Loading...
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
432static const struct vm_operations_struct kernfs_vm_ops = {
433 .open = kernfs_vma_open,
434 .fault = kernfs_vma_fault,
435 .page_mkwrite = kernfs_vma_page_mkwrite,
436 .access = kernfs_vma_access,
437};
438
439static int kernfs_fop_mmap(struct file *file, struct vm_area_struct *vma)
440{
441 struct kernfs_open_file *of = kernfs_of(file);
442 const struct kernfs_ops *ops;
443 int rc;
444
445 /*
446 * mmap path and of->mutex are prone to triggering spurious lockdep
447 * warnings and we don't want to add spurious locking dependency
448 * between the two. Check whether mmap is actually implemented
449 * without grabbing @of->mutex by testing HAS_MMAP flag. See the
450 * comment in kernfs_fop_open() for more details.
451 */
452 if (!(of->kn->flags & KERNFS_HAS_MMAP))
453 return -ENODEV;
454
455 mutex_lock(&of->mutex);
456
457 rc = -ENODEV;
458 if (!kernfs_get_active(of->kn))
459 goto out_unlock;
460
461 ops = kernfs_ops(of->kn);
462 rc = ops->mmap(of, vma);
463 if (rc)
464 goto out_put;
465
466 /*
467 * PowerPC's pci_mmap of legacy_mem uses shmem_zero_setup()
468 * to satisfy versions of X which crash if the mmap fails: that
469 * substitutes a new vm_file, and we don't then want bin_vm_ops.
470 */
471 if (vma->vm_file != file)
472 goto out_put;
473
474 rc = -EINVAL;
475 if (of->mmapped && of->vm_ops != vma->vm_ops)
476 goto out_put;
477
478 /*
479 * It is not possible to successfully wrap close.
480 * So error if someone is trying to use close.
481 */
482 if (vma->vm_ops && vma->vm_ops->close)
483 goto out_put;
484
485 rc = 0;
486 if (!of->mmapped) {
487 of->mmapped = true;
488 of_on(of)->nr_mmapped++;
489 of->vm_ops = vma->vm_ops;
490 }
491 vma->vm_ops = &kernfs_vm_ops;
492out_put:
493 kernfs_put_active(of->kn);
494out_unlock:
495 mutex_unlock(&of->mutex);
496
497 return rc;
498}
499
500/**
501 * kernfs_get_open_node - get or create kernfs_open_node
502 * @kn: target kernfs_node
503 * @of: kernfs_open_file for this instance of open
504 *
505 * If @kn->attr.open exists, increment its reference count; otherwise,
506 * create one. @of is chained to the files list.
507 *
508 * Locking:
509 * Kernel thread context (may sleep).
510 *
511 * Return:
512 * %0 on success, -errno on failure.
513 */
514static int kernfs_get_open_node(struct kernfs_node *kn,
515 struct kernfs_open_file *of)
516{
517 struct kernfs_open_node *on;
518 struct mutex *mutex;
519
520 mutex = kernfs_open_file_mutex_lock(kn);
521 on = kernfs_deref_open_node_locked(kn);
522
523 if (!on) {
524 /* not there, initialize a new one */
525 on = kzalloc(sizeof(*on), GFP_KERNEL);
526 if (!on) {
527 mutex_unlock(mutex);
528 return -ENOMEM;
529 }
530 atomic_set(&on->event, 1);
531 init_waitqueue_head(&on->poll);
532 INIT_LIST_HEAD(&on->files);
533 rcu_assign_pointer(kn->attr.open, on);
534 }
535
536 list_add_tail(&of->list, &on->files);
537 if (kn->flags & KERNFS_HAS_RELEASE)
538 on->nr_to_release++;
539
540 mutex_unlock(mutex);
541 return 0;
542}
543
544/**
545 * kernfs_unlink_open_file - Unlink @of from @kn.
546 *
547 * @kn: target kernfs_node
548 * @of: associated kernfs_open_file
549 * @open_failed: ->open() failed, cancel ->release()
550 *
551 * Unlink @of from list of @kn's associated open files. If list of
552 * associated open files becomes empty, disassociate and free
553 * kernfs_open_node.
554 *
555 * LOCKING:
556 * None.
557 */
558static void kernfs_unlink_open_file(struct kernfs_node *kn,
559 struct kernfs_open_file *of,
560 bool open_failed)
561{
562 struct kernfs_open_node *on;
563 struct mutex *mutex;
564
565 mutex = kernfs_open_file_mutex_lock(kn);
566
567 on = kernfs_deref_open_node_locked(kn);
568 if (!on) {
569 mutex_unlock(mutex);
570 return;
571 }
572
573 if (of) {
574 if (kn->flags & KERNFS_HAS_RELEASE) {
575 WARN_ON_ONCE(of->released == open_failed);
576 if (open_failed)
577 on->nr_to_release--;
578 }
579 if (of->mmapped)
580 on->nr_mmapped--;
581 list_del(&of->list);
582 }
583
584 if (list_empty(&on->files)) {
585 rcu_assign_pointer(kn->attr.open, NULL);
586 kfree_rcu(on, rcu_head);
587 }
588
589 mutex_unlock(mutex);
590}
591
592static int kernfs_fop_open(struct inode *inode, struct file *file)
593{
594 struct kernfs_node *kn = inode->i_private;
595 struct kernfs_root *root = kernfs_root(kn);
596 const struct kernfs_ops *ops;
597 struct kernfs_open_file *of;
598 bool has_read, has_write, has_mmap;
599 int error = -EACCES;
600
601 if (!kernfs_get_active(kn))
602 return -ENODEV;
603
604 ops = kernfs_ops(kn);
605
606 has_read = ops->seq_show || ops->read || ops->mmap;
607 has_write = ops->write || ops->mmap;
608 has_mmap = ops->mmap;
609
610 /* see the flag definition for details */
611 if (root->flags & KERNFS_ROOT_EXTRA_OPEN_PERM_CHECK) {
612 if ((file->f_mode & FMODE_WRITE) &&
613 (!(inode->i_mode & S_IWUGO) || !has_write))
614 goto err_out;
615
616 if ((file->f_mode & FMODE_READ) &&
617 (!(inode->i_mode & S_IRUGO) || !has_read))
618 goto err_out;
619 }
620
621 /* allocate a kernfs_open_file for the file */
622 error = -ENOMEM;
623 of = kzalloc(sizeof(struct kernfs_open_file), GFP_KERNEL);
624 if (!of)
625 goto err_out;
626
627 /*
628 * The following is done to give a different lockdep key to
629 * @of->mutex for files which implement mmap. This is a rather
630 * crude way to avoid false positive lockdep warning around
631 * mm->mmap_lock - mmap nests @of->mutex under mm->mmap_lock and
632 * reading /sys/block/sda/trace/act_mask grabs sr_mutex, under
633 * which mm->mmap_lock nests, while holding @of->mutex. As each
634 * open file has a separate mutex, it's okay as long as those don't
635 * happen on the same file. At this point, we can't easily give
636 * each file a separate locking class. Let's differentiate on
637 * whether the file has mmap or not for now.
