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 "xfs_trace.h"
43 #include <linux/gfp.h>
44 #include <linux/mpage.h>
45 #include <linux/pagevec.h>
46 #include <linux/writeback.h>
50 * Prime number of hash buckets since address is used as the key.
53 #define to_ioend_wq(v) (&xfs_ioend_wq[((unsigned long)v) % NVSYNC])
54 static wait_queue_head_t xfs_ioend_wq
[NVSYNC
];
61 for (i
= 0; i
< NVSYNC
; i
++)
62 init_waitqueue_head(&xfs_ioend_wq
[i
]);
69 wait_queue_head_t
*wq
= to_ioend_wq(ip
);
71 wait_event(*wq
, (atomic_read(&ip
->i_iocount
) == 0));
78 if (atomic_dec_and_test(&ip
->i_iocount
))
79 wake_up(to_ioend_wq(ip
));
89 struct buffer_head
*bh
, *head
;
91 *delalloc
= *unmapped
= *unwritten
= 0;
93 bh
= head
= page_buffers(page
);
95 if (buffer_uptodate(bh
) && !buffer_mapped(bh
))
97 else if (buffer_unwritten(bh
))
99 else if (buffer_delay(bh
))
101 } while ((bh
= bh
->b_this_page
) != head
);
104 STATIC
struct block_device
*
105 xfs_find_bdev_for_inode(
106 struct xfs_inode
*ip
)
108 struct xfs_mount
*mp
= ip
->i_mount
;
110 if (XFS_IS_REALTIME_INODE(ip
))
111 return mp
->m_rtdev_targp
->bt_bdev
;
113 return mp
->m_ddev_targp
->bt_bdev
;
117 * We're now finished for good with this ioend structure.
118 * Update the page state via the associated buffer_heads,
119 * release holds on the inode and bio, and finally free
120 * up memory. Do not use the ioend after this.
126 struct buffer_head
*bh
, *next
;
127 struct xfs_inode
*ip
= XFS_I(ioend
->io_inode
);
129 for (bh
= ioend
->io_buffer_head
; bh
; bh
= next
) {
130 next
= bh
->b_private
;
131 bh
->b_end_io(bh
, !ioend
->io_error
);
135 * Volume managers supporting multiple paths can send back ENODEV
136 * when the final path disappears. In this case continuing to fill
137 * the page cache with dirty data which cannot be written out is
138 * evil, so prevent that.
140 if (unlikely(ioend
->io_error
== -ENODEV
)) {
141 xfs_do_force_shutdown(ip
->i_mount
, SHUTDOWN_DEVICE_REQ
,
146 mempool_free(ioend
, xfs_ioend_pool
);
150 * If the end of the current ioend is beyond the current EOF,
151 * return the new EOF value, otherwise zero.
157 xfs_inode_t
*ip
= XFS_I(ioend
->io_inode
);
161 bsize
= ioend
->io_offset
+ ioend
->io_size
;
162 isize
= MAX(ip
->i_size
, ip
->i_new_size
);
163 isize
= MIN(isize
, bsize
);
164 return isize
> ip
->i_d
.di_size
? isize
: 0;
168 * Update on-disk file size now that data has been written to disk. The
169 * current in-memory file size is i_size. If a write is beyond eof i_new_size
170 * will be the intended file size until i_size is updated. If this write does
171 * not extend all the way to the valid file size then restrict this update to
172 * the end of the write.
174 * This function does not block as blocking on the inode lock in IO completion
175 * can lead to IO completion order dependency deadlocks.. If it can't get the
176 * inode ilock it will return EAGAIN. Callers must handle this.
182 xfs_inode_t
*ip
= XFS_I(ioend
->io_inode
);
185 ASSERT((ip
->i_d
.di_mode
& S_IFMT
) == S_IFREG
);
186 ASSERT(ioend
->io_type
!= IOMAP_READ
);
188 if (unlikely(ioend
->io_error
))
191 if (!xfs_ilock_nowait(ip
, XFS_ILOCK_EXCL
))
194 isize
= xfs_ioend_new_eof(ioend
);
196 ip
->i_d
.di_size
= isize
;
197 xfs_mark_inode_dirty(ip
);
200 xfs_iunlock(ip
, XFS_ILOCK_EXCL
);
205 * Schedule IO completion handling on a xfsdatad if this was
206 * the final hold on this ioend. If we are asked to wait,
207 * flush the workqueue.
214 if (atomic_dec_and_test(&ioend
->io_remaining
)) {
215 struct workqueue_struct
*wq
;
217 wq
= (ioend
->io_type
== IOMAP_UNWRITTEN
) ?
218 xfsconvertd_workqueue
: xfsdatad_workqueue
;
219 queue_work(wq
, &ioend
->io_work
);
226 * IO write completion.
230 struct work_struct
*work
)
232 xfs_ioend_t
*ioend
= container_of(work
, xfs_ioend_t
, io_work
);
233 struct xfs_inode
*ip
= XFS_I(ioend
->io_inode
);
237 * For unwritten extents we need to issue transactions to convert a
238 * range to normal written extens after the data I/O has finished.
240 if (ioend
->io_type
== IOMAP_UNWRITTEN
&&
241 likely(!ioend
->io_error
&& !XFS_FORCED_SHUTDOWN(ip
->i_mount
))) {
243 error
= xfs_iomap_write_unwritten(ip
, ioend
->io_offset
,
246 ioend
->io_error
= error
;
250 * We might have to update the on-disk file size after extending
253 if (ioend
->io_type
!= IOMAP_READ
) {
254 error
= xfs_setfilesize(ioend
);
255 ASSERT(!error
|| error
== EAGAIN
);
259 * If we didn't complete processing of the ioend, requeue it to the
260 * tail of the workqueue for another attempt later. Otherwise destroy
263 if (error
== EAGAIN
) {
264 atomic_inc(&ioend
->io_remaining
);
265 xfs_finish_ioend(ioend
, 0);
266 /* ensure we don't spin on blocked ioends */
269 xfs_destroy_ioend(ioend
);
273 * Allocate and initialise an IO completion structure.
