xref: /openbmc/linux/kernel/watch_queue.c (revision 3ddc8b84)
1 // SPDX-License-Identifier: GPL-2.0
2 /* Watch queue and general notification mechanism, built on pipes
3  *
4  * Copyright (C) 2020 Red Hat, Inc. All Rights Reserved.
5  * Written by David Howells (dhowells@redhat.com)
6  *
7  * See Documentation/core-api/watch_queue.rst
8  */
9 
10 #define pr_fmt(fmt) "watchq: " fmt
11 #include <linux/module.h>
12 #include <linux/init.h>
13 #include <linux/sched.h>
14 #include <linux/slab.h>
15 #include <linux/printk.h>
16 #include <linux/miscdevice.h>
17 #include <linux/fs.h>
18 #include <linux/mm.h>
19 #include <linux/pagemap.h>
20 #include <linux/poll.h>
21 #include <linux/uaccess.h>
22 #include <linux/vmalloc.h>
23 #include <linux/file.h>
24 #include <linux/security.h>
25 #include <linux/cred.h>
26 #include <linux/sched/signal.h>
27 #include <linux/watch_queue.h>
28 #include <linux/pipe_fs_i.h>
29 
30 MODULE_DESCRIPTION("Watch queue");
31 MODULE_AUTHOR("Red Hat, Inc.");
32 
33 #define WATCH_QUEUE_NOTE_SIZE 128
34 #define WATCH_QUEUE_NOTES_PER_PAGE (PAGE_SIZE / WATCH_QUEUE_NOTE_SIZE)
35 
36 /*
37  * This must be called under the RCU read-lock, which makes
38  * sure that the wqueue still exists. It can then take the lock,
39  * and check that the wqueue hasn't been destroyed, which in
40  * turn makes sure that the notification pipe still exists.
41  */
42 static inline bool lock_wqueue(struct watch_queue *wqueue)
43 {
44 	spin_lock_bh(&wqueue->lock);
45 	if (unlikely(!wqueue->pipe)) {
46 		spin_unlock_bh(&wqueue->lock);
47 		return false;
48 	}
49 	return true;
50 }
51 
52 static inline void unlock_wqueue(struct watch_queue *wqueue)
53 {
54 	spin_unlock_bh(&wqueue->lock);
55 }
56 
57 static void watch_queue_pipe_buf_release(struct pipe_inode_info *pipe,
58 					 struct pipe_buffer *buf)
59 {
60 	struct watch_queue *wqueue = (struct watch_queue *)buf->private;
61 	struct page *page;
62 	unsigned int bit;
63 
64 	/* We need to work out which note within the page this refers to, but
65 	 * the note might have been maximum size, so merely ANDing the offset
66 	 * off doesn't work.  OTOH, the note must've been more than zero size.
67 	 */
68 	bit = buf->offset + buf->len;
69 	if ((bit & (WATCH_QUEUE_NOTE_SIZE - 1)) == 0)
70 		bit -= WATCH_QUEUE_NOTE_SIZE;
71 	bit /= WATCH_QUEUE_NOTE_SIZE;
72 
73 	page = buf->page;
74 	bit += page->index;
75 
76 	set_bit(bit, wqueue->notes_bitmap);
77 	generic_pipe_buf_release(pipe, buf);
78 }
79 
80 // No try_steal function => no stealing
81 #define watch_queue_pipe_buf_try_steal NULL
82 
83 /* New data written to a pipe may be appended to a buffer with this type. */
84 static const struct pipe_buf_operations watch_queue_pipe_buf_ops = {
85 	.release	= watch_queue_pipe_buf_release,
86 	.try_steal	= watch_queue_pipe_buf_try_steal,
87 	.get		= generic_pipe_buf_get,
88 };
89 
90 /*
91  * Post a notification to a watch queue.
92  *
93  * Must be called with the RCU lock for reading, and the
94  * watch_queue lock held, which guarantees that the pipe
95  * hasn't been released.
