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 <linux/mpage.h>
42 #include <linux/pagevec.h>
43 #include <linux/writeback.h>
47 * Prime number of hash buckets since address is used as the key.
50 #define to_ioend_wq(v) (&xfs_ioend_wq[((unsigned long)v) % NVSYNC])
51 static wait_queue_head_t xfs_ioend_wq[NVSYNC];
58 for (i = 0; i < NVSYNC; i++)
59 init_waitqueue_head(&xfs_ioend_wq[i]);
66 wait_queue_head_t *wq = to_ioend_wq(ip);
68 wait_event(*wq, (atomic_read(&ip->i_iocount) == 0));
75 if (atomic_dec_and_test(&ip->i_iocount))
76 wake_up(to_ioend_wq(ip));
86 struct buffer_head *bh, *head;
88 *delalloc = *unmapped = *unwritten = 0;
90 bh = head = page_buffers(page);
92 if (buffer_uptodate(bh) && !buffer_mapped(bh))
94 else if (buffer_unwritten(bh))
96 else if (buffer_delay(bh))
98 } while ((bh = bh->b_this_page) != head);
101 #if defined(XFS_RW_TRACE)
110 loff_t isize = i_size_read(inode);
111 loff_t offset = page_offset(page);
112 int delalloc = -1, unmapped = -1, unwritten = -1;
114 if (page_has_buffers(page))
115 xfs_count_page_state(page, &delalloc, &unmapped, &unwritten);
121 ktrace_enter(ip->i_rwtrace,
122 (void *)((unsigned long)tag),
127 (void *)((unsigned long)((ip->i_d.di_size >> 32) & 0xffffffff)),
128 (void *)((unsigned long)(ip->i_d.di_size & 0xffffffff)),
129 (void *)((unsigned long)((isize >> 32) & 0xffffffff)),
130 (void *)((unsigned long)(isize & 0xffffffff)),
131 (void *)((unsigned long)((offset >> 32) & 0xffffffff)),
132 (void *)((unsigned long)(offset & 0xffffffff)),
133 (void *)((unsigned long)delalloc),
134 (void *)((unsigned long)unmapped),
135 (void *)((unsigned long)unwritten),
136 (void *)((unsigned long)current_pid()),
140 #define xfs_page_trace(tag, inode, page, pgoff)
143 STATIC struct block_device *
144 xfs_find_bdev_for_inode(
145 struct xfs_inode *ip)
147 struct xfs_mount *mp = ip->i_mount;
149 if (XFS_IS_REALTIME_INODE(ip))
150 return mp->m_rtdev_targp->bt_bdev;
152 return mp->m_ddev_targp->bt_bdev;
156 * We're now finished for good with this ioend structure.
157 * Update the page state via the associated buffer_heads,
158 * release holds on the inode and bio, and finally free
159 * up memory. Do not use the ioend after this.
165 struct buffer_head *bh, *next;
166 struct xfs_inode *ip = XFS_I(ioend->io_inode);
168 for (bh = ioend->io_buffer_head; bh; bh = next) {
169 next = bh->b_private;
170 bh->b_end_io(bh, !ioend->io_error);
174 * Volume managers supporting multiple paths can send back ENODEV
175 * when the final path disappears. In this case continuing to fill
176 * the page cache with dirty data which cannot be written out is
177 * evil, so prevent that.
179 if (unlikely(ioend->io_error == -ENODEV)) {
180 xfs_do_force_shutdown(ip->i_mount, SHUTDOWN_DEVICE_REQ,
185 mempool_free(ioend, xfs_ioend_pool);
189 * Update on-disk file size now that data has been written to disk.
190 * The current in-memory file size is i_size. If a write is beyond
191 * eof i_new_size will be the intended file size until i_size is
192 * updated. If this write does not extend all the way to the valid
193 * file size then restrict this update to the end of the write.
199 xfs_inode_t *ip = XFS_I(ioend->io_inode);
203 ASSERT((ip->i_d.di_mode & S_IFMT) == S_IFREG);
204 ASSERT(ioend->io_type != IOMAP_READ);
206 if (unlikely(ioend->io_error))
209 bsize = ioend->io_offset + ioend->io_size;
211 xfs_ilock(ip, XFS_ILOCK_EXCL);
213 isize = MAX(ip->i_size, ip->i_new_size);
214 isize = MIN(isize, bsize);
216 if (ip->i_d.di_size < isize) {
217 ip->i_d.di_size = isize;
218 ip->i_update_core = 1;
219 xfs_mark_inode_dirty_sync(ip);
222 xfs_iunlock(ip, XFS_ILOCK_EXCL);
226 * Buffered IO write completion for delayed allocate extents.
229 xfs_end_bio_delalloc(
230 struct work_struct *work)
233 container_of(work, xfs_ioend_t, io_work);
235 xfs_setfilesize(ioend);
236 xfs_destroy_ioend(ioend);
240 * Buffered IO write completion for regular, written extents.
