1 // SPDX-License-Identifier: GPL-2.0+
3 * Copyright (C) 2016 Oracle. All Rights Reserved.
8 #include "xfs_shared.h"
9 #include "xfs_format.h"
10 #include "xfs_log_format.h"
11 #include "xfs_trans_resv.h"
12 #include "xfs_mount.h"
13 #include "xfs_defer.h"
14 #include "xfs_da_format.h"
15 #include "xfs_da_btree.h"
16 #include "xfs_inode.h"
17 #include "xfs_trans.h"
18 #include "xfs_inode_item.h"
20 #include "xfs_bmap_util.h"
21 #include "xfs_error.h"
23 #include "xfs_dir2_priv.h"
24 #include "xfs_ioctl.h"
25 #include "xfs_trace.h"
27 #include "xfs_icache.h"
29 #include "xfs_btree.h"
30 #include "xfs_refcount_btree.h"
31 #include "xfs_refcount.h"
32 #include "xfs_bmap_btree.h"
33 #include "xfs_trans_space.h"
35 #include "xfs_alloc.h"
36 #include "xfs_quota_defs.h"
37 #include "xfs_quota.h"
38 #include "xfs_reflink.h"
39 #include "xfs_iomap.h"
40 #include "xfs_rmap_btree.h"
42 #include "xfs_ag_resv.h"
45 * Copy on Write of Shared Blocks
47 * XFS must preserve "the usual" file semantics even when two files share
48 * the same physical blocks. This means that a write to one file must not
49 * alter the blocks in a different file; the way that we'll do that is
50 * through the use of a copy-on-write mechanism. At a high level, that
51 * means that when we want to write to a shared block, we allocate a new
52 * block, write the data to the new block, and if that succeeds we map the
53 * new block into the file.
55 * XFS provides a "delayed allocation" mechanism that defers the allocation
56 * of disk blocks to dirty-but-not-yet-mapped file blocks as long as
57 * possible. This reduces fragmentation by enabling the filesystem to ask
58 * for bigger chunks less often, which is exactly what we want for CoW.
60 * The delalloc mechanism begins when the kernel wants to make a block
61 * writable (write_begin or page_mkwrite). If the offset is not mapped, we
62 * create a delalloc mapping, which is a regular in-core extent, but without
63 * a real startblock. (For delalloc mappings, the startblock encodes both
64 * a flag that this is a delalloc mapping, and a worst-case estimate of how
65 * many blocks might be required to put the mapping into the BMBT.) delalloc
66 * mappings are a reservation against the free space in the filesystem;
67 * adjacent mappings can also be combined into fewer larger mappings.
69 * As an optimization, the CoW extent size hint (cowextsz) creates
70 * outsized aligned delalloc reservations in the hope of landing out of
71 * order nearby CoW writes in a single extent on disk, thereby reducing
72 * fragmentation and improving future performance.
74 * D: --RRRRRRSSSRRRRRRRR--- (data fork)
75 * C: ------DDDDDDD--------- (CoW fork)
77 * When dirty pages are being written out (typically in writepage), the
78 * delalloc reservations are converted into unwritten mappings by
79 * allocating blocks and replacing the delalloc mapping with real ones.
80 * A delalloc mapping can be replaced by several unwritten ones if the
81 * free space is fragmented.
83 * D: --RRRRRRSSSRRRRRRRR---
84 * C: ------UUUUUUU---------
86 * We want to adapt the delalloc mechanism for copy-on-write, since the
87 * write paths are similar. The first two steps (creating the reservation
88 * and allocating the blocks) are exactly the same as delalloc except that
89 * the mappings must be stored in a separate CoW fork because we do not want
90 * to disturb the mapping in the data fork until we're sure that the write
91 * succeeded. IO completion in this case is the process of removing the old
92 * mapping from the data fork and moving the new mapping from the CoW fork to
93 * the data fork. This will be discussed shortly.
95 * For now, unaligned directio writes will be bounced back to the page cache.
96 * Block-aligned directio writes will use the same mechanism as buffered
99 * Just prior to submitting the actual disk write requests, we convert
100 * the extents representing the range of the file actually being written
101 * (as opposed to extra pieces created for the cowextsize hint) to real
102 * extents. This will become important in the next step:
104 * D: --RRRRRRSSSRRRRRRRR---
105 * C: ------UUrrUUU---------
107 * CoW remapping must be done after the data block write completes,
108 * because we don't want to destroy the old data fork map until we're sure
109 * the new block has been written. Since the new mappings are kept in a
110 * separate fork, we can simply iterate these mappings to find the ones
111 * that cover the file blocks that we just CoW'd. For each extent, simply
112 * unmap the corresponding range in the data fork, map the new range into
113 * the data fork, and remove the extent from the CoW fork. Because of
114 * the presence of the cowextsize hint, however, we must be careful
115 * only to remap the blocks that we've actually written out -- we must
116 * never remap delalloc reservations nor CoW staging blocks that have
117 * yet to be written. This corresponds exactly to the real extents in
120 * D: --RRRRRRrrSRRRRRRRR---
121 * C: ------UU--UUU---------
123 * Since the remapping operation can be applied to an arbitrary file
124 * range, we record the need for the remap step as a flag in the ioend
125 * instead of declaring a new IO type. This is required for direct io
126 * because we only have ioend for the whole dio, and we have to be able to
127 * remember the presence of unwritten blocks and CoW blocks with a single
128 * ioend structure. Better yet, the more ground we can cover with one
133 * Given an AG extent, find the lowest-numbered run of shared blocks
134 * within that range and return the range in fbno/flen. If
135 * find_end_of_shared is true, return the longest contiguous extent of
136 * shared blocks. If there are no shared extents, fbno and flen will
137 * be set to NULLAGBLOCK and 0, respectively.
