xref: /openbmc/linux/fs/xfs/xfs_reflink.c (revision f220d3eb)
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_da_format.h"
15 #include "xfs_da_btree.h"
16 #include "xfs_inode.h"
17 #include "xfs_trans.h"
18 #include "xfs_inode_item.h"
19 #include "xfs_bmap.h"
20 #include "xfs_bmap_util.h"
21 #include "xfs_error.h"
22 #include "xfs_dir2.h"
23 #include "xfs_dir2_priv.h"
24 #include "xfs_ioctl.h"
25 #include "xfs_trace.h"
26 #include "xfs_log.h"
27 #include "xfs_icache.h"
28 #include "xfs_pnfs.h"
29 #include "xfs_btree.h"
30 #include "xfs_refcount_btree.h"
31 #include "xfs_refcount.h"
32 #include "xfs_bmap_btree.h"
33 #include "xfs_trans_space.h"
34 #include "xfs_bit.h"
35 #include "xfs_alloc.h"
36 #include "xfs_quota_defs.h"
37 #include "xfs_quota.h"
38 #include "xfs_reflink.h"
39 #include "xfs_iomap.h"
40 #include "xfs_rmap_btree.h"
41 #include "xfs_sb.h"
42 #include "xfs_ag_resv.h"
43 
44 /*
45  * Copy on Write of Shared Blocks
46  *
47  * XFS must preserve "the usual" file semantics even when two files share
48  * the same physical blocks.  This means that a write to one file must not
49  * alter the blocks in a different file; the way that we'll do that is
50  * through the use of a copy-on-write mechanism.  At a high level, that
51  * means that when we want to write to a shared block, we allocate a new
52  * block, write the data to the new block, and if that succeeds we map the
53  * new block into the file.
54  *
55  * XFS provides a "delayed allocation" mechanism that defers the allocation
56  * of disk blocks to dirty-but-not-yet-mapped file blocks as long as
57  * possible.  This reduces fragmentation by enabling the filesystem to ask
58  * for bigger chunks less often, which is exactly what we want for CoW.
59  *
60  * The delalloc mechanism begins when the kernel wants to make a block
61  * writable (write_begin or page_mkwrite).  If the offset is not mapped, we
62  * create a delalloc mapping, which is a regular in-core extent, but without
63  * a real startblock.  (For delalloc mappings, the startblock encodes both
64  * a flag that this is a delalloc mapping, and a worst-case estimate of how
65  * many blocks might be required to put the mapping into the BMBT.)  delalloc
66  * mappings are a reservation against the free space in the filesystem;
67  * adjacent mappings can also be combined into fewer larger mappings.
68  *
69  * As an optimization, the CoW extent size hint (cowextsz) creates
70  * outsized aligned delalloc reservations in the hope of landing out of
71  * order nearby CoW writes in a single extent on disk, thereby reducing
72  * fragmentation and improving future performance.
73  *
74  * D: --RRRRRRSSSRRRRRRRR--- (data fork)
75  * C: ------DDDDDDD--------- (CoW fork)
76  *
77  * When dirty pages are being written out (typically in writepage), the
78  * delalloc reservations are converted into unwritten mappings by
79  * allocating blocks and replacing the delalloc mapping with real ones.
80  * A delalloc mapping can be replaced by several unwritten ones if the
81  * free space is fragmented.
82  *
83  * D: --RRRRRRSSSRRRRRRRR---
84  * C: ------UUUUUUU---------
85  *
86  * We want to adapt the delalloc mechanism for copy-on-write, since the
87  * write paths are similar.  The first two steps (creating the reservation
88  * and allocating the blocks) are exactly the same as delalloc except that
89  * the mappings must be stored in a separate CoW fork because we do not want
90  * to disturb the mapping in the data fork until we're sure that the write
91  * succeeded.  IO completion in this case is the process of removing the old
92  * mapping from the data fork and moving the new mapping from the CoW fork to
93  * the data fork.  This will be discussed shortly.
94  *
95  * For now, unaligned directio writes will be bounced back to the page cache.
96  * Block-aligned directio writes will use the same mechanism as buffered
97  * writes.
98  *
99  * Just prior to submitting the actual disk write requests, we convert
100  * the extents representing the range of the file actually being written
101  * (as opposed to extra pieces created for the cowextsize hint) to real
102  * extents.  This will become important in the next step:
103  *
104  * D: --RRRRRRSSSRRRRRRRR---
105  * C: ------UUrrUUU---------
106  *
107  * CoW remapping must be done after the data block write completes,
108  * because we don't want to destroy the old data fork map until we're sure
109  * the new block has been written.  Since the new mappings are kept in a
110  * separate fork, we can simply iterate these mappings to find the ones
111  * that cover the file blocks that we just CoW'd.  For each extent, simply
112  * unmap the corresponding range in the data fork, map the new range into
113  * the data fork, and remove the extent from the CoW fork.  Because of
114  * the presence of the cowextsize hint, however, we must be careful
115  * only to remap the blocks that we've actually written out --  we must
116  * never remap delalloc reservations nor CoW staging blocks that have
117  * yet to be written.  This corresponds exactly to the real extents in
118  * the CoW fork:
119  *
120  * D: --RRRRRRrrSRRRRRRRR---
121  * C: ------UU--UUU---------
122  *
123  * Since the remapping operation can be applied to an arbitrary file
124  * range, we record the need for the remap step as a flag in the ioend
125  * instead of declaring a new IO type.  This is required for direct io
126  * because we only have ioend for the whole dio, and we have to be able to
127  * remember the presence of unwritten blocks and CoW blocks with a single
128  * ioend structure.  Better yet, the more ground we can cover with one
129  * ioend, the better.
130  */
131 
132 /*
133  * Given an AG extent, find the lowest-numbered run of shared blocks
134  * within that range and return the range in fbno/flen.  If
135  * find_end_of_shared is true, return the longest contiguous extent of
136  * shared blocks.  If there are no shared extents, fbno and flen will
137  * be set to NULLAGBLOCK and 0, respectively.
138  */
139 int
140 xfs_reflink_find_shared(
141 	struct xfs_mount	*mp,
142 	struct xfs_trans	*tp,
143 	xfs_agnumber_t		agno,
144 	xfs_agblock_t		agbno,
145 	xfs_extlen_t		aglen,
146 	xfs_agblock_t		*fbno,
147 	xfs_extlen_t		*flen,
148 	bool			find_end_of_shared)
149 {
150 	struct xfs_buf		*agbp;
151 	struct xfs_btree_cur	*cur;
152 	int			error;
153 
154 	error = xfs_alloc_read_agf(mp, tp, agno, 0, &agbp);
155 	if (error)
156 		return error;
157 	if (!agbp)
158 		return -ENOMEM;
159 
160 	cur = xfs_refcountbt_init_cursor(mp, tp, agbp, agno);
161 
162 	error = xfs_refcount_find_shared(cur, agbno, aglen, fbno, flen,
163 			find_end_of_shared);
164 
165 	xfs_btree_del_cursor(cur, error);
166 
167 	xfs_trans_brelse(tp, agbp);
168 	return error;
169 }
170 
171 /*
172  * Trim the mapping to the next block where there's a change in the
173  * shared/unshared status.  More specifically, this means that we
174  * find the lowest-numbered extent of shared blocks that coincides with
175  * the given block mapping.  If the shared extent overlaps the start of
176  * the mapping, trim the mapping to the end of the shared extent.  If
177  * the shared region intersects the mapping, trim the mapping to the
178  * start of the shared extent.  If there are no shared regions that
179  * overlap, just return the original extent.
