xref: /openbmc/linux/fs/xfs/xfs_reflink.c (revision c16c6655)
1 // SPDX-License-Identifier: GPL-2.0+
2 /*
3  * Copyright (C) 2016 Oracle.  All Rights Reserved.
4  * Author: Darrick J. Wong <darrick.wong@oracle.com>
5  */
6 #include "xfs.h"
7 #include "xfs_fs.h"
8 #include "xfs_shared.h"
9 #include "xfs_format.h"
10 #include "xfs_log_format.h"
11 #include "xfs_trans_resv.h"
12 #include "xfs_mount.h"
13 #include "xfs_defer.h"
14 #include "xfs_inode.h"
15 #include "xfs_trans.h"
16 #include "xfs_bmap.h"
17 #include "xfs_bmap_util.h"
18 #include "xfs_trace.h"
19 #include "xfs_icache.h"
20 #include "xfs_btree.h"
21 #include "xfs_refcount_btree.h"
22 #include "xfs_refcount.h"
23 #include "xfs_bmap_btree.h"
24 #include "xfs_trans_space.h"
25 #include "xfs_bit.h"
26 #include "xfs_alloc.h"
27 #include "xfs_quota.h"
28 #include "xfs_reflink.h"
29 #include "xfs_iomap.h"
30 #include "xfs_sb.h"
31 #include "xfs_ag_resv.h"
32 
33 /*
34  * Copy on Write of Shared Blocks
35  *
36  * XFS must preserve "the usual" file semantics even when two files share
37  * the same physical blocks.  This means that a write to one file must not
38  * alter the blocks in a different file; the way that we'll do that is
39  * through the use of a copy-on-write mechanism.  At a high level, that
40  * means that when we want to write to a shared block, we allocate a new
41  * block, write the data to the new block, and if that succeeds we map the
42  * new block into the file.
43  *
44  * XFS provides a "delayed allocation" mechanism that defers the allocation
45  * of disk blocks to dirty-but-not-yet-mapped file blocks as long as
46  * possible.  This reduces fragmentation by enabling the filesystem to ask
47  * for bigger chunks less often, which is exactly what we want for CoW.
48  *
49  * The delalloc mechanism begins when the kernel wants to make a block
50  * writable (write_begin or page_mkwrite).  If the offset is not mapped, we
51  * create a delalloc mapping, which is a regular in-core extent, but without
52  * a real startblock.  (For delalloc mappings, the startblock encodes both
53  * a flag that this is a delalloc mapping, and a worst-case estimate of how
54  * many blocks might be required to put the mapping into the BMBT.)  delalloc
55  * mappings are a reservation against the free space in the filesystem;
56  * adjacent mappings can also be combined into fewer larger mappings.
57  *
58  * As an optimization, the CoW extent size hint (cowextsz) creates
59  * outsized aligned delalloc reservations in the hope of landing out of
60  * order nearby CoW writes in a single extent on disk, thereby reducing
61  * fragmentation and improving future performance.
62  *
63  * D: --RRRRRRSSSRRRRRRRR--- (data fork)
64  * C: ------DDDDDDD--------- (CoW fork)
65  *
66  * When dirty pages are being written out (typically in writepage), the
67  * delalloc reservations are converted into unwritten mappings by
68  * allocating blocks and replacing the delalloc mapping with real ones.
69  * A delalloc mapping can be replaced by several unwritten ones if the
70  * free space is fragmented.
71  *
72  * D: --RRRRRRSSSRRRRRRRR---
73  * C: ------UUUUUUU---------
74  *
75  * We want to adapt the delalloc mechanism for copy-on-write, since the
76  * write paths are similar.  The first two steps (creating the reservation
77  * and allocating the blocks) are exactly the same as delalloc except that
78  * the mappings must be stored in a separate CoW fork because we do not want
79  * to disturb the mapping in the data fork until we're sure that the write
80  * succeeded.  IO completion in this case is the process of removing the old
81  * mapping from the data fork and moving the new mapping from the CoW fork to
82  * the data fork.  This will be discussed shortly.
83  *
84  * For now, unaligned directio writes will be bounced back to the page cache.
85  * Block-aligned directio writes will use the same mechanism as buffered
86  * writes.
87  *
88  * Just prior to submitting the actual disk write requests, we convert
89  * the extents representing the range of the file actually being written
90  * (as opposed to extra pieces created for the cowextsize hint) to real
91  * extents.  This will become important in the next step:
92  *
93  * D: --RRRRRRSSSRRRRRRRR---
94  * C: ------UUrrUUU---------
95  *
96  * CoW remapping must be done after the data block write completes,
97  * because we don't want to destroy the old data fork map until we're sure
98  * the new block has been written.  Since the new mappings are kept in a
99  * separate fork, we can simply iterate these mappings to find the ones
100  * that cover the file blocks that we just CoW'd.  For each extent, simply
101  * unmap the corresponding range in the data fork, map the new range into
102  * the data fork, and remove the extent from the CoW fork.  Because of
103  * the presence of the cowextsize hint, however, we must be careful
104  * only to remap the blocks that we've actually written out --  we must
105  * never remap delalloc reservations nor CoW staging blocks that have
106  * yet to be written.  This corresponds exactly to the real extents in
107  * the CoW fork:
108  *
109  * D: --RRRRRRrrSRRRRRRRR---
110  * C: ------UU--UUU---------
111  *
112  * Since the remapping operation can be applied to an arbitrary file
113  * range, we record the need for the remap step as a flag in the ioend
114  * instead of declaring a new IO type.  This is required for direct io
115  * because we only have ioend for the whole dio, and we have to be able to
116  * remember the presence of unwritten blocks and CoW blocks with a single
117  * ioend structure.  Better yet, the more ground we can cover with one
118  * ioend, the better.
119  */
120 
121 /*
122  * Given an AG extent, find the lowest-numbered run of shared blocks
123  * within that range and return the range in fbno/flen.  If
124  * find_end_of_shared is true, return the longest contiguous extent of
125  * shared blocks.  If there are no shared extents, fbno and flen will
126  * be set to NULLAGBLOCK and 0, respectively.
127  */
128 int
129 xfs_reflink_find_shared(
130 	struct xfs_mount	*mp,
131 	struct xfs_trans	*tp,
132 	xfs_agnumber_t		agno,
133 	xfs_agblock_t		agbno,
134 	xfs_extlen_t		aglen,
135 	xfs_agblock_t		*fbno,
136 	xfs_extlen_t		*flen,
137 	bool			find_end_of_shared)
138 {
139 	struct xfs_buf		*agbp;
140 	struct xfs_btree_cur	*cur;
141 	int			error;
142 
143 	error = xfs_alloc_read_agf(mp, tp, agno, 0, &agbp);
144 	if (error)
145 		return error;
146 	if (!agbp)
147 		return -ENOMEM;
148 
149 	cur = xfs_refcountbt_init_cursor(mp, tp, agbp, agno);
150 
151 	error = xfs_refcount_find_shared(cur, agbno, aglen, fbno, flen,
152 			find_end_of_shared);
153 
154 	xfs_btree_del_cursor(cur, error);
155 
156 	xfs_trans_brelse(tp, agbp);
157 	return error;
158 }
159 
160 /*
161  * Trim the mapping to the next block where there's a change in the
162  * shared/unshared status.  More specifically, this means that we
163  * find the lowest-numbered extent of shared blocks that coincides with
164  * the given block mapping.  If the shared extent overlaps the start of
165  * the mapping, trim the mapping to the end of the shared extent.  If
166  * the shared region intersects the mapping, trim the mapping to the
167  * start of the shared extent.  If there are no shared regions that
168  * overlap, just return the original extent.
