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 #include <linux/sched/mm.h>
49 
50 #include <linux/uaccess.h>
51 
52 #include <rdma/ib.h>
53 #include <rdma/uverbs_std_types.h>
54 
55 #include "uverbs.h"
56 #include "core_priv.h"
57 #include "rdma_core.h"
58 
59 MODULE_AUTHOR("Roland Dreier");
60 MODULE_DESCRIPTION("InfiniBand userspace verbs access");
61 MODULE_LICENSE("Dual BSD/GPL");
62 
63 enum {
64 	IB_UVERBS_MAJOR       = 231,
65 	IB_UVERBS_BASE_MINOR  = 192,
66 	IB_UVERBS_MAX_DEVICES = RDMA_MAX_PORTS,
67 	IB_UVERBS_NUM_FIXED_MINOR = 32,
68 	IB_UVERBS_NUM_DYNAMIC_MINOR = IB_UVERBS_MAX_DEVICES - IB_UVERBS_NUM_FIXED_MINOR,
69 };
70 
71 #define IB_UVERBS_BASE_DEV	MKDEV(IB_UVERBS_MAJOR, IB_UVERBS_BASE_MINOR)
72 
73 static dev_t dynamic_uverbs_dev;
74 static struct class *uverbs_class;
75 
76 static DEFINE_IDA(uverbs_ida);
77 static void ib_uverbs_add_one(struct ib_device *device);
78 static void ib_uverbs_remove_one(struct ib_device *device, void *client_data);
79 
80 /*
81  * Must be called with the ufile->device->disassociate_srcu held, and the lock
82  * must be held until use of the ucontext is finished.
83  */
84 struct ib_ucontext *ib_uverbs_get_ucontext_file(struct ib_uverbs_file *ufile)
85 {
86 	/*
87 	 * We do not hold the hw_destroy_rwsem lock for this flow, instead
88 	 * srcu is used. It does not matter if someone races this with
89 	 * get_context, we get NULL or valid ucontext.
90 	 */
91 	struct ib_ucontext *ucontext = smp_load_acquire(&ufile->ucontext);
92 
93 	if (!srcu_dereference(ufile->device->ib_dev,
94 			      &ufile->device->disassociate_srcu))
95 		return ERR_PTR(-EIO);
96 
97 	if (!ucontext)
98 		return ERR_PTR(-EINVAL);
99 
100 	return ucontext;
101 }
102 EXPORT_SYMBOL(ib_uverbs_get_ucontext_file);
103 
104 int uverbs_dealloc_mw(struct ib_mw *mw)
105 {
106 	struct ib_pd *pd = mw->pd;
107 	int ret;
108 
109 	ret = mw->device->ops.dealloc_mw(mw);
110 	if (!ret)
111 		atomic_dec(&pd->usecnt);
112 	return ret;
113 }
114 
115 static void ib_uverbs_release_dev(struct device *device)
116 {
117 	struct ib_uverbs_device *dev =
118 			container_of(device, struct ib_uverbs_device, dev);
119 
120 	uverbs_destroy_api(dev->uapi);
121 	cleanup_srcu_struct(&dev->disassociate_srcu);
122 	kfree(dev);
123 }
124 
125 static void ib_uverbs_release_async_event_file(struct kref *ref)
126 {
127 	struct ib_uverbs_async_event_file *file =
128 		container_of(ref, struct ib_uverbs_async_event_file, ref);
129 
130 	kfree(file);
131 }
132 
133 void ib_uverbs_release_ucq(struct ib_uverbs_file *file,
134 			  struct ib_uverbs_completion_event_file *ev_file,
135 			  struct ib_ucq_object *uobj)
136 {
137 	struct ib_uverbs_event *evt, *tmp;
138 
139 	if (ev_file) {
140 		spin_lock_irq(&ev_file->ev_queue.lock);
141 		list_for_each_entry_safe(evt, tmp, &uobj->comp_list, obj_list) {
142 			list_del(&evt->list);
143 			kfree(evt);
144 		}
145 		spin_unlock_irq(&ev_file->ev_queue.lock);
146 
147 		uverbs_uobject_put(&ev_file->uobj);
148 	}
149 
150 	spin_lock_irq(&file->async_file->ev_queue.lock);
151 	list_for_each_entry_safe(evt, tmp, &uobj->async_list, obj_list) {
152 		list_del(&evt->list);
153 		kfree(evt);
154 	}
155 	spin_unlock_irq(&file->async_file->ev_queue.lock);
156 }
157 
158 void ib_uverbs_release_uevent(struct ib_uverbs_file *file,
159 			      struct ib_uevent_object *uobj)
160 {
161 	struct ib_uverbs_event *evt, *tmp;
162 
163 	spin_lock_irq(&file->async_file->ev_queue.lock);
164 	list_for_each_entry_safe(evt, tmp, &uobj->event_list, obj_list) {
165 		list_del(&evt->list);
166 		kfree(evt);
167 	}
168 	spin_unlock_irq(&file->async_file->ev_queue.lock);
169 }
170 
171 void ib_uverbs_detach_umcast(struct ib_qp *qp,
172 			     struct ib_uqp_object *uobj)
173 {
174 	struct ib_uverbs_mcast_entry *mcast, *tmp;
175 
176 	list_for_each_entry_safe(mcast, tmp, &uobj->mcast_list, list) {
177 		ib_detach_mcast(qp, &mcast->gid, mcast->lid);
178 		list_del(&mcast->list);
179 		kfree(mcast);
180 	}
181 }
182 
183 static void ib_uverbs_comp_dev(struct ib_uverbs_device *dev)
184 {
185 	complete(&dev->comp);
186 }
187 
188 void ib_uverbs_release_file(struct kref *ref)
189 {
190 	struct ib_uverbs_file *file =
191 		container_of(ref, struct ib_uverbs_file, ref);
192 	struct ib_device *ib_dev;
193 	int srcu_key;
194 
195 	release_ufile_idr_uobject(file);
196 
197 	srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
198 	ib_dev = srcu_dereference(file->device->ib_dev,
199 				  &file->device->disassociate_srcu);
200 	if (ib_dev && !ib_dev->ops.disassociate_ucontext)
201 		module_put(ib_dev->owner);
202 	srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
203 
204 	if (atomic_dec_and_test(&file->device->refcount))
205 		ib_uverbs_comp_dev(file->device);
206 
207 	put_device(&file->device->dev);
208 	kfree(file);
209 }
210 
211 static ssize_t ib_uverbs_event_read(struct ib_uverbs_event_queue *ev_queue,
212 				    struct ib_uverbs_file *uverbs_file,
213 				    struct file *filp, char __user *buf,
214 				    size_t count, loff_t *pos,
215 				    size_t eventsz)
216 {
217 	struct ib_uverbs_event *event;
218 	int ret = 0;
219 
220 	spin_lock_irq(&ev_queue->lock);
221 
222 	while (list_empty(&ev_queue->event_list)) {
223 		spin_unlock_irq(&ev_queue->lock);
224 
225 		if (filp->f_flags & O_NONBLOCK)
226 			return -EAGAIN;
227 
228 		if (wait_event_interruptible(ev_queue->poll_wait,
229 					     (!list_empty(&ev_queue->event_list) ||
230 			/* The barriers built into wait_event_interruptible()
231 			 * and wake_up() guarentee this will see the null set
232 			 * without using RCU
233 			 */
234 					     !uverbs_file->device->ib_dev)))
235 			return -ERESTARTSYS;
236 
237 		/* If device was disassociated and no event exists set an error */
