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"
59 * Copy on Write of Shared Blocks
61 * XFS must preserve "the usual" file semantics even when two files share
62 * the same physical blocks. This means that a write to one file must not
63 * alter the blocks in a different file; the way that we'll do that is
64 * through the use of a copy-on-write mechanism. At a high level, that
65 * means that when we want to write to a shared block, we allocate a new
66 * block, write the data to the new block, and if that succeeds we map the
67 * new block into the file.
69 * XFS provides a "delayed allocation" mechanism that defers the allocation
70 * of disk blocks to dirty-but-not-yet-mapped file blocks as long as
71 * possible. This reduces fragmentation by enabling the filesystem to ask
72 * for bigger chunks less often, which is exactly what we want for CoW.
74 * The delalloc mechanism begins when the kernel wants to make a block
75 * writable (write_begin or page_mkwrite). If the offset is not mapped, we
76 * create a delalloc mapping, which is a regular in-core extent, but without
77 * a real startblock. (For delalloc mappings, the startblock encodes both
78 * a flag that this is a delalloc mapping, and a worst-case estimate of how
79 * many blocks might be required to put the mapping into the BMBT.) delalloc
80 * mappings are a reservation against the free space in the filesystem;
81 * adjacent mappings can also be combined into fewer larger mappings.
83 * When dirty pages are being written out (typically in writepage), the
84 * delalloc reservations are converted into real mappings by allocating
85 * blocks and replacing the delalloc mapping with real ones. A delalloc
86 * mapping can be replaced by several real ones if the free space is
89 * We want to adapt the delalloc mechanism for copy-on-write, since the
90 * write paths are similar. The first two steps (creating the reservation
91 * and allocating the blocks) are exactly the same as delalloc except that
92 * the mappings must be stored in a separate CoW fork because we do not want
93 * to disturb the mapping in the data fork until we're sure that the write
94 * succeeded. IO completion in this case is the process of removing the old
95 * mapping from the data fork and moving the new mapping from the CoW fork to
96 * the data fork. This will be discussed shortly.
98 * For now, unaligned directio writes will be bounced back to the page cache.
99 * Block-aligned directio writes will use the same mechanism as buffered
102 * CoW remapping must be done after the data block write completes,
103 * because we don't want to destroy the old data fork map until we're sure
104 * the new block has been written. Since the new mappings are kept in a
105 * separate fork, we can simply iterate these mappings to find the ones
106 * that cover the file blocks that we just CoW'd. For each extent, simply
107 * unmap the corresponding range in the data fork, map the new range into
108 * the data fork, and remove the extent from the CoW fork.
110 * Since the remapping operation can be applied to an arbitrary file
111 * range, we record the need for the remap step as a flag in the ioend
112 * instead of declaring a new IO type. This is required for direct io
113 * because we only have ioend for the whole dio, and we have to be able to
114 * remember the presence of unwritten blocks and CoW blocks with a single
115 * ioend structure. Better yet, the more ground we can cover with one
120 * Given an AG extent, find the lowest-numbered run of shared blocks
121 * within that range and return the range in fbno/flen. If
122 * find_end_of_shared is true, return the longest contiguous extent of
123 * shared blocks. If there are no shared extents, fbno and flen will
124 * be set to NULLAGBLOCK and 0, respectively.
127 xfs_reflink_find_shared(
128 struct xfs_mount *mp,
134 bool find_end_of_shared)
136 struct xfs_buf *agbp;
137 struct xfs_btree_cur *cur;
140 error = xfs_alloc_read_agf(mp, NULL, agno, 0, &agbp);
144 cur = xfs_refcountbt_init_cursor(mp, NULL, agbp, agno, NULL);
146 error = xfs_refcount_find_shared(cur, agbno, aglen, fbno, flen,
149 xfs_btree_del_cursor(cur, error ? XFS_BTREE_ERROR : XFS_BTREE_NOERROR);
156 * Trim the mapping to the next block where there's a change in the
157 * shared/unshared status. More specifically, this means that we
158 * find the lowest-numbered extent of shared blocks that coincides with
159 * the given block mapping. If the shared extent overlaps the start of
160 * the mapping, trim the mapping to the end of the shared extent. If
161 * the shared region intersects the mapping, trim the mapping to the
162 * start of the shared extent. If there are no shared regions that
163 * overlap, just return the original extent.
166 xfs_reflink_trim_around_shared(
167 struct xfs_inode *ip,
168 struct xfs_bmbt_irec *irec,
179 /* Holes, unwritten, and delalloc extents cannot be shared */
180 if (!xfs_is_reflink_inode(ip) ||
182 irec->br_startblock == HOLESTARTBLOCK ||
183 irec->br_startblock == DELAYSTARTBLOCK) {
188 trace_xfs_reflink_trim_around_shared(ip, irec);
190 agno = XFS_FSB_TO_AGNO(ip->i_mount, irec->br_startblock);
191 agbno = XFS_FSB_TO_AGBNO(ip->i_mount, irec->br_startblock);
192 aglen = irec->br_blockcount;
194 error = xfs_reflink_find_shared(ip->i_mount, agno, agbno,
195 aglen, &fbno, &flen, true);
199 *shared = *trimmed = false;
200 if (fbno == NULLAGBLOCK) {
201 /* No shared blocks at all. */
203 } else if (fbno == agbno) {
205 * The start of this extent is shared. Truncate the
206 * mapping at the end of the shared region so that a
207 * subsequent iteration starts at the start of the
210 irec->br_blockcount = flen;
217 * There's a shared extent midway through this extent.
