Re: bug in 2.6.22-rc2: loop mount limited to one single iso image

["Ray Lee" <ray-lk@madrabbit.org>]

On 5/20/07, Ken Chen <kenchen@google.com> wrote:
> On 5/19/07, Ray Lee <ray-lk@madrabbit.org> wrote:
> > Yeah, that's the only one left. I was hoping it wasn't that one, as it
> > claimed to have been tested extensively. Guess it wasn't tested with
> > udev.
> >
> > Ken? Ball's in your court. As the patch isn't providing a killer
> > feature for 2.6.22, I'd suggest just reverting it for now until the
> > issues are ironed out.
>
> The real solution is to have the user space tool to create these
> device nodes in advance.

Maybe. But requiring people upgrade their userspace tools or setup for
2.6.22 isn't a reasonable solution.

> The original loop patch was coded such that when we open a loop device
> N, the immediate adjacent device "N + 1" is created.  This will keep
> "mount -o loop" happy because it typically does a linear scan to find
> a free device.  This might be somewhat hackary, but certainly will be
> backward compatible before user space solution is deployed.

Except userspace is currently expecting 8 loop nodes upon bootup.
Creating n+1 when n is opened is good, but racy if userspace tries to
mount serveral loop devices in parallel.

If the loop device instantiates 8 (or max_loop) upon init, then we're
compatible with how things are being done in 2.6.21 and earlier.

> However, the code was removed by Al in this commit:

> commit 07002e995638b83a6987180f43722a0eb39d4932
> Author: Al Viro <viro@zeniv.linux.org.uk>
> Date:   Sat May 12 16:23:15 2007 -0400

>   fix the dynamic allocation and probe in loop.c

No, backing that code out wasn't good enough -- the reporter tested
reverting both of Al's patches out and was still getting errors on
boot. It took reverting your original one out as well to make it work.

So, a compromise? Let's start with 8 (or max_loop) populated, and then
we can move forward separately with teaching userspace new loop
tricks.