1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (c) 2000-2005 Silicon Graphics, Inc.
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"
17 #include "xfs_bmap_util.h"
19 #include "xfs_dir2_priv.h"
20 #include "xfs_ioctl.h"
21 #include "xfs_trace.h"
23 #include "xfs_icache.h"
25 #include "xfs_iomap.h"
26 #include "xfs_reflink.h"
29 #include <linux/dax.h>
30 #include <linux/falloc.h>
31 #include <linux/backing-dev.h>
32 #include <linux/mman.h>
33 #include <linux/fadvise.h>
34 #include <linux/mount.h>
36 static const struct vm_operations_struct xfs_file_vm_ops;
39 * Decide if the given file range is aligned to the size of the fundamental
40 * allocation unit for the file.
43 xfs_is_falloc_aligned(
48 unsigned int alloc_unit = xfs_inode_alloc_unitsize(ip);
50 if (!is_power_of_2(alloc_unit))
51 return isaligned_64(pos, alloc_unit) &&
52 isaligned_64(len, alloc_unit);
54 return !((pos | len) & (alloc_unit - 1));
58 * Fsync operations on directories are much simpler than on regular files,
59 * as there is no file data to flush, and thus also no need for explicit
60 * cache flush operations, and there are no non-transaction metadata updates
61 * on directories either.
70 struct xfs_inode *ip = XFS_I(file->f_mapping->host);
72 trace_xfs_dir_fsync(ip);
73 return xfs_log_force_inode(ip);
81 if (!xfs_ipincount(ip))
83 if (datasync && !(ip->i_itemp->ili_fsync_fields & ~XFS_ILOG_TIMESTAMP))
85 return ip->i_itemp->ili_commit_seq;
89 * All metadata updates are logged, which means that we just have to flush the
90 * log up to the latest LSN that touched the inode.
92 * If we have concurrent fsync/fdatasync() calls, we need them to all block on
93 * the log force before we clear the ili_fsync_fields field. This ensures that
94 * we don't get a racing sync operation that does not wait for the metadata to
95 * hit the journal before returning. If we race with clearing ili_fsync_fields,
96 * then all that will happen is the log force will do nothing as the lsn will
97 * already be on disk. We can't race with setting ili_fsync_fields because that
98 * is done under XFS_ILOCK_EXCL, and that can't happen because we hold the lock
99 * shared until after the ili_fsync_fields is cleared.
103 struct xfs_inode *ip,
110 xfs_ilock(ip, XFS_ILOCK_SHARED);
111 seq = xfs_fsync_seq(ip, datasync);
113 error = xfs_log_force_seq(ip->i_mount, seq, XFS_LOG_SYNC,
116 spin_lock(&ip->i_itemp->ili_lock);
117 ip->i_itemp->ili_fsync_fields = 0;
118 spin_unlock(&ip->i_itemp->ili_lock);
120 xfs_iunlock(ip, XFS_ILOCK_SHARED);
131 struct xfs_inode *ip = XFS_I(file->f_mapping->host);
132 struct xfs_mount *mp = ip->i_mount;
136 trace_xfs_file_fsync(ip);
138 error = file_write_and_wait_range(file, start, end);
142 if (xfs_is_shutdown(mp))
145 xfs_iflags_clear(ip, XFS_ITRUNCATED);
148 * If we have an RT and/or log subvolume we need to make sure to flush
149 * the write cache the device used for file data first. This is to
150 * ensure newly written file data make it to disk before logging the new
151 * inode size in case of an extending write.
153 if (XFS_IS_REALTIME_INODE(ip))
154 error = blkdev_issue_flush(mp->m_rtdev_targp->bt_bdev);
155 else if (mp->m_logdev_targp != mp->m_ddev_targp)
156 error = blkdev_issue_flush(mp->m_ddev_targp->bt_bdev);
159 * Any inode that has dirty modifications in the log is pinned. The
160 * racy check here for a pinned inode will not catch modifications
161 * that happen concurrently to the fsync call, but fsync semantics
162 * only require to sync previously completed I/O.
164 if (xfs_ipincount(ip)) {
165 err2 = xfs_fsync_flush_log(ip, datasync, &log_flushed);
171 * If we only have a single device, and the log force about was
172 * a no-op we might have to flush the data device cache here.
173 * This can only happen for fdatasync/O_DSYNC if we were overwriting
174 * an already allocated file and thus do not have any metadata to
177 if (!log_flushed && !XFS_IS_REALTIME_INODE(ip) &&
178 mp->m_logdev_targp == mp->m_ddev_targp) {
179 err2 = blkdev_issue_flush(mp->m_ddev_targp->bt_bdev);
190 unsigned int lock_mode)
192 struct xfs_inode *ip = XFS_I(file_inode(iocb->ki_filp));
194 if (iocb->ki_flags & IOCB_NOWAIT) {
195 if (!xfs_ilock_nowait(ip, lock_mode))
198 xfs_ilock(ip, lock_mode);
205 xfs_ilock_iocb_for_write(
207 unsigned int *lock_mode)
210 struct xfs_inode *ip = XFS_I(file_inode(iocb->ki_filp));
212 ret = xfs_ilock_iocb(iocb, *lock_mode);
217 * If a reflink remap is in progress we always need to take the iolock
218 * exclusively to wait for it to finish.
