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"
41 #include <linux/dcache.h>
42 #include <linux/falloc.h>
43 #include <linux/pagevec.h>
44 #include <linux/backing-dev.h>
46 static const struct vm_operations_struct xfs_file_vm_ops;
49 * Locking primitives for read and write IO paths to ensure we consistently use
50 * and order the inode->i_mutex, ip->i_lock and ip->i_iolock.
57 if (type & XFS_IOLOCK_EXCL)
58 inode_lock(VFS_I(ip));
67 xfs_iunlock(ip, type);
68 if (type & XFS_IOLOCK_EXCL)
69 inode_unlock(VFS_I(ip));
77 xfs_ilock_demote(ip, type);
78 if (type & XFS_IOLOCK_EXCL)
79 inode_unlock(VFS_I(ip));
83 * xfs_iozero clears the specified range supplied via the page cache (except in
84 * the DAX case). Writes through the page cache will allocate blocks over holes,
85 * though the callers usually map the holes first and avoid them. If a block is
86 * not completely zeroed, then it will be read from disk before being partially
89 * In the DAX case, we can just directly write to the underlying pages. This
90 * will not allocate blocks, but will avoid holes and unwritten extents and so
91 * not do unnecessary work.
95 struct xfs_inode *ip, /* inode */
96 loff_t pos, /* offset in file */
97 size_t count) /* size of data to zero */
100 struct address_space *mapping;
104 mapping = VFS_I(ip)->i_mapping;
106 unsigned offset, bytes;
109 offset = (pos & (PAGE_SIZE -1)); /* Within page */
110 bytes = PAGE_SIZE - offset;
114 if (IS_DAX(VFS_I(ip))) {
115 status = dax_zero_page_range(VFS_I(ip), pos, bytes,
116 xfs_get_blocks_direct);
120 status = pagecache_write_begin(NULL, mapping, pos, bytes,
121 AOP_FLAG_UNINTERRUPTIBLE,
126 zero_user(page, offset, bytes);
128 status = pagecache_write_end(NULL, mapping, pos, bytes,
129 bytes, page, fsdata);
130 WARN_ON(status <= 0); /* can't return less than zero! */
141 xfs_update_prealloc_flags(
142 struct xfs_inode *ip,
143 enum xfs_prealloc_flags flags)
145 struct xfs_trans *tp;
148 tp = xfs_trans_alloc(ip->i_mount, XFS_TRANS_WRITEID);
149 error = xfs_trans_reserve(tp, &M_RES(ip->i_mount)->tr_writeid, 0, 0);
151 xfs_trans_cancel(tp);
155 xfs_ilock(ip, XFS_ILOCK_EXCL);
156 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
158 if (!(flags & XFS_PREALLOC_INVISIBLE)) {
159 VFS_I(ip)->i_mode &= ~S_ISUID;
160 if (VFS_I(ip)->i_mode & S_IXGRP)
161 VFS_I(ip)->i_mode &= ~S_ISGID;
162 xfs_trans_ichgtime(tp, ip, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
165 if (flags & XFS_PREALLOC_SET)
166 ip->i_d.di_flags |= XFS_DIFLAG_PREALLOC;
167 if (flags & XFS_PREALLOC_CLEAR)
168 ip->i_d.di_flags &= ~XFS_DIFLAG_PREALLOC;
170 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
171 if (flags & XFS_PREALLOC_SYNC)
172 xfs_trans_set_sync(tp);
173 return xfs_trans_commit(tp);
177 * Fsync operations on directories are much simpler than on regular files,
178 * as there is no file data to flush, and thus also no need for explicit
179 * cache flush operations, and there are no non-transaction metadata updates
180 * on directories either.
189 struct xfs_inode *ip = XFS_I(file->f_mapping->host);
190 struct xfs_mount *mp = ip->i_mount;
193 trace_xfs_dir_fsync(ip);
195 xfs_ilock(ip, XFS_ILOCK_SHARED);
196 if (xfs_ipincount(ip))
197 lsn = ip->i_itemp->ili_last_lsn;
198 xfs_iunlock(ip, XFS_ILOCK_SHARED);
202 return _xfs_log_force_lsn(mp, lsn, XFS_LOG_SYNC, NULL);
212 struct inode *inode = file->f_mapping->host;
213 struct xfs_inode *ip = XFS_I(inode);
214 struct xfs_mount *mp = ip->i_mount;
219 trace_xfs_file_fsync(ip);
221 error = filemap_write_and_wait_range(inode->i_mapping, start, end);
225 if (XFS_FORCED_SHUTDOWN(mp))
228 xfs_iflags_clear(ip, XFS_ITRUNCATED);
230 if (mp->m_flags & XFS_MOUNT_BARRIER) {
232 * If we have an RT and/or log subvolume we need to make sure
233 * to flush the write cache the device used for file data
234 * first. This is to ensure newly written file data make
235 * it to disk before logging the new inode size in case of
236 * an extending write.
238 if (XFS_IS_REALTIME_INODE(ip))
239 xfs_blkdev_issue_flush(mp->m_rtdev_targp);
240 else if (mp->m_logdev_targp != mp->m_ddev_targp)
241 xfs_blkdev_issue_flush(mp->m_ddev_targp);
245 * All metadata updates are logged, which means that we just have to
246 * flush the log up to the latest LSN that touched the inode. If we have
247 * concurrent fsync/fdatasync() calls, we need them to all block on the
248 * log force before we clear the ili_fsync_fields field. This ensures
249 * that we don't get a racing sync operation that does not wait for the
250 * metadata to hit the journal before returning. If we race with
251 * clearing the ili_fsync_fields, then all that will happen is the log
252 * force will do nothing as the lsn will already be on disk. We can't
253 * race with setting ili_fsync_fields because that is done under
254 * XFS_ILOCK_EXCL, and that can't happen because we hold the lock shared
255 * until after the ili_fsync_fields is cleared.
