xref: /linux/drivers/infiniband/core/uverbs_main.c (revision 3ea5eb68b9d624935108b5e696859304edfac202)
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 
76 static DEFINE_IDA(uverbs_ida);
77 static int ib_uverbs_add_one(struct ib_device *device);
78 static void ib_uverbs_remove_one(struct ib_device *device, void *client_data);
79 
80 static char *uverbs_devnode(const struct device *dev, umode_t *mode)
81 {
82 	if (mode)
83 		*mode = 0666;
84 	return kasprintf(GFP_KERNEL, "infiniband/%s", dev_name(dev));
85 }
86 
87 static const struct class uverbs_class = {
88 	.name = "infiniband_verbs",
89 	.devnode = uverbs_devnode,
90 };
91 
92 /*
93  * Must be called with the ufile->device->disassociate_srcu held, and the lock
94  * must be held until use of the ucontext is finished.
95  */
96 struct ib_ucontext *ib_uverbs_get_ucontext_file(struct ib_uverbs_file *ufile)
97 {
98 	/*
99 	 * We do not hold the hw_destroy_rwsem lock for this flow, instead
100 	 * srcu is used. It does not matter if someone races this with
101 	 * get_context, we get NULL or valid ucontext.
102 	 */
103 	struct ib_ucontext *ucontext = smp_load_acquire(&ufile->ucontext);
104 
105 	if (!srcu_dereference(ufile->device->ib_dev,
106 			      &ufile->device->disassociate_srcu))
107 		return ERR_PTR(-EIO);
108 
109 	if (!ucontext)
110 		return ERR_PTR(-EINVAL);
111 
112 	return ucontext;
113 }
114 EXPORT_SYMBOL(ib_uverbs_get_ucontext_file);
115 
116 int uverbs_dealloc_mw(struct ib_mw *mw)
117 {
118 	struct ib_pd *pd = mw->pd;
119 	int ret;
120 
121 	ret = mw->device->ops.dealloc_mw(mw);
122 	if (ret)
123 		return ret;
124 
125 	atomic_dec(&pd->usecnt);
126 	kfree(mw);
127 	return ret;
128 }
129 
130 static void ib_uverbs_release_dev(struct device *device)
131 {
132 	struct ib_uverbs_device *dev =
133 			container_of(device, struct ib_uverbs_device, dev);
134 
135 	uverbs_destroy_api(dev->uapi);
136 	cleanup_srcu_struct(&dev->disassociate_srcu);
137 	mutex_destroy(&dev->lists_mutex);
138 	mutex_destroy(&dev->xrcd_tree_mutex);
139 	kfree(dev);
140 }
141 
142 void ib_uverbs_release_ucq(struct ib_uverbs_completion_event_file *ev_file,
143 			   struct ib_ucq_object *uobj)
144 {
145 	struct ib_uverbs_event *evt, *tmp;
146 
147 	if (ev_file) {
148 		spin_lock_irq(&ev_file->ev_queue.lock);
149 		list_for_each_entry_safe(evt, tmp, &uobj->comp_list, obj_list) {
150 			list_del(&evt->list);
151 			kfree(evt);
152 		}
153 		spin_unlock_irq(&ev_file->ev_queue.lock);
154 
155 		uverbs_uobject_put(&ev_file->uobj);
156 	}
157 
158 	ib_uverbs_release_uevent(&uobj->uevent);
159 }
160 
161 void ib_uverbs_release_uevent(struct ib_uevent_object *uobj)
162 {
163 	struct ib_uverbs_async_event_file *async_file = uobj->event_file;
164 	struct ib_uverbs_event *evt, *tmp;
165 
166 	if (!async_file)
167 		return;
168 
169 	spin_lock_irq(&async_file->ev_queue.lock);
170 	list_for_each_entry_safe(evt, tmp, &uobj->event_list, obj_list) {
171 		list_del(&evt->list);
172 		kfree(evt);
173 	}
174 	spin_unlock_irq(&async_file->ev_queue.lock);
175 	uverbs_uobject_put(&async_file->uobj);
176 }
177 
178 void ib_uverbs_detach_umcast(struct ib_qp *qp,
179 			     struct ib_uqp_object *uobj)
180 {
181 	struct ib_uverbs_mcast_entry *mcast, *tmp;
182 
183 	list_for_each_entry_safe(mcast, tmp, &uobj->mcast_list, list) {
184 		ib_detach_mcast(qp, &mcast->gid, mcast->lid);
185 		list_del(&mcast->list);
186 		kfree(mcast);
187 	}
188 }
189 
190 static void ib_uverbs_comp_dev(struct ib_uverbs_device *dev)
191 {
192 	complete(&dev->comp);
193 }
194 
195 void ib_uverbs_release_file(struct kref *ref)
196 {
197 	struct ib_uverbs_file *file =
198 		container_of(ref, struct ib_uverbs_file, ref);
199 	struct ib_device *ib_dev;
200 	int srcu_key;
201 
202 	release_ufile_idr_uobject(file);
203 
204 	srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
205 	ib_dev = srcu_dereference(file->device->ib_dev,
206 				  &file->device->disassociate_srcu);
207 	if (ib_dev && !ib_dev->ops.disassociate_ucontext)
208 		module_put(ib_dev->ops.owner);
209 	srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
210 
211 	if (refcount_dec_and_test(&file->device->refcount))
212 		ib_uverbs_comp_dev(file->device);
213 
214 	if (file->default_async_file)
215 		uverbs_uobject_put(&file->default_async_file->uobj);
216 	put_device(&file->device->dev);
217 
218 	if (file->disassociate_page)
219 		__free_pages(file->disassociate_page, 0);
220 	mutex_destroy(&file->umap_lock);
221 	mutex_destroy(&file->ucontext_lock);
222 	kfree(file);
223 }
224 
225 static ssize_t ib_uverbs_event_read(struct ib_uverbs_event_queue *ev_queue,
226 				    struct file *filp, char __user *buf,
227 				    size_t count, loff_t *pos,
228 				    size_t eventsz)
229 {
230 	struct ib_uverbs_event *event;
231 	int ret = 0;
232 
233 	spin_lock_irq(&ev_queue->lock);
234 
235 	while (list_empty(&ev_queue->event_list)) {
