2 * Copyright (C) 2016 Oracle. All Rights Reserved.
4 * Author: Darrick J. Wong <darrick.wong@oracle.com>
6 * This program is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU General Public License
8 * as published by the Free Software Foundation; either version 2
9 * of the License, or (at your option) any later version.
11 * This program is distributed in the hope that it would be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 * GNU General Public License for more details.
16 * You should have received a copy of the GNU General Public License
17 * along with this program; if not, write the Free Software Foundation,
18 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301, USA.
22 #include "xfs_shared.h"
23 #include "xfs_format.h"
24 #include "xfs_log_format.h"
25 #include "xfs_trans_resv.h"
26 #include "xfs_mount.h"
27 #include "xfs_defer.h"
28 #include "xfs_da_format.h"
29 #include "xfs_da_btree.h"
30 #include "xfs_inode.h"
31 #include "xfs_trans.h"
32 #include "xfs_inode_item.h"
34 #include "xfs_bmap_util.h"
35 #include "xfs_error.h"
37 #include "xfs_dir2_priv.h"
38 #include "xfs_ioctl.h"
39 #include "xfs_trace.h"
41 #include "xfs_icache.h"
43 #include "xfs_btree.h"
44 #include "xfs_refcount_btree.h"
45 #include "xfs_refcount.h"
46 #include "xfs_bmap_btree.h"
47 #include "xfs_trans_space.h"
49 #include "xfs_alloc.h"
50 #include "xfs_quota_defs.h"
51 #include "xfs_quota.h"
52 #include "xfs_btree.h"
53 #include "xfs_bmap_btree.h"
54 #include "xfs_reflink.h"
55 #include "xfs_iomap.h"
56 #include "xfs_rmap_btree.h"
58 #include "xfs_ag_resv.h"
61 * Copy on Write of Shared Blocks
63 * XFS must preserve "the usual" file semantics even when two files share
64 * the same physical blocks. This means that a write to one file must not
65 * alter the blocks in a different file; the way that we'll do that is
66 * through the use of a copy-on-write mechanism. At a high level, that
67 * means that when we want to write to a shared block, we allocate a new
68 * block, write the data to the new block, and if that succeeds we map the
69 * new block into the file.
71 * XFS provides a "delayed allocation" mechanism that defers the allocation
72 * of disk blocks to dirty-but-not-yet-mapped file blocks as long as
73 * possible. This reduces fragmentation by enabling the filesystem to ask
74 * for bigger chunks less often, which is exactly what we want for CoW.
76 * The delalloc mechanism begins when the kernel wants to make a block
77 * writable (write_begin or page_mkwrite). If the offset is not mapped, we
78 * create a delalloc mapping, which is a regular in-core extent, but without
79 * a real startblock. (For delalloc mappings, the startblock encodes both
80 * a flag that this is a delalloc mapping, and a worst-case estimate of how
81 * many blocks might be required to put the mapping into the BMBT.) delalloc
82 * mappings are a reservation against the free space in the filesystem;
83 * adjacent mappings can also be combined into fewer larger mappings.
85 * As an optimization, the CoW extent size hint (cowextsz) creates
86 * outsized aligned delalloc reservations in the hope of landing out of
87 * order nearby CoW writes in a single extent on disk, thereby reducing
88 * fragmentation and improving future performance.
90 * D: --RRRRRRSSSRRRRRRRR--- (data fork)
91 * C: ------DDDDDDD--------- (CoW fork)
93 * When dirty pages are being written out (typically in writepage), the
94 * delalloc reservations are converted into unwritten mappings by
95 * allocating blocks and replacing the delalloc mapping with real ones.
96 * A delalloc mapping can be replaced by several unwritten ones if the
97 * free space is fragmented.
99 * D: --RRRRRRSSSRRRRRRRR---
100 * C: ------UUUUUUU---------
102 * We want to adapt the delalloc mechanism for copy-on-write, since the
103 * write paths are similar. The first two steps (creating the reservation
104 * and allocating the blocks) are exactly the same as delalloc except that
105 * the mappings must be stored in a separate CoW fork because we do not want
106 * to disturb the mapping in the data fork until we're sure that the write
107 * succeeded. IO completion in this case is the process of removing the old
108 * mapping from the data fork and moving the new mapping from the CoW fork to
109 * the data fork. This will be discussed shortly.
111 * For now, unaligned directio writes will be bounced back to the page cache.
112 * Block-aligned directio writes will use the same mechanism as buffered
115 * Just prior to submitting the actual disk write requests, we convert
116 * the extents representing the range of the file actually being written
117 * (as opposed to extra pieces created for the cowextsize hint) to real
118 * extents. This will become important in the next step:
120 * D: --RRRRRRSSSRRRRRRRR---
121 * C: ------UUrrUUU---------
123 * CoW remapping must be done after the data block write completes,
124 * because we don't want to destroy the old data fork map until we're sure
125 * the new block has been written. Since the new mappings are kept in a
126 * separate fork, we can simply iterate these mappings to find the ones
127 * that cover the file blocks that we just CoW'd. For each extent, simply
128 * unmap the corresponding range in the data fork, map the new range into
129 * the data fork, and remove the extent from the CoW fork. Because of
130 * the presence of the cowextsize hint, however, we must be careful
131 * only to remap the blocks that we've actually written out -- we must
132 * never remap delalloc reservations nor CoW staging blocks that have
133 * yet to be written. This corresponds exactly to the real extents in
136 * D: --RRRRRRrrSRRRRRRRR---
137 * C: ------UU--UUU---------
139 * Since the remapping operation can be applied to an arbitrary file
140 * range, we record the need for the remap step as a flag in the ioend
141 * instead of declaring a new IO type. This is required for direct io
142 * because we only have ioend for the whole dio, and we have to be able to
143 * remember the presence of unwritten blocks and CoW blocks with a single
144 * ioend structure. Better yet, the more ground we can cover with one
149 * Given an AG extent, find the lowest-numbered run of shared blocks
150 * within that range and return the range in fbno/flen. If
151 * find_end_of_shared is true, return the longest contiguous extent of
152 * shared blocks. If there are no shared extents, fbno and flen will
153 * be set to NULLAGBLOCK and 0, respectively.
