xref: /linux/drivers/infiniband/core/uverbs_main.c (revision 2c684d892bb2ee31cc48f4a8b91e86a0f15e82f9)
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 
37 #include <linux/module.h>
38 #include <linux/init.h>
39 #include <linux/device.h>
40 #include <linux/err.h>
41 #include <linux/fs.h>
42 #include <linux/poll.h>
43 #include <linux/sched.h>
44 #include <linux/file.h>
45 #include <linux/cdev.h>
46 #include <linux/anon_inodes.h>
47 #include <linux/slab.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 enum {
58 	IB_UVERBS_MAJOR       = 231,
59 	IB_UVERBS_BASE_MINOR  = 192,
60 	IB_UVERBS_MAX_DEVICES = 32
61 };
62 
63 #define IB_UVERBS_BASE_DEV	MKDEV(IB_UVERBS_MAJOR, IB_UVERBS_BASE_MINOR)
64 
65 static struct class *uverbs_class;
66 
67 DEFINE_SPINLOCK(ib_uverbs_idr_lock);
68 DEFINE_IDR(ib_uverbs_pd_idr);
69 DEFINE_IDR(ib_uverbs_mr_idr);
70 DEFINE_IDR(ib_uverbs_mw_idr);
71 DEFINE_IDR(ib_uverbs_ah_idr);
72 DEFINE_IDR(ib_uverbs_cq_idr);
73 DEFINE_IDR(ib_uverbs_qp_idr);
74 DEFINE_IDR(ib_uverbs_srq_idr);
75 DEFINE_IDR(ib_uverbs_xrcd_idr);
76 DEFINE_IDR(ib_uverbs_rule_idr);
77 
78 static DEFINE_SPINLOCK(map_lock);
79 static DECLARE_BITMAP(dev_map, IB_UVERBS_MAX_DEVICES);
80 
81 static ssize_t (*uverbs_cmd_table[])(struct ib_uverbs_file *file,
82 				     struct ib_device *ib_dev,
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_REREG_MR]		= ib_uverbs_rereg_mr,
92 	[IB_USER_VERBS_CMD_DEREG_MR]		= ib_uverbs_dereg_mr,
93 	[IB_USER_VERBS_CMD_ALLOC_MW]		= ib_uverbs_alloc_mw,
94 	[IB_USER_VERBS_CMD_DEALLOC_MW]		= ib_uverbs_dealloc_mw,
95 	[IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL] = ib_uverbs_create_comp_channel,
96 	[IB_USER_VERBS_CMD_CREATE_CQ]		= ib_uverbs_create_cq,
97 	[IB_USER_VERBS_CMD_RESIZE_CQ]		= ib_uverbs_resize_cq,
98 	[IB_USER_VERBS_CMD_POLL_CQ]		= ib_uverbs_poll_cq,
99 	[IB_USER_VERBS_CMD_REQ_NOTIFY_CQ]	= ib_uverbs_req_notify_cq,
100 	[IB_USER_VERBS_CMD_DESTROY_CQ]		= ib_uverbs_destroy_cq,
101 	[IB_USER_VERBS_CMD_CREATE_QP]		= ib_uverbs_create_qp,
102 	[IB_USER_VERBS_CMD_QUERY_QP]		= ib_uverbs_query_qp,
103 	[IB_USER_VERBS_CMD_MODIFY_QP]		= ib_uverbs_modify_qp,
104 	[IB_USER_VERBS_CMD_DESTROY_QP]		= ib_uverbs_destroy_qp,
105 	[IB_USER_VERBS_CMD_POST_SEND]		= ib_uverbs_post_send,
106 	[IB_USER_VERBS_CMD_POST_RECV]		= ib_uverbs_post_recv,
107 	[IB_USER_VERBS_CMD_POST_SRQ_RECV]	= ib_uverbs_post_srq_recv,
108 	[IB_USER_VERBS_CMD_CREATE_AH]		= ib_uverbs_create_ah,
109 	[IB_USER_VERBS_CMD_DESTROY_AH]		= ib_uverbs_destroy_ah,
110 	[IB_USER_VERBS_CMD_ATTACH_MCAST]	= ib_uverbs_attach_mcast,
111 	[IB_USER_VERBS_CMD_DETACH_MCAST]	= ib_uverbs_detach_mcast,
112 	[IB_USER_VERBS_CMD_CREATE_SRQ]		= ib_uverbs_create_srq,
113 	[IB_USER_VERBS_CMD_MODIFY_SRQ]		= ib_uverbs_modify_srq,
114 	[IB_USER_VERBS_CMD_QUERY_SRQ]		= ib_uverbs_query_srq,
115 	[IB_USER_VERBS_CMD_DESTROY_SRQ]		= ib_uverbs_destroy_srq,
116 	[IB_USER_VERBS_CMD_OPEN_XRCD]		= ib_uverbs_open_xrcd,
117 	[IB_USER_VERBS_CMD_CLOSE_XRCD]		= ib_uverbs_close_xrcd,
118 	[IB_USER_VERBS_CMD_CREATE_XSRQ]		= ib_uverbs_create_xsrq,
119 	[IB_USER_VERBS_CMD_OPEN_QP]		= ib_uverbs_open_qp,
120 };
121 
122 static int (*uverbs_ex_cmd_table[])(struct ib_uverbs_file *file,
123 				    struct ib_device *ib_dev,
124 				    struct ib_udata *ucore,
125 				    struct ib_udata *uhw) = {
126 	[IB_USER_VERBS_EX_CMD_CREATE_FLOW]	= ib_uverbs_ex_create_flow,
127 	[IB_USER_VERBS_EX_CMD_DESTROY_FLOW]	= ib_uverbs_ex_destroy_flow,
128 	[IB_USER_VERBS_EX_CMD_QUERY_DEVICE]	= ib_uverbs_ex_query_device,
129 	[IB_USER_VERBS_EX_CMD_CREATE_CQ]	= ib_uverbs_ex_create_cq,
130 	[IB_USER_VERBS_EX_CMD_CREATE_QP]        = ib_uverbs_ex_create_qp,
131 };
132 
133 static void ib_uverbs_add_one(struct ib_device *device);
134 static void ib_uverbs_remove_one(struct ib_device *device, void *client_data);
135 
136 static void ib_uverbs_release_dev(struct kobject *kobj)
137 {
138 	struct ib_uverbs_device *dev =
139 		container_of(kobj, struct ib_uverbs_device, kobj);
140 
141 	cleanup_srcu_struct(&dev->disassociate_srcu);
142 	kfree(dev);
143 }
144 
145 static struct kobj_type ib_uverbs_dev_ktype = {
146 	.release = ib_uverbs_release_dev,
147 };
148 
149 static void ib_uverbs_release_event_file(struct kref *ref)
150 {
151 	struct ib_uverbs_event_file *file =
152 		container_of(ref, struct ib_uverbs_event_file, ref);
153 
154 	kfree(file);
155 }
156 
157 void ib_uverbs_release_ucq(struct ib_uverbs_file *file,
158 			  struct ib_uverbs_event_file *ev_file,
159 			  struct ib_ucq_object *uobj)
160 {
161 	struct ib_uverbs_event *evt, *tmp;
162 
163 	if (ev_file) {
164 		spin_lock_irq(&ev_file->lock);
165 		list_for_each_entry_safe(evt, tmp, &uobj->comp_list, obj_list) {
166 			list_del(&evt->list);
167 			kfree(evt);
168 		}
169 		spin_unlock_irq(&ev_file->lock);
170 
171 		kref_put(&ev_file->ref, ib_uverbs_release_event_file);
