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)
272 file_accessed(iocb->ki_filp);
275 * Locking is a bit tricky here. If we take an exclusive lock for direct
276 * IO, we effectively serialise all new concurrent read IO to this file
277 * and block it behind IO that is currently in progress because IO in
278 * progress holds the IO lock shared. We only need to hold the lock
279 * exclusive to blow away the page cache, so only take lock exclusively
280 * if the page cache needs invalidation. This allows the normal direct
281 * IO case of no page cache pages to proceeed concurrently without
284 xfs_rw_ilock(ip, XFS_IOLOCK_SHARED);
285 if (mapping->nrpages) {
286 xfs_rw_iunlock(ip, XFS_IOLOCK_SHARED);
287 xfs_rw_ilock(ip, XFS_IOLOCK_EXCL);
290 * The generic dio code only flushes the range of the particular
291 * I/O. Because we take an exclusive lock here, this whole
292 * sequence is considerably more expensive for us. This has a
293 * noticeable performance impact for any file with cached pages,
294 * even when outside of the range of the particular I/O.
296 * Hence, amortize the cost of the lock against a full file
297 * flush and reduce the chances of repeated iolock cycles going
300 if (mapping->nrpages) {
301 ret = filemap_write_and_wait(mapping);
303 xfs_rw_iunlock(ip, XFS_IOLOCK_EXCL);
308 * Invalidate whole pages. This can return an error if
309 * we fail to invalidate a page, but this should never
310 * happen on XFS. Warn if it does fail.
312 ret = invalidate_inode_pages2(mapping);
316 xfs_rw_ilock_demote(ip, XFS_IOLOCK_EXCL);
320 ret = __blockdev_direct_IO(iocb, inode, target->bt_bdev, &data,
321 xfs_get_blocks_direct, NULL, NULL, 0);
324 iov_iter_advance(to, ret);
326 xfs_rw_iunlock(ip, XFS_IOLOCK_SHARED);
331 static noinline ssize_t
336 struct address_space *mapping = iocb->ki_filp->f_mapping;
337 struct inode *inode = mapping->host;
338 struct xfs_inode *ip = XFS_I(inode);
339 struct iov_iter data = *to;
340 size_t count = iov_iter_count(to);
343 trace_xfs_file_dax_read(ip, count, iocb->ki_pos);
346 return 0; /* skip atime */
348 xfs_rw_ilock(ip, XFS_IOLOCK_SHARED);
349 ret = dax_do_io(iocb, inode, &data, xfs_get_blocks_direct, NULL, 0);
352 iov_iter_advance(to, ret);
354 xfs_rw_iunlock(ip, XFS_IOLOCK_SHARED);
356 file_accessed(iocb->ki_filp);
361 xfs_file_buffered_aio_read(
365 struct xfs_inode *ip = XFS_I(file_inode(iocb->ki_filp));
368 trace_xfs_file_buffered_read(ip, iov_iter_count(to), iocb->ki_pos);
370 xfs_rw_ilock(ip, XFS_IOLOCK_SHARED);
371 ret = generic_file_read_iter(iocb, to);
372 xfs_rw_iunlock(ip, XFS_IOLOCK_SHARED);
382 struct inode *inode = file_inode(iocb->ki_filp);
383 struct xfs_mount *mp = XFS_I(inode)->i_mount;
386 XFS_STATS_INC(mp, xs_read_calls);
388 if (XFS_FORCED_SHUTDOWN(mp))
392 ret = xfs_file_dax_read(iocb, to);
393 else if (iocb->ki_flags & IOCB_DIRECT)
394 ret = xfs_file_dio_aio_read(iocb, to);
396 ret = xfs_file_buffered_aio_read(iocb, to);
399 XFS_STATS_ADD(mp, xs_read_bytes, ret);
404 xfs_file_splice_read(
407 struct pipe_inode_info *pipe,
411 struct xfs_inode *ip = XFS_I(infilp->f_mapping->host);
414 XFS_STATS_INC(ip->i_mount, xs_read_calls);
416 if (XFS_FORCED_SHUTDOWN(ip->i_mount))
419 trace_xfs_file_splice_read(ip, count, *ppos);
422 * DAX inodes cannot ues the page cache for splice, so we have to push
423 * them through the VFS IO path. This means it goes through
424 * ->read_iter, which for us takes the XFS_IOLOCK_SHARED. Hence we
425 * cannot lock the splice operation at this level for DAX inodes.
427 if (IS_DAX(VFS_I(ip))) {
428 ret = default_file_splice_read(infilp, ppos, pipe, count,
433 xfs_rw_ilock(ip, XFS_IOLOCK_SHARED);
434 ret = generic_file_splice_read(infilp, ppos, pipe, count, flags);
435 xfs_rw_iunlock(ip, XFS_IOLOCK_SHARED);
438 XFS_STATS_ADD(ip->i_mount, xs_read_bytes, ret);
443 * Zero any on disk space between the current EOF and the new, larger EOF.
