xref: /openbmc/linux/fs/xfs/xfs_reflink.c (revision 6dfcd296)
1 /*
2  * Copyright (C) 2016 Oracle.  All Rights Reserved.
3  *
4  * Author: Darrick J. Wong <darrick.wong@oracle.com>
5  *
6  * This program is free software; you can redistribute it and/or
7  * modify it under the terms of the GNU General Public License
8  * as published by the Free Software Foundation; either version 2
9  * of the License, or (at your option) any later version.
10  *
11  * This program is distributed in the hope that it would be useful,
12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
14  * GNU General Public License for more details.
15  *
16  * You should have received a copy of the GNU General Public License
17  * along with this program; if not, write the Free Software Foundation,
18  * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301, USA.
19  */
20 #include "xfs.h"
21 #include "xfs_fs.h"
22 #include "xfs_shared.h"
23 #include "xfs_format.h"
24 #include "xfs_log_format.h"
25 #include "xfs_trans_resv.h"
26 #include "xfs_mount.h"
27 #include "xfs_defer.h"
28 #include "xfs_da_format.h"
29 #include "xfs_da_btree.h"
30 #include "xfs_inode.h"
31 #include "xfs_trans.h"
32 #include "xfs_inode_item.h"
33 #include "xfs_bmap.h"
34 #include "xfs_bmap_util.h"
35 #include "xfs_error.h"
36 #include "xfs_dir2.h"
37 #include "xfs_dir2_priv.h"
38 #include "xfs_ioctl.h"
39 #include "xfs_trace.h"
40 #include "xfs_log.h"
41 #include "xfs_icache.h"
42 #include "xfs_pnfs.h"
43 #include "xfs_btree.h"
44 #include "xfs_refcount_btree.h"
45 #include "xfs_refcount.h"
46 #include "xfs_bmap_btree.h"
47 #include "xfs_trans_space.h"
48 #include "xfs_bit.h"
49 #include "xfs_alloc.h"
50 #include "xfs_quota_defs.h"
51 #include "xfs_quota.h"
52 #include "xfs_btree.h"
53 #include "xfs_bmap_btree.h"
54 #include "xfs_reflink.h"
55 #include "xfs_iomap.h"
56 #include "xfs_rmap_btree.h"
57 #include "xfs_sb.h"
58 #include "xfs_ag_resv.h"
59 
60 /*
61  * Copy on Write of Shared Blocks
62  *
63  * XFS must preserve "the usual" file semantics even when two files share
64  * the same physical blocks.  This means that a write to one file must not
65  * alter the blocks in a different file; the way that we'll do that is
66  * through the use of a copy-on-write mechanism.  At a high level, that
67  * means that when we want to write to a shared block, we allocate a new
68  * block, write the data to the new block, and if that succeeds we map the
69  * new block into the file.
70  *
71  * XFS provides a "delayed allocation" mechanism that defers the allocation
72  * of disk blocks to dirty-but-not-yet-mapped file blocks as long as
73  * possible.  This reduces fragmentation by enabling the filesystem to ask
74  * for bigger chunks less often, which is exactly what we want for CoW.
75  *
76  * The delalloc mechanism begins when the kernel wants to make a block
77  * writable (write_begin or page_mkwrite).  If the offset is not mapped, we
78  * create a delalloc mapping, which is a regular in-core extent, but without
79  * a real startblock.  (For delalloc mappings, the startblock encodes both
80  * a flag that this is a delalloc mapping, and a worst-case estimate of how
81  * many blocks might be required to put the mapping into the BMBT.)  delalloc
82  * mappings are a reservation against the free space in the filesystem;
83  * adjacent mappings can also be combined into fewer larger mappings.
84  *
85  * When dirty pages are being written out (typically in writepage), the
86  * delalloc reservations are converted into real mappings by allocating
87  * blocks and replacing the delalloc mapping with real ones.  A delalloc
88  * mapping can be replaced by several real ones if the free space is
89  * fragmented.
90  *
91  * We want to adapt the delalloc mechanism for copy-on-write, since the
92  * write paths are similar.  The first two steps (creating the reservation
93  * and allocating the blocks) are exactly the same as delalloc except that
94  * the mappings must be stored in a separate CoW fork because we do not want
95  * to disturb the mapping in the data fork until we're sure that the write
96  * succeeded.  IO completion in this case is the process of removing the old
97  * mapping from the data fork and moving the new mapping from the CoW fork to
98  * the data fork.  This will be discussed shortly.
99  *
100  * For now, unaligned directio writes will be bounced back to the page cache.
101  * Block-aligned directio writes will use the same mechanism as buffered
102  * writes.
103  *
104  * CoW remapping must be done after the data block write completes,
105  * because we don't want to destroy the old data fork map until we're sure
106  * the new block has been written.  Since the new mappings are kept in a
107  * separate fork, we can simply iterate these mappings to find the ones
108  * that cover the file blocks that we just CoW'd.  For each extent, simply
109  * unmap the corresponding range in the data fork, map the new range into
110  * the data fork, and remove the extent from the CoW fork.
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 	xfs_agnumber_t		agno,
132 	xfs_agblock_t		agbno,
133 	xfs_extlen_t		aglen,
134 	xfs_agblock_t		*fbno,
135 	xfs_extlen_t		*flen,
136 	bool			find_end_of_shared)
137 {
138 	struct xfs_buf		*agbp;
139 	struct xfs_btree_cur	*cur;
140 	int			error;
141 
142 	error = xfs_alloc_read_agf(mp, NULL, agno, 0, &agbp);
143 	if (error)
144 		return error;
145 
146 	cur = xfs_refcountbt_init_cursor(mp, NULL, agbp, agno, NULL);
147 
148 	error = xfs_refcount_find_shared(cur, agbno, aglen, fbno, flen,
149 			find_end_of_shared);
150 
151 	xfs_btree_del_cursor(cur, error ? XFS_BTREE_ERROR : XFS_BTREE_NOERROR);
152 
153 	xfs_buf_relse(agbp);
154 	return error;
155 }
156 
157 /*
158  * Trim the mapping to the next block where there's a change in the
159  * shared/unshared status.  More specifically, this means that we
160  * find the lowest-numbered extent of shared blocks that coincides with
161  * the given block mapping.  If the shared extent overlaps the start of
162  * the mapping, trim the mapping to the end of the shared extent.  If
163  * the shared region intersects the mapping, trim the mapping to the
164  * start of the shared extent.  If there are no shared regions that
165  * overlap, just return the original extent.
166  */
167 int
168 xfs_reflink_trim_around_shared(
169 	struct xfs_inode	*ip,
170 	struct xfs_bmbt_irec	*irec,
171 	bool			*shared,
172 	bool			*trimmed)
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_reflink_inode(ip) ||
183 	    ISUNWRITTEN(irec) ||
184 	    irec->br_startblock == HOLESTARTBLOCK ||
185 	    irec->br_startblock == DELAYSTARTBLOCK) {
186 		*shared = false;
187 		return 0;
188 	}
189 
190 	trace_xfs_reflink_trim_around_shared(ip, irec);
191 
192 	agno = XFS_FSB_TO_AGNO(ip->i_mount, irec->br_startblock);
193 	agbno = XFS_FSB_TO_AGBNO(ip->i_mount, irec->br_startblock);
194 	aglen = irec->br_blockcount;
195 
196 	error = xfs_reflink_find_shared(ip->i_mount, agno, agbno,
197 			aglen, &fbno, &flen, true);
198 	if (error)
199 		return error;
200 
201 	*shared = *trimmed = false;
202 	if (fbno == NULLAGBLOCK) {
203 		/* No shared blocks at all. */
204 		return 0;
205 	} else if (fbno == agbno) {
206 		/*
207 		 * The start of this extent is shared.  Truncate the
208 		 * mapping at the end of the shared region so that a
209 		 * subsequent iteration starts at the start of the
210 		 * unshared region.
211 		 */
212 		irec->br_blockcount = flen;
213 		*shared = true;
214 		if (flen != aglen)
215 			*trimmed = true;
216 		return 0;
217 	} else {
218 		/*
219 		 * There's a shared extent midway through this extent.
