xref: /openbmc/linux/fs/overlayfs/file.c (revision 16b0314a)
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 	struct kiocb *orig_iocb;
21 	struct fd fd;
22 };
23 
24 static struct kmem_cache *ovl_aio_request_cachep;
25 
26 static char ovl_whatisit(struct inode *inode, struct inode *realinode)
27 {
28 	if (realinode != ovl_inode_upper(inode))
29 		return 'l';
30 	if (ovl_has_upperdata(inode))
31 		return 'u';
32 	else
33 		return 'm';
34 }
35 
36 /* No atime modificaton nor notify on underlying */
37 #define OVL_OPEN_FLAGS (O_NOATIME | FMODE_NONOTIFY)
38 
39 static struct file *ovl_open_realfile(const struct file *file,
40 				      struct inode *realinode)
41 {
42 	struct inode *inode = file_inode(file);
43 	struct file *realfile;
44 	const struct cred *old_cred;
45 	int flags = file->f_flags | OVL_OPEN_FLAGS;
46 	int acc_mode = ACC_MODE(flags);
47 	int err;
48 
49 	if (flags & O_APPEND)
50 		acc_mode |= MAY_APPEND;
51 
52 	old_cred = ovl_override_creds(inode->i_sb);
53 	err = inode_permission(&init_user_ns, realinode, MAY_OPEN | acc_mode);
54 	if (err) {
55 		realfile = ERR_PTR(err);
56 	} else {
57 		if (!inode_owner_or_capable(&init_user_ns, realinode))
58 			flags &= ~O_NOATIME;
59 
60 		realfile = open_with_fake_path(&file->f_path, flags, realinode,
61 					       current_cred());
62 	}
63 	revert_creds(old_cred);
64 
65 	pr_debug("open(%p[%pD2/%c], 0%o) -> (%p, 0%o)\n",
66 		 file, file, ovl_whatisit(inode, realinode), file->f_flags,
67 		 realfile, IS_ERR(realfile) ? 0 : realfile->f_flags);
68 
69 	return realfile;
70 }
71 
72 #define OVL_SETFL_MASK (O_APPEND | O_NONBLOCK | O_NDELAY | O_DIRECT)
73 
74 static int ovl_change_flags(struct file *file, unsigned int flags)
75 {
76 	struct inode *inode = file_inode(file);
77 	int err;
78 
79 	flags &= OVL_SETFL_MASK;
80 
81 	if (((flags ^ file->f_flags) & O_APPEND) && IS_APPEND(inode))
82 		return -EPERM;
83 
84 	if (flags & O_DIRECT) {
85 		if (!file->f_mapping->a_ops ||
86 		    !file->f_mapping->a_ops->direct_IO)
87 			return -EINVAL;
88 	}
89 
90 	if (file->f_op->check_flags) {
91 		err = file->f_op->check_flags(flags);
92 		if (err)
93 			return err;
94 	}
95 
96 	spin_lock(&file->f_lock);
97 	file->f_flags = (file->f_flags & ~OVL_SETFL_MASK) | flags;
98 	spin_unlock(&file->f_lock);
99 
100 	return 0;
101 }
102 
103 static int ovl_real_fdget_meta(const struct file *file, struct fd *real,
104 			       bool allow_meta)
105 {
106 	struct inode *inode = file_inode(file);
107 	struct inode *realinode;
108 
109 	real->flags = 0;
110 	real->file = file->private_data;
111 
112 	if (allow_meta)
113 		realinode = ovl_inode_real(inode);
114 	else
115 		realinode = ovl_inode_realdata(inode);
116 
117 	/* Has it been copied up since we'd opened it? */
