• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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/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 MODULE_LICENSE("GPL");
33 
34 #define WATCH_QUEUE_NOTE_SIZE 128
35 #define WATCH_QUEUE_NOTES_PER_PAGE (PAGE_SIZE / WATCH_QUEUE_NOTE_SIZE)
36 
37 /*
38  * This must be called under the RCU read-lock, which makes
39  * sure that the wqueue still exists. It can then take the lock,
40  * and check that the wqueue hasn't been destroyed, which in
41  * turn makes sure that the notification pipe still exists.
42  */
lock_wqueue(struct watch_queue * wqueue)43 static inline bool lock_wqueue(struct watch_queue *wqueue)
44 {
45 	spin_lock_bh(&wqueue->lock);
46 	if (unlikely(wqueue->defunct)) {
47 		spin_unlock_bh(&wqueue->lock);
48 		return false;
49 	}
50 	return true;
51 }
52 
unlock_wqueue(struct watch_queue * wqueue)53 static inline void unlock_wqueue(struct watch_queue *wqueue)
54 {
55 	spin_unlock_bh(&wqueue->lock);
56 }
57 
watch_queue_pipe_buf_release(struct pipe_inode_info * pipe,struct pipe_buffer * buf)58 static void watch_queue_pipe_buf_release(struct pipe_inode_info *pipe,
59 					 struct pipe_buffer *buf)
60 {
61 	struct watch_queue *wqueue = (struct watch_queue *)buf->private;
62 	struct page *page;
63 	unsigned int bit;
64 
65 	/* We need to work out which note within the page this refers to, but
66 	 * the note might have been maximum size, so merely ANDing the offset
67 	 * off doesn't work.  OTOH, the note must've been more than zero size.
68 	 */
69 	bit = buf->offset + buf->len;
70 	if ((bit & (WATCH_QUEUE_NOTE_SIZE - 1)) == 0)
71 		bit -= WATCH_QUEUE_NOTE_SIZE;
72 	bit /= WATCH_QUEUE_NOTE_SIZE;
73 
74 	page = buf->page;
75 	bit += page->index;
76 
77 	set_bit(bit, wqueue->notes_bitmap);
78 	generic_pipe_buf_release(pipe, buf);
79 }
80 
81 // No try_steal function => no stealing
82 #define watch_queue_pipe_buf_try_steal NULL
83 
84 /* New data written to a pipe may be appended to a buffer with this type. */
85 static const struct pipe_buf_operations watch_queue_pipe_buf_ops = {
86 	.release	= watch_queue_pipe_buf_release,
87 	.try_steal	= watch_queue_pipe_buf_try_steal,
88 	.get		= generic_pipe_buf_get,
89 };
90 
91 /*
92  * Post a notification to a watch queue.
93  *
94  * Must be called with the RCU lock for reading, and the
95  * watch_queue lock held, which guarantees that the pipe
96  * hasn't been released.
97  */
post_one_notification(struct watch_queue * wqueue,struct watch_notification * n)98 static bool post_one_notification(struct watch_queue *wqueue,
99 				  struct watch_notification *n)
100 {
101 	void *p;
102 	struct pipe_inode_info *pipe = wqueue->pipe;
103 	struct pipe_buffer *buf;
104 	struct page *page;
105 	unsigned int head, tail, mask, note, offset, len;
106 	bool done = false;
107 
108 	if (!pipe)
109 		return false;
110 
111 	spin_lock_irq(&pipe->rd_wait.lock);
112 
113 	mask = pipe->ring_size - 1;
114 	head = pipe->head;
115 	tail = pipe->tail;
116 	if (pipe_full(head, tail, pipe->ring_size))
117 		goto lost;
118 
119 	note = find_first_bit(wqueue->notes_bitmap, wqueue->nr_notes);
120 	if (note >= wqueue->nr_notes)
121 		goto lost;
122 
123 	page = wqueue->notes[note / WATCH_QUEUE_NOTES_PER_PAGE];
124 	offset = note % WATCH_QUEUE_NOTES_PER_PAGE * WATCH_QUEUE_NOTE_SIZE;
125 	get_page(page);
126 	len = n->info & WATCH_INFO_LENGTH;
127 	p = kmap_atomic(page);
128 	memcpy(p + offset, n, len);
129 	kunmap_atomic(p);
130 
131 	buf = &pipe->bufs[head & mask];
132 	buf->page = page;
133 	buf->private = (unsigned long)wqueue;
134 	buf->ops = &watch_queue_pipe_buf_ops;
135 	buf->offset = offset;
136 	buf->len = len;
137 	buf->flags = PIPE_BUF_FLAG_WHOLE;
138 	smp_store_release(&pipe->head, head + 1); /* vs pipe_read() */
139 
140 	if (!test_and_clear_bit(note, wqueue->notes_bitmap)) {
141 		spin_unlock_irq(&pipe->rd_wait.lock);
142 		BUG();
143 	}
144 	wake_up_interruptible_sync_poll_locked(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM);
145 	done = true;
146 
147 out:
148 	spin_unlock_irq(&pipe->rd_wait.lock);
149 	if (done)
150 		kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
151 	return done;
152 
153 lost:
154 	buf = &pipe->bufs[(head - 1) & mask];
155 	buf->flags |= PIPE_BUF_FLAG_LOSS;
156 	goto out;
157 }
158 
159 /*
160  * Apply filter rules to a notification.
