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);
175 cur = xfs_refcountbt_init_cursor(mp, NULL, agbp, agno, NULL);
177 error = xfs_refcount_find_shared(cur, agbno, aglen, fbno, flen,
180 xfs_btree_del_cursor(cur, error ? XFS_BTREE_ERROR : XFS_BTREE_NOERROR);
187 * Trim the mapping to the next block where there's a change in the
188 * shared/unshared status. More specifically, this means that we
189 * find the lowest-numbered extent of shared blocks that coincides with
190 * the given block mapping. If the shared extent overlaps the start of
191 * the mapping, trim the mapping to the end of the shared extent. If
192 * the shared region intersects the mapping, trim the mapping to the
193 * start of the shared extent. If there are no shared regions that
194 * overlap, just return the original extent.
197 xfs_reflink_trim_around_shared(
198 struct xfs_inode *ip,
199 struct xfs_bmbt_irec *irec,
210 /* Holes, unwritten, and delalloc extents cannot be shared */
211 if (!xfs_is_reflink_inode(ip) ||
213 irec->br_startblock == HOLESTARTBLOCK ||
214 irec->br_startblock == DELAYSTARTBLOCK ||
215 isnullstartblock(irec->br_startblock)) {
220 trace_xfs_reflink_trim_around_shared(ip, irec);
222 agno = XFS_FSB_TO_AGNO(ip->i_mount, irec->br_startblock);
223 agbno = XFS_FSB_TO_AGBNO(ip->i_mount, irec->br_startblock);
224 aglen = irec->br_blockcount;
226 error = xfs_reflink_find_shared(ip->i_mount, agno, agbno,
227 aglen, &fbno, &flen, true);
231 *shared = *trimmed = false;
232 if (fbno == NULLAGBLOCK) {
233 /* No shared blocks at all. */
235 } else if (fbno == agbno) {
237 * The start of this extent is shared. Truncate the
238 * mapping at the end of the shared region so that a
239 * subsequent iteration starts at the start of the
242 irec->br_blockcount = flen;
249 * There's a shared extent midway through this extent.
250 * Truncate the mapping at the start of the shared
251 * extent so that a subsequent iteration starts at the
252 * start of the shared region.
254 irec->br_blockcount = fbno - agbno;
261 * Trim the passed in imap to the next shared/unshared extent boundary, and
262 * if imap->br_startoff points to a shared extent reserve space for it in the
263 * COW fork. In this case *shared is set to true, else to false.
265 * Note that imap will always contain the block numbers for the existing blocks
266 * in the data fork, as the upper layers need them for read-modify-write
270 xfs_reflink_reserve_cow(
271 struct xfs_inode *ip,
272 struct xfs_bmbt_irec *imap,
275 struct xfs_ifork *ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
276 struct xfs_bmbt_irec got;
278 bool eof = false, trimmed;
282 * Search the COW fork extent list first. This serves two purposes:
283 * first this implement the speculative preallocation using cowextisze,
284 * so that we also unshared block adjacent to shared blocks instead
285 * of just the shared blocks themselves. Second the lookup in the
286 * extent list is generally faster than going out to the shared extent
290 if (!xfs_iext_lookup_extent(ip, ifp, imap->br_startoff, &idx, &got))
292 if (!eof && got.br_startoff <= imap->br_startoff) {
293 trace_xfs_reflink_cow_found(ip, imap);
294 xfs_trim_extent(imap, got.br_startoff, got.br_blockcount);
300 /* Trim the mapping to the nearest shared extent boundary. */
301 error = xfs_reflink_trim_around_shared(ip, imap, shared, &trimmed);
305 /* Not shared? Just report the (potentially capped) extent. */
310 * Fork all the shared blocks from our write offset until the end of
313 error = xfs_qm_dqattach_locked(ip, 0);
317 error = xfs_bmapi_reserve_delalloc(ip, XFS_COW_FORK, imap->br_startoff,
318 imap->br_blockcount, 0, &got, &idx, eof);
319 if (error == -ENOSPC || error == -EDQUOT)
320 trace_xfs_reflink_cow_enospc(ip, imap);
324 trace_xfs_reflink_cow_alloc(ip, &got);
328 /* Convert part of an unwritten CoW extent to a real one. */
330 xfs_reflink_convert_cow_extent(
331 struct xfs_inode *ip,
332 struct xfs_bmbt_irec *imap,
333 xfs_fileoff_t offset_fsb,
334 xfs_filblks_t count_fsb,
335 struct xfs_defer_ops *dfops)
337 struct xfs_bmbt_irec irec = *imap;
338 xfs_fsblock_t first_block = NULLFSBLOCK;
341 if (imap->br_state == XFS_EXT_NORM)
344 xfs_trim_extent(&irec, offset_fsb, count_fsb);
345 trace_xfs_reflink_convert_cow(ip, &irec);
346 if (irec.br_blockcount == 0)
348 return xfs_bmapi_write(NULL, ip, irec.br_startoff, irec.br_blockcount,
349 XFS_BMAPI_COWFORK | XFS_BMAPI_CONVERT, &first_block,
350 0, &irec, &nimaps, dfops);
353 /* Convert all of the unwritten CoW extents in a file's range to real ones. */
355 xfs_reflink_convert_cow(
356 struct xfs_inode *ip,
360 struct xfs_bmbt_irec got;
361 struct xfs_defer_ops dfops;
362 struct xfs_mount *mp = ip->i_mount;