169  */
170 int
171 xfs_reflink_trim_around_shared(
172 	struct xfs_inode	*ip,
173 	struct xfs_bmbt_irec	*irec,
174 	bool			*shared)
175 {
176 	xfs_agnumber_t		agno;
177 	xfs_agblock_t		agbno;
178 	xfs_extlen_t		aglen;
179 	xfs_agblock_t		fbno;
180 	xfs_extlen_t		flen;
181 	int			error = 0;
182 
183 	/* Holes, unwritten, and delalloc extents cannot be shared */
184 	if (!xfs_is_cow_inode(ip) || !xfs_bmap_is_real_extent(irec)) {
185 		*shared = false;
186 		return 0;
187 	}
188 
189 	trace_xfs_reflink_trim_around_shared(ip, irec);
190 
191 	agno = XFS_FSB_TO_AGNO(ip->i_mount, irec->br_startblock);
192 	agbno = XFS_FSB_TO_AGBNO(ip->i_mount, irec->br_startblock);
193 	aglen = irec->br_blockcount;
194 
195 	error = xfs_reflink_find_shared(ip->i_mount, NULL, agno, agbno,
196 			aglen, &fbno, &flen, true);
197 	if (error)
198 		return error;
199 
200 	*shared = false;
201 	if (fbno == NULLAGBLOCK) {
202 		/* No shared blocks at all. */
203 		return 0;
204 	} else if (fbno == agbno) {
205 		/*
206 		 * The start of this extent is shared.  Truncate the
207 		 * mapping at the end of the shared region so that a
208 		 * subsequent iteration starts at the start of the
209 		 * unshared region.
210 		 */
211 		irec->br_blockcount = flen;
212 		*shared = true;
213 		return 0;
214 	} else {
215 		/*
216 		 * There's a shared extent midway through this extent.
217 		 * Truncate the mapping at the start of the shared
218 		 * extent so that a subsequent iteration starts at the
219 		 * start of the shared region.
220 		 */
221 		irec->br_blockcount = fbno - agbno;
222 		return 0;
223 	}
224 }
225 
226 bool
227 xfs_inode_need_cow(
228 	struct xfs_inode	*ip,
229 	struct xfs_bmbt_irec	*imap,
230 	bool			*shared)
231 {
232 	/* We can't update any real extents in always COW mode. */
233 	if (xfs_is_always_cow_inode(ip) &&
234 	    !isnullstartblock(imap->br_startblock)) {
235 		*shared = true;
236 		return 0;
237 	}
238 
239 	/* Trim the mapping to the nearest shared extent boundary. */
240 	return xfs_reflink_trim_around_shared(ip, imap, shared);
241 }
242 
243 static int
244 xfs_reflink_convert_cow_locked(
245 	struct xfs_inode	*ip,
246 	xfs_fileoff_t		offset_fsb,
247 	xfs_filblks_t		count_fsb)
248 {
249 	struct xfs_iext_cursor	icur;
250 	struct xfs_bmbt_irec	got;
251 	struct xfs_btree_cur	*dummy_cur = NULL;
252 	int			dummy_logflags;
253 	int			error = 0;
254 
255 	if (!xfs_iext_lookup_extent(ip, ip->i_cowfp, offset_fsb, &icur, &got))
256 		return 0;
257 
258 	do {
259 		if (got.br_startoff >= offset_fsb + count_fsb)
260 			break;
261 		if (got.br_state == XFS_EXT_NORM)
262 			continue;
263 		if (WARN_ON_ONCE(isnullstartblock(got.br_startblock)))
264 			return -EIO;
265 
266 		xfs_trim_extent(&got, offset_fsb, count_fsb);
267 		if (!got.br_blockcount)
268 			continue;
269 
270 		got.br_state = XFS_EXT_NORM;
271 		error = xfs_bmap_add_extent_unwritten_real(NULL, ip,
272 				XFS_COW_FORK, &icur, &dummy_cur, &got,
273 				&dummy_logflags);
274 		if (error)
275 			return error;
276 	} while (xfs_iext_next_extent(ip->i_cowfp, &icur, &got));
277 
278 	return error;
279 }
280 
281 /* Convert all of the unwritten CoW extents in a file's range to real ones. */
282 int
283 xfs_reflink_convert_cow(
284 	struct xfs_inode	*ip,
285 	xfs_off_t		offset,
286 	xfs_off_t		count)
287 {
288 	struct xfs_mount	*mp = ip->i_mount;
289 	xfs_fileoff_t		offset_fsb = XFS_B_TO_FSBT(mp, offset);
290 	xfs_fileoff_t		end_fsb = XFS_B_TO_FSB(mp, offset + count);
291 	xfs_filblks_t		count_fsb = end_fsb - offset_fsb;
292 	int			error;
293 
294 	ASSERT(count != 0);
295 
296 	xfs_ilock(ip, XFS_ILOCK_EXCL);
297 	error = xfs_reflink_convert_cow_locked(ip, offset_fsb, count_fsb);
298 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
299 	return error;
300 }
301 
302 /*
303  * Find the extent that maps the given range in the COW fork. Even if the extent
304  * is not shared we might have a preallocation for it in the COW fork. If so we
305  * use it that rather than trigger a new allocation.
306  */
307 static int
308 xfs_find_trim_cow_extent(
309 	struct xfs_inode	*ip,
310 	struct xfs_bmbt_irec	*imap,
311 	bool			*shared,
312 	bool			*found)
313 {
314 	xfs_fileoff_t		offset_fsb = imap->br_startoff;
315 	xfs_filblks_t		count_fsb = imap->br_blockcount;
316 	struct xfs_iext_cursor	icur;
317 	struct xfs_bmbt_irec	got;
318 
319 	*found = false;
320 
321 	/*
322 	 * If we don't find an overlapping extent, trim the range we need to
323 	 * allocate to fit the hole we found.
324 	 */
325 	if (!xfs_iext_lookup_extent(ip, ip->i_cowfp, offset_fsb, &icur, &got))
326 		got.br_startoff = offset_fsb + count_fsb;
327 	if (got.br_startoff > offset_fsb) {
328 		xfs_trim_extent(imap, imap->br_startoff,
329 				got.br_startoff - imap->br_startoff);
330 		return xfs_inode_need_cow(ip, imap, shared);
331 	}
332 
333 	*shared = true;
334 	if (isnullstartblock(got.br_startblock)) {
335 		xfs_trim_extent(imap, got.br_startoff, got.br_blockcount);
336 		return 0;
337 	}
338 
339 	/* real extent found - no need to allocate */
340 	xfs_trim_extent(&got, offset_fsb, count_fsb);
341 	*imap = got;
342 	*found = true;
343 	return 0;
344 }
345 
346 /* Allocate all CoW reservations covering a range of blocks in a file. */
347 int
348 xfs_reflink_allocate_cow(
349 	struct xfs_inode	*ip,
350 	struct xfs_bmbt_irec	*imap,
351 	bool			*shared,
352 	uint			*lockmode,
353 	bool			convert_now)
354 {
355 	struct xfs_mount	*mp = ip->i_mount;
356 	xfs_fileoff_t		offset_fsb = imap->br_startoff;
357 	xfs_filblks_t		count_fsb = imap->br_blockcount;
358 	struct xfs_trans	*tp;
359 	int			nimaps, error = 0;
360 	bool			found;
361 	xfs_filblks_t		resaligned;
362 	xfs_extlen_t		resblks = 0;
363 
364 	ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
365 	if (!ip->i_cowfp) {
366 		ASSERT(!xfs_is_reflink_inode(ip));
367 		xfs_ifork_init_cow(ip);
368 	}
369 
370 	error = xfs_find_trim_cow_extent(ip, imap, shared, &found);
371 	if (error || !*shared)
372 		return error;
373 	if (found)
374 		goto convert;
375 
376 	resaligned = xfs_aligned_fsb_count(imap->br_startoff,
377 		imap->br_blockcount, xfs_get_cowextsz_hint(ip));
378 	resblks = XFS_DIOSTRAT_SPACE_RES(mp, resaligned);
379 
380 	xfs_iunlock(ip, *lockmode);
381 	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, resblks, 0, 0, &tp);
382 	*lockmode = XFS_ILOCK_EXCL;
383 	xfs_ilock(ip, *lockmode);
384 
385 	if (error)
386 		return error;
387 
388 	error = xfs_qm_dqattach_locked(ip, false);
389 	if (error)
390 		goto out_trans_cancel;
391 
392 	/*
393 	 * Check for an overlapping extent again now that we dropped the ilock.