140 xfs_reflink_find_shared(
141 struct xfs_mount *mp,
142 struct xfs_trans *tp,
148 bool find_end_of_shared)
150 struct xfs_buf *agbp;
151 struct xfs_btree_cur *cur;
154 error = xfs_alloc_read_agf(mp, tp, agno, 0, &agbp);
160 cur = xfs_refcountbt_init_cursor(mp, tp, agbp, agno);
162 error = xfs_refcount_find_shared(cur, agbno, aglen, fbno, flen,
165 xfs_btree_del_cursor(cur, error);
167 xfs_trans_brelse(tp, agbp);
172 * Trim the mapping to the next block where there's a change in the
173 * shared/unshared status. More specifically, this means that we
174 * find the lowest-numbered extent of shared blocks that coincides with
175 * the given block mapping. If the shared extent overlaps the start of
176 * the mapping, trim the mapping to the end of the shared extent. If
177 * the shared region intersects the mapping, trim the mapping to the
178 * start of the shared extent. If there are no shared regions that
179 * overlap, just return the original extent.
182 xfs_reflink_trim_around_shared(
183 struct xfs_inode *ip,
184 struct xfs_bmbt_irec *irec,
194 /* Holes, unwritten, and delalloc extents cannot be shared */
195 if (!xfs_is_reflink_inode(ip) || !xfs_bmap_is_real_extent(irec)) {
200 trace_xfs_reflink_trim_around_shared(ip, irec);
202 agno = XFS_FSB_TO_AGNO(ip->i_mount, irec->br_startblock);
203 agbno = XFS_FSB_TO_AGBNO(ip->i_mount, irec->br_startblock);
204 aglen = irec->br_blockcount;
206 error = xfs_reflink_find_shared(ip->i_mount, NULL, agno, agbno,
207 aglen, &fbno, &flen, true);
212 if (fbno == NULLAGBLOCK) {
213 /* No shared blocks at all. */
215 } else if (fbno == agbno) {
217 * The start of this extent is shared. Truncate the
218 * mapping at the end of the shared region so that a
219 * subsequent iteration starts at the start of the
222 irec->br_blockcount = flen;
227 * There's a shared extent midway through this extent.
228 * Truncate the mapping at the start of the shared
229 * extent so that a subsequent iteration starts at the
230 * start of the shared region.
232 irec->br_blockcount = fbno - agbno;
238 * Trim the passed in imap to the next shared/unshared extent boundary, and
239 * if imap->br_startoff points to a shared extent reserve space for it in the
242 * Note that imap will always contain the block numbers for the existing blocks
243 * in the data fork, as the upper layers need them for read-modify-write
247 xfs_reflink_reserve_cow(
248 struct xfs_inode *ip,
249 struct xfs_bmbt_irec *imap)
251 struct xfs_ifork *ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
252 struct xfs_bmbt_irec got;
255 struct xfs_iext_cursor icur;
259 * Search the COW fork extent list first. This serves two purposes:
260 * first this implement the speculative preallocation using cowextisze,
261 * so that we also unshared block adjacent to shared blocks instead
262 * of just the shared blocks themselves. Second the lookup in the
263 * extent list is generally faster than going out to the shared extent
267 if (!xfs_iext_lookup_extent(ip, ifp, imap->br_startoff, &icur, &got))
269 if (!eof && got.br_startoff <= imap->br_startoff) {
270 trace_xfs_reflink_cow_found(ip, imap);
271 xfs_trim_extent(imap, got.br_startoff, got.br_blockcount);
275 /* Trim the mapping to the nearest shared extent boundary. */
276 error = xfs_reflink_trim_around_shared(ip, imap, &shared);
280 /* Not shared? Just report the (potentially capped) extent. */
285 * Fork all the shared blocks from our write offset until the end of
288 error = xfs_qm_dqattach_locked(ip, false);
292 error = xfs_bmapi_reserve_delalloc(ip, XFS_COW_FORK, imap->br_startoff,
293 imap->br_blockcount, 0, &got, &icur, eof);
294 if (error == -ENOSPC || error == -EDQUOT)
295 trace_xfs_reflink_cow_enospc(ip, imap);
299 trace_xfs_reflink_cow_alloc(ip, &got);
303 /* Convert part of an unwritten CoW extent to a real one. */
305 xfs_reflink_convert_cow_extent(
306 struct xfs_inode *ip,
307 struct xfs_bmbt_irec *imap,
308 xfs_fileoff_t offset_fsb,
309 xfs_filblks_t count_fsb)
313 if (imap->br_state == XFS_EXT_NORM)
316 xfs_trim_extent(imap, offset_fsb, count_fsb);
317 trace_xfs_reflink_convert_cow(ip, imap);
318 if (imap->br_blockcount == 0)
320 return xfs_bmapi_write(NULL, ip, imap->br_startoff, imap->br_blockcount,
321 XFS_BMAPI_COWFORK | XFS_BMAPI_CONVERT, 0, imap,
325 /* Convert all of the unwritten CoW extents in a file's range to real ones. */
327 xfs_reflink_convert_cow(
328 struct xfs_inode *ip,
332 struct xfs_mount *mp = ip->i_mount;
333 xfs_fileoff_t offset_fsb = XFS_B_TO_FSBT(mp, offset);
334 xfs_fileoff_t end_fsb = XFS_B_TO_FSB(mp, offset + count);
335 xfs_filblks_t count_fsb = end_fsb - offset_fsb;
336 struct xfs_bmbt_irec imap;
337 int nimaps = 1, error = 0;
341 xfs_ilock(ip, XFS_ILOCK_EXCL);
342 error = xfs_bmapi_write(NULL, ip, offset_fsb, count_fsb,
343 XFS_BMAPI_COWFORK | XFS_BMAPI_CONVERT |
344 XFS_BMAPI_CONVERT_ONLY, 0, &imap, &nimaps);
345 xfs_iunlock(ip, XFS_ILOCK_EXCL);
350 * Find the extent that maps the given range in the COW fork. Even if the extent
351 * is not shared we might have a preallocation for it in the COW fork. If so we
352 * use it that rather than trigger a new allocation.
