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 * When dirty pages are being written out (typically in writepage), the
86 * delalloc reservations are converted into real mappings by allocating
87 * blocks and replacing the delalloc mapping with real ones. A delalloc
88 * mapping can be replaced by several real ones if the free space is
91 * We want to adapt the delalloc mechanism for copy-on-write, since the
92 * write paths are similar. The first two steps (creating the reservation
93 * and allocating the blocks) are exactly the same as delalloc except that
94 * the mappings must be stored in a separate CoW fork because we do not want
95 * to disturb the mapping in the data fork until we're sure that the write
96 * succeeded. IO completion in this case is the process of removing the old
97 * mapping from the data fork and moving the new mapping from the CoW fork to
98 * the data fork. This will be discussed shortly.
100 * For now, unaligned directio writes will be bounced back to the page cache.
101 * Block-aligned directio writes will use the same mechanism as buffered
104 * CoW remapping must be done after the data block write completes,
105 * because we don't want to destroy the old data fork map until we're sure
106 * the new block has been written. Since the new mappings are kept in a
107 * separate fork, we can simply iterate these mappings to find the ones
108 * that cover the file blocks that we just CoW'd. For each extent, simply
109 * unmap the corresponding range in the data fork, map the new range into
110 * the data fork, and remove the extent from the CoW fork.
112 * Since the remapping operation can be applied to an arbitrary file
113 * range, we record the need for the remap step as a flag in the ioend
114 * instead of declaring a new IO type. This is required for direct io
115 * because we only have ioend for the whole dio, and we have to be able to
116 * remember the presence of unwritten blocks and CoW blocks with a single
117 * ioend structure. Better yet, the more ground we can cover with one
122 * Given an AG extent, find the lowest-numbered run of shared blocks
123 * within that range and return the range in fbno/flen. If
124 * find_end_of_shared is true, return the longest contiguous extent of
125 * shared blocks. If there are no shared extents, fbno and flen will
126 * be set to NULLAGBLOCK and 0, respectively.
129 xfs_reflink_find_shared(
130 struct xfs_mount *mp,
136 bool find_end_of_shared)
138 struct xfs_buf *agbp;
139 struct xfs_btree_cur *cur;
142 error = xfs_alloc_read_agf(mp, NULL, agno, 0, &agbp);
146 cur = xfs_refcountbt_init_cursor(mp, NULL, agbp, agno, NULL);
148 error = xfs_refcount_find_shared(cur, agbno, aglen, fbno, flen,
151 xfs_btree_del_cursor(cur, error ? XFS_BTREE_ERROR : XFS_BTREE_NOERROR);
158 * Trim the mapping to the next block where there's a change in the
159 * shared/unshared status. More specifically, this means that we
160 * find the lowest-numbered extent of shared blocks that coincides with
161 * the given block mapping. If the shared extent overlaps the start of
162 * the mapping, trim the mapping to the end of the shared extent. If
163 * the shared region intersects the mapping, trim the mapping to the
164 * start of the shared extent. If there are no shared regions that
165 * overlap, just return the original extent.
168 xfs_reflink_trim_around_shared(
169 struct xfs_inode *ip,
170 struct xfs_bmbt_irec *irec,
181 /* Holes, unwritten, and delalloc extents cannot be shared */
182 if (!xfs_is_reflink_inode(ip) ||
184 irec->br_startblock == HOLESTARTBLOCK ||
185 irec->br_startblock == DELAYSTARTBLOCK ||
186 isnullstartblock(irec->br_startblock)) {
191 trace_xfs_reflink_trim_around_shared(ip, irec);
193 agno = XFS_FSB_TO_AGNO(ip->i_mount, irec->br_startblock);
194 agbno = XFS_FSB_TO_AGBNO(ip->i_mount, irec->br_startblock);
195 aglen = irec->br_blockcount;
197 error = xfs_reflink_find_shared(ip->i_mount, agno, agbno,
198 aglen, &fbno, &flen, true);
202 *shared = *trimmed = false;
203 if (fbno == NULLAGBLOCK) {
204 /* No shared blocks at all. */
206 } else if (fbno == agbno) {
208 * The start of this extent is shared. Truncate the
209 * mapping at the end of the shared region so that a
210 * subsequent iteration starts at the start of the
213 irec->br_blockcount = flen;
220 * There's a shared extent midway through this extent.
221 * Truncate the mapping at the start of the shared
222 * extent so that a subsequent iteration starts at the
223 * start of the shared region.
225 irec->br_blockcount = fbno - agbno;
232 * Trim the passed in imap to the next shared/unshared extent boundary, and
233 * if imap->br_startoff points to a shared extent reserve space for it in the
234 * COW fork. In this case *shared is set to true, else to false.