394 	 */
395 	error = xfs_find_trim_cow_extent(ip, imap, shared, &found);
396 	if (error || !*shared)
397 		goto out_trans_cancel;
398 	if (found) {
399 		xfs_trans_cancel(tp);
400 		goto convert;
401 	}
402 
403 	error = xfs_trans_reserve_quota_nblks(tp, ip, resblks, 0,
404 			XFS_QMOPT_RES_REGBLKS);
405 	if (error)
406 		goto out_trans_cancel;
407 
408 	xfs_trans_ijoin(tp, ip, 0);
409 
410 	/* Allocate the entire reservation as unwritten blocks. */
411 	nimaps = 1;
412 	error = xfs_bmapi_write(tp, ip, imap->br_startoff, imap->br_blockcount,
413 			XFS_BMAPI_COWFORK | XFS_BMAPI_PREALLOC,
414 			resblks, imap, &nimaps);
415 	if (error)
416 		goto out_unreserve;
417 
418 	xfs_inode_set_cowblocks_tag(ip);
419 	error = xfs_trans_commit(tp);
420 	if (error)
421 		return error;
422 
423 	/*
424 	 * Allocation succeeded but the requested range was not even partially
425 	 * satisfied?  Bail out!
426 	 */
427 	if (nimaps == 0)
428 		return -ENOSPC;
429 convert:
430 	xfs_trim_extent(imap, offset_fsb, count_fsb);
431 	/*
432 	 * COW fork extents are supposed to remain unwritten until we're ready
433 	 * to initiate a disk write.  For direct I/O we are going to write the
434 	 * data and need the conversion, but for buffered writes we're done.
435 	 */
436 	if (!convert_now || imap->br_state == XFS_EXT_NORM)
437 		return 0;
438 	trace_xfs_reflink_convert_cow(ip, imap);
439 	return xfs_reflink_convert_cow_locked(ip, offset_fsb, count_fsb);
440 
441 out_unreserve:
442 	xfs_trans_unreserve_quota_nblks(tp, ip, (long)resblks, 0,
443 			XFS_QMOPT_RES_REGBLKS);
444 out_trans_cancel:
445 	xfs_trans_cancel(tp);
446 	return error;
447 }
448 
449 /*
450  * Cancel CoW reservations for some block range of an inode.
451  *
452  * If cancel_real is true this function cancels all COW fork extents for the
453  * inode; if cancel_real is false, real extents are not cleared.
454  *
455  * Caller must have already joined the inode to the current transaction. The
456  * inode will be joined to the transaction returned to the caller.
457  */
458 int
459 xfs_reflink_cancel_cow_blocks(
460 	struct xfs_inode		*ip,
461 	struct xfs_trans		**tpp,
462 	xfs_fileoff_t			offset_fsb,
463 	xfs_fileoff_t			end_fsb,
464 	bool				cancel_real)
465 {
466 	struct xfs_ifork		*ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
467 	struct xfs_bmbt_irec		got, del;
468 	struct xfs_iext_cursor		icur;
469 	int				error = 0;
470 
471 	if (!xfs_inode_has_cow_data(ip))
472 		return 0;
473 	if (!xfs_iext_lookup_extent_before(ip, ifp, &end_fsb, &icur, &got))
474 		return 0;
475 
476 	/* Walk backwards until we're out of the I/O range... */
477 	while (got.br_startoff + got.br_blockcount > offset_fsb) {
478 		del = got;
479 		xfs_trim_extent(&del, offset_fsb, end_fsb - offset_fsb);
480 
481 		/* Extent delete may have bumped ext forward */
482 		if (!del.br_blockcount) {
483 			xfs_iext_prev(ifp, &icur);
484 			goto next_extent;
485 		}
486 
487 		trace_xfs_reflink_cancel_cow(ip, &del);
488 
489 		if (isnullstartblock(del.br_startblock)) {
490 			error = xfs_bmap_del_extent_delay(ip, XFS_COW_FORK,
491 					&icur, &got, &del);
492 			if (error)
493 				break;
494 		} else if (del.br_state == XFS_EXT_UNWRITTEN || cancel_real) {
495 			ASSERT((*tpp)->t_firstblock == NULLFSBLOCK);
496 
497 			/* Free the CoW orphan record. */
498 			error = xfs_refcount_free_cow_extent(*tpp,
499 					del.br_startblock, del.br_blockcount);
500 			if (error)
501 				break;
502 
503 			xfs_bmap_add_free(*tpp, del.br_startblock,
504 					  del.br_blockcount, NULL);
505 
506 			/* Roll the transaction */
507 			error = xfs_defer_finish(tpp);
508 			if (error)
509 				break;
510 
511 			/* Remove the mapping from the CoW fork. */
512 			xfs_bmap_del_extent_cow(ip, &icur, &got, &del);
513 
514 			/* Remove the quota reservation */
515 			error = xfs_trans_reserve_quota_nblks(NULL, ip,
516 					-(long)del.br_blockcount, 0,
517 					XFS_QMOPT_RES_REGBLKS);
518 			if (error)
519 				break;
520 		} else {
521 			/* Didn't do anything, push cursor back. */
522 			xfs_iext_prev(ifp, &icur);
523 		}
524 next_extent:
525 		if (!xfs_iext_get_extent(ifp, &icur, &got))
526 			break;
527 	}
528 
529 	/* clear tag if cow fork is emptied */
530 	if (!ifp->if_bytes)
531 		xfs_inode_clear_cowblocks_tag(ip);
532 	return error;
533 }
534 
535 /*
536  * Cancel CoW reservations for some byte range of an inode.
537  *
538  * If cancel_real is true this function cancels all COW fork extents for the
539  * inode; if cancel_real is false, real extents are not cleared.
540  */
541 int
542 xfs_reflink_cancel_cow_range(
543 	struct xfs_inode	*ip,
544 	xfs_off_t		offset,
545 	xfs_off_t		count,
546 	bool			cancel_real)
547 {
548 	struct xfs_trans	*tp;
549 	xfs_fileoff_t		offset_fsb;
550 	xfs_fileoff_t		end_fsb;
551 	int			error;
552 
553 	trace_xfs_reflink_cancel_cow_range(ip, offset, count);
554 	ASSERT(ip->i_cowfp);
555 
556 	offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
557 	if (count == NULLFILEOFF)
558 		end_fsb = NULLFILEOFF;
559 	else
560 		end_fsb = XFS_B_TO_FSB(ip->i_mount, offset + count);
561 
562 	/* Start a rolling transaction to remove the mappings */
563 	error = xfs_trans_alloc(ip->i_mount, &M_RES(ip->i_mount)->tr_write,
564 			0, 0, 0, &tp);
565 	if (error)
566 		goto out;
567 
568 	xfs_ilock(ip, XFS_ILOCK_EXCL);
569 	xfs_trans_ijoin(tp, ip, 0);
570 
571 	/* Scrape out the old CoW reservations */
572 	error = xfs_reflink_cancel_cow_blocks(ip, &tp, offset_fsb, end_fsb,
573 			cancel_real);
574 	if (error)
575 		goto out_cancel;
576 
577 	error = xfs_trans_commit(tp);
578 
579 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
580 	return error;
581 
582 out_cancel:
583 	xfs_trans_cancel(tp);
584 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
585 out:
586 	trace_xfs_reflink_cancel_cow_range_error(ip, error, _RET_IP_);
587 	return error;
588 }
589 
590 /*
591  * Remap part of the CoW fork into the data fork.
592  *
593  * We aim to remap the range starting at @offset_fsb and ending at @end_fsb
594  * into the data fork; this function will remap what it can (at the end of the
595  * range) and update @end_fsb appropriately.  Each remap gets its own
596  * transaction because we can end up merging and splitting bmbt blocks for
597  * every remap operation and we'd like to keep the block reservation
598  * requirements as low as possible.
