xfs_reflink.c

   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_ag.h"
  31#include "xfs_ag_resv.h"
  32#include "xfs_health.h"
  33#include "xfs_rtrefcount_btree.h"
  34#include "xfs_rtalloc.h"
  35#include "xfs_rtgroup.h"
  36#include "xfs_metafile.h"
  37
  38/*
  39 * Copy on Write of Shared Blocks
  40 *
  41 * XFS must preserve "the usual" file semantics even when two files share
  42 * the same physical blocks.  This means that a write to one file must not
  43 * alter the blocks in a different file; the way that we'll do that is
  44 * through the use of a copy-on-write mechanism.  At a high level, that
  45 * means that when we want to write to a shared block, we allocate a new
  46 * block, write the data to the new block, and if that succeeds we map the
  47 * new block into the file.
  48 *
  49 * XFS provides a "delayed allocation" mechanism that defers the allocation
  50 * of disk blocks to dirty-but-not-yet-mapped file blocks as long as
  51 * possible.  This reduces fragmentation by enabling the filesystem to ask
  52 * for bigger chunks less often, which is exactly what we want for CoW.
  53 *
  54 * The delalloc mechanism begins when the kernel wants to make a block
  55 * writable (write_begin or page_mkwrite).  If the offset is not mapped, we
  56 * create a delalloc mapping, which is a regular in-core extent, but without
  57 * a real startblock.  (For delalloc mappings, the startblock encodes both
  58 * a flag that this is a delalloc mapping, and a worst-case estimate of how
  59 * many blocks might be required to put the mapping into the BMBT.)  delalloc
  60 * mappings are a reservation against the free space in the filesystem;
  61 * adjacent mappings can also be combined into fewer larger mappings.
  62 *
  63 * As an optimization, the CoW extent size hint (cowextsz) creates
  64 * outsized aligned delalloc reservations in the hope of landing out of
  65 * order nearby CoW writes in a single extent on disk, thereby reducing
  66 * fragmentation and improving future performance.
  67 *
  68 * D: --RRRRRRSSSRRRRRRRR--- (data fork)
  69 * C: ------DDDDDDD--------- (CoW fork)
  70 *
  71 * When dirty pages are being written out (typically in writepage), the
  72 * delalloc reservations are converted into unwritten mappings by
  73 * allocating blocks and replacing the delalloc mapping with real ones.
  74 * A delalloc mapping can be replaced by several unwritten ones if the
  75 * free space is fragmented.
  76 *
  77 * D: --RRRRRRSSSRRRRRRRR---
  78 * C: ------UUUUUUU---------
  79 *
  80 * We want to adapt the delalloc mechanism for copy-on-write, since the
  81 * write paths are similar.  The first two steps (creating the reservation
  82 * and allocating the blocks) are exactly the same as delalloc except that
  83 * the mappings must be stored in a separate CoW fork because we do not want
  84 * to disturb the mapping in the data fork until we're sure that the write
  85 * succeeded.  IO completion in this case is the process of removing the old
  86 * mapping from the data fork and moving the new mapping from the CoW fork to
  87 * the data fork.  This will be discussed shortly.
  88 *
  89 * For now, unaligned directio writes will be bounced back to the page cache.
  90 * Block-aligned directio writes will use the same mechanism as buffered
  91 * writes.
  92 *
  93 * Just prior to submitting the actual disk write requests, we convert
  94 * the extents representing the range of the file actually being written
  95 * (as opposed to extra pieces created for the cowextsize hint) to real
  96 * extents.  This will become important in the next step:
  97 *
  98 * D: --RRRRRRSSSRRRRRRRR---
  99 * C: ------UUrrUUU---------
 100 *
 101 * CoW remapping must be done after the data block write completes,
 102 * because we don't want to destroy the old data fork map until we're sure
 103 * the new block has been written.  Since the new mappings are kept in a
 104 * separate fork, we can simply iterate these mappings to find the ones
 105 * that cover the file blocks that we just CoW'd.  For each extent, simply
 106 * unmap the corresponding range in the data fork, map the new range into
 107 * the data fork, and remove the extent from the CoW fork.  Because of
 108 * the presence of the cowextsize hint, however, we must be careful
 109 * only to remap the blocks that we've actually written out --  we must
 110 * never remap delalloc reservations nor CoW staging blocks that have
 111 * yet to be written.  This corresponds exactly to the real extents in
 112 * the CoW fork:
 113 *
 114 * D: --RRRRRRrrSRRRRRRRR---
 115 * C: ------UU--UUU---------
 116 *
 117 * Since the remapping operation can be applied to an arbitrary file
 118 * range, we record the need for the remap step as a flag in the ioend
 119 * instead of declaring a new IO type.  This is required for direct io
 120 * because we only have ioend for the whole dio, and we have to be able to
 121 * remember the presence of unwritten blocks and CoW blocks with a single
 122 * ioend structure.  Better yet, the more ground we can cover with one
 123 * ioend, the better.
 124 */
 125
 126/*
 127 * Given a file mapping for the data device, find the lowest-numbered run of
 128 * shared blocks within that mapping and return it in shared_offset/shared_len.
 129 * The offset is relative to the start of irec.
 130 *
 131 * If find_end_of_shared is true, return the longest contiguous extent of shared
 132 * blocks.  If there are no shared extents, shared_offset and shared_len will be
 133 * set to 0;
 134 */
 135static int
 136xfs_reflink_find_shared(
 137	struct xfs_mount	*mp,
 138	struct xfs_trans	*tp,
 139	const struct xfs_bmbt_irec *irec,
 140	xfs_extlen_t		*shared_offset,
 141	xfs_extlen_t		*shared_len,
 142	bool			find_end_of_shared)
 143{
 144	struct xfs_buf		*agbp;
 145	struct xfs_perag	*pag;
 146	struct xfs_btree_cur	*cur;
 147	int			error;
 148	xfs_agblock_t		orig_bno, found_bno;
 149
 150	pag = xfs_perag_get(mp, XFS_FSB_TO_AGNO(mp, irec->br_startblock));
 151	orig_bno = XFS_FSB_TO_AGBNO(mp, irec->br_startblock);
 152
 153	error = xfs_alloc_read_agf(pag, tp, 0, &agbp);
 154	if (error)
 155		goto out;
 156
 157	cur = xfs_refcountbt_init_cursor(mp, tp, agbp, pag);
 158	error = xfs_refcount_find_shared(cur, orig_bno, irec->br_blockcount,
 159			&found_bno, shared_len, find_end_of_shared);
 160	xfs_btree_del_cursor(cur, error);
 161	xfs_trans_brelse(tp, agbp);
 162
 163	if (!error && *shared_len)
 164		*shared_offset = found_bno - orig_bno;
 165out:
 166	xfs_perag_put(pag);
 167	return error;
 168}
 169
 170/*
 171 * Given a file mapping for the rt device, find the lowest-numbered run of
 172 * shared blocks within that mapping and return it in shared_offset/shared_len.
 173 * The offset is relative to the start of irec.
 174 *
 175 * If find_end_of_shared is true, return the longest contiguous extent of shared
 176 * blocks.  If there are no shared extents, shared_offset and shared_len will be
 177 * set to 0;
 178 */
 179static int
 180xfs_reflink_find_rtshared(
 181	struct xfs_mount	*mp,
 182	struct xfs_trans	*tp,
 183	const struct xfs_bmbt_irec *irec,
 184	xfs_extlen_t		*shared_offset,
 185	xfs_extlen_t		*shared_len,
 186	bool			find_end_of_shared)
 187{
 188	struct xfs_rtgroup	*rtg;
 189	struct xfs_btree_cur	*cur;
 190	xfs_rgblock_t		orig_bno;
 191	xfs_agblock_t		found_bno;
 192	int			error;
 193
 194	BUILD_BUG_ON(NULLRGBLOCK != NULLAGBLOCK);
 195
 196	/*
 197	 * Note: this uses the not quite correct xfs_agblock_t type because
 198	 * xfs_refcount_find_shared is shared between the RT and data device
 199	 * refcount code.
 200	 */
 201	orig_bno = xfs_rtb_to_rgbno(mp, irec->br_startblock);
 202	rtg = xfs_rtgroup_get(mp, xfs_rtb_to_rgno(mp, irec->br_startblock));
 203
 204	xfs_rtgroup_lock(rtg, XFS_RTGLOCK_REFCOUNT);
 205	cur = xfs_rtrefcountbt_init_cursor(tp, rtg);
 206	error = xfs_refcount_find_shared(cur, orig_bno, irec->br_blockcount,
 207			&found_bno, shared_len, find_end_of_shared);
 208	xfs_btree_del_cursor(cur, error);
 209	xfs_rtgroup_unlock(rtg, XFS_RTGLOCK_REFCOUNT);
 210	xfs_rtgroup_put(rtg);
 211
 212	if (!error && *shared_len)
 213		*shared_offset = found_bno - orig_bno;
 214	return error;
 215}
 216
 217/*
 218 * Trim the mapping to the next block where there's a change in the
 219 * shared/unshared status.  More specifically, this means that we
 220 * find the lowest-numbered extent of shared blocks that coincides with
 221 * the given block mapping.  If the shared extent overlaps the start of
 222 * the mapping, trim the mapping to the end of the shared extent.  If
 223 * the shared region intersects the mapping, trim the mapping to the
 224 * start of the shared extent.  If there are no shared regions that
 225 * overlap, just return the original extent.
