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
19 #include "xfs_shared.h"
20 #include "xfs_format.h"
21 #include "xfs_log_format.h"
22 #include "xfs_trans_resv.h"
25 #include "xfs_mount.h"
26 #include "xfs_inode.h"
27 #include "xfs_trans.h"
28 #include "xfs_inode_item.h"
29 #include "xfs_alloc.h"
30 #include "xfs_error.h"
31 #include "xfs_iomap.h"
32 #include "xfs_trace.h"
34 #include "xfs_bmap_util.h"
35 #include "xfs_bmap_btree.h"
36 #include "xfs_dinode.h"
37 #include <linux/aio.h>
38 #include <linux/gfp.h>
39 #include <linux/mpage.h>
40 #include <linux/pagevec.h>
41 #include <linux/writeback.h>
49 struct buffer_head *bh, *head;
51 *delalloc = *unwritten = 0;
53 bh = head = page_buffers(page);
55 if (buffer_unwritten(bh))
57 else if (buffer_delay(bh))
59 } while ((bh = bh->b_this_page) != head);
62 STATIC struct block_device *
63 xfs_find_bdev_for_inode(
66 struct xfs_inode *ip = XFS_I(inode);
67 struct xfs_mount *mp = ip->i_mount;
69 if (XFS_IS_REALTIME_INODE(ip))
70 return mp->m_rtdev_targp->bt_bdev;
72 return mp->m_ddev_targp->bt_bdev;
76 * We're now finished for good with this ioend structure.
77 * Update the page state via the associated buffer_heads,
78 * release holds on the inode and bio, and finally free
79 * up memory. Do not use the ioend after this.
85 struct buffer_head *bh, *next;
87 for (bh = ioend->io_buffer_head; bh; bh = next) {
89 bh->b_end_io(bh, !ioend->io_error);
92 mempool_free(ioend, xfs_ioend_pool);
96 * Fast and loose check if this write could update the on-disk inode size.
98 static inline bool xfs_ioend_is_append(struct xfs_ioend *ioend)
100 return ioend->io_offset + ioend->io_size >
101 XFS_I(ioend->io_inode)->i_d.di_size;
105 xfs_setfilesize_trans_alloc(
106 struct xfs_ioend *ioend)
108 struct xfs_mount *mp = XFS_I(ioend->io_inode)->i_mount;
109 struct xfs_trans *tp;
112 tp = xfs_trans_alloc(mp, XFS_TRANS_FSYNC_TS);
114 error = xfs_trans_reserve(tp, &M_RES(mp)->tr_fsyncts, 0, 0);
116 xfs_trans_cancel(tp, 0);
120 ioend->io_append_trans = tp;
123 * We may pass freeze protection with a transaction. So tell lockdep
126 rwsem_release(&ioend->io_inode->i_sb->s_writers.lock_map[SB_FREEZE_FS-1],
129 * We hand off the transaction to the completion thread now, so
130 * clear the flag here.
132 current_restore_flags_nested(&tp->t_pflags, PF_FSTRANS);
137 * Update on-disk file size now that data has been written to disk.
141 struct xfs_ioend *ioend)
143 struct xfs_inode *ip = XFS_I(ioend->io_inode);
144 struct xfs_trans *tp = ioend->io_append_trans;
148 * The transaction may have been allocated in the I/O submission thread,
149 * thus we need to mark ourselves as beeing in a transaction manually.
150 * Similarly for freeze protection.
152 current_set_flags_nested(&tp->t_pflags, PF_FSTRANS);
153 rwsem_acquire_read(&VFS_I(ip)->i_sb->s_writers.lock_map[SB_FREEZE_FS-1],
156 xfs_ilock(ip, XFS_ILOCK_EXCL);
157 isize = xfs_new_eof(ip, ioend->io_offset + ioend->io_size);
159 xfs_iunlock(ip, XFS_ILOCK_EXCL);
160 xfs_trans_cancel(tp, 0);
164 trace_xfs_setfilesize(ip, ioend->io_offset, ioend->io_size);
166 ip->i_d.di_size = isize;
167 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
168 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
170 return xfs_trans_commit(tp, 0);
174 * Schedule IO completion handling on the final put of an ioend.
176 * If there is no work to do we might as well call it a day and free the
181 struct xfs_ioend *ioend)
183 if (atomic_dec_and_test(&ioend->io_remaining)) {
184 struct xfs_mount *mp = XFS_I(ioend->io_inode)->i_mount;
186 if (ioend->io_type == XFS_IO_UNWRITTEN)
187 queue_work(mp->m_unwritten_workqueue, &ioend->io_work);
188 else if (ioend->io_append_trans ||
189 (ioend->io_isdirect && xfs_ioend_is_append(ioend)))
190 queue_work(mp->m_data_workqueue, &ioend->io_work);
192 xfs_destroy_ioend(ioend);
197 * IO write completion.
201 struct work_struct *work)
203 xfs_ioend_t *ioend = container_of(work, xfs_ioend_t, io_work);
204 struct xfs_inode *ip = XFS_I(ioend->io_inode);
207 if (XFS_FORCED_SHUTDOWN(ip->i_mount)) {
208 ioend->io_error = -EIO;
215 * For unwritten extents we need to issue transactions to convert a
216 * range to normal written extens after the data I/O has finished.
218 if (ioend->io_type == XFS_IO_UNWRITTEN) {
219 error = xfs_iomap_write_unwritten(ip, ioend->io_offset,
221 } else if (ioend->io_isdirect && xfs_ioend_is_append(ioend)) {
223 * For direct I/O we do not know if we need to allocate blocks
224 * or not so we can't preallocate an append transaction as that
225 * results in nested reservations and log space deadlocks. Hence
226 * allocate the transaction here. While this is sub-optimal and
227 * can block IO completion for some time, we're stuck with doing
228 * it this way until we can pass the ioend to the direct IO
229 * allocation callbacks and avoid nesting that way.