236 		if (ev_queue->is_closed) {
237 			spin_unlock_irq(&ev_queue->lock);
238 			return -EIO;
239 		}
240 
241 		spin_unlock_irq(&ev_queue->lock);
242 		if (filp->f_flags & O_NONBLOCK)
243 			return -EAGAIN;
244 
245 		if (wait_event_interruptible(ev_queue->poll_wait,
246 					     (!list_empty(&ev_queue->event_list) ||
247 					      ev_queue->is_closed)))
248 			return -ERESTARTSYS;
249 
250 		spin_lock_irq(&ev_queue->lock);
251 	}
252 
253 	event = list_entry(ev_queue->event_list.next, struct ib_uverbs_event, list);
254 
255 	if (eventsz > count) {
256 		ret   = -EINVAL;
257 		event = NULL;
258 	} else {
259 		list_del(ev_queue->event_list.next);
260 		if (event->counter) {
261 			++(*event->counter);
262 			list_del(&event->obj_list);
263 		}
264 	}
265 
266 	spin_unlock_irq(&ev_queue->lock);
267 
268 	if (event) {
269 		if (copy_to_user(buf, event, eventsz))
270 			ret = -EFAULT;
271 		else
272 			ret = eventsz;
273 	}
274 
275 	kfree(event);
276 
277 	return ret;
278 }
279 
280 static ssize_t ib_uverbs_async_event_read(struct file *filp, char __user *buf,
281 					  size_t count, loff_t *pos)
282 {
283 	struct ib_uverbs_async_event_file *file = filp->private_data;
284 
285 	return ib_uverbs_event_read(&file->ev_queue, filp, 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, filp, buf, count,
296 				    pos,
297 				    sizeof(struct ib_uverbs_comp_event_desc));
298 }
299 
300 static __poll_t ib_uverbs_event_poll(struct ib_uverbs_event_queue *ev_queue,
301 					 struct file *filp,
302 					 struct poll_table_struct *wait)
303 {
304 	__poll_t pollflags = 0;
305 
306 	poll_wait(filp, &ev_queue->poll_wait, wait);
307 
308 	spin_lock_irq(&ev_queue->lock);
309 	if (!list_empty(&ev_queue->event_list))
310 		pollflags = EPOLLIN | EPOLLRDNORM;
311 	else if (ev_queue->is_closed)
312 		pollflags = EPOLLERR;
313 	spin_unlock_irq(&ev_queue->lock);
314 
315 	return pollflags;
316 }
317 
318 static __poll_t ib_uverbs_async_event_poll(struct file *filp,
319 					       struct poll_table_struct *wait)
320 {
321 	struct ib_uverbs_async_event_file *file = filp->private_data;
322 
323 	return ib_uverbs_event_poll(&file->ev_queue, filp, wait);
324 }
325 
326 static __poll_t ib_uverbs_comp_event_poll(struct file *filp,
327 					      struct poll_table_struct *wait)
328 {
329 	struct ib_uverbs_completion_event_file *comp_ev_file =
330 		filp->private_data;
331 
332 	return ib_uverbs_event_poll(&comp_ev_file->ev_queue, filp, wait);
333 }
334 
335 static int ib_uverbs_async_event_fasync(int fd, struct file *filp, int on)
336 {
337 	struct ib_uverbs_async_event_file *file = filp->private_data;
338 
339 	return fasync_helper(fd, filp, on, &file->ev_queue.async_queue);
340 }
341 
342 static int ib_uverbs_comp_event_fasync(int fd, struct file *filp, int on)
343 {
344 	struct ib_uverbs_completion_event_file *comp_ev_file =
345 		filp->private_data;
346 
347 	return fasync_helper(fd, filp, on, &comp_ev_file->ev_queue.async_queue);
348 }
349 
350 const struct file_operations uverbs_event_fops = {
351 	.owner	 = THIS_MODULE,
352 	.read	 = ib_uverbs_comp_event_read,
353 	.poll    = ib_uverbs_comp_event_poll,
354 	.release = uverbs_uobject_fd_release,
355 	.fasync  = ib_uverbs_comp_event_fasync,
356 };
357 
358 const struct file_operations uverbs_async_event_fops = {
359 	.owner	 = THIS_MODULE,
360 	.read	 = ib_uverbs_async_event_read,
361 	.poll    = ib_uverbs_async_event_poll,
362 	.release = uverbs_async_event_release,
363 	.fasync  = ib_uverbs_async_event_fasync,
364 };
365 
366 void ib_uverbs_comp_handler(struct ib_cq *cq, void *cq_context)
367 {
368 	struct ib_uverbs_event_queue   *ev_queue = cq_context;
369 	struct ib_ucq_object	       *uobj;
370 	struct ib_uverbs_event	       *entry;
371 	unsigned long			flags;
372 
373 	if (!ev_queue)
374 		return;
375 
376 	spin_lock_irqsave(&ev_queue->lock, flags);
377 	if (ev_queue->is_closed) {
378 		spin_unlock_irqrestore(&ev_queue->lock, flags);
379 		return;
380 	}
381 
382 	entry = kmalloc(sizeof(*entry), GFP_ATOMIC);
383 	if (!entry) {
384 		spin_unlock_irqrestore(&ev_queue->lock, flags);
385 		return;
386 	}
387 
388 	uobj = cq->uobject;
389 
390 	entry->desc.comp.cq_handle = cq->uobject->uevent.uobject.user_handle;
391 	entry->counter		   = &uobj->comp_events_reported;
392 
393 	list_add_tail(&entry->list, &ev_queue->event_list);
394 	list_add_tail(&entry->obj_list, &uobj->comp_list);
395 	spin_unlock_irqrestore(&ev_queue->lock, flags);
396 
397 	wake_up_interruptible(&ev_queue->poll_wait);
398 	kill_fasync(&ev_queue->async_queue, SIGIO, POLL_IN);
399 }
400 
401 void ib_uverbs_async_handler(struct ib_uverbs_async_event_file *async_file,
402 			     __u64 element, __u64 event,
403 			     struct list_head *obj_list, u32 *counter)
404 {
405 	struct ib_uverbs_event *entry;
406 	unsigned long flags;
407 