 226 */
 227int
 228xfs_reflink_trim_around_shared(
 229	struct xfs_inode	*ip,
 230	struct xfs_bmbt_irec	*irec,
 231	bool			*shared)
 232{
 233	struct xfs_mount	*mp = ip->i_mount;
 234	xfs_extlen_t		shared_offset, shared_len;
 235	int			error = 0;
 236
 237	/* Holes, unwritten, and delalloc extents cannot be shared */
 238	if (!xfs_is_reflink_inode(ip) || !xfs_bmap_is_written_extent(irec)) {
 239		*shared = false;
 240		return 0;
 241	}
 242
 243	trace_xfs_reflink_trim_around_shared(ip, irec);
 244
 245	if (XFS_IS_REALTIME_INODE(ip))
 246		error = xfs_reflink_find_rtshared(mp, NULL, irec,
 247				&shared_offset, &shared_len, true);
 248	else
 249		error = xfs_reflink_find_shared(mp, NULL, irec,
 250				&shared_offset, &shared_len, true);
 251	if (error)
 252		return error;
 253
 254	if (!shared_len) {
 255		/* No shared blocks at all. */
 256		*shared = false;
 257	} else if (!shared_offset) {
 258		/*
 259		 * The start of this mapping points to shared space.  Truncate
 260		 * the mapping at the end of the shared region so that a
 261		 * subsequent iteration starts at the start of the unshared
 262		 * region.
 263		 */
 264		irec->br_blockcount = shared_len;
 265		*shared = true;
 266	} else {
 267		/*
 268		 * There's a shared region that doesn't start at the beginning
 269		 * of the mapping.  Truncate the mapping at the start of the
 270		 * shared extent so that a subsequent iteration starts at the
 271		 * start of the shared region.
 272		 */
 273		irec->br_blockcount = shared_offset;
 274		*shared = false;
 275	}
 276	return 0;
 277}
 278
 279int
 280xfs_bmap_trim_cow(
 281	struct xfs_inode	*ip,
 282	struct xfs_bmbt_irec	*imap,
 283	bool			*shared)
 284{
 285	/* We can't update any real extents in always COW mode. */
 286	if (xfs_is_always_cow_inode(ip) &&
 287	    !isnullstartblock(imap->br_startblock)) {
 288		*shared = true;
 289		return 0;
 290	}
 291
 292	/* Trim the mapping to the nearest shared extent boundary. */
 293	return xfs_reflink_trim_around_shared(ip, imap, shared);
 294}
 295
 296int
 297xfs_reflink_convert_cow_locked(
 298	struct xfs_inode	*ip,
 299	xfs_fileoff_t		offset_fsb,
 300	xfs_filblks_t		count_fsb)
 301{
 302	struct xfs_iext_cursor	icur;
 303	struct xfs_bmbt_irec	got;
 304	struct xfs_btree_cur	*dummy_cur = NULL;
 305	int			dummy_logflags;
 306	int			error = 0;
 307
 308	if (!xfs_iext_lookup_extent(ip, ip->i_cowfp, offset_fsb, &icur, &got))
 309		return 0;
 310
 311	do {
 312		if (got.br_startoff >= offset_fsb + count_fsb)
 313			break;
 314		if (got.br_state == XFS_EXT_NORM)
 315			continue;
 316		if (WARN_ON_ONCE(isnullstartblock(got.br_startblock)))
 317			return -EIO;
 318
 319		xfs_trim_extent(&got, offset_fsb, count_fsb);
 320		if (!got.br_blockcount)
 321			continue;
 322
 323		got.br_state = XFS_EXT_NORM;
 324		error = xfs_bmap_add_extent_unwritten_real(NULL, ip,
 325				XFS_COW_FORK, &icur, &dummy_cur, &got,
 326				&dummy_logflags);
 327		if (error)
 328			return error;
 329	} while (xfs_iext_next_extent(ip->i_cowfp, &icur, &got));
 330
 331	return error;
 332}
 333
 334/* Convert all of the unwritten CoW extents in a file's range to real ones. */
 335int
 336xfs_reflink_convert_cow(
 337	struct xfs_inode	*ip,
 338	xfs_off_t		offset,
 339	xfs_off_t		count)
 340{
 341	struct xfs_mount	*mp = ip->i_mount;
 342	xfs_fileoff_t		offset_fsb = XFS_B_TO_FSBT(mp, offset);
 343	xfs_fileoff_t		end_fsb = XFS_B_TO_FSB(mp, offset + count);
 344	xfs_filblks_t		count_fsb = end_fsb - offset_fsb;
 345	int			error;
 346
 347	ASSERT(count != 0);
 348
 349	xfs_ilock(ip, XFS_ILOCK_EXCL);
 350	error = xfs_reflink_convert_cow_locked(ip, offset_fsb, count_fsb);
 351	xfs_iunlock(ip, XFS_ILOCK_EXCL);
 352	return error;
 353}
 354
 355/*
 356 * Find the extent that maps the given range in the COW fork. Even if the extent
 357 * is not shared we might have a preallocation for it in the COW fork. If so we
 358 * use it that rather than trigger a new allocation.
 359 */
 360static int
 361xfs_find_trim_cow_extent(
 362	struct xfs_inode	*ip,
 363	struct xfs_bmbt_irec	*imap,
 364	struct xfs_bmbt_irec	*cmap,
 365	bool			*shared,
 366	bool			*found)
 367{
 368	xfs_fileoff_t		offset_fsb = imap->br_startoff;
 369	xfs_filblks_t		count_fsb = imap->br_blockcount;
 370	struct xfs_iext_cursor	icur;
 371
 372	*found = false;
 373
 374	/*
 375	 * If we don't find an overlapping extent, trim the range we need to
 376	 * allocate to fit the hole we found.
 377	 */
 378	if (!xfs_iext_lookup_extent(ip, ip->i_cowfp, offset_fsb, &icur, cmap))
 379		cmap->br_startoff = offset_fsb + count_fsb;
 380	if (cmap->br_startoff > offset_fsb) {
 381		xfs_trim_extent(imap, imap->br_startoff,
 382				cmap->br_startoff - imap->br_startoff);
 383		return xfs_bmap_trim_cow(ip, imap, shared);
 384	}
 385
 386	*shared = true;
 387	if (isnullstartblock(cmap->br_startblock)) {
 388		xfs_trim_extent(imap, cmap->br_startoff, cmap->br_blockcount);
 389		return 0;
 390	}
 391
 392	/* real extent found - no need to allocate */
 393	xfs_trim_extent(cmap, offset_fsb, count_fsb);
 394	*found = true;
 395	return 0;
 396}
 397
 398static int
 399xfs_reflink_convert_unwritten(
 400	struct xfs_inode	*ip,
 401	struct xfs_bmbt_irec	*imap,
 402	struct xfs_bmbt_irec	*cmap,
 403	bool			convert_now)
 404{
 405	xfs_fileoff_t		offset_fsb = imap->br_startoff;
 406	xfs_filblks_t		count_fsb = imap->br_blockcount;
 407	int			error;
 408
 409	/*
 410	 * cmap might larger than imap due to cowextsize hint.
 411	 */
 412	xfs_trim_extent(cmap, offset_fsb, count_fsb);
 413
 414	/*
 415	 * COW fork extents are supposed to remain unwritten until we're ready
 416	 * to initiate a disk write.  For direct I/O we are going to write the
 417	 * data and need the conversion, but for buffered writes we're done.
 418	 */
 419	if (!convert_now || cmap->br_state == XFS_EXT_NORM)
 420		return 0;
 421
 422	trace_xfs_reflink_convert_cow(ip, cmap);
 423
 424	error = xfs_reflink_convert_cow_locked(ip, offset_fsb, count_fsb);
 425	if (!error)
 426		cmap->br_state = XFS_EXT_NORM;
 427
 428	return error;
 429}
 430
 431static int
 432xfs_reflink_fill_cow_hole(
 433	struct xfs_inode	*ip,
 434	struct xfs_bmbt_irec	*imap,
 435	struct xfs_bmbt_irec	*cmap,
 436	bool			*shared,
 437	uint			*lockmode,
 438	bool			convert_now)
 439{
 440	struct xfs_mount	*mp = ip->i_mount;
 441	struct xfs_trans	*tp;
 442	xfs_filblks_t		resaligned;
 443	unsigned int		dblocks = 0, rblocks = 0;
 444	int			nimaps;
 445	int			error;
 446	bool			found;
 447
 448	resaligned = xfs_aligned_fsb_count(imap->br_startoff,
 449		imap->br_blockcount, xfs_get_cowextsz_hint(ip));
 450	if (XFS_IS_REALTIME_INODE(ip)) {
 451		dblocks = XFS_DIOSTRAT_SPACE_RES(mp, 0);
 452		rblocks = resaligned;
 453	} else {
 454		dblocks = XFS_DIOSTRAT_SPACE_RES(mp, resaligned);
 455		rblocks = 0;
 456	}
 457
 458	xfs_iunlock(ip, *lockmode);
 459	*lockmode = 0;
 460
 461	error = xfs_trans_alloc_inode(ip, &M_RES(mp)->tr_write, dblocks,
 462			rblocks, false, &tp);
 463	if (error)
 464		return error;
 465
 466	*lockmode = XFS_ILOCK_EXCL;
 467
 468	error = xfs_find_trim_cow_extent(ip, imap, cmap, shared, &found);
 469	if (error || !*shared)
 470		goto out_trans_cancel;
 471
 472	if (found) {
 473		xfs_trans_cancel(tp);
 474		goto convert;
 475	}
 476
 477	/* Allocate the entire reservation as unwritten blocks. */
 478	nimaps = 1;
 479	error = xfs_bmapi_write(tp, ip, imap->br_startoff, imap->br_blockcount,
 480			XFS_BMAPI_COWFORK | XFS_BMAPI_PREALLOC, 0, cmap,
 481			&nimaps);
 482	if (error)
 483		goto out_trans_cancel;
 484
 485	xfs_inode_set_cowblocks_tag(ip);
 486	error = xfs_trans_commit(tp);
 487	if (error)
 488		return error;
 489
 490convert:
 491	return xfs_reflink_convert_unwritten(ip, imap, cmap, convert_now);
 492
 493out_trans_cancel:
 494	xfs_trans_cancel(tp);
 495	return error;
 496}
 497
 498static int
 499xfs_reflink_fill_delalloc(
 500	struct xfs_inode	*ip,
 501	struct xfs_bmbt_irec	*imap,
 502	struct xfs_bmbt_irec	*cmap,
 503	bool			*shared,
 504	uint			*lockmode,
 505	bool			convert_now)
 506{
 507	struct xfs_mount	*mp = ip->i_mount;
 508	struct xfs_trans	*tp;
 509	int			nimaps;
 510	int			error;
 511	bool			found;
 512
 513	do {
 514		xfs_iunlock(ip, *lockmode);
 515		*lockmode = 0;
 516
 517		error = xfs_trans_alloc_inode(ip, &M_RES(mp)->tr_write, 0, 0,
 518				false, &tp);
 519		if (error)
 520			return error;
 521
 522		*lockmode = XFS_ILOCK_EXCL;
 523
 524		error = xfs_find_trim_cow_extent(ip, imap, cmap, shared,
 525				&found);
 526		if (error || !*shared)
 527			goto out_trans_cancel;
 528
 529		if (found) {
 530			xfs_trans_cancel(tp);
 531			break;
 532		}
 533
 534		ASSERT(isnullstartblock(cmap->br_startblock) ||
 535		       cmap->br_startblock == DELAYSTARTBLOCK);
 536
 537		/*
 538		 * Replace delalloc reservation with an unwritten extent.