231 error = xfs_setfilesize_trans_alloc(ioend);
234 error = xfs_setfilesize(ioend);
235 } else if (ioend->io_append_trans) {
236 error = xfs_setfilesize(ioend);
238 ASSERT(!xfs_ioend_is_append(ioend));
243 ioend->io_error = -error;
244 xfs_destroy_ioend(ioend);
248 * Call IO completion handling in caller context on the final put of an ioend.
251 xfs_finish_ioend_sync(
252 struct xfs_ioend *ioend)
254 if (atomic_dec_and_test(&ioend->io_remaining))
255 xfs_end_io(&ioend->io_work);
259 * Allocate and initialise an IO completion structure.
260 * We need to track unwritten extent write completion here initially.
261 * We'll need to extend this for updating the ondisk inode size later
271 ioend = mempool_alloc(xfs_ioend_pool, GFP_NOFS);
274 * Set the count to 1 initially, which will prevent an I/O
275 * completion callback from happening before we have started
276 * all the I/O from calling the completion routine too early.
278 atomic_set(&ioend->io_remaining, 1);
279 ioend->io_isdirect = 0;
281 ioend->io_list = NULL;
282 ioend->io_type = type;
283 ioend->io_inode = inode;
284 ioend->io_buffer_head = NULL;
285 ioend->io_buffer_tail = NULL;
286 ioend->io_offset = 0;
288 ioend->io_append_trans = NULL;
290 INIT_WORK(&ioend->io_work, xfs_end_io);
298 struct xfs_bmbt_irec *imap,
302 struct xfs_inode *ip = XFS_I(inode);
303 struct xfs_mount *mp = ip->i_mount;
304 ssize_t count = 1 << inode->i_blkbits;
305 xfs_fileoff_t offset_fsb, end_fsb;
307 int bmapi_flags = XFS_BMAPI_ENTIRE;
310 if (XFS_FORCED_SHUTDOWN(mp))
311 return -XFS_ERROR(EIO);
313 if (type == XFS_IO_UNWRITTEN)
314 bmapi_flags |= XFS_BMAPI_IGSTATE;
316 if (!xfs_ilock_nowait(ip, XFS_ILOCK_SHARED)) {
318 return -XFS_ERROR(EAGAIN);
319 xfs_ilock(ip, XFS_ILOCK_SHARED);
322 ASSERT(ip->i_d.di_format != XFS_DINODE_FMT_BTREE ||
323 (ip->i_df.if_flags & XFS_IFEXTENTS));
324 ASSERT(offset <= mp->m_super->s_maxbytes);
326 if (offset + count > mp->m_super->s_maxbytes)
327 count = mp->m_super->s_maxbytes - offset;
328 end_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)offset + count);
329 offset_fsb = XFS_B_TO_FSBT(mp, offset);
330 error = xfs_bmapi_read(ip, offset_fsb, end_fsb - offset_fsb,
331 imap, &nimaps, bmapi_flags);
332 xfs_iunlock(ip, XFS_ILOCK_SHARED);
335 return -XFS_ERROR(error);
337 if (type == XFS_IO_DELALLOC &&
338 (!nimaps || isnullstartblock(imap->br_startblock))) {
339 error = xfs_iomap_write_allocate(ip, offset, imap);
341 trace_xfs_map_blocks_alloc(ip, offset, count, type, imap);
342 return -XFS_ERROR(error);
346 if (type == XFS_IO_UNWRITTEN) {
348 ASSERT(imap->br_startblock != HOLESTARTBLOCK);
349 ASSERT(imap->br_startblock != DELAYSTARTBLOCK);
353 trace_xfs_map_blocks_found(ip, offset, count, type, imap);
360 struct xfs_bmbt_irec *imap,
363 offset >>= inode->i_blkbits;
365 return offset >= imap->br_startoff &&
366 offset < imap->br_startoff + imap->br_blockcount;
370 * BIO completion handler for buffered IO.
377 xfs_ioend_t *ioend = bio->bi_private;
379 ASSERT(atomic_read(&bio->bi_cnt) >= 1);
380 ioend->io_error = test_bit(BIO_UPTODATE, &bio->bi_flags) ? 0 : error;
382 /* Toss bio and pass work off to an xfsdatad thread */
383 bio->bi_private = NULL;
384 bio->bi_end_io = NULL;
387 xfs_finish_ioend(ioend);
391 xfs_submit_ioend_bio(
392 struct writeback_control *wbc,
396 atomic_inc(&ioend->io_remaining);
397 bio->bi_private = ioend;
398 bio->bi_end_io = xfs_end_bio;
399 submit_bio(wbc->sync_mode == WB_SYNC_ALL ? WRITE_SYNC : WRITE, bio);
404 struct buffer_head *bh)
406 int nvecs = bio_get_nr_vecs(bh->b_bdev);
407 struct bio *bio = bio_alloc(GFP_NOIO, nvecs);
409 ASSERT(bio->bi_private == NULL);
410 bio->bi_iter.bi_sector = bh->b_blocknr * (bh->b_size >> 9);
411 bio->bi_bdev = bh->b_bdev;
416 xfs_start_buffer_writeback(
417 struct buffer_head *bh)
419 ASSERT(buffer_mapped(bh));
420 ASSERT(buffer_locked(bh));
421 ASSERT(!buffer_delay(bh));
422 ASSERT(!buffer_unwritten(bh));
424 mark_buffer_async_write(bh);
425 set_buffer_uptodate(bh);
426 clear_buffer_dirty(bh);
430 xfs_start_page_writeback(
435 ASSERT(PageLocked(page));
436 ASSERT(!PageWriteback(page));
439 * if the page was not fully cleaned, we need to ensure that the higher
440 * layers come back to it correctly. That means we need to keep the page
441 * dirty, and for WB_SYNC_ALL writeback we need to ensure the
442 * PAGECACHE_TAG_TOWRITE index mark is not removed so another attempt to
443 * write this page in this writeback sweep will be made.
