Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * linux/fs/ext4/file.c
4 *
5 * Copyright (C) 1992, 1993, 1994, 1995
6 * Remy Card (card@masi.ibp.fr)
7 * Laboratoire MASI - Institut Blaise Pascal
8 * Universite Pierre et Marie Curie (Paris VI)
9 *
10 * from
11 *
12 * linux/fs/minix/file.c
13 *
14 * Copyright (C) 1991, 1992 Linus Torvalds
15 *
16 * ext4 fs regular file handling primitives
17 *
18 * 64-bit file support on 64-bit platforms by Jakub Jelinek
19 * (jj@sunsite.ms.mff.cuni.cz)
20 */
21
22#include <linux/time.h>
23#include <linux/fs.h>
24#include <linux/iomap.h>
25#include <linux/mount.h>
26#include <linux/path.h>
27#include <linux/dax.h>
28#include <linux/quotaops.h>
29#include <linux/pagevec.h>
30#include <linux/uio.h>
31#include <linux/mman.h>
32#include <linux/backing-dev.h>
33#include "ext4.h"
34#include "ext4_jbd2.h"
35#include "xattr.h"
36#include "acl.h"
37#include "truncate.h"
38
39static bool ext4_dio_supported(struct inode *inode)
40{
41 if (IS_ENABLED(CONFIG_FS_ENCRYPTION) && IS_ENCRYPTED(inode))
42 return false;
43 if (fsverity_active(inode))
44 return false;
45 if (ext4_should_journal_data(inode))
46 return false;
47 if (ext4_has_inline_data(inode))
48 return false;
49 return true;
50}
51
52static ssize_t ext4_dio_read_iter(struct kiocb *iocb, struct iov_iter *to)
53{
54 ssize_t ret;
55 struct inode *inode = file_inode(iocb->ki_filp);
56
57 if (iocb->ki_flags & IOCB_NOWAIT) {
58 if (!inode_trylock_shared(inode))
59 return -EAGAIN;
60 } else {
61 inode_lock_shared(inode);
62 }
63
64 if (!ext4_dio_supported(inode)) {
65 inode_unlock_shared(inode);
66 /*
67 * Fallback to buffered I/O if the operation being performed on
68 * the inode is not supported by direct I/O. The IOCB_DIRECT
69 * flag needs to be cleared here in order to ensure that the
70 * direct I/O path within generic_file_read_iter() is not
71 * taken.
72 */
73 iocb->ki_flags &= ~IOCB_DIRECT;
74 return generic_file_read_iter(iocb, to);
75 }
76
77 ret = iomap_dio_rw(iocb, to, &ext4_iomap_ops, NULL,
78 is_sync_kiocb(iocb));
79 inode_unlock_shared(inode);
80
81 file_accessed(iocb->ki_filp);
82 return ret;
83}
84
85#ifdef CONFIG_FS_DAX
86static ssize_t ext4_dax_read_iter(struct kiocb *iocb, struct iov_iter *to)
87{
88 struct inode *inode = file_inode(iocb->ki_filp);
89 ssize_t ret;
90
91 if (iocb->ki_flags & IOCB_NOWAIT) {
92 if (!inode_trylock_shared(inode))
93 return -EAGAIN;
94 } else {
95 inode_lock_shared(inode);
96 }
97 /*
98 * Recheck under inode lock - at this point we are sure it cannot
99 * change anymore
100 */
101 if (!IS_DAX(inode)) {
102 inode_unlock_shared(inode);
103 /* Fallback to buffered IO in case we cannot support DAX */
104 return generic_file_read_iter(iocb, to);
105 }
106 ret = dax_iomap_rw(iocb, to, &ext4_iomap_ops);
107 inode_unlock_shared(inode);
108
109 file_accessed(iocb->ki_filp);
110 return ret;
111}
112#endif
113
114static ssize_t ext4_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
115{
116 struct inode *inode = file_inode(iocb->ki_filp);
117
118 if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
119 return -EIO;
120
121 if (!iov_iter_count(to))
122 return 0; /* skip atime */
123
124#ifdef CONFIG_FS_DAX
125 if (IS_DAX(inode))
126 return ext4_dax_read_iter(iocb, to);
127#endif
128 if (iocb->ki_flags & IOCB_DIRECT)
129 return ext4_dio_read_iter(iocb, to);
130
131 return generic_file_read_iter(iocb, to);
132}
133
134/*
135 * Called when an inode is released. Note that this is different
136 * from ext4_file_open: open gets called at every open, but release
137 * gets called only when /all/ the files are closed.
138 */
139static int ext4_release_file(struct inode *inode, struct file *filp)
140{
141 if (ext4_test_inode_state(inode, EXT4_STATE_DA_ALLOC_CLOSE)) {
142 ext4_alloc_da_blocks(inode);
143 ext4_clear_inode_state(inode, EXT4_STATE_DA_ALLOC_CLOSE);
144 }
145 /* if we are the last writer on the inode, drop the block reservation */
146 if ((filp->f_mode & FMODE_WRITE) &&
147 (atomic_read(&inode->i_writecount) == 1) &&
148 !EXT4_I(inode)->i_reserved_data_blocks) {
149 down_write(&EXT4_I(inode)->i_data_sem);
150 ext4_discard_preallocations(inode, 0);
151 up_write(&EXT4_I(inode)->i_data_sem);
152 }
153 if (is_dx(inode) && filp->private_data)
154 ext4_htree_free_dir_info(filp->private_data);
155
156 return 0;
157}
158
159/*
160 * This tests whether the IO in question is block-aligned or not.
161 * Ext4 utilizes unwritten extents when hole-filling during direct IO, and they
162 * are converted to written only after the IO is complete. Until they are
163 * mapped, these blocks appear as holes, so dio_zero_block() will assume that
164 * it needs to zero out portions of the start and/or end block. If 2 AIO
165 * threads are at work on the same unwritten block, they must be synchronized
166 * or one thread will zero the other's data, causing corruption.
167 */
168static bool
169ext4_unaligned_io(struct inode *inode, struct iov_iter *from, loff_t pos)
170{
171 struct super_block *sb = inode->i_sb;
172 unsigned long blockmask = sb->s_blocksize - 1;
173
174 if ((pos | iov_iter_alignment(from)) & blockmask)
175 return true;
176
177 return false;
178}
179
180static bool
181ext4_extending_io(struct inode *inode, loff_t offset, size_t len)
182{
183 if (offset + len > i_size_read(inode) ||
184 offset + len > EXT4_I(inode)->i_disksize)
185 return true;
186 return false;
187}
188
189/* Is IO overwriting allocated and initialized blocks? */
190static bool ext4_overwrite_io(struct inode *inode, loff_t pos, loff_t len)
191{
192 struct ext4_map_blocks map;
193 unsigned int blkbits = inode->i_blkbits;
194 int err, blklen;
195
196 if (pos + len > i_size_read(inode))
197 return false;
198
199 map.m_lblk = pos >> blkbits;
200 map.m_len = EXT4_MAX_BLOCKS(len, pos, blkbits);
201 blklen = map.m_len;
202
203 err = ext4_map_blocks(NULL, inode, &map, 0);
204 /*
205 * 'err==len' means that all of the blocks have been preallocated,
206 * regardless of whether they have been initialized or not. To exclude
207 * unwritten extents, we need to check m_flags.
208 */
209 return err == blklen && (map.m_flags & EXT4_MAP_MAPPED);
210}
211
212static ssize_t ext4_generic_write_checks(struct kiocb *iocb,
213 struct iov_iter *from)
214{
215 struct inode *inode = file_inode(iocb->ki_filp);
216 ssize_t ret;
217
218 if (unlikely(IS_IMMUTABLE(inode)))
219 return -EPERM;
220
221 ret = generic_write_checks(iocb, from);
222 if (ret <= 0)
223 return ret;
224
225 /*
226 * If we have encountered a bitmap-format file, the size limit
227 * is smaller than s_maxbytes, which is for extent-mapped files.
