2008-08-18 13:39:48

by Avi Kivity

[permalink] [raw]
Subject: Re: [Bug 11219] KVM modules break emergency reboot

(copying mingo)

(context: sysrq-B with kvm-intel.ko loaded doesn't work. on my machine,
it kills the sata interface, but the processor and network keeps working)

Strangely, the specs say:

> ? The INIT signal is blocked whenever a logical processor is in VMX
> root operation.
> It is not blocked in VMX non-root operation. Instead, INITs cause VM
> exits (see
> Section 21.3, ?Other Causes of VM Exits?).

So INIT (which is wired to the triple-fault processor output, it seems,
rather than RESET) is blocked and the machine is not reset completely.

So we need to disable vmx during native_machine_emergency_restart().
There are at least three ways of doing this:

- add a vmxoff sequence (with an exception handler) to
native_machine_emergency_restart(). while simplest, this will not
unblock INIT for other cpus

- add an emergency_restart notifier_block, and have kvm subscribe. This
has the disadvantage of being slightly complex, opening a tiny race
(emergency restart during kvm module initialization), and requiring IPIs
during emergency restart.

- move vmxon/vmxoff management out of the kvm module and into x86 core.
Bloats the core but reduces complexity. IPIs still required.

I think the notifier block is the way to go. Ingo, let me know what you
prefer.

--
error compiling committee.c: too many arguments to function


2008-08-18 14:04:40

by Ingo Molnar

[permalink] [raw]
Subject: Re: [Bug 11219] KVM modules break emergency reboot


* Avi Kivity <[email protected]> wrote:

> (copying mingo)
>
> (context: sysrq-B with kvm-intel.ko loaded doesn't work. on my machine,
> it kills the sata interface, but the processor and network keeps working)
>
> Strangely, the specs say:
>
>> • The INIT signal is blocked whenever a logical processor is in VMX
>> root operation.
>> It is not blocked in VMX non-root operation. Instead, INITs cause VM
>> exits (see
>> Section 21.3, “Other Causes of VM Exits”).
>
> So INIT (which is wired to the triple-fault processor output, it seems,
> rather than RESET) is blocked and the machine is not reset completely.
>
> So we need to disable vmx during native_machine_emergency_restart().
> There are at least three ways of doing this:
>
> - add a vmxoff sequence (with an exception handler) to
> native_machine_emergency_restart(). while simplest, this will not
> unblock INIT for other cpus
>
> - add an emergency_restart notifier_block, and have kvm subscribe. This
> has the disadvantage of being slightly complex, opening a tiny race
> (emergency restart during kvm module initialization), and requiring IPIs
> during emergency restart.
>
> - move vmxon/vmxoff management out of the kvm module and into x86 core.
> Bloats the core but reduces complexity. IPIs still required.
>
> I think the notifier block is the way to go. Ingo, let me know what you
> prefer.

notifier should be OK i think - sysrq-b is an emergency mechanism after
all.

btw., "echo b > /proc/sysrq-trigger" never worked reliably for me with
KVM also loaded.

Ingo