180  */
181 int
182 xfs_reflink_trim_around_shared(
183 	struct xfs_inode	*ip,
184 	struct xfs_bmbt_irec	*irec,
185 	bool			*shared,
186 	bool			*trimmed)
187 {
188 	xfs_agnumber_t		agno;
189 	xfs_agblock_t		agbno;
190 	xfs_extlen_t		aglen;
191 	xfs_agblock_t		fbno;
192 	xfs_extlen_t		flen;
193 	int			error = 0;
194 
195 	/* Holes, unwritten, and delalloc extents cannot be shared */
196 	if (!xfs_is_reflink_inode(ip) || !xfs_bmap_is_real_extent(irec)) {
197 		*shared = false;
198 		return 0;
199 	}
200 
201 	trace_xfs_reflink_trim_around_shared(ip, irec);
202 
203 	agno = XFS_FSB_TO_AGNO(ip->i_mount, irec->br_startblock);
204 	agbno = XFS_FSB_TO_AGBNO(ip->i_mount, irec->br_startblock);
205 	aglen = irec->br_blockcount;
206 
207 	error = xfs_reflink_find_shared(ip->i_mount, NULL, agno, agbno,
208 			aglen, &fbno, &flen, true);
209 	if (error)
210 		return error;
211 
212 	*shared = *trimmed = false;
213 	if (fbno == NULLAGBLOCK) {
214 		/* No shared blocks at all. */
215 		return 0;
216 	} else if (fbno == agbno) {
217 		/*
218 		 * The start of this extent is shared.  Truncate the
219 		 * mapping at the end of the shared region so that a
220 		 * subsequent iteration starts at the start of the
221 		 * unshared region.
222 		 */
223 		irec->br_blockcount = flen;
224 		*shared = true;
225 		if (flen != aglen)
226 			*trimmed = true;
227 		return 0;
228 	} else {
229 		/*
230 		 * There's a shared extent midway through this extent.
231 		 * Truncate the mapping at the start of the shared
232 		 * extent so that a subsequent iteration starts at the
233 		 * start of the shared region.
234 		 */
235 		irec->br_blockcount = fbno - agbno;
236 		*trimmed = true;
237 		return 0;
238 	}
239 }
240 
241 /*
242  * Trim the passed in imap to the next shared/unshared extent boundary, and
243  * if imap->br_startoff points to a shared extent reserve space for it in the
244  * COW fork.  In this case *shared is set to true, else to false.
245  *
246  * Note that imap will always contain the block numbers for the existing blocks
247  * in the data fork, as the upper layers need them for read-modify-write
248  * operations.
249  */
250 int
251 xfs_reflink_reserve_cow(
252 	struct xfs_inode	*ip,
253 	struct xfs_bmbt_irec	*imap,
254 	bool			*shared)
255 {
256 	struct xfs_ifork	*ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
257 	struct xfs_bmbt_irec	got;
258 	int			error = 0;
259 	bool			eof = false, trimmed;
260 	struct xfs_iext_cursor	icur;
261 
262 	/*
263 	 * Search the COW fork extent list first.  This serves two purposes:
264 	 * first this implement the speculative preallocation using cowextisze,
265 	 * so that we also unshared block adjacent to shared blocks instead
266 	 * of just the shared blocks themselves.  Second the lookup in the
267 	 * extent list is generally faster than going out to the shared extent
268 	 * tree.
269 	 */
270 
271 	if (!xfs_iext_lookup_extent(ip, ifp, imap->br_startoff, &icur, &got))
272 		eof = true;
273 	if (!eof && got.br_startoff <= imap->br_startoff) {
274 		trace_xfs_reflink_cow_found(ip, imap);
275 		xfs_trim_extent(imap, got.br_startoff, got.br_blockcount);
276 
277 		*shared = true;
278 		return 0;
279 	}
280 
281 	/* Trim the mapping to the nearest shared extent boundary. */
282 	error = xfs_reflink_trim_around_shared(ip, imap, shared, &trimmed);
283 	if (error)
284 		return error;
285 
286 	/* Not shared?  Just report the (potentially capped) extent. */
287 	if (!*shared)
288 		return 0;
289 
290 	/*
291 	 * Fork all the shared blocks from our write offset until the end of
292 	 * the extent.
293 	 */
294 	error = xfs_qm_dqattach_locked(ip, false);
295 	if (error)
296 		return error;
297 
298 	error = xfs_bmapi_reserve_delalloc(ip, XFS_COW_FORK, imap->br_startoff,
299 			imap->br_blockcount, 0, &got, &icur, eof);
300 	if (error == -ENOSPC || error == -EDQUOT)
301 		trace_xfs_reflink_cow_enospc(ip, imap);
302 	if (error)
303 		return error;
304 
305 	trace_xfs_reflink_cow_alloc(ip, &got);
306 	return 0;
307 }
308 
309 /* Convert part of an unwritten CoW extent to a real one. */
310 STATIC int
311 xfs_reflink_convert_cow_extent(
312 	struct xfs_inode		*ip,
313 	struct xfs_bmbt_irec		*imap,
314 	xfs_fileoff_t			offset_fsb,
315 	xfs_filblks_t			count_fsb)
316 {
317 	int				nimaps = 1;
318 
319 	if (imap->br_state == XFS_EXT_NORM)
320 		return 0;
321 
322 	xfs_trim_extent(imap, offset_fsb, count_fsb);
323 	trace_xfs_reflink_convert_cow(ip, imap);
324 	if (imap->br_blockcount == 0)
325 		return 0;
326 	return xfs_bmapi_write(NULL, ip, imap->br_startoff, imap->br_blockcount,
327 			XFS_BMAPI_COWFORK | XFS_BMAPI_CONVERT, 0, imap,
328 			&nimaps);
329 }
330 
331 /* Convert all of the unwritten CoW extents in a file's range to real ones. */
332 int
333 xfs_reflink_convert_cow(
334 	struct xfs_inode	*ip,
335 	xfs_off_t		offset,
336 	xfs_off_t		count)
337 {
338 	struct xfs_mount	*mp = ip->i_mount;
339 	xfs_fileoff_t		offset_fsb = XFS_B_TO_FSBT(mp, offset);
340 	xfs_fileoff_t		end_fsb = XFS_B_TO_FSB(mp, offset + count);
341 	xfs_filblks_t		count_fsb = end_fsb - offset_fsb;
342 	struct xfs_bmbt_irec	imap;
343 	int			nimaps = 1, error = 0;
344 
345 	ASSERT(count != 0);
346 
347 	xfs_ilock(ip, XFS_ILOCK_EXCL);
348 	error = xfs_bmapi_write(NULL, ip, offset_fsb, count_fsb,
349 			XFS_BMAPI_COWFORK | XFS_BMAPI_CONVERT |
350 			XFS_BMAPI_CONVERT_ONLY, 0, &imap, &nimaps);
351 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
352 	return error;
353 }
354 
355 /* Allocate all CoW reservations covering a range of blocks in a file. */
356 int
357 xfs_reflink_allocate_cow(
358 	struct xfs_inode	*ip,
359 	struct xfs_bmbt_irec	*imap,
360 	bool			*shared,
361 	uint			*lockmode)
362 {
363 	struct xfs_mount	*mp = ip->i_mount;
364 	xfs_fileoff_t		offset_fsb = imap->br_startoff;
365 	xfs_filblks_t		count_fsb = imap->br_blockcount;
366 	struct xfs_bmbt_irec	got;
367 	struct xfs_trans	*tp = NULL;
368 	int			nimaps, error = 0;
369 	bool			trimmed;
370 	xfs_filblks_t		resaligned;
371 	xfs_extlen_t		resblks = 0;
372 	struct xfs_iext_cursor	icur;
373 
374 retry:
375 	ASSERT(xfs_is_reflink_inode(ip));
376 	ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
377 
378 	/*
379 	 * Even if the extent is not shared we might have a preallocation for
380 	 * it in the COW fork.  If so use it.
