Re: [PATCH v5 3/3] i386: Add notify VM exit support

[Peter Xu <peterx@redhat.com>]

On Wed, Aug 17, 2022 at 10:08:45AM +0800, Chenyi Qiang wrote:
> There are cases that malicious virtual machine can cause CPU stuck (due
> to event windows don't open up), e.g., infinite loop in microcode when
> nested #AC (CVE-2015-5307). No event window means no event (NMI, SMI and
> IRQ) can be delivered. It leads the CPU to be unavailable to host or
> other VMs. Notify VM exit is introduced to mitigate such kind of
> attacks, which will generate a VM exit if no event window occurs in VM
> non-root mode for a specified amount of time (notify window).
> 
> A new KVM capability KVM_CAP_X86_NOTIFY_VMEXIT is exposed to user space
> so that the user can query the capability and set the expected notify
> window when creating VMs. The format of the argument when enabling this
> capability is as follows:
>   Bit 63:32 - notify window specified in qemu command
>   Bit 31:0  - some flags (e.g. KVM_X86_NOTIFY_VMEXIT_ENABLED is set to
>               enable the feature.)
> 
> Because there are some concerns, e.g. a notify VM exit may happen with
> VM_CONTEXT_INVALID set in exit qualification (no cases are anticipated
> that would set this bit), which means VM context is corrupted. To avoid
> the false positive and a well-behaved guest gets killed, make this
> feature disabled by default. Users can enable the feature by a new
> machine property:
>     qemu -machine notify_vmexit=on,notify_window=0 ...

The patch looks sane to me; I only read the KVM interface, though.  Worth
add a section to qemu-options.hx?  It'll also be worthwhile to mention the
valid range of notify_window and meaning of zero (IIUC that's also a valid
input, just use the hardware default window size).

Thanks,

> 
> A new KVM exit reason KVM_EXIT_NOTIFY is defined for notify VM exit. If
> it happens with VM_INVALID_CONTEXT, hypervisor exits to user space to
> inform the fatal case. Then user space can inject a SHUTDOWN event to
> the target vcpu. This is implemented by injecting a sythesized triple
> fault event.

-- 
Peter Xu