2 * Copyright (C) 2016 Oracle. All Rights Reserved.
4 * Author: Darrick J. Wong <darrick.wong@oracle.com>
6 * This program is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU General Public License
8 * as published by the Free Software Foundation; either version 2
9 * of the License, or (at your option) any later version.
11 * This program is distributed in the hope that it would be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 * GNU General Public License for more details.
16 * You should have received a copy of the GNU General Public License
17 * along with this program; if not, write the Free Software Foundation,
18 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA.
22 #include "xfs_shared.h"
23 #include "xfs_format.h"
24 #include "xfs_log_format.h"
25 #include "xfs_trans_resv.h"
26 #include "xfs_mount.h"
27 #include "xfs_defer.h"
28 #include "xfs_da_format.h"
29 #include "xfs_da_btree.h"
30 #include "xfs_inode.h"
31 #include "xfs_trans.h"
32 #include "xfs_inode_item.h"
34 #include "xfs_bmap_util.h"
35 #include "xfs_error.h"
37 #include "xfs_dir2_priv.h"
38 #include "xfs_ioctl.h"
39 #include "xfs_trace.h"
41 #include "xfs_icache.h"
43 #include "xfs_btree.h"
44 #include "xfs_refcount_btree.h"
45 #include "xfs_refcount.h"
46 #include "xfs_bmap_btree.h"
47 #include "xfs_trans_space.h"
49 #include "xfs_alloc.h"
50 #include "xfs_quota_defs.h"
51 #include "xfs_quota.h"
52 #include "xfs_btree.h"
53 #include "xfs_bmap_btree.h"
54 #include "xfs_reflink.h"
55 #include "xfs_iomap.h"
56 #include "xfs_rmap_btree.h"
58 #include "xfs_ag_resv.h"
61 * Copy on Write of Shared Blocks
63 * XFS must preserve "the usual" file semantics even when two files share
64 * the same physical blocks. This means that a write to one file must not
65 * alter the blocks in a different file; the way that we'll do that is
66 * through the use of a copy-on-write mechanism. At a high level, that
67 * means that when we want to write to a shared block, we allocate a new
68 * block, write the data to the new block, and if that succeeds we map the
69 * new block into the file.
71 * XFS provides a "delayed allocation" mechanism that defers the allocation
72 * of disk blocks to dirty-but-not-yet-mapped file blocks as long as
73 * possible. This reduces fragmentation by enabling the filesystem to ask
74 * for bigger chunks less often, which is exactly what we want for CoW.
76 * The delalloc mechanism begins when the kernel wants to make a block
77 * writable (write_begin or page_mkwrite). If the offset is not mapped, we
78 * create a delalloc mapping, which is a regular in-core extent, but without
79 * a real startblock. (For delalloc mappings, the startblock encodes both
80 * a flag that this is a delalloc mapping, and a worst-case estimate of how
81 * many blocks might be required to put the mapping into the BMBT.) delalloc
82 * mappings are a reservation against the free space in the filesystem;
83 * adjacent mappings can also be combined into fewer larger mappings.
85 * As an optimization, the CoW extent size hint (cowextsz) creates
86 * outsized aligned delalloc reservations in the hope of landing out of
87 * order nearby CoW writes in a single extent on disk, thereby reducing
88 * fragmentation and improving future performance.
90 * D: --RRRRRRSSSRRRRRRRR--- (data fork)
91 * C: ------DDDDDDD--------- (CoW fork)
93 * When dirty pages are being written out (typically in writepage), the
94 * delalloc reservations are converted into unwritten mappings by
95 * allocating blocks and replacing the delalloc mapping with real ones.
96 * A delalloc mapping can be replaced by several unwritten ones if the
97 * free space is fragmented.
99 * D: --RRRRRRSSSRRRRRRRR---
100 * C: ------UUUUUUU---------
102 * We want to adapt the delalloc mechanism for copy-on-write, since the
103 * write paths are similar. The first two steps (creating the reservation
104 * and allocating the blocks) are exactly the same as delalloc except that
105 * the mappings must be stored in a separate CoW fork because we do not want
106 * to disturb the mapping in the data fork until we're sure that the write
107 * succeeded. IO completion in this case is the process of removing the old
108 * mapping from the data fork and moving the new mapping from the CoW fork to
109 * the data fork. This will be discussed shortly.
111 * For now, unaligned directio writes will be bounced back to the page cache.
112 * Block-aligned directio writes will use the same mechanism as buffered
115 * Just prior to submitting the actual disk write requests, we convert
116 * the extents representing the range of the file actually being written
117 * (as opposed to extra pieces created for the cowextsize hint) to real
118 * extents. This will become important in the next step:
120 * D: --RRRRRRSSSRRRRRRRR---
121 * C: ------UUrrUUU---------
123 * CoW remapping must be done after the data block write completes,
124 * because we don't want to destroy the old data fork map until we're sure
125 * the new block has been written. Since the new mappings are kept in a
126 * separate fork, we can simply iterate these mappings to find the ones
127 * that cover the file blocks that we just CoW'd. For each extent, simply
128 * unmap the corresponding range in the data fork, map the new range into
129 * the data fork, and remove the extent from the CoW fork. Because of
130 * the presence of the cowextsize hint, however, we must be careful
131 * only to remap the blocks that we've actually written out -- we must
132 * never remap delalloc reservations nor CoW staging blocks that have
133 * yet to be written. This corresponds exactly to the real extents in
136 * D: --RRRRRRrrSRRRRRRRR---
137 * C: ------UU--UUU---------
139 * Since the remapping operation can be applied to an arbitrary file
140 * range, we record the need for the remap step as a flag in the ioend
141 * instead of declaring a new IO type. This is required for direct io
142 * because we only have ioend for the whole dio, and we have to be able to
143 * remember the presence of unwritten blocks and CoW blocks with a single
144 * ioend structure. Better yet, the more ground we can cover with one
149 * Given an AG extent, find the lowest-numbered run of shared blocks
150 * within that range and return the range in fbno/flen. If
151 * find_end_of_shared is true, return the longest contiguous extent of
152 * shared blocks. If there are no shared extents, fbno and flen will
153 * be set to NULLAGBLOCK and 0, respectively.