220 		 * Truncate the mapping at the start of the shared
221 		 * extent so that a subsequent iteration starts at the
222 		 * start of the shared region.
223 		 */
224 		irec->br_blockcount = fbno - agbno;
225 		*trimmed = true;
226 		return 0;
227 	}
228 }
229 
230 /* Create a CoW reservation for a range of blocks within a file. */
231 static int
232 __xfs_reflink_reserve_cow(
233 	struct xfs_inode	*ip,
234 	xfs_fileoff_t		*offset_fsb,
235 	xfs_fileoff_t		end_fsb,
236 	bool			*skipped)
237 {
238 	struct xfs_bmbt_irec	got, prev, imap;
239 	xfs_fileoff_t		orig_end_fsb;
240 	int			nimaps, eof = 0, error = 0;
241 	bool			shared = false, trimmed = false;
242 	xfs_extnum_t		idx;
243 	xfs_extlen_t		align;
244 
245 	/* Already reserved?  Skip the refcount btree access. */
246 	xfs_bmap_search_extents(ip, *offset_fsb, XFS_COW_FORK, &eof, &idx,
247 			&got, &prev);
248 	if (!eof && got.br_startoff <= *offset_fsb) {
249 		end_fsb = orig_end_fsb = got.br_startoff + got.br_blockcount;
250 		trace_xfs_reflink_cow_found(ip, &got);
251 		goto done;
252 	}
253 
254 	/* Read extent from the source file. */
255 	nimaps = 1;
256 	error = xfs_bmapi_read(ip, *offset_fsb, end_fsb - *offset_fsb,
257 			&imap, &nimaps, 0);
258 	if (error)
259 		goto out_unlock;
260 	ASSERT(nimaps == 1);
261 
262 	/* Trim the mapping to the nearest shared extent boundary. */
263 	error = xfs_reflink_trim_around_shared(ip, &imap, &shared, &trimmed);
264 	if (error)
265 		goto out_unlock;
266 
267 	end_fsb = orig_end_fsb = imap.br_startoff + imap.br_blockcount;
268 
269 	/* Not shared?  Just report the (potentially capped) extent. */
270 	if (!shared) {
271 		*skipped = true;
272 		goto done;
273 	}
274 
275 	/*
276 	 * Fork all the shared blocks from our write offset until the end of
277 	 * the extent.
278 	 */
279 	error = xfs_qm_dqattach_locked(ip, 0);
280 	if (error)
281 		goto out_unlock;
282 
283 	align = xfs_eof_alignment(ip, xfs_get_cowextsz_hint(ip));
284 	if (align)
285 		end_fsb = roundup_64(end_fsb, align);
286 
287 retry:
288 	error = xfs_bmapi_reserve_delalloc(ip, XFS_COW_FORK, *offset_fsb,
289 			end_fsb - *offset_fsb, &got,
290 			&prev, &idx, eof);
291 	switch (error) {
292 	case 0:
293 		break;
294 	case -ENOSPC:
295 	case -EDQUOT:
296 		/* retry without any preallocation */
297 		trace_xfs_reflink_cow_enospc(ip, &imap);
298 		if (end_fsb != orig_end_fsb) {
299 			end_fsb = orig_end_fsb;
300 			goto retry;
301 		}
302 		/*FALLTHRU*/
303 	default:
304 		goto out_unlock;
305 	}
306 
307 	if (end_fsb != orig_end_fsb)
308 		xfs_inode_set_cowblocks_tag(ip);
309 
310 	trace_xfs_reflink_cow_alloc(ip, &got);
311 done:
312 	*offset_fsb = end_fsb;
313 out_unlock:
314 	return error;
315 }
316 
317 /* Create a CoW reservation for part of a file. */
318 int
319 xfs_reflink_reserve_cow_range(
320 	struct xfs_inode	*ip,
321 	xfs_off_t		offset,
322 	xfs_off_t		count)
323 {
324 	struct xfs_mount	*mp = ip->i_mount;
325 	xfs_fileoff_t		offset_fsb, end_fsb;
326 	bool			skipped = false;
327 	int			error;
328 
329 	trace_xfs_reflink_reserve_cow_range(ip, offset, count);
330 
331 	offset_fsb = XFS_B_TO_FSBT(mp, offset);
332 	end_fsb = XFS_B_TO_FSB(mp, offset + count);
333 
334 	xfs_ilock(ip, XFS_ILOCK_EXCL);
335 	while (offset_fsb < end_fsb) {
336 		error = __xfs_reflink_reserve_cow(ip, &offset_fsb, end_fsb,
337 				&skipped);
338 		if (error) {
339 			trace_xfs_reflink_reserve_cow_range_error(ip, error,
340 				_RET_IP_);
341 			break;
342 		}
343 	}
344 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
345 
346 	return error;
347 }
348 
349 /* Allocate all CoW reservations covering a range of blocks in a file. */
350 static int
351 __xfs_reflink_allocate_cow(
352 	struct xfs_inode	*ip,
353 	xfs_fileoff_t		*offset_fsb,
354 	xfs_fileoff_t		end_fsb)
355 {
356 	struct xfs_mount	*mp = ip->i_mount;
357 	struct xfs_bmbt_irec	imap;
358 	struct xfs_defer_ops	dfops;
359 	struct xfs_trans	*tp;
360 	xfs_fsblock_t		first_block;
361 	xfs_fileoff_t		next_fsb;
362 	int			nimaps = 1, error;
363 	bool			skipped = false;
364 
365 	xfs_defer_init(&dfops, &first_block);
366 
367 	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, 0, 0,
368 			XFS_TRANS_RESERVE, &tp);
369 	if (error)
370 		return error;
371 
372 	xfs_ilock(ip, XFS_ILOCK_EXCL);
373 
374 	next_fsb = *offset_fsb;
375 	error = __xfs_reflink_reserve_cow(ip, &next_fsb, end_fsb, &skipped);
376 	if (error)
377 		goto out_trans_cancel;
378 
379 	if (skipped) {
380 		*offset_fsb = next_fsb;
381 		goto out_trans_cancel;
382 	}
383 
384 	xfs_trans_ijoin(tp, ip, 0);
385 	error = xfs_bmapi_write(tp, ip, *offset_fsb, next_fsb - *offset_fsb,
386 			XFS_BMAPI_COWFORK, &first_block,
387 			XFS_EXTENTADD_SPACE_RES(mp, XFS_DATA_FORK),
388 			&imap, &nimaps, &dfops);
389 	if (error)
390 		goto out_trans_cancel;
391 
392 	/* We might not have been able to map the whole delalloc extent */
393 	*offset_fsb = min(*offset_fsb + imap.br_blockcount, next_fsb);
394 
395 	error = xfs_defer_finish(&tp, &dfops, NULL);
396 	if (error)
397 		goto out_trans_cancel;
398 
399 	error = xfs_trans_commit(tp);
400 
401 out_unlock:
402 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
403 	return error;
404 out_trans_cancel:
405 	xfs_defer_cancel(&dfops);
406 	xfs_trans_cancel(tp);
407 	goto out_unlock;
408 }
409 
410 /* Allocate all CoW reservations covering a part of a file. */
411 int
412 xfs_reflink_allocate_cow_range(
413 	struct xfs_inode	*ip,
414 	xfs_off_t		offset,
415 	xfs_off_t		count)
416 {
417 	struct xfs_mount	*mp = ip->i_mount;
418 	xfs_fileoff_t		offset_fsb = XFS_B_TO_FSBT(mp, offset);
419 	xfs_fileoff_t		end_fsb = XFS_B_TO_FSB(mp, offset + count);
420 	int			error;
421 
422 	ASSERT(xfs_is_reflink_inode(ip));
423 
424 	trace_xfs_reflink_allocate_cow_range(ip, offset, count);
425 
426 	/*
427 	 * Make sure that the dquots are there.
