1 /*
2  * Copyright (c) 2005 Topspin Communications.  All rights reserved.
3  * Copyright (c) 2005, 2006 Cisco Systems.  All rights reserved.
4  * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
5  * Copyright (c) 2005 Voltaire, Inc. All rights reserved.
6  * Copyright (c) 2005 PathScale, Inc. All rights reserved.
7  *
8  * This software is available to you under a choice of one of two
9  * licenses.  You may choose to be licensed under the terms of the GNU
10  * General Public License (GPL) Version 2, available from the file
11  * COPYING in the main directory of this source tree, or the
12  * OpenIB.org BSD license below:
13  *
14  *     Redistribution and use in source and binary forms, with or
15  *     without modification, are permitted provided that the following
16  *     conditions are met:
17  *
18  *      - Redistributions of source code must retain the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer.
21  *
22  *      - Redistributions in binary form must reproduce the above
23  *        copyright notice, this list of conditions and the following
24  *        disclaimer in the documentation and/or other materials
25  *        provided with the distribution.
26  *
27  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
28  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
29  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
30  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
31  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
32  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
33  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
34  * SOFTWARE.
35  *
36  * $Id: uverbs_main.c 2733 2005-06-28 19:14:34Z roland $
37  */
38 
39 #include <linux/module.h>
40 #include <linux/init.h>
41 #include <linux/device.h>
42 #include <linux/err.h>
43 #include <linux/fs.h>
44 #include <linux/poll.h>
45 #include <linux/file.h>
46 #include <linux/mount.h>
47 #include <linux/cdev.h>
48 
49 #include <asm/uaccess.h>
50 
51 #include "uverbs.h"
52 
53 MODULE_AUTHOR("Roland Dreier");
54 MODULE_DESCRIPTION("InfiniBand userspace verbs access");
55 MODULE_LICENSE("Dual BSD/GPL");
56 
57 #define INFINIBANDEVENTFS_MAGIC	0x49426576	/* "IBev" */
58 
59 enum {
60 	IB_UVERBS_MAJOR       = 231,
61 	IB_UVERBS_BASE_MINOR  = 192,
62 	IB_UVERBS_MAX_DEVICES = 32
63 };
64 
65 #define IB_UVERBS_BASE_DEV	MKDEV(IB_UVERBS_MAJOR, IB_UVERBS_BASE_MINOR)
66 
67 static struct class *uverbs_class;
68 
69 DEFINE_SPINLOCK(ib_uverbs_idr_lock);
70 DEFINE_IDR(ib_uverbs_pd_idr);
71 DEFINE_IDR(ib_uverbs_mr_idr);
72 DEFINE_IDR(ib_uverbs_mw_idr);
73 DEFINE_IDR(ib_uverbs_ah_idr);
74 DEFINE_IDR(ib_uverbs_cq_idr);
75 DEFINE_IDR(ib_uverbs_qp_idr);
76 DEFINE_IDR(ib_uverbs_srq_idr);
77 
78 static spinlock_t map_lock;
79 static struct ib_uverbs_device *dev_table[IB_UVERBS_MAX_DEVICES];
80 static DECLARE_BITMAP(dev_map, IB_UVERBS_MAX_DEVICES);
81 
82 static ssize_t (*uverbs_cmd_table[])(struct ib_uverbs_file *file,
83 				     const char __user *buf, int in_len,
84 				     int out_len) = {
85 	[IB_USER_VERBS_CMD_GET_CONTEXT]   	= ib_uverbs_get_context,
86 	[IB_USER_VERBS_CMD_QUERY_DEVICE]  	= ib_uverbs_query_device,
87 	[IB_USER_VERBS_CMD_QUERY_PORT]    	= ib_uverbs_query_port,
88 	[IB_USER_VERBS_CMD_ALLOC_PD]      	= ib_uverbs_alloc_pd,
89 	[IB_USER_VERBS_CMD_DEALLOC_PD]    	= ib_uverbs_dealloc_pd,
90 	[IB_USER_VERBS_CMD_REG_MR]        	= ib_uverbs_reg_mr,
91 	[IB_USER_VERBS_CMD_DEREG_MR]      	= ib_uverbs_dereg_mr,
