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}
v6.9.4
   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_ag.h"
 
  31#include "xfs_ag_resv.h"
  32#include "xfs_health.h"
  33
  34/*
  35 * Copy on Write of Shared Blocks
  36 *
  37 * XFS must preserve "the usual" file semantics even when two files share
  38 * the same physical blocks.  This means that a write to one file must not
  39 * alter the blocks in a different file; the way that we'll do that is
  40 * through the use of a copy-on-write mechanism.  At a high level, that
  41 * means that when we want to write to a shared block, we allocate a new
  42 * block, write the data to the new block, and if that succeeds we map the
  43 * new block into the file.
  44 *
  45 * XFS provides a "delayed allocation" mechanism that defers the allocation
  46 * of disk blocks to dirty-but-not-yet-mapped file blocks as long as
  47 * possible.  This reduces fragmentation by enabling the filesystem to ask
  48 * for bigger chunks less often, which is exactly what we want for CoW.
  49 *
  50 * The delalloc mechanism begins when the kernel wants to make a block
  51 * writable (write_begin or page_mkwrite).  If the offset is not mapped, we
  52 * create a delalloc mapping, which is a regular in-core extent, but without
  53 * a real startblock.  (For delalloc mappings, the startblock encodes both
  54 * a flag that this is a delalloc mapping, and a worst-case estimate of how
  55 * many blocks might be required to put the mapping into the BMBT.)  delalloc
  56 * mappings are a reservation against the free space in the filesystem;
  57 * adjacent mappings can also be combined into fewer larger mappings.
  58 *
  59 * As an optimization, the CoW extent size hint (cowextsz) creates
  60 * outsized aligned delalloc reservations in the hope of landing out of
  61 * order nearby CoW writes in a single extent on disk, thereby reducing
  62 * fragmentation and improving future performance.
  63 *
  64 * D: --RRRRRRSSSRRRRRRRR--- (data fork)
  65 * C: ------DDDDDDD--------- (CoW fork)
  66 *
  67 * When dirty pages are being written out (typically in writepage), the
  68 * delalloc reservations are converted into unwritten mappings by
  69 * allocating blocks and replacing the delalloc mapping with real ones.
  70 * A delalloc mapping can be replaced by several unwritten ones if the
  71 * free space is fragmented.
  72 *
  73 * D: --RRRRRRSSSRRRRRRRR---
  74 * C: ------UUUUUUU---------
  75 *
  76 * We want to adapt the delalloc mechanism for copy-on-write, since the
  77 * write paths are similar.  The first two steps (creating the reservation
  78 * and allocating the blocks) are exactly the same as delalloc except that
  79 * the mappings must be stored in a separate CoW fork because we do not want
  80 * to disturb the mapping in the data fork until we're sure that the write
  81 * succeeded.  IO completion in this case is the process of removing the old
  82 * mapping from the data fork and moving the new mapping from the CoW fork to
  83 * the data fork.  This will be discussed shortly.
  84 *
  85 * For now, unaligned directio writes will be bounced back to the page cache.
  86 * Block-aligned directio writes will use the same mechanism as buffered
  87 * writes.
  88 *
  89 * Just prior to submitting the actual disk write requests, we convert
  90 * the extents representing the range of the file actually being written
  91 * (as opposed to extra pieces created for the cowextsize hint) to real
  92 * extents.  This will become important in the next step:
  93 *
  94 * D: --RRRRRRSSSRRRRRRRR---
  95 * C: ------UUrrUUU---------
  96 *
  97 * CoW remapping must be done after the data block write completes,
  98 * because we don't want to destroy the old data fork map until we're sure
  99 * the new block has been written.  Since the new mappings are kept in a
 100 * separate fork, we can simply iterate these mappings to find the ones
 101 * that cover the file blocks that we just CoW'd.  For each extent, simply
 102 * unmap the corresponding range in the data fork, map the new range into
 103 * the data fork, and remove the extent from the CoW fork.  Because of
 104 * the presence of the cowextsize hint, however, we must be careful
 105 * only to remap the blocks that we've actually written out --  we must
 106 * never remap delalloc reservations nor CoW staging blocks that have
 107 * yet to be written.  This corresponds exactly to the real extents in
 108 * the CoW fork:
 109 *
 110 * D: --RRRRRRrrSRRRRRRRR---
 111 * C: ------UU--UUU---------
 112 *
 113 * Since the remapping operation can be applied to an arbitrary file
 114 * range, we record the need for the remap step as a flag in the ioend
 115 * instead of declaring a new IO type.  This is required for direct io
 116 * because we only have ioend for the whole dio, and we have to be able to
 117 * remember the presence of unwritten blocks and CoW blocks with a single
 118 * ioend structure.  Better yet, the more ground we can cover with one
 119 * ioend, the better.
 120 */
 121
 122/*
 123 * Given an AG extent, find the lowest-numbered run of shared blocks
 124 * within that range and return the range in fbno/flen.  If
 125 * find_end_of_shared is true, return the longest contiguous extent of
 126 * shared blocks.  If there are no shared extents, fbno and flen will
 127 * be set to NULLAGBLOCK and 0, respectively.
 128 */
 129static int
 130xfs_reflink_find_shared(
 131	struct xfs_perag	*pag,
 132	struct xfs_trans	*tp,
 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(pag, tp, 0, &agbp);
 144	if (error)
 145		return error;
 146
 147	cur = xfs_refcountbt_init_cursor(pag->pag_mount, tp, agbp, pag);
 148
 149	error = xfs_refcount_find_shared(cur, agbno, aglen, fbno, flen,
 150			find_end_of_shared);
 151
 152	xfs_btree_del_cursor(cur, error);
 153
 154	xfs_trans_brelse(tp, agbp);
 155	return error;
 156}
 157
 158/*
 159 * Trim the mapping to the next block where there's a change in the
 160 * shared/unshared status.  More specifically, this means that we
 161 * find the lowest-numbered extent of shared blocks that coincides with
 162 * the given block mapping.  If the shared extent overlaps the start of
 163 * the mapping, trim the mapping to the end of the shared extent.  If
 164 * the shared region intersects the mapping, trim the mapping to the
 165 * start of the shared extent.  If there are no shared regions that
 166 * overlap, just return the original extent.
 167 */
 168int
 169xfs_reflink_trim_around_shared(
 170	struct xfs_inode	*ip,
 171	struct xfs_bmbt_irec	*irec,
 172	bool			*shared)
 
 173{
 174	struct xfs_mount	*mp = ip->i_mount;
 175	struct xfs_perag	*pag;
 176	xfs_agblock_t		agbno;
 177	xfs_extlen_t		aglen;
 178	xfs_agblock_t		fbno;
 179	xfs_extlen_t		flen;
 180	int			error = 0;
 181
 182	/* Holes, unwritten, and delalloc extents cannot be shared */
 183	if (!xfs_is_cow_inode(ip) || !xfs_bmap_is_written_extent(irec)) {
 
 
 
 
 184		*shared = false;
 185		return 0;
 186	}
 187
 188	trace_xfs_reflink_trim_around_shared(ip, irec);
 189
 190	pag = xfs_perag_get(mp, XFS_FSB_TO_AGNO(mp, irec->br_startblock));
 191	agbno = XFS_FSB_TO_AGBNO(mp, irec->br_startblock);
 192	aglen = irec->br_blockcount;
 193
 194	error = xfs_reflink_find_shared(pag, NULL, agbno, aglen, &fbno, &flen,
 195			true);
 196	xfs_perag_put(pag);
 197	if (error)
 198		return error;
 199
 200	*shared = false;
 201	if (fbno == NULLAGBLOCK) {
 202		/* No shared blocks at all. */
 203		return 0;
 204	}
 205
 206	if (fbno == agbno) {
 207		/*
 208		 * The start of this extent is shared.  Truncate the
 209		 * mapping at the end of the shared region so that a
 210		 * subsequent iteration starts at the start of the
 211		 * unshared region.
 212		 */
 213		irec->br_blockcount = flen;
 214		*shared = true;
 
 
 
 
 
 
 
 
 
 
 
 
 215		return 0;
 216	}
 217
 218	/*
 219	 * There's a shared extent midway through this extent.
 220	 * Truncate the mapping at the start of the shared
 221	 * extent so that a subsequent iteration starts at the
 222	 * start of the shared region.
 223	 */
 224	irec->br_blockcount = fbno - agbno;
 225	return 0;
 226}
 227
 
 
 
 
 
 
 
 
 
