Loading...
1/*
2 * fs/kernfs/file.c - kernfs file implementation
3 *
4 * Copyright (c) 2001-3 Patrick Mochel
5 * Copyright (c) 2007 SUSE Linux Products GmbH
6 * Copyright (c) 2007, 2013 Tejun Heo <tj@kernel.org>
7 *
8 * This file is released under the GPLv2.
9 */
10
11#include <linux/fs.h>
12#include <linux/seq_file.h>
13#include <linux/slab.h>
14#include <linux/poll.h>
15#include <linux/pagemap.h>
16#include <linux/sched/mm.h>
17#include <linux/fsnotify.h>
18
19#include "kernfs-internal.h"
20
21/*
22 * There's one kernfs_open_file for each open file and one kernfs_open_node
23 * for each kernfs_node with one or more open files.
24 *
25 * kernfs_node->attr.open points to kernfs_open_node. attr.open is
26 * protected by kernfs_open_node_lock.
27 *
28 * filp->private_data points to seq_file whose ->private points to
29 * kernfs_open_file. kernfs_open_files are chained at
30 * kernfs_open_node->files, which is protected by kernfs_open_file_mutex.
31 */
32static DEFINE_SPINLOCK(kernfs_open_node_lock);
33static DEFINE_MUTEX(kernfs_open_file_mutex);
34
35struct kernfs_open_node {
36 atomic_t refcnt;
37 atomic_t event;
38 wait_queue_head_t poll;
39 struct list_head files; /* goes through kernfs_open_file.list */
40};
41
42/*
43 * kernfs_notify() may be called from any context and bounces notifications
44 * through a work item. To minimize space overhead in kernfs_node, the
45 * pending queue is implemented as a singly linked list of kernfs_nodes.
46 * The list is terminated with the self pointer so that whether a
47 * kernfs_node is on the list or not can be determined by testing the next
48 * pointer for NULL.
49 */
50#define KERNFS_NOTIFY_EOL ((void *)&kernfs_notify_list)
51
52static DEFINE_SPINLOCK(kernfs_notify_lock);
53static struct kernfs_node *kernfs_notify_list = KERNFS_NOTIFY_EOL;
54
55static struct kernfs_open_file *kernfs_of(struct file *file)
56{
57 return ((struct seq_file *)file->private_data)->private;
58}
59
60/*
61 * Determine the kernfs_ops for the given kernfs_node. This function must
62 * be called while holding an active reference.
63 */
64static const struct kernfs_ops *kernfs_ops(struct kernfs_node *kn)
65{
66 if (kn->flags & KERNFS_LOCKDEP)
67 lockdep_assert_held(kn);
68 return kn->attr.ops;
69}
70
71/*
72 * As kernfs_seq_stop() is also called after kernfs_seq_start() or
73 * kernfs_seq_next() failure, it needs to distinguish whether it's stopping
74 * a seq_file iteration which is fully initialized with an active reference
75 * or an aborted kernfs_seq_start() due to get_active failure. The
76 * position pointer is the only context for each seq_file iteration and
77 * thus the stop condition should be encoded in it. As the return value is
78 * directly visible to userland, ERR_PTR(-ENODEV) is the only acceptable
79 * choice to indicate get_active failure.
80 *
81 * Unfortunately, this is complicated due to the optional custom seq_file
82 * operations which may return ERR_PTR(-ENODEV) too. kernfs_seq_stop()
83 * can't distinguish whether ERR_PTR(-ENODEV) is from get_active failure or
84 * custom seq_file operations and thus can't decide whether put_active
85 * should be performed or not only on ERR_PTR(-ENODEV).
86 *
87 * This is worked around by factoring out the custom seq_stop() and
88 * put_active part into kernfs_seq_stop_active(), skipping it from
89 * kernfs_seq_stop() if ERR_PTR(-ENODEV) while invoking it directly after
90 * custom seq_file operations fail with ERR_PTR(-ENODEV) - this ensures
91 * that kernfs_seq_stop_active() is skipped only after get_active failure.
92 */
93static void kernfs_seq_stop_active(struct seq_file *sf, void *v)
94{
95 struct kernfs_open_file *of = sf->private;
96 const struct kernfs_ops *ops = kernfs_ops(of->kn);
97
98 if (ops->seq_stop)
99 ops->seq_stop(sf, v);
100 kernfs_put_active(of->kn);
101}
102
103static void *kernfs_seq_start(struct seq_file *sf, loff_t *ppos)
104{
105 struct kernfs_open_file *of = sf->private;
106 const struct kernfs_ops *ops;
107
108 /*
109 * @of->mutex nests outside active ref and is primarily to ensure that
110 * the ops aren't called concurrently for the same open file.
111 */
112 mutex_lock(&of->mutex);
113 if (!kernfs_get_active(of->kn))
114 return ERR_PTR(-ENODEV);
115
116 ops = kernfs_ops(of->kn);
117 if (ops->seq_start) {
118 void *next = ops->seq_start(sf, ppos);
119 /* see the comment above kernfs_seq_stop_active() */
120 if (next == ERR_PTR(-ENODEV))
121 kernfs_seq_stop_active(sf, next);
122 return next;
123 } else {
124 /*
125 * The same behavior and code as single_open(). Returns
126 * !NULL if pos is at the beginning; otherwise, NULL.
127 */
128 return NULL + !*ppos;
129 }
130}
131
132static void *kernfs_seq_next(struct seq_file *sf, void *v, loff_t *ppos)
133{
134 struct kernfs_open_file *of = sf->private;
135 const struct kernfs_ops *ops = kernfs_ops(of->kn);
136
137 if (ops->seq_next) {
138 void *next = ops->seq_next(sf, v, ppos);
139 /* see the comment above kernfs_seq_stop_active() */
140 if (next == ERR_PTR(-ENODEV))
141 kernfs_seq_stop_active(sf, next);
142 return next;
143 } else {
144 /*
145 * The same behavior and code as single_open(), always
146 * terminate after the initial read.
147 */
148 ++*ppos;
149 return NULL;
150 }
151}
152
153static void kernfs_seq_stop(struct seq_file *sf, void *v)
154{
155 struct kernfs_open_file *of = sf->private;
156
157 if (v != ERR_PTR(-ENODEV))
158 kernfs_seq_stop_active(sf, v);
159 mutex_unlock(&of->mutex);
160}
161
162static int kernfs_seq_show(struct seq_file *sf, void *v)
163{
164 struct kernfs_open_file *of = sf->private;
165
166 of->event = atomic_read(&of->kn->attr.open->event);
167
168 return of->kn->attr.ops->seq_show(sf, v);
169}
170
171static const struct seq_operations kernfs_seq_ops = {
172 .start = kernfs_seq_start,
173 .next = kernfs_seq_next,
174 .stop = kernfs_seq_stop,
175 .show = kernfs_seq_show,
176};
177
178/*
179 * As reading a bin file can have side-effects, the exact offset and bytes
180 * specified in read(2) call should be passed to the read callback making
181 * it difficult to use seq_file. Implement simplistic custom buffering for
182 * bin files.
183 */
184static ssize_t kernfs_file_direct_read(struct kernfs_open_file *of,
185 char __user *user_buf, size_t count,
186 loff_t *ppos)
187{
188 ssize_t len = min_t(size_t, count, PAGE_SIZE);
189 const struct kernfs_ops *ops;
190 char *buf;
191
192 buf = of->prealloc_buf;
193 if (buf)
194 mutex_lock(&of->prealloc_mutex);
195 else
196 buf = kmalloc(len, GFP_KERNEL);
197 if (!buf)
198 return -ENOMEM;
199
200 /*
201 * @of->mutex nests outside active ref and is used both to ensure that
202 * the ops aren't called concurrently for the same open file.
203 */
204 mutex_lock(&of->mutex);
205 if (!kernfs_get_active(of->kn)) {
206 len = -ENODEV;
207 mutex_unlock(&of->mutex);
208 goto out_free;
209 }
210
211 of->event = atomic_read(&of->kn->attr.open->event);
212 ops = kernfs_ops(of->kn);
213 if (ops->read)
214 len = ops->read(of, buf, len, *ppos);
215 else
216 len = -EINVAL;
217
218 kernfs_put_active(of->kn);
219 mutex_unlock(&of->mutex);
220
221 if (len < 0)
222 goto out_free;
223
224 if (copy_to_user(user_buf, buf, len)) {
225 len = -EFAULT;
226 goto out_free;
227 }
228
229 *ppos += len;
230
231 out_free:
232 if (buf == of->prealloc_buf)
233 mutex_unlock(&of->prealloc_mutex);
234 else
235 kfree(buf);
236 return len;
237}
238
239/**
240 * kernfs_fop_read - kernfs vfs read callback
241 * @file: file pointer
242 * @user_buf: data to write
243 * @count: number of bytes
244 * @ppos: starting offset
245 */
246static ssize_t kernfs_fop_read(struct file *file, char __user *user_buf,
247 size_t count, loff_t *ppos)
248{
249 struct kernfs_open_file *of = kernfs_of(file);
250
251 if (of->kn->flags & KERNFS_HAS_SEQ_SHOW)
252 return seq_read(file, user_buf, count, ppos);
253 else
254 return kernfs_file_direct_read(of, user_buf, count, ppos);
255}
256
257/**
258 * kernfs_fop_write - kernfs vfs write callback
259 * @file: file pointer
260 * @user_buf: data to write
261 * @count: number of bytes
262 * @ppos: starting offset
263 *
264 * Copy data in from userland and pass it to the matching kernfs write
265 * operation.
