[Qemu-devel] Re: [PATCH 2/4] KVM: Rework VCPU state writeback API

[Marcelo Tosatti <mtosatti@redhat.com>]

On Tue, Mar 02, 2010 at 05:31:09PM +0100, Jan Kiszka wrote:
> Marcelo Tosatti wrote:
> > On Tue, Mar 02, 2010 at 09:00:04AM +0100, Jan Kiszka wrote:
> >> Marcelo Tosatti wrote:
> >>> On Mon, Mar 01, 2010 at 07:10:30PM +0100, Jan Kiszka wrote:
> >>>> This grand cleanup drops all reset and vmsave/load related
> >>>> synchronization points in favor of four(!) generic hooks:
> >>>>
> >>>> - cpu_synchronize_all_states in qemu_savevm_state_complete
> >>>>   (initial sync from kernel before vmsave)
> >>>> - cpu_synchronize_all_post_init in qemu_loadvm_state
> >>>>   (writeback after vmload)
> >>>> - cpu_synchronize_all_post_init in main after machine init
> >>>> - cpu_synchronize_all_post_reset in qemu_system_reset
> >>>>   (writeback after system reset)
> >>>>
> >>>> These writeback points + the existing one of VCPU exec after
> >>>> cpu_synchronize_state map on three levels of writeback:
> >>>>
> >>>> - KVM_PUT_RUNTIME_STATE (during runtime, other VCPUs continue to run)
> >>>> - KVM_PUT_RESET_STATE   (on synchronous system reset, all VCPUs stopped)
> >>>> - KVM_PUT_FULL_STATE    (on init or vmload, all VCPUs stopped as well)
> >>>>
> >>>> This level is passed to the arch-specific VCPU state writing function
> >>>> that will decide which concrete substates need to be written. That way,
> >>>> no writer of load, save or reset functions that interact with in-kernel
> >>>> KVM states will ever have to worry about synchronization again. That
> >>>> also means that a lot of reasons for races, segfaults and deadlocks are
> >>>> eliminated.
> >>>>
> >>>> cpu_synchronize_state remains untouched, just as Anthony suggested. We
> >>>> continue to need it before reading or writing of VCPU states that are
> >>>> also tracked by in-kernel KVM subsystems.
> >>>>
> >>>> Consequently, this patch removes many cpu_synchronize_state calls that
> >>>> are now redundant, just like remaining explicit register syncs.
> >>>>
> >>>> Signed-off-by: Jan Kiszka <jan.kiszka@siemens.com>
> >>> Jan,
> >>>
> >>> This patch breaks system reset of WinXP.32 install (more easily
> >>> reproducible without iothread enabled).
> >>>
> >>> Screenshot attached.
> >>>
> >> Strange - no issues with qemu-kvm? Any special command line switch? /me
> >> goes scrounging for some installation XP32 CD in the meantime...
> > 
> > No issues with qemu-kvm. Could not spot anything obvious.
> > 
> 
> And, of course, my WinXP installation did not trigger any reset issue,
> even in non-iothreaded mode. :(

Try kvm-autotest. I could not reproduce easily in hand run either.

This is not it, but still looks wrong:

In real mode emulation, kvm_get_sregs returns TR segment values of VMX
fields, while KVM caches them in vmx_vcpu->rmode.

diff --git a/arch/x86/kvm/vmx.c b/arch/x86/kvm/vmx.c
index 14873b9..898173a 100644
--- a/arch/x86/kvm/vmx.c
+++ b/arch/x86/kvm/vmx.c
@@ -1822,13 +1822,23 @@ static u64 vmx_get_segment_base(struct kvm_vcpu *vcpu, int seg)
 static void vmx_get_segment(struct kvm_vcpu *vcpu,
 			    struct kvm_segment *var, int seg)
 {
+	struct vcpu_vmx *vmx = to_vmx(vcpu);
 	struct kvm_vmx_segment_field *sf = &kvm_vmx_segment_fields[seg];
 	u32 ar;
 
+	if (vmx->rmode.vm86_active && seg == VCPU_SREG_TR) {
+		var->base = vmx->rmode.tr.base;
+		var->limit = vmx->rmode.tr.limit;
+		var->selector = vmx->rmode.tr.selector;
+		ar = vmx->rmode.tr.ar;
+		goto ar;
+	}
+
 	var->base = vmcs_readl(sf->base);
 	var->limit = vmcs_read32(sf->limit);
 	var->selector = vmcs_read16(sf->selector);
 	ar = vmcs_read32(sf->ar_bytes);
+ar:
 	if ((ar & AR_UNUSABLE_MASK) && !emulate_invalid_guest_state)
 		ar = 0;
 	var->type = ar & 15;