[PATCH 30/58] xfs: implement XFS_IOC_SWAPEXT when rmap btree is enabled

From: "Darrick J. Wong" <darrick.wong@oracle.com>
To: david@fromorbit.com, darrick.wong@oracle.com
Cc: linux-fsdevel@vger.kernel.org, xfs@oss.sgi.com
Subject: [PATCH 30/58] xfs: implement XFS_IOC_SWAPEXT when rmap btree is enabled
Date: Tue, 06 Oct 2015 21:58:25 -0700	[thread overview]
Message-ID: <20151007045825.30457.37098.stgit@birch.djwong.org> (raw)
In-Reply-To: <20151007045443.30457.47038.stgit@birch.djwong.org>

Implement extent swapping when reverse-mapping is enabled.

Signed-off-by: Darrick J. Wong <darrick.wong@oracle.com>
---
 fs/xfs/libxfs/xfs_btree.c      |   17 ++++++
 fs/xfs/libxfs/xfs_rmap.c       |  112 ++++++++++++++++++++++++++++++++++++++++
 fs/xfs/libxfs/xfs_rmap_btree.h |    8 +++
 fs/xfs/xfs_bmap_util.c         |   38 +++++++++-----
 4 files changed, 161 insertions(+), 14 deletions(-)

diff --git a/fs/xfs/libxfs/xfs_btree.c b/fs/xfs/libxfs/xfs_btree.c
index 13971c6..2d7dc8c 100644
--- a/fs/xfs/libxfs/xfs_btree.c
+++ b/fs/xfs/libxfs/xfs_btree.c
@@ -32,6 +32,7 @@
 #include "xfs_trace.h"
 #include "xfs_cksum.h"
 #include "xfs_alloc.h"
+#include "xfs_rmap_btree.h"
 
 /*
  * Cursor allocation zone.
@@ -3988,6 +3989,8 @@ xfs_btree_block_change_owner(
 	struct xfs_btree_block	*block;
 	struct xfs_buf		*bp;
 	union xfs_btree_ptr     rptr;
+	struct xfs_owner_info	old_oinfo, new_oinfo;
+	int			error;
 
 	/* do right sibling readahead */
 	xfs_btree_readahead(cur, level, XFS_BTCUR_RIGHTRA);
@@ -3999,6 +4002,20 @@ xfs_btree_block_change_owner(
 	else
 		block->bb_u.s.bb_owner = cpu_to_be32(new_owner);
 
+	/* change rmap owners (bmbt blocks only) */
+	if (cur->bc_flags & XFS_BTREE_LONG_PTRS) {
+		XFS_RMAP_INO_BMBT_OWNER(&old_oinfo,
+				cur->bc_private.b.ip->i_ino,
+				cur->bc_private.b.whichfork);
+		XFS_RMAP_INO_BMBT_OWNER(&new_oinfo,
+				new_owner,
+				cur->bc_private.b.whichfork);
+		error = xfs_rmap_change_bmbt_owner(cur, bp, &old_oinfo,
+				&new_oinfo);
+		if (error)
+			return error;
+	}
+
 	/*
 	 * If the block is a root block hosted in an inode, we might not have a
 	 * buffer pointer here and we shouldn't attempt to log the change as the
diff --git a/fs/xfs/libxfs/xfs_rmap.c b/fs/xfs/libxfs/xfs_rmap.c
index d821b1a..b2781f5 100644
--- a/fs/xfs/libxfs/xfs_rmap.c
+++ b/fs/xfs/libxfs/xfs_rmap.c
@@ -35,6 +35,8 @@
 #include "xfs_trace.h"
 #include "xfs_error.h"
 #include "xfs_extent_busy.h"
+#include "xfs_bmap.h"
+#include "xfs_bmap_btree.h"
 
