From mboxrd@z Thu Jan 1 00:00:00 1970 Received: from eggs.gnu.org ([208.118.235.92]:60929) by lists.gnu.org with esmtp (Exim 4.71) (envelope-from ) id 1Sow1A-0001G5-IL for qemu-devel@nongnu.org; Wed, 11 Jul 2012 08:27:58 -0400 Received: from Debian-exim by eggs.gnu.org with spam-scanned (Exim 4.71) (envelope-from ) id 1Sow0g-0004WJ-8d for qemu-devel@nongnu.org; Wed, 11 Jul 2012 08:27:44 -0400 Received: from mx1.redhat.com ([209.132.183.28]:45323) by eggs.gnu.org with esmtp (Exim 4.71) (envelope-from ) id 1Sow0f-0004Vu-VO for qemu-devel@nongnu.org; Wed, 11 Jul 2012 08:27:14 -0400 Message-ID: <4FFD7118.7050905@redhat.com> Date: Wed, 11 Jul 2012 14:27:04 +0200 From: Igor Mammedov MIME-Version: 1.0 References: <1340197164-9574-1-git-send-email-imammedo@redhat.com> <1340197164-9574-4-git-send-email-imammedo@redhat.com> <4FE1D1BB.6030709@suse.de> <4FFAB99E.1030203@redhat.com> <4FFAD54D.2050909@suse.de> <4FFC2F8C.5000806@redhat.com> In-Reply-To: Content-Type: text/plain; charset=ISO-8859-1; format=flowed Content-Transfer-Encoding: quoted-printable Subject: Re: [Qemu-devel] [PATCH 3/5] target-i386: call x86_cpu_realize() after APIC is initialized. List-Id: List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , To: liu ping fan Cc: peter.maydell@linaro.org, aliguori@us.ibm.com, ehabkost@redhat.com, jan.kiszka@siemens.com, mtosatti@redhat.com, qemu-devel@nongnu.org, mdroth@linux.vnet.ibm.com, blauwirbel@gmail.com, avi@redhat.com, pbonzini@redhat.com, =?ISO-8859-1?Q?Andreas_F=E4rber?= On 07/11/2012 09:35 AM, liu ping fan wrote: > On Tue, Jul 10, 2012 at 9:35 PM, Igor Mammedov wr= ote: >> On 07/09/2012 02:57 PM, Andreas F=E4rber wrote: >>> >>> Am 09.07.2012 12:59, schrieb igor: >>>> >>>> On 06/20/2012 03:35 PM, Andreas F=E4rber wrote: >>>>> >>>>> Am 20.06.2012 14:59, schrieb Igor Mammedov: >>>>>> >>>>>> It's not correct to make CPU runnable (i.e. calling x86_cpu_realiz= e()) >>>>>> when not all properties are set (APIC in this case). >>>>>> >>>>>> Fix it by calling x86_cpu_realize() at board level after APIC is >>>>>> initialized, right before cpu_reset(). >>>>>> >>>>>> Signed-off-by: Igor Mammedov >>>>>> --- >>>>>> hw/pc.c | 1 + >>>>>> target-i386/helper.c | 2 -- >>>>>> 2 files changed, 1 insertion(+), 2 deletions(-) >>>>>> >>>>>> diff --git a/hw/pc.c b/hw/pc.c >>>>>> index 8368701..8a662cf 100644 >>>>>> --- a/hw/pc.c >>>>>> +++ b/hw/pc.c >>>>>> @@ -948,6 +948,7 @@ static X86CPU *pc_new_cpu(const char *cpu_mode= l) >>>>>> env->apic_state =3D apic_init(env, env->cpuid_apic_id)= ; >>>>>> } >>>>>> qemu_register_reset(pc_cpu_reset, cpu); >>>>>> + x86_cpu_realize(OBJECT(cpu), NULL); >>>>>> pc_cpu_reset(cpu); >>>>>> return cpu; >>>>>> } >>>>>> diff --git a/target-i386/helper.c b/target-i386/helper.c >>>>>> index c52ec13..b38ea7f 100644 >>>>>> --- a/target-i386/helper.c >>>>>> +++ b/target-i386/helper.c >>>>>> @@ -1161,8 +1161,6 @@ X86CPU *cpu_x86_init(const char *cpu_model) >>>>>> return NULL; >>>>>> } >>>>>> >>>>>> - x86_cpu_realize(OBJECT(cpu), NULL); >>>>>> - >>>>>> return cpu; >>>>>> } >>>>>> >>>>> >>>>> This will require changes in linux-user and possibly bsd-user. Havi= ng a >>>>> cpu_realize() would probably help with avoiding #ifdef'ery. >>>>> Unfortunately deriving CPUState from DeviceState