qemu-devel
[Top][All Lists]
Advanced

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Qemu-devel] [RFC PATCH 2/2] spapr: -kernel: allow linking with spec


From: Andrew Jones
Subject: Re: [Qemu-devel] [RFC PATCH 2/2] spapr: -kernel: allow linking with specified addr
Date: Mon, 20 Jul 2015 15:09:20 +0200
User-agent: Mutt/1.5.23.1 (2014-03-12)

On Mon, Jul 20, 2015 at 08:47:53AM +0200, Thomas Huth wrote:
> On 20/07/15 07:01, David Gibson wrote:
> > On Fri, Jul 17, 2015 at 01:56:40PM +0200, Andrew Jones wrote:
> >> I've started playing with adding ppc support to kvm-unit-tests,
> >> using spapr for the machine model. I wanted to link the unit test
> >> at 0x400000 to match qemu's load address, making the unit test
> >> startup code simpler, but ended up with 0x800000 instead, due to
> >> how translate_kernel_address works. The translation makes sense
> >> for how Linux kernels are linked (always at 0xc0000000 or
> >> 0xc000000000000000), but for the unit test case we need to avoid
> >> adding the offset.
> >>
> >> Signed-off-by: Andrew Jones <address@hidden>
> >> ---
> >> Big RFC because I don't know if the "always at 0xc..." statement
> >> is 100% true for Linux, nor if this patch would break other stuff...
> > 
> > Yeah, I'm pretty dubious about this too, especially since I don't
> > entirely grasp what the load_elf() translation function is all about
> > anyway.
> 
> Well, AFAIK it's used to modify the addresses before the ELF loader uses
> the address for loading. For example if your ELF binary is linked at
> address 0x1000, the translate function would move your binary to
> 0x401000 instead so that it does not interfere with the SLOF firmware
> (which is loaded to address 0 IIRC).

This is correct, but the move isn't just to make sure we don't interfere
with SLOF, it's also to make sure we can load the kernel into main
memory. When the link address is 0xc...,  then we can't use vaddr ==
paddr. The Linux ppc64 kernel, for example, is linked at 0xc000000000000000.
So it gets pulled down with the mask, and then the 0x400000 offset is added
to get it above SLOF.

> 
> So I also think your fix here is wrong, Andrew. E.g. when you have a
> binary that is linked to address 0x1000, you don't want to bypass the
> translation step here since it then would clash with the firmware.

I set the unit test's text segment start at 0x400000, so QEMU still
loads it there with this patch, and thus it wouldn't clash with SLOF.
But, anyway, there's no need for SLOF in kvm-unit-tests, and I've
replaced it with a four byte boot loader, b 0x400000.

> 
> > That said, I suspect making your unit test assume a fixed load address
> > may not be the best idea - qemu or SLOF could change in future to move
> > things about, so it might be more robust to have your test copy itself
> > to address it wants to be at before executing.
> 

Well, the reason I want vaddr == paddr is to be able to run C code
without the MMU enabled, not to mention to avoid needing to do any sort
of reloc dance in what's supposed to be super simple code. I don't need
to worry about SLOF changing things, since I don't use it. If QEMU changes
the load address, then things will indeed break, but it would be a one line
Makefile fix in kvm-unit-tests to point the text segment to the new offset.

> +1
> 
> ... or you could try to get the elf_reloc code working for POWER, too
> (see include/hw/elf_ops.h). That way QEMU would take care of relocating
> your program. (you can peek at elf_apply_rela64() in
>  https://github.com/aik/SLOF/blob/master/lib/libelf/elf64.c
> if you want to know what basically has to be done for POWER relocations).
> 
>  Thomas
> 
> 

kvm-unit-tests doesn't load the unit test elf itself. It relies on QEMU's
-kernel parameter to get the "kernel" (the unit test) into memory.

Thanks,
drew



reply via email to

[Prev in Thread] Current Thread [Next in Thread]