172 	}
173 
174 	spin_lock_irq(&file->async_file->lock);
175 	list_for_each_entry_safe(evt, tmp, &uobj->async_list, obj_list) {
176 		list_del(&evt->list);
177 		kfree(evt);
178 	}
179 	spin_unlock_irq(&file->async_file->lock);
180 }
181 
182 void ib_uverbs_release_uevent(struct ib_uverbs_file *file,
183 			      struct ib_uevent_object *uobj)
184 {
185 	struct ib_uverbs_event *evt, *tmp;
186 
187 	spin_lock_irq(&file->async_file->lock);
188 	list_for_each_entry_safe(evt, tmp, &uobj->event_list, obj_list) {
189 		list_del(&evt->list);
190 		kfree(evt);
191 	}
192 	spin_unlock_irq(&file->async_file->lock);
193 }
194 
195 static void ib_uverbs_detach_umcast(struct ib_qp *qp,
196 				    struct ib_uqp_object *uobj)
197 {
198 	struct ib_uverbs_mcast_entry *mcast, *tmp;
199 
200 	list_for_each_entry_safe(mcast, tmp, &uobj->mcast_list, list) {
201 		ib_detach_mcast(qp, &mcast->gid, mcast->lid);
202 		list_del(&mcast->list);
203 		kfree(mcast);
204 	}
205 }
206 
207 static int ib_uverbs_cleanup_ucontext(struct ib_uverbs_file *file,
208 				      struct ib_ucontext *context)
209 {
210 	struct ib_uobject *uobj, *tmp;
211 
212 	context->closing = 1;
213 
214 	list_for_each_entry_safe(uobj, tmp, &context->ah_list, list) {
215 		struct ib_ah *ah = uobj->object;
216 
217 		idr_remove_uobj(&ib_uverbs_ah_idr, uobj);
218 		ib_destroy_ah(ah);
219 		kfree(uobj);
220 	}
221 
222 	/* Remove MWs before QPs, in order to support type 2A MWs. */
223 	list_for_each_entry_safe(uobj, tmp, &context->mw_list, list) {
224 		struct ib_mw *mw = uobj->object;
225 
226 		idr_remove_uobj(&ib_uverbs_mw_idr, uobj);
227 		ib_dealloc_mw(mw);
228 		kfree(uobj);
229 	}
230 
231 	list_for_each_entry_safe(uobj, tmp, &context->rule_list, list) {
232 		struct ib_flow *flow_id = uobj->object;
233 
234 		idr_remove_uobj(&ib_uverbs_rule_idr, uobj);
235 		ib_destroy_flow(flow_id);
236 		kfree(uobj);
237 	}
238 
239 	list_for_each_entry_safe(uobj, tmp, &context->qp_list, list) {
240 		struct ib_qp *qp = uobj->object;
241 		struct ib_uqp_object *uqp =
242 			container_of(uobj, struct ib_uqp_object, uevent.uobject);
243 
244 		idr_remove_uobj(&ib_uverbs_qp_idr, uobj);
245 		if (qp != qp->real_qp) {
246 			ib_close_qp(qp);
247 		} else {
248 			ib_uverbs_detach_umcast(qp, uqp);
249 			ib_destroy_qp(qp);
250 		}
251 		ib_uverbs_release_uevent(file, &uqp->uevent);
252 		kfree(uqp);
253 	}
254 
255 	list_for_each_entry_safe(uobj, tmp, &context->srq_list, list) {
256 		struct ib_srq *srq = uobj->object;
257 		struct ib_uevent_object *uevent =
258 			container_of(uobj, struct ib_uevent_object, uobject);
259 
260 		idr_remove_uobj(&ib_uverbs_srq_idr, uobj);
261 		ib_destroy_srq(srq);
262 		ib_uverbs_release_uevent(file, uevent);
263 		kfree(uevent);
264 	}
265 
266 	list_for_each_entry_safe(uobj, tmp, &context->cq_list, list) {
267 		struct ib_cq *cq = uobj->object;
268 		struct ib_uverbs_event_file *ev_file = cq->cq_context;
269 		struct ib_ucq_object *ucq =
270 			container_of(uobj, struct ib_ucq_object, uobject);
271 
272 		idr_remove_uobj(&ib_uverbs_cq_idr, uobj);
273 		ib_destroy_cq(cq);
274 		ib_uverbs_release_ucq(file, ev_file, ucq);
275 		kfree(ucq);
276 	}
277 
278 	list_for_each_entry_safe(uobj, tmp, &context->mr_list, list) {
279 		struct ib_mr *mr = uobj->object;
280 
281 		idr_remove_uobj(&ib_uverbs_mr_idr, uobj);
282 		ib_dereg_mr(mr);
283 		kfree(uobj);
284 	}
285 
286 	mutex_lock(&file->device->xrcd_tree_mutex);
287 	list_for_each_entry_safe(uobj, tmp, &context->xrcd_list, list) {
288 		struct ib_xrcd *xrcd = uobj->object;
289 		struct ib_uxrcd_object *uxrcd =
290 			container_of(uobj, struct ib_uxrcd_object, uobject);
291 
292 		idr_remove_uobj(&ib_uverbs_xrcd_idr, uobj);
293 		ib_uverbs_dealloc_xrcd(file->device, xrcd);
294 		kfree(uxrcd);
295 	}
296 	mutex_unlock(&file->device->xrcd_tree_mutex);
297 
298 	list_for_each_entry_safe(uobj, tmp, &context->pd_list, list) {
299 		struct ib_pd *pd = uobj->object;
300 
301 		idr_remove_uobj(&ib_uverbs_pd_idr, uobj);
302 		ib_dealloc_pd(pd);
303 		kfree(uobj);
304 	}
305 
306 	put_pid(context->tgid);
307 
308 	return context->device->dealloc_ucontext(context);
309 }
310 
311 static void ib_uverbs_comp_dev(struct ib_uverbs_device *dev)
312 {
313 	complete(&dev->comp);
314 }
315 
316 static void ib_uverbs_release_file(struct kref *ref)
317 {
318 	struct ib_uverbs_file *file =
319 		container_of(ref, struct ib_uverbs_file, ref);
320 	struct ib_device *ib_dev;
321 	int srcu_key;
322 
323 	srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
324 	ib_dev = srcu_dereference(file->device->ib_dev,
325 				  &file->device->disassociate_srcu);
326 	if (ib_dev && !ib_dev->disassociate_ucontext)
327 		module_put(ib_dev->owner);
328 	srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
329 
330 	if (atomic_dec_and_test(&file->device->refcount))
