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>
50 * Prime number of hash buckets since address is used as the key.
53 #define to_ioend_wq(v) (&xfs_ioend_wq[((unsigned long)v) % NVSYNC])
54 static wait_queue_head_t xfs_ioend_wq[NVSYNC];
61 for (i = 0; i < NVSYNC; i++)
62 init_waitqueue_head(&xfs_ioend_wq[i]);
69 wait_queue_head_t *wq = to_ioend_wq(ip);
71 wait_event(*wq, (atomic_read(&ip->i_iocount) == 0));
78 if (atomic_dec_and_test(&ip->i_iocount))
79 wake_up(to_ioend_wq(ip));
89 struct buffer_head *bh, *head;
91 *delalloc = *unmapped = *unwritten = 0;
93 bh = head = page_buffers(page);
95 if (buffer_uptodate(bh) && !buffer_mapped(bh))
97 else if (buffer_unwritten(bh))
99 else if (buffer_delay(bh))
101 } while ((bh = bh->b_this_page) != head);
104 STATIC struct block_device *
105 xfs_find_bdev_for_inode(
108 struct xfs_inode *ip = XFS_I(inode);
109 struct xfs_mount *mp = ip->i_mount;
111 if (XFS_IS_REALTIME_INODE(ip))
112 return mp->m_rtdev_targp->bt_bdev;
114 return mp->m_ddev_targp->bt_bdev;
118 * We're now finished for good with this ioend structure.
119 * Update the page state via the associated buffer_heads,
120 * release holds on the inode and bio, and finally free
121 * up memory. Do not use the ioend after this.
127 struct buffer_head *bh, *next;
128 struct xfs_inode *ip = XFS_I(ioend->io_inode);
130 for (bh = ioend->io_buffer_head; bh; bh = next) {
131 next = bh->b_private;
132 bh->b_end_io(bh, !ioend->io_error);
136 * Volume managers supporting multiple paths can send back ENODEV
137 * when the final path disappears. In this case continuing to fill
138 * the page cache with dirty data which cannot be written out is
139 * evil, so prevent that.
141 if (unlikely(ioend->io_error == -ENODEV)) {
142 xfs_do_force_shutdown(ip->i_mount, SHUTDOWN_DEVICE_REQ,
147 mempool_free(ioend, xfs_ioend_pool);
151 * If the end of the current ioend is beyond the current EOF,
152 * return the new EOF value, otherwise zero.
158 xfs_inode_t *ip = XFS_I(ioend->io_inode);
162 bsize = ioend->io_offset + ioend->io_size;
163 isize = MAX(ip->i_size, ip->i_new_size);
164 isize = MIN(isize, bsize);
165 return isize > ip->i_d.di_size ? isize : 0;
169 * Update on-disk file size now that data has been written to disk. The
170 * current in-memory file size is i_size. If a write is beyond eof i_new_size
171 * will be the intended file size until i_size is updated. If this write does
172 * not extend all the way to the valid file size then restrict this update to
173 * the end of the write.
175 * This function does not block as blocking on the inode lock in IO completion
176 * can lead to IO completion order dependency deadlocks.. If it can't get the
177 * inode ilock it will return EAGAIN. Callers must handle this.
183 xfs_inode_t *ip = XFS_I(ioend->io_inode);
186 ASSERT((ip->i_d.di_mode & S_IFMT) == S_IFREG);
187 ASSERT(ioend->io_type != IOMAP_READ);
189 if (unlikely(ioend->io_error))
192 if (!xfs_ilock_nowait(ip, XFS_ILOCK_EXCL))
195 isize = xfs_ioend_new_eof(ioend);
197 ip->i_d.di_size = isize;
198 xfs_mark_inode_dirty(ip);
201 xfs_iunlock(ip, XFS_ILOCK_EXCL);
206 * Schedule IO completion handling on a xfsdatad if this was
207 * the final hold on this ioend. If we are asked to wait,
208 * flush the workqueue.
215 if (atomic_dec_and_test(&ioend->io_remaining)) {
216 struct workqueue_struct *wq;
218 wq = (ioend->io_type == IOMAP_UNWRITTEN) ?
219 xfsconvertd_workqueue : xfsdatad_workqueue;
220 queue_work(wq, &ioend->io_work);
227 * IO write completion.
231 struct work_struct *work)
233 xfs_ioend_t *ioend = container_of(work, xfs_ioend_t, io_work);
234 struct xfs_inode *ip = XFS_I(ioend->io_inode);
238 * For unwritten extents we need to issue transactions to convert a
239 * range to normal written extens after the data I/O has finished.
241 if (ioend->io_type == IOMAP_UNWRITTEN &&
242 likely(!ioend->io_error && !XFS_FORCED_SHUTDOWN(ip->i_mount))) {
244 error = xfs_iomap_write_unwritten(ip, ioend->io_offset,
247 ioend->io_error = error;
251 * We might have to update the on-disk file size after extending
254 if (ioend->io_type != IOMAP_READ) {
255 error = xfs_setfilesize(ioend);
256 ASSERT(!error || error == EAGAIN);
260 * If we didn't complete processing of the ioend, requeue it to the
261 * tail of the workqueue for another attempt later. Otherwise destroy
264 if (error == EAGAIN) {
265 atomic_inc(&ioend->io_remaining);
266 xfs_finish_ioend(ioend, 0);
267 /* ensure we don't spin on blocked ioends */
270 xfs_destroy_ioend(ioend);
274 * Allocate and initialise an IO completion structure.
