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"
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>
35 static const struct vm_operations_struct xfs_file_vm_ops;
38 * Decide if the given file range is aligned to the size of the fundamental
39 * allocation unit for the file.
42 xfs_is_falloc_aligned(
47 struct xfs_mount *mp = ip->i_mount;
50 if (XFS_IS_REALTIME_INODE(ip)) {
51 if (!is_power_of_2(mp->m_sb.sb_rextsize)) {
55 rextbytes = XFS_FSB_TO_B(mp, mp->m_sb.sb_rextsize);
56 div_u64_rem(pos, rextbytes, &mod);
59 div_u64_rem(len, rextbytes, &mod);
62 mask = XFS_FSB_TO_B(mp, mp->m_sb.sb_rextsize) - 1;
64 mask = mp->m_sb.sb_blocksize - 1;
67 return !((pos | len) & mask);
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.
83 struct xfs_inode *ip = XFS_I(file->f_mapping->host);
85 trace_xfs_dir_fsync(ip);
86 return xfs_log_force_inode(ip);
94 if (!xfs_ipincount(ip))
96 if (datasync && !(ip->i_itemp->ili_fsync_fields & ~XFS_ILOG_TIMESTAMP))
98 return ip->i_itemp->ili_commit_seq;
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.
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.
116 struct xfs_inode *ip,
123 xfs_ilock(ip, XFS_ILOCK_SHARED);
124 seq = xfs_fsync_seq(ip, datasync);
126 error = xfs_log_force_seq(ip->i_mount, seq, XFS_LOG_SYNC,
129 spin_lock(&ip->i_itemp->ili_lock);
130 ip->i_itemp->ili_fsync_fields = 0;
131 spin_unlock(&ip->i_itemp->ili_lock);
133 xfs_iunlock(ip, XFS_ILOCK_SHARED);
144 struct xfs_inode *ip = XFS_I(file->f_mapping->host);
145 struct xfs_mount *mp = ip->i_mount;
149 trace_xfs_file_fsync(ip);
151 error = file_write_and_wait_range(file, start, end);
155 if (xfs_is_shutdown(mp))
158 xfs_iflags_clear(ip, XFS_ITRUNCATED);
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.
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);
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.
177 if (xfs_ipincount(ip)) {
178 err2 = xfs_fsync_flush_log(ip, datasync, &log_flushed);
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
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);
203 unsigned int lock_mode)
205 struct xfs_inode *ip = XFS_I(file_inode(iocb->ki_filp));
207 if (iocb->ki_flags & IOCB_NOWAIT) {
208 if (!xfs_ilock_nowait(ip, lock_mode))
211 xfs_ilock(ip, lock_mode);
218 xfs_ilock_iocb_for_write(
220 unsigned int *lock_mode)
223 struct xfs_inode *ip = XFS_I(file_inode(iocb->ki_filp));
225 ret = xfs_ilock_iocb(iocb, *lock_mode);
229 if (*lock_mode == XFS_IOLOCK_EXCL)
231 if (!xfs_iflags_test(ip, XFS_IREMAPPING))
234 xfs_iunlock(ip, *lock_mode);
235 *lock_mode = XFS_IOLOCK_EXCL;
236 return xfs_ilock_iocb(iocb, *lock_mode);
240 xfs_ilock_for_write_fault(
241 struct xfs_inode *ip)
243 /* get a shared lock if no remapping in progress */
244 xfs_ilock(ip, XFS_MMAPLOCK_SHARED);
245 if (!xfs_iflags_test(ip, XFS_IREMAPPING))
246 return XFS_MMAPLOCK_SHARED;
248 /* wait for remapping to complete */
249 xfs_iunlock(ip, XFS_MMAPLOCK_SHARED);
250 xfs_ilock(ip, XFS_MMAPLOCK_EXCL);
251 return XFS_MMAPLOCK_EXCL;
259 struct xfs_inode *ip = XFS_I(file_inode(iocb->ki_filp));
262 trace_xfs_file_direct_read(iocb, to);
264 if (!iov_iter_count(to))
265 return 0; /* skip atime */
267 file_accessed(iocb->ki_filp);
269 ret = xfs_ilock_iocb(iocb, XFS_IOLOCK_SHARED);
272 ret = iomap_dio_rw(iocb, to, &xfs_read_iomap_ops, NULL, 0, NULL, 0);
273 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
278 static noinline ssize_t
283 struct xfs_inode *ip = XFS_I(iocb->ki_filp->f_mapping->host);
286 trace_xfs_file_dax_read(iocb, to);
288 if (!iov_iter_count(to))
289 return 0; /* skip atime */
291 ret = xfs_ilock_iocb(iocb, XFS_IOLOCK_SHARED);
294 ret = dax_iomap_rw(iocb, to, &xfs_read_iomap_ops);
295 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
297 file_accessed(iocb->ki_filp);
302 xfs_file_buffered_read(
306 struct xfs_inode *ip = XFS_I(file_inode(iocb->ki_filp));
309 trace_xfs_file_buffered_read(iocb, to);
311 ret = xfs_ilock_iocb(iocb, XFS_IOLOCK_SHARED);
314 ret = generic_file_read_iter(iocb, to);
315 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
325 struct inode *inode = file_inode(iocb->ki_filp);
326 struct xfs_mount *mp = XFS_I(inode)->i_mount;
