Linux Audio

Check our new training course

Loading...
v4.10.11
 
   1/*
   2 * Copyright (C) 2016 Oracle.  All Rights Reserved.
   3 *
   4 * Author: Darrick J. Wong <darrick.wong@oracle.com>
   5 *
   6 * This program is free software; you can redistribute it and/or
   7 * modify it under the terms of the GNU General Public License
   8 * as published by the Free Software Foundation; either version 2
   9 * of the License, or (at your option) any later version.
  10 *
  11 * This program is distributed in the hope that it would be useful,
  12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  14 * GNU General Public License for more details.
  15 *
  16 * You should have received a copy of the GNU General Public License
  17 * along with this program; if not, write the Free Software Foundation,
  18 * Inc.,  51 Franklin St, Fifth Floor, Boston, MA  02110-1301, USA.
  19 */
  20#include "xfs.h"
  21#include "xfs_fs.h"
  22#include "xfs_shared.h"
  23#include "xfs_format.h"
  24#include "xfs_log_format.h"
  25#include "xfs_trans_resv.h"
  26#include "xfs_mount.h"
  27#include "xfs_defer.h"
  28#include "xfs_da_format.h"
  29#include "xfs_da_btree.h"
  30#include "xfs_inode.h"
  31#include "xfs_trans.h"
  32#include "xfs_inode_item.h"
  33#include "xfs_bmap.h"
  34#include "xfs_bmap_util.h"
  35#include "xfs_error.h"
  36#include "xfs_dir2.h"
  37#include "xfs_dir2_priv.h"
  38#include "xfs_ioctl.h"
  39#include "xfs_trace.h"
  40#include "xfs_log.h"
  41#include "xfs_icache.h"
  42#include "xfs_pnfs.h"
  43#include "xfs_btree.h"
  44#include "xfs_refcount_btree.h"
  45#include "xfs_refcount.h"
  46#include "xfs_bmap_btree.h"
  47#include "xfs_trans_space.h"
  48#include "xfs_bit.h"
  49#include "xfs_alloc.h"
  50#include "xfs_quota_defs.h"
  51#include "xfs_quota.h"
  52#include "xfs_btree.h"
  53#include "xfs_bmap_btree.h"
  54#include "xfs_reflink.h"
  55#include "xfs_iomap.h"
  56#include "xfs_rmap_btree.h"
  57#include "xfs_sb.h"
  58#include "xfs_ag_resv.h"
  59
  60/*
  61 * Copy on Write of Shared Blocks
  62 *
  63 * XFS must preserve "the usual" file semantics even when two files share
  64 * the same physical blocks.  This means that a write to one file must not
  65 * alter the blocks in a different file; the way that we'll do that is
  66 * through the use of a copy-on-write mechanism.  At a high level, that
  67 * means that when we want to write to a shared block, we allocate a new
  68 * block, write the data to the new block, and if that succeeds we map the
  69 * new block into the file.
  70 *
  71 * XFS provides a "delayed allocation" mechanism that defers the allocation
  72 * of disk blocks to dirty-but-not-yet-mapped file blocks as long as
  73 * possible.  This reduces fragmentation by enabling the filesystem to ask
  74 * for bigger chunks less often, which is exactly what we want for CoW.
  75 *
  76 * The delalloc mechanism begins when the kernel wants to make a block
  77 * writable (write_begin or page_mkwrite).  If the offset is not mapped, we
  78 * create a delalloc mapping, which is a regular in-core extent, but without
  79 * a real startblock.  (For delalloc mappings, the startblock encodes both
  80 * a flag that this is a delalloc mapping, and a worst-case estimate of how
  81 * many blocks might be required to put the mapping into the BMBT.)  delalloc
  82 * mappings are a reservation against the free space in the filesystem;
  83 * adjacent mappings can also be combined into fewer larger mappings.
  84 *
  85 * As an optimization, the CoW extent size hint (cowextsz) creates
  86 * outsized aligned delalloc reservations in the hope of landing out of
  87 * order nearby CoW writes in a single extent on disk, thereby reducing
  88 * fragmentation and improving future performance.
  89 *
  90 * D: --RRRRRRSSSRRRRRRRR--- (data fork)
  91 * C: ------DDDDDDD--------- (CoW fork)
  92 *
  93 * When dirty pages are being written out (typically in writepage), the
  94 * delalloc reservations are converted into unwritten mappings by
  95 * allocating blocks and replacing the delalloc mapping with real ones.
  96 * A delalloc mapping can be replaced by several unwritten ones if the
  97 * free space is fragmented.
  98 *
  99 * D: --RRRRRRSSSRRRRRRRR---
 100 * C: ------UUUUUUU---------
 101 *
 102 * We want to adapt the delalloc mechanism for copy-on-write, since the
 103 * write paths are similar.  The first two steps (creating the reservation
 104 * and allocating the blocks) are exactly the same as delalloc except that
 105 * the mappings must be stored in a separate CoW fork because we do not want
 106 * to disturb the mapping in the data fork until we're sure that the write
 107 * succeeded.  IO completion in this case is the process of removing the old
 108 * mapping from the data fork and moving the new mapping from the CoW fork to
 109 * the data fork.  This will be discussed shortly.
 110 *
 111 * For now, unaligned directio writes will be bounced back to the page cache.
 112 * Block-aligned directio writes will use the same mechanism as buffered
 113 * writes.
 114 *
 115 * Just prior to submitting the actual disk write requests, we convert
 116 * the extents representing the range of the file actually being written
 117 * (as opposed to extra pieces created for the cowextsize hint) to real
 118 * extents.  This will become important in the next step:
 119 *
 120 * D: --RRRRRRSSSRRRRRRRR---
 121 * C: ------UUrrUUU---------
 122 *
 123 * CoW remapping must be done after the data block write completes,
 124 * because we don't want to destroy the old data fork map until we're sure
 125 * the new block has been written.  Since the new mappings are kept in a
 126 * separate fork, we can simply iterate these mappings to find the ones
 127 * that cover the file blocks that we just CoW'd.  For each extent, simply
 128 * unmap the corresponding range in the data fork, map the new range into
 129 * the data fork, and remove the extent from the CoW fork.  Because of
 130 * the presence of the cowextsize hint, however, we must be careful
 131 * only to remap the blocks that we've actually written out --  we must
 132 * never remap delalloc reservations nor CoW staging blocks that have
 133 * yet to be written.  This corresponds exactly to the real extents in
 134 * the CoW fork:
 135 *
 136 * D: --RRRRRRrrSRRRRRRRR---
 137 * C: ------UU--UUU---------
 138 *
 139 * Since the remapping operation can be applied to an arbitrary file
 140 * range, we record the need for the remap step as a flag in the ioend
 141 * instead of declaring a new IO type.  This is required for direct io
 142 * because we only have ioend for the whole dio, and we have to be able to
 143 * remember the presence of unwritten blocks and CoW blocks with a single
 144 * ioend structure.  Better yet, the more ground we can cover with one
 145 * ioend, the better.
 146 */
 147
 148/*
 149 * Given an AG extent, find the lowest-numbered run of shared blocks
 150 * within that range and return the range in fbno/flen.  If
 151 * find_end_of_shared is true, return the longest contiguous extent of
 152 * shared blocks.  If there are no shared extents, fbno and flen will
 153 * be set to NULLAGBLOCK and 0, respectively.
 154 */
 155int
 156xfs_reflink_find_shared(
 157	struct xfs_mount	*mp,
 
 158	xfs_agnumber_t		agno,
 159	xfs_agblock_t		agbno,
 160	xfs_extlen_t		aglen,
 161	xfs_agblock_t		*fbno,
 162	xfs_extlen_t		*flen,
 163	bool			find_end_of_shared)
 164{
 165	struct xfs_buf		*agbp;
 166	struct xfs_btree_cur	*cur;
 167	int			error;
 168
 169	error = xfs_alloc_read_agf(mp, NULL, agno, 0, &agbp);
 170	if (error)
 171		return error;
 
 
 172
 173	cur = xfs_refcountbt_init_cursor(mp, NULL, agbp, agno, NULL);
 174
 175	error = xfs_refcount_find_shared(cur, agbno, aglen, fbno, flen,
 176			find_end_of_shared);
 177
 178	xfs_btree_del_cursor(cur, error ? XFS_BTREE_ERROR : XFS_BTREE_NOERROR);
 179
 180	xfs_buf_relse(agbp);
 181	return error;
 182}
 183
 184/*
 185 * Trim the mapping to the next block where there's a change in the
 186 * shared/unshared status.  More specifically, this means that we
 187 * find the lowest-numbered extent of shared blocks that coincides with
 188 * the given block mapping.  If the shared extent overlaps the start of
 189 * the mapping, trim the mapping to the end of the shared extent.  If
 190 * the shared region intersects the mapping, trim the mapping to the
 191 * start of the shared extent.  If there are no shared regions that
 192 * overlap, just return the original extent.
 193 */
 194int
 195xfs_reflink_trim_around_shared(
 196	struct xfs_inode	*ip,
 197	struct xfs_bmbt_irec	*irec,
 198	bool			*shared,
 199	bool			*trimmed)
 200{
 201	xfs_agnumber_t		agno;
 202	xfs_agblock_t		agbno;
 203	xfs_extlen_t		aglen;
 204	xfs_agblock_t		fbno;
 205	xfs_extlen_t		flen;
 206	int			error = 0;
 207
 208	/* Holes, unwritten, and delalloc extents cannot be shared */
 209	if (!xfs_is_reflink_inode(ip) ||
 210	    ISUNWRITTEN(irec) ||
 211	    irec->br_startblock == HOLESTARTBLOCK ||
 212	    irec->br_startblock == DELAYSTARTBLOCK ||
 213	    isnullstartblock(irec->br_startblock)) {
 214		*shared = false;
 215		return 0;
 216	}
 217
 218	trace_xfs_reflink_trim_around_shared(ip, irec);
 219
 220	agno = XFS_FSB_TO_AGNO(ip->i_mount, irec->br_startblock);
 221	agbno = XFS_FSB_TO_AGBNO(ip->i_mount, irec->br_startblock);
 222	aglen = irec->br_blockcount;
 223
 224	error = xfs_reflink_find_shared(ip->i_mount, agno, agbno,
 225			aglen, &fbno, &flen, true);
 226	if (error)
 227		return error;
 228
 229	*shared = *trimmed = false;
 230	if (fbno == NULLAGBLOCK) {
 231		/* No shared blocks at all. */
 232		return 0;
 233	} else if (fbno == agbno) {
 234		/*
 235		 * The start of this extent is shared.  Truncate the
 236		 * mapping at the end of the shared region so that a
 237		 * subsequent iteration starts at the start of the
 238		 * unshared region.
 239		 */
 240		irec->br_blockcount = flen;
 241		*shared = true;
 242		if (flen != aglen)
 243			*trimmed = true;
 244		return 0;
 245	} else {
 246		/*
 247		 * There's a shared extent midway through this extent.
 248		 * Truncate the mapping at the start of the shared
 249		 * extent so that a subsequent iteration starts at the
 250		 * start of the shared region.
 251		 */
 252		irec->br_blockcount = fbno - agbno;
 253		*trimmed = true;
 254		return 0;
 255	}
 256}
 257
 258/*
 259 * Trim the passed in imap to the next shared/unshared extent boundary, and
 260 * if imap->br_startoff points to a shared extent reserve space for it in the
 261 * COW fork.  In this case *shared is set to true, else to false.
 262 *
 263 * Note that imap will always contain the block numbers for the existing blocks
 264 * in the data fork, as the upper layers need them for read-modify-write
 265 * operations.
 266 */
 267int
 268xfs_reflink_reserve_cow(
 269	struct xfs_inode	*ip,
 270	struct xfs_bmbt_irec	*imap,
 271	bool			*shared)
 272{
 273	struct xfs_ifork	*ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
 274	struct xfs_bmbt_irec	got;
 275	int			error = 0;
 276	bool			eof = false, trimmed;
 277	xfs_extnum_t		idx;
 278
 279	/*
 280	 * Search the COW fork extent list first.  This serves two purposes:
 281	 * first this implement the speculative preallocation using cowextisze,
 282	 * so that we also unshared block adjacent to shared blocks instead
 283	 * of just the shared blocks themselves.  Second the lookup in the
 284	 * extent list is generally faster than going out to the shared extent
 285	 * tree.
 286	 */
 287
 288	if (!xfs_iext_lookup_extent(ip, ifp, imap->br_startoff, &idx, &got))
 289		eof = true;
 290	if (!eof && got.br_startoff <= imap->br_startoff) {
 291		trace_xfs_reflink_cow_found(ip, imap);
 292		xfs_trim_extent(imap, got.br_startoff, got.br_blockcount);
 293
 294		*shared = true;
 295		return 0;
 296	}
 297
 298	/* Trim the mapping to the nearest shared extent boundary. */
 299	error = xfs_reflink_trim_around_shared(ip, imap, shared, &trimmed);
 300	if (error)
 301		return error;
 302
 303	/* Not shared?  Just report the (potentially capped) extent. */
 304	if (!*shared)
 305		return 0;
 306
 307	/*
 308	 * Fork all the shared blocks from our write offset until the end of
 309	 * the extent.
 310	 */
 311	error = xfs_qm_dqattach_locked(ip, 0);
 312	if (error)
 313		return error;
 314
 315	error = xfs_bmapi_reserve_delalloc(ip, XFS_COW_FORK, imap->br_startoff,
 316			imap->br_blockcount, 0, &got, &idx, eof);
 317	if (error == -ENOSPC || error == -EDQUOT)
 318		trace_xfs_reflink_cow_enospc(ip, imap);
 319	if (error)
 320		return error;
 321
 322	trace_xfs_reflink_cow_alloc(ip, &got);
 323	return 0;
 324}
 325
 326/* Convert part of an unwritten CoW extent to a real one. */
 327STATIC int
 328xfs_reflink_convert_cow_extent(
 329	struct xfs_inode		*ip,
 330	struct xfs_bmbt_irec		*imap,
 331	xfs_fileoff_t			offset_fsb,
 332	xfs_filblks_t			count_fsb,
 333	struct xfs_defer_ops		*dfops)
 334{
 335	struct xfs_bmbt_irec		irec = *imap;
 336	xfs_fsblock_t			first_block;
 337	int				nimaps = 1;
 
 
 338
 339	if (imap->br_state == XFS_EXT_NORM)
 340		return 0;
 341
 342	xfs_trim_extent(&irec, offset_fsb, count_fsb);
 343	trace_xfs_reflink_convert_cow(ip, &irec);
 344	if (irec.br_blockcount == 0)
 345		return 0;
 346	return xfs_bmapi_write(NULL, ip, irec.br_startoff, irec.br_blockcount,
 347			XFS_BMAPI_COWFORK | XFS_BMAPI_CONVERT, &first_block,
 348			0, &irec, &nimaps, dfops);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 349}
 350
 351/* Convert all of the unwritten CoW extents in a file's range to real ones. */
 352int
 353xfs_reflink_convert_cow(
 354	struct xfs_inode	*ip,
 355	xfs_off_t		offset,
 356	xfs_off_t		count)
 357{
 358	struct xfs_bmbt_irec	got;
 359	struct xfs_defer_ops	dfops;
 360	struct xfs_mount	*mp = ip->i_mount;
 361	struct xfs_ifork	*ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
 362	xfs_fileoff_t		offset_fsb = XFS_B_TO_FSBT(mp, offset);
 363	xfs_fileoff_t		end_fsb = XFS_B_TO_FSB(mp, offset + count);
 364	xfs_extnum_t		idx;
 365	bool			found;
 366	int			error = 0;
 
 367
 368	xfs_ilock(ip, XFS_ILOCK_EXCL);
 
 
 
