xref: /linux/fs/xfs/xfs_reflink.c (revision fd639726bf15fca8ee1a00dce8e0096d0ad9bd18)
1 /*
2  * Copyright (C) 2016 Oracle.  All Rights Reserved.
3  *
4  * Author: Darrick J. Wong <darrick.wong@oracle.com>
5  *
6  * This program is free software; you can redistribute it and/or
7  * modify it under the terms of the GNU General Public License
8  * as published by the Free Software Foundation; either version 2
9  * of the License, or (at your option) any later version.
10  *
11  * This program is distributed in the hope that it would be useful,
12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
14  * GNU General Public License for more details.
15  *
16  * You should have received a copy of the GNU General Public License
17  * along with this program; if not, write the Free Software Foundation,
18  * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301, USA.
19  */
20 #include "xfs.h"
21 #include "xfs_fs.h"
22 #include "xfs_shared.h"
23 #include "xfs_format.h"
24 #include "xfs_log_format.h"
25 #include "xfs_trans_resv.h"
26 #include "xfs_mount.h"
27 #include "xfs_defer.h"
28 #include "xfs_da_format.h"
29 #include "xfs_da_btree.h"
30 #include "xfs_inode.h"
31 #include "xfs_trans.h"
32 #include "xfs_inode_item.h"
33 #include "xfs_bmap.h"
34 #include "xfs_bmap_util.h"
35 #include "xfs_error.h"
36 #include "xfs_dir2.h"
37 #include "xfs_dir2_priv.h"
38 #include "xfs_ioctl.h"
39 #include "xfs_trace.h"
40 #include "xfs_log.h"
41 #include "xfs_icache.h"
42 #include "xfs_pnfs.h"
43 #include "xfs_btree.h"
44 #include "xfs_refcount_btree.h"
45 #include "xfs_refcount.h"
46 #include "xfs_bmap_btree.h"
47 #include "xfs_trans_space.h"
48 #include "xfs_bit.h"
49 #include "xfs_alloc.h"
50 #include "xfs_quota_defs.h"
51 #include "xfs_quota.h"
52 #include "xfs_btree.h"
53 #include "xfs_bmap_btree.h"
54 #include "xfs_reflink.h"
55 #include "xfs_iomap.h"
56 #include "xfs_rmap_btree.h"
57 #include "xfs_sb.h"
58 #include "xfs_ag_resv.h"
59 
60 /*
61  * Copy on Write of Shared Blocks
62  *
63  * XFS must preserve "the usual" file semantics even when two files share
64  * the same physical blocks.  This means that a write to one file must not
65  * alter the blocks in a different file; the way that we'll do that is
66  * through the use of a copy-on-write mechanism.  At a high level, that
67  * means that when we want to write to a shared block, we allocate a new
68  * block, write the data to the new block, and if that succeeds we map the
69  * new block into the file.
70  *
71  * XFS provides a "delayed allocation" mechanism that defers the allocation
72  * of disk blocks to dirty-but-not-yet-mapped file blocks as long as
73  * possible.  This reduces fragmentation by enabling the filesystem to ask
74  * for bigger chunks less often, which is exactly what we want for CoW.
75  *
76  * The delalloc mechanism begins when the kernel wants to make a block
77  * writable (write_begin or page_mkwrite).  If the offset is not mapped, we
78  * create a delalloc mapping, which is a regular in-core extent, but without
79  * a real startblock.  (For delalloc mappings, the startblock encodes both
80  * a flag that this is a delalloc mapping, and a worst-case estimate of how
81  * many blocks might be required to put the mapping into the BMBT.)  delalloc
82  * mappings are a reservation against the free space in the filesystem;
83  * adjacent mappings can also be combined into fewer larger mappings.
84  *
85  * As an optimization, the CoW extent size hint (cowextsz) creates
86  * outsized aligned delalloc reservations in the hope of landing out of
87  * order nearby CoW writes in a single extent on disk, thereby reducing
88  * fragmentation and improving future performance.
89  *
90  * D: --RRRRRRSSSRRRRRRRR--- (data fork)
91  * C: ------DDDDDDD--------- (CoW fork)
92  *
93  * When dirty pages are being written out (typically in writepage), the
94  * delalloc reservations are converted into unwritten mappings by
95  * allocating blocks and replacing the delalloc mapping with real ones.
96  * A delalloc mapping can be replaced by several unwritten ones if the
97  * free space is fragmented.
98  *
99  * D: --RRRRRRSSSRRRRRRRR---
100  * C: ------UUUUUUU---------
101  *
102  * We want to adapt the delalloc mechanism for copy-on-write, since the
103  * write paths are similar.  The first two steps (creating the reservation
104  * and allocating the blocks) are exactly the same as delalloc except that
105  * the mappings must be stored in a separate CoW fork because we do not want
106  * to disturb the mapping in the data fork until we're sure that the write
107  * succeeded.  IO completion in this case is the process of removing the old
108  * mapping from the data fork and moving the new mapping from the CoW fork to
109  * the data fork.  This will be discussed shortly.
110  *
111  * For now, unaligned directio writes will be bounced back to the page cache.
112  * Block-aligned directio writes will use the same mechanism as buffered
113  * writes.
114  *
115  * Just prior to submitting the actual disk write requests, we convert
116  * the extents representing the range of the file actually being written
117  * (as opposed to extra pieces created for the cowextsize hint) to real
118  * extents.  This will become important in the next step:
119  *
120  * D: --RRRRRRSSSRRRRRRRR---
121  * C: ------UUrrUUU---------
122  *
123  * CoW remapping must be done after the data block write completes,
124  * because we don't want to destroy the old data fork map until we're sure
125  * the new block has been written.  Since the new mappings are kept in a
126  * separate fork, we can simply iterate these mappings to find the ones
127  * that cover the file blocks that we just CoW'd.  For each extent, simply
128  * unmap the corresponding range in the data fork, map the new range into
129  * the data fork, and remove the extent from the CoW fork.  Because of
130  * the presence of the cowextsize hint, however, we must be careful
131  * only to remap the blocks that we've actually written out --  we must
132  * never remap delalloc reservations nor CoW staging blocks that have
133  * yet to be written.  This corresponds exactly to the real extents in
134  * the CoW fork:
135  *
136  * D: --RRRRRRrrSRRRRRRRR---
137  * C: ------UU--UUU---------
138  *
139  * Since the remapping operation can be applied to an arbitrary file
140  * range, we record the need for the remap step as a flag in the ioend
141  * instead of declaring a new IO type.  This is required for direct io
142  * because we only have ioend for the whole dio, and we have to be able to
143  * remember the presence of unwritten blocks and CoW blocks with a single
144  * ioend structure.  Better yet, the more ground we can cover with one
145  * ioend, the better.
146  */
147 
148 /*
149  * Given an AG extent, find the lowest-numbered run of shared blocks
150  * within that range and return the range in fbno/flen.  If
151  * find_end_of_shared is true, return the longest contiguous extent of
152  * shared blocks.  If there are no shared extents, fbno and flen will
153  * be set to NULLAGBLOCK and 0, respectively.
154  */
155 int
156 xfs_reflink_find_shared(
157 	struct xfs_mount	*mp,
158 	struct xfs_trans	*tp,
159 	xfs_agnumber_t		agno,
160 	xfs_agblock_t		agbno,
161 	xfs_extlen_t		aglen,
162 	xfs_agblock_t		*fbno,
163 	xfs_extlen_t		*flen,
164 	bool			find_end_of_shared)
165 {
166 	struct xfs_buf		*agbp;
167 	struct xfs_btree_cur	*cur;
168 	int			error;
169 
170 	error = xfs_alloc_read_agf(mp, tp, agno, 0, &agbp);
171 	if (error)
172 		return error;
173 	if (!agbp)
174 		return -ENOMEM;
175 
176 	cur = xfs_refcountbt_init_cursor(mp, tp, agbp, agno, NULL);
177 
178 	error = xfs_refcount_find_shared(cur, agbno, aglen, fbno, flen,
179 			find_end_of_shared);
180 
181 	xfs_btree_del_cursor(cur, error ? XFS_BTREE_ERROR : XFS_BTREE_NOERROR);
182 
183 	xfs_trans_brelse(tp, agbp);
184 	return error;
185 }
186 
187 /*
188  * Trim the mapping to the next block where there's a change in the
189  * shared/unshared status.  More specifically, this means that we
190  * find the lowest-numbered extent of shared blocks that coincides with
191  * the given block mapping.  If the shared extent overlaps the start of
192  * the mapping, trim the mapping to the end of the shared extent.  If
193  * the shared region intersects the mapping, trim the mapping to the
194  * start of the shared extent.  If there are no shared regions that
195  * overlap, just return the original extent.