275 * We need to track unwritten extent write completion here initially.
276 * We'll need to extend this for updating the ondisk inode size later
286 ioend = mempool_alloc(xfs_ioend_pool, GFP_NOFS);
289 * Set the count to 1 initially, which will prevent an I/O
290 * completion callback from happening before we have started
291 * all the I/O from calling the completion routine too early.
293 atomic_set(&ioend->io_remaining, 1);
295 ioend->io_list = NULL;
296 ioend->io_type = type;
297 ioend->io_inode = inode;
298 ioend->io_buffer_head = NULL;
299 ioend->io_buffer_tail = NULL;
300 atomic_inc(&XFS_I(ioend->io_inode)->i_iocount);
301 ioend->io_offset = 0;
304 INIT_WORK(&ioend->io_work, xfs_end_io);
318 return -xfs_iomap(XFS_I(inode), offset, count, flags, mapp, &nmaps);
327 struct xfs_mount *mp = XFS_I(inode)->i_mount;
328 xfs_off_t iomap_offset = XFS_FSB_TO_B(mp, iomapp->iomap_offset);
329 xfs_off_t iomap_bsize = XFS_FSB_TO_B(mp, iomapp->iomap_bsize);
331 return offset >= iomap_offset &&
332 offset < iomap_offset + iomap_bsize;
336 * BIO completion handler for buffered IO.
343 xfs_ioend_t *ioend = bio->bi_private;
345 ASSERT(atomic_read(&bio->bi_cnt) >= 1);
346 ioend->io_error = test_bit(BIO_UPTODATE, &bio->bi_flags) ? 0 : error;
348 /* Toss bio and pass work off to an xfsdatad thread */
349 bio->bi_private = NULL;
350 bio->bi_end_io = NULL;
353 xfs_finish_ioend(ioend, 0);
357 xfs_submit_ioend_bio(
358 struct writeback_control *wbc,
362 atomic_inc(&ioend->io_remaining);
363 bio->bi_private = ioend;
364 bio->bi_end_io = xfs_end_bio;
367 * If the I/O is beyond EOF we mark the inode dirty immediately
368 * but don't update the inode size until I/O completion.
370 if (xfs_ioend_new_eof(ioend))
371 xfs_mark_inode_dirty(XFS_I(ioend->io_inode));
373 submit_bio(wbc->sync_mode == WB_SYNC_ALL ?
374 WRITE_SYNC_PLUG : WRITE, bio);
375 ASSERT(!bio_flagged(bio, BIO_EOPNOTSUPP));
381 struct buffer_head *bh)
384 int nvecs = bio_get_nr_vecs(bh->b_bdev);
387 bio = bio_alloc(GFP_NOIO, nvecs);
391 ASSERT(bio->bi_private == NULL);
392 bio->bi_sector = bh->b_blocknr * (bh->b_size >> 9);
393 bio->bi_bdev = bh->b_bdev;
399 xfs_start_buffer_writeback(
400 struct buffer_head *bh)
402 ASSERT(buffer_mapped(bh));
403 ASSERT(buffer_locked(bh));
404 ASSERT(!buffer_delay(bh));
405 ASSERT(!buffer_unwritten(bh));
407 mark_buffer_async_write(bh);
408 set_buffer_uptodate(bh);
409 clear_buffer_dirty(bh);
413 xfs_start_page_writeback(
418 ASSERT(PageLocked(page));
419 ASSERT(!PageWriteback(page));
421 clear_page_dirty_for_io(page);
422 set_page_writeback(page);
424 /* If no buffers on the page are to be written, finish it here */
426 end_page_writeback(page);
429 static inline int bio_add_buffer(struct bio *bio, struct buffer_head *bh)
431 return bio_add_page(bio, bh->b_page, bh->b_size, bh_offset(bh));
435 * Submit all of the bios for all of the ioends we have saved up, covering the
436 * initial writepage page and also any probed pages.
438 * Because we may have multiple ioends spanning a page, we need to start
439 * writeback on all the buffers before we submit them for I/O. If we mark the
440 * buffers as we got, then we can end up with a page that only has buffers
441 * marked async write and I/O complete on can occur before we mark the other
442 * buffers async write.
444 * The end result of this is that we trip a bug in end_page_writeback() because
445 * we call it twice for the one page as the code in end_buffer_async_write()
446 * assumes that all buffers on the page are started at the same time.
448 * The fix is two passes across the ioend list - one to start writeback on the
449 * buffer_heads, and then submit them for I/O on the second pass.
453 struct writeback_control *wbc,
456 xfs_ioend_t *head = ioend;
458 struct buffer_head *bh;
460 sector_t lastblock = 0;
462 /* Pass 1 - start writeback */
464 next = ioend->io_list;
465 for (bh = ioend->io_buffer_head; bh; bh = bh->b_private) {
466 xfs_start_buffer_writeback(bh);
468 } while ((ioend = next) != NULL);
470 /* Pass 2 - submit I/O */
473 next = ioend->io_list;
476 for (bh = ioend->io_buffer_head; bh; bh = bh->b_private) {
480 bio = xfs_alloc_ioend_bio(bh);
481 } else if (bh->b_blocknr != lastblock + 1) {
482 xfs_submit_ioend_bio(wbc, ioend, bio);
486 if (bio_add_buffer(bio, bh) != bh->b_size) {
487 xfs_submit_ioend_bio(wbc, ioend, bio);
491 lastblock = bh->b_blocknr;
494 xfs_submit_ioend_bio(wbc, ioend, bio);
495 xfs_finish_ioend(ioend, 0);
496 } while ((ioend = next) != NULL);
500 * Cancel submission of all buffer_heads so far in this endio.
