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_reflink.h"
53 #include "xfs_iomap.h"
54 #include "xfs_rmap_btree.h"
56 #include "xfs_ag_resv.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 * As an optimization, the CoW extent size hint (cowextsz) creates
84 * outsized aligned delalloc reservations in the hope of landing out of
85 * order nearby CoW writes in a single extent on disk, thereby reducing
86 * fragmentation and improving future performance.
88 * D: --RRRRRRSSSRRRRRRRR--- (data fork)
89 * C: ------DDDDDDD--------- (CoW fork)
91 * When dirty pages are being written out (typically in writepage), the
92 * delalloc reservations are converted into unwritten mappings by
93 * allocating blocks and replacing the delalloc mapping with real ones.
94 * A delalloc mapping can be replaced by several unwritten ones if the
95 * free space is fragmented.
97 * D: --RRRRRRSSSRRRRRRRR---
98 * C: ------UUUUUUU---------
100 * We want to adapt the delalloc mechanism for copy-on-write, since the
101 * write paths are similar. The first two steps (creating the reservation
102 * and allocating the blocks) are exactly the same as delalloc except that
103 * the mappings must be stored in a separate CoW fork because we do not want
104 * to disturb the mapping in the data fork until we're sure that the write
105 * succeeded. IO completion in this case is the process of removing the old
106 * mapping from the data fork and moving the new mapping from the CoW fork to
107 * the data fork. This will be discussed shortly.
109 * For now, unaligned directio writes will be bounced back to the page cache.
110 * Block-aligned directio writes will use the same mechanism as buffered
113 * Just prior to submitting the actual disk write requests, we convert
114 * the extents representing the range of the file actually being written
115 * (as opposed to extra pieces created for the cowextsize hint) to real
116 * extents. This will become important in the next step:
118 * D: --RRRRRRSSSRRRRRRRR---
119 * C: ------UUrrUUU---------
121 * CoW remapping must be done after the data block write completes,
122 * because we don't want to destroy the old data fork map until we're sure
123 * the new block has been written. Since the new mappings are kept in a
124 * separate fork, we can simply iterate these mappings to find the ones
125 * that cover the file blocks that we just CoW'd. For each extent, simply
126 * unmap the corresponding range in the data fork, map the new range into
127 * the data fork, and remove the extent from the CoW fork. Because of
128 * the presence of the cowextsize hint, however, we must be careful
129 * only to remap the blocks that we've actually written out -- we must
130 * never remap delalloc reservations nor CoW staging blocks that have
131 * yet to be written. This corresponds exactly to the real extents in
134 * D: --RRRRRRrrSRRRRRRRR---
135 * C: ------UU--UUU---------
137 * Since the remapping operation can be applied to an arbitrary file
138 * range, we record the need for the remap step as a flag in the ioend
139 * instead of declaring a new IO type. This is required for direct io
140 * because we only have ioend for the whole dio, and we have to be able to
141 * remember the presence of unwritten blocks and CoW blocks with a single
142 * ioend structure. Better yet, the more ground we can cover with one
147 * Given an AG extent, find the lowest-numbered run of shared blocks
148 * within that range and return the range in fbno/flen. If
149 * find_end_of_shared is true, return the longest contiguous extent of
150 * shared blocks. If there are no shared extents, fbno and flen will
151 * be set to NULLAGBLOCK and 0, respectively.
154 xfs_reflink_find_shared(
155 struct xfs_mount *mp,
156 struct xfs_trans *tp,
162 bool find_end_of_shared)
164 struct xfs_buf *agbp;
165 struct xfs_btree_cur *cur;
168 error = xfs_alloc_read_agf(mp, tp, agno, 0, &agbp);
174 cur = xfs_refcountbt_init_cursor(mp, tp, agbp, agno, NULL);
176 error = xfs_refcount_find_shared(cur, agbno, aglen, fbno, flen,
179 xfs_btree_del_cursor(cur, error ? XFS_BTREE_ERROR : XFS_BTREE_NOERROR);
181 xfs_trans_brelse(tp, agbp);
186 * Trim the mapping to the next block where there's a change in the
187 * shared/unshared status. More specifically, this means that we
188 * find the lowest-numbered extent of shared blocks that coincides with
189 * the given block mapping. If the shared extent overlaps the start of
190 * the mapping, trim the mapping to the end of the shared extent. If
191 * the shared region intersects the mapping, trim the mapping to the
192 * start of the shared extent. If there are no shared regions that
193 * overlap, just return the original extent.
196 xfs_reflink_trim_around_shared(
197 struct xfs_inode *ip,
198 struct xfs_bmbt_irec *irec,
209 /* Holes, unwritten, and delalloc extents cannot be shared */
210 if (!xfs_is_reflink_inode(ip) || !xfs_bmap_is_real_extent(irec)) {
215 trace_xfs_reflink_trim_around_shared(ip, irec);
217 agno = XFS_FSB_TO_AGNO(ip->i_mount, irec->br_startblock);
218 agbno = XFS_FSB_TO_AGBNO(ip->i_mount, irec->br_startblock);
219 aglen = irec->br_blockcount;
221 error = xfs_reflink_find_shared(ip->i_mount, NULL, agno, agbno,
222 aglen, &fbno, &flen, true);
226 *shared = *trimmed = false;
227 if (fbno == NULLAGBLOCK) {
228 /* No shared blocks at all. */
230 } else if (fbno == agbno) {
232 * The start of this extent is shared. Truncate the
233 * mapping at the end of the shared region so that a
234 * subsequent iteration starts at the start of the
237 irec->br_blockcount = flen;
244 * There's a shared extent midway through this extent.
