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 <rdma/ib.h>
52 
53 #include "uverbs.h"
54 
55 MODULE_AUTHOR("Roland Dreier");
56 MODULE_DESCRIPTION("InfiniBand userspace verbs access");
57 MODULE_LICENSE("Dual BSD/GPL");
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 DEFINE_IDR(ib_uverbs_xrcd_idr);
78 DEFINE_IDR(ib_uverbs_rule_idr);
79 
80 static DEFINE_SPINLOCK(map_lock);
81 static DECLARE_BITMAP(dev_map, IB_UVERBS_MAX_DEVICES);
82 
83 static ssize_t (*uverbs_cmd_table[])(struct ib_uverbs_file *file,
84 				     struct ib_device *ib_dev,
85 				     const char __user *buf, int in_len,
86 				     int out_len) = {
87 	[IB_USER_VERBS_CMD_GET_CONTEXT]		= ib_uverbs_get_context,
88 	[IB_USER_VERBS_CMD_QUERY_DEVICE]	= ib_uverbs_query_device,
89 	[IB_USER_VERBS_CMD_QUERY_PORT]		= ib_uverbs_query_port,
90 	[IB_USER_VERBS_CMD_ALLOC_PD]		= ib_uverbs_alloc_pd,
91 	[IB_USER_VERBS_CMD_DEALLOC_PD]		= ib_uverbs_dealloc_pd,
92 	[IB_USER_VERBS_CMD_REG_MR]		= ib_uverbs_reg_mr,
93 	[IB_USER_VERBS_CMD_REREG_MR]		= ib_uverbs_rereg_mr,
94 	[IB_USER_VERBS_CMD_DEREG_MR]		= ib_uverbs_dereg_mr,
95 	[IB_USER_VERBS_CMD_ALLOC_MW]		= ib_uverbs_alloc_mw,
96 	[IB_USER_VERBS_CMD_DEALLOC_MW]		= ib_uverbs_dealloc_mw,
97 	[IB_USER_VERBS_CMD_CREATE_COMP_CHANNEL] = ib_uverbs_create_comp_channel,
98 	[IB_USER_VERBS_CMD_CREATE_CQ]		= ib_uverbs_create_cq,
99 	[IB_USER_VERBS_CMD_RESIZE_CQ]		= ib_uverbs_resize_cq,
100 	[IB_USER_VERBS_CMD_POLL_CQ]		= ib_uverbs_poll_cq,
101 	[IB_USER_VERBS_CMD_REQ_NOTIFY_CQ]	= ib_uverbs_req_notify_cq,
102 	[IB_USER_VERBS_CMD_DESTROY_CQ]		= ib_uverbs_destroy_cq,
103 	[IB_USER_VERBS_CMD_CREATE_QP]		= ib_uverbs_create_qp,
104 	[IB_USER_VERBS_CMD_QUERY_QP]		= ib_uverbs_query_qp,
105 	[IB_USER_VERBS_CMD_MODIFY_QP]		= ib_uverbs_modify_qp,
106 	[IB_USER_VERBS_CMD_DESTROY_QP]		= ib_uverbs_destroy_qp,
107 	[IB_USER_VERBS_CMD_POST_SEND]		= ib_uverbs_post_send,
108 	[IB_USER_VERBS_CMD_POST_RECV]		= ib_uverbs_post_recv,
109 	[IB_USER_VERBS_CMD_POST_SRQ_RECV]	= ib_uverbs_post_srq_recv,
110 	[IB_USER_VERBS_CMD_CREATE_AH]		= ib_uverbs_create_ah,
111 	[IB_USER_VERBS_CMD_DESTROY_AH]		= ib_uverbs_destroy_ah,
112 	[IB_USER_VERBS_CMD_ATTACH_MCAST]	= ib_uverbs_attach_mcast,
113 	[IB_USER_VERBS_CMD_DETACH_MCAST]	= ib_uverbs_detach_mcast,
114 	[IB_USER_VERBS_CMD_CREATE_SRQ]		= ib_uverbs_create_srq,
115 	[IB_USER_VERBS_CMD_MODIFY_SRQ]		= ib_uverbs_modify_srq,
116 	[IB_USER_VERBS_CMD_QUERY_SRQ]		= ib_uverbs_query_srq,
117 	[IB_USER_VERBS_CMD_DESTROY_SRQ]		= ib_uverbs_destroy_srq,
118 	[IB_USER_VERBS_CMD_OPEN_XRCD]		= ib_uverbs_open_xrcd,
119 	[IB_USER_VERBS_CMD_CLOSE_XRCD]		= ib_uverbs_close_xrcd,
120 	[IB_USER_VERBS_CMD_CREATE_XSRQ]		= ib_uverbs_create_xsrq,
121 	[IB_USER_VERBS_CMD_OPEN_QP]		= ib_uverbs_open_qp,
122 };
123 
124 static int (*uverbs_ex_cmd_table[])(struct ib_uverbs_file *file,
125 				    struct ib_device *ib_dev,
126 				    struct ib_udata *ucore,
127 				    struct ib_udata *uhw) = {
128 	[IB_USER_VERBS_EX_CMD_CREATE_FLOW]	= ib_uverbs_ex_create_flow,
129 	[IB_USER_VERBS_EX_CMD_DESTROY_FLOW]	= ib_uverbs_ex_destroy_flow,
130 	[IB_USER_VERBS_EX_CMD_QUERY_DEVICE]	= ib_uverbs_ex_query_device,
131 	[IB_USER_VERBS_EX_CMD_CREATE_CQ]	= ib_uverbs_ex_create_cq,
132 	[IB_USER_VERBS_EX_CMD_CREATE_QP]        = ib_uverbs_ex_create_qp,
133 };
134 
135 static void ib_uverbs_add_one(struct ib_device *device);
136 static void ib_uverbs_remove_one(struct ib_device *device, void *client_data);
137 
138 int uverbs_dealloc_mw(struct ib_mw *mw)
139 {
140 	struct ib_pd *pd = mw->pd;
141 	int ret;
142 
143 	ret = mw->device->dealloc_mw(mw);
144 	if (!ret)
145 		atomic_dec(&pd->usecnt);
146 	return ret;
147 }
148 
149 static void ib_uverbs_release_dev(struct kobject *kobj)
150 {
151 	struct ib_uverbs_device *dev =
152 		container_of(kobj, struct ib_uverbs_device, kobj);
153 
154 	cleanup_srcu_struct(&dev->disassociate_srcu);
155 	kfree(dev);
156 }
157 
158 static struct kobj_type ib_uverbs_dev_ktype = {
159 	.release = ib_uverbs_release_dev,
160 };
161 
162 static void ib_uverbs_release_event_file(struct kref *ref)
163 {
164 	struct ib_uverbs_event_file *file =
165 		container_of(ref, struct ib_uverbs_event_file, ref);
166 
167 	kfree(file);
168 }
169 
170 void ib_uverbs_release_ucq(struct ib_uverbs_file *file,
171 			  struct ib_uverbs_event_file *ev_file,
172 			  struct ib_ucq_object *uobj)
173 {
174 	struct ib_uverbs_event *evt, *tmp;
175 
176 	if (ev_file) {
177 		spin_lock_irq(&ev_file->lock);
178 		list_for_each_entry_safe(evt, tmp, &uobj->comp_list, obj_list) {
179 			list_del(&evt->list);
180 			kfree(evt);
181 		}
182 		spin_unlock_irq(&ev_file->lock);
183 
184 		kref_put(&ev_file->ref, ib_uverbs_release_event_file);
185 	}
186 
187 	spin_lock_irq(&file->async_file->lock);
188 	list_for_each_entry_safe(evt, tmp, &uobj->async_list, obj_list) {
