Re: [PATCH v10 04/10] kexec_file: Add support for purgatory built as PIE.

[Thiago Jung Bauermann]

Am Mittwoch, 23. November 2016, 09:32:58 BRST schrieb Dave Young:
> On 11/22/16 at 11:44am, Thiago Jung Bauermann wrote:
> > Am Dienstag, 22. November 2016, 17:01:10 BRST schrieb Michael Ellerman:
> > > Thiago Jung Bauermann <bauerman@xxxxxxxxxxxxxxxxxx> writes:
> > > > Am Sonntag, 20. November 2016, 10:45:46 BRST schrieb Dave Young:
> > > >> On 11/10/16 at 01:27am, Thiago Jung Bauermann wrote:
> > > >> > powerpc's purgatory.ro has 12 relocation types when built as
> > > >> > a relocatable object. To implement support for them requires
> > > >> > arch_kexec_apply_relocations_add to duplicate a lot of code with
> > > >> > module_64.c:apply_relocate_add.
> > > >> > 
> > > >> > When built as a Position Independent Executable there are only 4
> > > >> > relocation types in purgatory.ro, so it becomes practical for the
> > > >> > powerpc
> > > >> > implementation of kexec_file to have its own relocation
> > > >> > implementation.
> > > >> > 
> > > >> > Also, the purgatory is an executable and not an intermediary output
> > > >> > from
> > > >> > the compiler so it makes sense conceptually that it is easier to
> > > >> > build
> > > >> > it as a PIE than as a partially linked object.
> > > >> > 
> > > >> > Apart from the greatly reduced number of relocations, there are two
> > > >> > differences between a relocatable object and a PIE:
> > > >> > 
> > > >> > 1. __kexec_load_purgatory needs to use the program headers rather
> > > >> > than
> > > >> > the
> > > >> > 
> > > >> >    section headers to figure out how to load the binary.
> > > >> > 
> > > >> > 2. Symbol values are absolute addresses instead of relative to the
> > > >> > 
> > > >> >    start of the section.
> > > >> > 
> > > >> > This patch adds the support needed in generic code for the
> > > >> > differences
> > > >> > above and allows powerpc to load and relocate a position
> > > >> > independent
> > > >> > purgatory.
> > > >> 
> > > >> [snip]
> > > >> 
> > > >> The kexec-tools machine_apply_elf_rel is pretty simple for ppc64, it
> > > >> is
> > > >> not that complex. So could you look into simplify your kexec_file
> > > >> implementation?
> > > > 
> > > > I can try, but there is one fundamental issue here: powerpc
> > > > position-dependent code relies more on relocations than x86
> > > > position-dependent code does, so there's a limit to how simple it can
> > > > be
> > > > made without switching to position- independent code. And it will
> > > > always
> > > > be more involved than it is on x86.
> > > 
> > > I think we need to go back to the drawing board on this one.
> > > 
> > > My hope was that building purgatory as PIE would reduce the amount of
> > > complexity, but instead it's just added more. Sorry for sending you in
> > > that direction.
> > 
> > It added complexity because in my series powerpc was using a PIE purgatory
> > but x86 kept using a partially-linked object (because of the problem I
> > mentioned I had when trying out a PIE x86 purgatory), so generic code
> > needed two purgatory loaders.
> > 
> > I'll see if I can make the PIE x86 purgatory to work so that generic code
> > can have only one loader implementation. Then it will indeed be simpler.
> Do we really need the PIE purgatory, after moving generic code out of
> x86, there will be no much benefit, no?

It still makes a big difference on powerpc, even after moving out the generic 
code. I just got the PIE purgatory working on x86 and it also simplifies the 
code there, so it's a win for both architectures.

I'll clean up the code and post tomorrow so that you can see what you think.

> Anyway, the first step should be
> making the purgatory code more generic so that it can be easier for
> other arches to support kexec_file in the future.

I'll try putting sha256.c in lib/purgatory/ as you suggested.

-- 
Thiago Jung Bauermann
IBM Linux Technology Center