2 * Copyright (c) 2000-2005 Silicon Graphics, Inc.
5 * This program is free software; you can redistribute it and/or
6 * modify it under the terms of the GNU General Public License as
7 * published by the Free Software Foundation.
9 * This program is distributed in the hope that it would be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write the Free Software Foundation,
16 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
24 #include "xfs_trans.h"
25 #include "xfs_mount.h"
26 #include "xfs_bmap_btree.h"
27 #include "xfs_dinode.h"
28 #include "xfs_inode.h"
29 #include "xfs_alloc.h"
30 #include "xfs_error.h"
32 #include "xfs_iomap.h"
33 #include "xfs_vnodeops.h"
34 #include "xfs_trace.h"
36 #include <linux/gfp.h>
37 #include <linux/mpage.h>
38 #include <linux/pagevec.h>
39 #include <linux/writeback.h>
43 * Prime number of hash buckets since address is used as the key.
46 #define to_ioend_wq(v) (&xfs_ioend_wq[((unsigned long)v) % NVSYNC])
47 static wait_queue_head_t xfs_ioend_wq[NVSYNC];
54 for (i = 0; i < NVSYNC; i++)
55 init_waitqueue_head(&xfs_ioend_wq[i]);
62 wait_queue_head_t *wq = to_ioend_wq(ip);
64 wait_event(*wq, (atomic_read(&ip->i_iocount) == 0));
71 if (atomic_dec_and_test(&ip->i_iocount))
72 wake_up(to_ioend_wq(ip));
81 struct buffer_head *bh, *head;
83 *delalloc = *unwritten = 0;
85 bh = head = page_buffers(page);
87 if (buffer_unwritten(bh))
89 else if (buffer_delay(bh))
91 } while ((bh = bh->b_this_page) != head);
94 STATIC struct block_device *
95 xfs_find_bdev_for_inode(
98 struct xfs_inode *ip = XFS_I(inode);
99 struct xfs_mount *mp = ip->i_mount;
101 if (XFS_IS_REALTIME_INODE(ip))
102 return mp->m_rtdev_targp->bt_bdev;
104 return mp->m_ddev_targp->bt_bdev;
108 * We're now finished for good with this ioend structure.
109 * Update the page state via the associated buffer_heads,
110 * release holds on the inode and bio, and finally free
111 * up memory. Do not use the ioend after this.
117 struct buffer_head *bh, *next;
118 struct xfs_inode *ip = XFS_I(ioend->io_inode);
120 for (bh = ioend->io_buffer_head; bh; bh = next) {
121 next = bh->b_private;
122 bh->b_end_io(bh, !ioend->io_error);
126 * Volume managers supporting multiple paths can send back ENODEV
127 * when the final path disappears. In this case continuing to fill
128 * the page cache with dirty data which cannot be written out is
129 * evil, so prevent that.
131 if (unlikely(ioend->io_error == -ENODEV)) {
132 xfs_do_force_shutdown(ip->i_mount, SHUTDOWN_DEVICE_REQ,
137 mempool_free(ioend, xfs_ioend_pool);
141 * If the end of the current ioend is beyond the current EOF,
142 * return the new EOF value, otherwise zero.
148 xfs_inode_t *ip = XFS_I(ioend->io_inode);
152 bsize = ioend->io_offset + ioend->io_size;
153 isize = MAX(ip->i_size, ip->i_new_size);
154 isize = MIN(isize, bsize);
155 return isize > ip->i_d.di_size ? isize : 0;
159 * Update on-disk file size now that data has been written to disk. The
160 * current in-memory file size is i_size. If a write is beyond eof i_new_size
161 * will be the intended file size until i_size is updated. If this write does
162 * not extend all the way to the valid file size then restrict this update to
163 * the end of the write.
165 * This function does not block as blocking on the inode lock in IO completion
166 * can lead to IO completion order dependency deadlocks.. If it can't get the
167 * inode ilock it will return EAGAIN. Callers must handle this.
173 xfs_inode_t *ip = XFS_I(ioend->io_inode);
176 if (unlikely(ioend->io_error))
179 if (!xfs_ilock_nowait(ip, XFS_ILOCK_EXCL))
182 isize = xfs_ioend_new_eof(ioend);
184 ip->i_d.di_size = isize;
185 xfs_mark_inode_dirty(ip);
188 xfs_iunlock(ip, XFS_ILOCK_EXCL);
193 * Schedule IO completion handling on the final put of an ioend.
197 struct xfs_ioend *ioend)
199 if (atomic_dec_and_test(&ioend->io_remaining)) {
200 if (ioend->io_type == IO_UNWRITTEN)
201 queue_work(xfsconvertd_workqueue, &ioend->io_work);
203 queue_work(xfsdatad_workqueue, &ioend->io_work);
208 * IO write completion.
212 struct work_struct *work)
214 xfs_ioend_t *ioend = container_of(work, xfs_ioend_t, io_work);
215 struct xfs_inode *ip = XFS_I(ioend->io_inode);
219 * For unwritten extents we need to issue transactions to convert a
220 * range to normal written extens after the data I/O has finished.
