1/*
   2 * Copyright (C) 2016 Oracle.  All Rights Reserved.
   3 *
   4 * Author: Darrick J. Wong <darrick.wong@oracle.com>
   5 *
   6 * This program is free software; you can redistribute it and/or
   7 * modify it under the terms of the GNU General Public License
   8 * as published by the Free Software Foundation; either version 2
   9 * of the License, or (at your option) any later version.
  10 *
  11 * This program is distributed in the hope that it would be useful,
  12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  14 * GNU General Public License for more details.
  15 *
  16 * You should have received a copy of the GNU General Public License
  17 * along with this program; if not, write the Free Software Foundation,
  18 * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301, USA.
  19 */
  20#include "xfs.h"
  21#include "xfs_fs.h"
  22#include "xfs_shared.h"
  23#include "xfs_format.h"
  24#include "xfs_log_format.h"
  25#include "xfs_trans_resv.h"
  26#include "xfs_mount.h"
  27#include "xfs_defer.h"
  28#include "xfs_da_format.h"
  29#include "xfs_da_btree.h"
  30#include "xfs_inode.h"
  31#include "xfs_trans.h"
  32#include "xfs_inode_item.h"
  33#include "xfs_bmap.h"
  34#include "xfs_bmap_util.h"
  35#include "xfs_error.h"
  36#include "xfs_dir2.h"
  37#include "xfs_dir2_priv.h"
  38#include "xfs_ioctl.h"
  39#include "xfs_trace.h"
  40#include "xfs_log.h"
  41#include "xfs_icache.h"
  42#include "xfs_pnfs.h"
  43#include "xfs_btree.h"
  44#include "xfs_refcount_btree.h"
  45#include "xfs_refcount.h"
  46#include "xfs_bmap_btree.h"
  47#include "xfs_trans_space.h"
  48#include "xfs_bit.h"
  49#include "xfs_alloc.h"
  50#include "xfs_quota_defs.h"
  51#include "xfs_quota.h"
  52#include "xfs_reflink.h"
  53#include "xfs_iomap.h"
  54#include "xfs_rmap_btree.h"
  55#include "xfs_sb.h"
  56#include "xfs_ag_resv.h"
  57
  58/*
  59 * Copy on Write of Shared Blocks
  60 *
  61 * XFS must preserve "the usual" file semantics even when two files share
  62 * the same physical blocks.  This means that a write to one file must not
  63 * alter the blocks in a different file; the way that we'll do that is
  64 * through the use of a copy-on-write mechanism.  At a high level, that
  65 * means that when we want to write to a shared block, we allocate a new
  66 * block, write the data to the new block, and if that succeeds we map the
  67 * new block into the file.
  68 *
  69 * XFS provides a "delayed allocation" mechanism that defers the allocation
  70 * of disk blocks to dirty-but-not-yet-mapped file blocks as long as
  71 * possible.  This reduces fragmentation by enabling the filesystem to ask
  72 * for bigger chunks less often, which is exactly what we want for CoW.
  73 *
  74 * The delalloc mechanism begins when the kernel wants to make a block
  75 * writable (write_begin or page_mkwrite).  If the offset is not mapped, we
  76 * create a delalloc mapping, which is a regular in-core extent, but without
  77 * a real startblock.  (For delalloc mappings, the startblock encodes both
  78 * a flag that this is a delalloc mapping, and a worst-case estimate of how
  79 * many blocks might be required to put the mapping into the BMBT.)  delalloc
  80 * mappings are a reservation against the free space in the filesystem;
  81 * adjacent mappings can also be combined into fewer larger mappings.
  82 *
  83 * As an optimization, the CoW extent size hint (cowextsz) creates
  84 * outsized aligned delalloc reservations in the hope of landing out of
  85 * order nearby CoW writes in a single extent on disk, thereby reducing
  86 * fragmentation and improving future performance.
  87 *
  88 * D: --RRRRRRSSSRRRRRRRR--- (data fork)
  89 * C: ------DDDDDDD--------- (CoW fork)
  90 *
  91 * When dirty pages are being written out (typically in writepage), the
  92 * delalloc reservations are converted into unwritten mappings by
  93 * allocating blocks and replacing the delalloc mapping with real ones.
  94 * A delalloc mapping can be replaced by several unwritten ones if the
  95 * free space is fragmented.
  96 *
  97 * D: --RRRRRRSSSRRRRRRRR---
  98 * C: ------UUUUUUU---------
  99 *
 100 * We want to adapt the delalloc mechanism for copy-on-write, since the
 101 * write paths are similar.  The first two steps (creating the reservation
 102 * and allocating the blocks) are exactly the same as delalloc except that
 103 * the mappings must be stored in a separate CoW fork because we do not want
 104 * to disturb the mapping in the data fork until we're sure that the write
 105 * succeeded.  IO completion in this case is the process of removing the old
 106 * mapping from the data fork and moving the new mapping from the CoW fork to
 107 * the data fork.  This will be discussed shortly.
 108 *
 109 * For now, unaligned directio writes will be bounced back to the page cache.
 110 * Block-aligned directio writes will use the same mechanism as buffered
 111 * writes.
 112 *
 113 * Just prior to submitting the actual disk write requests, we convert
 114 * the extents representing the range of the file actually being written
 115 * (as opposed to extra pieces created for the cowextsize hint) to real
 116 * extents.  This will become important in the next step:
 117 *
 118 * D: --RRRRRRSSSRRRRRRRR---
 119 * C: ------UUrrUUU---------
 120 *
 121 * CoW remapping must be done after the data block write completes,
 122 * because we don't want to destroy the old data fork map until we're sure
 123 * the new block has been written.  Since the new mappings are kept in a
 124 * separate fork, we can simply iterate these mappings to find the ones
 125 * that cover the file blocks that we just CoW'd.  For each extent, simply
 126 * unmap the corresponding range in the data fork, map the new range into
 127 * the data fork, and remove the extent from the CoW fork.  Because of
 128 * the presence of the cowextsize hint, however, we must be careful
 129 * only to remap the blocks that we've actually written out --  we must
 130 * never remap delalloc reservations nor CoW staging blocks that have
 131 * yet to be written.  This corresponds exactly to the real extents in
 132 * the CoW fork:
 133 *
 134 * D: --RRRRRRrrSRRRRRRRR---
 135 * C: ------UU--UUU---------
 136 *
 137 * Since the remapping operation can be applied to an arbitrary file
 138 * range, we record the need for the remap step as a flag in the ioend
 139 * instead of declaring a new IO type.  This is required for direct io
 140 * because we only have ioend for the whole dio, and we have to be able to
 141 * remember the presence of unwritten blocks and CoW blocks with a single
 142 * ioend structure.  Better yet, the more ground we can cover with one
 143 * ioend, the better.
 144 */
 145
 146/*
 147 * Given an AG extent, find the lowest-numbered run of shared blocks
 148 * within that range and return the range in fbno/flen.  If
 149 * find_end_of_shared is true, return the longest contiguous extent of
 150 * shared blocks.  If there are no shared extents, fbno and flen will
 151 * be set to NULLAGBLOCK and 0, respectively.
 152 */
 153int
 154xfs_reflink_find_shared(
 155	struct xfs_mount	*mp,
 156	struct xfs_trans	*tp,
 157	xfs_agnumber_t		agno,
 158	xfs_agblock_t		agbno,
 159	xfs_extlen_t		aglen,
 160	xfs_agblock_t		*fbno,
 161	xfs_extlen_t		*flen,
 162	bool			find_end_of_shared)
 163{
 164	struct xfs_buf		*agbp;
 165	struct xfs_btree_cur	*cur;
 166	int			error;
 167
 168	error = xfs_alloc_read_agf(mp, tp, agno, 0, &agbp);
 169	if (error)
 170		return error;
 171	if (!agbp)
 172		return -ENOMEM;
 173
 174	cur = xfs_refcountbt_init_cursor(mp, tp, agbp, agno, NULL);
 175
 176	error = xfs_refcount_find_shared(cur, agbno, aglen, fbno, flen,
 177			find_end_of_shared);
 178
 179	xfs_btree_del_cursor(cur, error ? XFS_BTREE_ERROR : XFS_BTREE_NOERROR);
 180
 181	xfs_trans_brelse(tp, agbp);
 182	return error;
 183}
 184
 185/*
 186 * Trim the mapping to the next block where there's a change in the
 187 * shared/unshared status.  More specifically, this means that we
 188 * find the lowest-numbered extent of shared blocks that coincides with
 189 * the given block mapping.  If the shared extent overlaps the start of
 190 * the mapping, trim the mapping to the end of the shared extent.  If
 191 * the shared region intersects the mapping, trim the mapping to the
 192 * start of the shared extent.  If there are no shared regions that
 193 * overlap, just return the original extent.
