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 #include <rdma/rdma_netlink.h>
55 
56 #include "uverbs.h"
57 #include "core_priv.h"
58 #include "rdma_core.h"
59 
60 MODULE_AUTHOR("Roland Dreier");
61 MODULE_DESCRIPTION("InfiniBand userspace verbs access");
62 MODULE_LICENSE("Dual BSD/GPL");
63 
64 enum {
65 	IB_UVERBS_MAJOR       = 231,
66 	IB_UVERBS_BASE_MINOR  = 192,
67 	IB_UVERBS_MAX_DEVICES = RDMA_MAX_PORTS,
68 	IB_UVERBS_NUM_FIXED_MINOR = 32,
69 	IB_UVERBS_NUM_DYNAMIC_MINOR = IB_UVERBS_MAX_DEVICES - IB_UVERBS_NUM_FIXED_MINOR,
70 };
71 
72 #define IB_UVERBS_BASE_DEV	MKDEV(IB_UVERBS_MAJOR, IB_UVERBS_BASE_MINOR)
73 
74 static dev_t dynamic_uverbs_dev;
75 static struct class *uverbs_class;
76 
77 static DEFINE_IDA(uverbs_ida);
78 static void ib_uverbs_add_one(struct ib_device *device);
79 static void ib_uverbs_remove_one(struct ib_device *device, void *client_data);
80 
81 /*
82  * Must be called with the ufile->device->disassociate_srcu held, and the lock
83  * must be held until use of the ucontext is finished.
84  */
85 struct ib_ucontext *ib_uverbs_get_ucontext_file(struct ib_uverbs_file *ufile)
86 {
87 	/*
88 	 * We do not hold the hw_destroy_rwsem lock for this flow, instead
89 	 * srcu is used. It does not matter if someone races this with
90 	 * get_context, we get NULL or valid ucontext.
91 	 */
92 	struct ib_ucontext *ucontext = smp_load_acquire(&ufile->ucontext);
93 
94 	if (!srcu_dereference(ufile->device->ib_dev,
95 			      &ufile->device->disassociate_srcu))
96 		return ERR_PTR(-EIO);
97 
98 	if (!ucontext)
99 		return ERR_PTR(-EINVAL);
100 
101 	return ucontext;
102 }
103 EXPORT_SYMBOL(ib_uverbs_get_ucontext_file);
104 
105 int uverbs_dealloc_mw(struct ib_mw *mw)
106 {
107 	struct ib_pd *pd = mw->pd;
108 	int ret;
109 
110 	ret = mw->device->ops.dealloc_mw(mw);
111 	if (!ret)
112 		atomic_dec(&pd->usecnt);
113 	return ret;
114 }
115 
116 static void ib_uverbs_release_dev(struct device *device)
117 {
118 	struct ib_uverbs_device *dev =
119 			container_of(device, struct ib_uverbs_device, dev);
120 
121 	uverbs_destroy_api(dev->uapi);
122 	cleanup_srcu_struct(&dev->disassociate_srcu);
123 	kfree(dev);
124 }
125 
126 static void ib_uverbs_release_async_event_file(struct kref *ref)
127 {
128 	struct ib_uverbs_async_event_file *file =
129 		container_of(ref, struct ib_uverbs_async_event_file, ref);
130 
131 	kfree(file);
132 }
133 
134 void ib_uverbs_release_ucq(struct ib_uverbs_file *file,
135 			  struct ib_uverbs_completion_event_file *ev_file,
136 			  struct ib_ucq_object *uobj)
137 {
138 	struct ib_uverbs_event *evt, *tmp;
139 
140 	if (ev_file) {
141 		spin_lock_irq(&ev_file->ev_queue.lock);
142 		list_for_each_entry_safe(evt, tmp, &uobj->comp_list, obj_list) {
143 			list_del(&evt->list);
144 			kfree(evt);
145 		}
146 		spin_unlock_irq(&ev_file->ev_queue.lock);
147 
148 		uverbs_uobject_put(&ev_file->uobj);
149 	}
150 
151 	spin_lock_irq(&file->async_file->ev_queue.lock);
152 	list_for_each_entry_safe(evt, tmp, &uobj->async_list, obj_list) {
153 		list_del(&evt->list);
154 		kfree(evt);
155 	}
156 	spin_unlock_irq(&file->async_file->ev_queue.lock);
157 }
158 
159 void ib_uverbs_release_uevent(struct ib_uverbs_file *file,
160 			      struct ib_uevent_object *uobj)
161 {
162 	struct ib_uverbs_event *evt, *tmp;
163 
164 	spin_lock_irq(&file->async_file->ev_queue.lock);
165 	list_for_each_entry_safe(evt, tmp, &uobj->event_list, obj_list) {
166 		list_del(&evt->list);
167 		kfree(evt);
168 	}
169 	spin_unlock_irq(&file->async_file->ev_queue.lock);
170 }
171 
172 void ib_uverbs_detach_umcast(struct ib_qp *qp,
173 			     struct ib_uqp_object *uobj)
174 {
175 	struct ib_uverbs_mcast_entry *mcast, *tmp;
176 
177 	list_for_each_entry_safe(mcast, tmp, &uobj->mcast_list, list) {
178 		ib_detach_mcast(qp, &mcast->gid, mcast->lid);
179 		list_del(&mcast->list);
180 		kfree(mcast);
181 	}
182 }
183 
184 static void ib_uverbs_comp_dev(struct ib_uverbs_device *dev)
185 {
186 	complete(&dev->comp);
187 }
188 
189 void ib_uverbs_release_file(struct kref *ref)
190 {
191 	struct ib_uverbs_file *file =
192 		container_of(ref, struct ib_uverbs_file, ref);
193 	struct ib_device *ib_dev;
194 	int srcu_key;
195 
196 	release_ufile_idr_uobject(file);
197 
198 	srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
199 	ib_dev = srcu_dereference(file->device->ib_dev,
200 				  &file->device->disassociate_srcu);
201 	if (ib_dev && !ib_dev->ops.disassociate_ucontext)
202 		module_put(ib_dev->ops.owner);
203 	srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
204 
205 	if (atomic_dec_and_test(&file->device->refcount))
206 		ib_uverbs_comp_dev(file->device);
207 
208 	if (file->async_file)
209 		kref_put(&file->async_file->ref,
210 			 ib_uverbs_release_async_event_file);
211 	put_device(&file->device->dev);
212 
213 	if (file->disassociate_page)
214 		__free_pages(file->disassociate_page, 0);
215 	kfree(file);
216 }
217 
218 static ssize_t ib_uverbs_event_read(struct ib_uverbs_event_queue *ev_queue,
219 				    struct ib_uverbs_file *uverbs_file,
220 				    struct file *filp, char __user *buf,
221 				    size_t count, loff_t *pos,
222 				    size_t eventsz)
223 {
224 	struct ib_uverbs_event *event;
225 	int ret = 0;
226 
227 	spin_lock_irq(&ev_queue->lock);
228 
229 	while (list_empty(&ev_queue->event_list)) {
230 		spin_unlock_irq(&ev_queue->lock);
231 
232 		if (filp->f_flags & O_NONBLOCK)
233 			return -EAGAIN;
234 
235 		if (wait_event_interruptible(ev_queue->poll_wait,
236 					     (!list_empty(&ev_queue->event_list) ||
237 			/* The barriers built into wait_event_interruptible()
238 			 * and wake_up() guarentee this will see the null set
239 			 * without using RCU
240 			 */
241 					     !uverbs_file->device->ib_dev)))
242 			return -ERESTARTSYS;
243 
244 		/* If device was disassociated and no event exists set an error */
245 		if (list_empty(&ev_queue->event_list) &&
246 		    !uverbs_file->device->ib_dev)
247 			return -EIO;
248 
249 		spin_lock_irq(&ev_queue->lock);
250 	}
251 
252 	event = list_entry(ev_queue->event_list.next, struct ib_uverbs_event, list);
253 
254 	if (eventsz > count) {
255 		ret   = -EINVAL;
256 		event = NULL;
257 	} else {
