From mboxrd@z Thu Jan 1 00:00:00 1970 From: Ian Campbell Subject: Re: [PATCH] [RFC PATCH] xen: vmx: Use an INT 2 call to process real NMI's instead of self_nmi() in VMEXIT handler Date: Tue, 13 Nov 2012 12:17:29 +0000 Message-ID: <1352809049.7491.70.camel@zakaz.uk.xensource.com> References: <50A23F3102000078000A80BE@nat28.tlf.novell.com> <20121113114714.GB44675@ocelot.phlegethon.org> <20121113120555.GC44675@ocelot.phlegethon.org> Mime-Version: 1.0 Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 7bit Return-path: In-Reply-To: <20121113120555.GC44675@ocelot.phlegethon.org> List-Unsubscribe: , List-Post: List-Help: List-Subscribe: , Sender: xen-devel-bounces@lists.xen.org Errors-To: xen-devel-bounces@lists.xen.org To: Tim Deegan Cc: Malcolm Crossley , "eddie.dong@intel.com" , "jun.nakajima@intel.com" , Jan Beulich , xen-devel List-Id: xen-devel@lists.xenproject.org On Tue, 2012-11-13 at 12:05 +0000, Tim Deegan wrote: > At 11:47 +0000 on 13 Nov (1352807234), Tim Deegan wrote: > > At 11:38 +0000 on 13 Nov (1352806689), Jan Beulich wrote: > > > > diff -r 62885b3c34c8 -r ea756059a8da xen/arch/x86/hvm/vmx/vmx.c > > > > --- a/xen/arch/x86/hvm/vmx/vmx.c > > > > +++ b/xen/arch/x86/hvm/vmx/vmx.c > > > > @@ -2442,7 +2442,7 @@ void vmx_vmexit_handler(struct cpu_user_ > > > > (X86_EVENTTYPE_NMI << 8) ) > > > > goto exit_and_crash; > > > > HVMTRACE_0D(NMI); > > > > - self_nmi(); /* Real NMI, vector 2: normal processing. */ > > > > + asm("int $2"); /* Real NMI, vector 2: normal processing. */ > > > > > > In any case - why can't you call do_nmi() directly from here? > > > > ... this is my doing. There used to be a call to do_nmi() here, but > > do_nmi() doesn't block NMIs, so you can't just call it here in case you > > get _another_ NMI while you're in the NMI handler. > > Oh wait, I see -- you're saying that this (20059:76a65bf2aa4d) is wrong > because NMIs are indeed blocked, and have been since the VMEXIT. > > In that case, I agree that we should just run the NMI handler, but first > I would really like to know what _unblocks_ NMIs in this case. Any of > the things I can think of (the next vmenter, the next iret, ??) will > need some handling to make sure they actually happen before, say, we > take this CPU into the idle loop... What about a little stub-asm return_from_nmi / reenable_nmis with something like: pushf pushq $__HYPERVISOR_CS pushq 1f iret 1: ... That should re-enable NMIs at whichever point we think is appropriate? Perhaps a little more work is needed to create a suitable frame under this one too, not sure what a vmexit frame looks like. I hope we aren't going to get into the NMI nesting problem described here: http://lwn.net/Articles/484932/ (I don't think so, no page-faults or break points in our NMI handlers) Ian.