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,
163 bool find_end_of_shared)
165 struct xfs_buf *agbp;
166 struct xfs_btree_cur *cur;
169 error = xfs_alloc_read_agf(mp, NULL, agno, 0, &agbp);
173 cur = xfs_refcountbt_init_cursor(mp, NULL, agbp, agno, NULL);
175 error = xfs_refcount_find_shared(cur, agbno, aglen, fbno, flen,
178 xfs_btree_del_cursor(cur, error ? XFS_BTREE_ERROR : XFS_BTREE_NOERROR);
185 * Trim the mapping to the next block where there's a change in the
186 * shared/unshared status. More specifically, this means that we
187 * find the lowest-numbered extent of shared blocks that coincides with
188 * the given block mapping. If the shared extent overlaps the start of
189 * the mapping, trim the mapping to the end of the shared extent. If
190 * the shared region intersects the mapping, trim the mapping to the
191 * start of the shared extent. If there are no shared regions that
192 * overlap, just return the original extent.
195 xfs_reflink_trim_around_shared(
196 struct xfs_inode *ip,
197 struct xfs_bmbt_irec *irec,
208 /* Holes, unwritten, and delalloc extents cannot be shared */
209 if (!xfs_is_reflink_inode(ip) ||
211 irec->br_startblock == HOLESTARTBLOCK ||
212 irec->br_startblock == DELAYSTARTBLOCK ||
213 isnullstartblock(irec->br_startblock)) {
218 trace_xfs_reflink_trim_around_shared(ip, irec);
220 agno = XFS_FSB_TO_AGNO(ip->i_mount, irec->br_startblock);
221 agbno = XFS_FSB_TO_AGBNO(ip->i_mount, irec->br_startblock);
222 aglen = irec->br_blockcount;
224 error = xfs_reflink_find_shared(ip->i_mount, agno, agbno,
225 aglen, &fbno, &flen, true);
229 *shared = *trimmed = false;
230 if (fbno == NULLAGBLOCK) {
231 /* No shared blocks at all. */
233 } else if (fbno == agbno) {
235 * The start of this extent is shared. Truncate the
236 * mapping at the end of the shared region so that a
237 * subsequent iteration starts at the start of the
240 irec->br_blockcount = flen;
247 * There's a shared extent midway through this extent.
248 * Truncate the mapping at the start of the shared
249 * extent so that a subsequent iteration starts at the
250 * start of the shared region.
252 irec->br_blockcount = fbno - agbno;
259 * Trim the passed in imap to the next shared/unshared extent boundary, and
260 * if imap->br_startoff points to a shared extent reserve space for it in the
261 * COW fork. In this case *shared is set to true, else to false.
263 * Note that imap will always contain the block numbers for the existing blocks
264 * in the data fork, as the upper layers need them for read-modify-write
268 xfs_reflink_reserve_cow(
269 struct xfs_inode *ip,
270 struct xfs_bmbt_irec *imap,
273 struct xfs_ifork *ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
274 struct xfs_bmbt_irec got;
276 bool eof = false, trimmed;
280 * Search the COW fork extent list first. This serves two purposes:
281 * first this implement the speculative preallocation using cowextisze,
282 * so that we also unshared block adjacent to shared blocks instead
283 * of just the shared blocks themselves. Second the lookup in the
284 * extent list is generally faster than going out to the shared extent
288 if (!xfs_iext_lookup_extent(ip, ifp, imap->br_startoff, &idx, &got))
290 if (!eof && got.br_startoff <= imap->br_startoff) {
291 trace_xfs_reflink_cow_found(ip, imap);
292 xfs_trim_extent(imap, got.br_startoff, got.br_blockcount);
298 /* Trim the mapping to the nearest shared extent boundary. */
299 error = xfs_reflink_trim_around_shared(ip, imap, shared, &trimmed);
303 /* Not shared? Just report the (potentially capped) extent. */
308 * Fork all the shared blocks from our write offset until the end of
311 error = xfs_qm_dqattach_locked(ip, 0);
315 error = xfs_bmapi_reserve_delalloc(ip, XFS_COW_FORK, imap->br_startoff,
316 imap->br_blockcount, 0, &got, &idx, eof);
317 if (error == -ENOSPC || error == -EDQUOT)
318 trace_xfs_reflink_cow_enospc(ip, imap);
