158 
159 	list_add_tail(&ctx->list, &file->ctx_list);
160 	return ctx;
161 
162 error:
163 	kfree(ctx);
164 	return NULL;
165 }
166 
167 static struct ucma_multicast* ucma_alloc_multicast(struct ucma_context *ctx)
168 {
169 	struct ucma_multicast *mc;
170 	int ret;
171 
172 	mc = kzalloc(sizeof(*mc), GFP_KERNEL);
173 	if (!mc)
174 		return NULL;
175 
176 	do {
177 		ret = idr_pre_get(&multicast_idr, GFP_KERNEL);
178 		if (!ret)
179 			goto error;
180 
181 		mutex_lock(&mut);
182 		ret = idr_get_new(&multicast_idr, mc, &mc->id);
183 		mutex_unlock(&mut);
184 	} while (ret == -EAGAIN);
185 
186 	if (ret)
187 		goto error;
188 
189 	mc->ctx = ctx;
190 	list_add_tail(&mc->list, &ctx->mc_list);
191 	return mc;
192 
193 error:
194 	kfree(mc);
195 	return NULL;
196 }
197 
198 static void ucma_copy_conn_event(struct rdma_ucm_conn_param *dst,
199 				 struct rdma_conn_param *src)
200 {
201 	if (src->private_data_len)
202 		memcpy(dst->private_data, src->private_data,
203 		       src->private_data_len);
204 	dst->private_data_len = src->private_data_len;
205 	dst->responder_resources =src->responder_resources;
206 	dst->initiator_depth = src->initiator_depth;
207 	dst->flow_control = src->flow_control;
208 	dst->retry_count = src->retry_count;
209 	dst->rnr_retry_count = src->rnr_retry_count;
210 	dst->srq = src->srq;
211 	dst->qp_num = src->qp_num;
212 }
213 
214 static void ucma_copy_ud_event(struct rdma_ucm_ud_param *dst,
215 			       struct rdma_ud_param *src)
216 {
217 	if (src->private_data_len)
218 		memcpy(dst->private_data, src->private_data,
219 		       src->private_data_len);
220 	dst->private_data_len = src->private_data_len;
221 	ib_copy_ah_attr_to_user(&dst->ah_attr, &src->ah_attr);
222 	dst->qp_num = src->qp_num;
223 	dst->qkey = src->qkey;
224 }
225 
226 static void ucma_set_event_context(struct ucma_context *ctx,
227 				   struct rdma_cm_event *event,
228 				   struct ucma_event *uevent)
229 {
230 	uevent->ctx = ctx;
231 	switch (event->event) {
232 	case RDMA_CM_EVENT_MULTICAST_JOIN:
233 	case RDMA_CM_EVENT_MULTICAST_ERROR:
234 		uevent->mc = (struct ucma_multicast *)
235 			     event->param.ud.private_data;
236 		uevent->resp.uid = uevent->mc->uid;
237 		uevent->resp.id = uevent->mc->id;
238 		break;
239 	default:
240 		uevent->resp.uid = ctx->uid;
241 		uevent->resp.id = ctx->id;
242 		break;
243 	}
244 }
245 
246 static int ucma_event_handler(struct rdma_cm_id *cm_id,
247 			      struct rdma_cm_event *event)
248 {
249 	struct ucma_event *uevent;
250 	struct ucma_context *ctx = cm_id->context;
251 	int ret = 0;
252 
253 	uevent = kzalloc(sizeof(*uevent), GFP_KERNEL);
254 	if (!uevent)
255 		return event->event == RDMA_CM_EVENT_CONNECT_REQUEST;
256 
257 	uevent->cm_id = cm_id;
258 	ucma_set_event_context(ctx, event, uevent);
259 	uevent->resp.event = event->event;
260 	uevent->resp.status = event->status;
261 	if (cm_id->ps == RDMA_PS_UDP || cm_id->ps == RDMA_PS_IPOIB)
262 		ucma_copy_ud_event(&uevent->resp.param.ud, &event->param.ud);
263 	else
264 		ucma_copy_conn_event(&uevent->resp.param.conn,
265 				     &event->param.conn);
266 
267 	mutex_lock(&ctx->file->mut);
268 	if (event->event == RDMA_CM_EVENT_CONNECT_REQUEST) {
269 		if (!ctx->backlog) {
270 			ret = -ENOMEM;
271 			kfree(uevent);
272 			goto out;
273 		}
274 		ctx->backlog--;
275 	} else if (!ctx->uid) {
276 		/*
277 		 * We ignore events for new connections until userspace has set
278 		 * their context.  This can only happen if an error occurs on a
279 		 * new connection before the user accepts it.  This is okay,
280 		 * since the accept will just fail later.