445 * This handles the normal case of zeroing the remainder of the last block in
446 * the file and the unusual case of zeroing blocks out beyond the size of the
447 * file. This second case only happens with fixed size extents and when the
448 * system crashes before the inode size was updated but after blocks were
451 * Expects the iolock to be held exclusive, and will take the ilock internally.
453 int /* error (positive) */
455 struct xfs_inode *ip,
456 xfs_off_t offset, /* starting I/O offset */
457 xfs_fsize_t isize, /* current inode size */
460 ASSERT(xfs_isilocked(ip, XFS_IOLOCK_EXCL));
461 ASSERT(offset > isize);
463 trace_xfs_zero_eof(ip, isize, offset - isize);
464 return xfs_zero_range(ip, isize, offset - isize, did_zeroing);
468 * Common pre-write limit and setup checks.
470 * Called with the iolocked held either shared and exclusive according to
471 * @iolock, and returns with it held. Might upgrade the iolock to exclusive
472 * if called for a direct write beyond i_size.
475 xfs_file_aio_write_checks(
477 struct iov_iter *from,
480 struct file *file = iocb->ki_filp;
481 struct inode *inode = file->f_mapping->host;
482 struct xfs_inode *ip = XFS_I(inode);
484 size_t count = iov_iter_count(from);
485 bool drained_dio = false;
488 error = generic_write_checks(iocb, from);
492 error = xfs_break_layouts(inode, iolock, true);
496 /* For changing security info in file_remove_privs() we need i_mutex */
497 if (*iolock == XFS_IOLOCK_SHARED && !IS_NOSEC(inode)) {
498 xfs_rw_iunlock(ip, *iolock);
499 *iolock = XFS_IOLOCK_EXCL;
500 xfs_rw_ilock(ip, *iolock);
504 * If the offset is beyond the size of the file, we need to zero any
505 * blocks that fall between the existing EOF and the start of this
506 * write. If zeroing is needed and we are currently holding the
507 * iolock shared, we need to update it to exclusive which implies
508 * having to redo all checks before.
510 * We need to serialise against EOF updates that occur in IO
511 * completions here. We want to make sure that nobody is changing the
512 * size while we do this check until we have placed an IO barrier (i.e.
513 * hold the XFS_IOLOCK_EXCL) that prevents new IO from being dispatched.
514 * The spinlock effectively forms a memory barrier once we have the
515 * XFS_IOLOCK_EXCL so we are guaranteed to see the latest EOF value
516 * and hence be able to correctly determine if we need to run zeroing.
518 spin_lock(&ip->i_flags_lock);
519 if (iocb->ki_pos > i_size_read(inode)) {
522 spin_unlock(&ip->i_flags_lock);
524 if (*iolock == XFS_IOLOCK_SHARED) {
525 xfs_rw_iunlock(ip, *iolock);
526 *iolock = XFS_IOLOCK_EXCL;
527 xfs_rw_ilock(ip, *iolock);
528 iov_iter_reexpand(from, count);
531 * We now have an IO submission barrier in place, but
532 * AIO can do EOF updates during IO completion and hence
533 * we now need to wait for all of them to drain. Non-AIO
534 * DIO will have drained before we are given the
535 * XFS_IOLOCK_EXCL, and so for most cases this wait is a
538 inode_dio_wait(inode);
542 error = xfs_zero_eof(ip, iocb->ki_pos, i_size_read(inode), &zero);
546 spin_unlock(&ip->i_flags_lock);
549 * Updating the timestamps will grab the ilock again from
550 * xfs_fs_dirty_inode, so we have to call it after dropping the
551 * lock above. Eventually we should look into a way to avoid
552 * the pointless lock roundtrip.
554 if (likely(!(file->f_mode & FMODE_NOCMTIME))) {
555 error = file_update_time(file);
561 * If we're writing the file then make sure to clear the setuid and
562 * setgid bits if the process is not being run by root. This keeps
563 * people from modifying setuid and setgid binaries.
565 if (!IS_NOSEC(inode))
566 return file_remove_privs(file);
571 * xfs_file_dio_aio_write - handle direct IO writes
573 * Lock the inode appropriately to prepare for and issue a direct IO write.
574 * By separating it from the buffered write path we remove all the tricky to
575 * follow locking changes and looping.
577 * If there are cached pages or we're extending the file, we need IOLOCK_EXCL
578 * until we're sure the bytes at the new EOF have been zeroed and/or the cached
579 * pages are flushed out.
581 * In most cases the direct IO writes will be done holding IOLOCK_SHARED
582 * allowing them to be done in parallel with reads and other direct IO writes.