355 xfs_find_trim_cow_extent(
356 struct xfs_inode *ip,
357 struct xfs_bmbt_irec *imap,
361 xfs_fileoff_t offset_fsb = imap->br_startoff;
362 xfs_filblks_t count_fsb = imap->br_blockcount;
363 struct xfs_iext_cursor icur;
364 struct xfs_bmbt_irec got;
369 * If we don't find an overlapping extent, trim the range we need to
370 * allocate to fit the hole we found.
372 if (!xfs_iext_lookup_extent(ip, ip->i_cowfp, offset_fsb, &icur, &got))
373 got.br_startoff = offset_fsb + count_fsb;
374 if (got.br_startoff > offset_fsb) {
375 xfs_trim_extent(imap, imap->br_startoff,
376 got.br_startoff - imap->br_startoff);
377 return xfs_reflink_trim_around_shared(ip, imap, shared);
381 if (isnullstartblock(got.br_startblock)) {
382 xfs_trim_extent(imap, got.br_startoff, got.br_blockcount);
386 /* real extent found - no need to allocate */
387 xfs_trim_extent(&got, offset_fsb, count_fsb);
393 /* Allocate all CoW reservations covering a range of blocks in a file. */
395 xfs_reflink_allocate_cow(
396 struct xfs_inode *ip,
397 struct xfs_bmbt_irec *imap,
401 struct xfs_mount *mp = ip->i_mount;
402 xfs_fileoff_t offset_fsb = imap->br_startoff;
403 xfs_filblks_t count_fsb = imap->br_blockcount;
404 struct xfs_trans *tp;
405 int nimaps, error = 0;
407 xfs_filblks_t resaligned;
408 xfs_extlen_t resblks = 0;
410 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
411 ASSERT(xfs_is_reflink_inode(ip));
413 error = xfs_find_trim_cow_extent(ip, imap, shared, &found);
414 if (error || !*shared)
419 resaligned = xfs_aligned_fsb_count(imap->br_startoff,
420 imap->br_blockcount, xfs_get_cowextsz_hint(ip));
421 resblks = XFS_DIOSTRAT_SPACE_RES(mp, resaligned);
423 xfs_iunlock(ip, *lockmode);
424 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, resblks, 0, 0, &tp);
425 *lockmode = XFS_ILOCK_EXCL;
426 xfs_ilock(ip, *lockmode);
431 error = xfs_qm_dqattach_locked(ip, false);
433 goto out_trans_cancel;
436 * Check for an overlapping extent again now that we dropped the ilock.
438 error = xfs_find_trim_cow_extent(ip, imap, shared, &found);
439 if (error || !*shared)
440 goto out_trans_cancel;
442 xfs_trans_cancel(tp);
446 error = xfs_trans_reserve_quota_nblks(tp, ip, resblks, 0,
447 XFS_QMOPT_RES_REGBLKS);
449 goto out_trans_cancel;
451 xfs_trans_ijoin(tp, ip, 0);
453 /* Allocate the entire reservation as unwritten blocks. */
455 error = xfs_bmapi_write(tp, ip, imap->br_startoff, imap->br_blockcount,
456 XFS_BMAPI_COWFORK | XFS_BMAPI_PREALLOC,
457 resblks, imap, &nimaps);
461 xfs_inode_set_cowblocks_tag(ip);
462 error = xfs_trans_commit(tp);
467 * Allocation succeeded but the requested range was not even partially
468 * satisfied? Bail out!
473 return xfs_reflink_convert_cow_extent(ip, imap, offset_fsb, count_fsb);
476 xfs_trans_unreserve_quota_nblks(tp, ip, (long)resblks, 0,
477 XFS_QMOPT_RES_REGBLKS);
479 xfs_trans_cancel(tp);
484 * Cancel CoW reservations for some block range of an inode.
486 * If cancel_real is true this function cancels all COW fork extents for the
487 * inode; if cancel_real is false, real extents are not cleared.
489 * Caller must have already joined the inode to the current transaction. The
490 * inode will be joined to the transaction returned to the caller.
493 xfs_reflink_cancel_cow_blocks(
494 struct xfs_inode *ip,
495 struct xfs_trans **tpp,
496 xfs_fileoff_t offset_fsb,
497 xfs_fileoff_t end_fsb,
500 struct xfs_ifork *ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
501 struct xfs_bmbt_irec got, del;
502 struct xfs_iext_cursor icur;
505 if (!xfs_inode_has_cow_data(ip))
507 if (!xfs_iext_lookup_extent_before(ip, ifp, &end_fsb, &icur, &got))
510 /* Walk backwards until we're out of the I/O range... */
511 while (got.br_startoff + got.br_blockcount > offset_fsb) {
513 xfs_trim_extent(&del, offset_fsb, end_fsb - offset_fsb);
515 /* Extent delete may have bumped ext forward */
516 if (!del.br_blockcount) {
517 xfs_iext_prev(ifp, &icur);
521 trace_xfs_reflink_cancel_cow(ip, &del);
523 if (isnullstartblock(del.br_startblock)) {
524 error = xfs_bmap_del_extent_delay(ip, XFS_COW_FORK,
528 } else if (del.br_state == XFS_EXT_UNWRITTEN || cancel_real) {
529 ASSERT((*tpp)->t_firstblock == NULLFSBLOCK);
531 /* Free the CoW orphan record. */
532 error = xfs_refcount_free_cow_extent(*tpp,
533 del.br_startblock, del.br_blockcount);
537 xfs_bmap_add_free(*tpp, del.br_startblock,
538 del.br_blockcount, NULL);
540 /* Roll the transaction */
541 error = xfs_defer_finish(tpp);
545 /* Remove the mapping from the CoW fork. */
546 xfs_bmap_del_extent_cow(ip, &icur, &got, &del);
548 /* Remove the quota reservation */
549 error = xfs_trans_reserve_quota_nblks(NULL, ip,
550 -(long)del.br_blockcount, 0,
551 XFS_QMOPT_RES_REGBLKS);
555 /* Didn't do anything, push cursor back. */
556 xfs_iext_prev(ifp, &icur);
559 if (!xfs_iext_get_extent(ifp, &icur, &got))
563 /* clear tag if cow fork is emptied */
565 xfs_inode_clear_cowblocks_tag(ip);
570 * Cancel CoW reservations for some byte range of an inode.
