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
25 #include "xfs_trans.h"
26 #include "xfs_dmapi.h"
27 #include "xfs_mount.h"
28 #include "xfs_bmap_btree.h"
29 #include "xfs_alloc_btree.h"
30 #include "xfs_ialloc_btree.h"
31 #include "xfs_dir2_sf.h"
32 #include "xfs_attr_sf.h"
33 #include "xfs_dinode.h"
34 #include "xfs_inode.h"
35 #include "xfs_alloc.h"
36 #include "xfs_btree.h"
37 #include "xfs_error.h"
39 #include "xfs_iomap.h"
40 #include "xfs_vnodeops.h"
41 #include "xfs_trace.h"
43 #include <linux/gfp.h>
44 #include <linux/mpage.h>
45 #include <linux/pagevec.h>
46 #include <linux/writeback.h>
49 * Types of I/O for bmap clustering and I/O completion tracking.
52 IO_READ, /* mapping for a read */
53 IO_DELAY, /* mapping covers delalloc region */
54 IO_UNWRITTEN, /* mapping covers allocated but uninitialized data */
55 IO_NEW /* just allocated */
59 * Prime number of hash buckets since address is used as the key.
62 #define to_ioend_wq(v) (&xfs_ioend_wq[((unsigned long)v) % NVSYNC])
63 static wait_queue_head_t xfs_ioend_wq[NVSYNC];
70 for (i = 0; i < NVSYNC; i++)
71 init_waitqueue_head(&xfs_ioend_wq[i]);
78 wait_queue_head_t *wq = to_ioend_wq(ip);
80 wait_event(*wq, (atomic_read(&ip->i_iocount) == 0));
87 if (atomic_dec_and_test(&ip->i_iocount))
88 wake_up(to_ioend_wq(ip));
98 struct buffer_head *bh, *head;
100 *delalloc = *unmapped = *unwritten = 0;
102 bh = head = page_buffers(page);
104 if (buffer_uptodate(bh) && !buffer_mapped(bh))
106 else if (buffer_unwritten(bh))
108 else if (buffer_delay(bh))
110 } while ((bh = bh->b_this_page) != head);
113 STATIC struct block_device *
114 xfs_find_bdev_for_inode(
117 struct xfs_inode *ip = XFS_I(inode);
118 struct xfs_mount *mp = ip->i_mount;
120 if (XFS_IS_REALTIME_INODE(ip))
121 return mp->m_rtdev_targp->bt_bdev;
123 return mp->m_ddev_targp->bt_bdev;
127 * We're now finished for good with this ioend structure.
128 * Update the page state via the associated buffer_heads,
129 * release holds on the inode and bio, and finally free
130 * up memory. Do not use the ioend after this.
136 struct buffer_head *bh, *next;
137 struct xfs_inode *ip = XFS_I(ioend->io_inode);
139 for (bh = ioend->io_buffer_head; bh; bh = next) {
140 next = bh->b_private;
141 bh->b_end_io(bh, !ioend->io_error);
145 * Volume managers supporting multiple paths can send back ENODEV
146 * when the final path disappears. In this case continuing to fill
147 * the page cache with dirty data which cannot be written out is
148 * evil, so prevent that.
150 if (unlikely(ioend->io_error == -ENODEV)) {
151 xfs_do_force_shutdown(ip->i_mount, SHUTDOWN_DEVICE_REQ,
156 mempool_free(ioend, xfs_ioend_pool);
160 * If the end of the current ioend is beyond the current EOF,
161 * return the new EOF value, otherwise zero.
167 xfs_inode_t *ip = XFS_I(ioend->io_inode);
171 bsize = ioend->io_offset + ioend->io_size;
172 isize = MAX(ip->i_size, ip->i_new_size);
173 isize = MIN(isize, bsize);
174 return isize > ip->i_d.di_size ? isize : 0;
178 * Update on-disk file size now that data has been written to disk. The
179 * current in-memory file size is i_size. If a write is beyond eof i_new_size
180 * will be the intended file size until i_size is updated. If this write does
181 * not extend all the way to the valid file size then restrict this update to
182 * the end of the write.
184 * This function does not block as blocking on the inode lock in IO completion
185 * can lead to IO completion order dependency deadlocks.. If it can't get the
186 * inode ilock it will return EAGAIN. Callers must handle this.