363 struct xfs_ifork *ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
364 xfs_fileoff_t offset_fsb = XFS_B_TO_FSBT(mp, offset);
365 xfs_fileoff_t end_fsb = XFS_B_TO_FSB(mp, offset + count);
370 xfs_ilock(ip, XFS_ILOCK_EXCL);
372 /* Convert all the extents to real from unwritten. */
373 for (found = xfs_iext_lookup_extent(ip, ifp, offset_fsb, &idx, &got);
374 found && got.br_startoff < end_fsb;
375 found = xfs_iext_get_extent(ifp, ++idx, &got)) {
376 error = xfs_reflink_convert_cow_extent(ip, &got, offset_fsb,
377 end_fsb - offset_fsb, &dfops);
383 xfs_iunlock(ip, XFS_ILOCK_EXCL);
387 /* Allocate all CoW reservations covering a range of blocks in a file. */
389 __xfs_reflink_allocate_cow(
390 struct xfs_inode *ip,
391 xfs_fileoff_t *offset_fsb,
392 xfs_fileoff_t end_fsb)
394 struct xfs_mount *mp = ip->i_mount;
395 struct xfs_bmbt_irec imap;
396 struct xfs_defer_ops dfops;
397 struct xfs_trans *tp;
398 xfs_fsblock_t first_block;
399 int nimaps = 1, error;
402 xfs_defer_init(&dfops, &first_block);
404 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, 0, 0,
405 XFS_TRANS_RESERVE, &tp);
409 xfs_ilock(ip, XFS_ILOCK_EXCL);
411 /* Read extent from the source file. */
413 error = xfs_bmapi_read(ip, *offset_fsb, end_fsb - *offset_fsb,
419 /* Make sure there's a CoW reservation for it. */
420 error = xfs_reflink_reserve_cow(ip, &imap, &shared);
422 goto out_trans_cancel;
425 *offset_fsb = imap.br_startoff + imap.br_blockcount;
426 goto out_trans_cancel;
429 /* Allocate the entire reservation as unwritten blocks. */
430 xfs_trans_ijoin(tp, ip, 0);
431 error = xfs_bmapi_write(tp, ip, imap.br_startoff, imap.br_blockcount,
432 XFS_BMAPI_COWFORK | XFS_BMAPI_PREALLOC, &first_block,
433 XFS_EXTENTADD_SPACE_RES(mp, XFS_DATA_FORK),
434 &imap, &nimaps, &dfops);
436 goto out_trans_cancel;
439 error = xfs_defer_finish(&tp, &dfops, NULL);
441 goto out_trans_cancel;
443 error = xfs_trans_commit(tp);
445 *offset_fsb = imap.br_startoff + imap.br_blockcount;
447 xfs_iunlock(ip, XFS_ILOCK_EXCL);
450 xfs_defer_cancel(&dfops);
451 xfs_trans_cancel(tp);
455 /* Allocate all CoW reservations covering a part of a file. */
457 xfs_reflink_allocate_cow_range(
458 struct xfs_inode *ip,
462 struct xfs_mount *mp = ip->i_mount;
463 xfs_fileoff_t offset_fsb = XFS_B_TO_FSBT(mp, offset);
464 xfs_fileoff_t end_fsb = XFS_B_TO_FSB(mp, offset + count);
467 ASSERT(xfs_is_reflink_inode(ip));
469 trace_xfs_reflink_allocate_cow_range(ip, offset, count);
472 * Make sure that the dquots are there.
474 error = xfs_qm_dqattach(ip, 0);
478 while (offset_fsb < end_fsb) {
479 error = __xfs_reflink_allocate_cow(ip, &offset_fsb, end_fsb);
481 trace_xfs_reflink_allocate_cow_range_error(ip, error,
487 /* Convert the CoW extents to regular. */
488 return xfs_reflink_convert_cow(ip, offset, count);
492 * Find the CoW reservation (and whether or not it needs block allocation)
493 * for a given byte offset of a file.
496 xfs_reflink_find_cow_mapping(
497 struct xfs_inode *ip,
499 struct xfs_bmbt_irec *imap,
502 struct xfs_bmbt_irec irec;
503 struct xfs_ifork *ifp;
504 struct xfs_bmbt_rec_host *gotp;
508 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL | XFS_ILOCK_SHARED));
509 ASSERT(xfs_is_reflink_inode(ip));
511 /* Find the extent in the CoW fork. */
512 ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
513 bno = XFS_B_TO_FSBT(ip->i_mount, offset);
514 gotp = xfs_iext_bno_to_ext(ifp, bno, &idx);
518 xfs_bmbt_get_all(gotp, &irec);
519 if (bno >= irec.br_startoff + irec.br_blockcount ||
520 bno < irec.br_startoff)
523 trace_xfs_reflink_find_cow_mapping(ip, offset, 1, XFS_IO_OVERWRITE,
526 /* If it's still delalloc, we must allocate later. */
528 *need_alloc = !!(isnullstartblock(irec.br_startblock));
534 * Trim an extent to end at the next CoW reservation past offset_fsb.
537 xfs_reflink_trim_irec_to_next_cow(
538 struct xfs_inode *ip,
539 xfs_fileoff_t offset_fsb,
540 struct xfs_bmbt_irec *imap)
542 struct xfs_bmbt_irec irec;
543 struct xfs_ifork *ifp;
544 struct xfs_bmbt_rec_host *gotp;
547 if (!xfs_is_reflink_inode(ip))
550 /* Find the extent in the CoW fork. */
551 ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
552 gotp = xfs_iext_bno_to_ext(ifp, offset_fsb, &idx);
555 xfs_bmbt_get_all(gotp, &irec);
557 /* This is the extent before; try sliding up one. */
558 if (irec.br_startoff < offset_fsb) {
560 if (idx >= xfs_iext_count(ifp))
562 gotp = xfs_iext_get_ext(ifp, idx);
563 xfs_bmbt_get_all(gotp, &irec);
566 if (irec.br_startoff >= imap->br_startoff + imap->br_blockcount)
569 imap->br_blockcount = irec.br_startoff - imap->br_startoff;
570 trace_xfs_reflink_trim_irec(ip, imap);
576 * Cancel CoW reservations for some block range of an inode.
