On Tue, 20 Nov 2018 22:03:59 +0100 Christian Borntraeger <borntraeger@xxxxxxxxxx> wrote: > On 11/20/2018 12:33 PM, Cornelia Huck wrote: > > On Mon, 19 Nov 2018 18:25:25 +0100 > > Michael Mueller <mimu@xxxxxxxxxxxxx> wrote: > > > >> Do not call __deliver_io() for adapter interruptions already > >> pending in the IPM. That is a double effort. They will > >> be processed as soon the vcpu control is given to SIE. > >> > >> Signed-off-by: Michael Mueller <mimu@xxxxxxxxxxxxx> > >> --- > >> arch/s390/kvm/interrupt.c | 54 ++++++++++++++++++++++------------------------- > >> 1 file changed, 25 insertions(+), 29 deletions(-) > > > > I think this patch does what it says on the tin, but I'm a bit lost as > > to the why. (Might make more sense with the gib.) > > > > Currently, we are trying to process any I/O interrupts, even if we'd > > get them delivered via the gisa, when we're out of the SIE anyway. > > IIRC, while this looks a bit like a belt-and-suspenders approach, it > > also prevented performance problems when the vcpu did not go back into > > the SIE immediately (it even may exit to userspace > > In fact, doing the inject when in SIE is likely better performance-wise. > Right now we "inject" the floating interrupt and then we handle > the requests. That can actually mean that it could take a while > until the interrupt is actually noticed by the guest (when > in SIE). If you now have a 2nd CPU enabled this interrupt could > have been delivered to the guest much earlier but it is "stuck" in > the local CPU. Hm, yes. Do we see any different effects if we have a guest with only one cpu (or only one cpu enabled for I/O interrupts?) Or does all of this even out in practice? > > Also, if you're ignoring the I/O interrupts pending in the ipm, you may > > end up delivering interrupts with a lower priority (higher isc) first. > > I'm not sure that's what we want. > > FWIW, LPAR has the same relaxation regarding priorities of subclasses. Interesting to know, thanks. What about restart etc. interrupts, as David has noted?