192 xfs_inode_t *ip = XFS_I(ioend->io_inode);
195 ASSERT((ip->i_d.di_mode & S_IFMT) == S_IFREG);
196 ASSERT(ioend->io_type != IO_READ);
198 if (unlikely(ioend->io_error))
201 if (!xfs_ilock_nowait(ip, XFS_ILOCK_EXCL))
204 isize = xfs_ioend_new_eof(ioend);
206 ip->i_d.di_size = isize;
207 xfs_mark_inode_dirty(ip);
210 xfs_iunlock(ip, XFS_ILOCK_EXCL);
215 * Schedule IO completion handling on a xfsdatad if this was
216 * the final hold on this ioend. If we are asked to wait,
217 * flush the workqueue.
224 if (atomic_dec_and_test(&ioend->io_remaining)) {
225 struct workqueue_struct *wq;
227 wq = (ioend->io_type == IO_UNWRITTEN) ?
228 xfsconvertd_workqueue : xfsdatad_workqueue;
229 queue_work(wq, &ioend->io_work);
236 * IO write completion.
240 struct work_struct *work)
242 xfs_ioend_t *ioend = container_of(work, xfs_ioend_t, io_work);
243 struct xfs_inode *ip = XFS_I(ioend->io_inode);
247 * For unwritten extents we need to issue transactions to convert a
248 * range to normal written extens after the data I/O has finished.
250 if (ioend->io_type == IO_UNWRITTEN &&
251 likely(!ioend->io_error && !XFS_FORCED_SHUTDOWN(ip->i_mount))) {
253 error = xfs_iomap_write_unwritten(ip, ioend->io_offset,
256 ioend->io_error = error;
260 * We might have to update the on-disk file size after extending
263 if (ioend->io_type != IO_READ) {
264 error = xfs_setfilesize(ioend);
265 ASSERT(!error || error == EAGAIN);
269 * If we didn't complete processing of the ioend, requeue it to the
270 * tail of the workqueue for another attempt later. Otherwise destroy
273 if (error == EAGAIN) {
274 atomic_inc(&ioend->io_remaining);
275 xfs_finish_ioend(ioend, 0);
276 /* ensure we don't spin on blocked ioends */
279 xfs_destroy_ioend(ioend);
283 * Allocate and initialise an IO completion structure.
284 * We need to track unwritten extent write completion here initially.
285 * We'll need to extend this for updating the ondisk inode size later
295 ioend = mempool_alloc(xfs_ioend_pool, GFP_NOFS);
298 * Set the count to 1 initially, which will prevent an I/O
299 * completion callback from happening before we have started
300 * all the I/O from calling the completion routine too early.
302 atomic_set(&ioend->io_remaining, 1);
304 ioend->io_list = NULL;
305 ioend->io_type = type;
306 ioend->io_inode = inode;
307 ioend->io_buffer_head = NULL;
308 ioend->io_buffer_tail = NULL;
309 atomic_inc(&XFS_I(ioend->io_inode)->i_iocount);
310 ioend->io_offset = 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, 0);
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, 0);
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,
622 if (PageWriteback(page))
625 if (page->mapping && PageDirty(page)) {
626 if (page_has_buffers(page)) {
627 struct buffer_head *bh, *head;
629 bh = head = page_buffers(page);
631 if (!buffer_uptodate(bh))
633 if (mapped != buffer_mapped(bh))
636 if (ret >= pg_offset)
638 } while ((bh = bh->b_this_page) != head);
640 ret = mapped ? 0 : PAGE_CACHE_SIZE;
649 struct page *startpage,
650 struct buffer_head *bh,
651 struct buffer_head *head,
655 pgoff_t tindex, tlast, tloff;
659 /* First sum forwards in this page */
661 if (!buffer_uptodate(bh) || (mapped != buffer_mapped(bh)))
664 } while ((bh = bh->b_this_page) != head);
666 /* if we reached the end of the page, sum forwards in following pages */
667 tlast = i_size_read(inode) >> PAGE_CACHE_SHIFT;
668 tindex = startpage->index + 1;
670 /* Prune this back to avoid pathological behavior */
671 tloff = min(tlast, startpage->index + 64);
673 pagevec_init(&pvec, 0);
674 while (!done && tindex <= tloff) {
675 unsigned len = min_t(pgoff_t, PAGEVEC_SIZE, tlast - tindex + 1);
677 if (!pagevec_lookup(&pvec, inode->i_mapping, tindex, len))
680 for (i = 0; i < pagevec_count(&pvec); i++) {
681 struct page *page = pvec.pages[i];
682 size_t pg_offset, pg_len = 0;
684 if (tindex == tlast) {
686 i_size_read(inode) & (PAGE_CACHE_SIZE - 1);
692 pg_offset = PAGE_CACHE_SIZE;
694 if (page->index == tindex && trylock_page(page)) {
695 pg_len = xfs_probe_page(page, pg_offset, mapped);
708 pagevec_release(&pvec);
716 * Test if a given page is suitable for writing as part of an unwritten
717 * or delayed allocate extent.
