Re: [Qemu-devel] [PATCH 6/6] kvm: Fix dirty tracking with large kernel page size

[Andreas Färber]

Am 04.03.2012 21:59, schrieb Alexander Graf:
> 
> 
> On 04.03.2012, at 21:31, Andreas Färber <address@hidden> wrote:
> 
>> Am 04.03.2012 21:25, schrieb Alexander Graf:
>>>
>>>
>>> On 04.03.2012, at 21:21, Andreas Färber <address@hidden> wrote:
>>>
>>>> Am 04.03.2012 19:46, schrieb Alexander Graf:
>>>>>
>>>>>
>>>>> On 04.03.2012, at 17:46, Andreas Färber <address@hidden> wrote:
>>>>>
>>>>>> Am 04.03.2012 12:53, schrieb Benjamin Herrenschmidt:
>>>>>>> On Sun, 2012-03-04 at 12:49 +0200, Avi Kivity wrote:
>>>>>>>> On 02/28/2012 11:48 PM, Benjamin Herrenschmidt wrote:
>>>>>>>>> On Tue, 2012-02-28 at 14:32 +0200, Avi Kivity wrote:
>>>>>>>>>
>>>>>>>>>> What if TARGET_PAGE_SIZE > getpagesize()?  Or is that impossible?
>>>>>>>>>
>>>>>>>>> We have yet to encounter such a case. It's not currently possible on
>>>>>>>>> power (some old embedded chips could do 1K and 2K page sizes in the 
>>>>>>>>> TLB
>>>>>>>>> iirc but we never supported that in Linux and it's being phased out in
>>>>>>>>> HW).
>>>>>>>>>
>>>>>>>>> I suggest that gets dealt with when/if it needs to, which means 
>>>>>>>>> probably
>>>>>>>>> never :-)
>>>>>>>>
>>>>>>>> Doesn't ppc support both 4k and 64k pages?  Suppose you run a 4k guest
>>>>>>>> on a 64k host?
>>>>>>>>
>>>>>>>> Maybe I'm misremembering or misunderstanding something.
>>>>>>
>>>>>>> TARGET_PAGE_SIZE in qemu is always 4k for powerpc, it's a compile time
>>>>>>> #define.
>>>>>>
>>>>>> Except for ppcemb-softmmu (1k), which is irrelevant for KVM AFAIU.
>>>>>>
>>>>>> Maybe just add an assert and be done with it?
>>>>>
>>>>> Assert for what? Linux page size of 64k is something perfectly normal on 
>>>>> ppc. The hardware can always do at least 4k maps however.
>>>>
>>>> g_assert(TARGET_PAGE_SIZE <= getpagesize())
>>>>
>>>> Just declare the above case as unsupported and abort if we encounter it.
>>>
>>> What I'm trying to tell you is that it's the default case on book3s ppc! ;)
>>
>> Exactly, which is why I'm saying just ignore the weird embedded case. :)
> 
> [...] So you're saying 'break for ppcemb'. Hrm. Not sure that'd be all that 
> great for 440, since there host pagesize is still 4k, but T_P_S is 1k.

Err, 1k <= 4k would still be supported. The way I see it, the only cases
breaking would be ppc with host page size < 4k, and ppcemb with host
page size < 1k (which I'm not aware of).

Is it realistic to expect virtualizing a Mac or pSeries to work on a
1k/2k bamboo? TCG would be unaffected AFAICT.

Andreas

-- 
SUSE LINUX Products GmbH, Maxfeldstr. 5, 90409 Nürnberg, Germany
GF: Jeff Hawn, Jennifer Guild, Felix Imendörffer; HRB 16746 AG Nürnberg