xref: /openbmc/linux/fs/xfs/xfs_reflink.c (revision 3d37ef41)
1 // SPDX-License-Identifier: GPL-2.0+
2 /*
3  * Copyright (C) 2016 Oracle.  All Rights Reserved.
4  * Author: Darrick J. Wong <darrick.wong@oracle.com>
5  */
6 #include "xfs.h"
7 #include "xfs_fs.h"
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_inode.h"
15 #include "xfs_trans.h"
16 #include "xfs_bmap.h"
17 #include "xfs_bmap_util.h"
18 #include "xfs_trace.h"
19 #include "xfs_icache.h"
20 #include "xfs_btree.h"
21 #include "xfs_refcount_btree.h"
22 #include "xfs_refcount.h"
23 #include "xfs_bmap_btree.h"
24 #include "xfs_trans_space.h"
25 #include "xfs_bit.h"
26 #include "xfs_alloc.h"
27 #include "xfs_quota.h"
28 #include "xfs_reflink.h"
29 #include "xfs_iomap.h"
30 #include "xfs_sb.h"
31 #include "xfs_ag_resv.h"
32 
33 /*
34  * Copy on Write of Shared Blocks
35  *
36  * XFS must preserve "the usual" file semantics even when two files share
37  * the same physical blocks.  This means that a write to one file must not
38  * alter the blocks in a different file; the way that we'll do that is
39  * through the use of a copy-on-write mechanism.  At a high level, that
40  * means that when we want to write to a shared block, we allocate a new
41  * block, write the data to the new block, and if that succeeds we map the
42  * new block into the file.
43  *
44  * XFS provides a "delayed allocation" mechanism that defers the allocation
45  * of disk blocks to dirty-but-not-yet-mapped file blocks as long as
46  * possible.  This reduces fragmentation by enabling the filesystem to ask
47  * for bigger chunks less often, which is exactly what we want for CoW.
48  *
49  * The delalloc mechanism begins when the kernel wants to make a block
50  * writable (write_begin or page_mkwrite).  If the offset is not mapped, we
51  * create a delalloc mapping, which is a regular in-core extent, but without
52  * a real startblock.  (For delalloc mappings, the startblock encodes both
53  * a flag that this is a delalloc mapping, and a worst-case estimate of how
54  * many blocks might be required to put the mapping into the BMBT.)  delalloc
55  * mappings are a reservation against the free space in the filesystem;
56  * adjacent mappings can also be combined into fewer larger mappings.
57  *
58  * As an optimization, the CoW extent size hint (cowextsz) creates
59  * outsized aligned delalloc reservations in the hope of landing out of
60  * order nearby CoW writes in a single extent on disk, thereby reducing
61  * fragmentation and improving future performance.
62  *
63  * D: --RRRRRRSSSRRRRRRRR--- (data fork)
64  * C: ------DDDDDDD--------- (CoW fork)
65  *
66  * When dirty pages are being written out (typically in writepage), the
67  * delalloc reservations are converted into unwritten mappings by
68  * allocating blocks and replacing the delalloc mapping with real ones.
69  * A delalloc mapping can be replaced by several unwritten ones if the
70  * free space is fragmented.
71  *
72  * D: --RRRRRRSSSRRRRRRRR---
73  * C: ------UUUUUUU---------
74  *
75  * We want to adapt the delalloc mechanism for copy-on-write, since the
76  * write paths are similar.  The first two steps (creating the reservation
77  * and allocating the blocks) are exactly the same as delalloc except that
78  * the mappings must be stored in a separate CoW fork because we do not want
79  * to disturb the mapping in the data fork until we're sure that the write
80  * succeeded.  IO completion in this case is the process of removing the old
81  * mapping from the data fork and moving the new mapping from the CoW fork to
82  * the data fork.  This will be discussed shortly.
83  *
84  * For now, unaligned directio writes will be bounced back to the page cache.
85  * Block-aligned directio writes will use the same mechanism as buffered
86  * writes.
87  *
88  * Just prior to submitting the actual disk write requests, we convert
89  * the extents representing the range of the file actually being written
90  * (as opposed to extra pieces created for the cowextsize hint) to real
91  * extents.  This will become important in the next step:
92  *
93  * D: --RRRRRRSSSRRRRRRRR---
94  * C: ------UUrrUUU---------
95  *
96  * CoW remapping must be done after the data block write completes,
97  * because we don't want to destroy the old data fork map until we're sure
98  * the new block has been written.  Since the new mappings are kept in a
99  * separate fork, we can simply iterate these mappings to find the ones
100  * that cover the file blocks that we just CoW'd.  For each extent, simply
101  * unmap the corresponding range in the data fork, map the new range into
102  * the data fork, and remove the extent from the CoW fork.  Because of
103  * the presence of the cowextsize hint, however, we must be careful
104  * only to remap the blocks that we've actually written out --  we must
105  * never remap delalloc reservations nor CoW staging blocks that have
106  * yet to be written.  This corresponds exactly to the real extents in
107  * the CoW fork:
108  *
109  * D: --RRRRRRrrSRRRRRRRR---
110  * C: ------UU--UUU---------
111  *
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
118  * ioend, the better.
119  */
120 
121 /*
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.
127  */
128 int
129 xfs_reflink_find_shared(
130 	struct xfs_mount	*mp,
131 	struct xfs_trans	*tp,
132 	xfs_agnumber_t		agno,
133 	xfs_agblock_t		agbno,
134 	xfs_extlen_t		aglen,
135 	xfs_agblock_t		*fbno,
136 	xfs_extlen_t		*flen,
137 	bool			find_end_of_shared)
138 {
139 	struct xfs_buf		*agbp;
140 	struct xfs_btree_cur	*cur;
141 	int			error;
142 
143 	error = xfs_alloc_read_agf(mp, tp, agno, 0, &agbp);
144 	if (error)
145 		return error;
146 
147 	cur = xfs_refcountbt_init_cursor(mp, tp, agbp, agno);
148 
149 	error = xfs_refcount_find_shared(cur, agbno, aglen, fbno, flen,
150 			find_end_of_shared);
151 
152 	xfs_btree_del_cursor(cur, error);
153 
154 	xfs_trans_brelse(tp, agbp);
155 	return error;
156 }
157 
158 /*
159  * Trim the mapping to the next block where there's a change in the
160  * shared/unshared status.  More specifically, this means that we
161  * find the lowest-numbered extent of shared blocks that coincides with
162  * the given block mapping.  If the shared extent overlaps the start of
163  * the mapping, trim the mapping to the end of the shared extent.  If
164  * the shared region intersects the mapping, trim the mapping to the
165  * start of the shared extent.  If there are no shared regions that
166  * overlap, just return the original extent.
167  */
168 int
169 xfs_reflink_trim_around_shared(
170 	struct xfs_inode	*ip,
171 	struct xfs_bmbt_irec	*irec,
172 	bool			*shared)
173 {
174 	xfs_agnumber_t		agno;
175 	xfs_agblock_t		agbno;
176 	xfs_extlen_t		aglen;
177 	xfs_agblock_t		fbno;
178 	xfs_extlen_t		flen;
179 	int			error = 0;
180 
181 	/* Holes, unwritten, and delalloc extents cannot be shared */
182 	if (!xfs_is_cow_inode(ip) || !xfs_bmap_is_written_extent(irec)) {
183 		*shared = false;
184 		return 0;
185 	}
186 
187 	trace_xfs_reflink_trim_around_shared(ip, irec);
188 
189 	agno = XFS_FSB_TO_AGNO(ip->i_mount, irec->br_startblock);
190 	agbno = XFS_FSB_TO_AGBNO(ip->i_mount, irec->br_startblock);
191 	aglen = irec->br_blockcount;
192 
193 	error = xfs_reflink_find_shared(ip->i_mount, NULL, agno, agbno,
194 			aglen, &fbno, &flen, true);
195 	if (error)
196 		return error;
197 
198 	*shared = false;
199 	if (fbno == NULLAGBLOCK) {
200 		/* No shared blocks at all. */
201 		return 0;
202 	} else if (fbno == agbno) {
203 		/*
204 		 * The start of this extent is shared.  Truncate the
205 		 * mapping at the end of the shared region so that a
206 		 * subsequent iteration starts at the start of the
207 		 * unshared region.