599  */
600 STATIC int
601 xfs_reflink_end_cow_extent(
602 	struct xfs_inode	*ip,
603 	xfs_fileoff_t		offset_fsb,
604 	xfs_fileoff_t		*end_fsb)
605 {
606 	struct xfs_bmbt_irec	got, del;
607 	struct xfs_iext_cursor	icur;
608 	struct xfs_mount	*mp = ip->i_mount;
609 	struct xfs_trans	*tp;
610 	struct xfs_ifork	*ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
611 	xfs_filblks_t		rlen;
612 	unsigned int		resblks;
613 	int			error;
614 
615 	/* No COW extents?  That's easy! */
616 	if (ifp->if_bytes == 0) {
617 		*end_fsb = offset_fsb;
618 		return 0;
619 	}
620 
621 	resblks = XFS_EXTENTADD_SPACE_RES(mp, XFS_DATA_FORK);
622 	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, resblks, 0,
623 			XFS_TRANS_RESERVE, &tp);
624 	if (error)
625 		return error;
626 
627 	/*
628 	 * Lock the inode.  We have to ijoin without automatic unlock because
629 	 * the lead transaction is the refcountbt record deletion; the data
630 	 * fork update follows as a deferred log item.
631 	 */
632 	xfs_ilock(ip, XFS_ILOCK_EXCL);
633 	xfs_trans_ijoin(tp, ip, 0);
634 
635 	/*
636 	 * In case of racing, overlapping AIO writes no COW extents might be
637 	 * left by the time I/O completes for the loser of the race.  In that
638 	 * case we are done.
639 	 */
640 	if (!xfs_iext_lookup_extent_before(ip, ifp, end_fsb, &icur, &got) ||
641 	    got.br_startoff + got.br_blockcount <= offset_fsb) {
642 		*end_fsb = offset_fsb;
643 		goto out_cancel;
644 	}
645 
646 	/*
647 	 * Structure copy @got into @del, then trim @del to the range that we
648 	 * were asked to remap.  We preserve @got for the eventual CoW fork
649 	 * deletion; from now on @del represents the mapping that we're
650 	 * actually remapping.
651 	 */
652 	del = got;
653 	xfs_trim_extent(&del, offset_fsb, *end_fsb - offset_fsb);
654 
655 	ASSERT(del.br_blockcount > 0);
656 
657 	/*
658 	 * Only remap real extents that contain data.  With AIO, speculative
659 	 * preallocations can leak into the range we are called upon, and we
660 	 * need to skip them.
661 	 */
662 	if (!xfs_bmap_is_real_extent(&got)) {
663 		*end_fsb = del.br_startoff;
664 		goto out_cancel;
665 	}
666 
667 	/* Unmap the old blocks in the data fork. */
668 	rlen = del.br_blockcount;
669 	error = __xfs_bunmapi(tp, ip, del.br_startoff, &rlen, 0, 1);
670 	if (error)
671 		goto out_cancel;
672 
673 	/* Trim the extent to whatever got unmapped. */
674 	xfs_trim_extent(&del, del.br_startoff + rlen, del.br_blockcount - rlen);
675 	trace_xfs_reflink_cow_remap(ip, &del);
676 
677 	/* Free the CoW orphan record. */
678 	error = xfs_refcount_free_cow_extent(tp, del.br_startblock,
679 			del.br_blockcount);
680 	if (error)
681 		goto out_cancel;
682 
683 	/* Map the new blocks into the data fork. */
684 	error = xfs_bmap_map_extent(tp, ip, &del);
685 	if (error)
686 		goto out_cancel;
687 
688 	/* Charge this new data fork mapping to the on-disk quota. */
689 	xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_DELBCOUNT,
690 			(long)del.br_blockcount);
691 
692 	/* Remove the mapping from the CoW fork. */
693 	xfs_bmap_del_extent_cow(ip, &icur, &got, &del);
694 
695 	error = xfs_trans_commit(tp);
696 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
697 	if (error)
698 		return error;
699 
700 	/* Update the caller about how much progress we made. */
701 	*end_fsb = del.br_startoff;
702 	return 0;
703 
704 out_cancel:
705 	xfs_trans_cancel(tp);
706 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
707 	return error;
708 }
709 
710 /*
711  * Remap parts of a file's data fork after a successful CoW.
712  */
713 int
714 xfs_reflink_end_cow(
715 	struct xfs_inode		*ip,
716 	xfs_off_t			offset,
717 	xfs_off_t			count)
718 {
719 	xfs_fileoff_t			offset_fsb;
720 	xfs_fileoff_t			end_fsb;
721 	int				error = 0;
722 
723 	trace_xfs_reflink_end_cow(ip, offset, count);
724 
725 	offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
726 	end_fsb = XFS_B_TO_FSB(ip->i_mount, offset + count);
727 
728 	/*
729 	 * Walk backwards until we're out of the I/O range.  The loop function
730 	 * repeatedly cycles the ILOCK to allocate one transaction per remapped
731 	 * extent.
732 	 *
733 	 * If we're being called by writeback then the the pages will still
734 	 * have PageWriteback set, which prevents races with reflink remapping
735 	 * and truncate.  Reflink remapping prevents races with writeback by
736 	 * taking the iolock and mmaplock before flushing the pages and
737 	 * remapping, which means there won't be any further writeback or page
738 	 * cache dirtying until the reflink completes.
739 	 *
740 	 * We should never have two threads issuing writeback for the same file
741 	 * region.  There are also have post-eof checks in the writeback
742 	 * preparation code so that we don't bother writing out pages that are
743 	 * about to be truncated.
744 	 *
745 	 * If we're being called as part of directio write completion, the dio
746 	 * count is still elevated, which reflink and truncate will wait for.
747 	 * Reflink remapping takes the iolock and mmaplock and waits for
748 	 * pending dio to finish, which should prevent any directio until the
749 	 * remap completes.  Multiple concurrent directio writes to the same
750 	 * region are handled by end_cow processing only occurring for the
751 	 * threads which succeed; the outcome of multiple overlapping direct
752 	 * writes is not well defined anyway.
753 	 *
754 	 * It's possible that a buffered write and a direct write could collide
755 	 * here (the buffered write stumbles in after the dio flushes and
756 	 * invalidates the page cache and immediately queues writeback), but we
757 	 * have never supported this 100%.  If either disk write succeeds the
758 	 * blocks will be remapped.
759 	 */
760 	while (end_fsb > offset_fsb && !error)
761 		error = xfs_reflink_end_cow_extent(ip, offset_fsb, &end_fsb);
762 
763 	if (error)
764 		trace_xfs_reflink_end_cow_error(ip, error, _RET_IP_);
765 	return error;
766 }
767 
768 /*
769  * Free leftover CoW reservations that didn't get cleaned out.
770  */
771 int
772 xfs_reflink_recover_cow(
773 	struct xfs_mount	*mp)
774 {
775 	xfs_agnumber_t		agno;
776 	int			error = 0;
777 
778 	if (!xfs_sb_version_hasreflink(&mp->m_sb))
779 		return 0;
780 
781 	for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
782 		error = xfs_refcount_recover_cow_leftovers(mp, agno);
783 		if (error)
784 			break;
785 	}
786 
787 	return error;
788 }
789 
790 /*
791  * Reflinking (Block) Ranges of Two Files Together
792  *
793  * First, ensure that the reflink flag is set on both inodes.  The flag is an
794  * optimization to avoid unnecessary refcount btree lookups in the write path.
795  *
796  * Now we can iteratively remap the range of extents (and holes) in src to the
797  * corresponding ranges in dest.  Let drange and srange denote the ranges of
798  * logical blocks in dest and src touched by the reflink operation.
799  *
800  * While the length of drange is greater than zero,
801  *    - Read src's bmbt at the start of srange ("imap")
802  *    - If imap doesn't exist, make imap appear to start at the end of srange
803  *      with zero length.
804  *    - If imap starts before srange, advance imap to start at srange.
805  *    - If imap goes beyond srange, truncate imap to end at the end of srange.
806  *    - Punch (imap start - srange start + imap len) blocks from dest at
807  *      offset (drange start).
808  *    - If imap points to a real range of pblks,
809  *         > Increase the refcount of the imap's pblks
810  *         > Map imap's pblks into dest at the offset
811  *           (drange start + imap start - srange start)
812  *    - Advance drange and srange by (imap start - srange start + imap len)
813  *
814  * Finally, if the reflink made dest longer, update both the in-core and
815  * on-disk file sizes.
816  *
817  * ASCII Art Demonstration:
818  *
819  * Let's say we want to reflink this source file:
820  *
821  * ----SSSSSSS-SSSSS----SSSSSS (src file)
822  *   <-------------------->
823  *
824  * into this destination file:
825  *
826  * --DDDDDDDDDDDDDDDDDDD--DDD (dest file)
827  *        <-------------------->
828  * '-' means a hole, and 'S' and 'D' are written blocks in the src and dest.