408 	if (!async_file)
409 		return;
410 
411 	spin_lock_irqsave(&async_file->ev_queue.lock, flags);
412 	if (async_file->ev_queue.is_closed) {
413 		spin_unlock_irqrestore(&async_file->ev_queue.lock, flags);
414 		return;
415 	}
416 
417 	entry = kmalloc(sizeof(*entry), GFP_ATOMIC);
418 	if (!entry) {
419 		spin_unlock_irqrestore(&async_file->ev_queue.lock, flags);
420 		return;
421 	}
422 
423 	entry->desc.async.element = element;
424 	entry->desc.async.event_type = event;
425 	entry->desc.async.reserved = 0;
426 	entry->counter = counter;
427 
428 	list_add_tail(&entry->list, &async_file->ev_queue.event_list);
429 	if (obj_list)
430 		list_add_tail(&entry->obj_list, obj_list);
431 	spin_unlock_irqrestore(&async_file->ev_queue.lock, flags);
432 
433 	wake_up_interruptible(&async_file->ev_queue.poll_wait);
434 	kill_fasync(&async_file->ev_queue.async_queue, SIGIO, POLL_IN);
435 }
436 
437 static void uverbs_uobj_event(struct ib_uevent_object *eobj,
438 			      struct ib_event *event)
439 {
440 	ib_uverbs_async_handler(eobj->event_file,
441 				eobj->uobject.user_handle, event->event,
442 				&eobj->event_list, &eobj->events_reported);
443 }
444 
445 void ib_uverbs_cq_event_handler(struct ib_event *event, void *context_ptr)
446 {
447 	uverbs_uobj_event(&event->element.cq->uobject->uevent, event);
448 }
449 
450 void ib_uverbs_qp_event_handler(struct ib_event *event, void *context_ptr)
451 {
452 	/* for XRC target qp's, check that qp is live */
453 	if (!event->element.qp->uobject)
454 		return;
455 
456 	uverbs_uobj_event(&event->element.qp->uobject->uevent, event);
457 }
458 
459 void ib_uverbs_wq_event_handler(struct ib_event *event, void *context_ptr)
460 {
461 	uverbs_uobj_event(&event->element.wq->uobject->uevent, event);
462 }
463 
464 void ib_uverbs_srq_event_handler(struct ib_event *event, void *context_ptr)
465 {
466 	uverbs_uobj_event(&event->element.srq->uobject->uevent, event);
467 }
468 
469 static void ib_uverbs_event_handler(struct ib_event_handler *handler,
470 				    struct ib_event *event)
471 {
472 	ib_uverbs_async_handler(
473 		container_of(handler, struct ib_uverbs_async_event_file,
474 			     event_handler),
475 		event->element.port_num, event->event, NULL, NULL);
476 }
477 
478 void ib_uverbs_init_event_queue(struct ib_uverbs_event_queue *ev_queue)
479 {
480 	spin_lock_init(&ev_queue->lock);
481 	INIT_LIST_HEAD(&ev_queue->event_list);
482 	init_waitqueue_head(&ev_queue->poll_wait);
483 	ev_queue->is_closed   = 0;
484 	ev_queue->async_queue = NULL;
485 }
486 
487 void ib_uverbs_init_async_event_file(
488 	struct ib_uverbs_async_event_file *async_file)
489 {
490 	struct ib_uverbs_file *uverbs_file = async_file->uobj.ufile;
491 	struct ib_device *ib_dev = async_file->uobj.context->device;
492 
493 	ib_uverbs_init_event_queue(&async_file->ev_queue);
494 
495 	/* The first async_event_file becomes the default one for the file. */
496 	mutex_lock(&uverbs_file->ucontext_lock);
497 	if (!uverbs_file->default_async_file) {
498 		/* Pairs with the put in ib_uverbs_release_file */
499 		uverbs_uobject_get(&async_file->uobj);
500 		smp_store_release(&uverbs_file->default_async_file, async_file);
501 	}
502 	mutex_unlock(&uverbs_file->ucontext_lock);
503 
504 	INIT_IB_EVENT_HANDLER(&async_file->event_handler, ib_dev,
505 			      ib_uverbs_event_handler);
506 	ib_register_event_handler(&async_file->event_handler);
507 }
508 
509 static ssize_t verify_hdr(struct ib_uverbs_cmd_hdr *hdr,
510 			  struct ib_uverbs_ex_cmd_hdr *ex_hdr, size_t count,
511 			  const struct uverbs_api_write_method *method_elm)
512 {
513 	if (method_elm->is_ex) {
514 		count -= sizeof(*hdr) + sizeof(*ex_hdr);
515 
516 		if ((hdr->in_words + ex_hdr->provider_in_words) * 8 != count)
517 			return -EINVAL;
518 
519 		if (hdr->in_words * 8 < method_elm->req_size)
520 			return -ENOSPC;
521 
522 		if (ex_hdr->cmd_hdr_reserved)
523 			return -EINVAL;
524 
525 		if (ex_hdr->response) {
526 			if (!hdr->out_words && !ex_hdr->provider_out_words)
527 				return -EINVAL;
528 
529 			if (hdr->out_words * 8 < method_elm->resp_size)
530 				return -ENOSPC;
531 
532 			if (!access_ok(u64_to_user_ptr(ex_hdr->response),
533 				       (hdr->out_words + ex_hdr->provider_out_words) * 8))
534 				return -EFAULT;
535 		} else {
536 			if (hdr->out_words || ex_hdr->provider_out_words)
537 				return -EINVAL;
538 		}
539 
540 		return 0;
541 	}
542 
543 	/* not extended command */
544 	if (hdr->in_words * 4 != count)
545 		return -EINVAL;
546 
547 	if (count < method_elm->req_size + sizeof(*hdr)) {
548 		/*
549 		 * rdma-core v18 and v19 have a bug where they send DESTROY_CQ
550 		 * with a 16 byte write instead of 24. Old kernels didn't
551 		 * check the size so they allowed this. Now that the size is
552 		 * checked provide a compatibility work around to not break
553 		 * those userspaces.