331 		ib_uverbs_comp_dev(file->device);
332 
333 	kfree(file);
334 }
335 
336 static ssize_t ib_uverbs_event_read(struct file *filp, char __user *buf,
337 				    size_t count, loff_t *pos)
338 {
339 	struct ib_uverbs_event_file *file = filp->private_data;
340 	struct ib_uverbs_event *event;
341 	int eventsz;
342 	int ret = 0;
343 
344 	spin_lock_irq(&file->lock);
345 
346 	while (list_empty(&file->event_list)) {
347 		spin_unlock_irq(&file->lock);
348 
349 		if (filp->f_flags & O_NONBLOCK)
350 			return -EAGAIN;
351 
352 		if (wait_event_interruptible(file->poll_wait,
353 					     (!list_empty(&file->event_list) ||
354 			/* The barriers built into wait_event_interruptible()
355 			 * and wake_up() guarentee this will see the null set
356 			 * without using RCU
357 			 */
358 					     !file->uverbs_file->device->ib_dev)))
359 			return -ERESTARTSYS;
360 
361 		/* If device was disassociated and no event exists set an error */
362 		if (list_empty(&file->event_list) &&
363 		    !file->uverbs_file->device->ib_dev)
364 			return -EIO;
365 
366 		spin_lock_irq(&file->lock);
367 	}
368 
369 	event = list_entry(file->event_list.next, struct ib_uverbs_event, list);
370 
371 	if (file->is_async)
372 		eventsz = sizeof (struct ib_uverbs_async_event_desc);
373 	else
374 		eventsz = sizeof (struct ib_uverbs_comp_event_desc);
375 
376 	if (eventsz > count) {
377 		ret   = -EINVAL;
378 		event = NULL;
379 	} else {
380 		list_del(file->event_list.next);
381 		if (event->counter) {
382 			++(*event->counter);
383 			list_del(&event->obj_list);
384 		}
385 	}
386 
387 	spin_unlock_irq(&file->lock);
388 
389 	if (event) {
390 		if (copy_to_user(buf, event, eventsz))
391 			ret = -EFAULT;
392 		else
393 			ret = eventsz;
394 	}
395 
396 	kfree(event);
397 
398 	return ret;
399 }
400 
401 static unsigned int ib_uverbs_event_poll(struct file *filp,
402 					 struct poll_table_struct *wait)
403 {
404 	unsigned int pollflags = 0;
405 	struct ib_uverbs_event_file *file = filp->private_data;
406 
407 	poll_wait(filp, &file->poll_wait, wait);
408 
409 	spin_lock_irq(&file->lock);
410 	if (!list_empty(&file->event_list))
411 		pollflags = POLLIN | POLLRDNORM;
412 	spin_unlock_irq(&file->lock);
413 
414 	return pollflags;
415 }
416 
417 static int ib_uverbs_event_fasync(int fd, struct file *filp, int on)
418 {
419 	struct ib_uverbs_event_file *file = filp->private_data;
420 
421 	return fasync_helper(fd, filp, on, &file->async_queue);
422 }
423 
424 static int ib_uverbs_event_close(struct inode *inode, struct file *filp)
425 {
426 	struct ib_uverbs_event_file *file = filp->private_data;
427 	struct ib_uverbs_event *entry, *tmp;
428 	int closed_already = 0;
429 
430 	mutex_lock(&file->uverbs_file->device->lists_mutex);
431 	spin_lock_irq(&file->lock);
432 	closed_already = file->is_closed;
433 	file->is_closed = 1;
434 	list_for_each_entry_safe(entry, tmp, &file->event_list, list) {
435 		if (entry->counter)
436 			list_del(&entry->obj_list);
437 		kfree(entry);
438 	}
439 	spin_unlock_irq(&file->lock);
440 	if (!closed_already) {
441 		list_del(&file->list);
442 		if (file->is_async)
443 			ib_unregister_event_handler(&file->uverbs_file->
444 				event_handler);
445 	}
446 	mutex_unlock(&file->uverbs_file->device->lists_mutex);
447 
448 	kref_put(&file->uverbs_file->ref, ib_uverbs_release_file);
449 	kref_put(&file->ref, ib_uverbs_release_event_file);
450 
451 	return 0;
452 }
453 
454 static const struct file_operations uverbs_event_fops = {
455 	.owner	 = THIS_MODULE,
456 	.read	 = ib_uverbs_event_read,
457 	.poll    = ib_uverbs_event_poll,
458 	.release = ib_uverbs_event_close,
459 	.fasync  = ib_uverbs_event_fasync,
460 	.llseek	 = no_llseek,
461 };
462 
463 void ib_uverbs_comp_handler(struct ib_cq *cq, void *cq_context)
464 {
465 	struct ib_uverbs_event_file    *file = cq_context;
466 	struct ib_ucq_object	       *uobj;
467 	struct ib_uverbs_event	       *entry;
468 	unsigned long			flags;
469 
470 	if (!file)
471 		return;
472 
473 	spin_lock_irqsave(&file->lock, flags);
474 	if (file->is_closed) {
475 		spin_unlock_irqrestore(&file->lock, flags);
476 		return;
477 	}
478 
479 	entry = kmalloc(sizeof *entry, GFP_ATOMIC);
480 	if (!entry) {
481 		spin_unlock_irqrestore(&file->lock, flags);
482 		return;
483 	}
484 
485 	uobj = container_of(cq->uobject, struct ib_ucq_object, uobject);
486 
487 	entry->desc.comp.cq_handle = cq->uobject->user_handle;
488 	entry->counter		   = &uobj->comp_events_reported;
489 
490 	list_add_tail(&entry->list, &file->event_list);
491 	list_add_tail(&entry->obj_list, &uobj->comp_list);
492 	spin_unlock_irqrestore(&file->lock, flags);
493 
494 	wake_up_interruptible(&file->poll_wait);
495 	kill_fasync(&file->async_queue, SIGIO, POLL_IN);
496 }
497 
498 static void ib_uverbs_async_handler(struct ib_uverbs_file *file,
499 				    __u64 element, __u64 event,
