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
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_alloc.h"
30 #include "xfs_error.h"
32 #include "xfs_iomap.h"
33 #include "xfs_vnodeops.h"
34 #include "xfs_trace.h"
36 #include <linux/gfp.h>
37 #include <linux/mpage.h>
38 #include <linux/pagevec.h>
39 #include <linux/writeback.h>
42 * Types of I/O for bmap clustering and I/O completion tracking.
45 IO_READ, /* mapping for a read */
46 IO_DELAY, /* mapping covers delalloc region */
47 IO_UNWRITTEN, /* mapping covers allocated but uninitialized data */
48 IO_NEW /* just allocated */
52 * Prime number of hash buckets since address is used as the key.
55 #define to_ioend_wq(v) (&xfs_ioend_wq[((unsigned long)v) % NVSYNC])
56 static wait_queue_head_t xfs_ioend_wq[NVSYNC];
63 for (i = 0; i < NVSYNC; i++)
64 init_waitqueue_head(&xfs_ioend_wq[i]);
71 wait_queue_head_t *wq = to_ioend_wq(ip);
73 wait_event(*wq, (atomic_read(&ip->i_iocount) == 0));
80 if (atomic_dec_and_test(&ip->i_iocount))
81 wake_up(to_ioend_wq(ip));
90 struct buffer_head *bh, *head;
92 *delalloc = *unwritten = 0;
94 bh = head = page_buffers(page);
96 if (buffer_unwritten(bh))
98 else if (buffer_delay(bh))
100 } while ((bh = bh->b_this_page) != head);
103 STATIC struct block_device *
104 xfs_find_bdev_for_inode(
107 struct xfs_inode *ip = XFS_I(inode);
108 struct xfs_mount *mp = ip->i_mount;
110 if (XFS_IS_REALTIME_INODE(ip))
111 return mp->m_rtdev_targp->bt_bdev;
113 return mp->m_ddev_targp->bt_bdev;
117 * We're now finished for good with this ioend structure.
118 * Update the page state via the associated buffer_heads,
119 * release holds on the inode and bio, and finally free
120 * up memory. Do not use the ioend after this.
126 struct buffer_head *bh, *next;
127 struct xfs_inode *ip = XFS_I(ioend->io_inode);
129 for (bh = ioend->io_buffer_head; bh; bh = next) {
130 next = bh->b_private;
131 bh->b_end_io(bh, !ioend->io_error);
135 * Volume managers supporting multiple paths can send back ENODEV
136 * when the final path disappears. In this case continuing to fill
137 * the page cache with dirty data which cannot be written out is
138 * evil, so prevent that.
140 if (unlikely(ioend->io_error == -ENODEV)) {
141 xfs_do_force_shutdown(ip->i_mount, SHUTDOWN_DEVICE_REQ,
146 mempool_free(ioend, xfs_ioend_pool);
150 * If the end of the current ioend is beyond the current EOF,
151 * return the new EOF value, otherwise zero.
157 xfs_inode_t *ip = XFS_I(ioend->io_inode);
161 bsize = ioend->io_offset + ioend->io_size;
162 isize = MAX(ip->i_size, ip->i_new_size);
163 isize = MIN(isize, bsize);
164 return isize > ip->i_d.di_size ? isize : 0;
168 * Update on-disk file size now that data has been written to disk. The
169 * current in-memory file size is i_size. If a write is beyond eof i_new_size
170 * will be the intended file size until i_size is updated. If this write does
171 * not extend all the way to the valid file size then restrict this update to
172 * the end of the write.
174 * This function does not block as blocking on the inode lock in IO completion
175 * can lead to IO completion order dependency deadlocks.. If it can't get the
176 * inode ilock it will return EAGAIN. Callers must handle this.
182 xfs_inode_t *ip = XFS_I(ioend->io_inode);
185 ASSERT((ip->i_d.di_mode & S_IFMT) == S_IFREG);
186 ASSERT(ioend->io_type != IO_READ);
188 if (unlikely(ioend->io_error))
191 if (!xfs_ilock_nowait(ip, XFS_ILOCK_EXCL))
194 isize = xfs_ioend_new_eof(ioend);
196 ip->i_d.di_size = isize;
197 xfs_mark_inode_dirty(ip);
200 xfs_iunlock(ip, XFS_ILOCK_EXCL);
205 * Schedule IO completion handling on the final put of an ioend.
209 struct xfs_ioend *ioend)
211 if (atomic_dec_and_test(&ioend->io_remaining)) {
212 if (ioend->io_type == IO_UNWRITTEN)
213 queue_work(xfsconvertd_workqueue, &ioend->io_work);
215 queue_work(xfsdatad_workqueue, &ioend->io_work);
220 * IO write completion.
