From mboxrd@z Thu Jan 1 00:00:00 1970 Received: from eggs.gnu.org ([208.118.235.92]:35650) by lists.gnu.org with esmtp (Exim 4.71) (envelope-from ) id 1TJlRS-00035a-5H for qemu-devel@nongnu.org; Thu, 04 Oct 2012 09:26:22 -0400 Received: from Debian-exim by eggs.gnu.org with spam-scanned (Exim 4.71) (envelope-from ) id 1TJlRL-00013a-MM for qemu-devel@nongnu.org; Thu, 04 Oct 2012 09:26:18 -0400 Received: from mx1.redhat.com ([209.132.183.28]:4580) by eggs.gnu.org with esmtp (Exim 4.71) (envelope-from ) id 1TJlRL-00013E-Dp for qemu-devel@nongnu.org; Thu, 04 Oct 2012 09:26:11 -0400 Date: Thu, 4 Oct 2012 15:25:59 +0200 From: Igor Mammedov Message-ID: <20121004152559.7e2993cd@nial.usersys.redhat.com> In-Reply-To: <20121004131027.GW15784@otherpad.lan.raisama.net> References: <1349192235-31895-8-git-send-email-imammedo@redhat.com> <20121002203845.GB4362@otherpad.lan.raisama.net> <20121003150339.GS15784@otherpad.lan.raisama.net> <506C57CE.9060002@redhat.com> <20121003182411.7075c0ef@nial.usersys.redhat.com> <20121003165434.GT15784@otherpad.lan.raisama.net> <20121004085322.599a19ab@nial.usersys.redhat.com> <20121004124341.GU15784@otherpad.lan.raisama.net> <20121004150119.5479fe32@nial.usersys.redhat.com> <20121004131027.GW15784@otherpad.lan.raisama.net> Mime-Version: 1.0 Content-Type: text/plain; charset=US-ASCII Content-Transfer-Encoding: 7bit Subject: Re: [Qemu-devel] [PATCH 07/23] target-i386: convert cpuid features into properties List-Id: List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , To: Eduardo Habkost Cc: aliguori@us.ibm.com, hpa@linux.intel.com, stefanha@linux.vnet.ibm.com, gleb@redhat.com, jan.kiszka@siemens.com, Don@cloudswitch.com, mtosatti@redhat.com, qemu-devel@nongnu.org, mdroth@linux.vnet.ibm.com, blauwirbel@gmail.com, avi@redhat.com, Paolo Bonzini , akong@redhat.com, lersek@redhat.com, afaerber@suse.de On Thu, 4 Oct 2012 10:10:27 -0300 Eduardo Habkost wrote: > On Thu, Oct 04, 2012 at 03:01:19PM +0200, Igor Mammedov wrote: > > On Thu, 4 Oct 2012 09:43:41 -0300 > > Eduardo Habkost wrote: > > > > > On Thu, Oct 04, 2012 at 08:53:22AM +0200, Igor Mammedov wrote: > > > > On Wed, 3 Oct 2012 13:54:34 -0300 > > > > Eduardo Habkost wrote: > > > > > > > > > On Wed, Oct 03, 2012 at 06:24:11PM +0200, Igor Mammedov wrote: > > > > > > On Wed, 03 Oct 2012 17:20:46 +0200 > > > > > > Paolo Bonzini wrote: > > > > > > > > > > > > > Il 03/10/2012 17:03, Eduardo Habkost ha scritto: > > > > > > > > On Tue, Oct 02, 2012 at 05:38:45PM -0300, Eduardo Habkost > > > > > > > > wrote: > > > > > > > >> (Now replying on the right thread, to keep the discussion in > > > > > > > >> the right place. I don't know how I ended up replying to a > > > > > > > >> pre-historic version of the patch, sorry.) > > > > > > > >> > > > > > > > >> On Tue, Oct 02, 2012 at 05:36:59PM +0200, Igor Mammedov > > > > > > > >> wrote: [...] > > > > > > > >>> @@ -1938,6 +2043,12 @@ static void x86_cpu_initfn(Object > > > > > > > >>> *obj) object_property_add(obj, "tsc-frequency", "int", > > > > > > > >>> x86_cpuid_get_tsc_freq, > > > > > > > >>> x86_cpuid_set_tsc_freq, NULL, NULL, > > > > > > > >>> NULL); > > > > > > > >>> + x86_register_cpuid_properties(obj, feature_name); > > > > > > > >>> + x86_register_cpuid_properties(obj, ext_feature_name); > > > > > > > >>> + x86_register_cpuid_properties(obj, ext2_feature_name); > > > > > > > >>> + x86_register_cpuid_properties(obj, ext3_feature_name); > > > > > > > >>> + x86_register_cpuid_properties(obj, kvm_feature_name); > > > > > > > >>> + x86_register_cpuid_properties(obj, svm_feature_name); > > > > > > > >> > > > > > > > >> Stupid question about