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 hand off the transaction to the completion thread now, so
128 * clear the flag here.
130 current_restore_flags_nested(&tp->t_pflags, PF_FSTRANS);
135 * Update on-disk file size now that data has been written to disk.
139 struct xfs_ioend *ioend)
141 struct xfs_inode *ip = XFS_I(ioend->io_inode);
142 struct xfs_trans *tp = ioend->io_append_trans;
146 * The transaction was allocated in the I/O submission thread,
147 * thus we need to mark ourselves as beeing in a transaction
150 current_set_flags_nested(&tp->t_pflags, PF_FSTRANS);
152 xfs_ilock(ip, XFS_ILOCK_EXCL);
153 isize = xfs_new_eof(ip, ioend->io_offset + ioend->io_size);
155 xfs_iunlock(ip, XFS_ILOCK_EXCL);
156 xfs_trans_cancel(tp, 0);
160 trace_xfs_setfilesize(ip, ioend->io_offset, ioend->io_size);
162 ip->i_d.di_size = isize;
163 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
164 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
166 return xfs_trans_commit(tp, 0);
170 * Schedule IO completion handling on the final put of an ioend.
172 * If there is no work to do we might as well call it a day and free the
177 struct xfs_ioend *ioend)
179 if (atomic_dec_and_test(&ioend->io_remaining)) {
180 struct xfs_mount *mp = XFS_I(ioend->io_inode)->i_mount;
182 if (ioend->io_type == IO_UNWRITTEN)
183 queue_work(mp->m_unwritten_workqueue, &ioend->io_work);
184 else if (ioend->io_append_trans)
185 queue_work(mp->m_data_workqueue, &ioend->io_work);
187 xfs_destroy_ioend(ioend);
192 * IO write completion.
196 struct work_struct *work)
198 xfs_ioend_t *ioend = container_of(work, xfs_ioend_t, io_work);
199 struct xfs_inode *ip = XFS_I(ioend->io_inode);
202 if (XFS_FORCED_SHUTDOWN(ip->i_mount)) {
203 ioend->io_error = -EIO;
210 * For unwritten extents we need to issue transactions to convert a
211 * range to normal written extens after the data I/O has finished.
213 if (ioend->io_type == IO_UNWRITTEN) {
215 * For buffered I/O we never preallocate a transaction when
216 * doing the unwritten extent conversion, but for direct I/O
217 * we do not know if we are converting an unwritten extent
218 * or not at the point where we preallocate the transaction.
220 if (ioend->io_append_trans) {
221 ASSERT(ioend->io_isdirect);
223 current_set_flags_nested(
224 &ioend->io_append_trans->t_pflags, PF_FSTRANS);
225 xfs_trans_cancel(ioend->io_append_trans, 0);
228 error = xfs_iomap_write_unwritten(ip, ioend->io_offset,
231 ioend->io_error = -error;
234 } else if (ioend->io_append_trans) {
235 error = xfs_setfilesize(ioend);
237 ioend->io_error = -error;
239 ASSERT(!xfs_ioend_is_append(ioend));
243 xfs_destroy_ioend(ioend);
247 * Call IO completion handling in caller context on the final put of an ioend.
250 xfs_finish_ioend_sync(
251 struct xfs_ioend *ioend)
253 if (atomic_dec_and_test(&ioend->io_remaining))
254 xfs_end_io(&ioend->io_work);
258 * Allocate and initialise an IO completion structure.
259 * We need to track unwritten extent write completion here initially.
260 * We'll need to extend this for updating the ondisk inode size later
270 ioend = mempool_alloc(xfs_ioend_pool, GFP_NOFS);
273 * Set the count to 1 initially, which will prevent an I/O
274 * completion callback from happening before we have started
275 * all the I/O from calling the completion routine too early.
