1 // SPDX-License-Identifier: GPL-2.0
2 /* Watch queue and general notification mechanism, built on pipes
4 * Copyright (C) 2020 Red Hat, Inc. All Rights Reserved.
5 * Written by David Howells (dhowells@redhat.com)
7 * See Documentation/watch_queue.rst
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>
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>
30 MODULE_DESCRIPTION("Watch queue");
31 MODULE_AUTHOR("Red Hat, Inc.");
32 MODULE_LICENSE("GPL");
34 #define WATCH_QUEUE_NOTE_SIZE 128
35 #define WATCH_QUEUE_NOTES_PER_PAGE (PAGE_SIZE / WATCH_QUEUE_NOTE_SIZE)
37 static void watch_queue_pipe_buf_release(struct pipe_inode_info *pipe,
38 struct pipe_buffer *buf)
40 struct watch_queue *wqueue = (struct watch_queue *)buf->private;
44 /* We need to work out which note within the page this refers to, but
45 * the note might have been maximum size, so merely ANDing the offset
46 * off doesn't work. OTOH, the note must've been more than zero size.
48 bit = buf->offset + buf->len;
49 if ((bit & (WATCH_QUEUE_NOTE_SIZE - 1)) == 0)
50 bit -= WATCH_QUEUE_NOTE_SIZE;
51 bit /= WATCH_QUEUE_NOTE_SIZE;
56 set_bit(bit, wqueue->notes_bitmap);
57 generic_pipe_buf_release(pipe, buf);
60 // No try_steal function => no stealing
61 #define watch_queue_pipe_buf_try_steal NULL
63 /* New data written to a pipe may be appended to a buffer with this type. */
64 static const struct pipe_buf_operations watch_queue_pipe_buf_ops = {
65 .release = watch_queue_pipe_buf_release,
66 .try_steal = watch_queue_pipe_buf_try_steal,
67 .get = generic_pipe_buf_get,
71 * Post a notification to a watch queue.
73 static bool post_one_notification(struct watch_queue *wqueue,
74 struct watch_notification *n)
77 struct pipe_inode_info *pipe = wqueue->pipe;
78 struct pipe_buffer *buf;
80 unsigned int head, tail, mask, note, offset, len;
86 spin_lock_irq(&pipe->rd_wait.lock);
91 mask = pipe->ring_size - 1;
94 if (pipe_full(head, tail, pipe->ring_size))
97 note = find_first_bit(wqueue->notes_bitmap, wqueue->nr_notes);
98 if (note >= wqueue->nr_notes)
101 page = wqueue->notes[note / WATCH_QUEUE_NOTES_PER_PAGE];
102 offset = note % WATCH_QUEUE_NOTES_PER_PAGE * WATCH_QUEUE_NOTE_SIZE;
104 len = n->info & WATCH_INFO_LENGTH;
105 p = kmap_atomic(page);
106 memcpy(p + offset, n, len);
109 buf = &pipe->bufs[head & mask];
111 buf->private = (unsigned long)wqueue;
112 buf->ops = &watch_queue_pipe_buf_ops;
113 buf->offset = offset;
115 buf->flags = PIPE_BUF_FLAG_WHOLE;
116 smp_store_release(&pipe->head, head + 1); /* vs pipe_read() */
118 if (!test_and_clear_bit(note, wqueue->notes_bitmap)) {
119 spin_unlock_irq(&pipe->rd_wait.lock);
122 wake_up_interruptible_sync_poll_locked(&pipe->rd_wait, EPOLLIN | EPOLLRDNORM);
126 spin_unlock_irq(&pipe->rd_wait.lock);
128 kill_fasync(&pipe->fasync_readers, SIGIO, POLL_IN);
132 buf = &pipe->bufs[(head - 1) & mask];
133 buf->flags |= PIPE_BUF_FLAG_LOSS;
138 * Apply filter rules to a notification.
140 static bool filter_watch_notification(const struct watch_filter *wf,
141 const struct watch_notification *n)
143 const struct watch_type_filter *wt;
144 unsigned int st_bits = sizeof(wt->subtype_filter[0]) * 8;
145 unsigned int st_index = n->subtype / st_bits;
146 unsigned int st_bit = 1U << (n->subtype % st_bits);
149 if (!test_bit(n->type, wf->type_filter))
152 for (i = 0; i < wf->nr_filters; i++) {
153 wt = &wf->filters[i];
154 if (n->type == wt->type &&
155 (wt->subtype_filter[st_index] & st_bit) &&
156 (n->info & wt->info_mask) == wt->info_filter)
160 return false; /* If there is a filter, the default is to reject. */
164 * __post_watch_notification - Post an event notification
165 * @wlist: The watch list to post the event to.