501 * Toss the endio too. Only ever called for the initial page
502 * in a writepage request, so only ever one page.
509 struct buffer_head *bh, *next_bh;
512 next = ioend->io_list;
513 bh = ioend->io_buffer_head;
515 next_bh = bh->b_private;
516 clear_buffer_async_write(bh);
518 } while ((bh = next_bh) != NULL);
520 xfs_ioend_wake(XFS_I(ioend->io_inode));
521 mempool_free(ioend, xfs_ioend_pool);
522 } while ((ioend = next) != NULL);
526 * Test to see if we've been building up a completion structure for
527 * earlier buffers -- if so, we try to append to this ioend if we
528 * can, otherwise we finish off any current ioend and start another.
529 * Return true if we've finished the given ioend.
534 struct buffer_head *bh,
537 xfs_ioend_t **result,
540 xfs_ioend_t *ioend = *result;
542 if (!ioend || need_ioend || type != ioend->io_type) {
543 xfs_ioend_t *previous = *result;
545 ioend = xfs_alloc_ioend(inode, type);
546 ioend->io_offset = offset;
547 ioend->io_buffer_head = bh;
548 ioend->io_buffer_tail = bh;
550 previous->io_list = ioend;
553 ioend->io_buffer_tail->b_private = bh;
554 ioend->io_buffer_tail = bh;
557 bh->b_private = NULL;
558 ioend->io_size += bh->b_size;
564 struct buffer_head *bh,
569 struct xfs_mount *m = XFS_I(inode)->i_mount;
570 xfs_off_t iomap_offset = XFS_FSB_TO_B(m, mp->iomap_offset);
572 ASSERT(mp->iomap_bn != IOMAP_DADDR_NULL);
574 bn = (mp->iomap_bn >> (inode->i_blkbits - BBSHIFT)) +
575 ((offset - iomap_offset) >> inode->i_blkbits);
577 ASSERT(bn || XFS_IS_REALTIME_INODE(XFS_I(inode)));
580 set_buffer_mapped(bh);
586 struct buffer_head *bh,
590 ASSERT(!(iomapp->iomap_flags & IOMAP_HOLE));
591 ASSERT(!(iomapp->iomap_flags & IOMAP_DELAY));
594 xfs_map_buffer(inode, bh, iomapp, offset);
595 bh->b_bdev = xfs_find_bdev_for_inode(inode);
596 set_buffer_mapped(bh);
597 clear_buffer_delay(bh);
598 clear_buffer_unwritten(bh);
602 * Look for a page at index that is suitable for clustering.
607 unsigned int pg_offset,
612 if (PageWriteback(page))
615 if (page->mapping && PageDirty(page)) {
616 if (page_has_buffers(page)) {
617 struct buffer_head *bh, *head;
619 bh = head = page_buffers(page);
621 if (!buffer_uptodate(bh))
623 if (mapped != buffer_mapped(bh))
626 if (ret >= pg_offset)
628 } while ((bh = bh->b_this_page) != head);
630 ret = mapped ? 0 : PAGE_CACHE_SIZE;
639 struct page *startpage,
640 struct buffer_head *bh,
641 struct buffer_head *head,
645 pgoff_t tindex, tlast, tloff;
649 /* First sum forwards in this page */
651 if (!buffer_uptodate(bh) || (mapped != buffer_mapped(bh)))
654 } while ((bh = bh->b_this_page) != head);
656 /* if we reached the end of the page, sum forwards in following pages */
657 tlast = i_size_read(inode) >> PAGE_CACHE_SHIFT;
658 tindex = startpage->index + 1;
660 /* Prune this back to avoid pathological behavior */
661 tloff = min(tlast, startpage->index + 64);
663 pagevec_init(&pvec, 0);
664 while (!done && tindex <= tloff) {
665 unsigned len = min_t(pgoff_t, PAGEVEC_SIZE, tlast - tindex + 1);
667 if (!pagevec_lookup(&pvec, inode->i_mapping, tindex, len))
670 for (i = 0; i < pagevec_count(&pvec); i++) {
671 struct page *page = pvec.pages[i];
672 size_t pg_offset, pg_len = 0;
674 if (tindex == tlast) {
676 i_size_read(inode) & (PAGE_CACHE_SIZE - 1);
682 pg_offset = PAGE_CACHE_SIZE;
684 if (page->index == tindex && trylock_page(page)) {
685 pg_len = xfs_probe_page(page, pg_offset, mapped);
698 pagevec_release(&pvec);
706 * Test if a given page is suitable for writing as part of an unwritten
707 * or delayed allocate extent.
