Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (c) 2000-2005 Silicon Graphics, Inc.
4 * All Rights Reserved.
5 */
6#include "xfs.h"
7#include "xfs_fs.h"
8#include "xfs_shared.h"
9#include "xfs_format.h"
10#include "xfs_log_format.h"
11#include "xfs_trans_resv.h"
12#include "xfs_mount.h"
13#include "xfs_inode.h"
14#include "xfs_trans.h"
15#include "xfs_inode_item.h"
16#include "xfs_bmap.h"
17#include "xfs_bmap_util.h"
18#include "xfs_dir2.h"
19#include "xfs_dir2_priv.h"
20#include "xfs_ioctl.h"
21#include "xfs_trace.h"
22#include "xfs_log.h"
23#include "xfs_icache.h"
24#include "xfs_pnfs.h"
25#include "xfs_iomap.h"
26#include "xfs_reflink.h"
27
28#include <linux/dax.h>
29#include <linux/falloc.h>
30#include <linux/backing-dev.h>
31#include <linux/mman.h>
32#include <linux/fadvise.h>
33#include <linux/mount.h>
34
35static const struct vm_operations_struct xfs_file_vm_ops;
36
37/*
38 * Decide if the given file range is aligned to the size of the fundamental
39 * allocation unit for the file.
40 */
41static bool
42xfs_is_falloc_aligned(
43 struct xfs_inode *ip,
44 loff_t pos,
45 long long int len)
46{
47 struct xfs_mount *mp = ip->i_mount;
48 uint64_t mask;
49
50 if (XFS_IS_REALTIME_INODE(ip)) {
51 if (!is_power_of_2(mp->m_sb.sb_rextsize)) {
52 u64 rextbytes;
53 u32 mod;
54
55 rextbytes = XFS_FSB_TO_B(mp, mp->m_sb.sb_rextsize);
56 div_u64_rem(pos, rextbytes, &mod);
57 if (mod)
58 return false;
59 div_u64_rem(len, rextbytes, &mod);
60 return mod == 0;
61 }
62 mask = XFS_FSB_TO_B(mp, mp->m_sb.sb_rextsize) - 1;
63 } else {
64 mask = mp->m_sb.sb_blocksize - 1;
65 }
66
67 return !((pos | len) & mask);
68}
69
70/*
71 * Fsync operations on directories are much simpler than on regular files,
72 * as there is no file data to flush, and thus also no need for explicit
73 * cache flush operations, and there are no non-transaction metadata updates
74 * on directories either.
75 */
76STATIC int
77xfs_dir_fsync(
78 struct file *file,
79 loff_t start,
80 loff_t end,
81 int datasync)
82{
83 struct xfs_inode *ip = XFS_I(file->f_mapping->host);
84
85 trace_xfs_dir_fsync(ip);
86 return xfs_log_force_inode(ip);
87}
88
89static xfs_csn_t
90xfs_fsync_seq(
91 struct xfs_inode *ip,
92 bool datasync)
93{
94 if (!xfs_ipincount(ip))
95 return 0;
96 if (datasync && !(ip->i_itemp->ili_fsync_fields & ~XFS_ILOG_TIMESTAMP))
97 return 0;
98 return ip->i_itemp->ili_commit_seq;
99}
100
101/*
102 * All metadata updates are logged, which means that we just have to flush the
103 * log up to the latest LSN that touched the inode.
104 *
105 * If we have concurrent fsync/fdatasync() calls, we need them to all block on
106 * the log force before we clear the ili_fsync_fields field. This ensures that
107 * we don't get a racing sync operation that does not wait for the metadata to
108 * hit the journal before returning. If we race with clearing ili_fsync_fields,
109 * then all that will happen is the log force will do nothing as the lsn will
110 * already be on disk. We can't race with setting ili_fsync_fields because that
111 * is done under XFS_ILOCK_EXCL, and that can't happen because we hold the lock
112 * shared until after the ili_fsync_fields is cleared.
113 */
114static int
115xfs_fsync_flush_log(
116 struct xfs_inode *ip,
117 bool datasync,
118 int *log_flushed)
119{
120 int error = 0;
121 xfs_csn_t seq;
122
123 xfs_ilock(ip, XFS_ILOCK_SHARED);
124 seq = xfs_fsync_seq(ip, datasync);
125 if (seq) {
126 error = xfs_log_force_seq(ip->i_mount, seq, XFS_LOG_SYNC,
127 log_flushed);
128
129 spin_lock(&ip->i_itemp->ili_lock);
130 ip->i_itemp->ili_fsync_fields = 0;
131 spin_unlock(&ip->i_itemp->ili_lock);
132 }
133 xfs_iunlock(ip, XFS_ILOCK_SHARED);
134 return error;
135}
136
137STATIC int
138xfs_file_fsync(
139 struct file *file,
140 loff_t start,
141 loff_t end,
142 int datasync)
143{
144 struct xfs_inode *ip = XFS_I(file->f_mapping->host);
145 struct xfs_mount *mp = ip->i_mount;
146 int error, err2;
147 int log_flushed = 0;
148
149 trace_xfs_file_fsync(ip);
150
151 error = file_write_and_wait_range(file, start, end);
152 if (error)
153 return error;
154
155 if (xfs_is_shutdown(mp))
156 return -EIO;
157
158 xfs_iflags_clear(ip, XFS_ITRUNCATED);
159
160 /*
161 * If we have an RT and/or log subvolume we need to make sure to flush
162 * the write cache the device used for file data first. This is to
163 * ensure newly written file data make it to disk before logging the new
164 * inode size in case of an extending write.
165 */
166 if (XFS_IS_REALTIME_INODE(ip))
167 error = blkdev_issue_flush(mp->m_rtdev_targp->bt_bdev);
168 else if (mp->m_logdev_targp != mp->m_ddev_targp)
169 error = blkdev_issue_flush(mp->m_ddev_targp->bt_bdev);
170
171 /*
172 * Any inode that has dirty modifications in the log is pinned. The
173 * racy check here for a pinned inode will not catch modifications
174 * that happen concurrently to the fsync call, but fsync semantics
175 * only require to sync previously completed I/O.
176 */
177 if (xfs_ipincount(ip)) {
178 err2 = xfs_fsync_flush_log(ip, datasync, &log_flushed);
179 if (err2 && !error)
180 error = err2;
181 }
182
183 /*
184 * If we only have a single device, and the log force about was
185 * a no-op we might have to flush the data device cache here.
186 * This can only happen for fdatasync/O_DSYNC if we were overwriting
187 * an already allocated file and thus do not have any metadata to
188 * commit.
189 */
190 if (!log_flushed && !XFS_IS_REALTIME_INODE(ip) &&
191 mp->m_logdev_targp == mp->m_ddev_targp) {
192 err2 = blkdev_issue_flush(mp->m_ddev_targp->bt_bdev);
193 if (err2 && !error)
194 error = err2;
195 }
196
197 return error;
198}
199
200static int
201xfs_ilock_iocb(
202 struct kiocb *iocb,
203 unsigned int lock_mode)
204{
205 struct xfs_inode *ip = XFS_I(file_inode(iocb->ki_filp));
206
207 if (iocb->ki_flags & IOCB_NOWAIT) {
208 if (!xfs_ilock_nowait(ip, lock_mode))
209 return -EAGAIN;
210 } else {
211 xfs_ilock(ip, lock_mode);
212 }
213
214 return 0;
215}
216
217STATIC ssize_t
218xfs_file_dio_read(
219 struct kiocb *iocb,
220 struct iov_iter *to)
221{
222 struct xfs_inode *ip = XFS_I(file_inode(iocb->ki_filp));
223 ssize_t ret;
224
225 trace_xfs_file_direct_read(iocb, to);
226
227 if (!iov_iter_count(to))
228 return 0; /* skip atime */
229
230 file_accessed(iocb->ki_filp);
231
232 ret = xfs_ilock_iocb(iocb, XFS_IOLOCK_SHARED);
233 if (ret)
234 return ret;
235 ret = iomap_dio_rw(iocb, to, &xfs_read_iomap_ops, NULL, 0, NULL, 0);
236 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
237
238 return ret;
239}
240
241static noinline ssize_t
242xfs_file_dax_read(
243 struct kiocb *iocb,
244 struct iov_iter *to)
245{
246 struct xfs_inode *ip = XFS_I(iocb->ki_filp->f_mapping->host);
247 ssize_t ret = 0;
248
249 trace_xfs_file_dax_read(iocb, to);
250
251 if (!iov_iter_count(to))
252 return 0; /* skip atime */
253
254 ret = xfs_ilock_iocb(iocb, XFS_IOLOCK_SHARED);
255 if (ret)
256 return ret;
257 ret = dax_iomap_rw(iocb, to, &xfs_read_iomap_ops);
258 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
259
260 file_accessed(iocb->ki_filp);
261 return ret;
262}
263
264STATIC ssize_t
265xfs_file_buffered_read(
266 struct kiocb *iocb,
267 struct iov_iter *to)
268{
269 struct xfs_inode *ip = XFS_I(file_inode(iocb->ki_filp));
270 ssize_t ret;
271
272 trace_xfs_file_buffered_read(iocb, to);
273
274 ret = xfs_ilock_iocb(iocb, XFS_IOLOCK_SHARED);
275 if (ret)
276 return ret;
277 ret = generic_file_read_iter(iocb, to);
278 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
279
280 return ret;
281}
282
283STATIC ssize_t
284xfs_file_read_iter(
285 struct kiocb *iocb,
286 struct iov_iter *to)
287{
288 struct inode *inode = file_inode(iocb->ki_filp);
289 struct xfs_mount *mp = XFS_I(inode)->i_mount;
290 ssize_t ret = 0;
291
292 XFS_STATS_INC(mp, xs_read_calls);
293
294 if (xfs_is_shutdown(mp))
295 return -EIO;
296
297 if (IS_DAX(inode))
298 ret = xfs_file_dax_read(iocb, to);
299 else if (iocb->ki_flags & IOCB_DIRECT)
300 ret = xfs_file_dio_read(iocb, to);
301 else
302 ret = xfs_file_buffered_read(iocb, to);
303
304 if (ret > 0)
305 XFS_STATS_ADD(mp, xs_read_bytes, ret);
306 return ret;
307}
308
309/*
310 * Common pre-write limit and setup checks.
311 *
312 * Called with the iolocked held either shared and exclusive according to
313 * @iolock, and returns with it held. Might upgrade the iolock to exclusive
314 * if called for a direct write beyond i_size.
315 */
316STATIC ssize_t
317xfs_file_write_checks(
318 struct kiocb *iocb,
319 struct iov_iter *from,
320 unsigned int *iolock)
321{
322 struct file *file = iocb->ki_filp;
323 struct inode *inode = file->f_mapping->host;
324 struct xfs_inode *ip = XFS_I(inode);
325 ssize_t error = 0;
326 size_t count = iov_iter_count(from);
327 bool drained_dio = false;
328 loff_t isize;
329
330restart:
331 error = generic_write_checks(iocb, from);
332 if (error <= 0)
333 return error;
334
335 if (iocb->ki_flags & IOCB_NOWAIT) {
336 error = break_layout(inode, false);
337 if (error == -EWOULDBLOCK)
338 error = -EAGAIN;
339 } else {
340 error = xfs_break_layouts(inode, iolock, BREAK_WRITE);
341 }
342
343 if (error)
344 return error;
345
346 /*
347 * For changing security info in file_remove_privs() we need i_rwsem
348 * exclusively.
349 */
350 if (*iolock == XFS_IOLOCK_SHARED && !IS_NOSEC(inode)) {
351 xfs_iunlock(ip, *iolock);
352 *iolock = XFS_IOLOCK_EXCL;
353 error = xfs_ilock_iocb(iocb, *iolock);
354 if (error) {
355 *iolock = 0;
356 return error;
357 }
358 goto restart;
359 }
360
361 /*
362 * If the offset is beyond the size of the file, we need to zero any
363 * blocks that fall between the existing EOF and the start of this
364 * write. If zeroing is needed and we are currently holding the iolock
365 * shared, we need to update it to exclusive which implies having to
366 * redo all checks before.
367 *
368 * We need to serialise against EOF updates that occur in IO completions
369 * here. We want to make sure that nobody is changing the size while we
370 * do this check until we have placed an IO barrier (i.e. hold the
371 * XFS_IOLOCK_EXCL) that prevents new IO from being dispatched. The
372 * spinlock effectively forms a memory barrier once we have the
373 * XFS_IOLOCK_EXCL so we are guaranteed to see the latest EOF value and
374 * hence be able to correctly determine if we need to run zeroing.
