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_shared.h"
20 #include "xfs_format.h"
21 #include "xfs_log_format.h"
22 #include "xfs_trans_resv.h"
25 #include "xfs_mount.h"
26 #include "xfs_inode.h"
27 #include "xfs_trans.h"
28 #include "xfs_inode_item.h"
29 #include "xfs_alloc.h"
30 #include "xfs_error.h"
31 #include "xfs_iomap.h"
32 #include "xfs_trace.h"
34 #include "xfs_bmap_util.h"
35 #include "xfs_bmap_btree.h"
36 #include "xfs_dinode.h"
37 #include <linux/aio.h>
38 #include <linux/gfp.h>
39 #include <linux/mpage.h>
40 #include <linux/pagevec.h>
41 #include <linux/writeback.h>
49 struct buffer_head *bh, *head;
51 *delalloc = *unwritten = 0;
53 bh = head = page_buffers(page);
55 if (buffer_unwritten(bh))
57 else if (buffer_delay(bh))
59 } while ((bh = bh->b_this_page) != head);
62 STATIC struct block_device *
63 xfs_find_bdev_for_inode(
66 struct xfs_inode *ip = XFS_I(inode);
67 struct xfs_mount *mp = ip->i_mount;
69 if (XFS_IS_REALTIME_INODE(ip))
70 return mp->m_rtdev_targp->bt_bdev;
72 return mp->m_ddev_targp->bt_bdev;
76 * We're now finished for good with this ioend structure.
77 * Update the page state via the associated buffer_heads,
78 * release holds on the inode and bio, and finally free
79 * up memory. Do not use the ioend after this.
85 struct buffer_head *bh, *next;
87 for (bh = ioend->io_buffer_head; bh; bh = next) {
89 bh->b_end_io(bh, !ioend->io_error);
92 mempool_free(ioend, xfs_ioend_pool);
96 * Fast and loose check if this write could update the on-disk inode size.
98 static inline bool xfs_ioend_is_append(struct xfs_ioend *ioend)
100 return ioend->io_offset + ioend->io_size >
101 XFS_I(ioend->io_inode)->i_d.di_size;
105 xfs_setfilesize_trans_alloc(
106 struct xfs_ioend *ioend)
108 struct xfs_mount *mp = XFS_I(ioend->io_inode)->i_mount;
109 struct xfs_trans *tp;
112 tp = xfs_trans_alloc(mp, XFS_TRANS_FSYNC_TS);
114 error = xfs_trans_reserve(tp, &M_RES(mp)->tr_fsyncts, 0, 0);
116 xfs_trans_cancel(tp, 0);
120 ioend->io_append_trans = tp;
123 * We may pass freeze protection with a transaction. So tell lockdep
126 rwsem_release(&ioend->io_inode->i_sb->s_writers.lock_map[SB_FREEZE_FS-1],
129 * We hand off the transaction to the completion thread now, so
130 * clear the flag here.
132 current_restore_flags_nested(&tp->t_pflags, PF_FSTRANS);
137 * Update on-disk file size now that data has been written to disk.
141 struct xfs_ioend *ioend)
143 struct xfs_inode *ip = XFS_I(ioend->io_inode);
144 struct xfs_trans *tp = ioend->io_append_trans;
148 * The transaction may have been allocated in the I/O submission thread,
149 * thus we need to mark ourselves as beeing in a transaction manually.
150 * Similarly for freeze protection.
152 current_set_flags_nested(&tp->t_pflags, PF_FSTRANS);
153 rwsem_acquire_read(&VFS_I(ip)->i_sb->s_writers.lock_map[SB_FREEZE_FS-1],
156 xfs_ilock(ip, XFS_ILOCK_EXCL);
157 isize = xfs_new_eof(ip, ioend->io_offset + ioend->io_size);
159 xfs_iunlock(ip, XFS_ILOCK_EXCL);
160 xfs_trans_cancel(tp, 0);
164 trace_xfs_setfilesize(ip, ioend->io_offset, ioend->io_size);
166 ip->i_d.di_size = isize;
167 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
168 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
170 return xfs_trans_commit(tp, 0);
174 * Schedule IO completion handling on the final put of an ioend.
176 * If there is no work to do we might as well call it a day and free the
181 struct xfs_ioend *ioend)
183 if (atomic_dec_and_test(&ioend->io_remaining)) {
184 struct xfs_mount *mp = XFS_I(ioend->io_inode)->i_mount;
186 if (ioend->io_type == XFS_IO_UNWRITTEN)
187 queue_work(mp->m_unwritten_workqueue, &ioend->io_work);
188 else if (ioend->io_append_trans ||
189 (ioend->io_isdirect && xfs_ioend_is_append(ioend)))
190 queue_work(mp->m_data_workqueue, &ioend->io_work);
192 xfs_destroy_ioend(ioend);
197 * IO write completion.
201 struct work_struct *work)
203 xfs_ioend_t *ioend = container_of(work, xfs_ioend_t, io_work);
204 struct xfs_inode *ip = XFS_I(ioend->io_inode);
207 if (XFS_FORCED_SHUTDOWN(ip->i_mount)) {
208 ioend->io_error = -EIO;
215 * For unwritten extents we need to issue transactions to convert a
216 * range to normal written extens after the data I/O has finished.
218 if (ioend->io_type == XFS_IO_UNWRITTEN) {
219 error = xfs_iomap_write_unwritten(ip, ioend->io_offset,
221 } else if (ioend->io_isdirect && xfs_ioend_is_append(ioend)) {
223 * For direct I/O we do not know if we need to allocate blocks
224 * or not so we can't preallocate an append transaction as that
225 * results in nested reservations and log space deadlocks. Hence
226 * allocate the transaction here. While this is sub-optimal and
227 * can block IO completion for some time, we're stuck with doing
228 * it this way until we can pass the ioend to the direct IO
229 * allocation callbacks and avoid nesting that way.