92 	[IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL] = ib_uverbs_create_comp_channel,
93 	[IB_USER_VERBS_CMD_CREATE_CQ]     	= ib_uverbs_create_cq,
94 	[IB_USER_VERBS_CMD_RESIZE_CQ]     	= ib_uverbs_resize_cq,
95 	[IB_USER_VERBS_CMD_POLL_CQ]     	= ib_uverbs_poll_cq,
96 	[IB_USER_VERBS_CMD_REQ_NOTIFY_CQ]     	= ib_uverbs_req_notify_cq,
97 	[IB_USER_VERBS_CMD_DESTROY_CQ]    	= ib_uverbs_destroy_cq,
98 	[IB_USER_VERBS_CMD_CREATE_QP]     	= ib_uverbs_create_qp,
99 	[IB_USER_VERBS_CMD_QUERY_QP]     	= ib_uverbs_query_qp,
100 	[IB_USER_VERBS_CMD_MODIFY_QP]     	= ib_uverbs_modify_qp,
101 	[IB_USER_VERBS_CMD_DESTROY_QP]    	= ib_uverbs_destroy_qp,
102 	[IB_USER_VERBS_CMD_POST_SEND]    	= ib_uverbs_post_send,
103 	[IB_USER_VERBS_CMD_POST_RECV]    	= ib_uverbs_post_recv,
104 	[IB_USER_VERBS_CMD_POST_SRQ_RECV]    	= ib_uverbs_post_srq_recv,
105 	[IB_USER_VERBS_CMD_CREATE_AH]    	= ib_uverbs_create_ah,
106 	[IB_USER_VERBS_CMD_DESTROY_AH]    	= ib_uverbs_destroy_ah,
107 	[IB_USER_VERBS_CMD_ATTACH_MCAST]  	= ib_uverbs_attach_mcast,
108 	[IB_USER_VERBS_CMD_DETACH_MCAST]  	= ib_uverbs_detach_mcast,
109 	[IB_USER_VERBS_CMD_CREATE_SRQ]    	= ib_uverbs_create_srq,
110 	[IB_USER_VERBS_CMD_MODIFY_SRQ]    	= ib_uverbs_modify_srq,
111 	[IB_USER_VERBS_CMD_QUERY_SRQ]     	= ib_uverbs_query_srq,
112 	[IB_USER_VERBS_CMD_DESTROY_SRQ]   	= ib_uverbs_destroy_srq,
113 };
114 
115 static struct vfsmount *uverbs_event_mnt;
116 
117 static void ib_uverbs_add_one(struct ib_device *device);
118 static void ib_uverbs_remove_one(struct ib_device *device);
119 
120 static void ib_uverbs_release_dev(struct kref *ref)
121 {
122 	struct ib_uverbs_device *dev =
123 		container_of(ref, struct ib_uverbs_device, ref);
124 
125 	complete(&dev->comp);
126 }
127 
128 void ib_uverbs_release_ucq(struct ib_uverbs_file *file,
129 			  struct ib_uverbs_event_file *ev_file,
130 			  struct ib_ucq_object *uobj)
131 {
132 	struct ib_uverbs_event *evt, *tmp;
133 
134 	if (ev_file) {
135 		spin_lock_irq(&ev_file->lock);
136 		list_for_each_entry_safe(evt, tmp, &uobj->comp_list, obj_list) {
137 			list_del(&evt->list);
138 			kfree(evt);
139 		}
140 		spin_unlock_irq(&ev_file->lock);
141 
142 		kref_put(&ev_file->ref, ib_uverbs_release_event_file);
143 	}
144 
145 	spin_lock_irq(&file->async_file->lock);
146 	list_for_each_entry_safe(evt, tmp, &uobj->async_list, obj_list) {
147 		list_del(&evt->list);
148 		kfree(evt);
149 	}
150 	spin_unlock_irq(&file->async_file->lock);
151 }
152 
153 void ib_uverbs_release_uevent(struct ib_uverbs_file *file,
154 			      struct ib_uevent_object *uobj)
155 {
156 	struct ib_uverbs_event *evt, *tmp;
157 
158 	spin_lock_irq(&file->async_file->lock);
159 	list_for_each_entry_safe(evt, tmp, &uobj->event_list, obj_list) {
160 		list_del(&evt->list);
161 		kfree(evt);
162 	}
163 	spin_unlock_irq(&file->async_file->lock);
164 }
165 
166 static void ib_uverbs_detach_umcast(struct ib_qp *qp,
167 				    struct ib_uqp_object *uobj)
168 {
169 	struct ib_uverbs_mcast_entry *mcast, *tmp;
170 
171 	list_for_each_entry_safe(mcast, tmp, &uobj->mcast_list, list) {
172 		ib_detach_mcast(qp, &mcast->gid, mcast->lid);
173 		list_del(&mcast->list);
174 		kfree(mcast);
175 	}
176 }
177 
178 static int ib_uverbs_cleanup_ucontext(struct ib_uverbs_file *file,
179 				      struct ib_ucontext *context)
180 {
181 	struct ib_uobject *uobj, *tmp;
182 
183 	if (!context)
184 		return 0;
185 
186 	context->closing = 1;
187 
188 	list_for_each_entry_safe(uobj, tmp, &context->ah_list, list) {
189 		struct ib_ah *ah = uobj->object;