 
 369
 370	/* Convert all the extents to real from unwritten. */
 371	for (found = xfs_iext_lookup_extent(ip, ifp, offset_fsb, &idx, &got);
 372	     found && got.br_startoff < end_fsb;
 373	     found = xfs_iext_get_extent(ifp, ++idx, &got)) {
 374		error = xfs_reflink_convert_cow_extent(ip, &got, offset_fsb,
 375				end_fsb - offset_fsb, &dfops);
 376		if (error)
 377			break;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 378	}
 379
 380	/* Finish up. */
 381	xfs_iunlock(ip, XFS_ILOCK_EXCL);
 382	return error;
 
 
 
 
 
 
 
 
 383}
 384
 385/* Allocate all CoW reservations covering a range of blocks in a file. */
 386static int
 387__xfs_reflink_allocate_cow(
 388	struct xfs_inode	*ip,
 389	xfs_fileoff_t		*offset_fsb,
 390	xfs_fileoff_t		end_fsb)
 
 
 391{
 392	struct xfs_mount	*mp = ip->i_mount;
 393	struct xfs_bmbt_irec	imap;
 394	struct xfs_defer_ops	dfops;
 395	struct xfs_trans	*tp;
 396	xfs_fsblock_t		first_block;
 397	int			nimaps = 1, error;
 398	bool			shared;
 
 399
 400	xfs_defer_init(&dfops, &first_block);
 
 
 
 
 401
 402	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, 0, 0,
 403			XFS_TRANS_RESERVE, &tp);
 404	if (error)
 405		return error;
 
 
 406
 407	xfs_ilock(ip, XFS_ILOCK_EXCL);
 
 
 
 
 
 
 
 408
 409	/* Read extent from the source file. */
 410	nimaps = 1;
 411	error = xfs_bmapi_read(ip, *offset_fsb, end_fsb - *offset_fsb,
 412			&imap, &nimaps, 0);
 413	if (error)
 414		goto out_unlock;
 415	ASSERT(nimaps == 1);
 416
 417	/* Make sure there's a CoW reservation for it. */
 418	error = xfs_reflink_reserve_cow(ip, &imap, &shared);
 419	if (error)
 420		goto out_trans_cancel;
 421
 422	if (!shared) {
 423		*offset_fsb = imap.br_startoff + imap.br_blockcount;
 
 
 
 424		goto out_trans_cancel;
 
 
 
 425	}
 426
 427	/* Allocate the entire reservation as unwritten blocks. */
 428	xfs_trans_ijoin(tp, ip, 0);
 429	error = xfs_bmapi_write(tp, ip, imap.br_startoff, imap.br_blockcount,
 430			XFS_BMAPI_COWFORK | XFS_BMAPI_PREALLOC, &first_block,
 431			XFS_EXTENTADD_SPACE_RES(mp, XFS_DATA_FORK),
 432			&imap, &nimaps, &dfops);
 433	if (error)
 434		goto out_trans_cancel;
 435
 436	/* Finish up. */
 437	error = xfs_defer_finish(&tp, &dfops, NULL);
 
 
 
 
 
 438	if (error)
 439		goto out_trans_cancel;
 440
 
 441	error = xfs_trans_commit(tp);
 442
 443	*offset_fsb = imap.br_startoff + imap.br_blockcount;
 444out_unlock:
 445	xfs_iunlock(ip, XFS_ILOCK_EXCL);
 446	return error;
 447out_trans_cancel:
 448	xfs_defer_cancel(&dfops);
 449	xfs_trans_cancel(tp);
 450	goto out_unlock;
 451}
 452
 453/* Allocate all CoW reservations covering a part of a file. */
 454int
 455xfs_reflink_allocate_cow_range(
 456	struct xfs_inode	*ip,
 457	xfs_off_t		offset,
 458	xfs_off_t		count)
 459{
 460	struct xfs_mount	*mp = ip->i_mount;
 461	xfs_fileoff_t		offset_fsb = XFS_B_TO_FSBT(mp, offset);
 462	xfs_fileoff_t		end_fsb = XFS_B_TO_FSB(mp, offset + count);
 463	int			error;
 464
 465	ASSERT(xfs_is_reflink_inode(ip));
 466
 467	trace_xfs_reflink_allocate_cow_range(ip, offset, count);
 468
 469	/*
 470	 * Make sure that the dquots are there.
 471	 */
 472	error = xfs_qm_dqattach(ip, 0);
 473	if (error)
 474		return error;
 475
 476	while (offset_fsb < end_fsb) {
 477		error = __xfs_reflink_allocate_cow(ip, &offset_fsb, end_fsb);
 478		if (error) {
 479			trace_xfs_reflink_allocate_cow_range_error(ip, error,
 480					_RET_IP_);
 481			return error;
 482		}
 483	}
 484
 485	/* Convert the CoW extents to regular. */
 486	return xfs_reflink_convert_cow(ip, offset, count);
 487}
 488
 489/*
 490 * Find the CoW reservation for a given byte offset of a file.
 491 */
 492bool
 493xfs_reflink_find_cow_mapping(
 494	struct xfs_inode		*ip,
 495	xfs_off_t			offset,
 496	struct xfs_bmbt_irec		*imap)
 497{
 498	struct xfs_ifork		*ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
 499	xfs_fileoff_t			offset_fsb;
 500	struct xfs_bmbt_irec		got;
 501	xfs_extnum_t			idx;
 502
 503	ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL | XFS_ILOCK_SHARED));
 504	ASSERT(xfs_is_reflink_inode(ip));
 505
 506	offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
 507	if (!xfs_iext_lookup_extent(ip, ifp, offset_fsb, &idx, &got))
 508		return false;
 509	if (got.br_startoff > offset_fsb)
 510		return false;
 511
 512	trace_xfs_reflink_find_cow_mapping(ip, offset, 1, XFS_IO_OVERWRITE,
 513			&got);
 514	*imap = got;
 515	return true;
 516}
 517
 518/*
 519 * Trim an extent to end at the next CoW reservation past offset_fsb.
 520 */
 521void
 522xfs_reflink_trim_irec_to_next_cow(
 523	struct xfs_inode		*ip,
 524	xfs_fileoff_t			offset_fsb,
 525	struct xfs_bmbt_irec		*imap)
 526{
 527	struct xfs_ifork		*ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
 528	struct xfs_bmbt_irec		got;
 529	xfs_extnum_t			idx;
 530
 531	if (!xfs_is_reflink_inode(ip))
 532		return;
 533
 534	/* Find the extent in the CoW fork. */
 535	if (!xfs_iext_lookup_extent(ip, ifp, offset_fsb, &idx, &got))
 536		return;
 537
 538	/* This is the extent before; try sliding up one. */
 539	if (got.br_startoff < offset_fsb) {
 540		if (!xfs_iext_get_extent(ifp, idx + 1, &got))
 541			return;
 542	}
 543
 544	if (got.br_startoff >= imap->br_startoff + imap->br_blockcount)
 545		return;
 546
 547	imap->br_blockcount = got.br_startoff - imap->br_startoff;
 548	trace_xfs_reflink_trim_irec(ip, imap);
 
 
 
 
 549}
 550
 551/*
 552 * Cancel CoW reservations for some block range of an inode.
 553 *
 554 * If cancel_real is true this function cancels all COW fork extents for the
 555 * inode; if cancel_real is false, real extents are not cleared.
 
 
 
 556 */
 557int
 558xfs_reflink_cancel_cow_blocks(
 559	struct xfs_inode		*ip,
 560	struct xfs_trans		**tpp,
 561	xfs_fileoff_t			offset_fsb,
 562	xfs_fileoff_t			end_fsb,
 563	bool				cancel_real)
 564{
 565	struct xfs_ifork		*ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
 566	struct xfs_bmbt_irec		got, del;
 567	xfs_extnum_t			idx;
 568	xfs_fsblock_t			firstfsb;
 569	struct xfs_defer_ops		dfops;
 570	int				error = 0;
 571
 572	if (!xfs_is_reflink_inode(ip))
 573		return 0;
 574	if (!xfs_iext_lookup_extent(ip, ifp, offset_fsb, &idx, &got))
 575		return 0;
 576
 577	while (got.br_startoff < end_fsb) {
 
 578		del = got;
 579		xfs_trim_extent(&del, offset_fsb, end_fsb - offset_fsb);
 
 
 
 
 
 
 
 580		trace_xfs_reflink_cancel_cow(ip, &del);
 581
 582		if (isnullstartblock(del.br_startblock)) {
 583			error = xfs_bmap_del_extent_delay(ip, XFS_COW_FORK,
 584					&idx, &got, &del);
 585			if (error)
 586				break;
 587		} else if (del.br_state == XFS_EXT_UNWRITTEN || cancel_real) {
 588			xfs_trans_ijoin(*tpp, ip, 0);
 589			xfs_defer_init(&dfops, &firstfsb);
 590
 591			/* Free the CoW orphan record. */
 592			error = xfs_refcount_free_cow_extent(ip->i_mount,
 593					&dfops, del.br_startblock,
 594					del.br_blockcount);
 595			if (error)
 596				break;
 597
 598			xfs_bmap_add_free(ip->i_mount, &dfops,
 599					del.br_startblock, del.br_blockcount,
 600					NULL);
 601
 602			/* Update quota accounting */
 603			xfs_trans_mod_dquot_byino(*tpp, ip, XFS_TRANS_DQ_BCOUNT,
 604					-(long)del.br_blockcount);
 605
 606			/* Roll the transaction */
 607			error = xfs_defer_finish(tpp, &dfops, ip);
 608			if (error) {
 609				xfs_defer_cancel(&dfops);
 610				break;
 611			}
 612
 613			/* Remove the mapping from the CoW fork. */
 614			xfs_bmap_del_extent_cow(ip, &idx, &got, &del);
 615		}
 616
 617		if (!xfs_iext_get_extent(ifp, ++idx, &got))
 
 
 
 
 
 
 
 
 
 
 
 618			break;
 619	}
 620
 621	/* clear tag if cow fork is emptied */
 622	if (!ifp->if_bytes)
 623		xfs_inode_clear_cowblocks_tag(ip);
 624
 625	return error;
 626}
 627
 628/*
 629 * Cancel CoW reservations for some byte range of an inode.
 630 *
 631 * If cancel_real is true this function cancels all COW fork extents for the
 632 * inode; if cancel_real is false, real extents are not cleared.
 633 */
 634int
 635xfs_reflink_cancel_cow_range(
 636	struct xfs_inode	*ip,
 637	xfs_off_t		offset,
 638	xfs_off_t		count,
 639	bool			cancel_real)
 640{
 641	struct xfs_trans	*tp;
 642	xfs_fileoff_t		offset_fsb;
 643	xfs_fileoff_t		end_fsb;
 644	int			error;
 645
 646	trace_xfs_reflink_cancel_cow_range(ip, offset, count);
 647	ASSERT(xfs_is_reflink_inode(ip));
 648
 649	offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
 650	if (count == NULLFILEOFF)
 651		end_fsb = NULLFILEOFF;
 652	else
 653		end_fsb = XFS_B_TO_FSB(ip->i_mount, offset + count);
 654
 655	/* Start a rolling transaction to remove the mappings */
 656	error = xfs_trans_alloc(ip->i_mount, &M_RES(ip->i_mount)->tr_write,
 657			0, 0, 0, &tp);
 658	if (error)
 659		goto out;
 660
 661	xfs_ilock(ip, XFS_ILOCK_EXCL);
 662	xfs_trans_ijoin(tp, ip, 0);
 663
 664	/* Scrape out the old CoW reservations */
 665	error = xfs_reflink_cancel_cow_blocks(ip, &tp, offset_fsb, end_fsb,
 666			cancel_real);
 667	if (error)
 668		goto out_cancel;
 669
 670	error = xfs_trans_commit(tp);
 671
 672	xfs_iunlock(ip, XFS_ILOCK_EXCL);
 673	return error;
 674
 675out_cancel:
 676	xfs_trans_cancel(tp);
 677	xfs_iunlock(ip, XFS_ILOCK_EXCL);
 678out:
 679	trace_xfs_reflink_cancel_cow_range_error(ip, error, _RET_IP_);
 680	return error;
 681}
 682
 683/*
 684 * Remap parts of a file's data fork after a successful CoW.
 
 
 
 
 
 
 
 685 */
 686int
 687xfs_reflink_end_cow(
 688	struct xfs_inode		*ip,
 689	xfs_off_t			offset,
 690	xfs_off_t			count)
 691{
 692	struct xfs_ifork		*ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
 693	struct xfs_bmbt_irec		got, del;
 694	struct xfs_trans		*tp;
 695	xfs_fileoff_t			offset_fsb;
 696	xfs_fileoff_t			end_fsb;
 697	xfs_fsblock_t			firstfsb;
 698	struct xfs_defer_ops		dfops;
 699	int				error;
 700	unsigned int			resblks;
 701	xfs_filblks_t			rlen;
 702	xfs_extnum_t			idx;
 703
 704	trace_xfs_reflink_end_cow(ip, offset, count);
 705
 706	/* No COW extents?  That's easy! */
 707	if (ifp->if_bytes == 0)
 
 708		return 0;
 
 709
 710	offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
 711	end_fsb = XFS_B_TO_FSB(ip->i_mount, offset + count);
 712
 713	/* Start a rolling transaction to switch the mappings */
 714	resblks = XFS_EXTENTADD_SPACE_RES(ip->i_mount, XFS_DATA_FORK);
 715	error = xfs_trans_alloc(ip->i_mount, &M_RES(ip->i_mount)->tr_write,
 716			resblks, 0, 0, &tp);
 717	if (error)
 718		goto out;
 719
 
 
 
 
 
 720	xfs_ilock(ip, XFS_ILOCK_EXCL);
 721	xfs_trans_ijoin(tp, ip, 0);
 722
 723	/* If there is a hole at end_fsb - 1 go to the previous extent */
 724	if (!xfs_iext_lookup_extent(ip, ifp, end_fsb - 1, &idx, &got) ||
 725	    got.br_startoff > end_fsb) {
 726		ASSERT(idx > 0);
 727		xfs_iext_get_extent(ifp, --idx, &got);
 
 
 
 
 728	}
 729
 730	/* Walk backwards until we're out of the I/O range... */
 731	while (got.br_startoff + got.br_blockcount > offset_fsb) {
 732		del = got;
 733		xfs_trim_extent(&del, offset_fsb, end_fsb - offset_fsb);
 734
 735		/* Extent delete may have bumped idx forward */
 736		if (!del.br_blockcount) {
 737			idx--;
 738			goto next_extent;
 739		}
 740
 741		ASSERT(!isnullstartblock(got.br_startblock));
 742
 743		/*
 744		 * Don't remap unwritten extents; these are
 745		 * speculatively preallocated CoW extents that have been
 746		 * allocated but have not yet been involved in a write.
 747		 */
 748		if (got.br_state == XFS_EXT_UNWRITTEN) {
 749			idx--;
 750			goto next_extent;
 751		}
 752
 753		/* Unmap the old blocks in the data fork. */
 754		xfs_defer_init(&dfops, &firstfsb);
 755		rlen = del.br_blockcount;
 756		error = __xfs_bunmapi(tp, ip, del.br_startoff, &rlen, 0, 1,
 757				&firstfsb, &dfops);
 758		if (error)
 759			goto out_defer;
 760
 761		/* Trim the extent to whatever got unmapped. */
 762		if (rlen) {
 763			xfs_trim_extent(&del, del.br_startoff + rlen,
 764				del.br_blockcount - rlen);
 765		}
 766		trace_xfs_reflink_cow_remap(ip, &del);
 767
 768		/* Free the CoW orphan record. */
 769		error = xfs_refcount_free_cow_extent(tp->t_mountp, &dfops,
 770				del.br_startblock, del.br_blockcount);
 771		if (error)
 772			goto out_defer;
 
 
 
 
 773
 774		/* Map the new blocks into the data fork. */
 775		error = xfs_bmap_map_extent(tp->t_mountp, &dfops, ip, &del);
 776		if (error)
 777			goto out_defer;
 