161  */
filter_watch_notification(const struct watch_filter * wf,const struct watch_notification * n)162 static bool filter_watch_notification(const struct watch_filter *wf,
163 				      const struct watch_notification *n)
164 {
165 	const struct watch_type_filter *wt;
166 	unsigned int st_bits = sizeof(wt->subtype_filter[0]) * 8;
167 	unsigned int st_index = n->subtype / st_bits;
168 	unsigned int st_bit = 1U << (n->subtype % st_bits);
169 	int i;
170 
171 	if (!test_bit(n->type, wf->type_filter))
172 		return false;
173 
174 	for (i = 0; i < wf->nr_filters; i++) {
175 		wt = &wf->filters[i];
176 		if (n->type == wt->type &&
177 		    (wt->subtype_filter[st_index] & st_bit) &&
178 		    (n->info & wt->info_mask) == wt->info_filter)
179 			return true;
180 	}
181 
182 	return false; /* If there is a filter, the default is to reject. */
183 }
184 
185 /**
186  * __post_watch_notification - Post an event notification
187  * @wlist: The watch list to post the event to.
188  * @n: The notification record to post.
189  * @cred: The creds of the process that triggered the notification.
190  * @id: The ID to match on the watch.
191  *
192  * Post a notification of an event into a set of watch queues and let the users
193  * know.
194  *
195  * The size of the notification should be set in n->info & WATCH_INFO_LENGTH and
196  * should be in units of sizeof(*n).
197  */
__post_watch_notification(struct watch_list * wlist,struct watch_notification * n,const struct cred * cred,u64 id)198 void __post_watch_notification(struct watch_list *wlist,
199 			       struct watch_notification *n,
200 			       const struct cred *cred,
201 			       u64 id)
202 {
203 	const struct watch_filter *wf;
204 	struct watch_queue *wqueue;
205 	struct watch *watch;
206 
207 	if (((n->info & WATCH_INFO_LENGTH) >> WATCH_INFO_LENGTH__SHIFT) == 0) {
208 		WARN_ON(1);
209 		return;
210 	}
211 
212 	rcu_read_lock();
213 
214 	hlist_for_each_entry_rcu(watch, &wlist->watchers, list_node) {
215 		if (watch->id != id)
216 			continue;
217 		n->info &= ~WATCH_INFO_ID;
218 		n->info |= watch->info_id;
219 
220 		wqueue = rcu_dereference(watch->queue);
221 		wf = rcu_dereference(wqueue->filter);
222 		if (wf && !filter_watch_notification(wf, n))
223 			continue;
224 
225 		if (security_post_notification(watch->cred, cred, n) < 0)
226 			continue;
227 
228 		if (lock_wqueue(wqueue)) {
229 			post_one_notification(wqueue, n);
230 			unlock_wqueue(wqueue);
231 		}
232 	}
233 
234 	rcu_read_unlock();
235 }
236 EXPORT_SYMBOL(__post_watch_notification);
237 
238 /*
239  * Allocate sufficient pages to preallocation for the requested number of
240  * notifications.
