18f5fd2a163b056eff89f038bdfa6ddfd495dc35
[releases.git] / file.c
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
39 /*
40  * Returns %true if the given DIO request should be attempted with DIO, or
41  * %false if it should fall back to buffered I/O.
42  *
43  * DIO isn't well specified; when it's unsupported (either due to the request
44  * being misaligned, or due to the file not supporting DIO at all), filesystems
45  * either fall back to buffered I/O or return EINVAL.  For files that don't use
46  * any special features like encryption or verity, ext4 has traditionally
47  * returned EINVAL for misaligned DIO.  iomap_dio_rw() uses this convention too.
48  * In this case, we should attempt the DIO, *not* fall back to buffered I/O.
49  *
50  * In contrast, in cases where DIO is unsupported due to ext4 features, ext4
51  * traditionally falls back to buffered I/O.
52  *
53  * This function implements the traditional ext4 behavior in all these cases.
54  */
55 static bool ext4_should_use_dio(struct kiocb *iocb, struct iov_iter *iter)
56 {
57         struct inode *inode = file_inode(iocb->ki_filp);
58         u32 dio_align = ext4_dio_alignment(inode);
59
60         if (dio_align == 0)
61                 return false;
62
63         if (dio_align == 1)
64                 return true;
65
66         return IS_ALIGNED(iocb->ki_pos | iov_iter_alignment(iter), dio_align);
67 }
68
69 static ssize_t ext4_dio_read_iter(struct kiocb *iocb, struct iov_iter *to)
70 {
71         ssize_t ret;
72         struct inode *inode = file_inode(iocb->ki_filp);
73
74         if (iocb->ki_flags & IOCB_NOWAIT) {
75                 if (!inode_trylock_shared(inode))
76                         return -EAGAIN;
77         } else {
78                 inode_lock_shared(inode);
79         }
80
81         if (!ext4_should_use_dio(iocb, to)) {
82                 inode_unlock_shared(inode);
83                 /*
84                  * Fallback to buffered I/O if the operation being performed on
85                  * the inode is not supported by direct I/O. The IOCB_DIRECT
86                  * flag needs to be cleared here in order to ensure that the
87                  * direct I/O path within generic_file_read_iter() is not
88                  * taken.
89                  */
90                 iocb->ki_flags &= ~IOCB_DIRECT;
91                 return generic_file_read_iter(iocb, to);
92         }
93
94         ret = iomap_dio_rw(iocb, to, &ext4_iomap_ops, NULL, 0, NULL, 0);
95         inode_unlock_shared(inode);
96
97         file_accessed(iocb->ki_filp);
98         return ret;
99 }
100
101 #ifdef CONFIG_FS_DAX
102 static ssize_t ext4_dax_read_iter(struct kiocb *iocb, struct iov_iter *to)
103 {
104         struct inode *inode = file_inode(iocb->ki_filp);
105         ssize_t ret;
106
107         if (iocb->ki_flags & IOCB_NOWAIT) {
108                 if (!inode_trylock_shared(inode))
109                         return -EAGAIN;
110         } else {
111                 inode_lock_shared(inode);
112         }
113         /*
114          * Recheck under inode lock - at this point we are sure it cannot
115          * change anymore
116          */
117         if (!IS_DAX(inode)) {
118                 inode_unlock_shared(inode);
119                 /* Fallback to buffered IO in case we cannot support DAX */
120                 return generic_file_read_iter(iocb, to);
121         }
122         ret = dax_iomap_rw(iocb, to, &ext4_iomap_ops);
123         inode_unlock_shared(inode);
124
125         file_accessed(iocb->ki_filp);
126         return ret;
127 }
128 #endif
129
130 static ssize_t ext4_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
131 {
132         struct inode *inode = file_inode(iocb->ki_filp);
133
134         if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
135                 return -EIO;
136
137         if (!iov_iter_count(to))
138                 return 0; /* skip atime */
139
140 #ifdef CONFIG_FS_DAX
141         if (IS_DAX(inode))
142                 return ext4_dax_read_iter(iocb, to);
143 #endif
144         if (iocb->ki_flags & IOCB_DIRECT)
145                 return ext4_dio_read_iter(iocb, to);
146
147         return generic_file_read_iter(iocb, to);
148 }
149
150 /*
151  * Called when an inode is released. Note that this is different
152  * from ext4_file_open: open gets called at every open, but release
153  * gets called only when /all/ the files are closed.
