xref: /openbmc/linux/fs/xfs/xfs_reflink.c (revision feac8c8b)
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_reflink.h"
53 #include "xfs_iomap.h"
54 #include "xfs_rmap_btree.h"
55 #include "xfs_sb.h"
56 #include "xfs_ag_resv.h"
57 
58 /*
59  * Copy on Write of Shared Blocks
60  *
61  * XFS must preserve "the usual" file semantics even when two files share
62  * the same physical blocks.  This means that a write to one file must not
63  * alter the blocks in a different file; the way that we'll do that is
64  * through the use of a copy-on-write mechanism.  At a high level, that
65  * means that when we want to write to a shared block, we allocate a new
66  * block, write the data to the new block, and if that succeeds we map the
67  * new block into the file.
68  *
69  * XFS provides a "delayed allocation" mechanism that defers the allocation
70  * of disk blocks to dirty-but-not-yet-mapped file blocks as long as
71  * possible.  This reduces fragmentation by enabling the filesystem to ask
72  * for bigger chunks less often, which is exactly what we want for CoW.
73  *
74  * The delalloc mechanism begins when the kernel wants to make a block
75  * writable (write_begin or page_mkwrite).  If the offset is not mapped, we
76  * create a delalloc mapping, which is a regular in-core extent, but without
77  * a real startblock.  (For delalloc mappings, the startblock encodes both
78  * a flag that this is a delalloc mapping, and a worst-case estimate of how
79  * many blocks might be required to put the mapping into the BMBT.)  delalloc
80  * mappings are a reservation against the free space in the filesystem;
81  * adjacent mappings can also be combined into fewer larger mappings.
82  *
83  * As an optimization, the CoW extent size hint (cowextsz) creates
84  * outsized aligned delalloc reservations in the hope of landing out of
85  * order nearby CoW writes in a single extent on disk, thereby reducing
86  * fragmentation and improving future performance.
87  *
88  * D: --RRRRRRSSSRRRRRRRR--- (data fork)
89  * C: ------DDDDDDD--------- (CoW fork)
90  *
91  * When dirty pages are being written out (typically in writepage), the
92  * delalloc reservations are converted into unwritten mappings by
93  * allocating blocks and replacing the delalloc mapping with real ones.
94  * A delalloc mapping can be replaced by several unwritten ones if the
95  * free space is fragmented.
96  *
97  * D: --RRRRRRSSSRRRRRRRR---
98  * C: ------UUUUUUU---------
99  *
100  * We want to adapt the delalloc mechanism for copy-on-write, since the
101  * write paths are similar.  The first two steps (creating the reservation
102  * and allocating the blocks) are exactly the same as delalloc except that
103  * the mappings must be stored in a separate CoW fork because we do not want
104  * to disturb the mapping in the data fork until we're sure that the write
105  * succeeded.  IO completion in this case is the process of removing the old
106  * mapping from the data fork and moving the new mapping from the CoW fork to
107  * the data fork.  This will be discussed shortly.
108  *
109  * For now, unaligned directio writes will be bounced back to the page cache.
110  * Block-aligned directio writes will use the same mechanism as buffered
111  * writes.
112  *
113  * Just prior to submitting the actual disk write requests, we convert
114  * the extents representing the range of the file actually being written
115  * (as opposed to extra pieces created for the cowextsize hint) to real
116  * extents.  This will become important in the next step:
117  *
118  * D: --RRRRRRSSSRRRRRRRR---
119  * C: ------UUrrUUU---------
120  *
121  * CoW remapping must be done after the data block write completes,
122  * because we don't want to destroy the old data fork map until we're sure
123  * the new block has been written.  Since the new mappings are kept in a
124  * separate fork, we can simply iterate these mappings to find the ones
125  * that cover the file blocks that we just CoW'd.  For each extent, simply
126  * unmap the corresponding range in the data fork, map the new range into
127  * the data fork, and remove the extent from the CoW fork.  Because of
128  * the presence of the cowextsize hint, however, we must be careful
129  * only to remap the blocks that we've actually written out --  we must
130  * never remap delalloc reservations nor CoW staging blocks that have
131  * yet to be written.  This corresponds exactly to the real extents in
132  * the CoW fork:
133  *
134  * D: --RRRRRRrrSRRRRRRRR---
135  * C: ------UU--UUU---------
136  *
137  * Since the remapping operation can be applied to an arbitrary file
138  * range, we record the need for the remap step as a flag in the ioend
139  * instead of declaring a new IO type.  This is required for direct io
140  * because we only have ioend for the whole dio, and we have to be able to
141  * remember the presence of unwritten blocks and CoW blocks with a single
142  * ioend structure.  Better yet, the more ground we can cover with one
143  * ioend, the better.
144  */
145 
146 /*
147  * Given an AG extent, find the lowest-numbered run of shared blocks
148  * within that range and return the range in fbno/flen.  If
149  * find_end_of_shared is true, return the longest contiguous extent of
150  * shared blocks.  If there are no shared extents, fbno and flen will
151  * be set to NULLAGBLOCK and 0, respectively.
152  */
153 int
154 xfs_reflink_find_shared(
155 	struct xfs_mount	*mp,
156 	struct xfs_trans	*tp,
157 	xfs_agnumber_t		agno,
158 	xfs_agblock_t		agbno,
159 	xfs_extlen_t		aglen,
160 	xfs_agblock_t		*fbno,
161 	xfs_extlen_t		*flen,
162 	bool			find_end_of_shared)
163 {
164 	struct xfs_buf		*agbp;
165 	struct xfs_btree_cur	*cur;
166 	int			error;
167 
168 	error = xfs_alloc_read_agf(mp, tp, agno, 0, &agbp);
169 	if (error)
170 		return error;
171 	if (!agbp)
172 		return -ENOMEM;
173 
174 	cur = xfs_refcountbt_init_cursor(mp, tp, agbp, agno, NULL);
175 
176 	error = xfs_refcount_find_shared(cur, agbno, aglen, fbno, flen,
177 			find_end_of_shared);
178 
179 	xfs_btree_del_cursor(cur, error ? XFS_BTREE_ERROR : XFS_BTREE_NOERROR);
180 
181 	xfs_trans_brelse(tp, agbp);
182 	return error;
183 }
184 
185 /*
186  * Trim the mapping to the next block where there's a change in the
187  * shared/unshared status.  More specifically, this means that we
188  * find the lowest-numbered extent of shared blocks that coincides with
189  * the given block mapping.  If the shared extent overlaps the start of
190  * the mapping, trim the mapping to the end of the shared extent.  If
191  * the shared region intersects the mapping, trim the mapping to the
192  * start of the shared extent.  If there are no shared regions that
193  * overlap, just return the original extent.
194  */
195 int
196 xfs_reflink_trim_around_shared(
197 	struct xfs_inode	*ip,
198 	struct xfs_bmbt_irec	*irec,
199 	bool			*shared,
200 	bool			*trimmed)
201 {
202 	xfs_agnumber_t		agno;
203 	xfs_agblock_t		agbno;
204 	xfs_extlen_t		aglen;
205 	xfs_agblock_t		fbno;
206 	xfs_extlen_t		flen;
207 	int			error = 0;
208 
209 	/* Holes, unwritten, and delalloc extents cannot be shared */
210 	if (!xfs_is_reflink_inode(ip) || !xfs_bmap_is_real_extent(irec)) {
211 		*shared = false;
212 		return 0;
213 	}
214 
215 	trace_xfs_reflink_trim_around_shared(ip, irec);
216 
217 	agno = XFS_FSB_TO_AGNO(ip->i_mount, irec->br_startblock);
218 	agbno = XFS_FSB_TO_AGBNO(ip->i_mount, irec->br_startblock);
219 	aglen = irec->br_blockcount;
220 
221 	error = xfs_reflink_find_shared(ip->i_mount, NULL, agno, agbno,
222 			aglen, &fbno, &flen, true);
223 	if (error)
224 		return error;
225 
226 	*shared = *trimmed = false;
227 	if (fbno == NULLAGBLOCK) {
228 		/* No shared blocks at all. */
229 		return 0;
230 	} else if (fbno == agbno) {
231 		/*
232 		 * The start of this extent is shared.  Truncate the
233 		 * mapping at the end of the shared region so that a
234 		 * subsequent iteration starts at the start of the
235 		 * unshared region.