281 		 */
282 		kfree(uevent);
283 		goto out;
284 	}
285 
286 	list_add_tail(&uevent->list, &ctx->file->event_list);
287 	wake_up_interruptible(&ctx->file->poll_wait);
288 out:
289 	mutex_unlock(&ctx->file->mut);
290 	return ret;
291 }
292 
293 static ssize_t ucma_get_event(struct ucma_file *file, const char __user *inbuf,
294 			      int in_len, int out_len)
295 {
296 	struct ucma_context *ctx;
297 	struct rdma_ucm_get_event cmd;
298 	struct ucma_event *uevent;
299 	int ret = 0;
300 	DEFINE_WAIT(wait);
301 
302 	if (out_len < sizeof uevent->resp)
303 		return -ENOSPC;
304 
305 	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
306 		return -EFAULT;
307 
308 	mutex_lock(&file->mut);
309 	while (list_empty(&file->event_list)) {
310 		mutex_unlock(&file->mut);
311 
312 		if (file->filp->f_flags & O_NONBLOCK)
313 			return -EAGAIN;
314 
315 		if (wait_event_interruptible(file->poll_wait,
316 					     !list_empty(&file->event_list)))
317 			return -ERESTARTSYS;
318 
319 		mutex_lock(&file->mut);
320 	}
321 
322 	uevent = list_entry(file->event_list.next, struct ucma_event, list);
323 
324 	if (uevent->resp.event == RDMA_CM_EVENT_CONNECT_REQUEST) {
325 		ctx = ucma_alloc_ctx(file);
326 		if (!ctx) {
327 			ret = -ENOMEM;
328 			goto done;
329 		}
330 		uevent->ctx->backlog++;
331 		ctx->cm_id = uevent->cm_id;
332 		ctx->cm_id->context = ctx;
333 		uevent->resp.id = ctx->id;
334 	}
335 
336 	if (copy_to_user((void __user *)(unsigned long)cmd.response,
337 			 &uevent->resp, sizeof uevent->resp)) {
338 		ret = -EFAULT;
339 		goto done;
340 	}
341 
342 	list_del(&uevent->list);
343 	uevent->ctx->events_reported++;
344 	if (uevent->mc)
345 		uevent->mc->events_reported++;
346 	kfree(uevent);
347 done:
348 	mutex_unlock(&file->mut);
349 	return ret;
350 }
351 
352 static ssize_t ucma_create_id(struct ucma_file *file,
353 				const char __user *inbuf,
354 				int in_len, int out_len)
355 {
356 	struct rdma_ucm_create_id cmd;
357 	struct rdma_ucm_create_id_resp resp;
358 	struct ucma_context *ctx;
359 	int ret;
360 
361 	if (out_len < sizeof(resp))
362 		return -ENOSPC;
363 
364 	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
365 		return -EFAULT;
366 
367 	mutex_lock(&file->mut);
368 	ctx = ucma_alloc_ctx(file);
369 	mutex_unlock(&file->mut);
370 	if (!ctx)
371 		return -ENOMEM;
372 
373 	ctx->uid = cmd.uid;
374 	ctx->cm_id = rdma_create_id(ucma_event_handler, ctx, cmd.ps);
375 	if (IS_ERR(ctx->cm_id)) {
376 		ret = PTR_ERR(ctx->cm_id);
377 		goto err1;
378 	}
379 
380 	resp.id = ctx->id;
381 	if (copy_to_user((void __user *)(unsigned long)cmd.response,
382 			 &resp, sizeof(resp))) {
383 		ret = -EFAULT;
384 		goto err2;
385 	}
386 	return 0;
387 
388 err2:
389 	rdma_destroy_id(ctx->cm_id);
390 err1:
391 	mutex_lock(&mut);
392 	idr_remove(&ctx_idr, ctx->id);
393 	mutex_unlock(&mut);
394 	kfree(ctx);
395 	return ret;
396 }
397 
398 static void ucma_cleanup_multicast(struct ucma_context *ctx)
399 {
400 	struct ucma_multicast *mc, *tmp;
401 
402 	mutex_lock(&mut);
403 	list_for_each_entry_safe(mc, tmp, &ctx->mc_list, list) {
404 		list_del(&mc->list);
405 		idr_remove(&multicast_idr, mc->id);
406 		kfree(mc);
407 	}
408 	mutex_unlock(&mut);
409 }
410 
411 static void ucma_cleanup_events(struct ucma_context *ctx)
412 {
413 	struct ucma_event *uevent, *tmp;
414 
415 	list_for_each_entry_safe(uevent, tmp, &ctx->file->event_list, list) {
416 		if (uevent->ctx != ctx)
417 			continue;
418 
419 		list_del(&uevent->list);
420 
421 		/* clear incoming connections. */
422 		if (uevent->resp.event == RDMA_CM_EVENT_CONNECT_REQUEST)
423 			rdma_destroy_id(uevent->cm_id);
424 
425 		kfree(uevent);
426 	}
427 }
428 
429 static void ucma_cleanup_mc_events(struct ucma_multicast *mc)
430 {
431 	struct ucma_event *uevent, *tmp;
432 
433 	list_for_each_entry_safe(uevent, tmp, &mc->ctx->file->event_list, list) {
434 		if (uevent->mc != mc)
435 			continue;
436 
437 		list_del(&uevent->list);
438 		kfree(uevent);
439 	}
440 }
441 
442 static int ucma_free_ctx(struct ucma_context *ctx)
443 {
444 	int events_reported;
445 
446 	/* No new events will be generated after destroying the id. */
447 	rdma_destroy_id(ctx->cm_id);
448 
449 	ucma_cleanup_multicast(ctx);
450 
451 	/* Cleanup events not yet reported to the user. */