446 clear_page_dirty_for_io(page);
447 set_page_writeback(page);
449 set_page_writeback_keepwrite(page);
453 /* If no buffers on the page are to be written, finish it here */
455 end_page_writeback(page);
458 static inline int xfs_bio_add_buffer(struct bio *bio, struct buffer_head *bh)
460 return bio_add_page(bio, bh->b_page, bh->b_size, bh_offset(bh));
464 * Submit all of the bios for all of the ioends we have saved up, covering the
465 * initial writepage page and also any probed pages.
467 * Because we may have multiple ioends spanning a page, we need to start
468 * writeback on all the buffers before we submit them for I/O. If we mark the
469 * buffers as we got, then we can end up with a page that only has buffers
470 * marked async write and I/O complete on can occur before we mark the other
471 * buffers async write.
473 * The end result of this is that we trip a bug in end_page_writeback() because
474 * we call it twice for the one page as the code in end_buffer_async_write()
475 * assumes that all buffers on the page are started at the same time.
477 * The fix is two passes across the ioend list - one to start writeback on the
478 * buffer_heads, and then submit them for I/O on the second pass.
480 * If @fail is non-zero, it means that we have a situation where some part of
481 * the submission process has failed after we have marked paged for writeback
482 * and unlocked them. In this situation, we need to fail the ioend chain rather
483 * than submit it to IO. This typically only happens on a filesystem shutdown.
487 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;
511 * If we are failing the IO now, just mark the ioend with an
512 * error and finish it. This will run IO completion immediately
513 * as there is only one reference to the ioend at this point in
517 ioend->io_error = -fail;
518 xfs_finish_ioend(ioend);
522 for (bh = ioend->io_buffer_head; bh; bh = bh->b_private) {
526 bio = xfs_alloc_ioend_bio(bh);
527 } else if (bh->b_blocknr != lastblock + 1) {
528 xfs_submit_ioend_bio(wbc, ioend, bio);
532 if (xfs_bio_add_buffer(bio, bh) != bh->b_size) {
533 xfs_submit_ioend_bio(wbc, ioend, bio);
537 lastblock = bh->b_blocknr;
540 xfs_submit_ioend_bio(wbc, ioend, bio);
541 xfs_finish_ioend(ioend);
542 } while ((ioend = next) != NULL);
546 * Cancel submission of all buffer_heads so far in this endio.
547 * Toss the endio too. Only ever called for the initial page
548 * in a writepage request, so only ever one page.
555 struct buffer_head *bh, *next_bh;
558 next = ioend->io_list;
559 bh = ioend->io_buffer_head;
561 next_bh = bh->b_private;
562 clear_buffer_async_write(bh);
564 } while ((bh = next_bh) != NULL);
566 mempool_free(ioend, xfs_ioend_pool);
567 } while ((ioend = next) != NULL);
571 * Test to see if we've been building up a completion structure for
572 * earlier buffers -- if so, we try to append to this ioend if we
573 * can, otherwise we finish off any current ioend and start another.
574 * Return true if we've finished the given ioend.
579 struct buffer_head *bh,
582 xfs_ioend_t **result,
585 xfs_ioend_t *ioend = *result;
587 if (!ioend || need_ioend || type != ioend->io_type) {
588 xfs_ioend_t *previous = *result;
590 ioend = xfs_alloc_ioend(inode, type);
591 ioend->io_offset = offset;
592 ioend->io_buffer_head = bh;
593 ioend->io_buffer_tail = bh;
595 previous->io_list = ioend;
598 ioend->io_buffer_tail->b_private = bh;
599 ioend->io_buffer_tail = bh;
602 bh->b_private = NULL;
603 ioend->io_size += bh->b_size;
609 struct buffer_head *bh,
610 struct xfs_bmbt_irec *imap,
614 struct xfs_mount *m = XFS_I(inode)->i_mount;
615 xfs_off_t iomap_offset = XFS_FSB_TO_B(m, imap->br_startoff);
616 xfs_daddr_t iomap_bn = xfs_fsb_to_db(XFS_I(inode), imap->br_startblock);
618 ASSERT(imap->br_startblock != HOLESTARTBLOCK);
619 ASSERT(imap->br_startblock != DELAYSTARTBLOCK);
621 bn = (iomap_bn >> (inode->i_blkbits - BBSHIFT)) +
622 ((offset - iomap_offset) >> inode->i_blkbits);
624 ASSERT(bn || XFS_IS_REALTIME_INODE(XFS_I(inode)));
627 set_buffer_mapped(bh);
633 struct buffer_head *bh,
634 struct xfs_bmbt_irec *imap,
637 ASSERT(imap->br_startblock != HOLESTARTBLOCK);
638 ASSERT(imap->br_startblock != DELAYSTARTBLOCK);
640 xfs_map_buffer(inode, bh, imap, offset);
641 set_buffer_mapped(bh);
642 clear_buffer_delay(bh);
643 clear_buffer_unwritten(bh);
647 * Test if a given page is suitable for writing as part of an unwritten
648 * or delayed allocate extent.
655 if (PageWriteback(page))
658 if (page->mapping && page_has_buffers(page)) {
659 struct buffer_head *bh, *head;
662 bh = head = page_buffers(page);
664 if (buffer_unwritten(bh))
665 acceptable += (type == XFS_IO_UNWRITTEN);
666 else if (buffer_delay(bh))
667 acceptable += (type == XFS_IO_DELALLOC);
668 else if (buffer_dirty(bh) && buffer_mapped(bh))
669 acceptable += (type == XFS_IO_OVERWRITE);
672 } while ((bh = bh->b_this_page) != head);
682 * Allocate & map buffers for page given the extent map. Write it out.