208 		 */
209 		irec->br_blockcount = flen;
210 		*shared = true;
211 		return 0;
212 	} else {
213 		/*
214 		 * There's a shared extent midway through this extent.
215 		 * Truncate the mapping at the start of the shared
216 		 * extent so that a subsequent iteration starts at the
217 		 * start of the shared region.
218 		 */
219 		irec->br_blockcount = fbno - agbno;
220 		return 0;
221 	}
222 }
223 
224 int
225 xfs_bmap_trim_cow(
226 	struct xfs_inode	*ip,
227 	struct xfs_bmbt_irec	*imap,
228 	bool			*shared)
229 {
230 	/* We can't update any real extents in always COW mode. */
231 	if (xfs_is_always_cow_inode(ip) &&
232 	    !isnullstartblock(imap->br_startblock)) {
233 		*shared = true;
234 		return 0;
235 	}
236 
237 	/* Trim the mapping to the nearest shared extent boundary. */
238 	return xfs_reflink_trim_around_shared(ip, imap, shared);
239 }
240 
241 static int
242 xfs_reflink_convert_cow_locked(
243 	struct xfs_inode	*ip,
244 	xfs_fileoff_t		offset_fsb,
245 	xfs_filblks_t		count_fsb)
246 {
247 	struct xfs_iext_cursor	icur;
248 	struct xfs_bmbt_irec	got;
249 	struct xfs_btree_cur	*dummy_cur = NULL;
250 	int			dummy_logflags;
251 	int			error = 0;
252 
253 	if (!xfs_iext_lookup_extent(ip, ip->i_cowfp, offset_fsb, &icur, &got))
254 		return 0;
255 
256 	do {
257 		if (got.br_startoff >= offset_fsb + count_fsb)
258 			break;
259 		if (got.br_state == XFS_EXT_NORM)
260 			continue;
261 		if (WARN_ON_ONCE(isnullstartblock(got.br_startblock)))
262 			return -EIO;
263 
264 		xfs_trim_extent(&got, offset_fsb, count_fsb);
265 		if (!got.br_blockcount)
266 			continue;
267 
268 		got.br_state = XFS_EXT_NORM;
269 		error = xfs_bmap_add_extent_unwritten_real(NULL, ip,
270 				XFS_COW_FORK, &icur, &dummy_cur, &got,
271 				&dummy_logflags);
272 		if (error)
273 			return error;
274 	} while (xfs_iext_next_extent(ip->i_cowfp, &icur, &got));
275 
276 	return error;
277 }
278 
279 /* Convert all of the unwritten CoW extents in a file's range to real ones. */
280 int
281 xfs_reflink_convert_cow(
282 	struct xfs_inode	*ip,
283 	xfs_off_t		offset,
284 	xfs_off_t		count)
285 {
286 	struct xfs_mount	*mp = ip->i_mount;
287 	xfs_fileoff_t		offset_fsb = XFS_B_TO_FSBT(mp, offset);
288 	xfs_fileoff_t		end_fsb = XFS_B_TO_FSB(mp, offset + count);
289 	xfs_filblks_t		count_fsb = end_fsb - offset_fsb;
290 	int			error;
291 
292 	ASSERT(count != 0);
293 
294 	xfs_ilock(ip, XFS_ILOCK_EXCL);
295 	error = xfs_reflink_convert_cow_locked(ip, offset_fsb, count_fsb);
296 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
297 	return error;
298 }
299 
300 /*
301  * Find the extent that maps the given range in the COW fork. Even if the extent
302  * is not shared we might have a preallocation for it in the COW fork. If so we
303  * use it that rather than trigger a new allocation.
304  */
305 static int
306 xfs_find_trim_cow_extent(
307 	struct xfs_inode	*ip,
308 	struct xfs_bmbt_irec	*imap,
309 	struct xfs_bmbt_irec	*cmap,
310 	bool			*shared,
311 	bool			*found)
312 {
313 	xfs_fileoff_t		offset_fsb = imap->br_startoff;
314 	xfs_filblks_t		count_fsb = imap->br_blockcount;
315 	struct xfs_iext_cursor	icur;
316 
317 	*found = false;
318 
319 	/*
320 	 * If we don't find an overlapping extent, trim the range we need to
321 	 * allocate to fit the hole we found.
322 	 */
323 	if (!xfs_iext_lookup_extent(ip, ip->i_cowfp, offset_fsb, &icur, cmap))
324 		cmap->br_startoff = offset_fsb + count_fsb;
325 	if (cmap->br_startoff > offset_fsb) {
326 		xfs_trim_extent(imap, imap->br_startoff,
327 				cmap->br_startoff - imap->br_startoff);
328 		return xfs_bmap_trim_cow(ip, imap, shared);
329 	}
330 
331 	*shared = true;
332 	if (isnullstartblock(cmap->br_startblock)) {
333 		xfs_trim_extent(imap, cmap->br_startoff, cmap->br_blockcount);
334 		return 0;
335 	}
336 
337 	/* real extent found - no need to allocate */
338 	xfs_trim_extent(cmap, offset_fsb, count_fsb);
339 	*found = true;
340 	return 0;
341 }
342 
343 /* Allocate all CoW reservations covering a range of blocks in a file. */
344 int
345 xfs_reflink_allocate_cow(
346 	struct xfs_inode	*ip,
347 	struct xfs_bmbt_irec	*imap,
348 	struct xfs_bmbt_irec	*cmap,
349 	bool			*shared,
350 	uint			*lockmode,
351 	bool			convert_now)
352 {
353 	struct xfs_mount	*mp = ip->i_mount;
354 	xfs_fileoff_t		offset_fsb = imap->br_startoff;
355 	xfs_filblks_t		count_fsb = imap->br_blockcount;
356 	struct xfs_trans	*tp;
357 	int			nimaps, error = 0;
358 	bool			found;
359 	xfs_filblks_t		resaligned;
360 	xfs_extlen_t		resblks = 0;
361 
362 	ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
363 	if (!ip->i_cowfp) {
364 		ASSERT(!xfs_is_reflink_inode(ip));
365 		xfs_ifork_init_cow(ip);
366 	}
367 
368 	error = xfs_find_trim_cow_extent(ip, imap, cmap, shared, &found);
369 	if (error || !*shared)
370 		return error;
371 	if (found)
372 		goto convert;
373 
374 	resaligned = xfs_aligned_fsb_count(imap->br_startoff,
375 		imap->br_blockcount, xfs_get_cowextsz_hint(ip));
376 	resblks = XFS_DIOSTRAT_SPACE_RES(mp, resaligned);
377 
378 	xfs_iunlock(ip, *lockmode);
379 	*lockmode = 0;
380 
381 	error = xfs_trans_alloc_inode(ip, &M_RES(mp)->tr_write, resblks, 0,
382 			false, &tp);
383 	if (error)
384 		return error;
385 
386 	*lockmode = XFS_ILOCK_EXCL;
387 
388 	/*
389 	 * Check for an overlapping extent again now that we dropped the ilock.