554 		 */
555 		if (hdr->command == IB_USER_VERBS_CMD_DESTROY_CQ &&
556 		    count == 16) {
557 			hdr->in_words = 6;
558 			return 0;
559 		}
560 		return -ENOSPC;
561 	}
562 	if (hdr->out_words * 4 < method_elm->resp_size)
563 		return -ENOSPC;
564 
565 	return 0;
566 }
567 
568 static ssize_t ib_uverbs_write(struct file *filp, const char __user *buf,
569 			     size_t count, loff_t *pos)
570 {
571 	struct ib_uverbs_file *file = filp->private_data;
572 	const struct uverbs_api_write_method *method_elm;
573 	struct uverbs_api *uapi = file->device->uapi;
574 	struct ib_uverbs_ex_cmd_hdr ex_hdr;
575 	struct ib_uverbs_cmd_hdr hdr;
576 	struct uverbs_attr_bundle bundle;
577 	int srcu_key;
578 	ssize_t ret;
579 
580 	if (!ib_safe_file_access(filp)) {
581 		pr_err_once("uverbs_write: process %d (%s) changed security contexts after opening file descriptor, this is not allowed.\n",
582 			    task_tgid_vnr(current), current->comm);
583 		return -EACCES;
584 	}
585 
586 	if (count < sizeof(hdr))
587 		return -EINVAL;
588 
589 	if (copy_from_user(&hdr, buf, sizeof(hdr)))
590 		return -EFAULT;
591 
592 	method_elm = uapi_get_method(uapi, hdr.command);
593 	if (IS_ERR(method_elm))
594 		return PTR_ERR(method_elm);
595 
596 	if (method_elm->is_ex) {
597 		if (count < (sizeof(hdr) + sizeof(ex_hdr)))
598 			return -EINVAL;
599 		if (copy_from_user(&ex_hdr, buf + sizeof(hdr), sizeof(ex_hdr)))
600 			return -EFAULT;
601 	}
602 
603 	ret = verify_hdr(&hdr, &ex_hdr, count, method_elm);
604 	if (ret)
605 		return ret;
606 
607 	srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
608 
609 	buf += sizeof(hdr);
610 
611 	memset(bundle.attr_present, 0, sizeof(bundle.attr_present));
612 	bundle.ufile = file;
613 	bundle.context = NULL; /* only valid if bundle has uobject */
614 	bundle.uobject = NULL;
615 	if (!method_elm->is_ex) {
616 		size_t in_len = hdr.in_words * 4 - sizeof(hdr);
617 		size_t out_len = hdr.out_words * 4;
618 		u64 response = 0;
619 
620 		if (method_elm->has_udata) {
621 			bundle.driver_udata.inlen =
622 				in_len - method_elm->req_size;
623 			in_len = method_elm->req_size;
624 			if (bundle.driver_udata.inlen)
625 				bundle.driver_udata.inbuf = buf + in_len;
626 			else
627 				bundle.driver_udata.inbuf = NULL;
628 		} else {
629 			memset(&bundle.driver_udata, 0,
630 			       sizeof(bundle.driver_udata));
631 		}
632 
633 		if (method_elm->has_resp) {
634 			/*
635 			 * The macros check that if has_resp is set
636 			 * then the command request structure starts
637 			 * with a '__aligned u64 response' member.
638 			 */
639 			ret = get_user(response, (const u64 __user *)buf);
640 			if (ret)
641 				goto out_unlock;
642 
643 			if (method_elm->has_udata) {
644 				bundle.driver_udata.outlen =
645 					out_len - method_elm->resp_size;
646 				out_len = method_elm->resp_size;
647 				if (bundle.driver_udata.outlen)
648 					bundle.driver_udata.outbuf =
649 						u64_to_user_ptr(response +
650 								out_len);
651 				else
652 					bundle.driver_udata.outbuf = NULL;
653 			}
654 		} else {
655 			bundle.driver_udata.outlen = 0;
656 			bundle.driver_udata.outbuf = NULL;
657 		}
658 
659 		ib_uverbs_init_udata_buf_or_null(
660 			&bundle.ucore, buf, u64_to_user_ptr(response),
661 			in_len, out_len);
662 	} else {
663 		buf += sizeof(ex_hdr);
664 
665 		ib_uverbs_init_udata_buf_or_null(&bundle.ucore, buf,
666 					u64_to_user_ptr(ex_hdr.response),
667 					hdr.in_words * 8, hdr.out_words * 8);
668 
669 		ib_uverbs_init_udata_buf_or_null(
670 			&bundle.driver_udata, buf + bundle.ucore.inlen,
671 			u64_to_user_ptr(ex_hdr.response) + bundle.ucore.outlen,
672 			ex_hdr.provider_in_words * 8,
673 			ex_hdr.provider_out_words * 8);
674 
675 	}
676 
677 	ret = method_elm->handler(&bundle);
678 	if (bundle.uobject)
679 		uverbs_finalize_object(bundle.uobject, UVERBS_ACCESS_NEW, true,
680 				       !ret, &bundle);
681 out_unlock:
682 	srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
683 	return (ret) ? : count;
684 }
685 
686 static const struct vm_operations_struct rdma_umap_ops;
687 
688 static int ib_uverbs_mmap(struct file *filp, struct vm_area_struct *vma)
689 {
690 	struct ib_uverbs_file *file = filp->private_data;
691 	struct ib_ucontext *ucontext;
692 	int ret = 0;
693 	int srcu_key;
694 
695 	srcu_key = srcu_read_lock(&file->device->disassociate_srcu);
696 	ucontext = ib_uverbs_get_ucontext_file(file);
697 	if (IS_ERR(ucontext)) {
698 		ret = PTR_ERR(ucontext);
699 		goto out;
700 	}
701 	vma->vm_ops = &rdma_umap_ops;
702 	ret = ucontext->device->ops.mmap(ucontext, vma);
703 out:
704 	srcu_read_unlock(&file->device->disassociate_srcu, srcu_key);
705 	return ret;
706 }
707 
708 /*
709  * The VMA has been dup'd, initialize the vm_private_data with a new tracking
710  * struct
711  */
712 static void rdma_umap_open(struct vm_area_struct *vma)
713 {
714 	struct ib_uverbs_file *ufile = vma->vm_file->private_data;
715 	struct rdma_umap_priv *opriv = vma->vm_private_data;
716 	struct rdma_umap_priv *priv;
717 
718 	if (!opriv)
719 		return;
720 
721 	/* We are racing with disassociation */
722 	if (!down_read_trylock(&ufile->hw_destroy_rwsem))
723 		goto out_zap;
724 	/*
725 	 * Disassociation already completed, the VMA should already be zapped.