683 * except for the original page of a writepage, this is called on
684 * delalloc/unwritten pages only, for the original page it is possible
685 * that the page has no mapping at all.
692 struct xfs_bmbt_irec *imap,
693 xfs_ioend_t **ioendp,
694 struct writeback_control *wbc)
696 struct buffer_head *bh, *head;
697 xfs_off_t end_offset;
698 unsigned long p_offset;
701 int count = 0, done = 0, uptodate = 1;
702 xfs_off_t offset = page_offset(page);
704 if (page->index != tindex)
706 if (!trylock_page(page))
708 if (PageWriteback(page))
709 goto fail_unlock_page;
710 if (page->mapping != inode->i_mapping)
711 goto fail_unlock_page;
712 if (!xfs_check_page_type(page, (*ioendp)->io_type))
713 goto fail_unlock_page;
716 * page_dirty is initially a count of buffers on the page before
717 * EOF and is decremented as we move each into a cleanable state.
721 * End offset is the highest offset that this page should represent.
722 * If we are on the last page, (end_offset & (PAGE_CACHE_SIZE - 1))
723 * will evaluate non-zero and be less than PAGE_CACHE_SIZE and
724 * hence give us the correct page_dirty count. On any other page,
725 * it will be zero and in that case we need page_dirty to be the
726 * count of buffers on the page.
728 end_offset = min_t(unsigned long long,
729 (xfs_off_t)(page->index + 1) << PAGE_CACHE_SHIFT,
733 * If the current map does not span the entire page we are about to try
734 * to write, then give up. The only way we can write a page that spans
735 * multiple mappings in a single writeback iteration is via the
736 * xfs_vm_writepage() function. Data integrity writeback requires the
737 * entire page to be written in a single attempt, otherwise the part of
738 * the page we don't write here doesn't get written as part of the data
741 * For normal writeback, we also don't attempt to write partial pages
742 * here as it simply means that write_cache_pages() will see it under
743 * writeback and ignore the page until some point in the future, at
744 * which time this will be the only page in the file that needs
745 * writeback. Hence for more optimal IO patterns, we should always
746 * avoid partial page writeback due to multiple mappings on a page here.
748 if (!xfs_imap_valid(inode, imap, end_offset))
749 goto fail_unlock_page;
751 len = 1 << inode->i_blkbits;
752 p_offset = min_t(unsigned long, end_offset & (PAGE_CACHE_SIZE - 1),
754 p_offset = p_offset ? roundup(p_offset, len) : PAGE_CACHE_SIZE;
755 page_dirty = p_offset / len;
757 bh = head = page_buffers(page);
759 if (offset >= end_offset)
761 if (!buffer_uptodate(bh))
763 if (!(PageUptodate(page) || buffer_uptodate(bh))) {
768 if (buffer_unwritten(bh) || buffer_delay(bh) ||
770 if (buffer_unwritten(bh))
771 type = XFS_IO_UNWRITTEN;
772 else if (buffer_delay(bh))
773 type = XFS_IO_DELALLOC;
775 type = XFS_IO_OVERWRITE;
777 if (!xfs_imap_valid(inode, imap, offset)) {
783 if (type != XFS_IO_OVERWRITE)
784 xfs_map_at_offset(inode, bh, imap, offset);
785 xfs_add_to_ioend(inode, bh, offset, type,
793 } while (offset += len, (bh = bh->b_this_page) != head);
795 if (uptodate && bh == head)
796 SetPageUptodate(page);
799 if (--wbc->nr_to_write <= 0 &&
800 wbc->sync_mode == WB_SYNC_NONE)
803 xfs_start_page_writeback(page, !page_dirty, count);
813 * Convert & write out a cluster of pages in the same extent as defined
814 * by mp and following the start page.
820 struct xfs_bmbt_irec *imap,
821 xfs_ioend_t **ioendp,
822 struct writeback_control *wbc,
828 pagevec_init(&pvec, 0);
829 while (!done && tindex <= tlast) {
830 unsigned len = min_t(pgoff_t, PAGEVEC_SIZE, tlast - tindex + 1);
832 if (!pagevec_lookup(&pvec, inode->i_mapping, tindex, len))
835 for (i = 0; i < pagevec_count(&pvec); i++) {
836 done = xfs_convert_page(inode, pvec.pages[i], tindex++,
842 pagevec_release(&pvec);
848 xfs_vm_invalidatepage(
853 trace_xfs_invalidatepage(page->mapping->host, page, offset,
855 block_invalidatepage(page, offset, length);
859 * If the page has delalloc buffers on it, we need to punch them out before we
860 * invalidate the page. If we don't, we leave a stale delalloc mapping on the
861 * inode that can trip a BUG() in xfs_get_blocks() later on if a direct IO read
862 * is done on that same region - the delalloc extent is returned when none is
863 * supposed to be there.
865 * We prevent this by truncating away the delalloc regions on the page before
866 * invalidating it. Because they are delalloc, we can do this without needing a
867 * transaction. Indeed - if we get ENOSPC errors, we have to be able to do this
868 * truncation without a transaction as there is no space left for block
869 * reservation (typically why we see a ENOSPC in writeback).
871 * This is not a performance critical path, so for now just do the punching a
872 * buffer head at a time.