244 struct work_struct *work)
247 container_of(work, xfs_ioend_t, io_work);
249 xfs_setfilesize(ioend);
250 xfs_destroy_ioend(ioend);
254 * IO write completion for unwritten extents.
256 * Issue transactions to convert a buffer range from unwritten
257 * to written extents.
260 xfs_end_bio_unwritten(
261 struct work_struct *work)
264 container_of(work, xfs_ioend_t, io_work);
265 struct xfs_inode *ip = XFS_I(ioend->io_inode);
266 xfs_off_t offset = ioend->io_offset;
267 size_t size = ioend->io_size;
269 if (likely(!ioend->io_error)) {
270 if (!XFS_FORCED_SHUTDOWN(ip->i_mount)) {
272 error = xfs_iomap_write_unwritten(ip, offset, size);
274 ioend->io_error = error;
276 xfs_setfilesize(ioend);
278 xfs_destroy_ioend(ioend);
282 * IO read completion for regular, written extents.
286 struct work_struct *work)
289 container_of(work, xfs_ioend_t, io_work);
291 xfs_destroy_ioend(ioend);
295 * Schedule IO completion handling on a xfsdatad if this was
296 * the final hold on this ioend. If we are asked to wait,
297 * flush the workqueue.
304 if (atomic_dec_and_test(&ioend->io_remaining)) {
305 struct workqueue_struct *wq = xfsdatad_workqueue;
306 if (ioend->io_work.func == xfs_end_bio_unwritten)
307 wq = xfsconvertd_workqueue;
309 queue_work(wq, &ioend->io_work);
316 * Allocate and initialise an IO completion structure.
317 * We need to track unwritten extent write completion here initially.
318 * We'll need to extend this for updating the ondisk inode size later
328 ioend = mempool_alloc(xfs_ioend_pool, GFP_NOFS);
331 * Set the count to 1 initially, which will prevent an I/O
332 * completion callback from happening before we have started
333 * all the I/O from calling the completion routine too early.
335 atomic_set(&ioend->io_remaining, 1);
337 ioend->io_list = NULL;
338 ioend->io_type = type;
339 ioend->io_inode = inode;
340 ioend->io_buffer_head = NULL;
341 ioend->io_buffer_tail = NULL;
342 atomic_inc(&XFS_I(ioend->io_inode)->i_iocount);
343 ioend->io_offset = 0;
346 if (type == IOMAP_UNWRITTEN)
347 INIT_WORK(&ioend->io_work, xfs_end_bio_unwritten);
348 else if (type == IOMAP_DELAY)
349 INIT_WORK(&ioend->io_work, xfs_end_bio_delalloc);
350 else if (type == IOMAP_READ)
351 INIT_WORK(&ioend->io_work, xfs_end_bio_read);
353 INIT_WORK(&ioend->io_work, xfs_end_bio_written);
368 return -xfs_iomap(XFS_I(inode), offset, count, flags, mapp, &nmaps);
376 return offset >= iomapp->iomap_offset &&
377 offset < iomapp->iomap_offset + iomapp->iomap_bsize;
381 * BIO completion handler for buffered IO.
388 xfs_ioend_t *ioend = bio->bi_private;
390 ASSERT(atomic_read(&bio->bi_cnt) >= 1);
391 ioend->io_error = test_bit(BIO_UPTODATE, &bio->bi_flags) ? 0 : error;
393 /* Toss bio and pass work off to an xfsdatad thread */
394 bio->bi_private = NULL;
395 bio->bi_end_io = NULL;
398 xfs_finish_ioend(ioend, 0);
402 xfs_submit_ioend_bio(
406 atomic_inc(&ioend->io_remaining);
408 bio->bi_private = ioend;
409 bio->bi_end_io = xfs_end_bio;
411 submit_bio(WRITE, bio);
412 ASSERT(!bio_flagged(bio, BIO_EOPNOTSUPP));
418 struct buffer_head *bh)
421 int nvecs = bio_get_nr_vecs(bh->b_bdev);
424 bio = bio_alloc(GFP_NOIO, nvecs);
428 ASSERT(bio->bi_private == NULL);
429 bio->bi_sector = bh->b_blocknr * (bh->b_size >> 9);
430 bio->bi_bdev = bh->b_bdev;
436 xfs_start_buffer_writeback(
437 struct buffer_head *bh)
439 ASSERT(buffer_mapped(bh));
440 ASSERT(buffer_locked(bh));
441 ASSERT(!buffer_delay(bh));
442 ASSERT(!buffer_unwritten(bh));
444 mark_buffer_async_write(bh);
445 set_buffer_uptodate(bh);
446 clear_buffer_dirty(bh);
450 xfs_start_page_writeback(
455 ASSERT(PageLocked(page));
456 ASSERT(!PageWriteback(page));
458 clear_page_dirty_for_io(page);
459 set_page_writeback(page);
461 /* If no buffers on the page are to be written, finish it here */
463 end_page_writeback(page);
466 static inline int bio_add_buffer(struct bio *bio, struct buffer_head *bh)
468 return bio_add_page(bio, bh->b_page, bh->b_size, bh_offset(bh));
472 * Submit all of the bios for all of the ioends we have saved up, covering the
473 * initial writepage page and also any probed pages.