 194 */
 195int
 196xfs_reflink_trim_around_shared(
 197	struct xfs_inode	*ip,
 198	struct xfs_bmbt_irec	*irec,
 199	bool			*shared,
 200	bool			*trimmed)
 201{
 202	xfs_agnumber_t		agno;
 203	xfs_agblock_t		agbno;
 204	xfs_extlen_t		aglen;
 205	xfs_agblock_t		fbno;
 206	xfs_extlen_t		flen;
 207	int			error = 0;
 208
 209	/* Holes, unwritten, and delalloc extents cannot be shared */
 210	if (!xfs_is_reflink_inode(ip) || !xfs_bmap_is_real_extent(irec)) {
 211		*shared = false;
 212		return 0;
 213	}
 214
 215	trace_xfs_reflink_trim_around_shared(ip, irec);
 216
 217	agno = XFS_FSB_TO_AGNO(ip->i_mount, irec->br_startblock);
 218	agbno = XFS_FSB_TO_AGBNO(ip->i_mount, irec->br_startblock);
 219	aglen = irec->br_blockcount;
 220
 221	error = xfs_reflink_find_shared(ip->i_mount, NULL, agno, agbno,
 222			aglen, &fbno, &flen, true);
 223	if (error)
 224		return error;
 225
 226	*shared = *trimmed = false;
 227	if (fbno == NULLAGBLOCK) {
 228		/* No shared blocks at all. */
 229		return 0;
 230	} else if (fbno == agbno) {
 231		/*
 232		 * The start of this extent is shared.  Truncate the
 233		 * mapping at the end of the shared region so that a
 234		 * subsequent iteration starts at the start of the
 235		 * unshared region.
 236		 */
 237		irec->br_blockcount = flen;
 238		*shared = true;
 239		if (flen != aglen)
 240			*trimmed = true;
 241		return 0;
 242	} else {
 243		/*
 244		 * There's a shared extent midway through this extent.
 245		 * Truncate the mapping at the start of the shared
 246		 * extent so that a subsequent iteration starts at the
 247		 * start of the shared region.
 248		 */
 249		irec->br_blockcount = fbno - agbno;
 250		*trimmed = true;
 251		return 0;
 252	}
 253}
 254
 255/*
 256 * Trim the passed in imap to the next shared/unshared extent boundary, and
 257 * if imap->br_startoff points to a shared extent reserve space for it in the
 258 * COW fork.  In this case *shared is set to true, else to false.
 259 *
 260 * Note that imap will always contain the block numbers for the existing blocks
 261 * in the data fork, as the upper layers need them for read-modify-write
 262 * operations.
 263 */
 264int
 265xfs_reflink_reserve_cow(
 266	struct xfs_inode	*ip,
 267	struct xfs_bmbt_irec	*imap,
 268	bool			*shared)
 269{
 270	struct xfs_ifork	*ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
 271	struct xfs_bmbt_irec	got;
 272	int			error = 0;
 273	bool			eof = false, trimmed;
 274	struct xfs_iext_cursor	icur;
 275
 276	/*
 277	 * Search the COW fork extent list first.  This serves two purposes:
 278	 * first this implement the speculative preallocation using cowextisze,
 279	 * so that we also unshared block adjacent to shared blocks instead
 280	 * of just the shared blocks themselves.  Second the lookup in the
 281	 * extent list is generally faster than going out to the shared extent
 282	 * tree.
 283	 */
 284
 285	if (!xfs_iext_lookup_extent(ip, ifp, imap->br_startoff, &icur, &got))
 286		eof = true;
 287	if (!eof && got.br_startoff <= imap->br_startoff) {
 288		trace_xfs_reflink_cow_found(ip, imap);
 289		xfs_trim_extent(imap, got.br_startoff, got.br_blockcount);
 290
 291		*shared = true;
 292		return 0;
 293	}
 294
 295	/* Trim the mapping to the nearest shared extent boundary. */
 296	error = xfs_reflink_trim_around_shared(ip, imap, shared, &trimmed);
 297	if (error)
 298		return error;
 299
 300	/* Not shared?  Just report the (potentially capped) extent. */
 301	if (!*shared)
 302		return 0;
 303
 304	/*
 305	 * Fork all the shared blocks from our write offset until the end of
 306	 * the extent.
 307	 */
 308	error = xfs_qm_dqattach_locked(ip, 0);
 309	if (error)
 310		return error;
 311
 312	error = xfs_bmapi_reserve_delalloc(ip, XFS_COW_FORK, imap->br_startoff,
 313			imap->br_blockcount, 0, &got, &icur, eof);
 314	if (error == -ENOSPC || error == -EDQUOT)
 315		trace_xfs_reflink_cow_enospc(ip, imap);
 316	if (error)
 317		return error;
 318
 319	trace_xfs_reflink_cow_alloc(ip, &got);
 320	return 0;
 321}
 322
 323/* Convert part of an unwritten CoW extent to a real one. */
 324STATIC int
 325xfs_reflink_convert_cow_extent(
 326	struct xfs_inode		*ip,
 327	struct xfs_bmbt_irec		*imap,
 328	xfs_fileoff_t			offset_fsb,
 329	xfs_filblks_t			count_fsb,
 330	struct xfs_defer_ops		*dfops)
 331{
 332	xfs_fsblock_t			first_block = NULLFSBLOCK;
 333	int				nimaps = 1;
 334
 335	if (imap->br_state == XFS_EXT_NORM)
 336		return 0;
 337
 338	xfs_trim_extent(imap, offset_fsb, count_fsb);
 339	trace_xfs_reflink_convert_cow(ip, imap);
 340	if (imap->br_blockcount == 0)
 341		return 0;
 342	return xfs_bmapi_write(NULL, ip, imap->br_startoff, imap->br_blockcount,
 343			XFS_BMAPI_COWFORK | XFS_BMAPI_CONVERT, &first_block,
 344			0, imap, &nimaps, dfops);
 345}
 346
 347/* Convert all of the unwritten CoW extents in a file's range to real ones. */
 348int
 349xfs_reflink_convert_cow(
 350	struct xfs_inode	*ip,
 351	xfs_off_t		offset,
 352	xfs_off_t		count)
 353{
 354	struct xfs_mount	*mp = ip->i_mount;
 355	xfs_fileoff_t		offset_fsb = XFS_B_TO_FSBT(mp, offset);
 356	xfs_fileoff_t		end_fsb = XFS_B_TO_FSB(mp, offset + count);
 357	xfs_filblks_t		count_fsb = end_fsb - offset_fsb;
 358	struct xfs_bmbt_irec	imap;
 359	struct xfs_defer_ops	dfops;
 360	xfs_fsblock_t		first_block = NULLFSBLOCK;
 361	int			nimaps = 1, error = 0;
 362
 363	ASSERT(count != 0);
 364
 365	xfs_ilock(ip, XFS_ILOCK_EXCL);
 366	error = xfs_bmapi_write(NULL, ip, offset_fsb, count_fsb,
 367			XFS_BMAPI_COWFORK | XFS_BMAPI_CONVERT |
 368			XFS_BMAPI_CONVERT_ONLY, &first_block, 0, &imap, &nimaps,
 369			&dfops);
 370	xfs_iunlock(ip, XFS_ILOCK_EXCL);
 371	return error;
 372}
 373
 374/* Allocate all CoW reservations covering a range of blocks in a file. */
 375int
 376xfs_reflink_allocate_cow(
 377	struct xfs_inode	*ip,
 378	struct xfs_bmbt_irec	*imap,
 379	bool			*shared,
 380	uint			*lockmode)
 381{
 382	struct xfs_mount	*mp = ip->i_mount;
 383	xfs_fileoff_t		offset_fsb = imap->br_startoff;
 384	xfs_filblks_t		count_fsb = imap->br_blockcount;
 385	struct xfs_bmbt_irec	got;
 386	struct xfs_defer_ops	dfops;
 387	struct xfs_trans	*tp = NULL;
 388	xfs_fsblock_t		first_block;
 389	int			nimaps, error = 0;
 390	bool			trimmed;
 391	xfs_filblks_t		resaligned;
 392	xfs_extlen_t		resblks = 0;
 393	struct xfs_iext_cursor	icur;
 394
 395retry:
 396	ASSERT(xfs_is_reflink_inode(ip));
 397	ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
 398
 399	/*
 400	 * Even if the extent is not shared we might have a preallocation for
 401	 * it in the COW fork.  If so use it.