452 	mutex_lock(&ctx->file->mut);
453 	ucma_cleanup_events(ctx);
454 	list_del(&ctx->list);
455 	mutex_unlock(&ctx->file->mut);
456 
457 	events_reported = ctx->events_reported;
458 	kfree(ctx);
459 	return events_reported;
460 }
461 
462 static ssize_t ucma_destroy_id(struct ucma_file *file, const char __user *inbuf,
463 			       int in_len, int out_len)
464 {
465 	struct rdma_ucm_destroy_id cmd;
466 	struct rdma_ucm_destroy_id_resp resp;
467 	struct ucma_context *ctx;
468 	int ret = 0;
469 
470 	if (out_len < sizeof(resp))
471 		return -ENOSPC;
472 
473 	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
474 		return -EFAULT;
475 
476 	mutex_lock(&mut);
477 	ctx = _ucma_find_context(cmd.id, file);
478 	if (!IS_ERR(ctx))
479 		idr_remove(&ctx_idr, ctx->id);
480 	mutex_unlock(&mut);
481 
482 	if (IS_ERR(ctx))
483 		return PTR_ERR(ctx);
484 
485 	ucma_put_ctx(ctx);
486 	wait_for_completion(&ctx->comp);
487 	resp.events_reported = ucma_free_ctx(ctx);
488 
489 	if (copy_to_user((void __user *)(unsigned long)cmd.response,
490 			 &resp, sizeof(resp)))
491 		ret = -EFAULT;
492 
493 	return ret;
494 }
495 
496 static ssize_t ucma_bind_addr(struct ucma_file *file, const char __user *inbuf,
497 			      int in_len, int out_len)
498 {
499 	struct rdma_ucm_bind_addr cmd;
500 	struct ucma_context *ctx;
501 	int ret;
502 
503 	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
504 		return -EFAULT;
505 
506 	ctx = ucma_get_ctx(file, cmd.id);
507 	if (IS_ERR(ctx))
508 		return PTR_ERR(ctx);
509 
510 	ret = rdma_bind_addr(ctx->cm_id, (struct sockaddr *) &cmd.addr);
511 	ucma_put_ctx(ctx);
512 	return ret;
513 }
514 
515 static ssize_t ucma_resolve_addr(struct ucma_file *file,
516 				 const char __user *inbuf,
517 				 int in_len, int out_len)
518 {
519 	struct rdma_ucm_resolve_addr cmd;
520 	struct ucma_context *ctx;
521 	int ret;
522 
523 	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
524 		return -EFAULT;
525 
526 	ctx = ucma_get_ctx(file, cmd.id);
527 	if (IS_ERR(ctx))
528 		return PTR_ERR(ctx);
529 
530 	ret = rdma_resolve_addr(ctx->cm_id, (struct sockaddr *) &cmd.src_addr,
531 				(struct sockaddr *) &cmd.dst_addr,
532 				cmd.timeout_ms);
533 	ucma_put_ctx(ctx);
534 	return ret;
535 }
536 
537 static ssize_t ucma_resolve_route(struct ucma_file *file,
538 				  const char __user *inbuf,
539 				  int in_len, int out_len)
540 {
541 	struct rdma_ucm_resolve_route cmd;
542 	struct ucma_context *ctx;
543 	int ret;
544 
545 	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
546 		return -EFAULT;
547 
548 	ctx = ucma_get_ctx(file, cmd.id);
549 	if (IS_ERR(ctx))
550 		return PTR_ERR(ctx);
551 
552 	ret = rdma_resolve_route(ctx->cm_id, cmd.timeout_ms);
553 	ucma_put_ctx(ctx);
554 	return ret;
555 }
556 
557 static void ucma_copy_ib_route(struct rdma_ucm_query_route_resp *resp,
558 			       struct rdma_route *route)
559 {
560 	struct rdma_dev_addr *dev_addr;
561 
562 	resp->num_paths = route->num_paths;
563 	switch (route->num_paths) {
564 	case 0:
565 		dev_addr = &route->addr.dev_addr;
566 		rdma_addr_get_dgid(dev_addr,
567 				   (union ib_gid *) &resp->ib_route[0].dgid);
568 		rdma_addr_get_sgid(dev_addr,
569 				   (union ib_gid *) &resp->ib_route[0].sgid);
570 		resp->ib_route[0].pkey = cpu_to_be16(ib_addr_get_pkey(dev_addr));
571 		break;
572 	case 2:
573 		ib_copy_path_rec_to_user(&resp->ib_route[1],
574 					 &route->path_rec[1]);
575 		/* fall through */
576 	case 1:
577 		ib_copy_path_rec_to_user(&resp->ib_route[0],
578 					 &route->path_rec[0]);
579 		break;
580 	default:
581 		break;
582 	}
583 }
584 
585 static ssize_t ucma_query_route(struct ucma_file *file,
586 				const char __user *inbuf,
587 				int in_len, int out_len)
588 {
589 	struct rdma_ucm_query_route cmd;
590 	struct rdma_ucm_query_route_resp resp;
591 	struct ucma_context *ctx;
592 	struct sockaddr *addr;
593 	int ret = 0;
594 
595 	if (out_len < sizeof(resp))
596 		return -ENOSPC;
597 
598 	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
599 		return -EFAULT;
600 
601 	ctx = ucma_get_ctx(file, cmd.id);
602 	if (IS_ERR(ctx))
603 		return PTR_ERR(ctx);
604 
605 	memset(&resp, 0, sizeof resp);
606 	addr = (struct sockaddr *) &ctx->cm_id->route.addr.src_addr;
607 	memcpy(&resp.src_addr, addr, addr->sa_family == AF_INET ?
