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"
23 #include "xfs_mount.h"
24 #include "xfs_inode.h"
25 #include "xfs_trans.h"
26 #include "xfs_inode_item.h"
27 #include "xfs_alloc.h"
28 #include "xfs_error.h"
29 #include "xfs_iomap.h"
30 #include "xfs_trace.h"
32 #include "xfs_bmap_util.h"
33 #include "xfs_bmap_btree.h"
34 #include <linux/gfp.h>
35 #include <linux/mpage.h>
36 #include <linux/pagevec.h>
37 #include <linux/writeback.h>
39 /* flags for direct write completions */
40 #define XFS_DIO_FLAG_UNWRITTEN (1 << 0)
41 #define XFS_DIO_FLAG_APPEND (1 << 1)
44 * structure owned by writepages passed to individual writepage calls
46 struct xfs_writepage_ctx {
47 struct xfs_bmbt_irec imap;
50 struct xfs_ioend *ioend;
60 struct buffer_head *bh, *head;
62 *delalloc = *unwritten = 0;
64 bh = head = page_buffers(page);
66 if (buffer_unwritten(bh))
68 else if (buffer_delay(bh))
70 } while ((bh = bh->b_this_page) != head);
74 xfs_find_bdev_for_inode(
77 struct xfs_inode *ip = XFS_I(inode);
78 struct xfs_mount *mp = ip->i_mount;
80 if (XFS_IS_REALTIME_INODE(ip))
81 return mp->m_rtdev_targp->bt_bdev;
83 return mp->m_ddev_targp->bt_bdev;
87 * We're now finished for good with this page. Update the page state via the
88 * associated buffer_heads, paying attention to the start and end offsets that
89 * we need to process on the page.
91 * Landmine Warning: bh->b_end_io() will call end_page_writeback() on the last
92 * buffer in the IO. Once it does this, it is unsafe to access the bufferhead or
93 * the page at all, as we may be racing with memory reclaim and it can free both
94 * the bufferhead chain and the page as it will see the page as clean and
98 xfs_finish_page_writeback(
100 struct bio_vec *bvec,
103 unsigned int end = bvec->bv_offset + bvec->bv_len - 1;
104 struct buffer_head *head, *bh, *next;
105 unsigned int off = 0;
108 ASSERT(bvec->bv_offset < PAGE_SIZE);
109 ASSERT((bvec->bv_offset & ((1 << inode->i_blkbits) - 1)) == 0);
110 ASSERT(end < PAGE_SIZE);
111 ASSERT((bvec->bv_len & ((1 << inode->i_blkbits) - 1)) == 0);
113 bh = head = page_buffers(bvec->bv_page);
117 next = bh->b_this_page;
118 if (off < bvec->bv_offset)
122 bh->b_end_io(bh, !error);
125 } while ((bh = next) != head);
129 * We're now finished for good with this ioend structure. Update the page
130 * state, release holds on bios, and finally free up memory. Do not use the
135 struct xfs_ioend *ioend,
138 struct inode *inode = ioend->io_inode;
139 struct bio *last = ioend->io_bio;
140 struct bio *bio, *next;
142 for (bio = &ioend->io_inline_bio; bio; bio = next) {
143 struct bio_vec *bvec;
147 * For the last bio, bi_private points to the ioend, so we
148 * need to explicitly end the iteration here.
153 next = bio->bi_private;
155 /* walk each page on bio, ending page IO on them */
156 bio_for_each_segment_all(bvec, bio, i)
157 xfs_finish_page_writeback(inode, bvec, error);
164 * Fast and loose check if this write could update the on-disk inode size.
166 static inline bool xfs_ioend_is_append(struct xfs_ioend *ioend)
168 return ioend->io_offset + ioend->io_size >
169 XFS_I(ioend->io_inode)->i_d.di_size;
173 xfs_setfilesize_trans_alloc(
174 struct xfs_ioend *ioend)
176 struct xfs_mount *mp = XFS_I(ioend->io_inode)->i_mount;
177 struct xfs_trans *tp;
180 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_fsyncts, 0, 0, 0, &tp);
184 ioend->io_append_trans = tp;
187 * We may pass freeze protection with a transaction. So tell lockdep
190 __sb_writers_release(ioend->io_inode->i_sb, SB_FREEZE_FS);
192 * We hand off the transaction to the completion thread now, so
193 * clear the flag here.
195 current_restore_flags_nested(&tp->t_pflags, PF_FSTRANS);
200 * Update on-disk file size now that data has been written to disk.
204 struct xfs_inode *ip,
205 struct xfs_trans *tp,
211 xfs_ilock(ip, XFS_ILOCK_EXCL);
212 isize = xfs_new_eof(ip, offset + size);
214 xfs_iunlock(ip, XFS_ILOCK_EXCL);
215 xfs_trans_cancel(tp);
219 trace_xfs_setfilesize(ip, offset, size);
221 ip->i_d.di_size = isize;
222 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
223 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
225 return xfs_trans_commit(tp);
229 xfs_setfilesize_ioend(
230 struct xfs_ioend *ioend,
233 struct xfs_inode *ip = XFS_I(ioend->io_inode);
234 struct xfs_trans *tp = ioend->io_append_trans;
237 * The transaction may have been allocated in the I/O submission thread,
238 * thus we need to mark ourselves as being in a transaction manually.
239 * Similarly for freeze protection.
