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
22 #include "xfs_trans.h"
23 #include "xfs_mount.h"
24 #include "xfs_bmap_btree.h"
25 #include "xfs_dinode.h"
26 #include "xfs_inode.h"
27 #include "xfs_inode_item.h"
28 #include "xfs_alloc.h"
29 #include "xfs_error.h"
30 #include "xfs_iomap.h"
31 #include "xfs_vnodeops.h"
32 #include "xfs_trace.h"
34 #include <linux/aio.h>
35 #include <linux/gfp.h>
36 #include <linux/mpage.h>
37 #include <linux/pagevec.h>
38 #include <linux/writeback.h>
46 struct buffer_head *bh, *head;
48 *delalloc = *unwritten = 0;
50 bh = head = page_buffers(page);
52 if (buffer_unwritten(bh))
54 else if (buffer_delay(bh))
56 } while ((bh = bh->b_this_page) != head);
59 STATIC struct block_device *
60 xfs_find_bdev_for_inode(
63 struct xfs_inode *ip = XFS_I(inode);
64 struct xfs_mount *mp = ip->i_mount;
66 if (XFS_IS_REALTIME_INODE(ip))
67 return mp->m_rtdev_targp->bt_bdev;
69 return mp->m_ddev_targp->bt_bdev;
73 * We're now finished for good with this ioend structure.
74 * Update the page state via the associated buffer_heads,
75 * release holds on the inode and bio, and finally free
76 * up memory. Do not use the ioend after this.
82 struct buffer_head *bh, *next;
84 for (bh = ioend->io_buffer_head; bh; bh = next) {
86 bh->b_end_io(bh, !ioend->io_error);
90 inode_dio_done(ioend->io_inode);
91 if (ioend->io_isasync) {
92 aio_complete(ioend->io_iocb, ioend->io_error ?
93 ioend->io_error : ioend->io_result, 0);
97 mempool_free(ioend, xfs_ioend_pool);
101 * Fast and loose check if this write could update the on-disk inode size.
103 static inline bool xfs_ioend_is_append(struct xfs_ioend *ioend)
105 return ioend->io_offset + ioend->io_size >
106 XFS_I(ioend->io_inode)->i_d.di_size;
110 xfs_setfilesize_trans_alloc(
111 struct xfs_ioend *ioend)
113 struct xfs_mount *mp = XFS_I(ioend->io_inode)->i_mount;
114 struct xfs_trans *tp;
117 tp = xfs_trans_alloc(mp, XFS_TRANS_FSYNC_TS);
119 error = xfs_trans_reserve(tp, 0, XFS_FSYNC_TS_LOG_RES(mp), 0, 0, 0);
121 xfs_trans_cancel(tp, 0);
125 ioend->io_append_trans = tp;
128 * We may pass freeze protection with a transaction. So tell lockdep
131 rwsem_release(&ioend->io_inode->i_sb->s_writers.lock_map[SB_FREEZE_FS-1],
134 * We hand off the transaction to the completion thread now, so
135 * clear the flag here.
137 current_restore_flags_nested(&tp->t_pflags, PF_FSTRANS);
142 * Update on-disk file size now that data has been written to disk.
146 struct xfs_ioend *ioend)
148 struct xfs_inode *ip = XFS_I(ioend->io_inode);
149 struct xfs_trans *tp = ioend->io_append_trans;
153 * The transaction may have been allocated in the I/O submission thread,
154 * thus we need to mark ourselves as beeing in a transaction manually.
155 * Similarly for freeze protection.
157 current_set_flags_nested(&tp->t_pflags, PF_FSTRANS);
158 rwsem_acquire_read(&VFS_I(ip)->i_sb->s_writers.lock_map[SB_FREEZE_FS-1],
161 xfs_ilock(ip, XFS_ILOCK_EXCL);
162 isize = xfs_new_eof(ip, ioend->io_offset + ioend->io_size);
164 xfs_iunlock(ip, XFS_ILOCK_EXCL);
165 xfs_trans_cancel(tp, 0);
169 trace_xfs_setfilesize(ip, ioend->io_offset, ioend->io_size);
171 ip->i_d.di_size = isize;
172 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
173 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
175 return xfs_trans_commit(tp, 0);
179 * Schedule IO completion handling on the final put of an ioend.
181 * If there is no work to do we might as well call it a day and free the
186 struct xfs_ioend *ioend)
188 if (atomic_dec_and_test(&ioend->io_remaining)) {
189 struct xfs_mount *mp = XFS_I(ioend->io_inode)->i_mount;
191 if (ioend->io_type == XFS_IO_UNWRITTEN)
192 queue_work(mp->m_unwritten_workqueue, &ioend->io_work);
193 else if (ioend->io_append_trans ||
194 (ioend->io_isdirect && xfs_ioend_is_append(ioend)))
195 queue_work(mp->m_data_workqueue, &ioend->io_work);
197 xfs_destroy_ioend(ioend);
202 * IO write completion.
206 struct work_struct *work)
208 xfs_ioend_t *ioend = container_of(work, xfs_ioend_t, io_work);
209 struct xfs_inode *ip = XFS_I(ioend->io_inode);
212 if (XFS_FORCED_SHUTDOWN(ip->i_mount)) {
213 ioend->io_error = -EIO;
220 * For unwritten extents we need to issue transactions to convert a
221 * range to normal written extens after the data I/O has finished.