726 	 */
727 	if (!ufile->ucontext)
728 		goto out_unlock;
729 
730 	priv = kzalloc(sizeof(*priv), GFP_KERNEL);
731 	if (!priv)
732 		goto out_unlock;
733 	rdma_umap_priv_init(priv, vma, opriv->entry);
734 
735 	up_read(&ufile->hw_destroy_rwsem);
736 	return;
737 
738 out_unlock:
739 	up_read(&ufile->hw_destroy_rwsem);
740 out_zap:
741 	/*
742 	 * We can't allow the VMA to be created with the actual IO pages, that
743 	 * would break our API contract, and it can't be stopped at this
744 	 * point, so zap it.
745 	 */
746 	vma->vm_private_data = NULL;
747 	zap_vma_ptes(vma, vma->vm_start, vma->vm_end - vma->vm_start);
748 }
749 
750 static void rdma_umap_close(struct vm_area_struct *vma)
751 {
752 	struct ib_uverbs_file *ufile = vma->vm_file->private_data;
753 	struct rdma_umap_priv *priv = vma->vm_private_data;
754 
755 	if (!priv)
756 		return;
757 
758 	/*
759 	 * The vma holds a reference on the struct file that created it, which
760 	 * in turn means that the ib_uverbs_file is guaranteed to exist at
761 	 * this point.
762 	 */
763 	mutex_lock(&ufile->umap_lock);
764 	if (priv->entry)
765 		rdma_user_mmap_entry_put(priv->entry);
766 
767 	list_del(&priv->list);
768 	mutex_unlock(&ufile->umap_lock);
769 	kfree(priv);
770 }
771 
772 /*
773  * Once the zap_vma_ptes has been called touches to the VMA will come here and
774  * we return a dummy writable zero page for all the pfns.
775  */
776 static vm_fault_t rdma_umap_fault(struct vm_fault *vmf)
777 {
778 	struct ib_uverbs_file *ufile = vmf->vma->vm_file->private_data;
779 	struct rdma_umap_priv *priv = vmf->vma->vm_private_data;
780 	vm_fault_t ret = 0;
781 
782 	if (!priv)
783 		return VM_FAULT_SIGBUS;
784 
785 	/* Read only pages can just use the system zero page. */
786 	if (!(vmf->vma->vm_flags & (VM_WRITE | VM_MAYWRITE))) {
787 		vmf->page = ZERO_PAGE(vmf->address);
788 		get_page(vmf->page);
789 		return 0;
790 	}
791 
792 	mutex_lock(&ufile->umap_lock);
793 	if (!ufile->disassociate_page)
794 		ufile->disassociate_page =
795 			alloc_pages(vmf->gfp_mask | __GFP_ZERO, 0);
796 
797 	if (ufile->disassociate_page) {
798 		/*
799 		 * This VMA is forced to always be shared so this doesn't have
800 		 * to worry about COW.
801 		 */
802 		vmf->page = ufile->disassociate_page;
803 		get_page(vmf->page);
804 	} else {
805 		ret = VM_FAULT_SIGBUS;
806 	}
807 	mutex_unlock(&ufile->umap_lock);
808 
809 	return ret;
810 }
811 
812 static const struct vm_operations_struct rdma_umap_ops = {
813 	.open = rdma_umap_open,
814 	.close = rdma_umap_close,
815 	.fault = rdma_umap_fault,
816 };
817 
818 void uverbs_user_mmap_disassociate(struct ib_uverbs_file *ufile)
819 {
820 	struct rdma_umap_priv *priv, *next_priv;
821 
822 	lockdep_assert_held(&ufile->hw_destroy_rwsem);
823 
824 	while (1) {
825 		struct mm_struct *mm = NULL;
826 
827 		/* Get an arbitrary mm pointer that hasn't been cleaned yet */
828 		mutex_lock(&ufile->umap_lock);
829 		while (!list_empty(&ufile->umaps)) {
830 			int ret;
831 
832 			priv = list_first_entry(&ufile->umaps,
833 						struct rdma_umap_priv, list);
834 			mm = priv->vma->vm_mm;
835 			ret = mmget_not_zero(mm);
836 			if (!ret) {
837 				list_del_init(&priv->list);
838 				if (priv->entry) {
839 					rdma_user_mmap_entry_put(priv->entry);
840 					priv->entry = NULL;
841 				}
842 				mm = NULL;
843 				continue;
844 			}
845 			break;
846 		}
847 		mutex_unlock(&ufile->umap_lock);
848 		if (!mm)
849 			return;
850 
851 		/*
852 		 * The umap_lock is nested under mmap_lock since it used within
853 		 * the vma_ops callbacks, so we have to clean the list one mm
854 		 * at a time to get the lock ordering right. Typically there
855 		 * will only be one mm, so no big deal.