231 error = xfs_setfilesize_trans_alloc(ioend);
234 error = xfs_setfilesize(ioend);
235 } else if (ioend->io_append_trans) {
236 error = xfs_setfilesize(ioend);
238 ASSERT(!xfs_ioend_is_append(ioend));
243 ioend->io_error = error;
244 xfs_destroy_ioend(ioend);
248 * Call IO completion handling in caller context on the final put of an ioend.
251 xfs_finish_ioend_sync(
252 struct xfs_ioend *ioend)
254 if (atomic_dec_and_test(&ioend->io_remaining))
255 xfs_end_io(&ioend->io_work);
259 * Allocate and initialise an IO completion structure.
260 * We need to track unwritten extent write completion here initially.
261 * We'll need to extend this for updating the ondisk inode size later
271 ioend = mempool_alloc(xfs_ioend_pool, GFP_NOFS);
274 * Set the count to 1 initially, which will prevent an I/O
275 * completion callback from happening before we have started
276 * all the I/O from calling the completion routine too early.
278 atomic_set(&ioend->io_remaining, 1);
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_append_trans = NULL;
290 INIT_WORK(&ioend->io_work, xfs_end_io);
298 struct xfs_bmbt_irec *imap,
302 struct xfs_inode *ip = XFS_I(inode);
303 struct xfs_mount *mp = ip->i_mount;
304 ssize_t count = 1 << inode->i_blkbits;
305 xfs_fileoff_t offset_fsb, end_fsb;
307 int bmapi_flags = XFS_BMAPI_ENTIRE;
310 if (XFS_FORCED_SHUTDOWN(mp))
313 if (type == XFS_IO_UNWRITTEN)
314 bmapi_flags |= XFS_BMAPI_IGSTATE;
316 if (!xfs_ilock_nowait(ip, XFS_ILOCK_SHARED)) {
319 xfs_ilock(ip, XFS_ILOCK_SHARED);
322 ASSERT(ip->i_d.di_format != XFS_DINODE_FMT_BTREE ||
323 (ip->i_df.if_flags & XFS_IFEXTENTS));
324 ASSERT(offset <= mp->m_super->s_maxbytes);
326 if (offset + count > mp->m_super->s_maxbytes)
327 count = mp->m_super->s_maxbytes - offset;
328 end_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)offset + count);
329 offset_fsb = XFS_B_TO_FSBT(mp, offset);
330 error = xfs_bmapi_read(ip, offset_fsb, end_fsb - offset_fsb,
331 imap, &nimaps, bmapi_flags);
332 xfs_iunlock(ip, XFS_ILOCK_SHARED);
337 if (type == XFS_IO_DELALLOC &&
338 (!nimaps || isnullstartblock(imap->br_startblock))) {
339 error = xfs_iomap_write_allocate(ip, offset, imap);
341 trace_xfs_map_blocks_alloc(ip, offset, count, type, imap);
346 if (type == XFS_IO_UNWRITTEN) {
348 ASSERT(imap->br_startblock != HOLESTARTBLOCK);
349 ASSERT(imap->br_startblock != DELAYSTARTBLOCK);
353 trace_xfs_map_blocks_found(ip, offset, count, type, imap);
360 struct xfs_bmbt_irec *imap,
363 offset >>= inode->i_blkbits;
365 return offset >= imap->br_startoff &&
366 offset < imap->br_startoff + imap->br_blockcount;
370 * BIO completion handler for buffered IO.
377 xfs_ioend_t *ioend = bio->bi_private;
379 ASSERT(atomic_read(&bio->bi_cnt) >= 1);
380 ioend->io_error = test_bit(BIO_UPTODATE, &bio->bi_flags) ? 0 : error;
382 /* Toss bio and pass work off to an xfsdatad thread */
383 bio->bi_private = NULL;
384 bio->bi_end_io = NULL;
387 xfs_finish_ioend(ioend);
391 xfs_submit_ioend_bio(
392 struct writeback_control *wbc,
396 atomic_inc(&ioend->io_remaining);
397 bio->bi_private = ioend;
398 bio->bi_end_io = xfs_end_bio;
399 submit_bio(wbc->sync_mode == WB_SYNC_ALL ? WRITE_SYNC : WRITE, bio);
404 struct buffer_head *bh)
406 int nvecs = bio_get_nr_vecs(bh->b_bdev);
407 struct bio *bio = bio_alloc(GFP_NOIO, nvecs);
409 ASSERT(bio->bi_private == NULL);
410 bio->bi_iter.bi_sector = bh->b_blocknr * (bh->b_size >> 9);
411 bio->bi_bdev = bh->b_bdev;
416 xfs_start_buffer_writeback(
417 struct buffer_head *bh)
419 ASSERT(buffer_mapped(bh));
420 ASSERT(buffer_locked(bh));
421 ASSERT(!buffer_delay(bh));
422 ASSERT(!buffer_unwritten(bh));
424 mark_buffer_async_write(bh);
425 set_buffer_uptodate(bh);
426 clear_buffer_dirty(bh);
430 xfs_start_page_writeback(
435 ASSERT(PageLocked(page));
436 ASSERT(!PageWriteback(page));
438 clear_page_dirty_for_io(page);
439 set_page_writeback(page);
441 /* If no buffers on the page are to be written, finish it here */
443 end_page_writeback(page);
446 static inline int xfs_bio_add_buffer(struct bio *bio, struct buffer_head *bh)
448 return bio_add_page(bio, bh->b_page, bh->b_size, bh_offset(bh));
452 * Submit all of the bios for all of the ioends we have saved up, covering the
453 * initial writepage page and also any probed pages.
