brintos

brintos / linux-shallow public Read only

0
0
Text · 48.0 KiB · 5bf6682 Raw
1704 lines · c
1// SPDX-License-Identifier: GPL-2.0+2/*3 * Copyright (C) 2016 Oracle.  All Rights Reserved.4 * Author: Darrick J. Wong <darrick.wong@oracle.com>5 */6#include "xfs.h"7#include "xfs_fs.h"8#include "xfs_shared.h"9#include "xfs_format.h"10#include "xfs_log_format.h"11#include "xfs_trans_resv.h"12#include "xfs_mount.h"13#include "xfs_defer.h"14#include "xfs_inode.h"15#include "xfs_trans.h"16#include "xfs_bmap.h"17#include "xfs_bmap_util.h"18#include "xfs_trace.h"19#include "xfs_icache.h"20#include "xfs_btree.h"21#include "xfs_refcount_btree.h"22#include "xfs_refcount.h"23#include "xfs_bmap_btree.h"24#include "xfs_trans_space.h"25#include "xfs_bit.h"26#include "xfs_alloc.h"27#include "xfs_quota.h"28#include "xfs_reflink.h"29#include "xfs_iomap.h"30#include "xfs_ag.h"31#include "xfs_ag_resv.h"32#include "xfs_health.h"33 34/*35 * Copy on Write of Shared Blocks36 *37 * XFS must preserve "the usual" file semantics even when two files share38 * the same physical blocks.  This means that a write to one file must not39 * alter the blocks in a different file; the way that we'll do that is40 * through the use of a copy-on-write mechanism.  At a high level, that41 * means that when we want to write to a shared block, we allocate a new42 * block, write the data to the new block, and if that succeeds we map the43 * new block into the file.44 *45 * XFS provides a "delayed allocation" mechanism that defers the allocation46 * of disk blocks to dirty-but-not-yet-mapped file blocks as long as47 * possible.  This reduces fragmentation by enabling the filesystem to ask48 * 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 block51 * writable (write_begin or page_mkwrite).  If the offset is not mapped, we52 * create a delalloc mapping, which is a regular in-core extent, but without53 * a real startblock.  (For delalloc mappings, the startblock encodes both54 * a flag that this is a delalloc mapping, and a worst-case estimate of how55 * many blocks might be required to put the mapping into the BMBT.)  delalloc56 * 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) creates60 * outsized aligned delalloc reservations in the hope of landing out of61 * order nearby CoW writes in a single extent on disk, thereby reducing62 * 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), the68 * delalloc reservations are converted into unwritten mappings by69 * allocating blocks and replacing the delalloc mapping with real ones.70 * A delalloc mapping can be replaced by several unwritten ones if the71 * 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 the77 * write paths are similar.  The first two steps (creating the reservation78 * and allocating the blocks) are exactly the same as delalloc except that79 * the mappings must be stored in a separate CoW fork because we do not want80 * to disturb the mapping in the data fork until we're sure that the write81 * succeeded.  IO completion in this case is the process of removing the old82 * mapping from the data fork and moving the new mapping from the CoW fork to83 * 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 buffered87 * writes.88 *89 * Just prior to submitting the actual disk write requests, we convert90 * the extents representing the range of the file actually being written91 * (as opposed to extra pieces created for the cowextsize hint) to real92 * 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 sure99 * the new block has been written.  Since the new mappings are kept in a100 * separate fork, we can simply iterate these mappings to find the ones101 * that cover the file blocks that we just CoW'd.  For each extent, simply102 * unmap the corresponding range in the data fork, map the new range into103 * the data fork, and remove the extent from the CoW fork.  Because of104 * the presence of the cowextsize hint, however, we must be careful105 * only to remap the blocks that we've actually written out --  we must106 * never remap delalloc reservations nor CoW staging blocks that have107 * yet to be written.  This corresponds exactly to the real extents in108 * the CoW fork:109 *110 * D: --RRRRRRrrSRRRRRRRR---111 * C: ------UU--UUU---------112 *113 * Since the remapping operation can be applied to an arbitrary file114 * range, we record the need for the remap step as a flag in the ioend115 * instead of declaring a new IO type.  This is required for direct io116 * because we only have ioend for the whole dio, and we have to be able to117 * remember the presence of unwritten blocks and CoW blocks with a single118 * ioend structure.  Better yet, the more ground we can cover with one119 * ioend, the better.120 */121 122/*123 * Given an AG extent, find the lowest-numbered run of shared blocks124 * within that range and return the range in fbno/flen.  If125 * find_end_of_shared is true, return the longest contiguous extent of126 * shared blocks.  If there are no shared extents, fbno and flen will127 * be set to NULLAGBLOCK and 0, respectively.128 */129static int130xfs_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 the160 * shared/unshared status.  More specifically, this means that we161 * find the lowest-numbered extent of shared blocks that coincides with162 * the given block mapping.  If the shared extent overlaps the start of163 * the mapping, trim the mapping to the end of the shared extent.  If164 * the shared region intersects the mapping, trim the mapping to the165 * start of the shared extent.  