96  */
97 static bool post_one_notification(struct watch_queue *wqueue,
98 				  struct watch_notification *n)
99 {
100 	void *p;
101 	struct pipe_inode_info *pipe = wqueue->pipe;
102 	struct pipe_buffer *buf;
103 	struct page *page;
104 	unsigned int head, tail, mask, note, offset, len;
105 	bool done = false;
106 
107 	spin_lock_irq(&pipe->rd_wait.lock);
108 
109 	mask = pipe->ring_size - 1;
110 	head = pipe->head;
111 	tail = pipe->tail;
112 	if (pipe_full(head, tail, pipe->ring_size))
113 		goto lost;
114 
115 	note = find_first_bit(wqueue->notes_bitmap, wqueue->nr_notes);
116 	if (note >= wqueue->nr_notes)
117 		goto lost;
118 
119 	page = wqueue->notes[note / WATCH_QUEUE_NOTES_PER_PAGE];
120 	offset = note % WATCH_QUEUE_NOTES_PER_PAGE * WATCH_QUEUE_NOTE_SIZE;
121 	get_page(page);
122 	len = n->info & WATCH_INFO_LENGTH;
123 	p = kmap_atomic(page);
124 	memcpy(p + offset, n, len);
125 	kunmap_atomic(p);
126 
127 	buf = &pipe->bufs[head & mask];
128 	buf->page = page;
129 	buf->private = (unsigned long)wqueue;
130 	buf->ops = &watch_queue_pipe_buf_ops;
131 	buf->offset = offset;
132 	buf->len = len;
133 	buf->flags = PIPE_BUF_FLAG_WHOLE;
134 	smp_store_release(&pipe->head, head + 1); /* vs pipe_read() */
135 
136 	if (!test_and_clear_bit(note, wqueue->notes_bitmap)) {
137 		spin_unlock_irq(&pipe->rd_wait.lock);
138 		BUG();
139 	}
140 	wake_up_interruptible_sync_poll_locked(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM);
141 	done = true;
142 
143 out:
144 	spin_unlock_irq(&pipe->rd_wait.lock);
145 	if (done)
146 		kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
147 	return done;
148 
149 lost:
150 	buf = &pipe->bufs[(head - 1) & mask];
151 	buf->flags |= PIPE_BUF_FLAG_LOSS;
152 	goto out;
153 }
154 
155 /*
156  * Apply filter rules to a notification.
157  */
158 static bool filter_watch_notification(const struct watch_filter *wf,
159 				      const struct watch_notification *n)
160 {
161 	const struct watch_type_filter *wt;
162 	unsigned int st_bits = sizeof(wt->subtype_filter[0]) * 8;
163 	unsigned int st_index = n->subtype / st_bits;
164 	unsigned int st_bit = 1U << (n->subtype % st_bits);
165 	int i;
166 
167 	if (!test_bit(n->type, wf->type_filter))
168 		return false;
169 
170 	for (i = 0; i < wf->nr_filters; i++) {
171 		wt = &wf->filters[i];
172 		if (n->type == wt->type &&
173 		    (wt->subtype_filter[st_index] & st_bit) &&
174 		    (n->info & wt->info_mask) == wt->info_filter)
175 			return true;
176 	}
177 
178 	return false; /* If there is a filter, the default is to reject. */
179 }
180 
181 /**
182  * __post_watch_notification - Post an event notification
183  * @wlist: The watch list to post the event to.
184  * @n: The notification record to post.
185  * @cred: The creds of the process that triggered the notification.
186  * @id: The ID to match on the watch.
187  *
188  * Post a notification of an event into a set of watch queues and let the users
189  * know.
190  *
191  * The size of the notification should be set in n->info & WATCH_INFO_LENGTH and
192  * should be in units of sizeof(*n).
193  */
194 void __post_watch_notification(struct watch_list *wlist,
195 			       struct watch_notification *n,
196 			       const struct cred *cred,
197 			       u64 id)
198 {
199 	const struct watch_filter *wf;
200 	struct watch_queue *wqueue;
201 	struct watch *watch;
202 
203 	if (((n->info & WATCH_INFO_LENGTH) >> WATCH_INFO_LENGTH__SHIFT) == 0) {
204 		WARN_ON(1);
205 		return;
206 	}
207 
208 	rcu_read_lock();
209 
210 	hlist_for_each_entry_rcu(watch, &wlist->watchers, list_node) {
211 		if (watch->id != id)
212 			continue;
213 		n->info &= ~WATCH_INFO_ID;
214 		n->info |= watch->info_id;
215 
216 		wqueue = rcu_dereference(watch->queue);
217 		wf = rcu_dereference(wqueue->filter);
218 		if (wf && !filter_watch_notification(wf, n))
219 			continue;
220 
221 		if (security_post_notification(watch->cred, cred, n) < 0)
222 			continue;
223 
224 		if (lock_wqueue(wqueue)) {
225 			post_one_notification(wqueue, n);
226 			unlock_wqueue(wqueue);
227 		}
228 	}
229 
230 	rcu_read_unlock();
231 }
232 EXPORT_SYMBOL(__post_watch_notification);
233 
234 /*
235  * Allocate sufficient pages to preallocation for the requested number of
236  * notifications.
