Re: [PATCH v3 0/3] 9p: add support for root file systems

[Changbin Du <changbin.du@gmail.com>]

On Mon, Jun 14, 2021 at 02:19:38AM -0700, Josh Triplett wrote:
> On Mon, Jun 14, 2021 at 06:01:44PM +0900, Dominique Martinet wrote:
> > Josh Triplett wrote on Mon, Jun 14, 2021 at 01:57:54AM -0700:
> > > On Mon, Jun 07, 2021 at 07:09:19AM +0800, Changbin Du wrote:
> > > > Just like cifs and nfs, this short series enables rootfs support for 9p.
> > > > Bellow is an example which mounts v9fs with tag 'r' as rootfs in qemu
> > > > guest via virtio transport.
> > > > 
> > > >   $ qemu-system-x86_64 -enable-kvm -cpu host -m 1024 \
> > > >         -virtfs local,path=$rootfs_dir,mount_tag=r,security_model=passthrough,id=r \
> > > >         -kernel /path/to/linux/arch/x86/boot/bzImage -nographic \
> > > >         -append "root=/dev/v9fs v9fsroot=r,trans=virtio rw console=ttyS0 3"
> > > 
> > > Rather than inventing a pseudo-device /dev/v9fs for this, would it
> > > potentially work to use the existing rootfstype and rootflags options
> > > for this? rootfstype already determines what filesystem should be used
> > > to mount the root, and rootflags already provides options for that
> > > filesystem.
> > > 
> > > For instance, for the above example:
> > > rootfstype=9p root=r rootflags=trans=virtio
> > > 
> > > That would require a bit of fiddling to make rootfstype=9p allow a root
> > > that's just the mount_tag. If that isn't an option, then even with
> > > root=/dev/v9fs I think it still makes sense to use the existing
> > > rootflags for "trans=virtio" rather than creating a new "v9fsroot"
> > > option for that.
> > 
> > This doesn't work as is because of the way the code is written, if
> > there's no block device associated with a root=x option right now it
> > will lead to kernel panic.
> > 
> > I replied with folks in Cc but there's another thread on linux-fsdevel@
> > with a more generic approach that will build a list of filesystems which
> > don't require such a block device (either hardcoded with virtiofs and 9p
> > or based on FS_REQUIRES_DEV), thread started there but there's a second
> > patch hidden an more discussion below:
> > https://lore.kernel.org/linux-fsdevel/20210608153524.GB504497@redhat.com/
> 
> The patch later on in that thread (either using a list of
> non-block-device filesystems or the version referenced elsewhere that
> uses a flag in the filesystem definition) looks really appealing! That's
> exactly what I was hoping for. That gets us closer to directly
> translating `mount -t type -o options rootdesc` into `rootfstype=type
> rootflags=options root=rootdesc` in the general case, rather than having
> special cases for different filesystems.
Bellow is all the parameters of non-block rootfs support.
 nfs:    nfsroot=[<server-ip>:]<root-dir>[,<nfs-options>]
 cifs:   cifsroot=//<server-ip>/<share>[,options]
 v9fs:   v9fsroot=<tag/ip>[,options]
 mtd:    root=mtd:<identifier> or root=ubi:<identifier>
 virtiofs: root=fstag:<tag>

The main problem is we lack a generic handing for non-block rootdev. I think
maybe we can unify all of above.
 non-block:  root=<type>:<identifier>
 blockdev:   root=<bock-dev-path> or root=<blockdev major/minor>

Then:
 nfs:    root=nfs:[<server-ip>:]<root-dir>  rootflags=[nfs-options]
 cifs:   root=cifs://<server-ip>/<share>    rootflags=[options]
 v9fs:   root=9p:<tag/ip>                   rootflags=[options]
 mtd:    root=mtd:<identifier>              rootflags=[options]
 ubi:    root=ubi:<identifier>              rootflags=[options]
 virtiofs: root=virtiofs:<tag>              rootflags=[options]

And maybe we can also remove all the special fs code out of init/do_mounts.c.

-- 
Cheers,
Changbin Du