xref: /openbmc/linux/fs/xfs/xfs_reflink.c (revision 4e1a33b1)
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 	xfs_agnumber_t		agno,
159 	xfs_agblock_t		agbno,
160 	xfs_extlen_t		aglen,
161 	xfs_agblock_t		*fbno,
162 	xfs_extlen_t		*flen,
163 	bool			find_end_of_shared)
164 {
165 	struct xfs_buf		*agbp;
166 	struct xfs_btree_cur	*cur;
167 	int			error;
168 
169 	error = xfs_alloc_read_agf(mp, NULL, agno, 0, &agbp);
170 	if (error)
171 		return error;
172 
173 	cur = xfs_refcountbt_init_cursor(mp, NULL, agbp, agno, NULL);
174 
175 	error = xfs_refcount_find_shared(cur, agbno, aglen, fbno, flen,
176 			find_end_of_shared);
177 
178 	xfs_btree_del_cursor(cur, error ? XFS_BTREE_ERROR : XFS_BTREE_NOERROR);
179 
180 	xfs_buf_relse(agbp);
181 	return error;
182 }
183 
184 /*
185  * Trim the mapping to the next block where there's a change in the
186  * shared/unshared status.  More specifically, this means that we
187  * find the lowest-numbered extent of shared blocks that coincides with
188  * the given block mapping.  If the shared extent overlaps the start of
189  * the mapping, trim the mapping to the end of the shared extent.  If
190  * the shared region intersects the mapping, trim the mapping to the
191  * start of the shared extent.  If there are no shared regions that
192  * overlap, just return the original extent.
193  */
194 int
195 xfs_reflink_trim_around_shared(
196 	struct xfs_inode	*ip,
197 	struct xfs_bmbt_irec	*irec,
198 	bool			*shared,
199 	bool			*trimmed)
200 {
201 	xfs_agnumber_t		agno;
202 	xfs_agblock_t		agbno;
203 	xfs_extlen_t		aglen;
204 	xfs_agblock_t		fbno;
205 	xfs_extlen_t		flen;
206 	int			error = 0;
207 
208 	/* Holes, unwritten, and delalloc extents cannot be shared */
209 	if (!xfs_is_reflink_inode(ip) ||
210 	    ISUNWRITTEN(irec) ||
211 	    irec->br_startblock == HOLESTARTBLOCK ||
212 	    irec->br_startblock == DELAYSTARTBLOCK ||
213 	    isnullstartblock(irec->br_startblock)) {
214 		*shared = false;
215 		return 0;
216 	}
217 
218 	trace_xfs_reflink_trim_around_shared(ip, irec);
219 
220 	agno = XFS_FSB_TO_AGNO(ip->i_mount, irec->br_startblock);
221 	agbno = XFS_FSB_TO_AGBNO(ip->i_mount, irec->br_startblock);
222 	aglen = irec->br_blockcount;
223 
224 	error = xfs_reflink_find_shared(ip->i_mount, agno, agbno,
225 			aglen, &fbno, &flen, true);
226 	if (error)
227 		return error;
228 
229 	*shared = *trimmed = false;
230 	if (fbno == NULLAGBLOCK) {
231 		/* No shared blocks at all. */
232 		return 0;
233 	} else if (fbno == agbno) {
234 		/*
235 		 * The start of this extent is shared.  Truncate the
236 		 * mapping at the end of the shared region so that a
237 		 * subsequent iteration starts at the start of the
238 		 * unshared region.
239 		 */
240 		irec->br_blockcount = flen;
241 		*shared = true;
242 		if (flen != aglen)
243 			*trimmed = true;
244 		return 0;
245 	} else {
246 		/*
247 		 * There's a shared extent midway through this extent.
248 		 * Truncate the mapping at the start of the shared
249 		 * extent so that a subsequent iteration starts at the
250 		 * start of the shared region.
251 		 */
252 		irec->br_blockcount = fbno - agbno;
253 		*trimmed = true;
254 		return 0;
255 	}
256 }
257 
258 /*
259  * Trim the passed in imap to the next shared/unshared extent boundary, and
260  * if imap->br_startoff points to a shared extent reserve space for it in the
261  * COW fork.  In this case *shared is set to true, else to false.
262  *
263  * Note that imap will always contain the block numbers for the existing blocks
264  * in the data fork, as the upper layers need them for read-modify-write
265  * operations.
266  */
267 int
268 xfs_reflink_reserve_cow(
269 	struct xfs_inode	*ip,
270 	struct xfs_bmbt_irec	*imap,
271 	bool			*shared)
272 {
273 	struct xfs_ifork	*ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
274 	struct xfs_bmbt_irec	got;
275 	int			error = 0;
276 	bool			eof = false, trimmed;
277 	xfs_extnum_t		idx;
278 
279 	/*
280 	 * Search the COW fork extent list first.  This serves two purposes:
281 	 * first this implement the speculative preallocation using cowextisze,
282 	 * so that we also unshared block adjacent to shared blocks instead
283 	 * of just the shared blocks themselves.  Second the lookup in the
284 	 * extent list is generally faster than going out to the shared extent
285 	 * tree.
286 	 */
287 
288 	if (!xfs_iext_lookup_extent(ip, ifp, imap->br_startoff, &idx, &got))
289 		eof = true;
290 	if (!eof && got.br_startoff <= imap->br_startoff) {
291 		trace_xfs_reflink_cow_found(ip, imap);
292 		xfs_trim_extent(imap, got.br_startoff, got.br_blockcount);
293 
294 		*shared = true;
295 		return 0;
296 	}
297 
298 	/* Trim the mapping to the nearest shared extent boundary. */
299 	error = xfs_reflink_trim_around_shared(ip, imap, shared, &trimmed);
300 	if (error)
301 		return error;
302 
303 	/* Not shared?  Just report the (potentially capped) extent. */
304 	if (!*shared)
305 		return 0;
306 
307 	/*
308 	 * Fork all the shared blocks from our write offset until the end of
309 	 * the extent.