 228int
 229xfs_bmap_trim_cow(
 230	struct xfs_inode	*ip,
 231	struct xfs_bmbt_irec	*imap,
 232	bool			*shared)
 233{
 234	/* We can't update any real extents in always COW mode. */
 235	if (xfs_is_always_cow_inode(ip) &&
 236	    !isnullstartblock(imap->br_startblock)) {
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 237		*shared = true;
 238		return 0;
 239	}
 240
 241	/* Trim the mapping to the nearest shared extent boundary. */
 242	return xfs_reflink_trim_around_shared(ip, imap, shared);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 243}
 244
 245static int
 246xfs_reflink_convert_cow_locked(
 247	struct xfs_inode	*ip,
 248	xfs_fileoff_t		offset_fsb,
 249	xfs_filblks_t		count_fsb)
 
 
 
 250{
 251	struct xfs_iext_cursor	icur;
 252	struct xfs_bmbt_irec	got;
 253	struct xfs_btree_cur	*dummy_cur = NULL;
 254	int			dummy_logflags;
 255	int			error = 0;
 256
 257	if (!xfs_iext_lookup_extent(ip, ip->i_cowfp, offset_fsb, &icur, &got))
 258		return 0;
 259
 260	do {
 261		if (got.br_startoff >= offset_fsb + count_fsb)
 262			break;
 263		if (got.br_state == XFS_EXT_NORM)
 264			continue;
 265		if (WARN_ON_ONCE(isnullstartblock(got.br_startblock)))
 266			return -EIO;
 267
 268		xfs_trim_extent(&got, offset_fsb, count_fsb);
 269		if (!got.br_blockcount)
 270			continue;
 271
 272		got.br_state = XFS_EXT_NORM;
 273		error = xfs_bmap_add_extent_unwritten_real(NULL, ip,
 274				XFS_COW_FORK, &icur, &dummy_cur, &got,
 275				&dummy_logflags);
 276		if (error)
 277			return error;
 278	} while (xfs_iext_next_extent(ip->i_cowfp, &icur, &got));
 279
 280	return error;
 281}
 282
 283/* Convert all of the unwritten CoW extents in a file's range to real ones. */
 284int
 285xfs_reflink_convert_cow(
 286	struct xfs_inode	*ip,
 287	xfs_off_t		offset,
 288	xfs_off_t		count)
 289{
 
 
 290	struct xfs_mount	*mp = ip->i_mount;
 
 291	xfs_fileoff_t		offset_fsb = XFS_B_TO_FSBT(mp, offset);
 292	xfs_fileoff_t		end_fsb = XFS_B_TO_FSB(mp, offset + count);
 293	xfs_filblks_t		count_fsb = end_fsb - offset_fsb;
 294	int			error;
 295
 296	ASSERT(count != 0);
 297
 298	xfs_ilock(ip, XFS_ILOCK_EXCL);
 299	error = xfs_reflink_convert_cow_locked(ip, offset_fsb, count_fsb);
 300	xfs_iunlock(ip, XFS_ILOCK_EXCL);
 301	return error;
 302}
 303
 304/*
 305 * Find the extent that maps the given range in the COW fork. Even if the extent
 306 * is not shared we might have a preallocation for it in the COW fork. If so we
 307 * use it that rather than trigger a new allocation.
 308 */
 309static int
 310xfs_find_trim_cow_extent(
 311	struct xfs_inode	*ip,
 312	struct xfs_bmbt_irec	*imap,
 313	struct xfs_bmbt_irec	*cmap,
 314	bool			*shared,
 315	bool			*found)
 316{
 317	xfs_fileoff_t		offset_fsb = imap->br_startoff;
 318	xfs_filblks_t		count_fsb = imap->br_blockcount;
 319	struct xfs_iext_cursor	icur;
 320
 321	*found = false;
 322
 323	/*
 324	 * If we don't find an overlapping extent, trim the range we need to
 325	 * allocate to fit the hole we found.
 326	 */
 327	if (!xfs_iext_lookup_extent(ip, ip->i_cowfp, offset_fsb, &icur, cmap))
 328		cmap->br_startoff = offset_fsb + count_fsb;
 329	if (cmap->br_startoff > offset_fsb) {
 330		xfs_trim_extent(imap, imap->br_startoff,
 331				cmap->br_startoff - imap->br_startoff);
 332		return xfs_bmap_trim_cow(ip, imap, shared);
 333	}
 334
 335	*shared = true;
 336	if (isnullstartblock(cmap->br_startblock)) {
 337		xfs_trim_extent(imap, cmap->br_startoff, cmap->br_blockcount);
 338		return 0;
 339	}
 340
 341	/* real extent found - no need to allocate */
 342	xfs_trim_extent(cmap, offset_fsb, count_fsb);
 343	*found = true;
 344	return 0;
 345}
 346
 347static int
 348xfs_reflink_convert_unwritten(
 349	struct xfs_inode	*ip,
 350	struct xfs_bmbt_irec	*imap,
 351	struct xfs_bmbt_irec	*cmap,
 352	bool			convert_now)
 353{
 354	xfs_fileoff_t		offset_fsb = imap->br_startoff;
 355	xfs_filblks_t		count_fsb = imap->br_blockcount;
 356	int			error;
 357
 358	/*
 359	 * cmap might larger than imap due to cowextsize hint.
 360	 */
 361	xfs_trim_extent(cmap, offset_fsb, count_fsb);
 362
 363	/*
 364	 * COW fork extents are supposed to remain unwritten until we're ready
 365	 * to initiate a disk write.  For direct I/O we are going to write the
 366	 * data and need the conversion, but for buffered writes we're done.
 367	 */
 368	if (!convert_now || cmap->br_state == XFS_EXT_NORM)
 369		return 0;
 370
 371	trace_xfs_reflink_convert_cow(ip, cmap);
 372
 373	error = xfs_reflink_convert_cow_locked(ip, offset_fsb, count_fsb);
 374	if (!error)
 375		cmap->br_state = XFS_EXT_NORM;
 376
 377	return error;
 378}
 379
 
 380static int
 381xfs_reflink_fill_cow_hole(
 382	struct xfs_inode	*ip,
 383	struct xfs_bmbt_irec	*imap,
 384	struct xfs_bmbt_irec	*cmap,
 385	bool			*shared,
 386	uint			*lockmode,
 387	bool			convert_now)
 388{
 389	struct xfs_mount	*mp = ip->i_mount;
 
 
 390	struct xfs_trans	*tp;
 391	xfs_filblks_t		resaligned;
 392	xfs_extlen_t		resblks;
 393	int			nimaps;
 394	int			error;
 395	bool			found;
 396
 397	resaligned = xfs_aligned_fsb_count(imap->br_startoff,
 398		imap->br_blockcount, xfs_get_cowextsz_hint(ip));
 399	resblks = XFS_DIOSTRAT_SPACE_RES(mp, resaligned);
 400
 401	xfs_iunlock(ip, *lockmode);
 402	*lockmode = 0;
 403
 404	error = xfs_trans_alloc_inode(ip, &M_RES(mp)->tr_write, resblks, 0,
 405			false, &tp);
 406	if (error)
 407		return error;
 408
 409	*lockmode = XFS_ILOCK_EXCL;
 
 
 
 
 
 
 
 
 410
 411	error = xfs_find_trim_cow_extent(ip, imap, cmap, shared, &found);
 412	if (error || !*shared)
 
 413		goto out_trans_cancel;
 414
 415	if (found) {
 416		xfs_trans_cancel(tp);
 417		goto convert;
 418	}
 419
 420	/* Allocate the entire reservation as unwritten blocks. */
 421	nimaps = 1;
 422	error = xfs_bmapi_write(tp, ip, imap->br_startoff, imap->br_blockcount,
 423			XFS_BMAPI_COWFORK | XFS_BMAPI_PREALLOC, 0, cmap,
 424			&nimaps);
 
 425	if (error)
 426		goto out_trans_cancel;
 427
 428	xfs_inode_set_cowblocks_tag(ip);
 429	error = xfs_trans_commit(tp);
 430	if (error)
 431		return error;
 432
 433	/*
 434	 * Allocation succeeded but the requested range was not even partially
 435	 * satisfied?  Bail out!
 436	 */
 437	if (nimaps == 0)
 438		return -ENOSPC;
 439
 440convert:
 441	return xfs_reflink_convert_unwritten(ip, imap, cmap, convert_now);
 442
 
 
 
 
 443out_trans_cancel:
 
 444	xfs_trans_cancel(tp);
 445	return error;
 446}
 447
 448static int
 449xfs_reflink_fill_delalloc(
 