329 XFS_STATS_INC(mp, xs_read_calls);
331 if (xfs_is_shutdown(mp))
335 ret = xfs_file_dax_read(iocb, to);
336 else if (iocb->ki_flags & IOCB_DIRECT)
337 ret = xfs_file_dio_read(iocb, to);
339 ret = xfs_file_buffered_read(iocb, to);
342 XFS_STATS_ADD(mp, xs_read_bytes, ret);
347 xfs_file_splice_read(
350 struct pipe_inode_info *pipe,
354 struct inode *inode = file_inode(in);
355 struct xfs_inode *ip = XFS_I(inode);
356 struct xfs_mount *mp = ip->i_mount;
359 XFS_STATS_INC(mp, xs_read_calls);
361 if (xfs_is_shutdown(mp))
364 trace_xfs_file_splice_read(ip, *ppos, len);
366 xfs_ilock(ip, XFS_IOLOCK_SHARED);
367 ret = filemap_splice_read(in, ppos, pipe, len, flags);
368 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
370 XFS_STATS_ADD(mp, xs_read_bytes, ret);
375 * Common pre-write limit and setup checks.
377 * Called with the iolocked held either shared and exclusive according to
378 * @iolock, and returns with it held. Might upgrade the iolock to exclusive
379 * if called for a direct write beyond i_size.
382 xfs_file_write_checks(
384 struct iov_iter *from,
385 unsigned int *iolock)
387 struct file *file = iocb->ki_filp;
388 struct inode *inode = file->f_mapping->host;
389 struct xfs_inode *ip = XFS_I(inode);
391 size_t count = iov_iter_count(from);
392 bool drained_dio = false;
396 error = generic_write_checks(iocb, from);
400 if (iocb->ki_flags & IOCB_NOWAIT) {
401 error = break_layout(inode, false);
402 if (error == -EWOULDBLOCK)
405 error = xfs_break_layouts(inode, iolock, BREAK_WRITE);
412 * For changing security info in file_remove_privs() we need i_rwsem
415 if (*iolock == XFS_IOLOCK_SHARED && !IS_NOSEC(inode)) {
416 xfs_iunlock(ip, *iolock);
417 *iolock = XFS_IOLOCK_EXCL;
418 error = xfs_ilock_iocb(iocb, *iolock);
427 * If the offset is beyond the size of the file, we need to zero any
428 * blocks that fall between the existing EOF and the start of this
429 * write. If zeroing is needed and we are currently holding the iolock
430 * shared, we need to update it to exclusive which implies having to
431 * redo all checks before.
433 * We need to serialise against EOF updates that occur in IO completions
434 * here. We want to make sure that nobody is changing the size while we
435 * do this check until we have placed an IO barrier (i.e. hold the
436 * XFS_IOLOCK_EXCL) that prevents new IO from being dispatched. The
437 * spinlock effectively forms a memory barrier once we have the
438 * XFS_IOLOCK_EXCL so we are guaranteed to see the latest EOF value and
439 * hence be able to correctly determine if we need to run zeroing.
441 * We can do an unlocked check here safely as IO completion can only
442 * extend EOF. Truncate is locked out at this point, so the EOF can
443 * not move backwards, only forwards. Hence we only need to take the
444 * slow path and spin locks when we are at or beyond the current EOF.
446 if (iocb->ki_pos <= i_size_read(inode))
449 spin_lock(&ip->i_flags_lock);
450 isize = i_size_read(inode);
451 if (iocb->ki_pos > isize) {
452 spin_unlock(&ip->i_flags_lock);
454 if (iocb->ki_flags & IOCB_NOWAIT)
458 if (*iolock == XFS_IOLOCK_SHARED) {
459 xfs_iunlock(ip, *iolock);
460 *iolock = XFS_IOLOCK_EXCL;
461 xfs_ilock(ip, *iolock);
462 iov_iter_reexpand(from, count);
465 * We now have an IO submission barrier in place, but
466 * AIO can do EOF updates during IO completion and hence
467 * we now need to wait for all of them to drain. Non-AIO
468 * DIO will have drained before we are given the
469 * XFS_IOLOCK_EXCL, and so for most cases this wait is a
472 inode_dio_wait(inode);
477 trace_xfs_zero_eof(ip, isize, iocb->ki_pos - isize);
478 error = xfs_zero_range(ip, isize, iocb->ki_pos - isize, NULL);
482 spin_unlock(&ip->i_flags_lock);
485 return kiocb_modified(iocb);
489 xfs_dio_write_end_io(
495 struct inode *inode = file_inode(iocb->ki_filp);
496 struct xfs_inode *ip = XFS_I(inode);
497 loff_t offset = iocb->ki_pos;
498 unsigned int nofs_flag;
500 trace_xfs_end_io_direct_write(ip, offset, size);
502 if (xfs_is_shutdown(ip->i_mount))
511 * Capture amount written on completion as we can't reliably account
512 * for it on submission.