241 current_set_flags_nested(&tp->t_pflags, PF_FSTRANS);
242 __sb_writers_acquired(VFS_I(ip)->i_sb, SB_FREEZE_FS);
244 /* we abort the update if there was an IO error */
246 xfs_trans_cancel(tp);
250 return xfs_setfilesize(ip, tp, ioend->io_offset, ioend->io_size);
254 * IO write completion.
258 struct work_struct *work)
260 struct xfs_ioend *ioend =
261 container_of(work, struct xfs_ioend, io_work);
262 struct xfs_inode *ip = XFS_I(ioend->io_inode);
263 int error = ioend->io_bio->bi_error;
266 * Set an error if the mount has shut down and proceed with end I/O
267 * processing so it can perform whatever cleanups are necessary.
269 if (XFS_FORCED_SHUTDOWN(ip->i_mount))
273 * For unwritten extents we need to issue transactions to convert a
274 * range to normal written extens after the data I/O has finished.
275 * Detecting and handling completion IO errors is done individually
276 * for each case as different cleanup operations need to be performed
279 if (ioend->io_type == XFS_IO_UNWRITTEN) {
282 error = xfs_iomap_write_unwritten(ip, ioend->io_offset,
284 } else if (ioend->io_append_trans) {
285 error = xfs_setfilesize_ioend(ioend, error);
287 ASSERT(!xfs_ioend_is_append(ioend));
291 xfs_destroy_ioend(ioend, error);
298 struct xfs_ioend *ioend = bio->bi_private;
299 struct xfs_mount *mp = XFS_I(ioend->io_inode)->i_mount;
301 if (ioend->io_type == XFS_IO_UNWRITTEN)
302 queue_work(mp->m_unwritten_workqueue, &ioend->io_work);
303 else if (ioend->io_append_trans)
304 queue_work(mp->m_data_workqueue, &ioend->io_work);
306 xfs_destroy_ioend(ioend, bio->bi_error);
313 struct xfs_bmbt_irec *imap,
316 struct xfs_inode *ip = XFS_I(inode);
317 struct xfs_mount *mp = ip->i_mount;
318 ssize_t count = 1 << inode->i_blkbits;
319 xfs_fileoff_t offset_fsb, end_fsb;
321 int bmapi_flags = XFS_BMAPI_ENTIRE;
324 if (XFS_FORCED_SHUTDOWN(mp))
327 if (type == XFS_IO_UNWRITTEN)
328 bmapi_flags |= XFS_BMAPI_IGSTATE;
330 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_super->s_maxbytes);
335 if (offset + count > mp->m_super->s_maxbytes)
336 count = mp->m_super->s_maxbytes - 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_read(ip, offset_fsb, end_fsb - offset_fsb,
340 imap, &nimaps, bmapi_flags);
341 xfs_iunlock(ip, XFS_ILOCK_SHARED);
346 if (type == XFS_IO_DELALLOC &&
347 (!nimaps || isnullstartblock(imap->br_startblock))) {
348 error = xfs_iomap_write_allocate(ip, offset, imap);
350 trace_xfs_map_blocks_alloc(ip, offset, count, type, imap);
355 if (type == XFS_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 xfs_start_buffer_writeback(
380 struct buffer_head *bh)
382 ASSERT(buffer_mapped(bh));
383 ASSERT(buffer_locked(bh));
384 ASSERT(!buffer_delay(bh));
385 ASSERT(!buffer_unwritten(bh));
387 mark_buffer_async_write(bh);
388 set_buffer_uptodate(bh);
389 clear_buffer_dirty(bh);
393 xfs_start_page_writeback(
397 ASSERT(PageLocked(page));
398 ASSERT(!PageWriteback(page));
401 * if the page was not fully cleaned, we need to ensure that the higher
402 * layers come back to it correctly. That means we need to keep the page
403 * dirty, and for WB_SYNC_ALL writeback we need to ensure the
404 * PAGECACHE_TAG_TOWRITE index mark is not removed so another attempt to
405 * write this page in this writeback sweep will be made.
408 clear_page_dirty_for_io(page);
409 set_page_writeback(page);
411 set_page_writeback_keepwrite(page);
416 static inline int xfs_bio_add_buffer(struct bio *bio, struct buffer_head *bh)
418 return bio_add_page(bio, bh->b_page, bh->b_size, bh_offset(bh));
422 * Submit the bio for an ioend. We are passed an ioend with a bio attached to
423 * it, and we submit that bio. The ioend may be used for multiple bio
424 * submissions, so we only want to allocate an append transaction for the ioend
425 * once. In the case of multiple bio submission, each bio will take an IO
426 * reference to the ioend to ensure that the ioend completion is only done once
427 * all bios have been submitted and the ioend is really done.