381 	 */
382 	if (xfs_iext_lookup_extent(ip, ip->i_cowfp, offset_fsb, &icur, &got) &&
383 	    got.br_startoff <= offset_fsb) {
384 		*shared = true;
385 
386 		/* If we have a real allocation in the COW fork we're done. */
387 		if (!isnullstartblock(got.br_startblock)) {
388 			xfs_trim_extent(&got, offset_fsb, count_fsb);
389 			*imap = got;
390 			goto convert;
391 		}
392 
393 		xfs_trim_extent(imap, got.br_startoff, got.br_blockcount);
394 	} else {
395 		error = xfs_reflink_trim_around_shared(ip, imap, shared, &trimmed);
396 		if (error || !*shared)
397 			goto out;
398 	}
399 
400 	if (!tp) {
401 		resaligned = xfs_aligned_fsb_count(imap->br_startoff,
402 			imap->br_blockcount, xfs_get_cowextsz_hint(ip));
403 		resblks = XFS_DIOSTRAT_SPACE_RES(mp, resaligned);
404 
405 		xfs_iunlock(ip, *lockmode);
406 		error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, resblks, 0, 0, &tp);
407 		*lockmode = XFS_ILOCK_EXCL;
408 		xfs_ilock(ip, *lockmode);
409 
410 		if (error)
411 			return error;
412 
413 		error = xfs_qm_dqattach_locked(ip, false);
414 		if (error)
415 			goto out;
416 		goto retry;
417 	}
418 
419 	error = xfs_trans_reserve_quota_nblks(tp, ip, resblks, 0,
420 			XFS_QMOPT_RES_REGBLKS);
421 	if (error)
422 		goto out;
423 
424 	xfs_trans_ijoin(tp, ip, 0);
425 
426 	nimaps = 1;
427 
428 	/* Allocate the entire reservation as unwritten blocks. */
429 	error = xfs_bmapi_write(tp, ip, imap->br_startoff, imap->br_blockcount,
430 			XFS_BMAPI_COWFORK | XFS_BMAPI_PREALLOC,
431 			resblks, imap, &nimaps);
432 	if (error)
433 		goto out_trans_cancel;
434 
435 	xfs_inode_set_cowblocks_tag(ip);
436 
437 	/* Finish up. */
438 	error = xfs_trans_commit(tp);
439 	if (error)
440 		return error;
441 
442 	/*
443 	 * Allocation succeeded but the requested range was not even partially
444 	 * satisfied?  Bail out!
445 	 */
446 	if (nimaps == 0)
447 		return -ENOSPC;
448 convert:
449 	return xfs_reflink_convert_cow_extent(ip, imap, offset_fsb, count_fsb);
450 out_trans_cancel:
451 	xfs_trans_unreserve_quota_nblks(tp, ip, (long)resblks, 0,
452 			XFS_QMOPT_RES_REGBLKS);
453 out:
454 	if (tp)
455 		xfs_trans_cancel(tp);
456 	return error;
457 }
458 
459 /*
460  * Cancel CoW reservations for some block range of an inode.
461  *
462  * If cancel_real is true this function cancels all COW fork extents for the
463  * inode; if cancel_real is false, real extents are not cleared.
464  *
465  * Caller must have already joined the inode to the current transaction. The
466  * inode will be joined to the transaction returned to the caller.
467  */
468 int
469 xfs_reflink_cancel_cow_blocks(
470 	struct xfs_inode		*ip,
471 	struct xfs_trans		**tpp,
472 	xfs_fileoff_t			offset_fsb,
473 	xfs_fileoff_t			end_fsb,
474 	bool				cancel_real)
475 {
476 	struct xfs_ifork		*ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
477 	struct xfs_bmbt_irec		got, del;
478 	struct xfs_iext_cursor		icur;
479 	int				error = 0;
480 
481 	if (!xfs_inode_has_cow_data(ip))
482 		return 0;
483 	if (!xfs_iext_lookup_extent_before(ip, ifp, &end_fsb, &icur, &got))
484 		return 0;
485 
486 	/* Walk backwards until we're out of the I/O range... */
487 	while (got.br_startoff + got.br_blockcount > offset_fsb) {
488 		del = got;
489 		xfs_trim_extent(&del, offset_fsb, end_fsb - offset_fsb);
490 
491 		/* Extent delete may have bumped ext forward */
492 		if (!del.br_blockcount) {
493 			xfs_iext_prev(ifp, &icur);
494 			goto next_extent;
495 		}
496 
497 		trace_xfs_reflink_cancel_cow(ip, &del);
498 
499 		if (isnullstartblock(del.br_startblock)) {
500 			error = xfs_bmap_del_extent_delay(ip, XFS_COW_FORK,
501 					&icur, &got, &del);
502 			if (error)
503 				break;
504 		} else if (del.br_state == XFS_EXT_UNWRITTEN || cancel_real) {
505 			ASSERT((*tpp)->t_firstblock == NULLFSBLOCK);
506 
507 			/* Free the CoW orphan record. */
508 			error = xfs_refcount_free_cow_extent(*tpp,
509 					del.br_startblock, del.br_blockcount);
510 			if (error)
511 				break;
512 
513 			xfs_bmap_add_free(*tpp, del.br_startblock,
514 					  del.br_blockcount, NULL);
515 
516 			/* Roll the transaction */
517 			error = xfs_defer_finish(tpp);
518 			if (error)
519 				break;
520 
521 			/* Remove the mapping from the CoW fork. */
522 			xfs_bmap_del_extent_cow(ip, &icur, &got, &del);
523 
524 			/* Remove the quota reservation */
525 			error = xfs_trans_reserve_quota_nblks(NULL, ip,
526 					-(long)del.br_blockcount, 0,
527 					XFS_QMOPT_RES_REGBLKS);
528 			if (error)
529 				break;
530 		} else {
531 			/* Didn't do anything, push cursor back. */
532 			xfs_iext_prev(ifp, &icur);
533 		}
534 next_extent:
535 		if (!xfs_iext_get_extent(ifp, &icur, &got))
536 			break;
537 	}
538 
539 	/* clear tag if cow fork is emptied */
540 	if (!ifp->if_bytes)
541 		xfs_inode_clear_cowblocks_tag(ip);
542 	return error;
543 }
544 
545 /*
546  * Cancel CoW reservations for some byte range of an inode.
547  *
548  * If cancel_real is true this function cancels all COW fork extents for the
549  * inode; if cancel_real is false, real extents are not cleared.
550  */
551 int
552 xfs_reflink_cancel_cow_range(
553 	struct xfs_inode	*ip,
554 	xfs_off_t		offset,
555 	xfs_off_t		count,
556 	bool			cancel_real)
557 {
558 	struct xfs_trans	*tp;
559 	xfs_fileoff_t		offset_fsb;
560 	xfs_fileoff_t		end_fsb;
561 	int			error;
562 
563 	trace_xfs_reflink_cancel_cow_range(ip, offset, count);
564 	ASSERT(xfs_is_reflink_inode(ip));
565 
566 	offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
567 	if (count == NULLFILEOFF)
568 		end_fsb = NULLFILEOFF;
569 	else
570 		end_fsb = XFS_B_TO_FSB(ip->i_mount, offset + count);
571 
572 	/* Start a rolling transaction to remove the mappings */
573 	error = xfs_trans_alloc(ip->i_mount, &M_RES(ip->i_mount)->tr_write,
574 			0, 0, XFS_TRANS_NOFS, &tp);
575 	if (error)
576 		goto out;
577 
578 	xfs_ilock(ip, XFS_ILOCK_EXCL);
579 	xfs_trans_ijoin(tp, ip, 0);
580 
581 	/* Scrape out the old CoW reservations */
582 	error = xfs_reflink_cancel_cow_blocks(ip, &tp, offset_fsb, end_fsb,
583 			cancel_real);
584 	if (error)
585 		goto out_cancel;
586 
587 	error = xfs_trans_commit(tp);
588 
589 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
590 	return error;
591 
592 out_cancel:
593 	xfs_trans_cancel(tp);
594 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
595 out:
596 	trace_xfs_reflink_cancel_cow_range_error(ip, error, _RET_IP_);
597 	return error;
598 }
599 
600 /*
601  * Remap parts of a file's data fork after a successful CoW.