274 * We need to track unwritten extent write completion here initially.
275 * We'll need to extend this for updating the ondisk inode size later
285 ioend
= mempool_alloc(xfs_ioend_pool
, GFP_NOFS
);
288 * Set the count to 1 initially, which will prevent an I/O
289 * completion callback from happening before we have started
290 * all the I/O from calling the completion routine too early.
292 atomic_set(&ioend
->io_remaining
, 1);
294 ioend
->io_list
= NULL
;
295 ioend
->io_type
= type
;
296 ioend
->io_inode
= inode
;
297 ioend
->io_buffer_head
= NULL
;
298 ioend
->io_buffer_tail
= NULL
;
299 atomic_inc(&XFS_I(ioend
->io_inode
)->i_iocount
);
300 ioend
->io_offset
= 0;
303 INIT_WORK(&ioend
->io_work
, xfs_end_io
);
317 return -xfs_iomap(XFS_I(inode
), offset
, count
, flags
, mapp
, &nmaps
);
325 return offset
>= iomapp
->iomap_offset
&&
326 offset
< iomapp
->iomap_offset
+ iomapp
->iomap_bsize
;
330 * BIO completion handler for buffered IO.
337 xfs_ioend_t
*ioend
= bio
->bi_private
;
339 ASSERT(atomic_read(&bio
->bi_cnt
) >= 1);
340 ioend
->io_error
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
) ? 0 : error
;
342 /* Toss bio and pass work off to an xfsdatad thread */
343 bio
->bi_private
= NULL
;
344 bio
->bi_end_io
= NULL
;
347 xfs_finish_ioend(ioend
, 0);
351 xfs_submit_ioend_bio(
352 struct writeback_control
*wbc
,
356 atomic_inc(&ioend
->io_remaining
);
357 bio
->bi_private
= ioend
;
358 bio
->bi_end_io
= xfs_end_bio
;
361 * If the I/O is beyond EOF we mark the inode dirty immediately
362 * but don't update the inode size until I/O completion.
364 if (xfs_ioend_new_eof(ioend
))
365 xfs_mark_inode_dirty(XFS_I(ioend
->io_inode
));
367 submit_bio(wbc
->sync_mode
== WB_SYNC_ALL
?
368 WRITE_SYNC_PLUG
: WRITE
, bio
);
369 ASSERT(!bio_flagged(bio
, BIO_EOPNOTSUPP
));
375 struct buffer_head
*bh
)
378 int nvecs
= bio_get_nr_vecs(bh
->b_bdev
);
381 bio
= bio_alloc(GFP_NOIO
, nvecs
);
385 ASSERT(bio
->bi_private
== NULL
);
386 bio
->bi_sector
= bh
->b_blocknr
* (bh
->b_size
>> 9);
387 bio
->bi_bdev
= bh
->b_bdev
;
393 xfs_start_buffer_writeback(
394 struct buffer_head
*bh
)
396 ASSERT(buffer_mapped(bh
));
397 ASSERT(buffer_locked(bh
));
398 ASSERT(!buffer_delay(bh
));
399 ASSERT(!buffer_unwritten(bh
));
401 mark_buffer_async_write(bh
);
402 set_buffer_uptodate(bh
);
403 clear_buffer_dirty(bh
);
407 xfs_start_page_writeback(
412 ASSERT(PageLocked(page
));
413 ASSERT(!PageWriteback(page
));
415 clear_page_dirty_for_io(page
);
416 set_page_writeback(page
);
418 /* If no buffers on the page are to be written, finish it here */
420 end_page_writeback(page
);
423 static inline int bio_add_buffer(struct bio
*bio
, struct buffer_head
*bh
)
425 return bio_add_page(bio
, bh
->b_page
, bh
->b_size
, bh_offset(bh
));
429 * Submit all of the bios for all of the ioends we have saved up, covering the
430 * initial writepage page and also any probed pages.
432 * Because we may have multiple ioends spanning a page, we need to start
433 * writeback on all the buffers before we submit them for I/O. If we mark the
434 * buffers as we got, then we can end up with a page that only has buffers
435 * marked async write and I/O complete on can occur before we mark the other
436 * buffers async write.
438 * The end result of this is that we trip a bug in end_page_writeback() because
439 * we call it twice for the one page as the code in end_buffer_async_write()
440 * assumes that all buffers on the page are started at the same time.
442 * The fix is two passes across the ioend list - one to start writeback on the
443 * buffer_heads, and then submit them for I/O on the second pass.
447 struct writeback_control
*wbc
,
450 xfs_ioend_t
*head
= ioend
;
452 struct buffer_head
*bh
;
454 sector_t lastblock
= 0;
456 /* Pass 1 - start writeback */
458 next
= ioend
->io_list
;
459 for (bh
= ioend
->io_buffer_head
; bh
; bh
= bh
->b_private
) {
460 xfs_start_buffer_writeback(bh
);
462 } while ((ioend
= next
) != NULL
);
464 /* Pass 2 - submit I/O */
467 next
= ioend
->io_list
;
470 for (bh
= ioend
->io_buffer_head
; bh
; bh
= bh
->b_private
) {
474 bio
= xfs_alloc_ioend_bio(bh
);
475 } else if (bh
->b_blocknr
!= lastblock
+ 1) {
476 xfs_submit_ioend_bio(wbc
, ioend
, bio
);
480 if (bio_add_buffer(bio
, bh
) != bh
->b_size
) {
481 xfs_submit_ioend_bio(wbc
, ioend
, bio
);
485 lastblock
= bh
->b_blocknr
;
488 xfs_submit_ioend_bio(wbc
, ioend
, bio
);
489 xfs_finish_ioend(ioend
, 0);
490 } while ((ioend
= next
) != NULL
);
494 * Cancel submission of all buffer_heads so far in this endio.