583 * However, if the IO is not aligned to filesystem blocks, the direct IO layer
584 * needs to do sub-block zeroing and that requires serialisation against other
585 * direct IOs to the same block. In this case we need to serialise the
586 * submission of the unaligned IOs so that we don't get racing block zeroing in
587 * the dio layer. To avoid the problem with aio, we also need to wait for
588 * outstanding IOs to complete so that unwritten extent conversion is completed
589 * before we try to map the overlapping block. This is currently implemented by
590 * hitting it with a big hammer (i.e. inode_dio_wait()).
592 * Returns with locks held indicated by @iolock and errors indicated by
593 * negative return values.
596 xfs_file_dio_aio_write(
598 struct iov_iter *from)
600 struct file *file = iocb->ki_filp;
601 struct address_space *mapping = file->f_mapping;
602 struct inode *inode = mapping->host;
603 struct xfs_inode *ip = XFS_I(inode);
604 struct xfs_mount *mp = ip->i_mount;
606 int unaligned_io = 0;
608 size_t count = iov_iter_count(from);
610 struct iov_iter data;
611 struct xfs_buftarg *target = XFS_IS_REALTIME_INODE(ip) ?
612 mp->m_rtdev_targp : mp->m_ddev_targp;
614 /* DIO must be aligned to device logical sector size */
615 if ((iocb->ki_pos | count) & target->bt_logical_sectormask)
618 /* "unaligned" here means not aligned to a filesystem block */
619 if ((iocb->ki_pos & mp->m_blockmask) ||
620 ((iocb->ki_pos + count) & mp->m_blockmask))
624 * We don't need to take an exclusive lock unless there page cache needs
625 * to be invalidated or unaligned IO is being executed. We don't need to
626 * consider the EOF extension case here because
627 * xfs_file_aio_write_checks() will relock the inode as necessary for
628 * EOF zeroing cases and fill out the new inode size as appropriate.
630 if (unaligned_io || mapping->nrpages)
631 iolock = XFS_IOLOCK_EXCL;
633 iolock = XFS_IOLOCK_SHARED;
634 xfs_rw_ilock(ip, iolock);
637 * Recheck if there are cached pages that need invalidate after we got
638 * the iolock to protect against other threads adding new pages while
639 * we were waiting for the iolock.
641 if (mapping->nrpages && iolock == XFS_IOLOCK_SHARED) {
642 xfs_rw_iunlock(ip, iolock);
643 iolock = XFS_IOLOCK_EXCL;
644 xfs_rw_ilock(ip, iolock);
647 ret = xfs_file_aio_write_checks(iocb, from, &iolock);
650 count = iov_iter_count(from);
651 end = iocb->ki_pos + count - 1;
654 * See xfs_file_dio_aio_read() for why we do a full-file flush here.
656 if (mapping->nrpages) {
657 ret = filemap_write_and_wait(VFS_I(ip)->i_mapping);
661 * Invalidate whole pages. This can return an error if we fail
662 * to invalidate a page, but this should never happen on XFS.
663 * Warn if it does fail.
665 ret = invalidate_inode_pages2(VFS_I(ip)->i_mapping);
671 * If we are doing unaligned IO, wait for all other IO to drain,
672 * otherwise demote the lock if we had to flush cached pages
675 inode_dio_wait(inode);
676 else if (iolock == XFS_IOLOCK_EXCL) {
677 xfs_rw_ilock_demote(ip, XFS_IOLOCK_EXCL);
678 iolock = XFS_IOLOCK_SHARED;
681 trace_xfs_file_direct_write(ip, count, iocb->ki_pos);
684 ret = __blockdev_direct_IO(iocb, inode, target->bt_bdev, &data,
685 xfs_get_blocks_direct, xfs_end_io_direct_write,
686 NULL, DIO_ASYNC_EXTEND);
688 /* see generic_file_direct_write() for why this is necessary */
689 if (mapping->nrpages) {
690 invalidate_inode_pages2_range(mapping,
691 iocb->ki_pos >> PAGE_SHIFT,
697 iov_iter_advance(from, ret);
700 xfs_rw_iunlock(ip, iolock);
703 * No fallback to buffered IO on errors for XFS, direct IO will either
704 * complete fully or fail.
706 ASSERT(ret < 0 || ret == count);
710 static noinline ssize_t
713 struct iov_iter *from)
715 struct address_space *mapping = iocb->ki_filp->f_mapping;
716 struct inode *inode = mapping->host;
717 struct xfs_inode *ip = XFS_I(inode);
718 struct xfs_mount *mp = ip->i_mount;
720 int unaligned_io = 0;
722 struct iov_iter data;
724 /* "unaligned" here means not aligned to a filesystem block */
725 if ((iocb->ki_pos & mp->m_blockmask) ||
726 ((iocb->ki_pos + iov_iter_count(from)) & mp->m_blockmask)) {
728 iolock = XFS_IOLOCK_EXCL;
729 } else if (mapping->nrpages) {
730 iolock = XFS_IOLOCK_EXCL;
732 iolock = XFS_IOLOCK_SHARED;
734 xfs_rw_ilock(ip, iolock);
736 ret = xfs_file_aio_write_checks(iocb, from, &iolock);
741 * Yes, even DAX files can have page cache attached to them: A zeroed
742 * page is inserted into the pagecache when we have to serve a write
743 * fault on a hole. It should never be dirtied and can simply be
744 * dropped from the pagecache once we get real data for the page.