310 	 */
311 	error = xfs_qm_dqattach_locked(ip, 0);
312 	if (error)
313 		return error;
314 
315 	error = xfs_bmapi_reserve_delalloc(ip, XFS_COW_FORK, imap->br_startoff,
316 			imap->br_blockcount, 0, &got, &idx, eof);
317 	if (error == -ENOSPC || error == -EDQUOT)
318 		trace_xfs_reflink_cow_enospc(ip, imap);
319 	if (error)
320 		return error;
321 
322 	trace_xfs_reflink_cow_alloc(ip, &got);
323 	return 0;
324 }
325 
326 /* Convert part of an unwritten CoW extent to a real one. */
327 STATIC int
328 xfs_reflink_convert_cow_extent(
329 	struct xfs_inode		*ip,
330 	struct xfs_bmbt_irec		*imap,
331 	xfs_fileoff_t			offset_fsb,
332 	xfs_filblks_t			count_fsb,
333 	struct xfs_defer_ops		*dfops)
334 {
335 	xfs_fsblock_t			first_block;
336 	int				nimaps = 1;
337 
338 	if (imap->br_state == XFS_EXT_NORM)
339 		return 0;
340 
341 	xfs_trim_extent(imap, offset_fsb, count_fsb);
342 	trace_xfs_reflink_convert_cow(ip, imap);
343 	if (imap->br_blockcount == 0)
344 		return 0;
345 	return xfs_bmapi_write(NULL, ip, imap->br_startoff, imap->br_blockcount,
346 			XFS_BMAPI_COWFORK | XFS_BMAPI_CONVERT, &first_block,
347 			0, imap, &nimaps, dfops);
348 }
349 
350 /* Convert all of the unwritten CoW extents in a file's range to real ones. */
351 int
352 xfs_reflink_convert_cow(
353 	struct xfs_inode	*ip,
354 	xfs_off_t		offset,
355 	xfs_off_t		count)
356 {
357 	struct xfs_bmbt_irec	got;
358 	struct xfs_defer_ops	dfops;
359 	struct xfs_mount	*mp = ip->i_mount;
360 	struct xfs_ifork	*ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
361 	xfs_fileoff_t		offset_fsb = XFS_B_TO_FSBT(mp, offset);
362 	xfs_fileoff_t		end_fsb = XFS_B_TO_FSB(mp, offset + count);
363 	xfs_extnum_t		idx;
364 	bool			found;
365 	int			error = 0;
366 
367 	xfs_ilock(ip, XFS_ILOCK_EXCL);
368 
369 	/* Convert all the extents to real from unwritten. */
370 	for (found = xfs_iext_lookup_extent(ip, ifp, offset_fsb, &idx, &got);
371 	     found && got.br_startoff < end_fsb;
372 	     found = xfs_iext_get_extent(ifp, ++idx, &got)) {
373 		error = xfs_reflink_convert_cow_extent(ip, &got, offset_fsb,
374 				end_fsb - offset_fsb, &dfops);
375 		if (error)
376 			break;
377 	}
378 
379 	/* Finish up. */
380 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
381 	return error;
382 }
383 
384 /* Allocate all CoW reservations covering a range of blocks in a file. */
385 int
386 xfs_reflink_allocate_cow(
387 	struct xfs_inode	*ip,
388 	struct xfs_bmbt_irec	*imap,
389 	bool			*shared,
390 	uint			*lockmode)
391 {
392 	struct xfs_mount	*mp = ip->i_mount;
393 	xfs_fileoff_t		offset_fsb = imap->br_startoff;
394 	xfs_filblks_t		count_fsb = imap->br_blockcount;
395 	struct xfs_bmbt_irec	got;
396 	struct xfs_defer_ops	dfops;
397 	struct xfs_trans	*tp = NULL;
398 	xfs_fsblock_t		first_block;
399 	int			nimaps, error = 0;
400 	bool			trimmed;
401 	xfs_filblks_t		resaligned;
402 	xfs_extlen_t		resblks = 0;
403 	xfs_extnum_t		idx;
404 
405 retry:
406 	ASSERT(xfs_is_reflink_inode(ip));
407 	ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL | XFS_ILOCK_SHARED));
408 
409 	/*
410 	 * Even if the extent is not shared we might have a preallocation for
411 	 * it in the COW fork.  If so use it.
412 	 */
413 	if (xfs_iext_lookup_extent(ip, ip->i_cowfp, offset_fsb, &idx, &got) &&
414 	    got.br_startoff <= offset_fsb) {
415 		*shared = true;
416 
417 		/* If we have a real allocation in the COW fork we're done. */
418 		if (!isnullstartblock(got.br_startblock)) {
419 			xfs_trim_extent(&got, offset_fsb, count_fsb);
420 			*imap = got;
421 			goto convert;
422 		}
423 
424 		xfs_trim_extent(imap, got.br_startoff, got.br_blockcount);
425 	} else {
426 		error = xfs_reflink_trim_around_shared(ip, imap, shared, &trimmed);
427 		if (error || !*shared)
428 			goto out;
429 	}
430 
431 	if (!tp) {
432 		resaligned = xfs_aligned_fsb_count(imap->br_startoff,
433 			imap->br_blockcount, xfs_get_cowextsz_hint(ip));
434 		resblks = XFS_DIOSTRAT_SPACE_RES(mp, resaligned);
435 
436 		xfs_iunlock(ip, *lockmode);
437 		error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, resblks, 0, 0, &tp);
438 		*lockmode = XFS_ILOCK_EXCL;
439 		xfs_ilock(ip, *lockmode);
440 
441 		if (error)
442 			return error;
443 
444 		error = xfs_qm_dqattach_locked(ip, 0);
445 		if (error)
446 			goto out;
447 		goto retry;
448 	}
449 
450 	error = xfs_trans_reserve_quota_nblks(tp, ip, resblks, 0,
451 			XFS_QMOPT_RES_REGBLKS);
452 	if (error)
453 		goto out;
454 
455 	xfs_trans_ijoin(tp, ip, 0);
456 
457 	xfs_defer_init(&dfops, &first_block);
458 	nimaps = 1;
459 
460 	/* Allocate the entire reservation as unwritten blocks. */
461 	error = xfs_bmapi_write(tp, ip, imap->br_startoff, imap->br_blockcount,
462 			XFS_BMAPI_COWFORK | XFS_BMAPI_PREALLOC, &first_block,
463 			resblks, imap, &nimaps, &dfops);
464 	if (error)
465 		goto out_bmap_cancel;
466 
467 	/* Finish up. */
468 	error = xfs_defer_finish(&tp, &dfops, NULL);
469 	if (error)
470 		goto out_bmap_cancel;
471 
472 	error = xfs_trans_commit(tp);
473 	if (error)
474 		return error;
475 convert:
476 	return xfs_reflink_convert_cow_extent(ip, imap, offset_fsb, count_fsb,
477 			&dfops);
478 out_bmap_cancel:
479 	xfs_defer_cancel(&dfops);
480 	xfs_trans_unreserve_quota_nblks(tp, ip, (long)resblks, 0,
481 			XFS_QMOPT_RES_REGBLKS);
482 out:
483 	if (tp)
484 		xfs_trans_cancel(tp);
485 	return error;
486 }
487 
488 /*
489  * Find the CoW reservation for a given byte offset of a file.
