On Fri, Sep 06, 2019 at 09:16:04PM +0200, Richard Weinberger wrote:
> On Fri, Sep 6, 2019 at 1:15 PM Stefan Hajnoczi <stefanha@redhat.com> wrote:
> >
> > Make it possible to boot directly from a virtiofs file system with tag
> > 'myfs' using the following kernel parameters:
> >
> >   rootfstype=virtiofs root=myfs rw
> >
> > Booting directly from virtiofs makes it possible to use a directory on
> > the host as the root file system.  This is convenient for testing and
> > situations where manipulating disk image files is cumbersome.
> >
> > Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
> > ---
> > This patch is based on linux-next (next-20190904) but should apply
> > cleanly to other virtiofs trees.
> >
> >  init/do_mounts.c | 10 ++++++++++
> >  1 file changed, 10 insertions(+)
> >
> > diff --git a/init/do_mounts.c b/init/do_mounts.c
> > index 9634ecf3743d..030be2f1999a 100644
> > --- a/init/do_mounts.c
> > +++ b/init/do_mounts.c
> > @@ -554,6 +554,16 @@ void __init mount_root(void)
> >                         change_floppy("root floppy");
> >         }
> >  #endif
> > +#ifdef CONFIG_VIRTIO_FS
> > +       if (root_fs_names && !strcmp(root_fs_names, "virtiofs")) {
> > +               if (!do_mount_root(root_device_name, "virtiofs",
> > +                                  root_mountflags, root_mount_data))
> > +                       return;
> > +
> > +               panic("VFS: Unable to mount root fs \"%s\" from virtiofs",
> > +                     root_device_name);
> > +       }
> > +#endif
> 
> I think you don't need this, you can abuse a hack for mtd/ubi in
> prepare_namespace().
> At least for 9p it works well:
> qemu-system-x86_64 -m 4G -M pc,accel=kvm -nographic -kernel
> arch/x86/boot/bzImage -append "rootfstype=9p
> rootflags=trans=virtio,version=9p2000.L root=mtdfake console=ttyS0 ro
> init=/bin/sh" -virtfs
> local,id=rootfs,path=/,security_model=passthrough,mount_tag=mtdfake

That is worse because:
1. The file system must be named "mtd*" or "ubi*".
2. When mounting fails you get confusing error messages about block
   devices and partitions.  These do not apply to virtio-fs or
   virtio-9p.

> If this works too for virtiofs I suggest to cleanup the hack and
> generalize it. B-)

Why mtd and ubi block devices even have a special case?  Maybe this code
was added because ROOT_DEV = name_to_dev_t(root_device_name) doesn't
work for "mtd:partition" device names so the regular CONFIG_BLOCK code
path doesn't work for these devices.

Given the ordering/fallback logic in prepare_namespace()/mount_root() I
don't feel comfortable changing other code paths.  It's likely to break
something.

If you or others have a concrete suggestion for how to generalize this
I'm happy to try implementing it.

Stefan