228 */
229 if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
230 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
231
232 if (iocb->ki_pos >= sbi->s_bitmap_maxbytes)
233 return -EFBIG;
234 iov_iter_truncate(from, sbi->s_bitmap_maxbytes - iocb->ki_pos);
235 }
236
237 return iov_iter_count(from);
238}
239
240static ssize_t ext4_write_checks(struct kiocb *iocb, struct iov_iter *from)
241{
242 ssize_t ret, count;
243
244 count = ext4_generic_write_checks(iocb, from);
245 if (count <= 0)
246 return count;
247
248 ret = file_modified(iocb->ki_filp);
249 if (ret)
250 return ret;
251 return count;
252}
253
254static ssize_t ext4_buffered_write_iter(struct kiocb *iocb,
255 struct iov_iter *from)
256{
257 ssize_t ret;
258 struct inode *inode = file_inode(iocb->ki_filp);
259
260 if (iocb->ki_flags & IOCB_NOWAIT)
261 return -EOPNOTSUPP;
262
263 inode_lock(inode);
264 ret = ext4_write_checks(iocb, from);
265 if (ret <= 0)
266 goto out;
267
268 current->backing_dev_info = inode_to_bdi(inode);
269 ret = generic_perform_write(iocb->ki_filp, from, iocb->ki_pos);
270 current->backing_dev_info = NULL;
271
272out:
273 inode_unlock(inode);
274 if (likely(ret > 0)) {
275 iocb->ki_pos += ret;
276 ret = generic_write_sync(iocb, ret);
277 }
278
279 return ret;
280}
281
282static ssize_t ext4_handle_inode_extension(struct inode *inode, loff_t offset,
283 ssize_t written, size_t count)
284{
285 handle_t *handle;
286 bool truncate = false;
287 u8 blkbits = inode->i_blkbits;
288 ext4_lblk_t written_blk, end_blk;
289 int ret;
290
291 /*
292 * Note that EXT4_I(inode)->i_disksize can get extended up to
293 * inode->i_size while the I/O was running due to writeback of delalloc
294 * blocks. But, the code in ext4_iomap_alloc() is careful to use
295 * zeroed/unwritten extents if this is possible; thus we won't leave
296 * uninitialized blocks in a file even if we didn't succeed in writing
297 * as much as we intended.
298 */
299 WARN_ON_ONCE(i_size_read(inode) < EXT4_I(inode)->i_disksize);
300 if (offset + count <= EXT4_I(inode)->i_disksize) {
301 /*
302 * We need to ensure that the inode is removed from the orphan
303 * list if it has been added prematurely, due to writeback of
304 * delalloc blocks.
305 */
306 if (!list_empty(&EXT4_I(inode)->i_orphan) && inode->i_nlink) {
307 handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
308
309 if (IS_ERR(handle)) {
310 ext4_orphan_del(NULL, inode);
311 return PTR_ERR(handle);
312 }
313
314 ext4_orphan_del(handle, inode);
315 ext4_journal_stop(handle);
316 }
317
318 return written;
319 }
320
321 if (written < 0)
322 goto truncate;
323
324 handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
325 if (IS_ERR(handle)) {
326 written = PTR_ERR(handle);
327 goto truncate;
328 }
329
330 if (ext4_update_inode_size(inode, offset + written)) {
331 ret = ext4_mark_inode_dirty(handle, inode);
332 if (unlikely(ret)) {
333 written = ret;
334 ext4_journal_stop(handle);
335 goto truncate;
336 }
337 }
338
339 /*
340 * We may need to truncate allocated but not written blocks beyond EOF.
341 */
342 written_blk = ALIGN(offset + written, 1 << blkbits);
343 end_blk = ALIGN(offset + count, 1 << blkbits);
344 if (written_blk < end_blk && ext4_can_truncate(inode))
345 truncate = true;
346
347 /*
348 * Remove the inode from the orphan list if it has been extended and
349 * everything went OK.
350 */
351 if (!truncate && inode->i_nlink)
352 ext4_orphan_del(handle, inode);
353 ext4_journal_stop(handle);
354
355 if (truncate) {
356truncate:
357 ext4_truncate_failed_write(inode);
358 /*
359 * If the truncate operation failed early, then the inode may
360 * still be on the orphan list. In that case, we need to try
361 * remove the inode from the in-memory linked list.
362 */
363 if (inode->i_nlink)
364 ext4_orphan_del(NULL, inode);
365 }
366
367 return written;
368}
369
370static int ext4_dio_write_end_io(struct kiocb *iocb, ssize_t size,
371 int error, unsigned int flags)
372{
373 loff_t offset = iocb->ki_pos;
374 struct inode *inode = file_inode(iocb->ki_filp);
375
376 if (error)
377 return error;
378
379 if (size && flags & IOMAP_DIO_UNWRITTEN)
380 return ext4_convert_unwritten_extents(NULL, inode,
381 offset, size);
382
383 return 0;
384}
385
386static const struct iomap_dio_ops ext4_dio_write_ops = {
387 .end_io = ext4_dio_write_end_io,
388};
389
390/*
391 * The intention here is to start with shared lock acquired then see if any
392 * condition requires an exclusive inode lock. If yes, then we restart the
393 * whole operation by releasing the shared lock and acquiring exclusive lock.
394 *
395 * - For unaligned_io we never take shared lock as it may cause data corruption
396 * when two unaligned IO tries to modify the same block e.g. while zeroing.
397 *
398 * - For extending writes case we don't take the shared lock, since it requires
399 * updating inode i_disksize and/or orphan handling with exclusive lock.
400 *
401 * - shared locking will only be true mostly with overwrites. Otherwise we will
402 * switch to exclusive i_rwsem lock.
403 */
404static ssize_t ext4_dio_write_checks(struct kiocb *iocb, struct iov_iter *from,
405 bool *ilock_shared, bool *extend)
406{
407 struct file *file = iocb->ki_filp;
408 struct inode *inode = file_inode(file);
409 loff_t offset;
410 size_t count;
411 ssize_t ret;
412
413restart:
414 ret = ext4_generic_write_checks(iocb, from);
415 if (ret <= 0)
416 goto out;
417
418 offset = iocb->ki_pos;
419 count = ret;
420 if (ext4_extending_io(inode, offset, count))
421 *extend = true;
422 /*
423 * Determine whether the IO operation will overwrite allocated
424 * and initialized blocks.
425 * We need exclusive i_rwsem for changing security info
426 * in file_modified().
427 */
428 if (*ilock_shared && (!IS_NOSEC(inode) || *extend ||
429 !ext4_overwrite_io(inode, offset, count))) {
430 if (iocb->ki_flags & IOCB_NOWAIT) {
431 ret = -EAGAIN;
432 goto out;
433 }
434 inode_unlock_shared(inode);
435 *ilock_shared = false;
436 inode_lock(inode);
437 goto restart;
438 }
439
440 ret = file_modified(file);
441 if (ret < 0)
442 goto out;
443
444 return count;
445out:
446 if (*ilock_shared)
447 inode_unlock_shared(inode);
448 else
449 inode_unlock(inode);
450 return ret;
451}
452
453static ssize_t ext4_dio_write_iter(struct kiocb *iocb, struct iov_iter *from)
454{
455 ssize_t ret;
456 handle_t *handle;
457 struct inode *inode = file_inode(iocb->ki_filp);
458 loff_t offset = iocb->ki_pos;
459 size_t count = iov_iter_count(from);
460 const struct iomap_ops *iomap_ops = &ext4_iomap_ops;
461 bool extend = false, unaligned_io = false;
462 bool ilock_shared = true;
463
464 /*
465 * We initially start with shared inode lock unless it is
466 * unaligned IO which needs exclusive lock anyways.