258 		list_del(ev_queue->event_list.next);
259 		if (event->counter) {
260 			++(*event->counter);
261 			list_del(&event->obj_list);
262 		}
263 	}
264 
265 	spin_unlock_irq(&ev_queue->lock);
266 
267 	if (event) {
268 		if (copy_to_user(buf, event, eventsz))
269 			ret = -EFAULT;
270 		else
271 			ret = eventsz;
272 	}
273 
274 	kfree(event);
275 
276 	return ret;
277 }
278 
279 static ssize_t ib_uverbs_async_event_read(struct file *filp, char __user *buf,
280 					  size_t count, loff_t *pos)
281 {
282 	struct ib_uverbs_async_event_file *file = filp->private_data;
283 
284 	return ib_uverbs_event_read(&file->ev_queue, file->uverbs_file, filp,
285 				    buf, count, pos,
286 				    sizeof(struct ib_uverbs_async_event_desc));
287 }
288 
289 static ssize_t ib_uverbs_comp_event_read(struct file *filp, char __user *buf,
290 					 size_t count, loff_t *pos)
291 {
292 	struct ib_uverbs_completion_event_file *comp_ev_file =
293 		filp->private_data;
294 
295 	return ib_uverbs_event_read(&comp_ev_file->ev_queue,
296 				    comp_ev_file->uobj.ufile, filp,
297 				    buf, count, pos,
298 				    sizeof(struct ib_uverbs_comp_event_desc));
299 }
300 
301 static __poll_t ib_uverbs_event_poll(struct ib_uverbs_event_queue *ev_queue,
302 					 struct file *filp,
303 					 struct poll_table_struct *wait)
304 {
305 	__poll_t pollflags = 0;
306 
307 	poll_wait(filp, &ev_queue->poll_wait, wait);
308 
309 	spin_lock_irq(&ev_queue->lock);
310 	if (!list_empty(&ev_queue->event_list))
311 		pollflags = EPOLLIN | EPOLLRDNORM;
312 	spin_unlock_irq(&ev_queue->lock);
313 
314 	return pollflags;
315 }
316 
317 static __poll_t ib_uverbs_async_event_poll(struct file *filp,
318 					       struct poll_table_struct *wait)
319 {
320 	return ib_uverbs_event_poll(filp->private_data, filp, wait);
321 }
322 
323 static __poll_t ib_uverbs_comp_event_poll(struct file *filp,
324 					      struct poll_table_struct *wait)
325 {
326 	struct ib_uverbs_completion_event_file *comp_ev_file =
327 		filp->private_data;
328 
329 	return ib_uverbs_event_poll(&comp_ev_file->ev_queue, filp, wait);
330 }
331 
332 static int ib_uverbs_async_event_fasync(int fd, struct file *filp, int on)
333 {
334 	struct ib_uverbs_event_queue *ev_queue = filp->private_data;
335 
336 	return fasync_helper(fd, filp, on, &ev_queue->async_queue);
337 }
338 
339 static int ib_uverbs_comp_event_fasync(int fd, struct file *filp, int on)
340 {
341 	struct ib_uverbs_completion_event_file *comp_ev_file =
342 		filp->private_data;
343 
344 	return fasync_helper(fd, filp, on, &comp_ev_file->ev_queue.async_queue);
345 }
346 
347 static int ib_uverbs_async_event_close(struct inode *inode, struct file *filp)
348 {
349 	struct ib_uverbs_async_event_file *file = filp->private_data;
350 	struct ib_uverbs_file *uverbs_file = file->uverbs_file;
351 	struct ib_uverbs_event *entry, *tmp;
352 	int closed_already = 0;
353 
354 	mutex_lock(&uverbs_file->device->lists_mutex);
355 	spin_lock_irq(&file->ev_queue.lock);
356 	closed_already = file->ev_queue.is_closed;
357 	file->ev_queue.is_closed = 1;
358 	list_for_each_entry_safe(entry, tmp, &file->ev_queue.event_list, list) {
359 		if (entry->counter)
360 			list_del(&entry->obj_list);
361 		kfree(entry);
362 	}
363 	spin_unlock_irq(&file->ev_queue.lock);
364 	if (!closed_already) {
365 		list_del(&file->list);
366 		ib_unregister_event_handler(&uverbs_file->event_handler);
367 	}
368 	mutex_unlock(&uverbs_file->device->lists_mutex);
369 
370 	kref_put(&uverbs_file->ref, ib_uverbs_release_file);
371 	kref_put(&file->ref, ib_uverbs_release_async_event_file);
372 
373 	return 0;
374 }
375 
376 static int ib_uverbs_comp_event_close(struct inode *inode, struct file *filp)
377 {
378 	struct ib_uobject *uobj = filp->private_data;
379 	struct ib_uverbs_completion_event_file *file = container_of(
380 		uobj, struct ib_uverbs_completion_event_file, uobj);
381 	struct ib_uverbs_event *entry, *tmp;
382 
383 	spin_lock_irq(&file->ev_queue.lock);
384 	list_for_each_entry_safe(entry, tmp, &file->ev_queue.event_list, list) {
385 		if (entry->counter)
386 			list_del(&entry->obj_list);
387 		kfree(entry);
388 	}
389 	file->ev_queue.is_closed = 1;
390 	spin_unlock_irq(&file->ev_queue.lock);
391 
392 	uverbs_close_fd(filp);
393 
394 	return 0;
395 }
396 
397 const struct file_operations uverbs_event_fops = {
398 	.owner	 = THIS_MODULE,
399 	.read	 = ib_uverbs_comp_event_read,
400 	.poll    = ib_uverbs_comp_event_poll,
401 	.release = ib_uverbs_comp_event_close,
402 	.fasync  = ib_uverbs_comp_event_fasync,
403 	.llseek	 = no_llseek,
404 };
405 
406 static const struct file_operations uverbs_async_event_fops = {
407 	.owner	 = THIS_MODULE,
408 	.read	 = ib_uverbs_async_event_read,
409 	.poll    = ib_uverbs_async_event_poll,
410 	.release = ib_uverbs_async_event_close,
411 	.fasync  = ib_uverbs_async_event_fasync,
412 	.llseek	 = no_llseek,
413 };
414 
415 void ib_uverbs_comp_handler(struct ib_cq *cq, void *cq_context)
416 {
417 	struct ib_uverbs_event_queue   *ev_queue = cq_context;
418 	struct ib_ucq_object	       *uobj;
419 	struct ib_uverbs_event	       *entry;
420 	unsigned long			flags;
421 
422 	if (!ev_queue)
423 		return;
424 
425 	spin_lock_irqsave(&ev_queue->lock, flags);
426 	if (ev_queue->is_closed) {
427 		spin_unlock_irqrestore(&ev_queue->lock, flags);
428 		return;
429 	}
430 
431 	entry = kmalloc(sizeof(*entry), GFP_ATOMIC);
432 	if (!entry) {
433 		spin_unlock_irqrestore(&ev_queue->lock, flags);
434 		return;
435 	}
436 
437 	uobj = container_of(cq->uobject, struct ib_ucq_object, uobject);
438 
439 	entry->desc.comp.cq_handle = cq->uobject->user_handle;
440 	entry->counter		   = &uobj->comp_events_reported;
441 
442 	list_add_tail(&entry->list, &ev_queue->event_list);
443 	list_add_tail(&entry->obj_list, &uobj->comp_list);
444 	spin_unlock_irqrestore(&ev_queue->lock, flags);
445 
446 	wake_up_interruptible(&ev_queue->poll_wait);
447 	kill_fasync(&ev_queue->async_queue, SIGIO, POLL_IN);
448 }
449 
450 static void ib_uverbs_async_handler(struct ib_uverbs_file *file,
451 				    __u64 element, __u64 event,
452 				    struct list_head *obj_list,
453 				    u32 *counter)
454 {
455 	struct ib_uverbs_event *entry;
456 	unsigned long flags;
457 
