xref: /openbmc/linux/fs/xfs/xfs_reflink.c (revision a977d045)
1 /*
2  * Copyright (C) 2016 Oracle.  All Rights Reserved.
3  *
4  * Author: Darrick J. Wong <darrick.wong@oracle.com>
5  *
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.
10  *
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.
15  *
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.
19  */
20 #include "xfs.h"
21 #include "xfs_fs.h"
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"
33 #include "xfs_bmap.h"
34 #include "xfs_bmap_util.h"
35 #include "xfs_error.h"
36 #include "xfs_dir2.h"
37 #include "xfs_dir2_priv.h"
38 #include "xfs_ioctl.h"
39 #include "xfs_trace.h"
40 #include "xfs_log.h"
41 #include "xfs_icache.h"
42 #include "xfs_pnfs.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"
48 #include "xfs_bit.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"
57 #include "xfs_sb.h"
58 #include "xfs_ag_resv.h"
59 
60 /*
61  * Copy on Write of Shared Blocks
62  *
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.
70  *
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.
75  *
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.
84  *
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.
89  *
90  * D: --RRRRRRSSSRRRRRRRR--- (data fork)
91  * C: ------DDDDDDD--------- (CoW fork)
92  *
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.
98  *
99  * D: --RRRRRRSSSRRRRRRRR---
100  * C: ------UUUUUUU---------
101  *
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.
110  *
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
113  * writes.
114  *
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:
119  *
120  * D: --RRRRRRSSSRRRRRRRR---
121  * C: ------UUrrUUU---------
122  *
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
134  * the CoW fork:
135  *
136  * D: --RRRRRRrrSRRRRRRRR---
137  * C: ------UU--UUU---------
138  *
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
145  * ioend, the better.
146  */
147 
148 /*
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.
154  */
155 int
156 xfs_reflink_find_shared(
157 	struct xfs_mount	*mp,
158 	struct xfs_trans	*tp,
159 	xfs_agnumber_t		agno,
160 	xfs_agblock_t		agbno,
161 	xfs_extlen_t		aglen,
162 	xfs_agblock_t		*fbno,
163 	xfs_extlen_t		*flen,
164 	bool			find_end_of_shared)
165 {
166 	struct xfs_buf		*agbp;
167 	struct xfs_btree_cur	*cur;
168 	int			error;
169 
170 	error = xfs_alloc_read_agf(mp, tp, agno, 0, &agbp);
171 	if (error)
172 		return error;
173 	if (!agbp)
174 		return -ENOMEM;
175 
176 	cur = xfs_refcountbt_init_cursor(mp, tp, agbp, agno, NULL);
177 
178 	error = xfs_refcount_find_shared(cur, agbno, aglen, fbno, flen,
179 			find_end_of_shared);
180 
181 	xfs_btree_del_cursor(cur, error ? XFS_BTREE_ERROR : XFS_BTREE_NOERROR);
182 
183 	xfs_trans_brelse(tp, agbp);
184 	return error;
185 }
186 
187 /*
188  * Trim the mapping to the next block where there's a change in the
189  * shared/unshared status.  More specifically, this means that we
190  * find the lowest-numbered extent of shared blocks that coincides with
191  * the given block mapping.  If the shared extent overlaps the start of
192  * the mapping, trim the mapping to the end of the shared extent.  If
193  * the shared region intersects the mapping, trim the mapping to the
194  * start of the shared extent.  If there are no shared regions that
195  * overlap, just return the original extent.
196  */
197 int
198 xfs_reflink_trim_around_shared(
199 	struct xfs_inode	*ip,
200 	struct xfs_bmbt_irec	*irec,
201 	bool			*shared,
202 	bool			*trimmed)
203 {
204 	xfs_agnumber_t		agno;
205 	xfs_agblock_t		agbno;
206 	xfs_extlen_t		aglen;
207 	xfs_agblock_t		fbno;
208 	xfs_extlen_t		flen;
209 	int			error = 0;
210 
211 	/* Holes, unwritten, and delalloc extents cannot be shared */
212 	if (!xfs_is_reflink_inode(ip) || !xfs_bmap_is_real_extent(irec)) {
213 		*shared = false;
214 		return 0;
215 	}
216 
217 	trace_xfs_reflink_trim_around_shared(ip, irec);
218 
219 	agno = XFS_FSB_TO_AGNO(ip->i_mount, irec->br_startblock);
220 	agbno = XFS_FSB_TO_AGBNO(ip->i_mount, irec->br_startblock);
221 	aglen = irec->br_blockcount;
222 
223 	error = xfs_reflink_find_shared(ip->i_mount, NULL, agno, agbno,
224 			aglen, &fbno, &flen, true);
225 	if (error)
226 		return error;
227 
228 	*shared = *trimmed = false;
229 	if (fbno == NULLAGBLOCK) {
230 		/* No shared blocks at all. */
231 		return 0;
232 	} else if (fbno == agbno) {
233 		/*
234 		 * The start of this extent is shared.  Truncate the
235 		 * mapping at the end of the shared region so that a
236 		 * subsequent iteration starts at the start of the
237 		 * unshared region.
238 		 */
239 		irec->br_blockcount = flen;
240 		*shared = true;
241 		if (flen != aglen)
242 			*trimmed = true;
243 		return 0;
244 	} else {
245 		/*
246 		 * There's a shared extent midway through this extent.
247 		 * Truncate the mapping at the start of the shared
248 		 * extent so that a subsequent iteration starts at the
249 		 * start of the shared region.
250 		 */
251 		irec->br_blockcount = fbno - agbno;
252 		*trimmed = true;
253 		return 0;
254 	}
255 }
256 
257 /*
258  * Trim the passed in imap to the next shared/unshared extent boundary, and
259  * if imap->br_startoff points to a shared extent reserve space for it in the
260  * COW fork.  In this case *shared is set to true, else to false.
261  *
262  * Note that imap will always contain the block numbers for the existing blocks
263  * in the data fork, as the upper layers need them for read-modify-write
264  * operations.
265  */
266 int
267 xfs_reflink_reserve_cow(
268 	struct xfs_inode	*ip,
269 	struct xfs_bmbt_irec	*imap,
270 	bool			*shared)
271 {
272 	struct xfs_ifork	*ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
273 	struct xfs_bmbt_irec	got;
274 	int			error = 0;
275 	bool			eof = false, trimmed;
276 	xfs_extnum_t		idx;
277 
278 	/*
279 	 * Search the COW fork extent list first.  This serves two purposes:
280 	 * first this implement the speculative preallocation using cowextisze,
281 	 * so that we also unshared block adjacent to shared blocks instead
282 	 * of just the shared blocks themselves.  Second the lookup in the
283 	 * extent list is generally faster than going out to the shared extent
284 	 * tree.
285 	 */
286 
287 	if (!xfs_iext_lookup_extent(ip, ifp, imap->br_startoff, &idx, &got))
288 		eof = true;
289 	if (!eof && got.br_startoff <= imap->br_startoff) {
290 		trace_xfs_reflink_cow_found(ip, imap);
291 		xfs_trim_extent(imap, got.br_startoff, got.br_blockcount);
292 
293 		*shared = true;
294 		return 0;
295 	}
296 
297 	/* Trim the mapping to the nearest shared extent boundary. */
298 	error = xfs_reflink_trim_around_shared(ip, imap, shared, &trimmed);
299 	if (error)
300 		return error;
301 
302 	/* Not shared?  Just report the (potentially capped) extent. */
303 	if (!*shared)
304 		return 0;
305 
306 	/*
307 	 * Fork all the shared blocks from our write offset until the end of
308 	 * the extent.
