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
19 #include "xfs_format.h"
20 #include "xfs_shared.h"
24 #include "xfs_trans.h"
25 #include "xfs_mount.h"
26 #include "xfs_bmap_btree.h"
27 #include "xfs_dinode.h"
28 #include "xfs_inode.h"
29 #include "xfs_inode_item.h"
30 #include "xfs_alloc.h"
31 #include "xfs_error.h"
32 #include "xfs_iomap.h"
33 #include "xfs_trace.h"
35 #include "xfs_bmap_util.h"
36 #include <linux/aio.h>
37 #include <linux/gfp.h>
38 #include <linux/mpage.h>
39 #include <linux/pagevec.h>
40 #include <linux/writeback.h>
48 struct buffer_head *bh, *head;
50 *delalloc = *unwritten = 0;
52 bh = head = page_buffers(page);
54 if (buffer_unwritten(bh))
56 else if (buffer_delay(bh))
58 } while ((bh = bh->b_this_page) != head);
61 STATIC struct block_device *
62 xfs_find_bdev_for_inode(
65 struct xfs_inode *ip = XFS_I(inode);
66 struct xfs_mount *mp = ip->i_mount;
68 if (XFS_IS_REALTIME_INODE(ip))
69 return mp->m_rtdev_targp->bt_bdev;
71 return mp->m_ddev_targp->bt_bdev;
75 * We're now finished for good with this ioend structure.
76 * Update the page state via the associated buffer_heads,
77 * release holds on the inode and bio, and finally free
78 * up memory. Do not use the ioend after this.
84 struct buffer_head *bh, *next;
86 for (bh = ioend->io_buffer_head; bh; bh = next) {
88 bh->b_end_io(bh, !ioend->io_error);
91 mempool_free(ioend, xfs_ioend_pool);
95 * Fast and loose check if this write could update the on-disk inode size.
97 static inline bool xfs_ioend_is_append(struct xfs_ioend *ioend)
99 return ioend->io_offset + ioend->io_size >
100 XFS_I(ioend->io_inode)->i_d.di_size;
104 xfs_setfilesize_trans_alloc(
105 struct xfs_ioend *ioend)
107 struct xfs_mount *mp = XFS_I(ioend->io_inode)->i_mount;
108 struct xfs_trans *tp;
111 tp = xfs_trans_alloc(mp, XFS_TRANS_FSYNC_TS);
113 error = xfs_trans_reserve(tp, &M_RES(mp)->tr_fsyncts, 0, 0);
115 xfs_trans_cancel(tp, 0);
119 ioend->io_append_trans = tp;
122 * We may pass freeze protection with a transaction. So tell lockdep
125 rwsem_release(&ioend->io_inode->i_sb->s_writers.lock_map[SB_FREEZE_FS-1],
128 * We hand off the transaction to the completion thread now, so
129 * clear the flag here.
131 current_restore_flags_nested(&tp->t_pflags, PF_FSTRANS);
136 * Update on-disk file size now that data has been written to disk.
140 struct xfs_ioend *ioend)
142 struct xfs_inode *ip = XFS_I(ioend->io_inode);
143 struct xfs_trans *tp = ioend->io_append_trans;
147 * The transaction may have been allocated in the I/O submission thread,
148 * thus we need to mark ourselves as beeing in a transaction manually.
149 * Similarly for freeze protection.
151 current_set_flags_nested(&tp->t_pflags, PF_FSTRANS);
152 rwsem_acquire_read(&VFS_I(ip)->i_sb->s_writers.lock_map[SB_FREEZE_FS-1],
155 xfs_ilock(ip, XFS_ILOCK_EXCL);
156 isize = xfs_new_eof(ip, ioend->io_offset + ioend->io_size);
158 xfs_iunlock(ip, XFS_ILOCK_EXCL);
159 xfs_trans_cancel(tp, 0);
163 trace_xfs_setfilesize(ip, ioend->io_offset, ioend->io_size);
165 ip->i_d.di_size = isize;
166 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
167 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
169 return xfs_trans_commit(tp, 0);
173 * Schedule IO completion handling on the final put of an ioend.
175 * If there is no work to do we might as well call it a day and free the
180 struct xfs_ioend *ioend)
182 if (atomic_dec_and_test(&ioend->io_remaining)) {
183 struct xfs_mount *mp = XFS_I(ioend->io_inode)->i_mount;
185 if (ioend->io_type == XFS_IO_UNWRITTEN)
186 queue_work(mp->m_unwritten_workqueue, &ioend->io_work);
187 else if (ioend->io_append_trans ||
188 (ioend->io_isdirect && xfs_ioend_is_append(ioend)))
189 queue_work(mp->m_data_workqueue, &ioend->io_work);
191 xfs_destroy_ioend(ioend);
196 * IO write completion.
200 struct work_struct *work)
202 xfs_ioend_t *ioend = container_of(work, xfs_ioend_t, io_work);
203 struct xfs_inode *ip = XFS_I(ioend->io_inode);
206 if (XFS_FORCED_SHUTDOWN(ip->i_mount)) {
207 ioend->io_error = -EIO;
214 * For unwritten extents we need to issue transactions to convert a
215 * range to normal written extens after the data I/O has finished.