458 	spin_lock_irqsave(&file->async_file->ev_queue.lock, flags);
459 	if (file->async_file->ev_queue.is_closed) {
460 		spin_unlock_irqrestore(&file->async_file->ev_queue.lock, flags);
461 		return;
462 	}
463 
464 	entry = kmalloc(sizeof(*entry), GFP_ATOMIC);
465 	if (!entry) {
466 		spin_unlock_irqrestore(&file->async_file->ev_queue.lock, flags);
467 		return;
468 	}
469 
470 	entry->desc.async.element    = element;
471 	entry->desc.async.event_type = event;
472 	entry->desc.async.reserved   = 0;
473 	entry->counter               = counter;
474 
475 	list_add_tail(&entry->list, &file->async_file->ev_queue.event_list);
476 	if (obj_list)
477 		list_add_tail(&entry->obj_list, obj_list);
478 	spin_unlock_irqrestore(&file->async_file->ev_queue.lock, flags);
479 
480 	wake_up_interruptible(&file->async_file->ev_queue.poll_wait);
481 	kill_fasync(&file->async_file->ev_queue.async_queue, SIGIO, POLL_IN);
482 }
483 
484 void ib_uverbs_cq_event_handler(struct ib_event *event, void *context_ptr)
485 {
486 	struct ib_ucq_object *uobj = container_of(event->element.cq->uobject,
487 						  struct ib_ucq_object, uobject);
488 
489 	ib_uverbs_async_handler(uobj->uobject.ufile, uobj->uobject.user_handle,
490 				event->event, &uobj->async_list,
491 				&uobj->async_events_reported);
492 }
493 
494 void ib_uverbs_qp_event_handler(struct ib_event *event, void *context_ptr)
495 {
496 	struct ib_uevent_object *uobj;
497 
498 	/* for XRC target qp's, check that qp is live */
499 	if (!event->element.qp->uobject)
500 		return;
501 
502 	uobj = container_of(event->element.qp->uobject,
503 			    struct ib_uevent_object, uobject);
504 
505 	ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
506 				event->event, &uobj->event_list,
507 				&uobj->events_reported);
508 }
509 
510 void ib_uverbs_wq_event_handler(struct ib_event *event, void *context_ptr)
511 {
512 	struct ib_uevent_object *uobj = container_of(event->element.wq->uobject,
513 						  struct ib_uevent_object, uobject);
514 
515 	ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
516 				event->event, &uobj->event_list,
517 				&uobj->events_reported);
518 }
519 
520 void ib_uverbs_srq_event_handler(struct ib_event *event, void *context_ptr)
521 {
522 	struct ib_uevent_object *uobj;
523 
524 	uobj = container_of(event->element.srq->uobject,
525 			    struct ib_uevent_object, uobject);
526 
527 	ib_uverbs_async_handler(context_ptr, uobj->uobject.user_handle,
528 				event->event, &uobj->event_list,
529 				&uobj->events_reported);
530 }
531 
532 void ib_uverbs_event_handler(struct ib_event_handler *handler,
533 			     struct ib_event *event)
534 {
535 	struct ib_uverbs_file *file =
536 		container_of(handler, struct ib_uverbs_file, event_handler);
537 
538 	ib_uverbs_async_handler(file, event->element.port_num, event->event,
539 				NULL, NULL);
540 }
541 
542 void ib_uverbs_free_async_event_file(struct ib_uverbs_file *file)
543 {
544 	kref_put(&file->async_file->ref, ib_uverbs_release_async_event_file);
545 	file->async_file = NULL;
546 }
547 
548 void ib_uverbs_init_event_queue(struct ib_uverbs_event_queue *ev_queue)
549 {
550 	spin_lock_init(&ev_queue->lock);
551 	INIT_LIST_HEAD(&ev_queue->event_list);
552 	init_waitqueue_head(&ev_queue->poll_wait);
553 	ev_queue->is_closed   = 0;
554 	ev_queue->async_queue = NULL;
555 }
556 
557 struct file *ib_uverbs_alloc_async_event_file(struct ib_uverbs_file *uverbs_file,
558 					      struct ib_device	*ib_dev)
559 {
560 	struct ib_uverbs_async_event_file *ev_file;
561 	struct file *filp;
562 
563 	ev_file = kzalloc(sizeof(*ev_file), GFP_KERNEL);
564 	if (!ev_file)
565 		return ERR_PTR(-ENOMEM);
566 
567 	ib_uverbs_init_event_queue(&ev_file->ev_queue);
568 	ev_file->uverbs_file = uverbs_file;
569 	kref_get(&ev_file->uverbs_file->ref);
570 	kref_init(&ev_file->ref);
571 	filp = anon_inode_getfile("[infinibandevent]", &uverbs_async_event_fops,
572 				  ev_file, O_RDONLY);
573 	if (IS_ERR(filp))
574 		goto err_put_refs;
575 
576 	mutex_lock(&uverbs_file->device->lists_mutex);
577 	list_add_tail(&ev_file->list,
578 		      &uverbs_file->device->uverbs_events_file_list);
579 	mutex_unlock(&uverbs_file->device->lists_mutex);
580 
581 	WARN_ON(uverbs_file->async_file);
582 	uverbs_file->async_file = ev_file;
583 	kref_get(&uverbs_file->async_file->ref);
584 	INIT_IB_EVENT_HANDLER(&uverbs_file->event_handler,
585 			      ib_dev,
586 			      ib_uverbs_event_handler);
587 	ib_register_event_handler(&uverbs_file->event_handler);
588 	/* At that point async file stuff was fully set */
589 
590 	return filp;
591 
592 err_put_refs:
593 	kref_put(&ev_file->uverbs_file->ref, ib_uverbs_release_file);
594 	kref_put(&ev_file->ref, ib_uverbs_release_async_event_file);
595 	return filp;
596 }
597 
598 static ssize_t verify_hdr(struct ib_uverbs_cmd_hdr *hdr,
599 			  struct ib_uverbs_ex_cmd_hdr *ex_hdr, size_t count,
600 			  const struct uverbs_api_write_method *method_elm)
601 {
602 	if (method_elm->is_ex) {
603 		count -= sizeof(*hdr) + sizeof(*ex_hdr);
604 
605 		if ((hdr->in_words + ex_hdr->provider_in_words) * 8 != count)
606 			return -EINVAL;
607 
608 		if (hdr->in_words * 8 < method_elm->req_size)
609 			return -ENOSPC;
610 
611 		if (ex_hdr->cmd_hdr_reserved)
612 			return -EINVAL;
613 
614 		if (ex_hdr->response) {
615 			if (!hdr->out_words && !ex_hdr->provider_out_words)
616 				return -EINVAL;
617 
618 			if (hdr->out_words * 8 < method_elm->resp_size)
619 				return -ENOSPC;
620 
621 			if (!access_ok(u64_to_user_ptr(ex_hdr->response),
622 				       (hdr->out_words + ex_hdr->provider_out_words) * 8))
623 				return -EFAULT;
624 		} else {
625 			if (hdr->out_words || ex_hdr->provider_out_words)
626 				return -EINVAL;
627 		}
628 
629 		return 0;
630 	}
631 
632 	/* not extended command */
633 	if (hdr->in_words * 4 != count)
634 		return -EINVAL;
635 
636 	if (count < method_elm->req_size + sizeof(hdr)) {
637 		/*
638 		 * rdma-core v18 and v19 have a bug where they send DESTROY_CQ
639 		 * with a 16 byte write instead of 24. Old kernels didn't
640 		 * check the size so they allowed this. Now that the size is
641 		 * checked provide a compatibility work around to not break
642 		 * those userspaces.
