Merge branch 'clk-imx' into clk-next
[linux-block.git] / kernel / watch_queue.c
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 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  */
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
53 static inline void unlock_wqueue(struct watch_queue *wqueue)
54 {
55         spin_unlock_bh(&wqueue->lock);
56 }
57
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  */
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  */
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  */
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  */
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         int ret, i, nr_pages;
249
250         if (!wqueue)
251                 return -ENODEV;
252         if (wqueue->notes)
253                 return -EBUSY;
254
255         if (nr_notes < 1 ||
256             nr_notes > 512) /* TODO: choose a better hard limit */
257                 return -EINVAL;
258
259         nr_pages = (nr_notes + WATCH_QUEUE_NOTES_PER_PAGE - 1);
260         nr_pages /= WATCH_QUEUE_NOTES_PER_PAGE;
261         user_bufs = account_pipe_buffers(pipe->user, pipe->nr_accounted, nr_pages);
262
263         if (nr_pages > pipe->max_usage &&
264             (too_many_pipe_buffers_hard(user_bufs) ||
265              too_many_pipe_buffers_soft(user_bufs)) &&
266             pipe_is_unprivileged_user()) {
267                 ret = -EPERM;
268                 goto error;
269         }
270
271         nr_notes = nr_pages * WATCH_QUEUE_NOTES_PER_PAGE;
272         ret = pipe_resize_ring(pipe, roundup_pow_of_two(nr_notes));
273         if (ret < 0)
274                 goto error;
275
276         ret = -ENOMEM;
277         pages = kcalloc(sizeof(struct page *), nr_pages, GFP_KERNEL);
278         if (!pages)
279                 goto error;
280
281         for (i = 0; i < nr_pages; i++) {
282                 pages[i] = alloc_page(GFP_KERNEL);
283                 if (!pages[i])
284                         goto error_p;
285                 pages[i]->index = i * WATCH_QUEUE_NOTES_PER_PAGE;
286         }
287
288         bitmap = bitmap_alloc(nr_notes, GFP_KERNEL);
289         if (!bitmap)
290                 goto error_p;
291
292         bitmap_fill(bitmap, nr_notes);
293         wqueue->notes = pages;
294         wqueue->notes_bitmap = bitmap;
295         wqueue->nr_pages = nr_pages;
296         wqueue->nr_notes = nr_notes;
297         return 0;
298
299 error_p:
300         while (--i >= 0)
301                 __free_page(pages[i]);
302         kfree(pages);
303 error:
304         (void) account_pipe_buffers(pipe->user, nr_pages, pipe->nr_accounted);
305         return ret;
306 }
307
308 /*
309  * Set the filter on a watch queue.
310  */
311 long watch_queue_set_filter(struct pipe_inode_info *pipe,
312                             struct watch_notification_filter __user *_filter)
313 {
314         struct watch_notification_type_filter *tf;
315         struct watch_notification_filter filter;
316         struct watch_type_filter *q;
317         struct watch_filter *wfilter;
318         struct watch_queue *wqueue = pipe->watch_queue;
319         int ret, nr_filter = 0, i;
320
321         if (!wqueue)
322                 return -ENODEV;
323
324         if (!_filter) {
325                 /* Remove the old filter */
326                 wfilter = NULL;
327                 goto set;
328         }
329
330         /* Grab the user's filter specification */
331         if (copy_from_user(&filter, _filter, sizeof(filter)) != 0)
332                 return -EFAULT;
333         if (filter.nr_filters == 0 ||
334             filter.nr_filters > 16 ||
335             filter.__reserved != 0)
336                 return -EINVAL;
337
338         tf = memdup_user(_filter->filters, filter.nr_filters * sizeof(*tf));
339         if (IS_ERR(tf))
340                 return PTR_ERR(tf);
341
342         ret = -EINVAL;
343         for (i = 0; i < filter.nr_filters; i++) {
344                 if ((tf[i].info_filter & ~tf[i].info_mask) ||
345                     tf[i].info_mask & WATCH_INFO_LENGTH)
346                         goto err_filter;
347                 /* Ignore any unknown types */
348                 if (tf[i].type >= WATCH_TYPE__NR)
349                         continue;
350                 nr_filter++;
351         }
352
353         /* Now we need to build the internal filter from only the relevant
354          * user-specified filters.
