RE: [Xen-devel] [PATCH 3/10] Add HVM support

["Tian, Kevin" <kevin.tian@xxxxxxxxx>]

>From: Keir Fraser [mailto:keir@xxxxxxxxxxxxx]
>Sent: 2007年7月10日 17:51
>
>Kevin,
>
>Is it documented anywhere in which CPU states VMCS synchronisation
>may be
>lost? I see it makes sense that S3 would do so since the CPUs are
>powered
>off in that state (although I don't see this documented anywhere -- maybe
>it's 'obvious'), but what about e.g., Deep Sleep (C states)?

I don't see either. Since S3 is triggered by I/O write to chipset, CPUs 
are immediately powered off as result and thus internal VMCS cache 
is also lost. All the CPU context favored by system software should 
be saved by system software before that I/O write. So I thought it as 
'obvious'. :-)

For deep sleep states, IPI is taken as break event to bring that 
CPU back to working state (C0).

>
>Shouldn't you VMXOFF before S3?
>

Yes, it should. Actually for normal S3 path, it's not necessary since all 
the CPUs are reset after resume. However it's required at least for:
        - failure to do S3 suspend
        - S1 support
In above two cases, CPUs are still powered while VMXON state 
makes physical CPU blocking INIT IPI which makes smp resume 
failed.

I just realized it recently. Do you like to piggyback fix together with 
this patch, or I may send out an incremental one later?

Thanks,
Kevin

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-devel