154  */
155 static int ext4_release_file(struct inode *inode, struct file *filp)
156 {
157         if (ext4_test_inode_state(inode, EXT4_STATE_DA_ALLOC_CLOSE)) {
158                 ext4_alloc_da_blocks(inode);
159                 ext4_clear_inode_state(inode, EXT4_STATE_DA_ALLOC_CLOSE);
160         }
161         /* if we are the last writer on the inode, drop the block reservation */
162         if ((filp->f_mode & FMODE_WRITE) &&
163                         (atomic_read(&inode->i_writecount) == 1) &&
164                         !EXT4_I(inode)->i_reserved_data_blocks) {
165                 down_write(&EXT4_I(inode)->i_data_sem);
166                 ext4_discard_preallocations(inode, 0);
167                 up_write(&EXT4_I(inode)->i_data_sem);
168         }
169         if (is_dx(inode) && filp->private_data)
170                 ext4_htree_free_dir_info(filp->private_data);
171
172         return 0;
173 }
174
175 /*
176  * This tests whether the IO in question is block-aligned or not.
177  * Ext4 utilizes unwritten extents when hole-filling during direct IO, and they
178  * are converted to written only after the IO is complete.  Until they are
179  * mapped, these blocks appear as holes, so dio_zero_block() will assume that
180  * it needs to zero out portions of the start and/or end block.  If 2 AIO
181  * threads are at work on the same unwritten block, they must be synchronized
182  * or one thread will zero the other's data, causing corruption.
183  */
184 static bool
185 ext4_unaligned_io(struct inode *inode, struct iov_iter *from, loff_t pos)
186 {
187         struct super_block *sb = inode->i_sb;
188         unsigned long blockmask = sb->s_blocksize - 1;
189
190         if ((pos | iov_iter_alignment(from)) & blockmask)
191                 return true;
192
193         return false;
194 }
195
196 static bool
197 ext4_extending_io(struct inode *inode, loff_t offset, size_t len)
198 {
199         if (offset + len > i_size_read(inode) ||
200             offset + len > EXT4_I(inode)->i_disksize)
201                 return true;
202         return false;
203 }
204
205 /* Is IO overwriting allocated and initialized blocks? */
206 static bool ext4_overwrite_io(struct inode *inode, loff_t pos, loff_t len)
207 {
208         struct ext4_map_blocks map;
209         unsigned int blkbits = inode->i_blkbits;
210         int err, blklen;
211
212         if (pos + len > i_size_read(inode))
213                 return false;
214
215         map.m_lblk = pos >> blkbits;
216         map.m_len = EXT4_MAX_BLOCKS(len, pos, blkbits);
217         blklen = map.m_len;
218
219         err = ext4_map_blocks(NULL, inode, &map, 0);
220         /*
221          * 'err==len' means that all of the blocks have been preallocated,
222          * regardless of whether they have been initialized or not. To exclude
223          * unwritten extents, we need to check m_flags.
224          */
225         return err == blklen && (map.m_flags & EXT4_MAP_MAPPED);
226 }
227
228 static ssize_t ext4_generic_write_checks(struct kiocb *iocb,
229                                          struct iov_iter *from)
230 {
231         struct inode *inode = file_inode(iocb->ki_filp);
232         ssize_t ret;
233
234         if (unlikely(IS_IMMUTABLE(inode)))
235                 return -EPERM;
236
237         ret = generic_write_checks(iocb, from);
238         if (ret <= 0)
239                 return ret;
240
241         /*
242          * If we have encountered a bitmap-format file, the size limit
243          * is smaller than s_maxbytes, which is for extent-mapped files.