238 		if (list_empty(&ev_queue->event_list) &&
239 		    !uverbs_file->device->ib_dev)
240 			return -EIO;
241 
242 		spin_lock_irq(&ev_queue->lock);
243 	}
244 
245 	event = list_entry(ev_queue->event_list.next, struct ib_uverbs_event, list);
246 
247 	if (eventsz > count) {
248 		ret   = -EINVAL;
249 		event = NULL;
250 	} else {
251 		list_del(ev_queue->event_list.next);
252 		if (event->counter) {
253 			++(*event->counter);
254 			list_del(&event->obj_list);
255 		}
256 	}
257 
258 	spin_unlock_irq(&ev_queue->lock);
259 
260 	if (event) {
261 		if (copy_to_user(buf, event, eventsz))
262 			ret = -EFAULT;
263 		else
264 			ret = eventsz;
265 	}
266 
267 	kfree(event);
268 
269 	return ret;
270 }
271 
272 static ssize_t ib_uverbs_async_event_read(struct file *filp, char __user *buf,
273 					  size_t count, loff_t *pos)
274 {
275 	struct ib_uverbs_async_event_file *file = filp->private_data;
276 
277 	return ib_uverbs_event_read(&file->ev_queue, file->uverbs_file, filp,
278 				    buf, count, pos,
279 				    sizeof(struct ib_uverbs_async_event_desc));
280 }
281 
282 static ssize_t ib_uverbs_comp_event_read(struct file *filp, char __user *buf,
283 					 size_t count, loff_t *pos)
284 {
285 	struct ib_uverbs_completion_event_file *comp_ev_file =
286 		filp->private_data;
287 
288 	return ib_uverbs_event_read(&comp_ev_file->ev_queue,
289 				    comp_ev_file->uobj.ufile, filp,
290 				    buf, count, pos,
291 				    sizeof(struct ib_uverbs_comp_event_desc));
292 }
293 
294 static __poll_t ib_uverbs_event_poll(struct ib_uverbs_event_queue *ev_queue,
295 					 struct file *filp,
296 					 struct poll_table_struct *wait)
297 {
298 	__poll_t pollflags = 0;
299 
300 	poll_wait(filp, &ev_queue->poll_wait, wait);
301 
302 	spin_lock_irq(&ev_queue->lock);
303 	if (!list_empty(&ev_queue->event_list))
304 		pollflags = EPOLLIN | EPOLLRDNORM;
305 	spin_unlock_irq(&ev_queue->lock);
306 
307 	return pollflags;
308 }
309 
310 static __poll_t ib_uverbs_async_event_poll(struct file *filp,
311 					       struct poll_table_struct *wait)
312 {
313 	return ib_uverbs_event_poll(filp->private_data, filp, wait);
314 }
315 
316 static __poll_t ib_uverbs_comp_event_poll(struct file *filp,
317 					      struct poll_table_struct *wait)
318 {
319 	struct ib_uverbs_completion_event_file *comp_ev_file =
320 		filp->private_data;
321 
322 	return ib_uverbs_event_poll(&comp_ev_file->ev_queue, filp, wait);
323 }
324 
325 static int ib_uverbs_async_event_fasync(int fd, struct file *filp, int on)
326 {
327 	struct ib_uverbs_event_queue *ev_queue = filp->private_data;
328 
329 	return fasync_helper(fd, filp, on, &ev_queue->async_queue);
330 }
331 
332 static int ib_uverbs_comp_event_fasync(int fd, struct file *filp, int on)
333 {
334 	struct ib_uverbs_completion_event_file *comp_ev_file =
335 		filp->private_data;
336 
337 	return fasync_helper(fd, filp, on, &comp_ev_file->ev_queue.async_queue);
338 }
339 
340 static int ib_uverbs_async_event_close(struct inode *inode, struct file *filp)
341 {
342 	struct ib_uverbs_async_event_file *file = filp->private_data;
343 	struct ib_uverbs_file *uverbs_file = file->uverbs_file;
344 	struct ib_uverbs_event *entry, *tmp;
345 	int closed_already = 0;
346 
347 	mutex_lock(&uverbs_file->device->lists_mutex);
348 	spin_lock_irq(&file->ev_queue.lock);
349 	closed_already = file->ev_queue.is_closed;
350 	file->ev_queue.is_closed = 1;
351 	list_for_each_entry_safe(entry, tmp, &file->ev_queue.event_list, list) {
352 		if (entry->counter)
353 			list_del(&entry->obj_list);
354 		kfree(entry);
355 	}
356 	spin_unlock_irq(&file->ev_queue.lock);
357 	if (!closed_already) {
358 		list_del(&file->list);
359 		ib_unregister_event_handler(&uverbs_file->event_handler);
360 	}
361 	mutex_unlock(&uverbs_file->device->lists_mutex);
362 
363 	kref_put(&uverbs_file->ref, ib_uverbs_release_file);
364 	kref_put(&file->ref, ib_uverbs_release_async_event_file);
365 
366 	return 0;
367 }
368 
369 static int ib_uverbs_comp_event_close(struct inode *inode, struct file *filp)
370 {
371 	struct ib_uobject *uobj = filp->private_data;
372 	struct ib_uverbs_completion_event_file *file = container_of(
373 		uobj, struct ib_uverbs_completion_event_file, uobj);
374 	struct ib_uverbs_event *entry, *tmp;
375 
376 	spin_lock_irq(&file->ev_queue.lock);
377 	list_for_each_entry_safe(entry, tmp, &file->ev_queue.event_list, list) {
378 		if (entry->counter)
379 			list_del(&entry->obj_list);
380 		kfree(entry);
381 	}
382 	file->ev_queue.is_closed = 1;
383 	spin_unlock_irq(&file->ev_queue.lock);
384 
385 	uverbs_close_fd(filp);
386 
387 	return 0;
388 }
389 
390 const struct file_operations uverbs_event_fops = {
391 	.owner	 = THIS_MODULE,
392 	.read	 = ib_uverbs_comp_event_read,
393 	.poll    = ib_uverbs_comp_event_poll,
394 	.release = ib_uverbs_comp_event_close,
395 	.fasync  = ib_uverbs_comp_event_fasync,
396 	.llseek	 = no_llseek,
397 };
398 
399 static const struct file_operations uverbs_async_event_fops = {
400 	.owner	 = THIS_MODULE,
401 	.read	 = ib_uverbs_async_event_read,
402 	.poll    = ib_uverbs_async_event_poll,
403 	.release = ib_uverbs_async_event_close,
404 	.fasync  = ib_uverbs_async_event_fasync,
405 	.llseek	 = no_llseek,
406 };
407 
408 void ib_uverbs_comp_handler(struct ib_cq *cq, void *cq_context)
409 {
410 	struct ib_uverbs_event_queue   *ev_queue = cq_context;
411 	struct ib_ucq_object	       *uobj;
412 	struct ib_uverbs_event	       *entry;
413 	unsigned long			flags;
414 
415 	if (!ev_queue)
416 		return;
417 
418 	spin_lock_irqsave(&ev_queue->lock, flags);
419 	if (ev_queue->is_closed) {
420 		spin_unlock_irqrestore(&ev_queue->lock, flags);
421 		return;
422 	}
423 