455 * Because we may have multiple ioends spanning a page, we need to start
456 * writeback on all the buffers before we submit them for I/O. If we mark the
457 * buffers as we got, then we can end up with a page that only has buffers
458 * marked async write and I/O complete on can occur before we mark the other
459 * buffers async write.
461 * The end result of this is that we trip a bug in end_page_writeback() because
462 * we call it twice for the one page as the code in end_buffer_async_write()
463 * assumes that all buffers on the page are started at the same time.
465 * The fix is two passes across the ioend list - one to start writeback on the
466 * buffer_heads, and then submit them for I/O on the second pass.
468 * If @fail is non-zero, it means that we have a situation where some part of
469 * the submission process has failed after we have marked paged for writeback
470 * and unlocked them. In this situation, we need to fail the ioend chain rather
471 * than submit it to IO. This typically only happens on a filesystem shutdown.
475 struct writeback_control *wbc,
479 xfs_ioend_t *head = ioend;
481 struct buffer_head *bh;
483 sector_t lastblock = 0;
485 /* Pass 1 - start writeback */
487 next = ioend->io_list;
488 for (bh = ioend->io_buffer_head; bh; bh = bh->b_private)
489 xfs_start_buffer_writeback(bh);
490 } while ((ioend = next) != NULL);
492 /* Pass 2 - submit I/O */
495 next = ioend->io_list;
499 * If we are failing the IO now, just mark the ioend with an
500 * error and finish it. This will run IO completion immediately
501 * as there is only one reference to the ioend at this point in
505 ioend->io_error = fail;
506 xfs_finish_ioend(ioend);
510 for (bh = ioend->io_buffer_head; bh; bh = bh->b_private) {
514 bio = xfs_alloc_ioend_bio(bh);
515 } else if (bh->b_blocknr != lastblock + 1) {
516 xfs_submit_ioend_bio(wbc, ioend, bio);
520 if (xfs_bio_add_buffer(bio, bh) != bh->b_size) {
521 xfs_submit_ioend_bio(wbc, ioend, bio);
525 lastblock = bh->b_blocknr;
528 xfs_submit_ioend_bio(wbc, ioend, bio);
529 xfs_finish_ioend(ioend);
530 } while ((ioend = next) != NULL);
534 * Cancel submission of all buffer_heads so far in this endio.
535 * Toss the endio too. Only ever called for the initial page
536 * in a writepage request, so only ever one page.
543 struct buffer_head *bh, *next_bh;
546 next = ioend->io_list;
547 bh = ioend->io_buffer_head;
549 next_bh = bh->b_private;
550 clear_buffer_async_write(bh);
552 * The unwritten flag is cleared when added to the
553 * ioend. We're not submitting for I/O so mark the
554 * buffer unwritten again for next time around.
556 if (ioend->io_type == XFS_IO_UNWRITTEN)
557 set_buffer_unwritten(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 contains at least one buffer of a given @type.
643 * If @check_all_buffers is true, then we walk all the buffers in the page to
644 * try to find one of the type passed in. If it is not set, then the caller only
645 * needs to check the first buffer on the page for a match.
651 bool check_all_buffers)
653 struct buffer_head *bh;
654 struct buffer_head *head;
656 if (PageWriteback(page))
660 if (!page_has_buffers(page))
663 bh = head = page_buffers(page);
665 if (buffer_unwritten(bh)) {
666 if (type == XFS_IO_UNWRITTEN)
668 } else if (buffer_delay(bh)) {
669 if (type == XFS_IO_DELALLOC)
671 } else if (buffer_dirty(bh) && buffer_mapped(bh)) {
672 if (type == XFS_IO_OVERWRITE)
676 /* If we are only checking the first buffer, we are done now. */
677 if (!check_all_buffers)
679 } while ((bh = bh->b_this_page) != head);
685 * Allocate & map buffers for page given the extent map. Write it out.
686 * except for the original page of a writepage, this is called on
687 * delalloc/unwritten pages only, for the original page it is possible
688 * that the page has no mapping at all.
695 struct xfs_bmbt_irec *imap,
696 xfs_ioend_t **ioendp,
697 struct writeback_control *wbc)
699 struct buffer_head *bh, *head;
700 xfs_off_t end_offset;
701 unsigned long p_offset;
704 int count = 0, done = 0, uptodate = 1;
705 xfs_off_t offset = page_offset(page);
707 if (page->index != tindex)
709 if (!trylock_page(page))
711 if (PageWriteback(page))
712 goto fail_unlock_page;
713 if (page->mapping != inode->i_mapping)
714 goto fail_unlock_page;
715 if (!xfs_check_page_type(page, (*ioendp)->io_type, false))
716 goto fail_unlock_page;
719 * page_dirty is initially a count of buffers on the page before
720 * EOF and is decremented as we move each into a cleanable state.
724 * End offset is the highest offset that this page should represent.
725 * If we are on the last page, (end_offset & (PAGE_CACHE_SIZE - 1))
726 * will evaluate non-zero and be less than PAGE_CACHE_SIZE and
727 * hence give us the correct page_dirty count. On any other page,
728 * it will be zero and in that case we need page_dirty to be the
729 * count of buffers on the page.
731 end_offset = min_t(unsigned long long,
732 (xfs_off_t)(page->index + 1) << PAGE_CACHE_SHIFT,
736 * If the current map does not span the entire page we are about to try
737 * to write, then give up. The only way we can write a page that spans
738 * multiple mappings in a single writeback iteration is via the
739 * xfs_vm_writepage() function. Data integrity writeback requires the
740 * entire page to be written in a single attempt, otherwise the part of
741 * the page we don't write here doesn't get written as part of the data
744 * For normal writeback, we also don't attempt to write partial pages
745 * here as it simply means that write_cache_pages() will see it under
746 * writeback and ignore the page until some point in the future, at
747 * which time this will be the only page in the file that needs
748 * writeback. Hence for more optimal IO patterns, we should always
749 * avoid partial page writeback due to multiple mappings on a page here.