244          */
245         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS))) {
246                 struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
247
248                 if (iocb->ki_pos >= sbi->s_bitmap_maxbytes)
249                         return -EFBIG;
250                 iov_iter_truncate(from, sbi->s_bitmap_maxbytes - iocb->ki_pos);
251         }
252
253         return iov_iter_count(from);
254 }
255
256 static ssize_t ext4_write_checks(struct kiocb *iocb, struct iov_iter *from)
257 {
258         ssize_t ret, count;
259
260         count = ext4_generic_write_checks(iocb, from);
261         if (count <= 0)
262                 return count;
263
264         ret = file_modified(iocb->ki_filp);
265         if (ret)
266                 return ret;
267         return count;
268 }
269
270 static ssize_t ext4_buffered_write_iter(struct kiocb *iocb,
271                                         struct iov_iter *from)
272 {
273         ssize_t ret;
274         struct inode *inode = file_inode(iocb->ki_filp);
275
276         if (iocb->ki_flags & IOCB_NOWAIT)
277                 return -EOPNOTSUPP;
278
279         inode_lock(inode);
280         ret = ext4_write_checks(iocb, from);
281         if (ret <= 0)
282                 goto out;
283
284         current->backing_dev_info = inode_to_bdi(inode);
285         ret = generic_perform_write(iocb, from);
286         current->backing_dev_info = NULL;
287
288 out:
289         inode_unlock(inode);
290         if (likely(ret > 0)) {
291                 iocb->ki_pos += ret;
292                 ret = generic_write_sync(iocb, ret);
293         }
294
295         return ret;
296 }
297
298 static ssize_t ext4_handle_inode_extension(struct inode *inode, loff_t offset,
299                                            ssize_t count)
300 {
301         handle_t *handle;
302
303         lockdep_assert_held_write(&inode->i_rwsem);
304         handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
305         if (IS_ERR(handle))
306                 return PTR_ERR(handle);
307
308         if (ext4_update_inode_size(inode, offset + count)) {
309                 int ret = ext4_mark_inode_dirty(handle, inode);
310                 if (unlikely(ret)) {
311                         ext4_journal_stop(handle);
312                         return ret;
313                 }
314         }
315
316         if (inode->i_nlink)
317                 ext4_orphan_del(handle, inode);
318         ext4_journal_stop(handle);
319
320         return count;
321 }
322
323 /*
324  * Clean up the inode after DIO or DAX extending write has completed and the
325  * inode size has been updated using ext4_handle_inode_extension().
326  */
327 static void ext4_inode_extension_cleanup(struct inode *inode, ssize_t count)
328 {
329         lockdep_assert_held_write(&inode->i_rwsem);
330         if (count < 0) {
331                 ext4_truncate_failed_write(inode);
332                 /*
333                  * If the truncate operation failed early, then the inode may
334                  * still be on the orphan list. In that case, we need to try
335                  * remove the inode from the in-memory linked list.
336                  */
337                 if (inode->i_nlink)
338                         ext4_orphan_del(NULL, inode);
339                 return;
340         }
341         /*
342          * If i_disksize got extended either due to writeback of delalloc
343          * blocks or extending truncate while the DIO was running we could fail
344          * to cleanup the orphan list in ext4_handle_inode_extension(). Do it
345          * now.
346          */
347         if (!list_empty(&EXT4_I(inode)->i_orphan) && inode->i_nlink) {
348                 handle_t *handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
349
350                 if (IS_ERR(handle)) {
351                         /*
352                          * The write has successfully completed. Not much to
353                          * do with the error here so just cleanup the orphan
354                          * list and hope for the best.
355                          */
356                         ext4_orphan_del(NULL, inode);
357                         return;
358                 }
359                 ext4_orphan_del(handle, inode);
360                 ext4_journal_stop(handle);
361         }
362 }
363
364 static int ext4_dio_write_end_io(struct kiocb *iocb, ssize_t size,
365                                  int error, unsigned int flags)
366 {
367         loff_t pos = iocb->ki_pos;
368         struct inode *inode = file_inode(iocb->ki_filp);
369
370         if (!error && size && flags & IOMAP_DIO_UNWRITTEN)
371                 error = ext4_convert_unwritten_extents(NULL, inode, pos, size);
372         if (error)
373                 return error;
374         /*
375          * Note that EXT4_I(inode)->i_disksize can get extended up to
376          * inode->i_size while the I/O was running due to writeback of delalloc
377          * blocks. But the code in ext4_iomap_alloc() is careful to use
378          * zeroed/unwritten extents if this is possible; thus we won't leave
379          * uninitialized blocks in a file even if we didn't succeed in writing
380          * as much as we intended. Also we can race with truncate or write
381          * expanding the file so we have to be a bit careful here.