475 * Because we may have multiple ioends spanning a page, we need to start
476 * writeback on all the buffers before we submit them for I/O. If we mark the
477 * buffers as we got, then we can end up with a page that only has buffers
478 * marked async write and I/O complete on can occur before we mark the other
479 * buffers async write.
481 * The end result of this is that we trip a bug in end_page_writeback() because
482 * we call it twice for the one page as the code in end_buffer_async_write()
483 * assumes that all buffers on the page are started at the same time.
485 * The fix is two passes across the ioend list - one to start writeback on the
486 * buffer_heads, and then submit them for I/O on the second pass.
492 xfs_ioend_t *head = ioend;
494 struct buffer_head *bh;
496 sector_t lastblock = 0;
498 /* Pass 1 - start writeback */
500 next = ioend->io_list;
501 for (bh = ioend->io_buffer_head; bh; bh = bh->b_private) {
502 xfs_start_buffer_writeback(bh);
504 } while ((ioend = next) != NULL);
506 /* Pass 2 - submit I/O */
509 next = ioend->io_list;
512 for (bh = ioend->io_buffer_head; bh; bh = bh->b_private) {
516 bio = xfs_alloc_ioend_bio(bh);
517 } else if (bh->b_blocknr != lastblock + 1) {
518 xfs_submit_ioend_bio(ioend, bio);
522 if (bio_add_buffer(bio, bh) != bh->b_size) {
523 xfs_submit_ioend_bio(ioend, bio);
527 lastblock = bh->b_blocknr;
530 xfs_submit_ioend_bio(ioend, bio);
531 xfs_finish_ioend(ioend, 0);
532 } while ((ioend = next) != NULL);
536 * Cancel submission of all buffer_heads so far in this endio.
537 * Toss the endio too. Only ever called for the initial page
538 * in a writepage request, so only ever one page.
545 struct buffer_head *bh, *next_bh;
548 next = ioend->io_list;
549 bh = ioend->io_buffer_head;
551 next_bh = bh->b_private;
552 clear_buffer_async_write(bh);
554 } while ((bh = next_bh) != NULL);
556 xfs_ioend_wake(XFS_I(ioend->io_inode));
557 mempool_free(ioend, xfs_ioend_pool);
558 } while ((ioend = next) != NULL);
562 * Test to see if we've been building up a completion structure for
563 * earlier buffers -- if so, we try to append to this ioend if we
564 * can, otherwise we finish off any current ioend and start another.
565 * Return true if we've finished the given ioend.
570 struct buffer_head *bh,
573 xfs_ioend_t **result,
576 xfs_ioend_t *ioend = *result;
578 if (!ioend || need_ioend || type != ioend->io_type) {
579 xfs_ioend_t *previous = *result;
581 ioend = xfs_alloc_ioend(inode, type);
582 ioend->io_offset = offset;
583 ioend->io_buffer_head = bh;
584 ioend->io_buffer_tail = bh;
586 previous->io_list = ioend;
589 ioend->io_buffer_tail->b_private = bh;
590 ioend->io_buffer_tail = bh;
593 bh->b_private = NULL;
594 ioend->io_size += bh->b_size;
599 struct buffer_head *bh,
606 ASSERT(mp->iomap_bn != IOMAP_DADDR_NULL);
608 bn = (mp->iomap_bn >> (block_bits - BBSHIFT)) +
609 ((offset - mp->iomap_offset) >> block_bits);
611 ASSERT(bn || (mp->iomap_flags & IOMAP_REALTIME));
614 set_buffer_mapped(bh);
619 struct buffer_head *bh,
624 ASSERT(!(iomapp->iomap_flags & IOMAP_HOLE));
625 ASSERT(!(iomapp->iomap_flags & IOMAP_DELAY));
628 xfs_map_buffer(bh, iomapp, offset, block_bits);
629 bh->b_bdev = iomapp->iomap_target->bt_bdev;
630 set_buffer_mapped(bh);
631 clear_buffer_delay(bh);
632 clear_buffer_unwritten(bh);
636 * Look for a page at index that is suitable for clustering.
