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 * Take care of zeroing post-EOF blocks when they might exist.
353 * Returns 0 if successfully, a negative error for a failure, or 1 if this
354 * function dropped the iolock and reacquired it exclusively and the caller
355 * needs to restart the write sanity checks.
358 xfs_file_write_zero_eof(
360 struct iov_iter *from,
361 unsigned int *iolock,
365 struct xfs_inode *ip = XFS_I(iocb->ki_filp->f_mapping->host);
370 * We need to serialise against EOF updates that occur in IO completions
371 * here. We want to make sure that nobody is changing the size while
372 * we do this check until we have placed an IO barrier (i.e. hold
373 * XFS_IOLOCK_EXCL) that prevents new IO from being dispatched. The
374 * spinlock effectively forms a memory barrier once we have
375 * XFS_IOLOCK_EXCL so we are guaranteed to see the latest EOF value and
376 * hence be able to correctly determine if we need to run zeroing.
378 spin_lock(&ip->i_flags_lock);
379 isize = i_size_read(VFS_I(ip));
380 if (iocb->ki_pos <= isize) {
381 spin_unlock(&ip->i_flags_lock);
384 spin_unlock(&ip->i_flags_lock);
386 if (iocb->ki_flags & IOCB_NOWAIT)
391 * If zeroing is needed and we are currently holding the iolock
392 * shared, we need to update it to exclusive which implies
393 * having to redo all checks before.
395 if (*iolock == XFS_IOLOCK_SHARED) {
396 xfs_iunlock(ip, *iolock);
397 *iolock = XFS_IOLOCK_EXCL;
398 xfs_ilock(ip, *iolock);
399 iov_iter_reexpand(from, count);
403 * We now have an IO submission barrier in place, but AIO can do
404 * EOF updates during IO completion and hence we now need to
405 * wait for all of them to drain. Non-AIO DIO will have drained
406 * before we are given the XFS_IOLOCK_EXCL, and so for most
407 * cases this wait is a no-op.
409 inode_dio_wait(VFS_I(ip));
414 trace_xfs_zero_eof(ip, isize, iocb->ki_pos - isize);
416 xfs_ilock(ip, XFS_MMAPLOCK_EXCL);
417 error = xfs_zero_range(ip, isize, iocb->ki_pos - isize, NULL);
418 xfs_iunlock(ip, XFS_MMAPLOCK_EXCL);
424 * Common pre-write limit and setup checks.
426 * Called with the iolock held either shared and exclusive according to
427 * @iolock, and returns with it held. Might upgrade the iolock to exclusive
428 * if called for a direct write beyond i_size.
431 xfs_file_write_checks(
433 struct iov_iter *from,
434 unsigned int *iolock)
436 struct inode *inode = iocb->ki_filp->f_mapping->host;
437 size_t count = iov_iter_count(from);
438 bool drained_dio = false;
442 error = generic_write_checks(iocb, from);
446 if (iocb->ki_flags & IOCB_NOWAIT) {
447 error = break_layout(inode, false);
448 if (error == -EWOULDBLOCK)
451 error = xfs_break_layouts(inode, iolock, BREAK_WRITE);
458 * For changing security info in file_remove_privs() we need i_rwsem
461 if (*iolock == XFS_IOLOCK_SHARED && !IS_NOSEC(inode)) {
462 xfs_iunlock(XFS_I(inode), *iolock);
463 *iolock = XFS_IOLOCK_EXCL;
464 error = xfs_ilock_iocb(iocb, *iolock);
473 * If the offset is beyond the size of the file, we need to zero all
474 * blocks that fall between the existing EOF and the start of this
477 * We can do an unlocked check for i_size here safely as I/O completion
478 * can only extend EOF. Truncate is locked out at this point, so the
479 * EOF can not move backwards, only forwards. Hence we only need to take
480 * the slow path when we are at or beyond the current EOF.
482 if (iocb->ki_pos > i_size_read(inode)) {
483 error = xfs_file_write_zero_eof(iocb, from, iolock, count,
491 return kiocb_modified(iocb);
495 xfs_dio_write_end_io(
501 struct inode *inode = file_inode(iocb->ki_filp);
502 struct xfs_inode *ip = XFS_I(inode);
503 loff_t offset = iocb->ki_pos;
504 unsigned int nofs_flag;
506 trace_xfs_end_io_direct_write(ip, offset, size);
508 if (xfs_is_shutdown(ip->i_mount))
517 * Capture amount written on completion as we can't reliably account
518 * for it on submission.