429 * If @fail is non-zero, it means that we have a situation where some part of
430 * the submission process has failed after we have marked paged for writeback
431 * and unlocked them. In this situation, we need to fail the bio and ioend
432 * rather than submit it to IO. This typically only happens on a filesystem
437 struct writeback_control *wbc,
438 struct xfs_ioend *ioend,
441 /* Reserve log space if we might write beyond the on-disk inode size. */
443 ioend->io_type != XFS_IO_UNWRITTEN &&
444 xfs_ioend_is_append(ioend) &&
445 !ioend->io_append_trans)
446 status = xfs_setfilesize_trans_alloc(ioend);
448 ioend->io_bio->bi_private = ioend;
449 ioend->io_bio->bi_end_io = xfs_end_bio;
452 * If we are failing the IO now, just mark the ioend with an
453 * error and finish it. This will run IO completion immediately
454 * as there is only one reference to the ioend at this point in
458 ioend->io_bio->bi_error = status;
459 bio_endio(ioend->io_bio);
463 submit_bio(wbc->sync_mode == WB_SYNC_ALL ? WRITE_SYNC : WRITE,
469 xfs_init_bio_from_bh(
471 struct buffer_head *bh)
473 bio->bi_iter.bi_sector = bh->b_blocknr * (bh->b_size >> 9);
474 bio->bi_bdev = bh->b_bdev;
477 static struct xfs_ioend *
482 struct buffer_head *bh)
484 struct xfs_ioend *ioend;
487 bio = bio_alloc_bioset(GFP_NOFS, BIO_MAX_PAGES, xfs_ioend_bioset);
488 xfs_init_bio_from_bh(bio, bh);
490 ioend = container_of(bio, struct xfs_ioend, io_inline_bio);
491 INIT_LIST_HEAD(&ioend->io_list);
492 ioend->io_type = type;
493 ioend->io_inode = inode;
495 ioend->io_offset = offset;
496 INIT_WORK(&ioend->io_work, xfs_end_io);
497 ioend->io_append_trans = NULL;
503 * Allocate a new bio, and chain the old bio to the new one.
505 * Note that we have to do perform the chaining in this unintuitive order
506 * so that the bi_private linkage is set up in the right direction for the
507 * traversal in xfs_destroy_ioend().
511 struct xfs_ioend *ioend,
512 struct writeback_control *wbc,
513 struct buffer_head *bh)
517 new = bio_alloc(GFP_NOFS, BIO_MAX_PAGES);
518 xfs_init_bio_from_bh(new, bh);
520 bio_chain(ioend->io_bio, new);
521 bio_get(ioend->io_bio); /* for xfs_destroy_ioend */
522 submit_bio(wbc->sync_mode == WB_SYNC_ALL ? WRITE_SYNC : WRITE,
528 * Test to see if we've been building up a completion structure for
529 * earlier buffers -- if so, we try to append to this ioend if we
530 * can, otherwise we finish off any current ioend and start another.
531 * Return the ioend we finished off so that the caller can submit it
532 * once it has finished processing the dirty page.
537 struct buffer_head *bh,
539 struct xfs_writepage_ctx *wpc,
540 struct writeback_control *wbc,
541 struct list_head *iolist)
543 if (!wpc->ioend || wpc->io_type != wpc->ioend->io_type ||
544 bh->b_blocknr != wpc->last_block + 1 ||
545 offset != wpc->ioend->io_offset + wpc->ioend->io_size) {
547 list_add(&wpc->ioend->io_list, iolist);
548 wpc->ioend = xfs_alloc_ioend(inode, wpc->io_type, offset, bh);
552 * If the buffer doesn't fit into the bio we need to allocate a new
553 * one. This shouldn't happen more than once for a given buffer.
555 while (xfs_bio_add_buffer(wpc->ioend->io_bio, bh) != bh->b_size)
556 xfs_chain_bio(wpc->ioend, wbc, bh);
558 wpc->ioend->io_size += bh->b_size;
559 wpc->last_block = bh->b_blocknr;
560 xfs_start_buffer_writeback(bh);
566 struct buffer_head *bh,
567 struct xfs_bmbt_irec *imap,
571 struct xfs_mount *m = XFS_I(inode)->i_mount;
572 xfs_off_t iomap_offset = XFS_FSB_TO_B(m, imap->br_startoff);
573 xfs_daddr_t iomap_bn = xfs_fsb_to_db(XFS_I(inode), imap->br_startblock);
575 ASSERT(imap->br_startblock != HOLESTARTBLOCK);
576 ASSERT(imap->br_startblock != DELAYSTARTBLOCK);
578 bn = (iomap_bn >> (inode->i_blkbits - BBSHIFT)) +
579 ((offset - iomap_offset) >> inode->i_blkbits);
581 ASSERT(bn || XFS_IS_REALTIME_INODE(XFS_I(inode)));
584 set_buffer_mapped(bh);
590 struct buffer_head *bh,
591 struct xfs_bmbt_irec *imap,
594 ASSERT(imap->br_startblock != HOLESTARTBLOCK);
595 ASSERT(imap->br_startblock != DELAYSTARTBLOCK);
597 xfs_map_buffer(inode, bh, imap, offset);
598 set_buffer_mapped(bh);
599 clear_buffer_delay(bh);
600 clear_buffer_unwritten(bh);
604 * Test if a given page contains at least one buffer of a given @type.
605 * If @check_all_buffers is true, then we walk all the buffers in the page to
606 * try to find one of the type passed in. If it is not set, then the caller only
607 * needs to check the first buffer on the page for a match.
