Re: incorrect layout of globals from head_64.S during kexec boot

[Daniel Kiper <dkiper@net-space.pl>]

On Thu, Jul 05, 2012 at 11:06:07PM +0200, Olaf Hering wrote:
>
> During kexec in a Xen PVonHVM guest the new kernel crashes most of the
> time in secondary_startup_64 because the content of phys_base is
> corrupted. Its not zero as expected but has some random other values.
> While debugging that crash I came up with the change below to inspect
> the memory around phys_base.
>
> It turned out that the globales are not in the expected memory location.
> An expected value such as phys_base_plus1 is shifted, but by a different
> amount during repeated kexec attempts. Up to now I havent figured out
> where this happens.
>
> My question is: were to put additional debug to trace the copying of the
> data section to its final destination? Is this a task of kexec -l or
> does that happen during decompressing? I suspect the latter. This is the
> console output before the crash (the crash happens in 'movq %rax, %cr3'):

Copy is done a few times durnig kexec/kdump but the most important
in this case, I think, is in relocate_kernel() function (look for
rep movsl or rep movsq and code around it). But I am a bit surprised
that kernel is decompressing itself. I always thought that it is done
during kexec/kdump load phase but maybe I am wrong. Could you send
me more info about your Linux Kernel version, kexec-tools version
and exact commands which you are using to load/exececute kernel?

Daniel