495 * Toss the endio too. Only ever called for the initial page
496 * in a writepage request, so only ever one page.
503 struct buffer_head
*bh
, *next_bh
;
506 next
= ioend
->io_list
;
507 bh
= ioend
->io_buffer_head
;
509 next_bh
= bh
->b_private
;
510 clear_buffer_async_write(bh
);
512 } while ((bh
= next_bh
) != NULL
);
514 xfs_ioend_wake(XFS_I(ioend
->io_inode
));
515 mempool_free(ioend
, xfs_ioend_pool
);
516 } while ((ioend
= next
) != NULL
);
520 * Test to see if we've been building up a completion structure for
521 * earlier buffers -- if so, we try to append to this ioend if we
522 * can, otherwise we finish off any current ioend and start another.
523 * Return true if we've finished the given ioend.
528 struct buffer_head
*bh
,
531 xfs_ioend_t
**result
,
534 xfs_ioend_t
*ioend
= *result
;
536 if (!ioend
|| need_ioend
|| type
!= ioend
->io_type
) {
537 xfs_ioend_t
*previous
= *result
;
539 ioend
= xfs_alloc_ioend(inode
, type
);
540 ioend
->io_offset
= offset
;
541 ioend
->io_buffer_head
= bh
;
542 ioend
->io_buffer_tail
= bh
;
544 previous
->io_list
= ioend
;
547 ioend
->io_buffer_tail
->b_private
= bh
;
548 ioend
->io_buffer_tail
= bh
;
551 bh
->b_private
= NULL
;
552 ioend
->io_size
+= bh
->b_size
;
557 struct buffer_head
*bh
,
564 ASSERT(mp
->iomap_bn
!= IOMAP_DADDR_NULL
);
566 bn
= (mp
->iomap_bn
>> (block_bits
- BBSHIFT
)) +
567 ((offset
- mp
->iomap_offset
) >> block_bits
);
569 ASSERT(bn
|| (mp
->iomap_flags
& IOMAP_REALTIME
));
572 set_buffer_mapped(bh
);
577 struct buffer_head
*bh
,
582 ASSERT(!(iomapp
->iomap_flags
& IOMAP_HOLE
));
583 ASSERT(!(iomapp
->iomap_flags
& IOMAP_DELAY
));
586 xfs_map_buffer(bh
, iomapp
, offset
, block_bits
);
587 bh
->b_bdev
= iomapp
->iomap_target
->bt_bdev
;
588 set_buffer_mapped(bh
);
589 clear_buffer_delay(bh
);
590 clear_buffer_unwritten(bh
);
594 * Look for a page at index that is suitable for clustering.
599 unsigned int pg_offset
,
604 if (PageWriteback(page
))
607 if (page
->mapping
&& PageDirty(page
)) {
608 if (page_has_buffers(page
)) {
609 struct buffer_head
*bh
, *head
;
611 bh
= head
= page_buffers(page
);
613 if (!buffer_uptodate(bh
))
615 if (mapped
!= buffer_mapped(bh
))
618 if (ret
>= pg_offset
)
620 } while ((bh
= bh
->b_this_page
) != head
);
622 ret
= mapped
? 0 : PAGE_CACHE_SIZE
;
631 struct page
*startpage
,
632 struct buffer_head
*bh
,
633 struct buffer_head
*head
,
637 pgoff_t tindex
, tlast
, tloff
;
641 /* First sum forwards in this page */
643 if (!buffer_uptodate(bh
) || (mapped
!= buffer_mapped(bh
)))
646 } while ((bh
= bh
->b_this_page
) != head
);
648 /* if we reached the end of the page, sum forwards in following pages */
649 tlast
= i_size_read(inode
) >> PAGE_CACHE_SHIFT
;
650 tindex
= startpage
->index
+ 1;
652 /* Prune this back to avoid pathological behavior */
653 tloff
= min(tlast
, startpage
->index
+ 64);
655 pagevec_init(&pvec
, 0);
656 while (!done
&& tindex
<= tloff
) {
657 unsigned len
= min_t(pgoff_t
, PAGEVEC_SIZE
, tlast
- tindex
+ 1);
659 if (!pagevec_lookup(&pvec
, inode
->i_mapping
, tindex
, len
))
662 for (i
= 0; i
< pagevec_count(&pvec
); i
++) {
663 struct page
*page
= pvec
.pages
[i
];
664 size_t pg_offset
, pg_len
= 0;
666 if (tindex
== tlast
) {
668 i_size_read(inode
) & (PAGE_CACHE_SIZE
- 1);
674 pg_offset
= PAGE_CACHE_SIZE
;
676 if (page
->index
== tindex
&& trylock_page(page
)) {
677 pg_len
= xfs_probe_page(page
, pg_offset
, mapped
);
690 pagevec_release(&pvec
);
698 * Test if a given page is suitable for writing as part of an unwritten
699 * or delayed allocate extent.