428 	 */
429 	error = xfs_qm_dqattach(ip, 0);
430 	if (error)
431 		return error;
432 
433 	while (offset_fsb < end_fsb) {
434 		error = __xfs_reflink_allocate_cow(ip, &offset_fsb, end_fsb);
435 		if (error) {
436 			trace_xfs_reflink_allocate_cow_range_error(ip, error,
437 					_RET_IP_);
438 			break;
439 		}
440 	}
441 
442 	return error;
443 }
444 
445 /*
446  * Find the CoW reservation (and whether or not it needs block allocation)
447  * for a given byte offset of a file.
448  */
449 bool
450 xfs_reflink_find_cow_mapping(
451 	struct xfs_inode		*ip,
452 	xfs_off_t			offset,
453 	struct xfs_bmbt_irec		*imap,
454 	bool				*need_alloc)
455 {
456 	struct xfs_bmbt_irec		irec;
457 	struct xfs_ifork		*ifp;
458 	struct xfs_bmbt_rec_host	*gotp;
459 	xfs_fileoff_t			bno;
460 	xfs_extnum_t			idx;
461 
462 	ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL | XFS_ILOCK_SHARED));
463 	ASSERT(xfs_is_reflink_inode(ip));
464 
465 	/* Find the extent in the CoW fork. */
466 	ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
467 	bno = XFS_B_TO_FSBT(ip->i_mount, offset);
468 	gotp = xfs_iext_bno_to_ext(ifp, bno, &idx);
469 	if (!gotp)
470 		return false;
471 
472 	xfs_bmbt_get_all(gotp, &irec);
473 	if (bno >= irec.br_startoff + irec.br_blockcount ||
474 	    bno < irec.br_startoff)
475 		return false;
476 
477 	trace_xfs_reflink_find_cow_mapping(ip, offset, 1, XFS_IO_OVERWRITE,
478 			&irec);
479 
480 	/* If it's still delalloc, we must allocate later. */
481 	*imap = irec;
482 	*need_alloc = !!(isnullstartblock(irec.br_startblock));
483 
484 	return true;
485 }
486 
487 /*
488  * Trim an extent to end at the next CoW reservation past offset_fsb.
489  */
490 int
491 xfs_reflink_trim_irec_to_next_cow(
492 	struct xfs_inode		*ip,
493 	xfs_fileoff_t			offset_fsb,
494 	struct xfs_bmbt_irec		*imap)
495 {
496 	struct xfs_bmbt_irec		irec;
497 	struct xfs_ifork		*ifp;
498 	struct xfs_bmbt_rec_host	*gotp;
499 	xfs_extnum_t			idx;
500 
501 	if (!xfs_is_reflink_inode(ip))
502 		return 0;
503 
504 	/* Find the extent in the CoW fork. */
505 	ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
506 	gotp = xfs_iext_bno_to_ext(ifp, offset_fsb, &idx);
507 	if (!gotp)
508 		return 0;
509 	xfs_bmbt_get_all(gotp, &irec);
510 
511 	/* This is the extent before; try sliding up one. */
512 	if (irec.br_startoff < offset_fsb) {
513 		idx++;
514 		if (idx >= ifp->if_bytes / sizeof(xfs_bmbt_rec_t))
515 			return 0;
516 		gotp = xfs_iext_get_ext(ifp, idx);
517 		xfs_bmbt_get_all(gotp, &irec);
518 	}
519 
520 	if (irec.br_startoff >= imap->br_startoff + imap->br_blockcount)
521 		return 0;
522 
523 	imap->br_blockcount = irec.br_startoff - imap->br_startoff;
524 	trace_xfs_reflink_trim_irec(ip, imap);
525 
526 	return 0;
527 }
528 
529 /*
530  * Cancel all pending CoW reservations for some block range of an inode.
531  */
532 int
533 xfs_reflink_cancel_cow_blocks(
534 	struct xfs_inode		*ip,
535 	struct xfs_trans		**tpp,
536 	xfs_fileoff_t			offset_fsb,
537 	xfs_fileoff_t			end_fsb)
538 {
539 	struct xfs_bmbt_irec		irec;
540 	xfs_filblks_t			count_fsb;
541 	xfs_fsblock_t			firstfsb;
542 	struct xfs_defer_ops		dfops;
543 	int				error = 0;
544 	int				nimaps;
545 
546 	if (!xfs_is_reflink_inode(ip))
547 		return 0;
548 
549 	/* Go find the old extent in the CoW fork. */
550 	while (offset_fsb < end_fsb) {
551 		nimaps = 1;
552 		count_fsb = (xfs_filblks_t)(end_fsb - offset_fsb);
553 		error = xfs_bmapi_read(ip, offset_fsb, count_fsb, &irec,
554 				&nimaps, XFS_BMAPI_COWFORK);
555 		if (error)
556 			break;
557 		ASSERT(nimaps == 1);
558 
559 		trace_xfs_reflink_cancel_cow(ip, &irec);
560 
561 		if (irec.br_startblock == DELAYSTARTBLOCK) {
562 			/* Free a delayed allocation. */
563 			xfs_mod_fdblocks(ip->i_mount, irec.br_blockcount,
564 					false);
565 			ip->i_delayed_blks -= irec.br_blockcount;
566 
567 			/* Remove the mapping from the CoW fork. */
568 			error = xfs_bunmapi_cow(ip, &irec);
569 			if (error)
570 				break;
571 		} else if (irec.br_startblock == HOLESTARTBLOCK) {
572 			/* empty */
573 		} else {
574 			xfs_trans_ijoin(*tpp, ip, 0);
575 			xfs_defer_init(&dfops, &firstfsb);
576 
577 			/* Free the CoW orphan record. */
578 			error = xfs_refcount_free_cow_extent(ip->i_mount,
579 					&dfops, irec.br_startblock,
580 					irec.br_blockcount);
581 			if (error)
582 				break;
583 
584 			xfs_bmap_add_free(ip->i_mount, &dfops,
585 					irec.br_startblock, irec.br_blockcount,
586 					NULL);
587 
588 			/* Update quota accounting */
589 			xfs_trans_mod_dquot_byino(*tpp, ip, XFS_TRANS_DQ_BCOUNT,
590 					-(long)irec.br_blockcount);
591 
592 			/* Roll the transaction */
593 			error = xfs_defer_finish(tpp, &dfops, ip);
594 			if (error) {
595 				xfs_defer_cancel(&dfops);
596 				break;
597 			}
598 
599 			/* Remove the mapping from the CoW fork. */
600 			error = xfs_bunmapi_cow(ip, &irec);
601 			if (error)
602 				break;
603 		}
604 
605 		/* Roll on... */
606 		offset_fsb = irec.br_startoff + irec.br_blockcount;
607 	}
608 
609 	return error;
610 }
611 
612 /*
613  * Cancel all pending CoW reservations for some byte range of an inode.
614  */
615 int
616 xfs_reflink_cancel_cow_range(
617 	struct xfs_inode	*ip,
618 	xfs_off_t		offset,
619 	xfs_off_t		count)
620 {
621 	struct xfs_trans	*tp;
622 	xfs_fileoff_t		offset_fsb;
623 	xfs_fileoff_t		end_fsb;
624 	int			error;
625 
626 	trace_xfs_reflink_cancel_cow_range(ip, offset, count);
627 	ASSERT(xfs_is_reflink_inode(ip));
628 
629 	offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
630 	if (count == NULLFILEOFF)
631 		end_fsb = NULLFILEOFF;
632 	else
633 		end_fsb = XFS_B_TO_FSB(ip->i_mount, offset + count);
634 
635 	/* Start a rolling transaction to remove the mappings */
636 	error = xfs_trans_alloc(ip->i_mount, &M_RES(ip->i_mount)->tr_write,
637 			0, 0, 0, &tp);
638 	if (error)
639 		goto out;
640 
641 	xfs_ilock(ip, XFS_ILOCK_EXCL);
642 	xfs_trans_ijoin(tp, ip, 0);
643 
644 	/* Scrape out the old CoW reservations */
645 	error = xfs_reflink_cancel_cow_blocks(ip, &tp, offset_fsb, end_fsb);
646 	if (error)
647 		goto out_cancel;
648 
649 	error = xfs_trans_commit(tp);
650 
651 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
652 	return error;
653 
654 out_cancel:
655 	xfs_trans_cancel(tp);
656 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
657 out:
658 	trace_xfs_reflink_cancel_cow_range_error(ip, error, _RET_IP_);
659 	return error;
660 }
661 
662 /*
663  * Remap parts of a file's data fork after a successful CoW.