277 atomic_set(&ioend->io_remaining, 1);
278 ioend->io_isasync = 0;
279 ioend->io_isdirect = 0;
281 ioend->io_list = NULL;
282 ioend->io_type = type;
283 ioend->io_inode = inode;
284 ioend->io_buffer_head = NULL;
285 ioend->io_buffer_tail = NULL;
286 ioend->io_offset = 0;
288 ioend->io_iocb = NULL;
289 ioend->io_result = 0;
290 ioend->io_append_trans = NULL;
292 INIT_WORK(&ioend->io_work, xfs_end_io);
300 struct xfs_bmbt_irec *imap,
304 struct xfs_inode *ip = XFS_I(inode);
305 struct xfs_mount *mp = ip->i_mount;
306 ssize_t count = 1 << inode->i_blkbits;
307 xfs_fileoff_t offset_fsb, end_fsb;
309 int bmapi_flags = XFS_BMAPI_ENTIRE;
312 if (XFS_FORCED_SHUTDOWN(mp))
313 return -XFS_ERROR(EIO);
315 if (type == IO_UNWRITTEN)
316 bmapi_flags |= XFS_BMAPI_IGSTATE;
318 if (!xfs_ilock_nowait(ip, XFS_ILOCK_SHARED)) {
320 return -XFS_ERROR(EAGAIN);
321 xfs_ilock(ip, XFS_ILOCK_SHARED);
324 ASSERT(ip->i_d.di_format != XFS_DINODE_FMT_BTREE ||
325 (ip->i_df.if_flags & XFS_IFEXTENTS));
326 ASSERT(offset <= mp->m_maxioffset);
328 if (offset + count > mp->m_maxioffset)
329 count = mp->m_maxioffset - offset;
330 end_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)offset + count);
331 offset_fsb = XFS_B_TO_FSBT(mp, offset);
332 error = xfs_bmapi_read(ip, offset_fsb, end_fsb - offset_fsb,
333 imap, &nimaps, bmapi_flags);
334 xfs_iunlock(ip, XFS_ILOCK_SHARED);
337 return -XFS_ERROR(error);
339 if (type == IO_DELALLOC &&
340 (!nimaps || isnullstartblock(imap->br_startblock))) {
341 error = xfs_iomap_write_allocate(ip, offset, count, imap);
343 trace_xfs_map_blocks_alloc(ip, offset, count, type, imap);
344 return -XFS_ERROR(error);
348 if (type == IO_UNWRITTEN) {
350 ASSERT(imap->br_startblock != HOLESTARTBLOCK);
351 ASSERT(imap->br_startblock != DELAYSTARTBLOCK);
355 trace_xfs_map_blocks_found(ip, offset, count, type, imap);
362 struct xfs_bmbt_irec *imap,
365 offset >>= inode->i_blkbits;
367 return offset >= imap->br_startoff &&
368 offset < imap->br_startoff + imap->br_blockcount;
372 * BIO completion handler for buffered IO.
379 xfs_ioend_t *ioend = bio->bi_private;
381 ASSERT(atomic_read(&bio->bi_cnt) >= 1);
382 ioend->io_error = test_bit(BIO_UPTODATE, &bio->bi_flags) ? 0 : error;
384 /* Toss bio and pass work off to an xfsdatad thread */
385 bio->bi_private = NULL;
386 bio->bi_end_io = NULL;
389 xfs_finish_ioend(ioend);
393 xfs_submit_ioend_bio(
394 struct writeback_control *wbc,
398 atomic_inc(&ioend->io_remaining);
399 bio->bi_private = ioend;
400 bio->bi_end_io = xfs_end_bio;
401 submit_bio(wbc->sync_mode == WB_SYNC_ALL ? WRITE_SYNC : WRITE, bio);
406 struct buffer_head *bh)
408 int nvecs = bio_get_nr_vecs(bh->b_bdev);
409 struct bio *bio = bio_alloc(GFP_NOIO, nvecs);
411 ASSERT(bio->bi_private == NULL);
412 bio->bi_sector = bh->b_blocknr * (bh->b_size >> 9);
413 bio->bi_bdev = bh->b_bdev;
418 xfs_start_buffer_writeback(
419 struct buffer_head *bh)
421 ASSERT(buffer_mapped(bh));
422 ASSERT(buffer_locked(bh));
423 ASSERT(!buffer_delay(bh));
424 ASSERT(!buffer_unwritten(bh));
426 mark_buffer_async_write(bh);
427 set_buffer_uptodate(bh);
428 clear_buffer_dirty(bh);
432 xfs_start_page_writeback(
437 ASSERT(PageLocked(page));
438 ASSERT(!PageWriteback(page));
440 clear_page_dirty_for_io(page);
441 set_page_writeback(page);
443 /* If no buffers on the page are to be written, finish it here */
445 end_page_writeback(page);
448 static inline int bio_add_buffer(struct bio *bio, struct buffer_head *bh)
450 return bio_add_page(bio, bh->b_page, bh->b_size, bh_offset(bh));
454 * Submit all of the bios for all of the ioends we have saved up, covering the
455 * initial writepage page and also any probed pages.
457 * Because we may have multiple ioends spanning a page, we need to start
458 * writeback on all the buffers before we submit them for I/O. If we mark the
459 * buffers as we got, then we can end up with a page that only has buffers
460 * marked async write and I/O complete on can occur before we mark the other
461 * buffers async write.
463 * The end result of this is that we trip a bug in end_page_writeback() because
464 * we call it twice for the one page as the code in end_buffer_async_write()
465 * assumes that all buffers on the page are started at the same time.
467 * The fix is two passes across the ioend list - one to start writeback on the
468 * buffer_heads, and then submit them for I/O on the second pass.