156 xfs_reflink_find_shared(
157 struct xfs_mount *mp,
158 struct xfs_trans *tp,
164 bool find_end_of_shared)
166 struct xfs_buf *agbp;
167 struct xfs_btree_cur *cur;
170 error = xfs_alloc_read_agf(mp, tp, agno, 0, &agbp);
176 cur = xfs_refcountbt_init_cursor(mp, tp, agbp, agno, NULL);
178 error = xfs_refcount_find_shared(cur, agbno, aglen, fbno, flen,
181 xfs_btree_del_cursor(cur, error ? XFS_BTREE_ERROR : XFS_BTREE_NOERROR);
183 xfs_trans_brelse(tp, agbp);
188 * Trim the mapping to the next block where there's a change in the
189 * shared/unshared status. More specifically, this means that we
190 * find the lowest-numbered extent of shared blocks that coincides with
191 * the given block mapping. If the shared extent overlaps the start of
192 * the mapping, trim the mapping to the end of the shared extent. If
193 * the shared region intersects the mapping, trim the mapping to the
194 * start of the shared extent. If there are no shared regions that
195 * overlap, just return the original extent.
198 xfs_reflink_trim_around_shared(
199 struct xfs_inode *ip,
200 struct xfs_bmbt_irec *irec,
211 /* Holes, unwritten, and delalloc extents cannot be shared */
212 if (!xfs_is_reflink_inode(ip) || !xfs_bmap_is_real_extent(irec)) {
217 trace_xfs_reflink_trim_around_shared(ip, irec);
219 agno = XFS_FSB_TO_AGNO(ip->i_mount, irec->br_startblock);
220 agbno = XFS_FSB_TO_AGBNO(ip->i_mount, irec->br_startblock);
221 aglen = irec->br_blockcount;
223 error = xfs_reflink_find_shared(ip->i_mount, NULL, agno, agbno,
224 aglen, &fbno, &flen, true);
228 *shared = *trimmed = false;
229 if (fbno == NULLAGBLOCK) {
230 /* No shared blocks at all. */
232 } else if (fbno == agbno) {
234 * The start of this extent is shared. Truncate the
235 * mapping at the end of the shared region so that a
236 * subsequent iteration starts at the start of the
239 irec->br_blockcount = flen;
246 * There's a shared extent midway through this extent.
247 * Truncate the mapping at the start of the shared
248 * extent so that a subsequent iteration starts at the
249 * start of the shared region.
251 irec->br_blockcount = fbno - agbno;
258 * Trim the passed in imap to the next shared/unshared extent boundary, and
259 * if imap->br_startoff points to a shared extent reserve space for it in the
260 * COW fork. In this case *shared is set to true, else to false.
262 * Note that imap will always contain the block numbers for the existing blocks
263 * in the data fork, as the upper layers need them for read-modify-write
267 xfs_reflink_reserve_cow(
268 struct xfs_inode *ip,
269 struct xfs_bmbt_irec *imap,
272 struct xfs_ifork *ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
273 struct xfs_bmbt_irec got;
275 bool eof = false, trimmed;
279 * Search the COW fork extent list first. This serves two purposes:
280 * first this implement the speculative preallocation using cowextisze,
281 * so that we also unshared block adjacent to shared blocks instead
282 * of just the shared blocks themselves. Second the lookup in the
283 * extent list is generally faster than going out to the shared extent
287 if (!xfs_iext_lookup_extent(ip, ifp, imap->br_startoff, &idx, &got))
289 if (!eof && got.br_startoff <= imap->br_startoff) {
290 trace_xfs_reflink_cow_found(ip, imap);
291 xfs_trim_extent(imap, got.br_startoff, got.br_blockcount);
297 /* Trim the mapping to the nearest shared extent boundary. */
298 error = xfs_reflink_trim_around_shared(ip, imap, shared, &trimmed);
302 /* Not shared? Just report the (potentially capped) extent. */
307 * Fork all the shared blocks from our write offset until the end of
310 error = xfs_qm_dqattach_locked(ip, 0);
314 error = xfs_bmapi_reserve_delalloc(ip, XFS_COW_FORK, imap->br_startoff,
315 imap->br_blockcount, 0, &got, &idx, eof);
316 if (error == -ENOSPC || error == -EDQUOT)
317 trace_xfs_reflink_cow_enospc(ip, imap);
321 trace_xfs_reflink_cow_alloc(ip, &got);
325 /* Convert part of an unwritten CoW extent to a real one. */
327 xfs_reflink_convert_cow_extent(
328 struct xfs_inode *ip,
329 struct xfs_bmbt_irec *imap,
330 xfs_fileoff_t offset_fsb,
331 xfs_filblks_t count_fsb,
332 struct xfs_defer_ops *dfops)
334 xfs_fsblock_t first_block = NULLFSBLOCK;
337 if (imap->br_state == XFS_EXT_NORM)
340 xfs_trim_extent(imap, offset_fsb, count_fsb);
341 trace_xfs_reflink_convert_cow(ip, imap);
342 if (imap->br_blockcount == 0)
344 return xfs_bmapi_write(NULL, ip, imap->br_startoff, imap->br_blockcount,
345 XFS_BMAPI_COWFORK | XFS_BMAPI_CONVERT, &first_block,
346 0, imap, &nimaps, dfops);
349 /* Convert all of the unwritten CoW extents in a file's range to real ones. */
351 xfs_reflink_convert_cow(
352 struct xfs_inode *ip,
356 struct xfs_bmbt_irec got;
357 struct xfs_defer_ops dfops;
358 struct xfs_mount *mp = ip->i_mount;
359 struct xfs_ifork *ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
360 xfs_fileoff_t offset_fsb = XFS_B_TO_FSBT(mp, offset);
361 xfs_fileoff_t end_fsb = XFS_B_TO_FSB(mp, offset + count);
366 xfs_ilock(ip, XFS_ILOCK_EXCL);
368 /* Convert all the extents to real from unwritten. */
369 for (found = xfs_iext_lookup_extent(ip, ifp, offset_fsb, &idx, &got);
370 found && got.br_startoff < end_fsb;
371 found = xfs_iext_get_extent(ifp, ++idx, &got)) {
372 error = xfs_reflink_convert_cow_extent(ip, &got, offset_fsb,
373 end_fsb - offset_fsb, &dfops);
379 xfs_iunlock(ip, XFS_ILOCK_EXCL);
383 /* Allocate all CoW reservations covering a range of blocks in a file. */
385 xfs_reflink_allocate_cow(
386 struct xfs_inode *ip,
387 struct xfs_bmbt_irec *imap,
391 struct xfs_mount *mp = ip->i_mount;
392 xfs_fileoff_t offset_fsb = imap->br_startoff;
393 xfs_filblks_t count_fsb = imap->br_blockcount;
394 struct xfs_bmbt_irec got;
395 struct xfs_defer_ops dfops;
396 struct xfs_trans *tp = NULL;
397 xfs_fsblock_t first_block;
398 int nimaps, error = 0;
400 xfs_filblks_t resaligned;
401 xfs_extlen_t resblks = 0;
405 ASSERT(xfs_is_reflink_inode(ip));
406 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL | XFS_ILOCK_SHARED));
409 * Even if the extent is not shared we might have a preallocation for
410 * it in the COW fork. If so use it.