236 * Note that imap will always contain the block numbers for the existing blocks
237 * in the data fork, as the upper layers need them for read-modify-write
241 xfs_reflink_reserve_cow(
242 struct xfs_inode *ip,
243 struct xfs_bmbt_irec *imap,
246 struct xfs_bmbt_irec got, prev;
247 xfs_fileoff_t end_fsb, orig_end_fsb;
248 int eof = 0, error = 0;
254 * Search the COW fork extent list first. This serves two purposes:
255 * first this implement the speculative preallocation using cowextisze,
256 * so that we also unshared block adjacent to shared blocks instead
257 * of just the shared blocks themselves. Second the lookup in the
258 * extent list is generally faster than going out to the shared extent
261 xfs_bmap_search_extents(ip, imap->br_startoff, XFS_COW_FORK, &eof, &idx,
263 if (!eof && got.br_startoff <= imap->br_startoff) {
264 trace_xfs_reflink_cow_found(ip, imap);
265 xfs_trim_extent(imap, got.br_startoff, got.br_blockcount);
271 /* Trim the mapping to the nearest shared extent boundary. */
272 error = xfs_reflink_trim_around_shared(ip, imap, shared, &trimmed);
276 /* Not shared? Just report the (potentially capped) extent. */
281 * Fork all the shared blocks from our write offset until the end of
284 error = xfs_qm_dqattach_locked(ip, 0);
288 end_fsb = orig_end_fsb = imap->br_startoff + imap->br_blockcount;
290 align = xfs_eof_alignment(ip, xfs_get_cowextsz_hint(ip));
292 end_fsb = roundup_64(end_fsb, align);
295 error = xfs_bmapi_reserve_delalloc(ip, XFS_COW_FORK, imap->br_startoff,
296 end_fsb - imap->br_startoff, &got, &prev, &idx, eof);
302 /* retry without any preallocation */
303 trace_xfs_reflink_cow_enospc(ip, imap);
304 if (end_fsb != orig_end_fsb) {
305 end_fsb = orig_end_fsb;
313 if (end_fsb != orig_end_fsb)
314 xfs_inode_set_cowblocks_tag(ip);
316 trace_xfs_reflink_cow_alloc(ip, &got);
320 /* Allocate all CoW reservations covering a range of blocks in a file. */
322 __xfs_reflink_allocate_cow(
323 struct xfs_inode *ip,
324 xfs_fileoff_t *offset_fsb,
325 xfs_fileoff_t end_fsb)
327 struct xfs_mount *mp = ip->i_mount;
328 struct xfs_bmbt_irec imap;
329 struct xfs_defer_ops dfops;
330 struct xfs_trans *tp;
331 xfs_fsblock_t first_block;
332 int nimaps = 1, error;
335 xfs_defer_init(&dfops, &first_block);
337 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, 0, 0,
338 XFS_TRANS_RESERVE, &tp);
342 xfs_ilock(ip, XFS_ILOCK_EXCL);
344 /* Read extent from the source file. */
346 error = xfs_bmapi_read(ip, *offset_fsb, end_fsb - *offset_fsb,
352 error = xfs_reflink_reserve_cow(ip, &imap, &shared);
354 goto out_trans_cancel;
357 *offset_fsb = imap.br_startoff + imap.br_blockcount;
358 goto out_trans_cancel;
361 xfs_trans_ijoin(tp, ip, 0);
362 error = xfs_bmapi_write(tp, ip, imap.br_startoff, imap.br_blockcount,
363 XFS_BMAPI_COWFORK, &first_block,
364 XFS_EXTENTADD_SPACE_RES(mp, XFS_DATA_FORK),
365 &imap, &nimaps, &dfops);
367 goto out_trans_cancel;
369 error = xfs_defer_finish(&tp, &dfops, NULL);
371 goto out_trans_cancel;
373 error = xfs_trans_commit(tp);
375 *offset_fsb = imap.br_startoff + imap.br_blockcount;
377 xfs_iunlock(ip, XFS_ILOCK_EXCL);
380 xfs_defer_cancel(&dfops);
381 xfs_trans_cancel(tp);
385 /* Allocate all CoW reservations covering a part of a file. */
387 xfs_reflink_allocate_cow_range(
388 struct xfs_inode *ip,
392 struct xfs_mount *mp = ip->i_mount;
393 xfs_fileoff_t offset_fsb = XFS_B_TO_FSBT(mp, offset);
394 xfs_fileoff_t end_fsb = XFS_B_TO_FSB(mp, offset + count);
397 ASSERT(xfs_is_reflink_inode(ip));
399 trace_xfs_reflink_allocate_cow_range(ip, offset, count);
402 * Make sure that the dquots are there.
404 error = xfs_qm_dqattach(ip, 0);
408 while (offset_fsb < end_fsb) {
409 error = __xfs_reflink_allocate_cow(ip, &offset_fsb, end_fsb);
411 trace_xfs_reflink_allocate_cow_range_error(ip, error,
421 * Find the CoW reservation (and whether or not it needs block allocation)
422 * for a given byte offset of a file.
425 xfs_reflink_find_cow_mapping(
426 struct xfs_inode *ip,
428 struct xfs_bmbt_irec *imap,
431 struct xfs_bmbt_irec irec;
432 struct xfs_ifork *ifp;
433 struct xfs_bmbt_rec_host *gotp;
437 ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL | XFS_ILOCK_SHARED));
438 ASSERT(xfs_is_reflink_inode(ip));
440 /* Find the extent in the CoW fork. */
441 ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
442 bno = XFS_B_TO_FSBT(ip->i_mount, offset);
443 gotp = xfs_iext_bno_to_ext(ifp, bno, &idx);
447 xfs_bmbt_get_all(gotp, &irec);
448 if (bno >= irec.br_startoff + irec.br_blockcount ||
449 bno < irec.br_startoff)
452 trace_xfs_reflink_find_cow_mapping(ip, offset, 1, XFS_IO_OVERWRITE,
455 /* If it's still delalloc, we must allocate later. */
457 *need_alloc = !!(isnullstartblock(irec.br_startblock));
463 * Trim an extent to end at the next CoW reservation past offset_fsb.
466 xfs_reflink_trim_irec_to_next_cow(
467 struct xfs_inode *ip,
468 xfs_fileoff_t offset_fsb,
469 struct xfs_bmbt_irec *imap)
471 struct xfs_bmbt_irec irec;
472 struct xfs_ifork *ifp;
473 struct xfs_bmbt_rec_host *gotp;
476 if (!xfs_is_reflink_inode(ip))
479 /* Find the extent in the CoW fork. */
480 ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
481 gotp = xfs_iext_bno_to_ext(ifp, offset_fsb, &idx);
484 xfs_bmbt_get_all(gotp, &irec);
486 /* This is the extent before; try sliding up one. */
487 if (irec.br_startoff < offset_fsb) {
489 if (idx >= ifp->if_bytes / sizeof(xfs_bmbt_rec_t))
491 gotp = xfs_iext_get_ext(ifp, idx);
492 xfs_bmbt_get_all(gotp, &irec);
495 if (irec.br_startoff >= imap->br_startoff + imap->br_blockcount)
498 imap->br_blockcount = irec.br_startoff - imap->br_startoff;
499 trace_xfs_reflink_trim_irec(ip, imap);
505 * Cancel all pending CoW reservations for some block range of an inode.