500 				    struct list_head *obj_list,
501 				    u32 *counter)
502 {
503 	struct ib_uverbs_event *entry;
504 	unsigned long flags;
505 
506 	spin_lock_irqsave(&file->async_file->lock, flags);
507 	if (file->async_file->is_closed) {
508 		spin_unlock_irqrestore(&file->async_file->lock, flags);
509 		return;
510 	}
511 
512 	entry = kmalloc(sizeof *entry, GFP_ATOMIC);
513 	if (!entry) {
514 		spin_unlock_irqrestore(&file->async_file->lock, flags);
515 		return;
516 	}
517 
518 	entry->desc.async.element    = element;
519 	entry->desc.async.event_type = event;
520 	entry->desc.async.reserved   = 0;
521 	entry->counter               = counter;
522 
523 	list_add_tail(&entry->list, &file->async_file->event_list);
524 	if (obj_list)
525 		list_add_tail(&entry->obj_list, obj_list);
526 	spin_unlock_irqrestore(&file->async_file->lock, flags);
527 
528 	wake_up_interruptible(&file->async_file->poll_wait);
529 	kill_fasync(&file->async_file->async_queue, SIGIO, POLL_IN);
530 }
531 
532 void ib_uverbs_cq_event_handler(struct ib_event *event, void *context_ptr)
533 {
534 	struct ib_ucq_object *uobj = container_of(event->element.cq->uobject,
535 						  struct ib_ucq_object, uobject);
536 
537 	ib_uverbs_async_handler(uobj->uverbs_file, uobj->uobject.user_handle,
538 				event->event, &uobj->async_list,
539 				&uobj->async_events_reported);
540 }
541 
542 void ib_uverbs_qp_event_handler(struct ib_event *event, void *context_ptr)
543 {
544 	struct ib_uevent_object *uobj;
545 
546 	/* for XRC target qp's, check that qp is live */
547 	if (!event->element.qp->uobject || !event->element.qp->uobject->live)
548 		return;
549 
550 	uobj = container_of(event->element.qp->uobject,
551 			    struct ib_uevent_object, uobject);
552 
553 	ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
554 				event->event, &uobj->event_list,
555 				&uobj->events_reported);
556 }
557 
558 void ib_uverbs_srq_event_handler(struct ib_event *event, void *context_ptr)
559 {
560 	struct ib_uevent_object *uobj;
561 
562 	uobj = container_of(event->element.srq->uobject,
563 			    struct ib_uevent_object, uobject);
564 
565 	ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
566 				event->event, &uobj->event_list,
567 				&uobj->events_reported);
568 }
569 
570 void ib_uverbs_event_handler(struct ib_event_handler *handler,
571 			     struct ib_event *event)
572 {
573 	struct ib_uverbs_file *file =
574 		container_of(handler, struct ib_uverbs_file, event_handler);
575 
576 	ib_uverbs_async_handler(file, event->element.port_num, event->event,
577 				NULL, NULL);
578 }
579 
580 void ib_uverbs_free_async_event_file(struct ib_uverbs_file *file)
581 {
582 	kref_put(&file->async_file->ref, ib_uverbs_release_event_file);
583 	file->async_file = NULL;
584 }
585 
586 struct file *ib_uverbs_alloc_event_file(struct ib_uverbs_file *uverbs_file,
587 					struct ib_device	*ib_dev,
588 					int is_async)
589 {
590 	struct ib_uverbs_event_file *ev_file;
591 	struct file *filp;
592 	int ret;
593 
594 	ev_file = kzalloc(sizeof(*ev_file), GFP_KERNEL);
595 	if (!ev_file)
596 		return ERR_PTR(-ENOMEM);
597 
598 	kref_init(&ev_file->ref);
599 	spin_lock_init(&ev_file->lock);
600 	INIT_LIST_HEAD(&ev_file->event_list);
601 	init_waitqueue_head(&ev_file->poll_wait);
602 	ev_file->uverbs_file = uverbs_file;
603 	kref_get(&ev_file->uverbs_file->ref);
604 	ev_file->async_queue = NULL;
605 	ev_file->is_closed   = 0;
606 
607 	filp = anon_inode_getfile("[infinibandevent]", &uverbs_event_fops,
608 				  ev_file, O_RDONLY);
609 	if (IS_ERR(filp))
610 		goto err_put_refs;
611 
612 	mutex_lock(&uverbs_file->device->lists_mutex);
613 	list_add_tail(&ev_file->list,
614 		      &uverbs_file->device->uverbs_events_file_list);
615 	mutex_unlock(&uverbs_file->device->lists_mutex);
616 
617 	if (is_async) {
618 		WARN_ON(uverbs_file->async_file);
619 		uverbs_file->async_file = ev_file;
620 		kref_get(&uverbs_file->async_file->ref);
621 		INIT_IB_EVENT_HANDLER(&uverbs_file->event_handler,
622 				      ib_dev,
623 				      ib_uverbs_event_handler);
624 		ret = ib_register_event_handler(&uverbs_file->event_handler);
625 		if (ret)
626 			goto err_put_file;
627 
628 		/* At that point async file stuff was fully set */
629 		ev_file->is_async = 1;
630 	}
631 
632 	return filp;
633 
634 err_put_file:
635 	fput(filp);
636 	kref_put(&uverbs_file->async_file->ref, ib_uverbs_release_event_file);
637 	uverbs_file->async_file = NULL;
638 	return ERR_PTR(ret);
639 
640 err_put_refs:
641 	kref_put(&ev_file->uverbs_file->ref, ib_uverbs_release_file);
642 	kref_put(&ev_file->ref, ib_uverbs_release_event_file);
643 	return filp;
644 }
645 
646 /*
647  * Look up a completion event file by FD.  If lookup is successful,
648  * takes a ref to the event file struct that it returns; if
649  * unsuccessful, returns NULL.