412 if (xfs_iext_lookup_extent(ip, ip->i_cowfp, offset_fsb, &idx, &got) &&
413 got.br_startoff <= offset_fsb) {
416 /* If we have a real allocation in the COW fork we're done. */
417 if (!isnullstartblock(got.br_startblock)) {
418 xfs_trim_extent(&got, offset_fsb, count_fsb);
423 xfs_trim_extent(imap, got.br_startoff, got.br_blockcount);
425 error = xfs_reflink_trim_around_shared(ip, imap, shared, &trimmed);
426 if (error || !*shared)
431 resaligned = xfs_aligned_fsb_count(imap->br_startoff,
432 imap->br_blockcount, xfs_get_cowextsz_hint(ip));
433 resblks = XFS_DIOSTRAT_SPACE_RES(mp, resaligned);
435 xfs_iunlock(ip, *lockmode);
436 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, resblks, 0, 0, &tp);
437 *lockmode = XFS_ILOCK_EXCL;
438 xfs_ilock(ip, *lockmode);
443 error = xfs_qm_dqattach_locked(ip, 0);
449 error = xfs_trans_reserve_quota_nblks(tp, ip, resblks, 0,
450 XFS_QMOPT_RES_REGBLKS);
454 xfs_trans_ijoin(tp, ip, 0);
456 xfs_defer_init(&dfops, &first_block);
459 /* Allocate the entire reservation as unwritten blocks. */
460 error = xfs_bmapi_write(tp, ip, imap->br_startoff, imap->br_blockcount,
461 XFS_BMAPI_COWFORK | XFS_BMAPI_PREALLOC, &first_block,
462 resblks, imap, &nimaps, &dfops);
464 goto out_bmap_cancel;
467 error = xfs_defer_finish(&tp, &dfops);
469 goto out_bmap_cancel;
471 error = xfs_trans_commit(tp);
475 return xfs_reflink_convert_cow_extent(ip, imap, offset_fsb, count_fsb,
478 xfs_defer_cancel(&dfops);
479 xfs_trans_unreserve_quota_nblks(tp, ip, (long)resblks, 0,
480 XFS_QMOPT_RES_REGBLKS);
483 xfs_trans_cancel(tp);
488 * Find the CoW reservation for a given byte offset of a file.
491 xfs_reflink_find_cow_mapping(
492 struct xfs_inode *ip,
494 struct xfs_bmbt_irec *imap)
496 struct xfs_ifork *ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
497 xfs_fileoff_t offset_fsb;
498 struct xfs_bmbt_irec got;
501 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL | XFS_ILOCK_SHARED));
502 ASSERT(xfs_is_reflink_inode(ip));
504 offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
505 if (!xfs_iext_lookup_extent(ip, ifp, offset_fsb, &idx, &got))
507 if (got.br_startoff > offset_fsb)
510 trace_xfs_reflink_find_cow_mapping(ip, offset, 1, XFS_IO_OVERWRITE,
517 * Trim an extent to end at the next CoW reservation past offset_fsb.
520 xfs_reflink_trim_irec_to_next_cow(
521 struct xfs_inode *ip,
522 xfs_fileoff_t offset_fsb,
523 struct xfs_bmbt_irec *imap)
525 struct xfs_ifork *ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
526 struct xfs_bmbt_irec got;
529 if (!xfs_is_reflink_inode(ip))
532 /* Find the extent in the CoW fork. */
533 if (!xfs_iext_lookup_extent(ip, ifp, offset_fsb, &idx, &got))
536 /* This is the extent before; try sliding up one. */
537 if (got.br_startoff < offset_fsb) {
538 if (!xfs_iext_get_extent(ifp, idx + 1, &got))
542 if (got.br_startoff >= imap->br_startoff + imap->br_blockcount)
545 imap->br_blockcount = got.br_startoff - imap->br_startoff;
546 trace_xfs_reflink_trim_irec(ip, imap);
550 * Cancel CoW reservations for some block range of an inode.
552 * If cancel_real is true this function cancels all COW fork extents for the
553 * inode; if cancel_real is false, real extents are not cleared.