514 XFS_STATS_ADD(ip->i_mount, xs_write_bytes, size);
517 * We can allocate memory here while doing writeback on behalf of
518 * memory reclaim. To avoid memory allocation deadlocks set the
519 * task-wide nofs context for the following operations.
521 nofs_flag = memalloc_nofs_save();
523 if (flags & IOMAP_DIO_COW) {
524 error = xfs_reflink_end_cow(ip, offset, size);
530 * Unwritten conversion updates the in-core isize after extent
531 * conversion but before updating the on-disk size. Updating isize any
532 * earlier allows a racing dio read to find unwritten extents before
533 * they are converted.
535 if (flags & IOMAP_DIO_UNWRITTEN) {
536 error = xfs_iomap_write_unwritten(ip, offset, size, true);
541 * We need to update the in-core inode size here so that we don't end up
542 * with the on-disk inode size being outside the in-core inode size. We
543 * have no other method of updating EOF for AIO, so always do it here
546 * We need to lock the test/set EOF update as we can be racing with
547 * other IO completions here to update the EOF. Failing to serialise
548 * here can result in EOF moving backwards and Bad Things Happen when
551 * As IO completion only ever extends EOF, we can do an unlocked check
552 * here to avoid taking the spinlock. If we land within the current EOF,
553 * then we do not need to do an extending update at all, and we don't
554 * need to take the lock to check this. If we race with an update moving
555 * EOF, then we'll either still be beyond EOF and need to take the lock,
556 * or we'll be within EOF and we don't need to take it at all.
558 if (offset + size <= i_size_read(inode))
561 spin_lock(&ip->i_flags_lock);
562 if (offset + size > i_size_read(inode)) {
563 i_size_write(inode, offset + size);
564 spin_unlock(&ip->i_flags_lock);
565 error = xfs_setfilesize(ip, offset, size);
567 spin_unlock(&ip->i_flags_lock);
571 memalloc_nofs_restore(nofs_flag);
575 static const struct iomap_dio_ops xfs_dio_write_ops = {
576 .end_io = xfs_dio_write_end_io,
580 * Handle block aligned direct I/O writes
582 static noinline ssize_t
583 xfs_file_dio_write_aligned(
584 struct xfs_inode *ip,
586 struct iov_iter *from)
588 unsigned int iolock = XFS_IOLOCK_SHARED;
591 ret = xfs_ilock_iocb_for_write(iocb, &iolock);
594 ret = xfs_file_write_checks(iocb, from, &iolock);
599 * We don't need to hold the IOLOCK exclusively across the IO, so demote
600 * the iolock back to shared if we had to take the exclusive lock in
601 * xfs_file_write_checks() for other reasons.
603 if (iolock == XFS_IOLOCK_EXCL) {
604 xfs_ilock_demote(ip, XFS_IOLOCK_EXCL);
605 iolock = XFS_IOLOCK_SHARED;
607 trace_xfs_file_direct_write(iocb, from);
608 ret = iomap_dio_rw(iocb, from, &xfs_direct_write_iomap_ops,
609 &xfs_dio_write_ops, 0, NULL, 0);
612 xfs_iunlock(ip, iolock);
617 * Handle block unaligned direct I/O writes
619 * In most cases direct I/O writes will be done holding IOLOCK_SHARED, allowing
620 * them to be done in parallel with reads and other direct I/O writes. However,
621 * if the I/O is not aligned to filesystem blocks, the direct I/O layer may need
622 * to do sub-block zeroing and that requires serialisation against other direct
623 * I/O to the same block. In this case we need to serialise the submission of
624 * the unaligned I/O so that we don't get racing block zeroing in the dio layer.
625 * In the case where sub-block zeroing is not required, we can do concurrent
626 * sub-block dios to the same block successfully.
628 * Optimistically submit the I/O using the shared lock first, but use the
629 * IOMAP_DIO_OVERWRITE_ONLY flag to tell the lower layers to return -EAGAIN
630 * if block allocation or partial block zeroing would be required. In that case
631 * we try again with the exclusive lock.