309 	 */
310 	error = xfs_qm_dqattach_locked(ip, 0);
311 	if (error)
312 		return error;
313 
314 	error = xfs_bmapi_reserve_delalloc(ip, XFS_COW_FORK, imap->br_startoff,
315 			imap->br_blockcount, 0, &got, &idx, eof);
316 	if (error == -ENOSPC || error == -EDQUOT)
317 		trace_xfs_reflink_cow_enospc(ip, imap);
318 	if (error)
319 		return error;
320 
321 	trace_xfs_reflink_cow_alloc(ip, &got);
322 	return 0;
323 }
324 
325 /* Convert part of an unwritten CoW extent to a real one. */
326 STATIC int
327 xfs_reflink_convert_cow_extent(
328 	struct xfs_inode		*ip,
329 	struct xfs_bmbt_irec		*imap,
330 	xfs_fileoff_t			offset_fsb,
331 	xfs_filblks_t			count_fsb,
332 	struct xfs_defer_ops		*dfops)
333 {
334 	xfs_fsblock_t			first_block = NULLFSBLOCK;
335 	int				nimaps = 1;
336 
337 	if (imap->br_state == XFS_EXT_NORM)
338 		return 0;
339 
340 	xfs_trim_extent(imap, offset_fsb, count_fsb);
341 	trace_xfs_reflink_convert_cow(ip, imap);
342 	if (imap->br_blockcount == 0)
343 		return 0;
344 	return xfs_bmapi_write(NULL, ip, imap->br_startoff, imap->br_blockcount,
345 			XFS_BMAPI_COWFORK | XFS_BMAPI_CONVERT, &first_block,
346 			0, imap, &nimaps, dfops);
347 }
348 
349 /* Convert all of the unwritten CoW extents in a file's range to real ones. */
350 int
351 xfs_reflink_convert_cow(
352 	struct xfs_inode	*ip,
353 	xfs_off_t		offset,
354 	xfs_off_t		count)
355 {
356 	struct xfs_bmbt_irec	got;
357 	struct xfs_defer_ops	dfops;
358 	struct xfs_mount	*mp = ip->i_mount;
359 	struct xfs_ifork	*ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
360 	xfs_fileoff_t		offset_fsb = XFS_B_TO_FSBT(mp, offset);
361 	xfs_fileoff_t		end_fsb = XFS_B_TO_FSB(mp, offset + count);
362 	xfs_extnum_t		idx;
363 	bool			found;
364 	int			error = 0;
365 
366 	xfs_ilock(ip, XFS_ILOCK_EXCL);
367 
368 	/* Convert all the extents to real from unwritten. */
369 	for (found = xfs_iext_lookup_extent(ip, ifp, offset_fsb, &idx, &got);
370 	     found && got.br_startoff < end_fsb;
371 	     found = xfs_iext_get_extent(ifp, ++idx, &got)) {
372 		error = xfs_reflink_convert_cow_extent(ip, &got, offset_fsb,
373 				end_fsb - offset_fsb, &dfops);
374 		if (error)
375 			break;
376 	}
377 
378 	/* Finish up. */
379 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
380 	return error;
381 }
382 
383 /* Allocate all CoW reservations covering a range of blocks in a file. */
384 int
385 xfs_reflink_allocate_cow(
386 	struct xfs_inode	*ip,
387 	struct xfs_bmbt_irec	*imap,
388 	bool			*shared,
389 	uint			*lockmode)
390 {
391 	struct xfs_mount	*mp = ip->i_mount;
392 	xfs_fileoff_t		offset_fsb = imap->br_startoff;
393 	xfs_filblks_t		count_fsb = imap->br_blockcount;
394 	struct xfs_bmbt_irec	got;
395 	struct xfs_defer_ops	dfops;
396 	struct xfs_trans	*tp = NULL;
397 	xfs_fsblock_t		first_block;
398 	int			nimaps, error = 0;
399 	bool			trimmed;
400 	xfs_filblks_t		resaligned;
401 	xfs_extlen_t		resblks = 0;
402 	xfs_extnum_t		idx;
403 
404 retry:
405 	ASSERT(xfs_is_reflink_inode(ip));
406 	ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL | XFS_ILOCK_SHARED));
407 
408 	/*
409 	 * Even if the extent is not shared we might have a preallocation for
410 	 * it in the COW fork.  If so use it.
411 	 */
412 	if (xfs_iext_lookup_extent(ip, ip->i_cowfp, offset_fsb, &idx, &got) &&
413 	    got.br_startoff <= offset_fsb) {
414 		*shared = true;
415 
416 		/* If we have a real allocation in the COW fork we're done. */
417 		if (!isnullstartblock(got.br_startblock)) {
418 			xfs_trim_extent(&got, offset_fsb, count_fsb);
419 			*imap = got;
420 			goto convert;
421 		}
422 
423 		xfs_trim_extent(imap, got.br_startoff, got.br_blockcount);
424 	} else {
425 		error = xfs_reflink_trim_around_shared(ip, imap, shared, &trimmed);
426 		if (error || !*shared)
427 			goto out;
428 	}
429 
430 	if (!tp) {
431 		resaligned = xfs_aligned_fsb_count(imap->br_startoff,
432 			imap->br_blockcount, xfs_get_cowextsz_hint(ip));
433 		resblks = XFS_DIOSTRAT_SPACE_RES(mp, resaligned);
434 
435 		xfs_iunlock(ip, *lockmode);
436 		error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, resblks, 0, 0, &tp);
437 		*lockmode = XFS_ILOCK_EXCL;
438 		xfs_ilock(ip, *lockmode);
439 
440 		if (error)
441 			return error;
442 
443 		error = xfs_qm_dqattach_locked(ip, 0);
444 		if (error)
445 			goto out;
446 		goto retry;
447 	}
448 
449 	error = xfs_trans_reserve_quota_nblks(tp, ip, resblks, 0,
450 			XFS_QMOPT_RES_REGBLKS);
451 	if (error)
452 		goto out;
453 
454 	xfs_trans_ijoin(tp, ip, 0);
455 
456 	xfs_defer_init(&dfops, &first_block);
457 	nimaps = 1;
458 
459 	/* Allocate the entire reservation as unwritten blocks. */
460 	error = xfs_bmapi_write(tp, ip, imap->br_startoff, imap->br_blockcount,
461 			XFS_BMAPI_COWFORK | XFS_BMAPI_PREALLOC, &first_block,
462 			resblks, imap, &nimaps, &dfops);
463 	if (error)
464 		goto out_bmap_cancel;
465 
466 	/* Finish up. */
467 	error = xfs_defer_finish(&tp, &dfops, NULL);
468 	if (error)
469 		goto out_bmap_cancel;
470 
471 	error = xfs_trans_commit(tp);
472 	if (error)
473 		return error;
474 convert:
475 	return xfs_reflink_convert_cow_extent(ip, imap, offset_fsb, count_fsb,
476 			&dfops);
477 out_bmap_cancel:
478 	xfs_defer_cancel(&dfops);
479 	xfs_trans_unreserve_quota_nblks(tp, ip, (long)resblks, 0,
480 			XFS_QMOPT_RES_REGBLKS);
481 out:
482 	if (tp)
483 		xfs_trans_cancel(tp);
484 	return error;
485 }
486 
487 /*
488  * Find the CoW reservation for a given byte offset of a file.