224 struct work_struct *work)
226 xfs_ioend_t *ioend = container_of(work, xfs_ioend_t, io_work);
227 struct xfs_inode *ip = XFS_I(ioend->io_inode);
231 * For unwritten extents we need to issue transactions to convert a
232 * range to normal written extens after the data I/O has finished.
234 if (ioend->io_type == IO_UNWRITTEN &&
235 likely(!ioend->io_error && !XFS_FORCED_SHUTDOWN(ip->i_mount))) {
237 error = xfs_iomap_write_unwritten(ip, ioend->io_offset,
240 ioend->io_error = error;
244 * We might have to update the on-disk file size after extending
247 if (ioend->io_type != IO_READ) {
248 error = xfs_setfilesize(ioend);
249 ASSERT(!error || error == EAGAIN);
253 * If we didn't complete processing of the ioend, requeue it to the
254 * tail of the workqueue for another attempt later. Otherwise destroy
257 if (error == EAGAIN) {
258 atomic_inc(&ioend->io_remaining);
259 xfs_finish_ioend(ioend);
260 /* ensure we don't spin on blocked ioends */
264 aio_complete(ioend->io_iocb, ioend->io_result, 0);
265 xfs_destroy_ioend(ioend);
270 * Call IO completion handling in caller context on the final put of an ioend.
273 xfs_finish_ioend_sync(
274 struct xfs_ioend *ioend)
276 if (atomic_dec_and_test(&ioend->io_remaining))
277 xfs_end_io(&ioend->io_work);
281 * Allocate and initialise an IO completion structure.
282 * We need to track unwritten extent write completion here initially.
283 * We'll need to extend this for updating the ondisk inode size later
293 ioend = mempool_alloc(xfs_ioend_pool, GFP_NOFS);
296 * Set the count to 1 initially, which will prevent an I/O
297 * completion callback from happening before we have started
298 * all the I/O from calling the completion routine too early.
300 atomic_set(&ioend->io_remaining, 1);
302 ioend->io_list = NULL;
303 ioend->io_type = type;
304 ioend->io_inode = inode;
305 ioend->io_buffer_head = NULL;
306 ioend->io_buffer_tail = NULL;
307 atomic_inc(&XFS_I(ioend->io_inode)->i_iocount);
308 ioend->io_offset = 0;
310 ioend->io_iocb = NULL;
311 ioend->io_result = 0;
313 INIT_WORK(&ioend->io_work, xfs_end_io);
322 struct xfs_bmbt_irec *imap,
328 return -xfs_iomap(XFS_I(inode), offset, count, flags, imap, &nmaps, &new);
334 struct xfs_bmbt_irec *imap,
337 offset >>= inode->i_blkbits;
339 return offset >= imap->br_startoff &&
340 offset < imap->br_startoff + imap->br_blockcount;
344 * BIO completion handler for buffered IO.
351 xfs_ioend_t *ioend = bio->bi_private;
353 ASSERT(atomic_read(&bio->bi_cnt) >= 1);
354 ioend->io_error = test_bit(BIO_UPTODATE, &bio->bi_flags) ? 0 : error;
356 /* Toss bio and pass work off to an xfsdatad thread */
357 bio->bi_private = NULL;
358 bio->bi_end_io = NULL;
361 xfs_finish_ioend(ioend);
365 xfs_submit_ioend_bio(
366 struct writeback_control *wbc,
370 atomic_inc(&ioend->io_remaining);
371 bio->bi_private = ioend;
372 bio->bi_end_io = xfs_end_bio;
375 * If the I/O is beyond EOF we mark the inode dirty immediately
376 * but don't update the inode size until I/O completion.
378 if (xfs_ioend_new_eof(ioend))
379 xfs_mark_inode_dirty(XFS_I(ioend->io_inode));
381 submit_bio(wbc->sync_mode == WB_SYNC_ALL ?
382 WRITE_SYNC_PLUG : WRITE, bio);
383 ASSERT(!bio_flagged(bio, BIO_EOPNOTSUPP));
389 struct buffer_head *bh)
392 int nvecs = bio_get_nr_vecs(bh->b_bdev);
395 bio = bio_alloc(GFP_NOIO, nvecs);
399 ASSERT(bio->bi_private == NULL);
400 bio->bi_sector = bh->b_blocknr * (bh->b_size >> 9);
401 bio->bi_bdev = bh->b_bdev;
407 xfs_start_buffer_writeback(
408 struct buffer_head *bh)
410 ASSERT(buffer_mapped(bh));
411 ASSERT(buffer_locked(bh));
412 ASSERT(!buffer_delay(bh));
413 ASSERT(!buffer_unwritten(bh));
415 mark_buffer_async_write(bh);
416 set_buffer_uptodate(bh);
417 clear_buffer_dirty(bh);
421 xfs_start_page_writeback(
426 ASSERT(PageLocked(page));
427 ASSERT(!PageWriteback(page));
429 clear_page_dirty_for_io(page);
430 set_page_writeback(page);
432 /* If no buffers on the page are to be written, finish it here */
434 end_page_writeback(page);
437 static inline int bio_add_buffer(struct bio *bio, struct buffer_head *bh)
439 return bio_add_page(bio, bh->b_page, bh->b_size, bh_offset(bh));
443 * Submit all of the bios for all of the ioends we have saved up, covering the
444 * initial writepage page and also any probed pages.
