Re: [PATCH v5 09/10] KVM: Disable manual dirty log when dirty ring enabled

[Peter Xu]

On Mon, Mar 22, 2021 at 02:55:44PM +0100, Paolo Bonzini wrote:
> On 22/03/21 10:17, Keqian Zhu wrote:
> > Hi Peter,
> > 
> > On 2021/3/11 4:33, Peter Xu wrote:
> > > KVM_CAP_MANUAL_DIRTY_LOG_PROTECT2 is for KVM_CLEAR_DIRTY_LOG, which is 
> > > only
> > > useful for KVM_GET_DIRTY_LOG.  Skip enabling it for kvm dirty ring.
> > > 
> > > More importantly, KVM_DIRTY_LOG_INITIALLY_SET will not wr-protect all the 
> > > pages
> > > initially, which is against how kvm dirty ring is used - there's no way 
> > > for kvm
> > > dirty ring to re-protect a page before it's notified as being written 
> > > first
> > > with a GFN entry in the ring!  So when KVM_DIRTY_LOG_INITIALLY_SET is 
> > > enabled
> > > with dirty ring, we'll see silent data loss after migration.
> > I feel a little regret that dirty ring can not work with 
> > KVM_DIRTY_LOG_INITIALLY_SET ...
> > With KVM_DIRTY_LOG_INITIALLY_SET, we can speedup dirty log start. More 
> > important, we can
> > enable dirty log gradually. For write fault based dirty log, it greatly 
> > reduces the side
> > effect of dirty log over guest.
> > 
> > I hope we can put forward another similar optimization under dirty ring 
> > mode. :)
> 
> Indeed, perhaps (even though KVM_GET_DIRTY_LOG does not make sense with
> dirty ring) we could allow KVM_CLEAR_DIRTY_LOG.

Right, KVM_CLEAR_DIRTY_LOG is a good interface to reuse so as to grant
userspace more flexibility to explicit wr-protect some guest pages.  However
that'll need kernel reworks - obviously when I worked on the kernel part I
didn't notice this issue..

To make it a complete work, IMHO we'll also need QEMU to completely drop the
whole dirty bitmap in all the layers, then I'd expect the dirty ring idea as a
whole start to make more difference on huge vms.  They all just need some more
work which should be based on this series.

Shall we make it a "TODO" though?  E.g., I can add a comment here mentioning
about this issue.  I still hope we can have the qemu series lands soon first,
since it'll be bigger project to fully complete it.  Paolo?

Thanks,

-- 
Peter Xu