On Tue, 5 Mar 2019 13:35:27 +0100 Christian Borntraeger <borntraeger@xxxxxxxxxx> wrote: > On 05.03.2019 11:56, Cornelia Huck wrote: > > On Tue, 5 Mar 2019 05:30:00 -0500 > > Christian Borntraeger <borntraeger@xxxxxxxxxx> wrote: > > > >> Folks, > >> > >> this is a very simple variant to disable halt polling when the KVM host > >> is already running virtualized. We could imagine more complex variants > >> (like tuning down the halt polling value) but this seems to do the trick > >> for some kvm deployment scenarios on s390x. (e.g. having multiple LPARS > >> with KVMs that are in itself already overcommitted). > > > > Yes, it makes sense to disable halt polling in that scenario, but I > > wonder whether we can make "disable halt polling for high steal times" > > more architecture-independent (can we obtain steal time quickly in some > > kind of architecture-independent way?) > > Not sure. I think an per-architecture callback should be simple enough > Having it generic is likely going to make it more complicated than necessary. Fair enough. > > > > > Btw, power seems to have its own halt polling code; not sure if it > > makes sense there as well. > > > >> This still needs tuning and the right default value, but I want to start > >> the discussion. > > > > Maybe optionally log a stat? > > I will have a look. >