166 * @n: The notification record to post.
167 * @cred: The creds of the process that triggered the notification.
168 * @id: The ID to match on the watch.
170 * Post a notification of an event into a set of watch queues and let the users
173 * The size of the notification should be set in n->info & WATCH_INFO_LENGTH and
174 * should be in units of sizeof(*n).
176 void __post_watch_notification(struct watch_list *wlist,
177 struct watch_notification *n,
178 const struct cred *cred,
181 const struct watch_filter *wf;
182 struct watch_queue *wqueue;
185 if (((n->info & WATCH_INFO_LENGTH) >> WATCH_INFO_LENGTH__SHIFT) == 0) {
192 hlist_for_each_entry_rcu(watch, &wlist->watchers, list_node) {
195 n->info &= ~WATCH_INFO_ID;
196 n->info |= watch->info_id;
198 wqueue = rcu_dereference(watch->queue);
199 wf = rcu_dereference(wqueue->filter);
200 if (wf && !filter_watch_notification(wf, n))
203 if (security_post_notification(watch->cred, cred, n) < 0)
206 post_one_notification(wqueue, n);
211 EXPORT_SYMBOL(__post_watch_notification);
214 * Allocate sufficient pages to preallocation for the requested number of
217 long watch_queue_set_size(struct pipe_inode_info *pipe, unsigned int nr_notes)
219 struct watch_queue *wqueue = pipe->watch_queue;
221 unsigned long *bitmap;
222 unsigned long user_bufs;
224 int ret, i, nr_pages;
232 nr_notes > 512) /* TODO: choose a better hard limit */
235 nr_pages = (nr_notes + WATCH_QUEUE_NOTES_PER_PAGE - 1);
236 nr_pages /= WATCH_QUEUE_NOTES_PER_PAGE;
237 user_bufs = account_pipe_buffers(pipe->user, pipe->nr_accounted, nr_pages);
239 if (nr_pages > pipe->max_usage &&
240 (too_many_pipe_buffers_hard(user_bufs) ||
241 too_many_pipe_buffers_soft(user_bufs)) &&
242 pipe_is_unprivileged_user()) {
247 nr_notes = nr_pages * WATCH_QUEUE_NOTES_PER_PAGE;
248 ret = pipe_resize_ring(pipe, roundup_pow_of_two(nr_notes));
252 pages = kcalloc(sizeof(struct page *), nr_pages, GFP_KERNEL);
256 for (i = 0; i < nr_pages; i++) {
257 pages[i] = alloc_page(GFP_KERNEL);
260 pages[i]->index = i * WATCH_QUEUE_NOTES_PER_PAGE;
263 bmsize = (nr_notes + BITS_PER_LONG - 1) / BITS_PER_LONG;
264 bmsize *= sizeof(unsigned long);
265 bitmap = kmalloc(bmsize, GFP_KERNEL);
269 memset(bitmap, 0xff, bmsize);
270 wqueue->notes = pages;
271 wqueue->notes_bitmap = bitmap;
272 wqueue->nr_pages = nr_pages;
273 wqueue->nr_notes = nr_notes;
278 __free_page(pages[i]);
281 (void) account_pipe_buffers(pipe->user, nr_pages, pipe->nr_accounted);
286 * Set the filter on a watch queue.
288 long watch_queue_set_filter(struct pipe_inode_info *pipe,
289 struct watch_notification_filter __user *_filter)
291 struct watch_notification_type_filter *tf;
292 struct watch_notification_filter filter;
293 struct watch_type_filter *q;
294 struct watch_filter *wfilter;
295 struct watch_queue *wqueue = pipe->watch_queue;
296 int ret, nr_filter = 0, i;
302 /* Remove the old filter */
307 /* Grab the user's filter specification */
308 if (copy_from_user(&filter, _filter, sizeof(filter)) != 0)
310 if (filter.nr_filters == 0 ||
311 filter.nr_filters > 16 ||
312 filter.__reserved != 0)
315 tf = memdup_user(_filter->filters, filter.nr_filters * sizeof(*tf));
320 for (i = 0; i < filter.nr_filters; i++) {
321 if ((tf[i].info_filter & ~tf[i].info_mask) ||
322 tf[i].info_mask & WATCH_INFO_LENGTH)
324 /* Ignore any unknown types */
325 if (tf[i].type >= WATCH_TYPE__NR)
330 /* Now we need to build the internal filter from only the relevant
331 * user-specified filters.
