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
22 #include "xfs_trans.h"
23 #include "xfs_mount.h"
24 #include "xfs_bmap_btree.h"
25 #include "xfs_dinode.h"
26 #include "xfs_inode.h"
27 #include "xfs_inode_item.h"
28 #include "xfs_alloc.h"
29 #include "xfs_error.h"
30 #include "xfs_iomap.h"
31 #include "xfs_vnodeops.h"
32 #include "xfs_trace.h"
34 #include <linux/gfp.h>
35 #include <linux/mpage.h>
36 #include <linux/pagevec.h>
37 #include <linux/writeback.h>
45 struct buffer_head
*bh
, *head
;
47 *delalloc
= *unwritten
= 0;
49 bh
= head
= page_buffers(page
);
51 if (buffer_unwritten(bh
))
53 else if (buffer_delay(bh
))
55 } while ((bh
= bh
->b_this_page
) != head
);
58 STATIC
struct block_device
*
59 xfs_find_bdev_for_inode(
62 struct xfs_inode
*ip
= XFS_I(inode
);
63 struct xfs_mount
*mp
= ip
->i_mount
;
65 if (XFS_IS_REALTIME_INODE(ip
))
66 return mp
->m_rtdev_targp
->bt_bdev
;
68 return mp
->m_ddev_targp
->bt_bdev
;
72 * We're now finished for good with this ioend structure.
73 * Update the page state via the associated buffer_heads,
74 * release holds on the inode and bio, and finally free
75 * up memory. Do not use the ioend after this.
81 struct buffer_head
*bh
, *next
;
83 for (bh
= ioend
->io_buffer_head
; bh
; bh
= next
) {
85 bh
->b_end_io(bh
, !ioend
->io_error
);
89 if (ioend
->io_isasync
) {
90 aio_complete(ioend
->io_iocb
, ioend
->io_error
?
91 ioend
->io_error
: ioend
->io_result
, 0);
93 inode_dio_done(ioend
->io_inode
);
96 mempool_free(ioend
, xfs_ioend_pool
);
100 * Fast and loose check if this write could update the on-disk inode size.
102 static inline bool xfs_ioend_is_append(struct xfs_ioend
*ioend
)
104 return ioend
->io_offset
+ ioend
->io_size
>
105 XFS_I(ioend
->io_inode
)->i_d
.di_size
;
109 xfs_setfilesize_trans_alloc(
110 struct xfs_ioend
*ioend
)
112 struct xfs_mount
*mp
= XFS_I(ioend
->io_inode
)->i_mount
;
113 struct xfs_trans
*tp
;
116 tp
= xfs_trans_alloc(mp
, XFS_TRANS_FSYNC_TS
);
118 error
= xfs_trans_reserve(tp
, 0, XFS_FSYNC_TS_LOG_RES(mp
), 0, 0, 0);
120 xfs_trans_cancel(tp
, 0);
124 ioend
->io_append_trans
= tp
;
127 * We may pass freeze protection with a transaction. So tell lockdep
130 rwsem_release(&ioend
->io_inode
->i_sb
->s_writers
.lock_map
[SB_FREEZE_FS
-1],
133 * We hand off the transaction to the completion thread now, so
134 * clear the flag here.
136 current_restore_flags_nested(&tp
->t_pflags
, PF_FSTRANS
);
141 * Update on-disk file size now that data has been written to disk.
145 struct xfs_ioend
*ioend
)
147 struct xfs_inode
*ip
= XFS_I(ioend
->io_inode
);
148 struct xfs_trans
*tp
= ioend
->io_append_trans
;
152 * The transaction may have been allocated in the I/O submission thread,
153 * thus we need to mark ourselves as beeing in a transaction manually.
154 * Similarly for freeze protection.
156 current_set_flags_nested(&tp
->t_pflags
, PF_FSTRANS
);
157 rwsem_acquire_read(&VFS_I(ip
)->i_sb
->s_writers
.lock_map
[SB_FREEZE_FS
-1],
160 xfs_ilock(ip
, XFS_ILOCK_EXCL
);
161 isize
= xfs_new_eof(ip
, ioend
->io_offset
+ ioend
->io_size
);
163 xfs_iunlock(ip
, XFS_ILOCK_EXCL
);
164 xfs_trans_cancel(tp
, 0);
168 trace_xfs_setfilesize(ip
, ioend
->io_offset
, ioend
->io_size
);
170 ip
->i_d
.di_size
= isize
;
171 xfs_trans_ijoin(tp
, ip
, XFS_ILOCK_EXCL
);
172 xfs_trans_log_inode(tp
, ip
, XFS_ILOG_CORE
);
174 return xfs_trans_commit(tp
, 0);
178 * Schedule IO completion handling on the final put of an ioend.
180 * If there is no work to do we might as well call it a day and free the
185 struct xfs_ioend
*ioend
)
187 if (atomic_dec_and_test(&ioend
->io_remaining
)) {
188 struct xfs_mount
*mp
= XFS_I(ioend
->io_inode
)->i_mount
;
190 if (ioend
->io_type
== XFS_IO_UNWRITTEN
)
191 queue_work(mp
->m_unwritten_workqueue
, &ioend
->io_work
);
192 else if (ioend
->io_append_trans
||
193 (ioend
->io_isdirect
&& xfs_ioend_is_append(ioend
)))
194 queue_work(mp
->m_data_workqueue
, &ioend
->io_work
);
196 xfs_destroy_ioend(ioend
);
201 * IO write completion.
205 struct work_struct
*work
)
207 xfs_ioend_t
*ioend
= container_of(work
, xfs_ioend_t
, io_work
);
208 struct xfs_inode
*ip
= XFS_I(ioend
->io_inode
);
211 if (XFS_FORCED_SHUTDOWN(ip
->i_mount
)) {
212 ioend
->io_error
= -EIO
;
219 * For unwritten extents we need to issue transactions to convert a
220 * range to normal written extens after the data I/O has finished.