714 if (PageWriteback(page))
717 if (page->mapping && page_has_buffers(page)) {
718 struct buffer_head *bh, *head;
721 bh = head = page_buffers(page);
723 if (buffer_unwritten(bh))
724 acceptable = (type == IOMAP_UNWRITTEN);
725 else if (buffer_delay(bh))
726 acceptable = (type == IOMAP_DELAY);
727 else if (buffer_dirty(bh) && buffer_mapped(bh))
728 acceptable = (type == IOMAP_NEW);
731 } while ((bh = bh->b_this_page) != head);
741 * Allocate & map buffers for page given the extent map. Write it out.
742 * except for the original page of a writepage, this is called on
743 * delalloc/unwritten pages only, for the original page it is possible
744 * that the page has no mapping at all.
752 xfs_ioend_t **ioendp,
753 struct writeback_control *wbc,
757 struct buffer_head *bh, *head;
758 xfs_off_t end_offset;
759 unsigned long p_offset;
762 int count = 0, done = 0, uptodate = 1;
763 xfs_off_t offset = page_offset(page);
765 if (page->index != tindex)
767 if (!trylock_page(page))
769 if (PageWriteback(page))
770 goto fail_unlock_page;
771 if (page->mapping != inode->i_mapping)
772 goto fail_unlock_page;
773 if (!xfs_is_delayed_page(page, (*ioendp)->io_type))
774 goto fail_unlock_page;
777 * page_dirty is initially a count of buffers on the page before
778 * EOF and is decremented as we move each into a cleanable state.
782 * End offset is the highest offset that this page should represent.
783 * If we are on the last page, (end_offset & (PAGE_CACHE_SIZE - 1))
784 * will evaluate non-zero and be less than PAGE_CACHE_SIZE and
785 * hence give us the correct page_dirty count. On any other page,
786 * it will be zero and in that case we need page_dirty to be the
787 * count of buffers on the page.
789 end_offset = min_t(unsigned long long,
790 (xfs_off_t)(page->index + 1) << PAGE_CACHE_SHIFT,
793 len = 1 << inode->i_blkbits;
794 p_offset = min_t(unsigned long, end_offset & (PAGE_CACHE_SIZE - 1),
796 p_offset = p_offset ? roundup(p_offset, len) : PAGE_CACHE_SIZE;
797 page_dirty = p_offset / len;
799 bh = head = page_buffers(page);
801 if (offset >= end_offset)
803 if (!buffer_uptodate(bh))
805 if (!(PageUptodate(page) || buffer_uptodate(bh))) {
810 if (buffer_unwritten(bh) || buffer_delay(bh)) {
811 if (buffer_unwritten(bh))
812 type = IOMAP_UNWRITTEN;
816 if (!xfs_iomap_valid(inode, mp, offset)) {
821 ASSERT(!(mp->iomap_flags & IOMAP_HOLE));
822 ASSERT(!(mp->iomap_flags & IOMAP_DELAY));
824 xfs_map_at_offset(inode, bh, mp, offset);
826 xfs_add_to_ioend(inode, bh, offset,
829 set_buffer_dirty(bh);
831 mark_buffer_dirty(bh);
837 if (buffer_mapped(bh) && all_bh && startio) {
839 xfs_add_to_ioend(inode, bh, offset,
847 } while (offset += len, (bh = bh->b_this_page) != head);
849 if (uptodate && bh == head)
850 SetPageUptodate(page);
855 if (wbc->nr_to_write <= 0)
858 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.
877 xfs_ioend_t **ioendp,
878 struct writeback_control *wbc,
886 pagevec_init(&pvec, 0);
887 while (!done && tindex <= tlast) {
888 unsigned len = min_t(pgoff_t, PAGEVEC_SIZE, tlast - tindex + 1);
890 if (!pagevec_lookup(&pvec, inode->i_mapping, tindex, len))
893 for (i = 0; i < pagevec_count(&pvec); i++) {
894 done = xfs_convert_page(inode, pvec.pages[i], tindex++,
895 iomapp, ioendp, wbc, startio, all_bh);
900 pagevec_release(&pvec);
906 xfs_vm_invalidatepage(
908 unsigned long offset)
910 trace_xfs_invalidatepage(page->mapping->host, page, offset);
911 block_invalidatepage(page, offset);
915 * If the page has delalloc buffers on it, we need to punch them out before we
916 * invalidate the page. If we don't, we leave a stale delalloc mapping on the
917 * inode that can trip a BUG() in xfs_get_blocks() later on if a direct IO read
918 * is done on that same region - the delalloc extent is returned when none is
919 * supposed to be there.
921 * We prevent this by truncating away the delalloc regions on the page before
922 * invalidating it. Because they are delalloc, we can do this without needing a
923 * transaction. Indeed - if we get ENOSPC errors, we have to be able to do this
924 * truncation without a transaction as there is no space left for block
925 * reservation (typically why we see a ENOSPC in writeback).
927 * This is not a performance critical path, so for now just do the punching a
928 * buffer head at a time.