608 				     sizeof(struct sockaddr_in) :
609 				     sizeof(struct sockaddr_in6));
610 	addr = (struct sockaddr *) &ctx->cm_id->route.addr.dst_addr;
611 	memcpy(&resp.dst_addr, addr, addr->sa_family == AF_INET ?
612 				     sizeof(struct sockaddr_in) :
613 				     sizeof(struct sockaddr_in6));
614 	if (!ctx->cm_id->device)
615 		goto out;
616 
617 	resp.node_guid = (__force __u64) ctx->cm_id->device->node_guid;
618 	resp.port_num = ctx->cm_id->port_num;
619 	switch (rdma_node_get_transport(ctx->cm_id->device->node_type)) {
620 	case RDMA_TRANSPORT_IB:
621 		ucma_copy_ib_route(&resp, &ctx->cm_id->route);
622 		break;
623 	default:
624 		break;
625 	}
626 
627 out:
628 	if (copy_to_user((void __user *)(unsigned long)cmd.response,
629 			 &resp, sizeof(resp)))
630 		ret = -EFAULT;
631 
632 	ucma_put_ctx(ctx);
633 	return ret;
634 }
635 
636 static void ucma_copy_conn_param(struct rdma_conn_param *dst,
637 				 struct rdma_ucm_conn_param *src)
638 {
639 	dst->private_data = src->private_data;
640 	dst->private_data_len = src->private_data_len;
641 	dst->responder_resources =src->responder_resources;
642 	dst->initiator_depth = src->initiator_depth;
643 	dst->flow_control = src->flow_control;
644 	dst->retry_count = src->retry_count;
645 	dst->rnr_retry_count = src->rnr_retry_count;
646 	dst->srq = src->srq;
647 	dst->qp_num = src->qp_num;
648 }
649 
650 static ssize_t ucma_connect(struct ucma_file *file, const char __user *inbuf,
651 			    int in_len, int out_len)
652 {
653 	struct rdma_ucm_connect cmd;
654 	struct rdma_conn_param conn_param;
655 	struct ucma_context *ctx;
656 	int ret;
657 
658 	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
659 		return -EFAULT;
660 
661 	if (!cmd.conn_param.valid)
662 		return -EINVAL;
663 
664 	ctx = ucma_get_ctx(file, cmd.id);
665 	if (IS_ERR(ctx))
666 		return PTR_ERR(ctx);
667 
668 	ucma_copy_conn_param(&conn_param, &cmd.conn_param);
669 	ret = rdma_connect(ctx->cm_id, &conn_param);
670 	ucma_put_ctx(ctx);
671 	return ret;
672 }
673 
674 static ssize_t ucma_listen(struct ucma_file *file, const char __user *inbuf,
675 			   int in_len, int out_len)
676 {
677 	struct rdma_ucm_listen cmd;
678 	struct ucma_context *ctx;
679 	int ret;
680 
681 	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
682 		return -EFAULT;
683 
684 	ctx = ucma_get_ctx(file, cmd.id);
685 	if (IS_ERR(ctx))
686 		return PTR_ERR(ctx);
687 
688 	ctx->backlog = cmd.backlog > 0 && cmd.backlog < UCMA_MAX_BACKLOG ?
689 		       cmd.backlog : UCMA_MAX_BACKLOG;
690 	ret = rdma_listen(ctx->cm_id, ctx->backlog);
691 	ucma_put_ctx(ctx);
692 	return ret;
693 }
694 
695 static ssize_t ucma_accept(struct ucma_file *file, const char __user *inbuf,
696 			   int in_len, int out_len)
697 {
698 	struct rdma_ucm_accept cmd;
699 	struct rdma_conn_param conn_param;
700 	struct ucma_context *ctx;
701 	int ret;
702 
703 	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
704 		return -EFAULT;
705 
706 	ctx = ucma_get_ctx(file, cmd.id);
707 	if (IS_ERR(ctx))
708 		return PTR_ERR(ctx);
709 
710 	if (cmd.conn_param.valid) {
711 		ctx->uid = cmd.uid;
712 		ucma_copy_conn_param(&conn_param, &cmd.conn_param);
713 		ret = rdma_accept(ctx->cm_id, &conn_param);
714 	} else
715 		ret = rdma_accept(ctx->cm_id, NULL);
716 
717 	ucma_put_ctx(ctx);
718 	return ret;
719 }
720 
721 static ssize_t ucma_reject(struct ucma_file *file, const char __user *inbuf,
722 			   int in_len, int out_len)
723 {
724 	struct rdma_ucm_reject cmd;
725 	struct ucma_context *ctx;
726 	int ret;
727 
728 	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
729 		return -EFAULT;
730 
731 	ctx = ucma_get_ctx(file, cmd.id);
732 	if (IS_ERR(ctx))
733 		return PTR_ERR(ctx);
734 
735 	ret = rdma_reject(ctx->cm_id, cmd.private_data, cmd.private_data_len);