236 		 */
237 		irec->br_blockcount = flen;
238 		*shared = true;
239 		if (flen != aglen)
240 			*trimmed = true;
241 		return 0;
242 	} else {
243 		/*
244 		 * There's a shared extent midway through this extent.
245 		 * Truncate the mapping at the start of the shared
246 		 * extent so that a subsequent iteration starts at the
247 		 * start of the shared region.
248 		 */
249 		irec->br_blockcount = fbno - agbno;
250 		*trimmed = true;
251 		return 0;
252 	}
253 }
254 
255 /*
256  * Trim the passed in imap to the next shared/unshared extent boundary, and
257  * if imap->br_startoff points to a shared extent reserve space for it in the
258  * COW fork.  In this case *shared is set to true, else to false.
259  *
260  * Note that imap will always contain the block numbers for the existing blocks
261  * in the data fork, as the upper layers need them for read-modify-write
262  * operations.
263  */
264 int
265 xfs_reflink_reserve_cow(
266 	struct xfs_inode	*ip,
267 	struct xfs_bmbt_irec	*imap,
268 	bool			*shared)
269 {
270 	struct xfs_ifork	*ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
271 	struct xfs_bmbt_irec	got;
272 	int			error = 0;
273 	bool			eof = false, trimmed;
274 	struct xfs_iext_cursor	icur;
275 
276 	/*
277 	 * Search the COW fork extent list first.  This serves two purposes:
278 	 * first this implement the speculative preallocation using cowextisze,
279 	 * so that we also unshared block adjacent to shared blocks instead
280 	 * of just the shared blocks themselves.  Second the lookup in the
281 	 * extent list is generally faster than going out to the shared extent
282 	 * tree.
283 	 */
284 
285 	if (!xfs_iext_lookup_extent(ip, ifp, imap->br_startoff, &icur, &got))
286 		eof = true;
287 	if (!eof && got.br_startoff <= imap->br_startoff) {
288 		trace_xfs_reflink_cow_found(ip, imap);
289 		xfs_trim_extent(imap, got.br_startoff, got.br_blockcount);
290 
291 		*shared = true;
292 		return 0;
293 	}
294 
295 	/* Trim the mapping to the nearest shared extent boundary. */
296 	error = xfs_reflink_trim_around_shared(ip, imap, shared, &trimmed);
297 	if (error)
298 		return error;
299 
300 	/* Not shared?  Just report the (potentially capped) extent. */
301 	if (!*shared)
302 		return 0;
303 
304 	/*
305 	 * Fork all the shared blocks from our write offset until the end of
306 	 * the extent.
307 	 */
308 	error = xfs_qm_dqattach_locked(ip, 0);
309 	if (error)
310 		return error;
311 
312 	error = xfs_bmapi_reserve_delalloc(ip, XFS_COW_FORK, imap->br_startoff,
313 			imap->br_blockcount, 0, &got, &icur, eof);
314 	if (error == -ENOSPC || error == -EDQUOT)
315 		trace_xfs_reflink_cow_enospc(ip, imap);
316 	if (error)
317 		return error;
318 
319 	trace_xfs_reflink_cow_alloc(ip, &got);
320 	return 0;
321 }
322 
323 /* Convert part of an unwritten CoW extent to a real one. */
324 STATIC int
325 xfs_reflink_convert_cow_extent(
326 	struct xfs_inode		*ip,
327 	struct xfs_bmbt_irec		*imap,
328 	xfs_fileoff_t			offset_fsb,
329 	xfs_filblks_t			count_fsb,
330 	struct xfs_defer_ops		*dfops)
331 {
332 	xfs_fsblock_t			first_block = NULLFSBLOCK;
333 	int				nimaps = 1;
334 
335 	if (imap->br_state == XFS_EXT_NORM)
336 		return 0;
337 
338 	xfs_trim_extent(imap, offset_fsb, count_fsb);
339 	trace_xfs_reflink_convert_cow(ip, imap);
340 	if (imap->br_blockcount == 0)
341 		return 0;
342 	return xfs_bmapi_write(NULL, ip, imap->br_startoff, imap->br_blockcount,
343 			XFS_BMAPI_COWFORK | XFS_BMAPI_CONVERT, &first_block,
344 			0, imap, &nimaps, dfops);
345 }
346 
347 /* Convert all of the unwritten CoW extents in a file's range to real ones. */
348 int
349 xfs_reflink_convert_cow(
350 	struct xfs_inode	*ip,
351 	xfs_off_t		offset,
352 	xfs_off_t		count)
353 {
354 	struct xfs_mount	*mp = ip->i_mount;
355 	xfs_fileoff_t		offset_fsb = XFS_B_TO_FSBT(mp, offset);
356 	xfs_fileoff_t		end_fsb = XFS_B_TO_FSB(mp, offset + count);
357 	xfs_filblks_t		count_fsb = end_fsb - offset_fsb;
358 	struct xfs_bmbt_irec	imap;
359 	struct xfs_defer_ops	dfops;
360 	xfs_fsblock_t		first_block = NULLFSBLOCK;
361 	int			nimaps = 1, error = 0;
362 
363 	ASSERT(count != 0);
364 
365 	xfs_ilock(ip, XFS_ILOCK_EXCL);
366 	error = xfs_bmapi_write(NULL, ip, offset_fsb, count_fsb,
367 			XFS_BMAPI_COWFORK | XFS_BMAPI_CONVERT |
368 			XFS_BMAPI_CONVERT_ONLY, &first_block, 0, &imap, &nimaps,
369 			&dfops);
370 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
371 	return error;
372 }
373 
374 /* Allocate all CoW reservations covering a range of blocks in a file. */
375 int
376 xfs_reflink_allocate_cow(
377 	struct xfs_inode	*ip,
378 	struct xfs_bmbt_irec	*imap,
379 	bool			*shared,
380 	uint			*lockmode)
381 {
382 	struct xfs_mount	*mp = ip->i_mount;
383 	xfs_fileoff_t		offset_fsb = imap->br_startoff;
384 	xfs_filblks_t		count_fsb = imap->br_blockcount;
385 	struct xfs_bmbt_irec	got;
386 	struct xfs_defer_ops	dfops;
387 	struct xfs_trans	*tp = NULL;
388 	xfs_fsblock_t		first_block;
389 	int			nimaps, error = 0;
390 	bool			trimmed;
391 	xfs_filblks_t		resaligned;
392 	xfs_extlen_t		resblks = 0;
393 	struct xfs_iext_cursor	icur;
394 
395 retry:
396 	ASSERT(xfs_is_reflink_inode(ip));
397 	ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL | XFS_ILOCK_SHARED));
398 
399 	/*
400 	 * Even if the extent is not shared we might have a preallocation for
401 	 * it in the COW fork.  If so use it.
