xref: /openbmc/linux/fs/overlayfs/file.c (revision 519b58bb)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2017 Red Hat, Inc.
4  */
5 
6 #include <linux/cred.h>
7 #include <linux/file.h>
8 #include <linux/mount.h>
9 #include <linux/xattr.h>
10 #include <linux/uio.h>
11 #include <linux/uaccess.h>
12 #include <linux/splice.h>
13 #include <linux/security.h>
14 #include <linux/mm.h>
15 #include <linux/fs.h>
16 #include "overlayfs.h"
17 
18 struct ovl_aio_req {
19 	struct kiocb iocb;
20 	refcount_t ref;
21 	struct kiocb *orig_iocb;
22 	struct fd fd;
23 };
24 
25 static struct kmem_cache *ovl_aio_request_cachep;
26 
27 static char ovl_whatisit(struct inode *inode, struct inode *realinode)
28 {
29 	if (realinode != ovl_inode_upper(inode))
30 		return 'l';
31 	if (ovl_has_upperdata(inode))
32 		return 'u';
33 	else
34 		return 'm';
35 }
36 
37 /* No atime modificaton nor notify on underlying */
38 #define OVL_OPEN_FLAGS (O_NOATIME | FMODE_NONOTIFY)
39 
40 static struct file *ovl_open_realfile(const struct file *file,
41 				      const struct path *realpath)
42 {
43 	struct inode *realinode = d_inode(realpath->dentry);
44 	struct inode *inode = file_inode(file);
45 	struct user_namespace *real_mnt_userns;
46 	struct file *realfile;
47 	const struct cred *old_cred;
48 	int flags = file->f_flags | OVL_OPEN_FLAGS;
49 	int acc_mode = ACC_MODE(flags);
50 	int err;
51 
52 	if (flags & O_APPEND)
53 		acc_mode |= MAY_APPEND;
54 
55 	old_cred = ovl_override_creds(inode->i_sb);
56 	real_mnt_userns = mnt_user_ns(realpath->mnt);
57 	err = inode_permission(real_mnt_userns, realinode, MAY_OPEN | acc_mode);
58 	if (err) {
59 		realfile = ERR_PTR(err);
60 	} else {
61 		if (!inode_owner_or_capable(real_mnt_userns, realinode))
62 			flags &= ~O_NOATIME;
63 
64 		realfile = open_with_fake_path(&file->f_path, flags, realinode,
65 					       current_cred());
66 	}
67 	revert_creds(old_cred);
68 
69 	pr_debug("open(%p[%pD2/%c], 0%o) -> (%p, 0%o)\n",
70 		 file, file, ovl_whatisit(inode, realinode), file->f_flags,
71 		 realfile, IS_ERR(realfile) ? 0 : realfile->f_flags);
72 
73 	return realfile;
74 }
75 
76 #define OVL_SETFL_MASK (O_APPEND | O_NONBLOCK | O_NDELAY | O_DIRECT)
77 
78 static int ovl_change_flags(struct file *file, unsigned int flags)
79 {
80 	struct inode *inode = file_inode(file);
81 	int err;
82 
83 	flags &= OVL_SETFL_MASK;
84 
85 	if (((flags ^ file->f_flags) & O_APPEND) && IS_APPEND(inode))
86 		return -EPERM;
87 
88 	if ((flags & O_DIRECT) && !(file->f_mode & FMODE_CAN_ODIRECT))
89 		return -EINVAL;
90 
91 	if (file->f_op->check_flags) {
92 		err = file->f_op->check_flags(flags);
93 		if (err)
94 			return err;
95 	}
96 
97 	spin_lock(&file->f_lock);
98 	file->f_flags = (file->f_flags & ~OVL_SETFL_MASK) | flags;
99 	spin_unlock(&file->f_lock);
100 
101 	return 0;
102 }
103 
