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 error = xfs_trans_alloc(ip->i_mount, &M_RES(ip->i_mount)->tr_writeid,
153 xfs_ilock(ip, XFS_ILOCK_EXCL);
154 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
156 if (!(flags & XFS_PREALLOC_INVISIBLE)) {
157 VFS_I(ip)->i_mode &= ~S_ISUID;
158 if (VFS_I(ip)->i_mode & S_IXGRP)
159 VFS_I(ip)->i_mode &= ~S_ISGID;
160 xfs_trans_ichgtime(tp, ip, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
163 if (flags & XFS_PREALLOC_SET)
164 ip->i_d.di_flags |= XFS_DIFLAG_PREALLOC;
165 if (flags & XFS_PREALLOC_CLEAR)
166 ip->i_d.di_flags &= ~XFS_DIFLAG_PREALLOC;
168 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
169 if (flags & XFS_PREALLOC_SYNC)
170 xfs_trans_set_sync(tp);
171 return xfs_trans_commit(tp);
175 * Fsync operations on directories are much simpler than on regular files,
176 * as there is no file data to flush, and thus also no need for explicit
177 * cache flush operations, and there are no non-transaction metadata updates
178 * on directories either.
187 struct xfs_inode *ip = XFS_I(file->f_mapping->host);
188 struct xfs_mount *mp = ip->i_mount;
191 trace_xfs_dir_fsync(ip);
193 xfs_ilock(ip, XFS_ILOCK_SHARED);
194 if (xfs_ipincount(ip))
195 lsn = ip->i_itemp->ili_last_lsn;
196 xfs_iunlock(ip, XFS_ILOCK_SHARED);
200 return _xfs_log_force_lsn(mp, lsn, XFS_LOG_SYNC, NULL);
210 struct inode *inode = file->f_mapping->host;
211 struct xfs_inode *ip = XFS_I(inode);
212 struct xfs_mount *mp = ip->i_mount;
217 trace_xfs_file_fsync(ip);
219 error = filemap_write_and_wait_range(inode->i_mapping, start, end);
223 if (XFS_FORCED_SHUTDOWN(mp))
226 xfs_iflags_clear(ip, XFS_ITRUNCATED);
228 if (mp->m_flags & XFS_MOUNT_BARRIER) {
230 * If we have an RT and/or log subvolume we need to make sure
231 * to flush the write cache the device used for file data
232 * first. This is to ensure newly written file data make
233 * it to disk before logging the new inode size in case of
234 * an extending write.
236 if (XFS_IS_REALTIME_INODE(ip))
237 xfs_blkdev_issue_flush(mp->m_rtdev_targp);
238 else if (mp->m_logdev_targp != mp->m_ddev_targp)
239 xfs_blkdev_issue_flush(mp->m_ddev_targp);
243 * All metadata updates are logged, which means that we just have to
244 * flush the log up to the latest LSN that touched the inode. If we have
245 * concurrent fsync/fdatasync() calls, we need them to all block on the
246 * log force before we clear the ili_fsync_fields field. This ensures
247 * that we don't get a racing sync operation that does not wait for the
248 * metadata to hit the journal before returning. If we race with
249 * clearing the ili_fsync_fields, then all that will happen is the log
250 * force will do nothing as the lsn will already be on disk. We can't
251 * race with setting ili_fsync_fields because that is done under
252 * XFS_ILOCK_EXCL, and that can't happen because we hold the lock shared
253 * until after the ili_fsync_fields is cleared.
255 xfs_ilock(ip, XFS_ILOCK_SHARED);
256 if (xfs_ipincount(ip)) {
258 (ip->i_itemp->ili_fsync_fields & ~XFS_ILOG_TIMESTAMP))
259 lsn = ip->i_itemp->ili_last_lsn;
263 error = _xfs_log_force_lsn(mp, lsn, XFS_LOG_SYNC, &log_flushed);
264 ip->i_itemp->ili_fsync_fields = 0;
266 xfs_iunlock(ip, XFS_ILOCK_SHARED);
269 * If we only have a single device, and the log force about was
270 * a no-op we might have to flush the data device cache here.
271 * This can only happen for fdatasync/O_DSYNC if we were overwriting
272 * an already allocated file and thus do not have any metadata to
275 if ((mp->m_flags & XFS_MOUNT_BARRIER) &&
276 mp->m_logdev_targp == mp->m_ddev_targp &&
277 !XFS_IS_REALTIME_INODE(ip) &&
279 xfs_blkdev_issue_flush(mp->m_ddev_targp);
289 struct file *file = iocb->ki_filp;
290 struct inode *inode = file->f_mapping->host;
291 struct xfs_inode *ip = XFS_I(inode);
292 struct xfs_mount *mp = ip->i_mount;
293 size_t size = iov_iter_count(to);
296 loff_t pos = iocb->ki_pos;