196  */
197 int
198 xfs_reflink_trim_around_shared(
199 	struct xfs_inode	*ip,
200 	struct xfs_bmbt_irec	*irec,
201 	bool			*shared,
202 	bool			*trimmed)
203 {
204 	xfs_agnumber_t		agno;
205 	xfs_agblock_t		agbno;
206 	xfs_extlen_t		aglen;
207 	xfs_agblock_t		fbno;
208 	xfs_extlen_t		flen;
209 	int			error = 0;
210 
211 	/* Holes, unwritten, and delalloc extents cannot be shared */
212 	if (!xfs_is_reflink_inode(ip) || !xfs_bmap_is_real_extent(irec)) {
213 		*shared = false;
214 		return 0;
215 	}
216 
217 	trace_xfs_reflink_trim_around_shared(ip, irec);
218 
219 	agno = XFS_FSB_TO_AGNO(ip->i_mount, irec->br_startblock);
220 	agbno = XFS_FSB_TO_AGBNO(ip->i_mount, irec->br_startblock);
221 	aglen = irec->br_blockcount;
222 
223 	error = xfs_reflink_find_shared(ip->i_mount, NULL, agno, agbno,
224 			aglen, &fbno, &flen, true);
225 	if (error)
226 		return error;
227 
228 	*shared = *trimmed = false;
229 	if (fbno == NULLAGBLOCK) {
230 		/* No shared blocks at all. */
231 		return 0;
232 	} else if (fbno == agbno) {
233 		/*
234 		 * The start of this extent is shared.  Truncate the
235 		 * mapping at the end of the shared region so that a
236 		 * subsequent iteration starts at the start of the
237 		 * unshared region.
238 		 */
239 		irec->br_blockcount = flen;
240 		*shared = true;
241 		if (flen != aglen)
242 			*trimmed = true;
243 		return 0;
244 	} else {
245 		/*
246 		 * There's a shared extent midway through this extent.
247 		 * Truncate the mapping at the start of the shared
248 		 * extent so that a subsequent iteration starts at the
249 		 * start of the shared region.
250 		 */
251 		irec->br_blockcount = fbno - agbno;
252 		*trimmed = true;
253 		return 0;
254 	}
255 }
256 
257 /*
258  * Trim the passed in imap to the next shared/unshared extent boundary, and
259  * if imap->br_startoff points to a shared extent reserve space for it in the
260  * COW fork.  In this case *shared is set to true, else to false.
261  *
262  * Note that imap will always contain the block numbers for the existing blocks
263  * in the data fork, as the upper layers need them for read-modify-write
264  * operations.
265  */
266 int
267 xfs_reflink_reserve_cow(
268 	struct xfs_inode	*ip,
269 	struct xfs_bmbt_irec	*imap,
270 	bool			*shared)
271 {
272 	struct xfs_ifork	*ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
273 	struct xfs_bmbt_irec	got;
274 	int			error = 0;
275 	bool			eof = false, trimmed;
276 	struct xfs_iext_cursor	icur;
277 
278 	/*
279 	 * Search the COW fork extent list first.  This serves two purposes:
280 	 * first this implement the speculative preallocation using cowextisze,
281 	 * so that we also unshared block adjacent to shared blocks instead
282 	 * of just the shared blocks themselves.  Second the lookup in the
283 	 * extent list is generally faster than going out to the shared extent
284 	 * tree.
285 	 */
286 
287 	if (!xfs_iext_lookup_extent(ip, ifp, imap->br_startoff, &icur, &got))
288 		eof = true;
289 	if (!eof && got.br_startoff <= imap->br_startoff) {
290 		trace_xfs_reflink_cow_found(ip, imap);
291 		xfs_trim_extent(imap, got.br_startoff, got.br_blockcount);
292 
293 		*shared = true;
294 		return 0;
295 	}
296 
297 	/* Trim the mapping to the nearest shared extent boundary. */
298 	error = xfs_reflink_trim_around_shared(ip, imap, shared, &trimmed);
299 	if (error)
300 		return error;
301 
302 	/* Not shared?  Just report the (potentially capped) extent. */
303 	if (!*shared)
304 		return 0;
305 
306 	/*
307 	 * Fork all the shared blocks from our write offset until the end of
308 	 * the extent.
309 	 */
310 	error = xfs_qm_dqattach_locked(ip, 0);
311 	if (error)
312 		return error;
313 
314 	error = xfs_bmapi_reserve_delalloc(ip, XFS_COW_FORK, imap->br_startoff,
315 			imap->br_blockcount, 0, &got, &icur, eof);
316 	if (error == -ENOSPC || error == -EDQUOT)
317 		trace_xfs_reflink_cow_enospc(ip, imap);
318 	if (error)
319 		return error;
320 
321 	trace_xfs_reflink_cow_alloc(ip, &got);
322 	return 0;
323 }
324 
325 /* Convert part of an unwritten CoW extent to a real one. */
326 STATIC int
327 xfs_reflink_convert_cow_extent(
328 	struct xfs_inode		*ip,
329 	struct xfs_bmbt_irec		*imap,
330 	xfs_fileoff_t			offset_fsb,
331 	xfs_filblks_t			count_fsb,
332 	struct xfs_defer_ops		*dfops)
333 {
334 	xfs_fsblock_t			first_block = NULLFSBLOCK;
335 	int				nimaps = 1;
336 
337 	if (imap->br_state == XFS_EXT_NORM)
338 		return 0;
339 
340 	xfs_trim_extent(imap, offset_fsb, count_fsb);
341 	trace_xfs_reflink_convert_cow(ip, imap);
342 	if (imap->br_blockcount == 0)
343 		return 0;
344 	return xfs_bmapi_write(NULL, ip, imap->br_startoff, imap->br_blockcount,
345 			XFS_BMAPI_COWFORK | XFS_BMAPI_CONVERT, &first_block,
346 			0, imap, &nimaps, dfops);
347 }
348 
349 /* Convert all of the unwritten CoW extents in a file's range to real ones. */
350 int
351 xfs_reflink_convert_cow(
352 	struct xfs_inode	*ip,
353 	xfs_off_t		offset,
354 	xfs_off_t		count)
355 {
356 	struct xfs_mount	*mp = ip->i_mount;
357 	xfs_fileoff_t		offset_fsb = XFS_B_TO_FSBT(mp, offset);
358 	xfs_fileoff_t		end_fsb = XFS_B_TO_FSB(mp, offset + count);
359 	xfs_filblks_t		count_fsb = end_fsb - offset_fsb;
360 	struct xfs_bmbt_irec	imap;
361 	struct xfs_defer_ops	dfops;
362 	xfs_fsblock_t		first_block = NULLFSBLOCK;
363 	int			nimaps = 1, error = 0;
364 
365 	ASSERT(count != 0);
366 
367 	xfs_ilock(ip, XFS_ILOCK_EXCL);
368 	error = xfs_bmapi_write(NULL, ip, offset_fsb, count_fsb,
369 			XFS_BMAPI_COWFORK | XFS_BMAPI_CONVERT |
370 			XFS_BMAPI_CONVERT_ONLY, &first_block, 0, &imap, &nimaps,
371 			&dfops);
372 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
373 	return error;
374 }
375 
376 /* Allocate all CoW reservations covering a range of blocks in a file. */
377 int
378 xfs_reflink_allocate_cow(
379 	struct xfs_inode	*ip,
380 	struct xfs_bmbt_irec	*imap,
381 	bool			*shared,
382 	uint			*lockmode)
383 {
384 	struct xfs_mount	*mp = ip->i_mount;
385 	xfs_fileoff_t		offset_fsb = imap->br_startoff;
386 	xfs_filblks_t		count_fsb = imap->br_blockcount;
387 	struct xfs_bmbt_irec	got;
388 	struct xfs_defer_ops	dfops;
389 	struct xfs_trans	*tp = NULL;
390 	xfs_fsblock_t		first_block;
391 	int			nimaps, error = 0;
392 	bool			trimmed;
393 	xfs_filblks_t		resaligned;
394 	xfs_extlen_t		resblks = 0;
395 	struct xfs_iext_cursor	icur;
396 
397 retry:
398 	ASSERT(xfs_is_reflink_inode(ip));
399 	ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL | XFS_ILOCK_SHARED));
400 
401 	/*
402 	 * Even if the extent is not shared we might have a preallocation for
403 	 * it in the COW fork.  If so use it.