218 * Truncate the mapping at the start of the shared
219 * extent so that a subsequent iteration starts at the
220 * start of the shared region.
222 irec->br_blockcount = fbno - agbno;
228 /* Create a CoW reservation for a range of blocks within a file. */
230 __xfs_reflink_reserve_cow(
231 struct xfs_inode *ip,
232 xfs_fileoff_t *offset_fsb,
233 xfs_fileoff_t end_fsb,
236 struct xfs_bmbt_irec got, prev, imap;
237 xfs_fileoff_t orig_end_fsb;
238 int nimaps, eof = 0, error = 0;
239 bool shared = false, trimmed = false;
242 /* Already reserved? Skip the refcount btree access. */
243 xfs_bmap_search_extents(ip, *offset_fsb, XFS_COW_FORK, &eof, &idx,
245 if (!eof && got.br_startoff <= *offset_fsb) {
246 end_fsb = orig_end_fsb = got.br_startoff + got.br_blockcount;
247 trace_xfs_reflink_cow_found(ip, &got);
251 /* Read extent from the source file. */
253 error = xfs_bmapi_read(ip, *offset_fsb, end_fsb - *offset_fsb,
259 /* Trim the mapping to the nearest shared extent boundary. */
260 error = xfs_reflink_trim_around_shared(ip, &imap, &shared, &trimmed);
264 end_fsb = orig_end_fsb = imap.br_startoff + imap.br_blockcount;
266 /* Not shared? Just report the (potentially capped) extent. */
273 * Fork all the shared blocks from our write offset until the end of
276 error = xfs_qm_dqattach_locked(ip, 0);
281 error = xfs_bmapi_reserve_delalloc(ip, XFS_COW_FORK, *offset_fsb,
282 end_fsb - *offset_fsb, &got,
289 /* retry without any preallocation */
290 trace_xfs_reflink_cow_enospc(ip, &imap);
291 if (end_fsb != orig_end_fsb) {
292 end_fsb = orig_end_fsb;
300 trace_xfs_reflink_cow_alloc(ip, &got);
302 *offset_fsb = end_fsb;
307 /* Create a CoW reservation for part of a file. */
309 xfs_reflink_reserve_cow_range(
310 struct xfs_inode *ip,
314 struct xfs_mount *mp = ip->i_mount;
315 xfs_fileoff_t offset_fsb, end_fsb;
316 bool skipped = false;
319 trace_xfs_reflink_reserve_cow_range(ip, offset, count);
321 offset_fsb = XFS_B_TO_FSBT(mp, offset);
322 end_fsb = XFS_B_TO_FSB(mp, offset + count);
324 xfs_ilock(ip, XFS_ILOCK_EXCL);
325 while (offset_fsb < end_fsb) {
326 error = __xfs_reflink_reserve_cow(ip, &offset_fsb, end_fsb,
329 trace_xfs_reflink_reserve_cow_range_error(ip, error,
334 xfs_iunlock(ip, XFS_ILOCK_EXCL);
339 /* Allocate all CoW reservations covering a range of blocks in a file. */
341 __xfs_reflink_allocate_cow(
342 struct xfs_inode *ip,
343 xfs_fileoff_t *offset_fsb,
344 xfs_fileoff_t end_fsb)
346 struct xfs_mount *mp = ip->i_mount;
347 struct xfs_bmbt_irec imap;
348 struct xfs_defer_ops dfops;
349 struct xfs_trans *tp;
350 xfs_fsblock_t first_block;
351 xfs_fileoff_t next_fsb;
352 int nimaps = 1, error;
353 bool skipped = false;
355 xfs_defer_init(&dfops, &first_block);
357 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, 0, 0,
358 XFS_TRANS_RESERVE, &tp);
362 xfs_ilock(ip, XFS_ILOCK_EXCL);
364 next_fsb = *offset_fsb;
365 error = __xfs_reflink_reserve_cow(ip, &next_fsb, end_fsb, &skipped);
367 goto out_trans_cancel;
370 *offset_fsb = next_fsb;
371 goto out_trans_cancel;
374 xfs_trans_ijoin(tp, ip, 0);
375 error = xfs_bmapi_write(tp, ip, *offset_fsb, next_fsb - *offset_fsb,
376 XFS_BMAPI_COWFORK, &first_block,
377 XFS_EXTENTADD_SPACE_RES(mp, XFS_DATA_FORK),
378 &imap, &nimaps, &dfops);
380 goto out_trans_cancel;
382 /* We might not have been able to map the whole delalloc extent */
383 *offset_fsb = min(*offset_fsb + imap.br_blockcount, next_fsb);
385 error = xfs_defer_finish(&tp, &dfops, NULL);
387 goto out_trans_cancel;
389 error = xfs_trans_commit(tp);
392 xfs_iunlock(ip, XFS_ILOCK_EXCL);
395 xfs_defer_cancel(&dfops);
396 xfs_trans_cancel(tp);
400 /* Allocate all CoW reservations covering a part of a file. */
402 xfs_reflink_allocate_cow_range(
403 struct xfs_inode *ip,
407 struct xfs_mount *mp = ip->i_mount;
408 xfs_fileoff_t offset_fsb = XFS_B_TO_FSBT(mp, offset);
409 xfs_fileoff_t end_fsb = XFS_B_TO_FSB(mp, offset + count);
412 ASSERT(xfs_is_reflink_inode(ip));
414 trace_xfs_reflink_allocate_cow_range(ip, offset, count);
417 * Make sure that the dquots are there.