390 	 */
391 	error = xfs_find_trim_cow_extent(ip, imap, cmap, shared, &found);
392 	if (error || !*shared)
393 		goto out_trans_cancel;
394 	if (found) {
395 		xfs_trans_cancel(tp);
396 		goto convert;
397 	}
398 
399 	/* Allocate the entire reservation as unwritten blocks. */
400 	nimaps = 1;
401 	error = xfs_bmapi_write(tp, ip, imap->br_startoff, imap->br_blockcount,
402 			XFS_BMAPI_COWFORK | XFS_BMAPI_PREALLOC, 0, cmap,
403 			&nimaps);
404 	if (error)
405 		goto out_trans_cancel;
406 
407 	xfs_inode_set_cowblocks_tag(ip);
408 	error = xfs_trans_commit(tp);
409 	if (error)
410 		return error;
411 
412 	/*
413 	 * Allocation succeeded but the requested range was not even partially
414 	 * satisfied?  Bail out!
415 	 */
416 	if (nimaps == 0)
417 		return -ENOSPC;
418 convert:
419 	xfs_trim_extent(cmap, offset_fsb, count_fsb);
420 	/*
421 	 * COW fork extents are supposed to remain unwritten until we're ready
422 	 * to initiate a disk write.  For direct I/O we are going to write the
423 	 * data and need the conversion, but for buffered writes we're done.
424 	 */
425 	if (!convert_now || cmap->br_state == XFS_EXT_NORM)
426 		return 0;
427 	trace_xfs_reflink_convert_cow(ip, cmap);
428 	return xfs_reflink_convert_cow_locked(ip, offset_fsb, count_fsb);
429 
430 out_trans_cancel:
431 	xfs_trans_cancel(tp);
432 	return error;
433 }
434 
435 /*
436  * Cancel CoW reservations for some block range of an inode.
437  *
438  * If cancel_real is true this function cancels all COW fork extents for the
439  * inode; if cancel_real is false, real extents are not cleared.
440  *
441  * Caller must have already joined the inode to the current transaction. The
442  * inode will be joined to the transaction returned to the caller.
443  */
444 int
445 xfs_reflink_cancel_cow_blocks(
446 	struct xfs_inode		*ip,
447 	struct xfs_trans		**tpp,
448 	xfs_fileoff_t			offset_fsb,
449 	xfs_fileoff_t			end_fsb,
450 	bool				cancel_real)
451 {
452 	struct xfs_ifork		*ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
453 	struct xfs_bmbt_irec		got, del;
454 	struct xfs_iext_cursor		icur;
455 	int				error = 0;
456 
457 	if (!xfs_inode_has_cow_data(ip))
458 		return 0;
459 	if (!xfs_iext_lookup_extent_before(ip, ifp, &end_fsb, &icur, &got))
460 		return 0;
461 
462 	/* Walk backwards until we're out of the I/O range... */
463 	while (got.br_startoff + got.br_blockcount > offset_fsb) {
464 		del = got;
465 		xfs_trim_extent(&del, offset_fsb, end_fsb - offset_fsb);
466 
467 		/* Extent delete may have bumped ext forward */
468 		if (!del.br_blockcount) {
469 			xfs_iext_prev(ifp, &icur);
470 			goto next_extent;
471 		}
472 
473 		trace_xfs_reflink_cancel_cow(ip, &del);
474 
475 		if (isnullstartblock(del.br_startblock)) {
476 			error = xfs_bmap_del_extent_delay(ip, XFS_COW_FORK,
477 					&icur, &got, &del);
478 			if (error)
479 				break;
480 		} else if (del.br_state == XFS_EXT_UNWRITTEN || cancel_real) {
481 			ASSERT((*tpp)->t_firstblock == NULLFSBLOCK);
482 
483 			/* Free the CoW orphan record. */
484 			xfs_refcount_free_cow_extent(*tpp, del.br_startblock,
485 					del.br_blockcount);
486 
487 			xfs_bmap_add_free(*tpp, del.br_startblock,
488 					  del.br_blockcount, NULL);
489 
490 			/* Roll the transaction */
491 			error = xfs_defer_finish(tpp);
492 			if (error)
493 				break;
494 
495 			/* Remove the mapping from the CoW fork. */
496 			xfs_bmap_del_extent_cow(ip, &icur, &got, &del);
497 
498 			/* Remove the quota reservation */
499 			error = xfs_quota_unreserve_blkres(ip,
500 					del.br_blockcount);
501 			if (error)
502 				break;
503 		} else {
504 			/* Didn't do anything, push cursor back. */
505 			xfs_iext_prev(ifp, &icur);
506 		}
507 next_extent:
508 		if (!xfs_iext_get_extent(ifp, &icur, &got))
509 			break;
510 	}
511 
512 	/* clear tag if cow fork is emptied */
513 	if (!ifp->if_bytes)
514 		xfs_inode_clear_cowblocks_tag(ip);
515 	return error;
516 }
517 
518 /*
519  * Cancel CoW reservations for some byte range of an inode.
520  *
521  * If cancel_real is true this function cancels all COW fork extents for the
522  * inode; if cancel_real is false, real extents are not cleared.
523  */
524 int
525 xfs_reflink_cancel_cow_range(
526 	struct xfs_inode	*ip,
527 	xfs_off_t		offset,
528 	xfs_off_t		count,
529 	bool			cancel_real)
530 {
531 	struct xfs_trans	*tp;
532 	xfs_fileoff_t		offset_fsb;
533 	xfs_fileoff_t		end_fsb;
534 	int			error;
535 
536 	trace_xfs_reflink_cancel_cow_range(ip, offset, count);
537 	ASSERT(ip->i_cowfp);
538 
539 	offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
540 	if (count == NULLFILEOFF)
541 		end_fsb = NULLFILEOFF;
542 	else
543 		end_fsb = XFS_B_TO_FSB(ip->i_mount, offset + count);
544 
545 	/* Start a rolling transaction to remove the mappings */
546 	error = xfs_trans_alloc(ip->i_mount, &M_RES(ip->i_mount)->tr_write,
547 			0, 0, 0, &tp);
548 	if (error)
549 		goto out;
550 
551 	xfs_ilock(ip, XFS_ILOCK_EXCL);
552 	xfs_trans_ijoin(tp, ip, 0);
553 
554 	/* Scrape out the old CoW reservations */
555 	error = xfs_reflink_cancel_cow_blocks(ip, &tp, offset_fsb, end_fsb,
556 			cancel_real);
557 	if (error)
558 		goto out_cancel;
559 
560 	error = xfs_trans_commit(tp);
561 
562 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
563 	return error;
564 
565 out_cancel:
566 	xfs_trans_cancel(tp);
567 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
568 out:
569 	trace_xfs_reflink_cancel_cow_range_error(ip, error, _RET_IP_);
570 	return error;
571 }
572 
573 /*
574  * Remap part of the CoW fork into the data fork.
575  *
576  * We aim to remap the range starting at @offset_fsb and ending at @end_fsb
577  * into the data fork; this function will remap what it can (at the end of the
578  * range) and update @end_fsb appropriately.  Each remap gets its own
579  * transaction because we can end up merging and splitting bmbt blocks for
580  * every remap operation and we'd like to keep the block reservation
581  * requirements as low as possible.