355          */
356         ret = -ENOMEM;
357         wfilter = kzalloc(struct_size(wfilter, filters, nr_filter), GFP_KERNEL);
358         if (!wfilter)
359                 goto err_filter;
360         wfilter->nr_filters = nr_filter;
361
362         q = wfilter->filters;
363         for (i = 0; i < filter.nr_filters; i++) {
364                 if (tf[i].type >= WATCH_TYPE__NR)
365                         continue;
366
367                 q->type                 = tf[i].type;
368                 q->info_filter          = tf[i].info_filter;
369                 q->info_mask            = tf[i].info_mask;
370                 q->subtype_filter[0]    = tf[i].subtype_filter[0];
371                 __set_bit(q->type, wfilter->type_filter);
372                 q++;
373         }
374
375         kfree(tf);
376 set:
377         pipe_lock(pipe);
378         wfilter = rcu_replace_pointer(wqueue->filter, wfilter,
379                                       lockdep_is_held(&pipe->mutex));
380         pipe_unlock(pipe);
381         if (wfilter)
382                 kfree_rcu(wfilter, rcu);
383         return 0;
384
385 err_filter:
386         kfree(tf);
387         return ret;
388 }
389
390 static void __put_watch_queue(struct kref *kref)
391 {
392         struct watch_queue *wqueue =
393                 container_of(kref, struct watch_queue, usage);
394         struct watch_filter *wfilter;
395         int i;
396
397         for (i = 0; i < wqueue->nr_pages; i++)
398                 __free_page(wqueue->notes[i]);
399         kfree(wqueue->notes);
400         bitmap_free(wqueue->notes_bitmap);
401
402         wfilter = rcu_access_pointer(wqueue->filter);
403         if (wfilter)
404                 kfree_rcu(wfilter, rcu);
405         kfree_rcu(wqueue, rcu);
406 }
407
408 /**
409  * put_watch_queue - Dispose of a ref on a watchqueue.
410  * @wqueue: The watch queue to unref.
411  */
412 void put_watch_queue(struct watch_queue *wqueue)
413 {
414         kref_put(&wqueue->usage, __put_watch_queue);
415 }
416 EXPORT_SYMBOL(put_watch_queue);
417
418 static void free_watch(struct rcu_head *rcu)
419 {
420         struct watch *watch = container_of(rcu, struct watch, rcu);
421
422         put_watch_queue(rcu_access_pointer(watch->queue));
423         atomic_dec(&watch->cred->user->nr_watches);
424         put_cred(watch->cred);
425         kfree(watch);
426 }
427
428 static void __put_watch(struct kref *kref)
429 {
430         struct watch *watch = container_of(kref, struct watch, usage);
431
432         call_rcu(&watch->rcu, free_watch);
433 }
434
435 /*
436  * Discard a watch.
437  */
438 static void put_watch(struct watch *watch)
439 {
440         kref_put(&watch->usage, __put_watch);
441 }
442
443 /**
444  * init_watch - Initialise a watch
445  * @watch: The watch to initialise.
446  * @wqueue: The queue to assign.
447  *
448  * Initialise a watch and set the watch queue.
