2 * Copyright (c) 2000-2005 Silicon Graphics, Inc.
5 * This program is free software; you can redistribute it and/or
6 * modify it under the terms of the GNU General Public License as
7 * published by the Free Software Foundation.
9 * This program is distributed in the hope that it would be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write the Free Software Foundation,
16 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
20 #include "xfs_shared.h"
21 #include "xfs_format.h"
22 #include "xfs_log_format.h"
23 #include "xfs_trans_resv.h"
24 #include "xfs_mount.h"
25 #include "xfs_da_format.h"
26 #include "xfs_da_btree.h"
27 #include "xfs_inode.h"
28 #include "xfs_trans.h"
29 #include "xfs_inode_item.h"
31 #include "xfs_bmap_util.h"
32 #include "xfs_error.h"
34 #include "xfs_dir2_priv.h"
35 #include "xfs_ioctl.h"
36 #include "xfs_trace.h"
38 #include "xfs_icache.h"
40 #include "xfs_iomap.h"
42 #include <linux/dcache.h>
43 #include <linux/falloc.h>
44 #include <linux/pagevec.h>
45 #include <linux/backing-dev.h>
47 static const struct vm_operations_struct xfs_file_vm_ops
;
50 * Locking primitives for read and write IO paths to ensure we consistently use
51 * and order the inode->i_mutex, ip->i_lock and ip->i_iolock.
58 if (type
& XFS_IOLOCK_EXCL
)
59 inode_lock(VFS_I(ip
));
68 xfs_iunlock(ip
, type
);
69 if (type
& XFS_IOLOCK_EXCL
)
70 inode_unlock(VFS_I(ip
));
78 xfs_ilock_demote(ip
, type
);
79 if (type
& XFS_IOLOCK_EXCL
)
80 inode_unlock(VFS_I(ip
));
84 * Clear the specified ranges to zero through either the pagecache or DAX.
85 * Holes and unwritten extents will be left as-is as they already are zeroed.
94 return iomap_zero_range(VFS_I(ip
), pos
, count
, NULL
, &xfs_iomap_ops
);
98 xfs_update_prealloc_flags(
100 enum xfs_prealloc_flags flags
)
102 struct xfs_trans
*tp
;
105 error
= xfs_trans_alloc(ip
->i_mount
, &M_RES(ip
->i_mount
)->tr_writeid
,
110 xfs_ilock(ip
, XFS_ILOCK_EXCL
);
111 xfs_trans_ijoin(tp
, ip
, XFS_ILOCK_EXCL
);
113 if (!(flags
& XFS_PREALLOC_INVISIBLE
)) {
114 VFS_I(ip
)->i_mode
&= ~S_ISUID
;
115 if (VFS_I(ip
)->i_mode
& S_IXGRP
)
116 VFS_I(ip
)->i_mode
&= ~S_ISGID
;
117 xfs_trans_ichgtime(tp
, ip
, XFS_ICHGTIME_MOD
| XFS_ICHGTIME_CHG
);
120 if (flags
& XFS_PREALLOC_SET
)
121 ip
->i_d
.di_flags
|= XFS_DIFLAG_PREALLOC
;
122 if (flags
& XFS_PREALLOC_CLEAR
)
123 ip
->i_d
.di_flags
&= ~XFS_DIFLAG_PREALLOC
;
125 xfs_trans_log_inode(tp
, ip
, XFS_ILOG_CORE
);
126 if (flags
& XFS_PREALLOC_SYNC
)
127 xfs_trans_set_sync(tp
);
128 return xfs_trans_commit(tp
);
132 * Fsync operations on directories are much simpler than on regular files,
133 * as there is no file data to flush, and thus also no need for explicit
134 * cache flush operations, and there are no non-transaction metadata updates
135 * on directories either.
144 struct xfs_inode
*ip
= XFS_I(file
->f_mapping
->host
);
145 struct xfs_mount
*mp
= ip
->i_mount
;
148 trace_xfs_dir_fsync(ip
);
150 xfs_ilock(ip
, XFS_ILOCK_SHARED
);
151 if (xfs_ipincount(ip
))
152 lsn
= ip
->i_itemp
->ili_last_lsn
;
153 xfs_iunlock(ip
, XFS_ILOCK_SHARED
);
157 return _xfs_log_force_lsn(mp
, lsn
, XFS_LOG_SYNC
, NULL
);
167 struct inode
*inode
= file
->f_mapping
->host
;
168 struct xfs_inode
*ip
= XFS_I(inode
);
169 struct xfs_mount
*mp
= ip
->i_mount
;
174 trace_xfs_file_fsync(ip
);
176 error
= filemap_write_and_wait_range(inode
->i_mapping
, start
, end
);
180 if (XFS_FORCED_SHUTDOWN(mp
))
183 xfs_iflags_clear(ip
, XFS_ITRUNCATED
);
185 if (mp
->m_flags
& XFS_MOUNT_BARRIER
) {
187 * If we have an RT and/or log subvolume we need to make sure
188 * to flush the write cache the device used for file data
189 * first. This is to ensure newly written file data make
190 * it to disk before logging the new inode size in case of
191 * an extending write.
193 if (XFS_IS_REALTIME_INODE(ip
))
194 xfs_blkdev_issue_flush(mp
->m_rtdev_targp
);
195 else if (mp
->m_logdev_targp
!= mp
->m_ddev_targp
)
196 xfs_blkdev_issue_flush(mp
->m_ddev_targp
);
200 * All metadata updates are logged, which means that we just have to
201 * flush the log up to the latest LSN that touched the inode. If we have
202 * concurrent fsync/fdatasync() calls, we need them to all block on the
203 * log force before we clear the ili_fsync_fields field. This ensures
204 * that we don't get a racing sync operation that does not wait for the
205 * metadata to hit the journal before returning. If we race with
206 * clearing the ili_fsync_fields, then all that will happen is the log
207 * force will do nothing as the lsn will already be on disk. We can't
208 * race with setting ili_fsync_fields because that is done under
209 * XFS_ILOCK_EXCL, and that can't happen because we hold the lock shared
210 * until after the ili_fsync_fields is cleared.