931 xfs_aops_discard_page(
934 struct inode *inode = page->mapping->host;
935 struct xfs_inode *ip = XFS_I(inode);
936 struct buffer_head *bh, *head;
937 loff_t offset = page_offset(page);
938 ssize_t len = 1 << inode->i_blkbits;
940 if (!xfs_is_delayed_page(page, IOMAP_DELAY))
943 if (XFS_FORCED_SHUTDOWN(ip->i_mount))
946 xfs_fs_cmn_err(CE_ALERT, ip->i_mount,
947 "page discard on page %p, inode 0x%llx, offset %llu.",
948 page, ip->i_ino, offset);
950 xfs_ilock(ip, XFS_ILOCK_EXCL);
951 bh = head = page_buffers(page);
954 xfs_fileoff_t offset_fsb;
955 xfs_bmbt_irec_t imap;
958 xfs_fsblock_t firstblock;
959 xfs_bmap_free_t flist;
961 if (!buffer_delay(bh))
964 offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
967 * Map the range first and check that it is a delalloc extent
968 * before trying to unmap the range. Otherwise we will be
969 * trying to remove a real extent (which requires a
970 * transaction) or a hole, which is probably a bad idea...
972 error = xfs_bmapi(NULL, ip, offset_fsb, 1,
973 XFS_BMAPI_ENTIRE, NULL, 0, &imap,
974 &nimaps, NULL, NULL);
977 /* something screwed, just bail */
978 if (!XFS_FORCED_SHUTDOWN(ip->i_mount)) {
979 xfs_fs_cmn_err(CE_ALERT, ip->i_mount,
980 "page discard failed delalloc mapping lookup.");
988 if (imap.br_startblock != DELAYSTARTBLOCK) {
989 /* been converted, ignore */
992 WARN_ON(imap.br_blockcount == 0);
995 * Note: while we initialise the firstblock/flist pair, they
996 * should never be used because blocks should never be
997 * allocated or freed for a delalloc extent and hence we need
998 * don't cancel or finish them after the xfs_bunmapi() call.
1000 xfs_bmap_init(&flist, &firstblock);
1001 error = xfs_bunmapi(NULL, ip, offset_fsb, 1, 0, 1, &firstblock,
1002 &flist, NULL, &done);
1004 ASSERT(!flist.xbf_count && !flist.xbf_first);
1006 /* something screwed, just bail */
1007 if (!XFS_FORCED_SHUTDOWN(ip->i_mount)) {
1008 xfs_fs_cmn_err(CE_ALERT, ip->i_mount,
1009 "page discard unable to remove delalloc mapping.");
1016 } while ((bh = bh->b_this_page) != head);
1018 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1020 xfs_vm_invalidatepage(page, 0);
1025 * Calling this without startio set means we are being asked to make a dirty
1026 * page ready for freeing it's buffers. When called with startio set then
1027 * we are coming from writepage.
1029 * When called with startio set it is important that we write the WHOLE
1031 * The bh->b_state's cannot know if any of the blocks or which block for
1032 * that matter are dirty due to mmap writes, and therefore bh uptodate is
1033 * only valid if the page itself isn't completely uptodate. Some layers
1034 * may clear the page dirty flag prior to calling write page, under the
1035 * assumption the entire page will be written out; by not writing out the
1036 * whole page the page can be reused before all valid dirty data is
1037 * written out. Note: in the case of a page that has been dirty'd by
1038 * mapwrite and but partially setup by block_prepare_write the
1039 * bh->b_states's will not agree and only ones setup by BPW/BCW will have
1040 * valid state, thus the whole page must be written out thing.
1044 xfs_page_state_convert(
1045 struct inode *inode,
1047 struct writeback_control *wbc,
1049 int unmapped) /* also implies page uptodate */
1051 struct buffer_head *bh, *head;
1053 xfs_ioend_t *ioend = NULL, *iohead = NULL;
1055 unsigned long p_offset = 0;
1057 __uint64_t end_offset;
1058 pgoff_t end_index, last_index, tlast;
1060 int flags, err, iomap_valid = 0, uptodate = 1;
1061 int page_dirty, count = 0;
1063 int all_bh = unmapped;
1066 if (wbc->sync_mode == WB_SYNC_NONE && wbc->nonblocking)
1067 trylock |= BMAPI_TRYLOCK;
1070 /* Is this page beyond the end of the file? */
1071 offset = i_size_read(inode);
1072 end_index = offset >> PAGE_CACHE_SHIFT;
1073 last_index = (offset - 1) >> PAGE_CACHE_SHIFT;
1074 if (page->index >= end_index) {
1075 if ((page->index >= end_index + 1) ||
1076 !(i_size_read(inode) & (PAGE_CACHE_SIZE - 1))) {
1084 * page_dirty is initially a count of buffers on the page before
1085 * EOF and is decremented as we move each into a cleanable state.
1089 * End offset is the highest offset that this page should represent.
1090 * If we are on the last page, (end_offset & (PAGE_CACHE_SIZE - 1))
1091 * will evaluate non-zero and be less than PAGE_CACHE_SIZE and
1092 * hence give us the correct page_dirty count. On any other page,
1093 * it will be zero and in that case we need page_dirty to be the
1094 * count of buffers on the page.
1096 end_offset = min_t(unsigned long long,
1097 (xfs_off_t)(page->index + 1) << PAGE_CACHE_SHIFT, offset);
1098 len = 1 << inode->i_blkbits;
1099 p_offset = min_t(unsigned long, end_offset & (PAGE_CACHE_SIZE - 1),
1101 p_offset = p_offset ? roundup(p_offset, len) : PAGE_CACHE_SIZE;
1102 page_dirty = p_offset / len;
1104 bh = head = page_buffers(page);
1105 offset = page_offset(page);
1109 /* TODO: cleanup count and page_dirty */
1112 if (offset >= end_offset)
1114 if (!buffer_uptodate(bh))
1116 if (!(PageUptodate(page) || buffer_uptodate(bh)) && !startio) {
1118 * the iomap is actually still valid, but the ioend
1119 * isn't. shouldn't happen too often.