490  */
491 bool
492 xfs_reflink_find_cow_mapping(
493 	struct xfs_inode		*ip,
494 	xfs_off_t			offset,
495 	struct xfs_bmbt_irec		*imap)
496 {
497 	struct xfs_ifork		*ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
498 	xfs_fileoff_t			offset_fsb;
499 	struct xfs_bmbt_irec		got;
500 	xfs_extnum_t			idx;
501 
502 	ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL | XFS_ILOCK_SHARED));
503 	ASSERT(xfs_is_reflink_inode(ip));
504 
505 	offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
506 	if (!xfs_iext_lookup_extent(ip, ifp, offset_fsb, &idx, &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 	xfs_extnum_t			idx;
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, &idx, &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_get_extent(ifp, idx + 1, &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 all pending CoW reservations for some block range of an inode.
552  */
553 int
554 xfs_reflink_cancel_cow_blocks(
555 	struct xfs_inode		*ip,
556 	struct xfs_trans		**tpp,
557 	xfs_fileoff_t			offset_fsb,
558 	xfs_fileoff_t			end_fsb)
559 {
560 	struct xfs_ifork		*ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
561 	struct xfs_bmbt_irec		got, del;
562 	xfs_extnum_t			idx;
563 	xfs_fsblock_t			firstfsb;
564 	struct xfs_defer_ops		dfops;
565 	int				error = 0;
566 
567 	if (!xfs_is_reflink_inode(ip))
568 		return 0;
569 	if (!xfs_iext_lookup_extent(ip, ifp, offset_fsb, &idx, &got))
570 		return 0;
571 
572 	while (got.br_startoff < end_fsb) {
573 		del = got;
574 		xfs_trim_extent(&del, offset_fsb, end_fsb - offset_fsb);
575 		trace_xfs_reflink_cancel_cow(ip, &del);
576 
577 		if (isnullstartblock(del.br_startblock)) {
578 			error = xfs_bmap_del_extent_delay(ip, XFS_COW_FORK,
579 					&idx, &got, &del);
580 			if (error)
581 				break;
582 		} else {
583 			xfs_trans_ijoin(*tpp, ip, 0);
584 			xfs_defer_init(&dfops, &firstfsb);
585 
586 			/* Free the CoW orphan record. */
587 			error = xfs_refcount_free_cow_extent(ip->i_mount,
588 					&dfops, del.br_startblock,
589 					del.br_blockcount);
590 			if (error)
591 				break;
592 
593 			xfs_bmap_add_free(ip->i_mount, &dfops,
594 					del.br_startblock, del.br_blockcount,
595 					NULL);
596 
597 			/* Update quota accounting */
598 			xfs_trans_mod_dquot_byino(*tpp, ip, XFS_TRANS_DQ_BCOUNT,
599 					-(long)del.br_blockcount);
600 
601 			/* Roll the transaction */
602 			error = xfs_defer_finish(tpp, &dfops, ip);
603 			if (error) {
604 				xfs_defer_cancel(&dfops);
605 				break;
606 			}
607 
608 			/* Remove the mapping from the CoW fork. */
609 			xfs_bmap_del_extent_cow(ip, &idx, &got, &del);
610 		}
611 
612 		if (!xfs_iext_get_extent(ifp, ++idx, &got))
613 			break;
614 	}
615 
616 	/* clear tag if cow fork is emptied */
617 	if (!ifp->if_bytes)
618 		xfs_inode_clear_cowblocks_tag(ip);
619 
620 	return error;
621 }
622 
623 /*
624  * Cancel all pending CoW reservations for some byte range of an inode.
625  */
626 int
627 xfs_reflink_cancel_cow_range(
628 	struct xfs_inode	*ip,
629 	xfs_off_t		offset,
630 	xfs_off_t		count)
631 {
632 	struct xfs_trans	*tp;
633 	xfs_fileoff_t		offset_fsb;
634 	xfs_fileoff_t		end_fsb;
635 	int			error;
636 
637 	trace_xfs_reflink_cancel_cow_range(ip, offset, count);
638 	ASSERT(xfs_is_reflink_inode(ip));
639 
640 	offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
641 	if (count == NULLFILEOFF)
642 		end_fsb = NULLFILEOFF;
643 	else
644 		end_fsb = XFS_B_TO_FSB(ip->i_mount, offset + count);
645 
646 	/* Start a rolling transaction to remove the mappings */
647 	error = xfs_trans_alloc(ip->i_mount, &M_RES(ip->i_mount)->tr_write,
648 			0, 0, 0, &tp);
649 	if (error)
650 		goto out;
651 
652 	xfs_ilock(ip, XFS_ILOCK_EXCL);
653 	xfs_trans_ijoin(tp, ip, 0);
654 
655 	/* Scrape out the old CoW reservations */
656 	error = xfs_reflink_cancel_cow_blocks(ip, &tp, offset_fsb, end_fsb);
657 	if (error)
658 		goto out_cancel;
659 
660 	error = xfs_trans_commit(tp);
661 
662 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
663 	return error;
664 
665 out_cancel:
666 	xfs_trans_cancel(tp);
667 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
668 out:
669 	trace_xfs_reflink_cancel_cow_range_error(ip, error, _RET_IP_);
670 	return error;
671 }
672 
673 /*
674  * Remap parts of a file's data fork after a successful CoW.
675  */
676 int
677 xfs_reflink_end_cow(
678 	struct xfs_inode		*ip,
679 	xfs_off_t			offset,
680 	xfs_off_t			count)
681 {
682 	struct xfs_ifork		*ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
683 	struct xfs_bmbt_irec		got, del;
684 	struct xfs_trans		*tp;
685 	xfs_fileoff_t			offset_fsb;
686 	xfs_fileoff_t			end_fsb;
687 	xfs_fsblock_t			firstfsb;
688 	struct xfs_defer_ops		dfops;
689 	int				error;
690 	unsigned int			resblks;
691 	xfs_filblks_t			rlen;
692 	xfs_extnum_t			idx;
693 
694 	trace_xfs_reflink_end_cow(ip, offset, count);
695 
696 	/* No COW extents?  That's easy! */
697 	if (ifp->if_bytes == 0)
698 		return 0;
699 
700 	offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
701 	end_fsb = XFS_B_TO_FSB(ip->i_mount, offset + count);
702 
703 	/* Start a rolling transaction to switch the mappings */
704 	resblks = XFS_EXTENTADD_SPACE_RES(ip->i_mount, XFS_DATA_FORK);
705 	error = xfs_trans_alloc(ip->i_mount, &M_RES(ip->i_mount)->tr_write,
706 			resblks, 0, 0, &tp);
707 	if (error)
708 		goto out;
709 
710 	xfs_ilock(ip, XFS_ILOCK_EXCL);
711 	xfs_trans_ijoin(tp, ip, 0);
712 
713 	/* If there is a hole at end_fsb - 1 go to the previous extent */
714 	if (!xfs_iext_lookup_extent(ip, ifp, end_fsb - 1, &idx, &got) ||
715 	    got.br_startoff > end_fsb) {
716 		ASSERT(idx > 0);
717 		xfs_iext_get_extent(ifp, --idx, &got);
718 	}
719 
720 	/* Walk backwards until we're out of the I/O range... */
721 	while (got.br_startoff + got.br_blockcount > offset_fsb) {
722 		del = got;
723 		xfs_trim_extent(&del, offset_fsb, end_fsb - offset_fsb);
724 
725 		/* Extent delete may have bumped idx forward */
726 		if (!del.br_blockcount) {
727 			idx--;
728 			goto next_extent;
729 		}
730 
731 		ASSERT(!isnullstartblock(got.br_startblock));
732 
733 		/*
734 		 * Don't remap unwritten extents; these are
735 		 * speculatively preallocated CoW extents that have been
736 		 * allocated but have not yet been involved in a write.