751 if (!xfs_imap_valid(inode, imap, end_offset))
752 goto fail_unlock_page;
754 len = 1 << inode->i_blkbits;
755 p_offset = min_t(unsigned long, end_offset & (PAGE_CACHE_SIZE - 1),
757 p_offset = p_offset ? roundup(p_offset, len) : PAGE_CACHE_SIZE;
758 page_dirty = p_offset / len;
761 * The moment we find a buffer that doesn't match our current type
762 * specification or can't be written, abort the loop and start
763 * writeback. As per the above xfs_imap_valid() check, only
764 * xfs_vm_writepage() can handle partial page writeback fully - we are
765 * limited here to the buffers that are contiguous with the current
766 * ioend, and hence a buffer we can't write breaks that contiguity and
767 * we have to defer the rest of the IO to xfs_vm_writepage().
769 bh = head = page_buffers(page);
771 if (offset >= end_offset)
773 if (!buffer_uptodate(bh))
775 if (!(PageUptodate(page) || buffer_uptodate(bh))) {
780 if (buffer_unwritten(bh) || buffer_delay(bh) ||
782 if (buffer_unwritten(bh))
783 type = XFS_IO_UNWRITTEN;
784 else if (buffer_delay(bh))
785 type = XFS_IO_DELALLOC;
787 type = XFS_IO_OVERWRITE;
790 * imap should always be valid because of the above
791 * partial page end_offset check on the imap.
793 ASSERT(xfs_imap_valid(inode, imap, offset));
796 if (type != XFS_IO_OVERWRITE)
797 xfs_map_at_offset(inode, bh, imap, offset);
798 xfs_add_to_ioend(inode, bh, offset, type,
807 } while (offset += len, (bh = bh->b_this_page) != head);
809 if (uptodate && bh == head)
810 SetPageUptodate(page);
813 if (--wbc->nr_to_write <= 0 &&
814 wbc->sync_mode == WB_SYNC_NONE)
817 xfs_start_page_writeback(page, !page_dirty, count);
827 * Convert & write out a cluster of pages in the same extent as defined
828 * by mp and following the start page.
834 struct xfs_bmbt_irec *imap,
835 xfs_ioend_t **ioendp,
836 struct writeback_control *wbc,
842 pagevec_init(&pvec, 0);
843 while (!done && tindex <= tlast) {
844 unsigned len = min_t(pgoff_t, PAGEVEC_SIZE, tlast - tindex + 1);
846 if (!pagevec_lookup(&pvec, inode->i_mapping, tindex, len))
849 for (i = 0; i < pagevec_count(&pvec); i++) {
850 done = xfs_convert_page(inode, pvec.pages[i], tindex++,
856 pagevec_release(&pvec);
862 xfs_vm_invalidatepage(
867 trace_xfs_invalidatepage(page->mapping->host, page, offset,
869 block_invalidatepage(page, offset, length);
873 * If the page has delalloc buffers on it, we need to punch them out before we
874 * invalidate the page. If we don't, we leave a stale delalloc mapping on the
875 * inode that can trip a BUG() in xfs_get_blocks() later on if a direct IO read
876 * is done on that same region - the delalloc extent is returned when none is
877 * supposed to be there.
879 * We prevent this by truncating away the delalloc regions on the page before
880 * invalidating it. Because they are delalloc, we can do this without needing a
881 * transaction. Indeed - if we get ENOSPC errors, we have to be able to do this
882 * truncation without a transaction as there is no space left for block
883 * reservation (typically why we see a ENOSPC in writeback).
885 * This is not a performance critical path, so for now just do the punching a
886 * buffer head at a time.
889 xfs_aops_discard_page(
892 struct inode *inode = page->mapping->host;
893 struct xfs_inode *ip = XFS_I(inode);
894 struct buffer_head *bh, *head;
895 loff_t offset = page_offset(page);
897 if (!xfs_check_page_type(page, XFS_IO_DELALLOC, true))
900 if (XFS_FORCED_SHUTDOWN(ip->i_mount))
903 xfs_alert(ip->i_mount,
904 "page discard on page %p, inode 0x%llx, offset %llu.",
905 page, ip->i_ino, offset);
907 xfs_ilock(ip, XFS_ILOCK_EXCL);
908 bh = head = page_buffers(page);
911 xfs_fileoff_t start_fsb;
913 if (!buffer_delay(bh))
916 start_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
917 error = xfs_bmap_punch_delalloc_range(ip, start_fsb, 1);
919 /* something screwed, just bail */
920 if (!XFS_FORCED_SHUTDOWN(ip->i_mount)) {
921 xfs_alert(ip->i_mount,
922 "page discard unable to remove delalloc mapping.");
927 offset += 1 << inode->i_blkbits;
929 } while ((bh = bh->b_this_page) != head);
931 xfs_iunlock(ip, XFS_ILOCK_EXCL);
933 xfs_vm_invalidatepage(page, 0, PAGE_CACHE_SIZE);
938 * Write out a dirty page.
940 * For delalloc space on the page we need to allocate space and flush it.
941 * For unwritten space on the page we need to start the conversion to
942 * regular allocated space.
943 * For any other dirty buffer heads on the page we should flush them.
948 struct writeback_control *wbc)
950 struct inode *inode = page->mapping->host;
951 struct buffer_head *bh, *head;
952 struct xfs_bmbt_irec imap;
953 xfs_ioend_t *ioend = NULL, *iohead = NULL;
956 __uint64_t end_offset;
957 pgoff_t end_index, last_index;
959 int err, imap_valid = 0, uptodate = 1;
963 trace_xfs_writepage(inode, page, 0, 0);
965 ASSERT(page_has_buffers(page));
968 * Refuse to write the page out if we are called from reclaim context.