245 * Truncate the mapping at the start of the shared
246 * extent so that a subsequent iteration starts at the
247 * start of the shared region.
249 irec->br_blockcount = fbno - agbno;
256 * Trim the passed in imap to the next shared/unshared extent boundary, and
257 * if imap->br_startoff points to a shared extent reserve space for it in the
258 * COW fork. In this case *shared is set to true, else to false.
260 * Note that imap will always contain the block numbers for the existing blocks
261 * in the data fork, as the upper layers need them for read-modify-write
265 xfs_reflink_reserve_cow(
266 struct xfs_inode *ip,
267 struct xfs_bmbt_irec *imap,
270 struct xfs_ifork *ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
271 struct xfs_bmbt_irec got;
273 bool eof = false, trimmed;
274 struct xfs_iext_cursor icur;
277 * Search the COW fork extent list first. This serves two purposes:
278 * first this implement the speculative preallocation using cowextisze,
279 * so that we also unshared block adjacent to shared blocks instead
280 * of just the shared blocks themselves. Second the lookup in the
281 * extent list is generally faster than going out to the shared extent
285 if (!xfs_iext_lookup_extent(ip, ifp, imap->br_startoff, &icur, &got))
287 if (!eof && got.br_startoff <= imap->br_startoff) {
288 trace_xfs_reflink_cow_found(ip, imap);
289 xfs_trim_extent(imap, got.br_startoff, got.br_blockcount);
295 /* Trim the mapping to the nearest shared extent boundary. */
296 error = xfs_reflink_trim_around_shared(ip, imap, shared, &trimmed);
300 /* Not shared? Just report the (potentially capped) extent. */
305 * Fork all the shared blocks from our write offset until the end of
308 error = xfs_qm_dqattach_locked(ip, 0);
312 error = xfs_bmapi_reserve_delalloc(ip, XFS_COW_FORK, imap->br_startoff,
313 imap->br_blockcount, 0, &got, &icur, eof);
314 if (error == -ENOSPC || error == -EDQUOT)
315 trace_xfs_reflink_cow_enospc(ip, imap);
319 trace_xfs_reflink_cow_alloc(ip, &got);
323 /* Convert part of an unwritten CoW extent to a real one. */
325 xfs_reflink_convert_cow_extent(
326 struct xfs_inode *ip,
327 struct xfs_bmbt_irec *imap,
328 xfs_fileoff_t offset_fsb,
329 xfs_filblks_t count_fsb,
330 struct xfs_defer_ops *dfops)
332 xfs_fsblock_t first_block = NULLFSBLOCK;
335 if (imap->br_state == XFS_EXT_NORM)
338 xfs_trim_extent(imap, offset_fsb, count_fsb);
339 trace_xfs_reflink_convert_cow(ip, imap);
340 if (imap->br_blockcount == 0)
342 return xfs_bmapi_write(NULL, ip, imap->br_startoff, imap->br_blockcount,
343 XFS_BMAPI_COWFORK | XFS_BMAPI_CONVERT, &first_block,
344 0, imap, &nimaps, dfops);
347 /* Convert all of the unwritten CoW extents in a file's range to real ones. */
349 xfs_reflink_convert_cow(
350 struct xfs_inode *ip,
354 struct xfs_mount *mp = ip->i_mount;
355 xfs_fileoff_t offset_fsb = XFS_B_TO_FSBT(mp, offset);
356 xfs_fileoff_t end_fsb = XFS_B_TO_FSB(mp, offset + count);
357 xfs_filblks_t count_fsb = end_fsb - offset_fsb;
358 struct xfs_bmbt_irec imap;
359 struct xfs_defer_ops dfops;
360 xfs_fsblock_t first_block = NULLFSBLOCK;
361 int nimaps = 1, error = 0;
365 xfs_ilock(ip, XFS_ILOCK_EXCL);
366 error = xfs_bmapi_write(NULL, ip, offset_fsb, count_fsb,
367 XFS_BMAPI_COWFORK | XFS_BMAPI_CONVERT |
368 XFS_BMAPI_CONVERT_ONLY, &first_block, 0, &imap, &nimaps,
370 xfs_iunlock(ip, XFS_ILOCK_EXCL);
374 /* Allocate all CoW reservations covering a range of blocks in a file. */
376 xfs_reflink_allocate_cow(
377 struct xfs_inode *ip,
378 struct xfs_bmbt_irec *imap,
382 struct xfs_mount *mp = ip->i_mount;
383 xfs_fileoff_t offset_fsb = imap->br_startoff;
384 xfs_filblks_t count_fsb = imap->br_blockcount;
385 struct xfs_bmbt_irec got;
386 struct xfs_defer_ops dfops;
387 struct xfs_trans *tp = NULL;
388 xfs_fsblock_t first_block;
389 int nimaps, error = 0;
391 xfs_filblks_t resaligned;
392 xfs_extlen_t resblks = 0;
393 struct xfs_iext_cursor icur;
396 ASSERT(xfs_is_reflink_inode(ip));
397 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL | XFS_ILOCK_SHARED));
400 * Even if the extent is not shared we might have a preallocation for
401 * it in the COW fork. If so use it.