322 trace_xfs_reflink_cow_alloc(ip, &got);
326 /* Convert part of an unwritten CoW extent to a real one. */
328 xfs_reflink_convert_cow_extent(
329 struct xfs_inode *ip,
330 struct xfs_bmbt_irec *imap,
331 xfs_fileoff_t offset_fsb,
332 xfs_filblks_t count_fsb,
333 struct xfs_defer_ops *dfops)
335 xfs_fsblock_t first_block;
338 if (imap->br_state == XFS_EXT_NORM)
341 xfs_trim_extent(imap, offset_fsb, count_fsb);
342 trace_xfs_reflink_convert_cow(ip, imap);
343 if (imap->br_blockcount == 0)
345 return xfs_bmapi_write(NULL, ip, imap->br_startoff, imap->br_blockcount,
346 XFS_BMAPI_COWFORK | XFS_BMAPI_CONVERT, &first_block,
347 0, imap, &nimaps, dfops);
350 /* Convert all of the unwritten CoW extents in a file's range to real ones. */
352 xfs_reflink_convert_cow(
353 struct xfs_inode *ip,
357 struct xfs_bmbt_irec got;
358 struct xfs_defer_ops dfops;
359 struct xfs_mount *mp = ip->i_mount;
360 struct xfs_ifork *ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
361 xfs_fileoff_t offset_fsb = XFS_B_TO_FSBT(mp, offset);
362 xfs_fileoff_t end_fsb = XFS_B_TO_FSB(mp, offset + count);
367 xfs_ilock(ip, XFS_ILOCK_EXCL);
369 /* Convert all the extents to real from unwritten. */
370 for (found = xfs_iext_lookup_extent(ip, ifp, offset_fsb, &idx, &got);
371 found && got.br_startoff < end_fsb;
372 found = xfs_iext_get_extent(ifp, ++idx, &got)) {
373 error = xfs_reflink_convert_cow_extent(ip, &got, offset_fsb,
374 end_fsb - offset_fsb, &dfops);
380 xfs_iunlock(ip, XFS_ILOCK_EXCL);
384 /* Allocate all CoW reservations covering a range of blocks in a file. */
386 xfs_reflink_allocate_cow(
387 struct xfs_inode *ip,
388 struct xfs_bmbt_irec *imap,
392 struct xfs_mount *mp = ip->i_mount;
393 xfs_fileoff_t offset_fsb = imap->br_startoff;
394 xfs_filblks_t count_fsb = imap->br_blockcount;
395 struct xfs_bmbt_irec got;
396 struct xfs_defer_ops dfops;
397 struct xfs_trans *tp = NULL;
398 xfs_fsblock_t first_block;
399 int nimaps, error = 0;
401 xfs_filblks_t resaligned;
402 xfs_extlen_t resblks = 0;
406 ASSERT(xfs_is_reflink_inode(ip));
407 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL | XFS_ILOCK_SHARED));
410 * Even if the extent is not shared we might have a preallocation for
411 * it in the COW fork. If so use it.
413 if (xfs_iext_lookup_extent(ip, ip->i_cowfp, offset_fsb, &idx, &got) &&
414 got.br_startoff <= offset_fsb) {
417 /* If we have a real allocation in the COW fork we're done. */
418 if (!isnullstartblock(got.br_startblock)) {
419 xfs_trim_extent(&got, offset_fsb, count_fsb);
424 xfs_trim_extent(imap, got.br_startoff, got.br_blockcount);
426 error = xfs_reflink_trim_around_shared(ip, imap, shared, &trimmed);
427 if (error || !*shared)
432 resaligned = xfs_aligned_fsb_count(imap->br_startoff,
433 imap->br_blockcount, xfs_get_cowextsz_hint(ip));
434 resblks = XFS_DIOSTRAT_SPACE_RES(mp, resaligned);
436 xfs_iunlock(ip, *lockmode);
437 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, resblks, 0, 0, &tp);
438 *lockmode = XFS_ILOCK_EXCL;
439 xfs_ilock(ip, *lockmode);
444 error = xfs_qm_dqattach_locked(ip, 0);
450 error = xfs_trans_reserve_quota_nblks(tp, ip, resblks, 0,
451 XFS_QMOPT_RES_REGBLKS);
455 xfs_trans_ijoin(tp, ip, 0);
457 xfs_defer_init(&dfops, &first_block);
460 /* Allocate the entire reservation as unwritten blocks. */
461 error = xfs_bmapi_write(tp, ip, imap->br_startoff, imap->br_blockcount,
462 XFS_BMAPI_COWFORK | XFS_BMAPI_PREALLOC, &first_block,
463 resblks, imap, &nimaps, &dfops);
465 goto out_bmap_cancel;
468 error = xfs_defer_finish(&tp, &dfops, NULL);
470 goto out_bmap_cancel;
472 error = xfs_trans_commit(tp);
476 return xfs_reflink_convert_cow_extent(ip, imap, offset_fsb, count_fsb,
479 xfs_defer_cancel(&dfops);
480 xfs_trans_unreserve_quota_nblks(tp, ip, (long)resblks, 0,
481 XFS_QMOPT_RES_REGBLKS);
484 xfs_trans_cancel(tp);
489 * Find the CoW reservation for a given byte offset of a file.