 450	struct xfs_inode	*ip,
 451	struct xfs_bmbt_irec	*imap,
 452	struct xfs_bmbt_irec	*cmap,
 453	bool			*shared,
 454	uint			*lockmode,
 455	bool			convert_now)
 456{
 457	struct xfs_mount	*mp = ip->i_mount;
 458	struct xfs_trans	*tp;
 459	int			nimaps;
 460	int			error;
 461	bool			found;
 462
 463	do {
 464		xfs_iunlock(ip, *lockmode);
 465		*lockmode = 0;
 466
 467		error = xfs_trans_alloc_inode(ip, &M_RES(mp)->tr_write, 0, 0,
 468				false, &tp);
 469		if (error)
 470			return error;
 471
 472		*lockmode = XFS_ILOCK_EXCL;
 
 
 
 
 
 473
 474		error = xfs_find_trim_cow_extent(ip, imap, cmap, shared,
 475				&found);
 476		if (error || !*shared)
 477			goto out_trans_cancel;
 478
 479		if (found) {
 480			xfs_trans_cancel(tp);
 481			break;
 482		}
 
 483
 484		ASSERT(isnullstartblock(cmap->br_startblock) ||
 485		       cmap->br_startblock == DELAYSTARTBLOCK);
 
 486
 487		/*
 488		 * Replace delalloc reservation with an unwritten extent.
 489		 */
 490		nimaps = 1;
 491		error = xfs_bmapi_write(tp, ip, cmap->br_startoff,
 492				cmap->br_blockcount,
 493				XFS_BMAPI_COWFORK | XFS_BMAPI_PREALLOC, 0,
 494				cmap, &nimaps);
 495		if (error)
 496			goto out_trans_cancel;
 
 
 
 497
 498		xfs_inode_set_cowblocks_tag(ip);
 499		error = xfs_trans_commit(tp);
 500		if (error)
 501			return error;
 502
 503		/*
 504		 * Allocation succeeded but the requested range was not even
 505		 * partially satisfied?  Bail out!
 506		 */
 507		if (nimaps == 0)
 508			return -ENOSPC;
 509	} while (cmap->br_startoff + cmap->br_blockcount <= imap->br_startoff);
 510
 511	return xfs_reflink_convert_unwritten(ip, imap, cmap, convert_now);
 512
 513out_trans_cancel:
 514	xfs_trans_cancel(tp);
 515	return error;
 516}
 517
 518/* Allocate all CoW reservations covering a range of blocks in a file. */
 519int
 520xfs_reflink_allocate_cow(
 521	struct xfs_inode	*ip,
 522	struct xfs_bmbt_irec	*imap,
 523	struct xfs_bmbt_irec	*cmap,
 524	bool			*shared,
 525	uint			*lockmode,
 526	bool			convert_now)
 527{
 528	int			error;
 529	bool			found;
 
 530
 531	xfs_assert_ilocked(ip, XFS_ILOCK_EXCL);
 532	if (!ip->i_cowfp) {
 533		ASSERT(!xfs_is_reflink_inode(ip));
 534		xfs_ifork_init_cow(ip);
 535	}
 536
 537	error = xfs_find_trim_cow_extent(ip, imap, cmap, shared, &found);
 538	if (error || !*shared)
 539		return error;
 540
 541	/* CoW fork has a real extent */
 542	if (found)
 543		return xfs_reflink_convert_unwritten(ip, imap, cmap,
 544				convert_now);
 
 545
 546	/*
 547	 * CoW fork does not have an extent and data extent is shared.
 548	 * Allocate a real extent in the CoW fork.
 549	 */
 550	if (cmap->br_startoff > imap->br_startoff)
 551		return xfs_reflink_fill_cow_hole(ip, imap, cmap, shared,
 552				lockmode, convert_now);
 553
 554	/*
 555	 * CoW fork has a delalloc reservation. Replace it with a real extent.
 556	 * There may or may not be a data fork mapping.
 557	 */
 558	if (isnullstartblock(cmap->br_startblock) ||
 559	    cmap->br_startblock == DELAYSTARTBLOCK)
 560		return xfs_reflink_fill_delalloc(ip, imap, cmap, shared,
 561				lockmode, convert_now);
 562
 563	/* Shouldn't get here. */
 564	ASSERT(0);
 565	return -EFSCORRUPTED;
 566}
 567
 568/*
 569 * Cancel CoW reservations for some block range of an inode.
 570 *
 571 * If cancel_real is true this function cancels all COW fork extents for the
 572 * inode; if cancel_real is false, real extents are not cleared.
 573 *
 574 * Caller must have already joined the inode to the current transaction. The
 575 * inode will be joined to the transaction returned to the caller.
 576 */
 577int
 578xfs_reflink_cancel_cow_blocks(
 579	struct xfs_inode		*ip,
 580	struct xfs_trans		**tpp,
 581	xfs_fileoff_t			offset_fsb,
 582	xfs_fileoff_t			end_fsb,
 583	bool				cancel_real)
 584{
 585	struct xfs_ifork		*ifp = xfs_ifork_ptr(ip, XFS_COW_FORK);
 586	struct xfs_bmbt_irec		got, del;
 587	struct xfs_iext_cursor		icur;
 
 
 588	int				error = 0;
 589
 590	if (!xfs_inode_has_cow_data(ip))
 591		return 0;
 592	if (!xfs_iext_lookup_extent_before(ip, ifp, &end_fsb, &icur, &got))
 593		return 0;
 594
 595	/* Walk backwards until we're out of the I/O range... */
 596	while (got.br_startoff + got.br_blockcount > offset_fsb) {
 597		del = got;
 598		xfs_trim_extent(&del, offset_fsb, end_fsb - offset_fsb);
 599
 600		/* Extent delete may have bumped ext forward */
 601		if (!del.br_blockcount) {
 602			xfs_iext_prev(ifp, &icur);
 603			goto next_extent;
 604		}
 605
 606		trace_xfs_reflink_cancel_cow(ip, &del);
 607
 608		if (isnullstartblock(del.br_startblock)) {
 609			error = xfs_bmap_del_extent_delay(ip, XFS_COW_FORK,
 610					&icur, &got, &del);
 611			if (error)
 612				break;
 613		} else if (del.br_state == XFS_EXT_UNWRITTEN || cancel_real) {
 614			ASSERT((*tpp)->t_highest_agno == NULLAGNUMBER);
 
 615
 616			/* Free the CoW orphan record. */
 617			xfs_refcount_free_cow_extent(*tpp, del.br_startblock,
 
 618					del.br_blockcount);
 619
 620			error = xfs_free_extent_later(*tpp, del.br_startblock,
 621					del.br_blockcount, NULL,
 622					XFS_AG_RESV_NONE, false);
 623			if (error)
 624				break;
 625
 
 
 
 
 
 
 
 
 626			/* Roll the transaction */
 627			error = xfs_defer_finish(tpp);
 628			if (error)
 
 629				break;
 
 630
 631			/* Remove the mapping from the CoW fork. */
 632			xfs_bmap_del_extent_cow(ip, &icur, &got, &del);
 
 633
 634			/* Remove the quota reservation */
 635			error = xfs_quota_unreserve_blkres(ip,
 636					del.br_blockcount);
 637			if (error)
 638				break;
 639		} else {
 640			/* Didn't do anything, push cursor back. */
 641			xfs_iext_prev(ifp, &icur);
 642		}
 643next_extent:
 644		if (!xfs_iext_get_extent(ifp, &icur, &got))
 645			break;
 646	}
 647
 648	/* clear tag if cow fork is emptied */
 649	if (!ifp->if_bytes)
 650		xfs_inode_clear_cowblocks_tag(ip);
 