724 if (PageWriteback(page))
727 if (page->mapping && page_has_buffers(page)) {
728 struct buffer_head *bh, *head;
731 bh = head = page_buffers(page);
733 if (buffer_unwritten(bh))
734 acceptable = (type == IO_UNWRITTEN);
735 else if (buffer_delay(bh))
736 acceptable = (type == IO_DELAY);
737 else if (buffer_dirty(bh) && buffer_mapped(bh))
738 acceptable = (type == IO_NEW);
741 } while ((bh = bh->b_this_page) != head);
751 * Allocate & map buffers for page given the extent map. Write it out.
752 * except for the original page of a writepage, this is called on
753 * delalloc/unwritten pages only, for the original page it is possible
754 * that the page has no mapping at all.
761 struct xfs_bmbt_irec *imap,
762 xfs_ioend_t **ioendp,
763 struct writeback_control *wbc,
767 struct buffer_head *bh, *head;
768 xfs_off_t end_offset;
769 unsigned long p_offset;
772 int count = 0, done = 0, uptodate = 1;
773 xfs_off_t offset = page_offset(page);
775 if (page->index != tindex)
777 if (!trylock_page(page))
779 if (PageWriteback(page))
780 goto fail_unlock_page;
781 if (page->mapping != inode->i_mapping)
782 goto fail_unlock_page;
783 if (!xfs_is_delayed_page(page, (*ioendp)->io_type))
784 goto fail_unlock_page;
787 * page_dirty is initially a count of buffers on the page before
788 * EOF and is decremented as we move each into a cleanable state.
792 * End offset is the highest offset that this page should represent.
793 * If we are on the last page, (end_offset & (PAGE_CACHE_SIZE - 1))
794 * will evaluate non-zero and be less than PAGE_CACHE_SIZE and
795 * hence give us the correct page_dirty count. On any other page,
796 * it will be zero and in that case we need page_dirty to be the
797 * count of buffers on the page.
799 end_offset = min_t(unsigned long long,
800 (xfs_off_t)(page->index + 1) << PAGE_CACHE_SHIFT,
803 len = 1 << inode->i_blkbits;
804 p_offset = min_t(unsigned long, end_offset & (PAGE_CACHE_SIZE - 1),
806 p_offset = p_offset ? roundup(p_offset, len) : PAGE_CACHE_SIZE;
807 page_dirty = p_offset / len;
809 bh = head = page_buffers(page);
811 if (offset >= end_offset)
813 if (!buffer_uptodate(bh))
815 if (!(PageUptodate(page) || buffer_uptodate(bh))) {
820 if (buffer_unwritten(bh) || buffer_delay(bh)) {
821 if (buffer_unwritten(bh))
826 if (!xfs_imap_valid(inode, imap, offset)) {
831 ASSERT(imap->br_startblock != HOLESTARTBLOCK);
832 ASSERT(imap->br_startblock != DELAYSTARTBLOCK);
834 xfs_map_at_offset(inode, bh, imap, offset);
836 xfs_add_to_ioend(inode, bh, offset,
839 set_buffer_dirty(bh);
841 mark_buffer_dirty(bh);
847 if (buffer_mapped(bh) && all_bh && startio) {
849 xfs_add_to_ioend(inode, bh, offset,
857 } while (offset += len, (bh = bh->b_this_page) != head);
859 if (uptodate && bh == head)
860 SetPageUptodate(page);
865 if (wbc->nr_to_write <= 0)
868 xfs_start_page_writeback(page, !page_dirty, count);
879 * Convert & write out a cluster of pages in the same extent as defined
880 * by mp and following the start page.
886 struct xfs_bmbt_irec *imap,
887 xfs_ioend_t **ioendp,
888 struct writeback_control *wbc,
896 pagevec_init(&pvec, 0);
897 while (!done && tindex <= tlast) {
898 unsigned len = min_t(pgoff_t, PAGEVEC_SIZE, tlast - tindex + 1);
900 if (!pagevec_lookup(&pvec, inode->i_mapping, tindex, len))
903 for (i = 0; i < pagevec_count(&pvec); i++) {
904 done = xfs_convert_page(inode, pvec.pages[i], tindex++,
905 imap, ioendp, wbc, startio, all_bh);
910 pagevec_release(&pvec);
916 xfs_vm_invalidatepage(
918 unsigned long offset)
920 trace_xfs_invalidatepage(page->mapping->host, page, offset);
921 block_invalidatepage(page, offset);
925 * If the page has delalloc buffers on it, we need to punch them out before we
926 * invalidate the page. If we don't, we leave a stale delalloc mapping on the
927 * inode that can trip a BUG() in xfs_get_blocks() later on if a direct IO read
928 * is done on that same region - the delalloc extent is returned when none is
929 * supposed to be there.
