Re: HVM pvops failures (one more qemu-dm.strace obtained)

[Jeremy Fitzhardinge <jeremy@goop.org>]

Andrew Lyon wrote:
> On Thu, Feb 26, 2009 at 11:35 AM, Ian Jackson <Ian.Jackson@eu.citrix.com> wrote:
>   
>> Jeremy Fitzhardinge writes ("Re: [Xen-devel] HVM pvops failures (one more qemu-dm.strace obtained)"):
>>     
>>> I was hoping Ian J would pop up and say "Its broken like this!  Fix it
>>> like this!".
>>>       
>> Yes, when I have a moment I'll try it myself but I've been buried in
>> the bowels of a nightmare USB emulation.  Soon I'll know whether the
>> latter works and thus whether I have any time to dig into pvops.
>>
>> Ian.
>>
>>     
>
> Has anybody made any progress on the qemu-dm / hvm issue with pv_ops?
> I spent most of yesterday going through the strace logs but I cannot
> see why it fails as it does.
>   

Thanks for looking into this.  The crash is a SIGBUS in qemu, shortly 
after an mmap of /proc/xen/privcmd, which is a suspicious combination.  
The SIGBUS means we managed to create a mapping where the kernel thinks 
there are no backing pages (at least in part).  It would be useful to 
know if any memory accesses to the privcmd mapping succeed, and what the 
offset of the failing one is.  (Assuming the SIGBUS is related to a 
privcmd mapping at all.)

I don't think privcmd mappings can be completely broken, because the 
qemu backend for pvfb works fine, and I assume that uses that path to 
map the framebuffer...

Also, have you tried using stub domains?  They should have no dependence 
on the dom0 kernel at all, so in theory they should work regardless.

    J