257 xfs_ilock(ip, XFS_ILOCK_SHARED);
258 if (xfs_ipincount(ip)) {
260 (ip->i_itemp->ili_fsync_fields & ~XFS_ILOG_TIMESTAMP))
261 lsn = ip->i_itemp->ili_last_lsn;
265 error = _xfs_log_force_lsn(mp, lsn, XFS_LOG_SYNC, &log_flushed);
266 ip->i_itemp->ili_fsync_fields = 0;
268 xfs_iunlock(ip, XFS_ILOCK_SHARED);
271 * If we only have a single device, and the log force about was
272 * a no-op we might have to flush the data device cache here.
273 * This can only happen for fdatasync/O_DSYNC if we were overwriting
274 * an already allocated file and thus do not have any metadata to
277 if ((mp->m_flags & XFS_MOUNT_BARRIER) &&
278 mp->m_logdev_targp == mp->m_ddev_targp &&
279 !XFS_IS_REALTIME_INODE(ip) &&
281 xfs_blkdev_issue_flush(mp->m_ddev_targp);
291 struct file *file = iocb->ki_filp;
292 struct inode *inode = file->f_mapping->host;
293 struct xfs_inode *ip = XFS_I(inode);
294 struct xfs_mount *mp = ip->i_mount;
295 size_t size = iov_iter_count(to);
299 loff_t pos = iocb->ki_pos;
301 XFS_STATS_INC(mp, xs_read_calls);
303 if (unlikely(iocb->ki_flags & IOCB_DIRECT))
304 ioflags |= XFS_IO_ISDIRECT;
305 if (file->f_mode & FMODE_NOCMTIME)
306 ioflags |= XFS_IO_INVIS;
308 if ((ioflags & XFS_IO_ISDIRECT) && !IS_DAX(inode)) {
309 xfs_buftarg_t *target =
310 XFS_IS_REALTIME_INODE(ip) ?
311 mp->m_rtdev_targp : mp->m_ddev_targp;
312 /* DIO must be aligned to device logical sector size */
313 if ((pos | size) & target->bt_logical_sectormask) {
314 if (pos == i_size_read(inode))
320 n = mp->m_super->s_maxbytes - pos;
321 if (n <= 0 || size == 0)
327 if (XFS_FORCED_SHUTDOWN(mp))
331 * Locking is a bit tricky here. If we take an exclusive lock for direct
332 * IO, we effectively serialise all new concurrent read IO to this file
333 * and block it behind IO that is currently in progress because IO in
334 * progress holds the IO lock shared. We only need to hold the lock
335 * exclusive to blow away the page cache, so only take lock exclusively
336 * if the page cache needs invalidation. This allows the normal direct
337 * IO case of no page cache pages to proceeed concurrently without
340 xfs_rw_ilock(ip, XFS_IOLOCK_SHARED);
341 if ((ioflags & XFS_IO_ISDIRECT) && inode->i_mapping->nrpages) {
342 xfs_rw_iunlock(ip, XFS_IOLOCK_SHARED);
343 xfs_rw_ilock(ip, XFS_IOLOCK_EXCL);
346 * The generic dio code only flushes the range of the particular
347 * I/O. Because we take an exclusive lock here, this whole
348 * sequence is considerably more expensive for us. This has a
349 * noticeable performance impact for any file with cached pages,
350 * even when outside of the range of the particular I/O.
352 * Hence, amortize the cost of the lock against a full file
353 * flush and reduce the chances of repeated iolock cycles going
356 if (inode->i_mapping->nrpages) {
357 ret = filemap_write_and_wait(VFS_I(ip)->i_mapping);
359 xfs_rw_iunlock(ip, XFS_IOLOCK_EXCL);
364 * Invalidate whole pages. This can return an error if
365 * we fail to invalidate a page, but this should never
366 * happen on XFS. Warn if it does fail.
368 ret = invalidate_inode_pages2(VFS_I(ip)->i_mapping);
372 xfs_rw_ilock_demote(ip, XFS_IOLOCK_EXCL);
375 trace_xfs_file_read(ip, size, pos, ioflags);
377 ret = generic_file_read_iter(iocb, to);
379 XFS_STATS_ADD(mp, xs_read_bytes, ret);
381 xfs_rw_iunlock(ip, XFS_IOLOCK_SHARED);
386 xfs_file_splice_read(
389 struct pipe_inode_info *pipe,
393 struct xfs_inode *ip = XFS_I(infilp->f_mapping->host);
397 XFS_STATS_INC(ip->i_mount, xs_read_calls);
399 if (infilp->f_mode & FMODE_NOCMTIME)
400 ioflags |= XFS_IO_INVIS;
402 if (XFS_FORCED_SHUTDOWN(ip->i_mount))
405 trace_xfs_file_splice_read(ip, count, *ppos, ioflags);
408 * DAX inodes cannot ues the page cache for splice, so we have to push
409 * them through the VFS IO path. This means it goes through
410 * ->read_iter, which for us takes the XFS_IOLOCK_SHARED. Hence we
411 * cannot lock the splice operation at this level for DAX inodes.
413 if (IS_DAX(VFS_I(ip))) {
414 ret = default_file_splice_read(infilp, ppos, pipe, count,
419 xfs_rw_ilock(ip, XFS_IOLOCK_SHARED);
420 ret = generic_file_splice_read(infilp, ppos, pipe, count, flags);
421 xfs_rw_iunlock(ip, XFS_IOLOCK_SHARED);
424 XFS_STATS_ADD(ip->i_mount, xs_read_bytes, ret);
429 * This routine is called to handle zeroing any space in the last block of the
430 * file that is beyond the EOF. We do this since the size is being increased
431 * without writing anything to that block and we don't want to read the
432 * garbage on the disk.