931 * We prevent this by truncating away the delalloc regions on the page before
932 * invalidating it. Because they are delalloc, we can do this without needing a
933 * transaction. Indeed - if we get ENOSPC errors, we have to be able to do this
934 * truncation without a transaction as there is no space left for block
935 * reservation (typically why we see a ENOSPC in writeback).
937 * This is not a performance critical path, so for now just do the punching a
938 * buffer head at a time.
941 xfs_aops_discard_page(
944 struct inode *inode = page->mapping->host;
945 struct xfs_inode *ip = XFS_I(inode);
946 struct buffer_head *bh, *head;
947 loff_t offset = page_offset(page);
948 ssize_t len = 1 << inode->i_blkbits;
950 if (!xfs_is_delayed_page(page, IO_DELAY))
953 if (XFS_FORCED_SHUTDOWN(ip->i_mount))
956 xfs_fs_cmn_err(CE_ALERT, ip->i_mount,
957 "page discard on page %p, inode 0x%llx, offset %llu.",
958 page, ip->i_ino, offset);
960 xfs_ilock(ip, XFS_ILOCK_EXCL);
961 bh = head = page_buffers(page);
964 xfs_fileoff_t offset_fsb;
965 xfs_bmbt_irec_t imap;
968 xfs_fsblock_t firstblock;
969 xfs_bmap_free_t flist;
971 if (!buffer_delay(bh))
974 offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
977 * Map the range first and check that it is a delalloc extent
978 * before trying to unmap the range. Otherwise we will be
979 * trying to remove a real extent (which requires a
980 * transaction) or a hole, which is probably a bad idea...
982 error = xfs_bmapi(NULL, ip, offset_fsb, 1,
983 XFS_BMAPI_ENTIRE, NULL, 0, &imap,
984 &nimaps, NULL, NULL);
987 /* something screwed, just bail */
988 if (!XFS_FORCED_SHUTDOWN(ip->i_mount)) {
989 xfs_fs_cmn_err(CE_ALERT, ip->i_mount,
990 "page discard failed delalloc mapping lookup.");
998 if (imap.br_startblock != DELAYSTARTBLOCK) {
999 /* been converted, ignore */
1002 WARN_ON(imap.br_blockcount == 0);
1005 * Note: while we initialise the firstblock/flist pair, they
1006 * should never be used because blocks should never be
1007 * allocated or freed for a delalloc extent and hence we need
1008 * don't cancel or finish them after the xfs_bunmapi() call.
1010 xfs_bmap_init(&flist, &firstblock);
1011 error = xfs_bunmapi(NULL, ip, offset_fsb, 1, 0, 1, &firstblock,
1012 &flist, NULL, &done);
1014 ASSERT(!flist.xbf_count && !flist.xbf_first);
1016 /* something screwed, just bail */
1017 if (!XFS_FORCED_SHUTDOWN(ip->i_mount)) {
1018 xfs_fs_cmn_err(CE_ALERT, ip->i_mount,
1019 "page discard unable to remove delalloc mapping.");
1026 } while ((bh = bh->b_this_page) != head);
1028 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1030 xfs_vm_invalidatepage(page, 0);
1035 * Calling this without startio set means we are being asked to make a dirty
1036 * page ready for freeing it's buffers. When called with startio set then
1037 * we are coming from writepage.
1039 * When called with startio set it is important that we write the WHOLE
1041 * The bh->b_state's cannot know if any of the blocks or which block for
1042 * that matter are dirty due to mmap writes, and therefore bh uptodate is
1043 * only valid if the page itself isn't completely uptodate. Some layers
1044 * may clear the page dirty flag prior to calling write page, under the
1045 * assumption the entire page will be written out; by not writing out the
1046 * whole page the page can be reused before all valid dirty data is
1047 * written out. Note: in the case of a page that has been dirty'd by
1048 * mapwrite and but partially setup by block_prepare_write the
1049 * bh->b_states's will not agree and only ones setup by BPW/BCW will have
1050 * valid state, thus the whole page must be written out thing.