572 * If cancel_real is true this function cancels all COW fork extents for the
573 * inode; if cancel_real is false, real extents are not cleared.
576 xfs_reflink_cancel_cow_range(
577 struct xfs_inode *ip,
582 struct xfs_trans *tp;
583 xfs_fileoff_t offset_fsb;
584 xfs_fileoff_t end_fsb;
587 trace_xfs_reflink_cancel_cow_range(ip, offset, count);
588 ASSERT(xfs_is_reflink_inode(ip));
590 offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
591 if (count == NULLFILEOFF)
592 end_fsb = NULLFILEOFF;
594 end_fsb = XFS_B_TO_FSB(ip->i_mount, offset + count);
596 /* Start a rolling transaction to remove the mappings */
597 error = xfs_trans_alloc(ip->i_mount, &M_RES(ip->i_mount)->tr_write,
598 0, 0, XFS_TRANS_NOFS, &tp);
602 xfs_ilock(ip, XFS_ILOCK_EXCL);
603 xfs_trans_ijoin(tp, ip, 0);
605 /* Scrape out the old CoW reservations */
606 error = xfs_reflink_cancel_cow_blocks(ip, &tp, offset_fsb, end_fsb,
611 error = xfs_trans_commit(tp);
613 xfs_iunlock(ip, XFS_ILOCK_EXCL);
617 xfs_trans_cancel(tp);
618 xfs_iunlock(ip, XFS_ILOCK_EXCL);
620 trace_xfs_reflink_cancel_cow_range_error(ip, error, _RET_IP_);
625 * Remap parts of a file's data fork after a successful CoW.
629 struct xfs_inode *ip,
633 struct xfs_ifork *ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
634 struct xfs_bmbt_irec got, del;
635 struct xfs_trans *tp;
636 xfs_fileoff_t offset_fsb;
637 xfs_fileoff_t end_fsb;
639 unsigned int resblks;
641 struct xfs_iext_cursor icur;
643 trace_xfs_reflink_end_cow(ip, offset, count);
645 /* No COW extents? That's easy! */
646 if (ifp->if_bytes == 0)
649 offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
650 end_fsb = XFS_B_TO_FSB(ip->i_mount, offset + count);
653 * Start a rolling transaction to switch the mappings. We're
654 * unlikely ever to have to remap 16T worth of single-block
655 * extents, so just cap the worst case extent count to 2^32-1.
656 * Stick a warning in just in case, and avoid 64-bit division.
658 BUILD_BUG_ON(MAX_RW_COUNT > UINT_MAX);
659 if (end_fsb - offset_fsb > UINT_MAX) {
660 error = -EFSCORRUPTED;
661 xfs_force_shutdown(ip->i_mount, SHUTDOWN_CORRUPT_INCORE);
665 resblks = XFS_NEXTENTADD_SPACE_RES(ip->i_mount,
666 (unsigned int)(end_fsb - offset_fsb),
668 error = xfs_trans_alloc(ip->i_mount, &M_RES(ip->i_mount)->tr_write,
669 resblks, 0, XFS_TRANS_RESERVE | XFS_TRANS_NOFS, &tp);
673 xfs_ilock(ip, XFS_ILOCK_EXCL);
674 xfs_trans_ijoin(tp, ip, 0);
677 * In case of racing, overlapping AIO writes no COW extents might be
678 * left by the time I/O completes for the loser of the race. In that
681 if (!xfs_iext_lookup_extent_before(ip, ifp, &end_fsb, &icur, &got))
684 /* Walk backwards until we're out of the I/O range... */
685 while (got.br_startoff + got.br_blockcount > offset_fsb) {
687 xfs_trim_extent(&del, offset_fsb, end_fsb - offset_fsb);
689 /* Extent delete may have bumped ext forward */
690 if (!del.br_blockcount)
694 * Only remap real extent that contain data. With AIO
695 * speculatively preallocations can leak into the range we
696 * are called upon, and we need to skip them.
698 if (!xfs_bmap_is_real_extent(&got))
701 /* Unmap the old blocks in the data fork. */
702 ASSERT(tp->t_firstblock == NULLFSBLOCK);
703 rlen = del.br_blockcount;
704 error = __xfs_bunmapi(tp, ip, del.br_startoff, &rlen, 0, 1);
708 /* Trim the extent to whatever got unmapped. */
710 xfs_trim_extent(&del, del.br_startoff + rlen,
711 del.br_blockcount - rlen);
713 trace_xfs_reflink_cow_remap(ip, &del);
715 /* Free the CoW orphan record. */
716 error = xfs_refcount_free_cow_extent(tp, del.br_startblock,
721 /* Map the new blocks into the data fork. */
722 error = xfs_bmap_map_extent(tp, ip, &del);
726 /* Charge this new data fork mapping to the on-disk quota. */
727 xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_DELBCOUNT,
728 (long)del.br_blockcount);
730 /* Remove the mapping from the CoW fork. */
731 xfs_bmap_del_extent_cow(ip, &icur, &got, &del);
733 error = xfs_defer_finish(&tp);
736 if (!xfs_iext_get_extent(ifp, &icur, &got))
740 if (!xfs_iext_prev_extent(ifp, &icur, &got))
744 error = xfs_trans_commit(tp);
745 xfs_iunlock(ip, XFS_ILOCK_EXCL);
751 xfs_trans_cancel(tp);
752 xfs_iunlock(ip, XFS_ILOCK_EXCL);
754 trace_xfs_reflink_end_cow_error(ip, error, _RET_IP_);
759 * Free leftover CoW reservations that didn't get cleaned out.
762 xfs_reflink_recover_cow(
763 struct xfs_mount *mp)
768 if (!xfs_sb_version_hasreflink(&mp->m_sb))
771 for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
772 error = xfs_refcount_recover_cow_leftovers(mp, agno);
781 * Reflinking (Block) Ranges of Two Files Together
783 * First, ensure that the reflink flag is set on both inodes. The flag is an
784 * optimization to avoid unnecessary refcount btree lookups in the write path.