746 * XXX: This is racy against mmap, and there's nothing we can do about
747 * it. dax_do_io() should really do this invalidation internally as
748 * it will know if we've allocated over a holei for this specific IO and
749 * if so it needs to update the mapping tree and invalidate existing
750 * PTEs over the newly allocated range. Remove this invalidation when
751 * dax_do_io() is fixed up.
753 if (mapping->nrpages) {
754 loff_t end = iocb->ki_pos + iov_iter_count(from) - 1;
756 ret = invalidate_inode_pages2_range(mapping,
757 iocb->ki_pos >> PAGE_SHIFT,
762 if (iolock == XFS_IOLOCK_EXCL && !unaligned_io) {
763 xfs_rw_ilock_demote(ip, XFS_IOLOCK_EXCL);
764 iolock = XFS_IOLOCK_SHARED;
767 trace_xfs_file_dax_write(ip, iov_iter_count(from), iocb->ki_pos);
770 ret = dax_do_io(iocb, inode, &data, xfs_get_blocks_direct,
771 xfs_end_io_direct_write, 0);
774 iov_iter_advance(from, ret);
777 xfs_rw_iunlock(ip, iolock);
782 xfs_file_buffered_aio_write(
784 struct iov_iter *from)
786 struct file *file = iocb->ki_filp;
787 struct address_space *mapping = file->f_mapping;
788 struct inode *inode = mapping->host;
789 struct xfs_inode *ip = XFS_I(inode);
792 int iolock = XFS_IOLOCK_EXCL;
794 xfs_rw_ilock(ip, iolock);
796 ret = xfs_file_aio_write_checks(iocb, from, &iolock);
800 /* We can write back this queue in page reclaim */
801 current->backing_dev_info = inode_to_bdi(inode);
804 trace_xfs_file_buffered_write(ip, iov_iter_count(from), iocb->ki_pos);
805 ret = iomap_file_buffered_write(iocb, from, &xfs_iomap_ops);
806 if (likely(ret >= 0))
810 * If we hit a space limit, try to free up some lingering preallocated
811 * space before returning an error. In the case of ENOSPC, first try to
812 * write back all dirty inodes to free up some of the excess reserved
813 * metadata space. This reduces the chances that the eofblocks scan
814 * waits on dirty mappings. Since xfs_flush_inodes() is serialized, this
815 * also behaves as a filter to prevent too many eofblocks scans from
816 * running at the same time.
818 if (ret == -EDQUOT && !enospc) {
819 enospc = xfs_inode_free_quota_eofblocks(ip);
822 } else if (ret == -ENOSPC && !enospc) {
823 struct xfs_eofblocks eofb = {0};
826 xfs_flush_inodes(ip->i_mount);
827 eofb.eof_scan_owner = ip->i_ino; /* for locking */
828 eofb.eof_flags = XFS_EOF_FLAGS_SYNC;
829 xfs_icache_free_eofblocks(ip->i_mount, &eofb);
833 current->backing_dev_info = NULL;
835 xfs_rw_iunlock(ip, iolock);
842 struct iov_iter *from)
844 struct file *file = iocb->ki_filp;
845 struct address_space *mapping = file->f_mapping;
846 struct inode *inode = mapping->host;
847 struct xfs_inode *ip = XFS_I(inode);
849 size_t ocount = iov_iter_count(from);
851 XFS_STATS_INC(ip->i_mount, xs_write_calls);
856 if (XFS_FORCED_SHUTDOWN(ip->i_mount))
860 ret = xfs_file_dax_write(iocb, from);
861 else if (iocb->ki_flags & IOCB_DIRECT)
862 ret = xfs_file_dio_aio_write(iocb, from);
864 ret = xfs_file_buffered_aio_write(iocb, from);
867 XFS_STATS_ADD(ip->i_mount, xs_write_bytes, ret);
869 /* Handle various SYNC-type writes */
870 ret = generic_write_sync(iocb, ret);
875 #define XFS_FALLOC_FL_SUPPORTED \
876 (FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE | \
877 FALLOC_FL_COLLAPSE_RANGE | FALLOC_FL_ZERO_RANGE | \
878 FALLOC_FL_INSERT_RANGE)
887 struct inode *inode = file_inode(file);
888 struct xfs_inode *ip = XFS_I(inode);
890 enum xfs_prealloc_flags flags = 0;
891 uint iolock = XFS_IOLOCK_EXCL;
893 bool do_file_insert = 0;
895 if (!S_ISREG(inode->i_mode))
897 if (mode & ~XFS_FALLOC_FL_SUPPORTED)
900 xfs_ilock(ip, iolock);
901 error = xfs_break_layouts(inode, &iolock, false);