375 *
376 * We can do an unlocked check here safely as IO completion can only
377 * extend EOF. Truncate is locked out at this point, so the EOF can
378 * not move backwards, only forwards. Hence we only need to take the
379 * slow path and spin locks when we are at or beyond the current EOF.
380 */
381 if (iocb->ki_pos <= i_size_read(inode))
382 goto out;
383
384 spin_lock(&ip->i_flags_lock);
385 isize = i_size_read(inode);
386 if (iocb->ki_pos > isize) {
387 spin_unlock(&ip->i_flags_lock);
388
389 if (iocb->ki_flags & IOCB_NOWAIT)
390 return -EAGAIN;
391
392 if (!drained_dio) {
393 if (*iolock == XFS_IOLOCK_SHARED) {
394 xfs_iunlock(ip, *iolock);
395 *iolock = XFS_IOLOCK_EXCL;
396 xfs_ilock(ip, *iolock);
397 iov_iter_reexpand(from, count);
398 }
399 /*
400 * We now have an IO submission barrier in place, but
401 * AIO can do EOF updates during IO completion and hence
402 * we now need to wait for all of them to drain. Non-AIO
403 * DIO will have drained before we are given the
404 * XFS_IOLOCK_EXCL, and so for most cases this wait is a
405 * no-op.
406 */
407 inode_dio_wait(inode);
408 drained_dio = true;
409 goto restart;
410 }
411
412 trace_xfs_zero_eof(ip, isize, iocb->ki_pos - isize);
413 error = xfs_zero_range(ip, isize, iocb->ki_pos - isize, NULL);
414 if (error)
415 return error;
416 } else
417 spin_unlock(&ip->i_flags_lock);
418
419out:
420 return kiocb_modified(iocb);
421}
422
423static int
424xfs_dio_write_end_io(
425 struct kiocb *iocb,
426 ssize_t size,
427 int error,
428 unsigned flags)
429{
430 struct inode *inode = file_inode(iocb->ki_filp);
431 struct xfs_inode *ip = XFS_I(inode);
432 loff_t offset = iocb->ki_pos;
433 unsigned int nofs_flag;
434
435 trace_xfs_end_io_direct_write(ip, offset, size);
436
437 if (xfs_is_shutdown(ip->i_mount))
438 return -EIO;
439
440 if (error)
441 return error;
442 if (!size)
443 return 0;
444
445 /*
446 * Capture amount written on completion as we can't reliably account
447 * for it on submission.
448 */
449 XFS_STATS_ADD(ip->i_mount, xs_write_bytes, size);
450
451 /*
452 * We can allocate memory here while doing writeback on behalf of
453 * memory reclaim. To avoid memory allocation deadlocks set the
454 * task-wide nofs context for the following operations.
455 */
456 nofs_flag = memalloc_nofs_save();
457
458 if (flags & IOMAP_DIO_COW) {
459 error = xfs_reflink_end_cow(ip, offset, size);
460 if (error)
461 goto out;
462 }
463
464 /*
465 * Unwritten conversion updates the in-core isize after extent
466 * conversion but before updating the on-disk size. Updating isize any
467 * earlier allows a racing dio read to find unwritten extents before
468 * they are converted.
469 */
470 if (flags & IOMAP_DIO_UNWRITTEN) {
471 error = xfs_iomap_write_unwritten(ip, offset, size, true);
472 goto out;
473 }
474
475 /*
476 * We need to update the in-core inode size here so that we don't end up
477 * with the on-disk inode size being outside the in-core inode size. We
478 * have no other method of updating EOF for AIO, so always do it here
479 * if necessary.
480 *
481 * We need to lock the test/set EOF update as we can be racing with
482 * other IO completions here to update the EOF. Failing to serialise
483 * here can result in EOF moving backwards and Bad Things Happen when
484 * that occurs.
485 *
486 * As IO completion only ever extends EOF, we can do an unlocked check
487 * here to avoid taking the spinlock. If we land within the current EOF,
488 * then we do not need to do an extending update at all, and we don't
489 * need to take the lock to check this. If we race with an update moving
490 * EOF, then we'll either still be beyond EOF and need to take the lock,
491 * or we'll be within EOF and we don't need to take it at all.
492 */
493 if (offset + size <= i_size_read(inode))
494 goto out;
495
496 spin_lock(&ip->i_flags_lock);
497 if (offset + size > i_size_read(inode)) {
498 i_size_write(inode, offset + size);
499 spin_unlock(&ip->i_flags_lock);
500 error = xfs_setfilesize(ip, offset, size);
501 } else {
502 spin_unlock(&ip->i_flags_lock);
503 }
504
505out:
506 memalloc_nofs_restore(nofs_flag);
507 return error;
508}
509
510static const struct iomap_dio_ops xfs_dio_write_ops = {
511 .end_io = xfs_dio_write_end_io,
512};
513
514/*
515 * Handle block aligned direct I/O writes
516 */
517static noinline ssize_t
518xfs_file_dio_write_aligned(
519 struct xfs_inode *ip,
520 struct kiocb *iocb,
521 struct iov_iter *from)
522{
523 unsigned int iolock = XFS_IOLOCK_SHARED;
524 ssize_t ret;
525
526 ret = xfs_ilock_iocb(iocb, iolock);
527 if (ret)
528 return ret;
529 ret = xfs_file_write_checks(iocb, from, &iolock);
530 if (ret)
531 goto out_unlock;
532
533 /*
534 * We don't need to hold the IOLOCK exclusively across the IO, so demote
535 * the iolock back to shared if we had to take the exclusive lock in
536 * xfs_file_write_checks() for other reasons.
537 */
538 if (iolock == XFS_IOLOCK_EXCL) {
539 xfs_ilock_demote(ip, XFS_IOLOCK_EXCL);
540 iolock = XFS_IOLOCK_SHARED;
541 }
542 trace_xfs_file_direct_write(iocb, from);
543 ret = iomap_dio_rw(iocb, from, &xfs_direct_write_iomap_ops,
544 &xfs_dio_write_ops, 0, NULL, 0);
545out_unlock:
546 if (iolock)
547 xfs_iunlock(ip, iolock);
548 return ret;
549}
550
551/*
552 * Handle block unaligned direct I/O writes
553 *
554 * In most cases direct I/O writes will be done holding IOLOCK_SHARED, allowing
555 * them to be done in parallel with reads and other direct I/O writes. However,
556 * if the I/O is not aligned to filesystem blocks, the direct I/O layer may need
557 * to do sub-block zeroing and that requires serialisation against other direct
558 * I/O to the same block. In this case we need to serialise the submission of
559 * the unaligned I/O so that we don't get racing block zeroing in the dio layer.
560 * In the case where sub-block zeroing is not required, we can do concurrent
561 * sub-block dios to the same block successfully.
562 *
563 * Optimistically submit the I/O using the shared lock first, but use the
564 * IOMAP_DIO_OVERWRITE_ONLY flag to tell the lower layers to return -EAGAIN
565 * if block allocation or partial block zeroing would be required. In that case
566 * we try again with the exclusive lock.
567 */
568static noinline ssize_t
569xfs_file_dio_write_unaligned(
570 struct xfs_inode *ip,
571 struct kiocb *iocb,
572 struct iov_iter *from)
573{
574 size_t isize = i_size_read(VFS_I(ip));
575 size_t count = iov_iter_count(from);
576 unsigned int iolock = XFS_IOLOCK_SHARED;
577 unsigned int flags = IOMAP_DIO_OVERWRITE_ONLY;
578 ssize_t ret;
579
580 /*
581 * Extending writes need exclusivity because of the sub-block zeroing
582 * that the DIO code always does for partial tail blocks beyond EOF, so
583 * don't even bother trying the fast path in this case.
584 */
585 if (iocb->ki_pos > isize || iocb->ki_pos + count >= isize) {
586 if (iocb->ki_flags & IOCB_NOWAIT)
587 return -EAGAIN;
588retry_exclusive:
589 iolock = XFS_IOLOCK_EXCL;
590 flags = IOMAP_DIO_FORCE_WAIT;
591 }
592
593 ret = xfs_ilock_iocb(iocb, iolock);
594 if (ret)
595 return ret;
596
597 /*
598 * We can't properly handle unaligned direct I/O to reflink files yet,
599 * as we can't unshare a partial block.
600 */
601 if (xfs_is_cow_inode(ip)) {
602 trace_xfs_reflink_bounce_dio_write(iocb, from);
603 ret = -ENOTBLK;
604 goto out_unlock;
605 }
606
607 ret = xfs_file_write_checks(iocb, from, &iolock);
608 if (ret)
609 goto out_unlock;
610
611 /*
612 * If we are doing exclusive unaligned I/O, this must be the only I/O
613 * in-flight. Otherwise we risk data corruption due to unwritten extent
614 * conversions from the AIO end_io handler. Wait for all other I/O to
615 * drain first.
616 */
617 if (flags & IOMAP_DIO_FORCE_WAIT)
618 inode_dio_wait(VFS_I(ip));
619
620 trace_xfs_file_direct_write(iocb, from);
621 ret = iomap_dio_rw(iocb, from, &xfs_direct_write_iomap_ops,
622 &xfs_dio_write_ops, flags, NULL, 0);
623
624 /*
625 * Retry unaligned I/O with exclusive blocking semantics if the DIO
626 * layer rejected it for mapping or locking reasons. If we are doing
627 * nonblocking user I/O, propagate the error.
628 */
629 if (ret == -EAGAIN && !(iocb->ki_flags & IOCB_NOWAIT)) {
630 ASSERT(flags & IOMAP_DIO_OVERWRITE_ONLY);
631 xfs_iunlock(ip, iolock);
632 goto retry_exclusive;
633 }
634
635out_unlock:
636 if (iolock)
637 xfs_iunlock(ip, iolock);
638 return ret;
639}
640
641static ssize_t
642xfs_file_dio_write(
643 struct kiocb *iocb,
644 struct iov_iter *from)
645{
646 struct xfs_inode *ip = XFS_I(file_inode(iocb->ki_filp));
647 struct xfs_buftarg *target = xfs_inode_buftarg(ip);
648 size_t count = iov_iter_count(from);
649
650 /* direct I/O must be aligned to device logical sector size */
651 if ((iocb->ki_pos | count) & target->bt_logical_sectormask)
652 return -EINVAL;
653 if ((iocb->ki_pos | count) & ip->i_mount->m_blockmask)
654 return xfs_file_dio_write_unaligned(ip, iocb, from);
655 return xfs_file_dio_write_aligned(ip, iocb, from);
656}
657
658static noinline ssize_t
659xfs_file_dax_write(
660 struct kiocb *iocb,
661 struct iov_iter *from)
662{
663 struct inode *inode = iocb->ki_filp->f_mapping->host;
664 struct xfs_inode *ip = XFS_I(inode);
665 unsigned int iolock = XFS_IOLOCK_EXCL;
666 ssize_t ret, error = 0;
667 loff_t pos;
668
669 ret = xfs_ilock_iocb(iocb, iolock);
670 if (ret)
671 return ret;
672 ret = xfs_file_write_checks(iocb, from, &iolock);
673 if (ret)
674 goto out;
675
676 pos = iocb->ki_pos;
677
678 trace_xfs_file_dax_write(iocb, from);
679 ret = dax_iomap_rw(iocb, from, &xfs_dax_write_iomap_ops);
680 if (ret > 0 && iocb->ki_pos > i_size_read(inode)) {
681 i_size_write(inode, iocb->ki_pos);
682 error = xfs_setfilesize(ip, pos, ret);
683 }
684out:
685 if (iolock)
686 xfs_iunlock(ip, iolock);
687 if (error)
688 return error;
689
690 if (ret > 0) {
691 XFS_STATS_ADD(ip->i_mount, xs_write_bytes, ret);
692
693 /* Handle various SYNC-type writes */
694 ret = generic_write_sync(iocb, ret);
695 }
696 return ret;
697}
698
699STATIC ssize_t
700xfs_file_buffered_write(
701 struct kiocb *iocb,
702 struct iov_iter *from)
703{
704 struct inode *inode = iocb->ki_filp->f_mapping->host;
705 struct xfs_inode *ip = XFS_I(inode);
706 ssize_t ret;
707 bool cleared_space = false;
708 unsigned int iolock;
709
710write_retry:
711 iolock = XFS_IOLOCK_EXCL;
712 ret = xfs_ilock_iocb(iocb, iolock);
713 if (ret)
714 return ret;
715
716 ret = xfs_file_write_checks(iocb, from, &iolock);
717 if (ret)
718 goto out;
719
720 /* We can write back this queue in page reclaim */
721 current->backing_dev_info = inode_to_bdi(inode);
722
723 trace_xfs_file_buffered_write(iocb, from);
724 ret = iomap_file_buffered_write(iocb, from,
725 &xfs_buffered_write_iomap_ops);
726 if (likely(ret >= 0))
727 iocb->ki_pos += ret;
728
729 /*
730 * If we hit a space limit, try to free up some lingering preallocated
731 * space before returning an error. In the case of ENOSPC, first try to
732 * write back all dirty inodes to free up some of the excess reserved
733 * metadata space. This reduces the chances that the eofblocks scan
734 * waits on dirty mappings. Since xfs_flush_inodes() is serialized, this
735 * also behaves as a filter to prevent too many eofblocks scans from
736 * running at the same time. Use a synchronous scan to increase the
737 * effectiveness of the scan.