786 * Now we can iteratively remap the range of extents (and holes) in src to the
787 * corresponding ranges in dest. Let drange and srange denote the ranges of
788 * logical blocks in dest and src touched by the reflink operation.
790 * While the length of drange is greater than zero,
791 * - Read src's bmbt at the start of srange ("imap")
792 * - If imap doesn't exist, make imap appear to start at the end of srange
794 * - If imap starts before srange, advance imap to start at srange.
795 * - If imap goes beyond srange, truncate imap to end at the end of srange.
796 * - Punch (imap start - srange start + imap len) blocks from dest at
797 * offset (drange start).
798 * - If imap points to a real range of pblks,
799 * > Increase the refcount of the imap's pblks
800 * > Map imap's pblks into dest at the offset
801 * (drange start + imap start - srange start)
802 * - Advance drange and srange by (imap start - srange start + imap len)
804 * Finally, if the reflink made dest longer, update both the in-core and
805 * on-disk file sizes.
807 * ASCII Art Demonstration:
809 * Let's say we want to reflink this source file:
811 * ----SSSSSSS-SSSSS----SSSSSS (src file)
812 * <-------------------->
814 * into this destination file:
816 * --DDDDDDDDDDDDDDDDDDD--DDD (dest file)
817 * <-------------------->
818 * '-' means a hole, and 'S' and 'D' are written blocks in the src and dest.
819 * Observe that the range has different logical offsets in either file.
821 * Consider that the first extent in the source file doesn't line up with our
822 * reflink range. Unmapping and remapping are separate operations, so we can
823 * unmap more blocks from the destination file than we remap.
825 * ----SSSSSSS-SSSSS----SSSSSS
827 * --DDDDD---------DDDDD--DDD
830 * Now remap the source extent into the destination file:
832 * ----SSSSSSS-SSSSS----SSSSSS
834 * --DDDDD--SSSSSSSDDDDD--DDD
837 * Do likewise with the second hole and extent in our range. Holes in the
838 * unmap range don't affect our operation.
840 * ----SSSSSSS-SSSSS----SSSSSS
842 * --DDDDD--SSSSSSS-SSSSS-DDD
845 * Finally, unmap and remap part of the third extent. This will increase the
846 * size of the destination file.
848 * ----SSSSSSS-SSSSS----SSSSSS
850 * --DDDDD--SSSSSSS-SSSSS----SSS
853 * Once we update the destination file's i_size, we're done.
857 * Ensure the reflink bit is set in both inodes.
860 xfs_reflink_set_inode_flag(
861 struct xfs_inode *src,
862 struct xfs_inode *dest)
864 struct xfs_mount *mp = src->i_mount;
866 struct xfs_trans *tp;
868 if (xfs_is_reflink_inode(src) && xfs_is_reflink_inode(dest))
871 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
875 /* Lock both files against IO */
876 if (src->i_ino == dest->i_ino)
877 xfs_ilock(src, XFS_ILOCK_EXCL);
879 xfs_lock_two_inodes(src, XFS_ILOCK_EXCL, dest, XFS_ILOCK_EXCL);
881 if (!xfs_is_reflink_inode(src)) {
882 trace_xfs_reflink_set_inode_flag(src);
883 xfs_trans_ijoin(tp, src, XFS_ILOCK_EXCL);
884 src->i_d.di_flags2 |= XFS_DIFLAG2_REFLINK;
885 xfs_trans_log_inode(tp, src, XFS_ILOG_CORE);
886 xfs_ifork_init_cow(src);
888 xfs_iunlock(src, XFS_ILOCK_EXCL);
890 if (src->i_ino == dest->i_ino)
893 if (!xfs_is_reflink_inode(dest)) {
894 trace_xfs_reflink_set_inode_flag(dest);
895 xfs_trans_ijoin(tp, dest, XFS_ILOCK_EXCL);
896 dest->i_d.di_flags2 |= XFS_DIFLAG2_REFLINK;
897 xfs_trans_log_inode(tp, dest, XFS_ILOG_CORE);
898 xfs_ifork_init_cow(dest);
900 xfs_iunlock(dest, XFS_ILOCK_EXCL);
903 error = xfs_trans_commit(tp);
909 trace_xfs_reflink_set_inode_flag_error(dest, error, _RET_IP_);
914 * Update destination inode size & cowextsize hint, if necessary.
917 xfs_reflink_update_dest(
918 struct xfs_inode *dest,
920 xfs_extlen_t cowextsize,
921 unsigned int remap_flags)
923 struct xfs_mount *mp = dest->i_mount;
924 struct xfs_trans *tp;
927 if (newlen <= i_size_read(VFS_I(dest)) && cowextsize == 0)
930 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
934 xfs_ilock(dest, XFS_ILOCK_EXCL);
935 xfs_trans_ijoin(tp, dest, XFS_ILOCK_EXCL);
937 if (newlen > i_size_read(VFS_I(dest))) {
938 trace_xfs_reflink_update_inode_size(dest, newlen);
939 i_size_write(VFS_I(dest), newlen);
940 dest->i_d.di_size = newlen;
944 dest->i_d.di_cowextsize = cowextsize;
945 dest->i_d.di_flags2 |= XFS_DIFLAG2_COWEXTSIZE;
948 xfs_trans_log_inode(tp, dest, XFS_ILOG_CORE);
950 error = xfs_trans_commit(tp);
956 trace_xfs_reflink_update_inode_size_error(dest, error, _RET_IP_);
961 * Do we have enough reserve in this AG to handle a reflink? The refcount
962 * btree already reserved all the space it needs, but the rmap btree can grow
963 * infinitely, so we won't allow more reflinks when the AG is down to the
967 xfs_reflink_ag_has_free_space(
968 struct xfs_mount *mp,
971 struct xfs_perag *pag;
974 if (!xfs_sb_version_hasrmapbt(&mp->m_sb))
977 pag = xfs_perag_get(mp, agno);
978 if (xfs_ag_resv_critical(pag, XFS_AG_RESV_RMAPBT) ||
979 xfs_ag_resv_critical(pag, XFS_AG_RESV_METADATA))
986 * Unmap a range of blocks from a file, then map other blocks into the hole.