467 */
468 if (ext4_unaligned_io(inode, from, offset)) {
469 unaligned_io = true;
470 ilock_shared = false;
471 }
472 /*
473 * Quick check here without any i_rwsem lock to see if it is extending
474 * IO. A more reliable check is done in ext4_dio_write_checks() with
475 * proper locking in place.
476 */
477 if (offset + count > i_size_read(inode))
478 ilock_shared = false;
479
480 if (iocb->ki_flags & IOCB_NOWAIT) {
481 if (ilock_shared) {
482 if (!inode_trylock_shared(inode))
483 return -EAGAIN;
484 } else {
485 if (!inode_trylock(inode))
486 return -EAGAIN;
487 }
488 } else {
489 if (ilock_shared)
490 inode_lock_shared(inode);
491 else
492 inode_lock(inode);
493 }
494
495 /* Fallback to buffered I/O if the inode does not support direct I/O. */
496 if (!ext4_dio_supported(inode)) {
497 if (ilock_shared)
498 inode_unlock_shared(inode);
499 else
500 inode_unlock(inode);
501 return ext4_buffered_write_iter(iocb, from);
502 }
503
504 ret = ext4_dio_write_checks(iocb, from, &ilock_shared, &extend);
505 if (ret <= 0)
506 return ret;
507
508 /* if we're going to block and IOCB_NOWAIT is set, return -EAGAIN */
509 if ((iocb->ki_flags & IOCB_NOWAIT) && (unaligned_io || extend)) {
510 ret = -EAGAIN;
511 goto out;
512 }
513
514 offset = iocb->ki_pos;
515 count = ret;
516
517 /*
518 * Unaligned direct IO must be serialized among each other as zeroing
519 * of partial blocks of two competing unaligned IOs can result in data
520 * corruption.
521 *
522 * So we make sure we don't allow any unaligned IO in flight.
523 * For IOs where we need not wait (like unaligned non-AIO DIO),
524 * below inode_dio_wait() may anyway become a no-op, since we start
525 * with exclusive lock.
526 */
527 if (unaligned_io)
528 inode_dio_wait(inode);
529
530 if (extend) {
531 handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
532 if (IS_ERR(handle)) {
533 ret = PTR_ERR(handle);
534 goto out;
535 }
536
537 ret = ext4_orphan_add(handle, inode);
538 if (ret) {
539 ext4_journal_stop(handle);
540 goto out;
541 }
542
543 ext4_journal_stop(handle);
544 }
545
546 if (ilock_shared)
547 iomap_ops = &ext4_iomap_overwrite_ops;
548 ret = iomap_dio_rw(iocb, from, iomap_ops, &ext4_dio_write_ops,
549 is_sync_kiocb(iocb) || unaligned_io || extend);
550 if (ret == -ENOTBLK)
551 ret = 0;
552
553 if (extend)
554 ret = ext4_handle_inode_extension(inode, offset, ret, count);
555
556out:
557 if (ilock_shared)
558 inode_unlock_shared(inode);
559 else
560 inode_unlock(inode);
561
562 if (ret >= 0 && iov_iter_count(from)) {
563 ssize_t err;
564 loff_t endbyte;
565
566 offset = iocb->ki_pos;
567 err = ext4_buffered_write_iter(iocb, from);
568 if (err < 0)
569 return err;
570
571 /*
572 * We need to ensure that the pages within the page cache for
573 * the range covered by this I/O are written to disk and
574 * invalidated. This is in attempt to preserve the expected
575 * direct I/O semantics in the case we fallback to buffered I/O
576 * to complete off the I/O request.
577 */
578 ret += err;
579 endbyte = offset + err - 1;
580 err = filemap_write_and_wait_range(iocb->ki_filp->f_mapping,
581 offset, endbyte);
582 if (!err)
583 invalidate_mapping_pages(iocb->ki_filp->f_mapping,
584 offset >> PAGE_SHIFT,
585 endbyte >> PAGE_SHIFT);
586 }
587
588 return ret;
589}
590
591#ifdef CONFIG_FS_DAX
592static ssize_t
593ext4_dax_write_iter(struct kiocb *iocb, struct iov_iter *from)
594{
595 ssize_t ret;
596 size_t count;
597 loff_t offset;
598 handle_t *handle;
599 bool extend = false;
600 struct inode *inode = file_inode(iocb->ki_filp);
601
602 if (iocb->ki_flags & IOCB_NOWAIT) {
603 if (!inode_trylock(inode))
604 return -EAGAIN;
605 } else {
606 inode_lock(inode);
607 }
608
609 ret = ext4_write_checks(iocb, from);
610 if (ret <= 0)
611 goto out;
612
613 offset = iocb->ki_pos;
614 count = iov_iter_count(from);
615
616 if (offset + count > EXT4_I(inode)->i_disksize) {
617 handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
618 if (IS_ERR(handle)) {
619 ret = PTR_ERR(handle);
620 goto out;
621 }
622
623 ret = ext4_orphan_add(handle, inode);
624 if (ret) {
625 ext4_journal_stop(handle);
626 goto out;
627 }
628
629 extend = true;
630 ext4_journal_stop(handle);
631 }
632
633 ret = dax_iomap_rw(iocb, from, &ext4_iomap_ops);
634
635 if (extend)
636 ret = ext4_handle_inode_extension(inode, offset, ret, count);
637out:
638 inode_unlock(inode);
639 if (ret > 0)
640 ret = generic_write_sync(iocb, ret);
641 return ret;
642}
643#endif
644
645static ssize_t
646ext4_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
647{
648 struct inode *inode = file_inode(iocb->ki_filp);
649
650 if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
651 return -EIO;
652
653#ifdef CONFIG_FS_DAX
654 if (IS_DAX(inode))
655 return ext4_dax_write_iter(iocb, from);
656#endif
657 if (iocb->ki_flags & IOCB_DIRECT)
658 return ext4_dio_write_iter(iocb, from);
659
660 return ext4_buffered_write_iter(iocb, from);
661}
662
663#ifdef CONFIG_FS_DAX
664static vm_fault_t ext4_dax_huge_fault(struct vm_fault *vmf,
665 enum page_entry_size pe_size)
666{
667 int error = 0;
668 vm_fault_t result;
669 int retries = 0;
670 handle_t *handle = NULL;
671 struct inode *inode = file_inode(vmf->vma->vm_file);
672 struct super_block *sb = inode->i_sb;
673
674 /*
675 * We have to distinguish real writes from writes which will result in a
676 * COW page; COW writes should *not* poke the journal (the file will not
677 * be changed). Doing so would cause unintended failures when mounted
678 * read-only.
679 *
680 * We check for VM_SHARED rather than vmf->cow_page since the latter is
681 * unset for pe_size != PE_SIZE_PTE (i.e. only in do_cow_fault); for
682 * other sizes, dax_iomap_fault will handle splitting / fallback so that
683 * we eventually come back with a COW page.