212 xfs_ilock(ip
, XFS_ILOCK_SHARED
);
213 if (xfs_ipincount(ip
)) {
215 (ip
->i_itemp
->ili_fsync_fields
& ~XFS_ILOG_TIMESTAMP
))
216 lsn
= ip
->i_itemp
->ili_last_lsn
;
220 error
= _xfs_log_force_lsn(mp
, lsn
, XFS_LOG_SYNC
, &log_flushed
);
221 ip
->i_itemp
->ili_fsync_fields
= 0;
223 xfs_iunlock(ip
, XFS_ILOCK_SHARED
);
226 * If we only have a single device, and the log force about was
227 * a no-op we might have to flush the data device cache here.
228 * This can only happen for fdatasync/O_DSYNC if we were overwriting
229 * an already allocated file and thus do not have any metadata to
232 if ((mp
->m_flags
& XFS_MOUNT_BARRIER
) &&
233 mp
->m_logdev_targp
== mp
->m_ddev_targp
&&
234 !XFS_IS_REALTIME_INODE(ip
) &&
236 xfs_blkdev_issue_flush(mp
->m_ddev_targp
);
242 xfs_file_dio_aio_read(
246 struct address_space
*mapping
= iocb
->ki_filp
->f_mapping
;
247 struct inode
*inode
= mapping
->host
;
248 struct xfs_inode
*ip
= XFS_I(inode
);
249 loff_t isize
= i_size_read(inode
);
250 size_t count
= iov_iter_count(to
);
251 struct iov_iter data
;
252 struct xfs_buftarg
*target
;
255 trace_xfs_file_direct_read(ip
, count
, iocb
->ki_pos
);
258 return 0; /* skip atime */
260 if (XFS_IS_REALTIME_INODE(ip
))
261 target
= ip
->i_mount
->m_rtdev_targp
;
263 target
= ip
->i_mount
->m_ddev_targp
;
265 /* DIO must be aligned to device logical sector size */
266 if ((iocb
->ki_pos
| count
) & target
->bt_logical_sectormask
) {
267 if (iocb
->ki_pos
== isize
)
273 * Locking is a bit tricky here. If we take an exclusive lock for direct
274 * IO, we effectively serialise all new concurrent read IO to this file
275 * and block it behind IO that is currently in progress because IO in
276 * progress holds the IO lock shared. We only need to hold the lock
277 * exclusive to blow away the page cache, so only take lock exclusively
278 * if the page cache needs invalidation. This allows the normal direct
279 * IO case of no page cache pages to proceeed concurrently without
282 xfs_rw_ilock(ip
, XFS_IOLOCK_SHARED
);
283 if (mapping
->nrpages
) {
284 xfs_rw_iunlock(ip
, XFS_IOLOCK_SHARED
);
285 xfs_rw_ilock(ip
, XFS_IOLOCK_EXCL
);
288 * The generic dio code only flushes the range of the particular
289 * I/O. Because we take an exclusive lock here, this whole
290 * sequence is considerably more expensive for us. This has a
291 * noticeable performance impact for any file with cached pages,
292 * even when outside of the range of the particular I/O.
294 * Hence, amortize the cost of the lock against a full file
295 * flush and reduce the chances of repeated iolock cycles going
298 if (mapping
->nrpages
) {
299 ret
= filemap_write_and_wait(mapping
);
301 xfs_rw_iunlock(ip
, XFS_IOLOCK_EXCL
);
306 * Invalidate whole pages. This can return an error if
307 * we fail to invalidate a page, but this should never
308 * happen on XFS. Warn if it does fail.
310 ret
= invalidate_inode_pages2(mapping
);
314 xfs_rw_ilock_demote(ip
, XFS_IOLOCK_EXCL
);
318 ret
= __blockdev_direct_IO(iocb
, inode
, target
->bt_bdev
, &data
,
319 xfs_get_blocks_direct
, NULL
, NULL
, 0);
322 iov_iter_advance(to
, ret
);
324 xfs_rw_iunlock(ip
, XFS_IOLOCK_SHARED
);
326 file_accessed(iocb
->ki_filp
);
330 static noinline ssize_t
335 struct address_space
*mapping
= iocb
->ki_filp
->f_mapping
;
336 struct inode
*inode
= mapping
->host
;
337 struct xfs_inode
*ip
= XFS_I(inode
);
338 struct iov_iter data
= *to
;
339 size_t count
= iov_iter_count(to
);
342 trace_xfs_file_dax_read(ip
, count
, iocb
->ki_pos
);
345 return 0; /* skip atime */
347 xfs_rw_ilock(ip
, XFS_IOLOCK_SHARED
);
348 ret
= dax_do_io(iocb
, inode
, &data
, xfs_get_blocks_direct
, NULL
, 0);
351 iov_iter_advance(to
, ret
);
353 xfs_rw_iunlock(ip
, XFS_IOLOCK_SHARED
);
355 file_accessed(iocb
->ki_filp
);
360 xfs_file_buffered_aio_read(
364 struct xfs_inode
*ip
= XFS_I(file_inode(iocb
->ki_filp
));
367 trace_xfs_file_buffered_read(ip
, iov_iter_count(to
), iocb
->ki_pos
);
369 xfs_rw_ilock(ip
, XFS_IOLOCK_SHARED
);
370 ret
= generic_file_read_iter(iocb
, to
);
371 xfs_rw_iunlock(ip
, XFS_IOLOCK_SHARED
);
381 struct inode
*inode
= file_inode(iocb
->ki_filp
);
382 struct xfs_mount
*mp
= XFS_I(inode
)->i_mount
;
385 XFS_STATS_INC(mp
, xs_read_calls
);
387 if (XFS_FORCED_SHUTDOWN(mp
))
391 ret
= xfs_file_dax_read(iocb
, to
);
392 else if (iocb
->ki_flags
& IOCB_DIRECT
)
393 ret
= xfs_file_dio_aio_read(iocb
, to
);
395 ret
= xfs_file_buffered_aio_read(iocb
, to
);
398 XFS_STATS_ADD(mp
, xs_read_bytes
, ret
);
403 xfs_file_splice_read(
406 struct pipe_inode_info
*pipe
,
410 struct xfs_inode
*ip
= XFS_I(infilp
->f_mapping
->host
);
413 XFS_STATS_INC(ip
->i_mount
, xs_read_calls
);
415 if (XFS_FORCED_SHUTDOWN(ip
->i_mount
))
418 trace_xfs_file_splice_read(ip
, count
, *ppos
);
421 * DAX inodes cannot ues the page cache for splice, so we have to push
422 * them through the VFS IO path. This means it goes through
423 * ->read_iter, which for us takes the XFS_IOLOCK_SHARED. Hence we
424 * cannot lock the splice operation at this level for DAX inodes.