419 error = xfs_qm_dqattach(ip, 0);
423 while (offset_fsb < end_fsb) {
424 error = __xfs_reflink_allocate_cow(ip, &offset_fsb, end_fsb);
426 trace_xfs_reflink_allocate_cow_range_error(ip, error,
436 * Find the CoW reservation (and whether or not it needs block allocation)
437 * for a given byte offset of a file.
440 xfs_reflink_find_cow_mapping(
441 struct xfs_inode *ip,
443 struct xfs_bmbt_irec *imap,
446 struct xfs_bmbt_irec irec;
447 struct xfs_ifork *ifp;
448 struct xfs_bmbt_rec_host *gotp;
452 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL | XFS_ILOCK_SHARED));
453 ASSERT(xfs_is_reflink_inode(ip));
455 /* Find the extent in the CoW fork. */
456 ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
457 bno = XFS_B_TO_FSBT(ip->i_mount, offset);
458 gotp = xfs_iext_bno_to_ext(ifp, bno, &idx);
462 xfs_bmbt_get_all(gotp, &irec);
463 if (bno >= irec.br_startoff + irec.br_blockcount ||
464 bno < irec.br_startoff)
467 trace_xfs_reflink_find_cow_mapping(ip, offset, 1, XFS_IO_OVERWRITE,
470 /* If it's still delalloc, we must allocate later. */
472 *need_alloc = !!(isnullstartblock(irec.br_startblock));
478 * Trim an extent to end at the next CoW reservation past offset_fsb.
481 xfs_reflink_trim_irec_to_next_cow(
482 struct xfs_inode *ip,
483 xfs_fileoff_t offset_fsb,
484 struct xfs_bmbt_irec *imap)
486 struct xfs_bmbt_irec irec;
487 struct xfs_ifork *ifp;
488 struct xfs_bmbt_rec_host *gotp;
491 if (!xfs_is_reflink_inode(ip))
494 /* Find the extent in the CoW fork. */
495 ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
496 gotp = xfs_iext_bno_to_ext(ifp, offset_fsb, &idx);
499 xfs_bmbt_get_all(gotp, &irec);
501 /* This is the extent before; try sliding up one. */
502 if (irec.br_startoff < offset_fsb) {
504 if (idx >= ifp->if_bytes / sizeof(xfs_bmbt_rec_t))
506 gotp = xfs_iext_get_ext(ifp, idx);
507 xfs_bmbt_get_all(gotp, &irec);
510 if (irec.br_startoff >= imap->br_startoff + imap->br_blockcount)
513 imap->br_blockcount = irec.br_startoff - imap->br_startoff;
514 trace_xfs_reflink_trim_irec(ip, imap);
520 * Cancel all pending CoW reservations for some block range of an inode.
523 xfs_reflink_cancel_cow_blocks(
524 struct xfs_inode *ip,
525 struct xfs_trans **tpp,
526 xfs_fileoff_t offset_fsb,
527 xfs_fileoff_t end_fsb)
529 struct xfs_bmbt_irec irec;
530 xfs_filblks_t count_fsb;
531 xfs_fsblock_t firstfsb;
532 struct xfs_defer_ops dfops;
536 if (!xfs_is_reflink_inode(ip))
539 /* Go find the old extent in the CoW fork. */
540 while (offset_fsb < end_fsb) {
542 count_fsb = (xfs_filblks_t)(end_fsb - offset_fsb);
543 error = xfs_bmapi_read(ip, offset_fsb, count_fsb, &irec,
544 &nimaps, XFS_BMAPI_COWFORK);
549 trace_xfs_reflink_cancel_cow(ip, &irec);
551 if (irec.br_startblock == DELAYSTARTBLOCK) {
552 /* Free a delayed allocation. */
553 xfs_mod_fdblocks(ip->i_mount, irec.br_blockcount,
555 ip->i_delayed_blks -= irec.br_blockcount;
557 /* Remove the mapping from the CoW fork. */
558 error = xfs_bunmapi_cow(ip, &irec);
561 } else if (irec.br_startblock == HOLESTARTBLOCK) {
564 xfs_trans_ijoin(*tpp, ip, 0);
565 xfs_defer_init(&dfops, &firstfsb);
567 /* Free the CoW orphan record. */
568 error = xfs_refcount_free_cow_extent(ip->i_mount,
569 &dfops, irec.br_startblock,
574 xfs_bmap_add_free(ip->i_mount, &dfops,
575 irec.br_startblock, irec.br_blockcount,
578 /* Update quota accounting */
579 xfs_trans_mod_dquot_byino(*tpp, ip, XFS_TRANS_DQ_BCOUNT,
580 -(long)irec.br_blockcount);
582 /* Roll the transaction */
583 error = xfs_defer_finish(tpp, &dfops, ip);
585 xfs_defer_cancel(&dfops);
589 /* Remove the mapping from the CoW fork. */
590 error = xfs_bunmapi_cow(ip, &irec);
596 offset_fsb = irec.br_startoff + irec.br_blockcount;
603 * Cancel all pending CoW reservations for some byte range of an inode.