189 		list_del(&evt->list);
190 		kfree(evt);
191 	}
192 	spin_unlock_irq(&file->async_file->lock);
193 }
194 
195 void ib_uverbs_release_uevent(struct ib_uverbs_file *file,
196 			      struct ib_uevent_object *uobj)
197 {
198 	struct ib_uverbs_event *evt, *tmp;
199 
200 	spin_lock_irq(&file->async_file->lock);
201 	list_for_each_entry_safe(evt, tmp, &uobj->event_list, obj_list) {
202 		list_del(&evt->list);
203 		kfree(evt);
204 	}
205 	spin_unlock_irq(&file->async_file->lock);
206 }
207 
208 static void ib_uverbs_detach_umcast(struct ib_qp *qp,
209 				    struct ib_uqp_object *uobj)
210 {
211 	struct ib_uverbs_mcast_entry *mcast, *tmp;
212 
213 	list_for_each_entry_safe(mcast, tmp, &uobj->mcast_list, list) {
214 		ib_detach_mcast(qp, &mcast->gid, mcast->lid);
215 		list_del(&mcast->list);
216 		kfree(mcast);
217 	}
218 }
219 
220 static int ib_uverbs_cleanup_ucontext(struct ib_uverbs_file *file,
221 				      struct ib_ucontext *context)
222 {
223 	struct ib_uobject *uobj, *tmp;
224 
225 	context->closing = 1;
226 
227 	list_for_each_entry_safe(uobj, tmp, &context->ah_list, list) {
228 		struct ib_ah *ah = uobj->object;
229 
230 		idr_remove_uobj(&ib_uverbs_ah_idr, uobj);
231 		ib_destroy_ah(ah);
232 		kfree(uobj);
233 	}
234 
235 	/* Remove MWs before QPs, in order to support type 2A MWs. */
236 	list_for_each_entry_safe(uobj, tmp, &context->mw_list, list) {
237 		struct ib_mw *mw = uobj->object;
238 
239 		idr_remove_uobj(&ib_uverbs_mw_idr, uobj);
240 		uverbs_dealloc_mw(mw);
241 		kfree(uobj);
242 	}
243 
244 	list_for_each_entry_safe(uobj, tmp, &context->rule_list, list) {
245 		struct ib_flow *flow_id = uobj->object;
246 
247 		idr_remove_uobj(&ib_uverbs_rule_idr, uobj);
248 		ib_destroy_flow(flow_id);
249 		kfree(uobj);
250 	}
251 
252 	list_for_each_entry_safe(uobj, tmp, &context->qp_list, list) {
253 		struct ib_qp *qp = uobj->object;
254 		struct ib_uqp_object *uqp =
255 			container_of(uobj, struct ib_uqp_object, uevent.uobject);
256 
257 		idr_remove_uobj(&ib_uverbs_qp_idr, uobj);
258 		if (qp != qp->real_qp) {
259 			ib_close_qp(qp);
260 		} else {
261 			ib_uverbs_detach_umcast(qp, uqp);
262 			ib_destroy_qp(qp);
263 		}
264 		ib_uverbs_release_uevent(file, &uqp->uevent);
265 		kfree(uqp);
266 	}
267 
268 	list_for_each_entry_safe(uobj, tmp, &context->srq_list, list) {
269 		struct ib_srq *srq = uobj->object;
270 		struct ib_uevent_object *uevent =
271 			container_of(uobj, struct ib_uevent_object, uobject);
272 
273 		idr_remove_uobj(&ib_uverbs_srq_idr, uobj);
274 		ib_destroy_srq(srq);
275 		ib_uverbs_release_uevent(file, uevent);
276 		kfree(uevent);
277 	}
278 
279 	list_for_each_entry_safe(uobj, tmp, &context->cq_list, list) {
280 		struct ib_cq *cq = uobj->object;
281 		struct ib_uverbs_event_file *ev_file = cq->cq_context;
282 		struct ib_ucq_object *ucq =
283 			container_of(uobj, struct ib_ucq_object, uobject);
284 
285 		idr_remove_uobj(&ib_uverbs_cq_idr, uobj);
286 		ib_destroy_cq(cq);
287 		ib_uverbs_release_ucq(file, ev_file, ucq);
288 		kfree(ucq);
289 	}
290 
291 	list_for_each_entry_safe(uobj, tmp, &context->mr_list, list) {
292 		struct ib_mr *mr = uobj->object;
293 
294 		idr_remove_uobj(&ib_uverbs_mr_idr, uobj);
295 		ib_dereg_mr(mr);
296 		kfree(uobj);
297 	}
298 
299 	mutex_lock(&file->device->xrcd_tree_mutex);
300 	list_for_each_entry_safe(uobj, tmp, &context->xrcd_list, list) {
301 		struct ib_xrcd *xrcd = uobj->object;
302 		struct ib_uxrcd_object *uxrcd =
303 			container_of(uobj, struct ib_uxrcd_object, uobject);
304 
305 		idr_remove_uobj(&ib_uverbs_xrcd_idr, uobj);
306 		ib_uverbs_dealloc_xrcd(file->device, xrcd);
307 		kfree(uxrcd);
308 	}
309 	mutex_unlock(&file->device->xrcd_tree_mutex);
310 
311 	list_for_each_entry_safe(uobj, tmp, &context->pd_list, list) {
312 		struct ib_pd *pd = uobj->object;
313 
314 		idr_remove_uobj(&ib_uverbs_pd_idr, uobj);
315 		ib_dealloc_pd(pd);
316 		kfree(uobj);
317 	}
318 
319 	put_pid(context->tgid);
320 
321 	return context->device->dealloc_ucontext(context);
322 }
323 
324 static void ib_uverbs_comp_dev(struct ib_uverbs_device *dev)
325 {
326 	complete(&dev->comp);
327 }
328 
329 static void ib_uverbs_release_file(struct kref *ref)
330 {
331 	struct ib_uverbs_file *file =
332 		container_of(ref, struct ib_uverbs_file, ref);
333 	struct ib_device *ib_dev;
334 	int srcu_key;
335 
336 	srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
337 	ib_dev = srcu_dereference(file->device->ib_dev,
338 				  &file->device->disassociate_srcu);
339 	if (ib_dev && !ib_dev->disassociate_ucontext)
340 		module_put(ib_dev->owner);
341 	srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
342 
343 	if (atomic_dec_and_test(&file->device->refcount))
344 		ib_uverbs_comp_dev(file->device);
345 
346 	kfree(file);
347 }
348 