664  */
665 int
666 xfs_reflink_end_cow(
667 	struct xfs_inode		*ip,
668 	xfs_off_t			offset,
669 	xfs_off_t			count)
670 {
671 	struct xfs_bmbt_irec		irec;
672 	struct xfs_bmbt_irec		uirec;
673 	struct xfs_trans		*tp;
674 	xfs_fileoff_t			offset_fsb;
675 	xfs_fileoff_t			end_fsb;
676 	xfs_filblks_t			count_fsb;
677 	xfs_fsblock_t			firstfsb;
678 	struct xfs_defer_ops		dfops;
679 	int				error;
680 	unsigned int			resblks;
681 	xfs_filblks_t			ilen;
682 	xfs_filblks_t			rlen;
683 	int				nimaps;
684 
685 	trace_xfs_reflink_end_cow(ip, offset, count);
686 
687 	offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
688 	end_fsb = XFS_B_TO_FSB(ip->i_mount, offset + count);
689 	count_fsb = (xfs_filblks_t)(end_fsb - offset_fsb);
690 
691 	/* Start a rolling transaction to switch the mappings */
692 	resblks = XFS_EXTENTADD_SPACE_RES(ip->i_mount, XFS_DATA_FORK);
693 	error = xfs_trans_alloc(ip->i_mount, &M_RES(ip->i_mount)->tr_write,
694 			resblks, 0, 0, &tp);
695 	if (error)
696 		goto out;
697 
698 	xfs_ilock(ip, XFS_ILOCK_EXCL);
699 	xfs_trans_ijoin(tp, ip, 0);
700 
701 	/* Go find the old extent in the CoW fork. */
702 	while (offset_fsb < end_fsb) {
703 		/* Read extent from the source file */
704 		nimaps = 1;
705 		count_fsb = (xfs_filblks_t)(end_fsb - offset_fsb);
706 		error = xfs_bmapi_read(ip, offset_fsb, count_fsb, &irec,
707 				&nimaps, XFS_BMAPI_COWFORK);
708 		if (error)
709 			goto out_cancel;
710 		ASSERT(nimaps == 1);
711 
712 		ASSERT(irec.br_startblock != DELAYSTARTBLOCK);
713 		trace_xfs_reflink_cow_remap(ip, &irec);
714 
715 		/*
716 		 * We can have a hole in the CoW fork if part of a directio
717 		 * write is CoW but part of it isn't.
718 		 */
719 		rlen = ilen = irec.br_blockcount;
720 		if (irec.br_startblock == HOLESTARTBLOCK)
721 			goto next_extent;
722 
723 		/* Unmap the old blocks in the data fork. */
724 		while (rlen) {
725 			xfs_defer_init(&dfops, &firstfsb);
726 			error = __xfs_bunmapi(tp, ip, irec.br_startoff,
727 					&rlen, 0, 1, &firstfsb, &dfops);
728 			if (error)
729 				goto out_defer;
730 
731 			/*
732 			 * Trim the extent to whatever got unmapped.
733 			 * Remember, bunmapi works backwards.
734 			 */
735 			uirec.br_startblock = irec.br_startblock + rlen;
736 			uirec.br_startoff = irec.br_startoff + rlen;
737 			uirec.br_blockcount = irec.br_blockcount - rlen;
738 			irec.br_blockcount = rlen;
739 			trace_xfs_reflink_cow_remap_piece(ip, &uirec);
740 
741 			/* Free the CoW orphan record. */
742 			error = xfs_refcount_free_cow_extent(tp->t_mountp,
743 					&dfops, uirec.br_startblock,
744 					uirec.br_blockcount);
745 			if (error)
746 				goto out_defer;
747 
748 			/* Map the new blocks into the data fork. */
749 			error = xfs_bmap_map_extent(tp->t_mountp, &dfops,
750 					ip, &uirec);
751 			if (error)
752 				goto out_defer;
753 
754 			/* Remove the mapping from the CoW fork. */
755 			error = xfs_bunmapi_cow(ip, &uirec);
756 			if (error)
757 				goto out_defer;
758 
759 			error = xfs_defer_finish(&tp, &dfops, ip);
760 			if (error)
761 				goto out_defer;
762 		}
763 
764 next_extent:
765 		/* Roll on... */
766 		offset_fsb = irec.br_startoff + ilen;
767 	}
768 
769 	error = xfs_trans_commit(tp);
770 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
771 	if (error)
772 		goto out;
773 	return 0;
774 
775 out_defer:
776 	xfs_defer_cancel(&dfops);
777 out_cancel:
778 	xfs_trans_cancel(tp);
779 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
780 out:
781 	trace_xfs_reflink_end_cow_error(ip, error, _RET_IP_);
782 	return error;
783 }
784 
785 /*
786  * Free leftover CoW reservations that didn't get cleaned out.
787  */
788 int
789 xfs_reflink_recover_cow(
790 	struct xfs_mount	*mp)
791 {
792 	xfs_agnumber_t		agno;
793 	int			error = 0;
794 
795 	if (!xfs_sb_version_hasreflink(&mp->m_sb))
796 		return 0;
797 
798 	for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
799 		error = xfs_refcount_recover_cow_leftovers(mp, agno);
800 		if (error)
801 			break;
802 	}
803 
804 	return error;
805 }
806 
807 /*
808  * Reflinking (Block) Ranges of Two Files Together
809  *
810  * First, ensure that the reflink flag is set on both inodes.  The flag is an
811  * optimization to avoid unnecessary refcount btree lookups in the write path.
812  *
813  * Now we can iteratively remap the range of extents (and holes) in src to the
814  * corresponding ranges in dest.  Let drange and srange denote the ranges of
815  * logical blocks in dest and src touched by the reflink operation.
816  *
817  * While the length of drange is greater than zero,
818  *    - Read src's bmbt at the start of srange ("imap")
819  *    - If imap doesn't exist, make imap appear to start at the end of srange
820  *      with zero length.
821  *    - If imap starts before srange, advance imap to start at srange.
822  *    - If imap goes beyond srange, truncate imap to end at the end of srange.
823  *    - Punch (imap start - srange start + imap len) blocks from dest at
824  *      offset (drange start).
825  *    - If imap points to a real range of pblks,
826  *         > Increase the refcount of the imap's pblks
827  *         > Map imap's pblks into dest at the offset
828  *           (drange start + imap start - srange start)
829  *    - Advance drange and srange by (imap start - srange start + imap len)
830  *
831  * Finally, if the reflink made dest longer, update both the in-core and
832  * on-disk file sizes.
833  *
834  * ASCII Art Demonstration:
835  *
836  * Let's say we want to reflink this source file:
837  *
838  * ----SSSSSSS-SSSSS----SSSSSS (src file)
839  *   <-------------------->
840  *
841  * into this destination file:
842  *
843  * --DDDDDDDDDDDDDDDDDDD--DDD (dest file)
844  *        <-------------------->
845  * '-' means a hole, and 'S' and 'D' are written blocks in the src and dest.
846  * Observe that the range has different logical offsets in either file.
847  *
848  * Consider that the first extent in the source file doesn't line up with our
849  * reflink range.  Unmapping  and remapping are separate operations, so we can
850  * unmap more blocks from the destination file than we remap.
851  *
852  * ----SSSSSSS-SSSSS----SSSSSS
853  *   <------->
854  * --DDDDD---------DDDDD--DDD
855  *        <------->
856  *
857  * Now remap the source extent into the destination file:
858  *
859  * ----SSSSSSS-SSSSS----SSSSSS
860  *   <------->
861  * --DDDDD--SSSSSSSDDDDD--DDD
862  *        <------->
863  *
864  * Do likewise with the second hole and extent in our range.  Holes in the
865  * unmap range don't affect our operation.