 651	return error;
 652}
 653
 654/*
 655 * Cancel CoW reservations for some byte range of an inode.
 656 *
 657 * If cancel_real is true this function cancels all COW fork extents for the
 658 * inode; if cancel_real is false, real extents are not cleared.
 659 */
 660int
 661xfs_reflink_cancel_cow_range(
 662	struct xfs_inode	*ip,
 663	xfs_off_t		offset,
 664	xfs_off_t		count,
 665	bool			cancel_real)
 666{
 667	struct xfs_trans	*tp;
 668	xfs_fileoff_t		offset_fsb;
 669	xfs_fileoff_t		end_fsb;
 670	int			error;
 671
 672	trace_xfs_reflink_cancel_cow_range(ip, offset, count);
 673	ASSERT(ip->i_cowfp);
 674
 675	offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
 676	if (count == NULLFILEOFF)
 677		end_fsb = NULLFILEOFF;
 678	else
 679		end_fsb = XFS_B_TO_FSB(ip->i_mount, offset + count);
 680
 681	/* Start a rolling transaction to remove the mappings */
 682	error = xfs_trans_alloc(ip->i_mount, &M_RES(ip->i_mount)->tr_write,
 683			0, 0, 0, &tp);
 684	if (error)
 685		goto out;
 686
 687	xfs_ilock(ip, XFS_ILOCK_EXCL);
 688	xfs_trans_ijoin(tp, ip, 0);
 689
 690	/* Scrape out the old CoW reservations */
 691	error = xfs_reflink_cancel_cow_blocks(ip, &tp, offset_fsb, end_fsb,
 692			cancel_real);
 693	if (error)
 694		goto out_cancel;
 695
 696	error = xfs_trans_commit(tp);
 697
 698	xfs_iunlock(ip, XFS_ILOCK_EXCL);
 699	return error;
 700
 701out_cancel:
 702	xfs_trans_cancel(tp);
 703	xfs_iunlock(ip, XFS_ILOCK_EXCL);
 704out:
 705	trace_xfs_reflink_cancel_cow_range_error(ip, error, _RET_IP_);
 706	return error;
 707}
 708
 709/*
 710 * Remap part of the CoW fork into the data fork.
 711 *
 712 * We aim to remap the range starting at @offset_fsb and ending at @end_fsb
 713 * into the data fork; this function will remap what it can (at the end of the
 714 * range) and update @end_fsb appropriately.  Each remap gets its own
 715 * transaction because we can end up merging and splitting bmbt blocks for
 716 * every remap operation and we'd like to keep the block reservation
 717 * requirements as low as possible.
 718 */
 719STATIC int
 720xfs_reflink_end_cow_extent(
 721	struct xfs_inode	*ip,
 722	xfs_fileoff_t		*offset_fsb,
 723	xfs_fileoff_t		end_fsb)
 724{
 725	struct xfs_iext_cursor	icur;
 726	struct xfs_bmbt_irec	got, del, data;
 727	struct xfs_mount	*mp = ip->i_mount;
 728	struct xfs_trans	*tp;
 729	struct xfs_ifork	*ifp = xfs_ifork_ptr(ip, XFS_COW_FORK);
 730	unsigned int		resblks;
 731	int			nmaps;
 732	int			error;
 
 
 
 
 
 733
 734	/* No COW extents?  That's easy! */
 735	if (ifp->if_bytes == 0) {
 736		*offset_fsb = end_fsb;
 737		return 0;
 738	}
 739
 740	resblks = XFS_EXTENTADD_SPACE_RES(mp, XFS_DATA_FORK);
 741	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, resblks, 0,
 742			XFS_TRANS_RESERVE, &tp);
 
 
 
 
 743	if (error)
 744		return error;
 745
 746	/*
 747	 * Lock the inode.  We have to ijoin without automatic unlock because
 748	 * the lead transaction is the refcountbt record deletion; the data
 749	 * fork update follows as a deferred log item.
 750	 */
 751	xfs_ilock(ip, XFS_ILOCK_EXCL);
 752	xfs_trans_ijoin(tp, ip, 0);
 753
 754	error = xfs_iext_count_may_overflow(ip, XFS_DATA_FORK,
 755			XFS_IEXT_REFLINK_END_COW_CNT);
 756	if (error == -EFBIG)
 757		error = xfs_iext_count_upgrade(tp, ip,
 758				XFS_IEXT_REFLINK_END_COW_CNT);
 759	if (error)
 760		goto out_cancel;
 761
 762	/*
 763	 * In case of racing, overlapping AIO writes no COW extents might be
 764	 * left by the time I/O completes for the loser of the race.  In that
 765	 * case we are done.
 766	 */
 767	if (!xfs_iext_lookup_extent(ip, ifp, *offset_fsb, &icur, &got) ||
 768	    got.br_startoff >= end_fsb) {
 769		*offset_fsb = end_fsb;
 770		goto out_cancel;
 771	}
 772
 773	/*
 774	 * Only remap real extents that contain data.  With AIO, speculative
 775	 * preallocations can leak into the range we are called upon, and we
 776	 * need to skip them.  Preserve @got for the eventual CoW fork
 777	 * deletion; from now on @del represents the mapping that we're
 778	 * actually remapping.
 779	 */
 780	while (!xfs_bmap_is_written_extent(&got)) {
 781		if (!xfs_iext_next_extent(ifp, &icur, &got) ||
 782		    got.br_startoff >= end_fsb) {
 783			*offset_fsb = end_fsb;
 784			goto out_cancel;
 785		}
 786	}
 787	del = got;
 788	xfs_trim_extent(&del, *offset_fsb, end_fsb - *offset_fsb);
 789
 790	/* Grab the corresponding mapping in the data fork. */
 791	nmaps = 1;
 792	error = xfs_bmapi_read(ip, del.br_startoff, del.br_blockcount, &data,
 793			&nmaps, 0);
 794	if (error)
 795		goto out_cancel;
 796
 797	/* We can only remap the smaller of the two extent sizes. */
 798	data.br_blockcount = min(data.br_blockcount, del.br_blockcount);
 799	del.br_blockcount = data.br_blockcount;
 800
 801	trace_xfs_reflink_cow_remap_from(ip, &del);
 802	trace_xfs_reflink_cow_remap_to(ip, &data);
 803
 804	if (xfs_bmap_is_real_extent(&data)) {
 805		/*
 806		 * If the extent we're remapping is backed by storage (written
 807		 * or not), unmap the extent and drop its refcount.
 
 808		 */
 809		xfs_bmap_unmap_extent(tp, ip, XFS_DATA_FORK, &data);
 810		xfs_refcount_decrease_extent(tp, &data);
 811		xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_BCOUNT,
 812				-data.br_blockcount);
 813	} else if (data.br_startblock == DELAYSTARTBLOCK) {
 814		int		done;
 815
 816		/*
 817		 * If the extent we're remapping is a delalloc reservation,
 818		 * we can use the regular bunmapi function to release the
 819		 * incore state.  Dropping the delalloc reservation takes care
 820		 * of the quota reservation for us.
 821		 */
 822		error = xfs_bunmapi(NULL, ip, data.br_startoff,
 823				data.br_blockcount, 0, 1, &done);
 824		if (error)
 825			goto out_cancel;
 826		ASSERT(done);
 827	}
 
 
 
 
 
 828
 829	/* Free the CoW orphan record. */
 830	xfs_refcount_free_cow_extent(tp, del.br_startblock, del.br_blockcount);
 
 
 
 831
 832	/* Map the new blocks into the data fork. */
 833	xfs_bmap_map_extent(tp, ip, XFS_DATA_FORK, &del);
 
 
 834
 835	/* Charge this new data fork mapping to the on-disk quota. */
 836	xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_DELBCOUNT,
 837			(long)del.br_blockcount);
 838
 839	/* Remove the mapping from the CoW fork. */
 840	xfs_bmap_del_extent_cow(ip, &icur, &got, &del);
 
 
 
 
 
 841
 842	error = xfs_trans_commit(tp);
 843	xfs_iunlock(ip, XFS_ILOCK_EXCL);
 844	if (error)
 845		return error;
 846
 847	/* Update the caller about how much progress we made. */
 848	*offset_fsb = del.br_startoff + del.br_blockcount;
 849	return 0;
 850
 851out_cancel:
 
 852	xfs_trans_cancel(tp);
 853	xfs_iunlock(ip, XFS_ILOCK_EXCL);
 