220 if (*lock_mode == XFS_IOLOCK_SHARED &&
221 xfs_iflags_test(ip, XFS_IREMAPPING)) {
222 xfs_iunlock(ip, *lock_mode);
223 *lock_mode = XFS_IOLOCK_EXCL;
224 return xfs_ilock_iocb(iocb, *lock_mode);
235 struct xfs_inode *ip = XFS_I(file_inode(iocb->ki_filp));
238 trace_xfs_file_direct_read(iocb, to);
240 if (!iov_iter_count(to))
241 return 0; /* skip atime */
243 file_accessed(iocb->ki_filp);
245 ret = xfs_ilock_iocb(iocb, XFS_IOLOCK_SHARED);
248 ret = iomap_dio_rw(iocb, to, &xfs_read_iomap_ops, NULL, 0, NULL, 0);
249 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
254 static noinline ssize_t
259 struct xfs_inode *ip = XFS_I(iocb->ki_filp->f_mapping->host);
262 trace_xfs_file_dax_read(iocb, to);
264 if (!iov_iter_count(to))
265 return 0; /* skip atime */
267 ret = xfs_ilock_iocb(iocb, XFS_IOLOCK_SHARED);
270 ret = dax_iomap_rw(iocb, to, &xfs_read_iomap_ops);
271 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
273 file_accessed(iocb->ki_filp);
278 xfs_file_buffered_read(
282 struct xfs_inode *ip = XFS_I(file_inode(iocb->ki_filp));
285 trace_xfs_file_buffered_read(iocb, to);
287 ret = xfs_ilock_iocb(iocb, XFS_IOLOCK_SHARED);
290 ret = generic_file_read_iter(iocb, to);
291 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
301 struct inode *inode = file_inode(iocb->ki_filp);
302 struct xfs_mount *mp = XFS_I(inode)->i_mount;
305 XFS_STATS_INC(mp, xs_read_calls);
307 if (xfs_is_shutdown(mp))
311 ret = xfs_file_dax_read(iocb, to);
312 else if (iocb->ki_flags & IOCB_DIRECT)
313 ret = xfs_file_dio_read(iocb, to);
315 ret = xfs_file_buffered_read(iocb, to);
318 XFS_STATS_ADD(mp, xs_read_bytes, ret);
323 xfs_file_splice_read(
326 struct pipe_inode_info *pipe,
330 struct inode *inode = file_inode(in);
331 struct xfs_inode *ip = XFS_I(inode);
332 struct xfs_mount *mp = ip->i_mount;
335 XFS_STATS_INC(mp, xs_read_calls);
337 if (xfs_is_shutdown(mp))
340 trace_xfs_file_splice_read(ip, *ppos, len);
342 xfs_ilock(ip, XFS_IOLOCK_SHARED);
343 ret = filemap_splice_read(in, ppos, pipe, len, flags);
344 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
346 XFS_STATS_ADD(mp, xs_read_bytes, ret);
351 * Common pre-write limit and setup checks.
353 * Called with the iolocked held either shared and exclusive according to
354 * @iolock, and returns with it held. Might upgrade the iolock to exclusive
355 * if called for a direct write beyond i_size.
358 xfs_file_write_checks(
360 struct iov_iter *from,
361 unsigned int *iolock)
363 struct file *file = iocb->ki_filp;
364 struct inode *inode = file->f_mapping->host;
365 struct xfs_inode *ip = XFS_I(inode);
367 size_t count = iov_iter_count(from);
368 bool drained_dio = false;
372 error = generic_write_checks(iocb, from);
376 if (iocb->ki_flags & IOCB_NOWAIT) {
377 error = break_layout(inode, false);
378 if (error == -EWOULDBLOCK)
381 error = xfs_break_layouts(inode, iolock, BREAK_WRITE);
388 * For changing security info in file_remove_privs() we need i_rwsem
391 if (*iolock == XFS_IOLOCK_SHARED && !IS_NOSEC(inode)) {
392 xfs_iunlock(ip, *iolock);
393 *iolock = XFS_IOLOCK_EXCL;
394 error = xfs_ilock_iocb(iocb, *iolock);
403 * If the offset is beyond the size of the file, we need to zero any
404 * blocks that fall between the existing EOF and the start of this
405 * write. If zeroing is needed and we are currently holding the iolock
406 * shared, we need to update it to exclusive which implies having to
407 * redo all checks before.
409 * We need to serialise against EOF updates that occur in IO completions
410 * here. We want to make sure that nobody is changing the size while we
411 * do this check until we have placed an IO barrier (i.e. hold the
412 * XFS_IOLOCK_EXCL) that prevents new IO from being dispatched. The
413 * spinlock effectively forms a memory barrier once we have the
414 * XFS_IOLOCK_EXCL so we are guaranteed to see the latest EOF value and
415 * hence be able to correctly determine if we need to run zeroing.