643 		 */
644 		if (hdr->command == IB_USER_VERBS_CMD_DESTROY_CQ &&
645 		    count == 16) {
646 			hdr->in_words = 6;
647 			return 0;
648 		}
649 		return -ENOSPC;
650 	}
651 	if (hdr->out_words * 4 < method_elm->resp_size)
652 		return -ENOSPC;
653 
654 	return 0;
655 }
656 
657 static ssize_t ib_uverbs_write(struct file *filp, const char __user *buf,
658 			     size_t count, loff_t *pos)
659 {
660 	struct ib_uverbs_file *file = filp->private_data;
661 	const struct uverbs_api_write_method *method_elm;
662 	struct uverbs_api *uapi = file->device->uapi;
663 	struct ib_uverbs_ex_cmd_hdr ex_hdr;
664 	struct ib_uverbs_cmd_hdr hdr;
665 	struct uverbs_attr_bundle bundle;
666 	int srcu_key;
667 	ssize_t ret;
668 
669 	if (!ib_safe_file_access(filp)) {
670 		pr_err_once("uverbs_write: process %d (%s) changed security contexts after opening file descriptor, this is not allowed.\n",
671 			    task_tgid_vnr(current), current->comm);
672 		return -EACCES;
673 	}
674 
675 	if (count < sizeof(hdr))
676 		return -EINVAL;
677 
678 	if (copy_from_user(&hdr, buf, sizeof(hdr)))
679 		return -EFAULT;
680 
681 	method_elm = uapi_get_method(uapi, hdr.command);
682 	if (IS_ERR(method_elm))
683 		return PTR_ERR(method_elm);
684 
685 	if (method_elm->is_ex) {
686 		if (count < (sizeof(hdr) + sizeof(ex_hdr)))
687 			return -EINVAL;
688 		if (copy_from_user(&ex_hdr, buf + sizeof(hdr), sizeof(ex_hdr)))
689 			return -EFAULT;
690 	}
691 
692 	ret = verify_hdr(&hdr, &ex_hdr, count, method_elm);
693 	if (ret)
694 		return ret;
695 
696 	srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
697 
698 	buf += sizeof(hdr);
699 
700 	memset(bundle.attr_present, 0, sizeof(bundle.attr_present));
701 	bundle.ufile = file;
702 	bundle.context = NULL; /* only valid if bundle has uobject */
703 	if (!method_elm->is_ex) {
704 		size_t in_len = hdr.in_words * 4 - sizeof(hdr);
705 		size_t out_len = hdr.out_words * 4;
706 		u64 response = 0;
707 
708 		if (method_elm->has_udata) {
709 			bundle.driver_udata.inlen =
710 				in_len - method_elm->req_size;
711 			in_len = method_elm->req_size;
712 			if (bundle.driver_udata.inlen)
713 				bundle.driver_udata.inbuf = buf + in_len;
714 			else
715 				bundle.driver_udata.inbuf = NULL;
716 		} else {
717 			memset(&bundle.driver_udata, 0,
718 			       sizeof(bundle.driver_udata));
719 		}
720 
721 		if (method_elm->has_resp) {
722 			/*
723 			 * The macros check that if has_resp is set
724 			 * then the command request structure starts
725 			 * with a '__aligned u64 response' member.
726 			 */
727 			ret = get_user(response, (const u64 __user *)buf);
728 			if (ret)
729 				goto out_unlock;
730 
731 			if (method_elm->has_udata) {
732 				bundle.driver_udata.outlen =
733 					out_len - method_elm->resp_size;
734 				out_len = method_elm->resp_size;
735 				if (bundle.driver_udata.outlen)
736 					bundle.driver_udata.outbuf =
737 						u64_to_user_ptr(response +
738 								out_len);
739 				else
740 					bundle.driver_udata.outbuf = NULL;
741 			}
742 		} else {
743 			bundle.driver_udata.outlen = 0;
744 			bundle.driver_udata.outbuf = NULL;
745 		}
746 
747 		ib_uverbs_init_udata_buf_or_null(
748 			&bundle.ucore, buf, u64_to_user_ptr(response),
749 			in_len, out_len);
750 	} else {
751 		buf += sizeof(ex_hdr);
752 
753 		ib_uverbs_init_udata_buf_or_null(&bundle.ucore, buf,
754 					u64_to_user_ptr(ex_hdr.response),
755 					hdr.in_words * 8, hdr.out_words * 8);
756 
757 		ib_uverbs_init_udata_buf_or_null(
758 			&bundle.driver_udata, buf + bundle.ucore.inlen,
759 			u64_to_user_ptr(ex_hdr.response) + bundle.ucore.outlen,
760 			ex_hdr.provider_in_words * 8,
761 			ex_hdr.provider_out_words * 8);
762 
763 	}
764 
765 	ret = method_elm->handler(&bundle);
766 out_unlock:
767 	srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
768 	return (ret) ? : count;
769 }
770 
771 static int ib_uverbs_mmap(struct file *filp, struct vm_area_struct *vma)
772 {
773 	struct ib_uverbs_file *file = filp->private_data;
774 	struct ib_ucontext *ucontext;
775 	int ret = 0;
776 	int srcu_key;
777 
778 	srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
779 	ucontext = ib_uverbs_get_ucontext_file(file);
780 	if (IS_ERR(ucontext)) {
781 		ret = PTR_ERR(ucontext);
782 		goto out;
783 	}
784 
785 	ret = ucontext->device->ops.mmap(ucontext, vma);
786 out:
787 	srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
788 	return ret;
789 }
790 
791 /*
792  * Each time we map IO memory into user space this keeps track of the mapping.