217 if (ioend->io_type == XFS_IO_UNWRITTEN) {
218 error = xfs_iomap_write_unwritten(ip, ioend->io_offset,
220 } else if (ioend->io_isdirect && xfs_ioend_is_append(ioend)) {
222 * For direct I/O we do not know if we need to allocate blocks
223 * or not so we can't preallocate an append transaction as that
224 * results in nested reservations and log space deadlocks. Hence
225 * allocate the transaction here. While this is sub-optimal and
226 * can block IO completion for some time, we're stuck with doing
227 * it this way until we can pass the ioend to the direct IO
228 * allocation callbacks and avoid nesting that way.
230 error = xfs_setfilesize_trans_alloc(ioend);
233 error = xfs_setfilesize(ioend);
234 } else if (ioend->io_append_trans) {
235 error = xfs_setfilesize(ioend);
237 ASSERT(!xfs_ioend_is_append(ioend));
242 ioend->io_error = -error;
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_isdirect = 0;
280 ioend->io_list = NULL;
281 ioend->io_type = type;
282 ioend->io_inode = inode;
283 ioend->io_buffer_head = NULL;
284 ioend->io_buffer_tail = NULL;
285 ioend->io_offset = 0;
287 ioend->io_append_trans = NULL;
289 INIT_WORK(&ioend->io_work, xfs_end_io);
297 struct xfs_bmbt_irec *imap,
301 struct xfs_inode *ip = XFS_I(inode);
302 struct xfs_mount *mp = ip->i_mount;
303 ssize_t count = 1 << inode->i_blkbits;
304 xfs_fileoff_t offset_fsb, end_fsb;
306 int bmapi_flags = XFS_BMAPI_ENTIRE;
309 if (XFS_FORCED_SHUTDOWN(mp))
310 return -XFS_ERROR(EIO);
312 if (type == XFS_IO_UNWRITTEN)
313 bmapi_flags |= XFS_BMAPI_IGSTATE;
315 if (!xfs_ilock_nowait(ip, XFS_ILOCK_SHARED)) {
317 return -XFS_ERROR(EAGAIN);
318 xfs_ilock(ip, XFS_ILOCK_SHARED);
321 ASSERT(ip->i_d.di_format != XFS_DINODE_FMT_BTREE ||
322 (ip->i_df.if_flags & XFS_IFEXTENTS));
323 ASSERT(offset <= mp->m_super->s_maxbytes);
325 if (offset + count > mp->m_super->s_maxbytes)
326 count = mp->m_super->s_maxbytes - offset;
327 end_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)offset + count);
328 offset_fsb = XFS_B_TO_FSBT(mp, offset);
329 error = xfs_bmapi_read(ip, offset_fsb, end_fsb - offset_fsb,
330 imap, &nimaps, bmapi_flags);
331 xfs_iunlock(ip, XFS_ILOCK_SHARED);
334 return -XFS_ERROR(error);
336 if (type == XFS_IO_DELALLOC &&
337 (!nimaps || isnullstartblock(imap->br_startblock))) {
338 error = xfs_iomap_write_allocate(ip, offset, imap);
340 trace_xfs_map_blocks_alloc(ip, offset, count, type, imap);
341 return -XFS_ERROR(error);
345 if (type == XFS_IO_UNWRITTEN) {
347 ASSERT(imap->br_startblock != HOLESTARTBLOCK);
348 ASSERT(imap->br_startblock != DELAYSTARTBLOCK);
352 trace_xfs_map_blocks_found(ip, offset, count, type, imap);
359 struct xfs_bmbt_irec *imap,
362 offset >>= inode->i_blkbits;
364 return offset >= imap->br_startoff &&
365 offset < imap->br_startoff + imap->br_blockcount;
369 * BIO completion handler for buffered IO.
376 xfs_ioend_t *ioend = bio->bi_private;
378 ASSERT(atomic_read(&bio->bi_cnt) >= 1);
379 ioend->io_error = test_bit(BIO_UPTODATE, &bio->bi_flags) ? 0 : error;
381 /* Toss bio and pass work off to an xfsdatad thread */
382 bio->bi_private = NULL;
383 bio->bi_end_io = NULL;
386 xfs_finish_ioend(ioend);
390 xfs_submit_ioend_bio(
391 struct writeback_control *wbc,
395 atomic_inc(&ioend->io_remaining);
396 bio->bi_private = ioend;
397 bio->bi_end_io = xfs_end_bio;
398 submit_bio(wbc->sync_mode == WB_SYNC_ALL ? WRITE_SYNC : WRITE, bio);
403 struct buffer_head *bh)
405 int nvecs = bio_get_nr_vecs(bh->b_bdev);
406 struct bio *bio = bio_alloc(GFP_NOIO, nvecs);
408 ASSERT(bio->bi_private == NULL);
409 bio->bi_sector = bh->b_blocknr * (bh->b_size >> 9);
410 bio->bi_bdev = bh->b_bdev;
415 xfs_start_buffer_writeback(
416 struct buffer_head *bh)
418 ASSERT(buffer_mapped(bh));
419 ASSERT(buffer_locked(bh));
420 ASSERT(!buffer_delay(bh));
421 ASSERT(!buffer_unwritten(bh));
423 mark_buffer_async_write(bh);
424 set_buffer_uptodate(bh);
425 clear_buffer_dirty(bh);
429 xfs_start_page_writeback(
434 ASSERT(PageLocked(page));
435 ASSERT(!PageWriteback(page));
437 clear_page_dirty_for_io(page);
438 set_page_writeback(page);
440 /* If no buffers on the page are to be written, finish it here */
442 end_page_writeback(page);
445 static inline int xfs_bio_add_buffer(struct bio *bio, struct buffer_head *bh)
447 return bio_add_page(bio, bh->b_page, bh->b_size, bh_offset(bh));
451 * Submit all of the bios for all of the ioends we have saved up, covering the
452 * initial writepage page and also any probed pages.