156 xfs_reflink_find_shared(
157 struct xfs_mount *mp,
163 bool find_end_of_shared)
165 struct xfs_buf *agbp;
166 struct xfs_btree_cur *cur;
169 error = xfs_alloc_read_agf(mp, NULL, agno, 0, &agbp);
173 cur = xfs_refcountbt_init_cursor(mp, NULL, agbp, agno, NULL);
175 error = xfs_refcount_find_shared(cur, agbno, aglen, fbno, flen,
178 xfs_btree_del_cursor(cur, error ? XFS_BTREE_ERROR : XFS_BTREE_NOERROR);
185 * Trim the mapping to the next block where there's a change in the
186 * shared/unshared status. More specifically, this means that we
187 * find the lowest-numbered extent of shared blocks that coincides with
188 * the given block mapping. If the shared extent overlaps the start of
189 * the mapping, trim the mapping to the end of the shared extent. If
190 * the shared region intersects the mapping, trim the mapping to the
191 * start of the shared extent. If there are no shared regions that
192 * overlap, just return the original extent.
195 xfs_reflink_trim_around_shared(
196 struct xfs_inode *ip,
197 struct xfs_bmbt_irec *irec,
208 /* Holes, unwritten, and delalloc extents cannot be shared */
209 if (!xfs_is_reflink_inode(ip) || !xfs_bmap_is_real_extent(irec)) {
214 trace_xfs_reflink_trim_around_shared(ip, irec);
216 agno = XFS_FSB_TO_AGNO(ip->i_mount, irec->br_startblock);
217 agbno = XFS_FSB_TO_AGBNO(ip->i_mount, irec->br_startblock);
218 aglen = irec->br_blockcount;
220 error = xfs_reflink_find_shared(ip->i_mount, agno, agbno,
221 aglen, &fbno, &flen, true);
225 *shared = *trimmed = false;
226 if (fbno == NULLAGBLOCK) {
227 /* No shared blocks at all. */
229 } else if (fbno == agbno) {
231 * The start of this extent is shared. Truncate the
232 * mapping at the end of the shared region so that a
233 * subsequent iteration starts at the start of the
236 irec->br_blockcount = flen;
243 * There's a shared extent midway through this extent.
244 * Truncate the mapping at the start of the shared
245 * extent so that a subsequent iteration starts at the
246 * start of the shared region.
248 irec->br_blockcount = fbno - agbno;
255 * Trim the passed in imap to the next shared/unshared extent boundary, and
256 * if imap->br_startoff points to a shared extent reserve space for it in the
257 * COW fork. In this case *shared is set to true, else to false.
259 * Note that imap will always contain the block numbers for the existing blocks
260 * in the data fork, as the upper layers need them for read-modify-write
264 xfs_reflink_reserve_cow(
265 struct xfs_inode *ip,
266 struct xfs_bmbt_irec *imap,
269 struct xfs_ifork *ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
270 struct xfs_bmbt_irec got;
272 bool eof = false, trimmed;
276 * Search the COW fork extent list first. This serves two purposes:
277 * first this implement the speculative preallocation using cowextisze,
278 * so that we also unshared block adjacent to shared blocks instead
279 * of just the shared blocks themselves. Second the lookup in the
280 * extent list is generally faster than going out to the shared extent
284 if (!xfs_iext_lookup_extent(ip, ifp, imap->br_startoff, &idx, &got))
286 if (!eof && got.br_startoff <= imap->br_startoff) {
287 trace_xfs_reflink_cow_found(ip, imap);
288 xfs_trim_extent(imap, got.br_startoff, got.br_blockcount);
294 /* Trim the mapping to the nearest shared extent boundary. */
295 error = xfs_reflink_trim_around_shared(ip, imap, shared, &trimmed);
299 /* Not shared? Just report the (potentially capped) extent. */
304 * Fork all the shared blocks from our write offset until the end of
307 error = xfs_qm_dqattach_locked(ip, 0);
311 error = xfs_bmapi_reserve_delalloc(ip, XFS_COW_FORK, imap->br_startoff,
312 imap->br_blockcount, 0, &got, &idx, eof);
313 if (error == -ENOSPC || error == -EDQUOT)
314 trace_xfs_reflink_cow_enospc(ip, imap);
318 trace_xfs_reflink_cow_alloc(ip, &got);
322 /* Convert part of an unwritten CoW extent to a real one. */
324 xfs_reflink_convert_cow_extent(
325 struct xfs_inode *ip,
326 struct xfs_bmbt_irec *imap,
327 xfs_fileoff_t offset_fsb,
328 xfs_filblks_t count_fsb,
329 struct xfs_defer_ops *dfops)
331 xfs_fsblock_t first_block;
334 if (imap->br_state == XFS_EXT_NORM)
337 xfs_trim_extent(imap, offset_fsb, count_fsb);
338 trace_xfs_reflink_convert_cow(ip, imap);
339 if (imap->br_blockcount == 0)
341 return xfs_bmapi_write(NULL, ip, imap->br_startoff, imap->br_blockcount,
342 XFS_BMAPI_COWFORK | XFS_BMAPI_CONVERT, &first_block,
343 0, imap, &nimaps, dfops);
346 /* Convert all of the unwritten CoW extents in a file's range to real ones. */
348 xfs_reflink_convert_cow(
349 struct xfs_inode *ip,
353 struct xfs_bmbt_irec got;
354 struct xfs_defer_ops dfops;
355 struct xfs_mount *mp = ip->i_mount;
356 struct xfs_ifork *ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
357 xfs_fileoff_t offset_fsb = XFS_B_TO_FSBT(mp, offset);
358 xfs_fileoff_t end_fsb = XFS_B_TO_FSB(mp, offset + count);
363 xfs_ilock(ip, XFS_ILOCK_EXCL);
365 /* Convert all the extents to real from unwritten. */
366 for (found = xfs_iext_lookup_extent(ip, ifp, offset_fsb, &idx, &got);
367 found && got.br_startoff < end_fsb;
368 found = xfs_iext_get_extent(ifp, ++idx, &got)) {
369 error = xfs_reflink_convert_cow_extent(ip, &got, offset_fsb,
370 end_fsb - offset_fsb, &dfops);
376 xfs_iunlock(ip, XFS_ILOCK_EXCL);
380 /* Allocate all CoW reservations covering a range of blocks in a file. */
382 xfs_reflink_allocate_cow(
383 struct xfs_inode *ip,
384 struct xfs_bmbt_irec *imap,
388 struct xfs_mount *mp = ip->i_mount;
389 xfs_fileoff_t offset_fsb = imap->br_startoff;
390 xfs_filblks_t count_fsb = imap->br_blockcount;
391 struct xfs_bmbt_irec got;
392 struct xfs_defer_ops dfops;
393 struct xfs_trans *tp = NULL;
394 xfs_fsblock_t first_block;
395 int nimaps, error = 0;
397 xfs_filblks_t resaligned;
398 xfs_extlen_t resblks = 0;
402 ASSERT(xfs_is_reflink_inode(ip));
403 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL | XFS_ILOCK_SHARED));
406 * Even if the extent is not shared we might have a preallocation for
407 * it in the COW fork. If so use it.