650  */
651 struct ib_uverbs_event_file *ib_uverbs_lookup_comp_file(int fd)
652 {
653 	struct ib_uverbs_event_file *ev_file = NULL;
654 	struct fd f = fdget(fd);
655 
656 	if (!f.file)
657 		return NULL;
658 
659 	if (f.file->f_op != &uverbs_event_fops)
660 		goto out;
661 
662 	ev_file = f.file->private_data;
663 	if (ev_file->is_async) {
664 		ev_file = NULL;
665 		goto out;
666 	}
667 
668 	kref_get(&ev_file->ref);
669 
670 out:
671 	fdput(f);
672 	return ev_file;
673 }
674 
675 static ssize_t ib_uverbs_write(struct file *filp, const char __user *buf,
676 			     size_t count, loff_t *pos)
677 {
678 	struct ib_uverbs_file *file = filp->private_data;
679 	struct ib_device *ib_dev;
680 	struct ib_uverbs_cmd_hdr hdr;
681 	__u32 flags;
682 	int srcu_key;
683 	ssize_t ret;
684 
685 	if (count < sizeof hdr)
686 		return -EINVAL;
687 
688 	if (copy_from_user(&hdr, buf, sizeof hdr))
689 		return -EFAULT;
690 
691 	srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
692 	ib_dev = srcu_dereference(file->device->ib_dev,
693 				  &file->device->disassociate_srcu);
694 	if (!ib_dev) {
695 		ret = -EIO;
696 		goto out;
697 	}
698 
699 	flags = (hdr.command &
700 		 IB_USER_VERBS_CMD_FLAGS_MASK) >> IB_USER_VERBS_CMD_FLAGS_SHIFT;
701 
702 	if (!flags) {
703 		__u32 command;
704 
705 		if (hdr.command & ~(__u32)(IB_USER_VERBS_CMD_FLAGS_MASK |
706 					   IB_USER_VERBS_CMD_COMMAND_MASK)) {
707 			ret = -EINVAL;
708 			goto out;
709 		}
710 
711 		command = hdr.command & IB_USER_VERBS_CMD_COMMAND_MASK;
712 
713 		if (command >= ARRAY_SIZE(uverbs_cmd_table) ||
714 		    !uverbs_cmd_table[command]) {
715 			ret = -EINVAL;
716 			goto out;
717 		}
718 
719 		if (!file->ucontext &&
720 		    command != IB_USER_VERBS_CMD_GET_CONTEXT) {
721 			ret = -EINVAL;
722 			goto out;
723 		}
724 
725 		if (!(ib_dev->uverbs_cmd_mask & (1ull << command))) {
726 			ret = -ENOSYS;
727 			goto out;
728 		}
729 
730 		if (hdr.in_words * 4 != count) {
731 			ret = -EINVAL;
732 			goto out;
733 		}
734 
735 		ret = uverbs_cmd_table[command](file, ib_dev,
736 						 buf + sizeof(hdr),
737 						 hdr.in_words * 4,
738 						 hdr.out_words * 4);
739 
740 	} else if (flags == IB_USER_VERBS_CMD_FLAG_EXTENDED) {
741 		__u32 command;
742 
743 		struct ib_uverbs_ex_cmd_hdr ex_hdr;
744 		struct ib_udata ucore;
745 		struct ib_udata uhw;
746 		size_t written_count = count;
747 
748 		if (hdr.command & ~(__u32)(IB_USER_VERBS_CMD_FLAGS_MASK |
749 					   IB_USER_VERBS_CMD_COMMAND_MASK)) {
750 			ret = -EINVAL;
751 			goto out;
752 		}
753 
754 		command = hdr.command & IB_USER_VERBS_CMD_COMMAND_MASK;
755 
756 		if (command >= ARRAY_SIZE(uverbs_ex_cmd_table) ||
757 		    !uverbs_ex_cmd_table[command]) {
758 			ret = -ENOSYS;
759 			goto out;
760 		}
761 
762 		if (!file->ucontext) {
763 			ret = -EINVAL;
764 			goto out;
765 		}
766 
767 		if (!(ib_dev->uverbs_ex_cmd_mask & (1ull << command))) {
768 			ret = -ENOSYS;
769 			goto out;
770 		}
771 
772 		if (count < (sizeof(hdr) + sizeof(ex_hdr))) {
773 			ret = -EINVAL;
774 			goto out;
775 		}
776 
777 		if (copy_from_user(&ex_hdr, buf + sizeof(hdr), sizeof(ex_hdr))) {
778 			ret = -EFAULT;
779 			goto out;
780 		}
781 
782 		count -= sizeof(hdr) + sizeof(ex_hdr);
783 		buf += sizeof(hdr) + sizeof(ex_hdr);
784 
785 		if ((hdr.in_words + ex_hdr.provider_in_words) * 8 != count) {
786 			ret = -EINVAL;
787 			goto out;
788 		}
789 
790 		if (ex_hdr.cmd_hdr_reserved) {
791 			ret = -EINVAL;
792 			goto out;
793 		}
794 
795 		if (ex_hdr.response) {
796 			if (!hdr.out_words && !ex_hdr.provider_out_words) {
797 				ret = -EINVAL;
798 				goto out;
799 			}
800 
801 			if (!access_ok(VERIFY_WRITE,
802 				       (void __user *) (unsigned long) ex_hdr.response,
803 				       (hdr.out_words + ex_hdr.provider_out_words) * 8)) {
804 				ret = -EFAULT;
805 				goto out;
806 			}
807 		} else {
808 			if (hdr.out_words || ex_hdr.provider_out_words) {
809 				ret = -EINVAL;
810 				goto out;
811 			}
812 		}
813 
814 		INIT_UDATA_BUF_OR_NULL(&ucore, buf, (unsigned long) ex_hdr.response,
815 				       hdr.in_words * 8, hdr.out_words * 8);
816 
817 		INIT_UDATA_BUF_OR_NULL(&uhw,
818 				       buf + ucore.inlen,
819 				       (unsigned long) ex_hdr.response + ucore.outlen,
820 				       ex_hdr.provider_in_words * 8,
821 				       ex_hdr.provider_out_words * 8);
822 
823 		ret = uverbs_ex_cmd_table[command](file,
824 						   ib_dev,
825 						   &ucore,
826 						   &uhw);
827 		if (!ret)
828 			ret = written_count;
829 	} else {
830 		ret = -ENOSYS;
831 	}
832 
833 out:
834 	srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
835 	return ret;
836 }
837 
838 static int ib_uverbs_mmap(struct file *filp, struct vm_area_struct *vma)
839 {
840 	struct ib_uverbs_file *file = filp->private_data;
841 	struct ib_device *ib_dev;
842 	int ret = 0;
843 	int srcu_key;
844 
845 	srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
846 	ib_dev = srcu_dereference(file->device->ib_dev,
847 				  &file->device->disassociate_srcu);
848 	if (!ib_dev) {
849 		ret = -EIO;
850 		goto out;
851 	}
852 
853 	if (!file->ucontext)
854 		ret = -ENODEV;
855 	else
856 		ret = ib_dev->mmap(file->ucontext, vma);
857 out:
858 	srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
859 	return ret;
860 }
861 
862 /*
863  * ib_uverbs_open() does not need the BKL:
864  *
865  *  - the ib_uverbs_device structures are properly reference counted and
866  *    everything else is purely local to the file being created, so
867  *    races against other open calls are not a problem;
868  *  - there is no ioctl method to race against;
869  *  - the open method will either immediately run -ENXIO, or all
870  *    required initialization will be done.
