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 | |
43 | static inline struct mutex *kernfs_open_file_mutex_ptr(struct kernfs_node *kn) |
44 | { |
45 | int idx = hash_ptr(ptr: kn, NR_KERNFS_LOCK_BITS); |
46 | |
47 | return &kernfs_locks->open_file_mutex[idx]; |
48 | } |
49 | |
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 | */ |
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 * |
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 | |
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 | */ |
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 | */ |
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(kn: of->kn); |
139 | |
140 | if (ops->seq_stop) |
141 | ops->seq_stop(sf, v); |
142 | kernfs_put_active(kn: of->kn); |
143 | } |
144 | |
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(kn: of->kn)) |
156 | return ERR_PTR(error: -ENODEV); |
157 | |
158 | ops = kernfs_ops(kn: 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(error: -ENODEV)) |
163 | kernfs_seq_stop_active(sf, v: next); |
164 | return next; |
165 | } |
166 | return single_start(sf, ppos); |
167 | } |
168 | |
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(kn: 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(error: -ENODEV)) |
178 | kernfs_seq_stop_active(sf, v: 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 | |
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(error: -ENODEV)) |
195 | kernfs_seq_stop_active(sf, v); |
196 | mutex_unlock(lock: &of->mutex); |
197 | } |
198 | |
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(v: &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 | */ |
221 | static ssize_t kernfs_file_read_iter(struct kiocb *iocb, struct iov_iter *iter) |
222 | { |
223 | struct kernfs_open_file *of = kernfs_of(file: 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(size: 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(kn: of->kn)) { |
242 | len = -ENODEV; |
243 | mutex_unlock(lock: &of->mutex); |
244 | goto out_free; |
245 | } |
246 | |
247 | of->event = atomic_read(v: &of_on(of)->event); |
248 | |
249 | ops = kernfs_ops(kn: 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(kn: of->kn); |
256 | mutex_unlock(lock: &of->mutex); |
257 | |
258 | if (len < 0) |
259 | goto out_free; |
260 | |
261 | if (copy_to_iter(addr: buf, bytes: len, i: 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(lock: &of->prealloc_mutex); |
271 | else |
272 | kfree(objp: buf); |
273 | return len; |
274 | } |
275 | |
276 | static ssize_t kernfs_fop_read_iter(struct kiocb *iocb, struct iov_iter *iter) |
277 | { |
278 | if (kernfs_of(file: 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 | */ |
293 | static ssize_t kernfs_fop_write_iter(struct kiocb *iocb, struct iov_iter *iter) |
294 | { |
295 | struct kernfs_open_file *of = kernfs_of(file: iocb->ki_filp); |
296 | ssize_t len = iov_iter_count(i: 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(size: len + 1, GFP_KERNEL); |
312 | if (!buf) |
313 | return -ENOMEM; |
314 | |
315 | if (copy_from_iter(addr: buf, bytes: len, i: 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(kn: of->kn)) { |
327 | mutex_unlock(lock: &of->mutex); |
328 | len = -ENODEV; |
329 | goto out_free; |
330 | } |
331 | |
332 | ops = kernfs_ops(kn: 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(kn: of->kn); |
339 | mutex_unlock(lock: &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(lock: &of->prealloc_mutex); |
347 | else |
348 | kfree(objp: buf); |
349 | return len; |
350 | } |
351 | |