proves a bit diffi= cult >>>>> in the meantime (it worked at one point, now there's lots of circul= ar >>>>> header dependencies), and realize support for Object got stopped. >>>>> >>>> As alternative to keep, I could leave x86_cpu_realize() in >>>> cpu_x86_init() and keep pc_cpu_reset() in pc_new_cpu(). That will re= sult >>>> in calling cpu_reset() 3 instead of 2 times. >>>> Later when apic_init is moved inside cpu.c, a pc_cpu_reset() in >>>> pc_new_cpu() would be unnecessary and could be cleaned up then. >>> >>> >>> Let me explain in more detail what I was thinking about: cpu_init() a= nd >>> cpu_x86_init() today return an initialized/realized object. I don't w= ant >>> bugs to creep into the user emulators because someone is not aware th= at >>> x86 is semantically differing from all other targets. >>> >>> What I did for a qemu-rl78 machine is to inline cpu_rl78_init() so th= at >>> I could put code in between, i.e., for x86: object_new(), APIC/BSP >>> stuff, x86_cpu_realize(). That way any addition to the realize functi= on >>> will still affect the user emulators. >>> The downside is that when we add x86 CPU subclasses we'd have to >>> remember to update two places. The solution to that would be to split >>> out a x86_cpu_new() function used from cpu_x86_init() and wherever yo= u >>> need it for the PC machine. Then the code is still maintainable in on= e >>> central place and you get to do your APIC cleanups, and we don't depe= nd >>> on a central realize implementation or device parent, what do you thi= nk? >> >> >> If you mean x86_cpu_new() =3D=3D pc_new_cpu() that calls cpu_x86_init(= ), >> then I'd like get rid of pc_new_cpu() completely, eventually replacing= it by >> cpu_x86_init() in hw/pc.c:pc_cpus_init(), something like this: >> >> -static X86CPU *pc_new_cpu(const char *cpu_model) >> -{ >> - X86CPU *cpu; >> - CPUX86State *env; >> - >> - cpu =3D cpu_x86_init(cpu_model); >> - if (cpu =3D=3D NULL) { >> - fprintf(stderr, "Unable to find x86 CPU definition\n"); >> - exit(1); >> - } >> - env =3D &cpu->env; >> - if ((env->cpuid_features & CPUID_APIC) || smp_cpus > 1) { >> - env->apic_state =3D apic_init(env, env->cpuid_apic_id); >> - } >> - cpu_reset(CPU(cpu)); >> - return cpu; >> -} >> - >> void pc_cpus_init(const char *cpu_model) >> { >> int i; >> @@ -950,7 +932,7 @@ void pc_cpus_init(const char *cpu_model) >> } >> >> for(i =3D 0; i < smp_cpus; i++) { >> - pc_new_cpu(cpu_model); >> + cpu_x86_init(cpu_model); >> } >> } >> >> goal I'm aiming at is to have a working cpu object that could be creat= ed >> using qdev_device_add without any adhoc calls. So in the end cpu_x86_i= nit() >> should become object_new(x86_cpu), [set props], realize() and nothing = else. > > Could we think apic's "creation + realize" as part of > x86_cpu_realize(), but not [set props]? > For the concept of building sub log unit inside chip. Yes, sure. Please look at https://github.com/imammedo/qemu/tree/x86_qom_apic it lacks apic_reset() from cpu_reset() but it is easy to add. > > Regards, > pingfan >> And maybe in some far future removing cpu_init -> cpu_x86_init() compl= etely. >> That would give us a single implementation of CPU one place (cpu.c) >> -- >> ----- >> Regards, >> Igor >> >> >> --=20 ----- Igor