If there are no shared regions that166 * overlap, just return the original extent.167 */168int169xfs_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 the209		 * mapping at the end of the shared region so that a210		 * subsequent iteration starts at the start of the211		 * 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 shared221	 * extent so that a subsequent iteration starts at the222	 * start of the shared region.223	 */224	irec->br_blockcount = fbno - agbno;225	return 0;226}227 228int229xfs_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 int246xfs_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. */284int285xfs_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 extent306 * is not shared we might have a preallocation for it in the COW fork. If so we307 * use it that rather than trigger a new allocation.308 */309static int310xfs_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 to325	 * 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 int348xfs_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 ready365	 * to initiate a disk write.  For direct I/O we are going to write the366	 * 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 int381xfs_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 433convert:434	return xfs_reflink_convert_unwritten(ip, imap, cmap, convert_now);435 436out_trans_cancel:437	xfs_trans_cancel(tp);438	return error;439}440 441static int442xfs_reflink_fill_delalloc(443	struct xfs_inode	*ip,444	struct xfs_bmbt_irec	*imap,445	struct xfs_bmbt_irec	*cmap,446	bool			*shared,447	uint			*lockmode,448	bool			convert_now)449{450	struct xfs_mount	*mp = ip->i_mount;451	struct xfs_trans	*tp;452	int			nimaps;453	int			error;454	bool			found;455 456	do {457		xfs_iunlock(ip, *lockmode);458		*lockmode = 0;459 460		error = xfs_trans_alloc_inode(ip, &M_RES(mp)->tr_write, 0, 0,461				false, &tp);462		if (error)463			return error;464 465		*lockmode = XFS_ILOCK_EXCL;466 467		error = xfs_find_trim_cow_extent(ip, imap, cmap, shared,468				&found);469		if (error || !*shared)470			goto out_trans_cancel;471 472		if (found) {473			xfs_trans_cancel(tp);474			break;475		}476 477		ASSERT(isnullstartblock(cmap->br_startblock) ||478		       cmap->br_startblock == DELAYSTARTBLOCK);479 480		/*481		 * Replace delalloc reservation with an unwritten extent.482		 */483		nimaps = 1;484		error = xfs_bmapi_write(tp, ip, cmap->br_startoff,485				cmap->br_blockcount,486				XFS_BMAPI_COWFORK | XFS_BMAPI_PREALLOC, 0,487				cmap, &nimaps);488		if (error)489			goto out_trans_cancel;490 491		xfs_inode_set_cowblocks_tag(ip);492		error = xfs_trans_commit(tp);493		if (error)494			return error;495	} while (cmap->br_startoff + cmap->br_blockcount <= imap->br_startoff);496 497	return xfs_reflink_convert_unwritten(ip, imap, cmap, convert_now);498 499out_trans_cancel:500	xfs_trans_cancel(tp);501	return error;502}503 504/* Allocate all CoW reservations covering a range of blocks in a file. */505int506xfs_reflink_allocate_cow(507	struct xfs_inode	*ip,508	struct xfs_bmbt_irec	*imap,509	struct xfs_bmbt_irec	*cmap,510	bool			*shared,511	uint			*lockmode,512	bool			convert_now)513{514	int			error;515	bool			found;516 517	xfs_assert_ilocked(ip, XFS_ILOCK_EXCL);518	if (!ip->i_cowfp) {519		ASSERT(!xfs_is_reflink_inode(ip));520		xfs_ifork_init_cow(ip);521	}522 523	error = xfs_find_trim_cow_extent(ip, imap, cmap, shared, &found);524	if (error || !*shared)525		return error;526 527	/* CoW fork has a real extent */528	if (found)529		return xfs_reflink_convert_unwritten(ip, imap, cmap,530				convert_now);531 532	/*533	 * CoW fork does not have an extent and data extent is shared.534	 * Allocate a real extent in the CoW fork.535	 */536	if (cmap->br_startoff > imap->br_startoff)537		return xfs_reflink_fill_cow_hole(ip, imap, cmap, shared,538				lockmode, convert_now);539 540	/*541	 * CoW fork has a delalloc reservation. Replace it with a real extent.542	 * There may or may not be a data fork mapping.543	 */544	if (isnullstartblock(cmap->br_startblock) ||545	    cmap->br_startblock == DELAYSTARTBLOCK)546		return xfs_reflink_fill_delalloc(ip, imap, cmap, shared,547				lockmode, convert_now);548 549	/* Shouldn't get here. */550	ASSERT(0);551	return -EFSCORRUPTED;552}553 554/*555 * Cancel CoW reservations for some block range of an inode.556 *557 * If cancel_real is true this function cancels all COW fork extents for the558 * inode; if cancel_real is false, real extents are not cleared.559 *560 * Caller must have already joined the inode to the current transaction. The561 * inode will be joined to the transaction returned to the caller.562 */563int564xfs_reflink_cancel_cow_blocks(565	struct xfs_inode		*ip,566	struct xfs_trans		**tpp,567	xfs_fileoff_t			offset_fsb,568	xfs_fileoff_t			end_fsb,569	bool				cancel_real)570{571	struct xfs_ifork		*ifp = xfs_ifork_ptr(ip, XFS_COW_FORK);572	struct xfs_bmbt_irec		got, del;573	struct xfs_iext_cursor		icur;574	int				error = 0;575 576	if (!xfs_inode_has_cow_data(ip))577		return 0;578	if (!xfs_iext_lookup_extent_before(ip, ifp, &end_fsb, &icur, &got))579		return 0;580 581	/* Walk backwards until we're out of the I/O range... */582	while (got.br_startoff + got.br_blockcount > offset_fsb) {583		del = got;584		xfs_trim_extent(&del, offset_fsb, end_fsb - offset_fsb);585 586		/* Extent delete may have bumped ext forward */587		if (!del.br_blockcount) {588			xfs_iext_prev(ifp, &icur);589			goto next_extent;590		}591 592		trace_xfs_reflink_cancel_cow(ip, &del);593 594		if (isnullstartblock(del.br_startblock)) {595			xfs_bmap_del_extent_delay(ip, XFS_COW_FORK, &icur, &got,596					&del);597		} else if (del.br_state == XFS_EXT_UNWRITTEN || cancel_real) {598			ASSERT((*tpp)->t_highest_agno == NULLAGNUMBER);599 600			/* Free the CoW orphan record. */601			xfs_refcount_free_cow_extent(*tpp, del.br_startblock,602					del.br_blockcount);603 604			error = xfs_free_extent_later(*tpp, del.br_startblock,605					del.br_blockcount, NULL,606					XFS_AG_RESV_NONE, 0);607			