[RFC] Handling VMEXITS with interrupts disabled?

[RFC] Handling VMEXITS with interrupts disabled?

[Anthony Liguori]

Howdy,

I had an idea for improving VMEXIT time by only saving/restoring 
sys{enter,call,ret} MSRs when exiting from kernel space code since as 
long as these instructions aren't executed on the host, everything 
should be fine.

This worked fine for MSR_IA32_SYSENTER* msrs, but not so much for the 
?TAR msrs.  I have hooks in vcpu_{get,put} but I suspect we're making a 
blocking call somewhere which is resulting in the scheduler being 
invoked (while we still have the vcpu).  This then results in generally 
badness in the host.

I'm still a little unclear about this though as I would think that if 
this could happen, it would create problems with VT since KVM another 
guest could run and overwrite the VMCS msr.

I'm I interpreting this correctly?  Does disabling preemption not 
guarantee that you won't potentially drop to userspace in your critical 
block?

Regards,

Anthony Liguori

-------------------------------------------------------------------------
This SF.net email is sponsored by DB2 Express
Download DB2 Express C - the FREE version of DB2 express and take
control of your XML. No limits. Just data. Click to get it now.
http://sourceforge.net/powerbar/db2/

Re: [RFC] Handling VMEXITS with interrupts disabled?

[Avi Kivity]

Anthony Liguori wrote:
> Howdy,
>
> I had an idea for improving VMEXIT time by only saving/restoring 
> sys{enter,call,ret} MSRs when exiting from kernel space code since as 
> long as these instructions aren't executed on the host, everything 
> should be fine.
>   

Yup.

> This worked fine for MSR_IA32_SYSENTER* msrs, but not so much for the 
> ?TAR msrs.  I have hooks in vcpu_{get,put} but I suspect we're making a 
> blocking call somewhere which is resulting in the scheduler being 
> invoked (while we still have the vcpu).  This then results in generally 
> badness in the host.
>   

You might compile the host with CONFIG_PREEMPT, which should get you a 
nice stacktrace after the "scheduling while atomic" message.

> I'm still a little unclear about this though as I would think that if 
> this could happen, it would create problems with VT since KVM another 
> guest could run and overwrite the VMCS msr.
>
> I'm I interpreting this correctly?  Does disabling preemption not 
> guarantee that you won't potentially drop to userspace in your critical 
> block?
>   

This is only checked under CONFIG_PREEMPT, otherwise preempt_disable 
(and get_cpu()) are no-ops.


I really need to enable CONFIG_PREEMPT myself...

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


-------------------------------------------------------------------------
This SF.net email is sponsored by DB2 Express
Download DB2 Express C - the FREE version of DB2 express and take
control of your XML. No limits. Just data. Click to get it now.
http://sourceforge.net/powerbar/db2/