 778
 779		/* Remove the mapping from the CoW fork. */
 780		xfs_bmap_del_extent_cow(ip, &idx, &got, &del);
 
 
 
 
 
 
 
 
 
 
 
 781
 782		error = xfs_defer_finish(&tp, &dfops, ip);
 783		if (error)
 784			goto out_defer;
 785next_extent:
 786		if (!xfs_iext_get_extent(ifp, idx, &got))
 787			break;
 788	}
 789
 790	error = xfs_trans_commit(tp);
 791	xfs_iunlock(ip, XFS_ILOCK_EXCL);
 792	if (error)
 793		goto out;
 
 
 
 794	return 0;
 795
 796out_defer:
 797	xfs_defer_cancel(&dfops);
 798	xfs_trans_cancel(tp);
 799	xfs_iunlock(ip, XFS_ILOCK_EXCL);
 800out:
 801	trace_xfs_reflink_end_cow_error(ip, error, _RET_IP_);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 802	return error;
 803}
 804
 805/*
 806 * Free leftover CoW reservations that didn't get cleaned out.
 807 */
 808int
 809xfs_reflink_recover_cow(
 810	struct xfs_mount	*mp)
 811{
 812	xfs_agnumber_t		agno;
 813	int			error = 0;
 814
 815	if (!xfs_sb_version_hasreflink(&mp->m_sb))
 816		return 0;
 817
 818	for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
 819		error = xfs_refcount_recover_cow_leftovers(mp, agno);
 820		if (error)
 821			break;
 822	}
 823
 824	return error;
 825}
 826
 827/*
 828 * Reflinking (Block) Ranges of Two Files Together
 829 *
 830 * First, ensure that the reflink flag is set on both inodes.  The flag is an
 831 * optimization to avoid unnecessary refcount btree lookups in the write path.
 832 *
 833 * Now we can iteratively remap the range of extents (and holes) in src to the
 834 * corresponding ranges in dest.  Let drange and srange denote the ranges of
 835 * logical blocks in dest and src touched by the reflink operation.
 836 *
 837 * While the length of drange is greater than zero,
 838 *    - Read src's bmbt at the start of srange ("imap")
 839 *    - If imap doesn't exist, make imap appear to start at the end of srange
 840 *      with zero length.
 841 *    - If imap starts before srange, advance imap to start at srange.
 842 *    - If imap goes beyond srange, truncate imap to end at the end of srange.
 843 *    - Punch (imap start - srange start + imap len) blocks from dest at
 844 *      offset (drange start).
 845 *    - If imap points to a real range of pblks,
 846 *         > Increase the refcount of the imap's pblks
 847 *         > Map imap's pblks into dest at the offset
 848 *           (drange start + imap start - srange start)
 849 *    - Advance drange and srange by (imap start - srange start + imap len)
 850 *
 851 * Finally, if the reflink made dest longer, update both the in-core and
 852 * on-disk file sizes.
 853 *
 854 * ASCII Art Demonstration:
 855 *
 856 * Let's say we want to reflink this source file:
 857 *
 858 * ----SSSSSSS-SSSSS----SSSSSS (src file)
 859 *   <-------------------->
 860 *
 861 * into this destination file:
 862 *
 863 * --DDDDDDDDDDDDDDDDDDD--DDD (dest file)
 864 *        <-------------------->
 865 * '-' means a hole, and 'S' and 'D' are written blocks in the src and dest.
 866 * Observe that the range has different logical offsets in either file.
 867 *
 868 * Consider that the first extent in the source file doesn't line up with our
 869 * reflink range.  Unmapping  and remapping are separate operations, so we can
 870 * unmap more blocks from the destination file than we remap.
 871 *
 872 * ----SSSSSSS-SSSSS----SSSSSS
 873 *   <------->
 874 * --DDDDD---------DDDDD--DDD
 875 *        <------->
 876 *
 877 * Now remap the source extent into the destination file:
 878 *
 879 * ----SSSSSSS-SSSSS----SSSSSS
 880 *   <------->
 881 * --DDDDD--SSSSSSSDDDDD--DDD
 882 *        <------->
 883 *
 884 * Do likewise with the second hole and extent in our range.  Holes in the
 885 * unmap range don't affect our operation.
 886 *
 887 * ----SSSSSSS-SSSSS----SSSSSS
 888 *            <---->
 889 * --DDDDD--SSSSSSS-SSSSS-DDD
 890 *                 <---->
 891 *
 892 * Finally, unmap and remap part of the third extent.  This will increase the
 893 * size of the destination file.
 894 *
 895 * ----SSSSSSS-SSSSS----SSSSSS
 896 *                  <----->
 897 * --DDDDD--SSSSSSS-SSSSS----SSS
 898 *                       <----->
 899 *
 900 * Once we update the destination file's i_size, we're done.
 901 */
 902
 903/*
 904 * Ensure the reflink bit is set in both inodes.
 905 */
 906STATIC int
 907xfs_reflink_set_inode_flag(
 908	struct xfs_inode	*src,
 909	struct xfs_inode	*dest)
 910{
 911	struct xfs_mount	*mp = src->i_mount;
 912	int			error;
 913	struct xfs_trans	*tp;
 914
 915	if (xfs_is_reflink_inode(src) && xfs_is_reflink_inode(dest))
 916		return 0;
 917
 918	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
 919	if (error)
 920		goto out_error;
 921
 922	/* Lock both files against IO */
 923	if (src->i_ino == dest->i_ino)
 924		xfs_ilock(src, XFS_ILOCK_EXCL);
 925	else
 926		xfs_lock_two_inodes(src, dest, XFS_ILOCK_EXCL);
 927
 928	if (!xfs_is_reflink_inode(src)) {
 929		trace_xfs_reflink_set_inode_flag(src);
 930		xfs_trans_ijoin(tp, src, XFS_ILOCK_EXCL);
 931		src->i_d.di_flags2 |= XFS_DIFLAG2_REFLINK;
 932		xfs_trans_log_inode(tp, src, XFS_ILOG_CORE);
 933		xfs_ifork_init_cow(src);
 934	} else
 935		xfs_iunlock(src, XFS_ILOCK_EXCL);
 936
 937	if (src->i_ino == dest->i_ino)
 938		goto commit_flags;
 939
 940	if (!xfs_is_reflink_inode(dest)) {
 941		trace_xfs_reflink_set_inode_flag(dest);
 942		xfs_trans_ijoin(tp, dest, XFS_ILOCK_EXCL);
 943		dest->i_d.di_flags2 |= XFS_DIFLAG2_REFLINK;
 944		xfs_trans_log_inode(tp, dest, XFS_ILOG_CORE);
 945		xfs_ifork_init_cow(dest);
 946	} else
 947		xfs_iunlock(dest, XFS_ILOCK_EXCL);
 948
 949commit_flags:
 950	error = xfs_trans_commit(tp);
 951	if (error)
 952		goto out_error;
 953	return error;
 954
 955out_error:
 956	trace_xfs_reflink_set_inode_flag_error(dest, error, _RET_IP_);
 957	return error;
 958}
 959
 960/*
 961 * Update destination inode size & cowextsize hint, if necessary.
 962 */
 963STATIC int
 964xfs_reflink_update_dest(
 965	struct xfs_inode	*dest,
 966	xfs_off_t		newlen,
 967	xfs_extlen_t		cowextsize,
 968	bool			is_dedupe)
 969{
 970	struct xfs_mount	*mp = dest->i_mount;
 971	struct xfs_trans	*tp;
 972	int			error;
 973
 974	if (is_dedupe && newlen <= i_size_read(VFS_I(dest)) && cowextsize == 0)
 975		return 0;
 976
 977	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
 978	if (error)
 979		goto out_error;
 980
 981	xfs_ilock(dest, XFS_ILOCK_EXCL);
 982	xfs_trans_ijoin(tp, dest, XFS_ILOCK_EXCL);
 983
 984	if (newlen > i_size_read(VFS_I(dest))) {
 985		trace_xfs_reflink_update_inode_size(dest, newlen);
 986		i_size_write(VFS_I(dest), newlen);
 987		dest->i_d.di_size = newlen;
 988	}
 989
 990	if (cowextsize) {
 991		dest->i_d.di_cowextsize = cowextsize;
 992		dest->i_d.di_flags2 |= XFS_DIFLAG2_COWEXTSIZE;
 993	}
 994
 995	if (!is_dedupe) {
 996		xfs_trans_ichgtime(tp, dest,
 997				   XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
 998	}
 999	xfs_trans_log_inode(tp, dest, XFS_ILOG_CORE);
1000
1001	error = xfs_trans_commit(tp);
1002	if (error)
1003		goto out_error;
1004	return error;
1005
1006out_error:
1007	trace_xfs_reflink_update_inode_size_error(dest, error, _RET_IP_);
1008	return error;
1009}
1010
1011/*
1012 * Do we have enough reserve in this AG to handle a reflink?  The refcount
1013 * btree already reserved all the space it needs, but the rmap btree can grow
1014 * infinitely, so we won't allow more reflinks when the AG is down to the
1015 * btree reserves.
1016 */
1017static int
1018xfs_reflink_ag_has_free_space(
1019	struct xfs_mount	*mp,
1020	xfs_agnumber_t		agno)
1021{
1022	struct xfs_perag	*pag;
1023	int			error = 0;
1024
1025	if (!xfs_sb_version_hasrmapbt(&mp->m_sb))
1026		return 0;
1027
1028	pag = xfs_perag_get(mp, agno);
1029	if (xfs_ag_resv_critical(pag, XFS_AG_RESV_AGFL) ||
1030	    xfs_ag_resv_critical(pag, XFS_AG_RESV_METADATA))
1031		error = -ENOSPC;
1032	xfs_perag_put(pag);
1033	return error;
1034}
1035
1036/*
1037 * Unmap a range of blocks from a file, then map other blocks into the hole.
1038 * The range to unmap is (destoff : destoff + srcioff + irec->br_blockcount).
1039 * The extent irec is mapped into dest at irec->br_startoff.
1040 */
1041STATIC int
1042xfs_reflink_remap_extent(
1043	struct xfs_inode	*ip,
1044	struct xfs_bmbt_irec	*irec,
1045	xfs_fileoff_t		destoff,
1046	xfs_off_t		new_isize)
1047{
1048	struct xfs_mount	*mp = ip->i_mount;
 
1049	struct xfs_trans	*tp;
1050	xfs_fsblock_t		firstfsb;
1051	unsigned int		resblks;
1052	struct xfs_defer_ops	dfops;
1053	struct xfs_bmbt_irec	uirec;
1054	bool			real_extent;
1055	xfs_filblks_t		rlen;
1056	xfs_filblks_t		unmap_len;
1057	xfs_off_t		newlen;
1058	int			error;
1059
1060	unmap_len = irec->br_startoff + irec->br_blockcount - destoff;
1061	trace_xfs_reflink_punch_range(ip, destoff, unmap_len);
1062
1063	/* Only remap normal extents. */
1064	real_extent =  (irec->br_startblock != HOLESTARTBLOCK &&
1065			irec->br_startblock != DELAYSTARTBLOCK &&
1066			!ISUNWRITTEN(irec));
1067
1068	/* No reflinking if we're low on space */
1069	if (real_extent) {
1070		error = xfs_reflink_ag_has_free_space(mp,
1071				XFS_FSB_TO_AGNO(mp, irec->br_startblock));
1072		if (error)
1073			goto out;
1074	}
1075
1076	/* Start a rolling transaction to switch the mappings */
1077	resblks = XFS_EXTENTADD_SPACE_RES(ip->i_mount, XFS_DATA_FORK);
1078	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, resblks, 0, 0, &tp);
1079	if (error)
1080		goto out;
1081
1082	xfs_ilock(ip, XFS_ILOCK_EXCL);
1083	xfs_trans_ijoin(tp, ip, 0);
1084
1085	/* If we're not just clearing space, then do we have enough quota? */
1086	if (real_extent) {
1087		error = xfs_trans_reserve_quota_nblks(tp, ip,
1088				irec->br_blockcount, 0, XFS_QMOPT_RES_REGBLKS);
1089		if (error)
1090			goto out_cancel;
1091	}
1092
1093	trace_xfs_reflink_remap(ip, irec->br_startoff,
1094				irec->br_blockcount, irec->br_startblock);
1095
1096	/* Unmap the old blocks in the data fork. */
1097	rlen = unmap_len;
1098	while (rlen) {
1099		xfs_defer_init(&dfops, &firstfsb);
1100		error = __xfs_bunmapi(tp, ip, destoff, &rlen, 0, 1,
1101				&firstfsb, &dfops);
1102		if (error)
1103			goto out_defer;
1104
1105		/*
1106		 * Trim the extent to whatever got unmapped.
1107		 * Remember, bunmapi works backwards.
1108		 */
1109		uirec.br_startblock = irec->br_startblock + rlen;
1110		uirec.br_startoff = irec->br_startoff + rlen;
1111		uirec.br_blockcount = unmap_len - rlen;
1112		unmap_len = rlen;
1113
1114		/* If this isn't a real mapping, we're done. */
1115		if (!real_extent || uirec.br_blockcount == 0)
1116			goto next_extent;
1117
1118		trace_xfs_reflink_remap(ip, uirec.br_startoff,
1119				uirec.br_blockcount, uirec.br_startblock);
1120
1121		/* Update the refcount tree */
1122		error = xfs_refcount_increase_extent(mp, &dfops, &uirec);
1123		if (error)
1124			goto out_defer;
1125
1126		/* Map the new blocks into the data fork. */
1127		error = xfs_bmap_map_extent(mp, &dfops, ip, &uirec);
1128		if (error)
1129			goto out_defer;
1130
1131		/* Update quota accounting. */
1132		xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_BCOUNT,
1133				uirec.br_blockcount);
1134
1135		/* Update dest isize if needed. */
1136		newlen = XFS_FSB_TO_B(mp,
1137				uirec.br_startoff + uirec.br_blockcount);
1138		newlen = min_t(xfs_off_t, newlen, new_isize);
1139		if (newlen > i_size_read(VFS_I(ip))) {
1140			trace_xfs_reflink_update_inode_size(ip, newlen);
1141			i_size_write(VFS_I(ip), newlen);
1142			ip->i_d.di_size = newlen;
1143			xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1144		}
1145
1146next_extent:
1147		/* Process all the deferred stuff. */
1148		error = xfs_defer_finish(&tp, &dfops, ip);
1149		if (error)
1150			goto out_defer;
1151	}
1152
1153	error = xfs_trans_commit(tp);
1154	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1155	if (error)
1156		goto out;
1157	return 0;
1158
1159out_defer:
1160	xfs_defer_cancel(&dfops);
1161out_cancel:
1162	xfs_trans_cancel(tp);
1163	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1164out:
1165	trace_xfs_reflink_remap_extent_error(ip, error, _RET_IP_);
1166	return error;
1167}
1168
1169/*
1170 * Iteratively remap one file's extents (and holes) to another's.
1171 */
1172STATIC int
1173xfs_reflink_remap_blocks(
1174	struct xfs_inode	*src,
1175	xfs_fileoff_t		srcoff,
1176	struct xfs_inode	*dest,
1177	xfs_fileoff_t		destoff,
1178	xfs_filblks_t		len,
1179	xfs_off_t		new_isize)
1180{
1181	struct xfs_bmbt_irec	imap;
 
 
 
 
 
 
1182	int			nimaps;
1183	int			error = 0;
1184	xfs_filblks_t		range_len;
 
 
 
1185
1186	/* drange = (destoff, destoff + len); srange = (srcoff, srcoff + len) */
1187	while (len) {
 
 
1188		trace_xfs_reflink_remap_blocks_loop(src, srcoff, len,
1189				dest, destoff);
 