402 	 */
403 	if (xfs_iext_lookup_extent(ip, ip->i_cowfp, offset_fsb, &icur, &got) &&
404 	    got.br_startoff <= offset_fsb) {
405 		*shared = true;
406 
407 		/* If we have a real allocation in the COW fork we're done. */
408 		if (!isnullstartblock(got.br_startblock)) {
409 			xfs_trim_extent(&got, offset_fsb, count_fsb);
410 			*imap = got;
411 			goto convert;
412 		}
413 
414 		xfs_trim_extent(imap, got.br_startoff, got.br_blockcount);
415 	} else {
416 		error = xfs_reflink_trim_around_shared(ip, imap, shared, &trimmed);
417 		if (error || !*shared)
418 			goto out;
419 	}
420 
421 	if (!tp) {
422 		resaligned = xfs_aligned_fsb_count(imap->br_startoff,
423 			imap->br_blockcount, xfs_get_cowextsz_hint(ip));
424 		resblks = XFS_DIOSTRAT_SPACE_RES(mp, resaligned);
425 
426 		xfs_iunlock(ip, *lockmode);
427 		error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, resblks, 0, 0, &tp);
428 		*lockmode = XFS_ILOCK_EXCL;
429 		xfs_ilock(ip, *lockmode);
430 
431 		if (error)
432 			return error;
433 
434 		error = xfs_qm_dqattach_locked(ip, 0);
435 		if (error)
436 			goto out;
437 		goto retry;
438 	}
439 
440 	error = xfs_trans_reserve_quota_nblks(tp, ip, resblks, 0,
441 			XFS_QMOPT_RES_REGBLKS);
442 	if (error)
443 		goto out;
444 
445 	xfs_trans_ijoin(tp, ip, 0);
446 
447 	xfs_defer_init(&dfops, &first_block);
448 	nimaps = 1;
449 
450 	/* Allocate the entire reservation as unwritten blocks. */
451 	error = xfs_bmapi_write(tp, ip, imap->br_startoff, imap->br_blockcount,
452 			XFS_BMAPI_COWFORK | XFS_BMAPI_PREALLOC, &first_block,
453 			resblks, imap, &nimaps, &dfops);
454 	if (error)
455 		goto out_bmap_cancel;
456 
457 	xfs_inode_set_cowblocks_tag(ip);
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 
468 	/*
469 	 * Allocation succeeded but the requested range was not even partially
470 	 * satisfied?  Bail out!
471 	 */
472 	if (nimaps == 0)
473 		return -ENOSPC;
474 convert:
475 	return xfs_reflink_convert_cow_extent(ip, imap, offset_fsb, count_fsb,
476 			&dfops);
477 out_bmap_cancel:
478 	xfs_defer_cancel(&dfops);
479 	xfs_trans_unreserve_quota_nblks(tp, ip, (long)resblks, 0,
480 			XFS_QMOPT_RES_REGBLKS);
481 out:
482 	if (tp)
483 		xfs_trans_cancel(tp);
484 	return error;
485 }
486 
487 /*
488  * Find the CoW reservation for a given byte offset of a file.
489  */
490 bool
491 xfs_reflink_find_cow_mapping(
492 	struct xfs_inode		*ip,
493 	xfs_off_t			offset,
494 	struct xfs_bmbt_irec		*imap)
495 {
496 	struct xfs_ifork		*ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
497 	xfs_fileoff_t			offset_fsb;
498 	struct xfs_bmbt_irec		got;
499 	struct xfs_iext_cursor		icur;
500 
501 	ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL | XFS_ILOCK_SHARED));
502 
503 	if (!xfs_is_reflink_inode(ip))
504 		return false;
505 	offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
506 	if (!xfs_iext_lookup_extent(ip, ifp, offset_fsb, &icur, &got))
507 		return false;
508 	if (got.br_startoff > offset_fsb)
509 		return false;
510 
511 	trace_xfs_reflink_find_cow_mapping(ip, offset, 1, XFS_IO_OVERWRITE,
512 			&got);
513 	*imap = got;
514 	return true;
515 }
516 
517 /*
518  * Trim an extent to end at the next CoW reservation past offset_fsb.
519  */
520 void
521 xfs_reflink_trim_irec_to_next_cow(
522 	struct xfs_inode		*ip,
523 	xfs_fileoff_t			offset_fsb,
524 	struct xfs_bmbt_irec		*imap)
525 {
526 	struct xfs_ifork		*ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
527 	struct xfs_bmbt_irec		got;
528 	struct xfs_iext_cursor		icur;
529 
530 	if (!xfs_is_reflink_inode(ip))
531 		return;
532 
533 	/* Find the extent in the CoW fork. */
534 	if (!xfs_iext_lookup_extent(ip, ifp, offset_fsb, &icur, &got))
535 		return;
536 
537 	/* This is the extent before; try sliding up one. */
538 	if (got.br_startoff < offset_fsb) {
539 		if (!xfs_iext_next_extent(ifp, &icur, &got))
540 			return;
541 	}
542 
543 	if (got.br_startoff >= imap->br_startoff + imap->br_blockcount)
544 		return;
545 
546 	imap->br_blockcount = got.br_startoff - imap->br_startoff;
547 	trace_xfs_reflink_trim_irec(ip, imap);
548 }
549 
550 /*
551  * Cancel CoW reservations for some block range of an inode.
552  *
553  * If cancel_real is true this function cancels all COW fork extents for the
554  * inode; if cancel_real is false, real extents are not cleared.
555  */
556 int
557 xfs_reflink_cancel_cow_blocks(
558 	struct xfs_inode		*ip,
559 	struct xfs_trans		**tpp,
560 	xfs_fileoff_t			offset_fsb,
561 	xfs_fileoff_t			end_fsb,
562 	bool				cancel_real)
563 {
564 	struct xfs_ifork		*ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
565 	struct xfs_bmbt_irec		got, del;
566 	struct xfs_iext_cursor		icur;
567 	xfs_fsblock_t			firstfsb;
568 	struct xfs_defer_ops		dfops;
569 	int				error = 0;
570 
571 	if (!xfs_is_reflink_inode(ip))
572 		return 0;
573 	if (!xfs_iext_lookup_extent_before(ip, ifp, &end_fsb, &icur, &got))
574 		return 0;
575 
576 	/* Walk backwards until we're out of the I/O range... */
577 	while (got.br_startoff + got.br_blockcount > offset_fsb) {
578 		del = got;
579 		xfs_trim_extent(&del, offset_fsb, end_fsb - offset_fsb);
580 
581 		/* Extent delete may have bumped ext forward */
582 		if (!del.br_blockcount) {
583 			xfs_iext_prev(ifp, &icur);
584 			goto next_extent;
585 		}
586 
587 		trace_xfs_reflink_cancel_cow(ip, &del);
588 
589 		if (isnullstartblock(del.br_startblock)) {
590 			error = xfs_bmap_del_extent_delay(ip, XFS_COW_FORK,
591 					&icur, &got, &del);
592 			if (error)
593 				break;
594 		} else if (del.br_state == XFS_EXT_UNWRITTEN || cancel_real) {
595 			xfs_trans_ijoin(*tpp, ip, 0);
596 			xfs_defer_init(&dfops, &firstfsb);
597 
598 			/* Free the CoW orphan record. */
599 			error = xfs_refcount_free_cow_extent(ip->i_mount,
600 					&dfops, del.br_startblock,
601 					del.br_blockcount);
602 			if (error)
603 				break;
604 
605 			xfs_bmap_add_free(ip->i_mount, &dfops,
606 					del.br_startblock, del.br_blockcount,
607 					NULL);
608 
609 			/* Roll the transaction */
610 			xfs_defer_ijoin(&dfops, ip);
611 			error = xfs_defer_finish(tpp, &dfops);
612 			if (error) {
613 				xfs_defer_cancel(&dfops);
614 				break;
615 			}
616 
617 			/* Remove the mapping from the CoW fork. */
618 			xfs_bmap_del_extent_cow(ip, &icur, &got, &del);
619 
620 			/* Remove the quota reservation */
621 			error = xfs_trans_reserve_quota_nblks(NULL, ip,
622 					-(long)del.br_blockcount, 0,
623 					XFS_QMOPT_RES_REGBLKS);
624 			if (error)
625 				break;
626 		} else {
627 			/* Didn't do anything, push cursor back. */
628 			xfs_iext_prev(ifp, &icur);
629 		}
630 next_extent:
631 		if (!xfs_iext_get_extent(ifp, &icur, &got))
632 			break;
633 	}
634 
635 	/* clear tag if cow fork is emptied */
636 	if (!ifp->if_bytes)
637 		xfs_inode_clear_cowblocks_tag(ip);
638 
639 	return error;
640 }
641 
642 /*
643  * Cancel CoW reservations for some byte range of an inode.
644  *
645  * If cancel_real is true this function cancels all COW fork extents for the
646  * inode; if cancel_real is false, real extents are not cleared.