222 if (ioend
->io_type
== XFS_IO_UNWRITTEN
) {
223 error
= xfs_iomap_write_unwritten(ip
, ioend
->io_offset
,
225 } else if (ioend
->io_isdirect
&& xfs_ioend_is_append(ioend
)) {
227 * For direct I/O we do not know if we need to allocate blocks
228 * or not so we can't preallocate an append transaction as that
229 * results in nested reservations and log space deadlocks. Hence
230 * allocate the transaction here. While this is sub-optimal and
231 * can block IO completion for some time, we're stuck with doing
232 * it this way until we can pass the ioend to the direct IO
233 * allocation callbacks and avoid nesting that way.
235 error
= xfs_setfilesize_trans_alloc(ioend
);
238 error
= xfs_setfilesize(ioend
);
239 } else if (ioend
->io_append_trans
) {
240 error
= xfs_setfilesize(ioend
);
242 ASSERT(!xfs_ioend_is_append(ioend
));
247 ioend
->io_error
= -error
;
248 xfs_destroy_ioend(ioend
);
252 * Call IO completion handling in caller context on the final put of an ioend.
255 xfs_finish_ioend_sync(
256 struct xfs_ioend
*ioend
)
258 if (atomic_dec_and_test(&ioend
->io_remaining
))
259 xfs_end_io(&ioend
->io_work
);
263 * Allocate and initialise an IO completion structure.
264 * We need to track unwritten extent write completion here initially.
265 * We'll need to extend this for updating the ondisk inode size later
275 ioend
= mempool_alloc(xfs_ioend_pool
, GFP_NOFS
);
278 * Set the count to 1 initially, which will prevent an I/O
279 * completion callback from happening before we have started
280 * all the I/O from calling the completion routine too early.
282 atomic_set(&ioend
->io_remaining
, 1);
283 ioend
->io_isasync
= 0;
284 ioend
->io_isdirect
= 0;
286 ioend
->io_list
= NULL
;
287 ioend
->io_type
= type
;
288 ioend
->io_inode
= inode
;
289 ioend
->io_buffer_head
= NULL
;
290 ioend
->io_buffer_tail
= NULL
;
291 ioend
->io_offset
= 0;
293 ioend
->io_iocb
= NULL
;
294 ioend
->io_result
= 0;
295 ioend
->io_append_trans
= NULL
;
297 INIT_WORK(&ioend
->io_work
, xfs_end_io
);
305 struct xfs_bmbt_irec
*imap
,
309 struct xfs_inode
*ip
= XFS_I(inode
);
310 struct xfs_mount
*mp
= ip
->i_mount
;
311 ssize_t count
= 1 << inode
->i_blkbits
;
312 xfs_fileoff_t offset_fsb
, end_fsb
;
314 int bmapi_flags
= XFS_BMAPI_ENTIRE
;
317 if (XFS_FORCED_SHUTDOWN(mp
))
318 return -XFS_ERROR(EIO
);
320 if (type
== XFS_IO_UNWRITTEN
)
321 bmapi_flags
|= XFS_BMAPI_IGSTATE
;
323 if (!xfs_ilock_nowait(ip
, XFS_ILOCK_SHARED
)) {
325 return -XFS_ERROR(EAGAIN
);
326 xfs_ilock(ip
, XFS_ILOCK_SHARED
);
329 ASSERT(ip
->i_d
.di_format
!= XFS_DINODE_FMT_BTREE
||
330 (ip
->i_df
.if_flags
& XFS_IFEXTENTS
));
331 ASSERT(offset
<= mp
->m_super
->s_maxbytes
);
333 if (offset
+ count
> mp
->m_super
->s_maxbytes
)
334 count
= mp
->m_super
->s_maxbytes
- offset
;
335 end_fsb
= XFS_B_TO_FSB(mp
, (xfs_ufsize_t
)offset
+ count
);
336 offset_fsb
= XFS_B_TO_FSBT(mp
, offset
);
337 error
= xfs_bmapi_read(ip
, offset_fsb
, end_fsb
- offset_fsb
,
338 imap
, &nimaps
, bmapi_flags
);
339 xfs_iunlock(ip
, XFS_ILOCK_SHARED
);
342 return -XFS_ERROR(error
);
344 if (type
== XFS_IO_DELALLOC
&&
345 (!nimaps
|| isnullstartblock(imap
->br_startblock
))) {
346 error
= xfs_iomap_write_allocate(ip
, offset
, count
, imap
);
348 trace_xfs_map_blocks_alloc(ip
, offset
, count
, type
, imap
);
349 return -XFS_ERROR(error
);
353 if (type
== XFS_IO_UNWRITTEN
) {
355 ASSERT(imap
->br_startblock
!= HOLESTARTBLOCK
);
356 ASSERT(imap
->br_startblock
!= DELAYSTARTBLOCK
);
360 trace_xfs_map_blocks_found(ip
, offset
, count
, type
, imap
);
367 struct xfs_bmbt_irec
*imap
,
370 offset
>>= inode
->i_blkbits
;
372 return offset
>= imap
->br_startoff
&&
373 offset
< imap
->br_startoff
+ imap
->br_blockcount
;
377 * BIO completion handler for buffered IO.