266 *
267 * There is no easy way for us to know if userspace is only doing a partial
268 * write, so we don't support them. We expect the entire buffer to come on
269 * the first write. Hint: if you're writing a value, first read the file,
270 * modify only the the value you're changing, then write entire buffer
271 * back.
272 */
273static ssize_t kernfs_fop_write(struct file *file, const char __user *user_buf,
274 size_t count, loff_t *ppos)
275{
276 struct kernfs_open_file *of = kernfs_of(file);
277 const struct kernfs_ops *ops;
278 ssize_t len;
279 char *buf;
280
281 if (of->atomic_write_len) {
282 len = count;
283 if (len > of->atomic_write_len)
284 return -E2BIG;
285 } else {
286 len = min_t(size_t, count, PAGE_SIZE);
287 }
288
289 buf = of->prealloc_buf;
290 if (buf)
291 mutex_lock(&of->prealloc_mutex);
292 else
293 buf = kmalloc(len + 1, GFP_KERNEL);
294 if (!buf)
295 return -ENOMEM;
296
297 if (copy_from_user(buf, user_buf, len)) {
298 len = -EFAULT;
299 goto out_free;
300 }
301 buf[len] = '\0'; /* guarantee string termination */
302
303 /*
304 * @of->mutex nests outside active ref and is used both to ensure that
305 * the ops aren't called concurrently for the same open file.
306 */
307 mutex_lock(&of->mutex);
308 if (!kernfs_get_active(of->kn)) {
309 mutex_unlock(&of->mutex);
310 len = -ENODEV;
311 goto out_free;
312 }
313
314 ops = kernfs_ops(of->kn);
315 if (ops->write)
316 len = ops->write(of, buf, len, *ppos);
317 else
318 len = -EINVAL;
319
320 kernfs_put_active(of->kn);
321 mutex_unlock(&of->mutex);
322
323 if (len > 0)
324 *ppos += len;
325
326out_free:
327 if (buf == of->prealloc_buf)
328 mutex_unlock(&of->prealloc_mutex);
329 else
330 kfree(buf);
331 return len;
332}
333
334static void kernfs_vma_open(struct vm_area_struct *vma)
335{
336 struct file *file = vma->vm_file;
337 struct kernfs_open_file *of = kernfs_of(file);
338
339 if (!of->vm_ops)
340 return;
341
342 if (!kernfs_get_active(of->kn))
343 return;
344
345 if (of->vm_ops->open)
346 of->vm_ops->open(vma);
347
348 kernfs_put_active(of->kn);
349}
350
351static int kernfs_vma_fault(struct vm_fault *vmf)
352{
353 struct file *file = vmf->vma->vm_file;
354 struct kernfs_open_file *of = kernfs_of(file);
355 int ret;
356
357 if (!of->vm_ops)
358 return VM_FAULT_SIGBUS;
359
360 if (!kernfs_get_active(of->kn))
361 return VM_FAULT_SIGBUS;
362
363 ret = VM_FAULT_SIGBUS;
364 if (of->vm_ops->fault)
365 ret = of->vm_ops->fault(vmf);
366
367 kernfs_put_active(of->kn);
368 return ret;
369}
370
371static int kernfs_vma_page_mkwrite(struct vm_fault *vmf)
372{
373 struct file *file = vmf->vma->vm_file;
374 struct kernfs_open_file *of = kernfs_of(file);
375 int ret;
376
377 if (!of->vm_ops)
378 return VM_FAULT_SIGBUS;
379
380 if (!kernfs_get_active(of->kn))
381 return VM_FAULT_SIGBUS;
382
383 ret = 0;
384 if (of->vm_ops->page_mkwrite)
385 ret = of->vm_ops->page_mkwrite(vmf);
386 else
387 file_update_time(file);
388
389 kernfs_put_active(of->kn);
390 return ret;
391}
392
393static int kernfs_vma_access(struct vm_area_struct *vma, unsigned long addr,
394 void *buf, int len, int write)
395{
396 struct file *file = vma->vm_file;
397 struct kernfs_open_file *of = kernfs_of(file);
398 int ret;
399
400 if (!of->vm_ops)
401 return -EINVAL;
402
403 if (!kernfs_get_active(of->kn))
404 return -EINVAL;
405
406 ret = -EINVAL;
407 if (of->vm_ops->access)
408 ret = of->vm_ops->access(vma, addr, buf, len, write);
409
410 kernfs_put_active(of->kn);
411 return ret;
412}
413
414#ifdef CONFIG_NUMA
415static int kernfs_vma_set_policy(struct vm_area_struct *vma,
416 struct mempolicy *new)
417{
418 struct file *file = vma->vm_file;
419 struct kernfs_open_file *of = kernfs_of(file);
420 int ret;
421
422 if (!of->vm_ops)
423 return 0;
424
425 if (!kernfs_get_active(of->kn))
426 return -EINVAL;
427
428 ret = 0;
429 if (of->vm_ops->set_policy)
430 ret = of->vm_ops->set_policy(vma, new);
431
432 kernfs_put_active(of->kn);
433 return ret;
434}
435
436static struct mempolicy *kernfs_vma_get_policy(struct vm_area_struct *vma,
437 unsigned long addr)
438{
439 struct file *file = vma->vm_file;
440 struct kernfs_open_file *of = kernfs_of(file);
441 struct mempolicy *pol;
442
443 if (!of->vm_ops)
444 return vma->vm_policy;
445
446 if (!kernfs_get_active(of->kn))
447 return vma->vm_policy;
448
449 pol = vma->vm_policy;
450 if (of->vm_ops->get_policy)
451 pol = of->vm_ops->get_policy(vma, addr);
452
453 kernfs_put_active(of->kn);
454 return pol;
455}
456
457#endif
458
459static const struct vm_operations_struct kernfs_vm_ops = {
460 .open = kernfs_vma_open,
461 .fault = kernfs_vma_fault,
462 .page_mkwrite = kernfs_vma_page_mkwrite,
463 .access = kernfs_vma_access,
464#ifdef CONFIG_NUMA
465 .set_policy = kernfs_vma_set_policy,
466 .get_policy = kernfs_vma_get_policy,
467#endif
468};
469
470static int kernfs_fop_mmap(struct file *file, struct vm_area_struct *vma)
471{
472 struct kernfs_open_file *of = kernfs_of(file);
473 const struct kernfs_ops *ops;
474 int rc;
475
476 /*
477 * mmap path and of->mutex are prone to triggering spurious lockdep
478 * warnings and we don't want to add spurious locking dependency
479 * between the two. Check whether mmap is actually implemented
480 * without grabbing @of->mutex by testing HAS_MMAP flag. See the
481 * comment in kernfs_file_open() for more details.
482 */
483 if (!(of->kn->flags & KERNFS_HAS_MMAP))
484 return -ENODEV;
485
486 mutex_lock(&of->mutex);
487
488 rc = -ENODEV;
489 if (!kernfs_get_active(of->kn))
490 goto out_unlock;
491
492 ops = kernfs_ops(of->kn);
493 rc = ops->mmap(of, vma);
494 if (rc)
495 goto out_put;
496
497 /*
498 * PowerPC's pci_mmap of legacy_mem uses shmem_zero_setup()
499 * to satisfy versions of X which crash if the mmap fails: that
500 * substitutes a new vm_file, and we don't then want bin_vm_ops.
501 */
502 if (vma->vm_file != file)
503 goto out_put;
504
505 rc = -EINVAL;
506 if (of->mmapped && of->vm_ops != vma->vm_ops)
507 goto out_put;
508
509 /*
510 * It is not possible to successfully wrap close.
511 * So error if someone is trying to use close.
512 */
513 rc = -EINVAL;
514 if (vma->vm_ops && vma->vm_ops->close)
515 goto out_put;
516
517 rc = 0;
518 of->mmapped = true;
519 of->vm_ops = vma->vm_ops;
520 vma->vm_ops = &kernfs_vm_ops;
521out_put:
522 kernfs_put_active(of->kn);
523out_unlock:
524 mutex_unlock(&of->mutex);
525
526 return rc;
527}
528
529/**
530 * kernfs_get_open_node - get or create kernfs_open_node
531 * @kn: target kernfs_node
532 * @of: kernfs_open_file for this instance of open
533 *
534 * If @kn->attr.open exists, increment its reference count; otherwise,
535 * create one. @of is chained to the files list.