613 bool check_all_buffers)
615 struct buffer_head *bh;
616 struct buffer_head *head;
618 if (PageWriteback(page))
622 if (!page_has_buffers(page))
625 bh = head = page_buffers(page);
627 if (buffer_unwritten(bh)) {
628 if (type == XFS_IO_UNWRITTEN)
630 } else if (buffer_delay(bh)) {
631 if (type == XFS_IO_DELALLOC)
633 } else if (buffer_dirty(bh) && buffer_mapped(bh)) {
634 if (type == XFS_IO_OVERWRITE)
638 /* If we are only checking the first buffer, we are done now. */
639 if (!check_all_buffers)
641 } while ((bh = bh->b_this_page) != head);
647 xfs_vm_invalidatepage(
652 trace_xfs_invalidatepage(page->mapping->host, page, offset,
654 block_invalidatepage(page, offset, length);
658 * If the page has delalloc buffers on it, we need to punch them out before we
659 * invalidate the page. If we don't, we leave a stale delalloc mapping on the
660 * inode that can trip a BUG() in xfs_get_blocks() later on if a direct IO read
661 * is done on that same region - the delalloc extent is returned when none is
662 * supposed to be there.
664 * We prevent this by truncating away the delalloc regions on the page before
665 * invalidating it. Because they are delalloc, we can do this without needing a
666 * transaction. Indeed - if we get ENOSPC errors, we have to be able to do this
667 * truncation without a transaction as there is no space left for block
668 * reservation (typically why we see a ENOSPC in writeback).
670 * This is not a performance critical path, so for now just do the punching a
671 * buffer head at a time.
674 xfs_aops_discard_page(
677 struct inode *inode = page->mapping->host;
678 struct xfs_inode *ip = XFS_I(inode);
679 struct buffer_head *bh, *head;
680 loff_t offset = page_offset(page);
682 if (!xfs_check_page_type(page, XFS_IO_DELALLOC, true))
685 if (XFS_FORCED_SHUTDOWN(ip->i_mount))
688 xfs_alert(ip->i_mount,
689 "page discard on page %p, inode 0x%llx, offset %llu.",
690 page, ip->i_ino, offset);
692 xfs_ilock(ip, XFS_ILOCK_EXCL);
693 bh = head = page_buffers(page);
696 xfs_fileoff_t start_fsb;
698 if (!buffer_delay(bh))
701 start_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
702 error = xfs_bmap_punch_delalloc_range(ip, start_fsb, 1);
704 /* something screwed, just bail */
705 if (!XFS_FORCED_SHUTDOWN(ip->i_mount)) {
706 xfs_alert(ip->i_mount,
707 "page discard unable to remove delalloc mapping.");
712 offset += 1 << inode->i_blkbits;
714 } while ((bh = bh->b_this_page) != head);
716 xfs_iunlock(ip, XFS_ILOCK_EXCL);
718 xfs_vm_invalidatepage(page, 0, PAGE_SIZE);
723 * We implement an immediate ioend submission policy here to avoid needing to
724 * chain multiple ioends and hence nest mempool allocations which can violate
725 * forward progress guarantees we need to provide. The current ioend we are
726 * adding buffers to is cached on the writepage context, and if the new buffer
727 * does not append to the cached ioend it will create a new ioend and cache that
730 * If a new ioend is created and cached, the old ioend is returned and queued
731 * locally for submission once the entire page is processed or an error has been
732 * detected. While ioends are submitted immediately after they are completed,
733 * batching optimisations are provided by higher level block plugging.
735 * At the end of a writeback pass, there will be a cached ioend remaining on the
736 * writepage context that the caller will need to submit.
740 struct xfs_writepage_ctx *wpc,
741 struct writeback_control *wbc,
745 __uint64_t end_offset)
747 LIST_HEAD(submit_list);
748 struct xfs_ioend *ioend, *next;
749 struct buffer_head *bh, *head;
750 ssize_t len = 1 << inode->i_blkbits;
755 bh = head = page_buffers(page);
756 offset = page_offset(page);
758 if (offset >= end_offset)
760 if (!buffer_uptodate(bh))
764 * set_page_dirty dirties all buffers in a page, independent
765 * of their state. The dirty state however is entirely
766 * meaningless for holes (!mapped && uptodate), so skip
767 * buffers covering holes here.
769 if (!buffer_mapped(bh) && buffer_uptodate(bh)) {
770 wpc->imap_valid = false;
774 if (buffer_unwritten(bh)) {
775 if (wpc->io_type != XFS_IO_UNWRITTEN) {
776 wpc->io_type = XFS_IO_UNWRITTEN;
777 wpc->imap_valid = false;
779 } else if (buffer_delay(bh)) {
780 if (wpc->io_type != XFS_IO_DELALLOC) {
781 wpc->io_type = XFS_IO_DELALLOC;
782 wpc->imap_valid = false;
784 } else if (buffer_uptodate(bh)) {
785 if (wpc->io_type != XFS_IO_OVERWRITE) {
786 wpc->io_type = XFS_IO_OVERWRITE;
787 wpc->imap_valid = false;
790 if (PageUptodate(page))
791 ASSERT(buffer_mapped(bh));
793 * This buffer is not uptodate and will not be
794 * written to disk. Ensure that we will put any
795 * subsequent writeable buffers into a new
798 wpc->imap_valid = false;
803 wpc->imap_valid = xfs_imap_valid(inode, &wpc->imap,
805 if (!wpc->imap_valid) {
806 error = xfs_map_blocks(inode, offset, &wpc->imap,
810 wpc->imap_valid = xfs_imap_valid(inode, &wpc->imap,
813 if (wpc->imap_valid) {
815 if (wpc->io_type != XFS_IO_OVERWRITE)
816 xfs_map_at_offset(inode, bh, &wpc->imap, offset);
817 xfs_add_to_ioend(inode, bh, offset, wpc, wbc, &submit_list);
821 } while (offset += len, ((bh = bh->b_this_page) != head));
823 if (uptodate && bh == head)
824 SetPageUptodate(page);
826 ASSERT(wpc->ioend || list_empty(&submit_list));
830 * On error, we have to fail the ioend here because we have locked
831 * buffers in the ioend. If we don't do this, we'll deadlock
832 * invalidating the page as that tries to lock the buffers on the page.