684 */
685 bool write = (vmf->flags & FAULT_FLAG_WRITE) &&
686 (vmf->vma->vm_flags & VM_SHARED);
687 pfn_t pfn;
688
689 if (write) {
690 sb_start_pagefault(sb);
691 file_update_time(vmf->vma->vm_file);
692 down_read(&EXT4_I(inode)->i_mmap_sem);
693retry:
694 handle = ext4_journal_start_sb(sb, EXT4_HT_WRITE_PAGE,
695 EXT4_DATA_TRANS_BLOCKS(sb));
696 if (IS_ERR(handle)) {
697 up_read(&EXT4_I(inode)->i_mmap_sem);
698 sb_end_pagefault(sb);
699 return VM_FAULT_SIGBUS;
700 }
701 } else {
702 down_read(&EXT4_I(inode)->i_mmap_sem);
703 }
704 result = dax_iomap_fault(vmf, pe_size, &pfn, &error, &ext4_iomap_ops);
705 if (write) {
706 ext4_journal_stop(handle);
707
708 if ((result & VM_FAULT_ERROR) && error == -ENOSPC &&
709 ext4_should_retry_alloc(sb, &retries))
710 goto retry;
711 /* Handling synchronous page fault? */
712 if (result & VM_FAULT_NEEDDSYNC)
713 result = dax_finish_sync_fault(vmf, pe_size, pfn);
714 up_read(&EXT4_I(inode)->i_mmap_sem);
715 sb_end_pagefault(sb);
716 } else {
717 up_read(&EXT4_I(inode)->i_mmap_sem);
718 }
719
720 return result;
721}
722
723static vm_fault_t ext4_dax_fault(struct vm_fault *vmf)
724{
725 return ext4_dax_huge_fault(vmf, PE_SIZE_PTE);
726}
727
728static const struct vm_operations_struct ext4_dax_vm_ops = {
729 .fault = ext4_dax_fault,
730 .huge_fault = ext4_dax_huge_fault,
731 .page_mkwrite = ext4_dax_fault,
732 .pfn_mkwrite = ext4_dax_fault,
733};
734#else
735#define ext4_dax_vm_ops ext4_file_vm_ops
736#endif
737
738static const struct vm_operations_struct ext4_file_vm_ops = {
739 .fault = ext4_filemap_fault,
740 .map_pages = filemap_map_pages,
741 .page_mkwrite = ext4_page_mkwrite,
742};
743
744static int ext4_file_mmap(struct file *file, struct vm_area_struct *vma)
745{
746 struct inode *inode = file->f_mapping->host;
747 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
748 struct dax_device *dax_dev = sbi->s_daxdev;
749
750 if (unlikely(ext4_forced_shutdown(sbi)))
751 return -EIO;
752
753 /*
754 * We don't support synchronous mappings for non-DAX files and
755 * for DAX files if underneath dax_device is not synchronous.
756 */
757 if (!daxdev_mapping_supported(vma, dax_dev))
758 return -EOPNOTSUPP;
759
760 file_accessed(file);
761 if (IS_DAX(file_inode(file))) {
762 vma->vm_ops = &ext4_dax_vm_ops;
763 vma->vm_flags |= VM_HUGEPAGE;
764 } else {
765 vma->vm_ops = &ext4_file_vm_ops;
766 }
767 return 0;
768}
769
770static int ext4_sample_last_mounted(struct super_block *sb,
771 struct vfsmount *mnt)
772{
773 struct ext4_sb_info *sbi = EXT4_SB(sb);
774 struct path path;
775 char buf[64], *cp;
776 handle_t *handle;
777 int err;
778
779 if (likely(sbi->s_mount_flags & EXT4_MF_MNTDIR_SAMPLED))
780 return 0;
781
782 if (sb_rdonly(sb) || !sb_start_intwrite_trylock(sb))
783 return 0;
784
785 sbi->s_mount_flags |= EXT4_MF_MNTDIR_SAMPLED;
786 /*
787 * Sample where the filesystem has been mounted and
788 * store it in the superblock for sysadmin convenience
789 * when trying to sort through large numbers of block
790 * devices or filesystem images.
791 */
792 memset(buf, 0, sizeof(buf));
793 path.mnt = mnt;
794 path.dentry = mnt->mnt_root;
795 cp = d_path(&path, buf, sizeof(buf));
796 err = 0;
797 if (IS_ERR(cp))
798 goto out;
799
800 handle = ext4_journal_start_sb(sb, EXT4_HT_MISC, 1);
801 err = PTR_ERR(handle);
802 if (IS_ERR(handle))
803 goto out;
804 BUFFER_TRACE(sbi->s_sbh, "get_write_access");
805 err = ext4_journal_get_write_access(handle, sbi->s_sbh);
806 if (err)
807 goto out_journal;
808 strlcpy(sbi->s_es->s_last_mounted, cp,
809 sizeof(sbi->s_es->s_last_mounted));
810 ext4_handle_dirty_super(handle, sb);
811out_journal:
812 ext4_journal_stop(handle);
813out:
814 sb_end_intwrite(sb);
815 return err;
816}
817
818static int ext4_file_open(struct inode *inode, struct file *filp)
819{
820 int ret;
821
822 if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
823 return -EIO;
824
825 ret = ext4_sample_last_mounted(inode->i_sb, filp->f_path.mnt);
826 if (ret)
827 return ret;
828
829 ret = fscrypt_file_open(inode, filp);
830 if (ret)
831 return ret;
832
833 ret = fsverity_file_open(inode, filp);
834 if (ret)
835 return ret;
836
837 /*
838 * Set up the jbd2_inode if we are opening the inode for
839 * writing and the journal is present
840 */
841 if (filp->f_mode & FMODE_WRITE) {
842 ret = ext4_inode_attach_jinode(inode);
843 if (ret < 0)
844 return ret;
845 }
846
847 filp->f_mode |= FMODE_NOWAIT;
848 return dquot_file_open(inode, filp);
849}
850
851/*
852 * ext4_llseek() handles both block-mapped and extent-mapped maxbytes values
853 * by calling generic_file_llseek_size() with the appropriate maxbytes
854 * value for each.
855 */
856loff_t ext4_llseek(struct file *file, loff_t offset, int whence)
857{
858 struct inode *inode = file->f_mapping->host;
859 loff_t maxbytes;
860
861 if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
862 maxbytes = EXT4_SB(inode->i_sb)->s_bitmap_maxbytes;
863 else
864 maxbytes = inode->i_sb->s_maxbytes;
865
866 switch (whence) {
867 default:
868 return generic_file_llseek_size(file, offset, whence,
869 maxbytes, i_size_read(inode));
870 case SEEK_HOLE:
871 inode_lock_shared(inode);
872 offset = iomap_seek_hole(inode, offset,
873 &ext4_iomap_report_ops);
874 inode_unlock_shared(inode);
875 break;
876 case SEEK_DATA:
877 inode_lock_shared(inode);
878 offset = iomap_seek_data(inode, offset,
879 &ext4_iomap_report_ops);
880 inode_unlock_shared(inode);
881 break;
882 }
883
884 if (offset < 0)
885 return offset;
886 return vfs_setpos(file, offset, maxbytes);
887}
888
889const struct file_operations ext4_file_operations = {
890 .llseek = ext4_llseek,
891 .read_iter = ext4_file_read_iter,
892 .write_iter = ext4_file_write_iter,
893 .iopoll = iomap_dio_iopoll,
894 .unlocked_ioctl = ext4_ioctl,
895#ifdef CONFIG_COMPAT
896 .compat_ioctl = ext4_compat_ioctl,
897#endif
898 .mmap = ext4_file_mmap,
899 .mmap_supported_flags = MAP_SYNC,
900 .open = ext4_file_open,
901 .release = ext4_release_file,
902 .fsync = ext4_sync_file,
903 .get_unmapped_area = thp_get_unmapped_area,
904 .splice_read = generic_file_splice_read,
905 .splice_write = iter_file_splice_write,
906 .fallocate = ext4_fallocate,
907};
908
909const struct inode_operations ext4_file_inode_operations = {
910 .setattr = ext4_setattr,
911 .getattr = ext4_file_getattr,
912 .listxattr = ext4_listxattr,
913 .get_acl = ext4_get_acl,
914 .set_acl = ext4_set_acl,
915 .fiemap = ext4_fiemap,
916};
917
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * linux/fs/ext4/file.c
4 *
5 * Copyright (C) 1992, 1993, 1994, 1995
6 * Remy Card (card@masi.ibp.fr)
7 * Laboratoire MASI - Institut Blaise Pascal
8 * Universite Pierre et Marie Curie (Paris VI)
9 *
10 * from
11 *
12 * linux/fs/minix/file.c
13 *
14 * Copyright (C) 1991, 1992 Linus Torvalds
15 *
16 * ext4 fs regular file handling primitives
17 *
18 * 64-bit file support on 64-bit platforms by Jakub Jelinek
19 * (jj@sunsite.ms.mff.cuni.cz)
20 */
21
22#include <linux/time.h>
23#include <linux/fs.h>
24#include <linux/iomap.h>
25#include <linux/mount.h>
26#include <linux/path.h>
27#include <linux/dax.h>
28#include <linux/quotaops.h>
29#include <linux/pagevec.h>
30#include <linux/uio.h>
31#include <linux/mman.h>
32#include <linux/backing-dev.h>
33#include "ext4.h"
34#include "ext4_jbd2.h"
35#include "xattr.h"
36#include "acl.h"
37#include "truncate.h"
38
39static bool ext4_dio_supported(struct inode *inode)
40{
41 if (IS_ENABLED(CONFIG_FS_ENCRYPTION) && IS_ENCRYPTED(inode))
42 return false;
43 if (fsverity_active(inode))
44 return false;
45 if (ext4_should_journal_data(inode))
46 return false;
47 if (ext4_has_inline_data(inode))
48 return false;
49 return true;
50}
51
52static ssize_t ext4_dio_read_iter(struct kiocb *iocb, struct iov_iter *to)
53{
54 ssize_t ret;
55 struct inode *inode = file_inode(iocb->ki_filp);
56
57 if (iocb->ki_flags & IOCB_NOWAIT) {
58 if (!inode_trylock_shared(inode))
59 return -EAGAIN;
60 } else {
61 inode_lock_shared(inode);
62 }
63
64 if (!ext4_dio_supported(inode)) {
65 inode_unlock_shared(inode);
66 /*
67 * Fallback to buffered I/O if the operation being performed on
68 * the inode is not supported by direct I/O. The IOCB_DIRECT
69 * flag needs to be cleared here in order to ensure that the
70 * direct I/O path within generic_file_read_iter() is not
71 * taken.