409 if (xfs_iext_lookup_extent(ip, ip->i_cowfp, offset_fsb, &idx, &got) &&
410 got.br_startoff <= offset_fsb) {
413 /* If we have a real allocation in the COW fork we're done. */
414 if (!isnullstartblock(got.br_startblock)) {
415 xfs_trim_extent(&got, offset_fsb, count_fsb);
420 xfs_trim_extent(imap, got.br_startoff, got.br_blockcount);
422 error = xfs_reflink_trim_around_shared(ip, imap, shared, &trimmed);
423 if (error || !*shared)
428 resaligned = xfs_aligned_fsb_count(imap->br_startoff,
429 imap->br_blockcount, xfs_get_cowextsz_hint(ip));
430 resblks = XFS_DIOSTRAT_SPACE_RES(mp, resaligned);
432 xfs_iunlock(ip, *lockmode);
433 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, resblks, 0, 0, &tp);
434 *lockmode = XFS_ILOCK_EXCL;
435 xfs_ilock(ip, *lockmode);
440 error = xfs_qm_dqattach_locked(ip, 0);
446 error = xfs_trans_reserve_quota_nblks(tp, ip, resblks, 0,
447 XFS_QMOPT_RES_REGBLKS);
451 xfs_trans_ijoin(tp, ip, 0);
453 xfs_defer_init(&dfops, &first_block);
456 /* Allocate the entire reservation as unwritten blocks. */
457 error = xfs_bmapi_write(tp, ip, imap->br_startoff, imap->br_blockcount,
458 XFS_BMAPI_COWFORK | XFS_BMAPI_PREALLOC, &first_block,
459 resblks, imap, &nimaps, &dfops);
461 goto out_bmap_cancel;
464 error = xfs_defer_finish(&tp, &dfops, NULL);
466 goto out_bmap_cancel;
468 error = xfs_trans_commit(tp);
472 return xfs_reflink_convert_cow_extent(ip, imap, offset_fsb, count_fsb,
475 xfs_defer_cancel(&dfops);
476 xfs_trans_unreserve_quota_nblks(tp, ip, (long)resblks, 0,
477 XFS_QMOPT_RES_REGBLKS);
480 xfs_trans_cancel(tp);
485 * Find the CoW reservation for a given byte offset of a file.
488 xfs_reflink_find_cow_mapping(
489 struct xfs_inode *ip,
491 struct xfs_bmbt_irec *imap)
493 struct xfs_ifork *ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
494 xfs_fileoff_t offset_fsb;
495 struct xfs_bmbt_irec got;
498 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL | XFS_ILOCK_SHARED));
499 ASSERT(xfs_is_reflink_inode(ip));
501 offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
502 if (!xfs_iext_lookup_extent(ip, ifp, offset_fsb, &idx, &got))
504 if (got.br_startoff > offset_fsb)
507 trace_xfs_reflink_find_cow_mapping(ip, offset, 1, XFS_IO_OVERWRITE,
514 * Trim an extent to end at the next CoW reservation past offset_fsb.
517 xfs_reflink_trim_irec_to_next_cow(
518 struct xfs_inode *ip,
519 xfs_fileoff_t offset_fsb,
520 struct xfs_bmbt_irec *imap)
522 struct xfs_ifork *ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
523 struct xfs_bmbt_irec got;
526 if (!xfs_is_reflink_inode(ip))
529 /* Find the extent in the CoW fork. */
530 if (!xfs_iext_lookup_extent(ip, ifp, offset_fsb, &idx, &got))
533 /* This is the extent before; try sliding up one. */
534 if (got.br_startoff < offset_fsb) {
535 if (!xfs_iext_get_extent(ifp, idx + 1, &got))
539 if (got.br_startoff >= imap->br_startoff + imap->br_blockcount)
542 imap->br_blockcount = got.br_startoff - imap->br_startoff;
543 trace_xfs_reflink_trim_irec(ip, imap);
547 * Cancel CoW reservations for some block range of an inode.
549 * If cancel_real is true this function cancels all COW fork extents for the
550 * inode; if cancel_real is false, real extents are not cleared.
