2017-07-31 19:25-0700, Wanpeng Li:
> From: Wanpeng Li <wanpeng.li@xxxxxxxxxxx>
>
> ------------[ cut here ]------------
> WARNING: CPU: 5 PID: 2288 at arch/x86/kvm/vmx.c:11124 nested_vmx_vmexit+0xd64/0xd70 [kvm_intel]
> CPU: 5 PID: 2288 Comm: qemu-system-x86 Not tainted 4.13.0-rc2+ #7
> RIP: 0010:nested_vmx_vmexit+0xd64/0xd70 [kvm_intel]
> Call Trace:
> vmx_check_nested_events+0x131/0x1f0 [kvm_intel]
> ? vmx_check_nested_events+0x131/0x1f0 [kvm_intel]
> kvm_arch_vcpu_ioctl_run+0x5dd/0x1be0 [kvm]
> ? vmx_vcpu_load+0x1be/0x220 [kvm_intel]
> ? kvm_arch_vcpu_load+0x62/0x230 [kvm]
> kvm_vcpu_ioctl+0x340/0x700 [kvm]
> ? kvm_vcpu_ioctl+0x340/0x700 [kvm]
> ? __fget+0xfc/0x210
> do_vfs_ioctl+0xa4/0x6a0
> ? __fget+0x11d/0x210
> SyS_ioctl+0x79/0x90
> do_syscall_64+0x8f/0x750
> ? trace_hardirqs_on_thunk+0x1a/0x1c
> entry_SYSCALL64_slow_path+0x25/0x25
>
> This can be reproduced by booting L1 guest w/ 'noapic' grub parameter, which
> means that tells the kernel to not make use of any IOAPICs that may be present
> in the system.
>
> Actually external_intr variable in nested_vmx_vmexit() is the req_int_win
> variable passed from vcpu_enter_guest() which means that the L0's userspace
> requests an irq window. I observed the scenario (!kvm_cpu_has_interrupt(vcpu) &&
> L0's userspace reqeusts an irq window) is true, so there is no interrupt which
> L1 requires to inject to L2, we should not attempt to emualte "Acknowledge
> interrupt on exit" for the irq window requirement in this scenario.
>
> This patch fixes it by not attempt to emulate "Acknowledge interrupt on exit"
> if there is no L1 requirement to inject an interrupt to L2.
>
> Cc: Paolo Bonzini <pbonzini@xxxxxxxxxx>
> Cc: Radim Krčmář <rkrcmar@xxxxxxxxxx>
> Signed-off-by: Wanpeng Li <wanpeng.li@xxxxxxxxxxx>
> ---
> v1 -> v2:
> * update patch description
> * check nested_exit_intr_ack_set() first
>
> diff --git a/arch/x86/kvm/vmx.c b/arch/x86/kvm/vmx.c
> @@ -11118,8 +11118,9 @@ static void nested_vmx_vmexit(struct kvm_vcpu *vcpu, u32 exit_reason,
>
> vmx_switch_vmcs(vcpu, &vmx->vmcs01);
>
> - if ((exit_reason == EXIT_REASON_EXTERNAL_INTERRUPT)
> - && nested_exit_intr_ack_set(vcpu)) {
> + if (nested_exit_intr_ack_set(vcpu) &&
> + exit_reason == EXIT_REASON_EXTERNAL_INTERRUPT &&
> + kvm_cpu_has_interrupt(vcpu)) {
This would work as a solution, but I don't think it's the correct
behavior.
SDM says that with acknowledge interrupt on exit, bit 31 of the VM-exit
interrupt information (valid interrupt) is always set to 1 on
EXIT_REASON_EXTERNAL_INTERRUPT. We don't want to break hypervisors
expecting an interrupt in that case, so we should do a userspace VM exit
when the window is open and then inject the userspace interrupt with a
VM exit.
The simplest thing that came to my mind is to:
diff --git a/arch/x86/kvm/vmx.c b/arch/x86/kvm/vmx.c
index 39a6222bf968..9ad0c882c4f5 100644
--- a/arch/x86/kvm/vmx.c
+++ b/arch/x86/kvm/vmx.c
@@ -10687,7 +10687,8 @@ static int vmx_check_nested_events(struct kvm_vcpu *vcpu, bool external_intr)
return 0;
}
- if ((kvm_cpu_has_interrupt(vcpu) || external_intr) &&
+ if ((kvm_cpu_has_interrupt(vcpu) ||
+ (external_intr && !nested_exit_intr_ack_set(vcpu))) &&
nested_exit_on_intr(vcpu)) {
if (vmx->nested.nested_run_pending)
return -EBUSY;
but I think it could break more ... actually, why was the window closed?
kvm_vcpu_ready_for_interrupt_injection() checks vmx_interrupt_allowed()
in order to decide need for the window, but vmx_check_nested_events()
doesn't care about that at all, so the window might just appear closed.
Would the following hunk help too?
diff --git a/arch/x86/kvm/vmx.c b/arch/x86/kvm/vmx.c
index 39a6222bf968..7e6caa9c225d 100644
--- a/arch/x86/kvm/vmx.c
+++ b/arch/x86/kvm/vmx.c
@@ -5567,8 +5567,10 @@ static int vmx_nmi_allowed(struct kvm_vcpu *vcpu)
static int vmx_interrupt_allowed(struct kvm_vcpu *vcpu)
{
- return (!to_vmx(vcpu)->nested.nested_run_pending &&
- vmcs_readl(GUEST_RFLAGS) & X86_EFLAGS_IF) &&
+ if (is_guest_mode(vcpu))
+ return !to_vmx(vcpu)->nested.nested_run_pending;
+
+ return vmcs_readl(GUEST_RFLAGS) & X86_EFLAGS_IF &&
!(vmcs_read32(GUEST_INTERRUPTIBILITY_INFO) &
(GUEST_INTR_STATE_STI | GUEST_INTR_STATE_MOV_SS));
}
(It doesn't prevent malicious userspace from hitting the WARN, though.)
Thanks.