492 xfs_reflink_find_cow_mapping(
493 struct xfs_inode *ip,
495 struct xfs_bmbt_irec *imap)
497 struct xfs_ifork *ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
498 xfs_fileoff_t offset_fsb;
499 struct xfs_bmbt_irec got;
502 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL | XFS_ILOCK_SHARED));
503 ASSERT(xfs_is_reflink_inode(ip));
505 offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
506 if (!xfs_iext_lookup_extent(ip, ifp, offset_fsb, &idx, &got))
508 if (got.br_startoff > offset_fsb)
511 trace_xfs_reflink_find_cow_mapping(ip, offset, 1, XFS_IO_OVERWRITE,
518 * Trim an extent to end at the next CoW reservation past offset_fsb.
521 xfs_reflink_trim_irec_to_next_cow(
522 struct xfs_inode *ip,
523 xfs_fileoff_t offset_fsb,
524 struct xfs_bmbt_irec *imap)
526 struct xfs_ifork *ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
527 struct xfs_bmbt_irec got;
530 if (!xfs_is_reflink_inode(ip))
533 /* Find the extent in the CoW fork. */
534 if (!xfs_iext_lookup_extent(ip, ifp, offset_fsb, &idx, &got))
537 /* This is the extent before; try sliding up one. */
538 if (got.br_startoff < offset_fsb) {
539 if (!xfs_iext_get_extent(ifp, idx + 1, &got))
543 if (got.br_startoff >= imap->br_startoff + imap->br_blockcount)
546 imap->br_blockcount = got.br_startoff - imap->br_startoff;
547 trace_xfs_reflink_trim_irec(ip, imap);
551 * Cancel CoW reservations for some block range of an inode.
553 * If cancel_real is true this function cancels all COW fork extents for the
554 * inode; if cancel_real is false, real extents are not cleared.
557 xfs_reflink_cancel_cow_blocks(
558 struct xfs_inode *ip,
559 struct xfs_trans **tpp,
560 xfs_fileoff_t offset_fsb,
561 xfs_fileoff_t end_fsb,
564 struct xfs_ifork *ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
565 struct xfs_bmbt_irec got, del;
567 xfs_fsblock_t firstfsb;
568 struct xfs_defer_ops dfops;
571 if (!xfs_is_reflink_inode(ip))
573 if (!xfs_iext_lookup_extent(ip, ifp, offset_fsb, &idx, &got))
576 while (got.br_startoff < end_fsb) {
578 xfs_trim_extent(&del, offset_fsb, end_fsb - offset_fsb);
579 trace_xfs_reflink_cancel_cow(ip, &del);
581 if (isnullstartblock(del.br_startblock)) {
582 error = xfs_bmap_del_extent_delay(ip, XFS_COW_FORK,
586 } else if (del.br_state == XFS_EXT_UNWRITTEN || cancel_real) {
587 xfs_trans_ijoin(*tpp, ip, 0);
588 xfs_defer_init(&dfops, &firstfsb);
590 /* Free the CoW orphan record. */
591 error = xfs_refcount_free_cow_extent(ip->i_mount,
592 &dfops, del.br_startblock,
597 xfs_bmap_add_free(ip->i_mount, &dfops,
598 del.br_startblock, del.br_blockcount,
601 /* Update quota accounting */
602 xfs_trans_mod_dquot_byino(*tpp, ip, XFS_TRANS_DQ_BCOUNT,
603 -(long)del.br_blockcount);
605 /* Roll the transaction */
606 error = xfs_defer_finish(tpp, &dfops, ip);
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);
712 /* Start a rolling transaction to switch the mappings */
713 resblks = XFS_EXTENTADD_SPACE_RES(ip->i_mount, XFS_DATA_FORK);
714 error = xfs_trans_alloc(ip->i_mount, &M_RES(ip->i_mount)->tr_write,
719 xfs_ilock(ip, XFS_ILOCK_EXCL);
720 xfs_trans_ijoin(tp, ip, 0);
722 /* If there is a hole at end_fsb - 1 go to the previous extent */
723 if (!xfs_iext_lookup_extent(ip, ifp, end_fsb - 1, &idx, &got) ||
724 got.br_startoff > end_fsb) {
726 xfs_iext_get_extent(ifp, --idx, &got);
729 /* Walk backwards until we're out of the I/O range... */
730 while (got.br_startoff + got.br_blockcount > offset_fsb) {
732 xfs_trim_extent(&del, offset_fsb, end_fsb - offset_fsb);
734 /* Extent delete may have bumped idx forward */
735 if (!del.br_blockcount) {
740 ASSERT(!isnullstartblock(got.br_startblock));
743 * Don't remap unwritten extents; these are
744 * speculatively preallocated CoW extents that have been
745 * allocated but have not yet been involved in a write.