647  */
648 int
649 xfs_reflink_cancel_cow_range(
650 	struct xfs_inode	*ip,
651 	xfs_off_t		offset,
652 	xfs_off_t		count,
653 	bool			cancel_real)
654 {
655 	struct xfs_trans	*tp;
656 	xfs_fileoff_t		offset_fsb;
657 	xfs_fileoff_t		end_fsb;
658 	int			error;
659 
660 	trace_xfs_reflink_cancel_cow_range(ip, offset, count);
661 	ASSERT(xfs_is_reflink_inode(ip));
662 
663 	offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
664 	if (count == NULLFILEOFF)
665 		end_fsb = NULLFILEOFF;
666 	else
667 		end_fsb = XFS_B_TO_FSB(ip->i_mount, offset + count);
668 
669 	/* Start a rolling transaction to remove the mappings */
670 	error = xfs_trans_alloc(ip->i_mount, &M_RES(ip->i_mount)->tr_write,
671 			0, 0, 0, &tp);
672 	if (error)
673 		goto out;
674 
675 	xfs_ilock(ip, XFS_ILOCK_EXCL);
676 	xfs_trans_ijoin(tp, ip, 0);
677 
678 	/* Scrape out the old CoW reservations */
679 	error = xfs_reflink_cancel_cow_blocks(ip, &tp, offset_fsb, end_fsb,
680 			cancel_real);
681 	if (error)
682 		goto out_cancel;
683 
684 	error = xfs_trans_commit(tp);
685 
686 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
687 	return error;
688 
689 out_cancel:
690 	xfs_trans_cancel(tp);
691 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
692 out:
693 	trace_xfs_reflink_cancel_cow_range_error(ip, error, _RET_IP_);
694 	return error;
695 }
696 
697 /*
698  * Remap parts of a file's data fork after a successful CoW.
699  */
700 int
701 xfs_reflink_end_cow(
702 	struct xfs_inode		*ip,
703 	xfs_off_t			offset,
704 	xfs_off_t			count)
705 {
706 	struct xfs_ifork		*ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
707 	struct xfs_bmbt_irec		got, del;
708 	struct xfs_trans		*tp;
709 	xfs_fileoff_t			offset_fsb;
710 	xfs_fileoff_t			end_fsb;
711 	xfs_fsblock_t			firstfsb;
712 	struct xfs_defer_ops		dfops;
713 	int				error;
714 	unsigned int			resblks;
715 	xfs_filblks_t			rlen;
716 	struct xfs_iext_cursor		icur;
717 
718 	trace_xfs_reflink_end_cow(ip, offset, count);
719 
720 	/* No COW extents?  That's easy! */
721 	if (ifp->if_bytes == 0)
722 		return 0;
723 
724 	offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
725 	end_fsb = XFS_B_TO_FSB(ip->i_mount, offset + count);
726 
727 	/*
728 	 * Start a rolling transaction to switch the mappings.  We're
729 	 * unlikely ever to have to remap 16T worth of single-block
730 	 * extents, so just cap the worst case extent count to 2^32-1.
731 	 * Stick a warning in just in case, and avoid 64-bit division.
732 	 */
733 	BUILD_BUG_ON(MAX_RW_COUNT > UINT_MAX);
734 	if (end_fsb - offset_fsb > UINT_MAX) {
735 		error = -EFSCORRUPTED;
736 		xfs_force_shutdown(ip->i_mount, SHUTDOWN_CORRUPT_INCORE);
737 		ASSERT(0);
738 		goto out;
739 	}
740 	resblks = XFS_NEXTENTADD_SPACE_RES(ip->i_mount,
741 			(unsigned int)(end_fsb - offset_fsb),
742 			XFS_DATA_FORK);
743 	error = xfs_trans_alloc(ip->i_mount, &M_RES(ip->i_mount)->tr_write,
744 			resblks, 0, XFS_TRANS_RESERVE, &tp);
745 	if (error)
746 		goto out;
747 
748 	xfs_ilock(ip, XFS_ILOCK_EXCL);
749 	xfs_trans_ijoin(tp, ip, 0);
750 
751 	/*
752 	 * In case of racing, overlapping AIO writes no COW extents might be
753 	 * left by the time I/O completes for the loser of the race.  In that
754 	 * case we are done.
755 	 */
756 	if (!xfs_iext_lookup_extent_before(ip, ifp, &end_fsb, &icur, &got))
757 		goto out_cancel;
758 
759 	/* Walk backwards until we're out of the I/O range... */
760 	while (got.br_startoff + got.br_blockcount > offset_fsb) {
761 		del = got;
762 		xfs_trim_extent(&del, offset_fsb, end_fsb - offset_fsb);
763 
764 		/* Extent delete may have bumped ext forward */
765 		if (!del.br_blockcount) {
766 			xfs_iext_prev(ifp, &icur);
767 			goto next_extent;
768 		}
769 
770 		ASSERT(!isnullstartblock(got.br_startblock));
771 
772 		/*
773 		 * Don't remap unwritten extents; these are
774 		 * speculatively preallocated CoW extents that have been
775 		 * allocated but have not yet been involved in a write.
776 		 */
777 		if (got.br_state == XFS_EXT_UNWRITTEN) {
778 			xfs_iext_prev(ifp, &icur);
779 			goto next_extent;
780 		}
781 
782 		/* Unmap the old blocks in the data fork. */
783 		xfs_defer_init(&dfops, &firstfsb);
784 		rlen = del.br_blockcount;
785 		error = __xfs_bunmapi(tp, ip, del.br_startoff, &rlen, 0, 1,
786 				&firstfsb, &dfops);
787 		if (error)
788 			goto out_defer;
789 
790 		/* Trim the extent to whatever got unmapped. */
791 		if (rlen) {
792 			xfs_trim_extent(&del, del.br_startoff + rlen,
793 				del.br_blockcount - rlen);
794 		}
795 		trace_xfs_reflink_cow_remap(ip, &del);
796 
797 		/* Free the CoW orphan record. */
798 		error = xfs_refcount_free_cow_extent(tp->t_mountp, &dfops,
799 				del.br_startblock, del.br_blockcount);
800 		if (error)
801 			goto out_defer;
802 
803 		/* Map the new blocks into the data fork. */
804 		error = xfs_bmap_map_extent(tp->t_mountp, &dfops, ip, &del);
805 		if (error)
806 			goto out_defer;
807 
808 		/* Charge this new data fork mapping to the on-disk quota. */
809 		xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_DELBCOUNT,
810 				(long)del.br_blockcount);
811 
812 		/* Remove the mapping from the CoW fork. */
813 		xfs_bmap_del_extent_cow(ip, &icur, &got, &del);
814 
815 		xfs_defer_ijoin(&dfops, ip);
816 		error = xfs_defer_finish(&tp, &dfops);
817 		if (error)
818 			goto out_defer;
819 next_extent:
820 		if (!xfs_iext_get_extent(ifp, &icur, &got))
821 			break;
822 	}
823 
824 	error = xfs_trans_commit(tp);
825 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
826 	if (error)
827 		goto out;
828 	return 0;
829 
830 out_defer:
831 	xfs_defer_cancel(&dfops);
832 out_cancel:
833 	xfs_trans_cancel(tp);
834 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
835 out:
836 	trace_xfs_reflink_end_cow_error(ip, error, _RET_IP_);
837 	return error;
838 }
839 
840 /*
841  * Free leftover CoW reservations that didn't get cleaned out.
842  */
843 int
844 xfs_reflink_recover_cow(
845 	struct xfs_mount	*mp)
846 {
847 	xfs_agnumber_t		agno;
848 	int			error = 0;
849 
850 	if (!xfs_sb_version_hasreflink(&mp->m_sb))
851 		return 0;
852 
853 	for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
854 		error = xfs_refcount_recover_cow_leftovers(mp, agno);
855 		if (error)
856 			break;
857 	}
858 
859 	return error;
860 }
861 
862 /*
863  * Reflinking (Block) Ranges of Two Files Together
864  *
865  * First, ensure that the reflink flag is set on both inodes.  The flag is an
866  * optimization to avoid unnecessary refcount btree lookups in the write path.
867  *
868  * Now we can iteratively remap the range of extents (and holes) in src to the
869  * corresponding ranges in dest.  Let drange and srange denote the ranges of
870  * logical blocks in dest and src touched by the reflink operation.