556 xfs_reflink_cancel_cow_blocks(
557 struct xfs_inode *ip,
558 struct xfs_trans **tpp,
559 xfs_fileoff_t offset_fsb,
560 xfs_fileoff_t end_fsb,
563 struct xfs_ifork *ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
564 struct xfs_bmbt_irec got, del;
566 xfs_fsblock_t firstfsb;
567 struct xfs_defer_ops dfops;
570 if (!xfs_is_reflink_inode(ip))
572 if (!xfs_iext_lookup_extent(ip, ifp, offset_fsb, &idx, &got))
575 while (got.br_startoff < end_fsb) {
577 xfs_trim_extent(&del, offset_fsb, end_fsb - offset_fsb);
578 trace_xfs_reflink_cancel_cow(ip, &del);
580 if (isnullstartblock(del.br_startblock)) {
581 error = xfs_bmap_del_extent_delay(ip, XFS_COW_FORK,
585 } else if (del.br_state == XFS_EXT_UNWRITTEN || cancel_real) {
586 xfs_trans_ijoin(*tpp, ip, 0);
587 xfs_defer_init(&dfops, &firstfsb);
589 /* Free the CoW orphan record. */
590 error = xfs_refcount_free_cow_extent(ip->i_mount,
591 &dfops, del.br_startblock,
596 xfs_bmap_add_free(ip->i_mount, &dfops,
597 del.br_startblock, del.br_blockcount,
600 /* Update quota accounting */
601 xfs_trans_mod_dquot_byino(*tpp, ip, XFS_TRANS_DQ_BCOUNT,
602 -(long)del.br_blockcount);
604 /* Roll the transaction */
605 xfs_defer_ijoin(&dfops, ip);
606 error = xfs_defer_finish(tpp, &dfops);
608 xfs_defer_cancel(&dfops);
612 /* Remove the mapping from the CoW fork. */
613 xfs_bmap_del_extent_cow(ip, &idx, &got, &del);
616 if (!xfs_iext_get_extent(ifp, ++idx, &got))
620 /* clear tag if cow fork is emptied */
622 xfs_inode_clear_cowblocks_tag(ip);
628 * Cancel CoW reservations for some byte range of an inode.
630 * If cancel_real is true this function cancels all COW fork extents for the
631 * inode; if cancel_real is false, real extents are not cleared.
634 xfs_reflink_cancel_cow_range(
635 struct xfs_inode *ip,
640 struct xfs_trans *tp;
641 xfs_fileoff_t offset_fsb;
642 xfs_fileoff_t end_fsb;
645 trace_xfs_reflink_cancel_cow_range(ip, offset, count);
646 ASSERT(xfs_is_reflink_inode(ip));
648 offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
649 if (count == NULLFILEOFF)
650 end_fsb = NULLFILEOFF;
652 end_fsb = XFS_B_TO_FSB(ip->i_mount, offset + count);
654 /* Start a rolling transaction to remove the mappings */
655 error = xfs_trans_alloc(ip->i_mount, &M_RES(ip->i_mount)->tr_write,
660 xfs_ilock(ip, XFS_ILOCK_EXCL);
661 xfs_trans_ijoin(tp, ip, 0);
663 /* Scrape out the old CoW reservations */
664 error = xfs_reflink_cancel_cow_blocks(ip, &tp, offset_fsb, end_fsb,
669 error = xfs_trans_commit(tp);
671 xfs_iunlock(ip, XFS_ILOCK_EXCL);
675 xfs_trans_cancel(tp);
676 xfs_iunlock(ip, XFS_ILOCK_EXCL);
678 trace_xfs_reflink_cancel_cow_range_error(ip, error, _RET_IP_);
683 * Remap parts of a file's data fork after a successful CoW.
687 struct xfs_inode *ip,
691 struct xfs_ifork *ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
692 struct xfs_bmbt_irec got, del;
693 struct xfs_trans *tp;
694 xfs_fileoff_t offset_fsb;
695 xfs_fileoff_t end_fsb;
696 xfs_fsblock_t firstfsb;
697 struct xfs_defer_ops dfops;
699 unsigned int resblks;
703 trace_xfs_reflink_end_cow(ip, offset, count);
705 /* No COW extents? That's easy! */
706 if (ifp->if_bytes == 0)
709 offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
710 end_fsb = XFS_B_TO_FSB(ip->i_mount, offset + count);
713 * Start a rolling transaction to switch the mappings. We're
714 * unlikely ever to have to remap 16T worth of single-block
715 * extents, so just cap the worst case extent count to 2^32-1.
716 * Stick a warning in just in case, and avoid 64-bit division.
718 BUILD_BUG_ON(MAX_RW_COUNT > UINT_MAX);
719 if (end_fsb - offset_fsb > UINT_MAX) {
720 error = -EFSCORRUPTED;
721 xfs_force_shutdown(ip->i_mount, SHUTDOWN_CORRUPT_INCORE);
725 resblks = XFS_NEXTENTADD_SPACE_RES(ip->i_mount,
726 (unsigned int)(end_fsb - offset_fsb),
728 error = xfs_trans_alloc(ip->i_mount, &M_RES(ip->i_mount)->tr_write,
733 xfs_ilock(ip, XFS_ILOCK_EXCL);
734 xfs_trans_ijoin(tp, ip, 0);
736 /* If there is a hole at end_fsb - 1 go to the previous extent */
737 if (!xfs_iext_lookup_extent(ip, ifp, end_fsb - 1, &idx, &got) ||
738 got.br_startoff > end_fsb) {
740 * In case of racing, overlapping AIO writes no COW extents
741 * might be left by the time I/O completes for the loser of
742 * the race. In that case we are done.
746 xfs_iext_get_extent(ifp, --idx, &got);
749 /* Walk backwards until we're out of the I/O range... */
750 while (got.br_startoff + got.br_blockcount > offset_fsb) {
752 xfs_trim_extent(&del, offset_fsb, end_fsb - offset_fsb);
754 /* Extent delete may have bumped idx forward */
755 if (!del.br_blockcount) {
760 ASSERT(!isnullstartblock(got.br_startblock));
763 * Don't remap unwritten extents; these are
764 * speculatively preallocated CoW extents that have been
765 * allocated but have not yet been involved in a write.