404 	 */
405 	if (xfs_iext_lookup_extent(ip, ip->i_cowfp, offset_fsb, &icur, &got) &&
406 	    got.br_startoff <= offset_fsb) {
407 		*shared = true;
408 
409 		/* If we have a real allocation in the COW fork we're done. */
410 		if (!isnullstartblock(got.br_startblock)) {
411 			xfs_trim_extent(&got, offset_fsb, count_fsb);
412 			*imap = got;
413 			goto convert;
414 		}
415 
416 		xfs_trim_extent(imap, got.br_startoff, got.br_blockcount);
417 	} else {
418 		error = xfs_reflink_trim_around_shared(ip, imap, shared, &trimmed);
419 		if (error || !*shared)
420 			goto out;
421 	}
422 
423 	if (!tp) {
424 		resaligned = xfs_aligned_fsb_count(imap->br_startoff,
425 			imap->br_blockcount, xfs_get_cowextsz_hint(ip));
426 		resblks = XFS_DIOSTRAT_SPACE_RES(mp, resaligned);
427 
428 		xfs_iunlock(ip, *lockmode);
429 		error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, resblks, 0, 0, &tp);
430 		*lockmode = XFS_ILOCK_EXCL;
431 		xfs_ilock(ip, *lockmode);
432 
433 		if (error)
434 			return error;
435 
436 		error = xfs_qm_dqattach_locked(ip, 0);
437 		if (error)
438 			goto out;
439 		goto retry;
440 	}
441 
442 	error = xfs_trans_reserve_quota_nblks(tp, ip, resblks, 0,
443 			XFS_QMOPT_RES_REGBLKS);
444 	if (error)
445 		goto out;
446 
447 	xfs_trans_ijoin(tp, ip, 0);
448 
449 	xfs_defer_init(&dfops, &first_block);
450 	nimaps = 1;
451 
452 	/* Allocate the entire reservation as unwritten blocks. */
453 	error = xfs_bmapi_write(tp, ip, imap->br_startoff, imap->br_blockcount,
454 			XFS_BMAPI_COWFORK | XFS_BMAPI_PREALLOC, &first_block,
455 			resblks, imap, &nimaps, &dfops);
456 	if (error)
457 		goto out_bmap_cancel;
458 
459 	/* Finish up. */
460 	error = xfs_defer_finish(&tp, &dfops);
461 	if (error)
462 		goto out_bmap_cancel;
463 
464 	error = xfs_trans_commit(tp);
465 	if (error)
466 		return error;
467 convert:
468 	return xfs_reflink_convert_cow_extent(ip, imap, offset_fsb, count_fsb,
469 			&dfops);
470 out_bmap_cancel:
471 	xfs_defer_cancel(&dfops);
472 	xfs_trans_unreserve_quota_nblks(tp, ip, (long)resblks, 0,
473 			XFS_QMOPT_RES_REGBLKS);
474 out:
475 	if (tp)
476 		xfs_trans_cancel(tp);
477 	return error;
478 }
479 
480 /*
481  * Find the CoW reservation for a given byte offset of a file.
482  */
483 bool
484 xfs_reflink_find_cow_mapping(
485 	struct xfs_inode		*ip,
486 	xfs_off_t			offset,
487 	struct xfs_bmbt_irec		*imap)
488 {
489 	struct xfs_ifork		*ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
490 	xfs_fileoff_t			offset_fsb;
491 	struct xfs_bmbt_irec		got;
492 	struct xfs_iext_cursor		icur;
493 
494 	ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL | XFS_ILOCK_SHARED));
495 	ASSERT(xfs_is_reflink_inode(ip));
496 
497 	offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
498 	if (!xfs_iext_lookup_extent(ip, ifp, offset_fsb, &icur, &got))
499 		return false;
500 	if (got.br_startoff > offset_fsb)
501 		return false;
502 
503 	trace_xfs_reflink_find_cow_mapping(ip, offset, 1, XFS_IO_OVERWRITE,
504 			&got);
505 	*imap = got;
506 	return true;
507 }
508 
509 /*
510  * Trim an extent to end at the next CoW reservation past offset_fsb.
511  */
512 void
513 xfs_reflink_trim_irec_to_next_cow(
514 	struct xfs_inode		*ip,
515 	xfs_fileoff_t			offset_fsb,
516 	struct xfs_bmbt_irec		*imap)
517 {
518 	struct xfs_ifork		*ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
519 	struct xfs_bmbt_irec		got;
520 	struct xfs_iext_cursor		icur;
521 
522 	if (!xfs_is_reflink_inode(ip))
523 		return;
524 
525 	/* Find the extent in the CoW fork. */
526 	if (!xfs_iext_lookup_extent(ip, ifp, offset_fsb, &icur, &got))
527 		return;
528 
529 	/* This is the extent before; try sliding up one. */
530 	if (got.br_startoff < offset_fsb) {
531 		if (!xfs_iext_next_extent(ifp, &icur, &got))
532 			return;
533 	}
534 
535 	if (got.br_startoff >= imap->br_startoff + imap->br_blockcount)
536 		return;
537 
538 	imap->br_blockcount = got.br_startoff - imap->br_startoff;
539 	trace_xfs_reflink_trim_irec(ip, imap);
540 }
541 
542 /*
543  * Cancel CoW reservations for some block range of an inode.
544  *
545  * If cancel_real is true this function cancels all COW fork extents for the
546  * inode; if cancel_real is false, real extents are not cleared.
