On 27/03/2017 17:46, Eduardo Habkost wrote:
On Mon, Mar 27, 2017 at 04:26:50PM +0200, Alexander Graf wrote:
KVM allows trap and emulate (read: NOP) of the MONITOR and MWAIT
instructions. There is work undergoing to enable actual execution
of these inside of KVM, but nobody really wants to expose the feature
to the guest by default, as it would eat up all of the host CPU.
Isn't this something that should be reported using
KVM_GET_EMULATED_CPUID? (QEMU still doesn't know how to use
KVM_GET_EMULATED_CPUID, however.)
Depends how you look at it. In KVM land there are basically 3 ways to
deal with MONITOR/MWAIT:
1) #VMEXIT on every execution, treat them as NOP
2) let the guest natively execute them (looks like a busy loop for
the host, but saves power)
3) be smart in KVM about it, add actual emulation and adaptively
allow for native mwait execution or emulated mwait which means we can
run inside host context
So today there is no streamlined way to actually notify the guest that
it's ok to execute MONITOR / MWAIT, even when we want to explicitly
leave the guest in guest context.
I'm not familiar with the variables involved in this decision.
How exactly would somebody (human or software) determine if it's
really ok to let the guest execute MONITOR / MWAIT?
Under what circumstances do you expect this to be used? Is this
just for debugging and development?
The main reason this is bubbling up at all are IPC intensive workloads.
Imagine you have the following:
CPU0 goes idle (waiting on something)
CPU1 wants to wake up CPU0 (because wait time is over)
In a normal KVM environment what you get is that
CPU0 calls HLT (going into KVM to do other work)
CPU1 triggers IPI via emulated APIC to wake up CPU0
However with actual native MWAIT what happens is
CPU0 calls MWAIT, stays in guest context ("wasting" CPU time)
CPU1 writes a byte to a memory location, waking up CPU0
With that scheme, you get your IPC latency down by a *huge* margin.
When is that useful? When you only have a single VM on your host for
example, so you want virtual machines for the sake of administration,
not for overprovisioning.
This patch adds a new -cpu parameter called "mwait" which - when
enabled - force enables the MONITOR / MWAIT CPUID flag, even when
the underlying accel framework does not explicitly advertise support.
If you really want something that makes QEMU ignore what the
accel code is reporting, I would prefer a syntax that could be
used for other features too, like "-cpu ...,monitor=force".
That sounds like a pretty nice idea and much more scalable. Let me see
if I can somehow pull that off :).
Alex