426 if (IS_DAX(VFS_I(ip
))) {
427 ret
= default_file_splice_read(infilp
, ppos
, pipe
, count
,
432 xfs_rw_ilock(ip
, XFS_IOLOCK_SHARED
);
433 ret
= generic_file_splice_read(infilp
, ppos
, pipe
, count
, flags
);
434 xfs_rw_iunlock(ip
, XFS_IOLOCK_SHARED
);
437 XFS_STATS_ADD(ip
->i_mount
, xs_read_bytes
, ret
);
442 * Zero any on disk space between the current EOF and the new, larger EOF.
444 * This handles the normal case of zeroing the remainder of the last block in
445 * the file and the unusual case of zeroing blocks out beyond the size of the
446 * file. This second case only happens with fixed size extents and when the
447 * system crashes before the inode size was updated but after blocks were
450 * Expects the iolock to be held exclusive, and will take the ilock internally.
452 int /* error (positive) */
454 struct xfs_inode
*ip
,
455 xfs_off_t offset
, /* starting I/O offset */
456 xfs_fsize_t isize
, /* current inode size */
459 ASSERT(xfs_isilocked(ip
, XFS_IOLOCK_EXCL
));
460 ASSERT(offset
> isize
);
462 trace_xfs_zero_eof(ip
, isize
, offset
- isize
);
463 return xfs_zero_range(ip
, isize
, offset
- isize
, did_zeroing
);
467 * Common pre-write limit and setup checks.
469 * Called with the iolocked held either shared and exclusive according to
470 * @iolock, and returns with it held. Might upgrade the iolock to exclusive
471 * if called for a direct write beyond i_size.
474 xfs_file_aio_write_checks(
476 struct iov_iter
*from
,
479 struct file
*file
= iocb
->ki_filp
;
480 struct inode
*inode
= file
->f_mapping
->host
;
481 struct xfs_inode
*ip
= XFS_I(inode
);
483 size_t count
= iov_iter_count(from
);
484 bool drained_dio
= false;
487 error
= generic_write_checks(iocb
, from
);
491 error
= xfs_break_layouts(inode
, iolock
, true);
495 /* For changing security info in file_remove_privs() we need i_mutex */
496 if (*iolock
== XFS_IOLOCK_SHARED
&& !IS_NOSEC(inode
)) {
497 xfs_rw_iunlock(ip
, *iolock
);
498 *iolock
= XFS_IOLOCK_EXCL
;
499 xfs_rw_ilock(ip
, *iolock
);
503 * If the offset is beyond the size of the file, we need to zero any
504 * blocks that fall between the existing EOF and the start of this
505 * write. If zeroing is needed and we are currently holding the
506 * iolock shared, we need to update it to exclusive which implies
507 * having to redo all checks before.
509 * We need to serialise against EOF updates that occur in IO
510 * completions here. We want to make sure that nobody is changing the
511 * size while we do this check until we have placed an IO barrier (i.e.
512 * hold the XFS_IOLOCK_EXCL) that prevents new IO from being dispatched.
513 * The spinlock effectively forms a memory barrier once we have the
514 * XFS_IOLOCK_EXCL so we are guaranteed to see the latest EOF value
515 * and hence be able to correctly determine if we need to run zeroing.
517 spin_lock(&ip
->i_flags_lock
);
518 if (iocb
->ki_pos
> i_size_read(inode
)) {
521 spin_unlock(&ip
->i_flags_lock
);
523 if (*iolock
== XFS_IOLOCK_SHARED
) {
524 xfs_rw_iunlock(ip
, *iolock
);
525 *iolock
= XFS_IOLOCK_EXCL
;
526 xfs_rw_ilock(ip
, *iolock
);
527 iov_iter_reexpand(from
, count
);
530 * We now have an IO submission barrier in place, but
531 * AIO can do EOF updates during IO completion and hence
532 * we now need to wait for all of them to drain. Non-AIO
533 * DIO will have drained before we are given the
534 * XFS_IOLOCK_EXCL, and so for most cases this wait is a
537 inode_dio_wait(inode
);
541 error
= xfs_zero_eof(ip
, iocb
->ki_pos
, i_size_read(inode
), &zero
);
545 spin_unlock(&ip
->i_flags_lock
);
548 * Updating the timestamps will grab the ilock again from
549 * xfs_fs_dirty_inode, so we have to call it after dropping the
550 * lock above. Eventually we should look into a way to avoid
551 * the pointless lock roundtrip.
553 if (likely(!(file
->f_mode
& FMODE_NOCMTIME
))) {
554 error
= file_update_time(file
);
560 * If we're writing the file then make sure to clear the setuid and
561 * setgid bits if the process is not being run by root. This keeps
562 * people from modifying setuid and setgid binaries.
564 if (!IS_NOSEC(inode
))
565 return file_remove_privs(file
);
570 * xfs_file_dio_aio_write - handle direct IO writes
572 * Lock the inode appropriately to prepare for and issue a direct IO write.
573 * By separating it from the buffered write path we remove all the tricky to
574 * follow locking changes and looping.
576 * If there are cached pages or we're extending the file, we need IOLOCK_EXCL
577 * until we're sure the bytes at the new EOF have been zeroed and/or the cached
578 * pages are flushed out.
580 * In most cases the direct IO writes will be done holding IOLOCK_SHARED
581 * allowing them to be done in parallel with reads and other direct IO writes.
582 * However, if the IO is not aligned to filesystem blocks, the direct IO layer
583 * needs to do sub-block zeroing and that requires serialisation against other
584 * direct IOs to the same block. In this case we need to serialise the
585 * submission of the unaligned IOs so that we don't get racing block zeroing in
586 * the dio layer. To avoid the problem with aio, we also need to wait for
587 * outstanding IOs to complete so that unwritten extent conversion is completed
588 * before we try to map the overlapping block. This is currently implemented by
589 * hitting it with a big hammer (i.e. inode_dio_wait()).