417 * We can do an unlocked check here safely as IO completion can only
418 * extend EOF. Truncate is locked out at this point, so the EOF can
419 * not move backwards, only forwards. Hence we only need to take the
420 * slow path and spin locks when we are at or beyond the current EOF.
422 if (iocb->ki_pos <= i_size_read(inode))
425 spin_lock(&ip->i_flags_lock);
426 isize = i_size_read(inode);
427 if (iocb->ki_pos > isize) {
428 spin_unlock(&ip->i_flags_lock);
430 if (iocb->ki_flags & IOCB_NOWAIT)
434 if (*iolock == XFS_IOLOCK_SHARED) {
435 xfs_iunlock(ip, *iolock);
436 *iolock = XFS_IOLOCK_EXCL;
437 xfs_ilock(ip, *iolock);
438 iov_iter_reexpand(from, count);
441 * We now have an IO submission barrier in place, but
442 * AIO can do EOF updates during IO completion and hence
443 * we now need to wait for all of them to drain. Non-AIO
444 * DIO will have drained before we are given the
445 * XFS_IOLOCK_EXCL, and so for most cases this wait is a
448 inode_dio_wait(inode);
453 trace_xfs_zero_eof(ip, isize, iocb->ki_pos - isize);
454 error = xfs_zero_range(ip, isize, iocb->ki_pos - isize, NULL);
458 spin_unlock(&ip->i_flags_lock);
461 return kiocb_modified(iocb);
465 xfs_dio_write_end_io(
471 struct inode *inode = file_inode(iocb->ki_filp);
472 struct xfs_inode *ip = XFS_I(inode);
473 loff_t offset = iocb->ki_pos;
474 unsigned int nofs_flag;
476 trace_xfs_end_io_direct_write(ip, offset, size);
478 if (xfs_is_shutdown(ip->i_mount))
487 * Capture amount written on completion as we can't reliably account
488 * for it on submission.
490 XFS_STATS_ADD(ip->i_mount, xs_write_bytes, size);
493 * We can allocate memory here while doing writeback on behalf of
494 * memory reclaim. To avoid memory allocation deadlocks set the
495 * task-wide nofs context for the following operations.
497 nofs_flag = memalloc_nofs_save();
499 if (flags & IOMAP_DIO_COW) {
500 error = xfs_reflink_end_cow(ip, offset, size);
506 * Unwritten conversion updates the in-core isize after extent
507 * conversion but before updating the on-disk size. Updating isize any
508 * earlier allows a racing dio read to find unwritten extents before
509 * they are converted.
511 if (flags & IOMAP_DIO_UNWRITTEN) {
512 error = xfs_iomap_write_unwritten(ip, offset, size, true);
517 * We need to update the in-core inode size here so that we don't end up
518 * with the on-disk inode size being outside the in-core inode size. We
519 * have no other method of updating EOF for AIO, so always do it here
522 * We need to lock the test/set EOF update as we can be racing with
523 * other IO completions here to update the EOF. Failing to serialise
524 * here can result in EOF moving backwards and Bad Things Happen when
527 * As IO completion only ever extends EOF, we can do an unlocked check
528 * here to avoid taking the spinlock. If we land within the current EOF,
529 * then we do not need to do an extending update at all, and we don't
530 * need to take the lock to check this. If we race with an update moving
531 * EOF, then we'll either still be beyond EOF and need to take the lock,
532 * or we'll be within EOF and we don't need to take it at all.
534 if (offset + size <= i_size_read(inode))
537 spin_lock(&ip->i_flags_lock);
538 if (offset + size > i_size_read(inode)) {
539 i_size_write(inode, offset + size);
540 spin_unlock(&ip->i_flags_lock);
541 error = xfs_setfilesize(ip, offset, size);
543 spin_unlock(&ip->i_flags_lock);
547 memalloc_nofs_restore(nofs_flag);
551 static const struct iomap_dio_ops xfs_dio_write_ops = {
552 .end_io = xfs_dio_write_end_io,
556 * Handle block aligned direct I/O writes
558 static noinline ssize_t
559 xfs_file_dio_write_aligned(
560 struct xfs_inode *ip,
562 struct iov_iter *from)
564 unsigned int iolock = XFS_IOLOCK_SHARED;
567 ret = xfs_ilock_iocb_for_write(iocb, &iolock);
570 ret = xfs_file_write_checks(iocb, from, &iolock);
575 * We don't need to hold the IOLOCK exclusively across the IO, so demote
576 * the iolock back to shared if we had to take the exclusive lock in
577 * xfs_file_write_checks() for other reasons.