334 wfilter = kzalloc(struct_size(wfilter, filters, nr_filter), GFP_KERNEL);
337 wfilter->nr_filters = nr_filter;
339 q = wfilter->filters;
340 for (i = 0; i < filter.nr_filters; i++) {
341 if (tf[i].type >= WATCH_TYPE__NR)
344 q->type = tf[i].type;
345 q->info_filter = tf[i].info_filter;
346 q->info_mask = tf[i].info_mask;
347 q->subtype_filter[0] = tf[i].subtype_filter[0];
348 __set_bit(q->type, wfilter->type_filter);
355 wfilter = rcu_replace_pointer(wqueue->filter, wfilter,
356 lockdep_is_held(&pipe->mutex));
359 kfree_rcu(wfilter, rcu);
367 static void __put_watch_queue(struct kref *kref)
369 struct watch_queue *wqueue =
370 container_of(kref, struct watch_queue, usage);
371 struct watch_filter *wfilter;
374 for (i = 0; i < wqueue->nr_pages; i++)
375 __free_page(wqueue->notes[i]);
376 kfree(wqueue->notes);
377 bitmap_free(wqueue->notes_bitmap);
379 wfilter = rcu_access_pointer(wqueue->filter);
381 kfree_rcu(wfilter, rcu);
382 kfree_rcu(wqueue, rcu);
386 * put_watch_queue - Dispose of a ref on a watchqueue.
387 * @wqueue: The watch queue to unref.
389 void put_watch_queue(struct watch_queue *wqueue)
391 kref_put(&wqueue->usage, __put_watch_queue);
393 EXPORT_SYMBOL(put_watch_queue);
395 static void free_watch(struct rcu_head *rcu)
397 struct watch *watch = container_of(rcu, struct watch, rcu);
399 put_watch_queue(rcu_access_pointer(watch->queue));
400 atomic_dec(&watch->cred->user->nr_watches);
401 put_cred(watch->cred);
405 static void __put_watch(struct kref *kref)
407 struct watch *watch = container_of(kref, struct watch, usage);
409 call_rcu(&watch->rcu, free_watch);
415 static void put_watch(struct watch *watch)
417 kref_put(&watch->usage, __put_watch);
421 * init_watch - Initialise a watch
422 * @watch: The watch to initialise.
423 * @wqueue: The queue to assign.
425 * Initialise a watch and set the watch queue.
427 void init_watch(struct watch *watch, struct watch_queue *wqueue)
429 kref_init(&watch->usage);
430 INIT_HLIST_NODE(&watch->list_node);
431 INIT_HLIST_NODE(&watch->queue_node);
432 rcu_assign_pointer(watch->queue, wqueue);
436 * add_watch_to_object - Add a watch on an object to a watch list
437 * @watch: The watch to add
438 * @wlist: The watch list to add to
440 * @watch->queue must have been set to point to the queue to post notifications
441 * to and the watch list of the object to be watched. @watch->cred must also
442 * have been set to the appropriate credentials and a ref taken on them.
444 * The caller must pin the queue and the list both and must hold the list
445 * locked against racing watch additions/removals.
447 int add_watch_to_object(struct watch *watch, struct watch_list *wlist)
449 struct watch_queue *wqueue = rcu_access_pointer(watch->queue);
452 hlist_for_each_entry(w, &wlist->watchers, list_node) {
453 struct watch_queue *wq = rcu_access_pointer(w->queue);
454 if (wqueue == wq && watch->id == w->id)
458 watch->cred = get_current_cred();
459 rcu_assign_pointer(watch->watch_list, wlist);
461 if (atomic_inc_return(&watch->cred->user->nr_watches) >
462 task_rlimit(current, RLIMIT_NOFILE)) {
463 atomic_dec(&watch->cred->user->nr_watches);
464 put_cred(watch->cred);
468 spin_lock_bh(&wqueue->lock);
469 kref_get(&wqueue->usage);
470 kref_get(&watch->usage);
471 hlist_add_head(&watch->queue_node, &wqueue->watches);
472 spin_unlock_bh(&wqueue->lock);
474 hlist_add_head(&watch->list_node, &wlist->watchers);
477 EXPORT_SYMBOL(add_watch_to_object);
480 * remove_watch_from_object - Remove a watch or all watches from an object.