1054 xfs_page_state_convert(
1055 struct inode *inode,
1057 struct writeback_control *wbc,
1059 int unmapped) /* also implies page uptodate */
1061 struct buffer_head *bh, *head;
1062 struct xfs_bmbt_irec imap;
1063 xfs_ioend_t *ioend = NULL, *iohead = NULL;
1065 unsigned long p_offset = 0;
1067 __uint64_t end_offset;
1068 pgoff_t end_index, last_index;
1070 int flags, err, imap_valid = 0, uptodate = 1;
1071 int page_dirty, count = 0;
1073 int all_bh = unmapped;
1076 if (wbc->sync_mode == WB_SYNC_NONE && wbc->nonblocking)
1077 trylock |= BMAPI_TRYLOCK;
1080 /* Is this page beyond the end of the file? */
1081 offset = i_size_read(inode);
1082 end_index = offset >> PAGE_CACHE_SHIFT;
1083 last_index = (offset - 1) >> PAGE_CACHE_SHIFT;
1084 if (page->index >= end_index) {
1085 if ((page->index >= end_index + 1) ||
1086 !(i_size_read(inode) & (PAGE_CACHE_SIZE - 1))) {
1094 * page_dirty is initially a count of buffers on the page before
1095 * EOF and is decremented as we move each into a cleanable state.
1099 * End offset is the highest offset that this page should represent.
1100 * If we are on the last page, (end_offset & (PAGE_CACHE_SIZE - 1))
1101 * will evaluate non-zero and be less than PAGE_CACHE_SIZE and
1102 * hence give us the correct page_dirty count. On any other page,
1103 * it will be zero and in that case we need page_dirty to be the
1104 * count of buffers on the page.
1106 end_offset = min_t(unsigned long long,
1107 (xfs_off_t)(page->index + 1) << PAGE_CACHE_SHIFT, offset);
1108 len = 1 << inode->i_blkbits;
1109 p_offset = min_t(unsigned long, end_offset & (PAGE_CACHE_SIZE - 1),
1111 p_offset = p_offset ? roundup(p_offset, len) : PAGE_CACHE_SIZE;
1112 page_dirty = p_offset / len;
1114 bh = head = page_buffers(page);
1115 offset = page_offset(page);
1119 /* TODO: cleanup count and page_dirty */
1122 if (offset >= end_offset)
1124 if (!buffer_uptodate(bh))
1126 if (!(PageUptodate(page) || buffer_uptodate(bh)) && !startio) {
1128 * the iomap is actually still valid, but the ioend
1129 * isn't. shouldn't happen too often.
1136 imap_valid = xfs_imap_valid(inode, &imap, offset);
1139 * First case, map an unwritten extent and prepare for
1140 * extent state conversion transaction on completion.
1142 * Second case, allocate space for a delalloc buffer.
1143 * We can return EAGAIN here in the release page case.
1145 * Third case, an unmapped buffer was found, and we are
1146 * in a path where we need to write the whole page out.
1148 if (buffer_unwritten(bh) || buffer_delay(bh) ||
1149 ((buffer_uptodate(bh) || PageUptodate(page)) &&
1150 !buffer_mapped(bh) && (unmapped || startio))) {
1154 * Make sure we don't use a read-only iomap
1156 if (flags == BMAPI_READ)
1159 if (buffer_unwritten(bh)) {
1160 type = IO_UNWRITTEN;
1161 flags = BMAPI_WRITE | BMAPI_IGNSTATE;
1162 } else if (buffer_delay(bh)) {
1164 flags = BMAPI_ALLOCATE | trylock;
1167 flags = BMAPI_WRITE | BMAPI_MMAP;
1172 * if we didn't have a valid mapping then we
1173 * need to ensure that we put the new mapping
1174 * in a new ioend structure. This needs to be
1175 * done to ensure that the ioends correctly
1176 * reflect the block mappings at io completion
1177 * for unwritten extent conversion.
1180 if (type == IO_NEW) {
1181 size = xfs_probe_cluster(inode,
1187 err = xfs_map_blocks(inode, offset, size,
1191 imap_valid = xfs_imap_valid(inode, &imap,
1195 xfs_map_at_offset(inode, bh, &imap, offset);
1197 xfs_add_to_ioend(inode, bh, offset,
1201 set_buffer_dirty(bh);
1203 mark_buffer_dirty(bh);
1208 } else if (buffer_uptodate(bh) && startio) {
1210 * we got here because the buffer is already mapped.
1211 * That means it must already have extents allocated
1212 * underneath it. Map the extent by reading it.