104 static int ovl_real_fdget_meta(const struct file *file, struct fd *real,
105 			       bool allow_meta)
106 {
107 	struct dentry *dentry = file_dentry(file);
108 	struct path realpath;
109 
110 	real->flags = 0;
111 	real->file = file->private_data;
112 
113 	if (allow_meta)
114 		ovl_path_real(dentry, &realpath);
115 	else
116 		ovl_path_realdata(dentry, &realpath);
117 
118 	/* Has it been copied up since we'd opened it? */
119 	if (unlikely(file_inode(real->file) != d_inode(realpath.dentry))) {
120 		real->flags = FDPUT_FPUT;
121 		real->file = ovl_open_realfile(file, &realpath);
122 
123 		return PTR_ERR_OR_ZERO(real->file);
124 	}
125 
126 	/* Did the flags change since open? */
127 	if (unlikely((file->f_flags ^ real->file->f_flags) & ~OVL_OPEN_FLAGS))
128 		return ovl_change_flags(real->file, file->f_flags);
129 
130 	return 0;
131 }
132 
133 static int ovl_real_fdget(const struct file *file, struct fd *real)
134 {
135 	if (d_is_dir(file_dentry(file))) {
136 		real->flags = 0;
137 		real->file = ovl_dir_real_file(file, false);
138 
139 		return PTR_ERR_OR_ZERO(real->file);
140 	}
141 
142 	return ovl_real_fdget_meta(file, real, false);
143 }
144 
145 static int ovl_open(struct inode *inode, struct file *file)
146 {
147 	struct dentry *dentry = file_dentry(file);
148 	struct file *realfile;
149 	struct path realpath;
150 	int err;
151 
152 	err = ovl_maybe_copy_up(dentry, file->f_flags);
153 	if (err)
154 		return err;
155 
156 	/* No longer need these flags, so don't pass them on to underlying fs */
157 	file->f_flags &= ~(O_CREAT | O_EXCL | O_NOCTTY | O_TRUNC);
158 
159 	ovl_path_realdata(dentry, &realpath);
160 	realfile = ovl_open_realfile(file, &realpath);
161 	if (IS_ERR(realfile))
162 		return PTR_ERR(realfile);
163 
164 	file->private_data = realfile;
165 
166 	return 0;
167 }
168 
169 static int ovl_release(struct inode *inode, struct file *file)
170 {
171 	fput(file->private_data);
172 
173 	return 0;
174 }
175 
176 static loff_t ovl_llseek(struct file *file, loff_t offset, int whence)
177 {
178 	struct inode *inode = file_inode(file);
179 	struct fd real;
180 	const struct cred *old_cred;
181 	loff_t ret;
182 
183 	/*
184 	 * The two special cases below do not need to involve real fs,
185 	 * so we can optimizing concurrent callers.
186 	 */
187 	if (offset == 0) {
188 		if (whence == SEEK_CUR)
189 			return file->f_pos;
190 
191 		if (whence == SEEK_SET)
192 			return vfs_setpos(file, 0, 0);
193 	}
194 
195 	ret = ovl_real_fdget(file, &real);
196 	if (ret)
197 		return ret;
198 
199 	/*
200 	 * Overlay file f_pos is the master copy that is preserved
201 	 * through copy up and modified on read/write, but only real
202 	 * fs knows how to SEEK_HOLE/SEEK_DATA and real fs may impose
203 	 * limitations that are more strict than ->s_maxbytes for specific
204 	 * files, so we use the real file to perform seeks.