454 * Because we may have multiple ioends spanning a page, we need to start
455 * writeback on all the buffers before we submit them for I/O. If we mark the
456 * buffers as we got, then we can end up with a page that only has buffers
457 * marked async write and I/O complete on can occur before we mark the other
458 * buffers async write.
460 * The end result of this is that we trip a bug in end_page_writeback() because
461 * we call it twice for the one page as the code in end_buffer_async_write()
462 * assumes that all buffers on the page are started at the same time.
464 * The fix is two passes across the ioend list - one to start writeback on the
465 * buffer_heads, and then submit them for I/O on the second pass.
467 * If @fail is non-zero, it means that we have a situation where some part of
468 * the submission process has failed after we have marked paged for writeback
469 * and unlocked them. In this situation, we need to fail the ioend chain rather
470 * than submit it to IO. This typically only happens on a filesystem shutdown.
474 struct writeback_control *wbc,
478 xfs_ioend_t *head = ioend;
480 struct buffer_head *bh;
482 sector_t lastblock = 0;
484 /* Pass 1 - start writeback */
486 next = ioend->io_list;
487 for (bh = ioend->io_buffer_head; bh; bh = bh->b_private)
488 xfs_start_buffer_writeback(bh);
489 } while ((ioend = next) != NULL);
491 /* Pass 2 - submit I/O */
494 next = ioend->io_list;
498 * If we are failing the IO now, just mark the ioend with an
499 * error and finish it. This will run IO completion immediately
500 * as there is only one reference to the ioend at this point in
504 ioend->io_error = -fail;
505 xfs_finish_ioend(ioend);
509 for (bh = ioend->io_buffer_head; bh; bh = bh->b_private) {
513 bio = xfs_alloc_ioend_bio(bh);
514 } else if (bh->b_blocknr != lastblock + 1) {
515 xfs_submit_ioend_bio(wbc, ioend, bio);
519 if (xfs_bio_add_buffer(bio, bh) != bh->b_size) {
520 xfs_submit_ioend_bio(wbc, ioend, bio);
524 lastblock = bh->b_blocknr;
527 xfs_submit_ioend_bio(wbc, ioend, bio);
528 xfs_finish_ioend(ioend);
529 } while ((ioend = next) != NULL);
533 * Cancel submission of all buffer_heads so far in this endio.
534 * Toss the endio too. Only ever called for the initial page
535 * in a writepage request, so only ever one page.
542 struct buffer_head *bh, *next_bh;
545 next = ioend->io_list;
546 bh = ioend->io_buffer_head;
548 next_bh = bh->b_private;
549 clear_buffer_async_write(bh);
551 } while ((bh = next_bh) != NULL);
553 mempool_free(ioend, xfs_ioend_pool);
554 } while ((ioend = next) != NULL);
558 * Test to see if we've been building up a completion structure for
559 * earlier buffers -- if so, we try to append to this ioend if we
560 * can, otherwise we finish off any current ioend and start another.
561 * Return true if we've finished the given ioend.
566 struct buffer_head *bh,
569 xfs_ioend_t **result,
572 xfs_ioend_t *ioend = *result;
574 if (!ioend || need_ioend || type != ioend->io_type) {
575 xfs_ioend_t *previous = *result;
577 ioend = xfs_alloc_ioend(inode, type);
578 ioend->io_offset = offset;
579 ioend->io_buffer_head = bh;
580 ioend->io_buffer_tail = bh;
582 previous->io_list = ioend;
585 ioend->io_buffer_tail->b_private = bh;
586 ioend->io_buffer_tail = bh;
589 bh->b_private = NULL;
590 ioend->io_size += bh->b_size;
596 struct buffer_head *bh,
597 struct xfs_bmbt_irec *imap,
601 struct xfs_mount *m = XFS_I(inode)->i_mount;
602 xfs_off_t iomap_offset = XFS_FSB_TO_B(m, imap->br_startoff);
603 xfs_daddr_t iomap_bn = xfs_fsb_to_db(XFS_I(inode), imap->br_startblock);
605 ASSERT(imap->br_startblock != HOLESTARTBLOCK);
606 ASSERT(imap->br_startblock != DELAYSTARTBLOCK);
608 bn = (iomap_bn >> (inode->i_blkbits - BBSHIFT)) +
609 ((offset - iomap_offset) >> inode->i_blkbits);
611 ASSERT(bn || XFS_IS_REALTIME_INODE(XFS_I(inode)));
614 set_buffer_mapped(bh);
620 struct buffer_head *bh,
621 struct xfs_bmbt_irec *imap,
624 ASSERT(imap->br_startblock != HOLESTARTBLOCK);
625 ASSERT(imap->br_startblock != DELAYSTARTBLOCK);
627 xfs_map_buffer(inode, bh, imap, offset);
628 set_buffer_mapped(bh);
629 clear_buffer_delay(bh);
630 clear_buffer_unwritten(bh);
634 * Test if a given page is suitable for writing as part of an unwritten
635 * or delayed allocate extent.
642 if (PageWriteback(page))
645 if (page->mapping && page_has_buffers(page)) {
646 struct buffer_head *bh, *head;
649 bh = head = page_buffers(page);
651 if (buffer_unwritten(bh))
652 acceptable += (type == XFS_IO_UNWRITTEN);
653 else if (buffer_delay(bh))
654 acceptable += (type == XFS_IO_DELALLOC);
655 else if (buffer_dirty(bh) && buffer_mapped(bh))
656 acceptable += (type == XFS_IO_OVERWRITE);
659 } while ((bh = bh->b_this_page) != head);
669 * Allocate & map buffers for page given the extent map. Write it out.