if (error)608				break;609 610			/* Roll the transaction */611			error = xfs_defer_finish(tpp);612			if (error)613				break;614 615			/* Remove the mapping from the CoW fork. */616			xfs_bmap_del_extent_cow(ip, &icur, &got, &del);617 618			/* Remove the quota reservation */619			xfs_quota_unreserve_blkres(ip, del.br_blockcount);620		} else {621			/* Didn't do anything, push cursor back. */622			xfs_iext_prev(ifp, &icur);623		}624next_extent:625		if (!xfs_iext_get_extent(ifp, &icur, &got))626			break;627	}628 629	/* clear tag if cow fork is emptied */630	if (!ifp->if_bytes)631		xfs_inode_clear_cowblocks_tag(ip);632	return error;633}634 635/*636 * Cancel CoW reservations for some byte range of an inode.637 *638 * If cancel_real is true this function cancels all COW fork extents for the639 * inode; if cancel_real is false, real extents are not cleared.640 */641int642xfs_reflink_cancel_cow_range(643	struct xfs_inode	*ip,644	xfs_off_t		offset,645	xfs_off_t		count,646	bool			cancel_real)647{648	struct xfs_trans	*tp;649	xfs_fileoff_t		offset_fsb;650	xfs_fileoff_t		end_fsb;651	int			error;652 653	trace_xfs_reflink_cancel_cow_range(ip, offset, count);654	ASSERT(ip->i_cowfp);655 656	offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);657	if (count == NULLFILEOFF)658		end_fsb = NULLFILEOFF;659	else660		end_fsb = XFS_B_TO_FSB(ip->i_mount, offset + count);661 662	/* Start a rolling transaction to remove the mappings */663	error = xfs_trans_alloc(ip->i_mount, &M_RES(ip->i_mount)->tr_write,664			0, 0, 0, &tp);665	if (error)666		goto out;667 668	xfs_ilock(ip, XFS_ILOCK_EXCL);669	xfs_trans_ijoin(tp, ip, 0);670 671	/* Scrape out the old CoW reservations */672	error = xfs_reflink_cancel_cow_blocks(ip, &tp, offset_fsb, end_fsb,673			cancel_real);674	if (error)675		goto out_cancel;676 677	error = xfs_trans_commit(tp);678 679	xfs_iunlock(ip, XFS_ILOCK_EXCL);680	return error;681 682out_cancel:683	xfs_trans_cancel(tp);684	xfs_iunlock(ip, XFS_ILOCK_EXCL);685out:686	trace_xfs_reflink_cancel_cow_range_error(ip, error, _RET_IP_);687	return error;688}689 690/*691 * Remap part of the CoW fork into the data fork.692 *693 * We aim to remap the range starting at @offset_fsb and ending at @end_fsb694 * into the data fork; this function will remap what it can (at the end of the695 * range) and update @end_fsb appropriately.  Each remap gets its own696 * transaction because we can end up merging and splitting bmbt blocks for697 * every remap operation and we'd like to keep the block reservation698 * requirements as low as possible.699 */700STATIC int701xfs_reflink_end_cow_extent(702	struct xfs_inode	*ip,703	xfs_fileoff_t		*offset_fsb,704	xfs_fileoff_t		end_fsb)705{706	struct xfs_iext_cursor	icur;707	struct xfs_bmbt_irec	got, del, data;708	struct xfs_mount	*mp = ip->i_mount;709	struct xfs_trans	*tp;710	struct xfs_ifork	*ifp = xfs_ifork_ptr(ip, XFS_COW_FORK);711	unsigned int		resblks;712	int			nmaps;713	int			error;714 715	resblks = XFS_EXTENTADD_SPACE_RES(mp, XFS_DATA_FORK);716	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, resblks, 0,717			XFS_TRANS_RESERVE, &tp);718	if (error)719		return error;720 721	/*722	 * Lock the inode.  We have to ijoin without automatic unlock because723	 * the lead transaction is the refcountbt record deletion; the data724	 * fork update follows as a deferred log item.725	 */726	xfs_ilock(ip, XFS_ILOCK_EXCL);727	xfs_trans_ijoin(tp, ip, 0);728 729	/*730	 * In case of racing, overlapping AIO writes no COW extents might be731	 * left by the time I/O completes for the loser of the race.  In that732	 * case we are done.733	 */734	if (!xfs_iext_lookup_extent(ip, ifp, *offset_fsb, &icur, &got) ||735	    got.br_startoff >= end_fsb) {736		*offset_fsb = end_fsb;737		goto out_cancel;738	}739 740	/*741	 * Only remap real extents that contain data.  With AIO, speculative742	 * preallocations can leak into the range we are called upon, and we743	 * need to skip them.  Preserve @got for the eventual CoW fork744	 * deletion; from now on @del represents the mapping that we're745	 * actually remapping.746	 */747	while (!xfs_bmap_is_written_extent(&got)) {748		if (!xfs_iext_next_extent(ifp, &icur, &got) ||749		    got.br_startoff >= end_fsb) {750			*offset_fsb = end_fsb;751			goto out_cancel;752		}753	}754	del = got;755	xfs_trim_extent(&del, *offset_fsb, end_fsb - *offset_fsb);756 757	error = xfs_iext_count_extend(tp, ip, XFS_DATA_FORK,758			XFS_IEXT_REFLINK_END_COW_CNT);759	if (error)760		goto out_cancel;761 762	/* Grab the corresponding mapping in the data fork. */763	nmaps = 1;764	error = xfs_bmapi_read(ip, del.br_startoff, del.br_blockcount, &data,765			&nmaps, 0);766	if (error)767		goto out_cancel;768 769	/* We can only remap the smaller of the two extent sizes. */770	data.br_blockcount = min(data.br_blockcount, del.br_blockcount);771	del.br_blockcount = data.br_blockcount;772 773	trace_xfs_reflink_cow_remap_from(ip, &del);774	trace_xfs_reflink_cow_remap_to(ip, &data);775 776	if (xfs_bmap_is_real_extent(&data)) {777		/*778		 * If the extent we're remapping is backed by storage (written779		 * or not), unmap the extent and drop its refcount.780		 */781		xfs_bmap_unmap_extent(tp, ip, XFS_DATA_FORK, &data);782		xfs_refcount_decrease_extent(tp, &data);783		xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_BCOUNT,784				-data.br_blockcount);785	} else if (data.br_startblock == DELAYSTARTBLOCK) {786		int		done;787 788		/*789		 * If the extent we're remapping is a delalloc reservation,790		 * we can use the regular bunmapi function to release the791		 * incore state.  