582  */
583 STATIC int
584 xfs_reflink_end_cow_extent(
585 	struct xfs_inode	*ip,
586 	xfs_fileoff_t		offset_fsb,
587 	xfs_fileoff_t		*end_fsb)
588 {
589 	struct xfs_bmbt_irec	got, del;
590 	struct xfs_iext_cursor	icur;
591 	struct xfs_mount	*mp = ip->i_mount;
592 	struct xfs_trans	*tp;
593 	struct xfs_ifork	*ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
594 	xfs_filblks_t		rlen;
595 	unsigned int		resblks;
596 	int			error;
597 
598 	/* No COW extents?  That's easy! */
599 	if (ifp->if_bytes == 0) {
600 		*end_fsb = offset_fsb;
601 		return 0;
602 	}
603 
604 	resblks = XFS_EXTENTADD_SPACE_RES(mp, XFS_DATA_FORK);
605 	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, resblks, 0,
606 			XFS_TRANS_RESERVE, &tp);
607 	if (error)
608 		return error;
609 
610 	/*
611 	 * Lock the inode.  We have to ijoin without automatic unlock because
612 	 * the lead transaction is the refcountbt record deletion; the data
613 	 * fork update follows as a deferred log item.
614 	 */
615 	xfs_ilock(ip, XFS_ILOCK_EXCL);
616 	xfs_trans_ijoin(tp, ip, 0);
617 
618 	error = xfs_iext_count_may_overflow(ip, XFS_DATA_FORK,
619 			XFS_IEXT_REFLINK_END_COW_CNT);
620 	if (error)
621 		goto out_cancel;
622 
623 	/*
624 	 * In case of racing, overlapping AIO writes no COW extents might be
625 	 * left by the time I/O completes for the loser of the race.  In that
626 	 * case we are done.
627 	 */
628 	if (!xfs_iext_lookup_extent_before(ip, ifp, end_fsb, &icur, &got) ||
629 	    got.br_startoff + got.br_blockcount <= offset_fsb) {
630 		*end_fsb = offset_fsb;
631 		goto out_cancel;
632 	}
633 
634 	/*
635 	 * Structure copy @got into @del, then trim @del to the range that we
636 	 * were asked to remap.  We preserve @got for the eventual CoW fork
637 	 * deletion; from now on @del represents the mapping that we're
638 	 * actually remapping.
639 	 */
640 	del = got;
641 	xfs_trim_extent(&del, offset_fsb, *end_fsb - offset_fsb);
642 
643 	ASSERT(del.br_blockcount > 0);
644 
645 	/*
646 	 * Only remap real extents that contain data.  With AIO, speculative
647 	 * preallocations can leak into the range we are called upon, and we
648 	 * need to skip them.
649 	 */
650 	if (!xfs_bmap_is_written_extent(&got)) {
651 		*end_fsb = del.br_startoff;
652 		goto out_cancel;
653 	}
654 
655 	/* Unmap the old blocks in the data fork. */
656 	rlen = del.br_blockcount;
657 	error = __xfs_bunmapi(tp, ip, del.br_startoff, &rlen, 0, 1);
658 	if (error)
659 		goto out_cancel;
660 
661 	/* Trim the extent to whatever got unmapped. */
662 	xfs_trim_extent(&del, del.br_startoff + rlen, del.br_blockcount - rlen);
663 	trace_xfs_reflink_cow_remap(ip, &del);
664 
665 	/* Free the CoW orphan record. */
666 	xfs_refcount_free_cow_extent(tp, del.br_startblock, del.br_blockcount);
667 
668 	/* Map the new blocks into the data fork. */
669 	xfs_bmap_map_extent(tp, ip, &del);
670 
671 	/* Charge this new data fork mapping to the on-disk quota. */
672 	xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_DELBCOUNT,
673 			(long)del.br_blockcount);
674 
675 	/* Remove the mapping from the CoW fork. */
676 	xfs_bmap_del_extent_cow(ip, &icur, &got, &del);
677 
678 	error = xfs_trans_commit(tp);
679 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
680 	if (error)
681 		return error;
682 
683 	/* Update the caller about how much progress we made. */
684 	*end_fsb = del.br_startoff;
685 	return 0;
686 
687 out_cancel:
688 	xfs_trans_cancel(tp);
689 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
690 	return error;
691 }
692 
693 /*
694  * Remap parts of a file's data fork after a successful CoW.
695  */
696 int
697 xfs_reflink_end_cow(
698 	struct xfs_inode		*ip,
699 	xfs_off_t			offset,
700 	xfs_off_t			count)
701 {
702 	xfs_fileoff_t			offset_fsb;
703 	xfs_fileoff_t			end_fsb;
704 	int				error = 0;
705 
706 	trace_xfs_reflink_end_cow(ip, offset, count);
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 	 * Walk backwards until we're out of the I/O range.  The loop function
713 	 * repeatedly cycles the ILOCK to allocate one transaction per remapped
714 	 * extent.
715 	 *
716 	 * If we're being called by writeback then the pages will still
717 	 * have PageWriteback set, which prevents races with reflink remapping
718 	 * and truncate.  Reflink remapping prevents races with writeback by
719 	 * taking the iolock and mmaplock before flushing the pages and
720 	 * remapping, which means there won't be any further writeback or page
721 	 * cache dirtying until the reflink completes.
722 	 *
723 	 * We should never have two threads issuing writeback for the same file
724 	 * region.  There are also have post-eof checks in the writeback
725 	 * preparation code so that we don't bother writing out pages that are
726 	 * about to be truncated.
727 	 *
728 	 * If we're being called as part of directio write completion, the dio
729 	 * count is still elevated, which reflink and truncate will wait for.
730 	 * Reflink remapping takes the iolock and mmaplock and waits for
731 	 * pending dio to finish, which should prevent any directio until the
732 	 * remap completes.  Multiple concurrent directio writes to the same
733 	 * region are handled by end_cow processing only occurring for the
734 	 * threads which succeed; the outcome of multiple overlapping direct
735 	 * writes is not well defined anyway.
736 	 *
737 	 * It's possible that a buffered write and a direct write could collide
738 	 * here (the buffered write stumbles in after the dio flushes and
739 	 * invalidates the page cache and immediately queues writeback), but we
740 	 * have never supported this 100%.  If either disk write succeeds the
741 	 * blocks will be remapped.
742 	 */
743 	while (end_fsb > offset_fsb && !error)
744 		error = xfs_reflink_end_cow_extent(ip, offset_fsb, &end_fsb);
745 
746 	if (error)
747 		trace_xfs_reflink_end_cow_error(ip, error, _RET_IP_);
748 	return error;
749 }
750 
751 /*
752  * Free leftover CoW reservations that didn't get cleaned out.
753  */
754 int
755 xfs_reflink_recover_cow(
756 	struct xfs_mount	*mp)
757 {
758 	xfs_agnumber_t		agno;
759 	int			error = 0;
760 
761 	if (!xfs_sb_version_hasreflink(&mp->m_sb))
762 		return 0;
763 
764 	for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
765 		error = xfs_refcount_recover_cow_leftovers(mp, agno);
766 		if (error)
767 			break;
768 	}
769 
770 	return error;
771 }
772 
773 /*
774  * Reflinking (Block) Ranges of Two Files Together
775  *
776  * First, ensure that the reflink flag is set on both inodes.  The flag is an
777  * optimization to avoid unnecessary refcount btree lookups in the write path.
778  *
779  * Now we can iteratively remap the range of extents (and holes) in src to the
780  * corresponding ranges in dest.  Let drange and srange denote the ranges of
781  * logical blocks in dest and src touched by the reflink operation.
782  *
783  * While the length of drange is greater than zero,
784  *    - Read src's bmbt at the start of srange ("imap")
785  *    - If imap doesn't exist, make imap appear to start at the end of srange
786  *      with zero length.
787  *    - If imap starts before srange, advance imap to start at srange.
788  *    - If imap goes beyond srange, truncate imap to end at the end of srange.