747 if (got.br_state == XFS_EXT_UNWRITTEN) {
752 /* Unmap the old blocks in the data fork. */
753 xfs_defer_init(&dfops, &firstfsb);
754 rlen = del.br_blockcount;
755 error = __xfs_bunmapi(tp, ip, del.br_startoff, &rlen, 0, 1,
760 /* Trim the extent to whatever got unmapped. */
762 xfs_trim_extent(&del, del.br_startoff + rlen,
763 del.br_blockcount - rlen);
765 trace_xfs_reflink_cow_remap(ip, &del);
767 /* Free the CoW orphan record. */
768 error = xfs_refcount_free_cow_extent(tp->t_mountp, &dfops,
769 del.br_startblock, del.br_blockcount);
773 /* Map the new blocks into the data fork. */
774 error = xfs_bmap_map_extent(tp->t_mountp, &dfops, ip, &del);
778 /* Remove the mapping from the CoW fork. */
779 xfs_bmap_del_extent_cow(ip, &idx, &got, &del);
781 error = xfs_defer_finish(&tp, &dfops, ip);
785 if (!xfs_iext_get_extent(ifp, idx, &got))
789 error = xfs_trans_commit(tp);
790 xfs_iunlock(ip, XFS_ILOCK_EXCL);
796 xfs_defer_cancel(&dfops);
797 xfs_trans_cancel(tp);
798 xfs_iunlock(ip, XFS_ILOCK_EXCL);
800 trace_xfs_reflink_end_cow_error(ip, error, _RET_IP_);
805 * Free leftover CoW reservations that didn't get cleaned out.
808 xfs_reflink_recover_cow(
809 struct xfs_mount *mp)
814 if (!xfs_sb_version_hasreflink(&mp->m_sb))
817 for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
818 error = xfs_refcount_recover_cow_leftovers(mp, agno);
827 * Reflinking (Block) Ranges of Two Files Together
829 * First, ensure that the reflink flag is set on both inodes. The flag is an
830 * optimization to avoid unnecessary refcount btree lookups in the write path.
832 * Now we can iteratively remap the range of extents (and holes) in src to the
833 * corresponding ranges in dest. Let drange and srange denote the ranges of
834 * logical blocks in dest and src touched by the reflink operation.
836 * While the length of drange is greater than zero,
837 * - Read src's bmbt at the start of srange ("imap")
838 * - If imap doesn't exist, make imap appear to start at the end of srange
840 * - If imap starts before srange, advance imap to start at srange.
841 * - If imap goes beyond srange, truncate imap to end at the end of srange.
842 * - Punch (imap start - srange start + imap len) blocks from dest at
843 * offset (drange start).
844 * - If imap points to a real range of pblks,
845 * > Increase the refcount of the imap's pblks
846 * > Map imap's pblks into dest at the offset
847 * (drange start + imap start - srange start)
848 * - Advance drange and srange by (imap start - srange start + imap len)
850 * Finally, if the reflink made dest longer, update both the in-core and
851 * on-disk file sizes.
853 * ASCII Art Demonstration:
855 * Let's say we want to reflink this source file:
857 * ----SSSSSSS-SSSSS----SSSSSS (src file)
858 * <-------------------->
860 * into this destination file:
862 * --DDDDDDDDDDDDDDDDDDD--DDD (dest file)
863 * <-------------------->
864 * '-' means a hole, and 'S' and 'D' are written blocks in the src and dest.
865 * Observe that the range has different logical offsets in either file.
867 * Consider that the first extent in the source file doesn't line up with our
868 * reflink range. Unmapping and remapping are separate operations, so we can
869 * unmap more blocks from the destination file than we remap.
871 * ----SSSSSSS-SSSSS----SSSSSS
873 * --DDDDD---------DDDDD--DDD
876 * Now remap the source extent into the destination file:
878 * ----SSSSSSS-SSSSS----SSSSSS
880 * --DDDDD--SSSSSSSDDDDD--DDD
883 * Do likewise with the second hole and extent in our range. Holes in the
884 * unmap range don't affect our operation.
886 * ----SSSSSSS-SSSSS----SSSSSS
888 * --DDDDD--SSSSSSS-SSSSS-DDD
891 * Finally, unmap and remap part of the third extent. This will increase the
892 * size of the destination file.
894 * ----SSSSSSS-SSSSS----SSSSSS
896 * --DDDDD--SSSSSSS-SSSSS----SSS
899 * Once we update the destination file's i_size, we're done.
903 * Ensure the reflink bit is set in both inodes.