638 *
639 * For similar reasons, writable and readonly files are given different
640 * lockdep key, because the writable file /sys/power/resume may call vfs
641 * lookup helpers for arbitrary paths and readonly files can be read by
642 * overlayfs from vfs helpers when sysfs is a lower layer of overalyfs.
643 *
644 * All three cases look the same. They're supposed to
645 * look that way and give @of->mutex different static lockdep keys.
646 */
647 if (has_mmap)
648 mutex_init(&of->mutex);
649 else if (file->f_mode & FMODE_WRITE)
650 mutex_init(&of->mutex);
651 else
652 mutex_init(&of->mutex);
653
654 of->kn = kn;
655 of->file = file;
656
657 /*
658 * Write path needs to atomic_write_len outside active reference.
659 * Cache it in open_file. See kernfs_fop_write_iter() for details.
660 */
661 of->atomic_write_len = ops->atomic_write_len;
662
663 error = -EINVAL;
664 /*
665 * ->seq_show is incompatible with ->prealloc,
666 * as seq_read does its own allocation.
667 * ->read must be used instead.
668 */
669 if (ops->prealloc && ops->seq_show)
670 goto err_free;
671 if (ops->prealloc) {
672 int len = of->atomic_write_len ?: PAGE_SIZE;
673 of->prealloc_buf = kmalloc(len + 1, GFP_KERNEL);
674 error = -ENOMEM;
675 if (!of->prealloc_buf)
676 goto err_free;
677 mutex_init(&of->prealloc_mutex);
678 }
679
680 /*
681 * Always instantiate seq_file even if read access doesn't use
682 * seq_file or is not requested. This unifies private data access
683 * and readable regular files are the vast majority anyway.
684 */
685 if (ops->seq_show)
686 error = seq_open(file, &kernfs_seq_ops);
687 else
688 error = seq_open(file, NULL);
689 if (error)
690 goto err_free;
691
692 of->seq_file = file->private_data;
693 of->seq_file->private = of;
694
695 /* seq_file clears PWRITE unconditionally, restore it if WRITE */
696 if (file->f_mode & FMODE_WRITE)
697 file->f_mode |= FMODE_PWRITE;
698
699 /* make sure we have open node struct */
700 error = kernfs_get_open_node(kn, of);
701 if (error)
702 goto err_seq_release;
703
704 if (ops->open) {
705 /* nobody has access to @of yet, skip @of->mutex */
706 error = ops->open(of);
707 if (error)
708 goto err_put_node;
709 }
710
711 /* open succeeded, put active references */
712 kernfs_put_active(kn);
713 return 0;
714
715err_put_node:
716 kernfs_unlink_open_file(kn, of, true);
717err_seq_release:
718 seq_release(inode, file);
719err_free:
720 kfree(of->prealloc_buf);
721 kfree(of);
722err_out:
723 kernfs_put_active(kn);
724 return error;
725}
726
727/* used from release/drain to ensure that ->release() is called exactly once */
728static void kernfs_release_file(struct kernfs_node *kn,
729 struct kernfs_open_file *of)
730{
731 /*
732 * @of is guaranteed to have no other file operations in flight and
733 * we just want to synchronize release and drain paths.
734 * @kernfs_open_file_mutex_ptr(kn) is enough. @of->mutex can't be used
735 * here because drain path may be called from places which can
736 * cause circular dependency.
737 */
738 lockdep_assert_held(kernfs_open_file_mutex_ptr(kn));
739
740 if (!of->released) {
741 /*
742 * A file is never detached without being released and we
743 * need to be able to release files which are deactivated
744 * and being drained. Don't use kernfs_ops().
745 */
746 kn->attr.ops->release(of);
747 of->released = true;
748 of_on(of)->nr_to_release--;
749 }
750}
751
752static int kernfs_fop_release(struct inode *inode, struct file *filp)
753{
754 struct kernfs_node *kn = inode->i_private;
755 struct kernfs_open_file *of = kernfs_of(filp);
756
757 if (kn->flags & KERNFS_HAS_RELEASE) {
758 struct mutex *mutex;
759
760 mutex = kernfs_open_file_mutex_lock(kn);
761 kernfs_release_file(kn, of);
762 mutex_unlock(mutex);
763 }
764
765 kernfs_unlink_open_file(kn, of, false);
766 seq_release(inode, filp);
767 kfree(of->prealloc_buf);
768 kfree(of);
769
770 return 0;
771}
772
773bool kernfs_should_drain_open_files(struct kernfs_node *kn)
774{
775 struct kernfs_open_node *on;
776 bool ret;
777
778 /*
779 * @kn being deactivated guarantees that @kn->attr.open can't change
780 * beneath us making the lockless test below safe.
