xref: /openbmc/linux/fs/kernfs/file.c (revision 1fb112ce)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * fs/kernfs/file.c - kernfs file implementation
4  *
5  * Copyright (c) 2001-3 Patrick Mochel
6  * Copyright (c) 2007 SUSE Linux Products GmbH
7  * Copyright (c) 2007, 2013 Tejun Heo <tj@kernel.org>
8  */
9 
10 #include <linux/fs.h>
11 #include <linux/seq_file.h>
12 #include <linux/slab.h>
13 #include <linux/poll.h>
14 #include <linux/pagemap.h>
15 #include <linux/sched/mm.h>
16 #include <linux/fsnotify.h>
17 #include <linux/uio.h>
18 
19 #include "kernfs-internal.h"
20 
21 struct kernfs_open_node {
22 	struct rcu_head		rcu_head;
23 	atomic_t		event;
24 	wait_queue_head_t	poll;
25 	struct list_head	files; /* goes through kernfs_open_file.list */
26 	unsigned int		nr_mmapped;
27 	unsigned int		nr_to_release;
28 };
29 
30 /*
31  * kernfs_notify() may be called from any context and bounces notifications
32  * through a work item.  To minimize space overhead in kernfs_node, the
33  * pending queue is implemented as a singly linked list of kernfs_nodes.
34  * The list is terminated with the self pointer so that whether a
35  * kernfs_node is on the list or not can be determined by testing the next
36  * pointer for %NULL.
37  */
38 #define KERNFS_NOTIFY_EOL			((void *)&kernfs_notify_list)
39 
40 static DEFINE_SPINLOCK(kernfs_notify_lock);
41 static struct kernfs_node *kernfs_notify_list = KERNFS_NOTIFY_EOL;
42 
kernfs_open_file_mutex_ptr(struct kernfs_node * kn)43 static inline struct mutex *kernfs_open_file_mutex_ptr(struct kernfs_node *kn)
44 {
45 	int idx = hash_ptr(kn, NR_KERNFS_LOCK_BITS);
46 
47 	return &kernfs_locks->open_file_mutex[idx];
48 }
49 
kernfs_open_file_mutex_lock(struct kernfs_node * kn)50 static inline struct mutex *kernfs_open_file_mutex_lock(struct kernfs_node *kn)
51 {
52 	struct mutex *lock;
53 
54 	lock = kernfs_open_file_mutex_ptr(kn);
55 
56 	mutex_lock(lock);
57 
58 	return lock;
59 }
60 
61 /**
62  * of_on - Get the kernfs_open_node of the specified kernfs_open_file
63  * @of: target kernfs_open_file
64  *
65  * Return: the kernfs_open_node of the kernfs_open_file
66  */
of_on(struct kernfs_open_file * of)67 static struct kernfs_open_node *of_on(struct kernfs_open_file *of)
68 {
69 	return rcu_dereference_protected(of->kn->attr.open,
70 					 !list_empty(&of->list));
71 }
72 
73 /**
74  * kernfs_deref_open_node_locked - Get kernfs_open_node corresponding to @kn
75  *
76  * @kn: target kernfs_node.
77  *
78  * Fetch and return ->attr.open of @kn when caller holds the
79  * kernfs_open_file_mutex_ptr(kn).
80  *
81  * Update of ->attr.open happens under kernfs_open_file_mutex_ptr(kn). So when
82  * the caller guarantees that this mutex is being held, other updaters can't
83  * change ->attr.open and this means that we can safely deref ->attr.open
84  * outside RCU read-side critical section.
85  *
86  * The caller needs to make sure that kernfs_open_file_mutex is held.
87  *
88  * Return: @kn->attr.open when kernfs_open_file_mutex is held.
89  */
90 static struct kernfs_open_node *
kernfs_deref_open_node_locked(struct kernfs_node * kn)91 kernfs_deref_open_node_locked(struct kernfs_node *kn)
92 {
93 	return rcu_dereference_protected(kn->attr.open,
94 				lockdep_is_held(kernfs_open_file_mutex_ptr(kn)));
95 }
96 
kernfs_of(struct file * file)97 static struct kernfs_open_file *kernfs_of(struct file *file)
98 {
99 	return ((struct seq_file *)file->private_data)->private;
100 }
101 
102 /*
103  * Determine the kernfs_ops for the given kernfs_node.  This function must
104  * be called while holding an active reference.
105  */
kernfs_ops(struct kernfs_node * kn)106 static const struct kernfs_ops *kernfs_ops(struct kernfs_node *kn)
107 {
108 	if (kn->flags & KERNFS_LOCKDEP)
109 		lockdep_assert_held(kn);
110 	return kn->attr.ops;
111 }
112 
113 /*
114  * As kernfs_seq_stop() is also called after kernfs_seq_start() or
115  * kernfs_seq_next() failure, it needs to distinguish whether it's stopping
116  * a seq_file iteration which is fully initialized with an active reference
117  * or an aborted kernfs_seq_start() due to get_active failure.  The
118  * position pointer is the only context for each seq_file iteration and
119  * thus the stop condition should be encoded in it.  As the return value is
120  * directly visible to userland, ERR_PTR(-ENODEV) is the only acceptable
121  * choice to indicate get_active failure.