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(kn: of->kn)) |
361 | return; |
362 | |
363 | if (of->vm_ops->open) |
364 | of->vm_ops->open(vma); |
365 | |
366 | kernfs_put_active(kn: of->kn); |
367 | } |
368 | |
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(kn: 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(kn: of->kn); |
386 | return ret; |
387 | } |
388 | |
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(kn: 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(kn: of->kn); |
408 | return ret; |
409 | } |
410 | |
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(kn: 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(kn: of->kn); |
429 | return ret; |
430 | } |
431 | |
432 | static const struct vm_operations_struct kernfs_vm_ops = { |
433 | .open = kernfs_vma_open, |
434 | .fault = kernfs_vma_fault, |
435 | .page_mkwrite = kernfs_vma_page_mkwrite, |
436 | .access = kernfs_vma_access, |
437 | }; |
438 | |
439 | static int kernfs_fop_mmap(struct file *file, struct vm_area_struct *vma) |
440 | { |
441 | struct kernfs_open_file *of = kernfs_of(file); |
442 | const struct kernfs_ops *ops; |
443 | int rc; |
444 | |
445 | /* |
446 | * mmap path and of->mutex are prone to triggering spurious lockdep |
447 | * warnings and we don't want to add spurious locking dependency |
448 | * between the two. Check whether mmap is actually implemented |
449 | * without grabbing @of->mutex by testing HAS_MMAP flag. See the |
450 | * comment in kernfs_file_open() for more details. |
451 | */ |
452 | if (!(of->kn->flags & KERNFS_HAS_MMAP)) |
453 | return -ENODEV; |
454 | |
455 | mutex_lock(&of->mutex); |
456 | |
457 | rc = -ENODEV; |
458 | if (!kernfs_get_active(kn: of->kn)) |
459 | goto out_unlock; |
460 | |
461 | ops = kernfs_ops(kn: of->kn); |
462 | rc = ops->mmap(of, vma); |
463 | if (rc) |
464 | goto out_put; |
465 | |
466 | /* |
467 | * PowerPC's pci_mmap of legacy_mem uses shmem_zero_setup() |
468 | * to satisfy versions of X which crash if the mmap fails: that |
469 | * substitutes a new vm_file, and we don't then want bin_vm_ops. |
470 | */ |
471 | if (vma->vm_file != file) |
472 | goto out_put; |
473 | |
474 | rc = -EINVAL; |
475 | if (of->mmapped && of->vm_ops != vma->vm_ops) |
476 | goto out_put; |
477 | |
478 | /* |
479 | * It is not possible to successfully wrap close. |
480 | * So error if someone is trying to use close. |
481 | */ |
482 | if (vma->vm_ops && vma->vm_ops->close) |
483 | goto out_put; |
484 | |
485 | rc = 0; |
486 | of->mmapped = true; |
487 | of_on(of)->nr_mmapped++; |
488 | of->vm_ops = vma->vm_ops; |
489 | vma->vm_ops = &kernfs_vm_ops; |
490 | out_put: |
491 | kernfs_put_active(kn: of->kn); |
492 | out_unlock: |
493 | mutex_unlock(lock: &of->mutex); |
494 | |
495 | return rc; |
496 | } |
497 | |
498 | /** |
499 | * kernfs_get_open_node - get or create kernfs_open_node |
500 | * @kn: target kernfs_node |
501 | * @of: kernfs_open_file for this instance of open |
502 | * |
503 | * If @kn->attr.open exists, increment its reference count; otherwise, |
504 | * create one. @of is chained to the files list. |
505 | * |
506 | * Locking: |
507 | * Kernel thread context (may sleep). |
508 | * |
509 | * Return: |
510 | * %0 on success, -errno on failure. |
511 | */ |
512 | static int kernfs_get_open_node(struct kernfs_node *kn, |
513 | struct kernfs_open_file *of) |
514 | { |
515 | struct kernfs_open_node *on; |
516 | struct mutex *mutex; |
517 | |
518 | mutex = kernfs_open_file_mutex_lock(kn); |
519 | on = kernfs_deref_open_node_locked(kn); |
520 | |
521 | if (!on) { |