547  */
548 int
549 xfs_reflink_cancel_cow_blocks(
550 	struct xfs_inode		*ip,
551 	struct xfs_trans		**tpp,
552 	xfs_fileoff_t			offset_fsb,
553 	xfs_fileoff_t			end_fsb,
554 	bool				cancel_real)
555 {
556 	struct xfs_ifork		*ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
557 	struct xfs_bmbt_irec		got, del;
558 	struct xfs_iext_cursor		icur;
559 	xfs_fsblock_t			firstfsb;
560 	struct xfs_defer_ops		dfops;
561 	int				error = 0;
562 
563 	if (!xfs_is_reflink_inode(ip))
564 		return 0;
565 	if (!xfs_iext_lookup_extent_before(ip, ifp, &end_fsb, &icur, &got))
566 		return 0;
567 
568 	/* Walk backwards until we're out of the I/O range... */
569 	while (got.br_startoff + got.br_blockcount > offset_fsb) {
570 		del = got;
571 		xfs_trim_extent(&del, offset_fsb, end_fsb - offset_fsb);
572 
573 		/* Extent delete may have bumped ext forward */
574 		if (!del.br_blockcount) {
575 			xfs_iext_prev(ifp, &icur);
576 			goto next_extent;
577 		}
578 
579 		trace_xfs_reflink_cancel_cow(ip, &del);
580 
581 		if (isnullstartblock(del.br_startblock)) {
582 			error = xfs_bmap_del_extent_delay(ip, XFS_COW_FORK,
583 					&icur, &got, &del);
584 			if (error)
585 				break;
586 		} else if (del.br_state == XFS_EXT_UNWRITTEN || cancel_real) {
587 			xfs_trans_ijoin(*tpp, ip, 0);
588 			xfs_defer_init(&dfops, &firstfsb);
589 
590 			/* Free the CoW orphan record. */
591 			error = xfs_refcount_free_cow_extent(ip->i_mount,
592 					&dfops, del.br_startblock,
593 					del.br_blockcount);
594 			if (error)
595 				break;
596 
597 			xfs_bmap_add_free(ip->i_mount, &dfops,
598 					del.br_startblock, del.br_blockcount,
599 					NULL);
600 
601 			/* Update quota accounting */
602 			xfs_trans_mod_dquot_byino(*tpp, ip, XFS_TRANS_DQ_BCOUNT,
603 					-(long)del.br_blockcount);
604 
605 			/* Roll the transaction */
606 			xfs_defer_ijoin(&dfops, ip);
607 			error = xfs_defer_finish(tpp, &dfops);
608 			if (error) {
609 				xfs_defer_cancel(&dfops);
610 				break;
611 			}
612 
613 			/* Remove the mapping from the CoW fork. */
614 			xfs_bmap_del_extent_cow(ip, &icur, &got, &del);
615 		}
616 next_extent:
617 		if (!xfs_iext_get_extent(ifp, &icur, &got))
618 			break;
619 	}
620 
621 	/* clear tag if cow fork is emptied */
622 	if (!ifp->if_bytes)
623 		xfs_inode_clear_cowblocks_tag(ip);
624 
625 	return error;
626 }
627 
628 /*
629  * Cancel CoW reservations for some byte range of an inode.
630  *
631  * If cancel_real is true this function cancels all COW fork extents for the
632  * inode; if cancel_real is false, real extents are not cleared.
633  */
634 int
635 xfs_reflink_cancel_cow_range(
636 	struct xfs_inode	*ip,
637 	xfs_off_t		offset,
638 	xfs_off_t		count,
639 	bool			cancel_real)
640 {
641 	struct xfs_trans	*tp;
642 	xfs_fileoff_t		offset_fsb;
643 	xfs_fileoff_t		end_fsb;
644 	int			error;
645 
646 	trace_xfs_reflink_cancel_cow_range(ip, offset, count);
647 	ASSERT(xfs_is_reflink_inode(ip));
648 
649 	offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
650 	if (count == NULLFILEOFF)
651 		end_fsb = NULLFILEOFF;
652 	else
653 		end_fsb = XFS_B_TO_FSB(ip->i_mount, offset + count);
654 
655 	/* Start a rolling transaction to remove the mappings */
656 	error = xfs_trans_alloc(ip->i_mount, &M_RES(ip->i_mount)->tr_write,
657 			0, 0, 0, &tp);
658 	if (error)
659 		goto out;
660 
661 	xfs_ilock(ip, XFS_ILOCK_EXCL);
662 	xfs_trans_ijoin(tp, ip, 0);
663 
664 	/* Scrape out the old CoW reservations */
665 	error = xfs_reflink_cancel_cow_blocks(ip, &tp, offset_fsb, end_fsb,
666 			cancel_real);
667 	if (error)
668 		goto out_cancel;
669 
670 	error = xfs_trans_commit(tp);
671 
672 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
673 	return error;
674 
675 out_cancel:
676 	xfs_trans_cancel(tp);
677 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
678 out:
679 	trace_xfs_reflink_cancel_cow_range_error(ip, error, _RET_IP_);
680 	return error;
681 }
682 
683 /*
684  * Remap parts of a file's data fork after a successful CoW.
685  */
686 int
687 xfs_reflink_end_cow(
688 	struct xfs_inode		*ip,
689 	xfs_off_t			offset,
690 	xfs_off_t			count)
691 {
692 	struct xfs_ifork		*ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
693 	struct xfs_bmbt_irec		got, del;
694 	struct xfs_trans		*tp;
695 	xfs_fileoff_t			offset_fsb;
696 	xfs_fileoff_t			end_fsb;
697 	xfs_fsblock_t			firstfsb;
698 	struct xfs_defer_ops		dfops;
699 	int				error;
700 	unsigned int			resblks;
701 	xfs_filblks_t			rlen;
702 	struct xfs_iext_cursor		icur;
703 
704 	trace_xfs_reflink_end_cow(ip, offset, count);
705 
706 	/* No COW extents?  That's easy! */
707 	if (ifp->if_bytes == 0)
708 		return 0;
709 
710 	offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
711 	end_fsb = XFS_B_TO_FSB(ip->i_mount, offset + count);
712 
713 	/*
714 	 * Start a rolling transaction to switch the mappings.  We're
715 	 * unlikely ever to have to remap 16T worth of single-block
716 	 * extents, so just cap the worst case extent count to 2^32-1.
717 	 * Stick a warning in just in case, and avoid 64-bit division.
718 	 */
719 	BUILD_BUG_ON(MAX_RW_COUNT > UINT_MAX);
720 	if (end_fsb - offset_fsb > UINT_MAX) {
721 		error = -EFSCORRUPTED;
722 		xfs_force_shutdown(ip->i_mount, SHUTDOWN_CORRUPT_INCORE);
723 		ASSERT(0);
724 		goto out;
725 	}
726 	resblks = XFS_NEXTENTADD_SPACE_RES(ip->i_mount,
727 			(unsigned int)(end_fsb - offset_fsb),
728 			XFS_DATA_FORK);
729 	error = xfs_trans_alloc(ip->i_mount, &M_RES(ip->i_mount)->tr_write,
730 			resblks, 0, 0, &tp);
731 	if (error)
732 		goto out;
733 
734 	xfs_ilock(ip, XFS_ILOCK_EXCL);
735 	xfs_trans_ijoin(tp, ip, 0);
736 
737 	/*
738 	 * In case of racing, overlapping AIO writes no COW extents might be
739 	 * left by the time I/O completes for the loser of the race.  In that
740 	 * case we are done.
741 	 */
742 	if (!xfs_iext_lookup_extent_before(ip, ifp, &end_fsb, &icur, &got))
743 		goto out_cancel;
744 
745 	/* Walk backwards until we're out of the I/O range... */
746 	while (got.br_startoff + got.br_blockcount > offset_fsb) {
747 		del = got;
748 		xfs_trim_extent(&del, offset_fsb, end_fsb - offset_fsb);
749 
750 		/* Extent delete may have bumped ext forward */
751 		if (!del.br_blockcount) {
752 			xfs_iext_prev(ifp, &icur);
753 			goto next_extent;
754 		}
755 
756 		ASSERT(!isnullstartblock(got.br_startblock));
757 
758 		/*
759 		 * Don't remap unwritten extents; these are
760 		 * speculatively preallocated CoW extents that have been
761 		 * allocated but have not yet been involved in a write.
