Re: Booting from PV disk driver (Was: Re: [PATCH 10/13] KVM: Wire up hypercall handler ..)

From: Avi Kivity <avi-atKUWr5tajBWk0Htik3J/w@public.gmane.org>
To: Arjan van de Ven <arjan-wEGCiKHe2LqWVfeAwA7xHQ@public.gmane.org>
Cc: kvm-devel-5NWGOfrQmneRv+LV9MX5uipxlwaOVQ5f@public.gmane.org
Subject: Re: Booting from PV disk driver (Was:	Re:	[PATCH	10/13] KVM: Wire up hypercall handler ..)
Date: Sun, 25 Feb 2007 08:52:03 +0200	[thread overview]
Message-ID: <45E13213.5070502@qumranet.com> (raw)
In-Reply-To: <1172385720.3265.12.camel-NIQFrBLA1CpScpXdPBN83iCwEArCW2h5@public.gmane.org>

Arjan van de Ven wrote:
>>> shared interrupts aren't a big deal in Linux. at all.
>>> In fact, sharing all PV interrupts to one number is a performance
>>> enhancement ;)
>>>   
>>>       
>> If you share a PV interrupt with a non-PV interrupt, then for each PV 
>> interrupt you have to check whether the non-PV interrupt fired.  That 
>> involves at least one expensive vmexit.
>>     
>
> hmm not really.
> the completion ring lives in guest memory, the guest can check it
> without vmexit entirely.
> The PV block driver only does 2 vm events normally (apart from boot time
> setup), one to inform the host new IO has been submitted in the submit
> ring, and one to receive an interrupt to get notified one or more
> completions are present in the completion ring. Nothing more....
>
>   

But doesn't the interrupt code have to check all interrupt sources on a 
shared line?  That means checking the non-PV interrupt as well, which 
involves a vmexit (for mmio read).

>> I agree that sharing PV interrupts is not very expensive (though I don't 
>> see why you call it an optimization - if you share 100 interrupts on one 
>> line you need to check 100 interrupt sources every time the interrupt 
>> fires).
>>     
>
> as long as the check is cheap that is actually fine; if in your 100
> sources scenario 10 have fired, you only get 1 vm event to deliver the
> irq, and then 100 polls of guest memory inside the guest. That should be
> cheaper than doing 10 real vm interrupts.
>   

It is possible (Xen does it) to batch interrupt delivery so that 
multiple interrupts can be delivered with one vmentry.  So you get the 
best of both worlds: one check for interrupt source, and one vmentry if 
multiple interrupts have fired.

-- 
Do not meddle in the internals of kernels, for they are subtle and quick to panic.

-------------------------------------------------------------------------
Take Surveys. Earn Cash. Influence the Future of IT
Join SourceForge.net's Techsay panel and you'll get the chance to share your
opinions on IT & business topics through brief surveys-and earn cash
http://www.techsay.com/default.php?page=join.php&p=sourceforge&CID=DEVDEV