434 STATIC int /* error (positive) */
436 struct xfs_inode *ip,
441 struct xfs_mount *mp = ip->i_mount;
442 xfs_fileoff_t last_fsb = XFS_B_TO_FSBT(mp, isize);
443 int zero_offset = XFS_B_FSB_OFFSET(mp, isize);
447 struct xfs_bmbt_irec imap;
449 xfs_ilock(ip, XFS_ILOCK_EXCL);
450 error = xfs_bmapi_read(ip, last_fsb, 1, &imap, &nimaps, 0);
451 xfs_iunlock(ip, XFS_ILOCK_EXCL);
458 * If the block underlying isize is just a hole, then there
459 * is nothing to zero.
461 if (imap.br_startblock == HOLESTARTBLOCK)
464 zero_len = mp->m_sb.sb_blocksize - zero_offset;
465 if (isize + zero_len > offset)
466 zero_len = offset - isize;
468 return xfs_iozero(ip, isize, zero_len);
472 * Zero any on disk space between the current EOF and the new, larger EOF.
474 * This handles the normal case of zeroing the remainder of the last block in
475 * the file and the unusual case of zeroing blocks out beyond the size of the
476 * file. This second case only happens with fixed size extents and when the
477 * system crashes before the inode size was updated but after blocks were
480 * Expects the iolock to be held exclusive, and will take the ilock internally.
482 int /* error (positive) */
484 struct xfs_inode *ip,
485 xfs_off_t offset, /* starting I/O offset */
486 xfs_fsize_t isize, /* current inode size */
489 struct xfs_mount *mp = ip->i_mount;
490 xfs_fileoff_t start_zero_fsb;
491 xfs_fileoff_t end_zero_fsb;
492 xfs_fileoff_t zero_count_fsb;
493 xfs_fileoff_t last_fsb;
494 xfs_fileoff_t zero_off;
495 xfs_fsize_t zero_len;
498 struct xfs_bmbt_irec imap;
500 ASSERT(xfs_isilocked(ip, XFS_IOLOCK_EXCL));
501 ASSERT(offset > isize);
503 trace_xfs_zero_eof(ip, isize, offset - isize);
506 * First handle zeroing the block on which isize resides.
508 * We only zero a part of that block so it is handled specially.
510 if (XFS_B_FSB_OFFSET(mp, isize) != 0) {
511 error = xfs_zero_last_block(ip, offset, isize, did_zeroing);
517 * Calculate the range between the new size and the old where blocks
518 * needing to be zeroed may exist.
520 * To get the block where the last byte in the file currently resides,
521 * we need to subtract one from the size and truncate back to a block
522 * boundary. We subtract 1 in case the size is exactly on a block
525 last_fsb = isize ? XFS_B_TO_FSBT(mp, isize - 1) : (xfs_fileoff_t)-1;
526 start_zero_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)isize);
527 end_zero_fsb = XFS_B_TO_FSBT(mp, offset - 1);
528 ASSERT((xfs_sfiloff_t)last_fsb < (xfs_sfiloff_t)start_zero_fsb);
529 if (last_fsb == end_zero_fsb) {
531 * The size was only incremented on its last block.
532 * We took care of that above, so just return.
537 ASSERT(start_zero_fsb <= end_zero_fsb);
538 while (start_zero_fsb <= end_zero_fsb) {
540 zero_count_fsb = end_zero_fsb - start_zero_fsb + 1;
542 xfs_ilock(ip, XFS_ILOCK_EXCL);
543 error = xfs_bmapi_read(ip, start_zero_fsb, zero_count_fsb,
545 xfs_iunlock(ip, XFS_ILOCK_EXCL);
551 if (imap.br_state == XFS_EXT_UNWRITTEN ||
552 imap.br_startblock == HOLESTARTBLOCK) {
553 start_zero_fsb = imap.br_startoff + imap.br_blockcount;
554 ASSERT(start_zero_fsb <= (end_zero_fsb + 1));
559 * There are blocks we need to zero.
561 zero_off = XFS_FSB_TO_B(mp, start_zero_fsb);
562 zero_len = XFS_FSB_TO_B(mp, imap.br_blockcount);
564 if ((zero_off + zero_len) > offset)
565 zero_len = offset - zero_off;
567 error = xfs_iozero(ip, zero_off, zero_len);
572 start_zero_fsb = imap.br_startoff + imap.br_blockcount;
573 ASSERT(start_zero_fsb <= (end_zero_fsb + 1));
580 * Common pre-write limit and setup checks.
582 * Called with the iolocked held either shared and exclusive according to
583 * @iolock, and returns with it held. Might upgrade the iolock to exclusive
584 * if called for a direct write beyond i_size.
587 xfs_file_aio_write_checks(
589 struct iov_iter *from,
592 struct file *file = iocb->ki_filp;
593 struct inode *inode = file->f_mapping->host;
594 struct xfs_inode *ip = XFS_I(inode);
596 size_t count = iov_iter_count(from);
597 bool drained_dio = false;
600 error = generic_write_checks(iocb, from);
604 error = xfs_break_layouts(inode, iolock, true);
608 /* For changing security info in file_remove_privs() we need i_mutex */
609 if (*iolock == XFS_IOLOCK_SHARED && !IS_NOSEC(inode)) {
610 xfs_rw_iunlock(ip, *iolock);
611 *iolock = XFS_IOLOCK_EXCL;
612 xfs_rw_ilock(ip, *iolock);
616 * If the offset is beyond the size of the file, we need to zero any
617 * blocks that fall between the existing EOF and the start of this
618 * write. If zeroing is needed and we are currently holding the
619 * iolock shared, we need to update it to exclusive which implies
620 * having to redo all checks before.
622 * We need to serialise against EOF updates that occur in IO
623 * completions here. We want to make sure that nobody is changing the
624 * size while we do this check until we have placed an IO barrier (i.e.