349 static ssize_t ib_uverbs_event_read(struct file *filp, char __user *buf,
350 				    size_t count, loff_t *pos)
351 {
352 	struct ib_uverbs_event_file *file = filp->private_data;
353 	struct ib_uverbs_event *event;
354 	int eventsz;
355 	int ret = 0;
356 
357 	spin_lock_irq(&file->lock);
358 
359 	while (list_empty(&file->event_list)) {
360 		spin_unlock_irq(&file->lock);
361 
362 		if (filp->f_flags & O_NONBLOCK)
363 			return -EAGAIN;
364 
365 		if (wait_event_interruptible(file->poll_wait,
366 					     (!list_empty(&file->event_list) ||
367 			/* The barriers built into wait_event_interruptible()
368 			 * and wake_up() guarentee this will see the null set
369 			 * without using RCU
370 			 */
371 					     !file->uverbs_file->device->ib_dev)))
372 			return -ERESTARTSYS;
373 
374 		/* If device was disassociated and no event exists set an error */
375 		if (list_empty(&file->event_list) &&
376 		    !file->uverbs_file->device->ib_dev)
377 			return -EIO;
378 
379 		spin_lock_irq(&file->lock);
380 	}
381 
382 	event = list_entry(file->event_list.next, struct ib_uverbs_event, list);
383 
384 	if (file->is_async)
385 		eventsz = sizeof (struct ib_uverbs_async_event_desc);
386 	else
387 		eventsz = sizeof (struct ib_uverbs_comp_event_desc);
388 
389 	if (eventsz > count) {
390 		ret   = -EINVAL;
391 		event = NULL;
392 	} else {
393 		list_del(file->event_list.next);
394 		if (event->counter) {
395 			++(*event->counter);
396 			list_del(&event->obj_list);
397 		}
398 	}
399 
400 	spin_unlock_irq(&file->lock);
401 
402 	if (event) {
403 		if (copy_to_user(buf, event, eventsz))
404 			ret = -EFAULT;
405 		else
406 			ret = eventsz;
407 	}
408 
409 	kfree(event);
410 
411 	return ret;
412 }
413 
414 static unsigned int ib_uverbs_event_poll(struct file *filp,
415 					 struct poll_table_struct *wait)
416 {
417 	unsigned int pollflags = 0;
418 	struct ib_uverbs_event_file *file = filp->private_data;
419 
420 	poll_wait(filp, &file->poll_wait, wait);
421 
422 	spin_lock_irq(&file->lock);
423 	if (!list_empty(&file->event_list))
424 		pollflags = POLLIN | POLLRDNORM;
425 	spin_unlock_irq(&file->lock);
426 
427 	return pollflags;
428 }
429 
430 static int ib_uverbs_event_fasync(int fd, struct file *filp, int on)
431 {
432 	struct ib_uverbs_event_file *file = filp->private_data;
433 
434 	return fasync_helper(fd, filp, on, &file->async_queue);
435 }
436 
437 static int ib_uverbs_event_close(struct inode *inode, struct file *filp)
438 {
439 	struct ib_uverbs_event_file *file = filp->private_data;
440 	struct ib_uverbs_event *entry, *tmp;
441 	int closed_already = 0;
442 
443 	mutex_lock(&file->uverbs_file->device->lists_mutex);
444 	spin_lock_irq(&file->lock);
445 	closed_already = file->is_closed;
446 	file->is_closed = 1;
447 	list_for_each_entry_safe(entry, tmp, &file->event_list, list) {
448 		if (entry->counter)
449 			list_del(&entry->obj_list);
450 		kfree(entry);
451 	}
452 	spin_unlock_irq(&file->lock);
453 	if (!closed_already) {
454 		list_del(&file->list);
455 		if (file->is_async)
456 			ib_unregister_event_handler(&file->uverbs_file->
457 				event_handler);
458 	}
459 	mutex_unlock(&file->uverbs_file->device->lists_mutex);
460 
461 	kref_put(&file->uverbs_file->ref, ib_uverbs_release_file);
462 	kref_put(&file->ref, ib_uverbs_release_event_file);
463 
464 	return 0;
465 }
466 
467 static const struct file_operations uverbs_event_fops = {
468 	.owner	 = THIS_MODULE,
469 	.read	 = ib_uverbs_event_read,
470 	.poll    = ib_uverbs_event_poll,
471 	.release = ib_uverbs_event_close,
472 	.fasync  = ib_uverbs_event_fasync,
473 	.llseek	 = no_llseek,
474 };
475 
476 void ib_uverbs_comp_handler(struct ib_cq *cq, void *cq_context)
477 {
478 	struct ib_uverbs_event_file    *file = cq_context;
479 	struct ib_ucq_object	       *uobj;
480 	struct ib_uverbs_event	       *entry;
481 	unsigned long			flags;
482 
483 	if (!file)
484 		return;
485 
486 	spin_lock_irqsave(&file->lock, flags);
487 	if (file->is_closed) {
488 		spin_unlock_irqrestore(&file->lock, flags);
489 		return;
490 	}
491 
492 	entry = kmalloc(sizeof *entry, GFP_ATOMIC);
493 	if (!entry) {
494 		spin_unlock_irqrestore(&file->lock, flags);
495 		return;
496 	}
497 
498 	uobj = container_of(cq->uobject, struct ib_ucq_object, uobject);
499 
500 	entry->desc.comp.cq_handle = cq->uobject->user_handle;
501 	entry->counter		   = &uobj->comp_events_reported;
502 
503 	list_add_tail(&entry->list, &file->event_list);
504 	list_add_tail(&entry->obj_list, &uobj->comp_list);
505 	spin_unlock_irqrestore(&file->lock, flags);
506 
507 	wake_up_interruptible(&file->poll_wait);
508 	kill_fasync(&file->async_queue, SIGIO, POLL_IN);
509 }
510 
511 static void ib_uverbs_async_handler(struct ib_uverbs_file *file,
512 				    __u64 element, __u64 event,