237  */
238 long watch_queue_set_size(struct pipe_inode_info *pipe, unsigned int nr_notes)
239 {
240 	struct watch_queue *wqueue = pipe->watch_queue;
241 	struct page **pages;
242 	unsigned long *bitmap;
243 	unsigned long user_bufs;
244 	int ret, i, nr_pages;
245 
246 	if (!wqueue)
247 		return -ENODEV;
248 	if (wqueue->notes)
249 		return -EBUSY;
250 
251 	if (nr_notes < 1 ||
252 	    nr_notes > 512) /* TODO: choose a better hard limit */
253 		return -EINVAL;
254 
255 	nr_pages = (nr_notes + WATCH_QUEUE_NOTES_PER_PAGE - 1);
256 	nr_pages /= WATCH_QUEUE_NOTES_PER_PAGE;
257 	user_bufs = account_pipe_buffers(pipe->user, pipe->nr_accounted, nr_pages);
258 
259 	if (nr_pages > pipe->max_usage &&
260 	    (too_many_pipe_buffers_hard(user_bufs) ||
261 	     too_many_pipe_buffers_soft(user_bufs)) &&
262 	    pipe_is_unprivileged_user()) {
263 		ret = -EPERM;
264 		goto error;
265 	}
266 
267 	nr_notes = nr_pages * WATCH_QUEUE_NOTES_PER_PAGE;
268 	ret = pipe_resize_ring(pipe, roundup_pow_of_two(nr_notes));
269 	if (ret < 0)
270 		goto error;
271 
272 	ret = -ENOMEM;
273 	pages = kcalloc(sizeof(struct page *), nr_pages, GFP_KERNEL);
274 	if (!pages)
275 		goto error;
276 
277 	for (i = 0; i < nr_pages; i++) {
278 		pages[i] = alloc_page(GFP_KERNEL);
279 		if (!pages[i])
280 			goto error_p;
281 		pages[i]->index = i * WATCH_QUEUE_NOTES_PER_PAGE;
282 	}
283 
284 	bitmap = bitmap_alloc(nr_notes, GFP_KERNEL);
285 	if (!bitmap)
286 		goto error_p;
287 
288 	bitmap_fill(bitmap, nr_notes);
289 	wqueue->notes = pages;
290 	wqueue->notes_bitmap = bitmap;
291 	wqueue->nr_pages = nr_pages;
292 	wqueue->nr_notes = nr_notes;
293 	return 0;
294 
295 error_p:
296 	while (--i >= 0)
297 		__free_page(pages[i]);
298 	kfree(pages);
299 error:
300 	(void) account_pipe_buffers(pipe->user, nr_pages, pipe->nr_accounted);
301 	return ret;
302 }
303 
304 /*
305  * Set the filter on a watch queue.
306  */
307 long watch_queue_set_filter(struct pipe_inode_info *pipe,
308 			    struct watch_notification_filter __user *_filter)
309 {
310 	struct watch_notification_type_filter *tf;
311 	struct watch_notification_filter filter;
312 	struct watch_type_filter *q;
313 	struct watch_filter *wfilter;
314 	struct watch_queue *wqueue = pipe->watch_queue;
315 	int ret, nr_filter = 0, i;
316 
317 	if (!wqueue)
318 		return -ENODEV;
319 
320 	if (!_filter) {
321 		/* Remove the old filter */
322 		wfilter = NULL;
323 		goto set;
324 	}
325 
326 	/* Grab the user's filter specification */
327 	if (copy_from_user(&filter, _filter, sizeof(filter)) != 0)
328 		return -EFAULT;
329 	if (filter.nr_filters == 0 ||
330 	    filter.nr_filters > 16 ||
331 	    filter.__reserved != 0)
332 		return -EINVAL;
333 
334 	tf = memdup_array_user(_filter->filters, filter.nr_filters, sizeof(*tf));
335 	if (IS_ERR(tf))
336 		return PTR_ERR(tf);
337 
338 	ret = -EINVAL;
339 	for (i = 0; i < filter.nr_filters; i++) {
340 		if ((tf[i].info_filter & ~tf[i].info_mask) ||
341 		    tf[i].info_mask & WATCH_INFO_LENGTH)
342 			goto err_filter;
343 		/* Ignore any unknown types */
344 		if (tf[i].type >= WATCH_TYPE__NR)
345 			continue;
346 		nr_filter++;
347 	}
348 
349 	/* Now we need to build the internal filter from only the relevant
350 	 * user-specified filters.