793  * When the device is hot-unplugged we 'zap' the mmaps in user space to point
794  * to the zero page and allow the hot unplug to proceed.
795  *
796  * This is necessary for cases like PCI physical hot unplug as the actual BAR
797  * memory may vanish after this and access to it from userspace could MCE.
798  *
799  * RDMA drivers supporting disassociation must have their user space designed
800  * to cope in some way with their IO pages going to the zero page.
801  */
802 struct rdma_umap_priv {
803 	struct vm_area_struct *vma;
804 	struct list_head list;
805 };
806 
807 static const struct vm_operations_struct rdma_umap_ops;
808 
809 static void rdma_umap_priv_init(struct rdma_umap_priv *priv,
810 				struct vm_area_struct *vma)
811 {
812 	struct ib_uverbs_file *ufile = vma->vm_file->private_data;
813 
814 	priv->vma = vma;
815 	vma->vm_private_data = priv;
816 	vma->vm_ops = &rdma_umap_ops;
817 
818 	mutex_lock(&ufile->umap_lock);
819 	list_add(&priv->list, &ufile->umaps);
820 	mutex_unlock(&ufile->umap_lock);
821 }
822 
823 /*
824  * The VMA has been dup'd, initialize the vm_private_data with a new tracking
825  * struct
826  */
827 static void rdma_umap_open(struct vm_area_struct *vma)
828 {
829 	struct ib_uverbs_file *ufile = vma->vm_file->private_data;
830 	struct rdma_umap_priv *opriv = vma->vm_private_data;
831 	struct rdma_umap_priv *priv;
832 
833 	if (!opriv)
834 		return;
835 
836 	/* We are racing with disassociation */
837 	if (!down_read_trylock(&ufile->hw_destroy_rwsem))
838 		goto out_zap;
839 	/*
840 	 * Disassociation already completed, the VMA should already be zapped.
841 	 */
842 	if (!ufile->ucontext)
843 		goto out_unlock;
844 
845 	priv = kzalloc(sizeof(*priv), GFP_KERNEL);
846 	if (!priv)
847 		goto out_unlock;
848 	rdma_umap_priv_init(priv, vma);
849 
850 	up_read(&ufile->hw_destroy_rwsem);
851 	return;
852 
853 out_unlock:
854 	up_read(&ufile->hw_destroy_rwsem);
855 out_zap:
856 	/*
857 	 * We can't allow the VMA to be created with the actual IO pages, that
858 	 * would break our API contract, and it can't be stopped at this
859 	 * point, so zap it.
860 	 */
861 	vma->vm_private_data = NULL;
862 	zap_vma_ptes(vma, vma->vm_start, vma->vm_end - vma->vm_start);
863 }
864 
865 static void rdma_umap_close(struct vm_area_struct *vma)
866 {
867 	struct ib_uverbs_file *ufile = vma->vm_file->private_data;
868 	struct rdma_umap_priv *priv = vma->vm_private_data;
869 
870 	if (!priv)
871 		return;
872 
873 	/*
874 	 * The vma holds a reference on the struct file that created it, which
875 	 * in turn means that the ib_uverbs_file is guaranteed to exist at
876 	 * this point.
877 	 */
878 	mutex_lock(&ufile->umap_lock);
879 	list_del(&priv->list);
880 	mutex_unlock(&ufile->umap_lock);
881 	kfree(priv);
882 }
883 
884 /*
885  * Once the zap_vma_ptes has been called touches to the VMA will come here and
886  * we return a dummy writable zero page for all the pfns.
887  */
888 static vm_fault_t rdma_umap_fault(struct vm_fault *vmf)
889 {
890 	struct ib_uverbs_file *ufile = vmf->vma->vm_file->private_data;
891 	struct rdma_umap_priv *priv = vmf->vma->vm_private_data;
892 	vm_fault_t ret = 0;
893 
894 	if (!priv)
895 		return VM_FAULT_SIGBUS;
896 
897 	/* Read only pages can just use the system zero page. */
898 	if (!(vmf->vma->vm_flags & (VM_WRITE | VM_MAYWRITE))) {
899 		vmf->page = ZERO_PAGE(vmf->address);
900 		get_page(vmf->page);
901 		return 0;
902 	}
903 
904 	mutex_lock(&ufile->umap_lock);
905 	if (!ufile->disassociate_page)
906 		ufile->disassociate_page =
907 			alloc_pages(vmf->gfp_mask | __GFP_ZERO, 0);
908 
909 	if (ufile->disassociate_page) {
910 		/*
911 		 * This VMA is forced to always be shared so this doesn't have
912 		 * to worry about COW.
913 		 */
914 		vmf->page = ufile->disassociate_page;
915 		get_page(vmf->page);
916 	} else {
917 		ret = VM_FAULT_SIGBUS;
918 	}
919 	mutex_unlock(&ufile->umap_lock);
920 
921 	return ret;
922 }
923 
924 static const struct vm_operations_struct rdma_umap_ops = {
925 	.open = rdma_umap_open,
926 	.close = rdma_umap_close,
927 	.fault = rdma_umap_fault,
928 };
929 
930 /*
931  * Map IO memory into a process. This is to be called by drivers as part of
932  * their mmap() functions if they wish to send something like PCI-E BAR memory
933  * to userspace.
934  */
935 int rdma_user_mmap_io(struct ib_ucontext *ucontext, struct vm_area_struct *vma,
936 		      unsigned long pfn, unsigned long size, pgprot_t prot)
937 {
938 	struct ib_uverbs_file *ufile = ucontext->ufile;
939 	struct rdma_umap_priv *priv;
940 
941 	if (!(vma->vm_flags & VM_SHARED))
942 		return -EINVAL;
943 
944 	if (vma->vm_end - vma->vm_start != size)
945 		return -EINVAL;
946 
947 	/* Driver is using this wrong, must be called by ib_uverbs_mmap */
948 	if (WARN_ON(!vma->vm_file ||
949 		    vma->vm_file->private_data != ufile))
950 		return -EINVAL;
951 	lockdep_assert_held(&ufile->device->disassociate_srcu);
952 
953 	priv = kzalloc(sizeof(*priv), GFP_KERNEL);
954 	if (!priv)
955 		return -ENOMEM;
956 
957 	vma->vm_page_prot = prot;
958 	if (io_remap_pfn_range(vma, vma->vm_start, pfn, size, prot)) {
959 		kfree(priv);
960 		return -EAGAIN;
961 	}
962 
963 	rdma_umap_priv_init(priv, vma);
964 	return 0;
965 }
966 EXPORT_SYMBOL(rdma_user_mmap_io);
967 
968 void uverbs_user_mmap_disassociate(struct ib_uverbs_file *ufile)
969 {
970 	struct rdma_umap_priv *priv, *next_priv;
971 
972 	lockdep_assert_held(&ufile->hw_destroy_rwsem);
973 
974 	while (1) {
975 		struct mm_struct *mm = NULL;
976 
977 		/* Get an arbitrary mm pointer that hasn't been cleaned yet */
978 		mutex_lock(&ufile->umap_lock);
979 		while (!list_empty(&ufile->umaps)) {
980 			int ret;
981 
982 			priv = list_first_entry(&ufile->umaps,
983 						struct rdma_umap_priv, list);
984 			mm = priv->vma->vm_mm;
985 			ret = mmget_not_zero(mm);
986 			if (!ret) {
987 				list_del_init(&priv->list);
988 				mm = NULL;
989 				continue;
990 			}
991 			break;
992 		}
993 		mutex_unlock(&ufile->umap_lock);
994 		if (!mm)
995 			return;
996 
997 		/*
998 		 * The umap_lock is nested under mmap_sem since it used within
999 		 * the vma_ops callbacks, so we have to clean the list one mm
1000 		 * at a time to get the lock ordering right. Typically there
1001 		 * will only be one mm, so no big deal.