384 xfs_ioend_t
*ioend
= bio
->bi_private
;
386 ASSERT(atomic_read(&bio
->bi_cnt
) >= 1);
387 ioend
->io_error
= test_bit(BIO_UPTODATE
, &bio
->bi_flags
) ? 0 : error
;
389 /* Toss bio and pass work off to an xfsdatad thread */
390 bio
->bi_private
= NULL
;
391 bio
->bi_end_io
= NULL
;
394 xfs_finish_ioend(ioend
);
398 xfs_submit_ioend_bio(
399 struct writeback_control
*wbc
,
403 atomic_inc(&ioend
->io_remaining
);
404 bio
->bi_private
= ioend
;
405 bio
->bi_end_io
= xfs_end_bio
;
406 submit_bio(wbc
->sync_mode
== WB_SYNC_ALL
? WRITE_SYNC
: WRITE
, bio
);
411 struct buffer_head
*bh
)
413 int nvecs
= bio_get_nr_vecs(bh
->b_bdev
);
414 struct bio
*bio
= bio_alloc(GFP_NOIO
, nvecs
);
416 ASSERT(bio
->bi_private
== NULL
);
417 bio
->bi_sector
= bh
->b_blocknr
* (bh
->b_size
>> 9);
418 bio
->bi_bdev
= bh
->b_bdev
;
423 xfs_start_buffer_writeback(
424 struct buffer_head
*bh
)
426 ASSERT(buffer_mapped(bh
));
427 ASSERT(buffer_locked(bh
));
428 ASSERT(!buffer_delay(bh
));
429 ASSERT(!buffer_unwritten(bh
));
431 mark_buffer_async_write(bh
);
432 set_buffer_uptodate(bh
);
433 clear_buffer_dirty(bh
);
437 xfs_start_page_writeback(
442 ASSERT(PageLocked(page
));
443 ASSERT(!PageWriteback(page
));
445 clear_page_dirty_for_io(page
);
446 set_page_writeback(page
);
448 /* If no buffers on the page are to be written, finish it here */
450 end_page_writeback(page
);
453 static inline int bio_add_buffer(struct bio
*bio
, struct buffer_head
*bh
)
455 return bio_add_page(bio
, bh
->b_page
, bh
->b_size
, bh_offset(bh
));
459 * Submit all of the bios for all of the ioends we have saved up, covering the
460 * initial writepage page and also any probed pages.
462 * Because we may have multiple ioends spanning a page, we need to start
463 * writeback on all the buffers before we submit them for I/O. If we mark the
464 * buffers as we got, then we can end up with a page that only has buffers
465 * marked async write and I/O complete on can occur before we mark the other
466 * buffers async write.
468 * The end result of this is that we trip a bug in end_page_writeback() because
469 * we call it twice for the one page as the code in end_buffer_async_write()
470 * assumes that all buffers on the page are started at the same time.
472 * The fix is two passes across the ioend list - one to start writeback on the
473 * buffer_heads, and then submit them for I/O on the second pass.
475 * If @fail is non-zero, it means that we have a situation where some part of
476 * the submission process has failed after we have marked paged for writeback
477 * and unlocked them. In this situation, we need to fail the ioend chain rather
478 * than submit it to IO. This typically only happens on a filesystem shutdown.
482 struct writeback_control
*wbc
,
486 xfs_ioend_t
*head
= ioend
;
488 struct buffer_head
*bh
;
490 sector_t lastblock
= 0;
492 /* Pass 1 - start writeback */
494 next
= ioend
->io_list
;
495 for (bh
= ioend
->io_buffer_head
; bh
; bh
= bh
->b_private
)
496 xfs_start_buffer_writeback(bh
);
497 } while ((ioend
= next
) != NULL
);
499 /* Pass 2 - submit I/O */
502 next
= ioend
->io_list
;
506 * If we are failing the IO now, just mark the ioend with an
507 * error and finish it. This will run IO completion immediately
508 * as there is only one reference to the ioend at this point in
512 ioend
->io_error
= -fail
;
513 xfs_finish_ioend(ioend
);
517 for (bh
= ioend
->io_buffer_head
; bh
; bh
= bh
->b_private
) {
521 bio
= xfs_alloc_ioend_bio(bh
);
522 } else if (bh
->b_blocknr
!= lastblock
+ 1) {
523 xfs_submit_ioend_bio(wbc
, ioend
, bio
);
527 if (bio_add_buffer(bio
, bh
) != bh
->b_size
) {
528 xfs_submit_ioend_bio(wbc
, ioend
, bio
);
532 lastblock
= bh
->b_blocknr
;
535 xfs_submit_ioend_bio(wbc
, ioend
, bio
);
536 xfs_finish_ioend(ioend
);
537 } while ((ioend
= next
) != NULL
);
541 * Cancel submission of all buffer_heads so far in this endio.
542 * Toss the endio too. Only ever called for the initial page
543 * in a writepage request, so only ever one page.
550 struct buffer_head
*bh
, *next_bh
;
553 next
= ioend
->io_list
;
554 bh
= ioend
->io_buffer_head
;
556 next_bh
= bh
->b_private
;
557 clear_buffer_async_write(bh
);
559 } while ((bh
= next_bh
) != NULL
);
561 mempool_free(ioend
, xfs_ioend_pool
);
562 } while ((ioend
= next
) != NULL
);
566 * Test to see if we've been building up a completion structure for
567 * earlier buffers -- if so, we try to append to this ioend if we
568 * can, otherwise we finish off any current ioend and start another.
569 * Return true if we've finished the given ioend.
574 struct buffer_head
*bh
,
577 xfs_ioend_t
**result
,
580 xfs_ioend_t
*ioend
= *result
;
582 if (!ioend
|| need_ioend
|| type
!= ioend
->io_type
) {
583 xfs_ioend_t
*previous
= *result
;
585 ioend
= xfs_alloc_ioend(inode
, type
);
586 ioend
->io_offset
= offset
;
587 ioend
->io_buffer_head
= bh
;
588 ioend
->io_buffer_tail
= bh
;
590 previous
->io_list
= ioend
;
593 ioend
->io_buffer_tail
->b_private
= bh
;
594 ioend
->io_buffer_tail
= bh
;
597 bh
->b_private
= NULL
;
598 ioend
->io_size
+= bh
->b_size
;
604 struct buffer_head
*bh
,
605 struct xfs_bmbt_irec
*imap
,
609 struct xfs_mount
*m
= XFS_I(inode
)->i_mount
;
610 xfs_off_t iomap_offset
= XFS_FSB_TO_B(m
, imap
->br_startoff
);
611 xfs_daddr_t iomap_bn
= xfs_fsb_to_db(XFS_I(inode
), imap
->br_startblock
);
613 ASSERT(imap
->br_startblock
!= HOLESTARTBLOCK
);
614 ASSERT(imap
->br_startblock
!= DELAYSTARTBLOCK
);
616 bn
= (iomap_bn
>> (inode
->i_blkbits
- BBSHIFT
)) +
617 ((offset
- iomap_offset
) >> inode
->i_blkbits
);
619 ASSERT(bn
|| XFS_IS_REALTIME_INODE(XFS_I(inode
)));
622 set_buffer_mapped(bh
);
628 struct buffer_head
*bh
,
629 struct xfs_bmbt_irec
*imap
,
632 ASSERT(imap
->br_startblock
!= HOLESTARTBLOCK
);
633 ASSERT(imap
->br_startblock
!= DELAYSTARTBLOCK
);
635 xfs_map_buffer(inode
, bh
, imap
, offset
);
636 set_buffer_mapped(bh
);
637 clear_buffer_delay(bh
);
638 clear_buffer_unwritten(bh
);
642 * Test if a given page is suitable for writing as part of an unwritten
643 * or delayed allocate extent.