1190		/* Read extent from the source file */
1191		nimaps = 1;
1192		xfs_ilock(src, XFS_ILOCK_EXCL);
1193		error = xfs_bmapi_read(src, srcoff, len, &imap, &nimaps, 0);
1194		xfs_iunlock(src, XFS_ILOCK_EXCL);
1195		if (error)
1196			goto err;
1197		ASSERT(nimaps == 1);
1198
1199		trace_xfs_reflink_remap_imap(src, srcoff, len, XFS_IO_OVERWRITE,
1200				&imap);
1201
1202		/* Translate imap into the destination file. */
1203		range_len = imap.br_startoff + imap.br_blockcount - srcoff;
1204		imap.br_startoff += destoff - srcoff;
1205
1206		/* Clear dest from destoff to the end of imap and map it in. */
1207		error = xfs_reflink_remap_extent(dest, &imap, destoff,
1208				new_isize);
1209		if (error)
1210			goto err;
1211
1212		if (fatal_signal_pending(current)) {
1213			error = -EINTR;
1214			goto err;
1215		}
1216
1217		/* Advance drange/srange */
1218		srcoff += range_len;
1219		destoff += range_len;
1220		len -= range_len;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1221	}
1222
1223	return 0;
 
1224
1225err:
1226	trace_xfs_reflink_remap_blocks_error(dest, error, _RET_IP_);
1227	return error;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1228}
1229
1230/*
1231 * Link a range of blocks from one file to another.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1232 */
1233int
1234xfs_reflink_remap_range(
1235	struct file		*file_in,
1236	loff_t			pos_in,
1237	struct file		*file_out,
1238	loff_t			pos_out,
1239	u64			len,
1240	bool			is_dedupe)
1241{
1242	struct inode		*inode_in = file_inode(file_in);
1243	struct xfs_inode	*src = XFS_I(inode_in);
1244	struct inode		*inode_out = file_inode(file_out);
1245	struct xfs_inode	*dest = XFS_I(inode_out);
1246	struct xfs_mount	*mp = src->i_mount;
1247	bool			same_inode = (inode_in == inode_out);
1248	xfs_fileoff_t		sfsbno, dfsbno;
1249	xfs_filblks_t		fsblen;
1250	xfs_extlen_t		cowextsize;
1251	ssize_t			ret;
1252
1253	if (!xfs_sb_version_hasreflink(&mp->m_sb))
1254		return -EOPNOTSUPP;
1255
1256	if (XFS_FORCED_SHUTDOWN(mp))
1257		return -EIO;
1258
1259	/* Lock both files against IO */
1260	lock_two_nondirectories(inode_in, inode_out);
 
 
1261	if (same_inode)
1262		xfs_ilock(src, XFS_MMAPLOCK_EXCL);
1263	else
1264		xfs_lock_two_inodes(src, dest, XFS_MMAPLOCK_EXCL);
 
1265
1266	/* Check file eligibility and prepare for block sharing. */
1267	ret = -EINVAL;
1268	/* Don't reflink realtime inodes */
1269	if (XFS_IS_REALTIME_INODE(src) || XFS_IS_REALTIME_INODE(dest))
1270		goto out_unlock;
1271
1272	/* Don't share DAX file data for now. */
1273	if (IS_DAX(inode_in) || IS_DAX(inode_out))
1274		goto out_unlock;
1275
1276	ret = vfs_clone_file_prep_inodes(inode_in, pos_in, inode_out, pos_out,
1277			&len, is_dedupe);
1278	if (ret <= 0)
1279		goto out_unlock;
1280
1281	trace_xfs_reflink_remap_range(src, pos_in, len, dest, pos_out);
1282
1283	/* Set flags and remap blocks. */
1284	ret = xfs_reflink_set_inode_flag(src, dest);
1285	if (ret)
1286		goto out_unlock;
1287
1288	dfsbno = XFS_B_TO_FSBT(mp, pos_out);
1289	sfsbno = XFS_B_TO_FSBT(mp, pos_in);
1290	fsblen = XFS_B_TO_FSB(mp, len);
1291	ret = xfs_reflink_remap_blocks(src, sfsbno, dest, dfsbno, fsblen,
1292			pos_out + len);
1293	if (ret)
1294		goto out_unlock;
1295
1296	/* Zap any page cache for the destination file's range. */
1297	truncate_inode_pages_range(&inode_out->i_data, pos_out,
1298				   PAGE_ALIGN(pos_out + len) - 1);
 
1299
1300	/*
1301	 * Carry the cowextsize hint from src to dest if we're sharing the
1302	 * entire source file to the entire destination file, the source file
1303	 * has a cowextsize hint, and the destination file does not.
1304	 */
1305	cowextsize = 0;
1306	if (pos_in == 0 && len == i_size_read(inode_in) &&
1307	    (src->i_d.di_flags2 & XFS_DIFLAG2_COWEXTSIZE) &&
1308	    pos_out == 0 && len >= i_size_read(inode_out) &&
1309	    !(dest->i_d.di_flags2 & XFS_DIFLAG2_COWEXTSIZE))
1310		cowextsize = src->i_d.di_cowextsize;
1311
1312	ret = xfs_reflink_update_dest(dest, pos_out + len, cowextsize,
1313			is_dedupe);
1314
 
1315out_unlock:
1316	xfs_iunlock(src, XFS_MMAPLOCK_EXCL);
1317	if (!same_inode)
1318		xfs_iunlock(dest, XFS_MMAPLOCK_EXCL);
1319	unlock_two_nondirectories(inode_in, inode_out);
1320	if (ret)
1321		trace_xfs_reflink_remap_range_error(dest, ret, _RET_IP_);
1322	return ret;
1323}
1324
1325/*
1326 * The user wants to preemptively CoW all shared blocks in this file,
1327 * which enables us to turn off the reflink flag.  Iterate all
1328 * extents which are not prealloc/delalloc to see which ranges are
1329 * mentioned in the refcount tree, then read those blocks into the
1330 * pagecache, dirty them, fsync them back out, and then we can update
1331 * the inode flag.  What happens if we run out of memory? :)
1332 */
1333STATIC int
1334xfs_reflink_dirty_extents(
1335	struct xfs_inode	*ip,
1336	xfs_fileoff_t		fbno,
1337	xfs_filblks_t		end,
1338	xfs_off_t		isize)
1339{
1340	struct xfs_mount	*mp = ip->i_mount;
1341	xfs_agnumber_t		agno;
1342	xfs_agblock_t		agbno;
1343	xfs_extlen_t		aglen;
1344	xfs_agblock_t		rbno;
1345	xfs_extlen_t		rlen;
1346	xfs_off_t		fpos;
1347	xfs_off_t		flen;
1348	struct xfs_bmbt_irec	map[2];
1349	int			nmaps;
1350	int			error = 0;
1351
1352	while (end - fbno > 0) {
1353		nmaps = 1;
1354		/*
1355		 * Look for extents in the file.  Skip holes, delalloc, or
1356		 * unwritten extents; they can't be reflinked.
1357		 */
1358		error = xfs_bmapi_read(ip, fbno, end - fbno, map, &nmaps, 0);
1359		if (error)
1360			goto out;
1361		if (nmaps == 0)
1362			break;
1363		if (map[0].br_startblock == HOLESTARTBLOCK ||
1364		    map[0].br_startblock == DELAYSTARTBLOCK ||
1365		    ISUNWRITTEN(&map[0]))
1366			goto next;
1367
1368		map[1] = map[0];
1369		while (map[1].br_blockcount) {
1370			agno = XFS_FSB_TO_AGNO(mp, map[1].br_startblock);
1371			agbno = XFS_FSB_TO_AGBNO(mp, map[1].br_startblock);
1372			aglen = map[1].br_blockcount;
1373
1374			error = xfs_reflink_find_shared(mp, agno, agbno, aglen,
1375					&rbno, &rlen, true);
1376			if (error)
1377				goto out;
1378			if (rbno == NULLAGBLOCK)
1379				break;
1380
1381			/* Dirty the pages */
1382			xfs_iunlock(ip, XFS_ILOCK_EXCL);
1383			fpos = XFS_FSB_TO_B(mp, map[1].br_startoff +
1384					(rbno - agbno));
1385			flen = XFS_FSB_TO_B(mp, rlen);
1386			if (fpos + flen > isize)
1387				flen = isize - fpos;
1388			error = iomap_file_dirty(VFS_I(ip), fpos, flen,
1389					&xfs_iomap_ops);
1390			xfs_ilock(ip, XFS_ILOCK_EXCL);
1391			if (error)
1392				goto out;
1393
1394			map[1].br_blockcount -= (rbno - agbno + rlen);
1395			map[1].br_startoff += (rbno - agbno + rlen);
1396			map[1].br_startblock += (rbno - agbno + rlen);
1397		}
1398
1399next:
1400		fbno = map[0].br_startoff + map[0].br_blockcount;
1401	}
1402out:
1403	return error;
1404}
1405
1406/* Clear the inode reflink flag if there are no shared extents. */
1407int
1408xfs_reflink_clear_inode_flag(
1409	struct xfs_inode	*ip,
1410	struct xfs_trans	**tpp)
 
1411{
1412	struct xfs_mount	*mp = ip->i_mount;
1413	xfs_fileoff_t		fbno;
1414	xfs_filblks_t		end;
1415	xfs_agnumber_t		agno;
1416	xfs_agblock_t		agbno;
1417	xfs_extlen_t		aglen;
1418	xfs_agblock_t		rbno;
1419	xfs_extlen_t		rlen;
1420	struct xfs_bmbt_irec	map;
1421	int			nmaps;
1422	int			error = 0;
1423
1424	ASSERT(xfs_is_reflink_inode(ip));
1425
1426	fbno = 0;
1427	end = XFS_B_TO_FSB(mp, i_size_read(VFS_I(ip)));
1428	while (end - fbno > 0) {
1429		nmaps = 1;
1430		/*
1431		 * Look for extents in the file.  Skip holes, delalloc, or
1432		 * unwritten extents; they can't be reflinked.
1433		 */
1434		error = xfs_bmapi_read(ip, fbno, end - fbno, &map, &nmaps, 0);
1435		if (error)
1436			return error;
1437		if (nmaps == 0)
1438			break;
1439		if (map.br_startblock == HOLESTARTBLOCK ||
1440		    map.br_startblock == DELAYSTARTBLOCK ||
1441		    ISUNWRITTEN(&map))
1442			goto next;
1443
1444		agno = XFS_FSB_TO_AGNO(mp, map.br_startblock);
1445		agbno = XFS_FSB_TO_AGBNO(mp, map.br_startblock);
1446		aglen = map.br_blockcount;
 
 
 
 
 
 
1447
1448		error = xfs_reflink_find_shared(mp, agno, agbno, aglen,
1449				&rbno, &rlen, false);
1450		if (error)
1451			return error;
1452		/* Is there still a shared block here? */
1453		if (rbno != NULLAGBLOCK)
 
1454			return 0;
 
1455next:
1456		fbno = map.br_startoff + map.br_blockcount;
1457	}
1458
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1459	/*
1460	 * We didn't find any shared blocks so turn off the reflink flag.
1461	 * First, get rid of any leftover CoW mappings.
1462	 */
1463	error = xfs_reflink_cancel_cow_blocks(ip, tpp, 0, NULLFILEOFF, true);
1464	if (error)
1465		return error;
1466
1467	/* Clear the inode flag. */
1468	trace_xfs_reflink_unset_inode_flag(ip);
1469	ip->i_d.di_flags2 &= ~XFS_DIFLAG2_REFLINK;
1470	xfs_inode_clear_cowblocks_tag(ip);
1471	xfs_trans_ijoin(*tpp, ip, 0);
1472	xfs_trans_log_inode(*tpp, ip, XFS_ILOG_CORE);
1473
1474	return error;
1475}
1476
1477/*
1478 * Clear the inode reflink flag if there are no shared extents and the size
1479 * hasn't changed.
1480 */
1481STATIC int
1482xfs_reflink_try_clear_inode_flag(
1483	struct xfs_inode	*ip)
1484{
1485	struct xfs_mount	*mp = ip->i_mount;
1486	struct xfs_trans	*tp;
1487	int			error = 0;
1488
1489	/* Start a rolling transaction to remove the mappings */
1490	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, 0, 0, 0, &tp);
1491	if (error)
1492		return error;
1493
1494	xfs_ilock(ip, XFS_ILOCK_EXCL);
1495	xfs_trans_ijoin(tp, ip, 0);
1496
1497	error = xfs_reflink_clear_inode_flag(ip, &tp);
1498	if (error)
1499		goto cancel;
1500
1501	error = xfs_trans_commit(tp);
1502	if (error)
1503		goto out;
1504
1505	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1506	return 0;
1507cancel:
1508	xfs_trans_cancel(tp);
1509out:
1510	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1511	return error;
1512}
1513
1514/*
1515 * Pre-COW all shared blocks within a given byte range of a file and turn off
1516 * the reflink flag if we unshare all of the file's blocks.
1517 */
1518int
1519xfs_reflink_unshare(
1520	struct xfs_inode	*ip,
1521	xfs_off_t		offset,
1522	xfs_off_t		len)
1523{
1524	struct xfs_mount	*mp = ip->i_mount;
1525	xfs_fileoff_t		fbno;
1526	xfs_filblks_t		end;
1527	xfs_off_t		isize;
1528	int			error;
1529
1530	if (!xfs_is_reflink_inode(ip))
1531		return 0;
1532
1533	trace_xfs_reflink_unshare(ip, offset, len);
1534
1535	inode_dio_wait(VFS_I(ip));
1536
1537	/* Try to CoW the selected ranges */
1538	xfs_ilock(ip, XFS_ILOCK_EXCL);
1539	fbno = XFS_B_TO_FSBT(mp, offset);
1540	isize = i_size_read(VFS_I(ip));
1541	end = XFS_B_TO_FSB(mp, offset + len);
1542	error = xfs_reflink_dirty_extents(ip, fbno, end, isize);
1543	if (error)
1544		goto out_unlock;
1545	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1546
1547	/* Wait for the IO to finish */
1548	error = filemap_write_and_wait(VFS_I(ip)->i_mapping);
1549	if (error)
1550		goto out;
1551
1552	/* Turn off the reflink flag if possible. */
1553	error = xfs_reflink_try_clear_inode_flag(ip);
1554	if (error)
1555		goto out;
1556
1557	return 0;
1558
1559out_unlock:
1560	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1561out:
1562	trace_xfs_reflink_unshare_error(ip, error, _RET_IP_);
1563	return error;
1564}
v5.4
   1// SPDX-License-Identifier: GPL-2.0+
   2/*
   3 * Copyright (C) 2016 Oracle.  All Rights Reserved.
 