553 xfs_reflink_cancel_cow_blocks(
554 struct xfs_inode *ip,
555 struct xfs_trans **tpp,
556 xfs_fileoff_t offset_fsb,
557 xfs_fileoff_t end_fsb,
560 struct xfs_ifork *ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
561 struct xfs_bmbt_irec got, del;
563 xfs_fsblock_t firstfsb;
564 struct xfs_defer_ops dfops;
567 if (!xfs_is_reflink_inode(ip))
569 if (!xfs_iext_lookup_extent(ip, ifp, offset_fsb, &idx, &got))
572 while (got.br_startoff < end_fsb) {
574 xfs_trim_extent(&del, offset_fsb, end_fsb - offset_fsb);
575 trace_xfs_reflink_cancel_cow(ip, &del);
577 if (isnullstartblock(del.br_startblock)) {
578 error = xfs_bmap_del_extent_delay(ip, XFS_COW_FORK,
582 } else if (del.br_state == XFS_EXT_UNWRITTEN || cancel_real) {
583 xfs_trans_ijoin(*tpp, ip, 0);
584 xfs_defer_init(&dfops, &firstfsb);
586 /* Free the CoW orphan record. */
587 error = xfs_refcount_free_cow_extent(ip->i_mount,
588 &dfops, del.br_startblock,
593 xfs_bmap_add_free(ip->i_mount, &dfops,
594 del.br_startblock, del.br_blockcount,
597 /* Update quota accounting */
598 xfs_trans_mod_dquot_byino(*tpp, ip, XFS_TRANS_DQ_BCOUNT,
599 -(long)del.br_blockcount);
601 /* Roll the transaction */
602 error = xfs_defer_finish(tpp, &dfops, ip);
604 xfs_defer_cancel(&dfops);
608 /* Remove the mapping from the CoW fork. */
609 xfs_bmap_del_extent_cow(ip, &idx, &got, &del);
612 if (!xfs_iext_get_extent(ifp, ++idx, &got))
616 /* clear tag if cow fork is emptied */
618 xfs_inode_clear_cowblocks_tag(ip);
624 * Cancel CoW reservations for some byte range of an inode.
626 * If cancel_real is true this function cancels all COW fork extents for the
627 * inode; if cancel_real is false, real extents are not cleared.
630 xfs_reflink_cancel_cow_range(
631 struct xfs_inode *ip,
636 struct xfs_trans *tp;
637 xfs_fileoff_t offset_fsb;
638 xfs_fileoff_t end_fsb;
641 trace_xfs_reflink_cancel_cow_range(ip, offset, count);
642 ASSERT(xfs_is_reflink_inode(ip));
644 offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
645 if (count == NULLFILEOFF)
646 end_fsb = NULLFILEOFF;
648 end_fsb = XFS_B_TO_FSB(ip->i_mount, offset + count);
650 /* Start a rolling transaction to remove the mappings */
651 error = xfs_trans_alloc(ip->i_mount, &M_RES(ip->i_mount)->tr_write,
656 xfs_ilock(ip, XFS_ILOCK_EXCL);
657 xfs_trans_ijoin(tp, ip, 0);
659 /* Scrape out the old CoW reservations */
660 error = xfs_reflink_cancel_cow_blocks(ip, &tp, offset_fsb, end_fsb,
665 error = xfs_trans_commit(tp);
667 xfs_iunlock(ip, XFS_ILOCK_EXCL);
671 xfs_trans_cancel(tp);
672 xfs_iunlock(ip, XFS_ILOCK_EXCL);
674 trace_xfs_reflink_cancel_cow_range_error(ip, error, _RET_IP_);
679 * Remap parts of a file's data fork after a successful CoW.
683 struct xfs_inode *ip,
687 struct xfs_ifork *ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
688 struct xfs_bmbt_irec got, del;
689 struct xfs_trans *tp;
690 xfs_fileoff_t offset_fsb;
691 xfs_fileoff_t end_fsb;
692 xfs_fsblock_t firstfsb;
693 struct xfs_defer_ops dfops;
695 unsigned int resblks;
699 trace_xfs_reflink_end_cow(ip, offset, count);
701 /* No COW extents? That's easy! */
702 if (ifp->if_bytes == 0)
705 offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
706 end_fsb = XFS_B_TO_FSB(ip->i_mount, offset + count);
709 * Start a rolling transaction to switch the mappings. We're
710 * unlikely ever to have to remap 16T worth of single-block
711 * extents, so just cap the worst case extent count to 2^32-1.
712 * Stick a warning in just in case, and avoid 64-bit division.
714 BUILD_BUG_ON(MAX_RW_COUNT > UINT_MAX);
715 if (end_fsb - offset_fsb > UINT_MAX) {
716 error = -EFSCORRUPTED;
717 xfs_force_shutdown(ip->i_mount, SHUTDOWN_CORRUPT_INCORE);
721 resblks = XFS_NEXTENTADD_SPACE_RES(ip->i_mount,
722 (unsigned int)(end_fsb - offset_fsb),
724 error = xfs_trans_alloc(ip->i_mount, &M_RES(ip->i_mount)->tr_write,
729 xfs_ilock(ip, XFS_ILOCK_EXCL);
730 xfs_trans_ijoin(tp, ip, 0);
732 /* If there is a hole at end_fsb - 1 go to the previous extent */
733 if (!xfs_iext_lookup_extent(ip, ifp, end_fsb - 1, &idx, &got) ||
734 got.br_startoff > end_fsb) {
736 xfs_iext_get_extent(ifp, --idx, &got);
739 /* Walk backwards until we're out of the I/O range... */
740 while (got.br_startoff + got.br_blockcount > offset_fsb) {
742 xfs_trim_extent(&del, offset_fsb, end_fsb - offset_fsb);
744 /* Extent delete may have bumped idx forward */
745 if (!del.br_blockcount) {
750 ASSERT(!isnullstartblock(got.br_startblock));
753 * Don't remap unwritten extents; these are
754 * speculatively preallocated CoW extents that have been
755 * allocated but have not yet been involved in a write.