298 XFS_STATS_INC(mp, xs_read_calls);
300 if ((iocb->ki_flags & IOCB_DIRECT) && !IS_DAX(inode)) {
301 xfs_buftarg_t *target =
302 XFS_IS_REALTIME_INODE(ip) ?
303 mp->m_rtdev_targp : mp->m_ddev_targp;
304 /* DIO must be aligned to device logical sector size */
305 if ((pos | size) & target->bt_logical_sectormask) {
306 if (pos == i_size_read(inode))
312 n = mp->m_super->s_maxbytes - pos;
313 if (n <= 0 || size == 0)
319 if (XFS_FORCED_SHUTDOWN(mp))
323 * Locking is a bit tricky here. If we take an exclusive lock for direct
324 * IO, we effectively serialise all new concurrent read IO to this file
325 * and block it behind IO that is currently in progress because IO in
326 * progress holds the IO lock shared. We only need to hold the lock
327 * exclusive to blow away the page cache, so only take lock exclusively
328 * if the page cache needs invalidation. This allows the normal direct
329 * IO case of no page cache pages to proceeed concurrently without
332 xfs_rw_ilock(ip, XFS_IOLOCK_SHARED);
333 if ((iocb->ki_flags & IOCB_DIRECT) && inode->i_mapping->nrpages) {
334 xfs_rw_iunlock(ip, XFS_IOLOCK_SHARED);
335 xfs_rw_ilock(ip, XFS_IOLOCK_EXCL);
338 * The generic dio code only flushes the range of the particular
339 * I/O. Because we take an exclusive lock here, this whole
340 * sequence is considerably more expensive for us. This has a
341 * noticeable performance impact for any file with cached pages,
342 * even when outside of the range of the particular I/O.
344 * Hence, amortize the cost of the lock against a full file
345 * flush and reduce the chances of repeated iolock cycles going
348 if (inode->i_mapping->nrpages) {
349 ret = filemap_write_and_wait(VFS_I(ip)->i_mapping);
351 xfs_rw_iunlock(ip, XFS_IOLOCK_EXCL);
356 * Invalidate whole pages. This can return an error if
357 * we fail to invalidate a page, but this should never
358 * happen on XFS. Warn if it does fail.
360 ret = invalidate_inode_pages2(VFS_I(ip)->i_mapping);
364 xfs_rw_ilock_demote(ip, XFS_IOLOCK_EXCL);
367 if (iocb->ki_flags & IOCB_DIRECT)
368 trace_xfs_file_direct_read(ip, size, pos);
370 trace_xfs_file_buffered_read(ip, size, pos);
372 ret = generic_file_read_iter(iocb, to);
374 XFS_STATS_ADD(mp, xs_read_bytes, ret);
376 xfs_rw_iunlock(ip, XFS_IOLOCK_SHARED);
381 xfs_file_splice_read(
384 struct pipe_inode_info *pipe,
388 struct xfs_inode *ip = XFS_I(infilp->f_mapping->host);
391 XFS_STATS_INC(ip->i_mount, xs_read_calls);
393 if (XFS_FORCED_SHUTDOWN(ip->i_mount))
396 trace_xfs_file_splice_read(ip, count, *ppos);
399 * DAX inodes cannot ues the page cache for splice, so we have to push
400 * them through the VFS IO path. This means it goes through
401 * ->read_iter, which for us takes the XFS_IOLOCK_SHARED. Hence we
402 * cannot lock the splice operation at this level for DAX inodes.
404 if (IS_DAX(VFS_I(ip))) {
405 ret = default_file_splice_read(infilp, ppos, pipe, count,
410 xfs_rw_ilock(ip, XFS_IOLOCK_SHARED);
411 ret = generic_file_splice_read(infilp, ppos, pipe, count, flags);
412 xfs_rw_iunlock(ip, XFS_IOLOCK_SHARED);
415 XFS_STATS_ADD(ip->i_mount, xs_read_bytes, ret);
420 * This routine is called to handle zeroing any space in the last block of the
421 * file that is beyond the EOF. We do this since the size is being increased
422 * without writing anything to that block and we don't want to read the
423 * garbage on the disk.
425 STATIC int /* error (positive) */
427 struct xfs_inode *ip,
432 struct xfs_mount *mp = ip->i_mount;
433 xfs_fileoff_t last_fsb = XFS_B_TO_FSBT(mp, isize);
434 int zero_offset = XFS_B_FSB_OFFSET(mp, isize);
438 struct xfs_bmbt_irec imap;
440 xfs_ilock(ip, XFS_ILOCK_EXCL);
441 error = xfs_bmapi_read(ip, last_fsb, 1, &imap, &nimaps, 0);
442 xfs_iunlock(ip, XFS_ILOCK_EXCL);
449 * If the block underlying isize is just a hole, then there
450 * is nothing to zero.
452 if (imap.br_startblock == HOLESTARTBLOCK)
455 zero_len = mp->m_sb.sb_blocksize - zero_offset;
456 if (isize + zero_len > offset)
457 zero_len = offset - isize;
459 return xfs_iozero(ip, isize, zero_len);
463 * Zero any on disk space between the current EOF and the new, larger EOF.
465 * This handles the normal case of zeroing the remainder of the last block in
466 * the file and the unusual case of zeroing blocks out beyond the size of the
467 * file. This second case only happens with fixed size extents and when the
468 * system crashes before the inode size was updated but after blocks were
471 * Expects the iolock to be held exclusive, and will take the ilock internally.
473 int /* error (positive) */
475 struct xfs_inode *ip,
476 xfs_off_t offset, /* starting I/O offset */
477 xfs_fsize_t isize, /* current inode size */
480 struct xfs_mount *mp = ip->i_mount;
481 xfs_fileoff_t start_zero_fsb;
482 xfs_fileoff_t end_zero_fsb;
483 xfs_fileoff_t zero_count_fsb;
484 xfs_fileoff_t last_fsb;
485 xfs_fileoff_t zero_off;
486 xfs_fsize_t zero_len;
489 struct xfs_bmbt_irec imap;
491 ASSERT(xfs_isilocked(ip, XFS_IOLOCK_EXCL));
492 ASSERT(offset > isize);
494 trace_xfs_zero_eof(ip, isize, offset - isize);
497 * First handle zeroing the block on which isize resides.