579 if (iolock == XFS_IOLOCK_EXCL) {
580 xfs_ilock_demote(ip, XFS_IOLOCK_EXCL);
581 iolock = XFS_IOLOCK_SHARED;
583 trace_xfs_file_direct_write(iocb, from);
584 ret = iomap_dio_rw(iocb, from, &xfs_direct_write_iomap_ops,
585 &xfs_dio_write_ops, 0, NULL, 0);
588 xfs_iunlock(ip, iolock);
593 * Handle block unaligned direct I/O writes
595 * In most cases direct I/O writes will be done holding IOLOCK_SHARED, allowing
596 * them to be done in parallel with reads and other direct I/O writes. However,
597 * if the I/O is not aligned to filesystem blocks, the direct I/O layer may need
598 * to do sub-block zeroing and that requires serialisation against other direct
599 * I/O to the same block. In this case we need to serialise the submission of
600 * the unaligned I/O so that we don't get racing block zeroing in the dio layer.
601 * In the case where sub-block zeroing is not required, we can do concurrent
602 * sub-block dios to the same block successfully.
604 * Optimistically submit the I/O using the shared lock first, but use the
605 * IOMAP_DIO_OVERWRITE_ONLY flag to tell the lower layers to return -EAGAIN
606 * if block allocation or partial block zeroing would be required. In that case
607 * we try again with the exclusive lock.
609 static noinline ssize_t
610 xfs_file_dio_write_unaligned(
611 struct xfs_inode *ip,
613 struct iov_iter *from)
615 size_t isize = i_size_read(VFS_I(ip));
616 size_t count = iov_iter_count(from);
617 unsigned int iolock = XFS_IOLOCK_SHARED;
618 unsigned int flags = IOMAP_DIO_OVERWRITE_ONLY;
622 * Extending writes need exclusivity because of the sub-block zeroing
623 * that the DIO code always does for partial tail blocks beyond EOF, so
624 * don't even bother trying the fast path in this case.
626 if (iocb->ki_pos > isize || iocb->ki_pos + count >= isize) {
627 if (iocb->ki_flags & IOCB_NOWAIT)
630 iolock = XFS_IOLOCK_EXCL;
631 flags = IOMAP_DIO_FORCE_WAIT;
634 ret = xfs_ilock_iocb_for_write(iocb, &iolock);
639 * We can't properly handle unaligned direct I/O to reflink files yet,
640 * as we can't unshare a partial block.
642 if (xfs_is_cow_inode(ip)) {
643 trace_xfs_reflink_bounce_dio_write(iocb, from);
648 ret = xfs_file_write_checks(iocb, from, &iolock);
653 * If we are doing exclusive unaligned I/O, this must be the only I/O
654 * in-flight. Otherwise we risk data corruption due to unwritten extent
655 * conversions from the AIO end_io handler. Wait for all other I/O to
658 if (flags & IOMAP_DIO_FORCE_WAIT)
659 inode_dio_wait(VFS_I(ip));
661 trace_xfs_file_direct_write(iocb, from);
662 ret = iomap_dio_rw(iocb, from, &xfs_direct_write_iomap_ops,
663 &xfs_dio_write_ops, flags, NULL, 0);
666 * Retry unaligned I/O with exclusive blocking semantics if the DIO
667 * layer rejected it for mapping or locking reasons. If we are doing
668 * nonblocking user I/O, propagate the error.
670 if (ret == -EAGAIN && !(iocb->ki_flags & IOCB_NOWAIT)) {
671 ASSERT(flags & IOMAP_DIO_OVERWRITE_ONLY);
672 xfs_iunlock(ip, iolock);
673 goto retry_exclusive;
678 xfs_iunlock(ip, iolock);
685 struct iov_iter *from)
687 struct xfs_inode *ip = XFS_I(file_inode(iocb->ki_filp));
688 struct xfs_buftarg *target = xfs_inode_buftarg(ip);
689 size_t count = iov_iter_count(from);
691 /* direct I/O must be aligned to device logical sector size */
692 if ((iocb->ki_pos | count) & target->bt_logical_sectormask)
694 if ((iocb->ki_pos | count) & ip->i_mount->m_blockmask)
695 return xfs_file_dio_write_unaligned(ip, iocb, from);
696 return xfs_file_dio_write_aligned(ip, iocb, from);
699 static noinline ssize_t
702 struct iov_iter *from)
704 struct inode *inode = iocb->ki_filp->f_mapping->host;
705 struct xfs_inode *ip = XFS_I(inode);
706 unsigned int iolock = XFS_IOLOCK_EXCL;
707 ssize_t ret, error = 0;
710 ret = xfs_ilock_iocb(iocb, iolock);
713 ret = xfs_file_write_checks(iocb, from, &iolock);
719 trace_xfs_file_dax_write(iocb, from);
720 ret = dax_iomap_rw(iocb, from, &xfs_dax_write_iomap_ops);
721 if (ret > 0 && iocb->ki_pos > i_size_read(inode)) {
722 i_size_write(inode, iocb->ki_pos);
723 error = xfs_setfilesize(ip, pos, ret);
727 xfs_iunlock(ip, iolock);
732 XFS_STATS_ADD(ip->i_mount, xs_write_bytes, ret);
734 /* Handle various SYNC-type writes */
735 ret = generic_write_sync(iocb, ret);
741 xfs_file_buffered_write(
743 struct iov_iter *from)
745 struct inode *inode = iocb->ki_filp->f_mapping->host;
746 struct xfs_inode *ip = XFS_I(inode);
748 bool cleared_space = false;
752 iolock = XFS_IOLOCK_EXCL;
753 ret = xfs_ilock_iocb(iocb, iolock);
757 ret = xfs_file_write_checks(iocb, from, &iolock);
761 trace_xfs_file_buffered_write(iocb, from);
762 ret = iomap_file_buffered_write(iocb, from,
763 &xfs_buffered_write_iomap_ops);
766 * If we hit a space limit, try to free up some lingering preallocated
767 * space before returning an error. In the case of ENOSPC, first try to
768 * write back all dirty inodes to free up some of the excess reserved
769 * metadata space. This reduces the chances that the eofblocks scan
770 * waits on dirty mappings. Since xfs_flush_inodes() is serialized, this
771 * also behaves as a filter to prevent too many eofblocks scans from
772 * running at the same time. Use a synchronous scan to increase the
773 * effectiveness of the scan.