222 if (ioend->io_type == IO_UNWRITTEN &&
223 likely(!ioend->io_error && !XFS_FORCED_SHUTDOWN(ip->i_mount))) {
225 error = xfs_iomap_write_unwritten(ip, ioend->io_offset,
228 ioend->io_error = error;
232 * We might have to update the on-disk file size after extending
235 error = xfs_setfilesize(ioend);
236 ASSERT(!error || error == EAGAIN);
239 * If we didn't complete processing of the ioend, requeue it to the
240 * tail of the workqueue for another attempt later. Otherwise destroy
243 if (error == EAGAIN) {
244 atomic_inc(&ioend->io_remaining);
245 xfs_finish_ioend(ioend);
246 /* ensure we don't spin on blocked ioends */
250 aio_complete(ioend->io_iocb, ioend->io_result, 0);
251 xfs_destroy_ioend(ioend);
256 * Call IO completion handling in caller context on the final put of an ioend.
259 xfs_finish_ioend_sync(
260 struct xfs_ioend *ioend)
262 if (atomic_dec_and_test(&ioend->io_remaining))
263 xfs_end_io(&ioend->io_work);
267 * Allocate and initialise an IO completion structure.
268 * We need to track unwritten extent write completion here initially.
269 * We'll need to extend this for updating the ondisk inode size later
279 ioend = mempool_alloc(xfs_ioend_pool, GFP_NOFS);
282 * Set the count to 1 initially, which will prevent an I/O
283 * completion callback from happening before we have started
284 * all the I/O from calling the completion routine too early.
286 atomic_set(&ioend->io_remaining, 1);
288 ioend->io_list = NULL;
289 ioend->io_type = type;
290 ioend->io_inode = inode;
291 ioend->io_buffer_head = NULL;
292 ioend->io_buffer_tail = NULL;
293 atomic_inc(&XFS_I(ioend->io_inode)->i_iocount);
294 ioend->io_offset = 0;
296 ioend->io_iocb = NULL;
297 ioend->io_result = 0;
299 INIT_WORK(&ioend->io_work, xfs_end_io);
308 struct xfs_bmbt_irec *imap,
312 struct xfs_inode *ip = XFS_I(inode);
313 struct xfs_mount *mp = ip->i_mount;
314 xfs_fileoff_t offset_fsb, end_fsb;
316 int bmapi_flags = XFS_BMAPI_ENTIRE;
319 if (XFS_FORCED_SHUTDOWN(mp))
320 return -XFS_ERROR(EIO);
322 if (type == IO_UNWRITTEN)
323 bmapi_flags |= XFS_BMAPI_IGSTATE;
325 if (!xfs_ilock_nowait(ip, XFS_ILOCK_SHARED)) {
327 return -XFS_ERROR(EAGAIN);
328 xfs_ilock(ip, XFS_ILOCK_SHARED);
331 ASSERT(ip->i_d.di_format != XFS_DINODE_FMT_BTREE ||
332 (ip->i_df.if_flags & XFS_IFEXTENTS));
333 ASSERT(offset <= mp->m_maxioffset);
335 if (offset + count > mp->m_maxioffset)
336 count = mp->m_maxioffset - offset;
337 end_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)offset + count);
338 offset_fsb = XFS_B_TO_FSBT(mp, offset);
339 error = xfs_bmapi(NULL, ip, offset_fsb, end_fsb - offset_fsb,
340 bmapi_flags, NULL, 0, imap, &nimaps, NULL);
341 xfs_iunlock(ip, XFS_ILOCK_SHARED);
344 return -XFS_ERROR(error);
346 if (type == IO_DELALLOC &&
347 (!nimaps || isnullstartblock(imap->br_startblock))) {
348 error = xfs_iomap_write_allocate(ip, offset, count, imap);
350 trace_xfs_map_blocks_alloc(ip, offset, count, type, imap);
351 return -XFS_ERROR(error);
355 if (type == IO_UNWRITTEN) {
357 ASSERT(imap->br_startblock != HOLESTARTBLOCK);
358 ASSERT(imap->br_startblock != DELAYSTARTBLOCK);
362 trace_xfs_map_blocks_found(ip, offset, count, type, imap);
369 struct xfs_bmbt_irec *imap,
372 offset >>= inode->i_blkbits;
374 return offset >= imap->br_startoff &&
375 offset < imap->br_startoff + imap->br_blockcount;
379 * BIO completion handler for buffered IO.
386 xfs_ioend_t *ioend = bio->bi_private;
388 ASSERT(atomic_read(&bio->bi_cnt) >= 1);
389 ioend->io_error = test_bit(BIO_UPTODATE, &bio->bi_flags) ? 0 : error;
391 /* Toss bio and pass work off to an xfsdatad thread */
392 bio->bi_private = NULL;
393 bio->bi_end_io = NULL;
396 xfs_finish_ioend(ioend);
400 xfs_submit_ioend_bio(
401 struct writeback_control *wbc,
405 atomic_inc(&ioend->io_remaining);
406 bio->bi_private = ioend;
407 bio->bi_end_io = xfs_end_bio;
410 * If the I/O is beyond EOF we mark the inode dirty immediately
411 * but don't update the inode size until I/O completion.