489  */
490 bool
491 xfs_reflink_find_cow_mapping(
492 	struct xfs_inode		*ip,
493 	xfs_off_t			offset,
494 	struct xfs_bmbt_irec		*imap)
495 {
496 	struct xfs_ifork		*ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
497 	xfs_fileoff_t			offset_fsb;
498 	struct xfs_bmbt_irec		got;
499 	xfs_extnum_t			idx;
500 
501 	ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL | XFS_ILOCK_SHARED));
502 	ASSERT(xfs_is_reflink_inode(ip));
503 
504 	offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
505 	if (!xfs_iext_lookup_extent(ip, ifp, offset_fsb, &idx, &got))
506 		return false;
507 	if (got.br_startoff > offset_fsb)
508 		return false;
509 
510 	trace_xfs_reflink_find_cow_mapping(ip, offset, 1, XFS_IO_OVERWRITE,
511 			&got);
512 	*imap = got;
513 	return true;
514 }
515 
516 /*
517  * Trim an extent to end at the next CoW reservation past offset_fsb.
518  */
519 void
520 xfs_reflink_trim_irec_to_next_cow(
521 	struct xfs_inode		*ip,
522 	xfs_fileoff_t			offset_fsb,
523 	struct xfs_bmbt_irec		*imap)
524 {
525 	struct xfs_ifork		*ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
526 	struct xfs_bmbt_irec		got;
527 	xfs_extnum_t			idx;
528 
529 	if (!xfs_is_reflink_inode(ip))
530 		return;
531 
532 	/* Find the extent in the CoW fork. */
533 	if (!xfs_iext_lookup_extent(ip, ifp, offset_fsb, &idx, &got))
534 		return;
535 
536 	/* This is the extent before; try sliding up one. */
537 	if (got.br_startoff < offset_fsb) {
538 		if (!xfs_iext_get_extent(ifp, idx + 1, &got))
539 			return;
540 	}
541 
542 	if (got.br_startoff >= imap->br_startoff + imap->br_blockcount)
543 		return;
544 
545 	imap->br_blockcount = got.br_startoff - imap->br_startoff;
546 	trace_xfs_reflink_trim_irec(ip, imap);
547 }
548 
549 /*
550  * Cancel CoW reservations for some block range of an inode.
551  *
552  * If cancel_real is true this function cancels all COW fork extents for the
553  * inode; if cancel_real is false, real extents are not cleared.
554  */
555 int
556 xfs_reflink_cancel_cow_blocks(
557 	struct xfs_inode		*ip,
558 	struct xfs_trans		**tpp,
559 	xfs_fileoff_t			offset_fsb,
560 	xfs_fileoff_t			end_fsb,
561 	bool				cancel_real)
562 {
563 	struct xfs_ifork		*ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
564 	struct xfs_bmbt_irec		got, del;
565 	xfs_extnum_t			idx;
566 	xfs_fsblock_t			firstfsb;
567 	struct xfs_defer_ops		dfops;
568 	int				error = 0;
569 
570 	if (!xfs_is_reflink_inode(ip))
571 		return 0;
572 	if (!xfs_iext_lookup_extent(ip, ifp, offset_fsb, &idx, &got))
573 		return 0;
574 
575 	while (got.br_startoff < end_fsb) {
576 		del = got;
577 		xfs_trim_extent(&del, offset_fsb, end_fsb - offset_fsb);
578 		trace_xfs_reflink_cancel_cow(ip, &del);
579 
580 		if (isnullstartblock(del.br_startblock)) {
581 			error = xfs_bmap_del_extent_delay(ip, XFS_COW_FORK,
582 					&idx, &got, &del);
583 			if (error)
584 				break;
585 		} else if (del.br_state == XFS_EXT_UNWRITTEN || cancel_real) {
586 			xfs_trans_ijoin(*tpp, ip, 0);
587 			xfs_defer_init(&dfops, &firstfsb);
588 
589 			/* Free the CoW orphan record. */
590 			error = xfs_refcount_free_cow_extent(ip->i_mount,
591 					&dfops, del.br_startblock,
592 					del.br_blockcount);
593 			if (error)
594 				break;
595 
596 			xfs_bmap_add_free(ip->i_mount, &dfops,
597 					del.br_startblock, del.br_blockcount,
598 					NULL);
599 
600 			/* Update quota accounting */
601 			xfs_trans_mod_dquot_byino(*tpp, ip, XFS_TRANS_DQ_BCOUNT,
602 					-(long)del.br_blockcount);
603 
604 			/* Roll the transaction */
605 			error = xfs_defer_finish(tpp, &dfops, ip);
606 			if (error) {
607 				xfs_defer_cancel(&dfops);
608 				break;
609 			}
610 
611 			/* Remove the mapping from the CoW fork. */
612 			xfs_bmap_del_extent_cow(ip, &idx, &got, &del);
613 		}
614 
615 		if (!xfs_iext_get_extent(ifp, ++idx, &got))
616 			break;
617 	}
618 
619 	/* clear tag if cow fork is emptied */
620 	if (!ifp->if_bytes)
621 		xfs_inode_clear_cowblocks_tag(ip);
622 
623 	return error;
624 }
625 
626 /*
627  * Cancel CoW reservations for some byte range of an inode.
628  *
629  * If cancel_real is true this function cancels all COW fork extents for the
630  * inode; if cancel_real is false, real extents are not cleared.
631  */
632 int
633 xfs_reflink_cancel_cow_range(
634 	struct xfs_inode	*ip,
635 	xfs_off_t		offset,
636 	xfs_off_t		count,
637 	bool			cancel_real)
638 {
639 	struct xfs_trans	*tp;
640 	xfs_fileoff_t		offset_fsb;
641 	xfs_fileoff_t		end_fsb;
642 	int			error;
643 
644 	trace_xfs_reflink_cancel_cow_range(ip, offset, count);
645 	ASSERT(xfs_is_reflink_inode(ip));
646 
647 	offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
648 	if (count == NULLFILEOFF)
649 		end_fsb = NULLFILEOFF;
650 	else
651 		end_fsb = XFS_B_TO_FSB(ip->i_mount, offset + count);
652 
653 	/* Start a rolling transaction to remove the mappings */
654 	error = xfs_trans_alloc(ip->i_mount, &M_RES(ip->i_mount)->tr_write,
655 			0, 0, 0, &tp);
656 	if (error)
657 		goto out;
658 
659 	xfs_ilock(ip, XFS_ILOCK_EXCL);
660 	xfs_trans_ijoin(tp, ip, 0);
661 
662 	/* Scrape out the old CoW reservations */
663 	error = xfs_reflink_cancel_cow_blocks(ip, &tp, offset_fsb, end_fsb,
664 			cancel_real);
665 	if (error)
666 		goto out_cancel;
667 
668 	error = xfs_trans_commit(tp);
669 
670 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
671 	return error;
672 
673 out_cancel:
674 	xfs_trans_cancel(tp);
675 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
676 out:
677 	trace_xfs_reflink_cancel_cow_range_error(ip, error, _RET_IP_);
678 	return error;
679 }
680 
681 /*
682  * Remap parts of a file's data fork after a successful CoW.