190 
191 		idr_remove_uobj(&ib_uverbs_ah_idr, uobj);
192 		ib_destroy_ah(ah);
193 		kfree(uobj);
194 	}
195 
196 	list_for_each_entry_safe(uobj, tmp, &context->qp_list, list) {
197 		struct ib_qp *qp = uobj->object;
198 		struct ib_uqp_object *uqp =
199 			container_of(uobj, struct ib_uqp_object, uevent.uobject);
200 
201 		idr_remove_uobj(&ib_uverbs_qp_idr, uobj);
202 		ib_uverbs_detach_umcast(qp, uqp);
203 		ib_destroy_qp(qp);
204 		ib_uverbs_release_uevent(file, &uqp->uevent);
205 		kfree(uqp);
206 	}
207 
208 	list_for_each_entry_safe(uobj, tmp, &context->cq_list, list) {
209 		struct ib_cq *cq = uobj->object;
210 		struct ib_uverbs_event_file *ev_file = cq->cq_context;
211 		struct ib_ucq_object *ucq =
212 			container_of(uobj, struct ib_ucq_object, uobject);
213 
214 		idr_remove_uobj(&ib_uverbs_cq_idr, uobj);
215 		ib_destroy_cq(cq);
216 		ib_uverbs_release_ucq(file, ev_file, ucq);
217 		kfree(ucq);
218 	}
219 
220 	list_for_each_entry_safe(uobj, tmp, &context->srq_list, list) {
221 		struct ib_srq *srq = uobj->object;
222 		struct ib_uevent_object *uevent =
223 			container_of(uobj, struct ib_uevent_object, uobject);
224 
225 		idr_remove_uobj(&ib_uverbs_srq_idr, uobj);
226 		ib_destroy_srq(srq);
227 		ib_uverbs_release_uevent(file, uevent);
228 		kfree(uevent);
229 	}
230 
231 	/* XXX Free MWs */
232 
233 	list_for_each_entry_safe(uobj, tmp, &context->mr_list, list) {
234 		struct ib_mr *mr = uobj->object;
235 
236 		idr_remove_uobj(&ib_uverbs_mr_idr, uobj);
237 		ib_dereg_mr(mr);
238 		kfree(uobj);
239 	}
240 
241 	list_for_each_entry_safe(uobj, tmp, &context->pd_list, list) {
242 		struct ib_pd *pd = uobj->object;
243 
244 		idr_remove_uobj(&ib_uverbs_pd_idr, uobj);
245 		ib_dealloc_pd(pd);
246 		kfree(uobj);
247 	}
248 
249 	return context->device->dealloc_ucontext(context);
250 }
251 
252 static void ib_uverbs_release_file(struct kref *ref)
253 {
254 	struct ib_uverbs_file *file =
255 		container_of(ref, struct ib_uverbs_file, ref);
256 
257 	module_put(file->device->ib_dev->owner);
258 	kref_put(&file->device->ref, ib_uverbs_release_dev);
259 
260 	kfree(file);
261 }
262 
263 static ssize_t ib_uverbs_event_read(struct file *filp, char __user *buf,
264 				    size_t count, loff_t *pos)
265 {
266 	struct ib_uverbs_event_file *file = filp->private_data;
267 	struct ib_uverbs_event *event;
268 	int eventsz;
269 	int ret = 0;
270 
271 	spin_lock_irq(&file->lock);
272 
273 	while (list_empty(&file->event_list)) {
274 		spin_unlock_irq(&file->lock);
275 
276 		if (filp->f_flags & O_NONBLOCK)
277 			return -EAGAIN;
278 
279 		if (wait_event_interruptible(file->poll_wait,
280 					     !list_empty(&file->event_list)))
281 			return -ERESTARTSYS;
282 
283 		spin_lock_irq(&file->lock);
284 	}
285 
286 	event = list_entry(file->event_list.next, struct ib_uverbs_event, list);
287 
288 	if (file->is_async)
289 		eventsz = sizeof (struct ib_uverbs_async_event_desc);
290 	else
291 		eventsz = sizeof (struct ib_uverbs_comp_event_desc);
292 
293 	if (eventsz > count) {
294 		ret   = -EINVAL;
295 		event = NULL;
296 	} else {
297 		list_del(file->event_list.next);
298 		if (event->counter) {
299 			++(*event->counter);
300 			list_del(&event->obj_list);
301 		}
302 	}
303 
304 	spin_unlock_irq(&file->lock);
305 
306 	if (event) {
307 		if (copy_to_user(buf, event, eventsz))
308 			ret = -EFAULT;
309 		else
310 			ret = eventsz;
311 	}
312 
313 	kfree(event);
314 
315 	return ret;
316 }
317 
318 static unsigned int ib_uverbs_event_poll(struct file *filp,