856 		 */
857 		mmap_read_lock(mm);
858 		mutex_lock(&ufile->umap_lock);
859 		list_for_each_entry_safe (priv, next_priv, &ufile->umaps,
860 					  list) {
861 			struct vm_area_struct *vma = priv->vma;
862 
863 			if (vma->vm_mm != mm)
864 				continue;
865 			list_del_init(&priv->list);
866 
867 			zap_vma_ptes(vma, vma->vm_start,
868 				     vma->vm_end - vma->vm_start);
869 
870 			if (priv->entry) {
871 				rdma_user_mmap_entry_put(priv->entry);
872 				priv->entry = NULL;
873 			}
874 		}
875 		mutex_unlock(&ufile->umap_lock);
876 		mmap_read_unlock(mm);
877 		mmput(mm);
878 	}
879 }
880 
881 /*
882  * ib_uverbs_open() does not need the BKL:
883  *
884  *  - the ib_uverbs_device structures are properly reference counted and
885  *    everything else is purely local to the file being created, so
886  *    races against other open calls are not a problem;
887  *  - there is no ioctl method to race against;
888  *  - the open method will either immediately run -ENXIO, or all
889  *    required initialization will be done.
890  */
891 static int ib_uverbs_open(struct inode *inode, struct file *filp)
892 {
893 	struct ib_uverbs_device *dev;
894 	struct ib_uverbs_file *file;
895 	struct ib_device *ib_dev;
896 	int ret;
897 	int module_dependent;
898 	int srcu_key;
899 
900 	dev = container_of(inode->i_cdev, struct ib_uverbs_device, cdev);
901 	if (!refcount_inc_not_zero(&dev->refcount))
902 		return -ENXIO;
903 
904 	get_device(&dev->dev);
905 	srcu_key = srcu_read_lock(&dev->disassociate_srcu);
906 	mutex_lock(&dev->lists_mutex);
907 	ib_dev = srcu_dereference(dev->ib_dev,
908 				  &dev->disassociate_srcu);
909 	if (!ib_dev) {
910 		ret = -EIO;
911 		goto err;
912 	}
913 
914 	if (!rdma_dev_access_netns(ib_dev, current->nsproxy->net_ns)) {
915 		ret = -EPERM;
916 		goto err;
917 	}
918 
919 	/* In case IB device supports disassociate ucontext, there is no hard
920 	 * dependency between uverbs device and its low level device.
921 	 */
922 	module_dependent = !(ib_dev->ops.disassociate_ucontext);
923 
924 	if (module_dependent) {
925 		if (!try_module_get(ib_dev->ops.owner)) {
926 			ret = -ENODEV;
927 			goto err;
928 		}
929 	}
930 
931 	file = kzalloc(sizeof(*file), GFP_KERNEL);
932 	if (!file) {
933 		ret = -ENOMEM;
934 		if (module_dependent)
935 			goto err_module;
936 
937 		goto err;
938 	}
939 
940 	file->device	 = dev;
941 	kref_init(&file->ref);
942 	mutex_init(&file->ucontext_lock);
943 
944 	spin_lock_init(&file->uobjects_lock);
945 	INIT_LIST_HEAD(&file->uobjects);
946 	init_rwsem(&file->hw_destroy_rwsem);
947 	mutex_init(&file->umap_lock);
948 	INIT_LIST_HEAD(&file->umaps);
949 
950 	filp->private_data = file;
951 	list_add_tail(&file->list, &dev->uverbs_file_list);
952 	mutex_unlock(&dev->lists_mutex);
953 	srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
954 
955 	setup_ufile_idr_uobject(file);
956 
957 	return stream_open(inode, filp);
958 
959 err_module:
960 	module_put(ib_dev->ops.owner);
961 
962 err:
963 	mutex_unlock(&dev->lists_mutex);
964 	srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
965 	if (refcount_dec_and_test(&dev->refcount))
966 		ib_uverbs_comp_dev(dev);
967 
968 	put_device(&dev->dev);
969 	return ret;
970 }
971 
972 static int ib_uverbs_close(struct inode *inode, struct file *filp)
973 {
974 	struct ib_uverbs_file *file = filp->private_data;
975 
976 	uverbs_destroy_ufile_hw(file, RDMA_REMOVE_CLOSE);
977 
978 	mutex_lock(&file->device->lists_mutex);
979 	list_del_init(&file->list);
980 	mutex_unlock(&file->device->lists_mutex);
981 
982 	kref_put(&file->ref, ib_uverbs_release_file);
983 
984 	return 0;
985 }
986 
987 static const struct file_operations uverbs_fops = {
988 	.owner	 = THIS_MODULE,
989 	.write	 = ib_uverbs_write,
990 	.open	 = ib_uverbs_open,
991 	.release = ib_uverbs_close,
992 	.unlocked_ioctl = ib_uverbs_ioctl,
993 	.compat_ioctl = compat_ptr_ioctl,
994 };
995 
996 static const struct file_operations uverbs_mmap_fops = {
997 	.owner	 = THIS_MODULE,
998 	.write	 = ib_uverbs_write,
999 	.mmap    = ib_uverbs_mmap,
1000 	.open	 = ib_uverbs_open,
1001 	.release = ib_uverbs_close,
1002 	.unlocked_ioctl = ib_uverbs_ioctl,
1003 	.compat_ioctl = compat_ptr_ioctl,
1004 };
1005 
1006 static int ib_uverbs_get_nl_info(struct ib_device *ibdev, void *client_data,
1007 				 struct ib_client_nl_info *res)
1008 {
1009 	struct ib_uverbs_device *uverbs_dev = client_data;
1010 	int ret;
1011 
1012 	if (res->port != -1)
1013 		return -EINVAL;
1014 
1015 	res->abi = ibdev->ops.uverbs_abi_ver;
1016 	res->cdev = &uverbs_dev->dev;
1017 
1018 	/*
1019 	 * To support DRIVER_ID binding in userspace some of the driver need
1020 	 * upgrading to expose their PCI dependent revision information
1021 	 * through get_context instead of relying on modalias matching. When
1022 	 * the drivers are fixed they can drop this flag.