499 * We only zero a part of that block so it is handled specially.
501 if (XFS_B_FSB_OFFSET(mp, isize) != 0) {
502 error = xfs_zero_last_block(ip, offset, isize, did_zeroing);
508 * Calculate the range between the new size and the old where blocks
509 * needing to be zeroed may exist.
511 * To get the block where the last byte in the file currently resides,
512 * we need to subtract one from the size and truncate back to a block
513 * boundary. We subtract 1 in case the size is exactly on a block
516 last_fsb = isize ? XFS_B_TO_FSBT(mp, isize - 1) : (xfs_fileoff_t)-1;
517 start_zero_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)isize);
518 end_zero_fsb = XFS_B_TO_FSBT(mp, offset - 1);
519 ASSERT((xfs_sfiloff_t)last_fsb < (xfs_sfiloff_t)start_zero_fsb);
520 if (last_fsb == end_zero_fsb) {
522 * The size was only incremented on its last block.
523 * We took care of that above, so just return.
528 ASSERT(start_zero_fsb <= end_zero_fsb);
529 while (start_zero_fsb <= end_zero_fsb) {
531 zero_count_fsb = end_zero_fsb - start_zero_fsb + 1;
533 xfs_ilock(ip, XFS_ILOCK_EXCL);
534 error = xfs_bmapi_read(ip, start_zero_fsb, zero_count_fsb,
536 xfs_iunlock(ip, XFS_ILOCK_EXCL);
542 if (imap.br_state == XFS_EXT_UNWRITTEN ||
543 imap.br_startblock == HOLESTARTBLOCK) {
544 start_zero_fsb = imap.br_startoff + imap.br_blockcount;
545 ASSERT(start_zero_fsb <= (end_zero_fsb + 1));
550 * There are blocks we need to zero.
552 zero_off = XFS_FSB_TO_B(mp, start_zero_fsb);
553 zero_len = XFS_FSB_TO_B(mp, imap.br_blockcount);
555 if ((zero_off + zero_len) > offset)
556 zero_len = offset - zero_off;
558 error = xfs_iozero(ip, zero_off, zero_len);
563 start_zero_fsb = imap.br_startoff + imap.br_blockcount;
564 ASSERT(start_zero_fsb <= (end_zero_fsb + 1));
571 * Common pre-write limit and setup checks.
573 * Called with the iolocked held either shared and exclusive according to
574 * @iolock, and returns with it held. Might upgrade the iolock to exclusive
575 * if called for a direct write beyond i_size.
578 xfs_file_aio_write_checks(
580 struct iov_iter *from,
583 struct file *file = iocb->ki_filp;
584 struct inode *inode = file->f_mapping->host;
585 struct xfs_inode *ip = XFS_I(inode);
587 size_t count = iov_iter_count(from);
588 bool drained_dio = false;
591 error = generic_write_checks(iocb, from);
595 error = xfs_break_layouts(inode, iolock, true);
599 /* For changing security info in file_remove_privs() we need i_mutex */
600 if (*iolock == XFS_IOLOCK_SHARED && !IS_NOSEC(inode)) {
601 xfs_rw_iunlock(ip, *iolock);
602 *iolock = XFS_IOLOCK_EXCL;
603 xfs_rw_ilock(ip, *iolock);
607 * If the offset is beyond the size of the file, we need to zero any
608 * blocks that fall between the existing EOF and the start of this
609 * write. If zeroing is needed and we are currently holding the
610 * iolock shared, we need to update it to exclusive which implies
611 * having to redo all checks before.
613 * We need to serialise against EOF updates that occur in IO
614 * completions here. We want to make sure that nobody is changing the
615 * size while we do this check until we have placed an IO barrier (i.e.
616 * hold the XFS_IOLOCK_EXCL) that prevents new IO from being dispatched.
617 * The spinlock effectively forms a memory barrier once we have the
618 * XFS_IOLOCK_EXCL so we are guaranteed to see the latest EOF value
619 * and hence be able to correctly determine if we need to run zeroing.
621 spin_lock(&ip->i_flags_lock);
622 if (iocb->ki_pos > i_size_read(inode)) {
625 spin_unlock(&ip->i_flags_lock);
627 if (*iolock == XFS_IOLOCK_SHARED) {
628 xfs_rw_iunlock(ip, *iolock);
629 *iolock = XFS_IOLOCK_EXCL;
630 xfs_rw_ilock(ip, *iolock);
631 iov_iter_reexpand(from, count);
634 * We now have an IO submission barrier in place, but
635 * AIO can do EOF updates during IO completion and hence
636 * we now need to wait for all of them to drain. Non-AIO
637 * DIO will have drained before we are given the
638 * XFS_IOLOCK_EXCL, and so for most cases this wait is a
641 inode_dio_wait(inode);
645 error = xfs_zero_eof(ip, iocb->ki_pos, i_size_read(inode), &zero);
649 spin_unlock(&ip->i_flags_lock);
652 * Updating the timestamps will grab the ilock again from
653 * xfs_fs_dirty_inode, so we have to call it after dropping the
654 * lock above. Eventually we should look into a way to avoid
655 * the pointless lock roundtrip.
657 if (likely(!(file->f_mode & FMODE_NOCMTIME))) {
658 error = file_update_time(file);
664 * If we're writing the file then make sure to clear the setuid and
665 * setgid bits if the process is not being run by root. This keeps
666 * people from modifying setuid and setgid binaries.