905 xfs_ilock(ip, XFS_MMAPLOCK_EXCL);
906 iolock |= XFS_MMAPLOCK_EXCL;
908 if (mode & FALLOC_FL_PUNCH_HOLE) {
909 error = xfs_free_file_space(ip, offset, len);
912 } else if (mode & FALLOC_FL_COLLAPSE_RANGE) {
913 unsigned blksize_mask = (1 << inode->i_blkbits) - 1;
915 if (offset & blksize_mask || len & blksize_mask) {
921 * There is no need to overlap collapse range with EOF,
922 * in which case it is effectively a truncate operation
924 if (offset + len >= i_size_read(inode)) {
929 new_size = i_size_read(inode) - len;
931 error = xfs_collapse_file_space(ip, offset, len);
934 } else if (mode & FALLOC_FL_INSERT_RANGE) {
935 unsigned blksize_mask = (1 << inode->i_blkbits) - 1;
937 new_size = i_size_read(inode) + len;
938 if (offset & blksize_mask || len & blksize_mask) {
943 /* check the new inode size does not wrap through zero */
944 if (new_size > inode->i_sb->s_maxbytes) {
949 /* Offset should be less than i_size */
950 if (offset >= i_size_read(inode)) {
956 flags |= XFS_PREALLOC_SET;
958 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
959 offset + len > i_size_read(inode)) {
960 new_size = offset + len;
961 error = inode_newsize_ok(inode, new_size);
966 if (mode & FALLOC_FL_ZERO_RANGE)
967 error = xfs_zero_file_space(ip, offset, len);
969 error = xfs_alloc_file_space(ip, offset, len,
975 if (file->f_flags & O_DSYNC)
976 flags |= XFS_PREALLOC_SYNC;
978 error = xfs_update_prealloc_flags(ip, flags);
982 /* Change file size if needed */
986 iattr.ia_valid = ATTR_SIZE;
987 iattr.ia_size = new_size;
988 error = xfs_setattr_size(ip, &iattr);
994 * Perform hole insertion now that the file size has been
995 * updated so that if we crash during the operation we don't
996 * leave shifted extents past EOF and hence losing access to
997 * the data that is contained within them.
1000 error = xfs_insert_file_space(ip, offset, len);
1003 xfs_iunlock(ip, iolock);
1010 struct inode *inode,
1013 if (!(file->f_flags & O_LARGEFILE) && i_size_read(inode) > MAX_NON_LFS)
1015 if (XFS_FORCED_SHUTDOWN(XFS_M(inode->i_sb)))
1022 struct inode *inode,
1025 struct xfs_inode *ip = XFS_I(inode);
1029 error = xfs_file_open(inode, file);
1034 * If there are any blocks, read-ahead block 0 as we're almost
1035 * certain to have the next operation be a read there.
1037 mode = xfs_ilock_data_map_shared(ip);
1038 if (ip->i_d.di_nextents > 0)
1039 xfs_dir3_data_readahead(ip, 0, -1);
1040 xfs_iunlock(ip, mode);
1046 struct inode *inode,
1049 return xfs_release(XFS_I(inode));
1055 struct dir_context *ctx)
1057 struct inode *inode = file_inode(file);
1058 xfs_inode_t *ip = XFS_I(inode);
1062 * The Linux API doesn't pass down the total size of the buffer
1063 * we read into down to the filesystem. With the filldir concept
1064 * it's not needed for correct information, but the XFS dir2 leaf
1065 * code wants an estimate of the buffer size to calculate it's
1066 * readahead window and size the buffers used for mapping to
1069 * Try to give it an estimate that's good enough, maybe at some
1070 * point we can change the ->readdir prototype to include the
1071 * buffer size. For now we use the current glibc buffer size.
1073 bufsize = (size_t)min_t(loff_t, 32768, ip->i_d.di_size);
1075 return xfs_readdir(ip, ctx, bufsize);
1079 * This type is designed to indicate the type of offset we would like
1080 * to search from page cache for xfs_seek_hole_data().
1088 * Lookup the desired type of offset from the given page.
1090 * On success, return true and the offset argument will point to the
1091 * start of the region that was found. Otherwise this function will
1092 * return false and keep the offset argument unchanged.