736 	ucma_put_ctx(ctx);
737 	return ret;
738 }
739 
740 static ssize_t ucma_disconnect(struct ucma_file *file, const char __user *inbuf,
741 			       int in_len, int out_len)
742 {
743 	struct rdma_ucm_disconnect cmd;
744 	struct ucma_context *ctx;
745 	int ret;
746 
747 	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
748 		return -EFAULT;
749 
750 	ctx = ucma_get_ctx(file, cmd.id);
751 	if (IS_ERR(ctx))
752 		return PTR_ERR(ctx);
753 
754 	ret = rdma_disconnect(ctx->cm_id);
755 	ucma_put_ctx(ctx);
756 	return ret;
757 }
758 
759 static ssize_t ucma_init_qp_attr(struct ucma_file *file,
760 				 const char __user *inbuf,
761 				 int in_len, int out_len)
762 {
763 	struct rdma_ucm_init_qp_attr cmd;
764 	struct ib_uverbs_qp_attr resp;
765 	struct ucma_context *ctx;
766 	struct ib_qp_attr qp_attr;
767 	int ret;
768 
769 	if (out_len < sizeof(resp))
770 		return -ENOSPC;
771 
772 	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
773 		return -EFAULT;
774 
775 	ctx = ucma_get_ctx(file, cmd.id);
776 	if (IS_ERR(ctx))
777 		return PTR_ERR(ctx);
778 
779 	resp.qp_attr_mask = 0;
780 	memset(&qp_attr, 0, sizeof qp_attr);
781 	qp_attr.qp_state = cmd.qp_state;
782 	ret = rdma_init_qp_attr(ctx->cm_id, &qp_attr, &resp.qp_attr_mask);
783 	if (ret)
784 		goto out;
785 
786 	ib_copy_qp_attr_to_user(&resp, &qp_attr);
787 	if (copy_to_user((void __user *)(unsigned long)cmd.response,
788 			 &resp, sizeof(resp)))
789 		ret = -EFAULT;
790 
791 out:
792 	ucma_put_ctx(ctx);
793 	return ret;
794 }
795 
796 static int ucma_set_option_id(struct ucma_context *ctx, int optname,
797 			      void *optval, size_t optlen)
798 {
799 	int ret = 0;
800 
801 	switch (optname) {
802 	case RDMA_OPTION_ID_TOS:
803 		if (optlen != sizeof(u8)) {
804 			ret = -EINVAL;
805 			break;
806 		}
807 		rdma_set_service_type(ctx->cm_id, *((u8 *) optval));
808 		break;
809 	default:
810 		ret = -ENOSYS;
811 	}
812 
813 	return ret;
814 }
815 
816 static int ucma_set_ib_path(struct ucma_context *ctx,
817 			    struct ib_path_rec_data *path_data, size_t optlen)
818 {
819 	struct ib_sa_path_rec sa_path;
820 	struct rdma_cm_event event;
821 	int ret;
822 
823 	if (optlen % sizeof(*path_data))
824 		return -EINVAL;
825 
826 	for (; optlen; optlen -= sizeof(*path_data), path_data++) {
827 		if (path_data->flags == (IB_PATH_GMP | IB_PATH_PRIMARY |
828 					 IB_PATH_BIDIRECTIONAL))
829 			break;
830 	}
831 
832 	if (!optlen)
833 		return -EINVAL;
834 
835 	ib_sa_unpack_path(path_data->path_rec, &sa_path);
836 	ret = rdma_set_ib_paths(ctx->cm_id, &sa_path, 1);
837 	if (ret)
838 		return ret;
839 
840 	memset(&event, 0, sizeof event);
841 	event.event = RDMA_CM_EVENT_ROUTE_RESOLVED;
842 	return ucma_event_handler(ctx->cm_id, &event);
843 }
844 
845 static int ucma_set_option_ib(struct ucma_context *ctx, int optname,
846 			      void *optval, size_t optlen)
847 {
848 	int ret;
849 
850 	switch (optname) {
851 	case RDMA_OPTION_IB_PATH:
852 		ret = ucma_set_ib_path(ctx, optval, optlen);
853 		break;
854 	default:
855 		ret = -ENOSYS;
856 	}
857 
858 	return ret;
859 }
860 
861 static int ucma_set_option_level(struct ucma_context *ctx, int level,
862 				 int optname, void *optval, size_t optlen)
863 {
864 	int ret;
865 
866 	switch (level) {
867 	case RDMA_OPTION_ID:
868 		ret = ucma_set_option_id(ctx, optname, optval, optlen);
869 		break;
870 	case RDMA_OPTION_IB:
871 		ret = ucma_set_option_ib(ctx, optname, optval, optlen);
872 		break;
873 	default:
874 		ret = -ENOSYS;
875 	}
876 
877 	return ret;
878 }
879 
880 static ssize_t ucma_set_option(struct ucma_file *file, const char __user *inbuf,
881 			       int in_len, int out_len)
882 {
883 	struct rdma_ucm_set_option cmd;
884 	struct ucma_context *ctx;
885 	void *optval;
886 	int ret;
887 