738 */
739 if (ret == -EDQUOT && !cleared_space) {
740 xfs_iunlock(ip, iolock);
741 xfs_blockgc_free_quota(ip, XFS_ICWALK_FLAG_SYNC);
742 cleared_space = true;
743 goto write_retry;
744 } else if (ret == -ENOSPC && !cleared_space) {
745 struct xfs_icwalk icw = {0};
746
747 cleared_space = true;
748 xfs_flush_inodes(ip->i_mount);
749
750 xfs_iunlock(ip, iolock);
751 icw.icw_flags = XFS_ICWALK_FLAG_SYNC;
752 xfs_blockgc_free_space(ip->i_mount, &icw);
753 goto write_retry;
754 }
755
756 current->backing_dev_info = NULL;
757out:
758 if (iolock)
759 xfs_iunlock(ip, iolock);
760
761 if (ret > 0) {
762 XFS_STATS_ADD(ip->i_mount, xs_write_bytes, ret);
763 /* Handle various SYNC-type writes */
764 ret = generic_write_sync(iocb, ret);
765 }
766 return ret;
767}
768
769STATIC ssize_t
770xfs_file_write_iter(
771 struct kiocb *iocb,
772 struct iov_iter *from)
773{
774 struct inode *inode = iocb->ki_filp->f_mapping->host;
775 struct xfs_inode *ip = XFS_I(inode);
776 ssize_t ret;
777 size_t ocount = iov_iter_count(from);
778
779 XFS_STATS_INC(ip->i_mount, xs_write_calls);
780
781 if (ocount == 0)
782 return 0;
783
784 if (xfs_is_shutdown(ip->i_mount))
785 return -EIO;
786
787 if (IS_DAX(inode))
788 return xfs_file_dax_write(iocb, from);
789
790 if (iocb->ki_flags & IOCB_DIRECT) {
791 /*
792 * Allow a directio write to fall back to a buffered
793 * write *only* in the case that we're doing a reflink
794 * CoW. In all other directio scenarios we do not
795 * allow an operation to fall back to buffered mode.
796 */
797 ret = xfs_file_dio_write(iocb, from);
798 if (ret != -ENOTBLK)
799 return ret;
800 }
801
802 return xfs_file_buffered_write(iocb, from);
803}
804
805static void
806xfs_wait_dax_page(
807 struct inode *inode)
808{
809 struct xfs_inode *ip = XFS_I(inode);
810
811 xfs_iunlock(ip, XFS_MMAPLOCK_EXCL);
812 schedule();
813 xfs_ilock(ip, XFS_MMAPLOCK_EXCL);
814}
815
816int
817xfs_break_dax_layouts(
818 struct inode *inode,
819 bool *retry)
820{
821 struct page *page;
822
823 ASSERT(xfs_isilocked(XFS_I(inode), XFS_MMAPLOCK_EXCL));
824
825 page = dax_layout_busy_page(inode->i_mapping);
826 if (!page)
827 return 0;
828
829 *retry = true;
830 return ___wait_var_event(&page->_refcount,
831 atomic_read(&page->_refcount) == 1, TASK_INTERRUPTIBLE,
832 0, 0, xfs_wait_dax_page(inode));
833}
834
835int
836xfs_break_layouts(
837 struct inode *inode,
838 uint *iolock,
839 enum layout_break_reason reason)
840{
841 bool retry;
842 int error;
843
844 ASSERT(xfs_isilocked(XFS_I(inode), XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL));
845
846 do {
847 retry = false;
848 switch (reason) {
849 case BREAK_UNMAP:
850 error = xfs_break_dax_layouts(inode, &retry);
851 if (error || retry)
852 break;
853 fallthrough;
854 case BREAK_WRITE:
855 error = xfs_break_leased_layouts(inode, iolock, &retry);
856 break;
857 default:
858 WARN_ON_ONCE(1);
859 error = -EINVAL;
860 }
861 } while (error == 0 && retry);
862
863 return error;
864}
865
866/* Does this file, inode, or mount want synchronous writes? */
867static inline bool xfs_file_sync_writes(struct file *filp)
868{
869 struct xfs_inode *ip = XFS_I(file_inode(filp));
870
871 if (xfs_has_wsync(ip->i_mount))
872 return true;
873 if (filp->f_flags & (__O_SYNC | O_DSYNC))
874 return true;
875 if (IS_SYNC(file_inode(filp)))
876 return true;
877
878 return false;
879}
880
881#define XFS_FALLOC_FL_SUPPORTED \
882 (FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE | \
883 FALLOC_FL_COLLAPSE_RANGE | FALLOC_FL_ZERO_RANGE | \
884 FALLOC_FL_INSERT_RANGE | FALLOC_FL_UNSHARE_RANGE)
885
886STATIC long
887xfs_file_fallocate(
888 struct file *file,
889 int mode,
890 loff_t offset,
891 loff_t len)
892{
893 struct inode *inode = file_inode(file);
894 struct xfs_inode *ip = XFS_I(inode);
895 long error;
896 uint iolock = XFS_IOLOCK_EXCL | XFS_MMAPLOCK_EXCL;
897 loff_t new_size = 0;
898 bool do_file_insert = false;
899
900 if (!S_ISREG(inode->i_mode))
901 return -EINVAL;
902 if (mode & ~XFS_FALLOC_FL_SUPPORTED)
903 return -EOPNOTSUPP;
904
905 xfs_ilock(ip, iolock);
906 error = xfs_break_layouts(inode, &iolock, BREAK_UNMAP);
907 if (error)
908 goto out_unlock;
909
910 /*
911 * Must wait for all AIO to complete before we continue as AIO can
912 * change the file size on completion without holding any locks we
913 * currently hold. We must do this first because AIO can update both
914 * the on disk and in memory inode sizes, and the operations that follow
915 * require the in-memory size to be fully up-to-date.
916 */
917 inode_dio_wait(inode);
918
919 /*
920 * Now AIO and DIO has drained we flush and (if necessary) invalidate
921 * the cached range over the first operation we are about to run.
922 *
923 * We care about zero and collapse here because they both run a hole
924 * punch over the range first. Because that can zero data, and the range
925 * of invalidation for the shift operations is much larger, we still do
926 * the required flush for collapse in xfs_prepare_shift().
927 *
928 * Insert has the same range requirements as collapse, and we extend the
929 * file first which can zero data. Hence insert has the same
930 * flush/invalidate requirements as collapse and so they are both
931 * handled at the right time by xfs_prepare_shift().
932 */
933 if (mode & (FALLOC_FL_PUNCH_HOLE | FALLOC_FL_ZERO_RANGE |
934 FALLOC_FL_COLLAPSE_RANGE)) {
935 error = xfs_flush_unmap_range(ip, offset, len);
936 if (error)
937 goto out_unlock;
938 }
939
940 error = file_modified(file);
941 if (error)
942 goto out_unlock;
943
944 if (mode & FALLOC_FL_PUNCH_HOLE) {
945 error = xfs_free_file_space(ip, offset, len);
946 if (error)
947 goto out_unlock;
948 } else if (mode & FALLOC_FL_COLLAPSE_RANGE) {
949 if (!xfs_is_falloc_aligned(ip, offset, len)) {
950 error = -EINVAL;
951 goto out_unlock;
952 }
953
954 /*
955 * There is no need to overlap collapse range with EOF,
956 * in which case it is effectively a truncate operation
957 */
958 if (offset + len >= i_size_read(inode)) {
959 error = -EINVAL;
960 goto out_unlock;
961 }
962
963 new_size = i_size_read(inode) - len;
964
965 error = xfs_collapse_file_space(ip, offset, len);
966 if (error)
967 goto out_unlock;
968 } else if (mode & FALLOC_FL_INSERT_RANGE) {
969 loff_t isize = i_size_read(inode);
970
971 if (!xfs_is_falloc_aligned(ip, offset, len)) {
972 error = -EINVAL;
973 goto out_unlock;
974 }
975
976 /*
977 * New inode size must not exceed ->s_maxbytes, accounting for
978 * possible signed overflow.
979 */
980 if (inode->i_sb->s_maxbytes - isize < len) {
981 error = -EFBIG;
982 goto out_unlock;
983 }
984 new_size = isize + len;
985
986 /* Offset should be less than i_size */
987 if (offset >= isize) {
988 error = -EINVAL;
989 goto out_unlock;
990 }
991 do_file_insert = true;
992 } else {
993 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
994 offset + len > i_size_read(inode)) {
995 new_size = offset + len;
996 error = inode_newsize_ok(inode, new_size);
997 if (error)
998 goto out_unlock;
999 }
1000
1001 if (mode & FALLOC_FL_ZERO_RANGE) {
1002 /*
1003 * Punch a hole and prealloc the range. We use a hole
1004 * punch rather than unwritten extent conversion for two
1005 * reasons:
1006 *
1007 * 1.) Hole punch handles partial block zeroing for us.
1008 * 2.) If prealloc returns ENOSPC, the file range is
1009 * still zero-valued by virtue of the hole punch.
1010 */
1011 unsigned int blksize = i_blocksize(inode);
1012
1013 trace_xfs_zero_file_space(ip);
1014
1015 error = xfs_free_file_space(ip, offset, len);
1016 if (error)
1017 goto out_unlock;
1018
1019 len = round_up(offset + len, blksize) -
1020 round_down(offset, blksize);
1021 offset = round_down(offset, blksize);
1022 } else if (mode & FALLOC_FL_UNSHARE_RANGE) {
1023 error = xfs_reflink_unshare(ip, offset, len);
1024 if (error)
1025 goto out_unlock;
1026 } else {
1027 /*
1028 * If always_cow mode we can't use preallocations and
1029 * thus should not create them.
1030 */
1031 if (xfs_is_always_cow_inode(ip)) {
1032 error = -EOPNOTSUPP;
1033 goto out_unlock;
1034 }
1035 }
1036
1037 if (!xfs_is_always_cow_inode(ip)) {
1038 error = xfs_alloc_file_space(ip, offset, len);
1039 if (error)
1040 goto out_unlock;
1041 }
1042 }
1043
1044 /* Change file size if needed */
1045 if (new_size) {
1046 struct iattr iattr;
1047
1048 iattr.ia_valid = ATTR_SIZE;
1049 iattr.ia_size = new_size;
1050 error = xfs_vn_setattr_size(file_mnt_user_ns(file),
1051 file_dentry(file), &iattr);
1052 if (error)
1053 goto out_unlock;
1054 }
1055
1056 /*
1057 * Perform hole insertion now that the file size has been
1058 * updated so that if we crash during the operation we don't
1059 * leave shifted extents past EOF and hence losing access to
1060 * the data that is contained within them.
