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
25 #include "xfs_trans.h"
26 #include "xfs_dmapi.h"
27 #include "xfs_mount.h"
28 #include "xfs_bmap_btree.h"
29 #include "xfs_alloc_btree.h"
30 #include "xfs_ialloc_btree.h"
31 #include "xfs_dir2_sf.h"
32 #include "xfs_attr_sf.h"
33 #include "xfs_dinode.h"
34 #include "xfs_inode.h"
35 #include "xfs_alloc.h"
36 #include "xfs_btree.h"
37 #include "xfs_error.h"
39 #include "xfs_iomap.h"
40 #include "xfs_vnodeops.h"
41 #include <linux/mpage.h>
42 #include <linux/pagevec.h>
43 #include <linux/writeback.h>
47 * Prime number of hash buckets since address is used as the key.
50 #define to_ioend_wq(v) (&xfs_ioend_wq[((unsigned long)v) % NVSYNC])
51 static wait_queue_head_t xfs_ioend_wq
[NVSYNC
];
58 for (i
= 0; i
< NVSYNC
; i
++)
59 init_waitqueue_head(&xfs_ioend_wq
[i
]);
66 wait_queue_head_t
*wq
= to_ioend_wq(ip
);
68 wait_event(*wq
, (atomic_read(&ip
->i_iocount
) == 0));
75 if (atomic_dec_and_test(&ip
->i_iocount
))
76 wake_up(to_ioend_wq(ip
));
86 struct buffer_head
*bh
, *head
;
88 *delalloc
= *unmapped
= *unwritten
= 0;
90 bh
= head
= page_buffers(page
);
92 if (buffer_uptodate(bh
) && !buffer_mapped(bh
))
94 else if (buffer_unwritten(bh
))
96 else if (buffer_delay(bh
))
98 } while ((bh
= bh
->b_this_page
) != head
);
101 #if defined(XFS_RW_TRACE)
110 loff_t isize
= i_size_read(inode
);
111 loff_t offset
= page_offset(page
);
112 int delalloc
= -1, unmapped
= -1, unwritten
= -1;
114 if (page_has_buffers(page
))
115 xfs_count_page_state(page
, &delalloc
, &unmapped
, &unwritten
);
121 ktrace_enter(ip
->i_rwtrace
,
122 (void *)((unsigned long)tag
),
127 (void *)((unsigned long)((ip
->i_d
.di_size
>> 32) & 0xffffffff)),
128 (void *)((unsigned long)(ip
->i_d
.di_size
& 0xffffffff)),
129 (void *)((unsigned long)((isize
>> 32) & 0xffffffff)),
130 (void *)((unsigned long)(isize
& 0xffffffff)),
131 (void *)((unsigned long)((offset
>> 32) & 0xffffffff)),
132 (void *)((unsigned long)(offset
& 0xffffffff)),
133 (void *)((unsigned long)delalloc
),
134 (void *)((unsigned long)unmapped
),
135 (void *)((unsigned long)unwritten
),
136 (void *)((unsigned long)current_pid()),
140 #define xfs_page_trace(tag, inode, page, pgoff)
143 STATIC
struct block_device
*
144 xfs_find_bdev_for_inode(
145 struct xfs_inode
*ip
)
147 struct xfs_mount
*mp
= ip
->i_mount
;
149 if (XFS_IS_REALTIME_INODE(ip
))
150 return mp
->m_rtdev_targp
->bt_bdev
;
152 return mp
->m_ddev_targp
->bt_bdev
;
156 * We're now finished for good with this ioend structure.
157 * Update the page state via the associated buffer_heads,
158 * release holds on the inode and bio, and finally free
159 * up memory. Do not use the ioend after this.
165 struct buffer_head
*bh
, *next
;
166 struct xfs_inode
*ip
= XFS_I(ioend
->io_inode
);
168 for (bh
= ioend
->io_buffer_head
; bh
; bh
= next
) {
169 next
= bh
->b_private
;
170 bh
->b_end_io(bh
, !ioend
->io_error
);
174 * Volume managers supporting multiple paths can send back ENODEV
175 * when the final path disappears. In this case continuing to fill
176 * the page cache with dirty data which cannot be written out is
177 * evil, so prevent that.
179 if (unlikely(ioend
->io_error
== -ENODEV
)) {
180 xfs_do_force_shutdown(ip
->i_mount
, SHUTDOWN_DEVICE_REQ
,
185 mempool_free(ioend
, xfs_ioend_pool
);
189 * If the end of the current ioend is beyond the current EOF,
190 * return the new EOF value, otherwise zero.
196 xfs_inode_t
*ip
= XFS_I(ioend
->io_inode
);
200 bsize
= ioend
->io_offset
+ ioend
->io_size
;
201 isize
= MAX(ip
->i_size
, ip
->i_new_size
);
202 isize
= MIN(isize
, bsize
);
203 return isize
> ip
->i_d
.di_size
? isize
: 0;
207 * Update on-disk file size now that data has been written to disk.
208 * The current in-memory file size is i_size. If a write is beyond
209 * eof i_new_size will be the intended file size until i_size is
210 * updated. If this write does not extend all the way to the valid
211 * file size then restrict this update to the end of the write.