1095 xfs_lookup_buffer_offset(
1100 loff_t lastoff = page_offset(page);
1102 struct buffer_head *bh, *head;
1104 bh = head = page_buffers(page);
1107 * Unwritten extents that have data in the page
1108 * cache covering them can be identified by the
1109 * BH_Unwritten state flag. Pages with multiple
1110 * buffers might have a mix of holes, data and
1111 * unwritten extents - any buffer with valid
1112 * data in it should have BH_Uptodate flag set
1115 if (buffer_unwritten(bh) ||
1116 buffer_uptodate(bh)) {
1117 if (type == DATA_OFF)
1120 if (type == HOLE_OFF)
1128 lastoff += bh->b_size;
1129 } while ((bh = bh->b_this_page) != head);
1135 * This routine is called to find out and return a data or hole offset
1136 * from the page cache for unwritten extents according to the desired
1137 * type for xfs_seek_hole_data().
1139 * The argument offset is used to tell where we start to search from the
1140 * page cache. Map is used to figure out the end points of the range to
1143 * Return true if the desired type of offset was found, and the argument
1144 * offset is filled with that address. Otherwise, return false and keep
1148 xfs_find_get_desired_pgoff(
1149 struct inode *inode,
1150 struct xfs_bmbt_irec *map,
1154 struct xfs_inode *ip = XFS_I(inode);
1155 struct xfs_mount *mp = ip->i_mount;
1156 struct pagevec pvec;
1160 loff_t startoff = *offset;
1161 loff_t lastoff = startoff;
1164 pagevec_init(&pvec, 0);
1166 index = startoff >> PAGE_SHIFT;
1167 endoff = XFS_FSB_TO_B(mp, map->br_startoff + map->br_blockcount);
1168 end = endoff >> PAGE_SHIFT;
1174 want = min_t(pgoff_t, end - index, PAGEVEC_SIZE);
1175 nr_pages = pagevec_lookup(&pvec, inode->i_mapping, index,
1178 * No page mapped into given range. If we are searching holes
1179 * and if this is the first time we got into the loop, it means
1180 * that the given offset is landed in a hole, return it.
1182 * If we have already stepped through some block buffers to find
1183 * holes but they all contains data. In this case, the last
1184 * offset is already updated and pointed to the end of the last
1185 * mapped page, if it does not reach the endpoint to search,
1186 * that means there should be a hole between them.
1188 if (nr_pages == 0) {
1189 /* Data search found nothing */
1190 if (type == DATA_OFF)
1193 ASSERT(type == HOLE_OFF);
1194 if (lastoff == startoff || lastoff < endoff) {
1202 * At lease we found one page. If this is the first time we
1203 * step into the loop, and if the first page index offset is
1204 * greater than the given search offset, a hole was found.
1206 if (type == HOLE_OFF && lastoff == startoff &&
1207 lastoff < page_offset(pvec.pages[0])) {
1212 for (i = 0; i < nr_pages; i++) {
1213 struct page *page = pvec.pages[i];
1217 * At this point, the page may be truncated or
1218 * invalidated (changing page->mapping to NULL),
1219 * or even swizzled back from swapper_space to tmpfs
1220 * file mapping. However, page->index will not change
1221 * because we have a reference on the page.
1223 * Searching done if the page index is out of range.
1224 * If the current offset is not reaches the end of
1225 * the specified search range, there should be a hole
1228 if (page->index > end) {
1229 if (type == HOLE_OFF && lastoff < endoff) {
1238 * Page truncated or invalidated(page->mapping == NULL).
1239 * We can freely skip it and proceed to check the next
1242 if (unlikely(page->mapping != inode->i_mapping)) {
1247 if (!page_has_buffers(page)) {
1252 found = xfs_lookup_buffer_offset(page, &b_offset, type);
1255 * The found offset may be less than the start
1256 * point to search if this is the first time to
1259 *offset = max_t(loff_t, startoff, b_offset);
1265 * We either searching data but nothing was found, or
1266 * searching hole but found a data buffer. In either
1267 * case, probably the next page contains the desired
1268 * things, update the last offset to it so.
1270 lastoff = page_offset(page) + PAGE_SIZE;
1275 * The number of returned pages less than our desired, search
1276 * done. In this case, nothing was found for searching data,
1277 * but we found a hole behind the last offset.
1279 if (nr_pages < want) {
1280 if (type == HOLE_OFF) {
1287 index = pvec.pages[i - 1]->index + 1;
1288 pagevec_release(&pvec);
1289 } while (index <= end);
1292 pagevec_release(&pvec);
1297 * caller must lock inode with xfs_ilock_data_map_shared,
1298 * can we craft an appropriate ASSERT?
1300 * end is because the VFS-level lseek interface is defined such that any
1301 * offset past i_size shall return -ENXIO, but we use this for quota code
1302 * which does not maintain i_size, and we want to SEEK_DATA past i_size.