319 					 struct poll_table_struct *wait)
320 {
321 	unsigned int pollflags = 0;
322 	struct ib_uverbs_event_file *file = filp->private_data;
323 
324 	poll_wait(filp, &file->poll_wait, wait);
325 
326 	spin_lock_irq(&file->lock);
327 	if (!list_empty(&file->event_list))
328 		pollflags = POLLIN | POLLRDNORM;
329 	spin_unlock_irq(&file->lock);
330 
331 	return pollflags;
332 }
333 
334 void ib_uverbs_release_event_file(struct kref *ref)
335 {
336 	struct ib_uverbs_event_file *file =
337 		container_of(ref, struct ib_uverbs_event_file, ref);
338 
339 	kfree(file);
340 }
341 
342 static int ib_uverbs_event_fasync(int fd, struct file *filp, int on)
343 {
344 	struct ib_uverbs_event_file *file = filp->private_data;
345 
346 	return fasync_helper(fd, filp, on, &file->async_queue);
347 }
348 
349 static int ib_uverbs_event_close(struct inode *inode, struct file *filp)
350 {
351 	struct ib_uverbs_event_file *file = filp->private_data;
352 	struct ib_uverbs_event *entry, *tmp;
353 
354 	spin_lock_irq(&file->lock);
355 	file->file = NULL;
356 	list_for_each_entry_safe(entry, tmp, &file->event_list, list) {
357 		if (entry->counter)
358 			list_del(&entry->obj_list);
359 		kfree(entry);
360 	}
361 	spin_unlock_irq(&file->lock);
362 
363 	ib_uverbs_event_fasync(-1, filp, 0);
364 
365 	if (file->is_async) {
366 		ib_unregister_event_handler(&file->uverbs_file->event_handler);
367 		kref_put(&file->uverbs_file->ref, ib_uverbs_release_file);
368 	}
369 	kref_put(&file->ref, ib_uverbs_release_event_file);
370 
371 	return 0;
372 }
373 
374 static const struct file_operations uverbs_event_fops = {
375 	.owner	 = THIS_MODULE,
376 	.read 	 = ib_uverbs_event_read,
377 	.poll    = ib_uverbs_event_poll,
378 	.release = ib_uverbs_event_close,
379 	.fasync  = ib_uverbs_event_fasync
380 };
381 
382 void ib_uverbs_comp_handler(struct ib_cq *cq, void *cq_context)
383 {
384 	struct ib_uverbs_event_file    *file = cq_context;
385 	struct ib_ucq_object	       *uobj;
386 	struct ib_uverbs_event	       *entry;
387 	unsigned long			flags;
388 
389 	if (!file)
390 		return;
391 
392 	spin_lock_irqsave(&file->lock, flags);
393 	if (!file->file) {
394 		spin_unlock_irqrestore(&file->lock, flags);
395 		return;
396 	}
397 
398 	entry = kmalloc(sizeof *entry, GFP_ATOMIC);
399 	if (!entry) {
400 		spin_unlock_irqrestore(&file->lock, flags);
401 		return;
402 	}
403 
404 	uobj = container_of(cq->uobject, struct ib_ucq_object, uobject);
405 
406 	entry->desc.comp.cq_handle = cq->uobject->user_handle;
407 	entry->counter		   = &uobj->comp_events_reported;
408 
409 	list_add_tail(&entry->list, &file->event_list);
410 	list_add_tail(&entry->obj_list, &uobj->comp_list);
411 	spin_unlock_irqrestore(&file->lock, flags);
412 
413 	wake_up_interruptible(&file->poll_wait);
414 	kill_fasync(&file->async_queue, SIGIO, POLL_IN);
415 }
416 
417 static void ib_uverbs_async_handler(struct ib_uverbs_file *file,
418 				    __u64 element, __u64 event,
419 				    struct list_head *obj_list,
420 				    u32 *counter)
421 {
422 	struct ib_uverbs_event *entry;
423 	unsigned long flags;
424 
425 	spin_lock_irqsave(&file->async_file->lock, flags);
426 	if (!file->async_file->file) {
427 		spin_unlock_irqrestore(&file->async_file->lock, flags);
428 		return;
429 	}
430 
431 	entry = kmalloc(sizeof *entry, GFP_ATOMIC);
432 	if (!entry) {
433 		spin_unlock_irqrestore(&file->async_file->lock, flags);
434 		return;
435 	}
436 
437 	entry->desc.async.element    = element;