218 xfs_inode_t
*ip
= XFS_I(ioend
->io_inode
);
221 ASSERT((ip
->i_d
.di_mode
& S_IFMT
) == S_IFREG
);
222 ASSERT(ioend
->io_type
!= IOMAP_READ
);
224 if (unlikely(ioend
->io_error
))
227 xfs_ilock(ip
, XFS_ILOCK_EXCL
);
228 isize
= xfs_ioend_new_eof(ioend
);
230 ip
->i_d
.di_size
= isize
;
231 xfs_mark_inode_dirty_sync(ip
);
234 xfs_iunlock(ip
, XFS_ILOCK_EXCL
);
238 * IO write completion.
242 struct work_struct
*work
)
245 container_of(work
, xfs_ioend_t
, io_work
);
246 struct xfs_inode
*ip
= XFS_I(ioend
->io_inode
);
249 * For unwritten extents we need to issue transactions to convert a
250 * range to normal written extens after the data I/O has finished.
252 if (ioend
->io_type
== IOMAP_UNWRITTEN
&&
253 likely(!ioend
->io_error
&& !XFS_FORCED_SHUTDOWN(ip
->i_mount
))) {
256 error
= xfs_iomap_write_unwritten(ip
, ioend
->io_offset
,
259 ioend
->io_error
= error
;
263 * We might have to update the on-disk file size after extending
266 if (ioend
->io_type
!= IOMAP_READ
)
267 xfs_setfilesize(ioend
);
268 xfs_destroy_ioend(ioend
);
272 * Schedule IO completion handling on a xfsdatad if this was
273 * the final hold on this ioend. If we are asked to wait,
274 * flush the workqueue.
281 if (atomic_dec_and_test(&ioend
->io_remaining
)) {
282 struct workqueue_struct
*wq
;
284 wq
= (ioend
->io_type
== IOMAP_UNWRITTEN
) ?
285 xfsconvertd_workqueue
: xfsdatad_workqueue
;
286 queue_work(wq
, &ioend
->io_work
);
293 * Allocate and initialise an IO completion structure.
294 * We need to track unwritten extent write completion here initially.
295 * We'll need to extend this for updating the ondisk inode size later
305 ioend
= mempool_alloc(xfs_ioend_pool
, GFP_NOFS
);
308 * Set the count to 1 initially, which will prevent an I/O
309 * completion callback from happening before we have started
310 * all the I/O from calling the completion routine too early.
312 atomic_set(&ioend
->io_remaining
, 1);
314 ioend
->io_list
= NULL
;
315 ioend
->io_type
= type
;
316 ioend
->io_inode
= inode
;
317 ioend
->io_buffer_head
= NULL
;
318 ioend
->io_buffer_tail
= NULL
;
319 atomic_inc(&XFS_I(ioend
->io_inode
)->i_iocount
);
320 ioend
->io_offset
= 0;
323 INIT_WORK(&ioend
->io_work
, xfs_end_io
);
337 return -xfs_iomap(XFS_I(inode
), offset
, count
, flags
, mapp
, &nmaps
);
345 return offset
>= iomapp
->iomap_offset
&&
346 offset
< iomapp
->iomap_offset
+ iomapp
->iomap_bsize
;
350 * BIO completion handler for buffered IO.
357 xfs_ioend_t
*ioend
= bio
->bi_private
;
359 ASSERT(atomic_read(&bio
->bi_cnt
) >= 1);
360 ioend
->io_error
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
) ? 0 : error
;
362 /* Toss bio and pass work off to an xfsdatad thread */
363 bio
->bi_private
= NULL
;
364 bio
->bi_end_io
= NULL
;
367 xfs_finish_ioend(ioend
, 0);
371 xfs_submit_ioend_bio(
372 struct writeback_control
*wbc
,
376 atomic_inc(&ioend
->io_remaining
);
377 bio
->bi_private
= ioend
;
378 bio
->bi_end_io
= xfs_end_bio
;
381 * If the I/O is beyond EOF we mark the inode dirty immediately
382 * but don't update the inode size until I/O completion.
384 if (xfs_ioend_new_eof(ioend
))
385 xfs_mark_inode_dirty_sync(XFS_I(ioend
->io_inode
));
387 submit_bio(wbc
->sync_mode
== WB_SYNC_ALL
?
388 WRITE_SYNC_PLUG
: WRITE
, bio
);
389 ASSERT(!bio_flagged(bio
, BIO_EOPNOTSUPP
));
395 struct buffer_head
*bh
)
398 int nvecs
= bio_get_nr_vecs(bh
->b_bdev
);
401 bio
= bio_alloc(GFP_NOIO
, nvecs
);
405 ASSERT(bio
->bi_private
== NULL
);
406 bio
->bi_sector
= bh
->b_blocknr
* (bh
->b_size
>> 9);
407 bio
->bi_bdev
= bh
->b_bdev
;
413 xfs_start_buffer_writeback(
414 struct buffer_head
*bh
)
416 ASSERT(buffer_mapped(bh
));
417 ASSERT(buffer_locked(bh
));
418 ASSERT(!buffer_delay(bh
));
419 ASSERT(!buffer_unwritten(bh
));
421 mark_buffer_async_write(bh
);
422 set_buffer_uptodate(bh
);
423 clear_buffer_dirty(bh
);
427 xfs_start_page_writeback(
432 ASSERT(PageLocked(page
));
433 ASSERT(!PageWriteback(page
));
435 clear_page_dirty_for_io(page
);
436 set_page_writeback(page
);
438 /* If no buffers on the page are to be written, finish it here */
440 end_page_writeback(page
);
443 static inline int bio_add_buffer(struct bio
*bio
, struct buffer_head
*bh
)
445 return bio_add_page(bio
, bh
->b_page
, bh
->b_size
, bh_offset(bh
));
449 * Submit all of the bios for all of the ioends we have saved up, covering the
450 * initial writepage page and also any probed pages.