625 * hold the XFS_IOLOCK_EXCL) that prevents new IO from being dispatched.
626 * The spinlock effectively forms a memory barrier once we have the
627 * XFS_IOLOCK_EXCL so we are guaranteed to see the latest EOF value
628 * and hence be able to correctly determine if we need to run zeroing.
630 spin_lock(&ip->i_flags_lock);
631 if (iocb->ki_pos > i_size_read(inode)) {
634 spin_unlock(&ip->i_flags_lock);
636 if (*iolock == XFS_IOLOCK_SHARED) {
637 xfs_rw_iunlock(ip, *iolock);
638 *iolock = XFS_IOLOCK_EXCL;
639 xfs_rw_ilock(ip, *iolock);
640 iov_iter_reexpand(from, count);
643 * We now have an IO submission barrier in place, but
644 * AIO can do EOF updates during IO completion and hence
645 * we now need to wait for all of them to drain. Non-AIO
646 * DIO will have drained before we are given the
647 * XFS_IOLOCK_EXCL, and so for most cases this wait is a
650 inode_dio_wait(inode);
654 error = xfs_zero_eof(ip, iocb->ki_pos, i_size_read(inode), &zero);
658 spin_unlock(&ip->i_flags_lock);
661 * Updating the timestamps will grab the ilock again from
662 * xfs_fs_dirty_inode, so we have to call it after dropping the
663 * lock above. Eventually we should look into a way to avoid
664 * the pointless lock roundtrip.
666 if (likely(!(file->f_mode & FMODE_NOCMTIME))) {
667 error = file_update_time(file);
673 * If we're writing the file then make sure to clear the setuid and
674 * setgid bits if the process is not being run by root. This keeps
675 * people from modifying setuid and setgid binaries.
677 if (!IS_NOSEC(inode))
678 return file_remove_privs(file);
683 * xfs_file_dio_aio_write - handle direct IO writes
685 * Lock the inode appropriately to prepare for and issue a direct IO write.
686 * By separating it from the buffered write path we remove all the tricky to
687 * follow locking changes and looping.
689 * If there are cached pages or we're extending the file, we need IOLOCK_EXCL
690 * until we're sure the bytes at the new EOF have been zeroed and/or the cached
691 * pages are flushed out.
693 * In most cases the direct IO writes will be done holding IOLOCK_SHARED
694 * allowing them to be done in parallel with reads and other direct IO writes.
695 * However, if the IO is not aligned to filesystem blocks, the direct IO layer
696 * needs to do sub-block zeroing and that requires serialisation against other
697 * direct IOs to the same block. In this case we need to serialise the
698 * submission of the unaligned IOs so that we don't get racing block zeroing in
699 * the dio layer. To avoid the problem with aio, we also need to wait for
700 * outstanding IOs to complete so that unwritten extent conversion is completed
701 * before we try to map the overlapping block. This is currently implemented by
702 * hitting it with a big hammer (i.e. inode_dio_wait()).
704 * Returns with locks held indicated by @iolock and errors indicated by
705 * negative return values.
708 xfs_file_dio_aio_write(
710 struct iov_iter *from)
712 struct file *file = iocb->ki_filp;
713 struct address_space *mapping = file->f_mapping;
714 struct inode *inode = mapping->host;
715 struct xfs_inode *ip = XFS_I(inode);
716 struct xfs_mount *mp = ip->i_mount;
718 int unaligned_io = 0;
720 size_t count = iov_iter_count(from);
722 struct iov_iter data;
723 struct xfs_buftarg *target = XFS_IS_REALTIME_INODE(ip) ?
724 mp->m_rtdev_targp : mp->m_ddev_targp;
726 /* DIO must be aligned to device logical sector size */
727 if (!IS_DAX(inode) &&
728 ((iocb->ki_pos | count) & target->bt_logical_sectormask))
731 /* "unaligned" here means not aligned to a filesystem block */
732 if ((iocb->ki_pos & mp->m_blockmask) ||
733 ((iocb->ki_pos + count) & mp->m_blockmask))
737 * We don't need to take an exclusive lock unless there page cache needs
738 * to be invalidated or unaligned IO is being executed. We don't need to
739 * consider the EOF extension case here because
740 * xfs_file_aio_write_checks() will relock the inode as necessary for
741 * EOF zeroing cases and fill out the new inode size as appropriate.
743 if (unaligned_io || mapping->nrpages)
744 iolock = XFS_IOLOCK_EXCL;
746 iolock = XFS_IOLOCK_SHARED;
747 xfs_rw_ilock(ip, iolock);
750 * Recheck if there are cached pages that need invalidate after we got
751 * the iolock to protect against other threads adding new pages while
752 * we were waiting for the iolock.
754 if (mapping->nrpages && iolock == XFS_IOLOCK_SHARED) {
755 xfs_rw_iunlock(ip, iolock);
756 iolock = XFS_IOLOCK_EXCL;
757 xfs_rw_ilock(ip, iolock);
760 ret = xfs_file_aio_write_checks(iocb, from, &iolock);
763 count = iov_iter_count(from);
764 end = iocb->ki_pos + count - 1;
767 * See xfs_file_read_iter() for why we do a full-file flush here.
769 if (mapping->nrpages) {
770 ret = filemap_write_and_wait(VFS_I(ip)->i_mapping);
774 * Invalidate whole pages. This can return an error if we fail
775 * to invalidate a page, but this should never happen on XFS.
776 * Warn if it does fail.