578 * If cancel_real is true this function cancels all COW fork extents for the
579 * inode; if cancel_real is false, real extents are not cleared.
582 xfs_reflink_cancel_cow_blocks(
583 struct xfs_inode *ip,
584 struct xfs_trans **tpp,
585 xfs_fileoff_t offset_fsb,
586 xfs_fileoff_t end_fsb,
589 struct xfs_ifork *ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
590 struct xfs_bmbt_irec got, prev, del;
592 xfs_fsblock_t firstfsb;
593 struct xfs_defer_ops dfops;
594 int error = 0, eof = 0;
596 if (!xfs_is_reflink_inode(ip))
599 xfs_bmap_search_extents(ip, offset_fsb, XFS_COW_FORK, &eof, &idx,
604 while (got.br_startoff < end_fsb) {
606 xfs_trim_extent(&del, offset_fsb, end_fsb - offset_fsb);
607 trace_xfs_reflink_cancel_cow(ip, &del);
609 if (isnullstartblock(del.br_startblock)) {
610 error = xfs_bmap_del_extent_delay(ip, XFS_COW_FORK,
614 } else if (del.br_state == XFS_EXT_UNWRITTEN || cancel_real) {
615 xfs_trans_ijoin(*tpp, ip, 0);
616 xfs_defer_init(&dfops, &firstfsb);
618 /* Free the CoW orphan record. */
619 error = xfs_refcount_free_cow_extent(ip->i_mount,
620 &dfops, del.br_startblock,
625 xfs_bmap_add_free(ip->i_mount, &dfops,
626 del.br_startblock, del.br_blockcount,
629 /* Update quota accounting */
630 xfs_trans_mod_dquot_byino(*tpp, ip, XFS_TRANS_DQ_BCOUNT,
631 -(long)del.br_blockcount);
633 /* Roll the transaction */
634 error = xfs_defer_finish(tpp, &dfops, ip);
636 xfs_defer_cancel(&dfops);
640 /* Remove the mapping from the CoW fork. */
641 xfs_bmap_del_extent_cow(ip, &idx, &got, &del);
644 if (++idx >= xfs_iext_count(ifp))
646 xfs_bmbt_get_all(xfs_iext_get_ext(ifp, idx), &got);
649 /* clear tag if cow fork is emptied */
651 xfs_inode_clear_cowblocks_tag(ip);
657 * Cancel CoW reservations for some byte range of an inode.
659 * If cancel_real is true this function cancels all COW fork extents for the
660 * inode; if cancel_real is false, real extents are not cleared.
663 xfs_reflink_cancel_cow_range(
664 struct xfs_inode *ip,
669 struct xfs_trans *tp;
670 xfs_fileoff_t offset_fsb;
671 xfs_fileoff_t end_fsb;
674 trace_xfs_reflink_cancel_cow_range(ip, offset, count);
675 ASSERT(xfs_is_reflink_inode(ip));
677 offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
678 if (count == NULLFILEOFF)
679 end_fsb = NULLFILEOFF;
681 end_fsb = XFS_B_TO_FSB(ip->i_mount, offset + count);
683 /* Start a rolling transaction to remove the mappings */
684 error = xfs_trans_alloc(ip->i_mount, &M_RES(ip->i_mount)->tr_write,
689 xfs_ilock(ip, XFS_ILOCK_EXCL);
690 xfs_trans_ijoin(tp, ip, 0);
692 /* Scrape out the old CoW reservations */
693 error = xfs_reflink_cancel_cow_blocks(ip, &tp, offset_fsb, end_fsb,
698 error = xfs_trans_commit(tp);
700 xfs_iunlock(ip, XFS_ILOCK_EXCL);
704 xfs_trans_cancel(tp);
705 xfs_iunlock(ip, XFS_ILOCK_EXCL);
707 trace_xfs_reflink_cancel_cow_range_error(ip, error, _RET_IP_);
712 * Remap parts of a file's data fork after a successful CoW.
716 struct xfs_inode *ip,
720 struct xfs_ifork *ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
721 struct xfs_bmbt_irec got, prev, del;
722 struct xfs_trans *tp;
723 xfs_fileoff_t offset_fsb;
724 xfs_fileoff_t end_fsb;
725 xfs_fsblock_t firstfsb;
726 struct xfs_defer_ops dfops;
728 unsigned int resblks;
732 trace_xfs_reflink_end_cow(ip, offset, count);
734 /* No COW extents? That's easy! */
735 if (ifp->if_bytes == 0)
738 offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
739 end_fsb = XFS_B_TO_FSB(ip->i_mount, offset + count);
742 * Start a rolling transaction to switch the mappings. We're
743 * unlikely ever to have to remap 16T worth of single-block
744 * extents, so just cap the worst case extent count to 2^32-1.
745 * Stick a warning in just in case, and avoid 64-bit division.
747 BUILD_BUG_ON(MAX_RW_COUNT > UINT_MAX);
748 if (end_fsb - offset_fsb > UINT_MAX) {
749 error = -EFSCORRUPTED;
750 xfs_force_shutdown(ip->i_mount, SHUTDOWN_CORRUPT_INCORE);
754 resblks = XFS_NEXTENTADD_SPACE_RES(ip->i_mount,
755 (unsigned int)(end_fsb - offset_fsb),
757 error = xfs_trans_alloc(ip->i_mount, &M_RES(ip->i_mount)->tr_write,
762 xfs_ilock(ip, XFS_ILOCK_EXCL);
763 xfs_trans_ijoin(tp, ip, 0);
765 xfs_bmap_search_extents(ip, end_fsb - 1, XFS_COW_FORK, &eof, &idx,
768 /* If there is a hole at end_fsb - 1 go to the previous extent */
769 if (eof || got.br_startoff > end_fsb) {
771 * In case of racing, overlapping AIO writes no COW extents
772 * might be left by the time I/O completes for the loser of
773 * the race. In that case we are done.