118 	if (unlikely(file_inode(real->file) != realinode)) {
119 		real->flags = FDPUT_FPUT;
120 		real->file = ovl_open_realfile(file, realinode);
121 
122 		return PTR_ERR_OR_ZERO(real->file);
123 	}
124 
125 	/* Did the flags change since open? */
126 	if (unlikely((file->f_flags ^ real->file->f_flags) & ~OVL_OPEN_FLAGS))
127 		return ovl_change_flags(real->file, file->f_flags);
128 
129 	return 0;
130 }
131 
132 static int ovl_real_fdget(const struct file *file, struct fd *real)
133 {
134 	if (d_is_dir(file_dentry(file))) {
135 		real->flags = 0;
136 		real->file = ovl_dir_real_file(file, false);
137 
138 		return PTR_ERR_OR_ZERO(real->file);
139 	}
140 
141 	return ovl_real_fdget_meta(file, real, false);
142 }
143 
144 static int ovl_open(struct inode *inode, struct file *file)
145 {
146 	struct file *realfile;
147 	int err;
148 
149 	err = ovl_maybe_copy_up(file_dentry(file), file->f_flags);
150 	if (err)
151 		return err;
152 
153 	/* No longer need these flags, so don't pass them on to underlying fs */
154 	file->f_flags &= ~(O_CREAT | O_EXCL | O_NOCTTY | O_TRUNC);
155 
156 	realfile = ovl_open_realfile(file, ovl_inode_realdata(inode));
157 	if (IS_ERR(realfile))
158 		return PTR_ERR(realfile);
159 
160 	file->private_data = realfile;
161 
162 	return 0;
163 }
164 
165 static int ovl_release(struct inode *inode, struct file *file)
166 {
167 	fput(file->private_data);
168 
169 	return 0;
170 }
171 
172 static loff_t ovl_llseek(struct file *file, loff_t offset, int whence)
173 {
174 	struct inode *inode = file_inode(file);
175 	struct fd real;
176 	const struct cred *old_cred;
177 	loff_t ret;
178 
179 	/*
180 	 * The two special cases below do not need to involve real fs,
181 	 * so we can optimizing concurrent callers.
182 	 */
183 	if (offset == 0) {
184 		if (whence == SEEK_CUR)
185 			return file->f_pos;
186 
187 		if (whence == SEEK_SET)
188 			return vfs_setpos(file, 0, 0);
189 	}
190 
191 	ret = ovl_real_fdget(file, &real);
192 	if (ret)
193 		return ret;
194 
195 	/*
196 	 * Overlay file f_pos is the master copy that is preserved
197 	 * through copy up and modified on read/write, but only real
198 	 * fs knows how to SEEK_HOLE/SEEK_DATA and real fs may impose
199 	 * limitations that are more strict than ->s_maxbytes for specific
200 	 * files, so we use the real file to perform seeks.
201 	 */
202 	ovl_inode_lock(inode);
203 	real.file->f_pos = file->f_pos;
204 
205 	old_cred = ovl_override_creds(inode->i_sb);
206 	ret = vfs_llseek(real.file, offset, whence);
207 	revert_creds(old_cred);
208 
209 	file->f_pos = real.file->f_pos;
210 	ovl_inode_unlock(inode);
211 
212 	fdput(real);
213 
214 	return ret;
215 }
216 
217 static void ovl_file_accessed(struct file *file)
218 {
219 	struct inode *inode, *upperinode;
220 
