From mboxrd@z Thu Jan 1 00:00:00 1970 Received: from eggs.gnu.org ([2001:4830:134:3::10]:49490) by lists.gnu.org with esmtp (Exim 4.71) (envelope-from ) id 1XRKJy-00082b-AB for qemu-devel@nongnu.org; Tue, 09 Sep 2014 08:15:00 -0400 Received: from Debian-exim by eggs.gnu.org with spam-scanned (Exim 4.71) (envelope-from ) id 1XRKJs-0003Uo-3H for qemu-devel@nongnu.org; Tue, 09 Sep 2014 08:14:54 -0400 Received: from mx1.redhat.com ([209.132.183.28]:40509) by eggs.gnu.org with esmtp (Exim 4.71) (envelope-from ) id 1XRKJr-0003UK-RN for qemu-devel@nongnu.org; Tue, 09 Sep 2014 08:14:48 -0400 Message-ID: <540EEF2C.9010607@redhat.com> Date: Tue, 09 Sep 2014 14:14:36 +0200 From: Paolo Bonzini MIME-Version: 1.0 References: <20140826071427.1672.48119.stgit@PASHA-ISP> <20140826071503.1672.32964.stgit@PASHA-ISP> <53FC5ADE.8030306@redhat.com> <002401cfc1f0$b5aefe50$210cfaf0$@Dovgaluk@ispras.ru> <53FDD09C.4040208@redhat.com> <002b01cfc1f7$4c954e90$e5bfebb0$@Dovgaluk@ispras.ru> <53FDDB7A.8020604@redhat.com> <000e01cfcc19$2302aba0$690802e0$@Dovgaluk@ispras.ru> In-Reply-To: <000e01cfcc19$2302aba0$690802e0$@Dovgaluk@ispras.ru> Content-Type: text/plain; charset=utf-8 Content-Transfer-Encoding: 7bit Subject: Re: [Qemu-devel] [PATCH 06/12] kvmvapic: fixing loading vmstate List-Id: List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , To: Pavel Dovgaluk , qemu-devel@nongnu.org Cc: zealot351@gmail.com, maria.klimushenkova@ispras.ru Il 09/09/2014 12:30, Pavel Dovgaluk ha scritto: >> From: Paolo Bonzini [mailto:pbonzini@redhat.com] >> Il 27/08/2014 15:03, Pavel Dovgaluk ha scritto: >>>>> Hmm, probably not. The bug would not be other timers accessing the >>>>> APIC, because that would also call apic_sync_vapic and the only effect >>>>> would be an extra useless synchronization. The bug would happen if the >>>>> APIC is accessed by the CPU before the timer has the occasion to run. >>> Sorry, but I don't understand which problem we will solve with apic_sync_vapic. >> >> Taking inspiration from what KVM does, the fix could be even simpler >> than a change state handler. run_on_cpu functions do not run while the >> VM is stopped, so the following should work: >> >> diff --git a/hw/intc/apic_common.c b/hw/intc/apic_common.c >> index ce3d903..81d1ad7 100644 >> --- a/hw/intc/apic_common.c >> +++ b/hw/intc/apic_common.c >> @@ -91,13 +91,20 @@ void apic_enable_tpr_access_reporting(DeviceState >> *dev, bool enable) >> } >> } >> >> +static void do_apic_enable_vapic(void *data) >> +{ >> + APICCommonState *s = APIC_COMMON(data); >> + APICCommonClass *info = APIC_COMMON_GET_CLASS(s); >> + >> + info->vapic_base_update(s); >> +} >> + >> void apic_enable_vapic(DeviceState *dev, hwaddr paddr) >> { >> APICCommonState *s = APIC_COMMON(dev); >> - APICCommonClass *info = APIC_COMMON_GET_CLASS(s); >> >> s->vapic_paddr = paddr; >> - info->vapic_base_update(s); >> + run_on_cpu(CPU(s->cpu), do_apic_enable_vapic, s); >> } > > I've tried this one and it doesn't work. > do_apic_enable_vapic runs on cpu, at the same time the VM state is loaded. > APIC state still remains broken because of this. You're right (in fact run_on_cpu is synchronous so the alternative would have been deadlock). A change state handler can work. I'll submit your patches to the migration maintainers, including an alternative fix for this VAPIC problem. Paolo