Re: [PATCH] kvm: optimize ISR lookups

["H. Peter Anvin" <hpa@zytor.com>]

On 05/21/2012 04:06 PM, Michael S. Tsirkin wrote:
> 
> I think the reason is __apic_read which now simply copies the registers
> out to guest, this code will become less straight-forward if it's not
> 1:1.
> 

It can still be 1:1, just drop the 12 bytes of completely useless
padding after each 32-bit datum.

>> And if you look deeper, then you'll notice that _ALL_ APIC registers
>> are on a 16 byte boundary (thanks Peter for pointing that out).
>>
>> So it's even more silly to have a 1:1 representation instead of
>> implementing the default emulated apic_read/write functions to access
>> (offset >> 2).
>>
>> And of course, that design decision causes lookups to be slow.
> 
> Yes, it might be one of the reasons why my patch helps so
> much: it adds a cache in front of this data structure.
> 

Well, *fix the fscking data structure first*.

> 
> So what you propose is in fact to rework apic registers at least for
> ISR,IRR,TMR to use a bitmap.
> I am fine with this suggestion but would like some feedback from kvm
> maintainers on where they want to go before I spend time on that.
> 

This should be a 20-minute hack.

	-hpa