Re: [PATCH] Properly init address_space->writeback_index

[Andrew Morton <akpm@linux-foundation.org>]

On Thu, 14 Aug 2008 15:35:33 -0400
Chris Mason <chris.mason@oracle.com> wrote:

> > 
> > Note that this assumes that a filesystem which implements
> > ->alloc_inode() will call inode_init_once() within its ->alloc_inode().
> > 
> > Which means, I think, that we can just remove inode_init_once()
> > altogether and move its initialisations into alloc_inode() along with
> > all the existing ones.
> > 
> > What do you think?
> > 
> 
> There's the silent breakage of out of tree FS,

We could leave

void inode_init_once(struct inode *inode)
{
	WARN_ON_ONCE(1);
}
EXPORT_SYMBOL(inode_init_once);

in place for a while.

> but, your patch seems
> much cleaner to me.

It has the downside that we'll need to reinitialise all those
list_heads on each inode allocation, whereas currently that only
happens at slab-page-allocation time.  Trade that off against all the
open-coded zeroings which got removed then it's probably a wash from a
performance POV.

Or not.  A memset followed by a sprinkle of random writes might be
faster than no-memset followed by a sprinkle of random writes.  Or not ;)