Re: Live patching for powerpc

[Jessica Yu]

+++ Miroslav Benes [14/04/16 15:28 +0200]:
> On Wed, 13 Apr 2016, Jessica Yu wrote:
>
> > +++ Miroslav Benes [13/04/16 15:01 +0200]:
> > > On Wed, 13 Apr 2016, Michael Ellerman wrote:
> > >
> > > > This series adds live patching support for powerpc (ppc64le only ATM).
> > > >
> > > > It's unchanged since the version I posted on March 24, with the exception
> > > > that
> > > > I've dropped the first patch, which was a testing-only patch.
> > > >
> > > > If there's no further comments I'll put this in a topic branch in the next
> > > > day
> > > > or two and Jiri & I will both merge that into next.
> > >
> > > Hi,
> > >
> > > I'll definitely give it a proper look today or tomorrow, but there is one
> > > thing that needs to be solved. The patch set from Jessica reworking
> > > relocations for live patching is now merged in our for-next branch. This
> > > means that we need to find out if there is something in struct
> > > mod_arch_specific for powerpc which needs to be preserved and do it.
> > >
> >
> > I took a look around the powerpc module.c code and it looks like the
> > mod_arch_specific stuff should be fine, since it is statically allocated
> > in the module struct (unlike the situation in s390, where
> > mod->arch.syminfo was vmalloc'd and we had to delay the free).
> > However I'm not familiar with the powerpc code so I need to dig around
> > a bit more to be 100% sure.
>
> I came to the same conclusion. There is only struct bug_entry *bug_table
> in mod_arch_specific but it looks unimportant wrt relocations.

Yeah, I think we are fine. As long as none of the values in
mod_arch_specific are "cleared," and I don't see that happening
anywhere.

> > A second concern I have is that apply_relocate_add() relies on
> > sections like .stubs and .toc (for 64-bit) and .init.plt and .plt
> > sections (for 32-bit). In order for apply_relocate_add() to work for
> > livepatch, we must make sure these sections aren't thrown away and are
> > not in init module memory since this memory will be freed at the end
> > of module load (see how INIT_OFFSET_MASK is used in kernel/module.c).
> > As long as these sections are placed in module core memory, we will be
> > OK. I need to think about this a bit more.
>
> I knew I shouldn't have opened arch/powerpc/kernel/module*.c.
>
> We could always hack sh_flags of those sections in
> module_arch_frob_sections() to make them stay.

I think we are fine here too. The onus would be on the patch build
tool (e.g., kpatch) to set the sh_flags to SHF_ALLOC, like we
already do to keep the klp relocation sections in memory :-)

For the 32-bit module code, I don't believe we would need to preserve
the .init.plt section for livepatch's call to apply_relocate_add(),
since relocations to init sections should've been applied during
module initialization, and we don't patch those types of functions.
Please correct me if my understanding is off.

Jessica