737 		 */
738 		if (got.br_state == XFS_EXT_UNWRITTEN) {
739 			idx--;
740 			goto next_extent;
741 		}
742 
743 		/* Unmap the old blocks in the data fork. */
744 		xfs_defer_init(&dfops, &firstfsb);
745 		rlen = del.br_blockcount;
746 		error = __xfs_bunmapi(tp, ip, del.br_startoff, &rlen, 0, 1,
747 				&firstfsb, &dfops);
748 		if (error)
749 			goto out_defer;
750 
751 		/* Trim the extent to whatever got unmapped. */
752 		if (rlen) {
753 			xfs_trim_extent(&del, del.br_startoff + rlen,
754 				del.br_blockcount - rlen);
755 		}
756 		trace_xfs_reflink_cow_remap(ip, &del);
757 
758 		/* Free the CoW orphan record. */
759 		error = xfs_refcount_free_cow_extent(tp->t_mountp, &dfops,
760 				del.br_startblock, del.br_blockcount);
761 		if (error)
762 			goto out_defer;
763 
764 		/* Map the new blocks into the data fork. */
765 		error = xfs_bmap_map_extent(tp->t_mountp, &dfops, ip, &del);
766 		if (error)
767 			goto out_defer;
768 
769 		/* Remove the mapping from the CoW fork. */
770 		xfs_bmap_del_extent_cow(ip, &idx, &got, &del);
771 
772 		error = xfs_defer_finish(&tp, &dfops, ip);
773 		if (error)
774 			goto out_defer;
775 next_extent:
776 		if (!xfs_iext_get_extent(ifp, idx, &got))
777 			break;
778 	}
779 
780 	error = xfs_trans_commit(tp);
781 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
782 	if (error)
783 		goto out;
784 	return 0;
785 
786 out_defer:
787 	xfs_defer_cancel(&dfops);
788 	xfs_trans_cancel(tp);
789 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
790 out:
791 	trace_xfs_reflink_end_cow_error(ip, error, _RET_IP_);
792 	return error;
793 }
794 
795 /*
796  * Free leftover CoW reservations that didn't get cleaned out.
797  */
798 int
799 xfs_reflink_recover_cow(
800 	struct xfs_mount	*mp)
801 {
802 	xfs_agnumber_t		agno;
803 	int			error = 0;
804 
805 	if (!xfs_sb_version_hasreflink(&mp->m_sb))
806 		return 0;
807 
808 	for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
809 		error = xfs_refcount_recover_cow_leftovers(mp, agno);
810 		if (error)
811 			break;
812 	}
813 
814 	return error;
815 }
816 
817 /*
818  * Reflinking (Block) Ranges of Two Files Together
819  *
820  * First, ensure that the reflink flag is set on both inodes.  The flag is an
821  * optimization to avoid unnecessary refcount btree lookups in the write path.
822  *
823  * Now we can iteratively remap the range of extents (and holes) in src to the
824  * corresponding ranges in dest.  Let drange and srange denote the ranges of
825  * logical blocks in dest and src touched by the reflink operation.
826  *
827  * While the length of drange is greater than zero,
828  *    - Read src's bmbt at the start of srange ("imap")
829  *    - If imap doesn't exist, make imap appear to start at the end of srange
830  *      with zero length.
831  *    - If imap starts before srange, advance imap to start at srange.
832  *    - If imap goes beyond srange, truncate imap to end at the end of srange.
833  *    - Punch (imap start - srange start + imap len) blocks from dest at
834  *      offset (drange start).
835  *    - If imap points to a real range of pblks,
836  *         > Increase the refcount of the imap's pblks
837  *         > Map imap's pblks into dest at the offset
838  *           (drange start + imap start - srange start)
839  *    - Advance drange and srange by (imap start - srange start + imap len)
840  *
841  * Finally, if the reflink made dest longer, update both the in-core and
842  * on-disk file sizes.
843  *
844  * ASCII Art Demonstration:
845  *
846  * Let's say we want to reflink this source file:
847  *
848  * ----SSSSSSS-SSSSS----SSSSSS (src file)
849  *   <-------------------->
850  *
851  * into this destination file:
852  *
853  * --DDDDDDDDDDDDDDDDDDD--DDD (dest file)
854  *        <-------------------->
855  * '-' means a hole, and 'S' and 'D' are written blocks in the src and dest.
856  * Observe that the range has different logical offsets in either file.
857  *
858  * Consider that the first extent in the source file doesn't line up with our
859  * reflink range.  Unmapping  and remapping are separate operations, so we can
860  * unmap more blocks from the destination file than we remap.
861  *
862  * ----SSSSSSS-SSSSS----SSSSSS
863  *   <------->
864  * --DDDDD---------DDDDD--DDD
865  *        <------->
866  *
867  * Now remap the source extent into the destination file:
868  *
869  * ----SSSSSSS-SSSSS----SSSSSS
870  *   <------->
871  * --DDDDD--SSSSSSSDDDDD--DDD
872  *        <------->
873  *
874  * Do likewise with the second hole and extent in our range.  Holes in the
875  * unmap range don't affect our operation.
876  *
877  * ----SSSSSSS-SSSSS----SSSSSS
878  *            <---->
879  * --DDDDD--SSSSSSS-SSSSS-DDD
880  *                 <---->
881  *
882  * Finally, unmap and remap part of the third extent.  This will increase the
883  * size of the destination file.
884  *
885  * ----SSSSSSS-SSSSS----SSSSSS
886  *                  <----->
887  * --DDDDD--SSSSSSS-SSSSS----SSS
888  *                       <----->
889  *
890  * Once we update the destination file's i_size, we're done.
891  */
892 
893 /*
894  * Ensure the reflink bit is set in both inodes.