 402	 */
 403	if (xfs_iext_lookup_extent(ip, ip->i_cowfp, offset_fsb, &icur, &got) &&
 404	    got.br_startoff <= offset_fsb) {
 405		*shared = true;
 406
 407		/* If we have a real allocation in the COW fork we're done. */
 408		if (!isnullstartblock(got.br_startblock)) {
 409			xfs_trim_extent(&got, offset_fsb, count_fsb);
 410			*imap = got;
 411			goto convert;
 412		}
 413
 414		xfs_trim_extent(imap, got.br_startoff, got.br_blockcount);
 415	} else {
 416		error = xfs_reflink_trim_around_shared(ip, imap, shared, &trimmed);
 417		if (error || !*shared)
 418			goto out;
 419	}
 420
 421	if (!tp) {
 422		resaligned = xfs_aligned_fsb_count(imap->br_startoff,
 423			imap->br_blockcount, xfs_get_cowextsz_hint(ip));
 424		resblks = XFS_DIOSTRAT_SPACE_RES(mp, resaligned);
 425
 426		xfs_iunlock(ip, *lockmode);
 427		error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, resblks, 0, 0, &tp);
 428		*lockmode = XFS_ILOCK_EXCL;
 429		xfs_ilock(ip, *lockmode);
 430
 431		if (error)
 432			return error;
 433
 434		error = xfs_qm_dqattach_locked(ip, 0);
 435		if (error)
 436			goto out;
 437		goto retry;
 438	}
 439
 440	error = xfs_trans_reserve_quota_nblks(tp, ip, resblks, 0,
 441			XFS_QMOPT_RES_REGBLKS);
 442	if (error)
 443		goto out;
 444
 445	xfs_trans_ijoin(tp, ip, 0);
 446
 447	xfs_defer_init(&dfops, &first_block);
 448	nimaps = 1;
 449
 450	/* Allocate the entire reservation as unwritten blocks. */
 451	error = xfs_bmapi_write(tp, ip, imap->br_startoff, imap->br_blockcount,
 452			XFS_BMAPI_COWFORK | XFS_BMAPI_PREALLOC, &first_block,
 453			resblks, imap, &nimaps, &dfops);
 454	if (error)
 455		goto out_bmap_cancel;
 456
 457	xfs_inode_set_cowblocks_tag(ip);
 458
 459	/* Finish up. */
 460	error = xfs_defer_finish(&tp, &dfops);
 461	if (error)
 462		goto out_bmap_cancel;
 463
 464	error = xfs_trans_commit(tp);
 465	if (error)
 466		return error;
 467
 468	/*
 469	 * Allocation succeeded but the requested range was not even partially
 470	 * satisfied?  Bail out!
 471	 */
 472	if (nimaps == 0)
 473		return -ENOSPC;
 474convert:
 475	return xfs_reflink_convert_cow_extent(ip, imap, offset_fsb, count_fsb,
 476			&dfops);
 477out_bmap_cancel:
 478	xfs_defer_cancel(&dfops);
 479	xfs_trans_unreserve_quota_nblks(tp, ip, (long)resblks, 0,
 480			XFS_QMOPT_RES_REGBLKS);
 481out:
 482	if (tp)
 483		xfs_trans_cancel(tp);
 484	return error;
 485}
 486
 487/*
 488 * Find the CoW reservation for a given byte offset of a file.
 489 */
 490bool
 491xfs_reflink_find_cow_mapping(
 492	struct xfs_inode		*ip,
 493	xfs_off_t			offset,
 494	struct xfs_bmbt_irec		*imap)
 495{
 496	struct xfs_ifork		*ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
 497	xfs_fileoff_t			offset_fsb;
 498	struct xfs_bmbt_irec		got;
 499	struct xfs_iext_cursor		icur;
 500
 501	ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL | XFS_ILOCK_SHARED));
 502
 503	if (!xfs_is_reflink_inode(ip))
 504		return false;
 505	offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
 506	if (!xfs_iext_lookup_extent(ip, ifp, offset_fsb, &icur, &got))
 507		return false;
 508	if (got.br_startoff > offset_fsb)
 509		return false;
 510
 511	trace_xfs_reflink_find_cow_mapping(ip, offset, 1, XFS_IO_OVERWRITE,
 512			&got);
 513	*imap = got;
 514	return true;
 515}
 516
 517/*
 518 * Trim an extent to end at the next CoW reservation past offset_fsb.
 519 */
 520void
 521xfs_reflink_trim_irec_to_next_cow(
 522	struct xfs_inode		*ip,
 523	xfs_fileoff_t			offset_fsb,
 524	struct xfs_bmbt_irec		*imap)
 525{
 526	struct xfs_ifork		*ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
 527	struct xfs_bmbt_irec		got;
 528	struct xfs_iext_cursor		icur;
 529
 530	if (!xfs_is_reflink_inode(ip))
 531		return;
 532
 533	/* Find the extent in the CoW fork. */
 534	if (!xfs_iext_lookup_extent(ip, ifp, offset_fsb, &icur, &got))
 535		return;
 536
 537	/* This is the extent before; try sliding up one. */
 538	if (got.br_startoff < offset_fsb) {
 539		if (!xfs_iext_next_extent(ifp, &icur, &got))
 540			return;
 541	}
 542
 543	if (got.br_startoff >= imap->br_startoff + imap->br_blockcount)
 544		return;
 545
 546	imap->br_blockcount = got.br_startoff - imap->br_startoff;
 547	trace_xfs_reflink_trim_irec(ip, imap);
 548}
 549
 550/*
 551 * Cancel CoW reservations for some block range of an inode.
 552 *
 553 * If cancel_real is true this function cancels all COW fork extents for the
 554 * inode; if cancel_real is false, real extents are not cleared.
 555 */
 556int
 557xfs_reflink_cancel_cow_blocks(
 558	struct xfs_inode		*ip,
 559	struct xfs_trans		**tpp,
 560	xfs_fileoff_t			offset_fsb,
 561	xfs_fileoff_t			end_fsb,
 562	bool				cancel_real)
 563{
 564	struct xfs_ifork		*ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
 565	struct xfs_bmbt_irec		got, del;
 566	struct xfs_iext_cursor		icur;
 567	xfs_fsblock_t			firstfsb;
 568	struct xfs_defer_ops		dfops;
 569	int				error = 0;
 570
 571	if (!xfs_is_reflink_inode(ip))
 572		return 0;
 573	if (!xfs_iext_lookup_extent_before(ip, ifp, &end_fsb, &icur, &got))
 574		return 0;
 575
 576	/* Walk backwards until we're out of the I/O range... */
 577	while (got.br_startoff + got.br_blockcount > offset_fsb) {
 578		del = got;
 579		xfs_trim_extent(&del, offset_fsb, end_fsb - offset_fsb);
 580
 581		/* Extent delete may have bumped ext forward */
 582		if (!del.br_blockcount) {
 583			xfs_iext_prev(ifp, &icur);
 584			goto next_extent;
 585		}
 586
 587		trace_xfs_reflink_cancel_cow(ip, &del);
 588
 589		if (isnullstartblock(del.br_startblock)) {
 590			error = xfs_bmap_del_extent_delay(ip, XFS_COW_FORK,
 591					&icur, &got, &del);
 592			if (error)
 593				break;
 594		} else if (del.br_state == XFS_EXT_UNWRITTEN || cancel_real) {
 595			xfs_trans_ijoin(*tpp, ip, 0);
 596			xfs_defer_init(&dfops, &firstfsb);
 597
 598			/* Free the CoW orphan record. */
 599			error = xfs_refcount_free_cow_extent(ip->i_mount,
 600					&dfops, del.br_startblock,
 601					del.br_blockcount);
 602			if (error)
 603				break;
 604
 605			xfs_bmap_add_free(ip->i_mount, &dfops,
 606					del.br_startblock, del.br_blockcount,
 607					NULL);
 608
 609			/* Roll the transaction */
 610			xfs_defer_ijoin(&dfops, ip);
 611			error = xfs_defer_finish(tpp, &dfops);
 612			if (error) {
 613				xfs_defer_cancel(&dfops);
 614				break;
 615			}
 616
 617			/* Remove the mapping from the CoW fork. */
 618			xfs_bmap_del_extent_cow(ip, &icur, &got, &del);
 619
 620			/* Remove the quota reservation */
 621			error = xfs_trans_reserve_quota_nblks(NULL, ip,
 622					-(long)del.br_blockcount, 0,
 623					XFS_QMOPT_RES_REGBLKS);
 624			if (error)
 625				break;
 626		} else {
 627			/* Didn't do anything, push cursor back. */
 628			xfs_iext_prev(ifp, &icur);
 629		}
 630next_extent:
 631		if (!xfs_iext_get_extent(ifp, &icur, &got))
 632			break;
 633	}
 634
 635	/* clear tag if cow fork is emptied */
 636	if (!ifp->if_bytes)
 637		xfs_inode_clear_cowblocks_tag(ip);
 638
 639	return error;
 640}
 641
 642/*
 643 * Cancel CoW reservations for some byte range of an inode.