829  * Observe that the range has different logical offsets in either file.
830  *
831  * Consider that the first extent in the source file doesn't line up with our
832  * reflink range.  Unmapping  and remapping are separate operations, so we can
833  * unmap more blocks from the destination file than we remap.
834  *
835  * ----SSSSSSS-SSSSS----SSSSSS
836  *   <------->
837  * --DDDDD---------DDDDD--DDD
838  *        <------->
839  *
840  * Now remap the source extent into the destination file:
841  *
842  * ----SSSSSSS-SSSSS----SSSSSS
843  *   <------->
844  * --DDDDD--SSSSSSSDDDDD--DDD
845  *        <------->
846  *
847  * Do likewise with the second hole and extent in our range.  Holes in the
848  * unmap range don't affect our operation.
849  *
850  * ----SSSSSSS-SSSSS----SSSSSS
851  *            <---->
852  * --DDDDD--SSSSSSS-SSSSS-DDD
853  *                 <---->
854  *
855  * Finally, unmap and remap part of the third extent.  This will increase the
856  * size of the destination file.
857  *
858  * ----SSSSSSS-SSSSS----SSSSSS
859  *                  <----->
860  * --DDDDD--SSSSSSS-SSSSS----SSS
861  *                       <----->
862  *
863  * Once we update the destination file's i_size, we're done.
864  */
865 
866 /*
867  * Ensure the reflink bit is set in both inodes.
868  */
869 STATIC int
870 xfs_reflink_set_inode_flag(
871 	struct xfs_inode	*src,
872 	struct xfs_inode	*dest)
873 {
874 	struct xfs_mount	*mp = src->i_mount;
875 	int			error;
876 	struct xfs_trans	*tp;
877 
878 	if (xfs_is_reflink_inode(src) && xfs_is_reflink_inode(dest))
879 		return 0;
880 
881 	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
882 	if (error)
883 		goto out_error;
884 
885 	/* Lock both files against IO */
886 	if (src->i_ino == dest->i_ino)
887 		xfs_ilock(src, XFS_ILOCK_EXCL);
888 	else
889 		xfs_lock_two_inodes(src, XFS_ILOCK_EXCL, dest, XFS_ILOCK_EXCL);
890 
891 	if (!xfs_is_reflink_inode(src)) {
892 		trace_xfs_reflink_set_inode_flag(src);
893 		xfs_trans_ijoin(tp, src, XFS_ILOCK_EXCL);
894 		src->i_d.di_flags2 |= XFS_DIFLAG2_REFLINK;
895 		xfs_trans_log_inode(tp, src, XFS_ILOG_CORE);
896 		xfs_ifork_init_cow(src);
897 	} else
898 		xfs_iunlock(src, XFS_ILOCK_EXCL);
899 
900 	if (src->i_ino == dest->i_ino)
901 		goto commit_flags;
902 
903 	if (!xfs_is_reflink_inode(dest)) {
904 		trace_xfs_reflink_set_inode_flag(dest);
905 		xfs_trans_ijoin(tp, dest, XFS_ILOCK_EXCL);
906 		dest->i_d.di_flags2 |= XFS_DIFLAG2_REFLINK;
907 		xfs_trans_log_inode(tp, dest, XFS_ILOG_CORE);
908 		xfs_ifork_init_cow(dest);
909 	} else
910 		xfs_iunlock(dest, XFS_ILOCK_EXCL);
911 
912 commit_flags:
913 	error = xfs_trans_commit(tp);
914 	if (error)
915 		goto out_error;
916 	return error;
917 
918 out_error:
919 	trace_xfs_reflink_set_inode_flag_error(dest, error, _RET_IP_);
920 	return error;
921 }
922 
923 /*
924  * Update destination inode size & cowextsize hint, if necessary.
925  */
926 int
927 xfs_reflink_update_dest(
928 	struct xfs_inode	*dest,
929 	xfs_off_t		newlen,
930 	xfs_extlen_t		cowextsize,
931 	unsigned int		remap_flags)
932 {
933 	struct xfs_mount	*mp = dest->i_mount;
934 	struct xfs_trans	*tp;
935 	int			error;
936 
937 	if (newlen <= i_size_read(VFS_I(dest)) && cowextsize == 0)
938 		return 0;
939 
940 	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
941 	if (error)
942 		goto out_error;
943 
944 	xfs_ilock(dest, XFS_ILOCK_EXCL);
945 	xfs_trans_ijoin(tp, dest, XFS_ILOCK_EXCL);
946 
947 	if (newlen > i_size_read(VFS_I(dest))) {
948 		trace_xfs_reflink_update_inode_size(dest, newlen);
949 		i_size_write(VFS_I(dest), newlen);
950 		dest->i_d.di_size = newlen;
951 	}
952 
953 	if (cowextsize) {
954 		dest->i_d.di_cowextsize = cowextsize;
955 		dest->i_d.di_flags2 |= XFS_DIFLAG2_COWEXTSIZE;
956 	}
957 
958 	xfs_trans_log_inode(tp, dest, XFS_ILOG_CORE);
959 
960 	error = xfs_trans_commit(tp);
961 	if (error)
962 		goto out_error;
963 	return error;
964 
965 out_error:
966 	trace_xfs_reflink_update_inode_size_error(dest, error, _RET_IP_);
967 	return error;
968 }
969 
970 /*
971  * Do we have enough reserve in this AG to handle a reflink?  The refcount
972  * btree already reserved all the space it needs, but the rmap btree can grow
973  * infinitely, so we won't allow more reflinks when the AG is down to the
974  * btree reserves.
975  */
976 static int
977 xfs_reflink_ag_has_free_space(
978 	struct xfs_mount	*mp,
979 	xfs_agnumber_t		agno)
980 {
981 	struct xfs_perag	*pag;
982 	int			error = 0;
983 
984 	if (!xfs_sb_version_hasrmapbt(&mp->m_sb))
985 		return 0;
986 
987 	pag = xfs_perag_get(mp, agno);
988 	if (xfs_ag_resv_critical(pag, XFS_AG_RESV_RMAPBT) ||
989 	    xfs_ag_resv_critical(pag, XFS_AG_RESV_METADATA))
990 		error = -ENOSPC;
991 	xfs_perag_put(pag);
992 	return error;
993 }
994 
995 /*
996  * Unmap a range of blocks from a file, then map other blocks into the hole.
997  * The range to unmap is (destoff : destoff + srcioff + irec->br_blockcount).
998  * The extent irec is mapped into dest at irec->br_startoff.
999  */
1000 STATIC int
1001 xfs_reflink_remap_extent(
1002 	struct xfs_inode	*ip,
1003 	struct xfs_bmbt_irec	*irec,
1004 	xfs_fileoff_t		destoff,
1005 	xfs_off_t		new_isize)
1006 {
1007 	struct xfs_mount	*mp = ip->i_mount;
1008 	bool			real_extent = xfs_bmap_is_real_extent(irec);
1009 	struct xfs_trans	*tp;
1010 	unsigned int		resblks;
1011 	struct xfs_bmbt_irec	uirec;
1012 	xfs_filblks_t		rlen;
1013 	xfs_filblks_t		unmap_len;
1014 	xfs_off_t		newlen;
1015 	int			error;
1016 
1017 	unmap_len = irec->br_startoff + irec->br_blockcount - destoff;
1018 	trace_xfs_reflink_punch_range(ip, destoff, unmap_len);
1019 
1020 	/* No reflinking if we're low on space */
1021 	if (real_extent) {
1022 		error = xfs_reflink_ag_has_free_space(mp,
1023 				XFS_FSB_TO_AGNO(mp, irec->br_startblock));
1024 		if (error)
1025 			goto out;
1026 	}
1027 
1028 	/* Start a rolling transaction to switch the mappings */
1029 	resblks = XFS_EXTENTADD_SPACE_RES(ip->i_mount, XFS_DATA_FORK);
1030 	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, resblks, 0, 0, &tp);
1031 	if (error)
1032 		goto out;
1033 
1034 	xfs_ilock(ip, XFS_ILOCK_EXCL);
1035 	xfs_trans_ijoin(tp, ip, 0);
1036 
1037 	/* If we're not just clearing space, then do we have enough quota? */
1038 	if (real_extent) {
1039 		error = xfs_trans_reserve_quota_nblks(tp, ip,
1040 				irec->br_blockcount, 0, XFS_QMOPT_RES_REGBLKS);
1041 		if (error)
1042 			goto out_cancel;
1043 	}
1044 
1045 	trace_xfs_reflink_remap(ip, irec->br_startoff,
1046 				irec->br_blockcount, irec->br_startblock);
1047 
1048 	/* Unmap the old blocks in the data fork. */
1049 	rlen = unmap_len;
1050 	while (rlen) {
1051 		ASSERT(tp->t_firstblock == NULLFSBLOCK);
1052 		error = __xfs_bunmapi(tp, ip, destoff, &rlen, 0, 1);
1053 		if (error)
1054 			goto out_cancel;
1055 
1056 		/*
1057 		 * Trim the extent to whatever got unmapped.