888 	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
889 		return -EFAULT;
890 
891 	ctx = ucma_get_ctx(file, cmd.id);
892 	if (IS_ERR(ctx))
893 		return PTR_ERR(ctx);
894 
895 	optval = kmalloc(cmd.optlen, GFP_KERNEL);
896 	if (!optval) {
897 		ret = -ENOMEM;
898 		goto out1;
899 	}
900 
901 	if (copy_from_user(optval, (void __user *) (unsigned long) cmd.optval,
902 			   cmd.optlen)) {
903 		ret = -EFAULT;
904 		goto out2;
905 	}
906 
907 	ret = ucma_set_option_level(ctx, cmd.level, cmd.optname, optval,
908 				    cmd.optlen);
909 out2:
910 	kfree(optval);
911 out1:
912 	ucma_put_ctx(ctx);
913 	return ret;
914 }
915 
916 static ssize_t ucma_notify(struct ucma_file *file, const char __user *inbuf,
917 			   int in_len, int out_len)
918 {
919 	struct rdma_ucm_notify cmd;
920 	struct ucma_context *ctx;
921 	int ret;
922 
923 	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
924 		return -EFAULT;
925 
926 	ctx = ucma_get_ctx(file, cmd.id);
927 	if (IS_ERR(ctx))
928 		return PTR_ERR(ctx);
929 
930 	ret = rdma_notify(ctx->cm_id, (enum ib_event_type) cmd.event);
931 	ucma_put_ctx(ctx);
932 	return ret;
933 }
934 
935 static ssize_t ucma_join_multicast(struct ucma_file *file,
936 				   const char __user *inbuf,
937 				   int in_len, int out_len)
938 {
939 	struct rdma_ucm_join_mcast cmd;
940 	struct rdma_ucm_create_id_resp resp;
941 	struct ucma_context *ctx;
942 	struct ucma_multicast *mc;
943 	int ret;
944 
945 	if (out_len < sizeof(resp))
946 		return -ENOSPC;
947 
948 	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
949 		return -EFAULT;
950 
951 	ctx = ucma_get_ctx(file, cmd.id);
952 	if (IS_ERR(ctx))
953 		return PTR_ERR(ctx);
954 
955 	mutex_lock(&file->mut);
956 	mc = ucma_alloc_multicast(ctx);
957 	if (!mc) {
958 		ret = -ENOMEM;
959 		goto err1;
960 	}
961 
962 	mc->uid = cmd.uid;
963 	memcpy(&mc->addr, &cmd.addr, sizeof cmd.addr);
964 	ret = rdma_join_multicast(ctx->cm_id, (struct sockaddr *) &mc->addr, mc);
965 	if (ret)
966 		goto err2;
967 
968 	resp.id = mc->id;
969 	if (copy_to_user((void __user *)(unsigned long)cmd.response,
970 			 &resp, sizeof(resp))) {
971 		ret = -EFAULT;
972 		goto err3;
973 	}
974 
975 	mutex_unlock(&file->mut);
976 	ucma_put_ctx(ctx);
977 	return 0;
978 
979 err3:
980 	rdma_leave_multicast(ctx->cm_id, (struct sockaddr *) &mc->addr);
981 	ucma_cleanup_mc_events(mc);
982 err2:
983 	mutex_lock(&mut);
984 	idr_remove(&multicast_idr, mc->id);
985 	mutex_unlock(&mut);
986 	list_del(&mc->list);
987 	kfree(mc);
988 err1:
989 	mutex_unlock(&file->mut);
990 	ucma_put_ctx(ctx);
991 	return ret;
992 }
993 
994 static ssize_t ucma_leave_multicast(struct ucma_file *file,
995 				    const char __user *inbuf,
996 				    int in_len, int out_len)
997 {
998 	struct rdma_ucm_destroy_id cmd;
999 	struct rdma_ucm_destroy_id_resp resp;
1000 	struct ucma_multicast *mc;
1001 	int ret = 0;
1002 
1003 	if (out_len < sizeof(resp))
1004 		return -ENOSPC;
1005 
1006 	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1007 		return -EFAULT;
1008 
1009 	mutex_lock(&mut);
1010 	mc = idr_find(&multicast_idr, cmd.id);
1011 	if (!mc)
1012 		mc = ERR_PTR(-ENOENT);
1013 	else if (mc->ctx->file != file)
1014 		mc = ERR_PTR(-EINVAL);
1015 	else {
1016 		idr_remove(&multicast_idr, mc->id);
1017 		atomic_inc(&mc->ctx->ref);
1018 	}
1019 	mutex_unlock(&mut);
1020 
1021 	if (IS_ERR(mc)) {
1022 		ret = PTR_ERR(mc);
1023 		goto out;
1024 	}
1025 
1026 	rdma_leave_multicast(mc->ctx->cm_id, (struct sockaddr *) &mc->addr);
1027 	mutex_lock(&mc->ctx->file->mut);
1028 	ucma_cleanup_mc_events(mc);
1029 	list_del(&mc->list);
1030 	mutex_unlock(&mc->ctx->file->mut);
1031 
1032 	ucma_put_ctx(mc->ctx);
1033 	resp.events_reported = mc->events_reported;
1034 	kfree(mc);
1035 
1036 	if (copy_to_user((void __user *)(unsigned long)cmd.response,