602  */
603 int
604 xfs_reflink_end_cow(
605 	struct xfs_inode		*ip,
606 	xfs_off_t			offset,
607 	xfs_off_t			count)
608 {
609 	struct xfs_ifork		*ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
610 	struct xfs_bmbt_irec		got, del;
611 	struct xfs_trans		*tp;
612 	xfs_fileoff_t			offset_fsb;
613 	xfs_fileoff_t			end_fsb;
614 	int				error;
615 	unsigned int			resblks;
616 	xfs_filblks_t			rlen;
617 	struct xfs_iext_cursor		icur;
618 
619 	trace_xfs_reflink_end_cow(ip, offset, count);
620 
621 	/* No COW extents?  That's easy! */
622 	if (ifp->if_bytes == 0)
623 		return 0;
624 
625 	offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
626 	end_fsb = XFS_B_TO_FSB(ip->i_mount, offset + count);
627 
628 	/*
629 	 * Start a rolling transaction to switch the mappings.  We're
630 	 * unlikely ever to have to remap 16T worth of single-block
631 	 * extents, so just cap the worst case extent count to 2^32-1.
632 	 * Stick a warning in just in case, and avoid 64-bit division.
633 	 */
634 	BUILD_BUG_ON(MAX_RW_COUNT > UINT_MAX);
635 	if (end_fsb - offset_fsb > UINT_MAX) {
636 		error = -EFSCORRUPTED;
637 		xfs_force_shutdown(ip->i_mount, SHUTDOWN_CORRUPT_INCORE);
638 		ASSERT(0);
639 		goto out;
640 	}
641 	resblks = XFS_NEXTENTADD_SPACE_RES(ip->i_mount,
642 			(unsigned int)(end_fsb - offset_fsb),
643 			XFS_DATA_FORK);
644 	error = xfs_trans_alloc(ip->i_mount, &M_RES(ip->i_mount)->tr_write,
645 			resblks, 0, XFS_TRANS_RESERVE | XFS_TRANS_NOFS, &tp);
646 	if (error)
647 		goto out;
648 
649 	xfs_ilock(ip, XFS_ILOCK_EXCL);
650 	xfs_trans_ijoin(tp, ip, 0);
651 
652 	/*
653 	 * In case of racing, overlapping AIO writes no COW extents might be
654 	 * left by the time I/O completes for the loser of the race.  In that
655 	 * case we are done.
656 	 */
657 	if (!xfs_iext_lookup_extent_before(ip, ifp, &end_fsb, &icur, &got))
658 		goto out_cancel;
659 
660 	/* Walk backwards until we're out of the I/O range... */
661 	while (got.br_startoff + got.br_blockcount > offset_fsb) {
662 		del = got;
663 		xfs_trim_extent(&del, offset_fsb, end_fsb - offset_fsb);
664 
665 		/* Extent delete may have bumped ext forward */
666 		if (!del.br_blockcount)
667 			goto prev_extent;
668 
669 		ASSERT(!isnullstartblock(got.br_startblock));
670 
671 		/*
672 		 * Don't remap unwritten extents; these are
673 		 * speculatively preallocated CoW extents that have been
674 		 * allocated but have not yet been involved in a write.
675 		 */
676 		if (got.br_state == XFS_EXT_UNWRITTEN)
677 			goto prev_extent;
678 
679 		/* Unmap the old blocks in the data fork. */
680 		ASSERT(tp->t_firstblock == NULLFSBLOCK);
681 		rlen = del.br_blockcount;
682 		error = __xfs_bunmapi(tp, ip, del.br_startoff, &rlen, 0, 1);
683 		if (error)
684 			goto out_cancel;
685 
686 		/* Trim the extent to whatever got unmapped. */
687 		if (rlen) {
688 			xfs_trim_extent(&del, del.br_startoff + rlen,
689 				del.br_blockcount - rlen);
690 		}
691 		trace_xfs_reflink_cow_remap(ip, &del);
692 
693 		/* Free the CoW orphan record. */
694 		error = xfs_refcount_free_cow_extent(tp, del.br_startblock,
695 				del.br_blockcount);
696 		if (error)
697 			goto out_cancel;
698 
699 		/* Map the new blocks into the data fork. */
700 		error = xfs_bmap_map_extent(tp, ip, &del);
701 		if (error)
702 			goto out_cancel;
703 
704 		/* Charge this new data fork mapping to the on-disk quota. */
705 		xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_DELBCOUNT,
706 				(long)del.br_blockcount);
707 
708 		/* Remove the mapping from the CoW fork. */
709 		xfs_bmap_del_extent_cow(ip, &icur, &got, &del);
710 
711 		error = xfs_defer_finish(&tp);
712 		if (error)
713 			goto out_cancel;
714 		if (!xfs_iext_get_extent(ifp, &icur, &got))
715 			break;
716 		continue;
717 prev_extent:
718 		if (!xfs_iext_prev_extent(ifp, &icur, &got))
719 			break;
720 	}
721 
722 	error = xfs_trans_commit(tp);
723 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
724 	if (error)
725 		goto out;
726 	return 0;
727 
728 out_cancel:
729 	xfs_trans_cancel(tp);
730 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
731 out:
732 	trace_xfs_reflink_end_cow_error(ip, error, _RET_IP_);
733 	return error;
734 }
735 
736 /*
737  * Free leftover CoW reservations that didn't get cleaned out.
738  */
739 int
740 xfs_reflink_recover_cow(
741 	struct xfs_mount	*mp)
742 {
743 	xfs_agnumber_t		agno;
744 	int			error = 0;
745 
746 	if (!xfs_sb_version_hasreflink(&mp->m_sb))
747 		return 0;
748 
749 	for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
750 		error = xfs_refcount_recover_cow_leftovers(mp, agno);
751 		if (error)
752 			break;
753 	}
754 
755 	return error;
756 }
757 
758 /*
759  * Reflinking (Block) Ranges of Two Files Together
760  *
761  * First, ensure that the reflink flag is set on both inodes.  The flag is an
762  * optimization to avoid unnecessary refcount btree lookups in the write path.
763  *
764  * Now we can iteratively remap the range of extents (and holes) in src to the
765  * corresponding ranges in dest.  Let drange and srange denote the ranges of
766  * logical blocks in dest and src touched by the reflink operation.
767  *
768  * While the length of drange is greater than zero,
769  *    - Read src's bmbt at the start of srange ("imap")
770  *    - If imap doesn't exist, make imap appear to start at the end of srange
771  *      with zero length.
772  *    - If imap starts before srange, advance imap to start at srange.
773  *    - If imap goes beyond srange, truncate imap to end at the end of srange.
774  *    - Punch (imap start - srange start + imap len) blocks from dest at
775  *      offset (drange start).
776  *    - If imap points to a real range of pblks,
777  *         > Increase the refcount of the imap's pblks
778  *         > Map imap's pblks into dest at the offset
779  *           (drange start + imap start - srange start)
780  *    - Advance drange and srange by (imap start - srange start + imap len)
781  *
782  * Finally, if the reflink made dest longer, update both the in-core and
783  * on-disk file sizes.