446 * Because we may have multiple ioends spanning a page, we need to start
447 * writeback on all the buffers before we submit them for I/O. If we mark the
448 * buffers as we got, then we can end up with a page that only has buffers
449 * marked async write and I/O complete on can occur before we mark the other
450 * buffers async write.
452 * The end result of this is that we trip a bug in end_page_writeback() because
453 * we call it twice for the one page as the code in end_buffer_async_write()
454 * assumes that all buffers on the page are started at the same time.
456 * The fix is two passes across the ioend list - one to start writeback on the
457 * buffer_heads, and then submit them for I/O on the second pass.
461 struct writeback_control *wbc,
464 xfs_ioend_t *head = ioend;
466 struct buffer_head *bh;
468 sector_t lastblock = 0;
470 /* Pass 1 - start writeback */
472 next = ioend->io_list;
473 for (bh = ioend->io_buffer_head; bh; bh = bh->b_private) {
474 xfs_start_buffer_writeback(bh);
476 } while ((ioend = next) != NULL);
478 /* Pass 2 - submit I/O */
481 next = ioend->io_list;
484 for (bh = ioend->io_buffer_head; bh; bh = bh->b_private) {
488 bio = xfs_alloc_ioend_bio(bh);
489 } else if (bh->b_blocknr != lastblock + 1) {
490 xfs_submit_ioend_bio(wbc, ioend, bio);
494 if (bio_add_buffer(bio, bh) != bh->b_size) {
495 xfs_submit_ioend_bio(wbc, ioend, bio);
499 lastblock = bh->b_blocknr;
502 xfs_submit_ioend_bio(wbc, ioend, bio);
503 xfs_finish_ioend(ioend);
504 } while ((ioend = next) != NULL);
508 * Cancel submission of all buffer_heads so far in this endio.
509 * Toss the endio too. Only ever called for the initial page
510 * in a writepage request, so only ever one page.
517 struct buffer_head *bh, *next_bh;
520 next = ioend->io_list;
521 bh = ioend->io_buffer_head;
523 next_bh = bh->b_private;
524 clear_buffer_async_write(bh);
526 } while ((bh = next_bh) != NULL);
528 xfs_ioend_wake(XFS_I(ioend->io_inode));
529 mempool_free(ioend, xfs_ioend_pool);
530 } while ((ioend = next) != NULL);
534 * Test to see if we've been building up a completion structure for
535 * earlier buffers -- if so, we try to append to this ioend if we
536 * can, otherwise we finish off any current ioend and start another.
537 * Return true if we've finished the given ioend.
542 struct buffer_head *bh,
545 xfs_ioend_t **result,
548 xfs_ioend_t *ioend = *result;
550 if (!ioend || need_ioend || type != ioend->io_type) {
551 xfs_ioend_t *previous = *result;
553 ioend = xfs_alloc_ioend(inode, type);
554 ioend->io_offset = offset;
555 ioend->io_buffer_head = bh;
556 ioend->io_buffer_tail = bh;
558 previous->io_list = ioend;
561 ioend->io_buffer_tail->b_private = bh;
562 ioend->io_buffer_tail = bh;
565 bh->b_private = NULL;
566 ioend->io_size += bh->b_size;
572 struct buffer_head *bh,
573 struct xfs_bmbt_irec *imap,
577 struct xfs_mount *m = XFS_I(inode)->i_mount;
578 xfs_off_t iomap_offset = XFS_FSB_TO_B(m, imap->br_startoff);
579 xfs_daddr_t iomap_bn = xfs_fsb_to_db(XFS_I(inode), imap->br_startblock);
581 ASSERT(imap->br_startblock != HOLESTARTBLOCK);
582 ASSERT(imap->br_startblock != DELAYSTARTBLOCK);
584 bn = (iomap_bn >> (inode->i_blkbits - BBSHIFT)) +
585 ((offset - iomap_offset) >> inode->i_blkbits);
587 ASSERT(bn || XFS_IS_REALTIME_INODE(XFS_I(inode)));
590 set_buffer_mapped(bh);
596 struct buffer_head *bh,
597 struct xfs_bmbt_irec *imap,
600 ASSERT(imap->br_startblock != HOLESTARTBLOCK);
601 ASSERT(imap->br_startblock != DELAYSTARTBLOCK);
604 xfs_map_buffer(inode, bh, imap, offset);
605 bh->b_bdev = xfs_find_bdev_for_inode(inode);
606 set_buffer_mapped(bh);
607 clear_buffer_delay(bh);
608 clear_buffer_unwritten(bh);
612 * Look for a page at index that is suitable for clustering.