1305 __xfs_seek_hole_data(
1306 struct inode *inode,
1311 struct xfs_inode *ip = XFS_I(inode);
1312 struct xfs_mount *mp = ip->i_mount;
1313 loff_t uninitialized_var(offset);
1314 xfs_fileoff_t fsbno;
1315 xfs_filblks_t lastbno;
1324 * Try to read extents from the first block indicated
1325 * by fsbno to the end block of the file.
1327 fsbno = XFS_B_TO_FSBT(mp, start);
1328 lastbno = XFS_B_TO_FSB(mp, end);
1331 struct xfs_bmbt_irec map[2];
1335 error = xfs_bmapi_read(ip, fsbno, lastbno - fsbno, map, &nmap,
1340 /* No extents at given offset, must be beyond EOF */
1346 for (i = 0; i < nmap; i++) {
1347 offset = max_t(loff_t, start,
1348 XFS_FSB_TO_B(mp, map[i].br_startoff));
1350 /* Landed in the hole we wanted? */
1351 if (whence == SEEK_HOLE &&
1352 map[i].br_startblock == HOLESTARTBLOCK)
1355 /* Landed in the data extent we wanted? */
1356 if (whence == SEEK_DATA &&
1357 (map[i].br_startblock == DELAYSTARTBLOCK ||
1358 (map[i].br_state == XFS_EXT_NORM &&
1359 !isnullstartblock(map[i].br_startblock))))
1363 * Landed in an unwritten extent, try to search
1364 * for hole or data from page cache.
1366 if (map[i].br_state == XFS_EXT_UNWRITTEN) {
1367 if (xfs_find_get_desired_pgoff(inode, &map[i],
1368 whence == SEEK_HOLE ? HOLE_OFF : DATA_OFF,
1375 * We only received one extent out of the two requested. This
1376 * means we've hit EOF and didn't find what we are looking for.
1380 * If we were looking for a hole, set offset to
1381 * the end of the file (i.e., there is an implicit
1382 * hole at the end of any file).
1384 if (whence == SEEK_HOLE) {
1389 * If we were looking for data, it's nowhere to be found
1391 ASSERT(whence == SEEK_DATA);
1399 * Nothing was found, proceed to the next round of search
1400 * if the next reading offset is not at or beyond EOF.
1402 fsbno = map[i - 1].br_startoff + map[i - 1].br_blockcount;
1403 start = XFS_FSB_TO_B(mp, fsbno);
1405 if (whence == SEEK_HOLE) {
1409 ASSERT(whence == SEEK_DATA);
1417 * If at this point we have found the hole we wanted, the returned
1418 * offset may be bigger than the file size as it may be aligned to
1419 * page boundary for unwritten extents. We need to deal with this
1420 * situation in particular.
1422 if (whence == SEEK_HOLE)
1423 offset = min_t(loff_t, offset, end);
1437 struct inode *inode = file->f_mapping->host;
1438 struct xfs_inode *ip = XFS_I(inode);
1439 struct xfs_mount *mp = ip->i_mount;
1444 if (XFS_FORCED_SHUTDOWN(mp))
1447 lock = xfs_ilock_data_map_shared(ip);
1449 end = i_size_read(inode);
1450 offset = __xfs_seek_hole_data(inode, start, end, whence);
1456 offset = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
1459 xfs_iunlock(ip, lock);
1476 return generic_file_llseek(file, offset, whence);
1479 return xfs_seek_hole_data(file, offset, whence);
1486 * Locking for serialisation of IO during page faults. This results in a lock
1490 * sb_start_pagefault(vfs, freeze)
1491 * i_mmaplock (XFS - truncate serialisation)
1493 * i_lock (XFS - extent map serialisation)
1497 * mmap()d file has taken write protection fault and is being made writable. We
1498 * can set the page state up correctly for a writable page, which means we can
1499 * do correct delalloc accounting (ENOSPC checking!) and unwritten extent
1503 xfs_filemap_page_mkwrite(
1504 struct vm_area_struct *vma,
1505 struct vm_fault *vmf)
1507 struct inode *inode = file_inode(vma->vm_file);
1510 trace_xfs_filemap_page_mkwrite(XFS_I(inode));
1512 sb_start_pagefault(inode->i_sb);
1513 file_update_time(vma->vm_file);
1514 xfs_ilock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1516 if (IS_DAX(inode)) {
1517 ret = dax_mkwrite(vma, vmf, xfs_get_blocks_dax_fault);
1519 ret = iomap_page_mkwrite(vma, vmf, &xfs_iomap_ops);
1520 ret = block_page_mkwrite_return(ret);
1523 xfs_iunlock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1524 sb_end_pagefault(inode->i_sb);
1531 struct vm_area_struct *vma,
1532 struct vm_fault *vmf)
1534 struct inode *inode = file_inode(vma->vm_file);
1537 trace_xfs_filemap_fault(XFS_I(inode));
1539 /* DAX can shortcut the normal fault path on write faults! */
1540 if ((vmf->flags & FAULT_FLAG_WRITE) && IS_DAX(inode))
1541 return xfs_filemap_page_mkwrite(vma, vmf);
1543 xfs_ilock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1544 if (IS_DAX(inode)) {
1546 * we do not want to trigger unwritten extent conversion on read
1547 * faults - that is unnecessary overhead and would also require
1548 * changes to xfs_get_blocks_direct() to map unwritten extent
1549 * ioend for conversion on read-only mappings.