784  *
785  * ASCII Art Demonstration:
786  *
787  * Let's say we want to reflink this source file:
788  *
789  * ----SSSSSSS-SSSSS----SSSSSS (src file)
790  *   <-------------------->
791  *
792  * into this destination file:
793  *
794  * --DDDDDDDDDDDDDDDDDDD--DDD (dest file)
795  *        <-------------------->
796  * '-' means a hole, and 'S' and 'D' are written blocks in the src and dest.
797  * Observe that the range has different logical offsets in either file.
798  *
799  * Consider that the first extent in the source file doesn't line up with our
800  * reflink range.  Unmapping  and remapping are separate operations, so we can
801  * unmap more blocks from the destination file than we remap.
802  *
803  * ----SSSSSSS-SSSSS----SSSSSS
804  *   <------->
805  * --DDDDD---------DDDDD--DDD
806  *        <------->
807  *
808  * Now remap the source extent into the destination file:
809  *
810  * ----SSSSSSS-SSSSS----SSSSSS
811  *   <------->
812  * --DDDDD--SSSSSSSDDDDD--DDD
813  *        <------->
814  *
815  * Do likewise with the second hole and extent in our range.  Holes in the
816  * unmap range don't affect our operation.
817  *
818  * ----SSSSSSS-SSSSS----SSSSSS
819  *            <---->
820  * --DDDDD--SSSSSSS-SSSSS-DDD
821  *                 <---->
822  *
823  * Finally, unmap and remap part of the third extent.  This will increase the
824  * size of the destination file.
825  *
826  * ----SSSSSSS-SSSSS----SSSSSS
827  *                  <----->
828  * --DDDDD--SSSSSSS-SSSSS----SSS
829  *                       <----->
830  *
831  * Once we update the destination file's i_size, we're done.
832  */
833 
834 /*
835  * Ensure the reflink bit is set in both inodes.
836  */
837 STATIC int
838 xfs_reflink_set_inode_flag(
839 	struct xfs_inode	*src,
840 	struct xfs_inode	*dest)
841 {
842 	struct xfs_mount	*mp = src->i_mount;
843 	int			error;
844 	struct xfs_trans	*tp;
845 
846 	if (xfs_is_reflink_inode(src) && xfs_is_reflink_inode(dest))
847 		return 0;
848 
849 	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
850 	if (error)
851 		goto out_error;
852 
853 	/* Lock both files against IO */
854 	if (src->i_ino == dest->i_ino)
855 		xfs_ilock(src, XFS_ILOCK_EXCL);
856 	else
857 		xfs_lock_two_inodes(src, XFS_ILOCK_EXCL, dest, XFS_ILOCK_EXCL);
858 
859 	if (!xfs_is_reflink_inode(src)) {
860 		trace_xfs_reflink_set_inode_flag(src);
861 		xfs_trans_ijoin(tp, src, XFS_ILOCK_EXCL);
862 		src->i_d.di_flags2 |= XFS_DIFLAG2_REFLINK;
863 		xfs_trans_log_inode(tp, src, XFS_ILOG_CORE);
864 		xfs_ifork_init_cow(src);
865 	} else
866 		xfs_iunlock(src, XFS_ILOCK_EXCL);
867 
868 	if (src->i_ino == dest->i_ino)
869 		goto commit_flags;
870 
871 	if (!xfs_is_reflink_inode(dest)) {
872 		trace_xfs_reflink_set_inode_flag(dest);
873 		xfs_trans_ijoin(tp, dest, XFS_ILOCK_EXCL);
874 		dest->i_d.di_flags2 |= XFS_DIFLAG2_REFLINK;
875 		xfs_trans_log_inode(tp, dest, XFS_ILOG_CORE);
876 		xfs_ifork_init_cow(dest);
877 	} else
878 		xfs_iunlock(dest, XFS_ILOCK_EXCL);
879 
880 commit_flags:
881 	error = xfs_trans_commit(tp);
882 	if (error)
883 		goto out_error;
884 	return error;
885 
886 out_error:
887 	trace_xfs_reflink_set_inode_flag_error(dest, error, _RET_IP_);
888 	return error;
889 }
890 
891 /*
892  * Update destination inode size & cowextsize hint, if necessary.
893  */
894 STATIC int
895 xfs_reflink_update_dest(
896 	struct xfs_inode	*dest,
897 	xfs_off_t		newlen,
898 	xfs_extlen_t		cowextsize,
899 	bool			is_dedupe)
900 {
901 	struct xfs_mount	*mp = dest->i_mount;
902 	struct xfs_trans	*tp;
903 	int			error;
904 
905 	if (is_dedupe && newlen <= i_size_read(VFS_I(dest)) && cowextsize == 0)
906 		return 0;
907 
908 	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
909 	if (error)
910 		goto out_error;
911 
912 	xfs_ilock(dest, XFS_ILOCK_EXCL);
913 	xfs_trans_ijoin(tp, dest, XFS_ILOCK_EXCL);
914 
915 	if (newlen > i_size_read(VFS_I(dest))) {
916 		trace_xfs_reflink_update_inode_size(dest, newlen);
917 		i_size_write(VFS_I(dest), newlen);
918 		dest->i_d.di_size = newlen;
919 	}
920 
921 	if (cowextsize) {
922 		dest->i_d.di_cowextsize = cowextsize;
923 		dest->i_d.di_flags2 |= XFS_DIFLAG2_COWEXTSIZE;
924 	}
925 
926 	if (!is_dedupe) {
927 		xfs_trans_ichgtime(tp, dest,
928 				   XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
929 	}
930 	xfs_trans_log_inode(tp, dest, XFS_ILOG_CORE);
931 
932 	error = xfs_trans_commit(tp);
933 	if (error)
934 		goto out_error;
935 	return error;
936 
937 out_error:
938 	trace_xfs_reflink_update_inode_size_error(dest, error, _RET_IP_);
939 	return error;
940 }
941 
942 /*
943  * Do we have enough reserve in this AG to handle a reflink?  The refcount
944  * btree already reserved all the space it needs, but the rmap btree can grow
945  * infinitely, so we won't allow more reflinks when the AG is down to the
946  * btree reserves.
947  */
948 static int
949 xfs_reflink_ag_has_free_space(
950 	struct xfs_mount	*mp,
951 	xfs_agnumber_t		agno)
952 {
953 	struct xfs_perag	*pag;
954 	int			error = 0;
955 
956 	if (!xfs_sb_version_hasrmapbt(&mp->m_sb))
957 		return 0;
958 
959 	pag = xfs_perag_get(mp, agno);
960 	if (xfs_ag_resv_critical(pag, XFS_AG_RESV_RMAPBT) ||
961 	    xfs_ag_resv_critical(pag, XFS_AG_RESV_METADATA))
962 		error = -ENOSPC;
963 	xfs_perag_put(pag);
964 	return error;
965 }
966 
967 /*
968  * Unmap a range of blocks from a file, then map other blocks into the hole.
969  * The range to unmap is (destoff : destoff + srcioff + irec->br_blockcount).
970  * The extent irec is mapped into dest at irec->br_startoff.