668 if (!IS_NOSEC(inode))
669 return file_remove_privs(file);
674 * xfs_file_dio_aio_write - handle direct IO writes
676 * Lock the inode appropriately to prepare for and issue a direct IO write.
677 * By separating it from the buffered write path we remove all the tricky to
678 * follow locking changes and looping.
680 * If there are cached pages or we're extending the file, we need IOLOCK_EXCL
681 * until we're sure the bytes at the new EOF have been zeroed and/or the cached
682 * pages are flushed out.
684 * In most cases the direct IO writes will be done holding IOLOCK_SHARED
685 * allowing them to be done in parallel with reads and other direct IO writes.
686 * However, if the IO is not aligned to filesystem blocks, the direct IO layer
687 * needs to do sub-block zeroing and that requires serialisation against other
688 * direct IOs to the same block. In this case we need to serialise the
689 * submission of the unaligned IOs so that we don't get racing block zeroing in
690 * the dio layer. To avoid the problem with aio, we also need to wait for
691 * outstanding IOs to complete so that unwritten extent conversion is completed
692 * before we try to map the overlapping block. This is currently implemented by
693 * hitting it with a big hammer (i.e. inode_dio_wait()).
695 * Returns with locks held indicated by @iolock and errors indicated by
696 * negative return values.
699 xfs_file_dio_aio_write(
701 struct iov_iter *from)
703 struct file *file = iocb->ki_filp;
704 struct address_space *mapping = file->f_mapping;
705 struct inode *inode = mapping->host;
706 struct xfs_inode *ip = XFS_I(inode);
707 struct xfs_mount *mp = ip->i_mount;
709 int unaligned_io = 0;
711 size_t count = iov_iter_count(from);
713 struct iov_iter data;
714 struct xfs_buftarg *target = XFS_IS_REALTIME_INODE(ip) ?
715 mp->m_rtdev_targp : mp->m_ddev_targp;
717 /* DIO must be aligned to device logical sector size */
718 if (!IS_DAX(inode) &&
719 ((iocb->ki_pos | count) & target->bt_logical_sectormask))
722 /* "unaligned" here means not aligned to a filesystem block */
723 if ((iocb->ki_pos & mp->m_blockmask) ||
724 ((iocb->ki_pos + count) & mp->m_blockmask))
728 * We don't need to take an exclusive lock unless there page cache needs
729 * to be invalidated or unaligned IO is being executed. We don't need to
730 * consider the EOF extension case here because
731 * xfs_file_aio_write_checks() will relock the inode as necessary for
732 * EOF zeroing cases and fill out the new inode size as appropriate.
734 if (unaligned_io || mapping->nrpages)
735 iolock = XFS_IOLOCK_EXCL;
737 iolock = XFS_IOLOCK_SHARED;
738 xfs_rw_ilock(ip, iolock);
741 * Recheck if there are cached pages that need invalidate after we got
742 * the iolock to protect against other threads adding new pages while
743 * we were waiting for the iolock.
745 if (mapping->nrpages && iolock == XFS_IOLOCK_SHARED) {
746 xfs_rw_iunlock(ip, iolock);
747 iolock = XFS_IOLOCK_EXCL;
748 xfs_rw_ilock(ip, iolock);
751 ret = xfs_file_aio_write_checks(iocb, from, &iolock);
754 count = iov_iter_count(from);
755 end = iocb->ki_pos + count - 1;
758 * See xfs_file_read_iter() for why we do a full-file flush here.
760 if (mapping->nrpages) {
761 ret = filemap_write_and_wait(VFS_I(ip)->i_mapping);
765 * Invalidate whole pages. This can return an error if we fail
766 * to invalidate a page, but this should never happen on XFS.
767 * Warn if it does fail.
769 ret = invalidate_inode_pages2(VFS_I(ip)->i_mapping);
775 * If we are doing unaligned IO, wait for all other IO to drain,
776 * otherwise demote the lock if we had to flush cached pages
779 inode_dio_wait(inode);
780 else if (iolock == XFS_IOLOCK_EXCL) {
781 xfs_rw_ilock_demote(ip, XFS_IOLOCK_EXCL);
782 iolock = XFS_IOLOCK_SHARED;
785 trace_xfs_file_direct_write(ip, count, iocb->ki_pos);
788 ret = mapping->a_ops->direct_IO(iocb, &data);
790 /* see generic_file_direct_write() for why this is necessary */
791 if (mapping->nrpages) {
792 invalidate_inode_pages2_range(mapping,
793 iocb->ki_pos >> PAGE_SHIFT,
799 iov_iter_advance(from, ret);
802 xfs_rw_iunlock(ip, iolock);
805 * No fallback to buffered IO on errors for XFS. DAX can result in
806 * partial writes, but direct IO will either complete fully or fail.
