Re: [PATCH 10/11] VFS: use d_alloc_parallel() in lookup_one_qstr_excl().

["NeilBrown" <neil@brown.name>]

On Wed, 13 Aug 2025, Al Viro wrote:
> On Tue, Aug 12, 2025 at 12:25:13PM +1000, NeilBrown wrote:
> 
> > + * If it is d_in_lookup() then these conditions can only be checked by the
> > + * file system when carrying out the intent (create or rename).
> 
> I do not understand.  In which cases would that happen and what would happen
> prior to that patch in the same cases?
> 

NFS (and I think it is only NFS) returns NULL from ->lookup() without
instantiating the dentry and without clearing DENTRY_PAR_LOOKUP if
passed "LOOKUP_CREATE | LOOKUP_EXCL" or "LOOKUP_RENAME_TARGET".

So when e.g. filename_create() calls lookup_one_qstr_excl() the result could
be a d_in_lookup() dentry.  It could be that the name exists on the
server, but the client hasn't bothered to check.  So determining that
the result wasn't ERR_PTR(-EEXIST) does NOT assure us that the name
doesn't exist.

The intent needs to be attempted, such as when do_mknodat() goes on to
call e.g.  vfs_create().  Only once that returns an error can we know if
the name existed.

i.e. the API promise:

+ *   Will return -EEXIST if name is found and LOOKUP_EXCL was passed.

must be understood against the background that the name might not be
found due to the lookup being short-circuited and not attempted.
The other promise:

+ *   Will return -ENOENT if name isn't found and LOOKUP_CREATE wasn't passed.

is currently safe from confusion, but I can imagine that one day a
LOOKUP_UNLINK intent could allow a filesystem to short-circuit the
lookup in do_unlinkat() and simply send an UNLINK request to a server
and return the result.

So I thought it worth highlighting the fact that these errors are
best-effort, and that d_in_lookup() is a real possibility.

NeilBrown