508 xfs_reflink_cancel_cow_blocks(
509 struct xfs_inode *ip,
510 struct xfs_trans **tpp,
511 xfs_fileoff_t offset_fsb,
512 xfs_fileoff_t end_fsb)
514 struct xfs_bmbt_irec irec;
515 xfs_filblks_t count_fsb;
516 xfs_fsblock_t firstfsb;
517 struct xfs_defer_ops dfops;
521 if (!xfs_is_reflink_inode(ip))
524 /* Go find the old extent in the CoW fork. */
525 while (offset_fsb < end_fsb) {
527 count_fsb = (xfs_filblks_t)(end_fsb - offset_fsb);
528 error = xfs_bmapi_read(ip, offset_fsb, count_fsb, &irec,
529 &nimaps, XFS_BMAPI_COWFORK);
534 trace_xfs_reflink_cancel_cow(ip, &irec);
536 if (irec.br_startblock == DELAYSTARTBLOCK) {
537 /* Free a delayed allocation. */
538 xfs_mod_fdblocks(ip->i_mount, irec.br_blockcount,
540 ip->i_delayed_blks -= irec.br_blockcount;
542 /* Remove the mapping from the CoW fork. */
543 error = xfs_bunmapi_cow(ip, &irec);
546 } else if (irec.br_startblock == HOLESTARTBLOCK) {
549 xfs_trans_ijoin(*tpp, ip, 0);
550 xfs_defer_init(&dfops, &firstfsb);
552 /* Free the CoW orphan record. */
553 error = xfs_refcount_free_cow_extent(ip->i_mount,
554 &dfops, irec.br_startblock,
559 xfs_bmap_add_free(ip->i_mount, &dfops,
560 irec.br_startblock, irec.br_blockcount,
563 /* Update quota accounting */
564 xfs_trans_mod_dquot_byino(*tpp, ip, XFS_TRANS_DQ_BCOUNT,
565 -(long)irec.br_blockcount);
567 /* Roll the transaction */
568 error = xfs_defer_finish(tpp, &dfops, ip);
570 xfs_defer_cancel(&dfops);
574 /* Remove the mapping from the CoW fork. */
575 error = xfs_bunmapi_cow(ip, &irec);
581 offset_fsb = irec.br_startoff + irec.br_blockcount;
588 * Cancel all pending CoW reservations for some byte range of an inode.
591 xfs_reflink_cancel_cow_range(
592 struct xfs_inode *ip,
596 struct xfs_trans *tp;
597 xfs_fileoff_t offset_fsb;
598 xfs_fileoff_t end_fsb;
601 trace_xfs_reflink_cancel_cow_range(ip, offset, count);
602 ASSERT(xfs_is_reflink_inode(ip));
604 offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
605 if (count == NULLFILEOFF)
606 end_fsb = NULLFILEOFF;
608 end_fsb = XFS_B_TO_FSB(ip->i_mount, offset + count);
610 /* Start a rolling transaction to remove the mappings */
611 error = xfs_trans_alloc(ip->i_mount, &M_RES(ip->i_mount)->tr_write,
616 xfs_ilock(ip, XFS_ILOCK_EXCL);
617 xfs_trans_ijoin(tp, ip, 0);
619 /* Scrape out the old CoW reservations */
620 error = xfs_reflink_cancel_cow_blocks(ip, &tp, offset_fsb, end_fsb);
624 error = xfs_trans_commit(tp);
626 xfs_iunlock(ip, XFS_ILOCK_EXCL);
630 xfs_trans_cancel(tp);
631 xfs_iunlock(ip, XFS_ILOCK_EXCL);
633 trace_xfs_reflink_cancel_cow_range_error(ip, error, _RET_IP_);
638 * Remap parts of a file's data fork after a successful CoW.
642 struct xfs_inode *ip,
646 struct xfs_bmbt_irec irec;
647 struct xfs_bmbt_irec uirec;
648 struct xfs_trans *tp;
649 xfs_fileoff_t offset_fsb;
650 xfs_fileoff_t end_fsb;
651 xfs_filblks_t count_fsb;
652 xfs_fsblock_t firstfsb;
653 struct xfs_defer_ops dfops;
655 unsigned int resblks;
660 trace_xfs_reflink_end_cow(ip, offset, count);
662 offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
663 end_fsb = XFS_B_TO_FSB(ip->i_mount, offset + count);
664 count_fsb = (xfs_filblks_t)(end_fsb - offset_fsb);
666 /* Start a rolling transaction to switch the mappings */
667 resblks = XFS_EXTENTADD_SPACE_RES(ip->i_mount, XFS_DATA_FORK);
668 error = xfs_trans_alloc(ip->i_mount, &M_RES(ip->i_mount)->tr_write,
673 xfs_ilock(ip, XFS_ILOCK_EXCL);
674 xfs_trans_ijoin(tp, ip, 0);
676 /* Go find the old extent in the CoW fork. */
677 while (offset_fsb < end_fsb) {
678 /* Read extent from the source file */
680 count_fsb = (xfs_filblks_t)(end_fsb - offset_fsb);
681 error = xfs_bmapi_read(ip, offset_fsb, count_fsb, &irec,
682 &nimaps, XFS_BMAPI_COWFORK);
687 ASSERT(irec.br_startblock != DELAYSTARTBLOCK);
688 trace_xfs_reflink_cow_remap(ip, &irec);
691 * We can have a hole in the CoW fork if part of a directio
692 * write is CoW but part of it isn't.