781 */
782 WARN_ON_ONCE(atomic_read(&kn->active) != KN_DEACTIVATED_BIAS);
783
784 rcu_read_lock();
785 on = rcu_dereference(kn->attr.open);
786 ret = on && (on->nr_mmapped || on->nr_to_release);
787 rcu_read_unlock();
788
789 return ret;
790}
791
792void kernfs_drain_open_files(struct kernfs_node *kn)
793{
794 struct kernfs_open_node *on;
795 struct kernfs_open_file *of;
796 struct mutex *mutex;
797
798 mutex = kernfs_open_file_mutex_lock(kn);
799 on = kernfs_deref_open_node_locked(kn);
800 if (!on) {
801 mutex_unlock(mutex);
802 return;
803 }
804
805 list_for_each_entry(of, &on->files, list) {
806 struct inode *inode = file_inode(of->file);
807
808 if (of->mmapped) {
809 unmap_mapping_range(inode->i_mapping, 0, 0, 1);
810 of->mmapped = false;
811 on->nr_mmapped--;
812 }
813
814 if (kn->flags & KERNFS_HAS_RELEASE)
815 kernfs_release_file(kn, of);
816 }
817
818 WARN_ON_ONCE(on->nr_mmapped || on->nr_to_release);
819 mutex_unlock(mutex);
820}
821
822/*
823 * Kernfs attribute files are pollable. The idea is that you read
824 * the content and then you use 'poll' or 'select' to wait for
825 * the content to change. When the content changes (assuming the
826 * manager for the kobject supports notification), poll will
827 * return EPOLLERR|EPOLLPRI, and select will return the fd whether
828 * it is waiting for read, write, or exceptions.
829 * Once poll/select indicates that the value has changed, you
830 * need to close and re-open the file, or seek to 0 and read again.
831 * Reminder: this only works for attributes which actively support
832 * it, and it is not possible to test an attribute from userspace
833 * to see if it supports poll (Neither 'poll' nor 'select' return
834 * an appropriate error code). When in doubt, set a suitable timeout value.
835 */
836__poll_t kernfs_generic_poll(struct kernfs_open_file *of, poll_table *wait)
837{
838 struct kernfs_open_node *on = of_on(of);
839
840 poll_wait(of->file, &on->poll, wait);
841
842 if (of->event != atomic_read(&on->event))
843 return DEFAULT_POLLMASK|EPOLLERR|EPOLLPRI;
844
845 return DEFAULT_POLLMASK;
846}
847
848static __poll_t kernfs_fop_poll(struct file *filp, poll_table *wait)
849{
850 struct kernfs_open_file *of = kernfs_of(filp);
851 struct kernfs_node *kn = kernfs_dentry_node(filp->f_path.dentry);
852 __poll_t ret;
853
854 if (!kernfs_get_active(kn))
855 return DEFAULT_POLLMASK|EPOLLERR|EPOLLPRI;
856
857 if (kn->attr.ops->poll)
858 ret = kn->attr.ops->poll(of, wait);
859 else
860 ret = kernfs_generic_poll(of, wait);
861
862 kernfs_put_active(kn);
863 return ret;
864}
865
866static loff_t kernfs_fop_llseek(struct file *file, loff_t offset, int whence)
867{
868 struct kernfs_open_file *of = kernfs_of(file);
869 const struct kernfs_ops *ops;
870 loff_t ret;
871
872 /*
873 * @of->mutex nests outside active ref and is primarily to ensure that
874 * the ops aren't called concurrently for the same open file.
875 */
876 mutex_lock(&of->mutex);
877 if (!kernfs_get_active(of->kn)) {
878 mutex_unlock(&of->mutex);
879 return -ENODEV;
880 }
881
882 ops = kernfs_ops(of->kn);
883 if (ops->llseek)
884 ret = ops->llseek(of, offset, whence);
885 else
886 ret = generic_file_llseek(file, offset, whence);
887
888 kernfs_put_active(of->kn);
889 mutex_unlock(&of->mutex);
890 return ret;
891}
892
893static void kernfs_notify_workfn(struct work_struct *work)
894{
895 struct kernfs_node *kn;
896 struct kernfs_super_info *info;
897 struct kernfs_root *root;
898repeat:
899 /* pop one off the notify_list */
900 spin_lock_irq(&kernfs_notify_lock);
901 kn = kernfs_notify_list;
902 if (kn == KERNFS_NOTIFY_EOL) {
903 spin_unlock_irq(&kernfs_notify_lock);
904 return;
905 }
906 kernfs_notify_list = kn->attr.notify_next;
907 kn->attr.notify_next = NULL;
908 spin_unlock_irq(&kernfs_notify_lock);
909
910 root = kernfs_root(kn);
911 /* kick fsnotify */
912
913 down_read(&root->kernfs_supers_rwsem);
914 list_for_each_entry(info, &kernfs_root(kn)->supers, node) {
915 struct kernfs_node *parent;
916 struct inode *p_inode = NULL;
917 struct inode *inode;
918 struct qstr name;
919
920 /*
921 * We want fsnotify_modify() on @kn but as the
922 * modifications aren't originating from userland don't
923 * have the matching @file available. Look up the inodes
924 * and generate the events manually.
925 */
926 inode = ilookup(info->sb, kernfs_ino(kn));
927 if (!inode)
928 continue;
929
930 name = (struct qstr)QSTR_INIT(kn->name, strlen(kn->name));
931 parent = kernfs_get_parent(kn);
932 if (parent) {
933 p_inode = ilookup(info->sb, kernfs_ino(parent));
934 if (p_inode) {
935 fsnotify(FS_MODIFY | FS_EVENT_ON_CHILD,
936 inode, FSNOTIFY_EVENT_INODE,
937 p_inode, &name, inode, 0);
938 iput(p_inode);
939 }
940
941 kernfs_put(parent);
942 }
943
944 if (!p_inode)
945 fsnotify_inode(inode, FS_MODIFY);
946
947 iput(inode);
948 }
949
950 up_read(&root->kernfs_supers_rwsem);
951 kernfs_put(kn);
952 goto repeat;
953}
954
955/**
956 * kernfs_notify - notify a kernfs file
957 * @kn: file to notify
958 *
959 * Notify @kn such that poll(2) on @kn wakes up. Maybe be called from any
960 * context.
