[Cluster-devel] [PATCH v2] GFS: write vfs inode attributes to disk

Benjamin Marzinski bmarzins at redhat.com
Fri May 30 05:19:41 UTC 2014


GFS wasn't ever updating mtime during mmaps to a file.  This patch makes
gfs_write_inode write the vfs inode attributes out to disk, so that they
get updated when the mmap data is written back to disk. It also makes
sure that the updated vfs timestamps don't get overwritten by older gfs
ones before they can be written out.

This patch is similar to my earlier version, but includes more checks in
gfs_write_inode to make sure that it can start a transaction, and changes
to inode_attr_in to keep updated vfs timestamps from getting overwritten.
These changes were made to resolve issues that caused bz #1066181 to fail
QA. I have verified that it is still possible to manually reset the inode
mtime to an earlier time using the touch command.

Signed-off-by: Benjamin Marzinski <bmarzins at redhat.com>
---
 gfs-kernel/src/gfs/inode.c     | 15 +++++++++------
 gfs-kernel/src/gfs/ops_super.c | 43 ++++++++++++++++++++++++++++++++++++++++--
 2 files changed, 50 insertions(+), 8 deletions(-)

diff --git a/gfs-kernel/src/gfs/inode.c b/gfs-kernel/src/gfs/inode.c
index 454366a..52b0819 100644
--- a/gfs-kernel/src/gfs/inode.c
+++ b/gfs-kernel/src/gfs/inode.c
@@ -46,7 +46,7 @@
  */
 
 static void
-inode_attr_in(struct gfs_inode *ip, struct inode *inode)
+inode_attr_in(struct gfs_inode *ip, struct inode *inode, int force)
 {
 	unsigned int mode;
 
@@ -93,9 +93,12 @@ inode_attr_in(struct gfs_inode *ip, struct inode *inode)
 	inode->i_uid = ip->i_di.di_uid;
 	inode->i_gid = ip->i_di.di_gid;
 	i_size_write(inode, ip->i_di.di_size);
-	inode->i_atime.tv_sec = ip->i_di.di_atime;
-	inode->i_mtime.tv_sec = ip->i_di.di_mtime;
-	inode->i_ctime.tv_sec = ip->i_di.di_ctime;
+	if (force || ip->i_di.di_atime > inode->i_atime.tv_sec)
+		inode->i_atime.tv_sec = ip->i_di.di_atime;
+	if (force || ip->i_di.di_mtime > inode->i_mtime.tv_sec)
+		inode->i_mtime.tv_sec = ip->i_di.di_mtime;
+	if (force || ip->i_di.di_ctime > inode->i_ctime.tv_sec)
+		inode->i_ctime.tv_sec = ip->i_di.di_ctime;
 	inode->i_atime.tv_nsec = inode->i_mtime.tv_nsec = inode->i_ctime.tv_nsec = 0;
 	inode->i_blocks = ip->i_di.di_blocks <<
 		(ip->i_sbd->sd_sb.sb_bsize_shift - GFS_BASIC_BLOCK_SHIFT);
@@ -125,7 +128,7 @@ gfs_inode_attr_in(struct gfs_inode *ip)
 
 	inode = gfs_iget(ip, NO_CREATE);
 	if (inode) {
-		inode_attr_in(ip, inode);
+		inode_attr_in(ip, inode, 0);
 		iput(inode);
 	}
 
@@ -185,7 +188,7 @@ gfs_iget(struct gfs_inode *ip, int create)
 	if (!tmp)
 		return NULL;
 
-	inode_attr_in(ip, tmp);
+	inode_attr_in(ip, tmp, 1);
 
 	/* Attach GFS-specific ops vectors */
 	if (ip->i_di.di_type == GFS_FILE_REG) {
diff --git a/gfs-kernel/src/gfs/ops_super.c b/gfs-kernel/src/gfs/ops_super.c
index c111a2e..10dbff7 100644
--- a/gfs-kernel/src/gfs/ops_super.c
+++ b/gfs-kernel/src/gfs/ops_super.c
@@ -39,6 +39,7 @@
 #include "super.h"
 #include "sys.h"
 #include "mount.h"
+#include "trans.h"
 
 /**
  * gfs_write_inode - Make sure the inode is stable on the disk
@@ -51,14 +52,52 @@
 static int
 gfs_write_inode(struct inode *inode, int sync)
 {
+	int ret = 0;
 	struct gfs_inode *ip = get_v2ip(inode);
+	struct buffer_head *dibh;
+	struct gfs_holder i_gh;
+	int need_unlock = 0;
+
+	if (!ip)
+		return 0;
 
 	atomic_inc(&ip->i_sbd->sd_ops_super);
 
-	if (ip && sync)
+	if (current->flags & PF_MEMALLOC || get_transaction)
+		goto do_flush;
+
+	if (!gfs_glock_is_locked_by_me(ip->i_gl)) {
+		ret = gfs_glock_nq_init(ip->i_gl, LM_ST_EXCLUSIVE, 0, &i_gh);
+		if (ret)
+			goto do_flush;
+		need_unlock = 1;
+	}
+	else if (!gfs_glock_is_held_excl(ip->i_gl))
+		goto do_flush;
+	/* Trans may require:
+	   one dinode block. */
+	ret = gfs_trans_begin(ip->i_sbd, 1, 0);
+	if (ret)
+		goto do_unlock;
+
+	ret = gfs_get_inode_buffer(ip, &dibh);
+	if (ret == 0) {
+		gfs_inode_attr_out(ip);
+		gfs_trans_add_bh(ip->i_gl, dibh);
+		gfs_dinode_out(&ip->i_di, dibh->b_data);
+		brelse(dibh);
+	}
+
+	gfs_trans_end(ip->i_sbd);
+
+do_unlock:
+	if (need_unlock)
+		gfs_glock_dq_uninit(&i_gh);
+do_flush:
+	if (sync)
 		gfs_log_flush_glock(ip->i_gl, 0);
 
-	return 0;
+	return ret;
 }
 
 /**
-- 
1.8.3.1




More information about the Cluster-devel mailing list