438 	entry->desc.async.event_type = event;
439 	entry->counter               = counter;
440 
441 	list_add_tail(&entry->list, &file->async_file->event_list);
442 	if (obj_list)
443 		list_add_tail(&entry->obj_list, obj_list);
444 	spin_unlock_irqrestore(&file->async_file->lock, flags);
445 
446 	wake_up_interruptible(&file->async_file->poll_wait);
447 	kill_fasync(&file->async_file->async_queue, SIGIO, POLL_IN);
448 }
449 
450 void ib_uverbs_cq_event_handler(struct ib_event *event, void *context_ptr)
451 {
452 	struct ib_ucq_object *uobj = container_of(event->element.cq->uobject,
453 						  struct ib_ucq_object, uobject);
454 
455 	ib_uverbs_async_handler(uobj->uverbs_file, uobj->uobject.user_handle,
456 				event->event, &uobj->async_list,
457 				&uobj->async_events_reported);
458 }
459 
460 void ib_uverbs_qp_event_handler(struct ib_event *event, void *context_ptr)
461 {
462 	struct ib_uevent_object *uobj;
463 
464 	uobj = container_of(event->element.qp->uobject,
465 			    struct ib_uevent_object, uobject);
466 
467 	ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
468 				event->event, &uobj->event_list,
469 				&uobj->events_reported);
470 }
471 
472 void ib_uverbs_srq_event_handler(struct ib_event *event, void *context_ptr)
473 {
474 	struct ib_uevent_object *uobj;
475 
476 	uobj = container_of(event->element.srq->uobject,
477 			    struct ib_uevent_object, uobject);
478 
479 	ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
480 				event->event, &uobj->event_list,
481 				&uobj->events_reported);
482 }
483 
484 void ib_uverbs_event_handler(struct ib_event_handler *handler,
485 			     struct ib_event *event)
486 {
487 	struct ib_uverbs_file *file =
488 		container_of(handler, struct ib_uverbs_file, event_handler);
489 
490 	ib_uverbs_async_handler(file, event->element.port_num, event->event,
491 				NULL, NULL);
492 }
493 
494 struct file *ib_uverbs_alloc_event_file(struct ib_uverbs_file *uverbs_file,
495 					int is_async, int *fd)
496 {
497 	struct ib_uverbs_event_file *ev_file;
498 	struct file *filp;
499 	int ret;
500 
501 	ev_file = kmalloc(sizeof *ev_file, GFP_KERNEL);
502 	if (!ev_file)
503 		return ERR_PTR(-ENOMEM);
504 
505 	kref_init(&ev_file->ref);
506 	spin_lock_init(&ev_file->lock);
507 	INIT_LIST_HEAD(&ev_file->event_list);
508 	init_waitqueue_head(&ev_file->poll_wait);
509 	ev_file->uverbs_file = uverbs_file;
510 	ev_file->async_queue = NULL;
511 	ev_file->is_async    = is_async;
512 
513 	*fd = get_unused_fd();
514 	if (*fd < 0) {
515 		ret = *fd;
516 		goto err;
517 	}
518 
519 	filp = get_empty_filp();
520 	if (!filp) {
521 		ret = -ENFILE;
522 		goto err_fd;
523 	}
524 
525 	ev_file->file      = filp;
526 
527 	/*
528 	 * fops_get() can't fail here, because we're coming from a
529 	 * system call on a uverbs file, which will already have a
530 	 * module reference.
531 	 */
532 	filp->f_op 	   = fops_get(&uverbs_event_fops);
533 	filp->f_path.mnt 	   = mntget(uverbs_event_mnt);
534 	filp->f_path.dentry 	   = dget(uverbs_event_mnt->mnt_root);
535 	filp->f_mapping    = filp->f_path.dentry->d_inode->i_mapping;
536 	filp->f_flags      = O_RDONLY;
537 	filp->f_mode       = FMODE_READ;
538 	filp->private_data = ev_file;
539 
540 	return filp;
541 
542 err_fd:
543 	put_unused_fd(*fd);
544 
545 err:
546 	kfree(ev_file);
547 	return ERR_PTR(ret);
548 }
549 
550 /*
551  * Look up a completion event file by FD.  If lookup is successful,
552  * takes a ref to the event file struct that it returns; if
553  * unsuccessful, returns NULL.