424 	entry = kmalloc(sizeof(*entry), GFP_ATOMIC);
425 	if (!entry) {
426 		spin_unlock_irqrestore(&ev_queue->lock, flags);
427 		return;
428 	}
429 
430 	uobj = container_of(cq->uobject, struct ib_ucq_object, uobject);
431 
432 	entry->desc.comp.cq_handle = cq->uobject->user_handle;
433 	entry->counter		   = &uobj->comp_events_reported;
434 
435 	list_add_tail(&entry->list, &ev_queue->event_list);
436 	list_add_tail(&entry->obj_list, &uobj->comp_list);
437 	spin_unlock_irqrestore(&ev_queue->lock, flags);
438 
439 	wake_up_interruptible(&ev_queue->poll_wait);
440 	kill_fasync(&ev_queue->async_queue, SIGIO, POLL_IN);
441 }
442 
443 static void ib_uverbs_async_handler(struct ib_uverbs_file *file,
444 				    __u64 element, __u64 event,
445 				    struct list_head *obj_list,
446 				    u32 *counter)
447 {
448 	struct ib_uverbs_event *entry;
449 	unsigned long flags;
450 
451 	spin_lock_irqsave(&file->async_file->ev_queue.lock, flags);
452 	if (file->async_file->ev_queue.is_closed) {
453 		spin_unlock_irqrestore(&file->async_file->ev_queue.lock, flags);
454 		return;
455 	}
456 
457 	entry = kmalloc(sizeof(*entry), GFP_ATOMIC);
458 	if (!entry) {
459 		spin_unlock_irqrestore(&file->async_file->ev_queue.lock, flags);
460 		return;
461 	}
462 
463 	entry->desc.async.element    = element;
464 	entry->desc.async.event_type = event;
465 	entry->desc.async.reserved   = 0;
466 	entry->counter               = counter;
467 
468 	list_add_tail(&entry->list, &file->async_file->ev_queue.event_list);
469 	if (obj_list)
470 		list_add_tail(&entry->obj_list, obj_list);
471 	spin_unlock_irqrestore(&file->async_file->ev_queue.lock, flags);
472 
473 	wake_up_interruptible(&file->async_file->ev_queue.poll_wait);
474 	kill_fasync(&file->async_file->ev_queue.async_queue, SIGIO, POLL_IN);
475 }
476 
477 void ib_uverbs_cq_event_handler(struct ib_event *event, void *context_ptr)
478 {
479 	struct ib_ucq_object *uobj = container_of(event->element.cq->uobject,
480 						  struct ib_ucq_object, uobject);
481 
482 	ib_uverbs_async_handler(uobj->uobject.ufile, uobj->uobject.user_handle,
483 				event->event, &uobj->async_list,
484 				&uobj->async_events_reported);
485 }
486 
487 void ib_uverbs_qp_event_handler(struct ib_event *event, void *context_ptr)
488 {
489 	struct ib_uevent_object *uobj;
490 
491 	/* for XRC target qp's, check that qp is live */
492 	if (!event->element.qp->uobject)
493 		return;
494 
495 	uobj = container_of(event->element.qp->uobject,
496 			    struct ib_uevent_object, uobject);
497 
498 	ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
499 				event->event, &uobj->event_list,
500 				&uobj->events_reported);
501 }
502 
503 void ib_uverbs_wq_event_handler(struct ib_event *event, void *context_ptr)
504 {
505 	struct ib_uevent_object *uobj = container_of(event->element.wq->uobject,
506 						  struct ib_uevent_object, uobject);
507 
508 	ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
509 				event->event, &uobj->event_list,
510 				&uobj->events_reported);
511 }
512 
513 void ib_uverbs_srq_event_handler(struct ib_event *event, void *context_ptr)
514 {
515 	struct ib_uevent_object *uobj;
516 
517 	uobj = container_of(event->element.srq->uobject,
518 			    struct ib_uevent_object, uobject);
519 
520 	ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
521 				event->event, &uobj->event_list,
522 				&uobj->events_reported);
523 }
524 
525 void ib_uverbs_event_handler(struct ib_event_handler *handler,
526 			     struct ib_event *event)
527 {
528 	struct ib_uverbs_file *file =
529 		container_of(handler, struct ib_uverbs_file, event_handler);
530 
531 	ib_uverbs_async_handler(file, event->element.port_num, event->event,
532 				NULL, NULL);
533 }
534 
535 void ib_uverbs_free_async_event_file(struct ib_uverbs_file *file)
536 {
537 	kref_put(&file->async_file->ref, ib_uverbs_release_async_event_file);
538 	file->async_file = NULL;
539 }
540 
541 void ib_uverbs_init_event_queue(struct ib_uverbs_event_queue *ev_queue)
542 {
543 	spin_lock_init(&ev_queue->lock);
544 	INIT_LIST_HEAD(&ev_queue->event_list);
545 	init_waitqueue_head(&ev_queue->poll_wait);
546 	ev_queue->is_closed   = 0;
547 	ev_queue->async_queue = NULL;
548 }
549 
550 struct file *ib_uverbs_alloc_async_event_file(struct ib_uverbs_file *uverbs_file,
551 					      struct ib_device	*ib_dev)
552 {
553 	struct ib_uverbs_async_event_file *ev_file;
554 	struct file *filp;
555 
556 	ev_file = kzalloc(sizeof(*ev_file), GFP_KERNEL);
557 	if (!ev_file)
558 		return ERR_PTR(-ENOMEM);
559 
560 	ib_uverbs_init_event_queue(&ev_file->ev_queue);
561 	ev_file->uverbs_file = uverbs_file;
562 	kref_get(&ev_file->uverbs_file->ref);
563 	kref_init(&ev_file->ref);
564 	filp = anon_inode_getfile("[infinibandevent]", &uverbs_async_event_fops,
565 				  ev_file, O_RDONLY);
566 	if (IS_ERR(filp))
567 		goto err_put_refs;
568 
569 	mutex_lock(&uverbs_file->device->lists_mutex);
570 	list_add_tail(&ev_file->list,
571 		      &uverbs_file->device->uverbs_events_file_list);
572 	mutex_unlock(&uverbs_file->device->lists_mutex);
573 
574 	WARN_ON(uverbs_file->async_file);
575 	uverbs_file->async_file = ev_file;
576 	kref_get(&uverbs_file->async_file->ref);
577 	INIT_IB_EVENT_HANDLER(&uverbs_file->event_handler,
578 			      ib_dev,
579 			      ib_uverbs_event_handler);
580 	ib_register_event_handler(&uverbs_file->event_handler);
581 	/* At that point async file stuff was fully set */
582 
583 	return filp;
584 
585 err_put_refs:
586 	kref_put(&ev_file->uverbs_file->ref, ib_uverbs_release_file);
587 	kref_put(&ev_file->ref, ib_uverbs_release_async_event_file);
588 	return filp;
589 }
590 
591 static ssize_t verify_hdr(struct ib_uverbs_cmd_hdr *hdr,
592 			  struct ib_uverbs_ex_cmd_hdr *ex_hdr, size_t count,