683  */
684 int
685 xfs_reflink_end_cow(
686 	struct xfs_inode		*ip,
687 	xfs_off_t			offset,
688 	xfs_off_t			count)
689 {
690 	struct xfs_ifork		*ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
691 	struct xfs_bmbt_irec		got, del;
692 	struct xfs_trans		*tp;
693 	xfs_fileoff_t			offset_fsb;
694 	xfs_fileoff_t			end_fsb;
695 	xfs_fsblock_t			firstfsb;
696 	struct xfs_defer_ops		dfops;
697 	int				error;
698 	unsigned int			resblks;
699 	xfs_filblks_t			rlen;
700 	xfs_extnum_t			idx;
701 
702 	trace_xfs_reflink_end_cow(ip, offset, count);
703 
704 	/* No COW extents?  That's easy! */
705 	if (ifp->if_bytes == 0)
706 		return 0;
707 
708 	offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
709 	end_fsb = XFS_B_TO_FSB(ip->i_mount, offset + count);
710 
711 	/*
712 	 * Start a rolling transaction to switch the mappings.  We're
713 	 * unlikely ever to have to remap 16T worth of single-block
714 	 * extents, so just cap the worst case extent count to 2^32-1.
715 	 * Stick a warning in just in case, and avoid 64-bit division.
716 	 */
717 	BUILD_BUG_ON(MAX_RW_COUNT > UINT_MAX);
718 	if (end_fsb - offset_fsb > UINT_MAX) {
719 		error = -EFSCORRUPTED;
720 		xfs_force_shutdown(ip->i_mount, SHUTDOWN_CORRUPT_INCORE);
721 		ASSERT(0);
722 		goto out;
723 	}
724 	resblks = XFS_NEXTENTADD_SPACE_RES(ip->i_mount,
725 			(unsigned int)(end_fsb - offset_fsb),
726 			XFS_DATA_FORK);
727 	error = xfs_trans_alloc(ip->i_mount, &M_RES(ip->i_mount)->tr_write,
728 			resblks, 0, 0, &tp);
729 	if (error)
730 		goto out;
731 
732 	xfs_ilock(ip, XFS_ILOCK_EXCL);
733 	xfs_trans_ijoin(tp, ip, 0);
734 
735 	/* If there is a hole at end_fsb - 1 go to the previous extent */
736 	if (!xfs_iext_lookup_extent(ip, ifp, end_fsb - 1, &idx, &got) ||
737 	    got.br_startoff > end_fsb) {
738 		ASSERT(idx > 0);
739 		xfs_iext_get_extent(ifp, --idx, &got);
740 	}
741 
742 	/* Walk backwards until we're out of the I/O range... */
743 	while (got.br_startoff + got.br_blockcount > offset_fsb) {
744 		del = got;
745 		xfs_trim_extent(&del, offset_fsb, end_fsb - offset_fsb);
746 
747 		/* Extent delete may have bumped idx forward */
748 		if (!del.br_blockcount) {
749 			idx--;
750 			goto next_extent;
751 		}
752 
753 		ASSERT(!isnullstartblock(got.br_startblock));
754 
755 		/*
756 		 * Don't remap unwritten extents; these are
757 		 * speculatively preallocated CoW extents that have been
758 		 * allocated but have not yet been involved in a write.
759 		 */
760 		if (got.br_state == XFS_EXT_UNWRITTEN) {
761 			idx--;
762 			goto next_extent;
763 		}
764 
765 		/* Unmap the old blocks in the data fork. */
766 		xfs_defer_init(&dfops, &firstfsb);
767 		rlen = del.br_blockcount;
768 		error = __xfs_bunmapi(tp, ip, del.br_startoff, &rlen, 0, 1,
769 				&firstfsb, &dfops);
770 		if (error)
771 			goto out_defer;
772 
773 		/* Trim the extent to whatever got unmapped. */
774 		if (rlen) {
775 			xfs_trim_extent(&del, del.br_startoff + rlen,
776 				del.br_blockcount - rlen);
777 		}
778 		trace_xfs_reflink_cow_remap(ip, &del);
779 
780 		/* Free the CoW orphan record. */
781 		error = xfs_refcount_free_cow_extent(tp->t_mountp, &dfops,
782 				del.br_startblock, del.br_blockcount);
783 		if (error)
784 			goto out_defer;
785 
786 		/* Map the new blocks into the data fork. */
787 		error = xfs_bmap_map_extent(tp->t_mountp, &dfops, ip, &del);
788 		if (error)
789 			goto out_defer;
790 
791 		/* Remove the mapping from the CoW fork. */
792 		xfs_bmap_del_extent_cow(ip, &idx, &got, &del);
793 
794 		error = xfs_defer_finish(&tp, &dfops, ip);
795 		if (error)
796 			goto out_defer;
797 next_extent:
798 		if (!xfs_iext_get_extent(ifp, idx, &got))
799 			break;
800 	}
801 
802 	error = xfs_trans_commit(tp);
803 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
804 	if (error)
805 		goto out;
806 	return 0;
807 
808 out_defer:
809 	xfs_defer_cancel(&dfops);
810 	xfs_trans_cancel(tp);
811 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
812 out:
813 	trace_xfs_reflink_end_cow_error(ip, error, _RET_IP_);
814 	return error;
815 }
816 
817 /*
818  * Free leftover CoW reservations that didn't get cleaned out.
819  */
820 int
821 xfs_reflink_recover_cow(
822 	struct xfs_mount	*mp)
823 {
824 	xfs_agnumber_t		agno;
825 	int			error = 0;
826 
827 	if (!xfs_sb_version_hasreflink(&mp->m_sb))
828 		return 0;
829 
830 	for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
831 		error = xfs_refcount_recover_cow_leftovers(mp, agno);
832 		if (error)
833 			break;
834 	}
835 
836 	return error;
837 }
838 
839 /*
840  * Reflinking (Block) Ranges of Two Files Together
841  *
842  * First, ensure that the reflink flag is set on both inodes.  The flag is an
843  * optimization to avoid unnecessary refcount btree lookups in the write path.
844  *
845  * Now we can iteratively remap the range of extents (and holes) in src to the
846  * corresponding ranges in dest.  Let drange and srange denote the ranges of
847  * logical blocks in dest and src touched by the reflink operation.
848  *
849  * While the length of drange is greater than zero,
850  *    - Read src's bmbt at the start of srange ("imap")
851  *    - If imap doesn't exist, make imap appear to start at the end of srange
852  *      with zero length.
853  *    - If imap starts before srange, advance imap to start at srange.
854  *    - If imap goes beyond srange, truncate imap to end at the end of srange.
855  *    - Punch (imap start - srange start + imap len) blocks from dest at
856  *      offset (drange start).
857  *    - If imap points to a real range of pblks,
858  *         > Increase the refcount of the imap's pblks
859  *         > Map imap's pblks into dest at the offset
860  *           (drange start + imap start - srange start)
861  *    - Advance drange and srange by (imap start - srange start + imap len)
862  *
863  * Finally, if the reflink made dest longer, update both the in-core and
864  * on-disk file sizes.