   4 * Author: Darrick J. Wong <darrick.wong@oracle.com>
 
 
 
 
 
 
 
 
 
 
 
 
 
 
   5 */
   6#include "xfs.h"
   7#include "xfs_fs.h"
   8#include "xfs_shared.h"
   9#include "xfs_format.h"
  10#include "xfs_log_format.h"
  11#include "xfs_trans_resv.h"
  12#include "xfs_mount.h"
  13#include "xfs_defer.h"
 
 
  14#include "xfs_inode.h"
  15#include "xfs_trans.h"
 
  16#include "xfs_bmap.h"
  17#include "xfs_bmap_util.h"
 
 
 
 
  18#include "xfs_trace.h"
 
  19#include "xfs_icache.h"
 
  20#include "xfs_btree.h"
  21#include "xfs_refcount_btree.h"
  22#include "xfs_refcount.h"
  23#include "xfs_bmap_btree.h"
  24#include "xfs_trans_space.h"
  25#include "xfs_bit.h"
  26#include "xfs_alloc.h"
 
  27#include "xfs_quota.h"
 
 
  28#include "xfs_reflink.h"
  29#include "xfs_iomap.h"
 
  30#include "xfs_sb.h"
  31#include "xfs_ag_resv.h"
  32
  33/*
  34 * Copy on Write of Shared Blocks
  35 *
  36 * XFS must preserve "the usual" file semantics even when two files share
  37 * the same physical blocks.  This means that a write to one file must not
  38 * alter the blocks in a different file; the way that we'll do that is
  39 * through the use of a copy-on-write mechanism.  At a high level, that
  40 * means that when we want to write to a shared block, we allocate a new
  41 * block, write the data to the new block, and if that succeeds we map the
  42 * new block into the file.
  43 *
  44 * XFS provides a "delayed allocation" mechanism that defers the allocation
  45 * of disk blocks to dirty-but-not-yet-mapped file blocks as long as
  46 * possible.  This reduces fragmentation by enabling the filesystem to ask
  47 * for bigger chunks less often, which is exactly what we want for CoW.
  48 *
  49 * The delalloc mechanism begins when the kernel wants to make a block
  50 * writable (write_begin or page_mkwrite).  If the offset is not mapped, we
  51 * create a delalloc mapping, which is a regular in-core extent, but without
  52 * a real startblock.  (For delalloc mappings, the startblock encodes both
  53 * a flag that this is a delalloc mapping, and a worst-case estimate of how
  54 * many blocks might be required to put the mapping into the BMBT.)  delalloc
  55 * mappings are a reservation against the free space in the filesystem;
  56 * adjacent mappings can also be combined into fewer larger mappings.
  57 *
  58 * As an optimization, the CoW extent size hint (cowextsz) creates
  59 * outsized aligned delalloc reservations in the hope of landing out of
  60 * order nearby CoW writes in a single extent on disk, thereby reducing
  61 * fragmentation and improving future performance.
  62 *
  63 * D: --RRRRRRSSSRRRRRRRR--- (data fork)
  64 * C: ------DDDDDDD--------- (CoW fork)
  65 *
  66 * When dirty pages are being written out (typically in writepage), the
  67 * delalloc reservations are converted into unwritten mappings by
  68 * allocating blocks and replacing the delalloc mapping with real ones.
  69 * A delalloc mapping can be replaced by several unwritten ones if the
  70 * free space is fragmented.
  71 *
  72 * D: --RRRRRRSSSRRRRRRRR---
  73 * C: ------UUUUUUU---------
  74 *
  75 * We want to adapt the delalloc mechanism for copy-on-write, since the
  76 * write paths are similar.  The first two steps (creating the reservation
  77 * and allocating the blocks) are exactly the same as delalloc except that
  78 * the mappings must be stored in a separate CoW fork because we do not want
  79 * to disturb the mapping in the data fork until we're sure that the write
  80 * succeeded.  IO completion in this case is the process of removing the old
  81 * mapping from the data fork and moving the new mapping from the CoW fork to
  82 * the data fork.  This will be discussed shortly.
  83 *
  84 * For now, unaligned directio writes will be bounced back to the page cache.
  85 * Block-aligned directio writes will use the same mechanism as buffered
  86 * writes.
  87 *
  88 * Just prior to submitting the actual disk write requests, we convert
  89 * the extents representing the range of the file actually being written
  90 * (as opposed to extra pieces created for the cowextsize hint) to real
  91 * extents.  This will become important in the next step:
  92 *
  93 * D: --RRRRRRSSSRRRRRRRR---
  94 * C: ------UUrrUUU---------
  95 *
  96 * CoW remapping must be done after the data block write completes,
  97 * because we don't want to destroy the old data fork map until we're sure
  98 * the new block has been written.  Since the new mappings are kept in a
  99 * separate fork, we can simply iterate these mappings to find the ones
 100 * that cover the file blocks that we just CoW'd.  For each extent, simply
 101 * unmap the corresponding range in the data fork, map the new range into
 102 * the data fork, and remove the extent from the CoW fork.  Because of
 103 * the presence of the cowextsize hint, however, we must be careful
 104 * only to remap the blocks that we've actually written out --  we must
 105 * never remap delalloc reservations nor CoW staging blocks that have
 106 * yet to be written.  This corresponds exactly to the real extents in
 107 * the CoW fork:
 108 *
 109 * D: --RRRRRRrrSRRRRRRRR---
 110 * C: ------UU--UUU---------
 111 *
 112 * Since the remapping operation can be applied to an arbitrary file
 113 * range, we record the need for the remap step as a flag in the ioend
 114 * instead of declaring a new IO type.  This is required for direct io
 115 * because we only have ioend for the whole dio, and we have to be able to
 116 * remember the presence of unwritten blocks and CoW blocks with a single
 117 * ioend structure.  Better yet, the more ground we can cover with one
 118 * ioend, the better.
 119 */
 120
 121/*
 122 * Given an AG extent, find the lowest-numbered run of shared blocks
 123 * within that range and return the range in fbno/flen.  If
 124 * find_end_of_shared is true, return the longest contiguous extent of
 125 * shared blocks.  If there are no shared extents, fbno and flen will
 126 * be set to NULLAGBLOCK and 0, respectively.
 127 */
 128int
 129xfs_reflink_find_shared(
 130	struct xfs_mount	*mp,
 131	struct xfs_trans	*tp,
 132	xfs_agnumber_t		agno,
 133	xfs_agblock_t		agbno,
 134	xfs_extlen_t		aglen,
 135	xfs_agblock_t		*fbno,
 136	xfs_extlen_t		*flen,
 137	bool			find_end_of_shared)
 138{
 139	struct xfs_buf		*agbp;
 140	struct xfs_btree_cur	*cur;
 141	int			error;
 142
 143	error = xfs_alloc_read_agf(mp, tp, agno, 0, &agbp);
 144	if (error)
 145		return error;
 146	if (!agbp)
 147		return -ENOMEM;
 148
 149	cur = xfs_refcountbt_init_cursor(mp, tp, agbp, agno);
 150
 151	error = xfs_refcount_find_shared(cur, agbno, aglen, fbno, flen,
 152			find_end_of_shared);
 153
 154	xfs_btree_del_cursor(cur, error);
 155
 156	xfs_trans_brelse(tp, agbp);
 157	return error;
 158}
 159
 160/*
 161 * Trim the mapping to the next block where there's a change in the
 162 * shared/unshared status.  More specifically, this means that we
 163 * find the lowest-numbered extent of shared blocks that coincides with
 164 * the given block mapping.  If the shared extent overlaps the start of
 165 * the mapping, trim the mapping to the end of the shared extent.  If
 166 * the shared region intersects the mapping, trim the mapping to the
 167 * start of the shared extent.  If there are no shared regions that
 168 * overlap, just return the original extent.
 169 */
 170int
 171xfs_reflink_trim_around_shared(
 172	struct xfs_inode	*ip,
 173	struct xfs_bmbt_irec	*irec,
 174	bool			*shared)
 
 175{
 176	xfs_agnumber_t		agno;
 177	xfs_agblock_t		agbno;
 178	xfs_extlen_t		aglen;
 179	xfs_agblock_t		fbno;
 180	xfs_extlen_t		flen;
 181	int			error = 0;
 182
 183	/* Holes, unwritten, and delalloc extents cannot be shared */
 184	if (!xfs_is_cow_inode(ip) || !xfs_bmap_is_real_extent(irec)) {
 
 
 
 
 185		*shared = false;
 186		return 0;
 187	}
 188
 189	trace_xfs_reflink_trim_around_shared(ip, irec);
 190
 191	agno = XFS_FSB_TO_AGNO(ip->i_mount, irec->br_startblock);
 192	agbno = XFS_FSB_TO_AGBNO(ip->i_mount, irec->br_startblock);
 193	aglen = irec->br_blockcount;
 194
 195	error = xfs_reflink_find_shared(ip->i_mount, NULL, agno, agbno,
 196			aglen, &fbno, &flen, true);
 197	if (error)
 198		return error;
 199
 200	*shared = false;
 201	if (fbno == NULLAGBLOCK) {
 202		/* No shared blocks at all. */
 203		return 0;
 204	} else if (fbno == agbno) {
 205		/*
 206		 * The start of this extent is shared.  Truncate the
 207		 * mapping at the end of the shared region so that a
 208		 * subsequent iteration starts at the start of the
 209		 * unshared region.
 210		 */
 211		irec->br_blockcount = flen;
 212		*shared = true;
 
 
 213		return 0;
 214	} else {
 215		/*
 216		 * There's a shared extent midway through this extent.
 217		 * Truncate the mapping at the start of the shared
 218		 * extent so that a subsequent iteration starts at the
 219		 * start of the shared region.
 220		 */
 221		irec->br_blockcount = fbno - agbno;
 
 222		return 0;
 223	}
 224}
 225
 226bool
 227xfs_inode_need_cow(
 
 
 
 
 
 
 
 
 
 228	struct xfs_inode	*ip,
 229	struct xfs_bmbt_irec	*imap,
 230	bool			*shared)
 231{
 232	/* We can't update any real extents in always COW mode. */
 233	if (xfs_is_always_cow_inode(ip) &&
 234	    !isnullstartblock(imap->br_startblock)) {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 235		*shared = true;
 236		return 0;
 237	}
 238
 239	/* Trim the mapping to the nearest shared extent boundary. */
 240	return xfs_reflink_trim_around_shared(ip, imap, shared);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 241}
 242
 243static int
 244xfs_reflink_convert_cow_locked(
 245	struct xfs_inode	*ip,
 246	xfs_fileoff_t		offset_fsb,
 247	xfs_filblks_t		count_fsb)
 
 
 
 248{
 249	struct xfs_iext_cursor	icur;
 250	struct xfs_bmbt_irec	got;
 251	struct xfs_btree_cur	*dummy_cur = NULL;
 252	int			dummy_logflags;
 253	int			error = 0;
 254
 255	if (!xfs_iext_lookup_extent(ip, ip->i_cowfp, offset_fsb, &icur, &got))
 256		return 0;
 257
 258	do {
 259		if (got.br_startoff >= offset_fsb + count_fsb)
 260			break;
 261		if (got.br_state == XFS_EXT_NORM)
 262			continue;
 263		if (WARN_ON_ONCE(isnullstartblock(got.br_startblock)))
 264			return -EIO;
 265
 266		xfs_trim_extent(&got, offset_fsb, count_fsb);
 267		if (!got.br_blockcount)
 268			continue;
 269
 270		got.br_state = XFS_EXT_NORM;
 271		error = xfs_bmap_add_extent_unwritten_real(NULL, ip,
 272				XFS_COW_FORK, &icur, &dummy_cur, &got,
 273				&dummy_logflags);
 274		if (error)
 275			return error;
 276	} while (xfs_iext_next_extent(ip->i_cowfp, &icur, &got));
 277
 278	return error;
 279}
 280
 281/* Convert all of the unwritten CoW extents in a file's range to real ones. */
 282int
 283xfs_reflink_convert_cow(
 284	struct xfs_inode	*ip,
 285	xfs_off_t		offset,
 286	xfs_off_t		count)
 287{
 
 
 288	struct xfs_mount	*mp = ip->i_mount;
 
 289	xfs_fileoff_t		offset_fsb = XFS_B_TO_FSBT(mp, offset);
 290	xfs_fileoff_t		end_fsb = XFS_B_TO_FSB(mp, offset + count);
 291	xfs_filblks_t		count_fsb = end_fsb - offset_fsb;
 292	int			error;
 293
 294	ASSERT(count != 0);
 295
 296	xfs_ilock(ip, XFS_ILOCK_EXCL);
 297	error = xfs_reflink_convert_cow_locked(ip, offset_fsb, count_fsb);
 298	xfs_iunlock(ip, XFS_ILOCK_EXCL);
 299	return error;
 300}
 301
 302/*
 303 * Find the extent that maps the given range in the COW fork. Even if the extent
 304 * is not shared we might have a preallocation for it in the COW fork. If so we
 305 * use it that rather than trigger a new allocation.
 306 */
 307static int
 308xfs_find_trim_cow_extent(
 309	struct xfs_inode	*ip,
 310	struct xfs_bmbt_irec	*imap,
 311	bool			*shared,
 312	bool			*found)
 313{
 314	xfs_fileoff_t		offset_fsb = imap->br_startoff;
 315	xfs_filblks_t		count_fsb = imap->br_blockcount;
 316	struct xfs_iext_cursor	icur;
 317	struct xfs_bmbt_irec	got;
 318
 319	*found = false;
 320
 321	/*
 322	 * If we don't find an overlapping extent, trim the range we need to
 323	 * allocate to fit the hole we found.
 324	 */
 325	if (!xfs_iext_lookup_extent(ip, ip->i_cowfp, offset_fsb, &icur, &got))
 326		got.br_startoff = offset_fsb + count_fsb;
 327	if (got.br_startoff > offset_fsb) {
 328		xfs_trim_extent(imap, imap->br_startoff,
 329				got.br_startoff - imap->br_startoff);
 330		return xfs_inode_need_cow(ip, imap, shared);
 331	}
 332
 333	*shared = true;
 334	if (isnullstartblock(got.br_startblock)) {
 335		xfs_trim_extent(imap, got.br_startoff, got.br_blockcount);
 336		return 0;
 337	}
 338
 339	/* real extent found - no need to allocate */
 340	xfs_trim_extent(&got, offset_fsb, count_fsb);
 341	*imap = got;
 342	*found = true;
 343	return 0;
 344}
 345
 346/* Allocate all CoW reservations covering a range of blocks in a file. */
 347int
 348xfs_reflink_allocate_cow(
 349	struct xfs_inode	*ip,
 350	struct xfs_bmbt_irec	*imap,
 351	bool			*shared,
 352	uint			*lockmode,
 353	bool			convert_now)
 354{
 355	struct xfs_mount	*mp = ip->i_mount;
 356	xfs_fileoff_t		offset_fsb = imap->br_startoff;
 357	xfs_filblks_t		count_fsb = imap->br_blockcount;
 358	struct xfs_trans	*tp;
 359	int			nimaps, error = 0;
 360	bool			found;
 361	xfs_filblks_t		resaligned;
 362	xfs_extlen_t		resblks = 0;
 363
 364	ASSERT(xfs_isilocked(ip, XFS_ILOCK_EXCL));
 365	if (!ip->i_cowfp) {
 366		ASSERT(!xfs_is_reflink_inode(ip));
 367		xfs_ifork_init_cow(ip);
 368	}
 369
 370	error = xfs_find_trim_cow_extent(ip, imap, shared, &found);
 371	if (error || !*shared)
 
 372		return error;
 373	if (found)
 374		goto convert;
 375
 376	resaligned = xfs_aligned_fsb_count(imap->br_startoff,
 377		imap->br_blockcount, xfs_get_cowextsz_hint(ip));
 378	resblks = XFS_DIOSTRAT_SPACE_RES(mp, resaligned);
 379
 380	xfs_iunlock(ip, *lockmode);
 381	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, resblks, 0, 0, &tp);
 382	*lockmode = XFS_ILOCK_EXCL;
 383	xfs_ilock(ip, *lockmode);
 384
 
 
 
 
 385	if (error)
 386		return error;
 
 387
 388	error = xfs_qm_dqattach_locked(ip, false);
 
 389	if (error)
 390		goto out_trans_cancel;
 391
 392	/*
 393	 * Check for an overlapping extent again now that we dropped the ilock.
 394	 */
 395	error = xfs_find_trim_cow_extent(ip, imap, shared, &found);
 396	if (error || !*shared)
 397		goto out_trans_cancel;
 398	if (found) {
 399		xfs_trans_cancel(tp);
 400		goto convert;
 401	}
 402
 403	error = xfs_trans_reserve_quota_nblks(tp, ip, resblks, 0,
 404			XFS_QMOPT_RES_REGBLKS);
 
 
 