591 * Returns with locks held indicated by @iolock and errors indicated by
592 * negative return values.
595 xfs_file_dio_aio_write(
597 struct iov_iter
*from
)
599 struct file
*file
= iocb
->ki_filp
;
600 struct address_space
*mapping
= file
->f_mapping
;
601 struct inode
*inode
= mapping
->host
;
602 struct xfs_inode
*ip
= XFS_I(inode
);
603 struct xfs_mount
*mp
= ip
->i_mount
;
605 int unaligned_io
= 0;
607 size_t count
= iov_iter_count(from
);
609 struct iov_iter data
;
610 struct xfs_buftarg
*target
= XFS_IS_REALTIME_INODE(ip
) ?
611 mp
->m_rtdev_targp
: mp
->m_ddev_targp
;
613 /* DIO must be aligned to device logical sector size */
614 if ((iocb
->ki_pos
| count
) & target
->bt_logical_sectormask
)
617 /* "unaligned" here means not aligned to a filesystem block */
618 if ((iocb
->ki_pos
& mp
->m_blockmask
) ||
619 ((iocb
->ki_pos
+ count
) & mp
->m_blockmask
))
623 * We don't need to take an exclusive lock unless there page cache needs
624 * to be invalidated or unaligned IO is being executed. We don't need to
625 * consider the EOF extension case here because
626 * xfs_file_aio_write_checks() will relock the inode as necessary for
627 * EOF zeroing cases and fill out the new inode size as appropriate.
629 if (unaligned_io
|| mapping
->nrpages
)
630 iolock
= XFS_IOLOCK_EXCL
;
632 iolock
= XFS_IOLOCK_SHARED
;
633 xfs_rw_ilock(ip
, iolock
);
636 * Recheck if there are cached pages that need invalidate after we got
637 * the iolock to protect against other threads adding new pages while
638 * we were waiting for the iolock.
640 if (mapping
->nrpages
&& iolock
== XFS_IOLOCK_SHARED
) {
641 xfs_rw_iunlock(ip
, iolock
);
642 iolock
= XFS_IOLOCK_EXCL
;
643 xfs_rw_ilock(ip
, iolock
);
646 ret
= xfs_file_aio_write_checks(iocb
, from
, &iolock
);
649 count
= iov_iter_count(from
);
650 end
= iocb
->ki_pos
+ count
- 1;
653 * See xfs_file_dio_aio_read() for why we do a full-file flush here.
655 if (mapping
->nrpages
) {
656 ret
= filemap_write_and_wait(VFS_I(ip
)->i_mapping
);
660 * Invalidate whole pages. This can return an error if we fail
661 * to invalidate a page, but this should never happen on XFS.
662 * Warn if it does fail.
664 ret
= invalidate_inode_pages2(VFS_I(ip
)->i_mapping
);
670 * If we are doing unaligned IO, wait for all other IO to drain,
671 * otherwise demote the lock if we had to flush cached pages
674 inode_dio_wait(inode
);
675 else if (iolock
== XFS_IOLOCK_EXCL
) {
676 xfs_rw_ilock_demote(ip
, XFS_IOLOCK_EXCL
);
677 iolock
= XFS_IOLOCK_SHARED
;
680 trace_xfs_file_direct_write(ip
, count
, iocb
->ki_pos
);
683 ret
= __blockdev_direct_IO(iocb
, inode
, target
->bt_bdev
, &data
,
684 xfs_get_blocks_direct
, xfs_end_io_direct_write
,
685 NULL
, DIO_ASYNC_EXTEND
);
687 /* see generic_file_direct_write() for why this is necessary */
688 if (mapping
->nrpages
) {
689 invalidate_inode_pages2_range(mapping
,
690 iocb
->ki_pos
>> PAGE_SHIFT
,
696 iov_iter_advance(from
, ret
);
699 xfs_rw_iunlock(ip
, iolock
);
702 * No fallback to buffered IO on errors for XFS, direct IO will either
703 * complete fully or fail.
705 ASSERT(ret
< 0 || ret
== count
);
709 static noinline ssize_t
712 struct iov_iter
*from
)
714 struct address_space
*mapping
= iocb
->ki_filp
->f_mapping
;
715 struct inode
*inode
= mapping
->host
;
716 struct xfs_inode
*ip
= XFS_I(inode
);
717 struct xfs_mount
*mp
= ip
->i_mount
;
719 int unaligned_io
= 0;
721 struct iov_iter data
;
723 /* "unaligned" here means not aligned to a filesystem block */
724 if ((iocb
->ki_pos
& mp
->m_blockmask
) ||
725 ((iocb
->ki_pos
+ iov_iter_count(from
)) & mp
->m_blockmask
)) {
727 iolock
= XFS_IOLOCK_EXCL
;
728 } else if (mapping
->nrpages
) {
729 iolock
= XFS_IOLOCK_EXCL
;
731 iolock
= XFS_IOLOCK_SHARED
;
733 xfs_rw_ilock(ip
, iolock
);
735 ret
= xfs_file_aio_write_checks(iocb
, from
, &iolock
);
740 * Yes, even DAX files can have page cache attached to them: A zeroed
741 * page is inserted into the pagecache when we have to serve a write
742 * fault on a hole. It should never be dirtied and can simply be
743 * dropped from the pagecache once we get real data for the page.
745 * XXX: This is racy against mmap, and there's nothing we can do about
746 * it. dax_do_io() should really do this invalidation internally as
747 * it will know if we've allocated over a holei for this specific IO and
748 * if so it needs to update the mapping tree and invalidate existing
749 * PTEs over the newly allocated range. Remove this invalidation when
750 * dax_do_io() is fixed up.