1058 		 * Remember, bunmapi works backwards.
1059 		 */
1060 		uirec.br_startblock = irec->br_startblock + rlen;
1061 		uirec.br_startoff = irec->br_startoff + rlen;
1062 		uirec.br_blockcount = unmap_len - rlen;
1063 		unmap_len = rlen;
1064 
1065 		/* If this isn't a real mapping, we're done. */
1066 		if (!real_extent || uirec.br_blockcount == 0)
1067 			goto next_extent;
1068 
1069 		trace_xfs_reflink_remap(ip, uirec.br_startoff,
1070 				uirec.br_blockcount, uirec.br_startblock);
1071 
1072 		/* Update the refcount tree */
1073 		error = xfs_refcount_increase_extent(tp, &uirec);
1074 		if (error)
1075 			goto out_cancel;
1076 
1077 		/* Map the new blocks into the data fork. */
1078 		error = xfs_bmap_map_extent(tp, ip, &uirec);
1079 		if (error)
1080 			goto out_cancel;
1081 
1082 		/* Update quota accounting. */
1083 		xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_BCOUNT,
1084 				uirec.br_blockcount);
1085 
1086 		/* Update dest isize if needed. */
1087 		newlen = XFS_FSB_TO_B(mp,
1088 				uirec.br_startoff + uirec.br_blockcount);
1089 		newlen = min_t(xfs_off_t, newlen, new_isize);
1090 		if (newlen > i_size_read(VFS_I(ip))) {
1091 			trace_xfs_reflink_update_inode_size(ip, newlen);
1092 			i_size_write(VFS_I(ip), newlen);
1093 			ip->i_d.di_size = newlen;
1094 			xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1095 		}
1096 
1097 next_extent:
1098 		/* Process all the deferred stuff. */
1099 		error = xfs_defer_finish(&tp);
1100 		if (error)
1101 			goto out_cancel;
1102 	}
1103 
1104 	error = xfs_trans_commit(tp);
1105 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1106 	if (error)
1107 		goto out;
1108 	return 0;
1109 
1110 out_cancel:
1111 	xfs_trans_cancel(tp);
1112 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1113 out:
1114 	trace_xfs_reflink_remap_extent_error(ip, error, _RET_IP_);
1115 	return error;
1116 }
1117 
1118 /*
1119  * Iteratively remap one file's extents (and holes) to another's.
1120  */
1121 int
1122 xfs_reflink_remap_blocks(
1123 	struct xfs_inode	*src,
1124 	loff_t			pos_in,
1125 	struct xfs_inode	*dest,
1126 	loff_t			pos_out,
1127 	loff_t			remap_len,
1128 	loff_t			*remapped)
1129 {
1130 	struct xfs_bmbt_irec	imap;
1131 	xfs_fileoff_t		srcoff;
1132 	xfs_fileoff_t		destoff;
1133 	xfs_filblks_t		len;
1134 	xfs_filblks_t		range_len;
1135 	xfs_filblks_t		remapped_len = 0;
1136 	xfs_off_t		new_isize = pos_out + remap_len;
1137 	int			nimaps;
1138 	int			error = 0;
1139 
1140 	destoff = XFS_B_TO_FSBT(src->i_mount, pos_out);
1141 	srcoff = XFS_B_TO_FSBT(src->i_mount, pos_in);
1142 	len = XFS_B_TO_FSB(src->i_mount, remap_len);
1143 
1144 	/* drange = (destoff, destoff + len); srange = (srcoff, srcoff + len) */
1145 	while (len) {
1146 		uint		lock_mode;
1147 
1148 		trace_xfs_reflink_remap_blocks_loop(src, srcoff, len,
1149 				dest, destoff);
1150 
1151 		/* Read extent from the source file */
1152 		nimaps = 1;
1153 		lock_mode = xfs_ilock_data_map_shared(src);
1154 		error = xfs_bmapi_read(src, srcoff, len, &imap, &nimaps, 0);
1155 		xfs_iunlock(src, lock_mode);
1156 		if (error)
1157 			break;
1158 		ASSERT(nimaps == 1);
1159 
1160 		trace_xfs_reflink_remap_imap(src, srcoff, len, XFS_DATA_FORK,
1161 				&imap);
1162 
1163 		/* Translate imap into the destination file. */
1164 		range_len = imap.br_startoff + imap.br_blockcount - srcoff;
1165 		imap.br_startoff += destoff - srcoff;
1166 
1167 		/* Clear dest from destoff to the end of imap and map it in. */
1168 		error = xfs_reflink_remap_extent(dest, &imap, destoff,
1169 				new_isize);
1170 		if (error)
1171 			break;
1172 
1173 		if (fatal_signal_pending(current)) {
1174 			error = -EINTR;
1175 			break;
1176 		}
1177 
1178 		/* Advance drange/srange */
1179 		srcoff += range_len;
1180 		destoff += range_len;
1181 		len -= range_len;
1182 		remapped_len += range_len;
1183 	}
1184 
1185 	if (error)
1186 		trace_xfs_reflink_remap_blocks_error(dest, error, _RET_IP_);
1187 	*remapped = min_t(loff_t, remap_len,
1188 			  XFS_FSB_TO_B(src->i_mount, remapped_len));
1189 	return error;
1190 }
1191 
1192 /*
1193  * Grab the exclusive iolock for a data copy from src to dest, making sure to
1194  * abide vfs locking order (lowest pointer value goes first) and breaking the
1195  * layout leases before proceeding.  The loop is needed because we cannot call
1196  * the blocking break_layout() with the iolocks held, and therefore have to
1197  * back out both locks.
1198  */
1199 static int
1200 xfs_iolock_two_inodes_and_break_layout(
1201 	struct inode		*src,
1202 	struct inode		*dest)
1203 {
1204 	int			error;
1205 
1206 	if (src > dest)
1207 		swap(src, dest);
1208 
1209 retry:
1210 	/* Wait to break both inodes' layouts before we start locking. */
1211 	error = break_layout(src, true);
1212 	if (error)
1213 		return error;
1214 	if (src != dest) {
1215 		error = break_layout(dest, true);
1216 		if (error)
1217 			return error;
1218 	}
1219 
1220 	/* Lock one inode and make sure nobody got in and leased it. */
1221 	inode_lock(src);
1222 	error = break_layout(src, false);
1223 	if (error) {
1224 		inode_unlock(src);
1225 		if (error == -EWOULDBLOCK)
1226 			goto retry;
1227 		return error;
1228 	}
1229 
1230 	if (src == dest)
1231 		return 0;
1232 
1233 	/* Lock the other inode and make sure nobody got in and leased it. */
1234 	inode_lock_nested(dest, I_MUTEX_NONDIR2);
1235 	error = break_layout(dest, false);
1236 	if (error) {
1237 		inode_unlock(src);
1238 		inode_unlock(dest);
1239 		if (error == -EWOULDBLOCK)
1240 			goto retry;
1241 		return error;
1242 	}
1243 
1244 	return 0;
1245 }
1246 
1247 /* Unlock both inodes after they've been prepped for a range clone. */
1248 void
1249 xfs_reflink_remap_unlock(
1250 	struct file		*file_in,
1251 	struct file		*file_out)
1252 {
1253 	struct inode		*inode_in = file_inode(file_in);
1254 	struct xfs_inode	*src = XFS_I(inode_in);
1255 	struct inode		*inode_out = file_inode(file_out);
1256 	struct xfs_inode	*dest = XFS_I(inode_out);
1257 	bool			same_inode = (inode_in == inode_out);
1258 
1259 	xfs_iunlock(dest, XFS_MMAPLOCK_EXCL);
1260 	if (!same_inode)
1261 		xfs_iunlock(src, XFS_MMAPLOCK_EXCL);
1262 	inode_unlock(inode_out);
1263 	if (!same_inode)
1264 		inode_unlock(inode_in);
1265 }
1266 
1267 /*
1268  * If we're reflinking to a point past the destination file's EOF, we must
1269  * zero any speculative post-EOF preallocations that sit between the old EOF
1270  * and the destination file offset.