670 * except for the original page of a writepage, this is called on
671 * delalloc/unwritten pages only, for the original page it is possible
672 * that the page has no mapping at all.
679 struct xfs_bmbt_irec *imap,
680 xfs_ioend_t **ioendp,
681 struct writeback_control *wbc)
683 struct buffer_head *bh, *head;
684 xfs_off_t end_offset;
685 unsigned long p_offset;
688 int count = 0, done = 0, uptodate = 1;
689 xfs_off_t offset = page_offset(page);
691 if (page->index != tindex)
693 if (!trylock_page(page))
695 if (PageWriteback(page))
696 goto fail_unlock_page;
697 if (page->mapping != inode->i_mapping)
698 goto fail_unlock_page;
699 if (!xfs_check_page_type(page, (*ioendp)->io_type))
700 goto fail_unlock_page;
703 * page_dirty is initially a count of buffers on the page before
704 * EOF and is decremented as we move each into a cleanable state.
708 * End offset is the highest offset that this page should represent.
709 * If we are on the last page, (end_offset & (PAGE_CACHE_SIZE - 1))
710 * will evaluate non-zero and be less than PAGE_CACHE_SIZE and
711 * hence give us the correct page_dirty count. On any other page,
712 * it will be zero and in that case we need page_dirty to be the
713 * count of buffers on the page.
715 end_offset = min_t(unsigned long long,
716 (xfs_off_t)(page->index + 1) << PAGE_CACHE_SHIFT,
720 * If the current map does not span the entire page we are about to try
721 * to write, then give up. The only way we can write a page that spans
722 * multiple mappings in a single writeback iteration is via the
723 * xfs_vm_writepage() function. Data integrity writeback requires the
724 * entire page to be written in a single attempt, otherwise the part of
725 * the page we don't write here doesn't get written as part of the data
728 * For normal writeback, we also don't attempt to write partial pages
729 * here as it simply means that write_cache_pages() will see it under
730 * writeback and ignore the page until some point in the future, at
731 * which time this will be the only page in the file that needs
732 * writeback. Hence for more optimal IO patterns, we should always
733 * avoid partial page writeback due to multiple mappings on a page here.
735 if (!xfs_imap_valid(inode, imap, end_offset))
736 goto fail_unlock_page;
738 len = 1 << inode->i_blkbits;
739 p_offset = min_t(unsigned long, end_offset & (PAGE_CACHE_SIZE - 1),
741 p_offset = p_offset ? roundup(p_offset, len) : PAGE_CACHE_SIZE;
742 page_dirty = p_offset / len;
744 bh = head = page_buffers(page);
746 if (offset >= end_offset)
748 if (!buffer_uptodate(bh))
750 if (!(PageUptodate(page) || buffer_uptodate(bh))) {
755 if (buffer_unwritten(bh) || buffer_delay(bh) ||
757 if (buffer_unwritten(bh))
758 type = XFS_IO_UNWRITTEN;
759 else if (buffer_delay(bh))
760 type = XFS_IO_DELALLOC;
762 type = XFS_IO_OVERWRITE;
764 if (!xfs_imap_valid(inode, imap, offset)) {
770 if (type != XFS_IO_OVERWRITE)
771 xfs_map_at_offset(inode, bh, imap, offset);
772 xfs_add_to_ioend(inode, bh, offset, type,
780 } while (offset += len, (bh = bh->b_this_page) != head);
782 if (uptodate && bh == head)
783 SetPageUptodate(page);
786 if (--wbc->nr_to_write <= 0 &&
787 wbc->sync_mode == WB_SYNC_NONE)
790 xfs_start_page_writeback(page, !page_dirty, count);
800 * Convert & write out a cluster of pages in the same extent as defined
801 * by mp and following the start page.
807 struct xfs_bmbt_irec *imap,
808 xfs_ioend_t **ioendp,
809 struct writeback_control *wbc,
815 pagevec_init(&pvec, 0);
816 while (!done && tindex <= tlast) {
817 unsigned len = min_t(pgoff_t, PAGEVEC_SIZE, tlast - tindex + 1);
819 if (!pagevec_lookup(&pvec, inode->i_mapping, tindex, len))
822 for (i = 0; i < pagevec_count(&pvec); i++) {
823 done = xfs_convert_page(inode, pvec.pages[i], tindex++,
829 pagevec_release(&pvec);
835 xfs_vm_invalidatepage(
840 trace_xfs_invalidatepage(page->mapping->host, page, offset,
842 block_invalidatepage(page, offset, length);
846 * If the page has delalloc buffers on it, we need to punch them out before we
847 * invalidate the page. If we don't, we leave a stale delalloc mapping on the
848 * inode that can trip a BUG() in xfs_get_blocks() later on if a direct IO read
849 * is done on that same region - the delalloc extent is returned when none is
850 * supposed to be there.
852 * We prevent this by truncating away the delalloc regions on the page before
853 * invalidating it. Because they are delalloc, we can do this without needing a
854 * transaction. Indeed - if we get ENOSPC errors, we have to be able to do this
855 * truncation without a transaction as there is no space left for block
856 * reservation (typically why we see a ENOSPC in writeback).
858 * This is not a performance critical path, so for now just do the punching a
859 * buffer head at a time.