413 if (xfs_ioend_new_eof(ioend))
414 xfs_mark_inode_dirty(XFS_I(ioend->io_inode));
416 submit_bio(wbc->sync_mode == WB_SYNC_ALL ?
417 WRITE_SYNC_PLUG : WRITE, bio);
422 struct buffer_head *bh)
424 int nvecs = bio_get_nr_vecs(bh->b_bdev);
425 struct bio *bio = bio_alloc(GFP_NOIO, nvecs);
427 ASSERT(bio->bi_private == NULL);
428 bio->bi_sector = bh->b_blocknr * (bh->b_size >> 9);
429 bio->bi_bdev = bh->b_bdev;
434 xfs_start_buffer_writeback(
435 struct buffer_head *bh)
437 ASSERT(buffer_mapped(bh));
438 ASSERT(buffer_locked(bh));
439 ASSERT(!buffer_delay(bh));
440 ASSERT(!buffer_unwritten(bh));
442 mark_buffer_async_write(bh);
443 set_buffer_uptodate(bh);
444 clear_buffer_dirty(bh);
448 xfs_start_page_writeback(
453 ASSERT(PageLocked(page));
454 ASSERT(!PageWriteback(page));
456 clear_page_dirty_for_io(page);
457 set_page_writeback(page);
459 /* If no buffers on the page are to be written, finish it here */
461 end_page_writeback(page);
464 static inline int bio_add_buffer(struct bio *bio, struct buffer_head *bh)
466 return bio_add_page(bio, bh->b_page, bh->b_size, bh_offset(bh));
470 * Submit all of the bios for all of the ioends we have saved up, covering the
471 * initial writepage page and also any probed pages.
473 * Because we may have multiple ioends spanning a page, we need to start
474 * writeback on all the buffers before we submit them for I/O. If we mark the
475 * buffers as we got, then we can end up with a page that only has buffers
476 * marked async write and I/O complete on can occur before we mark the other
477 * buffers async write.
479 * The end result of this is that we trip a bug in end_page_writeback() because
480 * we call it twice for the one page as the code in end_buffer_async_write()
481 * assumes that all buffers on the page are started at the same time.
483 * The fix is two passes across the ioend list - one to start writeback on the
484 * buffer_heads, and then submit them for I/O on the second pass.
488 struct writeback_control *wbc,
491 xfs_ioend_t *head = ioend;
493 struct buffer_head *bh;
495 sector_t lastblock = 0;
497 /* Pass 1 - start writeback */
499 next = ioend->io_list;
500 for (bh = ioend->io_buffer_head; bh; bh = bh->b_private)
501 xfs_start_buffer_writeback(bh);
502 } while ((ioend = next) != NULL);
504 /* Pass 2 - submit I/O */
507 next = ioend->io_list;
510 for (bh = ioend->io_buffer_head; bh; bh = bh->b_private) {
514 bio = xfs_alloc_ioend_bio(bh);
515 } else if (bh->b_blocknr != lastblock + 1) {
516 xfs_submit_ioend_bio(wbc, ioend, bio);
520 if (bio_add_buffer(bio, bh) != bh->b_size) {
521 xfs_submit_ioend_bio(wbc, ioend, bio);
525 lastblock = bh->b_blocknr;
528 xfs_submit_ioend_bio(wbc, ioend, bio);
529 xfs_finish_ioend(ioend);
530 } while ((ioend = next) != NULL);
534 * Cancel submission of all buffer_heads so far in this endio.
535 * Toss the endio too. Only ever called for the initial page
536 * in a writepage request, so only ever one page.
543 struct buffer_head *bh, *next_bh;
546 next = ioend->io_list;
547 bh = ioend->io_buffer_head;
549 next_bh = bh->b_private;
550 clear_buffer_async_write(bh);
552 } while ((bh = next_bh) != NULL);
554 xfs_ioend_wake(XFS_I(ioend->io_inode));
555 mempool_free(ioend, xfs_ioend_pool);
556 } while ((ioend = next) != NULL);
560 * Test to see if we've been building up a completion structure for
561 * earlier buffers -- if so, we try to append to this ioend if we
562 * can, otherwise we finish off any current ioend and start another.
563 * Return true if we've finished the given ioend.