833 * Also, because we may have set pages under writeback, we have to make
834 * sure we run IO completion to mark the error state of the IO
835 * appropriately, so we can't cancel the ioend directly here. That means
836 * we have to mark this page as under writeback if we included any
837 * buffers from it in the ioend chain so that completion treats it
840 * If we didn't include the page in the ioend, the on error we can
841 * simply discard and unlock it as there are no other users of the page
842 * or it's buffers right now. The caller will still need to trigger
843 * submission of outstanding ioends on the writepage context so they are
844 * treated correctly on error.
847 xfs_start_page_writeback(page, !error);
850 * Preserve the original error if there was one, otherwise catch
851 * submission errors here and propagate into subsequent ioend
854 list_for_each_entry_safe(ioend, next, &submit_list, io_list) {
857 list_del_init(&ioend->io_list);
858 error2 = xfs_submit_ioend(wbc, ioend, error);
859 if (error2 && !error)
863 xfs_aops_discard_page(page);
864 ClearPageUptodate(page);
868 * We can end up here with no error and nothing to write if we
869 * race with a partial page truncate on a sub-page block sized
870 * filesystem. In that case we need to mark the page clean.
872 xfs_start_page_writeback(page, 1);
873 end_page_writeback(page);
876 mapping_set_error(page->mapping, error);
881 * Write out a dirty page.
883 * For delalloc space on the page we need to allocate space and flush it.
884 * For unwritten space on the page we need to start the conversion to
885 * regular allocated space.
886 * For any other dirty buffer heads on the page we should flush them.
891 struct writeback_control *wbc,
894 struct xfs_writepage_ctx *wpc = data;
895 struct inode *inode = page->mapping->host;
897 __uint64_t end_offset;
900 trace_xfs_writepage(inode, page, 0, 0);
902 ASSERT(page_has_buffers(page));
905 * Refuse to write the page out if we are called from reclaim context.
907 * This avoids stack overflows when called from deeply used stacks in
908 * random callers for direct reclaim or memcg reclaim. We explicitly
909 * allow reclaim from kswapd as the stack usage there is relatively low.
911 * This should never happen except in the case of a VM regression so
914 if (WARN_ON_ONCE((current->flags & (PF_MEMALLOC|PF_KSWAPD)) ==
919 * Given that we do not allow direct reclaim to call us, we should
920 * never be called while in a filesystem transaction.
922 if (WARN_ON_ONCE(current->flags & PF_FSTRANS))
926 * Is this page beyond the end of the file?
928 * The page index is less than the end_index, adjust the end_offset
929 * to the highest offset that this page should represent.
930 * -----------------------------------------------------
931 * | file mapping | <EOF> |
932 * -----------------------------------------------------
933 * | Page ... | Page N-2 | Page N-1 | Page N | |
934 * ^--------------------------------^----------|--------
935 * | desired writeback range | see else |
936 * ---------------------------------^------------------|
938 offset = i_size_read(inode);
939 end_index = offset >> PAGE_SHIFT;
940 if (page->index < end_index)
941 end_offset = (xfs_off_t)(page->index + 1) << PAGE_SHIFT;
944 * Check whether the page to write out is beyond or straddles
946 * -------------------------------------------------------
947 * | file mapping | <EOF> |
948 * -------------------------------------------------------
949 * | Page ... | Page N-2 | Page N-1 | Page N | Beyond |
950 * ^--------------------------------^-----------|---------
952 * ---------------------------------^-----------|--------|
954 unsigned offset_into_page = offset & (PAGE_SIZE - 1);
957 * Skip the page if it is fully outside i_size, e.g. due to a
958 * truncate operation that is in progress. We must redirty the
959 * page so that reclaim stops reclaiming it. Otherwise
960 * xfs_vm_releasepage() is called on it and gets confused.
962 * Note that the end_index is unsigned long, it would overflow
963 * if the given offset is greater than 16TB on 32-bit system
964 * and if we do check the page is fully outside i_size or not
965 * via "if (page->index >= end_index + 1)" as "end_index + 1"
966 * will be evaluated to 0. Hence this page will be redirtied
967 * and be written out repeatedly which would result in an
968 * infinite loop, the user program that perform this operation
969 * will hang. Instead, we can verify this situation by checking
970 * if the page to write is totally beyond the i_size or if it's
971 * offset is just equal to the EOF.
973 if (page->index > end_index ||
974 (page->index == end_index && offset_into_page == 0))
978 * The page straddles i_size. It must be zeroed out on each
979 * and every writepage invocation because it may be mmapped.
980 * "A file is mapped in multiples of the page size. For a file
981 * that is not a multiple of the page size, the remaining
982 * memory is zeroed when mapped, and writes to that region are
983 * not written out to the file."