221 	if (file->f_flags & O_NOATIME)
222 		return;
223 
224 	inode = file_inode(file);
225 	upperinode = ovl_inode_upper(inode);
226 
227 	if (!upperinode)
228 		return;
229 
230 	if ((!timespec64_equal(&inode->i_mtime, &upperinode->i_mtime) ||
231 	     !timespec64_equal(&inode->i_ctime, &upperinode->i_ctime))) {
232 		inode->i_mtime = upperinode->i_mtime;
233 		inode->i_ctime = upperinode->i_ctime;
234 	}
235 
236 	touch_atime(&file->f_path);
237 }
238 
239 static rwf_t ovl_iocb_to_rwf(int ifl)
240 {
241 	rwf_t flags = 0;
242 
243 	if (ifl & IOCB_NOWAIT)
244 		flags |= RWF_NOWAIT;
245 	if (ifl & IOCB_HIPRI)
246 		flags |= RWF_HIPRI;
247 	if (ifl & IOCB_DSYNC)
248 		flags |= RWF_DSYNC;
249 	if (ifl & IOCB_SYNC)
250 		flags |= RWF_SYNC;
251 
252 	return flags;
253 }
254 
255 static void ovl_aio_cleanup_handler(struct ovl_aio_req *aio_req)
256 {
257 	struct kiocb *iocb = &aio_req->iocb;
258 	struct kiocb *orig_iocb = aio_req->orig_iocb;
259 
260 	if (iocb->ki_flags & IOCB_WRITE) {
261 		struct inode *inode = file_inode(orig_iocb->ki_filp);
262 
263 		/* Actually acquired in ovl_write_iter() */
264 		__sb_writers_acquired(file_inode(iocb->ki_filp)->i_sb,
265 				      SB_FREEZE_WRITE);
266 		file_end_write(iocb->ki_filp);
267 		ovl_copyattr(ovl_inode_real(inode), inode);
268 	}
269 
270 	orig_iocb->ki_pos = iocb->ki_pos;
271 	fdput(aio_req->fd);
272 	kmem_cache_free(ovl_aio_request_cachep, aio_req);
273 }
274 
275 static void ovl_aio_rw_complete(struct kiocb *iocb, long res, long res2)
276 {
277 	struct ovl_aio_req *aio_req = container_of(iocb,
278 						   struct ovl_aio_req, iocb);
279 	struct kiocb *orig_iocb = aio_req->orig_iocb;
280 
281 	ovl_aio_cleanup_handler(aio_req);
282 	orig_iocb->ki_complete(orig_iocb, res, res2);
283 }
284 
285 static ssize_t ovl_read_iter(struct kiocb *iocb, struct iov_iter *iter)
286 {
287 	struct file *file = iocb->ki_filp;
288 	struct fd real;
289 	const struct cred *old_cred;
290 	ssize_t ret;
291 
292 	if (!iov_iter_count(iter))
293 		return 0;
294 
295 	ret = ovl_real_fdget(file, &real);
296 	if (ret)
297 		return ret;
298 
299 	ret = -EINVAL;
300 	if (iocb->ki_flags & IOCB_DIRECT &&
301 	    (!real.file->f_mapping->a_ops ||
302 	     !real.file->f_mapping->a_ops->direct_IO))
303 		goto out_fdput;
304 
305 	old_cred = ovl_override_creds(file_inode(file)->i_sb);
306 	if (is_sync_kiocb(iocb)) {
307 		ret = vfs_iter_read(real.file, iter, &iocb->ki_pos,
308 				    ovl_iocb_to_rwf(iocb->ki_flags));
309 	} else {
310 		struct ovl_aio_req *aio_req;
311 
312 		ret = -ENOMEM;
313 		aio_req = kmem_cache_zalloc(ovl_aio_request_cachep, GFP_KERNEL);
314 		if (!aio_req)
315 			goto out;
316 
317 		aio_req->fd = real;
318 		real.flags = 0;
319 		aio_req->orig_iocb = iocb;
320 		kiocb_clone(&aio_req->iocb, iocb, real.file);
