Merge tag 'for-6.5-rc6-tag' of git://git.kernel.org/pub/scm/linux/kernel/git/kdave...
[platform/kernel/linux-rpi.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
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_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 }