 
 854	return error;
 855}
 856
 857/*
 858 * Remap parts of a file's data fork after a successful CoW.
 859 */
 860int
 861xfs_reflink_end_cow(
 862	struct xfs_inode		*ip,
 863	xfs_off_t			offset,
 864	xfs_off_t			count)
 865{
 866	xfs_fileoff_t			offset_fsb;
 867	xfs_fileoff_t			end_fsb;
 868	int				error = 0;
 869
 870	trace_xfs_reflink_end_cow(ip, offset, count);
 871
 872	offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
 873	end_fsb = XFS_B_TO_FSB(ip->i_mount, offset + count);
 874
 875	/*
 876	 * Walk forwards until we've remapped the I/O range.  The loop function
 877	 * repeatedly cycles the ILOCK to allocate one transaction per remapped
 878	 * extent.
 879	 *
 880	 * If we're being called by writeback then the pages will still
 881	 * have PageWriteback set, which prevents races with reflink remapping
 882	 * and truncate.  Reflink remapping prevents races with writeback by
 883	 * taking the iolock and mmaplock before flushing the pages and
 884	 * remapping, which means there won't be any further writeback or page
 885	 * cache dirtying until the reflink completes.
 886	 *
 887	 * We should never have two threads issuing writeback for the same file
 888	 * region.  There are also have post-eof checks in the writeback
 889	 * preparation code so that we don't bother writing out pages that are
 890	 * about to be truncated.
 891	 *
 892	 * If we're being called as part of directio write completion, the dio
 893	 * count is still elevated, which reflink and truncate will wait for.
 894	 * Reflink remapping takes the iolock and mmaplock and waits for
 895	 * pending dio to finish, which should prevent any directio until the
 896	 * remap completes.  Multiple concurrent directio writes to the same
 897	 * region are handled by end_cow processing only occurring for the
 898	 * threads which succeed; the outcome of multiple overlapping direct
 899	 * writes is not well defined anyway.
 900	 *
 901	 * It's possible that a buffered write and a direct write could collide
 902	 * here (the buffered write stumbles in after the dio flushes and
 903	 * invalidates the page cache and immediately queues writeback), but we
 904	 * have never supported this 100%.  If either disk write succeeds the
 905	 * blocks will be remapped.
 906	 */
 907	while (end_fsb > offset_fsb && !error)
 908		error = xfs_reflink_end_cow_extent(ip, &offset_fsb, end_fsb);
 909
 910	if (error)
 911		trace_xfs_reflink_end_cow_error(ip, error, _RET_IP_);
 912	return error;
 913}
 914
 915/*
 916 * Free all CoW staging blocks that are still referenced by the ondisk refcount
 917 * metadata.  The ondisk metadata does not track which inode created the
 918 * staging extent, so callers must ensure that there are no cached inodes with
 919 * live CoW staging extents.
 920 */
 921int
 922xfs_reflink_recover_cow(
 923	struct xfs_mount	*mp)
 924{
 925	struct xfs_perag	*pag;
 926	xfs_agnumber_t		agno;
 927	int			error = 0;
 928
 929	if (!xfs_has_reflink(mp))
 930		return 0;
 931
 932	for_each_perag(mp, agno, pag) {
 933		error = xfs_refcount_recover_cow_leftovers(mp, pag);
 934		if (error) {
 935			xfs_perag_rele(pag);
 936			break;
 937		}
 938	}
 939
 940	return error;
 941}
 942
 943/*
 944 * Reflinking (Block) Ranges of Two Files Together
 945 *
 946 * First, ensure that the reflink flag is set on both inodes.  The flag is an
 947 * optimization to avoid unnecessary refcount btree lookups in the write path.
 948 *
 949 * Now we can iteratively remap the range of extents (and holes) in src to the
 950 * corresponding ranges in dest.  Let drange and srange denote the ranges of
 951 * logical blocks in dest and src touched by the reflink operation.
 952 *
 953 * While the length of drange is greater than zero,
 954 *    - Read src's bmbt at the start of srange ("imap")
 955 *    - If imap doesn't exist, make imap appear to start at the end of srange
 956 *      with zero length.
 957 *    - If imap starts before srange, advance imap to start at srange.
 958 *    - If imap goes beyond srange, truncate imap to end at the end of srange.
 959 *    - Punch (imap start - srange start + imap len) blocks from dest at
 960 *      offset (drange start).
 961 *    - If imap points to a real range of pblks,
 962 *         > Increase the refcount of the imap's pblks
 963 *         > Map imap's pblks into dest at the offset
 964 *           (drange start + imap start - srange start)
 965 *    - Advance drange and srange by (imap start - srange start + imap len)
 966 *
 967 * Finally, if the reflink made dest longer, update both the in-core and
 968 * on-disk file sizes.
 969 *
 970 * ASCII Art Demonstration:
 971 *
 972 * Let's say we want to reflink this source file:
 973 *
 974 * ----SSSSSSS-SSSSS----SSSSSS (src file)
 975 *   <-------------------->
 976 *
 977 * into this destination file:
 978 *
 979 * --DDDDDDDDDDDDDDDDDDD--DDD (dest file)
 980 *        <-------------------->
 981 * '-' means a hole, and 'S' and 'D' are written blocks in the src and dest.
 982 * Observe that the range has different logical offsets in either file.
 983 *
 984 * Consider that the first extent in the source file doesn't line up with our
 985 * reflink range.  Unmapping  and remapping are separate operations, so we can
 986 * unmap more blocks from the destination file than we remap.
 987 *
 988 * ----SSSSSSS-SSSSS----SSSSSS
 989 *   <------->
 990 * --DDDDD---------DDDDD--DDD
 991 *        <------->
 992 *
 993 * Now remap the source extent into the destination file:
 994 *
 995 * ----SSSSSSS-SSSSS----SSSSSS
 996 *   <------->
 997 * --DDDDD--SSSSSSSDDDDD--DDD
 998 *        <------->
 999 *
1000 * Do likewise with the second hole and extent in our range.  Holes in the
1001 * unmap range don't affect our operation.
1002 *
1003 * ----SSSSSSS-SSSSS----SSSSSS
1004 *            <---->
1005 * --DDDDD--SSSSSSS-SSSSS-DDD
1006 *                 <---->
1007 *
1008 * Finally, unmap and remap part of the third extent.  This will increase the
1009 * size of the destination file.
1010 *
1011 * ----SSSSSSS-SSSSS----SSSSSS
1012 *                  <----->
1013 * --DDDDD--SSSSSSS-SSSSS----SSS
1014 *                       <----->
1015 *
1016 * Once we update the destination file's i_size, we're done.
1017 */
1018
1019/*
1020 * Ensure the reflink bit is set in both inodes.
1021 */
1022STATIC int
1023xfs_reflink_set_inode_flag(
1024	struct xfs_inode	*src,
1025	struct xfs_inode	*dest)
1026{
1027	struct xfs_mount	*mp = src->i_mount;
1028	int			error;
1029	struct xfs_trans	*tp;
1030
1031	if (xfs_is_reflink_inode(src) && xfs_is_reflink_inode(dest))
1032		return 0;
1033
1034	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
1035	if (error)
1036		goto out_error;
1037
1038	/* Lock both files against IO */
1039	if (src->i_ino == dest->i_ino)
1040		xfs_ilock(src, XFS_ILOCK_EXCL);
1041	else
1042		xfs_lock_two_inodes(src, XFS_ILOCK_EXCL, dest, XFS_ILOCK_EXCL);
1043
1044	if (!xfs_is_reflink_inode(src)) {
1045		trace_xfs_reflink_set_inode_flag(src);
1046		xfs_trans_ijoin(tp, src, XFS_ILOCK_EXCL);
1047		src->i_diflags2 |= XFS_DIFLAG2_REFLINK;
1048		xfs_trans_log_inode(tp, src, XFS_ILOG_CORE);
1049		xfs_ifork_init_cow(src);
1050	} else
1051		xfs_iunlock(src, XFS_ILOCK_EXCL);
1052
1053	if (src->i_ino == dest->i_ino)
1054		goto commit_flags;
1055
1056	if (!xfs_is_reflink_inode(dest)) {
1057		trace_xfs_reflink_set_inode_flag(dest);
1058		xfs_trans_ijoin(tp, dest, XFS_ILOCK_EXCL);
1059		dest->i_diflags2 |= XFS_DIFLAG2_REFLINK;
1060		xfs_trans_log_inode(tp, dest, XFS_ILOG_CORE);
1061		xfs_ifork_init_cow(dest);
1062	} else
1063		xfs_iunlock(dest, XFS_ILOCK_EXCL);
1064
1065commit_flags:
1066	error = xfs_trans_commit(tp);
1067	if (error)
1068		goto out_error;
1069	return error;
1070
1071out_error:
1072	trace_xfs_reflink_set_inode_flag_error(dest, error, _RET_IP_);
1073	return error;
1074}
1075
1076/*
1077 * Update destination inode size & cowextsize hint, if necessary.
1078 */
1079int
1080xfs_reflink_update_dest(
1081	struct xfs_inode	*dest,
1082	xfs_off_t		newlen,
1083	xfs_extlen_t		cowextsize,
1084	unsigned int		remap_flags)
1085{
1086	struct xfs_mount	*mp = dest->i_mount;
1087	struct xfs_trans	*tp;
1088	int			error;
1089
1090	if (newlen <= i_size_read(VFS_I(dest)) && cowextsize == 0)
1091		return 0;
1092
1093	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);
1094	if (error)
1095		goto out_error;
1096
1097	xfs_ilock(dest, XFS_ILOCK_EXCL);
1098	xfs_trans_ijoin(tp, dest, XFS_ILOCK_EXCL);
1099
1100	if (newlen > i_size_read(VFS_I(dest))) {
1101		trace_xfs_reflink_update_inode_size(dest, newlen);
1102		i_size_write(VFS_I(dest), newlen);
1103		dest->i_disk_size = newlen;
1104	}
1105
1106	if (cowextsize) {
1107		dest->i_cowextsize = cowextsize;
1108		dest->i_diflags2 |= XFS_DIFLAG2_COWEXTSIZE;
1109	}
1110
 