875 xfs_aops_discard_page(
878 struct inode *inode = page->mapping->host;
879 struct xfs_inode *ip = XFS_I(inode);
880 struct buffer_head *bh, *head;
881 loff_t offset = page_offset(page);
883 if (!xfs_check_page_type(page, XFS_IO_DELALLOC))
886 if (XFS_FORCED_SHUTDOWN(ip->i_mount))
889 xfs_alert(ip->i_mount,
890 "page discard on page %p, inode 0x%llx, offset %llu.",
891 page, ip->i_ino, offset);
893 xfs_ilock(ip, XFS_ILOCK_EXCL);
894 bh = head = page_buffers(page);
897 xfs_fileoff_t start_fsb;
899 if (!buffer_delay(bh))
902 start_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
903 error = xfs_bmap_punch_delalloc_range(ip, start_fsb, 1);
905 /* something screwed, just bail */
906 if (!XFS_FORCED_SHUTDOWN(ip->i_mount)) {
907 xfs_alert(ip->i_mount,
908 "page discard unable to remove delalloc mapping.");
913 offset += 1 << inode->i_blkbits;
915 } while ((bh = bh->b_this_page) != head);
917 xfs_iunlock(ip, XFS_ILOCK_EXCL);
919 xfs_vm_invalidatepage(page, 0, PAGE_CACHE_SIZE);
924 * Write out a dirty page.
926 * For delalloc space on the page we need to allocate space and flush it.
927 * For unwritten space on the page we need to start the conversion to
928 * regular allocated space.
929 * For any other dirty buffer heads on the page we should flush them.
934 struct writeback_control *wbc)
936 struct inode *inode = page->mapping->host;
937 struct buffer_head *bh, *head;
938 struct xfs_bmbt_irec imap;
939 xfs_ioend_t *ioend = NULL, *iohead = NULL;
942 __uint64_t end_offset;
943 pgoff_t end_index, last_index;
945 int err, imap_valid = 0, uptodate = 1;
949 trace_xfs_writepage(inode, page, 0, 0);
951 ASSERT(page_has_buffers(page));
954 * Refuse to write the page out if we are called from reclaim context.
956 * This avoids stack overflows when called from deeply used stacks in
957 * random callers for direct reclaim or memcg reclaim. We explicitly
958 * allow reclaim from kswapd as the stack usage there is relatively low.
960 * This should never happen except in the case of a VM regression so
963 if (WARN_ON_ONCE((current->flags & (PF_MEMALLOC|PF_KSWAPD)) ==
968 * Given that we do not allow direct reclaim to call us, we should
969 * never be called while in a filesystem transaction.
971 if (WARN_ON(current->flags & PF_FSTRANS))
974 /* Is this page beyond the end of the file? */
975 offset = i_size_read(inode);
976 end_index = offset >> PAGE_CACHE_SHIFT;
977 last_index = (offset - 1) >> PAGE_CACHE_SHIFT;
978 if (page->index >= end_index) {
979 unsigned offset_into_page = offset & (PAGE_CACHE_SIZE - 1);
982 * Skip the page if it is fully outside i_size, e.g. due to a
983 * truncate operation that is in progress. We must redirty the
984 * page so that reclaim stops reclaiming it. Otherwise
985 * xfs_vm_releasepage() is called on it and gets confused.
987 if (page->index >= end_index + 1 || offset_into_page == 0)
991 * The page straddles i_size. It must be zeroed out on each
992 * and every writepage invocation because it may be mmapped.
993 * "A file is mapped in multiples of the page size. For a file
994 * that is not a multiple of the page size, the remaining
995 * memory is zeroed when mapped, and writes to that region are
996 * not written out to the file."
998 zero_user_segment(page, offset_into_page, PAGE_CACHE_SIZE);
1001 end_offset = min_t(unsigned long long,
1002 (xfs_off_t)(page->index + 1) << PAGE_CACHE_SHIFT,
1004 len = 1 << inode->i_blkbits;
1006 bh = head = page_buffers(page);
1007 offset = page_offset(page);
1008 type = XFS_IO_OVERWRITE;
1010 if (wbc->sync_mode == WB_SYNC_NONE)
1016 if (offset >= end_offset)
1018 if (!buffer_uptodate(bh))
1022 * set_page_dirty dirties all buffers in a page, independent
1023 * of their state. The dirty state however is entirely
1024 * meaningless for holes (!mapped && uptodate), so skip
1025 * buffers covering holes here.
1027 if (!buffer_mapped(bh) && buffer_uptodate(bh)) {
1032 if (buffer_unwritten(bh)) {
1033 if (type != XFS_IO_UNWRITTEN) {
1034 type = XFS_IO_UNWRITTEN;
1037 } else if (buffer_delay(bh)) {
1038 if (type != XFS_IO_DELALLOC) {
1039 type = XFS_IO_DELALLOC;
1042 } else if (buffer_uptodate(bh)) {
1043 if (type != XFS_IO_OVERWRITE) {
1044 type = XFS_IO_OVERWRITE;
1048 if (PageUptodate(page))
1049 ASSERT(buffer_mapped(bh));
1051 * This buffer is not uptodate and will not be
1052 * written to disk. Ensure that we will put any
1053 * subsequent writeable buffers into a new
1061 imap_valid = xfs_imap_valid(inode, &imap, offset);
1064 * If we didn't have a valid mapping then we need to
1065 * put the new mapping into a separate ioend structure.
1066 * This ensures non-contiguous extents always have
1067 * separate ioends, which is particularly important
1068 * for unwritten extent conversion at I/O completion
1072 err = xfs_map_blocks(inode, offset, &imap, type,
1076 imap_valid = xfs_imap_valid(inode, &imap, offset);
1080 if (type != XFS_IO_OVERWRITE)
1081 xfs_map_at_offset(inode, bh, &imap, offset);
1082 xfs_add_to_ioend(inode, bh, offset, type, &ioend,
1090 } while (offset += len, ((bh = bh->b_this_page) != head));
1092 if (uptodate && bh == head)
1093 SetPageUptodate(page);
1095 xfs_start_page_writeback(page, 1, count);
1097 /* if there is no IO to be submitted for this page, we are done */
1104 * Any errors from this point onwards need tobe reported through the IO
1105 * completion path as we have marked the initial page as under writeback
1109 xfs_off_t end_index;
1111 end_index = imap.br_startoff + imap.br_blockcount;
1114 end_index <<= inode->i_blkbits;
1117 end_index = (end_index - 1) >> PAGE_CACHE_SHIFT;
1119 /* check against file size */
1120 if (end_index > last_index)
1121 end_index = last_index;
1123 xfs_cluster_write(inode, page->index + 1, &imap, &ioend,
1129 * Reserve log space if we might write beyond the on-disk inode size.