 644 *
 645 * If cancel_real is true this function cancels all COW fork extents for the
 646 * inode; if cancel_real is false, real extents are not cleared.
 647 */
 648int
 649xfs_reflink_cancel_cow_range(
 650	struct xfs_inode	*ip,
 651	xfs_off_t		offset,
 652	xfs_off_t		count,
 653	bool			cancel_real)
 654{
 655	struct xfs_trans	*tp;
 656	xfs_fileoff_t		offset_fsb;
 657	xfs_fileoff_t		end_fsb;
 658	int			error;
 659
 660	trace_xfs_reflink_cancel_cow_range(ip, offset, count);
 661	ASSERT(xfs_is_reflink_inode(ip));
 662
 663	offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
 664	if (count == NULLFILEOFF)
 665		end_fsb = NULLFILEOFF;
 666	else
 667		end_fsb = XFS_B_TO_FSB(ip->i_mount, offset + count);
 668
 669	/* Start a rolling transaction to remove the mappings */
 670	error = xfs_trans_alloc(ip->i_mount, &M_RES(ip->i_mount)->tr_write,
 671			0, 0, XFS_TRANS_NOFS, &tp);
 672	if (error)
 673		goto out;
 674
 675	xfs_ilock(ip, XFS_ILOCK_EXCL);
 676	xfs_trans_ijoin(tp, ip, 0);
 677
 678	/* Scrape out the old CoW reservations */
 679	error = xfs_reflink_cancel_cow_blocks(ip, &tp, offset_fsb, end_fsb,
 680			cancel_real);
 681	if (error)
 682		goto out_cancel;
 683
 684	error = xfs_trans_commit(tp);
 685
 686	xfs_iunlock(ip, XFS_ILOCK_EXCL);
 687	return error;
 688
 689out_cancel:
 690	xfs_trans_cancel(tp);
 691	xfs_iunlock(ip, XFS_ILOCK_EXCL);
 692out:
 693	trace_xfs_reflink_cancel_cow_range_error(ip, error, _RET_IP_);
 694	return error;
 695}
 696
 697/*
 698 * Remap parts of a file's data fork after a successful CoW.
 699 */
 700int
 701xfs_reflink_end_cow(
 702	struct xfs_inode		*ip,
 703	xfs_off_t			offset,
 704	xfs_off_t			count)
 705{
 706	struct xfs_ifork		*ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
 707	struct xfs_bmbt_irec		got, del;
 708	struct xfs_trans		*tp;
 709	xfs_fileoff_t			offset_fsb;
 710	xfs_fileoff_t			end_fsb;
 711	xfs_fsblock_t			firstfsb;
 712	struct xfs_defer_ops		dfops;
 713	int				error;
 714	unsigned int			resblks;
 715	xfs_filblks_t			rlen;
 716	struct xfs_iext_cursor		icur;
 717
 718	trace_xfs_reflink_end_cow(ip, offset, count);
 719
 720	/* No COW extents?  That's easy! */
 721	if (ifp->if_bytes == 0)
 722		return 0;
 723
 724	offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
 725	end_fsb = XFS_B_TO_FSB(ip->i_mount, offset + count);
 726
 727	/*
 728	 * Start a rolling transaction to switch the mappings.  We're
 729	 * unlikely ever to have to remap 16T worth of single-block
 730	 * extents, so just cap the worst case extent count to 2^32-1.
 731	 * Stick a warning in just in case, and avoid 64-bit division.
 732	 */
 733	BUILD_BUG_ON(MAX_RW_COUNT > UINT_MAX);
 734	if (end_fsb - offset_fsb > UINT_MAX) {
 735		error = -EFSCORRUPTED;
 736		xfs_force_shutdown(ip->i_mount, SHUTDOWN_CORRUPT_INCORE);
 737		ASSERT(0);
 738		goto out;
 739	}
 740	resblks = XFS_NEXTENTADD_SPACE_RES(ip->i_mount,
 741			(unsigned int)(end_fsb - offset_fsb),
 742			XFS_DATA_FORK);
 743	error = xfs_trans_alloc(ip->i_mount, &M_RES(ip->i_mount)->tr_write,
 744			resblks, 0, XFS_TRANS_RESERVE | XFS_TRANS_NOFS, &tp);
 745	if (error)
 746		goto out;
 747
 748	xfs_ilock(ip, XFS_ILOCK_EXCL);
 749	xfs_trans_ijoin(tp, ip, 0);
 750
 751	/*
 752	 * In case of racing, overlapping AIO writes no COW extents might be
 753	 * left by the time I/O completes for the loser of the race.  In that
 754	 * case we are done.
 755	 */
 756	if (!xfs_iext_lookup_extent_before(ip, ifp, &end_fsb, &icur, &got))
 757		goto out_cancel;
 758
 759	/* Walk backwards until we're out of the I/O range... */
 760	while (got.br_startoff + got.br_blockcount > offset_fsb) {
 761		del = got;
 762		xfs_trim_extent(&del, offset_fsb, end_fsb - offset_fsb);
 763
 764		/* Extent delete may have bumped ext forward */
 765		if (!del.br_blockcount)
 766			goto prev_extent;
 767
 768		ASSERT(!isnullstartblock(got.br_startblock));
 769
 770		/*
 771		 * Don't remap unwritten extents; these are
 772		 * speculatively preallocated CoW extents that have been
 773		 * allocated but have not yet been involved in a write.
 774		 */
 775		if (got.br_state == XFS_EXT_UNWRITTEN)
 776			goto prev_extent;
 777
 778		/* Unmap the old blocks in the data fork. */
 779		xfs_defer_init(&dfops, &firstfsb);
 780		rlen = del.br_blockcount;
 781		error = __xfs_bunmapi(tp, ip, del.br_startoff, &rlen, 0, 1,
 782				&firstfsb, &dfops);
 783		if (error)
 784			goto out_defer;
 785
 786		/* Trim the extent to whatever got unmapped. */
 787		if (rlen) {
 788			xfs_trim_extent(&del, del.br_startoff + rlen,
 789				del.br_blockcount - rlen);
 790		}
 791		trace_xfs_reflink_cow_remap(ip, &del);
 792
 793		/* Free the CoW orphan record. */
 794		error = xfs_refcount_free_cow_extent(tp->t_mountp, &dfops,
 795				del.br_startblock, del.br_blockcount);
 796		if (error)
 797			goto out_defer;
 798
 799		/* Map the new blocks into the data fork. */
 800		error = xfs_bmap_map_extent(tp->t_mountp, &dfops, ip, &del);
 801		if (error)
 802			goto out_defer;
 803
 804		/* Charge this new data fork mapping to the on-disk quota. */
 805		xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_DELBCOUNT,
 806				(long)del.br_blockcount);
 807
 808		/* Remove the mapping from the CoW fork. */
 809		xfs_bmap_del_extent_cow(ip, &icur, &got, &del);
 810
 811		xfs_defer_ijoin(&dfops, ip);
 812		error = xfs_defer_finish(&tp, &dfops);
 813		if (error)
 814			goto out_defer;
 815		if (!xfs_iext_get_extent(ifp, &icur, &got))
 816			break;
 817		continue;
 818prev_extent:
 819		if (!xfs_iext_prev_extent(ifp, &icur, &got))
 820			break;
 821	}
 822
 823	error = xfs_trans_commit(tp);
 824	xfs_iunlock(ip, XFS_ILOCK_EXCL);
 825	if (error)
 826		goto out;
 827	return 0;
 828
 829out_defer:
 830	xfs_defer_cancel(&dfops);
 831out_cancel:
 832	xfs_trans_cancel(tp);
 833	xfs_iunlock(ip, XFS_ILOCK_EXCL);
 834out:
 835	trace_xfs_reflink_end_cow_error(ip, error, _RET_IP_);
 836	return error;
 837}
 838
 839/*
 840 * Free leftover CoW reservations that didn't get cleaned out.