862 xfs_aops_discard_page(
865 struct inode *inode = page->mapping->host;
866 struct xfs_inode *ip = XFS_I(inode);
867 struct buffer_head *bh, *head;
868 loff_t offset = page_offset(page);
870 if (!xfs_check_page_type(page, XFS_IO_DELALLOC))
873 if (XFS_FORCED_SHUTDOWN(ip->i_mount))
876 xfs_alert(ip->i_mount,
877 "page discard on page %p, inode 0x%llx, offset %llu.",
878 page, ip->i_ino, offset);
880 xfs_ilock(ip, XFS_ILOCK_EXCL);
881 bh = head = page_buffers(page);
884 xfs_fileoff_t start_fsb;
886 if (!buffer_delay(bh))
889 start_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
890 error = xfs_bmap_punch_delalloc_range(ip, start_fsb, 1);
892 /* something screwed, just bail */
893 if (!XFS_FORCED_SHUTDOWN(ip->i_mount)) {
894 xfs_alert(ip->i_mount,
895 "page discard unable to remove delalloc mapping.");
900 offset += 1 << inode->i_blkbits;
902 } while ((bh = bh->b_this_page) != head);
904 xfs_iunlock(ip, XFS_ILOCK_EXCL);
906 xfs_vm_invalidatepage(page, 0, PAGE_CACHE_SIZE);
911 * Write out a dirty page.
913 * For delalloc space on the page we need to allocate space and flush it.
914 * For unwritten space on the page we need to start the conversion to
915 * regular allocated space.
916 * For any other dirty buffer heads on the page we should flush them.
921 struct writeback_control *wbc)
923 struct inode *inode = page->mapping->host;
924 struct buffer_head *bh, *head;
925 struct xfs_bmbt_irec imap;
926 xfs_ioend_t *ioend = NULL, *iohead = NULL;
929 __uint64_t end_offset;
930 pgoff_t end_index, last_index;
932 int err, imap_valid = 0, uptodate = 1;
936 trace_xfs_writepage(inode, page, 0, 0);
938 ASSERT(page_has_buffers(page));
941 * Refuse to write the page out if we are called from reclaim context.
943 * This avoids stack overflows when called from deeply used stacks in
944 * random callers for direct reclaim or memcg reclaim. We explicitly
945 * allow reclaim from kswapd as the stack usage there is relatively low.
947 * This should never happen except in the case of a VM regression so
950 if (WARN_ON_ONCE((current->flags & (PF_MEMALLOC|PF_KSWAPD)) ==
955 * Given that we do not allow direct reclaim to call us, we should
956 * never be called while in a filesystem transaction.
958 if (WARN_ON(current->flags & PF_FSTRANS))
961 /* Is this page beyond the end of the file? */
962 offset = i_size_read(inode);
963 end_index = offset >> PAGE_CACHE_SHIFT;
964 last_index = (offset - 1) >> PAGE_CACHE_SHIFT;
965 if (page->index >= end_index) {
966 unsigned offset_into_page = offset & (PAGE_CACHE_SIZE - 1);
969 * Skip the page if it is fully outside i_size, e.g. due to a
970 * truncate operation that is in progress. We must redirty the
971 * page so that reclaim stops reclaiming it. Otherwise
972 * xfs_vm_releasepage() is called on it and gets confused.
974 if (page->index >= end_index + 1 || offset_into_page == 0)
978 * The page straddles i_size. It must be zeroed out on each
979 * and every writepage invocation because it may be mmapped.
980 * "A file is mapped in multiples of the page size. For a file
981 * that is not a multiple of the page size, the remaining
982 * memory is zeroed when mapped, and writes to that region are
983 * not written out to the file."
985 zero_user_segment(page, offset_into_page, PAGE_CACHE_SIZE);
988 end_offset = min_t(unsigned long long,
989 (xfs_off_t)(page->index + 1) << PAGE_CACHE_SHIFT,
991 len = 1 << inode->i_blkbits;
993 bh = head = page_buffers(page);
994 offset = page_offset(page);
995 type = XFS_IO_OVERWRITE;
997 if (wbc->sync_mode == WB_SYNC_NONE)
1003 if (offset >= end_offset)
1005 if (!buffer_uptodate(bh))
1009 * set_page_dirty dirties all buffers in a page, independent
1010 * of their state. The dirty state however is entirely
1011 * meaningless for holes (!mapped && uptodate), so skip
1012 * buffers covering holes here.
1014 if (!buffer_mapped(bh) && buffer_uptodate(bh)) {
1019 if (buffer_unwritten(bh)) {
1020 if (type != XFS_IO_UNWRITTEN) {
1021 type = XFS_IO_UNWRITTEN;
1024 } else if (buffer_delay(bh)) {
1025 if (type != XFS_IO_DELALLOC) {
1026 type = XFS_IO_DELALLOC;
1029 } else if (buffer_uptodate(bh)) {
1030 if (type != XFS_IO_OVERWRITE) {
1031 type = XFS_IO_OVERWRITE;
1035 if (PageUptodate(page))
1036 ASSERT(buffer_mapped(bh));
1038 * This buffer is not uptodate and will not be
1039 * written to disk. Ensure that we will put any
1040 * subsequent writeable buffers into a new
1048 imap_valid = xfs_imap_valid(inode, &imap, offset);
1051 * If we didn't have a valid mapping then we need to
1052 * put the new mapping into a separate ioend structure.