 
 
 
1111	xfs_trans_log_inode(tp, dest, XFS_ILOG_CORE);
1112
1113	error = xfs_trans_commit(tp);
1114	if (error)
1115		goto out_error;
1116	return error;
1117
1118out_error:
1119	trace_xfs_reflink_update_inode_size_error(dest, error, _RET_IP_);
1120	return error;
1121}
1122
1123/*
1124 * Do we have enough reserve in this AG to handle a reflink?  The refcount
1125 * btree already reserved all the space it needs, but the rmap btree can grow
1126 * infinitely, so we won't allow more reflinks when the AG is down to the
1127 * btree reserves.
1128 */
1129static int
1130xfs_reflink_ag_has_free_space(
1131	struct xfs_mount	*mp,
1132	xfs_agnumber_t		agno)
1133{
1134	struct xfs_perag	*pag;
1135	int			error = 0;
1136
1137	if (!xfs_has_rmapbt(mp))
1138		return 0;
1139
1140	pag = xfs_perag_get(mp, agno);
1141	if (xfs_ag_resv_critical(pag, XFS_AG_RESV_RMAPBT) ||
1142	    xfs_ag_resv_critical(pag, XFS_AG_RESV_METADATA))
1143		error = -ENOSPC;
1144	xfs_perag_put(pag);
1145	return error;
1146}
1147
1148/*
1149 * Remap the given extent into the file.  The dmap blockcount will be set to
1150 * the number of blocks that were actually remapped.
 
1151 */
1152STATIC int
1153xfs_reflink_remap_extent(
1154	struct xfs_inode	*ip,
1155	struct xfs_bmbt_irec	*dmap,
 
1156	xfs_off_t		new_isize)
1157{
1158	struct xfs_bmbt_irec	smap;
1159	struct xfs_mount	*mp = ip->i_mount;
1160	struct xfs_trans	*tp;
 
 
 
 
 
 
 
1161	xfs_off_t		newlen;
1162	int64_t			qdelta = 0;
1163	unsigned int		resblks;
1164	bool			quota_reserved = true;
1165	bool			smap_real;
1166	bool			dmap_written = xfs_bmap_is_written_extent(dmap);
1167	int			iext_delta = 0;
1168	int			nimaps;
1169	int			error;
1170
1171	/*
1172	 * Start a rolling transaction to switch the mappings.
1173	 *
1174	 * Adding a written extent to the extent map can cause a bmbt split,
1175	 * and removing a mapped extent from the extent can cause a bmbt split.
1176	 * The two operations cannot both cause a split since they operate on
1177	 * the same index in the bmap btree, so we only need a reservation for
1178	 * one bmbt split if either thing is happening.  However, we haven't
1179	 * locked the inode yet, so we reserve assuming this is the case.
1180	 *
1181	 * The first allocation call tries to reserve enough space to handle
1182	 * mapping dmap into a sparse part of the file plus the bmbt split.  We
1183	 * haven't locked the inode or read the existing mapping yet, so we do
1184	 * not know for sure that we need the space.  This should succeed most
1185	 * of the time.
1186	 *
1187	 * If the first attempt fails, try again but reserving only enough
1188	 * space to handle a bmbt split.  This is the hard minimum requirement,
1189	 * and we revisit quota reservations later when we know more about what
1190	 * we're remapping.
1191	 */
1192	resblks = XFS_EXTENTADD_SPACE_RES(mp, XFS_DATA_FORK);
1193	error = xfs_trans_alloc_inode(ip, &M_RES(mp)->tr_write,
1194			resblks + dmap->br_blockcount, 0, false, &tp);
1195	if (error == -EDQUOT || error == -ENOSPC) {
1196		quota_reserved = false;
1197		error = xfs_trans_alloc_inode(ip, &M_RES(mp)->tr_write,
1198				resblks, 0, false, &tp);
1199	}
1200	if (error)
1201		goto out;
1202
1203	/*
1204	 * Read what's currently mapped in the destination file into smap.
1205	 * If smap isn't a hole, we will have to remove it before we can add
1206	 * dmap to the destination file.
1207	 */
1208	nimaps = 1;
1209	error = xfs_bmapi_read(ip, dmap->br_startoff, dmap->br_blockcount,
1210			&smap, &nimaps, 0);
1211	if (error)
1212		goto out_cancel;
1213	ASSERT(nimaps == 1 && smap.br_startoff == dmap->br_startoff);
1214	smap_real = xfs_bmap_is_real_extent(&smap);
1215
1216	/*
1217	 * We can only remap as many blocks as the smaller of the two extent
1218	 * maps, because we can only remap one extent at a time.
1219	 */
1220	dmap->br_blockcount = min(dmap->br_blockcount, smap.br_blockcount);
1221	ASSERT(dmap->br_blockcount == smap.br_blockcount);
1222
1223	trace_xfs_reflink_remap_extent_dest(ip, &smap);
1224
1225	/*
1226	 * Two extents mapped to the same physical block must not have
1227	 * different states; that's filesystem corruption.  Move on to the next
1228	 * extent if they're both holes or both the same physical extent.
1229	 */
1230	if (dmap->br_startblock == smap.br_startblock) {
1231		if (dmap->br_state != smap.br_state) {
1232			xfs_bmap_mark_sick(ip, XFS_DATA_FORK);
1233			error = -EFSCORRUPTED;
1234		}
1235		goto out_cancel;
1236	}
1237
1238	/* If both extents are unwritten, leave them alone. */
1239	if (dmap->br_state == XFS_EXT_UNWRITTEN &&
1240	    smap.br_state == XFS_EXT_UNWRITTEN)
1241		goto out_cancel;
 
1242
1243	/* No reflinking if the AG of the dest mapping is low on space. */
1244	if (dmap_written) {
1245		error = xfs_reflink_ag_has_free_space(mp,
1246				XFS_FSB_TO_AGNO(mp, dmap->br_startblock));
1247		if (error)
1248			goto out_cancel;
1249	}
1250
1251	/*
1252	 * Increase quota reservation if we think the quota block counter for
1253	 * this file could increase.
1254	 *
1255	 * If we are mapping a written extent into the file, we need to have
1256	 * enough quota block count reservation to handle the blocks in that
1257	 * extent.  We log only the delta to the quota block counts, so if the
1258	 * extent we're unmapping also has blocks allocated to it, we don't
1259	 * need a quota reservation for the extent itself.
1260	 *
1261	 * Note that if we're replacing a delalloc reservation with a written
1262	 * extent, we have to take the full quota reservation because removing
1263	 * the delalloc reservation gives the block count back to the quota
1264	 * count.  This is suboptimal, but the VFS flushed the dest range
1265	 * before we started.  That should have removed all the delalloc
1266	 * reservations, but we code defensively.
1267	 *
1268	 * xfs_trans_alloc_inode above already tried to grab an even larger
1269	 * quota reservation, and kicked off a blockgc scan if it couldn't.
1270	 * If we can't get a potentially smaller quota reservation now, we're
1271	 * done.
1272	 */
1273	if (!quota_reserved && !smap_real && dmap_written) {
1274		error = xfs_trans_reserve_quota_nblks(tp, ip,
1275				dmap->br_blockcount, 0, false);
1276		if (error)
1277			goto out_cancel;
1278	}
1279
1280	if (smap_real)
1281		++iext_delta;
1282
1283	if (dmap_written)
1284		++iext_delta;
1285
1286	error = xfs_iext_count_may_overflow(ip, XFS_DATA_FORK, iext_delta);
1287	if (error == -EFBIG)
1288		error = xfs_iext_count_upgrade(tp, ip, iext_delta);
1289	if (error)
1290		goto out_cancel;
1291
1292	if (smap_real) {
1293		/*
1294		 * If the extent we're unmapping is backed by storage (written
1295		 * or not), unmap the extent and drop its refcount.
1296		 */
1297		xfs_bmap_unmap_extent(tp, ip, XFS_DATA_FORK, &smap);
1298		xfs_refcount_decrease_extent(tp, &smap);
1299		qdelta -= smap.br_blockcount;
1300	} else if (smap.br_startblock == DELAYSTARTBLOCK) {
1301		int		done;
 