520 XFS_STATS_ADD(ip->i_mount, xs_write_bytes, size);
523 * We can allocate memory here while doing writeback on behalf of
524 * memory reclaim. To avoid memory allocation deadlocks set the
525 * task-wide nofs context for the following operations.
527 nofs_flag = memalloc_nofs_save();
529 if (flags & IOMAP_DIO_COW) {
530 error = xfs_reflink_end_cow(ip, offset, size);
536 * Unwritten conversion updates the in-core isize after extent
537 * conversion but before updating the on-disk size. Updating isize any
538 * earlier allows a racing dio read to find unwritten extents before
539 * they are converted.
541 if (flags & IOMAP_DIO_UNWRITTEN) {
542 error = xfs_iomap_write_unwritten(ip, offset, size, true);
547 * We need to update the in-core inode size here so that we don't end up
548 * with the on-disk inode size being outside the in-core inode size. We
549 * have no other method of updating EOF for AIO, so always do it here
552 * We need to lock the test/set EOF update as we can be racing with
553 * other IO completions here to update the EOF. Failing to serialise
554 * here can result in EOF moving backwards and Bad Things Happen when
557 * As IO completion only ever extends EOF, we can do an unlocked check
558 * here to avoid taking the spinlock. If we land within the current EOF,
559 * then we do not need to do an extending update at all, and we don't
560 * need to take the lock to check this. If we race with an update moving
561 * EOF, then we'll either still be beyond EOF and need to take the lock,
562 * or we'll be within EOF and we don't need to take it at all.
564 if (offset + size <= i_size_read(inode))
567 spin_lock(&ip->i_flags_lock);
568 if (offset + size > i_size_read(inode)) {
569 i_size_write(inode, offset + size);
570 spin_unlock(&ip->i_flags_lock);
571 error = xfs_setfilesize(ip, offset, size);
573 spin_unlock(&ip->i_flags_lock);
577 memalloc_nofs_restore(nofs_flag);
581 static const struct iomap_dio_ops xfs_dio_write_ops = {
582 .end_io = xfs_dio_write_end_io,
586 * Handle block aligned direct I/O writes
588 static noinline ssize_t
589 xfs_file_dio_write_aligned(
590 struct xfs_inode *ip,
592 struct iov_iter *from)
594 unsigned int iolock = XFS_IOLOCK_SHARED;
597 ret = xfs_ilock_iocb_for_write(iocb, &iolock);
600 ret = xfs_file_write_checks(iocb, from, &iolock);
605 * We don't need to hold the IOLOCK exclusively across the IO, so demote
606 * the iolock back to shared if we had to take the exclusive lock in
607 * xfs_file_write_checks() for other reasons.
609 if (iolock == XFS_IOLOCK_EXCL) {
610 xfs_ilock_demote(ip, XFS_IOLOCK_EXCL);
611 iolock = XFS_IOLOCK_SHARED;
613 trace_xfs_file_direct_write(iocb, from);
614 ret = iomap_dio_rw(iocb, from, &xfs_direct_write_iomap_ops,
615 &xfs_dio_write_ops, 0, NULL, 0);
618 xfs_iunlock(ip, iolock);
623 * Handle block unaligned direct I/O writes
625 * In most cases direct I/O writes will be done holding IOLOCK_SHARED, allowing
626 * them to be done in parallel with reads and other direct I/O writes. However,
627 * if the I/O is not aligned to filesystem blocks, the direct I/O layer may need
628 * to do sub-block zeroing and that requires serialisation against other direct
629 * I/O to the same block. In this case we need to serialise the submission of
630 * the unaligned I/O so that we don't get racing block zeroing in the dio layer.
631 * In the case where sub-block zeroing is not required, we can do concurrent
632 * sub-block dios to the same block successfully.
634 * Optimistically submit the I/O using the shared lock first, but use the
635 * IOMAP_DIO_OVERWRITE_ONLY flag to tell the lower layers to return -EAGAIN
636 * if block allocation or partial block zeroing would be required. In that case
637 * we try again with the exclusive lock.
639 static noinline ssize_t
640 xfs_file_dio_write_unaligned(
641 struct xfs_inode *ip,
643 struct iov_iter *from)
645 size_t isize = i_size_read(VFS_I(ip));
646 size_t count = iov_iter_count(from);
647 unsigned int iolock = XFS_IOLOCK_SHARED;
648 unsigned int flags = IOMAP_DIO_OVERWRITE_ONLY;
652 * Extending writes need exclusivity because of the sub-block zeroing
653 * that the DIO code always does for partial tail blocks beyond EOF, so
654 * don't even bother trying the fast path in this case.