871  */
872 static int ib_uverbs_open(struct inode *inode, struct file *filp)
873 {
874 	struct ib_uverbs_device *dev;
875 	struct ib_uverbs_file *file;
876 	struct ib_device *ib_dev;
877 	int ret;
878 	int module_dependent;
879 	int srcu_key;
880 
881 	dev = container_of(inode->i_cdev, struct ib_uverbs_device, cdev);
882 	if (!atomic_inc_not_zero(&dev->refcount))
883 		return -ENXIO;
884 
885 	srcu_key = srcu_read_lock(&dev->disassociate_srcu);
886 	mutex_lock(&dev->lists_mutex);
887 	ib_dev = srcu_dereference(dev->ib_dev,
888 				  &dev->disassociate_srcu);
889 	if (!ib_dev) {
890 		ret = -EIO;
891 		goto err;
892 	}
893 
894 	/* In case IB device supports disassociate ucontext, there is no hard
895 	 * dependency between uverbs device and its low level device.
896 	 */
897 	module_dependent = !(ib_dev->disassociate_ucontext);
898 
899 	if (module_dependent) {
900 		if (!try_module_get(ib_dev->owner)) {
901 			ret = -ENODEV;
902 			goto err;
903 		}
904 	}
905 
906 	file = kzalloc(sizeof(*file), GFP_KERNEL);
907 	if (!file) {
908 		ret = -ENOMEM;
909 		if (module_dependent)
910 			goto err_module;
911 
912 		goto err;
913 	}
914 
915 	file->device	 = dev;
916 	file->ucontext	 = NULL;
917 	file->async_file = NULL;
918 	kref_init(&file->ref);
919 	mutex_init(&file->mutex);
920 
921 	filp->private_data = file;
922 	kobject_get(&dev->kobj);
923 	list_add_tail(&file->list, &dev->uverbs_file_list);
924 	mutex_unlock(&dev->lists_mutex);
925 	srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
926 
927 	return nonseekable_open(inode, filp);
928 
929 err_module:
930 	module_put(ib_dev->owner);
931 
932 err:
933 	mutex_unlock(&dev->lists_mutex);
934 	srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
935 	if (atomic_dec_and_test(&dev->refcount))
936 		ib_uverbs_comp_dev(dev);
937 
938 	return ret;
939 }
940 
941 static int ib_uverbs_close(struct inode *inode, struct file *filp)
942 {
943 	struct ib_uverbs_file *file = filp->private_data;
944 	struct ib_uverbs_device *dev = file->device;
945 	struct ib_ucontext *ucontext = NULL;
946 
947 	mutex_lock(&file->device->lists_mutex);
948 	ucontext = file->ucontext;
949 	file->ucontext = NULL;
950 	if (!file->is_closed) {
951 		list_del(&file->list);
952 		file->is_closed = 1;
953 	}
954 	mutex_unlock(&file->device->lists_mutex);
955 	if (ucontext)
956 		ib_uverbs_cleanup_ucontext(file, ucontext);
957 
958 	if (file->async_file)
959 		kref_put(&file->async_file->ref, ib_uverbs_release_event_file);
960 
961 	kref_put(&file->ref, ib_uverbs_release_file);
962 	kobject_put(&dev->kobj);
963 
964 	return 0;
965 }
966 
967 static const struct file_operations uverbs_fops = {
968 	.owner	 = THIS_MODULE,
969 	.write	 = ib_uverbs_write,
970 	.open	 = ib_uverbs_open,
971 	.release = ib_uverbs_close,
972 	.llseek	 = no_llseek,
973 };
974 
975 static const struct file_operations uverbs_mmap_fops = {
976 	.owner	 = THIS_MODULE,
977 	.write	 = ib_uverbs_write,
978 	.mmap    = ib_uverbs_mmap,
979 	.open	 = ib_uverbs_open,
980 	.release = ib_uverbs_close,
981 	.llseek	 = no_llseek,
982 };
983 
984 static struct ib_client uverbs_client = {
985 	.name   = "uverbs",
986 	.add    = ib_uverbs_add_one,
987 	.remove = ib_uverbs_remove_one
988 };
989 
990 static ssize_t show_ibdev(struct device *device, struct device_attribute *attr,
991 			  char *buf)
992 {
993 	int ret = -ENODEV;
994 	int srcu_key;
995 	struct ib_uverbs_device *dev = dev_get_drvdata(device);
996 	struct ib_device *ib_dev;
997 
998 	if (!dev)
999 		return -ENODEV;
1000 
1001 	srcu_key = srcu_read_lock(&dev->disassociate_srcu);
1002 	ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu);
1003 	if (ib_dev)
1004 		ret = sprintf(buf, "%s\n", ib_dev->name);
1005 	srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1006 
1007 	return ret;
1008 }
1009 static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
1010 
1011 static ssize_t show_dev_abi_version(struct device *device,
1012 				    struct device_attribute *attr, char *buf)
1013 {
1014 	struct ib_uverbs_device *dev = dev_get_drvdata(device);
1015 	int ret = -ENODEV;
1016 	int srcu_key;
1017 	struct ib_device *ib_dev;
1018 
1019 	if (!dev)
1020 		return -ENODEV;
1021 	srcu_key = srcu_read_lock(&dev->disassociate_srcu);
1022 	ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu);
1023 	if (ib_dev)
1024 		ret = sprintf(buf, "%d\n", ib_dev->uverbs_abi_ver);
1025 	srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1026 
1027 	return ret;
1028 }
1029 static DEVICE_ATTR(abi_version, S_IRUGO, show_dev_abi_version, NULL);
1030 
1031 static CLASS_ATTR_STRING(abi_version, S_IRUGO,
1032 			 __stringify(IB_USER_VERBS_ABI_VERSION));
1033 
1034 static dev_t overflow_maj;
1035 static DECLARE_BITMAP(overflow_map, IB_UVERBS_MAX_DEVICES);
1036 
1037 /*
1038  * If we have more than IB_UVERBS_MAX_DEVICES, dynamically overflow by
1039  * requesting a new major number and doubling the number of max devices we
1040  * support. It's stupid, but simple.