536 *
537 * LOCKING:
538 * Kernel thread context (may sleep).
539 *
540 * RETURNS:
541 * 0 on success, -errno on failure.
542 */
543static int kernfs_get_open_node(struct kernfs_node *kn,
544 struct kernfs_open_file *of)
545{
546 struct kernfs_open_node *on, *new_on = NULL;
547
548 retry:
549 mutex_lock(&kernfs_open_file_mutex);
550 spin_lock_irq(&kernfs_open_node_lock);
551
552 if (!kn->attr.open && new_on) {
553 kn->attr.open = new_on;
554 new_on = NULL;
555 }
556
557 on = kn->attr.open;
558 if (on) {
559 atomic_inc(&on->refcnt);
560 list_add_tail(&of->list, &on->files);
561 }
562
563 spin_unlock_irq(&kernfs_open_node_lock);
564 mutex_unlock(&kernfs_open_file_mutex);
565
566 if (on) {
567 kfree(new_on);
568 return 0;
569 }
570
571 /* not there, initialize a new one and retry */
572 new_on = kmalloc(sizeof(*new_on), GFP_KERNEL);
573 if (!new_on)
574 return -ENOMEM;
575
576 atomic_set(&new_on->refcnt, 0);
577 atomic_set(&new_on->event, 1);
578 init_waitqueue_head(&new_on->poll);
579 INIT_LIST_HEAD(&new_on->files);
580 goto retry;
581}
582
583/**
584 * kernfs_put_open_node - put kernfs_open_node
585 * @kn: target kernfs_nodet
586 * @of: associated kernfs_open_file
587 *
588 * Put @kn->attr.open and unlink @of from the files list. If
589 * reference count reaches zero, disassociate and free it.
590 *
591 * LOCKING:
592 * None.
593 */
594static void kernfs_put_open_node(struct kernfs_node *kn,
595 struct kernfs_open_file *of)
596{
597 struct kernfs_open_node *on = kn->attr.open;
598 unsigned long flags;
599
600 mutex_lock(&kernfs_open_file_mutex);
601 spin_lock_irqsave(&kernfs_open_node_lock, flags);
602
603 if (of)
604 list_del(&of->list);
605
606 if (atomic_dec_and_test(&on->refcnt))
607 kn->attr.open = NULL;
608 else
609 on = NULL;
610
611 spin_unlock_irqrestore(&kernfs_open_node_lock, flags);
612 mutex_unlock(&kernfs_open_file_mutex);
613
614 kfree(on);
615}
616
617static int kernfs_fop_open(struct inode *inode, struct file *file)
618{
619 struct kernfs_node *kn = inode->i_private;
620 struct kernfs_root *root = kernfs_root(kn);
621 const struct kernfs_ops *ops;
622 struct kernfs_open_file *of;
623 bool has_read, has_write, has_mmap;
624 int error = -EACCES;
625
626 if (!kernfs_get_active(kn))
627 return -ENODEV;
628
629 ops = kernfs_ops(kn);
630
631 has_read = ops->seq_show || ops->read || ops->mmap;
632 has_write = ops->write || ops->mmap;
633 has_mmap = ops->mmap;
634
635 /* see the flag definition for details */
636 if (root->flags & KERNFS_ROOT_EXTRA_OPEN_PERM_CHECK) {
637 if ((file->f_mode & FMODE_WRITE) &&
638 (!(inode->i_mode & S_IWUGO) || !has_write))
639 goto err_out;
640
641 if ((file->f_mode & FMODE_READ) &&
642 (!(inode->i_mode & S_IRUGO) || !has_read))
643 goto err_out;
644 }
645
646 /* allocate a kernfs_open_file for the file */
647 error = -ENOMEM;
648 of = kzalloc(sizeof(struct kernfs_open_file), GFP_KERNEL);
649 if (!of)
650 goto err_out;
651
652 /*
653 * The following is done to give a different lockdep key to
654 * @of->mutex for files which implement mmap. This is a rather
655 * crude way to avoid false positive lockdep warning around
656 * mm->mmap_sem - mmap nests @of->mutex under mm->mmap_sem and
657 * reading /sys/block/sda/trace/act_mask grabs sr_mutex, under
658 * which mm->mmap_sem nests, while holding @of->mutex. As each
659 * open file has a separate mutex, it's okay as long as those don't
660 * happen on the same file. At this point, we can't easily give
661 * each file a separate locking class. Let's differentiate on
662 * whether the file has mmap or not for now.
663 *
664 * Both paths of the branch look the same. They're supposed to
665 * look that way and give @of->mutex different static lockdep keys.
666 */
667 if (has_mmap)
668 mutex_init(&of->mutex);
669 else
670 mutex_init(&of->mutex);
671
672 of->kn = kn;
673 of->file = file;
674
675 /*
676 * Write path needs to atomic_write_len outside active reference.
677 * Cache it in open_file. See kernfs_fop_write() for details.
678 */
679 of->atomic_write_len = ops->atomic_write_len;
680
681 error = -EINVAL;
682 /*
683 * ->seq_show is incompatible with ->prealloc,
684 * as seq_read does its own allocation.
685 * ->read must be used instead.
686 */
687 if (ops->prealloc && ops->seq_show)
688 goto err_free;
689 if (ops->prealloc) {
690 int len = of->atomic_write_len ?: PAGE_SIZE;
691 of->prealloc_buf = kmalloc(len + 1, GFP_KERNEL);
692 error = -ENOMEM;
693 if (!of->prealloc_buf)
694 goto err_free;
695 mutex_init(&of->prealloc_mutex);
696 }
697
698 /*
699 * Always instantiate seq_file even if read access doesn't use
700 * seq_file or is not requested. This unifies private data access
701 * and readable regular files are the vast majority anyway.
702 */
703 if (ops->seq_show)
704 error = seq_open(file, &kernfs_seq_ops);
705 else
706 error = seq_open(file, NULL);
707 if (error)
708 goto err_free;
709
710 of->seq_file = file->private_data;
711 of->seq_file->private = of;
712
713 /* seq_file clears PWRITE unconditionally, restore it if WRITE */
714 if (file->f_mode & FMODE_WRITE)
715 file->f_mode |= FMODE_PWRITE;
716
717 /* make sure we have open node struct */
718 error = kernfs_get_open_node(kn, of);
719 if (error)
720 goto err_seq_release;
721
722 if (ops->open) {
723 /* nobody has access to @of yet, skip @of->mutex */
724 error = ops->open(of);
725 if (error)
726 goto err_put_node;
727 }
728
729 /* open succeeded, put active references */
730 kernfs_put_active(kn);
731 return 0;
732
733err_put_node:
734 kernfs_put_open_node(kn, of);
735err_seq_release:
736 seq_release(inode, file);
737err_free:
738 kfree(of->prealloc_buf);
739 kfree(of);
740err_out:
741 kernfs_put_active(kn);
742 return error;
743}
744
745/* used from release/drain to ensure that ->release() is called exactly once */
746static void kernfs_release_file(struct kernfs_node *kn,
747 struct kernfs_open_file *of)
748{
749 /*
750 * @of is guaranteed to have no other file operations in flight and
751 * we just want to synchronize release and drain paths.
752 * @kernfs_open_file_mutex is enough. @of->mutex can't be used
753 * here because drain path may be called from places which can
754 * cause circular dependency.
755 */
756 lockdep_assert_held(&kernfs_open_file_mutex);
757
758 if (!of->released) {
759 /*
760 * A file is never detached without being released and we
761 * need to be able to release files which are deactivated
762 * and being drained. Don't use kernfs_ops().
763 */
764 kn->attr.ops->release(of);
765 of->released = true;
766 }
767}
768
769static int kernfs_fop_release(struct inode *inode, struct file *filp)
770{
771 struct kernfs_node *kn = inode->i_private;
772 struct kernfs_open_file *of = kernfs_of(filp);
773
774 if (kn->flags & KERNFS_HAS_RELEASE) {
775 mutex_lock(&kernfs_open_file_mutex);
776 kernfs_release_file(kn, of);
777 mutex_unlock(&kernfs_open_file_mutex);
778 }
779
780 kernfs_put_open_node(kn, of);
781 seq_release(inode, filp);
782 kfree(of->prealloc_buf);
783 kfree(of);
784
785 return 0;
786}
787
788void kernfs_drain_open_files(struct kernfs_node *kn)
789{
790 struct kernfs_open_node *on;
791 struct kernfs_open_file *of;
792
793 if (!(kn->flags & (KERNFS_HAS_MMAP | KERNFS_HAS_RELEASE)))
794 return;
795
796 spin_lock_irq(&kernfs_open_node_lock);
797 on = kn->attr.open;
798 if (on)
799 atomic_inc(&on->refcnt);
800 spin_unlock_irq(&kernfs_open_node_lock);
801 if (!on)
802 return;
803
804 mutex_lock(&kernfs_open_file_mutex);
805
806 list_for_each_entry(of, &on->files, list) {
807 struct inode *inode = file_inode(of->file);
808
809 if (kn->flags & KERNFS_HAS_MMAP)
810 unmap_mapping_range(inode->i_mapping, 0, 0, 1);
811
812 if (kn->flags & KERNFS_HAS_RELEASE)
813 kernfs_release_file(kn, of);
814 }
815
816 mutex_unlock(&kernfs_open_file_mutex);
817
818 kernfs_put_open_node(kn, NULL);
819}
820
821/*
822 * Kernfs attribute files are pollable. The idea is that you read
823 * the content and then you use 'poll' or 'select' to wait for
824 * the content to change. When the content changes (assuming the
825 * manager for the kobject supports notification), poll will
826 * return EPOLLERR|EPOLLPRI, and select will return the fd whether
827 * it is waiting for read, write, or exceptions.