472 struct writeback_control *wbc,
475 xfs_ioend_t *head = ioend;
477 struct buffer_head *bh;
479 sector_t lastblock = 0;
481 /* Pass 1 - start writeback */
483 next = ioend->io_list;
484 for (bh = ioend->io_buffer_head; bh; bh = bh->b_private)
485 xfs_start_buffer_writeback(bh);
486 } while ((ioend = next) != NULL);
488 /* Pass 2 - submit I/O */
491 next = ioend->io_list;
494 for (bh = ioend->io_buffer_head; bh; bh = bh->b_private) {
498 bio = xfs_alloc_ioend_bio(bh);
499 } else if (bh->b_blocknr != lastblock + 1) {
500 xfs_submit_ioend_bio(wbc, ioend, bio);
504 if (bio_add_buffer(bio, bh) != bh->b_size) {
505 xfs_submit_ioend_bio(wbc, ioend, bio);
509 lastblock = bh->b_blocknr;
512 xfs_submit_ioend_bio(wbc, ioend, bio);
513 xfs_finish_ioend(ioend);
514 } while ((ioend = next) != NULL);
518 * Cancel submission of all buffer_heads so far in this endio.
519 * Toss the endio too. Only ever called for the initial page
520 * in a writepage request, so only ever one page.
527 struct buffer_head *bh, *next_bh;
530 next = ioend->io_list;
531 bh = ioend->io_buffer_head;
533 next_bh = bh->b_private;
534 clear_buffer_async_write(bh);
536 } while ((bh = next_bh) != NULL);
538 mempool_free(ioend, xfs_ioend_pool);
539 } while ((ioend = next) != NULL);
543 * Test to see if we've been building up a completion structure for
544 * earlier buffers -- if so, we try to append to this ioend if we
545 * can, otherwise we finish off any current ioend and start another.
546 * Return true if we've finished the given ioend.
551 struct buffer_head *bh,
554 xfs_ioend_t **result,
557 xfs_ioend_t *ioend = *result;
559 if (!ioend || need_ioend || type != ioend->io_type) {
560 xfs_ioend_t *previous = *result;
562 ioend = xfs_alloc_ioend(inode, type);
563 ioend->io_offset = offset;
564 ioend->io_buffer_head = bh;
565 ioend->io_buffer_tail = bh;
567 previous->io_list = ioend;
570 ioend->io_buffer_tail->b_private = bh;
571 ioend->io_buffer_tail = bh;
574 bh->b_private = NULL;
575 ioend->io_size += bh->b_size;
581 struct buffer_head *bh,
582 struct xfs_bmbt_irec *imap,
586 struct xfs_mount *m = XFS_I(inode)->i_mount;
587 xfs_off_t iomap_offset = XFS_FSB_TO_B(m, imap->br_startoff);
588 xfs_daddr_t iomap_bn = xfs_fsb_to_db(XFS_I(inode), imap->br_startblock);
590 ASSERT(imap->br_startblock != HOLESTARTBLOCK);
591 ASSERT(imap->br_startblock != DELAYSTARTBLOCK);
593 bn = (iomap_bn >> (inode->i_blkbits - BBSHIFT)) +
594 ((offset - iomap_offset) >> inode->i_blkbits);
596 ASSERT(bn || XFS_IS_REALTIME_INODE(XFS_I(inode)));
599 set_buffer_mapped(bh);
605 struct buffer_head *bh,
606 struct xfs_bmbt_irec *imap,
609 ASSERT(imap->br_startblock != HOLESTARTBLOCK);
610 ASSERT(imap->br_startblock != DELAYSTARTBLOCK);
612 xfs_map_buffer(inode, bh, imap, offset);
613 set_buffer_mapped(bh);
614 clear_buffer_delay(bh);
615 clear_buffer_unwritten(bh);
619 * Test if a given page is suitable for writing as part of an unwritten
620 * or delayed allocate extent.
627 if (PageWriteback(page))
630 if (page->mapping && page_has_buffers(page)) {
631 struct buffer_head *bh, *head;
634 bh = head = page_buffers(page);
636 if (buffer_unwritten(bh))
637 acceptable += (type == IO_UNWRITTEN);
638 else if (buffer_delay(bh))
639 acceptable += (type == IO_DELALLOC);
640 else if (buffer_dirty(bh) && buffer_mapped(bh))
641 acceptable += (type == IO_OVERWRITE);
644 } while ((bh = bh->b_this_page) != head);
654 * Allocate & map buffers for page given the extent map. Write it out.
655 * except for the original page of a writepage, this is called on
656 * delalloc/unwritten pages only, for the original page it is possible
657 * that the page has no mapping at all.
664 struct xfs_bmbt_irec *imap,
665 xfs_ioend_t **ioendp,
666 struct writeback_control *wbc)
668 struct buffer_head *bh, *head;
669 xfs_off_t end_offset;
670 unsigned long p_offset;
673 int count = 0, done = 0, uptodate = 1;
674 xfs_off_t offset = page_offset(page);
676 if (page->index != tindex)
678 if (!trylock_page(page))
680 if (PageWriteback(page))
681 goto fail_unlock_page;
682 if (page->mapping != inode->i_mapping)
683 goto fail_unlock_page;
684 if (!xfs_check_page_type(page, (*ioendp)->io_type))
685 goto fail_unlock_page;
688 * page_dirty is initially a count of buffers on the page before
689 * EOF and is decremented as we move each into a cleanable state.
