Re: [PATCH v4 4/9] xen: introduce XEN_DOMCTL_devour

[Vitaly Kuznetsov <vkuznets@redhat.com>]

Julien Grall <julien.grall@linaro.org> writes:

> Hi Vitaly,
>
> On 03/12/2014 17:16, Vitaly Kuznetsov wrote:
>> New operation sets the 'recipient' domain which will recieve all
>
> s/recieve/receive/
>
>> memory pages from a particular domain and kills the original domain.
>>
>> Signed-off-by: Vitaly Kuznetsov <vkuznets@redhat.com>
>> ---
>> @@ -1764,13 +1765,32 @@ void free_domheap_pages(struct page_info *pg, unsigned int order)
>
> [..]
>
>> +        else
>> +        {
>> +            mfn = page_to_mfn(pg);
>> +            gmfn = mfn_to_gmfn(d, mfn);
>> +
>> +            page_set_owner(pg, NULL);
>> +            if ( assign_pages(d->recipient, pg, order, 0) )
>> +                /* assign_pages reports the error by itself */
>> +                goto out;
>> +
>> +            if ( guest_physmap_add_page(d->recipient, gmfn, mfn, order) )
>
> On ARM, mfn_to_gmfn will always return the mfn. This would result to
> add a 1:1 mapping in the recipient domain.
>
> But ... only DOM0 has its memory mapped 1:1. So this code may blow up
> the P2M of the recipient domain.

I know almost nothing about ARM so please bear with me. So, for a guest
domain the mapping is not 1:1 (so guest sees different addresses) but
mfn_to_gmfn() doesn't returs these addresses? I was under an impression
it is not x86-specific and I can see its usage in e.g. getdomaininfo(),
memory_exchange(),.. How can one figure out the mapping then?
Anyway, what I want to do here is: when this page is freed I want to
reassign it to our newly-created guest at the exact same address it was
mapped in the original domain. 

>
> I'm not an x86 expert, but this may also happen when the recipient
> domain is using translated page mode (i.e HVM/PVHM).

PVHVM is the main target here (as kexec is unsupported for PV) and it
kinda works. mfn_to_gmfn() returns gmfn != mfn.

BTW, what's the current state of affairs with kexec and ARM guest? I
suppose we should have similar problems: vcpu_info, event channels,
.. 

>
> Regards,

-- 
  Vitaly