On 2014-10-08 12:34, Paolo Bonzini wrote:
> Il 08/10/2014 12:29, Jan Kiszka ha scritto:
>>>> But it would write to the vmcs02, not to the shadow VMCS; the shadow
>>>> VMCS is active during copy_shadow_to_vmcs12/copy_vmcs12_to_shadow, and
>>>> at no other time. It is not clear to me how the VIRTUAL_INTR_PENDING
>>>> bit ended up from the vmcs02 (where it is perfectly fine) to the vmcs12.
>> Well, but somehow that bit ends up in vmcs12, that's a fact. Also that
>> the proble disappears when shadowing is disabled. Need to think about
>> the path again. Maybe there is just a bug, not a conceptual issue.
>
> Yeah, and at this point we cannot actually exclude a processor bug. Can
> you check that the bit is not in the shadow VMCS just before vmrun, or
> just after enable_irq_window?
>
> Having a kvm-unit-tests testcase could also be of some help.
As usual, this was a nasty race that involved some concurrent VCPUs and
proper host load, so hard to write unit tests...
diff --git a/arch/x86/kvm/vmx.c b/arch/x86/kvm/vmx.c
index 04fa1b8..d6bcaca 100644
--- a/arch/x86/kvm/vmx.c
+++ b/arch/x86/kvm/vmx.c
@@ -6417,6 +6417,8 @@ static void copy_shadow_to_vmcs12(struct vcpu_vmx *vmx)
const unsigned long *fields = shadow_read_write_fields;
const int num_fields = max_shadow_read_write_fields;
+ preempt_disable();
+
vmcs_load(shadow_vmcs);
for (i = 0; i < num_fields; i++) {
@@ -6440,6 +6442,8 @@ static void copy_shadow_to_vmcs12(struct vcpu_vmx *vmx)
vmcs_clear(shadow_vmcs);
vmcs_load(vmx->loaded_vmcs->vmcs);
+
+ preempt_enable();
}
static void copy_vmcs12_to_shadow(struct vcpu_vmx *vmx)
@@ -6457,6 +6461,8 @@ static void copy_vmcs12_to_shadow(struct vcpu_vmx *vmx)
u64 field_value = 0;
struct vmcs *shadow_vmcs = vmx->nested.current_shadow_vmcs;
+ preempt_disable();
+
vmcs_load(shadow_vmcs);
for (q = 0; q < ARRAY_SIZE(fields); q++) {
@@ -6483,6 +6489,8 @@ static void copy_vmcs12_to_shadow(struct vcpu_vmx *vmx)
vmcs_clear(shadow_vmcs);
vmcs_load(vmx->loaded_vmcs->vmcs);
+
+ preempt_enable();
}
/*
No proper patch yet because there might be a smarter approach without
using the preempt_disable() hammer. But the point is that we temporarily
load a vmcs without updating loaded_vmcs->vmcs. Now, if some other VCPU
is scheduling in right in the middle of this, the wrong vmcs will be
flushed and then reloaded - e.g. a non-shadow vmcs with that interrupt
window flag set...
Patch is currently under heavy load testing here, but it looks very good
as the bug was quickly reproducible before I applied it.
Jan
--
Siemens AG, Corporate Technology, CT RTC ITP SES-DE
Corporate Competence Center Embedded Linux
--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