694 rlen = ilen = irec.br_blockcount;
695 if (irec.br_startblock == HOLESTARTBLOCK)
698 /* Unmap the old blocks in the data fork. */
700 xfs_defer_init(&dfops, &firstfsb);
701 error = __xfs_bunmapi(tp, ip, irec.br_startoff,
702 &rlen, 0, 1, &firstfsb, &dfops);
707 * Trim the extent to whatever got unmapped.
708 * Remember, bunmapi works backwards.
710 uirec.br_startblock = irec.br_startblock + rlen;
711 uirec.br_startoff = irec.br_startoff + rlen;
712 uirec.br_blockcount = irec.br_blockcount - rlen;
713 irec.br_blockcount = rlen;
714 trace_xfs_reflink_cow_remap_piece(ip, &uirec);
716 /* Free the CoW orphan record. */
717 error = xfs_refcount_free_cow_extent(tp->t_mountp,
718 &dfops, uirec.br_startblock,
719 uirec.br_blockcount);
723 /* Map the new blocks into the data fork. */
724 error = xfs_bmap_map_extent(tp->t_mountp, &dfops,
729 /* Remove the mapping from the CoW fork. */
730 error = xfs_bunmapi_cow(ip, &uirec);
734 error = xfs_defer_finish(&tp, &dfops, ip);
741 offset_fsb = irec.br_startoff + ilen;
744 error = xfs_trans_commit(tp);
745 xfs_iunlock(ip, XFS_ILOCK_EXCL);
751 xfs_defer_cancel(&dfops);
753 xfs_trans_cancel(tp);
754 xfs_iunlock(ip, XFS_ILOCK_EXCL);
756 trace_xfs_reflink_end_cow_error(ip, error, _RET_IP_);
761 * Free leftover CoW reservations that didn't get cleaned out.
764 xfs_reflink_recover_cow(
765 struct xfs_mount *mp)
770 if (!xfs_sb_version_hasreflink(&mp->m_sb))
773 for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
774 error = xfs_refcount_recover_cow_leftovers(mp, agno);
783 * Reflinking (Block) Ranges of Two Files Together
785 * First, ensure that the reflink flag is set on both inodes. The flag is an
786 * optimization to avoid unnecessary refcount btree lookups in the write path.
788 * Now we can iteratively remap the range of extents (and holes) in src to the
789 * corresponding ranges in dest. Let drange and srange denote the ranges of
790 * logical blocks in dest and src touched by the reflink operation.
792 * While the length of drange is greater than zero,
793 * - Read src's bmbt at the start of srange ("imap")
794 * - If imap doesn't exist, make imap appear to start at the end of srange
796 * - If imap starts before srange, advance imap to start at srange.
797 * - If imap goes beyond srange, truncate imap to end at the end of srange.
798 * - Punch (imap start - srange start + imap len) blocks from dest at
799 * offset (drange start).
800 * - If imap points to a real range of pblks,
801 * > Increase the refcount of the imap's pblks
802 * > Map imap's pblks into dest at the offset
803 * (drange start + imap start - srange start)
804 * - Advance drange and srange by (imap start - srange start + imap len)
806 * Finally, if the reflink made dest longer, update both the in-core and
807 * on-disk file sizes.
809 * ASCII Art Demonstration:
811 * Let's say we want to reflink this source file:
813 * ----SSSSSSS-SSSSS----SSSSSS (src file)
814 * <-------------------->
816 * into this destination file:
818 * --DDDDDDDDDDDDDDDDDDD--DDD (dest file)
819 * <-------------------->
820 * '-' means a hole, and 'S' and 'D' are written blocks in the src and dest.
821 * Observe that the range has different logical offsets in either file.
823 * Consider that the first extent in the source file doesn't line up with our
824 * reflink range. Unmapping and remapping are separate operations, so we can
825 * unmap more blocks from the destination file than we remap.
827 * ----SSSSSSS-SSSSS----SSSSSS
829 * --DDDDD---------DDDDD--DDD
832 * Now remap the source extent into the destination file:
834 * ----SSSSSSS-SSSSS----SSSSSS
836 * --DDDDD--SSSSSSSDDDDD--DDD
839 * Do likewise with the second hole and extent in our range. Holes in the
840 * unmap range don't affect our operation.
842 * ----SSSSSSS-SSSSS----SSSSSS
844 * --DDDDD--SSSSSSS-SSSSS-DDD
847 * Finally, unmap and remap part of the third extent. This will increase the
848 * size of the destination file.
850 * ----SSSSSSS-SSSSS----SSSSSS
852 * --DDDDD--SSSSSSS-SSSSS----SSS
855 * Once we update the destination file's i_size, we're done.
859 * Ensure the reflink bit is set in both inodes.