452 * Because we may have multiple ioends spanning a page, we need to start
453 * writeback on all the buffers before we submit them for I/O. If we mark the
454 * buffers as we got, then we can end up with a page that only has buffers
455 * marked async write and I/O complete on can occur before we mark the other
456 * buffers async write.
458 * The end result of this is that we trip a bug in end_page_writeback() because
459 * we call it twice for the one page as the code in end_buffer_async_write()
460 * assumes that all buffers on the page are started at the same time.
462 * The fix is two passes across the ioend list - one to start writeback on the
463 * buffer_heads, and then submit them for I/O on the second pass.
467 struct writeback_control
*wbc
,
470 xfs_ioend_t
*head
= ioend
;
472 struct buffer_head
*bh
;
474 sector_t lastblock
= 0;
476 /* Pass 1 - start writeback */
478 next
= ioend
->io_list
;
479 for (bh
= ioend
->io_buffer_head
; bh
; bh
= bh
->b_private
) {
480 xfs_start_buffer_writeback(bh
);
482 } while ((ioend
= next
) != NULL
);
484 /* Pass 2 - submit I/O */
487 next
= ioend
->io_list
;
490 for (bh
= ioend
->io_buffer_head
; bh
; bh
= bh
->b_private
) {
494 bio
= xfs_alloc_ioend_bio(bh
);
495 } else if (bh
->b_blocknr
!= lastblock
+ 1) {
496 xfs_submit_ioend_bio(wbc
, ioend
, bio
);
500 if (bio_add_buffer(bio
, bh
) != bh
->b_size
) {
501 xfs_submit_ioend_bio(wbc
, ioend
, bio
);
505 lastblock
= bh
->b_blocknr
;
508 xfs_submit_ioend_bio(wbc
, ioend
, bio
);
509 xfs_finish_ioend(ioend
, 0);
510 } while ((ioend
= next
) != NULL
);
514 * Cancel submission of all buffer_heads so far in this endio.
515 * Toss the endio too. Only ever called for the initial page
516 * in a writepage request, so only ever one page.
523 struct buffer_head
*bh
, *next_bh
;
526 next
= ioend
->io_list
;
527 bh
= ioend
->io_buffer_head
;
529 next_bh
= bh
->b_private
;
530 clear_buffer_async_write(bh
);
532 } while ((bh
= next_bh
) != NULL
);
534 xfs_ioend_wake(XFS_I(ioend
->io_inode
));
535 mempool_free(ioend
, xfs_ioend_pool
);
536 } while ((ioend
= next
) != NULL
);
540 * Test to see if we've been building up a completion structure for
541 * earlier buffers -- if so, we try to append to this ioend if we
542 * can, otherwise we finish off any current ioend and start another.
543 * Return true if we've finished the given ioend.
548 struct buffer_head
*bh
,
551 xfs_ioend_t
**result
,
554 xfs_ioend_t
*ioend
= *result
;
556 if (!ioend
|| need_ioend
|| type
!= ioend
->io_type
) {
557 xfs_ioend_t
*previous
= *result
;
559 ioend
= xfs_alloc_ioend(inode
, type
);
560 ioend
->io_offset
= offset
;
561 ioend
->io_buffer_head
= bh
;
562 ioend
->io_buffer_tail
= bh
;
564 previous
->io_list
= ioend
;
567 ioend
->io_buffer_tail
->b_private
= bh
;
568 ioend
->io_buffer_tail
= bh
;
571 bh
->b_private
= NULL
;
572 ioend
->io_size
+= bh
->b_size
;
577 struct buffer_head
*bh
,
584 ASSERT(mp
->iomap_bn
!= IOMAP_DADDR_NULL
);
586 bn
= (mp
->iomap_bn
>> (block_bits
- BBSHIFT
)) +
587 ((offset
- mp
->iomap_offset
) >> block_bits
);
589 ASSERT(bn
|| (mp
->iomap_flags
& IOMAP_REALTIME
));
592 set_buffer_mapped(bh
);
597 struct buffer_head
*bh
,
602 ASSERT(!(iomapp
->iomap_flags
& IOMAP_HOLE
));
603 ASSERT(!(iomapp
->iomap_flags
& IOMAP_DELAY
));
606 xfs_map_buffer(bh
, iomapp
, offset
, block_bits
);
607 bh
->b_bdev
= iomapp
->iomap_target
->bt_bdev
;
608 set_buffer_mapped(bh
);
609 clear_buffer_delay(bh
);
610 clear_buffer_unwritten(bh
);
614 * Look for a page at index that is suitable for clustering.