 539		 */
 540		nimaps = 1;
 541		error = xfs_bmapi_write(tp, ip, cmap->br_startoff,
 542				cmap->br_blockcount,
 543				XFS_BMAPI_COWFORK | XFS_BMAPI_PREALLOC, 0,
 544				cmap, &nimaps);
 545		if (error)
 546			goto out_trans_cancel;
 547
 548		xfs_inode_set_cowblocks_tag(ip);
 549		error = xfs_trans_commit(tp);
 550		if (error)
 551			return error;
 552	} while (cmap->br_startoff + cmap->br_blockcount <= imap->br_startoff);
 553
 554	return xfs_reflink_convert_unwritten(ip, imap, cmap, convert_now);
 555
 556out_trans_cancel:
 557	xfs_trans_cancel(tp);
 558	return error;
 559}
 560
 561/* Allocate all CoW reservations covering a range of blocks in a file. */
 562int
 563xfs_reflink_allocate_cow(
 564	struct xfs_inode	*ip,
 565	struct xfs_bmbt_irec	*imap,
 566	struct xfs_bmbt_irec	*cmap,
 567	bool			*shared,
 568	uint			*lockmode,
 569	bool			convert_now)
 570{
 571	int			error;
 572	bool			found;
 573
 574	xfs_assert_ilocked(ip, XFS_ILOCK_EXCL);
 575	if (!ip->i_cowfp) {
 576		ASSERT(!xfs_is_reflink_inode(ip));
 577		xfs_ifork_init_cow(ip);
 578	}
 579
 580	error = xfs_find_trim_cow_extent(ip, imap, cmap, shared, &found);
 581	if (error || !*shared)
 582		return error;
 583
 584	/* CoW fork has a real extent */
 585	if (found)
 586		return xfs_reflink_convert_unwritten(ip, imap, cmap,
 587				convert_now);
 588
 589	/*
 590	 * CoW fork does not have an extent and data extent is shared.
 591	 * Allocate a real extent in the CoW fork.
 592	 */
 593	if (cmap->br_startoff > imap->br_startoff)
 594		return xfs_reflink_fill_cow_hole(ip, imap, cmap, shared,
 595				lockmode, convert_now);
 596
 597	/*
 598	 * CoW fork has a delalloc reservation. Replace it with a real extent.
 599	 * There may or may not be a data fork mapping.
 600	 */
 601	if (isnullstartblock(cmap->br_startblock) ||
 602	    cmap->br_startblock == DELAYSTARTBLOCK)
 603		return xfs_reflink_fill_delalloc(ip, imap, cmap, shared,
 604				lockmode, convert_now);
 605
 606	/* Shouldn't get here. */
 607	ASSERT(0);
 608	return -EFSCORRUPTED;
 609}
 610
 611/*
 612 * Cancel CoW reservations for some block range of an inode.
 613 *
 614 * If cancel_real is true this function cancels all COW fork extents for the
 615 * inode; if cancel_real is false, real extents are not cleared.
 616 *
 617 * Caller must have already joined the inode to the current transaction. The
 618 * inode will be joined to the transaction returned to the caller.
 619 */
 620int
 621xfs_reflink_cancel_cow_blocks(
 622	struct xfs_inode		*ip,
 623	struct xfs_trans		**tpp,
 624	xfs_fileoff_t			offset_fsb,
 625	xfs_fileoff_t			end_fsb,
 626	bool				cancel_real)
 627{
 628	struct xfs_ifork		*ifp = xfs_ifork_ptr(ip, XFS_COW_FORK);
 629	struct xfs_bmbt_irec		got, del;
 630	struct xfs_iext_cursor		icur;
 631	bool				isrt = XFS_IS_REALTIME_INODE(ip);
 632	int				error = 0;
 633
 634	if (!xfs_inode_has_cow_data(ip))
 635		return 0;
 636	if (!xfs_iext_lookup_extent_before(ip, ifp, &end_fsb, &icur, &got))
 637		return 0;
 638
 639	/* Walk backwards until we're out of the I/O range... */
 640	while (got.br_startoff + got.br_blockcount > offset_fsb) {
 641		del = got;
 642		xfs_trim_extent(&del, offset_fsb, end_fsb - offset_fsb);
 643
 644		/* Extent delete may have bumped ext forward */
 645		if (!del.br_blockcount) {
 646			xfs_iext_prev(ifp, &icur);
 647			goto next_extent;
 648		}
 649
 650		trace_xfs_reflink_cancel_cow(ip, &del);
 651
 652		if (isnullstartblock(del.br_startblock)) {
 653			xfs_bmap_del_extent_delay(ip, XFS_COW_FORK, &icur, &got,
 654					&del, 0);
 655		} else if (del.br_state == XFS_EXT_UNWRITTEN || cancel_real) {
 656			ASSERT((*tpp)->t_highest_agno == NULLAGNUMBER);
 657
 658			/* Free the CoW orphan record. */
 659			xfs_refcount_free_cow_extent(*tpp, isrt,
 660					del.br_startblock, del.br_blockcount);
 661
 662			error = xfs_free_extent_later(*tpp, del.br_startblock,
 663					del.br_blockcount, NULL,
 664					XFS_AG_RESV_NONE,
 665					isrt ? XFS_FREE_EXTENT_REALTIME : 0);
 666			if (error)
 667				break;
 668
 669			/* Roll the transaction */
 670			error = xfs_defer_finish(tpp);
 671			if (error)
 672				break;
 673
 674			/* Remove the mapping from the CoW fork. */
 675			xfs_bmap_del_extent_cow(ip, &icur, &got, &del);
 676
 677			/* Remove the quota reservation */
 678			xfs_quota_unreserve_blkres(ip, del.br_blockcount);
 679		} else {
 680			/* Didn't do anything, push cursor back. */
 681			xfs_iext_prev(ifp, &icur);
 682		}
 683next_extent:
 684		if (!xfs_iext_get_extent(ifp, &icur, &got))
 685			break;
 686	}
 687
 688	/* clear tag if cow fork is emptied */
 689	if (!ifp->if_bytes)
 690		xfs_inode_clear_cowblocks_tag(ip);
 691	return error;
 692}
 693
 694/*
 695 * Cancel CoW reservations for some byte range of an inode.
 696 *
 697 * If cancel_real is true this function cancels all COW fork extents for the
 698 * inode; if cancel_real is false, real extents are not cleared.