961 */
962void kernfs_notify(struct kernfs_node *kn)
963{
964 static DECLARE_WORK(kernfs_notify_work, kernfs_notify_workfn);
965 unsigned long flags;
966 struct kernfs_open_node *on;
967
968 if (WARN_ON(kernfs_type(kn) != KERNFS_FILE))
969 return;
970
971 /* kick poll immediately */
972 rcu_read_lock();
973 on = rcu_dereference(kn->attr.open);
974 if (on) {
975 atomic_inc(&on->event);
976 wake_up_interruptible(&on->poll);
977 }
978 rcu_read_unlock();
979
980 /* schedule work to kick fsnotify */
981 spin_lock_irqsave(&kernfs_notify_lock, flags);
982 if (!kn->attr.notify_next) {
983 kernfs_get(kn);
984 kn->attr.notify_next = kernfs_notify_list;
985 kernfs_notify_list = kn;
986 schedule_work(&kernfs_notify_work);
987 }
988 spin_unlock_irqrestore(&kernfs_notify_lock, flags);
989}
990EXPORT_SYMBOL_GPL(kernfs_notify);
991
992const struct file_operations kernfs_file_fops = {
993 .read_iter = kernfs_fop_read_iter,
994 .write_iter = kernfs_fop_write_iter,
995 .llseek = kernfs_fop_llseek,
996 .mmap = kernfs_fop_mmap,
997 .open = kernfs_fop_open,
998 .release = kernfs_fop_release,
999 .poll = kernfs_fop_poll,
1000 .fsync = noop_fsync,
1001 .splice_read = copy_splice_read,
1002 .splice_write = iter_file_splice_write,
1003};
1004
1005/**
1006 * __kernfs_create_file - kernfs internal function to create a file
1007 * @parent: directory to create the file in
1008 * @name: name of the file
1009 * @mode: mode of the file
1010 * @uid: uid of the file
1011 * @gid: gid of the file
1012 * @size: size of the file
1013 * @ops: kernfs operations for the file
1014 * @priv: private data for the file
1015 * @ns: optional namespace tag of the file
1016 * @key: lockdep key for the file's active_ref, %NULL to disable lockdep
1017 *
1018 * Return: the created node on success, ERR_PTR() value on error.
1019 */
1020struct kernfs_node *__kernfs_create_file(struct kernfs_node *parent,
1021 const char *name,
1022 umode_t mode, kuid_t uid, kgid_t gid,
1023 loff_t size,
1024 const struct kernfs_ops *ops,
1025 void *priv, const void *ns,
1026 struct lock_class_key *key)
1027{
1028 struct kernfs_node *kn;
1029 unsigned flags;
1030 int rc;
1031
1032 flags = KERNFS_FILE;
1033
1034 kn = kernfs_new_node(parent, name, (mode & S_IALLUGO) | S_IFREG,
1035 uid, gid, flags);
1036 if (!kn)
1037 return ERR_PTR(-ENOMEM);
1038
1039 kn->attr.ops = ops;
1040 kn->attr.size = size;
1041 kn->ns = ns;
1042 kn->priv = priv;
1043
1044#ifdef CONFIG_DEBUG_LOCK_ALLOC
1045 if (key) {
1046 lockdep_init_map(&kn->dep_map, "kn->active", key, 0);
1047 kn->flags |= KERNFS_LOCKDEP;
1048 }
1049#endif
1050
1051 /*
1052 * kn->attr.ops is accessible only while holding active ref. We
1053 * need to know whether some ops are implemented outside active
1054 * ref. Cache their existence in flags.
1055 */
1056 if (ops->seq_show)
1057 kn->flags |= KERNFS_HAS_SEQ_SHOW;
1058 if (ops->mmap)
1059 kn->flags |= KERNFS_HAS_MMAP;
1060 if (ops->release)
1061 kn->flags |= KERNFS_HAS_RELEASE;
1062
1063 rc = kernfs_add_one(kn);
1064 if (rc) {
1065 kernfs_put(kn);
1066 return ERR_PTR(rc);
1067 }
1068 return kn;
1069}
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
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_read_iter(struct kiocb *iocb, struct iov_iter *iter)
185{
186 struct kernfs_open_file *of = kernfs_of(iocb->ki_filp);
187 ssize_t len = min_t(size_t, iov_iter_count(iter), PAGE_SIZE);
188 const struct kernfs_ops *ops;
189 char *buf;
190
191 buf = of->prealloc_buf;
192 if (buf)
193 mutex_lock(&of->prealloc_mutex);
194 else
195 buf = kmalloc(len, GFP_KERNEL);
196 if (!buf)
197 return -ENOMEM;
198
199 /*
200 * @of->mutex nests outside active ref and is used both to ensure that
201 * the ops aren't called concurrently for the same open file.
202 */
203 mutex_lock(&of->mutex);
204 if (!kernfs_get_active(of->kn)) {
205 len = -ENODEV;
206 mutex_unlock(&of->mutex);
207 goto out_free;
208 }
209
210 of->event = atomic_read(&of->kn->attr.open->event);
211 ops = kernfs_ops(of->kn);
212 if (ops->read)
213 len = ops->read(of, buf, len, iocb->ki_pos);
214 else
215 len = -EINVAL;
216
217 kernfs_put_active(of->kn);
218 mutex_unlock(&of->mutex);
219
220 if (len < 0)
221 goto out_free;
222
223 if (copy_to_iter(buf, len, iter) != len) {
224 len = -EFAULT;
225 goto out_free;
226 }
227
228 iocb->ki_pos += len;
229
230 out_free:
231 if (buf == of->prealloc_buf)
232 mutex_unlock(&of->prealloc_mutex);
233 else
234 kfree(buf);
235 return len;
236}
237
238static ssize_t kernfs_fop_read_iter(struct kiocb *iocb, struct iov_iter *iter)
239{
240 if (kernfs_of(iocb->ki_filp)->kn->flags & KERNFS_HAS_SEQ_SHOW)
241 return seq_read_iter(iocb, iter);
242 return kernfs_file_read_iter(iocb, iter);
243}
244
245/*
246 * Copy data in from userland and pass it to the matching kernfs write
247 * operation.
248 *
249 * There is no easy way for us to know if userspace is only doing a partial
250 * write, so we don't support them. We expect the entire buffer to come on
251 * the first write. Hint: if you're writing a value, first read the file,
252 * modify only the the value you're changing, then write entire buffer
253 * back.
254 */
255static ssize_t kernfs_fop_write_iter(struct kiocb *iocb, struct iov_iter *iter)
256{
257 struct kernfs_open_file *of = kernfs_of(iocb->ki_filp);
258 ssize_t len = iov_iter_count(iter);
259 const struct kernfs_ops *ops;
260 char *buf;
261
262 if (of->atomic_write_len) {
263 if (len > of->atomic_write_len)
264 return -E2BIG;
265 } else {
266 len = min_t(size_t, len, PAGE_SIZE);
267 }
268
269 buf = of->prealloc_buf;
270 if (buf)
271 mutex_lock(&of->prealloc_mutex);
272 else
273 buf = kmalloc(len + 1, GFP_KERNEL);
274 if (!buf)
275 return -ENOMEM;
276
277 if (copy_from_iter(buf, len, iter) != len) {
278 len = -EFAULT;
279 goto out_free;
280 }
281 buf[len] = '\0'; /* guarantee string termination */
282
283 /*
284 * @of->mutex nests outside active ref and is used both to ensure that
285 * the ops aren't called concurrently for the same open file.