513 				    struct list_head *obj_list,
514 				    u32 *counter)
515 {
516 	struct ib_uverbs_event *entry;
517 	unsigned long flags;
518 
519 	spin_lock_irqsave(&file->async_file->lock, flags);
520 	if (file->async_file->is_closed) {
521 		spin_unlock_irqrestore(&file->async_file->lock, flags);
522 		return;
523 	}
524 
525 	entry = kmalloc(sizeof *entry, GFP_ATOMIC);
526 	if (!entry) {
527 		spin_unlock_irqrestore(&file->async_file->lock, flags);
528 		return;
529 	}
530 
531 	entry->desc.async.element    = element;
532 	entry->desc.async.event_type = event;
533 	entry->desc.async.reserved   = 0;
534 	entry->counter               = counter;
535 
536 	list_add_tail(&entry->list, &file->async_file->event_list);
537 	if (obj_list)
538 		list_add_tail(&entry->obj_list, obj_list);
539 	spin_unlock_irqrestore(&file->async_file->lock, flags);
540 
541 	wake_up_interruptible(&file->async_file->poll_wait);
542 	kill_fasync(&file->async_file->async_queue, SIGIO, POLL_IN);
543 }
544 
545 void ib_uverbs_cq_event_handler(struct ib_event *event, void *context_ptr)
546 {
547 	struct ib_ucq_object *uobj = container_of(event->element.cq->uobject,
548 						  struct ib_ucq_object, uobject);
549 
550 	ib_uverbs_async_handler(uobj->uverbs_file, uobj->uobject.user_handle,
551 				event->event, &uobj->async_list,
552 				&uobj->async_events_reported);
553 }
554 
555 void ib_uverbs_qp_event_handler(struct ib_event *event, void *context_ptr)
556 {
557 	struct ib_uevent_object *uobj;
558 
559 	/* for XRC target qp's, check that qp is live */
560 	if (!event->element.qp->uobject || !event->element.qp->uobject->live)
561 		return;
562 
563 	uobj = container_of(event->element.qp->uobject,
564 			    struct ib_uevent_object, uobject);
565 
566 	ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
567 				event->event, &uobj->event_list,
568 				&uobj->events_reported);
569 }
570 
571 void ib_uverbs_srq_event_handler(struct ib_event *event, void *context_ptr)
572 {
573 	struct ib_uevent_object *uobj;
574 
575 	uobj = container_of(event->element.srq->uobject,
576 			    struct ib_uevent_object, uobject);
577 
578 	ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
579 				event->event, &uobj->event_list,
580 				&uobj->events_reported);
581 }
582 
583 void ib_uverbs_event_handler(struct ib_event_handler *handler,
584 			     struct ib_event *event)
585 {
586 	struct ib_uverbs_file *file =
587 		container_of(handler, struct ib_uverbs_file, event_handler);
588 
589 	ib_uverbs_async_handler(file, event->element.port_num, event->event,
590 				NULL, NULL);
591 }
592 
593 void ib_uverbs_free_async_event_file(struct ib_uverbs_file *file)
594 {
595 	kref_put(&file->async_file->ref, ib_uverbs_release_event_file);
596 	file->async_file = NULL;
597 }
598 
599 struct file *ib_uverbs_alloc_event_file(struct ib_uverbs_file *uverbs_file,
600 					struct ib_device	*ib_dev,
601 					int is_async)
602 {
603 	struct ib_uverbs_event_file *ev_file;
604 	struct file *filp;
605 	int ret;
606 
607 	ev_file = kzalloc(sizeof(*ev_file), GFP_KERNEL);
608 	if (!ev_file)
609 		return ERR_PTR(-ENOMEM);
610 
611 	kref_init(&ev_file->ref);
612 	spin_lock_init(&ev_file->lock);
613 	INIT_LIST_HEAD(&ev_file->event_list);
614 	init_waitqueue_head(&ev_file->poll_wait);
615 	ev_file->uverbs_file = uverbs_file;
616 	kref_get(&ev_file->uverbs_file->ref);
617 	ev_file->async_queue = NULL;
618 	ev_file->is_closed   = 0;
619 
620 	filp = anon_inode_getfile("[infinibandevent]", &uverbs_event_fops,
621 				  ev_file, O_RDONLY);
622 	if (IS_ERR(filp))
623 		goto err_put_refs;
624 
625 	mutex_lock(&uverbs_file->device->lists_mutex);
626 	list_add_tail(&ev_file->list,
627 		      &uverbs_file->device->uverbs_events_file_list);
628 	mutex_unlock(&uverbs_file->device->lists_mutex);
629 
630 	if (is_async) {
631 		WARN_ON(uverbs_file->async_file);
632 		uverbs_file->async_file = ev_file;
633 		kref_get(&uverbs_file->async_file->ref);
634 		INIT_IB_EVENT_HANDLER(&uverbs_file->event_handler,
635 				      ib_dev,
636 				      ib_uverbs_event_handler);
637 		ret = ib_register_event_handler(&uverbs_file->event_handler);
638 		if (ret)
639 			goto err_put_file;
640 
641 		/* At that point async file stuff was fully set */
642 		ev_file->is_async = 1;
643 	}
644 
645 	return filp;
646 
647 err_put_file:
648 	fput(filp);
649 	kref_put(&uverbs_file->async_file->ref, ib_uverbs_release_event_file);
650 	uverbs_file->async_file = NULL;
651 	return ERR_PTR(ret);
652 
653 err_put_refs:
654 	kref_put(&ev_file->uverbs_file->ref, ib_uverbs_release_file);
655 	kref_put(&ev_file->ref, ib_uverbs_release_event_file);
656 	return filp;
657 }
658 
659 /*
660  * Look up a completion event file by FD.  If lookup is successful,
661  * takes a ref to the event file struct that it returns; if
662  * unsuccessful, returns NULL.