593 			  const struct uverbs_api_write_method *method_elm)
594 {
595 	if (method_elm->is_ex) {
596 		count -= sizeof(*hdr) + sizeof(*ex_hdr);
597 
598 		if ((hdr->in_words + ex_hdr->provider_in_words) * 8 != count)
599 			return -EINVAL;
600 
601 		if (hdr->in_words * 8 < method_elm->req_size)
602 			return -ENOSPC;
603 
604 		if (ex_hdr->cmd_hdr_reserved)
605 			return -EINVAL;
606 
607 		if (ex_hdr->response) {
608 			if (!hdr->out_words && !ex_hdr->provider_out_words)
609 				return -EINVAL;
610 
611 			if (hdr->out_words * 8 < method_elm->resp_size)
612 				return -ENOSPC;
613 
614 			if (!access_ok(u64_to_user_ptr(ex_hdr->response),
615 				       (hdr->out_words + ex_hdr->provider_out_words) * 8))
616 				return -EFAULT;
617 		} else {
618 			if (hdr->out_words || ex_hdr->provider_out_words)
619 				return -EINVAL;
620 		}
621 
622 		return 0;
623 	}
624 
625 	/* not extended command */
626 	if (hdr->in_words * 4 != count)
627 		return -EINVAL;
628 
629 	if (count < method_elm->req_size + sizeof(hdr)) {
630 		/*
631 		 * rdma-core v18 and v19 have a bug where they send DESTROY_CQ
632 		 * with a 16 byte write instead of 24. Old kernels didn't
633 		 * check the size so they allowed this. Now that the size is
634 		 * checked provide a compatibility work around to not break
635 		 * those userspaces.
636 		 */
637 		if (hdr->command == IB_USER_VERBS_CMD_DESTROY_CQ &&
638 		    count == 16) {
639 			hdr->in_words = 6;
640 			return 0;
641 		}
642 		return -ENOSPC;
643 	}
644 	if (hdr->out_words * 4 < method_elm->resp_size)
645 		return -ENOSPC;
646 
647 	return 0;
648 }
649 
650 static ssize_t ib_uverbs_write(struct file *filp, const char __user *buf,
651 			     size_t count, loff_t *pos)
652 {
653 	struct ib_uverbs_file *file = filp->private_data;
654 	const struct uverbs_api_write_method *method_elm;
655 	struct uverbs_api *uapi = file->device->uapi;
656 	struct ib_uverbs_ex_cmd_hdr ex_hdr;
657 	struct ib_uverbs_cmd_hdr hdr;
658 	struct uverbs_attr_bundle bundle;
659 	int srcu_key;
660 	ssize_t ret;
661 
662 	if (!ib_safe_file_access(filp)) {
663 		pr_err_once("uverbs_write: process %d (%s) changed security contexts after opening file descriptor, this is not allowed.\n",
664 			    task_tgid_vnr(current), current->comm);
665 		return -EACCES;
666 	}
667 
668 	if (count < sizeof(hdr))
669 		return -EINVAL;
670 
671 	if (copy_from_user(&hdr, buf, sizeof(hdr)))
672 		return -EFAULT;
673 
674 	method_elm = uapi_get_method(uapi, hdr.command);
675 	if (IS_ERR(method_elm))
676 		return PTR_ERR(method_elm);
677 
678 	if (method_elm->is_ex) {
679 		if (count < (sizeof(hdr) + sizeof(ex_hdr)))
680 			return -EINVAL;
681 		if (copy_from_user(&ex_hdr, buf + sizeof(hdr), sizeof(ex_hdr)))
682 			return -EFAULT;
683 	}
684 
685 	ret = verify_hdr(&hdr, &ex_hdr, count, method_elm);
686 	if (ret)
687 		return ret;
688 
689 	srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
690 
691 	buf += sizeof(hdr);
692 
693 	bundle.ufile = file;
694 	if (!method_elm->is_ex) {
695 		size_t in_len = hdr.in_words * 4 - sizeof(hdr);
696 		size_t out_len = hdr.out_words * 4;
697 		u64 response = 0;
698 
699 		if (method_elm->has_udata) {
700 			bundle.driver_udata.inlen =
701 				in_len - method_elm->req_size;
702 			in_len = method_elm->req_size;
703 			if (bundle.driver_udata.inlen)
704 				bundle.driver_udata.inbuf = buf + in_len;
705 			else
706 				bundle.driver_udata.inbuf = NULL;
707 		} else {
708 			memset(&bundle.driver_udata, 0,
709 			       sizeof(bundle.driver_udata));
710 		}
711 
712 		if (method_elm->has_resp) {
713 			/*
714 			 * The macros check that if has_resp is set
715 			 * then the command request structure starts
716 			 * with a '__aligned u64 response' member.
717 			 */
718 			ret = get_user(response, (const u64 *)buf);
719 			if (ret)
720 				goto out_unlock;
721 
722 			if (method_elm->has_udata) {
723 				bundle.driver_udata.outlen =
724 					out_len - method_elm->resp_size;
725 				out_len = method_elm->resp_size;
726 				if (bundle.driver_udata.outlen)
727 					bundle.driver_udata.outbuf =
728 						u64_to_user_ptr(response +
729 								out_len);
730 				else
731 					bundle.driver_udata.outbuf = NULL;
732 			}
733 		} else {
734 			bundle.driver_udata.outlen = 0;
735 			bundle.driver_udata.outbuf = NULL;
736 		}
737 
738 		ib_uverbs_init_udata_buf_or_null(
739 			&bundle.ucore, buf, u64_to_user_ptr(response),
740 			in_len, out_len);
741 	} else {
742 		buf += sizeof(ex_hdr);
743 
744 		ib_uverbs_init_udata_buf_or_null(&bundle.ucore, buf,
745 					u64_to_user_ptr(ex_hdr.response),
746 					hdr.in_words * 8, hdr.out_words * 8);
747 
748 		ib_uverbs_init_udata_buf_or_null(
749 			&bundle.driver_udata, buf + bundle.ucore.inlen,
750 			u64_to_user_ptr(ex_hdr.response) + bundle.ucore.outlen,
751 			ex_hdr.provider_in_words * 8,
752 			ex_hdr.provider_out_words * 8);
753 
754 	}
755 
756 	ret = method_elm->handler(&bundle);
757 out_unlock:
758 	srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
759 	return (ret) ? : count;
760 }
761 
762 static int ib_uverbs_mmap(struct file *filp, struct vm_area_struct *vma)
763 {
764 	struct ib_uverbs_file *file = filp->private_data;
765 	struct ib_ucontext *ucontext;
766 	int ret = 0;
767 	int srcu_key;
768 
769 	srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
770 	ucontext = ib_uverbs_get_ucontext_file(file);
771 	if (IS_ERR(ucontext)) {
772 		ret = PTR_ERR(ucontext);
773 		goto out;
774 	}
775 
776 	ret = ucontext->device->ops.mmap(ucontext, vma);
777 out:
778 	srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
779 	return ret;
780 }
781 
782 /*
783  * Each time we map IO memory into user space this keeps track of the mapping.