286 */
287 mutex_lock(&of->mutex);
288 if (!kernfs_get_active(of->kn)) {
289 mutex_unlock(&of->mutex);
290 len = -ENODEV;
291 goto out_free;
292 }
293
294 ops = kernfs_ops(of->kn);
295 if (ops->write)
296 len = ops->write(of, buf, len, iocb->ki_pos);
297 else
298 len = -EINVAL;
299
300 kernfs_put_active(of->kn);
301 mutex_unlock(&of->mutex);
302
303 if (len > 0)
304 iocb->ki_pos += len;
305
306out_free:
307 if (buf == of->prealloc_buf)
308 mutex_unlock(&of->prealloc_mutex);
309 else
310 kfree(buf);
311 return len;
312}
313
314static void kernfs_vma_open(struct vm_area_struct *vma)
315{
316 struct file *file = vma->vm_file;
317 struct kernfs_open_file *of = kernfs_of(file);
318
319 if (!of->vm_ops)
320 return;
321
322 if (!kernfs_get_active(of->kn))
323 return;
324
325 if (of->vm_ops->open)
326 of->vm_ops->open(vma);
327
328 kernfs_put_active(of->kn);
329}
330
331static vm_fault_t kernfs_vma_fault(struct vm_fault *vmf)
332{
333 struct file *file = vmf->vma->vm_file;
334 struct kernfs_open_file *of = kernfs_of(file);
335 vm_fault_t ret;
336
337 if (!of->vm_ops)
338 return VM_FAULT_SIGBUS;
339
340 if (!kernfs_get_active(of->kn))
341 return VM_FAULT_SIGBUS;
342
343 ret = VM_FAULT_SIGBUS;
344 if (of->vm_ops->fault)
345 ret = of->vm_ops->fault(vmf);
346
347 kernfs_put_active(of->kn);
348 return ret;
349}
350
351static vm_fault_t kernfs_vma_page_mkwrite(struct vm_fault *vmf)
352{
353 struct file *file = vmf->vma->vm_file;
354 struct kernfs_open_file *of = kernfs_of(file);
355 vm_fault_t 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 = 0;
364 if (of->vm_ops->page_mkwrite)
365 ret = of->vm_ops->page_mkwrite(vmf);
366 else
367 file_update_time(file);
368
369 kernfs_put_active(of->kn);
370 return ret;
371}
372
373static int kernfs_vma_access(struct vm_area_struct *vma, unsigned long addr,
374 void *buf, int len, int write)
375{
376 struct file *file = vma->vm_file;
377 struct kernfs_open_file *of = kernfs_of(file);
378 int ret;
379
380 if (!of->vm_ops)
381 return -EINVAL;
382
383 if (!kernfs_get_active(of->kn))
384 return -EINVAL;
385
386 ret = -EINVAL;
387 if (of->vm_ops->access)
388 ret = of->vm_ops->access(vma, addr, buf, len, write);
389
390 kernfs_put_active(of->kn);
391 return ret;
392}
393
394#ifdef CONFIG_NUMA
395static int kernfs_vma_set_policy(struct vm_area_struct *vma,
396 struct mempolicy *new)
397{
398 struct file *file = vma->vm_file;
399 struct kernfs_open_file *of = kernfs_of(file);
400 int ret;
401
402 if (!of->vm_ops)
403 return 0;
404
405 if (!kernfs_get_active(of->kn))
406 return -EINVAL;
407
408 ret = 0;
409 if (of->vm_ops->set_policy)
410 ret = of->vm_ops->set_policy(vma, new);
411
412 kernfs_put_active(of->kn);
413 return ret;
414}
415
416static struct mempolicy *kernfs_vma_get_policy(struct vm_area_struct *vma,
417 unsigned long addr)
418{
419 struct file *file = vma->vm_file;
420 struct kernfs_open_file *of = kernfs_of(file);
421 struct mempolicy *pol;
422
423 if (!of->vm_ops)
424 return vma->vm_policy;
425
426 if (!kernfs_get_active(of->kn))
427 return vma->vm_policy;
428
429 pol = vma->vm_policy;
430 if (of->vm_ops->get_policy)
431 pol = of->vm_ops->get_policy(vma, addr);
432
433 kernfs_put_active(of->kn);
434 return pol;
435}
436
437#endif
438
439static const struct vm_operations_struct kernfs_vm_ops = {
440 .open = kernfs_vma_open,
441 .fault = kernfs_vma_fault,
442 .page_mkwrite = kernfs_vma_page_mkwrite,
443 .access = kernfs_vma_access,
444#ifdef CONFIG_NUMA
445 .set_policy = kernfs_vma_set_policy,
446 .get_policy = kernfs_vma_get_policy,
447#endif
448};
449
450static int kernfs_fop_mmap(struct file *file, struct vm_area_struct *vma)
451{
452 struct kernfs_open_file *of = kernfs_of(file);
453 const struct kernfs_ops *ops;
454 int rc;
455
456 /*
457 * mmap path and of->mutex are prone to triggering spurious lockdep
458 * warnings and we don't want to add spurious locking dependency
459 * between the two. Check whether mmap is actually implemented
460 * without grabbing @of->mutex by testing HAS_MMAP flag. See the
461 * comment in kernfs_file_open() for more details.
462 */
463 if (!(of->kn->flags & KERNFS_HAS_MMAP))
464 return -ENODEV;
465
466 mutex_lock(&of->mutex);
467
468 rc = -ENODEV;
469 if (!kernfs_get_active(of->kn))
470 goto out_unlock;
471
472 ops = kernfs_ops(of->kn);
473 rc = ops->mmap(of, vma);
474 if (rc)
475 goto out_put;
476
477 /*
478 * PowerPC's pci_mmap of legacy_mem uses shmem_zero_setup()
479 * to satisfy versions of X which crash if the mmap fails: that
480 * substitutes a new vm_file, and we don't then want bin_vm_ops.