775 if (ret == -EDQUOT && !cleared_space) {
776 xfs_iunlock(ip, iolock);
777 xfs_blockgc_free_quota(ip, XFS_ICWALK_FLAG_SYNC);
778 cleared_space = true;
780 } else if (ret == -ENOSPC && !cleared_space) {
781 struct xfs_icwalk icw = {0};
783 cleared_space = true;
784 xfs_flush_inodes(ip->i_mount);
786 xfs_iunlock(ip, iolock);
787 icw.icw_flags = XFS_ICWALK_FLAG_SYNC;
788 xfs_blockgc_free_space(ip->i_mount, &icw);
794 xfs_iunlock(ip, iolock);
797 XFS_STATS_ADD(ip->i_mount, xs_write_bytes, ret);
798 /* Handle various SYNC-type writes */
799 ret = generic_write_sync(iocb, ret);
807 struct iov_iter *from)
809 struct inode *inode = iocb->ki_filp->f_mapping->host;
810 struct xfs_inode *ip = XFS_I(inode);
812 size_t ocount = iov_iter_count(from);
814 XFS_STATS_INC(ip->i_mount, xs_write_calls);
819 if (xfs_is_shutdown(ip->i_mount))
823 return xfs_file_dax_write(iocb, from);
825 if (iocb->ki_flags & IOCB_DIRECT) {
827 * Allow a directio write to fall back to a buffered
828 * write *only* in the case that we're doing a reflink
829 * CoW. In all other directio scenarios we do not
830 * allow an operation to fall back to buffered mode.
832 ret = xfs_file_dio_write(iocb, from);
837 return xfs_file_buffered_write(iocb, from);
840 /* Does this file, inode, or mount want synchronous writes? */
841 static inline bool xfs_file_sync_writes(struct file *filp)
843 struct xfs_inode *ip = XFS_I(file_inode(filp));
845 if (xfs_has_wsync(ip->i_mount))
847 if (filp->f_flags & (__O_SYNC | O_DSYNC))
849 if (IS_SYNC(file_inode(filp)))
855 #define XFS_FALLOC_FL_SUPPORTED \
856 (FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE | \
857 FALLOC_FL_COLLAPSE_RANGE | FALLOC_FL_ZERO_RANGE | \
858 FALLOC_FL_INSERT_RANGE | FALLOC_FL_UNSHARE_RANGE)
867 struct inode *inode = file_inode(file);
868 struct xfs_inode *ip = XFS_I(inode);
870 uint iolock = XFS_IOLOCK_EXCL | XFS_MMAPLOCK_EXCL;
872 bool do_file_insert = false;
874 if (!S_ISREG(inode->i_mode))
876 if (mode & ~XFS_FALLOC_FL_SUPPORTED)
879 xfs_ilock(ip, iolock);
880 error = xfs_break_layouts(inode, &iolock, BREAK_UNMAP);
885 * Must wait for all AIO to complete before we continue as AIO can
886 * change the file size on completion without holding any locks we
887 * currently hold. We must do this first because AIO can update both
888 * the on disk and in memory inode sizes, and the operations that follow
889 * require the in-memory size to be fully up-to-date.
891 inode_dio_wait(inode);
894 * Now AIO and DIO has drained we flush and (if necessary) invalidate
895 * the cached range over the first operation we are about to run.
897 * We care about zero and collapse here because they both run a hole
898 * punch over the range first. Because that can zero data, and the range
899 * of invalidation for the shift operations is much larger, we still do
900 * the required flush for collapse in xfs_prepare_shift().
902 * Insert has the same range requirements as collapse, and we extend the
903 * file first which can zero data. Hence insert has the same
904 * flush/invalidate requirements as collapse and so they are both
905 * handled at the right time by xfs_prepare_shift().
