Incorrect error message from erofs "backed by file" in 6.12-rc

[Christian Brauner]

On Tue, Oct 08, 2024 at 10:13:31AM GMT, Gao Xiang wrote:
> Hi Christian,
> 
> On 2024/10/7 19:35, Christian Brauner wrote:
> > On Sat, Oct 05, 2024 at 10:41:10PM GMT, Gao Xiang wrote:
> 
> ...
> 
> > > 
> > > Hi Christian, if possible, could you give some other
> > > idea to handle this case better? Many thanks!
> 
> Thanks for the reply!
> 
> > 
> > (1) Require that the path be qualified like:
> > 
> >      fsconfig(<fd>, FSCONFIG_SET_STRING, "source", "file:/home/lis/src/mountcfs/cfs", 0)
> > 
> >      and match on it in either erofs_*_get_tree() or by adding a custom
> >      function for the Opt_source/"source" parameter.
> 
> IMHO, Users could create names with the prefix `file:`,
> it's somewhat strange to define a fixed prefix by the
> definition of source path fc->source.
> 
> Although there could be some escape character likewise
> way, but I'm not sure if it's worthwhile to work out
> this in kernel.
> 
> > 
> > (2) Add a erofs specific "source-file" mount option. IOW, check that
> >      either "source-file" or "source" was specified but not both. You
> >      could even set fc->source to "source-file" value and fail if
> >      fc->source is already set. You get the idea.
> 
> I once thought to add a new mount option too, yet from
> the user perpertive, I think users may not care about
> the source type of an arbitary path, and the kernel also
> can parse the type of the source path directly... so..
> 
> 
> So.. I wonder if it's possible to add a new VFS interface
> like get_tree_bdev_by_dev() for filesystems to specify a
> device number rather than hardcoded hard-coded source path
> way, e.g. I could see the potential benefits other than
> the EROFS use case:
> 
>  - Filesystems can have other ways to get a bdev-based sb
>    in addition to the current hard-coded source path way;
> 
>  - Some pseudo fs can use this way to generate a fs from a
>    bdev.
> 
>  - Just like get_tree_nodev(), it doesn't strictly tie to
>    fc->source too.
> 
> Also EROFS could lookup_bdev() (this kAPI is already
> exported) itself to check if it uses get_tree_bdev_by_dev()
> or get_tree_nodev()... Does it sounds good?  Many thanks!

Sounds fair to me.