784  * When the device is hot-unplugged we 'zap' the mmaps in user space to point
785  * to the zero page and allow the hot unplug to proceed.
786  *
787  * This is necessary for cases like PCI physical hot unplug as the actual BAR
788  * memory may vanish after this and access to it from userspace could MCE.
789  *
790  * RDMA drivers supporting disassociation must have their user space designed
791  * to cope in some way with their IO pages going to the zero page.
792  */
793 struct rdma_umap_priv {
794 	struct vm_area_struct *vma;
795 	struct list_head list;
796 };
797 
798 static const struct vm_operations_struct rdma_umap_ops;
799 
800 static void rdma_umap_priv_init(struct rdma_umap_priv *priv,
801 				struct vm_area_struct *vma)
802 {
803 	struct ib_uverbs_file *ufile = vma->vm_file->private_data;
804 
805 	priv->vma = vma;
806 	vma->vm_private_data = priv;
807 	vma->vm_ops = &rdma_umap_ops;
808 
809 	mutex_lock(&ufile->umap_lock);
810 	list_add(&priv->list, &ufile->umaps);
811 	mutex_unlock(&ufile->umap_lock);
812 }
813 
814 /*
815  * The VMA has been dup'd, initialize the vm_private_data with a new tracking
816  * struct
817  */
818 static void rdma_umap_open(struct vm_area_struct *vma)
819 {
820 	struct ib_uverbs_file *ufile = vma->vm_file->private_data;
821 	struct rdma_umap_priv *opriv = vma->vm_private_data;
822 	struct rdma_umap_priv *priv;
823 
824 	if (!opriv)
825 		return;
826 
827 	/* We are racing with disassociation */
828 	if (!down_read_trylock(&ufile->hw_destroy_rwsem))
829 		goto out_zap;
830 	/*
831 	 * Disassociation already completed, the VMA should already be zapped.
832 	 */
833 	if (!ufile->ucontext)
834 		goto out_unlock;
835 
836 	priv = kzalloc(sizeof(*priv), GFP_KERNEL);
837 	if (!priv)
838 		goto out_unlock;
839 	rdma_umap_priv_init(priv, vma);
840 
841 	up_read(&ufile->hw_destroy_rwsem);
842 	return;
843 
844 out_unlock:
845 	up_read(&ufile->hw_destroy_rwsem);
846 out_zap:
847 	/*
848 	 * We can't allow the VMA to be created with the actual IO pages, that
849 	 * would break our API contract, and it can't be stopped at this
850 	 * point, so zap it.
851 	 */
852 	vma->vm_private_data = NULL;
853 	zap_vma_ptes(vma, vma->vm_start, vma->vm_end - vma->vm_start);
854 }
855 
856 static void rdma_umap_close(struct vm_area_struct *vma)
857 {
858 	struct ib_uverbs_file *ufile = vma->vm_file->private_data;
859 	struct rdma_umap_priv *priv = vma->vm_private_data;
860 
861 	if (!priv)
862 		return;
863 
864 	/*
865 	 * The vma holds a reference on the struct file that created it, which
866 	 * in turn means that the ib_uverbs_file is guaranteed to exist at
867 	 * this point.
868 	 */
869 	mutex_lock(&ufile->umap_lock);
870 	list_del(&priv->list);
871 	mutex_unlock(&ufile->umap_lock);
872 	kfree(priv);
873 }
874 
875 static const struct vm_operations_struct rdma_umap_ops = {
876 	.open = rdma_umap_open,
877 	.close = rdma_umap_close,
878 };
879 
880 static struct rdma_umap_priv *rdma_user_mmap_pre(struct ib_ucontext *ucontext,
881 						 struct vm_area_struct *vma,
882 						 unsigned long size)
883 {
884 	struct ib_uverbs_file *ufile = ucontext->ufile;
885 	struct rdma_umap_priv *priv;
886 
887 	if (vma->vm_end - vma->vm_start != size)
888 		return ERR_PTR(-EINVAL);
889 
890 	/* Driver is using this wrong, must be called by ib_uverbs_mmap */
891 	if (WARN_ON(!vma->vm_file ||
892 		    vma->vm_file->private_data != ufile))
893 		return ERR_PTR(-EINVAL);
894 	lockdep_assert_held(&ufile->device->disassociate_srcu);
895 
896 	priv = kzalloc(sizeof(*priv), GFP_KERNEL);
897 	if (!priv)
898 		return ERR_PTR(-ENOMEM);
899 	return priv;
900 }
901 
902 /*
903  * Map IO memory into a process. This is to be called by drivers as part of
904  * their mmap() functions if they wish to send something like PCI-E BAR memory
905  * to userspace.
906  */
907 int rdma_user_mmap_io(struct ib_ucontext *ucontext, struct vm_area_struct *vma,
908 		      unsigned long pfn, unsigned long size, pgprot_t prot)
909 {
910 	struct rdma_umap_priv *priv = rdma_user_mmap_pre(ucontext, vma, size);
911 
912 	if (IS_ERR(priv))
913 		return PTR_ERR(priv);
914 
915 	vma->vm_page_prot = prot;
916 	if (io_remap_pfn_range(vma, vma->vm_start, pfn, size, prot)) {
917 		kfree(priv);
918 		return -EAGAIN;
919 	}
920 
921 	rdma_umap_priv_init(priv, vma);
922 	return 0;
923 }
924 EXPORT_SYMBOL(rdma_user_mmap_io);
925 
926 /*
927  * The page case is here for a slightly different reason, the driver expects
928  * to be able to free the page it is sharing to user space when it destroys
929  * its ucontext, which means we need to zap the user space references.
930  *
931  * We could handle this differently by providing an API to allocate a shared
932  * page and then only freeing the shared page when the last ufile is
933  * destroyed.