382          */
383         if (pos + size <= READ_ONCE(EXT4_I(inode)->i_disksize) &&
384             pos + size <= i_size_read(inode))
385                 return size;
386         return ext4_handle_inode_extension(inode, pos, size);
387 }
388
389 static const struct iomap_dio_ops ext4_dio_write_ops = {
390         .end_io = ext4_dio_write_end_io,
391 };
392
393 /*
394  * The intention here is to start with shared lock acquired then see if any
395  * condition requires an exclusive inode lock. If yes, then we restart the
396  * whole operation by releasing the shared lock and acquiring exclusive lock.
397  *
398  * - For unaligned_io we never take shared lock as it may cause data corruption
399  *   when two unaligned IO tries to modify the same block e.g. while zeroing.
400  *
401  * - For extending writes case we don't take the shared lock, since it requires
402  *   updating inode i_disksize and/or orphan handling with exclusive lock.
403  *
404  * - shared locking will only be true mostly with overwrites. Otherwise we will
405  *   switch to exclusive i_rwsem lock.
406  */
407 static ssize_t ext4_dio_write_checks(struct kiocb *iocb, struct iov_iter *from,
408                                      bool *ilock_shared, bool *extend)
409 {
410         struct file *file = iocb->ki_filp;
411         struct inode *inode = file_inode(file);
412         loff_t offset;
413         size_t count;
414         ssize_t ret;
415
416 restart:
417         ret = ext4_generic_write_checks(iocb, from);
418         if (ret <= 0)
419                 goto out;
420
421         offset = iocb->ki_pos;
422         count = ret;
423         if (ext4_extending_io(inode, offset, count))
424                 *extend = true;
425         /*
426          * Determine whether the IO operation will overwrite allocated
427          * and initialized blocks.
428          * We need exclusive i_rwsem for changing security info
429          * in file_modified().
430          */
431         if (*ilock_shared && (!IS_NOSEC(inode) || *extend ||
432              !ext4_overwrite_io(inode, offset, count))) {
433                 if (iocb->ki_flags & IOCB_NOWAIT) {
434                         ret = -EAGAIN;
435                         goto out;
436                 }
437                 inode_unlock_shared(inode);
438                 *ilock_shared = false;
439                 inode_lock(inode);
440                 goto restart;
441         }
442
443         ret = file_modified(file);
444         if (ret < 0)
445                 goto out;
446
447         return count;
448 out:
449         if (*ilock_shared)
450                 inode_unlock_shared(inode);
451         else
452                 inode_unlock(inode);
453         return ret;
454 }
455
456 static ssize_t ext4_dio_write_iter(struct kiocb *iocb, struct iov_iter *from)
457 {
458         ssize_t ret;
459         handle_t *handle;
460         struct inode *inode = file_inode(iocb->ki_filp);
461         loff_t offset = iocb->ki_pos;
462         size_t count = iov_iter_count(from);
463         const struct iomap_ops *iomap_ops = &ext4_iomap_ops;
464         bool extend = false, unaligned_io = false;
465         bool ilock_shared = true;
466
467         /*
468          * We initially start with shared inode lock unless it is
469          * unaligned IO which needs exclusive lock anyways.
470          */
471         if (ext4_unaligned_io(inode, from, offset)) {
472                 unaligned_io = true;
473                 ilock_shared = false;
474         }
475         /*
476          * Quick check here without any i_rwsem lock to see if it is extending
477          * IO. A more reliable check is done in ext4_dio_write_checks() with
478          * proper locking in place.
479          */
480         if (offset + count > i_size_read(inode))
481                 ilock_shared = false;
482
483         if (iocb->ki_flags & IOCB_NOWAIT) {
484                 if (ilock_shared) {
485                         if (!inode_trylock_shared(inode))
486                                 return -EAGAIN;
487                 } else {
488                         if (!inode_trylock(inode))
489                                 return -EAGAIN;
490                 }
491         } else {
492                 if (ilock_shared)
493                         inode_lock_shared(inode);
494                 else
495                         inode_lock(inode);
496         }
497
498         /* Fallback to buffered I/O if the inode does not support direct I/O. */
499         if (!ext4_should_use_dio(iocb, from)) {
500                 if (ilock_shared)
501                         inode_unlock_shared(inode);
502                 else
503                         inode_unlock(inode);
504                 return ext4_buffered_write_iter(iocb, from);
505         }
506
507         ret = ext4_dio_write_checks(iocb, from, &ilock_shared, &extend);
508         if (ret <= 0)
509                 return ret;
510
511         /* if we're going to block and IOCB_NOWAIT is set, return -EAGAIN */
512         if ((iocb->ki_flags & IOCB_NOWAIT) && (unaligned_io || extend)) {
513                 ret = -EAGAIN;
514                 goto out;
515         }
516         /*
517          * Make sure inline data cannot be created anymore since we are going
518          * to allocate blocks for DIO. We know the inode does not have any
519          * inline data now because ext4_dio_supported() checked for that.