706 if (PageWriteback(page
))
709 if (page
->mapping
&& page_has_buffers(page
)) {
710 struct buffer_head
*bh
, *head
;
713 bh
= head
= page_buffers(page
);
715 if (buffer_unwritten(bh
))
716 acceptable
= (type
== IOMAP_UNWRITTEN
);
717 else if (buffer_delay(bh
))
718 acceptable
= (type
== IOMAP_DELAY
);
719 else if (buffer_dirty(bh
) && buffer_mapped(bh
))
720 acceptable
= (type
== IOMAP_NEW
);
723 } while ((bh
= bh
->b_this_page
) != head
);
733 * Allocate & map buffers for page given the extent map. Write it out.
734 * except for the original page of a writepage, this is called on
735 * delalloc/unwritten pages only, for the original page it is possible
736 * that the page has no mapping at all.
744 xfs_ioend_t
**ioendp
,
745 struct writeback_control
*wbc
,
749 struct buffer_head
*bh
, *head
;
750 xfs_off_t end_offset
;
751 unsigned long p_offset
;
753 int bbits
= inode
->i_blkbits
;
755 int count
= 0, done
= 0, uptodate
= 1;
756 xfs_off_t offset
= page_offset(page
);
758 if (page
->index
!= tindex
)
760 if (!trylock_page(page
))
762 if (PageWriteback(page
))
763 goto fail_unlock_page
;
764 if (page
->mapping
!= inode
->i_mapping
)
765 goto fail_unlock_page
;
766 if (!xfs_is_delayed_page(page
, (*ioendp
)->io_type
))
767 goto fail_unlock_page
;
770 * page_dirty is initially a count of buffers on the page before
771 * EOF and is decremented as we move each into a cleanable state.
775 * End offset is the highest offset that this page should represent.
776 * If we are on the last page, (end_offset & (PAGE_CACHE_SIZE - 1))
777 * will evaluate non-zero and be less than PAGE_CACHE_SIZE and
778 * hence give us the correct page_dirty count. On any other page,
779 * it will be zero and in that case we need page_dirty to be the
780 * count of buffers on the page.
782 end_offset
= min_t(unsigned long long,
783 (xfs_off_t
)(page
->index
+ 1) << PAGE_CACHE_SHIFT
,
786 len
= 1 << inode
->i_blkbits
;
787 p_offset
= min_t(unsigned long, end_offset
& (PAGE_CACHE_SIZE
- 1),
789 p_offset
= p_offset
? roundup(p_offset
, len
) : PAGE_CACHE_SIZE
;
790 page_dirty
= p_offset
/ len
;
792 bh
= head
= page_buffers(page
);
794 if (offset
>= end_offset
)
796 if (!buffer_uptodate(bh
))
798 if (!(PageUptodate(page
) || buffer_uptodate(bh
))) {
803 if (buffer_unwritten(bh
) || buffer_delay(bh
)) {
804 if (buffer_unwritten(bh
))
805 type
= IOMAP_UNWRITTEN
;
809 if (!xfs_iomap_valid(mp
, offset
)) {
814 ASSERT(!(mp
->iomap_flags
& IOMAP_HOLE
));
815 ASSERT(!(mp
->iomap_flags
& IOMAP_DELAY
));
817 xfs_map_at_offset(bh
, offset
, bbits
, mp
);
819 xfs_add_to_ioend(inode
, bh
, offset
,
822 set_buffer_dirty(bh
);
824 mark_buffer_dirty(bh
);
830 if (buffer_mapped(bh
) && all_bh
&& startio
) {
832 xfs_add_to_ioend(inode
, bh
, offset
,
840 } while (offset
+= len
, (bh
= bh
->b_this_page
) != head
);
842 if (uptodate
&& bh
== head
)
843 SetPageUptodate(page
);
848 if (wbc
->nr_to_write
<= 0)
851 xfs_start_page_writeback(page
, !page_dirty
, count
);
862 * Convert & write out a cluster of pages in the same extent as defined
863 * by mp and following the start page.
870 xfs_ioend_t
**ioendp
,
871 struct writeback_control
*wbc
,
879 pagevec_init(&pvec
, 0);
880 while (!done
&& tindex
<= tlast
) {
881 unsigned len
= min_t(pgoff_t
, PAGEVEC_SIZE
, tlast
- tindex
+ 1);
883 if (!pagevec_lookup(&pvec
, inode
->i_mapping
, tindex
, len
))
886 for (i
= 0; i
< pagevec_count(&pvec
); i
++) {
887 done
= xfs_convert_page(inode
, pvec
.pages
[i
], tindex
++,
888 iomapp
, ioendp
, wbc
, startio
, all_bh
);
893 pagevec_release(&pvec
);
899 xfs_vm_invalidatepage(
901 unsigned long offset
)
903 trace_xfs_invalidatepage(page
->mapping
->host
, page
, offset
);
904 block_invalidatepage(page
, offset
);
908 * If the page has delalloc buffers on it, we need to punch them out before we
909 * invalidate the page. If we don't, we leave a stale delalloc mapping on the
910 * inode that can trip a BUG() in xfs_get_blocks() later on if a direct IO read
911 * is done on that same region - the delalloc extent is returned when none is
912 * supposed to be there.
914 * We prevent this by truncating away the delalloc regions on the page before
915 * invalidating it. Because they are delalloc, we can do this without needing a
916 * transaction. Indeed - if we get ENOSPC errors, we have to be able to do this
917 * truncation without a transaction as there is no space left for block
918 * reservation (typically why we see a ENOSPC in writeback).
920 * This is not a performance critical path, so for now just do the punching a
921 * buffer head at a time.