205 	 */
206 	ovl_inode_lock(inode);
207 	real.file->f_pos = file->f_pos;
208 
209 	old_cred = ovl_override_creds(inode->i_sb);
210 	ret = vfs_llseek(real.file, offset, whence);
211 	revert_creds(old_cred);
212 
213 	file->f_pos = real.file->f_pos;
214 	ovl_inode_unlock(inode);
215 
216 	fdput(real);
217 
218 	return ret;
219 }
220 
221 static void ovl_file_accessed(struct file *file)
222 {
223 	struct inode *inode, *upperinode;
224 
225 	if (file->f_flags & O_NOATIME)
226 		return;
227 
228 	inode = file_inode(file);
229 	upperinode = ovl_inode_upper(inode);
230 
231 	if (!upperinode)
232 		return;
233 
234 	if ((!timespec64_equal(&inode->i_mtime, &upperinode->i_mtime) ||
235 	     !timespec64_equal(&inode->i_ctime, &upperinode->i_ctime))) {
236 		inode->i_mtime = upperinode->i_mtime;
237 		inode->i_ctime = upperinode->i_ctime;
238 	}
239 
240 	touch_atime(&file->f_path);
241 }
242 
243 static rwf_t ovl_iocb_to_rwf(int ifl)
244 {
245 	rwf_t flags = 0;
246 
247 	if (ifl & IOCB_NOWAIT)
248 		flags |= RWF_NOWAIT;
249 	if (ifl & IOCB_HIPRI)
250 		flags |= RWF_HIPRI;
251 	if (ifl & IOCB_DSYNC)
252 		flags |= RWF_DSYNC;
253 	if (ifl & IOCB_SYNC)
254 		flags |= RWF_SYNC;
255 
256 	return flags;
257 }
258 
259 static inline void ovl_aio_put(struct ovl_aio_req *aio_req)
260 {
261 	if (refcount_dec_and_test(&aio_req->ref)) {
262 		fdput(aio_req->fd);
263 		kmem_cache_free(ovl_aio_request_cachep, aio_req);
264 	}
265 }
266 
267 static void ovl_aio_cleanup_handler(struct ovl_aio_req *aio_req)
268 {
269 	struct kiocb *iocb = &aio_req->iocb;
270 	struct kiocb *orig_iocb = aio_req->orig_iocb;
271 
272 	if (iocb->ki_flags & IOCB_WRITE) {
273 		struct inode *inode = file_inode(orig_iocb->ki_filp);
274 
275 		/* Actually acquired in ovl_write_iter() */
276 		__sb_writers_acquired(file_inode(iocb->ki_filp)->i_sb,
277 				      SB_FREEZE_WRITE);
278 		file_end_write(iocb->ki_filp);
279 		ovl_copyattr(inode);
280 	}
281 
282 	orig_iocb->ki_pos = iocb->ki_pos;
283 	ovl_aio_put(aio_req);
284 }
285 
286 static void ovl_aio_rw_complete(struct kiocb *iocb, long res)
287 {
288 	struct ovl_aio_req *aio_req = container_of(iocb,
289 						   struct ovl_aio_req, iocb);
290 	struct kiocb *orig_iocb = aio_req->orig_iocb;
291 
292 	ovl_aio_cleanup_handler(aio_req);
293 	orig_iocb->ki_complete(orig_iocb, res);
294 }
295 
296 static ssize_t ovl_read_iter(struct kiocb *iocb, struct iov_iter *iter)
297 {
298 	struct file *file = iocb->ki_filp;
299 	struct fd real;
300 	const struct cred *old_cred;
301 	ssize_t ret;
302 
303 	if (!iov_iter_count(iter))
304 		return 0;
305 
306 	ret = ovl_real_fdget(file, &real);
307 	if (ret)
308 		return ret;
309 
310 	ret = -EINVAL;
311 	if (iocb->ki_flags & IOCB_DIRECT &&
312 	    !(real.file->f_mode & FMODE_CAN_ODIRECT))
313 		goto out_fdput;
314 
315 	old_cred = ovl_override_creds(file_inode(file)->i_sb);
316 	if (is_sync_kiocb(iocb)) {
317 		ret = vfs_iter_read(real.file, iter, &iocb->ki_pos,
318 				    ovl_iocb_to_rwf(iocb->ki_flags));
319 	} else {
320 		struct ovl_aio_req *aio_req;
321 