985 zero_user_segment(page, offset_into_page, PAGE_SIZE);
987 /* Adjust the end_offset to the end of file */
991 return xfs_writepage_map(wpc, wbc, inode, page, offset, end_offset);
994 redirty_page_for_writepage(wbc, page);
1002 struct writeback_control *wbc)
1004 struct xfs_writepage_ctx wpc = {
1005 .io_type = XFS_IO_INVALID,
1009 ret = xfs_do_writepage(page, wbc, &wpc);
1011 ret = xfs_submit_ioend(wbc, wpc.ioend, ret);
1017 struct address_space *mapping,
1018 struct writeback_control *wbc)
1020 struct xfs_writepage_ctx wpc = {
1021 .io_type = XFS_IO_INVALID,
1025 xfs_iflags_clear(XFS_I(mapping->host), XFS_ITRUNCATED);
1026 if (dax_mapping(mapping))
1027 return dax_writeback_mapping_range(mapping,
1028 xfs_find_bdev_for_inode(mapping->host), wbc);
1030 ret = write_cache_pages(mapping, wbc, xfs_do_writepage, &wpc);
1032 ret = xfs_submit_ioend(wbc, wpc.ioend, ret);
1037 * Called to move a page into cleanable state - and from there
1038 * to be released. The page should already be clean. We always
1039 * have buffer heads in this call.
1041 * Returns 1 if the page is ok to release, 0 otherwise.
1048 int delalloc, unwritten;
1050 trace_xfs_releasepage(page->mapping->host, page, 0, 0);
1053 * mm accommodates an old ext3 case where clean pages might not have had
1054 * the dirty bit cleared. Thus, it can send actual dirty pages to
1055 * ->releasepage() via shrink_active_list(). Conversely,
1056 * block_invalidatepage() can send pages that are still marked dirty
1057 * but otherwise have invalidated buffers.
1059 * We've historically freed buffers on the latter. Instead, quietly
1060 * filter out all dirty pages to avoid spurious buffer state warnings.
1061 * This can likely be removed once shrink_active_list() is fixed.
1063 if (PageDirty(page))
1066 xfs_count_page_state(page, &delalloc, &unwritten);
1068 if (WARN_ON_ONCE(delalloc))
1070 if (WARN_ON_ONCE(unwritten))
1073 return try_to_free_buffers(page);
1077 * When we map a DIO buffer, we may need to pass flags to
1078 * xfs_end_io_direct_write to tell it what kind of write IO we are doing.
1080 * Note that for DIO, an IO to the highest supported file block offset (i.e.
1081 * 2^63 - 1FSB bytes) will result in the offset + count overflowing a signed 64
1082 * bit variable. Hence if we see this overflow, we have to assume that the IO is
1083 * extending the file size. We won't know for sure until IO completion is run
1084 * and the actual max write offset is communicated to the IO completion
1089 struct inode *inode,
1090 struct buffer_head *bh_result,
1091 struct xfs_bmbt_irec *imap,
1094 uintptr_t *flags = (uintptr_t *)&bh_result->b_private;
1095 xfs_off_t size = bh_result->b_size;
1097 trace_xfs_get_blocks_map_direct(XFS_I(inode), offset, size,
1098 ISUNWRITTEN(imap) ? XFS_IO_UNWRITTEN : XFS_IO_OVERWRITE, imap);
1100 if (ISUNWRITTEN(imap)) {
1101 *flags |= XFS_DIO_FLAG_UNWRITTEN;
1102 set_buffer_defer_completion(bh_result);
1103 } else if (offset + size > i_size_read(inode) || offset + size < 0) {
1104 *flags |= XFS_DIO_FLAG_APPEND;
1105 set_buffer_defer_completion(bh_result);
1110 * If this is O_DIRECT or the mpage code calling tell them how large the mapping
1111 * is, so that we can avoid repeated get_blocks calls.
1113 * If the mapping spans EOF, then we have to break the mapping up as the mapping
1114 * for blocks beyond EOF must be marked new so that sub block regions can be
1115 * correctly zeroed. We can't do this for mappings within EOF unless the mapping
1116 * was just allocated or is unwritten, otherwise the callers would overwrite
1117 * existing data with zeros. Hence we have to split the mapping into a range up
1118 * to and including EOF, and a second mapping for beyond EOF.
1122 struct inode *inode,
1124 struct buffer_head *bh_result,
1125 struct xfs_bmbt_irec *imap,
1129 xfs_off_t mapping_size;
1131 mapping_size = imap->br_startoff + imap->br_blockcount - iblock;
1132 mapping_size <<= inode->i_blkbits;
1134 ASSERT(mapping_size > 0);
1135 if (mapping_size > size)
1136 mapping_size = size;
1137 if (offset < i_size_read(inode) &&
1138 offset + mapping_size >= i_size_read(inode)) {
1139 /* limit mapping to block that spans EOF */
1140 mapping_size = roundup_64(i_size_read(inode) - offset,
1141 1 << inode->i_blkbits);
1143 if (mapping_size > LONG_MAX)
1144 mapping_size = LONG_MAX;
1146 bh_result->b_size = mapping_size;
1151 struct inode *inode,
1153 struct buffer_head *bh_result,
1158 struct xfs_inode *ip = XFS_I(inode);
1159 struct xfs_mount *mp = ip->i_mount;
1160 xfs_fileoff_t offset_fsb, end_fsb;
1163 struct xfs_bmbt_irec imap;
1169 BUG_ON(create && !direct);
1171 if (XFS_FORCED_SHUTDOWN(mp))
1174 offset = (xfs_off_t)iblock << inode->i_blkbits;
1175 ASSERT(bh_result->b_size >= (1 << inode->i_blkbits));
1176 size = bh_result->b_size;
1178 if (!create && offset >= i_size_read(inode))
1182 * Direct I/O is usually done on preallocated files, so try getting
1183 * a block mapping without an exclusive lock first.
