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