971  */
972 STATIC int
973 xfs_reflink_remap_extent(
974 	struct xfs_inode	*ip,
975 	struct xfs_bmbt_irec	*irec,
976 	xfs_fileoff_t		destoff,
977 	xfs_off_t		new_isize)
978 {
979 	struct xfs_mount	*mp = ip->i_mount;
980 	bool			real_extent = xfs_bmap_is_real_extent(irec);
981 	struct xfs_trans	*tp;
982 	unsigned int		resblks;
983 	struct xfs_bmbt_irec	uirec;
984 	xfs_filblks_t		rlen;
985 	xfs_filblks_t		unmap_len;
986 	xfs_off_t		newlen;
987 	int			error;
988 
989 	unmap_len = irec->br_startoff + irec->br_blockcount - destoff;
990 	trace_xfs_reflink_punch_range(ip, destoff, unmap_len);
991 
992 	/* No reflinking if we're low on space */
993 	if (real_extent) {
994 		error = xfs_reflink_ag_has_free_space(mp,
995 				XFS_FSB_TO_AGNO(mp, irec->br_startblock));
996 		if (error)
997 			goto out;
998 	}
999 
1000 	/* Start a rolling transaction to switch the mappings */
1001 	resblks = XFS_EXTENTADD_SPACE_RES(ip->i_mount, XFS_DATA_FORK);
1002 	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, resblks, 0, 0, &tp);
1003 	if (error)
1004 		goto out;
1005 
1006 	xfs_ilock(ip, XFS_ILOCK_EXCL);
1007 	xfs_trans_ijoin(tp, ip, 0);
1008 
1009 	/* If we're not just clearing space, then do we have enough quota? */
1010 	if (real_extent) {
1011 		error = xfs_trans_reserve_quota_nblks(tp, ip,
1012 				irec->br_blockcount, 0, XFS_QMOPT_RES_REGBLKS);
1013 		if (error)
1014 			goto out_cancel;
1015 	}
1016 
1017 	trace_xfs_reflink_remap(ip, irec->br_startoff,
1018 				irec->br_blockcount, irec->br_startblock);
1019 
1020 	/* Unmap the old blocks in the data fork. */
1021 	rlen = unmap_len;
1022 	while (rlen) {
1023 		ASSERT(tp->t_firstblock == NULLFSBLOCK);
1024 		error = __xfs_bunmapi(tp, ip, destoff, &rlen, 0, 1);
1025 		if (error)
1026 			goto out_cancel;
1027 
1028 		/*
1029 		 * Trim the extent to whatever got unmapped.
1030 		 * Remember, bunmapi works backwards.
1031 		 */
1032 		uirec.br_startblock = irec->br_startblock + rlen;
1033 		uirec.br_startoff = irec->br_startoff + rlen;
1034 		uirec.br_blockcount = unmap_len - rlen;
1035 		unmap_len = rlen;
1036 
1037 		/* If this isn't a real mapping, we're done. */
1038 		if (!real_extent || uirec.br_blockcount == 0)
1039 			goto next_extent;
1040 
1041 		trace_xfs_reflink_remap(ip, uirec.br_startoff,
1042 				uirec.br_blockcount, uirec.br_startblock);
1043 
1044 		/* Update the refcount tree */
1045 		error = xfs_refcount_increase_extent(tp, &uirec);
1046 		if (error)
1047 			goto out_cancel;
1048 
1049 		/* Map the new blocks into the data fork. */
1050 		error = xfs_bmap_map_extent(tp, ip, &uirec);
1051 		if (error)
1052 			goto out_cancel;
1053 
1054 		/* Update quota accounting. */
1055 		xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_BCOUNT,
1056 				uirec.br_blockcount);
1057 
1058 		/* Update dest isize if needed. */
1059 		newlen = XFS_FSB_TO_B(mp,
1060 				uirec.br_startoff + uirec.br_blockcount);
1061 		newlen = min_t(xfs_off_t, newlen, new_isize);
1062 		if (newlen > i_size_read(VFS_I(ip))) {
1063 			trace_xfs_reflink_update_inode_size(ip, newlen);
1064 			i_size_write(VFS_I(ip), newlen);
1065 			ip->i_d.di_size = newlen;
1066 			xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1067 		}
1068 
1069 next_extent:
1070 		/* Process all the deferred stuff. */
1071 		error = xfs_defer_finish(&tp);
1072 		if (error)
1073 			goto out_cancel;
1074 	}
1075 
1076 	error = xfs_trans_commit(tp);
1077 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1078 	if (error)
1079 		goto out;
1080 	return 0;
1081 
1082 out_cancel:
1083 	xfs_trans_cancel(tp);
1084 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1085 out:
1086 	trace_xfs_reflink_remap_extent_error(ip, error, _RET_IP_);
1087 	return error;
1088 }
1089 
1090 /*
1091  * Iteratively remap one file's extents (and holes) to another's.
1092  */
1093 STATIC int
1094 xfs_reflink_remap_blocks(
1095 	struct xfs_inode	*src,
1096 	xfs_fileoff_t		srcoff,
1097 	struct xfs_inode	*dest,
1098 	xfs_fileoff_t		destoff,
1099 	xfs_filblks_t		len,
1100 	xfs_off_t		new_isize)
1101 {
1102 	struct xfs_bmbt_irec	imap;
1103 	int			nimaps;
1104 	int			error = 0;
1105 	xfs_filblks_t		range_len;
1106 
1107 	/* drange = (destoff, destoff + len); srange = (srcoff, srcoff + len) */
1108 	while (len) {
1109 		uint		lock_mode;
1110 
1111 		trace_xfs_reflink_remap_blocks_loop(src, srcoff, len,
1112 				dest, destoff);
1113 
1114 		/* Read extent from the source file */
1115 		nimaps = 1;
1116 		lock_mode = xfs_ilock_data_map_shared(src);
1117 		error = xfs_bmapi_read(src, srcoff, len, &imap, &nimaps, 0);
1118 		xfs_iunlock(src, lock_mode);
1119 		if (error)
1120 			goto err;
1121 		ASSERT(nimaps == 1);
1122 
1123 		trace_xfs_reflink_remap_imap(src, srcoff, len, XFS_IO_OVERWRITE,
1124 				&imap);
1125 
1126 		/* Translate imap into the destination file. */
1127 		range_len = imap.br_startoff + imap.br_blockcount - srcoff;
1128 		imap.br_startoff += destoff - srcoff;
1129 
1130 		/* Clear dest from destoff to the end of imap and map it in. */
1131 		error = xfs_reflink_remap_extent(dest, &imap, destoff,
1132 				new_isize);
1133 		if (error)
1134 			goto err;
1135 
1136 		if (fatal_signal_pending(current)) {
1137 			error = -EINTR;
1138 			goto err;
1139 		}
1140 
1141 		/* Advance drange/srange */
1142 		srcoff += range_len;
1143 		destoff += range_len;
1144 		len -= range_len;
1145 	}
1146 
1147 	return 0;
1148 
1149 err:
1150 	trace_xfs_reflink_remap_blocks_error(dest, error, _RET_IP_);
1151 	return error;
1152 }
1153 
1154 /*
1155  * Grab the exclusive iolock for a data copy from src to dest, making
1156  * sure to abide vfs locking order (lowest pointer value goes first) and
1157  * breaking the pnfs layout leases on dest before proceeding.  The loop
1158  * is needed because we cannot call the blocking break_layout() with the
1159  * src iolock held, and therefore have to back out both locks.
1160  */
1161 static int
1162 xfs_iolock_two_inodes_and_break_layout(
1163 	struct inode		*src,
1164 	struct inode		*dest)
1165 {
1166 	int			error;
1167 
1168 retry:
1169 	if (src < dest) {
1170 		inode_lock_shared(src);
1171 		inode_lock_nested(dest, I_MUTEX_NONDIR2);
1172 	} else {
1173 		/* src >= dest */
1174 		inode_lock(dest);
1175 	}
1176 
1177 	error = break_layout(dest, false);
1178 	if (error == -EWOULDBLOCK) {
1179 		inode_unlock(dest);
1180 		if (src < dest)
1181 			inode_unlock_shared(src);
1182 		error = break_layout(dest, true);
1183 		if (error)
1184 			return error;
1185 		goto retry;
1186 	}
1187 	if (error) {
1188 		inode_unlock(dest);
1189 		if (src < dest)
1190 			inode_unlock_shared(src);
1191 		return error;
1192 	}
1193 	if (src > dest)
1194 		inode_lock_shared_nested(src, I_MUTEX_NONDIR2);
1195 	return 0;
1196 }
1197 
1198 /*
1199  * Link a range of blocks from one file to another.