122  *
123  * Unfortunately, this is complicated due to the optional custom seq_file
124  * operations which may return ERR_PTR(-ENODEV) too.  kernfs_seq_stop()
125  * can't distinguish whether ERR_PTR(-ENODEV) is from get_active failure or
126  * custom seq_file operations and thus can't decide whether put_active
127  * should be performed or not only on ERR_PTR(-ENODEV).
128  *
129  * This is worked around by factoring out the custom seq_stop() and
130  * put_active part into kernfs_seq_stop_active(), skipping it from
131  * kernfs_seq_stop() if ERR_PTR(-ENODEV) while invoking it directly after
132  * custom seq_file operations fail with ERR_PTR(-ENODEV) - this ensures
133  * that kernfs_seq_stop_active() is skipped only after get_active failure.
134  */
kernfs_seq_stop_active(struct seq_file * sf,void * v)135 static void kernfs_seq_stop_active(struct seq_file *sf, void *v)
136 {
137 	struct kernfs_open_file *of = sf->private;
138 	const struct kernfs_ops *ops = kernfs_ops(of->kn);
139 
140 	if (ops->seq_stop)
141 		ops->seq_stop(sf, v);
142 	kernfs_put_active(of->kn);
143 }
144 
kernfs_seq_start(struct seq_file * sf,loff_t * ppos)145 static void *kernfs_seq_start(struct seq_file *sf, loff_t *ppos)
146 {
147 	struct kernfs_open_file *of = sf->private;
148 	const struct kernfs_ops *ops;
149 
150 	/*
151 	 * @of->mutex nests outside active ref and is primarily to ensure that
152 	 * the ops aren't called concurrently for the same open file.
153 	 */
154 	mutex_lock(&of->mutex);
155 	if (!kernfs_get_active(of->kn))
156 		return ERR_PTR(-ENODEV);
157 
158 	ops = kernfs_ops(of->kn);
159 	if (ops->seq_start) {
160 		void *next = ops->seq_start(sf, ppos);
161 		/* see the comment above kernfs_seq_stop_active() */
162 		if (next == ERR_PTR(-ENODEV))
163 			kernfs_seq_stop_active(sf, next);
164 		return next;
165 	}
166 	return single_start(sf, ppos);
167 }
168 
kernfs_seq_next(struct seq_file * sf,void * v,loff_t * ppos)169 static void *kernfs_seq_next(struct seq_file *sf, void *v, loff_t *ppos)
170 {
171 	struct kernfs_open_file *of = sf->private;
172 	const struct kernfs_ops *ops = kernfs_ops(of->kn);
173 
174 	if (ops->seq_next) {
175 		void *next = ops->seq_next(sf, v, ppos);
176 		/* see the comment above kernfs_seq_stop_active() */
177 		if (next == ERR_PTR(-ENODEV))
178 			kernfs_seq_stop_active(sf, next);
179 		return next;
180 	} else {
181 		/*
182 		 * The same behavior and code as single_open(), always
183 		 * terminate after the initial read.
184 		 */
185 		++*ppos;
186 		return NULL;
187 	}
188 }
189 
kernfs_seq_stop(struct seq_file * sf,void * v)190 static void kernfs_seq_stop(struct seq_file *sf, void *v)
191 {
192 	struct kernfs_open_file *of = sf->private;
193 
194 	if (v != ERR_PTR(-ENODEV))
195 		kernfs_seq_stop_active(sf, v);
196 	mutex_unlock(&of->mutex);
197 }
198 
kernfs_seq_show(struct seq_file * sf,void * v)199 static int kernfs_seq_show(struct seq_file *sf, void *v)
200 {
201 	struct kernfs_open_file *of = sf->private;
202 
203 	of->event = atomic_read(&of_on(of)->event);
204 
205 	return of->kn->attr.ops->seq_show(sf, v);
206 }
207 
208 static const struct seq_operations kernfs_seq_ops = {
209 	.start = kernfs_seq_start,
210 	.next = kernfs_seq_next,
211 	.stop = kernfs_seq_stop,
212 	.show = kernfs_seq_show,
213 };
214 
215 /*
216  * As reading a bin file can have side-effects, the exact offset and bytes
217  * specified in read(2) call should be passed to the read callback making
218  * it difficult to use seq_file.  Implement simplistic custom buffering for
219  * bin files.
220  */
kernfs_file_read_iter(struct kiocb * iocb,struct iov_iter * iter)221 static ssize_t kernfs_file_read_iter(struct kiocb *iocb, struct iov_iter *iter)
222 {
223 	struct kernfs_open_file *of = kernfs_of(iocb->ki_filp);
224 	ssize_t len = min_t(size_t, iov_iter_count(iter), PAGE_SIZE);
225 	const struct kernfs_ops *ops;
226 	char *buf;
227 
228 	buf = of->prealloc_buf;
229 	if (buf)
230 		mutex_lock(&of->prealloc_mutex);
231 	else
232 		buf = kmalloc(len, GFP_KERNEL);
233 	if (!buf)
234 		return -ENOMEM;
235 
236 	/*
237 	 * @of->mutex nests outside active ref and is used both to ensure that
238 	 * the ops aren't called concurrently for the same open file.