987 * The range to unmap is (destoff : destoff + srcioff + irec->br_blockcount).
988 * The extent irec is mapped into dest at irec->br_startoff.
991 xfs_reflink_remap_extent(
992 struct xfs_inode *ip,
993 struct xfs_bmbt_irec *irec,
994 xfs_fileoff_t destoff,
997 struct xfs_mount *mp = ip->i_mount;
998 bool real_extent = xfs_bmap_is_real_extent(irec);
999 struct xfs_trans *tp;
1000 unsigned int resblks;
1001 struct xfs_bmbt_irec uirec;
1003 xfs_filblks_t unmap_len;
1007 unmap_len = irec->br_startoff + irec->br_blockcount - destoff;
1008 trace_xfs_reflink_punch_range(ip, destoff, unmap_len);
1010 /* No reflinking if we're low on space */
1012 error = xfs_reflink_ag_has_free_space(mp,
1013 XFS_FSB_TO_AGNO(mp, irec->br_startblock));
1018 /* Start a rolling transaction to switch the mappings */
1019 resblks = XFS_EXTENTADD_SPACE_RES(ip->i_mount, XFS_DATA_FORK);
1020 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, resblks, 0, 0, &tp);
1024 xfs_ilock(ip, XFS_ILOCK_EXCL);
1025 xfs_trans_ijoin(tp, ip, 0);
1027 /* If we're not just clearing space, then do we have enough quota? */
1029 error = xfs_trans_reserve_quota_nblks(tp, ip,
1030 irec->br_blockcount, 0, XFS_QMOPT_RES_REGBLKS);
1035 trace_xfs_reflink_remap(ip, irec->br_startoff,
1036 irec->br_blockcount, irec->br_startblock);
1038 /* Unmap the old blocks in the data fork. */
1041 ASSERT(tp->t_firstblock == NULLFSBLOCK);
1042 error = __xfs_bunmapi(tp, ip, destoff, &rlen, 0, 1);
1047 * Trim the extent to whatever got unmapped.
1048 * Remember, bunmapi works backwards.
1050 uirec.br_startblock = irec->br_startblock + rlen;
1051 uirec.br_startoff = irec->br_startoff + rlen;
1052 uirec.br_blockcount = unmap_len - rlen;
1055 /* If this isn't a real mapping, we're done. */
1056 if (!real_extent || uirec.br_blockcount == 0)
1059 trace_xfs_reflink_remap(ip, uirec.br_startoff,
1060 uirec.br_blockcount, uirec.br_startblock);
1062 /* Update the refcount tree */
1063 error = xfs_refcount_increase_extent(tp, &uirec);
1067 /* Map the new blocks into the data fork. */
1068 error = xfs_bmap_map_extent(tp, ip, &uirec);
1072 /* Update quota accounting. */
1073 xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_BCOUNT,
1074 uirec.br_blockcount);
1076 /* Update dest isize if needed. */
1077 newlen = XFS_FSB_TO_B(mp,
1078 uirec.br_startoff + uirec.br_blockcount);
1079 newlen = min_t(xfs_off_t, newlen, new_isize);
1080 if (newlen > i_size_read(VFS_I(ip))) {
1081 trace_xfs_reflink_update_inode_size(ip, newlen);
1082 i_size_write(VFS_I(ip), newlen);
1083 ip->i_d.di_size = newlen;
1084 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1088 /* Process all the deferred stuff. */
1089 error = xfs_defer_finish(&tp);
1094 error = xfs_trans_commit(tp);
1095 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1101 xfs_trans_cancel(tp);
1102 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1104 trace_xfs_reflink_remap_extent_error(ip, error, _RET_IP_);
1109 * Iteratively remap one file's extents (and holes) to another's.
1112 xfs_reflink_remap_blocks(
1113 struct xfs_inode *src,
1115 struct xfs_inode *dest,
1120 struct xfs_bmbt_irec imap;
1121 xfs_fileoff_t srcoff;
1122 xfs_fileoff_t destoff;
1124 xfs_filblks_t range_len;
1125 xfs_filblks_t remapped_len = 0;
1126 xfs_off_t new_isize = pos_out + remap_len;
1130 destoff = XFS_B_TO_FSBT(src->i_mount, pos_out);
1131 srcoff = XFS_B_TO_FSBT(src->i_mount, pos_in);
1132 len = XFS_B_TO_FSB(src->i_mount, remap_len);
1134 /* drange = (destoff, destoff + len); srange = (srcoff, srcoff + len) */
1138 trace_xfs_reflink_remap_blocks_loop(src, srcoff, len,
1141 /* Read extent from the source file */
1143 lock_mode = xfs_ilock_data_map_shared(src);
1144 error = xfs_bmapi_read(src, srcoff, len, &imap, &nimaps, 0);
1145 xfs_iunlock(src, lock_mode);
1148 ASSERT(nimaps == 1);
1150 trace_xfs_reflink_remap_imap(src, srcoff, len, XFS_IO_OVERWRITE,
1153 /* Translate imap into the destination file. */
1154 range_len = imap.br_startoff + imap.br_blockcount - srcoff;
1155 imap.br_startoff += destoff - srcoff;
1157 /* Clear dest from destoff to the end of imap and map it in. */
1158 error = xfs_reflink_remap_extent(dest, &imap, destoff,
1163 if (fatal_signal_pending(current)) {
1168 /* Advance drange/srange */
1169 srcoff += range_len;
1170 destoff += range_len;
1172 remapped_len += range_len;
1176 trace_xfs_reflink_remap_blocks_error(dest, error, _RET_IP_);
1177 *remapped = min_t(loff_t, remap_len,
1178 XFS_FSB_TO_B(src->i_mount, remapped_len));
1183 * Grab the exclusive iolock for a data copy from src to dest, making
1184 * sure to abide vfs locking order (lowest pointer value goes first) and
1185 * breaking the pnfs layout leases on dest before proceeding. The loop
1186 * is needed because we cannot call the blocking break_layout() with the
1187 * src iolock held, and therefore have to back out both locks.