223 if (ioend->io_type == XFS_IO_UNWRITTEN) {
224 error = xfs_iomap_write_unwritten(ip, ioend->io_offset,
226 } else if (ioend->io_isdirect && xfs_ioend_is_append(ioend)) {
228 * For direct I/O we do not know if we need to allocate blocks
229 * or not so we can't preallocate an append transaction as that
230 * results in nested reservations and log space deadlocks. Hence
231 * allocate the transaction here. While this is sub-optimal and
232 * can block IO completion for some time, we're stuck with doing
233 * it this way until we can pass the ioend to the direct IO
234 * allocation callbacks and avoid nesting that way.
236 error = xfs_setfilesize_trans_alloc(ioend);
239 error = xfs_setfilesize(ioend);
240 } else if (ioend->io_append_trans) {
241 error = xfs_setfilesize(ioend);
243 ASSERT(!xfs_ioend_is_append(ioend));
248 ioend->io_error = -error;
249 xfs_destroy_ioend(ioend);
253 * Call IO completion handling in caller context on the final put of an ioend.
256 xfs_finish_ioend_sync(
257 struct xfs_ioend *ioend)
259 if (atomic_dec_and_test(&ioend->io_remaining))
260 xfs_end_io(&ioend->io_work);
264 * Allocate and initialise an IO completion structure.
265 * We need to track unwritten extent write completion here initially.
266 * We'll need to extend this for updating the ondisk inode size later
276 ioend = mempool_alloc(xfs_ioend_pool, GFP_NOFS);
279 * Set the count to 1 initially, which will prevent an I/O
280 * completion callback from happening before we have started
281 * all the I/O from calling the completion routine too early.
283 atomic_set(&ioend->io_remaining, 1);
284 ioend->io_isasync = 0;
285 ioend->io_isdirect = 0;
287 ioend->io_list = NULL;
288 ioend->io_type = type;
289 ioend->io_inode = inode;
290 ioend->io_buffer_head = NULL;
291 ioend->io_buffer_tail = NULL;
292 ioend->io_offset = 0;
294 ioend->io_iocb = NULL;
295 ioend->io_result = 0;
296 ioend->io_append_trans = NULL;
298 INIT_WORK(&ioend->io_work, xfs_end_io);
306 struct xfs_bmbt_irec *imap,
310 struct xfs_inode *ip = XFS_I(inode);
311 struct xfs_mount *mp = ip->i_mount;
312 ssize_t count = 1 << inode->i_blkbits;
313 xfs_fileoff_t offset_fsb, end_fsb;
315 int bmapi_flags = XFS_BMAPI_ENTIRE;
318 if (XFS_FORCED_SHUTDOWN(mp))
319 return -XFS_ERROR(EIO);
321 if (type == XFS_IO_UNWRITTEN)
322 bmapi_flags |= XFS_BMAPI_IGSTATE;
324 if (!xfs_ilock_nowait(ip, XFS_ILOCK_SHARED)) {
326 return -XFS_ERROR(EAGAIN);
327 xfs_ilock(ip, XFS_ILOCK_SHARED);
330 ASSERT(ip->i_d.di_format != XFS_DINODE_FMT_BTREE ||
331 (ip->i_df.if_flags & XFS_IFEXTENTS));
332 ASSERT(offset <= mp->m_super->s_maxbytes);
334 if (offset + count > mp->m_super->s_maxbytes)
335 count = mp->m_super->s_maxbytes - offset;
336 end_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)offset + count);
337 offset_fsb = XFS_B_TO_FSBT(mp, offset);
338 error = xfs_bmapi_read(ip, offset_fsb, end_fsb - offset_fsb,
339 imap, &nimaps, bmapi_flags);
340 xfs_iunlock(ip, XFS_ILOCK_SHARED);
343 return -XFS_ERROR(error);
345 if (type == XFS_IO_DELALLOC &&
346 (!nimaps || isnullstartblock(imap->br_startblock))) {
347 error = xfs_iomap_write_allocate(ip, offset, count, imap);
349 trace_xfs_map_blocks_alloc(ip, offset, count, type, imap);
350 return -XFS_ERROR(error);
354 if (type == XFS_IO_UNWRITTEN) {
356 ASSERT(imap->br_startblock != HOLESTARTBLOCK);
357 ASSERT(imap->br_startblock != DELAYSTARTBLOCK);
361 trace_xfs_map_blocks_found(ip, offset, count, type, imap);
368 struct xfs_bmbt_irec *imap,
371 offset >>= inode->i_blkbits;
373 return offset >= imap->br_startoff &&
374 offset < imap->br_startoff + imap->br_blockcount;
378 * BIO completion handler for buffered IO.