 699 */
 700int
 701xfs_reflink_cancel_cow_range(
 702	struct xfs_inode	*ip,
 703	xfs_off_t		offset,
 704	xfs_off_t		count,
 705	bool			cancel_real)
 706{
 707	struct xfs_trans	*tp;
 708	xfs_fileoff_t		offset_fsb;
 709	xfs_fileoff_t		end_fsb;
 710	int			error;
 711
 712	trace_xfs_reflink_cancel_cow_range(ip, offset, count);
 713	ASSERT(ip->i_cowfp);
 714
 715	offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
 716	if (count == NULLFILEOFF)
 717		end_fsb = NULLFILEOFF;
 718	else
 719		end_fsb = XFS_B_TO_FSB(ip->i_mount, offset + count);
 720
 721	/* Start a rolling transaction to remove the mappings */
 722	error = xfs_trans_alloc(ip->i_mount, &M_RES(ip->i_mount)->tr_write,
 723			0, 0, 0, &tp);
 724	if (error)
 725		goto out;
 726
 727	xfs_ilock(ip, XFS_ILOCK_EXCL);
 728	xfs_trans_ijoin(tp, ip, 0);
 729
 730	/* Scrape out the old CoW reservations */
 731	error = xfs_reflink_cancel_cow_blocks(ip, &tp, offset_fsb, end_fsb,
 732			cancel_real);
 733	if (error)
 734		goto out_cancel;
 735
 736	error = xfs_trans_commit(tp);
 737
 738	xfs_iunlock(ip, XFS_ILOCK_EXCL);
 739	return error;
 740
 741out_cancel:
 742	xfs_trans_cancel(tp);
 743	xfs_iunlock(ip, XFS_ILOCK_EXCL);
 744out:
 745	trace_xfs_reflink_cancel_cow_range_error(ip, error, _RET_IP_);
 746	return error;
 747}
 748
 749#ifdef CONFIG_XFS_QUOTA
 750/*
 751 * Update quota accounting for a remapping operation.  When we're remapping
 752 * something from the CoW fork to the data fork, we must update the quota
 753 * accounting for delayed allocations.  For remapping from the data fork to the
 754 * data fork, use regular block accounting.
 755 */
 756static inline void
 757xfs_reflink_update_quota(
 758	struct xfs_trans	*tp,
 759	struct xfs_inode	*ip,
 760	bool			is_cow,
 761	int64_t			blocks)
 762{
 763	unsigned int		qflag;
 764
 765	if (XFS_IS_REALTIME_INODE(ip)) {
 766		qflag = is_cow ? XFS_TRANS_DQ_DELRTBCOUNT :
 767				 XFS_TRANS_DQ_RTBCOUNT;
 768	} else {
 769		qflag = is_cow ? XFS_TRANS_DQ_DELBCOUNT :
 770				 XFS_TRANS_DQ_BCOUNT;
 771	}
 772	xfs_trans_mod_dquot_byino(tp, ip, qflag, blocks);
 773}
 774#else
 775# define xfs_reflink_update_quota(tp, ip, is_cow, blocks)	((void)0)
 776#endif
 777
 778/*
 779 * Remap part of the CoW fork into the data fork.
 780 *
 781 * We aim to remap the range starting at @offset_fsb and ending at @end_fsb
 782 * into the data fork; this function will remap what it can (at the end of the
 783 * range) and update @end_fsb appropriately.  Each remap gets its own
 784 * transaction because we can end up merging and splitting bmbt blocks for
 785 * every remap operation and we'd like to keep the block reservation
 786 * requirements as low as possible.
 787 */
 788STATIC int
 789xfs_reflink_end_cow_extent_locked(
 790	struct xfs_trans	*tp,
 791	struct xfs_inode	*ip,
 792	xfs_fileoff_t		*offset_fsb,
 793	xfs_fileoff_t		end_fsb)
 794{
 795	struct xfs_iext_cursor	icur;
 796	struct xfs_bmbt_irec	got, del, data;
 797	struct xfs_ifork	*ifp = xfs_ifork_ptr(ip, XFS_COW_FORK);
 798	int			nmaps;
 799	bool			isrt = XFS_IS_REALTIME_INODE(ip);
 800	int			error;
 801
 802	/*
 803	 * In case of racing, overlapping AIO writes no COW extents might be
 804	 * left by the time I/O completes for the loser of the race.  In that
 805	 * case we are done.
 806	 */
 807	if (!xfs_iext_lookup_extent(ip, ifp, *offset_fsb, &icur, &got) ||
 808	    got.br_startoff >= end_fsb) {
 809		*offset_fsb = end_fsb;
 810		return 0;
 811	}
 812
 813	/*
 814	 * Only remap real extents that contain data.  With AIO, speculative
 815	 * preallocations can leak into the range we are called upon, and we
 816	 * need to skip them.  Preserve @got for the eventual CoW fork
 817	 * deletion; from now on @del represents the mapping that we're
 818	 * actually remapping.
 819	 */
 820	while (!xfs_bmap_is_written_extent(&got)) {
 821		if (!xfs_iext_next_extent(ifp, &icur, &got) ||
 822		    got.br_startoff >= end_fsb) {
 823			*offset_fsb = end_fsb;
 824			return 0;
 825		}
 826	}
 827	del = got;
 828	xfs_trim_extent(&del, *offset_fsb, end_fsb - *offset_fsb);
 829
 830	error = xfs_iext_count_extend(tp, ip, XFS_DATA_FORK,
 831			XFS_IEXT_REFLINK_END_COW_CNT);
 832	if (error)
 833		return error;
 834
 835	/* Grab the corresponding mapping in the data fork. */
 836	nmaps = 1;
 837	error = xfs_bmapi_read(ip, del.br_startoff, del.br_blockcount, &data,
 838			&nmaps, 0);
 839	if (error)
 840		return error;
 841
 842	/* We can only remap the smaller of the two extent sizes. */
 843	data.br_blockcount = min(data.br_blockcount, del.br_blockcount);
 844	del.br_blockcount = data.br_blockcount;
 845
 846	trace_xfs_reflink_cow_remap_from(ip, &del);
 847	trace_xfs_reflink_cow_remap_to(ip, &data);
 848
 849	if (xfs_bmap_is_real_extent(&data)) {
 850		/*
 851		 * If the extent we're remapping is backed by storage (written
 852		 * or not), unmap the extent and drop its refcount.
 853		 */
 854		xfs_bmap_unmap_extent(tp, ip, XFS_DATA_FORK, &data);
 855		xfs_refcount_decrease_extent(tp, isrt, &data);
 856		xfs_reflink_update_quota(tp, ip, false, -data.br_blockcount);
 857	} else if (data.br_startblock == DELAYSTARTBLOCK) {
 858		int		done;
 859
 860		/*
 861		 * If the extent we're remapping is a delalloc reservation,
 862		 * we can use the regular bunmapi function to release the
 863		 * incore state.  Dropping the delalloc reservation takes care
 864		 * of the quota reservation for us.
 865		 */
 866		error = xfs_bunmapi(NULL, ip, data.br_startoff,
 867				data.br_blockcount, 0, 1, &done);
 868		if (error)
 869			return error;
 870		ASSERT(done);
 871	}
 872
 873	/* Free the CoW orphan record. */
 874	xfs_refcount_free_cow_extent(tp, isrt, del.br_startblock,
 875			del.br_blockcount);
 876
 877	/* Map the new blocks into the data fork. */
 878	xfs_bmap_map_extent(tp, ip, XFS_DATA_FORK, &del);
 879
 880	/* Charge this new data fork mapping to the on-disk quota. */
 881	xfs_reflink_update_quota(tp, ip, true, del.br_blockcount);
 882
 883	/* Remove the mapping from the CoW fork. */
 884	xfs_bmap_del_extent_cow(ip, &icur, &got, &del);
 885
 886	/* Update the caller about how much progress we made. */
 887	*offset_fsb = del.br_startoff + del.br_blockcount;
 888	return 0;
 889}
 890
 891/*
 892 * Remap part of the CoW fork into the data fork.
 893 *
 894 * We aim to remap the range starting at @offset_fsb and ending at @end_fsb
 895 * into the data fork; this function will remap what it can (at the end of the
 896 * range) and update @end_fsb appropriately.  Each remap gets its own
 897 * transaction because we can end up merging and splitting bmbt blocks for
 898 * every remap operation and we'd like to keep the block reservation
 899 * requirements as low as possible.
 900 */
 901STATIC int
 902xfs_reflink_end_cow_extent(
 903	struct xfs_inode	*ip,
 904	xfs_fileoff_t		*offset_fsb,
 905	xfs_fileoff_t		end_fsb)
 906{
 907	struct xfs_mount	*mp = ip->i_mount;
 908	struct xfs_trans	*tp;
 909	unsigned int		resblks;
 910	int			error;
 911
 912	resblks = XFS_EXTENTADD_SPACE_RES(mp, XFS_DATA_FORK);
 913	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, resblks, 0,
 914			XFS_TRANS_RESERVE, &tp);
 915	if (error)
 916		return error;
 917	xfs_ilock(ip, XFS_ILOCK_EXCL);
 918	xfs_trans_ijoin(tp, ip, 0);
 919
 920	error = xfs_reflink_end_cow_extent_locked(tp, ip, offset_fsb, end_fsb);
 921	if (error)
 922		xfs_trans_cancel(tp);
 923	else
 924		error = xfs_trans_commit(tp);
 925	xfs_iunlock(ip, XFS_ILOCK_EXCL);
 926	return error;
 927}
 928
 929/*
 930 * Remap parts of a file's data fork after a successful CoW.
 931 */
 932int
 933xfs_reflink_end_cow(
 934	struct xfs_inode		*ip,
 935	xfs_off_t			offset,
 936	xfs_off_t			count)
 937{
 938	xfs_fileoff_t			offset_fsb;
 939	xfs_fileoff_t			end_fsb;
 940	int				error = 0;
 941
 942	trace_xfs_reflink_end_cow(ip, offset, count);
 943
 944	offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
 945	end_fsb = XFS_B_TO_FSB(ip->i_mount, offset + count);
 946
 947	/*
 948	 * Walk forwards until we've remapped the I/O range.  The loop function
 949	 * repeatedly cycles the ILOCK to allocate one transaction per remapped
 950	 * extent.
 951	 *
 952	 * If we're being called by writeback then the pages will still
 953	 * have PageWriteback set, which prevents races with reflink remapping
 954	 * and truncate.  Reflink remapping prevents races with writeback by
 955	 * taking the iolock and mmaplock before flushing the pages and
 956	 * remapping, which means there won't be any further writeback or page
 957	 * cache dirtying until the reflink completes.