777 xfs_bmbt_get_all(xfs_iext_get_ext(ifp, --idx), &got);
780 /* Walk backwards until we're out of the I/O range... */
781 while (got.br_startoff + got.br_blockcount > offset_fsb) {
783 xfs_trim_extent(&del, offset_fsb, end_fsb - offset_fsb);
785 /* Extent delete may have bumped idx forward */
786 if (!del.br_blockcount) {
791 ASSERT(!isnullstartblock(got.br_startblock));
794 * Don't remap unwritten extents; these are
795 * speculatively preallocated CoW extents that have been
796 * allocated but have not yet been involved in a write.
798 if (got.br_state == XFS_EXT_UNWRITTEN) {
803 /* Unmap the old blocks in the data fork. */
804 xfs_defer_init(&dfops, &firstfsb);
805 rlen = del.br_blockcount;
806 error = __xfs_bunmapi(tp, ip, del.br_startoff, &rlen, 0, 1,
811 /* Trim the extent to whatever got unmapped. */
813 xfs_trim_extent(&del, del.br_startoff + rlen,
814 del.br_blockcount - rlen);
816 trace_xfs_reflink_cow_remap(ip, &del);
818 /* Free the CoW orphan record. */
819 error = xfs_refcount_free_cow_extent(tp->t_mountp, &dfops,
820 del.br_startblock, del.br_blockcount);
824 /* Map the new blocks into the data fork. */
825 error = xfs_bmap_map_extent(tp->t_mountp, &dfops, ip, &del);
829 /* Remove the mapping from the CoW fork. */
830 xfs_bmap_del_extent_cow(ip, &idx, &got, &del);
832 error = xfs_defer_finish(&tp, &dfops, ip);
839 xfs_bmbt_get_all(xfs_iext_get_ext(ifp, idx), &got);
842 error = xfs_trans_commit(tp);
843 xfs_iunlock(ip, XFS_ILOCK_EXCL);
849 xfs_defer_cancel(&dfops);
851 xfs_trans_cancel(tp);
852 xfs_iunlock(ip, XFS_ILOCK_EXCL);
854 trace_xfs_reflink_end_cow_error(ip, error, _RET_IP_);
859 * Free leftover CoW reservations that didn't get cleaned out.
862 xfs_reflink_recover_cow(
863 struct xfs_mount *mp)
868 if (!xfs_sb_version_hasreflink(&mp->m_sb))
871 for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
872 error = xfs_refcount_recover_cow_leftovers(mp, agno);
881 * Reflinking (Block) Ranges of Two Files Together
883 * First, ensure that the reflink flag is set on both inodes. The flag is an
884 * optimization to avoid unnecessary refcount btree lookups in the write path.
886 * Now we can iteratively remap the range of extents (and holes) in src to the
887 * corresponding ranges in dest. Let drange and srange denote the ranges of
888 * logical blocks in dest and src touched by the reflink operation.
890 * While the length of drange is greater than zero,
891 * - Read src's bmbt at the start of srange ("imap")
892 * - If imap doesn't exist, make imap appear to start at the end of srange
894 * - If imap starts before srange, advance imap to start at srange.
895 * - If imap goes beyond srange, truncate imap to end at the end of srange.
896 * - Punch (imap start - srange start + imap len) blocks from dest at
897 * offset (drange start).
898 * - If imap points to a real range of pblks,
899 * > Increase the refcount of the imap's pblks
900 * > Map imap's pblks into dest at the offset
901 * (drange start + imap start - srange start)
902 * - Advance drange and srange by (imap start - srange start + imap len)
904 * Finally, if the reflink made dest longer, update both the in-core and
905 * on-disk file sizes.
907 * ASCII Art Demonstration:
909 * Let's say we want to reflink this source file:
911 * ----SSSSSSS-SSSSS----SSSSSS (src file)
912 * <-------------------->
914 * into this destination file:
916 * --DDDDDDDDDDDDDDDDDDD--DDD (dest file)
917 * <-------------------->
918 * '-' means a hole, and 'S' and 'D' are written blocks in the src and dest.
919 * Observe that the range has different logical offsets in either file.
921 * Consider that the first extent in the source file doesn't line up with our
922 * reflink range. Unmapping and remapping are separate operations, so we can
923 * unmap more blocks from the destination file than we remap.
925 * ----SSSSSSS-SSSSS----SSSSSS
927 * --DDDDD---------DDDDD--DDD
930 * Now remap the source extent into the destination file:
932 * ----SSSSSSS-SSSSS----SSSSSS
934 * --DDDDD--SSSSSSSDDDDD--DDD
937 * Do likewise with the second hole and extent in our range. Holes in the
938 * unmap range don't affect our operation.
940 * ----SSSSSSS-SSSSS----SSSSSS
942 * --DDDDD--SSSSSSS-SSSSS-DDD
945 * Finally, unmap and remap part of the third extent. This will increase the
946 * size of the destination file.
948 * ----SSSSSSS-SSSSS----SSSSSS
950 * --DDDDD--SSSSSSS-SSSSS----SSS
953 * Once we update the destination file's i_size, we're done.
957 * Ensure the reflink bit is set in both inodes.
