Re: [libvirt] [Qemu-devel] CPU models and feature probing (was Re: [PATCH qom-cpu 00/16 v10] target-i386: convert CPU) features into properties

Am 11.02.2014 16:58, schrieb Anthony Liguori: > On Tue, Feb 11, 2014 at 7:25 AM, Eduardo Habkost <ehabkost(a)redhat.com&gt; wrote: >> On Tue, Feb 11, 2014 at 06:31:35AM -0800, Anthony Liguori wrote: >>> On Fri, Feb 7, 2014 at 2:55 AM, Paolo Bonzini <pbonzini(a)redhat.com&gt; wrote: >>>> Il 07/02/2014 11:16, Eduardo Habkost ha scritto: >>>> >>>>> You are not alone. I remember we spent lots of time trying to convince >>>>> Anthony to allow global properties and compat_props affect dynamic >>>>> properties not just static properties, and static properties were a big >>>>> deal due to reasons I didn't understand completely. Now I am hearing the >>>>> opposite message, and I don't understand the reasons for the change of >>>>> plans. I am confused. >>>> >>>> >>>> Picture me confused as well, but at the same I think I understand the >>>> reasons for the change of plans. >>> >>> There's no real convincing. It's just a question of code. >> >> I am sure there's a lot of convincing involved, even after the code is >> written (in this case, 15 months after the code was written). > > N.B. the code you refer to doesn't "make global propeties and > compat_props affect dynamic properties." It converts CPU properties > to static properties which I'm pretty sure I said many times is a > perfectly reasonable thing to do. > >>> There are >>> no defaults in classes for dynamic properties to modify. compat_props >>> are a nice mechanism, making them work for all properties is a >>> reasonable thing to do. >> >> That's exactly the opposite of what you said before[1]. But that isn't >> supposed to be a problem, I understand there may be change of plans (we >> should be able to change our minds). > > I think you're confusing a few things. You cannot make dynamic > properties work with globals today. Globals change class default > values and there are no class defaults for dynamic properties.[*] > > There's a perfectly valid discussion to have about whether we should > even have dynamic properties. It's certainly been a long time since > they were introduced and they haven't made their way into all that > many devices so it's reasonable to say that perhaps we'd be better off > without them. I would not object to a patch series that moved > properties to classes entirely provided it removed existing uses of > dynamic properties and didn't just introduce yet another mechanism. > > But compat properties as a concept could be made to work with dynamic > properties. They would have to be evaluated after instance init. > There's quite a few places they would end up touching I suspect. Erm, sorry, that is already implemented in qemu.git!? instance_post_init by Eduardo plus glue by me.

Andreas > > Another point of confusion worth mention is legacy properties since > this usually comes up in the discussion. Legacy properties (the > properties that are set/get as strings) are something that we should > try to avoid. They end up as strings on the wire and make it harder > to write client code. > > * I recognize that compat_props are implemented as globals. I'm > really talking about the current implementation of globals, not the > concept of -global which could be made with dynamic properties. > > Regards, > > Anthony Liguori > >> What I don't understand is the rejection of code that works, matches the >> style used by 200+ other source files, adds more useful introspectable >> information, done in the way that was suggested 16 months ago, because >> we have some rough idea about how a new grand design will look like in >> the far future. >> >> [1] http://lists.gnu.org/archive/html/qemu-devel/2013-06/msg00990.html >> >> -- >> Eduardo -- SUSE LINUX Products GmbH, Maxfeldstr. 5, 90409 Nürnberg, Germany GF: Jeff Hawn, Jennifer Guild, Felix Imendörffer; HRB 16746 AG Nürnberg