1053 * This ensures non-contiguous extents always have
1054 * separate ioends, which is particularly important
1055 * for unwritten extent conversion at I/O completion
1059 err = xfs_map_blocks(inode, offset, &imap, type,
1063 imap_valid = xfs_imap_valid(inode, &imap, offset);
1067 if (type != XFS_IO_OVERWRITE)
1068 xfs_map_at_offset(inode, bh, &imap, offset);
1069 xfs_add_to_ioend(inode, bh, offset, type, &ioend,
1077 } while (offset += len, ((bh = bh->b_this_page) != head));
1079 if (uptodate && bh == head)
1080 SetPageUptodate(page);
1082 xfs_start_page_writeback(page, 1, count);
1084 /* if there is no IO to be submitted for this page, we are done */
1091 * Any errors from this point onwards need tobe reported through the IO
1092 * completion path as we have marked the initial page as under writeback
1096 xfs_off_t end_index;
1098 end_index = imap.br_startoff + imap.br_blockcount;
1101 end_index <<= inode->i_blkbits;
1104 end_index = (end_index - 1) >> PAGE_CACHE_SHIFT;
1106 /* check against file size */
1107 if (end_index > last_index)
1108 end_index = last_index;
1110 xfs_cluster_write(inode, page->index + 1, &imap, &ioend,
1116 * Reserve log space if we might write beyond the on-disk inode size.
1119 if (ioend->io_type != XFS_IO_UNWRITTEN && xfs_ioend_is_append(ioend))
1120 err = xfs_setfilesize_trans_alloc(ioend);
1122 xfs_submit_ioend(wbc, iohead, err);
1128 xfs_cancel_ioend(iohead);
1133 xfs_aops_discard_page(page);
1134 ClearPageUptodate(page);
1139 redirty_page_for_writepage(wbc, page);
1146 struct address_space *mapping,
1147 struct writeback_control *wbc)
1149 xfs_iflags_clear(XFS_I(mapping->host), XFS_ITRUNCATED);
1150 return generic_writepages(mapping, wbc);
1154 * Called to move a page into cleanable state - and from there
1155 * to be released. The page should already be clean. We always
1156 * have buffer heads in this call.
1158 * Returns 1 if the page is ok to release, 0 otherwise.
1165 int delalloc, unwritten;
1167 trace_xfs_releasepage(page->mapping->host, page, 0, 0);
1169 xfs_count_page_state(page, &delalloc, &unwritten);
1171 if (WARN_ON(delalloc))
1173 if (WARN_ON(unwritten))
1176 return try_to_free_buffers(page);
1181 struct inode *inode,
1183 struct buffer_head *bh_result,
1187 struct xfs_inode *ip = XFS_I(inode);
1188 struct xfs_mount *mp = ip->i_mount;
1189 xfs_fileoff_t offset_fsb, end_fsb;
1192 struct xfs_bmbt_irec imap;
1198 if (XFS_FORCED_SHUTDOWN(mp))
1199 return -XFS_ERROR(EIO);
1201 offset = (xfs_off_t)iblock << inode->i_blkbits;
1202 ASSERT(bh_result->b_size >= (1 << inode->i_blkbits));
1203 size = bh_result->b_size;
1205 if (!create && direct && offset >= i_size_read(inode))
1209 * Direct I/O is usually done on preallocated files, so try getting
1210 * a block mapping without an exclusive lock first. For buffered
1211 * writes we already have the exclusive iolock anyway, so avoiding
1212 * a lock roundtrip here by taking the ilock exclusive from the
1213 * beginning is a useful micro optimization.
1215 if (create && !direct) {
1216 lockmode = XFS_ILOCK_EXCL;
1217 xfs_ilock(ip, lockmode);
1219 lockmode = xfs_ilock_map_shared(ip);
1222 ASSERT(offset <= mp->m_super->s_maxbytes);
1223 if (offset + size > mp->m_super->s_maxbytes)
1224 size = mp->m_super->s_maxbytes - offset;
1225 end_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)offset + size);
1226 offset_fsb = XFS_B_TO_FSBT(mp, offset);
1228 error = xfs_bmapi_read(ip, offset_fsb, end_fsb - offset_fsb,
1229 &imap, &nimaps, XFS_BMAPI_ENTIRE);
1235 (imap.br_startblock == HOLESTARTBLOCK ||
1236 imap.br_startblock == DELAYSTARTBLOCK))) {
1237 if (direct || xfs_get_extsz_hint(ip)) {
1239 * Drop the ilock in preparation for starting the block
1240 * allocation transaction. It will be retaken
1241 * exclusively inside xfs_iomap_write_direct for the
1242 * actual allocation.
1244 xfs_iunlock(ip, lockmode);
1245 error = xfs_iomap_write_direct(ip, offset, size,
1252 * Delalloc reservations do not require a transaction,
1253 * we can go on without dropping the lock here. If we
1254 * are allocating a new delalloc block, make sure that
1255 * we set the new flag so that we mark the buffer new so
1256 * that we know that it is newly allocated if the write
1259 if (nimaps && imap.br_startblock == HOLESTARTBLOCK)
1261 error = xfs_iomap_write_delay(ip, offset, size, &imap);
1265 xfs_iunlock(ip, lockmode);
1268 trace_xfs_get_blocks_alloc(ip, offset, size, 0, &imap);
1269 } else if (nimaps) {
1270 trace_xfs_get_blocks_found(ip, offset, size, 0, &imap);
1271 xfs_iunlock(ip, lockmode);
1273 trace_xfs_get_blocks_notfound(ip, offset, size);
1277 if (imap.br_startblock != HOLESTARTBLOCK &&
1278 imap.br_startblock != DELAYSTARTBLOCK) {
1280 * For unwritten extents do not report a disk address on
1281 * the read case (treat as if we're reading into a hole).