568 struct buffer_head *bh,
571 xfs_ioend_t **result,
574 xfs_ioend_t *ioend = *result;
576 if (!ioend || need_ioend || type != ioend->io_type) {
577 xfs_ioend_t *previous = *result;
579 ioend = xfs_alloc_ioend(inode, type);
580 ioend->io_offset = offset;
581 ioend->io_buffer_head = bh;
582 ioend->io_buffer_tail = bh;
584 previous->io_list = ioend;
587 ioend->io_buffer_tail->b_private = bh;
588 ioend->io_buffer_tail = bh;
591 bh->b_private = NULL;
592 ioend->io_size += bh->b_size;
598 struct buffer_head *bh,
599 struct xfs_bmbt_irec *imap,
603 struct xfs_mount *m = XFS_I(inode)->i_mount;
604 xfs_off_t iomap_offset = XFS_FSB_TO_B(m, imap->br_startoff);
605 xfs_daddr_t iomap_bn = xfs_fsb_to_db(XFS_I(inode), imap->br_startblock);
607 ASSERT(imap->br_startblock != HOLESTARTBLOCK);
608 ASSERT(imap->br_startblock != DELAYSTARTBLOCK);
610 bn = (iomap_bn >> (inode->i_blkbits - BBSHIFT)) +
611 ((offset - iomap_offset) >> inode->i_blkbits);
613 ASSERT(bn || XFS_IS_REALTIME_INODE(XFS_I(inode)));
616 set_buffer_mapped(bh);
622 struct buffer_head *bh,
623 struct xfs_bmbt_irec *imap,
626 ASSERT(imap->br_startblock != HOLESTARTBLOCK);
627 ASSERT(imap->br_startblock != DELAYSTARTBLOCK);
630 xfs_map_buffer(inode, bh, imap, offset);
631 bh->b_bdev = xfs_find_bdev_for_inode(inode);
632 set_buffer_mapped(bh);
633 clear_buffer_delay(bh);
634 clear_buffer_unwritten(bh);
638 * Look for a page at index that is suitable for clustering.
643 unsigned int pg_offset)
645 struct buffer_head *bh, *head;
648 if (PageWriteback(page))
650 if (!PageDirty(page))
654 if (!page_has_buffers(page))
657 bh = head = page_buffers(page);
659 if (!buffer_uptodate(bh))
661 if (!buffer_mapped(bh))
664 if (ret >= pg_offset)
666 } while ((bh = bh->b_this_page) != head);
674 struct page *startpage,
675 struct buffer_head *bh,
676 struct buffer_head *head)
679 pgoff_t tindex, tlast, tloff;
683 /* First sum forwards in this page */
685 if (!buffer_uptodate(bh) || !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);
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 == IO_UNWRITTEN);
759 else if (buffer_delay(bh))
760 acceptable = (type == IO_DELALLOC);
761 else if (buffer_dirty(bh) && buffer_mapped(bh))
762 acceptable = (type == IO_OVERWRITE);
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.
785 struct xfs_bmbt_irec *imap,
786 xfs_ioend_t **ioendp,
787 struct writeback_control *wbc,
790 struct buffer_head *bh, *head;
791 xfs_off_t end_offset;
792 unsigned long p_offset;
795 int count = 0, done = 0, uptodate = 1;
796 xfs_off_t offset = page_offset(page);
798 if (page->index != tindex)
800 if (!trylock_page(page))
802 if (PageWriteback(page))
803 goto fail_unlock_page;
804 if (page->mapping != inode->i_mapping)
805 goto fail_unlock_page;
806 if (!xfs_is_delayed_page(page, (*ioendp)->io_type))
807 goto fail_unlock_page;
810 * page_dirty is initially a count of buffers on the page before
811 * EOF and is decremented as we move each into a cleanable state.
815 * End offset is the highest offset that this page should represent.
816 * If we are on the last page, (end_offset & (PAGE_CACHE_SIZE - 1))
817 * will evaluate non-zero and be less than PAGE_CACHE_SIZE and
818 * hence give us the correct page_dirty count. On any other page,
819 * it will be zero and in that case we need page_dirty to be the
820 * count of buffers on the page.
822 end_offset = min_t(unsigned long long,
823 (xfs_off_t)(page->index + 1) << PAGE_CACHE_SHIFT,
826 len = 1 << inode->i_blkbits;
827 p_offset = min_t(unsigned long, end_offset & (PAGE_CACHE_SIZE - 1),
829 p_offset = p_offset ? roundup(p_offset, len) : PAGE_CACHE_SIZE;
830 page_dirty = p_offset / len;
832 bh = head = page_buffers(page);
834 if (offset >= end_offset)
836 if (!buffer_uptodate(bh))
838 if (!(PageUptodate(page) || buffer_uptodate(bh))) {
843 if (buffer_unwritten(bh) || buffer_delay(bh)) {
844 if (buffer_unwritten(bh))
849 if (!xfs_imap_valid(inode, imap, offset)) {
854 ASSERT(imap->br_startblock != HOLESTARTBLOCK);
855 ASSERT(imap->br_startblock != DELAYSTARTBLOCK);
857 xfs_map_at_offset(inode, bh, imap, offset);
858 xfs_add_to_ioend(inode, bh, offset, type,
865 if (buffer_mapped(bh) && all_bh) {
867 xfs_add_to_ioend(inode, bh, offset,
875 } while (offset += len, (bh = bh->b_this_page) != head);
877 if (uptodate && bh == head)
878 SetPageUptodate(page);
881 if (--wbc->nr_to_write <= 0 &&
882 wbc->sync_mode == WB_SYNC_NONE)
885 xfs_start_page_writeback(page, !page_dirty, count);
895 * Convert & write out a cluster of pages in the same extent as defined
896 * by mp and following the start page.