762 		 */
763 		if (got.br_state == XFS_EXT_UNWRITTEN) {
764 			xfs_iext_prev(ifp, &icur);
765 			goto next_extent;
766 		}
767 
768 		/* Unmap the old blocks in the data fork. */
769 		xfs_defer_init(&dfops, &firstfsb);
770 		rlen = del.br_blockcount;
771 		error = __xfs_bunmapi(tp, ip, del.br_startoff, &rlen, 0, 1,
772 				&firstfsb, &dfops);
773 		if (error)
774 			goto out_defer;
775 
776 		/* Trim the extent to whatever got unmapped. */
777 		if (rlen) {
778 			xfs_trim_extent(&del, del.br_startoff + rlen,
779 				del.br_blockcount - rlen);
780 		}
781 		trace_xfs_reflink_cow_remap(ip, &del);
782 
783 		/* Free the CoW orphan record. */
784 		error = xfs_refcount_free_cow_extent(tp->t_mountp, &dfops,
785 				del.br_startblock, del.br_blockcount);
786 		if (error)
787 			goto out_defer;
788 
789 		/* Map the new blocks into the data fork. */
790 		error = xfs_bmap_map_extent(tp->t_mountp, &dfops, ip, &del);
791 		if (error)
792 			goto out_defer;
793 
794 		/* Remove the mapping from the CoW fork. */
795 		xfs_bmap_del_extent_cow(ip, &icur, &got, &del);
796 
797 		xfs_defer_ijoin(&dfops, ip);
798 		error = xfs_defer_finish(&tp, &dfops);
799 		if (error)
800 			goto out_defer;
801 next_extent:
802 		if (!xfs_iext_get_extent(ifp, &icur, &got))
803 			break;
804 	}
805 
806 	error = xfs_trans_commit(tp);
807 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
808 	if (error)
809 		goto out;
810 	return 0;
811 
812 out_defer:
813 	xfs_defer_cancel(&dfops);
814 out_cancel:
815 	xfs_trans_cancel(tp);
816 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
817 out:
818 	trace_xfs_reflink_end_cow_error(ip, error, _RET_IP_);
819 	return error;
820 }
821 
822 /*
823  * Free leftover CoW reservations that didn't get cleaned out.
824  */
825 int
826 xfs_reflink_recover_cow(
827 	struct xfs_mount	*mp)
828 {
829 	xfs_agnumber_t		agno;
830 	int			error = 0;
831 
832 	if (!xfs_sb_version_hasreflink(&mp->m_sb))
833 		return 0;
834 
835 	for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
836 		error = xfs_refcount_recover_cow_leftovers(mp, agno);
837 		if (error)
838 			break;
839 	}
840 
841 	return error;
842 }
843 
844 /*
845  * Reflinking (Block) Ranges of Two Files Together
846  *
847  * First, ensure that the reflink flag is set on both inodes.  The flag is an
848  * optimization to avoid unnecessary refcount btree lookups in the write path.
849  *
850  * Now we can iteratively remap the range of extents (and holes) in src to the
851  * corresponding ranges in dest.  Let drange and srange denote the ranges of
852  * logical blocks in dest and src touched by the reflink operation.
853  *
854  * While the length of drange is greater than zero,
855  *    - Read src's bmbt at the start of srange ("imap")
856  *    - If imap doesn't exist, make imap appear to start at the end of srange
857  *      with zero length.
858  *    - If imap starts before srange, advance imap to start at srange.
859  *    - If imap goes beyond srange, truncate imap to end at the end of srange.
860  *    - Punch (imap start - srange start + imap len) blocks from dest at
861  *      offset (drange start).
862  *    - If imap points to a real range of pblks,
863  *         > Increase the refcount of the imap's pblks
864  *         > Map imap's pblks into dest at the offset
865  *           (drange start + imap start - srange start)
866  *    - Advance drange and srange by (imap start - srange start + imap len)
867  *
868  * Finally, if the reflink made dest longer, update both the in-core and
869  * on-disk file sizes.
870  *
871  * ASCII Art Demonstration:
872  *
873  * Let's say we want to reflink this source file:
874  *
875  * ----SSSSSSS-SSSSS----SSSSSS (src file)
876  *   <-------------------->
877  *
878  * into this destination file:
879  *
880  * --DDDDDDDDDDDDDDDDDDD--DDD (dest file)
881  *        <-------------------->
882  * '-' means a hole, and 'S' and 'D' are written blocks in the src and dest.
883  * Observe that the range has different logical offsets in either file.
884  *
885  * Consider that the first extent in the source file doesn't line up with our
886  * reflink range.  Unmapping  and remapping are separate operations, so we can
887  * unmap more blocks from the destination file than we remap.
888  *
889  * ----SSSSSSS-SSSSS----SSSSSS
890  *   <------->
891  * --DDDDD---------DDDDD--DDD
892  *        <------->
893  *
894  * Now remap the source extent into the destination file:
895  *
896  * ----SSSSSSS-SSSSS----SSSSSS
897  *   <------->
898  * --DDDDD--SSSSSSSDDDDD--DDD
899  *        <------->
900  *
901  * Do likewise with the second hole and extent in our range.  Holes in the
902  * unmap range don't affect our operation.
903  *
904  * ----SSSSSSS-SSSSS----SSSSSS
905  *            <---->
906  * --DDDDD--SSSSSSS-SSSSS-DDD
907  *                 <---->
908  *
909  * Finally, unmap and remap part of the third extent.  This will increase the
910  * size of the destination file.
911  *
912  * ----SSSSSSS-SSSSS----SSSSSS
913  *                  <----->
914  * --DDDDD--SSSSSSS-SSSSS----SSS
915  *                       <----->
916  *
917  * Once we update the destination file's i_size, we're done.
918  */
919 
920 /*
921  * Ensure the reflink bit is set in both inodes.
922  */
923 STATIC int
924 xfs_reflink_set_inode_flag(
925 	struct xfs_inode	*src,
926 	struct xfs_inode	*dest)
927 {
928 	struct xfs_mount	*mp = src->i_mount;
929 	int			error;
930 	struct xfs_trans	*tp;
931 
932 	if (xfs_is_reflink_inode(src) && xfs_is_reflink_inode(dest))
933 		return 0;
934 
935 	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
936 	if (error)
937 		goto out_error;
938 
939 	/* Lock both files against IO */
940 	if (src->i_ino == dest->i_ino)
941 		xfs_ilock(src, XFS_ILOCK_EXCL);
942 	else
943 		xfs_lock_two_inodes(src, dest, XFS_ILOCK_EXCL);
944 
945 	if (!xfs_is_reflink_inode(src)) {
946 		trace_xfs_reflink_set_inode_flag(src);
947 		xfs_trans_ijoin(tp, src, XFS_ILOCK_EXCL);
948 		src->i_d.di_flags2 |= XFS_DIFLAG2_REFLINK;
949 		xfs_trans_log_inode(tp, src, XFS_ILOG_CORE);
950 		xfs_ifork_init_cow(src);
951 	} else
952 		xfs_iunlock(src, XFS_ILOCK_EXCL);
953 
954 	if (src->i_ino == dest->i_ino)
955 		goto commit_flags;
956 
957 	if (!xfs_is_reflink_inode(dest)) {
958 		trace_xfs_reflink_set_inode_flag(dest);
959 		xfs_trans_ijoin(tp, dest, XFS_ILOCK_EXCL);
960 		dest->i_d.di_flags2 |= XFS_DIFLAG2_REFLINK;
961 		xfs_trans_log_inode(tp, dest, XFS_ILOG_CORE);
962 		xfs_ifork_init_cow(dest);
963 	} else
964 		xfs_iunlock(dest, XFS_ILOCK_EXCL);
965 
966 commit_flags:
967 	error = xfs_trans_commit(tp);
968 	if (error)
969 		goto out_error;
970 	return error;
971 
972 out_error:
973 	trace_xfs_reflink_set_inode_flag_error(dest, error, _RET_IP_);
974 	return error;
975 }
976 
977 /*
978  * Update destination inode size & cowextsize hint, if necessary.