866  *
867  * ----SSSSSSS-SSSSS----SSSSSS
868  *            <---->
869  * --DDDDD--SSSSSSS-SSSSS-DDD
870  *                 <---->
871  *
872  * Finally, unmap and remap part of the third extent.  This will increase the
873  * size of the destination file.
874  *
875  * ----SSSSSSS-SSSSS----SSSSSS
876  *                  <----->
877  * --DDDDD--SSSSSSS-SSSSS----SSS
878  *                       <----->
879  *
880  * Once we update the destination file's i_size, we're done.
881  */
882 
883 /*
884  * Ensure the reflink bit is set in both inodes.
885  */
886 STATIC int
887 xfs_reflink_set_inode_flag(
888 	struct xfs_inode	*src,
889 	struct xfs_inode	*dest)
890 {
891 	struct xfs_mount	*mp = src->i_mount;
892 	int			error;
893 	struct xfs_trans	*tp;
894 
895 	if (xfs_is_reflink_inode(src) && xfs_is_reflink_inode(dest))
896 		return 0;
897 
898 	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
899 	if (error)
900 		goto out_error;
901 
902 	/* Lock both files against IO */
903 	if (src->i_ino == dest->i_ino)
904 		xfs_ilock(src, XFS_ILOCK_EXCL);
905 	else
906 		xfs_lock_two_inodes(src, dest, XFS_ILOCK_EXCL);
907 
908 	if (!xfs_is_reflink_inode(src)) {
909 		trace_xfs_reflink_set_inode_flag(src);
910 		xfs_trans_ijoin(tp, src, XFS_ILOCK_EXCL);
911 		src->i_d.di_flags2 |= XFS_DIFLAG2_REFLINK;
912 		xfs_trans_log_inode(tp, src, XFS_ILOG_CORE);
913 		xfs_ifork_init_cow(src);
914 	} else
915 		xfs_iunlock(src, XFS_ILOCK_EXCL);
916 
917 	if (src->i_ino == dest->i_ino)
918 		goto commit_flags;
919 
920 	if (!xfs_is_reflink_inode(dest)) {
921 		trace_xfs_reflink_set_inode_flag(dest);
922 		xfs_trans_ijoin(tp, dest, XFS_ILOCK_EXCL);
923 		dest->i_d.di_flags2 |= XFS_DIFLAG2_REFLINK;
924 		xfs_trans_log_inode(tp, dest, XFS_ILOG_CORE);
925 		xfs_ifork_init_cow(dest);
926 	} else
927 		xfs_iunlock(dest, XFS_ILOCK_EXCL);
928 
929 commit_flags:
930 	error = xfs_trans_commit(tp);
931 	if (error)
932 		goto out_error;
933 	return error;
934 
935 out_error:
936 	trace_xfs_reflink_set_inode_flag_error(dest, error, _RET_IP_);
937 	return error;
938 }
939 
940 /*
941  * Update destination inode size & cowextsize hint, if necessary.
942  */
943 STATIC int
944 xfs_reflink_update_dest(
945 	struct xfs_inode	*dest,
946 	xfs_off_t		newlen,
947 	xfs_extlen_t		cowextsize)
948 {
949 	struct xfs_mount	*mp = dest->i_mount;
950 	struct xfs_trans	*tp;
951 	int			error;
952 
953 	if (newlen <= i_size_read(VFS_I(dest)) && cowextsize == 0)
954 		return 0;
955 
956 	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
957 	if (error)
958 		goto out_error;
959 
960 	xfs_ilock(dest, XFS_ILOCK_EXCL);
961 	xfs_trans_ijoin(tp, dest, XFS_ILOCK_EXCL);
962 
963 	if (newlen > i_size_read(VFS_I(dest))) {
964 		trace_xfs_reflink_update_inode_size(dest, newlen);
965 		i_size_write(VFS_I(dest), newlen);
966 		dest->i_d.di_size = newlen;
967 	}
968 
969 	if (cowextsize) {
970 		dest->i_d.di_cowextsize = cowextsize;
971 		dest->i_d.di_flags2 |= XFS_DIFLAG2_COWEXTSIZE;
972 	}
973 
974 	xfs_trans_log_inode(tp, dest, XFS_ILOG_CORE);
975 
976 	error = xfs_trans_commit(tp);
977 	if (error)
978 		goto out_error;
979 	return error;
980 
981 out_error:
982 	trace_xfs_reflink_update_inode_size_error(dest, error, _RET_IP_);
983 	return error;
984 }
985 
986 /*
987  * Do we have enough reserve in this AG to handle a reflink?  The refcount
988  * btree already reserved all the space it needs, but the rmap btree can grow
989  * infinitely, so we won't allow more reflinks when the AG is down to the
990  * btree reserves.
991  */
992 static int
993 xfs_reflink_ag_has_free_space(
994 	struct xfs_mount	*mp,
995 	xfs_agnumber_t		agno)
996 {
997 	struct xfs_perag	*pag;
998 	int			error = 0;
999 
1000 	if (!xfs_sb_version_hasrmapbt(&mp->m_sb))
1001 		return 0;
1002 
1003 	pag = xfs_perag_get(mp, agno);
1004 	if (xfs_ag_resv_critical(pag, XFS_AG_RESV_AGFL) ||
1005 	    xfs_ag_resv_critical(pag, XFS_AG_RESV_METADATA))
1006 		error = -ENOSPC;
1007 	xfs_perag_put(pag);
1008 	return error;
1009 }
1010 
1011 /*
1012  * Unmap a range of blocks from a file, then map other blocks into the hole.
1013  * The range to unmap is (destoff : destoff + srcioff + irec->br_blockcount).
1014  * The extent irec is mapped into dest at irec->br_startoff.
1015  */
1016 STATIC int
1017 xfs_reflink_remap_extent(
1018 	struct xfs_inode	*ip,
1019 	struct xfs_bmbt_irec	*irec,
1020 	xfs_fileoff_t		destoff,
1021 	xfs_off_t		new_isize)
1022 {
1023 	struct xfs_mount	*mp = ip->i_mount;
1024 	struct xfs_trans	*tp;
1025 	xfs_fsblock_t		firstfsb;
1026 	unsigned int		resblks;
1027 	struct xfs_defer_ops	dfops;
1028 	struct xfs_bmbt_irec	uirec;
1029 	bool			real_extent;
1030 	xfs_filblks_t		rlen;
1031 	xfs_filblks_t		unmap_len;
1032 	xfs_off_t		newlen;
1033 	int			error;
1034 
1035 	unmap_len = irec->br_startoff + irec->br_blockcount - destoff;
1036 	trace_xfs_reflink_punch_range(ip, destoff, unmap_len);
1037 
1038 	/* Only remap normal extents. */
1039 	real_extent =  (irec->br_startblock != HOLESTARTBLOCK &&
1040 			irec->br_startblock != DELAYSTARTBLOCK &&
1041 			!ISUNWRITTEN(irec));
1042 
1043 	/* No reflinking if we're low on space */
1044 	if (real_extent) {
1045 		error = xfs_reflink_ag_has_free_space(mp,
1046 				XFS_FSB_TO_AGNO(mp, irec->br_startblock));
1047 		if (error)
1048 			goto out;
1049 	}
1050 
1051 	/* Start a rolling transaction to switch the mappings */
1052 	resblks = XFS_EXTENTADD_SPACE_RES(ip->i_mount, XFS_DATA_FORK);
1053 	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, resblks, 0, 0, &tp);
1054 	if (error)
1055 		goto out;
1056 
1057 	xfs_ilock(ip, XFS_ILOCK_EXCL);
1058 	xfs_trans_ijoin(tp, ip, 0);
1059 
1060 	/* If we're not just clearing space, then do we have enough quota? */
1061 	if (real_extent) {
1062 		error = xfs_trans_reserve_quota_nblks(tp, ip,
1063 				irec->br_blockcount, 0, XFS_QMOPT_RES_REGBLKS);
1064 		if (error)
1065 			goto out_cancel;
1066 	}
1067 
1068 	trace_xfs_reflink_remap(ip, irec->br_startoff,
1069 				irec->br_blockcount, irec->br_startblock);
1070 
1071 	/* Unmap the old blocks in the data fork. */
1072 	rlen = unmap_len;
1073 	while (rlen) {
1074 		xfs_defer_init(&dfops, &firstfsb);
1075 		error = __xfs_bunmapi(tp, ip, destoff, &rlen, 0, 1,
1076 				&firstfsb, &dfops);
1077 		if (error)
1078 			goto out_defer;
1079 
1080 		/*
1081 		 * Trim the extent to whatever got unmapped.