693 * End offset is the highest offset that this page should represent.
694 * If we are on the last page, (end_offset & (PAGE_CACHE_SIZE - 1))
695 * will evaluate non-zero and be less than PAGE_CACHE_SIZE and
696 * hence give us the correct page_dirty count. On any other page,
697 * it will be zero and in that case we need page_dirty to be the
698 * count of buffers on the page.
700 end_offset = min_t(unsigned long long,
701 (xfs_off_t)(page->index + 1) << PAGE_CACHE_SHIFT,
704 len = 1 << inode->i_blkbits;
705 p_offset = min_t(unsigned long, end_offset & (PAGE_CACHE_SIZE - 1),
707 p_offset = p_offset ? roundup(p_offset, len) : PAGE_CACHE_SIZE;
708 page_dirty = p_offset / len;
710 bh = head = page_buffers(page);
712 if (offset >= end_offset)
714 if (!buffer_uptodate(bh))
716 if (!(PageUptodate(page) || buffer_uptodate(bh))) {
721 if (buffer_unwritten(bh) || buffer_delay(bh) ||
723 if (buffer_unwritten(bh))
725 else if (buffer_delay(bh))
730 if (!xfs_imap_valid(inode, imap, offset)) {
736 if (type != IO_OVERWRITE)
737 xfs_map_at_offset(inode, bh, imap, offset);
738 xfs_add_to_ioend(inode, bh, offset, type,
746 } while (offset += len, (bh = bh->b_this_page) != head);
748 if (uptodate && bh == head)
749 SetPageUptodate(page);
752 if (--wbc->nr_to_write <= 0 &&
753 wbc->sync_mode == WB_SYNC_NONE)
756 xfs_start_page_writeback(page, !page_dirty, count);
766 * Convert & write out a cluster of pages in the same extent as defined
767 * by mp and following the start page.
773 struct xfs_bmbt_irec *imap,
774 xfs_ioend_t **ioendp,
775 struct writeback_control *wbc,
781 pagevec_init(&pvec, 0);
782 while (!done && tindex <= tlast) {
783 unsigned len = min_t(pgoff_t, PAGEVEC_SIZE, tlast - tindex + 1);
785 if (!pagevec_lookup(&pvec, inode->i_mapping, tindex, len))
788 for (i = 0; i < pagevec_count(&pvec); i++) {
789 done = xfs_convert_page(inode, pvec.pages[i], tindex++,
795 pagevec_release(&pvec);
801 xfs_vm_invalidatepage(
803 unsigned long offset)
805 trace_xfs_invalidatepage(page->mapping->host, page, offset);
806 block_invalidatepage(page, offset);
810 * If the page has delalloc buffers on it, we need to punch them out before we
811 * invalidate the page. If we don't, we leave a stale delalloc mapping on the
812 * inode that can trip a BUG() in xfs_get_blocks() later on if a direct IO read
813 * is done on that same region - the delalloc extent is returned when none is
814 * supposed to be there.
816 * We prevent this by truncating away the delalloc regions on the page before
817 * invalidating it. Because they are delalloc, we can do this without needing a
818 * transaction. Indeed - if we get ENOSPC errors, we have to be able to do this
819 * truncation without a transaction as there is no space left for block
820 * reservation (typically why we see a ENOSPC in writeback).
822 * This is not a performance critical path, so for now just do the punching a
823 * buffer head at a time.
826 xfs_aops_discard_page(
829 struct inode *inode = page->mapping->host;
830 struct xfs_inode *ip = XFS_I(inode);
831 struct buffer_head *bh, *head;
832 loff_t offset = page_offset(page);
834 if (!xfs_check_page_type(page, IO_DELALLOC))
837 if (XFS_FORCED_SHUTDOWN(ip->i_mount))
840 xfs_alert(ip->i_mount,
841 "page discard on page %p, inode 0x%llx, offset %llu.",
842 page, ip->i_ino, offset);
844 xfs_ilock(ip, XFS_ILOCK_EXCL);
845 bh = head = page_buffers(page);
848 xfs_fileoff_t start_fsb;
850 if (!buffer_delay(bh))
853 start_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
854 error = xfs_bmap_punch_delalloc_range(ip, start_fsb, 1);
856 /* something screwed, just bail */
857 if (!XFS_FORCED_SHUTDOWN(ip->i_mount)) {
858 xfs_alert(ip->i_mount,
859 "page discard unable to remove delalloc mapping.");
864 offset += 1 << inode->i_blkbits;
866 } while ((bh = bh->b_this_page) != head);
868 xfs_iunlock(ip, XFS_ILOCK_EXCL);
870 xfs_vm_invalidatepage(page, 0);
875 * Write out a dirty page.