663  */
664 struct ib_uverbs_event_file *ib_uverbs_lookup_comp_file(int fd)
665 {
666 	struct ib_uverbs_event_file *ev_file = NULL;
667 	struct fd f = fdget(fd);
668 
669 	if (!f.file)
670 		return NULL;
671 
672 	if (f.file->f_op != &uverbs_event_fops)
673 		goto out;
674 
675 	ev_file = f.file->private_data;
676 	if (ev_file->is_async) {
677 		ev_file = NULL;
678 		goto out;
679 	}
680 
681 	kref_get(&ev_file->ref);
682 
683 out:
684 	fdput(f);
685 	return ev_file;
686 }
687 
688 static int verify_command_mask(struct ib_device *ib_dev, __u32 command)
689 {
690 	u64 mask;
691 
692 	if (command <= IB_USER_VERBS_CMD_OPEN_QP)
693 		mask = ib_dev->uverbs_cmd_mask;
694 	else
695 		mask = ib_dev->uverbs_ex_cmd_mask;
696 
697 	if (mask & ((u64)1 << command))
698 		return 0;
699 
700 	return -1;
701 }
702 
703 static ssize_t ib_uverbs_write(struct file *filp, const char __user *buf,
704 			     size_t count, loff_t *pos)
705 {
706 	struct ib_uverbs_file *file = filp->private_data;
707 	struct ib_device *ib_dev;
708 	struct ib_uverbs_cmd_hdr hdr;
709 	__u32 command;
710 	__u32 flags;
711 	int srcu_key;
712 	ssize_t ret;
713 
714 	if (WARN_ON_ONCE(!ib_safe_file_access(filp)))
715 		return -EACCES;
716 
717 	if (count < sizeof hdr)
718 		return -EINVAL;
719 
720 	if (copy_from_user(&hdr, buf, sizeof hdr))
721 		return -EFAULT;
722 
723 	srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
724 	ib_dev = srcu_dereference(file->device->ib_dev,
725 				  &file->device->disassociate_srcu);
726 	if (!ib_dev) {
727 		ret = -EIO;
728 		goto out;
729 	}
730 
731 	if (hdr.command & ~(__u32)(IB_USER_VERBS_CMD_FLAGS_MASK |
732 				   IB_USER_VERBS_CMD_COMMAND_MASK)) {
733 		ret = -EINVAL;
734 		goto out;
735 	}
736 
737 	command = hdr.command & IB_USER_VERBS_CMD_COMMAND_MASK;
738 	if (verify_command_mask(ib_dev, command)) {
739 		ret = -EOPNOTSUPP;
740 		goto out;
741 	}
742 
743 	if (!file->ucontext &&
744 	    command != IB_USER_VERBS_CMD_GET_CONTEXT) {
745 		ret = -EINVAL;
746 		goto out;
747 	}
748 
749 	flags = (hdr.command &
750 		 IB_USER_VERBS_CMD_FLAGS_MASK) >> IB_USER_VERBS_CMD_FLAGS_SHIFT;
751 
752 	if (!flags) {
753 		if (command >= ARRAY_SIZE(uverbs_cmd_table) ||
754 		    !uverbs_cmd_table[command]) {
755 			ret = -EINVAL;
756 			goto out;
757 		}
758 
759 		if (hdr.in_words * 4 != count) {
760 			ret = -EINVAL;
761 			goto out;
762 		}
763 
764 		ret = uverbs_cmd_table[command](file, ib_dev,
765 						 buf + sizeof(hdr),
766 						 hdr.in_words * 4,
767 						 hdr.out_words * 4);
768 
769 	} else if (flags == IB_USER_VERBS_CMD_FLAG_EXTENDED) {
770 		struct ib_uverbs_ex_cmd_hdr ex_hdr;
771 		struct ib_udata ucore;
772 		struct ib_udata uhw;
773 		size_t written_count = count;
774 
775 		if (command >= ARRAY_SIZE(uverbs_ex_cmd_table) ||
776 		    !uverbs_ex_cmd_table[command]) {
777 			ret = -ENOSYS;
778 			goto out;
779 		}
780 
781 		if (!file->ucontext) {
782 			ret = -EINVAL;
783 			goto out;
784 		}
785 
786 		if (count < (sizeof(hdr) + sizeof(ex_hdr))) {
787 			ret = -EINVAL;
788 			goto out;
789 		}
790 
791 		if (copy_from_user(&ex_hdr, buf + sizeof(hdr), sizeof(ex_hdr))) {
792 			ret = -EFAULT;
793 			goto out;
794 		}
795 
796 		count -= sizeof(hdr) + sizeof(ex_hdr);
797 		buf += sizeof(hdr) + sizeof(ex_hdr);
798 
799 		if ((hdr.in_words + ex_hdr.provider_in_words) * 8 != count) {
800 			ret = -EINVAL;
801 			goto out;
802 		}
803 
804 		if (ex_hdr.cmd_hdr_reserved) {
805 			ret = -EINVAL;
806 			goto out;
807 		}
808 
809 		if (ex_hdr.response) {
810 			if (!hdr.out_words && !ex_hdr.provider_out_words) {
811 				ret = -EINVAL;
812 				goto out;
813 			}
814 
815 			if (!access_ok(VERIFY_WRITE,
816 				       (void __user *) (unsigned long) ex_hdr.response,
817 				       (hdr.out_words + ex_hdr.provider_out_words) * 8)) {
818 				ret = -EFAULT;
819 				goto out;
820 			}
821 		} else {
822 			if (hdr.out_words || ex_hdr.provider_out_words) {
823 				ret = -EINVAL;
824 				goto out;
825 			}
826 		}
827 
828 		INIT_UDATA_BUF_OR_NULL(&ucore, buf, (unsigned long) ex_hdr.response,
829 				       hdr.in_words * 8, hdr.out_words * 8);
830 
831 		INIT_UDATA_BUF_OR_NULL(&uhw,
832 				       buf + ucore.inlen,
833 				       (unsigned long) ex_hdr.response + ucore.outlen,
834 				       ex_hdr.provider_in_words * 8,
835 				       ex_hdr.provider_out_words * 8);
836 
837 		ret = uverbs_ex_cmd_table[command](file,
838 						   ib_dev,
839 						   &ucore,
840 						   &uhw);
841 		if (!ret)
842 			ret = written_count;
843 	} else {
844 		ret = -ENOSYS;
845 	}
846 
847 out:
848 	srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
849 	return ret;
850 }
851 
852 static int ib_uverbs_mmap(struct file *filp, struct vm_area_struct *vma)
853 {
854 	struct ib_uverbs_file *file = filp->private_data;
855 	struct ib_device *ib_dev;
856 	int ret = 0;
857 	int srcu_key;
858 
859 	srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
860 	ib_dev = srcu_dereference(file->device->ib_dev,
861 				  &file->device->disassociate_srcu);
862 	if (!ib_dev) {
863 		ret = -EIO;
864 		goto out;
865 	}
866 
867 	if (!file->ucontext)
868 		ret = -ENODEV;
869 	else
870 		ret = ib_dev->mmap(file->ucontext, vma);
871 out:
872 	srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
873 	return ret;
874 }
875 
876 /*
877  * ib_uverbs_open() does not need the BKL:
878  *
879  *  - the ib_uverbs_device structures are properly reference counted and
880  *    everything else is purely local to the file being created, so
881  *    races against other open calls are not a problem;
882  *  - there is no ioctl method to race against;
883  *  - the open method will either immediately run -ENXIO, or all
884  *    required initialization will be done.
