Re: [RFC] Raising the UBI version

[Boris Brezillon <boris.brezillon@free-electrons.com>]

On Wed, 22 Jun 2016 16:21:36 +0200
Richard Weinberger <richard@nod.at> wrote:

> Am 22.06.2016 um 16:13 schrieb Boris Brezillon:
> >> /sys/class/ubi/version is the version of the UBI implementation,
> >> not the version of the attached UBI image.
> >> It will the here as soon you load the UBI module.  
> > 
> > Do we have /sys/class/ubi/ubiX/version for the UBI image version?  
> 
> No. That's why I plan to add /sys/class/ubi/ubiX/features_used to
> show which features the attached UBI image requested.
> And having a /sys/class/ubi/ubi/features which denotes what features
> the _implementation_ supports.

Still the version and features are encoding different things IMO.
Incrementing the on-flash version means that the on-flash format has
changed in an incompatible way, while features denotes the fact that
the existing format has been extended with new features but is backward
compatible.

> 
> > This is still unclear to me why we need to version the
> > user-space/kernel-space ABI, since it's supposed to be backward
> > compatible, so adding new features requires adding new ioctls and
> > keeping the old ones in a working state.
> > 
> > What is /sys/class/ubi/version actually encoding? Isn't it encoding the
> > fact that a specific UBI implementation is supporting all UBI on-flash
> > formats up to format version X (that was my understanding)?
> >   
> 
> Well, /sys/class/ubi/version exports UBI_VERSION from ubi-media.h.
> It is (ab)used to encode the ABI version *and* the on-flash version.

The on-flash versions supported by the implementation is a useful
information.

> The problem is that mtd-utils libubi will refuse to work with
> /sys/class/ubi/version unequal 1.
> 
> Here the gem from libubi:
>         if (read_positive_int(lib->ubi_version, &version))
>                 goto out_error;
>         if (version != LIBUBI_UBI_VERSION) {
>                 errmsg("this library was made for UBI version %d, but UBI "
>                        "version %d is detected\n", LIBUBI_UBI_VERSION, version);
>                 goto out_error;
>         }

And this is where the problem is: libubi does not make proper use of
this information. We should either have

	if (version >= LIBUBI_UBI_VERSION)

or, if we decide that version is a bitfield directly encoding which
versions are supported by the implementation

#define VERSION_SUPPORTED(version, x)	((version) & BIT((x)-1)))

	if (VERSION_SUPPORTED(version, 1))

> 
> This is why I want to hardcode it to 1.
> Everything else will break existing user space in some way.
> 10 years ago /sys/class/ubi/version seemed like a good idea
> but now it hits us hard.

Yes, I understand that, but this also means /sys/class/ubi/version is
just a dummy file which only purpose is to make libubi happy :).
If this is the case, then I think we should have another file encoding
the supported on-flash formats...