 841 */
 842int
 843xfs_reflink_recover_cow(
 844	struct xfs_mount	*mp)
 845{
 846	xfs_agnumber_t		agno;
 847	int			error = 0;
 848
 849	if (!xfs_sb_version_hasreflink(&mp->m_sb))
 850		return 0;
 851
 852	for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
 853		error = xfs_refcount_recover_cow_leftovers(mp, agno);
 854		if (error)
 855			break;
 856	}
 857
 858	return error;
 859}
 860
 861/*
 862 * Reflinking (Block) Ranges of Two Files Together
 863 *
 864 * First, ensure that the reflink flag is set on both inodes.  The flag is an
 865 * optimization to avoid unnecessary refcount btree lookups in the write path.
 866 *
 867 * Now we can iteratively remap the range of extents (and holes) in src to the
 868 * corresponding ranges in dest.  Let drange and srange denote the ranges of
 869 * logical blocks in dest and src touched by the reflink operation.
 870 *
 871 * While the length of drange is greater than zero,
 872 *    - Read src's bmbt at the start of srange ("imap")
 873 *    - If imap doesn't exist, make imap appear to start at the end of srange
 874 *      with zero length.
 875 *    - If imap starts before srange, advance imap to start at srange.
 876 *    - If imap goes beyond srange, truncate imap to end at the end of srange.
 877 *    - Punch (imap start - srange start + imap len) blocks from dest at
 878 *      offset (drange start).
 879 *    - If imap points to a real range of pblks,
 880 *         > Increase the refcount of the imap's pblks
 881 *         > Map imap's pblks into dest at the offset
 882 *           (drange start + imap start - srange start)
 883 *    - Advance drange and srange by (imap start - srange start + imap len)
 884 *
 885 * Finally, if the reflink made dest longer, update both the in-core and
 886 * on-disk file sizes.
 887 *
 888 * ASCII Art Demonstration:
 889 *
 890 * Let's say we want to reflink this source file:
 891 *
 892 * ----SSSSSSS-SSSSS----SSSSSS (src file)
 893 *   <-------------------->
 894 *
 895 * into this destination file:
 896 *
 897 * --DDDDDDDDDDDDDDDDDDD--DDD (dest file)
 898 *        <-------------------->
 899 * '-' means a hole, and 'S' and 'D' are written blocks in the src and dest.
 900 * Observe that the range has different logical offsets in either file.
 901 *
 902 * Consider that the first extent in the source file doesn't line up with our
 903 * reflink range.  Unmapping  and remapping are separate operations, so we can
 904 * unmap more blocks from the destination file than we remap.
 905 *
 906 * ----SSSSSSS-SSSSS----SSSSSS
 907 *   <------->
 908 * --DDDDD---------DDDDD--DDD
 909 *        <------->
 910 *
 911 * Now remap the source extent into the destination file:
 912 *
 913 * ----SSSSSSS-SSSSS----SSSSSS
 914 *   <------->
 915 * --DDDDD--SSSSSSSDDDDD--DDD
 916 *        <------->
 917 *
 918 * Do likewise with the second hole and extent in our range.  Holes in the
 919 * unmap range don't affect our operation.
 920 *
 921 * ----SSSSSSS-SSSSS----SSSSSS
 922 *            <---->
 923 * --DDDDD--SSSSSSS-SSSSS-DDD
 924 *                 <---->
 925 *
 926 * Finally, unmap and remap part of the third extent.  This will increase the
 927 * size of the destination file.
 928 *
 929 * ----SSSSSSS-SSSSS----SSSSSS
 930 *                  <----->
 931 * --DDDDD--SSSSSSS-SSSSS----SSS
 932 *                       <----->
 933 *
 934 * Once we update the destination file's i_size, we're done.
 935 */
 936
 937/*
 938 * Ensure the reflink bit is set in both inodes.
 939 */
 940STATIC int
 941xfs_reflink_set_inode_flag(
 942	struct xfs_inode	*src,
 943	struct xfs_inode	*dest)
 944{
 945	struct xfs_mount	*mp = src->i_mount;
 946	int			error;
 947	struct xfs_trans	*tp;
 948
 949	if (xfs_is_reflink_inode(src) && xfs_is_reflink_inode(dest))
 950		return 0;
 951
 952	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
 953	if (error)
 954		goto out_error;
 955
 956	/* Lock both files against IO */
 957	if (src->i_ino == dest->i_ino)
 958		xfs_ilock(src, XFS_ILOCK_EXCL);
 959	else
 960		xfs_lock_two_inodes(src, XFS_ILOCK_EXCL, dest, XFS_ILOCK_EXCL);
 961
 962	if (!xfs_is_reflink_inode(src)) {
 963		trace_xfs_reflink_set_inode_flag(src);
 964		xfs_trans_ijoin(tp, src, XFS_ILOCK_EXCL);
 965		src->i_d.di_flags2 |= XFS_DIFLAG2_REFLINK;
 966		xfs_trans_log_inode(tp, src, XFS_ILOG_CORE);
 967		xfs_ifork_init_cow(src);
 968	} else
 969		xfs_iunlock(src, XFS_ILOCK_EXCL);
 970
 971	if (src->i_ino == dest->i_ino)
 972		goto commit_flags;
 973
 974	if (!xfs_is_reflink_inode(dest)) {
 975		trace_xfs_reflink_set_inode_flag(dest);
 976		xfs_trans_ijoin(tp, dest, XFS_ILOCK_EXCL);
 977		dest->i_d.di_flags2 |= XFS_DIFLAG2_REFLINK;
 978		xfs_trans_log_inode(tp, dest, XFS_ILOG_CORE);
 979		xfs_ifork_init_cow(dest);
 980	} else
 981		xfs_iunlock(dest, XFS_ILOCK_EXCL);
 982
 983commit_flags:
 984	error = xfs_trans_commit(tp);
 985	if (error)
 986		goto out_error;
 987	return error;
 988
 989out_error:
 990	trace_xfs_reflink_set_inode_flag_error(dest, error, _RET_IP_);
 991	return error;
 992}
 993
 994/*
 995 * Update destination inode size & cowextsize hint, if necessary.
