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