239 	 */
240 	mutex_lock(&of->mutex);
241 	if (!kernfs_get_active(of->kn)) {
242 		len = -ENODEV;
243 		mutex_unlock(&of->mutex);
244 		goto out_free;
245 	}
246 
247 	of->event = atomic_read(&of_on(of)->event);
248 
249 	ops = kernfs_ops(of->kn);
250 	if (ops->read)
251 		len = ops->read(of, buf, len, iocb->ki_pos);
252 	else
253 		len = -EINVAL;
254 
255 	kernfs_put_active(of->kn);
256 	mutex_unlock(&of->mutex);
257 
258 	if (len < 0)
259 		goto out_free;
260 
261 	if (copy_to_iter(buf, len, iter) != len) {
262 		len = -EFAULT;
263 		goto out_free;
264 	}
265 
266 	iocb->ki_pos += len;
267 
268  out_free:
269 	if (buf == of->prealloc_buf)
270 		mutex_unlock(&of->prealloc_mutex);
271 	else
272 		kfree(buf);
273 	return len;
274 }
275 
kernfs_fop_read_iter(struct kiocb * iocb,struct iov_iter * iter)276 static ssize_t kernfs_fop_read_iter(struct kiocb *iocb, struct iov_iter *iter)
277 {
278 	if (kernfs_of(iocb->ki_filp)->kn->flags & KERNFS_HAS_SEQ_SHOW)
279 		return seq_read_iter(iocb, iter);
280 	return kernfs_file_read_iter(iocb, iter);
281 }
282 
283 /*
284  * Copy data in from userland and pass it to the matching kernfs write
285  * operation.
286  *
287  * There is no easy way for us to know if userspace is only doing a partial
288  * write, so we don't support them. We expect the entire buffer to come on
289  * the first write.  Hint: if you're writing a value, first read the file,
290  * modify only the value you're changing, then write entire buffer
291  * back.
292  */
kernfs_fop_write_iter(struct kiocb * iocb,struct iov_iter * iter)293 static ssize_t kernfs_fop_write_iter(struct kiocb *iocb, struct iov_iter *iter)
294 {
295 	struct kernfs_open_file *of = kernfs_of(iocb->ki_filp);
296 	ssize_t len = iov_iter_count(iter);
297 	const struct kernfs_ops *ops;
298 	char *buf;
299 
300 	if (of->atomic_write_len) {
301 		if (len > of->atomic_write_len)
302 			return -E2BIG;
303 	} else {
304 		len = min_t(size_t, len, PAGE_SIZE);
305 	}
306 
307 	buf = of->prealloc_buf;
308 	if (buf)
309 		mutex_lock(&of->prealloc_mutex);
310 	else
311 		buf = kmalloc(len + 1, GFP_KERNEL);
312 	if (!buf)
313 		return -ENOMEM;
314 
315 	if (copy_from_iter(buf, len, iter) != len) {
316 		len = -EFAULT;
317 		goto out_free;
318 	}
319 	buf[len] = '\0';	/* guarantee string termination */
320 
321 	/*
322 	 * @of->mutex nests outside active ref and is used both to ensure that
323 	 * the ops aren't called concurrently for the same open file.
324 	 */
325 	mutex_lock(&of->mutex);
326 	if (!kernfs_get_active(of->kn)) {
327 		mutex_unlock(&of->mutex);
328 		len = -ENODEV;
329 		goto out_free;
330 	}
331 
332 	ops = kernfs_ops(of->kn);
333 	if (ops->write)
334 		len = ops->write(of, buf, len, iocb->ki_pos);
335 	else
336 		len = -EINVAL;
337 
338 	kernfs_put_active(of->kn);
339 	mutex_unlock(&of->mutex);
340 
341 	if (len > 0)
342 		iocb->ki_pos += len;
343 
344 out_free:
345 	if (buf == of->prealloc_buf)
346 		mutex_unlock(&of->prealloc_mutex);
347 	else
348 		kfree(buf);
349 	return len;
350 }
351 
kernfs_vma_open(struct vm_area_struct * vma)352 static void kernfs_vma_open(struct vm_area_struct *vma)
353 {
354 	struct file *file = vma->vm_file;
355 	struct kernfs_open_file *of = kernfs_of(file);
356 
357 	if (!of->vm_ops)
358 		return;
359 
360 	if (!kernfs_get_active(of->kn))
361 		return;
362 
363 	if (of->vm_ops->open)
364 		of->vm_ops->open(vma);
365 
366 	kernfs_put_active(of->kn);
367 }
368 
kernfs_vma_fault(struct vm_fault * vmf)369 static vm_fault_t kernfs_vma_fault(struct vm_fault *vmf)
370 {
371 	struct file *file = vmf->vma->vm_file;
372 	struct kernfs_open_file *of = kernfs_of(file);
373 	vm_fault_t ret;
374 
375 	if (!of->vm_ops)
376 		return VM_FAULT_SIGBUS;
377 
378 	if (!kernfs_get_active(of->kn))
379 		return VM_FAULT_SIGBUS;
380 
381 	ret = VM_FAULT_SIGBUS;
382 	if (of->vm_ops->fault)
383 		ret = of->vm_ops->fault(vmf);
384 
385 	kernfs_put_active(of->kn);
386 	return ret;
387 }
388 
kernfs_vma_page_mkwrite(struct vm_fault * vmf)389 static vm_fault_t kernfs_vma_page_mkwrite(struct vm_fault *vmf)