979  */
980 STATIC int
981 xfs_reflink_update_dest(
982 	struct xfs_inode	*dest,
983 	xfs_off_t		newlen,
984 	xfs_extlen_t		cowextsize,
985 	bool			is_dedupe)
986 {
987 	struct xfs_mount	*mp = dest->i_mount;
988 	struct xfs_trans	*tp;
989 	int			error;
990 
991 	if (is_dedupe && newlen <= i_size_read(VFS_I(dest)) && cowextsize == 0)
992 		return 0;
993 
994 	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
995 	if (error)
996 		goto out_error;
997 
998 	xfs_ilock(dest, XFS_ILOCK_EXCL);
999 	xfs_trans_ijoin(tp, dest, XFS_ILOCK_EXCL);
1000 
1001 	if (newlen > i_size_read(VFS_I(dest))) {
1002 		trace_xfs_reflink_update_inode_size(dest, newlen);
1003 		i_size_write(VFS_I(dest), newlen);
1004 		dest->i_d.di_size = newlen;
1005 	}
1006 
1007 	if (cowextsize) {
1008 		dest->i_d.di_cowextsize = cowextsize;
1009 		dest->i_d.di_flags2 |= XFS_DIFLAG2_COWEXTSIZE;
1010 	}
1011 
1012 	if (!is_dedupe) {
1013 		xfs_trans_ichgtime(tp, dest,
1014 				   XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1015 	}
1016 	xfs_trans_log_inode(tp, dest, XFS_ILOG_CORE);
1017 
1018 	error = xfs_trans_commit(tp);
1019 	if (error)
1020 		goto out_error;
1021 	return error;
1022 
1023 out_error:
1024 	trace_xfs_reflink_update_inode_size_error(dest, error, _RET_IP_);
1025 	return error;
1026 }
1027 
1028 /*
1029  * Do we have enough reserve in this AG to handle a reflink?  The refcount
1030  * btree already reserved all the space it needs, but the rmap btree can grow
1031  * infinitely, so we won't allow more reflinks when the AG is down to the
1032  * btree reserves.
1033  */
1034 static int
1035 xfs_reflink_ag_has_free_space(
1036 	struct xfs_mount	*mp,
1037 	xfs_agnumber_t		agno)
1038 {
1039 	struct xfs_perag	*pag;
1040 	int			error = 0;
1041 
1042 	if (!xfs_sb_version_hasrmapbt(&mp->m_sb))
1043 		return 0;
1044 
1045 	pag = xfs_perag_get(mp, agno);
1046 	if (xfs_ag_resv_critical(pag, XFS_AG_RESV_AGFL) ||
1047 	    xfs_ag_resv_critical(pag, XFS_AG_RESV_METADATA))
1048 		error = -ENOSPC;
1049 	xfs_perag_put(pag);
1050 	return error;
1051 }
1052 
1053 /*
1054  * Unmap a range of blocks from a file, then map other blocks into the hole.
1055  * The range to unmap is (destoff : destoff + srcioff + irec->br_blockcount).
1056  * The extent irec is mapped into dest at irec->br_startoff.
1057  */
1058 STATIC int
1059 xfs_reflink_remap_extent(
1060 	struct xfs_inode	*ip,
1061 	struct xfs_bmbt_irec	*irec,
1062 	xfs_fileoff_t		destoff,
1063 	xfs_off_t		new_isize)
1064 {
1065 	struct xfs_mount	*mp = ip->i_mount;
1066 	bool			real_extent = xfs_bmap_is_real_extent(irec);
1067 	struct xfs_trans	*tp;
1068 	xfs_fsblock_t		firstfsb;
1069 	unsigned int		resblks;
1070 	struct xfs_defer_ops	dfops;
1071 	struct xfs_bmbt_irec	uirec;
1072 	xfs_filblks_t		rlen;
1073 	xfs_filblks_t		unmap_len;
1074 	xfs_off_t		newlen;
1075 	int			error;
1076 
1077 	unmap_len = irec->br_startoff + irec->br_blockcount - destoff;
1078 	trace_xfs_reflink_punch_range(ip, destoff, unmap_len);
1079 
1080 	/* No reflinking if we're low on space */
1081 	if (real_extent) {
1082 		error = xfs_reflink_ag_has_free_space(mp,
1083 				XFS_FSB_TO_AGNO(mp, irec->br_startblock));
1084 		if (error)
1085 			goto out;
1086 	}
1087 
1088 	/* Start a rolling transaction to switch the mappings */
1089 	resblks = XFS_EXTENTADD_SPACE_RES(ip->i_mount, XFS_DATA_FORK);
1090 	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, resblks, 0, 0, &tp);
1091 	if (error)
1092 		goto out;
1093 
1094 	xfs_ilock(ip, XFS_ILOCK_EXCL);
1095 	xfs_trans_ijoin(tp, ip, 0);
1096 
1097 	/* If we're not just clearing space, then do we have enough quota? */
1098 	if (real_extent) {
1099 		error = xfs_trans_reserve_quota_nblks(tp, ip,
1100 				irec->br_blockcount, 0, XFS_QMOPT_RES_REGBLKS);
1101 		if (error)
1102 			goto out_cancel;
1103 	}
1104 
1105 	trace_xfs_reflink_remap(ip, irec->br_startoff,
1106 				irec->br_blockcount, irec->br_startblock);
1107 
1108 	/* Unmap the old blocks in the data fork. */
1109 	rlen = unmap_len;
1110 	while (rlen) {
1111 		xfs_defer_init(&dfops, &firstfsb);
1112 		error = __xfs_bunmapi(tp, ip, destoff, &rlen, 0, 1,
1113 				&firstfsb, &dfops);
1114 		if (error)
1115 			goto out_defer;
1116 
1117 		/*
1118 		 * Trim the extent to whatever got unmapped.
1119 		 * Remember, bunmapi works backwards.
1120 		 */
1121 		uirec.br_startblock = irec->br_startblock + rlen;
1122 		uirec.br_startoff = irec->br_startoff + rlen;
1123 		uirec.br_blockcount = unmap_len - rlen;
1124 		unmap_len = rlen;
1125 
1126 		/* If this isn't a real mapping, we're done. */
1127 		if (!real_extent || uirec.br_blockcount == 0)
1128 			goto next_extent;
1129 
1130 		trace_xfs_reflink_remap(ip, uirec.br_startoff,
1131 				uirec.br_blockcount, uirec.br_startblock);
1132 
1133 		/* Update the refcount tree */
1134 		error = xfs_refcount_increase_extent(mp, &dfops, &uirec);
1135 		if (error)
1136 			goto out_defer;
1137 
1138 		/* Map the new blocks into the data fork. */
1139 		error = xfs_bmap_map_extent(mp, &dfops, ip, &uirec);
1140 		if (error)
1141 			goto out_defer;
1142 
1143 		/* Update quota accounting. */
1144 		xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_BCOUNT,
1145 				uirec.br_blockcount);
1146 
1147 		/* Update dest isize if needed. */
1148 		newlen = XFS_FSB_TO_B(mp,
1149 				uirec.br_startoff + uirec.br_blockcount);
1150 		newlen = min_t(xfs_off_t, newlen, new_isize);
1151 		if (newlen > i_size_read(VFS_I(ip))) {
1152 			trace_xfs_reflink_update_inode_size(ip, newlen);
1153 			i_size_write(VFS_I(ip), newlen);
1154 			ip->i_d.di_size = newlen;
1155 			xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1156 		}
1157 
1158 next_extent:
1159 		/* Process all the deferred stuff. */
1160 		xfs_defer_ijoin(&dfops, ip);
1161 		error = xfs_defer_finish(&tp, &dfops);
1162 		if (error)
1163 			goto out_defer;
1164 	}
1165 
1166 	error = xfs_trans_commit(tp);
1167 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1168 	if (error)
1169 		goto out;
1170 	return 0;
1171 
1172 out_defer:
1173 	xfs_defer_cancel(&dfops);
1174 out_cancel:
1175 	xfs_trans_cancel(tp);
1176 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1177 out:
1178 	trace_xfs_reflink_remap_extent_error(ip, error, _RET_IP_);
1179 	return error;
1180 }
1181 
1182 /*
1183  * Iteratively remap one file's extents (and holes) to another's.