877 * For delalloc space on the page we need to allocate space and flush it.
878 * For unwritten space on the page we need to start the conversion to
879 * regular allocated space.
880 * For any other dirty buffer heads on the page we should flush them.
885 struct writeback_control *wbc)
887 struct inode *inode = page->mapping->host;
888 struct buffer_head *bh, *head;
889 struct xfs_bmbt_irec imap;
890 xfs_ioend_t *ioend = NULL, *iohead = NULL;
893 __uint64_t end_offset;
894 pgoff_t end_index, last_index;
896 int err, imap_valid = 0, uptodate = 1;
900 trace_xfs_writepage(inode, page, 0);
902 ASSERT(page_has_buffers(page));
905 * Refuse to write the page out if we are called from reclaim context.
907 * This avoids stack overflows when called from deeply used stacks in
908 * random callers for direct reclaim or memcg reclaim. We explicitly
909 * allow reclaim from kswapd as the stack usage there is relatively low.
911 * This should never happen except in the case of a VM regression so
914 if (WARN_ON_ONCE((current->flags & (PF_MEMALLOC|PF_KSWAPD)) ==
919 * Given that we do not allow direct reclaim to call us, we should
920 * never be called while in a filesystem transaction.
922 if (WARN_ON(current->flags & PF_FSTRANS))
925 /* Is this page beyond the end of the file? */
926 offset = i_size_read(inode);
927 end_index = offset >> PAGE_CACHE_SHIFT;
928 last_index = (offset - 1) >> PAGE_CACHE_SHIFT;
929 if (page->index >= end_index) {
930 if ((page->index >= end_index + 1) ||
931 !(i_size_read(inode) & (PAGE_CACHE_SIZE - 1))) {
937 end_offset = min_t(unsigned long long,
938 (xfs_off_t)(page->index + 1) << PAGE_CACHE_SHIFT,
940 len = 1 << inode->i_blkbits;
942 bh = head = page_buffers(page);
943 offset = page_offset(page);
946 if (wbc->sync_mode == WB_SYNC_NONE)
952 if (offset >= end_offset)
954 if (!buffer_uptodate(bh))
958 * set_page_dirty dirties all buffers in a page, independent
959 * of their state. The dirty state however is entirely
960 * meaningless for holes (!mapped && uptodate), so skip
961 * buffers covering holes here.
963 if (!buffer_mapped(bh) && buffer_uptodate(bh)) {
968 if (buffer_unwritten(bh)) {
969 if (type != IO_UNWRITTEN) {
973 } else if (buffer_delay(bh)) {
974 if (type != IO_DELALLOC) {
978 } else if (buffer_uptodate(bh)) {
979 if (type != IO_OVERWRITE) {
984 if (PageUptodate(page))
985 ASSERT(buffer_mapped(bh));
987 * This buffer is not uptodate and will not be
988 * written to disk. Ensure that we will put any
989 * subsequent writeable buffers into a new
997 imap_valid = xfs_imap_valid(inode, &imap, offset);
1000 * If we didn't have a valid mapping then we need to
1001 * put the new mapping into a separate ioend structure.
1002 * This ensures non-contiguous extents always have
1003 * separate ioends, which is particularly important
1004 * for unwritten extent conversion at I/O completion
1008 err = xfs_map_blocks(inode, offset, &imap, type,
1012 imap_valid = xfs_imap_valid(inode, &imap, offset);
1016 if (type != IO_OVERWRITE)
1017 xfs_map_at_offset(inode, bh, &imap, offset);
1018 xfs_add_to_ioend(inode, bh, offset, type, &ioend,
1026 } while (offset += len, ((bh = bh->b_this_page) != head));
1028 if (uptodate && bh == head)
1029 SetPageUptodate(page);
1031 xfs_start_page_writeback(page, 1, count);
1033 if (ioend && imap_valid) {
1034 xfs_off_t end_index;
1036 end_index = imap.br_startoff + imap.br_blockcount;
1039 end_index <<= inode->i_blkbits;
1042 end_index = (end_index - 1) >> PAGE_CACHE_SHIFT;
1044 /* check against file size */
1045 if (end_index > last_index)
1046 end_index = last_index;
1048 xfs_cluster_write(inode, page->index + 1, &imap, &ioend,
1054 * Reserve log space if we might write beyond the on-disk
1057 if (ioend->io_type != IO_UNWRITTEN &&
1058 xfs_ioend_is_append(ioend)) {
1059 err = xfs_setfilesize_trans_alloc(ioend);
1064 xfs_submit_ioend(wbc, iohead);
1071 xfs_cancel_ioend(iohead);
1076 xfs_aops_discard_page(page);
1077 ClearPageUptodate(page);
1082 redirty_page_for_writepage(wbc, page);
1089 struct address_space *mapping,
1090 struct writeback_control *wbc)
1092 xfs_iflags_clear(XFS_I(mapping->host), XFS_ITRUNCATED);
1093 return generic_writepages(mapping, wbc);
1097 * Called to move a page into cleanable state - and from there
1098 * to be released. The page should already be clean. We always
1099 * have buffer heads in this call.
