Re: [Qemu-devel] Re: Suggestion - trap window-close of VM

[Struan Bartlett]

I think there's two good points here.

Of course it's true that, just because *(EIP-1) == 0xF4 and IF is 
cleared, it cannot be deduced with certainty that the previous 
instruction was an HLT and not a multi-byte instruction ending in 0xF4. 
I was wrong about that. Something like 'cli' then 'mov al,0xf4' might 
have been the previous instructions. You are right.

However, as a practical matter, I think it's true that this can't be a 
very common scenario. Before I added the 'IF-cleared' test to my patch I 
was routinely experiencing false positives (i.e. when closing the 
window, the patch would judge the guest as having halted when it clearly 
wasn't). But since I added the IF-cleared test to the patch I haven't 
personally experienced one single false positive (i.e. it has never 
so-far judged the guest as having halted when it wasn't) - and I must 
have tested it over fifty times so far. I understand though that this 
may not yet be good enough reliability for some people, but if it helps 
anybody at all I'd be pleased.

Anyhow, I can't say I find this patch's halt-detection method very 
satisfactory except as a purely interim measure. A solution to the APM 
bug would be much preferable. I'm looking at the APM Bios Interface 
Specification v1.2 now 
(http://www.phamhoaiviet.com/archives%5Cebooks%5Capmv12.rtf)

Struan

Leonardo E. Reiter wrote:

> If I understand this correctly, it's specifically doing a halt with 
> interrupts disabled.  In theory that should just halt the processor 
> indefinitely, since interrupts cannot be received.  I can't imagine 
> this being a common scenario for any software other than when it is 
> completely shut down, unless of course you are explicitly expecting a 
> non-maskable interrupt, which cannot be disabled, during a halt 
> instruction.  Again, I don't think it's all that common.
>
> - Leo Reiter
>
> Lennert Buytenhek wrote:
>
> > So now you're implying that disabling interrupts is something that
> > only very rarely happens, for example when shutting down the machine?
> > I don't think that is a very realistic assumption.