Dropping the delalloc reservation takes care792		 * of the quota reservation for us.793		 */794		error = xfs_bunmapi(NULL, ip, data.br_startoff,795				data.br_blockcount, 0, 1, &done);796		if (error)797			goto out_cancel;798		ASSERT(done);799	}800 801	/* Free the CoW orphan record. */802	xfs_refcount_free_cow_extent(tp, del.br_startblock, del.br_blockcount);803 804	/* Map the new blocks into the data fork. */805	xfs_bmap_map_extent(tp, ip, XFS_DATA_FORK, &del);806 807	/* Charge this new data fork mapping to the on-disk quota. */808	xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_DELBCOUNT,809			(long)del.br_blockcount);810 811	/* Remove the mapping from the CoW fork. */812	xfs_bmap_del_extent_cow(ip, &icur, &got, &del);813 814	error = xfs_trans_commit(tp);815	xfs_iunlock(ip, XFS_ILOCK_EXCL);816	if (error)817		return error;818 819	/* Update the caller about how much progress we made. */820	*offset_fsb = del.br_startoff + del.br_blockcount;821	return 0;822 823out_cancel:824	xfs_trans_cancel(tp);825	xfs_iunlock(ip, XFS_ILOCK_EXCL);826	return error;827}828 829/*830 * Remap parts of a file's data fork after a successful CoW.831 */832int833xfs_reflink_end_cow(834	struct xfs_inode		*ip,835	xfs_off_t			offset,836	xfs_off_t			count)837{838	xfs_fileoff_t			offset_fsb;839	xfs_fileoff_t			end_fsb;840	int				error = 0;841 842	trace_xfs_reflink_end_cow(ip, offset, count);843 844	offset_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);845	end_fsb = XFS_B_TO_FSB(ip->i_mount, offset + count);846 847	/*848	 * Walk forwards until we've remapped the I/O range.  The loop function849	 * repeatedly cycles the ILOCK to allocate one transaction per remapped850	 * extent.851	 *852	 * If we're being called by writeback then the pages will still853	 * have PageWriteback set, which prevents races with reflink remapping854	 * and truncate.  Reflink remapping prevents races with writeback by855	 * taking the iolock and mmaplock before flushing the pages and856	 * remapping, which means there won't be any further writeback or page857	 * cache dirtying until the reflink completes.858	 *859	 * We should never have two threads issuing writeback for the same file860	 * region.  There are also have post-eof checks in the writeback861	 * preparation code so that we don't bother writing out pages that are862	 * about to be truncated.863	 *864	 * If we're being called as part of directio write completion, the dio865	 * count is still elevated, which reflink and truncate will wait for.866	 * Reflink remapping takes the iolock and mmaplock and waits for867	 * pending dio to finish, which should prevent any directio until the868	 * remap completes.  Multiple concurrent directio writes to the same869	 * region are handled by end_cow processing only occurring for the870	 * threads which succeed; the outcome of multiple overlapping direct871	 * writes is not well defined anyway.872	 *873	 * It's possible that a buffered write and a direct write could collide874	 * here (the buffered write stumbles in after the dio flushes and875	 * invalidates the page cache and immediately queues writeback), but we876	 * have never supported this 100%.  If either disk write succeeds the877	 * blocks will be remapped.878	 */879	while (end_fsb > offset_fsb && !error)880		error = xfs_reflink_end_cow_extent(ip, &offset_fsb, end_fsb);881 882	if (error)883		trace_xfs_reflink_end_cow_error(ip, error, _RET_IP_);884	return error;885}886 887/*888 * Free all CoW staging blocks that are still referenced by the ondisk refcount889 * metadata.  The ondisk metadata does not track which inode created the890 * staging extent, so callers must ensure that there are no cached inodes with891 * live CoW staging extents.892 */893int894xfs_reflink_recover_cow(895	struct xfs_mount	*mp)896{897	struct xfs_perag	*pag;898	xfs_agnumber_t		agno;899	int			error = 0;900 901	if (!xfs_has_reflink(mp))902		return 0;903 904	for_each_perag(mp, agno, pag) {905		error = xfs_refcount_recover_cow_leftovers(mp, pag);906		if (error) {907			xfs_perag_rele(pag);908			break;909		}910	}911 912	return error;913}914 915/*916 * Reflinking (Block) Ranges of Two Files Together917 *918 * First, ensure that the reflink flag is set on both inodes.  The flag is an919 * optimization to avoid unnecessary refcount btree lookups in the write path.920 *921 * Now we can iteratively remap the range of extents (and holes) in src to the922 * corresponding ranges in dest.  Let drange and srange denote the ranges of923 * logical blocks in dest and src touched by the reflink operation.924 *925 * While the length of drange is greater than zero,926 *    - Read src's bmbt at the start of srange ("imap")927 *    - If imap doesn't exist, make imap appear to start at the end of srange928 *      with zero length.929 *    - If imap starts before srange, advance imap to start at srange.930 *    - If imap goes beyond srange, truncate imap to end at the end of srange.931 *    - Punch (imap start - srange start + imap len) blocks from dest at932 *      offset (drange start).933 *    - If imap points to a real range of pblks,934 *         > Increase the refcount of the imap's pblks935 *         > Map imap's pblks into dest at the offset936 *           (drange start + imap start - srange start)937 *    - Advance drange and srange by (imap start - srange start + imap len)938 *939 * Finally, if the reflink made dest longer, update both the in-core and940 * on-disk file sizes.941 *942 * ASCII Art Demonstration:943 *944 * Let's say we want to reflink this source file:945 *946 * ----SSSSSSS-SSSSS----SSSSSS (src file)947 *   <-------------------->948 *949 * into this destination file:950 *951 * --DDDDDDDDDDDDDDDDDDD--DDD (dest file)952 *        <-------------------->953 * '-' means a hole, and 'S' and 'D' are written blocks in the src and dest.954 * Observe that the range has different logical offsets in either file.955 *956 * Consider that the first extent in the source file doesn't line up with our957 * reflink range.  