520          */
521         ext4_clear_inode_state(inode, EXT4_STATE_MAY_INLINE_DATA);
522
523         offset = iocb->ki_pos;
524         count = ret;
525
526         /*
527          * Unaligned direct IO must be serialized among each other as zeroing
528          * of partial blocks of two competing unaligned IOs can result in data
529          * corruption.
530          *
531          * So we make sure we don't allow any unaligned IO in flight.
532          * For IOs where we need not wait (like unaligned non-AIO DIO),
533          * below inode_dio_wait() may anyway become a no-op, since we start
534          * with exclusive lock.
535          */
536         if (unaligned_io)
537                 inode_dio_wait(inode);
538
539         if (extend) {
540                 handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
541                 if (IS_ERR(handle)) {
542                         ret = PTR_ERR(handle);
543                         goto out;
544                 }
545
546                 ret = ext4_orphan_add(handle, inode);
547                 if (ret) {
548                         ext4_journal_stop(handle);
549                         goto out;
550                 }
551
552                 ext4_journal_stop(handle);
553         }
554
555         if (ilock_shared)
556                 iomap_ops = &ext4_iomap_overwrite_ops;
557         ret = iomap_dio_rw(iocb, from, iomap_ops, &ext4_dio_write_ops,
558                            (unaligned_io || extend) ? IOMAP_DIO_FORCE_WAIT : 0,
559                            NULL, 0);
560         if (ret == -ENOTBLK)
561                 ret = 0;
562         if (extend) {
563                 /*
564                  * We always perform extending DIO write synchronously so by
565                  * now the IO is completed and ext4_handle_inode_extension()
566                  * was called. Cleanup the inode in case of error or race with
567                  * writeback of delalloc blocks.
568                  */
569                 WARN_ON_ONCE(ret == -EIOCBQUEUED);
570                 ext4_inode_extension_cleanup(inode, ret);
571         }
572
573 out:
574         if (ilock_shared)
575                 inode_unlock_shared(inode);
576         else
577                 inode_unlock(inode);
578
579         if (ret >= 0 && iov_iter_count(from)) {
580                 ssize_t err;
581                 loff_t endbyte;
582
583                 offset = iocb->ki_pos;
584                 err = ext4_buffered_write_iter(iocb, from);
585                 if (err < 0)
586                         return err;
587
588                 /*
589                  * We need to ensure that the pages within the page cache for
590                  * the range covered by this I/O are written to disk and
591                  * invalidated. This is in attempt to preserve the expected
592                  * direct I/O semantics in the case we fallback to buffered I/O
593                  * to complete off the I/O request.