1283 if (create || !ISUNWRITTEN(&imap))
1284 xfs_map_buffer(inode, bh_result, &imap, offset);
1285 if (create && ISUNWRITTEN(&imap)) {
1287 bh_result->b_private = inode;
1288 set_buffer_defer_completion(bh_result);
1290 set_buffer_unwritten(bh_result);
1295 * If this is a realtime file, data may be on a different device.
1296 * to that pointed to from the buffer_head b_bdev currently.
1298 bh_result->b_bdev = xfs_find_bdev_for_inode(inode);
1301 * If we previously allocated a block out beyond eof and we are now
1302 * coming back to use it then we will need to flag it as new even if it
1303 * has a disk address.
1305 * With sub-block writes into unwritten extents we also need to mark
1306 * the buffer as new so that the unwritten parts of the buffer gets
1310 ((!buffer_mapped(bh_result) && !buffer_uptodate(bh_result)) ||
1311 (offset >= i_size_read(inode)) ||
1312 (new || ISUNWRITTEN(&imap))))
1313 set_buffer_new(bh_result);
1315 if (imap.br_startblock == DELAYSTARTBLOCK) {
1318 set_buffer_uptodate(bh_result);
1319 set_buffer_mapped(bh_result);
1320 set_buffer_delay(bh_result);
1325 * If this is O_DIRECT or the mpage code calling tell them how large
1326 * the mapping is, so that we can avoid repeated get_blocks calls.
1328 if (direct || size > (1 << inode->i_blkbits)) {
1329 xfs_off_t mapping_size;
1331 mapping_size = imap.br_startoff + imap.br_blockcount - iblock;
1332 mapping_size <<= inode->i_blkbits;
1334 ASSERT(mapping_size > 0);
1335 if (mapping_size > size)
1336 mapping_size = size;
1337 if (mapping_size > LONG_MAX)
1338 mapping_size = LONG_MAX;
1340 bh_result->b_size = mapping_size;
1346 xfs_iunlock(ip, lockmode);
1352 struct inode *inode,
1354 struct buffer_head *bh_result,
1357 return __xfs_get_blocks(inode, iblock, bh_result, create, 0);
1361 xfs_get_blocks_direct(
1362 struct inode *inode,
1364 struct buffer_head *bh_result,
1367 return __xfs_get_blocks(inode, iblock, bh_result, create, 1);
1371 * Complete a direct I/O write request.
1373 * If the private argument is non-NULL __xfs_get_blocks signals us that we
1374 * need to issue a transaction to convert the range from unwritten to written
1375 * extents. In case this is regular synchronous I/O we just call xfs_end_io
1376 * to do this and we are done. But in case this was a successful AIO
1377 * request this handler is called from interrupt context, from which we
1378 * can't start transactions. In that case offload the I/O completion to
1379 * the workqueues we also use for buffered I/O completion.
1382 xfs_end_io_direct_write(
1388 struct xfs_ioend *ioend = iocb->private;
1391 * While the generic direct I/O code updates the inode size, it does
1392 * so only after the end_io handler is called, which means our
1393 * end_io handler thinks the on-disk size is outside the in-core
1394 * size. To prevent this just update it a little bit earlier here.
1396 if (offset + size > i_size_read(ioend->io_inode))
1397 i_size_write(ioend->io_inode, offset + size);
1400 * blockdev_direct_IO can return an error even after the I/O
1401 * completion handler was called. Thus we need to protect
1402 * against double-freeing.
1404 iocb->private = NULL;
1406 ioend->io_offset = offset;
1407 ioend->io_size = size;
1408 if (private && size > 0)
1409 ioend->io_type = XFS_IO_UNWRITTEN;
1411 xfs_finish_ioend_sync(ioend);
1418 const struct iovec *iov,
1420 unsigned long nr_segs)
1422 struct inode *inode = iocb->ki_filp->f_mapping->host;
1423 struct block_device *bdev = xfs_find_bdev_for_inode(inode);
1424 struct xfs_ioend *ioend = NULL;
1428 size_t size = iov_length(iov, nr_segs);
1431 * We cannot preallocate a size update transaction here as we
1432 * don't know whether allocation is necessary or not. Hence we
1433 * can only tell IO completion that one is necessary if we are
1434 * not doing unwritten extent conversion.
1436 iocb->private = ioend = xfs_alloc_ioend(inode, XFS_IO_DIRECT);
1437 if (offset + size > XFS_I(inode)->i_d.di_size)
1438 ioend->io_isdirect = 1;
1440 ret = __blockdev_direct_IO(rw, iocb, inode, bdev, iov,
1442 xfs_get_blocks_direct,
1443 xfs_end_io_direct_write, NULL, 0);
1444 if (ret != -EIOCBQUEUED && iocb->private)
1445 goto out_destroy_ioend;
1447 ret = __blockdev_direct_IO(rw, iocb, inode, bdev, iov,
1449 xfs_get_blocks_direct,
1456 xfs_destroy_ioend(ioend);
1461 * Punch out the delalloc blocks we have already allocated.
1463 * Don't bother with xfs_setattr given that nothing can have made it to disk yet
1464 * as the page is still locked at this point.