619 unsigned int pg_offset
,
624 if (PageWriteback(page
))
627 if (page
->mapping
&& PageDirty(page
)) {
628 if (page_has_buffers(page
)) {
629 struct buffer_head
*bh
, *head
;
631 bh
= head
= page_buffers(page
);
633 if (!buffer_uptodate(bh
))
635 if (mapped
!= buffer_mapped(bh
))
638 if (ret
>= pg_offset
)
640 } while ((bh
= bh
->b_this_page
) != head
);
642 ret
= mapped
? 0 : PAGE_CACHE_SIZE
;
651 struct page
*startpage
,
652 struct buffer_head
*bh
,
653 struct buffer_head
*head
,
657 pgoff_t tindex
, tlast
, tloff
;
661 /* First sum forwards in this page */
663 if (!buffer_uptodate(bh
) || (mapped
!= buffer_mapped(bh
)))
666 } while ((bh
= bh
->b_this_page
) != head
);
668 /* if we reached the end of the page, sum forwards in following pages */
669 tlast
= i_size_read(inode
) >> PAGE_CACHE_SHIFT
;
670 tindex
= startpage
->index
+ 1;
672 /* Prune this back to avoid pathological behavior */
673 tloff
= min(tlast
, startpage
->index
+ 64);
675 pagevec_init(&pvec
, 0);
676 while (!done
&& tindex
<= tloff
) {
677 unsigned len
= min_t(pgoff_t
, PAGEVEC_SIZE
, tlast
- tindex
+ 1);
679 if (!pagevec_lookup(&pvec
, inode
->i_mapping
, tindex
, len
))
682 for (i
= 0; i
< pagevec_count(&pvec
); i
++) {
683 struct page
*page
= pvec
.pages
[i
];
684 size_t pg_offset
, pg_len
= 0;
686 if (tindex
== tlast
) {
688 i_size_read(inode
) & (PAGE_CACHE_SIZE
- 1);
694 pg_offset
= PAGE_CACHE_SIZE
;
696 if (page
->index
== tindex
&& trylock_page(page
)) {
697 pg_len
= xfs_probe_page(page
, pg_offset
, mapped
);
710 pagevec_release(&pvec
);
718 * Test if a given page is suitable for writing as part of an unwritten
719 * or delayed allocate extent.
726 if (PageWriteback(page
))
729 if (page
->mapping
&& page_has_buffers(page
)) {
730 struct buffer_head
*bh
, *head
;
733 bh
= head
= page_buffers(page
);
735 if (buffer_unwritten(bh
))
736 acceptable
= (type
== IOMAP_UNWRITTEN
);
737 else if (buffer_delay(bh
))
738 acceptable
= (type
== IOMAP_DELAY
);
739 else if (buffer_dirty(bh
) && buffer_mapped(bh
))
740 acceptable
= (type
== IOMAP_NEW
);
743 } while ((bh
= bh
->b_this_page
) != head
);
753 * Allocate & map buffers for page given the extent map. Write it out.
754 * except for the original page of a writepage, this is called on
755 * delalloc/unwritten pages only, for the original page it is possible
756 * that the page has no mapping at all.
764 xfs_ioend_t
**ioendp
,
765 struct writeback_control
*wbc
,
769 struct buffer_head
*bh
, *head
;
770 xfs_off_t end_offset
;
771 unsigned long p_offset
;
773 int bbits
= inode
->i_blkbits
;
775 int count
= 0, done
= 0, uptodate
= 1;
776 xfs_off_t offset
= page_offset(page
);
778 if (page
->index
!= tindex
)
780 if (!trylock_page(page
))
782 if (PageWriteback(page
))
783 goto fail_unlock_page
;
784 if (page
->mapping
!= inode
->i_mapping
)
785 goto fail_unlock_page
;
786 if (!xfs_is_delayed_page(page
, (*ioendp
)->io_type
))
787 goto fail_unlock_page
;
790 * page_dirty is initially a count of buffers on the page before
791 * EOF and is decremented as we move each into a cleanable state.
795 * End offset is the highest offset that this page should represent.
796 * If we are on the last page, (end_offset & (PAGE_CACHE_SIZE - 1))
797 * will evaluate non-zero and be less than PAGE_CACHE_SIZE and
798 * hence give us the correct page_dirty count. On any other page,
799 * it will be zero and in that case we need page_dirty to be the
800 * count of buffers on the page.
802 end_offset
= min_t(unsigned long long,
803 (xfs_off_t
)(page
->index
+ 1) << PAGE_CACHE_SHIFT
,
806 len
= 1 << inode
->i_blkbits
;
807 p_offset
= min_t(unsigned long, end_offset
& (PAGE_CACHE_SIZE
- 1),
809 p_offset
= p_offset
? roundup(p_offset
, len
) : PAGE_CACHE_SIZE
;
810 page_dirty
= p_offset
/ len
;
812 bh
= head
= page_buffers(page
);
814 if (offset
>= end_offset
)
816 if (!buffer_uptodate(bh
))
818 if (!(PageUptodate(page
) || buffer_uptodate(bh
))) {
823 if (buffer_unwritten(bh
) || buffer_delay(bh
)) {
824 if (buffer_unwritten(bh
))
825 type
= IOMAP_UNWRITTEN
;
829 if (!xfs_iomap_valid(mp
, offset
)) {
834 ASSERT(!(mp
->iomap_flags
& IOMAP_HOLE
));
835 ASSERT(!(mp
->iomap_flags
& IOMAP_DELAY
));
837 xfs_map_at_offset(bh
, offset
, bbits
, mp
);
839 xfs_add_to_ioend(inode
, bh
, offset
,
842 set_buffer_dirty(bh
);
844 mark_buffer_dirty(bh
);
850 if (buffer_mapped(bh
) && all_bh
&& startio
) {
852 xfs_add_to_ioend(inode
, bh
, offset
,
860 } while (offset
+= len
, (bh
= bh
->b_this_page
) != head
);
862 if (uptodate
&& bh
== head
)
863 SetPageUptodate(page
);
867 struct backing_dev_info
*bdi
;
869 bdi
= inode
->i_mapping
->backing_dev_info
;
871 if (bdi_write_congested(bdi
)) {
872 wbc
->encountered_congestion
= 1;
874 } else if (wbc
->nr_to_write
<= 0) {
878 xfs_start_page_writeback(page
, !page_dirty
, count
);
889 * Convert & write out a cluster of pages in the same extent as defined
890 * by mp and following the start page.