522 | /* not there, initialize a new one */ |
523 | on = kzalloc(size: sizeof(*on), GFP_KERNEL); |
524 | if (!on) { |
525 | mutex_unlock(lock: mutex); |
526 | return -ENOMEM; |
527 | } |
528 | atomic_set(v: &on->event, i: 1); |
529 | init_waitqueue_head(&on->poll); |
530 | INIT_LIST_HEAD(list: &on->files); |
531 | rcu_assign_pointer(kn->attr.open, on); |
532 | } |
533 | |
534 | list_add_tail(new: &of->list, head: &on->files); |
535 | if (kn->flags & KERNFS_HAS_RELEASE) |
536 | on->nr_to_release++; |
537 | |
538 | mutex_unlock(lock: mutex); |
539 | return 0; |
540 | } |
541 | |
542 | /** |
543 | * kernfs_unlink_open_file - Unlink @of from @kn. |
544 | * |
545 | * @kn: target kernfs_node |
546 | * @of: associated kernfs_open_file |
547 | * @open_failed: ->open() failed, cancel ->release() |
548 | * |
549 | * Unlink @of from list of @kn's associated open files. If list of |
550 | * associated open files becomes empty, disassociate and free |
551 | * kernfs_open_node. |
552 | * |
553 | * LOCKING: |
554 | * None. |
555 | */ |
556 | static void kernfs_unlink_open_file(struct kernfs_node *kn, |
557 | struct kernfs_open_file *of, |
558 | bool open_failed) |
559 | { |
560 | struct kernfs_open_node *on; |
561 | struct mutex *mutex; |
562 | |
563 | mutex = kernfs_open_file_mutex_lock(kn); |
564 | |
565 | on = kernfs_deref_open_node_locked(kn); |
566 | if (!on) { |
567 | mutex_unlock(lock: mutex); |
568 | return; |
569 | } |
570 | |
571 | if (of) { |
572 | if (kn->flags & KERNFS_HAS_RELEASE) { |
573 | WARN_ON_ONCE(of->released == open_failed); |
574 | if (open_failed) |
575 | on->nr_to_release--; |
576 | } |
577 | if (of->mmapped) |
578 | on->nr_mmapped--; |
579 | list_del(entry: &of->list); |
580 | } |
581 | |
582 | if (list_empty(head: &on->files)) { |
583 | rcu_assign_pointer(kn->attr.open, NULL); |
584 | kfree_rcu(on, rcu_head); |
585 | } |
586 | |
587 | mutex_unlock(lock: mutex); |
588 | } |
589 | |
590 | static int kernfs_fop_open(struct inode *inode, struct file *file) |
591 | { |
592 | struct kernfs_node *kn = inode->i_private; |
593 | struct kernfs_root *root = kernfs_root(kn); |
594 | const struct kernfs_ops *ops; |
595 | struct kernfs_open_file *of; |
596 | bool has_read, has_write, has_mmap; |
597 | int error = -EACCES; |
598 | |
599 | if (!kernfs_get_active(kn)) |
600 | return -ENODEV; |
601 | |
602 | ops = kernfs_ops(kn); |
603 | |
604 | has_read = ops->seq_show || ops->read || ops->mmap; |
605 | has_write = ops->write || ops->mmap; |
606 | has_mmap = ops->mmap; |
607 | |
608 | /* see the flag definition for details */ |
609 | if (root->flags & KERNFS_ROOT_EXTRA_OPEN_PERM_CHECK) { |
610 | if ((file->f_mode & FMODE_WRITE) && |
611 | (!(inode->i_mode & S_IWUGO) || !has_write)) |
612 | goto err_out; |
613 | |
614 | if ((file->f_mode & FMODE_READ) && |
615 | (!(inode->i_mode & S_IRUGO) || !has_read)) |
616 | goto err_out; |
617 | } |
618 | |
619 | /* allocate a kernfs_open_file for the file */ |
620 | error = -ENOMEM; |
621 | of = kzalloc(size: sizeof(struct kernfs_open_file), GFP_KERNEL); |
622 | if (!of) |
623 | goto err_out; |
624 | |
625 | /* |
626 | * The following is done to give a different lockdep key to |
627 | * @of->mutex for files which implement mmap. This is a rather |
628 | * crude way to avoid false positive lockdep warning around |
629 | * mm->mmap_lock - mmap nests @of->mutex under mm->mmap_lock and |
630 | * reading /sys/block/sda/trace/act_mask grabs sr_mutex, under |
631 | * which mm->mmap_lock nests, while holding @of->mutex. As each |
632 | * open file has a separate mutex, it's okay as long as those don't |