895  */
896 STATIC int
897 xfs_reflink_set_inode_flag(
898 	struct xfs_inode	*src,
899 	struct xfs_inode	*dest)
900 {
901 	struct xfs_mount	*mp = src->i_mount;
902 	int			error;
903 	struct xfs_trans	*tp;
904 
905 	if (xfs_is_reflink_inode(src) && xfs_is_reflink_inode(dest))
906 		return 0;
907 
908 	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
909 	if (error)
910 		goto out_error;
911 
912 	/* Lock both files against IO */
913 	if (src->i_ino == dest->i_ino)
914 		xfs_ilock(src, XFS_ILOCK_EXCL);
915 	else
916 		xfs_lock_two_inodes(src, dest, XFS_ILOCK_EXCL);
917 
918 	if (!xfs_is_reflink_inode(src)) {
919 		trace_xfs_reflink_set_inode_flag(src);
920 		xfs_trans_ijoin(tp, src, XFS_ILOCK_EXCL);
921 		src->i_d.di_flags2 |= XFS_DIFLAG2_REFLINK;
922 		xfs_trans_log_inode(tp, src, XFS_ILOG_CORE);
923 		xfs_ifork_init_cow(src);
924 	} else
925 		xfs_iunlock(src, XFS_ILOCK_EXCL);
926 
927 	if (src->i_ino == dest->i_ino)
928 		goto commit_flags;
929 
930 	if (!xfs_is_reflink_inode(dest)) {
931 		trace_xfs_reflink_set_inode_flag(dest);
932 		xfs_trans_ijoin(tp, dest, XFS_ILOCK_EXCL);
933 		dest->i_d.di_flags2 |= XFS_DIFLAG2_REFLINK;
934 		xfs_trans_log_inode(tp, dest, XFS_ILOG_CORE);
935 		xfs_ifork_init_cow(dest);
936 	} else
937 		xfs_iunlock(dest, XFS_ILOCK_EXCL);
938 
939 commit_flags:
940 	error = xfs_trans_commit(tp);
941 	if (error)
942 		goto out_error;
943 	return error;
944 
945 out_error:
946 	trace_xfs_reflink_set_inode_flag_error(dest, error, _RET_IP_);
947 	return error;
948 }
949 
950 /*
951  * Update destination inode size & cowextsize hint, if necessary.
952  */
953 STATIC int
954 xfs_reflink_update_dest(
955 	struct xfs_inode	*dest,
956 	xfs_off_t		newlen,
957 	xfs_extlen_t		cowextsize,
958 	bool			is_dedupe)
959 {
960 	struct xfs_mount	*mp = dest->i_mount;
961 	struct xfs_trans	*tp;
962 	int			error;
963 
964 	if (is_dedupe && newlen <= i_size_read(VFS_I(dest)) && cowextsize == 0)
965 		return 0;
966 
967 	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
968 	if (error)
969 		goto out_error;
970 
971 	xfs_ilock(dest, XFS_ILOCK_EXCL);
972 	xfs_trans_ijoin(tp, dest, XFS_ILOCK_EXCL);
973 
974 	if (newlen > i_size_read(VFS_I(dest))) {
975 		trace_xfs_reflink_update_inode_size(dest, newlen);
976 		i_size_write(VFS_I(dest), newlen);
977 		dest->i_d.di_size = newlen;
978 	}
979 
980 	if (cowextsize) {
981 		dest->i_d.di_cowextsize = cowextsize;
982 		dest->i_d.di_flags2 |= XFS_DIFLAG2_COWEXTSIZE;
983 	}
984 
985 	if (!is_dedupe) {
986 		xfs_trans_ichgtime(tp, dest,
987 				   XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
988 	}
989 	xfs_trans_log_inode(tp, dest, XFS_ILOG_CORE);
990 
991 	error = xfs_trans_commit(tp);
992 	if (error)
993 		goto out_error;
994 	return error;
995 
996 out_error:
997 	trace_xfs_reflink_update_inode_size_error(dest, error, _RET_IP_);
998 	return error;
999 }
1000 
1001 /*
1002  * Do we have enough reserve in this AG to handle a reflink?  The refcount
1003  * btree already reserved all the space it needs, but the rmap btree can grow
1004  * infinitely, so we won't allow more reflinks when the AG is down to the
1005  * btree reserves.
1006  */
1007 static int
1008 xfs_reflink_ag_has_free_space(
1009 	struct xfs_mount	*mp,
1010 	xfs_agnumber_t		agno)
1011 {
1012 	struct xfs_perag	*pag;
1013 	int			error = 0;
1014 
1015 	if (!xfs_sb_version_hasrmapbt(&mp->m_sb))
1016 		return 0;
1017 
1018 	pag = xfs_perag_get(mp, agno);
1019 	if (xfs_ag_resv_critical(pag, XFS_AG_RESV_AGFL) ||
1020 	    xfs_ag_resv_critical(pag, XFS_AG_RESV_METADATA))
1021 		error = -ENOSPC;
1022 	xfs_perag_put(pag);
1023 	return error;
1024 }
1025 
1026 /*
1027  * Unmap a range of blocks from a file, then map other blocks into the hole.
1028  * The range to unmap is (destoff : destoff + srcioff + irec->br_blockcount).
1029  * The extent irec is mapped into dest at irec->br_startoff.
1030  */
1031 STATIC int
1032 xfs_reflink_remap_extent(
1033 	struct xfs_inode	*ip,
1034 	struct xfs_bmbt_irec	*irec,
1035 	xfs_fileoff_t		destoff,
1036 	xfs_off_t		new_isize)
1037 {
1038 	struct xfs_mount	*mp = ip->i_mount;
1039 	struct xfs_trans	*tp;
1040 	xfs_fsblock_t		firstfsb;
1041 	unsigned int		resblks;
1042 	struct xfs_defer_ops	dfops;
1043 	struct xfs_bmbt_irec	uirec;
1044 	bool			real_extent;
1045 	xfs_filblks_t		rlen;
1046 	xfs_filblks_t		unmap_len;
1047 	xfs_off_t		newlen;
1048 	int			error;
1049 
1050 	unmap_len = irec->br_startoff + irec->br_blockcount - destoff;
1051 	trace_xfs_reflink_punch_range(ip, destoff, unmap_len);
1052 
1053 	/* Only remap normal extents. */
1054 	real_extent =  (irec->br_startblock != HOLESTARTBLOCK &&
1055 			irec->br_startblock != DELAYSTARTBLOCK &&
1056 			!ISUNWRITTEN(irec));
1057 
1058 	/* No reflinking if we're low on space */
1059 	if (real_extent) {
1060 		error = xfs_reflink_ag_has_free_space(mp,
1061 				XFS_FSB_TO_AGNO(mp, irec->br_startblock));
1062 		if (error)
1063 			goto out;
1064 	}
1065 
1066 	/* Start a rolling transaction to switch the mappings */
1067 	resblks = XFS_EXTENTADD_SPACE_RES(ip->i_mount, XFS_DATA_FORK);
1068 	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, resblks, 0, 0, &tp);
1069 	if (error)
1070 		goto out;
1071 
1072 	xfs_ilock(ip, XFS_ILOCK_EXCL);
1073 	xfs_trans_ijoin(tp, ip, 0);
1074 
1075 	/* If we're not just clearing space, then do we have enough quota? */
1076 	if (real_extent) {
1077 		error = xfs_trans_reserve_quota_nblks(tp, ip,
1078 				irec->br_blockcount, 0, XFS_QMOPT_RES_REGBLKS);
1079 		if (error)
1080 			goto out_cancel;
1081 	}
1082 
1083 	trace_xfs_reflink_remap(ip, irec->br_startoff,
1084 				irec->br_blockcount, irec->br_startblock);
1085 
1086 	/* Unmap the old blocks in the data fork. */
1087 	rlen = unmap_len;
1088 	while (rlen) {
1089 		xfs_defer_init(&dfops, &firstfsb);
1090 		error = __xfs_bunmapi(tp, ip, destoff, &rlen, 0, 1,
1091 				&firstfsb, &dfops);
1092 		if (error)
1093 			goto out_defer;
1094 
1095 		/*
1096 		 * Trim the extent to whatever got unmapped.