970 * This avoids stack overflows when called from deeply used stacks in
971 * random callers for direct reclaim or memcg reclaim. We explicitly
972 * allow reclaim from kswapd as the stack usage there is relatively low.
974 * This should never happen except in the case of a VM regression so
977 if (WARN_ON_ONCE((current->flags & (PF_MEMALLOC|PF_KSWAPD)) ==
982 * Given that we do not allow direct reclaim to call us, we should
983 * never be called while in a filesystem transaction.
985 if (WARN_ON_ONCE(current->flags & PF_FSTRANS))
988 /* Is this page beyond the end of the file? */
989 offset = i_size_read(inode);
990 end_index = offset >> PAGE_CACHE_SHIFT;
991 last_index = (offset - 1) >> PAGE_CACHE_SHIFT;
994 * The page index is less than the end_index, adjust the end_offset
995 * to the highest offset that this page should represent.
996 * -----------------------------------------------------
997 * | file mapping | <EOF> |
998 * -----------------------------------------------------
999 * | Page ... | Page N-2 | Page N-1 | Page N | |
1000 * ^--------------------------------^----------|--------
1001 * | desired writeback range | see else |
1002 * ---------------------------------^------------------|
1004 if (page->index < end_index)
1005 end_offset = (xfs_off_t)(page->index + 1) << PAGE_CACHE_SHIFT;
1008 * Check whether the page to write out is beyond or straddles
1010 * -------------------------------------------------------
1011 * | file mapping | <EOF> |
1012 * -------------------------------------------------------
1013 * | Page ... | Page N-2 | Page N-1 | Page N | Beyond |
1014 * ^--------------------------------^-----------|---------
1016 * ---------------------------------^-----------|--------|
1018 unsigned offset_into_page = offset & (PAGE_CACHE_SIZE - 1);
1021 * Skip the page if it is fully outside i_size, e.g. due to a
1022 * truncate operation that is in progress. We must redirty the
1023 * page so that reclaim stops reclaiming it. Otherwise
1024 * xfs_vm_releasepage() is called on it and gets confused.
1026 * Note that the end_index is unsigned long, it would overflow
1027 * if the given offset is greater than 16TB on 32-bit system
1028 * and if we do check the page is fully outside i_size or not
1029 * via "if (page->index >= end_index + 1)" as "end_index + 1"
1030 * will be evaluated to 0. Hence this page will be redirtied
1031 * and be written out repeatedly which would result in an
1032 * infinite loop, the user program that perform this operation
1033 * will hang. Instead, we can verify this situation by checking
1034 * if the page to write is totally beyond the i_size or if it's
1035 * offset is just equal to the EOF.
1037 if (page->index > end_index ||
1038 (page->index == end_index && offset_into_page == 0))
1042 * The page straddles i_size. It must be zeroed out on each
1043 * and every writepage invocation because it may be mmapped.
1044 * "A file is mapped in multiples of the page size. For a file
1045 * that is not a multiple of the page size, the remaining
1046 * memory is zeroed when mapped, and writes to that region are
1047 * not written out to the file."
1049 zero_user_segment(page, offset_into_page, PAGE_CACHE_SIZE);
1051 /* Adjust the end_offset to the end of file */
1052 end_offset = offset;
1055 len = 1 << inode->i_blkbits;
1057 bh = head = page_buffers(page);
1058 offset = page_offset(page);
1059 type = XFS_IO_OVERWRITE;
1061 if (wbc->sync_mode == WB_SYNC_NONE)
1067 if (offset >= end_offset)
1069 if (!buffer_uptodate(bh))
1073 * set_page_dirty dirties all buffers in a page, independent
1074 * of their state. The dirty state however is entirely
1075 * meaningless for holes (!mapped && uptodate), so skip
1076 * buffers covering holes here.
1078 if (!buffer_mapped(bh) && buffer_uptodate(bh)) {
1083 if (buffer_unwritten(bh)) {
1084 if (type != XFS_IO_UNWRITTEN) {
1085 type = XFS_IO_UNWRITTEN;
1088 } else if (buffer_delay(bh)) {
1089 if (type != XFS_IO_DELALLOC) {
1090 type = XFS_IO_DELALLOC;
1093 } else if (buffer_uptodate(bh)) {
1094 if (type != XFS_IO_OVERWRITE) {
1095 type = XFS_IO_OVERWRITE;
1099 if (PageUptodate(page))
1100 ASSERT(buffer_mapped(bh));
1102 * This buffer is not uptodate and will not be
1103 * written to disk. Ensure that we will put any
1104 * subsequent writeable buffers into a new
1112 imap_valid = xfs_imap_valid(inode, &imap, offset);
1115 * If we didn't have a valid mapping then we need to
1116 * put the new mapping into a separate ioend structure.
1117 * This ensures non-contiguous extents always have
1118 * separate ioends, which is particularly important
1119 * for unwritten extent conversion at I/O completion
1123 err = xfs_map_blocks(inode, offset, &imap, type,
1127 imap_valid = xfs_imap_valid(inode, &imap, offset);
1131 if (type != XFS_IO_OVERWRITE)
1132 xfs_map_at_offset(inode, bh, &imap, offset);
1133 xfs_add_to_ioend(inode, bh, offset, type, &ioend,
1141 } while (offset += len, ((bh = bh->b_this_page) != head));
1143 if (uptodate && bh == head)
1144 SetPageUptodate(page);
1146 xfs_start_page_writeback(page, 1, count);
1148 /* if there is no IO to be submitted for this page, we are done */
1155 * Any errors from this point onwards need tobe reported through the IO
1156 * completion path as we have marked the initial page as under writeback
1160 xfs_off_t end_index;
1162 end_index = imap.br_startoff + imap.br_blockcount;
1165 end_index <<= inode->i_blkbits;
1168 end_index = (end_index - 1) >> PAGE_CACHE_SHIFT;
1170 /* check against file size */
1171 if (end_index > last_index)
1172 end_index = last_index;
1174 xfs_cluster_write(inode, page->index + 1, &imap, &ioend,
1180 * Reserve log space if we might write beyond the on-disk inode size.