633 | * happen on the same file. At this point, we can't easily give |
634 | * each file a separate locking class. Let's differentiate on |
635 | * whether the file has mmap or not for now. |
636 | * |
637 | * Both paths of the branch look the same. They're supposed to |
638 | * look that way and give @of->mutex different static lockdep keys. |
639 | */ |
640 | if (has_mmap) |
641 | mutex_init(&of->mutex); |
642 | else |
643 | mutex_init(&of->mutex); |
644 | |
645 | of->kn = kn; |
646 | of->file = file; |
647 | |
648 | /* |
649 | * Write path needs to atomic_write_len outside active reference. |
650 | * Cache it in open_file. See kernfs_fop_write_iter() for details. |
651 | */ |
652 | of->atomic_write_len = ops->atomic_write_len; |
653 | |
654 | error = -EINVAL; |
655 | /* |
656 | * ->seq_show is incompatible with ->prealloc, |
657 | * as seq_read does its own allocation. |
658 | * ->read must be used instead. |
659 | */ |
660 | if (ops->prealloc && ops->seq_show) |
661 | goto err_free; |
662 | if (ops->prealloc) { |
663 | int len = of->atomic_write_len ?: PAGE_SIZE; |
664 | of->prealloc_buf = kmalloc(size: len + 1, GFP_KERNEL); |
665 | error = -ENOMEM; |
666 | if (!of->prealloc_buf) |
667 | goto err_free; |
668 | mutex_init(&of->prealloc_mutex); |
669 | } |
670 | |
671 | /* |
672 | * Always instantiate seq_file even if read access doesn't use |
673 | * seq_file or is not requested. This unifies private data access |
674 | * and readable regular files are the vast majority anyway. |
675 | */ |
676 | if (ops->seq_show) |
677 | error = seq_open(file, &kernfs_seq_ops); |
678 | else |
679 | error = seq_open(file, NULL); |
680 | if (error) |
681 | goto err_free; |
682 | |
683 | of->seq_file = file->private_data; |
684 | of->seq_file->private = of; |
685 | |
686 | /* seq_file clears PWRITE unconditionally, restore it if WRITE */ |
687 | if (file->f_mode & FMODE_WRITE) |
688 | file->f_mode |= FMODE_PWRITE; |
689 | |
690 | /* make sure we have open node struct */ |
691 | error = kernfs_get_open_node(kn, of); |
692 | if (error) |
693 | goto err_seq_release; |
694 | |
695 | if (ops->open) { |
696 | /* nobody has access to @of yet, skip @of->mutex */ |
697 | error = ops->open(of); |
698 | if (error) |
699 | goto err_put_node; |
700 | } |
701 | |
702 | /* open succeeded, put active references */ |
703 | kernfs_put_active(kn); |
704 | return 0; |
705 | |
706 | err_put_node: |
707 | kernfs_unlink_open_file(kn, of, open_failed: true); |
708 | err_seq_release: |
709 | seq_release(inode, file); |
710 | err_free: |
711 | kfree(objp: of->prealloc_buf); |
712 | kfree(objp: of); |
713 | err_out: |
714 | kernfs_put_active(kn); |
715 | return error; |
716 | } |
717 | |
718 | /* used from release/drain to ensure that ->release() is called exactly once */ |
719 | static void kernfs_release_file(struct kernfs_node *kn, |
720 | struct kernfs_open_file *of) |
721 | { |
722 | /* |
723 | * @of is guaranteed to have no other file operations in flight and |
724 | * we just want to synchronize release and drain paths. |
725 | * @kernfs_open_file_mutex_ptr(kn) is enough. @of->mutex can't be used |
726 | * here because drain path may be called from places which can |
727 | * cause circular dependency. |
728 | */ |
729 | lockdep_assert_held(kernfs_open_file_mutex_ptr(kn)); |
730 | |
731 | if (!of->released) { |
732 | /* |
733 | * A file is never detached without being released and we |
734 | * need to be able to release files which are deactivated |
735 | * and being drained. Don't use kernfs_ops(). |