554  */
555 struct ib_uverbs_event_file *ib_uverbs_lookup_comp_file(int fd)
556 {
557 	struct ib_uverbs_event_file *ev_file = NULL;
558 	struct file *filp;
559 
560 	filp = fget(fd);
561 	if (!filp)
562 		return NULL;
563 
564 	if (filp->f_op != &uverbs_event_fops)
565 		goto out;
566 
567 	ev_file = filp->private_data;
568 	if (ev_file->is_async) {
569 		ev_file = NULL;
570 		goto out;
571 	}
572 
573 	kref_get(&ev_file->ref);
574 
575 out:
576 	fput(filp);
577 	return ev_file;
578 }
579 
580 static ssize_t ib_uverbs_write(struct file *filp, const char __user *buf,
581 			     size_t count, loff_t *pos)
582 {
583 	struct ib_uverbs_file *file = filp->private_data;
584 	struct ib_uverbs_cmd_hdr hdr;
585 
586 	if (count < sizeof hdr)
587 		return -EINVAL;
588 
589 	if (copy_from_user(&hdr, buf, sizeof hdr))
590 		return -EFAULT;
591 
592 	if (hdr.in_words * 4 != count)
593 		return -EINVAL;
594 
595 	if (hdr.command < 0				||
596 	    hdr.command >= ARRAY_SIZE(uverbs_cmd_table) ||
597 	    !uverbs_cmd_table[hdr.command]		||
598 	    !(file->device->ib_dev->uverbs_cmd_mask & (1ull << hdr.command)))
599 		return -EINVAL;
600 
601 	if (!file->ucontext &&
602 	    hdr.command != IB_USER_VERBS_CMD_GET_CONTEXT)
603 		return -EINVAL;
604 
605 	return uverbs_cmd_table[hdr.command](file, buf + sizeof hdr,
606 					     hdr.in_words * 4, hdr.out_words * 4);
607 }
608 
609 static int ib_uverbs_mmap(struct file *filp, struct vm_area_struct *vma)
610 {
611 	struct ib_uverbs_file *file = filp->private_data;
612 
613 	if (!file->ucontext)
614 		return -ENODEV;
615 	else
616 		return file->device->ib_dev->mmap(file->ucontext, vma);
617 }
618 
619 static int ib_uverbs_open(struct inode *inode, struct file *filp)
620 {
621 	struct ib_uverbs_device *dev;
622 	struct ib_uverbs_file *file;
623 	int ret;
624 
625 	spin_lock(&map_lock);
626 	dev = dev_table[iminor(inode) - IB_UVERBS_BASE_MINOR];
627 	if (dev)
628 		kref_get(&dev->ref);
629 	spin_unlock(&map_lock);
630 
631 	if (!dev)
632 		return -ENXIO;
633 
634 	if (!try_module_get(dev->ib_dev->owner)) {
635 		ret = -ENODEV;
636 		goto err;
637 	}
638 
639 	file = kmalloc(sizeof *file, GFP_KERNEL);
640 	if (!file) {
641 		ret = -ENOMEM;
642 		goto err_module;
643 	}
644 
645 	file->device	 = dev;
646 	file->ucontext	 = NULL;
647 	file->async_file = NULL;
648 	kref_init(&file->ref);
649 	mutex_init(&file->mutex);
650 
651 	filp->private_data = file;
652 
653 	return 0;
654 
655 err_module:
656 	module_put(dev->ib_dev->owner);
657 
658 err:
659 	kref_put(&dev->ref, ib_uverbs_release_dev);
660 
661 	return ret;
662 }
663 
664 static int ib_uverbs_close(struct inode *inode, struct file *filp)
665 {
666 	struct ib_uverbs_file *file = filp->private_data;
667 
668 	ib_uverbs_cleanup_ucontext(file, file->ucontext);
669 
670 	if (file->async_file)
671 		kref_put(&file->async_file->ref, ib_uverbs_release_event_file);
672 
673 	kref_put(&file->ref, ib_uverbs_release_file);
674 
675 	return 0;
676 }
677 
678 static const struct file_operations uverbs_fops = {
679 	.owner 	 = THIS_MODULE,
680 	.write 	 = ib_uverbs_write,
681 	.open 	 = ib_uverbs_open,
682 	.release = ib_uverbs_close
683 };
684 
685 static const struct file_operations uverbs_mmap_fops = {
686 	.owner 	 = THIS_MODULE,
687 	.write 	 = ib_uverbs_write,
688 	.mmap    = ib_uverbs_mmap,
689 	.open 	 = ib_uverbs_open,