1183 if (ioend->io_type != XFS_IO_UNWRITTEN && xfs_ioend_is_append(ioend))
1184 err = xfs_setfilesize_trans_alloc(ioend);
1186 xfs_submit_ioend(wbc, iohead, err);
1192 xfs_cancel_ioend(iohead);
1197 xfs_aops_discard_page(page);
1198 ClearPageUptodate(page);
1203 redirty_page_for_writepage(wbc, page);
1210 struct address_space *mapping,
1211 struct writeback_control *wbc)
1213 xfs_iflags_clear(XFS_I(mapping->host), XFS_ITRUNCATED);
1214 return generic_writepages(mapping, wbc);
1218 * Called to move a page into cleanable state - and from there
1219 * to be released. The page should already be clean. We always
1220 * have buffer heads in this call.
1222 * Returns 1 if the page is ok to release, 0 otherwise.
1229 int delalloc, unwritten;
1231 trace_xfs_releasepage(page->mapping->host, page, 0, 0);
1233 xfs_count_page_state(page, &delalloc, &unwritten);
1235 if (WARN_ON_ONCE(delalloc))
1237 if (WARN_ON_ONCE(unwritten))
1240 return try_to_free_buffers(page);
1245 struct inode *inode,
1247 struct buffer_head *bh_result,
1251 struct xfs_inode *ip = XFS_I(inode);
1252 struct xfs_mount *mp = ip->i_mount;
1253 xfs_fileoff_t offset_fsb, end_fsb;
1256 struct xfs_bmbt_irec imap;
1262 if (XFS_FORCED_SHUTDOWN(mp))
1265 offset = (xfs_off_t)iblock << inode->i_blkbits;
1266 ASSERT(bh_result->b_size >= (1 << inode->i_blkbits));
1267 size = bh_result->b_size;
1269 if (!create && direct && offset >= i_size_read(inode))
1273 * Direct I/O is usually done on preallocated files, so try getting
1274 * a block mapping without an exclusive lock first. For buffered
1275 * writes we already have the exclusive iolock anyway, so avoiding
1276 * a lock roundtrip here by taking the ilock exclusive from the
1277 * beginning is a useful micro optimization.
1279 if (create && !direct) {
1280 lockmode = XFS_ILOCK_EXCL;
1281 xfs_ilock(ip, lockmode);
1283 lockmode = xfs_ilock_data_map_shared(ip);
1286 ASSERT(offset <= mp->m_super->s_maxbytes);
1287 if (offset + size > mp->m_super->s_maxbytes)
1288 size = mp->m_super->s_maxbytes - offset;
1289 end_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)offset + size);
1290 offset_fsb = XFS_B_TO_FSBT(mp, offset);
1292 error = xfs_bmapi_read(ip, offset_fsb, end_fsb - offset_fsb,
1293 &imap, &nimaps, XFS_BMAPI_ENTIRE);
1299 (imap.br_startblock == HOLESTARTBLOCK ||
1300 imap.br_startblock == DELAYSTARTBLOCK))) {
1301 if (direct || xfs_get_extsz_hint(ip)) {
1303 * Drop the ilock in preparation for starting the block
1304 * allocation transaction. It will be retaken
1305 * exclusively inside xfs_iomap_write_direct for the
1306 * actual allocation.
1308 xfs_iunlock(ip, lockmode);
1309 error = xfs_iomap_write_direct(ip, offset, size,
1316 * Delalloc reservations do not require a transaction,
1317 * we can go on without dropping the lock here. If we
1318 * are allocating a new delalloc block, make sure that
1319 * we set the new flag so that we mark the buffer new so
1320 * that we know that it is newly allocated if the write
1323 if (nimaps && imap.br_startblock == HOLESTARTBLOCK)
1325 error = xfs_iomap_write_delay(ip, offset, size, &imap);
1329 xfs_iunlock(ip, lockmode);
1332 trace_xfs_get_blocks_alloc(ip, offset, size, 0, &imap);
1333 } else if (nimaps) {
1334 trace_xfs_get_blocks_found(ip, offset, size, 0, &imap);
1335 xfs_iunlock(ip, lockmode);
1337 trace_xfs_get_blocks_notfound(ip, offset, size);
1341 if (imap.br_startblock != HOLESTARTBLOCK &&
1342 imap.br_startblock != DELAYSTARTBLOCK) {
1344 * For unwritten extents do not report a disk address on
1345 * the read case (treat as if we're reading into a hole).
1347 if (create || !ISUNWRITTEN(&imap))
1348 xfs_map_buffer(inode, bh_result, &imap, offset);
1349 if (create && ISUNWRITTEN(&imap)) {
1351 bh_result->b_private = inode;
1352 set_buffer_defer_completion(bh_result);
1354 set_buffer_unwritten(bh_result);
1359 * If this is a realtime file, data may be on a different device.
1360 * to that pointed to from the buffer_head b_bdev currently.
1362 bh_result->b_bdev = xfs_find_bdev_for_inode(inode);
1365 * If we previously allocated a block out beyond eof and we are now
1366 * coming back to use it then we will need to flag it as new even if it
1367 * has a disk address.