1097 		 * Remember, bunmapi works backwards.
1098 		 */
1099 		uirec.br_startblock = irec->br_startblock + rlen;
1100 		uirec.br_startoff = irec->br_startoff + rlen;
1101 		uirec.br_blockcount = unmap_len - rlen;
1102 		unmap_len = rlen;
1103 
1104 		/* If this isn't a real mapping, we're done. */
1105 		if (!real_extent || uirec.br_blockcount == 0)
1106 			goto next_extent;
1107 
1108 		trace_xfs_reflink_remap(ip, uirec.br_startoff,
1109 				uirec.br_blockcount, uirec.br_startblock);
1110 
1111 		/* Update the refcount tree */
1112 		error = xfs_refcount_increase_extent(mp, &dfops, &uirec);
1113 		if (error)
1114 			goto out_defer;
1115 
1116 		/* Map the new blocks into the data fork. */
1117 		error = xfs_bmap_map_extent(mp, &dfops, ip, &uirec);
1118 		if (error)
1119 			goto out_defer;
1120 
1121 		/* Update quota accounting. */
1122 		xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_BCOUNT,
1123 				uirec.br_blockcount);
1124 
1125 		/* Update dest isize if needed. */
1126 		newlen = XFS_FSB_TO_B(mp,
1127 				uirec.br_startoff + uirec.br_blockcount);
1128 		newlen = min_t(xfs_off_t, newlen, new_isize);
1129 		if (newlen > i_size_read(VFS_I(ip))) {
1130 			trace_xfs_reflink_update_inode_size(ip, newlen);
1131 			i_size_write(VFS_I(ip), newlen);
1132 			ip->i_d.di_size = newlen;
1133 			xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1134 		}
1135 
1136 next_extent:
1137 		/* Process all the deferred stuff. */
1138 		error = xfs_defer_finish(&tp, &dfops, ip);
1139 		if (error)
1140 			goto out_defer;
1141 	}
1142 
1143 	error = xfs_trans_commit(tp);
1144 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1145 	if (error)
1146 		goto out;
1147 	return 0;
1148 
1149 out_defer:
1150 	xfs_defer_cancel(&dfops);
1151 out_cancel:
1152 	xfs_trans_cancel(tp);
1153 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1154 out:
1155 	trace_xfs_reflink_remap_extent_error(ip, error, _RET_IP_);
1156 	return error;
1157 }
1158 
1159 /*
1160  * Iteratively remap one file's extents (and holes) to another's.
1161  */
1162 STATIC int
1163 xfs_reflink_remap_blocks(
1164 	struct xfs_inode	*src,
1165 	xfs_fileoff_t		srcoff,
1166 	struct xfs_inode	*dest,
1167 	xfs_fileoff_t		destoff,
1168 	xfs_filblks_t		len,
1169 	xfs_off_t		new_isize)
1170 {
1171 	struct xfs_bmbt_irec	imap;
1172 	int			nimaps;
1173 	int			error = 0;
1174 	xfs_filblks_t		range_len;
1175 
1176 	/* drange = (destoff, destoff + len); srange = (srcoff, srcoff + len) */
1177 	while (len) {
1178 		trace_xfs_reflink_remap_blocks_loop(src, srcoff, len,
1179 				dest, destoff);
1180 		/* Read extent from the source file */
1181 		nimaps = 1;
1182 		xfs_ilock(src, XFS_ILOCK_EXCL);
1183 		error = xfs_bmapi_read(src, srcoff, len, &imap, &nimaps, 0);
1184 		xfs_iunlock(src, XFS_ILOCK_EXCL);
1185 		if (error)
1186 			goto err;
1187 		ASSERT(nimaps == 1);
1188 
1189 		trace_xfs_reflink_remap_imap(src, srcoff, len, XFS_IO_OVERWRITE,
1190 				&imap);
1191 
1192 		/* Translate imap into the destination file. */
1193 		range_len = imap.br_startoff + imap.br_blockcount - srcoff;
1194 		imap.br_startoff += destoff - srcoff;
1195 
1196 		/* Clear dest from destoff to the end of imap and map it in. */
1197 		error = xfs_reflink_remap_extent(dest, &imap, destoff,
1198 				new_isize);
1199 		if (error)
1200 			goto err;
1201 
1202 		if (fatal_signal_pending(current)) {
1203 			error = -EINTR;
1204 			goto err;
1205 		}
1206 
1207 		/* Advance drange/srange */
1208 		srcoff += range_len;
1209 		destoff += range_len;
1210 		len -= range_len;
1211 	}
1212 
1213 	return 0;
1214 
1215 err:
1216 	trace_xfs_reflink_remap_blocks_error(dest, error, _RET_IP_);
1217 	return error;
1218 }
1219 
1220 /*
1221  * Link a range of blocks from one file to another.