321 		aio_req->iocb.ki_complete = ovl_aio_rw_complete;
322 		ret = vfs_iocb_iter_read(real.file, &aio_req->iocb, iter);
323 		if (ret != -EIOCBQUEUED)
324 			ovl_aio_cleanup_handler(aio_req);
325 	}
326 out:
327 	revert_creds(old_cred);
328 	ovl_file_accessed(file);
329 out_fdput:
330 	fdput(real);
331 
332 	return ret;
333 }
334 
335 static ssize_t ovl_write_iter(struct kiocb *iocb, struct iov_iter *iter)
336 {
337 	struct file *file = iocb->ki_filp;
338 	struct inode *inode = file_inode(file);
339 	struct fd real;
340 	const struct cred *old_cred;
341 	ssize_t ret;
342 	int ifl = iocb->ki_flags;
343 
344 	if (!iov_iter_count(iter))
345 		return 0;
346 
347 	inode_lock(inode);
348 	/* Update mode */
349 	ovl_copyattr(ovl_inode_real(inode), inode);
350 	ret = file_remove_privs(file);
351 	if (ret)
352 		goto out_unlock;
353 
354 	ret = ovl_real_fdget(file, &real);
355 	if (ret)
356 		goto out_unlock;
357 
358 	ret = -EINVAL;
359 	if (iocb->ki_flags & IOCB_DIRECT &&
360 	    (!real.file->f_mapping->a_ops ||
361 	     !real.file->f_mapping->a_ops->direct_IO))
362 		goto out_fdput;
363 
364 	if (!ovl_should_sync(OVL_FS(inode->i_sb)))
365 		ifl &= ~(IOCB_DSYNC | IOCB_SYNC);
366 
367 	old_cred = ovl_override_creds(file_inode(file)->i_sb);
368 	if (is_sync_kiocb(iocb)) {
369 		file_start_write(real.file);
370 		ret = vfs_iter_write(real.file, iter, &iocb->ki_pos,
371 				     ovl_iocb_to_rwf(ifl));
372 		file_end_write(real.file);
373 		/* Update size */
374 		ovl_copyattr(ovl_inode_real(inode), inode);
375 	} else {
376 		struct ovl_aio_req *aio_req;
377 
378 		ret = -ENOMEM;
379 		aio_req = kmem_cache_zalloc(ovl_aio_request_cachep, GFP_KERNEL);
380 		if (!aio_req)
381 			goto out;
382 
383 		file_start_write(real.file);
384 		/* Pacify lockdep, same trick as done in aio_write() */
385 		__sb_writers_release(file_inode(real.file)->i_sb,
386 				     SB_FREEZE_WRITE);
387 		aio_req->fd = real;
388 		real.flags = 0;
389 		aio_req->orig_iocb = iocb;
390 		kiocb_clone(&aio_req->iocb, iocb, real.file);
391 		aio_req->iocb.ki_flags = ifl;
392 		aio_req->iocb.ki_complete = ovl_aio_rw_complete;
393 		ret = vfs_iocb_iter_write(real.file, &aio_req->iocb, iter);
394 		if (ret != -EIOCBQUEUED)
395 			ovl_aio_cleanup_handler(aio_req);
396 	}
397 out:
398 	revert_creds(old_cred);
399 out_fdput:
400 	fdput(real);
401 
402 out_unlock:
403 	inode_unlock(inode);
404 
405 	return ret;
406 }
407 
408 /*
409  * Calling iter_file_splice_write() directly from overlay's f_op may deadlock
410  * due to lock order inversion between pipe->mutex in iter_file_splice_write()
411  * and file_start_write(real.file) in ovl_write_iter().
412  *
413  * So do everything ovl_write_iter() does and call iter_file_splice_write() on
414  * the real file.