865  *
866  * ASCII Art Demonstration:
867  *
868  * Let's say we want to reflink this source file:
869  *
870  * ----SSSSSSS-SSSSS----SSSSSS (src file)
871  *   <-------------------->
872  *
873  * into this destination file:
874  *
875  * --DDDDDDDDDDDDDDDDDDD--DDD (dest file)
876  *        <-------------------->
877  * '-' means a hole, and 'S' and 'D' are written blocks in the src and dest.
878  * Observe that the range has different logical offsets in either file.
879  *
880  * Consider that the first extent in the source file doesn't line up with our
881  * reflink range.  Unmapping  and remapping are separate operations, so we can
882  * unmap more blocks from the destination file than we remap.
883  *
884  * ----SSSSSSS-SSSSS----SSSSSS
885  *   <------->
886  * --DDDDD---------DDDDD--DDD
887  *        <------->
888  *
889  * Now remap the source extent into the destination file:
890  *
891  * ----SSSSSSS-SSSSS----SSSSSS
892  *   <------->
893  * --DDDDD--SSSSSSSDDDDD--DDD
894  *        <------->
895  *
896  * Do likewise with the second hole and extent in our range.  Holes in the
897  * unmap range don't affect our operation.
898  *
899  * ----SSSSSSS-SSSSS----SSSSSS
900  *            <---->
901  * --DDDDD--SSSSSSS-SSSSS-DDD
902  *                 <---->
903  *
904  * Finally, unmap and remap part of the third extent.  This will increase the
905  * size of the destination file.
906  *
907  * ----SSSSSSS-SSSSS----SSSSSS
908  *                  <----->
909  * --DDDDD--SSSSSSS-SSSSS----SSS
910  *                       <----->
911  *
912  * Once we update the destination file's i_size, we're done.
913  */
914 
915 /*
916  * Ensure the reflink bit is set in both inodes.
917  */
918 STATIC int
919 xfs_reflink_set_inode_flag(
920 	struct xfs_inode	*src,
921 	struct xfs_inode	*dest)
922 {
923 	struct xfs_mount	*mp = src->i_mount;
924 	int			error;
925 	struct xfs_trans	*tp;
926 
927 	if (xfs_is_reflink_inode(src) && xfs_is_reflink_inode(dest))
928 		return 0;
929 
930 	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
931 	if (error)
932 		goto out_error;
933 
934 	/* Lock both files against IO */
935 	if (src->i_ino == dest->i_ino)
936 		xfs_ilock(src, XFS_ILOCK_EXCL);
937 	else
938 		xfs_lock_two_inodes(src, dest, XFS_ILOCK_EXCL);
939 
940 	if (!xfs_is_reflink_inode(src)) {
941 		trace_xfs_reflink_set_inode_flag(src);
942 		xfs_trans_ijoin(tp, src, XFS_ILOCK_EXCL);
943 		src->i_d.di_flags2 |= XFS_DIFLAG2_REFLINK;
944 		xfs_trans_log_inode(tp, src, XFS_ILOG_CORE);
945 		xfs_ifork_init_cow(src);
946 	} else
947 		xfs_iunlock(src, XFS_ILOCK_EXCL);
948 
949 	if (src->i_ino == dest->i_ino)
950 		goto commit_flags;
951 
952 	if (!xfs_is_reflink_inode(dest)) {
953 		trace_xfs_reflink_set_inode_flag(dest);
954 		xfs_trans_ijoin(tp, dest, XFS_ILOCK_EXCL);
955 		dest->i_d.di_flags2 |= XFS_DIFLAG2_REFLINK;
956 		xfs_trans_log_inode(tp, dest, XFS_ILOG_CORE);
957 		xfs_ifork_init_cow(dest);
958 	} else
959 		xfs_iunlock(dest, XFS_ILOCK_EXCL);
960 
961 commit_flags:
962 	error = xfs_trans_commit(tp);
963 	if (error)
964 		goto out_error;
965 	return error;
966 
967 out_error:
968 	trace_xfs_reflink_set_inode_flag_error(dest, error, _RET_IP_);
969 	return error;
970 }
971 
972 /*
973  * Update destination inode size & cowextsize hint, if necessary.
974  */
975 STATIC int
976 xfs_reflink_update_dest(
977 	struct xfs_inode	*dest,
978 	xfs_off_t		newlen,
979 	xfs_extlen_t		cowextsize,
980 	bool			is_dedupe)
981 {
982 	struct xfs_mount	*mp = dest->i_mount;
983 	struct xfs_trans	*tp;
984 	int			error;
985 
986 	if (is_dedupe && newlen <= i_size_read(VFS_I(dest)) && cowextsize == 0)
987 		return 0;
988 
989 	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
990 	if (error)
991 		goto out_error;
992 
993 	xfs_ilock(dest, XFS_ILOCK_EXCL);
994 	xfs_trans_ijoin(tp, dest, XFS_ILOCK_EXCL);
995 
996 	if (newlen > i_size_read(VFS_I(dest))) {
997 		trace_xfs_reflink_update_inode_size(dest, newlen);
998 		i_size_write(VFS_I(dest), newlen);
999 		dest->i_d.di_size = newlen;
1000 	}
1001 
1002 	if (cowextsize) {
1003 		dest->i_d.di_cowextsize = cowextsize;
1004 		dest->i_d.di_flags2 |= XFS_DIFLAG2_COWEXTSIZE;
1005 	}
1006 
1007 	if (!is_dedupe) {
1008 		xfs_trans_ichgtime(tp, dest,
1009 				   XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1010 	}
1011 	xfs_trans_log_inode(tp, dest, XFS_ILOG_CORE);
1012 
1013 	error = xfs_trans_commit(tp);
1014 	if (error)
1015 		goto out_error;
1016 	return error;
1017 
1018 out_error:
1019 	trace_xfs_reflink_update_inode_size_error(dest, error, _RET_IP_);
1020 	return error;
1021 }
1022 
1023 /*
1024  * Do we have enough reserve in this AG to handle a reflink?  The refcount
1025  * btree already reserved all the space it needs, but the rmap btree can grow
1026  * infinitely, so we won't allow more reflinks when the AG is down to the
1027  * btree reserves.
1028  */
1029 static int
1030 xfs_reflink_ag_has_free_space(
1031 	struct xfs_mount	*mp,
1032 	xfs_agnumber_t		agno)
1033 {
1034 	struct xfs_perag	*pag;
1035 	int			error = 0;
1036 
1037 	if (!xfs_sb_version_hasrmapbt(&mp->m_sb))
1038 		return 0;
1039 
1040 	pag = xfs_perag_get(mp, agno);
1041 	if (xfs_ag_resv_critical(pag, XFS_AG_RESV_AGFL) ||
1042 	    xfs_ag_resv_critical(pag, XFS_AG_RESV_METADATA))
1043 		error = -ENOSPC;
1044 	xfs_perag_put(pag);
1045 	return error;
1046 }
1047 
1048 /*
1049  * Unmap a range of blocks from a file, then map other blocks into the hole.
1050  * The range to unmap is (destoff : destoff + srcioff + irec->br_blockcount).
1051  * The extent irec is mapped into dest at irec->br_startoff.