778 ret = invalidate_inode_pages2(VFS_I(ip)->i_mapping);
784 * If we are doing unaligned IO, wait for all other IO to drain,
785 * otherwise demote the lock if we had to flush cached pages
788 inode_dio_wait(inode);
789 else if (iolock == XFS_IOLOCK_EXCL) {
790 xfs_rw_ilock_demote(ip, XFS_IOLOCK_EXCL);
791 iolock = XFS_IOLOCK_SHARED;
794 trace_xfs_file_direct_write(ip, count, iocb->ki_pos, 0);
797 ret = mapping->a_ops->direct_IO(iocb, &data);
799 /* see generic_file_direct_write() for why this is necessary */
800 if (mapping->nrpages) {
801 invalidate_inode_pages2_range(mapping,
802 iocb->ki_pos >> PAGE_SHIFT,
808 iov_iter_advance(from, ret);
811 xfs_rw_iunlock(ip, iolock);
814 * No fallback to buffered IO on errors for XFS. DAX can result in
815 * partial writes, but direct IO will either complete fully or fail.
817 ASSERT(ret < 0 || ret == count || IS_DAX(VFS_I(ip)));
822 xfs_file_buffered_aio_write(
824 struct iov_iter *from)
826 struct file *file = iocb->ki_filp;
827 struct address_space *mapping = file->f_mapping;
828 struct inode *inode = mapping->host;
829 struct xfs_inode *ip = XFS_I(inode);
832 int iolock = XFS_IOLOCK_EXCL;
834 xfs_rw_ilock(ip, iolock);
836 ret = xfs_file_aio_write_checks(iocb, from, &iolock);
840 /* We can write back this queue in page reclaim */
841 current->backing_dev_info = inode_to_bdi(inode);
844 trace_xfs_file_buffered_write(ip, iov_iter_count(from),
846 ret = generic_perform_write(file, from, iocb->ki_pos);
847 if (likely(ret >= 0))
851 * If we hit a space limit, try to free up some lingering preallocated
852 * space before returning an error. In the case of ENOSPC, first try to
853 * write back all dirty inodes to free up some of the excess reserved
854 * metadata space. This reduces the chances that the eofblocks scan
855 * waits on dirty mappings. Since xfs_flush_inodes() is serialized, this
856 * also behaves as a filter to prevent too many eofblocks scans from
857 * running at the same time.
859 if (ret == -EDQUOT && !enospc) {
860 enospc = xfs_inode_free_quota_eofblocks(ip);
863 } else if (ret == -ENOSPC && !enospc) {
864 struct xfs_eofblocks eofb = {0};
867 xfs_flush_inodes(ip->i_mount);
868 eofb.eof_scan_owner = ip->i_ino; /* for locking */
869 eofb.eof_flags = XFS_EOF_FLAGS_SYNC;
870 xfs_icache_free_eofblocks(ip->i_mount, &eofb);
874 current->backing_dev_info = NULL;
876 xfs_rw_iunlock(ip, iolock);
883 struct iov_iter *from)
885 struct file *file = iocb->ki_filp;
886 struct address_space *mapping = file->f_mapping;
887 struct inode *inode = mapping->host;
888 struct xfs_inode *ip = XFS_I(inode);
890 size_t ocount = iov_iter_count(from);
892 XFS_STATS_INC(ip->i_mount, xs_write_calls);
897 if (XFS_FORCED_SHUTDOWN(ip->i_mount))
900 if ((iocb->ki_flags & IOCB_DIRECT) || IS_DAX(inode))
901 ret = xfs_file_dio_aio_write(iocb, from);
903 ret = xfs_file_buffered_aio_write(iocb, from);
906 XFS_STATS_ADD(ip->i_mount, xs_write_bytes, ret);
908 /* Handle various SYNC-type writes */
909 ret = generic_write_sync(iocb, ret);
914 #define XFS_FALLOC_FL_SUPPORTED \
915 (FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE | \
916 FALLOC_FL_COLLAPSE_RANGE | FALLOC_FL_ZERO_RANGE | \
917 FALLOC_FL_INSERT_RANGE)
926 struct inode *inode = file_inode(file);
927 struct xfs_inode *ip = XFS_I(inode);
929 enum xfs_prealloc_flags flags = 0;
930 uint iolock = XFS_IOLOCK_EXCL;
932 bool do_file_insert = 0;
934 if (!S_ISREG(inode->i_mode))
936 if (mode & ~XFS_FALLOC_FL_SUPPORTED)
939 xfs_ilock(ip, iolock);
940 error = xfs_break_layouts(inode, &iolock, false);
944 xfs_ilock(ip, XFS_MMAPLOCK_EXCL);
945 iolock |= XFS_MMAPLOCK_EXCL;
947 if (mode & FALLOC_FL_PUNCH_HOLE) {
948 error = xfs_free_file_space(ip, offset, len);
951 } else if (mode & FALLOC_FL_COLLAPSE_RANGE) {
952 unsigned blksize_mask = (1 << inode->i_blkbits) - 1;
954 if (offset & blksize_mask || len & blksize_mask) {
960 * There is no need to overlap collapse range with EOF,
961 * in which case it is effectively a truncate operation
963 if (offset + len >= i_size_read(inode)) {
968 new_size = i_size_read(inode) - len;
970 error = xfs_collapse_file_space(ip, offset, len);
973 } else if (mode & FALLOC_FL_INSERT_RANGE) {
974 unsigned blksize_mask = (1 << inode->i_blkbits) - 1;
976 new_size = i_size_read(inode) + len;
977 if (offset & blksize_mask || len & blksize_mask) {
982 /* check the new inode size does not wrap through zero */
983 if (new_size > inode->i_sb->s_maxbytes) {
988 /* Offset should be less than i_size */
989 if (offset >= i_size_read(inode)) {
995 flags |= XFS_PREALLOC_SET;
997 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
998 offset + len > i_size_read(inode)) {
999 new_size = offset + len;
1000 error = inode_newsize_ok(inode, new_size);
1005 if (mode & FALLOC_FL_ZERO_RANGE)
1006 error = xfs_zero_file_space(ip, offset, len);
1008 error = xfs_alloc_file_space(ip, offset, len,
1009 XFS_BMAPI_PREALLOC);
1014 if (file->f_flags & O_DSYNC)
1015 flags |= XFS_PREALLOC_SYNC;
1017 error = xfs_update_prealloc_flags(ip, flags);
1021 /* Change file size if needed */
1025 iattr.ia_valid = ATTR_SIZE;
1026 iattr.ia_size = new_size;
1027 error = xfs_setattr_size(ip, &iattr);
1033 * Perform hole insertion now that the file size has been
1034 * updated so that if we crash during the operation we don't
1035 * leave shifted extents past EOF and hence losing access to
1036 * the data that is contained within them.