871  *
872  * While the length of drange is greater than zero,
873  *    - Read src's bmbt at the start of srange ("imap")
874  *    - If imap doesn't exist, make imap appear to start at the end of srange
875  *      with zero length.
876  *    - If imap starts before srange, advance imap to start at srange.
877  *    - If imap goes beyond srange, truncate imap to end at the end of srange.
878  *    - Punch (imap start - srange start + imap len) blocks from dest at
879  *      offset (drange start).
880  *    - If imap points to a real range of pblks,
881  *         > Increase the refcount of the imap's pblks
882  *         > Map imap's pblks into dest at the offset
883  *           (drange start + imap start - srange start)
884  *    - Advance drange and srange by (imap start - srange start + imap len)
885  *
886  * Finally, if the reflink made dest longer, update both the in-core and
887  * on-disk file sizes.
888  *
889  * ASCII Art Demonstration:
890  *
891  * Let's say we want to reflink this source file:
892  *
893  * ----SSSSSSS-SSSSS----SSSSSS (src file)
894  *   <-------------------->
895  *
896  * into this destination file:
897  *
898  * --DDDDDDDDDDDDDDDDDDD--DDD (dest file)
899  *        <-------------------->
900  * '-' means a hole, and 'S' and 'D' are written blocks in the src and dest.
901  * Observe that the range has different logical offsets in either file.
902  *
903  * Consider that the first extent in the source file doesn't line up with our
904  * reflink range.  Unmapping  and remapping are separate operations, so we can
905  * unmap more blocks from the destination file than we remap.
906  *
907  * ----SSSSSSS-SSSSS----SSSSSS
908  *   <------->
909  * --DDDDD---------DDDDD--DDD
910  *        <------->
911  *
912  * Now remap the source extent into the destination file:
913  *
914  * ----SSSSSSS-SSSSS----SSSSSS
915  *   <------->
916  * --DDDDD--SSSSSSSDDDDD--DDD
917  *        <------->
918  *
919  * Do likewise with the second hole and extent in our range.  Holes in the
920  * unmap range don't affect our operation.
921  *
922  * ----SSSSSSS-SSSSS----SSSSSS
923  *            <---->
924  * --DDDDD--SSSSSSS-SSSSS-DDD
925  *                 <---->
926  *
927  * Finally, unmap and remap part of the third extent.  This will increase the
928  * size of the destination file.
929  *
930  * ----SSSSSSS-SSSSS----SSSSSS
931  *                  <----->
932  * --DDDDD--SSSSSSS-SSSSS----SSS
933  *                       <----->
934  *
935  * Once we update the destination file's i_size, we're done.
936  */
937 
938 /*
939  * Ensure the reflink bit is set in both inodes.
940  */
941 STATIC int
942 xfs_reflink_set_inode_flag(
943 	struct xfs_inode	*src,
944 	struct xfs_inode	*dest)
945 {
946 	struct xfs_mount	*mp = src->i_mount;
947 	int			error;
948 	struct xfs_trans	*tp;
949 
950 	if (xfs_is_reflink_inode(src) && xfs_is_reflink_inode(dest))
951 		return 0;
952 
953 	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
954 	if (error)
955 		goto out_error;
956 
957 	/* Lock both files against IO */
958 	if (src->i_ino == dest->i_ino)
959 		xfs_ilock(src, XFS_ILOCK_EXCL);
960 	else
961 		xfs_lock_two_inodes(src, XFS_ILOCK_EXCL, dest, XFS_ILOCK_EXCL);
962 
963 	if (!xfs_is_reflink_inode(src)) {
964 		trace_xfs_reflink_set_inode_flag(src);
965 		xfs_trans_ijoin(tp, src, XFS_ILOCK_EXCL);
966 		src->i_d.di_flags2 |= XFS_DIFLAG2_REFLINK;
967 		xfs_trans_log_inode(tp, src, XFS_ILOG_CORE);
968 		xfs_ifork_init_cow(src);
969 	} else
970 		xfs_iunlock(src, XFS_ILOCK_EXCL);
971 
972 	if (src->i_ino == dest->i_ino)
973 		goto commit_flags;
974 
975 	if (!xfs_is_reflink_inode(dest)) {
976 		trace_xfs_reflink_set_inode_flag(dest);
977 		xfs_trans_ijoin(tp, dest, XFS_ILOCK_EXCL);
978 		dest->i_d.di_flags2 |= XFS_DIFLAG2_REFLINK;
979 		xfs_trans_log_inode(tp, dest, XFS_ILOG_CORE);
980 		xfs_ifork_init_cow(dest);
981 	} else
982 		xfs_iunlock(dest, XFS_ILOCK_EXCL);
983 
984 commit_flags:
985 	error = xfs_trans_commit(tp);
986 	if (error)
987 		goto out_error;
988 	return error;
989 
990 out_error:
991 	trace_xfs_reflink_set_inode_flag_error(dest, error, _RET_IP_);
992 	return error;
993 }
994 
995 /*
996  * Update destination inode size & cowextsize hint, if necessary.
997  */
998 STATIC int
999 xfs_reflink_update_dest(
1000 	struct xfs_inode	*dest,
1001 	xfs_off_t		newlen,
1002 	xfs_extlen_t		cowextsize,
1003 	bool			is_dedupe)
1004 {
1005 	struct xfs_mount	*mp = dest->i_mount;
1006 	struct xfs_trans	*tp;
1007 	int			error;
1008 
1009 	if (is_dedupe && newlen <= i_size_read(VFS_I(dest)) && cowextsize == 0)
1010 		return 0;
1011 
1012 	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
1013 	if (error)
1014 		goto out_error;
1015 
1016 	xfs_ilock(dest, XFS_ILOCK_EXCL);
1017 	xfs_trans_ijoin(tp, dest, XFS_ILOCK_EXCL);
1018 
1019 	if (newlen > i_size_read(VFS_I(dest))) {
1020 		trace_xfs_reflink_update_inode_size(dest, newlen);
1021 		i_size_write(VFS_I(dest), newlen);
1022 		dest->i_d.di_size = newlen;
1023 	}
1024 
1025 	if (cowextsize) {
1026 		dest->i_d.di_cowextsize = cowextsize;
1027 		dest->i_d.di_flags2 |= XFS_DIFLAG2_COWEXTSIZE;
1028 	}
1029 
1030 	if (!is_dedupe) {
1031 		xfs_trans_ichgtime(tp, dest,
1032 				   XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1033 	}
1034 	xfs_trans_log_inode(tp, dest, XFS_ILOG_CORE);
1035 
1036 	error = xfs_trans_commit(tp);
1037 	if (error)
1038 		goto out_error;
1039 	return error;
1040 
1041 out_error:
1042 	trace_xfs_reflink_update_inode_size_error(dest, error, _RET_IP_);
1043 	return error;
1044 }
1045 
1046 /*
1047  * Do we have enough reserve in this AG to handle a reflink?  The refcount
1048  * btree already reserved all the space it needs, but the rmap btree can grow
1049  * infinitely, so we won't allow more reflinks when the AG is down to the
1050  * btree reserves.
1051  */
1052 static int
1053 xfs_reflink_ag_has_free_space(
1054 	struct xfs_mount	*mp,
1055 	xfs_agnumber_t		agno)
1056 {
1057 	struct xfs_perag	*pag;
1058 	int			error = 0;
1059 
1060 	if (!xfs_sb_version_hasrmapbt(&mp->m_sb))
1061 		return 0;
1062 
1063 	pag = xfs_perag_get(mp, agno);
1064 	if (xfs_ag_resv_critical(pag, XFS_AG_RESV_AGFL) ||
1065 	    xfs_ag_resv_critical(pag, XFS_AG_RESV_METADATA))
1066 		error = -ENOSPC;
1067 	xfs_perag_put(pag);
1068 	return error;
1069 }
1070 
1071 /*
1072  * Unmap a range of blocks from a file, then map other blocks into the hole.
1073  * The range to unmap is (destoff : destoff + srcioff + irec->br_blockcount).