862 xfs_reflink_set_inode_flag(
863 struct xfs_inode *src,
864 struct xfs_inode *dest)
866 struct xfs_mount *mp = src->i_mount;
868 struct xfs_trans *tp;
870 if (xfs_is_reflink_inode(src) && xfs_is_reflink_inode(dest))
873 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
877 /* Lock both files against IO */
878 if (src->i_ino == dest->i_ino)
879 xfs_ilock(src, XFS_ILOCK_EXCL);
881 xfs_lock_two_inodes(src, dest, XFS_ILOCK_EXCL);
883 if (!xfs_is_reflink_inode(src)) {
884 trace_xfs_reflink_set_inode_flag(src);
885 xfs_trans_ijoin(tp, src, XFS_ILOCK_EXCL);
886 src->i_d.di_flags2 |= XFS_DIFLAG2_REFLINK;
887 xfs_trans_log_inode(tp, src, XFS_ILOG_CORE);
888 xfs_ifork_init_cow(src);
890 xfs_iunlock(src, XFS_ILOCK_EXCL);
892 if (src->i_ino == dest->i_ino)
895 if (!xfs_is_reflink_inode(dest)) {
896 trace_xfs_reflink_set_inode_flag(dest);
897 xfs_trans_ijoin(tp, dest, XFS_ILOCK_EXCL);
898 dest->i_d.di_flags2 |= XFS_DIFLAG2_REFLINK;
899 xfs_trans_log_inode(tp, dest, XFS_ILOG_CORE);
900 xfs_ifork_init_cow(dest);
902 xfs_iunlock(dest, XFS_ILOCK_EXCL);
905 error = xfs_trans_commit(tp);
911 trace_xfs_reflink_set_inode_flag_error(dest, error, _RET_IP_);
916 * Update destination inode size & cowextsize hint, if necessary.
919 xfs_reflink_update_dest(
920 struct xfs_inode *dest,
922 xfs_extlen_t cowextsize)
924 struct xfs_mount *mp = dest->i_mount;
925 struct xfs_trans *tp;
928 if (newlen <= i_size_read(VFS_I(dest)) && cowextsize == 0)
931 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
935 xfs_ilock(dest, XFS_ILOCK_EXCL);
936 xfs_trans_ijoin(tp, dest, XFS_ILOCK_EXCL);
938 if (newlen > i_size_read(VFS_I(dest))) {
939 trace_xfs_reflink_update_inode_size(dest, newlen);
940 i_size_write(VFS_I(dest), newlen);
941 dest->i_d.di_size = newlen;
945 dest->i_d.di_cowextsize = cowextsize;
946 dest->i_d.di_flags2 |= XFS_DIFLAG2_COWEXTSIZE;
949 xfs_trans_log_inode(tp, dest, XFS_ILOG_CORE);
951 error = xfs_trans_commit(tp);
957 trace_xfs_reflink_update_inode_size_error(dest, error, _RET_IP_);
962 * Do we have enough reserve in this AG to handle a reflink? The refcount
963 * btree already reserved all the space it needs, but the rmap btree can grow
964 * infinitely, so we won't allow more reflinks when the AG is down to the
968 xfs_reflink_ag_has_free_space(
969 struct xfs_mount *mp,
972 struct xfs_perag *pag;
975 if (!xfs_sb_version_hasrmapbt(&mp->m_sb))
978 pag = xfs_perag_get(mp, agno);
979 if (xfs_ag_resv_critical(pag, XFS_AG_RESV_AGFL) ||
980 xfs_ag_resv_critical(pag, XFS_AG_RESV_METADATA))
987 * Unmap a range of blocks from a file, then map other blocks into the hole.
988 * The range to unmap is (destoff : destoff + srcioff + irec->br_blockcount).
989 * The extent irec is mapped into dest at irec->br_startoff.
992 xfs_reflink_remap_extent(
993 struct xfs_inode *ip,
994 struct xfs_bmbt_irec *irec,
995 xfs_fileoff_t destoff,
998 struct xfs_mount *mp = ip->i_mount;
999 struct xfs_trans *tp;
1000 xfs_fsblock_t firstfsb;
1001 unsigned int resblks;
1002 struct xfs_defer_ops dfops;
1003 struct xfs_bmbt_irec uirec;
1006 xfs_filblks_t unmap_len;
1010 unmap_len = irec->br_startoff + irec->br_blockcount - destoff;
1011 trace_xfs_reflink_punch_range(ip, destoff, unmap_len);
1013 /* Only remap normal extents. */
1014 real_extent = (irec->br_startblock != HOLESTARTBLOCK &&
1015 irec->br_startblock != DELAYSTARTBLOCK &&
1016 !ISUNWRITTEN(irec));
1018 /* No reflinking if we're low on space */
1020 error = xfs_reflink_ag_has_free_space(mp,
1021 XFS_FSB_TO_AGNO(mp, irec->br_startblock));
1026 /* Start a rolling transaction to switch the mappings */
1027 resblks = XFS_EXTENTADD_SPACE_RES(ip->i_mount, XFS_DATA_FORK);
1028 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, resblks, 0, 0, &tp);
1032 xfs_ilock(ip, XFS_ILOCK_EXCL);
1033 xfs_trans_ijoin(tp, ip, 0);
1035 /* If we're not just clearing space, then do we have enough quota? */
1037 error = xfs_trans_reserve_quota_nblks(tp, ip,
1038 irec->br_blockcount, 0, XFS_QMOPT_RES_REGBLKS);
1043 trace_xfs_reflink_remap(ip, irec->br_startoff,
1044 irec->br_blockcount, irec->br_startblock);
1046 /* Unmap the old blocks in the data fork. */
1049 xfs_defer_init(&dfops, &firstfsb);
1050 error = __xfs_bunmapi(tp, ip, destoff, &rlen, 0, 1,
1056 * Trim the extent to whatever got unmapped.
1057 * Remember, bunmapi works backwards.