449  */
450 void init_watch(struct watch *watch, struct watch_queue *wqueue)
451 {
452         kref_init(&watch->usage);
453         INIT_HLIST_NODE(&watch->list_node);
454         INIT_HLIST_NODE(&watch->queue_node);
455         rcu_assign_pointer(watch->queue, wqueue);
456 }
457
458 static int add_one_watch(struct watch *watch, struct watch_list *wlist, struct watch_queue *wqueue)
459 {
460         const struct cred *cred;
461         struct watch *w;
462
463         hlist_for_each_entry(w, &wlist->watchers, list_node) {
464                 struct watch_queue *wq = rcu_access_pointer(w->queue);
465                 if (wqueue == wq && watch->id == w->id)
466                         return -EBUSY;
467         }
468
469         cred = current_cred();
470         if (atomic_inc_return(&cred->user->nr_watches) > task_rlimit(current, RLIMIT_NOFILE)) {
471                 atomic_dec(&cred->user->nr_watches);
472                 return -EAGAIN;
473         }
474
475         watch->cred = get_cred(cred);
476         rcu_assign_pointer(watch->watch_list, wlist);
477
478         kref_get(&wqueue->usage);
479         kref_get(&watch->usage);
480         hlist_add_head(&watch->queue_node, &wqueue->watches);
481         hlist_add_head_rcu(&watch->list_node, &wlist->watchers);
482         return 0;
483 }
484
485 /**
486  * add_watch_to_object - Add a watch on an object to a watch list
487  * @watch: The watch to add
488  * @wlist: The watch list to add to
489  *
490  * @watch->queue must have been set to point to the queue to post notifications
491  * to and the watch list of the object to be watched.  @watch->cred must also
492  * have been set to the appropriate credentials and a ref taken on them.
493  *
494  * The caller must pin the queue and the list both and must hold the list
495  * locked against racing watch additions/removals.
496  */
497 int add_watch_to_object(struct watch *watch, struct watch_list *wlist)
498 {
499         struct watch_queue *wqueue;
500         int ret = -ENOENT;
501
502         rcu_read_lock();
503
504         wqueue = rcu_access_pointer(watch->queue);
505         if (lock_wqueue(wqueue)) {
506                 spin_lock(&wlist->lock);
507                 ret = add_one_watch(watch, wlist, wqueue);
508                 spin_unlock(&wlist->lock);
509                 unlock_wqueue(wqueue);
510         }
511
512         rcu_read_unlock();
513         return ret;
514 }
515 EXPORT_SYMBOL(add_watch_to_object);
516
517 /**
518  * remove_watch_from_object - Remove a watch or all watches from an object.
519  * @wlist: The watch list to remove from
520  * @wq: The watch queue of interest (ignored if @all is true)
521  * @id: The ID of the watch to remove (ignored if @all is true)
522  * @all: True to remove all objects
523  *
524  * Remove a specific watch or all watches from an object.  A notification is
525  * sent to the watcher to tell them that this happened.
526  */
527 int remove_watch_from_object(struct watch_list *wlist, struct watch_queue *wq,
528                              u64 id, bool all)
529 {
530         struct watch_notification_removal n;
531         struct watch_queue *wqueue;
532         struct watch *watch;
533         int ret = -EBADSLT;
534
535         rcu_read_lock();
536
537 again:
538         spin_lock(&wlist->lock);
539         hlist_for_each_entry(watch, &wlist->watchers, list_node) {
540                 if (all ||
541                     (watch->id == id && rcu_access_pointer(watch->queue) == wq))
542                         goto found;
543         }
544         spin_unlock(&wlist->lock);
545         goto out;
546
547 found:
548         ret = 0;
549         hlist_del_init_rcu(&watch->list_node);
550         rcu_assign_pointer(watch->watch_list, NULL);
551         spin_unlock(&wlist->lock);
552
553         /* We now own the reference on watch that used to belong to wlist. */
554
555         n.watch.type = WATCH_TYPE_META;
556         n.watch.subtype = WATCH_META_REMOVAL_NOTIFICATION;
557         n.watch.info = watch->info_id | watch_sizeof(n.watch);
558         n.id = id;
559         if (id != 0)
560                 n.watch.info = watch->info_id | watch_sizeof(n);
561
562         wqueue = rcu_dereference(watch->queue);
563
564         if (lock_wqueue(wqueue)) {
565                 post_one_notification(wqueue, &n.watch);
566
567                 if (!hlist_unhashed(&watch->queue_node)) {
568                         hlist_del_init_rcu(&watch->queue_node);
569                         put_watch(watch);
570                 }
571
572                 unlock_wqueue(wqueue);
573         }
574
575         if (wlist->release_watch) {
576                 void (*release_watch)(struct watch *);
577
578                 release_watch = wlist->release_watch;
579                 rcu_read_unlock();
580                 (*release_watch)(watch);
581                 rcu_read_lock();
582         }
583         put_watch(watch);
584
585         if (all && !hlist_empty(&wlist->watchers))
586                 goto again;
587 out:
588         rcu_read_unlock();
589         return ret;
590 }
591 EXPORT_SYMBOL(remove_watch_from_object);
592
593 /*
594  * Remove all the watches that are contributory to a queue.  This has the
595  * potential to race with removal of the watches by the destruction of the
596  * objects being watched or with the distribution of notifications.