241  */
watch_queue_set_size(struct pipe_inode_info * pipe,unsigned int nr_notes)242 long watch_queue_set_size(struct pipe_inode_info *pipe, unsigned int nr_notes)
243 {
244 	struct watch_queue *wqueue = pipe->watch_queue;
245 	struct page **pages;
246 	unsigned long *bitmap;
247 	unsigned long user_bufs;
248 	unsigned int bmsize;
249 	int ret, i, nr_pages;
250 
251 	if (!wqueue)
252 		return -ENODEV;
253 	if (wqueue->notes)
254 		return -EBUSY;
255 
256 	if (nr_notes < 1 ||
257 	    nr_notes > 512) /* TODO: choose a better hard limit */
258 		return -EINVAL;
259 
260 	nr_pages = (nr_notes + WATCH_QUEUE_NOTES_PER_PAGE - 1);
261 	nr_pages /= WATCH_QUEUE_NOTES_PER_PAGE;
262 	user_bufs = account_pipe_buffers(pipe->user, pipe->nr_accounted, nr_pages);
263 
264 	if (nr_pages > pipe->max_usage &&
265 	    (too_many_pipe_buffers_hard(user_bufs) ||
266 	     too_many_pipe_buffers_soft(user_bufs)) &&
267 	    pipe_is_unprivileged_user()) {
268 		ret = -EPERM;
269 		goto error;
270 	}
271 
272 	nr_notes = nr_pages * WATCH_QUEUE_NOTES_PER_PAGE;
273 	ret = pipe_resize_ring(pipe, roundup_pow_of_two(nr_notes));
274 	if (ret < 0)
275 		goto error;
276 
277 	ret = -ENOMEM;
278 	pages = kcalloc(sizeof(struct page *), nr_pages, GFP_KERNEL);
279 	if (!pages)
280 		goto error;
281 
282 	for (i = 0; i < nr_pages; i++) {
283 		pages[i] = alloc_page(GFP_KERNEL);
284 		if (!pages[i])
285 			goto error_p;
286 		pages[i]->index = i * WATCH_QUEUE_NOTES_PER_PAGE;
287 	}
288 
289 	bmsize = (nr_notes + BITS_PER_LONG - 1) / BITS_PER_LONG;
290 	bmsize *= sizeof(unsigned long);
291 	bitmap = kmalloc(bmsize, GFP_KERNEL);
292 	if (!bitmap)
293 		goto error_p;
294 
295 	memset(bitmap, 0xff, bmsize);
296 	wqueue->notes = pages;
297 	wqueue->notes_bitmap = bitmap;
298 	wqueue->nr_pages = nr_pages;
299 	wqueue->nr_notes = nr_notes;
300 	return 0;
301 
302 error_p:
303 	while (--i >= 0)
304 		__free_page(pages[i]);
305 	kfree(pages);
306 error:
307 	(void) account_pipe_buffers(pipe->user, nr_pages, pipe->nr_accounted);
308 	return ret;
309 }
310 
311 /*
312  * Set the filter on a watch queue.
313  */
watch_queue_set_filter(struct pipe_inode_info * pipe,struct watch_notification_filter __user * _filter)314 long watch_queue_set_filter(struct pipe_inode_info *pipe,
315 			    struct watch_notification_filter __user *_filter)
316 {
317 	struct watch_notification_type_filter *tf;
318 	struct watch_notification_filter filter;
319 	struct watch_type_filter *q;
320 	struct watch_filter *wfilter;
321 	struct watch_queue *wqueue = pipe->watch_queue;
322 	int ret, nr_filter = 0, i;
323 
324 	if (!wqueue)
325 		return -ENODEV;
326 
327 	if (!_filter) {
328 		/* Remove the old filter */
329 		wfilter = NULL;
330 		goto set;
331 	}
332 
333 	/* Grab the user's filter specification */
334 	if (copy_from_user(&filter, _filter, sizeof(filter)) != 0)
335 		return -EFAULT;
336 	if (filter.nr_filters == 0 ||
337 	    filter.nr_filters > 16 ||
338 	    filter.__reserved != 0)
339 		return -EINVAL;
340 
341 	tf = memdup_array_user(_filter->filters, filter.nr_filters, sizeof(*tf));
342 	if (IS_ERR(tf))
343 		return PTR_ERR(tf);
344 
345 	ret = -EINVAL;
346 	for (i = 0; i < filter.nr_filters; i++) {
347 		if ((tf[i].info_filter & ~tf[i].info_mask) ||
348 		    tf[i].info_mask & WATCH_INFO_LENGTH)
349 			goto err_filter;
350 		/* Ignore any unknown types */
351 		if (tf[i].type >= WATCH_TYPE__NR)
352 			continue;
353 		nr_filter++;
354 	}
355 
356 	/* Now we need to build the internal filter from only the relevant
357 	 * user-specified filters.