1059 uirec.br_startblock = irec->br_startblock + rlen;
1060 uirec.br_startoff = irec->br_startoff + rlen;
1061 uirec.br_blockcount = unmap_len - rlen;
1064 /* If this isn't a real mapping, we're done. */
1065 if (!real_extent || uirec.br_blockcount == 0)
1068 trace_xfs_reflink_remap(ip, uirec.br_startoff,
1069 uirec.br_blockcount, uirec.br_startblock);
1071 /* Update the refcount tree */
1072 error = xfs_refcount_increase_extent(mp, &dfops, &uirec);
1076 /* Map the new blocks into the data fork. */
1077 error = xfs_bmap_map_extent(mp, &dfops, ip, &uirec);
1081 /* Update quota accounting. */
1082 xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_BCOUNT,
1083 uirec.br_blockcount);
1085 /* Update dest isize if needed. */
1086 newlen = XFS_FSB_TO_B(mp,
1087 uirec.br_startoff + uirec.br_blockcount);
1088 newlen = min_t(xfs_off_t, newlen, new_isize);
1089 if (newlen > i_size_read(VFS_I(ip))) {
1090 trace_xfs_reflink_update_inode_size(ip, newlen);
1091 i_size_write(VFS_I(ip), newlen);
1092 ip->i_d.di_size = newlen;
1093 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1097 /* Process all the deferred stuff. */
1098 error = xfs_defer_finish(&tp, &dfops, ip);
1103 error = xfs_trans_commit(tp);
1104 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1110 xfs_defer_cancel(&dfops);
1112 xfs_trans_cancel(tp);
1113 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1115 trace_xfs_reflink_remap_extent_error(ip, error, _RET_IP_);
1120 * Iteratively remap one file's extents (and holes) to another's.
1123 xfs_reflink_remap_blocks(
1124 struct xfs_inode *src,
1125 xfs_fileoff_t srcoff,
1126 struct xfs_inode *dest,
1127 xfs_fileoff_t destoff,
1129 xfs_off_t new_isize)
1131 struct xfs_bmbt_irec imap;
1134 xfs_filblks_t range_len;
1136 /* drange = (destoff, destoff + len); srange = (srcoff, srcoff + len) */
1138 trace_xfs_reflink_remap_blocks_loop(src, srcoff, len,
1140 /* Read extent from the source file */
1142 xfs_ilock(src, XFS_ILOCK_EXCL);
1143 error = xfs_bmapi_read(src, srcoff, len, &imap, &nimaps, 0);
1144 xfs_iunlock(src, XFS_ILOCK_EXCL);
1147 ASSERT(nimaps == 1);
1149 trace_xfs_reflink_remap_imap(src, srcoff, len, XFS_IO_OVERWRITE,
1152 /* Translate imap into the destination file. */
1153 range_len = imap.br_startoff + imap.br_blockcount - srcoff;
1154 imap.br_startoff += destoff - srcoff;
1156 /* Clear dest from destoff to the end of imap and map it in. */
1157 error = xfs_reflink_remap_extent(dest, &imap, destoff,
1162 if (fatal_signal_pending(current)) {
1167 /* Advance drange/srange */
1168 srcoff += range_len;
1169 destoff += range_len;
1176 trace_xfs_reflink_remap_blocks_error(dest, error, _RET_IP_);
1181 * Read a page's worth of file data into the page cache. Return the page
1184 static struct page *
1186 struct inode *inode,
1189 struct address_space *mapping;
1193 n = offset >> PAGE_SHIFT;
1194 mapping = inode->i_mapping;
1195 page = read_mapping_page(mapping, n, NULL);
1198 if (!PageUptodate(page)) {
1200 return ERR_PTR(-EIO);
1207 * Compare extents of two files to see if they are the same.
1210 xfs_compare_extents(
1219 xfs_off_t dest_poff;
1222 struct page *src_page;
1223 struct page *dest_page;
1231 src_poff = srcoff & (PAGE_SIZE - 1);
1232 dest_poff = destoff & (PAGE_SIZE - 1);
1233 cmp_len = min(PAGE_SIZE - src_poff,
1234 PAGE_SIZE - dest_poff);
1235 cmp_len = min(cmp_len, len);
1236 ASSERT(cmp_len > 0);
1238 trace_xfs_reflink_compare_extents(XFS_I(src), srcoff, cmp_len,
1239 XFS_I(dest), destoff);
1241 src_page = xfs_get_page(src, srcoff);
1242 if (IS_ERR(src_page)) {
1243 error = PTR_ERR(src_page);
1246 dest_page = xfs_get_page(dest, destoff);
1247 if (IS_ERR(dest_page)) {
1248 error = PTR_ERR(dest_page);
1249 unlock_page(src_page);
1253 src_addr = kmap_atomic(src_page);
1254 dest_addr = kmap_atomic(dest_page);
1256 flush_dcache_page(src_page);
1257 flush_dcache_page(dest_page);
1259 if (memcmp(src_addr + src_poff, dest_addr + dest_poff, cmp_len))
1262 kunmap_atomic(dest_addr);
1263 kunmap_atomic(src_addr);
1264 unlock_page(dest_page);
1265 unlock_page(src_page);
1266 put_page(dest_page);
1281 trace_xfs_reflink_compare_extents_error(XFS_I(dest), error, _RET_IP_);
1286 * Link a range of blocks from one file to another.