 /*
  * Lookup the first record less than or equal to [bno, len, owner, offset]
@@ -859,3 +861,113 @@ xfs_rmap_resize(
 done:
 	return error;
 }
+
+/**
+ * Change ownership of a file's BMBT block reverse-mappings.
+ */
+int
+xfs_rmap_change_bmbt_owner(
+	struct xfs_btree_cur	*bcur,
+	struct xfs_buf		*bp,
+	struct xfs_owner_info	*old_owner,
+	struct xfs_owner_info	*new_owner)
+{
+	struct xfs_buf		*agfbp;
+	xfs_fsblock_t		fsbno;
+	xfs_agnumber_t		agno;
+	xfs_agblock_t		agbno;
+	int			error;
+
+	if (!xfs_sb_version_hasrmapbt(&bcur->bc_mp->m_sb) || !bp)
+		return 0;
+
+	fsbno = XFS_DADDR_TO_FSB(bcur->bc_mp, XFS_BUF_ADDR(bp));
+	agno = XFS_FSB_TO_AGNO(bcur->bc_mp, fsbno);
+	agbno = XFS_FSB_TO_AGBNO(bcur->bc_mp, fsbno);
+
+	error = xfs_read_agf(bcur->bc_mp, bcur->bc_tp, agno, 0, &agfbp);
+
+	error = xfs_rmap_free(bcur->bc_tp, agfbp, agno, agbno, 1, old_owner);
+	if (error)
+		goto err;
+
+	error = xfs_rmap_alloc(bcur->bc_tp, agfbp, agno, agbno, 1, new_owner);
+	if (error)
+		goto err;
+
+err:
+	xfs_trans_brelse(bcur->bc_tp, agfbp);
+	return error;
+}
+
+/**
+ * Change the ownership on a file's extent's reverse-mappings.
+ */
+int
+xfs_rmap_change_extent_owner(
+	struct xfs_mount	*mp,
+	struct xfs_inode	*ip,
+	xfs_ino_t		ino,
+	xfs_fileoff_t		isize,
+	struct xfs_trans	*tp,
+	int			whichfork,
+	xfs_ino_t		new_owner)
+{
+	struct xfs_bmbt_irec	imap;
+	struct xfs_btree_cur	*cur = NULL;
+	struct xfs_buf		*agfbp = NULL;
+	int			nimaps;
+	xfs_fileoff_t		offset;
+	xfs_filblks_t		len;
+	xfs_agnumber_t		agno;
+	int			flags = 0;
+	int			error;
+
+	if (!xfs_sb_version_hasrmapbt(&mp->m_sb))
+		return 0;
+
+	if (whichfork == XFS_ATTR_FORK)
+		flags |= XFS_BMAPI_ATTRFORK;
+
+	offset = 0;
+	len = XFS_B_TO_FSB(mp, isize);
+	nimaps = 1;
+	error = xfs_bmapi_read(ip, offset, len, &imap, &nimaps, flags);
+	while (error == 0 && nimaps > 0) {
+		if (imap.br_startblock == HOLESTARTBLOCK ||
+		    imap.br_startblock == DELAYSTARTBLOCK)
+			goto advloop;
+
+		agno = XFS_FSB_TO_AGNO(mp, imap.br_startblock);
+
+		error = xfs_read_agf(mp, tp, agno, 0, &agfbp);
+		if (error)
+			break;
+
+		cur = xfs_rmapbt_init_cursor(mp, tp, agfbp, agno);
+
+		error = xfs_rmap_delete(cur, ino, whichfork, &imap);
+		if (error)
+			break;
+		error = xfs_rmap_insert(cur, new_owner, whichfork, &imap);
+		if (error)
+			break;
+
+		xfs_btree_del_cursor(cur, XFS_BTREE_NOERROR);
+		cur = NULL;
+		xfs_trans_brelse(tp, agfbp);
+		agfbp = NULL;
+advloop:
+		offset += imap.br_blockcount;
+		len -= imap.br_blockcount;
+		nimaps = 1;
+		error = xfs_bmapi_read(ip, offset, len, &imap, &nimaps, flags);
+	}
+
+	if (cur)
+		xfs_btree_del_cursor(cur, error ? XFS_BTREE_ERROR :
+					  XFS_BTREE_NOERROR);
+	if (agfbp)
+		xfs_trans_brelse(tp, agfbp);
+	return error;
+}
diff --git a/fs/xfs/libxfs/xfs_rmap_btree.h b/fs/xfs/libxfs/xfs_rmap_btree.h
index 0131d9a..5d248b5 100644
--- a/fs/xfs/libxfs/xfs_rmap_btree.h
+++ b/fs/xfs/libxfs/xfs_rmap_btree.h
@@ -88,4 +88,12 @@ int xfs_rmap_slide(struct xfs_btree_cur *rcur, xfs_ino_t ino, int whichfork,
 int xfs_rmap_resize(struct xfs_btree_cur *rcur, xfs_ino_t ino, int whichfork,
 		struct xfs_bmbt_irec *PREV, long size_adj);
 