385 xfs_ioend_t *ioend = bio->bi_private;
387 ASSERT(atomic_read(&bio->bi_cnt) >= 1);
388 ioend->io_error = test_bit(BIO_UPTODATE, &bio->bi_flags) ? 0 : error;
390 /* Toss bio and pass work off to an xfsdatad thread */
391 bio->bi_private = NULL;
392 bio->bi_end_io = NULL;
395 xfs_finish_ioend(ioend);
399 xfs_submit_ioend_bio(
400 struct writeback_control *wbc,
404 atomic_inc(&ioend->io_remaining);
405 bio->bi_private = ioend;
406 bio->bi_end_io = xfs_end_bio;
407 submit_bio(wbc->sync_mode == WB_SYNC_ALL ? WRITE_SYNC : WRITE, bio);
412 struct buffer_head *bh)
414 int nvecs = bio_get_nr_vecs(bh->b_bdev);
415 struct bio *bio = bio_alloc(GFP_NOIO, nvecs);
417 ASSERT(bio->bi_private == NULL);
418 bio->bi_sector = bh->b_blocknr * (bh->b_size >> 9);
419 bio->bi_bdev = bh->b_bdev;
424 xfs_start_buffer_writeback(
425 struct buffer_head *bh)
427 ASSERT(buffer_mapped(bh));
428 ASSERT(buffer_locked(bh));
429 ASSERT(!buffer_delay(bh));
430 ASSERT(!buffer_unwritten(bh));
432 mark_buffer_async_write(bh);
433 set_buffer_uptodate(bh);
434 clear_buffer_dirty(bh);
438 xfs_start_page_writeback(
443 ASSERT(PageLocked(page));
444 ASSERT(!PageWriteback(page));
446 clear_page_dirty_for_io(page);
447 set_page_writeback(page);
449 /* If no buffers on the page are to be written, finish it here */
451 end_page_writeback(page);
454 static inline int bio_add_buffer(struct bio *bio, struct buffer_head *bh)
456 return bio_add_page(bio, bh->b_page, bh->b_size, bh_offset(bh));
460 * Submit all of the bios for all of the ioends we have saved up, covering the
461 * initial writepage page and also any probed pages.
463 * Because we may have multiple ioends spanning a page, we need to start
464 * writeback on all the buffers before we submit them for I/O. If we mark the
465 * buffers as we got, then we can end up with a page that only has buffers
466 * marked async write and I/O complete on can occur before we mark the other
467 * buffers async write.
469 * The end result of this is that we trip a bug in end_page_writeback() because
470 * we call it twice for the one page as the code in end_buffer_async_write()
471 * assumes that all buffers on the page are started at the same time.
473 * The fix is two passes across the ioend list - one to start writeback on the
474 * buffer_heads, and then submit them for I/O on the second pass.
476 * If @fail is non-zero, it means that we have a situation where some part of
477 * the submission process has failed after we have marked paged for writeback
478 * and unlocked them. In this situation, we need to fail the ioend chain rather
479 * than submit it to IO. This typically only happens on a filesystem shutdown.
483 struct writeback_control *wbc,
487 xfs_ioend_t *head = ioend;
489 struct buffer_head *bh;
491 sector_t lastblock = 0;
493 /* Pass 1 - start writeback */
495 next = ioend->io_list;
496 for (bh = ioend->io_buffer_head; bh; bh = bh->b_private)
497 xfs_start_buffer_writeback(bh);
498 } while ((ioend = next) != NULL);
500 /* Pass 2 - submit I/O */
503 next = ioend->io_list;
507 * If we are failing the IO now, just mark the ioend with an
508 * error and finish it. This will run IO completion immediately
509 * as there is only one reference to the ioend at this point in
513 ioend->io_error = -fail;
514 xfs_finish_ioend(ioend);
518 for (bh = ioend->io_buffer_head; bh; bh = bh->b_private) {
522 bio = xfs_alloc_ioend_bio(bh);
523 } else if (bh->b_blocknr != lastblock + 1) {
524 xfs_submit_ioend_bio(wbc, ioend, bio);
528 if (bio_add_buffer(bio, bh) != bh->b_size) {
529 xfs_submit_ioend_bio(wbc, ioend, bio);
533 lastblock = bh->b_blocknr;
536 xfs_submit_ioend_bio(wbc, ioend, bio);
537 xfs_finish_ioend(ioend);
538 } while ((ioend = next) != NULL);
542 * Cancel submission of all buffer_heads so far in this endio.
543 * Toss the endio too. Only ever called for the initial page
544 * in a writepage request, so only ever one page.
551 struct buffer_head *bh, *next_bh;
554 next = ioend->io_list;
555 bh = ioend->io_buffer_head;
557 next_bh = bh->b_private;
558 clear_buffer_async_write(bh);
560 } while ((bh = next_bh) != NULL);
562 mempool_free(ioend, xfs_ioend_pool);
563 } while ((ioend = next) != NULL);
567 * Test to see if we've been building up a completion structure for
568 * earlier buffers -- if so, we try to append to this ioend if we
569 * can, otherwise we finish off any current ioend and start another.
570 * Return true if we've finished the given ioend.
575 struct buffer_head *bh,
578 xfs_ioend_t **result,
581 xfs_ioend_t *ioend = *result;
583 if (!ioend || need_ioend || type != ioend->io_type) {
584 xfs_ioend_t *previous = *result;
586 ioend = xfs_alloc_ioend(inode, type);
587 ioend->io_offset = offset;
588 ioend->io_buffer_head = bh;
589 ioend->io_buffer_tail = bh;
591 previous->io_list = ioend;
594 ioend->io_buffer_tail->b_private = bh;
595 ioend->io_buffer_tail = bh;
598 bh->b_private = NULL;
599 ioend->io_size += bh->b_size;
605 struct buffer_head *bh,
606 struct xfs_bmbt_irec *imap,
610 struct xfs_mount *m = XFS_I(inode)->i_mount;
611 xfs_off_t iomap_offset = XFS_FSB_TO_B(m, imap->br_startoff);
612 xfs_daddr_t iomap_bn = xfs_fsb_to_db(XFS_I(inode), imap->br_startblock);
614 ASSERT(imap->br_startblock != HOLESTARTBLOCK);
615 ASSERT(imap->br_startblock != DELAYSTARTBLOCK);
617 bn = (iomap_bn >> (inode->i_blkbits - BBSHIFT)) +
618 ((offset - iomap_offset) >> inode->i_blkbits);
620 ASSERT(bn || XFS_IS_REALTIME_INODE(XFS_I(inode)));
623 set_buffer_mapped(bh);
629 struct buffer_head *bh,
630 struct xfs_bmbt_irec *imap,
633 ASSERT(imap->br_startblock != HOLESTARTBLOCK);
634 ASSERT(imap->br_startblock != DELAYSTARTBLOCK);
636 xfs_map_buffer(inode, bh, imap, offset);
637 set_buffer_mapped(bh);
638 clear_buffer_delay(bh);
639 clear_buffer_unwritten(bh);
643 * Test if a given page is suitable for writing as part of an unwritten
644 * or delayed allocate extent.