1101 * Returns 1 if the page is ok to release, 0 otherwise.
1108 int delalloc, unwritten;
1110 trace_xfs_releasepage(page->mapping->host, page, 0);
1112 xfs_count_page_state(page, &delalloc, &unwritten);
1114 if (WARN_ON(delalloc))
1116 if (WARN_ON(unwritten))
1119 return try_to_free_buffers(page);
1124 struct inode *inode,
1126 struct buffer_head *bh_result,
1130 struct xfs_inode *ip = XFS_I(inode);
1131 struct xfs_mount *mp = ip->i_mount;
1132 xfs_fileoff_t offset_fsb, end_fsb;
1135 struct xfs_bmbt_irec imap;
1141 if (XFS_FORCED_SHUTDOWN(mp))
1142 return -XFS_ERROR(EIO);
1144 offset = (xfs_off_t)iblock << inode->i_blkbits;
1145 ASSERT(bh_result->b_size >= (1 << inode->i_blkbits));
1146 size = bh_result->b_size;
1148 if (!create && direct && offset >= i_size_read(inode))
1152 * Direct I/O is usually done on preallocated files, so try getting
1153 * a block mapping without an exclusive lock first. For buffered
1154 * writes we already have the exclusive iolock anyway, so avoiding
1155 * a lock roundtrip here by taking the ilock exclusive from the
1156 * beginning is a useful micro optimization.
1158 if (create && !direct) {
1159 lockmode = XFS_ILOCK_EXCL;
1160 xfs_ilock(ip, lockmode);
1162 lockmode = xfs_ilock_map_shared(ip);
1165 ASSERT(offset <= mp->m_maxioffset);
1166 if (offset + size > mp->m_maxioffset)
1167 size = mp->m_maxioffset - offset;
1168 end_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)offset + size);
1169 offset_fsb = XFS_B_TO_FSBT(mp, offset);
1171 error = xfs_bmapi_read(ip, offset_fsb, end_fsb - offset_fsb,
1172 &imap, &nimaps, XFS_BMAPI_ENTIRE);
1178 (imap.br_startblock == HOLESTARTBLOCK ||
1179 imap.br_startblock == DELAYSTARTBLOCK))) {
1180 if (direct || xfs_get_extsz_hint(ip)) {
1182 * Drop the ilock in preparation for starting the block
1183 * allocation transaction. It will be retaken
1184 * exclusively inside xfs_iomap_write_direct for the
1185 * actual allocation.
1187 xfs_iunlock(ip, lockmode);
1188 error = xfs_iomap_write_direct(ip, offset, size,
1195 * Delalloc reservations do not require a transaction,
1196 * we can go on without dropping the lock here. If we
1197 * are allocating a new delalloc block, make sure that
1198 * we set the new flag so that we mark the buffer new so
1199 * that we know that it is newly allocated if the write
1202 if (nimaps && imap.br_startblock == HOLESTARTBLOCK)
1204 error = xfs_iomap_write_delay(ip, offset, size, &imap);
1208 xfs_iunlock(ip, lockmode);
1211 trace_xfs_get_blocks_alloc(ip, offset, size, 0, &imap);
1212 } else if (nimaps) {
1213 trace_xfs_get_blocks_found(ip, offset, size, 0, &imap);
1214 xfs_iunlock(ip, lockmode);
1216 trace_xfs_get_blocks_notfound(ip, offset, size);
1220 if (imap.br_startblock != HOLESTARTBLOCK &&
1221 imap.br_startblock != DELAYSTARTBLOCK) {
1223 * For unwritten extents do not report a disk address on
1224 * the read case (treat as if we're reading into a hole).
1226 if (create || !ISUNWRITTEN(&imap))
1227 xfs_map_buffer(inode, bh_result, &imap, offset);
1228 if (create && ISUNWRITTEN(&imap)) {
1230 bh_result->b_private = inode;
1231 set_buffer_unwritten(bh_result);
1236 * If this is a realtime file, data may be on a different device.
1237 * to that pointed to from the buffer_head b_bdev currently.
1239 bh_result->b_bdev = xfs_find_bdev_for_inode(inode);
1242 * If we previously allocated a block out beyond eof and we are now
1243 * coming back to use it then we will need to flag it as new even if it
1244 * has a disk address.