1023 	 */
1024 	if (!ibdev->ops.uverbs_no_driver_id_binding) {
1025 		ret = nla_put_u32(res->nl_msg, RDMA_NLDEV_ATTR_UVERBS_DRIVER_ID,
1026 				  ibdev->ops.driver_id);
1027 		if (ret)
1028 			return ret;
1029 	}
1030 	return 0;
1031 }
1032 
1033 static struct ib_client uverbs_client = {
1034 	.name   = "uverbs",
1035 	.no_kverbs_req = true,
1036 	.add    = ib_uverbs_add_one,
1037 	.remove = ib_uverbs_remove_one,
1038 	.get_nl_info = ib_uverbs_get_nl_info,
1039 };
1040 MODULE_ALIAS_RDMA_CLIENT("uverbs");
1041 
1042 static ssize_t ibdev_show(struct device *device, struct device_attribute *attr,
1043 			  char *buf)
1044 {
1045 	struct ib_uverbs_device *dev =
1046 			container_of(device, struct ib_uverbs_device, dev);
1047 	int ret = -ENODEV;
1048 	int srcu_key;
1049 	struct ib_device *ib_dev;
1050 
1051 	srcu_key = srcu_read_lock(&dev->disassociate_srcu);
1052 	ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu);
1053 	if (ib_dev)
1054 		ret = sysfs_emit(buf, "%s\n", dev_name(&ib_dev->dev));
1055 	srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1056 
1057 	return ret;
1058 }
1059 static DEVICE_ATTR_RO(ibdev);
1060 
1061 static ssize_t abi_version_show(struct device *device,
1062 				struct device_attribute *attr, char *buf)
1063 {
1064 	struct ib_uverbs_device *dev =
1065 			container_of(device, struct ib_uverbs_device, dev);
1066 	int ret = -ENODEV;
1067 	int srcu_key;
1068 	struct ib_device *ib_dev;
1069 
1070 	srcu_key = srcu_read_lock(&dev->disassociate_srcu);
1071 	ib_dev = srcu_dereference(dev->ib_dev, &dev->disassociate_srcu);
1072 	if (ib_dev)
1073 		ret = sysfs_emit(buf, "%u\n", ib_dev->ops.uverbs_abi_ver);
1074 	srcu_read_unlock(&dev->disassociate_srcu, srcu_key);
1075 
1076 	return ret;
1077 }
1078 static DEVICE_ATTR_RO(abi_version);
1079 
1080 static struct attribute *ib_dev_attrs[] = {
1081 	&dev_attr_abi_version.attr,
1082 	&dev_attr_ibdev.attr,
1083 	NULL,
1084 };
1085 
1086 static const struct attribute_group dev_attr_group = {
1087 	.attrs = ib_dev_attrs,
1088 };
1089 
1090 static CLASS_ATTR_STRING(abi_version, S_IRUGO,
1091 			 __stringify(IB_USER_VERBS_ABI_VERSION));
1092 
1093 static int ib_uverbs_create_uapi(struct ib_device *device,
1094 				 struct ib_uverbs_device *uverbs_dev)
1095 {
1096 	struct uverbs_api *uapi;
1097 
1098 	uapi = uverbs_alloc_api(device);
1099 	if (IS_ERR(uapi))
1100 		return PTR_ERR(uapi);
1101 
1102 	uverbs_dev->uapi = uapi;
1103 	return 0;
1104 }
1105 
1106 static int ib_uverbs_add_one(struct ib_device *device)
1107 {
1108 	int devnum;
1109 	dev_t base;
1110 	struct ib_uverbs_device *uverbs_dev;
1111 	int ret;
1112 
1113 	if (!device->ops.alloc_ucontext ||
1114 	    device->type == RDMA_DEVICE_TYPE_SMI)
1115 		return -EOPNOTSUPP;
1116 
1117 	uverbs_dev = kzalloc(sizeof(*uverbs_dev), GFP_KERNEL);
1118 	if (!uverbs_dev)
1119 		return -ENOMEM;
1120 
1121 	ret = init_srcu_struct(&uverbs_dev->disassociate_srcu);
1122 	if (ret) {
1123 		kfree(uverbs_dev);
1124 		return -ENOMEM;
1125 	}
1126 
1127 	device_initialize(&uverbs_dev->dev);
1128 	uverbs_dev->dev.class = &uverbs_class;
1129 	uverbs_dev->dev.parent = device->dev.parent;
1130 	uverbs_dev->dev.release = ib_uverbs_release_dev;
1131 	uverbs_dev->groups[0] = &dev_attr_group;
1132 	uverbs_dev->dev.groups = uverbs_dev->groups;
1133 	refcount_set(&uverbs_dev->refcount, 1);
1134 	init_completion(&uverbs_dev->comp);
1135 	uverbs_dev->xrcd_tree = RB_ROOT;
1136 	mutex_init(&uverbs_dev->xrcd_tree_mutex);
1137 	mutex_init(&uverbs_dev->lists_mutex);
1138 	INIT_LIST_HEAD(&uverbs_dev->uverbs_file_list);
1139 	rcu_assign_pointer(uverbs_dev->ib_dev, device);
1140 	uverbs_dev->num_comp_vectors = device->num_comp_vectors;
1141 
1142 	devnum = ida_alloc_max(&uverbs_ida, IB_UVERBS_MAX_DEVICES - 1,
1143 			       GFP_KERNEL);
1144 	if (devnum < 0) {
1145 		ret = -ENOMEM;
1146 		goto err;
1147 	}
1148 	uverbs_dev->devnum = devnum;
1149 	if (devnum >= IB_UVERBS_NUM_FIXED_MINOR)
1150 		base = dynamic_uverbs_dev + devnum - IB_UVERBS_NUM_FIXED_MINOR;
1151 	else
1152 		base = IB_UVERBS_BASE_DEV + devnum;
1153 
1154 	ret = ib_uverbs_create_uapi(device, uverbs_dev);
1155 	if (ret)
1156 		goto err_uapi;
1157 
1158 	uverbs_dev->dev.devt = base;
1159 	dev_set_name(&uverbs_dev->dev, "uverbs%d", uverbs_dev->devnum);
1160 
1161 	cdev_init(&uverbs_dev->cdev,
1162 		  device->ops.mmap ? &uverbs_mmap_fops : &uverbs_fops);
1163 	uverbs_dev->cdev.owner = THIS_MODULE;
1164 
1165 	ret = cdev_device_add(&uverbs_dev->cdev, &uverbs_dev->dev);
1166 	if (ret)
1167 		goto err_uapi;
1168 
1169 	ib_set_client_data(device, &uverbs_client, uverbs_dev);
1170 	return 0;
1171 
1172 err_uapi:
1173 	ida_free(&uverbs_ida, devnum);
1174 err:
1175 	if (refcount_dec_and_test(&uverbs_dev->refcount))
1176 		ib_uverbs_comp_dev(uverbs_dev);
1177 	wait_for_completion(&uverbs_dev->comp);
1178 	put_device(&uverbs_dev->dev);
1179 	return ret;
1180 }
1181 
1182 static void ib_uverbs_free_hw_resources(struct ib_uverbs_device *uverbs_dev,
1183 					struct ib_device *ib_dev)
1184 {
1185 	struct ib_uverbs_file *file;
1186 
1187 	/* Pending running commands to terminate */
1188 	uverbs_disassociate_api_pre(uverbs_dev);
1189 
1190 	mutex_lock(&uverbs_dev->lists_mutex);
1191 	while (!list_empty(&uverbs_dev->uverbs_file_list)) {
1192 		file = list_first_entry(&uverbs_dev->uverbs_file_list,
1193 					struct ib_uverbs_file, list);
1194 		list_del_init(&file->list);
1195 		kref_get(&file->ref);
1196 
1197 		/* We must release the mutex before going ahead and calling
1198 		 * uverbs_cleanup_ufile, as it might end up indirectly calling
1199 		 * uverbs_close, for example due to freeing the resources (e.g
1200 		 * mmput).