qdev: > > > > > > > >> > > > > > > > >> - qdev_prop_set_globals() is called from device_initfn() > > > > > > > >> - device_initfn() is called before the child class > > > > > > > >> instance_init() function (x86_cpu_initfn()) > > > > > > > >> - So, qdev_prop_set_globals() gets called before the CPU > > > > > > > >> class properties are registered. > > > > > > > >> > > > > > > > >> So this would defeat the whole point of all the work we're > > > > > > > >> doing, that is to allow compatibility bits to be set as > > > > > > > >> machine-type global properties. But I don't know what's the > > > > > > > >> right solution here. > > > > > > > >> > > > > > > > >> Should the qdev_prop_set_globals() call be moved to > > > > > > > >> qdev_init() instead? Should the CPU properties be registered > > > > > > > >> somewhere else? > > > > > > > > > > > > > > Properties should be registered (for all objects, not just > > > > > > > CPUs) in the instance_init function. This is device_initfn. > > > > > > > > > > > > > > I would add an instance_postinit function that is called at the > > > > > > > end of object_initialize_with_type, that is after > > > > > > > instance_init, and in the opposite order (i.e. from the leaf to > > > > > > > the root). > > > > > > > > > > > > You've meant something like that? > > > > > > > > > > > > > > > > That's almost exactly the same code I wrote here. :-) > > > > > > > > > > The only difference is that I added post_init to the struct Object > > > > > documentation comments, and added a unit test. The unit test > > > > > required the qdev-core/qdev split, so we could compile it without > > > > > bringing too many dependencies. I will submit it soon. > > > > > > > > > After irc discussion, Anthony suggested to use static properties > > > > instead of dynamic ones that we use now. > > > > > > > > But qdev_prop_set_globals() in device_initfn() is still causes > > > > problems even with static properties. > > > > > > > > For x86 CPU classes we were going dynamically generate CPU classes and > > > > store pointer to appropriate cpudef from builtin_x86_defs in class > > > > field for each CPU class and then init default feature words values > > > > from this field int x86_cpu_initfn(). > > > > > > > > However with qdev_prop_set_globals() in device_initfn() that is called > > > > before x86_cpu_initfn() it won't work because defaults in > > > > x86_cpu_initfn() will overwrite whatever was set by > > > > qdev_prop_set_globals(). > > > > > > We can set the default values on class_init, instead. The class_init > > > function for each CPU model can get the x86_def_t struct as the data > > > pointer. > > > > > > I still think that the interface to build the DeviceClass.props array on > > > class_init is really painful to use, but it's still doable. > > > > You mean dynamic building of DeviceClass.props arrays for each CPU > > sub-class? > > That's the only solution I see if we want to make all the CPU properties > static, yes. Well I could generate compile time arrays for every built-in cpu model and we can remove then x86_def_t struct & builtins altogether. Only 'host' would be left for dynamic generation then. > > I'm still not convinced we really need to do that, though. Maybe we can > make static only the ones we really need to be able to implement > machine-type-compatibility global properties? > > Machine-type compatibility global properties were the initial reason for > the static-properties requirement. We don't really need to allow _all_ > CPU features to be controlled by global properties, only the ones we > need for machine-type compatibility. >