656 if (iocb->ki_pos > isize || iocb->ki_pos + count >= isize) {
657 if (iocb->ki_flags & IOCB_NOWAIT)
660 iolock = XFS_IOLOCK_EXCL;
661 flags = IOMAP_DIO_FORCE_WAIT;
664 ret = xfs_ilock_iocb_for_write(iocb, &iolock);
669 * We can't properly handle unaligned direct I/O to reflink files yet,
670 * as we can't unshare a partial block.
672 if (xfs_is_cow_inode(ip)) {
673 trace_xfs_reflink_bounce_dio_write(iocb, from);
678 ret = xfs_file_write_checks(iocb, from, &iolock);
683 * If we are doing exclusive unaligned I/O, this must be the only I/O
684 * in-flight. Otherwise we risk data corruption due to unwritten extent
685 * conversions from the AIO end_io handler. Wait for all other I/O to
688 if (flags & IOMAP_DIO_FORCE_WAIT)
689 inode_dio_wait(VFS_I(ip));
691 trace_xfs_file_direct_write(iocb, from);
692 ret = iomap_dio_rw(iocb, from, &xfs_direct_write_iomap_ops,
693 &xfs_dio_write_ops, flags, NULL, 0);
696 * Retry unaligned I/O with exclusive blocking semantics if the DIO
697 * layer rejected it for mapping or locking reasons. If we are doing
698 * nonblocking user I/O, propagate the error.
700 if (ret == -EAGAIN && !(iocb->ki_flags & IOCB_NOWAIT)) {
701 ASSERT(flags & IOMAP_DIO_OVERWRITE_ONLY);
702 xfs_iunlock(ip, iolock);
703 goto retry_exclusive;
708 xfs_iunlock(ip, iolock);
715 struct iov_iter *from)
717 struct xfs_inode *ip = XFS_I(file_inode(iocb->ki_filp));
718 struct xfs_buftarg *target = xfs_inode_buftarg(ip);
719 size_t count = iov_iter_count(from);
721 /* direct I/O must be aligned to device logical sector size */
722 if ((iocb->ki_pos | count) & target->bt_logical_sectormask)
724 if ((iocb->ki_pos | count) & ip->i_mount->m_blockmask)
725 return xfs_file_dio_write_unaligned(ip, iocb, from);
726 return xfs_file_dio_write_aligned(ip, iocb, from);
729 static noinline ssize_t
732 struct iov_iter *from)
734 struct inode *inode = iocb->ki_filp->f_mapping->host;
735 struct xfs_inode *ip = XFS_I(inode);
736 unsigned int iolock = XFS_IOLOCK_EXCL;
737 ssize_t ret, error = 0;
740 ret = xfs_ilock_iocb(iocb, iolock);
743 ret = xfs_file_write_checks(iocb, from, &iolock);
749 trace_xfs_file_dax_write(iocb, from);
750 ret = dax_iomap_rw(iocb, from, &xfs_dax_write_iomap_ops);
751 if (ret > 0 && iocb->ki_pos > i_size_read(inode)) {
752 i_size_write(inode, iocb->ki_pos);
753 error = xfs_setfilesize(ip, pos, ret);
757 xfs_iunlock(ip, iolock);
762 XFS_STATS_ADD(ip->i_mount, xs_write_bytes, ret);
764 /* Handle various SYNC-type writes */
765 ret = generic_write_sync(iocb, ret);
771 xfs_file_buffered_write(
773 struct iov_iter *from)
775 struct inode *inode = iocb->ki_filp->f_mapping->host;
776 struct xfs_inode *ip = XFS_I(inode);
778 bool cleared_space = false;
782 iolock = XFS_IOLOCK_EXCL;
783 ret = xfs_ilock_iocb(iocb, iolock);
787 ret = xfs_file_write_checks(iocb, from, &iolock);
791 trace_xfs_file_buffered_write(iocb, from);
792 ret = iomap_file_buffered_write(iocb, from,
793 &xfs_buffered_write_iomap_ops, NULL);
796 * If we hit a space limit, try to free up some lingering preallocated
797 * space before returning an error. In the case of ENOSPC, first try to
798 * write back all dirty inodes to free up some of the excess reserved
799 * metadata space. This reduces the chances that the eofblocks scan
800 * waits on dirty mappings. Since xfs_flush_inodes() is serialized, this
801 * also behaves as a filter to prevent too many eofblocks scans from
802 * running at the same time. Use a synchronous scan to increase the
803 * effectiveness of the scan.