358 	 */
359 	ret = -ENOMEM;
360 	wfilter = kzalloc(struct_size(wfilter, filters, nr_filter), GFP_KERNEL);
361 	if (!wfilter)
362 		goto err_filter;
363 	wfilter->nr_filters = nr_filter;
364 
365 	q = wfilter->filters;
366 	for (i = 0; i < filter.nr_filters; i++) {
367 		if (tf[i].type >= WATCH_TYPE__NR)
368 			continue;
369 
370 		q->type			= tf[i].type;
371 		q->info_filter		= tf[i].info_filter;
372 		q->info_mask		= tf[i].info_mask;
373 		q->subtype_filter[0]	= tf[i].subtype_filter[0];
374 		__set_bit(q->type, wfilter->type_filter);
375 		q++;
376 	}
377 
378 	kfree(tf);
379 set:
380 	pipe_lock(pipe);
381 	wfilter = rcu_replace_pointer(wqueue->filter, wfilter,
382 				      lockdep_is_held(&pipe->mutex));
383 	pipe_unlock(pipe);
384 	if (wfilter)
385 		kfree_rcu(wfilter, rcu);
386 	return 0;
387 
388 err_filter:
389 	kfree(tf);
390 	return ret;
391 }
392 
__put_watch_queue(struct kref * kref)393 static void __put_watch_queue(struct kref *kref)
394 {
395 	struct watch_queue *wqueue =
396 		container_of(kref, struct watch_queue, usage);
397 	struct watch_filter *wfilter;
398 	int i;
399 
400 	for (i = 0; i < wqueue->nr_pages; i++)
401 		__free_page(wqueue->notes[i]);
402 	kfree(wqueue->notes);
403 	bitmap_free(wqueue->notes_bitmap);
404 
405 	wfilter = rcu_access_pointer(wqueue->filter);
406 	if (wfilter)
407 		kfree_rcu(wfilter, rcu);
408 	kfree_rcu(wqueue, rcu);
409 }
410 
411 /**
412  * put_watch_queue - Dispose of a ref on a watchqueue.
413  * @wqueue: The watch queue to unref.
414  */
put_watch_queue(struct watch_queue * wqueue)415 void put_watch_queue(struct watch_queue *wqueue)
416 {
417 	kref_put(&wqueue->usage, __put_watch_queue);
418 }
419 EXPORT_SYMBOL(put_watch_queue);
420 
free_watch(struct rcu_head * rcu)421 static void free_watch(struct rcu_head *rcu)
422 {
423 	struct watch *watch = container_of(rcu, struct watch, rcu);
424 
425 	put_watch_queue(rcu_access_pointer(watch->queue));
426 	atomic_dec(&watch->cred->user->nr_watches);
427 	put_cred(watch->cred);
428 	kfree(watch);
429 }
430 
__put_watch(struct kref * kref)431 static void __put_watch(struct kref *kref)
432 {
433 	struct watch *watch = container_of(kref, struct watch, usage);
434 
435 	call_rcu(&watch->rcu, free_watch);
436 }
437 
438 /*
439  * Discard a watch.
440  */
put_watch(struct watch * watch)441 static void put_watch(struct watch *watch)
442 {
443 	kref_put(&watch->usage, __put_watch);
444 }
445 
446 /**
447  * init_watch - Initialise a watch
448  * @watch: The watch to initialise.
449  * @wqueue: The queue to assign.
450  *
451  * Initialise a watch and set the watch queue.