415  */
416 static ssize_t ovl_splice_write(struct pipe_inode_info *pipe, struct file *out,
417 				loff_t *ppos, size_t len, unsigned int flags)
418 {
419 	struct fd real;
420 	const struct cred *old_cred;
421 	struct inode *inode = file_inode(out);
422 	struct inode *realinode = ovl_inode_real(inode);
423 	ssize_t ret;
424 
425 	inode_lock(inode);
426 	/* Update mode */
427 	ovl_copyattr(realinode, inode);
428 	ret = file_remove_privs(out);
429 	if (ret)
430 		goto out_unlock;
431 
432 	ret = ovl_real_fdget(out, &real);
433 	if (ret)
434 		goto out_unlock;
435 
436 	old_cred = ovl_override_creds(inode->i_sb);
437 	file_start_write(real.file);
438 
439 	ret = iter_file_splice_write(pipe, real.file, ppos, len, flags);
440 
441 	file_end_write(real.file);
442 	/* Update size */
443 	ovl_copyattr(realinode, inode);
444 	revert_creds(old_cred);
445 	fdput(real);
446 
447 out_unlock:
448 	inode_unlock(inode);
449 
450 	return ret;
451 }
452 
453 static int ovl_fsync(struct file *file, loff_t start, loff_t end, int datasync)
454 {
455 	struct fd real;
456 	const struct cred *old_cred;
457 	int ret;
458 
459 	ret = ovl_sync_status(OVL_FS(file_inode(file)->i_sb));
460 	if (ret <= 0)
461 		return ret;
462 
463 	ret = ovl_real_fdget_meta(file, &real, !datasync);
464 	if (ret)
465 		return ret;
466 
467 	/* Don't sync lower file for fear of receiving EROFS error */
468 	if (file_inode(real.file) == ovl_inode_upper(file_inode(file))) {
469 		old_cred = ovl_override_creds(file_inode(file)->i_sb);
470 		ret = vfs_fsync_range(real.file, start, end, datasync);
471 		revert_creds(old_cred);
472 	}
473 
474 	fdput(real);
475 
476 	return ret;
477 }
478 
479 static int ovl_mmap(struct file *file, struct vm_area_struct *vma)
480 {
481 	struct file *realfile = file->private_data;
482 	const struct cred *old_cred;
483 	int ret;
484 
485 	if (!realfile->f_op->mmap)
486 		return -ENODEV;
487 
488 	if (WARN_ON(file != vma->vm_file))
489 		return -EIO;
490 
491 	vma_set_file(vma, realfile);
492 
493 	old_cred = ovl_override_creds(file_inode(file)->i_sb);
494 	ret = call_mmap(vma->vm_file, vma);
495 	revert_creds(old_cred);
496 	ovl_file_accessed(file);
497 
498 	return ret;
499 }
500 
501 static long ovl_fallocate(struct file *file, int mode, loff_t offset, loff_t len)
502 {
503 	struct inode *inode = file_inode(file);
504 	struct fd real;
505 	const struct cred *old_cred;
506 	int ret;
507 
508 	ret = ovl_real_fdget(file, &real);
509 	if (ret)
510 		return ret;
511 
512 	old_cred = ovl_override_creds(file_inode(file)->i_sb);
513 	ret = vfs_fallocate(real.file, mode, offset, len);
514 	revert_creds(old_cred);
515 
516 	/* Update size */
517 	ovl_copyattr(ovl_inode_real(inode), inode);
518 
519 	fdput(real);
520 
521 	return ret;
522 }
523 
524 static int ovl_fadvise(struct file *file, loff_t offset, loff_t len, int advice)
525 {
526 	struct fd real;
527 	const struct cred *old_cred;
528 	int ret;
529 
530 	ret = ovl_real_fdget(file, &real);
531 	if (ret)
532 		return ret;
533 
534 	old_cred = ovl_override_creds(file_inode(file)->i_sb);
535 	ret = vfs_fadvise(real.file, offset, len, advice);
536 	revert_creds(old_cred);
537 
538 	fdput(real);
539 
540 	return ret;
541 }
542 
543 enum ovl_copyop {
544 	OVL_COPY,
545 	OVL_CLONE,
546 	OVL_DEDUPE,
547 };
548 
549 static loff_t ovl_copyfile(struct file *file_in, loff_t pos_in,
550 			    struct file *file_out, loff_t pos_out,
551 			    loff_t len, unsigned int flags, enum ovl_copyop op)
552 {
553 	struct inode *inode_out = file_inode(file_out);