805 if (ret == -EDQUOT && !cleared_space) {
806 xfs_iunlock(ip, iolock);
807 xfs_blockgc_free_quota(ip, XFS_ICWALK_FLAG_SYNC);
808 cleared_space = true;
810 } else if (ret == -ENOSPC && !cleared_space) {
811 struct xfs_icwalk icw = {0};
813 cleared_space = true;
814 xfs_flush_inodes(ip->i_mount);
816 xfs_iunlock(ip, iolock);
817 icw.icw_flags = XFS_ICWALK_FLAG_SYNC;
818 xfs_blockgc_free_space(ip->i_mount, &icw);
824 xfs_iunlock(ip, iolock);
827 XFS_STATS_ADD(ip->i_mount, xs_write_bytes, ret);
828 /* Handle various SYNC-type writes */
829 ret = generic_write_sync(iocb, ret);
837 struct iov_iter *from)
839 struct inode *inode = iocb->ki_filp->f_mapping->host;
840 struct xfs_inode *ip = XFS_I(inode);
842 size_t ocount = iov_iter_count(from);
844 XFS_STATS_INC(ip->i_mount, xs_write_calls);
849 if (xfs_is_shutdown(ip->i_mount))
853 return xfs_file_dax_write(iocb, from);
855 if (iocb->ki_flags & IOCB_ATOMIC) {
857 * Currently only atomic writing of a single FS block is
858 * supported. It would be possible to atomic write smaller than
859 * a FS block, but there is no requirement to support this.
860 * Note that iomap also does not support this yet.
862 if (ocount != ip->i_mount->m_sb.sb_blocksize)
864 ret = generic_atomic_write_valid(iocb, from);
869 if (iocb->ki_flags & IOCB_DIRECT) {
871 * Allow a directio write to fall back to a buffered
872 * write *only* in the case that we're doing a reflink
873 * CoW. In all other directio scenarios we do not
874 * allow an operation to fall back to buffered mode.
876 ret = xfs_file_dio_write(iocb, from);
881 return xfs_file_buffered_write(iocb, from);
884 /* Does this file, inode, or mount want synchronous writes? */
885 static inline bool xfs_file_sync_writes(struct file *filp)
887 struct xfs_inode *ip = XFS_I(file_inode(filp));
889 if (xfs_has_wsync(ip->i_mount))
891 if (filp->f_flags & (__O_SYNC | O_DSYNC))
893 if (IS_SYNC(file_inode(filp)))
907 struct inode *inode = file_inode(file);
909 if ((mode & FALLOC_FL_KEEP_SIZE) || offset + len <= i_size_read(inode))
911 *new_size = offset + len;
912 return inode_newsize_ok(inode, *new_size);
920 struct iattr iattr = {
921 .ia_valid = ATTR_SIZE,
927 return xfs_vn_setattr_size(file_mnt_idmap(file), file_dentry(file),
932 xfs_falloc_collapse_range(
937 struct inode *inode = file_inode(file);
938 loff_t new_size = i_size_read(inode) - len;
941 if (!xfs_is_falloc_aligned(XFS_I(inode), offset, len))
945 * There is no need to overlap collapse range with EOF, in which case it
946 * is effectively a truncate operation
948 if (offset + len >= i_size_read(inode))
951 error = xfs_collapse_file_space(XFS_I(inode), offset, len);
954 return xfs_falloc_setsize(file, new_size);
958 xfs_falloc_insert_range(
963 struct inode *inode = file_inode(file);
964 loff_t isize = i_size_read(inode);
967 if (!xfs_is_falloc_aligned(XFS_I(inode), offset, len))
971 * New inode size must not exceed ->s_maxbytes, accounting for
972 * possible signed overflow.
974 if (inode->i_sb->s_maxbytes - isize < len)
977 /* Offset should be less than i_size */
981 error = xfs_falloc_setsize(file, isize + len);
986 * Perform hole insertion now that the file size has been updated so
987 * that if we crash during the operation we don't leave shifted extents
988 * past EOF and hence losing access to the data that is contained within
991 return xfs_insert_file_space(XFS_I(inode), offset, len);
995 * Punch a hole and prealloc the range. We use a hole punch rather than
996 * unwritten extent conversion for two reasons:
998 * 1.) Hole punch handles partial block zeroing for us.
999 * 2.) If prealloc returns ENOSPC, the file range is still zero-valued by
1000 * virtue of the hole punch.