452  */
init_watch(struct watch * watch,struct watch_queue * wqueue)453 void init_watch(struct watch *watch, struct watch_queue *wqueue)
454 {
455 	kref_init(&watch->usage);
456 	INIT_HLIST_NODE(&watch->list_node);
457 	INIT_HLIST_NODE(&watch->queue_node);
458 	rcu_assign_pointer(watch->queue, wqueue);
459 }
460 
add_one_watch(struct watch * watch,struct watch_list * wlist,struct watch_queue * wqueue)461 static int add_one_watch(struct watch *watch, struct watch_list *wlist, struct watch_queue *wqueue)
462 {
463 	const struct cred *cred;
464 	struct watch *w;
465 
466 	hlist_for_each_entry(w, &wlist->watchers, list_node) {
467 		struct watch_queue *wq = rcu_access_pointer(w->queue);
468 		if (wqueue == wq && watch->id == w->id)
469 			return -EBUSY;
470 	}
471 
472 	cred = current_cred();
473 	if (atomic_inc_return(&cred->user->nr_watches) > task_rlimit(current, RLIMIT_NOFILE)) {
474 		atomic_dec(&cred->user->nr_watches);
475 		return -EAGAIN;
476 	}
477 
478 	watch->cred = get_cred(cred);
479 	rcu_assign_pointer(watch->watch_list, wlist);
480 
481 	kref_get(&wqueue->usage);
482 	kref_get(&watch->usage);
483 	hlist_add_head(&watch->queue_node, &wqueue->watches);
484 	hlist_add_head_rcu(&watch->list_node, &wlist->watchers);
485 	return 0;
486 }
487 
488 /**
489  * add_watch_to_object - Add a watch on an object to a watch list
490  * @watch: The watch to add
491  * @wlist: The watch list to add to
492  *
493  * @watch->queue must have been set to point to the queue to post notifications
494  * to and the watch list of the object to be watched.  @watch->cred must also
495  * have been set to the appropriate credentials and a ref taken on them.
496  *
497  * The caller must pin the queue and the list both and must hold the list
498  * locked against racing watch additions/removals.
499  */
add_watch_to_object(struct watch * watch,struct watch_list * wlist)500 int add_watch_to_object(struct watch *watch, struct watch_list *wlist)
501 {
502 	struct watch_queue *wqueue;
503 	int ret = -ENOENT;
504 
505 	rcu_read_lock();
506 
507 	wqueue = rcu_access_pointer(watch->queue);
508 	if (lock_wqueue(wqueue)) {
509 		spin_lock(&wlist->lock);
510 		ret = add_one_watch(watch, wlist, wqueue);
511 		spin_unlock(&wlist->lock);
512 		unlock_wqueue(wqueue);
513 	}
514 
515 	rcu_read_unlock();
516 	return ret;
517 }
518 EXPORT_SYMBOL(add_watch_to_object);
519 
520 /**
521  * remove_watch_from_object - Remove a watch or all watches from an object.
522  * @wlist: The watch list to remove from
523  * @wq: The watch queue of interest (ignored if @all is true)
524  * @id: The ID of the watch to remove (ignored if @all is true)
525  * @all: True to remove all objects
526  *
527  * Remove a specific watch or all watches from an object.  A notification is
528  * sent to the watcher to tell them that this happened.
529  */
remove_watch_from_object(struct watch_list * wlist,struct watch_queue * wq,u64 id,bool all)530 int remove_watch_from_object(struct watch_list *wlist, struct watch_queue *wq,
531 			     u64 id, bool all)
532 {
533 	struct watch_notification_removal n;
534 	struct watch_queue *wqueue;
535 	struct watch *watch;
536 	int ret = -EBADSLT;
537 
538 	rcu_read_lock();
539 
540 again:
541 	spin_lock(&wlist->lock);
542 	hlist_for_each_entry(watch, &wlist->watchers, list_node) {
543 		if (all ||
544 		    (watch->id == id && rcu_access_pointer(watch->queue) == wq))
545 			goto found;
546 	}
547 	spin_unlock(&wlist->lock);
548 	goto out;
549 
550 found:
551 	ret = 0;
552 	hlist_del_init_rcu(&watch->list_node);
553 	rcu_assign_pointer(watch->watch_list, NULL);
554 	spin_unlock(&wlist->lock);
555 
556 	/* We now own the reference on watch that used to belong to wlist. */
557 
558 	n.watch.type = WATCH_TYPE_META;
559 	n.watch.subtype = WATCH_META_REMOVAL_NOTIFICATION;
560 	n.watch.info = watch->info_id | watch_sizeof(n.watch);
561 	n.id = id;
562 	if (id != 0)
563 		n.watch.info = watch->info_id | watch_sizeof(n);
564 
565 	wqueue = rcu_dereference(watch->queue);
566 
567 	if (lock_wqueue(wqueue)) {
568 		post_one_notification(wqueue, &n.watch);
569 
570 		if (!hlist_unhashed(&watch->queue_node)) {
571 			hlist_del_init_rcu(&watch->queue_node);
572 			put_watch(watch);
573 		}
574 
575 		unlock_wqueue(wqueue);
576 	}
577 
578 	if (wlist->release_watch) {
579 		void (*release_watch)(struct watch *);
580 
581 		release_watch = wlist->release_watch;
582 		rcu_read_unlock();
583 		(*release_watch)(watch);
584 		rcu_read_lock();
585 	}
586 	put_watch(watch);
587 
588 	if (all && !hlist_empty(&wlist->watchers))
589 		goto again;
590 out:
591 	rcu_read_unlock();
592 	return ret;
593 }
594 EXPORT_SYMBOL(remove_watch_from_object);
595 
596 /*
597  * Remove all the watches that are contributory to a queue.  This has the
598  * potential to race with removal of the watches by the destruction of the
599  * objects being watched or with the distribution of notifications.