1002 		 */
1003 		down_read(&mm->mmap_sem);
1004 		if (!mmget_still_valid(mm))
1005 			goto skip_mm;
1006 		mutex_lock(&ufile->umap_lock);
1007 		list_for_each_entry_safe (priv, next_priv, &ufile->umaps,
1008 					  list) {
1009 			struct vm_area_struct *vma = priv->vma;
1010 
1011 			if (vma->vm_mm != mm)
1012 				continue;
1013 			list_del_init(&priv->list);
1014 
1015 			zap_vma_ptes(vma, vma->vm_start,
1016 				     vma->vm_end - vma->vm_start);
1017 		}
1018 		mutex_unlock(&ufile->umap_lock);
1019 	skip_mm:
1020 		up_read(&mm->mmap_sem);
1021 		mmput(mm);
1022 	}
1023 }
1024 
1025 /*
1026  * ib_uverbs_open() does not need the BKL:
1027  *
1028  *  - the ib_uverbs_device structures are properly reference counted and
1029  *    everything else is purely local to the file being created, so
1030  *    races against other open calls are not a problem;
1031  *  - there is no ioctl method to race against;
1032  *  - the open method will either immediately run -ENXIO, or all
1033  *    required initialization will be done.
1034  */
1035 static int ib_uverbs_open(struct inode *inode, struct file *filp)
1036 {
1037 	struct ib_uverbs_device *dev;
1038 	struct ib_uverbs_file *file;
1039 	struct ib_device *ib_dev;
1040 	int ret;
1041 	int module_dependent;
1042 	int srcu_key;
1043 
1044 	dev = container_of(inode->i_cdev, struct ib_uverbs_device, cdev);
1045 	if (!atomic_inc_not_zero(&dev->refcount))
1046 		return -ENXIO;
1047 
1048 	get_device(&dev->dev);
1049 	srcu_key = srcu_read_lock(&dev->disassociate_srcu);
1050 	mutex_lock(&dev->lists_mutex);
1051 	ib_dev = srcu_dereference(dev->ib_dev,
1052 				  &dev->disassociate_srcu);
1053 	if (!ib_dev) {
1054 		ret = -EIO;
1055 		goto err;
1056 	}
1057 
1058 	if (!rdma_dev_access_netns(ib_dev, current->nsproxy->net_ns)) {
1059 		ret = -EPERM;
1060 		goto err;
1061 	}
1062 
1063 	/* In case IB device supports disassociate ucontext, there is no hard
1064 	 * dependency between uverbs device and its low level device.
1065 	 */
1066 	module_dependent = !(ib_dev->ops.disassociate_ucontext);
1067 
1068 	if (module_dependent) {
1069 		if (!try_module_get(ib_dev->ops.owner)) {
1070 			ret = -ENODEV;
1071 			goto err;
1072 		}
1073 	}
1074 
1075 	file = kzalloc(sizeof(*file), GFP_KERNEL);
1076 	if (!file) {
1077 		ret = -ENOMEM;
1078 		if (module_dependent)
1079 			goto err_module;
1080 
1081 		goto err;
1082 	}
1083 
1084 	file->device	 = dev;
1085 	kref_init(&file->ref);
1086 	mutex_init(&file->ucontext_lock);
1087 
1088 	spin_lock_init(&file->uobjects_lock);
1089 	INIT_LIST_HEAD(&file->uobjects);
1090 	init_rwsem(&file->hw_destroy_rwsem);
1091 	mutex_init(&file->umap_lock);
1092 	INIT_LIST_HEAD(&file->umaps);
1093 
1094 	filp->private_data = file;
1095 	list_add_tail(&file->list, &dev->uverbs_file_list);
1096 	mutex_unlock(&dev->lists_mutex);
1097 	srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1098 
1099 	setup_ufile_idr_uobject(file);
1100 
1101 	return stream_open(inode, filp);
1102 
1103 err_module:
1104 	module_put(ib_dev->ops.owner);
1105 
1106 err:
1107 	mutex_unlock(&dev->lists_mutex);
1108 	srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1109 	if (atomic_dec_and_test(&dev->refcount))
1110 		ib_uverbs_comp_dev(dev);
1111 
1112 	put_device(&dev->dev);
1113 	return ret;
1114 }
1115 
1116 static int ib_uverbs_close(struct inode *inode, struct file *filp)
1117 {
1118 	struct ib_uverbs_file *file = filp->private_data;
1119 
1120 	uverbs_destroy_ufile_hw(file, RDMA_REMOVE_CLOSE);
1121 
1122 	mutex_lock(&file->device->lists_mutex);
1123 	list_del_init(&file->list);
1124 	mutex_unlock(&file->device->lists_mutex);
1125 
1126 	kref_put(&file->ref, ib_uverbs_release_file);
1127 
1128 	return 0;
1129 }
1130 
1131 static const struct file_operations uverbs_fops = {
1132 	.owner	 = THIS_MODULE,
1133 	.write	 = ib_uverbs_write,
1134 	.open	 = ib_uverbs_open,
1135 	.release = ib_uverbs_close,
1136 	.llseek	 = no_llseek,
1137 	.unlocked_ioctl = ib_uverbs_ioctl,
1138 	.compat_ioctl = ib_uverbs_ioctl,
1139 };
1140 
1141 static const struct file_operations uverbs_mmap_fops = {
1142 	.owner	 = THIS_MODULE,
1143 	.write	 = ib_uverbs_write,
1144 	.mmap    = ib_uverbs_mmap,
1145 	.open	 = ib_uverbs_open,
1146 	.release = ib_uverbs_close,
1147 	.llseek	 = no_llseek,
1148 	.unlocked_ioctl = ib_uverbs_ioctl,
1149 	.compat_ioctl = ib_uverbs_ioctl,
1150 };
1151 
1152 static int ib_uverbs_get_nl_info(struct ib_device *ibdev, void *client_data,
1153 				 struct ib_client_nl_info *res)
1154 {
1155 	struct ib_uverbs_device *uverbs_dev = client_data;
1156 	int ret;
1157 
1158 	if (res->port != -1)
1159 		return -EINVAL;
1160 
1161 	res->abi = ibdev->ops.uverbs_abi_ver;
1162 	res->cdev = &uverbs_dev->dev;
1163 
1164 	/*
1165 	 * To support DRIVER_ID binding in userspace some of the driver need
1166 	 * upgrading to expose their PCI dependent revision information
1167 	 * through get_context instead of relying on modalias matching. When
1168 	 * the drivers are fixed they can drop this flag.