481 */
482 if (vma->vm_file != file)
483 goto out_put;
484
485 rc = -EINVAL;
486 if (of->mmapped && of->vm_ops != vma->vm_ops)
487 goto out_put;
488
489 /*
490 * It is not possible to successfully wrap close.
491 * So error if someone is trying to use close.
492 */
493 rc = -EINVAL;
494 if (vma->vm_ops && vma->vm_ops->close)
495 goto out_put;
496
497 rc = 0;
498 of->mmapped = true;
499 of->vm_ops = vma->vm_ops;
500 vma->vm_ops = &kernfs_vm_ops;
501out_put:
502 kernfs_put_active(of->kn);
503out_unlock:
504 mutex_unlock(&of->mutex);
505
506 return rc;
507}
508
509/**
510 * kernfs_get_open_node - get or create kernfs_open_node
511 * @kn: target kernfs_node
512 * @of: kernfs_open_file for this instance of open
513 *
514 * If @kn->attr.open exists, increment its reference count; otherwise,
515 * create one. @of is chained to the files list.
516 *
517 * LOCKING:
518 * Kernel thread context (may sleep).
519 *
520 * RETURNS:
521 * 0 on success, -errno on failure.
522 */
523static int kernfs_get_open_node(struct kernfs_node *kn,
524 struct kernfs_open_file *of)
525{
526 struct kernfs_open_node *on, *new_on = NULL;
527
528 retry:
529 mutex_lock(&kernfs_open_file_mutex);
530 spin_lock_irq(&kernfs_open_node_lock);
531
532 if (!kn->attr.open && new_on) {
533 kn->attr.open = new_on;
534 new_on = NULL;
535 }
536
537 on = kn->attr.open;
538 if (on) {
539 atomic_inc(&on->refcnt);
540 list_add_tail(&of->list, &on->files);
541 }
542
543 spin_unlock_irq(&kernfs_open_node_lock);
544 mutex_unlock(&kernfs_open_file_mutex);
545
546 if (on) {
547 kfree(new_on);
548 return 0;
549 }
550
551 /* not there, initialize a new one and retry */
552 new_on = kmalloc(sizeof(*new_on), GFP_KERNEL);
553 if (!new_on)
554 return -ENOMEM;
555
556 atomic_set(&new_on->refcnt, 0);
557 atomic_set(&new_on->event, 1);
558 init_waitqueue_head(&new_on->poll);
559 INIT_LIST_HEAD(&new_on->files);
560 goto retry;
561}
562
563/**
564 * kernfs_put_open_node - put kernfs_open_node
565 * @kn: target kernfs_nodet
566 * @of: associated kernfs_open_file
567 *
568 * Put @kn->attr.open and unlink @of from the files list. If
569 * reference count reaches zero, disassociate and free it.
570 *
571 * LOCKING:
572 * None.
573 */
574static void kernfs_put_open_node(struct kernfs_node *kn,
575 struct kernfs_open_file *of)
576{
577 struct kernfs_open_node *on = kn->attr.open;
578 unsigned long flags;
579
580 mutex_lock(&kernfs_open_file_mutex);
581 spin_lock_irqsave(&kernfs_open_node_lock, flags);
582
583 if (of)
584 list_del(&of->list);
585
586 if (atomic_dec_and_test(&on->refcnt))
587 kn->attr.open = NULL;
588 else
589 on = NULL;
590
591 spin_unlock_irqrestore(&kernfs_open_node_lock, flags);
592 mutex_unlock(&kernfs_open_file_mutex);
593
594 kfree(on);
595}
596
597static int kernfs_fop_open(struct inode *inode, struct file *file)
598{
599 struct kernfs_node *kn = inode->i_private;
600 struct kernfs_root *root = kernfs_root(kn);
601 const struct kernfs_ops *ops;
602 struct kernfs_open_file *of;
603 bool has_read, has_write, has_mmap;
604 int error = -EACCES;
605
606 if (!kernfs_get_active(kn))
607 return -ENODEV;
608
609 ops = kernfs_ops(kn);
610
611 has_read = ops->seq_show || ops->read || ops->mmap;
612 has_write = ops->write || ops->mmap;
613 has_mmap = ops->mmap;
614
615 /* see the flag definition for details */
616 if (root->flags & KERNFS_ROOT_EXTRA_OPEN_PERM_CHECK) {
617 if ((file->f_mode & FMODE_WRITE) &&
618 (!(inode->i_mode & S_IWUGO) || !has_write))
619 goto err_out;
620
621 if ((file->f_mode & FMODE_READ) &&
622 (!(inode->i_mode & S_IRUGO) || !has_read))
623 goto err_out;
624 }
625
626 /* allocate a kernfs_open_file for the file */
627 error = -ENOMEM;
628 of = kzalloc(sizeof(struct kernfs_open_file), GFP_KERNEL);
629 if (!of)
630 goto err_out;
631
632 /*
633 * The following is done to give a different lockdep key to
634 * @of->mutex for files which implement mmap. This is a rather
635 * crude way to avoid false positive lockdep warning around
636 * mm->mmap_lock - mmap nests @of->mutex under mm->mmap_lock and
637 * reading /sys/block/sda/trace/act_mask grabs sr_mutex, under
638 * which mm->mmap_lock nests, while holding @of->mutex. As each
639 * open file has a separate mutex, it's okay as long as those don't
640 * happen on the same file. At this point, we can't easily give
641 * each file a separate locking class. Let's differentiate on
642 * whether the file has mmap or not for now.
643 *
644 * Both paths of the branch look the same. They're supposed to
645 * look that way and give @of->mutex different static lockdep keys.
646 */
647 if (has_mmap)
648 mutex_init(&of->mutex);
649 else
650 mutex_init(&of->mutex);
651
652 of->kn = kn;
653 of->file = file;
654
655 /*
656 * Write path needs to atomic_write_len outside active reference.
657 * Cache it in open_file. See kernfs_fop_write_iter() for details.
658 */
659 of->atomic_write_len = ops->atomic_write_len;
660
661 error = -EINVAL;
662 /*
663 * ->seq_show is incompatible with ->prealloc,
664 * as seq_read does its own allocation.