1185 lockmode = xfs_ilock_data_map_shared(ip);
1187 ASSERT(offset <= mp->m_super->s_maxbytes);
1188 if (offset + size > mp->m_super->s_maxbytes)
1189 size = mp->m_super->s_maxbytes - offset;
1190 end_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)offset + size);
1191 offset_fsb = XFS_B_TO_FSBT(mp, offset);
1193 error = xfs_bmapi_read(ip, offset_fsb, end_fsb - offset_fsb,
1194 &imap, &nimaps, XFS_BMAPI_ENTIRE);
1198 /* for DAX, we convert unwritten extents directly */
1201 (imap.br_startblock == HOLESTARTBLOCK ||
1202 imap.br_startblock == DELAYSTARTBLOCK) ||
1203 (IS_DAX(inode) && ISUNWRITTEN(&imap)))) {
1205 * xfs_iomap_write_direct() expects the shared lock. It
1206 * is unlocked on return.
1208 if (lockmode == XFS_ILOCK_EXCL)
1209 xfs_ilock_demote(ip, lockmode);
1211 error = xfs_iomap_write_direct(ip, offset, size,
1217 trace_xfs_get_blocks_alloc(ip, offset, size,
1218 ISUNWRITTEN(&imap) ? XFS_IO_UNWRITTEN
1219 : XFS_IO_DELALLOC, &imap);
1220 } else if (nimaps) {
1221 trace_xfs_get_blocks_found(ip, offset, size,
1222 ISUNWRITTEN(&imap) ? XFS_IO_UNWRITTEN
1223 : XFS_IO_OVERWRITE, &imap);
1224 xfs_iunlock(ip, lockmode);
1226 trace_xfs_get_blocks_notfound(ip, offset, size);
1230 if (IS_DAX(inode) && create) {
1231 ASSERT(!ISUNWRITTEN(&imap));
1232 /* zeroing is not needed at a higher layer */
1236 /* trim mapping down to size requested */
1237 xfs_map_trim_size(inode, iblock, bh_result, &imap, offset, size);
1240 * For unwritten extents do not report a disk address in the buffered
1241 * read case (treat as if we're reading into a hole).
1243 if (imap.br_startblock != HOLESTARTBLOCK &&
1244 imap.br_startblock != DELAYSTARTBLOCK &&
1245 (create || !ISUNWRITTEN(&imap))) {
1246 xfs_map_buffer(inode, bh_result, &imap, offset);
1247 if (ISUNWRITTEN(&imap))
1248 set_buffer_unwritten(bh_result);
1249 /* direct IO needs special help */
1252 ASSERT(!ISUNWRITTEN(&imap));
1254 xfs_map_direct(inode, bh_result, &imap, offset);
1259 * If this is a realtime file, data may be on a different device.
1260 * to that pointed to from the buffer_head b_bdev currently.
1262 bh_result->b_bdev = xfs_find_bdev_for_inode(inode);
1265 * If we previously allocated a block out beyond eof and we are now
1266 * coming back to use it then we will need to flag it as new even if it
1267 * has a disk address.
1269 * With sub-block writes into unwritten extents we also need to mark
1270 * the buffer as new so that the unwritten parts of the buffer gets
1274 ((!buffer_mapped(bh_result) && !buffer_uptodate(bh_result)) ||
1275 (offset >= i_size_read(inode)) ||
1276 (new || ISUNWRITTEN(&imap))))
1277 set_buffer_new(bh_result);
1279 BUG_ON(direct && imap.br_startblock == DELAYSTARTBLOCK);
1284 xfs_iunlock(ip, lockmode);
1290 struct inode *inode,
1292 struct buffer_head *bh_result,
1295 return __xfs_get_blocks(inode, iblock, bh_result, create, false, false);
1299 xfs_get_blocks_direct(
1300 struct inode *inode,
1302 struct buffer_head *bh_result,
1305 return __xfs_get_blocks(inode, iblock, bh_result, create, true, false);
1309 xfs_get_blocks_dax_fault(
1310 struct inode *inode,
1312 struct buffer_head *bh_result,
1315 return __xfs_get_blocks(inode, iblock, bh_result, create, true, true);
1319 * Complete a direct I/O write request.
1321 * xfs_map_direct passes us some flags in the private data to tell us what to
1322 * do. If no flags are set, then the write IO is an overwrite wholly within
1323 * the existing allocated file size and so there is nothing for us to do.
1325 * Note that in this case the completion can be called in interrupt context,
1326 * whereas if we have flags set we will always be called in task context
1327 * (i.e. from a workqueue).
1330 xfs_end_io_direct_write(
1336 struct inode *inode = file_inode(iocb->ki_filp);
1337 struct xfs_inode *ip = XFS_I(inode);
1338 struct xfs_mount *mp = ip->i_mount;
1339 uintptr_t flags = (uintptr_t)private;
1342 trace_xfs_end_io_direct_write(ip, offset, size);
1344 if (XFS_FORCED_SHUTDOWN(mp))
1351 * The flags tell us whether we are doing unwritten extent conversions
1352 * or an append transaction that updates the on-disk file size. These
1353 * cases are the only cases where we should *potentially* be needing
1354 * to update the VFS inode size.