1037 			 &resp, sizeof(resp)))
1038 		ret = -EFAULT;
1039 out:
1040 	return ret;
1041 }
1042 
1043 static void ucma_lock_files(struct ucma_file *file1, struct ucma_file *file2)
1044 {
1045 	/* Acquire mutex's based on pointer comparison to prevent deadlock. */
1046 	if (file1 < file2) {
1047 		mutex_lock(&file1->mut);
1048 		mutex_lock(&file2->mut);
1049 	} else {
1050 		mutex_lock(&file2->mut);
1051 		mutex_lock(&file1->mut);
1052 	}
1053 }
1054 
1055 static void ucma_unlock_files(struct ucma_file *file1, struct ucma_file *file2)
1056 {
1057 	if (file1 < file2) {
1058 		mutex_unlock(&file2->mut);
1059 		mutex_unlock(&file1->mut);
1060 	} else {
1061 		mutex_unlock(&file1->mut);
1062 		mutex_unlock(&file2->mut);
1063 	}
1064 }
1065 
1066 static void ucma_move_events(struct ucma_context *ctx, struct ucma_file *file)
1067 {
1068 	struct ucma_event *uevent, *tmp;
1069 
1070 	list_for_each_entry_safe(uevent, tmp, &ctx->file->event_list, list)
1071 		if (uevent->ctx == ctx)
1072 			list_move_tail(&uevent->list, &file->event_list);
1073 }
1074 
1075 static ssize_t ucma_migrate_id(struct ucma_file *new_file,
1076 			       const char __user *inbuf,
1077 			       int in_len, int out_len)
1078 {
1079 	struct rdma_ucm_migrate_id cmd;
1080 	struct rdma_ucm_migrate_resp resp;
1081 	struct ucma_context *ctx;
1082 	struct file *filp;
1083 	struct ucma_file *cur_file;
1084 	int ret = 0;
1085 
1086 	if (copy_from_user(&cmd, inbuf, sizeof(cmd)))
1087 		return -EFAULT;
1088 
1089 	/* Get current fd to protect against it being closed */
1090 	filp = fget(cmd.fd);
1091 	if (!filp)
1092 		return -ENOENT;
1093 
1094 	/* Validate current fd and prevent destruction of id. */
1095 	ctx = ucma_get_ctx(filp->private_data, cmd.id);
1096 	if (IS_ERR(ctx)) {
1097 		ret = PTR_ERR(ctx);
1098 		goto file_put;
1099 	}
1100 
1101 	cur_file = ctx->file;
1102 	if (cur_file == new_file) {
1103 		resp.events_reported = ctx->events_reported;
1104 		goto response;
1105 	}
1106 
1107 	/*
1108 	 * Migrate events between fd's, maintaining order, and avoiding new
1109 	 * events being added before existing events.
1110 	 */
1111 	ucma_lock_files(cur_file, new_file);
1112 	mutex_lock(&mut);
1113 
1114 	list_move_tail(&ctx->list, &new_file->ctx_list);
1115 	ucma_move_events(ctx, new_file);
1116 	ctx->file = new_file;
1117 	resp.events_reported = ctx->events_reported;
1118 
1119 	mutex_unlock(&mut);
1120 	ucma_unlock_files(cur_file, new_file);
1121 
1122 response:
1123 	if (copy_to_user((void __user *)(unsigned long)cmd.response,
1124 			 &resp, sizeof(resp)))
1125 		ret = -EFAULT;
1126 
1127 	ucma_put_ctx(ctx);
1128 file_put:
1129 	fput(filp);
1130 	return ret;
1131 }
1132 
1133 static ssize_t (*ucma_cmd_table[])(struct ucma_file *file,
1134 				   const char __user *inbuf,
1135 				   int in_len, int out_len) = {
1136 	[RDMA_USER_CM_CMD_CREATE_ID]	= ucma_create_id,
1137 	[RDMA_USER_CM_CMD_DESTROY_ID]	= ucma_destroy_id,
1138 	[RDMA_USER_CM_CMD_BIND_ADDR]	= ucma_bind_addr,
1139 	[RDMA_USER_CM_CMD_RESOLVE_ADDR]	= ucma_resolve_addr,
1140 	[RDMA_USER_CM_CMD_RESOLVE_ROUTE]= ucma_resolve_route,
1141 	[RDMA_USER_CM_CMD_QUERY_ROUTE]	= ucma_query_route,
1142 	[RDMA_USER_CM_CMD_CONNECT]	= ucma_connect,
1143 	[RDMA_USER_CM_CMD_LISTEN]	= ucma_listen,
1144 	[RDMA_USER_CM_CMD_ACCEPT]	= ucma_accept,
1145 	[RDMA_USER_CM_CMD_REJECT]	= ucma_reject,
1146 	[RDMA_USER_CM_CMD_DISCONNECT]	= ucma_disconnect,
1147 	[RDMA_USER_CM_CMD_INIT_QP_ATTR]	= ucma_init_qp_attr,
1148 	[RDMA_USER_CM_CMD_GET_EVENT]	= ucma_get_event,
1149 	[RDMA_USER_CM_CMD_GET_OPTION]	= NULL,
1150 	[RDMA_USER_CM_CMD_SET_OPTION]	= ucma_set_option,
1151 	[RDMA_USER_CM_CMD_NOTIFY]	= ucma_notify,
1152 	[RDMA_USER_CM_CMD_JOIN_MCAST]	= ucma_join_multicast,
1153 	[RDMA_USER_CM_CMD_LEAVE_MCAST]	= ucma_leave_multicast,