481 * @wlist: The watch list to remove from
482 * @wq: The watch queue of interest (ignored if @all is true)
483 * @id: The ID of the watch to remove (ignored if @all is true)
484 * @all: True to remove all objects
486 * Remove a specific watch or all watches from an object. A notification is
487 * sent to the watcher to tell them that this happened.
489 int remove_watch_from_object(struct watch_list *wlist, struct watch_queue *wq,
492 struct watch_notification_removal n;
493 struct watch_queue *wqueue;
500 spin_lock(&wlist->lock);
501 hlist_for_each_entry(watch, &wlist->watchers, list_node) {
503 (watch->id == id && rcu_access_pointer(watch->queue) == wq))
506 spin_unlock(&wlist->lock);
511 hlist_del_init_rcu(&watch->list_node);
512 rcu_assign_pointer(watch->watch_list, NULL);
513 spin_unlock(&wlist->lock);
515 /* We now own the reference on watch that used to belong to wlist. */
517 n.watch.type = WATCH_TYPE_META;
518 n.watch.subtype = WATCH_META_REMOVAL_NOTIFICATION;
519 n.watch.info = watch->info_id | watch_sizeof(n.watch);
522 n.watch.info = watch->info_id | watch_sizeof(n);
524 wqueue = rcu_dereference(watch->queue);
526 /* We don't need the watch list lock for the next bit as RCU is
527 * protecting *wqueue from deallocation.
530 post_one_notification(wqueue, &n.watch);
532 spin_lock_bh(&wqueue->lock);
534 if (!hlist_unhashed(&watch->queue_node)) {
535 hlist_del_init_rcu(&watch->queue_node);
539 spin_unlock_bh(&wqueue->lock);
542 if (wlist->release_watch) {
543 void (*release_watch)(struct watch *);
545 release_watch = wlist->release_watch;
547 (*release_watch)(watch);
552 if (all && !hlist_empty(&wlist->watchers))
558 EXPORT_SYMBOL(remove_watch_from_object);
561 * Remove all the watches that are contributory to a queue. This has the
562 * potential to race with removal of the watches by the destruction of the
563 * objects being watched or with the distribution of notifications.
565 void watch_queue_clear(struct watch_queue *wqueue)
567 struct watch_list *wlist;
572 spin_lock_bh(&wqueue->lock);
574 /* Prevent new notifications from being stored. */
575 wqueue->defunct = true;
577 while (!hlist_empty(&wqueue->watches)) {
578 watch = hlist_entry(wqueue->watches.first, struct watch, queue_node);
579 hlist_del_init_rcu(&watch->queue_node);
580 /* We now own a ref on the watch. */
581 spin_unlock_bh(&wqueue->lock);
583 /* We can't do the next bit under the queue lock as we need to
584 * get the list lock - which would cause a deadlock if someone
585 * was removing from the opposite direction at the same time or
586 * posting a notification.
588 wlist = rcu_dereference(watch->watch_list);
590 void (*release_watch)(struct watch *);
592 spin_lock(&wlist->lock);
594 release = !hlist_unhashed(&watch->list_node);
596 hlist_del_init_rcu(&watch->list_node);
597 rcu_assign_pointer(watch->watch_list, NULL);
599 /* We now own a second ref on the watch. */
602 release_watch = wlist->release_watch;
603 spin_unlock(&wlist->lock);
608 /* This might need to call dput(), so
609 * we have to drop all the locks.
611 (*release_watch)(watch);
619 spin_lock_bh(&wqueue->lock);
622 spin_unlock_bh(&wqueue->lock);
627 * get_watch_queue - Get a watch queue from its file descriptor.
628 * @fd: The fd to query.
630 struct watch_queue *get_watch_queue(int fd)
632 struct pipe_inode_info *pipe;
633 struct watch_queue *wqueue = ERR_PTR(-EINVAL);
638 pipe = get_pipe_info(f.file, false);
639 if (pipe && pipe->watch_queue) {
640 wqueue = pipe->watch_queue;
641 kref_get(&wqueue->usage);
648 EXPORT_SYMBOL(get_watch_queue);
651 * Initialise a watch queue
653 int watch_queue_init(struct pipe_inode_info *pipe)
655 struct watch_queue *wqueue;
657 wqueue = kzalloc(sizeof(*wqueue), GFP_KERNEL);
662 kref_init(&wqueue->usage);
663 spin_lock_init(&wqueue->lock);
664 INIT_HLIST_HEAD(&wqueue->watches);
666 pipe->watch_queue = wqueue;