789  *    - Punch (imap start - srange start + imap len) blocks from dest at
790  *      offset (drange start).
791  *    - If imap points to a real range of pblks,
792  *         > Increase the refcount of the imap's pblks
793  *         > Map imap's pblks into dest at the offset
794  *           (drange start + imap start - srange start)
795  *    - Advance drange and srange by (imap start - srange start + imap len)
796  *
797  * Finally, if the reflink made dest longer, update both the in-core and
798  * on-disk file sizes.
799  *
800  * ASCII Art Demonstration:
801  *
802  * Let's say we want to reflink this source file:
803  *
804  * ----SSSSSSS-SSSSS----SSSSSS (src file)
805  *   <-------------------->
806  *
807  * into this destination file:
808  *
809  * --DDDDDDDDDDDDDDDDDDD--DDD (dest file)
810  *        <-------------------->
811  * '-' means a hole, and 'S' and 'D' are written blocks in the src and dest.
812  * Observe that the range has different logical offsets in either file.
813  *
814  * Consider that the first extent in the source file doesn't line up with our
815  * reflink range.  Unmapping  and remapping are separate operations, so we can
816  * unmap more blocks from the destination file than we remap.
817  *
818  * ----SSSSSSS-SSSSS----SSSSSS
819  *   <------->
820  * --DDDDD---------DDDDD--DDD
821  *        <------->
822  *
823  * Now remap the source extent into the destination file:
824  *
825  * ----SSSSSSS-SSSSS----SSSSSS
826  *   <------->
827  * --DDDDD--SSSSSSSDDDDD--DDD
828  *        <------->
829  *
830  * Do likewise with the second hole and extent in our range.  Holes in the
831  * unmap range don't affect our operation.
832  *
833  * ----SSSSSSS-SSSSS----SSSSSS
834  *            <---->
835  * --DDDDD--SSSSSSS-SSSSS-DDD
836  *                 <---->
837  *
838  * Finally, unmap and remap part of the third extent.  This will increase the
839  * size of the destination file.
840  *
841  * ----SSSSSSS-SSSSS----SSSSSS
842  *                  <----->
843  * --DDDDD--SSSSSSS-SSSSS----SSS
844  *                       <----->
845  *
846  * Once we update the destination file's i_size, we're done.
847  */
848 
849 /*
850  * Ensure the reflink bit is set in both inodes.
851  */
852 STATIC int
853 xfs_reflink_set_inode_flag(
854 	struct xfs_inode	*src,
855 	struct xfs_inode	*dest)
856 {
857 	struct xfs_mount	*mp = src->i_mount;
858 	int			error;
859 	struct xfs_trans	*tp;
860 
861 	if (xfs_is_reflink_inode(src) && xfs_is_reflink_inode(dest))
862 		return 0;
863 
864 	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
865 	if (error)
866 		goto out_error;
867 
868 	/* Lock both files against IO */
869 	if (src->i_ino == dest->i_ino)
870 		xfs_ilock(src, XFS_ILOCK_EXCL);
871 	else
872 		xfs_lock_two_inodes(src, XFS_ILOCK_EXCL, dest, XFS_ILOCK_EXCL);
873 
874 	if (!xfs_is_reflink_inode(src)) {
875 		trace_xfs_reflink_set_inode_flag(src);
876 		xfs_trans_ijoin(tp, src, XFS_ILOCK_EXCL);
877 		src->i_diflags2 |= XFS_DIFLAG2_REFLINK;
878 		xfs_trans_log_inode(tp, src, XFS_ILOG_CORE);
879 		xfs_ifork_init_cow(src);
880 	} else
881 		xfs_iunlock(src, XFS_ILOCK_EXCL);
882 
883 	if (src->i_ino == dest->i_ino)
884 		goto commit_flags;
885 
886 	if (!xfs_is_reflink_inode(dest)) {
887 		trace_xfs_reflink_set_inode_flag(dest);
888 		xfs_trans_ijoin(tp, dest, XFS_ILOCK_EXCL);
889 		dest->i_diflags2 |= XFS_DIFLAG2_REFLINK;
890 		xfs_trans_log_inode(tp, dest, XFS_ILOG_CORE);
891 		xfs_ifork_init_cow(dest);
892 	} else
893 		xfs_iunlock(dest, XFS_ILOCK_EXCL);
894 
895 commit_flags:
896 	error = xfs_trans_commit(tp);
897 	if (error)
898 		goto out_error;
899 	return error;
900 
901 out_error:
902 	trace_xfs_reflink_set_inode_flag_error(dest, error, _RET_IP_);
903 	return error;
904 }
905 
906 /*
907  * Update destination inode size & cowextsize hint, if necessary.
908  */
909 int
910 xfs_reflink_update_dest(
911 	struct xfs_inode	*dest,
912 	xfs_off_t		newlen,
913 	xfs_extlen_t		cowextsize,
914 	unsigned int		remap_flags)
915 {
916 	struct xfs_mount	*mp = dest->i_mount;
917 	struct xfs_trans	*tp;
918 	int			error;
919 
920 	if (newlen <= i_size_read(VFS_I(dest)) && cowextsize == 0)
921 		return 0;
922 
923 	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
924 	if (error)
925 		goto out_error;
926 
927 	xfs_ilock(dest, XFS_ILOCK_EXCL);
928 	xfs_trans_ijoin(tp, dest, XFS_ILOCK_EXCL);
929 
930 	if (newlen > i_size_read(VFS_I(dest))) {
931 		trace_xfs_reflink_update_inode_size(dest, newlen);
932 		i_size_write(VFS_I(dest), newlen);
933 		dest->i_disk_size = newlen;
934 	}
935 
936 	if (cowextsize) {
937 		dest->i_cowextsize = cowextsize;
938 		dest->i_diflags2 |= XFS_DIFLAG2_COWEXTSIZE;
939 	}
940 
941 	xfs_trans_log_inode(tp, dest, XFS_ILOG_CORE);
942 
943 	error = xfs_trans_commit(tp);
944 	if (error)
945 		goto out_error;
946 	return error;
947 
948 out_error:
949 	trace_xfs_reflink_update_inode_size_error(dest, error, _RET_IP_);
950 	return error;
951 }
952 
953 /*
954  * Do we have enough reserve in this AG to handle a reflink?  The refcount
955  * btree already reserved all the space it needs, but the rmap btree can grow
956  * infinitely, so we won't allow more reflinks when the AG is down to the
957  * btree reserves.
958  */
959 static int
960 xfs_reflink_ag_has_free_space(
961 	struct xfs_mount	*mp,
962 	xfs_agnumber_t		agno)
963 {
964 	struct xfs_perag	*pag;
965 	int			error = 0;
966 
967 	if (!xfs_sb_version_hasrmapbt(&mp->m_sb))
968 		return 0;
969 
970 	pag = xfs_perag_get(mp, agno);
971 	if (xfs_ag_resv_critical(pag, XFS_AG_RESV_RMAPBT) ||
972 	    xfs_ag_resv_critical(pag, XFS_AG_RESV_METADATA))
973 		error = -ENOSPC;
974 	xfs_perag_put(pag);
975 	return error;
976 }
977 
978 /*
979  * Remap the given extent into the file.  The dmap blockcount will be set to
980  * the number of blocks that were actually remapped.