641 unsigned int pg_offset,
646 if (PageWriteback(page))
649 if (page->mapping && PageDirty(page)) {
650 if (page_has_buffers(page)) {
651 struct buffer_head *bh, *head;
653 bh = head = page_buffers(page);
655 if (!buffer_uptodate(bh))
657 if (mapped != buffer_mapped(bh))
660 if (ret >= pg_offset)
662 } while ((bh = bh->b_this_page) != head);
664 ret = mapped ? 0 : PAGE_CACHE_SIZE;
673 struct page *startpage,
674 struct buffer_head *bh,
675 struct buffer_head *head,
679 pgoff_t tindex, tlast, tloff;
683 /* First sum forwards in this page */
685 if (!buffer_uptodate(bh) || (mapped != buffer_mapped(bh)))
688 } while ((bh = bh->b_this_page) != head);
690 /* if we reached the end of the page, sum forwards in following pages */
691 tlast = i_size_read(inode) >> PAGE_CACHE_SHIFT;
692 tindex = startpage->index + 1;
694 /* Prune this back to avoid pathological behavior */
695 tloff = min(tlast, startpage->index + 64);
697 pagevec_init(&pvec, 0);
698 while (!done && tindex <= tloff) {
699 unsigned len = min_t(pgoff_t, PAGEVEC_SIZE, tlast - tindex + 1);
701 if (!pagevec_lookup(&pvec, inode->i_mapping, tindex, len))
704 for (i = 0; i < pagevec_count(&pvec); i++) {
705 struct page *page = pvec.pages[i];
706 size_t pg_offset, pg_len = 0;
708 if (tindex == tlast) {
710 i_size_read(inode) & (PAGE_CACHE_SIZE - 1);
716 pg_offset = PAGE_CACHE_SIZE;
718 if (page->index == tindex && trylock_page(page)) {
719 pg_len = xfs_probe_page(page, pg_offset, mapped);
732 pagevec_release(&pvec);
740 * Test if a given page is suitable for writing as part of an unwritten
741 * or delayed allocate extent.
748 if (PageWriteback(page))
751 if (page->mapping && page_has_buffers(page)) {
752 struct buffer_head *bh, *head;
755 bh = head = page_buffers(page);
757 if (buffer_unwritten(bh))
758 acceptable = (type == IOMAP_UNWRITTEN);
759 else if (buffer_delay(bh))
760 acceptable = (type == IOMAP_DELAY);
761 else if (buffer_dirty(bh) && buffer_mapped(bh))
762 acceptable = (type == IOMAP_NEW);
765 } while ((bh = bh->b_this_page) != head);
775 * Allocate & map buffers for page given the extent map. Write it out.
776 * except for the original page of a writepage, this is called on
777 * delalloc/unwritten pages only, for the original page it is possible
778 * that the page has no mapping at all.
786 xfs_ioend_t **ioendp,
787 struct writeback_control *wbc,
791 struct buffer_head *bh, *head;
792 xfs_off_t end_offset;
793 unsigned long p_offset;
795 int bbits = inode->i_blkbits;
797 int count = 0, done = 0, uptodate = 1;
798 xfs_off_t offset = page_offset(page);
800 if (page->index != tindex)
802 if (!trylock_page(page))
804 if (PageWriteback(page))
805 goto fail_unlock_page;
806 if (page->mapping != inode->i_mapping)
807 goto fail_unlock_page;
808 if (!xfs_is_delayed_page(page, (*ioendp)->io_type))
809 goto fail_unlock_page;
812 * page_dirty is initially a count of buffers on the page before
813 * EOF and is decremented as we move each into a cleanable state.
817 * End offset is the highest offset that this page should represent.
818 * If we are on the last page, (end_offset & (PAGE_CACHE_SIZE - 1))
819 * will evaluate non-zero and be less than PAGE_CACHE_SIZE and
820 * hence give us the correct page_dirty count. On any other page,
821 * it will be zero and in that case we need page_dirty to be the
822 * count of buffers on the page.
824 end_offset = min_t(unsigned long long,
825 (xfs_off_t)(page->index + 1) << PAGE_CACHE_SHIFT,
828 len = 1 << inode->i_blkbits;
829 p_offset = min_t(unsigned long, end_offset & (PAGE_CACHE_SIZE - 1),
831 p_offset = p_offset ? roundup(p_offset, len) : PAGE_CACHE_SIZE;
832 page_dirty = p_offset / len;
834 bh = head = page_buffers(page);
836 if (offset >= end_offset)
838 if (!buffer_uptodate(bh))
840 if (!(PageUptodate(page) || buffer_uptodate(bh))) {
845 if (buffer_unwritten(bh) || buffer_delay(bh)) {
846 if (buffer_unwritten(bh))
847 type = IOMAP_UNWRITTEN;
851 if (!xfs_iomap_valid(mp, offset)) {
856 ASSERT(!(mp->iomap_flags & IOMAP_HOLE));
857 ASSERT(!(mp->iomap_flags & IOMAP_DELAY));
859 xfs_map_at_offset(bh, offset, bbits, mp);
861 xfs_add_to_ioend(inode, bh, offset,
864 set_buffer_dirty(bh);
866 mark_buffer_dirty(bh);
872 if (buffer_mapped(bh) && all_bh && startio) {
874 xfs_add_to_ioend(inode, bh, offset,
882 } while (offset += len, (bh = bh->b_this_page) != head);
884 if (uptodate && bh == head)
885 SetPageUptodate(page);
889 struct backing_dev_info *bdi;
891 bdi = inode->i_mapping->backing_dev_info;
893 if (bdi_write_congested(bdi)) {
894 wbc->encountered_congestion = 1;
896 } else if (wbc->nr_to_write <= 0) {
900 xfs_start_page_writeback(page, !page_dirty, count);
911 * Convert & write out a cluster of pages in the same extent as defined
912 * by mp and following the start page.