1003 xfs_falloc_zero_range(
1009 struct inode *inode = file_inode(file);
1010 unsigned int blksize = i_blocksize(inode);
1011 loff_t new_size = 0;
1014 trace_xfs_zero_file_space(XFS_I(inode));
1016 error = xfs_falloc_newsize(file, mode, offset, len, &new_size);
1020 error = xfs_free_file_space(XFS_I(inode), offset, len);
1024 len = round_up(offset + len, blksize) - round_down(offset, blksize);
1025 offset = round_down(offset, blksize);
1026 error = xfs_alloc_file_space(XFS_I(inode), offset, len);
1029 return xfs_falloc_setsize(file, new_size);
1033 xfs_falloc_unshare_range(
1039 struct inode *inode = file_inode(file);
1040 loff_t new_size = 0;
1043 error = xfs_falloc_newsize(file, mode, offset, len, &new_size);
1047 error = xfs_reflink_unshare(XFS_I(inode), offset, len);
1051 error = xfs_alloc_file_space(XFS_I(inode), offset, len);
1054 return xfs_falloc_setsize(file, new_size);
1058 xfs_falloc_allocate_range(
1064 struct inode *inode = file_inode(file);
1065 loff_t new_size = 0;
1069 * If always_cow mode we can't use preallocations and thus should not
1072 if (xfs_is_always_cow_inode(XFS_I(inode)))
1075 error = xfs_falloc_newsize(file, mode, offset, len, &new_size);
1079 error = xfs_alloc_file_space(XFS_I(inode), offset, len);
1082 return xfs_falloc_setsize(file, new_size);
1085 #define XFS_FALLOC_FL_SUPPORTED \
1086 (FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE | \
1087 FALLOC_FL_COLLAPSE_RANGE | FALLOC_FL_ZERO_RANGE | \
1088 FALLOC_FL_INSERT_RANGE | FALLOC_FL_UNSHARE_RANGE)
1097 struct inode *inode = file_inode(file);
1098 struct xfs_inode *ip = XFS_I(inode);
1100 uint iolock = XFS_IOLOCK_EXCL | XFS_MMAPLOCK_EXCL;
1102 if (!S_ISREG(inode->i_mode))
1104 if (mode & ~XFS_FALLOC_FL_SUPPORTED)
1107 xfs_ilock(ip, iolock);
1108 error = xfs_break_layouts(inode, &iolock, BREAK_UNMAP);
1113 * Must wait for all AIO to complete before we continue as AIO can
1114 * change the file size on completion without holding any locks we
1115 * currently hold. We must do this first because AIO can update both
1116 * the on disk and in memory inode sizes, and the operations that follow
1117 * require the in-memory size to be fully up-to-date.
1119 inode_dio_wait(inode);
1121 error = file_modified(file);
1125 switch (mode & FALLOC_FL_MODE_MASK) {
1126 case FALLOC_FL_PUNCH_HOLE:
1127 error = xfs_free_file_space(ip, offset, len);
1129 case FALLOC_FL_COLLAPSE_RANGE:
1130 error = xfs_falloc_collapse_range(file, offset, len);
1132 case FALLOC_FL_INSERT_RANGE:
1133 error = xfs_falloc_insert_range(file, offset, len);
1135 case FALLOC_FL_ZERO_RANGE:
1136 error = xfs_falloc_zero_range(file, mode, offset, len);
1138 case FALLOC_FL_UNSHARE_RANGE:
1139 error = xfs_falloc_unshare_range(file, mode, offset, len);
1141 case FALLOC_FL_ALLOCATE_RANGE:
1142 error = xfs_falloc_allocate_range(file, mode, offset, len);
1145 error = -EOPNOTSUPP;
1149 if (!error && xfs_file_sync_writes(file))
1150 error = xfs_log_force_inode(ip);
1153 xfs_iunlock(ip, iolock);
1164 struct xfs_inode *ip = XFS_I(file_inode(file));
1169 * Operations creating pages in page cache need protection from hole
1170 * punching and similar ops
1172 if (advice == POSIX_FADV_WILLNEED) {
1173 lockflags = XFS_IOLOCK_SHARED;
1174 xfs_ilock(ip, lockflags);
1176 ret = generic_fadvise(file, start, end, advice);
1178 xfs_iunlock(ip, lockflags);
1183 xfs_file_remap_range(
1184 struct file *file_in,
1186 struct file *file_out,
1189 unsigned int remap_flags)
1191 struct inode *inode_in = file_inode(file_in);
1192 struct xfs_inode *src = XFS_I(inode_in);
1193 struct inode *inode_out = file_inode(file_out);
1194 struct xfs_inode *dest = XFS_I(inode_out);
1195 struct xfs_mount *mp = src->i_mount;
1196 loff_t remapped = 0;
1197 xfs_extlen_t cowextsize;
1200 if (remap_flags & ~(REMAP_FILE_DEDUP | REMAP_FILE_ADVISORY))
1203 if (!xfs_has_reflink(mp))
1206 if (xfs_is_shutdown(mp))
1209 /* Prepare and then clone file data. */
1210 ret = xfs_reflink_remap_prep(file_in, pos_in, file_out, pos_out,
1212 if (ret || len == 0)
1215 trace_xfs_reflink_remap_range(src, pos_in, len, dest, pos_out);
1217 ret = xfs_reflink_remap_blocks(src, pos_in, dest, pos_out, len,
1223 * Carry the cowextsize hint from src to dest if we're sharing the
1224 * entire source file to the entire destination file, the source file
1225 * has a cowextsize hint, and the destination file does not.
