On Mon, Apr 17 2023 at 16:50, Mark Rutland wrote:
> On Sat, Apr 15, 2023 at 01:44:49AM +0200, Thomas Gleixner wrote:
> I gave this a spin on arm64 (in a 64-vCPU VM on an M1 host), and it seems to
> work fine with a bunch of vCPUs being hotplugged off and on again randomly.
>
> FWIW:
>
> Tested-by: Mark Rutland <mark.rutland@xxxxxxx>
>
> I also hacked the code to have the dying CPU spin forever before the call to
> cpuhp_ap_report_dead(). In that case I see a warning, and that we don't call
> arch_cpuhp_cleanup_dead_cpu(), and that the CPU is marked as offline (per
> /sys/devices/system/cpu/$N/online).
Nice!
> As a tangent/aside, we might need to improve that for confidential compute
> architectures, and we might want to generically track cpus which might still be
> using kernel text/data. On arm64 we ensure that via our cpu_kill() callback
> (which'll use PSCI CPU_AFFINITY_INFO), but I'm not sure if TDX and/or SEV-SNP
> have a similar mechanism.
>
> Otherwise, a malicious hypervisor can pause a vCPU just before it leaves the
> kernel (e.g. immediately after the arch_cpuhp_cleanup_dead_cpu() call), wait
> for a kexec (or resuse of stack memroy), and unpause the vCPU to cause things
> to blow up.
There are a gazillion ways for a malicious hypervisor to blow up a
'squint enough to be confident' guest.
The real question is whether it can utilize such a blow up to extract
confidential information from the guest.
If not then it's just yet another way of DoS which is an "acceptable"
attack as it only affects availability but not confidentiality.
Thanks,
tglx