960 xfs_reflink_set_inode_flag(
961 struct xfs_inode *src,
962 struct xfs_inode *dest)
964 struct xfs_mount *mp = src->i_mount;
966 struct xfs_trans *tp;
968 if (xfs_is_reflink_inode(src) && xfs_is_reflink_inode(dest))
971 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
975 /* Lock both files against IO */
976 if (src->i_ino == dest->i_ino)
977 xfs_ilock(src, XFS_ILOCK_EXCL);
979 xfs_lock_two_inodes(src, dest, XFS_ILOCK_EXCL);
981 if (!xfs_is_reflink_inode(src)) {
982 trace_xfs_reflink_set_inode_flag(src);
983 xfs_trans_ijoin(tp, src, XFS_ILOCK_EXCL);
984 src->i_d.di_flags2 |= XFS_DIFLAG2_REFLINK;
985 xfs_trans_log_inode(tp, src, XFS_ILOG_CORE);
986 xfs_ifork_init_cow(src);
988 xfs_iunlock(src, XFS_ILOCK_EXCL);
990 if (src->i_ino == dest->i_ino)
993 if (!xfs_is_reflink_inode(dest)) {
994 trace_xfs_reflink_set_inode_flag(dest);
995 xfs_trans_ijoin(tp, dest, XFS_ILOCK_EXCL);
996 dest->i_d.di_flags2 |= XFS_DIFLAG2_REFLINK;
997 xfs_trans_log_inode(tp, dest, XFS_ILOG_CORE);
998 xfs_ifork_init_cow(dest);
1000 xfs_iunlock(dest, XFS_ILOCK_EXCL);
1003 error = xfs_trans_commit(tp);
1009 trace_xfs_reflink_set_inode_flag_error(dest, error, _RET_IP_);
1014 * Update destination inode size & cowextsize hint, if necessary.
1017 xfs_reflink_update_dest(
1018 struct xfs_inode *dest,
1020 xfs_extlen_t cowextsize,
1023 struct xfs_mount *mp = dest->i_mount;
1024 struct xfs_trans *tp;
1027 if (is_dedupe && newlen <= i_size_read(VFS_I(dest)) && cowextsize == 0)
1030 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
1034 xfs_ilock(dest, XFS_ILOCK_EXCL);
1035 xfs_trans_ijoin(tp, dest, XFS_ILOCK_EXCL);
1037 if (newlen > i_size_read(VFS_I(dest))) {
1038 trace_xfs_reflink_update_inode_size(dest, newlen);
1039 i_size_write(VFS_I(dest), newlen);
1040 dest->i_d.di_size = newlen;
1044 dest->i_d.di_cowextsize = cowextsize;
1045 dest->i_d.di_flags2 |= XFS_DIFLAG2_COWEXTSIZE;
1049 xfs_trans_ichgtime(tp, dest,
1050 XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1052 xfs_trans_log_inode(tp, dest, XFS_ILOG_CORE);
1054 error = xfs_trans_commit(tp);
1060 trace_xfs_reflink_update_inode_size_error(dest, error, _RET_IP_);
1065 * Do we have enough reserve in this AG to handle a reflink? The refcount
1066 * btree already reserved all the space it needs, but the rmap btree can grow
1067 * infinitely, so we won't allow more reflinks when the AG is down to the
1071 xfs_reflink_ag_has_free_space(
1072 struct xfs_mount *mp,
1073 xfs_agnumber_t agno)
1075 struct xfs_perag *pag;
1078 if (!xfs_sb_version_hasrmapbt(&mp->m_sb))
1081 pag = xfs_perag_get(mp, agno);
1082 if (xfs_ag_resv_critical(pag, XFS_AG_RESV_AGFL) ||
1083 xfs_ag_resv_critical(pag, XFS_AG_RESV_METADATA))
1090 * Unmap a range of blocks from a file, then map other blocks into the hole.
1091 * The range to unmap is (destoff : destoff + srcioff + irec->br_blockcount).
1092 * The extent irec is mapped into dest at irec->br_startoff.
1095 xfs_reflink_remap_extent(
1096 struct xfs_inode *ip,
1097 struct xfs_bmbt_irec *irec,
1098 xfs_fileoff_t destoff,
1099 xfs_off_t new_isize)
1101 struct xfs_mount *mp = ip->i_mount;
1102 struct xfs_trans *tp;
1103 xfs_fsblock_t firstfsb;
1104 unsigned int resblks;
1105 struct xfs_defer_ops dfops;
1106 struct xfs_bmbt_irec uirec;
1109 xfs_filblks_t unmap_len;
1114 unmap_len = irec->br_startoff + irec->br_blockcount - destoff;
1115 trace_xfs_reflink_punch_range(ip, destoff, unmap_len);
1117 /* Only remap normal extents. */
1118 real_extent = (irec->br_startblock != HOLESTARTBLOCK &&
1119 irec->br_startblock != DELAYSTARTBLOCK &&
1120 !ISUNWRITTEN(irec));
1122 /* No reflinking if we're low on space */
1124 error = xfs_reflink_ag_has_free_space(mp,
1125 XFS_FSB_TO_AGNO(mp, irec->br_startblock));
1130 /* Start a rolling transaction to switch the mappings */
1131 resblks = XFS_EXTENTADD_SPACE_RES(ip->i_mount, XFS_DATA_FORK);
1132 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, resblks, 0, 0, &tp);
1136 xfs_ilock(ip, XFS_ILOCK_EXCL);
1137 xfs_trans_ijoin(tp, ip, 0);
1140 * Reserve quota for this operation. We don't know if the first unmap
1141 * in the dest file will cause a bmap btree split, so we always reserve
1142 * at least enough blocks for that split. If the extent being mapped
1143 * in is written, we need to reserve quota for that too.
1145 qres = XFS_EXTENTADD_SPACE_RES(mp, XFS_DATA_FORK);
1147 qres += irec->br_blockcount;
1148 error = xfs_trans_reserve_quota_nblks(tp, ip, qres, 0,
1149 XFS_QMOPT_RES_REGBLKS);
1153 trace_xfs_reflink_remap(ip, irec->br_startoff,
1154 irec->br_blockcount, irec->br_startblock);
1156 /* Unmap the old blocks in the data fork. */
1159 xfs_defer_init(&dfops, &firstfsb);
1160 error = __xfs_bunmapi(tp, ip, destoff, &rlen, 0, 1,
1166 * Trim the extent to whatever got unmapped.