736 | */ |
737 | kn->attr.ops->release(of); |
738 | of->released = true; |
739 | of_on(of)->nr_to_release--; |
740 | } |
741 | } |
742 | |
743 | static int kernfs_fop_release(struct inode *inode, struct file *filp) |
744 | { |
745 | struct kernfs_node *kn = inode->i_private; |
746 | struct kernfs_open_file *of = kernfs_of(file: filp); |
747 | |
748 | if (kn->flags & KERNFS_HAS_RELEASE) { |
749 | struct mutex *mutex; |
750 | |
751 | mutex = kernfs_open_file_mutex_lock(kn); |
752 | kernfs_release_file(kn, of); |
753 | mutex_unlock(lock: mutex); |
754 | } |
755 | |
756 | kernfs_unlink_open_file(kn, of, open_failed: false); |
757 | seq_release(inode, filp); |
758 | kfree(objp: of->prealloc_buf); |
759 | kfree(objp: of); |
760 | |
761 | return 0; |
762 | } |
763 | |
764 | bool kernfs_should_drain_open_files(struct kernfs_node *kn) |
765 | { |
766 | struct kernfs_open_node *on; |
767 | bool ret; |
768 | |
769 | /* |
770 | * @kn being deactivated guarantees that @kn->attr.open can't change |
771 | * beneath us making the lockless test below safe. |
772 | */ |
773 | WARN_ON_ONCE(atomic_read(&kn->active) != KN_DEACTIVATED_BIAS); |
774 | |
775 | rcu_read_lock(); |
776 | on = rcu_dereference(kn->attr.open); |
777 | ret = on && (on->nr_mmapped || on->nr_to_release); |
778 | rcu_read_unlock(); |
779 | |
780 | return ret; |
781 | } |
782 | |
783 | void kernfs_drain_open_files(struct kernfs_node *kn) |
784 | { |
785 | struct kernfs_open_node *on; |
786 | struct kernfs_open_file *of; |
787 | struct mutex *mutex; |
788 | |
789 | mutex = kernfs_open_file_mutex_lock(kn); |
790 | on = kernfs_deref_open_node_locked(kn); |
791 | if (!on) { |
792 | mutex_unlock(lock: mutex); |
793 | return; |
794 | } |
795 | |
796 | list_for_each_entry(of, &on->files, list) { |
797 | struct inode *inode = file_inode(f: of->file); |
798 | |
799 | if (of->mmapped) { |
800 | unmap_mapping_range(mapping: inode->i_mapping, holebegin: 0, holelen: 0, even_cows: 1); |
801 | of->mmapped = false; |
802 | on->nr_mmapped--; |
803 | } |
804 | |
805 | if (kn->flags & KERNFS_HAS_RELEASE) |
806 | kernfs_release_file(kn, of); |
807 | } |
808 | |
809 | WARN_ON_ONCE(on->nr_mmapped || on->nr_to_release); |
810 | mutex_unlock(lock: mutex); |
811 | } |
812 | |
813 | /* |
814 | * Kernfs attribute files are pollable. The idea is that you read |
815 | * the content and then you use 'poll' or 'select' to wait for |
816 | * the content to change. When the content changes (assuming the |
817 | * manager for the kobject supports notification), poll will |
818 | * return EPOLLERR|EPOLLPRI, and select will return the fd whether |
819 | * it is waiting for read, write, or exceptions. |
820 | * Once poll/select indicates that the value has changed, you |
821 | * need to close and re-open the file, or seek to 0 and read again. |
822 | * Reminder: this only works for attributes which actively support |
823 | * it, and it is not possible to test an attribute from userspace |
824 | * to see if it supports poll (Neither 'poll' nor 'select' return |
825 | * an appropriate error code). When in doubt, set a suitable timeout value. |
826 | */ |
827 | __poll_t kernfs_generic_poll(struct kernfs_open_file *of, poll_table *wait) |
828 | { |
829 | struct kernfs_open_node *on = of_on(of); |
830 | |
831 | poll_wait(filp: of->file, wait_address: &on->poll, p: wait); |
832 | |
833 | if (of->event != atomic_read(v: &on->event)) |
834 | return DEFAULT_POLLMASK|EPOLLERR|EPOLLPRI; |
835 | |
836 | return DEFAULT_POLLMASK; |
837 | } |
838 | |
839 | static __poll_t kernfs_fop_poll(struct file *filp, poll_table *wait) |