1052  */
1053 STATIC int
1054 xfs_reflink_remap_extent(
1055 	struct xfs_inode	*ip,
1056 	struct xfs_bmbt_irec	*irec,
1057 	xfs_fileoff_t		destoff,
1058 	xfs_off_t		new_isize)
1059 {
1060 	struct xfs_mount	*mp = ip->i_mount;
1061 	bool			real_extent = xfs_bmap_is_real_extent(irec);
1062 	struct xfs_trans	*tp;
1063 	xfs_fsblock_t		firstfsb;
1064 	unsigned int		resblks;
1065 	struct xfs_defer_ops	dfops;
1066 	struct xfs_bmbt_irec	uirec;
1067 	xfs_filblks_t		rlen;
1068 	xfs_filblks_t		unmap_len;
1069 	xfs_off_t		newlen;
1070 	int			error;
1071 
1072 	unmap_len = irec->br_startoff + irec->br_blockcount - destoff;
1073 	trace_xfs_reflink_punch_range(ip, destoff, unmap_len);
1074 
1075 	/* No reflinking if we're low on space */
1076 	if (real_extent) {
1077 		error = xfs_reflink_ag_has_free_space(mp,
1078 				XFS_FSB_TO_AGNO(mp, irec->br_startblock));
1079 		if (error)
1080 			goto out;
1081 	}
1082 
1083 	/* Start a rolling transaction to switch the mappings */
1084 	resblks = XFS_EXTENTADD_SPACE_RES(ip->i_mount, XFS_DATA_FORK);
1085 	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, resblks, 0, 0, &tp);
1086 	if (error)
1087 		goto out;
1088 
1089 	xfs_ilock(ip, XFS_ILOCK_EXCL);
1090 	xfs_trans_ijoin(tp, ip, 0);
1091 
1092 	/* If we're not just clearing space, then do we have enough quota? */
1093 	if (real_extent) {
1094 		error = xfs_trans_reserve_quota_nblks(tp, ip,
1095 				irec->br_blockcount, 0, XFS_QMOPT_RES_REGBLKS);
1096 		if (error)
1097 			goto out_cancel;
1098 	}
1099 
1100 	trace_xfs_reflink_remap(ip, irec->br_startoff,
1101 				irec->br_blockcount, irec->br_startblock);
1102 
1103 	/* Unmap the old blocks in the data fork. */
1104 	rlen = unmap_len;
1105 	while (rlen) {
1106 		xfs_defer_init(&dfops, &firstfsb);
1107 		error = __xfs_bunmapi(tp, ip, destoff, &rlen, 0, 1,
1108 				&firstfsb, &dfops);
1109 		if (error)
1110 			goto out_defer;
1111 
1112 		/*
1113 		 * Trim the extent to whatever got unmapped.
1114 		 * Remember, bunmapi works backwards.
1115 		 */
1116 		uirec.br_startblock = irec->br_startblock + rlen;
1117 		uirec.br_startoff = irec->br_startoff + rlen;
1118 		uirec.br_blockcount = unmap_len - rlen;
1119 		unmap_len = rlen;
1120 
1121 		/* If this isn't a real mapping, we're done. */
1122 		if (!real_extent || uirec.br_blockcount == 0)
1123 			goto next_extent;
1124 
1125 		trace_xfs_reflink_remap(ip, uirec.br_startoff,
1126 				uirec.br_blockcount, uirec.br_startblock);
1127 
1128 		/* Update the refcount tree */
1129 		error = xfs_refcount_increase_extent(mp, &dfops, &uirec);
1130 		if (error)
1131 			goto out_defer;
1132 
1133 		/* Map the new blocks into the data fork. */
1134 		error = xfs_bmap_map_extent(mp, &dfops, ip, &uirec);
1135 		if (error)
1136 			goto out_defer;
1137 
1138 		/* Update quota accounting. */
1139 		xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_BCOUNT,
1140 				uirec.br_blockcount);
1141 
1142 		/* Update dest isize if needed. */
1143 		newlen = XFS_FSB_TO_B(mp,
1144 				uirec.br_startoff + uirec.br_blockcount);
1145 		newlen = min_t(xfs_off_t, newlen, new_isize);
1146 		if (newlen > i_size_read(VFS_I(ip))) {
1147 			trace_xfs_reflink_update_inode_size(ip, newlen);
1148 			i_size_write(VFS_I(ip), newlen);
1149 			ip->i_d.di_size = newlen;
1150 			xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1151 		}
1152 
1153 next_extent:
1154 		/* Process all the deferred stuff. */
1155 		error = xfs_defer_finish(&tp, &dfops, ip);
1156 		if (error)
1157 			goto out_defer;
1158 	}
1159 
1160 	error = xfs_trans_commit(tp);
1161 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1162 	if (error)
1163 		goto out;
1164 	return 0;
1165 
1166 out_defer:
1167 	xfs_defer_cancel(&dfops);
1168 out_cancel:
1169 	xfs_trans_cancel(tp);
1170 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1171 out:
1172 	trace_xfs_reflink_remap_extent_error(ip, error, _RET_IP_);
1173 	return error;
1174 }
1175 
1176 /*
1177  * Iteratively remap one file's extents (and holes) to another's.
1178  */
1179 STATIC int
1180 xfs_reflink_remap_blocks(
1181 	struct xfs_inode	*src,
1182 	xfs_fileoff_t		srcoff,
1183 	struct xfs_inode	*dest,
1184 	xfs_fileoff_t		destoff,
1185 	xfs_filblks_t		len,
1186 	xfs_off_t		new_isize)
1187 {
1188 	struct xfs_bmbt_irec	imap;
1189 	int			nimaps;
1190 	int			error = 0;
1191 	xfs_filblks_t		range_len;
1192 
1193 	/* drange = (destoff, destoff + len); srange = (srcoff, srcoff + len) */
1194 	while (len) {
1195 		trace_xfs_reflink_remap_blocks_loop(src, srcoff, len,
1196 				dest, destoff);
1197 		/* Read extent from the source file */
1198 		nimaps = 1;
1199 		xfs_ilock(src, XFS_ILOCK_EXCL);
1200 		error = xfs_bmapi_read(src, srcoff, len, &imap, &nimaps, 0);
1201 		xfs_iunlock(src, XFS_ILOCK_EXCL);
1202 		if (error)
1203 			goto err;
1204 		ASSERT(nimaps == 1);
1205 
1206 		trace_xfs_reflink_remap_imap(src, srcoff, len, XFS_IO_OVERWRITE,
1207 				&imap);
1208 
1209 		/* Translate imap into the destination file. */
1210 		range_len = imap.br_startoff + imap.br_blockcount - srcoff;
1211 		imap.br_startoff += destoff - srcoff;
1212 
1213 		/* Clear dest from destoff to the end of imap and map it in. */
1214 		error = xfs_reflink_remap_extent(dest, &imap, destoff,
1215 				new_isize);
1216 		if (error)
1217 			goto err;
1218 
1219 		if (fatal_signal_pending(current)) {
1220 			error = -EINTR;
1221 			goto err;
1222 		}
1223 
1224 		/* Advance drange/srange */
1225 		srcoff += range_len;
1226 		destoff += range_len;
1227 		len -= range_len;
1228 	}
1229 
1230 	return 0;
1231 
1232 err:
1233 	trace_xfs_reflink_remap_blocks_error(dest, error, _RET_IP_);
1234 	return error;
1235 }
1236 
1237 /*
1238  * Link a range of blocks from one file to another.