1061 */
1062 if (do_file_insert) {
1063 error = xfs_insert_file_space(ip, offset, len);
1064 if (error)
1065 goto out_unlock;
1066 }
1067
1068 if (xfs_file_sync_writes(file))
1069 error = xfs_log_force_inode(ip);
1070
1071out_unlock:
1072 xfs_iunlock(ip, iolock);
1073 return error;
1074}
1075
1076STATIC int
1077xfs_file_fadvise(
1078 struct file *file,
1079 loff_t start,
1080 loff_t end,
1081 int advice)
1082{
1083 struct xfs_inode *ip = XFS_I(file_inode(file));
1084 int ret;
1085 int lockflags = 0;
1086
1087 /*
1088 * Operations creating pages in page cache need protection from hole
1089 * punching and similar ops
1090 */
1091 if (advice == POSIX_FADV_WILLNEED) {
1092 lockflags = XFS_IOLOCK_SHARED;
1093 xfs_ilock(ip, lockflags);
1094 }
1095 ret = generic_fadvise(file, start, end, advice);
1096 if (lockflags)
1097 xfs_iunlock(ip, lockflags);
1098 return ret;
1099}
1100
1101STATIC loff_t
1102xfs_file_remap_range(
1103 struct file *file_in,
1104 loff_t pos_in,
1105 struct file *file_out,
1106 loff_t pos_out,
1107 loff_t len,
1108 unsigned int remap_flags)
1109{
1110 struct inode *inode_in = file_inode(file_in);
1111 struct xfs_inode *src = XFS_I(inode_in);
1112 struct inode *inode_out = file_inode(file_out);
1113 struct xfs_inode *dest = XFS_I(inode_out);
1114 struct xfs_mount *mp = src->i_mount;
1115 loff_t remapped = 0;
1116 xfs_extlen_t cowextsize;
1117 int ret;
1118
1119 if (remap_flags & ~(REMAP_FILE_DEDUP | REMAP_FILE_ADVISORY))
1120 return -EINVAL;
1121
1122 if (!xfs_has_reflink(mp))
1123 return -EOPNOTSUPP;
1124
1125 if (xfs_is_shutdown(mp))
1126 return -EIO;
1127
1128 /* Prepare and then clone file data. */
1129 ret = xfs_reflink_remap_prep(file_in, pos_in, file_out, pos_out,
1130 &len, remap_flags);
1131 if (ret || len == 0)
1132 return ret;
1133
1134 trace_xfs_reflink_remap_range(src, pos_in, len, dest, pos_out);
1135
1136 ret = xfs_reflink_remap_blocks(src, pos_in, dest, pos_out, len,
1137 &remapped);
1138 if (ret)
1139 goto out_unlock;
1140
1141 /*
1142 * Carry the cowextsize hint from src to dest if we're sharing the
1143 * entire source file to the entire destination file, the source file
1144 * has a cowextsize hint, and the destination file does not.
1145 */
1146 cowextsize = 0;
1147 if (pos_in == 0 && len == i_size_read(inode_in) &&
1148 (src->i_diflags2 & XFS_DIFLAG2_COWEXTSIZE) &&
1149 pos_out == 0 && len >= i_size_read(inode_out) &&
1150 !(dest->i_diflags2 & XFS_DIFLAG2_COWEXTSIZE))
1151 cowextsize = src->i_cowextsize;
1152
1153 ret = xfs_reflink_update_dest(dest, pos_out + len, cowextsize,
1154 remap_flags);
1155 if (ret)
1156 goto out_unlock;
1157
1158 if (xfs_file_sync_writes(file_in) || xfs_file_sync_writes(file_out))
1159 xfs_log_force_inode(dest);
1160out_unlock:
1161 xfs_iunlock2_io_mmap(src, dest);
1162 if (ret)
1163 trace_xfs_reflink_remap_range_error(dest, ret, _RET_IP_);
1164 return remapped > 0 ? remapped : ret;
1165}
1166
1167STATIC int
1168xfs_file_open(
1169 struct inode *inode,
1170 struct file *file)
1171{
1172 if (xfs_is_shutdown(XFS_M(inode->i_sb)))
1173 return -EIO;
1174 file->f_mode |= FMODE_NOWAIT | FMODE_BUF_RASYNC | FMODE_BUF_WASYNC;
1175 return generic_file_open(inode, file);
1176}
1177
1178STATIC int
1179xfs_dir_open(
1180 struct inode *inode,
1181 struct file *file)
1182{
1183 struct xfs_inode *ip = XFS_I(inode);
1184 unsigned int mode;
1185 int error;
1186
1187 error = xfs_file_open(inode, file);
1188 if (error)
1189 return error;
1190
1191 /*
1192 * If there are any blocks, read-ahead block 0 as we're almost
1193 * certain to have the next operation be a read there.
1194 */
1195 mode = xfs_ilock_data_map_shared(ip);
1196 if (ip->i_df.if_nextents > 0)
1197 error = xfs_dir3_data_readahead(ip, 0, 0);
1198 xfs_iunlock(ip, mode);
1199 return error;
1200}
1201
1202STATIC int
1203xfs_file_release(
1204 struct inode *inode,
1205 struct file *filp)
1206{
1207 return xfs_release(XFS_I(inode));
1208}
1209
1210STATIC int
1211xfs_file_readdir(
1212 struct file *file,
1213 struct dir_context *ctx)
1214{
1215 struct inode *inode = file_inode(file);
1216 xfs_inode_t *ip = XFS_I(inode);
1217 size_t bufsize;
1218
1219 /*
1220 * The Linux API doesn't pass down the total size of the buffer
1221 * we read into down to the filesystem. With the filldir concept
1222 * it's not needed for correct information, but the XFS dir2 leaf
1223 * code wants an estimate of the buffer size to calculate it's
1224 * readahead window and size the buffers used for mapping to
1225 * physical blocks.
1226 *
1227 * Try to give it an estimate that's good enough, maybe at some
1228 * point we can change the ->readdir prototype to include the
1229 * buffer size. For now we use the current glibc buffer size.
1230 */
1231 bufsize = (size_t)min_t(loff_t, XFS_READDIR_BUFSIZE, ip->i_disk_size);
1232
1233 return xfs_readdir(NULL, ip, ctx, bufsize);
1234}
1235
1236STATIC loff_t
1237xfs_file_llseek(
1238 struct file *file,
1239 loff_t offset,
1240 int whence)
1241{
1242 struct inode *inode = file->f_mapping->host;
1243
1244 if (xfs_is_shutdown(XFS_I(inode)->i_mount))
1245 return -EIO;
1246
1247 switch (whence) {
1248 default:
1249 return generic_file_llseek(file, offset, whence);
1250 case SEEK_HOLE:
1251 offset = iomap_seek_hole(inode, offset, &xfs_seek_iomap_ops);
1252 break;
1253 case SEEK_DATA:
1254 offset = iomap_seek_data(inode, offset, &xfs_seek_iomap_ops);
1255 break;
1256 }
1257
1258 if (offset < 0)
1259 return offset;
1260 return vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
1261}
1262
1263#ifdef CONFIG_FS_DAX
1264static inline vm_fault_t
1265xfs_dax_fault(
1266 struct vm_fault *vmf,
1267 enum page_entry_size pe_size,
1268 bool write_fault,
1269 pfn_t *pfn)
1270{
1271 return dax_iomap_fault(vmf, pe_size, pfn, NULL,
1272 (write_fault && !vmf->cow_page) ?
1273 &xfs_dax_write_iomap_ops :
1274 &xfs_read_iomap_ops);
1275}
1276#else
1277static inline vm_fault_t
1278xfs_dax_fault(
1279 struct vm_fault *vmf,
1280 enum page_entry_size pe_size,
1281 bool write_fault,
1282 pfn_t *pfn)
1283{
1284 ASSERT(0);
1285 return VM_FAULT_SIGBUS;
1286}
1287#endif
1288
1289/*
1290 * Locking for serialisation of IO during page faults. This results in a lock
1291 * ordering of:
1292 *
1293 * mmap_lock (MM)
1294 * sb_start_pagefault(vfs, freeze)
1295 * invalidate_lock (vfs/XFS_MMAPLOCK - truncate serialisation)
1296 * page_lock (MM)
1297 * i_lock (XFS - extent map serialisation)
1298 */
1299static vm_fault_t
1300__xfs_filemap_fault(
1301 struct vm_fault *vmf,
1302 enum page_entry_size pe_size,
1303 bool write_fault)
1304{
1305 struct inode *inode = file_inode(vmf->vma->vm_file);
1306 struct xfs_inode *ip = XFS_I(inode);
1307 vm_fault_t ret;
1308
1309 trace_xfs_filemap_fault(ip, pe_size, write_fault);
1310
1311 if (write_fault) {
1312 sb_start_pagefault(inode->i_sb);
1313 file_update_time(vmf->vma->vm_file);
1314 }
1315
1316 if (IS_DAX(inode)) {
1317 pfn_t pfn;
1318
1319 xfs_ilock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1320 ret = xfs_dax_fault(vmf, pe_size, write_fault, &pfn);
1321 if (ret & VM_FAULT_NEEDDSYNC)
1322 ret = dax_finish_sync_fault(vmf, pe_size, pfn);
1323 xfs_iunlock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1324 } else {
1325 if (write_fault) {
1326 xfs_ilock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1327 ret = iomap_page_mkwrite(vmf,
1328 &xfs_page_mkwrite_iomap_ops);
1329 xfs_iunlock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1330 } else {
1331 ret = filemap_fault(vmf);
1332 }
1333 }
1334
1335 if (write_fault)
1336 sb_end_pagefault(inode->i_sb);
1337 return ret;
1338}
1339
1340static inline bool
1341xfs_is_write_fault(
1342 struct vm_fault *vmf)
1343{
1344 return (vmf->flags & FAULT_FLAG_WRITE) &&
1345 (vmf->vma->vm_flags & VM_SHARED);
1346}
1347
1348static vm_fault_t
1349xfs_filemap_fault(
1350 struct vm_fault *vmf)
1351{
1352 /* DAX can shortcut the normal fault path on write faults! */
1353 return __xfs_filemap_fault(vmf, PE_SIZE_PTE,
1354 IS_DAX(file_inode(vmf->vma->vm_file)) &&
1355 xfs_is_write_fault(vmf));
1356}
1357
1358static vm_fault_t
1359xfs_filemap_huge_fault(
1360 struct vm_fault *vmf,
1361 enum page_entry_size pe_size)
1362{
1363 if (!IS_DAX(file_inode(vmf->vma->vm_file)))
1364 return VM_FAULT_FALLBACK;
1365
1366 /* DAX can shortcut the normal fault path on write faults! */
1367 return __xfs_filemap_fault(vmf, pe_size,
1368 xfs_is_write_fault(vmf));
1369}
1370
1371static vm_fault_t
1372xfs_filemap_page_mkwrite(
1373 struct vm_fault *vmf)
1374{
1375 return __xfs_filemap_fault(vmf, PE_SIZE_PTE, true);
1376}
1377
1378/*
1379 * pfn_mkwrite was originally intended to ensure we capture time stamp updates
1380 * on write faults. In reality, it needs to serialise against truncate and
1381 * prepare memory for writing so handle is as standard write fault.
1382 */
1383static vm_fault_t
1384xfs_filemap_pfn_mkwrite(
1385 struct vm_fault *vmf)
1386{
1387
1388 return __xfs_filemap_fault(vmf, PE_SIZE_PTE, true);
1389}
1390
1391static vm_fault_t
1392xfs_filemap_map_pages(
1393 struct vm_fault *vmf,
1394 pgoff_t start_pgoff,
1395 pgoff_t end_pgoff)
1396{
1397 struct inode *inode = file_inode(vmf->vma->vm_file);
1398 vm_fault_t ret;
1399
1400 xfs_ilock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1401 ret = filemap_map_pages(vmf, start_pgoff, end_pgoff);
1402 xfs_iunlock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1403 return ret;
1404}
1405
1406static const struct vm_operations_struct xfs_file_vm_ops = {
1407 .fault = xfs_filemap_fault,
1408 .huge_fault = xfs_filemap_huge_fault,
1409 .map_pages = xfs_filemap_map_pages,
1410 .page_mkwrite = xfs_filemap_page_mkwrite,
1411 .pfn_mkwrite = xfs_filemap_pfn_mkwrite,
1412};
1413
1414STATIC int
1415xfs_file_mmap(
1416 struct file *file,
1417 struct vm_area_struct *vma)
1418{
1419 struct inode *inode = file_inode(file);
1420 struct xfs_buftarg *target = xfs_inode_buftarg(XFS_I(inode));
1421
1422 /*
1423 * We don't support synchronous mappings for non-DAX files and
1424 * for DAX files if underneath dax_device is not synchronous.