1222  */
1223 int
1224 xfs_reflink_remap_range(
1225 	struct file		*file_in,
1226 	loff_t			pos_in,
1227 	struct file		*file_out,
1228 	loff_t			pos_out,
1229 	u64			len,
1230 	bool			is_dedupe)
1231 {
1232 	struct inode		*inode_in = file_inode(file_in);
1233 	struct xfs_inode	*src = XFS_I(inode_in);
1234 	struct inode		*inode_out = file_inode(file_out);
1235 	struct xfs_inode	*dest = XFS_I(inode_out);
1236 	struct xfs_mount	*mp = src->i_mount;
1237 	bool			same_inode = (inode_in == inode_out);
1238 	xfs_fileoff_t		sfsbno, dfsbno;
1239 	xfs_filblks_t		fsblen;
1240 	xfs_extlen_t		cowextsize;
1241 	ssize_t			ret;
1242 
1243 	if (!xfs_sb_version_hasreflink(&mp->m_sb))
1244 		return -EOPNOTSUPP;
1245 
1246 	if (XFS_FORCED_SHUTDOWN(mp))
1247 		return -EIO;
1248 
1249 	/* Lock both files against IO */
1250 	lock_two_nondirectories(inode_in, inode_out);
1251 	if (same_inode)
1252 		xfs_ilock(src, XFS_MMAPLOCK_EXCL);
1253 	else
1254 		xfs_lock_two_inodes(src, dest, XFS_MMAPLOCK_EXCL);
1255 
1256 	/* Check file eligibility and prepare for block sharing. */
1257 	ret = -EINVAL;
1258 	/* Don't reflink realtime inodes */
1259 	if (XFS_IS_REALTIME_INODE(src) || XFS_IS_REALTIME_INODE(dest))
1260 		goto out_unlock;
1261 
1262 	/* Don't share DAX file data for now. */
1263 	if (IS_DAX(inode_in) || IS_DAX(inode_out))
1264 		goto out_unlock;
1265 
1266 	ret = vfs_clone_file_prep_inodes(inode_in, pos_in, inode_out, pos_out,
1267 			&len, is_dedupe);
1268 	if (ret <= 0)
1269 		goto out_unlock;
1270 
1271 	trace_xfs_reflink_remap_range(src, pos_in, len, dest, pos_out);
1272 
1273 	/* Set flags and remap blocks. */
1274 	ret = xfs_reflink_set_inode_flag(src, dest);
1275 	if (ret)
1276 		goto out_unlock;
1277 
1278 	dfsbno = XFS_B_TO_FSBT(mp, pos_out);
1279 	sfsbno = XFS_B_TO_FSBT(mp, pos_in);
1280 	fsblen = XFS_B_TO_FSB(mp, len);
1281 	ret = xfs_reflink_remap_blocks(src, sfsbno, dest, dfsbno, fsblen,
1282 			pos_out + len);
1283 	if (ret)
1284 		goto out_unlock;
1285 
1286 	/* Zap any page cache for the destination file's range. */
1287 	truncate_inode_pages_range(&inode_out->i_data, pos_out,
1288 				   PAGE_ALIGN(pos_out + len) - 1);
1289 
1290 	/*
1291 	 * Carry the cowextsize hint from src to dest if we're sharing the
1292 	 * entire source file to the entire destination file, the source file
1293 	 * has a cowextsize hint, and the destination file does not.
1294 	 */
1295 	cowextsize = 0;
1296 	if (pos_in == 0 && len == i_size_read(inode_in) &&
1297 	    (src->i_d.di_flags2 & XFS_DIFLAG2_COWEXTSIZE) &&
1298 	    pos_out == 0 && len >= i_size_read(inode_out) &&
1299 	    !(dest->i_d.di_flags2 & XFS_DIFLAG2_COWEXTSIZE))
1300 		cowextsize = src->i_d.di_cowextsize;
1301 
1302 	ret = xfs_reflink_update_dest(dest, pos_out + len, cowextsize,
1303 			is_dedupe);
1304 
1305 out_unlock:
1306 	xfs_iunlock(src, XFS_MMAPLOCK_EXCL);
1307 	if (!same_inode)
1308 		xfs_iunlock(dest, XFS_MMAPLOCK_EXCL);
1309 	unlock_two_nondirectories(inode_in, inode_out);
1310 	if (ret)
1311 		trace_xfs_reflink_remap_range_error(dest, ret, _RET_IP_);
1312 	return ret;
1313 }
1314 
1315 /*
1316  * The user wants to preemptively CoW all shared blocks in this file,
1317  * which enables us to turn off the reflink flag.  Iterate all
1318  * extents which are not prealloc/delalloc to see which ranges are
1319  * mentioned in the refcount tree, then read those blocks into the
1320  * pagecache, dirty them, fsync them back out, and then we can update
1321  * the inode flag.  What happens if we run out of memory? :)
1322  */
1323 STATIC int
1324 xfs_reflink_dirty_extents(
1325 	struct xfs_inode	*ip,
1326 	xfs_fileoff_t		fbno,
1327 	xfs_filblks_t		end,
1328 	xfs_off_t		isize)
1329 {
1330 	struct xfs_mount	*mp = ip->i_mount;
1331 	xfs_agnumber_t		agno;
1332 	xfs_agblock_t		agbno;
1333 	xfs_extlen_t		aglen;
1334 	xfs_agblock_t		rbno;
1335 	xfs_extlen_t		rlen;
1336 	xfs_off_t		fpos;
1337 	xfs_off_t		flen;
1338 	struct xfs_bmbt_irec	map[2];
1339 	int			nmaps;
1340 	int			error = 0;
1341 
1342 	while (end - fbno > 0) {
1343 		nmaps = 1;
1344 		/*
1345 		 * Look for extents in the file.  Skip holes, delalloc, or
1346 		 * unwritten extents; they can't be reflinked.