633 static noinline ssize_t
634 xfs_file_dio_write_unaligned(
635 struct xfs_inode *ip,
637 struct iov_iter *from)
639 size_t isize = i_size_read(VFS_I(ip));
640 size_t count = iov_iter_count(from);
641 unsigned int iolock = XFS_IOLOCK_SHARED;
642 unsigned int flags = IOMAP_DIO_OVERWRITE_ONLY;
646 * Extending writes need exclusivity because of the sub-block zeroing
647 * that the DIO code always does for partial tail blocks beyond EOF, so
648 * don't even bother trying the fast path in this case.
650 if (iocb->ki_pos > isize || iocb->ki_pos + count >= isize) {
651 if (iocb->ki_flags & IOCB_NOWAIT)
654 iolock = XFS_IOLOCK_EXCL;
655 flags = IOMAP_DIO_FORCE_WAIT;
658 ret = xfs_ilock_iocb_for_write(iocb, &iolock);
663 * We can't properly handle unaligned direct I/O to reflink files yet,
664 * as we can't unshare a partial block.
666 if (xfs_is_cow_inode(ip)) {
667 trace_xfs_reflink_bounce_dio_write(iocb, from);
672 ret = xfs_file_write_checks(iocb, from, &iolock);
677 * If we are doing exclusive unaligned I/O, this must be the only I/O
678 * in-flight. Otherwise we risk data corruption due to unwritten extent
679 * conversions from the AIO end_io handler. Wait for all other I/O to
682 if (flags & IOMAP_DIO_FORCE_WAIT)
683 inode_dio_wait(VFS_I(ip));
685 trace_xfs_file_direct_write(iocb, from);
686 ret = iomap_dio_rw(iocb, from, &xfs_direct_write_iomap_ops,
687 &xfs_dio_write_ops, flags, NULL, 0);
690 * Retry unaligned I/O with exclusive blocking semantics if the DIO
691 * layer rejected it for mapping or locking reasons. If we are doing
692 * nonblocking user I/O, propagate the error.
694 if (ret == -EAGAIN && !(iocb->ki_flags & IOCB_NOWAIT)) {
695 ASSERT(flags & IOMAP_DIO_OVERWRITE_ONLY);
696 xfs_iunlock(ip, iolock);
697 goto retry_exclusive;
702 xfs_iunlock(ip, iolock);
709 struct iov_iter *from)
711 struct xfs_inode *ip = XFS_I(file_inode(iocb->ki_filp));
712 struct xfs_buftarg *target = xfs_inode_buftarg(ip);
713 size_t count = iov_iter_count(from);
715 /* direct I/O must be aligned to device logical sector size */
716 if ((iocb->ki_pos | count) & target->bt_logical_sectormask)
718 if ((iocb->ki_pos | count) & ip->i_mount->m_blockmask)
719 return xfs_file_dio_write_unaligned(ip, iocb, from);
720 return xfs_file_dio_write_aligned(ip, iocb, from);
723 static noinline ssize_t
726 struct iov_iter *from)
728 struct inode *inode = iocb->ki_filp->f_mapping->host;
729 struct xfs_inode *ip = XFS_I(inode);
730 unsigned int iolock = XFS_IOLOCK_EXCL;
731 ssize_t ret, error = 0;
734 ret = xfs_ilock_iocb(iocb, iolock);
737 ret = xfs_file_write_checks(iocb, from, &iolock);
743 trace_xfs_file_dax_write(iocb, from);
744 ret = dax_iomap_rw(iocb, from, &xfs_dax_write_iomap_ops);
745 if (ret > 0 && iocb->ki_pos > i_size_read(inode)) {
746 i_size_write(inode, iocb->ki_pos);
747 error = xfs_setfilesize(ip, pos, ret);
751 xfs_iunlock(ip, iolock);
756 XFS_STATS_ADD(ip->i_mount, xs_write_bytes, ret);
758 /* Handle various SYNC-type writes */
759 ret = generic_write_sync(iocb, ret);
765 xfs_file_buffered_write(
767 struct iov_iter *from)
769 struct inode *inode = iocb->ki_filp->f_mapping->host;
770 struct xfs_inode *ip = XFS_I(inode);
772 bool cleared_space = false;
776 iolock = XFS_IOLOCK_EXCL;
777 ret = xfs_ilock_iocb(iocb, iolock);
781 ret = xfs_file_write_checks(iocb, from, &iolock);
785 trace_xfs_file_buffered_write(iocb, from);
786 ret = iomap_file_buffered_write(iocb, from,
787 &xfs_buffered_write_iomap_ops);
790 * If we hit a space limit, try to free up some lingering preallocated
791 * space before returning an error. In the case of ENOSPC, first try to
792 * write back all dirty inodes to free up some of the excess reserved
793 * metadata space. This reduces the chances that the eofblocks scan
794 * waits on dirty mappings. Since xfs_flush_inodes() is serialized, this
795 * also behaves as a filter to prevent too many eofblocks scans from
796 * running at the same time. Use a synchronous scan to increase the
797 * effectiveness of the scan.