828 * Once poll/select indicates that the value has changed, you
829 * need to close and re-open the file, or seek to 0 and read again.
830 * Reminder: this only works for attributes which actively support
831 * it, and it is not possible to test an attribute from userspace
832 * to see if it supports poll (Neither 'poll' nor 'select' return
833 * an appropriate error code). When in doubt, set a suitable timeout value.
834 */
835static __poll_t kernfs_fop_poll(struct file *filp, poll_table *wait)
836{
837 struct kernfs_open_file *of = kernfs_of(filp);
838 struct kernfs_node *kn = kernfs_dentry_node(filp->f_path.dentry);
839 struct kernfs_open_node *on = kn->attr.open;
840
841 if (!kernfs_get_active(kn))
842 goto trigger;
843
844 poll_wait(filp, &on->poll, wait);
845
846 kernfs_put_active(kn);
847
848 if (of->event != atomic_read(&on->event))
849 goto trigger;
850
851 return DEFAULT_POLLMASK;
852
853 trigger:
854 return DEFAULT_POLLMASK|EPOLLERR|EPOLLPRI;
855}
856
857static void kernfs_notify_workfn(struct work_struct *work)
858{
859 struct kernfs_node *kn;
860 struct kernfs_open_node *on;
861 struct kernfs_super_info *info;
862repeat:
863 /* pop one off the notify_list */
864 spin_lock_irq(&kernfs_notify_lock);
865 kn = kernfs_notify_list;
866 if (kn == KERNFS_NOTIFY_EOL) {
867 spin_unlock_irq(&kernfs_notify_lock);
868 return;
869 }
870 kernfs_notify_list = kn->attr.notify_next;
871 kn->attr.notify_next = NULL;
872 spin_unlock_irq(&kernfs_notify_lock);
873
874 /* kick poll */
875 spin_lock_irq(&kernfs_open_node_lock);
876
877 on = kn->attr.open;
878 if (on) {
879 atomic_inc(&on->event);
880 wake_up_interruptible(&on->poll);
881 }
882
883 spin_unlock_irq(&kernfs_open_node_lock);
884
885 /* kick fsnotify */
886 mutex_lock(&kernfs_mutex);
887
888 list_for_each_entry(info, &kernfs_root(kn)->supers, node) {
889 struct kernfs_node *parent;
890 struct inode *inode;
891
892 /*
893 * We want fsnotify_modify() on @kn but as the
894 * modifications aren't originating from userland don't
895 * have the matching @file available. Look up the inodes
896 * and generate the events manually.
897 */
898 inode = ilookup(info->sb, kn->id.ino);
899 if (!inode)
900 continue;
901
902 parent = kernfs_get_parent(kn);
903 if (parent) {
904 struct inode *p_inode;
905
906 p_inode = ilookup(info->sb, parent->id.ino);
907 if (p_inode) {
908 fsnotify(p_inode, FS_MODIFY | FS_EVENT_ON_CHILD,
909 inode, FSNOTIFY_EVENT_INODE, kn->name, 0);
910 iput(p_inode);
911 }
912
913 kernfs_put(parent);
914 }
915
916 fsnotify(inode, FS_MODIFY, inode, FSNOTIFY_EVENT_INODE,
917 kn->name, 0);
918 iput(inode);
919 }
920
921 mutex_unlock(&kernfs_mutex);
922 kernfs_put(kn);
923 goto repeat;
924}
925
926/**
927 * kernfs_notify - notify a kernfs file
928 * @kn: file to notify
929 *
930 * Notify @kn such that poll(2) on @kn wakes up. Maybe be called from any
931 * context.
932 */
933void kernfs_notify(struct kernfs_node *kn)
934{
935 static DECLARE_WORK(kernfs_notify_work, kernfs_notify_workfn);
936 unsigned long flags;
937
938 if (WARN_ON(kernfs_type(kn) != KERNFS_FILE))
939 return;
940
941 spin_lock_irqsave(&kernfs_notify_lock, flags);
942 if (!kn->attr.notify_next) {
943 kernfs_get(kn);
944 kn->attr.notify_next = kernfs_notify_list;
945 kernfs_notify_list = kn;
946 schedule_work(&kernfs_notify_work);
947 }
948 spin_unlock_irqrestore(&kernfs_notify_lock, flags);
949}
950EXPORT_SYMBOL_GPL(kernfs_notify);
951
952const struct file_operations kernfs_file_fops = {
953 .read = kernfs_fop_read,
954 .write = kernfs_fop_write,
955 .llseek = generic_file_llseek,
956 .mmap = kernfs_fop_mmap,
957 .open = kernfs_fop_open,
958 .release = kernfs_fop_release,
959 .poll = kernfs_fop_poll,
960 .fsync = noop_fsync,
961};
962
963/**
964 * __kernfs_create_file - kernfs internal function to create a file
965 * @parent: directory to create the file in
966 * @name: name of the file
967 * @mode: mode of the file
968 * @size: size of the file
969 * @ops: kernfs operations for the file
970 * @priv: private data for the file
971 * @ns: optional namespace tag of the file
972 * @key: lockdep key for the file's active_ref, %NULL to disable lockdep
973 *
974 * Returns the created node on success, ERR_PTR() value on error.
975 */
976struct kernfs_node *__kernfs_create_file(struct kernfs_node *parent,
977 const char *name,
978 umode_t mode, loff_t size,
979 const struct kernfs_ops *ops,
980 void *priv, const void *ns,
981 struct lock_class_key *key)
982{
983 struct kernfs_node *kn;
984 unsigned flags;
985 int rc;
986
987 flags = KERNFS_FILE;
988
989 kn = kernfs_new_node(parent, name, (mode & S_IALLUGO) | S_IFREG, flags);
990 if (!kn)
991 return ERR_PTR(-ENOMEM);
992
993 kn->attr.ops = ops;
994 kn->attr.size = size;
995 kn->ns = ns;
996 kn->priv = priv;
997
998#ifdef CONFIG_DEBUG_LOCK_ALLOC
999 if (key) {
1000 lockdep_init_map(&kn->dep_map, "kn->count", key, 0);
1001 kn->flags |= KERNFS_LOCKDEP;
1002 }
1003#endif
1004
1005 /*
1006 * kn->attr.ops is accesible only while holding active ref. We
1007 * need to know whether some ops are implemented outside active
1008 * ref. Cache their existence in flags.
1009 */
1010 if (ops->seq_show)
1011 kn->flags |= KERNFS_HAS_SEQ_SHOW;
1012 if (ops->mmap)
1013 kn->flags |= KERNFS_HAS_MMAP;
1014 if (ops->release)
1015 kn->flags |= KERNFS_HAS_RELEASE;
1016
1017 rc = kernfs_add_one(kn);
1018 if (rc) {
1019 kernfs_put(kn);
1020 return ERR_PTR(rc);
1021 }
1022 return kn;
1023}
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * fs/kernfs/file.c - kernfs file implementation
4 *
5 * Copyright (c) 2001-3 Patrick Mochel
6 * Copyright (c) 2007 SUSE Linux Products GmbH
7 * Copyright (c) 2007, 2013 Tejun Heo <tj@kernel.org>
8 */
9
10#include <linux/fs.h>
11#include <linux/seq_file.h>
12#include <linux/slab.h>
13#include <linux/poll.h>
14#include <linux/pagemap.h>
15#include <linux/sched/mm.h>
16#include <linux/fsnotify.h>
17#include <linux/uio.h>
18
19#include "kernfs-internal.h"
20
21struct kernfs_open_node {
22 struct rcu_head rcu_head;
23 atomic_t event;
24 wait_queue_head_t poll;
25 struct list_head files; /* goes through kernfs_open_file.list */
26 unsigned int nr_mmapped;
27 unsigned int nr_to_release;
28};
29
30/*
31 * kernfs_notify() may be called from any context and bounces notifications
32 * through a work item. To minimize space overhead in kernfs_node, the
33 * pending queue is implemented as a singly linked list of kernfs_nodes.
34 * The list is terminated with the self pointer so that whether a
35 * kernfs_node is on the list or not can be determined by testing the next
36 * pointer for %NULL.