840 | { |
841 | struct kernfs_open_file *of = kernfs_of(file: filp); |
842 | struct kernfs_node *kn = kernfs_dentry_node(dentry: filp->f_path.dentry); |
843 | __poll_t ret; |
844 | |
845 | if (!kernfs_get_active(kn)) |
846 | return DEFAULT_POLLMASK|EPOLLERR|EPOLLPRI; |
847 | |
848 | if (kn->attr.ops->poll) |
849 | ret = kn->attr.ops->poll(of, wait); |
850 | else |
851 | ret = kernfs_generic_poll(of, wait); |
852 | |
853 | kernfs_put_active(kn); |
854 | return ret; |
855 | } |
856 | |
857 | static loff_t kernfs_fop_llseek(struct file *file, loff_t offset, int whence) |
858 | { |
859 | struct kernfs_open_file *of = kernfs_of(file); |
860 | const struct kernfs_ops *ops; |
861 | loff_t ret; |
862 | |
863 | /* |
864 | * @of->mutex nests outside active ref and is primarily to ensure that |
865 | * the ops aren't called concurrently for the same open file. |
866 | */ |
867 | mutex_lock(&of->mutex); |
868 | if (!kernfs_get_active(kn: of->kn)) { |
869 | mutex_unlock(lock: &of->mutex); |
870 | return -ENODEV; |
871 | } |
872 | |
873 | ops = kernfs_ops(kn: of->kn); |
874 | if (ops->llseek) |
875 | ret = ops->llseek(of, offset, whence); |
876 | else |
877 | ret = generic_file_llseek(file, offset, whence); |
878 | |
879 | kernfs_put_active(kn: of->kn); |
880 | mutex_unlock(lock: &of->mutex); |
881 | return ret; |
882 | } |
883 | |
884 | static void kernfs_notify_workfn(struct work_struct *work) |
885 | { |
886 | struct kernfs_node *kn; |
887 | struct kernfs_super_info *info; |
888 | struct kernfs_root *root; |
889 | repeat: |
890 | /* pop one off the notify_list */ |
891 | spin_lock_irq(lock: &kernfs_notify_lock); |
892 | kn = kernfs_notify_list; |
893 | if (kn == KERNFS_NOTIFY_EOL) { |
894 | spin_unlock_irq(lock: &kernfs_notify_lock); |
895 | return; |
896 | } |
897 | kernfs_notify_list = kn->attr.notify_next; |
898 | kn->attr.notify_next = NULL; |
899 | spin_unlock_irq(lock: &kernfs_notify_lock); |
900 | |
901 | root = kernfs_root(kn); |
902 | /* kick fsnotify */ |
903 | |
904 | down_read(sem: &root->kernfs_supers_rwsem); |
905 | list_for_each_entry(info, &kernfs_root(kn)->supers, node) { |
906 | struct kernfs_node *parent; |
907 | struct inode *p_inode = NULL; |
908 | struct inode *inode; |
909 | struct qstr name; |
910 | |
911 | /* |
912 | * We want fsnotify_modify() on @kn but as the |
913 | * modifications aren't originating from userland don't |
914 | * have the matching @file available. Look up the inodes |
915 | * and generate the events manually. |
916 | */ |
917 | inode = ilookup(sb: info->sb, ino: kernfs_ino(kn)); |
918 | if (!inode) |
919 | continue; |
920 | |
921 | name = (struct qstr)QSTR_INIT(kn->name, strlen(kn->name)); |
922 | parent = kernfs_get_parent(kn); |
923 | if (parent) { |
924 | p_inode = ilookup(sb: info->sb, ino: kernfs_ino(kn: parent)); |
925 | if (p_inode) { |
926 | fsnotify(FS_MODIFY | FS_EVENT_ON_CHILD, |
927 | data: inode, data_type: FSNOTIFY_EVENT_INODE, |
928 | dir: p_inode, name: &name, inode, cookie: 0); |
929 | iput(p_inode); |
930 | } |
931 | |
932 | kernfs_put(kn: parent); |
933 | } |
934 | |
935 | if (!p_inode) |
936 | fsnotify_inode(inode, FS_MODIFY); |
937 | |
938 | iput(inode); |
939 | } |
940 | |
941 | up_read(sem: &root->kernfs_supers_rwsem); |
942 | kernfs_put(kn); |
943 | goto repeat; |
944 | } |
945 | |
946 | /** |
947 | * kernfs_notify - notify a kernfs file |
948 | * @kn: file to notify |
949 | * |
950 | * Notify @kn such that poll(2) on @kn wakes up. Maybe be called from any |
951 | * context. |
952 | */ |