 996 */
 997STATIC int
 998xfs_reflink_update_dest(
 999	struct xfs_inode	*dest,
1000	xfs_off_t		newlen,
1001	xfs_extlen_t		cowextsize,
1002	bool			is_dedupe)
1003{
1004	struct xfs_mount	*mp = dest->i_mount;
1005	struct xfs_trans	*tp;
1006	int			error;
1007
1008	if (is_dedupe && newlen <= i_size_read(VFS_I(dest)) && cowextsize == 0)
1009		return 0;
1010
1011	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
1012	if (error)
1013		goto out_error;
1014
1015	xfs_ilock(dest, XFS_ILOCK_EXCL);
1016	xfs_trans_ijoin(tp, dest, XFS_ILOCK_EXCL);
1017
1018	if (newlen > i_size_read(VFS_I(dest))) {
1019		trace_xfs_reflink_update_inode_size(dest, newlen);
1020		i_size_write(VFS_I(dest), newlen);
1021		dest->i_d.di_size = newlen;
1022	}
1023
1024	if (cowextsize) {
1025		dest->i_d.di_cowextsize = cowextsize;
1026		dest->i_d.di_flags2 |= XFS_DIFLAG2_COWEXTSIZE;
1027	}
1028
1029	if (!is_dedupe) {
1030		xfs_trans_ichgtime(tp, dest,
1031				   XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
1032	}
1033	xfs_trans_log_inode(tp, dest, XFS_ILOG_CORE);
1034
1035	error = xfs_trans_commit(tp);
1036	if (error)
1037		goto out_error;
1038	return error;
1039
1040out_error:
1041	trace_xfs_reflink_update_inode_size_error(dest, error, _RET_IP_);
1042	return error;
1043}
1044
1045/*
1046 * Do we have enough reserve in this AG to handle a reflink?  The refcount
1047 * btree already reserved all the space it needs, but the rmap btree can grow
1048 * infinitely, so we won't allow more reflinks when the AG is down to the
1049 * btree reserves.
1050 */
1051static int
1052xfs_reflink_ag_has_free_space(
1053	struct xfs_mount	*mp,
1054	xfs_agnumber_t		agno)
1055{
1056	struct xfs_perag	*pag;
1057	int			error = 0;
1058
1059	if (!xfs_sb_version_hasrmapbt(&mp->m_sb))
1060		return 0;
1061
1062	pag = xfs_perag_get(mp, agno);
1063	if (xfs_ag_resv_critical(pag, XFS_AG_RESV_RMAPBT) ||
1064	    xfs_ag_resv_critical(pag, XFS_AG_RESV_METADATA))
1065		error = -ENOSPC;
1066	xfs_perag_put(pag);
1067	return error;
1068}
1069
1070/*
1071 * Unmap a range of blocks from a file, then map other blocks into the hole.
1072 * The range to unmap is (destoff : destoff + srcioff + irec->br_blockcount).
1073 * The extent irec is mapped into dest at irec->br_startoff.
1074 */
1075STATIC int
1076xfs_reflink_remap_extent(
1077	struct xfs_inode	*ip,
1078	struct xfs_bmbt_irec	*irec,
1079	xfs_fileoff_t		destoff,
1080	xfs_off_t		new_isize)
1081{
1082	struct xfs_mount	*mp = ip->i_mount;
1083	bool			real_extent = xfs_bmap_is_real_extent(irec);
1084	struct xfs_trans	*tp;
1085	xfs_fsblock_t		firstfsb;
1086	unsigned int		resblks;
1087	struct xfs_defer_ops	dfops;
1088	struct xfs_bmbt_irec	uirec;
1089	xfs_filblks_t		rlen;
1090	xfs_filblks_t		unmap_len;
1091	xfs_off_t		newlen;
1092	int			error;
1093
1094	unmap_len = irec->br_startoff + irec->br_blockcount - destoff;
1095	trace_xfs_reflink_punch_range(ip, destoff, unmap_len);
1096
1097	/* No reflinking if we're low on space */
1098	if (real_extent) {
1099		error = xfs_reflink_ag_has_free_space(mp,
1100				XFS_FSB_TO_AGNO(mp, irec->br_startblock));
1101		if (error)
1102			goto out;
1103	}
1104
1105	/* Start a rolling transaction to switch the mappings */
1106	resblks = XFS_EXTENTADD_SPACE_RES(ip->i_mount, XFS_DATA_FORK);
1107	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, resblks, 0, 0, &tp);
1108	if (error)
1109		goto out;
1110
1111	xfs_ilock(ip, XFS_ILOCK_EXCL);
1112	xfs_trans_ijoin(tp, ip, 0);
1113
1114	/* If we're not just clearing space, then do we have enough quota? */
1115	if (real_extent) {
1116		error = xfs_trans_reserve_quota_nblks(tp, ip,
1117				irec->br_blockcount, 0, XFS_QMOPT_RES_REGBLKS);
1118		if (error)
1119			goto out_cancel;
1120	}
1121
1122	trace_xfs_reflink_remap(ip, irec->br_startoff,
1123				irec->br_blockcount, irec->br_startblock);
1124
1125	/* Unmap the old blocks in the data fork. */
1126	rlen = unmap_len;
1127	while (rlen) {
1128		xfs_defer_init(&dfops, &firstfsb);
1129		error = __xfs_bunmapi(tp, ip, destoff, &rlen, 0, 1,
1130				&firstfsb, &dfops);
1131		if (error)
1132			goto out_defer;
1133
1134		/*
1135		 * Trim the extent to whatever got unmapped.
1136		 * Remember, bunmapi works backwards.
1137		 */
1138		uirec.br_startblock = irec->br_startblock + rlen;
1139		uirec.br_startoff = irec->br_startoff + rlen;
1140		uirec.br_blockcount = unmap_len - rlen;
1141		unmap_len = rlen;
1142
1143		/* If this isn't a real mapping, we're done. */
1144		if (!real_extent || uirec.br_blockcount == 0)
1145			goto next_extent;
1146
1147		trace_xfs_reflink_remap(ip, uirec.br_startoff,
1148				uirec.br_blockcount, uirec.br_startblock);
1149
1150		/* Update the refcount tree */
1151		error = xfs_refcount_increase_extent(mp, &dfops, &uirec);
1152		if (error)
1153			goto out_defer;
1154
1155		/* Map the new blocks into the data fork. */
1156		error = xfs_bmap_map_extent(mp, &dfops, ip, &uirec);
1157		if (error)
1158			goto out_defer;
1159
1160		/* Update quota accounting. */
1161		xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_BCOUNT,
1162				uirec.br_blockcount);
1163
1164		/* Update dest isize if needed. */
1165		newlen = XFS_FSB_TO_B(mp,
1166				uirec.br_startoff + uirec.br_blockcount);
1167		newlen = min_t(xfs_off_t, newlen, new_isize);
1168		if (newlen > i_size_read(VFS_I(ip))) {
1169			trace_xfs_reflink_update_inode_size(ip, newlen);
1170			i_size_write(VFS_I(ip), newlen);
1171			ip->i_d.di_size = newlen;
1172			xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1173		}
1174
1175next_extent:
1176		/* Process all the deferred stuff. */
1177		xfs_defer_ijoin(&dfops, ip);
1178		error = xfs_defer_finish(&tp, &dfops);
1179		if (error)
1180			goto out_defer;
1181	}
1182
1183	error = xfs_trans_commit(tp);
1184	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1185	if (error)
1186		goto out;
1187	return 0;
1188
1189out_defer:
1190	xfs_defer_cancel(&dfops);
1191out_cancel:
1192	xfs_trans_cancel(tp);
1193	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1194out:
1195	trace_xfs_reflink_remap_extent_error(ip, error, _RET_IP_);
1196	return error;
1197}
1198
1199/*
1200 * Iteratively remap one file's extents (and holes) to another's.
1201 */
1202STATIC int
1203xfs_reflink_remap_blocks(
1204	struct xfs_inode	*src,
1205	xfs_fileoff_t		srcoff,
1206	struct xfs_inode	*dest,
1207	xfs_fileoff_t		destoff,
1208	xfs_filblks_t		len,
1209	xfs_off_t		new_isize)
1210{
1211	struct xfs_bmbt_irec	imap;
1212	int			nimaps;
1213	int			error = 0;
1214	xfs_filblks_t		range_len;
1215
1216	/* drange = (destoff, destoff + len); srange = (srcoff, srcoff + len) */
1217	while (len) {
1218		uint		lock_mode;
1219
1220		trace_xfs_reflink_remap_blocks_loop(src, srcoff, len,
1221				dest, destoff);
1222
1223		/* Read extent from the source file */
1224		nimaps = 1;
1225		lock_mode = xfs_ilock_data_map_shared(src);
1226		error = xfs_bmapi_read(src, srcoff, len, &imap, &nimaps, 0);
1227		xfs_iunlock(src, lock_mode);
1228		if (error)
1229			goto err;
1230		ASSERT(nimaps == 1);
1231
1232		trace_xfs_reflink_remap_imap(src, srcoff, len, XFS_IO_OVERWRITE,
1233				&imap);
1234
1235		/* Translate imap into the destination file. */
1236		range_len = imap.br_startoff + imap.br_blockcount - srcoff;
1237		imap.br_startoff += destoff - srcoff;
1238
1239		/* Clear dest from destoff to the end of imap and map it in. */
1240		error = xfs_reflink_remap_extent(dest, &imap, destoff,
1241				new_isize);
1242		if (error)
1243			goto err;
1244
1245		if (fatal_signal_pending(current)) {
1246			error = -EINTR;
1247			goto err;
1248		}
1249
1250		/* Advance drange/srange */
1251		srcoff += range_len;
1252		destoff += range_len;
1253		len -= range_len;
1254	}
1255
1256	return 0;
1257
1258err:
1259	trace_xfs_reflink_remap_blocks_error(dest, error, _RET_IP_);
1260	return error;
1261}
1262
1263/*
1264 * Grab the exclusive iolock for a data copy from src to dest, making
1265 * sure to abide vfs locking order (lowest pointer value goes first) and
1266 * breaking the pnfs layout leases on dest before proceeding.  The loop
1267 * is needed because we cannot call the blocking break_layout() with the
1268 * src iolock held, and therefore have to back out both locks.