1169 	 */
1170 	if (!ibdev->ops.uverbs_no_driver_id_binding) {
1171 		ret = nla_put_u32(res->nl_msg, RDMA_NLDEV_ATTR_UVERBS_DRIVER_ID,
1172 				  ibdev->ops.driver_id);
1173 		if (ret)
1174 			return ret;
1175 	}
1176 	return 0;
1177 }
1178 
1179 static struct ib_client uverbs_client = {
1180 	.name   = "uverbs",
1181 	.no_kverbs_req = true,
1182 	.add    = ib_uverbs_add_one,
1183 	.remove = ib_uverbs_remove_one,
1184 	.get_nl_info = ib_uverbs_get_nl_info,
1185 };
1186 MODULE_ALIAS_RDMA_CLIENT("uverbs");
1187 
1188 static ssize_t ibdev_show(struct device *device, struct device_attribute *attr,
1189 			  char *buf)
1190 {
1191 	struct ib_uverbs_device *dev =
1192 			container_of(device, struct ib_uverbs_device, dev);
1193 	int ret = -ENODEV;
1194 	int srcu_key;
1195 	struct ib_device *ib_dev;
1196 
1197 	srcu_key = srcu_read_lock(&dev->disassociate_srcu);
1198 	ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu);
1199 	if (ib_dev)
1200 		ret = sprintf(buf, "%s\n", dev_name(&ib_dev->dev));
1201 	srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1202 
1203 	return ret;
1204 }
1205 static DEVICE_ATTR_RO(ibdev);
1206 
1207 static ssize_t abi_version_show(struct device *device,
1208 				struct device_attribute *attr, char *buf)
1209 {
1210 	struct ib_uverbs_device *dev =
1211 			container_of(device, struct ib_uverbs_device, dev);
1212 	int ret = -ENODEV;
1213 	int srcu_key;
1214 	struct ib_device *ib_dev;
1215 
1216 	srcu_key = srcu_read_lock(&dev->disassociate_srcu);
1217 	ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu);
1218 	if (ib_dev)
1219 		ret = sprintf(buf, "%u\n", ib_dev->ops.uverbs_abi_ver);
1220 	srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1221 
1222 	return ret;
1223 }
1224 static DEVICE_ATTR_RO(abi_version);
1225 
1226 static struct attribute *ib_dev_attrs[] = {
1227 	&dev_attr_abi_version.attr,
1228 	&dev_attr_ibdev.attr,
1229 	NULL,
1230 };
1231 
1232 static const struct attribute_group dev_attr_group = {
1233 	.attrs = ib_dev_attrs,
1234 };
1235 
1236 static CLASS_ATTR_STRING(abi_version, S_IRUGO,
1237 			 __stringify(IB_USER_VERBS_ABI_VERSION));
1238 
1239 static int ib_uverbs_create_uapi(struct ib_device *device,
1240 				 struct ib_uverbs_device *uverbs_dev)
1241 {
1242 	struct uverbs_api *uapi;
1243 
1244 	uapi = uverbs_alloc_api(device);
1245 	if (IS_ERR(uapi))
1246 		return PTR_ERR(uapi);
1247 
1248 	uverbs_dev->uapi = uapi;
1249 	return 0;
1250 }
1251 
1252 static void ib_uverbs_add_one(struct ib_device *device)
1253 {
1254 	int devnum;
1255 	dev_t base;
1256 	struct ib_uverbs_device *uverbs_dev;
1257 	int ret;
1258 
1259 	if (!device->ops.alloc_ucontext)
1260 		return;
1261 
1262 	uverbs_dev = kzalloc(sizeof(*uverbs_dev), GFP_KERNEL);
1263 	if (!uverbs_dev)
1264 		return;
1265 
1266 	ret = init_srcu_struct(&uverbs_dev->disassociate_srcu);
1267 	if (ret) {
1268 		kfree(uverbs_dev);
1269 		return;
1270 	}
1271 
1272 	device_initialize(&uverbs_dev->dev);
1273 	uverbs_dev->dev.class = uverbs_class;
1274 	uverbs_dev->dev.parent = device->dev.parent;
1275 	uverbs_dev->dev.release = ib_uverbs_release_dev;
1276 	uverbs_dev->groups[0] = &dev_attr_group;
1277 	uverbs_dev->dev.groups = uverbs_dev->groups;
1278 	atomic_set(&uverbs_dev->refcount, 1);
1279 	init_completion(&uverbs_dev->comp);
1280 	uverbs_dev->xrcd_tree = RB_ROOT;
1281 	mutex_init(&uverbs_dev->xrcd_tree_mutex);
1282 	mutex_init(&uverbs_dev->lists_mutex);
1283 	INIT_LIST_HEAD(&uverbs_dev->uverbs_file_list);
1284 	INIT_LIST_HEAD(&uverbs_dev->uverbs_events_file_list);
1285 	rcu_assign_pointer(uverbs_dev->ib_dev, device);
1286 	uverbs_dev->num_comp_vectors = device->num_comp_vectors;
1287 
1288 	devnum = ida_alloc_max(&uverbs_ida, IB_UVERBS_MAX_DEVICES - 1,
1289 			       GFP_KERNEL);
1290 	if (devnum < 0)
1291 		goto err;
1292 	uverbs_dev->devnum = devnum;
1293 	if (devnum >= IB_UVERBS_NUM_FIXED_MINOR)
1294 		base = dynamic_uverbs_dev + devnum - IB_UVERBS_NUM_FIXED_MINOR;
1295 	else
1296 		base = IB_UVERBS_BASE_DEV + devnum;
1297 
1298 	if (ib_uverbs_create_uapi(device, uverbs_dev))
1299 		goto err_uapi;
1300 
1301 	uverbs_dev->dev.devt = base;
1302 	dev_set_name(&uverbs_dev->dev, "uverbs%d", uverbs_dev->devnum);
1303 
1304 	cdev_init(&uverbs_dev->cdev,
1305 		  device->ops.mmap ? &uverbs_mmap_fops : &uverbs_fops);
1306 	uverbs_dev->cdev.owner = THIS_MODULE;
1307 
1308 	ret = cdev_device_add(&uverbs_dev->cdev, &uverbs_dev->dev);
1309 	if (ret)
1310 		goto err_uapi;
1311 
1312 	ib_set_client_data(device, &uverbs_client, uverbs_dev);
1313 	return;
1314 
1315 err_uapi:
1316 	ida_free(&uverbs_ida, devnum);
1317 err:
1318 	if (atomic_dec_and_test(&uverbs_dev->refcount))
1319 		ib_uverbs_comp_dev(uverbs_dev);
1320 	wait_for_completion(&uverbs_dev->comp);
1321 	put_device(&uverbs_dev->dev);
1322 	return;
1323 }
1324 
1325 static void ib_uverbs_free_hw_resources(struct ib_uverbs_device *uverbs_dev,
1326 					struct ib_device *ib_dev)
1327 {
1328 	struct ib_uverbs_file *file;
1329 	struct ib_uverbs_async_event_file *event_file;
1330 	struct ib_event event;
1331 
1332 	/* Pending running commands to terminate */
1333 	uverbs_disassociate_api_pre(uverbs_dev);
1334 	event.event = IB_EVENT_DEVICE_FATAL;
1335 	event.element.port_num = 0;
1336 	event.device = ib_dev;
1337 
1338 	mutex_lock(&uverbs_dev->lists_mutex);
1339 	while (!list_empty(&uverbs_dev->uverbs_file_list)) {
1340 		file = list_first_entry(&uverbs_dev->uverbs_file_list,
1341 					struct ib_uverbs_file, list);
1342 		list_del_init(&file->list);
1343 		kref_get(&file->ref);
1344 
1345 		/* We must release the mutex before going ahead and calling
1346 		 * uverbs_cleanup_ufile, as it might end up indirectly calling
1347 		 * uverbs_close, for example due to freeing the resources (e.g
1348 		 * mmput).