1041  */
1042 static int find_overflow_devnum(void)
1043 {
1044 	int ret;
1045 
1046 	if (!overflow_maj) {
1047 		ret = alloc_chrdev_region(&overflow_maj, 0, IB_UVERBS_MAX_DEVICES,
1048 					  "infiniband_verbs");
1049 		if (ret) {
1050 			printk(KERN_ERR "user_verbs: couldn't register dynamic device number\n");
1051 			return ret;
1052 		}
1053 	}
1054 
1055 	ret = find_first_zero_bit(overflow_map, IB_UVERBS_MAX_DEVICES);
1056 	if (ret >= IB_UVERBS_MAX_DEVICES)
1057 		return -1;
1058 
1059 	return ret;
1060 }
1061 
1062 static void ib_uverbs_add_one(struct ib_device *device)
1063 {
1064 	int devnum;
1065 	dev_t base;
1066 	struct ib_uverbs_device *uverbs_dev;
1067 	int ret;
1068 
1069 	if (!device->alloc_ucontext)
1070 		return;
1071 
1072 	uverbs_dev = kzalloc(sizeof *uverbs_dev, GFP_KERNEL);
1073 	if (!uverbs_dev)
1074 		return;
1075 
1076 	ret = init_srcu_struct(&uverbs_dev->disassociate_srcu);
1077 	if (ret) {
1078 		kfree(uverbs_dev);
1079 		return;
1080 	}
1081 
1082 	atomic_set(&uverbs_dev->refcount, 1);
1083 	init_completion(&uverbs_dev->comp);
1084 	uverbs_dev->xrcd_tree = RB_ROOT;
1085 	mutex_init(&uverbs_dev->xrcd_tree_mutex);
1086 	kobject_init(&uverbs_dev->kobj, &ib_uverbs_dev_ktype);
1087 	mutex_init(&uverbs_dev->lists_mutex);
1088 	INIT_LIST_HEAD(&uverbs_dev->uverbs_file_list);
1089 	INIT_LIST_HEAD(&uverbs_dev->uverbs_events_file_list);
1090 
1091 	spin_lock(&map_lock);
1092 	devnum = find_first_zero_bit(dev_map, IB_UVERBS_MAX_DEVICES);
1093 	if (devnum >= IB_UVERBS_MAX_DEVICES) {
1094 		spin_unlock(&map_lock);
1095 		devnum = find_overflow_devnum();
1096 		if (devnum < 0)
1097 			goto err;
1098 
1099 		spin_lock(&map_lock);
1100 		uverbs_dev->devnum = devnum + IB_UVERBS_MAX_DEVICES;
1101 		base = devnum + overflow_maj;
1102 		set_bit(devnum, overflow_map);
1103 	} else {
1104 		uverbs_dev->devnum = devnum;
1105 		base = devnum + IB_UVERBS_BASE_DEV;
1106 		set_bit(devnum, dev_map);
1107 	}
1108 	spin_unlock(&map_lock);
1109 
1110 	rcu_assign_pointer(uverbs_dev->ib_dev, device);
1111 	uverbs_dev->num_comp_vectors = device->num_comp_vectors;
1112 
1113 	cdev_init(&uverbs_dev->cdev, NULL);
1114 	uverbs_dev->cdev.owner = THIS_MODULE;
1115 	uverbs_dev->cdev.ops = device->mmap ? &uverbs_mmap_fops : &uverbs_fops;
1116 	uverbs_dev->cdev.kobj.parent = &uverbs_dev->kobj;
1117 	kobject_set_name(&uverbs_dev->cdev.kobj, "uverbs%d", uverbs_dev->devnum);
1118 	if (cdev_add(&uverbs_dev->cdev, base, 1))
1119 		goto err_cdev;
1120 
1121 	uverbs_dev->dev = device_create(uverbs_class, device->dma_device,
1122 					uverbs_dev->cdev.dev, uverbs_dev,
1123 					"uverbs%d", uverbs_dev->devnum);
1124 	if (IS_ERR(uverbs_dev->dev))
1125 		goto err_cdev;
1126 
1127 	if (device_create_file(uverbs_dev->dev, &dev_attr_ibdev))
1128 		goto err_class;
1129 	if (device_create_file(uverbs_dev->dev, &dev_attr_abi_version))
1130 		goto err_class;
1131 
1132 	ib_set_client_data(device, &uverbs_client, uverbs_dev);
1133 
1134 	return;
1135 
1136 err_class:
1137 	device_destroy(uverbs_class, uverbs_dev->cdev.dev);
1138 
1139 err_cdev:
1140 	cdev_del(&uverbs_dev->cdev);
1141 	if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
1142 		clear_bit(devnum, dev_map);
1143 	else
1144 		clear_bit(devnum, overflow_map);
1145 
1146 err:
1147 	if (atomic_dec_and_test(&uverbs_dev->refcount))
1148 		ib_uverbs_comp_dev(uverbs_dev);
1149 	wait_for_completion(&uverbs_dev->comp);
1150 	kobject_put(&uverbs_dev->kobj);
1151 	return;
1152 }
1153 
1154 static void ib_uverbs_free_hw_resources(struct ib_uverbs_device *uverbs_dev,
1155 					struct ib_device *ib_dev)
1156 {
1157 	struct ib_uverbs_file *file;
1158 	struct ib_uverbs_event_file *event_file;
1159 	struct ib_event event;
1160 
1161 	/* Pending running commands to terminate */
1162 	synchronize_srcu(&uverbs_dev->disassociate_srcu);
1163 	event.event = IB_EVENT_DEVICE_FATAL;
1164 	event.element.port_num = 0;
1165 	event.device = ib_dev;
1166 
1167 	mutex_lock(&uverbs_dev->lists_mutex);
1168 	while (!list_empty(&uverbs_dev->uverbs_file_list)) {
1169 		struct ib_ucontext *ucontext;
1170 
1171 		file = list_first_entry(&uverbs_dev->uverbs_file_list,
1172 					struct ib_uverbs_file, list);
1173 		file->is_closed = 1;
1174 		ucontext = file->ucontext;
1175 		list_del(&file->list);
1176 		file->ucontext = NULL;
1177 		kref_get(&file->ref);
1178 		mutex_unlock(&uverbs_dev->lists_mutex);
1179 		/* We must release the mutex before going ahead and calling
1180 		 * disassociate_ucontext. disassociate_ucontext might end up
1181 		 * indirectly calling uverbs_close, for example due to freeing
1182 		 * the resources (e.g mmput).