322 		ret = -ENOMEM;
323 		aio_req = kmem_cache_zalloc(ovl_aio_request_cachep, GFP_KERNEL);
324 		if (!aio_req)
325 			goto out;
326 
327 		aio_req->fd = real;
328 		real.flags = 0;
329 		aio_req->orig_iocb = iocb;
330 		kiocb_clone(&aio_req->iocb, iocb, real.file);
331 		aio_req->iocb.ki_complete = ovl_aio_rw_complete;
332 		refcount_set(&aio_req->ref, 2);
333 		ret = vfs_iocb_iter_read(real.file, &aio_req->iocb, iter);
334 		ovl_aio_put(aio_req);
335 		if (ret != -EIOCBQUEUED)
336 			ovl_aio_cleanup_handler(aio_req);
337 	}
338 out:
339 	revert_creds(old_cred);
340 	ovl_file_accessed(file);
341 out_fdput:
342 	fdput(real);
343 
344 	return ret;
345 }
346 
347 static ssize_t ovl_write_iter(struct kiocb *iocb, struct iov_iter *iter)
348 {
349 	struct file *file = iocb->ki_filp;
350 	struct inode *inode = file_inode(file);
351 	struct fd real;
352 	const struct cred *old_cred;
353 	ssize_t ret;
354 	int ifl = iocb->ki_flags;
355 
356 	if (!iov_iter_count(iter))
357 		return 0;
358 
359 	inode_lock(inode);
360 	/* Update mode */
361 	ovl_copyattr(inode);
362 	ret = file_remove_privs(file);
363 	if (ret)
364 		goto out_unlock;
365 
366 	ret = ovl_real_fdget(file, &real);
367 	if (ret)
368 		goto out_unlock;
369 
370 	ret = -EINVAL;
371 	if (iocb->ki_flags & IOCB_DIRECT &&
372 	    !(real.file->f_mode & FMODE_CAN_ODIRECT))
373 		goto out_fdput;
374 
375 	if (!ovl_should_sync(OVL_FS(inode->i_sb)))
376 		ifl &= ~(IOCB_DSYNC | IOCB_SYNC);
377 
378 	old_cred = ovl_override_creds(file_inode(file)->i_sb);
379 	if (is_sync_kiocb(iocb)) {
380 		file_start_write(real.file);
381 		ret = vfs_iter_write(real.file, iter, &iocb->ki_pos,
382 				     ovl_iocb_to_rwf(ifl));
383 		file_end_write(real.file);
384 		/* Update size */
385 		ovl_copyattr(inode);
386 	} else {
387 		struct ovl_aio_req *aio_req;
388 
389 		ret = -ENOMEM;
390 		aio_req = kmem_cache_zalloc(ovl_aio_request_cachep, GFP_KERNEL);
391 		if (!aio_req)
392 			goto out;
393 
394 		file_start_write(real.file);
395 		/* Pacify lockdep, same trick as done in aio_write() */
396 		__sb_writers_release(file_inode(real.file)->i_sb,
397 				     SB_FREEZE_WRITE);
398 		aio_req->fd = real;
399 		real.flags = 0;
400 		aio_req->orig_iocb = iocb;
401 		kiocb_clone(&aio_req->iocb, iocb, real.file);
402 		aio_req->iocb.ki_flags = ifl;
403 		aio_req->iocb.ki_complete = ovl_aio_rw_complete;
404 		refcount_set(&aio_req->ref, 2);
405 		ret = vfs_iocb_iter_write(real.file, &aio_req->iocb, iter);
406 		ovl_aio_put(aio_req);
407 		if (ret != -EIOCBQUEUED)
408 			ovl_aio_cleanup_handler(aio_req);
409 	}
410 out:
411 	revert_creds(old_cred);
412 out_fdput:
413 	fdput(real);
414 
415 out_unlock:
416 	inode_unlock(inode);
417 
418 	return ret;
419 }
420 
421 /*
422  * Calling iter_file_splice_write() directly from overlay's f_op may deadlock
423  * due to lock order inversion between pipe->mutex in iter_file_splice_write()
424  * and file_start_write(real.file) in ovl_write_iter().
425  *
426  * So do everything ovl_write_iter() does and call iter_file_splice_write() on
427  * the real file.