1551 ret = dax_fault(vma, vmf, xfs_get_blocks_dax_fault);
1553 ret = filemap_fault(vma, vmf);
1554 xfs_iunlock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1560 * Similar to xfs_filemap_fault(), the DAX fault path can call into here on
1561 * both read and write faults. Hence we need to handle both cases. There is no
1562 * ->pmd_mkwrite callout for huge pages, so we have a single function here to
1563 * handle both cases here. @flags carries the information on the type of fault
1567 xfs_filemap_pmd_fault(
1568 struct vm_area_struct *vma,
1573 struct inode *inode = file_inode(vma->vm_file);
1574 struct xfs_inode *ip = XFS_I(inode);
1578 return VM_FAULT_FALLBACK;
1580 trace_xfs_filemap_pmd_fault(ip);
1582 if (flags & FAULT_FLAG_WRITE) {
1583 sb_start_pagefault(inode->i_sb);
1584 file_update_time(vma->vm_file);
1587 xfs_ilock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1588 ret = dax_pmd_fault(vma, addr, pmd, flags, xfs_get_blocks_dax_fault);
1589 xfs_iunlock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1591 if (flags & FAULT_FLAG_WRITE)
1592 sb_end_pagefault(inode->i_sb);
1598 * pfn_mkwrite was originally inteneded to ensure we capture time stamp
1599 * updates on write faults. In reality, it's need to serialise against
1600 * truncate similar to page_mkwrite. Hence we cycle the XFS_MMAPLOCK_SHARED
1601 * to ensure we serialise the fault barrier in place.
1604 xfs_filemap_pfn_mkwrite(
1605 struct vm_area_struct *vma,
1606 struct vm_fault *vmf)
1609 struct inode *inode = file_inode(vma->vm_file);
1610 struct xfs_inode *ip = XFS_I(inode);
1611 int ret = VM_FAULT_NOPAGE;
1614 trace_xfs_filemap_pfn_mkwrite(ip);
1616 sb_start_pagefault(inode->i_sb);
1617 file_update_time(vma->vm_file);
1619 /* check if the faulting page hasn't raced with truncate */
1620 xfs_ilock(ip, XFS_MMAPLOCK_SHARED);
1621 size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
1622 if (vmf->pgoff >= size)
1623 ret = VM_FAULT_SIGBUS;
1624 else if (IS_DAX(inode))
1625 ret = dax_pfn_mkwrite(vma, vmf);
1626 xfs_iunlock(ip, XFS_MMAPLOCK_SHARED);
1627 sb_end_pagefault(inode->i_sb);
1632 static const struct vm_operations_struct xfs_file_vm_ops = {
1633 .fault = xfs_filemap_fault,
1634 .pmd_fault = xfs_filemap_pmd_fault,
1635 .map_pages = filemap_map_pages,
1636 .page_mkwrite = xfs_filemap_page_mkwrite,
1637 .pfn_mkwrite = xfs_filemap_pfn_mkwrite,
1643 struct vm_area_struct *vma)
1645 file_accessed(filp);
1646 vma->vm_ops = &xfs_file_vm_ops;
1647 if (IS_DAX(file_inode(filp)))
1648 vma->vm_flags |= VM_MIXEDMAP | VM_HUGEPAGE;
1652 const struct file_operations xfs_file_operations = {
1653 .llseek = xfs_file_llseek,
1654 .read_iter = xfs_file_read_iter,
1655 .write_iter = xfs_file_write_iter,
1656 .splice_read = xfs_file_splice_read,
1657 .splice_write = iter_file_splice_write,
1658 .unlocked_ioctl = xfs_file_ioctl,
1659 #ifdef CONFIG_COMPAT
1660 .compat_ioctl = xfs_file_compat_ioctl,
1662 .mmap = xfs_file_mmap,
1663 .open = xfs_file_open,
1664 .release = xfs_file_release,
1665 .fsync = xfs_file_fsync,
1666 .fallocate = xfs_file_fallocate,
1669 const struct file_operations xfs_dir_file_operations = {
1670 .open = xfs_dir_open,
1671 .read = generic_read_dir,
1672 .iterate_shared = xfs_file_readdir,
1673 .llseek = generic_file_llseek,
1674 .unlocked_ioctl = xfs_file_ioctl,
1675 #ifdef CONFIG_COMPAT
1676 .compat_ioctl = xfs_file_compat_ioctl,
1678 .fsync = xfs_dir_fsync,