597  */
598 void watch_queue_clear(struct watch_queue *wqueue)
599 {
600         struct watch_list *wlist;
601         struct watch *watch;
602         bool release;
603
604         rcu_read_lock();
605         spin_lock_bh(&wqueue->lock);
606
607         /* Prevent new notifications from being stored. */
608         wqueue->defunct = true;
609
610         while (!hlist_empty(&wqueue->watches)) {
611                 watch = hlist_entry(wqueue->watches.first, struct watch, queue_node);
612                 hlist_del_init_rcu(&watch->queue_node);
613                 /* We now own a ref on the watch. */
614                 spin_unlock_bh(&wqueue->lock);
615
616                 /* We can't do the next bit under the queue lock as we need to
617                  * get the list lock - which would cause a deadlock if someone
618                  * was removing from the opposite direction at the same time or
619                  * posting a notification.
620                  */
621                 wlist = rcu_dereference(watch->watch_list);
622                 if (wlist) {
623                         void (*release_watch)(struct watch *);
624
625                         spin_lock(&wlist->lock);
626
627                         release = !hlist_unhashed(&watch->list_node);
628                         if (release) {
629                                 hlist_del_init_rcu(&watch->list_node);
630                                 rcu_assign_pointer(watch->watch_list, NULL);
631
632                                 /* We now own a second ref on the watch. */
633                         }
634
635                         release_watch = wlist->release_watch;
636                         spin_unlock(&wlist->lock);
637
638                         if (release) {
639                                 if (release_watch) {
640                                         rcu_read_unlock();
641                                         /* This might need to call dput(), so
642                                          * we have to drop all the locks.
643                                          */
644                                         (*release_watch)(watch);
645                                         rcu_read_lock();
646                                 }
647                                 put_watch(watch);
648                         }
649                 }
650
651                 put_watch(watch);
652                 spin_lock_bh(&wqueue->lock);
653         }
654
655         spin_unlock_bh(&wqueue->lock);
656         rcu_read_unlock();
657 }
658
659 /**
660  * get_watch_queue - Get a watch queue from its file descriptor.
661  * @fd: The fd to query.
662  */
663 struct watch_queue *get_watch_queue(int fd)
664 {
665         struct pipe_inode_info *pipe;
666         struct watch_queue *wqueue = ERR_PTR(-EINVAL);
667         struct fd f;
668
669         f = fdget(fd);
670         if (f.file) {
671                 pipe = get_pipe_info(f.file, false);
672                 if (pipe && pipe->watch_queue) {
673                         wqueue = pipe->watch_queue;
674                         kref_get(&wqueue->usage);
675                 }
676                 fdput(f);
677         }
678
679         return wqueue;
680 }
681 EXPORT_SYMBOL(get_watch_queue);
682
683 /*
684  * Initialise a watch queue
685  */
686 int watch_queue_init(struct pipe_inode_info *pipe)
687 {
688         struct watch_queue *wqueue;
689
690         wqueue = kzalloc(sizeof(*wqueue), GFP_KERNEL);
691         if (!wqueue)
692                 return -ENOMEM;
693
694         wqueue->pipe = pipe;
695         kref_init(&wqueue->usage);
696         spin_lock_init(&wqueue->lock);
697         INIT_HLIST_HEAD(&wqueue->watches);
698
699         pipe->watch_queue = wqueue;
700         return 0;
701 }