428  */
429 static ssize_t ovl_splice_write(struct pipe_inode_info *pipe, struct file *out,
430 				loff_t *ppos, size_t len, unsigned int flags)
431 {
432 	struct fd real;
433 	const struct cred *old_cred;
434 	struct inode *inode = file_inode(out);
435 	ssize_t ret;
436 
437 	inode_lock(inode);
438 	/* Update mode */
439 	ovl_copyattr(inode);
440 	ret = file_remove_privs(out);
441 	if (ret)
442 		goto out_unlock;
443 
444 	ret = ovl_real_fdget(out, &real);
445 	if (ret)
446 		goto out_unlock;
447 
448 	old_cred = ovl_override_creds(inode->i_sb);
449 	file_start_write(real.file);
450 
451 	ret = iter_file_splice_write(pipe, real.file, ppos, len, flags);
452 
453 	file_end_write(real.file);
454 	/* Update size */
455 	ovl_copyattr(inode);
456 	revert_creds(old_cred);
457 	fdput(real);
458 
459 out_unlock:
460 	inode_unlock(inode);
461 
462 	return ret;
463 }
464 
465 static int ovl_fsync(struct file *file, loff_t start, loff_t end, int datasync)
466 {
467 	struct fd real;
468 	const struct cred *old_cred;
469 	int ret;
470 
471 	ret = ovl_sync_status(OVL_FS(file_inode(file)->i_sb));
472 	if (ret <= 0)
473 		return ret;
474 
475 	ret = ovl_real_fdget_meta(file, &real, !datasync);
476 	if (ret)
477 		return ret;
478 
479 	/* Don't sync lower file for fear of receiving EROFS error */
480 	if (file_inode(real.file) == ovl_inode_upper(file_inode(file))) {
481 		old_cred = ovl_override_creds(file_inode(file)->i_sb);
482 		ret = vfs_fsync_range(real.file, start, end, datasync);
483 		revert_creds(old_cred);
484 	}
485 
486 	fdput(real);
487 
488 	return ret;
489 }
490 
491 static int ovl_mmap(struct file *file, struct vm_area_struct *vma)
492 {
493 	struct file *realfile = file->private_data;
494 	const struct cred *old_cred;
495 	int ret;
496 
497 	if (!realfile->f_op->mmap)
498 		return -ENODEV;
499 
500 	if (WARN_ON(file != vma->vm_file))
501 		return -EIO;
502 
503 	vma_set_file(vma, realfile);
504 
505 	old_cred = ovl_override_creds(file_inode(file)->i_sb);
506 	ret = call_mmap(vma->vm_file, vma);
507 	revert_creds(old_cred);
508 	ovl_file_accessed(file);
509 
510 	return ret;
511 }
512 
513 static long ovl_fallocate(struct file *file, int mode, loff_t offset, loff_t len)
514 {
515 	struct inode *inode = file_inode(file);
516 	struct fd real;
517 	const struct cred *old_cred;
518 	int ret;
519 
520 	ret = ovl_real_fdget(file, &real);
521 	if (ret)
522 		return ret;
523 
524 	old_cred = ovl_override_creds(file_inode(file)->i_sb);
525 	ret = vfs_fallocate(real.file, mode, offset, len);
526 	revert_creds(old_cred);
527 
528 	/* Update size */
529 	ovl_copyattr(inode);
530 
531 	fdput(real);
532 
533 	return ret;
534 }
535 
536 static int ovl_fadvise(struct file *file, loff_t offset, loff_t len, int advice)
537 {
538 	struct fd real;
539 	const struct cred *old_cred;
540 	int ret;
541 
542 	ret = ovl_real_fdget(file, &real);
543 	if (ret)
544 		return ret;
545 
546 	old_cred = ovl_override_creds(file_inode(file)->i_sb);
547 	ret = vfs_fadvise(real.file, offset, len, advice);
548 	revert_creds(old_cred);
549 
550 	fdput(real);
551 
552 	return ret;
553 }
554 
555 enum ovl_copyop {
556 	OVL_COPY,
557 	OVL_CLONE,
558 	OVL_DEDUPE,
559 };
560 
561 static loff_t ovl_copyfile(struct file *file_in, loff_t pos_in,
562 			    struct file *file_out, loff_t pos_out,
563 			    loff_t len, unsigned int flags, enum ovl_copyop op)
564 {
565 	struct inode *inode_out = file_inode(file_out);
566 	struct fd real_in, real_out;