351 	 */
352 	ret = -ENOMEM;
353 	wfilter = kzalloc(struct_size(wfilter, filters, nr_filter), GFP_KERNEL);
354 	if (!wfilter)
355 		goto err_filter;
356 	wfilter->nr_filters = nr_filter;
357 
358 	q = wfilter->filters;
359 	for (i = 0; i < filter.nr_filters; i++) {
360 		if (tf[i].type >= WATCH_TYPE__NR)
361 			continue;
362 
363 		q->type			= tf[i].type;
364 		q->info_filter		= tf[i].info_filter;
365 		q->info_mask		= tf[i].info_mask;
366 		q->subtype_filter[0]	= tf[i].subtype_filter[0];
367 		__set_bit(q->type, wfilter->type_filter);
368 		q++;
369 	}
370 
371 	kfree(tf);
372 set:
373 	pipe_lock(pipe);
374 	wfilter = rcu_replace_pointer(wqueue->filter, wfilter,
375 				      lockdep_is_held(&pipe->mutex));
376 	pipe_unlock(pipe);
377 	if (wfilter)
378 		kfree_rcu(wfilter, rcu);
379 	return 0;
380 
381 err_filter:
382 	kfree(tf);
383 	return ret;
384 }
385 
386 static void __put_watch_queue(struct kref *kref)
387 {
388 	struct watch_queue *wqueue =
389 		container_of(kref, struct watch_queue, usage);
390 	struct watch_filter *wfilter;
391 	int i;
392 
393 	for (i = 0; i < wqueue->nr_pages; i++)
394 		__free_page(wqueue->notes[i]);
395 	kfree(wqueue->notes);
396 	bitmap_free(wqueue->notes_bitmap);
397 
398 	wfilter = rcu_access_pointer(wqueue->filter);
399 	if (wfilter)
400 		kfree_rcu(wfilter, rcu);
401 	kfree_rcu(wqueue, rcu);
402 }
403 
404 /**
405  * put_watch_queue - Dispose of a ref on a watchqueue.
406  * @wqueue: The watch queue to unref.
407  */
408 void put_watch_queue(struct watch_queue *wqueue)
409 {
410 	kref_put(&wqueue->usage, __put_watch_queue);
411 }
412 EXPORT_SYMBOL(put_watch_queue);
413 
414 static void free_watch(struct rcu_head *rcu)
415 {
416 	struct watch *watch = container_of(rcu, struct watch, rcu);
417 
418 	put_watch_queue(rcu_access_pointer(watch->queue));
419 	atomic_dec(&watch->cred->user->nr_watches);
420 	put_cred(watch->cred);
421 	kfree(watch);
422 }
423 
424 static void __put_watch(struct kref *kref)
425 {
426 	struct watch *watch = container_of(kref, struct watch, usage);
427 
428 	call_rcu(&watch->rcu, free_watch);
429 }
430 
431 /*
432  * Discard a watch.
433  */
434 static void put_watch(struct watch *watch)
435 {
436 	kref_put(&watch->usage, __put_watch);
437 }
438 
439 /**
440  * init_watch - Initialise a watch
441  * @watch: The watch to initialise.
442  * @wqueue: The queue to assign.
443  *
444  * Initialise a watch and set the watch queue.