767 if (got.br_state == XFS_EXT_UNWRITTEN) {
772 /* Unmap the old blocks in the data fork. */
773 xfs_defer_init(&dfops, &firstfsb);
774 rlen = del.br_blockcount;
775 error = __xfs_bunmapi(tp, ip, del.br_startoff, &rlen, 0, 1,
780 /* Trim the extent to whatever got unmapped. */
782 xfs_trim_extent(&del, del.br_startoff + rlen,
783 del.br_blockcount - rlen);
785 trace_xfs_reflink_cow_remap(ip, &del);
787 /* Free the CoW orphan record. */
788 error = xfs_refcount_free_cow_extent(tp->t_mountp, &dfops,
789 del.br_startblock, del.br_blockcount);
793 /* Map the new blocks into the data fork. */
794 error = xfs_bmap_map_extent(tp->t_mountp, &dfops, ip, &del);
798 /* Remove the mapping from the CoW fork. */
799 xfs_bmap_del_extent_cow(ip, &idx, &got, &del);
801 xfs_defer_ijoin(&dfops, ip);
802 error = xfs_defer_finish(&tp, &dfops);
806 if (!xfs_iext_get_extent(ifp, idx, &got))
810 error = xfs_trans_commit(tp);
811 xfs_iunlock(ip, XFS_ILOCK_EXCL);
817 xfs_defer_cancel(&dfops);
819 xfs_trans_cancel(tp);
820 xfs_iunlock(ip, XFS_ILOCK_EXCL);
822 trace_xfs_reflink_end_cow_error(ip, error, _RET_IP_);
827 * Free leftover CoW reservations that didn't get cleaned out.
830 xfs_reflink_recover_cow(
831 struct xfs_mount *mp)
836 if (!xfs_sb_version_hasreflink(&mp->m_sb))
839 for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
840 error = xfs_refcount_recover_cow_leftovers(mp, agno);
849 * Reflinking (Block) Ranges of Two Files Together
851 * First, ensure that the reflink flag is set on both inodes. The flag is an
852 * optimization to avoid unnecessary refcount btree lookups in the write path.
854 * Now we can iteratively remap the range of extents (and holes) in src to the
855 * corresponding ranges in dest. Let drange and srange denote the ranges of
856 * logical blocks in dest and src touched by the reflink operation.
858 * While the length of drange is greater than zero,
859 * - Read src's bmbt at the start of srange ("imap")
860 * - If imap doesn't exist, make imap appear to start at the end of srange
862 * - If imap starts before srange, advance imap to start at srange.
863 * - If imap goes beyond srange, truncate imap to end at the end of srange.
864 * - Punch (imap start - srange start + imap len) blocks from dest at
865 * offset (drange start).
866 * - If imap points to a real range of pblks,
867 * > Increase the refcount of the imap's pblks
868 * > Map imap's pblks into dest at the offset
869 * (drange start + imap start - srange start)
870 * - Advance drange and srange by (imap start - srange start + imap len)
872 * Finally, if the reflink made dest longer, update both the in-core and
873 * on-disk file sizes.
875 * ASCII Art Demonstration:
877 * Let's say we want to reflink this source file:
879 * ----SSSSSSS-SSSSS----SSSSSS (src file)
880 * <-------------------->
882 * into this destination file:
884 * --DDDDDDDDDDDDDDDDDDD--DDD (dest file)
885 * <-------------------->
886 * '-' means a hole, and 'S' and 'D' are written blocks in the src and dest.
887 * Observe that the range has different logical offsets in either file.
889 * Consider that the first extent in the source file doesn't line up with our
890 * reflink range. Unmapping and remapping are separate operations, so we can
891 * unmap more blocks from the destination file than we remap.
893 * ----SSSSSSS-SSSSS----SSSSSS
895 * --DDDDD---------DDDDD--DDD
898 * Now remap the source extent into the destination file:
900 * ----SSSSSSS-SSSSS----SSSSSS
902 * --DDDDD--SSSSSSSDDDDD--DDD
905 * Do likewise with the second hole and extent in our range. Holes in the
906 * unmap range don't affect our operation.
908 * ----SSSSSSS-SSSSS----SSSSSS
910 * --DDDDD--SSSSSSS-SSSSS-DDD
913 * Finally, unmap and remap part of the third extent. This will increase the
914 * size of the destination file.
916 * ----SSSSSSS-SSSSS----SSSSSS
918 * --DDDDD--SSSSSSS-SSSSS----SSS
921 * Once we update the destination file's i_size, we're done.
925 * Ensure the reflink bit is set in both inodes.
928 xfs_reflink_set_inode_flag(
929 struct xfs_inode *src,
930 struct xfs_inode *dest)
932 struct xfs_mount *mp = src->i_mount;
934 struct xfs_trans *tp;
936 if (xfs_is_reflink_inode(src) && xfs_is_reflink_inode(dest))
939 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
943 /* Lock both files against IO */
944 if (src->i_ino == dest->i_ino)
945 xfs_ilock(src, XFS_ILOCK_EXCL);
947 xfs_lock_two_inodes(src, dest, XFS_ILOCK_EXCL);
949 if (!xfs_is_reflink_inode(src)) {
950 trace_xfs_reflink_set_inode_flag(src);
951 xfs_trans_ijoin(tp, src, XFS_ILOCK_EXCL);
952 src->i_d.di_flags2 |= XFS_DIFLAG2_REFLINK;
953 xfs_trans_log_inode(tp, src, XFS_ILOG_CORE);
954 xfs_ifork_init_cow(src);
956 xfs_iunlock(src, XFS_ILOCK_EXCL);
958 if (src->i_ino == dest->i_ino)
961 if (!xfs_is_reflink_inode(dest)) {
962 trace_xfs_reflink_set_inode_flag(dest);
963 xfs_trans_ijoin(tp, dest, XFS_ILOCK_EXCL);
964 dest->i_d.di_flags2 |= XFS_DIFLAG2_REFLINK;
965 xfs_trans_log_inode(tp, dest, XFS_ILOG_CORE);
966 xfs_ifork_init_cow(dest);
968 xfs_iunlock(dest, XFS_ILOCK_EXCL);
971 error = xfs_trans_commit(tp);
977 trace_xfs_reflink_set_inode_flag_error(dest, error, _RET_IP_);
982 * Update destination inode size & cowextsize hint, if necessary.