902 struct xfs_bmbt_irec *imap,
903 xfs_ioend_t **ioendp,
904 struct writeback_control *wbc,
911 pagevec_init(&pvec, 0);
912 while (!done && tindex <= tlast) {
913 unsigned len = min_t(pgoff_t, PAGEVEC_SIZE, tlast - tindex + 1);
915 if (!pagevec_lookup(&pvec, inode->i_mapping, tindex, len))
918 for (i = 0; i < pagevec_count(&pvec); i++) {
919 done = xfs_convert_page(inode, pvec.pages[i], tindex++,
920 imap, ioendp, wbc, all_bh);
925 pagevec_release(&pvec);
931 xfs_vm_invalidatepage(
933 unsigned long offset)
935 trace_xfs_invalidatepage(page->mapping->host, page, offset);
936 block_invalidatepage(page, offset);
940 * If the page has delalloc buffers on it, we need to punch them out before we
941 * invalidate the page. If we don't, we leave a stale delalloc mapping on the
942 * inode that can trip a BUG() in xfs_get_blocks() later on if a direct IO read
943 * is done on that same region - the delalloc extent is returned when none is
944 * supposed to be there.
946 * We prevent this by truncating away the delalloc regions on the page before
947 * invalidating it. Because they are delalloc, we can do this without needing a
948 * transaction. Indeed - if we get ENOSPC errors, we have to be able to do this
949 * truncation without a transaction as there is no space left for block
950 * reservation (typically why we see a ENOSPC in writeback).
952 * This is not a performance critical path, so for now just do the punching a
953 * buffer head at a time.
956 xfs_aops_discard_page(
959 struct inode *inode = page->mapping->host;
960 struct xfs_inode *ip = XFS_I(inode);
961 struct buffer_head *bh, *head;
962 loff_t offset = page_offset(page);
964 if (!xfs_is_delayed_page(page, IO_DELALLOC))
967 if (XFS_FORCED_SHUTDOWN(ip->i_mount))
970 xfs_fs_cmn_err(CE_ALERT, ip->i_mount,
971 "page discard on page %p, inode 0x%llx, offset %llu.",
972 page, ip->i_ino, offset);
974 xfs_ilock(ip, XFS_ILOCK_EXCL);
975 bh = head = page_buffers(page);
978 xfs_fileoff_t start_fsb;
980 if (!buffer_delay(bh))
983 start_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
984 error = xfs_bmap_punch_delalloc_range(ip, start_fsb, 1);
986 /* something screwed, just bail */
987 if (!XFS_FORCED_SHUTDOWN(ip->i_mount)) {
988 xfs_fs_cmn_err(CE_ALERT, ip->i_mount,
989 "page discard unable to remove delalloc mapping.");
994 offset += 1 << inode->i_blkbits;
996 } while ((bh = bh->b_this_page) != head);
998 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1000 xfs_vm_invalidatepage(page, 0);
1005 * Write out a dirty page.
1007 * For delalloc space on the page we need to allocate space and flush it.
1008 * For unwritten space on the page we need to start the conversion to
1009 * regular allocated space.
1010 * For any other dirty buffer heads on the page we should flush them.
1012 * If we detect that a transaction would be required to flush the page, we
1013 * have to check the process flags first, if we are already in a transaction
1014 * or disk I/O during allocations is off, we need to fail the writepage and
1020 struct writeback_control *wbc)
1022 struct inode *inode = page->mapping->host;
1023 int delalloc, unwritten;
1024 struct buffer_head *bh, *head;
1025 struct xfs_bmbt_irec imap;
1026 xfs_ioend_t *ioend = NULL, *iohead = NULL;
1029 __uint64_t end_offset;
1030 pgoff_t end_index, last_index;
1032 int err, imap_valid = 0, uptodate = 1;
1035 int nonblocking = 0;
1037 trace_xfs_writepage(inode, page, 0);
1039 ASSERT(page_has_buffers(page));
1042 * Refuse to write the page out if we are called from reclaim context.
1044 * This avoids stack overflows when called from deeply used stacks in
1045 * random callers for direct reclaim or memcg reclaim. We explicitly
1046 * allow reclaim from kswapd as the stack usage there is relatively low.
1048 * This should really be done by the core VM, but until that happens
1049 * filesystems like XFS, btrfs and ext4 have to take care of this
1052 if ((current->flags & (PF_MEMALLOC|PF_KSWAPD)) == PF_MEMALLOC)
1056 * We need a transaction if there are delalloc or unwritten buffers
1059 * If we need a transaction and the process flags say we are already
1060 * in a transaction, or no IO is allowed then mark the page dirty
1061 * again and leave the page as is.