808 ASSERT(ret < 0 || ret == count || IS_DAX(VFS_I(ip)));
813 xfs_file_buffered_aio_write(
815 struct iov_iter *from)
817 struct file *file = iocb->ki_filp;
818 struct address_space *mapping = file->f_mapping;
819 struct inode *inode = mapping->host;
820 struct xfs_inode *ip = XFS_I(inode);
823 int iolock = XFS_IOLOCK_EXCL;
825 xfs_rw_ilock(ip, iolock);
827 ret = xfs_file_aio_write_checks(iocb, from, &iolock);
831 /* We can write back this queue in page reclaim */
832 current->backing_dev_info = inode_to_bdi(inode);
835 trace_xfs_file_buffered_write(ip, iov_iter_count(from), iocb->ki_pos);
836 ret = generic_perform_write(file, from, iocb->ki_pos);
837 if (likely(ret >= 0))
841 * If we hit a space limit, try to free up some lingering preallocated
842 * space before returning an error. In the case of ENOSPC, first try to
843 * write back all dirty inodes to free up some of the excess reserved
844 * metadata space. This reduces the chances that the eofblocks scan
845 * waits on dirty mappings. Since xfs_flush_inodes() is serialized, this
846 * also behaves as a filter to prevent too many eofblocks scans from
847 * running at the same time.
849 if (ret == -EDQUOT && !enospc) {
850 enospc = xfs_inode_free_quota_eofblocks(ip);
853 } else if (ret == -ENOSPC && !enospc) {
854 struct xfs_eofblocks eofb = {0};
857 xfs_flush_inodes(ip->i_mount);
858 eofb.eof_scan_owner = ip->i_ino; /* for locking */
859 eofb.eof_flags = XFS_EOF_FLAGS_SYNC;
860 xfs_icache_free_eofblocks(ip->i_mount, &eofb);
864 current->backing_dev_info = NULL;
866 xfs_rw_iunlock(ip, iolock);
873 struct iov_iter *from)
875 struct file *file = iocb->ki_filp;
876 struct address_space *mapping = file->f_mapping;
877 struct inode *inode = mapping->host;
878 struct xfs_inode *ip = XFS_I(inode);
880 size_t ocount = iov_iter_count(from);
882 XFS_STATS_INC(ip->i_mount, xs_write_calls);
887 if (XFS_FORCED_SHUTDOWN(ip->i_mount))
890 if ((iocb->ki_flags & IOCB_DIRECT) || IS_DAX(inode))
891 ret = xfs_file_dio_aio_write(iocb, from);
893 ret = xfs_file_buffered_aio_write(iocb, from);
896 XFS_STATS_ADD(ip->i_mount, xs_write_bytes, ret);
898 /* Handle various SYNC-type writes */
899 ret = generic_write_sync(iocb, ret);
904 #define XFS_FALLOC_FL_SUPPORTED \
905 (FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE | \
906 FALLOC_FL_COLLAPSE_RANGE | FALLOC_FL_ZERO_RANGE | \
907 FALLOC_FL_INSERT_RANGE)
916 struct inode *inode = file_inode(file);
917 struct xfs_inode *ip = XFS_I(inode);
919 enum xfs_prealloc_flags flags = 0;
920 uint iolock = XFS_IOLOCK_EXCL;
922 bool do_file_insert = 0;
924 if (!S_ISREG(inode->i_mode))
926 if (mode & ~XFS_FALLOC_FL_SUPPORTED)
929 xfs_ilock(ip, iolock);
930 error = xfs_break_layouts(inode, &iolock, false);
934 xfs_ilock(ip, XFS_MMAPLOCK_EXCL);
935 iolock |= XFS_MMAPLOCK_EXCL;
937 if (mode & FALLOC_FL_PUNCH_HOLE) {
938 error = xfs_free_file_space(ip, offset, len);
941 } else if (mode & FALLOC_FL_COLLAPSE_RANGE) {
942 unsigned blksize_mask = (1 << inode->i_blkbits) - 1;
944 if (offset & blksize_mask || len & blksize_mask) {
950 * There is no need to overlap collapse range with EOF,
951 * in which case it is effectively a truncate operation
953 if (offset + len >= i_size_read(inode)) {
958 new_size = i_size_read(inode) - len;
960 error = xfs_collapse_file_space(ip, offset, len);
963 } else if (mode & FALLOC_FL_INSERT_RANGE) {
964 unsigned blksize_mask = (1 << inode->i_blkbits) - 1;
966 new_size = i_size_read(inode) + len;
967 if (offset & blksize_mask || len & blksize_mask) {
972 /* check the new inode size does not wrap through zero */
973 if (new_size > inode->i_sb->s_maxbytes) {
978 /* Offset should be less than i_size */
979 if (offset >= i_size_read(inode)) {
985 flags |= XFS_PREALLOC_SET;
987 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
988 offset + len > i_size_read(inode)) {
989 new_size = offset + len;
990 error = inode_newsize_ok(inode, new_size);
995 if (mode & FALLOC_FL_ZERO_RANGE)
996 error = xfs_zero_file_space(ip, offset, len);
998 error = xfs_alloc_file_space(ip, offset, len,
1004 if (file->f_flags & O_DSYNC)
1005 flags |= XFS_PREALLOC_SYNC;
1007 error = xfs_update_prealloc_flags(ip, flags);
1011 /* Change file size if needed */
1015 iattr.ia_valid = ATTR_SIZE;
1016 iattr.ia_size = new_size;
1017 error = xfs_setattr_size(ip, &iattr);
1023 * Perform hole insertion now that the file size has been
1024 * updated so that if we crash during the operation we don't
1025 * leave shifted extents past EOF and hence losing access to
1026 * the data that is contained within them.
1029 error = xfs_insert_file_space(ip, offset, len);
1032 xfs_iunlock(ip, iolock);
1039 struct inode *inode,
1042 if (!(file->f_flags & O_LARGEFILE) && i_size_read(inode) > MAX_NON_LFS)
1044 if (XFS_FORCED_SHUTDOWN(XFS_M(inode->i_sb)))
1051 struct inode *inode,
1054 struct xfs_inode *ip = XFS_I(inode);
1058 error = xfs_file_open(inode, file);
1063 * If there are any blocks, read-ahead block 0 as we're almost
1064 * certain to have the next operation be a read there.