907 if (mode & (FALLOC_FL_PUNCH_HOLE | FALLOC_FL_ZERO_RANGE |
908 FALLOC_FL_COLLAPSE_RANGE)) {
909 error = xfs_flush_unmap_range(ip, offset, len);
914 error = file_modified(file);
918 if (mode & FALLOC_FL_PUNCH_HOLE) {
919 error = xfs_free_file_space(ip, offset, len);
922 } else if (mode & FALLOC_FL_COLLAPSE_RANGE) {
923 if (!xfs_is_falloc_aligned(ip, offset, len)) {
929 * There is no need to overlap collapse range with EOF,
930 * in which case it is effectively a truncate operation
932 if (offset + len >= i_size_read(inode)) {
937 new_size = i_size_read(inode) - len;
939 error = xfs_collapse_file_space(ip, offset, len);
942 } else if (mode & FALLOC_FL_INSERT_RANGE) {
943 loff_t isize = i_size_read(inode);
945 if (!xfs_is_falloc_aligned(ip, offset, len)) {
951 * New inode size must not exceed ->s_maxbytes, accounting for
952 * possible signed overflow.
954 if (inode->i_sb->s_maxbytes - isize < len) {
958 new_size = isize + len;
960 /* Offset should be less than i_size */
961 if (offset >= isize) {
965 do_file_insert = true;
967 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
968 offset + len > i_size_read(inode)) {
969 new_size = offset + len;
970 error = inode_newsize_ok(inode, new_size);
975 if (mode & FALLOC_FL_ZERO_RANGE) {
977 * Punch a hole and prealloc the range. We use a hole
978 * punch rather than unwritten extent conversion for two
981 * 1.) Hole punch handles partial block zeroing for us.
982 * 2.) If prealloc returns ENOSPC, the file range is
983 * still zero-valued by virtue of the hole punch.
985 unsigned int blksize = i_blocksize(inode);
987 trace_xfs_zero_file_space(ip);
989 error = xfs_free_file_space(ip, offset, len);
993 len = round_up(offset + len, blksize) -
994 round_down(offset, blksize);
995 offset = round_down(offset, blksize);
996 } else if (mode & FALLOC_FL_UNSHARE_RANGE) {
997 error = xfs_reflink_unshare(ip, offset, len);
1002 * If always_cow mode we can't use preallocations and
1003 * thus should not create them.
1005 if (xfs_is_always_cow_inode(ip)) {
1006 error = -EOPNOTSUPP;
1011 if (!xfs_is_always_cow_inode(ip)) {
1012 error = xfs_alloc_file_space(ip, offset, len);
1018 /* Change file size if needed */
1022 iattr.ia_valid = ATTR_SIZE;
1023 iattr.ia_size = new_size;
1024 error = xfs_vn_setattr_size(file_mnt_idmap(file),
1025 file_dentry(file), &iattr);
1031 * Perform hole insertion now that the file size has been
1032 * updated so that if we crash during the operation we don't
1033 * leave shifted extents past EOF and hence losing access to
1034 * the data that is contained within them.
1036 if (do_file_insert) {
1037 error = xfs_insert_file_space(ip, offset, len);
1042 if (xfs_file_sync_writes(file))
1043 error = xfs_log_force_inode(ip);
1046 xfs_iunlock(ip, iolock);
1057 struct xfs_inode *ip = XFS_I(file_inode(file));
1062 * Operations creating pages in page cache need protection from hole
1063 * punching and similar ops
1065 if (advice == POSIX_FADV_WILLNEED) {
1066 lockflags = XFS_IOLOCK_SHARED;
1067 xfs_ilock(ip, lockflags);
1069 ret = generic_fadvise(file, start, end, advice);
1071 xfs_iunlock(ip, lockflags);
1076 xfs_file_remap_range(
1077 struct file *file_in,
1079 struct file *file_out,
1082 unsigned int remap_flags)
1084 struct inode *inode_in = file_inode(file_in);
1085 struct xfs_inode *src = XFS_I(inode_in);
1086 struct inode *inode_out = file_inode(file_out);
1087 struct xfs_inode *dest = XFS_I(inode_out);
1088 struct xfs_mount *mp = src->i_mount;
1089 loff_t remapped = 0;
1090 xfs_extlen_t cowextsize;
1093 if (remap_flags & ~(REMAP_FILE_DEDUP | REMAP_FILE_ADVISORY))
1096 if (!xfs_has_reflink(mp))
1099 if (xfs_is_shutdown(mp))
1102 /* Prepare and then clone file data. */
1103 ret = xfs_reflink_remap_prep(file_in, pos_in, file_out, pos_out,
1105 if (ret || len == 0)
1108 trace_xfs_reflink_remap_range(src, pos_in, len, dest, pos_out);
1110 ret = xfs_reflink_remap_blocks(src, pos_in, dest, pos_out, len,
1116 * Carry the cowextsize hint from src to dest if we're sharing the
1117 * entire source file to the entire destination file, the source file
1118 * has a cowextsize hint, and the destination file does not.