1132 if (ioend->io_type != XFS_IO_UNWRITTEN && xfs_ioend_is_append(ioend))
1133 err = xfs_setfilesize_trans_alloc(ioend);
1135 xfs_submit_ioend(wbc, iohead, err);
1141 xfs_cancel_ioend(iohead);
1146 xfs_aops_discard_page(page);
1147 ClearPageUptodate(page);
1152 redirty_page_for_writepage(wbc, page);
1159 struct address_space *mapping,
1160 struct writeback_control *wbc)
1162 xfs_iflags_clear(XFS_I(mapping->host), XFS_ITRUNCATED);
1163 return generic_writepages(mapping, wbc);
1167 * Called to move a page into cleanable state - and from there
1168 * to be released. The page should already be clean. We always
1169 * have buffer heads in this call.
1171 * Returns 1 if the page is ok to release, 0 otherwise.
1178 int delalloc, unwritten;
1180 trace_xfs_releasepage(page->mapping->host, page, 0, 0);
1182 xfs_count_page_state(page, &delalloc, &unwritten);
1184 if (WARN_ON(delalloc))
1186 if (WARN_ON(unwritten))
1189 return try_to_free_buffers(page);
1194 struct inode *inode,
1196 struct buffer_head *bh_result,
1200 struct xfs_inode *ip = XFS_I(inode);
1201 struct xfs_mount *mp = ip->i_mount;
1202 xfs_fileoff_t offset_fsb, end_fsb;
1205 struct xfs_bmbt_irec imap;
1211 if (XFS_FORCED_SHUTDOWN(mp))
1212 return -XFS_ERROR(EIO);
1214 offset = (xfs_off_t)iblock << inode->i_blkbits;
1215 ASSERT(bh_result->b_size >= (1 << inode->i_blkbits));
1216 size = bh_result->b_size;
1218 if (!create && direct && offset >= i_size_read(inode))
1222 * Direct I/O is usually done on preallocated files, so try getting
1223 * a block mapping without an exclusive lock first. For buffered
1224 * writes we already have the exclusive iolock anyway, so avoiding
1225 * a lock roundtrip here by taking the ilock exclusive from the
1226 * beginning is a useful micro optimization.
1228 if (create && !direct) {
1229 lockmode = XFS_ILOCK_EXCL;
1230 xfs_ilock(ip, lockmode);
1232 lockmode = xfs_ilock_data_map_shared(ip);
1235 ASSERT(offset <= mp->m_super->s_maxbytes);
1236 if (offset + size > mp->m_super->s_maxbytes)
1237 size = mp->m_super->s_maxbytes - offset;
1238 end_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)offset + size);
1239 offset_fsb = XFS_B_TO_FSBT(mp, offset);
1241 error = xfs_bmapi_read(ip, offset_fsb, end_fsb - offset_fsb,
1242 &imap, &nimaps, XFS_BMAPI_ENTIRE);
1248 (imap.br_startblock == HOLESTARTBLOCK ||
1249 imap.br_startblock == DELAYSTARTBLOCK))) {
1250 if (direct || xfs_get_extsz_hint(ip)) {
1252 * Drop the ilock in preparation for starting the block
1253 * allocation transaction. It will be retaken
1254 * exclusively inside xfs_iomap_write_direct for the
1255 * actual allocation.
1257 xfs_iunlock(ip, lockmode);
1258 error = xfs_iomap_write_direct(ip, offset, size,
1265 * Delalloc reservations do not require a transaction,
1266 * we can go on without dropping the lock here. If we
1267 * are allocating a new delalloc block, make sure that
1268 * we set the new flag so that we mark the buffer new so
1269 * that we know that it is newly allocated if the write
1272 if (nimaps && imap.br_startblock == HOLESTARTBLOCK)
1274 error = xfs_iomap_write_delay(ip, offset, size, &imap);
1278 xfs_iunlock(ip, lockmode);
1281 trace_xfs_get_blocks_alloc(ip, offset, size, 0, &imap);
1282 } else if (nimaps) {
1283 trace_xfs_get_blocks_found(ip, offset, size, 0, &imap);
1284 xfs_iunlock(ip, lockmode);
1286 trace_xfs_get_blocks_notfound(ip, offset, size);
1290 if (imap.br_startblock != HOLESTARTBLOCK &&
1291 imap.br_startblock != DELAYSTARTBLOCK) {
1293 * For unwritten extents do not report a disk address on
1294 * the read case (treat as if we're reading into a hole).
1296 if (create || !ISUNWRITTEN(&imap))
1297 xfs_map_buffer(inode, bh_result, &imap, offset);
1298 if (create && ISUNWRITTEN(&imap)) {
1300 bh_result->b_private = inode;
1301 set_buffer_defer_completion(bh_result);
1303 set_buffer_unwritten(bh_result);
1308 * If this is a realtime file, data may be on a different device.
1309 * to that pointed to from the buffer_head b_bdev currently.
1311 bh_result->b_bdev = xfs_find_bdev_for_inode(inode);
1314 * If we previously allocated a block out beyond eof and we are now
1315 * coming back to use it then we will need to flag it as new even if it
1316 * has a disk address.
1318 * With sub-block writes into unwritten extents we also need to mark
1319 * the buffer as new so that the unwritten parts of the buffer gets
1323 ((!buffer_mapped(bh_result) && !buffer_uptodate(bh_result)) ||
1324 (offset >= i_size_read(inode)) ||
1325 (new || ISUNWRITTEN(&imap))))
1326 set_buffer_new(bh_result);
1328 if (imap.br_startblock == DELAYSTARTBLOCK) {
1331 set_buffer_uptodate(bh_result);
1332 set_buffer_mapped(bh_result);
1333 set_buffer_delay(bh_result);
1338 * If this is O_DIRECT or the mpage code calling tell them how large
1339 * the mapping is, so that we can avoid repeated get_blocks calls.