1167 * Remember, bunmapi works backwards.
1169 uirec.br_startblock = irec->br_startblock + rlen;
1170 uirec.br_startoff = irec->br_startoff + rlen;
1171 uirec.br_blockcount = unmap_len - rlen;
1172 uirec.br_state = irec->br_state;
1175 /* If this isn't a real mapping, we're done. */
1176 if (!real_extent || uirec.br_blockcount == 0)
1179 trace_xfs_reflink_remap(ip, uirec.br_startoff,
1180 uirec.br_blockcount, uirec.br_startblock);
1182 /* Update the refcount tree */
1183 error = xfs_refcount_increase_extent(mp, &dfops, &uirec);
1187 /* Map the new blocks into the data fork. */
1188 error = xfs_bmap_map_extent(mp, &dfops, ip, &uirec);
1192 /* Update quota accounting. */
1193 xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_BCOUNT,
1194 uirec.br_blockcount);
1196 /* Update dest isize if needed. */
1197 newlen = XFS_FSB_TO_B(mp,
1198 uirec.br_startoff + uirec.br_blockcount);
1199 newlen = min_t(xfs_off_t, newlen, new_isize);
1200 if (newlen > i_size_read(VFS_I(ip))) {
1201 trace_xfs_reflink_update_inode_size(ip, newlen);
1202 i_size_write(VFS_I(ip), newlen);
1203 ip->i_d.di_size = newlen;
1204 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1208 /* Process all the deferred stuff. */
1209 error = xfs_defer_finish(&tp, &dfops, ip);
1214 error = xfs_trans_commit(tp);
1215 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1221 xfs_defer_cancel(&dfops);
1223 xfs_trans_cancel(tp);
1224 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1226 trace_xfs_reflink_remap_extent_error(ip, error, _RET_IP_);
1231 * Iteratively remap one file's extents (and holes) to another's.
1234 xfs_reflink_remap_blocks(
1235 struct xfs_inode *src,
1236 xfs_fileoff_t srcoff,
1237 struct xfs_inode *dest,
1238 xfs_fileoff_t destoff,
1240 xfs_off_t new_isize)
1242 struct xfs_bmbt_irec imap;
1245 xfs_filblks_t range_len;
1247 /* drange = (destoff, destoff + len); srange = (srcoff, srcoff + len) */
1249 trace_xfs_reflink_remap_blocks_loop(src, srcoff, len,
1251 /* Read extent from the source file */
1253 xfs_ilock(src, XFS_ILOCK_EXCL);
1254 error = xfs_bmapi_read(src, srcoff, len, &imap, &nimaps, 0);
1255 xfs_iunlock(src, XFS_ILOCK_EXCL);
1258 ASSERT(nimaps == 1);
1260 trace_xfs_reflink_remap_imap(src, srcoff, len, XFS_IO_OVERWRITE,
1263 /* Translate imap into the destination file. */
1264 range_len = imap.br_startoff + imap.br_blockcount - srcoff;
1265 imap.br_startoff += destoff - srcoff;
1267 /* Clear dest from destoff to the end of imap and map it in. */
1268 error = xfs_reflink_remap_extent(dest, &imap, destoff,
1273 if (fatal_signal_pending(current)) {
1278 /* Advance drange/srange */
1279 srcoff += range_len;
1280 destoff += range_len;
1287 trace_xfs_reflink_remap_blocks_error(dest, error, _RET_IP_);
1292 * Read a page's worth of file data into the page cache. Return the page
1295 static struct page *
1297 struct inode *inode,
1300 struct address_space *mapping;
1304 n = offset >> PAGE_SHIFT;
1305 mapping = inode->i_mapping;
1306 page = read_mapping_page(mapping, n, NULL);
1309 if (!PageUptodate(page)) {
1311 return ERR_PTR(-EIO);
1318 * Compare extents of two files to see if they are the same.
1321 xfs_compare_extents(
1330 xfs_off_t dest_poff;
1333 struct page *src_page;
1334 struct page *dest_page;
1342 src_poff = srcoff & (PAGE_SIZE - 1);
1343 dest_poff = destoff & (PAGE_SIZE - 1);
1344 cmp_len = min(PAGE_SIZE - src_poff,
1345 PAGE_SIZE - dest_poff);
1346 cmp_len = min(cmp_len, len);
1347 ASSERT(cmp_len > 0);
1349 trace_xfs_reflink_compare_extents(XFS_I(src), srcoff, cmp_len,
1350 XFS_I(dest), destoff);
1352 src_page = xfs_get_page(src, srcoff);
1353 if (IS_ERR(src_page)) {
1354 error = PTR_ERR(src_page);
1357 dest_page = xfs_get_page(dest, destoff);
1358 if (IS_ERR(dest_page)) {
1359 error = PTR_ERR(dest_page);
1360 unlock_page(src_page);
1364 src_addr = kmap_atomic(src_page);
1365 dest_addr = kmap_atomic(dest_page);
1367 flush_dcache_page(src_page);
1368 flush_dcache_page(dest_page);
1370 if (memcmp(src_addr + src_poff, dest_addr + dest_poff, cmp_len))
1373 kunmap_atomic(dest_addr);
1374 kunmap_atomic(src_addr);
1375 unlock_page(dest_page);
1376 unlock_page(src_page);
1377 put_page(dest_page);
1392 trace_xfs_reflink_compare_extents_error(XFS_I(dest), error, _RET_IP_);
1397 * Link a range of blocks from one file to another.