1200  */
1201 int
1202 xfs_reflink_remap_range(
1203 	struct file		*file_in,
1204 	loff_t			pos_in,
1205 	struct file		*file_out,
1206 	loff_t			pos_out,
1207 	u64			len,
1208 	bool			is_dedupe)
1209 {
1210 	struct inode		*inode_in = file_inode(file_in);
1211 	struct xfs_inode	*src = XFS_I(inode_in);
1212 	struct inode		*inode_out = file_inode(file_out);
1213 	struct xfs_inode	*dest = XFS_I(inode_out);
1214 	struct xfs_mount	*mp = src->i_mount;
1215 	bool			same_inode = (inode_in == inode_out);
1216 	xfs_fileoff_t		sfsbno, dfsbno;
1217 	xfs_filblks_t		fsblen;
1218 	xfs_extlen_t		cowextsize;
1219 	ssize_t			ret;
1220 
1221 	if (!xfs_sb_version_hasreflink(&mp->m_sb))
1222 		return -EOPNOTSUPP;
1223 
1224 	if (XFS_FORCED_SHUTDOWN(mp))
1225 		return -EIO;
1226 
1227 	/* Lock both files against IO */
1228 	ret = xfs_iolock_two_inodes_and_break_layout(inode_in, inode_out);
1229 	if (ret)
1230 		return ret;
1231 	if (same_inode)
1232 		xfs_ilock(src, XFS_MMAPLOCK_EXCL);
1233 	else
1234 		xfs_lock_two_inodes(src, XFS_MMAPLOCK_SHARED, dest,
1235 				XFS_MMAPLOCK_EXCL);
1236 
1237 	/* Check file eligibility and prepare for block sharing. */
1238 	ret = -EINVAL;
1239 	/* Don't reflink realtime inodes */
1240 	if (XFS_IS_REALTIME_INODE(src) || XFS_IS_REALTIME_INODE(dest))
1241 		goto out_unlock;
1242 
1243 	/* Don't share DAX file data for now. */
1244 	if (IS_DAX(inode_in) || IS_DAX(inode_out))
1245 		goto out_unlock;
1246 
1247 	ret = vfs_clone_file_prep_inodes(inode_in, pos_in, inode_out, pos_out,
1248 			&len, is_dedupe);
1249 	if (ret <= 0)
1250 		goto out_unlock;
1251 
1252 	/* Attach dquots to dest inode before changing block map */
1253 	ret = xfs_qm_dqattach(dest);
1254 	if (ret)
1255 		goto out_unlock;
1256 
1257 	trace_xfs_reflink_remap_range(src, pos_in, len, dest, pos_out);
1258 
1259 	/*
1260 	 * Clear out post-eof preallocations because we don't have page cache
1261 	 * backing the delayed allocations and they'll never get freed on
1262 	 * their own.
1263 	 */
1264 	if (xfs_can_free_eofblocks(dest, true)) {
1265 		ret = xfs_free_eofblocks(dest);
1266 		if (ret)
1267 			goto out_unlock;
1268 	}
1269 
1270 	/* Set flags and remap blocks. */
1271 	ret = xfs_reflink_set_inode_flag(src, dest);
1272 	if (ret)
1273 		goto out_unlock;
1274 
1275 	dfsbno = XFS_B_TO_FSBT(mp, pos_out);
1276 	sfsbno = XFS_B_TO_FSBT(mp, pos_in);
1277 	fsblen = XFS_B_TO_FSB(mp, len);
1278 	ret = xfs_reflink_remap_blocks(src, sfsbno, dest, dfsbno, fsblen,
1279 			pos_out + len);
1280 	if (ret)
1281 		goto out_unlock;
1282 
1283 	/* Zap any page cache for the destination file's range. */
1284 	truncate_inode_pages_range(&inode_out->i_data, pos_out,
1285 				   PAGE_ALIGN(pos_out + len) - 1);
1286 
1287 	/*
1288 	 * Carry the cowextsize hint from src to dest if we're sharing the
1289 	 * entire source file to the entire destination file, the source file
1290 	 * has a cowextsize hint, and the destination file does not.
1291 	 */
1292 	cowextsize = 0;
1293 	if (pos_in == 0 && len == i_size_read(inode_in) &&
1294 	    (src->i_d.di_flags2 & XFS_DIFLAG2_COWEXTSIZE) &&
1295 	    pos_out == 0 && len >= i_size_read(inode_out) &&
1296 	    !(dest->i_d.di_flags2 & XFS_DIFLAG2_COWEXTSIZE))
1297 		cowextsize = src->i_d.di_cowextsize;
1298 
1299 	ret = xfs_reflink_update_dest(dest, pos_out + len, cowextsize,
1300 			is_dedupe);
1301 
1302 out_unlock:
1303 	xfs_iunlock(dest, XFS_MMAPLOCK_EXCL);
1304 	if (!same_inode)
1305 		xfs_iunlock(src, XFS_MMAPLOCK_SHARED);
1306 	inode_unlock(inode_out);
1307 	if (!same_inode)
1308 		inode_unlock_shared(inode_in);
1309 	if (ret)
1310 		trace_xfs_reflink_remap_range_error(dest, ret, _RET_IP_);
1311 	return ret;
1312 }
1313 
1314 /*
1315  * The user wants to preemptively CoW all shared blocks in this file,
1316  * which enables us to turn off the reflink flag.  Iterate all
1317  * extents which are not prealloc/delalloc to see which ranges are
1318  * mentioned in the refcount tree, then read those blocks into the
1319  * pagecache, dirty them, fsync them back out, and then we can update
1320  * the inode flag.  What happens if we run out of memory? :)
1321  */
1322 STATIC int
1323 xfs_reflink_dirty_extents(
1324 	struct xfs_inode	*ip,
1325 	xfs_fileoff_t		fbno,
1326 	xfs_filblks_t		end,
1327 	xfs_off_t		isize)
1328 {
1329 	struct xfs_mount	*mp = ip->i_mount;
1330 	xfs_agnumber_t		agno;
1331 	xfs_agblock_t		agbno;
1332 	xfs_extlen_t		aglen;
1333 	xfs_agblock_t		rbno;
1334 	xfs_extlen_t		rlen;
1335 	xfs_off_t		fpos;
1336 	xfs_off_t		flen;
1337 	struct xfs_bmbt_irec	map[2];
1338 	int			nmaps;
1339 	int			error = 0;
1340 
1341 	while (end - fbno > 0) {
1342 		nmaps = 1;
1343 		/*
1344 		 * Look for extents in the file.  Skip holes, delalloc, or
1345 		 * unwritten extents; they can't be reflinked.
