Re: [PATCH 05/15] Coalesce userspace/kernel irqchip interrupt injection logic.

From: Gleb Natapov <gleb@redhat.com>
To: Jan Kiszka <jan.kiszka@web.de>
Cc: Dmitry Eremin-Solenikov <dbaryshkov@gmail.com>,
	kvm@vger.kernel.org, Avi Kivity <avi@redhat.com>,
	Joerg Roedel <joerg.roedel@amd.com>,
	Alexander Graf <agraf@suse.de>,
	qemu-devel <qemu-devel@nongnu.org>
Subject: Re: [PATCH 05/15] Coalesce userspace/kernel irqchip interrupt injection logic.
Date: Sun, 19 Apr 2009 18:20:23 +0300	[thread overview]
Message-ID: <20090419152023.GC21459@redhat.com> (raw)
In-Reply-To: <49EB3DF5.6070601@web.de>

On Sun, Apr 19, 2009 at 05:06:29PM +0200, Jan Kiszka wrote:
> Jan Kiszka wrote:
> > Gleb Natapov wrote:
> >> On Sat, Apr 18, 2009 at 07:28:20PM +0300, Gleb Natapov wrote:
> >>>> So this patch may either expose a bug in the svm emulation of qemu or
> >>>> comes with a subtle regression that only triggers due to qemu's timing.
> >>>> This needs to be understood. Gleb, any progress on reproducing it on
> >>>> your side?
> >>>>
> >>> I reproduced it and I am debugging it. In my case the boot hangs on sti;hlt
> >>> sequence. Instrumentation thus far shows that at this point interrupts no longer
> >>> injected because ppr value is too big. Need to see why, but tpr handling
> >>> is not complete in qemu svm. May be this is the reason. Will know more
> >>> tomorrow.
> >>>
> >> I've looked into this and my conclusion is that if you are not going to
> >> develop SVM in qemu don't use it just yet.
> > 
> > We had a resource conflict regarding SVM capable AMD boxes and a tight
> > schedule, so we decided to pick qemu as initial development platform.
> > Turns out that this has was a bit too optimistic. :)
> > 
> >> QEMU doesn't handle exceptions
> >> during event injection properly. Actually it does not handle it at all,
> >> so if PF happens during interrupt injection interrupt is lost and, what
> >> worse, is never acked. If interrupt was high prio it blocks all other
> >> interrupts.
> >>
> >> The patch below adds exception handling during event injection. Valid
> >> flag removed from EVENTINJ only after successful injection and EVENTINJ
> >> is copied to EXITINTINFO on exit. Can you give it a try?
> > 
> > Ah, great, thanks. Will test.
> 
> I can confirm: patch below makes my kvm-in-qemu test case happy, too.
> Maybe you want to post this with changelog and signed-off to qemu-devel.
> 
Yeah, I'll reformat and submit.

--
			Gleb.