665 * ->read must be used instead.
666 */
667 if (ops->prealloc && ops->seq_show)
668 goto err_free;
669 if (ops->prealloc) {
670 int len = of->atomic_write_len ?: PAGE_SIZE;
671 of->prealloc_buf = kmalloc(len + 1, GFP_KERNEL);
672 error = -ENOMEM;
673 if (!of->prealloc_buf)
674 goto err_free;
675 mutex_init(&of->prealloc_mutex);
676 }
677
678 /*
679 * Always instantiate seq_file even if read access doesn't use
680 * seq_file or is not requested. This unifies private data access
681 * and readable regular files are the vast majority anyway.
682 */
683 if (ops->seq_show)
684 error = seq_open(file, &kernfs_seq_ops);
685 else
686 error = seq_open(file, NULL);
687 if (error)
688 goto err_free;
689
690 of->seq_file = file->private_data;
691 of->seq_file->private = of;
692
693 /* seq_file clears PWRITE unconditionally, restore it if WRITE */
694 if (file->f_mode & FMODE_WRITE)
695 file->f_mode |= FMODE_PWRITE;
696
697 /* make sure we have open node struct */
698 error = kernfs_get_open_node(kn, of);
699 if (error)
700 goto err_seq_release;
701
702 if (ops->open) {
703 /* nobody has access to @of yet, skip @of->mutex */
704 error = ops->open(of);
705 if (error)
706 goto err_put_node;
707 }
708
709 /* open succeeded, put active references */
710 kernfs_put_active(kn);
711 return 0;
712
713err_put_node:
714 kernfs_put_open_node(kn, of);
715err_seq_release:
716 seq_release(inode, file);
717err_free:
718 kfree(of->prealloc_buf);
719 kfree(of);
720err_out:
721 kernfs_put_active(kn);
722 return error;
723}
724
725/* used from release/drain to ensure that ->release() is called exactly once */
726static void kernfs_release_file(struct kernfs_node *kn,
727 struct kernfs_open_file *of)
728{
729 /*
730 * @of is guaranteed to have no other file operations in flight and
731 * we just want to synchronize release and drain paths.
732 * @kernfs_open_file_mutex is enough. @of->mutex can't be used
733 * here because drain path may be called from places which can
734 * cause circular dependency.
735 */
736 lockdep_assert_held(&kernfs_open_file_mutex);
737
738 if (!of->released) {
739 /*
740 * A file is never detached without being released and we
741 * need to be able to release files which are deactivated
742 * and being drained. Don't use kernfs_ops().
743 */
744 kn->attr.ops->release(of);
745 of->released = true;
746 }
747}
748
749static int kernfs_fop_release(struct inode *inode, struct file *filp)
750{
751 struct kernfs_node *kn = inode->i_private;
752 struct kernfs_open_file *of = kernfs_of(filp);
753
754 if (kn->flags & KERNFS_HAS_RELEASE) {
755 mutex_lock(&kernfs_open_file_mutex);
756 kernfs_release_file(kn, of);
757 mutex_unlock(&kernfs_open_file_mutex);
758 }
759
760 kernfs_put_open_node(kn, of);
761 seq_release(inode, filp);
762 kfree(of->prealloc_buf);
763 kfree(of);
764
765 return 0;
766}
767
768void kernfs_drain_open_files(struct kernfs_node *kn)
769{
770 struct kernfs_open_node *on;
771 struct kernfs_open_file *of;
772
773 if (!(kn->flags & (KERNFS_HAS_MMAP | KERNFS_HAS_RELEASE)))
774 return;
775
776 spin_lock_irq(&kernfs_open_node_lock);
777 on = kn->attr.open;
778 if (on)
779 atomic_inc(&on->refcnt);
780 spin_unlock_irq(&kernfs_open_node_lock);
781 if (!on)
782 return;
783
784 mutex_lock(&kernfs_open_file_mutex);
785
786 list_for_each_entry(of, &on->files, list) {
787 struct inode *inode = file_inode(of->file);
788
789 if (kn->flags & KERNFS_HAS_MMAP)
790 unmap_mapping_range(inode->i_mapping, 0, 0, 1);
791
792 if (kn->flags & KERNFS_HAS_RELEASE)
793 kernfs_release_file(kn, of);
794 }
795
796 mutex_unlock(&kernfs_open_file_mutex);
797
798 kernfs_put_open_node(kn, NULL);
799}
800
801/*
802 * Kernfs attribute files are pollable. The idea is that you read
803 * the content and then you use 'poll' or 'select' to wait for
804 * the content to change. When the content changes (assuming the
805 * manager for the kobject supports notification), poll will
806 * return EPOLLERR|EPOLLPRI, and select will return the fd whether
807 * it is waiting for read, write, or exceptions.
808 * Once poll/select indicates that the value has changed, you
809 * need to close and re-open the file, or seek to 0 and read again.
810 * Reminder: this only works for attributes which actively support
811 * it, and it is not possible to test an attribute from userspace
812 * to see if it supports poll (Neither 'poll' nor 'select' return
813 * an appropriate error code). When in doubt, set a suitable timeout value.