799 if (ret == -EDQUOT && !cleared_space) {
800 xfs_iunlock(ip, iolock);
801 xfs_blockgc_free_quota(ip, XFS_ICWALK_FLAG_SYNC);
802 cleared_space = true;
804 } else if (ret == -ENOSPC && !cleared_space) {
805 struct xfs_icwalk icw = {0};
807 cleared_space = true;
808 xfs_flush_inodes(ip->i_mount);
810 xfs_iunlock(ip, iolock);
811 icw.icw_flags = XFS_ICWALK_FLAG_SYNC;
812 xfs_blockgc_free_space(ip->i_mount, &icw);
818 xfs_iunlock(ip, iolock);
821 XFS_STATS_ADD(ip->i_mount, xs_write_bytes, ret);
822 /* Handle various SYNC-type writes */
823 ret = generic_write_sync(iocb, ret);
831 struct iov_iter *from)
833 struct inode *inode = iocb->ki_filp->f_mapping->host;
834 struct xfs_inode *ip = XFS_I(inode);
836 size_t ocount = iov_iter_count(from);
838 XFS_STATS_INC(ip->i_mount, xs_write_calls);
843 if (xfs_is_shutdown(ip->i_mount))
847 return xfs_file_dax_write(iocb, from);
849 if (iocb->ki_flags & IOCB_DIRECT) {
851 * Allow a directio write to fall back to a buffered
852 * write *only* in the case that we're doing a reflink
853 * CoW. In all other directio scenarios we do not
854 * allow an operation to fall back to buffered mode.
856 ret = xfs_file_dio_write(iocb, from);
861 return xfs_file_buffered_write(iocb, from);
868 struct xfs_inode *ip = XFS_I(inode);
870 xfs_iunlock(ip, XFS_MMAPLOCK_EXCL);
872 xfs_ilock(ip, XFS_MMAPLOCK_EXCL);
876 xfs_break_dax_layouts(
882 ASSERT(xfs_isilocked(XFS_I(inode), XFS_MMAPLOCK_EXCL));
884 page = dax_layout_busy_page(inode->i_mapping);
889 return ___wait_var_event(&page->_refcount,
890 atomic_read(&page->_refcount) == 1, TASK_INTERRUPTIBLE,
891 0, 0, xfs_wait_dax_page(inode));
898 enum layout_break_reason reason)
903 ASSERT(xfs_isilocked(XFS_I(inode), XFS_IOLOCK_SHARED|XFS_IOLOCK_EXCL));
909 error = xfs_break_dax_layouts(inode, &retry);
914 error = xfs_break_leased_layouts(inode, iolock, &retry);
920 } while (error == 0 && retry);
925 /* Does this file, inode, or mount want synchronous writes? */
926 static inline bool xfs_file_sync_writes(struct file *filp)
928 struct xfs_inode *ip = XFS_I(file_inode(filp));
930 if (xfs_has_wsync(ip->i_mount))
932 if (filp->f_flags & (__O_SYNC | O_DSYNC))
934 if (IS_SYNC(file_inode(filp)))
940 #define XFS_FALLOC_FL_SUPPORTED \
941 (FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE | \
942 FALLOC_FL_COLLAPSE_RANGE | FALLOC_FL_ZERO_RANGE | \
943 FALLOC_FL_INSERT_RANGE | FALLOC_FL_UNSHARE_RANGE)
952 struct inode *inode = file_inode(file);
953 struct xfs_inode *ip = XFS_I(inode);
955 uint iolock = XFS_IOLOCK_EXCL | XFS_MMAPLOCK_EXCL;
957 bool do_file_insert = false;
959 if (!S_ISREG(inode->i_mode))
961 if (mode & ~XFS_FALLOC_FL_SUPPORTED)
964 xfs_ilock(ip, iolock);
965 error = xfs_break_layouts(inode, &iolock, BREAK_UNMAP);
970 * Must wait for all AIO to complete before we continue as AIO can
971 * change the file size on completion without holding any locks we
972 * currently hold. We must do this first because AIO can update both
973 * the on disk and in memory inode sizes, and the operations that follow
974 * require the in-memory size to be fully up-to-date.
976 inode_dio_wait(inode);
979 * Now AIO and DIO has drained we flush and (if necessary) invalidate
980 * the cached range over the first operation we are about to run.