 958	 *
 959	 * We should never have two threads issuing writeback for the same file
 960	 * region.  There are also have post-eof checks in the writeback
 961	 * preparation code so that we don't bother writing out pages that are
 962	 * about to be truncated.
 963	 *
 964	 * If we're being called as part of directio write completion, the dio
 965	 * count is still elevated, which reflink and truncate will wait for.
 966	 * Reflink remapping takes the iolock and mmaplock and waits for
 967	 * pending dio to finish, which should prevent any directio until the
 968	 * remap completes.  Multiple concurrent directio writes to the same
 969	 * region are handled by end_cow processing only occurring for the
 970	 * threads which succeed; the outcome of multiple overlapping direct
 971	 * writes is not well defined anyway.
 972	 *
 973	 * It's possible that a buffered write and a direct write could collide
 974	 * here (the buffered write stumbles in after the dio flushes and
 975	 * invalidates the page cache and immediately queues writeback), but we
 976	 * have never supported this 100%.  If either disk write succeeds the
 977	 * blocks will be remapped.
 978	 */
 979	while (end_fsb > offset_fsb && !error)
 980		error = xfs_reflink_end_cow_extent(ip, &offset_fsb, end_fsb);
 981
 982	if (error)
 983		trace_xfs_reflink_end_cow_error(ip, error, _RET_IP_);
 984	return error;
 985}
 986
 987/*
 988 * Fully remap all of the file's data fork at once, which is the critical part
 989 * in achieving atomic behaviour.
 990 * The regular CoW end path does not use function as to keep the block
 991 * reservation per transaction as low as possible.
 992 */
 993int
 994xfs_reflink_end_atomic_cow(
 995	struct xfs_inode		*ip,
 996	xfs_off_t			offset,
 997	xfs_off_t			count)
 998{
 999	xfs_fileoff_t			offset_fsb;
1000	xfs_fileoff_t			end_fsb;
1001	int				error = 0;
1002	struct xfs_mount		*mp = ip->i_mount;
1003	struct xfs_trans		*tp;
1004	unsigned int			resblks;
1005
1006	trace_xfs_reflink_end_cow(ip, offset, count);
1007
1008	offset_fsb = XFS_B_TO_FSBT(mp, offset);
1009	end_fsb = XFS_B_TO_FSB(mp, offset + count);
1010
1011	/*
1012	 * Each remapping operation could cause a btree split, so in the worst
1013	 * case that's one for each block.
1014	 */
1015	resblks = (end_fsb - offset_fsb) *
1016			XFS_NEXTENTADD_SPACE_RES(mp, 1, XFS_DATA_FORK);
1017
1018	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_atomic_ioend, resblks, 0,
1019			XFS_TRANS_RESERVE, &tp);
1020	if (error)
1021		return error;
1022
1023	xfs_ilock(ip, XFS_ILOCK_EXCL);
1024	xfs_trans_ijoin(tp, ip, 0);
1025
1026	while (end_fsb > offset_fsb && !error) {
1027		error = xfs_reflink_end_cow_extent_locked(tp, ip, &offset_fsb,
1028				end_fsb);
1029	}
1030	if (error) {
1031		trace_xfs_reflink_end_cow_error(ip, error, _RET_IP_);
1032		goto out_cancel;
1033	}
1034	error = xfs_trans_commit(tp);
1035	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1036	return error;
1037out_cancel:
1038	xfs_trans_cancel(tp);
1039	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1040	return error;
1041}
1042
1043/* Compute the largest atomic write that we can complete through software. */
1044xfs_extlen_t
1045xfs_reflink_max_atomic_cow(
1046	struct xfs_mount	*mp)
1047{
1048	/* We cannot do any atomic writes without out of place writes. */
1049	if (!xfs_can_sw_atomic_write(mp))
1050		return 0;
1051
1052	/*
1053	 * Atomic write limits must always be a power-of-2, according to
1054	 * generic_atomic_write_valid.
1055	 */
1056	return rounddown_pow_of_two(xfs_calc_max_atomic_write_fsblocks(mp));
1057}
1058
1059/*
1060 * Free all CoW staging blocks that are still referenced by the ondisk refcount
1061 * metadata.  The ondisk metadata does not track which inode created the
1062 * staging extent, so callers must ensure that there are no cached inodes with
1063 * live CoW staging extents.
1064 */
1065int
1066xfs_reflink_recover_cow(
1067	struct xfs_mount	*mp)
1068{
1069	struct xfs_perag	*pag = NULL;
1070	struct xfs_rtgroup	*rtg = NULL;
1071	int			error = 0;
1072
1073	if (!xfs_has_reflink(mp))
1074		return 0;
1075
1076	while ((pag = xfs_perag_next(mp, pag))) {
1077		error = xfs_refcount_recover_cow_leftovers(pag_group(pag));
1078		if (error) {
1079			xfs_perag_rele(pag);
1080			return error;
1081		}
1082	}
1083
1084	while ((rtg = xfs_rtgroup_next(mp, rtg))) {
1085		error = xfs_refcount_recover_cow_leftovers(rtg_group(rtg));
1086		if (error) {
1087			xfs_rtgroup_rele(rtg);
1088			return error;
1089		}
1090	}
1091
1092	return 0;
1093}
1094
1095/*
1096 * Reflinking (Block) Ranges of Two Files Together
1097 *
1098 * First, ensure that the reflink flag is set on both inodes.  The flag is an
1099 * optimization to avoid unnecessary refcount btree lookups in the write path.
1100 *
1101 * Now we can iteratively remap the range of extents (and holes) in src to the
1102 * corresponding ranges in dest.  Let drange and srange denote the ranges of
1103 * logical blocks in dest and src touched by the reflink operation.
1104 *
1105 * While the length of drange is greater than zero,
1106 *    - Read src's bmbt at the start of srange ("imap")
1107 *    - If imap doesn't exist, make imap appear to start at the end of srange
1108 *      with zero length.
1109 *    - If imap starts before srange, advance imap to start at srange.
1110 *    - If imap goes beyond srange, truncate imap to end at the end of srange.
1111 *    - Punch (imap start - srange start + imap len) blocks from dest at
1112 *      offset (drange start).
1113 *    - If imap points to a real range of pblks,
1114 *         > Increase the refcount of the imap's pblks
1115 *         > Map imap's pblks into dest at the offset
1116 *           (drange start + imap start - srange start)
1117 *    - Advance drange and srange by (imap start - srange start + imap len)
1118 *
1119 * Finally, if the reflink made dest longer, update both the in-core and
1120 * on-disk file sizes.
1121 *
1122 * ASCII Art Demonstration:
1123 *
1124 * Let's say we want to reflink this source file:
1125 *
1126 * ----SSSSSSS-SSSSS----SSSSSS (src file)
1127 *   <-------------------->
1128 *
1129 * into this destination file:
1130 *
1131 * --DDDDDDDDDDDDDDDDDDD--DDD (dest file)
1132 *        <-------------------->
1133 * '-' means a hole, and 'S' and 'D' are written blocks in the src and dest.
1134 * Observe that the range has different logical offsets in either file.
1135 *
1136 * Consider that the first extent in the source file doesn't line up with our
1137 * reflink range.  Unmapping  and remapping are separate operations, so we can
1138 * unmap more blocks from the destination file than we remap.
1139 *
1140 * ----SSSSSSS-SSSSS----SSSSSS
1141 *   <------->
1142 * --DDDDD---------DDDDD--DDD
1143 *        <------->
1144 *
1145 * Now remap the source extent into the destination file:
1146 *
1147 * ----SSSSSSS-SSSSS----SSSSSS
1148 *   <------->
1149 * --DDDDD--SSSSSSSDDDDD--DDD
1150 *        <------->
1151 *
1152 * Do likewise with the second hole and extent in our range.  Holes in the
1153 * unmap range don't affect our operation.
1154 *
1155 * ----SSSSSSS-SSSSS----SSSSSS
1156 *            <---->
1157 * --DDDDD--SSSSSSS-SSSSS-DDD
1158 *                 <---->
1159 *
1160 * Finally, unmap and remap part of the third extent.  This will increase the
1161 * size of the destination file.
1162 *
1163 * ----SSSSSSS-SSSSS----SSSSSS
1164 *                  <----->
1165 * --DDDDD--SSSSSSS-SSSSS----SSS
1166 *                       <----->
1167 *
1168 * Once we update the destination file's i_size, we're done.
1169 */
1170
1171/*
1172 * Ensure the reflink bit is set in both inodes.