Unmapping  and remapping are separate operations, so we can958 * unmap more blocks from the destination file than we remap.959 *960 * ----SSSSSSS-SSSSS----SSSSSS961 *   <------->962 * --DDDDD---------DDDDD--DDD963 *        <------->964 *965 * Now remap the source extent into the destination file:966 *967 * ----SSSSSSS-SSSSS----SSSSSS968 *   <------->969 * --DDDDD--SSSSSSSDDDDD--DDD970 *        <------->971 *972 * Do likewise with the second hole and extent in our range.  Holes in the973 * unmap range don't affect our operation.974 *975 * ----SSSSSSS-SSSSS----SSSSSS976 *            <---->977 * --DDDDD--SSSSSSS-SSSSS-DDD978 *                 <---->979 *980 * Finally, unmap and remap part of the third extent.  This will increase the981 * size of the destination file.982 *983 * ----SSSSSSS-SSSSS----SSSSSS984 *                  <----->985 * --DDDDD--SSSSSSS-SSSSS----SSS986 *                       <----->987 *988 * Once we update the destination file's i_size, we're done.989 */990 991/*992 * Ensure the reflink bit is set in both inodes.993 */994STATIC int995xfs_reflink_set_inode_flag(996	struct xfs_inode	*src,997	struct xfs_inode	*dest)998{999	struct xfs_mount	*mp = src->i_mount;1000	int			error;1001	struct xfs_trans	*tp;1002 1003	if (xfs_is_reflink_inode(src) && xfs_is_reflink_inode(dest))1004		return 0;1005 1006	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);1007	if (error)1008		goto out_error;1009 1010	/* Lock both files against IO */1011	if (src->i_ino == dest->i_ino)1012		xfs_ilock(src, XFS_ILOCK_EXCL);1013	else1014		xfs_lock_two_inodes(src, XFS_ILOCK_EXCL, dest, XFS_ILOCK_EXCL);1015 1016	if (!xfs_is_reflink_inode(src)) {1017		trace_xfs_reflink_set_inode_flag(src);1018		xfs_trans_ijoin(tp, src, XFS_ILOCK_EXCL);1019		src->i_diflags2 |= XFS_DIFLAG2_REFLINK;1020		xfs_trans_log_inode(tp, src, XFS_ILOG_CORE);1021		xfs_ifork_init_cow(src);1022	} else1023		xfs_iunlock(src, XFS_ILOCK_EXCL);1024 1025	if (src->i_ino == dest->i_ino)1026		goto commit_flags;1027 1028	if (!xfs_is_reflink_inode(dest)) {1029		trace_xfs_reflink_set_inode_flag(dest);1030		xfs_trans_ijoin(tp, dest, XFS_ILOCK_EXCL);1031		dest->i_diflags2 |= XFS_DIFLAG2_REFLINK;1032		xfs_trans_log_inode(tp, dest, XFS_ILOG_CORE);1033		xfs_ifork_init_cow(dest);1034	} else1035		xfs_iunlock(dest, XFS_ILOCK_EXCL);1036 1037commit_flags:1038	error = xfs_trans_commit(tp);1039	if (error)1040		goto out_error;1041	return error;1042 1043out_error:1044	trace_xfs_reflink_set_inode_flag_error(dest, error, _RET_IP_);1045	return error;1046}1047 1048/*1049 * Update destination inode size & cowextsize hint, if necessary.1050 */1051int1052xfs_reflink_update_dest(1053	struct xfs_inode	*dest,1054	xfs_off_t		newlen,1055	xfs_extlen_t		cowextsize,1056	unsigned int		remap_flags)1057{1058	struct xfs_mount	*mp = dest->i_mount;1059	struct xfs_trans	*tp;1060	int			error;1061 1062	if (newlen <= i_size_read(VFS_I(dest)) && cowextsize == 0)1063		return 0;1064 1065	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_ichange, 0, 0, 0, &tp);1066	if (error)1067		goto out_error;1068 1069	xfs_ilock(dest, XFS_ILOCK_EXCL);1070	xfs_trans_ijoin(tp, dest, XFS_ILOCK_EXCL);1071 1072	if (newlen > i_size_read(VFS_I(dest))) {1073		trace_xfs_reflink_update_inode_size(dest, newlen);1074		i_size_write(VFS_I(dest), newlen);1075		dest->i_disk_size = newlen;1076	}1077 1078	if (cowextsize) {1079		dest->i_cowextsize = cowextsize;1080		dest->i_diflags2 |= XFS_DIFLAG2_COWEXTSIZE;1081	}1082 1083	xfs_trans_log_inode(tp, dest, XFS_ILOG_CORE);1084 1085	error = xfs_trans_commit(tp);1086	if (error)1087		goto out_error;1088	return error;1089 1090out_error:1091	trace_xfs_reflink_update_inode_size_error(dest, error, _RET_IP_);1092	return error;1093}1094 1095/*1096 * Do we have enough reserve in this AG to handle a reflink?  The refcount1097 * btree already reserved all the space it needs, but the rmap btree can grow1098 * infinitely, so we won't allow more reflinks when the AG is down to the1099 * btree reserves.1100 */1101static int1102xfs_reflink_ag_has_free_space(1103	struct xfs_mount	*mp,1104	xfs_agnumber_t		agno)1105{1106	struct xfs_perag	*pag;1107	int			error = 0;1108 1109	if (!xfs_has_rmapbt(mp))1110		return 0;1111 1112	pag = xfs_perag_get(mp, agno);1113	if (xfs_ag_resv_critical(pag, XFS_AG_RESV_RMAPBT) ||1114	    xfs_ag_resv_critical(pag, XFS_AG_RESV_METADATA))1115		error = -ENOSPC;1116	xfs_perag_put(pag);1117	return error;1118}1119 1120/*1121 * Remap the given extent into the file.  The dmap blockcount will be set to1122 * the number of blocks that were actually remapped.1123 */1124STATIC int1125xfs_reflink_remap_extent(1126	struct xfs_inode	*ip,1127	struct xfs_bmbt_irec	*dmap,1128	xfs_off_t		new_isize)1129{1130	struct xfs_bmbt_irec	smap;1131	struct xfs_mount	*mp = ip->i_mount;1132	struct xfs_trans	*tp;1133	xfs_off_t		newlen;1134	int64_t			qdelta = 0;1135	unsigned int		resblks;1136	bool			quota_reserved = true;1137	bool			smap_real;1138	bool			dmap_written = xfs_bmap_is_written_extent(dmap);1139	int			iext_delta = 0;1140	int			nimaps;1141	int			error;1142 1143	/*1144	 * Start a rolling transaction to switch the mappings.1145	 *1146	 * Adding a written extent to the extent map can cause a bmbt split,1147	 * and removing a mapped extent from the extent can cause a bmbt split.1148	 * The two operations cannot both cause a split since they operate on1149	 * the same index in the bmap btree, so we only need a reservation for1150	 * one bmbt split if either thing is happening.  However, we haven't1151	 * locked the inode yet, so we reserve assuming this is the case.1152	 *1153	 * The first allocation call tries to reserve enough space to handle1154	 * mapping dmap into a sparse part of the file plus the bmbt split.  