1400 xfs_reflink_remap_range(
1401 struct file *file_in,
1403 struct file *file_out,
1408 struct inode *inode_in = file_inode(file_in);
1409 struct xfs_inode *src = XFS_I(inode_in);
1410 struct inode *inode_out = file_inode(file_out);
1411 struct xfs_inode *dest = XFS_I(inode_out);
1412 struct xfs_mount *mp = src->i_mount;
1413 loff_t bs = inode_out->i_sb->s_blocksize;
1414 bool same_inode = (inode_in == inode_out);
1415 xfs_fileoff_t sfsbno, dfsbno;
1416 xfs_filblks_t fsblen;
1417 xfs_extlen_t cowextsize;
1422 if (!xfs_sb_version_hasreflink(&mp->m_sb))
1425 if (XFS_FORCED_SHUTDOWN(mp))
1428 /* Lock both files against IO */
1430 xfs_ilock(src, XFS_IOLOCK_EXCL);
1431 xfs_ilock(src, XFS_MMAPLOCK_EXCL);
1433 xfs_lock_two_inodes(src, dest, XFS_IOLOCK_EXCL);
1434 xfs_lock_two_inodes(src, dest, XFS_MMAPLOCK_EXCL);
1437 /* Don't touch certain kinds of inodes */
1439 if (IS_IMMUTABLE(inode_out))
1443 if (IS_SWAPFILE(inode_in) || IS_SWAPFILE(inode_out))
1447 /* Don't reflink dirs, pipes, sockets... */
1449 if (S_ISDIR(inode_in->i_mode) || S_ISDIR(inode_out->i_mode))
1452 if (S_ISFIFO(inode_in->i_mode) || S_ISFIFO(inode_out->i_mode))
1454 if (!S_ISREG(inode_in->i_mode) || !S_ISREG(inode_out->i_mode))
1457 /* Don't reflink realtime inodes */
1458 if (XFS_IS_REALTIME_INODE(src) || XFS_IS_REALTIME_INODE(dest))
1461 /* Don't share DAX file data for now. */
1462 if (IS_DAX(inode_in) || IS_DAX(inode_out))
1465 /* Are we going all the way to the end? */
1466 isize = i_size_read(inode_in);
1472 /* Zero length dedupe exits immediately; reflink goes to EOF. */
1478 len = isize - pos_in;
1481 /* Ensure offsets don't wrap and the input is inside i_size */
1482 if (pos_in + len < pos_in || pos_out + len < pos_out ||
1483 pos_in + len > isize)
1486 /* Don't allow dedupe past EOF in the dest file */
1490 disize = i_size_read(inode_out);
1491 if (pos_out >= disize || pos_out + len > disize)
1495 /* If we're linking to EOF, continue to the block boundary. */
1496 if (pos_in + len == isize)
1497 blen = ALIGN(isize, bs) - pos_in;
1501 /* Only reflink if we're aligned to block boundaries */
1502 if (!IS_ALIGNED(pos_in, bs) || !IS_ALIGNED(pos_in + blen, bs) ||
1503 !IS_ALIGNED(pos_out, bs) || !IS_ALIGNED(pos_out + blen, bs))
1506 /* Don't allow overlapped reflink within the same file */
1508 if (pos_out + blen > pos_in && pos_out < pos_in + blen)
1512 /* Wait for the completion of any pending IOs on both files */
1513 inode_dio_wait(inode_in);
1515 inode_dio_wait(inode_out);
1517 ret = filemap_write_and_wait_range(inode_in->i_mapping,
1518 pos_in, pos_in + len - 1);
1522 ret = filemap_write_and_wait_range(inode_out->i_mapping,
1523 pos_out, pos_out + len - 1);
1527 trace_xfs_reflink_remap_range(src, pos_in, len, dest, pos_out);
1530 * Check that the extents are the same.
1533 bool is_same = false;
1535 ret = xfs_compare_extents(inode_in, pos_in, inode_out, pos_out,
1545 ret = xfs_reflink_set_inode_flag(src, dest);
1550 * Invalidate the page cache so that we can clear any CoW mappings
1551 * in the destination file.
1553 truncate_inode_pages_range(&inode_out->i_data, pos_out,
1554 PAGE_ALIGN(pos_out + len) - 1);
1556 dfsbno = XFS_B_TO_FSBT(mp, pos_out);
1557 sfsbno = XFS_B_TO_FSBT(mp, pos_in);
1558 fsblen = XFS_B_TO_FSB(mp, len);
1559 ret = xfs_reflink_remap_blocks(src, sfsbno, dest, dfsbno, fsblen,
1565 * Carry the cowextsize hint from src to dest if we're sharing the
1566 * entire source file to the entire destination file, the source file
1567 * has a cowextsize hint, and the destination file does not.
1570 if (pos_in == 0 && len == i_size_read(inode_in) &&
1571 (src->i_d.di_flags2 & XFS_DIFLAG2_COWEXTSIZE) &&
1572 pos_out == 0 && len >= i_size_read(inode_out) &&
1573 !(dest->i_d.di_flags2 & XFS_DIFLAG2_COWEXTSIZE))
1574 cowextsize = src->i_d.di_cowextsize;
1576 ret = xfs_reflink_update_dest(dest, pos_out + len, cowextsize,
1580 xfs_iunlock(src, XFS_MMAPLOCK_EXCL);
1581 xfs_iunlock(src, XFS_IOLOCK_EXCL);
1582 if (src->i_ino != dest->i_ino) {
1583 xfs_iunlock(dest, XFS_MMAPLOCK_EXCL);
1584 xfs_iunlock(dest, XFS_IOLOCK_EXCL);
1587 trace_xfs_reflink_remap_range_error(dest, ret, _RET_IP_);
1592 * The user wants to preemptively CoW all shared blocks in this file,
1593 * which enables us to turn off the reflink flag. Iterate all
1594 * extents which are not prealloc/delalloc to see which ranges are
1595 * mentioned in the refcount tree, then read those blocks into the
1596 * pagecache, dirty them, fsync them back out, and then we can update
1597 * the inode flag. What happens if we run out of memory? :)
1600 xfs_reflink_dirty_extents(
1601 struct xfs_inode *ip,
1606 struct xfs_mount *mp = ip->i_mount;
1607 xfs_agnumber_t agno;
1608 xfs_agblock_t agbno;
1614 struct xfs_bmbt_irec map[2];
1618 while (end - fbno > 0) {
1621 * Look for extents in the file. Skip holes, delalloc, or
1622 * unwritten extents; they can't be reflinked.