606 xfs_reflink_cancel_cow_range(
607 struct xfs_inode *ip,
611 struct xfs_trans *tp;
612 xfs_fileoff_t offset_fsb;
613 xfs_fileoff_t end_fsb;
616 trace_xfs_reflink_cancel_cow_range(ip, offset, count);
618 offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
619 if (count == NULLFILEOFF)
620 end_fsb = NULLFILEOFF;
622 end_fsb = XFS_B_TO_FSB(ip->i_mount, offset + count);
624 /* Start a rolling transaction to remove the mappings */
625 error = xfs_trans_alloc(ip->i_mount, &M_RES(ip->i_mount)->tr_write,
630 xfs_ilock(ip, XFS_ILOCK_EXCL);
631 xfs_trans_ijoin(tp, ip, 0);
633 /* Scrape out the old CoW reservations */
634 error = xfs_reflink_cancel_cow_blocks(ip, &tp, offset_fsb, end_fsb);
638 error = xfs_trans_commit(tp);
640 xfs_iunlock(ip, XFS_ILOCK_EXCL);
644 xfs_trans_cancel(tp);
645 xfs_iunlock(ip, XFS_ILOCK_EXCL);
647 trace_xfs_reflink_cancel_cow_range_error(ip, error, _RET_IP_);
652 * Remap parts of a file's data fork after a successful CoW.
656 struct xfs_inode *ip,
660 struct xfs_bmbt_irec irec;
661 struct xfs_bmbt_irec uirec;
662 struct xfs_trans *tp;
663 xfs_fileoff_t offset_fsb;
664 xfs_fileoff_t end_fsb;
665 xfs_filblks_t count_fsb;
666 xfs_fsblock_t firstfsb;
667 struct xfs_defer_ops dfops;
669 unsigned int resblks;
674 trace_xfs_reflink_end_cow(ip, offset, count);
676 offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
677 end_fsb = XFS_B_TO_FSB(ip->i_mount, offset + count);
678 count_fsb = (xfs_filblks_t)(end_fsb - offset_fsb);
680 /* Start a rolling transaction to switch the mappings */
681 resblks = XFS_EXTENTADD_SPACE_RES(ip->i_mount, XFS_DATA_FORK);
682 error = xfs_trans_alloc(ip->i_mount, &M_RES(ip->i_mount)->tr_write,
687 xfs_ilock(ip, XFS_ILOCK_EXCL);
688 xfs_trans_ijoin(tp, ip, 0);
690 /* Go find the old extent in the CoW fork. */
691 while (offset_fsb < end_fsb) {
692 /* Read extent from the source file */
694 count_fsb = (xfs_filblks_t)(end_fsb - offset_fsb);
695 error = xfs_bmapi_read(ip, offset_fsb, count_fsb, &irec,
696 &nimaps, XFS_BMAPI_COWFORK);
701 ASSERT(irec.br_startblock != DELAYSTARTBLOCK);
702 trace_xfs_reflink_cow_remap(ip, &irec);
705 * We can have a hole in the CoW fork if part of a directio
706 * write is CoW but part of it isn't.
708 rlen = ilen = irec.br_blockcount;
709 if (irec.br_startblock == HOLESTARTBLOCK)
712 /* Unmap the old blocks in the data fork. */
714 xfs_defer_init(&dfops, &firstfsb);
715 error = __xfs_bunmapi(tp, ip, irec.br_startoff,
716 &rlen, 0, 1, &firstfsb, &dfops);
721 * Trim the extent to whatever got unmapped.
722 * Remember, bunmapi works backwards.
724 uirec.br_startblock = irec.br_startblock + rlen;
725 uirec.br_startoff = irec.br_startoff + rlen;
726 uirec.br_blockcount = irec.br_blockcount - rlen;
727 irec.br_blockcount = rlen;
728 trace_xfs_reflink_cow_remap_piece(ip, &uirec);
730 /* Free the CoW orphan record. */
731 error = xfs_refcount_free_cow_extent(tp->t_mountp,
732 &dfops, uirec.br_startblock,
733 uirec.br_blockcount);
737 /* Map the new blocks into the data fork. */
738 error = xfs_bmap_map_extent(tp->t_mountp, &dfops,
743 /* Remove the mapping from the CoW fork. */
744 error = xfs_bunmapi_cow(ip, &uirec);
748 error = xfs_defer_finish(&tp, &dfops, ip);
755 offset_fsb = irec.br_startoff + ilen;
758 error = xfs_trans_commit(tp);
759 xfs_iunlock(ip, XFS_ILOCK_EXCL);
765 xfs_defer_cancel(&dfops);
767 xfs_trans_cancel(tp);
768 xfs_iunlock(ip, XFS_ILOCK_EXCL);
770 trace_xfs_reflink_end_cow_error(ip, error, _RET_IP_);
775 * Free leftover CoW reservations that didn't get cleaned out.
778 xfs_reflink_recover_cow(
779 struct xfs_mount *mp)
784 if (!xfs_sb_version_hasreflink(&mp->m_sb))
787 for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
788 error = xfs_refcount_recover_cow_leftovers(mp, agno);
797 * Reflinking (Block) Ranges of Two Files Together
799 * First, ensure that the reflink flag is set on both inodes. The flag is an
800 * optimization to avoid unnecessary refcount btree lookups in the write path.
802 * Now we can iteratively remap the range of extents (and holes) in src to the
803 * corresponding ranges in dest. Let drange and srange denote the ranges of
804 * logical blocks in dest and src touched by the reflink operation.