554 	struct fd real_in, real_out;
555 	const struct cred *old_cred;
556 	loff_t ret;
557 
558 	ret = ovl_real_fdget(file_out, &real_out);
559 	if (ret)
560 		return ret;
561 
562 	ret = ovl_real_fdget(file_in, &real_in);
563 	if (ret) {
564 		fdput(real_out);
565 		return ret;
566 	}
567 
568 	old_cred = ovl_override_creds(file_inode(file_out)->i_sb);
569 	switch (op) {
570 	case OVL_COPY:
571 		ret = vfs_copy_file_range(real_in.file, pos_in,
572 					  real_out.file, pos_out, len, flags);
573 		break;
574 
575 	case OVL_CLONE:
576 		ret = vfs_clone_file_range(real_in.file, pos_in,
577 					   real_out.file, pos_out, len, flags);
578 		break;
579 
580 	case OVL_DEDUPE:
581 		ret = vfs_dedupe_file_range_one(real_in.file, pos_in,
582 						real_out.file, pos_out, len,
583 						flags);
584 		break;
585 	}
586 	revert_creds(old_cred);
587 
588 	/* Update size */
589 	ovl_copyattr(ovl_inode_real(inode_out), inode_out);
590 
591 	fdput(real_in);
592 	fdput(real_out);
593 
594 	return ret;
595 }
596 
597 static ssize_t ovl_copy_file_range(struct file *file_in, loff_t pos_in,
598 				   struct file *file_out, loff_t pos_out,
599 				   size_t len, unsigned int flags)
600 {
601 	return ovl_copyfile(file_in, pos_in, file_out, pos_out, len, flags,
602 			    OVL_COPY);
603 }
604 
605 static loff_t ovl_remap_file_range(struct file *file_in, loff_t pos_in,
606 				   struct file *file_out, loff_t pos_out,
607 				   loff_t len, unsigned int remap_flags)
608 {
609 	enum ovl_copyop op;
610 
611 	if (remap_flags & ~(REMAP_FILE_DEDUP | REMAP_FILE_ADVISORY))
612 		return -EINVAL;
613 
614 	if (remap_flags & REMAP_FILE_DEDUP)
615 		op = OVL_DEDUPE;
616 	else
617 		op = OVL_CLONE;
618 
619 	/*
620 	 * Don't copy up because of a dedupe request, this wouldn't make sense
621 	 * most of the time (data would be duplicated instead of deduplicated).
622 	 */
623 	if (op == OVL_DEDUPE &&
624 	    (!ovl_inode_upper(file_inode(file_in)) ||
625 	     !ovl_inode_upper(file_inode(file_out))))
626 		return -EPERM;
627 
628 	return ovl_copyfile(file_in, pos_in, file_out, pos_out, len,
629 			    remap_flags, op);
630 }
631 
632 static int ovl_flush(struct file *file, fl_owner_t id)
633 {
634 	struct fd real;
635 	const struct cred *old_cred;
636 	int err;
637 
638 	err = ovl_real_fdget(file, &real);
639 	if (err)
640 		return err;
641 
642 	if (real.file->f_op->flush) {
643 		old_cred = ovl_override_creds(file_inode(file)->i_sb);
644 		err = real.file->f_op->flush(real.file, id);
645 		revert_creds(old_cred);
646 	}
647 	fdput(real);
648 
649 	return err;
650 }
651 
652 const struct file_operations ovl_file_operations = {
653 	.open		= ovl_open,
654 	.release	= ovl_release,
655 	.llseek		= ovl_llseek,
656 	.read_iter	= ovl_read_iter,
657 	.write_iter	= ovl_write_iter,
658 	.fsync		= ovl_fsync,
659 	.mmap		= ovl_mmap,
660 	.fallocate	= ovl_fallocate,
661 	.fadvise	= ovl_fadvise,
662 	.flush		= ovl_flush,
663 	.splice_read    = generic_file_splice_read,
664 	.splice_write   = ovl_splice_write,
665 
666 	.copy_file_range	= ovl_copy_file_range,
667 	.remap_file_range	= ovl_remap_file_range,
668 };
669 
670 int __init ovl_aio_request_cache_init(void)
671 {
672 	ovl_aio_request_cachep = kmem_cache_create("ovl_aio_req",
673 						   sizeof(struct ovl_aio_req),
674 						   0, SLAB_HWCACHE_ALIGN, NULL);
675 	if (!ovl_aio_request_cachep)
676 		return -ENOMEM;
677 
678 	return 0;
679 }
680 
681 void ovl_aio_request_cache_destroy(void)
682 {
683 	kmem_cache_destroy(ovl_aio_request_cachep);
684 }
685