1269 */
1270static int
1271xfs_iolock_two_inodes_and_break_layout(
1272	struct inode		*src,
1273	struct inode		*dest)
1274{
1275	int			error;
1276
1277retry:
1278	if (src < dest) {
1279		inode_lock_shared(src);
1280		inode_lock_nested(dest, I_MUTEX_NONDIR2);
1281	} else {
1282		/* src >= dest */
1283		inode_lock(dest);
1284	}
1285
1286	error = break_layout(dest, false);
1287	if (error == -EWOULDBLOCK) {
1288		inode_unlock(dest);
1289		if (src < dest)
1290			inode_unlock_shared(src);
1291		error = break_layout(dest, true);
1292		if (error)
1293			return error;
1294		goto retry;
1295	}
1296	if (error) {
1297		inode_unlock(dest);
1298		if (src < dest)
1299			inode_unlock_shared(src);
1300		return error;
1301	}
1302	if (src > dest)
1303		inode_lock_shared_nested(src, I_MUTEX_NONDIR2);
1304	return 0;
1305}
1306
1307/*
1308 * Link a range of blocks from one file to another.
1309 */
1310int
1311xfs_reflink_remap_range(
1312	struct file		*file_in,
1313	loff_t			pos_in,
1314	struct file		*file_out,
1315	loff_t			pos_out,
1316	u64			len,
1317	bool			is_dedupe)
1318{
1319	struct inode		*inode_in = file_inode(file_in);
1320	struct xfs_inode	*src = XFS_I(inode_in);
1321	struct inode		*inode_out = file_inode(file_out);
1322	struct xfs_inode	*dest = XFS_I(inode_out);
1323	struct xfs_mount	*mp = src->i_mount;
1324	bool			same_inode = (inode_in == inode_out);
1325	xfs_fileoff_t		sfsbno, dfsbno;
1326	xfs_filblks_t		fsblen;
1327	xfs_extlen_t		cowextsize;
1328	ssize_t			ret;
1329
1330	if (!xfs_sb_version_hasreflink(&mp->m_sb))
1331		return -EOPNOTSUPP;
1332
1333	if (XFS_FORCED_SHUTDOWN(mp))
1334		return -EIO;
1335
1336	/* Lock both files against IO */
1337	ret = xfs_iolock_two_inodes_and_break_layout(inode_in, inode_out);
1338	if (ret)
1339		return ret;
1340	if (same_inode)
1341		xfs_ilock(src, XFS_MMAPLOCK_EXCL);
1342	else
1343		xfs_lock_two_inodes(src, XFS_MMAPLOCK_SHARED, dest,
1344				XFS_MMAPLOCK_EXCL);
1345
1346	/* Check file eligibility and prepare for block sharing. */
1347	ret = -EINVAL;
1348	/* Don't reflink realtime inodes */
1349	if (XFS_IS_REALTIME_INODE(src) || XFS_IS_REALTIME_INODE(dest))
1350		goto out_unlock;
1351
1352	/* Don't share DAX file data for now. */
1353	if (IS_DAX(inode_in) || IS_DAX(inode_out))
1354		goto out_unlock;
1355
1356	ret = vfs_clone_file_prep_inodes(inode_in, pos_in, inode_out, pos_out,
1357			&len, is_dedupe);
1358	if (ret <= 0)
1359		goto out_unlock;
1360
1361	/* Attach dquots to dest inode before changing block map */
1362	ret = xfs_qm_dqattach(dest, 0);
1363	if (ret)
1364		goto out_unlock;
1365
1366	trace_xfs_reflink_remap_range(src, pos_in, len, dest, pos_out);
1367
1368	/*
1369	 * Clear out post-eof preallocations because we don't have page cache
1370	 * backing the delayed allocations and they'll never get freed on
1371	 * their own.
1372	 */
1373	if (xfs_can_free_eofblocks(dest, true)) {
1374		ret = xfs_free_eofblocks(dest);
1375		if (ret)
1376			goto out_unlock;
1377	}
1378
1379	/* Set flags and remap blocks. */
1380	ret = xfs_reflink_set_inode_flag(src, dest);
1381	if (ret)
1382		goto out_unlock;
1383
1384	dfsbno = XFS_B_TO_FSBT(mp, pos_out);
1385	sfsbno = XFS_B_TO_FSBT(mp, pos_in);
1386	fsblen = XFS_B_TO_FSB(mp, len);
1387	ret = xfs_reflink_remap_blocks(src, sfsbno, dest, dfsbno, fsblen,
1388			pos_out + len);
1389	if (ret)
1390		goto out_unlock;
1391
1392	/* Zap any page cache for the destination file's range. */
1393	truncate_inode_pages_range(&inode_out->i_data, pos_out,
1394				   PAGE_ALIGN(pos_out + len) - 1);
1395
1396	/*
1397	 * Carry the cowextsize hint from src to dest if we're sharing the
1398	 * entire source file to the entire destination file, the source file
1399	 * has a cowextsize hint, and the destination file does not.
1400	 */
1401	cowextsize = 0;
1402	if (pos_in == 0 && len == i_size_read(inode_in) &&
1403	    (src->i_d.di_flags2 & XFS_DIFLAG2_COWEXTSIZE) &&
1404	    pos_out == 0 && len >= i_size_read(inode_out) &&
1405	    !(dest->i_d.di_flags2 & XFS_DIFLAG2_COWEXTSIZE))
1406		cowextsize = src->i_d.di_cowextsize;
1407
1408	ret = xfs_reflink_update_dest(dest, pos_out + len, cowextsize,
1409			is_dedupe);
1410
1411out_unlock:
1412	xfs_iunlock(dest, XFS_MMAPLOCK_EXCL);
1413	if (!same_inode)
1414		xfs_iunlock(src, XFS_MMAPLOCK_SHARED);
1415	inode_unlock(inode_out);
1416	if (!same_inode)
1417		inode_unlock_shared(inode_in);
1418	if (ret)
1419		trace_xfs_reflink_remap_range_error(dest, ret, _RET_IP_);
1420	return ret;
1421}
1422
1423/*
1424 * The user wants to preemptively CoW all shared blocks in this file,
1425 * which enables us to turn off the reflink flag.  Iterate all
1426 * extents which are not prealloc/delalloc to see which ranges are
1427 * mentioned in the refcount tree, then read those blocks into the
1428 * pagecache, dirty them, fsync them back out, and then we can update
1429 * the inode flag.  What happens if we run out of memory? :)
1430 */
1431STATIC int
1432xfs_reflink_dirty_extents(
1433	struct xfs_inode	*ip,
1434	xfs_fileoff_t		fbno,
1435	xfs_filblks_t		end,
1436	xfs_off_t		isize)
1437{
1438	struct xfs_mount	*mp = ip->i_mount;
1439	xfs_agnumber_t		agno;
1440	xfs_agblock_t		agbno;
1441	xfs_extlen_t		aglen;
1442	xfs_agblock_t		rbno;
1443	xfs_extlen_t		rlen;
1444	xfs_off_t		fpos;
1445	xfs_off_t		flen;
1446	struct xfs_bmbt_irec	map[2];
1447	int			nmaps;
1448	int			error = 0;
1449
1450	while (end - fbno > 0) {
1451		nmaps = 1;
1452		/*
1453		 * Look for extents in the file.  Skip holes, delalloc, or
1454		 * unwritten extents; they can't be reflinked.