37 */
38#define KERNFS_NOTIFY_EOL ((void *)&kernfs_notify_list)
39
40static DEFINE_SPINLOCK(kernfs_notify_lock);
41static struct kernfs_node *kernfs_notify_list = KERNFS_NOTIFY_EOL;
42
43static inline struct mutex *kernfs_open_file_mutex_ptr(struct kernfs_node *kn)
44{
45 int idx = hash_ptr(kn, NR_KERNFS_LOCK_BITS);
46
47 return &kernfs_locks->open_file_mutex[idx];
48}
49
50static inline struct mutex *kernfs_open_file_mutex_lock(struct kernfs_node *kn)
51{
52 struct mutex *lock;
53
54 lock = kernfs_open_file_mutex_ptr(kn);
55
56 mutex_lock(lock);
57
58 return lock;
59}
60
61/**
62 * of_on - Get the kernfs_open_node of the specified kernfs_open_file
63 * @of: target kernfs_open_file
64 *
65 * Return: the kernfs_open_node of the kernfs_open_file
66 */
67static struct kernfs_open_node *of_on(struct kernfs_open_file *of)
68{
69 return rcu_dereference_protected(of->kn->attr.open,
70 !list_empty(&of->list));
71}
72
73/**
74 * kernfs_deref_open_node_locked - Get kernfs_open_node corresponding to @kn
75 *
76 * @kn: target kernfs_node.
77 *
78 * Fetch and return ->attr.open of @kn when caller holds the
79 * kernfs_open_file_mutex_ptr(kn).
80 *
81 * Update of ->attr.open happens under kernfs_open_file_mutex_ptr(kn). So when
82 * the caller guarantees that this mutex is being held, other updaters can't
83 * change ->attr.open and this means that we can safely deref ->attr.open
84 * outside RCU read-side critical section.
85 *
86 * The caller needs to make sure that kernfs_open_file_mutex is held.
87 *
88 * Return: @kn->attr.open when kernfs_open_file_mutex is held.
89 */
90static struct kernfs_open_node *
91kernfs_deref_open_node_locked(struct kernfs_node *kn)
92{
93 return rcu_dereference_protected(kn->attr.open,
94 lockdep_is_held(kernfs_open_file_mutex_ptr(kn)));
95}
96
97static struct kernfs_open_file *kernfs_of(struct file *file)
98{
99 return ((struct seq_file *)file->private_data)->private;
100}
101
102/*
103 * Determine the kernfs_ops for the given kernfs_node. This function must
104 * be called while holding an active reference.
105 */
106static const struct kernfs_ops *kernfs_ops(struct kernfs_node *kn)
107{
108 if (kn->flags & KERNFS_LOCKDEP)
109 lockdep_assert_held(kn);
110 return kn->attr.ops;
111}
112
113/*
114 * As kernfs_seq_stop() is also called after kernfs_seq_start() or
115 * kernfs_seq_next() failure, it needs to distinguish whether it's stopping
116 * a seq_file iteration which is fully initialized with an active reference
117 * or an aborted kernfs_seq_start() due to get_active failure. The
118 * position pointer is the only context for each seq_file iteration and
119 * thus the stop condition should be encoded in it. As the return value is
120 * directly visible to userland, ERR_PTR(-ENODEV) is the only acceptable
121 * choice to indicate get_active failure.
122 *
123 * Unfortunately, this is complicated due to the optional custom seq_file
124 * operations which may return ERR_PTR(-ENODEV) too. kernfs_seq_stop()
125 * can't distinguish whether ERR_PTR(-ENODEV) is from get_active failure or
126 * custom seq_file operations and thus can't decide whether put_active
127 * should be performed or not only on ERR_PTR(-ENODEV).
128 *
129 * This is worked around by factoring out the custom seq_stop() and
130 * put_active part into kernfs_seq_stop_active(), skipping it from
131 * kernfs_seq_stop() if ERR_PTR(-ENODEV) while invoking it directly after
132 * custom seq_file operations fail with ERR_PTR(-ENODEV) - this ensures
133 * that kernfs_seq_stop_active() is skipped only after get_active failure.
134 */
135static void kernfs_seq_stop_active(struct seq_file *sf, void *v)
136{
137 struct kernfs_open_file *of = sf->private;
138 const struct kernfs_ops *ops = kernfs_ops(of->kn);
139
140 if (ops->seq_stop)
141 ops->seq_stop(sf, v);
142 kernfs_put_active(of->kn);
143}
144
145static void *kernfs_seq_start(struct seq_file *sf, loff_t *ppos)
146{
147 struct kernfs_open_file *of = sf->private;
148 const struct kernfs_ops *ops;
149
150 /*
151 * @of->mutex nests outside active ref and is primarily to ensure that
152 * the ops aren't called concurrently for the same open file.
153 */
154 mutex_lock(&of->mutex);
155 if (!kernfs_get_active(of->kn))
156 return ERR_PTR(-ENODEV);
157
158 ops = kernfs_ops(of->kn);
159 if (ops->seq_start) {
160 void *next = ops->seq_start(sf, ppos);
161 /* see the comment above kernfs_seq_stop_active() */
162 if (next == ERR_PTR(-ENODEV))
163 kernfs_seq_stop_active(sf, next);
164 return next;
165 }
166 return single_start(sf, ppos);
167}
168
169static void *kernfs_seq_next(struct seq_file *sf, void *v, loff_t *ppos)
170{
171 struct kernfs_open_file *of = sf->private;
172 const struct kernfs_ops *ops = kernfs_ops(of->kn);
173
174 if (ops->seq_next) {
175 void *next = ops->seq_next(sf, v, ppos);
176 /* see the comment above kernfs_seq_stop_active() */
177 if (next == ERR_PTR(-ENODEV))
178 kernfs_seq_stop_active(sf, next);
179 return next;
180 } else {
181 /*
182 * The same behavior and code as single_open(), always
183 * terminate after the initial read.
184 */
185 ++*ppos;
186 return NULL;
187 }
188}
189
190static void kernfs_seq_stop(struct seq_file *sf, void *v)
191{
192 struct kernfs_open_file *of = sf->private;
193
194 if (v != ERR_PTR(-ENODEV))
195 kernfs_seq_stop_active(sf, v);
196 mutex_unlock(&of->mutex);
197}
198
199static int kernfs_seq_show(struct seq_file *sf, void *v)
200{
201 struct kernfs_open_file *of = sf->private;
202
203 of->event = atomic_read(&of_on(of)->event);
204
205 return of->kn->attr.ops->seq_show(sf, v);
206}
207
208static const struct seq_operations kernfs_seq_ops = {
209 .start = kernfs_seq_start,
210 .next = kernfs_seq_next,
211 .stop = kernfs_seq_stop,
212 .show = kernfs_seq_show,
213};
214
215/*
216 * As reading a bin file can have side-effects, the exact offset and bytes
217 * specified in read(2) call should be passed to the read callback making
218 * it difficult to use seq_file. Implement simplistic custom buffering for
219 * bin files.
220 */
221static ssize_t kernfs_file_read_iter(struct kiocb *iocb, struct iov_iter *iter)
222{
223 struct kernfs_open_file *of = kernfs_of(iocb->ki_filp);
224 ssize_t len = min_t(size_t, iov_iter_count(iter), PAGE_SIZE);
225 const struct kernfs_ops *ops;
226 char *buf;
227
228 buf = of->prealloc_buf;
229 if (buf)
230 mutex_lock(&of->prealloc_mutex);
231 else
232 buf = kmalloc(len, GFP_KERNEL);
233 if (!buf)
234 return -ENOMEM;
235
236 /*
237 * @of->mutex nests outside active ref and is used both to ensure that
238 * the ops aren't called concurrently for the same open file.
239 */
240 mutex_lock(&of->mutex);
241 if (!kernfs_get_active(of->kn)) {
242 len = -ENODEV;
243 mutex_unlock(&of->mutex);
244 goto out_free;
245 }
246
247 of->event = atomic_read(&of_on(of)->event);
248
249 ops = kernfs_ops(of->kn);
250 if (ops->read)
251 len = ops->read(of, buf, len, iocb->ki_pos);
252 else
253 len = -EINVAL;
254
255 kernfs_put_active(of->kn);
256 mutex_unlock(&of->mutex);
257
258 if (len < 0)
259 goto out_free;
260
261 if (copy_to_iter(buf, len, iter) != len) {
262 len = -EFAULT;
263 goto out_free;
264 }
265
266 iocb->ki_pos += len;
267
268 out_free:
269 if (buf == of->prealloc_buf)
270 mutex_unlock(&of->prealloc_mutex);
271 else
272 kfree(buf);
273 return len;
274}
275
276static ssize_t kernfs_fop_read_iter(struct kiocb *iocb, struct iov_iter *iter)
277{
278 if (kernfs_of(iocb->ki_filp)->kn->flags & KERNFS_HAS_SEQ_SHOW)
279 return seq_read_iter(iocb, iter);
280 return kernfs_file_read_iter(iocb, iter);
281}
282
283/*
284 * Copy data in from userland and pass it to the matching kernfs write
285 * operation.