617 unsigned int pg_offset)
619 struct buffer_head *bh, *head;
622 if (PageWriteback(page))
624 if (!PageDirty(page))
628 if (!page_has_buffers(page))
631 bh = head = page_buffers(page);
633 if (!buffer_uptodate(bh))
635 if (!buffer_mapped(bh))
638 if (ret >= pg_offset)
640 } while ((bh = bh->b_this_page) != head);
648 struct page *startpage,
649 struct buffer_head *bh,
650 struct buffer_head *head)
653 pgoff_t tindex, tlast, tloff;
657 /* First sum forwards in this page */
659 if (!buffer_uptodate(bh) || !buffer_mapped(bh))
662 } while ((bh = bh->b_this_page) != head);
664 /* if we reached the end of the page, sum forwards in following pages */
665 tlast = i_size_read(inode) >> PAGE_CACHE_SHIFT;
666 tindex = startpage->index + 1;
668 /* Prune this back to avoid pathological behavior */
669 tloff = min(tlast, startpage->index + 64);
671 pagevec_init(&pvec, 0);
672 while (!done && tindex <= tloff) {
673 unsigned len = min_t(pgoff_t, PAGEVEC_SIZE, tlast - tindex + 1);
675 if (!pagevec_lookup(&pvec, inode->i_mapping, tindex, len))
678 for (i = 0; i < pagevec_count(&pvec); i++) {
679 struct page *page = pvec.pages[i];
680 size_t pg_offset, pg_len = 0;
682 if (tindex == tlast) {
684 i_size_read(inode) & (PAGE_CACHE_SIZE - 1);
690 pg_offset = PAGE_CACHE_SIZE;
692 if (page->index == tindex && trylock_page(page)) {
693 pg_len = xfs_probe_page(page, pg_offset);
706 pagevec_release(&pvec);
714 * Test if a given page is suitable for writing as part of an unwritten
715 * or delayed allocate extent.
722 if (PageWriteback(page))
725 if (page->mapping && page_has_buffers(page)) {
726 struct buffer_head *bh, *head;
729 bh = head = page_buffers(page);
731 if (buffer_unwritten(bh))
732 acceptable = (type == IO_UNWRITTEN);
733 else if (buffer_delay(bh))
734 acceptable = (type == IO_DELAY);
735 else if (buffer_dirty(bh) && buffer_mapped(bh))
736 acceptable = (type == IO_NEW);
739 } while ((bh = bh->b_this_page) != head);
749 * Allocate & map buffers for page given the extent map. Write it out.
750 * except for the original page of a writepage, this is called on
751 * delalloc/unwritten pages only, for the original page it is possible
752 * that the page has no mapping at all.
759 struct xfs_bmbt_irec *imap,
760 xfs_ioend_t **ioendp,
761 struct writeback_control *wbc,
764 struct buffer_head *bh, *head;
765 xfs_off_t end_offset;
766 unsigned long p_offset;
769 int count = 0, done = 0, uptodate = 1;
770 xfs_off_t offset = page_offset(page);
772 if (page->index != tindex)
774 if (!trylock_page(page))
776 if (PageWriteback(page))
777 goto fail_unlock_page;
778 if (page->mapping != inode->i_mapping)
779 goto fail_unlock_page;
780 if (!xfs_is_delayed_page(page, (*ioendp)->io_type))
781 goto fail_unlock_page;
784 * page_dirty is initially a count of buffers on the page before
785 * EOF and is decremented as we move each into a cleanable state.
789 * End offset is the highest offset that this page should represent.
790 * If we are on the last page, (end_offset & (PAGE_CACHE_SIZE - 1))
791 * will evaluate non-zero and be less than PAGE_CACHE_SIZE and
792 * hence give us the correct page_dirty count. On any other page,
793 * it will be zero and in that case we need page_dirty to be the
794 * count of buffers on the page.
796 end_offset = min_t(unsigned long long,
797 (xfs_off_t)(page->index + 1) << PAGE_CACHE_SHIFT,
800 len = 1 << inode->i_blkbits;
801 p_offset = min_t(unsigned long, end_offset & (PAGE_CACHE_SIZE - 1),
803 p_offset = p_offset ? roundup(p_offset, len) : PAGE_CACHE_SIZE;
804 page_dirty = p_offset / len;
806 bh = head = page_buffers(page);
808 if (offset >= end_offset)
810 if (!buffer_uptodate(bh))
812 if (!(PageUptodate(page) || buffer_uptodate(bh))) {
817 if (buffer_unwritten(bh) || buffer_delay(bh)) {
818 if (buffer_unwritten(bh))
823 if (!xfs_imap_valid(inode, imap, offset)) {
828 ASSERT(imap->br_startblock != HOLESTARTBLOCK);
829 ASSERT(imap->br_startblock != DELAYSTARTBLOCK);
831 xfs_map_at_offset(inode, bh, imap, offset);
832 xfs_add_to_ioend(inode, bh, offset, type,
839 if (buffer_mapped(bh) && all_bh) {
841 xfs_add_to_ioend(inode, bh, offset,
849 } while (offset += len, (bh = bh->b_this_page) != head);
851 if (uptodate && bh == head)
852 SetPageUptodate(page);
855 if (--wbc->nr_to_write <= 0 &&
856 wbc->sync_mode == WB_SYNC_NONE)
859 xfs_start_page_writeback(page, !page_dirty, count);
869 * Convert & write out a cluster of pages in the same extent as defined
870 * by mp and following the start page.