1063 xfs_count_page_state(page, &delalloc, &unwritten);
1064 if ((current->flags & PF_FSTRANS) && (delalloc || unwritten))
1067 /* Is this page beyond the end of the file? */
1068 offset = i_size_read(inode);
1069 end_index = offset >> PAGE_CACHE_SHIFT;
1070 last_index = (offset - 1) >> PAGE_CACHE_SHIFT;
1071 if (page->index >= end_index) {
1072 if ((page->index >= end_index + 1) ||
1073 !(i_size_read(inode) & (PAGE_CACHE_SIZE - 1))) {
1079 end_offset = min_t(unsigned long long,
1080 (xfs_off_t)(page->index + 1) << PAGE_CACHE_SHIFT,
1082 len = 1 << inode->i_blkbits;
1084 bh = head = page_buffers(page);
1085 offset = page_offset(page);
1086 type = IO_OVERWRITE;
1088 if (wbc->sync_mode == WB_SYNC_NONE && wbc->nonblocking)
1094 if (offset >= end_offset)
1096 if (!buffer_uptodate(bh))
1100 * set_page_dirty dirties all buffers in a page, independent
1101 * of their state. The dirty state however is entirely
1102 * meaningless for holes (!mapped && uptodate), so skip
1103 * buffers covering holes here.
1105 if (!buffer_mapped(bh) && buffer_uptodate(bh)) {
1111 imap_valid = xfs_imap_valid(inode, &imap, offset);
1113 if (buffer_unwritten(bh) || buffer_delay(bh)) {
1114 if (buffer_unwritten(bh)) {
1115 if (type != IO_UNWRITTEN) {
1116 type = IO_UNWRITTEN;
1119 } else if (buffer_delay(bh)) {
1120 if (type != IO_DELALLOC) {
1128 * If we didn't have a valid mapping then we
1129 * need to ensure that we put the new mapping
1130 * in a new ioend structure. This needs to be
1131 * done to ensure that the ioends correctly
1132 * reflect the block mappings at io completion
1133 * for unwritten extent conversion.
1136 err = xfs_map_blocks(inode, offset, len, &imap,
1140 imap_valid = xfs_imap_valid(inode, &imap,
1144 xfs_map_at_offset(inode, bh, &imap, offset);
1145 xfs_add_to_ioend(inode, bh, offset, type,
1149 } else if (buffer_uptodate(bh)) {
1151 * we got here because the buffer is already mapped.
1152 * That means it must already have extents allocated
1153 * underneath it. Map the extent by reading it.
1155 if (type != IO_OVERWRITE) {
1156 type = IO_OVERWRITE;
1161 size = xfs_probe_cluster(inode, page, bh, head);
1162 err = xfs_map_blocks(inode, offset, size,
1163 &imap, type, nonblocking);
1166 imap_valid = xfs_imap_valid(inode, &imap,
1173 xfs_add_to_ioend(inode, bh, offset, type,
1177 } else if (PageUptodate(page)) {
1178 ASSERT(buffer_mapped(bh));
1185 } while (offset += len, ((bh = bh->b_this_page) != head));
1187 if (uptodate && bh == head)
1188 SetPageUptodate(page);
1190 xfs_start_page_writeback(page, 1, count);
1192 if (ioend && imap_valid) {
1193 xfs_off_t end_index;
1195 end_index = imap.br_startoff + imap.br_blockcount;
1198 end_index <<= inode->i_blkbits;
1201 end_index = (end_index - 1) >> PAGE_CACHE_SHIFT;
1203 /* check against file size */
1204 if (end_index > last_index)
1205 end_index = last_index;
1207 xfs_cluster_write(inode, page->index + 1, &imap, &ioend,
1208 wbc, all_bh, end_index);
1212 xfs_submit_ioend(wbc, iohead);
1218 xfs_cancel_ioend(iohead);
1223 xfs_aops_discard_page(page);
1224 ClearPageUptodate(page);
1229 redirty_page_for_writepage(wbc, page);
1236 struct address_space *mapping,
1237 struct writeback_control *wbc)
1239 xfs_iflags_clear(XFS_I(mapping->host), XFS_ITRUNCATED);
1240 return generic_writepages(mapping, wbc);
1244 * Called to move a page into cleanable state - and from there
1245 * to be released. The page should already be clean. We always
1246 * have buffer heads in this call.
1248 * Returns 1 if the page is ok to release, 0 otherwise.