1173 */
1174STATIC int
1175xfs_reflink_set_inode_flag(
1176	struct xfs_inode	*src,
1177	struct xfs_inode	*dest)
1178{
1179	struct xfs_mount	*mp = src->i_mount;
1180	int			error;
1181	struct xfs_trans	*tp;
1182
1183	if (xfs_is_reflink_inode(src) && xfs_is_reflink_inode(dest))
1184		return 0;
1185
1186	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
1187	if (error)
1188		goto out_error;
1189
1190	/* Lock both files against IO */
1191	if (src->i_ino == dest->i_ino)
1192		xfs_ilock(src, XFS_ILOCK_EXCL);
1193	else
1194		xfs_lock_two_inodes(src, XFS_ILOCK_EXCL, dest, XFS_ILOCK_EXCL);
1195
1196	if (!xfs_is_reflink_inode(src)) {
1197		trace_xfs_reflink_set_inode_flag(src);
1198		xfs_trans_ijoin(tp, src, XFS_ILOCK_EXCL);
1199		src->i_diflags2 |= XFS_DIFLAG2_REFLINK;
1200		xfs_trans_log_inode(tp, src, XFS_ILOG_CORE);
1201		xfs_ifork_init_cow(src);
1202	} else
1203		xfs_iunlock(src, XFS_ILOCK_EXCL);
1204
1205	if (src->i_ino == dest->i_ino)
1206		goto commit_flags;
1207
1208	if (!xfs_is_reflink_inode(dest)) {
1209		trace_xfs_reflink_set_inode_flag(dest);
1210		xfs_trans_ijoin(tp, dest, XFS_ILOCK_EXCL);
1211		dest->i_diflags2 |= XFS_DIFLAG2_REFLINK;
1212		xfs_trans_log_inode(tp, dest, XFS_ILOG_CORE);
1213		xfs_ifork_init_cow(dest);
1214	} else
1215		xfs_iunlock(dest, XFS_ILOCK_EXCL);
1216
1217commit_flags:
1218	error = xfs_trans_commit(tp);
1219	if (error)
1220		goto out_error;
1221	return error;
1222
1223out_error:
1224	trace_xfs_reflink_set_inode_flag_error(dest, error, _RET_IP_);
1225	return error;
1226}
1227
1228/*
1229 * Update destination inode size & cowextsize hint, if necessary.
1230 */
1231int
1232xfs_reflink_update_dest(
1233	struct xfs_inode	*dest,
1234	xfs_off_t		newlen,
1235	xfs_extlen_t		cowextsize,
1236	unsigned int		remap_flags)
1237{
1238	struct xfs_mount	*mp = dest->i_mount;
1239	struct xfs_trans	*tp;
1240	int			error;
1241
1242	if (newlen <= i_size_read(VFS_I(dest)) && cowextsize == 0)
1243		return 0;
1244
1245	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
1246	if (error)
1247		goto out_error;
1248
1249	xfs_ilock(dest, XFS_ILOCK_EXCL);
1250	xfs_trans_ijoin(tp, dest, XFS_ILOCK_EXCL);
1251
1252	if (newlen > i_size_read(VFS_I(dest))) {
1253		trace_xfs_reflink_update_inode_size(dest, newlen);
1254		i_size_write(VFS_I(dest), newlen);
1255		dest->i_disk_size = newlen;
1256	}
1257
1258	if (cowextsize) {
1259		dest->i_cowextsize = cowextsize;
1260		dest->i_diflags2 |= XFS_DIFLAG2_COWEXTSIZE;
1261	}
1262
1263	xfs_trans_log_inode(tp, dest, XFS_ILOG_CORE);
1264
1265	error = xfs_trans_commit(tp);
1266	if (error)
1267		goto out_error;
1268	return error;
1269
1270out_error:
1271	trace_xfs_reflink_update_inode_size_error(dest, error, _RET_IP_);
1272	return error;
1273}
1274
1275/*
1276 * Do we have enough reserve in this AG to handle a reflink?  The refcount
1277 * btree already reserved all the space it needs, but the rmap btree can grow
1278 * infinitely, so we won't allow more reflinks when the AG is down to the
1279 * btree reserves.
1280 */
1281static int
1282xfs_reflink_ag_has_free_space(
1283	struct xfs_mount	*mp,
1284	struct xfs_inode	*ip,
1285	xfs_fsblock_t		fsb)
1286{
1287	struct xfs_perag	*pag;
1288	xfs_agnumber_t		agno;
1289	int			error = 0;
1290
1291	if (!xfs_has_rmapbt(mp))
1292		return 0;
1293	if (XFS_IS_REALTIME_INODE(ip)) {
1294		if (xfs_metafile_resv_critical(mp))
1295			return -ENOSPC;
1296		return 0;
1297	}
1298
1299	agno = XFS_FSB_TO_AGNO(mp, fsb);
1300	pag = xfs_perag_get(mp, agno);
1301	if (xfs_ag_resv_critical(pag, XFS_AG_RESV_RMAPBT) ||
1302	    xfs_ag_resv_critical(pag, XFS_AG_RESV_METADATA))
1303		error = -ENOSPC;
1304	xfs_perag_put(pag);
1305	return error;
1306}
1307
1308/*
1309 * Remap the given extent into the file.  The dmap blockcount will be set to
1310 * the number of blocks that were actually remapped.
1311 */
1312STATIC int
1313xfs_reflink_remap_extent(
1314	struct xfs_inode	*ip,
1315	struct xfs_bmbt_irec	*dmap,
1316	xfs_off_t		new_isize)
1317{
1318	struct xfs_bmbt_irec	smap;
1319	struct xfs_mount	*mp = ip->i_mount;
1320	struct xfs_trans	*tp;
1321	xfs_off_t		newlen;
1322	int64_t			qdelta = 0;
1323	unsigned int		dblocks, rblocks, resblks;
1324	bool			quota_reserved = true;
1325	bool			smap_real;
1326	bool			dmap_written = xfs_bmap_is_written_extent(dmap);
1327	bool			isrt = XFS_IS_REALTIME_INODE(ip);
1328	int			iext_delta = 0;
1329	int			nimaps;
1330	int			error;
1331
1332	/*
1333	 * Start a rolling transaction to switch the mappings.
1334	 *
1335	 * Adding a written extent to the extent map can cause a bmbt split,
1336	 * and removing a mapped extent from the extent can cause a bmbt split.
1337	 * The two operations cannot both cause a split since they operate on
1338	 * the same index in the bmap btree, so we only need a reservation for
1339	 * one bmbt split if either thing is happening.  However, we haven't
1340	 * locked the inode yet, so we reserve assuming this is the case.
1341	 *
1342	 * The first allocation call tries to reserve enough space to handle
1343	 * mapping dmap into a sparse part of the file plus the bmbt split.  We
1344	 * haven't locked the inode or read the existing mapping yet, so we do
1345	 * not know for sure that we need the space.  This should succeed most
1346	 * of the time.
1347	 *
1348	 * If the first attempt fails, try again but reserving only enough
1349	 * space to handle a bmbt split.  This is the hard minimum requirement,
1350	 * and we revisit quota reservations later when we know more about what
1351	 * we're remapping.
1352	 */
1353	resblks = XFS_EXTENTADD_SPACE_RES(mp, XFS_DATA_FORK);
1354	if (XFS_IS_REALTIME_INODE(ip)) {
1355		dblocks = resblks;
1356		rblocks = dmap->br_blockcount;
1357	} else {
1358		dblocks = resblks + dmap->br_blockcount;
1359		rblocks = 0;
1360	}
1361	error = xfs_trans_alloc_inode(ip, &M_RES(mp)->tr_write,
1362			dblocks, rblocks, false, &tp);
1363	if (error == -EDQUOT || error == -ENOSPC) {
1364		quota_reserved = false;
1365		error = xfs_trans_alloc_inode(ip, &M_RES(mp)->tr_write,
1366				resblks, 0, false, &tp);
1367	}
1368	if (error)
1369		goto out;
1370
1371	/*
1372	 * Read what's currently mapped in the destination file into smap.
1373	 * If smap isn't a hole, we will have to remove it before we can add
1374	 * dmap to the destination file.
1375	 */
1376	nimaps = 1;
1377	error = xfs_bmapi_read(ip, dmap->br_startoff, dmap->br_blockcount,
1378			&smap, &nimaps, 0);
1379	if (error)
1380		goto out_cancel;
1381	ASSERT(nimaps == 1 && smap.br_startoff == dmap->br_startoff);
1382	smap_real = xfs_bmap_is_real_extent(&smap);
1383
1384	/*
1385	 * We can only remap as many blocks as the smaller of the two extent
1386	 * maps, because we can only remap one extent at a time.
1387	 */
1388	dmap->br_blockcount = min(dmap->br_blockcount, smap.br_blockcount);
1389	ASSERT(dmap->br_blockcount == smap.br_blockcount);
1390
1391	trace_xfs_reflink_remap_extent_dest(ip, &smap);
1392
1393	/*
1394	 * Two extents mapped to the same physical block must not have
1395	 * different states; that's filesystem corruption.  Move on to the next
1396	 * extent if they're both holes or both the same physical extent.
1397	 */
1398	if (dmap->br_startblock == smap.br_startblock) {
1399		if (dmap->br_state != smap.br_state) {
1400			xfs_bmap_mark_sick(ip, XFS_DATA_FORK);
1401			error = -EFSCORRUPTED;
1402		}
1403		goto out_cancel;
1404	}
1405
1406	/* If both extents are unwritten, leave them alone. */
1407	if (dmap->br_state == XFS_EXT_UNWRITTEN &&
1408	    smap.br_state == XFS_EXT_UNWRITTEN)
1409		goto out_cancel;
1410
1411	/* No reflinking if the AG of the dest mapping is low on space. */
1412	if (dmap_written) {
1413		error = xfs_reflink_ag_has_free_space(mp, ip,
1414				dmap->br_startblock);
1415		if (error)
1416			goto out_cancel;
1417	}
1418
1419	/*
1420	 * Increase quota reservation if we think the quota block counter for
1421	 * this file could increase.
1422	 *
1423	 * If we are mapping a written extent into the file, we need to have
1424	 * enough quota block count reservation to handle the blocks in that
1425	 * extent.  We log only the delta to the quota block counts, so if the
1426	 * extent we're unmapping also has blocks allocated to it, we don't
1427	 * need a quota reservation for the extent itself.