690 	.release = ib_uverbs_close
691 };
692 
693 static struct ib_client uverbs_client = {
694 	.name   = "uverbs",
695 	.add    = ib_uverbs_add_one,
696 	.remove = ib_uverbs_remove_one
697 };
698 
699 static ssize_t show_ibdev(struct class_device *class_dev, char *buf)
700 {
701 	struct ib_uverbs_device *dev = class_get_devdata(class_dev);
702 
703 	if (!dev)
704 		return -ENODEV;
705 
706 	return sprintf(buf, "%s\n", dev->ib_dev->name);
707 }
708 static CLASS_DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
709 
710 static ssize_t show_dev_abi_version(struct class_device *class_dev, char *buf)
711 {
712 	struct ib_uverbs_device *dev = class_get_devdata(class_dev);
713 
714 	if (!dev)
715 		return -ENODEV;
716 
717 	return sprintf(buf, "%d\n", dev->ib_dev->uverbs_abi_ver);
718 }
719 static CLASS_DEVICE_ATTR(abi_version, S_IRUGO, show_dev_abi_version, NULL);
720 
721 static ssize_t show_abi_version(struct class *class, char *buf)
722 {
723 	return sprintf(buf, "%d\n", IB_USER_VERBS_ABI_VERSION);
724 }
725 static CLASS_ATTR(abi_version, S_IRUGO, show_abi_version, NULL);
726 
727 static void ib_uverbs_add_one(struct ib_device *device)
728 {
729 	struct ib_uverbs_device *uverbs_dev;
730 
731 	if (!device->alloc_ucontext)
732 		return;
733 
734 	uverbs_dev = kzalloc(sizeof *uverbs_dev, GFP_KERNEL);
735 	if (!uverbs_dev)
736 		return;
737 
738 	kref_init(&uverbs_dev->ref);
739 	init_completion(&uverbs_dev->comp);
740 
741 	spin_lock(&map_lock);
742 	uverbs_dev->devnum = find_first_zero_bit(dev_map, IB_UVERBS_MAX_DEVICES);
743 	if (uverbs_dev->devnum >= IB_UVERBS_MAX_DEVICES) {
744 		spin_unlock(&map_lock);
745 		goto err;
746 	}
747 	set_bit(uverbs_dev->devnum, dev_map);
748 	spin_unlock(&map_lock);
749 
750 	uverbs_dev->ib_dev           = device;
751 	uverbs_dev->num_comp_vectors = device->num_comp_vectors;
752 
753 	uverbs_dev->dev = cdev_alloc();
754 	if (!uverbs_dev->dev)
755 		goto err;
756 	uverbs_dev->dev->owner = THIS_MODULE;
757 	uverbs_dev->dev->ops = device->mmap ? &uverbs_mmap_fops : &uverbs_fops;
758 	kobject_set_name(&uverbs_dev->dev->kobj, "uverbs%d", uverbs_dev->devnum);
759 	if (cdev_add(uverbs_dev->dev, IB_UVERBS_BASE_DEV + uverbs_dev->devnum, 1))
760 		goto err_cdev;
761 
762 	uverbs_dev->class_dev = class_device_create(uverbs_class, NULL,
763 						    uverbs_dev->dev->dev,
764 						    device->dma_device,
765 						    "uverbs%d", uverbs_dev->devnum);
766 	if (IS_ERR(uverbs_dev->class_dev))
767 		goto err_cdev;
768 
769 	class_set_devdata(uverbs_dev->class_dev, uverbs_dev);
770 
771 	if (class_device_create_file(uverbs_dev->class_dev, &class_device_attr_ibdev))
772 		goto err_class;
773 	if (class_device_create_file(uverbs_dev->class_dev, &class_device_attr_abi_version))
774 		goto err_class;
775 
776 	spin_lock(&map_lock);
777 	dev_table[uverbs_dev->devnum] = uverbs_dev;
778 	spin_unlock(&map_lock);
779 
780 	ib_set_client_data(device, &uverbs_client, uverbs_dev);
781 
782 	return;
783 
784 err_class:
785 	class_device_destroy(uverbs_class, uverbs_dev->dev->dev);
786 
787 err_cdev:
788 	cdev_del(uverbs_dev->dev);
789 	clear_bit(uverbs_dev->devnum, dev_map);
790 
791 err:
792 	kref_put(&uverbs_dev->ref, ib_uverbs_release_dev);
793 	wait_for_completion(&uverbs_dev->comp);
794 	kfree(uverbs_dev);
795 	return;
796 }
797 
798 static void ib_uverbs_remove_one(struct ib_device *device)