1066 mode = xfs_ilock_data_map_shared(ip);
1067 if (ip->i_d.di_nextents > 0)
1068 xfs_dir3_data_readahead(ip, 0, -1);
1069 xfs_iunlock(ip, mode);
1075 struct inode *inode,
1078 return xfs_release(XFS_I(inode));
1084 struct dir_context *ctx)
1086 struct inode *inode = file_inode(file);
1087 xfs_inode_t *ip = XFS_I(inode);
1091 * The Linux API doesn't pass down the total size of the buffer
1092 * we read into down to the filesystem. With the filldir concept
1093 * it's not needed for correct information, but the XFS dir2 leaf
1094 * code wants an estimate of the buffer size to calculate it's
1095 * readahead window and size the buffers used for mapping to
1098 * Try to give it an estimate that's good enough, maybe at some
1099 * point we can change the ->readdir prototype to include the
1100 * buffer size. For now we use the current glibc buffer size.
1102 bufsize = (size_t)min_t(loff_t, 32768, ip->i_d.di_size);
1104 return xfs_readdir(ip, ctx, bufsize);
1108 * This type is designed to indicate the type of offset we would like
1109 * to search from page cache for xfs_seek_hole_data().
1117 * Lookup the desired type of offset from the given page.
1119 * On success, return true and the offset argument will point to the
1120 * start of the region that was found. Otherwise this function will
1121 * return false and keep the offset argument unchanged.
1124 xfs_lookup_buffer_offset(
1129 loff_t lastoff = page_offset(page);
1131 struct buffer_head *bh, *head;
1133 bh = head = page_buffers(page);
1136 * Unwritten extents that have data in the page
1137 * cache covering them can be identified by the
1138 * BH_Unwritten state flag. Pages with multiple
1139 * buffers might have a mix of holes, data and
1140 * unwritten extents - any buffer with valid
1141 * data in it should have BH_Uptodate flag set
1144 if (buffer_unwritten(bh) ||
1145 buffer_uptodate(bh)) {
1146 if (type == DATA_OFF)
1149 if (type == HOLE_OFF)
1157 lastoff += bh->b_size;
1158 } while ((bh = bh->b_this_page) != head);
1164 * This routine is called to find out and return a data or hole offset
1165 * from the page cache for unwritten extents according to the desired
1166 * type for xfs_seek_hole_data().
1168 * The argument offset is used to tell where we start to search from the
1169 * page cache. Map is used to figure out the end points of the range to
1172 * Return true if the desired type of offset was found, and the argument
1173 * offset is filled with that address. Otherwise, return false and keep
1177 xfs_find_get_desired_pgoff(
1178 struct inode *inode,
1179 struct xfs_bmbt_irec *map,
1183 struct xfs_inode *ip = XFS_I(inode);
1184 struct xfs_mount *mp = ip->i_mount;
1185 struct pagevec pvec;
1189 loff_t startoff = *offset;
1190 loff_t lastoff = startoff;
1193 pagevec_init(&pvec, 0);
1195 index = startoff >> PAGE_SHIFT;
1196 endoff = XFS_FSB_TO_B(mp, map->br_startoff + map->br_blockcount);
1197 end = endoff >> PAGE_SHIFT;
1203 want = min_t(pgoff_t, end - index, PAGEVEC_SIZE);
1204 nr_pages = pagevec_lookup(&pvec, inode->i_mapping, index,
1207 * No page mapped into given range. If we are searching holes
1208 * and if this is the first time we got into the loop, it means
1209 * that the given offset is landed in a hole, return it.
1211 * If we have already stepped through some block buffers to find
1212 * holes but they all contains data. In this case, the last
1213 * offset is already updated and pointed to the end of the last
1214 * mapped page, if it does not reach the endpoint to search,
1215 * that means there should be a hole between them.
1217 if (nr_pages == 0) {
1218 /* Data search found nothing */
1219 if (type == DATA_OFF)
1222 ASSERT(type == HOLE_OFF);
1223 if (lastoff == startoff || lastoff < endoff) {
1231 * At lease we found one page. If this is the first time we
1232 * step into the loop, and if the first page index offset is
1233 * greater than the given search offset, a hole was found.
1235 if (type == HOLE_OFF && lastoff == startoff &&
1236 lastoff < page_offset(pvec.pages[0])) {
1241 for (i = 0; i < nr_pages; i++) {
1242 struct page *page = pvec.pages[i];
1246 * At this point, the page may be truncated or
1247 * invalidated (changing page->mapping to NULL),
1248 * or even swizzled back from swapper_space to tmpfs
1249 * file mapping. However, page->index will not change
1250 * because we have a reference on the page.
1252 * Searching done if the page index is out of range.
1253 * If the current offset is not reaches the end of
1254 * the specified search range, there should be a hole
1257 if (page->index > end) {
1258 if (type == HOLE_OFF && lastoff < endoff) {
1267 * Page truncated or invalidated(page->mapping == NULL).
1268 * We can freely skip it and proceed to check the next
1271 if (unlikely(page->mapping != inode->i_mapping)) {
1276 if (!page_has_buffers(page)) {
1281 found = xfs_lookup_buffer_offset(page, &b_offset, type);
1284 * The found offset may be less than the start
1285 * point to search if this is the first time to
1288 *offset = max_t(loff_t, startoff, b_offset);
1294 * We either searching data but nothing was found, or
1295 * searching hole but found a data buffer. In either
1296 * case, probably the next page contains the desired
1297 * things, update the last offset to it so.
