Re: Re: Paravirtualizing bits of acpi access

[Jeremy Fitzhardinge <jeremy@goop.org>]

Bjorn Helgaas wrote:
> On Tuesday 24 March 2009 01:05:43 am Jeremy Fitzhardinge wrote:
>   
>> Tian, Kevin wrote:
>>     
>>> OK, then it's safe to avoid that change. I had thought that dom0's 0-1M
>>> is identity-mapped to machine 0-1M... :-)
>>>       
>> No, only the ISA 640k-1M region.
>>     
>
> I'm speaking out of turn here because I don't work on Xen or
> suspend/resume.  However, I do try to clean up random bits of
> ACPI, and I have to say this patch looks like a pain in the
> maintenance department.  Having tests for a specific hypervisor
> is unpleasant.  We don't want to end up with tests for a collection
> of hypervisors.

I agree.  If we start to see other hypervisor-specific changes in this 
area, we'd need to rethink this approach.  But I'm not inclined to add a 
layer of infrastructure to just deal with Xen. (IOW, abstract only when 
there's something to abstract.)

>   It looks like suspend becomes a weird hybrid of
> ACPI and Xen, which makes it harder to reason about future suspend
> changes.  And all this discussion about 640k-1M and dom0 identity
> mapping and "there's no special effort to remap it" and whether
> there are conflicts makes me nervous.  There's no way all those
> assumptions can be remembered or verified five years down the road.
>   

Well, I think Kevin was over-complicating things in his own mind.  The 
upshot is that the normal "running on bare metal" code can do its normal 
thing, and if we happen to be running under Xen we can make it a no-op.  
In other words, the acpi developers don't really need to worry about the 
"running under Xen case", for the most part.

The two changes this patch makes are, unfortunately, unavoidable:

   1. Turn the final "enter sleep" into a hypercall, so that Xen can do
      all the low-level context save/restore.  The normal baremetal case
      is still localized in acpica/hwsleep.c, but it (may) make an
      excursion into arch code to see if it should do something else, and
   2. Directly enter the sleep state, rather than save cpu context
      (which Xen does)

Though, come to think of it, perhaps there's no harm in letting the 
kernel do its own state-saving.  I'll check.

    J