1126 iomap_valid = xfs_iomap_valid(inode, &iomap, offset);
1129 * First case, map an unwritten extent and prepare for
1130 * extent state conversion transaction on completion.
1132 * Second case, allocate space for a delalloc buffer.
1133 * We can return EAGAIN here in the release page case.
1135 * Third case, an unmapped buffer was found, and we are
1136 * in a path where we need to write the whole page out.
1138 if (buffer_unwritten(bh) || buffer_delay(bh) ||
1139 ((buffer_uptodate(bh) || PageUptodate(page)) &&
1140 !buffer_mapped(bh) && (unmapped || startio))) {
1144 * Make sure we don't use a read-only iomap
1146 if (flags == BMAPI_READ)
1149 if (buffer_unwritten(bh)) {
1150 type = IOMAP_UNWRITTEN;
1151 flags = BMAPI_WRITE | BMAPI_IGNSTATE;
1152 } else if (buffer_delay(bh)) {
1154 flags = BMAPI_ALLOCATE | trylock;
1157 flags = BMAPI_WRITE | BMAPI_MMAP;
1162 * if we didn't have a valid mapping then we
1163 * need to ensure that we put the new mapping
1164 * in a new ioend structure. This needs to be
1165 * done to ensure that the ioends correctly
1166 * reflect the block mappings at io completion
1167 * for unwritten extent conversion.
1170 if (type == IOMAP_NEW) {
1171 size = xfs_probe_cluster(inode,
1177 err = xfs_map_blocks(inode, offset, size,
1181 iomap_valid = xfs_iomap_valid(inode, &iomap, offset);
1184 xfs_map_at_offset(inode, bh, &iomap, offset);
1186 xfs_add_to_ioend(inode, bh, offset,
1190 set_buffer_dirty(bh);
1192 mark_buffer_dirty(bh);
1197 } else if (buffer_uptodate(bh) && startio) {
1199 * we got here because the buffer is already mapped.
1200 * That means it must already have extents allocated
1201 * underneath it. Map the extent by reading it.
1203 if (!iomap_valid || flags != BMAPI_READ) {
1205 size = xfs_probe_cluster(inode, page, bh,
1207 err = xfs_map_blocks(inode, offset, size,
1211 iomap_valid = xfs_iomap_valid(inode, &iomap, offset);
1215 * We set the type to IOMAP_NEW in case we are doing a
1216 * small write at EOF that is extending the file but
1217 * without needing an allocation. We need to update the
1218 * file size on I/O completion in this case so it is
1219 * the same case as having just allocated a new extent
1220 * that we are writing into for the first time.
1223 if (trylock_buffer(bh)) {
1224 ASSERT(buffer_mapped(bh));
1227 xfs_add_to_ioend(inode, bh, offset, type,
1228 &ioend, !iomap_valid);
1234 } else if ((buffer_uptodate(bh) || PageUptodate(page)) &&
1235 (unmapped || startio)) {
1242 } while (offset += len, ((bh = bh->b_this_page) != head));
1244 if (uptodate && bh == head)
1245 SetPageUptodate(page);
1248 xfs_start_page_writeback(page, 1, count);
1250 if (ioend && iomap_valid) {
1251 struct xfs_mount *m = XFS_I(inode)->i_mount;
1252 xfs_off_t iomap_offset = XFS_FSB_TO_B(m, iomap.iomap_offset);
1253 xfs_off_t iomap_bsize = XFS_FSB_TO_B(m, iomap.iomap_bsize);
1255 offset = (iomap_offset + iomap_bsize - 1) >>
1257 tlast = min_t(pgoff_t, offset, last_index);
1258 xfs_cluster_write(inode, page->index + 1, &iomap, &ioend,
1259 wbc, startio, all_bh, tlast);
1263 xfs_submit_ioend(wbc, iohead);
1269 xfs_cancel_ioend(iohead);
1272 * If it's delalloc and we have nowhere to put it,
1273 * throw it away, unless the lower layers told
1276 if (err != -EAGAIN) {
1278 xfs_aops_discard_page(page);
1279 ClearPageUptodate(page);
1285 * writepage: Called from one of two places:
1287 * 1. we are flushing a delalloc buffer head.
1289 * 2. we are writing out a dirty page. Typically the page dirty
1290 * state is cleared before we get here. In this case is it
1291 * conceivable we have no buffer heads.
1293 * For delalloc space on the page we need to allocate space and
1294 * flush it. For unmapped buffer heads on the page we should
1295 * allocate space if the page is uptodate. For any other dirty
1296 * buffer heads on the page we should flush them.
1298 * If we detect that a transaction would be required to flush
1299 * the page, we have to check the process flags first, if we
1300 * are already in a transaction or disk I/O during allocations
1301 * is off, we need to fail the writepage and redirty the page.