1271  */
1272 static int
1273 xfs_reflink_zero_posteof(
1274 	struct xfs_inode	*ip,
1275 	loff_t			pos)
1276 {
1277 	loff_t			isize = i_size_read(VFS_I(ip));
1278 
1279 	if (pos <= isize)
1280 		return 0;
1281 
1282 	trace_xfs_zero_eof(ip, isize, pos - isize);
1283 	return iomap_zero_range(VFS_I(ip), isize, pos - isize, NULL,
1284 			&xfs_iomap_ops);
1285 }
1286 
1287 /*
1288  * Prepare two files for range cloning.  Upon a successful return both inodes
1289  * will have the iolock and mmaplock held, the page cache of the out file will
1290  * be truncated, and any leases on the out file will have been broken.  This
1291  * function borrows heavily from xfs_file_aio_write_checks.
1292  *
1293  * The VFS allows partial EOF blocks to "match" for dedupe even though it hasn't
1294  * checked that the bytes beyond EOF physically match. Hence we cannot use the
1295  * EOF block in the source dedupe range because it's not a complete block match,
1296  * hence can introduce a corruption into the file that has it's block replaced.
1297  *
1298  * In similar fashion, the VFS file cloning also allows partial EOF blocks to be
1299  * "block aligned" for the purposes of cloning entire files.  However, if the
1300  * source file range includes the EOF block and it lands within the existing EOF
1301  * of the destination file, then we can expose stale data from beyond the source
1302  * file EOF in the destination file.
1303  *
1304  * XFS doesn't support partial block sharing, so in both cases we have check
1305  * these cases ourselves. For dedupe, we can simply round the length to dedupe
1306  * down to the previous whole block and ignore the partial EOF block. While this
1307  * means we can't dedupe the last block of a file, this is an acceptible
1308  * tradeoff for simplicity on implementation.
1309  *
1310  * For cloning, we want to share the partial EOF block if it is also the new EOF
1311  * block of the destination file. If the partial EOF block lies inside the
1312  * existing destination EOF, then we have to abort the clone to avoid exposing
1313  * stale data in the destination file. Hence we reject these clone attempts with
1314  * -EINVAL in this case.
1315  */
1316 int
1317 xfs_reflink_remap_prep(
1318 	struct file		*file_in,
1319 	loff_t			pos_in,
1320 	struct file		*file_out,
1321 	loff_t			pos_out,
1322 	loff_t			*len,
1323 	unsigned int		remap_flags)
1324 {
1325 	struct inode		*inode_in = file_inode(file_in);
1326 	struct xfs_inode	*src = XFS_I(inode_in);
1327 	struct inode		*inode_out = file_inode(file_out);
1328 	struct xfs_inode	*dest = XFS_I(inode_out);
1329 	bool			same_inode = (inode_in == inode_out);
1330 	ssize_t			ret;
1331 
1332 	/* Lock both files against IO */
1333 	ret = xfs_iolock_two_inodes_and_break_layout(inode_in, inode_out);
1334 	if (ret)
1335 		return ret;
1336 	if (same_inode)
1337 		xfs_ilock(src, XFS_MMAPLOCK_EXCL);
1338 	else
1339 		xfs_lock_two_inodes(src, XFS_MMAPLOCK_EXCL, dest,
1340 				XFS_MMAPLOCK_EXCL);
1341 
1342 	/* Check file eligibility and prepare for block sharing. */
1343 	ret = -EINVAL;
1344 	/* Don't reflink realtime inodes */
1345 	if (XFS_IS_REALTIME_INODE(src) || XFS_IS_REALTIME_INODE(dest))
1346 		goto out_unlock;
1347 
1348 	/* Don't share DAX file data for now. */
1349 	if (IS_DAX(inode_in) || IS_DAX(inode_out))
1350 		goto out_unlock;
1351 
1352 	ret = generic_remap_file_range_prep(file_in, pos_in, file_out, pos_out,
1353 			len, remap_flags);
1354 	if (ret < 0 || *len == 0)
1355 		goto out_unlock;
1356 
1357 	/* Attach dquots to dest inode before changing block map */
1358 	ret = xfs_qm_dqattach(dest);
1359 	if (ret)
1360 		goto out_unlock;
1361 
1362 	/*
1363 	 * Zero existing post-eof speculative preallocations in the destination
1364 	 * file.
1365 	 */
1366 	ret = xfs_reflink_zero_posteof(dest, pos_out);
1367 	if (ret)
1368 		goto out_unlock;
1369 
1370 	/* Set flags and remap blocks. */
1371 	ret = xfs_reflink_set_inode_flag(src, dest);
1372 	if (ret)
1373 		goto out_unlock;
1374 
1375 	/*
1376 	 * If pos_out > EOF, we may have dirtied blocks between EOF and
1377 	 * pos_out. In that case, we need to extend the flush and unmap to cover
1378 	 * from EOF to the end of the copy length.
1379 	 */
1380 	if (pos_out > XFS_ISIZE(dest)) {
1381 		loff_t	flen = *len + (pos_out - XFS_ISIZE(dest));
1382 		ret = xfs_flush_unmap_range(dest, XFS_ISIZE(dest), flen);
1383 	} else {
1384 		ret = xfs_flush_unmap_range(dest, pos_out, *len);
1385 	}
1386 	if (ret)
1387 		goto out_unlock;
1388 
1389 	return 1;
1390 out_unlock:
1391 	xfs_reflink_remap_unlock(file_in, file_out);
1392 	return ret;
1393 }
1394 
1395 /*
1396  * The user wants to preemptively CoW all shared blocks in this file,
1397  * which enables us to turn off the reflink flag.  Iterate all
1398  * extents which are not prealloc/delalloc to see which ranges are
1399  * mentioned in the refcount tree, then read those blocks into the
1400  * pagecache, dirty them, fsync them back out, and then we can update
1401  * the inode flag.  What happens if we run out of memory? :)
1402  */
1403 STATIC int
1404 xfs_reflink_dirty_extents(
1405 	struct xfs_inode	*ip,
1406 	xfs_fileoff_t		fbno,
1407 	xfs_filblks_t		end,
1408 	xfs_off_t		isize)
1409 {
1410 	struct xfs_mount	*mp = ip->i_mount;
1411 	xfs_agnumber_t		agno;
1412 	xfs_agblock_t		agbno;
1413 	xfs_extlen_t		aglen;
1414 	xfs_agblock_t		rbno;
1415 	xfs_extlen_t		rlen;
1416 	xfs_off_t		fpos;
1417 	xfs_off_t		flen;
1418 	struct xfs_bmbt_irec	map[2];
1419 	int			nmaps;
1420 	int			error = 0;
1421 
1422 	while (end - fbno > 0) {
1423 		nmaps = 1;
1424 		/*
1425 		 * Look for extents in the file.  Skip holes, delalloc, or
1426 		 * unwritten extents; they can't be reflinked.