1246 * With sub-block writes into unwritten extents we also need to mark
1247 * the buffer as new so that the unwritten parts of the buffer gets
1251 ((!buffer_mapped(bh_result) && !buffer_uptodate(bh_result)) ||
1252 (offset >= i_size_read(inode)) ||
1253 (new || ISUNWRITTEN(&imap))))
1254 set_buffer_new(bh_result);
1256 if (imap.br_startblock == DELAYSTARTBLOCK) {
1259 set_buffer_uptodate(bh_result);
1260 set_buffer_mapped(bh_result);
1261 set_buffer_delay(bh_result);
1266 * If this is O_DIRECT or the mpage code calling tell them how large
1267 * the mapping is, so that we can avoid repeated get_blocks calls.
1269 if (direct || size > (1 << inode->i_blkbits)) {
1270 xfs_off_t mapping_size;
1272 mapping_size = imap.br_startoff + imap.br_blockcount - iblock;
1273 mapping_size <<= inode->i_blkbits;
1275 ASSERT(mapping_size > 0);
1276 if (mapping_size > size)
1277 mapping_size = size;
1278 if (mapping_size > LONG_MAX)
1279 mapping_size = LONG_MAX;
1281 bh_result->b_size = mapping_size;
1287 xfs_iunlock(ip, lockmode);
1293 struct inode *inode,
1295 struct buffer_head *bh_result,
1298 return __xfs_get_blocks(inode, iblock, bh_result, create, 0);
1302 xfs_get_blocks_direct(
1303 struct inode *inode,
1305 struct buffer_head *bh_result,
1308 return __xfs_get_blocks(inode, iblock, bh_result, create, 1);
1312 * Complete a direct I/O write request.
1314 * If the private argument is non-NULL __xfs_get_blocks signals us that we
1315 * need to issue a transaction to convert the range from unwritten to written
1316 * extents. In case this is regular synchronous I/O we just call xfs_end_io
1317 * to do this and we are done. But in case this was a successful AIO
1318 * request this handler is called from interrupt context, from which we
1319 * can't start transactions. In that case offload the I/O completion to
1320 * the workqueues we also use for buffered I/O completion.
1323 xfs_end_io_direct_write(
1331 struct xfs_ioend *ioend = iocb->private;
1334 * While the generic direct I/O code updates the inode size, it does
1335 * so only after the end_io handler is called, which means our
1336 * end_io handler thinks the on-disk size is outside the in-core
1337 * size. To prevent this just update it a little bit earlier here.
1339 if (offset + size > i_size_read(ioend->io_inode))
1340 i_size_write(ioend->io_inode, offset + size);
1343 * blockdev_direct_IO can return an error even after the I/O
1344 * completion handler was called. Thus we need to protect
1345 * against double-freeing.
1347 iocb->private = NULL;
1349 ioend->io_offset = offset;
1350 ioend->io_size = size;
1351 ioend->io_iocb = iocb;
1352 ioend->io_result = ret;
1353 if (private && size > 0)
1354 ioend->io_type = IO_UNWRITTEN;
1357 ioend->io_isasync = 1;
1358 xfs_finish_ioend(ioend);
1360 xfs_finish_ioend_sync(ioend);
1368 const struct iovec *iov,
1370 unsigned long nr_segs)
1372 struct inode *inode = iocb->ki_filp->f_mapping->host;
1373 struct block_device *bdev = xfs_find_bdev_for_inode(inode);
1374 struct xfs_ioend *ioend = NULL;
1378 size_t size = iov_length(iov, nr_segs);
1381 * We need to preallocate a transaction for a size update
1382 * here. In the case that this write both updates the size
1383 * and converts at least on unwritten extent we will cancel
1384 * the still clean transaction after the I/O has finished.
1386 iocb->private = ioend = xfs_alloc_ioend(inode, IO_DIRECT);
1387 if (offset + size > XFS_I(inode)->i_d.di_size) {
1388 ret = xfs_setfilesize_trans_alloc(ioend);
1390 goto out_destroy_ioend;
1391 ioend->io_isdirect = 1;
1394 ret = __blockdev_direct_IO(rw, iocb, inode, bdev, iov,
1396 xfs_get_blocks_direct,
1397 xfs_end_io_direct_write, NULL, 0);
1398 if (ret != -EIOCBQUEUED && iocb->private)
1399 goto out_trans_cancel;
1401 ret = __blockdev_direct_IO(rw, iocb, inode, bdev, iov,
1403 xfs_get_blocks_direct,
1410 if (ioend->io_append_trans) {
1411 current_set_flags_nested(&ioend->io_append_trans->t_pflags,
1413 xfs_trans_cancel(ioend->io_append_trans, 0);
1416 xfs_destroy_ioend(ioend);
1421 * Punch out the delalloc blocks we have already allocated.
1423 * Don't bother with xfs_setattr given that nothing can have made it to disk yet
1424 * as the page is still locked at this point.