1082 		 * Remember, bunmapi works backwards.
1083 		 */
1084 		uirec.br_startblock = irec->br_startblock + rlen;
1085 		uirec.br_startoff = irec->br_startoff + rlen;
1086 		uirec.br_blockcount = unmap_len - rlen;
1087 		unmap_len = rlen;
1088 
1089 		/* If this isn't a real mapping, we're done. */
1090 		if (!real_extent || uirec.br_blockcount == 0)
1091 			goto next_extent;
1092 
1093 		trace_xfs_reflink_remap(ip, uirec.br_startoff,
1094 				uirec.br_blockcount, uirec.br_startblock);
1095 
1096 		/* Update the refcount tree */
1097 		error = xfs_refcount_increase_extent(mp, &dfops, &uirec);
1098 		if (error)
1099 			goto out_defer;
1100 
1101 		/* Map the new blocks into the data fork. */
1102 		error = xfs_bmap_map_extent(mp, &dfops, ip, &uirec);
1103 		if (error)
1104 			goto out_defer;
1105 
1106 		/* Update quota accounting. */
1107 		xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_BCOUNT,
1108 				uirec.br_blockcount);
1109 
1110 		/* Update dest isize if needed. */
1111 		newlen = XFS_FSB_TO_B(mp,
1112 				uirec.br_startoff + uirec.br_blockcount);
1113 		newlen = min_t(xfs_off_t, newlen, new_isize);
1114 		if (newlen > i_size_read(VFS_I(ip))) {
1115 			trace_xfs_reflink_update_inode_size(ip, newlen);
1116 			i_size_write(VFS_I(ip), newlen);
1117 			ip->i_d.di_size = newlen;
1118 			xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1119 		}
1120 
1121 next_extent:
1122 		/* Process all the deferred stuff. */
1123 		error = xfs_defer_finish(&tp, &dfops, ip);
1124 		if (error)
1125 			goto out_defer;
1126 	}
1127 
1128 	error = xfs_trans_commit(tp);
1129 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1130 	if (error)
1131 		goto out;
1132 	return 0;
1133 
1134 out_defer:
1135 	xfs_defer_cancel(&dfops);
1136 out_cancel:
1137 	xfs_trans_cancel(tp);
1138 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1139 out:
1140 	trace_xfs_reflink_remap_extent_error(ip, error, _RET_IP_);
1141 	return error;
1142 }
1143 
1144 /*
1145  * Iteratively remap one file's extents (and holes) to another's.
1146  */
1147 STATIC int
1148 xfs_reflink_remap_blocks(
1149 	struct xfs_inode	*src,
1150 	xfs_fileoff_t		srcoff,
1151 	struct xfs_inode	*dest,
1152 	xfs_fileoff_t		destoff,
1153 	xfs_filblks_t		len,
1154 	xfs_off_t		new_isize)
1155 {
1156 	struct xfs_bmbt_irec	imap;
1157 	int			nimaps;
1158 	int			error = 0;
1159 	xfs_filblks_t		range_len;
1160 
1161 	/* drange = (destoff, destoff + len); srange = (srcoff, srcoff + len) */
1162 	while (len) {
1163 		trace_xfs_reflink_remap_blocks_loop(src, srcoff, len,
1164 				dest, destoff);
1165 		/* Read extent from the source file */
1166 		nimaps = 1;
1167 		xfs_ilock(src, XFS_ILOCK_EXCL);
1168 		error = xfs_bmapi_read(src, srcoff, len, &imap, &nimaps, 0);
1169 		xfs_iunlock(src, XFS_ILOCK_EXCL);
1170 		if (error)
1171 			goto err;
1172 		ASSERT(nimaps == 1);
1173 
1174 		trace_xfs_reflink_remap_imap(src, srcoff, len, XFS_IO_OVERWRITE,
1175 				&imap);
1176 
1177 		/* Translate imap into the destination file. */
1178 		range_len = imap.br_startoff + imap.br_blockcount - srcoff;
1179 		imap.br_startoff += destoff - srcoff;
1180 
1181 		/* Clear dest from destoff to the end of imap and map it in. */
1182 		error = xfs_reflink_remap_extent(dest, &imap, destoff,
1183 				new_isize);
1184 		if (error)
1185 			goto err;
1186 
1187 		if (fatal_signal_pending(current)) {
1188 			error = -EINTR;
1189 			goto err;
1190 		}
1191 
1192 		/* Advance drange/srange */
1193 		srcoff += range_len;
1194 		destoff += range_len;
1195 		len -= range_len;
1196 	}
1197 
1198 	return 0;
1199 
1200 err:
1201 	trace_xfs_reflink_remap_blocks_error(dest, error, _RET_IP_);
1202 	return error;
1203 }
1204 
1205 /*
1206  * Read a page's worth of file data into the page cache.  Return the page
1207  * locked.
1208  */
1209 static struct page *
1210 xfs_get_page(
1211 	struct inode	*inode,
1212 	xfs_off_t	offset)
1213 {
1214 	struct address_space	*mapping;
1215 	struct page		*page;
1216 	pgoff_t			n;
1217 
1218 	n = offset >> PAGE_SHIFT;
1219 	mapping = inode->i_mapping;
1220 	page = read_mapping_page(mapping, n, NULL);
1221 	if (IS_ERR(page))
1222 		return page;
1223 	if (!PageUptodate(page)) {
1224 		put_page(page);
1225 		return ERR_PTR(-EIO);
1226 	}
1227 	lock_page(page);
1228 	return page;
1229 }
1230 
1231 /*
1232  * Compare extents of two files to see if they are the same.
1233  */
1234 static int
1235 xfs_compare_extents(
1236 	struct inode	*src,
1237 	xfs_off_t	srcoff,
1238 	struct inode	*dest,
1239 	xfs_off_t	destoff,
1240 	xfs_off_t	len,
1241 	bool		*is_same)
1242 {
1243 	xfs_off_t	src_poff;
1244 	xfs_off_t	dest_poff;
1245 	void		*src_addr;
1246 	void		*dest_addr;
1247 	struct page	*src_page;
1248 	struct page	*dest_page;
1249 	xfs_off_t	cmp_len;
1250 	bool		same;
1251 	int		error;
1252 
1253 	error = -EINVAL;
1254 	same = true;
1255 	while (len) {
1256 		src_poff = srcoff & (PAGE_SIZE - 1);
1257 		dest_poff = destoff & (PAGE_SIZE - 1);
1258 		cmp_len = min(PAGE_SIZE - src_poff,
1259 			      PAGE_SIZE - dest_poff);
1260 		cmp_len = min(cmp_len, len);
1261 		ASSERT(cmp_len > 0);
1262 
1263 		trace_xfs_reflink_compare_extents(XFS_I(src), srcoff, cmp_len,
1264 				XFS_I(dest), destoff);
1265 
1266 		src_page = xfs_get_page(src, srcoff);
1267 		if (IS_ERR(src_page)) {
1268 			error = PTR_ERR(src_page);
1269 			goto out_error;
1270 		}
1271 		dest_page = xfs_get_page(dest, destoff);
1272 		if (IS_ERR(dest_page)) {
1273 			error = PTR_ERR(dest_page);
1274 			unlock_page(src_page);
1275 			put_page(src_page);
1276 			goto out_error;
1277 		}
1278 		src_addr = kmap_atomic(src_page);
1279 		dest_addr = kmap_atomic(dest_page);
1280 
1281 		flush_dcache_page(src_page);
1282 		flush_dcache_page(dest_page);
1283 
1284 		if (memcmp(src_addr + src_poff, dest_addr + dest_poff, cmp_len))
1285 			same = false;
1286 
1287 		kunmap_atomic(dest_addr);
1288 		kunmap_atomic(src_addr);
1289 		unlock_page(dest_page);
1290 		unlock_page(src_page);
1291 		put_page(dest_page);
1292 		put_page(src_page);
1293 
1294 		if (!same)
1295 			break;
1296 
1297 		srcoff += cmp_len;
1298 		destoff += cmp_len;
1299 		len -= cmp_len;
1300 	}
1301 
1302 	*is_same = same;
1303 	return 0;
1304 
1305 out_error:
1306 	trace_xfs_reflink_compare_extents_error(XFS_I(dest), error, _RET_IP_);
1307 	return error;
1308 }
1309 
1310 /*
1311  * Link a range of blocks from one file to another.