1307 struct writeback_control *wbc)
1311 int delalloc, unmapped, unwritten;
1312 struct inode *inode = page->mapping->host;
1314 trace_xfs_writepage(inode, page, 0);
1317 * We need a transaction if:
1318 * 1. There are delalloc buffers on the page
1319 * 2. The page is uptodate and we have unmapped buffers
1320 * 3. The page is uptodate and we have no buffers
1321 * 4. There are unwritten buffers on the page
1324 if (!page_has_buffers(page)) {
1328 xfs_count_page_state(page, &delalloc, &unmapped, &unwritten);
1329 if (!PageUptodate(page))
1331 need_trans = delalloc + unmapped + unwritten;
1335 * If we need a transaction and the process flags say
1336 * we are already in a transaction, or no IO is allowed
1337 * then mark the page dirty again and leave the page
1340 if (current_test_flags(PF_FSTRANS) && need_trans)
1344 * Delay hooking up buffer heads until we have
1345 * made our go/no-go decision.
1347 if (!page_has_buffers(page))
1348 create_empty_buffers(page, 1 << inode->i_blkbits, 0);
1352 * VM calculation for nr_to_write seems off. Bump it way
1353 * up, this gets simple streaming writes zippy again.
1354 * To be reviewed again after Jens' writeback changes.
1356 wbc->nr_to_write *= 4;
1359 * Convert delayed allocate, unwritten or unmapped space
1360 * to real space and flush out to disk.
1362 error = xfs_page_state_convert(inode, page, wbc, 1, unmapped);
1363 if (error == -EAGAIN)
1365 if (unlikely(error < 0))
1371 redirty_page_for_writepage(wbc, page);
1381 struct address_space *mapping,
1382 struct writeback_control *wbc)
1384 xfs_iflags_clear(XFS_I(mapping->host), XFS_ITRUNCATED);
1385 return generic_writepages(mapping, wbc);
1389 * Called to move a page into cleanable state - and from there
1390 * to be released. Possibly the page is already clean. We always
1391 * have buffer heads in this call.
1393 * Returns 0 if the page is ok to release, 1 otherwise.
1395 * Possible scenarios are:
1397 * 1. We are being called to release a page which has been written
1398 * to via regular I/O. buffer heads will be dirty and possibly
1399 * delalloc. If no delalloc buffer heads in this case then we
1400 * can just return zero.
1402 * 2. We are called to release a page which has been written via
1403 * mmap, all we need to do is ensure there is no delalloc
1404 * state in the buffer heads, if not we can let the caller
1405 * free them and we should come back later via writepage.
1412 struct inode *inode = page->mapping->host;
1413 int dirty, delalloc, unmapped, unwritten;
1414 struct writeback_control wbc = {
1415 .sync_mode = WB_SYNC_ALL,
1419 trace_xfs_releasepage(inode, page, 0);
1421 if (!page_has_buffers(page))
1424 xfs_count_page_state(page, &delalloc, &unmapped, &unwritten);
1425 if (!delalloc && !unwritten)
1428 if (!(gfp_mask & __GFP_FS))
1431 /* If we are already inside a transaction or the thread cannot
1432 * do I/O, we cannot release this page.
1434 if (current_test_flags(PF_FSTRANS))
1438 * Convert delalloc space to real space, do not flush the
1439 * data out to disk, that will be done by the caller.
1440 * Never need to allocate space here - we will always
1441 * come back to writepage in that case.
1443 dirty = xfs_page_state_convert(inode, page, &wbc, 0, 0);
1444 if (dirty == 0 && !unwritten)
1449 return try_to_free_buffers(page);
1454 struct inode *inode,
1456 struct buffer_head *bh_result,
1459 bmapi_flags_t flags)
1467 offset = (xfs_off_t)iblock << inode->i_blkbits;
1468 ASSERT(bh_result->b_size >= (1 << inode->i_blkbits));
1469 size = bh_result->b_size;
1471 if (!create && direct && offset >= i_size_read(inode))
1474 error = xfs_iomap(XFS_I(inode), offset, size,
1475 create ? flags : BMAPI_READ, &iomap, &niomap);
1481 if (iomap.iomap_bn != IOMAP_DADDR_NULL) {
1483 * For unwritten extents do not report a disk address on
1484 * the read case (treat as if we're reading into a hole).
1486 if (create || !(iomap.iomap_flags & IOMAP_UNWRITTEN))
1487 xfs_map_buffer(inode, bh_result, &iomap, offset);
1488 if (create && (iomap.iomap_flags & IOMAP_UNWRITTEN)) {
1490 bh_result->b_private = inode;
1491 set_buffer_unwritten(bh_result);
1496 * If this is a realtime file, data may be on a different device.
1497 * to that pointed to from the buffer_head b_bdev currently.
1499 bh_result->b_bdev = xfs_find_bdev_for_inode(inode);
1502 * If we previously allocated a block out beyond eof and we are now
1503 * coming back to use it then we will need to flag it as new even if it
1504 * has a disk address.