919 xfs_ioend_t **ioendp,
920 struct writeback_control *wbc,
928 pagevec_init(&pvec, 0);
929 while (!done && tindex <= tlast) {
930 unsigned len = min_t(pgoff_t, PAGEVEC_SIZE, tlast - tindex + 1);
932 if (!pagevec_lookup(&pvec, inode->i_mapping, tindex, len))
935 for (i = 0; i < pagevec_count(&pvec); i++) {
936 done = xfs_convert_page(inode, pvec.pages[i], tindex++,
937 iomapp, ioendp, wbc, startio, all_bh);
942 pagevec_release(&pvec);
948 * Calling this without startio set means we are being asked to make a dirty
949 * page ready for freeing it's buffers. When called with startio set then
950 * we are coming from writepage.
952 * When called with startio set it is important that we write the WHOLE
954 * The bh->b_state's cannot know if any of the blocks or which block for
955 * that matter are dirty due to mmap writes, and therefore bh uptodate is
956 * only valid if the page itself isn't completely uptodate. Some layers
957 * may clear the page dirty flag prior to calling write page, under the
958 * assumption the entire page will be written out; by not writing out the
959 * whole page the page can be reused before all valid dirty data is
960 * written out. Note: in the case of a page that has been dirty'd by
961 * mapwrite and but partially setup by block_prepare_write the
962 * bh->b_states's will not agree and only ones setup by BPW/BCW will have
963 * valid state, thus the whole page must be written out thing.
967 xfs_page_state_convert(
970 struct writeback_control *wbc,
972 int unmapped) /* also implies page uptodate */
974 struct buffer_head *bh, *head;
976 xfs_ioend_t *ioend = NULL, *iohead = NULL;
978 unsigned long p_offset = 0;
980 __uint64_t end_offset;
981 pgoff_t end_index, last_index, tlast;
983 int flags, err, iomap_valid = 0, uptodate = 1;
984 int page_dirty, count = 0;
986 int all_bh = unmapped;
989 if (wbc->sync_mode == WB_SYNC_NONE && wbc->nonblocking)
990 trylock |= BMAPI_TRYLOCK;
993 /* Is this page beyond the end of the file? */
994 offset = i_size_read(inode);
995 end_index = offset >> PAGE_CACHE_SHIFT;
996 last_index = (offset - 1) >> PAGE_CACHE_SHIFT;
997 if (page->index >= end_index) {
998 if ((page->index >= end_index + 1) ||
999 !(i_size_read(inode) & (PAGE_CACHE_SIZE - 1))) {
1007 * page_dirty is initially a count of buffers on the page before
1008 * EOF and is decremented as we move each into a cleanable state.
1012 * End offset is the highest offset that this page should represent.
1013 * If we are on the last page, (end_offset & (PAGE_CACHE_SIZE - 1))
1014 * will evaluate non-zero and be less than PAGE_CACHE_SIZE and
1015 * hence give us the correct page_dirty count. On any other page,
1016 * it will be zero and in that case we need page_dirty to be the
1017 * count of buffers on the page.
1019 end_offset = min_t(unsigned long long,
1020 (xfs_off_t)(page->index + 1) << PAGE_CACHE_SHIFT, offset);
1021 len = 1 << inode->i_blkbits;
1022 p_offset = min_t(unsigned long, end_offset & (PAGE_CACHE_SIZE - 1),
1024 p_offset = p_offset ? roundup(p_offset, len) : PAGE_CACHE_SIZE;
1025 page_dirty = p_offset / len;
1027 bh = head = page_buffers(page);
1028 offset = page_offset(page);
1032 /* TODO: cleanup count and page_dirty */
1035 if (offset >= end_offset)
1037 if (!buffer_uptodate(bh))
1039 if (!(PageUptodate(page) || buffer_uptodate(bh)) && !startio) {
1041 * the iomap is actually still valid, but the ioend
1042 * isn't. shouldn't happen too often.
1049 iomap_valid = xfs_iomap_valid(&iomap, offset);
1052 * First case, map an unwritten extent and prepare for
1053 * extent state conversion transaction on completion.
1055 * Second case, allocate space for a delalloc buffer.
1056 * We can return EAGAIN here in the release page case.
1058 * Third case, an unmapped buffer was found, and we are
1059 * in a path where we need to write the whole page out.