906 xfs_reflink_set_inode_flag(
907 struct xfs_inode *src,
908 struct xfs_inode *dest)
910 struct xfs_mount *mp = src->i_mount;
912 struct xfs_trans *tp;
914 if (xfs_is_reflink_inode(src) && xfs_is_reflink_inode(dest))
917 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
921 /* Lock both files against IO */
922 if (src->i_ino == dest->i_ino)
923 xfs_ilock(src, XFS_ILOCK_EXCL);
925 xfs_lock_two_inodes(src, dest, XFS_ILOCK_EXCL);
927 if (!xfs_is_reflink_inode(src)) {
928 trace_xfs_reflink_set_inode_flag(src);
929 xfs_trans_ijoin(tp, src, XFS_ILOCK_EXCL);
930 src->i_d.di_flags2 |= XFS_DIFLAG2_REFLINK;
931 xfs_trans_log_inode(tp, src, XFS_ILOG_CORE);
932 xfs_ifork_init_cow(src);
934 xfs_iunlock(src, XFS_ILOCK_EXCL);
936 if (src->i_ino == dest->i_ino)
939 if (!xfs_is_reflink_inode(dest)) {
940 trace_xfs_reflink_set_inode_flag(dest);
941 xfs_trans_ijoin(tp, dest, XFS_ILOCK_EXCL);
942 dest->i_d.di_flags2 |= XFS_DIFLAG2_REFLINK;
943 xfs_trans_log_inode(tp, dest, XFS_ILOG_CORE);
944 xfs_ifork_init_cow(dest);
946 xfs_iunlock(dest, XFS_ILOCK_EXCL);
949 error = xfs_trans_commit(tp);
955 trace_xfs_reflink_set_inode_flag_error(dest, error, _RET_IP_);
960 * Update destination inode size & cowextsize hint, if necessary.
963 xfs_reflink_update_dest(
964 struct xfs_inode *dest,
966 xfs_extlen_t cowextsize,
969 struct xfs_mount *mp = dest->i_mount;
970 struct xfs_trans *tp;
973 if (is_dedupe && newlen <= i_size_read(VFS_I(dest)) && cowextsize == 0)
976 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
980 xfs_ilock(dest, XFS_ILOCK_EXCL);
981 xfs_trans_ijoin(tp, dest, XFS_ILOCK_EXCL);
983 if (newlen > i_size_read(VFS_I(dest))) {
984 trace_xfs_reflink_update_inode_size(dest, newlen);
985 i_size_write(VFS_I(dest), newlen);
986 dest->i_d.di_size = newlen;
990 dest->i_d.di_cowextsize = cowextsize;
991 dest->i_d.di_flags2 |= XFS_DIFLAG2_COWEXTSIZE;
995 xfs_trans_ichgtime(tp, dest,
996 XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
998 xfs_trans_log_inode(tp, dest, XFS_ILOG_CORE);
1000 error = xfs_trans_commit(tp);
1006 trace_xfs_reflink_update_inode_size_error(dest, error, _RET_IP_);
1011 * Do we have enough reserve in this AG to handle a reflink? The refcount
1012 * btree already reserved all the space it needs, but the rmap btree can grow
1013 * infinitely, so we won't allow more reflinks when the AG is down to the
1017 xfs_reflink_ag_has_free_space(
1018 struct xfs_mount *mp,
1019 xfs_agnumber_t agno)
1021 struct xfs_perag *pag;
1024 if (!xfs_sb_version_hasrmapbt(&mp->m_sb))
1027 pag = xfs_perag_get(mp, agno);
1028 if (xfs_ag_resv_critical(pag, XFS_AG_RESV_AGFL) ||
1029 xfs_ag_resv_critical(pag, XFS_AG_RESV_METADATA))
1036 * Unmap a range of blocks from a file, then map other blocks into the hole.
1037 * The range to unmap is (destoff : destoff + srcioff + irec->br_blockcount).
1038 * The extent irec is mapped into dest at irec->br_startoff.