982 * We care about zero and collapse here because they both run a hole
983 * punch over the range first. Because that can zero data, and the range
984 * of invalidation for the shift operations is much larger, we still do
985 * the required flush for collapse in xfs_prepare_shift().
987 * Insert has the same range requirements as collapse, and we extend the
988 * file first which can zero data. Hence insert has the same
989 * flush/invalidate requirements as collapse and so they are both
990 * handled at the right time by xfs_prepare_shift().
992 if (mode & (FALLOC_FL_PUNCH_HOLE | FALLOC_FL_ZERO_RANGE |
993 FALLOC_FL_COLLAPSE_RANGE)) {
994 error = xfs_flush_unmap_range(ip, offset, len);
999 error = file_modified(file);
1003 if (mode & FALLOC_FL_PUNCH_HOLE) {
1004 error = xfs_free_file_space(ip, offset, len);
1007 } else if (mode & FALLOC_FL_COLLAPSE_RANGE) {
1008 if (!xfs_is_falloc_aligned(ip, offset, len)) {
1014 * There is no need to overlap collapse range with EOF,
1015 * in which case it is effectively a truncate operation
1017 if (offset + len >= i_size_read(inode)) {
1022 new_size = i_size_read(inode) - len;
1024 error = xfs_collapse_file_space(ip, offset, len);
1027 } else if (mode & FALLOC_FL_INSERT_RANGE) {
1028 loff_t isize = i_size_read(inode);
1030 if (!xfs_is_falloc_aligned(ip, offset, len)) {
1036 * New inode size must not exceed ->s_maxbytes, accounting for
1037 * possible signed overflow.
1039 if (inode->i_sb->s_maxbytes - isize < len) {
1043 new_size = isize + len;
1045 /* Offset should be less than i_size */
1046 if (offset >= isize) {
1050 do_file_insert = true;
1052 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
1053 offset + len > i_size_read(inode)) {
1054 new_size = offset + len;
1055 error = inode_newsize_ok(inode, new_size);
1060 if (mode & FALLOC_FL_ZERO_RANGE) {
1062 * Punch a hole and prealloc the range. We use a hole
1063 * punch rather than unwritten extent conversion for two
1066 * 1.) Hole punch handles partial block zeroing for us.
1067 * 2.) If prealloc returns ENOSPC, the file range is
1068 * still zero-valued by virtue of the hole punch.
1070 unsigned int blksize = i_blocksize(inode);
1072 trace_xfs_zero_file_space(ip);
1074 error = xfs_free_file_space(ip, offset, len);
1078 len = round_up(offset + len, blksize) -
1079 round_down(offset, blksize);
1080 offset = round_down(offset, blksize);
1081 } else if (mode & FALLOC_FL_UNSHARE_RANGE) {
1082 error = xfs_reflink_unshare(ip, offset, len);
1087 * If always_cow mode we can't use preallocations and
1088 * thus should not create them.
1090 if (xfs_is_always_cow_inode(ip)) {
1091 error = -EOPNOTSUPP;
1096 if (!xfs_is_always_cow_inode(ip)) {
1097 error = xfs_alloc_file_space(ip, offset, len);
1103 /* Change file size if needed */
1107 iattr.ia_valid = ATTR_SIZE;
1108 iattr.ia_size = new_size;
1109 error = xfs_vn_setattr_size(file_mnt_idmap(file),
1110 file_dentry(file), &iattr);
1116 * Perform hole insertion now that the file size has been
1117 * updated so that if we crash during the operation we don't
1118 * leave shifted extents past EOF and hence losing access to
1119 * the data that is contained within them.
1121 if (do_file_insert) {
1122 error = xfs_insert_file_space(ip, offset, len);
1127 if (xfs_file_sync_writes(file))
1128 error = xfs_log_force_inode(ip);
1131 xfs_iunlock(ip, iolock);
1142 struct xfs_inode *ip = XFS_I(file_inode(file));
1147 * Operations creating pages in page cache need protection from hole
1148 * punching and similar ops
1150 if (advice == POSIX_FADV_WILLNEED) {
1151 lockflags = XFS_IOLOCK_SHARED;
1152 xfs_ilock(ip, lockflags);
1154 ret = generic_fadvise(file, start, end, advice);
1156 xfs_iunlock(ip, lockflags);
1161 xfs_file_remap_range(
1162 struct file *file_in,
1164 struct file *file_out,
1167 unsigned int remap_flags)
1169 struct inode *inode_in = file_inode(file_in);
1170 struct xfs_inode *src = XFS_I(inode_in);
1171 struct inode *inode_out = file_inode(file_out);
1172 struct xfs_inode *dest = XFS_I(inode_out);
1173 struct xfs_mount *mp = src->i_mount;
1174 loff_t remapped = 0;
1175 xfs_extlen_t cowextsize;
1178 if (remap_flags & ~(REMAP_FILE_DEDUP | REMAP_FILE_ADVISORY))
1181 if (!xfs_has_reflink(mp))
1184 if (xfs_is_shutdown(mp))
1187 /* Prepare and then clone file data. */
1188 ret = xfs_reflink_remap_prep(file_in, pos_in, file_out, pos_out,
1190 if (ret || len == 0)
1193 trace_xfs_reflink_remap_range(src, pos_in, len, dest, pos_out);
1195 ret = xfs_reflink_remap_blocks(src, pos_in, dest, pos_out, len,
1201 * Carry the cowextsize hint from src to dest if we're sharing the
1202 * entire source file to the entire destination file, the source file
1203 * has a cowextsize hint, and the destination file does not.