403 if (xfs_iext_lookup_extent(ip, ip->i_cowfp, offset_fsb, &icur, &got) &&
404 got.br_startoff <= offset_fsb) {
407 /* If we have a real allocation in the COW fork we're done. */
408 if (!isnullstartblock(got.br_startblock)) {
409 xfs_trim_extent(&got, offset_fsb, count_fsb);
414 xfs_trim_extent(imap, got.br_startoff, got.br_blockcount);
416 error = xfs_reflink_trim_around_shared(ip, imap, shared, &trimmed);
417 if (error || !*shared)
422 resaligned = xfs_aligned_fsb_count(imap->br_startoff,
423 imap->br_blockcount, xfs_get_cowextsz_hint(ip));
424 resblks = XFS_DIOSTRAT_SPACE_RES(mp, resaligned);
426 xfs_iunlock(ip, *lockmode);
427 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, resblks, 0, 0, &tp);
428 *lockmode = XFS_ILOCK_EXCL;
429 xfs_ilock(ip, *lockmode);
434 error = xfs_qm_dqattach_locked(ip, 0);
440 error = xfs_trans_reserve_quota_nblks(tp, ip, resblks, 0,
441 XFS_QMOPT_RES_REGBLKS);
445 xfs_trans_ijoin(tp, ip, 0);
447 xfs_defer_init(&dfops, &first_block);
450 /* Allocate the entire reservation as unwritten blocks. */
451 error = xfs_bmapi_write(tp, ip, imap->br_startoff, imap->br_blockcount,
452 XFS_BMAPI_COWFORK | XFS_BMAPI_PREALLOC, &first_block,
453 resblks, imap, &nimaps, &dfops);
455 goto out_bmap_cancel;
458 error = xfs_defer_finish(&tp, &dfops);
460 goto out_bmap_cancel;
462 error = xfs_trans_commit(tp);
466 return xfs_reflink_convert_cow_extent(ip, imap, offset_fsb, count_fsb,
469 xfs_defer_cancel(&dfops);
470 xfs_trans_unreserve_quota_nblks(tp, ip, (long)resblks, 0,
471 XFS_QMOPT_RES_REGBLKS);
474 xfs_trans_cancel(tp);
479 * Find the CoW reservation for a given byte offset of a file.
482 xfs_reflink_find_cow_mapping(
483 struct xfs_inode *ip,
485 struct xfs_bmbt_irec *imap)
487 struct xfs_ifork *ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
488 xfs_fileoff_t offset_fsb;
489 struct xfs_bmbt_irec got;
490 struct xfs_iext_cursor icur;
492 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL | XFS_ILOCK_SHARED));
493 ASSERT(xfs_is_reflink_inode(ip));
495 offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
496 if (!xfs_iext_lookup_extent(ip, ifp, offset_fsb, &icur, &got))
498 if (got.br_startoff > offset_fsb)
501 trace_xfs_reflink_find_cow_mapping(ip, offset, 1, XFS_IO_OVERWRITE,
508 * Trim an extent to end at the next CoW reservation past offset_fsb.
511 xfs_reflink_trim_irec_to_next_cow(
512 struct xfs_inode *ip,
513 xfs_fileoff_t offset_fsb,
514 struct xfs_bmbt_irec *imap)
516 struct xfs_ifork *ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
517 struct xfs_bmbt_irec got;
518 struct xfs_iext_cursor icur;
520 if (!xfs_is_reflink_inode(ip))
523 /* Find the extent in the CoW fork. */
524 if (!xfs_iext_lookup_extent(ip, ifp, offset_fsb, &icur, &got))
527 /* This is the extent before; try sliding up one. */
528 if (got.br_startoff < offset_fsb) {
529 if (!xfs_iext_next_extent(ifp, &icur, &got))
533 if (got.br_startoff >= imap->br_startoff + imap->br_blockcount)
536 imap->br_blockcount = got.br_startoff - imap->br_startoff;
537 trace_xfs_reflink_trim_irec(ip, imap);
541 * Cancel CoW reservations for some block range of an inode.
543 * If cancel_real is true this function cancels all COW fork extents for the
544 * inode; if cancel_real is false, real extents are not cleared.
547 xfs_reflink_cancel_cow_blocks(
548 struct xfs_inode *ip,
549 struct xfs_trans **tpp,
550 xfs_fileoff_t offset_fsb,
551 xfs_fileoff_t end_fsb,
554 struct xfs_ifork *ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
555 struct xfs_bmbt_irec got, del;
556 struct xfs_iext_cursor icur;
557 xfs_fsblock_t firstfsb;
558 struct xfs_defer_ops dfops;
561 if (!xfs_is_reflink_inode(ip))
563 if (!xfs_iext_lookup_extent_before(ip, ifp, &end_fsb, &icur, &got))
566 /* Walk backwards until we're out of the I/O range... */
567 while (got.br_startoff + got.br_blockcount > offset_fsb) {
569 xfs_trim_extent(&del, offset_fsb, end_fsb - offset_fsb);
571 /* Extent delete may have bumped ext forward */
572 if (!del.br_blockcount) {
573 xfs_iext_prev(ifp, &icur);
577 trace_xfs_reflink_cancel_cow(ip, &del);
579 if (isnullstartblock(del.br_startblock)) {
580 error = xfs_bmap_del_extent_delay(ip, XFS_COW_FORK,
584 } else if (del.br_state == XFS_EXT_UNWRITTEN || cancel_real) {
585 xfs_trans_ijoin(*tpp, ip, 0);
586 xfs_defer_init(&dfops, &firstfsb);
588 /* Free the CoW orphan record. */
589 error = xfs_refcount_free_cow_extent(ip->i_mount,
590 &dfops, del.br_startblock,
595 xfs_bmap_add_free(ip->i_mount, &dfops,
596 del.br_startblock, del.br_blockcount,
599 /* Update quota accounting */
600 xfs_trans_mod_dquot_byino(*tpp, ip, XFS_TRANS_DQ_BCOUNT,
601 -(long)del.br_blockcount);
603 /* Roll the transaction */
604 xfs_defer_ijoin(&dfops, ip);
605 error = xfs_defer_finish(tpp, &dfops);
607 xfs_defer_cancel(&dfops);
611 /* Remove the mapping from the CoW fork. */
612 xfs_bmap_del_extent_cow(ip, &icur, &got, &del);
615 if (!xfs_iext_get_extent(ifp, &icur, &got))
619 /* clear tag if cow fork is emptied */
621 xfs_inode_clear_cowblocks_tag(ip);
627 * Cancel CoW reservations for some byte range of an inode.