+/* functions for changing rmap ownership */
+int xfs_rmap_change_extent_owner(struct xfs_mount *mp, struct xfs_inode *ip,
+		xfs_ino_t ino, xfs_fileoff_t isize, struct xfs_trans *tp,
+		int whichfork, xfs_ino_t new_owner);
+int xfs_rmap_change_bmbt_owner(struct xfs_btree_cur *bcur, struct xfs_buf *bp,
+		struct xfs_owner_info *old_owner,
+		struct xfs_owner_info *new_owner);
+
 #endif	/* __XFS_RMAP_BTREE_H__ */
diff --git a/fs/xfs/xfs_bmap_util.c b/fs/xfs/xfs_bmap_util.c
index f41a6f7..245a34a 100644
--- a/fs/xfs/xfs_bmap_util.c
+++ b/fs/xfs/xfs_bmap_util.c
@@ -40,6 +40,7 @@
 #include "xfs_trace.h"
 #include "xfs_icache.h"
 #include "xfs_log.h"
+#include "xfs_rmap_btree.h"
 
 /* Kernel only BMAP related definitions and functions */
 
@@ -1668,6 +1669,17 @@ xfs_swap_extent_flush(
 	return 0;
 }
 
+static int
+change_extent_owner(
+	struct xfs_inode	*ip,
+	struct xfs_trans	*tp,
+	int			whichfork,
+	struct xfs_inode	*tip)
+{
+	return xfs_rmap_change_extent_owner(ip->i_mount, ip, ip->i_ino,
+			ip->i_d.di_size, tp, whichfork, tip->i_ino);
+}
+
 int
 xfs_swap_extents(
 	xfs_inode_t	*ip,	/* target inode */
@@ -1684,19 +1696,7 @@ xfs_swap_extents(
 	int		taforkblks = 0;
 	__uint64_t	tmp;
 	int		lock_flags;
-
-	/*
-	 * We can't swap extents on rmap btree enabled filesystems yet
-	 * as there is no mechanism to update the owner of extents in
-	 * the rmap tree yet. Hence, for the moment, just reject attempts
-	 * to swap extents with EINVAL after emitting a warning once to remind
-	 * us this needs fixing.
-	 */
-	if (xfs_sb_version_hasrmapbt(&mp->m_sb)) {
-		WARN_ONCE(1,
-	"XFS: XFS_IOC_SWAPEXT not supported on RMAP enabled filesystems\n");
-		return -EINVAL;
-	}
+	struct xfs_trans_res	tres = {.tr_logres = 262144, .tr_logcount = 1, .tr_logflags = 0};
 
 	tempifp = kmem_alloc(sizeof(xfs_ifork_t), KM_MAYFAIL);
 	if (!tempifp) {
@@ -1734,7 +1734,9 @@ xfs_swap_extents(
 		goto out_unlock;
 
 	tp = xfs_trans_alloc(mp, XFS_TRANS_SWAPEXT);
-	error = xfs_trans_reserve(tp, &M_RES(mp)->tr_ichange, 0, 0);
+	/* XXX How do we create a potentially huge transaction here? */
+	/* error = xfs_trans_reserve(tp, &M_RES(mp)->tr_ichange, 0, 0); */
+	error = xfs_trans_reserve(tp, &tres, 0, 0);
 	if (error) {
 		xfs_trans_cancel(tp);
 		goto out_unlock;
@@ -1834,6 +1836,14 @@ xfs_swap_extents(
 			goto out_trans_cancel;
 	}
 
+	/* Change owners in the extent rmaps */
+	error = change_extent_owner(ip, tp, XFS_DATA_FORK, tip);
+	if (error)
+		goto out_trans_cancel;
+	error = change_extent_owner(tip, tp, XFS_DATA_FORK, ip);
+	if (error)
+		goto out_trans_cancel;
+
 	/*
 	 * Swap the data forks of the inodes
 	 */