806 * While the length of drange is greater than zero,
807 * - Read src's bmbt at the start of srange ("imap")
808 * - If imap doesn't exist, make imap appear to start at the end of srange
810 * - If imap starts before srange, advance imap to start at srange.
811 * - If imap goes beyond srange, truncate imap to end at the end of srange.
812 * - Punch (imap start - srange start + imap len) blocks from dest at
813 * offset (drange start).
814 * - If imap points to a real range of pblks,
815 * > Increase the refcount of the imap's pblks
816 * > Map imap's pblks into dest at the offset
817 * (drange start + imap start - srange start)
818 * - Advance drange and srange by (imap start - srange start + imap len)
820 * Finally, if the reflink made dest longer, update both the in-core and
821 * on-disk file sizes.
823 * ASCII Art Demonstration:
825 * Let's say we want to reflink this source file:
827 * ----SSSSSSS-SSSSS----SSSSSS (src file)
828 * <-------------------->
830 * into this destination file:
832 * --DDDDDDDDDDDDDDDDDDD--DDD (dest file)
833 * <-------------------->
834 * '-' means a hole, and 'S' and 'D' are written blocks in the src and dest.
835 * Observe that the range has different logical offsets in either file.
837 * Consider that the first extent in the source file doesn't line up with our
838 * reflink range. Unmapping and remapping are separate operations, so we can
839 * unmap more blocks from the destination file than we remap.
841 * ----SSSSSSS-SSSSS----SSSSSS
843 * --DDDDD---------DDDDD--DDD
846 * Now remap the source extent into the destination file:
848 * ----SSSSSSS-SSSSS----SSSSSS
850 * --DDDDD--SSSSSSSDDDDD--DDD
853 * Do likewise with the second hole and extent in our range. Holes in the
854 * unmap range don't affect our operation.
856 * ----SSSSSSS-SSSSS----SSSSSS
858 * --DDDDD--SSSSSSS-SSSSS-DDD
861 * Finally, unmap and remap part of the third extent. This will increase the
862 * size of the destination file.
864 * ----SSSSSSS-SSSSS----SSSSSS
866 * --DDDDD--SSSSSSS-SSSSS----SSS
869 * Once we update the destination file's i_size, we're done.
873 * Ensure the reflink bit is set in both inodes.
876 xfs_reflink_set_inode_flag(
877 struct xfs_inode *src,
878 struct xfs_inode *dest)
880 struct xfs_mount *mp = src->i_mount;
882 struct xfs_trans *tp;
884 if (xfs_is_reflink_inode(src) && xfs_is_reflink_inode(dest))
887 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
891 /* Lock both files against IO */
892 if (src->i_ino == dest->i_ino)
893 xfs_ilock(src, XFS_ILOCK_EXCL);
895 xfs_lock_two_inodes(src, dest, XFS_ILOCK_EXCL);
897 if (!xfs_is_reflink_inode(src)) {
898 trace_xfs_reflink_set_inode_flag(src);
899 xfs_trans_ijoin(tp, src, XFS_ILOCK_EXCL);
900 src->i_d.di_flags2 |= XFS_DIFLAG2_REFLINK;
901 xfs_trans_log_inode(tp, src, XFS_ILOG_CORE);
902 xfs_ifork_init_cow(src);
904 xfs_iunlock(src, XFS_ILOCK_EXCL);
906 if (src->i_ino == dest->i_ino)
909 if (!xfs_is_reflink_inode(dest)) {
910 trace_xfs_reflink_set_inode_flag(dest);
911 xfs_trans_ijoin(tp, dest, XFS_ILOCK_EXCL);
912 dest->i_d.di_flags2 |= XFS_DIFLAG2_REFLINK;
913 xfs_trans_log_inode(tp, dest, XFS_ILOG_CORE);
914 xfs_ifork_init_cow(dest);
916 xfs_iunlock(dest, XFS_ILOCK_EXCL);
919 error = xfs_trans_commit(tp);
925 trace_xfs_reflink_set_inode_flag_error(dest, error, _RET_IP_);
930 * Update destination inode size, if necessary.
933 xfs_reflink_update_dest(
934 struct xfs_inode *dest,
937 struct xfs_mount *mp = dest->i_mount;
938 struct xfs_trans *tp;
941 if (newlen <= i_size_read(VFS_I(dest)))
944 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
948 xfs_ilock(dest, XFS_ILOCK_EXCL);
949 xfs_trans_ijoin(tp, dest, XFS_ILOCK_EXCL);
951 trace_xfs_reflink_update_inode_size(dest, newlen);
952 i_size_write(VFS_I(dest), newlen);
953 dest->i_d.di_size = newlen;
954 xfs_trans_log_inode(tp, dest, XFS_ILOG_CORE);
956 error = xfs_trans_commit(tp);
962 trace_xfs_reflink_update_inode_size_error(dest, error, _RET_IP_);
967 * Unmap a range of blocks from a file, then map other blocks into the hole.
968 * The range to unmap is (destoff : destoff + srcioff + irec->br_blockcount).
969 * The extent irec is mapped into dest at irec->br_startoff.