650 if (PageWriteback(page
))
653 if (page
->mapping
&& page_has_buffers(page
)) {
654 struct buffer_head
*bh
, *head
;
657 bh
= head
= page_buffers(page
);
659 if (buffer_unwritten(bh
))
660 acceptable
+= (type
== XFS_IO_UNWRITTEN
);
661 else if (buffer_delay(bh
))
662 acceptable
+= (type
== XFS_IO_DELALLOC
);
663 else if (buffer_dirty(bh
) && buffer_mapped(bh
))
664 acceptable
+= (type
== XFS_IO_OVERWRITE
);
667 } while ((bh
= bh
->b_this_page
) != head
);
677 * Allocate & map buffers for page given the extent map. Write it out.
678 * except for the original page of a writepage, this is called on
679 * delalloc/unwritten pages only, for the original page it is possible
680 * that the page has no mapping at all.
687 struct xfs_bmbt_irec
*imap
,
688 xfs_ioend_t
**ioendp
,
689 struct writeback_control
*wbc
)
691 struct buffer_head
*bh
, *head
;
692 xfs_off_t end_offset
;
693 unsigned long p_offset
;
696 int count
= 0, done
= 0, uptodate
= 1;
697 xfs_off_t offset
= page_offset(page
);
699 if (page
->index
!= tindex
)
701 if (!trylock_page(page
))
703 if (PageWriteback(page
))
704 goto fail_unlock_page
;
705 if (page
->mapping
!= inode
->i_mapping
)
706 goto fail_unlock_page
;
707 if (!xfs_check_page_type(page
, (*ioendp
)->io_type
))
708 goto fail_unlock_page
;
711 * page_dirty is initially a count of buffers on the page before
712 * EOF and is decremented as we move each into a cleanable state.
716 * End offset is the highest offset that this page should represent.
717 * If we are on the last page, (end_offset & (PAGE_CACHE_SIZE - 1))
718 * will evaluate non-zero and be less than PAGE_CACHE_SIZE and
719 * hence give us the correct page_dirty count. On any other page,
720 * it will be zero and in that case we need page_dirty to be the
721 * count of buffers on the page.
723 end_offset
= min_t(unsigned long long,
724 (xfs_off_t
)(page
->index
+ 1) << PAGE_CACHE_SHIFT
,
727 len
= 1 << inode
->i_blkbits
;
728 p_offset
= min_t(unsigned long, end_offset
& (PAGE_CACHE_SIZE
- 1),
730 p_offset
= p_offset
? roundup(p_offset
, len
) : PAGE_CACHE_SIZE
;
731 page_dirty
= p_offset
/ len
;
733 bh
= head
= page_buffers(page
);
735 if (offset
>= end_offset
)
737 if (!buffer_uptodate(bh
))
739 if (!(PageUptodate(page
) || buffer_uptodate(bh
))) {
744 if (buffer_unwritten(bh
) || buffer_delay(bh
) ||
746 if (buffer_unwritten(bh
))
747 type
= XFS_IO_UNWRITTEN
;
748 else if (buffer_delay(bh
))
749 type
= XFS_IO_DELALLOC
;
751 type
= XFS_IO_OVERWRITE
;
753 if (!xfs_imap_valid(inode
, imap
, offset
)) {
759 if (type
!= XFS_IO_OVERWRITE
)
760 xfs_map_at_offset(inode
, bh
, imap
, offset
);
761 xfs_add_to_ioend(inode
, bh
, offset
, type
,
769 } while (offset
+= len
, (bh
= bh
->b_this_page
) != head
);
771 if (uptodate
&& bh
== head
)
772 SetPageUptodate(page
);
775 if (--wbc
->nr_to_write
<= 0 &&
776 wbc
->sync_mode
== WB_SYNC_NONE
)
779 xfs_start_page_writeback(page
, !page_dirty
, count
);
789 * Convert & write out a cluster of pages in the same extent as defined
790 * by mp and following the start page.
796 struct xfs_bmbt_irec
*imap
,
797 xfs_ioend_t
**ioendp
,
798 struct writeback_control
*wbc
,
804 pagevec_init(&pvec
, 0);
805 while (!done
&& tindex
<= tlast
) {
806 unsigned len
= min_t(pgoff_t
, PAGEVEC_SIZE
, tlast
- tindex
+ 1);
808 if (!pagevec_lookup(&pvec
, inode
->i_mapping
, tindex
, len
))
811 for (i
= 0; i
< pagevec_count(&pvec
); i
++) {
812 done
= xfs_convert_page(inode
, pvec
.pages
[i
], tindex
++,
818 pagevec_release(&pvec
);
824 xfs_vm_invalidatepage(
826 unsigned long offset
)
828 trace_xfs_invalidatepage(page
->mapping
->host
, page
, offset
);
829 block_invalidatepage(page
, offset
);
833 * If the page has delalloc buffers on it, we need to punch them out before we
834 * invalidate the page. If we don't, we leave a stale delalloc mapping on the
835 * inode that can trip a BUG() in xfs_get_blocks() later on if a direct IO read
836 * is done on that same region - the delalloc extent is returned when none is
837 * supposed to be there.