876 struct xfs_bmbt_irec *imap,
877 xfs_ioend_t **ioendp,
878 struct writeback_control *wbc,
885 pagevec_init(&pvec, 0);
886 while (!done && tindex <= tlast) {
887 unsigned len = min_t(pgoff_t, PAGEVEC_SIZE, tlast - tindex + 1);
889 if (!pagevec_lookup(&pvec, inode->i_mapping, tindex, len))
892 for (i = 0; i < pagevec_count(&pvec); i++) {
893 done = xfs_convert_page(inode, pvec.pages[i], tindex++,
894 imap, ioendp, wbc, all_bh);
899 pagevec_release(&pvec);
905 xfs_vm_invalidatepage(
907 unsigned long offset)
909 trace_xfs_invalidatepage(page->mapping->host, page, offset);
910 block_invalidatepage(page, offset);
914 * If the page has delalloc buffers on it, we need to punch them out before we
915 * invalidate the page. If we don't, we leave a stale delalloc mapping on the
916 * inode that can trip a BUG() in xfs_get_blocks() later on if a direct IO read
917 * is done on that same region - the delalloc extent is returned when none is
918 * supposed to be there.
920 * We prevent this by truncating away the delalloc regions on the page before
921 * invalidating it. Because they are delalloc, we can do this without needing a
922 * transaction. Indeed - if we get ENOSPC errors, we have to be able to do this
923 * truncation without a transaction as there is no space left for block
924 * reservation (typically why we see a ENOSPC in writeback).
926 * This is not a performance critical path, so for now just do the punching a
927 * buffer head at a time.
930 xfs_aops_discard_page(
933 struct inode *inode = page->mapping->host;
934 struct xfs_inode *ip = XFS_I(inode);
935 struct buffer_head *bh, *head;
936 loff_t offset = page_offset(page);
937 ssize_t len = 1 << inode->i_blkbits;
939 if (!xfs_is_delayed_page(page, IO_DELAY))
942 if (XFS_FORCED_SHUTDOWN(ip->i_mount))
945 xfs_fs_cmn_err(CE_ALERT, ip->i_mount,
946 "page discard on page %p, inode 0x%llx, offset %llu.",
947 page, ip->i_ino, offset);
949 xfs_ilock(ip, XFS_ILOCK_EXCL);
950 bh = head = page_buffers(page);
953 xfs_fileoff_t offset_fsb;
954 xfs_bmbt_irec_t imap;
957 xfs_fsblock_t firstblock;
958 xfs_bmap_free_t flist;
960 if (!buffer_delay(bh))
963 offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
966 * Map the range first and check that it is a delalloc extent
967 * before trying to unmap the range. Otherwise we will be
968 * trying to remove a real extent (which requires a
969 * transaction) or a hole, which is probably a bad idea...
971 error = xfs_bmapi(NULL, ip, offset_fsb, 1,
972 XFS_BMAPI_ENTIRE, NULL, 0, &imap,
976 /* something screwed, just bail */
977 if (!XFS_FORCED_SHUTDOWN(ip->i_mount)) {
978 xfs_fs_cmn_err(CE_ALERT, ip->i_mount,
979 "page discard failed delalloc mapping lookup.");
987 if (imap.br_startblock != DELAYSTARTBLOCK) {
988 /* been converted, ignore */
991 WARN_ON(imap.br_blockcount == 0);
994 * Note: while we initialise the firstblock/flist pair, they
995 * should never be used because blocks should never be
996 * allocated or freed for a delalloc extent and hence we need
997 * don't cancel or finish them after the xfs_bunmapi() call.
999 xfs_bmap_init(&flist, &firstblock);
1000 error = xfs_bunmapi(NULL, ip, offset_fsb, 1, 0, 1, &firstblock,
1003 ASSERT(!flist.xbf_count && !flist.xbf_first);
1005 /* something screwed, just bail */
1006 if (!XFS_FORCED_SHUTDOWN(ip->i_mount)) {
1007 xfs_fs_cmn_err(CE_ALERT, ip->i_mount,
1008 "page discard unable to remove delalloc mapping.");
1015 } while ((bh = bh->b_this_page) != head);
1017 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1019 xfs_vm_invalidatepage(page, 0);
1024 * Write out a dirty page.