1312  */
1313 int
1314 xfs_reflink_remap_range(
1315 	struct xfs_inode	*src,
1316 	xfs_off_t		srcoff,
1317 	struct xfs_inode	*dest,
1318 	xfs_off_t		destoff,
1319 	xfs_off_t		len,
1320 	unsigned int		flags)
1321 {
1322 	struct xfs_mount	*mp = src->i_mount;
1323 	xfs_fileoff_t		sfsbno, dfsbno;
1324 	xfs_filblks_t		fsblen;
1325 	int			error;
1326 	xfs_extlen_t		cowextsize;
1327 	bool			is_same;
1328 
1329 	if (!xfs_sb_version_hasreflink(&mp->m_sb))
1330 		return -EOPNOTSUPP;
1331 
1332 	if (XFS_FORCED_SHUTDOWN(mp))
1333 		return -EIO;
1334 
1335 	/* Don't reflink realtime inodes */
1336 	if (XFS_IS_REALTIME_INODE(src) || XFS_IS_REALTIME_INODE(dest))
1337 		return -EINVAL;
1338 
1339 	if (flags & ~XFS_REFLINK_ALL)
1340 		return -EINVAL;
1341 
1342 	trace_xfs_reflink_remap_range(src, srcoff, len, dest, destoff);
1343 
1344 	/* Lock both files against IO */
1345 	if (src->i_ino == dest->i_ino) {
1346 		xfs_ilock(src, XFS_IOLOCK_EXCL);
1347 		xfs_ilock(src, XFS_MMAPLOCK_EXCL);
1348 	} else {
1349 		xfs_lock_two_inodes(src, dest, XFS_IOLOCK_EXCL);
1350 		xfs_lock_two_inodes(src, dest, XFS_MMAPLOCK_EXCL);
1351 	}
1352 
1353 	/*
1354 	 * Check that the extents are the same.
1355 	 */
1356 	if (flags & XFS_REFLINK_DEDUPE) {
1357 		is_same = false;
1358 		error = xfs_compare_extents(VFS_I(src), srcoff, VFS_I(dest),
1359 				destoff, len, &is_same);
1360 		if (error)
1361 			goto out_error;
1362 		if (!is_same) {
1363 			error = -EBADE;
1364 			goto out_error;
1365 		}
1366 	}
1367 
1368 	error = xfs_reflink_set_inode_flag(src, dest);
1369 	if (error)
1370 		goto out_error;
1371 
1372 	/*
1373 	 * Invalidate the page cache so that we can clear any CoW mappings
1374 	 * in the destination file.
1375 	 */
1376 	truncate_inode_pages_range(&VFS_I(dest)->i_data, destoff,
1377 				   PAGE_ALIGN(destoff + len) - 1);
1378 
1379 	dfsbno = XFS_B_TO_FSBT(mp, destoff);
1380 	sfsbno = XFS_B_TO_FSBT(mp, srcoff);
1381 	fsblen = XFS_B_TO_FSB(mp, len);
1382 	error = xfs_reflink_remap_blocks(src, sfsbno, dest, dfsbno, fsblen,
1383 			destoff + len);
1384 	if (error)
1385 		goto out_error;
1386 
1387 	/*
1388 	 * Carry the cowextsize hint from src to dest if we're sharing the
1389 	 * entire source file to the entire destination file, the source file
1390 	 * has a cowextsize hint, and the destination file does not.
1391 	 */
1392 	cowextsize = 0;
1393 	if (srcoff == 0 && len == i_size_read(VFS_I(src)) &&
1394 	    (src->i_d.di_flags2 & XFS_DIFLAG2_COWEXTSIZE) &&
1395 	    destoff == 0 && len >= i_size_read(VFS_I(dest)) &&
1396 	    !(dest->i_d.di_flags2 & XFS_DIFLAG2_COWEXTSIZE))
1397 		cowextsize = src->i_d.di_cowextsize;
1398 
1399 	error = xfs_reflink_update_dest(dest, destoff + len, cowextsize);
1400 	if (error)
1401 		goto out_error;
1402 
1403 out_error:
1404 	xfs_iunlock(src, XFS_MMAPLOCK_EXCL);
1405 	xfs_iunlock(src, XFS_IOLOCK_EXCL);
1406 	if (src->i_ino != dest->i_ino) {
1407 		xfs_iunlock(dest, XFS_MMAPLOCK_EXCL);
1408 		xfs_iunlock(dest, XFS_IOLOCK_EXCL);
1409 	}
1410 	if (error)
1411 		trace_xfs_reflink_remap_range_error(dest, error, _RET_IP_);
1412 	return error;
1413 }
1414 
1415 /*
1416  * The user wants to preemptively CoW all shared blocks in this file,
1417  * which enables us to turn off the reflink flag.  Iterate all
1418  * extents which are not prealloc/delalloc to see which ranges are
1419  * mentioned in the refcount tree, then read those blocks into the
1420  * pagecache, dirty them, fsync them back out, and then we can update
1421  * the inode flag.  What happens if we run out of memory? :)
1422  */
1423 STATIC int
1424 xfs_reflink_dirty_extents(
1425 	struct xfs_inode	*ip,
1426 	xfs_fileoff_t		fbno,
1427 	xfs_filblks_t		end,
1428 	xfs_off_t		isize)
1429 {
1430 	struct xfs_mount	*mp = ip->i_mount;
1431 	xfs_agnumber_t		agno;
1432 	xfs_agblock_t		agbno;
1433 	xfs_extlen_t		aglen;
1434 	xfs_agblock_t		rbno;
1435 	xfs_extlen_t		rlen;
1436 	xfs_off_t		fpos;
1437 	xfs_off_t		flen;
1438 	struct xfs_bmbt_irec	map[2];
1439 	int			nmaps;
1440 	int			error = 0;
1441 
1442 	while (end - fbno > 0) {
1443 		nmaps = 1;
1444 		/*
1445 		 * Look for extents in the file.  Skip holes, delalloc, or
1446 		 * unwritten extents; they can't be reflinked.