1349 		 */
1350 		mutex_unlock(&uverbs_dev->lists_mutex);
1351 
1352 		ib_uverbs_event_handler(&file->event_handler, &event);
1353 		uverbs_destroy_ufile_hw(file, RDMA_REMOVE_DRIVER_REMOVE);
1354 		kref_put(&file->ref, ib_uverbs_release_file);
1355 
1356 		mutex_lock(&uverbs_dev->lists_mutex);
1357 	}
1358 
1359 	while (!list_empty(&uverbs_dev->uverbs_events_file_list)) {
1360 		event_file = list_first_entry(&uverbs_dev->
1361 					      uverbs_events_file_list,
1362 					      struct ib_uverbs_async_event_file,
1363 					      list);
1364 		spin_lock_irq(&event_file->ev_queue.lock);
1365 		event_file->ev_queue.is_closed = 1;
1366 		spin_unlock_irq(&event_file->ev_queue.lock);
1367 
1368 		list_del(&event_file->list);
1369 		ib_unregister_event_handler(
1370 			&event_file->uverbs_file->event_handler);
1371 		event_file->uverbs_file->event_handler.device =
1372 			NULL;
1373 
1374 		wake_up_interruptible(&event_file->ev_queue.poll_wait);
1375 		kill_fasync(&event_file->ev_queue.async_queue, SIGIO, POLL_IN);
1376 	}
1377 	mutex_unlock(&uverbs_dev->lists_mutex);
1378 
1379 	uverbs_disassociate_api(uverbs_dev->uapi);
1380 }
1381 
1382 static void ib_uverbs_remove_one(struct ib_device *device, void *client_data)
1383 {
1384 	struct ib_uverbs_device *uverbs_dev = client_data;
1385 	int wait_clients = 1;
1386 
1387 	if (!uverbs_dev)
1388 		return;
1389 
1390 	cdev_device_del(&uverbs_dev->cdev, &uverbs_dev->dev);
1391 	ida_free(&uverbs_ida, uverbs_dev->devnum);
1392 
1393 	if (device->ops.disassociate_ucontext) {
1394 		/* We disassociate HW resources and immediately return.
1395 		 * Userspace will see a EIO errno for all future access.
1396 		 * Upon returning, ib_device may be freed internally and is not
1397 		 * valid any more.
1398 		 * uverbs_device is still available until all clients close
1399 		 * their files, then the uverbs device ref count will be zero
1400 		 * and its resources will be freed.
1401 		 * Note: At this point no more files can be opened since the
1402 		 * cdev was deleted, however active clients can still issue
1403 		 * commands and close their open files.
1404 		 */
1405 		ib_uverbs_free_hw_resources(uverbs_dev, device);
1406 		wait_clients = 0;
1407 	}
1408 
1409 	if (atomic_dec_and_test(&uverbs_dev->refcount))
1410 		ib_uverbs_comp_dev(uverbs_dev);
1411 	if (wait_clients)
1412 		wait_for_completion(&uverbs_dev->comp);
1413 
1414 	put_device(&uverbs_dev->dev);
1415 }
1416 
1417 static char *uverbs_devnode(struct device *dev, umode_t *mode)
1418 {
1419 	if (mode)
1420 		*mode = 0666;
1421 	return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev));
1422 }
1423 
1424 static int __init ib_uverbs_init(void)
1425 {
1426 	int ret;
1427 
1428 	ret = register_chrdev_region(IB_UVERBS_BASE_DEV,
1429 				     IB_UVERBS_NUM_FIXED_MINOR,
1430 				     "infiniband_verbs");
1431 	if (ret) {
1432 		pr_err("user_verbs: couldn't register device number\n");
1433 		goto out;
1434 	}
1435 
1436 	ret = alloc_chrdev_region(&dynamic_uverbs_dev, 0,
1437 				  IB_UVERBS_NUM_DYNAMIC_MINOR,
1438 				  "infiniband_verbs");
1439 	if (ret) {
1440 		pr_err("couldn't register dynamic device number\n");
1441 		goto out_alloc;
1442 	}
1443 
1444 	uverbs_class = class_create(THIS_MODULE, "infiniband_verbs");
1445 	if (IS_ERR(uverbs_class)) {
1446 		ret = PTR_ERR(uverbs_class);
1447 		pr_err("user_verbs: couldn't create class infiniband_verbs\n");
1448 		goto out_chrdev;
1449 	}
1450 
1451 	uverbs_class->devnode = uverbs_devnode;
1452 
1453 	ret = class_create_file(uverbs_class, &class_attr_abi_version.attr);
1454 	if (ret) {
1455 		pr_err("user_verbs: couldn't create abi_version attribute\n");
1456 		goto out_class;
1457 	}
1458 
1459 	ret = ib_register_client(&uverbs_client);
1460 	if (ret) {
1461 		pr_err("user_verbs: couldn't register client\n");
1462 		goto out_class;
1463 	}
1464 
1465 	return 0;
1466 
1467 out_class:
1468 	class_destroy(uverbs_class);
1469 
1470 out_chrdev:
1471 	unregister_chrdev_region(dynamic_uverbs_dev,
1472 				 IB_UVERBS_NUM_DYNAMIC_MINOR);
1473 
1474 out_alloc:
1475 	unregister_chrdev_region(IB_UVERBS_BASE_DEV,
1476 				 IB_UVERBS_NUM_FIXED_MINOR);
1477 
1478 out:
1479 	return ret;
1480 }
1481 
1482 static void __exit ib_uverbs_cleanup(void)
1483 {
1484 	ib_unregister_client(&uverbs_client);
1485 	class_destroy(uverbs_class);
1486 	unregister_chrdev_region(IB_UVERBS_BASE_DEV,
1487 				 IB_UVERBS_NUM_FIXED_MINOR);
1488 	unregister_chrdev_region(dynamic_uverbs_dev,
1489 				 IB_UVERBS_NUM_DYNAMIC_MINOR);
1490 }
1491 
1492 module_init(ib_uverbs_init);
1493 module_exit(ib_uverbs_cleanup);
1494