390 {
391 	struct file *file = vmf->vma->vm_file;
392 	struct kernfs_open_file *of = kernfs_of(file);
393 	vm_fault_t ret;
394 
395 	if (!of->vm_ops)
396 		return VM_FAULT_SIGBUS;
397 
398 	if (!kernfs_get_active(of->kn))
399 		return VM_FAULT_SIGBUS;
400 
401 	ret = 0;
402 	if (of->vm_ops->page_mkwrite)
403 		ret = of->vm_ops->page_mkwrite(vmf);
404 	else
405 		file_update_time(file);
406 
407 	kernfs_put_active(of->kn);
408 	return ret;
409 }
410 
kernfs_vma_access(struct vm_area_struct * vma,unsigned long addr,void * buf,int len,int write)411 static int kernfs_vma_access(struct vm_area_struct *vma, unsigned long addr,
412 			     void *buf, int len, int write)
413 {
414 	struct file *file = vma->vm_file;
415 	struct kernfs_open_file *of = kernfs_of(file);
416 	int ret;
417 
418 	if (!of->vm_ops)
419 		return -EINVAL;
420 
421 	if (!kernfs_get_active(of->kn))
422 		return -EINVAL;
423 
424 	ret = -EINVAL;
425 	if (of->vm_ops->access)
426 		ret = of->vm_ops->access(vma, addr, buf, len, write);
427 
428 	kernfs_put_active(of->kn);
429 	return ret;
430 }
431 
432 #ifdef CONFIG_NUMA
kernfs_vma_set_policy(struct vm_area_struct * vma,struct mempolicy * new)433 static int kernfs_vma_set_policy(struct vm_area_struct *vma,
434 				 struct mempolicy *new)
435 {
436 	struct file *file = vma->vm_file;
437 	struct kernfs_open_file *of = kernfs_of(file);
438 	int ret;
439 
440 	if (!of->vm_ops)
441 		return 0;
442 
443 	if (!kernfs_get_active(of->kn))
444 		return -EINVAL;
445 
446 	ret = 0;
447 	if (of->vm_ops->set_policy)
448 		ret = of->vm_ops->set_policy(vma, new);
449 
450 	kernfs_put_active(of->kn);
451 	return ret;
452 }
453 
kernfs_vma_get_policy(struct vm_area_struct * vma,unsigned long addr)454 static struct mempolicy *kernfs_vma_get_policy(struct vm_area_struct *vma,
455 					       unsigned long addr)
456 {
457 	struct file *file = vma->vm_file;
458 	struct kernfs_open_file *of = kernfs_of(file);
459 	struct mempolicy *pol;
460 
461 	if (!of->vm_ops)
462 		return vma->vm_policy;
463 
464 	if (!kernfs_get_active(of->kn))
465 		return vma->vm_policy;
466 
467 	pol = vma->vm_policy;
468 	if (of->vm_ops->get_policy)
469 		pol = of->vm_ops->get_policy(vma, addr);
470 
471 	kernfs_put_active(of->kn);
472 	return pol;
473 }
474 
475 #endif
476 
477 static const struct vm_operations_struct kernfs_vm_ops = {
478 	.open		= kernfs_vma_open,
479 	.fault		= kernfs_vma_fault,
480 	.page_mkwrite	= kernfs_vma_page_mkwrite,
481 	.access		= kernfs_vma_access,
482 #ifdef CONFIG_NUMA
483 	.set_policy	= kernfs_vma_set_policy,
484 	.get_policy	= kernfs_vma_get_policy,
485 #endif
486 };
487 
kernfs_fop_mmap(struct file * file,struct vm_area_struct * vma)488 static int kernfs_fop_mmap(struct file *file, struct vm_area_struct *vma)
489 {
490 	struct kernfs_open_file *of = kernfs_of(file);
491 	const struct kernfs_ops *ops;
492 	int rc;
493 
494 	/*
495 	 * mmap path and of->mutex are prone to triggering spurious lockdep
496 	 * warnings and we don't want to add spurious locking dependency
497 	 * between the two.  Check whether mmap is actually implemented
498 	 * without grabbing @of->mutex by testing HAS_MMAP flag.  See the
499 	 * comment in kernfs_file_open() for more details.
500 	 */
501 	if (!(of->kn->flags & KERNFS_HAS_MMAP))
502 		return -ENODEV;
503 
504 	mutex_lock(&of->mutex);
505 
506 	rc = -ENODEV;
507 	if (!kernfs_get_active(of->kn))
508 		goto out_unlock;
509 
510 	ops = kernfs_ops(of->kn);
511 	rc = ops->mmap(of, vma);
512 	if (rc)
513 		goto out_put;
514 
515 	/*
516 	 * PowerPC's pci_mmap of legacy_mem uses shmem_zero_setup()
517 	 * to satisfy versions of X which crash if the mmap fails: that
518 	 * substitutes a new vm_file, and we don't then want bin_vm_ops.
519 	 */
520 	if (vma->vm_file != file)
521 		goto out_put;
522 
523 	rc = -EINVAL;
524 	if (of->mmapped && of->vm_ops != vma->vm_ops)
525 		goto out_put;
526 
527 	/*
528 	 * It is not possible to successfully wrap close.
529 	 * So error if someone is trying to use close.