1467 xfs_vm_kill_delalloc_range(
1468 struct inode *inode,
1472 struct xfs_inode *ip = XFS_I(inode);
1473 xfs_fileoff_t start_fsb;
1474 xfs_fileoff_t end_fsb;
1477 start_fsb = XFS_B_TO_FSB(ip->i_mount, start);
1478 end_fsb = XFS_B_TO_FSB(ip->i_mount, end);
1479 if (end_fsb <= start_fsb)
1482 xfs_ilock(ip, XFS_ILOCK_EXCL);
1483 error = xfs_bmap_punch_delalloc_range(ip, start_fsb,
1484 end_fsb - start_fsb);
1486 /* something screwed, just bail */
1487 if (!XFS_FORCED_SHUTDOWN(ip->i_mount)) {
1488 xfs_alert(ip->i_mount,
1489 "xfs_vm_write_failed: unable to clean up ino %lld",
1493 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1497 xfs_vm_write_failed(
1498 struct inode *inode,
1503 loff_t block_offset;
1506 loff_t from = pos & (PAGE_CACHE_SIZE - 1);
1507 loff_t to = from + len;
1508 struct buffer_head *bh, *head;
1511 * The request pos offset might be 32 or 64 bit, this is all fine
1512 * on 64-bit platform. However, for 64-bit pos request on 32-bit
1513 * platform, the high 32-bit will be masked off if we evaluate the
1514 * block_offset via (pos & PAGE_MASK) because the PAGE_MASK is
1515 * 0xfffff000 as an unsigned long, hence the result is incorrect
1516 * which could cause the following ASSERT failed in most cases.
1517 * In order to avoid this, we can evaluate the block_offset of the
1518 * start of the page by using shifts rather than masks the mismatch
1521 block_offset = (pos >> PAGE_CACHE_SHIFT) << PAGE_CACHE_SHIFT;
1523 ASSERT(block_offset + from == pos);
1525 head = page_buffers(page);
1527 for (bh = head; bh != head || !block_start;
1528 bh = bh->b_this_page, block_start = block_end,
1529 block_offset += bh->b_size) {
1530 block_end = block_start + bh->b_size;
1532 /* skip buffers before the write */
1533 if (block_end <= from)
1536 /* if the buffer is after the write, we're done */
1537 if (block_start >= to)
1540 if (!buffer_delay(bh))
1543 if (!buffer_new(bh) && block_offset < i_size_read(inode))
1546 xfs_vm_kill_delalloc_range(inode, block_offset,
1547 block_offset + bh->b_size);
1553 * This used to call block_write_begin(), but it unlocks and releases the page
1554 * on error, and we need that page to be able to punch stale delalloc blocks out
1555 * on failure. hence we copy-n-waste it here and call xfs_vm_write_failed() at
1556 * the appropriate point.
1561 struct address_space *mapping,
1565 struct page **pagep,
1568 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1572 ASSERT(len <= PAGE_CACHE_SIZE);
1574 page = grab_cache_page_write_begin(mapping, index,
1575 flags | AOP_FLAG_NOFS);
1579 status = __block_write_begin(page, pos, len, xfs_get_blocks);
1580 if (unlikely(status)) {
1581 struct inode *inode = mapping->host;
1583 xfs_vm_write_failed(inode, page, pos, len);
1586 if (pos + len > i_size_read(inode))
1587 truncate_pagecache(inode, i_size_read(inode));
1589 page_cache_release(page);
1598 * On failure, we only need to kill delalloc blocks beyond EOF because they
1599 * will never be written. For blocks within EOF, generic_write_end() zeros them
1600 * so they are safe to leave alone and be written with all the other valid data.
1605 struct address_space *mapping,
1614 ASSERT(len <= PAGE_CACHE_SIZE);
1616 ret = generic_write_end(file, mapping, pos, len, copied, page, fsdata);
1617 if (unlikely(ret < len)) {
1618 struct inode *inode = mapping->host;
1619 size_t isize = i_size_read(inode);
1620 loff_t to = pos + len;
1623 truncate_pagecache(inode, isize);
1624 xfs_vm_kill_delalloc_range(inode, isize, to);
1632 struct address_space *mapping,
1635 struct inode *inode = (struct inode *)mapping->host;
1636 struct xfs_inode *ip = XFS_I(inode);
1638 trace_xfs_vm_bmap(XFS_I(inode));
1639 xfs_ilock(ip, XFS_IOLOCK_SHARED);
1640 filemap_write_and_wait(mapping);
1641 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
1642 return generic_block_bmap(mapping, block, xfs_get_blocks);
1647 struct file *unused,
1650 return mpage_readpage(page, xfs_get_blocks);
1655 struct file *unused,
1656 struct address_space *mapping,
1657 struct list_head *pages,
1660 return mpage_readpages(mapping, pages, nr_pages, xfs_get_blocks);
1663 const struct address_space_operations xfs_address_space_operations = {
1664 .readpage = xfs_vm_readpage,
1665 .readpages = xfs_vm_readpages,
1666 .writepage = xfs_vm_writepage,
1667 .writepages = xfs_vm_writepages,
1668 .releasepage = xfs_vm_releasepage,
1669 .invalidatepage = xfs_vm_invalidatepage,
1670 .write_begin = xfs_vm_write_begin,
1671 .write_end = xfs_vm_write_end,
1672 .bmap = xfs_vm_bmap,
1673 .direct_IO = xfs_vm_direct_IO,
1674 .migratepage = buffer_migrate_page,
1675 .is_partially_uptodate = block_is_partially_uptodate,
1676 .error_remove_page = generic_error_remove_page,