897 xfs_ioend_t
**ioendp
,
898 struct writeback_control
*wbc
,
906 pagevec_init(&pvec
, 0);
907 while (!done
&& tindex
<= tlast
) {
908 unsigned len
= min_t(pgoff_t
, PAGEVEC_SIZE
, tlast
- tindex
+ 1);
910 if (!pagevec_lookup(&pvec
, inode
->i_mapping
, tindex
, len
))
913 for (i
= 0; i
< pagevec_count(&pvec
); i
++) {
914 done
= xfs_convert_page(inode
, pvec
.pages
[i
], tindex
++,
915 iomapp
, ioendp
, wbc
, startio
, all_bh
);
920 pagevec_release(&pvec
);
926 * Calling this without startio set means we are being asked to make a dirty
927 * page ready for freeing it's buffers. When called with startio set then
928 * we are coming from writepage.
930 * When called with startio set it is important that we write the WHOLE
932 * The bh->b_state's cannot know if any of the blocks or which block for
933 * that matter are dirty due to mmap writes, and therefore bh uptodate is
934 * only valid if the page itself isn't completely uptodate. Some layers
935 * may clear the page dirty flag prior to calling write page, under the
936 * assumption the entire page will be written out; by not writing out the
937 * whole page the page can be reused before all valid dirty data is
938 * written out. Note: in the case of a page that has been dirty'd by
939 * mapwrite and but partially setup by block_prepare_write the
940 * bh->b_states's will not agree and only ones setup by BPW/BCW will have
941 * valid state, thus the whole page must be written out thing.
945 xfs_page_state_convert(
948 struct writeback_control
*wbc
,
950 int unmapped
) /* also implies page uptodate */
952 struct buffer_head
*bh
, *head
;
954 xfs_ioend_t
*ioend
= NULL
, *iohead
= NULL
;
956 unsigned long p_offset
= 0;
958 __uint64_t end_offset
;
959 pgoff_t end_index
, last_index
, tlast
;
961 int flags
, err
, iomap_valid
= 0, uptodate
= 1;
962 int page_dirty
, count
= 0;
964 int all_bh
= unmapped
;
967 if (wbc
->sync_mode
== WB_SYNC_NONE
&& wbc
->nonblocking
)
968 trylock
|= BMAPI_TRYLOCK
;
971 /* Is this page beyond the end of the file? */
972 offset
= i_size_read(inode
);
973 end_index
= offset
>> PAGE_CACHE_SHIFT
;
974 last_index
= (offset
- 1) >> PAGE_CACHE_SHIFT
;
975 if (page
->index
>= end_index
) {
976 if ((page
->index
>= end_index
+ 1) ||
977 !(i_size_read(inode
) & (PAGE_CACHE_SIZE
- 1))) {
985 * page_dirty is initially a count of buffers on the page before
986 * EOF and is decremented as we move each into a cleanable state.
990 * End offset is the highest offset that this page should represent.
991 * If we are on the last page, (end_offset & (PAGE_CACHE_SIZE - 1))
992 * will evaluate non-zero and be less than PAGE_CACHE_SIZE and
993 * hence give us the correct page_dirty count. On any other page,
994 * it will be zero and in that case we need page_dirty to be the
995 * count of buffers on the page.
997 end_offset
= min_t(unsigned long long,
998 (xfs_off_t
)(page
->index
+ 1) << PAGE_CACHE_SHIFT
, offset
);
999 len
= 1 << inode
->i_blkbits
;
1000 p_offset
= min_t(unsigned long, end_offset
& (PAGE_CACHE_SIZE
- 1),
1002 p_offset
= p_offset
? roundup(p_offset
, len
) : PAGE_CACHE_SIZE
;
1003 page_dirty
= p_offset
/ len
;
1005 bh
= head
= page_buffers(page
);
1006 offset
= page_offset(page
);
1010 /* TODO: cleanup count and page_dirty */
1013 if (offset
>= end_offset
)
1015 if (!buffer_uptodate(bh
))
1017 if (!(PageUptodate(page
) || buffer_uptodate(bh
)) && !startio
) {
1019 * the iomap is actually still valid, but the ioend
1020 * isn't. shouldn't happen too often.
1027 iomap_valid
= xfs_iomap_valid(&iomap
, offset
);
1030 * First case, map an unwritten extent and prepare for
1031 * extent state conversion transaction on completion.
1033 * Second case, allocate space for a delalloc buffer.
1034 * We can return EAGAIN here in the release page case.
1036 * Third case, an unmapped buffer was found, and we are
1037 * in a path where we need to write the whole page out.