1428	 *
1429	 * Note that if we're replacing a delalloc reservation with a written
1430	 * extent, we have to take the full quota reservation because removing
1431	 * the delalloc reservation gives the block count back to the quota
1432	 * count.  This is suboptimal, but the VFS flushed the dest range
1433	 * before we started.  That should have removed all the delalloc
1434	 * reservations, but we code defensively.
1435	 *
1436	 * xfs_trans_alloc_inode above already tried to grab an even larger
1437	 * quota reservation, and kicked off a blockgc scan if it couldn't.
1438	 * If we can't get a potentially smaller quota reservation now, we're
1439	 * done.
1440	 */
1441	if (!quota_reserved && !smap_real && dmap_written) {
1442		if (XFS_IS_REALTIME_INODE(ip)) {
1443			dblocks = 0;
1444			rblocks = dmap->br_blockcount;
1445		} else {
1446			dblocks = dmap->br_blockcount;
1447			rblocks = 0;
1448		}
1449		error = xfs_trans_reserve_quota_nblks(tp, ip, dblocks, rblocks,
1450				false);
1451		if (error)
1452			goto out_cancel;
1453	}
1454
1455	if (smap_real)
1456		++iext_delta;
1457
1458	if (dmap_written)
1459		++iext_delta;
1460
1461	error = xfs_iext_count_extend(tp, ip, XFS_DATA_FORK, iext_delta);
1462	if (error)
1463		goto out_cancel;
1464
1465	if (smap_real) {
1466		/*
1467		 * If the extent we're unmapping is backed by storage (written
1468		 * or not), unmap the extent and drop its refcount.
1469		 */
1470		xfs_bmap_unmap_extent(tp, ip, XFS_DATA_FORK, &smap);
1471		xfs_refcount_decrease_extent(tp, isrt, &smap);
1472		qdelta -= smap.br_blockcount;
1473	} else if (smap.br_startblock == DELAYSTARTBLOCK) {
1474		int		done;
1475
1476		/*
1477		 * If the extent we're unmapping is a delalloc reservation,
1478		 * we can use the regular bunmapi function to release the
1479		 * incore state.  Dropping the delalloc reservation takes care
1480		 * of the quota reservation for us.
1481		 */
1482		error = xfs_bunmapi(NULL, ip, smap.br_startoff,
1483				smap.br_blockcount, 0, 1, &done);
1484		if (error)
1485			goto out_cancel;
1486		ASSERT(done);
1487	}
1488
1489	/*
1490	 * If the extent we're sharing is backed by written storage, increase
1491	 * its refcount and map it into the file.
1492	 */
1493	if (dmap_written) {
1494		xfs_refcount_increase_extent(tp, isrt, dmap);
1495		xfs_bmap_map_extent(tp, ip, XFS_DATA_FORK, dmap);
1496		qdelta += dmap->br_blockcount;
1497	}
1498
1499	xfs_reflink_update_quota(tp, ip, false, qdelta);
1500
1501	/* Update dest isize if needed. */
1502	newlen = XFS_FSB_TO_B(mp, dmap->br_startoff + dmap->br_blockcount);
1503	newlen = min_t(xfs_off_t, newlen, new_isize);
1504	if (newlen > i_size_read(VFS_I(ip))) {
1505		trace_xfs_reflink_update_inode_size(ip, newlen);
1506		i_size_write(VFS_I(ip), newlen);
1507		ip->i_disk_size = newlen;
1508		xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1509	}
1510
1511	/* Commit everything and unlock. */
1512	error = xfs_trans_commit(tp);
1513	goto out_unlock;
1514
1515out_cancel:
1516	xfs_trans_cancel(tp);
1517out_unlock:
1518	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1519out:
1520	if (error)
1521		trace_xfs_reflink_remap_extent_error(ip, error, _RET_IP_);
1522	return error;
1523}
1524
1525/* Remap a range of one file to the other. */
1526int
1527xfs_reflink_remap_blocks(
1528	struct xfs_inode	*src,
1529	loff_t			pos_in,
1530	struct xfs_inode	*dest,
1531	loff_t			pos_out,
1532	loff_t			remap_len,
1533	loff_t			*remapped)
1534{
1535	struct xfs_bmbt_irec	imap;
1536	struct xfs_mount	*mp = src->i_mount;
1537	xfs_fileoff_t		srcoff = XFS_B_TO_FSBT(mp, pos_in);
1538	xfs_fileoff_t		destoff = XFS_B_TO_FSBT(mp, pos_out);
1539	xfs_filblks_t		len;
1540	xfs_filblks_t		remapped_len = 0;
1541	xfs_off_t		new_isize = pos_out + remap_len;
1542	int			nimaps;
1543	int			error = 0;
1544
1545	len = min_t(xfs_filblks_t, XFS_B_TO_FSB(mp, remap_len),
1546			XFS_MAX_FILEOFF);
1547
1548	trace_xfs_reflink_remap_blocks(src, srcoff, len, dest, destoff);
1549
1550	while (len > 0) {
1551		unsigned int	lock_mode;
1552
1553		/* Read extent from the source file */
1554		nimaps = 1;
1555		lock_mode = xfs_ilock_data_map_shared(src);
1556		error = xfs_bmapi_read(src, srcoff, len, &imap, &nimaps, 0);
1557		xfs_iunlock(src, lock_mode);
1558		if (error)
1559			break;
1560		/*
1561		 * The caller supposedly flushed all dirty pages in the source
1562		 * file range, which means that writeback should have allocated
1563		 * or deleted all delalloc reservations in that range.  If we
1564		 * find one, that's a good sign that something is seriously
1565		 * wrong here.
1566		 */
1567		ASSERT(nimaps == 1 && imap.br_startoff == srcoff);
1568		if (imap.br_startblock == DELAYSTARTBLOCK) {
1569			ASSERT(imap.br_startblock != DELAYSTARTBLOCK);
1570			xfs_bmap_mark_sick(src, XFS_DATA_FORK);
1571			error = -EFSCORRUPTED;
1572			break;
1573		}
1574
1575		trace_xfs_reflink_remap_extent_src(src, &imap);
1576
1577		/* Remap into the destination file at the given offset. */
1578		imap.br_startoff = destoff;
1579		error = xfs_reflink_remap_extent(dest, &imap, new_isize);
1580		if (error)
1581			break;
1582
1583		if (fatal_signal_pending(current)) {
1584			error = -EINTR;
1585			break;
1586		}
1587
1588		/* Advance drange/srange */
1589		srcoff += imap.br_blockcount;
1590		destoff += imap.br_blockcount;
1591		len -= imap.br_blockcount;
1592		remapped_len += imap.br_blockcount;
1593		cond_resched();
1594	}
1595
1596	if (error)
1597		trace_xfs_reflink_remap_blocks_error(dest, error, _RET_IP_);
1598	*remapped = min_t(loff_t, remap_len,
1599			  XFS_FSB_TO_B(src->i_mount, remapped_len));
1600	return error;
1601}
1602
1603/*
1604 * If we're reflinking to a point past the destination file's EOF, we must
1605 * zero any speculative post-EOF preallocations that sit between the old EOF
1606 * and the destination file offset.
1607 */
1608static int
1609xfs_reflink_zero_posteof(
1610	struct xfs_inode	*ip,
1611	loff_t			pos)
1612{
1613	loff_t			isize = i_size_read(VFS_I(ip));
1614
1615	if (pos <= isize)
1616		return 0;
1617
1618	trace_xfs_zero_eof(ip, isize, pos - isize);
1619	return xfs_zero_range(ip, isize, pos - isize, NULL, NULL);
1620}
1621
1622/*
1623 * Prepare two files for range cloning.  Upon a successful return both inodes
1624 * will have the iolock and mmaplock held, the page cache of the out file will
1625 * be truncated, and any leases on the out file will have been broken.  This
1626 * function borrows heavily from xfs_file_aio_write_checks.
1627 *
1628 * The VFS allows partial EOF blocks to "match" for dedupe even though it hasn't
1629 * checked that the bytes beyond EOF physically match. Hence we cannot use the
1630 * EOF block in the source dedupe range because it's not a complete block match,
1631 * hence can introduce a corruption into the file that has it's block replaced.
1632 *
1633 * In similar fashion, the VFS file cloning also allows partial EOF blocks to be
1634 * "block aligned" for the purposes of cloning entire files.  However, if the
1635 * source file range includes the EOF block and it lands within the existing EOF
1636 * of the destination file, then we can expose stale data from beyond the source
1637 * file EOF in the destination file.
1638 *
1639 * XFS doesn't support partial block sharing, so in both cases we have check
1640 * these cases ourselves. For dedupe, we can simply round the length to dedupe
1641 * down to the previous whole block and ignore the partial EOF block. While this
1642 * means we can't dedupe the last block of a file, this is an acceptible
1643 * tradeoff for simplicity on implementation.
1644 *
1645 * For cloning, we want to share the partial EOF block if it is also the new EOF
1646 * block of the destination file. If the partial EOF block lies inside the
1647 * existing destination EOF, then we have to abort the clone to avoid exposing
1648 * stale data in the destination file. Hence we reject these clone attempts with
1649 * -EINVAL in this case.