530 	 */
531 	if (vma->vm_ops && vma->vm_ops->close)
532 		goto out_put;
533 
534 	rc = 0;
535 	if (!of->mmapped) {
536 		of->mmapped = true;
537 		of_on(of)->nr_mmapped++;
538 		of->vm_ops = vma->vm_ops;
539 	}
540 	vma->vm_ops = &kernfs_vm_ops;
541 out_put:
542 	kernfs_put_active(of->kn);
543 out_unlock:
544 	mutex_unlock(&of->mutex);
545 
546 	return rc;
547 }
548 
549 /**
550  *	kernfs_get_open_node - get or create kernfs_open_node
551  *	@kn: target kernfs_node
552  *	@of: kernfs_open_file for this instance of open
553  *
554  *	If @kn->attr.open exists, increment its reference count; otherwise,
555  *	create one.  @of is chained to the files list.
556  *
557  *	Locking:
558  *	Kernel thread context (may sleep).
559  *
560  *	Return:
561  *	%0 on success, -errno on failure.
562  */
kernfs_get_open_node(struct kernfs_node * kn,struct kernfs_open_file * of)563 static int kernfs_get_open_node(struct kernfs_node *kn,
564 				struct kernfs_open_file *of)
565 {
566 	struct kernfs_open_node *on;
567 	struct mutex *mutex;
568 
569 	mutex = kernfs_open_file_mutex_lock(kn);
570 	on = kernfs_deref_open_node_locked(kn);
571 
572 	if (!on) {
573 		/* not there, initialize a new one */
574 		on = kzalloc(sizeof(*on), GFP_KERNEL);
575 		if (!on) {
576 			mutex_unlock(mutex);
577 			return -ENOMEM;
578 		}
579 		atomic_set(&on->event, 1);
580 		init_waitqueue_head(&on->poll);
581 		INIT_LIST_HEAD(&on->files);
582 		rcu_assign_pointer(kn->attr.open, on);
583 	}
584 
585 	list_add_tail(&of->list, &on->files);
586 	if (kn->flags & KERNFS_HAS_RELEASE)
587 		on->nr_to_release++;
588 
589 	mutex_unlock(mutex);
590 	return 0;
591 }
592 
593 /**
594  *	kernfs_unlink_open_file - Unlink @of from @kn.
595  *
596  *	@kn: target kernfs_node
597  *	@of: associated kernfs_open_file
598  *	@open_failed: ->open() failed, cancel ->release()
599  *
600  *	Unlink @of from list of @kn's associated open files. If list of
601  *	associated open files becomes empty, disassociate and free
602  *	kernfs_open_node.
603  *
604  *	LOCKING:
605  *	None.
606  */
kernfs_unlink_open_file(struct kernfs_node * kn,struct kernfs_open_file * of,bool open_failed)607 static void kernfs_unlink_open_file(struct kernfs_node *kn,
608 				    struct kernfs_open_file *of,
609 				    bool open_failed)
610 {
611 	struct kernfs_open_node *on;
612 	struct mutex *mutex;
613 
614 	mutex = kernfs_open_file_mutex_lock(kn);
615 
616 	on = kernfs_deref_open_node_locked(kn);
617 	if (!on) {
618 		mutex_unlock(mutex);
619 		return;
620 	}
621 
622 	if (of) {
623 		if (kn->flags & KERNFS_HAS_RELEASE) {
624 			WARN_ON_ONCE(of->released == open_failed);
625 			if (open_failed)
626 				on->nr_to_release--;
627 		}
628 		if (of->mmapped)
629 			on->nr_mmapped--;
630 		list_del(&of->list);
631 	}
632 
633 	if (list_empty(&on->files)) {
634 		rcu_assign_pointer(kn->attr.open, NULL);
635 		kfree_rcu(on, rcu_head);
636 	}
637 
638 	mutex_unlock(mutex);
639 }
640 
kernfs_fop_open(struct inode * inode,struct file * file)641 static int kernfs_fop_open(struct inode *inode, struct file *file)
642 {
643 	struct kernfs_node *kn = inode->i_private;
644 	struct kernfs_root *root = kernfs_root(kn);
645 	const struct kernfs_ops *ops;
646 	struct kernfs_open_file *of;
647 	bool has_read, has_write, has_mmap;
648 	int error = -EACCES;
649 
650 	if (!kernfs_get_active(kn))
651 		return -ENODEV;
652 
653 	ops = kernfs_ops(kn);
654 
655 	has_read = ops->seq_show || ops->read || ops->mmap;
656 	has_write = ops->write || ops->mmap;
657 	has_mmap = ops->mmap;
658 
659 	/* see the flag definition for details */
660 	if (root->flags & KERNFS_ROOT_EXTRA_OPEN_PERM_CHECK) {
661 		if ((file->f_mode & FMODE_WRITE) &&
662 		    (!(inode->i_mode & S_IWUGO) || !has_write))
663 			goto err_out;
664 
665 		if ((file->f_mode & FMODE_READ) &&
666 		    (!(inode->i_mode & S_IRUGO) || !has_read))
667 			goto err_out;
668 	}
669 
670 	/* allocate a kernfs_open_file for the file */
671 	error = -ENOMEM;
672 	of = kzalloc(sizeof(struct kernfs_open_file), GFP_KERNEL);
673 	if (!of)
674 		goto err_out;
675 
676 	/*
677 	 * The following is done to give a different lockdep key to
678 	 * @of->mutex for files which implement mmap.  This is a rather
679 	 * crude way to avoid false positive lockdep warning around
680 	 * mm->mmap_lock - mmap nests @of->mutex under mm->mmap_lock and
681 	 * reading /sys/block/sda/trace/act_mask grabs sr_mutex, under
682 	 * which mm->mmap_lock nests, while holding @of->mutex.  As each
683 	 * open file has a separate mutex, it's okay as long as those don't
684 	 * happen on the same file.  At this point, we can't easily give
685 	 * each file a separate locking class.  Let's differentiate on
686 	 * whether the file has mmap or not for now.