757 if (got.br_state == XFS_EXT_UNWRITTEN) {
762 /* Unmap the old blocks in the data fork. */
763 xfs_defer_init(&dfops, &firstfsb);
764 rlen = del.br_blockcount;
765 error = __xfs_bunmapi(tp, ip, del.br_startoff, &rlen, 0, 1,
770 /* Trim the extent to whatever got unmapped. */
772 xfs_trim_extent(&del, del.br_startoff + rlen,
773 del.br_blockcount - rlen);
775 trace_xfs_reflink_cow_remap(ip, &del);
777 /* Free the CoW orphan record. */
778 error = xfs_refcount_free_cow_extent(tp->t_mountp, &dfops,
779 del.br_startblock, del.br_blockcount);
783 /* Map the new blocks into the data fork. */
784 error = xfs_bmap_map_extent(tp->t_mountp, &dfops, ip, &del);
788 /* Remove the mapping from the CoW fork. */
789 xfs_bmap_del_extent_cow(ip, &idx, &got, &del);
791 error = xfs_defer_finish(&tp, &dfops, ip);
795 if (!xfs_iext_get_extent(ifp, idx, &got))
799 error = xfs_trans_commit(tp);
800 xfs_iunlock(ip, XFS_ILOCK_EXCL);
806 xfs_defer_cancel(&dfops);
807 xfs_trans_cancel(tp);
808 xfs_iunlock(ip, XFS_ILOCK_EXCL);
810 trace_xfs_reflink_end_cow_error(ip, error, _RET_IP_);
815 * Free leftover CoW reservations that didn't get cleaned out.
818 xfs_reflink_recover_cow(
819 struct xfs_mount *mp)
824 if (!xfs_sb_version_hasreflink(&mp->m_sb))
827 for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
828 error = xfs_refcount_recover_cow_leftovers(mp, agno);
837 * Reflinking (Block) Ranges of Two Files Together
839 * First, ensure that the reflink flag is set on both inodes. The flag is an
840 * optimization to avoid unnecessary refcount btree lookups in the write path.
842 * Now we can iteratively remap the range of extents (and holes) in src to the
843 * corresponding ranges in dest. Let drange and srange denote the ranges of
844 * logical blocks in dest and src touched by the reflink operation.
846 * While the length of drange is greater than zero,
847 * - Read src's bmbt at the start of srange ("imap")
848 * - If imap doesn't exist, make imap appear to start at the end of srange
850 * - If imap starts before srange, advance imap to start at srange.
851 * - If imap goes beyond srange, truncate imap to end at the end of srange.
852 * - Punch (imap start - srange start + imap len) blocks from dest at
853 * offset (drange start).
854 * - If imap points to a real range of pblks,
855 * > Increase the refcount of the imap's pblks
856 * > Map imap's pblks into dest at the offset
857 * (drange start + imap start - srange start)
858 * - Advance drange and srange by (imap start - srange start + imap len)
860 * Finally, if the reflink made dest longer, update both the in-core and
861 * on-disk file sizes.
863 * ASCII Art Demonstration:
865 * Let's say we want to reflink this source file:
867 * ----SSSSSSS-SSSSS----SSSSSS (src file)
868 * <-------------------->
870 * into this destination file:
872 * --DDDDDDDDDDDDDDDDDDD--DDD (dest file)
873 * <-------------------->
874 * '-' means a hole, and 'S' and 'D' are written blocks in the src and dest.
875 * Observe that the range has different logical offsets in either file.
877 * Consider that the first extent in the source file doesn't line up with our
878 * reflink range. Unmapping and remapping are separate operations, so we can
879 * unmap more blocks from the destination file than we remap.
881 * ----SSSSSSS-SSSSS----SSSSSS
883 * --DDDDD---------DDDDD--DDD
886 * Now remap the source extent into the destination file:
888 * ----SSSSSSS-SSSSS----SSSSSS
890 * --DDDDD--SSSSSSSDDDDD--DDD
893 * Do likewise with the second hole and extent in our range. Holes in the
894 * unmap range don't affect our operation.
896 * ----SSSSSSS-SSSSS----SSSSSS
898 * --DDDDD--SSSSSSS-SSSSS-DDD
901 * Finally, unmap and remap part of the third extent. This will increase the
902 * size of the destination file.
904 * ----SSSSSSS-SSSSS----SSSSSS
906 * --DDDDD--SSSSSSS-SSSSS----SSS
909 * Once we update the destination file's i_size, we're done.
913 * Ensure the reflink bit is set in both inodes.
916 xfs_reflink_set_inode_flag(
917 struct xfs_inode *src,
918 struct xfs_inode *dest)
920 struct xfs_mount *mp = src->i_mount;
922 struct xfs_trans *tp;
924 if (xfs_is_reflink_inode(src) && xfs_is_reflink_inode(dest))
927 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
931 /* Lock both files against IO */
932 if (src->i_ino == dest->i_ino)
933 xfs_ilock(src, XFS_ILOCK_EXCL);
935 xfs_lock_two_inodes(src, dest, XFS_ILOCK_EXCL);
937 if (!xfs_is_reflink_inode(src)) {
938 trace_xfs_reflink_set_inode_flag(src);
939 xfs_trans_ijoin(tp, src, XFS_ILOCK_EXCL);
940 src->i_d.di_flags2 |= XFS_DIFLAG2_REFLINK;
941 xfs_trans_log_inode(tp, src, XFS_ILOG_CORE);
942 xfs_ifork_init_cow(src);
944 xfs_iunlock(src, XFS_ILOCK_EXCL);
946 if (src->i_ino == dest->i_ino)
949 if (!xfs_is_reflink_inode(dest)) {
950 trace_xfs_reflink_set_inode_flag(dest);
951 xfs_trans_ijoin(tp, dest, XFS_ILOCK_EXCL);
952 dest->i_d.di_flags2 |= XFS_DIFLAG2_REFLINK;
953 xfs_trans_log_inode(tp, dest, XFS_ILOG_CORE);
954 xfs_ifork_init_cow(dest);
956 xfs_iunlock(dest, XFS_ILOCK_EXCL);
959 error = xfs_trans_commit(tp);
965 trace_xfs_reflink_set_inode_flag_error(dest, error, _RET_IP_);
970 * Update destination inode size & cowextsize hint, if necessary.