1425 */
1426 if (!daxdev_mapping_supported(vma, target->bt_daxdev))
1427 return -EOPNOTSUPP;
1428
1429 file_accessed(file);
1430 vma->vm_ops = &xfs_file_vm_ops;
1431 if (IS_DAX(inode))
1432 vma->vm_flags |= VM_HUGEPAGE;
1433 return 0;
1434}
1435
1436const struct file_operations xfs_file_operations = {
1437 .llseek = xfs_file_llseek,
1438 .read_iter = xfs_file_read_iter,
1439 .write_iter = xfs_file_write_iter,
1440 .splice_read = generic_file_splice_read,
1441 .splice_write = iter_file_splice_write,
1442 .iopoll = iocb_bio_iopoll,
1443 .unlocked_ioctl = xfs_file_ioctl,
1444#ifdef CONFIG_COMPAT
1445 .compat_ioctl = xfs_file_compat_ioctl,
1446#endif
1447 .mmap = xfs_file_mmap,
1448 .mmap_supported_flags = MAP_SYNC,
1449 .open = xfs_file_open,
1450 .release = xfs_file_release,
1451 .fsync = xfs_file_fsync,
1452 .get_unmapped_area = thp_get_unmapped_area,
1453 .fallocate = xfs_file_fallocate,
1454 .fadvise = xfs_file_fadvise,
1455 .remap_file_range = xfs_file_remap_range,
1456};
1457
1458const struct file_operations xfs_dir_file_operations = {
1459 .open = xfs_dir_open,
1460 .read = generic_read_dir,
1461 .iterate_shared = xfs_file_readdir,
1462 .llseek = generic_file_llseek,
1463 .unlocked_ioctl = xfs_file_ioctl,
1464#ifdef CONFIG_COMPAT
1465 .compat_ioctl = xfs_file_compat_ioctl,
1466#endif
1467 .fsync = xfs_dir_fsync,
1468};
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Copyright (c) 2000-2005 Silicon Graphics, Inc.
4 * All Rights Reserved.
5 */
6#include "xfs.h"
7#include "xfs_fs.h"
8#include "xfs_shared.h"
9#include "xfs_format.h"
10#include "xfs_log_format.h"
11#include "xfs_trans_resv.h"
12#include "xfs_mount.h"
13#include "xfs_inode.h"
14#include "xfs_trans.h"
15#include "xfs_inode_item.h"
16#include "xfs_bmap.h"
17#include "xfs_bmap_util.h"
18#include "xfs_dir2.h"
19#include "xfs_dir2_priv.h"
20#include "xfs_ioctl.h"
21#include "xfs_trace.h"
22#include "xfs_log.h"
23#include "xfs_icache.h"
24#include "xfs_pnfs.h"
25#include "xfs_iomap.h"
26#include "xfs_reflink.h"
27
28#include <linux/falloc.h>
29#include <linux/backing-dev.h>
30#include <linux/mman.h>
31#include <linux/fadvise.h>
32
33static const struct vm_operations_struct xfs_file_vm_ops;
34
35int
36xfs_update_prealloc_flags(
37 struct xfs_inode *ip,
38 enum xfs_prealloc_flags flags)
39{
40 struct xfs_trans *tp;
41 int error;
42
43 error = xfs_trans_alloc(ip->i_mount, &M_RES(ip->i_mount)->tr_writeid,
44 0, 0, 0, &tp);
45 if (error)
46 return error;
47
48 xfs_ilock(ip, XFS_ILOCK_EXCL);
49 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
50
51 if (!(flags & XFS_PREALLOC_INVISIBLE)) {
52 VFS_I(ip)->i_mode &= ~S_ISUID;
53 if (VFS_I(ip)->i_mode & S_IXGRP)
54 VFS_I(ip)->i_mode &= ~S_ISGID;
55 xfs_trans_ichgtime(tp, ip, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
56 }
57
58 if (flags & XFS_PREALLOC_SET)
59 ip->i_d.di_flags |= XFS_DIFLAG_PREALLOC;
60 if (flags & XFS_PREALLOC_CLEAR)
61 ip->i_d.di_flags &= ~XFS_DIFLAG_PREALLOC;
62
63 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
64 if (flags & XFS_PREALLOC_SYNC)
65 xfs_trans_set_sync(tp);
66 return xfs_trans_commit(tp);
67}
68
69/*
70 * Fsync operations on directories are much simpler than on regular files,
71 * as there is no file data to flush, and thus also no need for explicit
72 * cache flush operations, and there are no non-transaction metadata updates
73 * on directories either.
74 */
75STATIC int
76xfs_dir_fsync(
77 struct file *file,
78 loff_t start,
79 loff_t end,
80 int datasync)
81{
82 struct xfs_inode *ip = XFS_I(file->f_mapping->host);
83
84 trace_xfs_dir_fsync(ip);
85 return xfs_log_force_inode(ip);
86}
87
88STATIC int
89xfs_file_fsync(
90 struct file *file,
91 loff_t start,
92 loff_t end,
93 int datasync)
94{
95 struct inode *inode = file->f_mapping->host;
96 struct xfs_inode *ip = XFS_I(inode);
97 struct xfs_inode_log_item *iip = ip->i_itemp;
98 struct xfs_mount *mp = ip->i_mount;
99 int error = 0;
100 int log_flushed = 0;
101 xfs_lsn_t lsn = 0;
102
103 trace_xfs_file_fsync(ip);
104
105 error = file_write_and_wait_range(file, start, end);
106 if (error)
107 return error;
108
109 if (XFS_FORCED_SHUTDOWN(mp))
110 return -EIO;
111
112 xfs_iflags_clear(ip, XFS_ITRUNCATED);
113
114 /*
115 * If we have an RT and/or log subvolume we need to make sure to flush
116 * the write cache the device used for file data first. This is to
117 * ensure newly written file data make it to disk before logging the new
118 * inode size in case of an extending write.
119 */
120 if (XFS_IS_REALTIME_INODE(ip))
121 xfs_blkdev_issue_flush(mp->m_rtdev_targp);
122 else if (mp->m_logdev_targp != mp->m_ddev_targp)
123 xfs_blkdev_issue_flush(mp->m_ddev_targp);
124
125 /*
126 * All metadata updates are logged, which means that we just have to
127 * flush the log up to the latest LSN that touched the inode. If we have
128 * concurrent fsync/fdatasync() calls, we need them to all block on the
129 * log force before we clear the ili_fsync_fields field. This ensures
130 * that we don't get a racing sync operation that does not wait for the
131 * metadata to hit the journal before returning. If we race with
132 * clearing the ili_fsync_fields, then all that will happen is the log
133 * force will do nothing as the lsn will already be on disk. We can't
134 * race with setting ili_fsync_fields because that is done under
135 * XFS_ILOCK_EXCL, and that can't happen because we hold the lock shared
136 * until after the ili_fsync_fields is cleared.
137 */
138 xfs_ilock(ip, XFS_ILOCK_SHARED);
139 if (xfs_ipincount(ip)) {
140 if (!datasync ||
141 (iip->ili_fsync_fields & ~XFS_ILOG_TIMESTAMP))
142 lsn = iip->ili_last_lsn;
143 }
144
145 if (lsn) {
146 error = xfs_log_force_lsn(mp, lsn, XFS_LOG_SYNC, &log_flushed);
147 spin_lock(&iip->ili_lock);
148 iip->ili_fsync_fields = 0;
149 spin_unlock(&iip->ili_lock);
150 }
151 xfs_iunlock(ip, XFS_ILOCK_SHARED);
152
153 /*
154 * If we only have a single device, and the log force about was
155 * a no-op we might have to flush the data device cache here.
156 * This can only happen for fdatasync/O_DSYNC if we were overwriting
157 * an already allocated file and thus do not have any metadata to
158 * commit.
159 */
160 if (!log_flushed && !XFS_IS_REALTIME_INODE(ip) &&
161 mp->m_logdev_targp == mp->m_ddev_targp)
162 xfs_blkdev_issue_flush(mp->m_ddev_targp);
163
164 return error;
165}
166
167STATIC ssize_t
168xfs_file_dio_aio_read(
169 struct kiocb *iocb,
170 struct iov_iter *to)
171{
172 struct xfs_inode *ip = XFS_I(file_inode(iocb->ki_filp));
173 size_t count = iov_iter_count(to);
174 ssize_t ret;
175
176 trace_xfs_file_direct_read(ip, count, iocb->ki_pos);
177
178 if (!count)
179 return 0; /* skip atime */
180
181 file_accessed(iocb->ki_filp);
182
183 if (iocb->ki_flags & IOCB_NOWAIT) {
184 if (!xfs_ilock_nowait(ip, XFS_IOLOCK_SHARED))
185 return -EAGAIN;
186 } else {
187 xfs_ilock(ip, XFS_IOLOCK_SHARED);
188 }
189 ret = iomap_dio_rw(iocb, to, &xfs_read_iomap_ops, NULL,
190 is_sync_kiocb(iocb));
191 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
192
193 return ret;
194}
195
196static noinline ssize_t
197xfs_file_dax_read(
198 struct kiocb *iocb,
199 struct iov_iter *to)
200{
201 struct xfs_inode *ip = XFS_I(iocb->ki_filp->f_mapping->host);
202 size_t count = iov_iter_count(to);
203 ssize_t ret = 0;
204
205 trace_xfs_file_dax_read(ip, count, iocb->ki_pos);
206
207 if (!count)
208 return 0; /* skip atime */
209
210 if (iocb->ki_flags & IOCB_NOWAIT) {
211 if (!xfs_ilock_nowait(ip, XFS_IOLOCK_SHARED))
212 return -EAGAIN;
213 } else {
214 xfs_ilock(ip, XFS_IOLOCK_SHARED);
215 }
216
217 ret = dax_iomap_rw(iocb, to, &xfs_read_iomap_ops);
218 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
219
220 file_accessed(iocb->ki_filp);
221 return ret;
222}
223
224STATIC ssize_t
225xfs_file_buffered_aio_read(
226 struct kiocb *iocb,
227 struct iov_iter *to)
228{
229 struct xfs_inode *ip = XFS_I(file_inode(iocb->ki_filp));
230 ssize_t ret;
231
232 trace_xfs_file_buffered_read(ip, iov_iter_count(to), iocb->ki_pos);
233
234 if (iocb->ki_flags & IOCB_NOWAIT) {
235 if (!xfs_ilock_nowait(ip, XFS_IOLOCK_SHARED))
236 return -EAGAIN;
237 } else {
238 xfs_ilock(ip, XFS_IOLOCK_SHARED);
239 }
240 ret = generic_file_read_iter(iocb, to);
241 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
242
243 return ret;
244}
245
246STATIC ssize_t
247xfs_file_read_iter(
248 struct kiocb *iocb,
249 struct iov_iter *to)
250{
251 struct inode *inode = file_inode(iocb->ki_filp);
252 struct xfs_mount *mp = XFS_I(inode)->i_mount;
253 ssize_t ret = 0;
254
255 XFS_STATS_INC(mp, xs_read_calls);
256
257 if (XFS_FORCED_SHUTDOWN(mp))
258 return -EIO;
259
260 if (IS_DAX(inode))
261 ret = xfs_file_dax_read(iocb, to);
262 else if (iocb->ki_flags & IOCB_DIRECT)
263 ret = xfs_file_dio_aio_read(iocb, to);
264 else
265 ret = xfs_file_buffered_aio_read(iocb, to);
266
267 if (ret > 0)
268 XFS_STATS_ADD(mp, xs_read_bytes, ret);
269 return ret;
270}
271
272/*
273 * Common pre-write limit and setup checks.
274 *
275 * Called with the iolocked held either shared and exclusive according to
276 * @iolock, and returns with it held. Might upgrade the iolock to exclusive
277 * if called for a direct write beyond i_size.
278 */
279STATIC ssize_t
280xfs_file_aio_write_checks(
281 struct kiocb *iocb,
282 struct iov_iter *from,
283 int *iolock)
284{
285 struct file *file = iocb->ki_filp;
286 struct inode *inode = file->f_mapping->host;
287 struct xfs_inode *ip = XFS_I(inode);
288 ssize_t error = 0;
289 size_t count = iov_iter_count(from);
290 bool drained_dio = false;
291 loff_t isize;
292
293restart:
294 error = generic_write_checks(iocb, from);
295 if (error <= 0)
296 return error;
297
298 error = xfs_break_layouts(inode, iolock, BREAK_WRITE);
299 if (error)
300 return error;
301
302 /*
303 * For changing security info in file_remove_privs() we need i_rwsem
304 * exclusively.