924 xfs_aops_discard_page(
927 struct inode
*inode
= page
->mapping
->host
;
928 struct xfs_inode
*ip
= XFS_I(inode
);
929 struct buffer_head
*bh
, *head
;
930 loff_t offset
= page_offset(page
);
931 ssize_t len
= 1 << inode
->i_blkbits
;
933 if (!xfs_is_delayed_page(page
, IOMAP_DELAY
))
936 if (XFS_FORCED_SHUTDOWN(ip
->i_mount
))
939 xfs_fs_cmn_err(CE_ALERT
, ip
->i_mount
,
940 "page discard on page %p, inode 0x%llx, offset %llu.",
941 page
, ip
->i_ino
, offset
);
943 xfs_ilock(ip
, XFS_ILOCK_EXCL
);
944 bh
= head
= page_buffers(page
);
947 xfs_fileoff_t offset_fsb
;
948 xfs_bmbt_irec_t imap
;
951 xfs_fsblock_t firstblock
;
952 xfs_bmap_free_t flist
;
954 if (!buffer_delay(bh
))
957 offset_fsb
= XFS_B_TO_FSBT(ip
->i_mount
, offset
);
960 * Map the range first and check that it is a delalloc extent
961 * before trying to unmap the range. Otherwise we will be
962 * trying to remove a real extent (which requires a
963 * transaction) or a hole, which is probably a bad idea...
965 error
= xfs_bmapi(NULL
, ip
, offset_fsb
, 1,
966 XFS_BMAPI_ENTIRE
, NULL
, 0, &imap
,
967 &nimaps
, NULL
, NULL
);
970 /* something screwed, just bail */
971 if (!XFS_FORCED_SHUTDOWN(ip
->i_mount
)) {
972 xfs_fs_cmn_err(CE_ALERT
, ip
->i_mount
,
973 "page discard failed delalloc mapping lookup.");
981 if (imap
.br_startblock
!= DELAYSTARTBLOCK
) {
982 /* been converted, ignore */
985 WARN_ON(imap
.br_blockcount
== 0);
988 * Note: while we initialise the firstblock/flist pair, they
989 * should never be used because blocks should never be
990 * allocated or freed for a delalloc extent and hence we need
991 * don't cancel or finish them after the xfs_bunmapi() call.
993 xfs_bmap_init(&flist
, &firstblock
);
994 error
= xfs_bunmapi(NULL
, ip
, offset_fsb
, 1, 0, 1, &firstblock
,
995 &flist
, NULL
, &done
);
997 ASSERT(!flist
.xbf_count
&& !flist
.xbf_first
);
999 /* something screwed, just bail */
1000 if (!XFS_FORCED_SHUTDOWN(ip
->i_mount
)) {
1001 xfs_fs_cmn_err(CE_ALERT
, ip
->i_mount
,
1002 "page discard unable to remove delalloc mapping.");
1009 } while ((bh
= bh
->b_this_page
) != head
);
1011 xfs_iunlock(ip
, XFS_ILOCK_EXCL
);
1013 xfs_vm_invalidatepage(page
, 0);
1018 * Calling this without startio set means we are being asked to make a dirty
1019 * page ready for freeing it's buffers. When called with startio set then
1020 * we are coming from writepage.
1022 * When called with startio set it is important that we write the WHOLE
1024 * The bh->b_state's cannot know if any of the blocks or which block for
1025 * that matter are dirty due to mmap writes, and therefore bh uptodate is
1026 * only valid if the page itself isn't completely uptodate. Some layers
1027 * may clear the page dirty flag prior to calling write page, under the
1028 * assumption the entire page will be written out; by not writing out the
1029 * whole page the page can be reused before all valid dirty data is
1030 * written out. Note: in the case of a page that has been dirty'd by
1031 * mapwrite and but partially setup by block_prepare_write the
1032 * bh->b_states's will not agree and only ones setup by BPW/BCW will have
1033 * valid state, thus the whole page must be written out thing.
1037 xfs_page_state_convert(
1038 struct inode
*inode
,
1040 struct writeback_control
*wbc
,
1042 int unmapped
) /* also implies page uptodate */
1044 struct buffer_head
*bh
, *head
;
1046 xfs_ioend_t
*ioend
= NULL
, *iohead
= NULL
;
1048 unsigned long p_offset
= 0;
1050 __uint64_t end_offset
;
1051 pgoff_t end_index
, last_index
, tlast
;
1053 int flags
, err
, iomap_valid
= 0, uptodate
= 1;
1054 int page_dirty
, count
= 0;
1056 int all_bh
= unmapped
;
1059 if (wbc
->sync_mode
== WB_SYNC_NONE
&& wbc
->nonblocking
)
1060 trylock
|= BMAPI_TRYLOCK
;
1063 /* Is this page beyond the end of the file? */
1064 offset
= i_size_read(inode
);
1065 end_index
= offset
>> PAGE_CACHE_SHIFT
;
1066 last_index
= (offset
- 1) >> PAGE_CACHE_SHIFT
;
1067 if (page
->index
>= end_index
) {
1068 if ((page
->index
>= end_index
+ 1) ||
1069 !(i_size_read(inode
) & (PAGE_CACHE_SIZE
- 1))) {
1077 * page_dirty is initially a count of buffers on the page before
1078 * EOF and is decremented as we move each into a cleanable state.
1082 * End offset is the highest offset that this page should represent.
1083 * If we are on the last page, (end_offset & (PAGE_CACHE_SIZE - 1))
1084 * will evaluate non-zero and be less than PAGE_CACHE_SIZE and
1085 * hence give us the correct page_dirty count. On any other page,
1086 * it will be zero and in that case we need page_dirty to be the
1087 * count of buffers on the page.