799 {
800 	struct ib_uverbs_device *uverbs_dev = ib_get_client_data(device, &uverbs_client);
801 
802 	if (!uverbs_dev)
803 		return;
804 
805 	class_set_devdata(uverbs_dev->class_dev, NULL);
806 	class_device_destroy(uverbs_class, uverbs_dev->dev->dev);
807 	cdev_del(uverbs_dev->dev);
808 
809 	spin_lock(&map_lock);
810 	dev_table[uverbs_dev->devnum] = NULL;
811 	spin_unlock(&map_lock);
812 
813 	clear_bit(uverbs_dev->devnum, dev_map);
814 
815 	kref_put(&uverbs_dev->ref, ib_uverbs_release_dev);
816 	wait_for_completion(&uverbs_dev->comp);
817 	kfree(uverbs_dev);
818 }
819 
820 static int uverbs_event_get_sb(struct file_system_type *fs_type, int flags,
821 			       const char *dev_name, void *data,
822 			       struct vfsmount *mnt)
823 {
824 	return get_sb_pseudo(fs_type, "infinibandevent:", NULL,
825 			     INFINIBANDEVENTFS_MAGIC, mnt);
826 }
827 
828 static struct file_system_type uverbs_event_fs = {
829 	/* No owner field so module can be unloaded */
830 	.name    = "infinibandeventfs",
831 	.get_sb  = uverbs_event_get_sb,
832 	.kill_sb = kill_litter_super
833 };
834 
835 static int __init ib_uverbs_init(void)
836 {
837 	int ret;
838 
839 	spin_lock_init(&map_lock);
840 
841 	ret = register_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES,
842 				     "infiniband_verbs");
843 	if (ret) {
844 		printk(KERN_ERR "user_verbs: couldn't register device number\n");
845 		goto out;
846 	}
847 
848 	uverbs_class = class_create(THIS_MODULE, "infiniband_verbs");
849 	if (IS_ERR(uverbs_class)) {
850 		ret = PTR_ERR(uverbs_class);
851 		printk(KERN_ERR "user_verbs: couldn't create class infiniband_verbs\n");
852 		goto out_chrdev;
853 	}
854 
855 	ret = class_create_file(uverbs_class, &class_attr_abi_version);
856 	if (ret) {
857 		printk(KERN_ERR "user_verbs: couldn't create abi_version attribute\n");
858 		goto out_class;
859 	}
860 
861 	ret = register_filesystem(&uverbs_event_fs);
862 	if (ret) {
863 		printk(KERN_ERR "user_verbs: couldn't register infinibandeventfs\n");
864 		goto out_class;
865 	}
866 
867 	uverbs_event_mnt = kern_mount(&uverbs_event_fs);
868 	if (IS_ERR(uverbs_event_mnt)) {
869 		ret = PTR_ERR(uverbs_event_mnt);
870 		printk(KERN_ERR "user_verbs: couldn't mount infinibandeventfs\n");
871 		goto out_fs;
872 	}
873 
874 	ret = ib_register_client(&uverbs_client);
875 	if (ret) {
876 		printk(KERN_ERR "user_verbs: couldn't register client\n");
877 		goto out_mnt;
878 	}
879 
880 	return 0;
881 
882 out_mnt:
883 	mntput(uverbs_event_mnt);
884 
885 out_fs:
886 	unregister_filesystem(&uverbs_event_fs);
887 
888 out_class:
889 	class_destroy(uverbs_class);
890 
891 out_chrdev:
892 	unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
893 
894 out:
895 	return ret;
896 }
897 
898 static void __exit ib_uverbs_cleanup(void)
899 {
900 	ib_unregister_client(&uverbs_client);
901 	mntput(uverbs_event_mnt);
902 	unregister_filesystem(&uverbs_event_fs);
903 	class_destroy(uverbs_class);
904 	unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
905 	idr_destroy(&ib_uverbs_pd_idr);
906 	idr_destroy(&ib_uverbs_mr_idr);
907 	idr_destroy(&ib_uverbs_mw_idr);
908 	idr_destroy(&ib_uverbs_ah_idr);
909 	idr_destroy(&ib_uverbs_cq_idr);
910 	idr_destroy(&ib_uverbs_qp_idr);
911 	idr_destroy(&ib_uverbs_srq_idr);
912 }
913 
914 module_init(ib_uverbs_init);
915 module_exit(ib_uverbs_cleanup);
916