1039 error = xfs_insert_file_space(ip, offset, len);
1042 xfs_iunlock(ip, iolock);
1049 struct inode *inode,
1052 if (!(file->f_flags & O_LARGEFILE) && i_size_read(inode) > MAX_NON_LFS)
1054 if (XFS_FORCED_SHUTDOWN(XFS_M(inode->i_sb)))
1061 struct inode *inode,
1064 struct xfs_inode *ip = XFS_I(inode);
1068 error = xfs_file_open(inode, file);
1073 * If there are any blocks, read-ahead block 0 as we're almost
1074 * certain to have the next operation be a read there.
1076 mode = xfs_ilock_data_map_shared(ip);
1077 if (ip->i_d.di_nextents > 0)
1078 xfs_dir3_data_readahead(ip, 0, -1);
1079 xfs_iunlock(ip, mode);
1085 struct inode *inode,
1088 return xfs_release(XFS_I(inode));
1094 struct dir_context *ctx)
1096 struct inode *inode = file_inode(file);
1097 xfs_inode_t *ip = XFS_I(inode);
1101 * The Linux API doesn't pass down the total size of the buffer
1102 * we read into down to the filesystem. With the filldir concept
1103 * it's not needed for correct information, but the XFS dir2 leaf
1104 * code wants an estimate of the buffer size to calculate it's
1105 * readahead window and size the buffers used for mapping to
1108 * Try to give it an estimate that's good enough, maybe at some
1109 * point we can change the ->readdir prototype to include the
1110 * buffer size. For now we use the current glibc buffer size.
1112 bufsize = (size_t)min_t(loff_t, 32768, ip->i_d.di_size);
1114 return xfs_readdir(ip, ctx, bufsize);
1118 * This type is designed to indicate the type of offset we would like
1119 * to search from page cache for xfs_seek_hole_data().
1127 * Lookup the desired type of offset from the given page.
1129 * On success, return true and the offset argument will point to the
1130 * start of the region that was found. Otherwise this function will
1131 * return false and keep the offset argument unchanged.
1134 xfs_lookup_buffer_offset(
1139 loff_t lastoff = page_offset(page);
1141 struct buffer_head *bh, *head;
1143 bh = head = page_buffers(page);
1146 * Unwritten extents that have data in the page
1147 * cache covering them can be identified by the
1148 * BH_Unwritten state flag. Pages with multiple
1149 * buffers might have a mix of holes, data and
1150 * unwritten extents - any buffer with valid
1151 * data in it should have BH_Uptodate flag set
1154 if (buffer_unwritten(bh) ||
1155 buffer_uptodate(bh)) {
1156 if (type == DATA_OFF)
1159 if (type == HOLE_OFF)
1167 lastoff += bh->b_size;
1168 } while ((bh = bh->b_this_page) != head);
1174 * This routine is called to find out and return a data or hole offset
1175 * from the page cache for unwritten extents according to the desired
1176 * type for xfs_seek_hole_data().
1178 * The argument offset is used to tell where we start to search from the
1179 * page cache. Map is used to figure out the end points of the range to
1182 * Return true if the desired type of offset was found, and the argument
1183 * offset is filled with that address. Otherwise, return false and keep
1187 xfs_find_get_desired_pgoff(
1188 struct inode *inode,
1189 struct xfs_bmbt_irec *map,
1193 struct xfs_inode *ip = XFS_I(inode);
1194 struct xfs_mount *mp = ip->i_mount;
1195 struct pagevec pvec;
1199 loff_t startoff = *offset;
1200 loff_t lastoff = startoff;
1203 pagevec_init(&pvec, 0);
1205 index = startoff >> PAGE_SHIFT;
1206 endoff = XFS_FSB_TO_B(mp, map->br_startoff + map->br_blockcount);
1207 end = endoff >> PAGE_SHIFT;
1213 want = min_t(pgoff_t, end - index, PAGEVEC_SIZE);
1214 nr_pages = pagevec_lookup(&pvec, inode->i_mapping, index,
1217 * No page mapped into given range. If we are searching holes
1218 * and if this is the first time we got into the loop, it means
1219 * that the given offset is landed in a hole, return it.
1221 * If we have already stepped through some block buffers to find
1222 * holes but they all contains data. In this case, the last
1223 * offset is already updated and pointed to the end of the last
1224 * mapped page, if it does not reach the endpoint to search,
1225 * that means there should be a hole between them.
1227 if (nr_pages == 0) {
1228 /* Data search found nothing */
1229 if (type == DATA_OFF)
1232 ASSERT(type == HOLE_OFF);
1233 if (lastoff == startoff || lastoff < endoff) {
1241 * At lease we found one page. If this is the first time we
1242 * step into the loop, and if the first page index offset is
1243 * greater than the given search offset, a hole was found.
1245 if (type == HOLE_OFF && lastoff == startoff &&
1246 lastoff < page_offset(pvec.pages[0])) {
1251 for (i = 0; i < nr_pages; i++) {
1252 struct page *page = pvec.pages[i];
1256 * At this point, the page may be truncated or
1257 * invalidated (changing page->mapping to NULL),
1258 * or even swizzled back from swapper_space to tmpfs
1259 * file mapping. However, page->index will not change
1260 * because we have a reference on the page.
