Re: [PATCH v2 3/3] ovl: fix the parsing of empty string mount parameters

From: Miklos Szeredi <miklos@szeredi.hu>
To: Christian Brauner <brauner@kernel.org>
Cc: Luis Henriques <lhenriques@suse.de>,
	"Theodore Ts'o" <tytso@mit.edu>,
	 Andreas Dilger <adilger.kernel@dilger.ca>,
	Alexander Viro <viro@zeniv.linux.org.uk>,
	 Jan Kara <jack@suse.cz>, Amir Goldstein <amir73il@gmail.com>,
	linux-ext4@vger.kernel.org,  linux-fsdevel@vger.kernel.org,
	linux-unionfs@vger.kernel.org,  linux-kernel@vger.kernel.org
Subject: Re: [PATCH v2 3/3] ovl: fix the parsing of empty string mount parameters
Date: Mon, 11 Mar 2024 15:39:39 +0100	[thread overview]
Message-ID: <CAJfpegsn-jMY2J8Wd2Q9qmZFqxR6fAwZ4auoK+-uyxaK+F-0rw@mail.gmail.com> (raw)
In-Reply-To: <20240311-weltmeere-gesiegt-798c4201c3f8@brauner>

On Mon, 11 Mar 2024 at 14:25, Christian Brauner <brauner@kernel.org> wrote:

> Yeah, so with that I do agree. But have you read my reply to the other
> thread? I'd like to hear your thoughs on that. The problem is that
> mount(8) currently does:
>
> fsconfig(3, FSCONFIG_SET_FLAG, "usrjquota", NULL, 0) = -1 EINVAL (Invalid argument)
>
> for both -o usrjquota and -o usrjquota=

For "-o usrjquota" this seems right.

For "-o usrjquota=" it doesn't.  Flags should never have that "=", so
this seems buggy in more than one ways.

> So we need a clear contract with userspace or the in-kernel solution
> proposed here. I see the following options:
>
> (1) Userspace must know that mount options such as "usrjquota" that can
>     have no value must be specified as "usrjquota=" when passed to
>     mount(8). This in turn means we need to tell Karel to update
>     mount(8) to recognize this and infer from "usrjquota=" that it must
>     be passed as FSCONFIG_SET_STRING.

Yes, this is what I'm thinking.  Of course this only works if there
are no backward compatibility issues, if "-o usrjquota" worked in the
past and some systems out there relied on this, then this is not
sufficient.
>
> (2) We use the proposed in-kernel solution where relevant filesystems
>     get the ability to declare this both as a string or as a flag value
>     in their parameter parsing code. That's not a VFS generic thing.
>     It's a per-fs thing.

This encourages inconsistency between filesystems, but if there's no
other way to preserve backward compatibility, then...

>
> (3) We burden mount(8) with knowing what mount options are string
>     options that are allowed to be empty. This is clearly the least
>     preferable one, imho.
>
> (4) We add a sentinel such as "usrjquota=default" or
>     "usrjquota=auto" as a VFS level keyword.

I don't really understand how this last one is supposed to fix the issue.

> In any case, we need to document what we want:
>
> https://github.com/brauner/man-pages-md/blob/main/fsconfig.md

What's the plan with these?  It would be good if "man fsconfig" would
finally work.

Thanks,
Miklos