981  */
982 STATIC int
983 xfs_reflink_remap_extent(
984 	struct xfs_inode	*ip,
985 	struct xfs_bmbt_irec	*dmap,
986 	xfs_off_t		new_isize)
987 {
988 	struct xfs_bmbt_irec	smap;
989 	struct xfs_mount	*mp = ip->i_mount;
990 	struct xfs_trans	*tp;
991 	xfs_off_t		newlen;
992 	int64_t			qdelta = 0;
993 	unsigned int		resblks;
994 	bool			quota_reserved = true;
995 	bool			smap_real;
996 	bool			dmap_written = xfs_bmap_is_written_extent(dmap);
997 	int			iext_delta = 0;
998 	int			nimaps;
999 	int			error;
1000 
1001 	/*
1002 	 * Start a rolling transaction to switch the mappings.
1003 	 *
1004 	 * Adding a written extent to the extent map can cause a bmbt split,
1005 	 * and removing a mapped extent from the extent can cause a bmbt split.
1006 	 * The two operations cannot both cause a split since they operate on
1007 	 * the same index in the bmap btree, so we only need a reservation for
1008 	 * one bmbt split if either thing is happening.  However, we haven't
1009 	 * locked the inode yet, so we reserve assuming this is the case.
1010 	 *
1011 	 * The first allocation call tries to reserve enough space to handle
1012 	 * mapping dmap into a sparse part of the file plus the bmbt split.  We
1013 	 * haven't locked the inode or read the existing mapping yet, so we do
1014 	 * not know for sure that we need the space.  This should succeed most
1015 	 * of the time.
1016 	 *
1017 	 * If the first attempt fails, try again but reserving only enough
1018 	 * space to handle a bmbt split.  This is the hard minimum requirement,
1019 	 * and we revisit quota reservations later when we know more about what
1020 	 * we're remapping.
1021 	 */
1022 	resblks = XFS_EXTENTADD_SPACE_RES(mp, XFS_DATA_FORK);
1023 	error = xfs_trans_alloc_inode(ip, &M_RES(mp)->tr_write,
1024 			resblks + dmap->br_blockcount, 0, false, &tp);
1025 	if (error == -EDQUOT || error == -ENOSPC) {
1026 		quota_reserved = false;
1027 		error = xfs_trans_alloc_inode(ip, &M_RES(mp)->tr_write,
1028 				resblks, 0, false, &tp);
1029 	}
1030 	if (error)
1031 		goto out;
1032 
1033 	/*
1034 	 * Read what's currently mapped in the destination file into smap.
1035 	 * If smap isn't a hole, we will have to remove it before we can add
1036 	 * dmap to the destination file.
1037 	 */
1038 	nimaps = 1;
1039 	error = xfs_bmapi_read(ip, dmap->br_startoff, dmap->br_blockcount,
1040 			&smap, &nimaps, 0);
1041 	if (error)
1042 		goto out_cancel;
1043 	ASSERT(nimaps == 1 && smap.br_startoff == dmap->br_startoff);
1044 	smap_real = xfs_bmap_is_real_extent(&smap);
1045 
1046 	/*
1047 	 * We can only remap as many blocks as the smaller of the two extent
1048 	 * maps, because we can only remap one extent at a time.
1049 	 */
1050 	dmap->br_blockcount = min(dmap->br_blockcount, smap.br_blockcount);
1051 	ASSERT(dmap->br_blockcount == smap.br_blockcount);
1052 
1053 	trace_xfs_reflink_remap_extent_dest(ip, &smap);
1054 
1055 	/*
1056 	 * Two extents mapped to the same physical block must not have
1057 	 * different states; that's filesystem corruption.  Move on to the next
1058 	 * extent if they're both holes or both the same physical extent.
1059 	 */
1060 	if (dmap->br_startblock == smap.br_startblock) {
1061 		if (dmap->br_state != smap.br_state)
1062 			error = -EFSCORRUPTED;
1063 		goto out_cancel;
1064 	}
1065 
1066 	/* If both extents are unwritten, leave them alone. */
1067 	if (dmap->br_state == XFS_EXT_UNWRITTEN &&
1068 	    smap.br_state == XFS_EXT_UNWRITTEN)
1069 		goto out_cancel;
1070 
1071 	/* No reflinking if the AG of the dest mapping is low on space. */
1072 	if (dmap_written) {
1073 		error = xfs_reflink_ag_has_free_space(mp,
1074 				XFS_FSB_TO_AGNO(mp, dmap->br_startblock));
1075 		if (error)
1076 			goto out_cancel;
1077 	}
1078 
1079 	/*
1080 	 * Increase quota reservation if we think the quota block counter for
1081 	 * this file could increase.
1082 	 *
1083 	 * If we are mapping a written extent into the file, we need to have
1084 	 * enough quota block count reservation to handle the blocks in that
1085 	 * extent.  We log only the delta to the quota block counts, so if the
1086 	 * extent we're unmapping also has blocks allocated to it, we don't
1087 	 * need a quota reservation for the extent itself.
1088 	 *
1089 	 * Note that if we're replacing a delalloc reservation with a written
1090 	 * extent, we have to take the full quota reservation because removing
1091 	 * the delalloc reservation gives the block count back to the quota
1092 	 * count.  This is suboptimal, but the VFS flushed the dest range
1093 	 * before we started.  That should have removed all the delalloc
1094 	 * reservations, but we code defensively.
1095 	 *
1096 	 * xfs_trans_alloc_inode above already tried to grab an even larger
1097 	 * quota reservation, and kicked off a blockgc scan if it couldn't.
1098 	 * If we can't get a potentially smaller quota reservation now, we're
1099 	 * done.
1100 	 */
1101 	if (!quota_reserved && !smap_real && dmap_written) {
1102 		error = xfs_trans_reserve_quota_nblks(tp, ip,
1103 				dmap->br_blockcount, 0, false);
1104 		if (error)
1105 			goto out_cancel;
1106 	}
1107 
1108 	if (smap_real)
1109 		++iext_delta;
1110 
1111 	if (dmap_written)
1112 		++iext_delta;
1113 
1114 	error = xfs_iext_count_may_overflow(ip, XFS_DATA_FORK, iext_delta);
1115 	if (error)
1116 		goto out_cancel;
1117 
1118 	if (smap_real) {
1119 		/*
1120 		 * If the extent we're unmapping is backed by storage (written
1121 		 * or not), unmap the extent and drop its refcount.
1122 		 */
1123 		xfs_bmap_unmap_extent(tp, ip, &smap);
1124 		xfs_refcount_decrease_extent(tp, &smap);
1125 		qdelta -= smap.br_blockcount;
1126 	} else if (smap.br_startblock == DELAYSTARTBLOCK) {
1127 		xfs_filblks_t	len = smap.br_blockcount;
1128 
1129 		/*
1130 		 * If the extent we're unmapping is a delalloc reservation,
1131 		 * we can use the regular bunmapi function to release the
1132 		 * incore state.  Dropping the delalloc reservation takes care
1133 		 * of the quota reservation for us.
1134 		 */
1135 		error = __xfs_bunmapi(NULL, ip, smap.br_startoff, &len, 0, 1);
1136 		if (error)
1137 			goto out_cancel;
1138 		ASSERT(len == 0);
1139 	}
1140 
1141 	/*
1142 	 * If the extent we're sharing is backed by written storage, increase
1143 	 * its refcount and map it into the file.