885  */
886 static int ib_uverbs_open(struct inode *inode, struct file *filp)
887 {
888 	struct ib_uverbs_device *dev;
889 	struct ib_uverbs_file *file;
890 	struct ib_device *ib_dev;
891 	int ret;
892 	int module_dependent;
893 	int srcu_key;
894 
895 	dev = container_of(inode->i_cdev, struct ib_uverbs_device, cdev);
896 	if (!atomic_inc_not_zero(&dev->refcount))
897 		return -ENXIO;
898 
899 	srcu_key = srcu_read_lock(&dev->disassociate_srcu);
900 	mutex_lock(&dev->lists_mutex);
901 	ib_dev = srcu_dereference(dev->ib_dev,
902 				  &dev->disassociate_srcu);
903 	if (!ib_dev) {
904 		ret = -EIO;
905 		goto err;
906 	}
907 
908 	/* In case IB device supports disassociate ucontext, there is no hard
909 	 * dependency between uverbs device and its low level device.
910 	 */
911 	module_dependent = !(ib_dev->disassociate_ucontext);
912 
913 	if (module_dependent) {
914 		if (!try_module_get(ib_dev->owner)) {
915 			ret = -ENODEV;
916 			goto err;
917 		}
918 	}
919 
920 	file = kzalloc(sizeof(*file), GFP_KERNEL);
921 	if (!file) {
922 		ret = -ENOMEM;
923 		if (module_dependent)
924 			goto err_module;
925 
926 		goto err;
927 	}
928 
929 	file->device	 = dev;
930 	file->ucontext	 = NULL;
931 	file->async_file = NULL;
932 	kref_init(&file->ref);
933 	mutex_init(&file->mutex);
934 
935 	filp->private_data = file;
936 	kobject_get(&dev->kobj);
937 	list_add_tail(&file->list, &dev->uverbs_file_list);
938 	mutex_unlock(&dev->lists_mutex);
939 	srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
940 
941 	return nonseekable_open(inode, filp);
942 
943 err_module:
944 	module_put(ib_dev->owner);
945 
946 err:
947 	mutex_unlock(&dev->lists_mutex);
948 	srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
949 	if (atomic_dec_and_test(&dev->refcount))
950 		ib_uverbs_comp_dev(dev);
951 
952 	return ret;
953 }
954 
955 static int ib_uverbs_close(struct inode *inode, struct file *filp)
956 {
957 	struct ib_uverbs_file *file = filp->private_data;
958 	struct ib_uverbs_device *dev = file->device;
959 	struct ib_ucontext *ucontext = NULL;
960 
961 	mutex_lock(&file->device->lists_mutex);
962 	ucontext = file->ucontext;
963 	file->ucontext = NULL;
964 	if (!file->is_closed) {
965 		list_del(&file->list);
966 		file->is_closed = 1;
967 	}
968 	mutex_unlock(&file->device->lists_mutex);
969 	if (ucontext)
970 		ib_uverbs_cleanup_ucontext(file, ucontext);
971 
972 	if (file->async_file)
973 		kref_put(&file->async_file->ref, ib_uverbs_release_event_file);
974 
975 	kref_put(&file->ref, ib_uverbs_release_file);
976 	kobject_put(&dev->kobj);
977 
978 	return 0;
979 }
980 
981 static const struct file_operations uverbs_fops = {
982 	.owner	 = THIS_MODULE,
983 	.write	 = ib_uverbs_write,
984 	.open	 = ib_uverbs_open,
985 	.release = ib_uverbs_close,
986 	.llseek	 = no_llseek,
987 };
988 
989 static const struct file_operations uverbs_mmap_fops = {
990 	.owner	 = THIS_MODULE,
991 	.write	 = ib_uverbs_write,
992 	.mmap    = ib_uverbs_mmap,
993 	.open	 = ib_uverbs_open,
994 	.release = ib_uverbs_close,
995 	.llseek	 = no_llseek,
996 };
997 
998 static struct ib_client uverbs_client = {
999 	.name   = "uverbs",
1000 	.add    = ib_uverbs_add_one,
1001 	.remove = ib_uverbs_remove_one
1002 };
1003 
1004 static ssize_t show_ibdev(struct device *device, struct device_attribute *attr,
1005 			  char *buf)
1006 {
1007 	int ret = -ENODEV;
1008 	int srcu_key;
1009 	struct ib_uverbs_device *dev = dev_get_drvdata(device);
1010 	struct ib_device *ib_dev;
1011 
1012 	if (!dev)
1013 		return -ENODEV;
1014 
1015 	srcu_key = srcu_read_lock(&dev->disassociate_srcu);
1016 	ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu);
1017 	if (ib_dev)
1018 		ret = sprintf(buf, "%s\n", ib_dev->name);
1019 	srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1020 
1021 	return ret;
1022 }
1023 static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
1024 
1025 static ssize_t show_dev_abi_version(struct device *device,
1026 				    struct device_attribute *attr, char *buf)
1027 {
1028 	struct ib_uverbs_device *dev = dev_get_drvdata(device);
1029 	int ret = -ENODEV;
1030 	int srcu_key;
1031 	struct ib_device *ib_dev;
1032 
1033 	if (!dev)
1034 		return -ENODEV;
1035 	srcu_key = srcu_read_lock(&dev->disassociate_srcu);
1036 	ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu);
1037 	if (ib_dev)
1038 		ret = sprintf(buf, "%d\n", ib_dev->uverbs_abi_ver);
1039 	srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1040 
1041 	return ret;
1042 }
1043 static DEVICE_ATTR(abi_version, S_IRUGO, show_dev_abi_version, NULL);
1044 
1045 static CLASS_ATTR_STRING(abi_version, S_IRUGO,
1046 			 __stringify(IB_USER_VERBS_ABI_VERSION));
1047 
1048 static dev_t overflow_maj;
1049 static DECLARE_BITMAP(overflow_map, IB_UVERBS_MAX_DEVICES);
1050 
1051 /*
1052  * If we have more than IB_UVERBS_MAX_DEVICES, dynamically overflow by
1053  * requesting a new major number and doubling the number of max devices we
1054  * support. It's stupid, but simple.