1089 end_offset
= min_t(unsigned long long,
1090 (xfs_off_t
)(page
->index
+ 1) << PAGE_CACHE_SHIFT
, offset
);
1091 len
= 1 << inode
->i_blkbits
;
1092 p_offset
= min_t(unsigned long, end_offset
& (PAGE_CACHE_SIZE
- 1),
1094 p_offset
= p_offset
? roundup(p_offset
, len
) : PAGE_CACHE_SIZE
;
1095 page_dirty
= p_offset
/ len
;
1097 bh
= head
= page_buffers(page
);
1098 offset
= page_offset(page
);
1102 /* TODO: cleanup count and page_dirty */
1105 if (offset
>= end_offset
)
1107 if (!buffer_uptodate(bh
))
1109 if (!(PageUptodate(page
) || buffer_uptodate(bh
)) && !startio
) {
1111 * the iomap is actually still valid, but the ioend
1112 * isn't. shouldn't happen too often.
1119 iomap_valid
= xfs_iomap_valid(&iomap
, offset
);
1122 * First case, map an unwritten extent and prepare for
1123 * extent state conversion transaction on completion.
1125 * Second case, allocate space for a delalloc buffer.
1126 * We can return EAGAIN here in the release page case.
1128 * Third case, an unmapped buffer was found, and we are
1129 * in a path where we need to write the whole page out.
1131 if (buffer_unwritten(bh
) || buffer_delay(bh
) ||
1132 ((buffer_uptodate(bh
) || PageUptodate(page
)) &&
1133 !buffer_mapped(bh
) && (unmapped
|| startio
))) {
1137 * Make sure we don't use a read-only iomap
1139 if (flags
== BMAPI_READ
)
1142 if (buffer_unwritten(bh
)) {
1143 type
= IOMAP_UNWRITTEN
;
1144 flags
= BMAPI_WRITE
| BMAPI_IGNSTATE
;
1145 } else if (buffer_delay(bh
)) {
1147 flags
= BMAPI_ALLOCATE
| trylock
;
1150 flags
= BMAPI_WRITE
| BMAPI_MMAP
;
1155 * if we didn't have a valid mapping then we
1156 * need to ensure that we put the new mapping
1157 * in a new ioend structure. This needs to be
1158 * done to ensure that the ioends correctly
1159 * reflect the block mappings at io completion
1160 * for unwritten extent conversion.
1163 if (type
== IOMAP_NEW
) {
1164 size
= xfs_probe_cluster(inode
,
1170 err
= xfs_map_blocks(inode
, offset
, size
,
1174 iomap_valid
= xfs_iomap_valid(&iomap
, offset
);
1177 xfs_map_at_offset(bh
, offset
,
1178 inode
->i_blkbits
, &iomap
);
1180 xfs_add_to_ioend(inode
, bh
, offset
,
1184 set_buffer_dirty(bh
);
1186 mark_buffer_dirty(bh
);
1191 } else if (buffer_uptodate(bh
) && startio
) {
1193 * we got here because the buffer is already mapped.
1194 * That means it must already have extents allocated
1195 * underneath it. Map the extent by reading it.
1197 if (!iomap_valid
|| flags
!= BMAPI_READ
) {
1199 size
= xfs_probe_cluster(inode
, page
, bh
,
1201 err
= xfs_map_blocks(inode
, offset
, size
,
1205 iomap_valid
= xfs_iomap_valid(&iomap
, offset
);
1209 * We set the type to IOMAP_NEW in case we are doing a
1210 * small write at EOF that is extending the file but
1211 * without needing an allocation. We need to update the
1212 * file size on I/O completion in this case so it is
1213 * the same case as having just allocated a new extent
1214 * that we are writing into for the first time.
1217 if (trylock_buffer(bh
)) {
1218 ASSERT(buffer_mapped(bh
));
1221 xfs_add_to_ioend(inode
, bh
, offset
, type
,
1222 &ioend
, !iomap_valid
);
1228 } else if ((buffer_uptodate(bh
) || PageUptodate(page
)) &&
1229 (unmapped
|| startio
)) {
1236 } while (offset
+= len
, ((bh
= bh
->b_this_page
) != head
));
1238 if (uptodate
&& bh
== head
)
1239 SetPageUptodate(page
);
1242 xfs_start_page_writeback(page
, 1, count
);
1244 if (ioend
&& iomap_valid
) {
1245 offset
= (iomap
.iomap_offset
+ iomap
.iomap_bsize
- 1) >>
1247 tlast
= min_t(pgoff_t
, offset
, last_index
);
1248 xfs_cluster_write(inode
, page
->index
+ 1, &iomap
, &ioend
,
1249 wbc
, startio
, all_bh
, tlast
);
1253 xfs_submit_ioend(wbc
, iohead
);
1259 xfs_cancel_ioend(iohead
);
1262 * If it's delalloc and we have nowhere to put it,
1263 * throw it away, unless the lower layers told
1266 if (err
!= -EAGAIN
) {
1268 xfs_aops_discard_page(page
);
1269 ClearPageUptodate(page
);
1275 * writepage: Called from one of two places:
1277 * 1. we are flushing a delalloc buffer head.
1279 * 2. we are writing out a dirty page. Typically the page dirty
1280 * state is cleared before we get here. In this case is it
1281 * conceivable we have no buffer heads.
1283 * For delalloc space on the page we need to allocate space and
1284 * flush it. For unmapped buffer heads on the page we should
1285 * allocate space if the page is uptodate. For any other dirty
1286 * buffer heads on the page we should flush them.
1288 * If we detect that a transaction would be required to flush
1289 * the page, we have to check the process flags first, if we
1290 * are already in a transaction or disk I/O during allocations
1291 * is off, we need to fail the writepage and redirty the page.
