| Message ID | 20190305103002.5801-1-borntraeger@de.ibm.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | disabling halt polling for nested virtualization | expand |
On 05.03.19 11:30, Christian Borntraeger 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). > This still needs tuning and the right default value, but I want to start > the discussion. > > > Christian Borntraeger (2): > KVM: polling: add architecture backend to disable polling > KVM: s390: provide kvm_arch_no_poll function > > arch/s390/include/asm/kvm_host.h | 6 ++++++ > arch/s390/kvm/Kconfig | 1 + > include/linux/kvm_host.h | 10 ++++++++++ > virt/kvm/Kconfig | 3 +++ > virt/kvm/kvm_main.c | 2 +- > 5 files changed, 21 insertions(+), 1 deletion(-) > Makes perfect sense to me.
On Tue, 5 Mar 2019 05:30:00 -0500 Christian Borntraeger <borntraeger@de.ibm.com> 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?) 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? > > > Christian Borntraeger (2): > KVM: polling: add architecture backend to disable polling > KVM: s390: provide kvm_arch_no_poll function > > arch/s390/include/asm/kvm_host.h | 6 ++++++ > arch/s390/kvm/Kconfig | 1 + > include/linux/kvm_host.h | 10 ++++++++++ > virt/kvm/Kconfig | 3 +++ > virt/kvm/kvm_main.c | 2 +- > 5 files changed, 21 insertions(+), 1 deletion(-) >
On 05.03.2019 11:56, Cornelia Huck wrote: > On Tue, 5 Mar 2019 05:30:00 -0500 > Christian Borntraeger <borntraeger@de.ibm.com> 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. > > 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.
On Tue, 5 Mar 2019 13:35:27 +0100 Christian Borntraeger <borntraeger@de.ibm.com> wrote: > On 05.03.2019 11:56, Cornelia Huck wrote: > > On Tue, 5 Mar 2019 05:30:00 -0500 > > Christian Borntraeger <borntraeger@de.ibm.com> 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. >
On 05/03/19 11:30, Christian Borntraeger 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). > This still needs tuning and the right default value, but I want to start > the discussion. > > > Christian Borntraeger (2): > KVM: polling: add architecture backend to disable polling > KVM: s390: provide kvm_arch_no_poll function > > arch/s390/include/asm/kvm_host.h | 6 ++++++ > arch/s390/kvm/Kconfig | 1 + > include/linux/kvm_host.h | 10 ++++++++++ > virt/kvm/Kconfig | 3 +++ > virt/kvm/kvm_main.c | 2 +- > 5 files changed, 21 insertions(+), 1 deletion(-) > Acked-by: Paolo Bonzini <pbonzini@redhat.com> Can you resubmit with a fix for the issue pointed out by the test robot? Thanks, Paolo
On 15.03.19 18:50, Paolo Bonzini wrote: > On 05/03/19 11:30, Christian Borntraeger 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). >> This still needs tuning and the right default value, but I want to start >> the discussion. >> >> >> Christian Borntraeger (2): >> KVM: polling: add architecture backend to disable polling >> KVM: s390: provide kvm_arch_no_poll function >> >> arch/s390/include/asm/kvm_host.h | 6 ++++++ >> arch/s390/kvm/Kconfig | 1 + >> include/linux/kvm_host.h | 10 ++++++++++ >> virt/kvm/Kconfig | 3 +++ >> virt/kvm/kvm_main.c | 2 +- >> 5 files changed, 21 insertions(+), 1 deletion(-) >> > > Acked-by: Paolo Bonzini <pbonzini@redhat.com> > > Can you resubmit with a fix for the issue pointed out by the test robot? Yes, this was based on a patch that was still missing in linux-next. Its now there. I will also add some instrumentation (kvm stat counter I think) in the next version.