1262 * Searching done if the page index is out of range.
1263 * If the current offset is not reaches the end of
1264 * the specified search range, there should be a hole
1267 if (page->index > end) {
1268 if (type == HOLE_OFF && lastoff < endoff) {
1277 * Page truncated or invalidated(page->mapping == NULL).
1278 * We can freely skip it and proceed to check the next
1281 if (unlikely(page->mapping != inode->i_mapping)) {
1286 if (!page_has_buffers(page)) {
1291 found = xfs_lookup_buffer_offset(page, &b_offset, type);
1294 * The found offset may be less than the start
1295 * point to search if this is the first time to
1298 *offset = max_t(loff_t, startoff, b_offset);
1304 * We either searching data but nothing was found, or
1305 * searching hole but found a data buffer. In either
1306 * case, probably the next page contains the desired
1307 * things, update the last offset to it so.
1309 lastoff = page_offset(page) + PAGE_SIZE;
1314 * The number of returned pages less than our desired, search
1315 * done. In this case, nothing was found for searching data,
1316 * but we found a hole behind the last offset.
1318 if (nr_pages < want) {
1319 if (type == HOLE_OFF) {
1326 index = pvec.pages[i - 1]->index + 1;
1327 pagevec_release(&pvec);
1328 } while (index <= end);
1331 pagevec_release(&pvec);
1336 * caller must lock inode with xfs_ilock_data_map_shared,
1337 * can we craft an appropriate ASSERT?
1339 * end is because the VFS-level lseek interface is defined such that any
1340 * offset past i_size shall return -ENXIO, but we use this for quota code
1341 * which does not maintain i_size, and we want to SEEK_DATA past i_size.
1344 __xfs_seek_hole_data(
1345 struct inode *inode,
1350 struct xfs_inode *ip = XFS_I(inode);
1351 struct xfs_mount *mp = ip->i_mount;
1352 loff_t uninitialized_var(offset);
1353 xfs_fileoff_t fsbno;
1354 xfs_filblks_t lastbno;
1363 * Try to read extents from the first block indicated
1364 * by fsbno to the end block of the file.
1366 fsbno = XFS_B_TO_FSBT(mp, start);
1367 lastbno = XFS_B_TO_FSB(mp, end);
1370 struct xfs_bmbt_irec map[2];
1374 error = xfs_bmapi_read(ip, fsbno, lastbno - fsbno, map, &nmap,
1379 /* No extents at given offset, must be beyond EOF */
1385 for (i = 0; i < nmap; i++) {
1386 offset = max_t(loff_t, start,
1387 XFS_FSB_TO_B(mp, map[i].br_startoff));
1389 /* Landed in the hole we wanted? */
1390 if (whence == SEEK_HOLE &&
1391 map[i].br_startblock == HOLESTARTBLOCK)
1394 /* Landed in the data extent we wanted? */
1395 if (whence == SEEK_DATA &&
1396 (map[i].br_startblock == DELAYSTARTBLOCK ||
1397 (map[i].br_state == XFS_EXT_NORM &&
1398 !isnullstartblock(map[i].br_startblock))))
1402 * Landed in an unwritten extent, try to search
1403 * for hole or data from page cache.
1405 if (map[i].br_state == XFS_EXT_UNWRITTEN) {
1406 if (xfs_find_get_desired_pgoff(inode, &map[i],
1407 whence == SEEK_HOLE ? HOLE_OFF : DATA_OFF,
1414 * We only received one extent out of the two requested. This
1415 * means we've hit EOF and didn't find what we are looking for.
1419 * If we were looking for a hole, set offset to
1420 * the end of the file (i.e., there is an implicit
1421 * hole at the end of any file).
1423 if (whence == SEEK_HOLE) {
1428 * If we were looking for data, it's nowhere to be found
1430 ASSERT(whence == SEEK_DATA);
1438 * Nothing was found, proceed to the next round of search
1439 * if the next reading offset is not at or beyond EOF.
1441 fsbno = map[i - 1].br_startoff + map[i - 1].br_blockcount;
1442 start = XFS_FSB_TO_B(mp, fsbno);
1444 if (whence == SEEK_HOLE) {
1448 ASSERT(whence == SEEK_DATA);
1456 * If at this point we have found the hole we wanted, the returned
1457 * offset may be bigger than the file size as it may be aligned to
1458 * page boundary for unwritten extents. We need to deal with this
1459 * situation in particular.