629 * If cancel_real is true this function cancels all COW fork extents for the
630 * inode; if cancel_real is false, real extents are not cleared.
633 xfs_reflink_cancel_cow_range(
634 struct xfs_inode *ip,
639 struct xfs_trans *tp;
640 xfs_fileoff_t offset_fsb;
641 xfs_fileoff_t end_fsb;
644 trace_xfs_reflink_cancel_cow_range(ip, offset, count);
645 ASSERT(xfs_is_reflink_inode(ip));
647 offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
648 if (count == NULLFILEOFF)
649 end_fsb = NULLFILEOFF;
651 end_fsb = XFS_B_TO_FSB(ip->i_mount, offset + count);
653 /* Start a rolling transaction to remove the mappings */
654 error = xfs_trans_alloc(ip->i_mount, &M_RES(ip->i_mount)->tr_write,
659 xfs_ilock(ip, XFS_ILOCK_EXCL);
660 xfs_trans_ijoin(tp, ip, 0);
662 /* Scrape out the old CoW reservations */
663 error = xfs_reflink_cancel_cow_blocks(ip, &tp, offset_fsb, end_fsb,
668 error = xfs_trans_commit(tp);
670 xfs_iunlock(ip, XFS_ILOCK_EXCL);
674 xfs_trans_cancel(tp);
675 xfs_iunlock(ip, XFS_ILOCK_EXCL);
677 trace_xfs_reflink_cancel_cow_range_error(ip, error, _RET_IP_);
682 * Remap parts of a file's data fork after a successful CoW.
686 struct xfs_inode *ip,
690 struct xfs_ifork *ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
691 struct xfs_bmbt_irec got, del;
692 struct xfs_trans *tp;
693 xfs_fileoff_t offset_fsb;
694 xfs_fileoff_t end_fsb;
695 xfs_fsblock_t firstfsb;
696 struct xfs_defer_ops dfops;
698 unsigned int resblks;
700 struct xfs_iext_cursor icur;
702 trace_xfs_reflink_end_cow(ip, offset, count);
704 /* No COW extents? That's easy! */
705 if (ifp->if_bytes == 0)
708 offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
709 end_fsb = XFS_B_TO_FSB(ip->i_mount, offset + count);
712 * Start a rolling transaction to switch the mappings. We're
713 * unlikely ever to have to remap 16T worth of single-block
714 * extents, so just cap the worst case extent count to 2^32-1.
715 * Stick a warning in just in case, and avoid 64-bit division.
717 BUILD_BUG_ON(MAX_RW_COUNT > UINT_MAX);
718 if (end_fsb - offset_fsb > UINT_MAX) {
719 error = -EFSCORRUPTED;
720 xfs_force_shutdown(ip->i_mount, SHUTDOWN_CORRUPT_INCORE);
724 resblks = XFS_NEXTENTADD_SPACE_RES(ip->i_mount,
725 (unsigned int)(end_fsb - offset_fsb),
727 error = xfs_trans_alloc(ip->i_mount, &M_RES(ip->i_mount)->tr_write,
732 xfs_ilock(ip, XFS_ILOCK_EXCL);
733 xfs_trans_ijoin(tp, ip, 0);
736 * In case of racing, overlapping AIO writes no COW extents might be
737 * left by the time I/O completes for the loser of the race. In that
740 if (!xfs_iext_lookup_extent_before(ip, ifp, &end_fsb, &icur, &got))
743 /* Walk backwards until we're out of the I/O range... */
744 while (got.br_startoff + got.br_blockcount > offset_fsb) {
746 xfs_trim_extent(&del, offset_fsb, end_fsb - offset_fsb);
748 /* Extent delete may have bumped ext forward */
749 if (!del.br_blockcount) {
750 xfs_iext_prev(ifp, &icur);
754 ASSERT(!isnullstartblock(got.br_startblock));
757 * Don't remap unwritten extents; these are
758 * speculatively preallocated CoW extents that have been
759 * allocated but have not yet been involved in a write.
761 if (got.br_state == XFS_EXT_UNWRITTEN) {
762 xfs_iext_prev(ifp, &icur);
766 /* Unmap the old blocks in the data fork. */
767 xfs_defer_init(&dfops, &firstfsb);
768 rlen = del.br_blockcount;
769 error = __xfs_bunmapi(tp, ip, del.br_startoff, &rlen, 0, 1,
774 /* Trim the extent to whatever got unmapped. */
776 xfs_trim_extent(&del, del.br_startoff + rlen,
777 del.br_blockcount - rlen);
779 trace_xfs_reflink_cow_remap(ip, &del);
781 /* Free the CoW orphan record. */
782 error = xfs_refcount_free_cow_extent(tp->t_mountp, &dfops,
783 del.br_startblock, del.br_blockcount);
787 /* Map the new blocks into the data fork. */
788 error = xfs_bmap_map_extent(tp->t_mountp, &dfops, ip, &del);
792 /* Remove the mapping from the CoW fork. */
793 xfs_bmap_del_extent_cow(ip, &icur, &got, &del);
795 xfs_defer_ijoin(&dfops, ip);
796 error = xfs_defer_finish(&tp, &dfops);
800 if (!xfs_iext_get_extent(ifp, &icur, &got))
804 error = xfs_trans_commit(tp);
805 xfs_iunlock(ip, XFS_ILOCK_EXCL);
811 xfs_defer_cancel(&dfops);
813 xfs_trans_cancel(tp);
814 xfs_iunlock(ip, XFS_ILOCK_EXCL);
816 trace_xfs_reflink_end_cow_error(ip, error, _RET_IP_);
821 * Free leftover CoW reservations that didn't get cleaned out.