305 */
306 if (*iolock == XFS_IOLOCK_SHARED && !IS_NOSEC(inode)) {
307 xfs_iunlock(ip, *iolock);
308 *iolock = XFS_IOLOCK_EXCL;
309 xfs_ilock(ip, *iolock);
310 goto restart;
311 }
312 /*
313 * If the offset is beyond the size of the file, we need to zero any
314 * blocks that fall between the existing EOF and the start of this
315 * write. If zeroing is needed and we are currently holding the
316 * iolock shared, we need to update it to exclusive which implies
317 * having to redo all checks before.
318 *
319 * We need to serialise against EOF updates that occur in IO
320 * completions here. We want to make sure that nobody is changing the
321 * size while we do this check until we have placed an IO barrier (i.e.
322 * hold the XFS_IOLOCK_EXCL) that prevents new IO from being dispatched.
323 * The spinlock effectively forms a memory barrier once we have the
324 * XFS_IOLOCK_EXCL so we are guaranteed to see the latest EOF value
325 * and hence be able to correctly determine if we need to run zeroing.
326 */
327 spin_lock(&ip->i_flags_lock);
328 isize = i_size_read(inode);
329 if (iocb->ki_pos > isize) {
330 spin_unlock(&ip->i_flags_lock);
331 if (!drained_dio) {
332 if (*iolock == XFS_IOLOCK_SHARED) {
333 xfs_iunlock(ip, *iolock);
334 *iolock = XFS_IOLOCK_EXCL;
335 xfs_ilock(ip, *iolock);
336 iov_iter_reexpand(from, count);
337 }
338 /*
339 * We now have an IO submission barrier in place, but
340 * AIO can do EOF updates during IO completion and hence
341 * we now need to wait for all of them to drain. Non-AIO
342 * DIO will have drained before we are given the
343 * XFS_IOLOCK_EXCL, and so for most cases this wait is a
344 * no-op.
345 */
346 inode_dio_wait(inode);
347 drained_dio = true;
348 goto restart;
349 }
350
351 trace_xfs_zero_eof(ip, isize, iocb->ki_pos - isize);
352 error = iomap_zero_range(inode, isize, iocb->ki_pos - isize,
353 NULL, &xfs_buffered_write_iomap_ops);
354 if (error)
355 return error;
356 } else
357 spin_unlock(&ip->i_flags_lock);
358
359 /*
360 * Updating the timestamps will grab the ilock again from
361 * xfs_fs_dirty_inode, so we have to call it after dropping the
362 * lock above. Eventually we should look into a way to avoid
363 * the pointless lock roundtrip.
364 */
365 return file_modified(file);
366}
367
368static int
369xfs_dio_write_end_io(
370 struct kiocb *iocb,
371 ssize_t size,
372 int error,
373 unsigned flags)
374{
375 struct inode *inode = file_inode(iocb->ki_filp);
376 struct xfs_inode *ip = XFS_I(inode);
377 loff_t offset = iocb->ki_pos;
378 unsigned int nofs_flag;
379
380 trace_xfs_end_io_direct_write(ip, offset, size);
381
382 if (XFS_FORCED_SHUTDOWN(ip->i_mount))
383 return -EIO;
384
385 if (error)
386 return error;
387 if (!size)
388 return 0;
389
390 /*
391 * Capture amount written on completion as we can't reliably account
392 * for it on submission.
393 */
394 XFS_STATS_ADD(ip->i_mount, xs_write_bytes, size);
395
396 /*
397 * We can allocate memory here while doing writeback on behalf of
398 * memory reclaim. To avoid memory allocation deadlocks set the
399 * task-wide nofs context for the following operations.
400 */
401 nofs_flag = memalloc_nofs_save();
402
403 if (flags & IOMAP_DIO_COW) {
404 error = xfs_reflink_end_cow(ip, offset, size);
405 if (error)
406 goto out;
407 }
408
409 /*
410 * Unwritten conversion updates the in-core isize after extent
411 * conversion but before updating the on-disk size. Updating isize any
412 * earlier allows a racing dio read to find unwritten extents before
413 * they are converted.
414 */
415 if (flags & IOMAP_DIO_UNWRITTEN) {
416 error = xfs_iomap_write_unwritten(ip, offset, size, true);
417 goto out;
418 }
419
420 /*
421 * We need to update the in-core inode size here so that we don't end up
422 * with the on-disk inode size being outside the in-core inode size. We
423 * have no other method of updating EOF for AIO, so always do it here
424 * if necessary.
425 *
426 * We need to lock the test/set EOF update as we can be racing with
427 * other IO completions here to update the EOF. Failing to serialise
428 * here can result in EOF moving backwards and Bad Things Happen when
429 * that occurs.
430 */
431 spin_lock(&ip->i_flags_lock);
432 if (offset + size > i_size_read(inode)) {
433 i_size_write(inode, offset + size);
434 spin_unlock(&ip->i_flags_lock);
435 error = xfs_setfilesize(ip, offset, size);
436 } else {
437 spin_unlock(&ip->i_flags_lock);
438 }
439
440out:
441 memalloc_nofs_restore(nofs_flag);
442 return error;
443}
444
445static const struct iomap_dio_ops xfs_dio_write_ops = {
446 .end_io = xfs_dio_write_end_io,
447};
448
449/*
450 * xfs_file_dio_aio_write - handle direct IO writes
451 *
452 * Lock the inode appropriately to prepare for and issue a direct IO write.
453 * By separating it from the buffered write path we remove all the tricky to
454 * follow locking changes and looping.
455 *
456 * If there are cached pages or we're extending the file, we need IOLOCK_EXCL
457 * until we're sure the bytes at the new EOF have been zeroed and/or the cached
458 * pages are flushed out.
459 *
460 * In most cases the direct IO writes will be done holding IOLOCK_SHARED
461 * allowing them to be done in parallel with reads and other direct IO writes.
462 * However, if the IO is not aligned to filesystem blocks, the direct IO layer
463 * needs to do sub-block zeroing and that requires serialisation against other
464 * direct IOs to the same block. In this case we need to serialise the
465 * submission of the unaligned IOs so that we don't get racing block zeroing in
466 * the dio layer. To avoid the problem with aio, we also need to wait for
467 * outstanding IOs to complete so that unwritten extent conversion is completed
468 * before we try to map the overlapping block. This is currently implemented by
469 * hitting it with a big hammer (i.e. inode_dio_wait()).
470 *
471 * Returns with locks held indicated by @iolock and errors indicated by
472 * negative return values.
473 */
474STATIC ssize_t
475xfs_file_dio_aio_write(
476 struct kiocb *iocb,
477 struct iov_iter *from)
478{
479 struct file *file = iocb->ki_filp;
480 struct address_space *mapping = file->f_mapping;
481 struct inode *inode = mapping->host;
482 struct xfs_inode *ip = XFS_I(inode);
483 struct xfs_mount *mp = ip->i_mount;
484 ssize_t ret = 0;
485 int unaligned_io = 0;
486 int iolock;
487 size_t count = iov_iter_count(from);
488 struct xfs_buftarg *target = xfs_inode_buftarg(ip);
489
490 /* DIO must be aligned to device logical sector size */
491 if ((iocb->ki_pos | count) & target->bt_logical_sectormask)
492 return -EINVAL;
493
494 /*
495 * Don't take the exclusive iolock here unless the I/O is unaligned to
496 * the file system block size. We don't need to consider the EOF
497 * extension case here because xfs_file_aio_write_checks() will relock
498 * the inode as necessary for EOF zeroing cases and fill out the new
499 * inode size as appropriate.
500 */
501 if ((iocb->ki_pos & mp->m_blockmask) ||
502 ((iocb->ki_pos + count) & mp->m_blockmask)) {
503 unaligned_io = 1;
504
505 /*
506 * We can't properly handle unaligned direct I/O to reflink
507 * files yet, as we can't unshare a partial block.
508 */
509 if (xfs_is_cow_inode(ip)) {
510 trace_xfs_reflink_bounce_dio_write(ip, iocb->ki_pos, count);
511 return -ENOTBLK;
512 }
513 iolock = XFS_IOLOCK_EXCL;
514 } else {
515 iolock = XFS_IOLOCK_SHARED;
516 }
517
518 if (iocb->ki_flags & IOCB_NOWAIT) {
519 /* unaligned dio always waits, bail */
520 if (unaligned_io)
521 return -EAGAIN;
522 if (!xfs_ilock_nowait(ip, iolock))
523 return -EAGAIN;
524 } else {
525 xfs_ilock(ip, iolock);
526 }
527
528 ret = xfs_file_aio_write_checks(iocb, from, &iolock);
529 if (ret)
530 goto out;
531 count = iov_iter_count(from);
532
533 /*
534 * If we are doing unaligned IO, we can't allow any other overlapping IO
535 * in-flight at the same time or we risk data corruption. Wait for all
536 * other IO to drain before we submit. If the IO is aligned, demote the
537 * iolock if we had to take the exclusive lock in
538 * xfs_file_aio_write_checks() for other reasons.
539 */
540 if (unaligned_io) {
541 inode_dio_wait(inode);
542 } else if (iolock == XFS_IOLOCK_EXCL) {
543 xfs_ilock_demote(ip, XFS_IOLOCK_EXCL);
544 iolock = XFS_IOLOCK_SHARED;
545 }
546
547 trace_xfs_file_direct_write(ip, count, iocb->ki_pos);
548 /*
549 * If unaligned, this is the only IO in-flight. Wait on it before we
550 * release the iolock to prevent subsequent overlapping IO.
551 */
552 ret = iomap_dio_rw(iocb, from, &xfs_direct_write_iomap_ops,
553 &xfs_dio_write_ops,
554 is_sync_kiocb(iocb) || unaligned_io);
555out:
556 xfs_iunlock(ip, iolock);
557
558 /*
559 * No fallback to buffered IO after short writes for XFS, direct I/O
560 * will either complete fully or return an error.
561 */
562 ASSERT(ret < 0 || ret == count);
563 return ret;
564}
565
566static noinline ssize_t
567xfs_file_dax_write(
568 struct kiocb *iocb,
569 struct iov_iter *from)
570{
571 struct inode *inode = iocb->ki_filp->f_mapping->host;
572 struct xfs_inode *ip = XFS_I(inode);
573 int iolock = XFS_IOLOCK_EXCL;
574 ssize_t ret, error = 0;
575 size_t count;
576 loff_t pos;
577
578 if (iocb->ki_flags & IOCB_NOWAIT) {
579 if (!xfs_ilock_nowait(ip, iolock))
580 return -EAGAIN;
581 } else {
582 xfs_ilock(ip, iolock);
583 }
584
585 ret = xfs_file_aio_write_checks(iocb, from, &iolock);
586 if (ret)
587 goto out;
588
589 pos = iocb->ki_pos;
590 count = iov_iter_count(from);
591
592 trace_xfs_file_dax_write(ip, count, pos);
593 ret = dax_iomap_rw(iocb, from, &xfs_direct_write_iomap_ops);
594 if (ret > 0 && iocb->ki_pos > i_size_read(inode)) {
595 i_size_write(inode, iocb->ki_pos);
596 error = xfs_setfilesize(ip, pos, ret);
597 }
598out:
599 xfs_iunlock(ip, iolock);
600 if (error)
601 return error;
602
603 if (ret > 0) {
604 XFS_STATS_ADD(ip->i_mount, xs_write_bytes, ret);
605
606 /* Handle various SYNC-type writes */
607 ret = generic_write_sync(iocb, ret);
608 }
609 return ret;
610}
611
612STATIC ssize_t
613xfs_file_buffered_aio_write(
614 struct kiocb *iocb,
615 struct iov_iter *from)
616{
617 struct file *file = iocb->ki_filp;
618 struct address_space *mapping = file->f_mapping;
619 struct inode *inode = mapping->host;
620 struct xfs_inode *ip = XFS_I(inode);
621 ssize_t ret;
622 int enospc = 0;
623 int iolock;
624
625 if (iocb->ki_flags & IOCB_NOWAIT)
626 return -EOPNOTSUPP;
627
628write_retry:
629 iolock = XFS_IOLOCK_EXCL;
630 xfs_ilock(ip, iolock);
631
632 ret = xfs_file_aio_write_checks(iocb, from, &iolock);
633 if (ret)
634 goto out;
635
636 /* We can write back this queue in page reclaim */
637 current->backing_dev_info = inode_to_bdi(inode);
638
639 trace_xfs_file_buffered_write(ip, iov_iter_count(from), iocb->ki_pos);
640 ret = iomap_file_buffered_write(iocb, from,
641 &xfs_buffered_write_iomap_ops);
642 if (likely(ret >= 0))
643 iocb->ki_pos += ret;
644
645 /*
646 * If we hit a space limit, try to free up some lingering preallocated
647 * space before returning an error. In the case of ENOSPC, first try to
648 * write back all dirty inodes to free up some of the excess reserved
649 * metadata space. This reduces the chances that the eofblocks scan
650 * waits on dirty mappings. Since xfs_flush_inodes() is serialized, this
651 * also behaves as a filter to prevent too many eofblocks scans from
652 * running at the same time.