814 */
815__poll_t kernfs_generic_poll(struct kernfs_open_file *of, poll_table *wait)
816{
817 struct kernfs_node *kn = kernfs_dentry_node(of->file->f_path.dentry);
818 struct kernfs_open_node *on = kn->attr.open;
819
820 poll_wait(of->file, &on->poll, wait);
821
822 if (of->event != atomic_read(&on->event))
823 return DEFAULT_POLLMASK|EPOLLERR|EPOLLPRI;
824
825 return DEFAULT_POLLMASK;
826}
827
828static __poll_t kernfs_fop_poll(struct file *filp, poll_table *wait)
829{
830 struct kernfs_open_file *of = kernfs_of(filp);
831 struct kernfs_node *kn = kernfs_dentry_node(filp->f_path.dentry);
832 __poll_t ret;
833
834 if (!kernfs_get_active(kn))
835 return DEFAULT_POLLMASK|EPOLLERR|EPOLLPRI;
836
837 if (kn->attr.ops->poll)
838 ret = kn->attr.ops->poll(of, wait);
839 else
840 ret = kernfs_generic_poll(of, wait);
841
842 kernfs_put_active(kn);
843 return ret;
844}
845
846static void kernfs_notify_workfn(struct work_struct *work)
847{
848 struct kernfs_node *kn;
849 struct kernfs_super_info *info;
850repeat:
851 /* pop one off the notify_list */
852 spin_lock_irq(&kernfs_notify_lock);
853 kn = kernfs_notify_list;
854 if (kn == KERNFS_NOTIFY_EOL) {
855 spin_unlock_irq(&kernfs_notify_lock);
856 return;
857 }
858 kernfs_notify_list = kn->attr.notify_next;
859 kn->attr.notify_next = NULL;
860 spin_unlock_irq(&kernfs_notify_lock);
861
862 /* kick fsnotify */
863 mutex_lock(&kernfs_mutex);
864
865 list_for_each_entry(info, &kernfs_root(kn)->supers, node) {
866 struct kernfs_node *parent;
867 struct inode *p_inode = NULL;
868 struct inode *inode;
869 struct qstr name;
870
871 /*
872 * We want fsnotify_modify() on @kn but as the
873 * modifications aren't originating from userland don't
874 * have the matching @file available. Look up the inodes
875 * and generate the events manually.
876 */
877 inode = ilookup(info->sb, kernfs_ino(kn));
878 if (!inode)
879 continue;
880
881 name = (struct qstr)QSTR_INIT(kn->name, strlen(kn->name));
882 parent = kernfs_get_parent(kn);
883 if (parent) {
884 p_inode = ilookup(info->sb, kernfs_ino(parent));
885 if (p_inode) {
886 fsnotify(FS_MODIFY | FS_EVENT_ON_CHILD,
887 inode, FSNOTIFY_EVENT_INODE,
888 p_inode, &name, inode, 0);
889 iput(p_inode);
890 }
891
892 kernfs_put(parent);
893 }
894
895 if (!p_inode)
896 fsnotify_inode(inode, FS_MODIFY);
897
898 iput(inode);
899 }
900
901 mutex_unlock(&kernfs_mutex);
902 kernfs_put(kn);
903 goto repeat;
904}
905
906/**
907 * kernfs_notify - notify a kernfs file
908 * @kn: file to notify
909 *
910 * Notify @kn such that poll(2) on @kn wakes up. Maybe be called from any
911 * context.
912 */
913void kernfs_notify(struct kernfs_node *kn)
914{
915 static DECLARE_WORK(kernfs_notify_work, kernfs_notify_workfn);
916 unsigned long flags;
917 struct kernfs_open_node *on;
918
919 if (WARN_ON(kernfs_type(kn) != KERNFS_FILE))
920 return;
921
922 /* kick poll immediately */
923 spin_lock_irqsave(&kernfs_open_node_lock, flags);
924 on = kn->attr.open;
925 if (on) {
926 atomic_inc(&on->event);
927 wake_up_interruptible(&on->poll);
928 }
929 spin_unlock_irqrestore(&kernfs_open_node_lock, flags);
930
931 /* schedule work to kick fsnotify */
932 spin_lock_irqsave(&kernfs_notify_lock, flags);
933 if (!kn->attr.notify_next) {
934 kernfs_get(kn);
935 kn->attr.notify_next = kernfs_notify_list;
936 kernfs_notify_list = kn;
937 schedule_work(&kernfs_notify_work);
938 }
939 spin_unlock_irqrestore(&kernfs_notify_lock, flags);
940}
941EXPORT_SYMBOL_GPL(kernfs_notify);
942
943const struct file_operations kernfs_file_fops = {
944 .read_iter = kernfs_fop_read_iter,
945 .write_iter = kernfs_fop_write_iter,
946 .llseek = generic_file_llseek,
947 .mmap = kernfs_fop_mmap,
948 .open = kernfs_fop_open,
949 .release = kernfs_fop_release,
950 .poll = kernfs_fop_poll,
951 .fsync = noop_fsync,
952 .splice_read = generic_file_splice_read,
953 .splice_write = iter_file_splice_write,
954};
955
956/**
957 * __kernfs_create_file - kernfs internal function to create a file
958 * @parent: directory to create the file in
959 * @name: name of the file
960 * @mode: mode of the file
961 * @uid: uid of the file
962 * @gid: gid of the file
963 * @size: size of the file
964 * @ops: kernfs operations for the file
965 * @priv: private data for the file
966 * @ns: optional namespace tag of the file
967 * @key: lockdep key for the file's active_ref, %NULL to disable lockdep
968 *
969 * Returns the created node on success, ERR_PTR() value on error.
970 */
971struct kernfs_node *__kernfs_create_file(struct kernfs_node *parent,
972 const char *name,
973 umode_t mode, kuid_t uid, kgid_t gid,
974 loff_t size,
975 const struct kernfs_ops *ops,
976 void *priv, const void *ns,
977 struct lock_class_key *key)
978{
979 struct kernfs_node *kn;
980 unsigned flags;
981 int rc;
982
983 flags = KERNFS_FILE;
984
985 kn = kernfs_new_node(parent, name, (mode & S_IALLUGO) | S_IFREG,
986 uid, gid, flags);
987 if (!kn)
988 return ERR_PTR(-ENOMEM);
989
990 kn->attr.ops = ops;
991 kn->attr.size = size;
992 kn->ns = ns;
993 kn->priv = priv;
994
995#ifdef CONFIG_DEBUG_LOCK_ALLOC
996 if (key) {
997 lockdep_init_map(&kn->dep_map, "kn->active", key, 0);
998 kn->flags |= KERNFS_LOCKDEP;
999 }
1000#endif
1001
1002 /*
1003 * kn->attr.ops is accesible only while holding active ref. We
1004 * need to know whether some ops are implemented outside active
1005 * ref. Cache their existence in flags.
1006 */
1007 if (ops->seq_show)
1008 kn->flags |= KERNFS_HAS_SEQ_SHOW;
1009 if (ops->mmap)
1010 kn->flags |= KERNFS_HAS_MMAP;
1011 if (ops->release)
1012 kn->flags |= KERNFS_HAS_RELEASE;
1013
1014 rc = kernfs_add_one(kn);
1015 if (rc) {
1016 kernfs_put(kn);
1017 return ERR_PTR(rc);
1018 }
1019 return kn;
1020}