1650 */
1651int
1652xfs_reflink_remap_prep(
1653	struct file		*file_in,
1654	loff_t			pos_in,
1655	struct file		*file_out,
1656	loff_t			pos_out,
1657	loff_t			*len,
1658	unsigned int		remap_flags)
1659{
1660	struct inode		*inode_in = file_inode(file_in);
1661	struct xfs_inode	*src = XFS_I(inode_in);
1662	struct inode		*inode_out = file_inode(file_out);
1663	struct xfs_inode	*dest = XFS_I(inode_out);
1664	int			ret;
1665
1666	/* Lock both files against IO */
1667	ret = xfs_ilock2_io_mmap(src, dest);
1668	if (ret)
1669		return ret;
1670
1671	/* Check file eligibility and prepare for block sharing. */
1672	ret = -EINVAL;
1673	/* Can't reflink between data and rt volumes */
1674	if (XFS_IS_REALTIME_INODE(src) != XFS_IS_REALTIME_INODE(dest))
1675		goto out_unlock;
1676
1677	/* Don't share DAX file data with non-DAX file. */
1678	if (IS_DAX(inode_in) != IS_DAX(inode_out))
1679		goto out_unlock;
1680
1681	if (!IS_DAX(inode_in))
1682		ret = generic_remap_file_range_prep(file_in, pos_in, file_out,
1683				pos_out, len, remap_flags);
1684	else
1685		ret = dax_remap_file_range_prep(file_in, pos_in, file_out,
1686				pos_out, len, remap_flags, &xfs_read_iomap_ops);
1687	if (ret || *len == 0)
1688		goto out_unlock;
1689
1690	/* Attach dquots to dest inode before changing block map */
1691	ret = xfs_qm_dqattach(dest);
1692	if (ret)
1693		goto out_unlock;
1694
1695	/*
1696	 * Zero existing post-eof speculative preallocations in the destination
1697	 * file.
1698	 */
1699	ret = xfs_reflink_zero_posteof(dest, pos_out);
1700	if (ret)
1701		goto out_unlock;
1702
1703	/* Set flags and remap blocks. */
1704	ret = xfs_reflink_set_inode_flag(src, dest);
1705	if (ret)
1706		goto out_unlock;
1707
1708	/*
1709	 * If pos_out > EOF, we may have dirtied blocks between EOF and
1710	 * pos_out. In that case, we need to extend the flush and unmap to cover
1711	 * from EOF to the end of the copy length.
1712	 */
1713	if (pos_out > XFS_ISIZE(dest)) {
1714		loff_t	flen = *len + (pos_out - XFS_ISIZE(dest));
1715		ret = xfs_flush_unmap_range(dest, XFS_ISIZE(dest), flen);
1716	} else {
1717		ret = xfs_flush_unmap_range(dest, pos_out, *len);
1718	}
1719	if (ret)
1720		goto out_unlock;
1721
1722	xfs_iflags_set(src, XFS_IREMAPPING);
1723	if (inode_in != inode_out)
1724		xfs_ilock_demote(src, XFS_IOLOCK_EXCL | XFS_MMAPLOCK_EXCL);
1725
1726	return 0;
1727out_unlock:
1728	xfs_iunlock2_io_mmap(src, dest);
1729	return ret;
1730}
1731
1732/* Does this inode need the reflink flag? */
1733int
1734xfs_reflink_inode_has_shared_extents(
1735	struct xfs_trans		*tp,
1736	struct xfs_inode		*ip,
1737	bool				*has_shared)
1738{
1739	struct xfs_bmbt_irec		got;
1740	struct xfs_mount		*mp = ip->i_mount;
1741	struct xfs_ifork		*ifp;
1742	struct xfs_iext_cursor		icur;
1743	bool				found;
1744	int				error;
1745
1746	ifp = xfs_ifork_ptr(ip, XFS_DATA_FORK);
1747	error = xfs_iread_extents(tp, ip, XFS_DATA_FORK);
1748	if (error)
1749		return error;
1750
1751	*has_shared = false;
1752	found = xfs_iext_lookup_extent(ip, ifp, 0, &icur, &got);
1753	while (found) {
1754		xfs_extlen_t		shared_offset, shared_len;
1755
1756		if (isnullstartblock(got.br_startblock) ||
1757		    got.br_state != XFS_EXT_NORM)
1758			goto next;
1759
1760		if (XFS_IS_REALTIME_INODE(ip))
1761			error = xfs_reflink_find_rtshared(mp, tp, &got,
1762					&shared_offset, &shared_len, false);
1763		else
1764			error = xfs_reflink_find_shared(mp, tp, &got,
1765					&shared_offset, &shared_len, false);
1766		if (error)
1767			return error;
1768
1769		/* Is there still a shared block here? */
1770		if (shared_len) {
1771			*has_shared = true;
1772			return 0;
1773		}
1774next:
1775		found = xfs_iext_next_extent(ifp, &icur, &got);
1776	}
1777
1778	return 0;
1779}
1780
1781/*
1782 * Clear the inode reflink flag if there are no shared extents.
1783 *
1784 * The caller is responsible for joining the inode to the transaction passed in.
1785 * The inode will be joined to the transaction that is returned to the caller.
1786 */
1787int
1788xfs_reflink_clear_inode_flag(
1789	struct xfs_inode	*ip,
1790	struct xfs_trans	**tpp)
1791{
1792	bool			needs_flag;
1793	int			error = 0;
1794
1795	ASSERT(xfs_is_reflink_inode(ip));
1796
1797	if (!xfs_can_free_cowblocks(ip))
1798		return 0;
1799
1800	error = xfs_reflink_inode_has_shared_extents(*tpp, ip, &needs_flag);
1801	if (error || needs_flag)
1802		return error;
1803
1804	/*
1805	 * We didn't find any shared blocks so turn off the reflink flag.
1806	 * First, get rid of any leftover CoW mappings.
1807	 */
1808	error = xfs_reflink_cancel_cow_blocks(ip, tpp, 0, XFS_MAX_FILEOFF,
1809			true);
1810	if (error)
1811		return error;
1812
1813	/* Clear the inode flag. */
1814	trace_xfs_reflink_unset_inode_flag(ip);
1815	ip->i_diflags2 &= ~XFS_DIFLAG2_REFLINK;
1816	xfs_inode_clear_cowblocks_tag(ip);
1817	xfs_trans_log_inode(*tpp, ip, XFS_ILOG_CORE);
1818
1819	return error;
1820}
1821
1822/*
1823 * Clear the inode reflink flag if there are no shared extents and the size
1824 * hasn't changed.
1825 */
1826STATIC int
1827xfs_reflink_try_clear_inode_flag(
1828	struct xfs_inode	*ip)
1829{
1830	struct xfs_mount	*mp = ip->i_mount;
1831	struct xfs_trans	*tp;
1832	int			error = 0;
1833
1834	/* Start a rolling transaction to remove the mappings */
1835	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, 0, 0, 0, &tp);
1836	if (error)
1837		return error;
1838
1839	xfs_ilock(ip, XFS_ILOCK_EXCL);
1840	xfs_trans_ijoin(tp, ip, 0);
1841
1842	error = xfs_reflink_clear_inode_flag(ip, &tp);
1843	if (error)
1844		goto cancel;
1845
1846	error = xfs_trans_commit(tp);
1847	if (error)
1848		goto out;
1849
1850	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1851	return 0;
1852cancel:
1853	xfs_trans_cancel(tp);
1854out:
1855	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1856	return error;
1857}
1858
1859/*
1860 * Pre-COW all shared blocks within a given byte range of a file and turn off
1861 * the reflink flag if we unshare all of the file's blocks.
1862 */
1863int
1864xfs_reflink_unshare(
1865	struct xfs_inode	*ip,
1866	xfs_off_t		offset,
1867	xfs_off_t		len)
1868{
1869	struct inode		*inode = VFS_I(ip);
1870	int			error;
1871
1872	if (!xfs_is_reflink_inode(ip))
1873		return 0;
1874
1875	trace_xfs_reflink_unshare(ip, offset, len);
1876
1877	inode_dio_wait(inode);
1878
1879	if (IS_DAX(inode))
1880		error = dax_file_unshare(inode, offset, len,
1881				&xfs_dax_write_iomap_ops);
1882	else
1883		error = iomap_file_unshare(inode, offset, len,
1884				&xfs_buffered_write_iomap_ops,
1885				&xfs_iomap_write_ops);
1886	if (error)
1887		goto out;
1888
1889	error = filemap_write_and_wait_range(inode->i_mapping, offset,
1890			offset + len - 1);
1891	if (error)
1892		goto out;
1893
1894	/* Turn off the reflink flag if possible. */
1895	error = xfs_reflink_try_clear_inode_flag(ip);
1896	if (error)
1897		goto out;
1898	return 0;
1899
1900out:
1901	trace_xfs_reflink_unshare_error(ip, error, _RET_IP_);
1902	return error;
1903}
1904
1905/*
1906 * Can we use reflink with this realtime extent size?  Note that we don't check
1907 * for rblocks > 0 here because this can be called as part of attaching a new
1908 * rt section.
1909 */
1910bool
1911xfs_reflink_supports_rextsize(
1912	struct xfs_mount	*mp,
1913	unsigned int		rextsize)
1914{
1915	/* reflink on the realtime device requires rtgroups */
1916	if (!xfs_has_rtgroups(mp))
1917	       return false;
1918
1919	/*
1920	 * Reflink doesn't support rt extent size larger than a single fsblock
1921	 * because we would have to perform CoW-around for unaligned write
1922	 * requests to guarantee that we always remap entire rt extents.
1923	 */
1924	if (rextsize != 1)
1925		return false;
1926
1927	return true;
1928}