651 if (PageWriteback(page))
654 if (page->mapping && page_has_buffers(page)) {
655 struct buffer_head *bh, *head;
658 bh = head = page_buffers(page);
660 if (buffer_unwritten(bh))
661 acceptable += (type == XFS_IO_UNWRITTEN);
662 else if (buffer_delay(bh))
663 acceptable += (type == XFS_IO_DELALLOC);
664 else if (buffer_dirty(bh) && buffer_mapped(bh))
665 acceptable += (type == XFS_IO_OVERWRITE);
668 } while ((bh = bh->b_this_page) != head);
678 * Allocate & map buffers for page given the extent map. Write it out.
679 * except for the original page of a writepage, this is called on
680 * delalloc/unwritten pages only, for the original page it is possible
681 * that the page has no mapping at all.
688 struct xfs_bmbt_irec *imap,
689 xfs_ioend_t **ioendp,
690 struct writeback_control *wbc)
692 struct buffer_head *bh, *head;
693 xfs_off_t end_offset;
694 unsigned long p_offset;
697 int count = 0, done = 0, uptodate = 1;
698 xfs_off_t offset = page_offset(page);
700 if (page->index != tindex)
702 if (!trylock_page(page))
704 if (PageWriteback(page))
705 goto fail_unlock_page;
706 if (page->mapping != inode->i_mapping)
707 goto fail_unlock_page;
708 if (!xfs_check_page_type(page, (*ioendp)->io_type))
709 goto fail_unlock_page;
712 * page_dirty is initially a count of buffers on the page before
713 * EOF and is decremented as we move each into a cleanable state.
717 * End offset is the highest offset that this page should represent.
718 * If we are on the last page, (end_offset & (PAGE_CACHE_SIZE - 1))
719 * will evaluate non-zero and be less than PAGE_CACHE_SIZE and
720 * hence give us the correct page_dirty count. On any other page,
721 * it will be zero and in that case we need page_dirty to be the
722 * count of buffers on the page.
724 end_offset = min_t(unsigned long long,
725 (xfs_off_t)(page->index + 1) << PAGE_CACHE_SHIFT,
728 len = 1 << inode->i_blkbits;
729 p_offset = min_t(unsigned long, end_offset & (PAGE_CACHE_SIZE - 1),
731 p_offset = p_offset ? roundup(p_offset, len) : PAGE_CACHE_SIZE;
732 page_dirty = p_offset / len;
734 bh = head = page_buffers(page);
736 if (offset >= end_offset)
738 if (!buffer_uptodate(bh))
740 if (!(PageUptodate(page) || buffer_uptodate(bh))) {
745 if (buffer_unwritten(bh) || buffer_delay(bh) ||
747 if (buffer_unwritten(bh))
748 type = XFS_IO_UNWRITTEN;
749 else if (buffer_delay(bh))
750 type = XFS_IO_DELALLOC;
752 type = XFS_IO_OVERWRITE;
754 if (!xfs_imap_valid(inode, imap, offset)) {
760 if (type != XFS_IO_OVERWRITE)
761 xfs_map_at_offset(inode, bh, imap, offset);
762 xfs_add_to_ioend(inode, bh, offset, type,
770 } while (offset += len, (bh = bh->b_this_page) != head);
772 if (uptodate && bh == head)
773 SetPageUptodate(page);
776 if (--wbc->nr_to_write <= 0 &&
777 wbc->sync_mode == WB_SYNC_NONE)
780 xfs_start_page_writeback(page, !page_dirty, count);
790 * Convert & write out a cluster of pages in the same extent as defined
791 * by mp and following the start page.
797 struct xfs_bmbt_irec *imap,
798 xfs_ioend_t **ioendp,
799 struct writeback_control *wbc,
805 pagevec_init(&pvec, 0);
806 while (!done && tindex <= tlast) {
807 unsigned len = min_t(pgoff_t, PAGEVEC_SIZE, tlast - tindex + 1);
809 if (!pagevec_lookup(&pvec, inode->i_mapping, tindex, len))
812 for (i = 0; i < pagevec_count(&pvec); i++) {
813 done = xfs_convert_page(inode, pvec.pages[i], tindex++,
819 pagevec_release(&pvec);
825 xfs_vm_invalidatepage(
827 unsigned long offset)
829 trace_xfs_invalidatepage(page->mapping->host, page, offset);
830 block_invalidatepage(page, offset);
834 * If the page has delalloc buffers on it, we need to punch them out before we
835 * invalidate the page. If we don't, we leave a stale delalloc mapping on the
836 * inode that can trip a BUG() in xfs_get_blocks() later on if a direct IO read
837 * is done on that same region - the delalloc extent is returned when none is
838 * supposed to be there.