1341 if (direct || size > (1 << inode->i_blkbits)) {
1342 xfs_off_t mapping_size;
1344 mapping_size = imap.br_startoff + imap.br_blockcount - iblock;
1345 mapping_size <<= inode->i_blkbits;
1347 ASSERT(mapping_size > 0);
1348 if (mapping_size > size)
1349 mapping_size = size;
1350 if (mapping_size > LONG_MAX)
1351 mapping_size = LONG_MAX;
1353 bh_result->b_size = mapping_size;
1359 xfs_iunlock(ip, lockmode);
1365 struct inode *inode,
1367 struct buffer_head *bh_result,
1370 return __xfs_get_blocks(inode, iblock, bh_result, create, 0);
1374 xfs_get_blocks_direct(
1375 struct inode *inode,
1377 struct buffer_head *bh_result,
1380 return __xfs_get_blocks(inode, iblock, bh_result, create, 1);
1384 * Complete a direct I/O write request.
1386 * If the private argument is non-NULL __xfs_get_blocks signals us that we
1387 * need to issue a transaction to convert the range from unwritten to written
1388 * extents. In case this is regular synchronous I/O we just call xfs_end_io
1389 * to do this and we are done. But in case this was a successful AIO
1390 * request this handler is called from interrupt context, from which we
1391 * can't start transactions. In that case offload the I/O completion to
1392 * the workqueues we also use for buffered I/O completion.
1395 xfs_end_io_direct_write(
1401 struct xfs_ioend *ioend = iocb->private;
1404 * While the generic direct I/O code updates the inode size, it does
1405 * so only after the end_io handler is called, which means our
1406 * end_io handler thinks the on-disk size is outside the in-core
1407 * size. To prevent this just update it a little bit earlier here.
1409 if (offset + size > i_size_read(ioend->io_inode))
1410 i_size_write(ioend->io_inode, offset + size);
1413 * blockdev_direct_IO can return an error even after the I/O
1414 * completion handler was called. Thus we need to protect
1415 * against double-freeing.
1417 iocb->private = NULL;
1419 ioend->io_offset = offset;
1420 ioend->io_size = size;
1421 if (private && size > 0)
1422 ioend->io_type = XFS_IO_UNWRITTEN;
1424 xfs_finish_ioend_sync(ioend);
1431 const struct iovec *iov,
1433 unsigned long nr_segs)
1435 struct inode *inode = iocb->ki_filp->f_mapping->host;
1436 struct block_device *bdev = xfs_find_bdev_for_inode(inode);
1437 struct xfs_ioend *ioend = NULL;
1441 size_t size = iov_length(iov, nr_segs);
1444 * We cannot preallocate a size update transaction here as we
1445 * don't know whether allocation is necessary or not. Hence we
1446 * can only tell IO completion that one is necessary if we are
1447 * not doing unwritten extent conversion.
1449 iocb->private = ioend = xfs_alloc_ioend(inode, XFS_IO_DIRECT);
1450 if (offset + size > XFS_I(inode)->i_d.di_size)
1451 ioend->io_isdirect = 1;
1453 ret = __blockdev_direct_IO(rw, iocb, inode, bdev, iov,
1455 xfs_get_blocks_direct,
1456 xfs_end_io_direct_write, NULL, 0);
1457 if (ret != -EIOCBQUEUED && iocb->private)
1458 goto out_destroy_ioend;
1460 ret = __blockdev_direct_IO(rw, iocb, inode, bdev, iov,
1462 xfs_get_blocks_direct,
1469 xfs_destroy_ioend(ioend);
1474 * Punch out the delalloc blocks we have already allocated.
1476 * Don't bother with xfs_setattr given that nothing can have made it to disk yet
1477 * as the page is still locked at this point.
1480 xfs_vm_kill_delalloc_range(
1481 struct inode *inode,
1485 struct xfs_inode *ip = XFS_I(inode);
1486 xfs_fileoff_t start_fsb;
1487 xfs_fileoff_t end_fsb;
1490 start_fsb = XFS_B_TO_FSB(ip->i_mount, start);
1491 end_fsb = XFS_B_TO_FSB(ip->i_mount, end);
1492 if (end_fsb <= start_fsb)
1495 xfs_ilock(ip, XFS_ILOCK_EXCL);
1496 error = xfs_bmap_punch_delalloc_range(ip, start_fsb,
1497 end_fsb - start_fsb);
1499 /* something screwed, just bail */
1500 if (!XFS_FORCED_SHUTDOWN(ip->i_mount)) {
1501 xfs_alert(ip->i_mount,
1502 "xfs_vm_write_failed: unable to clean up ino %lld",
1506 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1510 xfs_vm_write_failed(
1511 struct inode *inode,
1516 loff_t block_offset;
1519 loff_t from = pos & (PAGE_CACHE_SIZE - 1);
1520 loff_t to = from + len;
1521 struct buffer_head *bh, *head;
1524 * The request pos offset might be 32 or 64 bit, this is all fine
1525 * on 64-bit platform. However, for 64-bit pos request on 32-bit
1526 * platform, the high 32-bit will be masked off if we evaluate the
1527 * block_offset via (pos & PAGE_MASK) because the PAGE_MASK is
1528 * 0xfffff000 as an unsigned long, hence the result is incorrect
1529 * which could cause the following ASSERT failed in most cases.