1369 * With sub-block writes into unwritten extents we also need to mark
1370 * the buffer as new so that the unwritten parts of the buffer gets
1374 ((!buffer_mapped(bh_result) && !buffer_uptodate(bh_result)) ||
1375 (offset >= i_size_read(inode)) ||
1376 (new || ISUNWRITTEN(&imap))))
1377 set_buffer_new(bh_result);
1379 if (imap.br_startblock == DELAYSTARTBLOCK) {
1382 set_buffer_uptodate(bh_result);
1383 set_buffer_mapped(bh_result);
1384 set_buffer_delay(bh_result);
1389 * If this is O_DIRECT or the mpage code calling tell them how large
1390 * the mapping is, so that we can avoid repeated get_blocks calls.
1392 * If the mapping spans EOF, then we have to break the mapping up as the
1393 * mapping for blocks beyond EOF must be marked new so that sub block
1394 * regions can be correctly zeroed. We can't do this for mappings within
1395 * EOF unless the mapping was just allocated or is unwritten, otherwise
1396 * the callers would overwrite existing data with zeros. Hence we have
1397 * to split the mapping into a range up to and including EOF, and a
1398 * second mapping for beyond EOF.
1400 if (direct || size > (1 << inode->i_blkbits)) {
1401 xfs_off_t mapping_size;
1403 mapping_size = imap.br_startoff + imap.br_blockcount - iblock;
1404 mapping_size <<= inode->i_blkbits;
1406 ASSERT(mapping_size > 0);
1407 if (mapping_size > size)
1408 mapping_size = size;
1409 if (offset < i_size_read(inode) &&
1410 offset + mapping_size >= i_size_read(inode)) {
1411 /* limit mapping to block that spans EOF */
1412 mapping_size = roundup_64(i_size_read(inode) - offset,
1413 1 << inode->i_blkbits);
1415 if (mapping_size > LONG_MAX)
1416 mapping_size = LONG_MAX;
1418 bh_result->b_size = mapping_size;
1424 xfs_iunlock(ip, lockmode);
1430 struct inode *inode,
1432 struct buffer_head *bh_result,
1435 return __xfs_get_blocks(inode, iblock, bh_result, create, 0);
1439 xfs_get_blocks_direct(
1440 struct inode *inode,
1442 struct buffer_head *bh_result,
1445 return __xfs_get_blocks(inode, iblock, bh_result, create, 1);
1449 * Complete a direct I/O write request.
1451 * If the private argument is non-NULL __xfs_get_blocks signals us that we
1452 * need to issue a transaction to convert the range from unwritten to written
1453 * extents. In case this is regular synchronous I/O we just call xfs_end_io
1454 * to do this and we are done. But in case this was a successful AIO
1455 * request this handler is called from interrupt context, from which we
1456 * can't start transactions. In that case offload the I/O completion to
1457 * the workqueues we also use for buffered I/O completion.
1460 xfs_end_io_direct_write(
1466 struct xfs_ioend *ioend = iocb->private;
1469 * While the generic direct I/O code updates the inode size, it does
1470 * so only after the end_io handler is called, which means our
1471 * end_io handler thinks the on-disk size is outside the in-core
1472 * size. To prevent this just update it a little bit earlier here.
1474 if (offset + size > i_size_read(ioend->io_inode))
1475 i_size_write(ioend->io_inode, offset + size);
1478 * blockdev_direct_IO can return an error even after the I/O
1479 * completion handler was called. Thus we need to protect
1480 * against double-freeing.
1482 iocb->private = NULL;
1484 ioend->io_offset = offset;
1485 ioend->io_size = size;
1486 if (private && size > 0)
1487 ioend->io_type = XFS_IO_UNWRITTEN;
1489 xfs_finish_ioend_sync(ioend);
1496 struct iov_iter *iter,
1499 struct inode *inode = iocb->ki_filp->f_mapping->host;
1500 struct block_device *bdev = xfs_find_bdev_for_inode(inode);
1501 struct xfs_ioend *ioend = NULL;
1505 size_t size = iov_iter_count(iter);
1508 * We cannot preallocate a size update transaction here as we
1509 * don't know whether allocation is necessary or not. Hence we
1510 * can only tell IO completion that one is necessary if we are
1511 * not doing unwritten extent conversion.
1513 iocb->private = ioend = xfs_alloc_ioend(inode, XFS_IO_DIRECT);
1514 if (offset + size > XFS_I(inode)->i_d.di_size)
1515 ioend->io_isdirect = 1;
1517 ret = __blockdev_direct_IO(rw, iocb, inode, bdev, iter,
1518 offset, xfs_get_blocks_direct,
1519 xfs_end_io_direct_write, NULL,
1521 if (ret != -EIOCBQUEUED && iocb->private)
1522 goto out_destroy_ioend;
1524 ret = __blockdev_direct_IO(rw, iocb, inode, bdev, iter,
1525 offset, xfs_get_blocks_direct,
1532 xfs_destroy_ioend(ioend);
1537 * Punch out the delalloc blocks we have already allocated.
1539 * Don't bother with xfs_setattr given that nothing can have made it to disk yet
1540 * as the page is still locked at this point.