1297 struct writeback_control
*wbc
)
1301 int delalloc
, unmapped
, unwritten
;
1302 struct inode
*inode
= page
->mapping
->host
;
1304 trace_xfs_writepage(inode
, page
, 0);
1307 * We need a transaction if:
1308 * 1. There are delalloc buffers on the page
1309 * 2. The page is uptodate and we have unmapped buffers
1310 * 3. The page is uptodate and we have no buffers
1311 * 4. There are unwritten buffers on the page
1314 if (!page_has_buffers(page
)) {
1318 xfs_count_page_state(page
, &delalloc
, &unmapped
, &unwritten
);
1319 if (!PageUptodate(page
))
1321 need_trans
= delalloc
+ unmapped
+ unwritten
;
1325 * If we need a transaction and the process flags say
1326 * we are already in a transaction, or no IO is allowed
1327 * then mark the page dirty again and leave the page
1330 if (current_test_flags(PF_FSTRANS
) && need_trans
)
1334 * Delay hooking up buffer heads until we have
1335 * made our go/no-go decision.
1337 if (!page_has_buffers(page
))
1338 create_empty_buffers(page
, 1 << inode
->i_blkbits
, 0);
1342 * VM calculation for nr_to_write seems off. Bump it way
1343 * up, this gets simple streaming writes zippy again.
1344 * To be reviewed again after Jens' writeback changes.
1346 wbc
->nr_to_write
*= 4;
1349 * Convert delayed allocate, unwritten or unmapped space
1350 * to real space and flush out to disk.
1352 error
= xfs_page_state_convert(inode
, page
, wbc
, 1, unmapped
);
1353 if (error
== -EAGAIN
)
1355 if (unlikely(error
< 0))
1361 redirty_page_for_writepage(wbc
, page
);
1371 struct address_space
*mapping
,
1372 struct writeback_control
*wbc
)
1374 xfs_iflags_clear(XFS_I(mapping
->host
), XFS_ITRUNCATED
);
1375 return generic_writepages(mapping
, wbc
);
1379 * Called to move a page into cleanable state - and from there
1380 * to be released. Possibly the page is already clean. We always
1381 * have buffer heads in this call.
1383 * Returns 0 if the page is ok to release, 1 otherwise.
1385 * Possible scenarios are:
1387 * 1. We are being called to release a page which has been written
1388 * to via regular I/O. buffer heads will be dirty and possibly
1389 * delalloc. If no delalloc buffer heads in this case then we
1390 * can just return zero.
1392 * 2. We are called to release a page which has been written via
1393 * mmap, all we need to do is ensure there is no delalloc
1394 * state in the buffer heads, if not we can let the caller
1395 * free them and we should come back later via writepage.
1402 struct inode
*inode
= page
->mapping
->host
;
1403 int dirty
, delalloc
, unmapped
, unwritten
;
1404 struct writeback_control wbc
= {
1405 .sync_mode
= WB_SYNC_ALL
,
1409 trace_xfs_releasepage(inode
, page
, 0);
1411 if (!page_has_buffers(page
))
1414 xfs_count_page_state(page
, &delalloc
, &unmapped
, &unwritten
);
1415 if (!delalloc
&& !unwritten
)
1418 if (!(gfp_mask
& __GFP_FS
))
1421 /* If we are already inside a transaction or the thread cannot
1422 * do I/O, we cannot release this page.
1424 if (current_test_flags(PF_FSTRANS
))
1428 * Convert delalloc space to real space, do not flush the
1429 * data out to disk, that will be done by the caller.
1430 * Never need to allocate space here - we will always
1431 * come back to writepage in that case.
1433 dirty
= xfs_page_state_convert(inode
, page
, &wbc
, 0, 0);
1434 if (dirty
== 0 && !unwritten
)
1439 return try_to_free_buffers(page
);
1444 struct inode
*inode
,
1446 struct buffer_head
*bh_result
,
1449 bmapi_flags_t flags
)
1457 offset
= (xfs_off_t
)iblock
<< inode
->i_blkbits
;
1458 ASSERT(bh_result
->b_size
>= (1 << inode
->i_blkbits
));
1459 size
= bh_result
->b_size
;
1461 if (!create
&& direct
&& offset
>= i_size_read(inode
))
1464 error
= xfs_iomap(XFS_I(inode
), offset
, size
,
1465 create
? flags
: BMAPI_READ
, &iomap
, &niomap
);
1471 if (iomap
.iomap_bn
!= IOMAP_DADDR_NULL
) {
1473 * For unwritten extents do not report a disk address on
1474 * the read case (treat as if we're reading into a hole).
1476 if (create
|| !(iomap
.iomap_flags
& IOMAP_UNWRITTEN
)) {
1477 xfs_map_buffer(bh_result
, &iomap
, offset
,
1480 if (create
&& (iomap
.iomap_flags
& IOMAP_UNWRITTEN
)) {
1482 bh_result
->b_private
= inode
;
1483 set_buffer_unwritten(bh_result
);
1488 * If this is a realtime file, data may be on a different device.
1489 * to that pointed to from the buffer_head b_bdev currently.
1491 bh_result
->b_bdev
= iomap
.iomap_target
->bt_bdev
;
1494 * If we previously allocated a block out beyond eof and we are now
1495 * coming back to use it then we will need to flag it as new even if it
1496 * has a disk address.