1061 if (buffer_unwritten(bh) || buffer_delay(bh) ||
1062 ((buffer_uptodate(bh) || PageUptodate(page)) &&
1063 !buffer_mapped(bh) && (unmapped || startio))) {
1067 * Make sure we don't use a read-only iomap
1069 if (flags == BMAPI_READ)
1072 if (buffer_unwritten(bh)) {
1073 type = IOMAP_UNWRITTEN;
1074 flags = BMAPI_WRITE | BMAPI_IGNSTATE;
1075 } else if (buffer_delay(bh)) {
1077 flags = BMAPI_ALLOCATE | trylock;
1080 flags = BMAPI_WRITE | BMAPI_MMAP;
1085 * if we didn't have a valid mapping then we
1086 * need to ensure that we put the new mapping
1087 * in a new ioend structure. This needs to be
1088 * done to ensure that the ioends correctly
1089 * reflect the block mappings at io completion
1090 * for unwritten extent conversion.
1093 if (type == IOMAP_NEW) {
1094 size = xfs_probe_cluster(inode,
1100 err = xfs_map_blocks(inode, offset, size,
1104 iomap_valid = xfs_iomap_valid(&iomap, offset);
1107 xfs_map_at_offset(bh, offset,
1108 inode->i_blkbits, &iomap);
1110 xfs_add_to_ioend(inode, bh, offset,
1114 set_buffer_dirty(bh);
1116 mark_buffer_dirty(bh);
1121 } else if (buffer_uptodate(bh) && startio) {
1123 * we got here because the buffer is already mapped.
1124 * That means it must already have extents allocated
1125 * underneath it. Map the extent by reading it.
1127 if (!iomap_valid || flags != BMAPI_READ) {
1129 size = xfs_probe_cluster(inode, page, bh,
1131 err = xfs_map_blocks(inode, offset, size,
1135 iomap_valid = xfs_iomap_valid(&iomap, offset);
1139 * We set the type to IOMAP_NEW in case we are doing a
1140 * small write at EOF that is extending the file but
1141 * without needing an allocation. We need to update the
1142 * file size on I/O completion in this case so it is
1143 * the same case as having just allocated a new extent
1144 * that we are writing into for the first time.
1147 if (trylock_buffer(bh)) {
1148 ASSERT(buffer_mapped(bh));
1151 xfs_add_to_ioend(inode, bh, offset, type,
1152 &ioend, !iomap_valid);
1158 } else if ((buffer_uptodate(bh) || PageUptodate(page)) &&
1159 (unmapped || startio)) {
1166 } while (offset += len, ((bh = bh->b_this_page) != head));
1168 if (uptodate && bh == head)
1169 SetPageUptodate(page);
1172 xfs_start_page_writeback(page, 1, count);
1174 if (ioend && iomap_valid) {
1175 offset = (iomap.iomap_offset + iomap.iomap_bsize - 1) >>
1177 tlast = min_t(pgoff_t, offset, last_index);
1178 xfs_cluster_write(inode, page->index + 1, &iomap, &ioend,
1179 wbc, startio, all_bh, tlast);
1183 xfs_submit_ioend(iohead);
1189 xfs_cancel_ioend(iohead);
1192 * If it's delalloc and we have nowhere to put it,
1193 * throw it away, unless the lower layers told
1196 if (err != -EAGAIN) {
1198 block_invalidatepage(page, 0);
1199 ClearPageUptodate(page);
1205 * writepage: Called from one of two places:
1207 * 1. we are flushing a delalloc buffer head.
1209 * 2. we are writing out a dirty page. Typically the page dirty
1210 * state is cleared before we get here. In this case is it
1211 * conceivable we have no buffer heads.
1213 * For delalloc space on the page we need to allocate space and
1214 * flush it. For unmapped buffer heads on the page we should
1215 * allocate space if the page is uptodate. For any other dirty
1216 * buffer heads on the page we should flush them.
1218 * If we detect that a transaction would be required to flush
1219 * the page, we have to check the process flags first, if we
1220 * are already in a transaction or disk I/O during allocations
1221 * is off, we need to fail the writepage and redirty the page.
1227 struct writeback_control *wbc)
1231 int delalloc, unmapped, unwritten;
1232 struct inode *inode = page->mapping->host;
1234 xfs_page_trace(XFS_WRITEPAGE_ENTER, inode, page, 0);
1237 * We need a transaction if:
1238 * 1. There are delalloc buffers on the page
1239 * 2. The page is uptodate and we have unmapped buffers
1240 * 3. The page is uptodate and we have no buffers
1241 * 4. There are unwritten buffers on the page
1244 if (!page_has_buffers(page)) {
1248 xfs_count_page_state(page, &delalloc, &unmapped, &unwritten);
1249 if (!PageUptodate(page))
1251 need_trans = delalloc + unmapped + unwritten;
1255 * If we need a transaction and the process flags say
1256 * we are already in a transaction, or no IO is allowed
1257 * then mark the page dirty again and leave the page
1260 if (current_test_flags(PF_FSTRANS) && need_trans)
1264 * Delay hooking up buffer heads until we have
1265 * made our go/no-go decision.
1267 if (!page_has_buffers(page))
1268 create_empty_buffers(page, 1 << inode->i_blkbits, 0);
1272 * VM calculation for nr_to_write seems off. Bump it way
1273 * up, this gets simple streaming writes zippy again.