687 	 *
688 	 * Both paths of the branch look the same.  They're supposed to
689 	 * look that way and give @of->mutex different static lockdep keys.
690 	 */
691 	if (has_mmap)
692 		mutex_init(&of->mutex);
693 	else
694 		mutex_init(&of->mutex);
695 
696 	of->kn = kn;
697 	of->file = file;
698 
699 	/*
700 	 * Write path needs to atomic_write_len outside active reference.
701 	 * Cache it in open_file.  See kernfs_fop_write_iter() for details.
702 	 */
703 	of->atomic_write_len = ops->atomic_write_len;
704 
705 	error = -EINVAL;
706 	/*
707 	 * ->seq_show is incompatible with ->prealloc,
708 	 * as seq_read does its own allocation.
709 	 * ->read must be used instead.
710 	 */
711 	if (ops->prealloc && ops->seq_show)
712 		goto err_free;
713 	if (ops->prealloc) {
714 		int len = of->atomic_write_len ?: PAGE_SIZE;
715 		of->prealloc_buf = kmalloc(len + 1, GFP_KERNEL);
716 		error = -ENOMEM;
717 		if (!of->prealloc_buf)
718 			goto err_free;
719 		mutex_init(&of->prealloc_mutex);
720 	}
721 
722 	/*
723 	 * Always instantiate seq_file even if read access doesn't use
724 	 * seq_file or is not requested.  This unifies private data access
725 	 * and readable regular files are the vast majority anyway.
726 	 */
727 	if (ops->seq_show)
728 		error = seq_open(file, &kernfs_seq_ops);
729 	else
730 		error = seq_open(file, NULL);
731 	if (error)
732 		goto err_free;
733 
734 	of->seq_file = file->private_data;
735 	of->seq_file->private = of;
736 
737 	/* seq_file clears PWRITE unconditionally, restore it if WRITE */
738 	if (file->f_mode & FMODE_WRITE)
739 		file->f_mode |= FMODE_PWRITE;
740 
741 	/* make sure we have open node struct */
742 	error = kernfs_get_open_node(kn, of);
743 	if (error)
744 		goto err_seq_release;
745 
746 	if (ops->open) {
747 		/* nobody has access to @of yet, skip @of->mutex */
748 		error = ops->open(of);
749 		if (error)
750 			goto err_put_node;
751 	}
752 
753 	/* open succeeded, put active references */
754 	kernfs_put_active(kn);
755 	return 0;
756 
757 err_put_node:
758 	kernfs_unlink_open_file(kn, of, true);
759 err_seq_release:
760 	seq_release(inode, file);
761 err_free:
762 	kfree(of->prealloc_buf);
763 	kfree(of);
764 err_out:
765 	kernfs_put_active(kn);
766 	return error;
767 }
768 
769 /* used from release/drain to ensure that ->release() is called exactly once */
kernfs_release_file(struct kernfs_node * kn,struct kernfs_open_file * of)770 static void kernfs_release_file(struct kernfs_node *kn,
771 				struct kernfs_open_file *of)
772 {
773 	/*
774 	 * @of is guaranteed to have no other file operations in flight and
775 	 * we just want to synchronize release and drain paths.
776 	 * @kernfs_open_file_mutex_ptr(kn) is enough. @of->mutex can't be used
777 	 * here because drain path may be called from places which can
778 	 * cause circular dependency.
779 	 */
780 	lockdep_assert_held(kernfs_open_file_mutex_ptr(kn));
781 
782 	if (!of->released) {
783 		/*
784 		 * A file is never detached without being released and we
785 		 * need to be able to release files which are deactivated
786 		 * and being drained.  Don't use kernfs_ops().
787 		 */
788 		kn->attr.ops->release(of);
789 		of->released = true;
790 		of_on(of)->nr_to_release--;
791 	}
792 }
793 
kernfs_fop_release(struct inode * inode,struct file * filp)794 static int kernfs_fop_release(struct inode *inode, struct file *filp)
795 {
796 	struct kernfs_node *kn = inode->i_private;
797 	struct kernfs_open_file *of = kernfs_of(filp);
798 
799 	if (kn->flags & KERNFS_HAS_RELEASE) {
800 		struct mutex *mutex;
801 
802 		mutex = kernfs_open_file_mutex_lock(kn);
803 		kernfs_release_file(kn, of);
804 		mutex_unlock(mutex);
805 	}
806 
807 	kernfs_unlink_open_file(kn, of, false);
808 	seq_release(inode, filp);
809 	kfree(of->prealloc_buf);
810 	kfree(of);
811 
812 	return 0;
813 }
814 
kernfs_should_drain_open_files(struct kernfs_node * kn)815 bool kernfs_should_drain_open_files(struct kernfs_node *kn)
816 {
817 	struct kernfs_open_node *on;
818 	bool ret;
819 
820 	/*
821 	 * @kn being deactivated guarantees that @kn->attr.open can't change
822 	 * beneath us making the lockless test below safe.