1039 if (buffer_unwritten(bh
) || buffer_delay(bh
) ||
1040 ((buffer_uptodate(bh
) || PageUptodate(page
)) &&
1041 !buffer_mapped(bh
) && (unmapped
|| startio
))) {
1045 * Make sure we don't use a read-only iomap
1047 if (flags
== BMAPI_READ
)
1050 if (buffer_unwritten(bh
)) {
1051 type
= IOMAP_UNWRITTEN
;
1052 flags
= BMAPI_WRITE
| BMAPI_IGNSTATE
;
1053 } else if (buffer_delay(bh
)) {
1055 flags
= BMAPI_ALLOCATE
| trylock
;
1058 flags
= BMAPI_WRITE
| BMAPI_MMAP
;
1063 * if we didn't have a valid mapping then we
1064 * need to ensure that we put the new mapping
1065 * in a new ioend structure. This needs to be
1066 * done to ensure that the ioends correctly
1067 * reflect the block mappings at io completion
1068 * for unwritten extent conversion.
1071 if (type
== IOMAP_NEW
) {
1072 size
= xfs_probe_cluster(inode
,
1078 err
= xfs_map_blocks(inode
, offset
, size
,
1082 iomap_valid
= xfs_iomap_valid(&iomap
, offset
);
1085 xfs_map_at_offset(bh
, offset
,
1086 inode
->i_blkbits
, &iomap
);
1088 xfs_add_to_ioend(inode
, bh
, offset
,
1092 set_buffer_dirty(bh
);
1094 mark_buffer_dirty(bh
);
1099 } else if (buffer_uptodate(bh
) && startio
) {
1101 * we got here because the buffer is already mapped.
1102 * That means it must already have extents allocated
1103 * underneath it. Map the extent by reading it.
1105 if (!iomap_valid
|| flags
!= BMAPI_READ
) {
1107 size
= xfs_probe_cluster(inode
, page
, bh
,
1109 err
= xfs_map_blocks(inode
, offset
, size
,
1113 iomap_valid
= xfs_iomap_valid(&iomap
, offset
);
1117 * We set the type to IOMAP_NEW in case we are doing a
1118 * small write at EOF that is extending the file but
1119 * without needing an allocation. We need to update the
1120 * file size on I/O completion in this case so it is
1121 * the same case as having just allocated a new extent
1122 * that we are writing into for the first time.
1125 if (trylock_buffer(bh
)) {
1126 ASSERT(buffer_mapped(bh
));
1129 xfs_add_to_ioend(inode
, bh
, offset
, type
,
1130 &ioend
, !iomap_valid
);
1136 } else if ((buffer_uptodate(bh
) || PageUptodate(page
)) &&
1137 (unmapped
|| startio
)) {
1144 } while (offset
+= len
, ((bh
= bh
->b_this_page
) != head
));
1146 if (uptodate
&& bh
== head
)
1147 SetPageUptodate(page
);
1150 xfs_start_page_writeback(page
, 1, count
);
1152 if (ioend
&& iomap_valid
) {
1153 offset
= (iomap
.iomap_offset
+ iomap
.iomap_bsize
- 1) >>
1155 tlast
= min_t(pgoff_t
, offset
, last_index
);
1156 xfs_cluster_write(inode
, page
->index
+ 1, &iomap
, &ioend
,
1157 wbc
, startio
, all_bh
, tlast
);
1161 xfs_submit_ioend(wbc
, iohead
);
1167 xfs_cancel_ioend(iohead
);
1170 * If it's delalloc and we have nowhere to put it,
1171 * throw it away, unless the lower layers told
1174 if (err
!= -EAGAIN
) {
1176 block_invalidatepage(page
, 0);
1177 ClearPageUptodate(page
);
1183 * writepage: Called from one of two places:
1185 * 1. we are flushing a delalloc buffer head.
1187 * 2. we are writing out a dirty page. Typically the page dirty
1188 * state is cleared before we get here. In this case is it
1189 * conceivable we have no buffer heads.
1191 * For delalloc space on the page we need to allocate space and
1192 * flush it. For unmapped buffer heads on the page we should
1193 * allocate space if the page is uptodate. For any other dirty
1194 * buffer heads on the page we should flush them.
1196 * If we detect that a transaction would be required to flush
1197 * the page, we have to check the process flags first, if we
1198 * are already in a transaction or disk I/O during allocations
1199 * is off, we need to fail the writepage and redirty the page.
1205 struct writeback_control
*wbc
)
1209 int delalloc
, unmapped
, unwritten
;
1210 struct inode
*inode
= page
->mapping
->host
;
1212 xfs_page_trace(XFS_WRITEPAGE_ENTER
, inode
, page
, 0);
1215 * We need a transaction if:
1216 * 1. There are delalloc buffers on the page
1217 * 2. The page is uptodate and we have unmapped buffers
1218 * 3. The page is uptodate and we have no buffers
1219 * 4. There are unwritten buffers on the page
1222 if (!page_has_buffers(page
)) {
1226 xfs_count_page_state(page
, &delalloc
, &unmapped
, &unwritten
);
1227 if (!PageUptodate(page
))
1229 need_trans
= delalloc
+ unmapped
+ unwritten
;
1233 * If we need a transaction and the process flags say
1234 * we are already in a transaction, or no IO is allowed
1235 * then mark the page dirty again and leave the page
1238 if (current_test_flags(PF_FSTRANS
) && need_trans
)
1242 * Delay hooking up buffer heads until we have
1243 * made our go/no-go decision.
1245 if (!page_has_buffers(page
))
1246 create_empty_buffers(page
, 1 << inode
->i_blkbits
, 0);
1250 * VM calculation for nr_to_write seems off. Bump it way
1251 * up, this gets simple streaming writes zippy again.