 
 405	if (error)
 406		goto out_trans_cancel;
 407
 408	xfs_trans_ijoin(tp, ip, 0);
 409
 410	/* Allocate the entire reservation as unwritten blocks. */
 411	nimaps = 1;
 412	error = xfs_bmapi_write(tp, ip, imap->br_startoff, imap->br_blockcount,
 413			XFS_BMAPI_COWFORK | XFS_BMAPI_PREALLOC,
 414			resblks, imap, &nimaps);
 415	if (error)
 416		goto out_unreserve;
 417
 418	xfs_inode_set_cowblocks_tag(ip);
 419	error = xfs_trans_commit(tp);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 420	if (error)
 421		return error;
 422
 423	/*
 424	 * Allocation succeeded but the requested range was not even partially
 425	 * satisfied?  Bail out!
 426	 */
 427	if (nimaps == 0)
 428		return -ENOSPC;
 429convert:
 430	xfs_trim_extent(imap, offset_fsb, count_fsb);
 431	/*
 432	 * COW fork extents are supposed to remain unwritten until we're ready
 433	 * to initiate a disk write.  For direct I/O we are going to write the
 434	 * data and need the conversion, but for buffered writes we're done.
 435	 */
 436	if (!convert_now || imap->br_state == XFS_EXT_NORM)
 437		return 0;
 438	trace_xfs_reflink_convert_cow(ip, imap);
 439	return xfs_reflink_convert_cow_locked(ip, offset_fsb, count_fsb);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 440
 441out_unreserve:
 442	xfs_trans_unreserve_quota_nblks(tp, ip, (long)resblks, 0,
 443			XFS_QMOPT_RES_REGBLKS);
 444out_trans_cancel:
 445	xfs_trans_cancel(tp);
 446	return error;
 447}
 448
 449/*
 450 * Cancel CoW reservations for some block range of an inode.
 451 *
 452 * If cancel_real is true this function cancels all COW fork extents for the
 453 * inode; if cancel_real is false, real extents are not cleared.
 454 *
 455 * Caller must have already joined the inode to the current transaction. The
 456 * inode will be joined to the transaction returned to the caller.
 457 */
 458int
 459xfs_reflink_cancel_cow_blocks(
 460	struct xfs_inode		*ip,
 461	struct xfs_trans		**tpp,
 462	xfs_fileoff_t			offset_fsb,
 463	xfs_fileoff_t			end_fsb,
 464	bool				cancel_real)
 465{
 466	struct xfs_ifork		*ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
 467	struct xfs_bmbt_irec		got, del;
 468	struct xfs_iext_cursor		icur;
 
 
 469	int				error = 0;
 470
 471	if (!xfs_inode_has_cow_data(ip))
 472		return 0;
 473	if (!xfs_iext_lookup_extent_before(ip, ifp, &end_fsb, &icur, &got))
 474		return 0;
 475
 476	/* Walk backwards until we're out of the I/O range... */
 477	while (got.br_startoff + got.br_blockcount > offset_fsb) {
 478		del = got;
 479		xfs_trim_extent(&del, offset_fsb, end_fsb - offset_fsb);
 480
 481		/* Extent delete may have bumped ext forward */
 482		if (!del.br_blockcount) {
 483			xfs_iext_prev(ifp, &icur);
 484			goto next_extent;
 485		}
 486
 487		trace_xfs_reflink_cancel_cow(ip, &del);
 488
 489		if (isnullstartblock(del.br_startblock)) {
 490			error = xfs_bmap_del_extent_delay(ip, XFS_COW_FORK,
 491					&icur, &got, &del);
 492			if (error)
 493				break;
 494		} else if (del.br_state == XFS_EXT_UNWRITTEN || cancel_real) {
 495			ASSERT((*tpp)->t_firstblock == NULLFSBLOCK);
 
 496
 497			/* Free the CoW orphan record. */
 498			xfs_refcount_free_cow_extent(*tpp, del.br_startblock,
 
 499					del.br_blockcount);
 
 
 500
 501			xfs_bmap_add_free(*tpp, del.br_startblock,
 502					  del.br_blockcount, NULL);
 
 
 
 
 
 503
 504			/* Roll the transaction */
 505			error = xfs_defer_finish(tpp);
 506			if (error)
 
 507				break;
 
 508
 509			/* Remove the mapping from the CoW fork. */
 510			xfs_bmap_del_extent_cow(ip, &icur, &got, &del);
 
 511
 512			/* Remove the quota reservation */
 513			error = xfs_trans_reserve_quota_nblks(NULL, ip,
 514					-(long)del.br_blockcount, 0,
 515					XFS_QMOPT_RES_REGBLKS);
 516			if (error)
 517				break;
 518		} else {
 519			/* Didn't do anything, push cursor back. */
 520			xfs_iext_prev(ifp, &icur);
 521		}
 522next_extent:
 523		if (!xfs_iext_get_extent(ifp, &icur, &got))
 524			break;
 525	}
 526
 527	/* clear tag if cow fork is emptied */
 528	if (!ifp->if_bytes)
 529		xfs_inode_clear_cowblocks_tag(ip);
 
 530	return error;
 531}
 532
 533/*
 534 * Cancel CoW reservations for some byte range of an inode.
 535 *
 536 * If cancel_real is true this function cancels all COW fork extents for the
 537 * inode; if cancel_real is false, real extents are not cleared.
 538 */
 539int
 540xfs_reflink_cancel_cow_range(
 541	struct xfs_inode	*ip,
 542	xfs_off_t		offset,
 543	xfs_off_t		count,
 544	bool			cancel_real)
 545{
 546	struct xfs_trans	*tp;
 547	xfs_fileoff_t		offset_fsb;
 548	xfs_fileoff_t		end_fsb;
 549	int			error;
 550
 551	trace_xfs_reflink_cancel_cow_range(ip, offset, count);
 552	ASSERT(ip->i_cowfp);
 553
 554	offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
 555	if (count == NULLFILEOFF)
 556		end_fsb = NULLFILEOFF;
 557	else
 558		end_fsb = XFS_B_TO_FSB(ip->i_mount, offset + count);
 559
 560	/* Start a rolling transaction to remove the mappings */
 561	error = xfs_trans_alloc(ip->i_mount, &M_RES(ip->i_mount)->tr_write,
 562			0, 0, 0, &tp);
 563	if (error)
 564		goto out;
 565
 566	xfs_ilock(ip, XFS_ILOCK_EXCL);
 567	xfs_trans_ijoin(tp, ip, 0);
 568
 569	/* Scrape out the old CoW reservations */
 570	error = xfs_reflink_cancel_cow_blocks(ip, &tp, offset_fsb, end_fsb,
 571			cancel_real);
 572	if (error)
 573		goto out_cancel;
 574
 575	error = xfs_trans_commit(tp);
 576
 577	xfs_iunlock(ip, XFS_ILOCK_EXCL);
 578	return error;
 579
 580out_cancel:
 581	xfs_trans_cancel(tp);
 582	xfs_iunlock(ip, XFS_ILOCK_EXCL);
 583out:
 584	trace_xfs_reflink_cancel_cow_range_error(ip, error, _RET_IP_);
 585	return error;
 586}
 587
 588/*
 589 * Remap part of the CoW fork into the data fork.
 590 *
 591 * We aim to remap the range starting at @offset_fsb and ending at @end_fsb
 592 * into the data fork; this function will remap what it can (at the end of the
 593 * range) and update @end_fsb appropriately.  Each remap gets its own
 594 * transaction because we can end up merging and splitting bmbt blocks for
 595 * every remap operation and we'd like to keep the block reservation
 596 * requirements as low as possible.
 597 */
 598STATIC int
 599xfs_reflink_end_cow_extent(
 600	struct xfs_inode	*ip,
 601	xfs_fileoff_t		offset_fsb,
 602	xfs_fileoff_t		*end_fsb)
 603{
 604	struct xfs_bmbt_irec	got, del;
 605	struct xfs_iext_cursor	icur;
 606	struct xfs_mount	*mp = ip->i_mount;
 607	struct xfs_trans	*tp;
 608	struct xfs_ifork	*ifp = XFS_IFORK_PTR(ip, XFS_COW_FORK);
 609	xfs_filblks_t		rlen;
 610	unsigned int		resblks;
 611	int			error;
 
 
 
 
 
 612
 613	/* No COW extents?  That's easy! */
 614	if (ifp->if_bytes == 0) {
 615		*end_fsb = offset_fsb;
 616		return 0;
 617	}
 618
 619	resblks = XFS_EXTENTADD_SPACE_RES(mp, XFS_DATA_FORK);
 620	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, resblks, 0,
 621			XFS_TRANS_RESERVE, &tp);
 
 
 
 
 622	if (error)
 623		return error;
 624
 625	/*
 626	 * Lock the inode.  We have to ijoin without automatic unlock because
 627	 * the lead transaction is the refcountbt record deletion; the data
 628	 * fork update follows as a deferred log item.
 629	 */
 630	xfs_ilock(ip, XFS_ILOCK_EXCL);
 631	xfs_trans_ijoin(tp, ip, 0);
 632
 633	/*
 634	 * In case of racing, overlapping AIO writes no COW extents might be
 635	 * left by the time I/O completes for the loser of the race.  In that
 636	 * case we are done.
 637	 */
 638	if (!xfs_iext_lookup_extent_before(ip, ifp, end_fsb, &icur, &got) ||
 639	    got.br_startoff + got.br_blockcount <= offset_fsb) {
 640		*end_fsb = offset_fsb;
 641		goto out_cancel;
 642	}
 643
 644	/*
 645	 * Structure copy @got into @del, then trim @del to the range that we
 646	 * were asked to remap.  We preserve @got for the eventual CoW fork
 647	 * deletion; from now on @del represents the mapping that we're
 648	 * actually remapping.
 649	 */
 650	del = got;
 651	xfs_trim_extent(&del, offset_fsb, *end_fsb - offset_fsb);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 652
 653	ASSERT(del.br_blockcount > 0);
 
 
 
 
 
 654
 655	/*
 656	 * Only remap real extents that contain data.  With AIO, speculative
 657	 * preallocations can leak into the range we are called upon, and we
 658	 * need to skip them.
 659	 */
 660	if (!xfs_bmap_is_real_extent(&got)) {
 661		*end_fsb = del.br_startoff;
 662		goto out_cancel;
 663	}
 664
 665	/* Unmap the old blocks in the data fork. */
 666	rlen = del.br_blockcount;
 667	error = __xfs_bunmapi(tp, ip, del.br_startoff, &rlen, 0, 1);
 668	if (error)
 669		goto out_cancel;
 670
 671	/* Trim the extent to whatever got unmapped. */
 672	xfs_trim_extent(&del, del.br_startoff + rlen, del.br_blockcount - rlen);
 673	trace_xfs_reflink_cow_remap(ip, &del);
 674
 675	/* Free the CoW orphan record. */
 676	xfs_refcount_free_cow_extent(tp, del.br_startblock, del.br_blockcount);
 677
 678	/* Map the new blocks into the data fork. */
 679	xfs_bmap_map_extent(tp, ip, &del);
 680
 681	/* Charge this new data fork mapping to the on-disk quota. */
 682	xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_DELBCOUNT,
 683			(long)del.br_blockcount);
 684
 685	/* Remove the mapping from the CoW fork. */
 686	xfs_bmap_del_extent_cow(ip, &icur, &got, &del);
 
 
 
 
 
 687
 688	error = xfs_trans_commit(tp);
 689	xfs_iunlock(ip, XFS_ILOCK_EXCL);
 690	if (error)
 691		return error;
 692
 693	/* Update the caller about how much progress we made. */
 694	*end_fsb = del.br_startoff;
 695	return 0;
 696
 697out_cancel:
 
