Re: [Qemu-devel] [PATCH v2 0/2] kvm: x86 CPU power management

[Michael S. Tsirkin]

On Fri, Jun 15, 2018 at 03:32:27PM -0300, Marcelo Tosatti wrote:
> On Thu, Jun 14, 2018 at 02:37:28AM +0300, Michael S. Tsirkin wrote:
> > On Wed, Jun 13, 2018 at 07:34:53PM -0300, Marcelo Tosatti wrote:
> > > On Tue, Jun 12, 2018 at 09:47:11PM +0300, Michael S. Tsirkin wrote:
> > > > This adds ability to expose host CPU power management capabilities to
> > > > guests. For intel guests, this is sufficient for guest to enable
> > > > low power CPU power management. For AMD guests it isn't sufficient,
> > > > deeper C-states are entered using System-IO.
> > > > 
> > > > mwait based power management is tied closely to specifics of CPUID,
> > > > making migration challenging. At this point only the non-migrateable
> > > > -cpu host is supported.
> > > > 
> > > > With this patch applied, VM latency is within the noise of
> > > > baremetal for some benchmarks.
> > > > 
> > > > perf bench sched pipe results:
> > > > Before:
> > > >     6.452 sec
> > > > After:
> > > >     4.382 sec
> > > > Baremetal:
> > > >     4.136 sec
> > > > 
> > > > Michael S. Tsirkin (2):
> > > >   kvm: support -realtime cpu-pm=on|off
> > > >   i386/cpu: make -cpu host support monitor/mwait
> > > > 
> > > >  include/sysemu/sysemu.h |  1 +
> > > >  target/i386/cpu.h       |  9 +++++++++
> > > >  target/i386/cpu.c       | 19 ++++++++++++++-----
> > > >  target/i386/kvm.c       | 30 ++++++++++++++++++++++++++++++
> > > >  vl.c                    |  6 ++++++
> > > >  qemu-options.hx         |  9 +++++++--
> > > >  6 files changed, 67 insertions(+), 7 deletions(-)
> > > > 
> > > > -- 
> > > > MST
> > > 
> > > Hi Michael,
> > > 
> > > 1) Command line option interface
> > > 
> > > Why is this not an optional cpu feature such as the other features? 
> > > 
> > > 
> > > -cpu CPU,+mwait  
> > > 
> > > rather than a separate, architecture independent "-realtime 
> > > cpu-pm=on|off" 
> > > command line option?
> > 
> > Because it's not just a guest flag. With guest pm on, one guest
> > can severely affect the latency of others on the same host CPU.
> 
> How so ?

Look at drivers/idle/intel_idle.c
There are states with exit latencies of 
10000

> > > 2) Migration
> > > 
> > > Isnt it sufficient to check that both CPUID leafs are the same, 
> > > to allow migration ?
> > 
> > Not at the moment since linux guests use mwait hints and latency values
> > from a table in intel_idle.  If the host and guest models do not match,
> > mwait will get a wrong hint.
> > 
> > It will not do the right thing then!
> > 
> > You want exactly the same host CPU for it to work.
> > 
> > This isn't different from how -host cpu works generally.
> 
> Ok, makes sense.
> 
> > > 1. Check that the processor supports MONITOR and MWAIT. If
> > > CPUID.01H:ECX.MONITOR[bit 3] = 1, MONITOR and MWAIT are available at
> > > ring 0.
> > > 
> > > 2. Query the smallest and largest line size that MONITOR uses.
> > > Use CPUID.05H:EAX.smallest[bits 15:0];EBX.largest[bits15:0]. 
> > >