839 * We prevent this by truncating away the delalloc regions on the page before
840 * invalidating it. Because they are delalloc, we can do this without needing a
841 * transaction. Indeed - if we get ENOSPC errors, we have to be able to do this
842 * truncation without a transaction as there is no space left for block
843 * reservation (typically why we see a ENOSPC in writeback).
845 * This is not a performance critical path, so for now just do the punching a
846 * buffer head at a time.
849 xfs_aops_discard_page(
852 struct inode
*inode
= page
->mapping
->host
;
853 struct xfs_inode
*ip
= XFS_I(inode
);
854 struct buffer_head
*bh
, *head
;
855 loff_t offset
= page_offset(page
);
857 if (!xfs_check_page_type(page
, XFS_IO_DELALLOC
))
860 if (XFS_FORCED_SHUTDOWN(ip
->i_mount
))
863 xfs_alert(ip
->i_mount
,
864 "page discard on page %p, inode 0x%llx, offset %llu.",
865 page
, ip
->i_ino
, offset
);
867 xfs_ilock(ip
, XFS_ILOCK_EXCL
);
868 bh
= head
= page_buffers(page
);
871 xfs_fileoff_t start_fsb
;
873 if (!buffer_delay(bh
))
876 start_fsb
= XFS_B_TO_FSBT(ip
->i_mount
, offset
);
877 error
= xfs_bmap_punch_delalloc_range(ip
, start_fsb
, 1);
879 /* something screwed, just bail */
880 if (!XFS_FORCED_SHUTDOWN(ip
->i_mount
)) {
881 xfs_alert(ip
->i_mount
,
882 "page discard unable to remove delalloc mapping.");
887 offset
+= 1 << inode
->i_blkbits
;
889 } while ((bh
= bh
->b_this_page
) != head
);
891 xfs_iunlock(ip
, XFS_ILOCK_EXCL
);
893 xfs_vm_invalidatepage(page
, 0);
898 * Write out a dirty page.
900 * For delalloc space on the page we need to allocate space and flush it.
901 * For unwritten space on the page we need to start the conversion to
902 * regular allocated space.
903 * For any other dirty buffer heads on the page we should flush them.
908 struct writeback_control
*wbc
)
910 struct inode
*inode
= page
->mapping
->host
;
911 struct buffer_head
*bh
, *head
;
912 struct xfs_bmbt_irec imap
;
913 xfs_ioend_t
*ioend
= NULL
, *iohead
= NULL
;
916 __uint64_t end_offset
;
917 pgoff_t end_index
, last_index
;
919 int err
, imap_valid
= 0, uptodate
= 1;
923 trace_xfs_writepage(inode
, page
, 0);
925 ASSERT(page_has_buffers(page
));
928 * Refuse to write the page out if we are called from reclaim context.
930 * This avoids stack overflows when called from deeply used stacks in
931 * random callers for direct reclaim or memcg reclaim. We explicitly
932 * allow reclaim from kswapd as the stack usage there is relatively low.
934 * This should never happen except in the case of a VM regression so
937 if (WARN_ON_ONCE((current
->flags
& (PF_MEMALLOC
|PF_KSWAPD
)) ==
942 * Given that we do not allow direct reclaim to call us, we should
943 * never be called while in a filesystem transaction.
945 if (WARN_ON(current
->flags
& PF_FSTRANS
))
948 /* Is this page beyond the end of the file? */
949 offset
= i_size_read(inode
);
950 end_index
= offset
>> PAGE_CACHE_SHIFT
;
951 last_index
= (offset
- 1) >> PAGE_CACHE_SHIFT
;
952 if (page
->index
>= end_index
) {
953 unsigned offset_into_page
= offset
& (PAGE_CACHE_SIZE
- 1);
956 * Just skip the page if it is fully outside i_size, e.g. due
957 * to a truncate operation that is in progress.
959 if (page
->index
>= end_index
+ 1 || offset_into_page
== 0) {
965 * The page straddles i_size. It must be zeroed out on each
966 * and every writepage invocation because it may be mmapped.
967 * "A file is mapped in multiples of the page size. For a file
968 * that is not a multiple of the page size, the remaining
969 * memory is zeroed when mapped, and writes to that region are
970 * not written out to the file."
972 zero_user_segment(page
, offset_into_page
, PAGE_CACHE_SIZE
);
975 end_offset
= min_t(unsigned long long,
976 (xfs_off_t
)(page
->index
+ 1) << PAGE_CACHE_SHIFT
,
978 len
= 1 << inode
->i_blkbits
;
980 bh
= head
= page_buffers(page
);
981 offset
= page_offset(page
);
982 type
= XFS_IO_OVERWRITE
;
984 if (wbc
->sync_mode
== WB_SYNC_NONE
)
990 if (offset
>= end_offset
)
992 if (!buffer_uptodate(bh
))
996 * set_page_dirty dirties all buffers in a page, independent
997 * of their state. The dirty state however is entirely
998 * meaningless for holes (!mapped && uptodate), so skip
999 * buffers covering holes here.
1001 if (!buffer_mapped(bh
) && buffer_uptodate(bh
)) {
1006 if (buffer_unwritten(bh
)) {
1007 if (type
!= XFS_IO_UNWRITTEN
) {
1008 type
= XFS_IO_UNWRITTEN
;
1011 } else if (buffer_delay(bh
)) {
1012 if (type
!= XFS_IO_DELALLOC
) {
1013 type
= XFS_IO_DELALLOC
;
1016 } else if (buffer_uptodate(bh
)) {
1017 if (type
!= XFS_IO_OVERWRITE
) {
1018 type
= XFS_IO_OVERWRITE
;
1022 if (PageUptodate(page
))
1023 ASSERT(buffer_mapped(bh
));
1025 * This buffer is not uptodate and will not be
1026 * written to disk. Ensure that we will put any
1027 * subsequent writeable buffers into a new
1035 imap_valid
= xfs_imap_valid(inode
, &imap
, offset
);
1038 * If we didn't have a valid mapping then we need to
1039 * put the new mapping into a separate ioend structure.