1201 		 */
1202 		mutex_unlock(&uverbs_dev->lists_mutex);
1203 
1204 		uverbs_destroy_ufile_hw(file, RDMA_REMOVE_DRIVER_REMOVE);
1205 		kref_put(&file->ref, ib_uverbs_release_file);
1206 
1207 		mutex_lock(&uverbs_dev->lists_mutex);
1208 	}
1209 	mutex_unlock(&uverbs_dev->lists_mutex);
1210 
1211 	uverbs_disassociate_api(uverbs_dev->uapi);
1212 }
1213 
1214 static void ib_uverbs_remove_one(struct ib_device *device, void *client_data)
1215 {
1216 	struct ib_uverbs_device *uverbs_dev = client_data;
1217 	int wait_clients = 1;
1218 
1219 	cdev_device_del(&uverbs_dev->cdev, &uverbs_dev->dev);
1220 	ida_free(&uverbs_ida, uverbs_dev->devnum);
1221 
1222 	if (device->ops.disassociate_ucontext) {
1223 		/* We disassociate HW resources and immediately return.
1224 		 * Userspace will see a EIO errno for all future access.
1225 		 * Upon returning, ib_device may be freed internally and is not
1226 		 * valid any more.
1227 		 * uverbs_device is still available until all clients close
1228 		 * their files, then the uverbs device ref count will be zero
1229 		 * and its resources will be freed.
1230 		 * Note: At this point no more files can be opened since the
1231 		 * cdev was deleted, however active clients can still issue
1232 		 * commands and close their open files.
1233 		 */
1234 		ib_uverbs_free_hw_resources(uverbs_dev, device);
1235 		wait_clients = 0;
1236 	}
1237 
1238 	if (refcount_dec_and_test(&uverbs_dev->refcount))
1239 		ib_uverbs_comp_dev(uverbs_dev);
1240 	if (wait_clients)
1241 		wait_for_completion(&uverbs_dev->comp);
1242 
1243 	put_device(&uverbs_dev->dev);
1244 }
1245 
1246 static int __init ib_uverbs_init(void)
1247 {
1248 	int ret;
1249 
1250 	ret = register_chrdev_region(IB_UVERBS_BASE_DEV,
1251 				     IB_UVERBS_NUM_FIXED_MINOR,
1252 				     "infiniband_verbs");
1253 	if (ret) {
1254 		pr_err("user_verbs: couldn't register device number\n");
1255 		goto out;
1256 	}
1257 
1258 	ret = alloc_chrdev_region(&dynamic_uverbs_dev, 0,
1259 				  IB_UVERBS_NUM_DYNAMIC_MINOR,
1260 				  "infiniband_verbs");
1261 	if (ret) {
1262 		pr_err("couldn't register dynamic device number\n");
1263 		goto out_alloc;
1264 	}
1265 
1266 	ret = class_register(&uverbs_class);
1267 	if (ret) {
1268 		pr_err("user_verbs: couldn't create class infiniband_verbs\n");
1269 		goto out_chrdev;
1270 	}
1271 
1272 	ret = class_create_file(&uverbs_class, &class_attr_abi_version.attr);
1273 	if (ret) {
1274 		pr_err("user_verbs: couldn't create abi_version attribute\n");
1275 		goto out_class;
1276 	}
1277 
1278 	ret = ib_register_client(&uverbs_client);
1279 	if (ret) {
1280 		pr_err("user_verbs: couldn't register client\n");
1281 		goto out_class;
1282 	}
1283 
1284 	return 0;
1285 
1286 out_class:
1287 	class_unregister(&uverbs_class);
1288 
1289 out_chrdev:
1290 	unregister_chrdev_region(dynamic_uverbs_dev,
1291 				 IB_UVERBS_NUM_DYNAMIC_MINOR);
1292 
1293 out_alloc:
1294 	unregister_chrdev_region(IB_UVERBS_BASE_DEV,
1295 				 IB_UVERBS_NUM_FIXED_MINOR);
1296 
1297 out:
1298 	return ret;
1299 }
1300 
1301 static void __exit ib_uverbs_cleanup(void)
1302 {
1303 	ib_unregister_client(&uverbs_client);
1304 	class_unregister(&uverbs_class);
1305 	unregister_chrdev_region(IB_UVERBS_BASE_DEV,
1306 				 IB_UVERBS_NUM_FIXED_MINOR);
1307 	unregister_chrdev_region(dynamic_uverbs_dev,
1308 				 IB_UVERBS_NUM_DYNAMIC_MINOR);
1309 	mmu_notifier_synchronize();
1310 }
1311 
1312 module_init(ib_uverbs_init);
1313 module_exit(ib_uverbs_cleanup);
1314