752 if (mapping
->nrpages
) {
753 loff_t end
= iocb
->ki_pos
+ iov_iter_count(from
) - 1;
755 ret
= invalidate_inode_pages2_range(mapping
,
756 iocb
->ki_pos
>> PAGE_SHIFT
,
761 if (iolock
== XFS_IOLOCK_EXCL
&& !unaligned_io
) {
762 xfs_rw_ilock_demote(ip
, XFS_IOLOCK_EXCL
);
763 iolock
= XFS_IOLOCK_SHARED
;
766 trace_xfs_file_dax_write(ip
, iov_iter_count(from
), iocb
->ki_pos
);
769 ret
= dax_do_io(iocb
, inode
, &data
, xfs_get_blocks_direct
,
770 xfs_end_io_direct_write
, 0);
773 iov_iter_advance(from
, ret
);
776 xfs_rw_iunlock(ip
, iolock
);
781 xfs_file_buffered_aio_write(
783 struct iov_iter
*from
)
785 struct file
*file
= iocb
->ki_filp
;
786 struct address_space
*mapping
= file
->f_mapping
;
787 struct inode
*inode
= mapping
->host
;
788 struct xfs_inode
*ip
= XFS_I(inode
);
791 int iolock
= XFS_IOLOCK_EXCL
;
793 xfs_rw_ilock(ip
, iolock
);
795 ret
= xfs_file_aio_write_checks(iocb
, from
, &iolock
);
799 /* We can write back this queue in page reclaim */
800 current
->backing_dev_info
= inode_to_bdi(inode
);
803 trace_xfs_file_buffered_write(ip
, iov_iter_count(from
), iocb
->ki_pos
);
804 ret
= iomap_file_buffered_write(iocb
, from
, &xfs_iomap_ops
);
805 if (likely(ret
>= 0))
809 * If we hit a space limit, try to free up some lingering preallocated
810 * space before returning an error. In the case of ENOSPC, first try to
811 * write back all dirty inodes to free up some of the excess reserved
812 * metadata space. This reduces the chances that the eofblocks scan
813 * waits on dirty mappings. Since xfs_flush_inodes() is serialized, this
814 * also behaves as a filter to prevent too many eofblocks scans from
815 * running at the same time.
817 if (ret
== -EDQUOT
&& !enospc
) {
818 enospc
= xfs_inode_free_quota_eofblocks(ip
);
821 } else if (ret
== -ENOSPC
&& !enospc
) {
822 struct xfs_eofblocks eofb
= {0};
825 xfs_flush_inodes(ip
->i_mount
);
826 eofb
.eof_scan_owner
= ip
->i_ino
; /* for locking */
827 eofb
.eof_flags
= XFS_EOF_FLAGS_SYNC
;
828 xfs_icache_free_eofblocks(ip
->i_mount
, &eofb
);
832 current
->backing_dev_info
= NULL
;
834 xfs_rw_iunlock(ip
, iolock
);
841 struct iov_iter
*from
)
843 struct file
*file
= iocb
->ki_filp
;
844 struct address_space
*mapping
= file
->f_mapping
;
845 struct inode
*inode
= mapping
->host
;
846 struct xfs_inode
*ip
= XFS_I(inode
);
848 size_t ocount
= iov_iter_count(from
);
850 XFS_STATS_INC(ip
->i_mount
, xs_write_calls
);
855 if (XFS_FORCED_SHUTDOWN(ip
->i_mount
))
859 ret
= xfs_file_dax_write(iocb
, from
);
860 else if (iocb
->ki_flags
& IOCB_DIRECT
)
861 ret
= xfs_file_dio_aio_write(iocb
, from
);
863 ret
= xfs_file_buffered_aio_write(iocb
, from
);
866 XFS_STATS_ADD(ip
->i_mount
, xs_write_bytes
, ret
);
868 /* Handle various SYNC-type writes */
869 ret
= generic_write_sync(iocb
, ret
);
874 #define XFS_FALLOC_FL_SUPPORTED \
875 (FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE | \
876 FALLOC_FL_COLLAPSE_RANGE | FALLOC_FL_ZERO_RANGE | \
877 FALLOC_FL_INSERT_RANGE)
886 struct inode
*inode
= file_inode(file
);
887 struct xfs_inode
*ip
= XFS_I(inode
);
889 enum xfs_prealloc_flags flags
= 0;
890 uint iolock
= XFS_IOLOCK_EXCL
;
892 bool do_file_insert
= 0;
894 if (!S_ISREG(inode
->i_mode
))
896 if (mode
& ~XFS_FALLOC_FL_SUPPORTED
)
899 xfs_ilock(ip
, iolock
);
900 error
= xfs_break_layouts(inode
, &iolock
, false);
904 xfs_ilock(ip
, XFS_MMAPLOCK_EXCL
);
905 iolock
|= XFS_MMAPLOCK_EXCL
;
907 if (mode
& FALLOC_FL_PUNCH_HOLE
) {
908 error
= xfs_free_file_space(ip
, offset
, len
);
911 } else if (mode
& FALLOC_FL_COLLAPSE_RANGE
) {
912 unsigned blksize_mask
= (1 << inode
->i_blkbits
) - 1;
914 if (offset
& blksize_mask
|| len
& blksize_mask
) {
920 * There is no need to overlap collapse range with EOF,
921 * in which case it is effectively a truncate operation
923 if (offset
+ len
>= i_size_read(inode
)) {
928 new_size
= i_size_read(inode
) - len
;
930 error
= xfs_collapse_file_space(ip
, offset
, len
);
933 } else if (mode
& FALLOC_FL_INSERT_RANGE
) {
934 unsigned blksize_mask
= (1 << inode
->i_blkbits
) - 1;
936 new_size
= i_size_read(inode
) + len
;
937 if (offset
& blksize_mask
|| len
& blksize_mask
) {
942 /* check the new inode size does not wrap through zero */
943 if (new_size
> inode
->i_sb
->s_maxbytes
) {
948 /* Offset should be less than i_size */
949 if (offset
>= i_size_read(inode
)) {
955 flags
|= XFS_PREALLOC_SET
;
957 if (!(mode
& FALLOC_FL_KEEP_SIZE
) &&
958 offset
+ len
> i_size_read(inode
)) {
959 new_size
= offset
+ len
;
960 error
= inode_newsize_ok(inode
, new_size
);
965 if (mode
& FALLOC_FL_ZERO_RANGE
)
966 error
= xfs_zero_file_space(ip
, offset
, len
);
968 error
= xfs_alloc_file_space(ip
, offset
, len
,
974 if (file
->f_flags
& O_DSYNC
)
975 flags
|= XFS_PREALLOC_SYNC
;
977 error
= xfs_update_prealloc_flags(ip
, flags
);
981 /* Change file size if needed */
985 iattr
.ia_valid
= ATTR_SIZE
;
986 iattr
.ia_size
= new_size
;
987 error
= xfs_setattr_size(ip
, &iattr
);
993 * Perform hole insertion now that the file size has been
994 * updated so that if we crash during the operation we don't
995 * leave shifted extents past EOF and hence losing access to
996 * the data that is contained within them.