1624 error = xfs_bmapi_read(ip, fbno, end - fbno, map, &nmaps, 0);
1629 if (map[0].br_startblock == HOLESTARTBLOCK ||
1630 map[0].br_startblock == DELAYSTARTBLOCK ||
1631 ISUNWRITTEN(&map[0]))
1635 while (map[1].br_blockcount) {
1636 agno = XFS_FSB_TO_AGNO(mp, map[1].br_startblock);
1637 agbno = XFS_FSB_TO_AGBNO(mp, map[1].br_startblock);
1638 aglen = map[1].br_blockcount;
1640 error = xfs_reflink_find_shared(mp, agno, agbno, aglen,
1641 &rbno, &rlen, true);
1644 if (rbno == NULLAGBLOCK)
1647 /* Dirty the pages */
1648 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1649 fpos = XFS_FSB_TO_B(mp, map[1].br_startoff +
1651 flen = XFS_FSB_TO_B(mp, rlen);
1652 if (fpos + flen > isize)
1653 flen = isize - fpos;
1654 error = iomap_file_dirty(VFS_I(ip), fpos, flen,
1656 xfs_ilock(ip, XFS_ILOCK_EXCL);
1660 map[1].br_blockcount -= (rbno - agbno + rlen);
1661 map[1].br_startoff += (rbno - agbno + rlen);
1662 map[1].br_startblock += (rbno - agbno + rlen);
1666 fbno = map[0].br_startoff + map[0].br_blockcount;
1672 /* Clear the inode reflink flag if there are no shared extents. */
1674 xfs_reflink_clear_inode_flag(
1675 struct xfs_inode *ip,
1676 struct xfs_trans **tpp)
1678 struct xfs_mount *mp = ip->i_mount;
1681 xfs_agnumber_t agno;
1682 xfs_agblock_t agbno;
1686 struct xfs_bmbt_irec map;
1690 ASSERT(xfs_is_reflink_inode(ip));
1693 end = XFS_B_TO_FSB(mp, i_size_read(VFS_I(ip)));
1694 while (end - fbno > 0) {
1697 * Look for extents in the file. Skip holes, delalloc, or
1698 * unwritten extents; they can't be reflinked.
1700 error = xfs_bmapi_read(ip, fbno, end - fbno, &map, &nmaps, 0);
1705 if (map.br_startblock == HOLESTARTBLOCK ||
1706 map.br_startblock == DELAYSTARTBLOCK ||
1710 agno = XFS_FSB_TO_AGNO(mp, map.br_startblock);
1711 agbno = XFS_FSB_TO_AGBNO(mp, map.br_startblock);
1712 aglen = map.br_blockcount;
1714 error = xfs_reflink_find_shared(mp, agno, agbno, aglen,
1715 &rbno, &rlen, false);
1718 /* Is there still a shared block here? */
1719 if (rbno != NULLAGBLOCK)
1722 fbno = map.br_startoff + map.br_blockcount;
1726 * We didn't find any shared blocks so turn off the reflink flag.
1727 * First, get rid of any leftover CoW mappings.
1729 error = xfs_reflink_cancel_cow_blocks(ip, tpp, 0, NULLFILEOFF, true);
1733 /* Clear the inode flag. */
1734 trace_xfs_reflink_unset_inode_flag(ip);
1735 ip->i_d.di_flags2 &= ~XFS_DIFLAG2_REFLINK;
1736 xfs_inode_clear_cowblocks_tag(ip);
1737 xfs_trans_ijoin(*tpp, ip, 0);
1738 xfs_trans_log_inode(*tpp, ip, XFS_ILOG_CORE);
1744 * Clear the inode reflink flag if there are no shared extents and the size
1748 xfs_reflink_try_clear_inode_flag(
1749 struct xfs_inode *ip)
1751 struct xfs_mount *mp = ip->i_mount;
1752 struct xfs_trans *tp;
1755 /* Start a rolling transaction to remove the mappings */
1756 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, 0, 0, 0, &tp);
1760 xfs_ilock(ip, XFS_ILOCK_EXCL);
1761 xfs_trans_ijoin(tp, ip, 0);
1763 error = xfs_reflink_clear_inode_flag(ip, &tp);
1767 error = xfs_trans_commit(tp);
1771 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1774 xfs_trans_cancel(tp);
1776 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1781 * Pre-COW all shared blocks within a given byte range of a file and turn off
1782 * the reflink flag if we unshare all of the file's blocks.
1785 xfs_reflink_unshare(
1786 struct xfs_inode *ip,
1790 struct xfs_mount *mp = ip->i_mount;
1796 if (!xfs_is_reflink_inode(ip))
1799 trace_xfs_reflink_unshare(ip, offset, len);
1801 inode_dio_wait(VFS_I(ip));
1803 /* Try to CoW the selected ranges */
1804 xfs_ilock(ip, XFS_ILOCK_EXCL);
1805 fbno = XFS_B_TO_FSBT(mp, offset);
1806 isize = i_size_read(VFS_I(ip));
1807 end = XFS_B_TO_FSB(mp, offset + len);
1808 error = xfs_reflink_dirty_extents(ip, fbno, end, isize);
1811 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1813 /* Wait for the IO to finish */
1814 error = filemap_write_and_wait(VFS_I(ip)->i_mapping);
1818 /* Turn off the reflink flag if possible. */
1819 error = xfs_reflink_try_clear_inode_flag(ip);
1826 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1828 trace_xfs_reflink_unshare_error(ip, error, _RET_IP_);