72 */
73 iocb->ki_flags &= ~IOCB_DIRECT;
74 return generic_file_read_iter(iocb, to);
75 }
76
77 ret = iomap_dio_rw(iocb, to, &ext4_iomap_ops, NULL, 0);
78 inode_unlock_shared(inode);
79
80 file_accessed(iocb->ki_filp);
81 return ret;
82}
83
84#ifdef CONFIG_FS_DAX
85static ssize_t ext4_dax_read_iter(struct kiocb *iocb, struct iov_iter *to)
86{
87 struct inode *inode = file_inode(iocb->ki_filp);
88 ssize_t ret;
89
90 if (iocb->ki_flags & IOCB_NOWAIT) {
91 if (!inode_trylock_shared(inode))
92 return -EAGAIN;
93 } else {
94 inode_lock_shared(inode);
95 }
96 /*
97 * Recheck under inode lock - at this point we are sure it cannot
98 * change anymore
99 */
100 if (!IS_DAX(inode)) {
101 inode_unlock_shared(inode);
102 /* Fallback to buffered IO in case we cannot support DAX */
103 return generic_file_read_iter(iocb, to);
104 }
105 ret = dax_iomap_rw(iocb, to, &ext4_iomap_ops);
106 inode_unlock_shared(inode);
107
108 file_accessed(iocb->ki_filp);
109 return ret;
110}
111#endif
112
113static ssize_t ext4_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
114{
115 struct inode *inode = file_inode(iocb->ki_filp);
116
117 if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
118 return -EIO;
119
120 if (!iov_iter_count(to))
121 return 0; /* skip atime */
122
123#ifdef CONFIG_FS_DAX
124 if (IS_DAX(inode))
125 return ext4_dax_read_iter(iocb, to);
126#endif
127 if (iocb->ki_flags & IOCB_DIRECT)
128 return ext4_dio_read_iter(iocb, to);
129
130 return generic_file_read_iter(iocb, to);
131}
132
133/*
134 * Called when an inode is released. Note that this is different
135 * from ext4_file_open: open gets called at every open, but release
136 * gets called only when /all/ the files are closed.
137 */
138static int ext4_release_file(struct inode *inode, struct file *filp)
139{
140 if (ext4_test_inode_state(inode, EXT4_STATE_DA_ALLOC_CLOSE)) {
141 ext4_alloc_da_blocks(inode);
142 ext4_clear_inode_state(inode, EXT4_STATE_DA_ALLOC_CLOSE);
143 }
144 /* if we are the last writer on the inode, drop the block reservation */
145 if ((filp->f_mode & FMODE_WRITE) &&
146 (atomic_read(&inode->i_writecount) == 1) &&
147 !EXT4_I(inode)->i_reserved_data_blocks) {
148 down_write(&EXT4_I(inode)->i_data_sem);
149 ext4_discard_preallocations(inode, 0);
150 up_write(&EXT4_I(inode)->i_data_sem);
151 }
152 if (is_dx(inode) && filp->private_data)
153 ext4_htree_free_dir_info(filp->private_data);
154
155 return 0;
156}
157
158/*
159 * This tests whether the IO in question is block-aligned or not.
160 * Ext4 utilizes unwritten extents when hole-filling during direct IO, and they
161 * are converted to written only after the IO is complete. Until they are
162 * mapped, these blocks appear as holes, so dio_zero_block() will assume that
163 * it needs to zero out portions of the start and/or end block. If 2 AIO
164 * threads are at work on the same unwritten block, they must be synchronized
165 * or one thread will zero the other's data, causing corruption.
166 */
167static bool
168ext4_unaligned_io(struct inode *inode, struct iov_iter *from, loff_t pos)
169{
170 struct super_block *sb = inode->i_sb;
171 unsigned long blockmask = sb->s_blocksize - 1;
172
173 if ((pos | iov_iter_alignment(from)) & blockmask)
174 return true;
175
176 return false;
177}
178
179static bool
180ext4_extending_io(struct inode *inode, loff_t offset, size_t len)
181{
182 if (offset + len > i_size_read(inode) ||
183 offset + len > EXT4_I(inode)->i_disksize)
184 return true;
185 return false;
186}
187
188/* Is IO overwriting allocated and initialized blocks? */
189static bool ext4_overwrite_io(struct inode *inode, loff_t pos, loff_t len)
190{
191 struct ext4_map_blocks map;
192 unsigned int blkbits = inode->i_blkbits;
193 int err, blklen;
194
195 if (pos + len > i_size_read(inode))
196 return false;
197
198 map.m_lblk = pos >> blkbits;
199 map.m_len = EXT4_MAX_BLOCKS(len, pos, blkbits);
200 blklen = map.m_len;
201
202 err = ext4_map_blocks(NULL, inode, &map, 0);
203 /*
204 * 'err==len' means that all of the blocks have been preallocated,
205 * regardless of whether they have been initialized or not. To exclude
206 * unwritten extents, we need to check m_flags.
207 */
208 return err == blklen && (map.m_flags & EXT4_MAP_MAPPED);
209}
210
211static ssize_t ext4_generic_write_checks(struct kiocb *iocb,
212 struct iov_iter *from)
213{
214 struct inode *inode = file_inode(iocb->ki_filp);
215 ssize_t ret;
216
217 if (unlikely(IS_IMMUTABLE(inode)))
218 return -EPERM;
219
220 ret = generic_write_checks(iocb, from);
221 if (ret <= 0)
222 return ret;
223
224 /*
225 * If we have encountered a bitmap-format file, the size limit
226 * is smaller than s_maxbytes, which is for extent-mapped files.