1498 * With sub-block writes into unwritten extents we also need to mark
1499 * the buffer as new so that the unwritten parts of the buffer gets
1503 ((!buffer_mapped(bh_result
) && !buffer_uptodate(bh_result
)) ||
1504 (offset
>= i_size_read(inode
)) ||
1505 (iomap
.iomap_flags
& (IOMAP_NEW
|IOMAP_UNWRITTEN
))))
1506 set_buffer_new(bh_result
);
1508 if (iomap
.iomap_flags
& IOMAP_DELAY
) {
1511 set_buffer_uptodate(bh_result
);
1512 set_buffer_mapped(bh_result
);
1513 set_buffer_delay(bh_result
);
1517 if (direct
|| size
> (1 << inode
->i_blkbits
)) {
1518 ASSERT(iomap
.iomap_bsize
- iomap
.iomap_delta
> 0);
1519 offset
= min_t(xfs_off_t
,
1520 iomap
.iomap_bsize
- iomap
.iomap_delta
, size
);
1521 bh_result
->b_size
= (ssize_t
)min_t(xfs_off_t
, LONG_MAX
, offset
);
1529 struct inode
*inode
,
1531 struct buffer_head
*bh_result
,
1534 return __xfs_get_blocks(inode
, iblock
,
1535 bh_result
, create
, 0, BMAPI_WRITE
);
1539 xfs_get_blocks_direct(
1540 struct inode
*inode
,
1542 struct buffer_head
*bh_result
,
1545 return __xfs_get_blocks(inode
, iblock
,
1546 bh_result
, create
, 1, BMAPI_WRITE
|BMAPI_DIRECT
);
1556 xfs_ioend_t
*ioend
= iocb
->private;
1559 * Non-NULL private data means we need to issue a transaction to
1560 * convert a range from unwritten to written extents. This needs
1561 * to happen from process context but aio+dio I/O completion
1562 * happens from irq context so we need to defer it to a workqueue.
1563 * This is not necessary for synchronous direct I/O, but we do
1564 * it anyway to keep the code uniform and simpler.
1566 * Well, if only it were that simple. Because synchronous direct I/O
1567 * requires extent conversion to occur *before* we return to userspace,
1568 * we have to wait for extent conversion to complete. Look at the
1569 * iocb that has been passed to us to determine if this is AIO or
1570 * not. If it is synchronous, tell xfs_finish_ioend() to kick the
1571 * workqueue and wait for it to complete.
1573 * The core direct I/O code might be changed to always call the
1574 * completion handler in the future, in which case all this can
1577 ioend
->io_offset
= offset
;
1578 ioend
->io_size
= size
;
1579 if (ioend
->io_type
== IOMAP_READ
) {
1580 xfs_finish_ioend(ioend
, 0);
1581 } else if (private && size
> 0) {
1582 xfs_finish_ioend(ioend
, is_sync_kiocb(iocb
));
1585 * A direct I/O write ioend starts it's life in unwritten
1586 * state in case they map an unwritten extent. This write
1587 * didn't map an unwritten extent so switch it's completion
1590 ioend
->io_type
= IOMAP_NEW
;
1591 xfs_finish_ioend(ioend
, 0);
1595 * blockdev_direct_IO can return an error even after the I/O
1596 * completion handler was called. Thus we need to protect
1597 * against double-freeing.
1599 iocb
->private = NULL
;
1606 const struct iovec
*iov
,
1608 unsigned long nr_segs
)
1610 struct file
*file
= iocb
->ki_filp
;
1611 struct inode
*inode
= file
->f_mapping
->host
;
1612 struct block_device
*bdev
;
1615 bdev
= xfs_find_bdev_for_inode(XFS_I(inode
));
1617 iocb
->private = xfs_alloc_ioend(inode
, rw
== WRITE
?
1618 IOMAP_UNWRITTEN
: IOMAP_READ
);
1620 ret
= blockdev_direct_IO_no_locking(rw
, iocb
, inode
, bdev
, iov
,
1622 xfs_get_blocks_direct
,
1625 if (unlikely(ret
!= -EIOCBQUEUED
&& iocb
->private))
1626 xfs_destroy_ioend(iocb
->private);
1633 struct address_space
*mapping
,
1637 struct page
**pagep
,
1641 return block_write_begin(file
, mapping
, pos
, len
, flags
, pagep
, fsdata
,
1647 struct address_space
*mapping
,
1650 struct inode
*inode
= (struct inode
*)mapping
->host
;
1651 struct xfs_inode
*ip
= XFS_I(inode
);
1653 xfs_itrace_entry(XFS_I(inode
));
1654 xfs_ilock(ip
, XFS_IOLOCK_SHARED
);
1655 xfs_flush_pages(ip
, (xfs_off_t
)0, -1, 0, FI_REMAPF
);
1656 xfs_iunlock(ip
, XFS_IOLOCK_SHARED
);
1657 return generic_block_bmap(mapping
, block
, xfs_get_blocks
);
1662 struct file
*unused
,
1665 return mpage_readpage(page
, xfs_get_blocks
);
1670 struct file
*unused
,
1671 struct address_space
*mapping
,
1672 struct list_head
*pages
,
1675 return mpage_readpages(mapping
, pages
, nr_pages
, xfs_get_blocks
);
1678 const struct address_space_operations xfs_address_space_operations
= {
1679 .readpage
= xfs_vm_readpage
,
1680 .readpages
= xfs_vm_readpages
,
1681 .writepage
= xfs_vm_writepage
,
1682 .writepages
= xfs_vm_writepages
,
1683 .sync_page
= block_sync_page
,
1684 .releasepage
= xfs_vm_releasepage
,
1685 .invalidatepage
= xfs_vm_invalidatepage
,
1686 .write_begin
= xfs_vm_write_begin
,
1687 .write_end
= generic_write_end
,
1688 .bmap
= xfs_vm_bmap
,
1689 .direct_IO
= xfs_vm_direct_IO
,
1690 .migratepage
= buffer_migrate_page
,
1691 .is_partially_uptodate
= block_is_partially_uptodate
,
1692 .error_remove_page
= generic_error_remove_page
,