1252 * To be reviewed again after Jens' writeback changes.
1254 wbc
->nr_to_write
*= 4;
1257 * Convert delayed allocate, unwritten or unmapped space
1258 * to real space and flush out to disk.
1260 error
= xfs_page_state_convert(inode
, page
, wbc
, 1, unmapped
);
1261 if (error
== -EAGAIN
)
1263 if (unlikely(error
< 0))
1269 redirty_page_for_writepage(wbc
, page
);
1279 struct address_space
*mapping
,
1280 struct writeback_control
*wbc
)
1282 xfs_iflags_clear(XFS_I(mapping
->host
), XFS_ITRUNCATED
);
1283 return generic_writepages(mapping
, wbc
);
1287 * Called to move a page into cleanable state - and from there
1288 * to be released. Possibly the page is already clean. We always
1289 * have buffer heads in this call.
1291 * Returns 0 if the page is ok to release, 1 otherwise.
1293 * Possible scenarios are:
1295 * 1. We are being called to release a page which has been written
1296 * to via regular I/O. buffer heads will be dirty and possibly
1297 * delalloc. If no delalloc buffer heads in this case then we
1298 * can just return zero.
1300 * 2. We are called to release a page which has been written via
1301 * mmap, all we need to do is ensure there is no delalloc
1302 * state in the buffer heads, if not we can let the caller
1303 * free them and we should come back later via writepage.
1310 struct inode
*inode
= page
->mapping
->host
;
1311 int dirty
, delalloc
, unmapped
, unwritten
;
1312 struct writeback_control wbc
= {
1313 .sync_mode
= WB_SYNC_ALL
,
1317 xfs_page_trace(XFS_RELEASEPAGE_ENTER
, inode
, page
, 0);
1319 if (!page_has_buffers(page
))
1322 xfs_count_page_state(page
, &delalloc
, &unmapped
, &unwritten
);
1323 if (!delalloc
&& !unwritten
)
1326 if (!(gfp_mask
& __GFP_FS
))
1329 /* If we are already inside a transaction or the thread cannot
1330 * do I/O, we cannot release this page.
1332 if (current_test_flags(PF_FSTRANS
))
1336 * Convert delalloc space to real space, do not flush the
1337 * data out to disk, that will be done by the caller.
1338 * Never need to allocate space here - we will always
1339 * come back to writepage in that case.
1341 dirty
= xfs_page_state_convert(inode
, page
, &wbc
, 0, 0);
1342 if (dirty
== 0 && !unwritten
)
1347 return try_to_free_buffers(page
);
1352 struct inode
*inode
,
1354 struct buffer_head
*bh_result
,
1357 bmapi_flags_t flags
)
1365 offset
= (xfs_off_t
)iblock
<< inode
->i_blkbits
;
1366 ASSERT(bh_result
->b_size
>= (1 << inode
->i_blkbits
));
1367 size
= bh_result
->b_size
;
1369 if (!create
&& direct
&& offset
>= i_size_read(inode
))
1372 error
= xfs_iomap(XFS_I(inode
), offset
, size
,
1373 create
? flags
: BMAPI_READ
, &iomap
, &niomap
);
1379 if (iomap
.iomap_bn
!= IOMAP_DADDR_NULL
) {
1381 * For unwritten extents do not report a disk address on
1382 * the read case (treat as if we're reading into a hole).
1384 if (create
|| !(iomap
.iomap_flags
& IOMAP_UNWRITTEN
)) {
1385 xfs_map_buffer(bh_result
, &iomap
, offset
,
1388 if (create
&& (iomap
.iomap_flags
& IOMAP_UNWRITTEN
)) {
1390 bh_result
->b_private
= inode
;
1391 set_buffer_unwritten(bh_result
);
1396 * If this is a realtime file, data may be on a different device.
1397 * to that pointed to from the buffer_head b_bdev currently.
1399 bh_result
->b_bdev
= iomap
.iomap_target
->bt_bdev
;
1402 * If we previously allocated a block out beyond eof and we are now
1403 * coming back to use it then we will need to flag it as new even if it
1404 * has a disk address.
1406 * With sub-block writes into unwritten extents we also need to mark
1407 * the buffer as new so that the unwritten parts of the buffer gets
1411 ((!buffer_mapped(bh_result
) && !buffer_uptodate(bh_result
)) ||
1412 (offset
>= i_size_read(inode
)) ||
1413 (iomap
.iomap_flags
& (IOMAP_NEW
|IOMAP_UNWRITTEN
))))
1414 set_buffer_new(bh_result
);
1416 if (iomap
.iomap_flags
& IOMAP_DELAY
) {
1419 set_buffer_uptodate(bh_result
);
1420 set_buffer_mapped(bh_result
);
1421 set_buffer_delay(bh_result
);
1425 if (direct
|| size
> (1 << inode
->i_blkbits
)) {
1426 ASSERT(iomap
.iomap_bsize
- iomap
.iomap_delta
> 0);
1427 offset
= min_t(xfs_off_t
,
1428 iomap
.iomap_bsize
- iomap
.iomap_delta
, size
);
1429 bh_result
->b_size
= (ssize_t
)min_t(xfs_off_t
, LONG_MAX
, offset
);
1437 struct inode
*inode
,
1439 struct buffer_head
*bh_result
,
1442 return __xfs_get_blocks(inode
, iblock
,
1443 bh_result
, create
, 0, BMAPI_WRITE
);
1447 xfs_get_blocks_direct(
1448 struct inode
*inode
,
1450 struct buffer_head
*bh_result
,
1453 return __xfs_get_blocks(inode
, iblock
,
1454 bh_result
, create
, 1, BMAPI_WRITE
|BMAPI_DIRECT
);
1464 xfs_ioend_t
*ioend
= iocb
->private;
1467 * Non-NULL private data means we need to issue a transaction to
1468 * convert a range from unwritten to written extents. This needs
1469 * to happen from process context but aio+dio I/O completion
1470 * happens from irq context so we need to defer it to a workqueue.