1530 * In order to avoid this, we can evaluate the block_offset of the
1531 * start of the page by using shifts rather than masks the mismatch
1534 block_offset = (pos >> PAGE_CACHE_SHIFT) << PAGE_CACHE_SHIFT;
1536 ASSERT(block_offset + from == pos);
1538 head = page_buffers(page);
1540 for (bh = head; bh != head || !block_start;
1541 bh = bh->b_this_page, block_start = block_end,
1542 block_offset += bh->b_size) {
1543 block_end = block_start + bh->b_size;
1545 /* skip buffers before the write */
1546 if (block_end <= from)
1549 /* if the buffer is after the write, we're done */
1550 if (block_start >= to)
1553 if (!buffer_delay(bh))
1556 if (!buffer_new(bh) && block_offset < i_size_read(inode))
1559 xfs_vm_kill_delalloc_range(inode, block_offset,
1560 block_offset + bh->b_size);
1566 * This used to call block_write_begin(), but it unlocks and releases the page
1567 * on error, and we need that page to be able to punch stale delalloc blocks out
1568 * on failure. hence we copy-n-waste it here and call xfs_vm_write_failed() at
1569 * the appropriate point.
1574 struct address_space *mapping,
1578 struct page **pagep,
1581 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1585 ASSERT(len <= PAGE_CACHE_SIZE);
1587 page = grab_cache_page_write_begin(mapping, index, flags);
1591 status = __block_write_begin(page, pos, len, xfs_get_blocks);
1592 if (unlikely(status)) {
1593 struct inode *inode = mapping->host;
1595 xfs_vm_write_failed(inode, page, pos, len);
1598 if (pos + len > i_size_read(inode))
1599 truncate_pagecache(inode, i_size_read(inode));
1601 page_cache_release(page);
1610 * On failure, we only need to kill delalloc blocks beyond EOF because they
1611 * will never be written. For blocks within EOF, generic_write_end() zeros them
1612 * so they are safe to leave alone and be written with all the other valid data.
1617 struct address_space *mapping,
1626 ASSERT(len <= PAGE_CACHE_SIZE);
1628 ret = generic_write_end(file, mapping, pos, len, copied, page, fsdata);
1629 if (unlikely(ret < len)) {
1630 struct inode *inode = mapping->host;
1631 size_t isize = i_size_read(inode);
1632 loff_t to = pos + len;
1635 truncate_pagecache(inode, isize);
1636 xfs_vm_kill_delalloc_range(inode, isize, to);
1644 struct address_space *mapping,
1647 struct inode *inode = (struct inode *)mapping->host;
1648 struct xfs_inode *ip = XFS_I(inode);
1650 trace_xfs_vm_bmap(XFS_I(inode));
1651 xfs_ilock(ip, XFS_IOLOCK_SHARED);
1652 filemap_write_and_wait(mapping);
1653 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
1654 return generic_block_bmap(mapping, block, xfs_get_blocks);
1659 struct file *unused,
1662 return mpage_readpage(page, xfs_get_blocks);
1667 struct file *unused,
1668 struct address_space *mapping,
1669 struct list_head *pages,
1672 return mpage_readpages(mapping, pages, nr_pages, xfs_get_blocks);
1676 * This is basically a copy of __set_page_dirty_buffers() with one
1677 * small tweak: buffers beyond EOF do not get marked dirty. If we mark them
1678 * dirty, we'll never be able to clean them because we don't write buffers
1679 * beyond EOF, and that means we can't invalidate pages that span EOF
1680 * that have been marked dirty. Further, the dirty state can leak into
1681 * the file interior if the file is extended, resulting in all sorts of
1682 * bad things happening as the state does not match the underlying data.
1684 * XXX: this really indicates that bufferheads in XFS need to die. Warts like
1685 * this only exist because of bufferheads and how the generic code manages them.
1688 xfs_vm_set_page_dirty(
1691 struct address_space *mapping = page->mapping;
1692 struct inode *inode = mapping->host;
1697 if (unlikely(!mapping))
1698 return !TestSetPageDirty(page);
1700 end_offset = i_size_read(inode);
1701 offset = page_offset(page);
1703 spin_lock(&mapping->private_lock);
1704 if (page_has_buffers(page)) {
1705 struct buffer_head *head = page_buffers(page);
1706 struct buffer_head *bh = head;
1709 if (offset < end_offset)
1710 set_buffer_dirty(bh);
1711 bh = bh->b_this_page;
1712 offset += 1 << inode->i_blkbits;
1713 } while (bh != head);
1715 newly_dirty = !TestSetPageDirty(page);
1716 spin_unlock(&mapping->private_lock);
1719 /* sigh - __set_page_dirty() is static, so copy it here, too */
1720 unsigned long flags;
1722 spin_lock_irqsave(&mapping->tree_lock, flags);
1723 if (page->mapping) { /* Race with truncate? */
1724 WARN_ON_ONCE(!PageUptodate(page));
1725 account_page_dirtied(page, mapping);
1726 radix_tree_tag_set(&mapping->page_tree,
1727 page_index(page), PAGECACHE_TAG_DIRTY);
1729 spin_unlock_irqrestore(&mapping->tree_lock, flags);
1730 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1735 const struct address_space_operations xfs_address_space_operations = {
1736 .readpage = xfs_vm_readpage,
1737 .readpages = xfs_vm_readpages,
1738 .writepage = xfs_vm_writepage,
1739 .writepages = xfs_vm_writepages,
1740 .set_page_dirty = xfs_vm_set_page_dirty,
1741 .releasepage = xfs_vm_releasepage,
1742 .invalidatepage = xfs_vm_invalidatepage,
1743 .write_begin = xfs_vm_write_begin,
1744 .write_end = xfs_vm_write_end,
1745 .bmap = xfs_vm_bmap,
1746 .direct_IO = xfs_vm_direct_IO,
1747 .migratepage = buffer_migrate_page,
1748 .is_partially_uptodate = block_is_partially_uptodate,
1749 .error_remove_page = generic_error_remove_page,