600  */
watch_queue_clear(struct watch_queue * wqueue)601 void watch_queue_clear(struct watch_queue *wqueue)
602 {
603 	struct watch_list *wlist;
604 	struct watch *watch;
605 	bool release;
606 
607 	rcu_read_lock();
608 	spin_lock_bh(&wqueue->lock);
609 
610 	/* Prevent new notifications from being stored. */
611 	wqueue->defunct = true;
612 
613 	while (!hlist_empty(&wqueue->watches)) {
614 		watch = hlist_entry(wqueue->watches.first, struct watch, queue_node);
615 		hlist_del_init_rcu(&watch->queue_node);
616 		/* We now own a ref on the watch. */
617 		spin_unlock_bh(&wqueue->lock);
618 
619 		/* We can't do the next bit under the queue lock as we need to
620 		 * get the list lock - which would cause a deadlock if someone
621 		 * was removing from the opposite direction at the same time or
622 		 * posting a notification.
623 		 */
624 		wlist = rcu_dereference(watch->watch_list);
625 		if (wlist) {
626 			void (*release_watch)(struct watch *);
627 
628 			spin_lock(&wlist->lock);
629 
630 			release = !hlist_unhashed(&watch->list_node);
631 			if (release) {
632 				hlist_del_init_rcu(&watch->list_node);
633 				rcu_assign_pointer(watch->watch_list, NULL);
634 
635 				/* We now own a second ref on the watch. */
636 			}
637 
638 			release_watch = wlist->release_watch;
639 			spin_unlock(&wlist->lock);
640 
641 			if (release) {
642 				if (release_watch) {
643 					rcu_read_unlock();
644 					/* This might need to call dput(), so
645 					 * we have to drop all the locks.
646 					 */
647 					(*release_watch)(watch);
648 					rcu_read_lock();
649 				}
650 				put_watch(watch);
651 			}
652 		}
653 
654 		put_watch(watch);
655 		spin_lock_bh(&wqueue->lock);
656 	}
657 
658 	spin_unlock_bh(&wqueue->lock);
659 	rcu_read_unlock();
660 }
661 
662 /**
663  * get_watch_queue - Get a watch queue from its file descriptor.
664  * @fd: The fd to query.
665  */
get_watch_queue(int fd)666 struct watch_queue *get_watch_queue(int fd)
667 {
668 	struct pipe_inode_info *pipe;
669 	struct watch_queue *wqueue = ERR_PTR(-EINVAL);
670 	struct fd f;
671 
672 	f = fdget(fd);
673 	if (f.file) {
674 		pipe = get_pipe_info(f.file, false);
675 		if (pipe && pipe->watch_queue) {
676 			wqueue = pipe->watch_queue;
677 			kref_get(&wqueue->usage);
678 		}
679 		fdput(f);
680 	}
681 
682 	return wqueue;
683 }
684 EXPORT_SYMBOL(get_watch_queue);
685 
686 /*
687  * Initialise a watch queue
688  */
watch_queue_init(struct pipe_inode_info * pipe)689 int watch_queue_init(struct pipe_inode_info *pipe)
690 {
691 	struct watch_queue *wqueue;
692 
693 	wqueue = kzalloc(sizeof(*wqueue), GFP_KERNEL);
694 	if (!wqueue)
695 		return -ENOMEM;
696 
697 	wqueue->pipe = pipe;
698 	kref_init(&wqueue->usage);
699 	spin_lock_init(&wqueue->lock);
700 	INIT_HLIST_HEAD(&wqueue->watches);
701 
702 	pipe->watch_queue = wqueue;
703 	return 0;
704 }
705