1346 		 */
1347 		error = xfs_bmapi_read(ip, fbno, end - fbno, map, &nmaps, 0);
1348 		if (error)
1349 			goto out;
1350 		if (nmaps == 0)
1351 			break;
1352 		if (!xfs_bmap_is_real_extent(&map[0]))
1353 			goto next;
1354 
1355 		map[1] = map[0];
1356 		while (map[1].br_blockcount) {
1357 			agno = XFS_FSB_TO_AGNO(mp, map[1].br_startblock);
1358 			agbno = XFS_FSB_TO_AGBNO(mp, map[1].br_startblock);
1359 			aglen = map[1].br_blockcount;
1360 
1361 			error = xfs_reflink_find_shared(mp, NULL, agno, agbno,
1362 					aglen, &rbno, &rlen, true);
1363 			if (error)
1364 				goto out;
1365 			if (rbno == NULLAGBLOCK)
1366 				break;
1367 
1368 			/* Dirty the pages */
1369 			xfs_iunlock(ip, XFS_ILOCK_EXCL);
1370 			fpos = XFS_FSB_TO_B(mp, map[1].br_startoff +
1371 					(rbno - agbno));
1372 			flen = XFS_FSB_TO_B(mp, rlen);
1373 			if (fpos + flen > isize)
1374 				flen = isize - fpos;
1375 			error = iomap_file_dirty(VFS_I(ip), fpos, flen,
1376 					&xfs_iomap_ops);
1377 			xfs_ilock(ip, XFS_ILOCK_EXCL);
1378 			if (error)
1379 				goto out;
1380 
1381 			map[1].br_blockcount -= (rbno - agbno + rlen);
1382 			map[1].br_startoff += (rbno - agbno + rlen);
1383 			map[1].br_startblock += (rbno - agbno + rlen);
1384 		}
1385 
1386 next:
1387 		fbno = map[0].br_startoff + map[0].br_blockcount;
1388 	}
1389 out:
1390 	return error;
1391 }
1392 
1393 /* Does this inode need the reflink flag? */
1394 int
1395 xfs_reflink_inode_has_shared_extents(
1396 	struct xfs_trans		*tp,
1397 	struct xfs_inode		*ip,
1398 	bool				*has_shared)
1399 {
1400 	struct xfs_bmbt_irec		got;
1401 	struct xfs_mount		*mp = ip->i_mount;
1402 	struct xfs_ifork		*ifp;
1403 	xfs_agnumber_t			agno;
1404 	xfs_agblock_t			agbno;
1405 	xfs_extlen_t			aglen;
1406 	xfs_agblock_t			rbno;
1407 	xfs_extlen_t			rlen;
1408 	struct xfs_iext_cursor		icur;
1409 	bool				found;
1410 	int				error;
1411 
1412 	ifp = XFS_IFORK_PTR(ip, XFS_DATA_FORK);
1413 	if (!(ifp->if_flags & XFS_IFEXTENTS)) {
1414 		error = xfs_iread_extents(tp, ip, XFS_DATA_FORK);
1415 		if (error)
1416 			return error;
1417 	}
1418 
1419 	*has_shared = false;
1420 	found = xfs_iext_lookup_extent(ip, ifp, 0, &icur, &got);
1421 	while (found) {
1422 		if (isnullstartblock(got.br_startblock) ||
1423 		    got.br_state != XFS_EXT_NORM)
1424 			goto next;
1425 		agno = XFS_FSB_TO_AGNO(mp, got.br_startblock);
1426 		agbno = XFS_FSB_TO_AGBNO(mp, got.br_startblock);
1427 		aglen = got.br_blockcount;
1428 
1429 		error = xfs_reflink_find_shared(mp, tp, agno, agbno, aglen,
1430 				&rbno, &rlen, false);
1431 		if (error)
1432 			return error;
1433 		/* Is there still a shared block here? */
1434 		if (rbno != NULLAGBLOCK) {
1435 			*has_shared = true;
1436 			return 0;
1437 		}
1438 next:
1439 		found = xfs_iext_next_extent(ifp, &icur, &got);
1440 	}
1441 
1442 	return 0;
1443 }
1444 
1445 /*
1446  * Clear the inode reflink flag if there are no shared extents.
1447  *
1448  * The caller is responsible for joining the inode to the transaction passed in.
1449  * The inode will be joined to the transaction that is returned to the caller.
1450  */
1451 int
1452 xfs_reflink_clear_inode_flag(
1453 	struct xfs_inode	*ip,
1454 	struct xfs_trans	**tpp)
1455 {
1456 	bool			needs_flag;
1457 	int			error = 0;
1458 
1459 	ASSERT(xfs_is_reflink_inode(ip));
1460 
1461 	error = xfs_reflink_inode_has_shared_extents(*tpp, ip, &needs_flag);
1462 	if (error || needs_flag)
1463 		return error;
1464 
1465 	/*
1466 	 * We didn't find any shared blocks so turn off the reflink flag.
1467 	 * First, get rid of any leftover CoW mappings.
1468 	 */
1469 	error = xfs_reflink_cancel_cow_blocks(ip, tpp, 0, NULLFILEOFF, true);
1470 	if (error)
1471 		return error;
1472 
1473 	/* Clear the inode flag. */
1474 	trace_xfs_reflink_unset_inode_flag(ip);
1475 	ip->i_d.di_flags2 &= ~XFS_DIFLAG2_REFLINK;
1476 	xfs_inode_clear_cowblocks_tag(ip);
1477 	xfs_trans_log_inode(*tpp, ip, XFS_ILOG_CORE);
1478 
1479 	return error;
1480 }
1481 
1482 /*
1483  * Clear the inode reflink flag if there are no shared extents and the size
1484  * hasn't changed.
1485  */
1486 STATIC int
1487 xfs_reflink_try_clear_inode_flag(
1488 	struct xfs_inode	*ip)
1489 {
1490 	struct xfs_mount	*mp = ip->i_mount;
1491 	struct xfs_trans	*tp;
1492 	int			error = 0;
1493 
1494 	/* Start a rolling transaction to remove the mappings */
1495 	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, 0, 0, 0, &tp);
1496 	if (error)
1497 		return error;
1498 
1499 	xfs_ilock(ip, XFS_ILOCK_EXCL);
1500 	xfs_trans_ijoin(tp, ip, 0);
1501 
1502 	error = xfs_reflink_clear_inode_flag(ip, &tp);
1503 	if (error)
1504 		goto cancel;
1505 
1506 	error = xfs_trans_commit(tp);
1507 	if (error)
1508 		goto out;
1509 
1510 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1511 	return 0;
1512 cancel:
1513 	xfs_trans_cancel(tp);
1514 out:
1515 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1516 	return error;
1517 }
1518 
1519 /*
1520  * Pre-COW all shared blocks within a given byte range of a file and turn off
1521  * the reflink flag if we unshare all of the file's blocks.
1522  */
1523 int
1524 xfs_reflink_unshare(
1525 	struct xfs_inode	*ip,
1526 	xfs_off_t		offset,
1527 	xfs_off_t		len)
1528 {
1529 	struct xfs_mount	*mp = ip->i_mount;
1530 	xfs_fileoff_t		fbno;
1531 	xfs_filblks_t		end;
1532 	xfs_off_t		isize;
1533 	int			error;
1534 
1535 	if (!xfs_is_reflink_inode(ip))
1536 		return 0;
1537 
1538 	trace_xfs_reflink_unshare(ip, offset, len);
1539 
1540 	inode_dio_wait(VFS_I(ip));
1541 
1542 	/* Try to CoW the selected ranges */
1543 	xfs_ilock(ip, XFS_ILOCK_EXCL);
1544 	fbno = XFS_B_TO_FSBT(mp, offset);
1545 	isize = i_size_read(VFS_I(ip));
1546 	end = XFS_B_TO_FSB(mp, offset + len);
1547 	error = xfs_reflink_dirty_extents(ip, fbno, end, isize);
1548 	if (error)
1549 		goto out_unlock;
1550 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1551 
1552 	/* Wait for the IO to finish */
1553 	error = filemap_write_and_wait(VFS_I(ip)->i_mapping);
1554 	if (error)
1555 		goto out;
1556 
1557 	/* Turn off the reflink flag if possible. */
1558 	error = xfs_reflink_try_clear_inode_flag(ip);
1559 	if (error)
1560 		goto out;
1561 
1562 	return 0;
1563 
1564 out_unlock:
1565 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1566 out:
1567 	trace_xfs_reflink_unshare_error(ip, error, _RET_IP_);
1568 	return error;
1569 }
1570