1055  */
1056 static int find_overflow_devnum(void)
1057 {
1058 	int ret;
1059 
1060 	if (!overflow_maj) {
1061 		ret = alloc_chrdev_region(&overflow_maj, 0, IB_UVERBS_MAX_DEVICES,
1062 					  "infiniband_verbs");
1063 		if (ret) {
1064 			pr_err("user_verbs: couldn't register dynamic device number\n");
1065 			return ret;
1066 		}
1067 	}
1068 
1069 	ret = find_first_zero_bit(overflow_map, IB_UVERBS_MAX_DEVICES);
1070 	if (ret >= IB_UVERBS_MAX_DEVICES)
1071 		return -1;
1072 
1073 	return ret;
1074 }
1075 
1076 static void ib_uverbs_add_one(struct ib_device *device)
1077 {
1078 	int devnum;
1079 	dev_t base;
1080 	struct ib_uverbs_device *uverbs_dev;
1081 	int ret;
1082 
1083 	if (!device->alloc_ucontext)
1084 		return;
1085 
1086 	uverbs_dev = kzalloc(sizeof *uverbs_dev, GFP_KERNEL);
1087 	if (!uverbs_dev)
1088 		return;
1089 
1090 	ret = init_srcu_struct(&uverbs_dev->disassociate_srcu);
1091 	if (ret) {
1092 		kfree(uverbs_dev);
1093 		return;
1094 	}
1095 
1096 	atomic_set(&uverbs_dev->refcount, 1);
1097 	init_completion(&uverbs_dev->comp);
1098 	uverbs_dev->xrcd_tree = RB_ROOT;
1099 	mutex_init(&uverbs_dev->xrcd_tree_mutex);
1100 	kobject_init(&uverbs_dev->kobj, &ib_uverbs_dev_ktype);
1101 	mutex_init(&uverbs_dev->lists_mutex);
1102 	INIT_LIST_HEAD(&uverbs_dev->uverbs_file_list);
1103 	INIT_LIST_HEAD(&uverbs_dev->uverbs_events_file_list);
1104 
1105 	spin_lock(&map_lock);
1106 	devnum = find_first_zero_bit(dev_map, IB_UVERBS_MAX_DEVICES);
1107 	if (devnum >= IB_UVERBS_MAX_DEVICES) {
1108 		spin_unlock(&map_lock);
1109 		devnum = find_overflow_devnum();
1110 		if (devnum < 0)
1111 			goto err;
1112 
1113 		spin_lock(&map_lock);
1114 		uverbs_dev->devnum = devnum + IB_UVERBS_MAX_DEVICES;
1115 		base = devnum + overflow_maj;
1116 		set_bit(devnum, overflow_map);
1117 	} else {
1118 		uverbs_dev->devnum = devnum;
1119 		base = devnum + IB_UVERBS_BASE_DEV;
1120 		set_bit(devnum, dev_map);
1121 	}
1122 	spin_unlock(&map_lock);
1123 
1124 	rcu_assign_pointer(uverbs_dev->ib_dev, device);
1125 	uverbs_dev->num_comp_vectors = device->num_comp_vectors;
1126 
1127 	cdev_init(&uverbs_dev->cdev, NULL);
1128 	uverbs_dev->cdev.owner = THIS_MODULE;
1129 	uverbs_dev->cdev.ops = device->mmap ? &uverbs_mmap_fops : &uverbs_fops;
1130 	uverbs_dev->cdev.kobj.parent = &uverbs_dev->kobj;
1131 	kobject_set_name(&uverbs_dev->cdev.kobj, "uverbs%d", uverbs_dev->devnum);
1132 	if (cdev_add(&uverbs_dev->cdev, base, 1))
1133 		goto err_cdev;
1134 
1135 	uverbs_dev->dev = device_create(uverbs_class, device->dma_device,
1136 					uverbs_dev->cdev.dev, uverbs_dev,
1137 					"uverbs%d", uverbs_dev->devnum);
1138 	if (IS_ERR(uverbs_dev->dev))
1139 		goto err_cdev;
1140 
1141 	if (device_create_file(uverbs_dev->dev, &dev_attr_ibdev))
1142 		goto err_class;
1143 	if (device_create_file(uverbs_dev->dev, &dev_attr_abi_version))
1144 		goto err_class;
1145 
1146 	ib_set_client_data(device, &uverbs_client, uverbs_dev);
1147 
1148 	return;
1149 
1150 err_class:
1151 	device_destroy(uverbs_class, uverbs_dev->cdev.dev);
1152 
1153 err_cdev:
1154 	cdev_del(&uverbs_dev->cdev);
1155 	if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
1156 		clear_bit(devnum, dev_map);
1157 	else
1158 		clear_bit(devnum, overflow_map);
1159 
1160 err:
1161 	if (atomic_dec_and_test(&uverbs_dev->refcount))
1162 		ib_uverbs_comp_dev(uverbs_dev);
1163 	wait_for_completion(&uverbs_dev->comp);
1164 	kobject_put(&uverbs_dev->kobj);
1165 	return;
1166 }
1167 
1168 static void ib_uverbs_free_hw_resources(struct ib_uverbs_device *uverbs_dev,
1169 					struct ib_device *ib_dev)
1170 {
1171 	struct ib_uverbs_file *file;
1172 	struct ib_uverbs_event_file *event_file;
1173 	struct ib_event event;
1174 
1175 	/* Pending running commands to terminate */
1176 	synchronize_srcu(&uverbs_dev->disassociate_srcu);
1177 	event.event = IB_EVENT_DEVICE_FATAL;
1178 	event.element.port_num = 0;
1179 	event.device = ib_dev;
1180 
1181 	mutex_lock(&uverbs_dev->lists_mutex);
1182 	while (!list_empty(&uverbs_dev->uverbs_file_list)) {
1183 		struct ib_ucontext *ucontext;
1184 
1185 		file = list_first_entry(&uverbs_dev->uverbs_file_list,
1186 					struct ib_uverbs_file, list);
1187 		file->is_closed = 1;
1188 		ucontext = file->ucontext;
1189 		list_del(&file->list);
1190 		file->ucontext = NULL;
1191 		kref_get(&file->ref);
1192 		mutex_unlock(&uverbs_dev->lists_mutex);
1193 		/* We must release the mutex before going ahead and calling
1194 		 * disassociate_ucontext. disassociate_ucontext might end up
1195 		 * indirectly calling uverbs_close, for example due to freeing
1196 		 * the resources (e.g mmput).