We1155	 * haven't locked the inode or read the existing mapping yet, so we do1156	 * not know for sure that we need the space.  This should succeed most1157	 * of the time.1158	 *1159	 * If the first attempt fails, try again but reserving only enough1160	 * space to handle a bmbt split.  This is the hard minimum requirement,1161	 * and we revisit quota reservations later when we know more about what1162	 * we're remapping.1163	 */1164	resblks = XFS_EXTENTADD_SPACE_RES(mp, XFS_DATA_FORK);1165	error = xfs_trans_alloc_inode(ip, &M_RES(mp)->tr_write,1166			resblks + dmap->br_blockcount, 0, false, &tp);1167	if (error == -EDQUOT || error == -ENOSPC) {1168		quota_reserved = false;1169		error = xfs_trans_alloc_inode(ip, &M_RES(mp)->tr_write,1170				resblks, 0, false, &tp);1171	}1172	if (error)1173		goto out;1174 1175	/*1176	 * Read what's currently mapped in the destination file into smap.1177	 * If smap isn't a hole, we will have to remove it before we can add1178	 * dmap to the destination file.1179	 */1180	nimaps = 1;1181	error = xfs_bmapi_read(ip, dmap->br_startoff, dmap->br_blockcount,1182			&smap, &nimaps, 0);1183	if (error)1184		goto out_cancel;1185	ASSERT(nimaps == 1 && smap.br_startoff == dmap->br_startoff);1186	smap_real = xfs_bmap_is_real_extent(&smap);1187 1188	/*1189	 * We can only remap as many blocks as the smaller of the two extent1190	 * maps, because we can only remap one extent at a time.1191	 */1192	dmap->br_blockcount = min(dmap->br_blockcount, smap.br_blockcount);1193	ASSERT(dmap->br_blockcount == smap.br_blockcount);1194 1195	trace_xfs_reflink_remap_extent_dest(ip, &smap);1196 1197	/*1198	 * Two extents mapped to the same physical block must not have1199	 * different states; that's filesystem corruption.  Move on to the next1200	 * extent if they're both holes or both the same physical extent.1201	 */1202	if (dmap->br_startblock == smap.br_startblock) {1203		if (dmap->br_state != smap.br_state) {1204			xfs_bmap_mark_sick(ip, XFS_DATA_FORK);1205			error = -EFSCORRUPTED;1206		}1207		goto out_cancel;1208	}1209 1210	/* If both extents are unwritten, leave them alone. */1211	if (dmap->br_state == XFS_EXT_UNWRITTEN &&1212	    smap.br_state == XFS_EXT_UNWRITTEN)1213		goto out_cancel;1214 1215	/* No reflinking if the AG of the dest mapping is low on space. */1216	if (dmap_written) {1217		error = xfs_reflink_ag_has_free_space(mp,1218				XFS_FSB_TO_AGNO(mp, dmap->br_startblock));1219		if (error)1220			goto out_cancel;1221	}1222 1223	/*1224	 * Increase quota reservation if we think the quota block counter for1225	 * this file could increase.1226	 *1227	 * If we are mapping a written extent into the file, we need to have1228	 * enough quota block count reservation to handle the blocks in that1229	 * extent.  We log only the delta to the quota block counts, so if the1230	 * extent we're unmapping also has blocks allocated to it, we don't1231	 * need a quota reservation for the extent itself.1232	 *1233	 * Note that if we're replacing a delalloc reservation with a written1234	 * extent, we have to take the full quota reservation because removing1235	 * the delalloc reservation gives the block count back to the quota1236	 * count.  This is suboptimal, but the VFS flushed the dest range1237	 * before we started.  That should have removed all the delalloc1238	 * reservations, but we code defensively.1239	 *1240	 * xfs_trans_alloc_inode above already tried to grab an even larger1241	 * quota reservation, and kicked off a blockgc scan if it couldn't.1242	 * If we can't get a potentially smaller quota reservation now, we're1243	 * done.1244	 */1245	if (!quota_reserved && !smap_real && dmap_written) {1246		error = xfs_trans_reserve_quota_nblks(tp, ip,1247				dmap->br_blockcount, 0, false);1248		if (error)1249			goto out_cancel;1250	}1251 1252	if (smap_real)1253		++iext_delta;1254 1255	if (dmap_written)1256		++iext_delta;1257 1258	error = xfs_iext_count_extend(tp, ip, XFS_DATA_FORK, iext_delta);1259	if (error)1260		goto out_cancel;1261 1262	if (smap_real) {1263		/*1264		 * If the extent we're unmapping is backed by storage (written1265		 * or not), unmap the extent and drop its refcount.1266		 */1267		xfs_bmap_unmap_extent(tp, ip, XFS_DATA_FORK, &smap);1268		xfs_refcount_decrease_extent(tp, &smap);1269		qdelta -= smap.br_blockcount;1270	} else if (smap.br_startblock == DELAYSTARTBLOCK) {1271		int		done;1272 1273		/*1274		 * If the extent we're unmapping is a delalloc reservation,1275		 * we can use the regular bunmapi function to release the1276		 * incore state.  Dropping the delalloc reservation takes care1277		 * of the quota reservation for us.1278		 */1279		error = xfs_bunmapi(NULL, ip, smap.br_startoff,1280				smap.br_blockcount, 0, 1, &done);1281		if (error)1282			goto out_cancel;1283		ASSERT(done);1284	}1285 1286	/*1287	 * If the extent we're sharing is backed by written storage, increase1288	 * its refcount and map it into the file.1289	 */1290	if (dmap_written) {1291		xfs_refcount_increase_extent(tp, dmap);1292		xfs_bmap_map_extent(tp, ip, XFS_DATA_FORK, dmap);1293		qdelta += dmap->br_blockcount;1294	}1295 1296	xfs_trans_mod_dquot_byino(tp, ip, XFS_TRANS_DQ_BCOUNT, qdelta);1297 1298	/* Update dest isize if needed. */1299	newlen = XFS_FSB_TO_B(mp, dmap->br_startoff + dmap->br_blockcount);1300	newlen = min_t(xfs_off_t, newlen, new_isize);1301	if (newlen > i_size_read(VFS_I(ip))) {1302		trace_xfs_reflink_update_inode_size(ip, newlen);1303		i_size_write(VFS_I(ip), newlen);1304		ip->i_disk_size = newlen;1305		xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);1306	}1307 1308	/* Commit everything and unlock. */1309	error = xfs_trans_commit(tp);1310	goto out_unlock;1311 1312out_cancel:1313	xfs_trans_cancel(tp);1314out_unlock:1315	xfs_iunlock(ip, XFS_ILOCK_EXCL);1316out:1317	