445  */
446 void init_watch(struct watch *watch, struct watch_queue *wqueue)
447 {
448 	kref_init(&watch->usage);
449 	INIT_HLIST_NODE(&watch->list_node);
450 	INIT_HLIST_NODE(&watch->queue_node);
451 	rcu_assign_pointer(watch->queue, wqueue);
452 }
453 
454 static int add_one_watch(struct watch *watch, struct watch_list *wlist, struct watch_queue *wqueue)
455 {
456 	const struct cred *cred;
457 	struct watch *w;
458 
459 	hlist_for_each_entry(w, &wlist->watchers, list_node) {
460 		struct watch_queue *wq = rcu_access_pointer(w->queue);
461 		if (wqueue == wq && watch->id == w->id)
462 			return -EBUSY;
463 	}
464 
465 	cred = current_cred();
466 	if (atomic_inc_return(&cred->user->nr_watches) > task_rlimit(current, RLIMIT_NOFILE)) {
467 		atomic_dec(&cred->user->nr_watches);
468 		return -EAGAIN;
469 	}
470 
471 	watch->cred = get_cred(cred);
472 	rcu_assign_pointer(watch->watch_list, wlist);
473 
474 	kref_get(&wqueue->usage);
475 	kref_get(&watch->usage);
476 	hlist_add_head(&watch->queue_node, &wqueue->watches);
477 	hlist_add_head_rcu(&watch->list_node, &wlist->watchers);
478 	return 0;
479 }
480 
481 /**
482  * add_watch_to_object - Add a watch on an object to a watch list
483  * @watch: The watch to add
484  * @wlist: The watch list to add to
485  *
486  * @watch->queue must have been set to point to the queue to post notifications
487  * to and the watch list of the object to be watched.  @watch->cred must also
488  * have been set to the appropriate credentials and a ref taken on them.
489  *
490  * The caller must pin the queue and the list both and must hold the list
491  * locked against racing watch additions/removals.
492  */
493 int add_watch_to_object(struct watch *watch, struct watch_list *wlist)
494 {
495 	struct watch_queue *wqueue;
496 	int ret = -ENOENT;
497 
498 	rcu_read_lock();
499 
500 	wqueue = rcu_access_pointer(watch->queue);
501 	if (lock_wqueue(wqueue)) {
502 		spin_lock(&wlist->lock);
503 		ret = add_one_watch(watch, wlist, wqueue);
504 		spin_unlock(&wlist->lock);
505 		unlock_wqueue(wqueue);
506 	}
507 
508 	rcu_read_unlock();
509 	return ret;
510 }
511 EXPORT_SYMBOL(add_watch_to_object);
512 
513 /**
514  * remove_watch_from_object - Remove a watch or all watches from an object.
515  * @wlist: The watch list to remove from
516  * @wq: The watch queue of interest (ignored if @all is true)
517  * @id: The ID of the watch to remove (ignored if @all is true)
518  * @all: True to remove all objects
519  *
520  * Remove a specific watch or all watches from an object.  A notification is
521  * sent to the watcher to tell them that this happened.
522  */
523 int remove_watch_from_object(struct watch_list *wlist, struct watch_queue *wq,
524 			     u64 id, bool all)
525 {
526 	struct watch_notification_removal n;
527 	struct watch_queue *wqueue;
528 	struct watch *watch;
529 	int ret = -EBADSLT;
530 
531 	rcu_read_lock();
532 
533 again:
534 	spin_lock(&wlist->lock);
535 	hlist_for_each_entry(watch, &wlist->watchers, list_node) {
536 		if (all ||
537 		    (watch->id == id && rcu_access_pointer(watch->queue) == wq))
538 			goto found;
539 	}
540 	spin_unlock(&wlist->lock);
541 	goto out;
542 
543 found:
544 	ret = 0;
545 	hlist_del_init_rcu(&watch->list_node);
546 	rcu_assign_pointer(watch->watch_list, NULL);
547 	spin_unlock(&wlist->lock);
548 
549 	/* We now own the reference on watch that used to belong to wlist. */
550 
551 	n.watch.type = WATCH_TYPE_META;
552 	n.watch.subtype = WATCH_META_REMOVAL_NOTIFICATION;
553 	n.watch.info = watch->info_id | watch_sizeof(n.watch);
554 	n.id = id;
555 	if (id != 0)
556 		n.watch.info = watch->info_id | watch_sizeof(n);
557 
558 	wqueue = rcu_dereference(watch->queue);
559 
560 	if (lock_wqueue(wqueue)) {
561 		post_one_notification(wqueue, &n.watch);
562 
563 		if (!hlist_unhashed(&watch->queue_node)) {
564 			hlist_del_init_rcu(&watch->queue_node);
565 			put_watch(watch);
566 		}
567 
568 		unlock_wqueue(wqueue);
569 	}
570 
571 	if (wlist->release_watch) {
572 		void (*release_watch)(struct watch *);
573 
574 		release_watch = wlist->release_watch;
575 		rcu_read_unlock();
576 		(*release_watch)(watch);
577 		rcu_read_lock();
578 	}
579 	put_watch(watch);
580 
581 	if (all && !hlist_empty(&wlist->watchers))
582 		goto again;
583 out:
584 	rcu_read_unlock();
585 	return ret;
586 }
587 EXPORT_SYMBOL(remove_watch_from_object);
588 
589 /*
590  * Remove all the watches that are contributory to a queue.  This has the
591  * potential to race with removal of the watches by the destruction of the
592  * objects being watched or with the distribution of notifications.
