[PATCH 1/2] ubi: mount partitions specified in device tree

[Arnd Bergmann]

On Monday, June 20, 2016 10:26:11 AM CEST Richard Weinberger wrote:
> Am 20.06.2016 um 10:09 schrieb Arnd Bergmann:
> > A few more thoughts on various things that came up in the discussion
> > so far (so I don't have to reply to each of them separately)
> > 
> > - The /chosen and /aliases nodes do configure the OS to some degree,
> >   and historically we have used them to specify the device names that
> >   show up in the kernel (e.g. on MacOS and AIX) as well as the rootfs
> >   itself
> > 
> > - The device tree should give us enough information to identify the
> >   root partition by a GUID from the kernel. This works on all block
> >   devices (at least with MBR and EFI partition tables, probably many
> >   of the lesser known ones as well) and it should really also work
> >   on UBI.
> 
> It does.
> 
> > - MTD partitioning unfortunately does not seem to have a widely used
> >   method for identifying partitions from what's stored on the flash,
> >   unlike block devices, so whatever we normally have in the partition
> >   table has to be stored in DT. This really sucks, but I don't know
> >   what else to do about it.
> 
> Since blocks on a MTD can render bad you'd lose the table sooner or later.
> That's why we cannot store it on the MTD directly.
> Defining the table in DT is at least less ugly than using the mtdparts=
> kernel parameter.

Right, there would be no benefit in using the kernel command line,
it just moves the information to another place inside of the same DT
(the /chosen property).

I think you can normally rely on the first block being readable
on flash, in particular if you write it very rarely (when updating
the partition information), so it would be technically possible to
have a partition table in there, but in practice that's not how
things are done, so the argument is useless.

> > - The UBI design was originally meant to cover the entire flash
> >   device and provide the partitioning information inside. I have not
> >   looked at UBI in a long time (actually not since before embedded
> >   PPC moved to DT), but I would hope that this case works without
> >   any DT magic.
> 
> Of course it works without DT magic.
> 
> So, I don't see a urge to have UBI DT bindings.
> First I thought having an attach flag would be handy but it turned out
> that this behavior can easily achieved by naming. i.e. you a kernel
> command line such as "ubi.mtd=UBI" and name the to be attached MTD "UBI"
> in your DT.

Right. I think either way can work, and I see advantages in both.
The downside of relying just on the name is that mounting the rootfs
by GUID doesn't work until you specify another command line, which
is a bit annoying but not a huge problem.

	Arnd