1197 		 */
1198 		ib_uverbs_event_handler(&file->event_handler, &event);
1199 		if (ucontext) {
1200 			ib_dev->disassociate_ucontext(ucontext);
1201 			ib_uverbs_cleanup_ucontext(file, ucontext);
1202 		}
1203 
1204 		mutex_lock(&uverbs_dev->lists_mutex);
1205 		kref_put(&file->ref, ib_uverbs_release_file);
1206 	}
1207 
1208 	while (!list_empty(&uverbs_dev->uverbs_events_file_list)) {
1209 		event_file = list_first_entry(&uverbs_dev->
1210 					      uverbs_events_file_list,
1211 					      struct ib_uverbs_event_file,
1212 					      list);
1213 		spin_lock_irq(&event_file->lock);
1214 		event_file->is_closed = 1;
1215 		spin_unlock_irq(&event_file->lock);
1216 
1217 		list_del(&event_file->list);
1218 		if (event_file->is_async) {
1219 			ib_unregister_event_handler(&event_file->uverbs_file->
1220 						    event_handler);
1221 			event_file->uverbs_file->event_handler.device = NULL;
1222 		}
1223 
1224 		wake_up_interruptible(&event_file->poll_wait);
1225 		kill_fasync(&event_file->async_queue, SIGIO, POLL_IN);
1226 	}
1227 	mutex_unlock(&uverbs_dev->lists_mutex);
1228 }
1229 
1230 static void ib_uverbs_remove_one(struct ib_device *device, void *client_data)
1231 {
1232 	struct ib_uverbs_device *uverbs_dev = client_data;
1233 	int wait_clients = 1;
1234 
1235 	if (!uverbs_dev)
1236 		return;
1237 
1238 	dev_set_drvdata(uverbs_dev->dev, NULL);
1239 	device_destroy(uverbs_class, uverbs_dev->cdev.dev);
1240 	cdev_del(&uverbs_dev->cdev);
1241 
1242 	if (uverbs_dev->devnum < IB_UVERBS_MAX_DEVICES)
1243 		clear_bit(uverbs_dev->devnum, dev_map);
1244 	else
1245 		clear_bit(uverbs_dev->devnum - IB_UVERBS_MAX_DEVICES, overflow_map);
1246 
1247 	if (device->disassociate_ucontext) {
1248 		/* We disassociate HW resources and immediately return.
1249 		 * Userspace will see a EIO errno for all future access.
1250 		 * Upon returning, ib_device may be freed internally and is not
1251 		 * valid any more.
1252 		 * uverbs_device is still available until all clients close
1253 		 * their files, then the uverbs device ref count will be zero
1254 		 * and its resources will be freed.
1255 		 * Note: At this point no more files can be opened since the
1256 		 * cdev was deleted, however active clients can still issue
1257 		 * commands and close their open files.
1258 		 */
1259 		rcu_assign_pointer(uverbs_dev->ib_dev, NULL);
1260 		ib_uverbs_free_hw_resources(uverbs_dev, device);
1261 		wait_clients = 0;
1262 	}
1263 
1264 	if (atomic_dec_and_test(&uverbs_dev->refcount))
1265 		ib_uverbs_comp_dev(uverbs_dev);
1266 	if (wait_clients)
1267 		wait_for_completion(&uverbs_dev->comp);
1268 	kobject_put(&uverbs_dev->kobj);
1269 }
1270 
1271 static char *uverbs_devnode(struct device *dev, umode_t *mode)
1272 {
1273 	if (mode)
1274 		*mode = 0666;
1275 	return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev));
1276 }
1277 
1278 static int __init ib_uverbs_init(void)
1279 {
1280 	int ret;
1281 
1282 	ret = register_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES,
1283 				     "infiniband_verbs");
1284 	if (ret) {
1285 		pr_err("user_verbs: couldn't register device number\n");
1286 		goto out;
1287 	}
1288 
1289 	uverbs_class = class_create(THIS_MODULE, "infiniband_verbs");
1290 	if (IS_ERR(uverbs_class)) {
1291 		ret = PTR_ERR(uverbs_class);
1292 		pr_err("user_verbs: couldn't create class infiniband_verbs\n");
1293 		goto out_chrdev;
1294 	}
1295 
1296 	uverbs_class->devnode = uverbs_devnode;
1297 
1298 	ret = class_create_file(uverbs_class, &class_attr_abi_version.attr);
1299 	if (ret) {
1300 		pr_err("user_verbs: couldn't create abi_version attribute\n");
1301 		goto out_class;
1302 	}
1303 
1304 	ret = ib_register_client(&uverbs_client);
1305 	if (ret) {
1306 		pr_err("user_verbs: couldn't register client\n");
1307 		goto out_class;
1308 	}
1309 
1310 	return 0;
1311 
1312 out_class:
1313 	class_destroy(uverbs_class);
1314 
1315 out_chrdev:
1316 	unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
1317 
1318 out:
1319 	return ret;
1320 }
1321 
1322 static void __exit ib_uverbs_cleanup(void)
1323 {
1324 	ib_unregister_client(&uverbs_client);
1325 	class_destroy(uverbs_class);
1326 	unregister_chrdev_region(IB_UVERBS_BASE_DEV, IB_UVERBS_MAX_DEVICES);
1327 	if (overflow_maj)
1328 		unregister_chrdev_region(overflow_maj, IB_UVERBS_MAX_DEVICES);
1329 	idr_destroy(&ib_uverbs_pd_idr);
1330 	idr_destroy(&ib_uverbs_mr_idr);
1331 	idr_destroy(&ib_uverbs_mw_idr);
1332 	idr_destroy(&ib_uverbs_ah_idr);
1333 	idr_destroy(&ib_uverbs_cq_idr);
1334 	idr_destroy(&ib_uverbs_qp_idr);
1335 	idr_destroy(&ib_uverbs_srq_idr);
1336 }
1337 
1338 module_init(ib_uverbs_init);
1339 module_exit(ib_uverbs_cleanup);
1340