1074  * The extent irec is mapped into dest at irec->br_startoff.
1075  */
1076 STATIC int
1077 xfs_reflink_remap_extent(
1078 	struct xfs_inode	*ip,
1079 	struct xfs_bmbt_irec	*irec,
1080 	xfs_fileoff_t		destoff,
1081 	xfs_off_t		new_isize)
1082 {
1083 	struct xfs_mount	*mp = ip->i_mount;
1084 	bool			real_extent = xfs_bmap_is_real_extent(irec);
1085 	struct xfs_trans	*tp;
1086 	xfs_fsblock_t		firstfsb;
1087 	unsigned int		resblks;
1088 	struct xfs_defer_ops	dfops;
1089 	struct xfs_bmbt_irec	uirec;
1090 	xfs_filblks_t		rlen;
1091 	xfs_filblks_t		unmap_len;
1092 	xfs_off_t		newlen;
1093 	int			error;
1094 
1095 	unmap_len = irec->br_startoff + irec->br_blockcount - destoff;
1096 	trace_xfs_reflink_punch_range(ip, destoff, unmap_len);
1097 
1098 	/* No reflinking if we're low on space */
1099 	if (real_extent) {
1100 		error = xfs_reflink_ag_has_free_space(mp,
1101 				XFS_FSB_TO_AGNO(mp, irec->br_startblock));
1102 		if (error)
1103 			goto out;
1104 	}
1105 
1106 	/* Start a rolling transaction to switch the mappings */
1107 	resblks = XFS_EXTENTADD_SPACE_RES(ip->i_mount, XFS_DATA_FORK);
1108 	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, resblks, 0, 0, &tp);
1109 	if (error)
1110 		goto out;
1111 
1112 	xfs_ilock(ip, XFS_ILOCK_EXCL);
1113 	xfs_trans_ijoin(tp, ip, 0);
1114 
1115 	/* If we're not just clearing space, then do we have enough quota? */
1116 	if (real_extent) {
1117 		error = xfs_trans_reserve_quota_nblks(tp, ip,
1118 				irec->br_blockcount, 0, XFS_QMOPT_RES_REGBLKS);
1119 		if (error)
1120 			goto out_cancel;
1121 	}
1122 
1123 	trace_xfs_reflink_remap(ip, irec->br_startoff,
1124 				irec->br_blockcount, irec->br_startblock);
1125 
1126 	/* Unmap the old blocks in the data fork. */
1127 	rlen = unmap_len;
1128 	while (rlen) {
1129 		xfs_defer_init(&dfops, &firstfsb);
1130 		error = __xfs_bunmapi(tp, ip, destoff, &rlen, 0, 1,
1131 				&firstfsb, &dfops);
1132 		if (error)
1133 			goto out_defer;
1134 
1135 		/*
1136 		 * Trim the extent to whatever got unmapped.
1137 		 * Remember, bunmapi works backwards.
1138 		 */
1139 		uirec.br_startblock = irec->br_startblock + rlen;
1140 		uirec.br_startoff = irec->br_startoff + rlen;
1141 		uirec.br_blockcount = unmap_len - rlen;
1142 		unmap_len = rlen;
1143 
1144 		/* If this isn't a real mapping, we're done. */
1145 		if (!real_extent || uirec.br_blockcount == 0)
1146 			goto next_extent;
1147 
1148 		trace_xfs_reflink_remap(ip, uirec.br_startoff,
1149 				uirec.br_blockcount, uirec.br_startblock);
1150 
1151 		/* Update the refcount tree */
1152 		error = xfs_refcount_increase_extent(mp, &dfops, &uirec);
1153 		if (error)
1154 			goto out_defer;
1155 
1156 		/* Map the new blocks into the data fork. */
1157 		error = xfs_bmap_map_extent(mp, &dfops, ip, &uirec);
1158 		if (error)
1159 			goto out_defer;
1160 
1161 		/* Update quota accounting. */
1162 		xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_BCOUNT,
1163 				uirec.br_blockcount);
1164 
1165 		/* Update dest isize if needed. */
1166 		newlen = XFS_FSB_TO_B(mp,
1167 				uirec.br_startoff + uirec.br_blockcount);
1168 		newlen = min_t(xfs_off_t, newlen, new_isize);
1169 		if (newlen > i_size_read(VFS_I(ip))) {
1170 			trace_xfs_reflink_update_inode_size(ip, newlen);
1171 			i_size_write(VFS_I(ip), newlen);
1172 			ip->i_d.di_size = newlen;
1173 			xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1174 		}
1175 
1176 next_extent:
1177 		/* Process all the deferred stuff. */
1178 		xfs_defer_ijoin(&dfops, ip);
1179 		error = xfs_defer_finish(&tp, &dfops);
1180 		if (error)
1181 			goto out_defer;
1182 	}
1183 
1184 	error = xfs_trans_commit(tp);
1185 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1186 	if (error)
1187 		goto out;
1188 	return 0;
1189 
1190 out_defer:
1191 	xfs_defer_cancel(&dfops);
1192 out_cancel:
1193 	xfs_trans_cancel(tp);
1194 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1195 out:
1196 	trace_xfs_reflink_remap_extent_error(ip, error, _RET_IP_);
1197 	return error;
1198 }
1199 
1200 /*
1201  * Iteratively remap one file's extents (and holes) to another's.
1202  */
1203 STATIC int
1204 xfs_reflink_remap_blocks(
1205 	struct xfs_inode	*src,
1206 	xfs_fileoff_t		srcoff,
1207 	struct xfs_inode	*dest,
1208 	xfs_fileoff_t		destoff,
1209 	xfs_filblks_t		len,
1210 	xfs_off_t		new_isize)
1211 {
1212 	struct xfs_bmbt_irec	imap;
1213 	int			nimaps;
1214 	int			error = 0;
1215 	xfs_filblks_t		range_len;
1216 
1217 	/* drange = (destoff, destoff + len); srange = (srcoff, srcoff + len) */
1218 	while (len) {
1219 		uint		lock_mode;
1220 
1221 		trace_xfs_reflink_remap_blocks_loop(src, srcoff, len,
1222 				dest, destoff);
1223 
1224 		/* Read extent from the source file */
1225 		nimaps = 1;
1226 		lock_mode = xfs_ilock_data_map_shared(src);
1227 		error = xfs_bmapi_read(src, srcoff, len, &imap, &nimaps, 0);
1228 		xfs_iunlock(src, lock_mode);
1229 		if (error)
1230 			goto err;
1231 		ASSERT(nimaps == 1);
1232 
1233 		trace_xfs_reflink_remap_imap(src, srcoff, len, XFS_IO_OVERWRITE,
1234 				&imap);
1235 
1236 		/* Translate imap into the destination file. */
1237 		range_len = imap.br_startoff + imap.br_blockcount - srcoff;
1238 		imap.br_startoff += destoff - srcoff;
1239 
1240 		/* Clear dest from destoff to the end of imap and map it in. */
1241 		error = xfs_reflink_remap_extent(dest, &imap, destoff,
1242 				new_isize);
1243 		if (error)
1244 			goto err;
1245 
1246 		if (fatal_signal_pending(current)) {
1247 			error = -EINTR;
1248 			goto err;
1249 		}
1250 
1251 		/* Advance drange/srange */
1252 		srcoff += range_len;
1253 		destoff += range_len;
1254 		len -= range_len;
1255 	}
1256 
1257 	return 0;
1258 
1259 err:
1260 	trace_xfs_reflink_remap_blocks_error(dest, error, _RET_IP_);
1261 	return error;
1262 }
1263 
1264 /*
1265  * Grab the exclusive iolock for a data copy from src to dest, making
1266  * sure to abide vfs locking order (lowest pointer value goes first) and
1267  * breaking the pnfs layout leases on dest before proceeding.  The loop
1268  * is needed because we cannot call the blocking break_layout() with the
1269  * src iolock held, and therefore have to back out both locks.