1190 xfs_iolock_two_inodes_and_break_layout(
1198 inode_lock_shared(src);
1199 inode_lock_nested(dest, I_MUTEX_NONDIR2);
1205 error = break_layout(dest, false);
1206 if (error == -EWOULDBLOCK) {
1209 inode_unlock_shared(src);
1210 error = break_layout(dest, true);
1218 inode_unlock_shared(src);
1222 inode_lock_shared_nested(src, I_MUTEX_NONDIR2);
1226 /* Unlock both inodes after they've been prepped for a range clone. */
1228 xfs_reflink_remap_unlock(
1229 struct file *file_in,
1230 struct file *file_out)
1232 struct inode *inode_in = file_inode(file_in);
1233 struct xfs_inode *src = XFS_I(inode_in);
1234 struct inode *inode_out = file_inode(file_out);
1235 struct xfs_inode *dest = XFS_I(inode_out);
1236 bool same_inode = (inode_in == inode_out);
1238 xfs_iunlock(dest, XFS_MMAPLOCK_EXCL);
1240 xfs_iunlock(src, XFS_MMAPLOCK_SHARED);
1241 inode_unlock(inode_out);
1243 inode_unlock_shared(inode_in);
1247 * If we're reflinking to a point past the destination file's EOF, we must
1248 * zero any speculative post-EOF preallocations that sit between the old EOF
1249 * and the destination file offset.
1252 xfs_reflink_zero_posteof(
1253 struct xfs_inode *ip,
1256 loff_t isize = i_size_read(VFS_I(ip));
1261 trace_xfs_zero_eof(ip, isize, pos - isize);
1262 return iomap_zero_range(VFS_I(ip), isize, pos - isize, NULL,
1267 * Prepare two files for range cloning. Upon a successful return both inodes
1268 * will have the iolock and mmaplock held, the page cache of the out file will
1269 * be truncated, and any leases on the out file will have been broken. This
1270 * function borrows heavily from xfs_file_aio_write_checks.
1272 * The VFS allows partial EOF blocks to "match" for dedupe even though it hasn't
1273 * checked that the bytes beyond EOF physically match. Hence we cannot use the
1274 * EOF block in the source dedupe range because it's not a complete block match,
1275 * hence can introduce a corruption into the file that has it's block replaced.
1277 * In similar fashion, the VFS file cloning also allows partial EOF blocks to be
1278 * "block aligned" for the purposes of cloning entire files. However, if the
1279 * source file range includes the EOF block and it lands within the existing EOF
1280 * of the destination file, then we can expose stale data from beyond the source
1281 * file EOF in the destination file.
1283 * XFS doesn't support partial block sharing, so in both cases we have check
1284 * these cases ourselves. For dedupe, we can simply round the length to dedupe
1285 * down to the previous whole block and ignore the partial EOF block. While this
1286 * means we can't dedupe the last block of a file, this is an acceptible
1287 * tradeoff for simplicity on implementation.
1289 * For cloning, we want to share the partial EOF block if it is also the new EOF
1290 * block of the destination file. If the partial EOF block lies inside the
1291 * existing destination EOF, then we have to abort the clone to avoid exposing
1292 * stale data in the destination file. Hence we reject these clone attempts with
1293 * -EINVAL in this case.
1296 xfs_reflink_remap_prep(
1297 struct file *file_in,
1299 struct file *file_out,
1302 unsigned int remap_flags)
1304 struct inode *inode_in = file_inode(file_in);
1305 struct xfs_inode *src = XFS_I(inode_in);
1306 struct inode *inode_out = file_inode(file_out);
1307 struct xfs_inode *dest = XFS_I(inode_out);
1308 bool same_inode = (inode_in == inode_out);
1311 /* Lock both files against IO */
1312 ret = xfs_iolock_two_inodes_and_break_layout(inode_in, inode_out);
1316 xfs_ilock(src, XFS_MMAPLOCK_EXCL);
1318 xfs_lock_two_inodes(src, XFS_MMAPLOCK_SHARED, dest,
1321 /* Check file eligibility and prepare for block sharing. */
1323 /* Don't reflink realtime inodes */
1324 if (XFS_IS_REALTIME_INODE(src) || XFS_IS_REALTIME_INODE(dest))
1327 /* Don't share DAX file data for now. */
1328 if (IS_DAX(inode_in) || IS_DAX(inode_out))
1331 ret = generic_remap_file_range_prep(file_in, pos_in, file_out, pos_out,
1333 if (ret < 0 || *len == 0)
1336 /* Attach dquots to dest inode before changing block map */
1337 ret = xfs_qm_dqattach(dest);
1342 * Zero existing post-eof speculative preallocations in the destination
1345 ret = xfs_reflink_zero_posteof(dest, pos_out);
1349 /* Set flags and remap blocks. */
1350 ret = xfs_reflink_set_inode_flag(src, dest);
1354 /* Zap any page cache for the destination file's range. */
1355 truncate_inode_pages_range(&inode_out->i_data,
1356 round_down(pos_out, PAGE_SIZE),
1357 round_up(pos_out + *len, PAGE_SIZE) - 1);
1361 xfs_reflink_remap_unlock(file_in, file_out);
1366 * The user wants to preemptively CoW all shared blocks in this file,
1367 * which enables us to turn off the reflink flag. Iterate all
1368 * extents which are not prealloc/delalloc to see which ranges are
1369 * mentioned in the refcount tree, then read those blocks into the
1370 * pagecache, dirty them, fsync them back out, and then we can update
1371 * the inode flag. What happens if we run out of memory? :)
1374 xfs_reflink_dirty_extents(
1375 struct xfs_inode *ip,
1380 struct xfs_mount *mp = ip->i_mount;
1381 xfs_agnumber_t agno;
1382 xfs_agblock_t agbno;
1388 struct xfs_bmbt_irec map[2];
1392 while (end - fbno > 0) {
1395 * Look for extents in the file. Skip holes, delalloc, or
1396 * unwritten extents; they can't be reflinked.