593  */
594 void watch_queue_clear(struct watch_queue *wqueue)
595 {
596 	struct watch_list *wlist;
597 	struct watch *watch;
598 	bool release;
599 
600 	rcu_read_lock();
601 	spin_lock_bh(&wqueue->lock);
602 
603 	/*
604 	 * This pipe can be freed by callers like free_pipe_info().
605 	 * Removing this reference also prevents new notifications.
606 	 */
607 	wqueue->pipe = NULL;
608 
609 	while (!hlist_empty(&wqueue->watches)) {
610 		watch = hlist_entry(wqueue->watches.first, struct watch, queue_node);
611 		hlist_del_init_rcu(&watch->queue_node);
612 		/* We now own a ref on the watch. */
613 		spin_unlock_bh(&wqueue->lock);
614 
615 		/* We can't do the next bit under the queue lock as we need to
616 		 * get the list lock - which would cause a deadlock if someone
617 		 * was removing from the opposite direction at the same time or
618 		 * posting a notification.
619 		 */
620 		wlist = rcu_dereference(watch->watch_list);
621 		if (wlist) {
622 			void (*release_watch)(struct watch *);
623 
624 			spin_lock(&wlist->lock);
625 
626 			release = !hlist_unhashed(&watch->list_node);
627 			if (release) {
628 				hlist_del_init_rcu(&watch->list_node);
629 				rcu_assign_pointer(watch->watch_list, NULL);
630 
631 				/* We now own a second ref on the watch. */
632 			}
633 
634 			release_watch = wlist->release_watch;
635 			spin_unlock(&wlist->lock);
636 
637 			if (release) {
638 				if (release_watch) {
639 					rcu_read_unlock();
640 					/* This might need to call dput(), so
641 					 * we have to drop all the locks.
642 					 */
643 					(*release_watch)(watch);
644 					rcu_read_lock();
645 				}
646 				put_watch(watch);
647 			}
648 		}
649 
650 		put_watch(watch);
651 		spin_lock_bh(&wqueue->lock);
652 	}
653 
654 	spin_unlock_bh(&wqueue->lock);
655 	rcu_read_unlock();
656 }
657 
658 /**
659  * get_watch_queue - Get a watch queue from its file descriptor.
660  * @fd: The fd to query.
661  */
662 struct watch_queue *get_watch_queue(int fd)
663 {
664 	struct pipe_inode_info *pipe;
665 	struct watch_queue *wqueue = ERR_PTR(-EINVAL);
666 	struct fd f;
667 
668 	f = fdget(fd);
669 	if (f.file) {
670 		pipe = get_pipe_info(f.file, false);
671 		if (pipe && pipe->watch_queue) {
672 			wqueue = pipe->watch_queue;
673 			kref_get(&wqueue->usage);
674 		}
675 		fdput(f);
676 	}
677 
678 	return wqueue;
679 }
680 EXPORT_SYMBOL(get_watch_queue);
681 
682 /*
683  * Initialise a watch queue
684  */
685 int watch_queue_init(struct pipe_inode_info *pipe)
686 {
687 	struct watch_queue *wqueue;
688 
689 	wqueue = kzalloc(sizeof(*wqueue), GFP_KERNEL);
690 	if (!wqueue)
691 		return -ENOMEM;
692 
693 	wqueue->pipe = pipe;
694 	kref_init(&wqueue->usage);
695 	spin_lock_init(&wqueue->lock);
696 	INIT_HLIST_HEAD(&wqueue->watches);
697 
698 	pipe->watch_queue = wqueue;
699 	return 0;
700 }
701