Re: [patch 5/5] elf: Add support for loading ET_CKPT files

[Tejun Heo <tj@kernel.org>]

Hello,

On Wed, Oct 19, 2011 at 11:56:32PM +0400, Cyrill Gorcunov wrote:
> > I find that quite difficult to agree with.  We're talking about some
> > minor additions to prctl against updating exec path to do something it
> > was never designed to do + new binary format.
> 
> Tejun, regardless the mm_struct entries, what about mapping vdso pages at the
> place they had on snapshot time? How would we handle them?

We discussed this elsewhere but just for the record.

I haven't really looked into vdso but what the currently suggested
implementation isn't correct.  vdso is there so that kernel can
provide userspace code which may vary depending on kernel version and
the actual machine it's running on, so both saving vdso and restoring
verbatim and asking the kernel to map its vdso at certain address are
wrong.

The checkpointer would need to remember the vdso symtab from the
original process and then somehow build indirect jump table at the
same address which redirects to the vdso of the machine the target
process is being restored on.  I'm not sure about the details of the
implementation but it could be something like "please allow me to
write to this original vdso address and keep your vdso elsewhere".

I think this comes back to why one-stop-solution-in-kernel is a bad
idea for CR.  My impression is that when people say "that would
require too many small API updatesa", it usually means that they
haven't really thought about each necessary piece enough and just
hacked something up lumpy and fuzzy on the edges.  The thing is that
no matter how you lump them, those problems don't go away and
attacking things in small understandable API pieces not only ensures
the update itself makes sense and may be useful for others too but
also forces people to actually think about each problem.

Thank you.

-- 
tejun