567 	const struct cred *old_cred;
568 	loff_t ret;
569 
570 	ret = ovl_real_fdget(file_out, &real_out);
571 	if (ret)
572 		return ret;
573 
574 	ret = ovl_real_fdget(file_in, &real_in);
575 	if (ret) {
576 		fdput(real_out);
577 		return ret;
578 	}
579 
580 	old_cred = ovl_override_creds(file_inode(file_out)->i_sb);
581 	switch (op) {
582 	case OVL_COPY:
583 		ret = vfs_copy_file_range(real_in.file, pos_in,
584 					  real_out.file, pos_out, len, flags);
585 		break;
586 
587 	case OVL_CLONE:
588 		ret = vfs_clone_file_range(real_in.file, pos_in,
589 					   real_out.file, pos_out, len, flags);
590 		break;
591 
592 	case OVL_DEDUPE:
593 		ret = vfs_dedupe_file_range_one(real_in.file, pos_in,
594 						real_out.file, pos_out, len,
595 						flags);
596 		break;
597 	}
598 	revert_creds(old_cred);
599 
600 	/* Update size */
601 	ovl_copyattr(inode_out);
602 
603 	fdput(real_in);
604 	fdput(real_out);
605 
606 	return ret;
607 }
608 
609 static ssize_t ovl_copy_file_range(struct file *file_in, loff_t pos_in,
610 				   struct file *file_out, loff_t pos_out,
611 				   size_t len, unsigned int flags)
612 {
613 	return ovl_copyfile(file_in, pos_in, file_out, pos_out, len, flags,
614 			    OVL_COPY);
615 }
616 
617 static loff_t ovl_remap_file_range(struct file *file_in, loff_t pos_in,
618 				   struct file *file_out, loff_t pos_out,
619 				   loff_t len, unsigned int remap_flags)
620 {
621 	enum ovl_copyop op;
622 
623 	if (remap_flags & ~(REMAP_FILE_DEDUP | REMAP_FILE_ADVISORY))
624 		return -EINVAL;
625 
626 	if (remap_flags & REMAP_FILE_DEDUP)
627 		op = OVL_DEDUPE;
628 	else
629 		op = OVL_CLONE;
630 
631 	/*
632 	 * Don't copy up because of a dedupe request, this wouldn't make sense
633 	 * most of the time (data would be duplicated instead of deduplicated).
634 	 */
635 	if (op == OVL_DEDUPE &&
636 	    (!ovl_inode_upper(file_inode(file_in)) ||
637 	     !ovl_inode_upper(file_inode(file_out))))
638 		return -EPERM;
639 
640 	return ovl_copyfile(file_in, pos_in, file_out, pos_out, len,
641 			    remap_flags, op);
642 }
643 
644 static int ovl_flush(struct file *file, fl_owner_t id)
645 {
646 	struct fd real;
647 	const struct cred *old_cred;
648 	int err;
649 
650 	err = ovl_real_fdget(file, &real);
651 	if (err)
652 		return err;
653 
654 	if (real.file->f_op->flush) {
655 		old_cred = ovl_override_creds(file_inode(file)->i_sb);
656 		err = real.file->f_op->flush(real.file, id);
657 		revert_creds(old_cred);
658 	}
659 	fdput(real);
660 
661 	return err;
662 }
663 
664 const struct file_operations ovl_file_operations = {
665 	.open		= ovl_open,
666 	.release	= ovl_release,
667 	.llseek		= ovl_llseek,
668 	.read_iter	= ovl_read_iter,
669 	.write_iter	= ovl_write_iter,
670 	.fsync		= ovl_fsync,
671 	.mmap		= ovl_mmap,
672 	.fallocate	= ovl_fallocate,
673 	.fadvise	= ovl_fadvise,
674 	.flush		= ovl_flush,
675 	.splice_read    = generic_file_splice_read,
676 	.splice_write   = ovl_splice_write,
677 
678 	.copy_file_range	= ovl_copy_file_range,
679 	.remap_file_range	= ovl_remap_file_range,
680 };
681 
682 int __init ovl_aio_request_cache_init(void)
683 {
684 	ovl_aio_request_cachep = kmem_cache_create("ovl_aio_req",
685 						   sizeof(struct ovl_aio_req),
686 						   0, SLAB_HWCACHE_ALIGN, NULL);
687 	if (!ovl_aio_request_cachep)
688 		return -ENOMEM;
689 
690 	return 0;
691 }
692 
693 void ovl_aio_request_cache_destroy(void)
694 {
695 	kmem_cache_destroy(ovl_aio_request_cachep);
696 }
697