934  */
935 int rdma_user_mmap_page(struct ib_ucontext *ucontext,
936 			struct vm_area_struct *vma, struct page *page,
937 			unsigned long size)
938 {
939 	struct rdma_umap_priv *priv = rdma_user_mmap_pre(ucontext, vma, size);
940 
941 	if (IS_ERR(priv))
942 		return PTR_ERR(priv);
943 
944 	if (remap_pfn_range(vma, vma->vm_start, page_to_pfn(page), size,
945 			    vma->vm_page_prot)) {
946 		kfree(priv);
947 		return -EAGAIN;
948 	}
949 
950 	rdma_umap_priv_init(priv, vma);
951 	return 0;
952 }
953 EXPORT_SYMBOL(rdma_user_mmap_page);
954 
955 void uverbs_user_mmap_disassociate(struct ib_uverbs_file *ufile)
956 {
957 	struct rdma_umap_priv *priv, *next_priv;
958 
959 	lockdep_assert_held(&ufile->hw_destroy_rwsem);
960 
961 	while (1) {
962 		struct mm_struct *mm = NULL;
963 
964 		/* Get an arbitrary mm pointer that hasn't been cleaned yet */
965 		mutex_lock(&ufile->umap_lock);
966 		if (!list_empty(&ufile->umaps)) {
967 			mm = list_first_entry(&ufile->umaps,
968 					      struct rdma_umap_priv, list)
969 				     ->vma->vm_mm;
970 			mmget(mm);
971 		}
972 		mutex_unlock(&ufile->umap_lock);
973 		if (!mm)
974 			return;
975 
976 		/*
977 		 * The umap_lock is nested under mmap_sem since it used within
978 		 * the vma_ops callbacks, so we have to clean the list one mm
979 		 * at a time to get the lock ordering right. Typically there
980 		 * will only be one mm, so no big deal.
981 		 */
982 		down_write(&mm->mmap_sem);
983 		mutex_lock(&ufile->umap_lock);
984 		list_for_each_entry_safe (priv, next_priv, &ufile->umaps,
985 					  list) {
986 			struct vm_area_struct *vma = priv->vma;
987 
988 			if (vma->vm_mm != mm)
989 				continue;
990 			list_del_init(&priv->list);
991 
992 			zap_vma_ptes(vma, vma->vm_start,
993 				     vma->vm_end - vma->vm_start);
994 			vma->vm_flags &= ~(VM_SHARED | VM_MAYSHARE);
995 		}
996 		mutex_unlock(&ufile->umap_lock);
997 		up_write(&mm->mmap_sem);
998 		mmput(mm);
999 	}
1000 }
1001 
1002 /*
1003  * ib_uverbs_open() does not need the BKL:
1004  *
1005  *  - the ib_uverbs_device structures are properly reference counted and
1006  *    everything else is purely local to the file being created, so
1007  *    races against other open calls are not a problem;
1008  *  - there is no ioctl method to race against;
1009  *  - the open method will either immediately run -ENXIO, or all
1010  *    required initialization will be done.
1011  */
1012 static int ib_uverbs_open(struct inode *inode, struct file *filp)
1013 {
1014 	struct ib_uverbs_device *dev;
1015 	struct ib_uverbs_file *file;
1016 	struct ib_device *ib_dev;
1017 	int ret;
1018 	int module_dependent;
1019 	int srcu_key;
1020 
1021 	dev = container_of(inode->i_cdev, struct ib_uverbs_device, cdev);
1022 	if (!atomic_inc_not_zero(&dev->refcount))
1023 		return -ENXIO;
1024 
1025 	get_device(&dev->dev);
1026 	srcu_key = srcu_read_lock(&dev->disassociate_srcu);
1027 	mutex_lock(&dev->lists_mutex);
1028 	ib_dev = srcu_dereference(dev->ib_dev,
1029 				  &dev->disassociate_srcu);
1030 	if (!ib_dev) {
1031 		ret = -EIO;
1032 		goto err;
1033 	}
1034 
1035 	/* In case IB device supports disassociate ucontext, there is no hard
1036 	 * dependency between uverbs device and its low level device.
1037 	 */
1038 	module_dependent = !(ib_dev->ops.disassociate_ucontext);
1039 
1040 	if (module_dependent) {
1041 		if (!try_module_get(ib_dev->owner)) {
1042 			ret = -ENODEV;
1043 			goto err;
1044 		}
1045 	}
1046 
1047 	file = kzalloc(sizeof(*file), GFP_KERNEL);
1048 	if (!file) {
1049 		ret = -ENOMEM;
1050 		if (module_dependent)
1051 			goto err_module;
1052 
1053 		goto err;
1054 	}
1055 
1056 	file->device	 = dev;
1057 	kref_init(&file->ref);
1058 	mutex_init(&file->ucontext_lock);
1059 
1060 	spin_lock_init(&file->uobjects_lock);
1061 	INIT_LIST_HEAD(&file->uobjects);
1062 	init_rwsem(&file->hw_destroy_rwsem);
1063 	mutex_init(&file->umap_lock);
1064 	INIT_LIST_HEAD(&file->umaps);
1065 
1066 	filp->private_data = file;
1067 	list_add_tail(&file->list, &dev->uverbs_file_list);
1068 	mutex_unlock(&dev->lists_mutex);
1069 	srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1070 
1071 	setup_ufile_idr_uobject(file);
1072 
1073 	return nonseekable_open(inode, filp);
1074 
1075 err_module:
1076 	module_put(ib_dev->owner);
1077 
1078 err:
1079 	mutex_unlock(&dev->lists_mutex);
1080 	srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1081 	if (atomic_dec_and_test(&dev->refcount))
1082 		ib_uverbs_comp_dev(dev);
1083 
1084 	put_device(&dev->dev);
1085 	return ret;
1086 }
1087 
1088 static int ib_uverbs_close(struct inode *inode, struct file *filp)
1089 {
1090 	struct ib_uverbs_file *file = filp->private_data;
1091 
1092 	uverbs_destroy_ufile_hw(file, RDMA_REMOVE_CLOSE);
1093 
1094 	mutex_lock(&file->device->lists_mutex);
1095 	list_del_init(&file->list);
1096 	mutex_unlock(&file->device->lists_mutex);
1097 
1098 	if (file->async_file)
1099 		kref_put(&file->async_file->ref,
1100 			 ib_uverbs_release_async_event_file);
1101 
1102 	kref_put(&file->ref, ib_uverbs_release_file);
1103 
1104 	return 0;
1105 }
1106 
1107 static const struct file_operations uverbs_fops = {
1108 	.owner	 = THIS_MODULE,
1109 	.write	 = ib_uverbs_write,
1110 	.open	 = ib_uverbs_open,
1111 	.release = ib_uverbs_close,
1112 	.llseek	 = no_llseek,
1113 	.unlocked_ioctl = ib_uverbs_ioctl,
1114 	.compat_ioctl = ib_uverbs_ioctl,
1115 };
1116 
1117 static const struct file_operations uverbs_mmap_fops = {
1118 	.owner	 = THIS_MODULE,
1119 	.write	 = ib_uverbs_write,
1120 	.mmap    = ib_uverbs_mmap,
