[PATCH 12/13] xfs: add version 3 inode format with CRCs

[Christoph Hellwig]

On Thu, Feb 12, 2009 at 12:42:02PM -0600, Eric Sandeen wrote:
> Josef 'Jeff' Sipek wrote:
> > On Tue, Feb 10, 2009 at 03:22:53PM -0500, Christoph Hellwig wrote:
> > ...
> >> Index: xfs/fs/xfs/xfs_dinode.h
> >> ===================================================================
> >> --- xfs.orig/fs/xfs/xfs_dinode.h	2009-02-10 19:45:51.939069576 +0100
> >> +++ xfs/fs/xfs/xfs_dinode.h	2009-02-10 19:45:59.195068745 +0100
> >> @@ -69,11 +69,32 @@ typedef struct xfs_dinode {
> >>  
> >>  	/* di_next_unlinked is the only non-core field in the old dinode */
> >>  	__be32		di_next_unlinked;/* agi unlinked list ptr */
> >> -} __attribute__((packed)) xfs_dinode_t;
> >> +
> >> +	/* start of the extended dinode, writable fields */
> >> +	__be32		di_crc;		/* CRC of the inode */
> >> +	__be64		di_changecount;	/* number of attribute changes */
> >> +	__u8		di_pad2[16];	/* more padding for future expansion */
> >> +
> >> +	/* fields only written to during inode creation */
> >> +	xfs_timestamp_t	di_crtime;	/* time created */
> >> +	__be64		di_ino;		/* inode number */
> >> +	uuid_t		di_uuid;	/* UUID of the filesystem */
> >> +} xfs_dinode_t;
> > 
> > Hrm...removing the packed attribute... Eric, do you remember the ARM ABI
> > alignment rules? Regardless of ARM, are those fields aligned nicely? (From a
> > quick glance at the code looks ok.)
> 
> I'll just have to just test it again I suppose.  Removing packed is
> good, as long as the padding makes it all come out right.

The reason why the current dinode needs the packed attribute is that
it's end on an un-even 32bit word, and thus the last field might be
wrongly aligned on 64-bit big-endian platforms (same issue as the
bad features2 one).  With the CRC patchset we now have a dinode
that's properly aligned on a even 32bit word.

> 
> -=Eric
---end quoted text---