999 error
= xfs_insert_file_space(ip
, offset
, len
);
1002 xfs_iunlock(ip
, iolock
);
1009 struct inode
*inode
,
1012 if (!(file
->f_flags
& O_LARGEFILE
) && i_size_read(inode
) > MAX_NON_LFS
)
1014 if (XFS_FORCED_SHUTDOWN(XFS_M(inode
->i_sb
)))
1021 struct inode
*inode
,
1024 struct xfs_inode
*ip
= XFS_I(inode
);
1028 error
= xfs_file_open(inode
, file
);
1033 * If there are any blocks, read-ahead block 0 as we're almost
1034 * certain to have the next operation be a read there.
1036 mode
= xfs_ilock_data_map_shared(ip
);
1037 if (ip
->i_d
.di_nextents
> 0)
1038 xfs_dir3_data_readahead(ip
, 0, -1);
1039 xfs_iunlock(ip
, mode
);
1045 struct inode
*inode
,
1048 return xfs_release(XFS_I(inode
));
1054 struct dir_context
*ctx
)
1056 struct inode
*inode
= file_inode(file
);
1057 xfs_inode_t
*ip
= XFS_I(inode
);
1061 * The Linux API doesn't pass down the total size of the buffer
1062 * we read into down to the filesystem. With the filldir concept
1063 * it's not needed for correct information, but the XFS dir2 leaf
1064 * code wants an estimate of the buffer size to calculate it's
1065 * readahead window and size the buffers used for mapping to
1068 * Try to give it an estimate that's good enough, maybe at some
1069 * point we can change the ->readdir prototype to include the
1070 * buffer size. For now we use the current glibc buffer size.
1072 bufsize
= (size_t)min_t(loff_t
, 32768, ip
->i_d
.di_size
);
1074 return xfs_readdir(ip
, ctx
, bufsize
);
1078 * This type is designed to indicate the type of offset we would like
1079 * to search from page cache for xfs_seek_hole_data().
1087 * Lookup the desired type of offset from the given page.
1089 * On success, return true and the offset argument will point to the
1090 * start of the region that was found. Otherwise this function will
1091 * return false and keep the offset argument unchanged.
1094 xfs_lookup_buffer_offset(
1099 loff_t lastoff
= page_offset(page
);
1101 struct buffer_head
*bh
, *head
;
1103 bh
= head
= page_buffers(page
);
1106 * Unwritten extents that have data in the page
1107 * cache covering them can be identified by the
1108 * BH_Unwritten state flag. Pages with multiple
1109 * buffers might have a mix of holes, data and
1110 * unwritten extents - any buffer with valid
1111 * data in it should have BH_Uptodate flag set
1114 if (buffer_unwritten(bh
) ||
1115 buffer_uptodate(bh
)) {
1116 if (type
== DATA_OFF
)
1119 if (type
== HOLE_OFF
)
1127 lastoff
+= bh
->b_size
;
1128 } while ((bh
= bh
->b_this_page
) != head
);
1134 * This routine is called to find out and return a data or hole offset
1135 * from the page cache for unwritten extents according to the desired
1136 * type for xfs_seek_hole_data().
1138 * The argument offset is used to tell where we start to search from the
1139 * page cache. Map is used to figure out the end points of the range to
1142 * Return true if the desired type of offset was found, and the argument
1143 * offset is filled with that address. Otherwise, return false and keep
1147 xfs_find_get_desired_pgoff(
1148 struct inode
*inode
,
1149 struct xfs_bmbt_irec
*map
,
1153 struct xfs_inode
*ip
= XFS_I(inode
);
1154 struct xfs_mount
*mp
= ip
->i_mount
;
1155 struct pagevec pvec
;
1159 loff_t startoff
= *offset
;
1160 loff_t lastoff
= startoff
;
1163 pagevec_init(&pvec
, 0);
1165 index
= startoff
>> PAGE_SHIFT
;
1166 endoff
= XFS_FSB_TO_B(mp
, map
->br_startoff
+ map
->br_blockcount
);
1167 end
= endoff
>> PAGE_SHIFT
;
1173 want
= min_t(pgoff_t
, end
- index
, PAGEVEC_SIZE
);
1174 nr_pages
= pagevec_lookup(&pvec
, inode
->i_mapping
, index
,
1177 * No page mapped into given range. If we are searching holes
1178 * and if this is the first time we got into the loop, it means
1179 * that the given offset is landed in a hole, return it.
1181 * If we have already stepped through some block buffers to find
1182 * holes but they all contains data. In this case, the last
1183 * offset is already updated and pointed to the end of the last
1184 * mapped page, if it does not reach the endpoint to search,
1185 * that means there should be a hole between them.
1187 if (nr_pages
== 0) {
1188 /* Data search found nothing */
1189 if (type
== DATA_OFF
)
1192 ASSERT(type
== HOLE_OFF
);
1193 if (lastoff
== startoff
|| lastoff
< endoff
) {
1201 * At lease we found one page. If this is the first time we
1202 * step into the loop, and if the first page index offset is
1203 * greater than the given search offset, a hole was found.
1205 if (type
== HOLE_OFF
&& lastoff
== startoff
&&
1206 lastoff
< page_offset(pvec
.pages
[0])) {
1211 for (i
= 0; i
< nr_pages
; i
++) {
1212 struct page
*page
= pvec
.pages
[i
];
1216 * At this point, the page may be truncated or
1217 * invalidated (changing page->mapping to NULL),
1218 * or even swizzled back from swapper_space to tmpfs
1219 * file mapping. However, page->index will not change
1220 * because we have a reference on the page.
1222 * Searching done if the page index is out of range.
1223 * If the current offset is not reaches the end of
1224 * the specified search range, there should be a hole
1227 if (page
->index
> end
) {
1228 if (type
== HOLE_OFF
&& lastoff
< endoff
) {
1237 * Page truncated or invalidated(page->mapping == NULL).