1447 		 */
1448 		error = xfs_bmapi_read(ip, fbno, end - fbno, map, &nmaps, 0);
1449 		if (error)
1450 			goto out;
1451 		if (nmaps == 0)
1452 			break;
1453 		if (map[0].br_startblock == HOLESTARTBLOCK ||
1454 		    map[0].br_startblock == DELAYSTARTBLOCK ||
1455 		    ISUNWRITTEN(&map[0]))
1456 			goto next;
1457 
1458 		map[1] = map[0];
1459 		while (map[1].br_blockcount) {
1460 			agno = XFS_FSB_TO_AGNO(mp, map[1].br_startblock);
1461 			agbno = XFS_FSB_TO_AGBNO(mp, map[1].br_startblock);
1462 			aglen = map[1].br_blockcount;
1463 
1464 			error = xfs_reflink_find_shared(mp, agno, agbno, aglen,
1465 					&rbno, &rlen, true);
1466 			if (error)
1467 				goto out;
1468 			if (rbno == NULLAGBLOCK)
1469 				break;
1470 
1471 			/* Dirty the pages */
1472 			xfs_iunlock(ip, XFS_ILOCK_EXCL);
1473 			fpos = XFS_FSB_TO_B(mp, map[1].br_startoff +
1474 					(rbno - agbno));
1475 			flen = XFS_FSB_TO_B(mp, rlen);
1476 			if (fpos + flen > isize)
1477 				flen = isize - fpos;
1478 			error = iomap_file_dirty(VFS_I(ip), fpos, flen,
1479 					&xfs_iomap_ops);
1480 			xfs_ilock(ip, XFS_ILOCK_EXCL);
1481 			if (error)
1482 				goto out;
1483 
1484 			map[1].br_blockcount -= (rbno - agbno + rlen);
1485 			map[1].br_startoff += (rbno - agbno + rlen);
1486 			map[1].br_startblock += (rbno - agbno + rlen);
1487 		}
1488 
1489 next:
1490 		fbno = map[0].br_startoff + map[0].br_blockcount;
1491 	}
1492 out:
1493 	return error;
1494 }
1495 
1496 /* Clear the inode reflink flag if there are no shared extents. */
1497 int
1498 xfs_reflink_clear_inode_flag(
1499 	struct xfs_inode	*ip,
1500 	struct xfs_trans	**tpp)
1501 {
1502 	struct xfs_mount	*mp = ip->i_mount;
1503 	xfs_fileoff_t		fbno;
1504 	xfs_filblks_t		end;
1505 	xfs_agnumber_t		agno;
1506 	xfs_agblock_t		agbno;
1507 	xfs_extlen_t		aglen;
1508 	xfs_agblock_t		rbno;
1509 	xfs_extlen_t		rlen;
1510 	struct xfs_bmbt_irec	map;
1511 	int			nmaps;
1512 	int			error = 0;
1513 
1514 	ASSERT(xfs_is_reflink_inode(ip));
1515 
1516 	fbno = 0;
1517 	end = XFS_B_TO_FSB(mp, i_size_read(VFS_I(ip)));
1518 	while (end - fbno > 0) {
1519 		nmaps = 1;
1520 		/*
1521 		 * Look for extents in the file.  Skip holes, delalloc, or
1522 		 * unwritten extents; they can't be reflinked.
1523 		 */
1524 		error = xfs_bmapi_read(ip, fbno, end - fbno, &map, &nmaps, 0);
1525 		if (error)
1526 			return error;
1527 		if (nmaps == 0)
1528 			break;
1529 		if (map.br_startblock == HOLESTARTBLOCK ||
1530 		    map.br_startblock == DELAYSTARTBLOCK ||
1531 		    ISUNWRITTEN(&map))
1532 			goto next;
1533 
1534 		agno = XFS_FSB_TO_AGNO(mp, map.br_startblock);
1535 		agbno = XFS_FSB_TO_AGBNO(mp, map.br_startblock);
1536 		aglen = map.br_blockcount;
1537 
1538 		error = xfs_reflink_find_shared(mp, agno, agbno, aglen,
1539 				&rbno, &rlen, false);
1540 		if (error)
1541 			return error;
1542 		/* Is there still a shared block here? */
1543 		if (rbno != NULLAGBLOCK)
1544 			return 0;
1545 next:
1546 		fbno = map.br_startoff + map.br_blockcount;
1547 	}
1548 
1549 	/*
1550 	 * We didn't find any shared blocks so turn off the reflink flag.
1551 	 * First, get rid of any leftover CoW mappings.
1552 	 */
1553 	error = xfs_reflink_cancel_cow_blocks(ip, tpp, 0, NULLFILEOFF);
1554 	if (error)
1555 		return error;
1556 
1557 	/* Clear the inode flag. */
1558 	trace_xfs_reflink_unset_inode_flag(ip);
1559 	ip->i_d.di_flags2 &= ~XFS_DIFLAG2_REFLINK;
1560 	xfs_inode_clear_cowblocks_tag(ip);
1561 	xfs_trans_ijoin(*tpp, ip, 0);
1562 	xfs_trans_log_inode(*tpp, ip, XFS_ILOG_CORE);
1563 
1564 	return error;
1565 }
1566 
1567 /*
1568  * Clear the inode reflink flag if there are no shared extents and the size
1569  * hasn't changed.
1570  */
1571 STATIC int
1572 xfs_reflink_try_clear_inode_flag(
1573 	struct xfs_inode	*ip)
1574 {
1575 	struct xfs_mount	*mp = ip->i_mount;
1576 	struct xfs_trans	*tp;
1577 	int			error = 0;
1578 
1579 	/* Start a rolling transaction to remove the mappings */
1580 	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, 0, 0, 0, &tp);
1581 	if (error)
1582 		return error;
1583 
1584 	xfs_ilock(ip, XFS_ILOCK_EXCL);
1585 	xfs_trans_ijoin(tp, ip, 0);
1586 
1587 	error = xfs_reflink_clear_inode_flag(ip, &tp);
1588 	if (error)
1589 		goto cancel;
1590 
1591 	error = xfs_trans_commit(tp);
1592 	if (error)
1593 		goto out;
1594 
1595 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1596 	return 0;
1597 cancel:
1598 	xfs_trans_cancel(tp);
1599 out:
1600 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1601 	return error;
1602 }
1603 
1604 /*
1605  * Pre-COW all shared blocks within a given byte range of a file and turn off
1606  * the reflink flag if we unshare all of the file's blocks.
1607  */
1608 int
1609 xfs_reflink_unshare(
1610 	struct xfs_inode	*ip,
1611 	xfs_off_t		offset,
1612 	xfs_off_t		len)
1613 {
1614 	struct xfs_mount	*mp = ip->i_mount;
1615 	xfs_fileoff_t		fbno;
1616 	xfs_filblks_t		end;
1617 	xfs_off_t		isize;
1618 	int			error;
1619 
1620 	if (!xfs_is_reflink_inode(ip))
1621 		return 0;
1622 
1623 	trace_xfs_reflink_unshare(ip, offset, len);
1624 
1625 	inode_dio_wait(VFS_I(ip));
1626 
1627 	/* Try to CoW the selected ranges */
1628 	xfs_ilock(ip, XFS_ILOCK_EXCL);
1629 	fbno = XFS_B_TO_FSBT(mp, offset);
1630 	isize = i_size_read(VFS_I(ip));
1631 	end = XFS_B_TO_FSB(mp, offset + len);
1632 	error = xfs_reflink_dirty_extents(ip, fbno, end, isize);
1633 	if (error)
1634 		goto out_unlock;
1635 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1636 
1637 	/* Wait for the IO to finish */
1638 	error = filemap_write_and_wait(VFS_I(ip)->i_mapping);
1639 	if (error)
1640 		goto out;
1641 
1642 	/* Turn off the reflink flag if possible. */
1643 	error = xfs_reflink_try_clear_inode_flag(ip);
1644 	if (error)
1645 		goto out;
1646 
1647 	return 0;
1648 
1649 out_unlock:
1650 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1651 out:
1652 	trace_xfs_reflink_unshare_error(ip, error, _RET_IP_);
1653 	return error;
1654 }
1655 
1656 /*
1657  * Does this inode have any real CoW reservations?
1658  */
1659 bool
1660 xfs_reflink_has_real_cow_blocks(
1661 	struct xfs_inode		*ip)
1662 {
1663 	struct xfs_bmbt_irec		irec;
1664 	struct xfs_ifork		*ifp;
1665 	struct xfs_bmbt_rec_host	*gotp;
1666 	xfs_extnum_t			idx;
1667 
1668 	if (!xfs_is_reflink_inode(ip))
1669 		return false;
1670 
1671 	/* Go find the old extent in the CoW fork. */
1672 	ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
1673 	gotp = xfs_iext_bno_to_ext(ifp, 0, &idx);
1674 	while (gotp) {
1675 		xfs_bmbt_get_all(gotp, &irec);
1676 
1677 		if (!isnullstartblock(irec.br_startblock))
1678 			return true;
1679 
1680 		/* Roll on... */
1681 		idx++;
1682 		if (idx >= ifp->if_bytes / sizeof(xfs_bmbt_rec_t))
1683 			break;
1684 		gotp = xfs_iext_get_ext(ifp, idx);
1685 	}
1686 
1687 	return false;
1688 }
1689