[Top][All Lists]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Qemu-ppc] [PATCH 02/10] ppc/xive: introduce a XiveTCTX pointer unde

From: David Gibson
Subject: Re: [Qemu-ppc] [PATCH 02/10] ppc/xive: introduce a XiveTCTX pointer under PowerPCCPU
Date: Fri, 4 Jan 2019 16:25:38 +1100
User-agent: Mutt/1.10.1 (2018-07-13)

On Thu, Jan 03, 2019 at 06:44:30PM +0100, Cédric Le Goater wrote:
> On 1/3/19 4:57 AM, David Gibson wrote:
> > On Wed, Jan 02, 2019 at 06:57:35AM +0100, Cédric Le Goater wrote:
> >> which will be used by the machine only when the XIVE interrupt mode is
> >> in use.
> > 
> > I don't love the idea of putting a hook this specific into the
> > PowerPCCPU structure, though it might be the easiest path in the short
> > term.
> > 
> > A couple of approaches: 1) revisit my changes to allow for a pointer
> > to machine-defined per-cpu data.  
> ok but we still need two different presenters objects, one for each
> mode.
> > or 2) do we actually need a cpu to tctx pointer.
> > 
> > Expanding on (2) - here you use the pointer to find the right TIMA
> > state to access,
> yes. 
> > but that could also be handled by having different TIMA IO instances 
> > and mapping those individually to cpu_as.  
> It might work for QEMU but I am not sure how to implement the KVM 
> backend with such a design. We only have a single ram ptr mapping 
> for the machine in the KVM case.
> There are a couple of places where we need to loop on the XiveTCTX 
> (presenter, KVM) and we use the QEMU CPU list and the cpu->tctx for 
> that. One of the reasons we introduced the cpu->intc pointer some 
> time ago was to get rid of the ICP array. 
> I don't think we want to introduce a XiveTCTX array again.
> > On the interrupt delivery side I think a tctx to cpu link will suffice.  
> yes. that we already have. It is mostly used by KVM in fact. 
> > For sPAPR there might be complications with translating cpu numbers in
> > hcalls to the right tctx.
> The XiveTCTX is not used by the hcalls. We are fine on that side.
> The double pointer solution is what I prefer today, but if you are 
> uncomfortable with it, we can come back to the previous where I was 
> allocating a XiveTCTX child under the CPU and switching the presenter 
> objects at reset. The only issue remaining was the child unparenting 
> in the unrealize() method.

Hm, yes, I see your point.  For now I'm applying patches 1-3, and hope
to clean it up later with a revised version of my machine specific
per-cpu patch when I get the chance.

David Gibson                    | I'll have my music baroque, and my code
david AT gibson.dropbear.id.au  | minimalist, thank you.  NOT _the_ _other_
                                | _way_ _around_!

Attachment: signature.asc
Description: PGP signature

reply via email to

[Prev in Thread] Current Thread [Next in Thread]