1228 if (pos_in == 0 && len == i_size_read(inode_in) &&
1229 (src->i_diflags2 & XFS_DIFLAG2_COWEXTSIZE) &&
1230 pos_out == 0 && len >= i_size_read(inode_out) &&
1231 !(dest->i_diflags2 & XFS_DIFLAG2_COWEXTSIZE))
1232 cowextsize = src->i_cowextsize;
1234 ret = xfs_reflink_update_dest(dest, pos_out + len, cowextsize,
1239 if (xfs_file_sync_writes(file_in) || xfs_file_sync_writes(file_out))
1240 xfs_log_force_inode(dest);
1242 xfs_iunlock2_remapping(src, dest);
1244 trace_xfs_reflink_remap_range_error(dest, ret, _RET_IP_);
1246 * If the caller did not set CAN_SHORTEN, then it is not prepared to
1247 * handle partial results -- either the whole remap succeeds, or we
1248 * must say why it did not. In this case, any error should be returned
1251 if (ret && remapped < len && !(remap_flags & REMAP_FILE_CAN_SHORTEN))
1253 return remapped > 0 ? remapped : ret;
1258 struct inode *inode,
1261 if (xfs_is_shutdown(XFS_M(inode->i_sb)))
1263 file->f_mode |= FMODE_NOWAIT | FMODE_CAN_ODIRECT;
1264 if (xfs_inode_can_atomicwrite(XFS_I(inode)))
1265 file->f_mode |= FMODE_CAN_ATOMIC_WRITE;
1266 return generic_file_open(inode, file);
1271 struct inode *inode,
1274 struct xfs_inode *ip = XFS_I(inode);
1278 if (xfs_is_shutdown(ip->i_mount))
1280 error = generic_file_open(inode, file);
1285 * If there are any blocks, read-ahead block 0 as we're almost
1286 * certain to have the next operation be a read there.
1288 mode = xfs_ilock_data_map_shared(ip);
1289 if (ip->i_df.if_nextents > 0)
1290 error = xfs_dir3_data_readahead(ip, 0, 0);
1291 xfs_iunlock(ip, mode);
1296 * Don't bother propagating errors. We're just doing cleanup, and the caller
1297 * ignores the return value anyway.
1301 struct inode *inode,
1304 struct xfs_inode *ip = XFS_I(inode);
1305 struct xfs_mount *mp = ip->i_mount;
1308 * If this is a read-only mount or the file system has been shut down,
1309 * don't generate I/O.
1311 if (xfs_is_readonly(mp) || xfs_is_shutdown(mp))
1315 * If we previously truncated this file and removed old data in the
1316 * process, we want to initiate "early" writeout on the last close.
1317 * This is an attempt to combat the notorious NULL files problem which
1318 * is particularly noticeable from a truncate down, buffered (re-)write
1319 * (delalloc), followed by a crash. What we are effectively doing here
1320 * is significantly reducing the time window where we'd otherwise be
1321 * exposed to that problem.
1323 if (xfs_iflags_test_and_clear(ip, XFS_ITRUNCATED)) {
1324 xfs_iflags_clear(ip, XFS_EOFBLOCKS_RELEASED);
1325 if (ip->i_delayed_blks > 0)
1326 filemap_flush(inode->i_mapping);
1330 * XFS aggressively preallocates post-EOF space to generate contiguous
1331 * allocations for writers that append to the end of the file.
1333 * To support workloads that close and reopen the file frequently, these
1334 * preallocations usually persist after a close unless it is the first
1335 * close for the inode. This is a tradeoff to generate tightly packed
1336 * data layouts for unpacking tarballs or similar archives that write
1337 * one file after another without going back to it while keeping the
1338 * preallocation for files that have recurring open/write/close cycles.
1340 * This heuristic is skipped for inodes with the append-only flag as
1341 * that flag is rather pointless for inodes written only once.
1343 * There is no point in freeing blocks here for open but unlinked files
1344 * as they will be taken care of by the inactivation path soon.