286 *
287 * There is no easy way for us to know if userspace is only doing a partial
288 * write, so we don't support them. We expect the entire buffer to come on
289 * the first write. Hint: if you're writing a value, first read the file,
290 * modify only the value you're changing, then write entire buffer
291 * back.
292 */
293static ssize_t kernfs_fop_write_iter(struct kiocb *iocb, struct iov_iter *iter)
294{
295 struct kernfs_open_file *of = kernfs_of(iocb->ki_filp);
296 ssize_t len = iov_iter_count(iter);
297 const struct kernfs_ops *ops;
298 char *buf;
299
300 if (of->atomic_write_len) {
301 if (len > of->atomic_write_len)
302 return -E2BIG;
303 } else {
304 len = min_t(size_t, len, PAGE_SIZE);
305 }
306
307 buf = of->prealloc_buf;
308 if (buf)
309 mutex_lock(&of->prealloc_mutex);
310 else
311 buf = kmalloc(len + 1, GFP_KERNEL);
312 if (!buf)
313 return -ENOMEM;
314
315 if (copy_from_iter(buf, len, iter) != len) {
316 len = -EFAULT;
317 goto out_free;
318 }
319 buf[len] = '\0'; /* guarantee string termination */
320
321 /*
322 * @of->mutex nests outside active ref and is used both to ensure that
323 * the ops aren't called concurrently for the same open file.
324 */
325 mutex_lock(&of->mutex);
326 if (!kernfs_get_active(of->kn)) {
327 mutex_unlock(&of->mutex);
328 len = -ENODEV;
329 goto out_free;
330 }
331
332 ops = kernfs_ops(of->kn);
333 if (ops->write)
334 len = ops->write(of, buf, len, iocb->ki_pos);
335 else
336 len = -EINVAL;
337
338 kernfs_put_active(of->kn);
339 mutex_unlock(&of->mutex);
340
341 if (len > 0)
342 iocb->ki_pos += len;
343
344out_free:
345 if (buf == of->prealloc_buf)
346 mutex_unlock(&of->prealloc_mutex);
347 else
348 kfree(buf);
349 return len;
350}
351
352static void kernfs_vma_open(struct vm_area_struct *vma)
353{
354 struct file *file = vma->vm_file;
355 struct kernfs_open_file *of = kernfs_of(file);
356
357 if (!of->vm_ops)
358 return;
359
360 if (!kernfs_get_active(of->kn))
361 return;
362
363 if (of->vm_ops->open)
364 of->vm_ops->open(vma);
365
366 kernfs_put_active(of->kn);
367}
368
369static vm_fault_t kernfs_vma_fault(struct vm_fault *vmf)
370{
371 struct file *file = vmf->vma->vm_file;
372 struct kernfs_open_file *of = kernfs_of(file);
373 vm_fault_t ret;
374
375 if (!of->vm_ops)
376 return VM_FAULT_SIGBUS;
377
378 if (!kernfs_get_active(of->kn))
379 return VM_FAULT_SIGBUS;
380
381 ret = VM_FAULT_SIGBUS;
382 if (of->vm_ops->fault)
383 ret = of->vm_ops->fault(vmf);
384
385 kernfs_put_active(of->kn);
386 return ret;
387}
388
389static vm_fault_t kernfs_vma_page_mkwrite(struct vm_fault *vmf)
390{
391 struct file *file = vmf->vma->vm_file;
392 struct kernfs_open_file *of = kernfs_of(file);
393 vm_fault_t ret;
394
395 if (!of->vm_ops)
396 return VM_FAULT_SIGBUS;
397
398 if (!kernfs_get_active(of->kn))
399 return VM_FAULT_SIGBUS;
400
401 ret = 0;
402 if (of->vm_ops->page_mkwrite)
403 ret = of->vm_ops->page_mkwrite(vmf);
404 else
405 file_update_time(file);
406
407 kernfs_put_active(of->kn);
408 return ret;
409}
410
411static int kernfs_vma_access(struct vm_area_struct *vma, unsigned long addr,
412 void *buf, int len, int write)
413{
414 struct file *file = vma->vm_file;
415 struct kernfs_open_file *of = kernfs_of(file);
416 int ret;
417
418 if (!of->vm_ops)
419 return -EINVAL;
420
421 if (!kernfs_get_active(of->kn))
422 return -EINVAL;
423
424 ret = -EINVAL;
425 if (of->vm_ops->access)
426 ret = of->vm_ops->access(vma, addr, buf, len, write);
427
428 kernfs_put_active(of->kn);
429 return ret;
430}
431
432#ifdef CONFIG_NUMA
433static int kernfs_vma_set_policy(struct vm_area_struct *vma,
434 struct mempolicy *new)
435{
436 struct file *file = vma->vm_file;
437 struct kernfs_open_file *of = kernfs_of(file);
438 int ret;
439
440 if (!of->vm_ops)
441 return 0;
442
443 if (!kernfs_get_active(of->kn))
444 return -EINVAL;
445
446 ret = 0;
447 if (of->vm_ops->set_policy)
448 ret = of->vm_ops->set_policy(vma, new);
449
450 kernfs_put_active(of->kn);
451 return ret;
452}
453
454static struct mempolicy *kernfs_vma_get_policy(struct vm_area_struct *vma,
455 unsigned long addr)
456{
457 struct file *file = vma->vm_file;
458 struct kernfs_open_file *of = kernfs_of(file);
459 struct mempolicy *pol;
460
461 if (!of->vm_ops)
462 return vma->vm_policy;
463
464 if (!kernfs_get_active(of->kn))
465 return vma->vm_policy;
466
467 pol = vma->vm_policy;
468 if (of->vm_ops->get_policy)
469 pol = of->vm_ops->get_policy(vma, addr);
470
471 kernfs_put_active(of->kn);
472 return pol;
473}
474
475#endif
476
477static const struct vm_operations_struct kernfs_vm_ops = {
478 .open = kernfs_vma_open,
479 .fault = kernfs_vma_fault,
480 .page_mkwrite = kernfs_vma_page_mkwrite,
481 .access = kernfs_vma_access,
482#ifdef CONFIG_NUMA
483 .set_policy = kernfs_vma_set_policy,
484 .get_policy = kernfs_vma_get_policy,
485#endif
486};
487
488static int kernfs_fop_mmap(struct file *file, struct vm_area_struct *vma)
489{
490 struct kernfs_open_file *of = kernfs_of(file);
491 const struct kernfs_ops *ops;
492 int rc;
493
494 /*
495 * mmap path and of->mutex are prone to triggering spurious lockdep
496 * warnings and we don't want to add spurious locking dependency
497 * between the two. Check whether mmap is actually implemented
498 * without grabbing @of->mutex by testing HAS_MMAP flag. See the
499 * comment in kernfs_file_open() for more details.
500 */
501 if (!(of->kn->flags & KERNFS_HAS_MMAP))
502 return -ENODEV;
503
504 mutex_lock(&of->mutex);
505
506 rc = -ENODEV;
507 if (!kernfs_get_active(of->kn))
508 goto out_unlock;
509
510 ops = kernfs_ops(of->kn);
511 rc = ops->mmap(of, vma);
512 if (rc)
513 goto out_put;
514
515 /*
516 * PowerPC's pci_mmap of legacy_mem uses shmem_zero_setup()
517 * to satisfy versions of X which crash if the mmap fails: that
518 * substitutes a new vm_file, and we don't then want bin_vm_ops.
519 */
520 if (vma->vm_file != file)
521 goto out_put;
522
523 rc = -EINVAL;
524 if (of->mmapped && of->vm_ops != vma->vm_ops)
525 goto out_put;
526
527 /*
528 * It is not possible to successfully wrap close.
529 * So error if someone is trying to use close.
530 */
531 if (vma->vm_ops && vma->vm_ops->close)
532 goto out_put;
533
534 rc = 0;
535 of->mmapped = true;
536 of_on(of)->nr_mmapped++;
537 of->vm_ops = vma->vm_ops;
538 vma->vm_ops = &kernfs_vm_ops;
539out_put:
540 kernfs_put_active(of->kn);
541out_unlock:
542 mutex_unlock(&of->mutex);
543
544 return rc;
545}
546
547/**
548 * kernfs_get_open_node - get or create kernfs_open_node
549 * @kn: target kernfs_node
550 * @of: kernfs_open_file for this instance of open
551 *
552 * If @kn->attr.open exists, increment its reference count; otherwise,
553 * create one. @of is chained to the files list.
554 *
555 * Locking:
556 * Kernel thread context (may sleep).
557 *
558 * Return:
559 * %0 on success, -errno on failure.
