Re: [PATCH] Fix emergency_restart (sysrq-b) with kvm loaded on Intel hosts

[Avi Kivity <avi@qumranet.com>]

Ingo Molnar wrote:
> * Avi Kivity <avi@qumranet.com> wrote:
>
>   
>> Enabling Intel VT has the curious side effect whereby the INIT signal 
>> is blocked.  Rather than comment on the wisdom of this side effect, 
>> this patch adds an emergency restart reboot notifier, and modifies the 
>> kvm reboot notifier to disable VT on emergency reboot.
>>     
>
> looks good to me - i was bitten by that problem on a testbox.
>   

I'm a little worried about making emergency restart more complex.

Another thing that worries me is that emergency_restart() doesn't reset 
the box -- it sends INIT.  We could do better by using the ACPI FADT 
reset register (hopefully that's connected to RESET).

The ACPI spec says:

> 4.7.3.6 Reset Register
>
> The optional ACPI reset mechanism specifies a standard mechanism that 
> provides a complete system reset.
> When implemented, this mechanism must reset the entire system. This 
> includes processors, core logic, all
> buses, and all peripherals. From an OSPM perspective, asserting the 
> reset mechanism is the logical
> equivalent to power cycling the machine. Upon gaining control after a 
> reset, OSPM will perform actions in
> like manner to a cold boot.
> The reset mechanism is implemented via an 8-bit register described by 
> RESET_REG in the FADT (always
> accessed via the natural alignment and size described in RESET_REG). 
> To reset the machine, software will
> write a value (indicated in RESET_VALUE in FADT) to the reset 
> register. The RESET_REG field in the
> FADT indicates the location of the reset register.
> The reset register may exist only in I/O space, Memory space, or in 
> PCI Configuration space on a function
> in bus 0. Therefore, the Address_Space_ID value in RESET_REG must be 
> set to I/O space, Memory space,
> or PCI Configuration space (with a bus number of 0). As the register 
> is only 8 bits, Register_Bit_Width
> must be 8 and Register_Bit_Offset must be 0.
> The system must reset immediately following the write to this 
> register. OSPM assumes that the processor
> will not execute beyond the write instruction. OSPM should execute 
> spin loops on the CPUs in the system
> following a write to this register.

Which seems to be what we want? Maybe we should just try acpi_reboot() 
before the other stuff.

>   Acked-by: Ingo Molnar <mingo@elte.hu>
>
> Seems best to merge this via the KVM tree, right?
>
>   

I'm happy to do that, if everyone feels the patch is fine.

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