why is short-circuiting nfs_lookup() for mkdir(2) et.al. dependent upon v3 or later?

why is short-circuiting nfs_lookup() for mkdir(2) et.al. dependent upon v3 or later?

[Al Viro]

        In nfs_lookup() we have
        /*
         * If we're doing an exclusive create, optimize away the lookup
         * but don't hash the dentry.
         */
        if (nfs_is_exclusive_create(dir, flags) || flags & LOOKUP_RENAME_TARGET)
                return NULL;
OK, fair enough - we don't need to find out whether it's negative or not for
mkdir() et.al.; if it isn't, server will tell us to sod off and we can live
with not having it in cache - in the worst case, we'll have to do the same
lookup we'd skipped here at some later point.  Same for rename(2) destination -
if it wasn't in dcache, we are not going to bother with sillyrename anyway, and
that's the only thing where we might care about the destination.  If rename(2)
succeeds, we won't see whatever had been there anyway, and if it fails, we won't
lose anything from having lookup done later.

        What I don't get is why, unlike rename(2) target, mkdir(2) argument is
handled that way only for v3 and later.  It's been a long time since I looked
at NFSv2 servers, but shouldn't we get NFSERR_EXIST if the sucker turns out to
have already been there?

        What am I missing?

Re: why is short-circuiting nfs_lookup() for mkdir(2) et.al. dependent upon v3 or later?

[Trond Myklebust]

On Tue, 2021-03-30 at 00:48 +0000, Al Viro wrote:
>         In nfs_lookup() we have
>         /*
>          * If we're doing an exclusive create, optimize away the
> lookup
>          * but don't hash the dentry.
>          */
>         if (nfs_is_exclusive_create(dir, flags) || flags &
> LOOKUP_RENAME_TARGET)
>                 return NULL;
> OK, fair enough - we don't need to find out whether it's negative or
> not for
> mkdir() et.al.; if it isn't, server will tell us to sod off and we
> can live
> with not having it in cache - in the worst case, we'll have to do the
> same
> lookup we'd skipped here at some later point.  Same for rename(2)
> destination -
> if it wasn't in dcache, we are not going to bother with sillyrename
> anyway, and
> that's the only thing where we might care about the destination.  If
> rename(2)
> succeeds, we won't see whatever had been there anyway, and if it
> fails, we won't
> lose anything from having lookup done later.
> 
>         What I don't get is why, unlike rename(2) target, mkdir(2)
> argument is
> handled that way only for v3 and later.  It's been a long time since
> I looked
> at NFSv2 servers, but shouldn't we get NFSERR_EXIST if the sucker
> turns out to
> have already been there?
> 
>         What am I missing?

The check for NFS version > 2 is mainly there for the case of CREATE,
which does not have an exclusive create mode in NFSv2.
IOW: I believe we can indeed rely on LINK, MKDIR, and RENAME returning
NFSERR_EXIST, but not the more common case of CREATE.

-- 
Trond Myklebust
Linux NFS client maintainer, Hammerspace
trond.myklebust@hammerspace.com