1206 if (pos_in == 0 && len == i_size_read(inode_in) &&
1207 (src->i_diflags2 & XFS_DIFLAG2_COWEXTSIZE) &&
1208 pos_out == 0 && len >= i_size_read(inode_out) &&
1209 !(dest->i_diflags2 & XFS_DIFLAG2_COWEXTSIZE))
1210 cowextsize = src->i_cowextsize;
1212 ret = xfs_reflink_update_dest(dest, pos_out + len, cowextsize,
1217 if (xfs_file_sync_writes(file_in) || xfs_file_sync_writes(file_out))
1218 xfs_log_force_inode(dest);
1220 xfs_iunlock2_remapping(src, dest);
1222 trace_xfs_reflink_remap_range_error(dest, ret, _RET_IP_);
1223 return remapped > 0 ? remapped : ret;
1228 struct inode *inode,
1231 if (xfs_is_shutdown(XFS_M(inode->i_sb)))
1233 file->f_mode |= FMODE_NOWAIT | FMODE_BUF_RASYNC | FMODE_BUF_WASYNC |
1234 FMODE_DIO_PARALLEL_WRITE | FMODE_CAN_ODIRECT;
1235 return generic_file_open(inode, file);
1240 struct inode *inode,
1243 struct xfs_inode *ip = XFS_I(inode);
1247 error = xfs_file_open(inode, file);
1252 * If there are any blocks, read-ahead block 0 as we're almost
1253 * certain to have the next operation be a read there.
1255 mode = xfs_ilock_data_map_shared(ip);
1256 if (ip->i_df.if_nextents > 0)
1257 error = xfs_dir3_data_readahead(ip, 0, 0);
1258 xfs_iunlock(ip, mode);
1264 struct inode *inode,
1267 return xfs_release(XFS_I(inode));
1273 struct dir_context *ctx)
1275 struct inode *inode = file_inode(file);
1276 xfs_inode_t *ip = XFS_I(inode);
1280 * The Linux API doesn't pass down the total size of the buffer
1281 * we read into down to the filesystem. With the filldir concept
1282 * it's not needed for correct information, but the XFS dir2 leaf
1283 * code wants an estimate of the buffer size to calculate it's
1284 * readahead window and size the buffers used for mapping to
1287 * Try to give it an estimate that's good enough, maybe at some
1288 * point we can change the ->readdir prototype to include the
1289 * buffer size. For now we use the current glibc buffer size.
1291 bufsize = (size_t)min_t(loff_t, XFS_READDIR_BUFSIZE, ip->i_disk_size);
1293 return xfs_readdir(NULL, ip, ctx, bufsize);
1302 struct inode *inode = file->f_mapping->host;
1304 if (xfs_is_shutdown(XFS_I(inode)->i_mount))
1309 return generic_file_llseek(file, offset, whence);
1311 offset = iomap_seek_hole(inode, offset, &xfs_seek_iomap_ops);
1314 offset = iomap_seek_data(inode, offset, &xfs_seek_iomap_ops);
1320 return vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
1323 #ifdef CONFIG_FS_DAX
1324 static inline vm_fault_t
1326 struct vm_fault *vmf,
1331 return dax_iomap_fault(vmf, order, pfn, NULL,
1332 (write_fault && !vmf->cow_page) ?