1543 xfs_vm_kill_delalloc_range(
1544 struct inode *inode,
1548 struct xfs_inode *ip = XFS_I(inode);
1549 xfs_fileoff_t start_fsb;
1550 xfs_fileoff_t end_fsb;
1553 start_fsb = XFS_B_TO_FSB(ip->i_mount, start);
1554 end_fsb = XFS_B_TO_FSB(ip->i_mount, end);
1555 if (end_fsb <= start_fsb)
1558 xfs_ilock(ip, XFS_ILOCK_EXCL);
1559 error = xfs_bmap_punch_delalloc_range(ip, start_fsb,
1560 end_fsb - start_fsb);
1562 /* something screwed, just bail */
1563 if (!XFS_FORCED_SHUTDOWN(ip->i_mount)) {
1564 xfs_alert(ip->i_mount,
1565 "xfs_vm_write_failed: unable to clean up ino %lld",
1569 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1573 xfs_vm_write_failed(
1574 struct inode *inode,
1579 loff_t block_offset;
1582 loff_t from = pos & (PAGE_CACHE_SIZE - 1);
1583 loff_t to = from + len;
1584 struct buffer_head *bh, *head;
1587 * The request pos offset might be 32 or 64 bit, this is all fine
1588 * on 64-bit platform. However, for 64-bit pos request on 32-bit
1589 * platform, the high 32-bit will be masked off if we evaluate the
1590 * block_offset via (pos & PAGE_MASK) because the PAGE_MASK is
1591 * 0xfffff000 as an unsigned long, hence the result is incorrect
1592 * which could cause the following ASSERT failed in most cases.
1593 * In order to avoid this, we can evaluate the block_offset of the
1594 * start of the page by using shifts rather than masks the mismatch
1597 block_offset = (pos >> PAGE_CACHE_SHIFT) << PAGE_CACHE_SHIFT;
1599 ASSERT(block_offset + from == pos);
1601 head = page_buffers(page);
1603 for (bh = head; bh != head || !block_start;
1604 bh = bh->b_this_page, block_start = block_end,
1605 block_offset += bh->b_size) {
1606 block_end = block_start + bh->b_size;
1608 /* skip buffers before the write */
1609 if (block_end <= from)
1612 /* if the buffer is after the write, we're done */
1613 if (block_start >= to)
1616 if (!buffer_delay(bh))
1619 if (!buffer_new(bh) && block_offset < i_size_read(inode))
1622 xfs_vm_kill_delalloc_range(inode, block_offset,
1623 block_offset + bh->b_size);
1626 * This buffer does not contain data anymore. make sure anyone
1627 * who finds it knows that for certain.
1629 clear_buffer_delay(bh);
1630 clear_buffer_uptodate(bh);
1631 clear_buffer_mapped(bh);
1632 clear_buffer_new(bh);
1633 clear_buffer_dirty(bh);
1639 * This used to call block_write_begin(), but it unlocks and releases the page
1640 * on error, and we need that page to be able to punch stale delalloc blocks out
1641 * on failure. hence we copy-n-waste it here and call xfs_vm_write_failed() at
1642 * the appropriate point.
1647 struct address_space *mapping,
1651 struct page **pagep,
1654 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1658 ASSERT(len <= PAGE_CACHE_SIZE);
1660 page = grab_cache_page_write_begin(mapping, index, flags);
1664 status = __block_write_begin(page, pos, len, xfs_get_blocks);
1665 if (unlikely(status)) {
1666 struct inode *inode = mapping->host;
1667 size_t isize = i_size_read(inode);
1669 xfs_vm_write_failed(inode, page, pos, len);
1673 * If the write is beyond EOF, we only want to kill blocks
1674 * allocated in this write, not blocks that were previously
1675 * written successfully.
1677 if (pos + len > isize) {
1678 ssize_t start = max_t(ssize_t, pos, isize);
1680 truncate_pagecache_range(inode, start, pos + len);
1683 page_cache_release(page);
1692 * On failure, we only need to kill delalloc blocks beyond EOF in the range of
1693 * this specific write because they will never be written. Previous writes
1694 * beyond EOF where block allocation succeeded do not need to be trashed, so
1695 * only new blocks from this write should be trashed. For blocks within
1696 * EOF, generic_write_end() zeros them so they are safe to leave alone and be
1697 * written with all the other valid data.
1702 struct address_space *mapping,
1711 ASSERT(len <= PAGE_CACHE_SIZE);
1713 ret = generic_write_end(file, mapping, pos, len, copied, page, fsdata);
1714 if (unlikely(ret < len)) {
1715 struct inode *inode = mapping->host;
1716 size_t isize = i_size_read(inode);
1717 loff_t to = pos + len;
1720 /* only kill blocks in this write beyond EOF */
1723 xfs_vm_kill_delalloc_range(inode, isize, to);
1724 truncate_pagecache_range(inode, isize, to);
1732 struct address_space *mapping,
1735 struct inode *inode = (struct inode *)mapping->host;
1736 struct xfs_inode *ip = XFS_I(inode);
1738 trace_xfs_vm_bmap(XFS_I(inode));
1739 xfs_ilock(ip, XFS_IOLOCK_SHARED);
1740 filemap_write_and_wait(mapping);
1741 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
1742 return generic_block_bmap(mapping, block, xfs_get_blocks);
1747 struct file *unused,
1750 return mpage_readpage(page, xfs_get_blocks);
1755 struct file *unused,
1756 struct address_space *mapping,
1757 struct list_head *pages,
1760 return mpage_readpages(mapping, pages, nr_pages, xfs_get_blocks);
1763 const struct address_space_operations xfs_address_space_operations = {
1764 .readpage = xfs_vm_readpage,
1765 .readpages = xfs_vm_readpages,
1766 .writepage = xfs_vm_writepage,
1767 .writepages = xfs_vm_writepages,
1768 .releasepage = xfs_vm_releasepage,
1769 .invalidatepage = xfs_vm_invalidatepage,
1770 .write_begin = xfs_vm_write_begin,
1771 .write_end = xfs_vm_write_end,
1772 .bmap = xfs_vm_bmap,
1773 .direct_IO = xfs_vm_direct_IO,
1774 .migratepage = buffer_migrate_page,
1775 .is_partially_uptodate = block_is_partially_uptodate,
1776 .error_remove_page = generic_error_remove_page,