1274 * To be reviewed again after Jens' writeback changes.
1276 wbc->nr_to_write *= 4;
1279 * Convert delayed allocate, unwritten or unmapped space
1280 * to real space and flush out to disk.
1282 error = xfs_page_state_convert(inode, page, wbc, 1, unmapped);
1283 if (error == -EAGAIN)
1285 if (unlikely(error < 0))
1291 redirty_page_for_writepage(wbc, page);
1301 struct address_space *mapping,
1302 struct writeback_control *wbc)
1304 xfs_iflags_clear(XFS_I(mapping->host), XFS_ITRUNCATED);
1305 return generic_writepages(mapping, wbc);
1309 * Called to move a page into cleanable state - and from there
1310 * to be released. Possibly the page is already clean. We always
1311 * have buffer heads in this call.
1313 * Returns 0 if the page is ok to release, 1 otherwise.
1315 * Possible scenarios are:
1317 * 1. We are being called to release a page which has been written
1318 * to via regular I/O. buffer heads will be dirty and possibly
1319 * delalloc. If no delalloc buffer heads in this case then we
1320 * can just return zero.
1322 * 2. We are called to release a page which has been written via
1323 * mmap, all we need to do is ensure there is no delalloc
1324 * state in the buffer heads, if not we can let the caller
1325 * free them and we should come back later via writepage.
1332 struct inode *inode = page->mapping->host;
1333 int dirty, delalloc, unmapped, unwritten;
1334 struct writeback_control wbc = {
1335 .sync_mode = WB_SYNC_ALL,
1339 xfs_page_trace(XFS_RELEASEPAGE_ENTER, inode, page, 0);
1341 if (!page_has_buffers(page))
1344 xfs_count_page_state(page, &delalloc, &unmapped, &unwritten);
1345 if (!delalloc && !unwritten)
1348 if (!(gfp_mask & __GFP_FS))
1351 /* If we are already inside a transaction or the thread cannot
1352 * do I/O, we cannot release this page.
1354 if (current_test_flags(PF_FSTRANS))
1358 * Convert delalloc space to real space, do not flush the
1359 * data out to disk, that will be done by the caller.
1360 * Never need to allocate space here - we will always
1361 * come back to writepage in that case.
1363 dirty = xfs_page_state_convert(inode, page, &wbc, 0, 0);
1364 if (dirty == 0 && !unwritten)
1369 return try_to_free_buffers(page);
1374 struct inode *inode,
1376 struct buffer_head *bh_result,
1379 bmapi_flags_t flags)
1387 offset = (xfs_off_t)iblock << inode->i_blkbits;
1388 ASSERT(bh_result->b_size >= (1 << inode->i_blkbits));
1389 size = bh_result->b_size;
1391 if (!create && direct && offset >= i_size_read(inode))
1394 error = xfs_iomap(XFS_I(inode), offset, size,
1395 create ? flags : BMAPI_READ, &iomap, &niomap);
1401 if (iomap.iomap_bn != IOMAP_DADDR_NULL) {
1403 * For unwritten extents do not report a disk address on
1404 * the read case (treat as if we're reading into a hole).
1406 if (create || !(iomap.iomap_flags & IOMAP_UNWRITTEN)) {
1407 xfs_map_buffer(bh_result, &iomap, offset,
1410 if (create && (iomap.iomap_flags & IOMAP_UNWRITTEN)) {
1412 bh_result->b_private = inode;
1413 set_buffer_unwritten(bh_result);
1418 * If this is a realtime file, data may be on a different device.
1419 * to that pointed to from the buffer_head b_bdev currently.
1421 bh_result->b_bdev = iomap.iomap_target->bt_bdev;
1424 * If we previously allocated a block out beyond eof and we are now
1425 * coming back to use it then we will need to flag it as new even if it
1426 * has a disk address.
1428 * With sub-block writes into unwritten extents we also need to mark
1429 * the buffer as new so that the unwritten parts of the buffer gets
1433 ((!buffer_mapped(bh_result) && !buffer_uptodate(bh_result)) ||
1434 (offset >= i_size_read(inode)) ||
1435 (iomap.iomap_flags & (IOMAP_NEW|IOMAP_UNWRITTEN))))
1436 set_buffer_new(bh_result);
1438 if (iomap.iomap_flags & IOMAP_DELAY) {
1441 set_buffer_uptodate(bh_result);
1442 set_buffer_mapped(bh_result);
1443 set_buffer_delay(bh_result);
1447 if (direct || size > (1 << inode->i_blkbits)) {
1448 ASSERT(iomap.iomap_bsize - iomap.iomap_delta > 0);
1449 offset = min_t(xfs_off_t,
1450 iomap.iomap_bsize - iomap.iomap_delta, size);
1451 bh_result->b_size = (ssize_t)min_t(xfs_off_t, LONG_MAX, offset);
1459 struct inode *inode,
1461 struct buffer_head *bh_result,
1464 return __xfs_get_blocks(inode, iblock,
1465 bh_result, create, 0, BMAPI_WRITE);
1469 xfs_get_blocks_direct(
1470 struct inode *inode,
1472 struct buffer_head *bh_result,
1475 return __xfs_get_blocks(inode, iblock,
1476 bh_result, create, 1, BMAPI_WRITE|BMAPI_DIRECT);
1486 xfs_ioend_t *ioend = iocb->private;
1489 * Non-NULL private data means we need to issue a transaction to
1490 * convert a range from unwritten to written extents. This needs
1491 * to happen from process context but aio+dio I/O completion
1492 * happens from irq context so we need to defer it to a workqueue.
