Re: [PATCH] mtd: allow mtd and jffs2 when ARCH=um

[Rob Landley <rob@landley.net>]

On Tuesday 14 December 2010 19:19:05 Jason Lunz wrote:
> On Tue, Dec 14, 2010 at 06:49:02PM -0600, Rob Landley wrote:
> > The problem is that jffs2 is a filesystem, and thus something people
> > would really like to be able to loopback mount, but it's hardwired to
> > assume it's only ever stored on a certain type of hardware, and thus
> > requies incestuous knowledge of the erase granularity of the flash layer
> > in order to function.
>
> I assume you can turn your jffs2 image file into a block dev using
> losetup, then turn the corresponding loop device into an mtd device
> using block2mtd, at which point you ought to be able to mount it with
> jffs2. I've never tried it.

That is awesome and I'm not finding any documentation on it...  Ah:

  http://wiki.maemo.org/Modifying_the_root_image#Block_device_emulating_an_MTD_device

Wow that's awkward.  Let's see, that says...

  mknod /tmp/mtdblock0 b 31 0
  modprobe loop
  losetup /dev/loop0 rootfs.jffs2
  modprobe mtdblock
  modprobe block2mtd
  # Note the ,128KiB is needed (on 2.6.26 at least) to set the
  # eraseblock  size.
  echo "/dev/loop0,128KiB" > /sys/module/block2mtd/parameters/block2mtd
  modprobe jffs2
  mount -t jffs2 /tmp/mtdblock0 /media/jffs2

So the system isn't automatically loading mtdblock, udev isn't creating any 
/dev nodes for it even thought it is for loop, the associations are created by 
writing strings into a filesystem with a _notoriously_ unstable API (and it 
expects a three letter mixed case suffix to specify units), and despite 
"mtdblock0" I have no idea how that echo syntax would specify more than one 
association at a time.

I'm trying to figure out whether creating a shell script for this or trying to 
modify the busybox mount command would be a better approach to beating some 
sort of usability out of doing this.

Hmmm, mkfs.jffs2 is named as a mkfs but acts a a generator ala genext2fs.  How 
about...

  mkdir empty
  mkfs.jffs2 -r empty -o rootfs.jffs2 -e 128 -l -n

And it created an empty (zero byte) file.  That's nice.  If I touch a file in 
"empty" now 116 bytes.  Adding -p makes it one erase block, but adding "-p 
2048" doesn't specify a multiple of the erase block size, it instead rounded 
rounded it _down_ to that many bytes.  But --pad=$((2048*1024*1024)) was 
apparently ignored...?  How do I specify a _size_ for this thing so it has 
empty space I can write into after the fact?  (The man page says pad with 
0xFF.  Am I going to have to do this by hand?)

  mkdir empty
  touch empty/hello
  mkfs.jffs2 -r empty -o temp.jffs2 -e 128 -l -n -p
  python -c "import sys; sys.stdout.write(131072*63*chr(0xff))" >> temp.jffs2
  losetup /dev/loop0 temp.jffs2
  echo "/dev/loop0,128KiB" > /sys/module/block2mtd/parameters/block2mtd
  mount -t jffs2 mtdblock0 empty

Yay!  I have a filesystem!  And df claims it's 8 megabytes.

Woot.

> > What any of this has to do with UML is an open question.  I don't want to
> > require UML to loopback mount a jffs2 image, I want to be able to do it
> > from my host.  From my perspective, you're solving the wrong problem.
>
> There's more than just loopback-mounting jffs2 images. I use this to
> make entire uml+jffs2 virtual machines for testing purposes.

As I said, I'm almost exclusively using qemu/kvm these days, but this is 
fun...

> Jason

Rob
-- 
GPLv3: as worthy a successor as The Phantom Menace, as timely as Duke Nukem 
Forever, and as welcome as New Coke.