1239  */
1240 int
1241 xfs_reflink_remap_range(
1242 	struct file		*file_in,
1243 	loff_t			pos_in,
1244 	struct file		*file_out,
1245 	loff_t			pos_out,
1246 	u64			len,
1247 	bool			is_dedupe)
1248 {
1249 	struct inode		*inode_in = file_inode(file_in);
1250 	struct xfs_inode	*src = XFS_I(inode_in);
1251 	struct inode		*inode_out = file_inode(file_out);
1252 	struct xfs_inode	*dest = XFS_I(inode_out);
1253 	struct xfs_mount	*mp = src->i_mount;
1254 	bool			same_inode = (inode_in == inode_out);
1255 	xfs_fileoff_t		sfsbno, dfsbno;
1256 	xfs_filblks_t		fsblen;
1257 	xfs_extlen_t		cowextsize;
1258 	ssize_t			ret;
1259 
1260 	if (!xfs_sb_version_hasreflink(&mp->m_sb))
1261 		return -EOPNOTSUPP;
1262 
1263 	if (XFS_FORCED_SHUTDOWN(mp))
1264 		return -EIO;
1265 
1266 	/* Lock both files against IO */
1267 	lock_two_nondirectories(inode_in, inode_out);
1268 	if (same_inode)
1269 		xfs_ilock(src, XFS_MMAPLOCK_EXCL);
1270 	else
1271 		xfs_lock_two_inodes(src, dest, XFS_MMAPLOCK_EXCL);
1272 
1273 	/* Check file eligibility and prepare for block sharing. */
1274 	ret = -EINVAL;
1275 	/* Don't reflink realtime inodes */
1276 	if (XFS_IS_REALTIME_INODE(src) || XFS_IS_REALTIME_INODE(dest))
1277 		goto out_unlock;
1278 
1279 	/* Don't share DAX file data for now. */
1280 	if (IS_DAX(inode_in) || IS_DAX(inode_out))
1281 		goto out_unlock;
1282 
1283 	ret = vfs_clone_file_prep_inodes(inode_in, pos_in, inode_out, pos_out,
1284 			&len, is_dedupe);
1285 	if (ret <= 0)
1286 		goto out_unlock;
1287 
1288 	trace_xfs_reflink_remap_range(src, pos_in, len, dest, pos_out);
1289 
1290 	/* Set flags and remap blocks. */
1291 	ret = xfs_reflink_set_inode_flag(src, dest);
1292 	if (ret)
1293 		goto out_unlock;
1294 
1295 	dfsbno = XFS_B_TO_FSBT(mp, pos_out);
1296 	sfsbno = XFS_B_TO_FSBT(mp, pos_in);
1297 	fsblen = XFS_B_TO_FSB(mp, len);
1298 	ret = xfs_reflink_remap_blocks(src, sfsbno, dest, dfsbno, fsblen,
1299 			pos_out + len);
1300 	if (ret)
1301 		goto out_unlock;
1302 
1303 	/* Zap any page cache for the destination file's range. */
1304 	truncate_inode_pages_range(&inode_out->i_data, pos_out,
1305 				   PAGE_ALIGN(pos_out + len) - 1);
1306 
1307 	/*
1308 	 * Carry the cowextsize hint from src to dest if we're sharing the
1309 	 * entire source file to the entire destination file, the source file
1310 	 * has a cowextsize hint, and the destination file does not.
1311 	 */
1312 	cowextsize = 0;
1313 	if (pos_in == 0 && len == i_size_read(inode_in) &&
1314 	    (src->i_d.di_flags2 & XFS_DIFLAG2_COWEXTSIZE) &&
1315 	    pos_out == 0 && len >= i_size_read(inode_out) &&
1316 	    !(dest->i_d.di_flags2 & XFS_DIFLAG2_COWEXTSIZE))
1317 		cowextsize = src->i_d.di_cowextsize;
1318 
1319 	ret = xfs_reflink_update_dest(dest, pos_out + len, cowextsize,
1320 			is_dedupe);
1321 
1322 out_unlock:
1323 	xfs_iunlock(src, XFS_MMAPLOCK_EXCL);
1324 	if (!same_inode)
1325 		xfs_iunlock(dest, XFS_MMAPLOCK_EXCL);
1326 	unlock_two_nondirectories(inode_in, inode_out);
1327 	if (ret)
1328 		trace_xfs_reflink_remap_range_error(dest, ret, _RET_IP_);
1329 	return ret;
1330 }
1331 
1332 /*
1333  * The user wants to preemptively CoW all shared blocks in this file,
1334  * which enables us to turn off the reflink flag.  Iterate all
1335  * extents which are not prealloc/delalloc to see which ranges are
1336  * mentioned in the refcount tree, then read those blocks into the
1337  * pagecache, dirty them, fsync them back out, and then we can update
1338  * the inode flag.  What happens if we run out of memory? :)
1339  */
1340 STATIC int
1341 xfs_reflink_dirty_extents(
1342 	struct xfs_inode	*ip,
1343 	xfs_fileoff_t		fbno,
1344 	xfs_filblks_t		end,
1345 	xfs_off_t		isize)
1346 {
1347 	struct xfs_mount	*mp = ip->i_mount;
1348 	xfs_agnumber_t		agno;
1349 	xfs_agblock_t		agbno;
1350 	xfs_extlen_t		aglen;
1351 	xfs_agblock_t		rbno;
1352 	xfs_extlen_t		rlen;
1353 	xfs_off_t		fpos;
1354 	xfs_off_t		flen;
1355 	struct xfs_bmbt_irec	map[2];
1356 	int			nmaps;
1357 	int			error = 0;
1358 
1359 	while (end - fbno > 0) {
1360 		nmaps = 1;
1361 		/*
1362 		 * Look for extents in the file.  Skip holes, delalloc, or
1363 		 * unwritten extents; they can't be reflinked.