824 xfs_reflink_recover_cow(
825 struct xfs_mount *mp)
830 if (!xfs_sb_version_hasreflink(&mp->m_sb))
833 for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
834 error = xfs_refcount_recover_cow_leftovers(mp, agno);
843 * Reflinking (Block) Ranges of Two Files Together
845 * First, ensure that the reflink flag is set on both inodes. The flag is an
846 * optimization to avoid unnecessary refcount btree lookups in the write path.
848 * Now we can iteratively remap the range of extents (and holes) in src to the
849 * corresponding ranges in dest. Let drange and srange denote the ranges of
850 * logical blocks in dest and src touched by the reflink operation.
852 * While the length of drange is greater than zero,
853 * - Read src's bmbt at the start of srange ("imap")
854 * - If imap doesn't exist, make imap appear to start at the end of srange
856 * - If imap starts before srange, advance imap to start at srange.
857 * - If imap goes beyond srange, truncate imap to end at the end of srange.
858 * - Punch (imap start - srange start + imap len) blocks from dest at
859 * offset (drange start).
860 * - If imap points to a real range of pblks,
861 * > Increase the refcount of the imap's pblks
862 * > Map imap's pblks into dest at the offset
863 * (drange start + imap start - srange start)
864 * - Advance drange and srange by (imap start - srange start + imap len)
866 * Finally, if the reflink made dest longer, update both the in-core and
867 * on-disk file sizes.
869 * ASCII Art Demonstration:
871 * Let's say we want to reflink this source file:
873 * ----SSSSSSS-SSSSS----SSSSSS (src file)
874 * <-------------------->
876 * into this destination file:
878 * --DDDDDDDDDDDDDDDDDDD--DDD (dest file)
879 * <-------------------->
880 * '-' means a hole, and 'S' and 'D' are written blocks in the src and dest.
881 * Observe that the range has different logical offsets in either file.
883 * Consider that the first extent in the source file doesn't line up with our
884 * reflink range. Unmapping and remapping are separate operations, so we can
885 * unmap more blocks from the destination file than we remap.
887 * ----SSSSSSS-SSSSS----SSSSSS
889 * --DDDDD---------DDDDD--DDD
892 * Now remap the source extent into the destination file:
894 * ----SSSSSSS-SSSSS----SSSSSS
896 * --DDDDD--SSSSSSSDDDDD--DDD
899 * Do likewise with the second hole and extent in our range. Holes in the
900 * unmap range don't affect our operation.
902 * ----SSSSSSS-SSSSS----SSSSSS
904 * --DDDDD--SSSSSSS-SSSSS-DDD
907 * Finally, unmap and remap part of the third extent. This will increase the
908 * size of the destination file.
910 * ----SSSSSSS-SSSSS----SSSSSS
912 * --DDDDD--SSSSSSS-SSSSS----SSS
915 * Once we update the destination file's i_size, we're done.
919 * Ensure the reflink bit is set in both inodes.
922 xfs_reflink_set_inode_flag(
923 struct xfs_inode *src,
924 struct xfs_inode *dest)
926 struct xfs_mount *mp = src->i_mount;
928 struct xfs_trans *tp;
930 if (xfs_is_reflink_inode(src) && xfs_is_reflink_inode(dest))
933 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
937 /* Lock both files against IO */
938 if (src->i_ino == dest->i_ino)
939 xfs_ilock(src, XFS_ILOCK_EXCL);
941 xfs_lock_two_inodes(src, dest, XFS_ILOCK_EXCL);
943 if (!xfs_is_reflink_inode(src)) {
944 trace_xfs_reflink_set_inode_flag(src);
945 xfs_trans_ijoin(tp, src, XFS_ILOCK_EXCL);
946 src->i_d.di_flags2 |= XFS_DIFLAG2_REFLINK;
947 xfs_trans_log_inode(tp, src, XFS_ILOG_CORE);
948 xfs_ifork_init_cow(src);
950 xfs_iunlock(src, XFS_ILOCK_EXCL);
952 if (src->i_ino == dest->i_ino)
955 if (!xfs_is_reflink_inode(dest)) {
956 trace_xfs_reflink_set_inode_flag(dest);
957 xfs_trans_ijoin(tp, dest, XFS_ILOCK_EXCL);
958 dest->i_d.di_flags2 |= XFS_DIFLAG2_REFLINK;
959 xfs_trans_log_inode(tp, dest, XFS_ILOG_CORE);
960 xfs_ifork_init_cow(dest);
962 xfs_iunlock(dest, XFS_ILOCK_EXCL);
965 error = xfs_trans_commit(tp);
971 trace_xfs_reflink_set_inode_flag_error(dest, error, _RET_IP_);
976 * Update destination inode size & cowextsize hint, if necessary.