653 */
654 if (ret == -EDQUOT && !enospc) {
655 xfs_iunlock(ip, iolock);
656 enospc = xfs_inode_free_quota_eofblocks(ip);
657 if (enospc)
658 goto write_retry;
659 enospc = xfs_inode_free_quota_cowblocks(ip);
660 if (enospc)
661 goto write_retry;
662 iolock = 0;
663 } else if (ret == -ENOSPC && !enospc) {
664 struct xfs_eofblocks eofb = {0};
665
666 enospc = 1;
667 xfs_flush_inodes(ip->i_mount);
668
669 xfs_iunlock(ip, iolock);
670 eofb.eof_flags = XFS_EOF_FLAGS_SYNC;
671 xfs_icache_free_eofblocks(ip->i_mount, &eofb);
672 xfs_icache_free_cowblocks(ip->i_mount, &eofb);
673 goto write_retry;
674 }
675
676 current->backing_dev_info = NULL;
677out:
678 if (iolock)
679 xfs_iunlock(ip, iolock);
680
681 if (ret > 0) {
682 XFS_STATS_ADD(ip->i_mount, xs_write_bytes, ret);
683 /* Handle various SYNC-type writes */
684 ret = generic_write_sync(iocb, ret);
685 }
686 return ret;
687}
688
689STATIC ssize_t
690xfs_file_write_iter(
691 struct kiocb *iocb,
692 struct iov_iter *from)
693{
694 struct file *file = iocb->ki_filp;
695 struct address_space *mapping = file->f_mapping;
696 struct inode *inode = mapping->host;
697 struct xfs_inode *ip = XFS_I(inode);
698 ssize_t ret;
699 size_t ocount = iov_iter_count(from);
700
701 XFS_STATS_INC(ip->i_mount, xs_write_calls);
702
703 if (ocount == 0)
704 return 0;
705
706 if (XFS_FORCED_SHUTDOWN(ip->i_mount))
707 return -EIO;
708
709 if (IS_DAX(inode))
710 return xfs_file_dax_write(iocb, from);
711
712 if (iocb->ki_flags & IOCB_DIRECT) {
713 /*
714 * Allow a directio write to fall back to a buffered
715 * write *only* in the case that we're doing a reflink
716 * CoW. In all other directio scenarios we do not
717 * allow an operation to fall back to buffered mode.
718 */
719 ret = xfs_file_dio_aio_write(iocb, from);
720 if (ret != -ENOTBLK)
721 return ret;
722 }
723
724 return xfs_file_buffered_aio_write(iocb, from);
725}
726
727static void
728xfs_wait_dax_page(
729 struct inode *inode)
730{
731 struct xfs_inode *ip = XFS_I(inode);
732
733 xfs_iunlock(ip, XFS_MMAPLOCK_EXCL);
734 schedule();
735 xfs_ilock(ip, XFS_MMAPLOCK_EXCL);
736}
737
738static int
739xfs_break_dax_layouts(
740 struct inode *inode,
741 bool *retry)
742{
743 struct page *page;
744
745 ASSERT(xfs_isilocked(XFS_I(inode), XFS_MMAPLOCK_EXCL));
746
747 page = dax_layout_busy_page(inode->i_mapping);
748 if (!page)
749 return 0;
750
751 *retry = true;
752 return ___wait_var_event(&page->_refcount,
753 atomic_read(&page->_refcount) == 1, TASK_INTERRUPTIBLE,
754 0, 0, xfs_wait_dax_page(inode));
755}
756
757int
758xfs_break_layouts(
759 struct inode *inode,
760 uint *iolock,
761 enum layout_break_reason reason)
762{
763 bool retry;
764 int error;
765
766 ASSERT(xfs_isilocked(XFS_I(inode), XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL));
767
768 do {
769 retry = false;
770 switch (reason) {
771 case BREAK_UNMAP:
772 error = xfs_break_dax_layouts(inode, &retry);
773 if (error || retry)
774 break;
775 /* fall through */
776 case BREAK_WRITE:
777 error = xfs_break_leased_layouts(inode, iolock, &retry);
778 break;
779 default:
780 WARN_ON_ONCE(1);
781 error = -EINVAL;
782 }
783 } while (error == 0 && retry);
784
785 return error;
786}
787
788#define XFS_FALLOC_FL_SUPPORTED \
789 (FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE | \
790 FALLOC_FL_COLLAPSE_RANGE | FALLOC_FL_ZERO_RANGE | \
791 FALLOC_FL_INSERT_RANGE | FALLOC_FL_UNSHARE_RANGE)
792
793STATIC long
794xfs_file_fallocate(
795 struct file *file,
796 int mode,
797 loff_t offset,
798 loff_t len)
799{
800 struct inode *inode = file_inode(file);
801 struct xfs_inode *ip = XFS_I(inode);
802 long error;
803 enum xfs_prealloc_flags flags = 0;
804 uint iolock = XFS_IOLOCK_EXCL | XFS_MMAPLOCK_EXCL;
805 loff_t new_size = 0;
806 bool do_file_insert = false;
807
808 if (!S_ISREG(inode->i_mode))
809 return -EINVAL;
810 if (mode & ~XFS_FALLOC_FL_SUPPORTED)
811 return -EOPNOTSUPP;
812
813 xfs_ilock(ip, iolock);
814 error = xfs_break_layouts(inode, &iolock, BREAK_UNMAP);
815 if (error)
816 goto out_unlock;
817
818 /*
819 * Must wait for all AIO to complete before we continue as AIO can
820 * change the file size on completion without holding any locks we
821 * currently hold. We must do this first because AIO can update both
822 * the on disk and in memory inode sizes, and the operations that follow
823 * require the in-memory size to be fully up-to-date.
824 */
825 inode_dio_wait(inode);
826
827 /*
828 * Now AIO and DIO has drained we flush and (if necessary) invalidate
829 * the cached range over the first operation we are about to run.
830 *
831 * We care about zero and collapse here because they both run a hole
832 * punch over the range first. Because that can zero data, and the range
833 * of invalidation for the shift operations is much larger, we still do
834 * the required flush for collapse in xfs_prepare_shift().
835 *
836 * Insert has the same range requirements as collapse, and we extend the
837 * file first which can zero data. Hence insert has the same
838 * flush/invalidate requirements as collapse and so they are both
839 * handled at the right time by xfs_prepare_shift().
840 */
841 if (mode & (FALLOC_FL_PUNCH_HOLE | FALLOC_FL_ZERO_RANGE |
842 FALLOC_FL_COLLAPSE_RANGE)) {
843 error = xfs_flush_unmap_range(ip, offset, len);
844 if (error)
845 goto out_unlock;
846 }
847
848 if (mode & FALLOC_FL_PUNCH_HOLE) {
849 error = xfs_free_file_space(ip, offset, len);
850 if (error)
851 goto out_unlock;
852 } else if (mode & FALLOC_FL_COLLAPSE_RANGE) {
853 unsigned int blksize_mask = i_blocksize(inode) - 1;
854
855 if (offset & blksize_mask || len & blksize_mask) {
856 error = -EINVAL;
857 goto out_unlock;
858 }
859
860 /*
861 * There is no need to overlap collapse range with EOF,
862 * in which case it is effectively a truncate operation
863 */
864 if (offset + len >= i_size_read(inode)) {
865 error = -EINVAL;
866 goto out_unlock;
867 }
868
869 new_size = i_size_read(inode) - len;
870
871 error = xfs_collapse_file_space(ip, offset, len);
872 if (error)
873 goto out_unlock;
874 } else if (mode & FALLOC_FL_INSERT_RANGE) {
875 unsigned int blksize_mask = i_blocksize(inode) - 1;
876 loff_t isize = i_size_read(inode);
877
878 if (offset & blksize_mask || len & blksize_mask) {
879 error = -EINVAL;
880 goto out_unlock;
881 }
882
883 /*
884 * New inode size must not exceed ->s_maxbytes, accounting for
885 * possible signed overflow.
886 */
887 if (inode->i_sb->s_maxbytes - isize < len) {
888 error = -EFBIG;
889 goto out_unlock;
890 }
891 new_size = isize + len;
892
893 /* Offset should be less than i_size */
894 if (offset >= isize) {
895 error = -EINVAL;
896 goto out_unlock;
897 }
898 do_file_insert = true;
899 } else {
900 flags |= XFS_PREALLOC_SET;
901
902 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
903 offset + len > i_size_read(inode)) {
904 new_size = offset + len;
905 error = inode_newsize_ok(inode, new_size);
906 if (error)
907 goto out_unlock;
908 }
909
910 if (mode & FALLOC_FL_ZERO_RANGE) {
911 /*
912 * Punch a hole and prealloc the range. We use a hole
913 * punch rather than unwritten extent conversion for two
914 * reasons:
915 *
916 * 1.) Hole punch handles partial block zeroing for us.
917 * 2.) If prealloc returns ENOSPC, the file range is
918 * still zero-valued by virtue of the hole punch.
919 */
920 unsigned int blksize = i_blocksize(inode);
921
922 trace_xfs_zero_file_space(ip);
923
924 error = xfs_free_file_space(ip, offset, len);
925 if (error)
926 goto out_unlock;
927
928 len = round_up(offset + len, blksize) -
929 round_down(offset, blksize);
930 offset = round_down(offset, blksize);
931 } else if (mode & FALLOC_FL_UNSHARE_RANGE) {
932 error = xfs_reflink_unshare(ip, offset, len);
933 if (error)
934 goto out_unlock;
935 } else {
936 /*
937 * If always_cow mode we can't use preallocations and
938 * thus should not create them.
939 */
940 if (xfs_is_always_cow_inode(ip)) {
941 error = -EOPNOTSUPP;
942 goto out_unlock;
943 }
944 }
945
946 if (!xfs_is_always_cow_inode(ip)) {
947 error = xfs_alloc_file_space(ip, offset, len,
948 XFS_BMAPI_PREALLOC);
949 if (error)
950 goto out_unlock;
951 }
952 }
953
954 if (file->f_flags & O_DSYNC)
955 flags |= XFS_PREALLOC_SYNC;
956
957 error = xfs_update_prealloc_flags(ip, flags);
958 if (error)
959 goto out_unlock;
960
961 /* Change file size if needed */
962 if (new_size) {
963 struct iattr iattr;
964
965 iattr.ia_valid = ATTR_SIZE;
966 iattr.ia_size = new_size;
967 error = xfs_vn_setattr_size(file_dentry(file), &iattr);
968 if (error)
969 goto out_unlock;
970 }
971
972 /*
973 * Perform hole insertion now that the file size has been
974 * updated so that if we crash during the operation we don't
975 * leave shifted extents past EOF and hence losing access to
976 * the data that is contained within them.