1026 * For delalloc space on the page we need to allocate space and flush it.
1027 * For unwritten space on the page we need to start the conversion to
1028 * regular allocated space.
1029 * For any other dirty buffer heads on the page we should flush them.
1031 * If we detect that a transaction would be required to flush the page, we
1032 * have to check the process flags first, if we are already in a transaction
1033 * or disk I/O during allocations is off, we need to fail the writepage and
1039 struct writeback_control *wbc)
1041 struct inode *inode = page->mapping->host;
1042 int delalloc, unwritten;
1043 struct buffer_head *bh, *head;
1044 struct xfs_bmbt_irec imap;
1045 xfs_ioend_t *ioend = NULL, *iohead = NULL;
1048 __uint64_t end_offset;
1049 pgoff_t end_index, last_index;
1051 int flags, err, imap_valid = 0, uptodate = 1;
1055 trace_xfs_writepage(inode, page, 0);
1057 ASSERT(page_has_buffers(page));
1060 * Refuse to write the page out if we are called from reclaim context.
1062 * This avoids stack overflows when called from deeply used stacks in
1063 * random callers for direct reclaim or memcg reclaim. We explicitly
1064 * allow reclaim from kswapd as the stack usage there is relatively low.
1066 * This should really be done by the core VM, but until that happens
1067 * filesystems like XFS, btrfs and ext4 have to take care of this
1070 if ((current->flags & (PF_MEMALLOC|PF_KSWAPD)) == PF_MEMALLOC)
1074 * We need a transaction if there are delalloc or unwritten buffers
1077 * If we need a transaction and the process flags say we are already
1078 * in a transaction, or no IO is allowed then mark the page dirty
1079 * again and leave the page as is.
1081 xfs_count_page_state(page, &delalloc, &unwritten);
1082 if ((current->flags & PF_FSTRANS) && (delalloc || unwritten))
1085 /* Is this page beyond the end of the file? */
1086 offset = i_size_read(inode);
1087 end_index = offset >> PAGE_CACHE_SHIFT;
1088 last_index = (offset - 1) >> PAGE_CACHE_SHIFT;
1089 if (page->index >= end_index) {
1090 if ((page->index >= end_index + 1) ||
1091 !(i_size_read(inode) & (PAGE_CACHE_SIZE - 1))) {
1097 end_offset = min_t(unsigned long long,
1098 (xfs_off_t)(page->index + 1) << PAGE_CACHE_SHIFT,
1100 len = 1 << inode->i_blkbits;
1102 bh = head = page_buffers(page);
1103 offset = page_offset(page);
1108 if (offset >= end_offset)
1110 if (!buffer_uptodate(bh))
1114 * A hole may still be marked uptodate because discard_buffer
1115 * leaves the flag set.
1117 if (!buffer_mapped(bh) && buffer_uptodate(bh)) {
1118 ASSERT(!buffer_dirty(bh));
1124 imap_valid = xfs_imap_valid(inode, &imap, offset);
1126 if (buffer_unwritten(bh) || buffer_delay(bh)) {
1130 * Make sure we don't use a read-only iomap
1132 if (flags == BMAPI_READ)
1135 if (buffer_unwritten(bh)) {
1136 type = IO_UNWRITTEN;
1137 flags = BMAPI_WRITE | BMAPI_IGNSTATE;
1138 } else if (buffer_delay(bh)) {
1140 flags = BMAPI_ALLOCATE;
1142 if (wbc->sync_mode == WB_SYNC_NONE &&
1144 flags |= BMAPI_TRYLOCK;
1149 * If we didn't have a valid mapping then we
1150 * need to ensure that we put the new mapping
1151 * in a new ioend structure. This needs to be
1152 * done to ensure that the ioends correctly
1153 * reflect the block mappings at io completion
1154 * for unwritten extent conversion.
1157 err = xfs_map_blocks(inode, offset, len,
1161 imap_valid = xfs_imap_valid(inode, &imap,
1165 xfs_map_at_offset(inode, bh, &imap, offset);
1166 xfs_add_to_ioend(inode, bh, offset, type,
1170 } else if (buffer_uptodate(bh)) {
1172 * we got here because the buffer is already mapped.
1173 * That means it must already have extents allocated
1174 * underneath it. Map the extent by reading it.
1176 if (!imap_valid || flags != BMAPI_READ) {
1178 size = xfs_probe_cluster(inode, page, bh, head);
1179 err = xfs_map_blocks(inode, offset, size,
1183 imap_valid = xfs_imap_valid(inode, &imap,
1188 * We set the type to IO_NEW in case we are doing a
1189 * small write at EOF that is extending the file but
1190 * without needing an allocation. We need to update the
1191 * file size on I/O completion in this case so it is
1192 * the same case as having just allocated a new extent
1193 * that we are writing into for the first time.
