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