 698	xfs_trans_cancel(tp);
 699	xfs_iunlock(ip, XFS_ILOCK_EXCL);
 700	return error;
 701}
 702
 703/*
 704 * Remap parts of a file's data fork after a successful CoW.
 705 */
 706int
 707xfs_reflink_end_cow(
 708	struct xfs_inode		*ip,
 709	xfs_off_t			offset,
 710	xfs_off_t			count)
 711{
 712	xfs_fileoff_t			offset_fsb;
 713	xfs_fileoff_t			end_fsb;
 714	int				error = 0;
 715
 716	trace_xfs_reflink_end_cow(ip, offset, count);
 717
 718	offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
 719	end_fsb = XFS_B_TO_FSB(ip->i_mount, offset + count);
 720
 721	/*
 722	 * Walk backwards until we're out of the I/O range.  The loop function
 723	 * repeatedly cycles the ILOCK to allocate one transaction per remapped
 724	 * extent.
 725	 *
 726	 * If we're being called by writeback then the the pages will still
 727	 * have PageWriteback set, which prevents races with reflink remapping
 728	 * and truncate.  Reflink remapping prevents races with writeback by
 729	 * taking the iolock and mmaplock before flushing the pages and
 730	 * remapping, which means there won't be any further writeback or page
 731	 * cache dirtying until the reflink completes.
 732	 *
 733	 * We should never have two threads issuing writeback for the same file
 734	 * region.  There are also have post-eof checks in the writeback
 735	 * preparation code so that we don't bother writing out pages that are
 736	 * about to be truncated.
 737	 *
 738	 * If we're being called as part of directio write completion, the dio
 739	 * count is still elevated, which reflink and truncate will wait for.
 740	 * Reflink remapping takes the iolock and mmaplock and waits for
 741	 * pending dio to finish, which should prevent any directio until the
 742	 * remap completes.  Multiple concurrent directio writes to the same
 743	 * region are handled by end_cow processing only occurring for the
 744	 * threads which succeed; the outcome of multiple overlapping direct
 745	 * writes is not well defined anyway.
 746	 *
 747	 * It's possible that a buffered write and a direct write could collide
 748	 * here (the buffered write stumbles in after the dio flushes and
 749	 * invalidates the page cache and immediately queues writeback), but we
 750	 * have never supported this 100%.  If either disk write succeeds the
 751	 * blocks will be remapped.
 752	 */
 753	while (end_fsb > offset_fsb && !error)
 754		error = xfs_reflink_end_cow_extent(ip, offset_fsb, &end_fsb);
 755
 756	if (error)
 757		trace_xfs_reflink_end_cow_error(ip, error, _RET_IP_);
 758	return error;
 759}
 760
 761/*
 762 * Free leftover CoW reservations that didn't get cleaned out.
 763 */
 764int
 765xfs_reflink_recover_cow(
 766	struct xfs_mount	*mp)
 767{
 768	xfs_agnumber_t		agno;
 769	int			error = 0;
 770
 771	if (!xfs_sb_version_hasreflink(&mp->m_sb))
 772		return 0;
 773
 774	for (agno = 0; agno < mp->m_sb.sb_agcount; agno++) {
 775		error = xfs_refcount_recover_cow_leftovers(mp, agno);
 776		if (error)
 777			break;
 778	}
 779
 780	return error;
 781}
 782
 783/*
 784 * Reflinking (Block) Ranges of Two Files Together
 785 *
 786 * First, ensure that the reflink flag is set on both inodes.  The flag is an
 787 * optimization to avoid unnecessary refcount btree lookups in the write path.
 788 *
 789 * Now we can iteratively remap the range of extents (and holes) in src to the
 790 * corresponding ranges in dest.  Let drange and srange denote the ranges of
 791 * logical blocks in dest and src touched by the reflink operation.
 792 *
 793 * While the length of drange is greater than zero,
 794 *    - Read src's bmbt at the start of srange ("imap")
 795 *    - If imap doesn't exist, make imap appear to start at the end of srange
 796 *      with zero length.
 797 *    - If imap starts before srange, advance imap to start at srange.
 798 *    - If imap goes beyond srange, truncate imap to end at the end of srange.
 799 *    - Punch (imap start - srange start + imap len) blocks from dest at
 800 *      offset (drange start).
 801 *    - If imap points to a real range of pblks,
 802 *         > Increase the refcount of the imap's pblks
 803 *         > Map imap's pblks into dest at the offset
 804 *           (drange start + imap start - srange start)
 805 *    - Advance drange and srange by (imap start - srange start + imap len)
 806 *
 807 * Finally, if the reflink made dest longer, update both the in-core and
 808 * on-disk file sizes.
 809 *
 810 * ASCII Art Demonstration:
 811 *
 812 * Let's say we want to reflink this source file:
 813 *
 814 * ----SSSSSSS-SSSSS----SSSSSS (src file)
 815 *   <-------------------->
 816 *
 817 * into this destination file:
 818 *
 819 * --DDDDDDDDDDDDDDDDDDD--DDD (dest file)
 820 *        <-------------------->
 821 * '-' means a hole, and 'S' and 'D' are written blocks in the src and dest.
 822 * Observe that the range has different logical offsets in either file.
 823 *
 824 * Consider that the first extent in the source file doesn't line up with our
 825 * reflink range.  Unmapping  and remapping are separate operations, so we can
 826 * unmap more blocks from the destination file than we remap.
 827 *
 828 * ----SSSSSSS-SSSSS----SSSSSS
 829 *   <------->
 830 * --DDDDD---------DDDDD--DDD
 831 *        <------->
 832 *
 833 * Now remap the source extent into the destination file:
 834 *
 835 * ----SSSSSSS-SSSSS----SSSSSS
 836 *   <------->
 837 * --DDDDD--SSSSSSSDDDDD--DDD
 838 *        <------->
 839 *
 840 * Do likewise with the second hole and extent in our range.  Holes in the
 841 * unmap range don't affect our operation.
 842 *
 843 * ----SSSSSSS-SSSSS----SSSSSS
 844 *            <---->
 845 * --DDDDD--SSSSSSS-SSSSS-DDD
 846 *                 <---->
 847 *
 848 * Finally, unmap and remap part of the third extent.  This will increase the
 849 * size of the destination file.
 850 *
 851 * ----SSSSSSS-SSSSS----SSSSSS
 852 *                  <----->
 853 * --DDDDD--SSSSSSS-SSSSS----SSS
 854 *                       <----->
 855 *
 856 * Once we update the destination file's i_size, we're done.
 857 */
 858
 859/*
 860 * Ensure the reflink bit is set in both inodes.
 861 */
 862STATIC int
 863xfs_reflink_set_inode_flag(
 864	struct xfs_inode	*src,
 865	struct xfs_inode	*dest)
 866{
 867	struct xfs_mount	*mp = src->i_mount;
 868	int			error;
 869	struct xfs_trans	*tp;
 870
 871	if (xfs_is_reflink_inode(src) && xfs_is_reflink_inode(dest))
 872		return 0;
 873
 874	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
 875	if (error)
 876		goto out_error;
 877
 878	/* Lock both files against IO */
 879	if (src->i_ino == dest->i_ino)
 880		xfs_ilock(src, XFS_ILOCK_EXCL);
 881	else
 882		xfs_lock_two_inodes(src, XFS_ILOCK_EXCL, dest, XFS_ILOCK_EXCL);
 883
 884	if (!xfs_is_reflink_inode(src)) {
 885		trace_xfs_reflink_set_inode_flag(src);
 886		xfs_trans_ijoin(tp, src, XFS_ILOCK_EXCL);
 887		src->i_d.di_flags2 |= XFS_DIFLAG2_REFLINK;
 888		xfs_trans_log_inode(tp, src, XFS_ILOG_CORE);
 889		xfs_ifork_init_cow(src);
 890	} else
 891		xfs_iunlock(src, XFS_ILOCK_EXCL);
 892
 893	if (src->i_ino == dest->i_ino)
 894		goto commit_flags;
 895
 896	if (!xfs_is_reflink_inode(dest)) {
 897		trace_xfs_reflink_set_inode_flag(dest);
 898		xfs_trans_ijoin(tp, dest, XFS_ILOCK_EXCL);
 899		dest->i_d.di_flags2 |= XFS_DIFLAG2_REFLINK;
 900		xfs_trans_log_inode(tp, dest, XFS_ILOG_CORE);
 901		xfs_ifork_init_cow(dest);
 902	} else
 903		xfs_iunlock(dest, XFS_ILOCK_EXCL);
 904
 905commit_flags:
 906	error = xfs_trans_commit(tp);
 907	if (error)
 908		goto out_error;
 909	return error;
 910
 911out_error:
 912	trace_xfs_reflink_set_inode_flag_error(dest, error, _RET_IP_);
 913	return error;
 914}
 915
 916/*
 917 * Update destination inode size & cowextsize hint, if necessary.
 918 */
 919int
 920xfs_reflink_update_dest(
 921	struct xfs_inode	*dest,
 922	xfs_off_t		newlen,
 923	xfs_extlen_t		cowextsize,
 924	unsigned int		remap_flags)
 925{
 926	struct xfs_mount	*mp = dest->i_mount;
 927	struct xfs_trans	*tp;
 928	int			error;
 929
 930	if (newlen <= i_size_read(VFS_I(dest)) && cowextsize == 0)
 931		return 0;
 932
 933	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
 934	if (error)
 935		goto out_error;
 936
 937	xfs_ilock(dest, XFS_ILOCK_EXCL);
 938	xfs_trans_ijoin(tp, dest, XFS_ILOCK_EXCL);
 939
 940	if (newlen > i_size_read(VFS_I(dest))) {
 941		trace_xfs_reflink_update_inode_size(dest, newlen);
 942		i_size_write(VFS_I(dest), newlen);
 943		dest->i_d.di_size = newlen;
 944	}
 945
 946	if (cowextsize) {
 947		dest->i_d.di_cowextsize = cowextsize;
 948		dest->i_d.di_flags2 |= XFS_DIFLAG2_COWEXTSIZE;
 949	}
 950
 
 
 
 
 951	xfs_trans_log_inode(tp, dest, XFS_ILOG_CORE);
 952
 953	error = xfs_trans_commit(tp);
 954	if (error)
 955		goto out_error;
 956	return error;
 957
 958out_error:
 959	trace_xfs_reflink_update_inode_size_error(dest, error, _RET_IP_);
 960	return error;
 961}
 962
 963/*
 964 * Do we have enough reserve in this AG to handle a reflink?  The refcount
 965 * btree already reserved all the space it needs, but the rmap btree can grow
 966 * infinitely, so we won't allow more reflinks when the AG is down to the
 967 * btree reserves.
 968 */
 969static int
 970xfs_reflink_ag_has_free_space(
 971	struct xfs_mount	*mp,
 972	xfs_agnumber_t		agno)
 973{
 974	struct xfs_perag	*pag;
 975	int			error = 0;
 976
 977	if (!xfs_sb_version_hasrmapbt(&mp->m_sb))
 978		return 0;
 979
 980	pag = xfs_perag_get(mp, agno);
 981	if (xfs_ag_resv_critical(pag, XFS_AG_RESV_RMAPBT) ||
 982	    xfs_ag_resv_critical(pag, XFS_AG_RESV_METADATA))
 983		error = -ENOSPC;
 984	xfs_perag_put(pag);
 985	return error;
 986}
 987
 988/*
 989 * Unmap a range of blocks from a file, then map other blocks into the hole.
 990 * The range to unmap is (destoff : destoff + srcioff + irec->br_blockcount).
 991 * The extent irec is mapped into dest at irec->br_startoff.
 992 */
 993STATIC int
 994xfs_reflink_remap_extent(
 995	struct xfs_inode	*ip,
 996	struct xfs_bmbt_irec	*irec,
 997	xfs_fileoff_t		destoff,
 998	xfs_off_t		new_isize)
 999{
1000	struct xfs_mount	*mp = ip->i_mount;
1001	bool			real_extent = xfs_bmap_is_real_extent(irec);
1002	struct xfs_trans	*tp;
 
1003	unsigned int		resblks;
 
1004	struct xfs_bmbt_irec	uirec;
 
1005	xfs_filblks_t		rlen;
1006	xfs_filblks_t		unmap_len;
1007	xfs_off_t		newlen;
1008	int			error;
1009
1010	unmap_len = irec->br_startoff + irec->br_blockcount - destoff;
1011	trace_xfs_reflink_punch_range(ip, destoff, unmap_len);
1012
 
 
 
 
 
1013	/* No reflinking if we're low on space */
1014	if (real_extent) {
1015		error = xfs_reflink_ag_has_free_space(mp,
1016				XFS_FSB_TO_AGNO(mp, irec->br_startblock));
1017		if (error)
1018			goto out;
1019	}
1020
1021	/* Start a rolling transaction to switch the mappings */
1022	resblks = XFS_EXTENTADD_SPACE_RES(ip->i_mount, XFS_DATA_FORK);
1023	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, resblks, 0, 0, &tp);
1024	if (error)
1025		goto out;
1026
1027	xfs_ilock(ip, XFS_ILOCK_EXCL);
1028	xfs_trans_ijoin(tp, ip, 0);
1029
1030	/* If we're not just clearing space, then do we have enough quota? */
1031	if (real_extent) {
1032		error = xfs_trans_reserve_quota_nblks(tp, ip,
1033				irec->br_blockcount, 0, XFS_QMOPT_RES_REGBLKS);
1034		if (error)
1035			goto out_cancel;
1036	}
1037
1038	trace_xfs_reflink_remap(ip, irec->br_startoff,
1039				irec->br_blockcount, irec->br_startblock);
1040
1041	/* Unmap the old blocks in the data fork. */
1042	rlen = unmap_len;
1043	while (rlen) {
1044		ASSERT(tp->t_firstblock == NULLFSBLOCK);
1045		error = __xfs_bunmapi(tp, ip, destoff, &rlen, 0, 1);
 
1046		if (error)
1047			goto out_cancel;
1048
1049		/*
1050		 * Trim the extent to whatever got unmapped.
1051		 * Remember, bunmapi works backwards.
1052		 */
1053		uirec.br_startblock = irec->br_startblock + rlen;
1054		uirec.br_startoff = irec->br_startoff + rlen;
1055		uirec.br_blockcount = unmap_len - rlen;
1056		unmap_len = rlen;
1057
1058		/* If this isn't a real mapping, we're done. */
1059		if (!real_extent || uirec.br_blockcount == 0)
1060			goto next_extent;
1061
1062		trace_xfs_reflink_remap(ip, uirec.br_startoff,
1063				uirec.br_blockcount, uirec.br_startblock);
1064
1065		/* Update the refcount tree */
1066		xfs_refcount_increase_extent(tp, &uirec);
 
 
1067
1068		/* Map the new blocks into the data fork. */
1069		xfs_bmap_map_extent(tp, ip, &uirec);
 
 
1070
1071		/* Update quota accounting. */
1072		xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_BCOUNT,
1073				uirec.br_blockcount);
1074
1075		/* Update dest isize if needed. */
1076		newlen = XFS_FSB_TO_B(mp,
1077				uirec.br_startoff + uirec.br_blockcount);
1078		newlen = min_t(xfs_off_t, newlen, new_isize);
1079		if (newlen > i_size_read(VFS_I(ip))) {
1080			trace_xfs_reflink_update_inode_size(ip, newlen);
1081			i_size_write(VFS_I(ip), newlen);
1082			ip->i_d.di_size = newlen;
1083			xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1084		}
1085
1086next_extent:
1087		/* Process all the deferred stuff. */
1088		error = xfs_defer_finish(&tp);
1089		if (error)
1090			goto out_cancel;
1091	}
1092
1093	error = xfs_trans_commit(tp);
1094	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1095	if (error)
1096		goto out;
1097	return 0;
1098
 
 
1099out_cancel:
1100	xfs_trans_cancel(tp);
1101	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1102out:
1103	trace_xfs_reflink_remap_extent_error(ip, error, _RET_IP_);
1104	return error;
1105}
1106
1107/*
1108 * Iteratively remap one file's extents (and holes) to another's.
1109 */
1110int
1111xfs_reflink_remap_blocks(
1112	struct xfs_inode	*src,
1113	loff_t			pos_in,
1114	struct xfs_inode	*dest,
1115	loff_t			pos_out,
1116	loff_t			remap_len,
1117	loff_t			*remapped)
1118{
1119	struct xfs_bmbt_irec	imap;
1120	xfs_fileoff_t		srcoff;
1121	xfs_fileoff_t		destoff;
1122	xfs_filblks_t		len;
1123	xfs_filblks_t		range_len;
1124	xfs_filblks_t		remapped_len = 0;
1125	xfs_off_t		new_isize = pos_out + remap_len;
1126	int			nimaps;
1127	int			error = 0;
1128
1129	destoff = XFS_B_TO_FSBT(src->i_mount, pos_out);
1130	srcoff = XFS_B_TO_FSBT(src->i_mount, pos_in);
1131	len = XFS_B_TO_FSB(src->i_mount, remap_len);
1132
1133	/* drange = (destoff, destoff + len); srange = (srcoff, srcoff + len) */
1134	while (len) {
1135		uint		lock_mode;
1136
1137		trace_xfs_reflink_remap_blocks_loop(src, srcoff, len,
1138				dest, destoff);
1139
1140		/* Read extent from the source file */
1141		nimaps = 1;
1142		lock_mode = xfs_ilock_data_map_shared(src);
1143		error = xfs_bmapi_read(src, srcoff, len, &imap, &nimaps, 0);
1144		xfs_iunlock(src, lock_mode);
1145		if (error)
1146			break;
1147		ASSERT(nimaps == 1);
1148
1149		trace_xfs_reflink_remap_imap(src, srcoff, len, XFS_DATA_FORK,
1150				&imap);
1151
1152		/* Translate imap into the destination file. */
1153		range_len = imap.br_startoff + imap.br_blockcount - srcoff;
1154		imap.br_startoff += destoff - srcoff;
1155
1156		/* Clear dest from destoff to the end of imap and map it in. */
1157		error = xfs_reflink_remap_extent(dest, &imap, destoff,
1158				new_isize);
1159		if (error)
1160			break;
1161
1162		if (fatal_signal_pending(current)) {
1163			error = -EINTR;
1164			break;
1165		}
1166
1167		/* Advance drange/srange */
1168		srcoff += range_len;
1169		destoff += range_len;
1170		len -= range_len;
1171		remapped_len += range_len;
1172	}
1173
1174	if (error)
1175		trace_xfs_reflink_remap_blocks_error(dest, error, _RET_IP_);
1176	*remapped = min_t(loff_t, remap_len,
1177			  XFS_FSB_TO_B(src->i_mount, remapped_len));
1178	return error;
1179}
1180
1181/*
1182 * Grab the exclusive iolock for a data copy from src to dest, making sure to
1183 * abide vfs locking order (lowest pointer value goes first) and breaking the
1184 * layout leases before proceeding.  The loop is needed because we cannot call
1185 * the blocking break_layout() with the iolocks held, and therefore have to
1186 * back out both locks.
1187 */
1188static int
1189xfs_iolock_two_inodes_and_break_layout(
1190	struct inode		*src,
1191	struct inode		*dest)
1192{
1193	int			error;
1194
1195	if (src > dest)
1196		swap(src, dest);
1197
1198retry:
1199	/* Wait to break both inodes' layouts before we start locking. */
1200	error = break_layout(src, true);
1201	if (error)
1202		return error;
1203	if (src != dest) {
1204		error = break_layout(dest, true);
1205		if (error)
1206			return error;
1207	}
1208
1209	/* Lock one inode and make sure nobody got in and leased it. */
1210	inode_lock(src);
1211	error = break_layout(src, false);
1212	if (error) {
1213		inode_unlock(src);
1214		if (error == -EWOULDBLOCK)
1215			goto retry;
1216		return error;
1217	}
1218
1219	if (src == dest)
1220		return 0;
1221
1222	/* Lock the other inode and make sure nobody got in and leased it. */
1223	inode_lock_nested(dest, I_MUTEX_NONDIR2);
1224	error = break_layout(dest, false);
1225	if (error) {
1226		inode_unlock(src);
1227		inode_unlock(dest);
1228		if (error == -EWOULDBLOCK)
1229			goto retry;
1230		return error;
1231	}
1232
1233	return 0;
1234}
1235
1236/* Unlock both inodes after they've been prepped for a range clone. */
1237void
1238xfs_reflink_remap_unlock(
1239	struct file		*file_in,
1240	struct file		*file_out)
1241{
1242	struct inode		*inode_in = file_inode(file_in);
1243	struct xfs_inode	*src = XFS_I(inode_in);
1244	struct inode		*inode_out = file_inode(file_out);
1245	struct xfs_inode	*dest = XFS_I(inode_out);
1246	bool			same_inode = (inode_in == inode_out);
1247
1248	xfs_iunlock(dest, XFS_MMAPLOCK_EXCL);
1249	if (!same_inode)
1250		xfs_iunlock(src, XFS_MMAPLOCK_EXCL);
1251	inode_unlock(inode_out);
1252	if (!same_inode)
1253		inode_unlock(inode_in);
1254}
1255
1256/*
1257 * If we're reflinking to a point past the destination file's EOF, we must
1258 * zero any speculative post-EOF preallocations that sit between the old EOF
1259 * and the destination file offset.
1260 */
1261static int
1262xfs_reflink_zero_posteof(
1263	struct xfs_inode	*ip,
1264	loff_t			pos)
1265{
1266	loff_t			isize = i_size_read(VFS_I(ip));
1267
1268	if (pos <= isize)
1269		return 0;
1270
1271	trace_xfs_zero_eof(ip, isize, pos - isize);
1272	return iomap_zero_range(VFS_I(ip), isize, pos - isize, NULL,
1273			&xfs_iomap_ops);
1274}
1275
1276/*
1277 * Prepare two files for range cloning.  Upon a successful return both inodes
1278 * will have the iolock and mmaplock held, the page cache of the out file will
1279 * be truncated, and any leases on the out file will have been broken.  This
1280 * function borrows heavily from xfs_file_aio_write_checks.
1281 *
1282 * The VFS allows partial EOF blocks to "match" for dedupe even though it hasn't
1283 * checked that the bytes beyond EOF physically match. Hence we cannot use the
1284 * EOF block in the source dedupe range because it's not a complete block match,
1285 * hence can introduce a corruption into the file that has it's block replaced.
1286 *
1287 * In similar fashion, the VFS file cloning also allows partial EOF blocks to be
1288 * "block aligned" for the purposes of cloning entire files.  However, if the
1289 * source file range includes the EOF block and it lands within the existing EOF
1290 * of the destination file, then we can expose stale data from beyond the source
1291 * file EOF in the destination file.
1292 *
1293 * XFS doesn't support partial block sharing, so in both cases we have check
1294 * these cases ourselves. For dedupe, we can simply round the length to dedupe
1295 * down to the previous whole block and ignore the partial EOF block. While this
1296 * means we can't dedupe the last block of a file, this is an acceptible
1297 * tradeoff for simplicity on implementation.
1298 *
1299 * For cloning, we want to share the partial EOF block if it is also the new EOF
1300 * block of the destination file. If the partial EOF block lies inside the
1301 * existing destination EOF, then we have to abort the clone to avoid exposing
1302 * stale data in the destination file. Hence we reject these clone attempts with
1303 * -EINVAL in this case.
1304 */
1305int
1306xfs_reflink_remap_prep(
1307	struct file		*file_in,
1308	loff_t			pos_in,
1309	struct file		*file_out,
1310	loff_t			pos_out,
1311	loff_t			*len,
1312	unsigned int		remap_flags)
1313{
1314	struct inode		*inode_in = file_inode(file_in);
1315	struct xfs_inode	*src = XFS_I(inode_in);
1316	struct inode		*inode_out = file_inode(file_out);
1317	struct xfs_inode	*dest = XFS_I(inode_out);
 