1299 lastoff = page_offset(page) + PAGE_SIZE;
1304 * The number of returned pages less than our desired, search
1305 * done. In this case, nothing was found for searching data,
1306 * but we found a hole behind the last offset.
1308 if (nr_pages < want) {
1309 if (type == HOLE_OFF) {
1316 index = pvec.pages[i - 1]->index + 1;
1317 pagevec_release(&pvec);
1318 } while (index <= end);
1321 pagevec_release(&pvec);
1326 * caller must lock inode with xfs_ilock_data_map_shared,
1327 * can we craft an appropriate ASSERT?
1329 * end is because the VFS-level lseek interface is defined such that any
1330 * offset past i_size shall return -ENXIO, but we use this for quota code
1331 * which does not maintain i_size, and we want to SEEK_DATA past i_size.
1334 __xfs_seek_hole_data(
1335 struct inode *inode,
1340 struct xfs_inode *ip = XFS_I(inode);
1341 struct xfs_mount *mp = ip->i_mount;
1342 loff_t uninitialized_var(offset);
1343 xfs_fileoff_t fsbno;
1344 xfs_filblks_t lastbno;
1353 * Try to read extents from the first block indicated
1354 * by fsbno to the end block of the file.
1356 fsbno = XFS_B_TO_FSBT(mp, start);
1357 lastbno = XFS_B_TO_FSB(mp, end);
1360 struct xfs_bmbt_irec map[2];
1364 error = xfs_bmapi_read(ip, fsbno, lastbno - fsbno, map, &nmap,
1369 /* No extents at given offset, must be beyond EOF */
1375 for (i = 0; i < nmap; i++) {
1376 offset = max_t(loff_t, start,
1377 XFS_FSB_TO_B(mp, map[i].br_startoff));
1379 /* Landed in the hole we wanted? */
1380 if (whence == SEEK_HOLE &&
1381 map[i].br_startblock == HOLESTARTBLOCK)
1384 /* Landed in the data extent we wanted? */
1385 if (whence == SEEK_DATA &&
1386 (map[i].br_startblock == DELAYSTARTBLOCK ||
1387 (map[i].br_state == XFS_EXT_NORM &&
1388 !isnullstartblock(map[i].br_startblock))))
1392 * Landed in an unwritten extent, try to search
1393 * for hole or data from page cache.
1395 if (map[i].br_state == XFS_EXT_UNWRITTEN) {
1396 if (xfs_find_get_desired_pgoff(inode, &map[i],
1397 whence == SEEK_HOLE ? HOLE_OFF : DATA_OFF,
1404 * We only received one extent out of the two requested. This
1405 * means we've hit EOF and didn't find what we are looking for.
1409 * If we were looking for a hole, set offset to
1410 * the end of the file (i.e., there is an implicit
1411 * hole at the end of any file).
1413 if (whence == SEEK_HOLE) {
1418 * If we were looking for data, it's nowhere to be found
1420 ASSERT(whence == SEEK_DATA);
1428 * Nothing was found, proceed to the next round of search
1429 * if the next reading offset is not at or beyond EOF.
1431 fsbno = map[i - 1].br_startoff + map[i - 1].br_blockcount;
1432 start = XFS_FSB_TO_B(mp, fsbno);
1434 if (whence == SEEK_HOLE) {
1438 ASSERT(whence == SEEK_DATA);
1446 * If at this point we have found the hole we wanted, the returned
1447 * offset may be bigger than the file size as it may be aligned to
1448 * page boundary for unwritten extents. We need to deal with this
1449 * situation in particular.
1451 if (whence == SEEK_HOLE)
1452 offset = min_t(loff_t, offset, end);
1466 struct inode *inode = file->f_mapping->host;
1467 struct xfs_inode *ip = XFS_I(inode);
1468 struct xfs_mount *mp = ip->i_mount;
1473 if (XFS_FORCED_SHUTDOWN(mp))
1476 lock = xfs_ilock_data_map_shared(ip);
1478 end = i_size_read(inode);
1479 offset = __xfs_seek_hole_data(inode, start, end, whence);
1485 offset = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
1488 xfs_iunlock(ip, lock);
1505 return generic_file_llseek(file, offset, whence);
1508 return xfs_seek_hole_data(file, offset, whence);
1515 * Locking for serialisation of IO during page faults. This results in a lock
1519 * sb_start_pagefault(vfs, freeze)
1520 * i_mmaplock (XFS - truncate serialisation)
1522 * i_lock (XFS - extent map serialisation)
1526 * mmap()d file has taken write protection fault and is being made writable. We
1527 * can set the page state up correctly for a writable page, which means we can
1528 * do correct delalloc accounting (ENOSPC checking!) and unwritten extent
1532 xfs_filemap_page_mkwrite(
1533 struct vm_area_struct *vma,
1534 struct vm_fault *vmf)
1536 struct inode *inode = file_inode(vma->vm_file);
1539 trace_xfs_filemap_page_mkwrite(XFS_I(inode));
1541 sb_start_pagefault(inode->i_sb);
1542 file_update_time(vma->vm_file);
1543 xfs_ilock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1545 if (IS_DAX(inode)) {
1546 ret = __dax_mkwrite(vma, vmf, xfs_get_blocks_dax_fault);
1548 ret = block_page_mkwrite(vma, vmf, xfs_get_blocks);
1549 ret = block_page_mkwrite_return(ret);
1552 xfs_iunlock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1553 sb_end_pagefault(inode->i_sb);
1560 struct vm_area_struct *vma,
1561 struct vm_fault *vmf)
1563 struct inode *inode = file_inode(vma->vm_file);
1566 trace_xfs_filemap_fault(XFS_I(inode));
1568 /* DAX can shortcut the normal fault path on write faults! */
1569 if ((vmf->flags & FAULT_FLAG_WRITE) && IS_DAX(inode))
1570 return xfs_filemap_page_mkwrite(vma, vmf);
1572 xfs_ilock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1573 if (IS_DAX(inode)) {
1575 * we do not want to trigger unwritten extent conversion on read
1576 * faults - that is unnecessary overhead and would also require
1577 * changes to xfs_get_blocks_direct() to map unwritten extent
1578 * ioend for conversion on read-only mappings.