840 * We prevent this by truncating away the delalloc regions on the page before
841 * invalidating it. Because they are delalloc, we can do this without needing a
842 * transaction. Indeed - if we get ENOSPC errors, we have to be able to do this
843 * truncation without a transaction as there is no space left for block
844 * reservation (typically why we see a ENOSPC in writeback).
846 * This is not a performance critical path, so for now just do the punching a
847 * buffer head at a time.
850 xfs_aops_discard_page(
853 struct inode *inode = page->mapping->host;
854 struct xfs_inode *ip = XFS_I(inode);
855 struct buffer_head *bh, *head;
856 loff_t offset = page_offset(page);
858 if (!xfs_check_page_type(page, XFS_IO_DELALLOC))
861 if (XFS_FORCED_SHUTDOWN(ip->i_mount))
864 xfs_alert(ip->i_mount,
865 "page discard on page %p, inode 0x%llx, offset %llu.",
866 page, ip->i_ino, offset);
868 xfs_ilock(ip, XFS_ILOCK_EXCL);
869 bh = head = page_buffers(page);
872 xfs_fileoff_t start_fsb;
874 if (!buffer_delay(bh))
877 start_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
878 error = xfs_bmap_punch_delalloc_range(ip, start_fsb, 1);
880 /* something screwed, just bail */
881 if (!XFS_FORCED_SHUTDOWN(ip->i_mount)) {
882 xfs_alert(ip->i_mount,
883 "page discard unable to remove delalloc mapping.");
888 offset += 1 << inode->i_blkbits;
890 } while ((bh = bh->b_this_page) != head);
892 xfs_iunlock(ip, XFS_ILOCK_EXCL);
894 xfs_vm_invalidatepage(page, 0);
899 * Write out a dirty page.
901 * For delalloc space on the page we need to allocate space and flush it.
902 * For unwritten space on the page we need to start the conversion to
903 * regular allocated space.
904 * For any other dirty buffer heads on the page we should flush them.
909 struct writeback_control *wbc)
911 struct inode *inode = page->mapping->host;
912 struct buffer_head *bh, *head;
913 struct xfs_bmbt_irec imap;
914 xfs_ioend_t *ioend = NULL, *iohead = NULL;
917 __uint64_t end_offset;
918 pgoff_t end_index, last_index;
920 int err, imap_valid = 0, uptodate = 1;
924 trace_xfs_writepage(inode, page, 0);
926 ASSERT(page_has_buffers(page));
929 * Refuse to write the page out if we are called from reclaim context.
931 * This avoids stack overflows when called from deeply used stacks in
932 * random callers for direct reclaim or memcg reclaim. We explicitly
933 * allow reclaim from kswapd as the stack usage there is relatively low.
935 * This should never happen except in the case of a VM regression so
938 if (WARN_ON_ONCE((current->flags & (PF_MEMALLOC|PF_KSWAPD)) ==
943 * Given that we do not allow direct reclaim to call us, we should
944 * never be called while in a filesystem transaction.
946 if (WARN_ON(current->flags & PF_FSTRANS))
949 /* Is this page beyond the end of the file? */
950 offset = i_size_read(inode);
951 end_index = offset >> PAGE_CACHE_SHIFT;
952 last_index = (offset - 1) >> PAGE_CACHE_SHIFT;
953 if (page->index >= end_index) {
954 unsigned offset_into_page = offset & (PAGE_CACHE_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 if (page->index >= end_index + 1 || offset_into_page == 0)
966 * The page straddles i_size. It must be zeroed out on each
967 * and every writepage invocation because it may be mmapped.
968 * "A file is mapped in multiples of the page size. For a file
969 * that is not a multiple of the page size, the remaining
970 * memory is zeroed when mapped, and writes to that region are
971 * not written out to the file."
973 zero_user_segment(page, offset_into_page, PAGE_CACHE_SIZE);
976 end_offset = min_t(unsigned long long,
977 (xfs_off_t)(page->index + 1) << PAGE_CACHE_SHIFT,
979 len = 1 << inode->i_blkbits;
981 bh = head = page_buffers(page);
982 offset = page_offset(page);
983 type = XFS_IO_OVERWRITE;
985 if (wbc->sync_mode == WB_SYNC_NONE)
991 if (offset >= end_offset)
993 if (!buffer_uptodate(bh))
997 * set_page_dirty dirties all buffers in a page, independent
998 * of their state. The dirty state however is entirely
999 * meaningless for holes (!mapped && uptodate), so skip
1000 * buffers covering holes here.
1002 if (!buffer_mapped(bh) && buffer_uptodate(bh)) {
1007 if (buffer_unwritten(bh)) {
1008 if (type != XFS_IO_UNWRITTEN) {
1009 type = XFS_IO_UNWRITTEN;
1012 } else if (buffer_delay(bh)) {
1013 if (type != XFS_IO_DELALLOC) {
1014 type = XFS_IO_DELALLOC;
1017 } else if (buffer_uptodate(bh)) {
1018 if (type != XFS_IO_OVERWRITE) {
1019 type = XFS_IO_OVERWRITE;
1023 if (PageUptodate(page))
1024 ASSERT(buffer_mapped(bh));
1026 * This buffer is not uptodate and will not be
1027 * written to disk. Ensure that we will put any
1028 * subsequent writeable buffers into a new
1036 imap_valid = xfs_imap_valid(inode, &imap, offset);
1039 * If we didn't have a valid mapping then we need to
1040 * put the new mapping into a separate ioend structure.