227 */
228 if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
229 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
230
231 if (iocb->ki_pos >= sbi->s_bitmap_maxbytes)
232 return -EFBIG;
233 iov_iter_truncate(from, sbi->s_bitmap_maxbytes - iocb->ki_pos);
234 }
235
236 return iov_iter_count(from);
237}
238
239static ssize_t ext4_write_checks(struct kiocb *iocb, struct iov_iter *from)
240{
241 ssize_t ret, count;
242
243 count = ext4_generic_write_checks(iocb, from);
244 if (count <= 0)
245 return count;
246
247 ret = file_modified(iocb->ki_filp);
248 if (ret)
249 return ret;
250 return count;
251}
252
253static ssize_t ext4_buffered_write_iter(struct kiocb *iocb,
254 struct iov_iter *from)
255{
256 ssize_t ret;
257 struct inode *inode = file_inode(iocb->ki_filp);
258
259 if (iocb->ki_flags & IOCB_NOWAIT)
260 return -EOPNOTSUPP;
261
262 ext4_fc_start_update(inode);
263 inode_lock(inode);
264 ret = ext4_write_checks(iocb, from);
265 if (ret <= 0)
266 goto out;
267
268 current->backing_dev_info = inode_to_bdi(inode);
269 ret = generic_perform_write(iocb->ki_filp, from, iocb->ki_pos);
270 current->backing_dev_info = NULL;
271
272out:
273 inode_unlock(inode);
274 ext4_fc_stop_update(inode);
275 if (likely(ret > 0)) {
276 iocb->ki_pos += ret;
277 ret = generic_write_sync(iocb, ret);
278 }
279
280 return ret;
281}
282
283static ssize_t ext4_handle_inode_extension(struct inode *inode, loff_t offset,
284 ssize_t written, size_t count)
285{
286 handle_t *handle;
287 bool truncate = false;
288 u8 blkbits = inode->i_blkbits;
289 ext4_lblk_t written_blk, end_blk;
290 int ret;
291
292 /*
293 * Note that EXT4_I(inode)->i_disksize can get extended up to
294 * inode->i_size while the I/O was running due to writeback of delalloc
295 * blocks. But, the code in ext4_iomap_alloc() is careful to use
296 * zeroed/unwritten extents if this is possible; thus we won't leave
297 * uninitialized blocks in a file even if we didn't succeed in writing
298 * as much as we intended.
299 */
300 WARN_ON_ONCE(i_size_read(inode) < EXT4_I(inode)->i_disksize);
301 if (offset + count <= EXT4_I(inode)->i_disksize) {
302 /*
303 * We need to ensure that the inode is removed from the orphan
304 * list if it has been added prematurely, due to writeback of
305 * delalloc blocks.
306 */
307 if (!list_empty(&EXT4_I(inode)->i_orphan) && inode->i_nlink) {
308 handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
309
310 if (IS_ERR(handle)) {
311 ext4_orphan_del(NULL, inode);
312 return PTR_ERR(handle);
313 }
314
315 ext4_orphan_del(handle, inode);
316 ext4_journal_stop(handle);
317 }
318
319 return written;
320 }
321
322 if (written < 0)
323 goto truncate;
324
325 handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
326 if (IS_ERR(handle)) {
327 written = PTR_ERR(handle);
328 goto truncate;
329 }
330
331 if (ext4_update_inode_size(inode, offset + written)) {
332 ret = ext4_mark_inode_dirty(handle, inode);
333 if (unlikely(ret)) {
334 written = ret;
335 ext4_journal_stop(handle);
336 goto truncate;
337 }
338 }
339
340 /*
341 * We may need to truncate allocated but not written blocks beyond EOF.
342 */
343 written_blk = ALIGN(offset + written, 1 << blkbits);
344 end_blk = ALIGN(offset + count, 1 << blkbits);
345 if (written_blk < end_blk && ext4_can_truncate(inode))
346 truncate = true;
347
348 /*
349 * Remove the inode from the orphan list if it has been extended and
350 * everything went OK.
351 */
352 if (!truncate && inode->i_nlink)
353 ext4_orphan_del(handle, inode);
354 ext4_journal_stop(handle);
355
356 if (truncate) {
357truncate:
358 ext4_truncate_failed_write(inode);
359 /*
360 * If the truncate operation failed early, then the inode may
361 * still be on the orphan list. In that case, we need to try
362 * remove the inode from the in-memory linked list.
363 */
364 if (inode->i_nlink)
365 ext4_orphan_del(NULL, inode);
366 }
367
368 return written;
369}
370
371static int ext4_dio_write_end_io(struct kiocb *iocb, ssize_t size,
372 int error, unsigned int flags)
373{
374 loff_t pos = iocb->ki_pos;
375 struct inode *inode = file_inode(iocb->ki_filp);
376
377 if (error)
378 return error;
379
380 if (size && flags & IOMAP_DIO_UNWRITTEN) {
381 error = ext4_convert_unwritten_extents(NULL, inode, pos, size);
382 if (error < 0)
383 return error;
384 }
385 /*
386 * If we are extending the file, we have to update i_size here before
387 * page cache gets invalidated in iomap_dio_rw(). Otherwise racing
388 * buffered reads could zero out too much from page cache pages. Update
389 * of on-disk size will happen later in ext4_dio_write_iter() where
390 * we have enough information to also perform orphan list handling etc.
391 * Note that we perform all extending writes synchronously under
392 * i_rwsem held exclusively so i_size update is safe here in that case.
393 * If the write was not extending, we cannot see pos > i_size here
394 * because operations reducing i_size like truncate wait for all
395 * outstanding DIO before updating i_size.
396 */
397 pos += size;
398 if (pos > i_size_read(inode))
399 i_size_write(inode, pos);
400
401 return 0;
402}
403
404static const struct iomap_dio_ops ext4_dio_write_ops = {
405 .end_io = ext4_dio_write_end_io,
406};
407
408/*
409 * The intention here is to start with shared lock acquired then see if any
410 * condition requires an exclusive inode lock. If yes, then we restart the
411 * whole operation by releasing the shared lock and acquiring exclusive lock.
412 *
413 * - For unaligned_io we never take shared lock as it may cause data corruption
414 * when two unaligned IO tries to modify the same block e.g. while zeroing.
415 *
416 * - For extending writes case we don't take the shared lock, since it requires
417 * updating inode i_disksize and/or orphan handling with exclusive lock.
418 *
419 * - shared locking will only be true mostly with overwrites. Otherwise we will
420 * switch to exclusive i_rwsem lock.
421 */
422static ssize_t ext4_dio_write_checks(struct kiocb *iocb, struct iov_iter *from,
423 bool *ilock_shared, bool *extend)
424{
425 struct file *file = iocb->ki_filp;
426 struct inode *inode = file_inode(file);
427 loff_t offset;
428 size_t count;
429 ssize_t ret;
430
431restart:
432 ret = ext4_generic_write_checks(iocb, from);
433 if (ret <= 0)
434 goto out;
435
436 offset = iocb->ki_pos;
437 count = ret;
438 if (ext4_extending_io(inode, offset, count))
439 *extend = true;
440 /*
441 * Determine whether the IO operation will overwrite allocated
442 * and initialized blocks.
443 * We need exclusive i_rwsem for changing security info
444 * in file_modified().
445 */
446 if (*ilock_shared && (!IS_NOSEC(inode) || *extend ||
447 !ext4_overwrite_io(inode, offset, count))) {
448 if (iocb->ki_flags & IOCB_NOWAIT) {
449 ret = -EAGAIN;
450 goto out;
451 }
452 inode_unlock_shared(inode);
453 *ilock_shared = false;
454 inode_lock(inode);
455 goto restart;
456 }
457
458 ret = file_modified(file);
459 if (ret < 0)
460 goto out;
461
462 return count;
463out:
464 if (*ilock_shared)
465 inode_unlock_shared(inode);
466 else
467 inode_unlock(inode);
468 return ret;
469}
470
471static ssize_t ext4_dio_write_iter(struct kiocb *iocb, struct iov_iter *from)
472{
473 ssize_t ret;
474 handle_t *handle;
475 struct inode *inode = file_inode(iocb->ki_filp);
476 loff_t offset = iocb->ki_pos;
477 size_t count = iov_iter_count(from);
478 const struct iomap_ops *iomap_ops = &ext4_iomap_ops;
479 bool extend = false, unaligned_io = false;
480 bool ilock_shared = true;
481
482 /*
483 * We initially start with shared inode lock unless it is
484 * unaligned IO which needs exclusive lock anyways.
