Re: [Xen-devel] [PATCH v6 01/15] x86: properly calculate ELF end of image address

[Daniel Kiper <daniel.kiper@xxxxxxxxxx>]

On Mon, Sep 19, 2016 at 08:52:02AM -0600, Jan Beulich wrote:
> >>> On 19.09.16 at 15:56, <daniel.kiper@xxxxxxxxxx> wrote:
> > On Mon, Sep 19, 2016 at 05:14:07AM -0600, Jan Beulich wrote:

[...]

> >> So before taking this patch I'd really like to see proof that what gets
> >> done currently does actually go wrong in at least one case. So far I
> >
> > During initial work on this patch series I discovered that p_memsz in xen
> > ELF file is set to unreasonably huge value, IIRC, ~1 GiB. That is why at
> > first I enforced 16 MiB here (just temporary workaround) and later proposed
> > this patch. Sadly, right now I am not able to find commit which introduced
> > this issue. However, it seems that it was "fixed" later by another commit.
> >
> > Anyway, I am not sure why are you against, IMO, more reliable solution.
> > This way we would avoid in the future similar issues which I described
> > above.
>
> I'm not against anything if it gets properly explained. Here,
> however, you present some vague statements which even you
> can't verify right now as it looks.

OK, I did some more tests and found out that after patch "efi: build
xen.gz with EFI code" we have following xen ELF file:

Program Headers:
  Type           Offset   VirtAddr   PhysAddr   FileSiz  MemSiz     Flg Align
  LOAD           0x000080 0x00100000 0x00100000 0x20c120 0x3ff00000 RWE 0x40
                                                         ^^^^^^^^^^

because "nm -nr xen/xen-syms" emits:

ffff82d0c0000000 A ALT_START
ffff82d08034d000 A __2M_rwdata_end
ffff82d08034cc00 A _end
ffff82d08034cc00 B __per_cpu_data_end
ffff82d08034cc00 B __bss_end

[...]

ALT_START lives in xen/arch/x86/efi/relocs-dummy.S. relocs-dummy.S
provides __base_relocs_start and __base_relocs_end symbols which
are referenced in xen/arch/x86/efi/efi-boot.h:efi_arch_relocate_image().
Of course one can argue that maybe we should do some changes in
efi_arch_relocate_image() and/or xen/arch/x86/efi/relocs-dummy.S.
This is true. However, this is separate issue and we should consider
it as such.

"efi: build xen.gz with EFI code" patch clearly shows that current
method used to calculate <final-exec-addr> mkelf32 argument is based
on bogus assumptions and very fragile. So, IMO, it should be changed
to something which is more reliable. And my proposal looks quite good.

Daniel

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
https://lists.xen.org/xen-devel