1493 * This is not necessary for synchronous direct I/O, but we do
1494 * it anyway to keep the code uniform and simpler.
1496 * Well, if only it were that simple. Because synchronous direct I/O
1497 * requires extent conversion to occur *before* we return to userspace,
1498 * we have to wait for extent conversion to complete. Look at the
1499 * iocb that has been passed to us to determine if this is AIO or
1500 * not. If it is synchronous, tell xfs_finish_ioend() to kick the
1501 * workqueue and wait for it to complete.
1503 * The core direct I/O code might be changed to always call the
1504 * completion handler in the future, in which case all this can
1507 ioend->io_offset = offset;
1508 ioend->io_size = size;
1509 if (ioend->io_type == IOMAP_READ) {
1510 xfs_finish_ioend(ioend, 0);
1511 } else if (private && size > 0) {
1512 xfs_finish_ioend(ioend, is_sync_kiocb(iocb));
1515 * A direct I/O write ioend starts it's life in unwritten
1516 * state in case they map an unwritten extent. This write
1517 * didn't map an unwritten extent so switch it's completion
1520 INIT_WORK(&ioend->io_work, xfs_end_bio_written);
1521 xfs_finish_ioend(ioend, 0);
1525 * blockdev_direct_IO can return an error even after the I/O
1526 * completion handler was called. Thus we need to protect
1527 * against double-freeing.
1529 iocb->private = NULL;
1536 const struct iovec *iov,
1538 unsigned long nr_segs)
1540 struct file *file = iocb->ki_filp;
1541 struct inode *inode = file->f_mapping->host;
1542 struct block_device *bdev;
1545 bdev = xfs_find_bdev_for_inode(XFS_I(inode));
1548 iocb->private = xfs_alloc_ioend(inode, IOMAP_UNWRITTEN);
1549 ret = blockdev_direct_IO_own_locking(rw, iocb, inode,
1550 bdev, iov, offset, nr_segs,
1551 xfs_get_blocks_direct,
1554 iocb->private = xfs_alloc_ioend(inode, IOMAP_READ);
1555 ret = blockdev_direct_IO_no_locking(rw, iocb, inode,
1556 bdev, iov, offset, nr_segs,
1557 xfs_get_blocks_direct,
1561 if (unlikely(ret != -EIOCBQUEUED && iocb->private))
1562 xfs_destroy_ioend(iocb->private);
1569 struct address_space *mapping,
1573 struct page **pagep,
1577 return block_write_begin(file, mapping, pos, len, flags, pagep, fsdata,
1583 struct address_space *mapping,
1586 struct inode *inode = (struct inode *)mapping->host;
1587 struct xfs_inode *ip = XFS_I(inode);
1589 xfs_itrace_entry(XFS_I(inode));
1590 xfs_ilock(ip, XFS_IOLOCK_SHARED);
1591 xfs_flush_pages(ip, (xfs_off_t)0, -1, 0, FI_REMAPF);
1592 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
1593 return generic_block_bmap(mapping, block, xfs_get_blocks);
1598 struct file *unused,
1601 return mpage_readpage(page, xfs_get_blocks);
1606 struct file *unused,
1607 struct address_space *mapping,
1608 struct list_head *pages,
1611 return mpage_readpages(mapping, pages, nr_pages, xfs_get_blocks);
1615 xfs_vm_invalidatepage(
1617 unsigned long offset)
1619 xfs_page_trace(XFS_INVALIDPAGE_ENTER,
1620 page->mapping->host, page, offset);
1621 block_invalidatepage(page, offset);
1624 const struct address_space_operations xfs_address_space_operations = {
1625 .readpage = xfs_vm_readpage,
1626 .readpages = xfs_vm_readpages,
1627 .writepage = xfs_vm_writepage,
1628 .writepages = xfs_vm_writepages,
1629 .sync_page = block_sync_page,
1630 .releasepage = xfs_vm_releasepage,
1631 .invalidatepage = xfs_vm_invalidatepage,
1632 .write_begin = xfs_vm_write_begin,
1633 .write_end = generic_write_end,
1634 .bmap = xfs_vm_bmap,
1635 .direct_IO = xfs_vm_direct_IO,
1636 .migratepage = buffer_migrate_page,
1637 .is_partially_uptodate = block_is_partially_uptodate,
1638 .error_remove_page = generic_error_remove_page,