1040 * This ensures non-contiguous extents always have
1041 * separate ioends, which is particularly important
1042 * for unwritten extent conversion at I/O completion
1046 err
= xfs_map_blocks(inode
, offset
, &imap
, type
,
1050 imap_valid
= xfs_imap_valid(inode
, &imap
, offset
);
1054 if (type
!= XFS_IO_OVERWRITE
)
1055 xfs_map_at_offset(inode
, bh
, &imap
, offset
);
1056 xfs_add_to_ioend(inode
, bh
, offset
, type
, &ioend
,
1064 } while (offset
+= len
, ((bh
= bh
->b_this_page
) != head
));
1066 if (uptodate
&& bh
== head
)
1067 SetPageUptodate(page
);
1069 xfs_start_page_writeback(page
, 1, count
);
1071 /* if there is no IO to be submitted for this page, we are done */
1078 * Any errors from this point onwards need tobe reported through the IO
1079 * completion path as we have marked the initial page as under writeback
1083 xfs_off_t end_index
;
1085 end_index
= imap
.br_startoff
+ imap
.br_blockcount
;
1088 end_index
<<= inode
->i_blkbits
;
1091 end_index
= (end_index
- 1) >> PAGE_CACHE_SHIFT
;
1093 /* check against file size */
1094 if (end_index
> last_index
)
1095 end_index
= last_index
;
1097 xfs_cluster_write(inode
, page
->index
+ 1, &imap
, &ioend
,
1103 * Reserve log space if we might write beyond the on-disk inode size.
1106 if (ioend
->io_type
!= XFS_IO_UNWRITTEN
&& xfs_ioend_is_append(ioend
))
1107 err
= xfs_setfilesize_trans_alloc(ioend
);
1109 xfs_submit_ioend(wbc
, iohead
, err
);
1115 xfs_cancel_ioend(iohead
);
1120 xfs_aops_discard_page(page
);
1121 ClearPageUptodate(page
);
1126 redirty_page_for_writepage(wbc
, page
);
1133 struct address_space
*mapping
,
1134 struct writeback_control
*wbc
)
1136 xfs_iflags_clear(XFS_I(mapping
->host
), XFS_ITRUNCATED
);
1137 return generic_writepages(mapping
, wbc
);
1141 * Called to move a page into cleanable state - and from there
1142 * to be released. The page should already be clean. We always
1143 * have buffer heads in this call.
1145 * Returns 1 if the page is ok to release, 0 otherwise.
1152 int delalloc
, unwritten
;
1154 trace_xfs_releasepage(page
->mapping
->host
, page
, 0);
1156 xfs_count_page_state(page
, &delalloc
, &unwritten
);
1158 if (WARN_ON(delalloc
))
1160 if (WARN_ON(unwritten
))
1163 return try_to_free_buffers(page
);
1168 struct inode
*inode
,
1170 struct buffer_head
*bh_result
,
1174 struct xfs_inode
*ip
= XFS_I(inode
);
1175 struct xfs_mount
*mp
= ip
->i_mount
;
1176 xfs_fileoff_t offset_fsb
, end_fsb
;
1179 struct xfs_bmbt_irec imap
;
1185 if (XFS_FORCED_SHUTDOWN(mp
))
1186 return -XFS_ERROR(EIO
);
1188 offset
= (xfs_off_t
)iblock
<< inode
->i_blkbits
;
1189 ASSERT(bh_result
->b_size
>= (1 << inode
->i_blkbits
));
1190 size
= bh_result
->b_size
;
1192 if (!create
&& direct
&& offset
>= i_size_read(inode
))
1196 * Direct I/O is usually done on preallocated files, so try getting
1197 * a block mapping without an exclusive lock first. For buffered
1198 * writes we already have the exclusive iolock anyway, so avoiding
1199 * a lock roundtrip here by taking the ilock exclusive from the
1200 * beginning is a useful micro optimization.
1202 if (create
&& !direct
) {
1203 lockmode
= XFS_ILOCK_EXCL
;
1204 xfs_ilock(ip
, lockmode
);
1206 lockmode
= xfs_ilock_map_shared(ip
);
1209 ASSERT(offset
<= mp
->m_super
->s_maxbytes
);
1210 if (offset
+ size
> mp
->m_super
->s_maxbytes
)
1211 size
= mp
->m_super
->s_maxbytes
- offset
;
1212 end_fsb
= XFS_B_TO_FSB(mp
, (xfs_ufsize_t
)offset
+ size
);
1213 offset_fsb
= XFS_B_TO_FSBT(mp
, offset
);
1215 error
= xfs_bmapi_read(ip
, offset_fsb
, end_fsb
- offset_fsb
,
1216 &imap
, &nimaps
, XFS_BMAPI_ENTIRE
);
1222 (imap
.br_startblock
== HOLESTARTBLOCK
||
1223 imap
.br_startblock
== DELAYSTARTBLOCK
))) {
1224 if (direct
|| xfs_get_extsz_hint(ip
)) {
1226 * Drop the ilock in preparation for starting the block
1227 * allocation transaction. It will be retaken
1228 * exclusively inside xfs_iomap_write_direct for the
1229 * actual allocation.
1231 xfs_iunlock(ip
, lockmode
);
1232 error
= xfs_iomap_write_direct(ip
, offset
, size
,
1239 * Delalloc reservations do not require a transaction,
1240 * we can go on without dropping the lock here. If we
1241 * are allocating a new delalloc block, make sure that
1242 * we set the new flag so that we mark the buffer new so
1243 * that we know that it is newly allocated if the write
1246 if (nimaps
&& imap
.br_startblock
== HOLESTARTBLOCK
)
1248 error
= xfs_iomap_write_delay(ip
, offset
, size
, &imap
);
1252 xfs_iunlock(ip
, lockmode
);
1255 trace_xfs_get_blocks_alloc(ip
, offset
, size
, 0, &imap
);
1256 } else if (nimaps
) {
1257 trace_xfs_get_blocks_found(ip
, offset
, size
, 0, &imap
);
1258 xfs_iunlock(ip
, lockmode
);
1260 trace_xfs_get_blocks_notfound(ip
, offset
, size
);
1264 if (imap
.br_startblock
!= HOLESTARTBLOCK
&&
1265 imap
.br_startblock
!= DELAYSTARTBLOCK
) {
1267 * For unwritten extents do not report a disk address on
1268 * the read case (treat as if we're reading into a hole).
