Re: [Qemu-devel] [PATCH] kvm: align ram_size to page boundary

[Blue Swirl]

On Sun, Jun 17, 2012 at 12:54 PM, Avi Kivity <address@hidden> wrote:
> On 06/17/2012 03:43 PM, Blue Swirl wrote:
>> On Sun, Jun 17, 2012 at 11:51 AM, Avi Kivity <address@hidden> wrote:
>>> On 06/17/2012 02:47 PM, Jan Kiszka wrote:
>>>>>>
>>>>>> I think this should rather go into generic code.
>>>>>
>>>>> To be honest, I put this in kvm-specific code because vl.c doesn't have
>>>>> TARGET_PAGE_ALIGN.  Maybe we should have machine->page_size or
>>>>> machine->ram_alignment.
>>>>>
>>>>>> What sense does it make
>>>>>> to have partial pages with TCG?
>>>>>
>>>>> Why impose an artificial restriction?
>>>>
>>>> Beca...
>>>>
>>>>>
>>>>> (answer: to reduce differences among various accelerators)
>>>>>
>>>>
>>>> Oh, you found the answer. :)
>>>
>>> Reducing round-trips across the Internet.
>>>
>>>>
>>>> At least, it should be enforce for the x86 target, independent of the
>>>> accelerator.
>>>
>>> Yeah.  So there's machine->page_size or machine->ram_alignment.  Not
>>> sure which is best.
>>
>> The boards should make sure that the amount of RAM is feasible with
>> the board memory slots. It's not possible to put 256kb SIMMs to a slot
>> that expects 1GB DIMMs. We can allow some flexibility there though,
>> I'm not sure if the current chipsets would support very much memory if
>> we followed the docs to the letter.
>
> Right. And generally memory modules are sized a power of two, creating
> the silly "mega == 1048576" movement.
>
>>
>> Maybe strtosz() should just enforce 1MB granularity.
>
> strtosz() is much too general.  We could do it in vl.c without trouble.
>  However, it takes away our ability to emulate a "640k should be enough
> for everyone" machine.

Then how about current max of target page sizes: 8k? No machine should
want less than that.

>
>>
>> What about ballooning (memory hotplug?), can that reduce the memory by
>> smaller amount than page size?
>
> Ballooning removes individual pages, that has no effect on the slot size.
>
> --
> error compiling committee.c: too many arguments to function
>
>