973 xfs_reflink_update_dest(
974 struct xfs_inode *dest,
976 xfs_extlen_t cowextsize,
979 struct xfs_mount *mp = dest->i_mount;
980 struct xfs_trans *tp;
983 if (is_dedupe && newlen <= i_size_read(VFS_I(dest)) && cowextsize == 0)
986 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
990 xfs_ilock(dest, XFS_ILOCK_EXCL);
991 xfs_trans_ijoin(tp, dest, XFS_ILOCK_EXCL);
993 if (newlen > i_size_read(VFS_I(dest))) {
994 trace_xfs_reflink_update_inode_size(dest, newlen);
995 i_size_write(VFS_I(dest), newlen);
996 dest->i_d.di_size = newlen;
1000 dest->i_d.di_cowextsize = cowextsize;
1001 dest->i_d.di_flags2 |= XFS_DIFLAG2_COWEXTSIZE;
1005 xfs_trans_ichgtime(tp, dest,
1006 XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1008 xfs_trans_log_inode(tp, dest, XFS_ILOG_CORE);
1010 error = xfs_trans_commit(tp);
1016 trace_xfs_reflink_update_inode_size_error(dest, error, _RET_IP_);
1021 * Do we have enough reserve in this AG to handle a reflink? The refcount
1022 * btree already reserved all the space it needs, but the rmap btree can grow
1023 * infinitely, so we won't allow more reflinks when the AG is down to the
1027 xfs_reflink_ag_has_free_space(
1028 struct xfs_mount *mp,
1029 xfs_agnumber_t agno)
1031 struct xfs_perag *pag;
1034 if (!xfs_sb_version_hasrmapbt(&mp->m_sb))
1037 pag = xfs_perag_get(mp, agno);
1038 if (xfs_ag_resv_critical(pag, XFS_AG_RESV_AGFL) ||
1039 xfs_ag_resv_critical(pag, XFS_AG_RESV_METADATA))
1046 * Unmap a range of blocks from a file, then map other blocks into the hole.
1047 * The range to unmap is (destoff : destoff + srcioff + irec->br_blockcount).
1048 * The extent irec is mapped into dest at irec->br_startoff.
1051 xfs_reflink_remap_extent(
1052 struct xfs_inode *ip,
1053 struct xfs_bmbt_irec *irec,
1054 xfs_fileoff_t destoff,
1055 xfs_off_t new_isize)
1057 struct xfs_mount *mp = ip->i_mount;
1058 bool real_extent = xfs_bmap_is_real_extent(irec);
1059 struct xfs_trans *tp;
1060 xfs_fsblock_t firstfsb;
1061 unsigned int resblks;
1062 struct xfs_defer_ops dfops;
1063 struct xfs_bmbt_irec uirec;
1065 xfs_filblks_t unmap_len;
1069 unmap_len = irec->br_startoff + irec->br_blockcount - destoff;
1070 trace_xfs_reflink_punch_range(ip, destoff, unmap_len);
1072 /* No reflinking if we're low on space */
1074 error = xfs_reflink_ag_has_free_space(mp,
1075 XFS_FSB_TO_AGNO(mp, irec->br_startblock));
1080 /* Start a rolling transaction to switch the mappings */
1081 resblks = XFS_EXTENTADD_SPACE_RES(ip->i_mount, XFS_DATA_FORK);
1082 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, resblks, 0, 0, &tp);
1086 xfs_ilock(ip, XFS_ILOCK_EXCL);
1087 xfs_trans_ijoin(tp, ip, 0);
1089 /* If we're not just clearing space, then do we have enough quota? */
1091 error = xfs_trans_reserve_quota_nblks(tp, ip,
1092 irec->br_blockcount, 0, XFS_QMOPT_RES_REGBLKS);
1097 trace_xfs_reflink_remap(ip, irec->br_startoff,
1098 irec->br_blockcount, irec->br_startblock);
1100 /* Unmap the old blocks in the data fork. */
1103 xfs_defer_init(&dfops, &firstfsb);
1104 error = __xfs_bunmapi(tp, ip, destoff, &rlen, 0, 1,
1110 * Trim the extent to whatever got unmapped.
1111 * Remember, bunmapi works backwards.
1113 uirec.br_startblock = irec->br_startblock + rlen;
1114 uirec.br_startoff = irec->br_startoff + rlen;
1115 uirec.br_blockcount = unmap_len - rlen;
1118 /* If this isn't a real mapping, we're done. */
1119 if (!real_extent || uirec.br_blockcount == 0)
1122 trace_xfs_reflink_remap(ip, uirec.br_startoff,
1123 uirec.br_blockcount, uirec.br_startblock);
1125 /* Update the refcount tree */
1126 error = xfs_refcount_increase_extent(mp, &dfops, &uirec);
1130 /* Map the new blocks into the data fork. */
1131 error = xfs_bmap_map_extent(mp, &dfops, ip, &uirec);
1135 /* Update quota accounting. */
1136 xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_BCOUNT,
1137 uirec.br_blockcount);
1139 /* Update dest isize if needed. */
1140 newlen = XFS_FSB_TO_B(mp,
1141 uirec.br_startoff + uirec.br_blockcount);
1142 newlen = min_t(xfs_off_t, newlen, new_isize);
1143 if (newlen > i_size_read(VFS_I(ip))) {
1144 trace_xfs_reflink_update_inode_size(ip, newlen);
1145 i_size_write(VFS_I(ip), newlen);
1146 ip->i_d.di_size = newlen;
1147 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1151 /* Process all the deferred stuff. */
1152 error = xfs_defer_finish(&tp, &dfops, ip);
1157 error = xfs_trans_commit(tp);
1158 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1164 xfs_defer_cancel(&dfops);
1166 xfs_trans_cancel(tp);
1167 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1169 trace_xfs_reflink_remap_extent_error(ip, error, _RET_IP_);
1174 * Iteratively remap one file's extents (and holes) to another's.