1471 * This is not necessary for synchronous direct I/O, but we do
1472 * it anyway to keep the code uniform and simpler.
1474 * Well, if only it were that simple. Because synchronous direct I/O
1475 * requires extent conversion to occur *before* we return to userspace,
1476 * we have to wait for extent conversion to complete. Look at the
1477 * iocb that has been passed to us to determine if this is AIO or
1478 * not. If it is synchronous, tell xfs_finish_ioend() to kick the
1479 * workqueue and wait for it to complete.
1481 * The core direct I/O code might be changed to always call the
1482 * completion handler in the future, in which case all this can
1485 ioend
->io_offset
= offset
;
1486 ioend
->io_size
= size
;
1487 if (ioend
->io_type
== IOMAP_READ
) {
1488 xfs_finish_ioend(ioend
, 0);
1489 } else if (private && size
> 0) {
1490 xfs_finish_ioend(ioend
, is_sync_kiocb(iocb
));
1493 * A direct I/O write ioend starts it's life in unwritten
1494 * state in case they map an unwritten extent. This write
1495 * didn't map an unwritten extent so switch it's completion
1498 ioend
->io_type
= IOMAP_NEW
;
1499 xfs_finish_ioend(ioend
, 0);
1503 * blockdev_direct_IO can return an error even after the I/O
1504 * completion handler was called. Thus we need to protect
1505 * against double-freeing.
1507 iocb
->private = NULL
;
1514 const struct iovec
*iov
,
1516 unsigned long nr_segs
)
1518 struct file
*file
= iocb
->ki_filp
;
1519 struct inode
*inode
= file
->f_mapping
->host
;
1520 struct block_device
*bdev
;
1523 bdev
= xfs_find_bdev_for_inode(XFS_I(inode
));
1526 iocb
->private = xfs_alloc_ioend(inode
, IOMAP_UNWRITTEN
);
1527 ret
= blockdev_direct_IO_own_locking(rw
, iocb
, inode
,
1528 bdev
, iov
, offset
, nr_segs
,
1529 xfs_get_blocks_direct
,
1532 iocb
->private = xfs_alloc_ioend(inode
, IOMAP_READ
);
1533 ret
= blockdev_direct_IO_no_locking(rw
, iocb
, inode
,
1534 bdev
, iov
, offset
, nr_segs
,
1535 xfs_get_blocks_direct
,
1539 if (unlikely(ret
!= -EIOCBQUEUED
&& iocb
->private))
1540 xfs_destroy_ioend(iocb
->private);
1547 struct address_space
*mapping
,
1551 struct page
**pagep
,
1555 return block_write_begin(file
, mapping
, pos
, len
, flags
, pagep
, fsdata
,
1561 struct address_space
*mapping
,
1564 struct inode
*inode
= (struct inode
*)mapping
->host
;
1565 struct xfs_inode
*ip
= XFS_I(inode
);
1567 xfs_itrace_entry(XFS_I(inode
));
1568 xfs_ilock(ip
, XFS_IOLOCK_SHARED
);
1569 xfs_flush_pages(ip
, (xfs_off_t
)0, -1, 0, FI_REMAPF
);
1570 xfs_iunlock(ip
, XFS_IOLOCK_SHARED
);
1571 return generic_block_bmap(mapping
, block
, xfs_get_blocks
);
1576 struct file
*unused
,
1579 return mpage_readpage(page
, xfs_get_blocks
);
1584 struct file
*unused
,
1585 struct address_space
*mapping
,
1586 struct list_head
*pages
,
1589 return mpage_readpages(mapping
, pages
, nr_pages
, xfs_get_blocks
);
1593 xfs_vm_invalidatepage(
1595 unsigned long offset
)
1597 xfs_page_trace(XFS_INVALIDPAGE_ENTER
,
1598 page
->mapping
->host
, page
, offset
);
1599 block_invalidatepage(page
, offset
);
1602 const struct address_space_operations xfs_address_space_operations
= {
1603 .readpage
= xfs_vm_readpage
,
1604 .readpages
= xfs_vm_readpages
,
1605 .writepage
= xfs_vm_writepage
,
1606 .writepages
= xfs_vm_writepages
,
1607 .sync_page
= block_sync_page
,
1608 .releasepage
= xfs_vm_releasepage
,
1609 .invalidatepage
= xfs_vm_invalidatepage
,
1610 .write_begin
= xfs_vm_write_begin
,
1611 .write_end
= generic_write_end
,
1612 .bmap
= xfs_vm_bmap
,
1613 .direct_IO
= xfs_vm_direct_IO
,
1614 .migratepage
= buffer_migrate_page
,
1615 .is_partially_uptodate
= block_is_partially_uptodate
,
1616 .error_remove_page
= generic_error_remove_page
,