823 	 */
824 	WARN_ON_ONCE(atomic_read(&kn->active) != KN_DEACTIVATED_BIAS);
825 
826 	rcu_read_lock();
827 	on = rcu_dereference(kn->attr.open);
828 	ret = on && (on->nr_mmapped || on->nr_to_release);
829 	rcu_read_unlock();
830 
831 	return ret;
832 }
833 
kernfs_drain_open_files(struct kernfs_node * kn)834 void kernfs_drain_open_files(struct kernfs_node *kn)
835 {
836 	struct kernfs_open_node *on;
837 	struct kernfs_open_file *of;
838 	struct mutex *mutex;
839 
840 	mutex = kernfs_open_file_mutex_lock(kn);
841 	on = kernfs_deref_open_node_locked(kn);
842 	if (!on) {
843 		mutex_unlock(mutex);
844 		return;
845 	}
846 
847 	list_for_each_entry(of, &on->files, list) {
848 		struct inode *inode = file_inode(of->file);
849 
850 		if (of->mmapped) {
851 			unmap_mapping_range(inode->i_mapping, 0, 0, 1);
852 			of->mmapped = false;
853 			on->nr_mmapped--;
854 		}
855 
856 		if (kn->flags & KERNFS_HAS_RELEASE)
857 			kernfs_release_file(kn, of);
858 	}
859 
860 	WARN_ON_ONCE(on->nr_mmapped || on->nr_to_release);
861 	mutex_unlock(mutex);
862 }
863 
864 /*
865  * Kernfs attribute files are pollable.  The idea is that you read
866  * the content and then you use 'poll' or 'select' to wait for
867  * the content to change.  When the content changes (assuming the
868  * manager for the kobject supports notification), poll will
869  * return EPOLLERR|EPOLLPRI, and select will return the fd whether
870  * it is waiting for read, write, or exceptions.
871  * Once poll/select indicates that the value has changed, you
872  * need to close and re-open the file, or seek to 0 and read again.
873  * Reminder: this only works for attributes which actively support
874  * it, and it is not possible to test an attribute from userspace
875  * to see if it supports poll (Neither 'poll' nor 'select' return
876  * an appropriate error code).  When in doubt, set a suitable timeout value.
877  */
kernfs_generic_poll(struct kernfs_open_file * of,poll_table * wait)878 __poll_t kernfs_generic_poll(struct kernfs_open_file *of, poll_table *wait)
879 {
880 	struct kernfs_open_node *on = of_on(of);
881 
882 	poll_wait(of->file, &on->poll, wait);
883 
884 	if (of->event != atomic_read(&on->event))
885 		return DEFAULT_POLLMASK|EPOLLERR|EPOLLPRI;
886 
887 	return DEFAULT_POLLMASK;
888 }
889 
kernfs_fop_poll(struct file * filp,poll_table * wait)890 static __poll_t kernfs_fop_poll(struct file *filp, poll_table *wait)
891 {
892 	struct kernfs_open_file *of = kernfs_of(filp);
893 	struct kernfs_node *kn = kernfs_dentry_node(filp->f_path.dentry);
894 	__poll_t ret;
895 
896 	if (!kernfs_get_active(kn))
897 		return DEFAULT_POLLMASK|EPOLLERR|EPOLLPRI;
898 
899 	if (kn->attr.ops->poll)
900 		ret = kn->attr.ops->poll(of, wait);
901 	else
902 		ret = kernfs_generic_poll(of, wait);
903 
904 	kernfs_put_active(kn);
905 	return ret;
906 }
907 
kernfs_notify_workfn(struct work_struct * work)908 static void kernfs_notify_workfn(struct work_struct *work)
909 {
910 	struct kernfs_node *kn;
911 	struct kernfs_super_info *info;
912 	struct kernfs_root *root;
913 repeat:
914 	/* pop one off the notify_list */
915 	spin_lock_irq(&kernfs_notify_lock);
916 	kn = kernfs_notify_list;
917 	if (kn == KERNFS_NOTIFY_EOL) {
918 		spin_unlock_irq(&kernfs_notify_lock);
919 		return;
920 	}
921 	kernfs_notify_list = kn->attr.notify_next;
922 	kn->attr.notify_next = NULL;
923 	spin_unlock_irq(&kernfs_notify_lock);
924 
925 	root = kernfs_root(kn);
926 	/* kick fsnotify */
927 
928 	down_read(&root->kernfs_supers_rwsem);
929 	list_for_each_entry(info, &kernfs_root(kn)->supers, node) {
930 		struct kernfs_node *parent;
931 		struct inode *p_inode = NULL;
932 		struct inode *inode;
933 		struct qstr name;
934 
935 		/*
936 		 * We want fsnotify_modify() on @kn but as the
937 		 * modifications aren't originating from userland don't
938 		 * have the matching @file available.  Look up the inodes
939 		 * and generate the events manually.