972 xfs_reflink_remap_extent(
973 struct xfs_inode *ip,
974 struct xfs_bmbt_irec *irec,
975 xfs_fileoff_t destoff,
978 struct xfs_mount *mp = ip->i_mount;
979 struct xfs_trans *tp;
980 xfs_fsblock_t firstfsb;
981 unsigned int resblks;
982 struct xfs_defer_ops dfops;
983 struct xfs_bmbt_irec uirec;
986 xfs_filblks_t unmap_len;
990 unmap_len = irec->br_startoff + irec->br_blockcount - destoff;
991 trace_xfs_reflink_punch_range(ip, destoff, unmap_len);
993 /* Only remap normal extents. */
994 real_extent = (irec->br_startblock != HOLESTARTBLOCK &&
995 irec->br_startblock != DELAYSTARTBLOCK &&
998 /* Start a rolling transaction to switch the mappings */
999 resblks = XFS_EXTENTADD_SPACE_RES(ip->i_mount, XFS_DATA_FORK);
1000 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, resblks, 0, 0, &tp);
1004 xfs_ilock(ip, XFS_ILOCK_EXCL);
1005 xfs_trans_ijoin(tp, ip, 0);
1007 /* If we're not just clearing space, then do we have enough quota? */
1009 error = xfs_trans_reserve_quota_nblks(tp, ip,
1010 irec->br_blockcount, 0, XFS_QMOPT_RES_REGBLKS);
1015 trace_xfs_reflink_remap(ip, irec->br_startoff,
1016 irec->br_blockcount, irec->br_startblock);
1018 /* Unmap the old blocks in the data fork. */
1021 xfs_defer_init(&dfops, &firstfsb);
1022 error = __xfs_bunmapi(tp, ip, destoff, &rlen, 0, 1,
1028 * Trim the extent to whatever got unmapped.
1029 * Remember, bunmapi works backwards.
1031 uirec.br_startblock = irec->br_startblock + rlen;
1032 uirec.br_startoff = irec->br_startoff + rlen;
1033 uirec.br_blockcount = unmap_len - rlen;
1036 /* If this isn't a real mapping, we're done. */
1037 if (!real_extent || uirec.br_blockcount == 0)
1040 trace_xfs_reflink_remap(ip, uirec.br_startoff,
1041 uirec.br_blockcount, uirec.br_startblock);
1043 /* Update the refcount tree */
1044 error = xfs_refcount_increase_extent(mp, &dfops, &uirec);
1048 /* Map the new blocks into the data fork. */
1049 error = xfs_bmap_map_extent(mp, &dfops, ip, &uirec);
1053 /* Update quota accounting. */
1054 xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_BCOUNT,
1055 uirec.br_blockcount);
1057 /* Update dest isize if needed. */
1058 newlen = XFS_FSB_TO_B(mp,
1059 uirec.br_startoff + uirec.br_blockcount);
1060 newlen = min_t(xfs_off_t, newlen, new_isize);
1061 if (newlen > i_size_read(VFS_I(ip))) {
1062 trace_xfs_reflink_update_inode_size(ip, newlen);
1063 i_size_write(VFS_I(ip), newlen);
1064 ip->i_d.di_size = newlen;
1065 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1069 /* Process all the deferred stuff. */
1070 error = xfs_defer_finish(&tp, &dfops, ip);
1075 error = xfs_trans_commit(tp);
1076 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1082 xfs_defer_cancel(&dfops);
1084 xfs_trans_cancel(tp);
1085 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1087 trace_xfs_reflink_remap_extent_error(ip, error, _RET_IP_);
1092 * Iteratively remap one file's extents (and holes) to another's.
1095 xfs_reflink_remap_blocks(
1096 struct xfs_inode *src,
1097 xfs_fileoff_t srcoff,
1098 struct xfs_inode *dest,
1099 xfs_fileoff_t destoff,
1101 xfs_off_t new_isize)
1103 struct xfs_bmbt_irec imap;
1106 xfs_filblks_t range_len;
1108 /* drange = (destoff, destoff + len); srange = (srcoff, srcoff + len) */
1110 trace_xfs_reflink_remap_blocks_loop(src, srcoff, len,
1112 /* Read extent from the source file */
1114 xfs_ilock(src, XFS_ILOCK_EXCL);
1115 error = xfs_bmapi_read(src, srcoff, len, &imap, &nimaps, 0);
1116 xfs_iunlock(src, XFS_ILOCK_EXCL);
1119 ASSERT(nimaps == 1);
1121 trace_xfs_reflink_remap_imap(src, srcoff, len, XFS_IO_OVERWRITE,
1124 /* Translate imap into the destination file. */
1125 range_len = imap.br_startoff + imap.br_blockcount - srcoff;
1126 imap.br_startoff += destoff - srcoff;
1128 /* Clear dest from destoff to the end of imap and map it in. */
1129 error = xfs_reflink_remap_extent(dest, &imap, destoff,
1134 if (fatal_signal_pending(current)) {
1139 /* Advance drange/srange */
1140 srcoff += range_len;
1141 destoff += range_len;
1148 trace_xfs_reflink_remap_blocks_error(dest, error, _RET_IP_);
1153 * Read a page's worth of file data into the page cache. Return the page
1156 static struct page *
1158 struct inode *inode,
1161 struct address_space *mapping;
1165 n = offset >> PAGE_SHIFT;
1166 mapping = inode->i_mapping;
1167 page = read_mapping_page(mapping, n, NULL);
1170 if (!PageUptodate(page)) {
1172 return ERR_PTR(-EIO);
1179 * Compare extents of two files to see if they are the same.