1041 * This ensures non-contiguous extents always have
1042 * separate ioends, which is particularly important
1043 * for unwritten extent conversion at I/O completion
1047 err = xfs_map_blocks(inode, offset, &imap, type,
1051 imap_valid = xfs_imap_valid(inode, &imap, offset);
1055 if (type != XFS_IO_OVERWRITE)
1056 xfs_map_at_offset(inode, bh, &imap, offset);
1057 xfs_add_to_ioend(inode, bh, offset, type, &ioend,
1065 } while (offset += len, ((bh = bh->b_this_page) != head));
1067 if (uptodate && bh == head)
1068 SetPageUptodate(page);
1070 xfs_start_page_writeback(page, 1, count);
1072 /* if there is no IO to be submitted for this page, we are done */
1079 * Any errors from this point onwards need tobe reported through the IO
1080 * completion path as we have marked the initial page as under writeback
1084 xfs_off_t end_index;
1086 end_index = imap.br_startoff + imap.br_blockcount;
1089 end_index <<= inode->i_blkbits;
1092 end_index = (end_index - 1) >> PAGE_CACHE_SHIFT;
1094 /* check against file size */
1095 if (end_index > last_index)
1096 end_index = last_index;
1098 xfs_cluster_write(inode, page->index + 1, &imap, &ioend,
1104 * Reserve log space if we might write beyond the on-disk inode size.
1107 if (ioend->io_type != XFS_IO_UNWRITTEN && xfs_ioend_is_append(ioend))
1108 err = xfs_setfilesize_trans_alloc(ioend);
1110 xfs_submit_ioend(wbc, iohead, err);
1116 xfs_cancel_ioend(iohead);
1121 xfs_aops_discard_page(page);
1122 ClearPageUptodate(page);
1127 redirty_page_for_writepage(wbc, page);
1134 struct address_space *mapping,
1135 struct writeback_control *wbc)
1137 xfs_iflags_clear(XFS_I(mapping->host), XFS_ITRUNCATED);
1138 return generic_writepages(mapping, wbc);
1142 * Called to move a page into cleanable state - and from there
1143 * to be released. The page should already be clean. We always
1144 * have buffer heads in this call.
1146 * Returns 1 if the page is ok to release, 0 otherwise.
1153 int delalloc, unwritten;
1155 trace_xfs_releasepage(page->mapping->host, page, 0);
1157 xfs_count_page_state(page, &delalloc, &unwritten);
1159 if (WARN_ON(delalloc))
1161 if (WARN_ON(unwritten))
1164 return try_to_free_buffers(page);
1169 struct inode *inode,
1171 struct buffer_head *bh_result,
1175 struct xfs_inode *ip = XFS_I(inode);
1176 struct xfs_mount *mp = ip->i_mount;
1177 xfs_fileoff_t offset_fsb, end_fsb;
1180 struct xfs_bmbt_irec imap;
1186 if (XFS_FORCED_SHUTDOWN(mp))
1187 return -XFS_ERROR(EIO);
1189 offset = (xfs_off_t)iblock << inode->i_blkbits;
1190 ASSERT(bh_result->b_size >= (1 << inode->i_blkbits));
1191 size = bh_result->b_size;
1193 if (!create && direct && offset >= i_size_read(inode))
1197 * Direct I/O is usually done on preallocated files, so try getting
1198 * a block mapping without an exclusive lock first. For buffered
1199 * writes we already have the exclusive iolock anyway, so avoiding
1200 * a lock roundtrip here by taking the ilock exclusive from the
1201 * beginning is a useful micro optimization.
1203 if (create && !direct) {
1204 lockmode = XFS_ILOCK_EXCL;
1205 xfs_ilock(ip, lockmode);
1207 lockmode = xfs_ilock_map_shared(ip);
1210 ASSERT(offset <= mp->m_super->s_maxbytes);
1211 if (offset + size > mp->m_super->s_maxbytes)
1212 size = mp->m_super->s_maxbytes - offset;
1213 end_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)offset + size);
1214 offset_fsb = XFS_B_TO_FSBT(mp, offset);
1216 error = xfs_bmapi_read(ip, offset_fsb, end_fsb - offset_fsb,
1217 &imap, &nimaps, XFS_BMAPI_ENTIRE);
1223 (imap.br_startblock == HOLESTARTBLOCK ||
1224 imap.br_startblock == DELAYSTARTBLOCK))) {
1225 if (direct || xfs_get_extsz_hint(ip)) {
1227 * Drop the ilock in preparation for starting the block
1228 * allocation transaction. It will be retaken
1229 * exclusively inside xfs_iomap_write_direct for the
1230 * actual allocation.