1214 if (!imap_valid || flags != BMAPI_READ) {
1216 size = xfs_probe_cluster(inode, page, bh,
1218 err = xfs_map_blocks(inode, offset, size,
1222 imap_valid = xfs_imap_valid(inode, &imap,
1227 * We set the type to IO_NEW in case we are doing a
1228 * small write at EOF that is extending the file but
1229 * without needing an allocation. We need to update the
1230 * file size on I/O completion in this case so it is
1231 * the same case as having just allocated a new extent
1232 * that we are writing into for the first time.
1235 if (trylock_buffer(bh)) {
1236 ASSERT(buffer_mapped(bh));
1239 xfs_add_to_ioend(inode, bh, offset, type,
1240 &ioend, !imap_valid);
1246 } else if ((buffer_uptodate(bh) || PageUptodate(page)) &&
1247 (unmapped || startio)) {
1254 } while (offset += len, ((bh = bh->b_this_page) != head));
1256 if (uptodate && bh == head)
1257 SetPageUptodate(page);
1260 xfs_start_page_writeback(page, 1, count);
1262 if (ioend && imap_valid) {
1263 xfs_off_t end_index;
1265 end_index = imap.br_startoff + imap.br_blockcount;
1268 end_index <<= inode->i_blkbits;
1271 end_index = (end_index - 1) >> PAGE_CACHE_SHIFT;
1273 /* check against file size */
1274 if (end_index > last_index)
1275 end_index = last_index;
1277 xfs_cluster_write(inode, page->index + 1, &imap, &ioend,
1278 wbc, startio, all_bh, end_index);
1282 xfs_submit_ioend(wbc, iohead);
1288 xfs_cancel_ioend(iohead);
1291 * If it's delalloc and we have nowhere to put it,
1292 * throw it away, unless the lower layers told
1295 if (err != -EAGAIN) {
1297 xfs_aops_discard_page(page);
1298 ClearPageUptodate(page);
1304 * writepage: Called from one of two places:
1306 * 1. we are flushing a delalloc buffer head.
1308 * 2. we are writing out a dirty page. Typically the page dirty
1309 * state is cleared before we get here. In this case is it
1310 * conceivable we have no buffer heads.
1312 * For delalloc space on the page we need to allocate space and
1313 * flush it. For unmapped buffer heads on the page we should
1314 * allocate space if the page is uptodate. For any other dirty
1315 * buffer heads on the page we should flush them.
1317 * If we detect that a transaction would be required to flush
1318 * the page, we have to check the process flags first, if we
1319 * are already in a transaction or disk I/O during allocations
1320 * is off, we need to fail the writepage and redirty the page.
1326 struct writeback_control *wbc)
1330 int delalloc, unmapped, unwritten;
1331 struct inode *inode = page->mapping->host;
1333 trace_xfs_writepage(inode, page, 0);
1336 * Refuse to write the page out if we are called from reclaim context.
1338 * This is primarily to avoid stack overflows when called from deep
1339 * used stacks in random callers for direct reclaim, but disabling
1340 * reclaim for kswap is a nice side-effect as kswapd causes rather
1341 * suboptimal I/O patters, too.
1343 * This should really be done by the core VM, but until that happens
1344 * filesystems like XFS, btrfs and ext4 have to take care of this
1347 if (current->flags & PF_MEMALLOC)
1351 * We need a transaction if:
1352 * 1. There are delalloc buffers on the page
1353 * 2. The page is uptodate and we have unmapped buffers
1354 * 3. The page is uptodate and we have no buffers
1355 * 4. There are unwritten buffers on the page
1358 if (!page_has_buffers(page)) {
1362 xfs_count_page_state(page, &delalloc, &unmapped, &unwritten);
1363 if (!PageUptodate(page))
1365 need_trans = delalloc + unmapped + unwritten;
1369 * If we need a transaction and the process flags say
1370 * we are already in a transaction, or no IO is allowed
1371 * then mark the page dirty again and leave the page
1374 if (current_test_flags(PF_FSTRANS) && need_trans)
1378 * Delay hooking up buffer heads until we have
1379 * made our go/no-go decision.
1381 if (!page_has_buffers(page))
1382 create_empty_buffers(page, 1 << inode->i_blkbits, 0);
1385 * Convert delayed allocate, unwritten or unmapped space
1386 * to real space and flush out to disk.