985 xfs_reflink_update_dest(
986 struct xfs_inode *dest,
988 xfs_extlen_t cowextsize,
991 struct xfs_mount *mp = dest->i_mount;
992 struct xfs_trans *tp;
995 if (is_dedupe && newlen <= i_size_read(VFS_I(dest)) && cowextsize == 0)
998 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
1002 xfs_ilock(dest, XFS_ILOCK_EXCL);
1003 xfs_trans_ijoin(tp, dest, XFS_ILOCK_EXCL);
1005 if (newlen > i_size_read(VFS_I(dest))) {
1006 trace_xfs_reflink_update_inode_size(dest, newlen);
1007 i_size_write(VFS_I(dest), newlen);
1008 dest->i_d.di_size = newlen;
1012 dest->i_d.di_cowextsize = cowextsize;
1013 dest->i_d.di_flags2 |= XFS_DIFLAG2_COWEXTSIZE;
1017 xfs_trans_ichgtime(tp, dest,
1018 XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1020 xfs_trans_log_inode(tp, dest, XFS_ILOG_CORE);
1022 error = xfs_trans_commit(tp);
1028 trace_xfs_reflink_update_inode_size_error(dest, error, _RET_IP_);
1033 * Do we have enough reserve in this AG to handle a reflink? The refcount
1034 * btree already reserved all the space it needs, but the rmap btree can grow
1035 * infinitely, so we won't allow more reflinks when the AG is down to the
1039 xfs_reflink_ag_has_free_space(
1040 struct xfs_mount *mp,
1041 xfs_agnumber_t agno)
1043 struct xfs_perag *pag;
1046 if (!xfs_sb_version_hasrmapbt(&mp->m_sb))
1049 pag = xfs_perag_get(mp, agno);
1050 if (xfs_ag_resv_critical(pag, XFS_AG_RESV_AGFL) ||
1051 xfs_ag_resv_critical(pag, XFS_AG_RESV_METADATA))
1058 * Unmap a range of blocks from a file, then map other blocks into the hole.
1059 * The range to unmap is (destoff : destoff + srcioff + irec->br_blockcount).
1060 * The extent irec is mapped into dest at irec->br_startoff.
1063 xfs_reflink_remap_extent(
1064 struct xfs_inode *ip,
1065 struct xfs_bmbt_irec *irec,
1066 xfs_fileoff_t destoff,
1067 xfs_off_t new_isize)
1069 struct xfs_mount *mp = ip->i_mount;
1070 bool real_extent = xfs_bmap_is_real_extent(irec);
1071 struct xfs_trans *tp;
1072 xfs_fsblock_t firstfsb;
1073 unsigned int resblks;
1074 struct xfs_defer_ops dfops;
1075 struct xfs_bmbt_irec uirec;
1077 xfs_filblks_t unmap_len;
1081 unmap_len = irec->br_startoff + irec->br_blockcount - destoff;
1082 trace_xfs_reflink_punch_range(ip, destoff, unmap_len);
1084 /* No reflinking if we're low on space */
1086 error = xfs_reflink_ag_has_free_space(mp,
1087 XFS_FSB_TO_AGNO(mp, irec->br_startblock));
1092 /* Start a rolling transaction to switch the mappings */
1093 resblks = XFS_EXTENTADD_SPACE_RES(ip->i_mount, XFS_DATA_FORK);
1094 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, resblks, 0, 0, &tp);
1098 xfs_ilock(ip, XFS_ILOCK_EXCL);
1099 xfs_trans_ijoin(tp, ip, 0);
1101 /* If we're not just clearing space, then do we have enough quota? */
1103 error = xfs_trans_reserve_quota_nblks(tp, ip,
1104 irec->br_blockcount, 0, XFS_QMOPT_RES_REGBLKS);
1109 trace_xfs_reflink_remap(ip, irec->br_startoff,
1110 irec->br_blockcount, irec->br_startblock);
1112 /* Unmap the old blocks in the data fork. */
1115 xfs_defer_init(&dfops, &firstfsb);
1116 error = __xfs_bunmapi(tp, ip, destoff, &rlen, 0, 1,
1122 * Trim the extent to whatever got unmapped.
1123 * Remember, bunmapi works backwards.
1125 uirec.br_startblock = irec->br_startblock + rlen;
1126 uirec.br_startoff = irec->br_startoff + rlen;
1127 uirec.br_blockcount = unmap_len - rlen;
1130 /* If this isn't a real mapping, we're done. */
1131 if (!real_extent || uirec.br_blockcount == 0)
1134 trace_xfs_reflink_remap(ip, uirec.br_startoff,
1135 uirec.br_blockcount, uirec.br_startblock);
1137 /* Update the refcount tree */
1138 error = xfs_refcount_increase_extent(mp, &dfops, &uirec);
1142 /* Map the new blocks into the data fork. */
1143 error = xfs_bmap_map_extent(mp, &dfops, ip, &uirec);
1147 /* Update quota accounting. */
1148 xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_BCOUNT,
1149 uirec.br_blockcount);
1151 /* Update dest isize if needed. */
1152 newlen = XFS_FSB_TO_B(mp,
1153 uirec.br_startoff + uirec.br_blockcount);
1154 newlen = min_t(xfs_off_t, newlen, new_isize);
1155 if (newlen > i_size_read(VFS_I(ip))) {
1156 trace_xfs_reflink_update_inode_size(ip, newlen);
1157 i_size_write(VFS_I(ip), newlen);
1158 ip->i_d.di_size = newlen;
1159 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1163 /* Process all the deferred stuff. */
1164 xfs_defer_ijoin(&dfops, ip);
1165 error = xfs_defer_finish(&tp, &dfops);
1170 error = xfs_trans_commit(tp);
1171 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1177 xfs_defer_cancel(&dfops);
1179 xfs_trans_cancel(tp);
1180 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1182 trace_xfs_reflink_remap_extent_error(ip, error, _RET_IP_);
1187 * Iteratively remap one file's extents (and holes) to another's.