1121 	.open	 = ib_uverbs_open,
1122 	.release = ib_uverbs_close,
1123 	.llseek	 = no_llseek,
1124 	.unlocked_ioctl = ib_uverbs_ioctl,
1125 	.compat_ioctl = ib_uverbs_ioctl,
1126 };
1127 
1128 static struct ib_client uverbs_client = {
1129 	.name   = "uverbs",
1130 	.add    = ib_uverbs_add_one,
1131 	.remove = ib_uverbs_remove_one
1132 };
1133 
1134 static ssize_t ibdev_show(struct device *device, struct device_attribute *attr,
1135 			  char *buf)
1136 {
1137 	struct ib_uverbs_device *dev =
1138 			container_of(device, struct ib_uverbs_device, dev);
1139 	int ret = -ENODEV;
1140 	int srcu_key;
1141 	struct ib_device *ib_dev;
1142 
1143 	srcu_key = srcu_read_lock(&dev->disassociate_srcu);
1144 	ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu);
1145 	if (ib_dev)
1146 		ret = sprintf(buf, "%s\n", dev_name(&ib_dev->dev));
1147 	srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1148 
1149 	return ret;
1150 }
1151 static DEVICE_ATTR_RO(ibdev);
1152 
1153 static ssize_t abi_version_show(struct device *device,
1154 				struct device_attribute *attr, char *buf)
1155 {
1156 	struct ib_uverbs_device *dev =
1157 			container_of(device, struct ib_uverbs_device, dev);
1158 	int ret = -ENODEV;
1159 	int srcu_key;
1160 	struct ib_device *ib_dev;
1161 
1162 	srcu_key = srcu_read_lock(&dev->disassociate_srcu);
1163 	ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu);
1164 	if (ib_dev)
1165 		ret = sprintf(buf, "%d\n", ib_dev->uverbs_abi_ver);
1166 	srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1167 
1168 	return ret;
1169 }
1170 static DEVICE_ATTR_RO(abi_version);
1171 
1172 static struct attribute *ib_dev_attrs[] = {
1173 	&dev_attr_abi_version.attr,
1174 	&dev_attr_ibdev.attr,
1175 	NULL,
1176 };
1177 
1178 static const struct attribute_group dev_attr_group = {
1179 	.attrs = ib_dev_attrs,
1180 };
1181 
1182 static CLASS_ATTR_STRING(abi_version, S_IRUGO,
1183 			 __stringify(IB_USER_VERBS_ABI_VERSION));
1184 
1185 static int ib_uverbs_create_uapi(struct ib_device *device,
1186 				 struct ib_uverbs_device *uverbs_dev)
1187 {
1188 	struct uverbs_api *uapi;
1189 
1190 	uapi = uverbs_alloc_api(device);
1191 	if (IS_ERR(uapi))
1192 		return PTR_ERR(uapi);
1193 
1194 	uverbs_dev->uapi = uapi;
1195 	return 0;
1196 }
1197 
1198 static void ib_uverbs_add_one(struct ib_device *device)
1199 {
1200 	int devnum;
1201 	dev_t base;
1202 	struct ib_uverbs_device *uverbs_dev;
1203 	int ret;
1204 
1205 	if (!device->ops.alloc_ucontext)
1206 		return;
1207 
1208 	uverbs_dev = kzalloc(sizeof(*uverbs_dev), GFP_KERNEL);
1209 	if (!uverbs_dev)
1210 		return;
1211 
1212 	ret = init_srcu_struct(&uverbs_dev->disassociate_srcu);
1213 	if (ret) {
1214 		kfree(uverbs_dev);
1215 		return;
1216 	}
1217 
1218 	device_initialize(&uverbs_dev->dev);
1219 	uverbs_dev->dev.class = uverbs_class;
1220 	uverbs_dev->dev.parent = device->dev.parent;
1221 	uverbs_dev->dev.release = ib_uverbs_release_dev;
1222 	uverbs_dev->groups[0] = &dev_attr_group;
1223 	uverbs_dev->dev.groups = uverbs_dev->groups;
1224 	atomic_set(&uverbs_dev->refcount, 1);
1225 	init_completion(&uverbs_dev->comp);
1226 	uverbs_dev->xrcd_tree = RB_ROOT;
1227 	mutex_init(&uverbs_dev->xrcd_tree_mutex);
1228 	mutex_init(&uverbs_dev->lists_mutex);
1229 	INIT_LIST_HEAD(&uverbs_dev->uverbs_file_list);
1230 	INIT_LIST_HEAD(&uverbs_dev->uverbs_events_file_list);
1231 	rcu_assign_pointer(uverbs_dev->ib_dev, device);
1232 	uverbs_dev->num_comp_vectors = device->num_comp_vectors;
1233 
1234 	devnum = ida_alloc_max(&uverbs_ida, IB_UVERBS_MAX_DEVICES - 1,
1235 			       GFP_KERNEL);
1236 	if (devnum < 0)
1237 		goto err;
1238 	uverbs_dev->devnum = devnum;
1239 	if (devnum >= IB_UVERBS_NUM_FIXED_MINOR)
1240 		base = dynamic_uverbs_dev + devnum - IB_UVERBS_NUM_FIXED_MINOR;
1241 	else
1242 		base = IB_UVERBS_BASE_DEV + devnum;
1243 
1244 	if (ib_uverbs_create_uapi(device, uverbs_dev))
1245 		goto err_uapi;
1246 
1247 	uverbs_dev->dev.devt = base;
1248 	dev_set_name(&uverbs_dev->dev, "uverbs%d", uverbs_dev->devnum);
1249 
1250 	cdev_init(&uverbs_dev->cdev,
1251 		  device->ops.mmap ? &uverbs_mmap_fops : &uverbs_fops);
1252 	uverbs_dev->cdev.owner = THIS_MODULE;
1253 
1254 	ret = cdev_device_add(&uverbs_dev->cdev, &uverbs_dev->dev);
1255 	if (ret)
1256 		goto err_uapi;
1257 
1258 	ib_set_client_data(device, &uverbs_client, uverbs_dev);
1259 	return;
1260 
1261 err_uapi:
1262 	ida_free(&uverbs_ida, devnum);
1263 err:
1264 	if (atomic_dec_and_test(&uverbs_dev->refcount))
1265 		ib_uverbs_comp_dev(uverbs_dev);
1266 	wait_for_completion(&uverbs_dev->comp);
1267 	put_device(&uverbs_dev->dev);
1268 	return;
1269 }
1270 
1271 static void ib_uverbs_free_hw_resources(struct ib_uverbs_device *uverbs_dev,
1272 					struct ib_device *ib_dev)
1273 {
1274 	struct ib_uverbs_file *file;
1275 	struct ib_uverbs_async_event_file *event_file;
1276 	struct ib_event event;
1277 
1278 	/* Pending running commands to terminate */
1279 	uverbs_disassociate_api_pre(uverbs_dev);
1280 	event.event = IB_EVENT_DEVICE_FATAL;
1281 	event.element.port_num = 0;
1282 	event.device = ib_dev;
1283 
1284 	mutex_lock(&uverbs_dev->lists_mutex);
1285 	while (!list_empty(&uverbs_dev->uverbs_file_list)) {
1286 		file = list_first_entry(&uverbs_dev->uverbs_file_list,
1287 					struct ib_uverbs_file, list);
1288 		list_del_init(&file->list);
1289 		kref_get(&file->ref);
1290 
1291 		/* We must release the mutex before going ahead and calling
1292 		 * uverbs_cleanup_ufile, as it might end up indirectly calling
1293 		 * uverbs_close, for example due to freeing the resources (e.g
1294 		 * mmput).