1357 ASSERT(offset + size <= i_size_read(inode));
1362 * We need to update the in-core inode size here so that we don't end up
1363 * with the on-disk inode size being outside the in-core inode size. We
1364 * have no other method of updating EOF for AIO, so always do it here
1367 * We need to lock the test/set EOF update as we can be racing with
1368 * other IO completions here to update the EOF. Failing to serialise
1369 * here can result in EOF moving backwards and Bad Things Happen when
1372 spin_lock(&ip->i_flags_lock);
1373 if (offset + size > i_size_read(inode))
1374 i_size_write(inode, offset + size);
1375 spin_unlock(&ip->i_flags_lock);
1377 if (flags & XFS_DIO_FLAG_UNWRITTEN) {
1378 trace_xfs_end_io_direct_write_unwritten(ip, offset, size);
1380 error = xfs_iomap_write_unwritten(ip, offset, size);
1381 } else if (flags & XFS_DIO_FLAG_APPEND) {
1382 struct xfs_trans *tp;
1384 trace_xfs_end_io_direct_write_append(ip, offset, size);
1386 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_fsyncts, 0, 0, 0,
1389 error = xfs_setfilesize(ip, tp, offset, size);
1398 struct iov_iter *iter)
1401 * We just need the method present so that open/fcntl allow direct I/O.
1408 struct address_space *mapping,
1411 struct inode *inode = (struct inode *)mapping->host;
1412 struct xfs_inode *ip = XFS_I(inode);
1414 trace_xfs_vm_bmap(XFS_I(inode));
1415 xfs_ilock(ip, XFS_IOLOCK_SHARED);
1416 filemap_write_and_wait(mapping);
1417 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
1418 return generic_block_bmap(mapping, block, xfs_get_blocks);
1423 struct file *unused,
1426 trace_xfs_vm_readpage(page->mapping->host, 1);
1427 return mpage_readpage(page, xfs_get_blocks);
1432 struct file *unused,
1433 struct address_space *mapping,
1434 struct list_head *pages,
1437 trace_xfs_vm_readpages(mapping->host, nr_pages);
1438 return mpage_readpages(mapping, pages, nr_pages, xfs_get_blocks);
1442 * This is basically a copy of __set_page_dirty_buffers() with one
1443 * small tweak: buffers beyond EOF do not get marked dirty. If we mark them
1444 * dirty, we'll never be able to clean them because we don't write buffers
1445 * beyond EOF, and that means we can't invalidate pages that span EOF
1446 * that have been marked dirty. Further, the dirty state can leak into
1447 * the file interior if the file is extended, resulting in all sorts of
1448 * bad things happening as the state does not match the underlying data.
1450 * XXX: this really indicates that bufferheads in XFS need to die. Warts like
1451 * this only exist because of bufferheads and how the generic code manages them.
1454 xfs_vm_set_page_dirty(
1457 struct address_space *mapping = page->mapping;
1458 struct inode *inode = mapping->host;
1463 if (unlikely(!mapping))
1464 return !TestSetPageDirty(page);
1466 end_offset = i_size_read(inode);
1467 offset = page_offset(page);
1469 spin_lock(&mapping->private_lock);
1470 if (page_has_buffers(page)) {
1471 struct buffer_head *head = page_buffers(page);
1472 struct buffer_head *bh = head;
1475 if (offset < end_offset)
1476 set_buffer_dirty(bh);
1477 bh = bh->b_this_page;
1478 offset += 1 << inode->i_blkbits;
1479 } while (bh != head);
1482 * Lock out page->mem_cgroup migration to keep PageDirty
1483 * synchronized with per-memcg dirty page counters.
1485 lock_page_memcg(page);
1486 newly_dirty = !TestSetPageDirty(page);
1487 spin_unlock(&mapping->private_lock);
1490 /* sigh - __set_page_dirty() is static, so copy it here, too */
1491 unsigned long flags;
1493 spin_lock_irqsave(&mapping->tree_lock, flags);
1494 if (page->mapping) { /* Race with truncate? */
1495 WARN_ON_ONCE(!PageUptodate(page));
1496 account_page_dirtied(page, mapping);
1497 radix_tree_tag_set(&mapping->page_tree,
1498 page_index(page), PAGECACHE_TAG_DIRTY);
1500 spin_unlock_irqrestore(&mapping->tree_lock, flags);
1502 unlock_page_memcg(page);
1504 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1508 const struct address_space_operations xfs_address_space_operations = {
1509 .readpage = xfs_vm_readpage,
1510 .readpages = xfs_vm_readpages,
1511 .writepage = xfs_vm_writepage,
1512 .writepages = xfs_vm_writepages,
1513 .set_page_dirty = xfs_vm_set_page_dirty,
1514 .releasepage = xfs_vm_releasepage,
1515 .invalidatepage = xfs_vm_invalidatepage,
1516 .bmap = xfs_vm_bmap,
1517 .direct_IO = xfs_vm_direct_IO,
1518 .migratepage = buffer_migrate_page,
1519 .is_partially_uptodate = block_is_partially_uptodate,
1520 .error_remove_page = generic_error_remove_page,