1461 if (whence == SEEK_HOLE)
1462 offset = min_t(loff_t, offset, end);
1476 struct inode *inode = file->f_mapping->host;
1477 struct xfs_inode *ip = XFS_I(inode);
1478 struct xfs_mount *mp = ip->i_mount;
1483 if (XFS_FORCED_SHUTDOWN(mp))
1486 lock = xfs_ilock_data_map_shared(ip);
1488 end = i_size_read(inode);
1489 offset = __xfs_seek_hole_data(inode, start, end, whence);
1495 offset = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
1498 xfs_iunlock(ip, lock);
1515 return generic_file_llseek(file, offset, whence);
1518 return xfs_seek_hole_data(file, offset, whence);
1525 * Locking for serialisation of IO during page faults. This results in a lock
1529 * sb_start_pagefault(vfs, freeze)
1530 * i_mmaplock (XFS - truncate serialisation)
1532 * i_lock (XFS - extent map serialisation)
1536 * mmap()d file has taken write protection fault and is being made writable. We
1537 * can set the page state up correctly for a writable page, which means we can
1538 * do correct delalloc accounting (ENOSPC checking!) and unwritten extent
1542 xfs_filemap_page_mkwrite(
1543 struct vm_area_struct *vma,
1544 struct vm_fault *vmf)
1546 struct inode *inode = file_inode(vma->vm_file);
1549 trace_xfs_filemap_page_mkwrite(XFS_I(inode));
1551 sb_start_pagefault(inode->i_sb);
1552 file_update_time(vma->vm_file);
1553 xfs_ilock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1555 if (IS_DAX(inode)) {
1556 ret = __dax_mkwrite(vma, vmf, xfs_get_blocks_dax_fault, NULL);
1558 ret = block_page_mkwrite(vma, vmf, xfs_get_blocks);
1559 ret = block_page_mkwrite_return(ret);
1562 xfs_iunlock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1563 sb_end_pagefault(inode->i_sb);
1570 struct vm_area_struct *vma,
1571 struct vm_fault *vmf)
1573 struct inode *inode = file_inode(vma->vm_file);
1576 trace_xfs_filemap_fault(XFS_I(inode));
1578 /* DAX can shortcut the normal fault path on write faults! */
1579 if ((vmf->flags & FAULT_FLAG_WRITE) && IS_DAX(inode))
1580 return xfs_filemap_page_mkwrite(vma, vmf);
1582 xfs_ilock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1583 if (IS_DAX(inode)) {
1585 * we do not want to trigger unwritten extent conversion on read
1586 * faults - that is unnecessary overhead and would also require
1587 * changes to xfs_get_blocks_direct() to map unwritten extent
1588 * ioend for conversion on read-only mappings.
1590 ret = __dax_fault(vma, vmf, xfs_get_blocks_dax_fault, NULL);
1592 ret = filemap_fault(vma, vmf);
1593 xfs_iunlock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1599 * Similar to xfs_filemap_fault(), the DAX fault path can call into here on
1600 * both read and write faults. Hence we need to handle both cases. There is no
1601 * ->pmd_mkwrite callout for huge pages, so we have a single function here to
1602 * handle both cases here. @flags carries the information on the type of fault
1606 xfs_filemap_pmd_fault(
1607 struct vm_area_struct *vma,
1612 struct inode *inode = file_inode(vma->vm_file);
1613 struct xfs_inode *ip = XFS_I(inode);
1617 return VM_FAULT_FALLBACK;
1619 trace_xfs_filemap_pmd_fault(ip);
1621 if (flags & FAULT_FLAG_WRITE) {
1622 sb_start_pagefault(inode->i_sb);
1623 file_update_time(vma->vm_file);
1626 xfs_ilock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1627 ret = __dax_pmd_fault(vma, addr, pmd, flags, xfs_get_blocks_dax_fault,
1629 xfs_iunlock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1631 if (flags & FAULT_FLAG_WRITE)
1632 sb_end_pagefault(inode->i_sb);
1638 * pfn_mkwrite was originally inteneded to ensure we capture time stamp
1639 * updates on write faults. In reality, it's need to serialise against
1640 * truncate similar to page_mkwrite. Hence we cycle the XFS_MMAPLOCK_SHARED
1641 * to ensure we serialise the fault barrier in place.
1644 xfs_filemap_pfn_mkwrite(
1645 struct vm_area_struct *vma,
1646 struct vm_fault *vmf)
1649 struct inode *inode = file_inode(vma->vm_file);
1650 struct xfs_inode *ip = XFS_I(inode);
1651 int ret = VM_FAULT_NOPAGE;
1654 trace_xfs_filemap_pfn_mkwrite(ip);
1656 sb_start_pagefault(inode->i_sb);
1657 file_update_time(vma->vm_file);
1659 /* check if the faulting page hasn't raced with truncate */
1660 xfs_ilock(ip, XFS_MMAPLOCK_SHARED);
1661 size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
1662 if (vmf->pgoff >= size)
1663 ret = VM_FAULT_SIGBUS;
1664 else if (IS_DAX(inode))
1665 ret = dax_pfn_mkwrite(vma, vmf);
1666 xfs_iunlock(ip, XFS_MMAPLOCK_SHARED);
1667 sb_end_pagefault(inode->i_sb);
1672 static const struct vm_operations_struct xfs_file_vm_ops = {
1673 .fault = xfs_filemap_fault,
1674 .pmd_fault = xfs_filemap_pmd_fault,
1675 .map_pages = filemap_map_pages,
1676 .page_mkwrite = xfs_filemap_page_mkwrite,
1677 .pfn_mkwrite = xfs_filemap_pfn_mkwrite,
1683 struct vm_area_struct *vma)
1685 file_accessed(filp);
1686 vma->vm_ops = &xfs_file_vm_ops;
1687 if (IS_DAX(file_inode(filp)))
1688 vma->vm_flags |= VM_MIXEDMAP | VM_HUGEPAGE;
1692 const struct file_operations xfs_file_operations = {
1693 .llseek = xfs_file_llseek,
1694 .read_iter = xfs_file_read_iter,
1695 .write_iter = xfs_file_write_iter,
1696 .splice_read = xfs_file_splice_read,
1697 .splice_write = iter_file_splice_write,
1698 .unlocked_ioctl = xfs_file_ioctl,
1699 #ifdef CONFIG_COMPAT
1700 .compat_ioctl = xfs_file_compat_ioctl,
1702 .mmap = xfs_file_mmap,
1703 .open = xfs_file_open,
1704 .release = xfs_file_release,
1705 .fsync = xfs_file_fsync,
1706 .fallocate = xfs_file_fallocate,
1709 const struct file_operations xfs_dir_file_operations = {
1710 .open = xfs_dir_open,
1711 .read = generic_read_dir,
1712 .iterate_shared = xfs_file_readdir,
1713 .llseek = generic_file_llseek,
1714 .unlocked_ioctl = xfs_file_ioctl,
1715 #ifdef CONFIG_COMPAT
1716 .compat_ioctl = xfs_file_compat_ioctl,
1718 .fsync = xfs_dir_fsync,