594                  */
595                 ret += err;
596                 endbyte = offset + err - 1;
597                 err = filemap_write_and_wait_range(iocb->ki_filp->f_mapping,
598                                                    offset, endbyte);
599                 if (!err)
600                         invalidate_mapping_pages(iocb->ki_filp->f_mapping,
601                                                  offset >> PAGE_SHIFT,
602                                                  endbyte >> PAGE_SHIFT);
603         }
604
605         return ret;
606 }
607
608 #ifdef CONFIG_FS_DAX
609 static ssize_t
610 ext4_dax_write_iter(struct kiocb *iocb, struct iov_iter *from)
611 {
612         ssize_t ret;
613         size_t count;
614         loff_t offset;
615         handle_t *handle;
616         bool extend = false;
617         struct inode *inode = file_inode(iocb->ki_filp);
618
619         if (iocb->ki_flags & IOCB_NOWAIT) {
620                 if (!inode_trylock(inode))
621                         return -EAGAIN;
622         } else {
623                 inode_lock(inode);
624         }
625
626         ret = ext4_write_checks(iocb, from);
627         if (ret <= 0)
628                 goto out;
629
630         offset = iocb->ki_pos;
631         count = iov_iter_count(from);
632
633         if (offset + count > EXT4_I(inode)->i_disksize) {
634                 handle = ext4_journal_start(inode, EXT4_HT_INODE, 2);
635                 if (IS_ERR(handle)) {
636                         ret = PTR_ERR(handle);
637                         goto out;
638                 }
639
640                 ret = ext4_orphan_add(handle, inode);
641                 if (ret) {
642                         ext4_journal_stop(handle);
643                         goto out;
644                 }
645
646                 extend = true;
647                 ext4_journal_stop(handle);
648         }
649
650         ret = dax_iomap_rw(iocb, from, &ext4_iomap_ops);
651
652         if (extend) {
653                 ret = ext4_handle_inode_extension(inode, offset, ret);
654                 ext4_inode_extension_cleanup(inode, ret);
655         }
656 out:
657         inode_unlock(inode);
658         if (ret > 0)
659                 ret = generic_write_sync(iocb, ret);
660         return ret;
661 }
662 #endif
663
664 static ssize_t
665 ext4_file_write_iter(struct kiocb *iocb, struct iov_iter *from)
666 {
667         struct inode *inode = file_inode(iocb->ki_filp);
668
669         if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
670                 return -EIO;
671
672 #ifdef CONFIG_FS_DAX
673         if (IS_DAX(inode))
674                 return ext4_dax_write_iter(iocb, from);
675 #endif
676         if (iocb->ki_flags & IOCB_DIRECT)
677                 return ext4_dio_write_iter(iocb, from);
678         else
679                 return ext4_buffered_write_iter(iocb, from);
680 }
681
682 #ifdef CONFIG_FS_DAX
683 static vm_fault_t ext4_dax_huge_fault(struct vm_fault *vmf,
684                 enum page_entry_size pe_size)
685 {
686         int error = 0;
687         vm_fault_t result;
688         int retries = 0;
689         handle_t *handle = NULL;
690         struct inode *inode = file_inode(vmf->vma->vm_file);
691         struct super_block *sb = inode->i_sb;
692
693         /*
694          * We have to distinguish real writes from writes which will result in a
695          * COW page; COW writes should *not* poke the journal (the file will not
696          * be changed). Doing so would cause unintended failures when mounted
697          * read-only.
698          *
699          * We check for VM_SHARED rather than vmf->cow_page since the latter is
700          * unset for pe_size != PE_SIZE_PTE (i.e. only in do_cow_fault); for
701          * other sizes, dax_iomap_fault will handle splitting / fallback so that
702          * we eventually come back with a COW page.
703          */
704         bool write = (vmf->flags & FAULT_FLAG_WRITE) &&
705                 (vmf->vma->vm_flags & VM_SHARED);
706         struct address_space *mapping = vmf->vma->vm_file->f_mapping;
707         pfn_t pfn;
708
709         if (write) {
710                 sb_start_pagefault(sb);
711                 file_update_time(vmf->vma->vm_file);
712                 filemap_invalidate_lock_shared(mapping);
713 retry:
714                 handle = ext4_journal_start_sb(sb, EXT4_HT_WRITE_PAGE,
715                                                EXT4_DATA_TRANS_BLOCKS(sb));
716                 if (IS_ERR(handle)) {
717                         filemap_invalidate_unlock_shared(mapping);
718                         sb_end_pagefault(sb);
719                         return VM_FAULT_SIGBUS;
720                 }
721         } else {
722                 filemap_invalidate_lock_shared(mapping);
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                 filemap_invalidate_unlock_shared(mapping);
735                 sb_end_pagefault(sb);
736         } else {