1144 	 */
1145 	if (dmap_written) {
1146 		xfs_refcount_increase_extent(tp, dmap);
1147 		xfs_bmap_map_extent(tp, ip, dmap);
1148 		qdelta += dmap->br_blockcount;
1149 	}
1150 
1151 	xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_BCOUNT, qdelta);
1152 
1153 	/* Update dest isize if needed. */
1154 	newlen = XFS_FSB_TO_B(mp, dmap->br_startoff + dmap->br_blockcount);
1155 	newlen = min_t(xfs_off_t, newlen, new_isize);
1156 	if (newlen > i_size_read(VFS_I(ip))) {
1157 		trace_xfs_reflink_update_inode_size(ip, newlen);
1158 		i_size_write(VFS_I(ip), newlen);
1159 		ip->i_disk_size = newlen;
1160 		xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1161 	}
1162 
1163 	/* Commit everything and unlock. */
1164 	error = xfs_trans_commit(tp);
1165 	goto out_unlock;
1166 
1167 out_cancel:
1168 	xfs_trans_cancel(tp);
1169 out_unlock:
1170 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1171 out:
1172 	if (error)
1173 		trace_xfs_reflink_remap_extent_error(ip, error, _RET_IP_);
1174 	return error;
1175 }
1176 
1177 /* Remap a range of one file to the other. */
1178 int
1179 xfs_reflink_remap_blocks(
1180 	struct xfs_inode	*src,
1181 	loff_t			pos_in,
1182 	struct xfs_inode	*dest,
1183 	loff_t			pos_out,
1184 	loff_t			remap_len,
1185 	loff_t			*remapped)
1186 {
1187 	struct xfs_bmbt_irec	imap;
1188 	struct xfs_mount	*mp = src->i_mount;
1189 	xfs_fileoff_t		srcoff = XFS_B_TO_FSBT(mp, pos_in);
1190 	xfs_fileoff_t		destoff = XFS_B_TO_FSBT(mp, pos_out);
1191 	xfs_filblks_t		len;
1192 	xfs_filblks_t		remapped_len = 0;
1193 	xfs_off_t		new_isize = pos_out + remap_len;
1194 	int			nimaps;
1195 	int			error = 0;
1196 
1197 	len = min_t(xfs_filblks_t, XFS_B_TO_FSB(mp, remap_len),
1198 			XFS_MAX_FILEOFF);
1199 
1200 	trace_xfs_reflink_remap_blocks(src, srcoff, len, dest, destoff);
1201 
1202 	while (len > 0) {
1203 		unsigned int	lock_mode;
1204 
1205 		/* Read extent from the source file */
1206 		nimaps = 1;
1207 		lock_mode = xfs_ilock_data_map_shared(src);
1208 		error = xfs_bmapi_read(src, srcoff, len, &imap, &nimaps, 0);
1209 		xfs_iunlock(src, lock_mode);
1210 		if (error)
1211 			break;
1212 		/*
1213 		 * The caller supposedly flushed all dirty pages in the source
1214 		 * file range, which means that writeback should have allocated
1215 		 * or deleted all delalloc reservations in that range.  If we
1216 		 * find one, that's a good sign that something is seriously
1217 		 * wrong here.
1218 		 */
1219 		ASSERT(nimaps == 1 && imap.br_startoff == srcoff);
1220 		if (imap.br_startblock == DELAYSTARTBLOCK) {
1221 			ASSERT(imap.br_startblock != DELAYSTARTBLOCK);
1222 			error = -EFSCORRUPTED;
1223 			break;
1224 		}
1225 
1226 		trace_xfs_reflink_remap_extent_src(src, &imap);
1227 
1228 		/* Remap into the destination file at the given offset. */
1229 		imap.br_startoff = destoff;
1230 		error = xfs_reflink_remap_extent(dest, &imap, new_isize);
1231 		if (error)
1232 			break;
1233 
1234 		if (fatal_signal_pending(current)) {
1235 			error = -EINTR;
1236 			break;
1237 		}
1238 
1239 		/* Advance drange/srange */
1240 		srcoff += imap.br_blockcount;
1241 		destoff += imap.br_blockcount;
1242 		len -= imap.br_blockcount;
1243 		remapped_len += imap.br_blockcount;
1244 	}
1245 
1246 	if (error)
1247 		trace_xfs_reflink_remap_blocks_error(dest, error, _RET_IP_);
1248 	*remapped = min_t(loff_t, remap_len,
1249 			  XFS_FSB_TO_B(src->i_mount, remapped_len));
1250 	return error;
1251 }
1252 
1253 /*
1254  * If we're reflinking to a point past the destination file's EOF, we must
1255  * zero any speculative post-EOF preallocations that sit between the old EOF
1256  * and the destination file offset.
1257  */
1258 static int
1259 xfs_reflink_zero_posteof(
1260 	struct xfs_inode	*ip,
1261 	loff_t			pos)
1262 {
1263 	loff_t			isize = i_size_read(VFS_I(ip));
1264 
1265 	if (pos <= isize)
1266 		return 0;
1267 
1268 	trace_xfs_zero_eof(ip, isize, pos - isize);
1269 	return iomap_zero_range(VFS_I(ip), isize, pos - isize, NULL,
1270 			&xfs_buffered_write_iomap_ops);
1271 }
1272 
1273 /*
1274  * Prepare two files for range cloning.  Upon a successful return both inodes
1275  * will have the iolock and mmaplock held, the page cache of the out file will
1276  * be truncated, and any leases on the out file will have been broken.  This
1277  * function borrows heavily from xfs_file_aio_write_checks.
1278  *
1279  * The VFS allows partial EOF blocks to "match" for dedupe even though it hasn't
1280  * checked that the bytes beyond EOF physically match. Hence we cannot use the
1281  * EOF block in the source dedupe range because it's not a complete block match,
1282  * hence can introduce a corruption into the file that has it's block replaced.
1283  *
1284  * In similar fashion, the VFS file cloning also allows partial EOF blocks to be
1285  * "block aligned" for the purposes of cloning entire files.  However, if the
1286  * source file range includes the EOF block and it lands within the existing EOF
1287  * of the destination file, then we can expose stale data from beyond the source
1288  * file EOF in the destination file.
1289  *
1290  * XFS doesn't support partial block sharing, so in both cases we have check
1291  * these cases ourselves. For dedupe, we can simply round the length to dedupe
1292  * down to the previous whole block and ignore the partial EOF block. While this
1293  * means we can't dedupe the last block of a file, this is an acceptible
1294  * tradeoff for simplicity on implementation.
1295  *
1296  * For cloning, we want to share the partial EOF block if it is also the new EOF
1297  * block of the destination file. If the partial EOF block lies inside the
1298  * existing destination EOF, then we have to abort the clone to avoid exposing
1299  * stale data in the destination file. Hence we reject these clone attempts with
1300  * -EINVAL in this case.
1301  */
1302 int
1303 xfs_reflink_remap_prep(
1304 	struct file		*file_in,
1305 	loff_t			pos_in,
1306 	struct file		*file_out,
1307 	loff_t			pos_out,
1308 	loff_t			*len,
1309 	unsigned int		remap_flags)
1310 {
1311 	struct inode		*inode_in = file_inode(file_in);
1312 	struct xfs_inode	*src = XFS_I(inode_in);
1313 	struct inode		*inode_out = file_inode(file_out);
1314 	struct xfs_inode	*dest = XFS_I(inode_out);
1315 	int			ret;
1316 
1317 	/* Lock both files against IO */
1318 	ret = xfs_ilock2_io_mmap(src, dest);
1319 	if (ret)
1320 		return ret;
1321 
1322 	/* Check file eligibility and prepare for block sharing. */
1323 	ret = -EINVAL;
1324 	/* Don't reflink realtime inodes */
1325 	if (XFS_IS_REALTIME_INODE(src) || XFS_IS_REALTIME_INODE(dest))
1326 		goto out_unlock;
1327 
1328 	/* Don't share DAX file data for now. */
1329 	if (IS_DAX(inode_in) || IS_DAX(inode_out))
1330 		goto out_unlock;
1331 
1332 	ret = generic_remap_file_range_prep(file_in, pos_in, file_out, pos_out,
1333 			len, remap_flags);
1334 	if (ret || *len == 0)
1335 		goto out_unlock;
1336 
1337 	/* Attach dquots to dest inode before changing block map */
1338 	ret = xfs_qm_dqattach(dest);
1339 	if (ret)
1340 		goto out_unlock;
1341 
1342 	/*
1343 	 * Zero existing post-eof speculative preallocations in the destination
1344 	 * file.