1238 * We can freely skip it and proceed to check the next
1241 if (unlikely(page
->mapping
!= inode
->i_mapping
)) {
1246 if (!page_has_buffers(page
)) {
1251 found
= xfs_lookup_buffer_offset(page
, &b_offset
, type
);
1254 * The found offset may be less than the start
1255 * point to search if this is the first time to
1258 *offset
= max_t(loff_t
, startoff
, b_offset
);
1264 * We either searching data but nothing was found, or
1265 * searching hole but found a data buffer. In either
1266 * case, probably the next page contains the desired
1267 * things, update the last offset to it so.
1269 lastoff
= page_offset(page
) + PAGE_SIZE
;
1274 * The number of returned pages less than our desired, search
1275 * done. In this case, nothing was found for searching data,
1276 * but we found a hole behind the last offset.
1278 if (nr_pages
< want
) {
1279 if (type
== HOLE_OFF
) {
1286 index
= pvec
.pages
[i
- 1]->index
+ 1;
1287 pagevec_release(&pvec
);
1288 } while (index
<= end
);
1291 pagevec_release(&pvec
);
1296 * caller must lock inode with xfs_ilock_data_map_shared,
1297 * can we craft an appropriate ASSERT?
1299 * end is because the VFS-level lseek interface is defined such that any
1300 * offset past i_size shall return -ENXIO, but we use this for quota code
1301 * which does not maintain i_size, and we want to SEEK_DATA past i_size.
1304 __xfs_seek_hole_data(
1305 struct inode
*inode
,
1310 struct xfs_inode
*ip
= XFS_I(inode
);
1311 struct xfs_mount
*mp
= ip
->i_mount
;
1312 loff_t
uninitialized_var(offset
);
1313 xfs_fileoff_t fsbno
;
1314 xfs_filblks_t lastbno
;
1323 * Try to read extents from the first block indicated
1324 * by fsbno to the end block of the file.
1326 fsbno
= XFS_B_TO_FSBT(mp
, start
);
1327 lastbno
= XFS_B_TO_FSB(mp
, end
);
1330 struct xfs_bmbt_irec map
[2];
1334 error
= xfs_bmapi_read(ip
, fsbno
, lastbno
- fsbno
, map
, &nmap
,
1339 /* No extents at given offset, must be beyond EOF */
1345 for (i
= 0; i
< nmap
; i
++) {
1346 offset
= max_t(loff_t
, start
,
1347 XFS_FSB_TO_B(mp
, map
[i
].br_startoff
));
1349 /* Landed in the hole we wanted? */
1350 if (whence
== SEEK_HOLE
&&
1351 map
[i
].br_startblock
== HOLESTARTBLOCK
)
1354 /* Landed in the data extent we wanted? */
1355 if (whence
== SEEK_DATA
&&
1356 (map
[i
].br_startblock
== DELAYSTARTBLOCK
||
1357 (map
[i
].br_state
== XFS_EXT_NORM
&&
1358 !isnullstartblock(map
[i
].br_startblock
))))
1362 * Landed in an unwritten extent, try to search
1363 * for hole or data from page cache.
1365 if (map
[i
].br_state
== XFS_EXT_UNWRITTEN
) {
1366 if (xfs_find_get_desired_pgoff(inode
, &map
[i
],
1367 whence
== SEEK_HOLE
? HOLE_OFF
: DATA_OFF
,
1374 * We only received one extent out of the two requested. This
1375 * means we've hit EOF and didn't find what we are looking for.
1379 * If we were looking for a hole, set offset to
1380 * the end of the file (i.e., there is an implicit
1381 * hole at the end of any file).
1383 if (whence
== SEEK_HOLE
) {
1388 * If we were looking for data, it's nowhere to be found
1390 ASSERT(whence
== SEEK_DATA
);
1398 * Nothing was found, proceed to the next round of search
1399 * if the next reading offset is not at or beyond EOF.
1401 fsbno
= map
[i
- 1].br_startoff
+ map
[i
- 1].br_blockcount
;
1402 start
= XFS_FSB_TO_B(mp
, fsbno
);
1404 if (whence
== SEEK_HOLE
) {
1408 ASSERT(whence
== SEEK_DATA
);
1416 * If at this point we have found the hole we wanted, the returned
1417 * offset may be bigger than the file size as it may be aligned to
1418 * page boundary for unwritten extents. We need to deal with this
1419 * situation in particular.
1421 if (whence
== SEEK_HOLE
)
1422 offset
= min_t(loff_t
, offset
, end
);
1436 struct inode
*inode
= file
->f_mapping
->host
;
1437 struct xfs_inode
*ip
= XFS_I(inode
);
1438 struct xfs_mount
*mp
= ip
->i_mount
;
1443 if (XFS_FORCED_SHUTDOWN(mp
))
1446 lock
= xfs_ilock_data_map_shared(ip
);
1448 end
= i_size_read(inode
);
1449 offset
= __xfs_seek_hole_data(inode
, start
, end
, whence
);
1455 offset
= vfs_setpos(file
, offset
, inode
->i_sb
->s_maxbytes
);
1458 xfs_iunlock(ip
, lock
);
1475 return generic_file_llseek(file
, offset
, whence
);
1478 return xfs_seek_hole_data(file
, offset
, whence
);
1485 * Locking for serialisation of IO during page faults. This results in a lock
1489 * sb_start_pagefault(vfs, freeze)
1490 * i_mmaplock (XFS - truncate serialisation)
1492 * i_lock (XFS - extent map serialisation)
1496 * mmap()d file has taken write protection fault and is being made writable. We
1497 * can set the page state up correctly for a writable page, which means we can
1498 * do correct delalloc accounting (ENOSPC checking!) and unwritten extent
1502 xfs_filemap_page_mkwrite(
1503 struct vm_area_struct
*vma
,
1504 struct vm_fault
*vmf
)
1506 struct inode
*inode
= file_inode(vma
->vm_file
);
1509 trace_xfs_filemap_page_mkwrite(XFS_I(inode
));
1511 sb_start_pagefault(inode
->i_sb
);
1512 file_update_time(vma
->vm_file
);
1513 xfs_ilock(XFS_I(inode
), XFS_MMAPLOCK_SHARED
);
1515 if (IS_DAX(inode
)) {
1516 ret
= dax_mkwrite(vma
, vmf
, xfs_get_blocks_dax_fault
);
1518 ret
= iomap_page_mkwrite(vma
, vmf
, &xfs_iomap_ops
);
1519 ret
= block_page_mkwrite_return(ret
);
1522 xfs_iunlock(XFS_I(inode
), XFS_MMAPLOCK_SHARED
);
1523 sb_end_pagefault(inode
->i_sb
);
1530 struct vm_area_struct
*vma
,
1531 struct vm_fault
*vmf
)
1533 struct inode
*inode
= file_inode(vma
->vm_file
);
1536 trace_xfs_filemap_fault(XFS_I(inode
));
1538 /* DAX can shortcut the normal fault path on write faults! */
1539 if ((vmf
->flags
& FAULT_FLAG_WRITE
) && IS_DAX(inode
))
1540 return xfs_filemap_page_mkwrite(vma
, vmf
);
1542 xfs_ilock(XFS_I(inode
), XFS_MMAPLOCK_SHARED
);
1543 if (IS_DAX(inode
)) {
1545 * we do not want to trigger unwritten extent conversion on read
1546 * faults - that is unnecessary overhead and would also require
1547 * changes to xfs_get_blocks_direct() to map unwritten extent
1548 * ioend for conversion on read-only mappings.