485 */
486 if (ext4_unaligned_io(inode, from, offset)) {
487 unaligned_io = true;
488 ilock_shared = false;
489 }
490 /*
491 * Quick check here without any i_rwsem lock to see if it is extending
492 * IO. A more reliable check is done in ext4_dio_write_checks() with
493 * proper locking in place.
494 */
495 if (offset + count > i_size_read(inode))
496 ilock_shared = false;
497
498 if (iocb->ki_flags & IOCB_NOWAIT) {
499 if (ilock_shared) {
500 if (!inode_trylock_shared(inode))
501 return -EAGAIN;
502 } else {
503 if (!inode_trylock(inode))
504 return -EAGAIN;
505 }
506 } else {
507 if (ilock_shared)
508 inode_lock_shared(inode);
509 else
510 inode_lock(inode);
511 }
512
513 /* Fallback to buffered I/O if the inode does not support direct I/O. */
514 if (!ext4_dio_supported(inode)) {
515 if (ilock_shared)
516 inode_unlock_shared(inode);
517 else
518 inode_unlock(inode);
519 return ext4_buffered_write_iter(iocb, from);
520 }
521
522 ret = ext4_dio_write_checks(iocb, from, &ilock_shared, &extend);
523 if (ret <= 0)
524 return ret;
525
526 /* if we're going to block and IOCB_NOWAIT is set, return -EAGAIN */
527 if ((iocb->ki_flags & IOCB_NOWAIT) && (unaligned_io || extend)) {
528 ret = -EAGAIN;
529 goto out;
530 }
531
532 offset = iocb->ki_pos;
533 count = ret;
534
535 /*
536 * Unaligned direct IO must be serialized among each other as zeroing
537 * of partial blocks of two competing unaligned IOs can result in data
538 * corruption.
539 *
540 * So we make sure we don't allow any unaligned IO in flight.
541 * For IOs where we need not wait (like unaligned non-AIO DIO),
542 * below inode_dio_wait() may anyway become a no-op, since we start
543 * with exclusive lock.
544 */
545 if (unaligned_io)
546 inode_dio_wait(inode);
547
548 if (extend) {
549 handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
550 if (IS_ERR(handle)) {
551 ret = PTR_ERR(handle);
552 goto out;
553 }
554
555 ext4_fc_start_update(inode);
556 ret = ext4_orphan_add(handle, inode);
557 ext4_fc_stop_update(inode);
558 if (ret) {
559 ext4_journal_stop(handle);
560 goto out;
561 }
562
563 ext4_journal_stop(handle);
564 }
565
566 if (ilock_shared)
567 iomap_ops = &ext4_iomap_overwrite_ops;
568 ret = iomap_dio_rw(iocb, from, iomap_ops, &ext4_dio_write_ops,
569 (unaligned_io || extend) ? IOMAP_DIO_FORCE_WAIT : 0);
570 if (ret == -ENOTBLK)
571 ret = 0;
572
573 if (extend)
574 ret = ext4_handle_inode_extension(inode, offset, ret, count);
575
576out:
577 if (ilock_shared)
578 inode_unlock_shared(inode);
579 else
580 inode_unlock(inode);
581
582 if (ret >= 0 && iov_iter_count(from)) {
583 ssize_t err;
584 loff_t endbyte;
585
586 offset = iocb->ki_pos;
587 err = ext4_buffered_write_iter(iocb, from);
588 if (err < 0)
589 return err;
590
591 /*
592 * We need to ensure that the pages within the page cache for
593 * the range covered by this I/O are written to disk and
594 * invalidated. This is in attempt to preserve the expected
595 * direct I/O semantics in the case we fallback to buffered I/O
596 * to complete off the I/O request.
597 */
598 ret += err;
599 endbyte = offset + err - 1;
600 err = filemap_write_and_wait_range(iocb->ki_filp->f_mapping,
601 offset, endbyte);
602 if (!err)
603 invalidate_mapping_pages(iocb->ki_filp->f_mapping,
604 offset >> PAGE_SHIFT,
605 endbyte >> PAGE_SHIFT);
606 }
607
608 return ret;
609}
610
611#ifdef CONFIG_FS_DAX
612static ssize_t
613ext4_dax_write_iter(struct kiocb *iocb, struct iov_iter *from)
614{
615 ssize_t ret;
616 size_t count;
617 loff_t offset;
618 handle_t *handle;
619 bool extend = false;
620 struct inode *inode = file_inode(iocb->ki_filp);
621
622 if (iocb->ki_flags & IOCB_NOWAIT) {
623 if (!inode_trylock(inode))
624 return -EAGAIN;
625 } else {
626 inode_lock(inode);
627 }
628
629 ret = ext4_write_checks(iocb, from);
630 if (ret <= 0)
631 goto out;
632
633 offset = iocb->ki_pos;
634 count = iov_iter_count(from);
635
636 if (offset + count > EXT4_I(inode)->i_disksize) {
637 handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
638 if (IS_ERR(handle)) {
639 ret = PTR_ERR(handle);
640 goto out;
641 }
642
643 ret = ext4_orphan_add(handle, inode);
644 if (ret) {
645 ext4_journal_stop(handle);
646 goto out;
647 }
648
649 extend = true;
650 ext4_journal_stop(handle);
651 }
652
653 ret = dax_iomap_rw(iocb, from, &ext4_iomap_ops);
654
655 if (extend)
656 ret = ext4_handle_inode_extension(inode, offset, ret, count);
657out:
658 inode_unlock(inode);
659 if (ret > 0)
660 ret = generic_write_sync(iocb, ret);
661 return ret;
662}
663#endif
664
665static ssize_t
666ext4_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
667{
668 struct inode *inode = file_inode(iocb->ki_filp);
669
670 if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
671 return -EIO;
672
673#ifdef CONFIG_FS_DAX
674 if (IS_DAX(inode))
675 return ext4_dax_write_iter(iocb, from);
676#endif
677 if (iocb->ki_flags & IOCB_DIRECT)
678 return ext4_dio_write_iter(iocb, from);
679 else
680 return ext4_buffered_write_iter(iocb, from);
681}
682
683#ifdef CONFIG_FS_DAX
684static vm_fault_t ext4_dax_huge_fault(struct vm_fault *vmf,
685 enum page_entry_size pe_size)
686{
687 int error = 0;
688 vm_fault_t result;
689 int retries = 0;
690 handle_t *handle = NULL;
691 struct inode *inode = file_inode(vmf->vma->vm_file);
692 struct super_block *sb = inode->i_sb;
693
694 /*
695 * We have to distinguish real writes from writes which will result in a
696 * COW page; COW writes should *not* poke the journal (the file will not
697 * be changed). Doing so would cause unintended failures when mounted
698 * read-only.
699 *
700 * We check for VM_SHARED rather than vmf->cow_page since the latter is
701 * unset for pe_size != PE_SIZE_PTE (i.e. only in do_cow_fault); for
702 * other sizes, dax_iomap_fault will handle splitting / fallback so that
703 * we eventually come back with a COW page.
