From mboxrd@z Thu Jan  1 00:00:00 1970
From: Avi Kivity <avi@redhat.com>
Subject: Re: kvm vmload/vmsave vs tss.ist
Date: Thu, 25 Dec 2008 20:12:42 +0200
Message-ID: <4953CD1A.5090906@redhat.com>
References: <49539FD0.7070103@redhat.com> <20081225151757.GA25117@elte.hu> <4953AAE5.4000708@redhat.com> <20081225162107.GB14486@elte.hu> <20081225164232.GA25195@elte.hu> <4953C57F.70502@redhat.com> <20081225175821.GA23390@elte.hu>
Mime-Version: 1.0
Content-Type: text/plain; charset=ISO-8859-1; format=flowed
Content-Transfer-Encoding: 7bit
Cc: "H. Peter Anvin" <hpa@zytor.com>,
	Joerg Roedel <joerg.roedel@amd.com>,
	Benjamin Serebrin <benjamin.serebrin@amd.com>,
	linux-kernel <linux-kernel@vger.kernel.org>, kvm@vger.kernel.org,
	Alexander Graf <agraf@suse.de>,
	Arjan van de Ven <arjan@infradead.org>,
	Alexander van Heukelum <heukelum@fastmail.fm>
To: Ingo Molnar <mingo@elte.hu>
Return-path: <kvm-owner@vger.kernel.org>
Received: from mx2.redhat.com ([66.187.237.31]:36497 "EHLO mx2.redhat.com"
	rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP
	id S1754451AbYLYSMp (ORCPT <rfc822;kvm@vger.kernel.org>);
	Thu, 25 Dec 2008 13:12:45 -0500
In-Reply-To: <20081225175821.GA23390@elte.hu>
Sender: kvm-owner@vger.kernel.org
List-ID: <kvm.vger.kernel.org>

Ingo Molnar wrote:
>> I think it's enough to switch %rsp before incrementing irqcount, no?
>>     
>
> no - that would introduce a small race: if an exception (say an NMI or 
> MCE, or a debug trap) happens in that small window then the exception 
> context thinks that it's on the IRQ stack already, and would use the task 
> stack.
>
>   

I'm suggesting

    check irqcount
    if (wasnt_in_irq)
        rsp = irqstack
    ++irqcount

If the NMI happens before the increment, we'll switch the stack 
unconditionally, and if the NMI happens after the increment, then we 
won't switch the stack, but we're guaranteed to be on the irqstack 
anyway.  The window size is negative :)

Similarly, the exit path should be

    oldstack_reg = oldstack;
    --irqcount;
    rsp = oldstack_register;

To guarantee that by the time we decrement irqcount, we don't need the 
stack anymore.

-- 
I have a truly marvellous patch that fixes the bug which this
signature is too narrow to contain.