1232 xfs_iunlock(ip, lockmode);
1233 error = xfs_iomap_write_direct(ip, offset, size,
1240 * Delalloc reservations do not require a transaction,
1241 * we can go on without dropping the lock here. If we
1242 * are allocating a new delalloc block, make sure that
1243 * we set the new flag so that we mark the buffer new so
1244 * that we know that it is newly allocated if the write
1247 if (nimaps && imap.br_startblock == HOLESTARTBLOCK)
1249 error = xfs_iomap_write_delay(ip, offset, size, &imap);
1253 xfs_iunlock(ip, lockmode);
1256 trace_xfs_get_blocks_alloc(ip, offset, size, 0, &imap);
1257 } else if (nimaps) {
1258 trace_xfs_get_blocks_found(ip, offset, size, 0, &imap);
1259 xfs_iunlock(ip, lockmode);
1261 trace_xfs_get_blocks_notfound(ip, offset, size);
1265 if (imap.br_startblock != HOLESTARTBLOCK &&
1266 imap.br_startblock != DELAYSTARTBLOCK) {
1268 * For unwritten extents do not report a disk address on
1269 * the read case (treat as if we're reading into a hole).
1271 if (create || !ISUNWRITTEN(&imap))
1272 xfs_map_buffer(inode, bh_result, &imap, offset);
1273 if (create && ISUNWRITTEN(&imap)) {
1275 bh_result->b_private = inode;
1276 set_buffer_unwritten(bh_result);
1281 * If this is a realtime file, data may be on a different device.
1282 * to that pointed to from the buffer_head b_bdev currently.
1284 bh_result->b_bdev = xfs_find_bdev_for_inode(inode);
1287 * If we previously allocated a block out beyond eof and we are now
1288 * coming back to use it then we will need to flag it as new even if it
1289 * has a disk address.
1291 * With sub-block writes into unwritten extents we also need to mark
1292 * the buffer as new so that the unwritten parts of the buffer gets
1296 ((!buffer_mapped(bh_result) && !buffer_uptodate(bh_result)) ||
1297 (offset >= i_size_read(inode)) ||
1298 (new || ISUNWRITTEN(&imap))))
1299 set_buffer_new(bh_result);
1301 if (imap.br_startblock == DELAYSTARTBLOCK) {
1304 set_buffer_uptodate(bh_result);
1305 set_buffer_mapped(bh_result);
1306 set_buffer_delay(bh_result);
1311 * If this is O_DIRECT or the mpage code calling tell them how large
1312 * the mapping is, so that we can avoid repeated get_blocks calls.
1314 if (direct || size > (1 << inode->i_blkbits)) {
1315 xfs_off_t mapping_size;
1317 mapping_size = imap.br_startoff + imap.br_blockcount - iblock;
1318 mapping_size <<= inode->i_blkbits;
1320 ASSERT(mapping_size > 0);
1321 if (mapping_size > size)
1322 mapping_size = size;
1323 if (mapping_size > LONG_MAX)
1324 mapping_size = LONG_MAX;
1326 bh_result->b_size = mapping_size;
1332 xfs_iunlock(ip, lockmode);
1338 struct inode *inode,
1340 struct buffer_head *bh_result,
1343 return __xfs_get_blocks(inode, iblock, bh_result, create, 0);
1347 xfs_get_blocks_direct(
1348 struct inode *inode,
1350 struct buffer_head *bh_result,
1353 return __xfs_get_blocks(inode, iblock, bh_result, create, 1);
1357 * Complete a direct I/O write request.
1359 * If the private argument is non-NULL __xfs_get_blocks signals us that we
1360 * need to issue a transaction to convert the range from unwritten to written
1361 * extents. In case this is regular synchronous I/O we just call xfs_end_io
1362 * to do this and we are done. But in case this was a successful AIO
1363 * request this handler is called from interrupt context, from which we
1364 * can't start transactions. In that case offload the I/O completion to
1365 * the workqueues we also use for buffered I/O completion.
1368 xfs_end_io_direct_write(
1376 struct xfs_ioend *ioend = iocb->private;
1379 * While the generic direct I/O code updates the inode size, it does
1380 * so only after the end_io handler is called, which means our
1381 * end_io handler thinks the on-disk size is outside the in-core
1382 * size. To prevent this just update it a little bit earlier here.
1384 if (offset + size > i_size_read(ioend->io_inode))
1385 i_size_write(ioend->io_inode, offset + size);
1388 * blockdev_direct_IO can return an error even after the I/O
1389 * completion handler was called. Thus we need to protect
1390 * against double-freeing.
1392 iocb->private = NULL;
1394 ioend->io_offset = offset;
1395 ioend->io_size = size;
1396 ioend->io_iocb = iocb;
1397 ioend->io_result = ret;
1398 if (private && size > 0)
1399 ioend->io_type = XFS_IO_UNWRITTEN;
1402 ioend->io_isasync = 1;
1403 xfs_finish_ioend(ioend);
1405 xfs_finish_ioend_sync(ioend);
1413 const struct iovec *iov,
1415 unsigned long nr_segs)
1417 struct inode *inode = iocb->ki_filp->f_mapping->host;
1418 struct block_device *bdev = xfs_find_bdev_for_inode(inode);
1419 struct xfs_ioend *ioend = NULL;
1423 size_t size = iov_length(iov, nr_segs);
1426 * We cannot preallocate a size update transaction here as we
1427 * don't know whether allocation is necessary or not. Hence we
1428 * can only tell IO completion that one is necessary if we are
1429 * not doing unwritten extent conversion.
1431 iocb->private = ioend = xfs_alloc_ioend(inode, XFS_IO_DIRECT);
1432 if (offset + size > XFS_I(inode)->i_d.di_size)
1433 ioend->io_isdirect = 1;
1435 ret = __blockdev_direct_IO(rw, iocb, inode, bdev, iov,
1437 xfs_get_blocks_direct,
1438 xfs_end_io_direct_write, NULL, 0);
1439 if (ret != -EIOCBQUEUED && iocb->private)
1440 goto out_destroy_ioend;
1442 ret = __blockdev_direct_IO(rw, iocb, inode, bdev, iov,
1444 xfs_get_blocks_direct,
1451 xfs_destroy_ioend(ioend);
1456 * Punch out the delalloc blocks we have already allocated.