1184  */
1185 STATIC int
1186 xfs_reflink_remap_blocks(
1187 	struct xfs_inode	*src,
1188 	xfs_fileoff_t		srcoff,
1189 	struct xfs_inode	*dest,
1190 	xfs_fileoff_t		destoff,
1191 	xfs_filblks_t		len,
1192 	xfs_off_t		new_isize)
1193 {
1194 	struct xfs_bmbt_irec	imap;
1195 	int			nimaps;
1196 	int			error = 0;
1197 	xfs_filblks_t		range_len;
1198 
1199 	/* drange = (destoff, destoff + len); srange = (srcoff, srcoff + len) */
1200 	while (len) {
1201 		trace_xfs_reflink_remap_blocks_loop(src, srcoff, len,
1202 				dest, destoff);
1203 		/* Read extent from the source file */
1204 		nimaps = 1;
1205 		xfs_ilock(src, XFS_ILOCK_EXCL);
1206 		error = xfs_bmapi_read(src, srcoff, len, &imap, &nimaps, 0);
1207 		xfs_iunlock(src, XFS_ILOCK_EXCL);
1208 		if (error)
1209 			goto err;
1210 		ASSERT(nimaps == 1);
1211 
1212 		trace_xfs_reflink_remap_imap(src, srcoff, len, XFS_IO_OVERWRITE,
1213 				&imap);
1214 
1215 		/* Translate imap into the destination file. */
1216 		range_len = imap.br_startoff + imap.br_blockcount - srcoff;
1217 		imap.br_startoff += destoff - srcoff;
1218 
1219 		/* Clear dest from destoff to the end of imap and map it in. */
1220 		error = xfs_reflink_remap_extent(dest, &imap, destoff,
1221 				new_isize);
1222 		if (error)
1223 			goto err;
1224 
1225 		if (fatal_signal_pending(current)) {
1226 			error = -EINTR;
1227 			goto err;
1228 		}
1229 
1230 		/* Advance drange/srange */
1231 		srcoff += range_len;
1232 		destoff += range_len;
1233 		len -= range_len;
1234 	}
1235 
1236 	return 0;
1237 
1238 err:
1239 	trace_xfs_reflink_remap_blocks_error(dest, error, _RET_IP_);
1240 	return error;
1241 }
1242 
1243 /*
1244  * Link a range of blocks from one file to another.
1245  */
1246 int
1247 xfs_reflink_remap_range(
1248 	struct file		*file_in,
1249 	loff_t			pos_in,
1250 	struct file		*file_out,
1251 	loff_t			pos_out,
1252 	u64			len,
1253 	bool			is_dedupe)
1254 {
1255 	struct inode		*inode_in = file_inode(file_in);
1256 	struct xfs_inode	*src = XFS_I(inode_in);
1257 	struct inode		*inode_out = file_inode(file_out);
1258 	struct xfs_inode	*dest = XFS_I(inode_out);
1259 	struct xfs_mount	*mp = src->i_mount;
1260 	bool			same_inode = (inode_in == inode_out);
1261 	xfs_fileoff_t		sfsbno, dfsbno;
1262 	xfs_filblks_t		fsblen;
1263 	xfs_extlen_t		cowextsize;
1264 	ssize_t			ret;
1265 
1266 	if (!xfs_sb_version_hasreflink(&mp->m_sb))
1267 		return -EOPNOTSUPP;
1268 
1269 	if (XFS_FORCED_SHUTDOWN(mp))
1270 		return -EIO;
1271 
1272 	/* Lock both files against IO */
1273 	lock_two_nondirectories(inode_in, inode_out);
1274 	if (same_inode)
1275 		xfs_ilock(src, XFS_MMAPLOCK_EXCL);
1276 	else
1277 		xfs_lock_two_inodes(src, dest, XFS_MMAPLOCK_EXCL);
1278 
1279 	/* Check file eligibility and prepare for block sharing. */
1280 	ret = -EINVAL;
1281 	/* Don't reflink realtime inodes */
1282 	if (XFS_IS_REALTIME_INODE(src) || XFS_IS_REALTIME_INODE(dest))
1283 		goto out_unlock;
1284 
1285 	/* Don't share DAX file data for now. */
1286 	if (IS_DAX(inode_in) || IS_DAX(inode_out))
1287 		goto out_unlock;
1288 
1289 	ret = vfs_clone_file_prep_inodes(inode_in, pos_in, inode_out, pos_out,
1290 			&len, is_dedupe);
1291 	if (ret <= 0)
1292 		goto out_unlock;
1293 
1294 	trace_xfs_reflink_remap_range(src, pos_in, len, dest, pos_out);
1295 
1296 	/* Set flags and remap blocks. */
1297 	ret = xfs_reflink_set_inode_flag(src, dest);
1298 	if (ret)
1299 		goto out_unlock;
1300 
1301 	dfsbno = XFS_B_TO_FSBT(mp, pos_out);
1302 	sfsbno = XFS_B_TO_FSBT(mp, pos_in);
1303 	fsblen = XFS_B_TO_FSB(mp, len);
1304 	ret = xfs_reflink_remap_blocks(src, sfsbno, dest, dfsbno, fsblen,
1305 			pos_out + len);
1306 	if (ret)
1307 		goto out_unlock;
1308 
1309 	/* Zap any page cache for the destination file's range. */
1310 	truncate_inode_pages_range(&inode_out->i_data, pos_out,
1311 				   PAGE_ALIGN(pos_out + len) - 1);
1312 
1313 	/*
1314 	 * Carry the cowextsize hint from src to dest if we're sharing the
1315 	 * entire source file to the entire destination file, the source file
1316 	 * has a cowextsize hint, and the destination file does not.
1317 	 */
1318 	cowextsize = 0;
1319 	if (pos_in == 0 && len == i_size_read(inode_in) &&
1320 	    (src->i_d.di_flags2 & XFS_DIFLAG2_COWEXTSIZE) &&
1321 	    pos_out == 0 && len >= i_size_read(inode_out) &&
1322 	    !(dest->i_d.di_flags2 & XFS_DIFLAG2_COWEXTSIZE))
1323 		cowextsize = src->i_d.di_cowextsize;
1324 
1325 	ret = xfs_reflink_update_dest(dest, pos_out + len, cowextsize,
1326 			is_dedupe);
1327 
1328 out_unlock:
1329 	xfs_iunlock(src, XFS_MMAPLOCK_EXCL);
1330 	if (!same_inode)
1331 		xfs_iunlock(dest, XFS_MMAPLOCK_EXCL);
1332 	unlock_two_nondirectories(inode_in, inode_out);
1333 	if (ret)
1334 		trace_xfs_reflink_remap_range_error(dest, ret, _RET_IP_);
1335 	return ret;
1336 }
1337 
1338 /*
1339  * The user wants to preemptively CoW all shared blocks in this file,
1340  * which enables us to turn off the reflink flag.  Iterate all
1341  * extents which are not prealloc/delalloc to see which ranges are
1342  * mentioned in the refcount tree, then read those blocks into the
1343  * pagecache, dirty them, fsync them back out, and then we can update
1344  * the inode flag.  What happens if we run out of memory? :)
1345  */
1346 STATIC int
1347 xfs_reflink_dirty_extents(
1348 	struct xfs_inode	*ip,
1349 	xfs_fileoff_t		fbno,
1350 	xfs_filblks_t		end,
1351 	xfs_off_t		isize)
1352 {
1353 	struct xfs_mount	*mp = ip->i_mount;
1354 	xfs_agnumber_t		agno;
1355 	xfs_agblock_t		agbno;
1356 	xfs_extlen_t		aglen;
1357 	xfs_agblock_t		rbno;
1358 	xfs_extlen_t		rlen;
1359 	xfs_off_t		fpos;
1360 	xfs_off_t		flen;
1361 	struct xfs_bmbt_irec	map[2];
1362 	int			nmaps;
1363 	int			error = 0;
1364 
1365 	while (end - fbno > 0) {
1366 		nmaps = 1;
1367 		/*
1368 		 * Look for extents in the file.  Skip holes, delalloc, or
1369 		 * unwritten extents; they can't be reflinked.