1190 xfs_reflink_remap_blocks(
1191 struct xfs_inode *src,
1192 xfs_fileoff_t srcoff,
1193 struct xfs_inode *dest,
1194 xfs_fileoff_t destoff,
1196 xfs_off_t new_isize)
1198 struct xfs_bmbt_irec imap;
1201 xfs_filblks_t range_len;
1203 /* drange = (destoff, destoff + len); srange = (srcoff, srcoff + len) */
1205 trace_xfs_reflink_remap_blocks_loop(src, srcoff, len,
1207 /* Read extent from the source file */
1209 xfs_ilock(src, XFS_ILOCK_EXCL);
1210 error = xfs_bmapi_read(src, srcoff, len, &imap, &nimaps, 0);
1211 xfs_iunlock(src, XFS_ILOCK_EXCL);
1214 ASSERT(nimaps == 1);
1216 trace_xfs_reflink_remap_imap(src, srcoff, len, XFS_IO_OVERWRITE,
1219 /* Translate imap into the destination file. */
1220 range_len = imap.br_startoff + imap.br_blockcount - srcoff;
1221 imap.br_startoff += destoff - srcoff;
1223 /* Clear dest from destoff to the end of imap and map it in. */
1224 error = xfs_reflink_remap_extent(dest, &imap, destoff,
1229 if (fatal_signal_pending(current)) {
1234 /* Advance drange/srange */
1235 srcoff += range_len;
1236 destoff += range_len;
1243 trace_xfs_reflink_remap_blocks_error(dest, error, _RET_IP_);
1248 * Link a range of blocks from one file to another.
1251 xfs_reflink_remap_range(
1252 struct file *file_in,
1254 struct file *file_out,
1259 struct inode *inode_in = file_inode(file_in);
1260 struct xfs_inode *src = XFS_I(inode_in);
1261 struct inode *inode_out = file_inode(file_out);
1262 struct xfs_inode *dest = XFS_I(inode_out);
1263 struct xfs_mount *mp = src->i_mount;
1264 bool same_inode = (inode_in == inode_out);
1265 xfs_fileoff_t sfsbno, dfsbno;
1266 xfs_filblks_t fsblen;
1267 xfs_extlen_t cowextsize;
1270 if (!xfs_sb_version_hasreflink(&mp->m_sb))
1273 if (XFS_FORCED_SHUTDOWN(mp))
1276 /* Lock both files against IO */
1277 lock_two_nondirectories(inode_in, inode_out);
1279 xfs_ilock(src, XFS_MMAPLOCK_EXCL);
1281 xfs_lock_two_inodes(src, dest, XFS_MMAPLOCK_EXCL);
1283 /* Check file eligibility and prepare for block sharing. */
1285 /* Don't reflink realtime inodes */
1286 if (XFS_IS_REALTIME_INODE(src) || XFS_IS_REALTIME_INODE(dest))
1289 /* Don't share DAX file data for now. */
1290 if (IS_DAX(inode_in) || IS_DAX(inode_out))
1293 ret = vfs_clone_file_prep_inodes(inode_in, pos_in, inode_out, pos_out,
1298 trace_xfs_reflink_remap_range(src, pos_in, len, dest, pos_out);
1300 /* Set flags and remap blocks. */
1301 ret = xfs_reflink_set_inode_flag(src, dest);
1305 dfsbno = XFS_B_TO_FSBT(mp, pos_out);
1306 sfsbno = XFS_B_TO_FSBT(mp, pos_in);
1307 fsblen = XFS_B_TO_FSB(mp, len);
1308 ret = xfs_reflink_remap_blocks(src, sfsbno, dest, dfsbno, fsblen,
1313 /* Zap any page cache for the destination file's range. */
1314 truncate_inode_pages_range(&inode_out->i_data, pos_out,
1315 PAGE_ALIGN(pos_out + len) - 1);
1318 * Carry the cowextsize hint from src to dest if we're sharing the
1319 * entire source file to the entire destination file, the source file
1320 * has a cowextsize hint, and the destination file does not.
1323 if (pos_in == 0 && len == i_size_read(inode_in) &&
1324 (src->i_d.di_flags2 & XFS_DIFLAG2_COWEXTSIZE) &&
1325 pos_out == 0 && len >= i_size_read(inode_out) &&
1326 !(dest->i_d.di_flags2 & XFS_DIFLAG2_COWEXTSIZE))
1327 cowextsize = src->i_d.di_cowextsize;
1329 ret = xfs_reflink_update_dest(dest, pos_out + len, cowextsize,
1333 xfs_iunlock(src, XFS_MMAPLOCK_EXCL);
1335 xfs_iunlock(dest, XFS_MMAPLOCK_EXCL);
1336 unlock_two_nondirectories(inode_in, inode_out);
1338 trace_xfs_reflink_remap_range_error(dest, ret, _RET_IP_);
1343 * The user wants to preemptively CoW all shared blocks in this file,
1344 * which enables us to turn off the reflink flag. Iterate all
1345 * extents which are not prealloc/delalloc to see which ranges are
1346 * mentioned in the refcount tree, then read those blocks into the
1347 * pagecache, dirty them, fsync them back out, and then we can update
1348 * the inode flag. What happens if we run out of memory? :)
1351 xfs_reflink_dirty_extents(
1352 struct xfs_inode *ip,
1357 struct xfs_mount *mp = ip->i_mount;
1358 xfs_agnumber_t agno;
1359 xfs_agblock_t agbno;
1365 struct xfs_bmbt_irec map[2];
1369 while (end - fbno > 0) {
1372 * Look for extents in the file. Skip holes, delalloc, or
1373 * unwritten extents; they can't be reflinked.