1398 error = xfs_bmapi_read(ip, fbno, end - fbno, map, &nmaps, 0);
1403 if (!xfs_bmap_is_real_extent(&map[0]))
1407 while (map[1].br_blockcount) {
1408 agno = XFS_FSB_TO_AGNO(mp, map[1].br_startblock);
1409 agbno = XFS_FSB_TO_AGBNO(mp, map[1].br_startblock);
1410 aglen = map[1].br_blockcount;
1412 error = xfs_reflink_find_shared(mp, NULL, agno, agbno,
1413 aglen, &rbno, &rlen, true);
1416 if (rbno == NULLAGBLOCK)
1419 /* Dirty the pages */
1420 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1421 fpos = XFS_FSB_TO_B(mp, map[1].br_startoff +
1423 flen = XFS_FSB_TO_B(mp, rlen);
1424 if (fpos + flen > isize)
1425 flen = isize - fpos;
1426 error = iomap_file_dirty(VFS_I(ip), fpos, flen,
1428 xfs_ilock(ip, XFS_ILOCK_EXCL);
1432 map[1].br_blockcount -= (rbno - agbno + rlen);
1433 map[1].br_startoff += (rbno - agbno + rlen);
1434 map[1].br_startblock += (rbno - agbno + rlen);
1438 fbno = map[0].br_startoff + map[0].br_blockcount;
1444 /* Does this inode need the reflink flag? */
1446 xfs_reflink_inode_has_shared_extents(
1447 struct xfs_trans *tp,
1448 struct xfs_inode *ip,
1451 struct xfs_bmbt_irec got;
1452 struct xfs_mount *mp = ip->i_mount;
1453 struct xfs_ifork *ifp;
1454 xfs_agnumber_t agno;
1455 xfs_agblock_t agbno;
1459 struct xfs_iext_cursor icur;
1463 ifp = XFS_IFORK_PTR(ip, XFS_DATA_FORK);
1464 if (!(ifp->if_flags & XFS_IFEXTENTS)) {
1465 error = xfs_iread_extents(tp, ip, XFS_DATA_FORK);
1470 *has_shared = false;
1471 found = xfs_iext_lookup_extent(ip, ifp, 0, &icur, &got);
1473 if (isnullstartblock(got.br_startblock) ||
1474 got.br_state != XFS_EXT_NORM)
1476 agno = XFS_FSB_TO_AGNO(mp, got.br_startblock);
1477 agbno = XFS_FSB_TO_AGBNO(mp, got.br_startblock);
1478 aglen = got.br_blockcount;
1480 error = xfs_reflink_find_shared(mp, tp, agno, agbno, aglen,
1481 &rbno, &rlen, false);
1484 /* Is there still a shared block here? */
1485 if (rbno != NULLAGBLOCK) {
1490 found = xfs_iext_next_extent(ifp, &icur, &got);
1497 * Clear the inode reflink flag if there are no shared extents.
1499 * The caller is responsible for joining the inode to the transaction passed in.
1500 * The inode will be joined to the transaction that is returned to the caller.
1503 xfs_reflink_clear_inode_flag(
1504 struct xfs_inode *ip,
1505 struct xfs_trans **tpp)
1510 ASSERT(xfs_is_reflink_inode(ip));
1512 error = xfs_reflink_inode_has_shared_extents(*tpp, ip, &needs_flag);
1513 if (error || needs_flag)
1517 * We didn't find any shared blocks so turn off the reflink flag.
1518 * First, get rid of any leftover CoW mappings.
1520 error = xfs_reflink_cancel_cow_blocks(ip, tpp, 0, NULLFILEOFF, true);
1524 /* Clear the inode flag. */
1525 trace_xfs_reflink_unset_inode_flag(ip);
1526 ip->i_d.di_flags2 &= ~XFS_DIFLAG2_REFLINK;
1527 xfs_inode_clear_cowblocks_tag(ip);
1528 xfs_trans_log_inode(*tpp, ip, XFS_ILOG_CORE);
1534 * Clear the inode reflink flag if there are no shared extents and the size
1538 xfs_reflink_try_clear_inode_flag(
1539 struct xfs_inode *ip)
1541 struct xfs_mount *mp = ip->i_mount;
1542 struct xfs_trans *tp;
1545 /* Start a rolling transaction to remove the mappings */
1546 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, 0, 0, 0, &tp);
1550 xfs_ilock(ip, XFS_ILOCK_EXCL);
1551 xfs_trans_ijoin(tp, ip, 0);
1553 error = xfs_reflink_clear_inode_flag(ip, &tp);
1557 error = xfs_trans_commit(tp);
1561 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1564 xfs_trans_cancel(tp);
1566 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1571 * Pre-COW all shared blocks within a given byte range of a file and turn off
1572 * the reflink flag if we unshare all of the file's blocks.
1575 xfs_reflink_unshare(
1576 struct xfs_inode *ip,
1580 struct xfs_mount *mp = ip->i_mount;
1586 if (!xfs_is_reflink_inode(ip))
1589 trace_xfs_reflink_unshare(ip, offset, len);
1591 inode_dio_wait(VFS_I(ip));
1593 /* Try to CoW the selected ranges */
1594 xfs_ilock(ip, XFS_ILOCK_EXCL);
1595 fbno = XFS_B_TO_FSBT(mp, offset);
1596 isize = i_size_read(VFS_I(ip));
1597 end = XFS_B_TO_FSB(mp, offset + len);
1598 error = xfs_reflink_dirty_extents(ip, fbno, end, isize);
1601 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1603 /* Wait for the IO to finish */
1604 error = filemap_write_and_wait(VFS_I(ip)->i_mapping);
1608 /* Turn off the reflink flag if possible. */
1609 error = xfs_reflink_try_clear_inode_flag(ip);
1616 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1618 trace_xfs_reflink_unshare_error(ip, error, _RET_IP_);