Re: [Qemu-devel] [PATCH] KVM: MMU: lazily drop large spte

[Marcelo Tosatti <mtosatti@redhat.com>]

On Wed, Nov 14, 2012 at 12:33:50AM +0900, Takuya Yoshikawa wrote:
> Ccing live migration developers who should be interested in this work,
> 
> On Mon, 12 Nov 2012 21:10:32 -0200
> Marcelo Tosatti <mtosatti@redhat.com> wrote:
> 
> > On Mon, Nov 05, 2012 at 05:59:26PM +0800, Xiao Guangrong wrote:
> > > Do not drop large spte until it can be insteaded by small pages so that
> > > the guest can happliy read memory through it
> > > 
> > > The idea is from Avi:
> > > | As I mentioned before, write-protecting a large spte is a good idea,
> > > | since it moves some work from protect-time to fault-time, so it reduces
> > > | jitter.  This removes the need for the return value.
> > > 
> > > Signed-off-by: Xiao Guangrong <xiaoguangrong@linux.vnet.ibm.com>
> > > ---
> > >  arch/x86/kvm/mmu.c |   34 +++++++++-------------------------
> > >  1 files changed, 9 insertions(+), 25 deletions(-)
> > 
> > Its likely that other 4k pages are mapped read-write in the 2mb range 
> > covered by a read-only 2mb map. Therefore its not entirely useful to
> > map read-only. 
> > 
> > Can you measure an improvement with this change?
> 
> What we discussed at KVM Forum last week was about the jitter we could
> measure right after starting live migration: both Isaku and Chegu reported
> such jitter.
> 
> So if this patch reduces such jitter for some real workloads, by lazily
> dropping largepage mappings and saving read faults until that point, that
> would be very nice!
> 
> But sadly, what they measured included interactions with the outside of the
> guest, and the main cause was due to the big QEMU lock problem, they guessed.
> The order is so different that an improvement by a kernel side effort may not
> be seen easily.
> 
> FWIW: I am now changing the initial write protection by
> kvm_mmu_slot_remove_write_access() to rmap based as I proposed at KVM Forum.
> ftrace said that 1ms was improved to 250-350us by the change for 10GB guest.
> My code still drops largepage mappings, so the initial write protection time
> itself may not be a such big issue here, I think.
> 
> Again, if we can eliminate read faults to such an extent that guests can see
> measurable improvement, that should be very nice!
> 
> Any thoughts?
> 
> Thanks,
> 	Takuya

OK, makes sense. I'm worried about shadow / oos interactions 
with large read-only mappings (trying to remember what was the 
case exactly, it might be non-existant now).