1121 if (pos_in == 0 && len == i_size_read(inode_in) &&
1122 (src->i_diflags2 & XFS_DIFLAG2_COWEXTSIZE) &&
1123 pos_out == 0 && len >= i_size_read(inode_out) &&
1124 !(dest->i_diflags2 & XFS_DIFLAG2_COWEXTSIZE))
1125 cowextsize = src->i_cowextsize;
1127 ret = xfs_reflink_update_dest(dest, pos_out + len, cowextsize,
1132 if (xfs_file_sync_writes(file_in) || xfs_file_sync_writes(file_out))
1133 xfs_log_force_inode(dest);
1135 xfs_iunlock2_remapping(src, dest);
1137 trace_xfs_reflink_remap_range_error(dest, ret, _RET_IP_);
1138 return remapped > 0 ? remapped : ret;
1143 struct inode *inode,
1146 if (xfs_is_shutdown(XFS_M(inode->i_sb)))
1148 file->f_mode |= FMODE_NOWAIT | FMODE_CAN_ODIRECT;
1149 return generic_file_open(inode, file);
1154 struct inode *inode,
1157 struct xfs_inode *ip = XFS_I(inode);
1161 if (xfs_is_shutdown(ip->i_mount))
1163 error = generic_file_open(inode, file);
1168 * If there are any blocks, read-ahead block 0 as we're almost
1169 * certain to have the next operation be a read there.
1171 mode = xfs_ilock_data_map_shared(ip);
1172 if (ip->i_df.if_nextents > 0)
1173 error = xfs_dir3_data_readahead(ip, 0, 0);
1174 xfs_iunlock(ip, mode);
1180 struct inode *inode,
1183 return xfs_release(XFS_I(inode));
1189 struct dir_context *ctx)
1191 struct inode *inode = file_inode(file);
1192 xfs_inode_t *ip = XFS_I(inode);
1196 * The Linux API doesn't pass down the total size of the buffer
1197 * we read into down to the filesystem. With the filldir concept
1198 * it's not needed for correct information, but the XFS dir2 leaf
1199 * code wants an estimate of the buffer size to calculate it's
1200 * readahead window and size the buffers used for mapping to
1203 * Try to give it an estimate that's good enough, maybe at some
1204 * point we can change the ->readdir prototype to include the
1205 * buffer size. For now we use the current glibc buffer size.
1207 bufsize = (size_t)min_t(loff_t, XFS_READDIR_BUFSIZE, ip->i_disk_size);
1209 return xfs_readdir(NULL, ip, ctx, bufsize);
1218 struct inode *inode = file->f_mapping->host;
1220 if (xfs_is_shutdown(XFS_I(inode)->i_mount))
1225 return generic_file_llseek(file, offset, whence);
1227 offset = iomap_seek_hole(inode, offset, &xfs_seek_iomap_ops);
1230 offset = iomap_seek_data(inode, offset, &xfs_seek_iomap_ops);
1236 return vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
1239 static inline vm_fault_t
1240 xfs_dax_fault_locked(
1241 struct vm_fault *vmf,
1248 if (!IS_ENABLED(CONFIG_FS_DAX)) {
1250 return VM_FAULT_SIGBUS;
1252 ret = dax_iomap_fault(vmf, order, &pfn, NULL,
1253 (write_fault && !vmf->cow_page) ?
1254 &xfs_dax_write_iomap_ops :
1255 &xfs_read_iomap_ops);
1256 if (ret & VM_FAULT_NEEDDSYNC)
1257 ret = dax_finish_sync_fault(vmf, order, pfn);
1263 struct vm_fault *vmf,
1266 struct xfs_inode *ip = XFS_I(file_inode(vmf->vma->vm_file));
1269 xfs_ilock(ip, XFS_MMAPLOCK_SHARED);
1270 ret = xfs_dax_fault_locked(vmf, order, false);
1271 xfs_iunlock(ip, XFS_MMAPLOCK_SHARED);
1278 struct vm_fault *vmf,
1281 struct inode *inode = file_inode(vmf->vma->vm_file);
1282 struct xfs_inode *ip = XFS_I(inode);
1283 unsigned int lock_mode = XFS_MMAPLOCK_SHARED;
1286 sb_start_pagefault(inode->i_sb);
1287 file_update_time(vmf->vma->vm_file);
1290 * Normally we only need the shared mmaplock, but if a reflink remap is
1291 * in progress we take the exclusive lock to wait for the remap to
1292 * finish before taking a write fault.