1177 xfs_reflink_remap_blocks(
1178 struct xfs_inode *src,
1179 xfs_fileoff_t srcoff,
1180 struct xfs_inode *dest,
1181 xfs_fileoff_t destoff,
1183 xfs_off_t new_isize)
1185 struct xfs_bmbt_irec imap;
1188 xfs_filblks_t range_len;
1190 /* drange = (destoff, destoff + len); srange = (srcoff, srcoff + len) */
1192 trace_xfs_reflink_remap_blocks_loop(src, srcoff, len,
1194 /* Read extent from the source file */
1196 xfs_ilock(src, XFS_ILOCK_EXCL);
1197 error = xfs_bmapi_read(src, srcoff, len, &imap, &nimaps, 0);
1198 xfs_iunlock(src, XFS_ILOCK_EXCL);
1201 ASSERT(nimaps == 1);
1203 trace_xfs_reflink_remap_imap(src, srcoff, len, XFS_IO_OVERWRITE,
1206 /* Translate imap into the destination file. */
1207 range_len = imap.br_startoff + imap.br_blockcount - srcoff;
1208 imap.br_startoff += destoff - srcoff;
1210 /* Clear dest from destoff to the end of imap and map it in. */
1211 error = xfs_reflink_remap_extent(dest, &imap, destoff,
1216 if (fatal_signal_pending(current)) {
1221 /* Advance drange/srange */
1222 srcoff += range_len;
1223 destoff += range_len;
1230 trace_xfs_reflink_remap_blocks_error(dest, error, _RET_IP_);
1235 * Link a range of blocks from one file to another.
1238 xfs_reflink_remap_range(
1239 struct file *file_in,
1241 struct file *file_out,
1246 struct inode *inode_in = file_inode(file_in);
1247 struct xfs_inode *src = XFS_I(inode_in);
1248 struct inode *inode_out = file_inode(file_out);
1249 struct xfs_inode *dest = XFS_I(inode_out);
1250 struct xfs_mount *mp = src->i_mount;
1251 bool same_inode = (inode_in == inode_out);
1252 xfs_fileoff_t sfsbno, dfsbno;
1253 xfs_filblks_t fsblen;
1254 xfs_extlen_t cowextsize;
1257 if (!xfs_sb_version_hasreflink(&mp->m_sb))
1260 if (XFS_FORCED_SHUTDOWN(mp))
1263 /* Lock both files against IO */
1264 lock_two_nondirectories(inode_in, inode_out);
1266 xfs_ilock(src, XFS_MMAPLOCK_EXCL);
1268 xfs_lock_two_inodes(src, dest, XFS_MMAPLOCK_EXCL);
1270 /* Check file eligibility and prepare for block sharing. */
1272 /* Don't reflink realtime inodes */
1273 if (XFS_IS_REALTIME_INODE(src) || XFS_IS_REALTIME_INODE(dest))
1276 /* Don't share DAX file data for now. */
1277 if (IS_DAX(inode_in) || IS_DAX(inode_out))
1280 ret = vfs_clone_file_prep_inodes(inode_in, pos_in, inode_out, pos_out,
1285 trace_xfs_reflink_remap_range(src, pos_in, len, dest, pos_out);
1287 /* Set flags and remap blocks. */
1288 ret = xfs_reflink_set_inode_flag(src, dest);
1292 dfsbno = XFS_B_TO_FSBT(mp, pos_out);
1293 sfsbno = XFS_B_TO_FSBT(mp, pos_in);
1294 fsblen = XFS_B_TO_FSB(mp, len);
1295 ret = xfs_reflink_remap_blocks(src, sfsbno, dest, dfsbno, fsblen,
1300 /* Zap any page cache for the destination file's range. */
1301 truncate_inode_pages_range(&inode_out->i_data, pos_out,
1302 PAGE_ALIGN(pos_out + len) - 1);
1305 * Carry the cowextsize hint from src to dest if we're sharing the
1306 * entire source file to the entire destination file, the source file
1307 * has a cowextsize hint, and the destination file does not.
1310 if (pos_in == 0 && len == i_size_read(inode_in) &&
1311 (src->i_d.di_flags2 & XFS_DIFLAG2_COWEXTSIZE) &&
1312 pos_out == 0 && len >= i_size_read(inode_out) &&
1313 !(dest->i_d.di_flags2 & XFS_DIFLAG2_COWEXTSIZE))
1314 cowextsize = src->i_d.di_cowextsize;
1316 ret = xfs_reflink_update_dest(dest, pos_out + len, cowextsize,
1320 xfs_iunlock(src, XFS_MMAPLOCK_EXCL);
1322 xfs_iunlock(dest, XFS_MMAPLOCK_EXCL);
1323 unlock_two_nondirectories(inode_in, inode_out);
1325 trace_xfs_reflink_remap_range_error(dest, ret, _RET_IP_);
1330 * The user wants to preemptively CoW all shared blocks in this file,
1331 * which enables us to turn off the reflink flag. Iterate all
1332 * extents which are not prealloc/delalloc to see which ranges are
1333 * mentioned in the refcount tree, then read those blocks into the
1334 * pagecache, dirty them, fsync them back out, and then we can update
1335 * the inode flag. What happens if we run out of memory? :)
1338 xfs_reflink_dirty_extents(
1339 struct xfs_inode *ip,
1344 struct xfs_mount *mp = ip->i_mount;
1345 xfs_agnumber_t agno;
1346 xfs_agblock_t agbno;
1352 struct xfs_bmbt_irec map[2];
1356 while (end - fbno > 0) {
1359 * Look for extents in the file. Skip holes, delalloc, or
1360 * unwritten extents; they can't be reflinked.