1295 		 */
1296 		mutex_unlock(&uverbs_dev->lists_mutex);
1297 
1298 		ib_uverbs_event_handler(&file->event_handler, &event);
1299 		uverbs_destroy_ufile_hw(file, RDMA_REMOVE_DRIVER_REMOVE);
1300 		kref_put(&file->ref, ib_uverbs_release_file);
1301 
1302 		mutex_lock(&uverbs_dev->lists_mutex);
1303 	}
1304 
1305 	while (!list_empty(&uverbs_dev->uverbs_events_file_list)) {
1306 		event_file = list_first_entry(&uverbs_dev->
1307 					      uverbs_events_file_list,
1308 					      struct ib_uverbs_async_event_file,
1309 					      list);
1310 		spin_lock_irq(&event_file->ev_queue.lock);
1311 		event_file->ev_queue.is_closed = 1;
1312 		spin_unlock_irq(&event_file->ev_queue.lock);
1313 
1314 		list_del(&event_file->list);
1315 		ib_unregister_event_handler(
1316 			&event_file->uverbs_file->event_handler);
1317 		event_file->uverbs_file->event_handler.device =
1318 			NULL;
1319 
1320 		wake_up_interruptible(&event_file->ev_queue.poll_wait);
1321 		kill_fasync(&event_file->ev_queue.async_queue, SIGIO, POLL_IN);
1322 	}
1323 	mutex_unlock(&uverbs_dev->lists_mutex);
1324 
1325 	uverbs_disassociate_api(uverbs_dev->uapi);
1326 }
1327 
1328 static void ib_uverbs_remove_one(struct ib_device *device, void *client_data)
1329 {
1330 	struct ib_uverbs_device *uverbs_dev = client_data;
1331 	int wait_clients = 1;
1332 
1333 	if (!uverbs_dev)
1334 		return;
1335 
1336 	cdev_device_del(&uverbs_dev->cdev, &uverbs_dev->dev);
1337 	ida_free(&uverbs_ida, uverbs_dev->devnum);
1338 
1339 	if (device->ops.disassociate_ucontext) {
1340 		/* We disassociate HW resources and immediately return.
1341 		 * Userspace will see a EIO errno for all future access.
1342 		 * Upon returning, ib_device may be freed internally and is not
1343 		 * valid any more.
1344 		 * uverbs_device is still available until all clients close
1345 		 * their files, then the uverbs device ref count will be zero
1346 		 * and its resources will be freed.
1347 		 * Note: At this point no more files can be opened since the
1348 		 * cdev was deleted, however active clients can still issue
1349 		 * commands and close their open files.
1350 		 */
1351 		ib_uverbs_free_hw_resources(uverbs_dev, device);
1352 		wait_clients = 0;
1353 	}
1354 
1355 	if (atomic_dec_and_test(&uverbs_dev->refcount))
1356 		ib_uverbs_comp_dev(uverbs_dev);
1357 	if (wait_clients)
1358 		wait_for_completion(&uverbs_dev->comp);
1359 
1360 	put_device(&uverbs_dev->dev);
1361 }
1362 
1363 static char *uverbs_devnode(struct device *dev, umode_t *mode)
1364 {
1365 	if (mode)
1366 		*mode = 0666;
1367 	return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev));
1368 }
1369 
1370 static int __init ib_uverbs_init(void)
1371 {
1372 	int ret;
1373 
1374 	ret = register_chrdev_region(IB_UVERBS_BASE_DEV,
1375 				     IB_UVERBS_NUM_FIXED_MINOR,
1376 				     "infiniband_verbs");
1377 	if (ret) {
1378 		pr_err("user_verbs: couldn't register device number\n");
1379 		goto out;
1380 	}
1381 
1382 	ret = alloc_chrdev_region(&dynamic_uverbs_dev, 0,
1383 				  IB_UVERBS_NUM_DYNAMIC_MINOR,
1384 				  "infiniband_verbs");
1385 	if (ret) {
1386 		pr_err("couldn't register dynamic device number\n");
1387 		goto out_alloc;
1388 	}
1389 
1390 	uverbs_class = class_create(THIS_MODULE, "infiniband_verbs");
1391 	if (IS_ERR(uverbs_class)) {
1392 		ret = PTR_ERR(uverbs_class);
1393 		pr_err("user_verbs: couldn't create class infiniband_verbs\n");
1394 		goto out_chrdev;
1395 	}
1396 
1397 	uverbs_class->devnode = uverbs_devnode;
1398 
1399 	ret = class_create_file(uverbs_class, &class_attr_abi_version.attr);
1400 	if (ret) {
1401 		pr_err("user_verbs: couldn't create abi_version attribute\n");
1402 		goto out_class;
1403 	}
1404 
1405 	ret = ib_register_client(&uverbs_client);
1406 	if (ret) {
1407 		pr_err("user_verbs: couldn't register client\n");
1408 		goto out_class;
1409 	}
1410 
1411 	return 0;
1412 
1413 out_class:
1414 	class_destroy(uverbs_class);
1415 
1416 out_chrdev:
1417 	unregister_chrdev_region(dynamic_uverbs_dev,
1418 				 IB_UVERBS_NUM_DYNAMIC_MINOR);
1419 
1420 out_alloc:
1421 	unregister_chrdev_region(IB_UVERBS_BASE_DEV,
1422 				 IB_UVERBS_NUM_FIXED_MINOR);
1423 
1424 out:
1425 	return ret;
1426 }
1427 
1428 static void __exit ib_uverbs_cleanup(void)
1429 {
1430 	ib_unregister_client(&uverbs_client);
1431 	class_destroy(uverbs_class);
1432 	unregister_chrdev_region(IB_UVERBS_BASE_DEV,
1433 				 IB_UVERBS_NUM_FIXED_MINOR);
1434 	unregister_chrdev_region(dynamic_uverbs_dev,
1435 				 IB_UVERBS_NUM_DYNAMIC_MINOR);
1436 }
1437 
1438 module_init(ib_uverbs_init);
1439 module_exit(ib_uverbs_cleanup);
1440