1270  */
1271 static int
1272 xfs_iolock_two_inodes_and_break_layout(
1273 	struct inode		*src,
1274 	struct inode		*dest)
1275 {
1276 	int			error;
1277 
1278 retry:
1279 	if (src < dest) {
1280 		inode_lock_shared(src);
1281 		inode_lock_nested(dest, I_MUTEX_NONDIR2);
1282 	} else {
1283 		/* src >= dest */
1284 		inode_lock(dest);
1285 	}
1286 
1287 	error = break_layout(dest, false);
1288 	if (error == -EWOULDBLOCK) {
1289 		inode_unlock(dest);
1290 		if (src < dest)
1291 			inode_unlock_shared(src);
1292 		error = break_layout(dest, true);
1293 		if (error)
1294 			return error;
1295 		goto retry;
1296 	}
1297 	if (error) {
1298 		inode_unlock(dest);
1299 		if (src < dest)
1300 			inode_unlock_shared(src);
1301 		return error;
1302 	}
1303 	if (src > dest)
1304 		inode_lock_shared_nested(src, I_MUTEX_NONDIR2);
1305 	return 0;
1306 }
1307 
1308 /*
1309  * Link a range of blocks from one file to another.
1310  */
1311 int
1312 xfs_reflink_remap_range(
1313 	struct file		*file_in,
1314 	loff_t			pos_in,
1315 	struct file		*file_out,
1316 	loff_t			pos_out,
1317 	u64			len,
1318 	bool			is_dedupe)
1319 {
1320 	struct inode		*inode_in = file_inode(file_in);
1321 	struct xfs_inode	*src = XFS_I(inode_in);
1322 	struct inode		*inode_out = file_inode(file_out);
1323 	struct xfs_inode	*dest = XFS_I(inode_out);
1324 	struct xfs_mount	*mp = src->i_mount;
1325 	bool			same_inode = (inode_in == inode_out);
1326 	xfs_fileoff_t		sfsbno, dfsbno;
1327 	xfs_filblks_t		fsblen;
1328 	xfs_extlen_t		cowextsize;
1329 	ssize_t			ret;
1330 
1331 	if (!xfs_sb_version_hasreflink(&mp->m_sb))
1332 		return -EOPNOTSUPP;
1333 
1334 	if (XFS_FORCED_SHUTDOWN(mp))
1335 		return -EIO;
1336 
1337 	/* Lock both files against IO */
1338 	ret = xfs_iolock_two_inodes_and_break_layout(inode_in, inode_out);
1339 	if (ret)
1340 		return ret;
1341 	if (same_inode)
1342 		xfs_ilock(src, XFS_MMAPLOCK_EXCL);
1343 	else
1344 		xfs_lock_two_inodes(src, XFS_MMAPLOCK_SHARED, dest,
1345 				XFS_MMAPLOCK_EXCL);
1346 
1347 	/* Check file eligibility and prepare for block sharing. */
1348 	ret = -EINVAL;
1349 	/* Don't reflink realtime inodes */
1350 	if (XFS_IS_REALTIME_INODE(src) || XFS_IS_REALTIME_INODE(dest))
1351 		goto out_unlock;
1352 
1353 	/* Don't share DAX file data for now. */
1354 	if (IS_DAX(inode_in) || IS_DAX(inode_out))
1355 		goto out_unlock;
1356 
1357 	ret = vfs_clone_file_prep_inodes(inode_in, pos_in, inode_out, pos_out,
1358 			&len, is_dedupe);
1359 	if (ret <= 0)
1360 		goto out_unlock;
1361 
1362 	/* Attach dquots to dest inode before changing block map */
1363 	ret = xfs_qm_dqattach(dest, 0);
1364 	if (ret)
1365 		goto out_unlock;
1366 
1367 	trace_xfs_reflink_remap_range(src, pos_in, len, dest, pos_out);
1368 
1369 	/*
1370 	 * Clear out post-eof preallocations because we don't have page cache
1371 	 * backing the delayed allocations and they'll never get freed on
1372 	 * their own.
1373 	 */
1374 	if (xfs_can_free_eofblocks(dest, true)) {
1375 		ret = xfs_free_eofblocks(dest);
1376 		if (ret)
1377 			goto out_unlock;
1378 	}
1379 
1380 	/* Set flags and remap blocks. */
1381 	ret = xfs_reflink_set_inode_flag(src, dest);
1382 	if (ret)
1383 		goto out_unlock;
1384 
1385 	dfsbno = XFS_B_TO_FSBT(mp, pos_out);
1386 	sfsbno = XFS_B_TO_FSBT(mp, pos_in);
1387 	fsblen = XFS_B_TO_FSB(mp, len);
1388 	ret = xfs_reflink_remap_blocks(src, sfsbno, dest, dfsbno, fsblen,
1389 			pos_out + len);
1390 	if (ret)
1391 		goto out_unlock;
1392 
1393 	/* Zap any page cache for the destination file's range. */
1394 	truncate_inode_pages_range(&inode_out->i_data, pos_out,
1395 				   PAGE_ALIGN(pos_out + len) - 1);
1396 
1397 	/*
1398 	 * Carry the cowextsize hint from src to dest if we're sharing the
1399 	 * entire source file to the entire destination file, the source file
1400 	 * has a cowextsize hint, and the destination file does not.
1401 	 */
1402 	cowextsize = 0;
1403 	if (pos_in == 0 && len == i_size_read(inode_in) &&
1404 	    (src->i_d.di_flags2 & XFS_DIFLAG2_COWEXTSIZE) &&
1405 	    pos_out == 0 && len >= i_size_read(inode_out) &&
1406 	    !(dest->i_d.di_flags2 & XFS_DIFLAG2_COWEXTSIZE))
1407 		cowextsize = src->i_d.di_cowextsize;
1408 
1409 	ret = xfs_reflink_update_dest(dest, pos_out + len, cowextsize,
1410 			is_dedupe);
1411 
1412 out_unlock:
1413 	xfs_iunlock(dest, XFS_MMAPLOCK_EXCL);
1414 	if (!same_inode)
1415 		xfs_iunlock(src, XFS_MMAPLOCK_SHARED);
1416 	inode_unlock(inode_out);
1417 	if (!same_inode)
1418 		inode_unlock_shared(inode_in);
1419 	if (ret)
1420 		trace_xfs_reflink_remap_range_error(dest, ret, _RET_IP_);
1421 	return ret;
1422 }
1423 
1424 /*
1425  * The user wants to preemptively CoW all shared blocks in this file,
1426  * which enables us to turn off the reflink flag.  Iterate all
1427  * extents which are not prealloc/delalloc to see which ranges are
1428  * mentioned in the refcount tree, then read those blocks into the
1429  * pagecache, dirty them, fsync them back out, and then we can update
1430  * the inode flag.  What happens if we run out of memory? :)
1431  */
1432 STATIC int
1433 xfs_reflink_dirty_extents(
1434 	struct xfs_inode	*ip,
1435 	xfs_fileoff_t		fbno,
1436 	xfs_filblks_t		end,
1437 	xfs_off_t		isize)
1438 {
1439 	struct xfs_mount	*mp = ip->i_mount;
1440 	xfs_agnumber_t		agno;
1441 	xfs_agblock_t		agbno;
1442 	xfs_extlen_t		aglen;
1443 	xfs_agblock_t		rbno;
1444 	xfs_extlen_t		rlen;
1445 	xfs_off_t		fpos;
1446 	xfs_off_t		flen;
1447 	struct xfs_bmbt_irec	map[2];
1448 	int			nmaps;
1449 	int			error = 0;
1450 
1451 	while (end - fbno > 0) {
1452 		nmaps = 1;
1453 		/*
1454 		 * Look for extents in the file.  Skip holes, delalloc, or
1455 		 * unwritten extents; they can't be reflinked.