560 */
561static int kernfs_get_open_node(struct kernfs_node *kn,
562 struct kernfs_open_file *of)
563{
564 struct kernfs_open_node *on;
565 struct mutex *mutex;
566
567 mutex = kernfs_open_file_mutex_lock(kn);
568 on = kernfs_deref_open_node_locked(kn);
569
570 if (!on) {
571 /* not there, initialize a new one */
572 on = kzalloc(sizeof(*on), GFP_KERNEL);
573 if (!on) {
574 mutex_unlock(mutex);
575 return -ENOMEM;
576 }
577 atomic_set(&on->event, 1);
578 init_waitqueue_head(&on->poll);
579 INIT_LIST_HEAD(&on->files);
580 rcu_assign_pointer(kn->attr.open, on);
581 }
582
583 list_add_tail(&of->list, &on->files);
584 if (kn->flags & KERNFS_HAS_RELEASE)
585 on->nr_to_release++;
586
587 mutex_unlock(mutex);
588 return 0;
589}
590
591/**
592 * kernfs_unlink_open_file - Unlink @of from @kn.
593 *
594 * @kn: target kernfs_node
595 * @of: associated kernfs_open_file
596 * @open_failed: ->open() failed, cancel ->release()
597 *
598 * Unlink @of from list of @kn's associated open files. If list of
599 * associated open files becomes empty, disassociate and free
600 * kernfs_open_node.
601 *
602 * LOCKING:
603 * None.
604 */
605static void kernfs_unlink_open_file(struct kernfs_node *kn,
606 struct kernfs_open_file *of,
607 bool open_failed)
608{
609 struct kernfs_open_node *on;
610 struct mutex *mutex;
611
612 mutex = kernfs_open_file_mutex_lock(kn);
613
614 on = kernfs_deref_open_node_locked(kn);
615 if (!on) {
616 mutex_unlock(mutex);
617 return;
618 }
619
620 if (of) {
621 if (kn->flags & KERNFS_HAS_RELEASE) {
622 WARN_ON_ONCE(of->released == open_failed);
623 if (open_failed)
624 on->nr_to_release--;
625 }
626 if (of->mmapped)
627 on->nr_mmapped--;
628 list_del(&of->list);
629 }
630
631 if (list_empty(&on->files)) {
632 rcu_assign_pointer(kn->attr.open, NULL);
633 kfree_rcu(on, rcu_head);
634 }
635
636 mutex_unlock(mutex);
637}
638
639static int kernfs_fop_open(struct inode *inode, struct file *file)
640{
641 struct kernfs_node *kn = inode->i_private;
642 struct kernfs_root *root = kernfs_root(kn);
643 const struct kernfs_ops *ops;
644 struct kernfs_open_file *of;
645 bool has_read, has_write, has_mmap;
646 int error = -EACCES;
647
648 if (!kernfs_get_active(kn))
649 return -ENODEV;
650
651 ops = kernfs_ops(kn);
652
653 has_read = ops->seq_show || ops->read || ops->mmap;
654 has_write = ops->write || ops->mmap;
655 has_mmap = ops->mmap;
656
657 /* see the flag definition for details */
658 if (root->flags & KERNFS_ROOT_EXTRA_OPEN_PERM_CHECK) {
659 if ((file->f_mode & FMODE_WRITE) &&
660 (!(inode->i_mode & S_IWUGO) || !has_write))
661 goto err_out;
662
663 if ((file->f_mode & FMODE_READ) &&
664 (!(inode->i_mode & S_IRUGO) || !has_read))
665 goto err_out;
666 }
667
668 /* allocate a kernfs_open_file for the file */
669 error = -ENOMEM;
670 of = kzalloc(sizeof(struct kernfs_open_file), GFP_KERNEL);
671 if (!of)
672 goto err_out;
673
674 /*
675 * The following is done to give a different lockdep key to
676 * @of->mutex for files which implement mmap. This is a rather
677 * crude way to avoid false positive lockdep warning around
678 * mm->mmap_lock - mmap nests @of->mutex under mm->mmap_lock and
679 * reading /sys/block/sda/trace/act_mask grabs sr_mutex, under
680 * which mm->mmap_lock nests, while holding @of->mutex. As each
681 * open file has a separate mutex, it's okay as long as those don't
682 * happen on the same file. At this point, we can't easily give
683 * each file a separate locking class. Let's differentiate on
684 * whether the file has mmap or not for now.
685 *
686 * Both paths of the branch look the same. They're supposed to
687 * look that way and give @of->mutex different static lockdep keys.
688 */
689 if (has_mmap)
690 mutex_init(&of->mutex);
691 else
692 mutex_init(&of->mutex);
693
694 of->kn = kn;
695 of->file = file;
696
697 /*
698 * Write path needs to atomic_write_len outside active reference.
699 * Cache it in open_file. See kernfs_fop_write_iter() for details.
700 */
701 of->atomic_write_len = ops->atomic_write_len;
702
703 error = -EINVAL;
704 /*
705 * ->seq_show is incompatible with ->prealloc,
706 * as seq_read does its own allocation.
707 * ->read must be used instead.
708 */
709 if (ops->prealloc && ops->seq_show)
710 goto err_free;
711 if (ops->prealloc) {
712 int len = of->atomic_write_len ?: PAGE_SIZE;
713 of->prealloc_buf = kmalloc(len + 1, GFP_KERNEL);
714 error = -ENOMEM;
715 if (!of->prealloc_buf)
716 goto err_free;
717 mutex_init(&of->prealloc_mutex);
718 }
719
720 /*
721 * Always instantiate seq_file even if read access doesn't use
722 * seq_file or is not requested. This unifies private data access
723 * and readable regular files are the vast majority anyway.
724 */
725 if (ops->seq_show)
726 error = seq_open(file, &kernfs_seq_ops);
727 else
728 error = seq_open(file, NULL);
729 if (error)
730 goto err_free;
731
732 of->seq_file = file->private_data;
733 of->seq_file->private = of;
734
735 /* seq_file clears PWRITE unconditionally, restore it if WRITE */
736 if (file->f_mode & FMODE_WRITE)
737 file->f_mode |= FMODE_PWRITE;
738
739 /* make sure we have open node struct */
740 error = kernfs_get_open_node(kn, of);
741 if (error)
742 goto err_seq_release;
743
744 if (ops->open) {
745 /* nobody has access to @of yet, skip @of->mutex */
746 error = ops->open(of);
747 if (error)
748 goto err_put_node;
749 }
750
751 /* open succeeded, put active references */
752 kernfs_put_active(kn);
753 return 0;
754
755err_put_node:
756 kernfs_unlink_open_file(kn, of, true);
757err_seq_release:
758 seq_release(inode, file);
759err_free:
760 kfree(of->prealloc_buf);
761 kfree(of);
762err_out:
763 kernfs_put_active(kn);
764 return error;
765}
766
767/* used from release/drain to ensure that ->release() is called exactly once */
768static void kernfs_release_file(struct kernfs_node *kn,
769 struct kernfs_open_file *of)
770{
771 /*
772 * @of is guaranteed to have no other file operations in flight and
773 * we just want to synchronize release and drain paths.
774 * @kernfs_open_file_mutex_ptr(kn) is enough. @of->mutex can't be used
775 * here because drain path may be called from places which can
776 * cause circular dependency.
777 */
778 lockdep_assert_held(kernfs_open_file_mutex_ptr(kn));
779
780 if (!of->released) {
781 /*
782 * A file is never detached without being released and we
783 * need to be able to release files which are deactivated
784 * and being drained. Don't use kernfs_ops().
785 */
786 kn->attr.ops->release(of);
787 of->released = true;
788 of_on(of)->nr_to_release--;
789 }
790}
791
792static int kernfs_fop_release(struct inode *inode, struct file *filp)
793{
794 struct kernfs_node *kn = inode->i_private;
795 struct kernfs_open_file *of = kernfs_of(filp);
796
797 if (kn->flags & KERNFS_HAS_RELEASE) {
798 struct mutex *mutex;
799
800 mutex = kernfs_open_file_mutex_lock(kn);
801 kernfs_release_file(kn, of);
802 mutex_unlock(mutex);
803 }
804
805 kernfs_unlink_open_file(kn, of, false);
806 seq_release(inode, filp);
807 kfree(of->prealloc_buf);
808 kfree(of);
809
810 return 0;
811}
812
813bool kernfs_should_drain_open_files(struct kernfs_node *kn)
814{
815 struct kernfs_open_node *on;
816 bool ret;
817
818 /*
819 * @kn being deactivated guarantees that @kn->attr.open can't change
820 * beneath us making the lockless test below safe.