1506 * With sub-block writes into unwritten extents we also need to mark
1507 * the buffer as new so that the unwritten parts of the buffer gets
1511 ((!buffer_mapped(bh_result) && !buffer_uptodate(bh_result)) ||
1512 (offset >= i_size_read(inode)) ||
1513 (iomap.iomap_flags & (IOMAP_NEW|IOMAP_UNWRITTEN))))
1514 set_buffer_new(bh_result);
1516 if (iomap.iomap_flags & IOMAP_DELAY) {
1519 set_buffer_uptodate(bh_result);
1520 set_buffer_mapped(bh_result);
1521 set_buffer_delay(bh_result);
1525 if (direct || size > (1 << inode->i_blkbits)) {
1526 struct xfs_mount *mp = XFS_I(inode)->i_mount;
1527 xfs_off_t iomap_offset = XFS_FSB_TO_B(mp, iomap.iomap_offset);
1528 xfs_off_t iomap_delta = offset - iomap_offset;
1529 xfs_off_t iomap_bsize = XFS_FSB_TO_B(mp, iomap.iomap_bsize);
1531 ASSERT(iomap_bsize - iomap_delta > 0);
1532 offset = min_t(xfs_off_t,
1533 iomap_bsize - iomap_delta, size);
1534 bh_result->b_size = (ssize_t)min_t(xfs_off_t, LONG_MAX, offset);
1542 struct inode *inode,
1544 struct buffer_head *bh_result,
1547 return __xfs_get_blocks(inode, iblock,
1548 bh_result, create, 0, BMAPI_WRITE);
1552 xfs_get_blocks_direct(
1553 struct inode *inode,
1555 struct buffer_head *bh_result,
1558 return __xfs_get_blocks(inode, iblock,
1559 bh_result, create, 1, BMAPI_WRITE|BMAPI_DIRECT);
1569 xfs_ioend_t *ioend = iocb->private;
1572 * Non-NULL private data means we need to issue a transaction to
1573 * convert a range from unwritten to written extents. This needs
1574 * to happen from process context but aio+dio I/O completion
1575 * happens from irq context so we need to defer it to a workqueue.
1576 * This is not necessary for synchronous direct I/O, but we do
1577 * it anyway to keep the code uniform and simpler.
1579 * Well, if only it were that simple. Because synchronous direct I/O
1580 * requires extent conversion to occur *before* we return to userspace,
1581 * we have to wait for extent conversion to complete. Look at the
1582 * iocb that has been passed to us to determine if this is AIO or
1583 * not. If it is synchronous, tell xfs_finish_ioend() to kick the
1584 * workqueue and wait for it to complete.
1586 * The core direct I/O code might be changed to always call the
1587 * completion handler in the future, in which case all this can
1590 ioend->io_offset = offset;
1591 ioend->io_size = size;
1592 if (ioend->io_type == IOMAP_READ) {
1593 xfs_finish_ioend(ioend, 0);
1594 } else if (private && size > 0) {
1595 xfs_finish_ioend(ioend, is_sync_kiocb(iocb));
1598 * A direct I/O write ioend starts it's life in unwritten
1599 * state in case they map an unwritten extent. This write
1600 * didn't map an unwritten extent so switch it's completion
1603 ioend->io_type = IOMAP_NEW;
1604 xfs_finish_ioend(ioend, 0);
1608 * blockdev_direct_IO can return an error even after the I/O
1609 * completion handler was called. Thus we need to protect
1610 * against double-freeing.
1612 iocb->private = NULL;
1619 const struct iovec *iov,
1621 unsigned long nr_segs)
1623 struct file *file = iocb->ki_filp;
1624 struct inode *inode = file->f_mapping->host;
1625 struct block_device *bdev;
1628 bdev = xfs_find_bdev_for_inode(inode);
1630 iocb->private = xfs_alloc_ioend(inode, rw == WRITE ?
1631 IOMAP_UNWRITTEN : IOMAP_READ);
1633 ret = blockdev_direct_IO_no_locking(rw, iocb, inode, bdev, iov,
1635 xfs_get_blocks_direct,
1638 if (unlikely(ret != -EIOCBQUEUED && iocb->private))
1639 xfs_destroy_ioend(iocb->private);
1646 struct address_space *mapping,
1650 struct page **pagep,
1654 return block_write_begin(file, mapping, pos, len, flags, pagep, fsdata,
1660 struct address_space *mapping,
1663 struct inode *inode = (struct inode *)mapping->host;
1664 struct xfs_inode *ip = XFS_I(inode);
1666 xfs_itrace_entry(XFS_I(inode));
1667 xfs_ilock(ip, XFS_IOLOCK_SHARED);
1668 xfs_flush_pages(ip, (xfs_off_t)0, -1, 0, FI_REMAPF);
1669 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
1670 return generic_block_bmap(mapping, block, xfs_get_blocks);
1675 struct file *unused,
1678 return mpage_readpage(page, xfs_get_blocks);
1683 struct file *unused,
1684 struct address_space *mapping,
1685 struct list_head *pages,
1688 return mpage_readpages(mapping, pages, nr_pages, xfs_get_blocks);
1691 const struct address_space_operations xfs_address_space_operations = {
1692 .readpage = xfs_vm_readpage,
1693 .readpages = xfs_vm_readpages,
1694 .writepage = xfs_vm_writepage,
1695 .writepages = xfs_vm_writepages,
1696 .sync_page = block_sync_page,
1697 .releasepage = xfs_vm_releasepage,
1698 .invalidatepage = xfs_vm_invalidatepage,
1699 .write_begin = xfs_vm_write_begin,
1700 .write_end = generic_write_end,
1701 .bmap = xfs_vm_bmap,
1702 .direct_IO = xfs_vm_direct_IO,
1703 .migratepage = buffer_migrate_page,
1704 .is_partially_uptodate = block_is_partially_uptodate,
1705 .error_remove_page = generic_error_remove_page,