1041 xfs_reflink_remap_extent(
1042 struct xfs_inode *ip,
1043 struct xfs_bmbt_irec *irec,
1044 xfs_fileoff_t destoff,
1045 xfs_off_t new_isize)
1047 struct xfs_mount *mp = ip->i_mount;
1048 struct xfs_trans *tp;
1049 xfs_fsblock_t firstfsb;
1050 unsigned int resblks;
1051 struct xfs_defer_ops dfops;
1052 struct xfs_bmbt_irec uirec;
1055 xfs_filblks_t unmap_len;
1059 unmap_len = irec->br_startoff + irec->br_blockcount - destoff;
1060 trace_xfs_reflink_punch_range(ip, destoff, unmap_len);
1062 /* Only remap normal extents. */
1063 real_extent = (irec->br_startblock != HOLESTARTBLOCK &&
1064 irec->br_startblock != DELAYSTARTBLOCK &&
1065 !ISUNWRITTEN(irec));
1067 /* No reflinking if we're low on space */
1069 error = xfs_reflink_ag_has_free_space(mp,
1070 XFS_FSB_TO_AGNO(mp, irec->br_startblock));
1075 /* Start a rolling transaction to switch the mappings */
1076 resblks = XFS_EXTENTADD_SPACE_RES(ip->i_mount, XFS_DATA_FORK);
1077 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, resblks, 0, 0, &tp);
1081 xfs_ilock(ip, XFS_ILOCK_EXCL);
1082 xfs_trans_ijoin(tp, ip, 0);
1084 /* If we're not just clearing space, then do we have enough quota? */
1086 error = xfs_trans_reserve_quota_nblks(tp, ip,
1087 irec->br_blockcount, 0, XFS_QMOPT_RES_REGBLKS);
1092 trace_xfs_reflink_remap(ip, irec->br_startoff,
1093 irec->br_blockcount, irec->br_startblock);
1095 /* Unmap the old blocks in the data fork. */
1098 xfs_defer_init(&dfops, &firstfsb);
1099 error = __xfs_bunmapi(tp, ip, destoff, &rlen, 0, 1,
1105 * Trim the extent to whatever got unmapped.
1106 * Remember, bunmapi works backwards.
1108 uirec.br_startblock = irec->br_startblock + rlen;
1109 uirec.br_startoff = irec->br_startoff + rlen;
1110 uirec.br_blockcount = unmap_len - rlen;
1113 /* If this isn't a real mapping, we're done. */
1114 if (!real_extent || uirec.br_blockcount == 0)
1117 trace_xfs_reflink_remap(ip, uirec.br_startoff,
1118 uirec.br_blockcount, uirec.br_startblock);
1120 /* Update the refcount tree */
1121 error = xfs_refcount_increase_extent(mp, &dfops, &uirec);
1125 /* Map the new blocks into the data fork. */
1126 error = xfs_bmap_map_extent(mp, &dfops, ip, &uirec);
1130 /* Update quota accounting. */
1131 xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_BCOUNT,
1132 uirec.br_blockcount);
1134 /* Update dest isize if needed. */
1135 newlen = XFS_FSB_TO_B(mp,
1136 uirec.br_startoff + uirec.br_blockcount);
1137 newlen = min_t(xfs_off_t, newlen, new_isize);
1138 if (newlen > i_size_read(VFS_I(ip))) {
1139 trace_xfs_reflink_update_inode_size(ip, newlen);
1140 i_size_write(VFS_I(ip), newlen);
1141 ip->i_d.di_size = newlen;
1142 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1146 /* Process all the deferred stuff. */
1147 error = xfs_defer_finish(&tp, &dfops, ip);
1152 error = xfs_trans_commit(tp);
1153 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1159 xfs_defer_cancel(&dfops);
1161 xfs_trans_cancel(tp);
1162 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1164 trace_xfs_reflink_remap_extent_error(ip, error, _RET_IP_);
1169 * Iteratively remap one file's extents (and holes) to another's.
1172 xfs_reflink_remap_blocks(
1173 struct xfs_inode *src,
1174 xfs_fileoff_t srcoff,
1175 struct xfs_inode *dest,
1176 xfs_fileoff_t destoff,
1178 xfs_off_t new_isize)
1180 struct xfs_bmbt_irec imap;
1183 xfs_filblks_t range_len;
1185 /* drange = (destoff, destoff + len); srange = (srcoff, srcoff + len) */
1187 trace_xfs_reflink_remap_blocks_loop(src, srcoff, len,
1189 /* Read extent from the source file */
1191 xfs_ilock(src, XFS_ILOCK_EXCL);
1192 error = xfs_bmapi_read(src, srcoff, len, &imap, &nimaps, 0);
1193 xfs_iunlock(src, XFS_ILOCK_EXCL);
1196 ASSERT(nimaps == 1);
1198 trace_xfs_reflink_remap_imap(src, srcoff, len, XFS_IO_OVERWRITE,
1201 /* Translate imap into the destination file. */
1202 range_len = imap.br_startoff + imap.br_blockcount - srcoff;
1203 imap.br_startoff += destoff - srcoff;
1205 /* Clear dest from destoff to the end of imap and map it in. */
1206 error = xfs_reflink_remap_extent(dest, &imap, destoff,
1211 if (fatal_signal_pending(current)) {
1216 /* Advance drange/srange */
1217 srcoff += range_len;
1218 destoff += range_len;
1225 trace_xfs_reflink_remap_blocks_error(dest, error, _RET_IP_);
1230 * Link a range of blocks from one file to another.