1427 xfs_vm_kill_delalloc_range(
1428 struct inode *inode,
1432 struct xfs_inode *ip = XFS_I(inode);
1433 xfs_fileoff_t start_fsb;
1434 xfs_fileoff_t end_fsb;
1437 start_fsb = XFS_B_TO_FSB(ip->i_mount, start);
1438 end_fsb = XFS_B_TO_FSB(ip->i_mount, end);
1439 if (end_fsb <= start_fsb)
1442 xfs_ilock(ip, XFS_ILOCK_EXCL);
1443 error = xfs_bmap_punch_delalloc_range(ip, start_fsb,
1444 end_fsb - start_fsb);
1446 /* something screwed, just bail */
1447 if (!XFS_FORCED_SHUTDOWN(ip->i_mount)) {
1448 xfs_alert(ip->i_mount,
1449 "xfs_vm_write_failed: unable to clean up ino %lld",
1453 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1457 xfs_vm_write_failed(
1458 struct inode *inode,
1463 loff_t block_offset = pos & PAGE_MASK;
1466 loff_t from = pos & (PAGE_CACHE_SIZE - 1);
1467 loff_t to = from + len;
1468 struct buffer_head *bh, *head;
1470 ASSERT(block_offset + from == pos);
1472 head = page_buffers(page);
1474 for (bh = head; bh != head || !block_start;
1475 bh = bh->b_this_page, block_start = block_end,
1476 block_offset += bh->b_size) {
1477 block_end = block_start + bh->b_size;
1479 /* skip buffers before the write */
1480 if (block_end <= from)
1483 /* if the buffer is after the write, we're done */
1484 if (block_start >= to)
1487 if (!buffer_delay(bh))
1490 if (!buffer_new(bh) && block_offset < i_size_read(inode))
1493 xfs_vm_kill_delalloc_range(inode, block_offset,
1494 block_offset + bh->b_size);
1500 * This used to call block_write_begin(), but it unlocks and releases the page
1501 * on error, and we need that page to be able to punch stale delalloc blocks out
1502 * on failure. hence we copy-n-waste it here and call xfs_vm_write_failed() at
1503 * the appropriate point.
1508 struct address_space *mapping,
1512 struct page **pagep,
1515 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1519 ASSERT(len <= PAGE_CACHE_SIZE);
1521 page = grab_cache_page_write_begin(mapping, index,
1522 flags | AOP_FLAG_NOFS);
1526 status = __block_write_begin(page, pos, len, xfs_get_blocks);
1527 if (unlikely(status)) {
1528 struct inode *inode = mapping->host;
1530 xfs_vm_write_failed(inode, page, pos, len);
1533 if (pos + len > i_size_read(inode))
1534 truncate_pagecache(inode, pos + len, i_size_read(inode));
1536 page_cache_release(page);
1545 * On failure, we only need to kill delalloc blocks beyond EOF because they
1546 * will never be written. For blocks within EOF, generic_write_end() zeros them
1547 * so they are safe to leave alone and be written with all the other valid data.
1552 struct address_space *mapping,
1561 ASSERT(len <= PAGE_CACHE_SIZE);
1563 ret = generic_write_end(file, mapping, pos, len, copied, page, fsdata);
1564 if (unlikely(ret < len)) {
1565 struct inode *inode = mapping->host;
1566 size_t isize = i_size_read(inode);
1567 loff_t to = pos + len;
1570 truncate_pagecache(inode, to, isize);
1571 xfs_vm_kill_delalloc_range(inode, isize, to);
1579 struct address_space *mapping,
1582 struct inode *inode = (struct inode *)mapping->host;
1583 struct xfs_inode *ip = XFS_I(inode);
1585 trace_xfs_vm_bmap(XFS_I(inode));
1586 xfs_ilock(ip, XFS_IOLOCK_SHARED);
1587 xfs_flush_pages(ip, (xfs_off_t)0, -1, 0, FI_REMAPF);
1588 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
1589 return generic_block_bmap(mapping, block, xfs_get_blocks);
1594 struct file *unused,
1597 return mpage_readpage(page, xfs_get_blocks);
1602 struct file *unused,
1603 struct address_space *mapping,
1604 struct list_head *pages,
1607 return mpage_readpages(mapping, pages, nr_pages, xfs_get_blocks);
1610 const struct address_space_operations xfs_address_space_operations = {
1611 .readpage = xfs_vm_readpage,
1612 .readpages = xfs_vm_readpages,
1613 .writepage = xfs_vm_writepage,
1614 .writepages = xfs_vm_writepages,
1615 .releasepage = xfs_vm_releasepage,
1616 .invalidatepage = xfs_vm_invalidatepage,
1617 .write_begin = xfs_vm_write_begin,
1618 .write_end = xfs_vm_write_end,
1619 .bmap = xfs_vm_bmap,
1620 .direct_IO = xfs_vm_direct_IO,
1621 .migratepage = buffer_migrate_page,
1622 .is_partially_uptodate = block_is_partially_uptodate,
1623 .error_remove_page = generic_error_remove_page,