1458 * Don't bother with xfs_setattr given that nothing can have made it to disk yet
1459 * as the page is still locked at this point.
1462 xfs_vm_kill_delalloc_range(
1463 struct inode *inode,
1467 struct xfs_inode *ip = XFS_I(inode);
1468 xfs_fileoff_t start_fsb;
1469 xfs_fileoff_t end_fsb;
1472 start_fsb = XFS_B_TO_FSB(ip->i_mount, start);
1473 end_fsb = XFS_B_TO_FSB(ip->i_mount, end);
1474 if (end_fsb <= start_fsb)
1477 xfs_ilock(ip, XFS_ILOCK_EXCL);
1478 error = xfs_bmap_punch_delalloc_range(ip, start_fsb,
1479 end_fsb - start_fsb);
1481 /* something screwed, just bail */
1482 if (!XFS_FORCED_SHUTDOWN(ip->i_mount)) {
1483 xfs_alert(ip->i_mount,
1484 "xfs_vm_write_failed: unable to clean up ino %lld",
1488 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1492 xfs_vm_write_failed(
1493 struct inode *inode,
1498 loff_t block_offset = pos & PAGE_MASK;
1501 loff_t from = pos & (PAGE_CACHE_SIZE - 1);
1502 loff_t to = from + len;
1503 struct buffer_head *bh, *head;
1505 ASSERT(block_offset + from == pos);
1507 head = page_buffers(page);
1509 for (bh = head; bh != head || !block_start;
1510 bh = bh->b_this_page, block_start = block_end,
1511 block_offset += bh->b_size) {
1512 block_end = block_start + bh->b_size;
1514 /* skip buffers before the write */
1515 if (block_end <= from)
1518 /* if the buffer is after the write, we're done */
1519 if (block_start >= to)
1522 if (!buffer_delay(bh))
1525 if (!buffer_new(bh) && block_offset < i_size_read(inode))
1528 xfs_vm_kill_delalloc_range(inode, block_offset,
1529 block_offset + bh->b_size);
1535 * This used to call block_write_begin(), but it unlocks and releases the page
1536 * on error, and we need that page to be able to punch stale delalloc blocks out
1537 * on failure. hence we copy-n-waste it here and call xfs_vm_write_failed() at
1538 * the appropriate point.
1543 struct address_space *mapping,
1547 struct page **pagep,
1550 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1554 ASSERT(len <= PAGE_CACHE_SIZE);
1556 page = grab_cache_page_write_begin(mapping, index,
1557 flags | AOP_FLAG_NOFS);
1561 status = __block_write_begin(page, pos, len, xfs_get_blocks);
1562 if (unlikely(status)) {
1563 struct inode *inode = mapping->host;
1565 xfs_vm_write_failed(inode, page, pos, len);
1568 if (pos + len > i_size_read(inode))
1569 truncate_pagecache(inode, pos + len, i_size_read(inode));
1571 page_cache_release(page);
1580 * On failure, we only need to kill delalloc blocks beyond EOF because they
1581 * will never be written. For blocks within EOF, generic_write_end() zeros them
1582 * so they are safe to leave alone and be written with all the other valid data.
1587 struct address_space *mapping,
1596 ASSERT(len <= PAGE_CACHE_SIZE);
1598 ret = generic_write_end(file, mapping, pos, len, copied, page, fsdata);
1599 if (unlikely(ret < len)) {
1600 struct inode *inode = mapping->host;
1601 size_t isize = i_size_read(inode);
1602 loff_t to = pos + len;
1605 truncate_pagecache(inode, to, isize);
1606 xfs_vm_kill_delalloc_range(inode, isize, to);
1614 struct address_space *mapping,
1617 struct inode *inode = (struct inode *)mapping->host;
1618 struct xfs_inode *ip = XFS_I(inode);
1620 trace_xfs_vm_bmap(XFS_I(inode));
1621 xfs_ilock(ip, XFS_IOLOCK_SHARED);
1622 filemap_write_and_wait(mapping);
1623 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
1624 return generic_block_bmap(mapping, block, xfs_get_blocks);
1629 struct file *unused,
1632 return mpage_readpage(page, xfs_get_blocks);
1637 struct file *unused,
1638 struct address_space *mapping,
1639 struct list_head *pages,
1642 return mpage_readpages(mapping, pages, nr_pages, xfs_get_blocks);
1645 const struct address_space_operations xfs_address_space_operations = {
1646 .readpage = xfs_vm_readpage,
1647 .readpages = xfs_vm_readpages,
1648 .writepage = xfs_vm_writepage,
1649 .writepages = xfs_vm_writepages,
1650 .releasepage = xfs_vm_releasepage,
1651 .invalidatepage = xfs_vm_invalidatepage,
1652 .write_begin = xfs_vm_write_begin,
1653 .write_end = xfs_vm_write_end,
1654 .bmap = xfs_vm_bmap,
1655 .direct_IO = xfs_vm_direct_IO,
1656 .migratepage = buffer_migrate_page,
1657 .is_partially_uptodate = block_is_partially_uptodate,
1658 .error_remove_page = generic_error_remove_page,