1455		 */
1456		error = xfs_bmapi_read(ip, fbno, end - fbno, map, &nmaps, 0);
1457		if (error)
1458			goto out;
1459		if (nmaps == 0)
1460			break;
1461		if (!xfs_bmap_is_real_extent(&map[0]))
1462			goto next;
1463
1464		map[1] = map[0];
1465		while (map[1].br_blockcount) {
1466			agno = XFS_FSB_TO_AGNO(mp, map[1].br_startblock);
1467			agbno = XFS_FSB_TO_AGBNO(mp, map[1].br_startblock);
1468			aglen = map[1].br_blockcount;
1469
1470			error = xfs_reflink_find_shared(mp, NULL, agno, agbno,
1471					aglen, &rbno, &rlen, true);
1472			if (error)
1473				goto out;
1474			if (rbno == NULLAGBLOCK)
1475				break;
1476
1477			/* Dirty the pages */
1478			xfs_iunlock(ip, XFS_ILOCK_EXCL);
1479			fpos = XFS_FSB_TO_B(mp, map[1].br_startoff +
1480					(rbno - agbno));
1481			flen = XFS_FSB_TO_B(mp, rlen);
1482			if (fpos + flen > isize)
1483				flen = isize - fpos;
1484			error = iomap_file_dirty(VFS_I(ip), fpos, flen,
1485					&xfs_iomap_ops);
1486			xfs_ilock(ip, XFS_ILOCK_EXCL);
1487			if (error)
1488				goto out;
1489
1490			map[1].br_blockcount -= (rbno - agbno + rlen);
1491			map[1].br_startoff += (rbno - agbno + rlen);
1492			map[1].br_startblock += (rbno - agbno + rlen);
1493		}
1494
1495next:
1496		fbno = map[0].br_startoff + map[0].br_blockcount;
1497	}
1498out:
1499	return error;
1500}
1501
1502/* Does this inode need the reflink flag? */
1503int
1504xfs_reflink_inode_has_shared_extents(
1505	struct xfs_trans		*tp,
1506	struct xfs_inode		*ip,
1507	bool				*has_shared)
1508{
1509	struct xfs_bmbt_irec		got;
1510	struct xfs_mount		*mp = ip->i_mount;
1511	struct xfs_ifork		*ifp;
1512	xfs_agnumber_t			agno;
1513	xfs_agblock_t			agbno;
1514	xfs_extlen_t			aglen;
1515	xfs_agblock_t			rbno;
1516	xfs_extlen_t			rlen;
1517	struct xfs_iext_cursor		icur;
1518	bool				found;
1519	int				error;
1520
1521	ifp = XFS_IFORK_PTR(ip, XFS_DATA_FORK);
1522	if (!(ifp->if_flags & XFS_IFEXTENTS)) {
1523		error = xfs_iread_extents(tp, ip, XFS_DATA_FORK);
1524		if (error)
1525			return error;
1526	}
1527
1528	*has_shared = false;
1529	found = xfs_iext_lookup_extent(ip, ifp, 0, &icur, &got);
1530	while (found) {
1531		if (isnullstartblock(got.br_startblock) ||
1532		    got.br_state != XFS_EXT_NORM)
1533			goto next;
1534		agno = XFS_FSB_TO_AGNO(mp, got.br_startblock);
1535		agbno = XFS_FSB_TO_AGBNO(mp, got.br_startblock);
1536		aglen = got.br_blockcount;
1537
1538		error = xfs_reflink_find_shared(mp, tp, agno, agbno, aglen,
1539				&rbno, &rlen, false);
1540		if (error)
1541			return error;
1542		/* Is there still a shared block here? */
1543		if (rbno != NULLAGBLOCK) {
1544			*has_shared = true;
1545			return 0;
1546		}
1547next:
1548		found = xfs_iext_next_extent(ifp, &icur, &got);
1549	}
1550
1551	return 0;
1552}
1553
1554/* Clear the inode reflink flag if there are no shared extents. */
1555int
1556xfs_reflink_clear_inode_flag(
1557	struct xfs_inode	*ip,
1558	struct xfs_trans	**tpp)
1559{
1560	bool			needs_flag;
1561	int			error = 0;
1562
1563	ASSERT(xfs_is_reflink_inode(ip));
1564
1565	error = xfs_reflink_inode_has_shared_extents(*tpp, ip, &needs_flag);
1566	if (error || needs_flag)
1567		return error;
1568
1569	/*
1570	 * We didn't find any shared blocks so turn off the reflink flag.
1571	 * First, get rid of any leftover CoW mappings.
1572	 */
1573	error = xfs_reflink_cancel_cow_blocks(ip, tpp, 0, NULLFILEOFF, true);
1574	if (error)
1575		return error;
1576
1577	/* Clear the inode flag. */
1578	trace_xfs_reflink_unset_inode_flag(ip);
1579	ip->i_d.di_flags2 &= ~XFS_DIFLAG2_REFLINK;
1580	xfs_inode_clear_cowblocks_tag(ip);
1581	xfs_trans_ijoin(*tpp, ip, 0);
1582	xfs_trans_log_inode(*tpp, ip, XFS_ILOG_CORE);
1583
1584	return error;
1585}
1586
1587/*
1588 * Clear the inode reflink flag if there are no shared extents and the size
1589 * hasn't changed.
1590 */
1591STATIC int
1592xfs_reflink_try_clear_inode_flag(
1593	struct xfs_inode	*ip)
1594{
1595	struct xfs_mount	*mp = ip->i_mount;
1596	struct xfs_trans	*tp;
1597	int			error = 0;
1598
1599	/* Start a rolling transaction to remove the mappings */
1600	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, 0, 0, 0, &tp);
1601	if (error)
1602		return error;
1603
1604	xfs_ilock(ip, XFS_ILOCK_EXCL);
1605	xfs_trans_ijoin(tp, ip, 0);
1606
1607	error = xfs_reflink_clear_inode_flag(ip, &tp);
1608	if (error)
1609		goto cancel;
1610
1611	error = xfs_trans_commit(tp);
1612	if (error)
1613		goto out;
1614
1615	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1616	return 0;
1617cancel:
1618	xfs_trans_cancel(tp);
1619out:
1620	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1621	return error;
1622}
1623
1624/*
1625 * Pre-COW all shared blocks within a given byte range of a file and turn off
1626 * the reflink flag if we unshare all of the file's blocks.
1627 */
1628int
1629xfs_reflink_unshare(
1630	struct xfs_inode	*ip,
1631	xfs_off_t		offset,
1632	xfs_off_t		len)
1633{
1634	struct xfs_mount	*mp = ip->i_mount;
1635	xfs_fileoff_t		fbno;
1636	xfs_filblks_t		end;
1637	xfs_off_t		isize;
1638	int			error;
1639
1640	if (!xfs_is_reflink_inode(ip))
1641		return 0;
1642
1643	trace_xfs_reflink_unshare(ip, offset, len);
1644
1645	inode_dio_wait(VFS_I(ip));
1646
1647	/* Try to CoW the selected ranges */
1648	xfs_ilock(ip, XFS_ILOCK_EXCL);
1649	fbno = XFS_B_TO_FSBT(mp, offset);
1650	isize = i_size_read(VFS_I(ip));
1651	end = XFS_B_TO_FSB(mp, offset + len);
1652	error = xfs_reflink_dirty_extents(ip, fbno, end, isize);
1653	if (error)
1654		goto out_unlock;
1655	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1656
1657	/* Wait for the IO to finish */
1658	error = filemap_write_and_wait(VFS_I(ip)->i_mapping);
1659	if (error)
1660		goto out;
1661
1662	/* Turn off the reflink flag if possible. */
1663	error = xfs_reflink_try_clear_inode_flag(ip);
1664	if (error)
1665		goto out;
1666
1667	return 0;
1668
1669out_unlock:
1670	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1671out:
1672	trace_xfs_reflink_unshare_error(ip, error, _RET_IP_);
1673	return error;
1674}