977 */
978 if (do_file_insert)
979 error = xfs_insert_file_space(ip, offset, len);
980
981out_unlock:
982 xfs_iunlock(ip, iolock);
983 return error;
984}
985
986STATIC int
987xfs_file_fadvise(
988 struct file *file,
989 loff_t start,
990 loff_t end,
991 int advice)
992{
993 struct xfs_inode *ip = XFS_I(file_inode(file));
994 int ret;
995 int lockflags = 0;
996
997 /*
998 * Operations creating pages in page cache need protection from hole
999 * punching and similar ops
1000 */
1001 if (advice == POSIX_FADV_WILLNEED) {
1002 lockflags = XFS_IOLOCK_SHARED;
1003 xfs_ilock(ip, lockflags);
1004 }
1005 ret = generic_fadvise(file, start, end, advice);
1006 if (lockflags)
1007 xfs_iunlock(ip, lockflags);
1008 return ret;
1009}
1010
1011STATIC loff_t
1012xfs_file_remap_range(
1013 struct file *file_in,
1014 loff_t pos_in,
1015 struct file *file_out,
1016 loff_t pos_out,
1017 loff_t len,
1018 unsigned int remap_flags)
1019{
1020 struct inode *inode_in = file_inode(file_in);
1021 struct xfs_inode *src = XFS_I(inode_in);
1022 struct inode *inode_out = file_inode(file_out);
1023 struct xfs_inode *dest = XFS_I(inode_out);
1024 struct xfs_mount *mp = src->i_mount;
1025 loff_t remapped = 0;
1026 xfs_extlen_t cowextsize;
1027 int ret;
1028
1029 if (remap_flags & ~(REMAP_FILE_DEDUP | REMAP_FILE_ADVISORY))
1030 return -EINVAL;
1031
1032 if (!xfs_sb_version_hasreflink(&mp->m_sb))
1033 return -EOPNOTSUPP;
1034
1035 if (XFS_FORCED_SHUTDOWN(mp))
1036 return -EIO;
1037
1038 /* Prepare and then clone file data. */
1039 ret = xfs_reflink_remap_prep(file_in, pos_in, file_out, pos_out,
1040 &len, remap_flags);
1041 if (ret || len == 0)
1042 return ret;
1043
1044 trace_xfs_reflink_remap_range(src, pos_in, len, dest, pos_out);
1045
1046 ret = xfs_reflink_remap_blocks(src, pos_in, dest, pos_out, len,
1047 &remapped);
1048 if (ret)
1049 goto out_unlock;
1050
1051 /*
1052 * Carry the cowextsize hint from src to dest if we're sharing the
1053 * entire source file to the entire destination file, the source file
1054 * has a cowextsize hint, and the destination file does not.
1055 */
1056 cowextsize = 0;
1057 if (pos_in == 0 && len == i_size_read(inode_in) &&
1058 (src->i_d.di_flags2 & XFS_DIFLAG2_COWEXTSIZE) &&
1059 pos_out == 0 && len >= i_size_read(inode_out) &&
1060 !(dest->i_d.di_flags2 & XFS_DIFLAG2_COWEXTSIZE))
1061 cowextsize = src->i_d.di_cowextsize;
1062
1063 ret = xfs_reflink_update_dest(dest, pos_out + len, cowextsize,
1064 remap_flags);
1065 if (ret)
1066 goto out_unlock;
1067
1068 if (mp->m_flags & XFS_MOUNT_WSYNC)
1069 xfs_log_force_inode(dest);
1070out_unlock:
1071 xfs_iunlock2_io_mmap(src, dest);
1072 if (ret)
1073 trace_xfs_reflink_remap_range_error(dest, ret, _RET_IP_);
1074 return remapped > 0 ? remapped : ret;
1075}
1076
1077STATIC int
1078xfs_file_open(
1079 struct inode *inode,
1080 struct file *file)
1081{
1082 if (!(file->f_flags & O_LARGEFILE) && i_size_read(inode) > MAX_NON_LFS)
1083 return -EFBIG;
1084 if (XFS_FORCED_SHUTDOWN(XFS_M(inode->i_sb)))
1085 return -EIO;
1086 file->f_mode |= FMODE_NOWAIT | FMODE_BUF_RASYNC;
1087 return 0;
1088}
1089
1090STATIC int
1091xfs_dir_open(
1092 struct inode *inode,
1093 struct file *file)
1094{
1095 struct xfs_inode *ip = XFS_I(inode);
1096 int mode;
1097 int error;
1098
1099 error = xfs_file_open(inode, file);
1100 if (error)
1101 return error;
1102
1103 /*
1104 * If there are any blocks, read-ahead block 0 as we're almost
1105 * certain to have the next operation be a read there.
1106 */
1107 mode = xfs_ilock_data_map_shared(ip);
1108 if (ip->i_df.if_nextents > 0)
1109 error = xfs_dir3_data_readahead(ip, 0, 0);
1110 xfs_iunlock(ip, mode);
1111 return error;
1112}
1113
1114STATIC int
1115xfs_file_release(
1116 struct inode *inode,
1117 struct file *filp)
1118{
1119 return xfs_release(XFS_I(inode));
1120}
1121
1122STATIC int
1123xfs_file_readdir(
1124 struct file *file,
1125 struct dir_context *ctx)
1126{
1127 struct inode *inode = file_inode(file);
1128 xfs_inode_t *ip = XFS_I(inode);
1129 size_t bufsize;
1130
1131 /*
1132 * The Linux API doesn't pass down the total size of the buffer
1133 * we read into down to the filesystem. With the filldir concept
1134 * it's not needed for correct information, but the XFS dir2 leaf
1135 * code wants an estimate of the buffer size to calculate it's
1136 * readahead window and size the buffers used for mapping to
1137 * physical blocks.
1138 *
1139 * Try to give it an estimate that's good enough, maybe at some
1140 * point we can change the ->readdir prototype to include the
1141 * buffer size. For now we use the current glibc buffer size.
1142 */
1143 bufsize = (size_t)min_t(loff_t, XFS_READDIR_BUFSIZE, ip->i_d.di_size);
1144
1145 return xfs_readdir(NULL, ip, ctx, bufsize);
1146}
1147
1148STATIC loff_t
1149xfs_file_llseek(
1150 struct file *file,
1151 loff_t offset,
1152 int whence)
1153{
1154 struct inode *inode = file->f_mapping->host;
1155
1156 if (XFS_FORCED_SHUTDOWN(XFS_I(inode)->i_mount))
1157 return -EIO;
1158
1159 switch (whence) {
1160 default:
1161 return generic_file_llseek(file, offset, whence);
1162 case SEEK_HOLE:
1163 offset = iomap_seek_hole(inode, offset, &xfs_seek_iomap_ops);
1164 break;
1165 case SEEK_DATA:
1166 offset = iomap_seek_data(inode, offset, &xfs_seek_iomap_ops);
1167 break;
1168 }
1169
1170 if (offset < 0)
1171 return offset;
1172 return vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
1173}
1174
1175/*
1176 * Locking for serialisation of IO during page faults. This results in a lock
1177 * ordering of:
1178 *
1179 * mmap_lock (MM)
1180 * sb_start_pagefault(vfs, freeze)
1181 * i_mmaplock (XFS - truncate serialisation)
1182 * page_lock (MM)
1183 * i_lock (XFS - extent map serialisation)
1184 */
1185static vm_fault_t
1186__xfs_filemap_fault(
1187 struct vm_fault *vmf,
1188 enum page_entry_size pe_size,
1189 bool write_fault)
1190{
1191 struct inode *inode = file_inode(vmf->vma->vm_file);
1192 struct xfs_inode *ip = XFS_I(inode);
1193 vm_fault_t ret;
1194
1195 trace_xfs_filemap_fault(ip, pe_size, write_fault);
1196
1197 if (write_fault) {
1198 sb_start_pagefault(inode->i_sb);
1199 file_update_time(vmf->vma->vm_file);
1200 }
1201
1202 xfs_ilock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1203 if (IS_DAX(inode)) {
1204 pfn_t pfn;
1205
1206 ret = dax_iomap_fault(vmf, pe_size, &pfn, NULL,
1207 (write_fault && !vmf->cow_page) ?
1208 &xfs_direct_write_iomap_ops :
1209 &xfs_read_iomap_ops);
1210 if (ret & VM_FAULT_NEEDDSYNC)
1211 ret = dax_finish_sync_fault(vmf, pe_size, pfn);
1212 } else {
1213 if (write_fault)
1214 ret = iomap_page_mkwrite(vmf,
1215 &xfs_buffered_write_iomap_ops);
1216 else
1217 ret = filemap_fault(vmf);
1218 }
1219 xfs_iunlock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1220
1221 if (write_fault)
1222 sb_end_pagefault(inode->i_sb);
1223 return ret;
1224}
1225
1226static inline bool
1227xfs_is_write_fault(
1228 struct vm_fault *vmf)
1229{
1230 return (vmf->flags & FAULT_FLAG_WRITE) &&
1231 (vmf->vma->vm_flags & VM_SHARED);
1232}
1233
1234static vm_fault_t
1235xfs_filemap_fault(
1236 struct vm_fault *vmf)
1237{
1238 /* DAX can shortcut the normal fault path on write faults! */
1239 return __xfs_filemap_fault(vmf, PE_SIZE_PTE,
1240 IS_DAX(file_inode(vmf->vma->vm_file)) &&
1241 xfs_is_write_fault(vmf));
1242}
1243
1244static vm_fault_t
1245xfs_filemap_huge_fault(
1246 struct vm_fault *vmf,
1247 enum page_entry_size pe_size)
1248{
1249 if (!IS_DAX(file_inode(vmf->vma->vm_file)))
1250 return VM_FAULT_FALLBACK;
1251
1252 /* DAX can shortcut the normal fault path on write faults! */
1253 return __xfs_filemap_fault(vmf, pe_size,
1254 xfs_is_write_fault(vmf));
1255}
1256
1257static vm_fault_t
1258xfs_filemap_page_mkwrite(
1259 struct vm_fault *vmf)
1260{
1261 return __xfs_filemap_fault(vmf, PE_SIZE_PTE, true);
1262}
1263
1264/*
1265 * pfn_mkwrite was originally intended to ensure we capture time stamp updates
1266 * on write faults. In reality, it needs to serialise against truncate and
1267 * prepare memory for writing so handle is as standard write fault.
1268 */
1269static vm_fault_t
1270xfs_filemap_pfn_mkwrite(
1271 struct vm_fault *vmf)
1272{
1273
1274 return __xfs_filemap_fault(vmf, PE_SIZE_PTE, true);
1275}
1276
1277static void
1278xfs_filemap_map_pages(
1279 struct vm_fault *vmf,
1280 pgoff_t start_pgoff,
1281 pgoff_t end_pgoff)
1282{
1283 struct inode *inode = file_inode(vmf->vma->vm_file);
1284
1285 xfs_ilock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1286 filemap_map_pages(vmf, start_pgoff, end_pgoff);
1287 xfs_iunlock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1288}
1289
1290static const struct vm_operations_struct xfs_file_vm_ops = {
1291 .fault = xfs_filemap_fault,
1292 .huge_fault = xfs_filemap_huge_fault,
1293 .map_pages = xfs_filemap_map_pages,
1294 .page_mkwrite = xfs_filemap_page_mkwrite,
1295 .pfn_mkwrite = xfs_filemap_pfn_mkwrite,
1296};
1297
1298STATIC int
1299xfs_file_mmap(
1300 struct file *file,
1301 struct vm_area_struct *vma)
1302{
1303 struct inode *inode = file_inode(file);
1304 struct xfs_buftarg *target = xfs_inode_buftarg(XFS_I(inode));
1305
1306 /*
1307 * We don't support synchronous mappings for non-DAX files and
1308 * for DAX files if underneath dax_device is not synchronous.
1309 */
1310 if (!daxdev_mapping_supported(vma, target->bt_daxdev))
1311 return -EOPNOTSUPP;
1312
1313 file_accessed(file);
1314 vma->vm_ops = &xfs_file_vm_ops;
1315 if (IS_DAX(inode))
1316 vma->vm_flags |= VM_HUGEPAGE;
1317 return 0;
1318}
1319
1320const struct file_operations xfs_file_operations = {
1321 .llseek = xfs_file_llseek,
1322 .read_iter = xfs_file_read_iter,
1323 .write_iter = xfs_file_write_iter,
1324 .splice_read = generic_file_splice_read,
1325 .splice_write = iter_file_splice_write,
1326 .iopoll = iomap_dio_iopoll,
1327 .unlocked_ioctl = xfs_file_ioctl,
1328#ifdef CONFIG_COMPAT
1329 .compat_ioctl = xfs_file_compat_ioctl,
1330#endif
1331 .mmap = xfs_file_mmap,
1332 .mmap_supported_flags = MAP_SYNC,
1333 .open = xfs_file_open,
1334 .release = xfs_file_release,
1335 .fsync = xfs_file_fsync,
1336 .get_unmapped_area = thp_get_unmapped_area,
1337 .fallocate = xfs_file_fallocate,
1338 .fadvise = xfs_file_fadvise,
1339 .remap_file_range = xfs_file_remap_range,
1340};
1341
1342const struct file_operations xfs_dir_file_operations = {
1343 .open = xfs_dir_open,
1344 .read = generic_read_dir,
1345 .iterate_shared = xfs_file_readdir,
1346 .llseek = generic_file_llseek,
1347 .unlocked_ioctl = xfs_file_ioctl,
1348#ifdef CONFIG_COMPAT
1349 .compat_ioctl = xfs_file_compat_ioctl,
1350#endif
1351 .fsync = xfs_dir_fsync,
1352};