1289 xfs_reflink_remap_range(
1290 struct file *file_in,
1292 struct file *file_out,
1297 struct inode *inode_in = file_inode(file_in);
1298 struct xfs_inode *src = XFS_I(inode_in);
1299 struct inode *inode_out = file_inode(file_out);
1300 struct xfs_inode *dest = XFS_I(inode_out);
1301 struct xfs_mount *mp = src->i_mount;
1302 loff_t bs = inode_out->i_sb->s_blocksize;
1303 bool same_inode = (inode_in == inode_out);
1304 xfs_fileoff_t sfsbno, dfsbno;
1305 xfs_filblks_t fsblen;
1306 xfs_extlen_t cowextsize;
1311 if (!xfs_sb_version_hasreflink(&mp->m_sb))
1314 if (XFS_FORCED_SHUTDOWN(mp))
1317 /* Lock both files against IO */
1319 xfs_ilock(src, XFS_IOLOCK_EXCL);
1320 xfs_ilock(src, XFS_MMAPLOCK_EXCL);
1322 xfs_lock_two_inodes(src, dest, XFS_IOLOCK_EXCL);
1323 xfs_lock_two_inodes(src, dest, XFS_MMAPLOCK_EXCL);
1326 /* Don't touch certain kinds of inodes */
1328 if (IS_IMMUTABLE(inode_out))
1332 if (IS_SWAPFILE(inode_in) || IS_SWAPFILE(inode_out))
1336 /* Don't reflink dirs, pipes, sockets... */
1338 if (S_ISDIR(inode_in->i_mode) || S_ISDIR(inode_out->i_mode))
1341 if (S_ISFIFO(inode_in->i_mode) || S_ISFIFO(inode_out->i_mode))
1343 if (!S_ISREG(inode_in->i_mode) || !S_ISREG(inode_out->i_mode))
1346 /* Don't reflink realtime inodes */
1347 if (XFS_IS_REALTIME_INODE(src) || XFS_IS_REALTIME_INODE(dest))
1350 /* Don't share DAX file data for now. */
1351 if (IS_DAX(inode_in) || IS_DAX(inode_out))
1354 /* Are we going all the way to the end? */
1355 isize = i_size_read(inode_in);
1362 len = isize - pos_in;
1364 /* Ensure offsets don't wrap and the input is inside i_size */
1365 if (pos_in + len < pos_in || pos_out + len < pos_out ||
1366 pos_in + len > isize)
1369 /* Don't allow dedupe past EOF in the dest file */
1373 disize = i_size_read(inode_out);
1374 if (pos_out >= disize || pos_out + len > disize)
1378 /* If we're linking to EOF, continue to the block boundary. */
1379 if (pos_in + len == isize)
1380 blen = ALIGN(isize, bs) - pos_in;
1384 /* Only reflink if we're aligned to block boundaries */
1385 if (!IS_ALIGNED(pos_in, bs) || !IS_ALIGNED(pos_in + blen, bs) ||
1386 !IS_ALIGNED(pos_out, bs) || !IS_ALIGNED(pos_out + blen, bs))
1389 /* Don't allow overlapped reflink within the same file */
1391 if (pos_out + blen > pos_in && pos_out < pos_in + blen)
1395 /* Wait for the completion of any pending IOs on both files */
1396 inode_dio_wait(inode_in);
1398 inode_dio_wait(inode_out);
1400 ret = filemap_write_and_wait_range(inode_in->i_mapping,
1401 pos_in, pos_in + len - 1);
1405 ret = filemap_write_and_wait_range(inode_out->i_mapping,
1406 pos_out, pos_out + len - 1);
1410 trace_xfs_reflink_remap_range(src, pos_in, len, dest, pos_out);
1413 * Check that the extents are the same.
1416 bool is_same = false;
1418 ret = xfs_compare_extents(inode_in, pos_in, inode_out, pos_out,
1428 ret = xfs_reflink_set_inode_flag(src, dest);
1433 * Invalidate the page cache so that we can clear any CoW mappings
1434 * in the destination file.
1436 truncate_inode_pages_range(&inode_out->i_data, pos_out,
1437 PAGE_ALIGN(pos_out + len) - 1);
1439 dfsbno = XFS_B_TO_FSBT(mp, pos_out);
1440 sfsbno = XFS_B_TO_FSBT(mp, pos_in);
1441 fsblen = XFS_B_TO_FSB(mp, len);
1442 ret = xfs_reflink_remap_blocks(src, sfsbno, dest, dfsbno, fsblen,
1448 * Carry the cowextsize hint from src to dest if we're sharing the
1449 * entire source file to the entire destination file, the source file
1450 * has a cowextsize hint, and the destination file does not.
1453 if (pos_in == 0 && len == i_size_read(inode_in) &&
1454 (src->i_d.di_flags2 & XFS_DIFLAG2_COWEXTSIZE) &&
1455 pos_out == 0 && len >= i_size_read(inode_out) &&
1456 !(dest->i_d.di_flags2 & XFS_DIFLAG2_COWEXTSIZE))
1457 cowextsize = src->i_d.di_cowextsize;
1459 ret = xfs_reflink_update_dest(dest, pos_out + len, cowextsize);
1462 xfs_iunlock(src, XFS_MMAPLOCK_EXCL);
1463 xfs_iunlock(src, XFS_IOLOCK_EXCL);
1464 if (src->i_ino != dest->i_ino) {
1465 xfs_iunlock(dest, XFS_MMAPLOCK_EXCL);
1466 xfs_iunlock(dest, XFS_IOLOCK_EXCL);
1469 trace_xfs_reflink_remap_range_error(dest, ret, _RET_IP_);
1474 * The user wants to preemptively CoW all shared blocks in this file,
1475 * which enables us to turn off the reflink flag. Iterate all
1476 * extents which are not prealloc/delalloc to see which ranges are
1477 * mentioned in the refcount tree, then read those blocks into the
1478 * pagecache, dirty them, fsync them back out, and then we can update
1479 * the inode flag. What happens if we run out of memory? :)
1482 xfs_reflink_dirty_extents(
1483 struct xfs_inode *ip,
1488 struct xfs_mount *mp = ip->i_mount;
1489 xfs_agnumber_t agno;
1490 xfs_agblock_t agbno;
1496 struct xfs_bmbt_irec map[2];
1500 while (end - fbno > 0) {
1503 * Look for extents in the file. Skip holes, delalloc, or
1504 * unwritten extents; they can't be reflinked.