1333 &xfs_dax_write_iomap_ops :
1334 &xfs_read_iomap_ops);
1337 static inline vm_fault_t
1339 struct vm_fault *vmf,
1345 return VM_FAULT_SIGBUS;
1350 * Locking for serialisation of IO during page faults. This results in a lock
1354 * sb_start_pagefault(vfs, freeze)
1355 * invalidate_lock (vfs/XFS_MMAPLOCK - truncate serialisation)
1357 * i_lock (XFS - extent map serialisation)
1360 __xfs_filemap_fault(
1361 struct vm_fault *vmf,
1365 struct inode *inode = file_inode(vmf->vma->vm_file);
1366 struct xfs_inode *ip = XFS_I(inode);
1368 unsigned int lock_mode = 0;
1370 trace_xfs_filemap_fault(ip, order, write_fault);
1373 sb_start_pagefault(inode->i_sb);
1374 file_update_time(vmf->vma->vm_file);
1377 if (IS_DAX(inode) || write_fault)
1378 lock_mode = xfs_ilock_for_write_fault(XFS_I(inode));
1380 if (IS_DAX(inode)) {
1383 ret = xfs_dax_fault(vmf, order, write_fault, &pfn);
1384 if (ret & VM_FAULT_NEEDDSYNC)
1385 ret = dax_finish_sync_fault(vmf, order, pfn);
1386 } else if (write_fault) {
1387 ret = iomap_page_mkwrite(vmf, &xfs_page_mkwrite_iomap_ops);
1389 ret = filemap_fault(vmf);
1393 xfs_iunlock(XFS_I(inode), lock_mode);
1396 sb_end_pagefault(inode->i_sb);
1402 struct vm_fault *vmf)
1404 return (vmf->flags & FAULT_FLAG_WRITE) &&
1405 (vmf->vma->vm_flags & VM_SHARED);
1410 struct vm_fault *vmf)
1412 /* DAX can shortcut the normal fault path on write faults! */
1413 return __xfs_filemap_fault(vmf, 0,
1414 IS_DAX(file_inode(vmf->vma->vm_file)) &&
1415 xfs_is_write_fault(vmf));
1419 xfs_filemap_huge_fault(
1420 struct vm_fault *vmf,
1423 if (!IS_DAX(file_inode(vmf->vma->vm_file)))
1424 return VM_FAULT_FALLBACK;
1426 /* DAX can shortcut the normal fault path on write faults! */
1427 return __xfs_filemap_fault(vmf, order,
1428 xfs_is_write_fault(vmf));
1432 xfs_filemap_page_mkwrite(
1433 struct vm_fault *vmf)
1435 return __xfs_filemap_fault(vmf, 0, true);
1439 * pfn_mkwrite was originally intended to ensure we capture time stamp updates
1440 * on write faults. In reality, it needs to serialise against truncate and
1441 * prepare memory for writing so handle is as standard write fault.
1444 xfs_filemap_pfn_mkwrite(
1445 struct vm_fault *vmf)
1448 return __xfs_filemap_fault(vmf, 0, true);
1451 static const struct vm_operations_struct xfs_file_vm_ops = {
1452 .fault = xfs_filemap_fault,
1453 .huge_fault = xfs_filemap_huge_fault,
1454 .map_pages = filemap_map_pages,
1455 .page_mkwrite = xfs_filemap_page_mkwrite,
1456 .pfn_mkwrite = xfs_filemap_pfn_mkwrite,
1462 struct vm_area_struct *vma)
1464 struct inode *inode = file_inode(file);
1465 struct xfs_buftarg *target = xfs_inode_buftarg(XFS_I(inode));
1468 * We don't support synchronous mappings for non-DAX files and
1469 * for DAX files if underneath dax_device is not synchronous.
1471 if (!daxdev_mapping_supported(vma, target->bt_daxdev))
1474 file_accessed(file);
1475 vma->vm_ops = &xfs_file_vm_ops;
1477 vm_flags_set(vma, VM_HUGEPAGE);
1481 const struct file_operations xfs_file_operations = {
1482 .llseek = xfs_file_llseek,
1483 .read_iter = xfs_file_read_iter,
1484 .write_iter = xfs_file_write_iter,
1485 .splice_read = xfs_file_splice_read,
1486 .splice_write = iter_file_splice_write,
1487 .iopoll = iocb_bio_iopoll,
1488 .unlocked_ioctl = xfs_file_ioctl,
1489 #ifdef CONFIG_COMPAT
1490 .compat_ioctl = xfs_file_compat_ioctl,
1492 .mmap = xfs_file_mmap,
1493 .mmap_supported_flags = MAP_SYNC,
1494 .open = xfs_file_open,
1495 .release = xfs_file_release,
1496 .fsync = xfs_file_fsync,
1497 .get_unmapped_area = thp_get_unmapped_area,
1498 .fallocate = xfs_file_fallocate,
1499 .fadvise = xfs_file_fadvise,
1500 .remap_file_range = xfs_file_remap_range,
1503 const struct file_operations xfs_dir_file_operations = {
1504 .open = xfs_dir_open,
1505 .read = generic_read_dir,
1506 .iterate_shared = xfs_file_readdir,
1507 .llseek = generic_file_llseek,
1508 .unlocked_ioctl = xfs_file_ioctl,
1509 #ifdef CONFIG_COMPAT
1510 .compat_ioctl = xfs_file_compat_ioctl,
1512 .fsync = xfs_dir_fsync,