 
 
 
 
 
1302
1303		/*
1304		 * If the extent we're unmapping is a delalloc reservation,
1305		 * we can use the regular bunmapi function to release the
1306		 * incore state.  Dropping the delalloc reservation takes care
1307		 * of the quota reservation for us.
1308		 */
1309		error = xfs_bunmapi(NULL, ip, smap.br_startoff,
1310				smap.br_blockcount, 0, 1, &done);
1311		if (error)
1312			goto out_cancel;
1313		ASSERT(done);
1314	}
1315
1316	/*
1317	 * If the extent we're sharing is backed by written storage, increase
1318	 * its refcount and map it into the file.
1319	 */
1320	if (dmap_written) {
1321		xfs_refcount_increase_extent(tp, dmap);
1322		xfs_bmap_map_extent(tp, ip, XFS_DATA_FORK, dmap);
1323		qdelta += dmap->br_blockcount;
1324	}
1325
1326	xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_BCOUNT, qdelta);
 
 
 
 
 
 
 
 
 
1327
1328	/* Update dest isize if needed. */
1329	newlen = XFS_FSB_TO_B(mp, dmap->br_startoff + dmap->br_blockcount);
1330	newlen = min_t(xfs_off_t, newlen, new_isize);
1331	if (newlen > i_size_read(VFS_I(ip))) {
1332		trace_xfs_reflink_update_inode_size(ip, newlen);
1333		i_size_write(VFS_I(ip), newlen);
1334		ip->i_disk_size = newlen;
1335		xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
1336	}
1337
1338	/* Commit everything and unlock. */
1339	error = xfs_trans_commit(tp);
1340	goto out_unlock;
 
 
 
1341
 
 
1342out_cancel:
1343	xfs_trans_cancel(tp);
1344out_unlock:
1345	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1346out:
1347	if (error)
1348		trace_xfs_reflink_remap_extent_error(ip, error, _RET_IP_);
1349	return error;
1350}
1351
1352/* Remap a range of one file to the other. */
1353int
 
 
1354xfs_reflink_remap_blocks(
1355	struct xfs_inode	*src,
1356	loff_t			pos_in,
1357	struct xfs_inode	*dest,
1358	loff_t			pos_out,
1359	loff_t			remap_len,
1360	loff_t			*remapped)
1361{
1362	struct xfs_bmbt_irec	imap;
1363	struct xfs_mount	*mp = src->i_mount;
1364	xfs_fileoff_t		srcoff = XFS_B_TO_FSBT(mp, pos_in);
1365	xfs_fileoff_t		destoff = XFS_B_TO_FSBT(mp, pos_out);
1366	xfs_filblks_t		len;
1367	xfs_filblks_t		remapped_len = 0;
1368	xfs_off_t		new_isize = pos_out + remap_len;
1369	int			nimaps;
1370	int			error = 0;
 
1371
1372	len = min_t(xfs_filblks_t, XFS_B_TO_FSB(mp, remap_len),
1373			XFS_MAX_FILEOFF);
1374
1375	trace_xfs_reflink_remap_blocks(src, srcoff, len, dest, destoff);
1376
1377	while (len > 0) {
1378		unsigned int	lock_mode;
1379
1380		/* Read extent from the source file */
1381		nimaps = 1;
1382		lock_mode = xfs_ilock_data_map_shared(src);
1383		error = xfs_bmapi_read(src, srcoff, len, &imap, &nimaps, 0);
1384		xfs_iunlock(src, lock_mode);
1385		if (error)
1386			break;
1387		/*
1388		 * The caller supposedly flushed all dirty pages in the source
1389		 * file range, which means that writeback should have allocated
1390		 * or deleted all delalloc reservations in that range.  If we
1391		 * find one, that's a good sign that something is seriously
1392		 * wrong here.
1393		 */
1394		ASSERT(nimaps == 1 && imap.br_startoff == srcoff);
1395		if (imap.br_startblock == DELAYSTARTBLOCK) {
1396			ASSERT(imap.br_startblock != DELAYSTARTBLOCK);
1397			xfs_bmap_mark_sick(src, XFS_DATA_FORK);
1398			error = -EFSCORRUPTED;
1399			break;
1400		}
1401
1402		trace_xfs_reflink_remap_extent_src(src, &imap);
 
1403
1404		/* Remap into the destination file at the given offset. */
1405		imap.br_startoff = destoff;
1406		error = xfs_reflink_remap_extent(dest, &imap, new_isize);
 
 
 
 
1407		if (error)
1408			break;
1409
1410		if (fatal_signal_pending(current)) {
1411			error = -EINTR;
1412			break;
1413		}
1414
1415		/* Advance drange/srange */
1416		srcoff += imap.br_blockcount;
1417		destoff += imap.br_blockcount;
1418		len -= imap.br_blockcount;
1419		remapped_len += imap.br_blockcount;
1420	}
1421
1422	if (error)
1423		trace_xfs_reflink_remap_blocks_error(dest, error, _RET_IP_);
1424	*remapped = min_t(loff_t, remap_len,
1425			  XFS_FSB_TO_B(src->i_mount, remapped_len));
1426	return error;
1427}
1428
1429/*
1430 * If we're reflinking to a point past the destination file's EOF, we must
1431 * zero any speculative post-EOF preallocations that sit between the old EOF
1432 * and the destination file offset.
1433 */
1434static int
1435xfs_reflink_zero_posteof(
1436	struct xfs_inode	*ip,
1437	loff_t			pos)
1438{
1439	loff_t			isize = i_size_read(VFS_I(ip));
1440
1441	if (pos <= isize)
1442		return 0;
1443
1444	trace_xfs_zero_eof(ip, isize, pos - isize);
1445	return xfs_zero_range(ip, isize, pos - isize, NULL);
1446}
1447
1448/*
1449 * Prepare two files for range cloning.  Upon a successful return both inodes
1450 * will have the iolock and mmaplock held, the page cache of the out file will
1451 * be truncated, and any leases on the out file will have been broken.  This
1452 * function borrows heavily from xfs_file_aio_write_checks.
1453 *
1454 * The VFS allows partial EOF blocks to "match" for dedupe even though it hasn't
1455 * checked that the bytes beyond EOF physically match. Hence we cannot use the
1456 * EOF block in the source dedupe range because it's not a complete block match,
1457 * hence can introduce a corruption into the file that has it's block replaced.
1458 *
1459 * In similar fashion, the VFS file cloning also allows partial EOF blocks to be
1460 * "block aligned" for the purposes of cloning entire files.  However, if the
1461 * source file range includes the EOF block and it lands within the existing EOF
1462 * of the destination file, then we can expose stale data from beyond the source
1463 * file EOF in the destination file.
1464 *
1465 * XFS doesn't support partial block sharing, so in both cases we have check
1466 * these cases ourselves. For dedupe, we can simply round the length to dedupe
1467 * down to the previous whole block and ignore the partial EOF block. While this
1468 * means we can't dedupe the last block of a file, this is an acceptible
1469 * tradeoff for simplicity on implementation.
1470 *
1471 * For cloning, we want to share the partial EOF block if it is also the new EOF
1472 * block of the destination file. If the partial EOF block lies inside the
1473 * existing destination EOF, then we have to abort the clone to avoid exposing
1474 * stale data in the destination file. Hence we reject these clone attempts with
1475 * -EINVAL in this case.
1476 */
1477int
1478xfs_reflink_remap_prep(
1479	struct file		*file_in,
1480	loff_t			pos_in,
1481	struct file		*file_out,
1482	loff_t			pos_out,
1483	loff_t			*len,
1484	unsigned int		remap_flags)
1485{
1486	struct inode		*inode_in = file_inode(file_in);
1487	struct xfs_inode	*src = XFS_I(inode_in);
1488	struct inode		*inode_out = file_inode(file_out);
1489	struct xfs_inode	*dest = XFS_I(inode_out);
1490	int			ret;
 
 
 
 
 
 
 
 
 
 
 
1491
1492	/* Lock both files against IO */
1493	ret = xfs_ilock2_io_mmap(src, dest);
1494	if (ret)
1495		return ret;
 