1370 		 */
1371 		error = xfs_bmapi_read(ip, fbno, end - fbno, map, &nmaps, 0);
1372 		if (error)
1373 			goto out;
1374 		if (nmaps == 0)
1375 			break;
1376 		if (!xfs_bmap_is_real_extent(&map[0]))
1377 			goto next;
1378 
1379 		map[1] = map[0];
1380 		while (map[1].br_blockcount) {
1381 			agno = XFS_FSB_TO_AGNO(mp, map[1].br_startblock);
1382 			agbno = XFS_FSB_TO_AGBNO(mp, map[1].br_startblock);
1383 			aglen = map[1].br_blockcount;
1384 
1385 			error = xfs_reflink_find_shared(mp, NULL, agno, agbno,
1386 					aglen, &rbno, &rlen, true);
1387 			if (error)
1388 				goto out;
1389 			if (rbno == NULLAGBLOCK)
1390 				break;
1391 
1392 			/* Dirty the pages */
1393 			xfs_iunlock(ip, XFS_ILOCK_EXCL);
1394 			fpos = XFS_FSB_TO_B(mp, map[1].br_startoff +
1395 					(rbno - agbno));
1396 			flen = XFS_FSB_TO_B(mp, rlen);
1397 			if (fpos + flen > isize)
1398 				flen = isize - fpos;
1399 			error = iomap_file_dirty(VFS_I(ip), fpos, flen,
1400 					&xfs_iomap_ops);
1401 			xfs_ilock(ip, XFS_ILOCK_EXCL);
1402 			if (error)
1403 				goto out;
1404 
1405 			map[1].br_blockcount -= (rbno - agbno + rlen);
1406 			map[1].br_startoff += (rbno - agbno + rlen);
1407 			map[1].br_startblock += (rbno - agbno + rlen);
1408 		}
1409 
1410 next:
1411 		fbno = map[0].br_startoff + map[0].br_blockcount;
1412 	}
1413 out:
1414 	return error;
1415 }
1416 
1417 /* Does this inode need the reflink flag? */
1418 int
1419 xfs_reflink_inode_has_shared_extents(
1420 	struct xfs_trans		*tp,
1421 	struct xfs_inode		*ip,
1422 	bool				*has_shared)
1423 {
1424 	struct xfs_bmbt_irec		got;
1425 	struct xfs_mount		*mp = ip->i_mount;
1426 	struct xfs_ifork		*ifp;
1427 	xfs_agnumber_t			agno;
1428 	xfs_agblock_t			agbno;
1429 	xfs_extlen_t			aglen;
1430 	xfs_agblock_t			rbno;
1431 	xfs_extlen_t			rlen;
1432 	struct xfs_iext_cursor		icur;
1433 	bool				found;
1434 	int				error;
1435 
1436 	ifp = XFS_IFORK_PTR(ip, XFS_DATA_FORK);
1437 	if (!(ifp->if_flags & XFS_IFEXTENTS)) {
1438 		error = xfs_iread_extents(tp, ip, XFS_DATA_FORK);
1439 		if (error)
1440 			return error;
1441 	}
1442 
1443 	*has_shared = false;
1444 	found = xfs_iext_lookup_extent(ip, ifp, 0, &icur, &got);
1445 	while (found) {
1446 		if (isnullstartblock(got.br_startblock) ||
1447 		    got.br_state != XFS_EXT_NORM)
1448 			goto next;
1449 		agno = XFS_FSB_TO_AGNO(mp, got.br_startblock);
1450 		agbno = XFS_FSB_TO_AGBNO(mp, got.br_startblock);
1451 		aglen = got.br_blockcount;
1452 
1453 		error = xfs_reflink_find_shared(mp, tp, agno, agbno, aglen,
1454 				&rbno, &rlen, false);
1455 		if (error)
1456 			return error;
1457 		/* Is there still a shared block here? */
1458 		if (rbno != NULLAGBLOCK) {
1459 			*has_shared = true;
1460 			return 0;
1461 		}
1462 next:
1463 		found = xfs_iext_next_extent(ifp, &icur, &got);
1464 	}
1465 
1466 	return 0;
1467 }
1468 
1469 /* Clear the inode reflink flag if there are no shared extents. */
1470 int
1471 xfs_reflink_clear_inode_flag(
1472 	struct xfs_inode	*ip,
1473 	struct xfs_trans	**tpp)
1474 {
1475 	bool			needs_flag;
1476 	int			error = 0;
1477 
1478 	ASSERT(xfs_is_reflink_inode(ip));
1479 
1480 	error = xfs_reflink_inode_has_shared_extents(*tpp, ip, &needs_flag);
1481 	if (error || needs_flag)
1482 		return error;
1483 
1484 	/*
1485 	 * We didn't find any shared blocks so turn off the reflink flag.
1486 	 * First, get rid of any leftover CoW mappings.
1487 	 */
1488 	error = xfs_reflink_cancel_cow_blocks(ip, tpp, 0, NULLFILEOFF, true);
1489 	if (error)
1490 		return error;
1491 
1492 	/* Clear the inode flag. */
1493 	trace_xfs_reflink_unset_inode_flag(ip);
1494 	ip->i_d.di_flags2 &= ~XFS_DIFLAG2_REFLINK;
1495 	xfs_inode_clear_cowblocks_tag(ip);
1496 	xfs_trans_ijoin(*tpp, ip, 0);
1497 	xfs_trans_log_inode(*tpp, ip, XFS_ILOG_CORE);
1498 
1499 	return error;
1500 }
1501 
1502 /*
1503  * Clear the inode reflink flag if there are no shared extents and the size
1504  * hasn't changed.
1505  */
1506 STATIC int
1507 xfs_reflink_try_clear_inode_flag(
1508 	struct xfs_inode	*ip)
1509 {
1510 	struct xfs_mount	*mp = ip->i_mount;
1511 	struct xfs_trans	*tp;
1512 	int			error = 0;
1513 
1514 	/* Start a rolling transaction to remove the mappings */
1515 	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, 0, 0, 0, &tp);
1516 	if (error)
1517 		return error;
1518 
1519 	xfs_ilock(ip, XFS_ILOCK_EXCL);
1520 	xfs_trans_ijoin(tp, ip, 0);
1521 
1522 	error = xfs_reflink_clear_inode_flag(ip, &tp);
1523 	if (error)
1524 		goto cancel;
1525 
1526 	error = xfs_trans_commit(tp);
1527 	if (error)
1528 		goto out;
1529 
1530 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1531 	return 0;
1532 cancel:
1533 	xfs_trans_cancel(tp);
1534 out:
1535 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1536 	return error;
1537 }
1538 
1539 /*
1540  * Pre-COW all shared blocks within a given byte range of a file and turn off
1541  * the reflink flag if we unshare all of the file's blocks.
1542  */
1543 int
1544 xfs_reflink_unshare(
1545 	struct xfs_inode	*ip,
1546 	xfs_off_t		offset,
1547 	xfs_off_t		len)
1548 {
1549 	struct xfs_mount	*mp = ip->i_mount;
1550 	xfs_fileoff_t		fbno;
1551 	xfs_filblks_t		end;
1552 	xfs_off_t		isize;
1553 	int			error;
1554 
1555 	if (!xfs_is_reflink_inode(ip))
1556 		return 0;
1557 
1558 	trace_xfs_reflink_unshare(ip, offset, len);
1559 
1560 	inode_dio_wait(VFS_I(ip));
1561 
1562 	/* Try to CoW the selected ranges */
1563 	xfs_ilock(ip, XFS_ILOCK_EXCL);
1564 	fbno = XFS_B_TO_FSBT(mp, offset);
1565 	isize = i_size_read(VFS_I(ip));
1566 	end = XFS_B_TO_FSB(mp, offset + len);
1567 	error = xfs_reflink_dirty_extents(ip, fbno, end, isize);
1568 	if (error)
1569 		goto out_unlock;
1570 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1571 
1572 	/* Wait for the IO to finish */
1573 	error = filemap_write_and_wait(VFS_I(ip)->i_mapping);
1574 	if (error)
1575 		goto out;
1576 
1577 	/* Turn off the reflink flag if possible. */
1578 	error = xfs_reflink_try_clear_inode_flag(ip);
1579 	if (error)
1580 		goto out;
1581 
1582 	return 0;
1583 
1584 out_unlock:
1585 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1586 out:
1587 	trace_xfs_reflink_unshare_error(ip, error, _RET_IP_);
1588 	return error;
1589 }
1590