979 xfs_reflink_update_dest(
980 struct xfs_inode *dest,
982 xfs_extlen_t cowextsize,
985 struct xfs_mount *mp = dest->i_mount;
986 struct xfs_trans *tp;
989 if (is_dedupe && newlen <= i_size_read(VFS_I(dest)) && cowextsize == 0)
992 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
996 xfs_ilock(dest, XFS_ILOCK_EXCL);
997 xfs_trans_ijoin(tp, dest, XFS_ILOCK_EXCL);
999 if (newlen > i_size_read(VFS_I(dest))) {
1000 trace_xfs_reflink_update_inode_size(dest, newlen);
1001 i_size_write(VFS_I(dest), newlen);
1002 dest->i_d.di_size = newlen;
1006 dest->i_d.di_cowextsize = cowextsize;
1007 dest->i_d.di_flags2 |= XFS_DIFLAG2_COWEXTSIZE;
1011 xfs_trans_ichgtime(tp, dest,
1012 XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1014 xfs_trans_log_inode(tp, dest, XFS_ILOG_CORE);
1016 error = xfs_trans_commit(tp);
1022 trace_xfs_reflink_update_inode_size_error(dest, error, _RET_IP_);
1027 * Do we have enough reserve in this AG to handle a reflink? The refcount
1028 * btree already reserved all the space it needs, but the rmap btree can grow
1029 * infinitely, so we won't allow more reflinks when the AG is down to the
1033 xfs_reflink_ag_has_free_space(
1034 struct xfs_mount *mp,
1035 xfs_agnumber_t agno)
1037 struct xfs_perag *pag;
1040 if (!xfs_sb_version_hasrmapbt(&mp->m_sb))
1043 pag = xfs_perag_get(mp, agno);
1044 if (xfs_ag_resv_critical(pag, XFS_AG_RESV_AGFL) ||
1045 xfs_ag_resv_critical(pag, XFS_AG_RESV_METADATA))
1052 * Unmap a range of blocks from a file, then map other blocks into the hole.
1053 * The range to unmap is (destoff : destoff + srcioff + irec->br_blockcount).
1054 * The extent irec is mapped into dest at irec->br_startoff.
1057 xfs_reflink_remap_extent(
1058 struct xfs_inode *ip,
1059 struct xfs_bmbt_irec *irec,
1060 xfs_fileoff_t destoff,
1061 xfs_off_t new_isize)
1063 struct xfs_mount *mp = ip->i_mount;
1064 bool real_extent = xfs_bmap_is_real_extent(irec);
1065 struct xfs_trans *tp;
1066 xfs_fsblock_t firstfsb;
1067 unsigned int resblks;
1068 struct xfs_defer_ops dfops;
1069 struct xfs_bmbt_irec uirec;
1071 xfs_filblks_t unmap_len;
1075 unmap_len = irec->br_startoff + irec->br_blockcount - destoff;
1076 trace_xfs_reflink_punch_range(ip, destoff, unmap_len);
1078 /* No reflinking if we're low on space */
1080 error = xfs_reflink_ag_has_free_space(mp,
1081 XFS_FSB_TO_AGNO(mp, irec->br_startblock));
1086 /* Start a rolling transaction to switch the mappings */
1087 resblks = XFS_EXTENTADD_SPACE_RES(ip->i_mount, XFS_DATA_FORK);
1088 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, resblks, 0, 0, &tp);
1092 xfs_ilock(ip, XFS_ILOCK_EXCL);
1093 xfs_trans_ijoin(tp, ip, 0);
1095 /* If we're not just clearing space, then do we have enough quota? */
1097 error = xfs_trans_reserve_quota_nblks(tp, ip,
1098 irec->br_blockcount, 0, XFS_QMOPT_RES_REGBLKS);
1103 trace_xfs_reflink_remap(ip, irec->br_startoff,
1104 irec->br_blockcount, irec->br_startblock);
1106 /* Unmap the old blocks in the data fork. */
1109 xfs_defer_init(&dfops, &firstfsb);
1110 error = __xfs_bunmapi(tp, ip, destoff, &rlen, 0, 1,
1116 * Trim the extent to whatever got unmapped.
1117 * Remember, bunmapi works backwards.
1119 uirec.br_startblock = irec->br_startblock + rlen;
1120 uirec.br_startoff = irec->br_startoff + rlen;
1121 uirec.br_blockcount = unmap_len - rlen;
1124 /* If this isn't a real mapping, we're done. */
1125 if (!real_extent || uirec.br_blockcount == 0)
1128 trace_xfs_reflink_remap(ip, uirec.br_startoff,
1129 uirec.br_blockcount, uirec.br_startblock);
1131 /* Update the refcount tree */
1132 error = xfs_refcount_increase_extent(mp, &dfops, &uirec);
1136 /* Map the new blocks into the data fork. */
1137 error = xfs_bmap_map_extent(mp, &dfops, ip, &uirec);
1141 /* Update quota accounting. */
1142 xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_BCOUNT,
1143 uirec.br_blockcount);
1145 /* Update dest isize if needed. */
1146 newlen = XFS_FSB_TO_B(mp,
1147 uirec.br_startoff + uirec.br_blockcount);
1148 newlen = min_t(xfs_off_t, newlen, new_isize);
1149 if (newlen > i_size_read(VFS_I(ip))) {
1150 trace_xfs_reflink_update_inode_size(ip, newlen);
1151 i_size_write(VFS_I(ip), newlen);
1152 ip->i_d.di_size = newlen;
1153 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1157 /* Process all the deferred stuff. */
1158 xfs_defer_ijoin(&dfops, ip);
1159 error = xfs_defer_finish(&tp, &dfops);
1164 error = xfs_trans_commit(tp);
1165 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1171 xfs_defer_cancel(&dfops);
1173 xfs_trans_cancel(tp);
1174 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1176 trace_xfs_reflink_remap_extent_error(ip, error, _RET_IP_);
1181 * Iteratively remap one file's extents (and holes) to another's.