1294 xfs_ilock(ip, XFS_MMAPLOCK_SHARED);
1295 if (xfs_iflags_test(ip, XFS_IREMAPPING)) {
1296 xfs_iunlock(ip, XFS_MMAPLOCK_SHARED);
1297 xfs_ilock(ip, XFS_MMAPLOCK_EXCL);
1298 lock_mode = XFS_MMAPLOCK_EXCL;
1302 ret = xfs_dax_fault_locked(vmf, order, true);
1304 ret = iomap_page_mkwrite(vmf, &xfs_page_mkwrite_iomap_ops);
1305 xfs_iunlock(ip, lock_mode);
1307 sb_end_pagefault(inode->i_sb);
1312 * Locking for serialisation of IO during page faults. This results in a lock
1316 * sb_start_pagefault(vfs, freeze)
1317 * invalidate_lock (vfs/XFS_MMAPLOCK - truncate serialisation)
1319 * i_lock (XFS - extent map serialisation)
1322 __xfs_filemap_fault(
1323 struct vm_fault *vmf,
1327 struct inode *inode = file_inode(vmf->vma->vm_file);
1329 trace_xfs_filemap_fault(XFS_I(inode), order, write_fault);
1332 return xfs_write_fault(vmf, order);
1334 return xfs_dax_read_fault(vmf, order);
1335 return filemap_fault(vmf);
1340 struct vm_fault *vmf)
1342 return (vmf->flags & FAULT_FLAG_WRITE) &&
1343 (vmf->vma->vm_flags & VM_SHARED);
1348 struct vm_fault *vmf)
1350 /* DAX can shortcut the normal fault path on write faults! */
1351 return __xfs_filemap_fault(vmf, 0,
1352 IS_DAX(file_inode(vmf->vma->vm_file)) &&
1353 xfs_is_write_fault(vmf));
1357 xfs_filemap_huge_fault(
1358 struct vm_fault *vmf,
1361 if (!IS_DAX(file_inode(vmf->vma->vm_file)))
1362 return VM_FAULT_FALLBACK;
1364 /* DAX can shortcut the normal fault path on write faults! */
1365 return __xfs_filemap_fault(vmf, order,
1366 xfs_is_write_fault(vmf));
1370 xfs_filemap_page_mkwrite(
1371 struct vm_fault *vmf)
1373 return __xfs_filemap_fault(vmf, 0, true);
1377 * pfn_mkwrite was originally intended to ensure we capture time stamp updates
1378 * on write faults. In reality, it needs to serialise against truncate and
1379 * prepare memory for writing so handle is as standard write fault.
1382 xfs_filemap_pfn_mkwrite(
1383 struct vm_fault *vmf)
1386 return __xfs_filemap_fault(vmf, 0, true);
1389 static const struct vm_operations_struct xfs_file_vm_ops = {
1390 .fault = xfs_filemap_fault,
1391 .huge_fault = xfs_filemap_huge_fault,
1392 .map_pages = filemap_map_pages,
1393 .page_mkwrite = xfs_filemap_page_mkwrite,
1394 .pfn_mkwrite = xfs_filemap_pfn_mkwrite,
1400 struct vm_area_struct *vma)
1402 struct inode *inode = file_inode(file);
1403 struct xfs_buftarg *target = xfs_inode_buftarg(XFS_I(inode));
1406 * We don't support synchronous mappings for non-DAX files and
1407 * for DAX files if underneath dax_device is not synchronous.
1409 if (!daxdev_mapping_supported(vma, target->bt_daxdev))
1412 file_accessed(file);
1413 vma->vm_ops = &xfs_file_vm_ops;
1415 vm_flags_set(vma, VM_HUGEPAGE);
1419 const struct file_operations xfs_file_operations = {
1420 .llseek = xfs_file_llseek,
1421 .read_iter = xfs_file_read_iter,
1422 .write_iter = xfs_file_write_iter,
1423 .splice_read = xfs_file_splice_read,
1424 .splice_write = iter_file_splice_write,
1425 .iopoll = iocb_bio_iopoll,
1426 .unlocked_ioctl = xfs_file_ioctl,
1427 #ifdef CONFIG_COMPAT
1428 .compat_ioctl = xfs_file_compat_ioctl,
1430 .mmap = xfs_file_mmap,
1431 .open = xfs_file_open,
1432 .release = xfs_file_release,
1433 .fsync = xfs_file_fsync,
1434 .get_unmapped_area = thp_get_unmapped_area,
1435 .fallocate = xfs_file_fallocate,
1436 .fadvise = xfs_file_fadvise,
1437 .remap_file_range = xfs_file_remap_range,
1438 .fop_flags = FOP_MMAP_SYNC | FOP_BUFFER_RASYNC |
1439 FOP_BUFFER_WASYNC | FOP_DIO_PARALLEL_WRITE,
1442 const struct file_operations xfs_dir_file_operations = {
1443 .open = xfs_dir_open,
1444 .read = generic_read_dir,
1445 .iterate_shared = xfs_file_readdir,
1446 .llseek = generic_file_llseek,
1447 .unlocked_ioctl = xfs_file_ioctl,
1448 #ifdef CONFIG_COMPAT
1449 .compat_ioctl = xfs_file_compat_ioctl,
1451 .fsync = xfs_dir_fsync,