1182 xfs_compare_extents(
1191 xfs_off_t dest_poff;
1194 struct page *src_page;
1195 struct page *dest_page;
1203 src_poff = srcoff & (PAGE_SIZE - 1);
1204 dest_poff = destoff & (PAGE_SIZE - 1);
1205 cmp_len = min(PAGE_SIZE - src_poff,
1206 PAGE_SIZE - dest_poff);
1207 cmp_len = min(cmp_len, len);
1208 ASSERT(cmp_len > 0);
1210 trace_xfs_reflink_compare_extents(XFS_I(src), srcoff, cmp_len,
1211 XFS_I(dest), destoff);
1213 src_page = xfs_get_page(src, srcoff);
1214 if (IS_ERR(src_page)) {
1215 error = PTR_ERR(src_page);
1218 dest_page = xfs_get_page(dest, destoff);
1219 if (IS_ERR(dest_page)) {
1220 error = PTR_ERR(dest_page);
1221 unlock_page(src_page);
1225 src_addr = kmap_atomic(src_page);
1226 dest_addr = kmap_atomic(dest_page);
1228 flush_dcache_page(src_page);
1229 flush_dcache_page(dest_page);
1231 if (memcmp(src_addr + src_poff, dest_addr + dest_poff, cmp_len))
1234 kunmap_atomic(dest_addr);
1235 kunmap_atomic(src_addr);
1236 unlock_page(dest_page);
1237 unlock_page(src_page);
1238 put_page(dest_page);
1253 trace_xfs_reflink_compare_extents_error(XFS_I(dest), error, _RET_IP_);
1258 * Link a range of blocks from one file to another.
1261 xfs_reflink_remap_range(
1262 struct xfs_inode *src,
1264 struct xfs_inode *dest,
1269 struct xfs_mount *mp = src->i_mount;
1270 xfs_fileoff_t sfsbno, dfsbno;
1271 xfs_filblks_t fsblen;
1275 if (!xfs_sb_version_hasreflink(&mp->m_sb))
1278 if (XFS_FORCED_SHUTDOWN(mp))
1281 /* Don't reflink realtime inodes */
1282 if (XFS_IS_REALTIME_INODE(src) || XFS_IS_REALTIME_INODE(dest))
1285 if (flags & ~XFS_REFLINK_ALL)
1288 trace_xfs_reflink_remap_range(src, srcoff, len, dest, destoff);
1290 /* Lock both files against IO */
1291 if (src->i_ino == dest->i_ino) {
1292 xfs_ilock(src, XFS_IOLOCK_EXCL);
1293 xfs_ilock(src, XFS_MMAPLOCK_EXCL);
1295 xfs_lock_two_inodes(src, dest, XFS_IOLOCK_EXCL);
1296 xfs_lock_two_inodes(src, dest, XFS_MMAPLOCK_EXCL);
1300 * Check that the extents are the same.
1302 if (flags & XFS_REFLINK_DEDUPE) {
1304 error = xfs_compare_extents(VFS_I(src), srcoff, VFS_I(dest),
1305 destoff, len, &is_same);
1314 error = xfs_reflink_set_inode_flag(src, dest);
1319 * Invalidate the page cache so that we can clear any CoW mappings
1320 * in the destination file.
1322 truncate_inode_pages_range(&VFS_I(dest)->i_data, destoff,
1323 PAGE_ALIGN(destoff + len) - 1);
1325 dfsbno = XFS_B_TO_FSBT(mp, destoff);
1326 sfsbno = XFS_B_TO_FSBT(mp, srcoff);
1327 fsblen = XFS_B_TO_FSB(mp, len);
1328 error = xfs_reflink_remap_blocks(src, sfsbno, dest, dfsbno, fsblen,
1333 error = xfs_reflink_update_dest(dest, destoff + len);
1338 xfs_iunlock(src, XFS_MMAPLOCK_EXCL);
1339 xfs_iunlock(src, XFS_IOLOCK_EXCL);
1340 if (src->i_ino != dest->i_ino) {
1341 xfs_iunlock(dest, XFS_MMAPLOCK_EXCL);
1342 xfs_iunlock(dest, XFS_IOLOCK_EXCL);
1345 trace_xfs_reflink_remap_range_error(dest, error, _RET_IP_);
1350 * The user wants to preemptively CoW all shared blocks in this file,
1351 * which enables us to turn off the reflink flag. Iterate all
1352 * extents which are not prealloc/delalloc to see which ranges are
1353 * mentioned in the refcount tree, then read those blocks into the
1354 * pagecache, dirty them, fsync them back out, and then we can update
1355 * the inode flag. What happens if we run out of memory? :)
1358 xfs_reflink_dirty_extents(
1359 struct xfs_inode *ip,
1364 struct xfs_mount *mp = ip->i_mount;
1365 xfs_agnumber_t agno;
1366 xfs_agblock_t agbno;
1372 struct xfs_bmbt_irec map[2];
1376 while (end - fbno > 0) {
1379 * Look for extents in the file. Skip holes, delalloc, or
1380 * unwritten extents; they can't be reflinked.