1550 ret
= dax_fault(vma
, vmf
, xfs_get_blocks_dax_fault
);
1552 ret
= filemap_fault(vma
, vmf
);
1553 xfs_iunlock(XFS_I(inode
), XFS_MMAPLOCK_SHARED
);
1559 * Similar to xfs_filemap_fault(), the DAX fault path can call into here on
1560 * both read and write faults. Hence we need to handle both cases. There is no
1561 * ->pmd_mkwrite callout for huge pages, so we have a single function here to
1562 * handle both cases here. @flags carries the information on the type of fault
1566 xfs_filemap_pmd_fault(
1567 struct vm_area_struct
*vma
,
1572 struct inode
*inode
= file_inode(vma
->vm_file
);
1573 struct xfs_inode
*ip
= XFS_I(inode
);
1577 return VM_FAULT_FALLBACK
;
1579 trace_xfs_filemap_pmd_fault(ip
);
1581 if (flags
& FAULT_FLAG_WRITE
) {
1582 sb_start_pagefault(inode
->i_sb
);
1583 file_update_time(vma
->vm_file
);
1586 xfs_ilock(XFS_I(inode
), XFS_MMAPLOCK_SHARED
);
1587 ret
= dax_pmd_fault(vma
, addr
, pmd
, flags
, xfs_get_blocks_dax_fault
);
1588 xfs_iunlock(XFS_I(inode
), XFS_MMAPLOCK_SHARED
);
1590 if (flags
& FAULT_FLAG_WRITE
)
1591 sb_end_pagefault(inode
->i_sb
);
1597 * pfn_mkwrite was originally inteneded to ensure we capture time stamp
1598 * updates on write faults. In reality, it's need to serialise against
1599 * truncate similar to page_mkwrite. Hence we cycle the XFS_MMAPLOCK_SHARED
1600 * to ensure we serialise the fault barrier in place.
1603 xfs_filemap_pfn_mkwrite(
1604 struct vm_area_struct
*vma
,
1605 struct vm_fault
*vmf
)
1608 struct inode
*inode
= file_inode(vma
->vm_file
);
1609 struct xfs_inode
*ip
= XFS_I(inode
);
1610 int ret
= VM_FAULT_NOPAGE
;
1613 trace_xfs_filemap_pfn_mkwrite(ip
);
1615 sb_start_pagefault(inode
->i_sb
);
1616 file_update_time(vma
->vm_file
);
1618 /* check if the faulting page hasn't raced with truncate */
1619 xfs_ilock(ip
, XFS_MMAPLOCK_SHARED
);
1620 size
= (i_size_read(inode
) + PAGE_SIZE
- 1) >> PAGE_SHIFT
;
1621 if (vmf
->pgoff
>= size
)
1622 ret
= VM_FAULT_SIGBUS
;
1623 else if (IS_DAX(inode
))
1624 ret
= dax_pfn_mkwrite(vma
, vmf
);
1625 xfs_iunlock(ip
, XFS_MMAPLOCK_SHARED
);
1626 sb_end_pagefault(inode
->i_sb
);
1631 static const struct vm_operations_struct xfs_file_vm_ops
= {
1632 .fault
= xfs_filemap_fault
,
1633 .pmd_fault
= xfs_filemap_pmd_fault
,
1634 .map_pages
= filemap_map_pages
,
1635 .page_mkwrite
= xfs_filemap_page_mkwrite
,
1636 .pfn_mkwrite
= xfs_filemap_pfn_mkwrite
,
1642 struct vm_area_struct
*vma
)
1644 file_accessed(filp
);
1645 vma
->vm_ops
= &xfs_file_vm_ops
;
1646 if (IS_DAX(file_inode(filp
)))
1647 vma
->vm_flags
|= VM_MIXEDMAP
| VM_HUGEPAGE
;
1651 const struct file_operations xfs_file_operations
= {
1652 .llseek
= xfs_file_llseek
,
1653 .read_iter
= xfs_file_read_iter
,
1654 .write_iter
= xfs_file_write_iter
,
1655 .splice_read
= xfs_file_splice_read
,
1656 .splice_write
= iter_file_splice_write
,
1657 .unlocked_ioctl
= xfs_file_ioctl
,
1658 #ifdef CONFIG_COMPAT
1659 .compat_ioctl
= xfs_file_compat_ioctl
,
1661 .mmap
= xfs_file_mmap
,
1662 .open
= xfs_file_open
,
1663 .release
= xfs_file_release
,
1664 .fsync
= xfs_file_fsync
,
1665 .fallocate
= xfs_file_fallocate
,
1668 const struct file_operations xfs_dir_file_operations
= {
1669 .open
= xfs_dir_open
,
1670 .read
= generic_read_dir
,
1671 .iterate_shared
= xfs_file_readdir
,
1672 .llseek
= generic_file_llseek
,
1673 .unlocked_ioctl
= xfs_file_ioctl
,
1674 #ifdef CONFIG_COMPAT
1675 .compat_ioctl
= xfs_file_compat_ioctl
,
1677 .fsync
= xfs_dir_fsync
,