1154 	[RDMA_USER_CM_CMD_MIGRATE_ID]	= ucma_migrate_id
1155 };
1156 
1157 static ssize_t ucma_write(struct file *filp, const char __user *buf,
1158 			  size_t len, loff_t *pos)
1159 {
1160 	struct ucma_file *file = filp->private_data;
1161 	struct rdma_ucm_cmd_hdr hdr;
1162 	ssize_t ret;
1163 
1164 	if (len < sizeof(hdr))
1165 		return -EINVAL;
1166 
1167 	if (copy_from_user(&hdr, buf, sizeof(hdr)))
1168 		return -EFAULT;
1169 
1170 	if (hdr.cmd < 0 || hdr.cmd >= ARRAY_SIZE(ucma_cmd_table))
1171 		return -EINVAL;
1172 
1173 	if (hdr.in + sizeof(hdr) > len)
1174 		return -EINVAL;
1175 
1176 	if (!ucma_cmd_table[hdr.cmd])
1177 		return -ENOSYS;
1178 
1179 	ret = ucma_cmd_table[hdr.cmd](file, buf + sizeof(hdr), hdr.in, hdr.out);
1180 	if (!ret)
1181 		ret = len;
1182 
1183 	return ret;
1184 }
1185 
1186 static unsigned int ucma_poll(struct file *filp, struct poll_table_struct *wait)
1187 {
1188 	struct ucma_file *file = filp->private_data;
1189 	unsigned int mask = 0;
1190 
1191 	poll_wait(filp, &file->poll_wait, wait);
1192 
1193 	if (!list_empty(&file->event_list))
1194 		mask = POLLIN | POLLRDNORM;
1195 
1196 	return mask;
1197 }
1198 
1199 /*
1200  * ucma_open() does not need the BKL:
1201  *
1202  *  - no global state is referred to;
1203  *  - there is no ioctl method to race against;
1204  *  - no further module initialization is required for open to work
1205  *    after the device is registered.
1206  */
1207 static int ucma_open(struct inode *inode, struct file *filp)
1208 {
1209 	struct ucma_file *file;
1210 
1211 	file = kmalloc(sizeof *file, GFP_KERNEL);
1212 	if (!file)
1213 		return -ENOMEM;
1214 
1215 	INIT_LIST_HEAD(&file->event_list);
1216 	INIT_LIST_HEAD(&file->ctx_list);
1217 	init_waitqueue_head(&file->poll_wait);
1218 	mutex_init(&file->mut);
1219 
1220 	filp->private_data = file;
1221 	file->filp = filp;
1222 	return 0;
1223 }
1224 
1225 static int ucma_close(struct inode *inode, struct file *filp)
1226 {
1227 	struct ucma_file *file = filp->private_data;
1228 	struct ucma_context *ctx, *tmp;
1229 
1230 	mutex_lock(&file->mut);
1231 	list_for_each_entry_safe(ctx, tmp, &file->ctx_list, list) {
1232 		mutex_unlock(&file->mut);
1233 
1234 		mutex_lock(&mut);
1235 		idr_remove(&ctx_idr, ctx->id);
1236 		mutex_unlock(&mut);
1237 
1238 		ucma_free_ctx(ctx);
1239 		mutex_lock(&file->mut);
1240 	}
1241 	mutex_unlock(&file->mut);
1242 	kfree(file);
1243 	return 0;
1244 }
1245 
1246 static const struct file_operations ucma_fops = {
1247 	.owner 	 = THIS_MODULE,
1248 	.open 	 = ucma_open,
1249 	.release = ucma_close,
1250 	.write	 = ucma_write,
1251 	.poll    = ucma_poll,
1252 };
1253 
1254 static struct miscdevice ucma_misc = {
1255 	.minor	= MISC_DYNAMIC_MINOR,
1256 	.name	= "rdma_cm",
1257 	.fops	= &ucma_fops,
1258 };
1259 
1260 static ssize_t show_abi_version(struct device *dev,
1261 				struct device_attribute *attr,
1262 				char *buf)
1263 {
1264 	return sprintf(buf, "%d\n", RDMA_USER_CM_ABI_VERSION);
1265 }
1266 static DEVICE_ATTR(abi_version, S_IRUGO, show_abi_version, NULL);
1267 
1268 static int __init ucma_init(void)
1269 {
1270 	int ret;
1271 
1272 	ret = misc_register(&ucma_misc);
1273 	if (ret)
1274 		return ret;
1275 
1276 	ret = device_create_file(ucma_misc.this_device, &dev_attr_abi_version);
1277 	if (ret) {
1278 		printk(KERN_ERR "rdma_ucm: couldn't create abi_version attr\n");
1279 		goto err;
1280 	}
1281 	return 0;
1282 err:
1283 	misc_deregister(&ucma_misc);
1284 	return ret;
1285 }
1286 
1287 static void __exit ucma_cleanup(void)
1288 {
1289 	device_remove_file(ucma_misc.this_device, &dev_attr_abi_version);
1290 	misc_deregister(&ucma_misc);
1291 	idr_destroy(&ctx_idr);
1292 }
1293 
1294 module_init(ucma_init);
1295 module_exit(ucma_cleanup);
1296