>From c93beb4367027f52979297dafae5728093f6a5d6 Mon Sep 17 00:00:00 2001
From: Bob Peterson <rpeterso@redhat.com>
Date: Fri, 19 Aug 2011 15:57:03 -0500
Subject: [PATCH 48/56] libgfs2: Make rebuild functions not re-read ip
Before this patch, the libgfs2 rebuild functions in structures.c
were not consistent about reading in the gfs2_inode in-core struct.
Some of the functions did and some didn't. The previous patch to
make fsck.gfs2 operate on gfs1 introduced a problem because the code
was changed to re-read the inum and statfs files if they were rebuilt.
But since the functions already did the read, it was a double-read.
A double-read is a bad thing because inode writes are now done on
the last inode put, and the double-read messed up the counter.
(Number of inode gets should match puts). This patch makes all the
build functions consistent: they all put the inode after building,
except for root and master which have no parent inodes.
There was also a minor problem whereby leaf searches were getting
the number of entries from the directory buffer rather than from
the dinode itself. (The buffer may be out of date, but the dinode
should always be current). So there were circumstances where you
could add a new dirent to "master" but the count of entries was not
correct in the buffer, which caused the search to stop prematurely.
Before this patch, the libgfs2 rebuild functions in structures.c
were not consistent about reading in the gfs2_inode in-core struct.
Some of the functions did and some didn't. The previous patch to
make fsck.gfs2 operate on gfs1 introduced a problem because the code
was changed to re-read the inum and statfs files if they were rebuilt.
But since the functions already did the read, it was a double-read.
A double-read is a bad thing because inode writes are now done on
the last inode put, and the double-read messed up the counter.
(Number of inode gets should match puts). This patch makes all the
build functions consistent: they all put the inode after building,
except for root and master which have no parent inodes.
There was also a minor problem whereby leaf searches were getting
the number of entries from the directory buffer rather than from
the dinode itself. (The buffer may be out of date, but the dinode
should always be current). So there were circumstances where you
could add a new dirent to "master" but the count of entries was not
correct in the buffer, which caused the search to stop prematurely.
diff --git a/gfs2/convert/gfs2_convert.c b/gfs2/convert/gfs2_convert.c
index 9cc3258..43d0093 100644
--- a/gfs2/convert/gfs2_convert.c
+++ b/gfs2/convert/gfs2_convert.c
@@ -2158,9 +2158,12 @@ int main(int argc, char **argv)
build_per_node(&sb2);
/* Create the empty inode number file */
build_inum(&sb2); /* Does not do inode_put */
+ gfs2_lookupi(sb2.master_dir, "inum", 4, &sb2.md.inum);
/* Create the statfs file */
build_statfs(&sb2); /* Does not do inode_put */
+ gfs2_lookupi(sb2.master_dir, "statfs", 6, &sb2.md.statfs);
+ do_init_statfs(&sb2);
/* Create the resource group index file */
build_rindex(&sb2);
/* Create the quota file */
diff --git a/gfs2/fsck/initialize.c b/gfs2/fsck/initialize.c
index cee1f49..75b363b 100644
--- a/gfs2/fsck/initialize.c
+++ b/gfs2/fsck/initialize.c
@@ -392,6 +392,7 @@ static int rebuild_master(struct gfs2_sbd *sdp)
IF2DT(S_IFREG | 0600));
} else {
build_inum(sdp);
+ gfs2_lookupi(sdp->master_dir, "inum", 4, &sdp->md.inum);
}
libgfs2: Make rebuild functions not re-read ip