1580 ret = __dax_fault(vma, vmf, xfs_get_blocks_dax_fault);
1582 ret = filemap_fault(vma, vmf);
1583 xfs_iunlock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1589 * Similar to xfs_filemap_fault(), the DAX fault path can call into here on
1590 * both read and write faults. Hence we need to handle both cases. There is no
1591 * ->pmd_mkwrite callout for huge pages, so we have a single function here to
1592 * handle both cases here. @flags carries the information on the type of fault
1596 xfs_filemap_pmd_fault(
1597 struct vm_area_struct *vma,
1602 struct inode *inode = file_inode(vma->vm_file);
1603 struct xfs_inode *ip = XFS_I(inode);
1607 return VM_FAULT_FALLBACK;
1609 trace_xfs_filemap_pmd_fault(ip);
1611 if (flags & FAULT_FLAG_WRITE) {
1612 sb_start_pagefault(inode->i_sb);
1613 file_update_time(vma->vm_file);
1616 xfs_ilock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1617 ret = __dax_pmd_fault(vma, addr, pmd, flags, xfs_get_blocks_dax_fault);
1618 xfs_iunlock(XFS_I(inode), XFS_MMAPLOCK_SHARED);
1620 if (flags & FAULT_FLAG_WRITE)
1621 sb_end_pagefault(inode->i_sb);
1627 * pfn_mkwrite was originally inteneded to ensure we capture time stamp
1628 * updates on write faults. In reality, it's need to serialise against
1629 * truncate similar to page_mkwrite. Hence we cycle the XFS_MMAPLOCK_SHARED
1630 * to ensure we serialise the fault barrier in place.
1633 xfs_filemap_pfn_mkwrite(
1634 struct vm_area_struct *vma,
1635 struct vm_fault *vmf)
1638 struct inode *inode = file_inode(vma->vm_file);
1639 struct xfs_inode *ip = XFS_I(inode);
1640 int ret = VM_FAULT_NOPAGE;
1643 trace_xfs_filemap_pfn_mkwrite(ip);
1645 sb_start_pagefault(inode->i_sb);
1646 file_update_time(vma->vm_file);
1648 /* check if the faulting page hasn't raced with truncate */
1649 xfs_ilock(ip, XFS_MMAPLOCK_SHARED);
1650 size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
1651 if (vmf->pgoff >= size)
1652 ret = VM_FAULT_SIGBUS;
1653 else if (IS_DAX(inode))
1654 ret = dax_pfn_mkwrite(vma, vmf);
1655 xfs_iunlock(ip, XFS_MMAPLOCK_SHARED);
1656 sb_end_pagefault(inode->i_sb);
1661 static const struct vm_operations_struct xfs_file_vm_ops = {
1662 .fault = xfs_filemap_fault,
1663 .pmd_fault = xfs_filemap_pmd_fault,
1664 .map_pages = filemap_map_pages,
1665 .page_mkwrite = xfs_filemap_page_mkwrite,
1666 .pfn_mkwrite = xfs_filemap_pfn_mkwrite,
1672 struct vm_area_struct *vma)
1674 file_accessed(filp);
1675 vma->vm_ops = &xfs_file_vm_ops;
1676 if (IS_DAX(file_inode(filp)))
1677 vma->vm_flags |= VM_MIXEDMAP | VM_HUGEPAGE;
1681 const struct file_operations xfs_file_operations = {
1682 .llseek = xfs_file_llseek,
1683 .read_iter = xfs_file_read_iter,
1684 .write_iter = xfs_file_write_iter,
1685 .splice_read = xfs_file_splice_read,
1686 .splice_write = iter_file_splice_write,
1687 .unlocked_ioctl = xfs_file_ioctl,
1688 #ifdef CONFIG_COMPAT
1689 .compat_ioctl = xfs_file_compat_ioctl,
1691 .mmap = xfs_file_mmap,
1692 .open = xfs_file_open,
1693 .release = xfs_file_release,
1694 .fsync = xfs_file_fsync,
1695 .fallocate = xfs_file_fallocate,
1698 const struct file_operations xfs_dir_file_operations = {
1699 .open = xfs_dir_open,
1700 .read = generic_read_dir,
1701 .iterate_shared = xfs_file_readdir,
1702 .llseek = generic_file_llseek,
1703 .unlocked_ioctl = xfs_file_ioctl,
1704 #ifdef CONFIG_COMPAT
1705 .compat_ioctl = xfs_file_compat_ioctl,
1707 .fsync = xfs_dir_fsync,