Re: [RFC PATCH 0/8] experimental struct filename followups

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

On Mon, Jan 12, 2026 at 11:00:10AM +0100, Christian Brauner wrote:
> On Thu, Jan 08, 2026 at 07:41:53AM +0000, Al Viro wrote:
> > This series switches the filename-consuming primitives to variants
> > that leave dropping the reference(s) to caller.  These days it's
> > fairly painless, and results look simpler wrt lifetime rules:
> > 	* with 3 exceptions, all instances have constructors and destructors
> > happen in the same scope (via CLASS(filename...), at that)
> > 	* CLASS(filename_consume) has no users left, could be dropped.
> > 	* exceptions are:
> > 		* audit dropping the references it stashed in audit_names
> > 		* fsconfig(2) creating and dropping references in two subcommands
> > 		* fs_lookup_param() playing silly buggers.
> > 	  That's it.
> > If we go that way, this will certainly get reordered back into the main series
> > and have several commits in there ripped apart and folded into these ones.
> > E.g. no sense to convert do_renameat2() et.al. to filename_consume, only to
> > have that followed by the first 6 commits here, etc.
> > 
> > For now I've put those into #experimental.filename, on top of #work.filename.
> > Comments would be very welcome...
> 
> Yeah, that looks nice. I like this a lot more than having calleee
> consume it.
> Reviewed-by: Christian Brauner <brauner@kernel.org>

FWIW, I've folded that into #work.filename and reordered the things to a somewhat
saner shape.  Will post the updated series shortly.

Open questions:

	* Exports.  Currently we have getname_kernel() and putname()
exported, while the rest of importers is not.  There is exactly one
module using those - ksmbd, and both users in it are doing only one
thing to resulting filename: passing it to vfs_path_parent_lookup().
No other callers of vfs_path_parent_lookup() exist.

	Options:
A) replace vfs_path_parent_lookup() with
int path_parent_root(const char *filename, unsigned int flags,
                     struct path *parent, struct qstr *last, int *type,
		     const struct path *root)
{
	CLASS(filename_kernel, name)(filename);
	return  __filename_parentat(AT_FDCWD, name, flags, parent, last,
					    type, root);
}
have that exported and used in fs/smb/server/vfs.c instead of
vfs_path_parent_lookup(); unexport getname_kernel() and putname().

B) make the rest of importers (well, CLASS(filename...), really) usable
for modules as well.  Then we probably want to publish filename_lookup()
as well.  Unattractive, IMO...

	* LOOKUP_EMPTY.  IMO putting it into LOOKUP_... space had been
a mistake.  We are actually pretty close to being able to extract it
from there; I'd love to make getname_flags() take boolean instead of this
"unsigned int, but we only care about one bit in it" thing.
	There's only one obstacle - modular callers of user_path_at() that
would possibly pass LOOKUP_EMPTY in flags.  Right now there's none (in-tree,
that is)- we have only 3 modular callers in the first place and they pass
only 0 or LOOKUP_FOLLOW in flags.
	I would rather provide user_path_maybe_null() if/when such beasts
appear; it's saner for any new API anyway.  I really wish we'd done it that
way for fspick(2), but it's too late to change now...
	Note that existing non-modular callers are better off with
CLASS(filename_...) + filename_lookup().  Modular ones can't do that,
due to the lack of exports; TBH, I'd rather keep that machinery internal
to the core kernel...

	Comments?