1388 error = xfs_page_state_convert(inode, page, wbc, 1, unmapped);
1389 if (error == -EAGAIN)
1391 if (unlikely(error < 0))
1397 redirty_page_for_writepage(wbc, page);
1407 struct address_space *mapping,
1408 struct writeback_control *wbc)
1410 xfs_iflags_clear(XFS_I(mapping->host), XFS_ITRUNCATED);
1411 return generic_writepages(mapping, wbc);
1415 * Called to move a page into cleanable state - and from there
1416 * to be released. Possibly the page is already clean. We always
1417 * have buffer heads in this call.
1419 * Returns 0 if the page is ok to release, 1 otherwise.
1421 * Possible scenarios are:
1423 * 1. We are being called to release a page which has been written
1424 * to via regular I/O. buffer heads will be dirty and possibly
1425 * delalloc. If no delalloc buffer heads in this case then we
1426 * can just return zero.
1428 * 2. We are called to release a page which has been written via
1429 * mmap, all we need to do is ensure there is no delalloc
1430 * state in the buffer heads, if not we can let the caller
1431 * free them and we should come back later via writepage.
1438 struct inode *inode = page->mapping->host;
1439 int dirty, delalloc, unmapped, unwritten;
1440 struct writeback_control wbc = {
1441 .sync_mode = WB_SYNC_ALL,
1445 trace_xfs_releasepage(inode, page, 0);
1447 if (!page_has_buffers(page))
1450 xfs_count_page_state(page, &delalloc, &unmapped, &unwritten);
1451 if (!delalloc && !unwritten)
1454 if (!(gfp_mask & __GFP_FS))
1457 /* If we are already inside a transaction or the thread cannot
1458 * do I/O, we cannot release this page.
1460 if (current_test_flags(PF_FSTRANS))
1464 * Convert delalloc space to real space, do not flush the
1465 * data out to disk, that will be done by the caller.
1466 * Never need to allocate space here - we will always
1467 * come back to writepage in that case.
1469 dirty = xfs_page_state_convert(inode, page, &wbc, 0, 0);
1470 if (dirty == 0 && !unwritten)
1475 return try_to_free_buffers(page);
1480 struct inode *inode,
1482 struct buffer_head *bh_result,
1485 bmapi_flags_t flags)
1487 struct xfs_bmbt_irec imap;
1494 offset = (xfs_off_t)iblock << inode->i_blkbits;
1495 ASSERT(bh_result->b_size >= (1 << inode->i_blkbits));
1496 size = bh_result->b_size;
1498 if (!create && direct && offset >= i_size_read(inode))
1501 error = xfs_iomap(XFS_I(inode), offset, size,
1502 create ? flags : BMAPI_READ, &imap, &nimap, &new);
1508 if (imap.br_startblock != HOLESTARTBLOCK &&
1509 imap.br_startblock != DELAYSTARTBLOCK) {
1511 * For unwritten extents do not report a disk address on
1512 * the read case (treat as if we're reading into a hole).
1514 if (create || !ISUNWRITTEN(&imap))
1515 xfs_map_buffer(inode, bh_result, &imap, offset);
1516 if (create && ISUNWRITTEN(&imap)) {
1518 bh_result->b_private = inode;
1519 set_buffer_unwritten(bh_result);
1524 * If this is a realtime file, data may be on a different device.
1525 * to that pointed to from the buffer_head b_bdev currently.
1527 bh_result->b_bdev = xfs_find_bdev_for_inode(inode);
1530 * If we previously allocated a block out beyond eof and we are now
1531 * coming back to use it then we will need to flag it as new even if it
1532 * has a disk address.
1534 * With sub-block writes into unwritten extents we also need to mark
1535 * the buffer as new so that the unwritten parts of the buffer gets
1539 ((!buffer_mapped(bh_result) && !buffer_uptodate(bh_result)) ||
1540 (offset >= i_size_read(inode)) ||
1541 (new || ISUNWRITTEN(&imap))))
1542 set_buffer_new(bh_result);
1544 if (imap.br_startblock == DELAYSTARTBLOCK) {
1547 set_buffer_uptodate(bh_result);
1548 set_buffer_mapped(bh_result);
1549 set_buffer_delay(bh_result);
1554 * If this is O_DIRECT or the mpage code calling tell them how large
1555 * the mapping is, so that we can avoid repeated get_blocks calls.