1270 if (create
|| !ISUNWRITTEN(&imap
))
1271 xfs_map_buffer(inode
, bh_result
, &imap
, offset
);
1272 if (create
&& ISUNWRITTEN(&imap
)) {
1274 bh_result
->b_private
= inode
;
1275 set_buffer_unwritten(bh_result
);
1280 * If this is a realtime file, data may be on a different device.
1281 * to that pointed to from the buffer_head b_bdev currently.
1283 bh_result
->b_bdev
= xfs_find_bdev_for_inode(inode
);
1286 * If we previously allocated a block out beyond eof and we are now
1287 * coming back to use it then we will need to flag it as new even if it
1288 * has a disk address.
1290 * With sub-block writes into unwritten extents we also need to mark
1291 * the buffer as new so that the unwritten parts of the buffer gets
1295 ((!buffer_mapped(bh_result
) && !buffer_uptodate(bh_result
)) ||
1296 (offset
>= i_size_read(inode
)) ||
1297 (new || ISUNWRITTEN(&imap
))))
1298 set_buffer_new(bh_result
);
1300 if (imap
.br_startblock
== DELAYSTARTBLOCK
) {
1303 set_buffer_uptodate(bh_result
);
1304 set_buffer_mapped(bh_result
);
1305 set_buffer_delay(bh_result
);
1310 * If this is O_DIRECT or the mpage code calling tell them how large
1311 * the mapping is, so that we can avoid repeated get_blocks calls.
1313 if (direct
|| size
> (1 << inode
->i_blkbits
)) {
1314 xfs_off_t mapping_size
;
1316 mapping_size
= imap
.br_startoff
+ imap
.br_blockcount
- iblock
;
1317 mapping_size
<<= inode
->i_blkbits
;
1319 ASSERT(mapping_size
> 0);
1320 if (mapping_size
> size
)
1321 mapping_size
= size
;
1322 if (mapping_size
> LONG_MAX
)
1323 mapping_size
= LONG_MAX
;
1325 bh_result
->b_size
= mapping_size
;
1331 xfs_iunlock(ip
, lockmode
);
1337 struct inode
*inode
,
1339 struct buffer_head
*bh_result
,
1342 return __xfs_get_blocks(inode
, iblock
, bh_result
, create
, 0);
1346 xfs_get_blocks_direct(
1347 struct inode
*inode
,
1349 struct buffer_head
*bh_result
,
1352 return __xfs_get_blocks(inode
, iblock
, bh_result
, create
, 1);
1356 * Complete a direct I/O write request.
1358 * If the private argument is non-NULL __xfs_get_blocks signals us that we
1359 * need to issue a transaction to convert the range from unwritten to written
1360 * extents. In case this is regular synchronous I/O we just call xfs_end_io
1361 * to do this and we are done. But in case this was a successful AIO
1362 * request this handler is called from interrupt context, from which we
1363 * can't start transactions. In that case offload the I/O completion to
1364 * the workqueues we also use for buffered I/O completion.
1367 xfs_end_io_direct_write(
1375 struct xfs_ioend
*ioend
= iocb
->private;
1378 * While the generic direct I/O code updates the inode size, it does
1379 * so only after the end_io handler is called, which means our
1380 * end_io handler thinks the on-disk size is outside the in-core
1381 * size. To prevent this just update it a little bit earlier here.
1383 if (offset
+ size
> i_size_read(ioend
->io_inode
))
1384 i_size_write(ioend
->io_inode
, offset
+ size
);
1387 * blockdev_direct_IO can return an error even after the I/O
1388 * completion handler was called. Thus we need to protect
1389 * against double-freeing.
1391 iocb
->private = NULL
;
1393 ioend
->io_offset
= offset
;
1394 ioend
->io_size
= size
;
1395 ioend
->io_iocb
= iocb
;
1396 ioend
->io_result
= ret
;
1397 if (private && size
> 0)
1398 ioend
->io_type
= XFS_IO_UNWRITTEN
;
1401 ioend
->io_isasync
= 1;
1402 xfs_finish_ioend(ioend
);
1404 xfs_finish_ioend_sync(ioend
);
1412 const struct iovec
*iov
,
1414 unsigned long nr_segs
)
1416 struct inode
*inode
= iocb
->ki_filp
->f_mapping
->host
;
1417 struct block_device
*bdev
= xfs_find_bdev_for_inode(inode
);
1418 struct xfs_ioend
*ioend
= NULL
;
1422 size_t size
= iov_length(iov
, nr_segs
);
1425 * We cannot preallocate a size update transaction here as we
1426 * don't know whether allocation is necessary or not. Hence we
1427 * can only tell IO completion that one is necessary if we are
1428 * not doing unwritten extent conversion.
1430 iocb
->private = ioend
= xfs_alloc_ioend(inode
, XFS_IO_DIRECT
);
1431 if (offset
+ size
> XFS_I(inode
)->i_d
.di_size
)
1432 ioend
->io_isdirect
= 1;
1434 ret
= __blockdev_direct_IO(rw
, iocb
, inode
, bdev
, iov
,
1436 xfs_get_blocks_direct
,
1437 xfs_end_io_direct_write
, NULL
, 0);
1438 if (ret
!= -EIOCBQUEUED
&& iocb
->private)
1439 goto out_destroy_ioend
;
1441 ret
= __blockdev_direct_IO(rw
, iocb
, inode
, bdev
, iov
,
1443 xfs_get_blocks_direct
,
1450 xfs_destroy_ioend(ioend
);
1455 * Punch out the delalloc blocks we have already allocated.