1456 		 */
1457 		error = xfs_bmapi_read(ip, fbno, end - fbno, map, &nmaps, 0);
1458 		if (error)
1459 			goto out;
1460 		if (nmaps == 0)
1461 			break;
1462 		if (!xfs_bmap_is_real_extent(&map[0]))
1463 			goto next;
1464 
1465 		map[1] = map[0];
1466 		while (map[1].br_blockcount) {
1467 			agno = XFS_FSB_TO_AGNO(mp, map[1].br_startblock);
1468 			agbno = XFS_FSB_TO_AGBNO(mp, map[1].br_startblock);
1469 			aglen = map[1].br_blockcount;
1470 
1471 			error = xfs_reflink_find_shared(mp, NULL, agno, agbno,
1472 					aglen, &rbno, &rlen, true);
1473 			if (error)
1474 				goto out;
1475 			if (rbno == NULLAGBLOCK)
1476 				break;
1477 
1478 			/* Dirty the pages */
1479 			xfs_iunlock(ip, XFS_ILOCK_EXCL);
1480 			fpos = XFS_FSB_TO_B(mp, map[1].br_startoff +
1481 					(rbno - agbno));
1482 			flen = XFS_FSB_TO_B(mp, rlen);
1483 			if (fpos + flen > isize)
1484 				flen = isize - fpos;
1485 			error = iomap_file_dirty(VFS_I(ip), fpos, flen,
1486 					&xfs_iomap_ops);
1487 			xfs_ilock(ip, XFS_ILOCK_EXCL);
1488 			if (error)
1489 				goto out;
1490 
1491 			map[1].br_blockcount -= (rbno - agbno + rlen);
1492 			map[1].br_startoff += (rbno - agbno + rlen);
1493 			map[1].br_startblock += (rbno - agbno + rlen);
1494 		}
1495 
1496 next:
1497 		fbno = map[0].br_startoff + map[0].br_blockcount;
1498 	}
1499 out:
1500 	return error;
1501 }
1502 
1503 /* Does this inode need the reflink flag? */
1504 int
1505 xfs_reflink_inode_has_shared_extents(
1506 	struct xfs_trans		*tp,
1507 	struct xfs_inode		*ip,
1508 	bool				*has_shared)
1509 {
1510 	struct xfs_bmbt_irec		got;
1511 	struct xfs_mount		*mp = ip->i_mount;
1512 	struct xfs_ifork		*ifp;
1513 	xfs_agnumber_t			agno;
1514 	xfs_agblock_t			agbno;
1515 	xfs_extlen_t			aglen;
1516 	xfs_agblock_t			rbno;
1517 	xfs_extlen_t			rlen;
1518 	struct xfs_iext_cursor		icur;
1519 	bool				found;
1520 	int				error;
1521 
1522 	ifp = XFS_IFORK_PTR(ip, XFS_DATA_FORK);
1523 	if (!(ifp->if_flags & XFS_IFEXTENTS)) {
1524 		error = xfs_iread_extents(tp, ip, XFS_DATA_FORK);
1525 		if (error)
1526 			return error;
1527 	}
1528 
1529 	*has_shared = false;
1530 	found = xfs_iext_lookup_extent(ip, ifp, 0, &icur, &got);
1531 	while (found) {
1532 		if (isnullstartblock(got.br_startblock) ||
1533 		    got.br_state != XFS_EXT_NORM)
1534 			goto next;
1535 		agno = XFS_FSB_TO_AGNO(mp, got.br_startblock);
1536 		agbno = XFS_FSB_TO_AGBNO(mp, got.br_startblock);
1537 		aglen = got.br_blockcount;
1538 
1539 		error = xfs_reflink_find_shared(mp, tp, agno, agbno, aglen,
1540 				&rbno, &rlen, false);
1541 		if (error)
1542 			return error;
1543 		/* Is there still a shared block here? */
1544 		if (rbno != NULLAGBLOCK) {
1545 			*has_shared = true;
1546 			return 0;
1547 		}
1548 next:
1549 		found = xfs_iext_next_extent(ifp, &icur, &got);
1550 	}
1551 
1552 	return 0;
1553 }
1554 
1555 /* Clear the inode reflink flag if there are no shared extents. */
1556 int
1557 xfs_reflink_clear_inode_flag(
1558 	struct xfs_inode	*ip,
1559 	struct xfs_trans	**tpp)
1560 {
1561 	bool			needs_flag;
1562 	int			error = 0;
1563 
1564 	ASSERT(xfs_is_reflink_inode(ip));
1565 
1566 	error = xfs_reflink_inode_has_shared_extents(*tpp, ip, &needs_flag);
1567 	if (error || needs_flag)
1568 		return error;
1569 
1570 	/*
1571 	 * We didn't find any shared blocks so turn off the reflink flag.
1572 	 * First, get rid of any leftover CoW mappings.
1573 	 */
1574 	error = xfs_reflink_cancel_cow_blocks(ip, tpp, 0, NULLFILEOFF, true);
1575 	if (error)
1576 		return error;
1577 
1578 	/* Clear the inode flag. */
1579 	trace_xfs_reflink_unset_inode_flag(ip);
1580 	ip->i_d.di_flags2 &= ~XFS_DIFLAG2_REFLINK;
1581 	xfs_inode_clear_cowblocks_tag(ip);
1582 	xfs_trans_ijoin(*tpp, ip, 0);
1583 	xfs_trans_log_inode(*tpp, ip, XFS_ILOG_CORE);
1584 
1585 	return error;
1586 }
1587 
1588 /*
1589  * Clear the inode reflink flag if there are no shared extents and the size
1590  * hasn't changed.
1591  */
1592 STATIC int
1593 xfs_reflink_try_clear_inode_flag(
1594 	struct xfs_inode	*ip)
1595 {
1596 	struct xfs_mount	*mp = ip->i_mount;
1597 	struct xfs_trans	*tp;
1598 	int			error = 0;
1599 
1600 	/* Start a rolling transaction to remove the mappings */
1601 	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, 0, 0, 0, &tp);
1602 	if (error)
1603 		return error;
1604 
1605 	xfs_ilock(ip, XFS_ILOCK_EXCL);
1606 	xfs_trans_ijoin(tp, ip, 0);
1607 
1608 	error = xfs_reflink_clear_inode_flag(ip, &tp);
1609 	if (error)
1610 		goto cancel;
1611 
1612 	error = xfs_trans_commit(tp);
1613 	if (error)
1614 		goto out;
1615 
1616 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1617 	return 0;
1618 cancel:
1619 	xfs_trans_cancel(tp);
1620 out:
1621 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1622 	return error;
1623 }
1624 
1625 /*
1626  * Pre-COW all shared blocks within a given byte range of a file and turn off
1627  * the reflink flag if we unshare all of the file's blocks.
1628  */
1629 int
1630 xfs_reflink_unshare(
1631 	struct xfs_inode	*ip,
1632 	xfs_off_t		offset,
1633 	xfs_off_t		len)
1634 {
1635 	struct xfs_mount	*mp = ip->i_mount;
1636 	xfs_fileoff_t		fbno;
1637 	xfs_filblks_t		end;
1638 	xfs_off_t		isize;
1639 	int			error;
1640 
1641 	if (!xfs_is_reflink_inode(ip))
1642 		return 0;
1643 
1644 	trace_xfs_reflink_unshare(ip, offset, len);
1645 
1646 	inode_dio_wait(VFS_I(ip));
1647 
1648 	/* Try to CoW the selected ranges */
1649 	xfs_ilock(ip, XFS_ILOCK_EXCL);
1650 	fbno = XFS_B_TO_FSBT(mp, offset);
1651 	isize = i_size_read(VFS_I(ip));
1652 	end = XFS_B_TO_FSB(mp, offset + len);
1653 	error = xfs_reflink_dirty_extents(ip, fbno, end, isize);
1654 	if (error)
1655 		goto out_unlock;
1656 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1657 
1658 	/* Wait for the IO to finish */
1659 	error = filemap_write_and_wait(VFS_I(ip)->i_mapping);
1660 	if (error)
1661 		goto out;
1662 
1663 	/* Turn off the reflink flag if possible. */
1664 	error = xfs_reflink_try_clear_inode_flag(ip);
1665 	if (error)
1666 		goto out;
1667 
1668 	return 0;
1669 
1670 out_unlock:
1671 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1672 out:
1673 	trace_xfs_reflink_unshare_error(ip, error, _RET_IP_);
1674 	return error;
1675 }
1676