On Thu, May 31, 2012 at 07:04:42AM +0000, Wrobel Heinz-R39252 wrote:
> Michael,
>
> > On Wed, 2012-05-30 at 16:33 +0200, Steffen Rumler wrote:
> > > I've found the following root cause:
> > >
> > > (6) Unfortunately, the trampoline code (do_plt_call()) is using
> > register r11 to setup the jump.
> > > It looks like the prologue and epilogue are using also the
> > register r11, in order to point to the previous stack frame.
> > > This is a conflict !!! The trampoline code is damaging the
> > content of r11.
> >
> > Hi Steffen,
> >
> > Great bug report!
> >
> > I can't quite work out what the standards say, the versions I'm looking
> > at are probably old anyway.
>
> The ABI supplement from https://www.power.org/resources/downloads/ is
> explicit about r11 being a requirement for the statically lined save/restore
> functions in section 3.3.4 on page 59.
>
> This means that any trampoline code must not ever use r11 or it can't be used
> to get to such save/restore functions safely from far away.
I believe that the basic premise is that you should provide a directly
reachable copy
of the save/rstore functions, even if this means that you need several copies
of the functions.
>
> Unfortunately the same doc and predecessors show r11 in all basic examples
> for PLT/trampoline code AFAICS, which is likely why all trampoline code uses
> r11 in any known case.
>
> I would guess that it was never envisioned that compiler generated code would
> be in a different section than save/restore functions, i.e., the Linux module
> "__init" assumptions for Power break the ABI. Or does the ABI break the
> __init concept?!
>
> Using r12 in the trampoline seems to be the obvious solution for module
> loading.
>
> But what about other code loading done? If, e.g., a user runs any app from
> bash it gets loaded and relocated and trampolines might get set up somehow.
I don't think so. The linker/whatever should generate a copy of the
save/restore functions for every
executable code area (shared library), and probably more than one copy if the
text becomes too large.
For 64 bit code, these functions are actually inserted by the linker.
[Ok, I just recompiled my 64 bit kernel with -Os and I see that vmlinux gets
one copy
of the save/restore functions and every module also gets its copy.]
This makes sense, really these functions are there for a compromise between
locality
and performance, there should be one per code section, otherwise the cache line
used by the trampoline negates a large part of their advantage.
>
> Wouldn't we have to find fix ANY trampoline code generator remotely related
> to a basic Power Architecture Linux?
> Or is it a basic assumption for anything but modules that compiler generated
> code may never ever be outside the .text section? I am not sure that would be
> a safe assumption.
>
> Isn't this problem going beyond just module loading for Power Architecture
> Linux?
I don't think so. It really seems to be a 32 bit kernel problem.
Regards,
Gabriel
_______________________________________________
Linuxppc-dev mailing list
Linuxppc-dev@lists.ozlabs.org
https://lists.ozlabs.org/listinfo/linuxppc-dev