1196 if (trylock_buffer(bh)) {
1199 xfs_add_to_ioend(inode, bh, offset, type,
1200 &ioend, !imap_valid);
1205 } else if (PageUptodate(page)) {
1206 ASSERT(buffer_mapped(bh));
1213 } while (offset += len, ((bh = bh->b_this_page) != head));
1215 if (uptodate && bh == head)
1216 SetPageUptodate(page);
1218 xfs_start_page_writeback(page, 1, count);
1220 if (ioend && imap_valid) {
1221 xfs_off_t end_index;
1223 end_index = imap.br_startoff + imap.br_blockcount;
1226 end_index <<= inode->i_blkbits;
1229 end_index = (end_index - 1) >> PAGE_CACHE_SHIFT;
1231 /* check against file size */
1232 if (end_index > last_index)
1233 end_index = last_index;
1235 xfs_cluster_write(inode, page->index + 1, &imap, &ioend,
1236 wbc, all_bh, end_index);
1240 xfs_submit_ioend(wbc, iohead);
1246 xfs_cancel_ioend(iohead);
1251 xfs_aops_discard_page(page);
1252 ClearPageUptodate(page);
1257 redirty_page_for_writepage(wbc, page);
1264 struct address_space *mapping,
1265 struct writeback_control *wbc)
1267 xfs_iflags_clear(XFS_I(mapping->host), XFS_ITRUNCATED);
1268 return generic_writepages(mapping, wbc);
1272 * Called to move a page into cleanable state - and from there
1273 * to be released. The page should already be clean. We always
1274 * have buffer heads in this call.
1276 * Returns 1 if the page is ok to release, 0 otherwise.
1283 int delalloc, unwritten;
1285 trace_xfs_releasepage(page->mapping->host, page, 0);
1287 xfs_count_page_state(page, &delalloc, &unwritten);
1289 if (WARN_ON(delalloc))
1291 if (WARN_ON(unwritten))
1294 return try_to_free_buffers(page);
1299 struct inode *inode,
1301 struct buffer_head *bh_result,
1305 int flags = create ? BMAPI_WRITE : BMAPI_READ;
1306 struct xfs_bmbt_irec imap;
1313 offset = (xfs_off_t)iblock << inode->i_blkbits;
1314 ASSERT(bh_result->b_size >= (1 << inode->i_blkbits));
1315 size = bh_result->b_size;
1317 if (!create && direct && offset >= i_size_read(inode))
1320 if (direct && create)
1321 flags |= BMAPI_DIRECT;
1323 error = xfs_iomap(XFS_I(inode), offset, size, flags, &imap, &nimap,
1330 if (imap.br_startblock != HOLESTARTBLOCK &&
1331 imap.br_startblock != DELAYSTARTBLOCK) {
1333 * For unwritten extents do not report a disk address on
1334 * the read case (treat as if we're reading into a hole).
1336 if (create || !ISUNWRITTEN(&imap))
1337 xfs_map_buffer(inode, bh_result, &imap, offset);
1338 if (create && ISUNWRITTEN(&imap)) {
1340 bh_result->b_private = inode;
1341 set_buffer_unwritten(bh_result);
1346 * If this is a realtime file, data may be on a different device.
1347 * to that pointed to from the buffer_head b_bdev currently.
1349 bh_result->b_bdev = xfs_find_bdev_for_inode(inode);
1352 * If we previously allocated a block out beyond eof and we are now
1353 * coming back to use it then we will need to flag it as new even if it
1354 * has a disk address.
1356 * With sub-block writes into unwritten extents we also need to mark
1357 * the buffer as new so that the unwritten parts of the buffer gets
1361 ((!buffer_mapped(bh_result) && !buffer_uptodate(bh_result)) ||
1362 (offset >= i_size_read(inode)) ||
1363 (new || ISUNWRITTEN(&imap))))
1364 set_buffer_new(bh_result);
1366 if (imap.br_startblock == DELAYSTARTBLOCK) {
1369 set_buffer_uptodate(bh_result);
1370 set_buffer_mapped(bh_result);
1371 set_buffer_delay(bh_result);
1376 * If this is O_DIRECT or the mpage code calling tell them how large
1377 * the mapping is, so that we can avoid repeated get_blocks calls.
1379 if (direct || size > (1 << inode->i_blkbits)) {
1380 xfs_off_t mapping_size;
1382 mapping_size = imap.br_startoff + imap.br_blockcount - iblock;
1383 mapping_size <<= inode->i_blkbits;
1385 ASSERT(mapping_size > 0);
1386 if (mapping_size > size)
1387 mapping_size = size;
1388 if (mapping_size > LONG_MAX)
1389 mapping_size = LONG_MAX;
1391 bh_result->b_size = mapping_size;
1399 struct inode *inode,
1401 struct buffer_head *bh_result,
1404 return __xfs_get_blocks(inode, iblock, bh_result, create, 0);
1408 xfs_get_blocks_direct(
1409 struct inode *inode,
1411 struct buffer_head *bh_result,
1414 return __xfs_get_blocks(inode, iblock, bh_result, create, 1);
1418 * Complete a direct I/O write request.
