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