1362 error = xfs_bmapi_read(ip, fbno, end - fbno, map, &nmaps, 0);
1367 if (!xfs_bmap_is_real_extent(&map[0]))
1371 while (map[1].br_blockcount) {
1372 agno = XFS_FSB_TO_AGNO(mp, map[1].br_startblock);
1373 agbno = XFS_FSB_TO_AGBNO(mp, map[1].br_startblock);
1374 aglen = map[1].br_blockcount;
1376 error = xfs_reflink_find_shared(mp, agno, agbno, aglen,
1377 &rbno, &rlen, true);
1380 if (rbno == NULLAGBLOCK)
1383 /* Dirty the pages */
1384 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1385 fpos = XFS_FSB_TO_B(mp, map[1].br_startoff +
1387 flen = XFS_FSB_TO_B(mp, rlen);
1388 if (fpos + flen > isize)
1389 flen = isize - fpos;
1390 error = iomap_file_dirty(VFS_I(ip), fpos, flen,
1392 xfs_ilock(ip, XFS_ILOCK_EXCL);
1396 map[1].br_blockcount -= (rbno - agbno + rlen);
1397 map[1].br_startoff += (rbno - agbno + rlen);
1398 map[1].br_startblock += (rbno - agbno + rlen);
1402 fbno = map[0].br_startoff + map[0].br_blockcount;
1408 /* Clear the inode reflink flag if there are no shared extents. */
1410 xfs_reflink_clear_inode_flag(
1411 struct xfs_inode *ip,
1412 struct xfs_trans **tpp)
1414 struct xfs_mount *mp = ip->i_mount;
1417 xfs_agnumber_t agno;
1418 xfs_agblock_t agbno;
1422 struct xfs_bmbt_irec map;
1426 ASSERT(xfs_is_reflink_inode(ip));
1429 end = XFS_B_TO_FSB(mp, i_size_read(VFS_I(ip)));
1430 while (end - fbno > 0) {
1433 * Look for extents in the file. Skip holes, delalloc, or
1434 * unwritten extents; they can't be reflinked.
1436 error = xfs_bmapi_read(ip, fbno, end - fbno, &map, &nmaps, 0);
1441 if (!xfs_bmap_is_real_extent(&map))
1444 agno = XFS_FSB_TO_AGNO(mp, map.br_startblock);
1445 agbno = XFS_FSB_TO_AGBNO(mp, map.br_startblock);
1446 aglen = map.br_blockcount;
1448 error = xfs_reflink_find_shared(mp, agno, agbno, aglen,
1449 &rbno, &rlen, false);
1452 /* Is there still a shared block here? */
1453 if (rbno != NULLAGBLOCK)
1456 fbno = map.br_startoff + map.br_blockcount;
1460 * We didn't find any shared blocks so turn off the reflink flag.
1461 * First, get rid of any leftover CoW mappings.
1463 error = xfs_reflink_cancel_cow_blocks(ip, tpp, 0, NULLFILEOFF, true);
1467 /* Clear the inode flag. */
1468 trace_xfs_reflink_unset_inode_flag(ip);
1469 ip->i_d.di_flags2 &= ~XFS_DIFLAG2_REFLINK;
1470 xfs_inode_clear_cowblocks_tag(ip);
1471 xfs_trans_ijoin(*tpp, ip, 0);
1472 xfs_trans_log_inode(*tpp, ip, XFS_ILOG_CORE);
1478 * Clear the inode reflink flag if there are no shared extents and the size
1482 xfs_reflink_try_clear_inode_flag(
1483 struct xfs_inode *ip)
1485 struct xfs_mount *mp = ip->i_mount;
1486 struct xfs_trans *tp;
1489 /* Start a rolling transaction to remove the mappings */
1490 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, 0, 0, 0, &tp);
1494 xfs_ilock(ip, XFS_ILOCK_EXCL);
1495 xfs_trans_ijoin(tp, ip, 0);
1497 error = xfs_reflink_clear_inode_flag(ip, &tp);
1501 error = xfs_trans_commit(tp);
1505 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1508 xfs_trans_cancel(tp);
1510 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1515 * Pre-COW all shared blocks within a given byte range of a file and turn off
1516 * the reflink flag if we unshare all of the file's blocks.
1519 xfs_reflink_unshare(
1520 struct xfs_inode *ip,
1524 struct xfs_mount *mp = ip->i_mount;
1530 if (!xfs_is_reflink_inode(ip))
1533 trace_xfs_reflink_unshare(ip, offset, len);
1535 inode_dio_wait(VFS_I(ip));
1537 /* Try to CoW the selected ranges */
1538 xfs_ilock(ip, XFS_ILOCK_EXCL);
1539 fbno = XFS_B_TO_FSBT(mp, offset);
1540 isize = i_size_read(VFS_I(ip));
1541 end = XFS_B_TO_FSB(mp, offset + len);
1542 error = xfs_reflink_dirty_extents(ip, fbno, end, isize);
1545 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1547 /* Wait for the IO to finish */
1548 error = filemap_write_and_wait(VFS_I(ip)->i_mapping);
1552 /* Turn off the reflink flag if possible. */
1553 error = xfs_reflink_try_clear_inode_flag(ip);
1560 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1562 trace_xfs_reflink_unshare_error(ip, error, _RET_IP_);