1233 xfs_reflink_remap_range(
1234 struct file *file_in,
1236 struct file *file_out,
1241 struct inode *inode_in = file_inode(file_in);
1242 struct xfs_inode *src = XFS_I(inode_in);
1243 struct inode *inode_out = file_inode(file_out);
1244 struct xfs_inode *dest = XFS_I(inode_out);
1245 struct xfs_mount *mp = src->i_mount;
1246 bool same_inode = (inode_in == inode_out);
1247 xfs_fileoff_t sfsbno, dfsbno;
1248 xfs_filblks_t fsblen;
1249 xfs_extlen_t cowextsize;
1252 if (!xfs_sb_version_hasreflink(&mp->m_sb))
1255 if (XFS_FORCED_SHUTDOWN(mp))
1258 /* Lock both files against IO */
1259 lock_two_nondirectories(inode_in, inode_out);
1261 xfs_ilock(src, XFS_MMAPLOCK_EXCL);
1263 xfs_lock_two_inodes(src, dest, XFS_MMAPLOCK_EXCL);
1265 /* Check file eligibility and prepare for block sharing. */
1267 /* Don't reflink realtime inodes */
1268 if (XFS_IS_REALTIME_INODE(src) || XFS_IS_REALTIME_INODE(dest))
1271 /* Don't share DAX file data for now. */
1272 if (IS_DAX(inode_in) || IS_DAX(inode_out))
1275 ret = vfs_clone_file_prep_inodes(inode_in, pos_in, inode_out, pos_out,
1280 trace_xfs_reflink_remap_range(src, pos_in, len, dest, pos_out);
1282 /* Set flags and remap blocks. */
1283 ret = xfs_reflink_set_inode_flag(src, dest);
1287 dfsbno = XFS_B_TO_FSBT(mp, pos_out);
1288 sfsbno = XFS_B_TO_FSBT(mp, pos_in);
1289 fsblen = XFS_B_TO_FSB(mp, len);
1290 ret = xfs_reflink_remap_blocks(src, sfsbno, dest, dfsbno, fsblen,
1295 /* Zap any page cache for the destination file's range. */
1296 truncate_inode_pages_range(&inode_out->i_data, pos_out,
1297 PAGE_ALIGN(pos_out + len) - 1);
1300 * Carry the cowextsize hint from src to dest if we're sharing the
1301 * entire source file to the entire destination file, the source file
1302 * has a cowextsize hint, and the destination file does not.
1305 if (pos_in == 0 && len == i_size_read(inode_in) &&
1306 (src->i_d.di_flags2 & XFS_DIFLAG2_COWEXTSIZE) &&
1307 pos_out == 0 && len >= i_size_read(inode_out) &&
1308 !(dest->i_d.di_flags2 & XFS_DIFLAG2_COWEXTSIZE))
1309 cowextsize = src->i_d.di_cowextsize;
1311 ret = xfs_reflink_update_dest(dest, pos_out + len, cowextsize,
1315 xfs_iunlock(src, XFS_MMAPLOCK_EXCL);
1317 xfs_iunlock(dest, XFS_MMAPLOCK_EXCL);
1318 unlock_two_nondirectories(inode_in, inode_out);
1320 trace_xfs_reflink_remap_range_error(dest, ret, _RET_IP_);
1325 * The user wants to preemptively CoW all shared blocks in this file,
1326 * which enables us to turn off the reflink flag. Iterate all
1327 * extents which are not prealloc/delalloc to see which ranges are
1328 * mentioned in the refcount tree, then read those blocks into the
1329 * pagecache, dirty them, fsync them back out, and then we can update
1330 * the inode flag. What happens if we run out of memory? :)
1333 xfs_reflink_dirty_extents(
1334 struct xfs_inode *ip,
1339 struct xfs_mount *mp = ip->i_mount;
1340 xfs_agnumber_t agno;
1341 xfs_agblock_t agbno;
1347 struct xfs_bmbt_irec map[2];
1351 while (end - fbno > 0) {
1354 * Look for extents in the file. Skip holes, delalloc, or
1355 * unwritten extents; they can't be reflinked.