1506 error = xfs_bmapi_read(ip, fbno, end - fbno, map, &nmaps, 0);
1511 if (map[0].br_startblock == HOLESTARTBLOCK ||
1512 map[0].br_startblock == DELAYSTARTBLOCK ||
1513 ISUNWRITTEN(&map[0]))
1517 while (map[1].br_blockcount) {
1518 agno = XFS_FSB_TO_AGNO(mp, map[1].br_startblock);
1519 agbno = XFS_FSB_TO_AGBNO(mp, map[1].br_startblock);
1520 aglen = map[1].br_blockcount;
1522 error = xfs_reflink_find_shared(mp, agno, agbno, aglen,
1523 &rbno, &rlen, true);
1526 if (rbno == NULLAGBLOCK)
1529 /* Dirty the pages */
1530 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1531 fpos = XFS_FSB_TO_B(mp, map[1].br_startoff +
1533 flen = XFS_FSB_TO_B(mp, rlen);
1534 if (fpos + flen > isize)
1535 flen = isize - fpos;
1536 error = iomap_file_dirty(VFS_I(ip), fpos, flen,
1538 xfs_ilock(ip, XFS_ILOCK_EXCL);
1542 map[1].br_blockcount -= (rbno - agbno + rlen);
1543 map[1].br_startoff += (rbno - agbno + rlen);
1544 map[1].br_startblock += (rbno - agbno + rlen);
1548 fbno = map[0].br_startoff + map[0].br_blockcount;
1554 /* Clear the inode reflink flag if there are no shared extents. */
1556 xfs_reflink_clear_inode_flag(
1557 struct xfs_inode *ip,
1558 struct xfs_trans **tpp)
1560 struct xfs_mount *mp = ip->i_mount;
1563 xfs_agnumber_t agno;
1564 xfs_agblock_t agbno;
1568 struct xfs_bmbt_irec map;
1572 ASSERT(xfs_is_reflink_inode(ip));
1575 end = XFS_B_TO_FSB(mp, i_size_read(VFS_I(ip)));
1576 while (end - fbno > 0) {
1579 * Look for extents in the file. Skip holes, delalloc, or
1580 * unwritten extents; they can't be reflinked.
1582 error = xfs_bmapi_read(ip, fbno, end - fbno, &map, &nmaps, 0);
1587 if (map.br_startblock == HOLESTARTBLOCK ||
1588 map.br_startblock == DELAYSTARTBLOCK ||
1592 agno = XFS_FSB_TO_AGNO(mp, map.br_startblock);
1593 agbno = XFS_FSB_TO_AGBNO(mp, map.br_startblock);
1594 aglen = map.br_blockcount;
1596 error = xfs_reflink_find_shared(mp, agno, agbno, aglen,
1597 &rbno, &rlen, false);
1600 /* Is there still a shared block here? */
1601 if (rbno != NULLAGBLOCK)
1604 fbno = map.br_startoff + map.br_blockcount;
1608 * We didn't find any shared blocks so turn off the reflink flag.
1609 * First, get rid of any leftover CoW mappings.
1611 error = xfs_reflink_cancel_cow_blocks(ip, tpp, 0, NULLFILEOFF);
1615 /* Clear the inode flag. */
1616 trace_xfs_reflink_unset_inode_flag(ip);
1617 ip->i_d.di_flags2 &= ~XFS_DIFLAG2_REFLINK;
1618 xfs_inode_clear_cowblocks_tag(ip);
1619 xfs_trans_ijoin(*tpp, ip, 0);
1620 xfs_trans_log_inode(*tpp, ip, XFS_ILOG_CORE);
1626 * Clear the inode reflink flag if there are no shared extents and the size
1630 xfs_reflink_try_clear_inode_flag(
1631 struct xfs_inode *ip)
1633 struct xfs_mount *mp = ip->i_mount;
1634 struct xfs_trans *tp;
1637 /* Start a rolling transaction to remove the mappings */
1638 error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, 0, 0, 0, &tp);
1642 xfs_ilock(ip, XFS_ILOCK_EXCL);
1643 xfs_trans_ijoin(tp, ip, 0);
1645 error = xfs_reflink_clear_inode_flag(ip, &tp);
1649 error = xfs_trans_commit(tp);
1653 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1656 xfs_trans_cancel(tp);
1658 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1663 * Pre-COW all shared blocks within a given byte range of a file and turn off
1664 * the reflink flag if we unshare all of the file's blocks.
1667 xfs_reflink_unshare(
1668 struct xfs_inode *ip,
1672 struct xfs_mount *mp = ip->i_mount;
1678 if (!xfs_is_reflink_inode(ip))
1681 trace_xfs_reflink_unshare(ip, offset, len);
1683 inode_dio_wait(VFS_I(ip));
1685 /* Try to CoW the selected ranges */
1686 xfs_ilock(ip, XFS_ILOCK_EXCL);
1687 fbno = XFS_B_TO_FSBT(mp, offset);
1688 isize = i_size_read(VFS_I(ip));
1689 end = XFS_B_TO_FSB(mp, offset + len);
1690 error = xfs_reflink_dirty_extents(ip, fbno, end, isize);
1693 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1695 /* Wait for the IO to finish */
1696 error = filemap_write_and_wait(VFS_I(ip)->i_mapping);
1700 /* Turn off the reflink flag if possible. */
1701 error = xfs_reflink_try_clear_inode_flag(ip);
1708 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1710 trace_xfs_reflink_unshare_error(ip, error, _RET_IP_);
1715 * Does this inode have any real CoW reservations?
1718 xfs_reflink_has_real_cow_blocks(
1719 struct xfs_inode *ip)
1721 struct xfs_bmbt_irec irec;
1722 struct xfs_ifork *ifp;
1723 struct xfs_bmbt_rec_host *gotp;
1726 if (!xfs_is_reflink_inode(ip))
1729 /* Go find the old extent in the CoW fork. */
1730 ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
1731 gotp = xfs_iext_bno_to_ext(ifp, 0, &idx);
1733 xfs_bmbt_get_all(gotp, &irec);
1735 if (!isnullstartblock(irec.br_startblock))
1740 if (idx >= ifp->if_bytes / sizeof(xfs_bmbt_rec_t))
1742 gotp = xfs_iext_get_ext(ifp, idx);