1318	bool			same_inode = (inode_in == inode_out);
 
 
 
1319	ssize_t			ret;
1320
 
 
 
 
 
 
1321	/* Lock both files against IO */
1322	ret = xfs_iolock_two_inodes_and_break_layout(inode_in, inode_out);
1323	if (ret)
1324		return ret;
1325	if (same_inode)
1326		xfs_ilock(src, XFS_MMAPLOCK_EXCL);
1327	else
1328		xfs_lock_two_inodes(src, XFS_MMAPLOCK_EXCL, dest,
1329				XFS_MMAPLOCK_EXCL);
1330
1331	/* Check file eligibility and prepare for block sharing. */
1332	ret = -EINVAL;
1333	/* Don't reflink realtime inodes */
1334	if (XFS_IS_REALTIME_INODE(src) || XFS_IS_REALTIME_INODE(dest))
1335		goto out_unlock;
1336
1337	/* Don't share DAX file data for now. */
1338	if (IS_DAX(inode_in) || IS_DAX(inode_out))
1339		goto out_unlock;
1340
1341	ret = generic_remap_file_range_prep(file_in, pos_in, file_out, pos_out,
1342			len, remap_flags);
1343	if (ret < 0 || *len == 0)
1344		goto out_unlock;
1345
1346	/* Attach dquots to dest inode before changing block map */
1347	ret = xfs_qm_dqattach(dest);
 
 
1348	if (ret)
1349		goto out_unlock;
1350
1351	/*
1352	 * Zero existing post-eof speculative preallocations in the destination
1353	 * file.
1354	 */
1355	ret = xfs_reflink_zero_posteof(dest, pos_out);
1356	if (ret)
1357		goto out_unlock;
1358
1359	/* Set flags and remap blocks. */
1360	ret = xfs_reflink_set_inode_flag(src, dest);
1361	if (ret)
1362		goto out_unlock;
1363
1364	/*
1365	 * If pos_out > EOF, we may have dirtied blocks between EOF and
1366	 * pos_out. In that case, we need to extend the flush and unmap to cover
1367	 * from EOF to the end of the copy length.
1368	 */
1369	if (pos_out > XFS_ISIZE(dest)) {
1370		loff_t	flen = *len + (pos_out - XFS_ISIZE(dest));
1371		ret = xfs_flush_unmap_range(dest, XFS_ISIZE(dest), flen);
1372	} else {
1373		ret = xfs_flush_unmap_range(dest, pos_out, *len);
1374	}
1375	if (ret)
1376		goto out_unlock;
 
1377
1378	return 1;
1379out_unlock:
1380	xfs_reflink_remap_unlock(file_in, file_out);
 
 
 
 
 
1381	return ret;
1382}
1383
1384/*
1385 * The user wants to preemptively CoW all shared blocks in this file,
1386 * which enables us to turn off the reflink flag.  Iterate all
1387 * extents which are not prealloc/delalloc to see which ranges are
1388 * mentioned in the refcount tree, then read those blocks into the
1389 * pagecache, dirty them, fsync them back out, and then we can update
1390 * the inode flag.  What happens if we run out of memory? :)
1391 */
1392STATIC int
1393xfs_reflink_dirty_extents(
1394	struct xfs_inode	*ip,
1395	xfs_fileoff_t		fbno,
1396	xfs_filblks_t		end,
1397	xfs_off_t		isize)
1398{
1399	struct xfs_mount	*mp = ip->i_mount;
1400	xfs_agnumber_t		agno;
1401	xfs_agblock_t		agbno;
1402	xfs_extlen_t		aglen;
1403	xfs_agblock_t		rbno;
1404	xfs_extlen_t		rlen;
1405	xfs_off_t		fpos;
1406	xfs_off_t		flen;
1407	struct xfs_bmbt_irec	map[2];
1408	int			nmaps;
1409	int			error = 0;
1410
1411	while (end - fbno > 0) {
1412		nmaps = 1;
1413		/*
1414		 * Look for extents in the file.  Skip holes, delalloc, or
1415		 * unwritten extents; they can't be reflinked.
1416		 */
1417		error = xfs_bmapi_read(ip, fbno, end - fbno, map, &nmaps, 0);
1418		if (error)
1419			goto out;
1420		if (nmaps == 0)
1421			break;
1422		if (!xfs_bmap_is_real_extent(&map[0]))
 
 
1423			goto next;
1424
1425		map[1] = map[0];
1426		while (map[1].br_blockcount) {
1427			agno = XFS_FSB_TO_AGNO(mp, map[1].br_startblock);
1428			agbno = XFS_FSB_TO_AGBNO(mp, map[1].br_startblock);
1429			aglen = map[1].br_blockcount;
1430
1431			error = xfs_reflink_find_shared(mp, NULL, agno, agbno,
1432					aglen, &rbno, &rlen, true);
1433			if (error)
1434				goto out;
1435			if (rbno == NULLAGBLOCK)
1436				break;
1437
1438			/* Dirty the pages */
1439			xfs_iunlock(ip, XFS_ILOCK_EXCL);
1440			fpos = XFS_FSB_TO_B(mp, map[1].br_startoff +
1441					(rbno - agbno));
1442			flen = XFS_FSB_TO_B(mp, rlen);
1443			if (fpos + flen > isize)
1444				flen = isize - fpos;
1445			error = iomap_file_dirty(VFS_I(ip), fpos, flen,
1446					&xfs_iomap_ops);
1447			xfs_ilock(ip, XFS_ILOCK_EXCL);
1448			if (error)
1449				goto out;
1450
1451			map[1].br_blockcount -= (rbno - agbno + rlen);
1452			map[1].br_startoff += (rbno - agbno + rlen);
1453			map[1].br_startblock += (rbno - agbno + rlen);
1454		}
1455
1456next:
1457		fbno = map[0].br_startoff + map[0].br_blockcount;
1458	}
1459out:
1460	return error;
1461}
1462
1463/* Does this inode need the reflink flag? */
1464int
1465xfs_reflink_inode_has_shared_extents(
1466	struct xfs_trans		*tp,
1467	struct xfs_inode		*ip,
1468	bool				*has_shared)
1469{
1470	struct xfs_bmbt_irec		got;
1471	struct xfs_mount		*mp = ip->i_mount;
1472	struct xfs_ifork		*ifp;
1473	xfs_agnumber_t			agno;
1474	xfs_agblock_t			agbno;
1475	xfs_extlen_t			aglen;
1476	xfs_agblock_t			rbno;
1477	xfs_extlen_t			rlen;
1478	struct xfs_iext_cursor		icur;
1479	bool				found;
1480	int				error;
 
 
1481
1482	ifp = XFS_IFORK_PTR(ip, XFS_DATA_FORK);
1483	if (!(ifp->if_flags & XFS_IFEXTENTS)) {
1484		error = xfs_iread_extents(tp, ip, XFS_DATA_FORK);
 
 
 
 
 
 
1485		if (error)
1486			return error;
1487	}
 
 
 
 
 
1488
1489	*has_shared = false;
1490	found = xfs_iext_lookup_extent(ip, ifp, 0, &icur, &got);
1491	while (found) {
1492		if (isnullstartblock(got.br_startblock) ||
1493		    got.br_state != XFS_EXT_NORM)
1494			goto next;
1495		agno = XFS_FSB_TO_AGNO(mp, got.br_startblock);
1496		agbno = XFS_FSB_TO_AGBNO(mp, got.br_startblock);
1497		aglen = got.br_blockcount;
1498
1499		error = xfs_reflink_find_shared(mp, tp, agno, agbno, aglen,
1500				&rbno, &rlen, false);
1501		if (error)
1502			return error;
1503		/* Is there still a shared block here? */
1504		if (rbno != NULLAGBLOCK) {
1505			*has_shared = true;
1506			return 0;
1507		}
1508next:
1509		found = xfs_iext_next_extent(ifp, &icur, &got);
1510	}
1511
1512	return 0;
1513}
1514
1515/*
1516 * Clear the inode reflink flag if there are no shared extents.
1517 *
1518 * The caller is responsible for joining the inode to the transaction passed in.
1519 * The inode will be joined to the transaction that is returned to the caller.
1520 */
1521int
1522xfs_reflink_clear_inode_flag(
1523	struct xfs_inode	*ip,
1524	struct xfs_trans	**tpp)
1525{
1526	bool			needs_flag;
1527	int			error = 0;
1528
1529	ASSERT(xfs_is_reflink_inode(ip));
1530
1531	error = xfs_reflink_inode_has_shared_extents(*tpp, ip, &needs_flag);
1532	if (error || needs_flag)
1533		return error;
1534
1535	/*
1536	 * We didn't find any shared blocks so turn off the reflink flag.
1537	 * First, get rid of any leftover CoW mappings.
1538	 */
1539	error = xfs_reflink_cancel_cow_blocks(ip, tpp, 0, NULLFILEOFF, true);
1540	if (error)
1541		return error;
1542
1543	/* Clear the inode flag. */
1544	trace_xfs_reflink_unset_inode_flag(ip);
1545	ip->i_d.di_flags2 &= ~XFS_DIFLAG2_REFLINK;
1546	xfs_inode_clear_cowblocks_tag(ip);
 
1547	xfs_trans_log_inode(*tpp, ip, XFS_ILOG_CORE);
1548
1549	return error;
1550}
1551
1552/*
1553 * Clear the inode reflink flag if there are no shared extents and the size
1554 * hasn't changed.
1555 */
1556STATIC int
1557xfs_reflink_try_clear_inode_flag(
1558	struct xfs_inode	*ip)
1559{
1560	struct xfs_mount	*mp = ip->i_mount;
1561	struct xfs_trans	*tp;
1562	int			error = 0;
1563
1564	/* Start a rolling transaction to remove the mappings */
1565	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, 0, 0, 0, &tp);
1566	if (error)
1567		return error;
1568
1569	xfs_ilock(ip, XFS_ILOCK_EXCL);
1570	xfs_trans_ijoin(tp, ip, 0);
1571
1572	error = xfs_reflink_clear_inode_flag(ip, &tp);
1573	if (error)
1574		goto cancel;
1575
1576	error = xfs_trans_commit(tp);
1577	if (error)
1578		goto out;
1579
1580	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1581	return 0;
1582cancel:
1583	xfs_trans_cancel(tp);
1584out:
1585	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1586	return error;
1587}
1588
1589/*
1590 * Pre-COW all shared blocks within a given byte range of a file and turn off
1591 * the reflink flag if we unshare all of the file's blocks.
1592 */
1593int
1594xfs_reflink_unshare(
1595	struct xfs_inode	*ip,
1596	xfs_off_t		offset,
1597	xfs_off_t		len)
1598{
1599	struct xfs_mount	*mp = ip->i_mount;
1600	xfs_fileoff_t		fbno;
1601	xfs_filblks_t		end;
1602	xfs_off_t		isize;
1603	int			error;
1604
1605	if (!xfs_is_reflink_inode(ip))
1606		return 0;
1607
1608	trace_xfs_reflink_unshare(ip, offset, len);
1609
1610	inode_dio_wait(VFS_I(ip));
1611
1612	/* Try to CoW the selected ranges */
1613	xfs_ilock(ip, XFS_ILOCK_EXCL);
1614	fbno = XFS_B_TO_FSBT(mp, offset);
1615	isize = i_size_read(VFS_I(ip));
1616	end = XFS_B_TO_FSB(mp, offset + len);
1617	error = xfs_reflink_dirty_extents(ip, fbno, end, isize);
1618	if (error)
1619		goto out_unlock;
1620	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1621
1622	/* Wait for the IO to finish */
1623	error = filemap_write_and_wait(VFS_I(ip)->i_mapping);
1624	if (error)
1625		goto out;
1626
1627	/* Turn off the reflink flag if possible. */
1628	error = xfs_reflink_try_clear_inode_flag(ip);
1629	if (error)
1630		goto out;
1631
1632	return 0;
1633
1634out_unlock:
1635	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1636out:
1637	trace_xfs_reflink_unshare_error(ip, error, _RET_IP_);
1638	return error;
1639}