1357 error = xfs_bmapi_read(ip, fbno, end - fbno, map, &nmaps, 0);
1362 if (map[0].br_startblock == HOLESTARTBLOCK ||
1363 map[0].br_startblock == DELAYSTARTBLOCK ||
1364 ISUNWRITTEN(&map[0]))
1368 while (map[1].br_blockcount) {
1369 agno = XFS_FSB_TO_AGNO(mp, map[1].br_startblock);
1370 agbno = XFS_FSB_TO_AGBNO(mp, map[1].br_startblock);
1371 aglen = map[1].br_blockcount;
1373 error = xfs_reflink_find_shared(mp, agno, agbno, aglen,
1374 &rbno, &rlen, true);
1377 if (rbno == NULLAGBLOCK)
1380 /* Dirty the pages */
1381 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1382 fpos = XFS_FSB_TO_B(mp, map[1].br_startoff +
1384 flen = XFS_FSB_TO_B(mp, rlen);
1385 if (fpos + flen > isize)
1386 flen = isize - fpos;
1387 error = iomap_file_dirty(VFS_I(ip), fpos, flen,
1389 xfs_ilock(ip, XFS_ILOCK_EXCL);
1393 map[1].br_blockcount -= (rbno - agbno + rlen);
1394 map[1].br_startoff += (rbno - agbno + rlen);
1395 map[1].br_startblock += (rbno - agbno + rlen);
1399 fbno = map[0].br_startoff + map[0].br_blockcount;
1405 /* Clear the inode reflink flag if there are no shared extents. */
1407 xfs_reflink_clear_inode_flag(
1408 struct xfs_inode *ip,
1409 struct xfs_trans **tpp)
1411 struct xfs_mount *mp = ip->i_mount;
1414 xfs_agnumber_t agno;
1415 xfs_agblock_t agbno;
1419 struct xfs_bmbt_irec map;
1423 ASSERT(xfs_is_reflink_inode(ip));
1426 end = XFS_B_TO_FSB(mp, i_size_read(VFS_I(ip)));
1427 while (end - fbno > 0) {
1430 * Look for extents in the file. Skip holes, delalloc, or
1431 * unwritten extents; they can't be reflinked.
1433 error = xfs_bmapi_read(ip, fbno, end - fbno, &map, &nmaps, 0);
1438 if (map.br_startblock == HOLESTARTBLOCK ||
1439 map.br_startblock == DELAYSTARTBLOCK ||
1443 agno = XFS_FSB_TO_AGNO(mp, map.br_startblock);
1444 agbno = XFS_FSB_TO_AGBNO(mp, map.br_startblock);
1445 aglen = map.br_blockcount;
1447 error = xfs_reflink_find_shared(mp, agno, agbno, aglen,
1448 &rbno, &rlen, false);
1451 /* Is there still a shared block here? */
1452 if (rbno != NULLAGBLOCK)
1455 fbno = map.br_startoff + map.br_blockcount;
1459 * We didn't find any shared blocks so turn off the reflink flag.
1460 * First, get rid of any leftover CoW mappings.
1462 error = xfs_reflink_cancel_cow_blocks(ip, tpp, 0, NULLFILEOFF, true);
1466 /* Clear the inode flag. */
1467 trace_xfs_reflink_unset_inode_flag(ip);
1468 ip->i_d.di_flags2 &= ~XFS_DIFLAG2_REFLINK;
1469 xfs_inode_clear_cowblocks_tag(ip);
1470 xfs_trans_ijoin(*tpp, ip, 0);
1471 xfs_trans_log_inode(*tpp, ip, XFS_ILOG_CORE);
1477 * Clear the inode reflink flag if there are no shared extents and the size
1481 xfs_reflink_try_clear_inode_flag(
1482 struct xfs_inode *ip)
1484 struct xfs_mount *mp = ip->i_mount;
1485 struct xfs_trans *tp;
1488 /* Start a rolling transaction to remove the mappings */
1489 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, 0, 0, 0, &tp);
1493 xfs_ilock(ip, XFS_ILOCK_EXCL);
1494 xfs_trans_ijoin(tp, ip, 0);
1496 error = xfs_reflink_clear_inode_flag(ip, &tp);
1500 error = xfs_trans_commit(tp);
1504 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1507 xfs_trans_cancel(tp);
1509 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1514 * Pre-COW all shared blocks within a given byte range of a file and turn off
1515 * the reflink flag if we unshare all of the file's blocks.
1518 xfs_reflink_unshare(
1519 struct xfs_inode *ip,
1523 struct xfs_mount *mp = ip->i_mount;
1529 if (!xfs_is_reflink_inode(ip))
1532 trace_xfs_reflink_unshare(ip, offset, len);
1534 inode_dio_wait(VFS_I(ip));
1536 /* Try to CoW the selected ranges */
1537 xfs_ilock(ip, XFS_ILOCK_EXCL);
1538 fbno = XFS_B_TO_FSBT(mp, offset);
1539 isize = i_size_read(VFS_I(ip));
1540 end = XFS_B_TO_FSB(mp, offset + len);
1541 error = xfs_reflink_dirty_extents(ip, fbno, end, isize);
1544 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1546 /* Wait for the IO to finish */
1547 error = filemap_write_and_wait(VFS_I(ip)->i_mapping);
1551 /* Turn off the reflink flag if possible. */
1552 error = xfs_reflink_try_clear_inode_flag(ip);
1559 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1561 trace_xfs_reflink_unshare_error(ip, error, _RET_IP_);