1255 int delalloc, unwritten;
1257 trace_xfs_releasepage(page->mapping->host, page, 0);
1259 xfs_count_page_state(page, &delalloc, &unwritten);
1261 if (WARN_ON(delalloc))
1263 if (WARN_ON(unwritten))
1266 return try_to_free_buffers(page);
1271 struct inode *inode,
1273 struct buffer_head *bh_result,
1277 struct xfs_inode *ip = XFS_I(inode);
1278 struct xfs_mount *mp = ip->i_mount;
1279 xfs_fileoff_t offset_fsb, end_fsb;
1282 struct xfs_bmbt_irec imap;
1288 if (XFS_FORCED_SHUTDOWN(mp))
1289 return -XFS_ERROR(EIO);
1291 offset = (xfs_off_t)iblock << inode->i_blkbits;
1292 ASSERT(bh_result->b_size >= (1 << inode->i_blkbits));
1293 size = bh_result->b_size;
1295 if (!create && direct && offset >= i_size_read(inode))
1299 lockmode = XFS_ILOCK_EXCL;
1300 xfs_ilock(ip, lockmode);
1302 lockmode = xfs_ilock_map_shared(ip);
1305 ASSERT(offset <= mp->m_maxioffset);
1306 if (offset + size > mp->m_maxioffset)
1307 size = mp->m_maxioffset - offset;
1308 end_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)offset + size);
1309 offset_fsb = XFS_B_TO_FSBT(mp, offset);
1311 error = xfs_bmapi(NULL, ip, offset_fsb, end_fsb - offset_fsb,
1312 XFS_BMAPI_ENTIRE, NULL, 0, &imap, &nimaps, NULL);
1318 (imap.br_startblock == HOLESTARTBLOCK ||
1319 imap.br_startblock == DELAYSTARTBLOCK))) {
1321 error = xfs_iomap_write_direct(ip, offset, size,
1324 error = xfs_iomap_write_delay(ip, offset, size, &imap);
1329 trace_xfs_get_blocks_alloc(ip, offset, size, 0, &imap);
1330 } else if (nimaps) {
1331 trace_xfs_get_blocks_found(ip, offset, size, 0, &imap);
1333 trace_xfs_get_blocks_notfound(ip, offset, size);
1336 xfs_iunlock(ip, lockmode);
1338 if (imap.br_startblock != HOLESTARTBLOCK &&
1339 imap.br_startblock != DELAYSTARTBLOCK) {
1341 * For unwritten extents do not report a disk address on
1342 * the read case (treat as if we're reading into a hole).
1344 if (create || !ISUNWRITTEN(&imap))
1345 xfs_map_buffer(inode, bh_result, &imap, offset);
1346 if (create && ISUNWRITTEN(&imap)) {
1348 bh_result->b_private = inode;
1349 set_buffer_unwritten(bh_result);
1354 * If this is a realtime file, data may be on a different device.
1355 * to that pointed to from the buffer_head b_bdev currently.
1357 bh_result->b_bdev = xfs_find_bdev_for_inode(inode);
1360 * If we previously allocated a block out beyond eof and we are now
1361 * coming back to use it then we will need to flag it as new even if it
1362 * has a disk address.
1364 * With sub-block writes into unwritten extents we also need to mark
1365 * the buffer as new so that the unwritten parts of the buffer gets
1369 ((!buffer_mapped(bh_result) && !buffer_uptodate(bh_result)) ||
1370 (offset >= i_size_read(inode)) ||
1371 (new || ISUNWRITTEN(&imap))))
1372 set_buffer_new(bh_result);
1374 if (imap.br_startblock == DELAYSTARTBLOCK) {
1377 set_buffer_uptodate(bh_result);
1378 set_buffer_mapped(bh_result);
1379 set_buffer_delay(bh_result);
1384 * If this is O_DIRECT or the mpage code calling tell them how large
1385 * the mapping is, so that we can avoid repeated get_blocks calls.
1387 if (direct || size > (1 << inode->i_blkbits)) {
1388 xfs_off_t mapping_size;
1390 mapping_size = imap.br_startoff + imap.br_blockcount - iblock;
1391 mapping_size <<= inode->i_blkbits;
1393 ASSERT(mapping_size > 0);
1394 if (mapping_size > size)
1395 mapping_size = size;
1396 if (mapping_size > LONG_MAX)
1397 mapping_size = LONG_MAX;
1399 bh_result->b_size = mapping_size;
1405 xfs_iunlock(ip, lockmode);
1411 struct inode *inode,
1413 struct buffer_head *bh_result,
1416 return __xfs_get_blocks(inode, iblock, bh_result, create, 0);
1420 xfs_get_blocks_direct(
1421 struct inode *inode,
1423 struct buffer_head *bh_result,
1426 return __xfs_get_blocks(inode, iblock, bh_result, create, 1);
1430 * Complete a direct I/O write request.
1432 * If the private argument is non-NULL __xfs_get_blocks signals us that we
1433 * need to issue a transaction to convert the range from unwritten to written
1434 * extents. In case this is regular synchronous I/O we just call xfs_end_io
1435 * to do this and we are done. But in case this was a successfull AIO
1436 * request this handler is called from interrupt context, from which we
1437 * can't start transactions. In that case offload the I/O completion to
1438 * the workqueues we also use for buffered I/O completion.
1441 xfs_end_io_direct_write(
1449 struct xfs_ioend *ioend = iocb->private;
1452 * blockdev_direct_IO can return an error even after the I/O
1453 * completion handler was called. Thus we need to protect
1454 * against double-freeing.