1184 xfs_reflink_remap_blocks(
1185 struct xfs_inode *src,
1186 xfs_fileoff_t srcoff,
1187 struct xfs_inode *dest,
1188 xfs_fileoff_t destoff,
1190 xfs_off_t new_isize)
1192 struct xfs_bmbt_irec imap;
1195 xfs_filblks_t range_len;
1197 /* drange = (destoff, destoff + len); srange = (srcoff, srcoff + len) */
1199 trace_xfs_reflink_remap_blocks_loop(src, srcoff, len,
1201 /* Read extent from the source file */
1203 xfs_ilock(src, XFS_ILOCK_EXCL);
1204 error = xfs_bmapi_read(src, srcoff, len, &imap, &nimaps, 0);
1205 xfs_iunlock(src, XFS_ILOCK_EXCL);
1208 ASSERT(nimaps == 1);
1210 trace_xfs_reflink_remap_imap(src, srcoff, len, XFS_IO_OVERWRITE,
1213 /* Translate imap into the destination file. */
1214 range_len = imap.br_startoff + imap.br_blockcount - srcoff;
1215 imap.br_startoff += destoff - srcoff;
1217 /* Clear dest from destoff to the end of imap and map it in. */
1218 error = xfs_reflink_remap_extent(dest, &imap, destoff,
1223 if (fatal_signal_pending(current)) {
1228 /* Advance drange/srange */
1229 srcoff += range_len;
1230 destoff += range_len;
1237 trace_xfs_reflink_remap_blocks_error(dest, error, _RET_IP_);
1242 * Link a range of blocks from one file to another.
1245 xfs_reflink_remap_range(
1246 struct file *file_in,
1248 struct file *file_out,
1253 struct inode *inode_in = file_inode(file_in);
1254 struct xfs_inode *src = XFS_I(inode_in);
1255 struct inode *inode_out = file_inode(file_out);
1256 struct xfs_inode *dest = XFS_I(inode_out);
1257 struct xfs_mount *mp = src->i_mount;
1258 bool same_inode = (inode_in == inode_out);
1259 xfs_fileoff_t sfsbno, dfsbno;
1260 xfs_filblks_t fsblen;
1261 xfs_extlen_t cowextsize;
1264 if (!xfs_sb_version_hasreflink(&mp->m_sb))
1267 if (XFS_FORCED_SHUTDOWN(mp))
1270 /* Lock both files against IO */
1271 lock_two_nondirectories(inode_in, inode_out);
1273 xfs_ilock(src, XFS_MMAPLOCK_EXCL);
1275 xfs_lock_two_inodes(src, dest, XFS_MMAPLOCK_EXCL);
1277 /* Check file eligibility and prepare for block sharing. */
1279 /* Don't reflink realtime inodes */
1280 if (XFS_IS_REALTIME_INODE(src) || XFS_IS_REALTIME_INODE(dest))
1283 /* Don't share DAX file data for now. */
1284 if (IS_DAX(inode_in) || IS_DAX(inode_out))
1287 ret = vfs_clone_file_prep_inodes(inode_in, pos_in, inode_out, pos_out,
1292 trace_xfs_reflink_remap_range(src, pos_in, len, dest, pos_out);
1294 /* Set flags and remap blocks. */
1295 ret = xfs_reflink_set_inode_flag(src, dest);
1299 dfsbno = XFS_B_TO_FSBT(mp, pos_out);
1300 sfsbno = XFS_B_TO_FSBT(mp, pos_in);
1301 fsblen = XFS_B_TO_FSB(mp, len);
1302 ret = xfs_reflink_remap_blocks(src, sfsbno, dest, dfsbno, fsblen,
1307 /* Zap any page cache for the destination file's range. */
1308 truncate_inode_pages_range(&inode_out->i_data, pos_out,
1309 PAGE_ALIGN(pos_out + len) - 1);
1312 * Carry the cowextsize hint from src to dest if we're sharing the
1313 * entire source file to the entire destination file, the source file
1314 * has a cowextsize hint, and the destination file does not.
1317 if (pos_in == 0 && len == i_size_read(inode_in) &&
1318 (src->i_d.di_flags2 & XFS_DIFLAG2_COWEXTSIZE) &&
1319 pos_out == 0 && len >= i_size_read(inode_out) &&
1320 !(dest->i_d.di_flags2 & XFS_DIFLAG2_COWEXTSIZE))
1321 cowextsize = src->i_d.di_cowextsize;
1323 ret = xfs_reflink_update_dest(dest, pos_out + len, cowextsize,
1327 xfs_iunlock(src, XFS_MMAPLOCK_EXCL);
1329 xfs_iunlock(dest, XFS_MMAPLOCK_EXCL);
1330 unlock_two_nondirectories(inode_in, inode_out);
1332 trace_xfs_reflink_remap_range_error(dest, ret, _RET_IP_);
1337 * The user wants to preemptively CoW all shared blocks in this file,
1338 * which enables us to turn off the reflink flag. Iterate all
1339 * extents which are not prealloc/delalloc to see which ranges are
1340 * mentioned in the refcount tree, then read those blocks into the
1341 * pagecache, dirty them, fsync them back out, and then we can update
1342 * the inode flag. What happens if we run out of memory? :)
1345 xfs_reflink_dirty_extents(
1346 struct xfs_inode *ip,
1351 struct xfs_mount *mp = ip->i_mount;
1352 xfs_agnumber_t agno;
1353 xfs_agblock_t agbno;
1359 struct xfs_bmbt_irec map[2];
1363 while (end - fbno > 0) {
1366 * Look for extents in the file. Skip holes, delalloc, or
1367 * unwritten extents; they can't be reflinked.