1457 * Don't bother with xfs_setattr given that nothing can have made it to disk yet
1458 * as the page is still locked at this point.
1461 xfs_vm_kill_delalloc_range(
1462 struct inode
*inode
,
1466 struct xfs_inode
*ip
= XFS_I(inode
);
1467 xfs_fileoff_t start_fsb
;
1468 xfs_fileoff_t end_fsb
;
1471 start_fsb
= XFS_B_TO_FSB(ip
->i_mount
, start
);
1472 end_fsb
= XFS_B_TO_FSB(ip
->i_mount
, end
);
1473 if (end_fsb
<= start_fsb
)
1476 xfs_ilock(ip
, XFS_ILOCK_EXCL
);
1477 error
= xfs_bmap_punch_delalloc_range(ip
, start_fsb
,
1478 end_fsb
- start_fsb
);
1480 /* something screwed, just bail */
1481 if (!XFS_FORCED_SHUTDOWN(ip
->i_mount
)) {
1482 xfs_alert(ip
->i_mount
,
1483 "xfs_vm_write_failed: unable to clean up ino %lld",
1487 xfs_iunlock(ip
, XFS_ILOCK_EXCL
);
1491 xfs_vm_write_failed(
1492 struct inode
*inode
,
1497 loff_t block_offset
= pos
& PAGE_MASK
;
1500 loff_t from
= pos
& (PAGE_CACHE_SIZE
- 1);
1501 loff_t to
= from
+ len
;
1502 struct buffer_head
*bh
, *head
;
1504 ASSERT(block_offset
+ from
== pos
);
1506 head
= page_buffers(page
);
1508 for (bh
= head
; bh
!= head
|| !block_start
;
1509 bh
= bh
->b_this_page
, block_start
= block_end
,
1510 block_offset
+= bh
->b_size
) {
1511 block_end
= block_start
+ bh
->b_size
;
1513 /* skip buffers before the write */
1514 if (block_end
<= from
)
1517 /* if the buffer is after the write, we're done */
1518 if (block_start
>= to
)
1521 if (!buffer_delay(bh
))
1524 if (!buffer_new(bh
) && block_offset
< i_size_read(inode
))
1527 xfs_vm_kill_delalloc_range(inode
, block_offset
,
1528 block_offset
+ bh
->b_size
);
1534 * This used to call block_write_begin(), but it unlocks and releases the page
1535 * on error, and we need that page to be able to punch stale delalloc blocks out
1536 * on failure. hence we copy-n-waste it here and call xfs_vm_write_failed() at
1537 * the appropriate point.
1542 struct address_space
*mapping
,
1546 struct page
**pagep
,
1549 pgoff_t index
= pos
>> PAGE_CACHE_SHIFT
;
1553 ASSERT(len
<= PAGE_CACHE_SIZE
);
1555 page
= grab_cache_page_write_begin(mapping
, index
,
1556 flags
| AOP_FLAG_NOFS
);
1560 status
= __block_write_begin(page
, pos
, len
, xfs_get_blocks
);
1561 if (unlikely(status
)) {
1562 struct inode
*inode
= mapping
->host
;
1564 xfs_vm_write_failed(inode
, page
, pos
, len
);
1567 if (pos
+ len
> i_size_read(inode
))
1568 truncate_pagecache(inode
, pos
+ len
, i_size_read(inode
));
1570 page_cache_release(page
);
1579 * On failure, we only need to kill delalloc blocks beyond EOF because they
1580 * will never be written. For blocks within EOF, generic_write_end() zeros them
1581 * so they are safe to leave alone and be written with all the other valid data.
1586 struct address_space
*mapping
,
1595 ASSERT(len
<= PAGE_CACHE_SIZE
);
1597 ret
= generic_write_end(file
, mapping
, pos
, len
, copied
, page
, fsdata
);
1598 if (unlikely(ret
< len
)) {
1599 struct inode
*inode
= mapping
->host
;
1600 size_t isize
= i_size_read(inode
);
1601 loff_t to
= pos
+ len
;
1604 truncate_pagecache(inode
, to
, isize
);
1605 xfs_vm_kill_delalloc_range(inode
, isize
, to
);
1613 struct address_space
*mapping
,
1616 struct inode
*inode
= (struct inode
*)mapping
->host
;
1617 struct xfs_inode
*ip
= XFS_I(inode
);
1619 trace_xfs_vm_bmap(XFS_I(inode
));
1620 xfs_ilock(ip
, XFS_IOLOCK_SHARED
);
1621 filemap_write_and_wait(mapping
);
1622 xfs_iunlock(ip
, XFS_IOLOCK_SHARED
);
1623 return generic_block_bmap(mapping
, block
, xfs_get_blocks
);
1628 struct file
*unused
,
1631 return mpage_readpage(page
, xfs_get_blocks
);
1636 struct file
*unused
,
1637 struct address_space
*mapping
,
1638 struct list_head
*pages
,
1641 return mpage_readpages(mapping
, pages
, nr_pages
, xfs_get_blocks
);
1644 const struct address_space_operations xfs_address_space_operations
= {
1645 .readpage
= xfs_vm_readpage
,
1646 .readpages
= xfs_vm_readpages
,
1647 .writepage
= xfs_vm_writepage
,
1648 .writepages
= xfs_vm_writepages
,
1649 .releasepage
= xfs_vm_releasepage
,
1650 .invalidatepage
= xfs_vm_invalidatepage
,
1651 .write_begin
= xfs_vm_write_begin
,
1652 .write_end
= xfs_vm_write_end
,
1653 .bmap
= xfs_vm_bmap
,
1654 .direct_IO
= xfs_vm_direct_IO
,
1655 .migratepage
= buffer_migrate_page
,
1656 .is_partially_uptodate
= block_is_partially_uptodate
,
1657 .error_remove_page
= generic_error_remove_page
,