1456 iocb->private = NULL;
1458 ioend->io_offset = offset;
1459 ioend->io_size = size;
1460 if (private && size > 0)
1461 ioend->io_type = IO_UNWRITTEN;
1465 * If we are converting an unwritten extent we need to delay
1466 * the AIO completion until after the unwrittent extent
1467 * conversion has completed, otherwise do it ASAP.
1469 if (ioend->io_type == IO_UNWRITTEN) {
1470 ioend->io_iocb = iocb;
1471 ioend->io_result = ret;
1473 aio_complete(iocb, ret, 0);
1475 xfs_finish_ioend(ioend);
1477 xfs_finish_ioend_sync(ioend);
1485 const struct iovec *iov,
1487 unsigned long nr_segs)
1489 struct inode *inode = iocb->ki_filp->f_mapping->host;
1490 struct block_device *bdev = xfs_find_bdev_for_inode(inode);
1494 iocb->private = xfs_alloc_ioend(inode, IO_DIRECT);
1496 ret = __blockdev_direct_IO(rw, iocb, inode, bdev, iov,
1498 xfs_get_blocks_direct,
1499 xfs_end_io_direct_write, NULL, 0);
1500 if (ret != -EIOCBQUEUED && iocb->private)
1501 xfs_destroy_ioend(iocb->private);
1503 ret = __blockdev_direct_IO(rw, iocb, inode, bdev, iov,
1505 xfs_get_blocks_direct,
1513 xfs_vm_write_failed(
1514 struct address_space *mapping,
1517 struct inode *inode = mapping->host;
1519 if (to > inode->i_size) {
1521 * punch out the delalloc blocks we have already allocated. We
1522 * don't call xfs_setattr() to do this as we may be in the
1523 * middle of a multi-iovec write and so the vfs inode->i_size
1524 * will not match the xfs ip->i_size and so it will zero too
1525 * much. Hence we jus truncate the page cache to zero what is
1526 * necessary and punch the delalloc blocks directly.
1528 struct xfs_inode *ip = XFS_I(inode);
1529 xfs_fileoff_t start_fsb;
1530 xfs_fileoff_t end_fsb;
1533 truncate_pagecache(inode, to, inode->i_size);
1536 * Check if there are any blocks that are outside of i_size
1537 * that need to be trimmed back.
1539 start_fsb = XFS_B_TO_FSB(ip->i_mount, inode->i_size) + 1;
1540 end_fsb = XFS_B_TO_FSB(ip->i_mount, to);
1541 if (end_fsb <= start_fsb)
1544 xfs_ilock(ip, XFS_ILOCK_EXCL);
1545 error = xfs_bmap_punch_delalloc_range(ip, start_fsb,
1546 end_fsb - start_fsb);
1548 /* something screwed, just bail */
1549 if (!XFS_FORCED_SHUTDOWN(ip->i_mount)) {
1550 xfs_fs_cmn_err(CE_ALERT, ip->i_mount,
1551 "xfs_vm_write_failed: unable to clean up ino %lld",
1555 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1562 struct address_space *mapping,
1566 struct page **pagep,
1571 ret = block_write_begin(mapping, pos, len, flags | AOP_FLAG_NOFS,
1572 pagep, xfs_get_blocks);
1574 xfs_vm_write_failed(mapping, pos + len);
1581 struct address_space *mapping,
1590 ret = generic_write_end(file, mapping, pos, len, copied, page, fsdata);
1591 if (unlikely(ret < len))
1592 xfs_vm_write_failed(mapping, pos + len);
1598 struct address_space *mapping,
1601 struct inode *inode = (struct inode *)mapping->host;
1602 struct xfs_inode *ip = XFS_I(inode);
1604 trace_xfs_vm_bmap(XFS_I(inode));
1605 xfs_ilock(ip, XFS_IOLOCK_SHARED);
1606 xfs_flush_pages(ip, (xfs_off_t)0, -1, 0, FI_REMAPF);
1607 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
1608 return generic_block_bmap(mapping, block, xfs_get_blocks);
1613 struct file *unused,
1616 return mpage_readpage(page, xfs_get_blocks);
1621 struct file *unused,
1622 struct address_space *mapping,
1623 struct list_head *pages,
1626 return mpage_readpages(mapping, pages, nr_pages, xfs_get_blocks);
1629 const struct address_space_operations xfs_address_space_operations = {
1630 .readpage = xfs_vm_readpage,
1631 .readpages = xfs_vm_readpages,
1632 .writepage = xfs_vm_writepage,
1633 .writepages = xfs_vm_writepages,
1634 .sync_page = block_sync_page,
1635 .releasepage = xfs_vm_releasepage,
1636 .invalidatepage = xfs_vm_invalidatepage,
1637 .write_begin = xfs_vm_write_begin,
1638 .write_end = xfs_vm_write_end,
1639 .bmap = xfs_vm_bmap,
1640 .direct_IO = xfs_vm_direct_IO,
1641 .migratepage = buffer_migrate_page,
1642 .is_partially_uptodate = block_is_partially_uptodate,
1643 .error_remove_page = generic_error_remove_page,