1346 * When releasing a read-only context, don't flush data or trim post-EOF
1347 * blocks. This avoids open/read/close workloads from removing EOF
1348 * blocks that other writers depend upon to reduce fragmentation.
1350 * If we can't get the iolock just skip truncating the blocks past EOF
1351 * because we could deadlock with the mmap_lock otherwise. We'll get
1352 * another chance to drop them once the last reference to the inode is
1353 * dropped, so we'll never leak blocks permanently.
1355 if (inode->i_nlink &&
1356 (file->f_mode & FMODE_WRITE) &&
1357 !(ip->i_diflags & XFS_DIFLAG_APPEND) &&
1358 !xfs_iflags_test(ip, XFS_EOFBLOCKS_RELEASED) &&
1359 xfs_ilock_nowait(ip, XFS_IOLOCK_EXCL)) {
1360 if (xfs_can_free_eofblocks(ip) &&
1361 !xfs_iflags_test_and_set(ip, XFS_EOFBLOCKS_RELEASED))
1362 xfs_free_eofblocks(ip);
1363 xfs_iunlock(ip, XFS_IOLOCK_EXCL);
1372 struct dir_context *ctx)
1374 struct inode *inode = file_inode(file);
1375 xfs_inode_t *ip = XFS_I(inode);
1379 * The Linux API doesn't pass down the total size of the buffer
1380 * we read into down to the filesystem. With the filldir concept
1381 * it's not needed for correct information, but the XFS dir2 leaf
1382 * code wants an estimate of the buffer size to calculate it's
1383 * readahead window and size the buffers used for mapping to
1386 * Try to give it an estimate that's good enough, maybe at some
1387 * point we can change the ->readdir prototype to include the
1388 * buffer size. For now we use the current glibc buffer size.
1390 bufsize = (size_t)min_t(loff_t, XFS_READDIR_BUFSIZE, ip->i_disk_size);
1392 return xfs_readdir(NULL, ip, ctx, bufsize);
1401 struct inode *inode = file->f_mapping->host;
1403 if (xfs_is_shutdown(XFS_I(inode)->i_mount))
1408 return generic_file_llseek(file, offset, whence);
1410 offset = iomap_seek_hole(inode, offset, &xfs_seek_iomap_ops);
1413 offset = iomap_seek_data(inode, offset, &xfs_seek_iomap_ops);
1419 return vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
1422 static inline vm_fault_t
1423 xfs_dax_fault_locked(
1424 struct vm_fault *vmf,
1431 if (!IS_ENABLED(CONFIG_FS_DAX)) {
1433 return VM_FAULT_SIGBUS;
1435 ret = dax_iomap_fault(vmf, order, &pfn, NULL,
1436 (write_fault && !vmf->cow_page) ?
1437 &xfs_dax_write_iomap_ops :
1438 &xfs_read_iomap_ops);
1439 if (ret & VM_FAULT_NEEDDSYNC)
1440 ret = dax_finish_sync_fault(vmf, order, pfn);
1446 struct vm_fault *vmf,
1449 struct xfs_inode *ip = XFS_I(file_inode(vmf->vma->vm_file));
1452 trace_xfs_read_fault(ip, order);
1454 xfs_ilock(ip, XFS_MMAPLOCK_SHARED);
1455 ret = xfs_dax_fault_locked(vmf, order, false);
1456 xfs_iunlock(ip, XFS_MMAPLOCK_SHARED);
1462 * Locking for serialisation of IO during page faults. This results in a lock
1466 * sb_start_pagefault(vfs, freeze)
1467 * invalidate_lock (vfs/XFS_MMAPLOCK - truncate serialisation)
1469 * i_lock (XFS - extent map serialisation)
1473 struct vm_fault *vmf,
1476 struct inode *inode = file_inode(vmf->vma->vm_file);
1477 struct xfs_inode *ip = XFS_I(inode);
1478 unsigned int lock_mode = XFS_MMAPLOCK_SHARED;
1481 trace_xfs_write_fault(ip, order);
1483 sb_start_pagefault(inode->i_sb);
1484 file_update_time(vmf->vma->vm_file);
1487 * Normally we only need the shared mmaplock, but if a reflink remap is
1488 * in progress we take the exclusive lock to wait for the remap to
1489 * finish before taking a write fault.