1369 error = xfs_bmapi_read(ip, fbno, end - fbno, map, &nmaps, 0);
1374 if (!xfs_bmap_is_real_extent(&map[0]))
1378 while (map[1].br_blockcount) {
1379 agno = XFS_FSB_TO_AGNO(mp, map[1].br_startblock);
1380 agbno = XFS_FSB_TO_AGBNO(mp, map[1].br_startblock);
1381 aglen = map[1].br_blockcount;
1383 error = xfs_reflink_find_shared(mp, NULL, agno, agbno,
1384 aglen, &rbno, &rlen, true);
1387 if (rbno == NULLAGBLOCK)
1390 /* Dirty the pages */
1391 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1392 fpos = XFS_FSB_TO_B(mp, map[1].br_startoff +
1394 flen = XFS_FSB_TO_B(mp, rlen);
1395 if (fpos + flen > isize)
1396 flen = isize - fpos;
1397 error = iomap_file_dirty(VFS_I(ip), fpos, flen,
1399 xfs_ilock(ip, XFS_ILOCK_EXCL);
1403 map[1].br_blockcount -= (rbno - agbno + rlen);
1404 map[1].br_startoff += (rbno - agbno + rlen);
1405 map[1].br_startblock += (rbno - agbno + rlen);
1409 fbno = map[0].br_startoff + map[0].br_blockcount;
1415 /* Does this inode need the reflink flag? */
1417 xfs_reflink_inode_has_shared_extents(
1418 struct xfs_trans *tp,
1419 struct xfs_inode *ip,
1422 struct xfs_bmbt_irec got;
1423 struct xfs_mount *mp = ip->i_mount;
1424 struct xfs_ifork *ifp;
1425 xfs_agnumber_t agno;
1426 xfs_agblock_t agbno;
1430 struct xfs_iext_cursor icur;
1434 ifp = XFS_IFORK_PTR(ip, XFS_DATA_FORK);
1435 if (!(ifp->if_flags & XFS_IFEXTENTS)) {
1436 error = xfs_iread_extents(tp, ip, XFS_DATA_FORK);
1441 *has_shared = false;
1442 found = xfs_iext_lookup_extent(ip, ifp, 0, &icur, &got);
1444 if (isnullstartblock(got.br_startblock) ||
1445 got.br_state != XFS_EXT_NORM)
1447 agno = XFS_FSB_TO_AGNO(mp, got.br_startblock);
1448 agbno = XFS_FSB_TO_AGBNO(mp, got.br_startblock);
1449 aglen = got.br_blockcount;
1451 error = xfs_reflink_find_shared(mp, tp, agno, agbno, aglen,
1452 &rbno, &rlen, false);
1455 /* Is there still a shared block here? */
1456 if (rbno != NULLAGBLOCK) {
1461 found = xfs_iext_next_extent(ifp, &icur, &got);
1467 /* Clear the inode reflink flag if there are no shared extents. */
1469 xfs_reflink_clear_inode_flag(
1470 struct xfs_inode *ip,
1471 struct xfs_trans **tpp)
1476 ASSERT(xfs_is_reflink_inode(ip));
1478 error = xfs_reflink_inode_has_shared_extents(*tpp, ip, &needs_flag);
1479 if (error || needs_flag)
1483 * We didn't find any shared blocks so turn off the reflink flag.
1484 * First, get rid of any leftover CoW mappings.
1486 error = xfs_reflink_cancel_cow_blocks(ip, tpp, 0, NULLFILEOFF, true);
1490 /* Clear the inode flag. */
1491 trace_xfs_reflink_unset_inode_flag(ip);
1492 ip->i_d.di_flags2 &= ~XFS_DIFLAG2_REFLINK;
1493 xfs_inode_clear_cowblocks_tag(ip);
1494 xfs_trans_ijoin(*tpp, ip, 0);
1495 xfs_trans_log_inode(*tpp, ip, XFS_ILOG_CORE);
1501 * Clear the inode reflink flag if there are no shared extents and the size
1505 xfs_reflink_try_clear_inode_flag(
1506 struct xfs_inode *ip)
1508 struct xfs_mount *mp = ip->i_mount;
1509 struct xfs_trans *tp;
1512 /* Start a rolling transaction to remove the mappings */
1513 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, 0, 0, 0, &tp);
1517 xfs_ilock(ip, XFS_ILOCK_EXCL);
1518 xfs_trans_ijoin(tp, ip, 0);
1520 error = xfs_reflink_clear_inode_flag(ip, &tp);
1524 error = xfs_trans_commit(tp);
1528 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1531 xfs_trans_cancel(tp);
1533 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1538 * Pre-COW all shared blocks within a given byte range of a file and turn off
1539 * the reflink flag if we unshare all of the file's blocks.
1542 xfs_reflink_unshare(
1543 struct xfs_inode *ip,
1547 struct xfs_mount *mp = ip->i_mount;
1553 if (!xfs_is_reflink_inode(ip))
1556 trace_xfs_reflink_unshare(ip, offset, len);
1558 inode_dio_wait(VFS_I(ip));
1560 /* Try to CoW the selected ranges */
1561 xfs_ilock(ip, XFS_ILOCK_EXCL);
1562 fbno = XFS_B_TO_FSBT(mp, offset);
1563 isize = i_size_read(VFS_I(ip));
1564 end = XFS_B_TO_FSB(mp, offset + len);
1565 error = xfs_reflink_dirty_extents(ip, fbno, end, isize);
1568 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1570 /* Wait for the IO to finish */
1571 error = filemap_write_and_wait(VFS_I(ip)->i_mapping);
1575 /* Turn off the reflink flag if possible. */
1576 error = xfs_reflink_try_clear_inode_flag(ip);
1583 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1585 trace_xfs_reflink_unshare_error(ip, error, _RET_IP_);