if (error)1318		trace_xfs_reflink_remap_extent_error(ip, error, _RET_IP_);1319	return error;1320}1321 1322/* Remap a range of one file to the other. */1323int1324xfs_reflink_remap_blocks(1325	struct xfs_inode	*src,1326	loff_t			pos_in,1327	struct xfs_inode	*dest,1328	loff_t			pos_out,1329	loff_t			remap_len,1330	loff_t			*remapped)1331{1332	struct xfs_bmbt_irec	imap;1333	struct xfs_mount	*mp = src->i_mount;1334	xfs_fileoff_t		srcoff = XFS_B_TO_FSBT(mp, pos_in);1335	xfs_fileoff_t		destoff = XFS_B_TO_FSBT(mp, pos_out);1336	xfs_filblks_t		len;1337	xfs_filblks_t		remapped_len = 0;1338	xfs_off_t		new_isize = pos_out + remap_len;1339	int			nimaps;1340	int			error = 0;1341 1342	len = min_t(xfs_filblks_t, XFS_B_TO_FSB(mp, remap_len),1343			XFS_MAX_FILEOFF);1344 1345	trace_xfs_reflink_remap_blocks(src, srcoff, len, dest, destoff);1346 1347	while (len > 0) {1348		unsigned int	lock_mode;1349 1350		/* Read extent from the source file */1351		nimaps = 1;1352		lock_mode = xfs_ilock_data_map_shared(src);1353		error = xfs_bmapi_read(src, srcoff, len, &imap, &nimaps, 0);1354		xfs_iunlock(src, lock_mode);1355		if (error)1356			break;1357		/*1358		 * The caller supposedly flushed all dirty pages in the source1359		 * file range, which means that writeback should have allocated1360		 * or deleted all delalloc reservations in that range.  If we1361		 * find one, that's a good sign that something is seriously1362		 * wrong here.1363		 */1364		ASSERT(nimaps == 1 && imap.br_startoff == srcoff);1365		if (imap.br_startblock == DELAYSTARTBLOCK) {1366			ASSERT(imap.br_startblock != DELAYSTARTBLOCK);1367			xfs_bmap_mark_sick(src, XFS_DATA_FORK);1368			error = -EFSCORRUPTED;1369			break;1370		}1371 1372		trace_xfs_reflink_remap_extent_src(src, &imap);1373 1374		/* Remap into the destination file at the given offset. */1375		imap.br_startoff = destoff;1376		error = xfs_reflink_remap_extent(dest, &imap, new_isize);1377		if (error)1378			break;1379 1380		if (fatal_signal_pending(current)) {1381			error = -EINTR;1382			break;1383		}1384 1385		/* Advance drange/srange */1386		srcoff += imap.br_blockcount;1387		destoff += imap.br_blockcount;1388		len -= imap.br_blockcount;1389		remapped_len += imap.br_blockcount;1390		cond_resched();1391	}1392 1393	if (error)1394		trace_xfs_reflink_remap_blocks_error(dest, error, _RET_IP_);1395	*remapped = min_t(loff_t, remap_len,1396			  XFS_FSB_TO_B(src->i_mount, remapped_len));1397	return error;1398}1399 1400/*1401 * If we're reflinking to a point past the destination file's EOF, we must1402 * zero any speculative post-EOF preallocations that sit between the old EOF1403 * and the destination file offset.1404 */1405static int1406xfs_reflink_zero_posteof(1407	struct xfs_inode	*ip,1408	loff_t			pos)1409{1410	loff_t			isize = i_size_read(VFS_I(ip));1411 1412	if (pos <= isize)1413		return 0;1414 1415	trace_xfs_zero_eof(ip, isize, pos - isize);1416	return xfs_zero_range(ip, isize, pos - isize, NULL);1417}1418 1419/*1420 * Prepare two files for range cloning.  Upon a successful return both inodes1421 * will have the iolock and mmaplock held, the page cache of the out file will1422 * be truncated, and any leases on the out file will have been broken.  This1423 * function borrows heavily from xfs_file_aio_write_checks.1424 *1425 * The VFS allows partial EOF blocks to "match" for dedupe even though it hasn't1426 * checked that the bytes beyond EOF physically match. Hence we cannot use the1427 * EOF block in the source dedupe range because it's not a complete block match,1428 * hence can introduce a corruption into the file that has it's block replaced.1429 *1430 * In similar fashion, the VFS file cloning also allows partial EOF blocks to be1431 * "block aligned" for the purposes of cloning entire files.  However, if the1432 * source file range includes the EOF block and it lands within the existing EOF1433 * of the destination file, then we can expose stale data from beyond the source1434 * file EOF in the destination file.1435 *1436 * XFS doesn't support partial block sharing, so in both cases we have check1437 * these cases ourselves. For dedupe, we can simply round the length to dedupe1438 * down to the previous whole block and ignore the partial EOF block. While this1439 * means we can't dedupe the last block of a file, this is an acceptible1440 * tradeoff for simplicity on implementation.1441 *1442 * For cloning, we want to share the partial EOF block if it is also the new EOF1443 * block of the destination file. If the partial EOF block lies inside the1444 * existing destination EOF, then we have to abort the clone to avoid exposing1445 * stale data in the destination file. Hence we reject these clone attempts with1446 * -EINVAL in this case.1447 */1448int1449xfs_reflink_remap_prep(1450	struct file		*file_in,1451	loff_t			pos_in,1452	struct file		*file_out,1453	loff_t			pos_out,1454	loff_t			*len,1455	unsigned int		remap_flags)1456{1457	struct inode		*inode_in = file_inode(file_in);1458	struct xfs_inode	*src = XFS_I(inode_in);1459	struct inode		*inode_out = file_inode(file_out);1460	struct xfs_inode	*dest = XFS_I(inode_out);1461	int			ret;1462 1463	/* Lock both files against IO */1464	ret = xfs_ilock2_io_mmap(src, dest);1465	if (ret)1466		return ret;1467 1468	/* Check file eligibility and prepare for block sharing. */1469	ret = -EINVAL;1470	/* Don't reflink realtime inodes */1471	if (XFS_IS_REALTIME_INODE(src) || XFS_IS_REALTIME_INODE(dest))1472		goto out_unlock;1473 1474	/* Don't share DAX file data with non-DAX file. */1475	if (IS_DAX(inode_in) != IS_DAX(inode_out))1476		goto out_unlock;1477 1478	if (!IS_DAX(inode_in))1479		ret = generic_remap_file_range_prep(file_in, pos_in, file_out,1480				pos_out, len, remap_flags);1481	else1482		ret = dax_remap_file_range_prep(file_in, pos_in, file_out,1483				pos_out, len, remap_flags, &xfs_read_iomap_ops);1484	if (ret || *len == 0)1485		goto out_unlock;1486 1487	/* Attach dquots to dest inode before changing block map */1488	ret = xfs_qm_dqattach(dest);1489	if (ret)1490		goto out_unlock;1491 1492	/*1493	 * Zero existing post-eof speculative preallocations in the destination1494	 * file.1495	 */1496	ret = xfs_reflink_zero_posteof(dest, pos_out);1497	if (ret)1498		goto out_unlock;1499 1500	/* Set flags and remap blocks. */1501	ret = xfs_reflink_set_inode_flag(src, dest);1502	if (ret)1503		goto out_unlock;1504 1505	/*1506	 * If pos_out > EOF, we may have dirtied blocks between EOF and1507	 * pos_out. In that case, we need to extend the flush and unmap to cover1508	 * from EOF to the end of the copy length.1509	 */1510	if (pos_out > XFS_ISIZE(dest)) {1511		loff_t	flen = *len + (pos_out - XFS_ISIZE(dest));1512		ret = xfs_flush_unmap_range(dest, XFS_ISIZE(dest), flen);1513	} else {1514		ret = xfs_flush_unmap_range(dest, pos_out, *len);1515	}1516	if (ret)1517		goto out_unlock;1518 1519	xfs_iflags_set(src, XFS_IREMAPPING);1520	if (inode_in != inode_out)1521		xfs_ilock_demote(src, XFS_IOLOCK_EXCL | XFS_MMAPLOCK_EXCL);1522 1523	return 0;1524out_unlock:1525	xfs_iunlock2_io_mmap(src, dest);1526	return ret;1527}1528 1529/* Does this inode need the reflink flag? */1530int1531xfs_reflink_inode_has_shared_extents(1532	struct xfs_trans		*tp,1533	struct xfs_inode		*ip,1534	bool				*has_shared)1535{1536	struct xfs_bmbt_irec		got;1537	struct xfs_mount		*mp = ip->i_mount;1538	struct xfs_ifork		*ifp;1539	struct xfs_iext_cursor		icur;1540	bool				found;1541	int				error;1542 1543	ifp = xfs_ifork_ptr(ip, XFS_DATA_FORK);1544	error = xfs_iread_extents(tp, ip, XFS_DATA_FORK);1545	if (error)1546		return error;1547 1548	*has_shared = false;1549	found = xfs_iext_lookup_extent(ip, ifp, 0, &icur, &got);1550	while (found) {1551		struct xfs_perag	*pag;1552		xfs_agblock_t		agbno;1553		xfs_extlen_t		aglen;1554		xfs_agblock_t		rbno;1555		xfs_extlen_t		rlen;1556 1557		if (isnullstartblock(got.br_startblock) ||1558		    got.br_state != XFS_EXT_NORM)1559			goto next;1560 1561		pag = xfs_perag_get(mp, XFS_FSB_TO_AGNO(mp, got.br_startblock));1562		agbno = XFS_FSB_TO_AGBNO(mp, got.br_startblock);1563		aglen = got.br_blockcount;1564		error = xfs_reflink_find_shared(pag, tp, agbno, aglen,1565				&rbno, &rlen, false);1566		xfs_perag_put(pag);1567		if (error)1568			return error;1569 1570		/* Is there still a shared block here? */1571		if (rbno != NULLAGBLOCK) {1572			*has_shared = true;1573			return 0;1574		}1575next:1576		found = xfs_iext_next_extent(ifp, &icur, &got);1577	}1578 1579	return 0;1580}1581 1582/*1583 * Clear the inode reflink flag if there are no shared extents.1584 *1585 * The caller is responsible for joining the inode to the transaction passed in.1586 * The inode will be joined to the transaction that is returned to the caller.1587 */1588int1589xfs_reflink_clear_inode_flag(1590	struct xfs_inode	*ip,1591	struct xfs_trans	**tpp)1592{1593	bool			needs_flag;1594	int			error = 0;1595 1596	ASSERT(xfs_is_reflink_inode(ip));1597 1598	if (!xfs_can_free_cowblocks(ip))1599		return 0;1600 1601	error = xfs_reflink_inode_has_shared_extents(*tpp, ip, &needs_flag);1602	if (error || needs_flag)1603		return error;1604 1605	/*1606	 * We didn't find any shared blocks so turn off the reflink flag.1607	 * First, get rid of any leftover CoW mappings.1608	 */1609	error = xfs_reflink_cancel_cow_blocks(ip, tpp, 0, XFS_MAX_FILEOFF,1610			true);1611	if (error)1612		return error;1613 1614	/* Clear the inode flag. */1615	trace_xfs_reflink_unset_inode_flag(ip);1616	ip->i_diflags2 &= ~XFS_DIFLAG2_REFLINK;1617	xfs_inode_clear_cowblocks_tag(ip);1618	xfs_trans_log_inode(*tpp, ip, XFS_ILOG_CORE);1619 1620	return error;1621}1622 1623/*1624 * Clear the inode reflink flag if there are no shared extents and the size1625 * hasn't changed.1626 */1627STATIC int1628xfs_reflink_try_clear_inode_flag(1629	struct xfs_inode	*ip)1630{1631	struct xfs_mount	*mp = ip->i_mount;1632	struct xfs_trans	*tp;1633	int			error = 0;1634 1635	/* Start a rolling transaction to remove the mappings */1636	error = xfs_trans_alloc(mp, &M_RES(mp)->tr_write, 0, 0, 0, &tp);1637	if (error)1638		return error;1639 1640	xfs_ilock(ip, XFS_ILOCK_EXCL);1641	xfs_trans_ijoin(tp, ip, 0);1642 1643	error = xfs_reflink_clear_inode_flag(ip, &tp);1644	if (error)1645		goto cancel;1646 1647	error = xfs_trans_commit(tp);1648	if (error)1649		goto out;1650 1651	xfs_iunlock(ip, XFS_ILOCK_EXCL);1652	return 0;1653cancel:1654	xfs_trans_cancel(tp);1655out:1656	xfs_iunlock(ip, XFS_ILOCK_EXCL);1657	return error;1658}1659 1660/*1661 * Pre-COW all shared blocks within a given byte range of a file and turn off1662 * the reflink flag if we unshare all of the file's blocks.1663 */1664int1665xfs_reflink_unshare(1666	struct xfs_inode	*ip,1667	xfs_off_t		offset,1668	xfs_off_t		len)1669{1670	struct inode		*inode = VFS_I(ip);1671	int			error;1672 1673	if (!xfs_is_reflink_inode(ip))1674		return 0;1675 1676	trace_xfs_reflink_unshare(ip, offset, len);1677 1678	inode_dio_wait(inode);1679 1680	if (IS_DAX(inode))1681		error = dax_file_unshare(inode, offset, len,1682				&xfs_dax_write_iomap_ops);1683	else1684		error = iomap_file_unshare(inode, offset, len,1685				&xfs_buffered_write_iomap_ops);1686	if (error)1687		goto out;1688 1689	error = filemap_write_and_wait_range(inode->i_mapping, offset,1690			offset + len - 1);1691	if (error)1692		goto out;1693 1694	/* Turn off the reflink flag if possible. */1695	error = xfs_reflink_try_clear_inode_flag(ip);1696	if (error)1697		goto out;1698	return 0;1699 1700out:1701	trace_xfs_reflink_unshare_error(ip, error, _RET_IP_);1702	return error;1703}1704