704 */
705 bool write = (vmf->flags & FAULT_FLAG_WRITE) &&
706 (vmf->vma->vm_flags & VM_SHARED);
707 pfn_t pfn;
708
709 if (write) {
710 sb_start_pagefault(sb);
711 file_update_time(vmf->vma->vm_file);
712 down_read(&EXT4_I(inode)->i_mmap_sem);
713retry:
714 handle = ext4_journal_start_sb(sb, EXT4_HT_WRITE_PAGE,
715 EXT4_DATA_TRANS_BLOCKS(sb));
716 if (IS_ERR(handle)) {
717 up_read(&EXT4_I(inode)->i_mmap_sem);
718 sb_end_pagefault(sb);
719 return VM_FAULT_SIGBUS;
720 }
721 } else {
722 down_read(&EXT4_I(inode)->i_mmap_sem);
723 }
724 result = dax_iomap_fault(vmf, pe_size, &pfn, &error, &ext4_iomap_ops);
725 if (write) {
726 ext4_journal_stop(handle);
727
728 if ((result & VM_FAULT_ERROR) && error == -ENOSPC &&
729 ext4_should_retry_alloc(sb, &retries))
730 goto retry;
731 /* Handling synchronous page fault? */
732 if (result & VM_FAULT_NEEDDSYNC)
733 result = dax_finish_sync_fault(vmf, pe_size, pfn);
734 up_read(&EXT4_I(inode)->i_mmap_sem);
735 sb_end_pagefault(sb);
736 } else {
737 up_read(&EXT4_I(inode)->i_mmap_sem);
738 }
739
740 return result;
741}
742
743static vm_fault_t ext4_dax_fault(struct vm_fault *vmf)
744{
745 return ext4_dax_huge_fault(vmf, PE_SIZE_PTE);
746}
747
748static const struct vm_operations_struct ext4_dax_vm_ops = {
749 .fault = ext4_dax_fault,
750 .huge_fault = ext4_dax_huge_fault,
751 .page_mkwrite = ext4_dax_fault,
752 .pfn_mkwrite = ext4_dax_fault,
753};
754#else
755#define ext4_dax_vm_ops ext4_file_vm_ops
756#endif
757
758static const struct vm_operations_struct ext4_file_vm_ops = {
759 .fault = ext4_filemap_fault,
760 .map_pages = filemap_map_pages,
761 .page_mkwrite = ext4_page_mkwrite,
762};
763
764static int ext4_file_mmap(struct file *file, struct vm_area_struct *vma)
765{
766 struct inode *inode = file->f_mapping->host;
767 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
768 struct dax_device *dax_dev = sbi->s_daxdev;
769
770 if (unlikely(ext4_forced_shutdown(sbi)))
771 return -EIO;
772
773 /*
774 * We don't support synchronous mappings for non-DAX files and
775 * for DAX files if underneath dax_device is not synchronous.
776 */
777 if (!daxdev_mapping_supported(vma, dax_dev))
778 return -EOPNOTSUPP;
779
780 file_accessed(file);
781 if (IS_DAX(file_inode(file))) {
782 vma->vm_ops = &ext4_dax_vm_ops;
783 vma->vm_flags |= VM_HUGEPAGE;
784 } else {
785 vma->vm_ops = &ext4_file_vm_ops;
786 }
787 return 0;
788}
789
790static int ext4_sample_last_mounted(struct super_block *sb,
791 struct vfsmount *mnt)
792{
793 struct ext4_sb_info *sbi = EXT4_SB(sb);
794 struct path path;
795 char buf[64], *cp;
796 handle_t *handle;
797 int err;
798
799 if (likely(ext4_test_mount_flag(sb, EXT4_MF_MNTDIR_SAMPLED)))
800 return 0;
801
802 if (sb_rdonly(sb) || !sb_start_intwrite_trylock(sb))
803 return 0;
804
805 ext4_set_mount_flag(sb, EXT4_MF_MNTDIR_SAMPLED);
806 /*
807 * Sample where the filesystem has been mounted and
808 * store it in the superblock for sysadmin convenience
809 * when trying to sort through large numbers of block
810 * devices or filesystem images.
811 */
812 memset(buf, 0, sizeof(buf));
813 path.mnt = mnt;
814 path.dentry = mnt->mnt_root;
815 cp = d_path(&path, buf, sizeof(buf));
816 err = 0;
817 if (IS_ERR(cp))
818 goto out;
819
820 handle = ext4_journal_start_sb(sb, EXT4_HT_MISC, 1);
821 err = PTR_ERR(handle);
822 if (IS_ERR(handle))
823 goto out;
824 BUFFER_TRACE(sbi->s_sbh, "get_write_access");
825 err = ext4_journal_get_write_access(handle, sbi->s_sbh);
826 if (err)
827 goto out_journal;
828 lock_buffer(sbi->s_sbh);
829 strncpy(sbi->s_es->s_last_mounted, cp,
830 sizeof(sbi->s_es->s_last_mounted));
831 ext4_superblock_csum_set(sb);
832 unlock_buffer(sbi->s_sbh);
833 ext4_handle_dirty_metadata(handle, NULL, sbi->s_sbh);
834out_journal:
835 ext4_journal_stop(handle);
836out:
837 sb_end_intwrite(sb);
838 return err;
839}
840
841static int ext4_file_open(struct inode *inode, struct file *filp)
842{
843 int ret;
844
845 if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
846 return -EIO;
847
848 ret = ext4_sample_last_mounted(inode->i_sb, filp->f_path.mnt);
849 if (ret)
850 return ret;
851
852 ret = fscrypt_file_open(inode, filp);
853 if (ret)
854 return ret;
855
856 ret = fsverity_file_open(inode, filp);
857 if (ret)
858 return ret;
859
860 /*
861 * Set up the jbd2_inode if we are opening the inode for
862 * writing and the journal is present
863 */
864 if (filp->f_mode & FMODE_WRITE) {
865 ret = ext4_inode_attach_jinode(inode);
866 if (ret < 0)
867 return ret;
868 }
869
870 filp->f_mode |= FMODE_NOWAIT | FMODE_BUF_RASYNC;
871 return dquot_file_open(inode, filp);
872}
873
874/*
875 * ext4_llseek() handles both block-mapped and extent-mapped maxbytes values
876 * by calling generic_file_llseek_size() with the appropriate maxbytes
877 * value for each.
878 */
879loff_t ext4_llseek(struct file *file, loff_t offset, int whence)
880{
881 struct inode *inode = file->f_mapping->host;
882 loff_t maxbytes;
883
884 if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
885 maxbytes = EXT4_SB(inode->i_sb)->s_bitmap_maxbytes;
886 else
887 maxbytes = inode->i_sb->s_maxbytes;
888
889 switch (whence) {
890 default:
891 return generic_file_llseek_size(file, offset, whence,
892 maxbytes, i_size_read(inode));
893 case SEEK_HOLE:
894 inode_lock_shared(inode);
895 offset = iomap_seek_hole(inode, offset,
896 &ext4_iomap_report_ops);
897 inode_unlock_shared(inode);
898 break;
899 case SEEK_DATA:
900 inode_lock_shared(inode);
901 offset = iomap_seek_data(inode, offset,
902 &ext4_iomap_report_ops);
903 inode_unlock_shared(inode);
904 break;
905 }
906
907 if (offset < 0)
908 return offset;
909 return vfs_setpos(file, offset, maxbytes);
910}
911
912const struct file_operations ext4_file_operations = {
913 .llseek = ext4_llseek,
914 .read_iter = ext4_file_read_iter,
915 .write_iter = ext4_file_write_iter,
916 .iopoll = iomap_dio_iopoll,
917 .unlocked_ioctl = ext4_ioctl,
918#ifdef CONFIG_COMPAT
919 .compat_ioctl = ext4_compat_ioctl,
920#endif
921 .mmap = ext4_file_mmap,
922 .mmap_supported_flags = MAP_SYNC,
923 .open = ext4_file_open,
924 .release = ext4_release_file,
925 .fsync = ext4_sync_file,
926 .get_unmapped_area = thp_get_unmapped_area,
927 .splice_read = generic_file_splice_read,
928 .splice_write = iter_file_splice_write,
929 .fallocate = ext4_fallocate,
930};
931
932const struct inode_operations ext4_file_inode_operations = {
933 .setattr = ext4_setattr,
934 .getattr = ext4_file_getattr,
935 .listxattr = ext4_listxattr,
936 .get_acl = ext4_get_acl,
937 .set_acl = ext4_set_acl,
938 .fiemap = ext4_fiemap,
939 .fileattr_get = ext4_fileattr_get,
940 .fileattr_set = ext4_fileattr_set,
941};
942