821 */
822 WARN_ON_ONCE(atomic_read(&kn->active) != KN_DEACTIVATED_BIAS);
823
824 rcu_read_lock();
825 on = rcu_dereference(kn->attr.open);
826 ret = on && (on->nr_mmapped || on->nr_to_release);
827 rcu_read_unlock();
828
829 return ret;
830}
831
832void kernfs_drain_open_files(struct kernfs_node *kn)
833{
834 struct kernfs_open_node *on;
835 struct kernfs_open_file *of;
836 struct mutex *mutex;
837
838 mutex = kernfs_open_file_mutex_lock(kn);
839 on = kernfs_deref_open_node_locked(kn);
840 if (!on) {
841 mutex_unlock(mutex);
842 return;
843 }
844
845 list_for_each_entry(of, &on->files, list) {
846 struct inode *inode = file_inode(of->file);
847
848 if (of->mmapped) {
849 unmap_mapping_range(inode->i_mapping, 0, 0, 1);
850 of->mmapped = false;
851 on->nr_mmapped--;
852 }
853
854 if (kn->flags & KERNFS_HAS_RELEASE)
855 kernfs_release_file(kn, of);
856 }
857
858 WARN_ON_ONCE(on->nr_mmapped || on->nr_to_release);
859 mutex_unlock(mutex);
860}
861
862/*
863 * Kernfs attribute files are pollable. The idea is that you read
864 * the content and then you use 'poll' or 'select' to wait for
865 * the content to change. When the content changes (assuming the
866 * manager for the kobject supports notification), poll will
867 * return EPOLLERR|EPOLLPRI, and select will return the fd whether
868 * it is waiting for read, write, or exceptions.
869 * Once poll/select indicates that the value has changed, you
870 * need to close and re-open the file, or seek to 0 and read again.
871 * Reminder: this only works for attributes which actively support
872 * it, and it is not possible to test an attribute from userspace
873 * to see if it supports poll (Neither 'poll' nor 'select' return
874 * an appropriate error code). When in doubt, set a suitable timeout value.
875 */
876__poll_t kernfs_generic_poll(struct kernfs_open_file *of, poll_table *wait)
877{
878 struct kernfs_open_node *on = of_on(of);
879
880 poll_wait(of->file, &on->poll, wait);
881
882 if (of->event != atomic_read(&on->event))
883 return DEFAULT_POLLMASK|EPOLLERR|EPOLLPRI;
884
885 return DEFAULT_POLLMASK;
886}
887
888static __poll_t kernfs_fop_poll(struct file *filp, poll_table *wait)
889{
890 struct kernfs_open_file *of = kernfs_of(filp);
891 struct kernfs_node *kn = kernfs_dentry_node(filp->f_path.dentry);
892 __poll_t ret;
893
894 if (!kernfs_get_active(kn))
895 return DEFAULT_POLLMASK|EPOLLERR|EPOLLPRI;
896
897 if (kn->attr.ops->poll)
898 ret = kn->attr.ops->poll(of, wait);
899 else
900 ret = kernfs_generic_poll(of, wait);
901
902 kernfs_put_active(kn);
903 return ret;
904}
905
906static void kernfs_notify_workfn(struct work_struct *work)
907{
908 struct kernfs_node *kn;
909 struct kernfs_super_info *info;
910 struct kernfs_root *root;
911repeat:
912 /* pop one off the notify_list */
913 spin_lock_irq(&kernfs_notify_lock);
914 kn = kernfs_notify_list;
915 if (kn == KERNFS_NOTIFY_EOL) {
916 spin_unlock_irq(&kernfs_notify_lock);
917 return;
918 }
919 kernfs_notify_list = kn->attr.notify_next;
920 kn->attr.notify_next = NULL;
921 spin_unlock_irq(&kernfs_notify_lock);
922
923 root = kernfs_root(kn);
924 /* kick fsnotify */
925 down_write(&root->kernfs_rwsem);
926
927 list_for_each_entry(info, &kernfs_root(kn)->supers, node) {
928 struct kernfs_node *parent;
929 struct inode *p_inode = NULL;
930 struct inode *inode;
931 struct qstr name;
932
933 /*
934 * We want fsnotify_modify() on @kn but as the
935 * modifications aren't originating from userland don't
936 * have the matching @file available. Look up the inodes
937 * and generate the events manually.
938 */
939 inode = ilookup(info->sb, kernfs_ino(kn));
940 if (!inode)
941 continue;
942
943 name = (struct qstr)QSTR_INIT(kn->name, strlen(kn->name));
944 parent = kernfs_get_parent(kn);
945 if (parent) {
946 p_inode = ilookup(info->sb, kernfs_ino(parent));
947 if (p_inode) {
948 fsnotify(FS_MODIFY | FS_EVENT_ON_CHILD,
949 inode, FSNOTIFY_EVENT_INODE,
950 p_inode, &name, inode, 0);
951 iput(p_inode);
952 }
953
954 kernfs_put(parent);
955 }
956
957 if (!p_inode)
958 fsnotify_inode(inode, FS_MODIFY);
959
960 iput(inode);
961 }
962
963 up_write(&root->kernfs_rwsem);
964 kernfs_put(kn);
965 goto repeat;
966}
967
968/**
969 * kernfs_notify - notify a kernfs file
970 * @kn: file to notify
971 *
972 * Notify @kn such that poll(2) on @kn wakes up. Maybe be called from any
973 * context.
974 */
975void kernfs_notify(struct kernfs_node *kn)
976{
977 static DECLARE_WORK(kernfs_notify_work, kernfs_notify_workfn);
978 unsigned long flags;
979 struct kernfs_open_node *on;
980
981 if (WARN_ON(kernfs_type(kn) != KERNFS_FILE))
982 return;
983
984 /* kick poll immediately */
985 rcu_read_lock();
986 on = rcu_dereference(kn->attr.open);
987 if (on) {
988 atomic_inc(&on->event);
989 wake_up_interruptible(&on->poll);
990 }
991 rcu_read_unlock();
992
993 /* schedule work to kick fsnotify */
994 spin_lock_irqsave(&kernfs_notify_lock, flags);
995 if (!kn->attr.notify_next) {
996 kernfs_get(kn);
997 kn->attr.notify_next = kernfs_notify_list;
998 kernfs_notify_list = kn;
999 schedule_work(&kernfs_notify_work);
1000 }
1001 spin_unlock_irqrestore(&kernfs_notify_lock, flags);
1002}
1003EXPORT_SYMBOL_GPL(kernfs_notify);
1004
1005const struct file_operations kernfs_file_fops = {
1006 .read_iter = kernfs_fop_read_iter,
1007 .write_iter = kernfs_fop_write_iter,
1008 .llseek = generic_file_llseek,
1009 .mmap = kernfs_fop_mmap,
1010 .open = kernfs_fop_open,
1011 .release = kernfs_fop_release,
1012 .poll = kernfs_fop_poll,
1013 .fsync = noop_fsync,
1014 .splice_read = generic_file_splice_read,
1015 .splice_write = iter_file_splice_write,
1016};
1017
1018/**
1019 * __kernfs_create_file - kernfs internal function to create a file
1020 * @parent: directory to create the file in
1021 * @name: name of the file
1022 * @mode: mode of the file
1023 * @uid: uid of the file
1024 * @gid: gid of the file
1025 * @size: size of the file
1026 * @ops: kernfs operations for the file
1027 * @priv: private data for the file
1028 * @ns: optional namespace tag of the file
1029 * @key: lockdep key for the file's active_ref, %NULL to disable lockdep
1030 *
1031 * Return: the created node on success, ERR_PTR() value on error.
1032 */
1033struct kernfs_node *__kernfs_create_file(struct kernfs_node *parent,
1034 const char *name,
1035 umode_t mode, kuid_t uid, kgid_t gid,
1036 loff_t size,
1037 const struct kernfs_ops *ops,
1038 void *priv, const void *ns,
1039 struct lock_class_key *key)
1040{
1041 struct kernfs_node *kn;
1042 unsigned flags;
1043 int rc;
1044
1045 flags = KERNFS_FILE;
1046
1047 kn = kernfs_new_node(parent, name, (mode & S_IALLUGO) | S_IFREG,
1048 uid, gid, flags);
1049 if (!kn)
1050 return ERR_PTR(-ENOMEM);
1051
1052 kn->attr.ops = ops;
1053 kn->attr.size = size;
1054 kn->ns = ns;
1055 kn->priv = priv;
1056
1057#ifdef CONFIG_DEBUG_LOCK_ALLOC
1058 if (key) {
1059 lockdep_init_map(&kn->dep_map, "kn->active", key, 0);
1060 kn->flags |= KERNFS_LOCKDEP;
1061 }
1062#endif
1063
1064 /*
1065 * kn->attr.ops is accessible only while holding active ref. We
1066 * need to know whether some ops are implemented outside active
1067 * ref. Cache their existence in flags.
1068 */
1069 if (ops->seq_show)
1070 kn->flags |= KERNFS_HAS_SEQ_SHOW;
1071 if (ops->mmap)
1072 kn->flags |= KERNFS_HAS_MMAP;
1073 if (ops->release)
1074 kn->flags |= KERNFS_HAS_RELEASE;
1075
1076 rc = kernfs_add_one(kn);
1077 if (rc) {
1078 kernfs_put(kn);
1079 return ERR_PTR(rc);
1080 }
1081 return kn;
1082}