1420 * If the private argument is non-NULL __xfs_get_blocks signals us that we
1421 * need to issue a transaction to convert the range from unwritten to written
1422 * extents. In case this is regular synchronous I/O we just call xfs_end_io
1423 * to do this and we are done. But in case this was a successfull AIO
1424 * request this handler is called from interrupt context, from which we
1425 * can't start transactions. In that case offload the I/O completion to
1426 * the workqueues we also use for buffered I/O completion.
1429 xfs_end_io_direct_write(
1437 struct xfs_ioend *ioend = iocb->private;
1440 * blockdev_direct_IO can return an error even after the I/O
1441 * completion handler was called. Thus we need to protect
1442 * against double-freeing.
1444 iocb->private = NULL;
1446 ioend->io_offset = offset;
1447 ioend->io_size = size;
1448 if (private && size > 0)
1449 ioend->io_type = IO_UNWRITTEN;
1453 * If we are converting an unwritten extent we need to delay
1454 * the AIO completion until after the unwrittent extent
1455 * conversion has completed, otherwise do it ASAP.
1457 if (ioend->io_type == IO_UNWRITTEN) {
1458 ioend->io_iocb = iocb;
1459 ioend->io_result = ret;
1461 aio_complete(iocb, ret, 0);
1463 xfs_finish_ioend(ioend);
1465 xfs_finish_ioend_sync(ioend);
1473 const struct iovec *iov,
1475 unsigned long nr_segs)
1477 struct inode *inode = iocb->ki_filp->f_mapping->host;
1478 struct block_device *bdev = xfs_find_bdev_for_inode(inode);
1482 iocb->private = xfs_alloc_ioend(inode, IO_NEW);
1484 ret = __blockdev_direct_IO(rw, iocb, inode, bdev, iov,
1486 xfs_get_blocks_direct,
1487 xfs_end_io_direct_write, NULL, 0);
1488 if (ret != -EIOCBQUEUED && iocb->private)
1489 xfs_destroy_ioend(iocb->private);
1491 ret = __blockdev_direct_IO(rw, iocb, inode, bdev, iov,
1493 xfs_get_blocks_direct,
1501 xfs_vm_write_failed(
1502 struct address_space *mapping,
1505 struct inode *inode = mapping->host;
1507 if (to > inode->i_size) {
1509 .ia_valid = ATTR_SIZE | ATTR_FORCE,
1510 .ia_size = inode->i_size,
1512 xfs_setattr(XFS_I(inode), &ia, XFS_ATTR_NOLOCK);
1519 struct address_space *mapping,
1523 struct page **pagep,
1528 ret = block_write_begin(mapping, pos, len, flags | AOP_FLAG_NOFS,
1529 pagep, xfs_get_blocks);
1531 xfs_vm_write_failed(mapping, pos + len);
1538 struct address_space *mapping,
1547 ret = generic_write_end(file, mapping, pos, len, copied, page, fsdata);
1548 if (unlikely(ret < len))
1549 xfs_vm_write_failed(mapping, pos + len);
1555 struct address_space *mapping,
1558 struct inode *inode = (struct inode *)mapping->host;
1559 struct xfs_inode *ip = XFS_I(inode);
1561 trace_xfs_vm_bmap(XFS_I(inode));
1562 xfs_ilock(ip, XFS_IOLOCK_SHARED);
1563 xfs_flush_pages(ip, (xfs_off_t)0, -1, 0, FI_REMAPF);
1564 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
1565 return generic_block_bmap(mapping, block, xfs_get_blocks);
1570 struct file *unused,
1573 return mpage_readpage(page, xfs_get_blocks);
1578 struct file *unused,
1579 struct address_space *mapping,
1580 struct list_head *pages,
1583 return mpage_readpages(mapping, pages, nr_pages, xfs_get_blocks);
1586 const struct address_space_operations xfs_address_space_operations = {
1587 .readpage = xfs_vm_readpage,
1588 .readpages = xfs_vm_readpages,
1589 .writepage = xfs_vm_writepage,
1590 .writepages = xfs_vm_writepages,
1591 .sync_page = block_sync_page,
1592 .releasepage = xfs_vm_releasepage,
1593 .invalidatepage = xfs_vm_invalidatepage,
1594 .write_begin = xfs_vm_write_begin,
1595 .write_end = xfs_vm_write_end,
1596 .bmap = xfs_vm_bmap,
1597 .direct_IO = xfs_vm_direct_IO,
1598 .migratepage = buffer_migrate_page,
1599 .is_partially_uptodate = block_is_partially_uptodate,
1600 .error_remove_page = generic_error_remove_page,