1382 error = xfs_bmapi_read(ip, fbno, end - fbno, map, &nmaps, 0);
1387 if (map[0].br_startblock == HOLESTARTBLOCK ||
1388 map[0].br_startblock == DELAYSTARTBLOCK ||
1389 ISUNWRITTEN(&map[0]))
1393 while (map[1].br_blockcount) {
1394 agno = XFS_FSB_TO_AGNO(mp, map[1].br_startblock);
1395 agbno = XFS_FSB_TO_AGBNO(mp, map[1].br_startblock);
1396 aglen = map[1].br_blockcount;
1398 error = xfs_reflink_find_shared(mp, agno, agbno, aglen,
1399 &rbno, &rlen, true);
1402 if (rbno == NULLAGBLOCK)
1405 /* Dirty the pages */
1406 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1407 fpos = XFS_FSB_TO_B(mp, map[1].br_startoff +
1409 flen = XFS_FSB_TO_B(mp, rlen);
1410 if (fpos + flen > isize)
1411 flen = isize - fpos;
1412 error = iomap_file_dirty(VFS_I(ip), fpos, flen,
1414 xfs_ilock(ip, XFS_ILOCK_EXCL);
1418 map[1].br_blockcount -= (rbno - agbno + rlen);
1419 map[1].br_startoff += (rbno - agbno + rlen);
1420 map[1].br_startblock += (rbno - agbno + rlen);
1424 fbno = map[0].br_startoff + map[0].br_blockcount;
1430 /* Clear the inode reflink flag if there are no shared extents. */
1432 xfs_reflink_clear_inode_flag(
1433 struct xfs_inode *ip,
1434 struct xfs_trans **tpp)
1436 struct xfs_mount *mp = ip->i_mount;
1439 xfs_agnumber_t agno;
1440 xfs_agblock_t agbno;
1444 struct xfs_bmbt_irec map[2];
1448 if (!(ip->i_d.di_flags2 & XFS_DIFLAG2_REFLINK))
1452 end = XFS_B_TO_FSB(mp, i_size_read(VFS_I(ip)));
1453 while (end - fbno > 0) {
1456 * Look for extents in the file. Skip holes, delalloc, or
1457 * unwritten extents; they can't be reflinked.
1459 error = xfs_bmapi_read(ip, fbno, end - fbno, map, &nmaps, 0);
1464 if (map[0].br_startblock == HOLESTARTBLOCK ||
1465 map[0].br_startblock == DELAYSTARTBLOCK ||
1466 ISUNWRITTEN(&map[0]))
1470 while (map[1].br_blockcount) {
1471 agno = XFS_FSB_TO_AGNO(mp, map[1].br_startblock);
1472 agbno = XFS_FSB_TO_AGBNO(mp, map[1].br_startblock);
1473 aglen = map[1].br_blockcount;
1475 error = xfs_reflink_find_shared(mp, agno, agbno, aglen,
1476 &rbno, &rlen, false);
1479 /* Is there still a shared block here? */
1480 if (rbno != NULLAGBLOCK)
1483 map[1].br_blockcount -= aglen;
1484 map[1].br_startoff += aglen;
1485 map[1].br_startblock += aglen;
1489 fbno = map[0].br_startoff + map[0].br_blockcount;
1493 * We didn't find any shared blocks so turn off the reflink flag.
1494 * First, get rid of any leftover CoW mappings.
1496 error = xfs_reflink_cancel_cow_blocks(ip, tpp, 0, NULLFILEOFF);
1500 /* Clear the inode flag. */
1501 trace_xfs_reflink_unset_inode_flag(ip);
1502 ip->i_d.di_flags2 &= ~XFS_DIFLAG2_REFLINK;
1503 xfs_trans_ijoin(*tpp, ip, 0);
1504 xfs_trans_log_inode(*tpp, ip, XFS_ILOG_CORE);
1510 * Clear the inode reflink flag if there are no shared extents and the size
1514 xfs_reflink_try_clear_inode_flag(
1515 struct xfs_inode *ip,
1516 xfs_off_t old_isize)
1518 struct xfs_mount *mp = ip->i_mount;
1519 struct xfs_trans *tp;
1522 /* Start a rolling transaction to remove the mappings */
1523 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, 0, 0, 0, &tp);
1527 xfs_ilock(ip, XFS_ILOCK_EXCL);
1528 xfs_trans_ijoin(tp, ip, 0);
1530 if (old_isize != i_size_read(VFS_I(ip)))
1533 error = xfs_reflink_clear_inode_flag(ip, &tp);
1537 error = xfs_trans_commit(tp);
1541 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1544 xfs_trans_cancel(tp);
1546 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1551 * Pre-COW all shared blocks within a given byte range of a file and turn off
1552 * the reflink flag if we unshare all of the file's blocks.
1555 xfs_reflink_unshare(
1556 struct xfs_inode *ip,
1560 struct xfs_mount *mp = ip->i_mount;
1566 if (!xfs_is_reflink_inode(ip))
1569 trace_xfs_reflink_unshare(ip, offset, len);
1571 inode_dio_wait(VFS_I(ip));
1573 /* Try to CoW the selected ranges */
1574 xfs_ilock(ip, XFS_ILOCK_EXCL);
1575 fbno = XFS_B_TO_FSB(mp, offset);
1576 isize = i_size_read(VFS_I(ip));
1577 end = XFS_B_TO_FSB(mp, offset + len);
1578 error = xfs_reflink_dirty_extents(ip, fbno, end, isize);
1581 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1583 /* Wait for the IO to finish */
1584 error = filemap_write_and_wait(VFS_I(ip)->i_mapping);
1588 /* Turn off the reflink flag if we unshared the whole file */
1589 if (offset == 0 && len == isize) {
1590 error = xfs_reflink_try_clear_inode_flag(ip, isize);
1598 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1600 trace_xfs_reflink_unshare_error(ip, error, _RET_IP_);