1491 xfs_ilock(ip, XFS_MMAPLOCK_SHARED);
1492 if (xfs_iflags_test(ip, XFS_IREMAPPING)) {
1493 xfs_iunlock(ip, XFS_MMAPLOCK_SHARED);
1494 xfs_ilock(ip, XFS_MMAPLOCK_EXCL);
1495 lock_mode = XFS_MMAPLOCK_EXCL;
1499 ret = xfs_dax_fault_locked(vmf, order, true);
1501 ret = iomap_page_mkwrite(vmf, &xfs_buffered_write_iomap_ops);
1502 xfs_iunlock(ip, lock_mode);
1504 sb_end_pagefault(inode->i_sb);
1510 struct vm_fault *vmf)
1512 return (vmf->flags & FAULT_FLAG_WRITE) &&
1513 (vmf->vma->vm_flags & VM_SHARED);
1518 struct vm_fault *vmf)
1520 struct inode *inode = file_inode(vmf->vma->vm_file);
1522 /* DAX can shortcut the normal fault path on write faults! */
1523 if (IS_DAX(inode)) {
1524 if (xfs_is_write_fault(vmf))
1525 return xfs_write_fault(vmf, 0);
1526 return xfs_dax_read_fault(vmf, 0);
1529 trace_xfs_read_fault(XFS_I(inode), 0);
1530 return filemap_fault(vmf);
1534 xfs_filemap_huge_fault(
1535 struct vm_fault *vmf,
1538 if (!IS_DAX(file_inode(vmf->vma->vm_file)))
1539 return VM_FAULT_FALLBACK;
1541 /* DAX can shortcut the normal fault path on write faults! */
1542 if (xfs_is_write_fault(vmf))
1543 return xfs_write_fault(vmf, order);
1544 return xfs_dax_read_fault(vmf, order);
1548 xfs_filemap_page_mkwrite(
1549 struct vm_fault *vmf)
1551 return xfs_write_fault(vmf, 0);
1555 * pfn_mkwrite was originally intended to ensure we capture time stamp updates
1556 * on write faults. In reality, it needs to serialise against truncate and
1557 * prepare memory for writing so handle is as standard write fault.
1560 xfs_filemap_pfn_mkwrite(
1561 struct vm_fault *vmf)
1563 return xfs_write_fault(vmf, 0);
1566 static const struct vm_operations_struct xfs_file_vm_ops = {
1567 .fault = xfs_filemap_fault,
1568 .huge_fault = xfs_filemap_huge_fault,
1569 .map_pages = filemap_map_pages,
1570 .page_mkwrite = xfs_filemap_page_mkwrite,
1571 .pfn_mkwrite = xfs_filemap_pfn_mkwrite,
1577 struct vm_area_struct *vma)
1579 struct inode *inode = file_inode(file);
1580 struct xfs_buftarg *target = xfs_inode_buftarg(XFS_I(inode));
1583 * We don't support synchronous mappings for non-DAX files and
1584 * for DAX files if underneath dax_device is not synchronous.
1586 if (!daxdev_mapping_supported(vma, target->bt_daxdev))
1589 file_accessed(file);
1590 vma->vm_ops = &xfs_file_vm_ops;
1592 vm_flags_set(vma, VM_HUGEPAGE);
1596 const struct file_operations xfs_file_operations = {
1597 .llseek = xfs_file_llseek,
1598 .read_iter = xfs_file_read_iter,
1599 .write_iter = xfs_file_write_iter,
1600 .splice_read = xfs_file_splice_read,
1601 .splice_write = iter_file_splice_write,
1602 .iopoll = iocb_bio_iopoll,
1603 .unlocked_ioctl = xfs_file_ioctl,
1604 #ifdef CONFIG_COMPAT
1605 .compat_ioctl = xfs_file_compat_ioctl,
1607 .mmap = xfs_file_mmap,
1608 .open = xfs_file_open,
1609 .release = xfs_file_release,
1610 .fsync = xfs_file_fsync,
1611 .get_unmapped_area = thp_get_unmapped_area,
1612 .fallocate = xfs_file_fallocate,
1613 .fadvise = xfs_file_fadvise,
1614 .remap_file_range = xfs_file_remap_range,
1615 .fop_flags = FOP_MMAP_SYNC | FOP_BUFFER_RASYNC |
1616 FOP_BUFFER_WASYNC | FOP_DIO_PARALLEL_WRITE,
1619 const struct file_operations xfs_dir_file_operations = {
1620 .open = xfs_dir_open,
1621 .read = generic_read_dir,
1622 .iterate_shared = xfs_file_readdir,
1623 .llseek = generic_file_llseek,
1624 .unlocked_ioctl = xfs_file_ioctl,
1625 #ifdef CONFIG_COMPAT
1626 .compat_ioctl = xfs_file_compat_ioctl,
1628 .fsync = xfs_dir_fsync,