1557 if (direct || size > (1 << inode->i_blkbits)) {
1558 xfs_off_t mapping_size;
1560 mapping_size = imap.br_startoff + imap.br_blockcount - iblock;
1561 mapping_size <<= inode->i_blkbits;
1563 ASSERT(mapping_size > 0);
1564 if (mapping_size > size)
1565 mapping_size = size;
1566 if (mapping_size > LONG_MAX)
1567 mapping_size = LONG_MAX;
1569 bh_result->b_size = mapping_size;
1577 struct inode *inode,
1579 struct buffer_head *bh_result,
1582 return __xfs_get_blocks(inode, iblock,
1583 bh_result, create, 0, BMAPI_WRITE);
1587 xfs_get_blocks_direct(
1588 struct inode *inode,
1590 struct buffer_head *bh_result,
1593 return __xfs_get_blocks(inode, iblock,
1594 bh_result, create, 1, BMAPI_WRITE|BMAPI_DIRECT);
1604 xfs_ioend_t *ioend = iocb->private;
1607 * Non-NULL private data means we need to issue a transaction to
1608 * convert a range from unwritten to written extents. This needs
1609 * to happen from process context but aio+dio I/O completion
1610 * happens from irq context so we need to defer it to a workqueue.
1611 * This is not necessary for synchronous direct I/O, but we do
1612 * it anyway to keep the code uniform and simpler.
1614 * Well, if only it were that simple. Because synchronous direct I/O
1615 * requires extent conversion to occur *before* we return to userspace,
1616 * we have to wait for extent conversion to complete. Look at the
1617 * iocb that has been passed to us to determine if this is AIO or
1618 * not. If it is synchronous, tell xfs_finish_ioend() to kick the
1619 * workqueue and wait for it to complete.
1621 * The core direct I/O code might be changed to always call the
1622 * completion handler in the future, in which case all this can
1625 ioend->io_offset = offset;
1626 ioend->io_size = size;
1627 if (ioend->io_type == IO_READ) {
1628 xfs_finish_ioend(ioend, 0);
1629 } else if (private && size > 0) {
1630 xfs_finish_ioend(ioend, is_sync_kiocb(iocb));
1633 * A direct I/O write ioend starts it's life in unwritten
1634 * state in case they map an unwritten extent. This write
1635 * didn't map an unwritten extent so switch it's completion
1638 ioend->io_type = IO_NEW;
1639 xfs_finish_ioend(ioend, 0);
1643 * blockdev_direct_IO can return an error even after the I/O
1644 * completion handler was called. Thus we need to protect
1645 * against double-freeing.
1647 iocb->private = NULL;
1654 const struct iovec *iov,
1656 unsigned long nr_segs)
1658 struct file *file = iocb->ki_filp;
1659 struct inode *inode = file->f_mapping->host;
1660 struct block_device *bdev;
1663 bdev = xfs_find_bdev_for_inode(inode);
1665 iocb->private = xfs_alloc_ioend(inode, rw == WRITE ?
1666 IO_UNWRITTEN : IO_READ);
1668 ret = blockdev_direct_IO_no_locking(rw, iocb, inode, bdev, iov,
1670 xfs_get_blocks_direct,
1673 if (unlikely(ret != -EIOCBQUEUED && iocb->private))
1674 xfs_destroy_ioend(iocb->private);
1681 struct address_space *mapping,
1685 struct page **pagep,
1689 return block_write_begin(file, mapping, pos, len, flags, pagep, fsdata,
1695 struct address_space *mapping,
1698 struct inode *inode = (struct inode *)mapping->host;
1699 struct xfs_inode *ip = XFS_I(inode);
1701 xfs_itrace_entry(XFS_I(inode));
1702 xfs_ilock(ip, XFS_IOLOCK_SHARED);
1703 xfs_flush_pages(ip, (xfs_off_t)0, -1, 0, FI_REMAPF);
1704 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
1705 return generic_block_bmap(mapping, block, xfs_get_blocks);
1710 struct file *unused,
1713 return mpage_readpage(page, xfs_get_blocks);
1718 struct file *unused,
1719 struct address_space *mapping,
1720 struct list_head *pages,
1723 return mpage_readpages(mapping, pages, nr_pages, xfs_get_blocks);
1726 const struct address_space_operations xfs_address_space_operations = {
1727 .readpage = xfs_vm_readpage,
1728 .readpages = xfs_vm_readpages,
1729 .writepage = xfs_vm_writepage,
1730 .writepages = xfs_vm_writepages,
1731 .sync_page = block_sync_page,
1732 .releasepage = xfs_vm_releasepage,
1733 .invalidatepage = xfs_vm_invalidatepage,
1734 .write_begin = xfs_vm_write_begin,
1735 .write_end = generic_write_end,
1736 .bmap = xfs_vm_bmap,
1737 .direct_IO = xfs_vm_direct_IO,
1738 .migratepage = buffer_migrate_page,
1739 .is_partially_uptodate = block_is_partially_uptodate,
1740 .error_remove_page = generic_error_remove_page,