 
1496
1497	/* Check file eligibility and prepare for block sharing. */
1498	ret = -EINVAL;
1499	/* Don't reflink realtime inodes */
1500	if (XFS_IS_REALTIME_INODE(src) || XFS_IS_REALTIME_INODE(dest))
1501		goto out_unlock;
1502
1503	/* Don't share DAX file data with non-DAX file. */
1504	if (IS_DAX(inode_in) != IS_DAX(inode_out))
1505		goto out_unlock;
1506
1507	if (!IS_DAX(inode_in))
1508		ret = generic_remap_file_range_prep(file_in, pos_in, file_out,
1509				pos_out, len, remap_flags);
1510	else
1511		ret = dax_remap_file_range_prep(file_in, pos_in, file_out,
1512				pos_out, len, remap_flags, &xfs_read_iomap_ops);
1513	if (ret || *len == 0)
1514		goto out_unlock;
1515
1516	/* Attach dquots to dest inode before changing block map */
1517	ret = xfs_qm_dqattach(dest);
 
 
1518	if (ret)
1519		goto out_unlock;
1520
1521	/*
1522	 * Zero existing post-eof speculative preallocations in the destination
1523	 * file.
1524	 */
1525	ret = xfs_reflink_zero_posteof(dest, pos_out);
1526	if (ret)
1527		goto out_unlock;
1528
1529	/* Set flags and remap blocks. */
1530	ret = xfs_reflink_set_inode_flag(src, dest);
1531	if (ret)
1532		goto out_unlock;
1533
1534	/*
1535	 * If pos_out > EOF, we may have dirtied blocks between EOF and
1536	 * pos_out. In that case, we need to extend the flush and unmap to cover
1537	 * from EOF to the end of the copy length.
1538	 */
1539	if (pos_out > XFS_ISIZE(dest)) {
1540		loff_t	flen = *len + (pos_out - XFS_ISIZE(dest));
1541		ret = xfs_flush_unmap_range(dest, XFS_ISIZE(dest), flen);
1542	} else {
1543		ret = xfs_flush_unmap_range(dest, pos_out, *len);
1544	}
1545	if (ret)
1546		goto out_unlock;
1547
1548	xfs_iflags_set(src, XFS_IREMAPPING);
1549	if (inode_in != inode_out)
1550		xfs_ilock_demote(src, XFS_IOLOCK_EXCL | XFS_MMAPLOCK_EXCL);
1551
1552	return 0;
1553out_unlock:
1554	xfs_iunlock2_io_mmap(src, dest);
 
 
 
 
 
1555	return ret;
1556}
1557
1558/* Does this inode need the reflink flag? */
1559int
1560xfs_reflink_inode_has_shared_extents(
1561	struct xfs_trans		*tp,
1562	struct xfs_inode		*ip,
1563	bool				*has_shared)
 
 
 
 
 
 
 
 
1564{
1565	struct xfs_bmbt_irec		got;
1566	struct xfs_mount		*mp = ip->i_mount;
1567	struct xfs_ifork		*ifp;
1568	struct xfs_iext_cursor		icur;
1569	bool				found;
1570	int				error;
 
 
 
 
 
1571
1572	ifp = xfs_ifork_ptr(ip, XFS_DATA_FORK);
1573	error = xfs_iread_extents(tp, ip, XFS_DATA_FORK);
1574	if (error)
1575		return error;
 
 
 
 
 
 
 
 
 
 
 
1576
1577	*has_shared = false;
1578	found = xfs_iext_lookup_extent(ip, ifp, 0, &icur, &got);
1579	while (found) {
1580		struct xfs_perag	*pag;
1581		xfs_agblock_t		agbno;
1582		xfs_extlen_t		aglen;
1583		xfs_agblock_t		rbno;
1584		xfs_extlen_t		rlen;
1585
1586		if (isnullstartblock(got.br_startblock) ||
1587		    got.br_state != XFS_EXT_NORM)
1588			goto next;
 
 
 
1589
1590		pag = xfs_perag_get(mp, XFS_FSB_TO_AGNO(mp, got.br_startblock));
1591		agbno = XFS_FSB_TO_AGBNO(mp, got.br_startblock);
1592		aglen = got.br_blockcount;
1593		error = xfs_reflink_find_shared(pag, tp, agbno, aglen,
1594				&rbno, &rlen, false);
1595		xfs_perag_put(pag);
1596		if (error)
1597			return error;
 
 
 
 
1598
1599		/* Is there still a shared block here? */
1600		if (rbno != NULLAGBLOCK) {
1601			*has_shared = true;
1602			return 0;
1603		}
 
1604next:
1605		found = xfs_iext_next_extent(ifp, &icur, &got);
1606	}
1607
1608	return 0;
1609}
1610
1611/*
1612 * Clear the inode reflink flag if there are no shared extents.
1613 *
1614 * The caller is responsible for joining the inode to the transaction passed in.
1615 * The inode will be joined to the transaction that is returned to the caller.
1616 */
1617int
1618xfs_reflink_clear_inode_flag(
1619	struct xfs_inode	*ip,
1620	struct xfs_trans	**tpp)
1621{
1622	bool			needs_flag;
 
 
 
 
 
 
 
 
 
1623	int			error = 0;
1624
1625	ASSERT(xfs_is_reflink_inode(ip));
1626
1627	error = xfs_reflink_inode_has_shared_extents(*tpp, ip, &needs_flag);
1628	if (error || needs_flag)
1629		return error;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1630
1631	/*
1632	 * We didn't find any shared blocks so turn off the reflink flag.
1633	 * First, get rid of any leftover CoW mappings.
1634	 */
1635	error = xfs_reflink_cancel_cow_blocks(ip, tpp, 0, XFS_MAX_FILEOFF,
1636			true);
1637	if (error)
1638		return error;
1639
1640	/* Clear the inode flag. */
1641	trace_xfs_reflink_unset_inode_flag(ip);
1642	ip->i_diflags2 &= ~XFS_DIFLAG2_REFLINK;
1643	xfs_inode_clear_cowblocks_tag(ip);
 
1644	xfs_trans_log_inode(*tpp, ip, XFS_ILOG_CORE);
1645
1646	return error;
1647}
1648
1649/*
1650 * Clear the inode reflink flag if there are no shared extents and the size
1651 * hasn't changed.
1652 */
1653STATIC int
1654xfs_reflink_try_clear_inode_flag(
1655	struct xfs_inode	*ip)
1656{
1657	struct xfs_mount	*mp = ip->i_mount;
1658	struct xfs_trans	*tp;
1659	int			error = 0;
1660
1661	/* Start a rolling transaction to remove the mappings */
1662	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, 0, 0, 0, &tp);
1663	if (error)
1664		return error;
1665
1666	xfs_ilock(ip, XFS_ILOCK_EXCL);
1667	xfs_trans_ijoin(tp, ip, 0);
1668
1669	error = xfs_reflink_clear_inode_flag(ip, &tp);
1670	if (error)
1671		goto cancel;
1672
1673	error = xfs_trans_commit(tp);
1674	if (error)
1675		goto out;
1676
1677	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1678	return 0;
1679cancel:
1680	xfs_trans_cancel(tp);
1681out:
1682	xfs_iunlock(ip, XFS_ILOCK_EXCL);
1683	return error;
1684}
1685
1686/*
1687 * Pre-COW all shared blocks within a given byte range of a file and turn off
1688 * the reflink flag if we unshare all of the file's blocks.
1689 */
1690int
1691xfs_reflink_unshare(
1692	struct xfs_inode	*ip,
1693	xfs_off_t		offset,
1694	xfs_off_t		len)
1695{
1696	struct inode		*inode = VFS_I(ip);
 
 
 
1697	int			error;
1698
1699	if (!xfs_is_reflink_inode(ip))
1700		return 0;
1701
1702	trace_xfs_reflink_unshare(ip, offset, len);
1703
1704	inode_dio_wait(inode);
1705
1706	if (IS_DAX(inode))
1707		error = dax_file_unshare(inode, offset, len,
1708				&xfs_dax_write_iomap_ops);
1709	else
1710		error = iomap_file_unshare(inode, offset, len,
1711				&xfs_buffered_write_iomap_ops);
1712	if (error)
1713		goto out;
 
1714
1715	error = filemap_write_and_wait_range(inode->i_mapping, offset,
1716			offset + len - 1);
1717	if (error)
1718		goto out;
1719
1720	/* Turn off the reflink flag if possible. */
1721	error = xfs_reflink_try_clear_inode_flag(ip);
1722	if (error)
1723		goto out;
 
1724	return 0;
1725
 
 
1726out:
1727	trace_xfs_reflink_unshare_error(ip, error, _RET_IP_);
1728	return error;
1729}