940 		 */
941 		inode = ilookup(info->sb, kernfs_ino(kn));
942 		if (!inode)
943 			continue;
944 
945 		name = (struct qstr)QSTR_INIT(kn->name, strlen(kn->name));
946 		parent = kernfs_get_parent(kn);
947 		if (parent) {
948 			p_inode = ilookup(info->sb, kernfs_ino(parent));
949 			if (p_inode) {
950 				fsnotify(FS_MODIFY | FS_EVENT_ON_CHILD,
951 					 inode, FSNOTIFY_EVENT_INODE,
952 					 p_inode, &name, inode, 0);
953 				iput(p_inode);
954 			}
955 
956 			kernfs_put(parent);
957 		}
958 
959 		if (!p_inode)
960 			fsnotify_inode(inode, FS_MODIFY);
961 
962 		iput(inode);
963 	}
964 
965 	up_read(&root->kernfs_supers_rwsem);
966 	kernfs_put(kn);
967 	goto repeat;
968 }
969 
970 /**
971  * kernfs_notify - notify a kernfs file
972  * @kn: file to notify
973  *
974  * Notify @kn such that poll(2) on @kn wakes up.  Maybe be called from any
975  * context.
976  */
kernfs_notify(struct kernfs_node * kn)977 void kernfs_notify(struct kernfs_node *kn)
978 {
979 	static DECLARE_WORK(kernfs_notify_work, kernfs_notify_workfn);
980 	unsigned long flags;
981 	struct kernfs_open_node *on;
982 
983 	if (WARN_ON(kernfs_type(kn) != KERNFS_FILE))
984 		return;
985 
986 	/* kick poll immediately */
987 	rcu_read_lock();
988 	on = rcu_dereference(kn->attr.open);
989 	if (on) {
990 		atomic_inc(&on->event);
991 		wake_up_interruptible(&on->poll);
992 	}
993 	rcu_read_unlock();
994 
995 	/* schedule work to kick fsnotify */
996 	spin_lock_irqsave(&kernfs_notify_lock, flags);
997 	if (!kn->attr.notify_next) {
998 		kernfs_get(kn);
999 		kn->attr.notify_next = kernfs_notify_list;
1000 		kernfs_notify_list = kn;
1001 		schedule_work(&kernfs_notify_work);
1002 	}
1003 	spin_unlock_irqrestore(&kernfs_notify_lock, flags);
1004 }
1005 EXPORT_SYMBOL_GPL(kernfs_notify);
1006 
1007 const struct file_operations kernfs_file_fops = {
1008 	.read_iter	= kernfs_fop_read_iter,
1009 	.write_iter	= kernfs_fop_write_iter,
1010 	.llseek		= generic_file_llseek,
1011 	.mmap		= kernfs_fop_mmap,
1012 	.open		= kernfs_fop_open,
1013 	.release	= kernfs_fop_release,
1014 	.poll		= kernfs_fop_poll,
1015 	.fsync		= noop_fsync,
1016 	.splice_read	= copy_splice_read,
1017 	.splice_write	= iter_file_splice_write,
1018 };
1019 
1020 /**
1021  * __kernfs_create_file - kernfs internal function to create a file
1022  * @parent: directory to create the file in
1023  * @name: name of the file
1024  * @mode: mode of the file
1025  * @uid: uid of the file
1026  * @gid: gid of the file
1027  * @size: size of the file
1028  * @ops: kernfs operations for the file
1029  * @priv: private data for the file
1030  * @ns: optional namespace tag of the file
1031  * @key: lockdep key for the file's active_ref, %NULL to disable lockdep
1032  *
1033  * Return: the created node on success, ERR_PTR() value on error.
1034  */
__kernfs_create_file(struct kernfs_node * parent,const char * name,umode_t mode,kuid_t uid,kgid_t gid,loff_t size,const struct kernfs_ops * ops,void * priv,const void * ns,struct lock_class_key * key)1035 struct kernfs_node *__kernfs_create_file(struct kernfs_node *parent,
1036 					 const char *name,
1037 					 umode_t mode, kuid_t uid, kgid_t gid,
1038 					 loff_t size,
1039 					 const struct kernfs_ops *ops,
1040 					 void *priv, const void *ns,
1041 					 struct lock_class_key *key)
1042 {
1043 	struct kernfs_node *kn;
1044 	unsigned flags;
1045 	int rc;
1046 
1047 	flags = KERNFS_FILE;
1048 
1049 	kn = kernfs_new_node(parent, name, (mode & S_IALLUGO) | S_IFREG,
1050 			     uid, gid, flags);
1051 	if (!kn)
1052 		return ERR_PTR(-ENOMEM);
1053 
1054 	kn->attr.ops = ops;
1055 	kn->attr.size = size;
1056 	kn->ns = ns;
1057 	kn->priv = priv;
1058 
1059 #ifdef CONFIG_DEBUG_LOCK_ALLOC
1060 	if (key) {
1061 		lockdep_init_map(&kn->dep_map, "kn->active", key, 0);
1062 		kn->flags |= KERNFS_LOCKDEP;
1063 	}
1064 #endif
1065 
1066 	/*
1067 	 * kn->attr.ops is accessible only while holding active ref.  We
1068 	 * need to know whether some ops are implemented outside active
1069 	 * ref.  Cache their existence in flags.
1070 	 */
1071 	if (ops->seq_show)
1072 		kn->flags |= KERNFS_HAS_SEQ_SHOW;
1073 	if (ops->mmap)
1074 		kn->flags |= KERNFS_HAS_MMAP;
1075 	if (ops->release)
1076 		kn->flags |= KERNFS_HAS_RELEASE;
1077 
1078 	rc = kernfs_add_one(kn);
1079 	if (rc) {
1080 		kernfs_put(kn);
1081 		return ERR_PTR(rc);
1082 	}
1083 	return kn;
1084 }
1085