1347 		 */
1348 		error = xfs_bmapi_read(ip, fbno, end - fbno, map, &nmaps, 0);
1349 		if (error)
1350 			goto out;
1351 		if (nmaps == 0)
1352 			break;
1353 		if (map[0].br_startblock == HOLESTARTBLOCK ||
1354 		    map[0].br_startblock == DELAYSTARTBLOCK ||
1355 		    ISUNWRITTEN(&map[0]))
1356 			goto next;
1357 
1358 		map[1] = map[0];
1359 		while (map[1].br_blockcount) {
1360 			agno = XFS_FSB_TO_AGNO(mp, map[1].br_startblock);
1361 			agbno = XFS_FSB_TO_AGBNO(mp, map[1].br_startblock);
1362 			aglen = map[1].br_blockcount;
1363 
1364 			error = xfs_reflink_find_shared(mp, agno, agbno, aglen,
1365 					&rbno, &rlen, true);
1366 			if (error)
1367 				goto out;
1368 			if (rbno == NULLAGBLOCK)
1369 				break;
1370 
1371 			/* Dirty the pages */
1372 			xfs_iunlock(ip, XFS_ILOCK_EXCL);
1373 			fpos = XFS_FSB_TO_B(mp, map[1].br_startoff +
1374 					(rbno - agbno));
1375 			flen = XFS_FSB_TO_B(mp, rlen);
1376 			if (fpos + flen > isize)
1377 				flen = isize - fpos;
1378 			error = iomap_file_dirty(VFS_I(ip), fpos, flen,
1379 					&xfs_iomap_ops);
1380 			xfs_ilock(ip, XFS_ILOCK_EXCL);
1381 			if (error)
1382 				goto out;
1383 
1384 			map[1].br_blockcount -= (rbno - agbno + rlen);
1385 			map[1].br_startoff += (rbno - agbno + rlen);
1386 			map[1].br_startblock += (rbno - agbno + rlen);
1387 		}
1388 
1389 next:
1390 		fbno = map[0].br_startoff + map[0].br_blockcount;
1391 	}
1392 out:
1393 	return error;
1394 }
1395 
1396 /* Clear the inode reflink flag if there are no shared extents. */
1397 int
1398 xfs_reflink_clear_inode_flag(
1399 	struct xfs_inode	*ip,
1400 	struct xfs_trans	**tpp)
1401 {
1402 	struct xfs_mount	*mp = ip->i_mount;
1403 	xfs_fileoff_t		fbno;
1404 	xfs_filblks_t		end;
1405 	xfs_agnumber_t		agno;
1406 	xfs_agblock_t		agbno;
1407 	xfs_extlen_t		aglen;
1408 	xfs_agblock_t		rbno;
1409 	xfs_extlen_t		rlen;
1410 	struct xfs_bmbt_irec	map;
1411 	int			nmaps;
1412 	int			error = 0;
1413 
1414 	ASSERT(xfs_is_reflink_inode(ip));
1415 
1416 	fbno = 0;
1417 	end = XFS_B_TO_FSB(mp, i_size_read(VFS_I(ip)));
1418 	while (end - fbno > 0) {
1419 		nmaps = 1;
1420 		/*
1421 		 * Look for extents in the file.  Skip holes, delalloc, or
1422 		 * unwritten extents; they can't be reflinked.
1423 		 */
1424 		error = xfs_bmapi_read(ip, fbno, end - fbno, &map, &nmaps, 0);
1425 		if (error)
1426 			return error;
1427 		if (nmaps == 0)
1428 			break;
1429 		if (map.br_startblock == HOLESTARTBLOCK ||
1430 		    map.br_startblock == DELAYSTARTBLOCK ||
1431 		    ISUNWRITTEN(&map))
1432 			goto next;
1433 
1434 		agno = XFS_FSB_TO_AGNO(mp, map.br_startblock);
1435 		agbno = XFS_FSB_TO_AGBNO(mp, map.br_startblock);
1436 		aglen = map.br_blockcount;
1437 
1438 		error = xfs_reflink_find_shared(mp, agno, agbno, aglen,
1439 				&rbno, &rlen, false);
1440 		if (error)
1441 			return error;
1442 		/* Is there still a shared block here? */
1443 		if (rbno != NULLAGBLOCK)
1444 			return 0;
1445 next:
1446 		fbno = map.br_startoff + map.br_blockcount;
1447 	}
1448 
1449 	/*
1450 	 * We didn't find any shared blocks so turn off the reflink flag.
1451 	 * First, get rid of any leftover CoW mappings.
1452 	 */
1453 	error = xfs_reflink_cancel_cow_blocks(ip, tpp, 0, NULLFILEOFF);
1454 	if (error)
1455 		return error;
1456 
1457 	/* Clear the inode flag. */
1458 	trace_xfs_reflink_unset_inode_flag(ip);
1459 	ip->i_d.di_flags2 &= ~XFS_DIFLAG2_REFLINK;
1460 	xfs_inode_clear_cowblocks_tag(ip);
1461 	xfs_trans_ijoin(*tpp, ip, 0);
1462 	xfs_trans_log_inode(*tpp, ip, XFS_ILOG_CORE);
1463 
1464 	return error;
1465 }
1466 
1467 /*
1468  * Clear the inode reflink flag if there are no shared extents and the size
1469  * hasn't changed.
1470  */
1471 STATIC int
1472 xfs_reflink_try_clear_inode_flag(
1473 	struct xfs_inode	*ip)
1474 {
1475 	struct xfs_mount	*mp = ip->i_mount;
1476 	struct xfs_trans	*tp;
1477 	int			error = 0;
1478 
1479 	/* Start a rolling transaction to remove the mappings */
1480 	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, 0, 0, 0, &tp);
1481 	if (error)
1482 		return error;
1483 
1484 	xfs_ilock(ip, XFS_ILOCK_EXCL);
1485 	xfs_trans_ijoin(tp, ip, 0);
1486 
1487 	error = xfs_reflink_clear_inode_flag(ip, &tp);
1488 	if (error)
1489 		goto cancel;
1490 
1491 	error = xfs_trans_commit(tp);
1492 	if (error)
1493 		goto out;
1494 
1495 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1496 	return 0;
1497 cancel:
1498 	xfs_trans_cancel(tp);
1499 out:
1500 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1501 	return error;
1502 }
1503 
1504 /*
1505  * Pre-COW all shared blocks within a given byte range of a file and turn off
1506  * the reflink flag if we unshare all of the file's blocks.
1507  */
1508 int
1509 xfs_reflink_unshare(
1510 	struct xfs_inode	*ip,
1511 	xfs_off_t		offset,
1512 	xfs_off_t		len)
1513 {
1514 	struct xfs_mount	*mp = ip->i_mount;
1515 	xfs_fileoff_t		fbno;
1516 	xfs_filblks_t		end;
1517 	xfs_off_t		isize;
1518 	int			error;
1519 
1520 	if (!xfs_is_reflink_inode(ip))
1521 		return 0;
1522 
1523 	trace_xfs_reflink_unshare(ip, offset, len);
1524 
1525 	inode_dio_wait(VFS_I(ip));
1526 
1527 	/* Try to CoW the selected ranges */
1528 	xfs_ilock(ip, XFS_ILOCK_EXCL);
1529 	fbno = XFS_B_TO_FSBT(mp, offset);
1530 	isize = i_size_read(VFS_I(ip));
1531 	end = XFS_B_TO_FSB(mp, offset + len);
1532 	error = xfs_reflink_dirty_extents(ip, fbno, end, isize);
1533 	if (error)
1534 		goto out_unlock;
1535 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1536 
1537 	/* Wait for the IO to finish */
1538 	error = filemap_write_and_wait(VFS_I(ip)->i_mapping);
1539 	if (error)
1540 		goto out;
1541 
1542 	/* Turn off the reflink flag if possible. */
1543 	error = xfs_reflink_try_clear_inode_flag(ip);
1544 	if (error)
1545 		goto out;
1546 
1547 	return 0;
1548 
1549 out_unlock:
1550 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1551 out:
1552 	trace_xfs_reflink_unshare_error(ip, error, _RET_IP_);
1553 	return error;
1554 }
1555