1183 		 */
1184 		ib_uverbs_event_handler(&file->event_handler, &event);
1185 		if (ucontext) {
1186 			ib_dev->disassociate_ucontext(ucontext);
1187 			ib_uverbs_cleanup_ucontext(file, ucontext);
1188 		}
1189 
1190 		mutex_lock(&uverbs_dev->lists_mutex);
1191 		kref_put(&file->ref, ib_uverbs_release_file);
1192 	}
1193 
1194 	while (!list_empty(&uverbs_dev->uverbs_events_file_list)) {
1195 		event_file = list_first_entry(&uverbs_dev->
1196 					      uverbs_events_file_list,
1197 					      struct ib_uverbs_event_file,
1198 					      list);
1199 		spin_lock_irq(&event_file->lock);
1200 		event_file->is_closed = 1;
1201 		spin_unlock_irq(&event_file->lock);
1202 
1203 		list_del(&event_file->list);
1204 		if (event_file->is_async) {
1205 			ib_unregister_event_handler(&event_file->uverbs_file->
1206 						    event_handler);
1207 			event_file->uverbs_file->event_handler.device = NULL;
1208 		}
1209 
1210 		wake_up_interruptible(&event_file->poll_wait);
1211 		kill_fasync(&event_file->async_queue, SIGIO, POLL_IN);
1212 	}
1213 	mutex_unlock(&uverbs_dev->lists_mutex);
1214 }
1215 
1216 static void ib_uverbs_remove_one(struct ib_device *device, void *client_data)
1217 {
1218 	struct ib_uverbs_device *uverbs_dev = client_data;
1219 	int wait_clients = 1;
1220 
1221 	if (!uverbs_dev)
1222 		return;
1223 
1224 	dev_set_drvdata(uverbs_dev->dev, NULL);
1225 	device_destroy(uverbs_class, uverbs_dev->cdev.dev);
1226 	cdev_del(&uverbs_dev->cdev);
1227 
1228 	if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
1229 		clear_bit(uverbs_dev->devnum, dev_map);
1230 	else
1231 		clear_bit(uverbs_dev->devnum - IB_UVERBS_MAX_DEVICES, overflow_map);
1232 
1233 	if (device->disassociate_ucontext) {
1234 		/* We disassociate HW resources and immediately return.
1235 		 * Userspace will see a EIO errno for all future access.
1236 		 * Upon returning, ib_device may be freed internally and is not
1237 		 * valid any more.
1238 		 * uverbs_device is still available until all clients close
1239 		 * their files, then the uverbs device ref count will be zero
1240 		 * and its resources will be freed.
1241 		 * Note: At this point no more files can be opened since the
1242 		 * cdev was deleted, however active clients can still issue
1243 		 * commands and close their open files.
1244 		 */
1245 		rcu_assign_pointer(uverbs_dev->ib_dev, NULL);
1246 		ib_uverbs_free_hw_resources(uverbs_dev, device);
1247 		wait_clients = 0;
1248 	}
1249 
1250 	if (atomic_dec_and_test(&uverbs_dev->refcount))
1251 		ib_uverbs_comp_dev(uverbs_dev);
1252 	if (wait_clients)
1253 		wait_for_completion(&uverbs_dev->comp);
1254 	kobject_put(&uverbs_dev->kobj);
1255 }
1256 
1257 static char *uverbs_devnode(struct device *dev, umode_t *mode)
1258 {
1259 	if (mode)
1260 		*mode = 0666;
1261 	return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev));
1262 }
1263 
1264 static int __init ib_uverbs_init(void)
1265 {
1266 	int ret;
1267 
1268 	ret = register_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES,
1269 				     "infiniband_verbs");
1270 	if (ret) {
1271 		printk(KERN_ERR "user_verbs: couldn't register device number\n");
1272 		goto out;
1273 	}
1274 
1275 	uverbs_class = class_create(THIS_MODULE, "infiniband_verbs");
1276 	if (IS_ERR(uverbs_class)) {
1277 		ret = PTR_ERR(uverbs_class);
1278 		printk(KERN_ERR "user_verbs: couldn't create class infiniband_verbs\n");
1279 		goto out_chrdev;
1280 	}
1281 
1282 	uverbs_class->devnode = uverbs_devnode;
1283 
1284 	ret = class_create_file(uverbs_class, &class_attr_abi_version.attr);
1285 	if (ret) {
1286 		printk(KERN_ERR "user_verbs: couldn't create abi_version attribute\n");
1287 		goto out_class;
1288 	}
1289 
1290 	ret = ib_register_client(&uverbs_client);
1291 	if (ret) {
1292 		printk(KERN_ERR "user_verbs: couldn't register client\n");
1293 		goto out_class;
1294 	}
1295 
1296 	return 0;
1297 
1298 out_class:
1299 	class_destroy(uverbs_class);
1300 
1301 out_chrdev:
1302 	unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
1303 
1304 out:
1305 	return ret;
1306 }
1307 
1308 static void __exit ib_uverbs_cleanup(void)
1309 {
1310 	ib_unregister_client(&uverbs_client);
1311 	class_destroy(uverbs_class);
1312 	unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
1313 	if (overflow_maj)
1314 		unregister_chrdev_region(overflow_maj, IB_UVERBS_MAX_DEVICES);
1315 	idr_destroy(&ib_uverbs_pd_idr);
1316 	idr_destroy(&ib_uverbs_mr_idr);
1317 	idr_destroy(&ib_uverbs_mw_idr);
1318 	idr_destroy(&ib_uverbs_ah_idr);
1319 	idr_destroy(&ib_uverbs_cq_idr);
1320 	idr_destroy(&ib_uverbs_qp_idr);
1321 	idr_destroy(&ib_uverbs_srq_idr);
1322 }
1323 
1324 module_init(ib_uverbs_init);
1325 module_exit(ib_uverbs_cleanup);
1326