Re: [Qemu-devel] [PATCH v4 0/6] save/restore on Xen

[Anthony Liguori <anthony@codemonkey.ws>]

On 01/23/2012 04:47 AM, Stefano Stabellini wrote:
> On Fri, 20 Jan 2012, Jan Kiszka wrote:
>> On 2012-01-20 18:20, Stefano Stabellini wrote:
>>> Hi all,
>>> this is the fourth version of the Xen save/restore patch series.
>>> We have been discussing this issue for quite a while on #qemu and
>>> qemu-devel:
>>>
>>>
>>> http://marc.info/?l=qemu-devel&m=132346828427314&w=2
>>> http://marc.info/?l=qemu-devel&m=132377734605464&w=2
>>>
>>>
>>> A few different approaches were proposed to achieve the goal
>>> of a working save/restore with upstream Qemu on Xen, however after
>>> prototyping some of them I came up with yet another solution, that I
>>> think leads to the best results with the less amount of code
>>> duplications and ugliness.
>>> Far from saying that this patch series is an example of elegance and
>>> simplicity, but it is closer to acceptable anything else I have seen so
>>> far.
>>>
>>> What's new is that Qemu is going to keep track of its own physmap on
>>> xenstore, so that Xen can be fully aware of the changes Qemu makes to
>>> the guest's memory map at any time.
>>> This is all handled by Xen or Xen support in Qemu internally and can be
>>> used to solve our save/restore framebuffer problem.
>>>
>>> > From the Qemu common code POV, we still need to avoid saving the guest's
>>> ram when running on Xen, and we need to avoid resetting the videoram on
>>> restore (that is a benefit to the generic Qemu case too, because it
>>> saves few cpu cycles).
>>
>> For my understanding: Refraining from the memset is required as the
>> already restored vram would then be overwritten?
>
> Yep
>
>> Or what is the ordering
>> of init, RAM restore, and initial device reset now?
>
> RAM restore (done by Xen)
>
> physmap rebuild (done by xen_hvm_init in qemu)
> pc_init()
> qemu_system_reset()
> load_vmstate()

That's your problem.  You don't want to do load_vmstate().  You just want to 
load the device model, not RAM.

You should have a separate load_device_state() function and mark anything that 
is RAM as RAM when doing savevm registration.  Better yet, mark devices as 
devices since that's what you really care about.

Regards,

Anthony Liguori