1364 		 */
1365 		error = xfs_bmapi_read(ip, fbno, end - fbno, map, &nmaps, 0);
1366 		if (error)
1367 			goto out;
1368 		if (nmaps == 0)
1369 			break;
1370 		if (!xfs_bmap_is_real_extent(&map[0]))
1371 			goto next;
1372 
1373 		map[1] = map[0];
1374 		while (map[1].br_blockcount) {
1375 			agno = XFS_FSB_TO_AGNO(mp, map[1].br_startblock);
1376 			agbno = XFS_FSB_TO_AGBNO(mp, map[1].br_startblock);
1377 			aglen = map[1].br_blockcount;
1378 
1379 			error = xfs_reflink_find_shared(mp, NULL, agno, agbno,
1380 					aglen, &rbno, &rlen, true);
1381 			if (error)
1382 				goto out;
1383 			if (rbno == NULLAGBLOCK)
1384 				break;
1385 
1386 			/* Dirty the pages */
1387 			xfs_iunlock(ip, XFS_ILOCK_EXCL);
1388 			fpos = XFS_FSB_TO_B(mp, map[1].br_startoff +
1389 					(rbno - agbno));
1390 			flen = XFS_FSB_TO_B(mp, rlen);
1391 			if (fpos + flen > isize)
1392 				flen = isize - fpos;
1393 			error = iomap_file_dirty(VFS_I(ip), fpos, flen,
1394 					&xfs_iomap_ops);
1395 			xfs_ilock(ip, XFS_ILOCK_EXCL);
1396 			if (error)
1397 				goto out;
1398 
1399 			map[1].br_blockcount -= (rbno - agbno + rlen);
1400 			map[1].br_startoff += (rbno - agbno + rlen);
1401 			map[1].br_startblock += (rbno - agbno + rlen);
1402 		}
1403 
1404 next:
1405 		fbno = map[0].br_startoff + map[0].br_blockcount;
1406 	}
1407 out:
1408 	return error;
1409 }
1410 
1411 /* Does this inode need the reflink flag? */
1412 int
1413 xfs_reflink_inode_has_shared_extents(
1414 	struct xfs_trans		*tp,
1415 	struct xfs_inode		*ip,
1416 	bool				*has_shared)
1417 {
1418 	struct xfs_bmbt_irec		got;
1419 	struct xfs_mount		*mp = ip->i_mount;
1420 	struct xfs_ifork		*ifp;
1421 	xfs_agnumber_t			agno;
1422 	xfs_agblock_t			agbno;
1423 	xfs_extlen_t			aglen;
1424 	xfs_agblock_t			rbno;
1425 	xfs_extlen_t			rlen;
1426 	xfs_extnum_t			idx;
1427 	bool				found;
1428 	int				error;
1429 
1430 	ifp = XFS_IFORK_PTR(ip, XFS_DATA_FORK);
1431 	if (!(ifp->if_flags & XFS_IFEXTENTS)) {
1432 		error = xfs_iread_extents(tp, ip, XFS_DATA_FORK);
1433 		if (error)
1434 			return error;
1435 	}
1436 
1437 	*has_shared = false;
1438 	found = xfs_iext_lookup_extent(ip, ifp, 0, &idx, &got);
1439 	while (found) {
1440 		if (isnullstartblock(got.br_startblock) ||
1441 		    got.br_state != XFS_EXT_NORM)
1442 			goto next;
1443 		agno = XFS_FSB_TO_AGNO(mp, got.br_startblock);
1444 		agbno = XFS_FSB_TO_AGBNO(mp, got.br_startblock);
1445 		aglen = got.br_blockcount;
1446 
1447 		error = xfs_reflink_find_shared(mp, tp, agno, agbno, aglen,
1448 				&rbno, &rlen, false);
1449 		if (error)
1450 			return error;
1451 		/* Is there still a shared block here? */
1452 		if (rbno != NULLAGBLOCK) {
1453 			*has_shared = true;
1454 			return 0;
1455 		}
1456 next:
1457 		found = xfs_iext_get_extent(ifp, ++idx, &got);
1458 	}
1459 
1460 	return 0;
1461 }
1462 
1463 /* Clear the inode reflink flag if there are no shared extents. */
1464 int
1465 xfs_reflink_clear_inode_flag(
1466 	struct xfs_inode	*ip,
1467 	struct xfs_trans	**tpp)
1468 {
1469 	bool			needs_flag;
1470 	int			error = 0;
1471 
1472 	ASSERT(xfs_is_reflink_inode(ip));
1473 
1474 	error = xfs_reflink_inode_has_shared_extents(*tpp, ip, &needs_flag);
1475 	if (error || needs_flag)
1476 		return error;
1477 
1478 	/*
1479 	 * We didn't find any shared blocks so turn off the reflink flag.
1480 	 * First, get rid of any leftover CoW mappings.
1481 	 */
1482 	error = xfs_reflink_cancel_cow_blocks(ip, tpp, 0, NULLFILEOFF, true);
1483 	if (error)
1484 		return error;
1485 
1486 	/* Clear the inode flag. */
1487 	trace_xfs_reflink_unset_inode_flag(ip);
1488 	ip->i_d.di_flags2 &= ~XFS_DIFLAG2_REFLINK;
1489 	xfs_inode_clear_cowblocks_tag(ip);
1490 	xfs_trans_ijoin(*tpp, ip, 0);
1491 	xfs_trans_log_inode(*tpp, ip, XFS_ILOG_CORE);
1492 
1493 	return error;
1494 }
1495 
1496 /*
1497  * Clear the inode reflink flag if there are no shared extents and the size
1498  * hasn't changed.
1499  */
1500 STATIC int
1501 xfs_reflink_try_clear_inode_flag(
1502 	struct xfs_inode	*ip)
1503 {
1504 	struct xfs_mount	*mp = ip->i_mount;
1505 	struct xfs_trans	*tp;
1506 	int			error = 0;
1507 
1508 	/* Start a rolling transaction to remove the mappings */
1509 	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, 0, 0, 0, &tp);
1510 	if (error)
1511 		return error;
1512 
1513 	xfs_ilock(ip, XFS_ILOCK_EXCL);
1514 	xfs_trans_ijoin(tp, ip, 0);
1515 
1516 	error = xfs_reflink_clear_inode_flag(ip, &tp);
1517 	if (error)
1518 		goto cancel;
1519 
1520 	error = xfs_trans_commit(tp);
1521 	if (error)
1522 		goto out;
1523 
1524 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1525 	return 0;
1526 cancel:
1527 	xfs_trans_cancel(tp);
1528 out:
1529 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1530 	return error;
1531 }
1532 
1533 /*
1534  * Pre-COW all shared blocks within a given byte range of a file and turn off
1535  * the reflink flag if we unshare all of the file's blocks.
1536  */
1537 int
1538 xfs_reflink_unshare(
1539 	struct xfs_inode	*ip,
1540 	xfs_off_t		offset,
1541 	xfs_off_t		len)
1542 {
1543 	struct xfs_mount	*mp = ip->i_mount;
1544 	xfs_fileoff_t		fbno;
1545 	xfs_filblks_t		end;
1546 	xfs_off_t		isize;
1547 	int			error;
1548 
1549 	if (!xfs_is_reflink_inode(ip))
1550 		return 0;
1551 
1552 	trace_xfs_reflink_unshare(ip, offset, len);
1553 
1554 	inode_dio_wait(VFS_I(ip));
1555 
1556 	/* Try to CoW the selected ranges */
1557 	xfs_ilock(ip, XFS_ILOCK_EXCL);
1558 	fbno = XFS_B_TO_FSBT(mp, offset);
1559 	isize = i_size_read(VFS_I(ip));
1560 	end = XFS_B_TO_FSB(mp, offset + len);
1561 	error = xfs_reflink_dirty_extents(ip, fbno, end, isize);
1562 	if (error)
1563 		goto out_unlock;
1564 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1565 
1566 	/* Wait for the IO to finish */
1567 	error = filemap_write_and_wait(VFS_I(ip)->i_mapping);
1568 	if (error)
1569 		goto out;
1570 
1571 	/* Turn off the reflink flag if possible. */
1572 	error = xfs_reflink_try_clear_inode_flag(ip);
1573 	if (error)
1574 		goto out;
1575 
1576 	return 0;
1577 
1578 out_unlock:
1579 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1580 out:
1581 	trace_xfs_reflink_unshare_error(ip, error, _RET_IP_);
1582 	return error;
1583 }
1584