1375 error = xfs_bmapi_read(ip, fbno, end - fbno, map, &nmaps, 0);
1380 if (!xfs_bmap_is_real_extent(&map[0]))
1384 while (map[1].br_blockcount) {
1385 agno = XFS_FSB_TO_AGNO(mp, map[1].br_startblock);
1386 agbno = XFS_FSB_TO_AGBNO(mp, map[1].br_startblock);
1387 aglen = map[1].br_blockcount;
1389 error = xfs_reflink_find_shared(mp, NULL, agno, agbno,
1390 aglen, &rbno, &rlen, true);
1393 if (rbno == NULLAGBLOCK)
1396 /* Dirty the pages */
1397 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1398 fpos = XFS_FSB_TO_B(mp, map[1].br_startoff +
1400 flen = XFS_FSB_TO_B(mp, rlen);
1401 if (fpos + flen > isize)
1402 flen = isize - fpos;
1403 error = iomap_file_dirty(VFS_I(ip), fpos, flen,
1405 xfs_ilock(ip, XFS_ILOCK_EXCL);
1409 map[1].br_blockcount -= (rbno - agbno + rlen);
1410 map[1].br_startoff += (rbno - agbno + rlen);
1411 map[1].br_startblock += (rbno - agbno + rlen);
1415 fbno = map[0].br_startoff + map[0].br_blockcount;
1421 /* Does this inode need the reflink flag? */
1423 xfs_reflink_inode_has_shared_extents(
1424 struct xfs_trans *tp,
1425 struct xfs_inode *ip,
1428 struct xfs_bmbt_irec got;
1429 struct xfs_mount *mp = ip->i_mount;
1430 struct xfs_ifork *ifp;
1431 xfs_agnumber_t agno;
1432 xfs_agblock_t agbno;
1440 ifp = XFS_IFORK_PTR(ip, XFS_DATA_FORK);
1441 if (!(ifp->if_flags & XFS_IFEXTENTS)) {
1442 error = xfs_iread_extents(tp, ip, XFS_DATA_FORK);
1447 *has_shared = false;
1448 found = xfs_iext_lookup_extent(ip, ifp, 0, &idx, &got);
1450 if (isnullstartblock(got.br_startblock) ||
1451 got.br_state != XFS_EXT_NORM)
1453 agno = XFS_FSB_TO_AGNO(mp, got.br_startblock);
1454 agbno = XFS_FSB_TO_AGBNO(mp, got.br_startblock);
1455 aglen = got.br_blockcount;
1457 error = xfs_reflink_find_shared(mp, tp, agno, agbno, aglen,
1458 &rbno, &rlen, false);
1461 /* Is there still a shared block here? */
1462 if (rbno != NULLAGBLOCK) {
1467 found = xfs_iext_get_extent(ifp, ++idx, &got);
1473 /* Clear the inode reflink flag if there are no shared extents. */
1475 xfs_reflink_clear_inode_flag(
1476 struct xfs_inode *ip,
1477 struct xfs_trans **tpp)
1482 ASSERT(xfs_is_reflink_inode(ip));
1484 error = xfs_reflink_inode_has_shared_extents(*tpp, ip, &needs_flag);
1485 if (error || needs_flag)
1489 * We didn't find any shared blocks so turn off the reflink flag.
1490 * First, get rid of any leftover CoW mappings.
1492 error = xfs_reflink_cancel_cow_blocks(ip, tpp, 0, NULLFILEOFF, true);
1496 /* Clear the inode flag. */
1497 trace_xfs_reflink_unset_inode_flag(ip);
1498 ip->i_d.di_flags2 &= ~XFS_DIFLAG2_REFLINK;
1499 xfs_inode_clear_cowblocks_tag(ip);
1500 xfs_trans_ijoin(*tpp, ip, 0);
1501 xfs_trans_log_inode(*tpp, ip, XFS_ILOG_CORE);
1507 * Clear the inode reflink flag if there are no shared extents and the size
1511 xfs_reflink_try_clear_inode_flag(
1512 struct xfs_inode *ip)
1514 struct xfs_mount *mp = ip->i_mount;
1515 struct xfs_trans *tp;
1518 /* Start a rolling transaction to remove the mappings */
1519 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, 0, 0, 0, &tp);
1523 xfs_ilock(ip, XFS_ILOCK_EXCL);
1524 xfs_trans_ijoin(tp, ip, 0);
1526 error = xfs_reflink_clear_inode_flag(ip, &tp);
1530 error = xfs_trans_commit(tp);
1534 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1537 xfs_trans_cancel(tp);
1539 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1544 * Pre-COW all shared blocks within a given byte range of a file and turn off
1545 * the reflink flag if we unshare all of the file's blocks.
1548 xfs_reflink_unshare(
1549 struct xfs_inode *ip,
1553 struct xfs_mount *mp = ip->i_mount;
1559 if (!xfs_is_reflink_inode(ip))
1562 trace_xfs_reflink_unshare(ip, offset, len);
1564 inode_dio_wait(VFS_I(ip));
1566 /* Try to CoW the selected ranges */
1567 xfs_ilock(ip, XFS_ILOCK_EXCL);
1568 fbno = XFS_B_TO_FSBT(mp, offset);
1569 isize = i_size_read(VFS_I(ip));
1570 end = XFS_B_TO_FSB(mp, offset + len);
1571 error = xfs_reflink_dirty_extents(ip, fbno, end, isize);
1574 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1576 /* Wait for the IO to finish */
1577 error = filemap_write_and_wait(VFS_I(ip)->i_mapping);
1581 /* Turn off the reflink flag if possible. */
1582 error = xfs_reflink_try_clear_inode_flag(ip);
1589 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1591 trace_xfs_reflink_unshare_error(ip, error, _RET_IP_);