1345 	 */
1346 	ret = xfs_reflink_zero_posteof(dest, pos_out);
1347 	if (ret)
1348 		goto out_unlock;
1349 
1350 	/* Set flags and remap blocks. */
1351 	ret = xfs_reflink_set_inode_flag(src, dest);
1352 	if (ret)
1353 		goto out_unlock;
1354 
1355 	/*
1356 	 * If pos_out > EOF, we may have dirtied blocks between EOF and
1357 	 * pos_out. In that case, we need to extend the flush and unmap to cover
1358 	 * from EOF to the end of the copy length.
1359 	 */
1360 	if (pos_out > XFS_ISIZE(dest)) {
1361 		loff_t	flen = *len + (pos_out - XFS_ISIZE(dest));
1362 		ret = xfs_flush_unmap_range(dest, XFS_ISIZE(dest), flen);
1363 	} else {
1364 		ret = xfs_flush_unmap_range(dest, pos_out, *len);
1365 	}
1366 	if (ret)
1367 		goto out_unlock;
1368 
1369 	return 0;
1370 out_unlock:
1371 	xfs_iunlock2_io_mmap(src, dest);
1372 	return ret;
1373 }
1374 
1375 /* Does this inode need the reflink flag? */
1376 int
1377 xfs_reflink_inode_has_shared_extents(
1378 	struct xfs_trans		*tp,
1379 	struct xfs_inode		*ip,
1380 	bool				*has_shared)
1381 {
1382 	struct xfs_bmbt_irec		got;
1383 	struct xfs_mount		*mp = ip->i_mount;
1384 	struct xfs_ifork		*ifp;
1385 	xfs_agnumber_t			agno;
1386 	xfs_agblock_t			agbno;
1387 	xfs_extlen_t			aglen;
1388 	xfs_agblock_t			rbno;
1389 	xfs_extlen_t			rlen;
1390 	struct xfs_iext_cursor		icur;
1391 	bool				found;
1392 	int				error;
1393 
1394 	ifp = XFS_IFORK_PTR(ip, XFS_DATA_FORK);
1395 	error = xfs_iread_extents(tp, ip, XFS_DATA_FORK);
1396 	if (error)
1397 		return error;
1398 
1399 	*has_shared = false;
1400 	found = xfs_iext_lookup_extent(ip, ifp, 0, &icur, &got);
1401 	while (found) {
1402 		if (isnullstartblock(got.br_startblock) ||
1403 		    got.br_state != XFS_EXT_NORM)
1404 			goto next;
1405 		agno = XFS_FSB_TO_AGNO(mp, got.br_startblock);
1406 		agbno = XFS_FSB_TO_AGBNO(mp, got.br_startblock);
1407 		aglen = got.br_blockcount;
1408 
1409 		error = xfs_reflink_find_shared(mp, tp, agno, agbno, aglen,
1410 				&rbno, &rlen, false);
1411 		if (error)
1412 			return error;
1413 		/* Is there still a shared block here? */
1414 		if (rbno != NULLAGBLOCK) {
1415 			*has_shared = true;
1416 			return 0;
1417 		}
1418 next:
1419 		found = xfs_iext_next_extent(ifp, &icur, &got);
1420 	}
1421 
1422 	return 0;
1423 }
1424 
1425 /*
1426  * Clear the inode reflink flag if there are no shared extents.
1427  *
1428  * The caller is responsible for joining the inode to the transaction passed in.
1429  * The inode will be joined to the transaction that is returned to the caller.
1430  */
1431 int
1432 xfs_reflink_clear_inode_flag(
1433 	struct xfs_inode	*ip,
1434 	struct xfs_trans	**tpp)
1435 {
1436 	bool			needs_flag;
1437 	int			error = 0;
1438 
1439 	ASSERT(xfs_is_reflink_inode(ip));
1440 
1441 	error = xfs_reflink_inode_has_shared_extents(*tpp, ip, &needs_flag);
1442 	if (error || needs_flag)
1443 		return error;
1444 
1445 	/*
1446 	 * We didn't find any shared blocks so turn off the reflink flag.
1447 	 * First, get rid of any leftover CoW mappings.
1448 	 */
1449 	error = xfs_reflink_cancel_cow_blocks(ip, tpp, 0, XFS_MAX_FILEOFF,
1450 			true);
1451 	if (error)
1452 		return error;
1453 
1454 	/* Clear the inode flag. */
1455 	trace_xfs_reflink_unset_inode_flag(ip);
1456 	ip->i_diflags2 &= ~XFS_DIFLAG2_REFLINK;
1457 	xfs_inode_clear_cowblocks_tag(ip);
1458 	xfs_trans_log_inode(*tpp, ip, XFS_ILOG_CORE);
1459 
1460 	return error;
1461 }
1462 
1463 /*
1464  * Clear the inode reflink flag if there are no shared extents and the size
1465  * hasn't changed.
1466  */
1467 STATIC int
1468 xfs_reflink_try_clear_inode_flag(
1469 	struct xfs_inode	*ip)
1470 {
1471 	struct xfs_mount	*mp = ip->i_mount;
1472 	struct xfs_trans	*tp;
1473 	int			error = 0;
1474 
1475 	/* Start a rolling transaction to remove the mappings */
1476 	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, 0, 0, 0, &tp);
1477 	if (error)
1478 		return error;
1479 
1480 	xfs_ilock(ip, XFS_ILOCK_EXCL);
1481 	xfs_trans_ijoin(tp, ip, 0);
1482 
1483 	error = xfs_reflink_clear_inode_flag(ip, &tp);
1484 	if (error)
1485 		goto cancel;
1486 
1487 	error = xfs_trans_commit(tp);
1488 	if (error)
1489 		goto out;
1490 
1491 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1492 	return 0;
1493 cancel:
1494 	xfs_trans_cancel(tp);
1495 out:
1496 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1497 	return error;
1498 }
1499 
1500 /*
1501  * Pre-COW all shared blocks within a given byte range of a file and turn off
1502  * the reflink flag if we unshare all of the file's blocks.
1503  */
1504 int
1505 xfs_reflink_unshare(
1506 	struct xfs_inode	*ip,
1507 	xfs_off_t		offset,
1508 	xfs_off_t		len)
1509 {
1510 	struct inode		*inode = VFS_I(ip);
1511 	int			error;
1512 
1513 	if (!xfs_is_reflink_inode(ip))
1514 		return 0;
1515 
1516 	trace_xfs_reflink_unshare(ip, offset, len);
1517 
1518 	inode_dio_wait(inode);
1519 
1520 	error = iomap_file_unshare(inode, offset, len,
1521 			&xfs_buffered_write_iomap_ops);
1522 	if (error)
1523 		goto out;
1524 
1525 	error = filemap_write_and_wait_range(inode->i_mapping, offset,
1526 			offset + len - 1);
1527 	if (error)
1528 		goto out;
1529 
1530 	/* Turn off the reflink flag if possible. */
1531 	error = xfs_reflink_try_clear_inode_flag(ip);
1532 	if (error)
1533 		goto out;
1534 	return 0;
1535 
1536 out:
1537 	trace_xfs_reflink_unshare_error(ip, error, _RET_IP_);
1538 	return error;
1539 }
1540