1427 		 */
1428 		error = xfs_bmapi_read(ip, fbno, end - fbno, map, &nmaps, 0);
1429 		if (error)
1430 			goto out;
1431 		if (nmaps == 0)
1432 			break;
1433 		if (!xfs_bmap_is_real_extent(&map[0]))
1434 			goto next;
1435 
1436 		map[1] = map[0];
1437 		while (map[1].br_blockcount) {
1438 			agno = XFS_FSB_TO_AGNO(mp, map[1].br_startblock);
1439 			agbno = XFS_FSB_TO_AGBNO(mp, map[1].br_startblock);
1440 			aglen = map[1].br_blockcount;
1441 
1442 			error = xfs_reflink_find_shared(mp, NULL, agno, agbno,
1443 					aglen, &rbno, &rlen, true);
1444 			if (error)
1445 				goto out;
1446 			if (rbno == NULLAGBLOCK)
1447 				break;
1448 
1449 			/* Dirty the pages */
1450 			xfs_iunlock(ip, XFS_ILOCK_EXCL);
1451 			fpos = XFS_FSB_TO_B(mp, map[1].br_startoff +
1452 					(rbno - agbno));
1453 			flen = XFS_FSB_TO_B(mp, rlen);
1454 			if (fpos + flen > isize)
1455 				flen = isize - fpos;
1456 			error = iomap_file_dirty(VFS_I(ip), fpos, flen,
1457 					&xfs_iomap_ops);
1458 			xfs_ilock(ip, XFS_ILOCK_EXCL);
1459 			if (error)
1460 				goto out;
1461 
1462 			map[1].br_blockcount -= (rbno - agbno + rlen);
1463 			map[1].br_startoff += (rbno - agbno + rlen);
1464 			map[1].br_startblock += (rbno - agbno + rlen);
1465 		}
1466 
1467 next:
1468 		fbno = map[0].br_startoff + map[0].br_blockcount;
1469 	}
1470 out:
1471 	return error;
1472 }
1473 
1474 /* Does this inode need the reflink flag? */
1475 int
1476 xfs_reflink_inode_has_shared_extents(
1477 	struct xfs_trans		*tp,
1478 	struct xfs_inode		*ip,
1479 	bool				*has_shared)
1480 {
1481 	struct xfs_bmbt_irec		got;
1482 	struct xfs_mount		*mp = ip->i_mount;
1483 	struct xfs_ifork		*ifp;
1484 	xfs_agnumber_t			agno;
1485 	xfs_agblock_t			agbno;
1486 	xfs_extlen_t			aglen;
1487 	xfs_agblock_t			rbno;
1488 	xfs_extlen_t			rlen;
1489 	struct xfs_iext_cursor		icur;
1490 	bool				found;
1491 	int				error;
1492 
1493 	ifp = XFS_IFORK_PTR(ip, XFS_DATA_FORK);
1494 	if (!(ifp->if_flags & XFS_IFEXTENTS)) {
1495 		error = xfs_iread_extents(tp, ip, XFS_DATA_FORK);
1496 		if (error)
1497 			return error;
1498 	}
1499 
1500 	*has_shared = false;
1501 	found = xfs_iext_lookup_extent(ip, ifp, 0, &icur, &got);
1502 	while (found) {
1503 		if (isnullstartblock(got.br_startblock) ||
1504 		    got.br_state != XFS_EXT_NORM)
1505 			goto next;
1506 		agno = XFS_FSB_TO_AGNO(mp, got.br_startblock);
1507 		agbno = XFS_FSB_TO_AGBNO(mp, got.br_startblock);
1508 		aglen = got.br_blockcount;
1509 
1510 		error = xfs_reflink_find_shared(mp, tp, agno, agbno, aglen,
1511 				&rbno, &rlen, false);
1512 		if (error)
1513 			return error;
1514 		/* Is there still a shared block here? */
1515 		if (rbno != NULLAGBLOCK) {
1516 			*has_shared = true;
1517 			return 0;
1518 		}
1519 next:
1520 		found = xfs_iext_next_extent(ifp, &icur, &got);
1521 	}
1522 
1523 	return 0;
1524 }
1525 
1526 /*
1527  * Clear the inode reflink flag if there are no shared extents.
1528  *
1529  * The caller is responsible for joining the inode to the transaction passed in.
1530  * The inode will be joined to the transaction that is returned to the caller.
1531  */
1532 int
1533 xfs_reflink_clear_inode_flag(
1534 	struct xfs_inode	*ip,
1535 	struct xfs_trans	**tpp)
1536 {
1537 	bool			needs_flag;
1538 	int			error = 0;
1539 
1540 	ASSERT(xfs_is_reflink_inode(ip));
1541 
1542 	error = xfs_reflink_inode_has_shared_extents(*tpp, ip, &needs_flag);
1543 	if (error || needs_flag)
1544 		return error;
1545 
1546 	/*
1547 	 * We didn't find any shared blocks so turn off the reflink flag.
1548 	 * First, get rid of any leftover CoW mappings.
1549 	 */
1550 	error = xfs_reflink_cancel_cow_blocks(ip, tpp, 0, NULLFILEOFF, true);
1551 	if (error)
1552 		return error;
1553 
1554 	/* Clear the inode flag. */
1555 	trace_xfs_reflink_unset_inode_flag(ip);
1556 	ip->i_d.di_flags2 &= ~XFS_DIFLAG2_REFLINK;
1557 	xfs_inode_clear_cowblocks_tag(ip);
1558 	xfs_trans_log_inode(*tpp, ip, XFS_ILOG_CORE);
1559 
1560 	return error;
1561 }
1562 
1563 /*
1564  * Clear the inode reflink flag if there are no shared extents and the size
1565  * hasn't changed.
1566  */
1567 STATIC int
1568 xfs_reflink_try_clear_inode_flag(
1569 	struct xfs_inode	*ip)
1570 {
1571 	struct xfs_mount	*mp = ip->i_mount;
1572 	struct xfs_trans	*tp;
1573 	int			error = 0;
1574 
1575 	/* Start a rolling transaction to remove the mappings */
1576 	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, 0, 0, 0, &tp);
1577 	if (error)
1578 		return error;
1579 
1580 	xfs_ilock(ip, XFS_ILOCK_EXCL);
1581 	xfs_trans_ijoin(tp, ip, 0);
1582 
1583 	error = xfs_reflink_clear_inode_flag(ip, &tp);
1584 	if (error)
1585 		goto cancel;
1586 
1587 	error = xfs_trans_commit(tp);
1588 	if (error)
1589 		goto out;
1590 
1591 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1592 	return 0;
1593 cancel:
1594 	xfs_trans_cancel(tp);
1595 out:
1596 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1597 	return error;
1598 }
1599 
1600 /*
1601  * Pre-COW all shared blocks within a given byte range of a file and turn off
1602  * the reflink flag if we unshare all of the file's blocks.
1603  */
1604 int
1605 xfs_reflink_unshare(
1606 	struct xfs_inode	*ip,
1607 	xfs_off_t		offset,
1608 	xfs_off_t		len)
1609 {
1610 	struct xfs_mount	*mp = ip->i_mount;
1611 	xfs_fileoff_t		fbno;
1612 	xfs_filblks_t		end;
1613 	xfs_off_t		isize;
1614 	int			error;
1615 
1616 	if (!xfs_is_reflink_inode(ip))
1617 		return 0;
1618 
1619 	trace_xfs_reflink_unshare(ip, offset, len);
1620 
1621 	inode_dio_wait(VFS_I(ip));
1622 
1623 	/* Try to CoW the selected ranges */
1624 	xfs_ilock(ip, XFS_ILOCK_EXCL);
1625 	fbno = XFS_B_TO_FSBT(mp, offset);
1626 	isize = i_size_read(VFS_I(ip));
1627 	end = XFS_B_TO_FSB(mp, offset + len);
1628 	error = xfs_reflink_dirty_extents(ip, fbno, end, isize);
1629 	if (error)
1630 		goto out_unlock;
1631 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1632 
1633 	/* Wait for the IO to finish */
1634 	error = filemap_write_and_wait(VFS_I(ip)->i_mapping);
1635 	if (error)
1636 		goto out;
1637 
1638 	/* Turn off the reflink flag if possible. */
1639 	error = xfs_reflink_try_clear_inode_flag(ip);
1640 	if (error)
1641 		goto out;
1642 
1643 	return 0;
1644 
1645 out_unlock:
1646 	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1647 out:
1648 	trace_xfs_reflink_unshare_error(ip, error, _RET_IP_);
1649 	return error;
1650 }
1651