953 | void kernfs_notify(struct kernfs_node *kn) |
954 | { |
955 | static DECLARE_WORK(kernfs_notify_work, kernfs_notify_workfn); |
956 | unsigned long flags; |
957 | struct kernfs_open_node *on; |
958 | |
959 | if (WARN_ON(kernfs_type(kn) != KERNFS_FILE)) |
960 | return; |
961 | |
962 | /* kick poll immediately */ |
963 | rcu_read_lock(); |
964 | on = rcu_dereference(kn->attr.open); |
965 | if (on) { |
966 | atomic_inc(v: &on->event); |
967 | wake_up_interruptible(&on->poll); |
968 | } |
969 | rcu_read_unlock(); |
970 | |
971 | /* schedule work to kick fsnotify */ |
972 | spin_lock_irqsave(&kernfs_notify_lock, flags); |
973 | if (!kn->attr.notify_next) { |
974 | kernfs_get(kn); |
975 | kn->attr.notify_next = kernfs_notify_list; |
976 | kernfs_notify_list = kn; |
977 | schedule_work(work: &kernfs_notify_work); |
978 | } |
979 | spin_unlock_irqrestore(lock: &kernfs_notify_lock, flags); |
980 | } |
981 | EXPORT_SYMBOL_GPL(kernfs_notify); |
982 | |
983 | const struct file_operations kernfs_file_fops = { |
984 | .read_iter = kernfs_fop_read_iter, |
985 | .write_iter = kernfs_fop_write_iter, |
986 | .llseek = kernfs_fop_llseek, |
987 | .mmap = kernfs_fop_mmap, |
988 | .open = kernfs_fop_open, |
989 | .release = kernfs_fop_release, |
990 | .poll = kernfs_fop_poll, |
991 | .fsync = noop_fsync, |
992 | .splice_read = copy_splice_read, |
993 | .splice_write = iter_file_splice_write, |
994 | }; |
995 | |
996 | /** |
997 | * __kernfs_create_file - kernfs internal function to create a file |
998 | * @parent: directory to create the file in |
999 | * @name: name of the file |
1000 | * @mode: mode of the file |
1001 | * @uid: uid of the file |
1002 | * @gid: gid of the file |
1003 | * @size: size of the file |
1004 | * @ops: kernfs operations for the file |
1005 | * @priv: private data for the file |
1006 | * @ns: optional namespace tag of the file |
1007 | * @key: lockdep key for the file's active_ref, %NULL to disable lockdep |
1008 | * |
1009 | * Return: the created node on success, ERR_PTR() value on error. |
1010 | */ |
1011 | struct kernfs_node *__kernfs_create_file(struct kernfs_node *parent, |
1012 | const char *name, |
1013 | umode_t mode, kuid_t uid, kgid_t gid, |
1014 | loff_t size, |
1015 | const struct kernfs_ops *ops, |
1016 | void *priv, const void *ns, |
1017 | struct lock_class_key *key) |
1018 | { |
1019 | struct kernfs_node *kn; |
1020 | unsigned flags; |
1021 | int rc; |
1022 | |
1023 | flags = KERNFS_FILE; |
1024 | |
1025 | kn = kernfs_new_node(parent, name, mode: (mode & S_IALLUGO) | S_IFREG, |
1026 | uid, gid, flags); |
1027 | if (!kn) |
1028 | return ERR_PTR(error: -ENOMEM); |
1029 | |
1030 | kn->attr.ops = ops; |
1031 | kn->attr.size = size; |
1032 | kn->ns = ns; |
1033 | kn->priv = priv; |
1034 | |
1035 | #ifdef CONFIG_DEBUG_LOCK_ALLOC |
1036 | if (key) { |
1037 | lockdep_init_map(lock: &kn->dep_map, name: "kn->active" , key, subclass: 0); |
1038 | kn->flags |= KERNFS_LOCKDEP; |
1039 | } |
1040 | #endif |
1041 | |
1042 | /* |
1043 | * kn->attr.ops is accessible only while holding active ref. We |
1044 | * need to know whether some ops are implemented outside active |
1045 | * ref. Cache their existence in flags. |
1046 | */ |
1047 | if (ops->seq_show) |
1048 | kn->flags |= KERNFS_HAS_SEQ_SHOW; |
1049 | if (ops->mmap) |
1050 | kn->flags |= KERNFS_HAS_MMAP; |
1051 | if (ops->release) |
1052 | kn->flags |= KERNFS_HAS_RELEASE; |
1053 | |
1054 | rc = kernfs_add_one(kn); |
1055 | if (rc) { |
1056 | kernfs_put(kn); |
1057 | return ERR_PTR(error: rc); |
1058 | } |
1059 | return kn; |
1060 | } |
1061 | |