Re: [PATCH] namei: clear nd->root.mnt before O_CREAT unlazy

[Brian Foster <bfoster@redhat.com>]

On Fri, Jan 07, 2022 at 05:51:52PM +0000, Al Viro wrote:
> On Fri, Jan 07, 2022 at 12:32:17PM -0500, Brian Foster wrote:
> 
> > > Other problems here (aside of whitespace damage - was that a
> > > cut'n'paste of some kind?  Looks like 8859-1 NBSP for each
> > > leading space...) are
> > 
> > Hmm.. I don't see any whitespace damage, even if I pull the patch back
> > from the mailing list into my tree..?
> 
> That had occured in Ian's reply, almost certainly.  Looks like whatever
> he's using for MUA (Evolution?) is misconfigured into doing whitespace
> damage - his next mail (in utf8, rather than 8859-1) had a scattering of
> U+00A0 in it...  Frankly, I'd never seen a decent GUI MUA, so I've no
> real experience with that thing and no suggestions on how to fix that.
> 
> > > 	* misleading name of the new helper - it sounds like
> > > "non-RCU side of complete_walk()" and that's not what it does
> > 
> > The intent was the opposite, of course. :P I'm not sure how you infer
> > the above from _rcu(), but I'll name the helper whatever. Suggestions?
> 
> s/non-// in the above (I really had been half-asleep).  What I'm
> saying is that this name invites an assumption that in RCU case
> complete_walk() is equivalent to it.  Which is wrong - that's
> what complete_walk() does as the first step if it needs to get
> out of RCU mode.
> 

Ah, Ok. I see your point. I suppose we could call it something like
__complete_walk() or __complete_walk_rcu() just to assert that it's a
subcomponent of complete_walk(). Alternatively, we could leave the
complete_walk() bits alone and create a slightly duplicate helper for
the create path (minus LOOKUP_CACHED), or just open code those bits in
the create path without a helper, or perhaps create a lower level helper
along the lines of:

static inline bool nd_reset_root_and_unlazy(struct nameidata *nd)
{
        /*      
         * We don't want to zero nd->root for scoped-lookups or
         * externally-managed nd->root.
         */
        if (!(nd->state & ND_ROOT_PRESET))
                if (!(nd->flags & LOOKUP_IS_SCOPED))
                        nd->root.mnt = NULL; 
        return try_to_unlazy(nd);
}

... and then just leave the LOOKUP_RCU check and LOOKUP_CACHED handling
open coded in the callers as they are now. Hm?

Brian