Bug 7955

Reply-To: akpm@linux-foundation.org

On Tue, 6 Feb 2007 18:52:13 -0800 bugme-daemon@bugzilla.kernel.org wrote:

> http://bugzilla.kernel.org/show_bug.cgi?id=7955
> 
>            Summary: rootfstype used with linuxrc of different fs type fails
>                     to start
>     Kernel Version: 2.6.12
>             Status: NEW
>           Severity: normal
>              Owner: fs_other@kernel-bugs.osdl.org
>          Submitter: ken.cole@bownetworks.com
> 
> 
> Most recent kernel where this bug did *NOT* occur:
> Distribution:
> Mandriva Corporate 3.0
> Hardware Environment:
> x86
> Software Environment:
> Mandriva Corp 30
> Problem Description:
> 
> Steps to reproduce:
> (1) Create initial ram disk image file using one filesystem (say, for eg. ext2)
> (2) Create linuxrc file that redirects /dev/root to other filesystem (say ext3)
>          ...
>          echo Mounting /proc filesystem
>          mount -t proc /proc /proc
>          echo Creating device files
>          mkdevices /dev
>          echo Mounting sysfs
>          mount -t sysfs none /sys
>          echo Creating root device
>          echo 0x0306 > /proc/sys/kernel/real-root-dev
>          mkrootdev /dev/root
>          umount /sys
>          umount /proc
> (3) Set bootloader append line to include "root=3:6 rootfstype=ext3"
>    where 3:6 is the real root filesystem
> (4) Boot with the above initrd image and append line.
> Result:
> VFS: Cannot open root device "ram" or unknown_block(1,0)
> 
> My expectation was that initrd would be mounted ext2 followed by execution of
> linuxrc nash script followed by mounting device "3:6" using ext3 type.  However,
> the initrd was detected and loaded as ext2 but then gave error listed above.
> 
> The 'rootfstype' option seems to confuse the initial RAM disk load which had
> previously loaded the "ext2" RAM disk image.  The following patch to
> init/do_mounts.c allows the initrd image to remain ext2 while the real file
> system type obeys 'rootfstype' append line option.
> 
>  void __init mount_block_root(char *name, int flags)
>  {
> +static char const default_fs[] = "ext2\000\000";
>         char *fs_names = __getname();
>         char *p;
>         char b[BDEVNAME_SIZE];
> 
> -       get_fs_names(fs_names);
> +        if (   strstr(name, ".old") )
> +            memcpy(fs_names, default_fs, sizeof(default_fs));
> +        else
> +           get_fs_names(fs_names);
>  retry:
>         for (p = fs_names; *p; p += strlen(p)+1) {
>                 int err = do_mount_root(name, p, flags, root_mount_data);
> 

We'd of course prefer to not hardcode the fstype like this.  Perhaps someone
else has time to look into this, or to suggest a solution?

I understand your concern about hard-coding and having time to figure a better
approach.  A slightly better hard-coding (not tested) would be to use default
filesystem type "auto" because I understand that that choice checks filesystem
magic for selecting filesystem for the mount command.

If "auto\000\000" is also not acceptable then I would have to dig deeper to
propose another approach.  Any guidance as to a minimal requirement for
acceptance into the kernel would be appreciated (i.e. is auto acceptable?). 
Also, I am not yet familiar with the "tmpfs" technique for initrd images and I
would want that to work too for any proposed solution.

Ken,
You can start with autofs approach and if it works - just submit it (and get
full code review :) And if your solution won't be optimal - there will be debate
and good suggestions during this process.

--Natalie

Ken, any updates from you on this issue?
Thanks.