Re: [PATCH v7 1/4] spinlock: A new lockref structure for lockless update of refcount

[Al Viro <viro@ZenIV.linux.org.uk>]

On Sun, Sep 01, 2013 at 01:13:06AM +0100, Al Viro wrote:
> > Actually, right now I'm debugging a variant that avoids local buffers; use
> > is %pD3 for grandparent/parent/name, etc., up to %pD4.  %pd is equivalent
> > to %pD1 (just the dentry name).  Keep in mind that things like NFS use
> > a _lot_ of what would be %pD2 in debugging printks and the string can grow
> > fairly long, so I'd rather live with widen() than mess with local buffers
> > here.  I'll send an updated variant when I'm more or less satisfied with
> > it...
> 
> Seems to be working...  This doesn't include the metric arseload of
> conversions in fs/*/* - just the sprintf part.

FWIW, now that I've looked at more users (and we do have a shitload
of those), it seems that we need the following set:
	dentry name
	dentry path 2--4 levels deep (most of the users want 2 right now,
but that's mostly a matter of arguments being too painful to type for
deeper ones)
	same for file - sure, we can just pass file->f_path.dentry,
but there's a lot of such guys and I'd like to reduce the amount of
places where we have ->f_path.dentry mentioned in source.

Suggestions regarding formats to use would be welcome.  For now I'm using
pd/pd<n>/pD/pD<n>, the latter two being for struct file, but I'd gladly
take anything prettier than that.