737                 filemap_invalidate_unlock_shared(mapping);
738         }
739
740         return result;
741 }
742
743 static vm_fault_t ext4_dax_fault(struct vm_fault *vmf)
744 {
745         return ext4_dax_huge_fault(vmf, PE_SIZE_PTE);
746 }
747
748 static 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
758 static const struct vm_operations_struct ext4_file_vm_ops = {
759         .fault          = filemap_fault,
760         .map_pages      = filemap_map_pages,
761         .page_mkwrite   = ext4_page_mkwrite,
762 };
763
764 static 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
790 static 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, sb, sbi->s_sbh,
826                                             EXT4_JTR_NONE);
827         if (err)
828                 goto out_journal;
829         lock_buffer(sbi->s_sbh);
830         strncpy(sbi->s_es->s_last_mounted, cp,
831                 sizeof(sbi->s_es->s_last_mounted));
832         ext4_superblock_csum_set(sb);
833         unlock_buffer(sbi->s_sbh);
834         ext4_handle_dirty_metadata(handle, NULL, sbi->s_sbh);
835 out_journal:
836         ext4_journal_stop(handle);
837 out:
838         sb_end_intwrite(sb);
839         return err;
840 }
841
842 static int ext4_file_open(struct inode *inode, struct file *filp)
843 {
844         int ret;
845
846         if (unlikely(ext4_forced_shutdown(EXT4_SB(inode->i_sb))))
847                 return -EIO;
848
849         ret = ext4_sample_last_mounted(inode->i_sb, filp->f_path.mnt);
850         if (ret)
851                 return ret;
852
853         ret = fscrypt_file_open(inode, filp);
854         if (ret)
855                 return ret;
856
857         ret = fsverity_file_open(inode, filp);
858         if (ret)
859                 return ret;
860
861         /*
862          * Set up the jbd2_inode if we are opening the inode for
863          * writing and the journal is present
864          */
865         if (filp->f_mode & FMODE_WRITE) {
866                 ret = ext4_inode_attach_jinode(inode);
867                 if (ret < 0)
868                         return ret;
869         }
870
871         filp->f_mode |= FMODE_NOWAIT | FMODE_BUF_RASYNC;
872         return dquot_file_open(inode, filp);
873 }
874
875 /*
876  * ext4_llseek() handles both block-mapped and extent-mapped maxbytes values
877  * by calling generic_file_llseek_size() with the appropriate maxbytes
878  * value for each.
879  */
880 loff_t ext4_llseek(struct file *file, loff_t offset, int whence)
881 {
882         struct inode *inode = file->f_mapping->host;
883         loff_t maxbytes;
884
885         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
886                 maxbytes = EXT4_SB(inode->i_sb)->s_bitmap_maxbytes;
887         else
888                 maxbytes = inode->i_sb->s_maxbytes;
889
890         switch (whence) {
891         default:
892                 return generic_file_llseek_size(file, offset, whence,
893                                                 maxbytes, i_size_read(inode));
894         case SEEK_HOLE:
895                 inode_lock_shared(inode);
896                 offset = iomap_seek_hole(inode, offset,
897                                          &ext4_iomap_report_ops);
898                 inode_unlock_shared(inode);
899                 break;
900         case SEEK_DATA:
901                 inode_lock_shared(inode);
902                 offset = iomap_seek_data(inode, offset,
903                                          &ext4_iomap_report_ops);
904                 inode_unlock_shared(inode);
905                 break;
906         }
907
908         if (offset < 0)
909                 return offset;
910         return vfs_setpos(file, offset, maxbytes);
911 }
912
913 const struct file_operations ext4_file_operations = {
914         .llseek         = ext4_llseek,
915         .read_iter      = ext4_file_read_iter,
916         .write_iter     = ext4_file_write_iter,
917         .iopoll         = iocb_bio_iopoll,
918         .unlocked_ioctl = ext4_ioctl,
919 #ifdef CONFIG_COMPAT
920         .compat_ioctl   = ext4_compat_ioctl,
921 #endif
922         .mmap           = ext4_file_mmap,
923         .mmap_supported_flags = MAP_SYNC,
924         .open           = ext4_file_open,
925         .release        = ext4_release_file,
926         .fsync          = ext4_sync_file,
927         .get_unmapped_area = thp_get_unmapped_area,
928         .splice_read    = generic_file_splice_read,
929         .splice_write   = iter_file_splice_write,
930         .fallocate      = ext4_fallocate,
931 };
932
933 const struct inode_operations ext4_file_inode_operations = {
934         .setattr        = ext4_setattr,
935         .getattr        = ext4_file_getattr,
936         .listxattr      = ext4_listxattr,
937         .get_acl        = ext4_get_acl,
938         .set_acl        = ext4_set_acl,
939         .fiemap         = ext4_fiemap,
940         .fileattr_get   = ext4_fileattr_get,
941         .fileattr_set   = ext4_fileattr_set,
942 };
943